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 26774 : pub async fn request_redo(
501 26774 : &self,
502 26774 : key: pageserver_api::key::Key,
503 26774 : lsn: Lsn,
504 26774 : base_img: Option<(Lsn, bytes::Bytes)>,
505 26774 : records: Vec<(Lsn, wal_decoder::models::record::NeonWalRecord)>,
506 26774 : pg_version: PgMajorVersion,
507 26774 : redo_attempt_type: RedoAttemptType,
508 26774 : ) -> Result<bytes::Bytes, walredo::Error> {
509 26774 : 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 26774 : Self::Test(mgr) => {
516 26774 : mgr.request_redo(key, lsn, base_img, records, pg_version, redo_attempt_type)
517 26774 : .await
518 : }
519 : }
520 26774 : }
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 : ctx,
1209 3 : )?;
1210 3 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
1211 :
1212 3 : if !disk_consistent_lsn.is_valid() {
1213 : // As opposed to normal timelines which get initialised with a disk consitent LSN
1214 : // via initdb, imported timelines start from 0. If the import task stops before
1215 : // it advances disk consitent LSN, allow it to resume.
1216 0 : let in_progress_import = import_pgdata
1217 0 : .as_ref()
1218 0 : .map(|import| !import.is_done())
1219 0 : .unwrap_or(false);
1220 0 : if !in_progress_import {
1221 0 : anyhow::bail!("Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn");
1222 0 : }
1223 3 : }
1224 :
1225 3 : assert_eq!(
1226 : disk_consistent_lsn,
1227 3 : metadata.disk_consistent_lsn(),
1228 0 : "these are used interchangeably"
1229 : );
1230 :
1231 3 : timeline.remote_client.init_upload_queue(&index_part)?;
1232 :
1233 3 : timeline
1234 3 : .load_layer_map(disk_consistent_lsn, index_part)
1235 3 : .await
1236 3 : .with_context(|| {
1237 0 : format!("Failed to load layermap for timeline {tenant_id}/{timeline_id}")
1238 0 : })?;
1239 :
1240 : // When unarchiving, we've mostly likely lost the heatmap generated prior
1241 : // to the archival operation. To allow warming this timeline up, generate
1242 : // a previous heatmap which contains all visible layers in the layer map.
1243 : // This previous heatmap will be used whenever a fresh heatmap is generated
1244 : // for the timeline.
1245 3 : if self.conf.generate_unarchival_heatmap && matches!(cause, LoadTimelineCause::Unoffload) {
1246 0 : let mut tline_ending_at = Some((&timeline, timeline.get_last_record_lsn()));
1247 0 : while let Some((tline, end_lsn)) = tline_ending_at {
1248 0 : let unarchival_heatmap = tline.generate_unarchival_heatmap(end_lsn).await;
1249 : // Another unearchived timeline might have generated a heatmap for this ancestor.
1250 : // If the current branch point greater than the previous one use the the heatmap
1251 : // we just generated - it should include more layers.
1252 0 : if !tline.should_keep_previous_heatmap(end_lsn) {
1253 0 : tline
1254 0 : .previous_heatmap
1255 0 : .store(Some(Arc::new(unarchival_heatmap)));
1256 0 : } else {
1257 0 : tracing::info!("Previous heatmap preferred. Dropping unarchival heatmap.")
1258 : }
1259 :
1260 0 : match tline.ancestor_timeline() {
1261 0 : Some(ancestor) => {
1262 0 : if ancestor.update_layer_visibility().await.is_err() {
1263 : // Ancestor timeline is shutting down.
1264 0 : break;
1265 0 : }
1266 :
1267 0 : tline_ending_at = Some((ancestor, tline.get_ancestor_lsn()));
1268 : }
1269 0 : None => {
1270 0 : tline_ending_at = None;
1271 0 : }
1272 : }
1273 : }
1274 3 : }
1275 :
1276 0 : match import_pgdata {
1277 0 : Some(import_pgdata) if !import_pgdata.is_done() => {
1278 0 : let mut guard = self.timelines_creating.lock().unwrap();
1279 0 : if !guard.insert(timeline_id) {
1280 : // We should never try and load the same timeline twice during startup
1281 0 : unreachable!("Timeline {tenant_id}/{timeline_id} is already being created")
1282 0 : }
1283 0 : let timeline_create_guard = TimelineCreateGuard {
1284 0 : _tenant_gate_guard: self.gate.enter()?,
1285 0 : owning_tenant: self.clone(),
1286 0 : timeline_id,
1287 0 : idempotency,
1288 : // The users of this specific return value don't need the timline_path in there.
1289 0 : timeline_path: timeline
1290 0 : .conf
1291 0 : .timeline_path(&timeline.tenant_shard_id, &timeline.timeline_id),
1292 : };
1293 0 : Ok(TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
1294 0 : TimelineInitAndSyncNeedsSpawnImportPgdata {
1295 0 : timeline,
1296 0 : import_pgdata,
1297 0 : guard: timeline_create_guard,
1298 0 : },
1299 0 : ))
1300 : }
1301 : Some(_) | None => {
1302 : {
1303 3 : let mut timelines_accessor = self.timelines.lock().unwrap();
1304 3 : match timelines_accessor.entry(timeline_id) {
1305 : // We should never try and load the same timeline twice during startup
1306 : Entry::Occupied(_) => {
1307 0 : unreachable!(
1308 : "Timeline {tenant_id}/{timeline_id} already exists in the tenant map"
1309 : );
1310 : }
1311 3 : Entry::Vacant(v) => {
1312 3 : v.insert(Arc::clone(&timeline));
1313 3 : timeline.maybe_spawn_flush_loop();
1314 3 : }
1315 : }
1316 : }
1317 :
1318 3 : if disk_consistent_lsn.is_valid() {
1319 : // Sanity check: a timeline should have some content.
1320 : // Exception: importing timelines might not yet have any
1321 3 : anyhow::ensure!(
1322 3 : ancestor.is_some()
1323 2 : || timeline
1324 2 : .layers
1325 2 : .read(LayerManagerLockHolder::LoadLayerMap)
1326 2 : .await
1327 2 : .layer_map()
1328 2 : .expect(
1329 2 : "currently loading, layer manager cannot be shutdown already"
1330 : )
1331 2 : .iter_historic_layers()
1332 2 : .next()
1333 2 : .is_some(),
1334 0 : "Timeline has no ancestor and no layer files"
1335 : );
1336 0 : }
1337 :
1338 3 : Ok(TimelineInitAndSyncResult::ReadyToActivate)
1339 : }
1340 : }
1341 3 : }
1342 :
1343 : /// Attach a tenant that's available in cloud storage.
1344 : ///
1345 : /// This returns quickly, after just creating the in-memory object
1346 : /// Tenant struct and launching a background task to download
1347 : /// the remote index files. On return, the tenant is most likely still in
1348 : /// Attaching state, and it will become Active once the background task
1349 : /// finishes. You can use wait_until_active() to wait for the task to
1350 : /// complete.
1351 : ///
1352 : #[allow(clippy::too_many_arguments)]
1353 0 : pub(crate) fn spawn(
1354 0 : conf: &'static PageServerConf,
1355 0 : tenant_shard_id: TenantShardId,
1356 0 : resources: TenantSharedResources,
1357 0 : attached_conf: AttachedTenantConf,
1358 0 : shard_identity: ShardIdentity,
1359 0 : init_order: Option<InitializationOrder>,
1360 0 : mode: SpawnMode,
1361 0 : ctx: &RequestContext,
1362 0 : ) -> Result<Arc<TenantShard>, GlobalShutDown> {
1363 0 : let wal_redo_manager =
1364 0 : WalRedoManager::new(PostgresRedoManager::new(conf, tenant_shard_id))?;
1365 :
1366 : let TenantSharedResources {
1367 0 : broker_client,
1368 0 : remote_storage,
1369 0 : deletion_queue_client,
1370 0 : l0_flush_global_state,
1371 0 : basebackup_cache,
1372 0 : feature_resolver,
1373 0 : } = resources;
1374 :
1375 0 : let attach_mode = attached_conf.location.attach_mode;
1376 0 : let generation = attached_conf.location.generation;
1377 :
1378 0 : let tenant = Arc::new(TenantShard::new(
1379 0 : TenantState::Attaching,
1380 0 : conf,
1381 0 : attached_conf,
1382 0 : shard_identity,
1383 0 : Some(wal_redo_manager),
1384 0 : tenant_shard_id,
1385 0 : remote_storage.clone(),
1386 0 : deletion_queue_client,
1387 0 : l0_flush_global_state,
1388 0 : basebackup_cache,
1389 0 : feature_resolver,
1390 : ));
1391 :
1392 : // The attach task will carry a GateGuard, so that shutdown() reliably waits for it to drop out if
1393 : // we shut down while attaching.
1394 0 : let attach_gate_guard = tenant
1395 0 : .gate
1396 0 : .enter()
1397 0 : .expect("We just created the TenantShard: nothing else can have shut it down yet");
1398 :
1399 : // Do all the hard work in the background
1400 0 : let tenant_clone = Arc::clone(&tenant);
1401 0 : let ctx = ctx.detached_child(TaskKind::Attach, DownloadBehavior::Warn);
1402 0 : task_mgr::spawn(
1403 0 : &tokio::runtime::Handle::current(),
1404 0 : TaskKind::Attach,
1405 0 : tenant_shard_id,
1406 0 : None,
1407 0 : "attach tenant",
1408 0 : async move {
1409 :
1410 0 : info!(
1411 : ?attach_mode,
1412 0 : "Attaching tenant"
1413 : );
1414 :
1415 0 : let _gate_guard = attach_gate_guard;
1416 :
1417 : // Is this tenant being spawned as part of process startup?
1418 0 : let starting_up = init_order.is_some();
1419 0 : scopeguard::defer! {
1420 : if starting_up {
1421 : TENANT.startup_complete.inc();
1422 : }
1423 : }
1424 :
1425 0 : fn make_broken_or_stopping(t: &TenantShard, err: anyhow::Error) {
1426 0 : t.state.send_modify(|state| match state {
1427 : // TODO: the old code alluded to DeleteTenantFlow sometimes setting
1428 : // TenantState::Stopping before we get here, but this may be outdated.
1429 : // Let's find out with a testing assertion. If this doesn't fire, and the
1430 : // logs don't show this happening in production, remove the Stopping cases.
1431 0 : TenantState::Stopping{..} if cfg!(any(test, feature = "testing")) => {
1432 0 : panic!("unexpected TenantState::Stopping during attach")
1433 : }
1434 : // If the tenant is cancelled, assume the error was caused by cancellation.
1435 0 : TenantState::Attaching if t.cancel.is_cancelled() => {
1436 0 : info!("attach cancelled, setting tenant state to Stopping: {err}");
1437 : // NB: progress None tells `set_stopping` that attach has cancelled.
1438 0 : *state = TenantState::Stopping { progress: None };
1439 : }
1440 : // According to the old code, DeleteTenantFlow may already have set this to
1441 : // Stopping. Retain its progress.
1442 : // TODO: there is no DeleteTenantFlow. Is this still needed? See above.
1443 0 : TenantState::Stopping { progress } if t.cancel.is_cancelled() => {
1444 0 : assert!(progress.is_some(), "concurrent attach cancellation");
1445 0 : info!("attach cancelled, already Stopping: {err}");
1446 : }
1447 : // Mark the tenant as broken.
1448 : TenantState::Attaching | TenantState::Stopping { .. } => {
1449 0 : error!("attach failed, setting tenant state to Broken (was {state}): {err:?}");
1450 0 : *state = TenantState::broken_from_reason(err.to_string())
1451 : }
1452 : // The attach task owns the tenant state until activated.
1453 0 : state => panic!("invalid tenant state {state} during attach: {err:?}"),
1454 0 : });
1455 0 : }
1456 :
1457 : // TODO: should also be rejecting tenant conf changes that violate this check.
1458 0 : if let Err(e) = crate::tenant::storage_layer::inmemory_layer::IndexEntry::validate_checkpoint_distance(tenant_clone.get_checkpoint_distance()) {
1459 0 : make_broken_or_stopping(&tenant_clone, anyhow::anyhow!(e));
1460 0 : return Ok(());
1461 0 : }
1462 :
1463 0 : let mut init_order = init_order;
1464 : // take the completion because initial tenant loading will complete when all of
1465 : // these tasks complete.
1466 0 : let _completion = init_order
1467 0 : .as_mut()
1468 0 : .and_then(|x| x.initial_tenant_load.take());
1469 0 : let remote_load_completion = init_order
1470 0 : .as_mut()
1471 0 : .and_then(|x| x.initial_tenant_load_remote.take());
1472 :
1473 : enum AttachType<'a> {
1474 : /// We are attaching this tenant lazily in the background.
1475 : Warmup {
1476 : _permit: tokio::sync::SemaphorePermit<'a>,
1477 : during_startup: bool
1478 : },
1479 : /// We are attaching this tenant as soon as we can, because for example an
1480 : /// endpoint tried to access it.
1481 : OnDemand,
1482 : /// During normal operations after startup, we are attaching a tenant, and
1483 : /// eager attach was requested.
1484 : Normal,
1485 : }
1486 :
1487 0 : let attach_type = if matches!(mode, SpawnMode::Lazy) {
1488 : // Before doing any I/O, wait for at least one of:
1489 : // - A client attempting to access to this tenant (on-demand loading)
1490 : // - A permit becoming available in the warmup semaphore (background warmup)
1491 :
1492 0 : tokio::select!(
1493 0 : permit = tenant_clone.activate_now_sem.acquire() => {
1494 0 : let _ = permit.expect("activate_now_sem is never closed");
1495 0 : tracing::info!("Activating tenant (on-demand)");
1496 0 : AttachType::OnDemand
1497 : },
1498 0 : permit = conf.concurrent_tenant_warmup.inner().acquire() => {
1499 0 : let _permit = permit.expect("concurrent_tenant_warmup semaphore is never closed");
1500 0 : tracing::info!("Activating tenant (warmup)");
1501 0 : AttachType::Warmup {
1502 0 : _permit,
1503 0 : during_startup: init_order.is_some()
1504 0 : }
1505 : }
1506 0 : _ = tenant_clone.cancel.cancelled() => {
1507 : // This is safe, but should be pretty rare: it is interesting if a tenant
1508 : // stayed in Activating for such a long time that shutdown found it in
1509 : // that state.
1510 0 : tracing::info!(state=%tenant_clone.current_state(), "Tenant shut down before activation");
1511 : // Set the tenant to Stopping to signal `set_stopping` that we're done.
1512 0 : make_broken_or_stopping(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"));
1513 0 : return Ok(());
1514 : },
1515 : )
1516 : } else {
1517 : // SpawnMode::{Create,Eager} always cause jumping ahead of the
1518 : // concurrent_tenant_warmup queue
1519 0 : AttachType::Normal
1520 : };
1521 :
1522 0 : let preload = match &mode {
1523 : SpawnMode::Eager | SpawnMode::Lazy => {
1524 0 : let _preload_timer = TENANT.preload.start_timer();
1525 0 : let res = tenant_clone
1526 0 : .preload(&remote_storage, task_mgr::shutdown_token())
1527 0 : .await;
1528 0 : match res {
1529 0 : Ok(p) => Some(p),
1530 0 : Err(e) => {
1531 0 : make_broken_or_stopping(&tenant_clone, anyhow::anyhow!(e));
1532 0 : return Ok(());
1533 : }
1534 : }
1535 : }
1536 :
1537 : };
1538 :
1539 : // Remote preload is complete.
1540 0 : drop(remote_load_completion);
1541 :
1542 :
1543 : // We will time the duration of the attach phase unless this is a creation (attach will do no work)
1544 0 : let attach_start = std::time::Instant::now();
1545 0 : let attached = {
1546 0 : let _attach_timer = Some(TENANT.attach.start_timer());
1547 0 : tenant_clone.attach(preload, &ctx).await
1548 : };
1549 0 : let attach_duration = attach_start.elapsed();
1550 0 : _ = tenant_clone.attach_wal_lag_cooldown.set(WalLagCooldown::new(attach_start, attach_duration));
1551 :
1552 0 : match attached {
1553 : Ok(()) => {
1554 0 : info!("attach finished, activating");
1555 0 : tenant_clone.activate(broker_client, None, &ctx);
1556 : }
1557 0 : Err(e) => make_broken_or_stopping(&tenant_clone, anyhow::anyhow!(e)),
1558 : }
1559 :
1560 : // If we are doing an opportunistic warmup attachment at startup, initialize
1561 : // logical size at the same time. This is better than starting a bunch of idle tenants
1562 : // with cold caches and then coming back later to initialize their logical sizes.
1563 : //
1564 : // It also prevents the warmup proccess competing with the concurrency limit on
1565 : // logical size calculations: if logical size calculation semaphore is saturated,
1566 : // then warmup will wait for that before proceeding to the next tenant.
1567 0 : if matches!(attach_type, AttachType::Warmup { during_startup: true, .. }) {
1568 0 : let mut futs: FuturesUnordered<_> = tenant_clone.timelines.lock().unwrap().values().cloned().map(|t| t.await_initial_logical_size()).collect();
1569 0 : tracing::info!("Waiting for initial logical sizes while warming up...");
1570 0 : while futs.next().await.is_some() {}
1571 0 : tracing::info!("Warm-up complete");
1572 0 : }
1573 :
1574 0 : Ok(())
1575 0 : }
1576 0 : .instrument(tracing::info_span!(parent: None, "attach", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), gen=?generation)),
1577 : );
1578 0 : Ok(tenant)
1579 0 : }
1580 :
1581 : #[instrument(skip_all)]
1582 : pub(crate) async fn preload(
1583 : self: &Arc<Self>,
1584 : remote_storage: &GenericRemoteStorage,
1585 : cancel: CancellationToken,
1586 : ) -> anyhow::Result<TenantPreload> {
1587 : span::debug_assert_current_span_has_tenant_id();
1588 : // Get list of remote timelines
1589 : // download index files for every tenant timeline
1590 : info!("listing remote timelines");
1591 : let (mut remote_timeline_ids, other_keys) = remote_timeline_client::list_remote_timelines(
1592 : remote_storage,
1593 : self.tenant_shard_id,
1594 : cancel.clone(),
1595 : )
1596 : .await?;
1597 :
1598 : let tenant_manifest = match download_tenant_manifest(
1599 : remote_storage,
1600 : &self.tenant_shard_id,
1601 : self.generation,
1602 : &cancel,
1603 : )
1604 : .await
1605 : {
1606 : Ok((tenant_manifest, _, _)) => Some(tenant_manifest),
1607 : Err(DownloadError::NotFound) => None,
1608 : Err(err) => return Err(err.into()),
1609 : };
1610 :
1611 : info!(
1612 : "found {} timelines ({} offloaded timelines)",
1613 : remote_timeline_ids.len(),
1614 : tenant_manifest
1615 : .as_ref()
1616 3 : .map(|m| m.offloaded_timelines.len())
1617 : .unwrap_or(0)
1618 : );
1619 :
1620 : for k in other_keys {
1621 : warn!("Unexpected non timeline key {k}");
1622 : }
1623 :
1624 : // Avoid downloading IndexPart of offloaded timelines.
1625 : let mut offloaded_with_prefix = HashSet::new();
1626 : if let Some(tenant_manifest) = &tenant_manifest {
1627 : for offloaded in tenant_manifest.offloaded_timelines.iter() {
1628 : if remote_timeline_ids.remove(&offloaded.timeline_id) {
1629 : offloaded_with_prefix.insert(offloaded.timeline_id);
1630 : } else {
1631 : // We'll take care later of timelines in the manifest without a prefix
1632 : }
1633 : }
1634 : }
1635 :
1636 : // TODO(vlad): Could go to S3 if the secondary is freezing cold and hasn't even
1637 : // pulled the first heatmap. Not entirely necessary since the storage controller
1638 : // will kick the secondary in any case and cause a download.
1639 : let maybe_heatmap_at = self.read_on_disk_heatmap().await;
1640 :
1641 : let timelines = self
1642 : .load_timelines_metadata(
1643 : remote_timeline_ids,
1644 : remote_storage,
1645 : maybe_heatmap_at,
1646 : cancel,
1647 : )
1648 : .await?;
1649 :
1650 : Ok(TenantPreload {
1651 : tenant_manifest,
1652 : timelines: timelines
1653 : .into_iter()
1654 3 : .map(|(id, tl)| (id, Some(tl)))
1655 0 : .chain(offloaded_with_prefix.into_iter().map(|id| (id, None)))
1656 : .collect(),
1657 : })
1658 : }
1659 :
1660 119 : async fn read_on_disk_heatmap(&self) -> Option<(HeatMapTenant, std::time::Instant)> {
1661 119 : if !self.conf.load_previous_heatmap {
1662 0 : return None;
1663 119 : }
1664 :
1665 119 : let on_disk_heatmap_path = self.conf.tenant_heatmap_path(&self.tenant_shard_id);
1666 119 : match tokio::fs::read_to_string(on_disk_heatmap_path).await {
1667 0 : Ok(heatmap) => match serde_json::from_str::<HeatMapTenant>(&heatmap) {
1668 0 : Ok(heatmap) => Some((heatmap, std::time::Instant::now())),
1669 0 : Err(err) => {
1670 0 : error!("Failed to deserialize old heatmap: {err}");
1671 0 : None
1672 : }
1673 : },
1674 119 : Err(err) => match err.kind() {
1675 119 : std::io::ErrorKind::NotFound => None,
1676 : _ => {
1677 0 : error!("Unexpected IO error reading old heatmap: {err}");
1678 0 : None
1679 : }
1680 : },
1681 : }
1682 119 : }
1683 :
1684 : ///
1685 : /// Background task that downloads all data for a tenant and brings it to Active state.
1686 : ///
1687 : /// No background tasks are started as part of this routine.
1688 : ///
1689 119 : async fn attach(
1690 119 : self: &Arc<TenantShard>,
1691 119 : preload: Option<TenantPreload>,
1692 119 : ctx: &RequestContext,
1693 119 : ) -> anyhow::Result<()> {
1694 119 : span::debug_assert_current_span_has_tenant_id();
1695 :
1696 119 : failpoint_support::sleep_millis_async!("before-attaching-tenant");
1697 :
1698 119 : let Some(preload) = preload else {
1699 0 : anyhow::bail!(
1700 0 : "local-only deployment is no longer supported, https://github.com/neondatabase/neon/issues/5624"
1701 : );
1702 : };
1703 :
1704 119 : let mut offloaded_timeline_ids = HashSet::new();
1705 119 : let mut offloaded_timelines_list = Vec::new();
1706 119 : if let Some(tenant_manifest) = &preload.tenant_manifest {
1707 3 : for timeline_manifest in tenant_manifest.offloaded_timelines.iter() {
1708 0 : let timeline_id = timeline_manifest.timeline_id;
1709 0 : let offloaded_timeline =
1710 0 : OffloadedTimeline::from_manifest(self.tenant_shard_id, timeline_manifest);
1711 0 : offloaded_timelines_list.push((timeline_id, Arc::new(offloaded_timeline)));
1712 0 : offloaded_timeline_ids.insert(timeline_id);
1713 0 : }
1714 116 : }
1715 : // Complete deletions for offloaded timeline id's from manifest.
1716 : // The manifest will be uploaded later in this function.
1717 119 : offloaded_timelines_list
1718 119 : .retain(|(offloaded_id, offloaded)| {
1719 : // Existence of a timeline is finally determined by the existence of an index-part.json in remote storage.
1720 : // If there is dangling references in another location, they need to be cleaned up.
1721 0 : let delete = !preload.timelines.contains_key(offloaded_id);
1722 0 : if delete {
1723 0 : tracing::info!("Removing offloaded timeline {offloaded_id} from manifest as no remote prefix was found");
1724 0 : offloaded.defuse_for_tenant_drop();
1725 0 : }
1726 0 : !delete
1727 0 : });
1728 :
1729 119 : let mut timelines_to_resume_deletions = vec![];
1730 :
1731 119 : let mut remote_index_and_client = HashMap::new();
1732 119 : let mut timeline_ancestors = HashMap::new();
1733 119 : let mut existent_timelines = HashSet::new();
1734 122 : for (timeline_id, preload) in preload.timelines {
1735 3 : let Some(preload) = preload else { continue };
1736 : // This is an invariant of the `preload` function's API
1737 3 : assert!(!offloaded_timeline_ids.contains(&timeline_id));
1738 3 : let index_part = match preload.index_part {
1739 3 : Ok(i) => {
1740 3 : debug!("remote index part exists for timeline {timeline_id}");
1741 : // We found index_part on the remote, this is the standard case.
1742 3 : existent_timelines.insert(timeline_id);
1743 3 : i
1744 : }
1745 : Err(DownloadError::NotFound) => {
1746 : // There is no index_part on the remote. We only get here
1747 : // if there is some prefix for the timeline in the remote storage.
1748 : // This can e.g. be the initdb.tar.zst archive, maybe a
1749 : // remnant from a prior incomplete creation or deletion attempt.
1750 : // Delete the local directory as the deciding criterion for a
1751 : // timeline's existence is presence of index_part.
1752 0 : info!(%timeline_id, "index_part not found on remote");
1753 0 : continue;
1754 : }
1755 0 : Err(DownloadError::Fatal(why)) => {
1756 : // If, while loading one remote timeline, we saw an indication that our generation
1757 : // number is likely invalid, then we should not load the whole tenant.
1758 0 : error!(%timeline_id, "Fatal error loading timeline: {why}");
1759 0 : anyhow::bail!(why.to_string());
1760 : }
1761 0 : Err(e) => {
1762 : // Some (possibly ephemeral) error happened during index_part download.
1763 : // Pretend the timeline exists to not delete the timeline directory,
1764 : // as it might be a temporary issue and we don't want to re-download
1765 : // everything after it resolves.
1766 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1767 :
1768 0 : existent_timelines.insert(timeline_id);
1769 0 : continue;
1770 : }
1771 : };
1772 3 : match index_part {
1773 3 : MaybeDeletedIndexPart::IndexPart(index_part) => {
1774 3 : timeline_ancestors.insert(timeline_id, index_part.metadata.clone());
1775 3 : remote_index_and_client.insert(
1776 3 : timeline_id,
1777 3 : (index_part, preload.client, preload.previous_heatmap),
1778 3 : );
1779 3 : }
1780 0 : MaybeDeletedIndexPart::Deleted(index_part) => {
1781 0 : info!(
1782 0 : "timeline {} is deleted, picking to resume deletion",
1783 : timeline_id
1784 : );
1785 0 : timelines_to_resume_deletions.push((timeline_id, index_part, preload.client));
1786 : }
1787 : }
1788 : }
1789 :
1790 119 : let mut gc_blocks = HashMap::new();
1791 :
1792 : // For every timeline, download the metadata file, scan the local directory,
1793 : // and build a layer map that contains an entry for each remote and local
1794 : // layer file.
1795 119 : let sorted_timelines = tree_sort_timelines(timeline_ancestors, |m| m.ancestor_timeline())?;
1796 122 : for (timeline_id, remote_metadata) in sorted_timelines {
1797 3 : let (index_part, remote_client, previous_heatmap) = remote_index_and_client
1798 3 : .remove(&timeline_id)
1799 3 : .expect("just put it in above");
1800 :
1801 3 : if let Some(blocking) = index_part.gc_blocking.as_ref() {
1802 : // could just filter these away, but it helps while testing
1803 0 : anyhow::ensure!(
1804 0 : !blocking.reasons.is_empty(),
1805 0 : "index_part for {timeline_id} is malformed: it should not have gc blocking with zero reasons"
1806 : );
1807 0 : let prev = gc_blocks.insert(timeline_id, blocking.reasons);
1808 0 : assert!(prev.is_none());
1809 3 : }
1810 :
1811 : // TODO again handle early failure
1812 3 : let effect = self
1813 3 : .load_remote_timeline(
1814 3 : timeline_id,
1815 3 : index_part,
1816 3 : remote_metadata,
1817 3 : previous_heatmap,
1818 3 : self.get_timeline_resources_for(remote_client),
1819 3 : LoadTimelineCause::Attach,
1820 3 : ctx,
1821 3 : )
1822 3 : .await
1823 3 : .with_context(|| {
1824 0 : format!(
1825 0 : "failed to load remote timeline {} for tenant {}",
1826 0 : timeline_id, self.tenant_shard_id
1827 : )
1828 0 : })?;
1829 :
1830 3 : match effect {
1831 3 : TimelineInitAndSyncResult::ReadyToActivate => {
1832 3 : // activation happens later, on Tenant::activate
1833 3 : }
1834 : TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
1835 : TimelineInitAndSyncNeedsSpawnImportPgdata {
1836 0 : timeline,
1837 0 : import_pgdata,
1838 0 : guard,
1839 : },
1840 : ) => {
1841 0 : let timeline_id = timeline.timeline_id;
1842 0 : let import_task_gate = Gate::default();
1843 0 : let import_task_guard = import_task_gate.enter().unwrap();
1844 0 : let import_task_handle =
1845 0 : tokio::task::spawn(self.clone().create_timeline_import_pgdata_task(
1846 0 : timeline.clone(),
1847 0 : import_pgdata,
1848 0 : guard,
1849 0 : import_task_guard,
1850 0 : ctx.detached_child(TaskKind::ImportPgdata, DownloadBehavior::Warn),
1851 : ));
1852 :
1853 0 : let prev = self.timelines_importing.lock().unwrap().insert(
1854 0 : timeline_id,
1855 0 : Arc::new(ImportingTimeline {
1856 0 : timeline: timeline.clone(),
1857 0 : import_task_handle,
1858 0 : import_task_gate,
1859 0 : delete_progress: TimelineDeleteProgress::default(),
1860 0 : }),
1861 0 : );
1862 :
1863 0 : assert!(prev.is_none());
1864 : }
1865 : }
1866 : }
1867 :
1868 : // At this point we've initialized all timelines and are tracking them.
1869 : // Now compute the layer visibility for all (not offloaded) timelines.
1870 119 : let compute_visiblity_for = {
1871 119 : let timelines_accessor = self.timelines.lock().unwrap();
1872 119 : let mut timelines_offloaded_accessor = self.timelines_offloaded.lock().unwrap();
1873 :
1874 119 : timelines_offloaded_accessor.extend(offloaded_timelines_list.into_iter());
1875 :
1876 : // Before activation, populate each Timeline's GcInfo with information about its children
1877 119 : self.initialize_gc_info(&timelines_accessor, &timelines_offloaded_accessor, None);
1878 :
1879 119 : timelines_accessor.values().cloned().collect::<Vec<_>>()
1880 : };
1881 :
1882 122 : for tl in compute_visiblity_for {
1883 3 : tl.update_layer_visibility().await.with_context(|| {
1884 0 : format!(
1885 0 : "failed initial timeline visibility computation {} for tenant {}",
1886 0 : tl.timeline_id, self.tenant_shard_id
1887 : )
1888 0 : })?;
1889 : }
1890 :
1891 : // Walk through deleted timelines, resume deletion
1892 119 : for (timeline_id, index_part, remote_timeline_client) in timelines_to_resume_deletions {
1893 0 : remote_timeline_client
1894 0 : .init_upload_queue_stopped_to_continue_deletion(&index_part)
1895 0 : .context("init queue stopped")
1896 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1897 :
1898 0 : DeleteTimelineFlow::resume_deletion(
1899 0 : Arc::clone(self),
1900 0 : timeline_id,
1901 0 : &index_part.metadata,
1902 0 : remote_timeline_client,
1903 0 : ctx,
1904 : )
1905 0 : .instrument(tracing::info_span!("timeline_delete", %timeline_id))
1906 0 : .await
1907 0 : .context("resume_deletion")
1908 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1909 : }
1910 :
1911 : // Stash the preloaded tenant manifest, and upload a new manifest if changed.
1912 : //
1913 : // NB: this must happen after the tenant is fully populated above. In particular the
1914 : // offloaded timelines, which are included in the manifest.
1915 : {
1916 119 : let mut guard = self.remote_tenant_manifest.lock().await;
1917 119 : assert!(guard.is_none(), "tenant manifest set before preload"); // first populated here
1918 119 : *guard = preload.tenant_manifest;
1919 : }
1920 119 : self.maybe_upload_tenant_manifest().await?;
1921 :
1922 : // The local filesystem contents are a cache of what's in the remote IndexPart;
1923 : // IndexPart is the source of truth.
1924 119 : self.clean_up_timelines(&existent_timelines)?;
1925 :
1926 119 : self.gc_block.set_scanned(gc_blocks);
1927 :
1928 119 : fail::fail_point!("attach-before-activate", |_| {
1929 0 : anyhow::bail!("attach-before-activate");
1930 0 : });
1931 119 : failpoint_support::sleep_millis_async!("attach-before-activate-sleep", &self.cancel);
1932 :
1933 119 : info!("Done");
1934 :
1935 119 : Ok(())
1936 119 : }
1937 :
1938 : /// Check for any local timeline directories that are temporary, or do not correspond to a
1939 : /// timeline that still exists: this can happen if we crashed during a deletion/creation, or
1940 : /// if a timeline was deleted while the tenant was attached to a different pageserver.
1941 119 : fn clean_up_timelines(&self, existent_timelines: &HashSet<TimelineId>) -> anyhow::Result<()> {
1942 119 : let timelines_dir = self.conf.timelines_path(&self.tenant_shard_id);
1943 :
1944 119 : let entries = match timelines_dir.read_dir_utf8() {
1945 119 : Ok(d) => d,
1946 0 : Err(e) => {
1947 0 : if e.kind() == std::io::ErrorKind::NotFound {
1948 0 : return Ok(());
1949 : } else {
1950 0 : return Err(e).context("list timelines directory for tenant");
1951 : }
1952 : }
1953 : };
1954 :
1955 123 : for entry in entries {
1956 4 : let entry = entry.context("read timeline dir entry")?;
1957 4 : let entry_path = entry.path();
1958 :
1959 4 : let purge = if crate::is_temporary(entry_path) {
1960 0 : true
1961 : } else {
1962 4 : match TimelineId::try_from(entry_path.file_name()) {
1963 4 : Ok(i) => {
1964 : // Purge if the timeline ID does not exist in remote storage: remote storage is the authority.
1965 4 : !existent_timelines.contains(&i)
1966 : }
1967 0 : Err(e) => {
1968 0 : tracing::warn!(
1969 0 : "Unparseable directory in timelines directory: {entry_path}, ignoring ({e})"
1970 : );
1971 : // Do not purge junk: if we don't recognize it, be cautious and leave it for a human.
1972 0 : false
1973 : }
1974 : }
1975 : };
1976 :
1977 4 : if purge {
1978 1 : tracing::info!("Purging stale timeline dentry {entry_path}");
1979 1 : if let Err(e) = match entry.file_type() {
1980 1 : Ok(t) => if t.is_dir() {
1981 1 : std::fs::remove_dir_all(entry_path)
1982 : } else {
1983 0 : std::fs::remove_file(entry_path)
1984 : }
1985 1 : .or_else(fs_ext::ignore_not_found),
1986 0 : Err(e) => Err(e),
1987 : } {
1988 0 : tracing::warn!("Failed to purge stale timeline dentry {entry_path}: {e}");
1989 1 : }
1990 3 : }
1991 : }
1992 :
1993 119 : Ok(())
1994 119 : }
1995 :
1996 : /// Get sum of all remote timelines sizes
1997 : ///
1998 : /// This function relies on the index_part instead of listing the remote storage
1999 0 : pub fn remote_size(&self) -> u64 {
2000 0 : let mut size = 0;
2001 :
2002 0 : for timeline in self.list_timelines() {
2003 0 : size += timeline.remote_client.get_remote_physical_size();
2004 0 : }
2005 :
2006 0 : size
2007 0 : }
2008 :
2009 : #[instrument(skip_all, fields(timeline_id=%timeline_id))]
2010 : #[allow(clippy::too_many_arguments)]
2011 : async fn load_remote_timeline(
2012 : self: &Arc<Self>,
2013 : timeline_id: TimelineId,
2014 : index_part: IndexPart,
2015 : remote_metadata: TimelineMetadata,
2016 : previous_heatmap: Option<PreviousHeatmap>,
2017 : resources: TimelineResources,
2018 : cause: LoadTimelineCause,
2019 : ctx: &RequestContext,
2020 : ) -> anyhow::Result<TimelineInitAndSyncResult> {
2021 : span::debug_assert_current_span_has_tenant_id();
2022 :
2023 : info!("downloading index file for timeline {}", timeline_id);
2024 : tokio::fs::create_dir_all(self.conf.timeline_path(&self.tenant_shard_id, &timeline_id))
2025 : .await
2026 : .context("Failed to create new timeline directory")?;
2027 :
2028 : let ancestor = if let Some(ancestor_id) = remote_metadata.ancestor_timeline() {
2029 : let timelines = self.timelines.lock().unwrap();
2030 : Some(Arc::clone(timelines.get(&ancestor_id).ok_or_else(
2031 0 : || {
2032 0 : anyhow::anyhow!(
2033 0 : "cannot find ancestor timeline {ancestor_id} for timeline {timeline_id}"
2034 : )
2035 0 : },
2036 : )?))
2037 : } else {
2038 : None
2039 : };
2040 :
2041 : self.timeline_init_and_sync(
2042 : timeline_id,
2043 : resources,
2044 : index_part,
2045 : remote_metadata,
2046 : previous_heatmap,
2047 : ancestor,
2048 : cause,
2049 : ctx,
2050 : )
2051 : .await
2052 : }
2053 :
2054 119 : async fn load_timelines_metadata(
2055 119 : self: &Arc<TenantShard>,
2056 119 : timeline_ids: HashSet<TimelineId>,
2057 119 : remote_storage: &GenericRemoteStorage,
2058 119 : heatmap: Option<(HeatMapTenant, std::time::Instant)>,
2059 119 : cancel: CancellationToken,
2060 119 : ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
2061 119 : let mut timeline_heatmaps = heatmap.map(|h| (h.0.into_timelines_index(), h.1));
2062 :
2063 119 : let mut part_downloads = JoinSet::new();
2064 122 : for timeline_id in timeline_ids {
2065 3 : let cancel_clone = cancel.clone();
2066 :
2067 3 : let previous_timeline_heatmap = timeline_heatmaps.as_mut().and_then(|hs| {
2068 0 : hs.0.remove(&timeline_id).map(|h| PreviousHeatmap::Active {
2069 0 : heatmap: h,
2070 0 : read_at: hs.1,
2071 0 : end_lsn: None,
2072 0 : })
2073 0 : });
2074 3 : part_downloads.spawn(
2075 3 : self.load_timeline_metadata(
2076 3 : timeline_id,
2077 3 : remote_storage.clone(),
2078 3 : previous_timeline_heatmap,
2079 3 : cancel_clone,
2080 : )
2081 3 : .instrument(info_span!("download_index_part", %timeline_id)),
2082 : );
2083 : }
2084 :
2085 119 : let mut timeline_preloads: HashMap<TimelineId, TimelinePreload> = HashMap::new();
2086 :
2087 : loop {
2088 122 : tokio::select!(
2089 122 : next = part_downloads.join_next() => {
2090 122 : match next {
2091 3 : Some(result) => {
2092 3 : let preload = result.context("join preload task")?;
2093 3 : timeline_preloads.insert(preload.timeline_id, preload);
2094 : },
2095 : None => {
2096 119 : break;
2097 : }
2098 : }
2099 : },
2100 122 : _ = cancel.cancelled() => {
2101 0 : anyhow::bail!("Cancelled while waiting for remote index download")
2102 : }
2103 : )
2104 : }
2105 :
2106 119 : Ok(timeline_preloads)
2107 119 : }
2108 :
2109 3 : fn build_timeline_client(
2110 3 : &self,
2111 3 : timeline_id: TimelineId,
2112 3 : remote_storage: GenericRemoteStorage,
2113 3 : ) -> RemoteTimelineClient {
2114 3 : RemoteTimelineClient::new(
2115 3 : remote_storage.clone(),
2116 3 : self.deletion_queue_client.clone(),
2117 3 : self.conf,
2118 3 : self.tenant_shard_id,
2119 3 : timeline_id,
2120 3 : self.generation,
2121 3 : &self.tenant_conf.load().location,
2122 : )
2123 3 : }
2124 :
2125 3 : fn load_timeline_metadata(
2126 3 : self: &Arc<TenantShard>,
2127 3 : timeline_id: TimelineId,
2128 3 : remote_storage: GenericRemoteStorage,
2129 3 : previous_heatmap: Option<PreviousHeatmap>,
2130 3 : cancel: CancellationToken,
2131 3 : ) -> impl Future<Output = TimelinePreload> + use<> {
2132 3 : let client = self.build_timeline_client(timeline_id, remote_storage);
2133 3 : async move {
2134 3 : debug_assert_current_span_has_tenant_and_timeline_id();
2135 3 : debug!("starting index part download");
2136 :
2137 3 : let index_part = client.download_index_file(&cancel).await;
2138 :
2139 3 : debug!("finished index part download");
2140 :
2141 3 : TimelinePreload {
2142 3 : client,
2143 3 : timeline_id,
2144 3 : index_part,
2145 3 : previous_heatmap,
2146 3 : }
2147 3 : }
2148 3 : }
2149 :
2150 0 : fn check_to_be_archived_has_no_unarchived_children(
2151 0 : timeline_id: TimelineId,
2152 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
2153 0 : ) -> Result<(), TimelineArchivalError> {
2154 0 : let children: Vec<TimelineId> = timelines
2155 0 : .iter()
2156 0 : .filter_map(|(id, entry)| {
2157 0 : if entry.get_ancestor_timeline_id() != Some(timeline_id) {
2158 0 : return None;
2159 0 : }
2160 0 : if entry.is_archived() == Some(true) {
2161 0 : return None;
2162 0 : }
2163 0 : Some(*id)
2164 0 : })
2165 0 : .collect();
2166 :
2167 0 : if !children.is_empty() {
2168 0 : return Err(TimelineArchivalError::HasUnarchivedChildren(children));
2169 0 : }
2170 0 : Ok(())
2171 0 : }
2172 :
2173 0 : fn check_ancestor_of_to_be_unarchived_is_not_archived(
2174 0 : ancestor_timeline_id: TimelineId,
2175 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
2176 0 : offloaded_timelines: &std::sync::MutexGuard<
2177 0 : '_,
2178 0 : HashMap<TimelineId, Arc<OffloadedTimeline>>,
2179 0 : >,
2180 0 : ) -> Result<(), TimelineArchivalError> {
2181 0 : let has_archived_parent =
2182 0 : if let Some(ancestor_timeline) = timelines.get(&ancestor_timeline_id) {
2183 0 : ancestor_timeline.is_archived() == Some(true)
2184 0 : } else if offloaded_timelines.contains_key(&ancestor_timeline_id) {
2185 0 : true
2186 : } else {
2187 0 : error!("ancestor timeline {ancestor_timeline_id} not found");
2188 0 : if cfg!(debug_assertions) {
2189 0 : panic!("ancestor timeline {ancestor_timeline_id} not found");
2190 0 : }
2191 0 : return Err(TimelineArchivalError::NotFound);
2192 : };
2193 0 : if has_archived_parent {
2194 0 : return Err(TimelineArchivalError::HasArchivedParent(
2195 0 : ancestor_timeline_id,
2196 0 : ));
2197 0 : }
2198 0 : Ok(())
2199 0 : }
2200 :
2201 0 : fn check_to_be_unarchived_timeline_has_no_archived_parent(
2202 0 : timeline: &Arc<Timeline>,
2203 0 : ) -> Result<(), TimelineArchivalError> {
2204 0 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
2205 0 : if ancestor_timeline.is_archived() == Some(true) {
2206 0 : return Err(TimelineArchivalError::HasArchivedParent(
2207 0 : ancestor_timeline.timeline_id,
2208 0 : ));
2209 0 : }
2210 0 : }
2211 0 : Ok(())
2212 0 : }
2213 :
2214 : /// Loads the specified (offloaded) timeline from S3 and attaches it as a loaded timeline
2215 : ///
2216 : /// Counterpart to [`offload_timeline`].
