Line data Source code
1 : //! Timeline repository implementation that keeps old data in layer files, and
2 : //! the recent changes in ephemeral files.
3 : //!
4 : //! See tenant/*_layer.rs files. The functions here are responsible for locating
5 : //! the correct layer for the get/put call, walking back the timeline branching
6 : //! history as needed.
7 : //!
8 : //! The files are stored in the .neon/tenants/<tenant_id>/timelines/<timeline_id>
9 : //! directory. See docs/pageserver-storage.md for how the files are managed.
10 : //! In addition to the layer files, there is a metadata file in the same
11 : //! directory that contains information about the timeline, in particular its
12 : //! parent timeline, and the last LSN that has been written to disk.
13 : //!
14 :
15 : use std::collections::hash_map::Entry;
16 : use std::collections::{BTreeMap, HashMap, HashSet};
17 : use std::fmt::{Debug, Display};
18 : use std::fs::File;
19 : use std::future::Future;
20 : use std::sync::atomic::{AtomicBool, AtomicU64, Ordering};
21 : use std::sync::{Arc, Mutex, Weak};
22 : use std::time::{Duration, Instant, SystemTime};
23 : use std::{fmt, fs};
24 :
25 : use anyhow::{Context, bail};
26 : use arc_swap::ArcSwap;
27 : use camino::{Utf8Path, Utf8PathBuf};
28 : use chrono::NaiveDateTime;
29 : use enumset::EnumSet;
30 : use futures::StreamExt;
31 : use futures::stream::FuturesUnordered;
32 : use itertools::Itertools as _;
33 : use once_cell::sync::Lazy;
34 : pub use pageserver_api::models::TenantState;
35 : use pageserver_api::models::{self, RelSizeMigration};
36 : use pageserver_api::models::{
37 : CompactInfoResponse, TimelineArchivalState, TimelineState, TopTenantShardItem,
38 : WalRedoManagerStatus,
39 : };
40 : use pageserver_api::shard::{ShardIdentity, ShardStripeSize, TenantShardId};
41 : use postgres_ffi::PgMajorVersion;
42 : use remote_storage::{DownloadError, GenericRemoteStorage, TimeoutOrCancel};
43 : use remote_timeline_client::index::GcCompactionState;
44 : use remote_timeline_client::manifest::{
45 : LATEST_TENANT_MANIFEST_VERSION, OffloadedTimelineManifest, TenantManifest,
46 : };
47 : use remote_timeline_client::{
48 : FAILED_REMOTE_OP_RETRIES, FAILED_UPLOAD_WARN_THRESHOLD, UploadQueueNotReadyError,
49 : download_tenant_manifest,
50 : };
51 : use secondary::heatmap::{HeatMapTenant, HeatMapTimeline};
52 : use storage_broker::BrokerClientChannel;
53 : use timeline::compaction::{CompactionOutcome, GcCompactionQueue};
54 : use timeline::import_pgdata::ImportingTimeline;
55 : use timeline::layer_manager::LayerManagerLockHolder;
56 : use timeline::offload::{OffloadError, offload_timeline};
57 : use timeline::{
58 : CompactFlags, CompactOptions, CompactionError, PreviousHeatmap, ShutdownMode, import_pgdata,
59 : };
60 : use tokio::io::BufReader;
61 : use tokio::sync::{Notify, Semaphore, watch};
62 : use tokio::task::JoinSet;
63 : use tokio_util::sync::CancellationToken;
64 : use tracing::*;
65 : use upload_queue::NotInitialized;
66 : use utils::circuit_breaker::CircuitBreaker;
67 : use utils::crashsafe::path_with_suffix_extension;
68 : use utils::sync::gate::{Gate, GateGuard};
69 : use utils::timeout::{TimeoutCancellableError, timeout_cancellable};
70 : use utils::try_rcu::ArcSwapExt;
71 : use utils::zstd::{create_zst_tarball, extract_zst_tarball};
72 : use utils::{backoff, completion, failpoint_support, fs_ext, pausable_failpoint};
73 :
74 : use self::config::{AttachedLocationConfig, AttachmentMode, LocationConf};
75 : use self::metadata::TimelineMetadata;
76 : use self::mgr::{GetActiveTenantError, GetTenantError};
77 : use self::remote_timeline_client::upload::{upload_index_part, upload_tenant_manifest};
78 : use self::remote_timeline_client::{RemoteTimelineClient, WaitCompletionError};
79 : use self::timeline::uninit::{TimelineCreateGuard, TimelineExclusionError, UninitializedTimeline};
80 : use self::timeline::{
81 : EvictionTaskTenantState, GcCutoffs, TimelineDeleteProgress, TimelineResources, WaitLsnError,
82 : };
83 : use crate::basebackup_cache::BasebackupCache;
84 : use crate::config::PageServerConf;
85 : use crate::context;
86 : use crate::context::RequestContextBuilder;
87 : use crate::context::{DownloadBehavior, RequestContext};
88 : use crate::deletion_queue::{DeletionQueueClient, DeletionQueueError};
89 : use crate::feature_resolver::{FeatureResolver, TenantFeatureResolver};
90 : use crate::l0_flush::L0FlushGlobalState;
91 : use crate::metrics::{
92 : BROKEN_TENANTS_SET, CIRCUIT_BREAKERS_BROKEN, CIRCUIT_BREAKERS_UNBROKEN, CONCURRENT_INITDBS,
93 : INITDB_RUN_TIME, INITDB_SEMAPHORE_ACQUISITION_TIME, TENANT, TENANT_OFFLOADED_TIMELINES,
94 : TENANT_STATE_METRIC, TENANT_SYNTHETIC_SIZE_METRIC, TIMELINE_STATE_METRIC,
95 : remove_tenant_metrics,
96 : };
97 : use crate::task_mgr::TaskKind;
98 : use crate::tenant::config::LocationMode;
99 : use crate::tenant::gc_result::GcResult;
100 : pub use crate::tenant::remote_timeline_client::index::IndexPart;
101 : use crate::tenant::remote_timeline_client::{
102 : INITDB_PATH, MaybeDeletedIndexPart, remote_initdb_archive_path,
103 : };
104 : use crate::tenant::storage_layer::{DeltaLayer, ImageLayer};
105 : use crate::tenant::timeline::delete::DeleteTimelineFlow;
106 : use crate::tenant::timeline::uninit::cleanup_timeline_directory;
107 : use crate::virtual_file::VirtualFile;
108 : use crate::walingest::WalLagCooldown;
109 : use crate::walredo::{PostgresRedoManager, RedoAttemptType};
110 : use crate::{InitializationOrder, TEMP_FILE_SUFFIX, import_datadir, span, task_mgr, walredo};
111 :
112 0 : static INIT_DB_SEMAPHORE: Lazy<Semaphore> = Lazy::new(|| Semaphore::new(8));
113 : use utils::crashsafe;
114 : use utils::generation::Generation;
115 : use utils::id::TimelineId;
116 : use utils::lsn::{Lsn, RecordLsn};
117 :
118 : pub mod blob_io;
119 : pub mod block_io;
120 : pub mod vectored_blob_io;
121 :
122 : pub mod disk_btree;
123 : pub(crate) mod ephemeral_file;
124 : pub mod layer_map;
125 :
126 : pub mod metadata;
127 : pub mod remote_timeline_client;
128 : pub mod storage_layer;
129 :
130 : pub mod checks;
131 : pub mod config;
132 : pub mod mgr;
133 : pub mod secondary;
134 : pub mod tasks;
135 : pub mod upload_queue;
136 :
137 : pub(crate) mod timeline;
138 :
139 : pub mod size;
140 :
141 : mod gc_block;
142 : mod gc_result;
143 : pub(crate) mod throttle;
144 :
145 : #[cfg(test)]
146 : pub mod debug;
147 :
148 : pub(crate) use timeline::{LogicalSizeCalculationCause, PageReconstructError, Timeline};
149 :
150 : pub(crate) use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
151 : // re-export for use in walreceiver
152 : pub use crate::tenant::timeline::WalReceiverInfo;
153 :
154 : /// The "tenants" part of `tenants/<tenant>/timelines...`
155 : pub const TENANTS_SEGMENT_NAME: &str = "tenants";
156 :
157 : /// Parts of the `.neon/tenants/<tenant_id>/timelines/<timeline_id>` directory prefix.
158 : pub const TIMELINES_SEGMENT_NAME: &str = "timelines";
159 :
160 : /// References to shared objects that are passed into each tenant, such
161 : /// as the shared remote storage client and process initialization state.
162 : #[derive(Clone)]
163 : pub struct TenantSharedResources {
164 : pub broker_client: storage_broker::BrokerClientChannel,
165 : pub remote_storage: GenericRemoteStorage,
166 : pub deletion_queue_client: DeletionQueueClient,
167 : pub l0_flush_global_state: L0FlushGlobalState,
168 : pub basebackup_cache: Arc<BasebackupCache>,
169 : pub feature_resolver: FeatureResolver,
170 : }
171 :
172 : /// A [`TenantShard`] is really an _attached_ tenant. The configuration
173 : /// for an attached tenant is a subset of the [`LocationConf`], represented
174 : /// in this struct.
175 : #[derive(Clone)]
176 : pub(super) struct AttachedTenantConf {
177 : tenant_conf: pageserver_api::models::TenantConfig,
178 : location: AttachedLocationConfig,
179 : /// The deadline before which we are blocked from GC so that
180 : /// leases have a chance to be renewed.
181 : lsn_lease_deadline: Option<tokio::time::Instant>,
182 : }
183 :
184 : impl AttachedTenantConf {
185 119 : fn new(
186 119 : conf: &'static PageServerConf,
187 119 : tenant_conf: pageserver_api::models::TenantConfig,
188 119 : location: AttachedLocationConfig,
189 119 : ) -> Self {
190 : // Sets a deadline before which we cannot proceed to GC due to lsn lease.
191 : //
192 : // We do this as the leases mapping are not persisted to disk. By delaying GC by lease
193 : // length, we guarantee that all the leases we granted before will have a chance to renew
194 : // when we run GC for the first time after restart / transition from AttachedMulti to AttachedSingle.
195 119 : let lsn_lease_deadline = if location.attach_mode == AttachmentMode::Single {
196 119 : Some(
197 119 : tokio::time::Instant::now()
198 119 : + TenantShard::get_lsn_lease_length_impl(conf, &tenant_conf),
199 119 : )
200 : } else {
201 : // We don't use `lsn_lease_deadline` to delay GC in AttachedMulti and AttachedStale
202 : // because we don't do GC in these modes.
203 0 : None
204 : };
205 :
206 119 : Self {
207 119 : tenant_conf,
208 119 : location,
209 119 : lsn_lease_deadline,
210 119 : }
211 119 : }
212 :
213 119 : fn try_from(
214 119 : conf: &'static PageServerConf,
215 119 : location_conf: LocationConf,
216 119 : ) -> anyhow::Result<Self> {
217 119 : match &location_conf.mode {
218 119 : LocationMode::Attached(attach_conf) => {
219 119 : Ok(Self::new(conf, location_conf.tenant_conf, *attach_conf))
220 : }
221 : LocationMode::Secondary(_) => {
222 0 : anyhow::bail!(
223 0 : "Attempted to construct AttachedTenantConf from a LocationConf in secondary mode"
224 : )
225 : }
226 : }
227 119 : }
228 :
229 380 : fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
230 380 : self.lsn_lease_deadline
231 380 : .map(|d| tokio::time::Instant::now() < d)
232 380 : .unwrap_or(false)
233 380 : }
234 : }
235 : struct TimelinePreload {
236 : timeline_id: TimelineId,
237 : client: RemoteTimelineClient,
238 : index_part: Result<MaybeDeletedIndexPart, DownloadError>,
239 : previous_heatmap: Option<PreviousHeatmap>,
240 : }
241 :
242 : pub(crate) struct TenantPreload {
243 : /// The tenant manifest from remote storage, or None if no manifest was found.
244 : tenant_manifest: Option<TenantManifest>,
245 : /// Map from timeline ID to a possible timeline preload. It is None iff the timeline is offloaded according to the manifest.
246 : timelines: HashMap<TimelineId, Option<TimelinePreload>>,
247 : }
248 :
249 : /// When we spawn a tenant, there is a special mode for tenant creation that
250 : /// avoids trying to read anything from remote storage.
251 : pub(crate) enum SpawnMode {
252 : /// Activate as soon as possible
253 : Eager,
254 : /// Lazy activation in the background, with the option to skip the queue if the need comes up
255 : Lazy,
256 : }
257 :
258 : ///
259 : /// Tenant consists of multiple timelines. Keep them in a hash table.
260 : ///
261 : pub struct TenantShard {
262 : // Global pageserver config parameters
263 : pub conf: &'static PageServerConf,
264 :
265 : /// The value creation timestamp, used to measure activation delay, see:
266 : /// <https://github.com/neondatabase/neon/issues/4025>
267 : constructed_at: Instant,
268 :
269 : state: watch::Sender<TenantState>,
270 :
271 : // Overridden tenant-specific config parameters.
272 : // We keep pageserver_api::models::TenantConfig sturct here to preserve the information
273 : // about parameters that are not set.
274 : // This is necessary to allow global config updates.
275 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
276 :
277 : tenant_shard_id: TenantShardId,
278 :
279 : // The detailed sharding information, beyond the number/count in tenant_shard_id
280 : shard_identity: ShardIdentity,
281 :
282 : /// The remote storage generation, used to protect S3 objects from split-brain.
283 : /// Does not change over the lifetime of the [`TenantShard`] object.
284 : ///
285 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
286 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
287 : generation: Generation,
288 :
289 : timelines: Mutex<HashMap<TimelineId, Arc<Timeline>>>,
290 :
291 : /// During timeline creation, we first insert the TimelineId to the
292 : /// creating map, then `timelines`, then remove it from the creating map.
293 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
294 : timelines_creating: std::sync::Mutex<HashSet<TimelineId>>,
295 :
296 : /// Possibly offloaded and archived timelines
297 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
298 : timelines_offloaded: Mutex<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
299 :
300 : /// Tracks the timelines that are currently importing into this tenant shard.
301 : ///
302 : /// Note that importing timelines are also present in [`Self::timelines_creating`].
303 : /// Keep this in mind when ordering lock acquisition.
304 : ///
305 : /// Lifetime:
306 : /// * An imported timeline is created while scanning the bucket on tenant attach
307 : /// if the index part contains an `import_pgdata` entry and said field marks the import
308 : /// as in progress.
309 : /// * Imported timelines are removed when the storage controller calls the post timeline
310 : /// import activation endpoint.
311 : timelines_importing: std::sync::Mutex<HashMap<TimelineId, Arc<ImportingTimeline>>>,
312 :
313 : /// The last tenant manifest known to be in remote storage. None if the manifest has not yet
314 : /// been either downloaded or uploaded. Always Some after tenant attach.
315 : ///
316 : /// Initially populated during tenant attach, updated via `maybe_upload_tenant_manifest`.
317 : ///
318 : /// Do not modify this directly. It is used to check whether a new manifest needs to be
319 : /// uploaded. The manifest is constructed in `build_tenant_manifest`, and uploaded via
320 : /// `maybe_upload_tenant_manifest`.
321 : remote_tenant_manifest: tokio::sync::Mutex<Option<TenantManifest>>,
322 :
323 : // This mutex prevents creation of new timelines during GC.
324 : // Adding yet another mutex (in addition to `timelines`) is needed because holding
325 : // `timelines` mutex during all GC iteration
326 : // may block for a long time `get_timeline`, `get_timelines_state`,... and other operations
327 : // with timelines, which in turn may cause dropping replication connection, expiration of wait_for_lsn
328 : // timeout...
329 : gc_cs: tokio::sync::Mutex<()>,
330 : walredo_mgr: Option<Arc<WalRedoManager>>,
331 :
332 : /// Provides access to timeline data sitting in the remote storage.
333 : pub(crate) remote_storage: GenericRemoteStorage,
334 :
335 : /// Access to global deletion queue for when this tenant wants to schedule a deletion.
336 : deletion_queue_client: DeletionQueueClient,
337 :
338 : /// A channel to send async requests to prepare a basebackup for the basebackup cache.
339 : basebackup_cache: Arc<BasebackupCache>,
340 :
341 : /// Cached logical sizes updated updated on each [`TenantShard::gather_size_inputs`].
342 : cached_logical_sizes: tokio::sync::Mutex<HashMap<(TimelineId, Lsn), u64>>,
343 : cached_synthetic_tenant_size: Arc<AtomicU64>,
344 :
345 : eviction_task_tenant_state: tokio::sync::Mutex<EvictionTaskTenantState>,
346 :
347 : /// Track repeated failures to compact, so that we can back off.
348 : /// Overhead of mutex is acceptable because compaction is done with a multi-second period.
349 : compaction_circuit_breaker: std::sync::Mutex<CircuitBreaker>,
350 :
351 : /// Signals the tenant compaction loop that there is L0 compaction work to be done.
352 : pub(crate) l0_compaction_trigger: Arc<Notify>,
353 :
354 : /// Scheduled gc-compaction tasks.
355 : scheduled_compaction_tasks: std::sync::Mutex<HashMap<TimelineId, Arc<GcCompactionQueue>>>,
356 :
357 : /// If the tenant is in Activating state, notify this to encourage it
358 : /// to proceed to Active as soon as possible, rather than waiting for lazy
359 : /// background warmup.
360 : pub(crate) activate_now_sem: tokio::sync::Semaphore,
361 :
362 : /// Time it took for the tenant to activate. Zero if not active yet.
363 : attach_wal_lag_cooldown: Arc<std::sync::OnceLock<WalLagCooldown>>,
364 :
365 : // Cancellation token fires when we have entered shutdown(). This is a parent of
366 : // Timelines' cancellation token.
367 : pub(crate) cancel: CancellationToken,
368 :
369 : // Users of the TenantShard such as the page service must take this Gate to avoid
370 : // trying to use a TenantShard which is shutting down.
371 : pub(crate) gate: Gate,
372 :
373 : /// Throttle applied at the top of [`Timeline::get`].
374 : /// All [`TenantShard::timelines`] of a given [`TenantShard`] instance share the same [`throttle::Throttle`] instance.
375 : pub(crate) pagestream_throttle: Arc<throttle::Throttle>,
376 :
377 : pub(crate) pagestream_throttle_metrics: Arc<crate::metrics::tenant_throttling::Pagestream>,
378 :
379 : /// An ongoing timeline detach concurrency limiter.
380 : ///
381 : /// As a tenant will likely be restarted as part of timeline detach ancestor it makes no sense
382 : /// to have two running at the same time. A different one can be started if an earlier one
383 : /// has failed for whatever reason.
384 : ongoing_timeline_detach: std::sync::Mutex<Option<(TimelineId, utils::completion::Barrier)>>,
385 :
386 : /// `index_part.json` based gc blocking reason tracking.
387 : ///
388 : /// New gc iterations must start a new iteration by acquiring `GcBlock::start` before
389 : /// proceeding.
390 : pub(crate) gc_block: gc_block::GcBlock,
391 :
392 : l0_flush_global_state: L0FlushGlobalState,
393 :
394 : pub(crate) feature_resolver: Arc<TenantFeatureResolver>,
395 : }
396 : impl std::fmt::Debug for TenantShard {
397 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
398 0 : write!(f, "{} ({})", self.tenant_shard_id, self.current_state())
399 0 : }
400 : }
401 :
402 : pub(crate) enum WalRedoManager {
403 : Prod(WalredoManagerId, PostgresRedoManager),
404 : #[cfg(test)]
405 : Test(harness::TestRedoManager),
406 : }
407 :
408 : #[derive(thiserror::Error, Debug)]
409 : #[error("pageserver is shutting down")]
410 : pub(crate) struct GlobalShutDown;
411 :
412 : impl WalRedoManager {
413 0 : pub(crate) fn new(mgr: PostgresRedoManager) -> Result<Arc<Self>, GlobalShutDown> {
414 0 : let id = WalredoManagerId::next();
415 0 : let arc = Arc::new(Self::Prod(id, mgr));
416 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
417 0 : match &mut *guard {
418 0 : Some(map) => {
419 0 : map.insert(id, Arc::downgrade(&arc));
420 0 : Ok(arc)
421 : }
422 0 : None => Err(GlobalShutDown),
423 : }
424 0 : }
425 : }
426 :
427 : impl Drop for WalRedoManager {
428 5 : fn drop(&mut self) {
429 5 : match self {
430 0 : Self::Prod(id, _) => {
431 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
432 0 : if let Some(map) = &mut *guard {
433 0 : map.remove(id).expect("new() registers, drop() unregisters");
434 0 : }
435 : }
436 : #[cfg(test)]
437 5 : Self::Test(_) => {
438 5 : // Not applicable to test redo manager
439 5 : }
440 : }
441 5 : }
442 : }
443 :
444 : /// Global registry of all walredo managers so that [`crate::shutdown_pageserver`] can shut down
445 : /// the walredo processes outside of the regular order.
446 : ///
447 : /// This is necessary to work around a systemd bug where it freezes if there are
448 : /// walredo processes left => <https://github.com/neondatabase/cloud/issues/11387>
449 : #[allow(clippy::type_complexity)]
450 : pub(crate) static WALREDO_MANAGERS: once_cell::sync::Lazy<
451 : Mutex<Option<HashMap<WalredoManagerId, Weak<WalRedoManager>>>>,
452 0 : > = once_cell::sync::Lazy::new(|| Mutex::new(Some(HashMap::new())));
453 : #[derive(PartialEq, Eq, Hash, Clone, Copy, Debug)]
454 : pub(crate) struct WalredoManagerId(u64);
455 : impl WalredoManagerId {
456 0 : pub fn next() -> Self {
457 : static NEXT: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(1);
458 0 : let id = NEXT.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
459 0 : if id == 0 {
460 0 : panic!(
461 0 : "WalredoManagerId::new() returned 0, indicating wraparound, risking it's no longer unique"
462 : );
463 0 : }
464 0 : Self(id)
465 0 : }
466 : }
467 :
468 : #[cfg(test)]
469 : impl From<harness::TestRedoManager> for WalRedoManager {
470 119 : fn from(mgr: harness::TestRedoManager) -> Self {
471 119 : Self::Test(mgr)
472 119 : }
473 : }
474 :
475 : impl WalRedoManager {
476 3 : pub(crate) async fn shutdown(&self) -> bool {
477 3 : match self {
478 0 : Self::Prod(_, mgr) => mgr.shutdown().await,
479 : #[cfg(test)]
480 : Self::Test(_) => {
481 : // Not applicable to test redo manager
482 3 : true
483 : }
484 : }
485 3 : }
486 :
487 0 : pub(crate) fn maybe_quiesce(&self, idle_timeout: Duration) {
488 0 : match self {
489 0 : Self::Prod(_, mgr) => mgr.maybe_quiesce(idle_timeout),
490 : #[cfg(test)]
491 0 : Self::Test(_) => {
492 0 : // Not applicable to test redo manager
493 0 : }
494 : }
495 0 : }
496 :
497 : /// # Cancel-Safety
498 : ///
499 : /// This method is cancellation-safe.
500 26649 : pub async fn request_redo(
501 26649 : &self,
502 26649 : key: pageserver_api::key::Key,
503 26649 : lsn: Lsn,
504 26649 : base_img: Option<(Lsn, bytes::Bytes)>,
505 26649 : records: Vec<(Lsn, wal_decoder::models::record::NeonWalRecord)>,
506 26649 : pg_version: PgMajorVersion,
507 26649 : redo_attempt_type: RedoAttemptType,
508 26649 : ) -> Result<bytes::Bytes, walredo::Error> {
509 26649 : match self {
510 0 : Self::Prod(_, mgr) => {
511 0 : mgr.request_redo(key, lsn, base_img, records, pg_version, redo_attempt_type)
512 0 : .await
513 : }
514 : #[cfg(test)]
515 26649 : Self::Test(mgr) => {
516 26649 : mgr.request_redo(key, lsn, base_img, records, pg_version, redo_attempt_type)
517 26649 : .await
518 : }
519 : }
520 26649 : }
521 :
522 0 : pub(crate) fn status(&self) -> Option<WalRedoManagerStatus> {
523 0 : match self {
524 0 : WalRedoManager::Prod(_, m) => Some(m.status()),
525 : #[cfg(test)]
526 0 : WalRedoManager::Test(_) => None,
527 : }
528 0 : }
529 : }
530 :
531 : /// A very lightweight memory representation of an offloaded timeline.
532 : ///
533 : /// We need to store the list of offloaded timelines so that we can perform operations on them,
534 : /// like unoffloading them, or (at a later date), decide to perform flattening.
535 : /// This type has a much smaller memory impact than [`Timeline`], and thus we can store many
536 : /// more offloaded timelines than we can manage ones that aren't.
537 : pub struct OffloadedTimeline {
538 : pub tenant_shard_id: TenantShardId,
539 : pub timeline_id: TimelineId,
540 : pub ancestor_timeline_id: Option<TimelineId>,
541 : /// Whether to retain the branch lsn at the ancestor or not
542 : pub ancestor_retain_lsn: Option<Lsn>,
543 :
544 : /// When the timeline was archived.
545 : ///
546 : /// Present for future flattening deliberations.
547 : pub archived_at: NaiveDateTime,
548 :
549 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
550 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
551 : pub delete_progress: TimelineDeleteProgress,
552 :
553 : /// Part of the `OffloadedTimeline` object's lifecycle: this needs to be set before we drop it
554 : pub deleted_from_ancestor: AtomicBool,
555 :
556 : _metrics_guard: OffloadedTimelineMetricsGuard,
557 : }
558 :
559 : /// Increases the offloaded timeline count metric when created, and decreases when dropped.
560 : struct OffloadedTimelineMetricsGuard;
561 :
562 : impl OffloadedTimelineMetricsGuard {
563 1 : fn new() -> Self {
564 1 : TIMELINE_STATE_METRIC
565 1 : .with_label_values(&["offloaded"])
566 1 : .inc();
567 1 : Self
568 1 : }
569 : }
570 :
571 : impl Drop for OffloadedTimelineMetricsGuard {
572 1 : fn drop(&mut self) {
573 1 : TIMELINE_STATE_METRIC
574 1 : .with_label_values(&["offloaded"])
575 1 : .dec();
576 1 : }
577 : }
578 :
579 : impl OffloadedTimeline {
580 : /// Obtains an offloaded timeline from a given timeline object.
581 : ///
582 : /// Returns `None` if the `archived_at` flag couldn't be obtained, i.e.
583 : /// the timeline is not in a stopped state.
584 : /// Panics if the timeline is not archived.
585 1 : fn from_timeline(timeline: &Timeline) -> Result<Self, UploadQueueNotReadyError> {
586 1 : let (ancestor_retain_lsn, ancestor_timeline_id) =
587 1 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
588 1 : let ancestor_lsn = timeline.get_ancestor_lsn();
589 1 : let ancestor_timeline_id = ancestor_timeline.timeline_id;
590 1 : let mut gc_info = ancestor_timeline.gc_info.write().unwrap();
591 1 : gc_info.insert_child(timeline.timeline_id, ancestor_lsn, MaybeOffloaded::Yes);
592 1 : (Some(ancestor_lsn), Some(ancestor_timeline_id))
593 : } else {
594 0 : (None, None)
595 : };
596 1 : let archived_at = timeline
597 1 : .remote_client
598 1 : .archived_at_stopped_queue()?
599 1 : .expect("must be called on an archived timeline");
600 1 : Ok(Self {
601 1 : tenant_shard_id: timeline.tenant_shard_id,
602 1 : timeline_id: timeline.timeline_id,
603 1 : ancestor_timeline_id,
604 1 : ancestor_retain_lsn,
605 1 : archived_at,
606 1 :
607 1 : delete_progress: timeline.delete_progress.clone(),
608 1 : deleted_from_ancestor: AtomicBool::new(false),
609 1 :
610 1 : _metrics_guard: OffloadedTimelineMetricsGuard::new(),
611 1 : })
612 1 : }
613 0 : fn from_manifest(tenant_shard_id: TenantShardId, manifest: &OffloadedTimelineManifest) -> Self {
614 : // We expect to reach this case in tenant loading, where the `retain_lsn` is populated in the parent's `gc_info`
615 : // by the `initialize_gc_info` function.
616 : let OffloadedTimelineManifest {
617 0 : timeline_id,
618 0 : ancestor_timeline_id,
619 0 : ancestor_retain_lsn,
620 0 : archived_at,
621 0 : } = *manifest;
622 0 : Self {
623 0 : tenant_shard_id,
624 0 : timeline_id,
625 0 : ancestor_timeline_id,
626 0 : ancestor_retain_lsn,
627 0 : archived_at,
628 0 : delete_progress: TimelineDeleteProgress::default(),
629 0 : deleted_from_ancestor: AtomicBool::new(false),
630 0 : _metrics_guard: OffloadedTimelineMetricsGuard::new(),
631 0 : }
632 0 : }
633 1 : fn manifest(&self) -> OffloadedTimelineManifest {
634 : let Self {
635 1 : timeline_id,
636 1 : ancestor_timeline_id,
637 1 : ancestor_retain_lsn,
638 1 : archived_at,
639 : ..
640 1 : } = self;
641 1 : OffloadedTimelineManifest {
642 1 : timeline_id: *timeline_id,
643 1 : ancestor_timeline_id: *ancestor_timeline_id,
644 1 : ancestor_retain_lsn: *ancestor_retain_lsn,
645 1 : archived_at: *archived_at,
646 1 : }
647 1 : }
648 : /// Delete this timeline's retain_lsn from its ancestor, if present in the given tenant
649 0 : fn delete_from_ancestor_with_timelines(
650 0 : &self,
651 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
652 0 : ) {
653 0 : if let (Some(_retain_lsn), Some(ancestor_timeline_id)) =
654 0 : (self.ancestor_retain_lsn, self.ancestor_timeline_id)
655 : {
656 0 : if let Some((_, ancestor_timeline)) = timelines
657 0 : .iter()
658 0 : .find(|(tid, _tl)| **tid == ancestor_timeline_id)
659 : {
660 0 : let removal_happened = ancestor_timeline
661 0 : .gc_info
662 0 : .write()
663 0 : .unwrap()
664 0 : .remove_child_offloaded(self.timeline_id);
665 0 : if !removal_happened {
666 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
667 0 : "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
668 0 : }
669 0 : }
670 0 : }
671 0 : self.deleted_from_ancestor.store(true, Ordering::Release);
672 0 : }
673 : /// Call [`Self::delete_from_ancestor_with_timelines`] instead if possible.
674 : ///
675 : /// As the entire tenant is being dropped, don't bother deregistering the `retain_lsn` from the ancestor.
676 1 : fn defuse_for_tenant_drop(&self) {
677 1 : self.deleted_from_ancestor.store(true, Ordering::Release);
678 1 : }
679 : }
680 :
681 : impl fmt::Debug for OffloadedTimeline {
682 0 : fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
683 0 : write!(f, "OffloadedTimeline<{}>", self.timeline_id)
684 0 : }
685 : }
686 :
687 : impl Drop for OffloadedTimeline {
688 1 : fn drop(&mut self) {
689 1 : if !self.deleted_from_ancestor.load(Ordering::Acquire) {
690 0 : tracing::warn!(
691 0 : "offloaded timeline {} was dropped without having cleaned it up at the ancestor",
692 : self.timeline_id
693 : );
694 1 : }
695 1 : }
696 : }
697 :
698 : #[derive(Copy, Clone, PartialEq, Eq, Hash, Debug)]
699 : pub enum MaybeOffloaded {
700 : Yes,
701 : No,
702 : }
703 :
704 : #[derive(Clone, Debug)]
705 : pub enum TimelineOrOffloaded {
706 : Timeline(Arc<Timeline>),
707 : Offloaded(Arc<OffloadedTimeline>),
708 : Importing(Arc<ImportingTimeline>),
709 : }
710 :
711 : impl TimelineOrOffloaded {
712 0 : pub fn arc_ref(&self) -> TimelineOrOffloadedArcRef<'_> {
713 0 : match self {
714 0 : TimelineOrOffloaded::Timeline(timeline) => {
715 0 : TimelineOrOffloadedArcRef::Timeline(timeline)
716 : }
717 0 : TimelineOrOffloaded::Offloaded(offloaded) => {
718 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded)
719 : }
720 0 : TimelineOrOffloaded::Importing(importing) => {
721 0 : TimelineOrOffloadedArcRef::Importing(importing)
722 : }
723 : }
724 0 : }
725 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
726 0 : self.arc_ref().tenant_shard_id()
727 0 : }
728 0 : pub fn timeline_id(&self) -> TimelineId {
729 0 : self.arc_ref().timeline_id()
730 0 : }
731 1 : pub fn delete_progress(&self) -> &Arc<tokio::sync::Mutex<DeleteTimelineFlow>> {
732 1 : match self {
733 1 : TimelineOrOffloaded::Timeline(timeline) => &timeline.delete_progress,
734 0 : TimelineOrOffloaded::Offloaded(offloaded) => &offloaded.delete_progress,
735 0 : TimelineOrOffloaded::Importing(importing) => &importing.delete_progress,
736 : }
737 1 : }
738 0 : fn maybe_remote_client(&self) -> Option<Arc<RemoteTimelineClient>> {
739 0 : match self {
740 0 : TimelineOrOffloaded::Timeline(timeline) => Some(timeline.remote_client.clone()),
741 0 : TimelineOrOffloaded::Offloaded(_offloaded) => None,
742 0 : TimelineOrOffloaded::Importing(importing) => {
743 0 : Some(importing.timeline.remote_client.clone())
744 : }
745 : }
746 0 : }
747 : }
748 :
749 : pub enum TimelineOrOffloadedArcRef<'a> {
750 : Timeline(&'a Arc<Timeline>),
751 : Offloaded(&'a Arc<OffloadedTimeline>),
752 : Importing(&'a Arc<ImportingTimeline>),
753 : }
754 :
755 : impl TimelineOrOffloadedArcRef<'_> {
756 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
757 0 : match self {
758 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.tenant_shard_id,
759 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.tenant_shard_id,
760 0 : TimelineOrOffloadedArcRef::Importing(importing) => importing.timeline.tenant_shard_id,
761 : }
762 0 : }
763 0 : pub fn timeline_id(&self) -> TimelineId {
764 0 : match self {
765 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.timeline_id,
766 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.timeline_id,
767 0 : TimelineOrOffloadedArcRef::Importing(importing) => importing.timeline.timeline_id,
768 : }
769 0 : }
770 : }
771 :
772 : impl<'a> From<&'a Arc<Timeline>> for TimelineOrOffloadedArcRef<'a> {
773 0 : fn from(timeline: &'a Arc<Timeline>) -> Self {
774 0 : Self::Timeline(timeline)
775 0 : }
776 : }
777 :
778 : impl<'a> From<&'a Arc<OffloadedTimeline>> for TimelineOrOffloadedArcRef<'a> {
779 0 : fn from(timeline: &'a Arc<OffloadedTimeline>) -> Self {
780 0 : Self::Offloaded(timeline)
781 0 : }
782 : }
783 :
784 : impl<'a> From<&'a Arc<ImportingTimeline>> for TimelineOrOffloadedArcRef<'a> {
785 0 : fn from(timeline: &'a Arc<ImportingTimeline>) -> Self {
786 0 : Self::Importing(timeline)
787 0 : }
788 : }
789 :
790 : #[derive(Debug, thiserror::Error, PartialEq, Eq)]
791 : pub enum GetTimelineError {
792 : #[error("Timeline is shutting down")]
793 : ShuttingDown,
794 : #[error("Timeline {tenant_id}/{timeline_id} is not active, state: {state:?}")]
795 : NotActive {
796 : tenant_id: TenantShardId,
797 : timeline_id: TimelineId,
798 : state: TimelineState,
799 : },
800 : #[error("Timeline {tenant_id}/{timeline_id} was not found")]
801 : NotFound {
802 : tenant_id: TenantShardId,
803 : timeline_id: TimelineId,
804 : },
805 : }
806 :
807 : #[derive(Debug, thiserror::Error)]
808 : pub enum LoadLocalTimelineError {
809 : #[error("FailedToLoad")]
810 : Load(#[source] anyhow::Error),
811 : #[error("FailedToResumeDeletion")]
812 : ResumeDeletion(#[source] anyhow::Error),
813 : }
814 :
815 : #[derive(thiserror::Error)]
816 : pub enum DeleteTimelineError {
817 : #[error("NotFound")]
818 : NotFound,
819 :
820 : #[error("HasChildren")]
821 : HasChildren(Vec<TimelineId>),
822 :
823 : #[error("Timeline deletion is already in progress")]
824 : AlreadyInProgress(Arc<tokio::sync::Mutex<DeleteTimelineFlow>>),
825 :
826 : #[error("Cancelled")]
827 : Cancelled,
828 :
829 : #[error(transparent)]
830 : Other(#[from] anyhow::Error),
831 : }
832 :
833 : impl Debug for DeleteTimelineError {
834 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
835 0 : match self {
836 0 : Self::NotFound => write!(f, "NotFound"),
837 0 : Self::HasChildren(c) => f.debug_tuple("HasChildren").field(c).finish(),
838 0 : Self::AlreadyInProgress(_) => f.debug_tuple("AlreadyInProgress").finish(),
839 0 : Self::Cancelled => f.debug_tuple("Cancelled").finish(),
840 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
841 : }
842 0 : }
843 : }
844 :
845 : #[derive(thiserror::Error)]
846 : pub enum TimelineArchivalError {
847 : #[error("NotFound")]
848 : NotFound,
849 :
850 : #[error("Timeout")]
851 : Timeout,
852 :
853 : #[error("Cancelled")]
854 : Cancelled,
855 :
856 : #[error("ancestor is archived: {}", .0)]
857 : HasArchivedParent(TimelineId),
858 :
859 : #[error("HasUnarchivedChildren")]
860 : HasUnarchivedChildren(Vec<TimelineId>),
861 :
862 : #[error("Timeline archival is already in progress")]
863 : AlreadyInProgress,
864 :
865 : #[error(transparent)]
866 : Other(anyhow::Error),
867 : }
868 :
869 : #[derive(thiserror::Error, Debug)]
870 : pub(crate) enum TenantManifestError {
871 : #[error("Remote storage error: {0}")]
872 : RemoteStorage(anyhow::Error),
873 :
874 : #[error("Cancelled")]
875 : Cancelled,
876 : }
877 :
878 : impl From<TenantManifestError> for TimelineArchivalError {
879 0 : fn from(e: TenantManifestError) -> Self {
880 0 : match e {
881 0 : TenantManifestError::RemoteStorage(e) => Self::Other(e),
882 0 : TenantManifestError::Cancelled => Self::Cancelled,
883 : }
884 0 : }
885 : }
886 :
887 : impl Debug for TimelineArchivalError {
888 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
889 0 : match self {
890 0 : Self::NotFound => write!(f, "NotFound"),
891 0 : Self::Timeout => write!(f, "Timeout"),
892 0 : Self::Cancelled => write!(f, "Cancelled"),
893 0 : Self::HasArchivedParent(p) => f.debug_tuple("HasArchivedParent").field(p).finish(),
894 0 : Self::HasUnarchivedChildren(c) => {
895 0 : f.debug_tuple("HasUnarchivedChildren").field(c).finish()
896 : }
897 0 : Self::AlreadyInProgress => f.debug_tuple("AlreadyInProgress").finish(),
898 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
899 : }
900 0 : }
901 : }
902 :
903 : pub enum SetStoppingError {
904 : AlreadyStopping(completion::Barrier),
905 : Broken,
906 : }
907 :
908 : impl Debug for SetStoppingError {
909 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
910 0 : match self {
911 0 : Self::AlreadyStopping(_) => f.debug_tuple("AlreadyStopping").finish(),
912 0 : Self::Broken => write!(f, "Broken"),
913 : }
914 0 : }
915 : }
916 :
917 : #[derive(thiserror::Error, Debug)]
918 : pub(crate) enum FinalizeTimelineImportError {
919 : #[error("Import task not done yet")]
920 : ImportTaskStillRunning,
921 : #[error("Shutting down")]
922 : ShuttingDown,
923 : }
924 :
925 : /// Arguments to [`TenantShard::create_timeline`].
926 : ///
927 : /// Not usable as an idempotency key for timeline creation because if [`CreateTimelineParamsBranch::ancestor_start_lsn`]
928 : /// is `None`, the result of the timeline create call is not deterministic.
929 : ///
930 : /// See [`CreateTimelineIdempotency`] for an idempotency key.
931 : #[derive(Debug)]
932 : pub(crate) enum CreateTimelineParams {
933 : Bootstrap(CreateTimelineParamsBootstrap),
934 : Branch(CreateTimelineParamsBranch),
935 : ImportPgdata(CreateTimelineParamsImportPgdata),
936 : }
937 :
938 : #[derive(Debug)]
939 : pub(crate) struct CreateTimelineParamsBootstrap {
940 : pub(crate) new_timeline_id: TimelineId,
941 : pub(crate) existing_initdb_timeline_id: Option<TimelineId>,
942 : pub(crate) pg_version: PgMajorVersion,
943 : }
944 :
945 : /// NB: See comment on [`CreateTimelineIdempotency::Branch`] for why there's no `pg_version` here.
946 : #[derive(Debug)]
947 : pub(crate) struct CreateTimelineParamsBranch {
948 : pub(crate) new_timeline_id: TimelineId,
949 : pub(crate) ancestor_timeline_id: TimelineId,
950 : pub(crate) ancestor_start_lsn: Option<Lsn>,
951 : }
952 :
953 : #[derive(Debug)]
954 : pub(crate) struct CreateTimelineParamsImportPgdata {
955 : pub(crate) new_timeline_id: TimelineId,
956 : pub(crate) location: import_pgdata::index_part_format::Location,
957 : pub(crate) idempotency_key: import_pgdata::index_part_format::IdempotencyKey,
958 : }
959 :
960 : /// What is used to determine idempotency of a [`TenantShard::create_timeline`] call in [`TenantShard::start_creating_timeline`] in [`TenantShard::start_creating_timeline`].
961 : ///
962 : /// Each [`Timeline`] object holds [`Self`] as an immutable property in [`Timeline::create_idempotency`].
963 : ///
964 : /// We lower timeline creation requests to [`Self`], and then use [`PartialEq::eq`] to compare [`Timeline::create_idempotency`] with the request.
965 : /// If they are equal, we return a reference to the existing timeline, otherwise it's an idempotency conflict.
966 : ///
967 : /// There is special treatment for [`Self::FailWithConflict`] to always return an idempotency conflict.
968 : /// It would be nice to have more advanced derive macros to make that special treatment declarative.
969 : ///
970 : /// Notes:
971 : /// - Unlike [`CreateTimelineParams`], ancestor LSN is fixed, so, branching will be at a deterministic LSN.
972 : /// - We make some trade-offs though, e.g., [`CreateTimelineParamsBootstrap::existing_initdb_timeline_id`]
973 : /// is not considered for idempotency. We can improve on this over time if we deem it necessary.
974 : ///
975 : #[derive(Debug, Clone, PartialEq, Eq)]
976 : pub(crate) enum CreateTimelineIdempotency {
977 : /// NB: special treatment, see comment in [`Self`].
978 : FailWithConflict,
979 : Bootstrap {
980 : pg_version: PgMajorVersion,
981 : },
982 : /// NB: branches always have the same `pg_version` as their ancestor.
983 : /// While [`pageserver_api::models::TimelineCreateRequestMode::Branch::pg_version`]
984 : /// exists as a field, and is set by cplane, it has always been ignored by pageserver when
985 : /// determining the child branch pg_version.
986 : Branch {
987 : ancestor_timeline_id: TimelineId,
988 : ancestor_start_lsn: Lsn,
989 : },
990 : ImportPgdata(CreatingTimelineIdempotencyImportPgdata),
991 : }
992 :
993 : #[derive(Debug, Clone, PartialEq, Eq)]
994 : pub(crate) struct CreatingTimelineIdempotencyImportPgdata {
995 : idempotency_key: import_pgdata::index_part_format::IdempotencyKey,
996 : }
997 :
998 : /// What is returned by [`TenantShard::start_creating_timeline`].
999 : #[must_use]
1000 : enum StartCreatingTimelineResult {
1001 : CreateGuard(TimelineCreateGuard),
1002 : Idempotent(Arc<Timeline>),
1003 : }
1004 :
1005 : #[allow(clippy::large_enum_variant, reason = "TODO")]
1006 : enum TimelineInitAndSyncResult {
1007 : ReadyToActivate,
1008 : NeedsSpawnImportPgdata(TimelineInitAndSyncNeedsSpawnImportPgdata),
1009 : }
1010 :
1011 : #[must_use]
1012 : struct TimelineInitAndSyncNeedsSpawnImportPgdata {
1013 : timeline: Arc<Timeline>,
1014 : import_pgdata: import_pgdata::index_part_format::Root,
1015 : guard: TimelineCreateGuard,
1016 : }
1017 :
1018 : /// What is returned by [`TenantShard::create_timeline`].
1019 : enum CreateTimelineResult {
1020 : Created(Arc<Timeline>),
1021 : Idempotent(Arc<Timeline>),
1022 : /// IMPORTANT: This [`Arc<Timeline>`] object is not in [`TenantShard::timelines`] when
1023 : /// we return this result, nor will this concrete object ever be added there.
1024 : /// Cf method comment on [`TenantShard::create_timeline_import_pgdata`].
1025 : ImportSpawned(Arc<Timeline>),
1026 : }
1027 :
1028 : impl CreateTimelineResult {
1029 0 : fn discriminant(&self) -> &'static str {
1030 0 : match self {
1031 0 : Self::Created(_) => "Created",
1032 0 : Self::Idempotent(_) => "Idempotent",
1033 0 : Self::ImportSpawned(_) => "ImportSpawned",
1034 : }
1035 0 : }
1036 0 : fn timeline(&self) -> &Arc<Timeline> {
1037 0 : match self {
1038 0 : Self::Created(t) | Self::Idempotent(t) | Self::ImportSpawned(t) => t,
1039 : }
1040 0 : }
1041 : /// Unit test timelines aren't activated, test has to do it if it needs to.
1042 : #[cfg(test)]
1043 118 : fn into_timeline_for_test(self) -> Arc<Timeline> {
1044 118 : match self {
1045 118 : Self::Created(t) | Self::Idempotent(t) | Self::ImportSpawned(t) => t,
1046 : }
1047 118 : }
1048 : }
1049 :
1050 : #[derive(thiserror::Error, Debug)]
1051 : pub enum CreateTimelineError {
1052 : #[error("creation of timeline with the given ID is in progress")]
1053 : AlreadyCreating,
1054 : #[error("timeline already exists with different parameters")]
1055 : Conflict,
1056 : #[error(transparent)]
1057 : AncestorLsn(anyhow::Error),
1058 : #[error("ancestor timeline is not active")]
1059 : AncestorNotActive,
1060 : #[error("ancestor timeline is archived")]
1061 : AncestorArchived,
1062 : #[error("tenant shutting down")]
1063 : ShuttingDown,
1064 : #[error(transparent)]
1065 : Other(#[from] anyhow::Error),
1066 : }
1067 :
1068 : #[derive(thiserror::Error, Debug)]
1069 : pub enum InitdbError {
1070 : #[error("Operation was cancelled")]
1071 : Cancelled,
1072 : #[error(transparent)]
1073 : Other(anyhow::Error),
1074 : #[error(transparent)]
1075 : Inner(postgres_initdb::Error),
1076 : }
1077 :
1078 : enum CreateTimelineCause {
1079 : Load,
1080 : Delete,
1081 : }
1082 :
1083 : #[allow(clippy::large_enum_variant, reason = "TODO")]
1084 : enum LoadTimelineCause {
1085 : Attach,
1086 : Unoffload,
1087 : }
1088 :
1089 : #[derive(thiserror::Error, Debug)]
1090 : pub(crate) enum GcError {
1091 : // The tenant is shutting down
1092 : #[error("tenant shutting down")]
1093 : TenantCancelled,
1094 :
1095 : // The tenant is shutting down
1096 : #[error("timeline shutting down")]
1097 : TimelineCancelled,
1098 :
1099 : // The tenant is in a state inelegible to run GC
1100 : #[error("not active")]
1101 : NotActive,
1102 :
1103 : // A requested GC cutoff LSN was invalid, for example it tried to move backwards
1104 : #[error("not active")]
1105 : BadLsn { why: String },
1106 :
1107 : // A remote storage error while scheduling updates after compaction
1108 : #[error(transparent)]
1109 : Remote(anyhow::Error),
1110 :
1111 : // An error reading while calculating GC cutoffs
1112 : #[error(transparent)]
1113 : GcCutoffs(PageReconstructError),
1114 :
1115 : // If GC was invoked for a particular timeline, this error means it didn't exist
1116 : #[error("timeline not found")]
1117 : TimelineNotFound,
1118 : }
1119 :
1120 : impl From<PageReconstructError> for GcError {
1121 0 : fn from(value: PageReconstructError) -> Self {
1122 0 : match value {
1123 0 : PageReconstructError::Cancelled => Self::TimelineCancelled,
1124 0 : other => Self::GcCutoffs(other),
1125 : }
1126 0 : }
1127 : }
1128 :
1129 : impl From<NotInitialized> for GcError {
1130 0 : fn from(value: NotInitialized) -> Self {
1131 0 : match value {
1132 0 : NotInitialized::Uninitialized => GcError::Remote(value.into()),
1133 0 : NotInitialized::Stopped | NotInitialized::ShuttingDown => GcError::TimelineCancelled,
1134 : }
1135 0 : }
1136 : }
1137 :
1138 : impl From<timeline::layer_manager::Shutdown> for GcError {
1139 0 : fn from(_: timeline::layer_manager::Shutdown) -> Self {
1140 0 : GcError::TimelineCancelled
1141 0 : }
1142 : }
1143 :
1144 : #[derive(thiserror::Error, Debug)]
1145 : pub(crate) enum LoadConfigError {
1146 : #[error("TOML deserialization error: '{0}'")]
1147 : DeserializeToml(#[from] toml_edit::de::Error),
1148 :
1149 : #[error("Config not found at {0}")]
1150 : NotFound(Utf8PathBuf),
1151 : }
1152 :
1153 : impl TenantShard {
1154 : /// Yet another helper for timeline initialization.
1155 : ///
1156 : /// - Initializes the Timeline struct and inserts it into the tenant's hash map
1157 : /// - Scans the local timeline directory for layer files and builds the layer map
1158 : /// - Downloads remote index file and adds remote files to the layer map
1159 : /// - Schedules remote upload tasks for any files that are present locally but missing from remote storage.
1160 : ///
1161 : /// If the operation fails, the timeline is left in the tenant's hash map in Broken state. On success,
1162 : /// it is marked as Active.
1163 : #[allow(clippy::too_many_arguments)]
1164 3 : async fn timeline_init_and_sync(
1165 3 : self: &Arc<Self>,
1166 3 : timeline_id: TimelineId,
1167 3 : resources: TimelineResources,
1168 3 : index_part: IndexPart,
1169 3 : metadata: TimelineMetadata,
1170 3 : previous_heatmap: Option<PreviousHeatmap>,
1171 3 : ancestor: Option<Arc<Timeline>>,
1172 3 : cause: LoadTimelineCause,
1173 3 : ctx: &RequestContext,
1174 3 : ) -> anyhow::Result<TimelineInitAndSyncResult> {
1175 3 : let tenant_id = self.tenant_shard_id;
1176 :
1177 3 : let import_pgdata = index_part.import_pgdata.clone();
1178 3 : let idempotency = match &import_pgdata {
1179 0 : Some(import_pgdata) => {
1180 0 : CreateTimelineIdempotency::ImportPgdata(CreatingTimelineIdempotencyImportPgdata {
1181 0 : idempotency_key: import_pgdata.idempotency_key().clone(),
1182 0 : })
1183 : }
1184 : None => {
1185 3 : if metadata.ancestor_timeline().is_none() {
1186 2 : CreateTimelineIdempotency::Bootstrap {
1187 2 : pg_version: metadata.pg_version(),
1188 2 : }
1189 : } else {
1190 1 : CreateTimelineIdempotency::Branch {
1191 1 : ancestor_timeline_id: metadata.ancestor_timeline().unwrap(),
1192 1 : ancestor_start_lsn: metadata.ancestor_lsn(),
1193 1 : }
1194 : }
1195 : }
1196 : };
1197 :
1198 3 : let (timeline, _timeline_ctx) = self.create_timeline_struct(
1199 3 : timeline_id,
1200 3 : &metadata,
1201 3 : previous_heatmap,
1202 3 : ancestor.clone(),
1203 3 : resources,
1204 3 : CreateTimelineCause::Load,
1205 3 : idempotency.clone(),
1206 3 : index_part.gc_compaction.clone(),
1207 3 : index_part.rel_size_migration.clone(),
1208 3 : index_part.rel_size_migrated_at,
1209 3 : ctx,
1210 3 : )?;
1211 3 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
1212 :
1213 3 : if !disk_consistent_lsn.is_valid() {
1214 : // As opposed to normal timelines which get initialised with a disk consitent LSN
1215 : // via initdb, imported timelines start from 0. If the import task stops before
1216 : // it advances disk consitent LSN, allow it to resume.
1217 0 : let in_progress_import = import_pgdata
1218 0 : .as_ref()
1219 0 : .map(|import| !import.is_done())
1220 0 : .unwrap_or(false);
1221 0 : if !in_progress_import {
1222 0 : anyhow::bail!("Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn");
1223 0 : }
1224 3 : }
1225 :
1226 3 : assert_eq!(
1227 : disk_consistent_lsn,
1228 3 : metadata.disk_consistent_lsn(),
1229 0 : "these are used interchangeably"
1230 : );
1231 :
1232 3 : timeline.remote_client.init_upload_queue(&index_part)?;
1233 :
1234 3 : timeline
1235 3 : .load_layer_map(disk_consistent_lsn, index_part)
1236 3 : .await
1237 3 : .with_context(|| {
1238 0 : format!("Failed to load layermap for timeline {tenant_id}/{timeline_id}")
1239 0 : })?;
1240 :
1241 : // When unarchiving, we've mostly likely lost the heatmap generated prior
1242 : // to the archival operation. To allow warming this timeline up, generate
1243 : // a previous heatmap which contains all visible layers in the layer map.
1244 : // This previous heatmap will be used whenever a fresh heatmap is generated
1245 : // for the timeline.
1246 3 : if self.conf.generate_unarchival_heatmap && matches!(cause, LoadTimelineCause::Unoffload) {
1247 0 : let mut tline_ending_at = Some((&timeline, timeline.get_last_record_lsn()));
1248 0 : while let Some((tline, end_lsn)) = tline_ending_at {
1249 0 : let unarchival_heatmap = tline.generate_unarchival_heatmap(end_lsn).await;
1250 : // Another unearchived timeline might have generated a heatmap for this ancestor.
1251 : // If the current branch point greater than the previous one use the the heatmap
1252 : // we just generated - it should include more layers.
1253 0 : if !tline.should_keep_previous_heatmap(end_lsn) {
1254 0 : tline
1255 0 : .previous_heatmap
1256 0 : .store(Some(Arc::new(unarchival_heatmap)));
1257 0 : } else {
1258 0 : tracing::info!("Previous heatmap preferred. Dropping unarchival heatmap.")
1259 : }
1260 :
1261 0 : match tline.ancestor_timeline() {
1262 0 : Some(ancestor) => {
1263 0 : if ancestor.update_layer_visibility().await.is_err() {
1264 : // Ancestor timeline is shutting down.
1265 0 : break;
1266 0 : }
1267 :
1268 0 : tline_ending_at = Some((ancestor, tline.get_ancestor_lsn()));
1269 : }
1270 0 : None => {
1271 0 : tline_ending_at = None;
1272 0 : }
1273 : }
1274 : }
1275 3 : }
1276 :
1277 0 : match import_pgdata {
1278 0 : Some(import_pgdata) if !import_pgdata.is_done() => {
1279 0 : let mut guard = self.timelines_creating.lock().unwrap();
1280 0 : if !guard.insert(timeline_id) {
1281 : // We should never try and load the same timeline twice during startup
1282 0 : unreachable!("Timeline {tenant_id}/{timeline_id} is already being created")
1283 0 : }
1284 0 : let timeline_create_guard = TimelineCreateGuard {
1285 0 : _tenant_gate_guard: self.gate.enter()?,
1286 0 : owning_tenant: self.clone(),
1287 0 : timeline_id,
1288 0 : idempotency,
1289 : // The users of this specific return value don't need the timline_path in there.
1290 0 : timeline_path: timeline
1291 0 : .conf
1292 0 : .timeline_path(&timeline.tenant_shard_id, &timeline.timeline_id),
1293 : };
1294 0 : Ok(TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
1295 0 : TimelineInitAndSyncNeedsSpawnImportPgdata {
1296 0 : timeline,
1297 0 : import_pgdata,
1298 0 : guard: timeline_create_guard,
1299 0 : },
1300 0 : ))
1301 : }
1302 : Some(_) | None => {
1303 : {
1304 3 : let mut timelines_accessor = self.timelines.lock().unwrap();
1305 3 : match timelines_accessor.entry(timeline_id) {
1306 : // We should never try and load the same timeline twice during startup
1307 : Entry::Occupied(_) => {
1308 0 : unreachable!(
1309 : "Timeline {tenant_id}/{timeline_id} already exists in the tenant map"
1310 : );
1311 : }
1312 3 : Entry::Vacant(v) => {
1313 3 : v.insert(Arc::clone(&timeline));
1314 3 : timeline.maybe_spawn_flush_loop();
1315 3 : }
1316 : }
1317 : }
1318 :
1319 3 : if disk_consistent_lsn.is_valid() {
1320 : // Sanity check: a timeline should have some content.
1321 : // Exception: importing timelines might not yet have any
1322 3 : anyhow::ensure!(
1323 3 : ancestor.is_some()
1324 2 : || timeline
1325 2 : .layers
1326 2 : .read(LayerManagerLockHolder::LoadLayerMap)
1327 2 : .await
1328 2 : .layer_map()
1329 2 : .expect(
1330 2 : "currently loading, layer manager cannot be shutdown already"
1331 : )
1332 2 : .iter_historic_layers()
1333 2 : .next()
1334 2 : .is_some(),
1335 0 : "Timeline has no ancestor and no layer files"
1336 : );
1337 0 : }
1338 :
1339 3 : Ok(TimelineInitAndSyncResult::ReadyToActivate)
1340 : }
1341 : }
1342 3 : }
1343 :
1344 : /// Attach a tenant that's available in cloud storage.
1345 : ///
1346 : /// This returns quickly, after just creating the in-memory object
1347 : /// Tenant struct and launching a background task to download
1348 : /// the remote index files. On return, the tenant is most likely still in
1349 : /// Attaching state, and it will become Active once the background task
1350 : /// finishes. You can use wait_until_active() to wait for the task to
1351 : /// complete.
1352 : ///
1353 : #[allow(clippy::too_many_arguments)]
1354 0 : pub(crate) fn spawn(
1355 0 : conf: &'static PageServerConf,
1356 0 : tenant_shard_id: TenantShardId,
1357 0 : resources: TenantSharedResources,
1358 0 : attached_conf: AttachedTenantConf,
1359 0 : shard_identity: ShardIdentity,
1360 0 : init_order: Option<InitializationOrder>,
1361 0 : mode: SpawnMode,
1362 0 : ctx: &RequestContext,
1363 0 : ) -> Result<Arc<TenantShard>, GlobalShutDown> {
1364 0 : let wal_redo_manager =
1365 0 : WalRedoManager::new(PostgresRedoManager::new(conf, tenant_shard_id))?;
1366 :
1367 : let TenantSharedResources {
1368 0 : broker_client,
1369 0 : remote_storage,
1370 0 : deletion_queue_client,
1371 0 : l0_flush_global_state,
1372 0 : basebackup_cache,
1373 0 : feature_resolver,
1374 0 : } = resources;
1375 :
1376 0 : let attach_mode = attached_conf.location.attach_mode;
1377 0 : let generation = attached_conf.location.generation;
1378 :
1379 0 : let tenant = Arc::new(TenantShard::new(
1380 0 : TenantState::Attaching,
1381 0 : conf,
1382 0 : attached_conf,
1383 0 : shard_identity,
1384 0 : Some(wal_redo_manager),
1385 0 : tenant_shard_id,
1386 0 : remote_storage.clone(),
1387 0 : deletion_queue_client,
1388 0 : l0_flush_global_state,
1389 0 : basebackup_cache,
1390 0 : feature_resolver,
1391 : ));
1392 :
1393 : // The attach task will carry a GateGuard, so that shutdown() reliably waits for it to drop out if
1394 : // we shut down while attaching.
1395 0 : let attach_gate_guard = tenant
1396 0 : .gate
1397 0 : .enter()
1398 0 : .expect("We just created the TenantShard: nothing else can have shut it down yet");
1399 :
1400 : // Do all the hard work in the background
1401 0 : let tenant_clone = Arc::clone(&tenant);
1402 0 : let ctx = ctx.detached_child(TaskKind::Attach, DownloadBehavior::Warn);
1403 0 : task_mgr::spawn(
1404 0 : &tokio::runtime::Handle::current(),
1405 0 : TaskKind::Attach,
1406 0 : tenant_shard_id,
1407 0 : None,
1408 0 : "attach tenant",
1409 0 : async move {
1410 :
1411 0 : info!(
1412 : ?attach_mode,
1413 0 : "Attaching tenant"
1414 : );
1415 :
1416 0 : let _gate_guard = attach_gate_guard;
1417 :
1418 : // Is this tenant being spawned as part of process startup?
1419 0 : let starting_up = init_order.is_some();
1420 0 : scopeguard::defer! {
1421 : if starting_up {
1422 : TENANT.startup_complete.inc();
1423 : }
1424 : }
1425 :
1426 0 : fn make_broken_or_stopping(t: &TenantShard, err: anyhow::Error) {
1427 0 : t.state.send_modify(|state| match state {
1428 : // TODO: the old code alluded to DeleteTenantFlow sometimes setting
1429 : // TenantState::Stopping before we get here, but this may be outdated.
1430 : // Let's find out with a testing assertion. If this doesn't fire, and the
1431 : // logs don't show this happening in production, remove the Stopping cases.
1432 0 : TenantState::Stopping{..} if cfg!(any(test, feature = "testing")) => {
1433 0 : panic!("unexpected TenantState::Stopping during attach")
1434 : }
1435 : // If the tenant is cancelled, assume the error was caused by cancellation.
1436 0 : TenantState::Attaching if t.cancel.is_cancelled() => {
1437 0 : info!("attach cancelled, setting tenant state to Stopping: {err}");
1438 : // NB: progress None tells `set_stopping` that attach has cancelled.
1439 0 : *state = TenantState::Stopping { progress: None };
1440 : }
1441 : // According to the old code, DeleteTenantFlow may already have set this to
1442 : // Stopping. Retain its progress.
1443 : // TODO: there is no DeleteTenantFlow. Is this still needed? See above.
1444 0 : TenantState::Stopping { progress } if t.cancel.is_cancelled() => {
1445 0 : assert!(progress.is_some(), "concurrent attach cancellation");
1446 0 : info!("attach cancelled, already Stopping: {err}");
1447 : }
1448 : // Mark the tenant as broken.
1449 : TenantState::Attaching | TenantState::Stopping { .. } => {
1450 0 : error!("attach failed, setting tenant state to Broken (was {state}): {err:?}");
1451 0 : *state = TenantState::broken_from_reason(err.to_string())
1452 : }
1453 : // The attach task owns the tenant state until activated.
1454 0 : state => panic!("invalid tenant state {state} during attach: {err:?}"),
1455 0 : });
1456 0 : }
1457 :
1458 : // TODO: should also be rejecting tenant conf changes that violate this check.
1459 0 : if let Err(e) = crate::tenant::storage_layer::inmemory_layer::IndexEntry::validate_checkpoint_distance(tenant_clone.get_checkpoint_distance()) {
1460 0 : make_broken_or_stopping(&tenant_clone, anyhow::anyhow!(e));
1461 0 : return Ok(());
1462 0 : }
1463 :
1464 0 : let mut init_order = init_order;
1465 : // take the completion because initial tenant loading will complete when all of
1466 : // these tasks complete.
1467 0 : let _completion = init_order
1468 0 : .as_mut()
1469 0 : .and_then(|x| x.initial_tenant_load.take());
1470 0 : let remote_load_completion = init_order
1471 0 : .as_mut()
1472 0 : .and_then(|x| x.initial_tenant_load_remote.take());
1473 :
1474 : enum AttachType<'a> {
1475 : /// We are attaching this tenant lazily in the background.
1476 : Warmup {
1477 : _permit: tokio::sync::SemaphorePermit<'a>,
1478 : during_startup: bool
1479 : },
1480 : /// We are attaching this tenant as soon as we can, because for example an
1481 : /// endpoint tried to access it.