2217 0 : async fn unoffload_timeline(
2218 0 : self: &Arc<Self>,
2219 0 : timeline_id: TimelineId,
2220 0 : broker_client: storage_broker::BrokerClientChannel,
2221 0 : ctx: RequestContext,
2222 0 : ) -> Result<Arc<Timeline>, TimelineArchivalError> {
2223 0 : info!("unoffloading timeline");
2224 :
2225 : // We activate the timeline below manually, so this must be called on an active tenant.
2226 : // We expect callers of this function to ensure this.
2227 0 : match self.current_state() {
2228 : TenantState::Activating { .. }
2229 : | TenantState::Attaching
2230 : | TenantState::Broken { .. } => {
2231 0 : panic!("Timeline expected to be active")
2232 : }
2233 0 : TenantState::Stopping { .. } => return Err(TimelineArchivalError::Cancelled),
2234 0 : TenantState::Active => {}
2235 : }
2236 0 : let cancel = self.cancel.clone();
2237 :
2238 : // Protect against concurrent attempts to use this TimelineId
2239 : // We don't care much about idempotency, as it's ensured a layer above.
2240 0 : let allow_offloaded = true;
2241 0 : let _create_guard = self
2242 0 : .create_timeline_create_guard(
2243 0 : timeline_id,
2244 0 : CreateTimelineIdempotency::FailWithConflict,
2245 0 : allow_offloaded,
2246 : )
2247 0 : .map_err(|err| match err {
2248 0 : TimelineExclusionError::AlreadyCreating => TimelineArchivalError::AlreadyInProgress,
2249 : TimelineExclusionError::AlreadyExists { .. } => {
2250 0 : TimelineArchivalError::Other(anyhow::anyhow!("Timeline already exists"))
2251 : }
2252 0 : TimelineExclusionError::Other(e) => TimelineArchivalError::Other(e),
2253 0 : TimelineExclusionError::ShuttingDown => TimelineArchivalError::Cancelled,
2254 0 : })?;
2255 :
2256 0 : let timeline_preload = self
2257 0 : .load_timeline_metadata(
2258 0 : timeline_id,
2259 0 : self.remote_storage.clone(),
2260 0 : None,
2261 0 : cancel.clone(),
2262 0 : )
2263 0 : .await;
2264 :
2265 0 : let index_part = match timeline_preload.index_part {
2266 0 : Ok(index_part) => {
2267 0 : debug!("remote index part exists for timeline {timeline_id}");
2268 0 : index_part
2269 : }
2270 : Err(DownloadError::NotFound) => {
2271 0 : error!(%timeline_id, "index_part not found on remote");
2272 0 : return Err(TimelineArchivalError::NotFound);
2273 : }
2274 0 : Err(DownloadError::Cancelled) => return Err(TimelineArchivalError::Cancelled),
2275 0 : Err(e) => {
2276 : // Some (possibly ephemeral) error happened during index_part download.
2277 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
2278 0 : return Err(TimelineArchivalError::Other(
2279 0 : anyhow::Error::new(e).context("downloading index_part from remote storage"),
2280 0 : ));
2281 : }
2282 : };
2283 0 : let index_part = match index_part {
2284 0 : MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
2285 0 : MaybeDeletedIndexPart::Deleted(_index_part) => {
2286 0 : info!("timeline is deleted according to index_part.json");
2287 0 : return Err(TimelineArchivalError::NotFound);
2288 : }
2289 : };
2290 0 : let remote_metadata = index_part.metadata.clone();
2291 0 : let timeline_resources = self.build_timeline_resources(timeline_id);
2292 0 : self.load_remote_timeline(
2293 0 : timeline_id,
2294 0 : index_part,
2295 0 : remote_metadata,
2296 0 : None,
2297 0 : timeline_resources,
2298 0 : LoadTimelineCause::Unoffload,
2299 0 : &ctx,
2300 0 : )
2301 0 : .await
2302 0 : .with_context(|| {
2303 0 : format!(
2304 0 : "failed to load remote timeline {} for tenant {}",
2305 0 : timeline_id, self.tenant_shard_id
2306 : )
2307 0 : })
2308 0 : .map_err(TimelineArchivalError::Other)?;
2309 :
2310 0 : let timeline = {
2311 0 : let timelines = self.timelines.lock().unwrap();
2312 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2313 0 : warn!("timeline not available directly after attach");
2314 : // This is not a panic because no locks are held between `load_remote_timeline`
2315 : // which puts the timeline into timelines, and our look into the timeline map.
2316 0 : return Err(TimelineArchivalError::Other(anyhow::anyhow!(
2317 0 : "timeline not available directly after attach"
2318 0 : )));
2319 : };
2320 0 : let mut offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2321 0 : match offloaded_timelines.remove(&timeline_id) {
2322 0 : Some(offloaded) => {
2323 0 : offloaded.delete_from_ancestor_with_timelines(&timelines);
2324 0 : }
2325 0 : None => warn!("timeline already removed from offloaded timelines"),
2326 : }
2327 :
2328 0 : self.initialize_gc_info(&timelines, &offloaded_timelines, Some(timeline_id));
2329 :
2330 0 : Arc::clone(timeline)
2331 : };
2332 :
2333 : // Upload new list of offloaded timelines to S3
2334 0 : self.maybe_upload_tenant_manifest().await?;
2335 :
2336 : // Activate the timeline (if it makes sense)
2337 0 : if !(timeline.is_broken() || timeline.is_stopping()) {
2338 0 : let background_jobs_can_start = None;
2339 0 : timeline.activate(
2340 0 : self.clone(),
2341 0 : broker_client.clone(),
2342 0 : background_jobs_can_start,
2343 0 : &ctx.with_scope_timeline(&timeline),
2344 0 : );
2345 0 : }
2346 :
2347 0 : info!("timeline unoffloading complete");
2348 0 : Ok(timeline)
2349 0 : }
2350 :
2351 0 : pub(crate) async fn apply_timeline_archival_config(
2352 0 : self: &Arc<Self>,
2353 0 : timeline_id: TimelineId,
2354 0 : new_state: TimelineArchivalState,
2355 0 : broker_client: storage_broker::BrokerClientChannel,
2356 0 : ctx: RequestContext,
2357 0 : ) -> Result<(), TimelineArchivalError> {
2358 0 : info!("setting timeline archival config");
2359 : // First part: figure out what is needed to do, and do validation
2360 0 : let timeline_or_unarchive_offloaded = 'outer: {
2361 0 : let timelines = self.timelines.lock().unwrap();
2362 :
2363 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2364 0 : let offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2365 0 : let Some(offloaded) = offloaded_timelines.get(&timeline_id) else {
2366 0 : return Err(TimelineArchivalError::NotFound);
2367 : };
2368 0 : if new_state == TimelineArchivalState::Archived {
2369 : // It's offloaded already, so nothing to do
2370 0 : return Ok(());
2371 0 : }
2372 0 : if let Some(ancestor_timeline_id) = offloaded.ancestor_timeline_id {
2373 0 : Self::check_ancestor_of_to_be_unarchived_is_not_archived(
2374 0 : ancestor_timeline_id,
2375 0 : &timelines,
2376 0 : &offloaded_timelines,
2377 0 : )?;
2378 0 : }
2379 0 : break 'outer None;
2380 : };
2381 :
2382 : // Do some validation. We release the timelines lock below, so there is potential
2383 : // for race conditions: these checks are more present to prevent misunderstandings of
2384 : // the API's capabilities, instead of serving as the sole way to defend their invariants.
2385 0 : match new_state {
2386 : TimelineArchivalState::Unarchived => {
2387 0 : Self::check_to_be_unarchived_timeline_has_no_archived_parent(timeline)?
2388 : }
2389 : TimelineArchivalState::Archived => {
2390 0 : Self::check_to_be_archived_has_no_unarchived_children(timeline_id, &timelines)?
2391 : }
2392 : }
2393 0 : Some(Arc::clone(timeline))
2394 : };
2395 :
2396 : // Second part: unoffload timeline (if needed)
2397 0 : let timeline = if let Some(timeline) = timeline_or_unarchive_offloaded {
2398 0 : timeline
2399 : } else {
2400 : // Turn offloaded timeline into a non-offloaded one
2401 0 : self.unoffload_timeline(timeline_id, broker_client, ctx)
2402 0 : .await?
2403 : };
2404 :
2405 : // Third part: upload new timeline archival state and block until it is present in S3
2406 0 : let upload_needed = match timeline
2407 0 : .remote_client
2408 0 : .schedule_index_upload_for_timeline_archival_state(new_state)
2409 : {
2410 0 : Ok(upload_needed) => upload_needed,
2411 0 : Err(e) => {
2412 0 : if timeline.cancel.is_cancelled() {
2413 0 : return Err(TimelineArchivalError::Cancelled);
2414 : } else {
2415 0 : return Err(TimelineArchivalError::Other(e));
2416 : }
2417 : }
2418 : };
2419 :
2420 0 : if upload_needed {
2421 0 : info!("Uploading new state");
2422 : const MAX_WAIT: Duration = Duration::from_secs(10);
2423 0 : let Ok(v) =
2424 0 : tokio::time::timeout(MAX_WAIT, timeline.remote_client.wait_completion()).await
2425 : else {
2426 0 : tracing::warn!("reached timeout for waiting on upload queue");
2427 0 : return Err(TimelineArchivalError::Timeout);
2428 : };
2429 0 : v.map_err(|e| match e {
2430 0 : WaitCompletionError::NotInitialized(e) => {
2431 0 : TimelineArchivalError::Other(anyhow::anyhow!(e))
2432 : }
2433 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2434 0 : TimelineArchivalError::Cancelled
2435 : }
2436 0 : })?;
2437 0 : }
2438 0 : Ok(())
2439 0 : }
2440 :
2441 1 : pub fn get_offloaded_timeline(
2442 1 : &self,
2443 1 : timeline_id: TimelineId,
2444 1 : ) -> Result<Arc<OffloadedTimeline>, GetTimelineError> {
2445 1 : self.timelines_offloaded
2446 1 : .lock()
2447 1 : .unwrap()
2448 1 : .get(&timeline_id)
2449 1 : .map(Arc::clone)
2450 1 : .ok_or(GetTimelineError::NotFound {
2451 1 : tenant_id: self.tenant_shard_id,
2452 1 : timeline_id,
2453 1 : })
2454 1 : }
2455 :
2456 2 : pub(crate) fn tenant_shard_id(&self) -> TenantShardId {
2457 2 : self.tenant_shard_id
2458 2 : }
2459 :
2460 : /// Get Timeline handle for given Neon timeline ID.
2461 : /// This function is idempotent. It doesn't change internal state in any way.
2462 111 : pub fn get_timeline(
2463 111 : &self,
2464 111 : timeline_id: TimelineId,
2465 111 : active_only: bool,
2466 111 : ) -> Result<Arc<Timeline>, GetTimelineError> {
2467 111 : let timelines_accessor = self.timelines.lock().unwrap();
2468 111 : let timeline = timelines_accessor
2469 111 : .get(&timeline_id)
2470 111 : .ok_or(GetTimelineError::NotFound {
2471 111 : tenant_id: self.tenant_shard_id,
2472 111 : timeline_id,
2473 111 : })?;
2474 :
2475 110 : if active_only && !timeline.is_active() {
2476 0 : Err(GetTimelineError::NotActive {
2477 0 : tenant_id: self.tenant_shard_id,
2478 0 : timeline_id,
2479 0 : state: timeline.current_state(),
2480 0 : })
2481 : } else {
2482 110 : Ok(Arc::clone(timeline))
2483 : }
2484 111 : }
2485 :
2486 : /// Lists timelines the tenant contains.
2487 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2488 3 : pub fn list_timelines(&self) -> Vec<Arc<Timeline>> {
2489 3 : self.timelines
2490 3 : .lock()
2491 3 : .unwrap()
2492 3 : .values()
2493 3 : .map(Arc::clone)
2494 3 : .collect()
2495 3 : }
2496 :
2497 : /// Lists timelines the tenant contains.
2498 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2499 0 : pub fn list_importing_timelines(&self) -> Vec<Arc<ImportingTimeline>> {
2500 0 : self.timelines_importing
2501 0 : .lock()
2502 0 : .unwrap()
2503 0 : .values()
2504 0 : .map(Arc::clone)
2505 0 : .collect()
2506 0 : }
2507 :
2508 : /// Lists timelines the tenant manages, including offloaded ones.
2509 : ///
2510 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2511 0 : pub fn list_timelines_and_offloaded(
2512 0 : &self,
2513 0 : ) -> (Vec<Arc<Timeline>>, Vec<Arc<OffloadedTimeline>>) {
2514 0 : let timelines = self
2515 0 : .timelines
2516 0 : .lock()
2517 0 : .unwrap()
2518 0 : .values()
2519 0 : .map(Arc::clone)
2520 0 : .collect();
2521 0 : let offloaded = self
2522 0 : .timelines_offloaded
2523 0 : .lock()
2524 0 : .unwrap()
2525 0 : .values()
2526 0 : .map(Arc::clone)
2527 0 : .collect();
2528 0 : (timelines, offloaded)
2529 0 : }
2530 :
2531 0 : pub fn list_timeline_ids(&self) -> Vec<TimelineId> {
2532 0 : self.timelines.lock().unwrap().keys().cloned().collect()
2533 0 : }
2534 :
2535 : /// This is used by tests & import-from-basebackup.
2536 : ///
2537 : /// The returned [`UninitializedTimeline`] contains no data nor metadata and it is in
2538 : /// a state that will fail [`TenantShard::load_remote_timeline`] because `disk_consistent_lsn=Lsn(0)`.
2539 : ///
2540 : /// The caller is responsible for getting the timeline into a state that will be accepted
2541 : /// by [`TenantShard::load_remote_timeline`] / [`TenantShard::attach`].
2542 : /// Then they may call [`UninitializedTimeline::finish_creation`] to add the timeline
2543 : /// to the [`TenantShard::timelines`].
2544 : ///
2545 : /// Tests should use `TenantShard::create_test_timeline` to set up the minimum required metadata keys.
2546 115 : pub(crate) async fn create_empty_timeline(
2547 115 : self: &Arc<Self>,
2548 115 : new_timeline_id: TimelineId,
2549 115 : initdb_lsn: Lsn,
2550 115 : pg_version: PgMajorVersion,
2551 115 : ctx: &RequestContext,
2552 115 : ) -> anyhow::Result<(UninitializedTimeline, RequestContext)> {
2553 115 : anyhow::ensure!(
2554 115 : self.is_active(),
2555 0 : "Cannot create empty timelines on inactive tenant"
2556 : );
2557 :
2558 : // Protect against concurrent attempts to use this TimelineId
2559 115 : let create_guard = match self
2560 115 : .start_creating_timeline(new_timeline_id, CreateTimelineIdempotency::FailWithConflict)
2561 115 : .await?
2562 : {
2563 114 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2564 : StartCreatingTimelineResult::Idempotent(_) => {
2565 0 : unreachable!("FailWithConflict implies we get an error instead")
2566 : }
2567 : };
2568 :
2569 114 : let new_metadata = TimelineMetadata::new(
2570 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2571 : // make it valid, before calling finish_creation()
2572 114 : Lsn(0),
2573 114 : None,
2574 114 : None,
2575 114 : Lsn(0),
2576 114 : initdb_lsn,
2577 114 : initdb_lsn,
2578 114 : pg_version,
2579 : );
2580 114 : self.prepare_new_timeline(
2581 114 : new_timeline_id,
2582 114 : &new_metadata,
2583 114 : create_guard,
2584 114 : initdb_lsn,
2585 114 : None,
2586 114 : None,
2587 114 : ctx,
2588 114 : )
2589 114 : .await
2590 115 : }
2591 :
2592 : /// Helper for unit tests to create an empty timeline.
2593 : ///
2594 : /// The timeline is has state value `Active` but its background loops are not running.
2595 : // This makes the various functions which anyhow::ensure! for Active state work in tests.
2596 : // Our current tests don't need the background loops.
2597 : #[cfg(test)]
2598 110 : pub async fn create_test_timeline(
2599 110 : self: &Arc<Self>,
2600 110 : new_timeline_id: TimelineId,
2601 110 : initdb_lsn: Lsn,
2602 110 : pg_version: PgMajorVersion,
2603 110 : ctx: &RequestContext,
2604 110 : ) -> anyhow::Result<Arc<Timeline>> {
2605 110 : let (uninit_tl, ctx) = self
2606 110 : .create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2607 110 : .await?;
2608 110 : let tline = uninit_tl.raw_timeline().expect("we just created it");
2609 110 : assert_eq!(tline.get_last_record_lsn(), Lsn(0));
2610 :
2611 : // Setup minimum keys required for the timeline to be usable.
2612 110 : let mut modification = tline.begin_modification(initdb_lsn);
2613 110 : modification
2614 110 : .init_empty_test_timeline()
2615 110 : .context("init_empty_test_timeline")?;
2616 110 : modification
2617 110 : .commit(&ctx)
2618 110 : .await
2619 110 : .context("commit init_empty_test_timeline modification")?;
2620 :
2621 : // Flush to disk so that uninit_tl's check for valid disk_consistent_lsn passes.
2622 110 : tline.maybe_spawn_flush_loop();
2623 110 : tline.freeze_and_flush().await.context("freeze_and_flush")?;
2624 :
2625 : // Make sure the freeze_and_flush reaches remote storage.
2626 110 : tline.remote_client.wait_completion().await.unwrap();
2627 :
2628 110 : let tl = uninit_tl.finish_creation().await?;
2629 : // The non-test code would call tl.activate() here.
2630 110 : tl.set_state(TimelineState::Active);
2631 110 : Ok(tl)
2632 110 : }
2633 :
2634 : /// Helper for unit tests to create a timeline with some pre-loaded states.
2635 : #[cfg(test)]
2636 : #[allow(clippy::too_many_arguments)]
2637 24 : pub async fn create_test_timeline_with_layers(
2638 24 : self: &Arc<Self>,
2639 24 : new_timeline_id: TimelineId,
2640 24 : initdb_lsn: Lsn,
2641 24 : pg_version: PgMajorVersion,
2642 24 : ctx: &RequestContext,
2643 24 : in_memory_layer_desc: Vec<timeline::InMemoryLayerTestDesc>,
2644 24 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
2645 24 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
2646 24 : end_lsn: Lsn,
2647 24 : ) -> anyhow::Result<Arc<Timeline>> {
2648 : use checks::check_valid_layermap;
2649 : use itertools::Itertools;
2650 :
2651 24 : let tline = self
2652 24 : .create_test_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2653 24 : .await?;
2654 24 : tline.force_advance_lsn(end_lsn);
2655 71 : for deltas in delta_layer_desc {
2656 47 : tline
2657 47 : .force_create_delta_layer(deltas, Some(initdb_lsn), ctx)
2658 47 : .await?;
2659 : }
2660 58 : for (lsn, images) in image_layer_desc {
2661 34 : tline
2662 34 : .force_create_image_layer(lsn, images, Some(initdb_lsn), ctx)
2663 34 : .await?;
2664 : }
2665 28 : for in_memory in in_memory_layer_desc {
2666 4 : tline
2667 4 : .force_create_in_memory_layer(in_memory, Some(initdb_lsn), ctx)
2668 4 : .await?;
2669 : }
2670 24 : let layer_names = tline
2671 24 : .layers
2672 24 : .read(LayerManagerLockHolder::Testing)
2673 24 : .await
2674 24 : .layer_map()
2675 24 : .unwrap()
2676 24 : .iter_historic_layers()
2677 105 : .map(|layer| layer.layer_name())
2678 24 : .collect_vec();
2679 24 : if let Some(err) = check_valid_layermap(&layer_names) {
2680 0 : bail!("invalid layermap: {err}");
2681 24 : }
2682 24 : Ok(tline)
2683 24 : }
2684 :
2685 : /// Create a new timeline.
2686 : ///
2687 : /// Returns the new timeline ID and reference to its Timeline object.
2688 : ///
2689 : /// If the caller specified the timeline ID to use (`new_timeline_id`), and timeline with
2690 : /// the same timeline ID already exists, returns CreateTimelineError::AlreadyExists.
2691 : #[allow(clippy::too_many_arguments)]
2692 0 : pub(crate) async fn create_timeline(
2693 0 : self: &Arc<TenantShard>,
2694 0 : params: CreateTimelineParams,
2695 0 : broker_client: storage_broker::BrokerClientChannel,
2696 0 : ctx: &RequestContext,
2697 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
2698 0 : if !self.is_active() {
2699 0 : if matches!(self.current_state(), TenantState::Stopping { .. }) {
2700 0 : return Err(CreateTimelineError::ShuttingDown);
2701 : } else {
2702 0 : return Err(CreateTimelineError::Other(anyhow::anyhow!(
2703 0 : "Cannot create timelines on inactive tenant"
2704 0 : )));
2705 : }
2706 0 : }
2707 :
2708 0 : let _gate = self
2709 0 : .gate
2710 0 : .enter()
2711 0 : .map_err(|_| CreateTimelineError::ShuttingDown)?;
2712 :
2713 0 : let result: CreateTimelineResult = match params {
2714 : CreateTimelineParams::Bootstrap(CreateTimelineParamsBootstrap {
2715 0 : new_timeline_id,
2716 0 : existing_initdb_timeline_id,
2717 0 : pg_version,
2718 : }) => {
2719 0 : self.bootstrap_timeline(
2720 0 : new_timeline_id,
2721 0 : pg_version,
2722 0 : existing_initdb_timeline_id,
2723 0 : ctx,
2724 0 : )
2725 0 : .await?
2726 : }
2727 : CreateTimelineParams::Branch(CreateTimelineParamsBranch {
2728 0 : new_timeline_id,
2729 0 : ancestor_timeline_id,
2730 0 : mut ancestor_start_lsn,
2731 : }) => {
2732 0 : let ancestor_timeline = self
2733 0 : .get_timeline(ancestor_timeline_id, false)
2734 0 : .context("Cannot branch off the timeline that's not present in pageserver")?;
2735 :
2736 : // instead of waiting around, just deny the request because ancestor is not yet
2737 : // ready for other purposes either.
2738 0 : if !ancestor_timeline.is_active() {
2739 0 : return Err(CreateTimelineError::AncestorNotActive);
2740 0 : }
2741 :
2742 0 : if ancestor_timeline.is_archived() == Some(true) {
2743 0 : info!("tried to branch archived timeline");
2744 0 : return Err(CreateTimelineError::AncestorArchived);
2745 0 : }
2746 :
2747 0 : if let Some(lsn) = ancestor_start_lsn.as_mut() {
2748 0 : *lsn = lsn.align();
2749 :
2750 0 : let ancestor_ancestor_lsn = ancestor_timeline.get_ancestor_lsn();
2751 0 : if ancestor_ancestor_lsn > *lsn {
2752 : // can we safely just branch from the ancestor instead?
2753 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
2754 0 : "invalid start lsn {} for ancestor timeline {}: less than timeline ancestor lsn {}",
2755 0 : lsn,
2756 0 : ancestor_timeline_id,
2757 0 : ancestor_ancestor_lsn,
2758 0 : )));
2759 0 : }
2760 :
2761 : // Wait for the WAL to arrive and be processed on the parent branch up
2762 : // to the requested branch point. The repository code itself doesn't
2763 : // require it, but if we start to receive WAL on the new timeline,
2764 : // decoding the new WAL might need to look up previous pages, relation
2765 : // sizes etc. and that would get confused if the previous page versions
2766 : // are not in the repository yet.
2767 0 : ancestor_timeline
2768 0 : .wait_lsn(
2769 0 : *lsn,
2770 0 : timeline::WaitLsnWaiter::Tenant,
2771 0 : timeline::WaitLsnTimeout::Default,
2772 0 : ctx,
2773 0 : )
2774 0 : .await
2775 0 : .map_err(|e| match e {
2776 0 : e @ (WaitLsnError::Timeout(_) | WaitLsnError::BadState { .. }) => {
2777 0 : CreateTimelineError::AncestorLsn(anyhow::anyhow!(e))
2778 : }
2779 0 : WaitLsnError::Shutdown => CreateTimelineError::ShuttingDown,
2780 0 : })?;
2781 0 : }
2782 :
2783 0 : self.branch_timeline(&ancestor_timeline, new_timeline_id, ancestor_start_lsn, ctx)
2784 0 : .await?
2785 : }
2786 0 : CreateTimelineParams::ImportPgdata(params) => {
2787 0 : self.create_timeline_import_pgdata(params, ctx).await?
2788 : }
2789 : };
2790 :
2791 : // At this point we have dropped our guard on [`Self::timelines_creating`], and
2792 : // the timeline is visible in [`Self::timelines`], but it is _not_ durable yet. We must
2793 : // not send a success to the caller until it is. The same applies to idempotent retries.
2794 : //
2795 : // TODO: the timeline is already visible in [`Self::timelines`]; a caller could incorrectly
2796 : // assume that, because they can see the timeline via API, that the creation is done and
2797 : // that it is durable. Ideally, we would keep the timeline hidden (in [`Self::timelines_creating`])
2798 : // until it is durable, e.g., by extending the time we hold the creation guard. This also
2799 : // interacts with UninitializedTimeline and is generally a bit tricky.
2800 : //
2801 : // To re-emphasize: the only correct way to create a timeline is to repeat calling the
2802 : // creation API until it returns success. Only then is durability guaranteed.
2803 0 : info!(creation_result=%result.discriminant(), "waiting for timeline to be durable");
2804 0 : result
2805 0 : .timeline()
2806 0 : .remote_client
2807 0 : .wait_completion()
2808 0 : .await
2809 0 : .map_err(|e| match e {
2810 : WaitCompletionError::NotInitialized(
2811 0 : e, // If the queue is already stopped, it's a shutdown error.
2812 0 : ) if e.is_stopping() => CreateTimelineError::ShuttingDown,
2813 : WaitCompletionError::NotInitialized(_) => {
2814 : // This is a bug: we should never try to wait for uploads before initializing the timeline
2815 0 : debug_assert!(false);
2816 0 : CreateTimelineError::Other(anyhow::anyhow!("timeline not initialized"))
2817 : }
2818 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2819 0 : CreateTimelineError::ShuttingDown
2820 : }
2821 0 : })?;
2822 :
2823 : // The creating task is responsible for activating the timeline.
2824 : // We do this after `wait_completion()` so that we don't spin up tasks that start
2825 : // doing stuff before the IndexPart is durable in S3, which is done by the previous section.
2826 0 : let activated_timeline = match result {
2827 0 : CreateTimelineResult::Created(timeline) => {
2828 0 : timeline.activate(
2829 0 : self.clone(),
2830 0 : broker_client,
2831 0 : None,
2832 0 : &ctx.with_scope_timeline(&timeline),
2833 : );
2834 0 : timeline
2835 : }
2836 0 : CreateTimelineResult::Idempotent(timeline) => {
2837 0 : info!(
2838 0 : "request was deemed idempotent, activation will be done by the creating task"
2839 : );
2840 0 : timeline
2841 : }
2842 0 : CreateTimelineResult::ImportSpawned(timeline) => {
2843 0 : info!(
2844 0 : "import task spawned, timeline will become visible and activated once the import is done"
2845 : );
2846 0 : timeline
2847 : }
2848 : };
2849 :
2850 0 : Ok(activated_timeline)
2851 0 : }
2852 :
2853 : /// The returned [`Arc<Timeline>`] is NOT in the [`TenantShard::timelines`] map until the import
2854 : /// completes in the background. A DIFFERENT [`Arc<Timeline>`] will be inserted into the
2855 : /// [`TenantShard::timelines`] map when the import completes.
2856 : /// We only return an [`Arc<Timeline>`] here so the API handler can create a [`pageserver_api::models::TimelineInfo`]
2857 : /// for the response.
2858 0 : async fn create_timeline_import_pgdata(
2859 0 : self: &Arc<Self>,
2860 0 : params: CreateTimelineParamsImportPgdata,
2861 0 : ctx: &RequestContext,
2862 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
2863 : let CreateTimelineParamsImportPgdata {
2864 0 : new_timeline_id,
2865 0 : location,
2866 0 : idempotency_key,
2867 0 : } = params;
2868 :
2869 0 : let started_at = chrono::Utc::now().naive_utc();
2870 :
2871 : //
2872 : // There's probably a simpler way to upload an index part, but, remote_timeline_client
2873 : // is the canonical way we do it.
2874 : // - create an empty timeline in-memory
2875 : // - use its remote_timeline_client to do the upload
2876 : // - dispose of the uninit timeline
2877 : // - keep the creation guard alive
2878 :
2879 0 : let timeline_create_guard = match self
2880 0 : .start_creating_timeline(
2881 0 : new_timeline_id,
2882 0 : CreateTimelineIdempotency::ImportPgdata(CreatingTimelineIdempotencyImportPgdata {
2883 0 : idempotency_key: idempotency_key.clone(),
2884 0 : }),
2885 0 : )
2886 0 : .await?
2887 : {
2888 0 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2889 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
2890 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
2891 : }
2892 : };
2893 :
2894 0 : let (mut uninit_timeline, timeline_ctx) = {
2895 0 : let this = &self;
2896 0 : let initdb_lsn = Lsn(0);
2897 0 : async move {
2898 0 : let new_metadata = TimelineMetadata::new(
2899 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2900 : // make it valid, before calling finish_creation()
2901 0 : Lsn(0),
2902 0 : None,
2903 0 : None,
2904 0 : Lsn(0),
2905 0 : initdb_lsn,
2906 0 : initdb_lsn,
2907 0 : PgMajorVersion::PG15,
2908 : );
2909 0 : this.prepare_new_timeline(
2910 0 : new_timeline_id,
2911 0 : &new_metadata,
2912 0 : timeline_create_guard,
2913 0 : initdb_lsn,
2914 0 : None,
2915 0 : None,
2916 0 : ctx,
2917 0 : )
2918 0 : .await
2919 0 : }
2920 : }
2921 0 : .await?;
2922 :
2923 0 : let in_progress = import_pgdata::index_part_format::InProgress {
2924 0 : idempotency_key,
2925 0 : location,
2926 0 : started_at,
2927 0 : };
2928 0 : let index_part = import_pgdata::index_part_format::Root::V1(
2929 0 : import_pgdata::index_part_format::V1::InProgress(in_progress),
2930 0 : );
2931 0 : uninit_timeline
2932 0 : .raw_timeline()
2933 0 : .unwrap()
2934 0 : .remote_client
2935 0 : .schedule_index_upload_for_import_pgdata_state_update(Some(index_part.clone()))?;
2936 :
2937 : // wait_completion happens in caller
2938 :
2939 0 : let (timeline, timeline_create_guard) = uninit_timeline.finish_creation_myself();
2940 :
2941 0 : let import_task_gate = Gate::default();
2942 0 : let import_task_guard = import_task_gate.enter().unwrap();
2943 :
2944 0 : let import_task_handle = tokio::spawn(self.clone().create_timeline_import_pgdata_task(
2945 0 : timeline.clone(),
2946 0 : index_part,
2947 0 : timeline_create_guard,
2948 0 : import_task_guard,
2949 0 : timeline_ctx.detached_child(TaskKind::ImportPgdata, DownloadBehavior::Warn),
2950 : ));
2951 :
2952 0 : let prev = self.timelines_importing.lock().unwrap().insert(
2953 0 : timeline.timeline_id,
2954 0 : Arc::new(ImportingTimeline {
2955 0 : timeline: timeline.clone(),
2956 0 : import_task_handle,
2957 0 : import_task_gate,
2958 0 : delete_progress: TimelineDeleteProgress::default(),
2959 0 : }),
2960 0 : );
2961 :
2962 : // Idempotency is enforced higher up the stack
2963 0 : assert!(prev.is_none());
2964 :
2965 : // NB: the timeline doesn't exist in self.timelines at this point
2966 0 : Ok(CreateTimelineResult::ImportSpawned(timeline))
2967 0 : }
2968 :
2969 : /// Finalize the import of a timeline on this shard by marking it complete in
2970 : /// the index part. If the import task hasn't finished yet, returns an error.
2971 : ///
2972 : /// This method is idempotent. If the import was finalized once, the next call
2973 : /// will be a no-op.
2974 0 : pub(crate) async fn finalize_importing_timeline(
2975 0 : &self,
2976 0 : timeline_id: TimelineId,
2977 0 : ) -> Result<(), FinalizeTimelineImportError> {
2978 0 : let timeline = {
2979 0 : let locked = self.timelines_importing.lock().unwrap();
2980 0 : match locked.get(&timeline_id) {
2981 0 : Some(importing_timeline) => {
2982 0 : if !importing_timeline.import_task_handle.is_finished() {
2983 0 : return Err(FinalizeTimelineImportError::ImportTaskStillRunning);
2984 0 : }
2985 :
2986 0 : importing_timeline.timeline.clone()
2987 : }
2988 : None => {
2989 0 : return Ok(());
2990 : }
2991 : }
2992 : };
2993 :
2994 0 : timeline
2995 0 : .remote_client
2996 0 : .schedule_index_upload_for_import_pgdata_finalize()
2997 0 : .map_err(|_err| FinalizeTimelineImportError::ShuttingDown)?;
2998 0 : timeline
2999 0 : .remote_client
3000 0 : .wait_completion()
3001 0 : .await
3002 0 : .map_err(|_err| FinalizeTimelineImportError::ShuttingDown)?;
3003 :
3004 0 : self.timelines_importing
3005 0 : .lock()
3006 0 : .unwrap()
3007 0 : .remove(&timeline_id);
3008 :
3009 0 : Ok(())
3010 0 : }
3011 :
3012 : #[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))]
3013 : async fn create_timeline_import_pgdata_task(
3014 : self: Arc<TenantShard>,
3015 : timeline: Arc<Timeline>,
3016 : index_part: import_pgdata::index_part_format::Root,
3017 : timeline_create_guard: TimelineCreateGuard,
3018 : _import_task_guard: GateGuard,
3019 : ctx: RequestContext,
3020 : ) {
3021 : debug_assert_current_span_has_tenant_and_timeline_id();
3022 : info!("starting");
3023 : scopeguard::defer! {info!("exiting")};
3024 :
3025 : let res = self
3026 : .create_timeline_import_pgdata_task_impl(
3027 : timeline,
3028 : index_part,
3029 : timeline_create_guard,
3030 : ctx,
3031 : )
3032 : .await;
3033 : if let Err(err) = &res {
3034 : error!(?err, "task failed");
3035 : // TODO sleep & retry, sensitive to tenant shutdown
3036 : // TODO: allow timeline deletion requests => should cancel the task
3037 : }
3038 : }
3039 :
3040 0 : async fn create_timeline_import_pgdata_task_impl(
3041 0 : self: Arc<TenantShard>,
3042 0 : timeline: Arc<Timeline>,
3043 0 : index_part: import_pgdata::index_part_format::Root,
3044 0 : _timeline_create_guard: TimelineCreateGuard,
3045 0 : ctx: RequestContext,
3046 0 : ) -> Result<(), anyhow::Error> {
3047 0 : info!("importing pgdata");
3048 0 : let ctx = ctx.with_scope_timeline(&timeline);
3049 0 : import_pgdata::doit(&timeline, index_part, &ctx, self.cancel.clone())
3050 0 : .await
3051 0 : .context("import")?;
3052 0 : info!("import done - waiting for activation");
3053 :
3054 0 : anyhow::Ok(())
3055 0 : }
3056 :
3057 0 : pub(crate) async fn delete_timeline(
3058 0 : self: Arc<Self>,
3059 0 : timeline_id: TimelineId,
3060 0 : ) -> Result<(), DeleteTimelineError> {
3061 0 : DeleteTimelineFlow::run(&self, timeline_id).await?;
3062 :
3063 0 : Ok(())
3064 0 : }
3065 :
3066 : /// perform one garbage collection iteration, removing old data files from disk.
3067 : /// this function is periodically called by gc task.
3068 : /// also it can be explicitly requested through page server api 'do_gc' command.
3069 : ///
3070 : /// `target_timeline_id` specifies the timeline to GC, or None for all.
3071 : ///
3072 : /// The `horizon` an `pitr` parameters determine how much WAL history needs to be retained.
3073 : /// Also known as the retention period, or the GC cutoff point. `horizon` specifies
3074 : /// the amount of history, as LSN difference from current latest LSN on each timeline.
3075 : /// `pitr` specifies the same as a time difference from the current time. The effective
3076 : /// GC cutoff point is determined conservatively by either `horizon` and `pitr`, whichever
3077 : /// requires more history to be retained.
3078 : //
3079 377 : pub(crate) async fn gc_iteration(
3080 377 : &self,
3081 377 : target_timeline_id: Option<TimelineId>,
3082 377 : horizon: u64,
3083 377 : pitr: Duration,
3084 377 : cancel: &CancellationToken,
3085 377 : ctx: &RequestContext,
3086 377 : ) -> Result<GcResult, GcError> {
3087 : // Don't start doing work during shutdown
3088 377 : if let TenantState::Stopping { .. } = self.current_state() {
3089 0 : return Ok(GcResult::default());
3090 377 : }
3091 :
3092 : // there is a global allowed_error for this
3093 377 : if !self.is_active() {
3094 0 : return Err(GcError::NotActive);
3095 377 : }
3096 :
3097 : {
3098 377 : let conf = self.tenant_conf.load();
3099 :
3100 : // If we may not delete layers, then simply skip GC. Even though a tenant
3101 : // in AttachedMulti state could do GC and just enqueue the blocked deletions,
3102 : // the only advantage to doing it is to perhaps shrink the LayerMap metadata
3103 : // a bit sooner than we would achieve by waiting for AttachedSingle status.
3104 377 : if !conf.location.may_delete_layers_hint() {
3105 0 : info!("Skipping GC in location state {:?}", conf.location);
3106 0 : return Ok(GcResult::default());
3107 377 : }
3108 :
3109 377 : if conf.is_gc_blocked_by_lsn_lease_deadline() {
3110 0 : info!("Skipping GC because lsn lease deadline is not reached");
3111 0 : return Ok(GcResult::default());
3112 377 : }
3113 : }
3114 :
3115 377 : let _guard = match self.gc_block.start().await {
3116 377 : Ok(guard) => guard,
3117 0 : Err(reasons) => {
3118 0 : info!("Skipping GC: {reasons}");
3119 0 : return Ok(GcResult::default());
3120 : }
3121 : };
3122 :
3123 377 : self.gc_iteration_internal(target_timeline_id, horizon, pitr, cancel, ctx)
3124 377 : .await
3125 377 : }
3126 :
3127 : /// Performs one compaction iteration. Called periodically from the compaction loop. Returns
3128 : /// whether another compaction is needed, if we still have pending work or if we yield for
3129 : /// immediate L0 compaction.
3130 : ///
3131 : /// Compaction can also be explicitly requested for a timeline via the HTTP API.
3132 0 : async fn compaction_iteration(
3133 0 : self: &Arc<Self>,
3134 0 : cancel: &CancellationToken,
3135 0 : ctx: &RequestContext,
3136 0 : ) -> Result<CompactionOutcome, CompactionError> {
3137 : // Don't compact inactive tenants.
3138 0 : if !self.is_active() {
3139 0 : return Ok(CompactionOutcome::Skipped);
3140 0 : }
3141 :
3142 : // Don't compact tenants that can't upload layers. We don't check `may_delete_layers_hint`,
3143 : // since we need to compact L0 even in AttachedMulti to bound read amplification.
3144 0 : let location = self.tenant_conf.load().location;
3145 0 : if !location.may_upload_layers_hint() {
3146 0 : info!("skipping compaction in location state {location:?}");
3147 0 : return Ok(CompactionOutcome::Skipped);
3148 0 : }
3149 :
3150 : // Don't compact if the circuit breaker is tripped.
3151 0 : if self.compaction_circuit_breaker.lock().unwrap().is_broken() {
3152 0 : info!("skipping compaction due to previous failures");
3153 0 : return Ok(CompactionOutcome::Skipped);
3154 0 : }
3155 :
3156 : // Collect all timelines to compact, along with offload instructions and L0 counts.
3157 0 : let mut compact: Vec<Arc<Timeline>> = Vec::new();
3158 0 : let mut offload: HashSet<TimelineId> = HashSet::new();
3159 0 : let mut l0_counts: HashMap<TimelineId, usize> = HashMap::new();
3160 :
3161 : {
3162 0 : let offload_enabled = self.get_timeline_offloading_enabled();
3163 0 : let timelines = self.timelines.lock().unwrap();
3164 0 : for (&timeline_id, timeline) in timelines.iter() {
3165 : // Skip inactive timelines.
3166 0 : if !timeline.is_active() {
3167 0 : continue;
3168 0 : }
3169 :
3170 : // Schedule the timeline for compaction.
3171 0 : compact.push(timeline.clone());
3172 :
3173 : // Schedule the timeline for offloading if eligible.
3174 0 : let can_offload = offload_enabled
3175 0 : && timeline.can_offload().0
3176 0 : && !timelines
3177 0 : .iter()
3178 0 : .any(|(_, tli)| tli.get_ancestor_timeline_id() == Some(timeline_id));
3179 0 : if can_offload {
3180 0 : offload.insert(timeline_id);
3181 0 : }
3182 : }
3183 : } // release timelines lock
3184 :
3185 0 : for timeline in &compact {
3186 : // Collect L0 counts. Can't await while holding lock above.
3187 0 : if let Ok(lm) = timeline
3188 0 : .layers
3189 0 : .read(LayerManagerLockHolder::Compaction)
3190 0 : .await
3191 0 : .layer_map()
3192 0 : {
3193 0 : l0_counts.insert(timeline.timeline_id, lm.level0_deltas().len());
3194 0 : }
3195 : }
3196 :
3197 : // Pass 1: L0 compaction across all timelines, in order of L0 count. We prioritize this to
3198 : // bound read amplification.
3199 : //
3200 : // TODO: this may spin on one or more ingest-heavy timelines, starving out image/GC
3201 : // compaction and offloading. We leave that as a potential problem to solve later. Consider
3202 : // splitting L0 and image/GC compaction to separate background jobs.
3203 0 : if self.get_compaction_l0_first() {
3204 0 : let compaction_threshold = self.get_compaction_threshold();
3205 0 : let compact_l0 = compact
3206 0 : .iter()
3207 0 : .map(|tli| (tli, l0_counts.get(&tli.timeline_id).copied().unwrap_or(0)))
3208 0 : .filter(|&(_, l0)| l0 >= compaction_threshold)
3209 0 : .sorted_by_key(|&(_, l0)| l0)
3210 0 : .rev()
3211 0 : .map(|(tli, _)| tli.clone())
3212 0 : .collect_vec();
3213 :
3214 0 : let mut has_pending_l0 = false;
3215 0 : for timeline in compact_l0 {
3216 0 : let ctx = &ctx.with_scope_timeline(&timeline);
3217 : // NB: don't set CompactFlags::YieldForL0, since this is an L0-only compaction pass.
3218 0 : let outcome = timeline
3219 0 : .compact(cancel, CompactFlags::OnlyL0Compaction.into(), ctx)
3220 0 : .instrument(info_span!("compact_timeline", timeline_id = %timeline.timeline_id))
3221 0 : .await
3222 0 : .inspect_err(|err| self.maybe_trip_compaction_breaker(err))?;
3223 0 : match outcome {
3224 0 : CompactionOutcome::Done => {}
3225 0 : CompactionOutcome::Skipped => {}
3226 0 : CompactionOutcome::Pending => has_pending_l0 = true,
3227 0 : CompactionOutcome::YieldForL0 => has_pending_l0 = true,
3228 : }
3229 : }
3230 0 : if has_pending_l0 {
3231 0 : return Ok(CompactionOutcome::YieldForL0); // do another pass
3232 0 : }
3233 0 : }
3234 :
3235 : // Pass 2: image compaction and timeline offloading. If any timelines have accumulated more
3236 : // L0 layers, they may also be compacted here. Image compaction will yield if there is
3237 : // pending L0 compaction on any tenant timeline.
3238 : //
3239 : // TODO: consider ordering timelines by some priority, e.g. time since last full compaction,
3240 : // amount of L1 delta debt or garbage, offload-eligible timelines first, etc.
3241 0 : let mut has_pending = false;
3242 0 : for timeline in compact {
3243 0 : if !timeline.is_active() {
3244 0 : continue;
3245 0 : }
3246 0 : let ctx = &ctx.with_scope_timeline(&timeline);
3247 :
3248 : // Yield for L0 if the separate L0 pass is enabled (otherwise there's no point).
3249 0 : let mut flags = EnumSet::default();
3250 0 : if self.get_compaction_l0_first() {
3251 0 : flags |= CompactFlags::YieldForL0;
3252 0 : }
3253 :
3254 0 : let mut outcome = timeline
3255 0 : .compact(cancel, flags, ctx)
3256 0 : .instrument(info_span!("compact_timeline", timeline_id = %timeline.timeline_id))
3257 0 : .await
3258 0 : .inspect_err(|err| self.maybe_trip_compaction_breaker(err))?;
3259 :
3260 : // If we're done compacting, check the scheduled GC compaction queue for more work.
3261 0 : if outcome == CompactionOutcome::Done {
3262 0 : let queue = {
3263 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3264 0 : guard
3265 0 : .entry(timeline.timeline_id)
3266 0 : .or_insert_with(|| Arc::new(GcCompactionQueue::new()))
3267 0 : .clone()
3268 : };
3269 0 : let gc_compaction_strategy = self
3270 0 : .feature_resolver
3271 0 : .evaluate_multivariate("gc-comapction-strategy")
3272 0 : .ok();
3273 0 : let span = if let Some(gc_compaction_strategy) = gc_compaction_strategy {
3274 0 : info_span!("gc_compact_timeline", timeline_id = %timeline.timeline_id, strategy = %gc_compaction_strategy)
3275 : } else {
3276 0 : info_span!("gc_compact_timeline", timeline_id = %timeline.timeline_id)
3277 : };
3278 0 : outcome = queue
3279 0 : .iteration(cancel, ctx, &self.gc_block, &timeline)
3280 0 : .instrument(span)
3281 0 : .await?;
3282 0 : }
3283 :
3284 : // If we're done compacting, offload the timeline if requested.
3285 0 : if outcome == CompactionOutcome::Done && offload.contains(&timeline.timeline_id) {
3286 0 : pausable_failpoint!("before-timeline-auto-offload");
3287 0 : offload_timeline(self, &timeline)
3288 0 : .instrument(info_span!("offload_timeline", timeline_id = %timeline.timeline_id))
3289 0 : .await
3290 0 : .or_else(|err| match err {
3291 : // Ignore this, we likely raced with unarchival.
3292 0 : OffloadError::NotArchived => Ok(()),
3293 0 : OffloadError::AlreadyInProgress => Ok(()),
3294 0 : OffloadError::Cancelled => Err(CompactionError::new_cancelled()),
3295 : // don't break the anyhow chain
3296 0 : OffloadError::Other(err) => Err(CompactionError::Other(err)),
3297 0 : })?;
3298 0 : }
3299 :
3300 0 : match outcome {
3301 0 : CompactionOutcome::Done => {}
3302 0 : CompactionOutcome::Skipped => {}
3303 0 : CompactionOutcome::Pending => has_pending = true,
3304 : // This mostly makes sense when the L0-only pass above is enabled, since there's
3305 : // otherwise no guarantee that we'll start with the timeline that has high L0.
3306 0 : CompactionOutcome::YieldForL0 => return Ok(CompactionOutcome::YieldForL0),
3307 : }
3308 : }
3309 :
3310 : // Success! Untrip the breaker if necessary.
3311 0 : self.compaction_circuit_breaker
3312 0 : .lock()
3313 0 : .unwrap()
3314 0 : .success(&CIRCUIT_BREAKERS_UNBROKEN);
3315 :
3316 0 : match has_pending {
3317 0 : true => Ok(CompactionOutcome::Pending),
3318 0 : false => Ok(CompactionOutcome::Done),
3319 : }
3320 0 : }
3321 :
3322 : /// Trips the compaction circuit breaker if appropriate.
3323 0 : pub(crate) fn maybe_trip_compaction_breaker(&self, err: &CompactionError) {
3324 0 : if err.is_cancel() {
3325 0 : return;
3326 0 : }
3327 0 : self.compaction_circuit_breaker
3328 0 : .lock()
3329 0 : .unwrap()
3330 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, err);
3331 0 : }
3332 :
3333 : /// Cancel scheduled compaction tasks
3334 0 : pub(crate) fn cancel_scheduled_compaction(&self, timeline_id: TimelineId) {
3335 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3336 0 : if let Some(q) = guard.get_mut(&timeline_id) {
3337 0 : q.cancel_scheduled();
3338 0 : }
3339 0 : }
3340 :
3341 0 : pub(crate) fn get_scheduled_compaction_tasks(
3342 0 : &self,
3343 0 : timeline_id: TimelineId,
3344 0 : ) -> Vec<CompactInfoResponse> {
3345 0 : let res = {
3346 0 : let guard = self.scheduled_compaction_tasks.lock().unwrap();
3347 0 : guard.get(&timeline_id).map(|q| q.remaining_jobs())
3348 : };
3349 0 : let Some((running, remaining)) = res else {
3350 0 : return Vec::new();
3351 : };
3352 0 : let mut result = Vec::new();
3353 0 : if let Some((id, running)) = running {
3354 0 : result.extend(running.into_compact_info_resp(id, true));
3355 0 : }
3356 0 : for (id, job) in remaining {
3357 0 : result.extend(job.into_compact_info_resp(id, false));
3358 0 : }
3359 0 : result
3360 0 : }
3361 :
3362 : /// Schedule a compaction task for a timeline.
3363 0 : pub(crate) async fn schedule_compaction(
3364 0 : &self,
3365 0 : timeline_id: TimelineId,
3366 0 : options: CompactOptions,
3367 0 : ) -> anyhow::Result<tokio::sync::oneshot::Receiver<()>> {
3368 0 : let (tx, rx) = tokio::sync::oneshot::channel();
3369 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3370 0 : let q = guard
3371 0 : .entry(timeline_id)
3372 0 : .or_insert_with(|| Arc::new(GcCompactionQueue::new()));
3373 0 : q.schedule_manual_compaction(options, Some(tx));
3374 0 : Ok(rx)
3375 0 : }
3376 :
3377 : /// Performs periodic housekeeping, via the tenant housekeeping background task.
3378 0 : async fn housekeeping(&self) {
3379 : // Call through to all timelines to freeze ephemeral layers as needed. This usually happens
3380 : // during ingest, but we don't want idle timelines to hold open layers for too long.
3381 : //
3382 : // We don't do this if the tenant can't upload layers (i.e. it's in stale attachment mode).
3383 : // We don't run compaction in this case either, and don't want to keep flushing tiny L0
3384 : // layers that won't be compacted down.
3385 0 : if self.tenant_conf.load().location.may_upload_layers_hint() {
3386 0 : let timelines = self
3387 0 : .timelines
3388 0 : .lock()
3389 0 : .unwrap()
3390 0 : .values()
3391 0 : .filter(|tli| tli.is_active())
3392 0 : .cloned()
3393 0 : .collect_vec();
3394 :
3395 0 : for timeline in timelines {
3396 : // Include a span with the timeline ID. The parent span already has the tenant ID.
3397 0 : let span =
3398 0 : info_span!("maybe_freeze_ephemeral_layer", timeline_id = %timeline.timeline_id);
3399 0 : timeline
3400 0 : .maybe_freeze_ephemeral_layer()
3401 0 : .instrument(span)
3402 0 : .await;
3403 : }
3404 0 : }
3405 :
3406 : // Shut down walredo if idle.
3407 : const WALREDO_IDLE_TIMEOUT: Duration = Duration::from_secs(180);
3408 0 : if let Some(ref walredo_mgr) = self.walredo_mgr {
3409 0 : walredo_mgr.maybe_quiesce(WALREDO_IDLE_TIMEOUT);
3410 0 : }
3411 :
3412 : // Update the feature resolver with the latest tenant-spcific data.
3413 0 : self.feature_resolver.refresh_properties_and_flags(self);
3414 0 : }
3415 :
3416 0 : pub fn timeline_has_no_attached_children(&self, timeline_id: TimelineId) -> bool {
3417 0 : let timelines = self.timelines.lock().unwrap();
3418 0 : !timelines
3419 0 : .iter()
3420 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(timeline_id))
3421 0 : }
3422 :
3423 1371 : pub fn current_state(&self) -> TenantState {
3424 1371 : self.state.borrow().clone()
3425 1371 : }
3426 :
3427 990 : pub fn is_active(&self) -> bool {
3428 990 : self.current_state() == TenantState::Active
3429 990 : }
3430 :
3431 0 : pub fn generation(&self) -> Generation {
3432 0 : self.generation
3433 0 : }
3434 :
3435 0 : pub(crate) fn wal_redo_manager_status(&self) -> Option<WalRedoManagerStatus> {
3436 0 : self.walredo_mgr.as_ref().and_then(|mgr| mgr.status())
3437 0 : }
3438 :
3439 : /// Changes tenant status to active, unless shutdown was already requested.
3440 : ///
3441 : /// `background_jobs_can_start` is an optional barrier set to a value during pageserver startup
3442 : /// to delay background jobs. Background jobs can be started right away when None is given.
3443 0 : fn activate(
3444 0 : self: &Arc<Self>,
3445 0 : broker_client: BrokerClientChannel,
3446 0 : background_jobs_can_start: Option<&completion::Barrier>,
3447 0 : ctx: &RequestContext,
3448 0 : ) {
3449 0 : span::debug_assert_current_span_has_tenant_id();
3450 :
3451 0 : let mut activating = false;
3452 0 : self.state.send_modify(|current_state| {
3453 : use pageserver_api::models::ActivatingFrom;
3454 0 : match &*current_state {
3455 : TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
3456 0 : panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {current_state:?}");
3457 : }
3458 0 : TenantState::Attaching => {
3459 0 : *current_state = TenantState::Activating(ActivatingFrom::Attaching);
3460 0 : }
3461 : }
3462 0 : debug!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), "Activating tenant");
3463 0 : activating = true;
3464 : // Continue outside the closure. We need to grab timelines.lock()
3465 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
3466 0 : });
3467 :
3468 0 : if activating {
3469 0 : let timelines_accessor = self.timelines.lock().unwrap();
3470 0 : let timelines_offloaded_accessor = self.timelines_offloaded.lock().unwrap();
3471 0 : let timelines_to_activate = timelines_accessor
3472 0 : .values()
3473 0 : .filter(|timeline| !(timeline.is_broken() || timeline.is_stopping()));
3474 :
3475 : // Spawn gc and compaction loops. The loops will shut themselves
3476 : // down when they notice that the tenant is inactive.
3477 0 : tasks::start_background_loops(self, background_jobs_can_start);
3478 :
3479 0 : let mut activated_timelines = 0;
3480 :
3481 0 : for timeline in timelines_to_activate {
3482 0 : timeline.activate(
3483 0 : self.clone(),
3484 0 : broker_client.clone(),
3485 0 : background_jobs_can_start,
3486 0 : &ctx.with_scope_timeline(timeline),
3487 0 : );
3488 0 : activated_timelines += 1;
3489 0 : }
3490 :
3491 0 : let tid = self.tenant_shard_id.tenant_id.to_string();
3492 0 : let shard_id = self.tenant_shard_id.shard_slug().to_string();
3493 0 : let offloaded_timeline_count = timelines_offloaded_accessor.len();
3494 0 : TENANT_OFFLOADED_TIMELINES
3495 0 : .with_label_values(&[&tid, &shard_id])
3496 0 : .set(offloaded_timeline_count as u64);
3497 :
3498 0 : self.state.send_modify(move |current_state| {
3499 0 : assert!(
3500 0 : matches!(current_state, TenantState::Activating(_)),
3501 0 : "set_stopping and set_broken wait for us to leave Activating state",
3502 : );
3503 0 : *current_state = TenantState::Active;
3504 :
3505 0 : let elapsed = self.constructed_at.elapsed();
3506 0 : let total_timelines = timelines_accessor.len();
3507 :
3508 : // log a lot of stuff, because some tenants sometimes suffer from user-visible
3509 : // times to activate. see https://github.com/neondatabase/neon/issues/4025
3510 0 : info!(
3511 0 : since_creation_millis = elapsed.as_millis(),
3512 0 : tenant_id = %self.tenant_shard_id.tenant_id,
3513 0 : shard_id = %self.tenant_shard_id.shard_slug(),
3514 : activated_timelines,
3515 : total_timelines,
3516 0 : post_state = <&'static str>::from(&*current_state),
3517 0 : "activation attempt finished"
3518 : );
3519 :
3520 0 : TENANT.activation.observe(elapsed.as_secs_f64());
3521 0 : });
3522 0 : }
3523 0 : }
3524 :
3525 : /// Shutdown the tenant and join all of the spawned tasks.
3526 : ///
3527 : /// The method caters for all use-cases:
3528 : /// - pageserver shutdown (freeze_and_flush == true)
3529 : /// - detach + ignore (freeze_and_flush == false)
3530 : ///
3531 : /// This will attempt to shutdown even if tenant is broken.
3532 : ///
3533 : /// `shutdown_progress` is a [`completion::Barrier`] for the shutdown initiated by this call.
3534 : /// If the tenant is already shutting down, we return a clone of the first shutdown call's
3535 : /// `Barrier` as an `Err`. This not-first caller can use the returned barrier to join with
3536 : /// the ongoing shutdown.
3537 3 : async fn shutdown(
3538 3 : &self,
3539 3 : shutdown_progress: completion::Barrier,
3540 3 : shutdown_mode: timeline::ShutdownMode,
3541 3 : ) -> Result<(), completion::Barrier> {
3542 3 : span::debug_assert_current_span_has_tenant_id();
3543 :
3544 : // Set tenant (and its timlines) to Stoppping state.
3545 : //
3546 : // Since we can only transition into Stopping state after activation is complete,
3547 : // run it in a JoinSet so all tenants have a chance to stop before we get SIGKILLed.
3548 : //
3549 : // Transitioning tenants to Stopping state has a couple of non-obvious side effects:
3550 : // 1. Lock out any new requests to the tenants.
3551 : // 2. Signal cancellation to WAL receivers (we wait on it below).
3552 : // 3. Signal cancellation for other tenant background loops.
3553 : // 4. ???
3554 : //
3555 : // The waiting for the cancellation is not done uniformly.
3556 : // We certainly wait for WAL receivers to shut down.
3557 : // That is necessary so that no new data comes in before the freeze_and_flush.
3558 : // But the tenant background loops are joined-on in our caller.
3559 : // It's mesed up.
3560 : // we just ignore the failure to stop
3561 :
3562 : // If we're still attaching, fire the cancellation token early to drop out: this
3563 : // will prevent us flushing, but ensures timely shutdown if some I/O during attach
3564 : // is very slow.
3565 3 : let shutdown_mode = if matches!(self.current_state(), TenantState::Attaching) {
3566 0 : self.cancel.cancel();
3567 :
3568 : // Having fired our cancellation token, do not try and flush timelines: their cancellation tokens
3569 : // are children of ours, so their flush loops will have shut down already
3570 0 : timeline::ShutdownMode::Hard
3571 : } else {
3572 3 : shutdown_mode
3573 : };
3574 :
3575 3 : match self.set_stopping(shutdown_progress).await {
3576 3 : Ok(()) => {}
3577 0 : Err(SetStoppingError::Broken) => {
3578 0 : // assume that this is acceptable
3579 0 : }
3580 0 : Err(SetStoppingError::AlreadyStopping(other)) => {
3581 : // give caller the option to wait for this this shutdown
3582 0 : info!("Tenant::shutdown: AlreadyStopping");
3583 0 : return Err(other);
3584 : }
3585 : };
3586 :
3587 3 : let mut js = tokio::task::JoinSet::new();
3588 : {
3589 3 : let timelines = self.timelines.lock().unwrap();
3590 3 : timelines.values().for_each(|timeline| {
3591 3 : let timeline = Arc::clone(timeline);
3592 3 : let timeline_id = timeline.timeline_id;
3593 3 : let span = tracing::info_span!("timeline_shutdown", %timeline_id, ?shutdown_mode);
3594 3 : js.spawn(async move { timeline.shutdown(shutdown_mode).instrument(span).await });
3595 3 : });
3596 : }
3597 : {
3598 3 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
3599 3 : timelines_offloaded.values().for_each(|timeline| {
3600 0 : timeline.defuse_for_tenant_drop();
3601 0 : });
3602 : }
3603 : {
3604 3 : let mut timelines_importing = self.timelines_importing.lock().unwrap();
3605 3 : timelines_importing
3606 3 : .drain()
3607 3 : .for_each(|(timeline_id, importing_timeline)| {
3608 0 : let span = tracing::info_span!("importing_timeline_shutdown", %timeline_id);
3609 0 : js.spawn(async move { importing_timeline.shutdown().instrument(span).await });
3610 0 : });
3611 : }
3612 : // test_long_timeline_create_then_tenant_delete is leaning on this message
3613 3 : tracing::info!("Waiting for timelines...");
3614 6 : while let Some(res) = js.join_next().await {
3615 0 : match res {
3616 3 : Ok(()) => {}
3617 0 : Err(je) if je.is_cancelled() => unreachable!("no cancelling used"),
3618 0 : Err(je) if je.is_panic() => { /* logged already */ }
3619 0 : Err(je) => warn!("unexpected JoinError: {je:?}"),
3620 : }
3621 : }
3622 :
3623 3 : if let ShutdownMode::Reload = shutdown_mode {
3624 0 : tracing::info!("Flushing deletion queue");
3625 0 : if let Err(e) = self.deletion_queue_client.flush().await {
3626 0 : match e {
3627 0 : DeletionQueueError::ShuttingDown => {
3628 0 : // This is the only error we expect for now. In the future, if more error
3629 0 : // variants are added, we should handle them here.
3630 0 : }
3631 : }
3632 0 : }
3633 3 : }
3634 :
3635 : // We cancel the Tenant's cancellation token _after_ the timelines have all shut down. This permits
3636 : // them to continue to do work during their shutdown methods, e.g. flushing data.
3637 3 : tracing::debug!("Cancelling CancellationToken");
3638 3 : self.cancel.cancel();
3639 :
3640 : // shutdown all tenant and timeline tasks: gc, compaction, page service
3641 : // No new tasks will be started for this tenant because it's in `Stopping` state.
3642 : //
3643 : // this will additionally shutdown and await all timeline tasks.
3644 3 : tracing::debug!("Waiting for tasks...");
3645 3 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), None).await;
3646 :
3647 3 : if let Some(walredo_mgr) = self.walredo_mgr.as_ref() {
3648 3 : walredo_mgr.shutdown().await;
3649 0 : }
3650 :
3651 : // Wait for any in-flight operations to complete
3652 3 : self.gate.close().await;
3653 :
3654 3 : remove_tenant_metrics(&self.tenant_shard_id);
3655 :
3656 3 : Ok(())
3657 3 : }
3658 :
3659 : /// Change tenant status to Stopping, to mark that it is being shut down.
3660 : ///
3661 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
3662 : ///
3663 : /// This function is not cancel-safe!
3664 3 : async fn set_stopping(&self, progress: completion::Barrier) -> Result<(), SetStoppingError> {
3665 3 : let mut rx = self.state.subscribe();
3666 :
3667 : // cannot stop before we're done activating, so wait out until we're done activating
3668 3 : rx.wait_for(|state| match state {
3669 : TenantState::Activating(_) | TenantState::Attaching => {
3670 0 : info!("waiting for {state} to turn Active|Broken|Stopping");
3671 0 : false
3672 : }
3673 3 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3674 3 : })
3675 3 : .await
3676 3 : .expect("cannot drop self.state while on a &self method");
3677 :
3678 : // we now know we're done activating, let's see whether this task is the winner to transition into Stopping
3679 3 : let mut err = None;
3680 3 : let stopping = self.state.send_if_modified(|current_state| match current_state {
3681 : TenantState::Activating(_) | TenantState::Attaching => {
3682 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
3683 : }
3684 : TenantState::Active => {
3685 : // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
3686 : // are created after the transition to Stopping. That's harmless, as the Timelines
3687 : // won't be accessible to anyone afterwards, because the Tenant is in Stopping state.
3688 3 : *current_state = TenantState::Stopping { progress: Some(progress) };
3689 : // Continue stopping outside the closure. We need to grab timelines.lock()
3690 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
3691 3 : true
3692 : }
3693 : TenantState::Stopping { progress: None } => {
3694 : // An attach was cancelled, and the attach transitioned the tenant from Attaching to
3695 : // Stopping(None) to let us know it exited. Register our progress and continue.
3696 0 : *current_state = TenantState::Stopping { progress: Some(progress) };
3697 0 : true
3698 : }
3699 0 : TenantState::Broken { reason, .. } => {
3700 0 : info!(
3701 0 : "Cannot set tenant to Stopping state, it is in Broken state due to: {reason}"
3702 : );
3703 0 : err = Some(SetStoppingError::Broken);
3704 0 : false
3705 : }
3706 0 : TenantState::Stopping { progress: Some(progress) } => {
3707 0 : info!("Tenant is already in Stopping state");
3708 0 : err = Some(SetStoppingError::AlreadyStopping(progress.clone()));
3709 0 : false
3710 : }
3711 3 : });
3712 3 : match (stopping, err) {
3713 3 : (true, None) => {} // continue
3714 0 : (false, Some(err)) => return Err(err),
3715 0 : (true, Some(_)) => unreachable!(
3716 : "send_if_modified closure must error out if not transitioning to Stopping"
3717 : ),
3718 0 : (false, None) => unreachable!(
3719 : "send_if_modified closure must return true if transitioning to Stopping"
3720 : ),
3721 : }
3722 :
3723 3 : let timelines_accessor = self.timelines.lock().unwrap();
3724 3 : let not_broken_timelines = timelines_accessor
3725 3 : .values()
3726 3 : .filter(|timeline| !timeline.is_broken());
3727 6 : for timeline in not_broken_timelines {
3728 3 : timeline.set_state(TimelineState::Stopping);
3729 3 : }
3730 3 : Ok(())
3731 3 : }
3732 :
3733 : /// Method for tenant::mgr to transition us into Broken state in case of a late failure in
3734 : /// `remove_tenant_from_memory`
3735 : ///
3736 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
3737 : ///
3738 : /// In tests, we also use this to set tenants to Broken state on purpose.
3739 0 : pub(crate) async fn set_broken(&self, reason: String) {
3740 0 : let mut rx = self.state.subscribe();
3741 :
3742 : // The load & attach routines own the tenant state until it has reached `Active`.
3743 : // So, wait until it's done.
3744 0 : rx.wait_for(|state| match state {
3745 : TenantState::Activating(_) | TenantState::Attaching => {
3746 0 : info!(
3747 0 : "waiting for {} to turn Active|Broken|Stopping",
3748 0 : <&'static str>::from(state)
3749 : );
3750 0 : false
3751 : }
3752 0 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3753 0 : })
3754 0 : .await
3755 0 : .expect("cannot drop self.state while on a &self method");
3756 :
3757 : // we now know we're done activating, let's see whether this task is the winner to transition into Broken
3758 0 : self.set_broken_no_wait(reason)
3759 0 : }
3760 :
3761 0 : pub(crate) fn set_broken_no_wait(&self, reason: impl Display) {
3762 0 : let reason = reason.to_string();
3763 0 : self.state.send_modify(|current_state| {
3764 0 : match *current_state {
3765 : TenantState::Activating(_) | TenantState::Attaching => {
3766 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
3767 : }
3768 : TenantState::Active => {
3769 0 : if cfg!(feature = "testing") {
3770 0 : warn!("Changing Active tenant to Broken state, reason: {}", reason);
3771 0 : *current_state = TenantState::broken_from_reason(reason);
3772 : } else {
3773 0 : unreachable!("not allowed to call set_broken on Active tenants in non-testing builds")
3774 : }
3775 : }
3776 : TenantState::Broken { .. } => {
3777 0 : warn!("Tenant is already in Broken state");
3778 : }
3779 : // This is the only "expected" path, any other path is a bug.
3780 : TenantState::Stopping { .. } => {
3781 0 : warn!(
3782 0 : "Marking Stopping tenant as Broken state, reason: {}",
3783 : reason
3784 : );
3785 0 : *current_state = TenantState::broken_from_reason(reason);
3786 : }
3787 : }
3788 0 : });
3789 0 : }
3790 :
3791 0 : pub fn subscribe_for_state_updates(&self) -> watch::Receiver<TenantState> {
3792 0 : self.state.subscribe()
3793 0 : }
3794 :
3795 : /// The activate_now semaphore is initialized with zero units. As soon as
3796 : /// we add a unit, waiters will be able to acquire a unit and proceed.
3797 0 : pub(crate) fn activate_now(&self) {
3798 0 : self.activate_now_sem.add_permits(1);
3799 0 : }
3800 :
3801 0 : pub(crate) async fn wait_to_become_active(
3802 0 : &self,
3803 0 : timeout: Duration,
3804 0 : ) -> Result<(), GetActiveTenantError> {
3805 0 : let mut receiver = self.state.subscribe();
3806 : loop {
3807 0 : let current_state = receiver.borrow_and_update().clone();
3808 0 : match current_state {
3809 : TenantState::Attaching | TenantState::Activating(_) => {
3810 : // in these states, there's a chance that we can reach ::Active
3811 0 : self.activate_now();
3812 0 : match timeout_cancellable(timeout, &self.cancel, receiver.changed()).await {
3813 0 : Ok(r) => {
3814 0 : r.map_err(
3815 : |_e: tokio::sync::watch::error::RecvError|
3816 : // Tenant existed but was dropped: report it as non-existent
3817 0 : GetActiveTenantError::NotFound(GetTenantError::ShardNotFound(self.tenant_shard_id))
3818 0 : )?
3819 : }
3820 : Err(TimeoutCancellableError::Cancelled) => {
3821 0 : return Err(GetActiveTenantError::Cancelled);
3822 : }
3823 : Err(TimeoutCancellableError::Timeout) => {
3824 0 : return Err(GetActiveTenantError::WaitForActiveTimeout {
3825 0 : latest_state: Some(self.current_state()),
3826 0 : wait_time: timeout,
3827 0 : });
3828 : }
3829 : }
3830 : }
3831 : TenantState::Active => {
3832 0 : return Ok(());
3833 : }
3834 0 : TenantState::Broken { reason, .. } => {
3835 : // This is fatal, and reported distinctly from the general case of "will never be active" because
3836 : // it's logically a 500 to external API users (broken is always a bug).
3837 0 : return Err(GetActiveTenantError::Broken(reason));
3838 : }
3839 : TenantState::Stopping { .. } => {
3840 : // There's no chance the tenant can transition back into ::Active
3841 0 : return Err(GetActiveTenantError::WillNotBecomeActive(current_state));
3842 : }
3843 : }
3844 : }
3845 0 : }
3846 :
3847 0 : pub(crate) fn get_attach_mode(&self) -> AttachmentMode {
3848 0 : self.tenant_conf.load().location.attach_mode
3849 0 : }
3850 :
3851 : /// For API access: generate a LocationConfig equivalent to the one that would be used to
3852 : /// create a Tenant in the same state. Do not use this in hot paths: it's for relatively
3853 : /// rare external API calls, like a reconciliation at startup.
3854 0 : pub(crate) fn get_location_conf(&self) -> models::LocationConfig {
3855 0 : let attached_tenant_conf = self.tenant_conf.load();
3856 :
3857 0 : let location_config_mode = match attached_tenant_conf.location.attach_mode {
3858 0 : AttachmentMode::Single => models::LocationConfigMode::AttachedSingle,
3859 0 : AttachmentMode::Multi => models::LocationConfigMode::AttachedMulti,
3860 0 : AttachmentMode::Stale => models::LocationConfigMode::AttachedStale,
3861 : };
3862 :
3863 0 : models::LocationConfig {
3864 0 : mode: location_config_mode,
3865 0 : generation: self.generation.into(),
3866 0 : secondary_conf: None,
3867 0 : shard_number: self.shard_identity.number.0,
3868 0 : shard_count: self.shard_identity.count.literal(),
3869 0 : shard_stripe_size: self.shard_identity.stripe_size.0,
3870 0 : tenant_conf: attached_tenant_conf.tenant_conf.clone(),
3871 0 : }
3872 0 : }
3873 :
3874 0 : pub(crate) fn get_tenant_shard_id(&self) -> &TenantShardId {
3875 0 : &self.tenant_shard_id
3876 0 : }
3877 :
3878 0 : pub(crate) fn get_shard_identity(&self) -> ShardIdentity {
3879 0 : self.shard_identity
3880 0 : }
3881 :
3882 120 : pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
3883 120 : self.shard_identity.stripe_size
3884 120 : }
3885 :
3886 0 : pub(crate) fn get_generation(&self) -> Generation {
3887 0 : self.generation
3888 0 : }
3889 :
3890 : /// This function partially shuts down the tenant (it shuts down the Timelines) and is fallible,
3891 : /// and can leave the tenant in a bad state if it fails. The caller is responsible for
3892 : /// resetting this tenant to a valid state if we fail.
3893 0 : pub(crate) async fn split_prepare(
3894 0 : &self,
3895 0 : child_shards: &Vec<TenantShardId>,
3896 0 : ) -> anyhow::Result<()> {
3897 0 : let (timelines, offloaded) = {
3898 0 : let timelines = self.timelines.lock().unwrap();
3899 0 : let offloaded = self.timelines_offloaded.lock().unwrap();
3900 0 : (timelines.clone(), offloaded.clone())
3901 0 : };
3902 0 : let timelines_iter = timelines
3903 0 : .values()
3904 0 : .map(TimelineOrOffloadedArcRef::<'_>::from)
3905 0 : .chain(
3906 0 : offloaded
3907 0 : .values()
3908 0 : .map(TimelineOrOffloadedArcRef::<'_>::from),
3909 : );
3910 0 : for timeline in timelines_iter {
3911 : // We do not block timeline creation/deletion during splits inside the pageserver: it is up to higher levels
3912 : // to ensure that they do not start a split if currently in the process of doing these.
3913 :
3914 0 : let timeline_id = timeline.timeline_id();
3915 :
3916 0 : if let TimelineOrOffloadedArcRef::Timeline(timeline) = timeline {
3917 : // Upload an index from the parent: this is partly to provide freshness for the
3918 : // child tenants that will copy it, and partly for general ease-of-debugging: there will
3919 : // always be a parent shard index in the same generation as we wrote the child shard index.
3920 0 : tracing::info!(%timeline_id, "Uploading index");
3921 0 : timeline
3922 0 : .remote_client
3923 0 : .schedule_index_upload_for_file_changes()?;
3924 0 : timeline.remote_client.wait_completion().await?;
3925 0 : }
3926 :
3927 0 : let remote_client = match timeline {
3928 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.remote_client.clone(),
3929 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => {
3930 0 : let remote_client = self
3931 0 : .build_timeline_client(offloaded.timeline_id, self.remote_storage.clone());
3932 0 : Arc::new(remote_client)
3933 : }
3934 : TimelineOrOffloadedArcRef::Importing(_) => {
3935 0 : unreachable!("Importing timelines are not included in the iterator")
3936 : }
3937 : };
3938 :
3939 : // Shut down the timeline's remote client: this means that the indices we write
3940 : // for child shards will not be invalidated by the parent shard deleting layers.
3941 0 : tracing::info!(%timeline_id, "Shutting down remote storage client");
3942 0 : remote_client.shutdown().await;
3943 :
3944 : // Download methods can still be used after shutdown, as they don't flow through the remote client's
3945 : // queue. In principal the RemoteTimelineClient could provide this without downloading it, but this
3946 : // operation is rare, so it's simpler to just download it (and robustly guarantees that the index
3947 : // we use here really is the remotely persistent one).
3948 0 : tracing::info!(%timeline_id, "Downloading index_part from parent");
3949 0 : let result = remote_client
3950 0 : .download_index_file(&self.cancel)
3951 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))
3952 0 : .await?;
3953 0 : let index_part = match result {
3954 : MaybeDeletedIndexPart::Deleted(_) => {
3955 0 : anyhow::bail!("Timeline deletion happened concurrently with split")
3956 : }
3957 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
3958 : };
3959 :
3960 : // A shard split may not take place while a timeline import is on-going
3961 : // for the tenant. Timeline imports run as part of each tenant shard
3962 : // and rely on the sharding scheme to split the work among pageservers.
3963 : // If we were to split in the middle of this process, we would have to
3964 : // either ensure that it's driven to completion on the old shard set
3965 : // or transfer it to the new shard set. It's technically possible, but complex.
3966 0 : match index_part.import_pgdata {
3967 0 : Some(ref import) if !import.is_done() => {
3968 0 : anyhow::bail!(
3969 0 : "Cannot split due to import with idempotency key: {:?}",
3970 0 : import.idempotency_key()
3971 : );
3972 : }
3973 0 : Some(_) | None => {
3974 0 : // fallthrough
3975 0 : }
3976 : }
3977 :
3978 0 : for child_shard in child_shards {
3979 0 : tracing::info!(%timeline_id, "Uploading index_part for child {}", child_shard.to_index());
3980 0 : upload_index_part(
3981 0 : &self.remote_storage,
3982 0 : child_shard,
3983 0 : &timeline_id,
3984 0 : self.generation,
3985 0 : &index_part,
3986 0 : &self.cancel,
3987 0 : )
3988 0 : .await?;
3989 : }
3990 : }
3991 :
3992 0 : let tenant_manifest = self.build_tenant_manifest();
3993 0 : for child_shard in child_shards {
3994 0 : tracing::info!(
3995 0 : "Uploading tenant manifest for child {}",
3996 0 : child_shard.to_index()
3997 : );
3998 0 : upload_tenant_manifest(
3999 0 : &self.remote_storage,
4000 0 : child_shard,
4001 0 : self.generation,
4002 0 : &tenant_manifest,
4003 0 : &self.cancel,
4004 0 : )
4005 0 : .await?;
4006 : }
4007 :
4008 0 : Ok(())
4009 0 : }
4010 :
4011 0 : pub(crate) fn get_sizes(&self) -> TopTenantShardItem {
4012 0 : let mut result = TopTenantShardItem {
4013 0 : id: self.tenant_shard_id,
4014 0 : resident_size: 0,
4015 0 : physical_size: 0,
4016 0 : max_logical_size: 0,
4017 0 : max_logical_size_per_shard: 0,
4018 0 : };
4019 :
4020 0 : for timeline in self.timelines.lock().unwrap().values() {
4021 0 : result.resident_size += timeline.metrics.resident_physical_size_gauge.get();
4022 0 :
4023 0 : result.physical_size += timeline
4024 0 : .remote_client
4025 0 : .metrics
4026 0 : .remote_physical_size_gauge
4027 0 : .get();
4028 0 : result.max_logical_size = std::cmp::max(
4029 0 : result.max_logical_size,
4030 0 : timeline.metrics.current_logical_size_gauge.get(),
4031 0 : );
4032 0 : }
4033 :
4034 0 : result.max_logical_size_per_shard = result
4035 0 : .max_logical_size
4036 0 : .div_ceil(self.tenant_shard_id.shard_count.count() as u64);
4037 :
4038 0 : result
4039 0 : }
4040 : }
4041 :
4042 : /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
4043 : /// perform a topological sort, so that the parent of each timeline comes
4044 : /// before the children.
4045 : /// E extracts the ancestor from T
4046 : /// This allows for T to be different. It can be TimelineMetadata, can be Timeline itself, etc.
4047 119 : fn tree_sort_timelines<T, E>(
4048 119 : timelines: HashMap<TimelineId, T>,
4049 119 : extractor: E,
4050 119 : ) -> anyhow::Result<Vec<(TimelineId, T)>>
4051 119 : where
4052 119 : E: Fn(&T) -> Option<TimelineId>,
4053 : {
4054 119 : let mut result = Vec::with_capacity(timelines.len());
4055 :
4056 119 : let mut now = Vec::with_capacity(timelines.len());
4057 : // (ancestor, children)
4058 119 : let mut later: HashMap<TimelineId, Vec<(TimelineId, T)>> =
4059 119 : HashMap::with_capacity(timelines.len());
4060 :
4061 122 : for (timeline_id, value) in timelines {
4062 3 : if let Some(ancestor_id) = extractor(&value) {
4063 1 : let children = later.entry(ancestor_id).or_default();
4064 1 : children.push((timeline_id, value));
4065 2 : } else {
4066 2 : now.push((timeline_id, value));
4067 2 : }
4068 : }
4069 :
4070 122 : while let Some((timeline_id, metadata)) = now.pop() {
4071 3 : result.push((timeline_id, metadata));
4072 : // All children of this can be loaded now
4073 3 : if let Some(mut children) = later.remove(&timeline_id) {
4074 1 : now.append(&mut children);
4075 2 : }
4076 : }
4077 :
4078 : // All timelines should be visited now. Unless there were timelines with missing ancestors.
4079 119 : if !later.is_empty() {
4080 0 : for (missing_id, orphan_ids) in later {
4081 0 : for (orphan_id, _) in orphan_ids {
4082 0 : error!(
4083 0 : "could not load timeline {orphan_id} because its ancestor timeline {missing_id} could not be loaded"
4084 : );
4085 : }
4086 : }
4087 0 : bail!("could not load tenant because some timelines are missing ancestors");
4088 119 : }
4089 :
4090 119 : Ok(result)
4091 119 : }
4092 :
4093 : impl TenantShard {
4094 0 : pub fn tenant_specific_overrides(&self) -> pageserver_api::models::TenantConfig {
4095 0 : self.tenant_conf.load().tenant_conf.clone()
4096 0 : }
4097 :
4098 0 : pub fn effective_config(&self) -> pageserver_api::config::TenantConfigToml {
4099 0 : self.tenant_specific_overrides()
4100 0 : .merge(self.conf.default_tenant_conf.clone())
4101 0 : }
4102 :
4103 0 : pub fn get_checkpoint_distance(&self) -> u64 {
4104 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4105 0 : tenant_conf
4106 0 : .checkpoint_distance
4107 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
4108 0 : }
4109 :
4110 0 : pub fn get_checkpoint_timeout(&self) -> Duration {
4111 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4112 0 : tenant_conf
4113 0 : .checkpoint_timeout
4114 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
4115 0 : }
4116 :
4117 0 : pub fn get_compaction_target_size(&self) -> u64 {
4118 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4119 0 : tenant_conf
4120 0 : .compaction_target_size
4121 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
4122 0 : }
4123 :
4124 0 : pub fn get_compaction_period(&self) -> Duration {
4125 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4126 0 : tenant_conf
4127 0 : .compaction_period
4128 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
4129 0 : }
4130 :
4131 0 : pub fn get_compaction_threshold(&self) -> usize {
4132 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4133 0 : tenant_conf
4134 0 : .compaction_threshold
4135 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
4136 0 : }
4137 :
4138 0 : pub fn get_rel_size_v2_enabled(&self) -> bool {
4139 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4140 0 : tenant_conf
4141 0 : .rel_size_v2_enabled
4142 0 : .unwrap_or(self.conf.default_tenant_conf.rel_size_v2_enabled)
4143 0 : }
4144 :
4145 0 : pub fn get_compaction_upper_limit(&self) -> usize {
4146 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4147 0 : tenant_conf
4148 0 : .compaction_upper_limit
4149 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_upper_limit)
4150 0 : }
4151 :
4152 0 : pub fn get_compaction_l0_first(&self) -> bool {
4153 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4154 0 : tenant_conf
4155 0 : .compaction_l0_first
4156 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_first)
4157 0 : }
4158 :
4159 121 : pub fn get_gc_horizon(&self) -> u64 {
4160 121 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4161 121 : tenant_conf
4162 121 : .gc_horizon
4163 121 : .unwrap_or(self.conf.default_tenant_conf.gc_horizon)
4164 121 : }
4165 :
4166 0 : pub fn get_gc_period(&self) -> Duration {
4167 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4168 0 : tenant_conf
4169 0 : .gc_period
4170 0 : .unwrap_or(self.conf.default_tenant_conf.gc_period)
4171 0 : }
4172 :
4173 0 : pub fn get_image_creation_threshold(&self) -> usize {
4174 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4175 0 : tenant_conf
4176 0 : .image_creation_threshold
4177 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
4178 0 : }
4179 :
4180 : // HADRON
4181 0 : pub fn get_image_creation_timeout(&self) -> Option<Duration> {
4182 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4183 0 : tenant_conf.image_layer_force_creation_period.or(self
4184 0 : .conf
4185 0 : .default_tenant_conf
4186 0 : .image_layer_force_creation_period)
4187 0 : }
4188 :
4189 2 : pub fn get_pitr_interval(&self) -> Duration {
4190 2 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4191 2 : tenant_conf
4192 2 : .pitr_interval
4193 2 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
4194 2 : }
4195 :
4196 0 : pub fn get_min_resident_size_override(&self) -> Option<u64> {
4197 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4198 0 : tenant_conf
4199 0 : .min_resident_size_override
4200 0 : .or(self.conf.default_tenant_conf.min_resident_size_override)
4201 0 : }
4202 :
4203 0 : pub fn get_heatmap_period(&self) -> Option<Duration> {
4204 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4205 0 : let heatmap_period = tenant_conf
4206 0 : .heatmap_period
4207 0 : .unwrap_or(self.conf.default_tenant_conf.heatmap_period);
4208 0 : if heatmap_period.is_zero() {
4209 0 : None
4210 : } else {
4211 0 : Some(heatmap_period)
4212 : }
4213 0 : }
4214 :
4215 0 : pub fn get_lsn_lease_length(&self) -> Duration {
4216 0 : Self::get_lsn_lease_length_impl(self.conf, &self.tenant_conf.load().tenant_conf)
4217 0 : }
4218 :
4219 119 : pub fn get_lsn_lease_length_impl(
4220 119 : conf: &'static PageServerConf,
4221 119 : tenant_conf: &pageserver_api::models::TenantConfig,
4222 119 : ) -> Duration {
4223 119 : tenant_conf
4224 119 : .lsn_lease_length
4225 119 : .unwrap_or(conf.default_tenant_conf.lsn_lease_length)
4226 119 : }
4227 :
4228 0 : pub fn get_timeline_offloading_enabled(&self) -> bool {
4229 0 : if self.conf.timeline_offloading {
4230 0 : return true;
4231 0 : }
4232 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4233 0 : tenant_conf
4234 0 : .timeline_offloading
4235 0 : .unwrap_or(self.conf.default_tenant_conf.timeline_offloading)
4236 0 : }
4237 :
4238 : /// Generate an up-to-date TenantManifest based on the state of this Tenant.