1482 : OnDemand,
1483 : /// During normal operations after startup, we are attaching a tenant, and
1484 : /// eager attach was requested.
1485 : Normal,
1486 : }
1487 :
1488 0 : let attach_type = if matches!(mode, SpawnMode::Lazy) {
1489 : // Before doing any I/O, wait for at least one of:
1490 : // - A client attempting to access to this tenant (on-demand loading)
1491 : // - A permit becoming available in the warmup semaphore (background warmup)
1492 :
1493 0 : tokio::select!(
1494 0 : permit = tenant_clone.activate_now_sem.acquire() => {
1495 0 : let _ = permit.expect("activate_now_sem is never closed");
1496 0 : tracing::info!("Activating tenant (on-demand)");
1497 0 : AttachType::OnDemand
1498 : },
1499 0 : permit = conf.concurrent_tenant_warmup.inner().acquire() => {
1500 0 : let _permit = permit.expect("concurrent_tenant_warmup semaphore is never closed");
1501 0 : tracing::info!("Activating tenant (warmup)");
1502 0 : AttachType::Warmup {
1503 0 : _permit,
1504 0 : during_startup: init_order.is_some()
1505 0 : }
1506 : }
1507 0 : _ = tenant_clone.cancel.cancelled() => {
1508 : // This is safe, but should be pretty rare: it is interesting if a tenant
1509 : // stayed in Activating for such a long time that shutdown found it in
1510 : // that state.
1511 0 : tracing::info!(state=%tenant_clone.current_state(), "Tenant shut down before activation");
1512 : // Set the tenant to Stopping to signal `set_stopping` that we're done.
1513 0 : make_broken_or_stopping(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"));
1514 0 : return Ok(());
1515 : },
1516 : )
1517 : } else {
1518 : // SpawnMode::{Create,Eager} always cause jumping ahead of the
1519 : // concurrent_tenant_warmup queue
1520 0 : AttachType::Normal
1521 : };
1522 :
1523 0 : let preload = match &mode {
1524 : SpawnMode::Eager | SpawnMode::Lazy => {
1525 0 : let _preload_timer = TENANT.preload.start_timer();
1526 0 : let res = tenant_clone
1527 0 : .preload(&remote_storage, task_mgr::shutdown_token())
1528 0 : .await;
1529 0 : match res {
1530 0 : Ok(p) => Some(p),
1531 0 : Err(e) => {
1532 0 : make_broken_or_stopping(&tenant_clone, anyhow::anyhow!(e));
1533 0 : return Ok(());
1534 : }
1535 : }
1536 : }
1537 :
1538 : };
1539 :
1540 : // Remote preload is complete.
1541 0 : drop(remote_load_completion);
1542 :
1543 :
1544 : // We will time the duration of the attach phase unless this is a creation (attach will do no work)
1545 0 : let attach_start = std::time::Instant::now();
1546 0 : let attached = {
1547 0 : let _attach_timer = Some(TENANT.attach.start_timer());
1548 0 : tenant_clone.attach(preload, &ctx).await
1549 : };
1550 0 : let attach_duration = attach_start.elapsed();
1551 0 : _ = tenant_clone.attach_wal_lag_cooldown.set(WalLagCooldown::new(attach_start, attach_duration));
1552 :
1553 0 : match attached {
1554 : Ok(()) => {
1555 0 : info!("attach finished, activating");
1556 0 : tenant_clone.activate(broker_client, None, &ctx);
1557 : }
1558 0 : Err(e) => make_broken_or_stopping(&tenant_clone, anyhow::anyhow!(e)),
1559 : }
1560 :
1561 : // If we are doing an opportunistic warmup attachment at startup, initialize
1562 : // logical size at the same time. This is better than starting a bunch of idle tenants
1563 : // with cold caches and then coming back later to initialize their logical sizes.
1564 : //
1565 : // It also prevents the warmup proccess competing with the concurrency limit on
1566 : // logical size calculations: if logical size calculation semaphore is saturated,
1567 : // then warmup will wait for that before proceeding to the next tenant.
1568 0 : if matches!(attach_type, AttachType::Warmup { during_startup: true, .. }) {
1569 0 : let mut futs: FuturesUnordered<_> = tenant_clone.timelines.lock().unwrap().values().cloned().map(|t| t.await_initial_logical_size()).collect();
1570 0 : tracing::info!("Waiting for initial logical sizes while warming up...");
1571 0 : while futs.next().await.is_some() {}
1572 0 : tracing::info!("Warm-up complete");
1573 0 : }
1574 :
1575 0 : Ok(())
1576 0 : }
1577 0 : .instrument(tracing::info_span!(parent: None, "attach", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), gen=?generation)),
1578 : );
1579 0 : Ok(tenant)
1580 0 : }
1581 :
1582 : #[instrument(skip_all)]
1583 : pub(crate) async fn preload(
1584 : self: &Arc<Self>,
1585 : remote_storage: &GenericRemoteStorage,
1586 : cancel: CancellationToken,
1587 : ) -> anyhow::Result<TenantPreload> {
1588 : span::debug_assert_current_span_has_tenant_id();
1589 : // Get list of remote timelines
1590 : // download index files for every tenant timeline
1591 : info!("listing remote timelines");
1592 : let (mut remote_timeline_ids, other_keys) = remote_timeline_client::list_remote_timelines(
1593 : remote_storage,
1594 : self.tenant_shard_id,
1595 : cancel.clone(),
1596 : )
1597 : .await?;
1598 :
1599 : let tenant_manifest = match download_tenant_manifest(
1600 : remote_storage,
1601 : &self.tenant_shard_id,
1602 : self.generation,
1603 : &cancel,
1604 : )
1605 : .await
1606 : {
1607 : Ok((tenant_manifest, _, _)) => Some(tenant_manifest),
1608 : Err(DownloadError::NotFound) => None,
1609 : Err(err) => return Err(err.into()),
1610 : };
1611 :
1612 : info!(
1613 : "found {} timelines ({} offloaded timelines)",
1614 : remote_timeline_ids.len(),
1615 : tenant_manifest
1616 : .as_ref()
1617 3 : .map(|m| m.offloaded_timelines.len())
1618 : .unwrap_or(0)
1619 : );
1620 :
1621 : for k in other_keys {
1622 : warn!("Unexpected non timeline key {k}");
1623 : }
1624 :
1625 : // Avoid downloading IndexPart of offloaded timelines.
1626 : let mut offloaded_with_prefix = HashSet::new();
1627 : if let Some(tenant_manifest) = &tenant_manifest {
1628 : for offloaded in tenant_manifest.offloaded_timelines.iter() {
1629 : if remote_timeline_ids.remove(&offloaded.timeline_id) {
1630 : offloaded_with_prefix.insert(offloaded.timeline_id);
1631 : } else {
1632 : // We'll take care later of timelines in the manifest without a prefix
1633 : }
1634 : }
1635 : }
1636 :
1637 : // TODO(vlad): Could go to S3 if the secondary is freezing cold and hasn't even
1638 : // pulled the first heatmap. Not entirely necessary since the storage controller
1639 : // will kick the secondary in any case and cause a download.
1640 : let maybe_heatmap_at = self.read_on_disk_heatmap().await;
1641 :
1642 : let timelines = self
1643 : .load_timelines_metadata(
1644 : remote_timeline_ids,
1645 : remote_storage,
1646 : maybe_heatmap_at,
1647 : cancel,
1648 : )
1649 : .await?;
1650 :
1651 : Ok(TenantPreload {
1652 : tenant_manifest,
1653 : timelines: timelines
1654 : .into_iter()
1655 3 : .map(|(id, tl)| (id, Some(tl)))
1656 0 : .chain(offloaded_with_prefix.into_iter().map(|id| (id, None)))
1657 : .collect(),
1658 : })
1659 : }
1660 :
1661 119 : async fn read_on_disk_heatmap(&self) -> Option<(HeatMapTenant, std::time::Instant)> {
1662 119 : if !self.conf.load_previous_heatmap {
1663 0 : return None;
1664 119 : }
1665 :
1666 119 : let on_disk_heatmap_path = self.conf.tenant_heatmap_path(&self.tenant_shard_id);
1667 119 : match tokio::fs::read_to_string(on_disk_heatmap_path).await {
1668 0 : Ok(heatmap) => match serde_json::from_str::<HeatMapTenant>(&heatmap) {
1669 0 : Ok(heatmap) => Some((heatmap, std::time::Instant::now())),
1670 0 : Err(err) => {
1671 0 : error!("Failed to deserialize old heatmap: {err}");
1672 0 : None
1673 : }
1674 : },
1675 119 : Err(err) => match err.kind() {
1676 119 : std::io::ErrorKind::NotFound => None,
1677 : _ => {
1678 0 : error!("Unexpected IO error reading old heatmap: {err}");
1679 0 : None
1680 : }
1681 : },
1682 : }
1683 119 : }
1684 :
1685 : ///
1686 : /// Background task that downloads all data for a tenant and brings it to Active state.
1687 : ///
1688 : /// No background tasks are started as part of this routine.
1689 : ///
1690 119 : async fn attach(
1691 119 : self: &Arc<TenantShard>,
1692 119 : preload: Option<TenantPreload>,
1693 119 : ctx: &RequestContext,
1694 119 : ) -> anyhow::Result<()> {
1695 119 : span::debug_assert_current_span_has_tenant_id();
1696 :
1697 119 : failpoint_support::sleep_millis_async!("before-attaching-tenant");
1698 :
1699 119 : let Some(preload) = preload else {
1700 0 : anyhow::bail!(
1701 0 : "local-only deployment is no longer supported, https://github.com/neondatabase/neon/issues/5624"
1702 : );
1703 : };
1704 :
1705 119 : let mut offloaded_timeline_ids = HashSet::new();
1706 119 : let mut offloaded_timelines_list = Vec::new();
1707 119 : if let Some(tenant_manifest) = &preload.tenant_manifest {
1708 3 : for timeline_manifest in tenant_manifest.offloaded_timelines.iter() {
1709 0 : let timeline_id = timeline_manifest.timeline_id;
1710 0 : let offloaded_timeline =
1711 0 : OffloadedTimeline::from_manifest(self.tenant_shard_id, timeline_manifest);
1712 0 : offloaded_timelines_list.push((timeline_id, Arc::new(offloaded_timeline)));
1713 0 : offloaded_timeline_ids.insert(timeline_id);
1714 0 : }
1715 116 : }
1716 : // Complete deletions for offloaded timeline id's from manifest.
1717 : // The manifest will be uploaded later in this function.
1718 119 : offloaded_timelines_list
1719 119 : .retain(|(offloaded_id, offloaded)| {
1720 : // Existence of a timeline is finally determined by the existence of an index-part.json in remote storage.
1721 : // If there is dangling references in another location, they need to be cleaned up.
1722 0 : let delete = !preload.timelines.contains_key(offloaded_id);
1723 0 : if delete {
1724 0 : tracing::info!("Removing offloaded timeline {offloaded_id} from manifest as no remote prefix was found");
1725 0 : offloaded.defuse_for_tenant_drop();
1726 0 : }
1727 0 : !delete
1728 0 : });
1729 :
1730 119 : let mut timelines_to_resume_deletions = vec![];
1731 :
1732 119 : let mut remote_index_and_client = HashMap::new();
1733 119 : let mut timeline_ancestors = HashMap::new();
1734 119 : let mut existent_timelines = HashSet::new();
1735 122 : for (timeline_id, preload) in preload.timelines {
1736 3 : let Some(preload) = preload else { continue };
1737 : // This is an invariant of the `preload` function's API
1738 3 : assert!(!offloaded_timeline_ids.contains(&timeline_id));
1739 3 : let index_part = match preload.index_part {
1740 3 : Ok(i) => {
1741 3 : debug!("remote index part exists for timeline {timeline_id}");
1742 : // We found index_part on the remote, this is the standard case.
1743 3 : existent_timelines.insert(timeline_id);
1744 3 : i
1745 : }
1746 : Err(DownloadError::NotFound) => {
1747 : // There is no index_part on the remote. We only get here
1748 : // if there is some prefix for the timeline in the remote storage.
1749 : // This can e.g. be the initdb.tar.zst archive, maybe a
1750 : // remnant from a prior incomplete creation or deletion attempt.
1751 : // Delete the local directory as the deciding criterion for a
1752 : // timeline's existence is presence of index_part.
1753 0 : info!(%timeline_id, "index_part not found on remote");
1754 0 : continue;
1755 : }
1756 0 : Err(DownloadError::Fatal(why)) => {
1757 : // If, while loading one remote timeline, we saw an indication that our generation
1758 : // number is likely invalid, then we should not load the whole tenant.
1759 0 : error!(%timeline_id, "Fatal error loading timeline: {why}");
1760 0 : anyhow::bail!(why.to_string());
1761 : }
1762 0 : Err(e) => {
1763 : // Some (possibly ephemeral) error happened during index_part download.
1764 : // Pretend the timeline exists to not delete the timeline directory,
1765 : // as it might be a temporary issue and we don't want to re-download
1766 : // everything after it resolves.
1767 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1768 :
1769 0 : existent_timelines.insert(timeline_id);
1770 0 : continue;
1771 : }
1772 : };
1773 3 : match index_part {
1774 3 : MaybeDeletedIndexPart::IndexPart(index_part) => {
1775 3 : timeline_ancestors.insert(timeline_id, index_part.metadata.clone());
1776 3 : remote_index_and_client.insert(
1777 3 : timeline_id,
1778 3 : (index_part, preload.client, preload.previous_heatmap),
1779 3 : );
1780 3 : }
1781 0 : MaybeDeletedIndexPart::Deleted(index_part) => {
1782 0 : info!(
1783 0 : "timeline {} is deleted, picking to resume deletion",
1784 : timeline_id
1785 : );
1786 0 : timelines_to_resume_deletions.push((timeline_id, index_part, preload.client));
1787 : }
1788 : }
1789 : }
1790 :
1791 119 : let mut gc_blocks = HashMap::new();
1792 :
1793 : // For every timeline, download the metadata file, scan the local directory,
1794 : // and build a layer map that contains an entry for each remote and local
1795 : // layer file.
1796 119 : let sorted_timelines = tree_sort_timelines(timeline_ancestors, |m| m.ancestor_timeline())?;
1797 122 : for (timeline_id, remote_metadata) in sorted_timelines {
1798 3 : let (index_part, remote_client, previous_heatmap) = remote_index_and_client
1799 3 : .remove(&timeline_id)
1800 3 : .expect("just put it in above");
1801 :
1802 3 : if let Some(blocking) = index_part.gc_blocking.as_ref() {
1803 : // could just filter these away, but it helps while testing
1804 0 : anyhow::ensure!(
1805 0 : !blocking.reasons.is_empty(),
1806 0 : "index_part for {timeline_id} is malformed: it should not have gc blocking with zero reasons"
1807 : );
1808 0 : let prev = gc_blocks.insert(timeline_id, blocking.reasons);
1809 0 : assert!(prev.is_none());
1810 3 : }
1811 :
1812 : // TODO again handle early failure
1813 3 : let effect = self
1814 3 : .load_remote_timeline(
1815 3 : timeline_id,
1816 3 : index_part,
1817 3 : remote_metadata,
1818 3 : previous_heatmap,
1819 3 : self.get_timeline_resources_for(remote_client),
1820 3 : LoadTimelineCause::Attach,
1821 3 : ctx,
1822 3 : )
1823 3 : .await
1824 3 : .with_context(|| {
1825 0 : format!(
1826 0 : "failed to load remote timeline {} for tenant {}",
1827 0 : timeline_id, self.tenant_shard_id
1828 : )
1829 0 : })?;
1830 :
1831 3 : match effect {
1832 3 : TimelineInitAndSyncResult::ReadyToActivate => {
1833 3 : // activation happens later, on Tenant::activate
1834 3 : }
1835 : TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
1836 : TimelineInitAndSyncNeedsSpawnImportPgdata {
1837 0 : timeline,
1838 0 : import_pgdata,
1839 0 : guard,
1840 : },
1841 : ) => {
1842 0 : let timeline_id = timeline.timeline_id;
1843 0 : let import_task_gate = Gate::default();
1844 0 : let import_task_guard = import_task_gate.enter().unwrap();
1845 0 : let import_task_handle =
1846 0 : tokio::task::spawn(self.clone().create_timeline_import_pgdata_task(
1847 0 : timeline.clone(),
1848 0 : import_pgdata,
1849 0 : guard,
1850 0 : import_task_guard,
1851 0 : ctx.detached_child(TaskKind::ImportPgdata, DownloadBehavior::Warn),
1852 : ));
1853 :
1854 0 : let prev = self.timelines_importing.lock().unwrap().insert(
1855 0 : timeline_id,
1856 0 : Arc::new(ImportingTimeline {
1857 0 : timeline: timeline.clone(),
1858 0 : import_task_handle,
1859 0 : import_task_gate,
1860 0 : delete_progress: TimelineDeleteProgress::default(),
1861 0 : }),
1862 0 : );
1863 :
1864 0 : assert!(prev.is_none());
1865 : }
1866 : }
1867 : }
1868 :
1869 : // At this point we've initialized all timelines and are tracking them.
1870 : // Now compute the layer visibility for all (not offloaded) timelines.
1871 119 : let compute_visiblity_for = {
1872 119 : let timelines_accessor = self.timelines.lock().unwrap();
1873 119 : let mut timelines_offloaded_accessor = self.timelines_offloaded.lock().unwrap();
1874 :
1875 119 : timelines_offloaded_accessor.extend(offloaded_timelines_list.into_iter());
1876 :
1877 : // Before activation, populate each Timeline's GcInfo with information about its children
1878 119 : self.initialize_gc_info(&timelines_accessor, &timelines_offloaded_accessor, None);
1879 :
1880 119 : timelines_accessor.values().cloned().collect::<Vec<_>>()
1881 : };
1882 :
1883 122 : for tl in compute_visiblity_for {
1884 3 : tl.update_layer_visibility().await.with_context(|| {
1885 0 : format!(
1886 0 : "failed initial timeline visibility computation {} for tenant {}",
1887 0 : tl.timeline_id, self.tenant_shard_id
1888 : )
1889 0 : })?;
1890 : }
1891 :
1892 : // Walk through deleted timelines, resume deletion
1893 119 : for (timeline_id, index_part, remote_timeline_client) in timelines_to_resume_deletions {
1894 0 : remote_timeline_client
1895 0 : .init_upload_queue_stopped_to_continue_deletion(&index_part)
1896 0 : .context("init queue stopped")
1897 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1898 :
1899 0 : DeleteTimelineFlow::resume_deletion(
1900 0 : Arc::clone(self),
1901 0 : timeline_id,
1902 0 : &index_part.metadata,
1903 0 : remote_timeline_client,
1904 0 : ctx,
1905 : )
1906 0 : .instrument(tracing::info_span!("timeline_delete", %timeline_id))
1907 0 : .await
1908 0 : .context("resume_deletion")
1909 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1910 : }
1911 :
1912 : // Stash the preloaded tenant manifest, and upload a new manifest if changed.
1913 : //
1914 : // NB: this must happen after the tenant is fully populated above. In particular the
1915 : // offloaded timelines, which are included in the manifest.
1916 : {
1917 119 : let mut guard = self.remote_tenant_manifest.lock().await;
1918 119 : assert!(guard.is_none(), "tenant manifest set before preload"); // first populated here
1919 119 : *guard = preload.tenant_manifest;
1920 : }
1921 119 : self.maybe_upload_tenant_manifest().await?;
1922 :
1923 : // The local filesystem contents are a cache of what's in the remote IndexPart;
1924 : // IndexPart is the source of truth.
1925 119 : self.clean_up_timelines(&existent_timelines)?;
1926 :
1927 119 : self.gc_block.set_scanned(gc_blocks);
1928 :
1929 119 : fail::fail_point!("attach-before-activate", |_| {
1930 0 : anyhow::bail!("attach-before-activate");
1931 0 : });
1932 119 : failpoint_support::sleep_millis_async!("attach-before-activate-sleep", &self.cancel);
1933 :
1934 119 : info!("Done");
1935 :
1936 119 : Ok(())
1937 119 : }
1938 :
1939 : /// Check for any local timeline directories that are temporary, or do not correspond to a
1940 : /// timeline that still exists: this can happen if we crashed during a deletion/creation, or
1941 : /// if a timeline was deleted while the tenant was attached to a different pageserver.
1942 119 : fn clean_up_timelines(&self, existent_timelines: &HashSet<TimelineId>) -> anyhow::Result<()> {
1943 119 : let timelines_dir = self.conf.timelines_path(&self.tenant_shard_id);
1944 :
1945 119 : let entries = match timelines_dir.read_dir_utf8() {
1946 119 : Ok(d) => d,
1947 0 : Err(e) => {
1948 0 : if e.kind() == std::io::ErrorKind::NotFound {
1949 0 : return Ok(());
1950 : } else {
1951 0 : return Err(e).context("list timelines directory for tenant");
1952 : }
1953 : }
1954 : };
1955 :
1956 123 : for entry in entries {
1957 4 : let entry = entry.context("read timeline dir entry")?;
1958 4 : let entry_path = entry.path();
1959 :
1960 4 : let purge = if crate::is_temporary(entry_path) {
1961 0 : true
1962 : } else {
1963 4 : match TimelineId::try_from(entry_path.file_name()) {
1964 4 : Ok(i) => {
1965 : // Purge if the timeline ID does not exist in remote storage: remote storage is the authority.
1966 4 : !existent_timelines.contains(&i)
1967 : }
1968 0 : Err(e) => {
1969 0 : tracing::warn!(
1970 0 : "Unparseable directory in timelines directory: {entry_path}, ignoring ({e})"
1971 : );
1972 : // Do not purge junk: if we don't recognize it, be cautious and leave it for a human.
1973 0 : false
1974 : }
1975 : }
1976 : };
1977 :
1978 4 : if purge {
1979 1 : tracing::info!("Purging stale timeline dentry {entry_path}");
1980 1 : if let Err(e) = match entry.file_type() {
1981 1 : Ok(t) => if t.is_dir() {
1982 1 : std::fs::remove_dir_all(entry_path)
1983 : } else {
1984 0 : std::fs::remove_file(entry_path)
1985 : }
1986 1 : .or_else(fs_ext::ignore_not_found),
1987 0 : Err(e) => Err(e),
1988 : } {
1989 0 : tracing::warn!("Failed to purge stale timeline dentry {entry_path}: {e}");
1990 1 : }
1991 3 : }
1992 : }
1993 :
1994 119 : Ok(())
1995 119 : }
1996 :
1997 : /// Get sum of all remote timelines sizes
1998 : ///
1999 : /// This function relies on the index_part instead of listing the remote storage
2000 0 : pub fn remote_size(&self) -> u64 {
2001 0 : let mut size = 0;
2002 :
2003 0 : for timeline in self.list_timelines() {
2004 0 : size += timeline.remote_client.get_remote_physical_size();
2005 0 : }
2006 :
2007 0 : size
2008 0 : }
2009 :
2010 : #[instrument(skip_all, fields(timeline_id=%timeline_id))]
2011 : #[allow(clippy::too_many_arguments)]
2012 : async fn load_remote_timeline(
2013 : self: &Arc<Self>,
2014 : timeline_id: TimelineId,
2015 : index_part: IndexPart,
2016 : remote_metadata: TimelineMetadata,
2017 : previous_heatmap: Option<PreviousHeatmap>,
2018 : resources: TimelineResources,
2019 : cause: LoadTimelineCause,
2020 : ctx: &RequestContext,
2021 : ) -> anyhow::Result<TimelineInitAndSyncResult> {
2022 : span::debug_assert_current_span_has_tenant_id();
2023 :
2024 : info!("downloading index file for timeline {}", timeline_id);
2025 : tokio::fs::create_dir_all(self.conf.timeline_path(&self.tenant_shard_id, &timeline_id))
2026 : .await
2027 : .context("Failed to create new timeline directory")?;
2028 :
2029 : let ancestor = if let Some(ancestor_id) = remote_metadata.ancestor_timeline() {
2030 : let timelines = self.timelines.lock().unwrap();
2031 : Some(Arc::clone(timelines.get(&ancestor_id).ok_or_else(
2032 0 : || {
2033 0 : anyhow::anyhow!(
2034 0 : "cannot find ancestor timeline {ancestor_id} for timeline {timeline_id}"
2035 : )
2036 0 : },
2037 : )?))
2038 : } else {
2039 : None
2040 : };
2041 :
2042 : self.timeline_init_and_sync(
2043 : timeline_id,
2044 : resources,
2045 : index_part,
2046 : remote_metadata,
2047 : previous_heatmap,
2048 : ancestor,
2049 : cause,
2050 : ctx,
2051 : )
2052 : .await
2053 : }
2054 :
2055 119 : async fn load_timelines_metadata(
2056 119 : self: &Arc<TenantShard>,
2057 119 : timeline_ids: HashSet<TimelineId>,
2058 119 : remote_storage: &GenericRemoteStorage,
2059 119 : heatmap: Option<(HeatMapTenant, std::time::Instant)>,
2060 119 : cancel: CancellationToken,
2061 119 : ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
2062 119 : let mut timeline_heatmaps = heatmap.map(|h| (h.0.into_timelines_index(), h.1));
2063 :
2064 119 : let mut part_downloads = JoinSet::new();
2065 122 : for timeline_id in timeline_ids {
2066 3 : let cancel_clone = cancel.clone();
2067 :
2068 3 : let previous_timeline_heatmap = timeline_heatmaps.as_mut().and_then(|hs| {
2069 0 : hs.0.remove(&timeline_id).map(|h| PreviousHeatmap::Active {
2070 0 : heatmap: h,
2071 0 : read_at: hs.1,
2072 0 : end_lsn: None,
2073 0 : })
2074 0 : });
2075 3 : part_downloads.spawn(
2076 3 : self.load_timeline_metadata(
2077 3 : timeline_id,
2078 3 : remote_storage.clone(),
2079 3 : previous_timeline_heatmap,
2080 3 : cancel_clone,
2081 : )
2082 3 : .instrument(info_span!("download_index_part", %timeline_id)),
2083 : );
2084 : }
2085 :
2086 119 : let mut timeline_preloads: HashMap<TimelineId, TimelinePreload> = HashMap::new();
2087 :
2088 : loop {
2089 122 : tokio::select!(
2090 122 : next = part_downloads.join_next() => {
2091 122 : match next {
2092 3 : Some(result) => {
2093 3 : let preload = result.context("join preload task")?;
2094 3 : timeline_preloads.insert(preload.timeline_id, preload);
2095 : },
2096 : None => {
2097 119 : break;
2098 : }
2099 : }
2100 : },
2101 122 : _ = cancel.cancelled() => {
2102 0 : anyhow::bail!("Cancelled while waiting for remote index download")
2103 : }
2104 : )
2105 : }
2106 :
2107 119 : Ok(timeline_preloads)
2108 119 : }
2109 :
2110 3 : fn build_timeline_client(
2111 3 : &self,
2112 3 : timeline_id: TimelineId,
2113 3 : remote_storage: GenericRemoteStorage,
2114 3 : ) -> RemoteTimelineClient {
2115 3 : RemoteTimelineClient::new(
2116 3 : remote_storage.clone(),
2117 3 : self.deletion_queue_client.clone(),
2118 3 : self.conf,
2119 3 : self.tenant_shard_id,
2120 3 : timeline_id,
2121 3 : self.generation,
2122 3 : &self.tenant_conf.load().location,
2123 : )
2124 3 : }
2125 :
2126 3 : fn load_timeline_metadata(
2127 3 : self: &Arc<TenantShard>,
2128 3 : timeline_id: TimelineId,
2129 3 : remote_storage: GenericRemoteStorage,
2130 3 : previous_heatmap: Option<PreviousHeatmap>,
2131 3 : cancel: CancellationToken,
2132 3 : ) -> impl Future<Output = TimelinePreload> + use<> {
2133 3 : let client = self.build_timeline_client(timeline_id, remote_storage);
2134 3 : async move {
2135 3 : debug_assert_current_span_has_tenant_and_timeline_id();
2136 3 : debug!("starting index part download");
2137 :
2138 3 : let index_part = client.download_index_file(&cancel).await;
2139 :
2140 3 : debug!("finished index part download");
2141 :
2142 3 : TimelinePreload {
2143 3 : client,
2144 3 : timeline_id,
2145 3 : index_part,
2146 3 : previous_heatmap,
2147 3 : }
2148 3 : }
2149 3 : }
2150 :
2151 0 : fn check_to_be_archived_has_no_unarchived_children(
2152 0 : timeline_id: TimelineId,
2153 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
2154 0 : ) -> Result<(), TimelineArchivalError> {
2155 0 : let children: Vec<TimelineId> = timelines
2156 0 : .iter()
2157 0 : .filter_map(|(id, entry)| {
2158 0 : if entry.get_ancestor_timeline_id() != Some(timeline_id) {
2159 0 : return None;
2160 0 : }
2161 0 : if entry.is_archived() == Some(true) {
2162 0 : return None;
2163 0 : }
2164 0 : Some(*id)
2165 0 : })
2166 0 : .collect();
2167 :
2168 0 : if !children.is_empty() {
2169 0 : return Err(TimelineArchivalError::HasUnarchivedChildren(children));
2170 0 : }
2171 0 : Ok(())
2172 0 : }
2173 :
2174 0 : fn check_ancestor_of_to_be_unarchived_is_not_archived(
2175 0 : ancestor_timeline_id: TimelineId,
2176 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
2177 0 : offloaded_timelines: &std::sync::MutexGuard<
2178 0 : '_,
2179 0 : HashMap<TimelineId, Arc<OffloadedTimeline>>,
2180 0 : >,
2181 0 : ) -> Result<(), TimelineArchivalError> {
2182 0 : let has_archived_parent =
2183 0 : if let Some(ancestor_timeline) = timelines.get(&ancestor_timeline_id) {
2184 0 : ancestor_timeline.is_archived() == Some(true)
2185 0 : } else if offloaded_timelines.contains_key(&ancestor_timeline_id) {
2186 0 : true
2187 : } else {
2188 0 : error!("ancestor timeline {ancestor_timeline_id} not found");
2189 0 : if cfg!(debug_assertions) {
2190 0 : panic!("ancestor timeline {ancestor_timeline_id} not found");
2191 0 : }
2192 0 : return Err(TimelineArchivalError::NotFound);
2193 : };
2194 0 : if has_archived_parent {
2195 0 : return Err(TimelineArchivalError::HasArchivedParent(
2196 0 : ancestor_timeline_id,
2197 0 : ));
2198 0 : }
2199 0 : Ok(())
2200 0 : }
2201 :
2202 0 : fn check_to_be_unarchived_timeline_has_no_archived_parent(
2203 0 : timeline: &Arc<Timeline>,
2204 0 : ) -> Result<(), TimelineArchivalError> {
2205 0 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
2206 0 : if ancestor_timeline.is_archived() == Some(true) {
2207 0 : return Err(TimelineArchivalError::HasArchivedParent(
2208 0 : ancestor_timeline.timeline_id,
2209 0 : ));
2210 0 : }
2211 0 : }
2212 0 : Ok(())
2213 0 : }
2214 :
2215 : /// Loads the specified (offloaded) timeline from S3 and attaches it as a loaded timeline
2216 : ///
2217 : /// Counterpart to [`offload_timeline`].
2218 0 : async fn unoffload_timeline(
2219 0 : self: &Arc<Self>,
2220 0 : timeline_id: TimelineId,
2221 0 : broker_client: storage_broker::BrokerClientChannel,
2222 0 : ctx: RequestContext,
2223 0 : ) -> Result<Arc<Timeline>, TimelineArchivalError> {
2224 0 : info!("unoffloading timeline");
2225 :
2226 : // We activate the timeline below manually, so this must be called on an active tenant.
2227 : // We expect callers of this function to ensure this.
2228 0 : match self.current_state() {
2229 : TenantState::Activating { .. }
2230 : | TenantState::Attaching
2231 : | TenantState::Broken { .. } => {
2232 0 : panic!("Timeline expected to be active")
2233 : }
2234 0 : TenantState::Stopping { .. } => return Err(TimelineArchivalError::Cancelled),
2235 0 : TenantState::Active => {}
2236 : }
2237 0 : let cancel = self.cancel.clone();
2238 :
2239 : // Protect against concurrent attempts to use this TimelineId
2240 : // We don't care much about idempotency, as it's ensured a layer above.
2241 0 : let allow_offloaded = true;
2242 0 : let _create_guard = self
2243 0 : .create_timeline_create_guard(
2244 0 : timeline_id,
2245 0 : CreateTimelineIdempotency::FailWithConflict,
2246 0 : allow_offloaded,
2247 : )
2248 0 : .map_err(|err| match err {
2249 0 : TimelineExclusionError::AlreadyCreating => TimelineArchivalError::AlreadyInProgress,
2250 : TimelineExclusionError::AlreadyExists { .. } => {
2251 0 : TimelineArchivalError::Other(anyhow::anyhow!("Timeline already exists"))
2252 : }
2253 0 : TimelineExclusionError::Other(e) => TimelineArchivalError::Other(e),
2254 0 : TimelineExclusionError::ShuttingDown => TimelineArchivalError::Cancelled,
2255 0 : })?;
2256 :
2257 0 : let timeline_preload = self
2258 0 : .load_timeline_metadata(
2259 0 : timeline_id,
2260 0 : self.remote_storage.clone(),
2261 0 : None,
2262 0 : cancel.clone(),
2263 0 : )
2264 0 : .await;
2265 :
2266 0 : let index_part = match timeline_preload.index_part {
2267 0 : Ok(index_part) => {
2268 0 : debug!("remote index part exists for timeline {timeline_id}");
2269 0 : index_part
2270 : }
2271 : Err(DownloadError::NotFound) => {
2272 0 : error!(%timeline_id, "index_part not found on remote");
2273 0 : return Err(TimelineArchivalError::NotFound);
2274 : }
2275 0 : Err(DownloadError::Cancelled) => return Err(TimelineArchivalError::Cancelled),
2276 0 : Err(e) => {
2277 : // Some (possibly ephemeral) error happened during index_part download.
2278 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
2279 0 : return Err(TimelineArchivalError::Other(
2280 0 : anyhow::Error::new(e).context("downloading index_part from remote storage"),
2281 0 : ));
2282 : }
2283 : };
2284 0 : let index_part = match index_part {
2285 0 : MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
2286 0 : MaybeDeletedIndexPart::Deleted(_index_part) => {
2287 0 : info!("timeline is deleted according to index_part.json");
2288 0 : return Err(TimelineArchivalError::NotFound);
2289 : }
2290 : };
2291 0 : let remote_metadata = index_part.metadata.clone();
2292 0 : let timeline_resources = self.build_timeline_resources(timeline_id);
2293 0 : self.load_remote_timeline(
2294 0 : timeline_id,
2295 0 : index_part,
2296 0 : remote_metadata,
2297 0 : None,
2298 0 : timeline_resources,
2299 0 : LoadTimelineCause::Unoffload,
2300 0 : &ctx,
2301 0 : )
2302 0 : .await
2303 0 : .with_context(|| {
2304 0 : format!(
2305 0 : "failed to load remote timeline {} for tenant {}",
2306 0 : timeline_id, self.tenant_shard_id
2307 : )
2308 0 : })
2309 0 : .map_err(TimelineArchivalError::Other)?;
2310 :
2311 0 : let timeline = {
2312 0 : let timelines = self.timelines.lock().unwrap();
2313 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2314 0 : warn!("timeline not available directly after attach");
2315 : // This is not a panic because no locks are held between `load_remote_timeline`
2316 : // which puts the timeline into timelines, and our look into the timeline map.
2317 0 : return Err(TimelineArchivalError::Other(anyhow::anyhow!(
2318 0 : "timeline not available directly after attach"
2319 0 : )));
2320 : };
2321 0 : let mut offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2322 0 : match offloaded_timelines.remove(&timeline_id) {
2323 0 : Some(offloaded) => {
2324 0 : offloaded.delete_from_ancestor_with_timelines(&timelines);
2325 0 : }
2326 0 : None => warn!("timeline already removed from offloaded timelines"),
2327 : }
2328 :
2329 0 : self.initialize_gc_info(&timelines, &offloaded_timelines, Some(timeline_id));
2330 :
2331 0 : Arc::clone(timeline)
2332 : };
2333 :
2334 : // Upload new list of offloaded timelines to S3
2335 0 : self.maybe_upload_tenant_manifest().await?;
2336 :
2337 : // Activate the timeline (if it makes sense)
2338 0 : if !(timeline.is_broken() || timeline.is_stopping()) {
2339 0 : let background_jobs_can_start = None;
2340 0 : timeline.activate(
2341 0 : self.clone(),
2342 0 : broker_client.clone(),
2343 0 : background_jobs_can_start,
2344 0 : &ctx.with_scope_timeline(&timeline),
2345 0 : );
2346 0 : }
2347 :
2348 0 : info!("timeline unoffloading complete");
2349 0 : Ok(timeline)
2350 0 : }
2351 :
2352 0 : pub(crate) async fn apply_timeline_archival_config(
2353 0 : self: &Arc<Self>,
2354 0 : timeline_id: TimelineId,
2355 0 : new_state: TimelineArchivalState,
2356 0 : broker_client: storage_broker::BrokerClientChannel,
2357 0 : ctx: RequestContext,
2358 0 : ) -> Result<(), TimelineArchivalError> {
2359 0 : info!("setting timeline archival config");
2360 : // First part: figure out what is needed to do, and do validation
2361 0 : let timeline_or_unarchive_offloaded = 'outer: {
2362 0 : let timelines = self.timelines.lock().unwrap();
2363 :
2364 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2365 0 : let offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2366 0 : let Some(offloaded) = offloaded_timelines.get(&timeline_id) else {
2367 0 : return Err(TimelineArchivalError::NotFound);
2368 : };
2369 0 : if new_state == TimelineArchivalState::Archived {
2370 : // It's offloaded already, so nothing to do
2371 0 : return Ok(());
2372 0 : }
2373 0 : if let Some(ancestor_timeline_id) = offloaded.ancestor_timeline_id {
2374 0 : Self::check_ancestor_of_to_be_unarchived_is_not_archived(
2375 0 : ancestor_timeline_id,
2376 0 : &timelines,
2377 0 : &offloaded_timelines,
2378 0 : )?;
2379 0 : }
2380 0 : break 'outer None;
2381 : };
2382 :
2383 : // Do some validation. We release the timelines lock below, so there is potential
2384 : // for race conditions: these checks are more present to prevent misunderstandings of
2385 : // the API's capabilities, instead of serving as the sole way to defend their invariants.
2386 0 : match new_state {
2387 : TimelineArchivalState::Unarchived => {
2388 0 : Self::check_to_be_unarchived_timeline_has_no_archived_parent(timeline)?
2389 : }
2390 : TimelineArchivalState::Archived => {
2391 0 : Self::check_to_be_archived_has_no_unarchived_children(timeline_id, &timelines)?
2392 : }
2393 : }
2394 0 : Some(Arc::clone(timeline))
2395 : };
2396 :
2397 : // Second part: unoffload timeline (if needed)
2398 0 : let timeline = if let Some(timeline) = timeline_or_unarchive_offloaded {
2399 0 : timeline
2400 : } else {
2401 : // Turn offloaded timeline into a non-offloaded one
2402 0 : self.unoffload_timeline(timeline_id, broker_client, ctx)
2403 0 : .await?
2404 : };
2405 :
2406 : // Third part: upload new timeline archival state and block until it is present in S3
2407 0 : let upload_needed = match timeline
2408 0 : .remote_client
2409 0 : .schedule_index_upload_for_timeline_archival_state(new_state)
2410 : {
2411 0 : Ok(upload_needed) => upload_needed,
2412 0 : Err(e) => {
2413 0 : if timeline.cancel.is_cancelled() {
2414 0 : return Err(TimelineArchivalError::Cancelled);
2415 : } else {
2416 0 : return Err(TimelineArchivalError::Other(e));
2417 : }
2418 : }
2419 : };
2420 :
2421 0 : if upload_needed {
2422 0 : info!("Uploading new state");
2423 : const MAX_WAIT: Duration = Duration::from_secs(10);
2424 0 : let Ok(v) =
2425 0 : tokio::time::timeout(MAX_WAIT, timeline.remote_client.wait_completion()).await
2426 : else {
2427 0 : tracing::warn!("reached timeout for waiting on upload queue");
2428 0 : return Err(TimelineArchivalError::Timeout);
2429 : };
2430 0 : v.map_err(|e| match e {
2431 0 : WaitCompletionError::NotInitialized(e) => {
2432 0 : TimelineArchivalError::Other(anyhow::anyhow!(e))
2433 : }
2434 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2435 0 : TimelineArchivalError::Cancelled
2436 : }
2437 0 : })?;
2438 0 : }
2439 0 : Ok(())
2440 0 : }
2441 :
2442 1 : pub fn get_offloaded_timeline(
2443 1 : &self,
2444 1 : timeline_id: TimelineId,
2445 1 : ) -> Result<Arc<OffloadedTimeline>, GetTimelineError> {
2446 1 : self.timelines_offloaded
2447 1 : .lock()
2448 1 : .unwrap()
2449 1 : .get(&timeline_id)
2450 1 : .map(Arc::clone)
2451 1 : .ok_or(GetTimelineError::NotFound {
2452 1 : tenant_id: self.tenant_shard_id,
2453 1 : timeline_id,
2454 1 : })
2455 1 : }
2456 :
2457 2 : pub(crate) fn tenant_shard_id(&self) -> TenantShardId {
2458 2 : self.tenant_shard_id
2459 2 : }
2460 :
2461 : /// Get Timeline handle for given Neon timeline ID.
2462 : /// This function is idempotent. It doesn't change internal state in any way.
2463 111 : pub fn get_timeline(
2464 111 : &self,
2465 111 : timeline_id: TimelineId,
2466 111 : active_only: bool,
2467 111 : ) -> Result<Arc<Timeline>, GetTimelineError> {
2468 111 : let timelines_accessor = self.timelines.lock().unwrap();
2469 111 : let timeline = timelines_accessor
2470 111 : .get(&timeline_id)
2471 111 : .ok_or(GetTimelineError::NotFound {
2472 111 : tenant_id: self.tenant_shard_id,
2473 111 : timeline_id,
2474 111 : })?;
2475 :
2476 110 : if active_only && !timeline.is_active() {
2477 0 : Err(GetTimelineError::NotActive {
2478 0 : tenant_id: self.tenant_shard_id,
2479 0 : timeline_id,
2480 0 : state: timeline.current_state(),
2481 0 : })
2482 : } else {
2483 110 : Ok(Arc::clone(timeline))
2484 : }
2485 111 : }
2486 :
2487 : /// Lists timelines the tenant contains.
2488 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2489 3 : pub fn list_timelines(&self) -> Vec<Arc<Timeline>> {
2490 3 : self.timelines
2491 3 : .lock()
2492 3 : .unwrap()
2493 3 : .values()
2494 3 : .map(Arc::clone)
2495 3 : .collect()
2496 3 : }
2497 :
2498 : /// Lists timelines the tenant contains.
2499 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2500 0 : pub fn list_importing_timelines(&self) -> Vec<Arc<ImportingTimeline>> {
2501 0 : self.timelines_importing
2502 0 : .lock()
2503 0 : .unwrap()
2504 0 : .values()
2505 0 : .map(Arc::clone)
2506 0 : .collect()
2507 0 : }
2508 :
2509 : /// Lists timelines the tenant manages, including offloaded ones.
2510 : ///
2511 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2512 0 : pub fn list_timelines_and_offloaded(
2513 0 : &self,
2514 0 : ) -> (Vec<Arc<Timeline>>, Vec<Arc<OffloadedTimeline>>) {
2515 0 : let timelines = self
2516 0 : .timelines
2517 0 : .lock()
2518 0 : .unwrap()
2519 0 : .values()
2520 0 : .map(Arc::clone)
2521 0 : .collect();
2522 0 : let offloaded = self
2523 0 : .timelines_offloaded
2524 0 : .lock()
2525 0 : .unwrap()
2526 0 : .values()
2527 0 : .map(Arc::clone)
2528 0 : .collect();
2529 0 : (timelines, offloaded)
2530 0 : }
2531 :
2532 0 : pub fn list_timeline_ids(&self) -> Vec<TimelineId> {
2533 0 : self.timelines.lock().unwrap().keys().cloned().collect()
2534 0 : }
2535 :
2536 : /// This is used by tests & import-from-basebackup.
2537 : ///
2538 : /// The returned [`UninitializedTimeline`] contains no data nor metadata and it is in
2539 : /// a state that will fail [`TenantShard::load_remote_timeline`] because `disk_consistent_lsn=Lsn(0)`.
2540 : ///
2541 : /// The caller is responsible for getting the timeline into a state that will be accepted
2542 : /// by [`TenantShard::load_remote_timeline`] / [`TenantShard::attach`].
2543 : /// Then they may call [`UninitializedTimeline::finish_creation`] to add the timeline
2544 : /// to the [`TenantShard::timelines`].
2545 : ///
2546 : /// Tests should use `TenantShard::create_test_timeline` to set up the minimum required metadata keys.
2547 115 : pub(crate) async fn create_empty_timeline(
2548 115 : self: &Arc<Self>,
2549 115 : new_timeline_id: TimelineId,
2550 115 : initdb_lsn: Lsn,
2551 115 : pg_version: PgMajorVersion,
2552 115 : ctx: &RequestContext,
2553 115 : ) -> anyhow::Result<(UninitializedTimeline, RequestContext)> {
2554 115 : anyhow::ensure!(
2555 115 : self.is_active(),
2556 0 : "Cannot create empty timelines on inactive tenant"
2557 : );
2558 :
2559 : // Protect against concurrent attempts to use this TimelineId
2560 115 : let create_guard = match self
2561 115 : .start_creating_timeline(new_timeline_id, CreateTimelineIdempotency::FailWithConflict)
2562 115 : .await?
2563 : {
2564 114 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2565 : StartCreatingTimelineResult::Idempotent(_) => {
2566 0 : unreachable!("FailWithConflict implies we get an error instead")
2567 : }
2568 : };
2569 :
2570 114 : let new_metadata = TimelineMetadata::new(
2571 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2572 : // make it valid, before calling finish_creation()
2573 114 : Lsn(0),
2574 114 : None,
2575 114 : None,
2576 114 : Lsn(0),
2577 114 : initdb_lsn,
2578 114 : initdb_lsn,
2579 114 : pg_version,
2580 : );
2581 114 : self.prepare_new_timeline(
2582 114 : new_timeline_id,
2583 114 : &new_metadata,
2584 114 : create_guard,
2585 114 : initdb_lsn,
2586 114 : None,
2587 114 : None,
2588 114 : None,
2589 114 : ctx,
2590 114 : )
2591 114 : .await
2592 115 : }
2593 :
2594 : /// Helper for unit tests to create an empty timeline.
2595 : ///
2596 : /// The timeline is has state value `Active` but its background loops are not running.
2597 : // This makes the various functions which anyhow::ensure! for Active state work in tests.
2598 : // Our current tests don't need the background loops.
2599 : #[cfg(test)]
2600 110 : pub async fn create_test_timeline(
2601 110 : self: &Arc<Self>,
2602 110 : new_timeline_id: TimelineId,
2603 110 : initdb_lsn: Lsn,
2604 110 : pg_version: PgMajorVersion,
2605 110 : ctx: &RequestContext,
2606 110 : ) -> anyhow::Result<Arc<Timeline>> {
2607 110 : let (uninit_tl, ctx) = self
2608 110 : .create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2609 110 : .await?;
2610 110 : let tline = uninit_tl.raw_timeline().expect("we just created it");
2611 110 : assert_eq!(tline.get_last_record_lsn(), Lsn(0));
2612 :
2613 : // Setup minimum keys required for the timeline to be usable.
2614 110 : let mut modification = tline.begin_modification(initdb_lsn);
2615 110 : modification
2616 110 : .init_empty_test_timeline()
2617 110 : .context("init_empty_test_timeline")?;
2618 110 : modification
2619 110 : .commit(&ctx)
2620 110 : .await
2621 110 : .context("commit init_empty_test_timeline modification")?;
2622 :
2623 : // Flush to disk so that uninit_tl's check for valid disk_consistent_lsn passes.
2624 110 : tline.maybe_spawn_flush_loop();
2625 110 : tline.freeze_and_flush().await.context("freeze_and_flush")?;
2626 :
2627 : // Make sure the freeze_and_flush reaches remote storage.
2628 110 : tline.remote_client.wait_completion().await.unwrap();
2629 :
2630 110 : let tl = uninit_tl.finish_creation().await?;
2631 : // The non-test code would call tl.activate() here.
2632 110 : tl.set_state(TimelineState::Active);
2633 110 : Ok(tl)
2634 110 : }
2635 :
2636 : /// Helper for unit tests to create a timeline with some pre-loaded states.
2637 : #[cfg(test)]
2638 : #[allow(clippy::too_many_arguments)]
2639 24 : pub async fn create_test_timeline_with_layers(
2640 24 : self: &Arc<Self>,
2641 24 : new_timeline_id: TimelineId,
2642 24 : initdb_lsn: Lsn,
2643 24 : pg_version: PgMajorVersion,
2644 24 : ctx: &RequestContext,
2645 24 : in_memory_layer_desc: Vec<timeline::InMemoryLayerTestDesc>,
2646 24 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
2647 24 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
2648 24 : end_lsn: Lsn,
2649 24 : ) -> anyhow::Result<Arc<Timeline>> {
2650 : use checks::check_valid_layermap;
2651 : use itertools::Itertools;
2652 :
2653 24 : let tline = self
2654 24 : .create_test_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2655 24 : .await?;
2656 24 : tline.force_advance_lsn(end_lsn);
2657 71 : for deltas in delta_layer_desc {
2658 47 : tline
2659 47 : .force_create_delta_layer(deltas, Some(initdb_lsn), ctx)
2660 47 : .await?;
2661 : }
2662 58 : for (lsn, images) in image_layer_desc {
2663 34 : tline
2664 34 : .force_create_image_layer(lsn, images, Some(initdb_lsn), ctx)
2665 34 : .await?;
2666 : }
2667 28 : for in_memory in in_memory_layer_desc {
2668 4 : tline
2669 4 : .force_create_in_memory_layer(in_memory, Some(initdb_lsn), ctx)
2670 4 : .await?;
2671 : }
2672 24 : let layer_names = tline
2673 24 : .layers
2674 24 : .read(LayerManagerLockHolder::Testing)
2675 24 : .await
2676 24 : .layer_map()
2677 24 : .unwrap()
2678 24 : .iter_historic_layers()
2679 105 : .map(|layer| layer.layer_name())
2680 24 : .collect_vec();
2681 24 : if let Some(err) = check_valid_layermap(&layer_names) {
2682 0 : bail!("invalid layermap: {err}");
2683 24 : }
2684 24 : Ok(tline)
2685 24 : }
2686 :
2687 : /// Create a new timeline.
2688 : ///
2689 : /// Returns the new timeline ID and reference to its Timeline object.
2690 : ///
2691 : /// If the caller specified the timeline ID to use (`new_timeline_id`), and timeline with
2692 : /// the same timeline ID already exists, returns CreateTimelineError::AlreadyExists.
2693 : #[allow(clippy::too_many_arguments)]
2694 0 : pub(crate) async fn create_timeline(
2695 0 : self: &Arc<TenantShard>,
2696 0 : params: CreateTimelineParams,
2697 0 : broker_client: storage_broker::BrokerClientChannel,
2698 0 : ctx: &RequestContext,
2699 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
2700 0 : if !self.is_active() {
2701 0 : if matches!(self.current_state(), TenantState::Stopping { .. }) {
2702 0 : return Err(CreateTimelineError::ShuttingDown);
2703 : } else {
2704 0 : return Err(CreateTimelineError::Other(anyhow::anyhow!(
2705 0 : "Cannot create timelines on inactive tenant"
2706 0 : )));
2707 : }
2708 0 : }
2709 :
2710 0 : let _gate = self
2711 0 : .gate
2712 0 : .enter()
2713 0 : .map_err(|_| CreateTimelineError::ShuttingDown)?;
2714 :
2715 0 : let result: CreateTimelineResult = match params {
2716 : CreateTimelineParams::Bootstrap(CreateTimelineParamsBootstrap {
2717 0 : new_timeline_id,
2718 0 : existing_initdb_timeline_id,
2719 0 : pg_version,
2720 : }) => {
2721 0 : self.bootstrap_timeline(
2722 0 : new_timeline_id,
2723 0 : pg_version,
2724 0 : existing_initdb_timeline_id,
2725 0 : ctx,
2726 0 : )
2727 0 : .await?
2728 : }
2729 : CreateTimelineParams::Branch(CreateTimelineParamsBranch {
2730 0 : new_timeline_id,
2731 0 : ancestor_timeline_id,
2732 0 : mut ancestor_start_lsn,
2733 : }) => {
2734 0 : let ancestor_timeline = self
2735 0 : .get_timeline(ancestor_timeline_id, false)
2736 0 : .context("Cannot branch off the timeline that's not present in pageserver")?;
2737 :
2738 : // instead of waiting around, just deny the request because ancestor is not yet
2739 : // ready for other purposes either.
2740 0 : if !ancestor_timeline.is_active() {
2741 0 : return Err(CreateTimelineError::AncestorNotActive);
2742 0 : }
2743 :
2744 0 : if ancestor_timeline.is_archived() == Some(true) {
2745 0 : info!("tried to branch archived timeline");
2746 0 : return Err(CreateTimelineError::AncestorArchived);
2747 0 : }
2748 :
2749 0 : if let Some(lsn) = ancestor_start_lsn.as_mut() {
2750 0 : *lsn = lsn.align();
2751 :
2752 0 : let ancestor_ancestor_lsn = ancestor_timeline.get_ancestor_lsn();
2753 0 : if ancestor_ancestor_lsn > *lsn {
2754 : // can we safely just branch from the ancestor instead?
2755 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
2756 0 : "invalid start lsn {} for ancestor timeline {}: less than timeline ancestor lsn {}",
2757 0 : lsn,
2758 0 : ancestor_timeline_id,
2759 0 : ancestor_ancestor_lsn,
2760 0 : )));
2761 0 : }
2762 :
2763 : // Wait for the WAL to arrive and be processed on the parent branch up
2764 : // to the requested branch point. The repository code itself doesn't
2765 : // require it, but if we start to receive WAL on the new timeline,
2766 : // decoding the new WAL might need to look up previous pages, relation
2767 : // sizes etc. and that would get confused if the previous page versions
2768 : // are not in the repository yet.
2769 0 : ancestor_timeline
2770 0 : .wait_lsn(
2771 0 : *lsn,
2772 0 : timeline::WaitLsnWaiter::Tenant,
2773 0 : timeline::WaitLsnTimeout::Default,
2774 0 : ctx,
2775 0 : )
2776 0 : .await
2777 0 : .map_err(|e| match e {
2778 0 : e @ (WaitLsnError::Timeout(_) | WaitLsnError::BadState { .. }) => {
2779 0 : CreateTimelineError::AncestorLsn(anyhow::anyhow!(e))
2780 : }
2781 0 : WaitLsnError::Shutdown => CreateTimelineError::ShuttingDown,
2782 0 : })?;
2783 0 : }
2784 :
2785 0 : self.branch_timeline(&ancestor_timeline, new_timeline_id, ancestor_start_lsn, ctx)
2786 0 : .await?
2787 : }
2788 0 : CreateTimelineParams::ImportPgdata(params) => {
2789 0 : self.create_timeline_import_pgdata(params, ctx).await?
2790 : }
2791 : };
2792 :
2793 : // At this point we have dropped our guard on [`Self::timelines_creating`], and
2794 : // the timeline is visible in [`Self::timelines`], but it is _not_ durable yet. We must
2795 : // not send a success to the caller until it is. The same applies to idempotent retries.
2796 : //
2797 : // TODO: the timeline is already visible in [`Self::timelines`]; a caller could incorrectly
2798 : // assume that, because they can see the timeline via API, that the creation is done and
2799 : // that it is durable. Ideally, we would keep the timeline hidden (in [`Self::timelines_creating`])
2800 : // until it is durable, e.g., by extending the time we hold the creation guard. This also
2801 : // interacts with UninitializedTimeline and is generally a bit tricky.
2802 : //
2803 : // To re-emphasize: the only correct way to create a timeline is to repeat calling the
2804 : // creation API until it returns success. Only then is durability guaranteed.
2805 0 : info!(creation_result=%result.discriminant(), "waiting for timeline to be durable");
2806 0 : result
2807 0 : .timeline()
2808 0 : .remote_client
2809 0 : .wait_completion()
2810 0 : .await
2811 0 : .map_err(|e| match e {
2812 : WaitCompletionError::NotInitialized(
2813 0 : e, // If the queue is already stopped, it's a shutdown error.
2814 0 : ) if e.is_stopping() => CreateTimelineError::ShuttingDown,
2815 : WaitCompletionError::NotInitialized(_) => {
2816 : // This is a bug: we should never try to wait for uploads before initializing the timeline
2817 0 : debug_assert!(false);
2818 0 : CreateTimelineError::Other(anyhow::anyhow!("timeline not initialized"))
2819 : }
2820 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2821 0 : CreateTimelineError::ShuttingDown
2822 : }
2823 0 : })?;
2824 :
2825 : // The creating task is responsible for activating the timeline.
2826 : // We do this after `wait_completion()` so that we don't spin up tasks that start
2827 : // doing stuff before the IndexPart is durable in S3, which is done by the previous section.
2828 0 : let activated_timeline = match result {
2829 0 : CreateTimelineResult::Created(timeline) => {
2830 0 : timeline.activate(
2831 0 : self.clone(),
2832 0 : broker_client,
2833 0 : None,
2834 0 : &ctx.with_scope_timeline(&timeline),
2835 : );
2836 0 : timeline
2837 : }
2838 0 : CreateTimelineResult::Idempotent(timeline) => {
2839 0 : info!(
2840 0 : "request was deemed idempotent, activation will be done by the creating task"
2841 : );
2842 0 : timeline
2843 : }
2844 0 : CreateTimelineResult::ImportSpawned(timeline) => {
2845 0 : info!(
2846 0 : "import task spawned, timeline will become visible and activated once the import is done"
2847 : );
2848 0 : timeline
2849 : }
2850 : };
2851 :
2852 0 : Ok(activated_timeline)
2853 0 : }
2854 :
2855 : /// The returned [`Arc<Timeline>`] is NOT in the [`TenantShard::timelines`] map until the import
2856 : /// completes in the background. A DIFFERENT [`Arc<Timeline>`] will be inserted into the
2857 : /// [`TenantShard::timelines`] map when the import completes.
2858 : /// We only return an [`Arc<Timeline>`] here so the API handler can create a [`pageserver_api::models::TimelineInfo`]
2859 : /// for the response.
2860 0 : async fn create_timeline_import_pgdata(
2861 0 : self: &Arc<Self>,
2862 0 : params: CreateTimelineParamsImportPgdata,
2863 0 : ctx: &RequestContext,
2864 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
2865 : let CreateTimelineParamsImportPgdata {
2866 0 : new_timeline_id,
2867 0 : location,
2868 0 : idempotency_key,
2869 0 : } = params;
2870 :
2871 0 : let started_at = chrono::Utc::now().naive_utc();
2872 :
2873 : //
2874 : // There's probably a simpler way to upload an index part, but, remote_timeline_client
2875 : // is the canonical way we do it.
2876 : // - create an empty timeline in-memory
2877 : // - use its remote_timeline_client to do the upload
2878 : // - dispose of the uninit timeline
2879 : // - keep the creation guard alive
2880 :
2881 0 : let timeline_create_guard = match self
2882 0 : .start_creating_timeline(
2883 0 : new_timeline_id,
2884 0 : CreateTimelineIdempotency::ImportPgdata(CreatingTimelineIdempotencyImportPgdata {
2885 0 : idempotency_key: idempotency_key.clone(),
2886 0 : }),
2887 0 : )
2888 0 : .await?
2889 : {
2890 0 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2891 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
2892 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
2893 : }
2894 : };
2895 :
2896 0 : let (mut uninit_timeline, timeline_ctx) = {
2897 0 : let this = &self;
2898 0 : let initdb_lsn = Lsn(0);
2899 0 : async move {
2900 0 : let new_metadata = TimelineMetadata::new(
2901 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2902 : // make it valid, before calling finish_creation()
2903 0 : Lsn(0),
2904 0 : None,
2905 0 : None,
2906 0 : Lsn(0),
2907 0 : initdb_lsn,
2908 0 : initdb_lsn,
2909 0 : PgMajorVersion::PG15,
2910 : );
2911 0 : this.prepare_new_timeline(
2912 0 : new_timeline_id,
2913 0 : &new_metadata,
2914 0 : timeline_create_guard,
2915 0 : initdb_lsn,
2916 0 : None,
2917 0 : None,
2918 0 : None,
2919 0 : ctx,
2920 0 : )
2921 0 : .await
2922 0 : }
2923 : }
2924 0 : .await?;
2925 :
2926 0 : let in_progress = import_pgdata::index_part_format::InProgress {
2927 0 : idempotency_key,
2928 0 : location,
2929 0 : started_at,
2930 0 : };
2931 0 : let index_part = import_pgdata::index_part_format::Root::V1(
2932 0 : import_pgdata::index_part_format::V1::InProgress(in_progress),
2933 0 : );
2934 0 : uninit_timeline
2935 0 : .raw_timeline()
2936 0 : .unwrap()
2937 0 : .remote_client
2938 0 : .schedule_index_upload_for_import_pgdata_state_update(Some(index_part.clone()))?;
2939 :
2940 : // wait_completion happens in caller
2941 :
2942 0 : let (timeline, timeline_create_guard) = uninit_timeline.finish_creation_myself();
2943 :
2944 0 : let import_task_gate = Gate::default();
2945 0 : let import_task_guard = import_task_gate.enter().unwrap();
2946 :
2947 0 : let import_task_handle = tokio::spawn(self.clone().create_timeline_import_pgdata_task(
2948 0 : timeline.clone(),
2949 0 : index_part,
2950 0 : timeline_create_guard,
2951 0 : import_task_guard,
2952 0 : timeline_ctx.detached_child(TaskKind::ImportPgdata, DownloadBehavior::Warn),
2953 : ));
2954 :
2955 0 : let prev = self.timelines_importing.lock().unwrap().insert(
2956 0 : timeline.timeline_id,
2957 0 : Arc::new(ImportingTimeline {
2958 0 : timeline: timeline.clone(),
2959 0 : import_task_handle,
2960 0 : import_task_gate,
2961 0 : delete_progress: TimelineDeleteProgress::default(),
2962 0 : }),
2963 0 : );
2964 :
2965 : // Idempotency is enforced higher up the stack
2966 0 : assert!(prev.is_none());
2967 :
2968 : // NB: the timeline doesn't exist in self.timelines at this point
2969 0 : Ok(CreateTimelineResult::ImportSpawned(timeline))
2970 0 : }
2971 :
2972 : /// Finalize the import of a timeline on this shard by marking it complete in
2973 : /// the index part. If the import task hasn't finished yet, returns an error.
2974 : ///
2975 : /// This method is idempotent. If the import was finalized once, the next call
2976 : /// will be a no-op.