4239 120 : fn build_tenant_manifest(&self) -> TenantManifest {
4240 : // Collect the offloaded timelines, and sort them for deterministic output.
4241 120 : let offloaded_timelines = self
4242 120 : .timelines_offloaded
4243 120 : .lock()
4244 120 : .unwrap()
4245 120 : .values()
4246 120 : .map(|tli| tli.manifest())
4247 120 : .sorted_by_key(|m| m.timeline_id)
4248 120 : .collect_vec();
4249 :
4250 120 : TenantManifest {
4251 120 : version: LATEST_TENANT_MANIFEST_VERSION,
4252 120 : stripe_size: Some(self.get_shard_stripe_size()),
4253 120 : offloaded_timelines,
4254 120 : }
4255 120 : }
4256 :
4257 1 : pub fn update_tenant_config<
4258 1 : F: Fn(
4259 1 : pageserver_api::models::TenantConfig,
4260 1 : ) -> anyhow::Result<pageserver_api::models::TenantConfig>,
4261 1 : >(
4262 1 : &self,
4263 1 : update: F,
4264 1 : ) -> anyhow::Result<pageserver_api::models::TenantConfig> {
4265 : // Use read-copy-update in order to avoid overwriting the location config
4266 : // state if this races with [`TenantShard::set_new_location_config`]. Note that
4267 : // this race is not possible if both request types come from the storage
4268 : // controller (as they should!) because an exclusive op lock is required
4269 : // on the storage controller side.
4270 :
4271 1 : self.tenant_conf
4272 1 : .try_rcu(|attached_conf| -> Result<_, anyhow::Error> {
4273 1 : Ok(Arc::new(AttachedTenantConf {
4274 1 : tenant_conf: update(attached_conf.tenant_conf.clone())?,
4275 1 : location: attached_conf.location,
4276 1 : lsn_lease_deadline: attached_conf.lsn_lease_deadline,
4277 : }))
4278 1 : })?;
4279 :
4280 1 : let updated = self.tenant_conf.load();
4281 :
4282 1 : self.tenant_conf_updated(&updated.tenant_conf);
4283 : // Don't hold self.timelines.lock() during the notifies.
4284 : // There's no risk of deadlock right now, but there could be if we consolidate
4285 : // mutexes in struct Timeline in the future.
4286 1 : let timelines = self.list_timelines();
4287 1 : for timeline in timelines {
4288 0 : timeline.tenant_conf_updated(&updated);
4289 0 : }
4290 :
4291 1 : Ok(updated.tenant_conf.clone())
4292 1 : }
4293 :
4294 0 : pub(crate) fn set_new_location_config(&self, new_conf: AttachedTenantConf) {
4295 0 : let new_tenant_conf = new_conf.tenant_conf.clone();
4296 :
4297 0 : self.tenant_conf.store(Arc::new(new_conf.clone()));
4298 :
4299 0 : self.tenant_conf_updated(&new_tenant_conf);
4300 : // Don't hold self.timelines.lock() during the notifies.
4301 : // There's no risk of deadlock right now, but there could be if we consolidate
4302 : // mutexes in struct Timeline in the future.
4303 0 : let timelines = self.list_timelines();
4304 0 : for timeline in timelines {
4305 0 : timeline.tenant_conf_updated(&new_conf);
4306 0 : }
4307 0 : }
4308 :
4309 120 : fn get_pagestream_throttle_config(
4310 120 : psconf: &'static PageServerConf,
4311 120 : overrides: &pageserver_api::models::TenantConfig,
4312 120 : ) -> throttle::Config {
4313 120 : overrides
4314 120 : .timeline_get_throttle
4315 120 : .clone()
4316 120 : .unwrap_or(psconf.default_tenant_conf.timeline_get_throttle.clone())
4317 120 : }
4318 :
4319 1 : pub(crate) fn tenant_conf_updated(&self, new_conf: &pageserver_api::models::TenantConfig) {
4320 1 : let conf = Self::get_pagestream_throttle_config(self.conf, new_conf);
4321 1 : self.pagestream_throttle.reconfigure(conf)
4322 1 : }
4323 :
4324 : /// Helper function to create a new Timeline struct.
4325 : ///
4326 : /// The returned Timeline is in Loading state. The caller is responsible for
4327 : /// initializing any on-disk state, and for inserting the Timeline to the 'timelines'
4328 : /// map.
4329 : ///
4330 : /// `validate_ancestor == false` is used when a timeline is created for deletion
4331 : /// and we might not have the ancestor present anymore which is fine for to be
4332 : /// deleted timelines.
4333 : #[allow(clippy::too_many_arguments)]
4334 235 : fn create_timeline_struct(
4335 235 : &self,
4336 235 : new_timeline_id: TimelineId,
4337 235 : new_metadata: &TimelineMetadata,
4338 235 : previous_heatmap: Option<PreviousHeatmap>,
4339 235 : ancestor: Option<Arc<Timeline>>,
4340 235 : resources: TimelineResources,
4341 235 : cause: CreateTimelineCause,
4342 235 : create_idempotency: CreateTimelineIdempotency,
4343 235 : gc_compaction_state: Option<GcCompactionState>,
4344 235 : rel_size_v2_status: Option<RelSizeMigration>,
4345 235 : ctx: &RequestContext,
4346 235 : ) -> anyhow::Result<(Arc<Timeline>, RequestContext)> {
4347 235 : let state = match cause {
4348 : CreateTimelineCause::Load => {
4349 235 : let ancestor_id = new_metadata.ancestor_timeline();
4350 235 : anyhow::ensure!(
4351 235 : ancestor_id == ancestor.as_ref().map(|t| t.timeline_id),
4352 0 : "Timeline's {new_timeline_id} ancestor {ancestor_id:?} was not found"
4353 : );
4354 235 : TimelineState::Loading
4355 : }
4356 0 : CreateTimelineCause::Delete => TimelineState::Stopping,
4357 : };
4358 :
4359 235 : let pg_version = new_metadata.pg_version();
4360 :
4361 235 : let timeline = Timeline::new(
4362 235 : self.conf,
4363 235 : Arc::clone(&self.tenant_conf),
4364 235 : new_metadata,
4365 235 : previous_heatmap,
4366 235 : ancestor,
4367 235 : new_timeline_id,
4368 235 : self.tenant_shard_id,
4369 235 : self.generation,
4370 235 : self.shard_identity,
4371 235 : self.walredo_mgr.clone(),
4372 235 : resources,
4373 235 : pg_version,
4374 235 : state,
4375 235 : self.attach_wal_lag_cooldown.clone(),
4376 235 : create_idempotency,
4377 235 : gc_compaction_state,
4378 235 : rel_size_v2_status,
4379 235 : self.cancel.child_token(),
4380 : );
4381 :
4382 235 : let timeline_ctx = RequestContextBuilder::from(ctx)
4383 235 : .scope(context::Scope::new_timeline(&timeline))
4384 235 : .detached_child();
4385 :
4386 235 : Ok((timeline, timeline_ctx))
4387 235 : }
4388 :
4389 : /// [`TenantShard::shutdown`] must be called before dropping the returned [`TenantShard`] object
4390 : /// to ensure proper cleanup of background tasks and metrics.
4391 : //
4392 : // Allow too_many_arguments because a constructor's argument list naturally grows with the
4393 : // number of attributes in the struct: breaking these out into a builder wouldn't be helpful.
4394 : #[allow(clippy::too_many_arguments)]
4395 119 : fn new(
4396 119 : state: TenantState,
4397 119 : conf: &'static PageServerConf,
4398 119 : attached_conf: AttachedTenantConf,
4399 119 : shard_identity: ShardIdentity,
4400 119 : walredo_mgr: Option<Arc<WalRedoManager>>,
4401 119 : tenant_shard_id: TenantShardId,
4402 119 : remote_storage: GenericRemoteStorage,
4403 119 : deletion_queue_client: DeletionQueueClient,
4404 119 : l0_flush_global_state: L0FlushGlobalState,
4405 119 : basebackup_cache: Arc<BasebackupCache>,
4406 119 : feature_resolver: FeatureResolver,
4407 119 : ) -> TenantShard {
4408 119 : assert!(!attached_conf.location.generation.is_none());
4409 :
4410 119 : let (state, mut rx) = watch::channel(state);
4411 :
4412 119 : tokio::spawn(async move {
4413 : // reflect tenant state in metrics:
4414 : // - global per tenant state: TENANT_STATE_METRIC
4415 : // - "set" of broken tenants: BROKEN_TENANTS_SET
4416 : //
4417 : // set of broken tenants should not have zero counts so that it remains accessible for
4418 : // alerting.
4419 :
4420 118 : let tid = tenant_shard_id.to_string();
4421 118 : let shard_id = tenant_shard_id.shard_slug().to_string();
4422 118 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
4423 :
4424 236 : fn inspect_state(state: &TenantState) -> ([&'static str; 1], bool) {
4425 236 : ([state.into()], matches!(state, TenantState::Broken { .. }))
4426 236 : }
4427 :
4428 118 : let mut tuple = inspect_state(&rx.borrow_and_update());
4429 :
4430 118 : let is_broken = tuple.1;
4431 118 : let mut counted_broken = if is_broken {
4432 : // add the id to the set right away, there should not be any updates on the channel
4433 : // after before tenant is removed, if ever
4434 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
4435 0 : true
4436 : } else {
4437 118 : false
4438 : };
4439 :
4440 : loop {
4441 236 : let labels = &tuple.0;
4442 236 : let current = TENANT_STATE_METRIC.with_label_values(labels);
4443 236 : current.inc();
4444 :
4445 236 : if rx.changed().await.is_err() {
4446 : // tenant has been dropped
4447 7 : current.dec();
4448 7 : drop(BROKEN_TENANTS_SET.remove_label_values(set_key));
4449 7 : break;
4450 118 : }
4451 :
4452 118 : current.dec();
4453 118 : tuple = inspect_state(&rx.borrow_and_update());
4454 :
4455 118 : let is_broken = tuple.1;
4456 118 : if is_broken && !counted_broken {
4457 0 : counted_broken = true;
4458 0 : // insert the tenant_id (back) into the set while avoiding needless counter
4459 0 : // access
4460 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
4461 118 : }
4462 : }
4463 7 : });
4464 :
4465 119 : TenantShard {
4466 119 : tenant_shard_id,
4467 119 : shard_identity,
4468 119 : generation: attached_conf.location.generation,
4469 119 : conf,
4470 119 : // using now here is good enough approximation to catch tenants with really long
4471 119 : // activation times.
4472 119 : constructed_at: Instant::now(),
4473 119 : timelines: Mutex::new(HashMap::new()),
4474 119 : timelines_creating: Mutex::new(HashSet::new()),
4475 119 : timelines_offloaded: Mutex::new(HashMap::new()),
4476 119 : timelines_importing: Mutex::new(HashMap::new()),
4477 119 : remote_tenant_manifest: Default::default(),
4478 119 : gc_cs: tokio::sync::Mutex::new(()),
4479 119 : walredo_mgr,
4480 119 : remote_storage,
4481 119 : deletion_queue_client,
4482 119 : state,
4483 119 : cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
4484 119 : cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
4485 119 : eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
4486 119 : compaction_circuit_breaker: std::sync::Mutex::new(CircuitBreaker::new(
4487 119 : format!("compaction-{tenant_shard_id}"),
4488 119 : 5,
4489 119 : // Compaction can be a very expensive operation, and might leak disk space. It also ought
4490 119 : // to be infallible, as long as remote storage is available. So if it repeatedly fails,
4491 119 : // use an extremely long backoff.
4492 119 : Some(Duration::from_secs(3600 * 24)),
4493 119 : )),
4494 119 : l0_compaction_trigger: Arc::new(Notify::new()),
4495 119 : scheduled_compaction_tasks: Mutex::new(Default::default()),
4496 119 : activate_now_sem: tokio::sync::Semaphore::new(0),
4497 119 : attach_wal_lag_cooldown: Arc::new(std::sync::OnceLock::new()),
4498 119 : cancel: CancellationToken::default(),
4499 119 : gate: Gate::default(),
4500 119 : pagestream_throttle: Arc::new(throttle::Throttle::new(
4501 119 : TenantShard::get_pagestream_throttle_config(conf, &attached_conf.tenant_conf),
4502 119 : )),
4503 119 : pagestream_throttle_metrics: Arc::new(
4504 119 : crate::metrics::tenant_throttling::Pagestream::new(&tenant_shard_id),
4505 119 : ),
4506 119 : tenant_conf: Arc::new(ArcSwap::from_pointee(attached_conf)),
4507 119 : ongoing_timeline_detach: std::sync::Mutex::default(),
4508 119 : gc_block: Default::default(),
4509 119 : l0_flush_global_state,
4510 119 : basebackup_cache,
4511 119 : feature_resolver: Arc::new(TenantFeatureResolver::new(
4512 119 : feature_resolver,
4513 119 : tenant_shard_id.tenant_id,
4514 119 : )),
4515 119 : }
4516 119 : }
4517 :
4518 : /// Locate and load config
4519 0 : pub(super) fn load_tenant_config(
4520 0 : conf: &'static PageServerConf,
4521 0 : tenant_shard_id: &TenantShardId,
4522 0 : ) -> Result<LocationConf, LoadConfigError> {
4523 0 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4524 :
4525 0 : info!("loading tenant configuration from {config_path}");
4526 :
4527 : // load and parse file
4528 0 : let config = fs::read_to_string(&config_path).map_err(|e| {
4529 0 : match e.kind() {
4530 : std::io::ErrorKind::NotFound => {
4531 : // The config should almost always exist for a tenant directory:
4532 : // - When attaching a tenant, the config is the first thing we write
4533 : // - When detaching a tenant, we atomically move the directory to a tmp location
4534 : // before deleting contents.
4535 : //
4536 : // The very rare edge case that can result in a missing config is if we crash during attach
4537 : // between creating directory and writing config. Callers should handle that as if the
4538 : // directory didn't exist.
4539 :
4540 0 : LoadConfigError::NotFound(config_path)
4541 : }
4542 : _ => {
4543 : // No IO errors except NotFound are acceptable here: other kinds of error indicate local storage or permissions issues
4544 : // that we cannot cleanly recover
4545 0 : crate::virtual_file::on_fatal_io_error(&e, "Reading tenant config file")
4546 : }
4547 : }
4548 0 : })?;
4549 :
4550 0 : Ok(toml_edit::de::from_str::<LocationConf>(&config)?)
4551 0 : }
4552 :
4553 : /// Stores a tenant location config to disk.
4554 : ///
4555 : /// NB: make sure to call `ShardIdentity::assert_equal` before persisting a new config, to avoid
4556 : /// changes to shard parameters that may result in data corruption.
4557 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4558 : pub(super) async fn persist_tenant_config(
4559 : conf: &'static PageServerConf,
4560 : tenant_shard_id: &TenantShardId,
4561 : location_conf: &LocationConf,
4562 : ) -> std::io::Result<()> {
4563 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4564 :
4565 : Self::persist_tenant_config_at(tenant_shard_id, &config_path, location_conf).await
4566 : }
4567 :
4568 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4569 : pub(super) async fn persist_tenant_config_at(
4570 : tenant_shard_id: &TenantShardId,
4571 : config_path: &Utf8Path,
4572 : location_conf: &LocationConf,
4573 : ) -> std::io::Result<()> {
4574 : debug!("persisting tenantconf to {config_path}");
4575 :
4576 : let mut conf_content = r#"# This file contains a specific per-tenant's config.
4577 : # It is read in case of pageserver restart.
4578 : "#
4579 : .to_string();
4580 :
4581 0 : fail::fail_point!("tenant-config-before-write", |_| {
4582 0 : Err(std::io::Error::other("tenant-config-before-write"))
4583 0 : });
4584 :
4585 : // Convert the config to a toml file.
4586 : conf_content +=
4587 : &toml_edit::ser::to_string_pretty(&location_conf).expect("Config serialization failed");
4588 :
4589 : let temp_path = path_with_suffix_extension(config_path, TEMP_FILE_SUFFIX);
4590 :
4591 : let conf_content = conf_content.into_bytes();
4592 : VirtualFile::crashsafe_overwrite(config_path.to_owned(), temp_path, conf_content).await
4593 : }
4594 :
4595 : //
4596 : // How garbage collection works:
4597 : //
4598 : // +--bar------------->
4599 : // /
4600 : // +----+-----foo---------------->
4601 : // /
4602 : // ----main--+-------------------------->
4603 : // \
4604 : // +-----baz-------->
4605 : //
4606 : //
4607 : // 1. Grab 'gc_cs' mutex to prevent new timelines from being created while Timeline's
4608 : // `gc_infos` are being refreshed
4609 : // 2. Scan collected timelines, and on each timeline, make note of the
4610 : // all the points where other timelines have been branched off.
4611 : // We will refrain from removing page versions at those LSNs.
4612 : // 3. For each timeline, scan all layer files on the timeline.
4613 : // Remove all files for which a newer file exists and which
4614 : // don't cover any branch point LSNs.
4615 : //
4616 : // TODO:
4617 : // - if a relation has a non-incremental persistent layer on a child branch, then we
4618 : // don't need to keep that in the parent anymore. But currently
4619 : // we do.
4620 377 : async fn gc_iteration_internal(
4621 377 : &self,
4622 377 : target_timeline_id: Option<TimelineId>,
4623 377 : horizon: u64,
4624 377 : pitr: Duration,
4625 377 : cancel: &CancellationToken,
4626 377 : ctx: &RequestContext,
4627 377 : ) -> Result<GcResult, GcError> {
4628 377 : let mut totals: GcResult = Default::default();
4629 377 : let now = Instant::now();
4630 :
4631 377 : let gc_timelines = self
4632 377 : .refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4633 377 : .await?;
4634 :
4635 377 : failpoint_support::sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
4636 :
4637 : // If there is nothing to GC, we don't want any messages in the INFO log.
4638 377 : if !gc_timelines.is_empty() {
4639 377 : info!("{} timelines need GC", gc_timelines.len());
4640 : } else {
4641 0 : debug!("{} timelines need GC", gc_timelines.len());
4642 : }
4643 :
4644 : // Perform GC for each timeline.
4645 : //
4646 : // Note that we don't hold the `TenantShard::gc_cs` lock here because we don't want to delay the
4647 : // branch creation task, which requires the GC lock. A GC iteration can run concurrently
4648 : // with branch creation.
4649 : //
4650 : // See comments in [`TenantShard::branch_timeline`] for more information about why branch
4651 : // creation task can run concurrently with timeline's GC iteration.
4652 754 : for timeline in gc_timelines {
4653 377 : if cancel.is_cancelled() {
4654 : // We were requested to shut down. Stop and return with the progress we
4655 : // made.
4656 0 : break;
4657 377 : }
4658 377 : let result = match timeline.gc().await {
4659 : Err(GcError::TimelineCancelled) => {
4660 0 : if target_timeline_id.is_some() {
4661 : // If we were targetting this specific timeline, surface cancellation to caller
4662 0 : return Err(GcError::TimelineCancelled);
4663 : } else {
4664 : // A timeline may be shutting down independently of the tenant's lifecycle: we should
4665 : // skip past this and proceed to try GC on other timelines.
4666 0 : continue;
4667 : }
4668 : }
4669 377 : r => r?,
4670 : };
4671 377 : totals += result;
4672 : }
4673 :
4674 377 : totals.elapsed = now.elapsed();
4675 377 : Ok(totals)
4676 377 : }
4677 :
4678 : /// Refreshes the Timeline::gc_info for all timelines, returning the
4679 : /// vector of timelines which have [`Timeline::get_last_record_lsn`] past
4680 : /// [`TenantShard::get_gc_horizon`].
4681 : ///
4682 : /// This is usually executed as part of periodic gc, but can now be triggered more often.
4683 2 : pub(crate) async fn refresh_gc_info(
4684 2 : &self,
4685 2 : cancel: &CancellationToken,
4686 2 : ctx: &RequestContext,
4687 2 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4688 : // since this method can now be called at different rates than the configured gc loop, it
4689 : // might be that these configuration values get applied faster than what it was previously,
4690 : // since these were only read from the gc task.
4691 2 : let horizon = self.get_gc_horizon();
4692 2 : let pitr = self.get_pitr_interval();
4693 :
4694 : // refresh all timelines
4695 2 : let target_timeline_id = None;
4696 :
4697 2 : self.refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4698 2 : .await
4699 2 : }
4700 :
4701 : /// Populate all Timelines' `GcInfo` with information about their children. We do not set the
4702 : /// PITR cutoffs here, because that requires I/O: this is done later, before GC, by [`Self::refresh_gc_info_internal`]
4703 : ///
4704 : /// Subsequently, parent-child relationships are updated incrementally inside [`Timeline::new`] and [`Timeline::drop`].
4705 119 : fn initialize_gc_info(
4706 119 : &self,
4707 119 : timelines: &std::sync::MutexGuard<HashMap<TimelineId, Arc<Timeline>>>,
4708 119 : timelines_offloaded: &std::sync::MutexGuard<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
4709 119 : restrict_to_timeline: Option<TimelineId>,
4710 119 : ) {
4711 119 : if restrict_to_timeline.is_none() {
4712 : // This function must be called before activation: after activation timeline create/delete operations
4713 : // might happen, and this function is not safe to run concurrently with those.
4714 119 : assert!(!self.is_active());
4715 0 : }
4716 :
4717 : // Scan all timelines. For each timeline, remember the timeline ID and
4718 : // the branch point where it was created.
4719 119 : let mut all_branchpoints: BTreeMap<TimelineId, Vec<(Lsn, TimelineId, MaybeOffloaded)>> =
4720 119 : BTreeMap::new();
4721 119 : timelines.iter().for_each(|(timeline_id, timeline_entry)| {
4722 3 : if let Some(ancestor_timeline_id) = &timeline_entry.get_ancestor_timeline_id() {
4723 1 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4724 1 : ancestor_children.push((
4725 1 : timeline_entry.get_ancestor_lsn(),
4726 1 : *timeline_id,
4727 1 : MaybeOffloaded::No,
4728 1 : ));
4729 2 : }
4730 3 : });
4731 119 : timelines_offloaded
4732 119 : .iter()
4733 119 : .for_each(|(timeline_id, timeline_entry)| {
4734 0 : let Some(ancestor_timeline_id) = &timeline_entry.ancestor_timeline_id else {
4735 0 : return;
4736 : };
4737 0 : let Some(retain_lsn) = timeline_entry.ancestor_retain_lsn else {
4738 0 : return;
4739 : };
4740 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4741 0 : ancestor_children.push((retain_lsn, *timeline_id, MaybeOffloaded::Yes));
4742 0 : });
4743 :
4744 : // The number of bytes we always keep, irrespective of PITR: this is a constant across timelines
4745 119 : let horizon = self.get_gc_horizon();
4746 :
4747 : // Populate each timeline's GcInfo with information about its child branches
4748 119 : let timelines_to_write = if let Some(timeline_id) = restrict_to_timeline {
4749 0 : itertools::Either::Left(timelines.get(&timeline_id).into_iter())
4750 : } else {
4751 119 : itertools::Either::Right(timelines.values())
4752 : };
4753 122 : for timeline in timelines_to_write {
4754 3 : let mut branchpoints: Vec<(Lsn, TimelineId, MaybeOffloaded)> = all_branchpoints
4755 3 : .remove(&timeline.timeline_id)
4756 3 : .unwrap_or_default();
4757 :
4758 3 : branchpoints.sort_by_key(|b| b.0);
4759 :
4760 3 : let mut target = timeline.gc_info.write().unwrap();
4761 :
4762 3 : target.retain_lsns = branchpoints;
4763 :
4764 3 : let space_cutoff = timeline
4765 3 : .get_last_record_lsn()
4766 3 : .checked_sub(horizon)
4767 3 : .unwrap_or(Lsn(0));
4768 :
4769 3 : target.cutoffs = GcCutoffs {
4770 3 : space: space_cutoff,
4771 3 : time: None,
4772 3 : };
4773 : }
4774 119 : }
4775 :
4776 379 : async fn refresh_gc_info_internal(
4777 379 : &self,
4778 379 : target_timeline_id: Option<TimelineId>,
4779 379 : horizon: u64,
4780 379 : pitr: Duration,
4781 379 : cancel: &CancellationToken,
4782 379 : ctx: &RequestContext,
4783 379 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4784 : // before taking the gc_cs lock, do the heavier weight finding of gc_cutoff points for
4785 : // currently visible timelines.
4786 379 : let timelines = self
4787 379 : .timelines
4788 379 : .lock()
4789 379 : .unwrap()
4790 379 : .values()
4791 1663 : .filter(|tl| match target_timeline_id.as_ref() {
4792 1655 : Some(target) => &tl.timeline_id == target,
4793 8 : None => true,
4794 1663 : })
4795 379 : .cloned()
4796 379 : .collect::<Vec<_>>();
4797 :
4798 379 : if target_timeline_id.is_some() && timelines.is_empty() {
4799 : // We were to act on a particular timeline and it wasn't found
4800 0 : return Err(GcError::TimelineNotFound);
4801 379 : }
4802 :
4803 379 : let mut gc_cutoffs: HashMap<TimelineId, GcCutoffs> =
4804 379 : HashMap::with_capacity(timelines.len());
4805 :
4806 : // Ensures all timelines use the same start time when computing the time cutoff.
4807 379 : let now_ts_for_pitr_calc = SystemTime::now();
4808 385 : for timeline in timelines.iter() {
4809 385 : let ctx = &ctx.with_scope_timeline(timeline);
4810 385 : let cutoff = timeline
4811 385 : .get_last_record_lsn()
4812 385 : .checked_sub(horizon)
4813 385 : .unwrap_or(Lsn(0));
4814 :
4815 385 : let cutoffs = timeline
4816 385 : .find_gc_cutoffs(now_ts_for_pitr_calc, cutoff, pitr, cancel, ctx)
4817 385 : .await?;
4818 385 : let old = gc_cutoffs.insert(timeline.timeline_id, cutoffs);
4819 385 : assert!(old.is_none());
4820 : }
4821 :
4822 379 : if !self.is_active() || self.cancel.is_cancelled() {
4823 0 : return Err(GcError::TenantCancelled);
4824 379 : }
4825 :
4826 : // grab mutex to prevent new timelines from being created here; avoid doing long operations
4827 : // because that will stall branch creation.
4828 379 : let gc_cs = self.gc_cs.lock().await;
4829 :
4830 : // Ok, we now know all the branch points.
4831 : // Update the GC information for each timeline.
4832 379 : let mut gc_timelines = Vec::with_capacity(timelines.len());
4833 764 : for timeline in timelines {
4834 : // We filtered the timeline list above
4835 385 : if let Some(target_timeline_id) = target_timeline_id {
4836 377 : assert_eq!(target_timeline_id, timeline.timeline_id);
4837 8 : }
4838 :
4839 : {
4840 385 : let mut target = timeline.gc_info.write().unwrap();
4841 :
4842 : // Cull any expired leases
4843 385 : let now = SystemTime::now();
4844 385 : target.leases.retain(|_, lease| !lease.is_expired(&now));
4845 :
4846 385 : timeline
4847 385 : .metrics
4848 385 : .valid_lsn_lease_count_gauge
4849 385 : .set(target.leases.len() as u64);
4850 :
4851 : // Look up parent's PITR cutoff to update the child's knowledge of whether it is within parent's PITR
4852 385 : if let Some(ancestor_id) = timeline.get_ancestor_timeline_id() {
4853 56 : if let Some(ancestor_gc_cutoffs) = gc_cutoffs.get(&ancestor_id) {
4854 6 : target.within_ancestor_pitr =
4855 6 : Some(timeline.get_ancestor_lsn()) >= ancestor_gc_cutoffs.time;
4856 50 : }
4857 329 : }
4858 :
4859 : // Update metrics that depend on GC state
4860 385 : timeline
4861 385 : .metrics
4862 385 : .archival_size
4863 385 : .set(if target.within_ancestor_pitr {
4864 0 : timeline.metrics.current_logical_size_gauge.get()
4865 : } else {
4866 385 : 0
4867 : });
4868 385 : if let Some(time_cutoff) = target.cutoffs.time {
4869 319 : timeline.metrics.pitr_history_size.set(
4870 319 : timeline
4871 319 : .get_last_record_lsn()
4872 319 : .checked_sub(time_cutoff)
4873 319 : .unwrap_or_default()
4874 319 : .0,
4875 319 : );
4876 319 : }
4877 :
4878 : // Apply the cutoffs we found to the Timeline's GcInfo. Why might we _not_ have cutoffs for a timeline?
4879 : // - this timeline was created while we were finding cutoffs
4880 : // - lsn for timestamp search fails for this timeline repeatedly
4881 385 : if let Some(cutoffs) = gc_cutoffs.get(&timeline.timeline_id) {
4882 385 : let original_cutoffs = target.cutoffs.clone();
4883 : // GC cutoffs should never go back
4884 385 : target.cutoffs = GcCutoffs {
4885 385 : space: cutoffs.space.max(original_cutoffs.space),
4886 385 : time: cutoffs.time.max(original_cutoffs.time),
4887 385 : }
4888 0 : }
4889 : }
4890 :
4891 385 : gc_timelines.push(timeline);
4892 : }
4893 379 : drop(gc_cs);
4894 379 : Ok(gc_timelines)
4895 379 : }
4896 :
4897 : /// A substitute for `branch_timeline` for use in unit tests.
4898 : /// The returned timeline will have state value `Active` to make various `anyhow::ensure!()`
4899 : /// calls pass, but, we do not actually call `.activate()` under the hood. So, none of the
4900 : /// timeline background tasks are launched, except the flush loop.
4901 : #[cfg(test)]
4902 119 : async fn branch_timeline_test(
4903 119 : self: &Arc<Self>,
4904 119 : src_timeline: &Arc<Timeline>,
4905 119 : dst_id: TimelineId,
4906 119 : ancestor_lsn: Option<Lsn>,
4907 119 : ctx: &RequestContext,
4908 119 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
4909 119 : let tl = self
4910 119 : .branch_timeline_impl(src_timeline, dst_id, ancestor_lsn, ctx)
4911 119 : .await?
4912 117 : .into_timeline_for_test();
4913 117 : tl.set_state(TimelineState::Active);
4914 117 : Ok(tl)
4915 119 : }
4916 :
4917 : /// Helper for unit tests to branch a timeline with some pre-loaded states.
4918 : #[cfg(test)]
4919 : #[allow(clippy::too_many_arguments)]
4920 6 : pub async fn branch_timeline_test_with_layers(
4921 6 : self: &Arc<Self>,
4922 6 : src_timeline: &Arc<Timeline>,
4923 6 : dst_id: TimelineId,
4924 6 : ancestor_lsn: Option<Lsn>,
4925 6 : ctx: &RequestContext,
4926 6 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
4927 6 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
4928 6 : end_lsn: Lsn,
4929 6 : ) -> anyhow::Result<Arc<Timeline>> {
4930 : use checks::check_valid_layermap;
4931 : use itertools::Itertools;
4932 :
4933 6 : let tline = self
4934 6 : .branch_timeline_test(src_timeline, dst_id, ancestor_lsn, ctx)
4935 6 : .await?;
4936 6 : let ancestor_lsn = if let Some(ancestor_lsn) = ancestor_lsn {
4937 6 : ancestor_lsn
4938 : } else {
4939 0 : tline.get_last_record_lsn()
4940 : };
4941 6 : assert!(end_lsn >= ancestor_lsn);
4942 6 : tline.force_advance_lsn(end_lsn);
4943 9 : for deltas in delta_layer_desc {
4944 3 : tline
4945 3 : .force_create_delta_layer(deltas, Some(ancestor_lsn), ctx)
4946 3 : .await?;
4947 : }
4948 8 : for (lsn, images) in image_layer_desc {
4949 2 : tline
4950 2 : .force_create_image_layer(lsn, images, Some(ancestor_lsn), ctx)
4951 2 : .await?;
4952 : }
4953 6 : let layer_names = tline
4954 6 : .layers
4955 6 : .read(LayerManagerLockHolder::Testing)
4956 6 : .await
4957 6 : .layer_map()
4958 6 : .unwrap()
4959 6 : .iter_historic_layers()
4960 6 : .map(|layer| layer.layer_name())
4961 6 : .collect_vec();
4962 6 : if let Some(err) = check_valid_layermap(&layer_names) {
4963 0 : bail!("invalid layermap: {err}");
4964 6 : }
4965 6 : Ok(tline)
4966 6 : }
4967 :
4968 : /// Branch an existing timeline.
4969 0 : async fn branch_timeline(
4970 0 : self: &Arc<Self>,
4971 0 : src_timeline: &Arc<Timeline>,
4972 0 : dst_id: TimelineId,
4973 0 : start_lsn: Option<Lsn>,
4974 0 : ctx: &RequestContext,
4975 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4976 0 : self.branch_timeline_impl(src_timeline, dst_id, start_lsn, ctx)
4977 0 : .await
4978 0 : }
4979 :
4980 119 : async fn branch_timeline_impl(
4981 119 : self: &Arc<Self>,
4982 119 : src_timeline: &Arc<Timeline>,
4983 119 : dst_id: TimelineId,
4984 119 : start_lsn: Option<Lsn>,
4985 119 : ctx: &RequestContext,
4986 119 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4987 119 : let src_id = src_timeline.timeline_id;
4988 :
4989 : // We will validate our ancestor LSN in this function. Acquire the GC lock so that
4990 : // this check cannot race with GC, and the ancestor LSN is guaranteed to remain
4991 : // valid while we are creating the branch.
4992 119 : let _gc_cs = self.gc_cs.lock().await;
4993 :
4994 : // If no start LSN is specified, we branch the new timeline from the source timeline's last record LSN
4995 119 : let start_lsn = start_lsn.unwrap_or_else(|| {
4996 1 : let lsn = src_timeline.get_last_record_lsn();
4997 1 : info!("branching timeline {dst_id} from timeline {src_id} at last record LSN: {lsn}");
4998 1 : lsn
4999 1 : });
5000 :
5001 : // we finally have determined the ancestor_start_lsn, so we can get claim exclusivity now
5002 119 : let timeline_create_guard = match self
5003 119 : .start_creating_timeline(
5004 119 : dst_id,
5005 119 : CreateTimelineIdempotency::Branch {
5006 119 : ancestor_timeline_id: src_timeline.timeline_id,
5007 119 : ancestor_start_lsn: start_lsn,
5008 119 : },
5009 119 : )
5010 119 : .await?
5011 : {
5012 119 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
5013 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
5014 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
5015 : }
5016 : };
5017 :
5018 : // Ensure that `start_lsn` is valid, i.e. the LSN is within the PITR
5019 : // horizon on the source timeline
5020 : //
5021 : // We check it against both the planned GC cutoff stored in 'gc_info',
5022 : // and the 'latest_gc_cutoff' of the last GC that was performed. The
5023 : // planned GC cutoff in 'gc_info' is normally larger than
5024 : // 'applied_gc_cutoff_lsn', but beware of corner cases like if you just
5025 : // changed the GC settings for the tenant to make the PITR window
5026 : // larger, but some of the data was already removed by an earlier GC
5027 : // iteration.
5028 :
5029 : // check against last actual 'latest_gc_cutoff' first
5030 119 : let applied_gc_cutoff_lsn = src_timeline.get_applied_gc_cutoff_lsn();
5031 : {
5032 119 : let gc_info = src_timeline.gc_info.read().unwrap();
5033 119 : let planned_cutoff = gc_info.min_cutoff();
5034 119 : if gc_info.lsn_covered_by_lease(start_lsn) {
5035 0 : tracing::info!(
5036 0 : "skipping comparison of {start_lsn} with gc cutoff {} and planned gc cutoff {planned_cutoff} due to lsn lease",
5037 0 : *applied_gc_cutoff_lsn
5038 : );
5039 : } else {
5040 119 : src_timeline
5041 119 : .check_lsn_is_in_scope(start_lsn, &applied_gc_cutoff_lsn)
5042 119 : .context(format!(
5043 119 : "invalid branch start lsn: less than latest GC cutoff {}",
5044 119 : *applied_gc_cutoff_lsn,
5045 : ))
5046 119 : .map_err(CreateTimelineError::AncestorLsn)?;
5047 :
5048 : // and then the planned GC cutoff
5049 117 : if start_lsn < planned_cutoff {
5050 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
5051 0 : "invalid branch start lsn: less than planned GC cutoff {planned_cutoff}"
5052 0 : )));
5053 117 : }
5054 : }
5055 : }
5056 :
5057 : //
5058 : // The branch point is valid, and we are still holding the 'gc_cs' lock
5059 : // so that GC cannot advance the GC cutoff until we are finished.
5060 : // Proceed with the branch creation.
5061 : //
5062 :
5063 : // Determine prev-LSN for the new timeline. We can only determine it if
5064 : // the timeline was branched at the current end of the source timeline.
5065 : let RecordLsn {
5066 117 : last: src_last,
5067 117 : prev: src_prev,
5068 117 : } = src_timeline.get_last_record_rlsn();
5069 117 : let dst_prev = if src_last == start_lsn {
5070 108 : Some(src_prev)
5071 : } else {
5072 9 : None
5073 : };
5074 :
5075 : // Create the metadata file, noting the ancestor of the new timeline.
5076 : // There is initially no data in it, but all the read-calls know to look
5077 : // into the ancestor.
5078 117 : let metadata = TimelineMetadata::new(
5079 117 : start_lsn,
5080 117 : dst_prev,
5081 117 : Some(src_id),
5082 117 : start_lsn,
5083 117 : *src_timeline.applied_gc_cutoff_lsn.read(), // FIXME: should we hold onto this guard longer?
5084 117 : src_timeline.initdb_lsn,
5085 117 : src_timeline.pg_version,
5086 : );
5087 :
5088 117 : let (uninitialized_timeline, _timeline_ctx) = self
5089 117 : .prepare_new_timeline(
5090 117 : dst_id,
5091 117 : &metadata,
5092 117 : timeline_create_guard,
5093 117 : start_lsn + 1,
5094 117 : Some(Arc::clone(src_timeline)),
5095 117 : Some(src_timeline.get_rel_size_v2_status()),
5096 117 : ctx,
5097 117 : )
5098 117 : .await?;
5099 :
5100 117 : let new_timeline = uninitialized_timeline.finish_creation().await?;
5101 :
5102 : // Root timeline gets its layers during creation and uploads them along with the metadata.
5103 : // A branch timeline though, when created, can get no writes for some time, hence won't get any layers created.
5104 : // We still need to upload its metadata eagerly: if other nodes `attach` the tenant and miss this timeline, their GC
5105 : // could get incorrect information and remove more layers, than needed.
5106 : // See also https://github.com/neondatabase/neon/issues/3865
5107 117 : new_timeline
5108 117 : .remote_client
5109 117 : .schedule_index_upload_for_full_metadata_update(&metadata)
5110 117 : .context("branch initial metadata upload")?;
5111 :
5112 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
5113 :
5114 117 : Ok(CreateTimelineResult::Created(new_timeline))
5115 119 : }
5116 :
5117 : /// For unit tests, make this visible so that other modules can directly create timelines
5118 : #[cfg(test)]
5119 : #[tracing::instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
5120 : pub(crate) async fn bootstrap_timeline_test(
5121 : self: &Arc<Self>,
5122 : timeline_id: TimelineId,
5123 : pg_version: PgMajorVersion,
5124 : load_existing_initdb: Option<TimelineId>,
5125 : ctx: &RequestContext,
5126 : ) -> anyhow::Result<Arc<Timeline>> {
5127 : self.bootstrap_timeline(timeline_id, pg_version, load_existing_initdb, ctx)
5128 : .await
5129 : .map_err(anyhow::Error::new)
5130 1 : .map(|r| r.into_timeline_for_test())
5131 : }
5132 :
5133 : /// Get exclusive access to the timeline ID for creation.
5134 : ///
5135 : /// Timeline-creating code paths must use this function before making changes
5136 : /// to in-memory or persistent state.
5137 : ///
5138 : /// The `state` parameter is a description of the timeline creation operation
5139 : /// we intend to perform.
5140 : /// If the timeline was already created in the meantime, we check whether this
5141 : /// request conflicts or is idempotent , based on `state`.
5142 235 : async fn start_creating_timeline(
5143 235 : self: &Arc<Self>,
5144 235 : new_timeline_id: TimelineId,
5145 235 : idempotency: CreateTimelineIdempotency,
5146 235 : ) -> Result<StartCreatingTimelineResult, CreateTimelineError> {
5147 235 : let allow_offloaded = false;
5148 235 : match self.create_timeline_create_guard(new_timeline_id, idempotency, allow_offloaded) {
5149 234 : Ok(create_guard) => {
5150 234 : pausable_failpoint!("timeline-creation-after-uninit");
5151 234 : Ok(StartCreatingTimelineResult::CreateGuard(create_guard))
5152 : }
5153 0 : Err(TimelineExclusionError::ShuttingDown) => Err(CreateTimelineError::ShuttingDown),
5154 : Err(TimelineExclusionError::AlreadyCreating) => {
5155 : // Creation is in progress, we cannot create it again, and we cannot
5156 : // check if this request matches the existing one, so caller must try
5157 : // again later.
5158 0 : Err(CreateTimelineError::AlreadyCreating)
5159 : }
5160 0 : Err(TimelineExclusionError::Other(e)) => Err(CreateTimelineError::Other(e)),
5161 : Err(TimelineExclusionError::AlreadyExists {
5162 0 : existing: TimelineOrOffloaded::Offloaded(_existing),
5163 : ..
5164 : }) => {
5165 0 : info!("timeline already exists but is offloaded");
5166 0 : Err(CreateTimelineError::Conflict)
5167 : }
5168 : Err(TimelineExclusionError::AlreadyExists {
5169 0 : existing: TimelineOrOffloaded::Importing(_existing),
5170 : ..
5171 : }) => {
5172 : // If there's a timeline already importing, then we would hit
5173 : // the [`TimelineExclusionError::AlreadyCreating`] branch above.
5174 0 : unreachable!("Importing timelines hold the creation guard")
5175 : }
5176 : Err(TimelineExclusionError::AlreadyExists {
5177 1 : existing: TimelineOrOffloaded::Timeline(existing),
5178 1 : arg,
5179 : }) => {
5180 : {
5181 1 : let existing = &existing.create_idempotency;
5182 1 : let _span = info_span!("idempotency_check", ?existing, ?arg).entered();
5183 1 : debug!("timeline already exists");
5184 :
5185 1 : match (existing, &arg) {
5186 : // FailWithConflict => no idempotency check
5187 : (CreateTimelineIdempotency::FailWithConflict, _)
5188 : | (_, CreateTimelineIdempotency::FailWithConflict) => {
5189 1 : warn!("timeline already exists, failing request");
5190 1 : return Err(CreateTimelineError::Conflict);
5191 : }
5192 : // Idempotent <=> CreateTimelineIdempotency is identical
5193 0 : (x, y) if x == y => {
5194 0 : info!(
5195 0 : "timeline already exists and idempotency matches, succeeding request"
5196 : );
5197 : // fallthrough
5198 : }
5199 : (_, _) => {
5200 0 : warn!("idempotency conflict, failing request");
5201 0 : return Err(CreateTimelineError::Conflict);
5202 : }
5203 : }
5204 : }
5205 :
5206 0 : Ok(StartCreatingTimelineResult::Idempotent(existing))
5207 : }
5208 : }
5209 235 : }
5210 :
5211 0 : async fn upload_initdb(
5212 0 : &self,
5213 0 : timelines_path: &Utf8PathBuf,
5214 0 : pgdata_path: &Utf8PathBuf,
5215 0 : timeline_id: &TimelineId,
5216 0 : ) -> anyhow::Result<()> {
5217 0 : let temp_path = timelines_path.join(format!(
5218 0 : "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
5219 0 : ));
5220 :
5221 0 : scopeguard::defer! {
5222 : if let Err(e) = fs::remove_file(&temp_path) {
5223 : error!("Failed to remove temporary initdb archive '{temp_path}': {e}");
5224 : }
5225 : }
5226 :
5227 0 : let (pgdata_zstd, tar_zst_size) = create_zst_tarball(pgdata_path, &temp_path).await?;
5228 : const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
5229 0 : if tar_zst_size > INITDB_TAR_ZST_WARN_LIMIT {
5230 0 : warn!(
5231 0 : "compressed {temp_path} size of {tar_zst_size} is above limit {INITDB_TAR_ZST_WARN_LIMIT}."
5232 : );
5233 0 : }
5234 :
5235 0 : pausable_failpoint!("before-initdb-upload");
5236 :
5237 0 : backoff::retry(
5238 0 : || async {
5239 0 : self::remote_timeline_client::upload_initdb_dir(
5240 0 : &self.remote_storage,
5241 0 : &self.tenant_shard_id.tenant_id,
5242 0 : timeline_id,
5243 0 : pgdata_zstd.try_clone().await?,
5244 0 : tar_zst_size,
5245 0 : &self.cancel,
5246 : )
5247 0 : .await
5248 0 : },
5249 : |_| false,
5250 : 3,
5251 : u32::MAX,
5252 0 : "persist_initdb_tar_zst",
5253 0 : &self.cancel,
5254 : )
5255 0 : .await
5256 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
5257 0 : .and_then(|x| x)
5258 0 : }
5259 :
5260 : /// - run initdb to init temporary instance and get bootstrap data
5261 : /// - after initialization completes, tar up the temp dir and upload it to S3.
5262 1 : async fn bootstrap_timeline(
5263 1 : self: &Arc<Self>,
5264 1 : timeline_id: TimelineId,
5265 1 : pg_version: PgMajorVersion,
5266 1 : load_existing_initdb: Option<TimelineId>,
5267 1 : ctx: &RequestContext,
5268 1 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
5269 1 : let timeline_create_guard = match self
5270 1 : .start_creating_timeline(
5271 1 : timeline_id,
5272 1 : CreateTimelineIdempotency::Bootstrap { pg_version },
5273 1 : )
5274 1 : .await?
5275 : {
5276 1 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
5277 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
5278 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
5279 : }
5280 : };
5281 :
5282 : // create a `tenant/{tenant_id}/timelines/basebackup-{timeline_id}.{TEMP_FILE_SUFFIX}/`
5283 : // temporary directory for basebackup files for the given timeline.
5284 :
5285 1 : let timelines_path = self.conf.timelines_path(&self.tenant_shard_id);
5286 1 : let pgdata_path = path_with_suffix_extension(
5287 1 : timelines_path.join(format!("basebackup-{timeline_id}")),
5288 1 : TEMP_FILE_SUFFIX,
5289 : );
5290 :
5291 : // Remove whatever was left from the previous runs: safe because TimelineCreateGuard guarantees
5292 : // we won't race with other creations or existent timelines with the same path.
5293 1 : if pgdata_path.exists() {
5294 0 : fs::remove_dir_all(&pgdata_path).with_context(|| {
5295 0 : format!("Failed to remove already existing initdb directory: {pgdata_path}")
5296 0 : })?;
5297 0 : tracing::info!("removed previous attempt's temporary initdb directory '{pgdata_path}'");
5298 1 : }
5299 :
5300 : // this new directory is very temporary, set to remove it immediately after bootstrap, we don't need it
5301 1 : let pgdata_path_deferred = pgdata_path.clone();
5302 1 : scopeguard::defer! {
5303 : if let Err(e) = fs::remove_dir_all(&pgdata_path_deferred).or_else(fs_ext::ignore_not_found) {
5304 : // this is unlikely, but we will remove the directory on pageserver restart or another bootstrap call
5305 : error!("Failed to remove temporary initdb directory '{pgdata_path_deferred}': {e}");
5306 : } else {
5307 : tracing::info!("removed temporary initdb directory '{pgdata_path_deferred}'");
5308 : }
5309 : }
5310 1 : if let Some(existing_initdb_timeline_id) = load_existing_initdb {
5311 1 : if existing_initdb_timeline_id != timeline_id {
5312 0 : let source_path = &remote_initdb_archive_path(
5313 0 : &self.tenant_shard_id.tenant_id,
5314 0 : &existing_initdb_timeline_id,
5315 0 : );
5316 0 : let dest_path =
5317 0 : &remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &timeline_id);
5318 :
5319 : // if this fails, it will get retried by retried control plane requests
5320 0 : self.remote_storage
5321 0 : .copy_object(source_path, dest_path, &self.cancel)
5322 0 : .await
5323 0 : .context("copy initdb tar")?;
5324 1 : }
5325 1 : let (initdb_tar_zst_path, initdb_tar_zst) =
5326 1 : self::remote_timeline_client::download_initdb_tar_zst(
5327 1 : self.conf,
5328 1 : &self.remote_storage,
5329 1 : &self.tenant_shard_id,
5330 1 : &existing_initdb_timeline_id,
5331 1 : &self.cancel,
5332 1 : )
5333 1 : .await
5334 1 : .context("download initdb tar")?;
5335 :
5336 1 : scopeguard::defer! {
5337 : if let Err(e) = fs::remove_file(&initdb_tar_zst_path) {
5338 : error!("Failed to remove temporary initdb archive '{initdb_tar_zst_path}': {e}");
5339 : }
5340 : }
5341 :
5342 1 : let buf_read =
5343 1 : BufReader::with_capacity(remote_timeline_client::BUFFER_SIZE, initdb_tar_zst);
5344 1 : extract_zst_tarball(&pgdata_path, buf_read)
5345 1 : .await
5346 1 : .context("extract initdb tar")?;
5347 : } else {
5348 : // Init temporarily repo to get bootstrap data, this creates a directory in the `pgdata_path` path
5349 0 : run_initdb(self.conf, &pgdata_path, pg_version, &self.cancel)
5350 0 : .await
5351 0 : .context("run initdb")?;
5352 :
5353 : // Upload the created data dir to S3
5354 0 : if self.tenant_shard_id().is_shard_zero() {
5355 0 : self.upload_initdb(&timelines_path, &pgdata_path, &timeline_id)
5356 0 : .await?;
5357 0 : }
5358 : }
5359 1 : let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
5360 :
5361 : // Import the contents of the data directory at the initial checkpoint
5362 : // LSN, and any WAL after that.
5363 : // Initdb lsn will be equal to last_record_lsn which will be set after import.
5364 : // Because we know it upfront avoid having an option or dummy zero value by passing it to the metadata.
5365 1 : let new_metadata = TimelineMetadata::new(
5366 1 : Lsn(0),
5367 1 : None,
5368 1 : None,
5369 1 : Lsn(0),
5370 1 : pgdata_lsn,
5371 1 : pgdata_lsn,
5372 1 : pg_version,
5373 : );
5374 1 : let (mut raw_timeline, timeline_ctx) = self
5375 1 : .prepare_new_timeline(
5376 1 : timeline_id,
5377 1 : &new_metadata,
5378 1 : timeline_create_guard,
5379 1 : pgdata_lsn,
5380 1 : None,
5381 1 : None,
5382 1 : ctx,
5383 1 : )
5384 1 : .await?;
5385 :
5386 1 : let tenant_shard_id = raw_timeline.owning_tenant.tenant_shard_id;
5387 1 : raw_timeline
5388 1 : .write(|unfinished_timeline| async move {
5389 1 : import_datadir::import_timeline_from_postgres_datadir(
5390 1 : &unfinished_timeline,
5391 1 : &pgdata_path,
5392 1 : pgdata_lsn,
5393 1 : &timeline_ctx,
5394 1 : )
5395 1 : .await
5396 1 : .with_context(|| {
5397 0 : format!(
5398 0 : "Failed to import pgdatadir for timeline {tenant_shard_id}/{timeline_id}"
5399 : )
5400 0 : })?;
5401 :
5402 1 : fail::fail_point!("before-checkpoint-new-timeline", |_| {
5403 0 : Err(CreateTimelineError::Other(anyhow::anyhow!(
5404 0 : "failpoint before-checkpoint-new-timeline"
5405 0 : )))
5406 0 : });
5407 :
5408 1 : Ok(())
5409 2 : })
5410 1 : .await?;
5411 :
5412 : // All done!
5413 1 : let timeline = raw_timeline.finish_creation().await?;
5414 :
5415 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
5416 :
5417 1 : Ok(CreateTimelineResult::Created(timeline))
5418 1 : }
5419 :
5420 232 : fn build_timeline_remote_client(&self, timeline_id: TimelineId) -> RemoteTimelineClient {
5421 232 : RemoteTimelineClient::new(
5422 232 : self.remote_storage.clone(),
5423 232 : self.deletion_queue_client.clone(),
5424 232 : self.conf,
5425 232 : self.tenant_shard_id,
5426 232 : timeline_id,
5427 232 : self.generation,
5428 232 : &self.tenant_conf.load().location,
5429 : )
5430 232 : }
5431 :
5432 : /// Builds required resources for a new timeline.
5433 232 : fn build_timeline_resources(&self, timeline_id: TimelineId) -> TimelineResources {
5434 232 : let remote_client = self.build_timeline_remote_client(timeline_id);
5435 232 : self.get_timeline_resources_for(remote_client)
5436 232 : }
5437 :
5438 : /// Builds timeline resources for the given remote client.
5439 235 : fn get_timeline_resources_for(&self, remote_client: RemoteTimelineClient) -> TimelineResources {
5440 235 : TimelineResources {
5441 235 : remote_client,
5442 235 : pagestream_throttle: self.pagestream_throttle.clone(),
5443 235 : pagestream_throttle_metrics: self.pagestream_throttle_metrics.clone(),
5444 235 : l0_compaction_trigger: self.l0_compaction_trigger.clone(),
5445 235 : l0_flush_global_state: self.l0_flush_global_state.clone(),
5446 235 : basebackup_cache: self.basebackup_cache.clone(),
5447 235 : feature_resolver: self.feature_resolver.clone(),
5448 235 : }
5449 235 : }
5450 :
5451 : /// Creates intermediate timeline structure and its files.
5452 : ///
5453 : /// An empty layer map is initialized, and new data and WAL can be imported starting
5454 : /// at 'disk_consistent_lsn'. After any initial data has been imported, call
5455 : /// `finish_creation` to insert the Timeline into the timelines map.
5456 : #[allow(clippy::too_many_arguments)]
5457 232 : async fn prepare_new_timeline<'a>(
5458 232 : &'a self,
5459 232 : new_timeline_id: TimelineId,
5460 232 : new_metadata: &TimelineMetadata,
5461 232 : create_guard: TimelineCreateGuard,
5462 232 : start_lsn: Lsn,
5463 232 : ancestor: Option<Arc<Timeline>>,
5464 232 : rel_size_v2_status: Option<RelSizeMigration>,
5465 232 : ctx: &RequestContext,
5466 232 : ) -> anyhow::Result<(UninitializedTimeline<'a>, RequestContext)> {
5467 232 : let tenant_shard_id = self.tenant_shard_id;
5468 :
5469 232 : let resources = self.build_timeline_resources(new_timeline_id);
5470 232 : resources
5471 232 : .remote_client
5472 232 : .init_upload_queue_for_empty_remote(new_metadata, rel_size_v2_status.clone())?;
5473 :
5474 232 : let (timeline_struct, timeline_ctx) = self
5475 232 : .create_timeline_struct(
5476 232 : new_timeline_id,
5477 232 : new_metadata,
5478 232 : None,
5479 232 : ancestor,
5480 232 : resources,
5481 232 : CreateTimelineCause::Load,
5482 232 : create_guard.idempotency.clone(),
5483 232 : None,
5484 232 : rel_size_v2_status,
5485 232 : ctx,
5486 : )
5487 232 : .context("Failed to create timeline data structure")?;
5488 :
5489 232 : timeline_struct.init_empty_layer_map(start_lsn);
5490 :
5491 232 : if let Err(e) = self
5492 232 : .create_timeline_files(&create_guard.timeline_path)
5493 232 : .await
5494 : {
5495 0 : error!(
5496 0 : "Failed to create initial files for timeline {tenant_shard_id}/{new_timeline_id}, cleaning up: {e:?}"
5497 : );
5498 0 : cleanup_timeline_directory(create_guard);
5499 0 : return Err(e);
5500 232 : }
5501 :
5502 232 : debug!(
5503 0 : "Successfully created initial files for timeline {tenant_shard_id}/{new_timeline_id}"
5504 : );
5505 :
5506 232 : Ok((
5507 232 : UninitializedTimeline::new(
5508 232 : self,
5509 232 : new_timeline_id,
5510 232 : Some((timeline_struct, create_guard)),
5511 232 : ),
5512 232 : timeline_ctx,
5513 232 : ))
5514 232 : }
5515 :
5516 232 : async fn create_timeline_files(&self, timeline_path: &Utf8Path) -> anyhow::Result<()> {
5517 232 : crashsafe::create_dir(timeline_path).context("Failed to create timeline directory")?;
5518 :
5519 232 : fail::fail_point!("after-timeline-dir-creation", |_| {
5520 0 : anyhow::bail!("failpoint after-timeline-dir-creation");
5521 0 : });
5522 :
5523 232 : Ok(())
5524 232 : }
5525 :
5526 : /// Get a guard that provides exclusive access to the timeline directory, preventing
5527 : /// concurrent attempts to create the same timeline.
5528 : ///
5529 : /// The `allow_offloaded` parameter controls whether to tolerate the existence of
5530 : /// offloaded timelines or not.
5531 235 : fn create_timeline_create_guard(
5532 235 : self: &Arc<Self>,
5533 235 : timeline_id: TimelineId,
5534 235 : idempotency: CreateTimelineIdempotency,
5535 235 : allow_offloaded: bool,
5536 235 : ) -> Result<TimelineCreateGuard, TimelineExclusionError> {
5537 235 : let tenant_shard_id = self.tenant_shard_id;
5538 :
5539 235 : let timeline_path = self.conf.timeline_path(&tenant_shard_id, &timeline_id);
5540 :
5541 235 : let create_guard = TimelineCreateGuard::new(
5542 235 : self,
5543 235 : timeline_id,
5544 235 : timeline_path.clone(),
5545 235 : idempotency,
5546 235 : allow_offloaded,
5547 1 : )?;
5548 :
5549 : // At this stage, we have got exclusive access to in-memory state for this timeline ID
5550 : // for creation.
5551 : // A timeline directory should never exist on disk already:
5552 : // - a previous failed creation would have cleaned up after itself
5553 : // - a pageserver restart would clean up timeline directories that don't have valid remote state
5554 : //
5555 : // Therefore it is an unexpected internal error to encounter a timeline directory already existing here,
5556 : // this error may indicate a bug in cleanup on failed creations.
5557 234 : if timeline_path.exists() {
5558 0 : return Err(TimelineExclusionError::Other(anyhow::anyhow!(
5559 0 : "Timeline directory already exists! This is a bug."
5560 0 : )));
5561 234 : }
5562 :
5563 234 : Ok(create_guard)
5564 235 : }
5565 :
5566 : /// Gathers inputs from all of the timelines to produce a sizing model input.
5567 : ///
5568 : /// Future is cancellation safe. Only one calculation can be running at once per tenant.
5569 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5570 : pub async fn gather_size_inputs(
5571 : &self,
5572 : // `max_retention_period` overrides the cutoff that is used to calculate the size
5573 : // (only if it is shorter than the real cutoff).
5574 : max_retention_period: Option<u64>,
5575 : cause: LogicalSizeCalculationCause,
5576 : cancel: &CancellationToken,
5577 : ctx: &RequestContext,
5578 : ) -> Result<size::ModelInputs, size::CalculateSyntheticSizeError> {
5579 : let logical_sizes_at_once = self
5580 : .conf
5581 : .concurrent_tenant_size_logical_size_queries
5582 : .inner();
5583 :
5584 : // TODO: Having a single mutex block concurrent reads is not great for performance.
5585 : //
5586 : // But the only case where we need to run multiple of these at once is when we
5587 : // request a size for a tenant manually via API, while another background calculation
5588 : // is in progress (which is not a common case).
5589 : //
5590 : // See more for on the issue #2748 condenced out of the initial PR review.
5591 : let mut shared_cache = tokio::select! {
5592 : locked = self.cached_logical_sizes.lock() => locked,
5593 : _ = cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5594 : _ = self.cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5595 : };
5596 :
5597 : size::gather_inputs(
5598 : self,
5599 : logical_sizes_at_once,
5600 : max_retention_period,
5601 : &mut shared_cache,
5602 : cause,
5603 : cancel,
5604 : ctx,
5605 : )
5606 : .await
5607 : }
5608 :
5609 : /// Calculate synthetic tenant size and cache the result.
5610 : /// This is periodically called by background worker.
5611 : /// result is cached in tenant struct
5612 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5613 : pub async fn calculate_synthetic_size(
5614 : &self,
5615 : cause: LogicalSizeCalculationCause,
5616 : cancel: &CancellationToken,
5617 : ctx: &RequestContext,
5618 : ) -> Result<u64, size::CalculateSyntheticSizeError> {
5619 : let inputs = self.gather_size_inputs(None, cause, cancel, ctx).await?;
5620 :
5621 : let size = inputs.calculate();
5622 :
5623 : self.set_cached_synthetic_size(size);
5624 :
5625 : Ok(size)
5626 : }
5627 :
5628 : /// Cache given synthetic size and update the metric value
5629 0 : pub fn set_cached_synthetic_size(&self, size: u64) {
5630 0 : self.cached_synthetic_tenant_size
5631 0 : .store(size, Ordering::Relaxed);
5632 :
5633 : // Only shard zero should be calculating synthetic sizes
5634 0 : debug_assert!(self.shard_identity.is_shard_zero());
5635 :
5636 0 : TENANT_SYNTHETIC_SIZE_METRIC
5637 0 : .get_metric_with_label_values(&[&self.tenant_shard_id.tenant_id.to_string()])
5638 0 : .unwrap()
5639 0 : .set(size);
5640 0 : }
5641 :
5642 0 : pub fn cached_synthetic_size(&self) -> u64 {
5643 0 : self.cached_synthetic_tenant_size.load(Ordering::Relaxed)
5644 0 : }
5645 :
5646 : /// Flush any in-progress layers, schedule uploads, and wait for uploads to complete.
5647 : ///
5648 : /// This function can take a long time: callers should wrap it in a timeout if calling
5649 : /// from an external API handler.
5650 : ///
5651 : /// Cancel-safety: cancelling this function may leave I/O running, but such I/O is
5652 : /// still bounded by tenant/timeline shutdown.
5653 : #[tracing::instrument(skip_all)]
5654 : pub(crate) async fn flush_remote(&self) -> anyhow::Result<()> {
5655 : let timelines = self.timelines.lock().unwrap().clone();
5656 :
5657 0 : async fn flush_timeline(_gate: GateGuard, timeline: Arc<Timeline>) -> anyhow::Result<()> {
5658 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Flushing...");
5659 0 : timeline.freeze_and_flush().await?;
5660 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Waiting for uploads...");
5661 0 : timeline.remote_client.wait_completion().await?;
5662 :
5663 0 : Ok(())
5664 0 : }
5665 :
5666 : // We do not use a JoinSet for these tasks, because we don't want them to be
5667 : // aborted when this function's future is cancelled: they should stay alive
5668 : // holding their GateGuard until they complete, to ensure their I/Os complete
5669 : // before Timeline shutdown completes.
5670 : let mut results = FuturesUnordered::new();
5671 :
5672 : for (_timeline_id, timeline) in timelines {
5673 : // Run each timeline's flush in a task holding the timeline's gate: this
5674 : // means that if this function's future is cancelled, the Timeline shutdown
5675 : // will still wait for any I/O in here to complete.
5676 : let Ok(gate) = timeline.gate.enter() else {
5677 : continue;
5678 : };
5679 0 : let jh = tokio::task::spawn(async move { flush_timeline(gate, timeline).await });
5680 : results.push(jh);
5681 : }
5682 :
5683 : while let Some(r) = results.next().await {
5684 : if let Err(e) = r {
5685 : if !e.is_cancelled() && !e.is_panic() {
5686 : tracing::error!("unexpected join error: {e:?}");
5687 : }
5688 : }
5689 : }
5690 :
5691 : // The flushes we did above were just writes, but the TenantShard might have had
5692 : // pending deletions as well from recent compaction/gc: we want to flush those
5693 : // as well. This requires flushing the global delete queue. This is cheap
5694 : // because it's typically a no-op.
5695 : match self.deletion_queue_client.flush_execute().await {
5696 : Ok(_) => {}
5697 : Err(DeletionQueueError::ShuttingDown) => {}
5698 : }
5699 :
5700 : Ok(())
5701 : }
5702 :
5703 0 : pub(crate) fn get_tenant_conf(&self) -> pageserver_api::models::TenantConfig {
5704 0 : self.tenant_conf.load().tenant_conf.clone()
5705 0 : }
5706 :
5707 : /// How much local storage would this tenant like to have? It can cope with
5708 : /// less than this (via eviction and on-demand downloads), but this function enables
5709 : /// the TenantShard to advertise how much storage it would prefer to have to provide fast I/O
5710 : /// by keeping important things on local disk.
5711 : ///
5712 : /// This is a heuristic, not a guarantee: tenants that are long-idle will actually use less
5713 : /// than they report here, due to layer eviction. Tenants with many active branches may
5714 : /// actually use more than they report here.
5715 0 : pub(crate) fn local_storage_wanted(&self) -> u64 {
5716 0 : let timelines = self.timelines.lock().unwrap();
5717 :
5718 : // Heuristic: we use the max() of the timelines' visible sizes, rather than the sum. This
5719 : // reflects the observation that on tenants with multiple large branches, typically only one
5720 : // of them is used actively enough to occupy space on disk.
5721 0 : timelines
5722 0 : .values()
5723 0 : .map(|t| t.metrics.visible_physical_size_gauge.get())
5724 0 : .max()
5725 0 : .unwrap_or(0)
5726 0 : }
5727 :
5728 : /// HADRON
5729 : /// Return the visible size of all timelines in this tenant.
5730 0 : pub(crate) fn get_visible_size(&self) -> u64 {
5731 0 : let timelines = self.timelines.lock().unwrap();
5732 0 : timelines
5733 0 : .values()
5734 0 : .map(|t| t.metrics.visible_physical_size_gauge.get())
5735 0 : .sum()
5736 0 : }
5737 :
5738 : /// Builds a new tenant manifest, and uploads it if it differs from the last-known tenant
5739 : /// manifest in `Self::remote_tenant_manifest`.
5740 : ///
5741 : /// TODO: instead of requiring callers to remember to call `maybe_upload_tenant_manifest` after
5742 : /// changing any `TenantShard` state that's included in the manifest, consider making the manifest
5743 : /// the authoritative source of data with an API that automatically uploads on changes. Revisit
5744 : /// this when the manifest is more widely used and we have a better idea of the data model.
5745 120 : pub(crate) async fn maybe_upload_tenant_manifest(&self) -> Result<(), TenantManifestError> {
5746 : // Multiple tasks may call this function concurrently after mutating the TenantShard runtime
5747 : // state, affecting the manifest generated by `build_tenant_manifest`. We use an async mutex
5748 : // to serialize these callers. `eq_ignoring_version` acts as a slightly inefficient but
5749 : // simple coalescing mechanism.
5750 120 : let mut guard = tokio::select! {
5751 120 : guard = self.remote_tenant_manifest.lock() => guard,
5752 120 : _ = self.cancel.cancelled() => return Err(TenantManifestError::Cancelled),
5753 : };
5754 :
5755 : // Build a new manifest.
5756 120 : let manifest = self.build_tenant_manifest();
5757 :
5758 : // Check if the manifest has changed. We ignore the version number here, to avoid
5759 : // uploading every manifest on version number bumps.
5760 120 : if let Some(old) = guard.as_ref() {
5761 4 : if manifest.eq_ignoring_version(old) {
5762 3 : return Ok(());
5763 1 : }
5764 116 : }
5765 :
5766 : // Update metrics
5767 117 : let tid = self.tenant_shard_id.to_string();
5768 117 : let shard_id = self.tenant_shard_id.shard_slug().to_string();
5769 117 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
5770 117 : TENANT_OFFLOADED_TIMELINES
5771 117 : .with_label_values(set_key)
5772 117 : .set(manifest.offloaded_timelines.len() as u64);
5773 :
5774 : // Upload the manifest. Remote storage does no retries internally, so retry here.
5775 117 : match backoff::retry(
5776 117 : || async {
5777 117 : upload_tenant_manifest(
5778 117 : &self.remote_storage,
5779 117 : &self.tenant_shard_id,
5780 117 : self.generation,
5781 117 : &manifest,
5782 117 : &self.cancel,
5783 117 : )
5784 117 : .await
5785 234 : },
5786 0 : |_| self.cancel.is_cancelled(),
5787 : FAILED_UPLOAD_WARN_THRESHOLD,
5788 : FAILED_REMOTE_OP_RETRIES,
5789 117 : "uploading tenant manifest",
5790 117 : &self.cancel,
5791 : )
5792 117 : .await
5793 : {
5794 0 : None => Err(TenantManifestError::Cancelled),
5795 0 : Some(Err(_)) if self.cancel.is_cancelled() => Err(TenantManifestError::Cancelled),
5796 0 : Some(Err(e)) => Err(TenantManifestError::RemoteStorage(e)),
5797 : Some(Ok(_)) => {
5798 : // Store the successfully uploaded manifest, so that future callers can avoid
5799 : // re-uploading the same thing.
5800 117 : *guard = Some(manifest);
5801 :
5802 117 : Ok(())
5803 : }
5804 : }
5805 120 : }
5806 : }
5807 :
5808 : /// Create the cluster temporarily in 'initdbpath' directory inside the repository
5809 : /// to get bootstrap data for timeline initialization.
5810 0 : async fn run_initdb(
5811 0 : conf: &'static PageServerConf,
5812 0 : initdb_target_dir: &Utf8Path,
5813 0 : pg_version: PgMajorVersion,
5814 0 : cancel: &CancellationToken,
5815 0 : ) -> Result<(), InitdbError> {
5816 0 : let initdb_bin_path = conf
5817 0 : .pg_bin_dir(pg_version)
5818 0 : .map_err(InitdbError::Other)?
5819 0 : .join("initdb");
5820 0 : let initdb_lib_dir = conf.pg_lib_dir(pg_version).map_err(InitdbError::Other)?;
5821 0 : info!(
5822 0 : "running {} in {}, libdir: {}",
5823 : initdb_bin_path, initdb_target_dir, initdb_lib_dir,
5824 : );
5825 :
5826 0 : let _permit = {
5827 0 : let _timer = INITDB_SEMAPHORE_ACQUISITION_TIME.start_timer();
5828 0 : INIT_DB_SEMAPHORE.acquire().await
5829 : };
5830 :
5831 0 : CONCURRENT_INITDBS.inc();
5832 0 : scopeguard::defer! {
5833 : CONCURRENT_INITDBS.dec();
5834 : }
5835 :
5836 0 : let _timer = INITDB_RUN_TIME.start_timer();
5837 0 : let res = postgres_initdb::do_run_initdb(postgres_initdb::RunInitdbArgs {
5838 0 : superuser: &conf.superuser,
5839 0 : locale: &conf.locale,
5840 0 : initdb_bin: &initdb_bin_path,
5841 0 : pg_version,
5842 0 : library_search_path: &initdb_lib_dir,
5843 0 : pgdata: initdb_target_dir,
5844 0 : })
5845 0 : .await
5846 0 : .map_err(InitdbError::Inner);
5847 :
5848 : // This isn't true cancellation support, see above. Still return an error to
5849 : // excercise the cancellation code path.
5850 0 : if cancel.is_cancelled() {
5851 0 : return Err(InitdbError::Cancelled);
5852 0 : }
5853 :
5854 0 : res
5855 0 : }
5856 :
5857 : /// Dump contents of a layer file to stdout.
5858 0 : pub async fn dump_layerfile_from_path(
5859 0 : path: &Utf8Path,
5860 0 : verbose: bool,
5861 0 : ctx: &RequestContext,
5862 0 : ) -> anyhow::Result<()> {
5863 : use std::os::unix::fs::FileExt;
5864 :
5865 : // All layer files start with a two-byte "magic" value, to identify the kind of
5866 : // file.
5867 0 : let file = File::open(path)?;
5868 0 : let mut header_buf = [0u8; 2];
5869 0 : file.read_exact_at(&mut header_buf, 0)?;
5870 :
5871 0 : match u16::from_be_bytes(header_buf) {
5872 : crate::IMAGE_FILE_MAGIC => {
5873 0 : ImageLayer::new_for_path(path, file)?
5874 0 : .dump(verbose, ctx)
5875 0 : .await?
5876 : }
5877 : crate::DELTA_FILE_MAGIC => {
5878 0 : DeltaLayer::new_for_path(path, file)?
5879 0 : .dump(verbose, ctx)
5880 0 : .await?
5881 : }
5882 0 : magic => bail!("unrecognized magic identifier: {:?}", magic),
5883 : }
5884 :
5885 0 : Ok(())
5886 0 : }
5887 :
5888 : #[cfg(test)]
5889 : pub(crate) mod harness {
5890 : use bytes::{Bytes, BytesMut};
5891 : use hex_literal::hex;
5892 : use once_cell::sync::OnceCell;
5893 : use pageserver_api::key::Key;
5894 : use pageserver_api::models::ShardParameters;
5895 : use pageserver_api::shard::ShardIndex;
5896 : use utils::id::TenantId;
5897 : use utils::logging;
5898 : use wal_decoder::models::record::NeonWalRecord;
5899 :
5900 : use super::*;
5901 : use crate::deletion_queue::mock::MockDeletionQueue;
5902 : use crate::l0_flush::L0FlushConfig;
5903 : use crate::walredo::apply_neon;
5904 :
5905 : pub const TIMELINE_ID: TimelineId =
5906 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
5907 : pub const NEW_TIMELINE_ID: TimelineId =
5908 : TimelineId::from_array(hex!("AA223344556677881122334455667788"));
5909 :
5910 : /// Convenience function to create a page image with given string as the only content
5911 2514385 : pub fn test_img(s: &str) -> Bytes {
5912 2514385 : let mut buf = BytesMut::new();
5913 2514385 : buf.extend_from_slice(s.as_bytes());
5914 2514385 : buf.resize(64, 0);
5915 :
5916 2514385 : buf.freeze()
5917 2514385 : }
5918 :
5919 : pub struct TenantHarness {
5920 : pub conf: &'static PageServerConf,
5921 : pub tenant_conf: pageserver_api::models::TenantConfig,
5922 : pub tenant_shard_id: TenantShardId,
5923 : pub shard_identity: ShardIdentity,
5924 : pub generation: Generation,
5925 : pub shard: ShardIndex,
5926 : pub remote_storage: GenericRemoteStorage,
5927 : pub remote_fs_dir: Utf8PathBuf,
5928 : pub deletion_queue: MockDeletionQueue,
5929 : }
5930 :
5931 : static LOG_HANDLE: OnceCell<()> = OnceCell::new();
5932 :
5933 131 : pub(crate) fn setup_logging() {
5934 131 : LOG_HANDLE.get_or_init(|| {
5935 125 : logging::init(
5936 125 : logging::LogFormat::Test,
5937 : // enable it in case the tests exercise code paths that use
5938 : // debug_assert_current_span_has_tenant_and_timeline_id
5939 125 : logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
5940 125 : logging::Output::Stdout,
5941 : )
5942 125 : .expect("Failed to init test logging");
5943 125 : });
5944 131 : }
5945 :
5946 : impl TenantHarness {
5947 119 : pub async fn create_custom(
5948 119 : test_name: &'static str,
5949 119 : tenant_conf: pageserver_api::models::TenantConfig,
5950 119 : tenant_id: TenantId,
5951 119 : shard_identity: ShardIdentity,
5952 119 : generation: Generation,
5953 119 : ) -> anyhow::Result<Self> {
5954 119 : setup_logging();
5955 :
5956 119 : let repo_dir = PageServerConf::test_repo_dir(test_name);
5957 119 : let _ = fs::remove_dir_all(&repo_dir);
5958 119 : fs::create_dir_all(&repo_dir)?;
5959 :
5960 119 : let conf = PageServerConf::dummy_conf(repo_dir);
5961 : // Make a static copy of the config. This can never be free'd, but that's
5962 : // OK in a test.
5963 119 : let conf: &'static PageServerConf = Box::leak(Box::new(conf));
5964 :
5965 119 : let shard = shard_identity.shard_index();
5966 119 : let tenant_shard_id = TenantShardId {
5967 119 : tenant_id,
5968 119 : shard_number: shard.shard_number,
5969 119 : shard_count: shard.shard_count,
5970 119 : };
5971 119 : fs::create_dir_all(conf.tenant_path(&tenant_shard_id))?;
5972 119 : fs::create_dir_all(conf.timelines_path(&tenant_shard_id))?;
5973 :
5974 : use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
5975 119 : let remote_fs_dir = conf.workdir.join("localfs");
5976 119 : std::fs::create_dir_all(&remote_fs_dir).unwrap();
5977 119 : let config = RemoteStorageConfig {
5978 119 : storage: RemoteStorageKind::LocalFs {
5979 119 : local_path: remote_fs_dir.clone(),
5980 119 : },
5981 119 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
5982 119 : small_timeout: RemoteStorageConfig::DEFAULT_SMALL_TIMEOUT,
5983 119 : };
5984 119 : let remote_storage = GenericRemoteStorage::from_config(&config).await.unwrap();
5985 119 : let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()));
5986 :
5987 119 : Ok(Self {
5988 119 : conf,
5989 119 : tenant_conf,
5990 119 : tenant_shard_id,
5991 119 : shard_identity,
5992 119 : generation,
5993 119 : shard,
5994 119 : remote_storage,
5995 119 : remote_fs_dir,
5996 119 : deletion_queue,
5997 119 : })
5998 119 : }
5999 :
6000 110 : pub async fn create(test_name: &'static str) -> anyhow::Result<Self> {
6001 : // Disable automatic GC and compaction to make the unit tests more deterministic.
6002 : // The tests perform them manually if needed.
6003 110 : let tenant_conf = pageserver_api::models::TenantConfig {
6004 110 : gc_period: Some(Duration::ZERO),
6005 110 : compaction_period: Some(Duration::ZERO),
6006 110 : ..Default::default()
6007 110 : };
6008 110 : let tenant_id = TenantId::generate();
6009 110 : let shard = ShardIdentity::unsharded();
6010 110 : Self::create_custom(
6011 110 : test_name,
6012 110 : tenant_conf,
6013 110 : tenant_id,
6014 110 : shard,
6015 110 : Generation::new(0xdeadbeef),
6016 110 : )
6017 110 : .await
6018 110 : }
6019 :
6020 10 : pub fn span(&self) -> tracing::Span {
6021 10 : info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
6022 10 : }
6023 :
6024 119 : pub(crate) async fn load(&self) -> (Arc<TenantShard>, RequestContext) {
6025 119 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error)
6026 119 : .with_scope_unit_test();
6027 : (
6028 119 : self.do_try_load(&ctx)
6029 119 : .await
6030 119 : .expect("failed to load test tenant"),
6031 119 : ctx,
6032 : )
6033 119 : }
6034 :
6035 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
6036 : pub(crate) async fn do_try_load_with_redo(
6037 : &self,
6038 : walredo_mgr: Arc<WalRedoManager>,
6039 : ctx: &RequestContext,
6040 : ) -> anyhow::Result<Arc<TenantShard>> {
6041 : let (basebackup_cache, _) = BasebackupCache::new(Utf8PathBuf::new(), None);
6042 :
6043 : let tenant = Arc::new(TenantShard::new(
6044 : TenantState::Attaching,
6045 : self.conf,
6046 : AttachedTenantConf::try_from(
6047 : self.conf,
6048 : LocationConf::attached_single(
6049 : self.tenant_conf.clone(),
6050 : self.generation,
6051 : ShardParameters::default(),
6052 : ),
6053 : )
6054 : .unwrap(),
6055 : self.shard_identity,
6056 : Some(walredo_mgr),
6057 : self.tenant_shard_id,
6058 : self.remote_storage.clone(),
6059 : self.deletion_queue.new_client(),
6060 : // TODO: ideally we should run all unit tests with both configs
6061 : L0FlushGlobalState::new(L0FlushConfig::default()),
6062 : basebackup_cache,
6063 : FeatureResolver::new_disabled(),
6064 : ));
6065 :
6066 : let preload = tenant
6067 : .preload(&self.remote_storage, CancellationToken::new())
6068 : .await?;
6069 : tenant.attach(Some(preload), ctx).await?;
6070 :
6071 : tenant.state.send_replace(TenantState::Active);
6072 : for timeline in tenant.timelines.lock().unwrap().values() {
6073 : timeline.set_state(TimelineState::Active);
6074 : }
6075 : Ok(tenant)
6076 : }
6077 :
6078 119 : pub(crate) async fn do_try_load(
6079 119 : &self,
6080 119 : ctx: &RequestContext,
6081 119 : ) -> anyhow::Result<Arc<TenantShard>> {
6082 119 : let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
6083 119 : self.do_try_load_with_redo(walredo_mgr, ctx).await
6084 119 : }
6085 :
6086 1 : pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
6087 1 : self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
6088 1 : }
6089 : }
6090 :
6091 : // Mock WAL redo manager that doesn't do much
6092 : pub(crate) struct TestRedoManager;
6093 :
6094 : impl TestRedoManager {
6095 : /// # Cancel-Safety
6096 : ///
6097 : /// This method is cancellation-safe.
6098 26774 : pub async fn request_redo(
6099 26774 : &self,
6100 26774 : key: Key,
6101 26774 : lsn: Lsn,
6102 26774 : base_img: Option<(Lsn, Bytes)>,
6103 26774 : records: Vec<(Lsn, NeonWalRecord)>,
6104 26774 : _pg_version: PgMajorVersion,
6105 26774 : _redo_attempt_type: RedoAttemptType,
6106 26774 : ) -> Result<Bytes, walredo::Error> {
6107 1403510 : let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
6108 26774 : if records_neon {
6109 : // For Neon wal records, we can decode without spawning postgres, so do so.
6110 26774 : let mut page = match (base_img, records.first()) {
6111 13029 : (Some((_lsn, img)), _) => {
6112 13029 : let mut page = BytesMut::new();
6113 13029 : page.extend_from_slice(&img);
6114 13029 : page
6115 : }
6116 13745 : (_, Some((_lsn, rec))) if rec.will_init() => BytesMut::new(),
6117 : _ => {
6118 0 : panic!("Neon WAL redo requires base image or will init record");
6119 : }
6120 : };
6121 :
6122 1430283 : for (record_lsn, record) in records {
6123 1403510 : apply_neon::apply_in_neon(&record, record_lsn, key, &mut page)?;
6124 : }
6125 26773 : Ok(page.freeze())
6126 : } else {
6127 : // We never spawn a postgres walredo process in unit tests: just log what we might have done.