2977 0 : pub(crate) async fn finalize_importing_timeline(
2978 0 : &self,
2979 0 : timeline_id: TimelineId,
2980 0 : ) -> Result<(), FinalizeTimelineImportError> {
2981 0 : let timeline = {
2982 0 : let locked = self.timelines_importing.lock().unwrap();
2983 0 : match locked.get(&timeline_id) {
2984 0 : Some(importing_timeline) => {
2985 0 : if !importing_timeline.import_task_handle.is_finished() {
2986 0 : return Err(FinalizeTimelineImportError::ImportTaskStillRunning);
2987 0 : }
2988 :
2989 0 : importing_timeline.timeline.clone()
2990 : }
2991 : None => {
2992 0 : return Ok(());
2993 : }
2994 : }
2995 : };
2996 :
2997 0 : timeline
2998 0 : .remote_client
2999 0 : .schedule_index_upload_for_import_pgdata_finalize()
3000 0 : .map_err(|_err| FinalizeTimelineImportError::ShuttingDown)?;
3001 0 : timeline
3002 0 : .remote_client
3003 0 : .wait_completion()
3004 0 : .await
3005 0 : .map_err(|_err| FinalizeTimelineImportError::ShuttingDown)?;
3006 :
3007 0 : self.timelines_importing
3008 0 : .lock()
3009 0 : .unwrap()
3010 0 : .remove(&timeline_id);
3011 :
3012 0 : Ok(())
3013 0 : }
3014 :
3015 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), timeline_id=%timeline.timeline_id))]
3016 : async fn create_timeline_import_pgdata_task(
3017 : self: Arc<TenantShard>,
3018 : timeline: Arc<Timeline>,
3019 : index_part: import_pgdata::index_part_format::Root,
3020 : timeline_create_guard: TimelineCreateGuard,
3021 : _import_task_guard: GateGuard,
3022 : ctx: RequestContext,
3023 : ) {
3024 : debug_assert_current_span_has_tenant_and_timeline_id();
3025 : info!("starting");
3026 : scopeguard::defer! {info!("exiting")};
3027 :
3028 : let res = self
3029 : .create_timeline_import_pgdata_task_impl(
3030 : timeline,
3031 : index_part,
3032 : timeline_create_guard,
3033 : ctx,
3034 : )
3035 : .await;
3036 : if let Err(err) = &res {
3037 : error!(?err, "task failed");
3038 : // TODO sleep & retry, sensitive to tenant shutdown
3039 : // TODO: allow timeline deletion requests => should cancel the task
3040 : }
3041 : }
3042 :
3043 0 : async fn create_timeline_import_pgdata_task_impl(
3044 0 : self: Arc<TenantShard>,
3045 0 : timeline: Arc<Timeline>,
3046 0 : index_part: import_pgdata::index_part_format::Root,
3047 0 : _timeline_create_guard: TimelineCreateGuard,
3048 0 : ctx: RequestContext,
3049 0 : ) -> Result<(), anyhow::Error> {
3050 0 : info!("importing pgdata");
3051 0 : let ctx = ctx.with_scope_timeline(&timeline);
3052 0 : import_pgdata::doit(&timeline, index_part, &ctx, self.cancel.clone())
3053 0 : .await
3054 0 : .context("import")?;
3055 0 : info!("import done - waiting for activation");
3056 :
3057 0 : anyhow::Ok(())
3058 0 : }
3059 :
3060 0 : pub(crate) async fn delete_timeline(
3061 0 : self: Arc<Self>,
3062 0 : timeline_id: TimelineId,
3063 0 : ) -> Result<(), DeleteTimelineError> {
3064 0 : DeleteTimelineFlow::run(&self, timeline_id).await?;
3065 :
3066 0 : Ok(())
3067 0 : }
3068 :
3069 : /// perform one garbage collection iteration, removing old data files from disk.
3070 : /// this function is periodically called by gc task.
3071 : /// also it can be explicitly requested through page server api 'do_gc' command.
3072 : ///
3073 : /// `target_timeline_id` specifies the timeline to GC, or None for all.
3074 : ///
3075 : /// The `horizon` an `pitr` parameters determine how much WAL history needs to be retained.
3076 : /// Also known as the retention period, or the GC cutoff point. `horizon` specifies
3077 : /// the amount of history, as LSN difference from current latest LSN on each timeline.
3078 : /// `pitr` specifies the same as a time difference from the current time. The effective
3079 : /// GC cutoff point is determined conservatively by either `horizon` and `pitr`, whichever
3080 : /// requires more history to be retained.
3081 : //
3082 377 : pub(crate) async fn gc_iteration(
3083 377 : &self,
3084 377 : target_timeline_id: Option<TimelineId>,
3085 377 : horizon: u64,
3086 377 : pitr: Duration,
3087 377 : cancel: &CancellationToken,
3088 377 : ctx: &RequestContext,
3089 377 : ) -> Result<GcResult, GcError> {
3090 : // Don't start doing work during shutdown
3091 377 : if let TenantState::Stopping { .. } = self.current_state() {
3092 0 : return Ok(GcResult::default());
3093 377 : }
3094 :
3095 : // there is a global allowed_error for this
3096 377 : if !self.is_active() {
3097 0 : return Err(GcError::NotActive);
3098 377 : }
3099 :
3100 : {
3101 377 : let conf = self.tenant_conf.load();
3102 :
3103 : // If we may not delete layers, then simply skip GC. Even though a tenant
3104 : // in AttachedMulti state could do GC and just enqueue the blocked deletions,
3105 : // the only advantage to doing it is to perhaps shrink the LayerMap metadata
3106 : // a bit sooner than we would achieve by waiting for AttachedSingle status.
3107 377 : if !conf.location.may_delete_layers_hint() {
3108 0 : info!("Skipping GC in location state {:?}", conf.location);
3109 0 : return Ok(GcResult::default());
3110 377 : }
3111 :
3112 377 : if conf.is_gc_blocked_by_lsn_lease_deadline() {
3113 0 : info!("Skipping GC because lsn lease deadline is not reached");
3114 0 : return Ok(GcResult::default());
3115 377 : }
3116 : }
3117 :
3118 377 : let _guard = match self.gc_block.start().await {
3119 377 : Ok(guard) => guard,
3120 0 : Err(reasons) => {
3121 0 : info!("Skipping GC: {reasons}");
3122 0 : return Ok(GcResult::default());
3123 : }
3124 : };
3125 :
3126 377 : self.gc_iteration_internal(target_timeline_id, horizon, pitr, cancel, ctx)
3127 377 : .await
3128 377 : }
3129 :
3130 : /// Performs one compaction iteration. Called periodically from the compaction loop. Returns
3131 : /// whether another compaction is needed, if we still have pending work or if we yield for
3132 : /// immediate L0 compaction.
3133 : ///
3134 : /// Compaction can also be explicitly requested for a timeline via the HTTP API.
3135 0 : async fn compaction_iteration(
3136 0 : self: &Arc<Self>,
3137 0 : cancel: &CancellationToken,
3138 0 : ctx: &RequestContext,
3139 0 : ) -> Result<CompactionOutcome, CompactionError> {
3140 : // Don't compact inactive tenants.
3141 0 : if !self.is_active() {
3142 0 : return Ok(CompactionOutcome::Skipped);
3143 0 : }
3144 :
3145 : // Don't compact tenants that can't upload layers. We don't check `may_delete_layers_hint`,
3146 : // since we need to compact L0 even in AttachedMulti to bound read amplification.
3147 0 : let location = self.tenant_conf.load().location;
3148 0 : if !location.may_upload_layers_hint() {
3149 0 : info!("skipping compaction in location state {location:?}");
3150 0 : return Ok(CompactionOutcome::Skipped);
3151 0 : }
3152 :
3153 : // Don't compact if the circuit breaker is tripped.
3154 0 : if self.compaction_circuit_breaker.lock().unwrap().is_broken() {
3155 0 : info!("skipping compaction due to previous failures");
3156 0 : return Ok(CompactionOutcome::Skipped);
3157 0 : }
3158 :
3159 : // Collect all timelines to compact, along with offload instructions and L0 counts.
3160 0 : let mut compact: Vec<Arc<Timeline>> = Vec::new();
3161 0 : let mut offload: HashSet<TimelineId> = HashSet::new();
3162 0 : let mut l0_counts: HashMap<TimelineId, usize> = HashMap::new();
3163 :
3164 : {
3165 0 : let offload_enabled = self.get_timeline_offloading_enabled();
3166 0 : let timelines = self.timelines.lock().unwrap();
3167 0 : for (&timeline_id, timeline) in timelines.iter() {
3168 : // Skip inactive timelines.
3169 0 : if !timeline.is_active() {
3170 0 : continue;
3171 0 : }
3172 :
3173 : // Schedule the timeline for compaction.
3174 0 : compact.push(timeline.clone());
3175 :
3176 : // Schedule the timeline for offloading if eligible.
3177 0 : let can_offload = offload_enabled
3178 0 : && timeline.can_offload().0
3179 0 : && !timelines
3180 0 : .iter()
3181 0 : .any(|(_, tli)| tli.get_ancestor_timeline_id() == Some(timeline_id));
3182 0 : if can_offload {
3183 0 : offload.insert(timeline_id);
3184 0 : }
3185 : }
3186 : } // release timelines lock
3187 :
3188 0 : for timeline in &compact {
3189 : // Collect L0 counts. Can't await while holding lock above.
3190 0 : if let Ok(lm) = timeline
3191 0 : .layers
3192 0 : .read(LayerManagerLockHolder::Compaction)
3193 0 : .await
3194 0 : .layer_map()
3195 0 : {
3196 0 : l0_counts.insert(timeline.timeline_id, lm.level0_deltas().len());
3197 0 : }
3198 : }
3199 :
3200 : // Pass 1: L0 compaction across all timelines, in order of L0 count. We prioritize this to
3201 : // bound read amplification.
3202 : //
3203 : // TODO: this may spin on one or more ingest-heavy timelines, starving out image/GC
3204 : // compaction and offloading. We leave that as a potential problem to solve later. Consider
3205 : // splitting L0 and image/GC compaction to separate background jobs.
3206 0 : if self.get_compaction_l0_first() {
3207 0 : let compaction_threshold = self.get_compaction_threshold();
3208 0 : let compact_l0 = compact
3209 0 : .iter()
3210 0 : .map(|tli| (tli, l0_counts.get(&tli.timeline_id).copied().unwrap_or(0)))
3211 0 : .filter(|&(_, l0)| l0 >= compaction_threshold)
3212 0 : .sorted_by_key(|&(_, l0)| l0)
3213 0 : .rev()
3214 0 : .map(|(tli, _)| tli.clone())
3215 0 : .collect_vec();
3216 :
3217 0 : let mut has_pending_l0 = false;
3218 0 : for timeline in compact_l0 {
3219 0 : let ctx = &ctx.with_scope_timeline(&timeline);
3220 : // NB: don't set CompactFlags::YieldForL0, since this is an L0-only compaction pass.
3221 0 : let outcome = timeline
3222 0 : .compact(cancel, CompactFlags::OnlyL0Compaction.into(), ctx)
3223 0 : .instrument(info_span!("compact_timeline", timeline_id = %timeline.timeline_id))
3224 0 : .await
3225 0 : .inspect_err(|err| self.maybe_trip_compaction_breaker(err))?;
3226 0 : match outcome {
3227 0 : CompactionOutcome::Done => {}
3228 0 : CompactionOutcome::Skipped => {}
3229 0 : CompactionOutcome::Pending => has_pending_l0 = true,
3230 0 : CompactionOutcome::YieldForL0 => has_pending_l0 = true,
3231 : }
3232 : }
3233 0 : if has_pending_l0 {
3234 0 : return Ok(CompactionOutcome::YieldForL0); // do another pass
3235 0 : }
3236 0 : }
3237 :
3238 : // Pass 2: image compaction and timeline offloading. If any timelines have accumulated more
3239 : // L0 layers, they may also be compacted here. Image compaction will yield if there is
3240 : // pending L0 compaction on any tenant timeline.
3241 : //
3242 : // TODO: consider ordering timelines by some priority, e.g. time since last full compaction,
3243 : // amount of L1 delta debt or garbage, offload-eligible timelines first, etc.
3244 0 : let mut has_pending = false;
3245 0 : for timeline in compact {
3246 0 : if !timeline.is_active() {
3247 0 : continue;
3248 0 : }
3249 0 : let ctx = &ctx.with_scope_timeline(&timeline);
3250 :
3251 : // Yield for L0 if the separate L0 pass is enabled (otherwise there's no point).
3252 0 : let mut flags = EnumSet::default();
3253 0 : if self.get_compaction_l0_first() {
3254 0 : flags |= CompactFlags::YieldForL0;
3255 0 : }
3256 :
3257 0 : let mut outcome = timeline
3258 0 : .compact(cancel, flags, ctx)
3259 0 : .instrument(info_span!("compact_timeline", timeline_id = %timeline.timeline_id))
3260 0 : .await
3261 0 : .inspect_err(|err| self.maybe_trip_compaction_breaker(err))?;
3262 :
3263 : // If we're done compacting, check the scheduled GC compaction queue for more work.
3264 0 : if outcome == CompactionOutcome::Done {
3265 0 : let queue = {
3266 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3267 0 : guard
3268 0 : .entry(timeline.timeline_id)
3269 0 : .or_insert_with(|| Arc::new(GcCompactionQueue::new()))
3270 0 : .clone()
3271 : };
3272 0 : let gc_compaction_strategy = self
3273 0 : .feature_resolver
3274 0 : .evaluate_multivariate("gc-comapction-strategy")
3275 0 : .ok();
3276 0 : let span = if let Some(gc_compaction_strategy) = gc_compaction_strategy {
3277 0 : info_span!("gc_compact_timeline", timeline_id = %timeline.timeline_id, strategy = %gc_compaction_strategy)
3278 : } else {
3279 0 : info_span!("gc_compact_timeline", timeline_id = %timeline.timeline_id)
3280 : };
3281 0 : outcome = queue
3282 0 : .iteration(cancel, ctx, &self.gc_block, &timeline)
3283 0 : .instrument(span)
3284 0 : .await?;
3285 0 : }
3286 :
3287 : // If we're done compacting, offload the timeline if requested.
3288 0 : if outcome == CompactionOutcome::Done && offload.contains(&timeline.timeline_id) {
3289 0 : pausable_failpoint!("before-timeline-auto-offload");
3290 0 : offload_timeline(self, &timeline)
3291 0 : .instrument(info_span!("offload_timeline", timeline_id = %timeline.timeline_id))
3292 0 : .await
3293 0 : .or_else(|err| match err {
3294 : // Ignore this, we likely raced with unarchival.
3295 0 : OffloadError::NotArchived => Ok(()),
3296 0 : OffloadError::AlreadyInProgress => Ok(()),
3297 0 : OffloadError::Cancelled => Err(CompactionError::new_cancelled()),
3298 : // don't break the anyhow chain
3299 0 : OffloadError::Other(err) => Err(CompactionError::Other(err)),
3300 0 : })?;
3301 0 : }
3302 :
3303 0 : match outcome {
3304 0 : CompactionOutcome::Done => {}
3305 0 : CompactionOutcome::Skipped => {}
3306 0 : CompactionOutcome::Pending => has_pending = true,
3307 : // This mostly makes sense when the L0-only pass above is enabled, since there's
3308 : // otherwise no guarantee that we'll start with the timeline that has high L0.
3309 0 : CompactionOutcome::YieldForL0 => return Ok(CompactionOutcome::YieldForL0),
3310 : }
3311 : }
3312 :
3313 : // Success! Untrip the breaker if necessary.
3314 0 : self.compaction_circuit_breaker
3315 0 : .lock()
3316 0 : .unwrap()
3317 0 : .success(&CIRCUIT_BREAKERS_UNBROKEN);
3318 :
3319 0 : match has_pending {
3320 0 : true => Ok(CompactionOutcome::Pending),
3321 0 : false => Ok(CompactionOutcome::Done),
3322 : }
3323 0 : }
3324 :
3325 : /// Trips the compaction circuit breaker if appropriate.
3326 0 : pub(crate) fn maybe_trip_compaction_breaker(&self, err: &CompactionError) {
3327 0 : if err.is_cancel() {
3328 0 : return;
3329 0 : }
3330 0 : self.compaction_circuit_breaker
3331 0 : .lock()
3332 0 : .unwrap()
3333 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, err);
3334 0 : }
3335 :
3336 : /// Cancel scheduled compaction tasks
3337 0 : pub(crate) fn cancel_scheduled_compaction(&self, timeline_id: TimelineId) {
3338 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3339 0 : if let Some(q) = guard.get_mut(&timeline_id) {
3340 0 : q.cancel_scheduled();
3341 0 : }
3342 0 : }
3343 :
3344 0 : pub(crate) fn get_scheduled_compaction_tasks(
3345 0 : &self,
3346 0 : timeline_id: TimelineId,
3347 0 : ) -> Vec<CompactInfoResponse> {
3348 0 : let res = {
3349 0 : let guard = self.scheduled_compaction_tasks.lock().unwrap();
3350 0 : guard.get(&timeline_id).map(|q| q.remaining_jobs())
3351 : };
3352 0 : let Some((running, remaining)) = res else {
3353 0 : return Vec::new();
3354 : };
3355 0 : let mut result = Vec::new();
3356 0 : if let Some((id, running)) = running {
3357 0 : result.extend(running.into_compact_info_resp(id, true));
3358 0 : }
3359 0 : for (id, job) in remaining {
3360 0 : result.extend(job.into_compact_info_resp(id, false));
3361 0 : }
3362 0 : result
3363 0 : }
3364 :
3365 : /// Schedule a compaction task for a timeline.
3366 0 : pub(crate) async fn schedule_compaction(
3367 0 : &self,
3368 0 : timeline_id: TimelineId,
3369 0 : options: CompactOptions,
3370 0 : ) -> anyhow::Result<tokio::sync::oneshot::Receiver<()>> {
3371 0 : let (tx, rx) = tokio::sync::oneshot::channel();
3372 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3373 0 : let q = guard
3374 0 : .entry(timeline_id)
3375 0 : .or_insert_with(|| Arc::new(GcCompactionQueue::new()));
3376 0 : q.schedule_manual_compaction(options, Some(tx));
3377 0 : Ok(rx)
3378 0 : }
3379 :
3380 : /// Performs periodic housekeeping, via the tenant housekeeping background task.
3381 0 : async fn housekeeping(&self) {
3382 : // Call through to all timelines to freeze ephemeral layers as needed. This usually happens
3383 : // during ingest, but we don't want idle timelines to hold open layers for too long.
3384 : //
3385 : // We don't do this if the tenant can't upload layers (i.e. it's in stale attachment mode).
3386 : // We don't run compaction in this case either, and don't want to keep flushing tiny L0
3387 : // layers that won't be compacted down.
3388 0 : if self.tenant_conf.load().location.may_upload_layers_hint() {
3389 0 : let timelines = self
3390 0 : .timelines
3391 0 : .lock()
3392 0 : .unwrap()
3393 0 : .values()
3394 0 : .filter(|tli| tli.is_active())
3395 0 : .cloned()
3396 0 : .collect_vec();
3397 :
3398 0 : for timeline in timelines {
3399 : // Include a span with the timeline ID. The parent span already has the tenant ID.
3400 0 : let span =
3401 0 : info_span!("maybe_freeze_ephemeral_layer", timeline_id = %timeline.timeline_id);
3402 0 : timeline
3403 0 : .maybe_freeze_ephemeral_layer()
3404 0 : .instrument(span)
3405 0 : .await;
3406 : }
3407 0 : }
3408 :
3409 : // Shut down walredo if idle.
3410 : const WALREDO_IDLE_TIMEOUT: Duration = Duration::from_secs(180);
3411 0 : if let Some(ref walredo_mgr) = self.walredo_mgr {
3412 0 : walredo_mgr.maybe_quiesce(WALREDO_IDLE_TIMEOUT);
3413 0 : }
3414 :
3415 : // Update the feature resolver with the latest tenant-spcific data.
3416 0 : self.feature_resolver.refresh_properties_and_flags(self);
3417 0 : }
3418 :
3419 0 : pub fn timeline_has_no_attached_children(&self, timeline_id: TimelineId) -> bool {
3420 0 : let timelines = self.timelines.lock().unwrap();
3421 0 : !timelines
3422 0 : .iter()
3423 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(timeline_id))
3424 0 : }
3425 :
3426 1371 : pub fn current_state(&self) -> TenantState {
3427 1371 : self.state.borrow().clone()
3428 1371 : }
3429 :
3430 990 : pub fn is_active(&self) -> bool {
3431 990 : self.current_state() == TenantState::Active
3432 990 : }
3433 :
3434 0 : pub fn generation(&self) -> Generation {
3435 0 : self.generation
3436 0 : }
3437 :
3438 0 : pub(crate) fn wal_redo_manager_status(&self) -> Option<WalRedoManagerStatus> {
3439 0 : self.walredo_mgr.as_ref().and_then(|mgr| mgr.status())
3440 0 : }
3441 :
3442 : /// Changes tenant status to active, unless shutdown was already requested.
3443 : ///
3444 : /// `background_jobs_can_start` is an optional barrier set to a value during pageserver startup
3445 : /// to delay background jobs. Background jobs can be started right away when None is given.
3446 0 : fn activate(
3447 0 : self: &Arc<Self>,
3448 0 : broker_client: BrokerClientChannel,
3449 0 : background_jobs_can_start: Option<&completion::Barrier>,
3450 0 : ctx: &RequestContext,
3451 0 : ) {
3452 0 : span::debug_assert_current_span_has_tenant_id();
3453 :
3454 0 : let mut activating = false;
3455 0 : self.state.send_modify(|current_state| {
3456 : use pageserver_api::models::ActivatingFrom;
3457 0 : match &*current_state {
3458 : TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
3459 0 : panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {current_state:?}");
3460 : }
3461 0 : TenantState::Attaching => {
3462 0 : *current_state = TenantState::Activating(ActivatingFrom::Attaching);
3463 0 : }
3464 : }
3465 0 : debug!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), "Activating tenant");
3466 0 : activating = true;
3467 : // Continue outside the closure. We need to grab timelines.lock()
3468 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
3469 0 : });
3470 :
3471 0 : if activating {
3472 0 : let timelines_accessor = self.timelines.lock().unwrap();
3473 0 : let timelines_offloaded_accessor = self.timelines_offloaded.lock().unwrap();
3474 0 : let timelines_to_activate = timelines_accessor
3475 0 : .values()
3476 0 : .filter(|timeline| !(timeline.is_broken() || timeline.is_stopping()));
3477 :
3478 : // Spawn gc and compaction loops. The loops will shut themselves
3479 : // down when they notice that the tenant is inactive.
3480 0 : tasks::start_background_loops(self, background_jobs_can_start);
3481 :
3482 0 : let mut activated_timelines = 0;
3483 :
3484 0 : for timeline in timelines_to_activate {
3485 0 : timeline.activate(
3486 0 : self.clone(),
3487 0 : broker_client.clone(),
3488 0 : background_jobs_can_start,
3489 0 : &ctx.with_scope_timeline(timeline),
3490 0 : );
3491 0 : activated_timelines += 1;
3492 0 : }
3493 :
3494 0 : let tid = self.tenant_shard_id.tenant_id.to_string();
3495 0 : let shard_id = self.tenant_shard_id.shard_slug().to_string();
3496 0 : let offloaded_timeline_count = timelines_offloaded_accessor.len();
3497 0 : TENANT_OFFLOADED_TIMELINES
3498 0 : .with_label_values(&[&tid, &shard_id])
3499 0 : .set(offloaded_timeline_count as u64);
3500 :
3501 0 : self.state.send_modify(move |current_state| {
3502 0 : assert!(
3503 0 : matches!(current_state, TenantState::Activating(_)),
3504 0 : "set_stopping and set_broken wait for us to leave Activating state",
3505 : );
3506 0 : *current_state = TenantState::Active;
3507 :
3508 0 : let elapsed = self.constructed_at.elapsed();
3509 0 : let total_timelines = timelines_accessor.len();
3510 :
3511 : // log a lot of stuff, because some tenants sometimes suffer from user-visible
3512 : // times to activate. see https://github.com/neondatabase/neon/issues/4025
3513 0 : info!(
3514 0 : since_creation_millis = elapsed.as_millis(),
3515 0 : tenant_id = %self.tenant_shard_id.tenant_id,
3516 0 : shard_id = %self.tenant_shard_id.shard_slug(),
3517 : activated_timelines,
3518 : total_timelines,
3519 0 : post_state = <&'static str>::from(&*current_state),
3520 0 : "activation attempt finished"
3521 : );
3522 :
3523 0 : TENANT.activation.observe(elapsed.as_secs_f64());
3524 0 : });
3525 0 : }
3526 0 : }
3527 :
3528 : /// Shutdown the tenant and join all of the spawned tasks.
3529 : ///
3530 : /// The method caters for all use-cases:
3531 : /// - pageserver shutdown (freeze_and_flush == true)
3532 : /// - detach + ignore (freeze_and_flush == false)
3533 : ///
3534 : /// This will attempt to shutdown even if tenant is broken.
3535 : ///
3536 : /// `shutdown_progress` is a [`completion::Barrier`] for the shutdown initiated by this call.
3537 : /// If the tenant is already shutting down, we return a clone of the first shutdown call's
3538 : /// `Barrier` as an `Err`. This not-first caller can use the returned barrier to join with
3539 : /// the ongoing shutdown.
3540 3 : async fn shutdown(
3541 3 : &self,
3542 3 : shutdown_progress: completion::Barrier,
3543 3 : shutdown_mode: timeline::ShutdownMode,
3544 3 : ) -> Result<(), completion::Barrier> {
3545 3 : span::debug_assert_current_span_has_tenant_id();
3546 :
3547 : // Set tenant (and its timlines) to Stoppping state.
3548 : //
3549 : // Since we can only transition into Stopping state after activation is complete,
3550 : // run it in a JoinSet so all tenants have a chance to stop before we get SIGKILLed.
3551 : //
3552 : // Transitioning tenants to Stopping state has a couple of non-obvious side effects:
3553 : // 1. Lock out any new requests to the tenants.
3554 : // 2. Signal cancellation to WAL receivers (we wait on it below).
3555 : // 3. Signal cancellation for other tenant background loops.
3556 : // 4. ???
3557 : //
3558 : // The waiting for the cancellation is not done uniformly.
3559 : // We certainly wait for WAL receivers to shut down.
3560 : // That is necessary so that no new data comes in before the freeze_and_flush.
3561 : // But the tenant background loops are joined-on in our caller.
3562 : // It's mesed up.
3563 : // we just ignore the failure to stop
3564 :
3565 : // If we're still attaching, fire the cancellation token early to drop out: this
3566 : // will prevent us flushing, but ensures timely shutdown if some I/O during attach
3567 : // is very slow.
3568 3 : let shutdown_mode = if matches!(self.current_state(), TenantState::Attaching) {
3569 0 : self.cancel.cancel();
3570 :
3571 : // Having fired our cancellation token, do not try and flush timelines: their cancellation tokens
3572 : // are children of ours, so their flush loops will have shut down already
3573 0 : timeline::ShutdownMode::Hard
3574 : } else {
3575 3 : shutdown_mode
3576 : };
3577 :
3578 3 : match self.set_stopping(shutdown_progress).await {
3579 3 : Ok(()) => {}
3580 0 : Err(SetStoppingError::Broken) => {
3581 0 : // assume that this is acceptable
3582 0 : }
3583 0 : Err(SetStoppingError::AlreadyStopping(other)) => {
3584 : // give caller the option to wait for this this shutdown
3585 0 : info!("Tenant::shutdown: AlreadyStopping");
3586 0 : return Err(other);
3587 : }
3588 : };
3589 :
3590 3 : let mut js = tokio::task::JoinSet::new();
3591 : {
3592 3 : let timelines = self.timelines.lock().unwrap();
3593 3 : timelines.values().for_each(|timeline| {
3594 3 : let timeline = Arc::clone(timeline);
3595 3 : let timeline_id = timeline.timeline_id;
3596 3 : let span = tracing::info_span!("timeline_shutdown", %timeline_id, ?shutdown_mode);
3597 3 : js.spawn(async move { timeline.shutdown(shutdown_mode).instrument(span).await });
3598 3 : });
3599 : }
3600 : {
3601 3 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
3602 3 : timelines_offloaded.values().for_each(|timeline| {
3603 0 : timeline.defuse_for_tenant_drop();
3604 0 : });
3605 : }
3606 : {
3607 3 : let mut timelines_importing = self.timelines_importing.lock().unwrap();
3608 3 : timelines_importing
3609 3 : .drain()
3610 3 : .for_each(|(timeline_id, importing_timeline)| {
3611 0 : let span = tracing::info_span!("importing_timeline_shutdown", %timeline_id);
3612 0 : js.spawn(async move { importing_timeline.shutdown().instrument(span).await });
3613 0 : });
3614 : }
3615 : // test_long_timeline_create_then_tenant_delete is leaning on this message
3616 3 : tracing::info!("Waiting for timelines...");
3617 6 : while let Some(res) = js.join_next().await {
3618 0 : match res {
3619 3 : Ok(()) => {}
3620 0 : Err(je) if je.is_cancelled() => unreachable!("no cancelling used"),
3621 0 : Err(je) if je.is_panic() => { /* logged already */ }
3622 0 : Err(je) => warn!("unexpected JoinError: {je:?}"),
3623 : }
3624 : }
3625 :
3626 3 : if let ShutdownMode::Reload = shutdown_mode {
3627 0 : tracing::info!("Flushing deletion queue");
3628 0 : if let Err(e) = self.deletion_queue_client.flush().await {
3629 0 : match e {
3630 0 : DeletionQueueError::ShuttingDown => {
3631 0 : // This is the only error we expect for now. In the future, if more error
3632 0 : // variants are added, we should handle them here.
3633 0 : }
3634 : }
3635 0 : }
3636 3 : }
3637 :
3638 : // We cancel the Tenant's cancellation token _after_ the timelines have all shut down. This permits
3639 : // them to continue to do work during their shutdown methods, e.g. flushing data.
3640 3 : tracing::debug!("Cancelling CancellationToken");
3641 3 : self.cancel.cancel();
3642 :
3643 : // shutdown all tenant and timeline tasks: gc, compaction, page service
3644 : // No new tasks will be started for this tenant because it's in `Stopping` state.
3645 : //
3646 : // this will additionally shutdown and await all timeline tasks.
3647 3 : tracing::debug!("Waiting for tasks...");
3648 3 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), None).await;
3649 :
3650 3 : if let Some(walredo_mgr) = self.walredo_mgr.as_ref() {
3651 3 : walredo_mgr.shutdown().await;
3652 0 : }
3653 :
3654 : // Wait for any in-flight operations to complete
3655 3 : self.gate.close().await;
3656 :
3657 3 : remove_tenant_metrics(&self.tenant_shard_id);
3658 :
3659 3 : Ok(())
3660 3 : }
3661 :
3662 : /// Change tenant status to Stopping, to mark that it is being shut down.
3663 : ///
3664 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
3665 : ///
3666 : /// This function is not cancel-safe!
3667 3 : async fn set_stopping(&self, progress: completion::Barrier) -> Result<(), SetStoppingError> {
3668 3 : let mut rx = self.state.subscribe();
3669 :
3670 : // cannot stop before we're done activating, so wait out until we're done activating
3671 3 : rx.wait_for(|state| match state {
3672 : TenantState::Activating(_) | TenantState::Attaching => {
3673 0 : info!("waiting for {state} to turn Active|Broken|Stopping");
3674 0 : false
3675 : }
3676 3 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3677 3 : })
3678 3 : .await
3679 3 : .expect("cannot drop self.state while on a &self method");
3680 :
3681 : // we now know we're done activating, let's see whether this task is the winner to transition into Stopping
3682 3 : let mut err = None;
3683 3 : let stopping = self.state.send_if_modified(|current_state| match current_state {
3684 : TenantState::Activating(_) | TenantState::Attaching => {
3685 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
3686 : }
3687 : TenantState::Active => {
3688 : // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
3689 : // are created after the transition to Stopping. That's harmless, as the Timelines
3690 : // won't be accessible to anyone afterwards, because the Tenant is in Stopping state.
3691 3 : *current_state = TenantState::Stopping { progress: Some(progress) };
3692 : // Continue stopping outside the closure. We need to grab timelines.lock()
3693 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
3694 3 : true
3695 : }
3696 : TenantState::Stopping { progress: None } => {
3697 : // An attach was cancelled, and the attach transitioned the tenant from Attaching to
3698 : // Stopping(None) to let us know it exited. Register our progress and continue.
3699 0 : *current_state = TenantState::Stopping { progress: Some(progress) };
3700 0 : true
3701 : }
3702 0 : TenantState::Broken { reason, .. } => {
3703 0 : info!(
3704 0 : "Cannot set tenant to Stopping state, it is in Broken state due to: {reason}"
3705 : );
3706 0 : err = Some(SetStoppingError::Broken);
3707 0 : false
3708 : }
3709 0 : TenantState::Stopping { progress: Some(progress) } => {
3710 0 : info!("Tenant is already in Stopping state");
3711 0 : err = Some(SetStoppingError::AlreadyStopping(progress.clone()));
3712 0 : false
3713 : }
3714 3 : });
3715 3 : match (stopping, err) {
3716 3 : (true, None) => {} // continue
3717 0 : (false, Some(err)) => return Err(err),
3718 0 : (true, Some(_)) => unreachable!(
3719 : "send_if_modified closure must error out if not transitioning to Stopping"
3720 : ),
3721 0 : (false, None) => unreachable!(
3722 : "send_if_modified closure must return true if transitioning to Stopping"
3723 : ),
3724 : }
3725 :
3726 3 : let timelines_accessor = self.timelines.lock().unwrap();
3727 3 : let not_broken_timelines = timelines_accessor
3728 3 : .values()
3729 3 : .filter(|timeline| !timeline.is_broken());
3730 6 : for timeline in not_broken_timelines {
3731 3 : timeline.set_state(TimelineState::Stopping);
3732 3 : }
3733 3 : Ok(())
3734 3 : }
3735 :
3736 : /// Method for tenant::mgr to transition us into Broken state in case of a late failure in
3737 : /// `remove_tenant_from_memory`
3738 : ///
3739 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
3740 : ///
3741 : /// In tests, we also use this to set tenants to Broken state on purpose.
3742 0 : pub(crate) async fn set_broken(&self, reason: String) {
3743 0 : let mut rx = self.state.subscribe();
3744 :
3745 : // The load & attach routines own the tenant state until it has reached `Active`.
3746 : // So, wait until it's done.
3747 0 : rx.wait_for(|state| match state {
3748 : TenantState::Activating(_) | TenantState::Attaching => {
3749 0 : info!(
3750 0 : "waiting for {} to turn Active|Broken|Stopping",
3751 0 : <&'static str>::from(state)
3752 : );
3753 0 : false
3754 : }
3755 0 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3756 0 : })
3757 0 : .await
3758 0 : .expect("cannot drop self.state while on a &self method");
3759 :
3760 : // we now know we're done activating, let's see whether this task is the winner to transition into Broken
3761 0 : self.set_broken_no_wait(reason)
3762 0 : }
3763 :
3764 0 : pub(crate) fn set_broken_no_wait(&self, reason: impl Display) {
3765 0 : let reason = reason.to_string();
3766 0 : self.state.send_modify(|current_state| {
3767 0 : match *current_state {
3768 : TenantState::Activating(_) | TenantState::Attaching => {
3769 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
3770 : }
3771 : TenantState::Active => {
3772 0 : if cfg!(feature = "testing") {
3773 0 : warn!("Changing Active tenant to Broken state, reason: {}", reason);
3774 0 : *current_state = TenantState::broken_from_reason(reason);
3775 : } else {
3776 0 : unreachable!("not allowed to call set_broken on Active tenants in non-testing builds")
3777 : }
3778 : }
3779 : TenantState::Broken { .. } => {
3780 0 : warn!("Tenant is already in Broken state");
3781 : }
3782 : // This is the only "expected" path, any other path is a bug.
3783 : TenantState::Stopping { .. } => {
3784 0 : warn!(
3785 0 : "Marking Stopping tenant as Broken state, reason: {}",
3786 : reason
3787 : );
3788 0 : *current_state = TenantState::broken_from_reason(reason);
3789 : }
3790 : }
3791 0 : });
3792 0 : }
3793 :
3794 0 : pub fn subscribe_for_state_updates(&self) -> watch::Receiver<TenantState> {
3795 0 : self.state.subscribe()
3796 0 : }
3797 :
3798 : /// The activate_now semaphore is initialized with zero units. As soon as
3799 : /// we add a unit, waiters will be able to acquire a unit and proceed.
3800 0 : pub(crate) fn activate_now(&self) {
3801 0 : self.activate_now_sem.add_permits(1);
3802 0 : }
3803 :
3804 0 : pub(crate) async fn wait_to_become_active(
3805 0 : &self,
3806 0 : timeout: Duration,
3807 0 : ) -> Result<(), GetActiveTenantError> {
3808 0 : let mut receiver = self.state.subscribe();
3809 : loop {
3810 0 : let current_state = receiver.borrow_and_update().clone();
3811 0 : match current_state {
3812 : TenantState::Attaching | TenantState::Activating(_) => {
3813 : // in these states, there's a chance that we can reach ::Active
3814 0 : self.activate_now();
3815 0 : match timeout_cancellable(timeout, &self.cancel, receiver.changed()).await {
3816 0 : Ok(r) => {
3817 0 : r.map_err(
3818 : |_e: tokio::sync::watch::error::RecvError|
3819 : // Tenant existed but was dropped: report it as non-existent
3820 0 : GetActiveTenantError::NotFound(GetTenantError::ShardNotFound(self.tenant_shard_id))
3821 0 : )?
3822 : }
3823 : Err(TimeoutCancellableError::Cancelled) => {
3824 0 : return Err(GetActiveTenantError::Cancelled);
3825 : }
3826 : Err(TimeoutCancellableError::Timeout) => {
3827 0 : return Err(GetActiveTenantError::WaitForActiveTimeout {
3828 0 : latest_state: Some(self.current_state()),
3829 0 : wait_time: timeout,
3830 0 : });
3831 : }
3832 : }
3833 : }
3834 : TenantState::Active => {
3835 0 : return Ok(());
3836 : }
3837 0 : TenantState::Broken { reason, .. } => {
3838 : // This is fatal, and reported distinctly from the general case of "will never be active" because
3839 : // it's logically a 500 to external API users (broken is always a bug).
3840 0 : return Err(GetActiveTenantError::Broken(reason));
3841 : }
3842 : TenantState::Stopping { .. } => {
3843 : // There's no chance the tenant can transition back into ::Active
3844 0 : return Err(GetActiveTenantError::WillNotBecomeActive(current_state));
3845 : }
3846 : }
3847 : }
3848 0 : }
3849 :
3850 0 : pub(crate) fn get_attach_mode(&self) -> AttachmentMode {
3851 0 : self.tenant_conf.load().location.attach_mode
3852 0 : }
3853 :
3854 : /// For API access: generate a LocationConfig equivalent to the one that would be used to
3855 : /// create a Tenant in the same state. Do not use this in hot paths: it's for relatively
3856 : /// rare external API calls, like a reconciliation at startup.
3857 0 : pub(crate) fn get_location_conf(&self) -> models::LocationConfig {
3858 0 : let attached_tenant_conf = self.tenant_conf.load();
3859 :
3860 0 : let location_config_mode = match attached_tenant_conf.location.attach_mode {
3861 0 : AttachmentMode::Single => models::LocationConfigMode::AttachedSingle,
3862 0 : AttachmentMode::Multi => models::LocationConfigMode::AttachedMulti,
3863 0 : AttachmentMode::Stale => models::LocationConfigMode::AttachedStale,
3864 : };
3865 :
3866 0 : models::LocationConfig {
3867 0 : mode: location_config_mode,
3868 0 : generation: self.generation.into(),
3869 0 : secondary_conf: None,
3870 0 : shard_number: self.shard_identity.number.0,
3871 0 : shard_count: self.shard_identity.count.literal(),
3872 0 : shard_stripe_size: self.shard_identity.stripe_size.0,
3873 0 : tenant_conf: attached_tenant_conf.tenant_conf.clone(),
3874 0 : }
3875 0 : }
3876 :
3877 0 : pub(crate) fn get_tenant_shard_id(&self) -> &TenantShardId {
3878 0 : &self.tenant_shard_id
3879 0 : }
3880 :
3881 0 : pub(crate) fn get_shard_identity(&self) -> ShardIdentity {
3882 0 : self.shard_identity
3883 0 : }
3884 :
3885 120 : pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
3886 120 : self.shard_identity.stripe_size
3887 120 : }
3888 :
3889 0 : pub(crate) fn get_generation(&self) -> Generation {
3890 0 : self.generation
3891 0 : }
3892 :
3893 : /// This function partially shuts down the tenant (it shuts down the Timelines) and is fallible,
3894 : /// and can leave the tenant in a bad state if it fails. The caller is responsible for
3895 : /// resetting this tenant to a valid state if we fail.
3896 0 : pub(crate) async fn split_prepare(
3897 0 : &self,
3898 0 : child_shards: &Vec<TenantShardId>,
3899 0 : ) -> anyhow::Result<()> {
3900 0 : let (timelines, offloaded) = {
3901 0 : let timelines = self.timelines.lock().unwrap();
3902 0 : let offloaded = self.timelines_offloaded.lock().unwrap();
3903 0 : (timelines.clone(), offloaded.clone())
3904 0 : };
3905 0 : let timelines_iter = timelines
3906 0 : .values()
3907 0 : .map(TimelineOrOffloadedArcRef::<'_>::from)
3908 0 : .chain(
3909 0 : offloaded
3910 0 : .values()
3911 0 : .map(TimelineOrOffloadedArcRef::<'_>::from),
3912 : );
3913 0 : for timeline in timelines_iter {
3914 : // We do not block timeline creation/deletion during splits inside the pageserver: it is up to higher levels
3915 : // to ensure that they do not start a split if currently in the process of doing these.
3916 :
3917 0 : let timeline_id = timeline.timeline_id();
3918 :
3919 0 : if let TimelineOrOffloadedArcRef::Timeline(timeline) = timeline {
3920 : // Upload an index from the parent: this is partly to provide freshness for the
3921 : // child tenants that will copy it, and partly for general ease-of-debugging: there will
3922 : // always be a parent shard index in the same generation as we wrote the child shard index.
3923 0 : tracing::info!(%timeline_id, "Uploading index");
3924 0 : timeline
3925 0 : .remote_client
3926 0 : .schedule_index_upload_for_file_changes()?;
3927 0 : timeline.remote_client.wait_completion().await?;
3928 0 : }
3929 :
3930 0 : let remote_client = match timeline {
3931 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.remote_client.clone(),
3932 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => {
3933 0 : let remote_client = self
3934 0 : .build_timeline_client(offloaded.timeline_id, self.remote_storage.clone());
3935 0 : Arc::new(remote_client)
3936 : }
3937 : TimelineOrOffloadedArcRef::Importing(_) => {
3938 0 : unreachable!("Importing timelines are not included in the iterator")
3939 : }
3940 : };
3941 :
3942 : // Shut down the timeline's remote client: this means that the indices we write
3943 : // for child shards will not be invalidated by the parent shard deleting layers.
3944 0 : tracing::info!(%timeline_id, "Shutting down remote storage client");
3945 0 : remote_client.shutdown().await;
3946 :
3947 : // Download methods can still be used after shutdown, as they don't flow through the remote client's
3948 : // queue. In principal the RemoteTimelineClient could provide this without downloading it, but this
3949 : // operation is rare, so it's simpler to just download it (and robustly guarantees that the index
3950 : // we use here really is the remotely persistent one).
3951 0 : tracing::info!(%timeline_id, "Downloading index_part from parent");
3952 0 : let result = remote_client
3953 0 : .download_index_file(&self.cancel)
3954 0 : .instrument(info_span!("download_index_file", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))
3955 0 : .await?;
3956 0 : let index_part = match result {
3957 : MaybeDeletedIndexPart::Deleted(_) => {
3958 0 : anyhow::bail!("Timeline deletion happened concurrently with split")
3959 : }
3960 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
3961 : };
3962 :
3963 : // A shard split may not take place while a timeline import is on-going
3964 : // for the tenant. Timeline imports run as part of each tenant shard
3965 : // and rely on the sharding scheme to split the work among pageservers.
3966 : // If we were to split in the middle of this process, we would have to
3967 : // either ensure that it's driven to completion on the old shard set
3968 : // or transfer it to the new shard set. It's technically possible, but complex.
3969 0 : match index_part.import_pgdata {
3970 0 : Some(ref import) if !import.is_done() => {
3971 0 : anyhow::bail!(
3972 0 : "Cannot split due to import with idempotency key: {:?}",
3973 0 : import.idempotency_key()
3974 : );
3975 : }
3976 0 : Some(_) | None => {
3977 0 : // fallthrough
3978 0 : }
3979 : }
3980 :
3981 0 : for child_shard in child_shards {
3982 0 : tracing::info!(%timeline_id, "Uploading index_part for child {}", child_shard.to_index());
3983 0 : upload_index_part(
3984 0 : &self.remote_storage,
3985 0 : child_shard,
3986 0 : &timeline_id,
3987 0 : self.generation,
3988 0 : &index_part,
3989 0 : &self.cancel,
3990 0 : )
3991 0 : .await?;
3992 : }
3993 : }
3994 :
3995 0 : let tenant_manifest = self.build_tenant_manifest();
3996 0 : for child_shard in child_shards {
3997 0 : tracing::info!(
3998 0 : "Uploading tenant manifest for child {}",
3999 0 : child_shard.to_index()
4000 : );
4001 0 : upload_tenant_manifest(
4002 0 : &self.remote_storage,
4003 0 : child_shard,
4004 0 : self.generation,
4005 0 : &tenant_manifest,
4006 0 : &self.cancel,
4007 0 : )
4008 0 : .await?;
4009 : }
4010 :
4011 0 : Ok(())
4012 0 : }
4013 :
4014 0 : pub(crate) fn get_sizes(&self) -> TopTenantShardItem {
4015 0 : let mut result = TopTenantShardItem {
4016 0 : id: self.tenant_shard_id,
4017 0 : resident_size: 0,
4018 0 : physical_size: 0,
4019 0 : max_logical_size: 0,
4020 0 : max_logical_size_per_shard: 0,
4021 0 : };
4022 :
4023 0 : for timeline in self.timelines.lock().unwrap().values() {
4024 0 : result.resident_size += timeline.metrics.resident_physical_size_gauge.get();
4025 0 :
4026 0 : result.physical_size += timeline
4027 0 : .remote_client
4028 0 : .metrics
4029 0 : .remote_physical_size_gauge
4030 0 : .get();
4031 0 : result.max_logical_size = std::cmp::max(
4032 0 : result.max_logical_size,
4033 0 : timeline.metrics.current_logical_size_gauge.get(),
4034 0 : );
4035 0 : }
4036 :
4037 0 : result.max_logical_size_per_shard = result
4038 0 : .max_logical_size
4039 0 : .div_ceil(self.tenant_shard_id.shard_count.count() as u64);
4040 :
4041 0 : result
4042 0 : }
4043 : }
4044 :
4045 : /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
4046 : /// perform a topological sort, so that the parent of each timeline comes
4047 : /// before the children.
4048 : /// E extracts the ancestor from T
4049 : /// This allows for T to be different. It can be TimelineMetadata, can be Timeline itself, etc.
4050 119 : fn tree_sort_timelines<T, E>(
4051 119 : timelines: HashMap<TimelineId, T>,
4052 119 : extractor: E,
4053 119 : ) -> anyhow::Result<Vec<(TimelineId, T)>>
4054 119 : where
4055 119 : E: Fn(&T) -> Option<TimelineId>,
4056 : {
4057 119 : let mut result = Vec::with_capacity(timelines.len());
4058 :
4059 119 : let mut now = Vec::with_capacity(timelines.len());
4060 : // (ancestor, children)
4061 119 : let mut later: HashMap<TimelineId, Vec<(TimelineId, T)>> =
4062 119 : HashMap::with_capacity(timelines.len());
4063 :
4064 122 : for (timeline_id, value) in timelines {
4065 3 : if let Some(ancestor_id) = extractor(&value) {
4066 1 : let children = later.entry(ancestor_id).or_default();
4067 1 : children.push((timeline_id, value));
4068 2 : } else {
4069 2 : now.push((timeline_id, value));
4070 2 : }
4071 : }
4072 :
4073 122 : while let Some((timeline_id, metadata)) = now.pop() {
4074 3 : result.push((timeline_id, metadata));
4075 : // All children of this can be loaded now
4076 3 : if let Some(mut children) = later.remove(&timeline_id) {
4077 1 : now.append(&mut children);
4078 2 : }
4079 : }
4080 :
4081 : // All timelines should be visited now. Unless there were timelines with missing ancestors.
4082 119 : if !later.is_empty() {
4083 0 : for (missing_id, orphan_ids) in later {
4084 0 : for (orphan_id, _) in orphan_ids {
4085 0 : error!(
4086 0 : "could not load timeline {orphan_id} because its ancestor timeline {missing_id} could not be loaded"
4087 : );
4088 : }
4089 : }
4090 0 : bail!("could not load tenant because some timelines are missing ancestors");
4091 119 : }
4092 :
4093 119 : Ok(result)
4094 119 : }
4095 :
4096 : impl TenantShard {
4097 0 : pub fn tenant_specific_overrides(&self) -> pageserver_api::models::TenantConfig {
4098 0 : self.tenant_conf.load().tenant_conf.clone()
4099 0 : }
4100 :
4101 0 : pub fn effective_config(&self) -> pageserver_api::config::TenantConfigToml {
4102 0 : self.tenant_specific_overrides()
4103 0 : .merge(self.conf.default_tenant_conf.clone())
4104 0 : }
4105 :
4106 0 : pub fn get_checkpoint_distance(&self) -> u64 {
4107 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4108 0 : tenant_conf
4109 0 : .checkpoint_distance
4110 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
4111 0 : }
4112 :
4113 0 : pub fn get_checkpoint_timeout(&self) -> Duration {
4114 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4115 0 : tenant_conf
4116 0 : .checkpoint_timeout
4117 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
4118 0 : }
4119 :
4120 0 : pub fn get_compaction_target_size(&self) -> u64 {
4121 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4122 0 : tenant_conf
4123 0 : .compaction_target_size
4124 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
4125 0 : }
4126 :
4127 0 : pub fn get_compaction_period(&self) -> Duration {
4128 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4129 0 : tenant_conf
4130 0 : .compaction_period
4131 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
4132 0 : }
4133 :
4134 0 : pub fn get_compaction_threshold(&self) -> usize {
4135 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4136 0 : tenant_conf
4137 0 : .compaction_threshold
4138 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
4139 0 : }
4140 :
4141 0 : pub fn get_rel_size_v2_enabled(&self) -> bool {
4142 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4143 0 : tenant_conf
4144 0 : .rel_size_v2_enabled
4145 0 : .unwrap_or(self.conf.default_tenant_conf.rel_size_v2_enabled)
4146 0 : }
4147 :
4148 0 : pub fn get_compaction_upper_limit(&self) -> usize {
4149 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4150 0 : tenant_conf
4151 0 : .compaction_upper_limit
4152 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_upper_limit)
4153 0 : }
4154 :
4155 0 : pub fn get_compaction_l0_first(&self) -> bool {
4156 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4157 0 : tenant_conf
4158 0 : .compaction_l0_first
4159 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_first)
4160 0 : }
4161 :
4162 121 : pub fn get_gc_horizon(&self) -> u64 {
4163 121 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4164 121 : tenant_conf
4165 121 : .gc_horizon
4166 121 : .unwrap_or(self.conf.default_tenant_conf.gc_horizon)
4167 121 : }
4168 :
4169 0 : pub fn get_gc_period(&self) -> Duration {
4170 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4171 0 : tenant_conf
4172 0 : .gc_period
4173 0 : .unwrap_or(self.conf.default_tenant_conf.gc_period)
4174 0 : }
4175 :
4176 0 : pub fn get_image_creation_threshold(&self) -> usize {
4177 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4178 0 : tenant_conf
4179 0 : .image_creation_threshold
4180 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
4181 0 : }
4182 :
4183 : // HADRON
4184 0 : pub fn get_image_creation_timeout(&self) -> Option<Duration> {
4185 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4186 0 : tenant_conf.image_layer_force_creation_period.or(self
4187 0 : .conf
4188 0 : .default_tenant_conf
4189 0 : .image_layer_force_creation_period)
4190 0 : }
4191 :
4192 2 : pub fn get_pitr_interval(&self) -> Duration {
4193 2 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4194 2 : tenant_conf
4195 2 : .pitr_interval
4196 2 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
4197 2 : }
4198 :
4199 0 : pub fn get_min_resident_size_override(&self) -> Option<u64> {
4200 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4201 0 : tenant_conf
4202 0 : .min_resident_size_override
4203 0 : .or(self.conf.default_tenant_conf.min_resident_size_override)
4204 0 : }
4205 :
4206 0 : pub fn get_heatmap_period(&self) -> Option<Duration> {
4207 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4208 0 : let heatmap_period = tenant_conf
4209 0 : .heatmap_period
4210 0 : .unwrap_or(self.conf.default_tenant_conf.heatmap_period);
4211 0 : if heatmap_period.is_zero() {
4212 0 : None
4213 : } else {
4214 0 : Some(heatmap_period)
4215 : }
4216 0 : }
4217 :
4218 0 : pub fn get_lsn_lease_length(&self) -> Duration {
4219 0 : Self::get_lsn_lease_length_impl(self.conf, &self.tenant_conf.load().tenant_conf)
4220 0 : }
4221 :
4222 119 : pub fn get_lsn_lease_length_impl(
4223 119 : conf: &'static PageServerConf,
4224 119 : tenant_conf: &pageserver_api::models::TenantConfig,
4225 119 : ) -> Duration {
4226 119 : tenant_conf
4227 119 : .lsn_lease_length
4228 119 : .unwrap_or(conf.default_tenant_conf.lsn_lease_length)
4229 119 : }
4230 :
4231 0 : pub fn get_timeline_offloading_enabled(&self) -> bool {
4232 0 : if self.conf.timeline_offloading {
4233 0 : return true;
4234 0 : }
4235 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4236 0 : tenant_conf
4237 0 : .timeline_offloading
4238 0 : .unwrap_or(self.conf.default_tenant_conf.timeline_offloading)
4239 0 : }
4240 :
4241 : /// Generate an up-to-date TenantManifest based on the state of this Tenant.
4242 120 : fn build_tenant_manifest(&self) -> TenantManifest {
4243 : // Collect the offloaded timelines, and sort them for deterministic output.
4244 120 : let offloaded_timelines = self
4245 120 : .timelines_offloaded
4246 120 : .lock()
4247 120 : .unwrap()
4248 120 : .values()
4249 120 : .map(|tli| tli.manifest())
4250 120 : .sorted_by_key(|m| m.timeline_id)
4251 120 : .collect_vec();
4252 :
4253 120 : TenantManifest {
4254 120 : version: LATEST_TENANT_MANIFEST_VERSION,
4255 120 : stripe_size: Some(self.get_shard_stripe_size()),
4256 120 : offloaded_timelines,
4257 120 : }
4258 120 : }
4259 :
4260 1 : pub fn update_tenant_config<
4261 1 : F: Fn(
4262 1 : pageserver_api::models::TenantConfig,
4263 1 : ) -> anyhow::Result<pageserver_api::models::TenantConfig>,
4264 1 : >(
4265 1 : &self,
4266 1 : update: F,
4267 1 : ) -> anyhow::Result<pageserver_api::models::TenantConfig> {
4268 : // Use read-copy-update in order to avoid overwriting the location config
4269 : // state if this races with [`TenantShard::set_new_location_config`]. Note that
4270 : // this race is not possible if both request types come from the storage
4271 : // controller (as they should!) because an exclusive op lock is required
4272 : // on the storage controller side.
4273 :
4274 1 : self.tenant_conf
4275 1 : .try_rcu(|attached_conf| -> Result<_, anyhow::Error> {
4276 1 : Ok(Arc::new(AttachedTenantConf {
4277 1 : tenant_conf: update(attached_conf.tenant_conf.clone())?,
4278 1 : location: attached_conf.location,
4279 1 : lsn_lease_deadline: attached_conf.lsn_lease_deadline,
4280 : }))
4281 1 : })?;
4282 :
4283 1 : let updated = self.tenant_conf.load();
4284 :
4285 1 : self.tenant_conf_updated(&updated.tenant_conf);
4286 : // Don't hold self.timelines.lock() during the notifies.
4287 : // There's no risk of deadlock right now, but there could be if we consolidate
4288 : // mutexes in struct Timeline in the future.
4289 1 : let timelines = self.list_timelines();
4290 1 : for timeline in timelines {
4291 0 : timeline.tenant_conf_updated(&updated);
4292 0 : }
4293 :
4294 1 : Ok(updated.tenant_conf.clone())
4295 1 : }
4296 :
4297 0 : pub(crate) fn set_new_location_config(&self, new_conf: AttachedTenantConf) {
4298 0 : let new_tenant_conf = new_conf.tenant_conf.clone();
4299 :
4300 0 : self.tenant_conf.store(Arc::new(new_conf.clone()));
4301 :
4302 0 : self.tenant_conf_updated(&new_tenant_conf);
4303 : // Don't hold self.timelines.lock() during the notifies.
4304 : // There's no risk of deadlock right now, but there could be if we consolidate
4305 : // mutexes in struct Timeline in the future.
4306 0 : let timelines = self.list_timelines();
4307 0 : for timeline in timelines {
4308 0 : timeline.tenant_conf_updated(&new_conf);
4309 0 : }
4310 0 : }
4311 :
4312 120 : fn get_pagestream_throttle_config(
4313 120 : psconf: &'static PageServerConf,
4314 120 : overrides: &pageserver_api::models::TenantConfig,
4315 120 : ) -> throttle::Config {
4316 120 : overrides
4317 120 : .timeline_get_throttle
4318 120 : .clone()
4319 120 : .unwrap_or(psconf.default_tenant_conf.timeline_get_throttle.clone())
4320 120 : }
4321 :
4322 1 : pub(crate) fn tenant_conf_updated(&self, new_conf: &pageserver_api::models::TenantConfig) {
4323 1 : let conf = Self::get_pagestream_throttle_config(self.conf, new_conf);
4324 1 : self.pagestream_throttle.reconfigure(conf)
4325 1 : }
4326 :
4327 : /// Helper function to create a new Timeline struct.
4328 : ///
4329 : /// The returned Timeline is in Loading state. The caller is responsible for
4330 : /// initializing any on-disk state, and for inserting the Timeline to the 'timelines'
4331 : /// map.
4332 : ///
4333 : /// `validate_ancestor == false` is used when a timeline is created for deletion
4334 : /// and we might not have the ancestor present anymore which is fine for to be
4335 : /// deleted timelines.
4336 : #[allow(clippy::too_many_arguments)]
4337 235 : fn create_timeline_struct(
4338 235 : &self,
4339 235 : new_timeline_id: TimelineId,
4340 235 : new_metadata: &TimelineMetadata,
4341 235 : previous_heatmap: Option<PreviousHeatmap>,
4342 235 : ancestor: Option<Arc<Timeline>>,
4343 235 : resources: TimelineResources,
4344 235 : cause: CreateTimelineCause,
4345 235 : create_idempotency: CreateTimelineIdempotency,
4346 235 : gc_compaction_state: Option<GcCompactionState>,
4347 235 : rel_size_v2_status: Option<RelSizeMigration>,
4348 235 : rel_size_migrated_at: Option<Lsn>,
4349 235 : ctx: &RequestContext,
4350 235 : ) -> anyhow::Result<(Arc<Timeline>, RequestContext)> {
4351 235 : let state = match cause {
4352 : CreateTimelineCause::Load => {
4353 235 : let ancestor_id = new_metadata.ancestor_timeline();
4354 235 : anyhow::ensure!(
4355 235 : ancestor_id == ancestor.as_ref().map(|t| t.timeline_id),
4356 0 : "Timeline's {new_timeline_id} ancestor {ancestor_id:?} was not found"
4357 : );
4358 235 : TimelineState::Loading
4359 : }
4360 0 : CreateTimelineCause::Delete => TimelineState::Stopping,
4361 : };
4362 :
4363 235 : let pg_version = new_metadata.pg_version();
4364 :
4365 235 : let timeline = Timeline::new(
4366 235 : self.conf,
4367 235 : Arc::clone(&self.tenant_conf),
4368 235 : new_metadata,
4369 235 : previous_heatmap,
4370 235 : ancestor,
4371 235 : new_timeline_id,
4372 235 : self.tenant_shard_id,
4373 235 : self.generation,
4374 235 : self.shard_identity,
4375 235 : self.walredo_mgr.clone(),
4376 235 : resources,
4377 235 : pg_version,
4378 235 : state,
4379 235 : self.attach_wal_lag_cooldown.clone(),
4380 235 : create_idempotency,
4381 235 : gc_compaction_state,
4382 235 : rel_size_v2_status,
4383 235 : rel_size_migrated_at,
4384 235 : self.cancel.child_token(),
4385 : );
4386 :
4387 235 : let timeline_ctx = RequestContextBuilder::from(ctx)
4388 235 : .scope(context::Scope::new_timeline(&timeline))
4389 235 : .detached_child();
4390 :
4391 235 : Ok((timeline, timeline_ctx))
4392 235 : }
4393 :
4394 : /// [`TenantShard::shutdown`] must be called before dropping the returned [`TenantShard`] object
4395 : /// to ensure proper cleanup of background tasks and metrics.
4396 : //
4397 : // Allow too_many_arguments because a constructor's argument list naturally grows with the
4398 : // number of attributes in the struct: breaking these out into a builder wouldn't be helpful.
4399 : #[allow(clippy::too_many_arguments)]
4400 119 : fn new(
4401 119 : state: TenantState,
4402 119 : conf: &'static PageServerConf,
4403 119 : attached_conf: AttachedTenantConf,
4404 119 : shard_identity: ShardIdentity,
4405 119 : walredo_mgr: Option<Arc<WalRedoManager>>,
4406 119 : tenant_shard_id: TenantShardId,
4407 119 : remote_storage: GenericRemoteStorage,
4408 119 : deletion_queue_client: DeletionQueueClient,
4409 119 : l0_flush_global_state: L0FlushGlobalState,
4410 119 : basebackup_cache: Arc<BasebackupCache>,
4411 119 : feature_resolver: FeatureResolver,
4412 119 : ) -> TenantShard {
4413 119 : assert!(!attached_conf.location.generation.is_none());
4414 :
4415 119 : let (state, mut rx) = watch::channel(state);
4416 :
4417 119 : tokio::spawn(async move {
4418 : // reflect tenant state in metrics:
4419 : // - global per tenant state: TENANT_STATE_METRIC
4420 : // - "set" of broken tenants: BROKEN_TENANTS_SET
4421 : //
4422 : // set of broken tenants should not have zero counts so that it remains accessible for
4423 : // alerting.
4424 :
4425 119 : let tid = tenant_shard_id.to_string();
4426 119 : let shard_id = tenant_shard_id.shard_slug().to_string();
4427 119 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
4428 :
4429 235 : fn inspect_state(state: &TenantState) -> ([&'static str; 1], bool) {
4430 235 : ([state.into()], matches!(state, TenantState::Broken { .. }))
4431 235 : }
4432 :
4433 119 : let mut tuple = inspect_state(&rx.borrow_and_update());
4434 :
4435 119 : let is_broken = tuple.1;
4436 119 : let mut counted_broken = if is_broken {
4437 : // add the id to the set right away, there should not be any updates on the channel
4438 : // after before tenant is removed, if ever
4439 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
4440 0 : true
4441 : } else {
4442 119 : false
4443 : };
4444 :
4445 : loop {
4446 235 : let labels = &tuple.0;
4447 235 : let current = TENANT_STATE_METRIC.with_label_values(labels);
4448 235 : current.inc();
4449 :
4450 235 : if rx.changed().await.is_err() {
4451 : // tenant has been dropped
4452 7 : current.dec();
4453 7 : drop(BROKEN_TENANTS_SET.remove_label_values(set_key));
4454 7 : break;
4455 116 : }
4456 :
4457 116 : current.dec();
4458 116 : tuple = inspect_state(&rx.borrow_and_update());
4459 :
4460 116 : let is_broken = tuple.1;
4461 116 : if is_broken && !counted_broken {
4462 0 : counted_broken = true;
4463 0 : // insert the tenant_id (back) into the set while avoiding needless counter
4464 0 : // access
4465 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
4466 116 : }
4467 : }
4468 7 : });
4469 :
4470 119 : TenantShard {
4471 119 : tenant_shard_id,
4472 119 : shard_identity,
4473 119 : generation: attached_conf.location.generation,
4474 119 : conf,
4475 119 : // using now here is good enough approximation to catch tenants with really long
4476 119 : // activation times.
4477 119 : constructed_at: Instant::now(),
4478 119 : timelines: Mutex::new(HashMap::new()),
4479 119 : timelines_creating: Mutex::new(HashSet::new()),
4480 119 : timelines_offloaded: Mutex::new(HashMap::new()),
4481 119 : timelines_importing: Mutex::new(HashMap::new()),
4482 119 : remote_tenant_manifest: Default::default(),
4483 119 : gc_cs: tokio::sync::Mutex::new(()),
4484 119 : walredo_mgr,
4485 119 : remote_storage,
4486 119 : deletion_queue_client,
4487 119 : state,
4488 119 : cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
4489 119 : cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
4490 119 : eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
4491 119 : compaction_circuit_breaker: std::sync::Mutex::new(CircuitBreaker::new(
4492 119 : format!("compaction-{tenant_shard_id}"),
4493 119 : 5,
4494 119 : // Compaction can be a very expensive operation, and might leak disk space. It also ought
4495 119 : // to be infallible, as long as remote storage is available. So if it repeatedly fails,
4496 119 : // use an extremely long backoff.
4497 119 : Some(Duration::from_secs(3600 * 24)),
4498 119 : )),
4499 119 : l0_compaction_trigger: Arc::new(Notify::new()),
4500 119 : scheduled_compaction_tasks: Mutex::new(Default::default()),
4501 119 : activate_now_sem: tokio::sync::Semaphore::new(0),
4502 119 : attach_wal_lag_cooldown: Arc::new(std::sync::OnceLock::new()),
4503 119 : cancel: CancellationToken::default(),
4504 119 : gate: Gate::default(),
4505 119 : pagestream_throttle: Arc::new(throttle::Throttle::new(
4506 119 : TenantShard::get_pagestream_throttle_config(conf, &attached_conf.tenant_conf),
4507 119 : )),
4508 119 : pagestream_throttle_metrics: Arc::new(
4509 119 : crate::metrics::tenant_throttling::Pagestream::new(&tenant_shard_id),
4510 119 : ),
4511 119 : tenant_conf: Arc::new(ArcSwap::from_pointee(attached_conf)),
4512 119 : ongoing_timeline_detach: std::sync::Mutex::default(),
4513 119 : gc_block: Default::default(),
4514 119 : l0_flush_global_state,
4515 119 : basebackup_cache,
4516 119 : feature_resolver: Arc::new(TenantFeatureResolver::new(
4517 119 : feature_resolver,
4518 119 : tenant_shard_id.tenant_id,
4519 119 : )),
4520 119 : }
4521 119 : }
4522 :
4523 : /// Locate and load config
4524 0 : pub(super) fn load_tenant_config(
4525 0 : conf: &'static PageServerConf,
4526 0 : tenant_shard_id: &TenantShardId,
4527 0 : ) -> Result<LocationConf, LoadConfigError> {
4528 0 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4529 :
4530 0 : info!("loading tenant configuration from {config_path}");
4531 :
4532 : // load and parse file
4533 0 : let config = fs::read_to_string(&config_path).map_err(|e| {
4534 0 : match e.kind() {
4535 : std::io::ErrorKind::NotFound => {
4536 : // The config should almost always exist for a tenant directory:
4537 : // - When attaching a tenant, the config is the first thing we write
4538 : // - When detaching a tenant, we atomically move the directory to a tmp location
4539 : // before deleting contents.
4540 : //
4541 : // The very rare edge case that can result in a missing config is if we crash during attach
4542 : // between creating directory and writing config. Callers should handle that as if the
4543 : // directory didn't exist.
4544 :
4545 0 : LoadConfigError::NotFound(config_path)
4546 : }
4547 : _ => {
4548 : // No IO errors except NotFound are acceptable here: other kinds of error indicate local storage or permissions issues
4549 : // that we cannot cleanly recover
4550 0 : crate::virtual_file::on_fatal_io_error(&e, "Reading tenant config file")
4551 : }
4552 : }
4553 0 : })?;
4554 :
4555 0 : Ok(toml_edit::de::from_str::<LocationConf>(&config)?)
4556 0 : }
4557 :
4558 : /// Stores a tenant location config to disk.
4559 : ///
4560 : /// NB: make sure to call `ShardIdentity::assert_equal` before persisting a new config, to avoid
4561 : /// changes to shard parameters that may result in data corruption.
4562 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4563 : pub(super) async fn persist_tenant_config(
4564 : conf: &'static PageServerConf,
4565 : tenant_shard_id: &TenantShardId,
4566 : location_conf: &LocationConf,
4567 : ) -> std::io::Result<()> {
4568 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4569 :
4570 : Self::persist_tenant_config_at(tenant_shard_id, &config_path, location_conf).await
4571 : }
4572 :
4573 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4574 : pub(super) async fn persist_tenant_config_at(
4575 : tenant_shard_id: &TenantShardId,
4576 : config_path: &Utf8Path,
4577 : location_conf: &LocationConf,
4578 : ) -> std::io::Result<()> {
4579 : debug!("persisting tenantconf to {config_path}");
4580 :
4581 : let mut conf_content = r#"# This file contains a specific per-tenant's config.
4582 : # It is read in case of pageserver restart.
4583 : "#
4584 : .to_string();
4585 :
4586 0 : fail::fail_point!("tenant-config-before-write", |_| {
4587 0 : Err(std::io::Error::other("tenant-config-before-write"))
4588 0 : });
4589 :
4590 : // Convert the config to a toml file.
4591 : conf_content +=
4592 : &toml_edit::ser::to_string_pretty(&location_conf).expect("Config serialization failed");
4593 :
4594 : let temp_path = path_with_suffix_extension(config_path, TEMP_FILE_SUFFIX);
4595 :
4596 : let conf_content = conf_content.into_bytes();
4597 : VirtualFile::crashsafe_overwrite(config_path.to_owned(), temp_path, conf_content).await
4598 : }
4599 :
4600 : //
4601 : // How garbage collection works:
4602 : //
4603 : // +--bar------------->
4604 : // /
4605 : // +----+-----foo---------------->
4606 : // /
4607 : // ----main--+-------------------------->
4608 : // \
4609 : // +-----baz-------->
4610 : //
4611 : //
4612 : // 1. Grab 'gc_cs' mutex to prevent new timelines from being created while Timeline's
4613 : // `gc_infos` are being refreshed
4614 : // 2. Scan collected timelines, and on each timeline, make note of the
4615 : // all the points where other timelines have been branched off.
4616 : // We will refrain from removing page versions at those LSNs.
4617 : // 3. For each timeline, scan all layer files on the timeline.
4618 : // Remove all files for which a newer file exists and which
4619 : // don't cover any branch point LSNs.
4620 : //
4621 : // TODO:
4622 : // - if a relation has a non-incremental persistent layer on a child branch, then we
4623 : // don't need to keep that in the parent anymore. But currently
4624 : // we do.
4625 377 : async fn gc_iteration_internal(
4626 377 : &self,
4627 377 : target_timeline_id: Option<TimelineId>,
4628 377 : horizon: u64,
4629 377 : pitr: Duration,
4630 377 : cancel: &CancellationToken,
4631 377 : ctx: &RequestContext,
4632 377 : ) -> Result<GcResult, GcError> {
4633 377 : let mut totals: GcResult = Default::default();
4634 377 : let now = Instant::now();
4635 :
4636 377 : let gc_timelines = self
4637 377 : .refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4638 377 : .await?;
4639 :
4640 377 : failpoint_support::sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
4641 :
4642 : // If there is nothing to GC, we don't want any messages in the INFO log.
4643 377 : if !gc_timelines.is_empty() {
4644 377 : info!("{} timelines need GC", gc_timelines.len());
4645 : } else {
4646 0 : debug!("{} timelines need GC", gc_timelines.len());
4647 : }
4648 :
4649 : // Perform GC for each timeline.
4650 : //
4651 : // Note that we don't hold the `TenantShard::gc_cs` lock here because we don't want to delay the
4652 : // branch creation task, which requires the GC lock. A GC iteration can run concurrently
4653 : // with branch creation.
4654 : //
4655 : // See comments in [`TenantShard::branch_timeline`] for more information about why branch
4656 : // creation task can run concurrently with timeline's GC iteration.
4657 754 : for timeline in gc_timelines {
4658 377 : if cancel.is_cancelled() {
4659 : // We were requested to shut down. Stop and return with the progress we
4660 : // made.
4661 0 : break;
4662 377 : }
4663 377 : let result = match timeline.gc().await {
4664 : Err(GcError::TimelineCancelled) => {
4665 0 : if target_timeline_id.is_some() {
4666 : // If we were targetting this specific timeline, surface cancellation to caller
4667 0 : return Err(GcError::TimelineCancelled);
4668 : } else {
4669 : // A timeline may be shutting down independently of the tenant's lifecycle: we should
4670 : // skip past this and proceed to try GC on other timelines.
4671 0 : continue;
4672 : }
4673 : }
4674 377 : r => r?,
4675 : };
4676 377 : totals += result;
4677 : }
4678 :
4679 377 : totals.elapsed = now.elapsed();
4680 377 : Ok(totals)
4681 377 : }
4682 :
4683 : /// Refreshes the Timeline::gc_info for all timelines, returning the
4684 : /// vector of timelines which have [`Timeline::get_last_record_lsn`] past
4685 : /// [`TenantShard::get_gc_horizon`].
4686 : ///
4687 : /// This is usually executed as part of periodic gc, but can now be triggered more often.
4688 2 : pub(crate) async fn refresh_gc_info(
4689 2 : &self,
4690 2 : cancel: &CancellationToken,
4691 2 : ctx: &RequestContext,
4692 2 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4693 : // since this method can now be called at different rates than the configured gc loop, it
4694 : // might be that these configuration values get applied faster than what it was previously,
4695 : // since these were only read from the gc task.
4696 2 : let horizon = self.get_gc_horizon();
4697 2 : let pitr = self.get_pitr_interval();
4698 :
4699 : // refresh all timelines
4700 2 : let target_timeline_id = None;
4701 :
4702 2 : self.refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4703 2 : .await
4704 2 : }
4705 :
4706 : /// Populate all Timelines' `GcInfo` with information about their children. We do not set the
4707 : /// PITR cutoffs here, because that requires I/O: this is done later, before GC, by [`Self::refresh_gc_info_internal`]
4708 : ///
4709 : /// Subsequently, parent-child relationships are updated incrementally inside [`Timeline::new`] and [`Timeline::drop`].
4710 119 : fn initialize_gc_info(
4711 119 : &self,
4712 119 : timelines: &std::sync::MutexGuard<HashMap<TimelineId, Arc<Timeline>>>,
4713 119 : timelines_offloaded: &std::sync::MutexGuard<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
4714 119 : restrict_to_timeline: Option<TimelineId>,
4715 119 : ) {
4716 119 : if restrict_to_timeline.is_none() {
4717 : // This function must be called before activation: after activation timeline create/delete operations
4718 : // might happen, and this function is not safe to run concurrently with those.
4719 119 : assert!(!self.is_active());
4720 0 : }
4721 :
4722 : // Scan all timelines. For each timeline, remember the timeline ID and
4723 : // the branch point where it was created.
4724 119 : let mut all_branchpoints: BTreeMap<TimelineId, Vec<(Lsn, TimelineId, MaybeOffloaded)>> =
4725 119 : BTreeMap::new();
4726 119 : timelines.iter().for_each(|(timeline_id, timeline_entry)| {
4727 3 : if let Some(ancestor_timeline_id) = &timeline_entry.get_ancestor_timeline_id() {
4728 1 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4729 1 : ancestor_children.push((
4730 1 : timeline_entry.get_ancestor_lsn(),
4731 1 : *timeline_id,
4732 1 : MaybeOffloaded::No,
4733 1 : ));
4734 2 : }
4735 3 : });
4736 119 : timelines_offloaded
4737 119 : .iter()
4738 119 : .for_each(|(timeline_id, timeline_entry)| {
4739 0 : let Some(ancestor_timeline_id) = &timeline_entry.ancestor_timeline_id else {
4740 0 : return;
4741 : };
4742 0 : let Some(retain_lsn) = timeline_entry.ancestor_retain_lsn else {
4743 0 : return;
4744 : };
4745 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4746 0 : ancestor_children.push((retain_lsn, *timeline_id, MaybeOffloaded::Yes));
4747 0 : });
4748 :
4749 : // The number of bytes we always keep, irrespective of PITR: this is a constant across timelines
4750 119 : let horizon = self.get_gc_horizon();
4751 :
4752 : // Populate each timeline's GcInfo with information about its child branches
4753 119 : let timelines_to_write = if let Some(timeline_id) = restrict_to_timeline {
4754 0 : itertools::Either::Left(timelines.get(&timeline_id).into_iter())
4755 : } else {
4756 119 : itertools::Either::Right(timelines.values())
4757 : };
4758 122 : for timeline in timelines_to_write {
4759 3 : let mut branchpoints: Vec<(Lsn, TimelineId, MaybeOffloaded)> = all_branchpoints
4760 3 : .remove(&timeline.timeline_id)
4761 3 : .unwrap_or_default();
4762 :
4763 3 : branchpoints.sort_by_key(|b| b.0);
4764 :
4765 3 : let mut target = timeline.gc_info.write().unwrap();
4766 :
4767 3 : target.retain_lsns = branchpoints;
4768 :
4769 3 : let space_cutoff = timeline
4770 3 : .get_last_record_lsn()
4771 3 : .checked_sub(horizon)
4772 3 : .unwrap_or(Lsn(0));
4773 :
4774 3 : target.cutoffs = GcCutoffs {
4775 3 : space: space_cutoff,
4776 3 : time: None,
4777 3 : };
4778 : }
4779 119 : }
4780 :
4781 379 : async fn refresh_gc_info_internal(
4782 379 : &self,
4783 379 : target_timeline_id: Option<TimelineId>,
4784 379 : horizon: u64,
4785 379 : pitr: Duration,
4786 379 : cancel: &CancellationToken,
4787 379 : ctx: &RequestContext,
4788 379 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4789 : // before taking the gc_cs lock, do the heavier weight finding of gc_cutoff points for
4790 : // currently visible timelines.
4791 379 : let timelines = self
4792 379 : .timelines
4793 379 : .lock()
4794 379 : .unwrap()
4795 379 : .values()
4796 1663 : .filter(|tl| match target_timeline_id.as_ref() {
4797 1655 : Some(target) => &tl.timeline_id == target,
4798 8 : None => true,
4799 1663 : })
4800 379 : .cloned()
4801 379 : .collect::<Vec<_>>();
4802 :
4803 379 : if target_timeline_id.is_some() && timelines.is_empty() {
4804 : // We were to act on a particular timeline and it wasn't found
4805 0 : return Err(GcError::TimelineNotFound);
4806 379 : }
4807 :
4808 379 : let mut gc_cutoffs: HashMap<TimelineId, GcCutoffs> =
4809 379 : HashMap::with_capacity(timelines.len());
4810 :
4811 : // Ensures all timelines use the same start time when computing the time cutoff.
4812 379 : let now_ts_for_pitr_calc = SystemTime::now();
4813 385 : for timeline in timelines.iter() {
4814 385 : let ctx = &ctx.with_scope_timeline(timeline);
4815 385 : let cutoff = timeline
4816 385 : .get_last_record_lsn()
4817 385 : .checked_sub(horizon)
4818 385 : .unwrap_or(Lsn(0));
4819 :
4820 385 : let cutoffs = timeline
4821 385 : .find_gc_cutoffs(now_ts_for_pitr_calc, cutoff, pitr, cancel, ctx)
4822 385 : .await?;
4823 385 : let old = gc_cutoffs.insert(timeline.timeline_id, cutoffs);
4824 385 : assert!(old.is_none());
4825 : }
4826 :
4827 379 : if !self.is_active() || self.cancel.is_cancelled() {
4828 0 : return Err(GcError::TenantCancelled);
4829 379 : }
4830 :
4831 : // grab mutex to prevent new timelines from being created here; avoid doing long operations
4832 : // because that will stall branch creation.
4833 379 : let gc_cs = self.gc_cs.lock().await;
4834 :
4835 : // Ok, we now know all the branch points.
4836 : // Update the GC information for each timeline.
4837 379 : let mut gc_timelines = Vec::with_capacity(timelines.len());
4838 764 : for timeline in timelines {
4839 : // We filtered the timeline list above
4840 385 : if let Some(target_timeline_id) = target_timeline_id {
4841 377 : assert_eq!(target_timeline_id, timeline.timeline_id);
4842 8 : }
4843 :
4844 : {
4845 385 : let mut target = timeline.gc_info.write().unwrap();
4846 :
4847 : // Cull any expired leases
4848 385 : let now = SystemTime::now();
4849 385 : target.leases.retain(|_, lease| !lease.is_expired(&now));
4850 :
4851 385 : timeline
4852 385 : .metrics
4853 385 : .valid_lsn_lease_count_gauge
4854 385 : .set(target.leases.len() as u64);
4855 :
4856 : // Look up parent's PITR cutoff to update the child's knowledge of whether it is within parent's PITR
4857 385 : if let Some(ancestor_id) = timeline.get_ancestor_timeline_id() {
4858 56 : if let Some(ancestor_gc_cutoffs) = gc_cutoffs.get(&ancestor_id) {
4859 6 : target.within_ancestor_pitr =
4860 6 : Some(timeline.get_ancestor_lsn()) >= ancestor_gc_cutoffs.time;
4861 50 : }
4862 329 : }
4863 :
4864 : // Update metrics that depend on GC state
4865 385 : timeline
4866 385 : .metrics
4867 385 : .archival_size
4868 385 : .set(if target.within_ancestor_pitr {
4869 0 : timeline.metrics.current_logical_size_gauge.get()
4870 : } else {
4871 385 : 0
4872 : });
4873 385 : if let Some(time_cutoff) = target.cutoffs.time {
4874 319 : timeline.metrics.pitr_history_size.set(
4875 319 : timeline
4876 319 : .get_last_record_lsn()
4877 319 : .checked_sub(time_cutoff)
4878 319 : .unwrap_or_default()
4879 319 : .0,
4880 319 : );
4881 319 : }
4882 :
4883 : // Apply the cutoffs we found to the Timeline's GcInfo. Why might we _not_ have cutoffs for a timeline?
4884 : // - this timeline was created while we were finding cutoffs
4885 : // - lsn for timestamp search fails for this timeline repeatedly
4886 385 : if let Some(cutoffs) = gc_cutoffs.get(&timeline.timeline_id) {
4887 385 : let original_cutoffs = target.cutoffs.clone();
4888 : // GC cutoffs should never go back
4889 385 : target.cutoffs = GcCutoffs {
4890 385 : space: cutoffs.space.max(original_cutoffs.space),
4891 385 : time: cutoffs.time.max(original_cutoffs.time),
4892 385 : }
4893 0 : }
4894 : }
4895 :
4896 385 : gc_timelines.push(timeline);
4897 : }
4898 379 : drop(gc_cs);
4899 379 : Ok(gc_timelines)
4900 379 : }
4901 :
4902 : /// A substitute for `branch_timeline` for use in unit tests.
4903 : /// The returned timeline will have state value `Active` to make various `anyhow::ensure!()`
4904 : /// calls pass, but, we do not actually call `.activate()` under the hood. So, none of the
4905 : /// timeline background tasks are launched, except the flush loop.
4906 : #[cfg(test)]
4907 119 : async fn branch_timeline_test(
4908 119 : self: &Arc<Self>,
4909 119 : src_timeline: &Arc<Timeline>,
4910 119 : dst_id: TimelineId,
4911 119 : ancestor_lsn: Option<Lsn>,
4912 119 : ctx: &RequestContext,
4913 119 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
4914 119 : let tl = self
4915 119 : .branch_timeline_impl(src_timeline, dst_id, ancestor_lsn, ctx)
4916 119 : .await?
4917 117 : .into_timeline_for_test();
4918 117 : tl.set_state(TimelineState::Active);
4919 117 : Ok(tl)
4920 119 : }
4921 :
4922 : /// Helper for unit tests to branch a timeline with some pre-loaded states.
4923 : #[cfg(test)]
4924 : #[allow(clippy::too_many_arguments)]
4925 6 : pub async fn branch_timeline_test_with_layers(
4926 6 : self: &Arc<Self>,
4927 6 : src_timeline: &Arc<Timeline>,
4928 6 : dst_id: TimelineId,
4929 6 : ancestor_lsn: Option<Lsn>,
4930 6 : ctx: &RequestContext,
4931 6 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
4932 6 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
4933 6 : end_lsn: Lsn,
4934 6 : ) -> anyhow::Result<Arc<Timeline>> {
4935 : use checks::check_valid_layermap;
4936 : use itertools::Itertools;
4937 :
4938 6 : let tline = self
4939 6 : .branch_timeline_test(src_timeline, dst_id, ancestor_lsn, ctx)
4940 6 : .await?;
4941 6 : let ancestor_lsn = if let Some(ancestor_lsn) = ancestor_lsn {
4942 6 : ancestor_lsn
4943 : } else {
4944 0 : tline.get_last_record_lsn()
4945 : };
4946 6 : assert!(end_lsn >= ancestor_lsn);
4947 6 : tline.force_advance_lsn(end_lsn);
4948 9 : for deltas in delta_layer_desc {
4949 3 : tline
4950 3 : .force_create_delta_layer(deltas, Some(ancestor_lsn), ctx)
4951 3 : .await?;
4952 : }
4953 8 : for (lsn, images) in image_layer_desc {
4954 2 : tline
4955 2 : .force_create_image_layer(lsn, images, Some(ancestor_lsn), ctx)
4956 2 : .await?;
4957 : }
4958 6 : let layer_names = tline
4959 6 : .layers
4960 6 : .read(LayerManagerLockHolder::Testing)
4961 6 : .await
4962 6 : .layer_map()
4963 6 : .unwrap()
4964 6 : .iter_historic_layers()
4965 6 : .map(|layer| layer.layer_name())
4966 6 : .collect_vec();
4967 6 : if let Some(err) = check_valid_layermap(&layer_names) {
4968 0 : bail!("invalid layermap: {err}");
4969 6 : }
4970 6 : Ok(tline)
4971 6 : }
4972 :
4973 : /// Branch an existing timeline.
4974 0 : async fn branch_timeline(
4975 0 : self: &Arc<Self>,
4976 0 : src_timeline: &Arc<Timeline>,
4977 0 : dst_id: TimelineId,
4978 0 : start_lsn: Option<Lsn>,
4979 0 : ctx: &RequestContext,
4980 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4981 0 : self.branch_timeline_impl(src_timeline, dst_id, start_lsn, ctx)
4982 0 : .await
4983 0 : }
4984 :
4985 119 : async fn branch_timeline_impl(
4986 119 : self: &Arc<Self>,
4987 119 : src_timeline: &Arc<Timeline>,
4988 119 : dst_id: TimelineId,
4989 119 : start_lsn: Option<Lsn>,
4990 119 : ctx: &RequestContext,
4991 119 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4992 119 : let src_id = src_timeline.timeline_id;
4993 :
4994 : // We will validate our ancestor LSN in this function. Acquire the GC lock so that
4995 : // this check cannot race with GC, and the ancestor LSN is guaranteed to remain
4996 : // valid while we are creating the branch.
4997 119 : let _gc_cs = self.gc_cs.lock().await;
4998 :
4999 : // If no start LSN is specified, we branch the new timeline from the source timeline's last record LSN
5000 119 : let start_lsn = start_lsn.unwrap_or_else(|| {
5001 1 : let lsn = src_timeline.get_last_record_lsn();
5002 1 : info!("branching timeline {dst_id} from timeline {src_id} at last record LSN: {lsn}");
5003 1 : lsn
5004 1 : });
5005 :
5006 : // we finally have determined the ancestor_start_lsn, so we can get claim exclusivity now
5007 119 : let timeline_create_guard = match self
5008 119 : .start_creating_timeline(
5009 119 : dst_id,
5010 119 : CreateTimelineIdempotency::Branch {
5011 119 : ancestor_timeline_id: src_timeline.timeline_id,
5012 119 : ancestor_start_lsn: start_lsn,
5013 119 : },
5014 119 : )
5015 119 : .await?
5016 : {
5017 119 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
5018 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
5019 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
5020 : }
5021 : };
5022 :
5023 : // Ensure that `start_lsn` is valid, i.e. the LSN is within the PITR
5024 : // horizon on the source timeline
5025 : //
5026 : // We check it against both the planned GC cutoff stored in 'gc_info',
5027 : // and the 'latest_gc_cutoff' of the last GC that was performed. The
5028 : // planned GC cutoff in 'gc_info' is normally larger than
5029 : // 'applied_gc_cutoff_lsn', but beware of corner cases like if you just
5030 : // changed the GC settings for the tenant to make the PITR window
5031 : // larger, but some of the data was already removed by an earlier GC
5032 : // iteration.
5033 :
5034 : // check against last actual 'latest_gc_cutoff' first
5035 119 : let applied_gc_cutoff_lsn = src_timeline.get_applied_gc_cutoff_lsn();
5036 : {
5037 119 : let gc_info = src_timeline.gc_info.read().unwrap();
5038 119 : let planned_cutoff = gc_info.min_cutoff();
5039 119 : if gc_info.lsn_covered_by_lease(start_lsn) {
5040 0 : tracing::info!(
5041 0 : "skipping comparison of {start_lsn} with gc cutoff {} and planned gc cutoff {planned_cutoff} due to lsn lease",
5042 0 : *applied_gc_cutoff_lsn
5043 : );
5044 : } else {
5045 119 : src_timeline
5046 119 : .check_lsn_is_in_scope(start_lsn, &applied_gc_cutoff_lsn)
5047 119 : .context(format!(
5048 119 : "invalid branch start lsn: less than latest GC cutoff {}",
5049 119 : *applied_gc_cutoff_lsn,
5050 : ))
5051 119 : .map_err(CreateTimelineError::AncestorLsn)?;
5052 :
5053 : // and then the planned GC cutoff
5054 117 : if start_lsn < planned_cutoff {
5055 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
5056 0 : "invalid branch start lsn: less than planned GC cutoff {planned_cutoff}"
5057 0 : )));
5058 117 : }
5059 : }
5060 : }
5061 :
5062 : //
5063 : // The branch point is valid, and we are still holding the 'gc_cs' lock
5064 : // so that GC cannot advance the GC cutoff until we are finished.
5065 : // Proceed with the branch creation.
5066 : //
5067 :
5068 : // Determine prev-LSN for the new timeline. We can only determine it if
5069 : // the timeline was branched at the current end of the source timeline.
5070 : let RecordLsn {
5071 117 : last: src_last,
5072 117 : prev: src_prev,
5073 117 : } = src_timeline.get_last_record_rlsn();
5074 117 : let dst_prev = if src_last == start_lsn {
5075 108 : Some(src_prev)
5076 : } else {
5077 9 : None
5078 : };
5079 :
5080 : // Create the metadata file, noting the ancestor of the new timeline.
5081 : // There is initially no data in it, but all the read-calls know to look
5082 : // into the ancestor.
5083 117 : let metadata = TimelineMetadata::new(
5084 117 : start_lsn,
5085 117 : dst_prev,
5086 117 : Some(src_id),
5087 117 : start_lsn,
5088 117 : *src_timeline.applied_gc_cutoff_lsn.read(), // FIXME: should we hold onto this guard longer?
5089 117 : src_timeline.initdb_lsn,
5090 117 : src_timeline.pg_version,
5091 : );
5092 :
5093 117 : let (rel_size_v2_status, rel_size_migrated_at) = src_timeline.get_rel_size_v2_status();
5094 117 : let (uninitialized_timeline, _timeline_ctx) = self
5095 117 : .prepare_new_timeline(
5096 117 : dst_id,
5097 117 : &metadata,
5098 117 : timeline_create_guard,
5099 117 : start_lsn + 1,
5100 117 : Some(Arc::clone(src_timeline)),
5101 117 : Some(rel_size_v2_status),
5102 117 : rel_size_migrated_at,
5103 117 : ctx,
5104 117 : )
5105 117 : .await?;
5106 :
5107 117 : let new_timeline = uninitialized_timeline.finish_creation().await?;
5108 :
5109 : // Root timeline gets its layers during creation and uploads them along with the metadata.
5110 : // A branch timeline though, when created, can get no writes for some time, hence won't get any layers created.
5111 : // We still need to upload its metadata eagerly: if other nodes `attach` the tenant and miss this timeline, their GC
5112 : // could get incorrect information and remove more layers, than needed.
5113 : // See also https://github.com/neondatabase/neon/issues/3865
5114 117 : new_timeline
5115 117 : .remote_client
5116 117 : .schedule_index_upload_for_full_metadata_update(&metadata)
5117 117 : .context("branch initial metadata upload")?;
5118 :
5119 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
5120 :
5121 117 : Ok(CreateTimelineResult::Created(new_timeline))
5122 119 : }
5123 :
5124 : /// For unit tests, make this visible so that other modules can directly create timelines
5125 : #[cfg(test)]
5126 : #[tracing::instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
5127 : pub(crate) async fn bootstrap_timeline_test(
5128 : self: &Arc<Self>,
5129 : timeline_id: TimelineId,
5130 : pg_version: PgMajorVersion,
5131 : load_existing_initdb: Option<TimelineId>,
5132 : ctx: &RequestContext,
5133 : ) -> anyhow::Result<Arc<Timeline>> {
5134 : self.bootstrap_timeline(timeline_id, pg_version, load_existing_initdb, ctx)
5135 : .await
5136 : .map_err(anyhow::Error::new)
5137 1 : .map(|r| r.into_timeline_for_test())
5138 : }
5139 :
5140 : /// Get exclusive access to the timeline ID for creation.
5141 : ///
5142 : /// Timeline-creating code paths must use this function before making changes
5143 : /// to in-memory or persistent state.
5144 : ///
5145 : /// The `state` parameter is a description of the timeline creation operation
5146 : /// we intend to perform.
5147 : /// If the timeline was already created in the meantime, we check whether this
5148 : /// request conflicts or is idempotent , based on `state`.
5149 235 : async fn start_creating_timeline(
5150 235 : self: &Arc<Self>,
5151 235 : new_timeline_id: TimelineId,
5152 235 : idempotency: CreateTimelineIdempotency,
5153 235 : ) -> Result<StartCreatingTimelineResult, CreateTimelineError> {
5154 235 : let allow_offloaded = false;
5155 235 : match self.create_timeline_create_guard(new_timeline_id, idempotency, allow_offloaded) {
5156 234 : Ok(create_guard) => {
5157 234 : pausable_failpoint!("timeline-creation-after-uninit");
5158 234 : Ok(StartCreatingTimelineResult::CreateGuard(create_guard))
5159 : }
5160 0 : Err(TimelineExclusionError::ShuttingDown) => Err(CreateTimelineError::ShuttingDown),
5161 : Err(TimelineExclusionError::AlreadyCreating) => {
5162 : // Creation is in progress, we cannot create it again, and we cannot
5163 : // check if this request matches the existing one, so caller must try
5164 : // again later.
5165 0 : Err(CreateTimelineError::AlreadyCreating)
5166 : }
5167 0 : Err(TimelineExclusionError::Other(e)) => Err(CreateTimelineError::Other(e)),
5168 : Err(TimelineExclusionError::AlreadyExists {
5169 0 : existing: TimelineOrOffloaded::Offloaded(_existing),
5170 : ..
5171 : }) => {
5172 0 : info!("timeline already exists but is offloaded");
5173 0 : Err(CreateTimelineError::Conflict)
5174 : }
5175 : Err(TimelineExclusionError::AlreadyExists {
5176 0 : existing: TimelineOrOffloaded::Importing(_existing),
5177 : ..
5178 : }) => {
5179 : // If there's a timeline already importing, then we would hit
5180 : // the [`TimelineExclusionError::AlreadyCreating`] branch above.
5181 0 : unreachable!("Importing timelines hold the creation guard")
5182 : }
5183 : Err(TimelineExclusionError::AlreadyExists {
5184 1 : existing: TimelineOrOffloaded::Timeline(existing),
5185 1 : arg,
5186 : }) => {
5187 : {
5188 1 : let existing = &existing.create_idempotency;
5189 1 : let _span = info_span!("idempotency_check", ?existing, ?arg).entered();
5190 1 : debug!("timeline already exists");
5191 :
5192 1 : match (existing, &arg) {
5193 : // FailWithConflict => no idempotency check
5194 : (CreateTimelineIdempotency::FailWithConflict, _)
5195 : | (_, CreateTimelineIdempotency::FailWithConflict) => {
5196 1 : warn!("timeline already exists, failing request");
5197 1 : return Err(CreateTimelineError::Conflict);
5198 : }
5199 : // Idempotent <=> CreateTimelineIdempotency is identical
5200 0 : (x, y) if x == y => {
5201 0 : info!(
5202 0 : "timeline already exists and idempotency matches, succeeding request"
5203 : );
5204 : // fallthrough
5205 : }
5206 : (_, _) => {
5207 0 : warn!("idempotency conflict, failing request");
5208 0 : return Err(CreateTimelineError::Conflict);
5209 : }
5210 : }
5211 : }
5212 :
5213 0 : Ok(StartCreatingTimelineResult::Idempotent(existing))
5214 : }
5215 : }
5216 235 : }
5217 :
5218 0 : async fn upload_initdb(
5219 0 : &self,
5220 0 : timelines_path: &Utf8PathBuf,
5221 0 : pgdata_path: &Utf8PathBuf,
5222 0 : timeline_id: &TimelineId,
5223 0 : ) -> anyhow::Result<()> {
5224 0 : let temp_path = timelines_path.join(format!(
5225 0 : "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
5226 0 : ));
5227 :
5228 0 : scopeguard::defer! {
5229 : if let Err(e) = fs::remove_file(&temp_path) {
5230 : error!("Failed to remove temporary initdb archive '{temp_path}': {e}");
5231 : }
5232 : }
5233 :
5234 0 : let (pgdata_zstd, tar_zst_size) = create_zst_tarball(pgdata_path, &temp_path).await?;
5235 : const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
5236 0 : if tar_zst_size > INITDB_TAR_ZST_WARN_LIMIT {
5237 0 : warn!(
5238 0 : "compressed {temp_path} size of {tar_zst_size} is above limit {INITDB_TAR_ZST_WARN_LIMIT}."
5239 : );
5240 0 : }
5241 :
5242 0 : pausable_failpoint!("before-initdb-upload");
5243 :
5244 0 : backoff::retry(
5245 0 : || async {
5246 0 : self::remote_timeline_client::upload_initdb_dir(
5247 0 : &self.remote_storage,
5248 0 : &self.tenant_shard_id.tenant_id,
5249 0 : timeline_id,
5250 0 : pgdata_zstd.try_clone().await?,
5251 0 : tar_zst_size,
5252 0 : &self.cancel,
5253 : )
5254 0 : .await
5255 0 : },
5256 : |_| false,
5257 : 3,
5258 : u32::MAX,
5259 0 : "persist_initdb_tar_zst",
5260 0 : &self.cancel,
5261 : )
5262 0 : .await
5263 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
5264 0 : .and_then(|x| x)
5265 0 : }
5266 :
5267 : /// - run initdb to init temporary instance and get bootstrap data
5268 : /// - after initialization completes, tar up the temp dir and upload it to S3.
5269 1 : async fn bootstrap_timeline(
5270 1 : self: &Arc<Self>,
5271 1 : timeline_id: TimelineId,
5272 1 : pg_version: PgMajorVersion,
5273 1 : load_existing_initdb: Option<TimelineId>,
5274 1 : ctx: &RequestContext,
5275 1 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
5276 1 : let timeline_create_guard = match self
5277 1 : .start_creating_timeline(
5278 1 : timeline_id,
5279 1 : CreateTimelineIdempotency::Bootstrap { pg_version },
5280 1 : )
5281 1 : .await?
5282 : {
5283 1 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
5284 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
5285 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
5286 : }
5287 : };
5288 :
5289 : // create a `tenant/{tenant_id}/timelines/basebackup-{timeline_id}.{TEMP_FILE_SUFFIX}/`
5290 : // temporary directory for basebackup files for the given timeline.
5291 :
5292 1 : let timelines_path = self.conf.timelines_path(&self.tenant_shard_id);
5293 1 : let pgdata_path = path_with_suffix_extension(
5294 1 : timelines_path.join(format!("basebackup-{timeline_id}")),
5295 1 : TEMP_FILE_SUFFIX,
5296 : );
5297 :
5298 : // Remove whatever was left from the previous runs: safe because TimelineCreateGuard guarantees
5299 : // we won't race with other creations or existent timelines with the same path.
5300 1 : if pgdata_path.exists() {
5301 0 : fs::remove_dir_all(&pgdata_path).with_context(|| {
5302 0 : format!("Failed to remove already existing initdb directory: {pgdata_path}")
5303 0 : })?;
5304 0 : tracing::info!("removed previous attempt's temporary initdb directory '{pgdata_path}'");
5305 1 : }
5306 :
5307 : // this new directory is very temporary, set to remove it immediately after bootstrap, we don't need it
5308 1 : let pgdata_path_deferred = pgdata_path.clone();
5309 1 : scopeguard::defer! {
5310 : if let Err(e) = fs::remove_dir_all(&pgdata_path_deferred).or_else(fs_ext::ignore_not_found) {
5311 : // this is unlikely, but we will remove the directory on pageserver restart or another bootstrap call
5312 : error!("Failed to remove temporary initdb directory '{pgdata_path_deferred}': {e}");
5313 : } else {
5314 : tracing::info!("removed temporary initdb directory '{pgdata_path_deferred}'");
5315 : }
5316 : }
5317 1 : if let Some(existing_initdb_timeline_id) = load_existing_initdb {
5318 1 : if existing_initdb_timeline_id != timeline_id {
5319 0 : let source_path = &remote_initdb_archive_path(
5320 0 : &self.tenant_shard_id.tenant_id,
5321 0 : &existing_initdb_timeline_id,
5322 0 : );
5323 0 : let dest_path =
5324 0 : &remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &timeline_id);
5325 :
5326 : // if this fails, it will get retried by retried control plane requests
5327 0 : self.remote_storage
5328 0 : .copy_object(source_path, dest_path, &self.cancel)
5329 0 : .await
5330 0 : .context("copy initdb tar")?;
5331 1 : }
5332 1 : let (initdb_tar_zst_path, initdb_tar_zst) =
5333 1 : self::remote_timeline_client::download_initdb_tar_zst(
5334 1 : self.conf,
5335 1 : &self.remote_storage,
5336 1 : &self.tenant_shard_id,
5337 1 : &existing_initdb_timeline_id,
5338 1 : &self.cancel,
5339 1 : )
5340 1 : .await
5341 1 : .context("download initdb tar")?;
5342 :
5343 1 : scopeguard::defer! {
5344 : if let Err(e) = fs::remove_file(&initdb_tar_zst_path) {
5345 : error!("Failed to remove temporary initdb archive '{initdb_tar_zst_path}': {e}");
5346 : }
5347 : }
5348 :
5349 1 : let buf_read =
5350 1 : BufReader::with_capacity(remote_timeline_client::BUFFER_SIZE, initdb_tar_zst);
5351 1 : extract_zst_tarball(&pgdata_path, buf_read)
5352 1 : .await
5353 1 : .context("extract initdb tar")?;
5354 : } else {
5355 : // Init temporarily repo to get bootstrap data, this creates a directory in the `pgdata_path` path
5356 0 : run_initdb(self.conf, &pgdata_path, pg_version, &self.cancel)
5357 0 : .await
5358 0 : .context("run initdb")?;
5359 :
5360 : // Upload the created data dir to S3
5361 0 : if self.tenant_shard_id().is_shard_zero() {
5362 0 : self.upload_initdb(&timelines_path, &pgdata_path, &timeline_id)
5363 0 : .await?;
5364 0 : }
5365 : }
5366 1 : let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
5367 :
5368 : // Import the contents of the data directory at the initial checkpoint
5369 : // LSN, and any WAL after that.
5370 : // Initdb lsn will be equal to last_record_lsn which will be set after import.
5371 : // Because we know it upfront avoid having an option or dummy zero value by passing it to the metadata.
5372 1 : let new_metadata = TimelineMetadata::new(
5373 1 : Lsn(0),
5374 1 : None,
5375 1 : None,
5376 1 : Lsn(0),
5377 1 : pgdata_lsn,
5378 1 : pgdata_lsn,
5379 1 : pg_version,
5380 : );
5381 1 : let (mut raw_timeline, timeline_ctx) = self
5382 1 : .prepare_new_timeline(
5383 1 : timeline_id,
5384 1 : &new_metadata,
5385 1 : timeline_create_guard,
5386 1 : pgdata_lsn,
5387 1 : None,
5388 1 : None,
5389 1 : None,
5390 1 : ctx,
5391 1 : )
5392 1 : .await?;
5393 :
5394 1 : let tenant_shard_id = raw_timeline.owning_tenant.tenant_shard_id;
5395 1 : raw_timeline
5396 1 : .write(|unfinished_timeline| async move {
5397 1 : import_datadir::import_timeline_from_postgres_datadir(
5398 1 : &unfinished_timeline,
5399 1 : &pgdata_path,
5400 1 : pgdata_lsn,
5401 1 : &timeline_ctx,
5402 1 : )
5403 1 : .await
5404 1 : .with_context(|| {
5405 0 : format!(
5406 0 : "Failed to import pgdatadir for timeline {tenant_shard_id}/{timeline_id}"
5407 : )
5408 0 : })?;
5409 :
5410 1 : fail::fail_point!("before-checkpoint-new-timeline", |_| {
5411 0 : Err(CreateTimelineError::Other(anyhow::anyhow!(
5412 0 : "failpoint before-checkpoint-new-timeline"
5413 0 : )))
5414 0 : });
5415 :
5416 1 : Ok(())
5417 2 : })
5418 1 : .await?;
5419 :
5420 : // All done!
5421 1 : let timeline = raw_timeline.finish_creation().await?;
5422 :
5423 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
5424 :
5425 1 : Ok(CreateTimelineResult::Created(timeline))
5426 1 : }
5427 :
5428 232 : fn build_timeline_remote_client(&self, timeline_id: TimelineId) -> RemoteTimelineClient {
5429 232 : RemoteTimelineClient::new(
5430 232 : self.remote_storage.clone(),
5431 232 : self.deletion_queue_client.clone(),
5432 232 : self.conf,
5433 232 : self.tenant_shard_id,
5434 232 : timeline_id,
5435 232 : self.generation,
5436 232 : &self.tenant_conf.load().location,
5437 : )
5438 232 : }
5439 :
5440 : /// Builds required resources for a new timeline.
5441 232 : fn build_timeline_resources(&self, timeline_id: TimelineId) -> TimelineResources {
5442 232 : let remote_client = self.build_timeline_remote_client(timeline_id);
5443 232 : self.get_timeline_resources_for(remote_client)
5444 232 : }
5445 :
5446 : /// Builds timeline resources for the given remote client.
5447 235 : fn get_timeline_resources_for(&self, remote_client: RemoteTimelineClient) -> TimelineResources {
5448 235 : TimelineResources {
5449 235 : remote_client,
5450 235 : pagestream_throttle: self.pagestream_throttle.clone(),
5451 235 : pagestream_throttle_metrics: self.pagestream_throttle_metrics.clone(),
5452 235 : l0_compaction_trigger: self.l0_compaction_trigger.clone(),
5453 235 : l0_flush_global_state: self.l0_flush_global_state.clone(),
5454 235 : basebackup_cache: self.basebackup_cache.clone(),
5455 235 : feature_resolver: self.feature_resolver.clone(),
5456 235 : }
5457 235 : }
5458 :
5459 : /// Creates intermediate timeline structure and its files.
5460 : ///
5461 : /// An empty layer map is initialized, and new data and WAL can be imported starting
5462 : /// at 'disk_consistent_lsn'. After any initial data has been imported, call
5463 : /// `finish_creation` to insert the Timeline into the timelines map.
5464 : #[allow(clippy::too_many_arguments)]
5465 232 : async fn prepare_new_timeline<'a>(
5466 232 : &'a self,
5467 232 : new_timeline_id: TimelineId,
5468 232 : new_metadata: &TimelineMetadata,
5469 232 : create_guard: TimelineCreateGuard,
5470 232 : start_lsn: Lsn,
5471 232 : ancestor: Option<Arc<Timeline>>,
5472 232 : rel_size_v2_status: Option<RelSizeMigration>,
5473 232 : rel_size_migrated_at: Option<Lsn>,
5474 232 : ctx: &RequestContext,
5475 232 : ) -> anyhow::Result<(UninitializedTimeline<'a>, RequestContext)> {
5476 232 : let tenant_shard_id = self.tenant_shard_id;
5477 :
5478 232 : let resources = self.build_timeline_resources(new_timeline_id);
5479 232 : resources.remote_client.init_upload_queue_for_empty_remote(
5480 232 : new_metadata,
5481 232 : rel_size_v2_status.clone(),
5482 232 : rel_size_migrated_at,
5483 0 : )?;
5484 :
5485 232 : let (timeline_struct, timeline_ctx) = self
5486 232 : .create_timeline_struct(
5487 232 : new_timeline_id,
5488 232 : new_metadata,
5489 232 : None,
5490 232 : ancestor,
5491 232 : resources,
5492 232 : CreateTimelineCause::Load,
5493 232 : create_guard.idempotency.clone(),
5494 232 : None,
5495 232 : rel_size_v2_status,
5496 232 : rel_size_migrated_at,
5497 232 : ctx,
5498 : )
5499 232 : .context("Failed to create timeline data structure")?;
5500 :
5501 232 : timeline_struct.init_empty_layer_map(start_lsn);
5502 :
5503 232 : if let Err(e) = self
5504 232 : .create_timeline_files(&create_guard.timeline_path)
5505 232 : .await
5506 : {
5507 0 : error!(
5508 0 : "Failed to create initial files for timeline {tenant_shard_id}/{new_timeline_id}, cleaning up: {e:?}"
5509 : );
5510 0 : cleanup_timeline_directory(create_guard);
5511 0 : return Err(e);
5512 232 : }
5513 :
5514 232 : debug!(
5515 0 : "Successfully created initial files for timeline {tenant_shard_id}/{new_timeline_id}"
5516 : );
5517 :
5518 232 : Ok((
5519 232 : UninitializedTimeline::new(
5520 232 : self,
5521 232 : new_timeline_id,
5522 232 : Some((timeline_struct, create_guard)),
5523 232 : ),
5524 232 : timeline_ctx,
5525 232 : ))
5526 232 : }
5527 :
5528 232 : async fn create_timeline_files(&self, timeline_path: &Utf8Path) -> anyhow::Result<()> {
5529 232 : crashsafe::create_dir(timeline_path).context("Failed to create timeline directory")?;
5530 :
5531 232 : fail::fail_point!("after-timeline-dir-creation", |_| {
5532 0 : anyhow::bail!("failpoint after-timeline-dir-creation");
5533 0 : });
5534 :
5535 232 : Ok(())
5536 232 : }
5537 :
5538 : /// Get a guard that provides exclusive access to the timeline directory, preventing
5539 : /// concurrent attempts to create the same timeline.
5540 : ///
5541 : /// The `allow_offloaded` parameter controls whether to tolerate the existence of
5542 : /// offloaded timelines or not.
5543 235 : fn create_timeline_create_guard(
5544 235 : self: &Arc<Self>,
5545 235 : timeline_id: TimelineId,
5546 235 : idempotency: CreateTimelineIdempotency,
5547 235 : allow_offloaded: bool,
5548 235 : ) -> Result<TimelineCreateGuard, TimelineExclusionError> {
5549 235 : let tenant_shard_id = self.tenant_shard_id;
5550 :
5551 235 : let timeline_path = self.conf.timeline_path(&tenant_shard_id, &timeline_id);
5552 :
5553 235 : let create_guard = TimelineCreateGuard::new(
5554 235 : self,
5555 235 : timeline_id,
5556 235 : timeline_path.clone(),
5557 235 : idempotency,
5558 235 : allow_offloaded,
5559 1 : )?;
5560 :
5561 : // At this stage, we have got exclusive access to in-memory state for this timeline ID
5562 : // for creation.
5563 : // A timeline directory should never exist on disk already:
5564 : // - a previous failed creation would have cleaned up after itself
5565 : // - a pageserver restart would clean up timeline directories that don't have valid remote state
5566 : //
5567 : // Therefore it is an unexpected internal error to encounter a timeline directory already existing here,
5568 : // this error may indicate a bug in cleanup on failed creations.
5569 234 : if timeline_path.exists() {
5570 0 : return Err(TimelineExclusionError::Other(anyhow::anyhow!(
5571 0 : "Timeline directory already exists! This is a bug."
5572 0 : )));
5573 234 : }
5574 :
5575 234 : Ok(create_guard)
5576 235 : }
5577 :
5578 : /// Gathers inputs from all of the timelines to produce a sizing model input.
5579 : ///
5580 : /// Future is cancellation safe. Only one calculation can be running at once per tenant.
5581 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5582 : pub async fn gather_size_inputs(
5583 : &self,
5584 : // `max_retention_period` overrides the cutoff that is used to calculate the size
5585 : // (only if it is shorter than the real cutoff).
5586 : max_retention_period: Option<u64>,
5587 : cause: LogicalSizeCalculationCause,
5588 : cancel: &CancellationToken,
5589 : ctx: &RequestContext,
5590 : ) -> Result<size::ModelInputs, size::CalculateSyntheticSizeError> {
5591 : let logical_sizes_at_once = self
5592 : .conf
5593 : .concurrent_tenant_size_logical_size_queries
5594 : .inner();
5595 :
5596 : // TODO: Having a single mutex block concurrent reads is not great for performance.
5597 : //
5598 : // But the only case where we need to run multiple of these at once is when we
5599 : // request a size for a tenant manually via API, while another background calculation
5600 : // is in progress (which is not a common case).
5601 : //
5602 : // See more for on the issue #2748 condenced out of the initial PR review.
5603 : let mut shared_cache = tokio::select! {
5604 : locked = self.cached_logical_sizes.lock() => locked,
5605 : _ = cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5606 : _ = self.cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5607 : };
5608 :
5609 : size::gather_inputs(
5610 : self,
5611 : logical_sizes_at_once,
5612 : max_retention_period,
5613 : &mut shared_cache,
5614 : cause,
5615 : cancel,
5616 : ctx,
5617 : )
5618 : .await
5619 : }
5620 :
5621 : /// Calculate synthetic tenant size and cache the result.
5622 : /// This is periodically called by background worker.
5623 : /// result is cached in tenant struct
5624 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5625 : pub async fn calculate_synthetic_size(
5626 : &self,
5627 : cause: LogicalSizeCalculationCause,
5628 : cancel: &CancellationToken,
5629 : ctx: &RequestContext,
5630 : ) -> Result<u64, size::CalculateSyntheticSizeError> {
5631 : let inputs = self.gather_size_inputs(None, cause, cancel, ctx).await?;
5632 :
5633 : let size = inputs.calculate();
5634 :
5635 : self.set_cached_synthetic_size(size);
5636 :
5637 : Ok(size)
5638 : }
5639 :
5640 : /// Cache given synthetic size and update the metric value
5641 0 : pub fn set_cached_synthetic_size(&self, size: u64) {
5642 0 : self.cached_synthetic_tenant_size
5643 0 : .store(size, Ordering::Relaxed);
5644 :
5645 : // Only shard zero should be calculating synthetic sizes
5646 0 : debug_assert!(self.shard_identity.is_shard_zero());
5647 :
5648 0 : TENANT_SYNTHETIC_SIZE_METRIC
5649 0 : .get_metric_with_label_values(&[&self.tenant_shard_id.tenant_id.to_string()])
5650 0 : .unwrap()
5651 0 : .set(size);
5652 0 : }
5653 :
5654 0 : pub fn cached_synthetic_size(&self) -> u64 {
5655 0 : self.cached_synthetic_tenant_size.load(Ordering::Relaxed)
5656 0 : }
5657 :
5658 : /// Flush any in-progress layers, schedule uploads, and wait for uploads to complete.
5659 : ///
5660 : /// This function can take a long time: callers should wrap it in a timeout if calling
5661 : /// from an external API handler.
5662 : ///
5663 : /// Cancel-safety: cancelling this function may leave I/O running, but such I/O is
5664 : /// still bounded by tenant/timeline shutdown.
5665 : #[tracing::instrument(skip_all)]
5666 : pub(crate) async fn flush_remote(&self) -> anyhow::Result<()> {
5667 : let timelines = self.timelines.lock().unwrap().clone();
5668 :
5669 0 : async fn flush_timeline(_gate: GateGuard, timeline: Arc<Timeline>) -> anyhow::Result<()> {
5670 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Flushing...");
5671 0 : timeline.freeze_and_flush().await?;
5672 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Waiting for uploads...");
5673 0 : timeline.remote_client.wait_completion().await?;
5674 :
5675 0 : Ok(())
5676 0 : }
5677 :
5678 : // We do not use a JoinSet for these tasks, because we don't want them to be
5679 : // aborted when this function's future is cancelled: they should stay alive
5680 : // holding their GateGuard until they complete, to ensure their I/Os complete
5681 : // before Timeline shutdown completes.
5682 : let mut results = FuturesUnordered::new();
5683 :
5684 : for (_timeline_id, timeline) in timelines {
5685 : // Run each timeline's flush in a task holding the timeline's gate: this
5686 : // means that if this function's future is cancelled, the Timeline shutdown
5687 : // will still wait for any I/O in here to complete.
5688 : let Ok(gate) = timeline.gate.enter() else {
5689 : continue;
5690 : };
5691 0 : let jh = tokio::task::spawn(async move { flush_timeline(gate, timeline).await });
5692 : results.push(jh);
5693 : }
5694 :
5695 : while let Some(r) = results.next().await {
5696 : if let Err(e) = r {
5697 : if !e.is_cancelled() && !e.is_panic() {
5698 : tracing::error!("unexpected join error: {e:?}");
5699 : }
5700 : }
5701 : }
5702 :
5703 : // The flushes we did above were just writes, but the TenantShard might have had
5704 : // pending deletions as well from recent compaction/gc: we want to flush those
5705 : // as well. This requires flushing the global delete queue. This is cheap
5706 : // because it's typically a no-op.
5707 : match self.deletion_queue_client.flush_execute().await {
5708 : Ok(_) => {}
5709 : Err(DeletionQueueError::ShuttingDown) => {}
5710 : }
5711 :
5712 : Ok(())
5713 : }
5714 :
5715 0 : pub(crate) fn get_tenant_conf(&self) -> pageserver_api::models::TenantConfig {
5716 0 : self.tenant_conf.load().tenant_conf.clone()
5717 0 : }
5718 :
5719 : /// How much local storage would this tenant like to have? It can cope with
5720 : /// less than this (via eviction and on-demand downloads), but this function enables
5721 : /// the TenantShard to advertise how much storage it would prefer to have to provide fast I/O
5722 : /// by keeping important things on local disk.
5723 : ///
5724 : /// This is a heuristic, not a guarantee: tenants that are long-idle will actually use less
5725 : /// than they report here, due to layer eviction. Tenants with many active branches may
5726 : /// actually use more than they report here.
5727 0 : pub(crate) fn local_storage_wanted(&self) -> u64 {
5728 0 : let timelines = self.timelines.lock().unwrap();
5729 :
5730 : // Heuristic: we use the max() of the timelines' visible sizes, rather than the sum. This
5731 : // reflects the observation that on tenants with multiple large branches, typically only one
5732 : // of them is used actively enough to occupy space on disk.
5733 0 : timelines
5734 0 : .values()
5735 0 : .map(|t| t.metrics.visible_physical_size_gauge.get())
5736 0 : .max()
5737 0 : .unwrap_or(0)
5738 0 : }
5739 :
5740 : /// HADRON
5741 : /// Return the visible size of all timelines in this tenant.
5742 0 : pub(crate) fn get_visible_size(&self) -> u64 {
5743 0 : let timelines = self.timelines.lock().unwrap();
5744 0 : timelines
5745 0 : .values()
5746 0 : .map(|t| t.metrics.visible_physical_size_gauge.get())
5747 0 : .sum()
5748 0 : }
5749 :
5750 : /// Builds a new tenant manifest, and uploads it if it differs from the last-known tenant
5751 : /// manifest in `Self::remote_tenant_manifest`.
5752 : ///
5753 : /// TODO: instead of requiring callers to remember to call `maybe_upload_tenant_manifest` after
5754 : /// changing any `TenantShard` state that's included in the manifest, consider making the manifest
5755 : /// the authoritative source of data with an API that automatically uploads on changes. Revisit
5756 : /// this when the manifest is more widely used and we have a better idea of the data model.
5757 120 : pub(crate) async fn maybe_upload_tenant_manifest(&self) -> Result<(), TenantManifestError> {
5758 : // Multiple tasks may call this function concurrently after mutating the TenantShard runtime
5759 : // state, affecting the manifest generated by `build_tenant_manifest`. We use an async mutex
5760 : // to serialize these callers. `eq_ignoring_version` acts as a slightly inefficient but
5761 : // simple coalescing mechanism.
5762 120 : let mut guard = tokio::select! {
5763 120 : guard = self.remote_tenant_manifest.lock() => guard,
5764 120 : _ = self.cancel.cancelled() => return Err(TenantManifestError::Cancelled),
5765 : };
5766 :
5767 : // Build a new manifest.
5768 120 : let manifest = self.build_tenant_manifest();
5769 :
5770 : // Check if the manifest has changed. We ignore the version number here, to avoid
5771 : // uploading every manifest on version number bumps.
5772 120 : if let Some(old) = guard.as_ref() {
5773 4 : if manifest.eq_ignoring_version(old) {
5774 3 : return Ok(());
5775 1 : }
5776 116 : }
5777 :
5778 : // Update metrics
5779 117 : let tid = self.tenant_shard_id.to_string();
5780 117 : let shard_id = self.tenant_shard_id.shard_slug().to_string();
5781 117 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
5782 117 : TENANT_OFFLOADED_TIMELINES
5783 117 : .with_label_values(set_key)
5784 117 : .set(manifest.offloaded_timelines.len() as u64);
5785 :
5786 : // Upload the manifest. Remote storage does no retries internally, so retry here.
5787 117 : match backoff::retry(
5788 117 : || async {
5789 117 : upload_tenant_manifest(
5790 117 : &self.remote_storage,
5791 117 : &self.tenant_shard_id,
5792 117 : self.generation,
5793 117 : &manifest,
5794 117 : &self.cancel,
5795 117 : )
5796 117 : .await
5797 234 : },
5798 0 : |_| self.cancel.is_cancelled(),
5799 : FAILED_UPLOAD_WARN_THRESHOLD,
5800 : FAILED_REMOTE_OP_RETRIES,
5801 117 : "uploading tenant manifest",
5802 117 : &self.cancel,
5803 : )
5804 117 : .await
5805 : {
5806 0 : None => Err(TenantManifestError::Cancelled),
5807 0 : Some(Err(_)) if self.cancel.is_cancelled() => Err(TenantManifestError::Cancelled),
5808 0 : Some(Err(e)) => Err(TenantManifestError::RemoteStorage(e)),
5809 : Some(Ok(_)) => {
5810 : // Store the successfully uploaded manifest, so that future callers can avoid
5811 : // re-uploading the same thing.
5812 117 : *guard = Some(manifest);
5813 :
5814 117 : Ok(())
5815 : }
5816 : }
5817 120 : }
5818 : }
5819 :
5820 : /// Create the cluster temporarily in 'initdbpath' directory inside the repository
5821 : /// to get bootstrap data for timeline initialization.
5822 0 : async fn run_initdb(
5823 0 : conf: &'static PageServerConf,
5824 0 : initdb_target_dir: &Utf8Path,
5825 0 : pg_version: PgMajorVersion,
5826 0 : cancel: &CancellationToken,
5827 0 : ) -> Result<(), InitdbError> {
5828 0 : let initdb_bin_path = conf
5829 0 : .pg_bin_dir(pg_version)
5830 0 : .map_err(InitdbError::Other)?
5831 0 : .join("initdb");
5832 0 : let initdb_lib_dir = conf.pg_lib_dir(pg_version).map_err(InitdbError::Other)?;
5833 0 : info!(
5834 0 : "running {} in {}, libdir: {}",
5835 : initdb_bin_path, initdb_target_dir, initdb_lib_dir,
5836 : );
5837 :
5838 0 : let _permit = {
5839 0 : let _timer = INITDB_SEMAPHORE_ACQUISITION_TIME.start_timer();
5840 0 : INIT_DB_SEMAPHORE.acquire().await
5841 : };
5842 :
5843 0 : CONCURRENT_INITDBS.inc();
5844 0 : scopeguard::defer! {
5845 : CONCURRENT_INITDBS.dec();
5846 : }
5847 :
5848 0 : let _timer = INITDB_RUN_TIME.start_timer();
5849 0 : let res = postgres_initdb::do_run_initdb(postgres_initdb::RunInitdbArgs {
5850 0 : superuser: &conf.superuser,
5851 0 : locale: &conf.locale,
5852 0 : initdb_bin: &initdb_bin_path,
5853 0 : pg_version,
5854 0 : library_search_path: &initdb_lib_dir,
5855 0 : pgdata: initdb_target_dir,
5856 0 : })
5857 0 : .await
5858 0 : .map_err(InitdbError::Inner);
5859 :
5860 : // This isn't true cancellation support, see above. Still return an error to
5861 : // excercise the cancellation code path.
5862 0 : if cancel.is_cancelled() {
5863 0 : return Err(InitdbError::Cancelled);
5864 0 : }
5865 :
5866 0 : res
5867 0 : }
5868 :
5869 : /// Dump contents of a layer file to stdout.
5870 0 : pub async fn dump_layerfile_from_path(
5871 0 : path: &Utf8Path,
5872 0 : verbose: bool,
5873 0 : ctx: &RequestContext,
5874 0 : ) -> anyhow::Result<()> {
5875 : use std::os::unix::fs::FileExt;
5876 :
5877 : // All layer files start with a two-byte "magic" value, to identify the kind of
5878 : // file.
5879 0 : let file = File::open(path)?;
5880 0 : let mut header_buf = [0u8; 2];
5881 0 : file.read_exact_at(&mut header_buf, 0)?;
5882 :
5883 0 : match u16::from_be_bytes(header_buf) {
5884 : crate::IMAGE_FILE_MAGIC => {
5885 0 : ImageLayer::new_for_path(path, file)?
5886 0 : .dump(verbose, ctx)
5887 0 : .await?
5888 : }
5889 : crate::DELTA_FILE_MAGIC => {
5890 0 : DeltaLayer::new_for_path(path, file)?
5891 0 : .dump(verbose, ctx)
5892 0 : .await?
5893 : }
5894 0 : magic => bail!("unrecognized magic identifier: {:?}", magic),
5895 : }
5896 :
5897 0 : Ok(())
5898 0 : }
5899 :
5900 : #[cfg(test)]
5901 : pub(crate) mod harness {
5902 : use bytes::{Bytes, BytesMut};
5903 : use hex_literal::hex;
5904 : use once_cell::sync::OnceCell;
5905 : use pageserver_api::key::Key;
5906 : use pageserver_api::models::ShardParameters;
5907 : use pageserver_api::shard::ShardIndex;
5908 : use utils::id::TenantId;
5909 : use utils::logging;
5910 : use wal_decoder::models::record::NeonWalRecord;
5911 :
5912 : use super::*;
5913 : use crate::deletion_queue::mock::MockDeletionQueue;
5914 : use crate::l0_flush::L0FlushConfig;
5915 : use crate::walredo::apply_neon;
5916 :
5917 : pub const TIMELINE_ID: TimelineId =
5918 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
5919 : pub const NEW_TIMELINE_ID: TimelineId =
5920 : TimelineId::from_array(hex!("AA223344556677881122334455667788"));
5921 :
5922 : /// Convenience function to create a page image with given string as the only content
5923 2514409 : pub fn test_img(s: &str) -> Bytes {
5924 2514409 : let mut buf = BytesMut::new();
5925 2514409 : buf.extend_from_slice(s.as_bytes());
5926 2514409 : buf.resize(64, 0);
5927 :
5928 2514409 : buf.freeze()
5929 2514409 : }
5930 :
5931 : pub struct TenantHarness {
5932 : pub conf: &'static PageServerConf,
5933 : pub tenant_conf: pageserver_api::models::TenantConfig,
5934 : pub tenant_shard_id: TenantShardId,
5935 : pub shard_identity: ShardIdentity,
5936 : pub generation: Generation,
5937 : pub shard: ShardIndex,
5938 : pub remote_storage: GenericRemoteStorage,
5939 : pub remote_fs_dir: Utf8PathBuf,
5940 : pub deletion_queue: MockDeletionQueue,
5941 : }
5942 :
5943 : static LOG_HANDLE: OnceCell<()> = OnceCell::new();
5944 :
5945 131 : pub(crate) fn setup_logging() {
5946 131 : LOG_HANDLE.get_or_init(|| {
5947 125 : logging::init(
5948 125 : logging::LogFormat::Test,
5949 : // enable it in case the tests exercise code paths that use
5950 : // debug_assert_current_span_has_tenant_and_timeline_id
5951 125 : logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
5952 125 : logging::Output::Stdout,
5953 : )
5954 125 : .expect("Failed to init test logging");
5955 125 : });
5956 131 : }
5957 :
5958 : impl TenantHarness {
5959 119 : pub async fn create_custom(
5960 119 : test_name: &'static str,
5961 119 : tenant_conf: pageserver_api::models::TenantConfig,
5962 119 : tenant_id: TenantId,
5963 119 : shard_identity: ShardIdentity,
5964 119 : generation: Generation,
5965 119 : ) -> anyhow::Result<Self> {
5966 119 : setup_logging();
5967 :
5968 119 : let repo_dir = PageServerConf::test_repo_dir(test_name);
5969 119 : let _ = fs::remove_dir_all(&repo_dir);
5970 119 : fs::create_dir_all(&repo_dir)?;
5971 :
5972 119 : let conf = PageServerConf::dummy_conf(repo_dir);
5973 : // Make a static copy of the config. This can never be free'd, but that's
5974 : // OK in a test.