6128 0 : let s = format!(
6129 0 : "redo for {} to get to {}, with {} and {} records",
6130 : key,
6131 : lsn,
6132 0 : if base_img.is_some() {
6133 0 : "base image"
6134 : } else {
6135 0 : "no base image"
6136 : },
6137 0 : records.len()
6138 : );
6139 0 : println!("{s}");
6140 :
6141 0 : Ok(test_img(&s))
6142 : }
6143 26774 : }
6144 : }
6145 : }
6146 :
6147 : #[cfg(test)]
6148 : mod tests {
6149 : use std::collections::{BTreeMap, BTreeSet};
6150 :
6151 : use bytes::{Bytes, BytesMut};
6152 : use hex_literal::hex;
6153 : use itertools::Itertools;
6154 : #[cfg(feature = "testing")]
6155 : use models::CompactLsnRange;
6156 : use pageserver_api::key::{
6157 : AUX_KEY_PREFIX, Key, NON_INHERITED_RANGE, RELATION_SIZE_PREFIX, repl_origin_key,
6158 : };
6159 : use pageserver_api::keyspace::KeySpace;
6160 : #[cfg(feature = "testing")]
6161 : use pageserver_api::keyspace::KeySpaceRandomAccum;
6162 : use pageserver_api::models::{CompactionAlgorithm, CompactionAlgorithmSettings, LsnLease};
6163 : use pageserver_compaction::helpers::overlaps_with;
6164 : #[cfg(feature = "testing")]
6165 : use rand::SeedableRng;
6166 : #[cfg(feature = "testing")]
6167 : use rand::rngs::StdRng;
6168 : use rand::{Rng, thread_rng};
6169 : #[cfg(feature = "testing")]
6170 : use std::ops::Range;
6171 : use storage_layer::{IoConcurrency, PersistentLayerKey};
6172 : use tests::storage_layer::ValuesReconstructState;
6173 : use tests::timeline::{GetVectoredError, ShutdownMode};
6174 : #[cfg(feature = "testing")]
6175 : use timeline::GcInfo;
6176 : #[cfg(feature = "testing")]
6177 : use timeline::InMemoryLayerTestDesc;
6178 : #[cfg(feature = "testing")]
6179 : use timeline::compaction::{KeyHistoryRetention, KeyLogAtLsn};
6180 : use timeline::{CompactOptions, DeltaLayerTestDesc, VersionedKeySpaceQuery};
6181 : use utils::id::TenantId;
6182 : use utils::shard::{ShardCount, ShardNumber};
6183 : #[cfg(feature = "testing")]
6184 : use wal_decoder::models::record::NeonWalRecord;
6185 : use wal_decoder::models::value::Value;
6186 :
6187 : use super::*;
6188 : use crate::DEFAULT_PG_VERSION;
6189 : use crate::keyspace::KeySpaceAccum;
6190 : use crate::tenant::harness::*;
6191 : use crate::tenant::timeline::CompactFlags;
6192 :
6193 : static TEST_KEY: Lazy<Key> =
6194 10 : Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));
6195 :
6196 : #[cfg(feature = "testing")]
6197 : struct TestTimelineSpecification {
6198 : start_lsn: Lsn,
6199 : last_record_lsn: Lsn,
6200 :
6201 : in_memory_layers_shape: Vec<(Range<Key>, Range<Lsn>)>,
6202 : delta_layers_shape: Vec<(Range<Key>, Range<Lsn>)>,
6203 : image_layers_shape: Vec<(Range<Key>, Lsn)>,
6204 :
6205 : gap_chance: u8,
6206 : will_init_chance: u8,
6207 : }
6208 :
6209 : #[cfg(feature = "testing")]
6210 : struct Storage {
6211 : storage: HashMap<(Key, Lsn), Value>,
6212 : start_lsn: Lsn,
6213 : }
6214 :
6215 : #[cfg(feature = "testing")]
6216 : impl Storage {
6217 32000 : fn get(&self, key: Key, lsn: Lsn) -> Bytes {
6218 : use bytes::BufMut;
6219 :
6220 32000 : let mut crnt_lsn = lsn;
6221 32000 : let mut got_base = false;
6222 :
6223 32000 : let mut acc = Vec::new();
6224 :
6225 2831871 : while crnt_lsn >= self.start_lsn {
6226 2831871 : if let Some(value) = self.storage.get(&(key, crnt_lsn)) {
6227 1421172 : acc.push(value.clone());
6228 :
6229 1402881 : match value {
6230 1402881 : Value::WalRecord(NeonWalRecord::Test { will_init, .. }) => {
6231 1402881 : if *will_init {
6232 13709 : got_base = true;
6233 13709 : break;
6234 1389172 : }
6235 : }
6236 : Value::Image(_) => {
6237 18291 : got_base = true;
6238 18291 : break;
6239 : }
6240 0 : _ => unreachable!(),
6241 : }
6242 1410699 : }
6243 :
6244 2799871 : crnt_lsn = crnt_lsn.checked_sub(1u64).unwrap();
6245 : }
6246 :
6247 32000 : assert!(
6248 32000 : got_base,
6249 0 : "Input data was incorrect. No base image for {key}@{lsn}"
6250 : );
6251 :
6252 32000 : tracing::debug!("Wal redo depth for {key}@{lsn} is {}", acc.len());
6253 :
6254 32000 : let mut blob = BytesMut::new();
6255 1421172 : for value in acc.into_iter().rev() {
6256 1402881 : match value {
6257 1402881 : Value::WalRecord(NeonWalRecord::Test { append, .. }) => {
6258 1402881 : blob.extend_from_slice(append.as_bytes());
6259 1402881 : }
6260 18291 : Value::Image(img) => {
6261 18291 : blob.put(img);
6262 18291 : }
6263 0 : _ => unreachable!(),
6264 : }
6265 : }
6266 :
6267 32000 : blob.into()
6268 32000 : }
6269 : }
6270 :
6271 : #[cfg(feature = "testing")]
6272 : #[allow(clippy::too_many_arguments)]
6273 1 : async fn randomize_timeline(
6274 1 : tenant: &Arc<TenantShard>,
6275 1 : new_timeline_id: TimelineId,
6276 1 : pg_version: PgMajorVersion,
6277 1 : spec: TestTimelineSpecification,
6278 1 : random: &mut rand::rngs::StdRng,
6279 1 : ctx: &RequestContext,
6280 1 : ) -> anyhow::Result<(Arc<Timeline>, Storage, Vec<Lsn>)> {
6281 1 : let mut storage: HashMap<(Key, Lsn), Value> = HashMap::default();
6282 1 : let mut interesting_lsns = vec![spec.last_record_lsn];
6283 :
6284 2 : for (key_range, lsn_range) in spec.in_memory_layers_shape.iter() {
6285 2 : let mut lsn = lsn_range.start;
6286 202 : while lsn < lsn_range.end {
6287 200 : let mut key = key_range.start;
6288 21018 : while key < key_range.end {
6289 20818 : let gap = random.gen_range(1..=100) <= spec.gap_chance;
6290 20818 : let will_init = random.gen_range(1..=100) <= spec.will_init_chance;
6291 :
6292 20818 : if gap {
6293 1018 : continue;
6294 19800 : }
6295 :
6296 19800 : let record = if will_init {
6297 191 : Value::WalRecord(NeonWalRecord::wal_init(format!("[wil_init {key}@{lsn}]")))
6298 : } else {
6299 19609 : Value::WalRecord(NeonWalRecord::wal_append(format!("[delta {key}@{lsn}]")))
6300 : };
6301 :
6302 19800 : storage.insert((key, lsn), record);
6303 :
6304 19800 : key = key.next();
6305 : }
6306 200 : lsn = Lsn(lsn.0 + 1);
6307 : }
6308 :
6309 : // Stash some interesting LSN for future use
6310 6 : for offset in [0, 5, 100].iter() {
6311 6 : if *offset == 0 {
6312 2 : interesting_lsns.push(lsn_range.start);
6313 2 : } else {
6314 4 : let below = lsn_range.start.checked_sub(*offset);
6315 4 : match below {
6316 4 : Some(v) if v >= spec.start_lsn => {
6317 4 : interesting_lsns.push(v);
6318 4 : }
6319 0 : _ => {}
6320 : }
6321 :
6322 4 : let above = Lsn(lsn_range.start.0 + offset);
6323 4 : interesting_lsns.push(above);
6324 : }
6325 : }
6326 : }
6327 :
6328 3 : for (key_range, lsn_range) in spec.delta_layers_shape.iter() {
6329 3 : let mut lsn = lsn_range.start;
6330 315 : while lsn < lsn_range.end {
6331 312 : let mut key = key_range.start;
6332 11112 : while key < key_range.end {
6333 10800 : let gap = random.gen_range(1..=100) <= spec.gap_chance;
6334 10800 : let will_init = random.gen_range(1..=100) <= spec.will_init_chance;
6335 :
6336 10800 : if gap {
6337 504 : continue;
6338 10296 : }
6339 :
6340 10296 : let record = if will_init {
6341 103 : Value::WalRecord(NeonWalRecord::wal_init(format!("[wil_init {key}@{lsn}]")))
6342 : } else {
6343 10193 : Value::WalRecord(NeonWalRecord::wal_append(format!("[delta {key}@{lsn}]")))
6344 : };
6345 :
6346 10296 : storage.insert((key, lsn), record);
6347 :
6348 10296 : key = key.next();
6349 : }
6350 312 : lsn = Lsn(lsn.0 + 1);
6351 : }
6352 :
6353 : // Stash some interesting LSN for future use
6354 9 : for offset in [0, 5, 100].iter() {
6355 9 : if *offset == 0 {
6356 3 : interesting_lsns.push(lsn_range.start);
6357 3 : } else {
6358 6 : let below = lsn_range.start.checked_sub(*offset);
6359 6 : match below {
6360 6 : Some(v) if v >= spec.start_lsn => {
6361 3 : interesting_lsns.push(v);
6362 3 : }
6363 3 : _ => {}
6364 : }
6365 :
6366 6 : let above = Lsn(lsn_range.start.0 + offset);
6367 6 : interesting_lsns.push(above);
6368 : }
6369 : }
6370 : }
6371 :
6372 3 : for (key_range, lsn) in spec.image_layers_shape.iter() {
6373 3 : let mut key = key_range.start;
6374 142 : while key < key_range.end {
6375 139 : let blob = Bytes::from(format!("[image {key}@{lsn}]"));
6376 139 : let record = Value::Image(blob.clone());
6377 139 : storage.insert((key, *lsn), record);
6378 139 :
6379 139 : key = key.next();
6380 139 : }
6381 :
6382 : // Stash some interesting LSN for future use
6383 9 : for offset in [0, 5, 100].iter() {
6384 9 : if *offset == 0 {
6385 3 : interesting_lsns.push(*lsn);
6386 3 : } else {
6387 6 : let below = lsn.checked_sub(*offset);
6388 6 : match below {
6389 6 : Some(v) if v >= spec.start_lsn => {
6390 4 : interesting_lsns.push(v);
6391 4 : }
6392 2 : _ => {}
6393 : }
6394 :
6395 6 : let above = Lsn(lsn.0 + offset);
6396 6 : interesting_lsns.push(above);
6397 : }
6398 : }
6399 : }
6400 :
6401 1 : let in_memory_test_layers = {
6402 1 : let mut acc = Vec::new();
6403 :
6404 2 : for (key_range, lsn_range) in spec.in_memory_layers_shape.iter() {
6405 2 : let mut data = Vec::new();
6406 :
6407 2 : let mut lsn = lsn_range.start;
6408 202 : while lsn < lsn_range.end {
6409 200 : let mut key = key_range.start;
6410 20000 : while key < key_range.end {
6411 19800 : if let Some(record) = storage.get(&(key, lsn)) {
6412 19800 : data.push((key, lsn, record.clone()));
6413 19800 : }
6414 :
6415 19800 : key = key.next();
6416 : }
6417 200 : lsn = Lsn(lsn.0 + 1);
6418 : }
6419 :
6420 2 : acc.push(InMemoryLayerTestDesc {
6421 2 : data,
6422 2 : lsn_range: lsn_range.clone(),
6423 2 : is_open: false,
6424 2 : })
6425 : }
6426 :
6427 1 : acc
6428 : };
6429 :
6430 1 : let delta_test_layers = {
6431 1 : let mut acc = Vec::new();
6432 :
6433 3 : for (key_range, lsn_range) in spec.delta_layers_shape.iter() {
6434 3 : let mut data = Vec::new();
6435 :
6436 3 : let mut lsn = lsn_range.start;
6437 315 : while lsn < lsn_range.end {
6438 312 : let mut key = key_range.start;
6439 10608 : while key < key_range.end {
6440 10296 : if let Some(record) = storage.get(&(key, lsn)) {
6441 10296 : data.push((key, lsn, record.clone()));
6442 10296 : }
6443 :
6444 10296 : key = key.next();
6445 : }
6446 312 : lsn = Lsn(lsn.0 + 1);
6447 : }
6448 :
6449 3 : acc.push(DeltaLayerTestDesc {
6450 3 : data,
6451 3 : lsn_range: lsn_range.clone(),
6452 3 : key_range: key_range.clone(),
6453 3 : })
6454 : }
6455 :
6456 1 : acc
6457 : };
6458 :
6459 1 : let image_test_layers = {
6460 1 : let mut acc = Vec::new();
6461 :
6462 3 : for (key_range, lsn) in spec.image_layers_shape.iter() {
6463 3 : let mut data = Vec::new();
6464 :
6465 3 : let mut key = key_range.start;
6466 142 : while key < key_range.end {
6467 139 : if let Some(record) = storage.get(&(key, *lsn)) {
6468 139 : let blob = match record {
6469 139 : Value::Image(blob) => blob.clone(),
6470 0 : _ => unreachable!(),
6471 : };
6472 :
6473 139 : data.push((key, blob));
6474 0 : }
6475 :
6476 139 : key = key.next();
6477 : }
6478 :
6479 3 : acc.push((*lsn, data));
6480 : }
6481 :
6482 1 : acc
6483 : };
6484 :
6485 1 : let tline = tenant
6486 1 : .create_test_timeline_with_layers(
6487 1 : new_timeline_id,
6488 1 : spec.start_lsn,
6489 1 : pg_version,
6490 1 : ctx,
6491 1 : in_memory_test_layers,
6492 1 : delta_test_layers,
6493 1 : image_test_layers,
6494 1 : spec.last_record_lsn,
6495 1 : )
6496 1 : .await?;
6497 :
6498 1 : Ok((
6499 1 : tline,
6500 1 : Storage {
6501 1 : storage,
6502 1 : start_lsn: spec.start_lsn,
6503 1 : },
6504 1 : interesting_lsns,
6505 1 : ))
6506 1 : }
6507 :
6508 : #[tokio::test]
6509 1 : async fn test_basic() -> anyhow::Result<()> {
6510 1 : let (tenant, ctx) = TenantHarness::create("test_basic").await?.load().await;
6511 1 : let tline = tenant
6512 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6513 1 : .await?;
6514 :
6515 1 : let mut writer = tline.writer().await;
6516 1 : writer
6517 1 : .put(
6518 1 : *TEST_KEY,
6519 1 : Lsn(0x10),
6520 1 : &Value::Image(test_img("foo at 0x10")),
6521 1 : &ctx,
6522 1 : )
6523 1 : .await?;
6524 1 : writer.finish_write(Lsn(0x10));
6525 1 : drop(writer);
6526 :
6527 1 : let mut writer = tline.writer().await;
6528 1 : writer
6529 1 : .put(
6530 1 : *TEST_KEY,
6531 1 : Lsn(0x20),
6532 1 : &Value::Image(test_img("foo at 0x20")),
6533 1 : &ctx,
6534 1 : )
6535 1 : .await?;
6536 1 : writer.finish_write(Lsn(0x20));
6537 1 : drop(writer);
6538 :
6539 1 : assert_eq!(
6540 1 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
6541 1 : test_img("foo at 0x10")
6542 : );
6543 1 : assert_eq!(
6544 1 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
6545 1 : test_img("foo at 0x10")
6546 : );
6547 1 : assert_eq!(
6548 1 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
6549 1 : test_img("foo at 0x20")
6550 : );
6551 :
6552 2 : Ok(())
6553 1 : }
6554 :
6555 : #[tokio::test]
6556 1 : async fn no_duplicate_timelines() -> anyhow::Result<()> {
6557 1 : let (tenant, ctx) = TenantHarness::create("no_duplicate_timelines")
6558 1 : .await?
6559 1 : .load()
6560 1 : .await;
6561 1 : let _ = tenant
6562 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6563 1 : .await?;
6564 :
6565 1 : match tenant
6566 1 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6567 1 : .await
6568 1 : {
6569 1 : Ok(_) => panic!("duplicate timeline creation should fail"),
6570 1 : Err(e) => assert_eq!(
6571 1 : e.to_string(),
6572 1 : "timeline already exists with different parameters".to_string()
6573 1 : ),
6574 1 : }
6575 1 :
6576 1 : Ok(())
6577 1 : }
6578 :
6579 : /// Convenience function to create a page image with given string as the only content
6580 5 : pub fn test_value(s: &str) -> Value {
6581 5 : let mut buf = BytesMut::new();
6582 5 : buf.extend_from_slice(s.as_bytes());
6583 5 : Value::Image(buf.freeze())
6584 5 : }
6585 :
6586 : ///
6587 : /// Test branch creation
6588 : ///
6589 : #[tokio::test]
6590 1 : async fn test_branch() -> anyhow::Result<()> {
6591 : use std::str::from_utf8;
6592 :
6593 1 : let (tenant, ctx) = TenantHarness::create("test_branch").await?.load().await;
6594 1 : let tline = tenant
6595 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6596 1 : .await?;
6597 1 : let mut writer = tline.writer().await;
6598 :
6599 : #[allow(non_snake_case)]
6600 1 : let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
6601 : #[allow(non_snake_case)]
6602 1 : let TEST_KEY_B: Key = Key::from_hex("110000000033333333444444445500000002").unwrap();
6603 :
6604 : // Insert a value on the timeline
6605 1 : writer
6606 1 : .put(TEST_KEY_A, Lsn(0x20), &test_value("foo at 0x20"), &ctx)
6607 1 : .await?;
6608 1 : writer
6609 1 : .put(TEST_KEY_B, Lsn(0x20), &test_value("foobar at 0x20"), &ctx)
6610 1 : .await?;
6611 1 : writer.finish_write(Lsn(0x20));
6612 :
6613 1 : writer
6614 1 : .put(TEST_KEY_A, Lsn(0x30), &test_value("foo at 0x30"), &ctx)
6615 1 : .await?;
6616 1 : writer.finish_write(Lsn(0x30));
6617 1 : writer
6618 1 : .put(TEST_KEY_A, Lsn(0x40), &test_value("foo at 0x40"), &ctx)
6619 1 : .await?;
6620 1 : writer.finish_write(Lsn(0x40));
6621 :
6622 : //assert_current_logical_size(&tline, Lsn(0x40));
6623 :
6624 : // Branch the history, modify relation differently on the new timeline
6625 1 : tenant
6626 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x30)), &ctx)
6627 1 : .await?;
6628 1 : let newtline = tenant
6629 1 : .get_timeline(NEW_TIMELINE_ID, true)
6630 1 : .expect("Should have a local timeline");
6631 1 : let mut new_writer = newtline.writer().await;
6632 1 : new_writer
6633 1 : .put(TEST_KEY_A, Lsn(0x40), &test_value("bar at 0x40"), &ctx)
6634 1 : .await?;
6635 1 : new_writer.finish_write(Lsn(0x40));
6636 :
6637 : // Check page contents on both branches
6638 1 : assert_eq!(
6639 1 : from_utf8(&tline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
6640 : "foo at 0x40"
6641 : );
6642 1 : assert_eq!(
6643 1 : from_utf8(&newtline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
6644 : "bar at 0x40"
6645 : );
6646 1 : assert_eq!(
6647 1 : from_utf8(&newtline.get(TEST_KEY_B, Lsn(0x40), &ctx).await?)?,
6648 : "foobar at 0x20"
6649 : );
6650 :
6651 : //assert_current_logical_size(&tline, Lsn(0x40));
6652 :
6653 2 : Ok(())
6654 1 : }
6655 :
6656 10 : async fn make_some_layers(
6657 10 : tline: &Timeline,
6658 10 : start_lsn: Lsn,
6659 10 : ctx: &RequestContext,
6660 10 : ) -> anyhow::Result<()> {
6661 10 : let mut lsn = start_lsn;
6662 : {
6663 10 : let mut writer = tline.writer().await;
6664 : // Create a relation on the timeline
6665 10 : writer
6666 10 : .put(
6667 10 : *TEST_KEY,
6668 10 : lsn,
6669 10 : &Value::Image(test_img(&format!("foo at {lsn}"))),
6670 10 : ctx,
6671 10 : )
6672 10 : .await?;
6673 10 : writer.finish_write(lsn);
6674 10 : lsn += 0x10;
6675 10 : writer
6676 10 : .put(
6677 10 : *TEST_KEY,
6678 10 : lsn,
6679 10 : &Value::Image(test_img(&format!("foo at {lsn}"))),
6680 10 : ctx,
6681 10 : )
6682 10 : .await?;
6683 10 : writer.finish_write(lsn);
6684 10 : lsn += 0x10;
6685 : }
6686 10 : tline.freeze_and_flush().await?;
6687 : {
6688 10 : let mut writer = tline.writer().await;
6689 10 : writer
6690 10 : .put(
6691 10 : *TEST_KEY,
6692 10 : lsn,
6693 10 : &Value::Image(test_img(&format!("foo at {lsn}"))),
6694 10 : ctx,
6695 10 : )
6696 10 : .await?;
6697 10 : writer.finish_write(lsn);
6698 10 : lsn += 0x10;
6699 10 : writer
6700 10 : .put(
6701 10 : *TEST_KEY,
6702 10 : lsn,
6703 10 : &Value::Image(test_img(&format!("foo at {lsn}"))),
6704 10 : ctx,
6705 10 : )
6706 10 : .await?;
6707 10 : writer.finish_write(lsn);
6708 : }
6709 10 : tline.freeze_and_flush().await.map_err(|e| e.into())
6710 10 : }
6711 :
6712 : #[tokio::test]
6713 1 : async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
6714 1 : let (tenant, ctx) =
6715 1 : TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
6716 1 : .await?
6717 1 : .load()
6718 1 : .await;
6719 1 : let tline = tenant
6720 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6721 1 : .await?;
6722 1 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6723 :
6724 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
6725 : // FIXME: this doesn't actually remove any layer currently, given how the flushing
6726 : // and compaction works. But it does set the 'cutoff' point so that the cross check
6727 : // below should fail.
6728 1 : tenant
6729 1 : .gc_iteration(
6730 1 : Some(TIMELINE_ID),
6731 1 : 0x10,
6732 1 : Duration::ZERO,
6733 1 : &CancellationToken::new(),
6734 1 : &ctx,
6735 1 : )
6736 1 : .await?;
6737 :
6738 : // try to branch at lsn 25, should fail because we already garbage collected the data
6739 1 : match tenant
6740 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
6741 1 : .await
6742 1 : {
6743 1 : Ok(_) => panic!("branching should have failed"),
6744 1 : Err(err) => {
6745 1 : let CreateTimelineError::AncestorLsn(err) = err else {
6746 1 : panic!("wrong error type")
6747 1 : };
6748 1 : assert!(err.to_string().contains("invalid branch start lsn"));
6749 1 : assert!(
6750 1 : err.source()
6751 1 : .unwrap()
6752 1 : .to_string()
6753 1 : .contains("we might've already garbage collected needed data")
6754 1 : )
6755 1 : }
6756 1 : }
6757 1 :
6758 1 : Ok(())
6759 1 : }
6760 :
6761 : #[tokio::test]
6762 1 : async fn test_prohibit_branch_creation_on_pre_initdb_lsn() -> anyhow::Result<()> {
6763 1 : let (tenant, ctx) =
6764 1 : TenantHarness::create("test_prohibit_branch_creation_on_pre_initdb_lsn")
6765 1 : .await?
6766 1 : .load()
6767 1 : .await;
6768 :
6769 1 : let tline = tenant
6770 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x50), DEFAULT_PG_VERSION, &ctx)
6771 1 : .await?;
6772 : // try to branch at lsn 0x25, should fail because initdb lsn is 0x50
6773 1 : match tenant
6774 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
6775 1 : .await
6776 1 : {
6777 1 : Ok(_) => panic!("branching should have failed"),
6778 1 : Err(err) => {
6779 1 : let CreateTimelineError::AncestorLsn(err) = err else {
6780 1 : panic!("wrong error type");
6781 1 : };
6782 1 : assert!(&err.to_string().contains("invalid branch start lsn"));
6783 1 : assert!(
6784 1 : &err.source()
6785 1 : .unwrap()
6786 1 : .to_string()
6787 1 : .contains("is earlier than latest GC cutoff")
6788 1 : );
6789 1 : }
6790 1 : }
6791 1 :
6792 1 : Ok(())
6793 1 : }
6794 :
6795 : /*
6796 : // FIXME: This currently fails to error out. Calling GC doesn't currently
6797 : // remove the old value, we'd need to work a little harder
6798 : #[tokio::test]
6799 : async fn test_prohibit_get_for_garbage_collected_data() -> anyhow::Result<()> {
6800 : let repo =
6801 : RepoHarness::create("test_prohibit_get_for_garbage_collected_data")?
6802 : .load();
6803 :
6804 : let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION)?;
6805 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6806 :
6807 : repo.gc_iteration(Some(TIMELINE_ID), 0x10, Duration::ZERO)?;
6808 : let applied_gc_cutoff_lsn = tline.get_applied_gc_cutoff_lsn();
6809 : assert!(*applied_gc_cutoff_lsn > Lsn(0x25));
6810 : match tline.get(*TEST_KEY, Lsn(0x25)) {
6811 : Ok(_) => panic!("request for page should have failed"),
6812 : Err(err) => assert!(err.to_string().contains("not found at")),
6813 : }
6814 : Ok(())
6815 : }
6816 : */
6817 :
6818 : #[tokio::test]
6819 1 : async fn test_get_branchpoints_from_an_inactive_timeline() -> anyhow::Result<()> {
6820 1 : let (tenant, ctx) =
6821 1 : TenantHarness::create("test_get_branchpoints_from_an_inactive_timeline")
6822 1 : .await?
6823 1 : .load()
6824 1 : .await;
6825 1 : let tline = tenant
6826 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6827 1 : .await?;
6828 1 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6829 :
6830 1 : tenant
6831 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6832 1 : .await?;
6833 1 : let newtline = tenant
6834 1 : .get_timeline(NEW_TIMELINE_ID, true)
6835 1 : .expect("Should have a local timeline");
6836 :
6837 1 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6838 :
6839 1 : tline.set_broken("test".to_owned());
6840 :
6841 1 : tenant
6842 1 : .gc_iteration(
6843 1 : Some(TIMELINE_ID),
6844 1 : 0x10,
6845 1 : Duration::ZERO,
6846 1 : &CancellationToken::new(),
6847 1 : &ctx,
6848 1 : )
6849 1 : .await?;
6850 :
6851 : // The branchpoints should contain all timelines, even ones marked
6852 : // as Broken.
6853 : {
6854 1 : let branchpoints = &tline.gc_info.read().unwrap().retain_lsns;
6855 1 : assert_eq!(branchpoints.len(), 1);
6856 1 : assert_eq!(
6857 1 : branchpoints[0],
6858 : (Lsn(0x40), NEW_TIMELINE_ID, MaybeOffloaded::No)
6859 : );
6860 : }
6861 :
6862 : // You can read the key from the child branch even though the parent is
6863 : // Broken, as long as you don't need to access data from the parent.
6864 1 : assert_eq!(
6865 1 : newtline.get(*TEST_KEY, Lsn(0x70), &ctx).await?,
6866 1 : test_img(&format!("foo at {}", Lsn(0x70)))
6867 : );
6868 :
6869 : // This needs to traverse to the parent, and fails.
6870 1 : let err = newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await.unwrap_err();
6871 1 : assert!(
6872 1 : err.to_string().starts_with(&format!(
6873 1 : "bad state on timeline {}: Broken",
6874 1 : tline.timeline_id
6875 1 : )),
6876 0 : "{err}"
6877 : );
6878 :
6879 2 : Ok(())
6880 1 : }
6881 :
6882 : #[tokio::test]
6883 1 : async fn test_retain_data_in_parent_which_is_needed_for_child() -> anyhow::Result<()> {
6884 1 : let (tenant, ctx) =
6885 1 : TenantHarness::create("test_retain_data_in_parent_which_is_needed_for_child")
6886 1 : .await?
6887 1 : .load()
6888 1 : .await;
6889 1 : let tline = tenant
6890 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6891 1 : .await?;
6892 1 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6893 :
6894 1 : tenant
6895 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6896 1 : .await?;
6897 1 : let newtline = tenant
6898 1 : .get_timeline(NEW_TIMELINE_ID, true)
6899 1 : .expect("Should have a local timeline");
6900 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
6901 1 : tenant
6902 1 : .gc_iteration(
6903 1 : Some(TIMELINE_ID),
6904 1 : 0x10,
6905 1 : Duration::ZERO,
6906 1 : &CancellationToken::new(),
6907 1 : &ctx,
6908 1 : )
6909 1 : .await?;
6910 1 : assert!(newtline.get(*TEST_KEY, Lsn(0x25), &ctx).await.is_ok());
6911 :
6912 2 : Ok(())
6913 1 : }
6914 : #[tokio::test]
6915 1 : async fn test_parent_keeps_data_forever_after_branching() -> anyhow::Result<()> {
6916 1 : let (tenant, ctx) = TenantHarness::create("test_parent_keeps_data_forever_after_branching")
6917 1 : .await?
6918 1 : .load()
6919 1 : .await;
6920 1 : let tline = tenant
6921 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6922 1 : .await?;
6923 1 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6924 :
6925 1 : tenant
6926 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6927 1 : .await?;
6928 1 : let newtline = tenant
6929 1 : .get_timeline(NEW_TIMELINE_ID, true)
6930 1 : .expect("Should have a local timeline");
6931 :
6932 1 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6933 :
6934 : // run gc on parent
6935 1 : tenant
6936 1 : .gc_iteration(
6937 1 : Some(TIMELINE_ID),
6938 1 : 0x10,
6939 1 : Duration::ZERO,
6940 1 : &CancellationToken::new(),
6941 1 : &ctx,
6942 1 : )
6943 1 : .await?;
6944 :
6945 : // Check that the data is still accessible on the branch.
6946 1 : assert_eq!(
6947 1 : newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
6948 1 : test_img(&format!("foo at {}", Lsn(0x40)))
6949 : );
6950 :
6951 2 : Ok(())
6952 1 : }
6953 :
6954 : #[tokio::test]
6955 1 : async fn timeline_load() -> anyhow::Result<()> {
6956 : const TEST_NAME: &str = "timeline_load";
6957 1 : let harness = TenantHarness::create(TEST_NAME).await?;
6958 : {
6959 1 : let (tenant, ctx) = harness.load().await;
6960 1 : let tline = tenant
6961 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x7000), DEFAULT_PG_VERSION, &ctx)
6962 1 : .await?;
6963 1 : make_some_layers(tline.as_ref(), Lsn(0x8000), &ctx).await?;
6964 : // so that all uploads finish & we can call harness.load() below again
6965 1 : tenant
6966 1 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
6967 1 : .instrument(harness.span())
6968 1 : .await
6969 1 : .ok()
6970 1 : .unwrap();
6971 : }
6972 :
6973 1 : let (tenant, _ctx) = harness.load().await;
6974 1 : tenant
6975 1 : .get_timeline(TIMELINE_ID, true)
6976 1 : .expect("cannot load timeline");
6977 :
6978 2 : Ok(())
6979 1 : }
6980 :
6981 : #[tokio::test]
6982 1 : async fn timeline_load_with_ancestor() -> anyhow::Result<()> {
6983 : const TEST_NAME: &str = "timeline_load_with_ancestor";
6984 1 : let harness = TenantHarness::create(TEST_NAME).await?;
6985 : // create two timelines
6986 : {
6987 1 : let (tenant, ctx) = harness.load().await;
6988 1 : let tline = tenant
6989 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6990 1 : .await?;
6991 :
6992 1 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6993 :
6994 1 : let child_tline = tenant
6995 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6996 1 : .await?;
6997 1 : child_tline.set_state(TimelineState::Active);
6998 :
6999 1 : let newtline = tenant
7000 1 : .get_timeline(NEW_TIMELINE_ID, true)
7001 1 : .expect("Should have a local timeline");
7002 :
7003 1 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
7004 :
7005 : // so that all uploads finish & we can call harness.load() below again
7006 1 : tenant
7007 1 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
7008 1 : .instrument(harness.span())
7009 1 : .await
7010 1 : .ok()
7011 1 : .unwrap();
7012 : }
7013 :
7014 : // check that both of them are initially unloaded
7015 1 : let (tenant, _ctx) = harness.load().await;
7016 :
7017 : // check that both, child and ancestor are loaded
7018 1 : let _child_tline = tenant
7019 1 : .get_timeline(NEW_TIMELINE_ID, true)
7020 1 : .expect("cannot get child timeline loaded");
7021 :
7022 1 : let _ancestor_tline = tenant
7023 1 : .get_timeline(TIMELINE_ID, true)
7024 1 : .expect("cannot get ancestor timeline loaded");
7025 :
7026 2 : Ok(())
7027 1 : }
7028 :
7029 : #[tokio::test]
7030 1 : async fn delta_layer_dumping() -> anyhow::Result<()> {
7031 : use storage_layer::AsLayerDesc;
7032 1 : let (tenant, ctx) = TenantHarness::create("test_layer_dumping")
7033 1 : .await?
7034 1 : .load()
7035 1 : .await;
7036 1 : let tline = tenant
7037 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7038 1 : .await?;
7039 1 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
7040 :
7041 1 : let layer_map = tline.layers.read(LayerManagerLockHolder::Testing).await;
7042 1 : let level0_deltas = layer_map
7043 1 : .layer_map()?
7044 1 : .level0_deltas()
7045 1 : .iter()
7046 2 : .map(|desc| layer_map.get_from_desc(desc))
7047 1 : .collect::<Vec<_>>();
7048 :
7049 1 : assert!(!level0_deltas.is_empty());
7050 :
7051 3 : for delta in level0_deltas {
7052 1 : // Ensure we are dumping a delta layer here
7053 2 : assert!(delta.layer_desc().is_delta);
7054 2 : delta.dump(true, &ctx).await.unwrap();
7055 1 : }
7056 1 :
7057 1 : Ok(())
7058 1 : }
7059 :
7060 : #[tokio::test]
7061 1 : async fn test_images() -> anyhow::Result<()> {
7062 1 : let (tenant, ctx) = TenantHarness::create("test_images").await?.load().await;
7063 1 : let tline = tenant
7064 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
7065 1 : .await?;
7066 :
7067 1 : let mut writer = tline.writer().await;
7068 1 : writer
7069 1 : .put(
7070 1 : *TEST_KEY,
7071 1 : Lsn(0x10),
7072 1 : &Value::Image(test_img("foo at 0x10")),
7073 1 : &ctx,
7074 1 : )
7075 1 : .await?;
7076 1 : writer.finish_write(Lsn(0x10));
7077 1 : drop(writer);
7078 :
7079 1 : tline.freeze_and_flush().await?;
7080 1 : tline
7081 1 : .compact(&CancellationToken::new(), EnumSet::default(), &ctx)
7082 1 : .await?;
7083 :
7084 1 : let mut writer = tline.writer().await;
7085 1 : writer
7086 1 : .put(
7087 1 : *TEST_KEY,
7088 1 : Lsn(0x20),
7089 1 : &Value::Image(test_img("foo at 0x20")),
7090 1 : &ctx,
7091 1 : )
7092 1 : .await?;
7093 1 : writer.finish_write(Lsn(0x20));
7094 1 : drop(writer);
7095 :
7096 1 : tline.freeze_and_flush().await?;
7097 1 : tline
7098 1 : .compact(&CancellationToken::new(), EnumSet::default(), &ctx)
7099 1 : .await?;
7100 :
7101 1 : let mut writer = tline.writer().await;
7102 1 : writer
7103 1 : .put(
7104 1 : *TEST_KEY,
7105 1 : Lsn(0x30),
7106 1 : &Value::Image(test_img("foo at 0x30")),
7107 1 : &ctx,
7108 1 : )
7109 1 : .await?;
7110 1 : writer.finish_write(Lsn(0x30));
7111 1 : drop(writer);
7112 :
7113 1 : tline.freeze_and_flush().await?;
7114 1 : tline
7115 1 : .compact(&CancellationToken::new(), EnumSet::default(), &ctx)
7116 1 : .await?;
7117 :
7118 1 : let mut writer = tline.writer().await;
7119 1 : writer
7120 1 : .put(
7121 1 : *TEST_KEY,
7122 1 : Lsn(0x40),
7123 1 : &Value::Image(test_img("foo at 0x40")),
7124 1 : &ctx,
7125 1 : )
7126 1 : .await?;
7127 1 : writer.finish_write(Lsn(0x40));
7128 1 : drop(writer);
7129 :
7130 1 : tline.freeze_and_flush().await?;
7131 1 : tline
7132 1 : .compact(&CancellationToken::new(), EnumSet::default(), &ctx)
7133 1 : .await?;
7134 :
7135 1 : assert_eq!(
7136 1 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
7137 1 : test_img("foo at 0x10")
7138 : );
7139 1 : assert_eq!(
7140 1 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
7141 1 : test_img("foo at 0x10")
7142 : );
7143 1 : assert_eq!(
7144 1 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
7145 1 : test_img("foo at 0x20")
7146 : );
7147 1 : assert_eq!(
7148 1 : tline.get(*TEST_KEY, Lsn(0x30), &ctx).await?,
7149 1 : test_img("foo at 0x30")
7150 : );
7151 1 : assert_eq!(
7152 1 : tline.get(*TEST_KEY, Lsn(0x40), &ctx).await?,
7153 1 : test_img("foo at 0x40")
7154 : );
7155 :
7156 2 : Ok(())
7157 1 : }
7158 :
7159 2 : async fn bulk_insert_compact_gc(
7160 2 : tenant: &TenantShard,
7161 2 : timeline: &Arc<Timeline>,
7162 2 : ctx: &RequestContext,
7163 2 : lsn: Lsn,
7164 2 : repeat: usize,
7165 2 : key_count: usize,
7166 2 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
7167 2 : let compact = true;
7168 2 : bulk_insert_maybe_compact_gc(tenant, timeline, ctx, lsn, repeat, key_count, compact).await
7169 2 : }
7170 :
7171 4 : async fn bulk_insert_maybe_compact_gc(
7172 4 : tenant: &TenantShard,
7173 4 : timeline: &Arc<Timeline>,
7174 4 : ctx: &RequestContext,
7175 4 : mut lsn: Lsn,
7176 4 : repeat: usize,
7177 4 : key_count: usize,
7178 4 : compact: bool,
7179 4 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
7180 4 : let mut inserted: HashMap<Key, BTreeSet<Lsn>> = Default::default();
7181 :
7182 4 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7183 4 : let mut blknum = 0;
7184 :
7185 : // Enforce that key range is monotonously increasing
7186 4 : let mut keyspace = KeySpaceAccum::new();
7187 :
7188 4 : let cancel = CancellationToken::new();
7189 :
7190 4 : for _ in 0..repeat {
7191 200 : for _ in 0..key_count {
7192 2000000 : test_key.field6 = blknum;
7193 2000000 : let mut writer = timeline.writer().await;
7194 2000000 : writer
7195 2000000 : .put(
7196 2000000 : test_key,
7197 2000000 : lsn,
7198 2000000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
7199 2000000 : ctx,
7200 2000000 : )
7201 2000000 : .await?;
7202 2000000 : inserted.entry(test_key).or_default().insert(lsn);
7203 2000000 : writer.finish_write(lsn);
7204 2000000 : drop(writer);
7205 :
7206 2000000 : keyspace.add_key(test_key);
7207 :
7208 2000000 : lsn = Lsn(lsn.0 + 0x10);
7209 2000000 : blknum += 1;
7210 : }
7211 :
7212 200 : timeline.freeze_and_flush().await?;
7213 200 : if compact {
7214 : // this requires timeline to be &Arc<Timeline>
7215 100 : timeline.compact(&cancel, EnumSet::default(), ctx).await?;
7216 100 : }
7217 :
7218 : // this doesn't really need to use the timeline_id target, but it is closer to what it
7219 : // originally was.
7220 200 : let res = tenant
7221 200 : .gc_iteration(Some(timeline.timeline_id), 0, Duration::ZERO, &cancel, ctx)
7222 200 : .await?;
7223 :
7224 200 : assert_eq!(res.layers_removed, 0, "this never removes anything");
7225 : }
7226 :
7227 4 : Ok(inserted)
7228 4 : }
7229 :
7230 : //
7231 : // Insert 1000 key-value pairs with increasing keys, flush, compact, GC.
7232 : // Repeat 50 times.
7233 : //
7234 : #[tokio::test]
7235 1 : async fn test_bulk_insert() -> anyhow::Result<()> {
7236 1 : let harness = TenantHarness::create("test_bulk_insert").await?;
7237 1 : let (tenant, ctx) = harness.load().await;
7238 1 : let tline = tenant
7239 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
7240 1 : .await?;
7241 :
7242 1 : let lsn = Lsn(0x10);
7243 1 : bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
7244 :
7245 2 : Ok(())
7246 1 : }
7247 :
7248 : // Test the vectored get real implementation against a simple sequential implementation.
7249 : //
7250 : // The test generates a keyspace by repeatedly flushing the in-memory layer and compacting.
7251 : // Projected to 2D the key space looks like below. Lsn grows upwards on the Y axis and keys
7252 : // grow to the right on the X axis.
7253 : // [Delta]
7254 : // [Delta]
7255 : // [Delta]
7256 : // [Delta]
7257 : // ------------ Image ---------------
7258 : //
7259 : // After layer generation we pick the ranges to query as follows:
7260 : // 1. The beginning of each delta layer
7261 : // 2. At the seam between two adjacent delta layers
7262 : //
7263 : // There's one major downside to this test: delta layers only contains images,
7264 : // so the search can stop at the first delta layer and doesn't traverse any deeper.
7265 : #[tokio::test]
7266 1 : async fn test_get_vectored() -> anyhow::Result<()> {
7267 1 : let harness = TenantHarness::create("test_get_vectored").await?;
7268 1 : let (tenant, ctx) = harness.load().await;
7269 1 : let io_concurrency = IoConcurrency::spawn_for_test();
7270 1 : let tline = tenant
7271 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
7272 1 : .await?;
7273 :
7274 1 : let lsn = Lsn(0x10);
7275 1 : let inserted = bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
7276 :
7277 1 : let guard = tline.layers.read(LayerManagerLockHolder::Testing).await;
7278 1 : let lm = guard.layer_map()?;
7279 :
7280 1 : lm.dump(true, &ctx).await?;
7281 :
7282 1 : let mut reads = Vec::new();
7283 1 : let mut prev = None;
7284 6 : lm.iter_historic_layers().for_each(|desc| {
7285 6 : if !desc.is_delta() {
7286 1 : prev = Some(desc.clone());
7287 1 : return;
7288 5 : }
7289 :
7290 5 : let start = desc.key_range.start;
7291 5 : let end = desc
7292 5 : .key_range
7293 5 : .start
7294 5 : .add(tenant.conf.max_get_vectored_keys.get() as u32);
7295 5 : reads.push(KeySpace {
7296 5 : ranges: vec![start..end],
7297 5 : });
7298 :
7299 5 : if let Some(prev) = &prev {
7300 5 : if !prev.is_delta() {
7301 5 : return;
7302 0 : }
7303 :
7304 0 : let first_range = Key {
7305 0 : field6: prev.key_range.end.field6 - 4,
7306 0 : ..prev.key_range.end
7307 0 : }..prev.key_range.end;
7308 :
7309 0 : let second_range = desc.key_range.start..Key {
7310 0 : field6: desc.key_range.start.field6 + 4,
7311 0 : ..desc.key_range.start
7312 0 : };
7313 :
7314 0 : reads.push(KeySpace {
7315 0 : ranges: vec![first_range, second_range],
7316 0 : });
7317 0 : };
7318 :
7319 0 : prev = Some(desc.clone());
7320 6 : });
7321 :
7322 1 : drop(guard);
7323 :
7324 : // Pick a big LSN such that we query over all the changes.
7325 1 : let reads_lsn = Lsn(u64::MAX - 1);
7326 :
7327 6 : for read in reads {
7328 5 : info!("Doing vectored read on {:?}", read);
7329 1 :
7330 5 : let query = VersionedKeySpaceQuery::uniform(read.clone(), reads_lsn);
7331 1 :
7332 5 : let vectored_res = tline
7333 5 : .get_vectored_impl(
7334 5 : query,
7335 5 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7336 5 : &ctx,
7337 5 : )
7338 5 : .await;
7339 1 :
7340 5 : let mut expected_lsns: HashMap<Key, Lsn> = Default::default();
7341 5 : let mut expect_missing = false;
7342 5 : let mut key = read.start().unwrap();
7343 165 : while key != read.end().unwrap() {
7344 160 : if let Some(lsns) = inserted.get(&key) {
7345 160 : let expected_lsn = lsns.iter().rfind(|lsn| **lsn <= reads_lsn);
7346 160 : match expected_lsn {
7347 160 : Some(lsn) => {
7348 160 : expected_lsns.insert(key, *lsn);
7349 160 : }
7350 1 : None => {
7351 1 : expect_missing = true;
7352 1 : break;
7353 1 : }
7354 1 : }
7355 1 : } else {
7356 1 : expect_missing = true;
7357 1 : break;
7358 1 : }
7359 1 :
7360 160 : key = key.next();
7361 1 : }
7362 1 :
7363 5 : if expect_missing {
7364 1 : assert!(matches!(vectored_res, Err(GetVectoredError::MissingKey(_))));
7365 1 : } else {
7366 160 : for (key, image) in vectored_res? {
7367 160 : let expected_lsn = expected_lsns.get(&key).expect("determined above");
7368 160 : let expected_image = test_img(&format!("{} at {}", key.field6, expected_lsn));
7369 160 : assert_eq!(image?, expected_image);
7370 1 : }
7371 1 : }
7372 1 : }
7373 1 :
7374 1 : Ok(())
7375 1 : }
7376 :
7377 : #[tokio::test]
7378 1 : async fn test_get_vectored_aux_files() -> anyhow::Result<()> {
7379 1 : let harness = TenantHarness::create("test_get_vectored_aux_files").await?;
7380 :
7381 1 : let (tenant, ctx) = harness.load().await;
7382 1 : let io_concurrency = IoConcurrency::spawn_for_test();
7383 1 : let (tline, ctx) = tenant
7384 1 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
7385 1 : .await?;
7386 1 : let tline = tline.raw_timeline().unwrap();
7387 :
7388 1 : let mut modification = tline.begin_modification(Lsn(0x1000));
7389 1 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
7390 1 : modification.set_lsn(Lsn(0x1008))?;
7391 1 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
7392 1 : modification.commit(&ctx).await?;
7393 :
7394 1 : let child_timeline_id = TimelineId::generate();
7395 1 : tenant
7396 1 : .branch_timeline_test(
7397 1 : tline,
7398 1 : child_timeline_id,
7399 1 : Some(tline.get_last_record_lsn()),
7400 1 : &ctx,
7401 1 : )
7402 1 : .await?;
7403 :
7404 1 : let child_timeline = tenant
7405 1 : .get_timeline(child_timeline_id, true)
7406 1 : .expect("Should have the branched timeline");
7407 :
7408 1 : let aux_keyspace = KeySpace {
7409 1 : ranges: vec![NON_INHERITED_RANGE],
7410 1 : };
7411 1 : let read_lsn = child_timeline.get_last_record_lsn();
7412 :
7413 1 : let query = VersionedKeySpaceQuery::uniform(aux_keyspace.clone(), read_lsn);
7414 :
7415 1 : let vectored_res = child_timeline
7416 1 : .get_vectored_impl(
7417 1 : query,
7418 1 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7419 1 : &ctx,
7420 1 : )
7421 1 : .await;
7422 :
7423 1 : let images = vectored_res?;
7424 1 : assert!(images.is_empty());
7425 2 : Ok(())
7426 1 : }
7427 :
7428 : // Test that vectored get handles layer gaps correctly
7429 : // by advancing into the next ancestor timeline if required.