5975 119 : let conf: &'static PageServerConf = Box::leak(Box::new(conf));
5976 :
5977 119 : let shard = shard_identity.shard_index();
5978 119 : let tenant_shard_id = TenantShardId {
5979 119 : tenant_id,
5980 119 : shard_number: shard.shard_number,
5981 119 : shard_count: shard.shard_count,
5982 119 : };
5983 119 : fs::create_dir_all(conf.tenant_path(&tenant_shard_id))?;
5984 119 : fs::create_dir_all(conf.timelines_path(&tenant_shard_id))?;
5985 :
5986 : use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
5987 119 : let remote_fs_dir = conf.workdir.join("localfs");
5988 119 : std::fs::create_dir_all(&remote_fs_dir).unwrap();
5989 119 : let config = RemoteStorageConfig {
5990 119 : storage: RemoteStorageKind::LocalFs {
5991 119 : local_path: remote_fs_dir.clone(),
5992 119 : },
5993 119 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
5994 119 : small_timeout: RemoteStorageConfig::DEFAULT_SMALL_TIMEOUT,
5995 119 : };
5996 119 : let remote_storage = GenericRemoteStorage::from_config(&config).await.unwrap();
5997 119 : let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()));
5998 :
5999 119 : Ok(Self {
6000 119 : conf,
6001 119 : tenant_conf,
6002 119 : tenant_shard_id,
6003 119 : shard_identity,
6004 119 : generation,
6005 119 : shard,
6006 119 : remote_storage,
6007 119 : remote_fs_dir,
6008 119 : deletion_queue,
6009 119 : })
6010 119 : }
6011 :
6012 110 : pub async fn create(test_name: &'static str) -> anyhow::Result<Self> {
6013 : // Disable automatic GC and compaction to make the unit tests more deterministic.
6014 : // The tests perform them manually if needed.
6015 110 : let tenant_conf = pageserver_api::models::TenantConfig {
6016 110 : gc_period: Some(Duration::ZERO),
6017 110 : compaction_period: Some(Duration::ZERO),
6018 110 : ..Default::default()
6019 110 : };
6020 110 : let tenant_id = TenantId::generate();
6021 110 : let shard = ShardIdentity::unsharded();
6022 110 : Self::create_custom(
6023 110 : test_name,
6024 110 : tenant_conf,
6025 110 : tenant_id,
6026 110 : shard,
6027 110 : Generation::new(0xdeadbeef),
6028 110 : )
6029 110 : .await
6030 110 : }
6031 :
6032 10 : pub fn span(&self) -> tracing::Span {
6033 10 : info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
6034 10 : }
6035 :
6036 119 : pub(crate) async fn load(&self) -> (Arc<TenantShard>, RequestContext) {
6037 119 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error)
6038 119 : .with_scope_unit_test();
6039 : (
6040 119 : self.do_try_load(&ctx)
6041 119 : .await
6042 119 : .expect("failed to load test tenant"),
6043 119 : ctx,
6044 : )
6045 119 : }
6046 :
6047 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
6048 : pub(crate) async fn do_try_load_with_redo(
6049 : &self,
6050 : walredo_mgr: Arc<WalRedoManager>,
6051 : ctx: &RequestContext,
6052 : ) -> anyhow::Result<Arc<TenantShard>> {
6053 : let (basebackup_cache, _) = BasebackupCache::new(Utf8PathBuf::new(), None);
6054 :
6055 : let tenant = Arc::new(TenantShard::new(
6056 : TenantState::Attaching,
6057 : self.conf,
6058 : AttachedTenantConf::try_from(
6059 : self.conf,
6060 : LocationConf::attached_single(
6061 : self.tenant_conf.clone(),
6062 : self.generation,
6063 : ShardParameters::default(),
6064 : ),
6065 : )
6066 : .unwrap(),
6067 : self.shard_identity,
6068 : Some(walredo_mgr),
6069 : self.tenant_shard_id,
6070 : self.remote_storage.clone(),
6071 : self.deletion_queue.new_client(),
6072 : // TODO: ideally we should run all unit tests with both configs
6073 : L0FlushGlobalState::new(L0FlushConfig::default()),
6074 : basebackup_cache,
6075 : FeatureResolver::new_disabled(),
6076 : ));
6077 :
6078 : let preload = tenant
6079 : .preload(&self.remote_storage, CancellationToken::new())
6080 : .await?;
6081 : tenant.attach(Some(preload), ctx).await?;
6082 :
6083 : tenant.state.send_replace(TenantState::Active);
6084 : for timeline in tenant.timelines.lock().unwrap().values() {
6085 : timeline.set_state(TimelineState::Active);
6086 : }
6087 : Ok(tenant)
6088 : }
6089 :
6090 119 : pub(crate) async fn do_try_load(
6091 119 : &self,
6092 119 : ctx: &RequestContext,
6093 119 : ) -> anyhow::Result<Arc<TenantShard>> {
6094 119 : let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
6095 119 : self.do_try_load_with_redo(walredo_mgr, ctx).await
6096 119 : }
6097 :
6098 1 : pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
6099 1 : self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
6100 1 : }
6101 : }
6102 :
6103 : // Mock WAL redo manager that doesn't do much
6104 : pub(crate) struct TestRedoManager;
6105 :
6106 : impl TestRedoManager {
6107 : /// # Cancel-Safety
6108 : ///
6109 : /// This method is cancellation-safe.
6110 26649 : pub async fn request_redo(
6111 26649 : &self,
6112 26649 : key: Key,
6113 26649 : lsn: Lsn,
6114 26649 : base_img: Option<(Lsn, Bytes)>,
6115 26649 : records: Vec<(Lsn, NeonWalRecord)>,
6116 26649 : _pg_version: PgMajorVersion,
6117 26649 : _redo_attempt_type: RedoAttemptType,
6118 26649 : ) -> Result<Bytes, walredo::Error> {
6119 1401844 : let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
6120 26649 : if records_neon {
6121 : // For Neon wal records, we can decode without spawning postgres, so do so.
6122 26649 : let mut page = match (base_img, records.first()) {
6123 12781 : (Some((_lsn, img)), _) => {
6124 12781 : let mut page = BytesMut::new();
6125 12781 : page.extend_from_slice(&img);
6126 12781 : page
6127 : }
6128 13868 : (_, Some((_lsn, rec))) if rec.will_init() => BytesMut::new(),
6129 : _ => {
6130 0 : panic!("Neon WAL redo requires base image or will init record");
6131 : }
6132 : };
6133 :
6134 1428492 : for (record_lsn, record) in records {
6135 1401844 : apply_neon::apply_in_neon(&record, record_lsn, key, &mut page)?;
6136 : }
6137 26648 : Ok(page.freeze())
6138 : } else {
6139 : // We never spawn a postgres walredo process in unit tests: just log what we might have done.
6140 0 : let s = format!(
6141 0 : "redo for {} to get to {}, with {} and {} records",
6142 : key,
6143 : lsn,
6144 0 : if base_img.is_some() {
6145 0 : "base image"
6146 : } else {
6147 0 : "no base image"
6148 : },
6149 0 : records.len()
6150 : );
6151 0 : println!("{s}");
6152 :
6153 0 : Ok(test_img(&s))
6154 : }
6155 26649 : }
6156 : }
6157 : }
6158 :
6159 : #[cfg(test)]
6160 : mod tests {
6161 : use std::collections::{BTreeMap, BTreeSet};
6162 :
6163 : use bytes::{Bytes, BytesMut};
6164 : use hex_literal::hex;
6165 : use itertools::Itertools;
6166 : #[cfg(feature = "testing")]
6167 : use models::CompactLsnRange;
6168 : use pageserver_api::key::{
6169 : AUX_KEY_PREFIX, Key, NON_INHERITED_RANGE, RELATION_SIZE_PREFIX, repl_origin_key,
6170 : };
6171 : use pageserver_api::keyspace::KeySpace;
6172 : #[cfg(feature = "testing")]
6173 : use pageserver_api::keyspace::KeySpaceRandomAccum;
6174 : use pageserver_api::models::{CompactionAlgorithm, CompactionAlgorithmSettings, LsnLease};
6175 : use pageserver_compaction::helpers::overlaps_with;
6176 : use rand::Rng;
6177 : #[cfg(feature = "testing")]
6178 : use rand::SeedableRng;
6179 : #[cfg(feature = "testing")]
6180 : use rand::rngs::StdRng;
6181 : #[cfg(feature = "testing")]
6182 : use std::ops::Range;
6183 : use storage_layer::{IoConcurrency, PersistentLayerKey};
6184 : use tests::storage_layer::ValuesReconstructState;
6185 : use tests::timeline::{GetVectoredError, ShutdownMode};
6186 : #[cfg(feature = "testing")]
6187 : use timeline::GcInfo;
6188 : #[cfg(feature = "testing")]
6189 : use timeline::InMemoryLayerTestDesc;
6190 : #[cfg(feature = "testing")]
6191 : use timeline::compaction::{KeyHistoryRetention, KeyLogAtLsn};
6192 : use timeline::{CompactOptions, DeltaLayerTestDesc, VersionedKeySpaceQuery};
6193 : use utils::id::TenantId;
6194 : use utils::shard::{ShardCount, ShardNumber};
6195 : #[cfg(feature = "testing")]
6196 : use wal_decoder::models::record::NeonWalRecord;
6197 : use wal_decoder::models::value::Value;
6198 :
6199 : use super::*;
6200 : use crate::DEFAULT_PG_VERSION;
6201 : use crate::keyspace::KeySpaceAccum;
6202 : use crate::tenant::harness::*;
6203 : use crate::tenant::timeline::CompactFlags;
6204 :
6205 : static TEST_KEY: Lazy<Key> =
6206 10 : Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));
6207 :
6208 : #[cfg(feature = "testing")]
6209 : struct TestTimelineSpecification {
6210 : start_lsn: Lsn,
6211 : last_record_lsn: Lsn,
6212 :
6213 : in_memory_layers_shape: Vec<(Range<Key>, Range<Lsn>)>,
6214 : delta_layers_shape: Vec<(Range<Key>, Range<Lsn>)>,
6215 : image_layers_shape: Vec<(Range<Key>, Lsn)>,
6216 :
6217 : gap_chance: u8,
6218 : will_init_chance: u8,
6219 : }
6220 :
6221 : #[cfg(feature = "testing")]
6222 : struct Storage {
6223 : storage: HashMap<(Key, Lsn), Value>,
6224 : start_lsn: Lsn,
6225 : }
6226 :
6227 : #[cfg(feature = "testing")]
6228 : impl Storage {
6229 32000 : fn get(&self, key: Key, lsn: Lsn) -> Bytes {
6230 : use bytes::BufMut;
6231 :
6232 32000 : let mut crnt_lsn = lsn;
6233 32000 : let mut got_base = false;
6234 :
6235 32000 : let mut acc = Vec::new();
6236 :
6237 2798140 : while crnt_lsn >= self.start_lsn {
6238 2798140 : if let Some(value) = self.storage.get(&(key, crnt_lsn)) {
6239 1419383 : acc.push(value.clone());
6240 :
6241 1401215 : match value {
6242 1401215 : Value::WalRecord(NeonWalRecord::Test { will_init, .. }) => {
6243 1401215 : if *will_init {
6244 13832 : got_base = true;
6245 13832 : break;
6246 1387383 : }
6247 : }
6248 : Value::Image(_) => {
6249 18168 : got_base = true;
6250 18168 : break;
6251 : }
6252 0 : _ => unreachable!(),
6253 : }
6254 1378757 : }
6255 :
6256 2766140 : crnt_lsn = crnt_lsn.checked_sub(1u64).unwrap();
6257 : }
6258 :
6259 32000 : assert!(
6260 32000 : got_base,
6261 0 : "Input data was incorrect. No base image for {key}@{lsn}"
6262 : );
6263 :
6264 32000 : tracing::debug!("Wal redo depth for {key}@{lsn} is {}", acc.len());
6265 :
6266 32000 : let mut blob = BytesMut::new();
6267 1419383 : for value in acc.into_iter().rev() {
6268 1401215 : match value {
6269 1401215 : Value::WalRecord(NeonWalRecord::Test { append, .. }) => {
6270 1401215 : blob.extend_from_slice(append.as_bytes());
6271 1401215 : }
6272 18168 : Value::Image(img) => {
6273 18168 : blob.put(img);
6274 18168 : }
6275 0 : _ => unreachable!(),
6276 : }
6277 : }
6278 :
6279 32000 : blob.into()
6280 32000 : }
6281 : }
6282 :
6283 : #[cfg(feature = "testing")]
6284 : #[allow(clippy::too_many_arguments)]
6285 1 : async fn randomize_timeline(
6286 1 : tenant: &Arc<TenantShard>,
6287 1 : new_timeline_id: TimelineId,
6288 1 : pg_version: PgMajorVersion,
6289 1 : spec: TestTimelineSpecification,
6290 1 : random: &mut rand::rngs::StdRng,
6291 1 : ctx: &RequestContext,
6292 1 : ) -> anyhow::Result<(Arc<Timeline>, Storage, Vec<Lsn>)> {
6293 1 : let mut storage: HashMap<(Key, Lsn), Value> = HashMap::default();
6294 1 : let mut interesting_lsns = vec![spec.last_record_lsn];
6295 :
6296 2 : for (key_range, lsn_range) in spec.in_memory_layers_shape.iter() {
6297 2 : let mut lsn = lsn_range.start;
6298 202 : while lsn < lsn_range.end {
6299 200 : let mut key = key_range.start;
6300 21018 : while key < key_range.end {
6301 20818 : let gap = random.random_range(1..=100) <= spec.gap_chance;
6302 20818 : let will_init = random.random_range(1..=100) <= spec.will_init_chance;
6303 :
6304 20818 : if gap {
6305 1018 : continue;
6306 19800 : }
6307 :
6308 19800 : let record = if will_init {
6309 191 : Value::WalRecord(NeonWalRecord::wal_init(format!("[wil_init {key}@{lsn}]")))
6310 : } else {
6311 19609 : Value::WalRecord(NeonWalRecord::wal_append(format!("[delta {key}@{lsn}]")))
6312 : };
6313 :
6314 19800 : storage.insert((key, lsn), record);
6315 :
6316 19800 : key = key.next();
6317 : }
6318 200 : lsn = Lsn(lsn.0 + 1);
6319 : }
6320 :
6321 : // Stash some interesting LSN for future use
6322 6 : for offset in [0, 5, 100].iter() {
6323 6 : if *offset == 0 {
6324 2 : interesting_lsns.push(lsn_range.start);
6325 2 : } else {
6326 4 : let below = lsn_range.start.checked_sub(*offset);
6327 4 : match below {
6328 4 : Some(v) if v >= spec.start_lsn => {
6329 4 : interesting_lsns.push(v);
6330 4 : }
6331 0 : _ => {}
6332 : }
6333 :
6334 4 : let above = Lsn(lsn_range.start.0 + offset);
6335 4 : interesting_lsns.push(above);
6336 : }
6337 : }
6338 : }
6339 :
6340 3 : for (key_range, lsn_range) in spec.delta_layers_shape.iter() {
6341 3 : let mut lsn = lsn_range.start;
6342 315 : while lsn < lsn_range.end {
6343 312 : let mut key = key_range.start;
6344 11112 : while key < key_range.end {
6345 10800 : let gap = random.random_range(1..=100) <= spec.gap_chance;
6346 10800 : let will_init = random.random_range(1..=100) <= spec.will_init_chance;
6347 :
6348 10800 : if gap {
6349 504 : continue;
6350 10296 : }
6351 :
6352 10296 : let record = if will_init {
6353 103 : Value::WalRecord(NeonWalRecord::wal_init(format!("[wil_init {key}@{lsn}]")))
6354 : } else {
6355 10193 : Value::WalRecord(NeonWalRecord::wal_append(format!("[delta {key}@{lsn}]")))
6356 : };
6357 :
6358 10296 : storage.insert((key, lsn), record);
6359 :
6360 10296 : key = key.next();
6361 : }
6362 312 : lsn = Lsn(lsn.0 + 1);
6363 : }
6364 :
6365 : // Stash some interesting LSN for future use
6366 9 : for offset in [0, 5, 100].iter() {
6367 9 : if *offset == 0 {
6368 3 : interesting_lsns.push(lsn_range.start);
6369 3 : } else {
6370 6 : let below = lsn_range.start.checked_sub(*offset);
6371 6 : match below {
6372 6 : Some(v) if v >= spec.start_lsn => {
6373 3 : interesting_lsns.push(v);
6374 3 : }
6375 3 : _ => {}
6376 : }
6377 :
6378 6 : let above = Lsn(lsn_range.start.0 + offset);
6379 6 : interesting_lsns.push(above);
6380 : }
6381 : }
6382 : }
6383 :
6384 3 : for (key_range, lsn) in spec.image_layers_shape.iter() {
6385 3 : let mut key = key_range.start;
6386 142 : while key < key_range.end {
6387 139 : let blob = Bytes::from(format!("[image {key}@{lsn}]"));
6388 139 : let record = Value::Image(blob.clone());
6389 139 : storage.insert((key, *lsn), record);
6390 139 :
6391 139 : key = key.next();
6392 139 : }
6393 :
6394 : // Stash some interesting LSN for future use
6395 9 : for offset in [0, 5, 100].iter() {
6396 9 : if *offset == 0 {
6397 3 : interesting_lsns.push(*lsn);
6398 3 : } else {
6399 6 : let below = lsn.checked_sub(*offset);
6400 6 : match below {
6401 6 : Some(v) if v >= spec.start_lsn => {
6402 4 : interesting_lsns.push(v);
6403 4 : }
6404 2 : _ => {}
6405 : }
6406 :
6407 6 : let above = Lsn(lsn.0 + offset);
6408 6 : interesting_lsns.push(above);
6409 : }
6410 : }
6411 : }
6412 :
6413 1 : let in_memory_test_layers = {
6414 1 : let mut acc = Vec::new();
6415 :
6416 2 : for (key_range, lsn_range) in spec.in_memory_layers_shape.iter() {
6417 2 : let mut data = Vec::new();
6418 :
6419 2 : let mut lsn = lsn_range.start;
6420 202 : while lsn < lsn_range.end {
6421 200 : let mut key = key_range.start;
6422 20000 : while key < key_range.end {
6423 19800 : if let Some(record) = storage.get(&(key, lsn)) {
6424 19800 : data.push((key, lsn, record.clone()));
6425 19800 : }
6426 :
6427 19800 : key = key.next();
6428 : }
6429 200 : lsn = Lsn(lsn.0 + 1);
6430 : }
6431 :
6432 2 : acc.push(InMemoryLayerTestDesc {
6433 2 : data,
6434 2 : lsn_range: lsn_range.clone(),
6435 2 : is_open: false,
6436 2 : })
6437 : }
6438 :
6439 1 : acc
6440 : };
6441 :
6442 1 : let delta_test_layers = {
6443 1 : let mut acc = Vec::new();
6444 :
6445 3 : for (key_range, lsn_range) in spec.delta_layers_shape.iter() {
6446 3 : let mut data = Vec::new();
6447 :
6448 3 : let mut lsn = lsn_range.start;
6449 315 : while lsn < lsn_range.end {
6450 312 : let mut key = key_range.start;
6451 10608 : while key < key_range.end {
6452 10296 : if let Some(record) = storage.get(&(key, lsn)) {
6453 10296 : data.push((key, lsn, record.clone()));
6454 10296 : }
6455 :
6456 10296 : key = key.next();
6457 : }
6458 312 : lsn = Lsn(lsn.0 + 1);
6459 : }
6460 :
6461 3 : acc.push(DeltaLayerTestDesc {
6462 3 : data,
6463 3 : lsn_range: lsn_range.clone(),
6464 3 : key_range: key_range.clone(),
6465 3 : })
6466 : }
6467 :
6468 1 : acc
6469 : };
6470 :
6471 1 : let image_test_layers = {
6472 1 : let mut acc = Vec::new();
6473 :
6474 3 : for (key_range, lsn) in spec.image_layers_shape.iter() {
6475 3 : let mut data = Vec::new();
6476 :
6477 3 : let mut key = key_range.start;
6478 142 : while key < key_range.end {
6479 139 : if let Some(record) = storage.get(&(key, *lsn)) {
6480 139 : let blob = match record {
6481 139 : Value::Image(blob) => blob.clone(),
6482 0 : _ => unreachable!(),
6483 : };
6484 :
6485 139 : data.push((key, blob));
6486 0 : }
6487 :
6488 139 : key = key.next();
6489 : }
6490 :
6491 3 : acc.push((*lsn, data));
6492 : }
6493 :
6494 1 : acc
6495 : };
6496 :
6497 1 : let tline = tenant
6498 1 : .create_test_timeline_with_layers(
6499 1 : new_timeline_id,
6500 1 : spec.start_lsn,
6501 1 : pg_version,
6502 1 : ctx,
6503 1 : in_memory_test_layers,
6504 1 : delta_test_layers,
6505 1 : image_test_layers,
6506 1 : spec.last_record_lsn,
6507 1 : )
6508 1 : .await?;
6509 :
6510 1 : Ok((
6511 1 : tline,
6512 1 : Storage {
6513 1 : storage,
6514 1 : start_lsn: spec.start_lsn,
6515 1 : },
6516 1 : interesting_lsns,
6517 1 : ))
6518 1 : }
6519 :
6520 : #[tokio::test]
6521 1 : async fn test_basic() -> anyhow::Result<()> {
6522 1 : let (tenant, ctx) = TenantHarness::create("test_basic").await?.load().await;
6523 1 : let tline = tenant
6524 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6525 1 : .await?;
6526 :
6527 1 : let mut writer = tline.writer().await;
6528 1 : writer
6529 1 : .put(
6530 1 : *TEST_KEY,
6531 1 : Lsn(0x10),
6532 1 : &Value::Image(test_img("foo at 0x10")),
6533 1 : &ctx,
6534 1 : )
6535 1 : .await?;
6536 1 : writer.finish_write(Lsn(0x10));
6537 1 : drop(writer);
6538 :
6539 1 : let mut writer = tline.writer().await;
6540 1 : writer
6541 1 : .put(
6542 1 : *TEST_KEY,
6543 1 : Lsn(0x20),
6544 1 : &Value::Image(test_img("foo at 0x20")),
6545 1 : &ctx,
6546 1 : )
6547 1 : .await?;
6548 1 : writer.finish_write(Lsn(0x20));
6549 1 : drop(writer);
6550 :
6551 1 : assert_eq!(
6552 1 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
6553 1 : test_img("foo at 0x10")
6554 : );
6555 1 : assert_eq!(
6556 1 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
6557 1 : test_img("foo at 0x10")
6558 : );
6559 1 : assert_eq!(
6560 1 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
6561 1 : test_img("foo at 0x20")
6562 : );
6563 :
6564 2 : Ok(())
6565 1 : }
6566 :
6567 : #[tokio::test]
6568 1 : async fn no_duplicate_timelines() -> anyhow::Result<()> {
6569 1 : let (tenant, ctx) = TenantHarness::create("no_duplicate_timelines")
6570 1 : .await?
6571 1 : .load()
6572 1 : .await;
6573 1 : let _ = tenant
6574 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6575 1 : .await?;
6576 :
6577 1 : match tenant
6578 1 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6579 1 : .await
6580 1 : {
6581 1 : Ok(_) => panic!("duplicate timeline creation should fail"),
6582 1 : Err(e) => assert_eq!(
6583 1 : e.to_string(),
6584 1 : "timeline already exists with different parameters".to_string()
6585 1 : ),
6586 1 : }
6587 1 :
6588 1 : Ok(())
6589 1 : }
6590 :
6591 : /// Convenience function to create a page image with given string as the only content
6592 5 : pub fn test_value(s: &str) -> Value {
6593 5 : let mut buf = BytesMut::new();
6594 5 : buf.extend_from_slice(s.as_bytes());
6595 5 : Value::Image(buf.freeze())
6596 5 : }
6597 :
6598 : ///
6599 : /// Test branch creation
6600 : ///
6601 : #[tokio::test]
6602 1 : async fn test_branch() -> anyhow::Result<()> {
6603 : use std::str::from_utf8;
6604 :
6605 1 : let (tenant, ctx) = TenantHarness::create("test_branch").await?.load().await;
6606 1 : let tline = tenant
6607 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6608 1 : .await?;
6609 1 : let mut writer = tline.writer().await;
6610 :
6611 : #[allow(non_snake_case)]
6612 1 : let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
6613 : #[allow(non_snake_case)]
6614 1 : let TEST_KEY_B: Key = Key::from_hex("110000000033333333444444445500000002").unwrap();
6615 :
6616 : // Insert a value on the timeline
6617 1 : writer
6618 1 : .put(TEST_KEY_A, Lsn(0x20), &test_value("foo at 0x20"), &ctx)
6619 1 : .await?;
6620 1 : writer
6621 1 : .put(TEST_KEY_B, Lsn(0x20), &test_value("foobar at 0x20"), &ctx)
6622 1 : .await?;
6623 1 : writer.finish_write(Lsn(0x20));
6624 :
6625 1 : writer
6626 1 : .put(TEST_KEY_A, Lsn(0x30), &test_value("foo at 0x30"), &ctx)
6627 1 : .await?;
6628 1 : writer.finish_write(Lsn(0x30));
6629 1 : writer
6630 1 : .put(TEST_KEY_A, Lsn(0x40), &test_value("foo at 0x40"), &ctx)
6631 1 : .await?;
6632 1 : writer.finish_write(Lsn(0x40));
6633 :
6634 : //assert_current_logical_size(&tline, Lsn(0x40));
6635 :
6636 : // Branch the history, modify relation differently on the new timeline
6637 1 : tenant
6638 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x30)), &ctx)
6639 1 : .await?;
6640 1 : let newtline = tenant
6641 1 : .get_timeline(NEW_TIMELINE_ID, true)
6642 1 : .expect("Should have a local timeline");
6643 1 : let mut new_writer = newtline.writer().await;
6644 1 : new_writer
6645 1 : .put(TEST_KEY_A, Lsn(0x40), &test_value("bar at 0x40"), &ctx)
6646 1 : .await?;
6647 1 : new_writer.finish_write(Lsn(0x40));
6648 :
6649 : // Check page contents on both branches
6650 1 : assert_eq!(
6651 1 : from_utf8(&tline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
6652 : "foo at 0x40"
6653 : );
6654 1 : assert_eq!(
6655 1 : from_utf8(&newtline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
6656 : "bar at 0x40"
6657 : );
6658 1 : assert_eq!(
6659 1 : from_utf8(&newtline.get(TEST_KEY_B, Lsn(0x40), &ctx).await?)?,
6660 : "foobar at 0x20"
6661 : );
6662 :
6663 : //assert_current_logical_size(&tline, Lsn(0x40));
6664 :
6665 2 : Ok(())
6666 1 : }
6667 :
6668 10 : async fn make_some_layers(
6669 10 : tline: &Timeline,
6670 10 : start_lsn: Lsn,
6671 10 : ctx: &RequestContext,
6672 10 : ) -> anyhow::Result<()> {
6673 10 : let mut lsn = start_lsn;
6674 : {
6675 10 : let mut writer = tline.writer().await;
6676 : // Create a relation on the timeline
6677 10 : writer
6678 10 : .put(
6679 10 : *TEST_KEY,
6680 10 : lsn,
6681 10 : &Value::Image(test_img(&format!("foo at {lsn}"))),
6682 10 : ctx,
6683 10 : )
6684 10 : .await?;
6685 10 : writer.finish_write(lsn);
6686 10 : lsn += 0x10;
6687 10 : writer
6688 10 : .put(
6689 10 : *TEST_KEY,
6690 10 : lsn,
6691 10 : &Value::Image(test_img(&format!("foo at {lsn}"))),
6692 10 : ctx,
6693 10 : )
6694 10 : .await?;
6695 10 : writer.finish_write(lsn);
6696 10 : lsn += 0x10;
6697 : }
6698 10 : tline.freeze_and_flush().await?;
6699 : {
6700 10 : let mut writer = tline.writer().await;
6701 10 : writer
6702 10 : .put(
6703 10 : *TEST_KEY,
6704 10 : lsn,
6705 10 : &Value::Image(test_img(&format!("foo at {lsn}"))),
6706 10 : ctx,
6707 10 : )
6708 10 : .await?;
6709 10 : writer.finish_write(lsn);
6710 10 : lsn += 0x10;
6711 10 : writer
6712 10 : .put(
6713 10 : *TEST_KEY,
6714 10 : lsn,
6715 10 : &Value::Image(test_img(&format!("foo at {lsn}"))),
6716 10 : ctx,
6717 10 : )
6718 10 : .await?;
6719 10 : writer.finish_write(lsn);
6720 : }
6721 10 : tline.freeze_and_flush().await.map_err(|e| e.into())
6722 10 : }
6723 :
6724 : #[tokio::test]
6725 1 : async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
6726 1 : let (tenant, ctx) =
6727 1 : TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
6728 1 : .await?
6729 1 : .load()
6730 1 : .await;
6731 1 : let tline = tenant
6732 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6733 1 : .await?;
6734 1 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6735 :
6736 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
6737 : // FIXME: this doesn't actually remove any layer currently, given how the flushing
6738 : // and compaction works. But it does set the 'cutoff' point so that the cross check
6739 : // below should fail.
6740 1 : tenant
6741 1 : .gc_iteration(
6742 1 : Some(TIMELINE_ID),
6743 1 : 0x10,
6744 1 : Duration::ZERO,
6745 1 : &CancellationToken::new(),
6746 1 : &ctx,
6747 1 : )
6748 1 : .await?;
6749 :
6750 : // try to branch at lsn 25, should fail because we already garbage collected the data
6751 1 : match tenant
6752 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
6753 1 : .await
6754 1 : {
6755 1 : Ok(_) => panic!("branching should have failed"),
6756 1 : Err(err) => {
6757 1 : let CreateTimelineError::AncestorLsn(err) = err else {
6758 1 : panic!("wrong error type")
6759 1 : };
6760 1 : assert!(err.to_string().contains("invalid branch start lsn"));
6761 1 : assert!(
6762 1 : err.source()
6763 1 : .unwrap()
6764 1 : .to_string()
6765 1 : .contains("we might've already garbage collected needed data")
6766 1 : )
6767 1 : }
6768 1 : }
6769 1 :
6770 1 : Ok(())
6771 1 : }
6772 :
6773 : #[tokio::test]
6774 1 : async fn test_prohibit_branch_creation_on_pre_initdb_lsn() -> anyhow::Result<()> {
6775 1 : let (tenant, ctx) =
6776 1 : TenantHarness::create("test_prohibit_branch_creation_on_pre_initdb_lsn")
6777 1 : .await?
6778 1 : .load()
6779 1 : .await;
6780 :
6781 1 : let tline = tenant
6782 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x50), DEFAULT_PG_VERSION, &ctx)
6783 1 : .await?;
6784 : // try to branch at lsn 0x25, should fail because initdb lsn is 0x50
6785 1 : match tenant
6786 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
6787 1 : .await
6788 1 : {
6789 1 : Ok(_) => panic!("branching should have failed"),
6790 1 : Err(err) => {
6791 1 : let CreateTimelineError::AncestorLsn(err) = err else {
6792 1 : panic!("wrong error type");
6793 1 : };
6794 1 : assert!(&err.to_string().contains("invalid branch start lsn"));
6795 1 : assert!(
6796 1 : &err.source()
6797 1 : .unwrap()
6798 1 : .to_string()
6799 1 : .contains("is earlier than latest GC cutoff")
6800 1 : );
6801 1 : }
6802 1 : }
6803 1 :
6804 1 : Ok(())
6805 1 : }
6806 :
6807 : /*
6808 : // FIXME: This currently fails to error out. Calling GC doesn't currently
6809 : // remove the old value, we'd need to work a little harder
6810 : #[tokio::test]
6811 : async fn test_prohibit_get_for_garbage_collected_data() -> anyhow::Result<()> {
6812 : let repo =
6813 : RepoHarness::create("test_prohibit_get_for_garbage_collected_data")?
6814 : .load();
6815 :
6816 : let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION)?;
6817 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6818 :
6819 : repo.gc_iteration(Some(TIMELINE_ID), 0x10, Duration::ZERO)?;
6820 : let applied_gc_cutoff_lsn = tline.get_applied_gc_cutoff_lsn();
6821 : assert!(*applied_gc_cutoff_lsn > Lsn(0x25));
6822 : match tline.get(*TEST_KEY, Lsn(0x25)) {
6823 : Ok(_) => panic!("request for page should have failed"),
6824 : Err(err) => assert!(err.to_string().contains("not found at")),
6825 : }
6826 : Ok(())
6827 : }
6828 : */
6829 :
6830 : #[tokio::test]
6831 1 : async fn test_get_branchpoints_from_an_inactive_timeline() -> anyhow::Result<()> {
6832 1 : let (tenant, ctx) =
6833 1 : TenantHarness::create("test_get_branchpoints_from_an_inactive_timeline")
6834 1 : .await?
6835 1 : .load()
6836 1 : .await;
6837 1 : let tline = tenant
6838 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6839 1 : .await?;
6840 1 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6841 :
6842 1 : tenant
6843 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6844 1 : .await?;
6845 1 : let newtline = tenant
6846 1 : .get_timeline(NEW_TIMELINE_ID, true)
6847 1 : .expect("Should have a local timeline");
6848 :
6849 1 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6850 :
6851 1 : tline.set_broken("test".to_owned());
6852 :
6853 1 : tenant
6854 1 : .gc_iteration(
6855 1 : Some(TIMELINE_ID),
6856 1 : 0x10,
6857 1 : Duration::ZERO,
6858 1 : &CancellationToken::new(),
6859 1 : &ctx,
6860 1 : )
6861 1 : .await?;
6862 :
6863 : // The branchpoints should contain all timelines, even ones marked
6864 : // as Broken.
6865 : {
6866 1 : let branchpoints = &tline.gc_info.read().unwrap().retain_lsns;
6867 1 : assert_eq!(branchpoints.len(), 1);
6868 1 : assert_eq!(
6869 1 : branchpoints[0],
6870 : (Lsn(0x40), NEW_TIMELINE_ID, MaybeOffloaded::No)
6871 : );
6872 : }
6873 :
6874 : // You can read the key from the child branch even though the parent is
6875 : // Broken, as long as you don't need to access data from the parent.
6876 1 : assert_eq!(
6877 1 : newtline.get(*TEST_KEY, Lsn(0x70), &ctx).await?,
6878 1 : test_img(&format!("foo at {}", Lsn(0x70)))
6879 : );
6880 :
6881 : // This needs to traverse to the parent, and fails.
6882 1 : let err = newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await.unwrap_err();
6883 1 : assert!(
6884 1 : err.to_string().starts_with(&format!(
6885 1 : "bad state on timeline {}: Broken",
6886 1 : tline.timeline_id
6887 1 : )),
6888 0 : "{err}"
6889 : );
6890 :
6891 2 : Ok(())
6892 1 : }
6893 :
6894 : #[tokio::test]
6895 1 : async fn test_retain_data_in_parent_which_is_needed_for_child() -> anyhow::Result<()> {
6896 1 : let (tenant, ctx) =
6897 1 : TenantHarness::create("test_retain_data_in_parent_which_is_needed_for_child")
6898 1 : .await?
6899 1 : .load()
6900 1 : .await;
6901 1 : let tline = tenant
6902 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6903 1 : .await?;
6904 1 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6905 :
6906 1 : tenant
6907 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6908 1 : .await?;
6909 1 : let newtline = tenant
6910 1 : .get_timeline(NEW_TIMELINE_ID, true)
6911 1 : .expect("Should have a local timeline");
6912 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
6913 1 : tenant
6914 1 : .gc_iteration(
6915 1 : Some(TIMELINE_ID),
6916 1 : 0x10,
6917 1 : Duration::ZERO,
6918 1 : &CancellationToken::new(),
6919 1 : &ctx,
6920 1 : )
6921 1 : .await?;
6922 1 : assert!(newtline.get(*TEST_KEY, Lsn(0x25), &ctx).await.is_ok());
6923 :
6924 2 : Ok(())
6925 1 : }
6926 : #[tokio::test]
6927 1 : async fn test_parent_keeps_data_forever_after_branching() -> anyhow::Result<()> {
6928 1 : let (tenant, ctx) = TenantHarness::create("test_parent_keeps_data_forever_after_branching")
6929 1 : .await?
6930 1 : .load()
6931 1 : .await;
6932 1 : let tline = tenant
6933 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6934 1 : .await?;
6935 1 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6936 :
6937 1 : tenant
6938 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6939 1 : .await?;
6940 1 : let newtline = tenant
6941 1 : .get_timeline(NEW_TIMELINE_ID, true)
6942 1 : .expect("Should have a local timeline");
6943 :
6944 1 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6945 :
6946 : // run gc on parent
6947 1 : tenant
6948 1 : .gc_iteration(
6949 1 : Some(TIMELINE_ID),
6950 1 : 0x10,
6951 1 : Duration::ZERO,
6952 1 : &CancellationToken::new(),
6953 1 : &ctx,
6954 1 : )
6955 1 : .await?;
6956 :
6957 : // Check that the data is still accessible on the branch.
6958 1 : assert_eq!(
6959 1 : newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
6960 1 : test_img(&format!("foo at {}", Lsn(0x40)))
6961 : );
6962 :
6963 2 : Ok(())
6964 1 : }
6965 :
6966 : #[tokio::test]
6967 1 : async fn timeline_load() -> anyhow::Result<()> {
6968 : const TEST_NAME: &str = "timeline_load";
6969 1 : let harness = TenantHarness::create(TEST_NAME).await?;
6970 : {
6971 1 : let (tenant, ctx) = harness.load().await;
6972 1 : let tline = tenant
6973 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x7000), DEFAULT_PG_VERSION, &ctx)
6974 1 : .await?;
6975 1 : make_some_layers(tline.as_ref(), Lsn(0x8000), &ctx).await?;
6976 : // so that all uploads finish & we can call harness.load() below again
6977 1 : tenant
6978 1 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
6979 1 : .instrument(harness.span())
6980 1 : .await
6981 1 : .ok()
6982 1 : .unwrap();
6983 : }
6984 :
6985 1 : let (tenant, _ctx) = harness.load().await;
6986 1 : tenant
6987 1 : .get_timeline(TIMELINE_ID, true)
6988 1 : .expect("cannot load timeline");
6989 :
6990 2 : Ok(())
6991 1 : }
6992 :
6993 : #[tokio::test]
6994 1 : async fn timeline_load_with_ancestor() -> anyhow::Result<()> {
6995 : const TEST_NAME: &str = "timeline_load_with_ancestor";
6996 1 : let harness = TenantHarness::create(TEST_NAME).await?;
6997 : // create two timelines
6998 : {
6999 1 : let (tenant, ctx) = harness.load().await;
7000 1 : let tline = tenant
7001 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7002 1 : .await?;
7003 :
7004 1 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
7005 :
7006 1 : let child_tline = tenant
7007 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
7008 1 : .await?;
7009 1 : child_tline.set_state(TimelineState::Active);
7010 :
7011 1 : let newtline = tenant
7012 1 : .get_timeline(NEW_TIMELINE_ID, true)
7013 1 : .expect("Should have a local timeline");
7014 :
7015 1 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
7016 :
7017 : // so that all uploads finish & we can call harness.load() below again
7018 1 : tenant
7019 1 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
7020 1 : .instrument(harness.span())
7021 1 : .await
7022 1 : .ok()
7023 1 : .unwrap();
7024 : }
7025 :
7026 : // check that both of them are initially unloaded
7027 1 : let (tenant, _ctx) = harness.load().await;
7028 :
7029 : // check that both, child and ancestor are loaded
7030 1 : let _child_tline = tenant
7031 1 : .get_timeline(NEW_TIMELINE_ID, true)
7032 1 : .expect("cannot get child timeline loaded");
7033 :
7034 1 : let _ancestor_tline = tenant
7035 1 : .get_timeline(TIMELINE_ID, true)
7036 1 : .expect("cannot get ancestor timeline loaded");
7037 :
7038 2 : Ok(())
7039 1 : }
7040 :
7041 : #[tokio::test]
7042 1 : async fn delta_layer_dumping() -> anyhow::Result<()> {
7043 : use storage_layer::AsLayerDesc;
7044 1 : let (tenant, ctx) = TenantHarness::create("test_layer_dumping")
7045 1 : .await?
7046 1 : .load()
7047 1 : .await;
7048 1 : let tline = tenant
7049 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7050 1 : .await?;
7051 1 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
7052 :
7053 1 : let layer_map = tline.layers.read(LayerManagerLockHolder::Testing).await;
7054 1 : let level0_deltas = layer_map
7055 1 : .layer_map()?
7056 1 : .level0_deltas()
7057 1 : .iter()
7058 2 : .map(|desc| layer_map.get_from_desc(desc))
7059 1 : .collect::<Vec<_>>();
7060 :
7061 1 : assert!(!level0_deltas.is_empty());
7062 :
7063 3 : for delta in level0_deltas {
7064 1 : // Ensure we are dumping a delta layer here
7065 2 : assert!(delta.layer_desc().is_delta);
7066 2 : delta.dump(true, &ctx).await.unwrap();
7067 1 : }
7068 1 :
7069 1 : Ok(())
7070 1 : }
7071 :
7072 : #[tokio::test]
7073 1 : async fn test_images() -> anyhow::Result<()> {
7074 1 : let (tenant, ctx) = TenantHarness::create("test_images").await?.load().await;
7075 1 : let tline = tenant
7076 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
7077 1 : .await?;
7078 :
7079 1 : let mut writer = tline.writer().await;
7080 1 : writer
7081 1 : .put(
7082 1 : *TEST_KEY,
7083 1 : Lsn(0x10),
7084 1 : &Value::Image(test_img("foo at 0x10")),
7085 1 : &ctx,
7086 1 : )
7087 1 : .await?;
7088 1 : writer.finish_write(Lsn(0x10));
7089 1 : drop(writer);
7090 :
7091 1 : tline.freeze_and_flush().await?;
7092 1 : tline
7093 1 : .compact(&CancellationToken::new(), EnumSet::default(), &ctx)
7094 1 : .await?;
7095 :
7096 1 : let mut writer = tline.writer().await;
7097 1 : writer
7098 1 : .put(
7099 1 : *TEST_KEY,
7100 1 : Lsn(0x20),
7101 1 : &Value::Image(test_img("foo at 0x20")),
7102 1 : &ctx,
7103 1 : )
7104 1 : .await?;
7105 1 : writer.finish_write(Lsn(0x20));
7106 1 : drop(writer);
7107 :
7108 1 : tline.freeze_and_flush().await?;
7109 1 : tline
7110 1 : .compact(&CancellationToken::new(), EnumSet::default(), &ctx)
7111 1 : .await?;
7112 :
7113 1 : let mut writer = tline.writer().await;
7114 1 : writer
7115 1 : .put(
7116 1 : *TEST_KEY,
7117 1 : Lsn(0x30),
7118 1 : &Value::Image(test_img("foo at 0x30")),
7119 1 : &ctx,
7120 1 : )
7121 1 : .await?;
7122 1 : writer.finish_write(Lsn(0x30));
7123 1 : drop(writer);
7124 :
7125 1 : tline.freeze_and_flush().await?;
7126 1 : tline
7127 1 : .compact(&CancellationToken::new(), EnumSet::default(), &ctx)
7128 1 : .await?;
7129 :
7130 1 : let mut writer = tline.writer().await;
7131 1 : writer
7132 1 : .put(
7133 1 : *TEST_KEY,
7134 1 : Lsn(0x40),
7135 1 : &Value::Image(test_img("foo at 0x40")),
7136 1 : &ctx,
7137 1 : )
7138 1 : .await?;
7139 1 : writer.finish_write(Lsn(0x40));
7140 1 : drop(writer);
7141 :
7142 1 : tline.freeze_and_flush().await?;
7143 1 : tline
7144 1 : .compact(&CancellationToken::new(), EnumSet::default(), &ctx)
7145 1 : .await?;
7146 :
7147 1 : assert_eq!(
7148 1 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
7149 1 : test_img("foo at 0x10")
7150 : );
7151 1 : assert_eq!(
7152 1 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
7153 1 : test_img("foo at 0x10")
7154 : );
7155 1 : assert_eq!(
7156 1 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
7157 1 : test_img("foo at 0x20")
7158 : );
7159 1 : assert_eq!(
7160 1 : tline.get(*TEST_KEY, Lsn(0x30), &ctx).await?,
7161 1 : test_img("foo at 0x30")
7162 : );
7163 1 : assert_eq!(
7164 1 : tline.get(*TEST_KEY, Lsn(0x40), &ctx).await?,
7165 1 : test_img("foo at 0x40")
7166 : );
7167 :
7168 2 : Ok(())
7169 1 : }
7170 :
7171 2 : async fn bulk_insert_compact_gc(
7172 2 : tenant: &TenantShard,
7173 2 : timeline: &Arc<Timeline>,
7174 2 : ctx: &RequestContext,
7175 2 : lsn: Lsn,
7176 2 : repeat: usize,
7177 2 : key_count: usize,
7178 2 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
7179 2 : let compact = true;
7180 2 : bulk_insert_maybe_compact_gc(tenant, timeline, ctx, lsn, repeat, key_count, compact).await
7181 2 : }
7182 :
7183 4 : async fn bulk_insert_maybe_compact_gc(
7184 4 : tenant: &TenantShard,
7185 4 : timeline: &Arc<Timeline>,
7186 4 : ctx: &RequestContext,
7187 4 : mut lsn: Lsn,
7188 4 : repeat: usize,
7189 4 : key_count: usize,
7190 4 : compact: bool,
7191 4 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
7192 4 : let mut inserted: HashMap<Key, BTreeSet<Lsn>> = Default::default();
7193 :
7194 4 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7195 4 : let mut blknum = 0;
7196 :
7197 : // Enforce that key range is monotonously increasing
7198 4 : let mut keyspace = KeySpaceAccum::new();
7199 :
7200 4 : let cancel = CancellationToken::new();
7201 :
7202 4 : for _ in 0..repeat {
7203 200 : for _ in 0..key_count {
7204 2000000 : test_key.field6 = blknum;
7205 2000000 : let mut writer = timeline.writer().await;
7206 2000000 : writer
7207 2000000 : .put(
7208 2000000 : test_key,
7209 2000000 : lsn,
7210 2000000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
7211 2000000 : ctx,
7212 2000000 : )
7213 2000000 : .await?;
7214 2000000 : inserted.entry(test_key).or_default().insert(lsn);
7215 2000000 : writer.finish_write(lsn);
7216 2000000 : drop(writer);
7217 :
7218 2000000 : keyspace.add_key(test_key);
7219 :
7220 2000000 : lsn = Lsn(lsn.0 + 0x10);
7221 2000000 : blknum += 1;
7222 : }
7223 :
7224 200 : timeline.freeze_and_flush().await?;
7225 200 : if compact {
7226 : // this requires timeline to be &Arc<Timeline>
7227 100 : timeline.compact(&cancel, EnumSet::default(), ctx).await?;
7228 100 : }
7229 :
7230 : // this doesn't really need to use the timeline_id target, but it is closer to what it
7231 : // originally was.
7232 200 : let res = tenant
7233 200 : .gc_iteration(Some(timeline.timeline_id), 0, Duration::ZERO, &cancel, ctx)
7234 200 : .await?;
7235 :
7236 200 : assert_eq!(res.layers_removed, 0, "this never removes anything");
7237 : }
7238 :
7239 4 : Ok(inserted)
7240 4 : }
7241 :
7242 : //
7243 : // Insert 1000 key-value pairs with increasing keys, flush, compact, GC.
7244 : // Repeat 50 times.
7245 : //
7246 : #[tokio::test]
7247 1 : async fn test_bulk_insert() -> anyhow::Result<()> {
7248 1 : let harness = TenantHarness::create("test_bulk_insert").await?;
7249 1 : let (tenant, ctx) = harness.load().await;
7250 1 : let tline = tenant
7251 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
7252 1 : .await?;
7253 :
7254 1 : let lsn = Lsn(0x10);
7255 1 : bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
7256 :
7257 2 : Ok(())
7258 1 : }
7259 :
7260 : // Test the vectored get real implementation against a simple sequential implementation.
7261 : //
7262 : // The test generates a keyspace by repeatedly flushing the in-memory layer and compacting.
7263 : // Projected to 2D the key space looks like below. Lsn grows upwards on the Y axis and keys
7264 : // grow to the right on the X axis.
7265 : // [Delta]
7266 : // [Delta]
7267 : // [Delta]
7268 : // [Delta]
7269 : // ------------ Image ---------------
7270 : //
7271 : // After layer generation we pick the ranges to query as follows:
7272 : // 1. The beginning of each delta layer
7273 : // 2. At the seam between two adjacent delta layers
7274 : //
7275 : // There's one major downside to this test: delta layers only contains images,
7276 : // so the search can stop at the first delta layer and doesn't traverse any deeper.
7277 : #[tokio::test]
7278 1 : async fn test_get_vectored() -> anyhow::Result<()> {
7279 1 : let harness = TenantHarness::create("test_get_vectored").await?;
7280 1 : let (tenant, ctx) = harness.load().await;
7281 1 : let io_concurrency = IoConcurrency::spawn_for_test();
7282 1 : let tline = tenant
7283 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
7284 1 : .await?;
7285 :
7286 1 : let lsn = Lsn(0x10);
7287 1 : let inserted = bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
7288 :
7289 1 : let guard = tline.layers.read(LayerManagerLockHolder::Testing).await;
7290 1 : let lm = guard.layer_map()?;
7291 :
7292 1 : lm.dump(true, &ctx).await?;
7293 :
7294 1 : let mut reads = Vec::new();
7295 1 : let mut prev = None;
7296 6 : lm.iter_historic_layers().for_each(|desc| {
7297 6 : if !desc.is_delta() {
7298 1 : prev = Some(desc.clone());
7299 1 : return;
7300 5 : }
7301 :
7302 5 : let start = desc.key_range.start;
7303 5 : let end = desc
7304 5 : .key_range
7305 5 : .start
7306 5 : .add(tenant.conf.max_get_vectored_keys.get() as u32);
7307 5 : reads.push(KeySpace {
7308 5 : ranges: vec![start..end],
7309 5 : });
7310 :
7311 5 : if let Some(prev) = &prev {
7312 5 : if !prev.is_delta() {
7313 5 : return;
7314 0 : }
7315 :
7316 0 : let first_range = Key {
7317 0 : field6: prev.key_range.end.field6 - 4,
7318 0 : ..prev.key_range.end
7319 0 : }..prev.key_range.end;
7320 :
7321 0 : let second_range = desc.key_range.start..Key {
7322 0 : field6: desc.key_range.start.field6 + 4,
7323 0 : ..desc.key_range.start
7324 0 : };
7325 :
7326 0 : reads.push(KeySpace {
7327 0 : ranges: vec![first_range, second_range],
7328 0 : });
7329 0 : };
7330 :
7331 0 : prev = Some(desc.clone());
7332 6 : });
7333 :
7334 1 : drop(guard);
7335 :
7336 : // Pick a big LSN such that we query over all the changes.
7337 1 : let reads_lsn = Lsn(u64::MAX - 1);
7338 :
7339 6 : for read in reads {
7340 5 : info!("Doing vectored read on {:?}", read);
7341 1 :
7342 5 : let query = VersionedKeySpaceQuery::uniform(read.clone(), reads_lsn);
7343 1 :
7344 5 : let vectored_res = tline
7345 5 : .get_vectored_impl(
7346 5 : query,
7347 5 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7348 5 : &ctx,
7349 5 : )
7350 5 : .await;
7351 1 :
7352 5 : let mut expected_lsns: HashMap<Key, Lsn> = Default::default();
7353 5 : let mut expect_missing = false;
7354 5 : let mut key = read.start().unwrap();
7355 165 : while key != read.end().unwrap() {
7356 160 : if let Some(lsns) = inserted.get(&key) {
7357 160 : let expected_lsn = lsns.iter().rfind(|lsn| **lsn <= reads_lsn);
7358 160 : match expected_lsn {
7359 160 : Some(lsn) => {
7360 160 : expected_lsns.insert(key, *lsn);
7361 160 : }
7362 1 : None => {
7363 1 : expect_missing = true;
7364 1 : break;
7365 1 : }
7366 1 : }
7367 1 : } else {
7368 1 : expect_missing = true;
7369 1 : break;
7370 1 : }
7371 1 :
7372 160 : key = key.next();
7373 1 : }
7374 1 :
7375 5 : if expect_missing {
7376 1 : assert!(matches!(vectored_res, Err(GetVectoredError::MissingKey(_))));
7377 1 : } else {
7378 160 : for (key, image) in vectored_res? {
7379 160 : let expected_lsn = expected_lsns.get(&key).expect("determined above");
7380 160 : let expected_image = test_img(&format!("{} at {}", key.field6, expected_lsn));
7381 160 : assert_eq!(image?, expected_image);
7382 1 : }
7383 1 : }
7384 1 : }
7385 1 :
7386 1 : Ok(())
7387 1 : }
7388 :
7389 : #[tokio::test]
7390 1 : async fn test_get_vectored_aux_files() -> anyhow::Result<()> {
7391 1 : let harness = TenantHarness::create("test_get_vectored_aux_files").await?;
7392 :
7393 1 : let (tenant, ctx) = harness.load().await;
7394 1 : let io_concurrency = IoConcurrency::spawn_for_test();
7395 1 : let (tline, ctx) = tenant
7396 1 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
7397 1 : .await?;
7398 1 : let tline = tline.raw_timeline().unwrap();
7399 :
7400 1 : let mut modification = tline.begin_modification(Lsn(0x1000));
7401 1 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
7402 1 : modification.set_lsn(Lsn(0x1008))?;
7403 1 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
7404 1 : modification.commit(&ctx).await?;
7405 :
7406 1 : let child_timeline_id = TimelineId::generate();
7407 1 : tenant
7408 1 : .branch_timeline_test(
7409 1 : tline,
7410 1 : child_timeline_id,
7411 1 : Some(tline.get_last_record_lsn()),
7412 1 : &ctx,
7413 1 : )
7414 1 : .await?;
7415 :
7416 1 : let child_timeline = tenant
7417 1 : .get_timeline(child_timeline_id, true)
7418 1 : .expect("Should have the branched timeline");
7419 :
7420 1 : let aux_keyspace = KeySpace {
7421 1 : ranges: vec![NON_INHERITED_RANGE],
7422 1 : };
7423 1 : let read_lsn = child_timeline.get_last_record_lsn();
7424 :
7425 1 : let query = VersionedKeySpaceQuery::uniform(aux_keyspace.clone(), read_lsn);
7426 :
7427 1 : let vectored_res = child_timeline
7428 1 : .get_vectored_impl(
7429 1 : query,
7430 1 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7431 1 : &ctx,
7432 1 : )
7433 1 : .await;
7434 :
7435 1 : let images = vectored_res?;
7436 1 : assert!(images.is_empty());
7437 2 : Ok(())
7438 1 : }
7439 :
7440 : // Test that vectored get handles layer gaps correctly
7441 : // by advancing into the next ancestor timeline if required.
7442 : //
7443 : // The test generates timelines that look like the diagram below.
7444 : // We leave a gap in one of the L1 layers at `gap_at_key` (`/` in the diagram).
7445 : // The reconstruct data for that key lies in the ancestor timeline (`X` in the diagram).
7446 : //
7447 : // ```
7448 : //-------------------------------+
7449 : // ... |
7450 : // [ L1 ] |
7451 : // [ / L1 ] | Child Timeline
7452 : // ... |
7453 : // ------------------------------+
7454 : // [ X L1 ] | Parent Timeline
7455 : // ------------------------------+
7456 : // ```
7457 : #[tokio::test]
7458 1 : async fn test_get_vectored_key_gap() -> anyhow::Result<()> {
7459 1 : let tenant_conf = pageserver_api::models::TenantConfig {
7460 1 : // Make compaction deterministic
7461 1 : gc_period: Some(Duration::ZERO),
7462 1 : compaction_period: Some(Duration::ZERO),
7463 1 : // Encourage creation of L1 layers
7464 1 : checkpoint_distance: Some(16 * 1024),
7465 1 : compaction_target_size: Some(8 * 1024),
7466 1 : ..Default::default()
7467 1 : };
7468 :
7469 1 : let harness = TenantHarness::create_custom(
7470 1 : "test_get_vectored_key_gap",
7471 1 : tenant_conf,
7472 1 : TenantId::generate(),
7473 1 : ShardIdentity::unsharded(),
7474 1 : Generation::new(0xdeadbeef),
7475 1 : )
7476 1 : .await?;
7477 1 : let (tenant, ctx) = harness.load().await;
7478 1 : let io_concurrency = IoConcurrency::spawn_for_test();
7479 :
7480 1 : let mut current_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7481 1 : let gap_at_key = current_key.add(100);
7482 1 : let mut current_lsn = Lsn(0x10);
7483 :
7484 : const KEY_COUNT: usize = 10_000;
7485 :
7486 1 : let timeline_id = TimelineId::generate();
7487 1 : let current_timeline = tenant
7488 1 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
7489 1 : .await?;
7490 :
7491 1 : current_lsn += 0x100;
7492 :
7493 1 : let mut writer = current_timeline.writer().await;
7494 1 : writer
7495 1 : .put(
7496 1 : gap_at_key,
7497 1 : current_lsn,
7498 1 : &Value::Image(test_img(&format!("{gap_at_key} at {current_lsn}"))),
7499 1 : &ctx,
7500 1 : )
7501 1 : .await?;
7502 1 : writer.finish_write(current_lsn);
7503 1 : drop(writer);
7504 :
7505 1 : let mut latest_lsns = HashMap::new();
7506 1 : latest_lsns.insert(gap_at_key, current_lsn);
7507 :
7508 1 : current_timeline.freeze_and_flush().await?;
7509 :
7510 1 : let child_timeline_id = TimelineId::generate();
7511 :
7512 1 : tenant
7513 1 : .branch_timeline_test(
7514 1 : ¤t_timeline,
7515 1 : child_timeline_id,
7516 1 : Some(current_lsn),
7517 1 : &ctx,
7518 1 : )
7519 1 : .await?;
7520 1 : let child_timeline = tenant
7521 1 : .get_timeline(child_timeline_id, true)
7522 1 : .expect("Should have the branched timeline");
7523 :
7524 10001 : for i in 0..KEY_COUNT {
7525 10000 : if current_key == gap_at_key {
7526 1 : current_key = current_key.next();
7527 1 : continue;
7528 9999 : }
7529 :
7530 9999 : current_lsn += 0x10;
7531 :
7532 9999 : let mut writer = child_timeline.writer().await;
7533 9999 : writer
7534 9999 : .put(
7535 9999 : current_key,
7536 9999 : current_lsn,
7537 9999 : &Value::Image(test_img(&format!("{current_key} at {current_lsn}"))),
7538 9999 : &ctx,
7539 9999 : )
7540 9999 : .await?;
7541 9999 : writer.finish_write(current_lsn);
7542 9999 : drop(writer);
7543 :
7544 9999 : latest_lsns.insert(current_key, current_lsn);
7545 9999 : current_key = current_key.next();
7546 :
7547 : // Flush every now and then to encourage layer file creation.
7548 9999 : if i % 500 == 0 {
7549 20 : child_timeline.freeze_and_flush().await?;
7550 9979 : }
7551 : }
7552 :
7553 1 : child_timeline.freeze_and_flush().await?;
7554 1 : let mut flags = EnumSet::new();
7555 1 : flags.insert(CompactFlags::ForceRepartition);
7556 1 : child_timeline
7557 1 : .compact(&CancellationToken::new(), flags, &ctx)
7558 1 : .await?;
7559 :
7560 1 : let key_near_end = {
7561 1 : let mut tmp = current_key;
7562 1 : tmp.field6 -= 10;
7563 1 : tmp
7564 : };
7565 :
7566 1 : let key_near_gap = {
7567 1 : let mut tmp = gap_at_key;
7568 1 : tmp.field6 -= 10;
7569 1 : tmp
7570 : };
7571 :
7572 1 : let read = KeySpace {
7573 1 : ranges: vec![key_near_gap..gap_at_key.next(), key_near_end..current_key],
7574 1 : };
7575 :
7576 1 : let query = VersionedKeySpaceQuery::uniform(read.clone(), current_lsn);
7577 :
7578 1 : let results = child_timeline
7579 1 : .get_vectored_impl(
7580 1 : query,
7581 1 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7582 1 : &ctx,
7583 1 : )
7584 1 : .await?;
7585 :
7586 22 : for (key, img_res) in results {
7587 21 : let expected = test_img(&format!("{} at {}", key, latest_lsns[&key]));
7588 21 : assert_eq!(img_res?, expected);
7589 1 : }
7590 1 :
7591 1 : Ok(())
7592 1 : }
7593 :
7594 : // Test that vectored get descends into ancestor timelines correctly and
7595 : // does not return an image that's newer than requested.
7596 : //
7597 : // The diagram below ilustrates an interesting case. We have a parent timeline
7598 : // (top of the Lsn range) and a child timeline. The request key cannot be reconstructed
7599 : // from the child timeline, so the parent timeline must be visited. When advacing into
7600 : // the child timeline, the read path needs to remember what the requested Lsn was in
7601 : // order to avoid returning an image that's too new. The test below constructs such
7602 : // a timeline setup and does a few queries around the Lsn of each page image.
7603 : // ```
7604 : // LSN
7605 : // ^
7606 : // |
7607 : // |
7608 : // 500 | --------------------------------------> branch point
7609 : // 400 | X
7610 : // 300 | X
7611 : // 200 | --------------------------------------> requested lsn
7612 : // 100 | X
7613 : // |---------------------------------------> Key
7614 : // |
7615 : // ------> requested key
7616 : //
7617 : // Legend:
7618 : // * X - page images
7619 : // ```
7620 : #[tokio::test]
7621 1 : async fn test_get_vectored_ancestor_descent() -> anyhow::Result<()> {
7622 1 : let harness = TenantHarness::create("test_get_vectored_on_lsn_axis").await?;
7623 1 : let (tenant, ctx) = harness.load().await;
7624 1 : let io_concurrency = IoConcurrency::spawn_for_test();
7625 :
7626 1 : let start_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7627 1 : let end_key = start_key.add(1000);
7628 1 : let child_gap_at_key = start_key.add(500);
7629 1 : let mut parent_gap_lsns: BTreeMap<Lsn, String> = BTreeMap::new();
7630 :
7631 1 : let mut current_lsn = Lsn(0x10);
7632 :
7633 1 : let timeline_id = TimelineId::generate();
7634 1 : let parent_timeline = tenant
7635 1 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
7636 1 : .await?;
7637 :
7638 1 : current_lsn += 0x100;
7639 :
7640 4 : for _ in 0..3 {
7641 3 : let mut key = start_key;
7642 3003 : while key < end_key {
7643 3000 : current_lsn += 0x10;
7644 :
7645 3000 : let image_value = format!("{child_gap_at_key} at {current_lsn}");
7646 :
7647 3000 : let mut writer = parent_timeline.writer().await;
7648 3000 : writer
7649 3000 : .put(
7650 3000 : key,
7651 3000 : current_lsn,
7652 3000 : &Value::Image(test_img(&image_value)),
7653 3000 : &ctx,
7654 3000 : )
7655 3000 : .await?;
7656 3000 : writer.finish_write(current_lsn);
7657 :
7658 3000 : if key == child_gap_at_key {
7659 3 : parent_gap_lsns.insert(current_lsn, image_value);
7660 2997 : }
7661 :
7662 3000 : key = key.next();
7663 : }
7664 :
7665 3 : parent_timeline.freeze_and_flush().await?;
7666 : }
7667 :
7668 1 : let child_timeline_id = TimelineId::generate();
7669 :
7670 1 : let child_timeline = tenant
7671 1 : .branch_timeline_test(&parent_timeline, child_timeline_id, Some(current_lsn), &ctx)
7672 1 : .await?;
7673 :
7674 1 : let mut key = start_key;
7675 1001 : while key < end_key {
7676 1000 : if key == child_gap_at_key {
7677 1 : key = key.next();
7678 1 : continue;
7679 999 : }
7680 :
7681 999 : current_lsn += 0x10;
7682 :
7683 999 : let mut writer = child_timeline.writer().await;
7684 999 : writer
7685 999 : .put(
7686 999 : key,
7687 999 : current_lsn,
7688 999 : &Value::Image(test_img(&format!("{key} at {current_lsn}"))),
7689 999 : &ctx,
7690 999 : )
7691 999 : .await?;
7692 999 : writer.finish_write(current_lsn);
7693 :
7694 999 : key = key.next();
7695 : }
7696 :
7697 1 : child_timeline.freeze_and_flush().await?;
7698 :
7699 1 : let lsn_offsets: [i64; 5] = [-10, -1, 0, 1, 10];
7700 1 : let mut query_lsns = Vec::new();
7701 3 : for image_lsn in parent_gap_lsns.keys().rev() {
7702 18 : for offset in lsn_offsets {
7703 15 : query_lsns.push(Lsn(image_lsn
7704 15 : .0
7705 15 : .checked_add_signed(offset)
7706 15 : .expect("Shouldn't overflow")));
7707 15 : }
7708 1 : }
7709 1 :
7710 16 : for query_lsn in query_lsns {
7711 15 : let query = VersionedKeySpaceQuery::uniform(
7712 15 : KeySpace {
7713 15 : ranges: vec![child_gap_at_key..child_gap_at_key.next()],
7714 15 : },
7715 15 : query_lsn,
7716 1 : );
7717 1 :
7718 15 : let results = child_timeline
7719 15 : .get_vectored_impl(
7720 15 : query,
7721 15 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7722 15 : &ctx,
7723 15 : )
7724 15 : .await;
7725 1 :
7726 15 : let expected_item = parent_gap_lsns
7727 15 : .iter()
7728 15 : .rev()
7729 34 : .find(|(lsn, _)| **lsn <= query_lsn);
7730 1 :
7731 15 : info!(
7732 1 : "Doing vectored read at LSN {}. Expecting image to be: {:?}",
7733 1 : query_lsn, expected_item
7734 1 : );
7735 1 :
7736 15 : match expected_item {
7737 13 : Some((_, img_value)) => {
7738 13 : let key_results = results.expect("No vectored get error expected");
7739 13 : let key_result = &key_results[&child_gap_at_key];
7740 13 : let returned_img = key_result
7741 13 : .as_ref()
7742 13 : .expect("No page reconstruct error expected");
7743 1 :
7744 13 : info!(
7745 1 : "Vectored read at LSN {} returned image {}",
7746 1 : query_lsn,
7747 1 : std::str::from_utf8(returned_img)?
7748 1 : );
7749 13 : assert_eq!(*returned_img, test_img(img_value));
7750 1 : }
7751 1 : None => {
7752 2 : assert!(matches!(results, Err(GetVectoredError::MissingKey(_))));
7753 1 : }
7754 1 : }
7755 1 : }
7756 1 :
7757 1 : Ok(())
7758 1 : }
7759 :
7760 : #[tokio::test]
7761 1 : async fn test_random_updates() -> anyhow::Result<()> {
7762 1 : let names_algorithms = [
7763 1 : ("test_random_updates_legacy", CompactionAlgorithm::Legacy),
7764 1 : ("test_random_updates_tiered", CompactionAlgorithm::Tiered),
7765 1 : ];
7766 3 : for (name, algorithm) in names_algorithms {
7767 2 : test_random_updates_algorithm(name, algorithm).await?;
7768 1 : }
7769 1 : Ok(())
7770 1 : }
7771 :
7772 2 : async fn test_random_updates_algorithm(
7773 2 : name: &'static str,
7774 2 : compaction_algorithm: CompactionAlgorithm,
7775 2 : ) -> anyhow::Result<()> {
7776 2 : let mut harness = TenantHarness::create(name).await?;
7777 2 : harness.tenant_conf.compaction_algorithm = Some(CompactionAlgorithmSettings {
7778 2 : kind: compaction_algorithm,
7779 2 : });
7780 2 : let (tenant, ctx) = harness.load().await;
7781 2 : let tline = tenant
7782 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7783 2 : .await?;
7784 :
7785 : const NUM_KEYS: usize = 1000;
7786 2 : let cancel = CancellationToken::new();
7787 :
7788 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7789 2 : let mut test_key_end = test_key;
7790 2 : test_key_end.field6 = NUM_KEYS as u32;
7791 2 : tline.add_extra_test_dense_keyspace(KeySpace::single(test_key..test_key_end));
7792 :
7793 2 : let mut keyspace = KeySpaceAccum::new();
7794 :
7795 : // Track when each page was last modified. Used to assert that
7796 : // a read sees the latest page version.
7797 2 : let mut updated = [Lsn(0); NUM_KEYS];
7798 :
7799 2 : let mut lsn = Lsn(0x10);
7800 : #[allow(clippy::needless_range_loop)]
7801 2002 : for blknum in 0..NUM_KEYS {
7802 2000 : lsn = Lsn(lsn.0 + 0x10);
7803 2000 : test_key.field6 = blknum as u32;
7804 2000 : let mut writer = tline.writer().await;
7805 2000 : writer
7806 2000 : .put(
7807 2000 : test_key,
7808 2000 : lsn,
7809 2000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
7810 2000 : &ctx,
7811 2000 : )
7812 2000 : .await?;
7813 2000 : writer.finish_write(lsn);
7814 2000 : updated[blknum] = lsn;
7815 2000 : drop(writer);
7816 :
7817 2000 : keyspace.add_key(test_key);
7818 : }
7819 :
7820 102 : for _ in 0..50 {
7821 100100 : for _ in 0..NUM_KEYS {
7822 100000 : lsn = Lsn(lsn.0 + 0x10);
7823 100000 : let blknum = rand::rng().random_range(0..NUM_KEYS);
7824 100000 : test_key.field6 = blknum as u32;
7825 100000 : let mut writer = tline.writer().await;
7826 100000 : writer
7827 100000 : .put(
7828 100000 : test_key,
7829 100000 : lsn,
7830 100000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
7831 100000 : &ctx,
7832 100000 : )
7833 100000 : .await?;
7834 100000 : writer.finish_write(lsn);
7835 100000 : drop(writer);
7836 100000 : updated[blknum] = lsn;
7837 : }
7838 :
7839 : // Read all the blocks
7840 100000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7841 100000 : test_key.field6 = blknum as u32;
7842 100000 : assert_eq!(
7843 100000 : tline.get(test_key, lsn, &ctx).await?,
7844 100000 : test_img(&format!("{blknum} at {last_lsn}"))
7845 : );
7846 : }
7847 :
7848 : // Perform a cycle of flush, and GC
7849 100 : tline.freeze_and_flush().await?;
7850 100 : tenant
7851 100 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7852 100 : .await?;
7853 : }
7854 :
7855 2 : Ok(())
7856 2 : }
7857 :
7858 : #[tokio::test]
7859 1 : async fn test_traverse_branches() -> anyhow::Result<()> {
7860 1 : let (tenant, ctx) = TenantHarness::create("test_traverse_branches")
7861 1 : .await?
7862 1 : .load()
7863 1 : .await;
7864 1 : let mut tline = tenant
7865 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7866 1 : .await?;
7867 :
7868 : const NUM_KEYS: usize = 1000;
7869 :
7870 1 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7871 :
7872 1 : let mut keyspace = KeySpaceAccum::new();
7873 :
7874 1 : let cancel = CancellationToken::new();
7875 :
7876 : // Track when each page was last modified. Used to assert that
7877 : // a read sees the latest page version.
7878 1 : let mut updated = [Lsn(0); NUM_KEYS];
7879 :
7880 1 : let mut lsn = Lsn(0x10);
7881 1 : #[allow(clippy::needless_range_loop)]
7882 1001 : for blknum in 0..NUM_KEYS {
7883 1000 : lsn = Lsn(lsn.0 + 0x10);
7884 1000 : test_key.field6 = blknum as u32;
7885 1000 : let mut writer = tline.writer().await;
7886 1000 : writer
7887 1000 : .put(
7888 1000 : test_key,
7889 1000 : lsn,
7890 1000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
7891 1000 : &ctx,
7892 1000 : )
7893 1000 : .await?;
7894 1000 : writer.finish_write(lsn);
7895 1000 : updated[blknum] = lsn;
7896 1000 : drop(writer);
7897 1 :
7898 1000 : keyspace.add_key(test_key);
7899 1 : }
7900 1 :
7901 51 : for _ in 0..50 {
7902 50 : let new_tline_id = TimelineId::generate();
7903 50 : tenant
7904 50 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7905 50 : .await?;
7906 50 : tline = tenant
7907 50 : .get_timeline(new_tline_id, true)
7908 50 : .expect("Should have the branched timeline");
7909 1 :
7910 50050 : for _ in 0..NUM_KEYS {
7911 50000 : lsn = Lsn(lsn.0 + 0x10);
7912 50000 : let blknum = rand::rng().random_range(0..NUM_KEYS);
7913 50000 : test_key.field6 = blknum as u32;
7914 50000 : let mut writer = tline.writer().await;
7915 50000 : writer
7916 50000 : .put(
7917 50000 : test_key,
7918 50000 : lsn,
7919 50000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
7920 50000 : &ctx,
7921 50000 : )
7922 50000 : .await?;
7923 50000 : println!("updating {blknum} at {lsn}");
7924 50000 : writer.finish_write(lsn);
7925 50000 : drop(writer);
7926 50000 : updated[blknum] = lsn;
7927 1 : }
7928 1 :
7929 1 : // Read all the blocks
7930 50000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7931 50000 : test_key.field6 = blknum as u32;
7932 50000 : assert_eq!(
7933 50000 : tline.get(test_key, lsn, &ctx).await?,
7934 50000 : test_img(&format!("{blknum} at {last_lsn}"))
7935 1 : );
7936 1 : }
7937 1 :
7938 1 : // Perform a cycle of flush, compact, and GC
7939 50 : tline.freeze_and_flush().await?;
7940 50 : tline.compact(&cancel, EnumSet::default(), &ctx).await?;
7941 50 : tenant
7942 50 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7943 50 : .await?;
7944 1 : }
7945 1 :
7946 1 : Ok(())
7947 1 : }
7948 :
7949 : #[tokio::test]
7950 1 : async fn test_traverse_ancestors() -> anyhow::Result<()> {
7951 1 : let (tenant, ctx) = TenantHarness::create("test_traverse_ancestors")
7952 1 : .await?
7953 1 : .load()
7954 1 : .await;
7955 1 : let mut tline = tenant
7956 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7957 1 : .await?;
7958 :
7959 : const NUM_KEYS: usize = 100;
7960 : const NUM_TLINES: usize = 50;
7961 :
7962 1 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7963 : // Track page mutation lsns across different timelines.
7964 1 : let mut updated = [[Lsn(0); NUM_KEYS]; NUM_TLINES];
7965 :
7966 1 : let mut lsn = Lsn(0x10);
7967 :
7968 1 : #[allow(clippy::needless_range_loop)]
7969 51 : for idx in 0..NUM_TLINES {
7970 50 : let new_tline_id = TimelineId::generate();
7971 50 : tenant
7972 50 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7973 50 : .await?;
7974 50 : tline = tenant
7975 50 : .get_timeline(new_tline_id, true)
7976 50 : .expect("Should have the branched timeline");
7977 1 :
7978 5050 : for _ in 0..NUM_KEYS {
7979 5000 : lsn = Lsn(lsn.0 + 0x10);
7980 5000 : let blknum = rand::rng().random_range(0..NUM_KEYS);
7981 5000 : test_key.field6 = blknum as u32;
7982 5000 : let mut writer = tline.writer().await;
7983 5000 : writer
7984 5000 : .put(
7985 5000 : test_key,
7986 5000 : lsn,
7987 5000 : &Value::Image(test_img(&format!("{idx} {blknum} at {lsn}"))),
7988 5000 : &ctx,
7989 5000 : )
7990 5000 : .await?;
7991 5000 : println!("updating [{idx}][{blknum}] at {lsn}");
7992 5000 : writer.finish_write(lsn);
7993 5000 : drop(writer);
7994 5000 : updated[idx][blknum] = lsn;
7995 1 : }
7996 1 : }
7997 1 :
7998 1 : // Read pages from leaf timeline across all ancestors.
7999 50 : for (idx, lsns) in updated.iter().enumerate() {
8000 5000 : for (blknum, lsn) in lsns.iter().enumerate() {
8001 1 : // Skip empty mutations.
8002 5000 : if lsn.0 == 0 {
8003 1815 : continue;
8004 3185 : }
8005 3185 : println!("checking [{idx}][{blknum}] at {lsn}");
8006 3185 : test_key.field6 = blknum as u32;
8007 3185 : assert_eq!(
8008 3185 : tline.get(test_key, *lsn, &ctx).await?,
8009 3185 : test_img(&format!("{idx} {blknum} at {lsn}"))
8010 1 : );
8011 1 : }
8012 1 : }
8013 1 : Ok(())
8014 1 : }
8015 :
8016 : #[tokio::test]
8017 1 : async fn test_write_at_initdb_lsn_takes_optimization_code_path() -> anyhow::Result<()> {
8018 1 : let (tenant, ctx) = TenantHarness::create("test_empty_test_timeline_is_usable")
8019 1 : .await?
8020 1 : .load()
8021 1 : .await;
8022 :
8023 1 : let initdb_lsn = Lsn(0x20);
8024 1 : let (utline, ctx) = tenant
8025 1 : .create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)
8026 1 : .await?;
8027 1 : let tline = utline.raw_timeline().unwrap();
8028 :
8029 : // Spawn flush loop now so that we can set the `expect_initdb_optimization`
8030 1 : tline.maybe_spawn_flush_loop();
8031 :
8032 : // Make sure the timeline has the minimum set of required keys for operation.
8033 : // The only operation you can always do on an empty timeline is to `put` new data.
8034 : // Except if you `put` at `initdb_lsn`.
8035 : // In that case, there's an optimization to directly create image layers instead of delta layers.
8036 : // It uses `repartition()`, which assumes some keys to be present.
8037 : // Let's make sure the test timeline can handle that case.
8038 : {
8039 1 : let mut state = tline.flush_loop_state.lock().unwrap();
8040 1 : assert_eq!(
8041 : timeline::FlushLoopState::Running {
8042 : expect_initdb_optimization: false,
8043 : initdb_optimization_count: 0,
8044 : },
8045 1 : *state
8046 : );
8047 1 : *state = timeline::FlushLoopState::Running {
8048 1 : expect_initdb_optimization: true,
8049 1 : initdb_optimization_count: 0,
8050 1 : };
8051 : }
8052 :
8053 : // Make writes at the initdb_lsn. When we flush it below, it should be handled by the optimization.
8054 : // As explained above, the optimization requires some keys to be present.
8055 : // As per `create_empty_timeline` documentation, use init_empty to set them.
8056 : // This is what `create_test_timeline` does, by the way.
8057 1 : let mut modification = tline.begin_modification(initdb_lsn);
8058 1 : modification
8059 1 : .init_empty_test_timeline()
8060 1 : .context("init_empty_test_timeline")?;
8061 1 : modification
8062 1 : .commit(&ctx)
8063 1 : .await
8064 1 : .context("commit init_empty_test_timeline modification")?;
8065 :
8066 : // Do the flush. The flush code will check the expectations that we set above.
8067 1 : tline.freeze_and_flush().await?;
8068 :
8069 : // assert freeze_and_flush exercised the initdb optimization
8070 1 : {
8071 1 : let state = tline.flush_loop_state.lock().unwrap();
8072 1 : let timeline::FlushLoopState::Running {
8073 1 : expect_initdb_optimization,
8074 1 : initdb_optimization_count,
8075 1 : } = *state
8076 1 : else {
8077 1 : panic!("unexpected state: {:?}", *state);
8078 1 : };
8079 1 : assert!(expect_initdb_optimization);
8080 1 : assert!(initdb_optimization_count > 0);
8081 1 : }
8082 1 : Ok(())
8083 1 : }
8084 :
8085 : #[tokio::test]
8086 1 : async fn test_create_guard_crash() -> anyhow::Result<()> {
8087 1 : let name = "test_create_guard_crash";
8088 1 : let harness = TenantHarness::create(name).await?;
8089 : {
8090 1 : let (tenant, ctx) = harness.load().await;
8091 1 : let (tline, _ctx) = tenant
8092 1 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
8093 1 : .await?;
8094 : // Leave the timeline ID in [`TenantShard::timelines_creating`] to exclude attempting to create it again
8095 1 : let raw_tline = tline.raw_timeline().unwrap();
8096 1 : raw_tline
8097 1 : .shutdown(super::timeline::ShutdownMode::Hard)
8098 1 : .instrument(info_span!("test_shutdown", tenant_id=%raw_tline.tenant_shard_id, shard_id=%raw_tline.tenant_shard_id.shard_slug(), timeline_id=%TIMELINE_ID))
8099 1 : .await;
8100 1 : std::mem::forget(tline);
8101 : }
8102 :
8103 1 : let (tenant, _) = harness.load().await;
8104 1 : match tenant.get_timeline(TIMELINE_ID, false) {
8105 0 : Ok(_) => panic!("timeline should've been removed during load"),
8106 1 : Err(e) => {
8107 1 : assert_eq!(
8108 : e,
8109 1 : GetTimelineError::NotFound {
8110 1 : tenant_id: tenant.tenant_shard_id,
8111 1 : timeline_id: TIMELINE_ID,
8112 1 : }
8113 : )
8114 : }
8115 : }
8116 :
8117 1 : assert!(
8118 1 : !harness
8119 1 : .conf
8120 1 : .timeline_path(&tenant.tenant_shard_id, &TIMELINE_ID)
8121 1 : .exists()
8122 : );
8123 :
8124 2 : Ok(())
8125 1 : }
8126 :
8127 : #[tokio::test]
8128 1 : async fn test_read_at_max_lsn() -> anyhow::Result<()> {
8129 1 : let names_algorithms = [
8130 1 : ("test_read_at_max_lsn_legacy", CompactionAlgorithm::Legacy),
8131 1 : ("test_read_at_max_lsn_tiered", CompactionAlgorithm::Tiered),
8132 1 : ];
8133 3 : for (name, algorithm) in names_algorithms {
8134 2 : test_read_at_max_lsn_algorithm(name, algorithm).await?;
8135 1 : }
8136 1 : Ok(())
8137 1 : }
8138 :
8139 2 : async fn test_read_at_max_lsn_algorithm(
8140 2 : name: &'static str,
8141 2 : compaction_algorithm: CompactionAlgorithm,
8142 2 : ) -> anyhow::Result<()> {
8143 2 : let mut harness = TenantHarness::create(name).await?;
8144 2 : harness.tenant_conf.compaction_algorithm = Some(CompactionAlgorithmSettings {
8145 2 : kind: compaction_algorithm,
8146 2 : });
8147 2 : let (tenant, ctx) = harness.load().await;
8148 2 : let tline = tenant
8149 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
8150 2 : .await?;
8151 :
8152 2 : let lsn = Lsn(0x10);
8153 2 : let compact = false;
8154 2 : bulk_insert_maybe_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000, compact).await?;
8155 :
8156 2 : let test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
8157 2 : let read_lsn = Lsn(u64::MAX - 1);
8158 :
8159 2 : let result = tline.get(test_key, read_lsn, &ctx).await;
8160 2 : assert!(result.is_ok(), "result is not Ok: {}", result.unwrap_err());
8161 :
8162 2 : Ok(())
8163 2 : }
8164 :
8165 : #[tokio::test]
8166 1 : async fn test_metadata_scan() -> anyhow::Result<()> {
8167 1 : let harness = TenantHarness::create("test_metadata_scan").await?;
8168 1 : let (tenant, ctx) = harness.load().await;
8169 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8170 1 : let tline = tenant
8171 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
8172 1 : .await?;
8173 :
8174 : const NUM_KEYS: usize = 1000;
8175 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
8176 :
8177 1 : let cancel = CancellationToken::new();
8178 :
8179 1 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
8180 1 : base_key.field1 = AUX_KEY_PREFIX;
8181 1 : let mut test_key = base_key;
8182 :
8183 : // Track when each page was last modified. Used to assert that
8184 : // a read sees the latest page version.
8185 1 : let mut updated = [Lsn(0); NUM_KEYS];
8186 :
8187 1 : let mut lsn = Lsn(0x10);
8188 : #[allow(clippy::needless_range_loop)]
8189 1001 : for blknum in 0..NUM_KEYS {
8190 1000 : lsn = Lsn(lsn.0 + 0x10);
8191 1000 : test_key.field6 = (blknum * STEP) as u32;
8192 1000 : let mut writer = tline.writer().await;
8193 1000 : writer
8194 1000 : .put(
8195 1000 : test_key,
8196 1000 : lsn,
8197 1000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
8198 1000 : &ctx,
8199 1000 : )
8200 1000 : .await?;
8201 1000 : writer.finish_write(lsn);
8202 1000 : updated[blknum] = lsn;
8203 1000 : drop(writer);
8204 : }
8205 :
8206 1 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
8207 :
8208 12 : for iter in 0..=10 {
8209 1 : // Read all the blocks
8210 11000 : for (blknum, last_lsn) in updated.iter().enumerate() {
8211 11000 : test_key.field6 = (blknum * STEP) as u32;
8212 11000 : assert_eq!(
8213 11000 : tline.get(test_key, lsn, &ctx).await?,
8214 11000 : test_img(&format!("{blknum} at {last_lsn}"))
8215 1 : );
8216 1 : }
8217 1 :
8218 11 : let mut cnt = 0;
8219 11 : let query = VersionedKeySpaceQuery::uniform(keyspace.clone(), lsn);
8220 1 :
8221 11000 : for (key, value) in tline
8222 11 : .get_vectored_impl(
8223 11 : query,
8224 11 : &mut ValuesReconstructState::new(io_concurrency.clone()),
8225 11 : &ctx,
8226 11 : )
8227 11 : .await?
8228 1 : {
8229 11000 : let blknum = key.field6 as usize;
8230 11000 : let value = value?;
8231 11000 : assert!(blknum % STEP == 0);
8232 11000 : let blknum = blknum / STEP;
8233 11000 : assert_eq!(
8234 1 : value,
8235 11000 : test_img(&format!("{} at {}", blknum, updated[blknum]))
8236 1 : );
8237 11000 : cnt += 1;
8238 1 : }
8239 1 :
8240 11 : assert_eq!(cnt, NUM_KEYS);
8241 1 :
8242 11011 : for _ in 0..NUM_KEYS {
8243 11000 : lsn = Lsn(lsn.0 + 0x10);
8244 11000 : let blknum = rand::rng().random_range(0..NUM_KEYS);
8245 11000 : test_key.field6 = (blknum * STEP) as u32;
8246 11000 : let mut writer = tline.writer().await;
8247 11000 : writer
8248 11000 : .put(
8249 11000 : test_key,
8250 11000 : lsn,
8251 11000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
8252 11000 : &ctx,
8253 11000 : )
8254 11000 : .await?;
8255 11000 : writer.finish_write(lsn);
8256 11000 : drop(writer);
8257 11000 : updated[blknum] = lsn;
8258 1 : }
8259 1 :
8260 1 : // Perform two cycles of flush, compact, and GC
8261 33 : for round in 0..2 {
8262 22 : tline.freeze_and_flush().await?;
8263 22 : tline
8264 22 : .compact(
8265 22 : &cancel,
8266 22 : if iter % 5 == 0 && round == 0 {
8267 3 : let mut flags = EnumSet::new();
8268 3 : flags.insert(CompactFlags::ForceImageLayerCreation);
8269 3 : flags.insert(CompactFlags::ForceRepartition);
8270 3 : flags
8271 1 : } else {
8272 19 : EnumSet::empty()
8273 1 : },
8274 22 : &ctx,
8275 1 : )
8276 22 : .await?;
8277 22 : tenant
8278 22 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
8279 22 : .await?;
8280 1 : }
8281 1 : }
8282 1 :
8283 1 : Ok(())
8284 1 : }
8285 :
8286 : #[tokio::test]
8287 1 : async fn test_metadata_compaction_trigger() -> anyhow::Result<()> {
8288 1 : let harness = TenantHarness::create("test_metadata_compaction_trigger").await?;
8289 1 : let (tenant, ctx) = harness.load().await;
8290 1 : let tline = tenant
8291 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
8292 1 : .await?;
8293 :
8294 1 : let cancel = CancellationToken::new();
8295 :
8296 1 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
8297 1 : base_key.field1 = AUX_KEY_PREFIX;
8298 1 : let test_key = base_key;
8299 1 : let mut lsn = Lsn(0x10);
8300 :
8301 21 : for _ in 0..20 {
8302 20 : lsn = Lsn(lsn.0 + 0x10);
8303 20 : let mut writer = tline.writer().await;
8304 20 : writer
8305 20 : .put(
8306 20 : test_key,
8307 20 : lsn,
8308 20 : &Value::Image(test_img(&format!("{} at {}", 0, lsn))),
8309 20 : &ctx,
8310 20 : )
8311 20 : .await?;
8312 20 : writer.finish_write(lsn);
8313 20 : drop(writer);
8314 20 : tline.freeze_and_flush().await?; // force create a delta layer
8315 : }
8316 :
8317 1 : let before_num_l0_delta_files = tline
8318 1 : .layers
8319 1 : .read(LayerManagerLockHolder::Testing)
8320 1 : .await
8321 1 : .layer_map()?
8322 1 : .level0_deltas()
8323 1 : .len();
8324 :
8325 1 : tline.compact(&cancel, EnumSet::default(), &ctx).await?;
8326 :
8327 1 : let after_num_l0_delta_files = tline
8328 1 : .layers
8329 1 : .read(LayerManagerLockHolder::Testing)
8330 1 : .await
8331 1 : .layer_map()?
8332 1 : .level0_deltas()
8333 1 : .len();
8334 :
8335 1 : assert!(
8336 1 : after_num_l0_delta_files < before_num_l0_delta_files,
8337 0 : "after_num_l0_delta_files={after_num_l0_delta_files}, before_num_l0_delta_files={before_num_l0_delta_files}"
8338 : );
8339 :
8340 1 : assert_eq!(
8341 1 : tline.get(test_key, lsn, &ctx).await?,
8342 1 : test_img(&format!("{} at {}", 0, lsn))
8343 : );
8344 :
8345 2 : Ok(())
8346 1 : }
8347 :
8348 : #[tokio::test]
8349 1 : async fn test_aux_file_e2e() {
8350 1 : let harness = TenantHarness::create("test_aux_file_e2e").await.unwrap();
8351 :
8352 1 : let (tenant, ctx) = harness.load().await;
8353 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8354 :
8355 1 : let mut lsn = Lsn(0x08);
8356 :
8357 1 : let tline: Arc<Timeline> = tenant
8358 1 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
8359 1 : .await
8360 1 : .unwrap();
8361 :
8362 : {
8363 1 : lsn += 8;
8364 1 : let mut modification = tline.begin_modification(lsn);
8365 1 : modification
8366 1 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
8367 1 : .await
8368 1 : .unwrap();
8369 1 : modification.commit(&ctx).await.unwrap();
8370 : }
8371 :
8372 : // we can read everything from the storage
8373 1 : let files = tline
8374 1 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
8375 1 : .await
8376 1 : .unwrap();
8377 1 : assert_eq!(
8378 1 : files.get("pg_logical/mappings/test1"),
8379 1 : Some(&bytes::Bytes::from_static(b"first"))
8380 : );
8381 :
8382 : {
8383 1 : lsn += 8;
8384 1 : let mut modification = tline.begin_modification(lsn);
8385 1 : modification
8386 1 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
8387 1 : .await
8388 1 : .unwrap();
8389 1 : modification.commit(&ctx).await.unwrap();
8390 : }
8391 :
8392 1 : let files = tline
8393 1 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
8394 1 : .await
8395 1 : .unwrap();
8396 1 : assert_eq!(
8397 1 : files.get("pg_logical/mappings/test2"),
8398 1 : Some(&bytes::Bytes::from_static(b"second"))
8399 : );
8400 :
8401 1 : let child = tenant
8402 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(lsn), &ctx)
8403 1 : .await
8404 1 : .unwrap();
8405 :
8406 1 : let files = child
8407 1 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
8408 1 : .await
8409 1 : .unwrap();
8410 1 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
8411 1 : assert_eq!(files.get("pg_logical/mappings/test2"), None);
8412 1 : }
8413 :
8414 : #[tokio::test]
8415 1 : async fn test_repl_origin_tombstones() {
8416 1 : let harness = TenantHarness::create("test_repl_origin_tombstones")
8417 1 : .await
8418 1 : .unwrap();
8419 :
8420 1 : let (tenant, ctx) = harness.load().await;
8421 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8422 :
8423 1 : let mut lsn = Lsn(0x08);
8424 :
8425 1 : let tline: Arc<Timeline> = tenant
8426 1 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
8427 1 : .await
8428 1 : .unwrap();
8429 :
8430 1 : let repl_lsn = Lsn(0x10);
8431 : {
8432 1 : lsn += 8;
8433 1 : let mut modification = tline.begin_modification(lsn);
8434 1 : modification.put_for_unit_test(repl_origin_key(2), Value::Image(Bytes::new()));
8435 1 : modification.set_replorigin(1, repl_lsn).await.unwrap();
8436 1 : modification.commit(&ctx).await.unwrap();
8437 : }
8438 :
8439 : // we can read everything from the storage
8440 1 : let repl_origins = tline
8441 1 : .get_replorigins(lsn, &ctx, io_concurrency.clone())
8442 1 : .await
8443 1 : .unwrap();
8444 1 : assert_eq!(repl_origins.len(), 1);
8445 1 : assert_eq!(repl_origins[&1], lsn);
8446 :
8447 : {
8448 1 : lsn += 8;
8449 1 : let mut modification = tline.begin_modification(lsn);
8450 1 : modification.put_for_unit_test(
8451 1 : repl_origin_key(3),
8452 1 : Value::Image(Bytes::copy_from_slice(b"cannot_decode_this")),
8453 : );
8454 1 : modification.commit(&ctx).await.unwrap();
8455 : }
8456 1 : let result = tline
8457 1 : .get_replorigins(lsn, &ctx, io_concurrency.clone())
8458 1 : .await;
8459 1 : assert!(result.is_err());
8460 1 : }
8461 :
8462 : #[tokio::test]
8463 1 : async fn test_metadata_image_creation() -> anyhow::Result<()> {
8464 1 : let harness = TenantHarness::create("test_metadata_image_creation").await?;
8465 1 : let (tenant, ctx) = harness.load().await;
8466 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8467 1 : let tline = tenant
8468 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
8469 1 : .await?;
8470 :
8471 : const NUM_KEYS: usize = 1000;
8472 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
8473 :
8474 1 : let cancel = CancellationToken::new();
8475 :
8476 1 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8477 1 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
8478 1 : let mut test_key = base_key;
8479 1 : let mut lsn = Lsn(0x10);
8480 :
8481 4 : async fn scan_with_statistics(
8482 4 : tline: &Timeline,
8483 4 : keyspace: &KeySpace,
8484 4 : lsn: Lsn,
8485 4 : ctx: &RequestContext,
8486 4 : io_concurrency: IoConcurrency,
8487 4 : ) -> anyhow::Result<(BTreeMap<Key, Result<Bytes, PageReconstructError>>, usize)> {
8488 4 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
8489 4 : let query = VersionedKeySpaceQuery::uniform(keyspace.clone(), lsn);
8490 4 : let res = tline
8491 4 : .get_vectored_impl(query, &mut reconstruct_state, ctx)
8492 4 : .await?;
8493 4 : Ok((res, reconstruct_state.get_delta_layers_visited() as usize))
8494 4 : }
8495 :
8496 1001 : for blknum in 0..NUM_KEYS {
8497 1000 : lsn = Lsn(lsn.0 + 0x10);
8498 1000 : test_key.field6 = (blknum * STEP) as u32;
8499 1000 : let mut writer = tline.writer().await;
8500 1000 : writer
8501 1000 : .put(
8502 1000 : test_key,
8503 1000 : lsn,
8504 1000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
8505 1000 : &ctx,
8506 1000 : )
8507 1000 : .await?;
8508 1000 : writer.finish_write(lsn);
8509 1000 : drop(writer);
8510 : }
8511 :
8512 1 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
8513 :
8514 11 : for iter in 1..=10 {
8515 10010 : for _ in 0..NUM_KEYS {
8516 10000 : lsn = Lsn(lsn.0 + 0x10);
8517 10000 : let blknum = rand::rng().random_range(0..NUM_KEYS);
8518 10000 : test_key.field6 = (blknum * STEP) as u32;
8519 10000 : let mut writer = tline.writer().await;
8520 10000 : writer
8521 10000 : .put(
8522 10000 : test_key,
8523 10000 : lsn,
8524 10000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
8525 10000 : &ctx,
8526 10000 : )
8527 10000 : .await?;
8528 10000 : writer.finish_write(lsn);
8529 10000 : drop(writer);
8530 1 : }
8531 1 :
8532 10 : tline.freeze_and_flush().await?;
8533 1 : // Force layers to L1
8534 10 : tline
8535 10 : .compact(
8536 10 : &cancel,
8537 10 : {
8538 10 : let mut flags = EnumSet::new();
8539 10 : flags.insert(CompactFlags::ForceL0Compaction);
8540 10 : flags
8541 10 : },
8542 10 : &ctx,
8543 10 : )
8544 10 : .await?;
8545 1 :
8546 10 : if iter % 5 == 0 {
8547 2 : let scan_lsn = Lsn(lsn.0 + 1);
8548 2 : info!("scanning at {}", scan_lsn);
8549 2 : let (_, before_delta_file_accessed) =
8550 2 : scan_with_statistics(&tline, &keyspace, scan_lsn, &ctx, io_concurrency.clone())
8551 2 : .await?;
8552 2 : tline
8553 2 : .compact(
8554 2 : &cancel,
8555 2 : {
8556 2 : let mut flags = EnumSet::new();
8557 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
8558 2 : flags.insert(CompactFlags::ForceRepartition);
8559 2 : flags.insert(CompactFlags::ForceL0Compaction);
8560 2 : flags
8561 2 : },
8562 2 : &ctx,
8563 2 : )
8564 2 : .await?;
8565 2 : let (_, after_delta_file_accessed) =
8566 2 : scan_with_statistics(&tline, &keyspace, scan_lsn, &ctx, io_concurrency.clone())
8567 2 : .await?;
8568 2 : assert!(
8569 2 : after_delta_file_accessed < before_delta_file_accessed,
8570 1 : "after_delta_file_accessed={after_delta_file_accessed}, before_delta_file_accessed={before_delta_file_accessed}"
8571 1 : );
8572 1 : // Given that we already produced an image layer, there should be no delta layer needed for the scan, but still setting a low threshold there for unforeseen circumstances.
8573 2 : assert!(
8574 2 : after_delta_file_accessed <= 2,
8575 1 : "after_delta_file_accessed={after_delta_file_accessed}"
8576 1 : );
8577 8 : }
8578 1 : }
8579 1 :
8580 1 : Ok(())
8581 1 : }
8582 :
8583 : #[tokio::test]
8584 1 : async fn test_vectored_missing_data_key_reads() -> anyhow::Result<()> {
8585 1 : let harness = TenantHarness::create("test_vectored_missing_data_key_reads").await?;
8586 1 : let (tenant, ctx) = harness.load().await;
8587 :
8588 1 : let base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
8589 1 : let base_key_child = Key::from_hex("000000000033333333444444445500000001").unwrap();
8590 1 : let base_key_nonexist = Key::from_hex("000000000033333333444444445500000002").unwrap();
8591 :
8592 1 : let tline = tenant
8593 1 : .create_test_timeline_with_layers(
8594 1 : TIMELINE_ID,
8595 1 : Lsn(0x10),
8596 1 : DEFAULT_PG_VERSION,
8597 1 : &ctx,
8598 1 : Vec::new(), // in-memory layers
8599 1 : Vec::new(), // delta layers
8600 1 : vec![(Lsn(0x20), vec![(base_key, test_img("data key 1"))])], // image layers
8601 1 : Lsn(0x20), // it's fine to not advance LSN to 0x30 while using 0x30 to get below because `get_vectored_impl` does not wait for LSN
8602 1 : )
8603 1 : .await?;
8604 1 : tline.add_extra_test_dense_keyspace(KeySpace::single(base_key..(base_key_nonexist.next())));
8605 :
8606 1 : let child = tenant
8607 1 : .branch_timeline_test_with_layers(
8608 1 : &tline,
8609 1 : NEW_TIMELINE_ID,
8610 1 : Some(Lsn(0x20)),
8611 1 : &ctx,
8612 1 : Vec::new(), // delta layers
8613 1 : vec![(Lsn(0x30), vec![(base_key_child, test_img("data key 2"))])], // image layers
8614 1 : Lsn(0x30),
8615 1 : )
8616 1 : .await
8617 1 : .unwrap();
8618 :
8619 1 : let lsn = Lsn(0x30);
8620 :
8621 : // test vectored get on parent timeline
8622 1 : assert_eq!(
8623 1 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
8624 1 : Some(test_img("data key 1"))
8625 : );
8626 1 : assert!(
8627 1 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx)
8628 1 : .await
8629 1 : .unwrap_err()
8630 1 : .is_missing_key_error()
8631 : );
8632 1 : assert!(
8633 1 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx)
8634 1 : .await
8635 1 : .unwrap_err()
8636 1 : .is_missing_key_error()
8637 : );
8638 :
8639 : // test vectored get on child timeline
8640 1 : assert_eq!(
8641 1 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
8642 1 : Some(test_img("data key 1"))
8643 : );
8644 1 : assert_eq!(
8645 1 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
8646 1 : Some(test_img("data key 2"))
8647 : );
8648 1 : assert!(
8649 1 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx)
8650 1 : .await
8651 1 : .unwrap_err()
8652 1 : .is_missing_key_error()
8653 : );
8654 :
8655 2 : Ok(())
8656 1 : }
8657 :
8658 : #[tokio::test]
8659 1 : async fn test_vectored_missing_metadata_key_reads() -> anyhow::Result<()> {
8660 1 : let harness = TenantHarness::create("test_vectored_missing_metadata_key_reads").await?;
8661 1 : let (tenant, ctx) = harness.load().await;
8662 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8663 :
8664 1 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8665 1 : let base_key_child = Key::from_hex("620000000033333333444444445500000001").unwrap();
8666 1 : let base_key_nonexist = Key::from_hex("620000000033333333444444445500000002").unwrap();
8667 1 : let base_key_overwrite = Key::from_hex("620000000033333333444444445500000003").unwrap();
8668 :
8669 1 : let base_inherited_key = Key::from_hex("610000000033333333444444445500000000").unwrap();
8670 1 : let base_inherited_key_child =
8671 1 : Key::from_hex("610000000033333333444444445500000001").unwrap();
8672 1 : let base_inherited_key_nonexist =
8673 1 : Key::from_hex("610000000033333333444444445500000002").unwrap();
8674 1 : let base_inherited_key_overwrite =
8675 1 : Key::from_hex("610000000033333333444444445500000003").unwrap();
8676 :
8677 1 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
8678 1 : assert_eq!(base_inherited_key.field1, RELATION_SIZE_PREFIX);
8679 :
8680 1 : let tline = tenant
8681 1 : .create_test_timeline_with_layers(
8682 1 : TIMELINE_ID,
8683 1 : Lsn(0x10),
8684 1 : DEFAULT_PG_VERSION,
8685 1 : &ctx,
8686 1 : Vec::new(), // in-memory layers
8687 1 : Vec::new(), // delta layers
8688 1 : vec![(
8689 1 : Lsn(0x20),
8690 1 : vec![
8691 1 : (base_inherited_key, test_img("metadata inherited key 1")),
8692 1 : (
8693 1 : base_inherited_key_overwrite,
8694 1 : test_img("metadata key overwrite 1a"),
8695 1 : ),
8696 1 : (base_key, test_img("metadata key 1")),
8697 1 : (base_key_overwrite, test_img("metadata key overwrite 1b")),
8698 1 : ],
8699 1 : )], // image layers
8700 1 : Lsn(0x20), // it's fine to not advance LSN to 0x30 while using 0x30 to get below because `get_vectored_impl` does not wait for LSN
8701 1 : )
8702 1 : .await?;
8703 :
8704 1 : let child = tenant
8705 1 : .branch_timeline_test_with_layers(
8706 1 : &tline,
8707 1 : NEW_TIMELINE_ID,
8708 1 : Some(Lsn(0x20)),
8709 1 : &ctx,
8710 1 : Vec::new(), // delta layers
8711 1 : vec![(
8712 1 : Lsn(0x30),
8713 1 : vec![
8714 1 : (
8715 1 : base_inherited_key_child,
8716 1 : test_img("metadata inherited key 2"),
8717 1 : ),
8718 1 : (
8719 1 : base_inherited_key_overwrite,
8720 1 : test_img("metadata key overwrite 2a"),
8721 1 : ),
8722 1 : (base_key_child, test_img("metadata key 2")),
8723 1 : (base_key_overwrite, test_img("metadata key overwrite 2b")),
8724 1 : ],
8725 1 : )], // image layers
8726 1 : Lsn(0x30),
8727 1 : )
8728 1 : .await
8729 1 : .unwrap();
8730 :
8731 1 : let lsn = Lsn(0x30);
8732 :
8733 : // test vectored get on parent timeline
8734 1 : assert_eq!(
8735 1 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
8736 1 : Some(test_img("metadata key 1"))
8737 : );
8738 1 : assert_eq!(
8739 1 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx).await?,
8740 : None
8741 : );
8742 1 : assert_eq!(
8743 1 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx).await?,
8744 : None
8745 : );
8746 1 : assert_eq!(
8747 1 : get_vectored_impl_wrapper(&tline, base_key_overwrite, lsn, &ctx).await?,
8748 1 : Some(test_img("metadata key overwrite 1b"))
8749 : );
8750 1 : assert_eq!(
8751 1 : get_vectored_impl_wrapper(&tline, base_inherited_key, lsn, &ctx).await?,
8752 1 : Some(test_img("metadata inherited key 1"))
8753 : );
8754 1 : assert_eq!(
8755 1 : get_vectored_impl_wrapper(&tline, base_inherited_key_child, lsn, &ctx).await?,
8756 : None
8757 : );
8758 1 : assert_eq!(
8759 1 : get_vectored_impl_wrapper(&tline, base_inherited_key_nonexist, lsn, &ctx).await?,
8760 : None
8761 : );
8762 1 : assert_eq!(
8763 1 : get_vectored_impl_wrapper(&tline, base_inherited_key_overwrite, lsn, &ctx).await?,
8764 1 : Some(test_img("metadata key overwrite 1a"))
8765 : );
8766 :
8767 : // test vectored get on child timeline
8768 1 : assert_eq!(
8769 1 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
8770 : None
8771 : );
8772 1 : assert_eq!(
8773 1 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
8774 1 : Some(test_img("metadata key 2"))
8775 : );
8776 1 : assert_eq!(
8777 1 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx).await?,
8778 : None
8779 : );
8780 1 : assert_eq!(
8781 1 : get_vectored_impl_wrapper(&child, base_inherited_key, lsn, &ctx).await?,
8782 1 : Some(test_img("metadata inherited key 1"))
8783 : );
8784 1 : assert_eq!(
8785 1 : get_vectored_impl_wrapper(&child, base_inherited_key_child, lsn, &ctx).await?,
8786 1 : Some(test_img("metadata inherited key 2"))
8787 : );
8788 1 : assert_eq!(
8789 1 : get_vectored_impl_wrapper(&child, base_inherited_key_nonexist, lsn, &ctx).await?,
8790 : None
8791 : );
8792 1 : assert_eq!(
8793 1 : get_vectored_impl_wrapper(&child, base_key_overwrite, lsn, &ctx).await?,
8794 1 : Some(test_img("metadata key overwrite 2b"))
8795 : );
8796 1 : assert_eq!(
8797 1 : get_vectored_impl_wrapper(&child, base_inherited_key_overwrite, lsn, &ctx).await?,
8798 1 : Some(test_img("metadata key overwrite 2a"))
8799 : );
8800 :
8801 : // test vectored scan on parent timeline
8802 1 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency.clone());
8803 1 : let query =
8804 1 : VersionedKeySpaceQuery::uniform(KeySpace::single(Key::metadata_key_range()), lsn);
8805 1 : let res = tline
8806 1 : .get_vectored_impl(query, &mut reconstruct_state, &ctx)
8807 1 : .await?;
8808 :
8809 1 : assert_eq!(
8810 1 : res.into_iter()
8811 4 : .map(|(k, v)| (k, v.unwrap()))
8812 1 : .collect::<Vec<_>>(),
8813 1 : vec![
8814 1 : (base_inherited_key, test_img("metadata inherited key 1")),
8815 1 : (
8816 1 : base_inherited_key_overwrite,
8817 1 : test_img("metadata key overwrite 1a")
8818 1 : ),
8819 1 : (base_key, test_img("metadata key 1")),
8820 1 : (base_key_overwrite, test_img("metadata key overwrite 1b")),
8821 : ]
8822 : );
8823 :
8824 : // test vectored scan on child timeline
8825 1 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency.clone());
8826 1 : let query =
8827 1 : VersionedKeySpaceQuery::uniform(KeySpace::single(Key::metadata_key_range()), lsn);
8828 1 : let res = child
8829 1 : .get_vectored_impl(query, &mut reconstruct_state, &ctx)
8830 1 : .await?;
8831 :
8832 1 : assert_eq!(
8833 1 : res.into_iter()
8834 5 : .map(|(k, v)| (k, v.unwrap()))
8835 1 : .collect::<Vec<_>>(),
8836 1 : vec![
8837 1 : (base_inherited_key, test_img("metadata inherited key 1")),
8838 1 : (
8839 1 : base_inherited_key_child,
8840 1 : test_img("metadata inherited key 2")
8841 1 : ),
8842 1 : (
8843 1 : base_inherited_key_overwrite,
8844 1 : test_img("metadata key overwrite 2a")
8845 1 : ),
8846 1 : (base_key_child, test_img("metadata key 2")),
8847 1 : (base_key_overwrite, test_img("metadata key overwrite 2b")),
8848 : ]
8849 : );
8850 :
8851 2 : Ok(())
8852 1 : }
8853 :
8854 28 : async fn get_vectored_impl_wrapper(
8855 28 : tline: &Arc<Timeline>,
8856 28 : key: Key,
8857 28 : lsn: Lsn,
8858 28 : ctx: &RequestContext,
8859 28 : ) -> Result<Option<Bytes>, GetVectoredError> {
8860 28 : let io_concurrency = IoConcurrency::spawn_from_conf(
8861 28 : tline.conf.get_vectored_concurrent_io,
8862 28 : tline.gate.enter().unwrap(),
8863 : );
8864 28 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
8865 28 : let query = VersionedKeySpaceQuery::uniform(KeySpace::single(key..key.next()), lsn);
8866 28 : let mut res = tline
8867 28 : .get_vectored_impl(query, &mut reconstruct_state, ctx)
8868 28 : .await?;
8869 25 : Ok(res.pop_last().map(|(k, v)| {
8870 16 : assert_eq!(k, key);
8871 16 : v.unwrap()
8872 16 : }))
8873 28 : }
8874 :
8875 : #[tokio::test]
8876 1 : async fn test_metadata_tombstone_reads() -> anyhow::Result<()> {
8877 1 : let harness = TenantHarness::create("test_metadata_tombstone_reads").await?;
8878 1 : let (tenant, ctx) = harness.load().await;
8879 1 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
8880 1 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8881 1 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8882 1 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
8883 :
8884 : // We emulate the situation that the compaction algorithm creates an image layer that removes the tombstones
8885 : // Lsn 0x30 key0, key3, no key1+key2
8886 : // Lsn 0x20 key1+key2 tomestones
8887 : // Lsn 0x10 key1 in image, key2 in delta
8888 1 : let tline = tenant
8889 1 : .create_test_timeline_with_layers(
8890 1 : TIMELINE_ID,
8891 1 : Lsn(0x10),
8892 1 : DEFAULT_PG_VERSION,
8893 1 : &ctx,
8894 1 : Vec::new(), // in-memory layers
8895 1 : // delta layers
8896 1 : vec![
8897 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8898 1 : Lsn(0x10)..Lsn(0x20),
8899 1 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8900 1 : ),
8901 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8902 1 : Lsn(0x20)..Lsn(0x30),
8903 1 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8904 1 : ),
8905 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8906 1 : Lsn(0x20)..Lsn(0x30),
8907 1 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8908 1 : ),
8909 1 : ],
8910 1 : // image layers
8911 1 : vec![
8912 1 : (Lsn(0x10), vec![(key1, test_img("metadata key 1"))]),
8913 1 : (
8914 1 : Lsn(0x30),
8915 1 : vec![
8916 1 : (key0, test_img("metadata key 0")),
8917 1 : (key3, test_img("metadata key 3")),
8918 1 : ],
8919 1 : ),
8920 1 : ],
8921 1 : Lsn(0x30),
8922 1 : )
8923 1 : .await?;
8924 :
8925 1 : let lsn = Lsn(0x30);
8926 1 : let old_lsn = Lsn(0x20);
8927 :
8928 1 : assert_eq!(
8929 1 : get_vectored_impl_wrapper(&tline, key0, lsn, &ctx).await?,
8930 1 : Some(test_img("metadata key 0"))
8931 : );
8932 1 : assert_eq!(
8933 1 : get_vectored_impl_wrapper(&tline, key1, lsn, &ctx).await?,
8934 : None,
8935 : );
8936 1 : assert_eq!(
8937 1 : get_vectored_impl_wrapper(&tline, key2, lsn, &ctx).await?,
8938 : None,
8939 : );
8940 1 : assert_eq!(
8941 1 : get_vectored_impl_wrapper(&tline, key1, old_lsn, &ctx).await?,
8942 1 : Some(Bytes::new()),
8943 : );
8944 1 : assert_eq!(
8945 1 : get_vectored_impl_wrapper(&tline, key2, old_lsn, &ctx).await?,
8946 1 : Some(Bytes::new()),
8947 : );
8948 1 : assert_eq!(
8949 1 : get_vectored_impl_wrapper(&tline, key3, lsn, &ctx).await?,
8950 1 : Some(test_img("metadata key 3"))
8951 : );
8952 :
8953 2 : Ok(())
8954 1 : }
8955 :
8956 : #[tokio::test]
8957 1 : async fn test_metadata_tombstone_image_creation() {
8958 1 : let harness = TenantHarness::create("test_metadata_tombstone_image_creation")
8959 1 : .await
8960 1 : .unwrap();
8961 1 : let (tenant, ctx) = harness.load().await;
8962 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8963 :
8964 1 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
8965 1 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8966 1 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8967 1 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
8968 :
8969 1 : let tline = tenant
8970 1 : .create_test_timeline_with_layers(
8971 1 : TIMELINE_ID,
8972 1 : Lsn(0x10),
8973 1 : DEFAULT_PG_VERSION,
8974 1 : &ctx,
8975 1 : Vec::new(), // in-memory layers
8976 1 : // delta layers
8977 1 : vec![
8978 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8979 1 : Lsn(0x10)..Lsn(0x20),
8980 1 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8981 1 : ),
8982 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8983 1 : Lsn(0x20)..Lsn(0x30),
8984 1 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8985 1 : ),
8986 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8987 1 : Lsn(0x20)..Lsn(0x30),
8988 1 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8989 1 : ),
8990 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8991 1 : Lsn(0x30)..Lsn(0x40),
8992 1 : vec![
8993 1 : (key0, Lsn(0x30), Value::Image(test_img("metadata key 0"))),
8994 1 : (key3, Lsn(0x30), Value::Image(test_img("metadata key 3"))),
8995 1 : ],
8996 1 : ),
8997 1 : ],
8998 1 : // image layers
8999 1 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
9000 1 : Lsn(0x40),
9001 1 : )
9002 1 : .await
9003 1 : .unwrap();
9004 :
9005 1 : let cancel = CancellationToken::new();
9006 :
9007 : // Image layer creation happens on the disk_consistent_lsn so we need to force set it now.
9008 1 : tline.force_set_disk_consistent_lsn(Lsn(0x40));
9009 1 : tline
9010 1 : .compact(
9011 1 : &cancel,
9012 1 : {
9013 1 : let mut flags = EnumSet::new();
9014 1 : flags.insert(CompactFlags::ForceImageLayerCreation);
9015 1 : flags.insert(CompactFlags::ForceRepartition);
9016 1 : flags
9017 1 : },
9018 1 : &ctx,
9019 1 : )
9020 1 : .await
9021 1 : .unwrap();
9022 : // Image layers are created at repartition LSN
9023 1 : let images = tline
9024 1 : .inspect_image_layers(Lsn(0x40), &ctx, io_concurrency.clone())
9025 1 : .await
9026 1 : .unwrap()
9027 1 : .into_iter()
9028 9 : .filter(|(k, _)| k.is_metadata_key())
9029 1 : .collect::<Vec<_>>();
9030 1 : assert_eq!(images.len(), 2); // the image layer should only contain two existing keys, tombstones should be removed.
9031 1 : }
9032 :
9033 : #[tokio::test]
9034 1 : async fn test_metadata_tombstone_empty_image_creation() {
9035 1 : let harness = TenantHarness::create("test_metadata_tombstone_empty_image_creation")
9036 1 : .await
9037 1 : .unwrap();
9038 1 : let (tenant, ctx) = harness.load().await;
9039 1 : let io_concurrency = IoConcurrency::spawn_for_test();
9040 :
9041 1 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
9042 1 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
9043 :
9044 1 : let tline = tenant
9045 1 : .create_test_timeline_with_layers(
9046 1 : TIMELINE_ID,
9047 1 : Lsn(0x10),
9048 1 : DEFAULT_PG_VERSION,
9049 1 : &ctx,
9050 1 : Vec::new(), // in-memory layers
9051 1 : // delta layers
9052 1 : vec![
9053 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9054 1 : Lsn(0x10)..Lsn(0x20),
9055 1 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
9056 1 : ),
9057 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9058 1 : Lsn(0x20)..Lsn(0x30),
9059 1 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
9060 1 : ),
9061 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9062 1 : Lsn(0x20)..Lsn(0x30),
9063 1 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
9064 1 : ),
9065 1 : ],
9066 1 : // image layers
9067 1 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
9068 1 : Lsn(0x30),
9069 1 : )
9070 1 : .await
9071 1 : .unwrap();
9072 :
9073 1 : let cancel = CancellationToken::new();
9074 :
9075 1 : tline
9076 1 : .compact(
9077 1 : &cancel,
9078 1 : {
9079 1 : let mut flags = EnumSet::new();
9080 1 : flags.insert(CompactFlags::ForceImageLayerCreation);
9081 1 : flags.insert(CompactFlags::ForceRepartition);
9082 1 : flags
9083 1 : },
9084 1 : &ctx,
9085 1 : )
9086 1 : .await
9087 1 : .unwrap();
9088 :
9089 : // Image layers are created at last_record_lsn
9090 1 : let images = tline
9091 1 : .inspect_image_layers(Lsn(0x30), &ctx, io_concurrency.clone())
9092 1 : .await
9093 1 : .unwrap()
9094 1 : .into_iter()
9095 7 : .filter(|(k, _)| k.is_metadata_key())
9096 1 : .collect::<Vec<_>>();
9097 1 : assert_eq!(images.len(), 0); // the image layer should not contain tombstones, or it is not created
9098 1 : }
9099 :
9100 : #[tokio::test]
9101 1 : async fn test_simple_bottom_most_compaction_images() -> anyhow::Result<()> {
9102 1 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_images").await?;
9103 1 : let (tenant, ctx) = harness.load().await;
9104 1 : let io_concurrency = IoConcurrency::spawn_for_test();
9105 :
9106 51 : fn get_key(id: u32) -> Key {
9107 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9108 51 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9109 51 : key.field6 = id;
9110 51 : key
9111 51 : }
9112 :
9113 : // We create
9114 : // - one bottom-most image layer,
9115 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
9116 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
9117 : // - a delta layer D3 above the horizon.
9118 : //
9119 : // | D3 |
9120 : // | D1 |
9121 : // -| |-- gc horizon -----------------
9122 : // | | | D2 |
9123 : // --------- img layer ------------------
9124 : //
9125 : // What we should expact from this compaction is:
9126 : // | D3 |
9127 : // | Part of D1 |
9128 : // --------- img layer with D1+D2 at GC horizon------------------
9129 :
9130 : // img layer at 0x10
9131 1 : let img_layer = (0..10)
9132 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9133 1 : .collect_vec();
9134 :
9135 1 : let delta1 = vec![
9136 1 : (
9137 1 : get_key(1),
9138 1 : Lsn(0x20),
9139 1 : Value::Image(Bytes::from("value 1@0x20")),
9140 1 : ),
9141 1 : (
9142 1 : get_key(2),
9143 1 : Lsn(0x30),
9144 1 : Value::Image(Bytes::from("value 2@0x30")),
9145 1 : ),
9146 1 : (
9147 1 : get_key(3),
9148 1 : Lsn(0x40),
9149 1 : Value::Image(Bytes::from("value 3@0x40")),
9150 1 : ),
9151 : ];
9152 1 : let delta2 = vec![
9153 1 : (
9154 1 : get_key(5),
9155 1 : Lsn(0x20),
9156 1 : Value::Image(Bytes::from("value 5@0x20")),
9157 1 : ),
9158 1 : (
9159 1 : get_key(6),
9160 1 : Lsn(0x20),
9161 1 : Value::Image(Bytes::from("value 6@0x20")),
9162 1 : ),
9163 : ];
9164 1 : let delta3 = vec![
9165 1 : (
9166 1 : get_key(8),
9167 1 : Lsn(0x48),
9168 1 : Value::Image(Bytes::from("value 8@0x48")),
9169 1 : ),
9170 1 : (
9171 1 : get_key(9),
9172 1 : Lsn(0x48),
9173 1 : Value::Image(Bytes::from("value 9@0x48")),
9174 1 : ),
9175 : ];
9176 :
9177 1 : let tline = tenant
9178 1 : .create_test_timeline_with_layers(
9179 1 : TIMELINE_ID,
9180 1 : Lsn(0x10),
9181 1 : DEFAULT_PG_VERSION,
9182 1 : &ctx,
9183 1 : Vec::new(), // in-memory layers
9184 1 : vec![
9185 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
9186 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
9187 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
9188 1 : ], // delta layers
9189 1 : vec![(Lsn(0x10), img_layer)], // image layers
9190 1 : Lsn(0x50),
9191 1 : )
9192 1 : .await?;
9193 : {
9194 1 : tline
9195 1 : .applied_gc_cutoff_lsn
9196 1 : .lock_for_write()
9197 1 : .store_and_unlock(Lsn(0x30))
9198 1 : .wait()
9199 1 : .await;
9200 : // Update GC info
9201 1 : let mut guard = tline.gc_info.write().unwrap();
9202 1 : guard.cutoffs.time = Some(Lsn(0x30));
9203 1 : guard.cutoffs.space = Lsn(0x30);
9204 : }
9205 :
9206 1 : let expected_result = [
9207 1 : Bytes::from_static(b"value 0@0x10"),
9208 1 : Bytes::from_static(b"value 1@0x20"),
9209 1 : Bytes::from_static(b"value 2@0x30"),
9210 1 : Bytes::from_static(b"value 3@0x40"),
9211 1 : Bytes::from_static(b"value 4@0x10"),
9212 1 : Bytes::from_static(b"value 5@0x20"),
9213 1 : Bytes::from_static(b"value 6@0x20"),
9214 1 : Bytes::from_static(b"value 7@0x10"),
9215 1 : Bytes::from_static(b"value 8@0x48"),
9216 1 : Bytes::from_static(b"value 9@0x48"),
9217 1 : ];
9218 :
9219 10 : for (idx, expected) in expected_result.iter().enumerate() {
9220 10 : assert_eq!(
9221 10 : tline
9222 10 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9223 10 : .await
9224 10 : .unwrap(),
9225 : expected
9226 : );
9227 : }
9228 :
9229 1 : let cancel = CancellationToken::new();
9230 1 : tline
9231 1 : .compact_with_gc(
9232 1 : &cancel,
9233 1 : CompactOptions::default_for_gc_compaction_unit_tests(),
9234 1 : &ctx,
9235 1 : )
9236 1 : .await
9237 1 : .unwrap();
9238 :
9239 10 : for (idx, expected) in expected_result.iter().enumerate() {
9240 10 : assert_eq!(
9241 10 : tline
9242 10 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9243 10 : .await
9244 10 : .unwrap(),
9245 : expected
9246 : );
9247 : }
9248 :
9249 : // Check if the image layer at the GC horizon contains exactly what we want
9250 1 : let image_at_gc_horizon = tline
9251 1 : .inspect_image_layers(Lsn(0x30), &ctx, io_concurrency.clone())
9252 1 : .await
9253 1 : .unwrap()
9254 1 : .into_iter()
9255 17 : .filter(|(k, _)| k.is_metadata_key())
9256 1 : .collect::<Vec<_>>();
9257 :
9258 1 : assert_eq!(image_at_gc_horizon.len(), 10);
9259 1 : let expected_result = [
9260 1 : Bytes::from_static(b"value 0@0x10"),
9261 1 : Bytes::from_static(b"value 1@0x20"),
9262 1 : Bytes::from_static(b"value 2@0x30"),
9263 1 : Bytes::from_static(b"value 3@0x10"),
9264 1 : Bytes::from_static(b"value 4@0x10"),
9265 1 : Bytes::from_static(b"value 5@0x20"),
9266 1 : Bytes::from_static(b"value 6@0x20"),
9267 1 : Bytes::from_static(b"value 7@0x10"),
9268 1 : Bytes::from_static(b"value 8@0x10"),
9269 1 : Bytes::from_static(b"value 9@0x10"),
9270 1 : ];
9271 11 : for idx in 0..10 {
9272 10 : assert_eq!(
9273 10 : image_at_gc_horizon[idx],
9274 10 : (get_key(idx as u32), expected_result[idx].clone())
9275 : );
9276 : }
9277 :
9278 : // Check if old layers are removed / new layers have the expected LSN
9279 1 : let all_layers = inspect_and_sort(&tline, None).await;
9280 1 : assert_eq!(
9281 : all_layers,
9282 1 : vec![
9283 : // Image layer at GC horizon
9284 1 : PersistentLayerKey {
9285 1 : key_range: Key::MIN..Key::MAX,
9286 1 : lsn_range: Lsn(0x30)..Lsn(0x31),
9287 1 : is_delta: false
9288 1 : },
9289 : // The delta layer below the horizon
9290 1 : PersistentLayerKey {
9291 1 : key_range: get_key(3)..get_key(4),
9292 1 : lsn_range: Lsn(0x30)..Lsn(0x48),
9293 1 : is_delta: true
9294 1 : },
9295 : // The delta3 layer that should not be picked for the compaction
9296 1 : PersistentLayerKey {
9297 1 : key_range: get_key(8)..get_key(10),
9298 1 : lsn_range: Lsn(0x48)..Lsn(0x50),
9299 1 : is_delta: true
9300 1 : }
9301 : ]
9302 : );
9303 :
9304 : // increase GC horizon and compact again
9305 : {
9306 1 : tline
9307 1 : .applied_gc_cutoff_lsn
9308 1 : .lock_for_write()
9309 1 : .store_and_unlock(Lsn(0x40))
9310 1 : .wait()
9311 1 : .await;
9312 : // Update GC info
9313 1 : let mut guard = tline.gc_info.write().unwrap();
9314 1 : guard.cutoffs.time = Some(Lsn(0x40));
9315 1 : guard.cutoffs.space = Lsn(0x40);
9316 : }
9317 1 : tline
9318 1 : .compact_with_gc(
9319 1 : &cancel,
9320 1 : CompactOptions::default_for_gc_compaction_unit_tests(),
9321 1 : &ctx,
9322 1 : )
9323 1 : .await
9324 1 : .unwrap();
9325 :
9326 2 : Ok(())
9327 1 : }
9328 :
9329 : #[cfg(feature = "testing")]
9330 : #[tokio::test]
9331 1 : async fn test_neon_test_record() -> anyhow::Result<()> {
9332 1 : let harness = TenantHarness::create("test_neon_test_record").await?;
9333 1 : let (tenant, ctx) = harness.load().await;
9334 :
9335 17 : fn get_key(id: u32) -> Key {
9336 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9337 17 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9338 17 : key.field6 = id;
9339 17 : key
9340 17 : }
9341 :
9342 1 : let delta1 = vec![
9343 1 : (
9344 1 : get_key(1),
9345 1 : Lsn(0x20),
9346 1 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
9347 1 : ),
9348 1 : (
9349 1 : get_key(1),
9350 1 : Lsn(0x30),
9351 1 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
9352 1 : ),
9353 1 : (get_key(2), Lsn(0x10), Value::Image("0x10".into())),
9354 1 : (
9355 1 : get_key(2),
9356 1 : Lsn(0x20),
9357 1 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
9358 1 : ),
9359 1 : (
9360 1 : get_key(2),
9361 1 : Lsn(0x30),
9362 1 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
9363 1 : ),
9364 1 : (get_key(3), Lsn(0x10), Value::Image("0x10".into())),
9365 1 : (
9366 1 : get_key(3),
9367 1 : Lsn(0x20),
9368 1 : Value::WalRecord(NeonWalRecord::wal_clear("c")),
9369 1 : ),
9370 1 : (get_key(4), Lsn(0x10), Value::Image("0x10".into())),
9371 1 : (
9372 1 : get_key(4),
9373 1 : Lsn(0x20),
9374 1 : Value::WalRecord(NeonWalRecord::wal_init("i")),
9375 1 : ),
9376 1 : (
9377 1 : get_key(4),
9378 1 : Lsn(0x30),
9379 1 : Value::WalRecord(NeonWalRecord::wal_append_conditional("j", "i")),
9380 1 : ),
9381 1 : (
9382 1 : get_key(5),
9383 1 : Lsn(0x20),
9384 1 : Value::WalRecord(NeonWalRecord::wal_init("1")),
9385 1 : ),
9386 1 : (
9387 1 : get_key(5),
9388 1 : Lsn(0x30),
9389 1 : Value::WalRecord(NeonWalRecord::wal_append_conditional("j", "2")),
9390 1 : ),
9391 : ];
9392 1 : let image1 = vec![(get_key(1), "0x10".into())];
9393 :
9394 1 : let tline = tenant
9395 1 : .create_test_timeline_with_layers(
9396 1 : TIMELINE_ID,
9397 1 : Lsn(0x10),
9398 1 : DEFAULT_PG_VERSION,
9399 1 : &ctx,
9400 1 : Vec::new(), // in-memory layers
9401 1 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
9402 1 : Lsn(0x10)..Lsn(0x40),
9403 1 : delta1,
9404 1 : )], // delta layers
9405 1 : vec![(Lsn(0x10), image1)], // image layers
9406 1 : Lsn(0x50),
9407 1 : )
9408 1 : .await?;
9409 :
9410 1 : assert_eq!(
9411 1 : tline.get(get_key(1), Lsn(0x50), &ctx).await?,
9412 1 : Bytes::from_static(b"0x10,0x20,0x30")
9413 : );
9414 1 : assert_eq!(
9415 1 : tline.get(get_key(2), Lsn(0x50), &ctx).await?,
9416 1 : Bytes::from_static(b"0x10,0x20,0x30")
9417 : );
9418 :
9419 : // Need to remove the limit of "Neon WAL redo requires base image".