7430 : //
7431 : // The test generates timelines that look like the diagram below.
7432 : // We leave a gap in one of the L1 layers at `gap_at_key` (`/` in the diagram).
7433 : // The reconstruct data for that key lies in the ancestor timeline (`X` in the diagram).
7434 : //
7435 : // ```
7436 : //-------------------------------+
7437 : // ... |
7438 : // [ L1 ] |
7439 : // [ / L1 ] | Child Timeline
7440 : // ... |
7441 : // ------------------------------+
7442 : // [ X L1 ] | Parent Timeline
7443 : // ------------------------------+
7444 : // ```
7445 : #[tokio::test]
7446 1 : async fn test_get_vectored_key_gap() -> anyhow::Result<()> {
7447 1 : let tenant_conf = pageserver_api::models::TenantConfig {
7448 1 : // Make compaction deterministic
7449 1 : gc_period: Some(Duration::ZERO),
7450 1 : compaction_period: Some(Duration::ZERO),
7451 1 : // Encourage creation of L1 layers
7452 1 : checkpoint_distance: Some(16 * 1024),
7453 1 : compaction_target_size: Some(8 * 1024),
7454 1 : ..Default::default()
7455 1 : };
7456 :
7457 1 : let harness = TenantHarness::create_custom(
7458 1 : "test_get_vectored_key_gap",
7459 1 : tenant_conf,
7460 1 : TenantId::generate(),
7461 1 : ShardIdentity::unsharded(),
7462 1 : Generation::new(0xdeadbeef),
7463 1 : )
7464 1 : .await?;
7465 1 : let (tenant, ctx) = harness.load().await;
7466 1 : let io_concurrency = IoConcurrency::spawn_for_test();
7467 :
7468 1 : let mut current_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7469 1 : let gap_at_key = current_key.add(100);
7470 1 : let mut current_lsn = Lsn(0x10);
7471 :
7472 : const KEY_COUNT: usize = 10_000;
7473 :
7474 1 : let timeline_id = TimelineId::generate();
7475 1 : let current_timeline = tenant
7476 1 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
7477 1 : .await?;
7478 :
7479 1 : current_lsn += 0x100;
7480 :
7481 1 : let mut writer = current_timeline.writer().await;
7482 1 : writer
7483 1 : .put(
7484 1 : gap_at_key,
7485 1 : current_lsn,
7486 1 : &Value::Image(test_img(&format!("{gap_at_key} at {current_lsn}"))),
7487 1 : &ctx,
7488 1 : )
7489 1 : .await?;
7490 1 : writer.finish_write(current_lsn);
7491 1 : drop(writer);
7492 :
7493 1 : let mut latest_lsns = HashMap::new();
7494 1 : latest_lsns.insert(gap_at_key, current_lsn);
7495 :
7496 1 : current_timeline.freeze_and_flush().await?;
7497 :
7498 1 : let child_timeline_id = TimelineId::generate();
7499 :
7500 1 : tenant
7501 1 : .branch_timeline_test(
7502 1 : ¤t_timeline,
7503 1 : child_timeline_id,
7504 1 : Some(current_lsn),
7505 1 : &ctx,
7506 1 : )
7507 1 : .await?;
7508 1 : let child_timeline = tenant
7509 1 : .get_timeline(child_timeline_id, true)
7510 1 : .expect("Should have the branched timeline");
7511 :
7512 10001 : for i in 0..KEY_COUNT {
7513 10000 : if current_key == gap_at_key {
7514 1 : current_key = current_key.next();
7515 1 : continue;
7516 9999 : }
7517 :
7518 9999 : current_lsn += 0x10;
7519 :
7520 9999 : let mut writer = child_timeline.writer().await;
7521 9999 : writer
7522 9999 : .put(
7523 9999 : current_key,
7524 9999 : current_lsn,
7525 9999 : &Value::Image(test_img(&format!("{current_key} at {current_lsn}"))),
7526 9999 : &ctx,
7527 9999 : )
7528 9999 : .await?;
7529 9999 : writer.finish_write(current_lsn);
7530 9999 : drop(writer);
7531 :
7532 9999 : latest_lsns.insert(current_key, current_lsn);
7533 9999 : current_key = current_key.next();
7534 :
7535 : // Flush every now and then to encourage layer file creation.
7536 9999 : if i % 500 == 0 {
7537 20 : child_timeline.freeze_and_flush().await?;
7538 9979 : }
7539 : }
7540 :
7541 1 : child_timeline.freeze_and_flush().await?;
7542 1 : let mut flags = EnumSet::new();
7543 1 : flags.insert(CompactFlags::ForceRepartition);
7544 1 : child_timeline
7545 1 : .compact(&CancellationToken::new(), flags, &ctx)
7546 1 : .await?;
7547 :
7548 1 : let key_near_end = {
7549 1 : let mut tmp = current_key;
7550 1 : tmp.field6 -= 10;
7551 1 : tmp
7552 : };
7553 :
7554 1 : let key_near_gap = {
7555 1 : let mut tmp = gap_at_key;
7556 1 : tmp.field6 -= 10;
7557 1 : tmp
7558 : };
7559 :
7560 1 : let read = KeySpace {
7561 1 : ranges: vec![key_near_gap..gap_at_key.next(), key_near_end..current_key],
7562 1 : };
7563 :
7564 1 : let query = VersionedKeySpaceQuery::uniform(read.clone(), current_lsn);
7565 :
7566 1 : let results = child_timeline
7567 1 : .get_vectored_impl(
7568 1 : query,
7569 1 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7570 1 : &ctx,
7571 1 : )
7572 1 : .await?;
7573 :
7574 22 : for (key, img_res) in results {
7575 21 : let expected = test_img(&format!("{} at {}", key, latest_lsns[&key]));
7576 21 : assert_eq!(img_res?, expected);
7577 1 : }
7578 1 :
7579 1 : Ok(())
7580 1 : }
7581 :
7582 : // Test that vectored get descends into ancestor timelines correctly and
7583 : // does not return an image that's newer than requested.
7584 : //
7585 : // The diagram below ilustrates an interesting case. We have a parent timeline
7586 : // (top of the Lsn range) and a child timeline. The request key cannot be reconstructed
7587 : // from the child timeline, so the parent timeline must be visited. When advacing into
7588 : // the child timeline, the read path needs to remember what the requested Lsn was in
7589 : // order to avoid returning an image that's too new. The test below constructs such
7590 : // a timeline setup and does a few queries around the Lsn of each page image.
7591 : // ```
7592 : // LSN
7593 : // ^
7594 : // |
7595 : // |
7596 : // 500 | --------------------------------------> branch point
7597 : // 400 | X
7598 : // 300 | X
7599 : // 200 | --------------------------------------> requested lsn
7600 : // 100 | X
7601 : // |---------------------------------------> Key
7602 : // |
7603 : // ------> requested key
7604 : //
7605 : // Legend:
7606 : // * X - page images
7607 : // ```
7608 : #[tokio::test]
7609 1 : async fn test_get_vectored_ancestor_descent() -> anyhow::Result<()> {
7610 1 : let harness = TenantHarness::create("test_get_vectored_on_lsn_axis").await?;
7611 1 : let (tenant, ctx) = harness.load().await;
7612 1 : let io_concurrency = IoConcurrency::spawn_for_test();
7613 :
7614 1 : let start_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7615 1 : let end_key = start_key.add(1000);
7616 1 : let child_gap_at_key = start_key.add(500);
7617 1 : let mut parent_gap_lsns: BTreeMap<Lsn, String> = BTreeMap::new();
7618 :
7619 1 : let mut current_lsn = Lsn(0x10);
7620 :
7621 1 : let timeline_id = TimelineId::generate();
7622 1 : let parent_timeline = tenant
7623 1 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
7624 1 : .await?;
7625 :
7626 1 : current_lsn += 0x100;
7627 :
7628 4 : for _ in 0..3 {
7629 3 : let mut key = start_key;
7630 3003 : while key < end_key {
7631 3000 : current_lsn += 0x10;
7632 :
7633 3000 : let image_value = format!("{child_gap_at_key} at {current_lsn}");
7634 :
7635 3000 : let mut writer = parent_timeline.writer().await;
7636 3000 : writer
7637 3000 : .put(
7638 3000 : key,
7639 3000 : current_lsn,
7640 3000 : &Value::Image(test_img(&image_value)),
7641 3000 : &ctx,
7642 3000 : )
7643 3000 : .await?;
7644 3000 : writer.finish_write(current_lsn);
7645 :
7646 3000 : if key == child_gap_at_key {
7647 3 : parent_gap_lsns.insert(current_lsn, image_value);
7648 2997 : }
7649 :
7650 3000 : key = key.next();
7651 : }
7652 :
7653 3 : parent_timeline.freeze_and_flush().await?;
7654 : }
7655 :
7656 1 : let child_timeline_id = TimelineId::generate();
7657 :
7658 1 : let child_timeline = tenant
7659 1 : .branch_timeline_test(&parent_timeline, child_timeline_id, Some(current_lsn), &ctx)
7660 1 : .await?;
7661 :
7662 1 : let mut key = start_key;
7663 1001 : while key < end_key {
7664 1000 : if key == child_gap_at_key {
7665 1 : key = key.next();
7666 1 : continue;
7667 999 : }
7668 :
7669 999 : current_lsn += 0x10;
7670 :
7671 999 : let mut writer = child_timeline.writer().await;
7672 999 : writer
7673 999 : .put(
7674 999 : key,
7675 999 : current_lsn,
7676 999 : &Value::Image(test_img(&format!("{key} at {current_lsn}"))),
7677 999 : &ctx,
7678 999 : )
7679 999 : .await?;
7680 999 : writer.finish_write(current_lsn);
7681 :
7682 999 : key = key.next();
7683 : }
7684 :
7685 1 : child_timeline.freeze_and_flush().await?;
7686 :
7687 1 : let lsn_offsets: [i64; 5] = [-10, -1, 0, 1, 10];
7688 1 : let mut query_lsns = Vec::new();
7689 3 : for image_lsn in parent_gap_lsns.keys().rev() {
7690 18 : for offset in lsn_offsets {
7691 15 : query_lsns.push(Lsn(image_lsn
7692 15 : .0
7693 15 : .checked_add_signed(offset)
7694 15 : .expect("Shouldn't overflow")));
7695 15 : }
7696 1 : }
7697 1 :
7698 16 : for query_lsn in query_lsns {
7699 15 : let query = VersionedKeySpaceQuery::uniform(
7700 15 : KeySpace {
7701 15 : ranges: vec![child_gap_at_key..child_gap_at_key.next()],
7702 15 : },
7703 15 : query_lsn,
7704 1 : );
7705 1 :
7706 15 : let results = child_timeline
7707 15 : .get_vectored_impl(
7708 15 : query,
7709 15 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7710 15 : &ctx,
7711 15 : )
7712 15 : .await;
7713 1 :
7714 15 : let expected_item = parent_gap_lsns
7715 15 : .iter()
7716 15 : .rev()
7717 34 : .find(|(lsn, _)| **lsn <= query_lsn);
7718 1 :
7719 15 : info!(
7720 1 : "Doing vectored read at LSN {}. Expecting image to be: {:?}",
7721 1 : query_lsn, expected_item
7722 1 : );
7723 1 :
7724 15 : match expected_item {
7725 13 : Some((_, img_value)) => {
7726 13 : let key_results = results.expect("No vectored get error expected");
7727 13 : let key_result = &key_results[&child_gap_at_key];
7728 13 : let returned_img = key_result
7729 13 : .as_ref()
7730 13 : .expect("No page reconstruct error expected");
7731 1 :
7732 13 : info!(
7733 1 : "Vectored read at LSN {} returned image {}",
7734 1 : query_lsn,
7735 1 : std::str::from_utf8(returned_img)?
7736 1 : );
7737 13 : assert_eq!(*returned_img, test_img(img_value));
7738 1 : }
7739 1 : None => {
7740 2 : assert!(matches!(results, Err(GetVectoredError::MissingKey(_))));
7741 1 : }
7742 1 : }
7743 1 : }
7744 1 :
7745 1 : Ok(())
7746 1 : }
7747 :
7748 : #[tokio::test]
7749 1 : async fn test_random_updates() -> anyhow::Result<()> {
7750 1 : let names_algorithms = [
7751 1 : ("test_random_updates_legacy", CompactionAlgorithm::Legacy),
7752 1 : ("test_random_updates_tiered", CompactionAlgorithm::Tiered),
7753 1 : ];
7754 3 : for (name, algorithm) in names_algorithms {
7755 2 : test_random_updates_algorithm(name, algorithm).await?;
7756 1 : }
7757 1 : Ok(())
7758 1 : }
7759 :
7760 2 : async fn test_random_updates_algorithm(
7761 2 : name: &'static str,
7762 2 : compaction_algorithm: CompactionAlgorithm,
7763 2 : ) -> anyhow::Result<()> {
7764 2 : let mut harness = TenantHarness::create(name).await?;
7765 2 : harness.tenant_conf.compaction_algorithm = Some(CompactionAlgorithmSettings {
7766 2 : kind: compaction_algorithm,
7767 2 : });
7768 2 : let (tenant, ctx) = harness.load().await;
7769 2 : let tline = tenant
7770 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7771 2 : .await?;
7772 :
7773 : const NUM_KEYS: usize = 1000;
7774 2 : let cancel = CancellationToken::new();
7775 :
7776 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7777 2 : let mut test_key_end = test_key;
7778 2 : test_key_end.field6 = NUM_KEYS as u32;
7779 2 : tline.add_extra_test_dense_keyspace(KeySpace::single(test_key..test_key_end));
7780 :
7781 2 : let mut keyspace = KeySpaceAccum::new();
7782 :
7783 : // Track when each page was last modified. Used to assert that
7784 : // a read sees the latest page version.
7785 2 : let mut updated = [Lsn(0); NUM_KEYS];
7786 :
7787 2 : let mut lsn = Lsn(0x10);
7788 : #[allow(clippy::needless_range_loop)]
7789 2002 : for blknum in 0..NUM_KEYS {
7790 2000 : lsn = Lsn(lsn.0 + 0x10);
7791 2000 : test_key.field6 = blknum as u32;
7792 2000 : let mut writer = tline.writer().await;
7793 2000 : writer
7794 2000 : .put(
7795 2000 : test_key,
7796 2000 : lsn,
7797 2000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
7798 2000 : &ctx,
7799 2000 : )
7800 2000 : .await?;
7801 2000 : writer.finish_write(lsn);
7802 2000 : updated[blknum] = lsn;
7803 2000 : drop(writer);
7804 :
7805 2000 : keyspace.add_key(test_key);
7806 : }
7807 :
7808 102 : for _ in 0..50 {
7809 100100 : for _ in 0..NUM_KEYS {
7810 100000 : lsn = Lsn(lsn.0 + 0x10);
7811 100000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7812 100000 : test_key.field6 = blknum as u32;
7813 100000 : let mut writer = tline.writer().await;
7814 100000 : writer
7815 100000 : .put(
7816 100000 : test_key,
7817 100000 : lsn,
7818 100000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
7819 100000 : &ctx,
7820 100000 : )
7821 100000 : .await?;
7822 100000 : writer.finish_write(lsn);
7823 100000 : drop(writer);
7824 100000 : updated[blknum] = lsn;
7825 : }
7826 :
7827 : // Read all the blocks
7828 100000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7829 100000 : test_key.field6 = blknum as u32;
7830 100000 : assert_eq!(
7831 100000 : tline.get(test_key, lsn, &ctx).await?,
7832 100000 : test_img(&format!("{blknum} at {last_lsn}"))
7833 : );
7834 : }
7835 :
7836 : // Perform a cycle of flush, and GC
7837 100 : tline.freeze_and_flush().await?;
7838 100 : tenant
7839 100 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7840 100 : .await?;
7841 : }
7842 :
7843 2 : Ok(())
7844 2 : }
7845 :
7846 : #[tokio::test]
7847 1 : async fn test_traverse_branches() -> anyhow::Result<()> {
7848 1 : let (tenant, ctx) = TenantHarness::create("test_traverse_branches")
7849 1 : .await?
7850 1 : .load()
7851 1 : .await;
7852 1 : let mut tline = tenant
7853 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7854 1 : .await?;
7855 :
7856 : const NUM_KEYS: usize = 1000;
7857 :
7858 1 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7859 :
7860 1 : let mut keyspace = KeySpaceAccum::new();
7861 :
7862 1 : let cancel = CancellationToken::new();
7863 :
7864 : // Track when each page was last modified. Used to assert that
7865 : // a read sees the latest page version.
7866 1 : let mut updated = [Lsn(0); NUM_KEYS];
7867 :
7868 1 : let mut lsn = Lsn(0x10);
7869 1 : #[allow(clippy::needless_range_loop)]
7870 1001 : for blknum in 0..NUM_KEYS {
7871 1000 : lsn = Lsn(lsn.0 + 0x10);
7872 1000 : test_key.field6 = blknum as u32;
7873 1000 : let mut writer = tline.writer().await;
7874 1000 : writer
7875 1000 : .put(
7876 1000 : test_key,
7877 1000 : lsn,
7878 1000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
7879 1000 : &ctx,
7880 1000 : )
7881 1000 : .await?;
7882 1000 : writer.finish_write(lsn);
7883 1000 : updated[blknum] = lsn;
7884 1000 : drop(writer);
7885 1 :
7886 1000 : keyspace.add_key(test_key);
7887 1 : }
7888 1 :
7889 51 : for _ in 0..50 {
7890 50 : let new_tline_id = TimelineId::generate();
7891 50 : tenant
7892 50 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7893 50 : .await?;
7894 50 : tline = tenant
7895 50 : .get_timeline(new_tline_id, true)
7896 50 : .expect("Should have the branched timeline");
7897 1 :
7898 50050 : for _ in 0..NUM_KEYS {
7899 50000 : lsn = Lsn(lsn.0 + 0x10);
7900 50000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7901 50000 : test_key.field6 = blknum as u32;
7902 50000 : let mut writer = tline.writer().await;
7903 50000 : writer
7904 50000 : .put(
7905 50000 : test_key,
7906 50000 : lsn,
7907 50000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
7908 50000 : &ctx,
7909 50000 : )
7910 50000 : .await?;
7911 50000 : println!("updating {blknum} at {lsn}");
7912 50000 : writer.finish_write(lsn);
7913 50000 : drop(writer);
7914 50000 : updated[blknum] = lsn;
7915 1 : }
7916 1 :
7917 1 : // Read all the blocks
7918 50000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7919 50000 : test_key.field6 = blknum as u32;
7920 50000 : assert_eq!(
7921 50000 : tline.get(test_key, lsn, &ctx).await?,
7922 50000 : test_img(&format!("{blknum} at {last_lsn}"))
7923 1 : );
7924 1 : }
7925 1 :
7926 1 : // Perform a cycle of flush, compact, and GC
7927 50 : tline.freeze_and_flush().await?;
7928 50 : tline.compact(&cancel, EnumSet::default(), &ctx).await?;
7929 50 : tenant
7930 50 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7931 50 : .await?;
7932 1 : }
7933 1 :
7934 1 : Ok(())
7935 1 : }
7936 :
7937 : #[tokio::test]
7938 1 : async fn test_traverse_ancestors() -> anyhow::Result<()> {
7939 1 : let (tenant, ctx) = TenantHarness::create("test_traverse_ancestors")
7940 1 : .await?
7941 1 : .load()
7942 1 : .await;
7943 1 : let mut tline = tenant
7944 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7945 1 : .await?;
7946 :
7947 : const NUM_KEYS: usize = 100;
7948 : const NUM_TLINES: usize = 50;
7949 :
7950 1 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7951 : // Track page mutation lsns across different timelines.
7952 1 : let mut updated = [[Lsn(0); NUM_KEYS]; NUM_TLINES];
7953 :
7954 1 : let mut lsn = Lsn(0x10);
7955 :
7956 1 : #[allow(clippy::needless_range_loop)]
7957 51 : for idx in 0..NUM_TLINES {
7958 50 : let new_tline_id = TimelineId::generate();
7959 50 : tenant
7960 50 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7961 50 : .await?;
7962 50 : tline = tenant
7963 50 : .get_timeline(new_tline_id, true)
7964 50 : .expect("Should have the branched timeline");
7965 1 :
7966 5050 : for _ in 0..NUM_KEYS {
7967 5000 : lsn = Lsn(lsn.0 + 0x10);
7968 5000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7969 5000 : test_key.field6 = blknum as u32;
7970 5000 : let mut writer = tline.writer().await;
7971 5000 : writer
7972 5000 : .put(
7973 5000 : test_key,
7974 5000 : lsn,
7975 5000 : &Value::Image(test_img(&format!("{idx} {blknum} at {lsn}"))),
7976 5000 : &ctx,
7977 5000 : )
7978 5000 : .await?;
7979 5000 : println!("updating [{idx}][{blknum}] at {lsn}");
7980 5000 : writer.finish_write(lsn);
7981 5000 : drop(writer);
7982 5000 : updated[idx][blknum] = lsn;
7983 1 : }
7984 1 : }
7985 1 :
7986 1 : // Read pages from leaf timeline across all ancestors.
7987 50 : for (idx, lsns) in updated.iter().enumerate() {
7988 5000 : for (blknum, lsn) in lsns.iter().enumerate() {
7989 1 : // Skip empty mutations.
7990 5000 : if lsn.0 == 0 {
7991 1839 : continue;
7992 3161 : }
7993 3161 : println!("checking [{idx}][{blknum}] at {lsn}");
7994 3161 : test_key.field6 = blknum as u32;
7995 3161 : assert_eq!(
7996 3161 : tline.get(test_key, *lsn, &ctx).await?,
7997 3161 : test_img(&format!("{idx} {blknum} at {lsn}"))
7998 1 : );
7999 1 : }
8000 1 : }
8001 1 : Ok(())
8002 1 : }
8003 :
8004 : #[tokio::test]
8005 1 : async fn test_write_at_initdb_lsn_takes_optimization_code_path() -> anyhow::Result<()> {
8006 1 : let (tenant, ctx) = TenantHarness::create("test_empty_test_timeline_is_usable")
8007 1 : .await?
8008 1 : .load()
8009 1 : .await;
8010 :
8011 1 : let initdb_lsn = Lsn(0x20);
8012 1 : let (utline, ctx) = tenant
8013 1 : .create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)
8014 1 : .await?;
8015 1 : let tline = utline.raw_timeline().unwrap();
8016 :
8017 : // Spawn flush loop now so that we can set the `expect_initdb_optimization`
8018 1 : tline.maybe_spawn_flush_loop();
8019 :
8020 : // Make sure the timeline has the minimum set of required keys for operation.
8021 : // The only operation you can always do on an empty timeline is to `put` new data.
8022 : // Except if you `put` at `initdb_lsn`.
8023 : // In that case, there's an optimization to directly create image layers instead of delta layers.
8024 : // It uses `repartition()`, which assumes some keys to be present.
8025 : // Let's make sure the test timeline can handle that case.
8026 : {
8027 1 : let mut state = tline.flush_loop_state.lock().unwrap();
8028 1 : assert_eq!(
8029 : timeline::FlushLoopState::Running {
8030 : expect_initdb_optimization: false,
8031 : initdb_optimization_count: 0,
8032 : },
8033 1 : *state
8034 : );
8035 1 : *state = timeline::FlushLoopState::Running {
8036 1 : expect_initdb_optimization: true,
8037 1 : initdb_optimization_count: 0,
8038 1 : };
8039 : }
8040 :
8041 : // Make writes at the initdb_lsn. When we flush it below, it should be handled by the optimization.
8042 : // As explained above, the optimization requires some keys to be present.
8043 : // As per `create_empty_timeline` documentation, use init_empty to set them.
8044 : // This is what `create_test_timeline` does, by the way.
8045 1 : let mut modification = tline.begin_modification(initdb_lsn);
8046 1 : modification
8047 1 : .init_empty_test_timeline()
8048 1 : .context("init_empty_test_timeline")?;
8049 1 : modification
8050 1 : .commit(&ctx)
8051 1 : .await
8052 1 : .context("commit init_empty_test_timeline modification")?;
8053 :
8054 : // Do the flush. The flush code will check the expectations that we set above.
8055 1 : tline.freeze_and_flush().await?;
8056 :
8057 : // assert freeze_and_flush exercised the initdb optimization
8058 1 : {
8059 1 : let state = tline.flush_loop_state.lock().unwrap();
8060 1 : let timeline::FlushLoopState::Running {
8061 1 : expect_initdb_optimization,
8062 1 : initdb_optimization_count,
8063 1 : } = *state
8064 1 : else {
8065 1 : panic!("unexpected state: {:?}", *state);
8066 1 : };
8067 1 : assert!(expect_initdb_optimization);
8068 1 : assert!(initdb_optimization_count > 0);
8069 1 : }
8070 1 : Ok(())
8071 1 : }
8072 :
8073 : #[tokio::test]
8074 1 : async fn test_create_guard_crash() -> anyhow::Result<()> {
8075 1 : let name = "test_create_guard_crash";
8076 1 : let harness = TenantHarness::create(name).await?;
8077 : {
8078 1 : let (tenant, ctx) = harness.load().await;
8079 1 : let (tline, _ctx) = tenant
8080 1 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
8081 1 : .await?;
8082 : // Leave the timeline ID in [`TenantShard::timelines_creating`] to exclude attempting to create it again
8083 1 : let raw_tline = tline.raw_timeline().unwrap();
8084 1 : raw_tline
8085 1 : .shutdown(super::timeline::ShutdownMode::Hard)
8086 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))
8087 1 : .await;
8088 1 : std::mem::forget(tline);
8089 : }
8090 :
8091 1 : let (tenant, _) = harness.load().await;
8092 1 : match tenant.get_timeline(TIMELINE_ID, false) {
8093 0 : Ok(_) => panic!("timeline should've been removed during load"),
8094 1 : Err(e) => {
8095 1 : assert_eq!(
8096 : e,
8097 1 : GetTimelineError::NotFound {
8098 1 : tenant_id: tenant.tenant_shard_id,
8099 1 : timeline_id: TIMELINE_ID,
8100 1 : }
8101 : )
8102 : }
8103 : }
8104 :
8105 1 : assert!(
8106 1 : !harness
8107 1 : .conf
8108 1 : .timeline_path(&tenant.tenant_shard_id, &TIMELINE_ID)
8109 1 : .exists()
8110 : );
8111 :
8112 2 : Ok(())
8113 1 : }
8114 :
8115 : #[tokio::test]
8116 1 : async fn test_read_at_max_lsn() -> anyhow::Result<()> {
8117 1 : let names_algorithms = [
8118 1 : ("test_read_at_max_lsn_legacy", CompactionAlgorithm::Legacy),
8119 1 : ("test_read_at_max_lsn_tiered", CompactionAlgorithm::Tiered),
8120 1 : ];
8121 3 : for (name, algorithm) in names_algorithms {
8122 2 : test_read_at_max_lsn_algorithm(name, algorithm).await?;
8123 1 : }
8124 1 : Ok(())
8125 1 : }
8126 :
8127 2 : async fn test_read_at_max_lsn_algorithm(
8128 2 : name: &'static str,
8129 2 : compaction_algorithm: CompactionAlgorithm,
8130 2 : ) -> anyhow::Result<()> {
8131 2 : let mut harness = TenantHarness::create(name).await?;
8132 2 : harness.tenant_conf.compaction_algorithm = Some(CompactionAlgorithmSettings {
8133 2 : kind: compaction_algorithm,
8134 2 : });
8135 2 : let (tenant, ctx) = harness.load().await;
8136 2 : let tline = tenant
8137 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
8138 2 : .await?;
8139 :
8140 2 : let lsn = Lsn(0x10);
8141 2 : let compact = false;
8142 2 : bulk_insert_maybe_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000, compact).await?;
8143 :
8144 2 : let test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
8145 2 : let read_lsn = Lsn(u64::MAX - 1);
8146 :
8147 2 : let result = tline.get(test_key, read_lsn, &ctx).await;
8148 2 : assert!(result.is_ok(), "result is not Ok: {}", result.unwrap_err());
8149 :
8150 2 : Ok(())
8151 2 : }
8152 :
8153 : #[tokio::test]
8154 1 : async fn test_metadata_scan() -> anyhow::Result<()> {
8155 1 : let harness = TenantHarness::create("test_metadata_scan").await?;
8156 1 : let (tenant, ctx) = harness.load().await;
8157 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8158 1 : let tline = tenant
8159 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
8160 1 : .await?;
8161 :
8162 : const NUM_KEYS: usize = 1000;
8163 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
8164 :
8165 1 : let cancel = CancellationToken::new();
8166 :
8167 1 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
8168 1 : base_key.field1 = AUX_KEY_PREFIX;
8169 1 : let mut test_key = base_key;
8170 :
8171 : // Track when each page was last modified. Used to assert that
8172 : // a read sees the latest page version.
8173 1 : let mut updated = [Lsn(0); NUM_KEYS];
8174 :
8175 1 : let mut lsn = Lsn(0x10);
8176 : #[allow(clippy::needless_range_loop)]
8177 1001 : for blknum in 0..NUM_KEYS {
8178 1000 : lsn = Lsn(lsn.0 + 0x10);
8179 1000 : test_key.field6 = (blknum * STEP) as u32;
8180 1000 : let mut writer = tline.writer().await;
8181 1000 : writer
8182 1000 : .put(
8183 1000 : test_key,
8184 1000 : lsn,
8185 1000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
8186 1000 : &ctx,
8187 1000 : )
8188 1000 : .await?;
8189 1000 : writer.finish_write(lsn);
8190 1000 : updated[blknum] = lsn;
8191 1000 : drop(writer);
8192 : }
8193 :
8194 1 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
8195 :
8196 12 : for iter in 0..=10 {
8197 1 : // Read all the blocks
8198 11000 : for (blknum, last_lsn) in updated.iter().enumerate() {
8199 11000 : test_key.field6 = (blknum * STEP) as u32;
8200 11000 : assert_eq!(
8201 11000 : tline.get(test_key, lsn, &ctx).await?,
8202 11000 : test_img(&format!("{blknum} at {last_lsn}"))
8203 1 : );
8204 1 : }
8205 1 :
8206 11 : let mut cnt = 0;
8207 11 : let query = VersionedKeySpaceQuery::uniform(keyspace.clone(), lsn);
8208 1 :
8209 11000 : for (key, value) in tline
8210 11 : .get_vectored_impl(
8211 11 : query,
8212 11 : &mut ValuesReconstructState::new(io_concurrency.clone()),
8213 11 : &ctx,
8214 11 : )
8215 11 : .await?
8216 1 : {
8217 11000 : let blknum = key.field6 as usize;
8218 11000 : let value = value?;
8219 11000 : assert!(blknum % STEP == 0);
8220 11000 : let blknum = blknum / STEP;
8221 11000 : assert_eq!(
8222 1 : value,
8223 11000 : test_img(&format!("{} at {}", blknum, updated[blknum]))
8224 1 : );
8225 11000 : cnt += 1;
8226 1 : }
8227 1 :
8228 11 : assert_eq!(cnt, NUM_KEYS);
8229 1 :
8230 11011 : for _ in 0..NUM_KEYS {
8231 11000 : lsn = Lsn(lsn.0 + 0x10);
8232 11000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
8233 11000 : test_key.field6 = (blknum * STEP) as u32;
8234 11000 : let mut writer = tline.writer().await;
8235 11000 : writer
8236 11000 : .put(
8237 11000 : test_key,
8238 11000 : lsn,
8239 11000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
8240 11000 : &ctx,
8241 11000 : )
8242 11000 : .await?;
8243 11000 : writer.finish_write(lsn);
8244 11000 : drop(writer);
8245 11000 : updated[blknum] = lsn;
8246 1 : }
8247 1 :
8248 1 : // Perform two cycles of flush, compact, and GC
8249 33 : for round in 0..2 {
8250 22 : tline.freeze_and_flush().await?;
8251 22 : tline
8252 22 : .compact(
8253 22 : &cancel,
8254 22 : if iter % 5 == 0 && round == 0 {
8255 3 : let mut flags = EnumSet::new();
8256 3 : flags.insert(CompactFlags::ForceImageLayerCreation);
8257 3 : flags.insert(CompactFlags::ForceRepartition);
8258 3 : flags
8259 1 : } else {
8260 19 : EnumSet::empty()
8261 1 : },
8262 22 : &ctx,
8263 1 : )
8264 22 : .await?;
8265 22 : tenant
8266 22 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
8267 22 : .await?;
8268 1 : }
8269 1 : }
8270 1 :
8271 1 : Ok(())
8272 1 : }
8273 :
8274 : #[tokio::test]
8275 1 : async fn test_metadata_compaction_trigger() -> anyhow::Result<()> {
8276 1 : let harness = TenantHarness::create("test_metadata_compaction_trigger").await?;
8277 1 : let (tenant, ctx) = harness.load().await;
8278 1 : let tline = tenant
8279 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
8280 1 : .await?;
8281 :
8282 1 : let cancel = CancellationToken::new();
8283 :
8284 1 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
8285 1 : base_key.field1 = AUX_KEY_PREFIX;
8286 1 : let test_key = base_key;
8287 1 : let mut lsn = Lsn(0x10);
8288 :
8289 21 : for _ in 0..20 {
8290 20 : lsn = Lsn(lsn.0 + 0x10);
8291 20 : let mut writer = tline.writer().await;
8292 20 : writer
8293 20 : .put(
8294 20 : test_key,
8295 20 : lsn,
8296 20 : &Value::Image(test_img(&format!("{} at {}", 0, lsn))),
8297 20 : &ctx,
8298 20 : )
8299 20 : .await?;
8300 20 : writer.finish_write(lsn);
8301 20 : drop(writer);
8302 20 : tline.freeze_and_flush().await?; // force create a delta layer
8303 : }
8304 :
8305 1 : let before_num_l0_delta_files = tline
8306 1 : .layers
8307 1 : .read(LayerManagerLockHolder::Testing)
8308 1 : .await
8309 1 : .layer_map()?
8310 1 : .level0_deltas()
8311 1 : .len();
8312 :
8313 1 : tline.compact(&cancel, EnumSet::default(), &ctx).await?;
8314 :
8315 1 : let after_num_l0_delta_files = tline
8316 1 : .layers
8317 1 : .read(LayerManagerLockHolder::Testing)
8318 1 : .await
8319 1 : .layer_map()?
8320 1 : .level0_deltas()
8321 1 : .len();
8322 :
8323 1 : assert!(
8324 1 : after_num_l0_delta_files < before_num_l0_delta_files,
8325 0 : "after_num_l0_delta_files={after_num_l0_delta_files}, before_num_l0_delta_files={before_num_l0_delta_files}"
8326 : );
8327 :
8328 1 : assert_eq!(
8329 1 : tline.get(test_key, lsn, &ctx).await?,
8330 1 : test_img(&format!("{} at {}", 0, lsn))
8331 : );
8332 :
8333 2 : Ok(())
8334 1 : }
8335 :
8336 : #[tokio::test]
8337 1 : async fn test_aux_file_e2e() {
8338 1 : let harness = TenantHarness::create("test_aux_file_e2e").await.unwrap();
8339 :
8340 1 : let (tenant, ctx) = harness.load().await;
8341 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8342 :
8343 1 : let mut lsn = Lsn(0x08);
8344 :
8345 1 : let tline: Arc<Timeline> = tenant
8346 1 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
8347 1 : .await
8348 1 : .unwrap();
8349 :
8350 : {
8351 1 : lsn += 8;
8352 1 : let mut modification = tline.begin_modification(lsn);
8353 1 : modification
8354 1 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
8355 1 : .await
8356 1 : .unwrap();
8357 1 : modification.commit(&ctx).await.unwrap();
8358 : }
8359 :
8360 : // we can read everything from the storage
8361 1 : let files = tline
8362 1 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
8363 1 : .await
8364 1 : .unwrap();
8365 1 : assert_eq!(
8366 1 : files.get("pg_logical/mappings/test1"),
8367 1 : Some(&bytes::Bytes::from_static(b"first"))
8368 : );
8369 :
8370 : {
8371 1 : lsn += 8;
8372 1 : let mut modification = tline.begin_modification(lsn);
8373 1 : modification
8374 1 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
8375 1 : .await
8376 1 : .unwrap();
8377 1 : modification.commit(&ctx).await.unwrap();
8378 : }
8379 :
8380 1 : let files = tline
8381 1 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
8382 1 : .await
8383 1 : .unwrap();
8384 1 : assert_eq!(
8385 1 : files.get("pg_logical/mappings/test2"),
8386 1 : Some(&bytes::Bytes::from_static(b"second"))
8387 : );
8388 :
8389 1 : let child = tenant
8390 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(lsn), &ctx)
8391 1 : .await
8392 1 : .unwrap();
8393 :
8394 1 : let files = child
8395 1 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
8396 1 : .await
8397 1 : .unwrap();
8398 1 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
8399 1 : assert_eq!(files.get("pg_logical/mappings/test2"), None);
8400 1 : }
8401 :
8402 : #[tokio::test]
8403 1 : async fn test_repl_origin_tombstones() {
8404 1 : let harness = TenantHarness::create("test_repl_origin_tombstones")
8405 1 : .await
8406 1 : .unwrap();
8407 :
8408 1 : let (tenant, ctx) = harness.load().await;
8409 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8410 :
8411 1 : let mut lsn = Lsn(0x08);
8412 :
8413 1 : let tline: Arc<Timeline> = tenant
8414 1 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
8415 1 : .await
8416 1 : .unwrap();
8417 :
8418 1 : let repl_lsn = Lsn(0x10);
8419 : {
8420 1 : lsn += 8;
8421 1 : let mut modification = tline.begin_modification(lsn);
8422 1 : modification.put_for_unit_test(repl_origin_key(2), Value::Image(Bytes::new()));
8423 1 : modification.set_replorigin(1, repl_lsn).await.unwrap();
8424 1 : modification.commit(&ctx).await.unwrap();
8425 : }
8426 :
8427 : // we can read everything from the storage
8428 1 : let repl_origins = tline
8429 1 : .get_replorigins(lsn, &ctx, io_concurrency.clone())
8430 1 : .await
8431 1 : .unwrap();
8432 1 : assert_eq!(repl_origins.len(), 1);
8433 1 : assert_eq!(repl_origins[&1], lsn);
8434 :
8435 : {
8436 1 : lsn += 8;
8437 1 : let mut modification = tline.begin_modification(lsn);
8438 1 : modification.put_for_unit_test(
8439 1 : repl_origin_key(3),
8440 1 : Value::Image(Bytes::copy_from_slice(b"cannot_decode_this")),
8441 : );
8442 1 : modification.commit(&ctx).await.unwrap();
8443 : }
8444 1 : let result = tline
8445 1 : .get_replorigins(lsn, &ctx, io_concurrency.clone())
8446 1 : .await;
8447 1 : assert!(result.is_err());
8448 1 : }
8449 :
8450 : #[tokio::test]
8451 1 : async fn test_metadata_image_creation() -> anyhow::Result<()> {
8452 1 : let harness = TenantHarness::create("test_metadata_image_creation").await?;
8453 1 : let (tenant, ctx) = harness.load().await;
8454 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8455 1 : let tline = tenant
8456 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
8457 1 : .await?;
8458 :
8459 : const NUM_KEYS: usize = 1000;
8460 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
8461 :
8462 1 : let cancel = CancellationToken::new();
8463 :
8464 1 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8465 1 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
8466 1 : let mut test_key = base_key;
8467 1 : let mut lsn = Lsn(0x10);
8468 :
8469 4 : async fn scan_with_statistics(
8470 4 : tline: &Timeline,
8471 4 : keyspace: &KeySpace,
8472 4 : lsn: Lsn,
8473 4 : ctx: &RequestContext,
8474 4 : io_concurrency: IoConcurrency,
8475 4 : ) -> anyhow::Result<(BTreeMap<Key, Result<Bytes, PageReconstructError>>, usize)> {
8476 4 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
8477 4 : let query = VersionedKeySpaceQuery::uniform(keyspace.clone(), lsn);
8478 4 : let res = tline
8479 4 : .get_vectored_impl(query, &mut reconstruct_state, ctx)
8480 4 : .await?;
8481 4 : Ok((res, reconstruct_state.get_delta_layers_visited() as usize))
8482 4 : }
8483 :
8484 1001 : for blknum in 0..NUM_KEYS {
8485 1000 : lsn = Lsn(lsn.0 + 0x10);
8486 1000 : test_key.field6 = (blknum * STEP) as u32;
8487 1000 : let mut writer = tline.writer().await;
8488 1000 : writer
8489 1000 : .put(
8490 1000 : test_key,
8491 1000 : lsn,
8492 1000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
8493 1000 : &ctx,
8494 1000 : )
8495 1000 : .await?;
8496 1000 : writer.finish_write(lsn);
8497 1000 : drop(writer);
8498 : }
8499 :
8500 1 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
8501 :
8502 11 : for iter in 1..=10 {
8503 10010 : for _ in 0..NUM_KEYS {
8504 10000 : lsn = Lsn(lsn.0 + 0x10);
8505 10000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
8506 10000 : test_key.field6 = (blknum * STEP) as u32;
8507 10000 : let mut writer = tline.writer().await;
8508 10000 : writer
8509 10000 : .put(
8510 10000 : test_key,
8511 10000 : lsn,
8512 10000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
8513 10000 : &ctx,
8514 10000 : )
8515 10000 : .await?;
8516 10000 : writer.finish_write(lsn);
8517 10000 : drop(writer);
8518 1 : }
8519 1 :
8520 10 : tline.freeze_and_flush().await?;
8521 1 : // Force layers to L1
8522 10 : tline
8523 10 : .compact(
8524 10 : &cancel,
8525 10 : {
8526 10 : let mut flags = EnumSet::new();
8527 10 : flags.insert(CompactFlags::ForceL0Compaction);
8528 10 : flags
8529 10 : },
8530 10 : &ctx,
8531 10 : )
8532 10 : .await?;
8533 1 :
8534 10 : if iter % 5 == 0 {
8535 2 : let scan_lsn = Lsn(lsn.0 + 1);
8536 2 : info!("scanning at {}", scan_lsn);
8537 2 : let (_, before_delta_file_accessed) =
8538 2 : scan_with_statistics(&tline, &keyspace, scan_lsn, &ctx, io_concurrency.clone())
8539 2 : .await?;
8540 2 : tline
8541 2 : .compact(
8542 2 : &cancel,
8543 2 : {
8544 2 : let mut flags = EnumSet::new();
8545 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
8546 2 : flags.insert(CompactFlags::ForceRepartition);
8547 2 : flags.insert(CompactFlags::ForceL0Compaction);
8548 2 : flags
8549 2 : },
8550 2 : &ctx,
8551 2 : )
8552 2 : .await?;
8553 2 : let (_, after_delta_file_accessed) =
8554 2 : scan_with_statistics(&tline, &keyspace, scan_lsn, &ctx, io_concurrency.clone())
8555 2 : .await?;
8556 2 : assert!(
8557 2 : after_delta_file_accessed < before_delta_file_accessed,
8558 1 : "after_delta_file_accessed={after_delta_file_accessed}, before_delta_file_accessed={before_delta_file_accessed}"
8559 1 : );
8560 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.