9420 :
9421 1 : assert_eq!(
9422 1 : tline.get(get_key(3), Lsn(0x50), &ctx).await?,
9423 1 : Bytes::from_static(b"c")
9424 : );
9425 1 : assert_eq!(
9426 1 : tline.get(get_key(4), Lsn(0x50), &ctx).await?,
9427 1 : Bytes::from_static(b"ij")
9428 : );
9429 :
9430 : // Manual testing required: currently, read errors will panic the process in debug mode. So we
9431 : // cannot enable this assertion in the unit test.
9432 : // assert!(tline.get(get_key(5), Lsn(0x50), &ctx).await.is_err());
9433 :
9434 2 : Ok(())
9435 1 : }
9436 :
9437 : #[tokio::test]
9438 1 : async fn test_lsn_lease() -> anyhow::Result<()> {
9439 1 : let (tenant, ctx) = TenantHarness::create("test_lsn_lease")
9440 1 : .await
9441 1 : .unwrap()
9442 1 : .load()
9443 1 : .await;
9444 : // set a non-zero lease length to test the feature
9445 1 : tenant
9446 1 : .update_tenant_config(|mut conf| {
9447 1 : conf.lsn_lease_length = Some(LsnLease::DEFAULT_LENGTH);
9448 1 : Ok(conf)
9449 1 : })
9450 1 : .unwrap();
9451 :
9452 1 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
9453 :
9454 1 : let end_lsn = Lsn(0x100);
9455 1 : let image_layers = (0x20..=0x90)
9456 1 : .step_by(0x10)
9457 8 : .map(|n| (Lsn(n), vec![(key, test_img(&format!("data key at {n:x}")))]))
9458 1 : .collect();
9459 :
9460 1 : let timeline = tenant
9461 1 : .create_test_timeline_with_layers(
9462 1 : TIMELINE_ID,
9463 1 : Lsn(0x10),
9464 1 : DEFAULT_PG_VERSION,
9465 1 : &ctx,
9466 1 : Vec::new(), // in-memory layers
9467 1 : Vec::new(),
9468 1 : image_layers,
9469 1 : end_lsn,
9470 1 : )
9471 1 : .await?;
9472 :
9473 1 : let leased_lsns = [0x30, 0x50, 0x70];
9474 1 : let mut leases = Vec::new();
9475 3 : leased_lsns.iter().for_each(|n| {
9476 3 : leases.push(
9477 3 : timeline
9478 3 : .init_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)
9479 3 : .expect("lease request should succeed"),
9480 : );
9481 3 : });
9482 :
9483 1 : let updated_lease_0 = timeline
9484 1 : .renew_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)
9485 1 : .expect("lease renewal should succeed");
9486 1 : assert_eq!(
9487 1 : updated_lease_0.valid_until, leases[0].valid_until,
9488 0 : " Renewing with shorter lease should not change the lease."
9489 : );
9490 :
9491 1 : let updated_lease_1 = timeline
9492 1 : .renew_lsn_lease(
9493 1 : Lsn(leased_lsns[1]),
9494 1 : timeline.get_lsn_lease_length() * 2,
9495 1 : &ctx,
9496 1 : )
9497 1 : .expect("lease renewal should succeed");
9498 1 : assert!(
9499 1 : updated_lease_1.valid_until > leases[1].valid_until,
9500 0 : "Renewing with a long lease should renew lease with later expiration time."
9501 : );
9502 :
9503 : // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
9504 1 : info!(
9505 0 : "applied_gc_cutoff_lsn: {}",
9506 0 : *timeline.get_applied_gc_cutoff_lsn()
9507 : );
9508 1 : timeline.force_set_disk_consistent_lsn(end_lsn);
9509 :
9510 1 : let res = tenant
9511 1 : .gc_iteration(
9512 1 : Some(TIMELINE_ID),
9513 1 : 0,
9514 1 : Duration::ZERO,
9515 1 : &CancellationToken::new(),
9516 1 : &ctx,
9517 1 : )
9518 1 : .await
9519 1 : .unwrap();
9520 :
9521 : // Keeping everything <= Lsn(0x80) b/c leases:
9522 : // 0/10: initdb layer
9523 : // (0/20..=0/70).step_by(0x10): image layers added when creating the timeline.
9524 1 : assert_eq!(res.layers_needed_by_leases, 7);
9525 : // Keeping 0/90 b/c it is the latest layer.
9526 1 : assert_eq!(res.layers_not_updated, 1);
9527 : // Removed 0/80.
9528 1 : assert_eq!(res.layers_removed, 1);
9529 :
9530 : // Make lease on a already GC-ed LSN.
9531 : // 0/80 does not have a valid lease + is below latest_gc_cutoff
9532 1 : assert!(Lsn(0x80) < *timeline.get_applied_gc_cutoff_lsn());
9533 1 : timeline
9534 1 : .init_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx)
9535 1 : .expect_err("lease request on GC-ed LSN should fail");
9536 :
9537 : // Should still be able to renew a currently valid lease
9538 : // Assumption: original lease to is still valid for 0/50.
9539 : // (use `Timeline::init_lsn_lease` for testing so it always does validation)
9540 1 : timeline
9541 1 : .init_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)
9542 1 : .expect("lease renewal with validation should succeed");
9543 :
9544 2 : Ok(())
9545 1 : }
9546 :
9547 : #[tokio::test]
9548 1 : async fn test_failed_flush_should_not_update_disk_consistent_lsn() -> anyhow::Result<()> {
9549 : //
9550 : // Setup
9551 : //
9552 1 : let harness = TenantHarness::create_custom(
9553 1 : "test_failed_flush_should_not_upload_disk_consistent_lsn",
9554 1 : pageserver_api::models::TenantConfig::default(),
9555 1 : TenantId::generate(),
9556 1 : ShardIdentity::new(ShardNumber(0), ShardCount(4), ShardStripeSize(128)).unwrap(),
9557 1 : Generation::new(1),
9558 1 : )
9559 1 : .await?;
9560 1 : let (tenant, ctx) = harness.load().await;
9561 :
9562 1 : let timeline = tenant
9563 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
9564 1 : .await?;
9565 1 : assert_eq!(timeline.get_shard_identity().count, ShardCount(4));
9566 1 : let mut writer = timeline.writer().await;
9567 1 : writer
9568 1 : .put(
9569 1 : *TEST_KEY,
9570 1 : Lsn(0x20),
9571 1 : &Value::Image(test_img("foo at 0x20")),
9572 1 : &ctx,
9573 1 : )
9574 1 : .await?;
9575 1 : writer.finish_write(Lsn(0x20));
9576 1 : drop(writer);
9577 1 : timeline.freeze_and_flush().await.unwrap();
9578 :
9579 1 : timeline.remote_client.wait_completion().await.unwrap();
9580 1 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
9581 1 : let remote_consistent_lsn = timeline.get_remote_consistent_lsn_projected();
9582 1 : assert_eq!(Some(disk_consistent_lsn), remote_consistent_lsn);
9583 :
9584 : //
9585 : // Test
9586 : //
9587 :
9588 1 : let mut writer = timeline.writer().await;
9589 1 : writer
9590 1 : .put(
9591 1 : *TEST_KEY,
9592 1 : Lsn(0x30),
9593 1 : &Value::Image(test_img("foo at 0x30")),
9594 1 : &ctx,
9595 1 : )
9596 1 : .await?;
9597 1 : writer.finish_write(Lsn(0x30));
9598 1 : drop(writer);
9599 :
9600 1 : fail::cfg(
9601 : "flush-layer-before-update-remote-consistent-lsn",
9602 1 : "return()",
9603 : )
9604 1 : .unwrap();
9605 :
9606 1 : let flush_res = timeline.freeze_and_flush().await;
9607 : // if flush failed, the disk/remote consistent LSN should not be updated
9608 1 : assert!(flush_res.is_err());
9609 1 : assert_eq!(disk_consistent_lsn, timeline.get_disk_consistent_lsn());
9610 1 : assert_eq!(
9611 : remote_consistent_lsn,
9612 1 : timeline.get_remote_consistent_lsn_projected()
9613 : );
9614 :
9615 2 : Ok(())
9616 1 : }
9617 :
9618 : #[cfg(feature = "testing")]
9619 : #[tokio::test]
9620 1 : async fn test_simple_bottom_most_compaction_deltas_1() -> anyhow::Result<()> {
9621 2 : test_simple_bottom_most_compaction_deltas_helper(
9622 2 : "test_simple_bottom_most_compaction_deltas_1",
9623 2 : false,
9624 2 : )
9625 2 : .await
9626 1 : }
9627 :
9628 : #[cfg(feature = "testing")]
9629 : #[tokio::test]
9630 1 : async fn test_simple_bottom_most_compaction_deltas_2() -> anyhow::Result<()> {
9631 2 : test_simple_bottom_most_compaction_deltas_helper(
9632 2 : "test_simple_bottom_most_compaction_deltas_2",
9633 2 : true,
9634 2 : )
9635 2 : .await
9636 1 : }
9637 :
9638 : #[cfg(feature = "testing")]
9639 2 : async fn test_simple_bottom_most_compaction_deltas_helper(
9640 2 : test_name: &'static str,
9641 2 : use_delta_bottom_layer: bool,
9642 2 : ) -> anyhow::Result<()> {
9643 2 : let harness = TenantHarness::create(test_name).await?;
9644 2 : let (tenant, ctx) = harness.load().await;
9645 :
9646 138 : fn get_key(id: u32) -> Key {
9647 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9648 138 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9649 138 : key.field6 = id;
9650 138 : key
9651 138 : }
9652 :
9653 : // We create
9654 : // - one bottom-most image layer,
9655 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
9656 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
9657 : // - a delta layer D3 above the horizon.
9658 : //
9659 : // | D3 |
9660 : // | D1 |
9661 : // -| |-- gc horizon -----------------
9662 : // | | | D2 |
9663 : // --------- img layer ------------------
9664 : //
9665 : // What we should expact from this compaction is:
9666 : // | D3 |
9667 : // | Part of D1 |
9668 : // --------- img layer with D1+D2 at GC horizon------------------
9669 :
9670 : // img layer at 0x10
9671 2 : let img_layer = (0..10)
9672 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9673 2 : .collect_vec();
9674 : // or, delta layer at 0x10 if `use_delta_bottom_layer` is true
9675 2 : let delta4 = (0..10)
9676 20 : .map(|id| {
9677 20 : (
9678 20 : get_key(id),
9679 20 : Lsn(0x08),
9680 20 : Value::WalRecord(NeonWalRecord::wal_init(format!("value {id}@0x10"))),
9681 20 : )
9682 20 : })
9683 2 : .collect_vec();
9684 :
9685 2 : let delta1 = vec![
9686 2 : (
9687 2 : get_key(1),
9688 2 : Lsn(0x20),
9689 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9690 2 : ),
9691 2 : (
9692 2 : get_key(2),
9693 2 : Lsn(0x30),
9694 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9695 2 : ),
9696 2 : (
9697 2 : get_key(3),
9698 2 : Lsn(0x28),
9699 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9700 2 : ),
9701 2 : (
9702 2 : get_key(3),
9703 2 : Lsn(0x30),
9704 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9705 2 : ),
9706 2 : (
9707 2 : get_key(3),
9708 2 : Lsn(0x40),
9709 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9710 2 : ),
9711 : ];
9712 2 : let delta2 = vec![
9713 2 : (
9714 2 : get_key(5),
9715 2 : Lsn(0x20),
9716 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9717 2 : ),
9718 2 : (
9719 2 : get_key(6),
9720 2 : Lsn(0x20),
9721 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9722 2 : ),
9723 : ];
9724 2 : let delta3 = vec![
9725 2 : (
9726 2 : get_key(8),
9727 2 : Lsn(0x48),
9728 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9729 2 : ),
9730 2 : (
9731 2 : get_key(9),
9732 2 : Lsn(0x48),
9733 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9734 2 : ),
9735 : ];
9736 :
9737 2 : let tline = if use_delta_bottom_layer {
9738 1 : tenant
9739 1 : .create_test_timeline_with_layers(
9740 1 : TIMELINE_ID,
9741 1 : Lsn(0x08),
9742 1 : DEFAULT_PG_VERSION,
9743 1 : &ctx,
9744 1 : Vec::new(), // in-memory layers
9745 1 : vec![
9746 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9747 1 : Lsn(0x08)..Lsn(0x10),
9748 1 : delta4,
9749 1 : ),
9750 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9751 1 : Lsn(0x20)..Lsn(0x48),
9752 1 : delta1,
9753 1 : ),
9754 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9755 1 : Lsn(0x20)..Lsn(0x48),
9756 1 : delta2,
9757 1 : ),
9758 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9759 1 : Lsn(0x48)..Lsn(0x50),
9760 1 : delta3,
9761 1 : ),
9762 1 : ], // delta layers
9763 1 : vec![], // image layers
9764 1 : Lsn(0x50),
9765 1 : )
9766 1 : .await?
9767 : } else {
9768 1 : tenant
9769 1 : .create_test_timeline_with_layers(
9770 1 : TIMELINE_ID,
9771 1 : Lsn(0x10),
9772 1 : DEFAULT_PG_VERSION,
9773 1 : &ctx,
9774 1 : Vec::new(), // in-memory layers
9775 1 : vec![
9776 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9777 1 : Lsn(0x10)..Lsn(0x48),
9778 1 : delta1,
9779 1 : ),
9780 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9781 1 : Lsn(0x10)..Lsn(0x48),
9782 1 : delta2,
9783 1 : ),
9784 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9785 1 : Lsn(0x48)..Lsn(0x50),
9786 1 : delta3,
9787 1 : ),
9788 1 : ], // delta layers
9789 1 : vec![(Lsn(0x10), img_layer)], // image layers
9790 1 : Lsn(0x50),
9791 1 : )
9792 1 : .await?
9793 : };
9794 : {
9795 2 : tline
9796 2 : .applied_gc_cutoff_lsn
9797 2 : .lock_for_write()
9798 2 : .store_and_unlock(Lsn(0x30))
9799 2 : .wait()
9800 2 : .await;
9801 : // Update GC info
9802 2 : let mut guard = tline.gc_info.write().unwrap();
9803 2 : *guard = GcInfo {
9804 2 : retain_lsns: vec![],
9805 2 : cutoffs: GcCutoffs {
9806 2 : time: Some(Lsn(0x30)),
9807 2 : space: Lsn(0x30),
9808 2 : },
9809 2 : leases: Default::default(),
9810 2 : within_ancestor_pitr: false,
9811 2 : };
9812 : }
9813 :
9814 2 : let expected_result = [
9815 2 : Bytes::from_static(b"value 0@0x10"),
9816 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9817 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9818 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9819 2 : Bytes::from_static(b"value 4@0x10"),
9820 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9821 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9822 2 : Bytes::from_static(b"value 7@0x10"),
9823 2 : Bytes::from_static(b"value 8@0x10@0x48"),
9824 2 : Bytes::from_static(b"value 9@0x10@0x48"),
9825 2 : ];
9826 :
9827 2 : let expected_result_at_gc_horizon = [
9828 2 : Bytes::from_static(b"value 0@0x10"),
9829 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9830 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9831 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
9832 2 : Bytes::from_static(b"value 4@0x10"),
9833 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9834 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9835 2 : Bytes::from_static(b"value 7@0x10"),
9836 2 : Bytes::from_static(b"value 8@0x10"),
9837 2 : Bytes::from_static(b"value 9@0x10"),
9838 2 : ];
9839 :
9840 22 : for idx in 0..10 {
9841 20 : assert_eq!(
9842 20 : tline
9843 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9844 20 : .await
9845 20 : .unwrap(),
9846 20 : &expected_result[idx]
9847 : );
9848 20 : assert_eq!(
9849 20 : tline
9850 20 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
9851 20 : .await
9852 20 : .unwrap(),
9853 20 : &expected_result_at_gc_horizon[idx]
9854 : );
9855 : }
9856 :
9857 2 : let cancel = CancellationToken::new();
9858 2 : tline
9859 2 : .compact_with_gc(
9860 2 : &cancel,
9861 2 : CompactOptions::default_for_gc_compaction_unit_tests(),
9862 2 : &ctx,
9863 2 : )
9864 2 : .await
9865 2 : .unwrap();
9866 :
9867 22 : for idx in 0..10 {
9868 20 : assert_eq!(
9869 20 : tline
9870 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9871 20 : .await
9872 20 : .unwrap(),
9873 20 : &expected_result[idx]
9874 : );
9875 20 : assert_eq!(
9876 20 : tline
9877 20 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
9878 20 : .await
9879 20 : .unwrap(),
9880 20 : &expected_result_at_gc_horizon[idx]
9881 : );
9882 : }
9883 :
9884 : // increase GC horizon and compact again
9885 : {
9886 2 : tline
9887 2 : .applied_gc_cutoff_lsn
9888 2 : .lock_for_write()
9889 2 : .store_and_unlock(Lsn(0x40))
9890 2 : .wait()
9891 2 : .await;
9892 : // Update GC info
9893 2 : let mut guard = tline.gc_info.write().unwrap();
9894 2 : guard.cutoffs.time = Some(Lsn(0x40));
9895 2 : guard.cutoffs.space = Lsn(0x40);
9896 : }
9897 2 : tline
9898 2 : .compact_with_gc(
9899 2 : &cancel,
9900 2 : CompactOptions::default_for_gc_compaction_unit_tests(),
9901 2 : &ctx,
9902 2 : )
9903 2 : .await
9904 2 : .unwrap();
9905 :
9906 2 : Ok(())
9907 2 : }
9908 :
9909 : #[cfg(feature = "testing")]
9910 : #[tokio::test]
9911 1 : async fn test_generate_key_retention() -> anyhow::Result<()> {
9912 1 : let harness = TenantHarness::create("test_generate_key_retention").await?;
9913 1 : let (tenant, ctx) = harness.load().await;
9914 1 : let tline = tenant
9915 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
9916 1 : .await?;
9917 1 : tline.force_advance_lsn(Lsn(0x70));
9918 1 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
9919 1 : let history = vec![
9920 1 : (
9921 1 : key,
9922 1 : Lsn(0x10),
9923 1 : Value::WalRecord(NeonWalRecord::wal_init("0x10")),
9924 1 : ),
9925 1 : (
9926 1 : key,
9927 1 : Lsn(0x20),
9928 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9929 1 : ),
9930 1 : (
9931 1 : key,
9932 1 : Lsn(0x30),
9933 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9934 1 : ),
9935 1 : (
9936 1 : key,
9937 1 : Lsn(0x40),
9938 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9939 1 : ),
9940 1 : (
9941 1 : key,
9942 1 : Lsn(0x50),
9943 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
9944 1 : ),
9945 1 : (
9946 1 : key,
9947 1 : Lsn(0x60),
9948 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9949 1 : ),
9950 1 : (
9951 1 : key,
9952 1 : Lsn(0x70),
9953 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9954 1 : ),
9955 1 : (
9956 1 : key,
9957 1 : Lsn(0x80),
9958 1 : Value::Image(Bytes::copy_from_slice(
9959 1 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9960 1 : )),
9961 1 : ),
9962 1 : (
9963 1 : key,
9964 1 : Lsn(0x90),
9965 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9966 1 : ),
9967 : ];
9968 1 : let res = tline
9969 1 : .generate_key_retention(
9970 1 : key,
9971 1 : &history,
9972 1 : Lsn(0x60),
9973 1 : &[Lsn(0x20), Lsn(0x40), Lsn(0x50)],
9974 1 : 3,
9975 1 : None,
9976 1 : true,
9977 1 : )
9978 1 : .await
9979 1 : .unwrap();
9980 1 : let expected_res = KeyHistoryRetention {
9981 1 : below_horizon: vec![
9982 1 : (
9983 1 : Lsn(0x20),
9984 1 : KeyLogAtLsn(vec![(
9985 1 : Lsn(0x20),
9986 1 : Value::Image(Bytes::from_static(b"0x10;0x20")),
9987 1 : )]),
9988 1 : ),
9989 1 : (
9990 1 : Lsn(0x40),
9991 1 : KeyLogAtLsn(vec![
9992 1 : (
9993 1 : Lsn(0x30),
9994 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9995 1 : ),
9996 1 : (
9997 1 : Lsn(0x40),
9998 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9999 1 : ),
10000 1 : ]),
10001 1 : ),
10002 1 : (
10003 1 : Lsn(0x50),
10004 1 : KeyLogAtLsn(vec![(
10005 1 : Lsn(0x50),
10006 1 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40;0x50")),
10007 1 : )]),
10008 1 : ),
10009 1 : (
10010 1 : Lsn(0x60),
10011 1 : KeyLogAtLsn(vec![(
10012 1 : Lsn(0x60),
10013 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
10014 1 : )]),
10015 1 : ),
10016 1 : ],
10017 1 : above_horizon: KeyLogAtLsn(vec![
10018 1 : (
10019 1 : Lsn(0x70),
10020 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
10021 1 : ),
10022 1 : (
10023 1 : Lsn(0x80),
10024 1 : Value::Image(Bytes::copy_from_slice(
10025 1 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
10026 1 : )),
10027 1 : ),
10028 1 : (
10029 1 : Lsn(0x90),
10030 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
10031 1 : ),
10032 1 : ]),
10033 1 : };
10034 1 : assert_eq!(res, expected_res);
10035 :
10036 : // We expect GC-compaction to run with the original GC. This would create a situation that
10037 : // the original GC algorithm removes some delta layers b/c there are full image coverage,
10038 : // therefore causing some keys to have an incomplete history below the lowest retain LSN.
10039 : // For example, we have
10040 : // ```plain
10041 : // init delta @ 0x10, image @ 0x20, delta @ 0x30 (gc_horizon), image @ 0x40.
10042 : // ```
10043 : // Now the GC horizon moves up, and we have
10044 : // ```plain
10045 : // init delta @ 0x10, image @ 0x20, delta @ 0x30, image @ 0x40 (gc_horizon)
10046 : // ```
10047 : // The original GC algorithm kicks in, and removes delta @ 0x10, image @ 0x20.
10048 : // We will end up with
10049 : // ```plain
10050 : // delta @ 0x30, image @ 0x40 (gc_horizon)
10051 : // ```
10052 : // Now we run the GC-compaction, and this key does not have a full history.
10053 : // We should be able to handle this partial history and drop everything before the
10054 : // gc_horizon image.
10055 :
10056 1 : let history = vec![
10057 1 : (
10058 1 : key,
10059 1 : Lsn(0x20),
10060 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
10061 1 : ),
10062 1 : (
10063 1 : key,
10064 1 : Lsn(0x30),
10065 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
10066 1 : ),
10067 1 : (
10068 1 : key,
10069 1 : Lsn(0x40),
10070 1 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
10071 1 : ),
10072 1 : (
10073 1 : key,
10074 1 : Lsn(0x50),
10075 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
10076 1 : ),
10077 1 : (
10078 1 : key,
10079 1 : Lsn(0x60),
10080 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
10081 1 : ),
10082 1 : (
10083 1 : key,
10084 1 : Lsn(0x70),
10085 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
10086 1 : ),
10087 1 : (
10088 1 : key,
10089 1 : Lsn(0x80),
10090 1 : Value::Image(Bytes::copy_from_slice(
10091 1 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
10092 1 : )),
10093 1 : ),
10094 1 : (
10095 1 : key,
10096 1 : Lsn(0x90),
10097 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
10098 1 : ),
10099 : ];
10100 1 : let res = tline
10101 1 : .generate_key_retention(
10102 1 : key,
10103 1 : &history,
10104 1 : Lsn(0x60),
10105 1 : &[Lsn(0x40), Lsn(0x50)],
10106 1 : 3,
10107 1 : None,
10108 1 : true,
10109 1 : )
10110 1 : .await
10111 1 : .unwrap();
10112 1 : let expected_res = KeyHistoryRetention {
10113 1 : below_horizon: vec![
10114 1 : (
10115 1 : Lsn(0x40),
10116 1 : KeyLogAtLsn(vec![(
10117 1 : Lsn(0x40),
10118 1 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
10119 1 : )]),
10120 1 : ),
10121 1 : (
10122 1 : Lsn(0x50),
10123 1 : KeyLogAtLsn(vec![(
10124 1 : Lsn(0x50),
10125 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
10126 1 : )]),
10127 1 : ),
10128 1 : (
10129 1 : Lsn(0x60),
10130 1 : KeyLogAtLsn(vec![(
10131 1 : Lsn(0x60),
10132 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
10133 1 : )]),
10134 1 : ),
10135 1 : ],
10136 1 : above_horizon: KeyLogAtLsn(vec![
10137 1 : (
10138 1 : Lsn(0x70),
10139 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
10140 1 : ),
10141 1 : (
10142 1 : Lsn(0x80),
10143 1 : Value::Image(Bytes::copy_from_slice(
10144 1 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
10145 1 : )),
10146 1 : ),
10147 1 : (
10148 1 : Lsn(0x90),
10149 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
10150 1 : ),
10151 1 : ]),
10152 1 : };
10153 1 : assert_eq!(res, expected_res);
10154 :
10155 : // In case of branch compaction, the branch itself does not have the full history, and we need to provide
10156 : // the ancestor image in the test case.
10157 :
10158 1 : let history = vec![
10159 1 : (
10160 1 : key,
10161 1 : Lsn(0x20),
10162 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
10163 1 : ),
10164 1 : (
10165 1 : key,
10166 1 : Lsn(0x30),
10167 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
10168 1 : ),
10169 1 : (
10170 1 : key,
10171 1 : Lsn(0x40),
10172 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
10173 1 : ),
10174 1 : (
10175 1 : key,
10176 1 : Lsn(0x70),
10177 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
10178 1 : ),
10179 : ];
10180 1 : let res = tline
10181 1 : .generate_key_retention(
10182 1 : key,
10183 1 : &history,
10184 1 : Lsn(0x60),
10185 1 : &[],
10186 1 : 3,
10187 1 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
10188 1 : true,
10189 1 : )
10190 1 : .await
10191 1 : .unwrap();
10192 1 : let expected_res = KeyHistoryRetention {
10193 1 : below_horizon: vec![(
10194 1 : Lsn(0x60),
10195 1 : KeyLogAtLsn(vec![(
10196 1 : Lsn(0x60),
10197 1 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")), // use the ancestor image to reconstruct the page
10198 1 : )]),
10199 1 : )],
10200 1 : above_horizon: KeyLogAtLsn(vec![(
10201 1 : Lsn(0x70),
10202 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
10203 1 : )]),
10204 1 : };
10205 1 : assert_eq!(res, expected_res);
10206 :
10207 1 : let history = vec![
10208 1 : (
10209 1 : key,
10210 1 : Lsn(0x20),
10211 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
10212 1 : ),
10213 1 : (
10214 1 : key,
10215 1 : Lsn(0x40),
10216 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
10217 1 : ),
10218 1 : (
10219 1 : key,
10220 1 : Lsn(0x60),
10221 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
10222 1 : ),
10223 1 : (
10224 1 : key,
10225 1 : Lsn(0x70),
10226 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
10227 1 : ),
10228 : ];
10229 1 : let res = tline
10230 1 : .generate_key_retention(
10231 1 : key,
10232 1 : &history,
10233 1 : Lsn(0x60),
10234 1 : &[Lsn(0x30)],
10235 1 : 3,
10236 1 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
10237 1 : true,
10238 1 : )
10239 1 : .await
10240 1 : .unwrap();
10241 1 : let expected_res = KeyHistoryRetention {
10242 1 : below_horizon: vec![
10243 1 : (
10244 1 : Lsn(0x30),
10245 1 : KeyLogAtLsn(vec![(
10246 1 : Lsn(0x20),
10247 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
10248 1 : )]),
10249 1 : ),
10250 1 : (
10251 1 : Lsn(0x60),
10252 1 : KeyLogAtLsn(vec![(
10253 1 : Lsn(0x60),
10254 1 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x40;0x60")),
10255 1 : )]),
10256 1 : ),
10257 1 : ],
10258 1 : above_horizon: KeyLogAtLsn(vec![(
10259 1 : Lsn(0x70),
10260 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
10261 1 : )]),
10262 1 : };
10263 1 : assert_eq!(res, expected_res);
10264 :
10265 2 : Ok(())
10266 1 : }
10267 :
10268 : #[cfg(feature = "testing")]
10269 : #[tokio::test]
10270 1 : async fn test_simple_bottom_most_compaction_with_retain_lsns() -> anyhow::Result<()> {
10271 1 : let harness =
10272 1 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns").await?;
10273 1 : let (tenant, ctx) = harness.load().await;
10274 :
10275 259 : fn get_key(id: u32) -> Key {
10276 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
10277 259 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
10278 259 : key.field6 = id;
10279 259 : key
10280 259 : }
10281 :
10282 1 : let img_layer = (0..10)
10283 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10284 1 : .collect_vec();
10285 :
10286 1 : let delta1 = vec![
10287 1 : (
10288 1 : get_key(1),
10289 1 : Lsn(0x20),
10290 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10291 1 : ),
10292 1 : (
10293 1 : get_key(2),
10294 1 : Lsn(0x30),
10295 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10296 1 : ),
10297 1 : (
10298 1 : get_key(3),
10299 1 : Lsn(0x28),
10300 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
10301 1 : ),
10302 1 : (
10303 1 : get_key(3),
10304 1 : Lsn(0x30),
10305 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10306 1 : ),
10307 1 : (
10308 1 : get_key(3),
10309 1 : Lsn(0x40),
10310 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
10311 1 : ),
10312 : ];
10313 1 : let delta2 = vec![
10314 1 : (
10315 1 : get_key(5),
10316 1 : Lsn(0x20),
10317 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10318 1 : ),
10319 1 : (
10320 1 : get_key(6),
10321 1 : Lsn(0x20),
10322 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10323 1 : ),
10324 : ];
10325 1 : let delta3 = vec![
10326 1 : (
10327 1 : get_key(8),
10328 1 : Lsn(0x48),
10329 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10330 1 : ),
10331 1 : (
10332 1 : get_key(9),
10333 1 : Lsn(0x48),
10334 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10335 1 : ),
10336 : ];
10337 :
10338 1 : let tline = tenant
10339 1 : .create_test_timeline_with_layers(
10340 1 : TIMELINE_ID,
10341 1 : Lsn(0x10),
10342 1 : DEFAULT_PG_VERSION,
10343 1 : &ctx,
10344 1 : Vec::new(), // in-memory layers
10345 1 : vec![
10346 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
10347 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
10348 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
10349 1 : ], // delta layers
10350 1 : vec![(Lsn(0x10), img_layer)], // image layers
10351 1 : Lsn(0x50),
10352 1 : )
10353 1 : .await?;
10354 : {
10355 1 : tline
10356 1 : .applied_gc_cutoff_lsn
10357 1 : .lock_for_write()
10358 1 : .store_and_unlock(Lsn(0x30))
10359 1 : .wait()
10360 1 : .await;
10361 : // Update GC info
10362 1 : let mut guard = tline.gc_info.write().unwrap();
10363 1 : *guard = GcInfo {
10364 1 : retain_lsns: vec![
10365 1 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
10366 1 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
10367 1 : ],
10368 1 : cutoffs: GcCutoffs {
10369 1 : time: Some(Lsn(0x30)),
10370 1 : space: Lsn(0x30),
10371 1 : },
10372 1 : leases: Default::default(),
10373 1 : within_ancestor_pitr: false,
10374 1 : };
10375 : }
10376 :
10377 1 : let expected_result = [
10378 1 : Bytes::from_static(b"value 0@0x10"),
10379 1 : Bytes::from_static(b"value 1@0x10@0x20"),
10380 1 : Bytes::from_static(b"value 2@0x10@0x30"),
10381 1 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
10382 1 : Bytes::from_static(b"value 4@0x10"),
10383 1 : Bytes::from_static(b"value 5@0x10@0x20"),
10384 1 : Bytes::from_static(b"value 6@0x10@0x20"),
10385 1 : Bytes::from_static(b"value 7@0x10"),
10386 1 : Bytes::from_static(b"value 8@0x10@0x48"),
10387 1 : Bytes::from_static(b"value 9@0x10@0x48"),
10388 1 : ];
10389 :
10390 1 : let expected_result_at_gc_horizon = [
10391 1 : Bytes::from_static(b"value 0@0x10"),
10392 1 : Bytes::from_static(b"value 1@0x10@0x20"),
10393 1 : Bytes::from_static(b"value 2@0x10@0x30"),
10394 1 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
10395 1 : Bytes::from_static(b"value 4@0x10"),
10396 1 : Bytes::from_static(b"value 5@0x10@0x20"),
10397 1 : Bytes::from_static(b"value 6@0x10@0x20"),
10398 1 : Bytes::from_static(b"value 7@0x10"),
10399 1 : Bytes::from_static(b"value 8@0x10"),
10400 1 : Bytes::from_static(b"value 9@0x10"),
10401 1 : ];
10402 :
10403 1 : let expected_result_at_lsn_20 = [
10404 1 : Bytes::from_static(b"value 0@0x10"),
10405 1 : Bytes::from_static(b"value 1@0x10@0x20"),
10406 1 : Bytes::from_static(b"value 2@0x10"),
10407 1 : Bytes::from_static(b"value 3@0x10"),
10408 1 : Bytes::from_static(b"value 4@0x10"),
10409 1 : Bytes::from_static(b"value 5@0x10@0x20"),
10410 1 : Bytes::from_static(b"value 6@0x10@0x20"),
10411 1 : Bytes::from_static(b"value 7@0x10"),
10412 1 : Bytes::from_static(b"value 8@0x10"),
10413 1 : Bytes::from_static(b"value 9@0x10"),
10414 1 : ];
10415 :
10416 1 : let expected_result_at_lsn_10 = [
10417 1 : Bytes::from_static(b"value 0@0x10"),
10418 1 : Bytes::from_static(b"value 1@0x10"),
10419 1 : Bytes::from_static(b"value 2@0x10"),
10420 1 : Bytes::from_static(b"value 3@0x10"),
10421 1 : Bytes::from_static(b"value 4@0x10"),
10422 1 : Bytes::from_static(b"value 5@0x10"),
10423 1 : Bytes::from_static(b"value 6@0x10"),
10424 1 : Bytes::from_static(b"value 7@0x10"),
10425 1 : Bytes::from_static(b"value 8@0x10"),
10426 1 : Bytes::from_static(b"value 9@0x10"),
10427 1 : ];
10428 :
10429 6 : let verify_result = || async {
10430 6 : let gc_horizon = {
10431 6 : let gc_info = tline.gc_info.read().unwrap();
10432 6 : gc_info.cutoffs.time.unwrap_or_default()
10433 : };
10434 66 : for idx in 0..10 {
10435 60 : assert_eq!(
10436 60 : tline
10437 60 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10438 60 : .await
10439 60 : .unwrap(),
10440 60 : &expected_result[idx]
10441 : );
10442 60 : assert_eq!(
10443 60 : tline
10444 60 : .get(get_key(idx as u32), gc_horizon, &ctx)
10445 60 : .await
10446 60 : .unwrap(),
10447 60 : &expected_result_at_gc_horizon[idx]
10448 : );
10449 60 : assert_eq!(
10450 60 : tline
10451 60 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
10452 60 : .await
10453 60 : .unwrap(),
10454 60 : &expected_result_at_lsn_20[idx]
10455 : );
10456 60 : assert_eq!(
10457 60 : tline
10458 60 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
10459 60 : .await
10460 60 : .unwrap(),
10461 60 : &expected_result_at_lsn_10[idx]
10462 : );
10463 : }
10464 12 : };
10465 :
10466 1 : verify_result().await;
10467 :
10468 1 : let cancel = CancellationToken::new();
10469 1 : let mut dryrun_flags = EnumSet::new();
10470 1 : dryrun_flags.insert(CompactFlags::DryRun);
10471 :
10472 1 : tline
10473 1 : .compact_with_gc(
10474 1 : &cancel,
10475 1 : CompactOptions {
10476 1 : flags: dryrun_flags,
10477 1 : ..CompactOptions::default_for_gc_compaction_unit_tests()
10478 1 : },
10479 1 : &ctx,
10480 1 : )
10481 1 : .await
10482 1 : .unwrap();
10483 : // We expect layer map to be the same b/c the dry run flag, but we don't know whether there will be other background jobs
10484 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
10485 1 : verify_result().await;
10486 :
10487 1 : tline
10488 1 : .compact_with_gc(
10489 1 : &cancel,
10490 1 : CompactOptions::default_for_gc_compaction_unit_tests(),
10491 1 : &ctx,
10492 1 : )
10493 1 : .await
10494 1 : .unwrap();
10495 1 : verify_result().await;
10496 :
10497 : // compact again
10498 1 : tline
10499 1 : .compact_with_gc(
10500 1 : &cancel,
10501 1 : CompactOptions::default_for_gc_compaction_unit_tests(),
10502 1 : &ctx,
10503 1 : )
10504 1 : .await
10505 1 : .unwrap();
10506 1 : verify_result().await;
10507 :
10508 : // increase GC horizon and compact again
10509 : {
10510 1 : tline
10511 1 : .applied_gc_cutoff_lsn
10512 1 : .lock_for_write()
10513 1 : .store_and_unlock(Lsn(0x38))
10514 1 : .wait()
10515 1 : .await;
10516 : // Update GC info
10517 1 : let mut guard = tline.gc_info.write().unwrap();
10518 1 : guard.cutoffs.time = Some(Lsn(0x38));
10519 1 : guard.cutoffs.space = Lsn(0x38);
10520 : }
10521 1 : tline
10522 1 : .compact_with_gc(
10523 1 : &cancel,
10524 1 : CompactOptions::default_for_gc_compaction_unit_tests(),
10525 1 : &ctx,
10526 1 : )
10527 1 : .await
10528 1 : .unwrap();
10529 1 : verify_result().await; // no wals between 0x30 and 0x38, so we should obtain the same result
10530 :
10531 : // not increasing the GC horizon and compact again
10532 1 : tline
10533 1 : .compact_with_gc(
10534 1 : &cancel,
10535 1 : CompactOptions::default_for_gc_compaction_unit_tests(),
10536 1 : &ctx,
10537 1 : )
10538 1 : .await
10539 1 : .unwrap();
10540 1 : verify_result().await;
10541 :
10542 2 : Ok(())
10543 1 : }
10544 :
10545 : #[cfg(feature = "testing")]
10546 : #[tokio::test]
10547 1 : async fn test_simple_bottom_most_compaction_with_retain_lsns_single_key() -> anyhow::Result<()>
10548 1 : {
10549 1 : let harness =
10550 1 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns_single_key")
10551 1 : .await?;
10552 1 : let (tenant, ctx) = harness.load().await;
10553 :
10554 176 : fn get_key(id: u32) -> Key {
10555 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
10556 176 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
10557 176 : key.field6 = id;
10558 176 : key
10559 176 : }
10560 :
10561 1 : let img_layer = (0..10)
10562 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10563 1 : .collect_vec();
10564 :
10565 1 : let delta1 = vec![
10566 1 : (
10567 1 : get_key(1),
10568 1 : Lsn(0x20),
10569 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10570 1 : ),
10571 1 : (
10572 1 : get_key(1),
10573 1 : Lsn(0x28),
10574 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
10575 1 : ),
10576 : ];
10577 1 : let delta2 = vec![
10578 1 : (
10579 1 : get_key(1),
10580 1 : Lsn(0x30),
10581 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10582 1 : ),
10583 1 : (
10584 1 : get_key(1),
10585 1 : Lsn(0x38),
10586 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
10587 1 : ),
10588 : ];
10589 1 : let delta3 = vec![
10590 1 : (
10591 1 : get_key(8),
10592 1 : Lsn(0x48),
10593 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10594 1 : ),
10595 1 : (
10596 1 : get_key(9),
10597 1 : Lsn(0x48),
10598 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10599 1 : ),
10600 : ];
10601 :
10602 1 : let tline = tenant
10603 1 : .create_test_timeline_with_layers(
10604 1 : TIMELINE_ID,
10605 1 : Lsn(0x10),
10606 1 : DEFAULT_PG_VERSION,
10607 1 : &ctx,
10608 1 : Vec::new(), // in-memory layers
10609 1 : vec![
10610 1 : // delta1 and delta 2 only contain a single key but multiple updates
10611 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x30), delta1),
10612 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
10613 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x50), delta3),
10614 1 : ], // delta layers
10615 1 : vec![(Lsn(0x10), img_layer)], // image layers
10616 1 : Lsn(0x50),
10617 1 : )
10618 1 : .await?;
10619 : {
10620 1 : tline
10621 1 : .applied_gc_cutoff_lsn
10622 1 : .lock_for_write()
10623 1 : .store_and_unlock(Lsn(0x30))
10624 1 : .wait()
10625 1 : .await;
10626 : // Update GC info
10627 1 : let mut guard = tline.gc_info.write().unwrap();
10628 1 : *guard = GcInfo {
10629 1 : retain_lsns: vec![
10630 1 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
10631 1 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
10632 1 : ],
10633 1 : cutoffs: GcCutoffs {
10634 1 : time: Some(Lsn(0x30)),
10635 1 : space: Lsn(0x30),
10636 1 : },
10637 1 : leases: Default::default(),
10638 1 : within_ancestor_pitr: false,
10639 1 : };
10640 : }
10641 :
10642 1 : let expected_result = [
10643 1 : Bytes::from_static(b"value 0@0x10"),
10644 1 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
10645 1 : Bytes::from_static(b"value 2@0x10"),
10646 1 : Bytes::from_static(b"value 3@0x10"),
10647 1 : Bytes::from_static(b"value 4@0x10"),
10648 1 : Bytes::from_static(b"value 5@0x10"),
10649 1 : Bytes::from_static(b"value 6@0x10"),
10650 1 : Bytes::from_static(b"value 7@0x10"),
10651 1 : Bytes::from_static(b"value 8@0x10@0x48"),
10652 1 : Bytes::from_static(b"value 9@0x10@0x48"),
10653 1 : ];
10654 :
10655 1 : let expected_result_at_gc_horizon = [
10656 1 : Bytes::from_static(b"value 0@0x10"),
10657 1 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
10658 1 : Bytes::from_static(b"value 2@0x10"),
10659 1 : Bytes::from_static(b"value 3@0x10"),
10660 1 : Bytes::from_static(b"value 4@0x10"),
10661 1 : Bytes::from_static(b"value 5@0x10"),
10662 1 : Bytes::from_static(b"value 6@0x10"),
10663 1 : Bytes::from_static(b"value 7@0x10"),
10664 1 : Bytes::from_static(b"value 8@0x10"),
10665 1 : Bytes::from_static(b"value 9@0x10"),
10666 1 : ];
10667 :
10668 1 : let expected_result_at_lsn_20 = [
10669 1 : Bytes::from_static(b"value 0@0x10"),
10670 1 : Bytes::from_static(b"value 1@0x10@0x20"),
10671 1 : Bytes::from_static(b"value 2@0x10"),
10672 1 : Bytes::from_static(b"value 3@0x10"),
10673 1 : Bytes::from_static(b"value 4@0x10"),
10674 1 : Bytes::from_static(b"value 5@0x10"),
10675 1 : Bytes::from_static(b"value 6@0x10"),
10676 1 : Bytes::from_static(b"value 7@0x10"),
10677 1 : Bytes::from_static(b"value 8@0x10"),
10678 1 : Bytes::from_static(b"value 9@0x10"),
10679 1 : ];
10680 :
10681 1 : let expected_result_at_lsn_10 = [
10682 1 : Bytes::from_static(b"value 0@0x10"),
10683 1 : Bytes::from_static(b"value 1@0x10"),
10684 1 : Bytes::from_static(b"value 2@0x10"),
10685 1 : Bytes::from_static(b"value 3@0x10"),
10686 1 : Bytes::from_static(b"value 4@0x10"),
10687 1 : Bytes::from_static(b"value 5@0x10"),
10688 1 : Bytes::from_static(b"value 6@0x10"),
10689 1 : Bytes::from_static(b"value 7@0x10"),
10690 1 : Bytes::from_static(b"value 8@0x10"),
10691 1 : Bytes::from_static(b"value 9@0x10"),
10692 1 : ];
10693 :
10694 4 : let verify_result = || async {
10695 4 : let gc_horizon = {
10696 4 : let gc_info = tline.gc_info.read().unwrap();
10697 4 : gc_info.cutoffs.time.unwrap_or_default()
10698 : };
10699 44 : for idx in 0..10 {
10700 40 : assert_eq!(
10701 40 : tline
10702 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10703 40 : .await
10704 40 : .unwrap(),
10705 40 : &expected_result[idx]
10706 : );
10707 40 : assert_eq!(
10708 40 : tline
10709 40 : .get(get_key(idx as u32), gc_horizon, &ctx)
10710 40 : .await
10711 40 : .unwrap(),
10712 40 : &expected_result_at_gc_horizon[idx]
10713 : );
10714 40 : assert_eq!(
10715 40 : tline
10716 40 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
10717 40 : .await
10718 40 : .unwrap(),
10719 40 : &expected_result_at_lsn_20[idx]
10720 : );
10721 40 : assert_eq!(
10722 40 : tline
10723 40 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
10724 40 : .await
10725 40 : .unwrap(),
10726 40 : &expected_result_at_lsn_10[idx]
10727 : );
10728 : }
10729 8 : };
10730 :
10731 1 : verify_result().await;
10732 :
10733 1 : let cancel = CancellationToken::new();
10734 1 : let mut dryrun_flags = EnumSet::new();
10735 1 : dryrun_flags.insert(CompactFlags::DryRun);
10736 :
10737 1 : tline
10738 1 : .compact_with_gc(
10739 1 : &cancel,
10740 1 : CompactOptions {
10741 1 : flags: dryrun_flags,
10742 1 : ..CompactOptions::default_for_gc_compaction_unit_tests()
10743 1 : },
10744 1 : &ctx,
10745 1 : )
10746 1 : .await
10747 1 : .unwrap();
10748 : // We expect layer map to be the same b/c the dry run flag, but we don't know whether there will be other background jobs
10749 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
10750 1 : verify_result().await;
10751 :
10752 1 : tline
10753 1 : .compact_with_gc(
10754 1 : &cancel,
10755 1 : CompactOptions::default_for_gc_compaction_unit_tests(),
10756 1 : &ctx,
10757 1 : )
10758 1 : .await
10759 1 : .unwrap();
10760 1 : verify_result().await;
10761 :
10762 : // compact again
10763 1 : tline
10764 1 : .compact_with_gc(
10765 1 : &cancel,
10766 1 : CompactOptions::default_for_gc_compaction_unit_tests(),
10767 1 : &ctx,
10768 1 : )
10769 1 : .await
10770 1 : .unwrap();
10771 1 : verify_result().await;
10772 :
10773 2 : Ok(())
10774 1 : }
10775 :
10776 : #[cfg(feature = "testing")]
10777 : #[tokio::test]
10778 1 : async fn test_simple_bottom_most_compaction_on_branch() -> anyhow::Result<()> {
10779 : use models::CompactLsnRange;
10780 :
10781 1 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_on_branch").await?;
10782 1 : let (tenant, ctx) = harness.load().await;
10783 :
10784 83 : fn get_key(id: u32) -> Key {
10785 83 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
10786 83 : key.field6 = id;
10787 83 : key
10788 83 : }
10789 :
10790 1 : let img_layer = (0..10)
10791 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10792 1 : .collect_vec();
10793 :
10794 1 : let delta1 = vec![
10795 1 : (
10796 1 : get_key(1),
10797 1 : Lsn(0x20),
10798 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10799 1 : ),
10800 1 : (
10801 1 : get_key(2),
10802 1 : Lsn(0x30),
10803 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10804 1 : ),
10805 1 : (
10806 1 : get_key(3),
10807 1 : Lsn(0x28),
10808 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
10809 1 : ),
10810 1 : (
10811 1 : get_key(3),
10812 1 : Lsn(0x30),
10813 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10814 1 : ),
10815 1 : (
10816 1 : get_key(3),
10817 1 : Lsn(0x40),
10818 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
10819 1 : ),
10820 : ];
10821 1 : let delta2 = vec![
10822 1 : (
10823 1 : get_key(5),
10824 1 : Lsn(0x20),
10825 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10826 1 : ),
10827 1 : (
10828 1 : get_key(6),
10829 1 : Lsn(0x20),
10830 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10831 1 : ),
10832 : ];
10833 1 : let delta3 = vec![
10834 1 : (
10835 1 : get_key(8),
10836 1 : Lsn(0x48),
10837 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10838 1 : ),
10839 1 : (
10840 1 : get_key(9),
10841 1 : Lsn(0x48),
10842 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10843 1 : ),
10844 : ];
10845 :
10846 1 : let parent_tline = tenant
10847 1 : .create_test_timeline_with_layers(
10848 1 : TIMELINE_ID,
10849 1 : Lsn(0x10),
10850 1 : DEFAULT_PG_VERSION,
10851 1 : &ctx,
10852 1 : vec![], // in-memory layers
10853 1 : vec![], // delta layers
10854 1 : vec![(Lsn(0x18), img_layer)], // image layers
10855 1 : Lsn(0x18),
10856 1 : )
10857 1 : .await?;
10858 :
10859 1 : parent_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
10860 :
10861 1 : let branch_tline = tenant
10862 1 : .branch_timeline_test_with_layers(
10863 1 : &parent_tline,
10864 1 : NEW_TIMELINE_ID,
10865 1 : Some(Lsn(0x18)),
10866 1 : &ctx,
10867 1 : vec![
10868 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
10869 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
10870 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
10871 1 : ], // delta layers
10872 1 : vec![], // image layers
10873 1 : Lsn(0x50),
10874 1 : )
10875 1 : .await?;
10876 :
10877 1 : branch_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
10878 :
10879 : {
10880 1 : parent_tline
10881 1 : .applied_gc_cutoff_lsn
10882 1 : .lock_for_write()
10883 1 : .store_and_unlock(Lsn(0x10))
10884 1 : .wait()
10885 1 : .await;
10886 : // Update GC info
10887 1 : let mut guard = parent_tline.gc_info.write().unwrap();
10888 1 : *guard = GcInfo {
10889 1 : retain_lsns: vec![(Lsn(0x18), branch_tline.timeline_id, MaybeOffloaded::No)],
10890 1 : cutoffs: GcCutoffs {
10891 1 : time: Some(Lsn(0x10)),
10892 1 : space: Lsn(0x10),
10893 1 : },
10894 1 : leases: Default::default(),
10895 1 : within_ancestor_pitr: false,
10896 1 : };
10897 : }
10898 :
10899 : {
10900 1 : branch_tline
10901 1 : .applied_gc_cutoff_lsn
10902 1 : .lock_for_write()
10903 1 : .store_and_unlock(Lsn(0x50))
10904 1 : .wait()
10905 1 : .await;
10906 : // Update GC info
10907 1 : let mut guard = branch_tline.gc_info.write().unwrap();
10908 1 : *guard = GcInfo {
10909 1 : retain_lsns: vec![(Lsn(0x40), branch_tline.timeline_id, MaybeOffloaded::No)],
10910 1 : cutoffs: GcCutoffs {
10911 1 : time: Some(Lsn(0x50)),
10912 1 : space: Lsn(0x50),
10913 1 : },
10914 1 : leases: Default::default(),
10915 1 : within_ancestor_pitr: false,
10916 1 : };
10917 : }
10918 :
10919 1 : let expected_result_at_gc_horizon = [
10920 1 : Bytes::from_static(b"value 0@0x10"),
10921 1 : Bytes::from_static(b"value 1@0x10@0x20"),
10922 1 : Bytes::from_static(b"value 2@0x10@0x30"),
10923 1 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
10924 1 : Bytes::from_static(b"value 4@0x10"),
10925 1 : Bytes::from_static(b"value 5@0x10@0x20"),
10926 1 : Bytes::from_static(b"value 6@0x10@0x20"),
10927 1 : Bytes::from_static(b"value 7@0x10"),
10928 1 : Bytes::from_static(b"value 8@0x10@0x48"),
10929 1 : Bytes::from_static(b"value 9@0x10@0x48"),
10930 1 : ];
10931 :
10932 1 : let expected_result_at_lsn_40 = [
10933 1 : Bytes::from_static(b"value 0@0x10"),
10934 1 : Bytes::from_static(b"value 1@0x10@0x20"),
10935 1 : Bytes::from_static(b"value 2@0x10@0x30"),
10936 1 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
10937 1 : Bytes::from_static(b"value 4@0x10"),
10938 1 : Bytes::from_static(b"value 5@0x10@0x20"),
10939 1 : Bytes::from_static(b"value 6@0x10@0x20"),
10940 1 : Bytes::from_static(b"value 7@0x10"),
10941 1 : Bytes::from_static(b"value 8@0x10"),
10942 1 : Bytes::from_static(b"value 9@0x10"),
10943 1 : ];
10944 :
10945 3 : let verify_result = || async {
10946 33 : for idx in 0..10 {
10947 30 : assert_eq!(
10948 30 : branch_tline
10949 30 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10950 30 : .await
10951 30 : .unwrap(),
10952 30 : &expected_result_at_gc_horizon[idx]
10953 : );
10954 30 : assert_eq!(
10955 30 : branch_tline
10956 30 : .get(get_key(idx as u32), Lsn(0x40), &ctx)
10957 30 : .await
10958 30 : .unwrap(),
10959 30 : &expected_result_at_lsn_40[idx]
10960 : );
10961 : }
10962 6 : };
10963 :
10964 1 : verify_result().await;
10965 :
10966 1 : let cancel = CancellationToken::new();
10967 1 : branch_tline
10968 1 : .compact_with_gc(
10969 1 : &cancel,
10970 1 : CompactOptions::default_for_gc_compaction_unit_tests(),
10971 1 : &ctx,
10972 1 : )
10973 1 : .await
10974 1 : .unwrap();
10975 :
10976 1 : verify_result().await;
10977 :
10978 : // Piggyback a compaction with above_lsn. Ensure it works correctly when the specified LSN intersects with the layer files.
10979 : // Now we already have a single large delta layer, so the compaction min_layer_lsn should be the same as ancestor LSN (0x18).
10980 1 : branch_tline
10981 1 : .compact_with_gc(
10982 1 : &cancel,
10983 1 : CompactOptions {
10984 1 : compact_lsn_range: Some(CompactLsnRange::above(Lsn(0x40))),
10985 1 : ..CompactOptions::default_for_gc_compaction_unit_tests()
10986 1 : },
10987 1 : &ctx,
10988 1 : )
10989 1 : .await
10990 1 : .unwrap();
10991 :
10992 1 : verify_result().await;
10993 :
10994 2 : Ok(())
10995 1 : }
10996 :
10997 : // Regression test for https://github.com/neondatabase/neon/issues/9012
10998 : // Create an image arrangement where we have to read at different LSN ranges
10999 : // from a delta layer. This is achieved by overlapping an image layer on top of
11000 : // a delta layer. Like so:
11001 : //
11002 : // A B
11003 : // +----------------+ -> delta_layer
11004 : // | | ^ lsn
11005 : // | =========|-> nested_image_layer |
11006 : // | C | |
11007 : // +----------------+ |
11008 : // ======== -> baseline_image_layer +-------> key
11009 : //
11010 : //
11011 : // When querying the key range [A, B) we need to read at different LSN ranges
11012 : // for [A, C) and [C, B). This test checks that the described edge case is handled correctly.
11013 : #[cfg(feature = "testing")]
11014 : #[tokio::test]
11015 1 : async fn test_vectored_read_with_nested_image_layer() -> anyhow::Result<()> {
11016 1 : let harness = TenantHarness::create("test_vectored_read_with_nested_image_layer").await?;
11017 1 : let (tenant, ctx) = harness.load().await;
11018 :
11019 1 : let will_init_keys = [2, 6];
11020 22 : fn get_key(id: u32) -> Key {
11021 22 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
11022 22 : key.field6 = id;
11023 22 : key
11024 22 : }
11025 :
11026 1 : let mut expected_key_values = HashMap::new();
11027 :
11028 1 : let baseline_image_layer_lsn = Lsn(0x10);
11029 1 : let mut baseline_img_layer = Vec::new();
11030 6 : for i in 0..5 {
11031 5 : let key = get_key(i);
11032 5 : let value = format!("value {i}@{baseline_image_layer_lsn}");
11033 :
11034 5 : let removed = expected_key_values.insert(key, value.clone());
11035 5 : assert!(removed.is_none());
11036 :
11037 5 : baseline_img_layer.push((key, Bytes::from(value)));
11038 : }
11039 :
11040 1 : let nested_image_layer_lsn = Lsn(0x50);
11041 1 : let mut nested_img_layer = Vec::new();
11042 6 : for i in 5..10 {
11043 5 : let key = get_key(i);
11044 5 : let value = format!("value {i}@{nested_image_layer_lsn}");
11045 :
11046 5 : let removed = expected_key_values.insert(key, value.clone());
11047 5 : assert!(removed.is_none());
11048 :
11049 5 : nested_img_layer.push((key, Bytes::from(value)));
11050 : }
11051 :
11052 1 : let mut delta_layer_spec = Vec::default();
11053 1 : let delta_layer_start_lsn = Lsn(0x20);
11054 1 : let mut delta_layer_end_lsn = delta_layer_start_lsn;
11055 :
11056 11 : for i in 0..10 {
11057 10 : let key = get_key(i);
11058 10 : let key_in_nested = nested_img_layer
11059 10 : .iter()
11060 40 : .any(|(key_with_img, _)| *key_with_img == key);
11061 10 : let lsn = {
11062 10 : if key_in_nested {
11063 5 : Lsn(nested_image_layer_lsn.0 + 0x10)
11064 : } else {
11065 5 : delta_layer_start_lsn
11066 : }
11067 : };
11068 :
11069 10 : let will_init = will_init_keys.contains(&i);
11070 10 : if will_init {
11071 2 : delta_layer_spec.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init(""))));
11072 2 :
11073 2 : expected_key_values.insert(key, "".to_string());
11074 8 : } else {
11075 8 : let delta = format!("@{lsn}");
11076 8 : delta_layer_spec.push((
11077 8 : key,
11078 8 : lsn,
11079 8 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
11080 8 : ));
11081 8 :
11082 8 : expected_key_values
11083 8 : .get_mut(&key)
11084 8 : .expect("An image exists for each key")
11085 8 : .push_str(delta.as_str());
11086 8 : }
11087 10 : delta_layer_end_lsn = std::cmp::max(delta_layer_start_lsn, lsn);
11088 : }
11089 :
11090 1 : delta_layer_end_lsn = Lsn(delta_layer_end_lsn.0 + 1);
11091 :
11092 1 : assert!(
11093 1 : nested_image_layer_lsn > delta_layer_start_lsn
11094 1 : && nested_image_layer_lsn < delta_layer_end_lsn
11095 : );
11096 :
11097 1 : let tline = tenant
11098 1 : .create_test_timeline_with_layers(
11099 1 : TIMELINE_ID,
11100 1 : baseline_image_layer_lsn,
11101 1 : DEFAULT_PG_VERSION,
11102 1 : &ctx,
11103 1 : vec![], // in-memory layers
11104 1 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
11105 1 : delta_layer_start_lsn..delta_layer_end_lsn,
11106 1 : delta_layer_spec,
11107 1 : )], // delta layers
11108 1 : vec![
11109 1 : (baseline_image_layer_lsn, baseline_img_layer),
11110 1 : (nested_image_layer_lsn, nested_img_layer),
11111 1 : ], // image layers
11112 1 : delta_layer_end_lsn,
11113 1 : )
11114 1 : .await?;
11115 :
11116 1 : let query = VersionedKeySpaceQuery::uniform(
11117 1 : KeySpace::single(get_key(0)..get_key(10)),
11118 1 : delta_layer_end_lsn,
11119 : );
11120 :
11121 1 : let results = tline
11122 1 : .get_vectored(query, IoConcurrency::sequential(), &ctx)
11123 1 : .await
11124 1 : .expect("No vectored errors");
11125 11 : for (key, res) in results {
11126 10 : let value = res.expect("No key errors");
11127 10 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
11128 10 : assert_eq!(value, Bytes::from(expected_value));
11129 1 : }
11130 1 :
11131 1 : Ok(())
11132 1 : }
11133 :
11134 : #[cfg(feature = "testing")]
11135 : #[tokio::test]
11136 1 : async fn test_vectored_read_with_image_layer_inside_inmem() -> anyhow::Result<()> {
11137 1 : let harness =
11138 1 : TenantHarness::create("test_vectored_read_with_image_layer_inside_inmem").await?;
11139 1 : let (tenant, ctx) = harness.load().await;
11140 :
11141 1 : let will_init_keys = [2, 6];
11142 32 : fn get_key(id: u32) -> Key {
11143 32 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
11144 32 : key.field6 = id;
11145 32 : key
11146 32 : }
11147 :
11148 1 : let mut expected_key_values = HashMap::new();
11149 :
11150 1 : let baseline_image_layer_lsn = Lsn(0x10);
11151 1 : let mut baseline_img_layer = Vec::new();
11152 6 : for i in 0..5 {
11153 5 : let key = get_key(i);
11154 5 : let value = format!("value {i}@{baseline_image_layer_lsn}");
11155 :
11156 5 : let removed = expected_key_values.insert(key, value.clone());
11157 5 : assert!(removed.is_none());
11158 :
11159 5 : baseline_img_layer.push((key, Bytes::from(value)));
11160 : }
11161 :
11162 1 : let nested_image_layer_lsn = Lsn(0x50);
11163 1 : let mut nested_img_layer = Vec::new();
11164 6 : for i in 5..10 {
11165 5 : let key = get_key(i);
11166 5 : let value = format!("value {i}@{nested_image_layer_lsn}");
11167 :
11168 5 : let removed = expected_key_values.insert(key, value.clone());
11169 5 : assert!(removed.is_none());
11170 :
11171 5 : nested_img_layer.push((key, Bytes::from(value)));
11172 : }
11173 :
11174 1 : let frozen_layer = {
11175 1 : let lsn_range = Lsn(0x40)..Lsn(0x60);
11176 1 : let mut data = Vec::new();
11177 11 : for i in 0..10 {
11178 10 : let key = get_key(i);
11179 10 : let key_in_nested = nested_img_layer
11180 10 : .iter()
11181 40 : .any(|(key_with_img, _)| *key_with_img == key);
11182 10 : let lsn = {
11183 10 : if key_in_nested {
11184 5 : Lsn(nested_image_layer_lsn.0 + 5)
11185 : } else {
11186 5 : lsn_range.start
11187 : }
11188 : };
11189 :
11190 10 : let will_init = will_init_keys.contains(&i);
11191 10 : if will_init {
11192 2 : data.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init(""))));
11193 2 :
11194 2 : expected_key_values.insert(key, "".to_string());
11195 8 : } else {
11196 8 : let delta = format!("@{lsn}");
11197 8 : data.push((
11198 8 : key,
11199 8 : lsn,
11200 8 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
11201 8 : ));
11202 8 :
11203 8 : expected_key_values
11204 8 : .get_mut(&key)
11205 8 : .expect("An image exists for each key")
11206 8 : .push_str(delta.as_str());
11207 8 : }
11208 : }
11209 :
11210 1 : InMemoryLayerTestDesc {
11211 1 : lsn_range,
11212 1 : is_open: false,
11213 1 : data,
11214 1 : }
11215 : };
11216 :
11217 1 : let (open_layer, last_record_lsn) = {
11218 1 : let start_lsn = Lsn(0x70);
11219 1 : let mut data = Vec::new();
11220 1 : let mut end_lsn = Lsn(0);
11221 11 : for i in 0..10 {
11222 10 : let key = get_key(i);
11223 10 : let lsn = Lsn(start_lsn.0 + i as u64);
11224 10 : let delta = format!("@{lsn}");
11225 10 : data.push((
11226 10 : key,
11227 10 : lsn,
11228 10 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
11229 10 : ));
11230 10 :
11231 10 : expected_key_values
11232 10 : .get_mut(&key)
11233 10 : .expect("An image exists for each key")
11234 10 : .push_str(delta.as_str());
11235 10 :
11236 10 : end_lsn = std::cmp::max(end_lsn, lsn);
11237 10 : }
11238 :
11239 1 : (
11240 1 : InMemoryLayerTestDesc {
11241 1 : lsn_range: start_lsn..Lsn::MAX,
11242 1 : is_open: true,
11243 1 : data,
11244 1 : },
11245 1 : end_lsn,
11246 1 : )
11247 : };
11248 :
11249 1 : assert!(
11250 1 : nested_image_layer_lsn > frozen_layer.lsn_range.start
11251 1 : && nested_image_layer_lsn < frozen_layer.lsn_range.end
11252 : );
11253 :
11254 1 : let tline = tenant
11255 1 : .create_test_timeline_with_layers(
11256 1 : TIMELINE_ID,
11257 1 : baseline_image_layer_lsn,
11258 1 : DEFAULT_PG_VERSION,
11259 1 : &ctx,
11260 1 : vec![open_layer, frozen_layer], // in-memory layers
11261 1 : Vec::new(), // delta layers
11262 1 : vec![
11263 1 : (baseline_image_layer_lsn, baseline_img_layer),
11264 1 : (nested_image_layer_lsn, nested_img_layer),
11265 1 : ], // image layers
11266 1 : last_record_lsn,
11267 1 : )
11268 1 : .await?;
11269 :
11270 1 : let query = VersionedKeySpaceQuery::uniform(
11271 1 : KeySpace::single(get_key(0)..get_key(10)),
11272 1 : last_record_lsn,
11273 : );
11274 :
11275 1 : let results = tline
11276 1 : .get_vectored(query, IoConcurrency::sequential(), &ctx)
11277 1 : .await
11278 1 : .expect("No vectored errors");
11279 11 : for (key, res) in results {
11280 10 : let value = res.expect("No key errors");
11281 10 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
11282 10 : assert_eq!(value, Bytes::from(expected_value.clone()));
11283 1 :
11284 10 : tracing::info!("key={key} value={expected_value}");
11285 1 : }
11286 1 :
11287 1 : Ok(())
11288 1 : }
11289 :
11290 : // A randomized read path test. Generates a layer map according to a deterministic
11291 : // specification. Fills the (key, LSN) space in random manner and then performs
11292 : // random scattered queries validating the results against in-memory storage.
11293 : //
11294 : // See this internal Notion page for a diagram of the layer map:
11295 : // https://www.notion.so/neondatabase/Read-Path-Unit-Testing-Fuzzing-1d1f189e0047806c8e5cd37781b0a350?pvs=4
11296 : //
11297 : // A fuzzing mode is also supported. In this mode, the test will use a random
11298 : // seed instead of a hardcoded one. Use it in conjunction with `cargo stress`
11299 : // to run multiple instances in parallel:
11300 : //
11301 : // $ RUST_BACKTRACE=1 RUST_LOG=INFO \
11302 : // cargo stress --package=pageserver --features=testing,fuzz-read-path --release -- test_read_path
11303 : #[cfg(feature = "testing")]
11304 : #[tokio::test]
11305 1 : async fn test_read_path() -> anyhow::Result<()> {
11306 : use rand::seq::IndexedRandom;
11307 :
11308 1 : let seed = if cfg!(feature = "fuzz-read-path") {
11309 0 : let seed: u64 = rand::rng().random();
11310 0 : seed
11311 : } else {
11312 : // Use a hard-coded seed when not in fuzzing mode.
11313 : // Note that with the current approach results are not reproducible
11314 : // accross platforms and Rust releases.
11315 : const SEED: u64 = 0;
11316 1 : SEED
11317 : };
11318 :
11319 1 : let mut random = StdRng::seed_from_u64(seed);
11320 :
11321 1 : let (queries, will_init_chance, gap_chance) = if cfg!(feature = "fuzz-read-path") {
11322 : const QUERIES: u64 = 5000;
11323 0 : let will_init_chance: u8 = random.random_range(0..=10);
11324 0 : let gap_chance: u8 = random.random_range(0..=50);
11325 :
11326 0 : (QUERIES, will_init_chance, gap_chance)
11327 : } else {
11328 : const QUERIES: u64 = 1000;
11329 : const WILL_INIT_CHANCE: u8 = 1;
11330 : const GAP_CHANCE: u8 = 5;
11331 :
11332 1 : (QUERIES, WILL_INIT_CHANCE, GAP_CHANCE)
11333 : };
11334 :
11335 1 : let harness = TenantHarness::create("test_read_path").await?;
11336 1 : let (tenant, ctx) = harness.load().await;
11337 :
11338 1 : tracing::info!("Using random seed: {seed}");
11339 1 : tracing::info!(%will_init_chance, %gap_chance, "Fill params");
11340 :
11341 : // Define the layer map shape. Note that this part is not randomized.
11342 :
11343 : const KEY_DIMENSION_SIZE: u32 = 99;
11344 1 : let start_key = Key::from_hex("110000000033333333444444445500000000").unwrap();
11345 1 : let end_key = start_key.add(KEY_DIMENSION_SIZE);
11346 1 : let total_key_range = start_key..end_key;
11347 1 : let total_key_range_size = end_key.to_i128() - start_key.to_i128();
11348 1 : let total_start_lsn = Lsn(104);
11349 1 : let last_record_lsn = Lsn(504);
11350 :
11351 1 : assert!(total_key_range_size % 3 == 0);
11352 :
11353 1 : let in_memory_layers_shape = vec![
11354 1 : (total_key_range.clone(), Lsn(304)..Lsn(400)),
11355 1 : (total_key_range.clone(), Lsn(400)..last_record_lsn),
11356 : ];
11357 :
11358 1 : let delta_layers_shape = vec![
11359 1 : (
11360 1 : start_key..(start_key.add((total_key_range_size / 3) as u32)),
11361 1 : Lsn(200)..Lsn(304),
11362 1 : ),
11363 1 : (
11364 1 : (start_key.add((total_key_range_size / 3) as u32))
11365 1 : ..(start_key.add((total_key_range_size * 2 / 3) as u32)),
11366 1 : Lsn(200)..Lsn(304),
11367 1 : ),
11368 1 : (
11369 1 : (start_key.add((total_key_range_size * 2 / 3) as u32))
11370 1 : ..(start_key.add(total_key_range_size as u32)),
11371 1 : Lsn(200)..Lsn(304),
11372 1 : ),
11373 : ];
11374 :
11375 1 : let image_layers_shape = vec![
11376 1 : (
11377 1 : start_key.add((total_key_range_size * 2 / 3 - 10) as u32)
11378 1 : ..start_key.add((total_key_range_size * 2 / 3 + 10) as u32),
11379 1 : Lsn(456),
11380 1 : ),
11381 1 : (
11382 1 : start_key.add((total_key_range_size / 3 - 10) as u32)
11383 1 : ..start_key.add((total_key_range_size / 3 + 10) as u32),
11384 1 : Lsn(256),
11385 1 : ),
11386 1 : (total_key_range.clone(), total_start_lsn),
11387 : ];
11388 :
11389 1 : let specification = TestTimelineSpecification {
11390 1 : start_lsn: total_start_lsn,
11391 1 : last_record_lsn,
11392 1 : in_memory_layers_shape,
11393 1 : delta_layers_shape,
11394 1 : image_layers_shape,
11395 1 : gap_chance,
11396 1 : will_init_chance,
11397 1 : };
11398 :
11399 : // Create and randomly fill in the layers according to the specification
11400 1 : let (tline, storage, interesting_lsns) = randomize_timeline(
11401 1 : &tenant,
11402 1 : TIMELINE_ID,
11403 1 : DEFAULT_PG_VERSION,
11404 1 : specification,
11405 1 : &mut random,
11406 1 : &ctx,
11407 1 : )
11408 1 : .await?;
11409 :
11410 : // Now generate queries based on the interesting lsns that we've collected.
11411 : //
11412 : // While there's still room in the query, pick and interesting LSN and a random
11413 : // key. Then roll the dice to see if the next key should also be included in
11414 : // the query. When the roll fails, break the "batch" and pick another point in the
11415 : // (key, LSN) space.
11416 :
11417 : const PICK_NEXT_CHANCE: u8 = 50;
11418 1 : for _ in 0..queries {
11419 1000 : let query = {
11420 1000 : let mut keyspaces_at_lsn: HashMap<Lsn, KeySpaceRandomAccum> = HashMap::default();
11421 1000 : let mut used_keys: HashSet<Key> = HashSet::default();
11422 1 :
11423 22401 : while used_keys.len() < tenant.conf.max_get_vectored_keys.get() {
11424 21401 : let selected_lsn = interesting_lsns.choose(&mut random).expect("not empty");
11425 21401 : let mut selected_key =
11426 21401 : start_key.add(random.random_range(0..KEY_DIMENSION_SIZE));
11427 1 :
11428 37626 : while used_keys.len() < tenant.conf.max_get_vectored_keys.get() {
11429 37104 : if used_keys.contains(&selected_key)
11430 32144 : || selected_key >= start_key.add(KEY_DIMENSION_SIZE)
11431 1 : {
11432 5104 : break;
11433 32000 : }
11434 1 :
11435 32000 : keyspaces_at_lsn
11436 32000 : .entry(*selected_lsn)
11437 32000 : .or_default()
11438 32000 : .add_key(selected_key);
11439 32000 : used_keys.insert(selected_key);
11440 1 :
11441 32000 : let pick_next = random.random_range(0..=100) <= PICK_NEXT_CHANCE;
11442 32000 : if pick_next {
11443 16225 : selected_key = selected_key.next();
11444 16225 : } else {
11445 15775 : break;
11446 1 : }
11447 1 : }
11448 1 : }
11449 1 :
11450 1000 : VersionedKeySpaceQuery::scattered(
11451 1000 : keyspaces_at_lsn
11452 1000 : .into_iter()
11453 11916 : .map(|(lsn, acc)| (lsn, acc.to_keyspace()))
11454 1000 : .collect(),
11455 1 : )
11456 1 : };
11457 1 :
11458 1 : // Run the query and validate the results
11459 1 :
11460 1000 : let results = tline
11461 1000 : .get_vectored(query.clone(), IoConcurrency::Sequential, &ctx)
11462 1000 : .await;
11463 1 :
11464 1000 : let blobs = match results {
11465 1000 : Ok(ok) => ok,
11466 1 : Err(err) => {
11467 1 : panic!("seed={seed} Error returned for query {query}: {err}");
11468 1 : }
11469 1 : };
11470 1 :
11471 32000 : for (key, key_res) in blobs.into_iter() {
11472 32000 : match key_res {
11473 32000 : Ok(blob) => {
11474 32000 : let requested_at_lsn = query.map_key_to_lsn(&key);
11475 32000 : let expected = storage.get(key, requested_at_lsn);
11476 1 :
11477 32000 : if blob != expected {
11478 1 : tracing::error!(
11479 1 : "seed={seed} Mismatch for {key}@{requested_at_lsn} from query: {query}"
11480 1 : );
11481 32000 : }
11482 1 :
11483 32000 : assert_eq!(blob, expected);
11484 1 : }
11485 1 : Err(err) => {
11486 1 : let requested_at_lsn = query.map_key_to_lsn(&key);
11487 1 :
11488 1 : panic!(
11489 1 : "seed={seed} Error returned for {key}@{requested_at_lsn} from query {query}: {err}"
11490 1 : );
11491 1 : }
11492 1 : }
11493 1 : }
11494 1 : }
11495 1 :
11496 1 : Ok(())
11497 1 : }
11498 :
11499 107 : fn sort_layer_key(k1: &PersistentLayerKey, k2: &PersistentLayerKey) -> std::cmp::Ordering {
11500 107 : (
11501 107 : k1.is_delta,
11502 107 : k1.key_range.start,
11503 107 : k1.key_range.end,
11504 107 : k1.lsn_range.start,
11505 107 : k1.lsn_range.end,
11506 107 : )
11507 107 : .cmp(&(
11508 107 : k2.is_delta,
11509 107 : k2.key_range.start,
11510 107 : k2.key_range.end,
11511 107 : k2.lsn_range.start,
11512 107 : k2.lsn_range.end,
11513 107 : ))
11514 107 : }
11515 :
11516 12 : async fn inspect_and_sort(
11517 12 : tline: &Arc<Timeline>,
11518 12 : filter: Option<std::ops::Range<Key>>,
11519 12 : ) -> Vec<PersistentLayerKey> {
11520 12 : let mut all_layers = tline.inspect_historic_layers().await.unwrap();
11521 12 : if let Some(filter) = filter {
11522 54 : all_layers.retain(|layer| overlaps_with(&layer.key_range, &filter));
11523 1 : }
11524 12 : all_layers.sort_by(sort_layer_key);
11525 12 : all_layers
11526 12 : }
11527 :
11528 : #[cfg(feature = "testing")]
11529 11 : fn check_layer_map_key_eq(
11530 11 : mut left: Vec<PersistentLayerKey>,
11531 11 : mut right: Vec<PersistentLayerKey>,
11532 11 : ) {
11533 11 : left.sort_by(sort_layer_key);
11534 11 : right.sort_by(sort_layer_key);
11535 11 : if left != right {
11536 0 : eprintln!("---LEFT---");
11537 0 : for left in left.iter() {
11538 0 : eprintln!("{left}");
11539 0 : }
11540 0 : eprintln!("---RIGHT---");
11541 0 : for right in right.iter() {
11542 0 : eprintln!("{right}");
11543 0 : }
11544 0 : assert_eq!(left, right);
11545 11 : }
11546 11 : }
11547 :
11548 : #[cfg(feature = "testing")]
11549 : #[tokio::test]
11550 1 : async fn test_simple_partial_bottom_most_compaction() -> anyhow::Result<()> {
11551 1 : let harness = TenantHarness::create("test_simple_partial_bottom_most_compaction").await?;
11552 1 : let (tenant, ctx) = harness.load().await;
11553 :
11554 91 : fn get_key(id: u32) -> Key {
11555 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
11556 91 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
11557 91 : key.field6 = id;
11558 91 : key
11559 91 : }
11560 :
11561 : // img layer at 0x10
11562 1 : let img_layer = (0..10)
11563 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
11564 1 : .collect_vec();
11565 :
11566 1 : let delta1 = vec![
11567 1 : (
11568 1 : get_key(1),
11569 1 : Lsn(0x20),
11570 1 : Value::Image(Bytes::from("value 1@0x20")),
11571 1 : ),
11572 1 : (
11573 1 : get_key(2),
11574 1 : Lsn(0x30),
11575 1 : Value::Image(Bytes::from("value 2@0x30")),
11576 1 : ),
11577 1 : (
11578 1 : get_key(3),
11579 1 : Lsn(0x40),
11580 1 : Value::Image(Bytes::from("value 3@0x40")),
11581 1 : ),
11582 : ];
11583 1 : let delta2 = vec![
11584 1 : (
11585 1 : get_key(5),
11586 1 : Lsn(0x20),
11587 1 : Value::Image(Bytes::from("value 5@0x20")),
11588 1 : ),
11589 1 : (
11590 1 : get_key(6),
11591 1 : Lsn(0x20),
11592 1 : Value::Image(Bytes::from("value 6@0x20")),
11593 1 : ),
11594 : ];
11595 1 : let delta3 = vec![
11596 1 : (
11597 1 : get_key(8),
11598 1 : Lsn(0x48),
11599 1 : Value::Image(Bytes::from("value 8@0x48")),
11600 1 : ),
11601 1 : (
11602 1 : get_key(9),
11603 1 : Lsn(0x48),
11604 1 : Value::Image(Bytes::from("value 9@0x48")),
11605 1 : ),
11606 : ];
11607 :
11608 1 : let tline = tenant
11609 1 : .create_test_timeline_with_layers(
11610 1 : TIMELINE_ID,
11611 1 : Lsn(0x10),
11612 1 : DEFAULT_PG_VERSION,
11613 1 : &ctx,
11614 1 : vec![], // in-memory layers
11615 1 : vec![
11616 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
11617 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
11618 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
11619 1 : ], // delta layers
11620 1 : vec![(Lsn(0x10), img_layer)], // image layers
11621 1 : Lsn(0x50),
11622 1 : )
11623 1 : .await?;
11624 :
11625 : {
11626 1 : tline
11627 1 : .applied_gc_cutoff_lsn
11628 1 : .lock_for_write()
11629 1 : .store_and_unlock(Lsn(0x30))
11630 1 : .wait()
11631 1 : .await;
11632 : // Update GC info
11633 1 : let mut guard = tline.gc_info.write().unwrap();
11634 1 : *guard = GcInfo {
11635 1 : retain_lsns: vec![(Lsn(0x20), tline.timeline_id, MaybeOffloaded::No)],
11636 1 : cutoffs: GcCutoffs {
11637 1 : time: Some(Lsn(0x30)),
11638 1 : space: Lsn(0x30),
11639 1 : },
11640 1 : leases: Default::default(),
11641 1 : within_ancestor_pitr: false,
11642 1 : };
11643 : }
11644 :
11645 1 : let cancel = CancellationToken::new();
11646 :
11647 : // Do a partial compaction on key range 0..2
11648 1 : tline
11649 1 : .compact_with_gc(
11650 1 : &cancel,
11651 1 : CompactOptions {
11652 1 : flags: EnumSet::new(),
11653 1 : compact_key_range: Some((get_key(0)..get_key(2)).into()),
11654 1 : ..CompactOptions::default_for_gc_compaction_unit_tests()
11655 1 : },
11656 1 : &ctx,
11657 1 : )
11658 1 : .await
11659 1 : .unwrap();
11660 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11661 1 : check_layer_map_key_eq(
11662 1 : all_layers,
11663 1 : vec![
11664 : // newly-generated image layer for the partial compaction range 0-2
11665 1 : PersistentLayerKey {
11666 1 : key_range: get_key(0)..get_key(2),
11667 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11668 1 : is_delta: false,
11669 1 : },
11670 1 : PersistentLayerKey {
11671 1 : key_range: get_key(0)..get_key(10),
11672 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
11673 1 : is_delta: false,
11674 1 : },
11675 : // delta1 is split and the second part is rewritten
11676 1 : PersistentLayerKey {
11677 1 : key_range: get_key(2)..get_key(4),
11678 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11679 1 : is_delta: true,
11680 1 : },
11681 1 : PersistentLayerKey {
11682 1 : key_range: get_key(5)..get_key(7),
11683 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11684 1 : is_delta: true,
11685 1 : },
11686 1 : PersistentLayerKey {
11687 1 : key_range: get_key(8)..get_key(10),
11688 1 : lsn_range: Lsn(0x48)..Lsn(0x50),
11689 1 : is_delta: true,
11690 1 : },
11691 : ],
11692 : );
11693 :
11694 : // Do a partial compaction on key range 2..4
11695 1 : tline
11696 1 : .compact_with_gc(
11697 1 : &cancel,
11698 1 : CompactOptions {
11699 1 : flags: EnumSet::new(),
11700 1 : compact_key_range: Some((get_key(2)..get_key(4)).into()),
11701 1 : ..CompactOptions::default_for_gc_compaction_unit_tests()
11702 1 : },
11703 1 : &ctx,
11704 1 : )
11705 1 : .await
11706 1 : .unwrap();
11707 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11708 1 : check_layer_map_key_eq(
11709 1 : all_layers,
11710 1 : vec![
11711 1 : PersistentLayerKey {
11712 1 : key_range: get_key(0)..get_key(2),
11713 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11714 1 : is_delta: false,
11715 1 : },
11716 1 : PersistentLayerKey {
11717 1 : key_range: get_key(0)..get_key(10),
11718 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
11719 1 : is_delta: false,
11720 1 : },
11721 : // image layer generated for the compaction range 2-4
11722 1 : PersistentLayerKey {
11723 1 : key_range: get_key(2)..get_key(4),
11724 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11725 1 : is_delta: false,
11726 1 : },
11727 : // we have key2/key3 above the retain_lsn, so we still need this delta layer
11728 1 : PersistentLayerKey {
11729 1 : key_range: get_key(2)..get_key(4),
11730 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11731 1 : is_delta: true,
11732 1 : },
11733 1 : PersistentLayerKey {
11734 1 : key_range: get_key(5)..get_key(7),
11735 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11736 1 : is_delta: true,
11737 1 : },
11738 1 : PersistentLayerKey {
11739 1 : key_range: get_key(8)..get_key(10),
11740 1 : lsn_range: Lsn(0x48)..Lsn(0x50),
11741 1 : is_delta: true,
11742 1 : },
11743 : ],
11744 : );
11745 :
11746 : // Do a partial compaction on key range 4..9
11747 1 : tline
11748 1 : .compact_with_gc(
11749 1 : &cancel,
11750 1 : CompactOptions {
11751 1 : flags: EnumSet::new(),
11752 1 : compact_key_range: Some((get_key(4)..get_key(9)).into()),
11753 1 : ..CompactOptions::default_for_gc_compaction_unit_tests()
11754 1 : },
11755 1 : &ctx,
11756 1 : )
11757 1 : .await
11758 1 : .unwrap();
11759 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11760 1 : check_layer_map_key_eq(
11761 1 : all_layers,
11762 1 : vec![
11763 1 : PersistentLayerKey {
11764 1 : key_range: get_key(0)..get_key(2),
11765 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11766 1 : is_delta: false,
11767 1 : },
11768 1 : PersistentLayerKey {
11769 1 : key_range: get_key(0)..get_key(10),
11770 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
11771 1 : is_delta: false,
11772 1 : },
11773 1 : PersistentLayerKey {
11774 1 : key_range: get_key(2)..get_key(4),
11775 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11776 1 : is_delta: false,
11777 1 : },
11778 1 : PersistentLayerKey {
11779 1 : key_range: get_key(2)..get_key(4),
11780 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11781 1 : is_delta: true,
11782 1 : },
11783 : // image layer generated for this compaction range
11784 1 : PersistentLayerKey {
11785 1 : key_range: get_key(4)..get_key(9),
11786 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11787 1 : is_delta: false,
11788 1 : },
11789 1 : PersistentLayerKey {
11790 1 : key_range: get_key(8)..get_key(10),
11791 1 : lsn_range: Lsn(0x48)..Lsn(0x50),
11792 1 : is_delta: true,
11793 1 : },
11794 : ],
11795 : );
11796 :
11797 : // Do a partial compaction on key range 9..10
11798 1 : tline
11799 1 : .compact_with_gc(
11800 1 : &cancel,
11801 1 : CompactOptions {
11802 1 : flags: EnumSet::new(),
11803 1 : compact_key_range: Some((get_key(9)..get_key(10)).into()),
11804 1 : ..CompactOptions::default_for_gc_compaction_unit_tests()
11805 1 : },
11806 1 : &ctx,
11807 1 : )
11808 1 : .await
11809 1 : .unwrap();
11810 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11811 1 : check_layer_map_key_eq(
11812 1 : all_layers,
11813 1 : vec![
11814 1 : PersistentLayerKey {
11815 1 : key_range: get_key(0)..get_key(2),
11816 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11817 1 : is_delta: false,
11818 1 : },
11819 1 : PersistentLayerKey {
11820 1 : key_range: get_key(0)..get_key(10),
11821 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
11822 1 : is_delta: false,
11823 1 : },
11824 1 : PersistentLayerKey {
11825 1 : key_range: get_key(2)..get_key(4),
11826 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11827 1 : is_delta: false,
11828 1 : },
11829 1 : PersistentLayerKey {
11830 1 : key_range: get_key(2)..get_key(4),
11831 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11832 1 : is_delta: true,
11833 1 : },
11834 1 : PersistentLayerKey {
11835 1 : key_range: get_key(4)..get_key(9),
11836 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11837 1 : is_delta: false,
11838 1 : },
11839 : // image layer generated for the compaction range
11840 1 : PersistentLayerKey {
11841 1 : key_range: get_key(9)..get_key(10),
11842 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11843 1 : is_delta: false,
11844 1 : },
11845 1 : PersistentLayerKey {
11846 1 : key_range: get_key(8)..get_key(10),
11847 1 : lsn_range: Lsn(0x48)..Lsn(0x50),
11848 1 : is_delta: true,
11849 1 : },
11850 : ],
11851 : );
11852 :
11853 : // Do a partial compaction on key range 0..10, all image layers below LSN 20 can be replaced with new ones.
11854 1 : tline
11855 1 : .compact_with_gc(
11856 1 : &cancel,
11857 1 : CompactOptions {
11858 1 : flags: EnumSet::new(),
11859 1 : compact_key_range: Some((get_key(0)..get_key(10)).into()),
11860 1 : ..CompactOptions::default_for_gc_compaction_unit_tests()
11861 1 : },
11862 1 : &ctx,
11863 1 : )
11864 1 : .await
11865 1 : .unwrap();
11866 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11867 1 : check_layer_map_key_eq(
11868 1 : all_layers,
11869 1 : vec![
11870 : // aha, we removed all unnecessary image/delta layers and got a very clean layer map!
11871 1 : PersistentLayerKey {
11872 1 : key_range: get_key(0)..get_key(10),
11873 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11874 1 : is_delta: false,
11875 1 : },
11876 1 : PersistentLayerKey {
11877 1 : key_range: get_key(2)..get_key(4),
11878 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11879 1 : is_delta: true,
11880 1 : },
11881 1 : PersistentLayerKey {
11882 1 : key_range: get_key(8)..get_key(10),
11883 1 : lsn_range: Lsn(0x48)..Lsn(0x50),
11884 1 : is_delta: true,
11885 1 : },
11886 : ],
11887 : );
11888 2 : Ok(())
11889 1 : }
11890 :
11891 : #[cfg(feature = "testing")]
11892 : #[tokio::test]
11893 1 : async fn test_timeline_offload_retain_lsn() -> anyhow::Result<()> {
11894 1 : let harness = TenantHarness::create("test_timeline_offload_retain_lsn")
11895 1 : .await
11896 1 : .unwrap();
11897 1 : let (tenant, ctx) = harness.load().await;
11898 1 : let tline_parent = tenant
11899 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
11900 1 : .await
11901 1 : .unwrap();
11902 1 : let tline_child = tenant
11903 1 : .branch_timeline_test(&tline_parent, NEW_TIMELINE_ID, Some(Lsn(0x20)), &ctx)
11904 1 : .await
11905 1 : .unwrap();
11906 : {
11907 1 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
11908 1 : assert_eq!(
11909 1 : gc_info_parent.retain_lsns,
11910 1 : vec![(Lsn(0x20), tline_child.timeline_id, MaybeOffloaded::No)]
11911 : );
11912 : }
11913 : // We have to directly call the remote_client instead of using the archive function to avoid constructing broker client...
11914 1 : tline_child
11915 1 : .remote_client
11916 1 : .schedule_index_upload_for_timeline_archival_state(TimelineArchivalState::Archived)
11917 1 : .unwrap();
11918 1 : tline_child.remote_client.wait_completion().await.unwrap();
11919 1 : offload_timeline(&tenant, &tline_child)
11920 1 : .instrument(tracing::info_span!(parent: None, "offload_test", tenant_id=%"test", shard_id=%"test", timeline_id=%"test"))
11921 1 : .await.unwrap();
11922 1 : let child_timeline_id = tline_child.timeline_id;
11923 1 : Arc::try_unwrap(tline_child).unwrap();
11924 :
11925 : {
11926 1 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
11927 1 : assert_eq!(
11928 1 : gc_info_parent.retain_lsns,
11929 1 : vec![(Lsn(0x20), child_timeline_id, MaybeOffloaded::Yes)]
11930 : );
11931 : }
11932 :
11933 1 : tenant
11934 1 : .get_offloaded_timeline(child_timeline_id)
11935 1 : .unwrap()
11936 1 : .defuse_for_tenant_drop();
11937 :
11938 2 : Ok(())
11939 1 : }
11940 :
11941 : #[cfg(feature = "testing")]
11942 : #[tokio::test]
11943 1 : async fn test_simple_bottom_most_compaction_above_lsn() -> anyhow::Result<()> {
11944 1 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_above_lsn").await?;
11945 1 : let (tenant, ctx) = harness.load().await;
11946 :
11947 148 : fn get_key(id: u32) -> Key {
11948 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
11949 148 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
11950 148 : key.field6 = id;
11951 148 : key
11952 148 : }
11953 :
11954 1 : let img_layer = (0..10)
11955 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
11956 1 : .collect_vec();
11957 :
11958 1 : let delta1 = vec![(
11959 1 : get_key(1),
11960 1 : Lsn(0x20),
11961 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
11962 1 : )];
11963 1 : let delta4 = vec![(
11964 1 : get_key(1),
11965 1 : Lsn(0x28),
11966 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
11967 1 : )];
11968 1 : let delta2 = vec![
11969 1 : (
11970 1 : get_key(1),
11971 1 : Lsn(0x30),
11972 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
11973 1 : ),
11974 1 : (
11975 1 : get_key(1),
11976 1 : Lsn(0x38),
11977 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
11978 1 : ),
11979 : ];
11980 1 : let delta3 = vec![
11981 1 : (
11982 1 : get_key(8),
11983 1 : Lsn(0x48),
11984 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
11985 1 : ),
11986 1 : (
11987 1 : get_key(9),
11988 1 : Lsn(0x48),
11989 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
11990 1 : ),
11991 : ];
11992 :
11993 1 : let tline = tenant
11994 1 : .create_test_timeline_with_layers(
11995 1 : TIMELINE_ID,
11996 1 : Lsn(0x10),
11997 1 : DEFAULT_PG_VERSION,
11998 1 : &ctx,
11999 1 : vec![], // in-memory layers
12000 1 : vec![
12001 1 : // delta1/2/4 only contain a single key but multiple updates
12002 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x28), delta1),
12003 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
12004 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x28)..Lsn(0x30), delta4),
12005 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta3),
12006 1 : ], // delta layers
12007 1 : vec![(Lsn(0x10), img_layer)], // image layers
12008 1 : Lsn(0x50),
12009 1 : )
12010 1 : .await?;
12011 : {
12012 1 : tline
12013 1 : .applied_gc_cutoff_lsn
12014 1 : .lock_for_write()
12015 1 : .store_and_unlock(Lsn(0x30))
12016 1 : .wait()
12017 1 : .await;
12018 : // Update GC info
12019 1 : let mut guard = tline.gc_info.write().unwrap();
12020 1 : *guard = GcInfo {
12021 1 : retain_lsns: vec![
12022 1 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
12023 1 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
12024 1 : ],
12025 1 : cutoffs: GcCutoffs {
12026 1 : time: Some(Lsn(0x30)),
12027 1 : space: Lsn(0x30),
12028 1 : },
12029 1 : leases: Default::default(),
12030 1 : within_ancestor_pitr: false,
12031 1 : };
12032 : }
12033 :
12034 1 : let expected_result = [
12035 1 : Bytes::from_static(b"value 0@0x10"),
12036 1 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
12037 1 : Bytes::from_static(b"value 2@0x10"),
12038 1 : Bytes::from_static(b"value 3@0x10"),
12039 1 : Bytes::from_static(b"value 4@0x10"),
12040 1 : Bytes::from_static(b"value 5@0x10"),
12041 1 : Bytes::from_static(b"value 6@0x10"),
12042 1 : Bytes::from_static(b"value 7@0x10"),
12043 1 : Bytes::from_static(b"value 8@0x10@0x48"),
12044 1 : Bytes::from_static(b"value 9@0x10@0x48"),
12045 1 : ];
12046 :
12047 1 : let expected_result_at_gc_horizon = [
12048 1 : Bytes::from_static(b"value 0@0x10"),
12049 1 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
12050 1 : Bytes::from_static(b"value 2@0x10"),
12051 1 : Bytes::from_static(b"value 3@0x10"),
12052 1 : Bytes::from_static(b"value 4@0x10"),
12053 1 : Bytes::from_static(b"value 5@0x10"),
12054 1 : Bytes::from_static(b"value 6@0x10"),
12055 1 : Bytes::from_static(b"value 7@0x10"),
12056 1 : Bytes::from_static(b"value 8@0x10"),
12057 1 : Bytes::from_static(b"value 9@0x10"),
12058 1 : ];
12059 :
12060 1 : let expected_result_at_lsn_20 = [
12061 1 : Bytes::from_static(b"value 0@0x10"),
12062 1 : Bytes::from_static(b"value 1@0x10@0x20"),
12063 1 : Bytes::from_static(b"value 2@0x10"),
12064 1 : Bytes::from_static(b"value 3@0x10"),
12065 1 : Bytes::from_static(b"value 4@0x10"),
12066 1 : Bytes::from_static(b"value 5@0x10"),
12067 1 : Bytes::from_static(b"value 6@0x10"),
12068 1 : Bytes::from_static(b"value 7@0x10"),
12069 1 : Bytes::from_static(b"value 8@0x10"),
12070 1 : Bytes::from_static(b"value 9@0x10"),
12071 1 : ];
12072 :
12073 1 : let expected_result_at_lsn_10 = [
12074 1 : Bytes::from_static(b"value 0@0x10"),
12075 1 : Bytes::from_static(b"value 1@0x10"),
12076 1 : Bytes::from_static(b"value 2@0x10"),
12077 1 : Bytes::from_static(b"value 3@0x10"),
12078 1 : Bytes::from_static(b"value 4@0x10"),
12079 1 : Bytes::from_static(b"value 5@0x10"),
12080 1 : Bytes::from_static(b"value 6@0x10"),
12081 1 : Bytes::from_static(b"value 7@0x10"),
12082 1 : Bytes::from_static(b"value 8@0x10"),
12083 1 : Bytes::from_static(b"value 9@0x10"),
12084 1 : ];
12085 :
12086 3 : let verify_result = || async {
12087 3 : let gc_horizon = {
12088 3 : let gc_info = tline.gc_info.read().unwrap();
12089 3 : gc_info.cutoffs.time.unwrap_or_default()
12090 : };
12091 33 : for idx in 0..10 {
12092 30 : assert_eq!(
12093 30 : tline
12094 30 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
12095 30 : .await
12096 30 : .unwrap(),
12097 30 : &expected_result[idx]
12098 : );
12099 30 : assert_eq!(
12100 30 : tline
12101 30 : .get(get_key(idx as u32), gc_horizon, &ctx)
12102 30 : .await
12103 30 : .unwrap(),
12104 30 : &expected_result_at_gc_horizon[idx]
12105 : );
12106 30 : assert_eq!(
12107 30 : tline
12108 30 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
12109 30 : .await
12110 30 : .unwrap(),
12111 30 : &expected_result_at_lsn_20[idx]
12112 : );
12113 30 : assert_eq!(
12114 30 : tline
12115 30 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
12116 30 : .await
12117 30 : .unwrap(),
12118 30 : &expected_result_at_lsn_10[idx]
12119 : );
12120 : }
12121 6 : };
12122 :
12123 1 : verify_result().await;
12124 :
12125 1 : let cancel = CancellationToken::new();
12126 1 : tline
12127 1 : .compact_with_gc(
12128 1 : &cancel,
12129 1 : CompactOptions {
12130 1 : compact_lsn_range: Some(CompactLsnRange::above(Lsn(0x28))),
12131 1 : ..CompactOptions::default_for_gc_compaction_unit_tests()
12132 1 : },
12133 1 : &ctx,
12134 1 : )
12135 1 : .await
12136 1 : .unwrap();
12137 1 : verify_result().await;
12138 :
12139 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12140 1 : check_layer_map_key_eq(
12141 1 : all_layers,
12142 1 : vec![
12143 : // The original image layer, not compacted
12144 1 : PersistentLayerKey {
12145 1 : key_range: get_key(0)..get_key(10),
12146 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
12147 1 : is_delta: false,
12148 1 : },
12149 : // Delta layer below the specified above_lsn not compacted
12150 1 : PersistentLayerKey {
12151 1 : key_range: get_key(1)..get_key(2),
12152 1 : lsn_range: Lsn(0x20)..Lsn(0x28),
12153 1 : is_delta: true,
12154 1 : },
12155 : // Delta layer compacted above the LSN
12156 1 : PersistentLayerKey {
12157 1 : key_range: get_key(1)..get_key(10),
12158 1 : lsn_range: Lsn(0x28)..Lsn(0x50),
12159 1 : is_delta: true,
12160 1 : },
12161 : ],
12162 : );
12163 :
12164 : // compact again
12165 1 : tline
12166 1 : .compact_with_gc(
12167 1 : &cancel,
12168 1 : CompactOptions::default_for_gc_compaction_unit_tests(),
12169 1 : &ctx,
12170 1 : )
12171 1 : .await
12172 1 : .unwrap();
12173 1 : verify_result().await;
12174 :
12175 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12176 1 : check_layer_map_key_eq(
12177 1 : all_layers,
12178 1 : vec![
12179 : // The compacted image layer (full key range)
12180 1 : PersistentLayerKey {
12181 1 : key_range: Key::MIN..Key::MAX,
12182 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
12183 1 : is_delta: false,
12184 1 : },
12185 : // All other data in the delta layer
12186 1 : PersistentLayerKey {
12187 1 : key_range: get_key(1)..get_key(10),
12188 1 : lsn_range: Lsn(0x10)..Lsn(0x50),
12189 1 : is_delta: true,
12190 1 : },
12191 : ],
12192 : );
12193 :
12194 2 : Ok(())
12195 1 : }
12196 :
12197 : #[cfg(feature = "testing")]
12198 : #[tokio::test]
12199 1 : async fn test_simple_bottom_most_compaction_rectangle() -> anyhow::Result<()> {
12200 1 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_rectangle").await?;
12201 1 : let (tenant, ctx) = harness.load().await;
12202 :
12203 254 : fn get_key(id: u32) -> Key {
12204 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
12205 254 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
12206 254 : key.field6 = id;
12207 254 : key
12208 254 : }
12209 :
12210 1 : let img_layer = (0..10)
12211 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
12212 1 : .collect_vec();
12213 :
12214 1 : let delta1 = vec![(
12215 1 : get_key(1),
12216 1 : Lsn(0x20),
12217 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
12218 1 : )];
12219 1 : let delta4 = vec![(
12220 1 : get_key(1),
12221 1 : Lsn(0x28),
12222 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
12223 1 : )];
12224 1 : let delta2 = vec![
12225 1 : (
12226 1 : get_key(1),
12227 1 : Lsn(0x30),
12228 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
12229 1 : ),
12230 1 : (
12231 1 : get_key(1),
12232 1 : Lsn(0x38),
12233 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
12234 1 : ),
12235 : ];
12236 1 : let delta3 = vec![
12237 1 : (
12238 1 : get_key(8),
12239 1 : Lsn(0x48),
12240 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
12241 1 : ),
12242 1 : (
12243 1 : get_key(9),
12244 1 : Lsn(0x48),
12245 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
12246 1 : ),
12247 : ];
12248 :
12249 1 : let tline = tenant
12250 1 : .create_test_timeline_with_layers(
12251 1 : TIMELINE_ID,
12252 1 : Lsn(0x10),
12253 1 : DEFAULT_PG_VERSION,
12254 1 : &ctx,
12255 1 : vec![], // in-memory layers
12256 1 : vec![
12257 1 : // delta1/2/4 only contain a single key but multiple updates
12258 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x28), delta1),
12259 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
12260 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x28)..Lsn(0x30), delta4),
12261 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta3),
12262 1 : ], // delta layers
12263 1 : vec![(Lsn(0x10), img_layer)], // image layers
12264 1 : Lsn(0x50),
12265 1 : )
12266 1 : .await?;
12267 : {
12268 1 : tline
12269 1 : .applied_gc_cutoff_lsn
12270 1 : .lock_for_write()
12271 1 : .store_and_unlock(Lsn(0x30))
12272 1 : .wait()
12273 1 : .await;
12274 : // Update GC info
12275 1 : let mut guard = tline.gc_info.write().unwrap();
12276 1 : *guard = GcInfo {
12277 1 : retain_lsns: vec![
12278 1 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
12279 1 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
12280 1 : ],
12281 1 : cutoffs: GcCutoffs {
12282 1 : time: Some(Lsn(0x30)),
12283 1 : space: Lsn(0x30),
12284 1 : },
12285 1 : leases: Default::default(),
12286 1 : within_ancestor_pitr: false,
12287 1 : };
12288 : }
12289 :
12290 1 : let expected_result = [
12291 1 : Bytes::from_static(b"value 0@0x10"),
12292 1 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
12293 1 : Bytes::from_static(b"value 2@0x10"),
12294 1 : Bytes::from_static(b"value 3@0x10"),
12295 1 : Bytes::from_static(b"value 4@0x10"),
12296 1 : Bytes::from_static(b"value 5@0x10"),
12297 1 : Bytes::from_static(b"value 6@0x10"),
12298 1 : Bytes::from_static(b"value 7@0x10"),
12299 1 : Bytes::from_static(b"value 8@0x10@0x48"),
12300 1 : Bytes::from_static(b"value 9@0x10@0x48"),
12301 1 : ];
12302 :
12303 1 : let expected_result_at_gc_horizon = [
12304 1 : Bytes::from_static(b"value 0@0x10"),
12305 1 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
12306 1 : Bytes::from_static(b"value 2@0x10"),
12307 1 : Bytes::from_static(b"value 3@0x10"),
12308 1 : Bytes::from_static(b"value 4@0x10"),
12309 1 : Bytes::from_static(b"value 5@0x10"),
12310 1 : Bytes::from_static(b"value 6@0x10"),
12311 1 : Bytes::from_static(b"value 7@0x10"),
12312 1 : Bytes::from_static(b"value 8@0x10"),
12313 1 : Bytes::from_static(b"value 9@0x10"),
12314 1 : ];
12315 :
12316 1 : let expected_result_at_lsn_20 = [
12317 1 : Bytes::from_static(b"value 0@0x10"),
12318 1 : Bytes::from_static(b"value 1@0x10@0x20"),
12319 1 : Bytes::from_static(b"value 2@0x10"),
12320 1 : Bytes::from_static(b"value 3@0x10"),
12321 1 : Bytes::from_static(b"value 4@0x10"),
12322 1 : Bytes::from_static(b"value 5@0x10"),
12323 1 : Bytes::from_static(b"value 6@0x10"),
12324 1 : Bytes::from_static(b"value 7@0x10"),
12325 1 : Bytes::from_static(b"value 8@0x10"),
12326 1 : Bytes::from_static(b"value 9@0x10"),
12327 1 : ];
12328 :
12329 1 : let expected_result_at_lsn_10 = [
12330 1 : Bytes::from_static(b"value 0@0x10"),
12331 1 : Bytes::from_static(b"value 1@0x10"),
12332 1 : Bytes::from_static(b"value 2@0x10"),
12333 1 : Bytes::from_static(b"value 3@0x10"),
12334 1 : Bytes::from_static(b"value 4@0x10"),
12335 1 : Bytes::from_static(b"value 5@0x10"),
12336 1 : Bytes::from_static(b"value 6@0x10"),
12337 1 : Bytes::from_static(b"value 7@0x10"),
12338 1 : Bytes::from_static(b"value 8@0x10"),
12339 1 : Bytes::from_static(b"value 9@0x10"),
12340 1 : ];
12341 :
12342 5 : let verify_result = || async {
12343 5 : let gc_horizon = {
12344 5 : let gc_info = tline.gc_info.read().unwrap();
12345 5 : gc_info.cutoffs.time.unwrap_or_default()
12346 : };
12347 55 : for idx in 0..10 {
12348 50 : assert_eq!(
12349 50 : tline
12350 50 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
12351 50 : .await
12352 50 : .unwrap(),
12353 50 : &expected_result[idx]
12354 : );
12355 50 : assert_eq!(
12356 50 : tline
12357 50 : .get(get_key(idx as u32), gc_horizon, &ctx)
12358 50 : .await
12359 50 : .unwrap(),
12360 50 : &expected_result_at_gc_horizon[idx]
12361 : );
12362 50 : assert_eq!(
12363 50 : tline
12364 50 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
12365 50 : .await
12366 50 : .unwrap(),
12367 50 : &expected_result_at_lsn_20[idx]
12368 : );
12369 50 : assert_eq!(
12370 50 : tline
12371 50 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
12372 50 : .await
12373 50 : .unwrap(),
12374 50 : &expected_result_at_lsn_10[idx]
12375 : );
12376 : }
12377 10 : };
12378 :
12379 1 : verify_result().await;
12380 :
12381 1 : let cancel = CancellationToken::new();
12382 :
12383 1 : tline
12384 1 : .compact_with_gc(
12385 1 : &cancel,
12386 1 : CompactOptions {
12387 1 : compact_key_range: Some((get_key(0)..get_key(2)).into()),
12388 1 : compact_lsn_range: Some((Lsn(0x20)..Lsn(0x28)).into()),
12389 1 : ..CompactOptions::default_for_gc_compaction_unit_tests()
12390 1 : },
12391 1 : &ctx,
12392 1 : )
12393 1 : .await
12394 1 : .unwrap();
12395 1 : verify_result().await;
12396 :
12397 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12398 1 : check_layer_map_key_eq(
12399 1 : all_layers,
12400 1 : vec![
12401 : // The original image layer, not compacted
12402 1 : PersistentLayerKey {
12403 1 : key_range: get_key(0)..get_key(10),
12404 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
12405 1 : is_delta: false,
12406 1 : },
12407 : // According the selection logic, we select all layers with start key <= 0x28, so we would merge the layer 0x20-0x28 and
12408 : // the layer 0x28-0x30 into one.
12409 1 : PersistentLayerKey {
12410 1 : key_range: get_key(1)..get_key(2),
12411 1 : lsn_range: Lsn(0x20)..Lsn(0x30),
12412 1 : is_delta: true,
12413 1 : },
12414 : // Above the upper bound and untouched
12415 1 : PersistentLayerKey {
12416 1 : key_range: get_key(1)..get_key(2),
12417 1 : lsn_range: Lsn(0x30)..Lsn(0x50),
12418 1 : is_delta: true,
12419 1 : },
12420 : // This layer is untouched
12421 1 : PersistentLayerKey {
12422 1 : key_range: get_key(8)..get_key(10),
12423 1 : lsn_range: Lsn(0x30)..Lsn(0x50),
12424 1 : is_delta: true,
12425 1 : },
12426 : ],
12427 : );
12428 :
12429 1 : tline
12430 1 : .compact_with_gc(
12431 1 : &cancel,
12432 1 : CompactOptions {
12433 1 : compact_key_range: Some((get_key(3)..get_key(8)).into()),
12434 1 : compact_lsn_range: Some((Lsn(0x28)..Lsn(0x40)).into()),
12435 1 : ..CompactOptions::default_for_gc_compaction_unit_tests()
12436 1 : },
12437 1 : &ctx,
12438 1 : )
12439 1 : .await
12440 1 : .unwrap();
12441 1 : verify_result().await;
12442 :
12443 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12444 1 : check_layer_map_key_eq(
12445 1 : all_layers,
12446 1 : vec![
12447 : // The original image layer, not compacted
12448 1 : PersistentLayerKey {
12449 1 : key_range: get_key(0)..get_key(10),
12450 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
12451 1 : is_delta: false,
12452 1 : },
12453 : // Not in the compaction key range, uncompacted
12454 1 : PersistentLayerKey {
12455 1 : key_range: get_key(1)..get_key(2),
12456 1 : lsn_range: Lsn(0x20)..Lsn(0x30),
12457 1 : is_delta: true,
12458 1 : },
12459 : // Not in the compaction key range, uncompacted but need rewrite because the delta layer overlaps with the range
12460 1 : PersistentLayerKey {
12461 1 : key_range: get_key(1)..get_key(2),
12462 1 : lsn_range: Lsn(0x30)..Lsn(0x50),
12463 1 : is_delta: true,
12464 1 : },
12465 : // Note that when we specify the LSN upper bound to be 0x40, the compaction algorithm will not try to cut the layer
12466 : // horizontally in half. Instead, it will include all LSNs that overlap with 0x40. So the real max_lsn of the compaction
12467 : // becomes 0x50.
12468 1 : PersistentLayerKey {
12469 1 : key_range: get_key(8)..get_key(10),
12470 1 : lsn_range: Lsn(0x30)..Lsn(0x50),
12471 1 : is_delta: true,
12472 1 : },
12473 : ],
12474 : );
12475 :
12476 : // compact again
12477 1 : tline
12478 1 : .compact_with_gc(
12479 1 : &cancel,
12480 1 : CompactOptions {
12481 1 : compact_key_range: Some((get_key(0)..get_key(5)).into()),
12482 1 : compact_lsn_range: Some((Lsn(0x20)..Lsn(0x50)).into()),
12483 1 : ..CompactOptions::default_for_gc_compaction_unit_tests()
12484 1 : },
12485 1 : &ctx,
12486 1 : )
12487 1 : .await
12488 1 : .unwrap();
12489 1 : verify_result().await;
12490 :
12491 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12492 1 : check_layer_map_key_eq(
12493 1 : all_layers,
12494 1 : vec![
12495 : // The original image layer, not compacted
12496 1 : PersistentLayerKey {
12497 1 : key_range: get_key(0)..get_key(10),
12498 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
12499 1 : is_delta: false,
12500 1 : },
12501 : // The range gets compacted
12502 1 : PersistentLayerKey {
12503 1 : key_range: get_key(1)..get_key(2),
12504 1 : lsn_range: Lsn(0x20)..Lsn(0x50),
12505 1 : is_delta: true,
12506 1 : },
12507 : // Not touched during this iteration of compaction
12508 1 : PersistentLayerKey {
12509 1 : key_range: get_key(8)..get_key(10),
12510 1 : lsn_range: Lsn(0x30)..Lsn(0x50),
12511 1 : is_delta: true,
12512 1 : },
12513 : ],
12514 : );
12515 :
12516 : // final full compaction
12517 1 : tline
12518 1 : .compact_with_gc(
12519 1 : &cancel,
12520 1 : CompactOptions::default_for_gc_compaction_unit_tests(),
12521 1 : &ctx,
12522 1 : )
12523 1 : .await
12524 1 : .unwrap();
12525 1 : verify_result().await;
12526 :
12527 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12528 1 : check_layer_map_key_eq(
12529 1 : all_layers,
12530 1 : vec![
12531 : // The compacted image layer (full key range)
12532 1 : PersistentLayerKey {
12533 1 : key_range: Key::MIN..Key::MAX,
12534 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
12535 1 : is_delta: false,
12536 1 : },
12537 : // All other data in the delta layer
12538 1 : PersistentLayerKey {
12539 1 : key_range: get_key(1)..get_key(10),
12540 1 : lsn_range: Lsn(0x10)..Lsn(0x50),
12541 1 : is_delta: true,
12542 1 : },
12543 : ],
12544 : );
12545 :
12546 2 : Ok(())
12547 1 : }
12548 :
12549 : #[cfg(feature = "testing")]
12550 : #[tokio::test]
12551 1 : async fn test_bottom_most_compation_redo_failure() -> anyhow::Result<()> {
12552 1 : let harness = TenantHarness::create("test_bottom_most_compation_redo_failure").await?;
12553 1 : let (tenant, ctx) = harness.load().await;
12554 :
12555 13 : fn get_key(id: u32) -> Key {
12556 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
12557 13 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
12558 13 : key.field6 = id;
12559 13 : key
12560 13 : }
12561 :
12562 1 : let img_layer = (0..10)
12563 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
12564 1 : .collect_vec();
12565 :
12566 1 : let delta1 = vec![
12567 1 : (
12568 1 : get_key(1),
12569 1 : Lsn(0x20),
12570 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
12571 1 : ),
12572 1 : (
12573 1 : get_key(1),
12574 1 : Lsn(0x24),
12575 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x24")),
12576 1 : ),
12577 1 : (
12578 1 : get_key(1),
12579 1 : Lsn(0x28),
12580 1 : // This record will fail to redo
12581 1 : Value::WalRecord(NeonWalRecord::wal_append_conditional("@0x28", "???")),
12582 1 : ),
12583 : ];
12584 :
12585 1 : let tline = tenant
12586 1 : .create_test_timeline_with_layers(
12587 1 : TIMELINE_ID,
12588 1 : Lsn(0x10),
12589 1 : DEFAULT_PG_VERSION,
12590 1 : &ctx,
12591 1 : vec![], // in-memory layers
12592 1 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
12593 1 : Lsn(0x20)..Lsn(0x30),
12594 1 : delta1,
12595 1 : )], // delta layers
12596 1 : vec![(Lsn(0x10), img_layer)], // image layers
12597 1 : Lsn(0x50),
12598 1 : )
12599 1 : .await?;
12600 : {
12601 1 : tline
12602 1 : .applied_gc_cutoff_lsn
12603 1 : .lock_for_write()
12604 1 : .store_and_unlock(Lsn(0x30))
12605 1 : .wait()
12606 1 : .await;
12607 : // Update GC info
12608 1 : let mut guard = tline.gc_info.write().unwrap();
12609 1 : *guard = GcInfo {
12610 1 : retain_lsns: vec![],
12611 1 : cutoffs: GcCutoffs {
12612 1 : time: Some(Lsn(0x30)),
12613 1 : space: Lsn(0x30),
12614 1 : },
12615 1 : leases: Default::default(),
12616 1 : within_ancestor_pitr: false,
12617 1 : };
12618 : }
12619 :
12620 1 : let cancel = CancellationToken::new();
12621 :
12622 : // Compaction will fail, but should not fire any critical error.
12623 : // Gc-compaction currently cannot figure out what keys are not in the keyspace during the compaction
12624 : // process. It will always try to redo the logs it reads and if it doesn't work, fail the entire
12625 : // compaction job. Tracked in <https://github.com/neondatabase/neon/issues/10395>.
12626 1 : let res = tline
12627 1 : .compact_with_gc(
12628 1 : &cancel,
12629 1 : CompactOptions {
12630 1 : compact_key_range: None,
12631 1 : compact_lsn_range: None,
12632 1 : ..CompactOptions::default_for_gc_compaction_unit_tests()
12633 1 : },
12634 1 : &ctx,
12635 1 : )
12636 1 : .await;
12637 1 : assert!(res.is_err());
12638 :
12639 2 : Ok(())
12640 1 : }
12641 :
12642 : #[cfg(feature = "testing")]
12643 : #[tokio::test]
12644 1 : async fn test_synthetic_size_calculation_with_invisible_branches() -> anyhow::Result<()> {
12645 : use pageserver_api::models::TimelineVisibilityState;
12646 :
12647 : use crate::tenant::size::gather_inputs;
12648 :
12649 1 : let tenant_conf = pageserver_api::models::TenantConfig {
12650 1 : // Ensure that we don't compute gc_cutoffs (which needs reading the layer files)
12651 1 : pitr_interval: Some(Duration::ZERO),
12652 1 : ..Default::default()
12653 1 : };
12654 1 : let harness = TenantHarness::create_custom(
12655 1 : "test_synthetic_size_calculation_with_invisible_branches",
12656 1 : tenant_conf,
12657 1 : TenantId::generate(),
12658 1 : ShardIdentity::unsharded(),
12659 1 : Generation::new(0xdeadbeef),
12660 1 : )
12661 1 : .await?;
12662 1 : let (tenant, ctx) = harness.load().await;
12663 1 : let main_tline = tenant
12664 1 : .create_test_timeline_with_layers(
12665 1 : TIMELINE_ID,
12666 1 : Lsn(0x10),
12667 1 : DEFAULT_PG_VERSION,
12668 1 : &ctx,
12669 1 : vec![],
12670 1 : vec![],
12671 1 : vec![],
12672 1 : Lsn(0x100),
12673 1 : )
12674 1 : .await?;
12675 :
12676 1 : let snapshot1 = TimelineId::from_array(hex!("11223344556677881122334455667790"));
12677 1 : tenant
12678 1 : .branch_timeline_test_with_layers(
12679 1 : &main_tline,
12680 1 : snapshot1,
12681 1 : Some(Lsn(0x20)),
12682 1 : &ctx,
12683 1 : vec![],
12684 1 : vec![],
12685 1 : Lsn(0x50),
12686 1 : )
12687 1 : .await?;
12688 1 : let snapshot2 = TimelineId::from_array(hex!("11223344556677881122334455667791"));
12689 1 : tenant
12690 1 : .branch_timeline_test_with_layers(
12691 1 : &main_tline,
12692 1 : snapshot2,
12693 1 : Some(Lsn(0x30)),
12694 1 : &ctx,
12695 1 : vec![],
12696 1 : vec![],
12697 1 : Lsn(0x50),
12698 1 : )
12699 1 : .await?;
12700 1 : let snapshot3 = TimelineId::from_array(hex!("11223344556677881122334455667792"));
12701 1 : tenant
12702 1 : .branch_timeline_test_with_layers(
12703 1 : &main_tline,
12704 1 : snapshot3,
12705 1 : Some(Lsn(0x40)),
12706 1 : &ctx,
12707 1 : vec![],
12708 1 : vec![],
12709 1 : Lsn(0x50),
12710 1 : )
12711 1 : .await?;
12712 1 : let limit = Arc::new(Semaphore::new(1));
12713 1 : let max_retention_period = None;
12714 1 : let mut logical_size_cache = HashMap::new();
12715 1 : let cause = LogicalSizeCalculationCause::EvictionTaskImitation;
12716 1 : let cancel = CancellationToken::new();
12717 :
12718 1 : let inputs = gather_inputs(
12719 1 : &tenant,
12720 1 : &limit,
12721 1 : max_retention_period,
12722 1 : &mut logical_size_cache,
12723 1 : cause,
12724 1 : &cancel,
12725 1 : &ctx,
12726 : )
12727 1 : .instrument(info_span!(
12728 : "gather_inputs",
12729 : tenant_id = "unknown",
12730 : shard_id = "unknown",
12731 : ))
12732 1 : .await?;
12733 : use crate::tenant::size::{LsnKind, ModelInputs, SegmentMeta};
12734 : use LsnKind::*;
12735 : use tenant_size_model::Segment;
12736 1 : let ModelInputs { mut segments, .. } = inputs;
12737 15 : segments.retain(|s| s.timeline_id == TIMELINE_ID);
12738 6 : for segment in segments.iter_mut() {
12739 6 : segment.segment.parent = None; // We don't care about the parent for the test
12740 6 : segment.segment.size = None; // We don't care about the size for the test
12741 6 : }
12742 1 : assert_eq!(
12743 : segments,
12744 : [
12745 : SegmentMeta {
12746 : segment: Segment {
12747 : parent: None,
12748 : lsn: 0x10,
12749 : size: None,
12750 : needed: false,
12751 : },
12752 : timeline_id: TIMELINE_ID,
12753 : kind: BranchStart,
12754 : },
12755 : SegmentMeta {
12756 : segment: Segment {
12757 : parent: None,
12758 : lsn: 0x20,
12759 : size: None,
12760 : needed: false,
12761 : },
12762 : timeline_id: TIMELINE_ID,
12763 : kind: BranchPoint,
12764 : },
12765 : SegmentMeta {
12766 : segment: Segment {
12767 : parent: None,
12768 : lsn: 0x30,
12769 : size: None,
12770 : needed: false,
12771 : },
12772 : timeline_id: TIMELINE_ID,
12773 : kind: BranchPoint,
12774 : },
12775 : SegmentMeta {
12776 : segment: Segment {
12777 : parent: None,
12778 : lsn: 0x40,
12779 : size: None,
12780 : needed: false,
12781 : },
12782 : timeline_id: TIMELINE_ID,
12783 : kind: BranchPoint,
12784 : },
12785 : SegmentMeta {
12786 : segment: Segment {
12787 : parent: None,
12788 : lsn: 0x100,
12789 : size: None,
12790 : needed: false,
12791 : },
12792 : timeline_id: TIMELINE_ID,
12793 : kind: GcCutOff,
12794 : }, // we need to retain everything above the last branch point
12795 : SegmentMeta {
12796 : segment: Segment {
12797 : parent: None,
12798 : lsn: 0x100,
12799 : size: None,
12800 : needed: true,
12801 : },
12802 : timeline_id: TIMELINE_ID,
12803 : kind: BranchEnd,
12804 : },
12805 : ]
12806 : );
12807 :
12808 1 : main_tline
12809 1 : .remote_client
12810 1 : .schedule_index_upload_for_timeline_invisible_state(
12811 1 : TimelineVisibilityState::Invisible,
12812 0 : )?;
12813 1 : main_tline.remote_client.wait_completion().await?;
12814 1 : let inputs = gather_inputs(
12815 1 : &tenant,
12816 1 : &limit,
12817 1 : max_retention_period,
12818 1 : &mut logical_size_cache,
12819 1 : cause,
12820 1 : &cancel,
12821 1 : &ctx,
12822 : )
12823 1 : .instrument(info_span!(
12824 : "gather_inputs",
12825 : tenant_id = "unknown",
12826 : shard_id = "unknown",
12827 : ))
12828 1 : .await?;
12829 1 : let ModelInputs { mut segments, .. } = inputs;
12830 14 : segments.retain(|s| s.timeline_id == TIMELINE_ID);
12831 5 : for segment in segments.iter_mut() {
12832 5 : segment.segment.parent = None; // We don't care about the parent for the test
12833 5 : segment.segment.size = None; // We don't care about the size for the test
12834 5 : }
12835 1 : assert_eq!(
12836 : segments,
12837 : [
12838 : SegmentMeta {
12839 : segment: Segment {
12840 : parent: None,
12841 : lsn: 0x10,
12842 : size: None,
12843 : needed: false,
12844 : },
12845 : timeline_id: TIMELINE_ID,
12846 : kind: BranchStart,
12847 : },
12848 : SegmentMeta {
12849 : segment: Segment {
12850 : parent: None,
12851 : lsn: 0x20,
12852 : size: None,
12853 : needed: false,
12854 : },
12855 : timeline_id: TIMELINE_ID,
12856 : kind: BranchPoint,
12857 : },
12858 : SegmentMeta {
12859 : segment: Segment {
12860 : parent: None,
12861 : lsn: 0x30,
12862 : size: None,
12863 : needed: false,
12864 : },
12865 : timeline_id: TIMELINE_ID,
12866 : kind: BranchPoint,
12867 : },
12868 : SegmentMeta {
12869 : segment: Segment {
12870 : parent: None,
12871 : lsn: 0x40,
12872 : size: None,
12873 : needed: false,
12874 : },
12875 : timeline_id: TIMELINE_ID,
12876 : kind: BranchPoint,
12877 : },
12878 : SegmentMeta {
12879 : segment: Segment {
12880 : parent: None,
12881 : lsn: 0x40, // Branch end LSN == last branch point LSN
12882 : size: None,
12883 : needed: true,
12884 : },
12885 : timeline_id: TIMELINE_ID,
12886 : kind: BranchEnd,
12887 : },
12888 : ]
12889 : );
12890 :
12891 2 : Ok(())
12892 1 : }
12893 :
12894 : #[tokio::test]
12895 1 : async fn test_get_force_image_creation_lsn() -> anyhow::Result<()> {
12896 1 : let tenant_conf = pageserver_api::models::TenantConfig {
12897 1 : pitr_interval: Some(Duration::from_secs(7 * 3600)),
12898 1 : image_layer_force_creation_period: Some(Duration::from_secs(3600)),
12899 1 : ..Default::default()
12900 1 : };
12901 :
12902 1 : let tenant_id = TenantId::generate();
12903 :
12904 1 : let harness = TenantHarness::create_custom(
12905 1 : "test_get_force_image_creation_lsn",
12906 1 : tenant_conf,
12907 1 : tenant_id,
12908 1 : ShardIdentity::unsharded(),
12909 1 : Generation::new(1),
12910 1 : )
12911 1 : .await?;
12912 1 : let (tenant, ctx) = harness.load().await;
12913 1 : let timeline = tenant
12914 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
12915 1 : .await?;
12916 1 : timeline.gc_info.write().unwrap().cutoffs.time = Some(Lsn(100));
12917 : {
12918 1 : let writer = timeline.writer().await;
12919 1 : writer.finish_write(Lsn(5000));
12920 : }
12921 :
12922 1 : let image_creation_lsn = timeline.get_force_image_creation_lsn().unwrap();
12923 1 : assert_eq!(image_creation_lsn, Lsn(4300));
12924 2 : Ok(())
12925 1 : }
12926 : }
|