8561 2 : assert!(
8562 2 : after_delta_file_accessed <= 2,
8563 1 : "after_delta_file_accessed={after_delta_file_accessed}"
8564 1 : );
8565 8 : }
8566 1 : }
8567 1 :
8568 1 : Ok(())
8569 1 : }
8570 :
8571 : #[tokio::test]
8572 1 : async fn test_vectored_missing_data_key_reads() -> anyhow::Result<()> {
8573 1 : let harness = TenantHarness::create("test_vectored_missing_data_key_reads").await?;
8574 1 : let (tenant, ctx) = harness.load().await;
8575 :
8576 1 : let base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
8577 1 : let base_key_child = Key::from_hex("000000000033333333444444445500000001").unwrap();
8578 1 : let base_key_nonexist = Key::from_hex("000000000033333333444444445500000002").unwrap();
8579 :
8580 1 : let tline = tenant
8581 1 : .create_test_timeline_with_layers(
8582 1 : TIMELINE_ID,
8583 1 : Lsn(0x10),
8584 1 : DEFAULT_PG_VERSION,
8585 1 : &ctx,
8586 1 : Vec::new(), // in-memory layers
8587 1 : Vec::new(), // delta layers
8588 1 : vec![(Lsn(0x20), vec![(base_key, test_img("data key 1"))])], // image layers
8589 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
8590 1 : )
8591 1 : .await?;
8592 1 : tline.add_extra_test_dense_keyspace(KeySpace::single(base_key..(base_key_nonexist.next())));
8593 :
8594 1 : let child = tenant
8595 1 : .branch_timeline_test_with_layers(
8596 1 : &tline,
8597 1 : NEW_TIMELINE_ID,
8598 1 : Some(Lsn(0x20)),
8599 1 : &ctx,
8600 1 : Vec::new(), // delta layers
8601 1 : vec![(Lsn(0x30), vec![(base_key_child, test_img("data key 2"))])], // image layers
8602 1 : Lsn(0x30),
8603 1 : )
8604 1 : .await
8605 1 : .unwrap();
8606 :
8607 1 : let lsn = Lsn(0x30);
8608 :
8609 : // test vectored get on parent timeline
8610 1 : assert_eq!(
8611 1 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
8612 1 : Some(test_img("data key 1"))
8613 : );
8614 1 : assert!(
8615 1 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx)
8616 1 : .await
8617 1 : .unwrap_err()
8618 1 : .is_missing_key_error()
8619 : );
8620 1 : assert!(
8621 1 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx)
8622 1 : .await
8623 1 : .unwrap_err()
8624 1 : .is_missing_key_error()
8625 : );
8626 :
8627 : // test vectored get on child timeline
8628 1 : assert_eq!(
8629 1 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
8630 1 : Some(test_img("data key 1"))
8631 : );
8632 1 : assert_eq!(
8633 1 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
8634 1 : Some(test_img("data key 2"))
8635 : );
8636 1 : assert!(
8637 1 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx)
8638 1 : .await
8639 1 : .unwrap_err()
8640 1 : .is_missing_key_error()
8641 : );
8642 :
8643 2 : Ok(())
8644 1 : }
8645 :
8646 : #[tokio::test]
8647 1 : async fn test_vectored_missing_metadata_key_reads() -> anyhow::Result<()> {
8648 1 : let harness = TenantHarness::create("test_vectored_missing_metadata_key_reads").await?;
8649 1 : let (tenant, ctx) = harness.load().await;
8650 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8651 :
8652 1 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8653 1 : let base_key_child = Key::from_hex("620000000033333333444444445500000001").unwrap();
8654 1 : let base_key_nonexist = Key::from_hex("620000000033333333444444445500000002").unwrap();
8655 1 : let base_key_overwrite = Key::from_hex("620000000033333333444444445500000003").unwrap();
8656 :
8657 1 : let base_inherited_key = Key::from_hex("610000000033333333444444445500000000").unwrap();
8658 1 : let base_inherited_key_child =
8659 1 : Key::from_hex("610000000033333333444444445500000001").unwrap();
8660 1 : let base_inherited_key_nonexist =
8661 1 : Key::from_hex("610000000033333333444444445500000002").unwrap();
8662 1 : let base_inherited_key_overwrite =
8663 1 : Key::from_hex("610000000033333333444444445500000003").unwrap();
8664 :
8665 1 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
8666 1 : assert_eq!(base_inherited_key.field1, RELATION_SIZE_PREFIX);
8667 :
8668 1 : let tline = tenant
8669 1 : .create_test_timeline_with_layers(
8670 1 : TIMELINE_ID,
8671 1 : Lsn(0x10),
8672 1 : DEFAULT_PG_VERSION,
8673 1 : &ctx,
8674 1 : Vec::new(), // in-memory layers
8675 1 : Vec::new(), // delta layers
8676 1 : vec![(
8677 1 : Lsn(0x20),
8678 1 : vec![
8679 1 : (base_inherited_key, test_img("metadata inherited key 1")),
8680 1 : (
8681 1 : base_inherited_key_overwrite,
8682 1 : test_img("metadata key overwrite 1a"),
8683 1 : ),
8684 1 : (base_key, test_img("metadata key 1")),
8685 1 : (base_key_overwrite, test_img("metadata key overwrite 1b")),
8686 1 : ],
8687 1 : )], // image layers
8688 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
8689 1 : )
8690 1 : .await?;
8691 :
8692 1 : let child = tenant
8693 1 : .branch_timeline_test_with_layers(
8694 1 : &tline,
8695 1 : NEW_TIMELINE_ID,
8696 1 : Some(Lsn(0x20)),
8697 1 : &ctx,
8698 1 : Vec::new(), // delta layers
8699 1 : vec![(
8700 1 : Lsn(0x30),
8701 1 : vec![
8702 1 : (
8703 1 : base_inherited_key_child,
8704 1 : test_img("metadata inherited key 2"),
8705 1 : ),
8706 1 : (
8707 1 : base_inherited_key_overwrite,
8708 1 : test_img("metadata key overwrite 2a"),
8709 1 : ),
8710 1 : (base_key_child, test_img("metadata key 2")),
8711 1 : (base_key_overwrite, test_img("metadata key overwrite 2b")),
8712 1 : ],
8713 1 : )], // image layers
8714 1 : Lsn(0x30),
8715 1 : )
8716 1 : .await
8717 1 : .unwrap();
8718 :
8719 1 : let lsn = Lsn(0x30);
8720 :
8721 : // test vectored get on parent timeline
8722 1 : assert_eq!(
8723 1 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
8724 1 : Some(test_img("metadata key 1"))
8725 : );
8726 1 : assert_eq!(
8727 1 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx).await?,
8728 : None
8729 : );
8730 1 : assert_eq!(
8731 1 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx).await?,
8732 : None
8733 : );
8734 1 : assert_eq!(
8735 1 : get_vectored_impl_wrapper(&tline, base_key_overwrite, lsn, &ctx).await?,
8736 1 : Some(test_img("metadata key overwrite 1b"))
8737 : );
8738 1 : assert_eq!(
8739 1 : get_vectored_impl_wrapper(&tline, base_inherited_key, lsn, &ctx).await?,
8740 1 : Some(test_img("metadata inherited key 1"))
8741 : );
8742 1 : assert_eq!(
8743 1 : get_vectored_impl_wrapper(&tline, base_inherited_key_child, lsn, &ctx).await?,
8744 : None
8745 : );
8746 1 : assert_eq!(
8747 1 : get_vectored_impl_wrapper(&tline, base_inherited_key_nonexist, lsn, &ctx).await?,
8748 : None
8749 : );
8750 1 : assert_eq!(
8751 1 : get_vectored_impl_wrapper(&tline, base_inherited_key_overwrite, lsn, &ctx).await?,
8752 1 : Some(test_img("metadata key overwrite 1a"))
8753 : );
8754 :
8755 : // test vectored get on child timeline
8756 1 : assert_eq!(
8757 1 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
8758 : None
8759 : );
8760 1 : assert_eq!(
8761 1 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
8762 1 : Some(test_img("metadata key 2"))
8763 : );
8764 1 : assert_eq!(
8765 1 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx).await?,
8766 : None
8767 : );
8768 1 : assert_eq!(
8769 1 : get_vectored_impl_wrapper(&child, base_inherited_key, lsn, &ctx).await?,
8770 1 : Some(test_img("metadata inherited key 1"))
8771 : );
8772 1 : assert_eq!(
8773 1 : get_vectored_impl_wrapper(&child, base_inherited_key_child, lsn, &ctx).await?,
8774 1 : Some(test_img("metadata inherited key 2"))
8775 : );
8776 1 : assert_eq!(
8777 1 : get_vectored_impl_wrapper(&child, base_inherited_key_nonexist, lsn, &ctx).await?,
8778 : None
8779 : );
8780 1 : assert_eq!(
8781 1 : get_vectored_impl_wrapper(&child, base_key_overwrite, lsn, &ctx).await?,
8782 1 : Some(test_img("metadata key overwrite 2b"))
8783 : );
8784 1 : assert_eq!(
8785 1 : get_vectored_impl_wrapper(&child, base_inherited_key_overwrite, lsn, &ctx).await?,
8786 1 : Some(test_img("metadata key overwrite 2a"))
8787 : );
8788 :
8789 : // test vectored scan on parent timeline
8790 1 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency.clone());
8791 1 : let query =
8792 1 : VersionedKeySpaceQuery::uniform(KeySpace::single(Key::metadata_key_range()), lsn);
8793 1 : let res = tline
8794 1 : .get_vectored_impl(query, &mut reconstruct_state, &ctx)
8795 1 : .await?;
8796 :
8797 1 : assert_eq!(
8798 1 : res.into_iter()
8799 4 : .map(|(k, v)| (k, v.unwrap()))
8800 1 : .collect::<Vec<_>>(),
8801 1 : vec![
8802 1 : (base_inherited_key, test_img("metadata inherited key 1")),
8803 1 : (
8804 1 : base_inherited_key_overwrite,
8805 1 : test_img("metadata key overwrite 1a")
8806 1 : ),
8807 1 : (base_key, test_img("metadata key 1")),
8808 1 : (base_key_overwrite, test_img("metadata key overwrite 1b")),
8809 : ]
8810 : );
8811 :
8812 : // test vectored scan on child timeline
8813 1 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency.clone());
8814 1 : let query =
8815 1 : VersionedKeySpaceQuery::uniform(KeySpace::single(Key::metadata_key_range()), lsn);
8816 1 : let res = child
8817 1 : .get_vectored_impl(query, &mut reconstruct_state, &ctx)
8818 1 : .await?;
8819 :
8820 1 : assert_eq!(
8821 1 : res.into_iter()
8822 5 : .map(|(k, v)| (k, v.unwrap()))
8823 1 : .collect::<Vec<_>>(),
8824 1 : vec![
8825 1 : (base_inherited_key, test_img("metadata inherited key 1")),
8826 1 : (
8827 1 : base_inherited_key_child,
8828 1 : test_img("metadata inherited key 2")
8829 1 : ),
8830 1 : (
8831 1 : base_inherited_key_overwrite,
8832 1 : test_img("metadata key overwrite 2a")
8833 1 : ),
8834 1 : (base_key_child, test_img("metadata key 2")),
8835 1 : (base_key_overwrite, test_img("metadata key overwrite 2b")),
8836 : ]
8837 : );
8838 :
8839 2 : Ok(())
8840 1 : }
8841 :
8842 28 : async fn get_vectored_impl_wrapper(
8843 28 : tline: &Arc<Timeline>,
8844 28 : key: Key,
8845 28 : lsn: Lsn,
8846 28 : ctx: &RequestContext,
8847 28 : ) -> Result<Option<Bytes>, GetVectoredError> {
8848 28 : let io_concurrency = IoConcurrency::spawn_from_conf(
8849 28 : tline.conf.get_vectored_concurrent_io,
8850 28 : tline.gate.enter().unwrap(),
8851 : );
8852 28 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
8853 28 : let query = VersionedKeySpaceQuery::uniform(KeySpace::single(key..key.next()), lsn);
8854 28 : let mut res = tline
8855 28 : .get_vectored_impl(query, &mut reconstruct_state, ctx)
8856 28 : .await?;
8857 25 : Ok(res.pop_last().map(|(k, v)| {
8858 16 : assert_eq!(k, key);
8859 16 : v.unwrap()
8860 16 : }))
8861 28 : }
8862 :
8863 : #[tokio::test]
8864 1 : async fn test_metadata_tombstone_reads() -> anyhow::Result<()> {
8865 1 : let harness = TenantHarness::create("test_metadata_tombstone_reads").await?;
8866 1 : let (tenant, ctx) = harness.load().await;
8867 1 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
8868 1 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8869 1 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8870 1 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
8871 :
8872 : // We emulate the situation that the compaction algorithm creates an image layer that removes the tombstones
8873 : // Lsn 0x30 key0, key3, no key1+key2
8874 : // Lsn 0x20 key1+key2 tomestones
8875 : // Lsn 0x10 key1 in image, key2 in delta
8876 1 : let tline = tenant
8877 1 : .create_test_timeline_with_layers(
8878 1 : TIMELINE_ID,
8879 1 : Lsn(0x10),
8880 1 : DEFAULT_PG_VERSION,
8881 1 : &ctx,
8882 1 : Vec::new(), // in-memory layers
8883 1 : // delta layers
8884 1 : vec![
8885 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8886 1 : Lsn(0x10)..Lsn(0x20),
8887 1 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8888 1 : ),
8889 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8890 1 : Lsn(0x20)..Lsn(0x30),
8891 1 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8892 1 : ),
8893 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8894 1 : Lsn(0x20)..Lsn(0x30),
8895 1 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8896 1 : ),
8897 1 : ],
8898 1 : // image layers
8899 1 : vec![
8900 1 : (Lsn(0x10), vec![(key1, test_img("metadata key 1"))]),
8901 1 : (
8902 1 : Lsn(0x30),
8903 1 : vec![
8904 1 : (key0, test_img("metadata key 0")),
8905 1 : (key3, test_img("metadata key 3")),
8906 1 : ],
8907 1 : ),
8908 1 : ],
8909 1 : Lsn(0x30),
8910 1 : )
8911 1 : .await?;
8912 :
8913 1 : let lsn = Lsn(0x30);
8914 1 : let old_lsn = Lsn(0x20);
8915 :
8916 1 : assert_eq!(
8917 1 : get_vectored_impl_wrapper(&tline, key0, lsn, &ctx).await?,
8918 1 : Some(test_img("metadata key 0"))
8919 : );
8920 1 : assert_eq!(
8921 1 : get_vectored_impl_wrapper(&tline, key1, lsn, &ctx).await?,
8922 : None,
8923 : );
8924 1 : assert_eq!(
8925 1 : get_vectored_impl_wrapper(&tline, key2, lsn, &ctx).await?,
8926 : None,
8927 : );
8928 1 : assert_eq!(
8929 1 : get_vectored_impl_wrapper(&tline, key1, old_lsn, &ctx).await?,
8930 1 : Some(Bytes::new()),
8931 : );
8932 1 : assert_eq!(
8933 1 : get_vectored_impl_wrapper(&tline, key2, old_lsn, &ctx).await?,
8934 1 : Some(Bytes::new()),
8935 : );
8936 1 : assert_eq!(
8937 1 : get_vectored_impl_wrapper(&tline, key3, lsn, &ctx).await?,
8938 1 : Some(test_img("metadata key 3"))
8939 : );
8940 :
8941 2 : Ok(())
8942 1 : }
8943 :
8944 : #[tokio::test]
8945 1 : async fn test_metadata_tombstone_image_creation() {
8946 1 : let harness = TenantHarness::create("test_metadata_tombstone_image_creation")
8947 1 : .await
8948 1 : .unwrap();
8949 1 : let (tenant, ctx) = harness.load().await;
8950 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8951 :
8952 1 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
8953 1 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8954 1 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8955 1 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
8956 :
8957 1 : let tline = tenant
8958 1 : .create_test_timeline_with_layers(
8959 1 : TIMELINE_ID,
8960 1 : Lsn(0x10),
8961 1 : DEFAULT_PG_VERSION,
8962 1 : &ctx,
8963 1 : Vec::new(), // in-memory layers
8964 1 : // delta layers
8965 1 : vec![
8966 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8967 1 : Lsn(0x10)..Lsn(0x20),
8968 1 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8969 1 : ),
8970 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8971 1 : Lsn(0x20)..Lsn(0x30),
8972 1 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8973 1 : ),
8974 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8975 1 : Lsn(0x20)..Lsn(0x30),
8976 1 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8977 1 : ),
8978 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8979 1 : Lsn(0x30)..Lsn(0x40),
8980 1 : vec![
8981 1 : (key0, Lsn(0x30), Value::Image(test_img("metadata key 0"))),
8982 1 : (key3, Lsn(0x30), Value::Image(test_img("metadata key 3"))),
8983 1 : ],
8984 1 : ),
8985 1 : ],
8986 1 : // image layers
8987 1 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
8988 1 : Lsn(0x40),
8989 1 : )
8990 1 : .await
8991 1 : .unwrap();
8992 :
8993 1 : let cancel = CancellationToken::new();
8994 :
8995 : // Image layer creation happens on the disk_consistent_lsn so we need to force set it now.
8996 1 : tline.force_set_disk_consistent_lsn(Lsn(0x40));
8997 1 : tline
8998 1 : .compact(
8999 1 : &cancel,
9000 1 : {
9001 1 : let mut flags = EnumSet::new();
9002 1 : flags.insert(CompactFlags::ForceImageLayerCreation);
9003 1 : flags.insert(CompactFlags::ForceRepartition);
9004 1 : flags
9005 1 : },
9006 1 : &ctx,
9007 1 : )
9008 1 : .await
9009 1 : .unwrap();
9010 : // Image layers are created at repartition LSN
9011 1 : let images = tline
9012 1 : .inspect_image_layers(Lsn(0x40), &ctx, io_concurrency.clone())
9013 1 : .await
9014 1 : .unwrap()
9015 1 : .into_iter()
9016 9 : .filter(|(k, _)| k.is_metadata_key())
9017 1 : .collect::<Vec<_>>();
9018 1 : assert_eq!(images.len(), 2); // the image layer should only contain two existing keys, tombstones should be removed.
9019 1 : }
9020 :
9021 : #[tokio::test]
9022 1 : async fn test_metadata_tombstone_empty_image_creation() {
9023 1 : let harness = TenantHarness::create("test_metadata_tombstone_empty_image_creation")
9024 1 : .await
9025 1 : .unwrap();
9026 1 : let (tenant, ctx) = harness.load().await;
9027 1 : let io_concurrency = IoConcurrency::spawn_for_test();
9028 :
9029 1 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
9030 1 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
9031 :
9032 1 : let tline = tenant
9033 1 : .create_test_timeline_with_layers(
9034 1 : TIMELINE_ID,
9035 1 : Lsn(0x10),
9036 1 : DEFAULT_PG_VERSION,
9037 1 : &ctx,
9038 1 : Vec::new(), // in-memory layers
9039 1 : // delta layers
9040 1 : vec![
9041 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9042 1 : Lsn(0x10)..Lsn(0x20),
9043 1 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
9044 1 : ),
9045 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9046 1 : Lsn(0x20)..Lsn(0x30),
9047 1 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
9048 1 : ),
9049 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9050 1 : Lsn(0x20)..Lsn(0x30),
9051 1 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
9052 1 : ),
9053 1 : ],
9054 1 : // image layers
9055 1 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
9056 1 : Lsn(0x30),
9057 1 : )
9058 1 : .await
9059 1 : .unwrap();
9060 :
9061 1 : let cancel = CancellationToken::new();
9062 :
9063 1 : tline
9064 1 : .compact(
9065 1 : &cancel,
9066 1 : {
9067 1 : let mut flags = EnumSet::new();
9068 1 : flags.insert(CompactFlags::ForceImageLayerCreation);
9069 1 : flags.insert(CompactFlags::ForceRepartition);
9070 1 : flags
9071 1 : },
9072 1 : &ctx,
9073 1 : )
9074 1 : .await
9075 1 : .unwrap();
9076 :
9077 : // Image layers are created at last_record_lsn
9078 1 : let images = tline
9079 1 : .inspect_image_layers(Lsn(0x30), &ctx, io_concurrency.clone())
9080 1 : .await
9081 1 : .unwrap()
9082 1 : .into_iter()
9083 7 : .filter(|(k, _)| k.is_metadata_key())
9084 1 : .collect::<Vec<_>>();
9085 1 : assert_eq!(images.len(), 0); // the image layer should not contain tombstones, or it is not created
9086 1 : }
9087 :
9088 : #[tokio::test]
9089 1 : async fn test_simple_bottom_most_compaction_images() -> anyhow::Result<()> {
9090 1 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_images").await?;
9091 1 : let (tenant, ctx) = harness.load().await;
9092 1 : let io_concurrency = IoConcurrency::spawn_for_test();
9093 :
9094 51 : fn get_key(id: u32) -> Key {
9095 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9096 51 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9097 51 : key.field6 = id;
9098 51 : key
9099 51 : }
9100 :
9101 : // We create
9102 : // - one bottom-most image layer,
9103 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
9104 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
9105 : // - a delta layer D3 above the horizon.
9106 : //
9107 : // | D3 |
9108 : // | D1 |
9109 : // -| |-- gc horizon -----------------
9110 : // | | | D2 |
9111 : // --------- img layer ------------------
9112 : //
9113 : // What we should expact from this compaction is:
9114 : // | D3 |
9115 : // | Part of D1 |
9116 : // --------- img layer with D1+D2 at GC horizon------------------
9117 :
9118 : // img layer at 0x10
9119 1 : let img_layer = (0..10)
9120 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9121 1 : .collect_vec();
9122 :
9123 1 : let delta1 = vec![
9124 1 : (
9125 1 : get_key(1),
9126 1 : Lsn(0x20),
9127 1 : Value::Image(Bytes::from("value 1@0x20")),
9128 1 : ),
9129 1 : (
9130 1 : get_key(2),
9131 1 : Lsn(0x30),
9132 1 : Value::Image(Bytes::from("value 2@0x30")),
9133 1 : ),
9134 1 : (
9135 1 : get_key(3),
9136 1 : Lsn(0x40),
9137 1 : Value::Image(Bytes::from("value 3@0x40")),
9138 1 : ),
9139 : ];
9140 1 : let delta2 = vec![
9141 1 : (
9142 1 : get_key(5),
9143 1 : Lsn(0x20),
9144 1 : Value::Image(Bytes::from("value 5@0x20")),
9145 1 : ),
9146 1 : (
9147 1 : get_key(6),
9148 1 : Lsn(0x20),
9149 1 : Value::Image(Bytes::from("value 6@0x20")),
9150 1 : ),
9151 : ];
9152 1 : let delta3 = vec![
9153 1 : (
9154 1 : get_key(8),
9155 1 : Lsn(0x48),
9156 1 : Value::Image(Bytes::from("value 8@0x48")),
9157 1 : ),
9158 1 : (
9159 1 : get_key(9),
9160 1 : Lsn(0x48),
9161 1 : Value::Image(Bytes::from("value 9@0x48")),
9162 1 : ),
9163 : ];
9164 :
9165 1 : let tline = tenant
9166 1 : .create_test_timeline_with_layers(
9167 1 : TIMELINE_ID,
9168 1 : Lsn(0x10),
9169 1 : DEFAULT_PG_VERSION,
9170 1 : &ctx,
9171 1 : Vec::new(), // in-memory layers
9172 1 : vec![
9173 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
9174 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
9175 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
9176 1 : ], // delta layers
9177 1 : vec![(Lsn(0x10), img_layer)], // image layers
9178 1 : Lsn(0x50),
9179 1 : )
9180 1 : .await?;
9181 : {
9182 1 : tline
9183 1 : .applied_gc_cutoff_lsn
9184 1 : .lock_for_write()
9185 1 : .store_and_unlock(Lsn(0x30))
9186 1 : .wait()
9187 1 : .await;
9188 : // Update GC info
9189 1 : let mut guard = tline.gc_info.write().unwrap();
9190 1 : guard.cutoffs.time = Some(Lsn(0x30));
9191 1 : guard.cutoffs.space = Lsn(0x30);
9192 : }
9193 :
9194 1 : let expected_result = [
9195 1 : Bytes::from_static(b"value 0@0x10"),
9196 1 : Bytes::from_static(b"value 1@0x20"),
9197 1 : Bytes::from_static(b"value 2@0x30"),
9198 1 : Bytes::from_static(b"value 3@0x40"),
9199 1 : Bytes::from_static(b"value 4@0x10"),
9200 1 : Bytes::from_static(b"value 5@0x20"),
9201 1 : Bytes::from_static(b"value 6@0x20"),
9202 1 : Bytes::from_static(b"value 7@0x10"),
9203 1 : Bytes::from_static(b"value 8@0x48"),
9204 1 : Bytes::from_static(b"value 9@0x48"),
9205 1 : ];
9206 :
9207 10 : for (idx, expected) in expected_result.iter().enumerate() {
9208 10 : assert_eq!(
9209 10 : tline
9210 10 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9211 10 : .await
9212 10 : .unwrap(),
9213 : expected
9214 : );
9215 : }
9216 :
9217 1 : let cancel = CancellationToken::new();
9218 1 : tline
9219 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9220 1 : .await
9221 1 : .unwrap();
9222 :
9223 10 : for (idx, expected) in expected_result.iter().enumerate() {
9224 10 : assert_eq!(
9225 10 : tline
9226 10 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9227 10 : .await
9228 10 : .unwrap(),
9229 : expected
9230 : );
9231 : }
9232 :
9233 : // Check if the image layer at the GC horizon contains exactly what we want
9234 1 : let image_at_gc_horizon = tline
9235 1 : .inspect_image_layers(Lsn(0x30), &ctx, io_concurrency.clone())
9236 1 : .await
9237 1 : .unwrap()
9238 1 : .into_iter()
9239 17 : .filter(|(k, _)| k.is_metadata_key())
9240 1 : .collect::<Vec<_>>();
9241 :
9242 1 : assert_eq!(image_at_gc_horizon.len(), 10);
9243 1 : let expected_result = [
9244 1 : Bytes::from_static(b"value 0@0x10"),
9245 1 : Bytes::from_static(b"value 1@0x20"),
9246 1 : Bytes::from_static(b"value 2@0x30"),
9247 1 : Bytes::from_static(b"value 3@0x10"),
9248 1 : Bytes::from_static(b"value 4@0x10"),
9249 1 : Bytes::from_static(b"value 5@0x20"),
9250 1 : Bytes::from_static(b"value 6@0x20"),
9251 1 : Bytes::from_static(b"value 7@0x10"),
9252 1 : Bytes::from_static(b"value 8@0x10"),
9253 1 : Bytes::from_static(b"value 9@0x10"),
9254 1 : ];
9255 11 : for idx in 0..10 {
9256 10 : assert_eq!(
9257 10 : image_at_gc_horizon[idx],
9258 10 : (get_key(idx as u32), expected_result[idx].clone())
9259 : );
9260 : }
9261 :
9262 : // Check if old layers are removed / new layers have the expected LSN
9263 1 : let all_layers = inspect_and_sort(&tline, None).await;
9264 1 : assert_eq!(
9265 : all_layers,
9266 1 : vec![
9267 : // Image layer at GC horizon
9268 1 : PersistentLayerKey {
9269 1 : key_range: Key::MIN..Key::MAX,
9270 1 : lsn_range: Lsn(0x30)..Lsn(0x31),
9271 1 : is_delta: false
9272 1 : },
9273 : // The delta layer below the horizon
9274 1 : PersistentLayerKey {
9275 1 : key_range: get_key(3)..get_key(4),
9276 1 : lsn_range: Lsn(0x30)..Lsn(0x48),
9277 1 : is_delta: true
9278 1 : },
9279 : // The delta3 layer that should not be picked for the compaction
9280 1 : PersistentLayerKey {
9281 1 : key_range: get_key(8)..get_key(10),
9282 1 : lsn_range: Lsn(0x48)..Lsn(0x50),
9283 1 : is_delta: true
9284 1 : }
9285 : ]
9286 : );
9287 :
9288 : // increase GC horizon and compact again
9289 : {
9290 1 : tline
9291 1 : .applied_gc_cutoff_lsn
9292 1 : .lock_for_write()
9293 1 : .store_and_unlock(Lsn(0x40))
9294 1 : .wait()
9295 1 : .await;
9296 : // Update GC info
9297 1 : let mut guard = tline.gc_info.write().unwrap();
9298 1 : guard.cutoffs.time = Some(Lsn(0x40));
9299 1 : guard.cutoffs.space = Lsn(0x40);
9300 : }
9301 1 : tline
9302 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9303 1 : .await
9304 1 : .unwrap();
9305 :
9306 2 : Ok(())
9307 1 : }
9308 :
9309 : #[cfg(feature = "testing")]
9310 : #[tokio::test]
9311 1 : async fn test_neon_test_record() -> anyhow::Result<()> {
9312 1 : let harness = TenantHarness::create("test_neon_test_record").await?;
9313 1 : let (tenant, ctx) = harness.load().await;
9314 :
9315 17 : fn get_key(id: u32) -> Key {
9316 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9317 17 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9318 17 : key.field6 = id;
9319 17 : key
9320 17 : }
9321 :
9322 1 : let delta1 = vec![
9323 1 : (
9324 1 : get_key(1),
9325 1 : Lsn(0x20),
9326 1 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
9327 1 : ),
9328 1 : (
9329 1 : get_key(1),
9330 1 : Lsn(0x30),
9331 1 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
9332 1 : ),
9333 1 : (get_key(2), Lsn(0x10), Value::Image("0x10".into())),
9334 1 : (
9335 1 : get_key(2),
9336 1 : Lsn(0x20),
9337 1 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
9338 1 : ),
9339 1 : (
9340 1 : get_key(2),
9341 1 : Lsn(0x30),
9342 1 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
9343 1 : ),
9344 1 : (get_key(3), Lsn(0x10), Value::Image("0x10".into())),
9345 1 : (
9346 1 : get_key(3),
9347 1 : Lsn(0x20),
9348 1 : Value::WalRecord(NeonWalRecord::wal_clear("c")),
9349 1 : ),
9350 1 : (get_key(4), Lsn(0x10), Value::Image("0x10".into())),
9351 1 : (
9352 1 : get_key(4),
9353 1 : Lsn(0x20),
9354 1 : Value::WalRecord(NeonWalRecord::wal_init("i")),
9355 1 : ),
9356 1 : (
9357 1 : get_key(4),
9358 1 : Lsn(0x30),
9359 1 : Value::WalRecord(NeonWalRecord::wal_append_conditional("j", "i")),
9360 1 : ),
9361 1 : (
9362 1 : get_key(5),
9363 1 : Lsn(0x20),
9364 1 : Value::WalRecord(NeonWalRecord::wal_init("1")),
9365 1 : ),
9366 1 : (
9367 1 : get_key(5),
9368 1 : Lsn(0x30),
9369 1 : Value::WalRecord(NeonWalRecord::wal_append_conditional("j", "2")),
9370 1 : ),
9371 : ];
9372 1 : let image1 = vec![(get_key(1), "0x10".into())];
9373 :
9374 1 : let tline = tenant
9375 1 : .create_test_timeline_with_layers(
9376 1 : TIMELINE_ID,
9377 1 : Lsn(0x10),
9378 1 : DEFAULT_PG_VERSION,
9379 1 : &ctx,
9380 1 : Vec::new(), // in-memory layers
9381 1 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
9382 1 : Lsn(0x10)..Lsn(0x40),
9383 1 : delta1,
9384 1 : )], // delta layers
9385 1 : vec![(Lsn(0x10), image1)], // image layers
9386 1 : Lsn(0x50),
9387 1 : )
9388 1 : .await?;
9389 :
9390 1 : assert_eq!(
9391 1 : tline.get(get_key(1), Lsn(0x50), &ctx).await?,
9392 1 : Bytes::from_static(b"0x10,0x20,0x30")
9393 : );
9394 1 : assert_eq!(
9395 1 : tline.get(get_key(2), Lsn(0x50), &ctx).await?,
9396 1 : Bytes::from_static(b"0x10,0x20,0x30")
9397 : );
9398 :
9399 : // Need to remove the limit of "Neon WAL redo requires base image".
9400 :
9401 1 : assert_eq!(
9402 1 : tline.get(get_key(3), Lsn(0x50), &ctx).await?,
9403 1 : Bytes::from_static(b"c")
9404 : );
9405 1 : assert_eq!(
9406 1 : tline.get(get_key(4), Lsn(0x50), &ctx).await?,
9407 1 : Bytes::from_static(b"ij")
9408 : );
9409 :
9410 : // Manual testing required: currently, read errors will panic the process in debug mode. So we
9411 : // cannot enable this assertion in the unit test.
9412 : // assert!(tline.get(get_key(5), Lsn(0x50), &ctx).await.is_err());
9413 :
9414 2 : Ok(())
9415 1 : }
9416 :
9417 : #[tokio::test]
9418 1 : async fn test_lsn_lease() -> anyhow::Result<()> {
9419 1 : let (tenant, ctx) = TenantHarness::create("test_lsn_lease")
9420 1 : .await
9421 1 : .unwrap()
9422 1 : .load()
9423 1 : .await;
9424 : // set a non-zero lease length to test the feature
9425 1 : tenant
9426 1 : .update_tenant_config(|mut conf| {
9427 1 : conf.lsn_lease_length = Some(LsnLease::DEFAULT_LENGTH);
9428 1 : Ok(conf)
9429 1 : })
9430 1 : .unwrap();
9431 :
9432 1 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
9433 :
9434 1 : let end_lsn = Lsn(0x100);
9435 1 : let image_layers = (0x20..=0x90)
9436 1 : .step_by(0x10)
9437 8 : .map(|n| (Lsn(n), vec![(key, test_img(&format!("data key at {n:x}")))]))
9438 1 : .collect();
9439 :
9440 1 : let timeline = tenant
9441 1 : .create_test_timeline_with_layers(
9442 1 : TIMELINE_ID,
9443 1 : Lsn(0x10),
9444 1 : DEFAULT_PG_VERSION,
9445 1 : &ctx,
9446 1 : Vec::new(), // in-memory layers
9447 1 : Vec::new(),
9448 1 : image_layers,
9449 1 : end_lsn,
9450 1 : )
9451 1 : .await?;
9452 :
9453 1 : let leased_lsns = [0x30, 0x50, 0x70];
9454 1 : let mut leases = Vec::new();
9455 3 : leased_lsns.iter().for_each(|n| {
9456 3 : leases.push(
9457 3 : timeline
9458 3 : .init_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)
9459 3 : .expect("lease request should succeed"),
9460 : );
9461 3 : });
9462 :
9463 1 : let updated_lease_0 = timeline
9464 1 : .renew_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)
9465 1 : .expect("lease renewal should succeed");
9466 1 : assert_eq!(
9467 1 : updated_lease_0.valid_until, leases[0].valid_until,
9468 0 : " Renewing with shorter lease should not change the lease."
9469 : );
9470 :
9471 1 : let updated_lease_1 = timeline
9472 1 : .renew_lsn_lease(
9473 1 : Lsn(leased_lsns[1]),
9474 1 : timeline.get_lsn_lease_length() * 2,
9475 1 : &ctx,
9476 1 : )
9477 1 : .expect("lease renewal should succeed");
9478 1 : assert!(
9479 1 : updated_lease_1.valid_until > leases[1].valid_until,
9480 0 : "Renewing with a long lease should renew lease with later expiration time."
9481 : );
9482 :
9483 : // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
9484 1 : info!(
9485 0 : "applied_gc_cutoff_lsn: {}",
9486 0 : *timeline.get_applied_gc_cutoff_lsn()
9487 : );
9488 1 : timeline.force_set_disk_consistent_lsn(end_lsn);
9489 :
9490 1 : let res = tenant
9491 1 : .gc_iteration(
9492 1 : Some(TIMELINE_ID),
9493 1 : 0,
9494 1 : Duration::ZERO,
9495 1 : &CancellationToken::new(),
9496 1 : &ctx,
9497 1 : )
9498 1 : .await
9499 1 : .unwrap();
9500 :
9501 : // Keeping everything <= Lsn(0x80) b/c leases:
9502 : // 0/10: initdb layer
9503 : // (0/20..=0/70).step_by(0x10): image layers added when creating the timeline.
9504 1 : assert_eq!(res.layers_needed_by_leases, 7);
9505 : // Keeping 0/90 b/c it is the latest layer.
9506 1 : assert_eq!(res.layers_not_updated, 1);
9507 : // Removed 0/80.
9508 1 : assert_eq!(res.layers_removed, 1);
9509 :
9510 : // Make lease on a already GC-ed LSN.
9511 : // 0/80 does not have a valid lease + is below latest_gc_cutoff
9512 1 : assert!(Lsn(0x80) < *timeline.get_applied_gc_cutoff_lsn());
9513 1 : timeline
9514 1 : .init_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx)
9515 1 : .expect_err("lease request on GC-ed LSN should fail");
9516 :
9517 : // Should still be able to renew a currently valid lease
9518 : // Assumption: original lease to is still valid for 0/50.
9519 : // (use `Timeline::init_lsn_lease` for testing so it always does validation)
9520 1 : timeline
9521 1 : .init_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)
9522 1 : .expect("lease renewal with validation should succeed");
9523 :
9524 2 : Ok(())
9525 1 : }
9526 :
9527 : #[tokio::test]
9528 1 : async fn test_failed_flush_should_not_update_disk_consistent_lsn() -> anyhow::Result<()> {
9529 : //
9530 : // Setup
9531 : //
9532 1 : let harness = TenantHarness::create_custom(
9533 1 : "test_failed_flush_should_not_upload_disk_consistent_lsn",
9534 1 : pageserver_api::models::TenantConfig::default(),
9535 1 : TenantId::generate(),
9536 1 : ShardIdentity::new(ShardNumber(0), ShardCount(4), ShardStripeSize(128)).unwrap(),
9537 1 : Generation::new(1),
9538 1 : )
9539 1 : .await?;
9540 1 : let (tenant, ctx) = harness.load().await;
9541 :
9542 1 : let timeline = tenant
9543 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
9544 1 : .await?;
9545 1 : assert_eq!(timeline.get_shard_identity().count, ShardCount(4));
9546 1 : let mut writer = timeline.writer().await;
9547 1 : writer
9548 1 : .put(
9549 1 : *TEST_KEY,
9550 1 : Lsn(0x20),
9551 1 : &Value::Image(test_img("foo at 0x20")),
9552 1 : &ctx,
9553 1 : )
9554 1 : .await?;
9555 1 : writer.finish_write(Lsn(0x20));
9556 1 : drop(writer);
9557 1 : timeline.freeze_and_flush().await.unwrap();
9558 :
9559 1 : timeline.remote_client.wait_completion().await.unwrap();
9560 1 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
9561 1 : let remote_consistent_lsn = timeline.get_remote_consistent_lsn_projected();
9562 1 : assert_eq!(Some(disk_consistent_lsn), remote_consistent_lsn);
9563 :
9564 : //
9565 : // Test
9566 : //
9567 :
9568 1 : let mut writer = timeline.writer().await;
9569 1 : writer
9570 1 : .put(
9571 1 : *TEST_KEY,
9572 1 : Lsn(0x30),
9573 1 : &Value::Image(test_img("foo at 0x30")),
9574 1 : &ctx,
9575 1 : )
9576 1 : .await?;
9577 1 : writer.finish_write(Lsn(0x30));
9578 1 : drop(writer);
9579 :
9580 1 : fail::cfg(
9581 : "flush-layer-before-update-remote-consistent-lsn",
9582 1 : "return()",
9583 : )
9584 1 : .unwrap();
9585 :
9586 1 : let flush_res = timeline.freeze_and_flush().await;
9587 : // if flush failed, the disk/remote consistent LSN should not be updated
9588 1 : assert!(flush_res.is_err());
9589 1 : assert_eq!(disk_consistent_lsn, timeline.get_disk_consistent_lsn());
9590 1 : assert_eq!(
9591 : remote_consistent_lsn,
9592 1 : timeline.get_remote_consistent_lsn_projected()
9593 : );
9594 :
9595 2 : Ok(())
9596 1 : }
9597 :
9598 : #[cfg(feature = "testing")]
9599 : #[tokio::test]
9600 1 : async fn test_simple_bottom_most_compaction_deltas_1() -> anyhow::Result<()> {
9601 2 : test_simple_bottom_most_compaction_deltas_helper(
9602 2 : "test_simple_bottom_most_compaction_deltas_1",
9603 2 : false,
9604 2 : )
9605 2 : .await
9606 1 : }
9607 :
9608 : #[cfg(feature = "testing")]
9609 : #[tokio::test]
9610 1 : async fn test_simple_bottom_most_compaction_deltas_2() -> anyhow::Result<()> {
9611 2 : test_simple_bottom_most_compaction_deltas_helper(
9612 2 : "test_simple_bottom_most_compaction_deltas_2",
9613 2 : true,
9614 2 : )
9615 2 : .await
9616 1 : }
9617 :
9618 : #[cfg(feature = "testing")]
9619 2 : async fn test_simple_bottom_most_compaction_deltas_helper(
9620 2 : test_name: &'static str,
9621 2 : use_delta_bottom_layer: bool,
9622 2 : ) -> anyhow::Result<()> {
9623 2 : let harness = TenantHarness::create(test_name).await?;
9624 2 : let (tenant, ctx) = harness.load().await;
9625 :
9626 138 : fn get_key(id: u32) -> Key {
9627 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9628 138 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9629 138 : key.field6 = id;
9630 138 : key
9631 138 : }
9632 :
9633 : // We create
9634 : // - one bottom-most image layer,
9635 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
9636 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
9637 : // - a delta layer D3 above the horizon.
9638 : //
9639 : // | D3 |
9640 : // | D1 |
9641 : // -| |-- gc horizon -----------------
9642 : // | | | D2 |
9643 : // --------- img layer ------------------
9644 : //
9645 : // What we should expact from this compaction is:
9646 : // | D3 |
9647 : // | Part of D1 |
9648 : // --------- img layer with D1+D2 at GC horizon------------------
9649 :
9650 : // img layer at 0x10
9651 2 : let img_layer = (0..10)
9652 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9653 2 : .collect_vec();
9654 : // or, delta layer at 0x10 if `use_delta_bottom_layer` is true
9655 2 : let delta4 = (0..10)
9656 20 : .map(|id| {
9657 20 : (
9658 20 : get_key(id),
9659 20 : Lsn(0x08),
9660 20 : Value::WalRecord(NeonWalRecord::wal_init(format!("value {id}@0x10"))),
9661 20 : )
9662 20 : })
9663 2 : .collect_vec();
9664 :
9665 2 : let delta1 = vec![
9666 2 : (
9667 2 : get_key(1),
9668 2 : Lsn(0x20),
9669 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9670 2 : ),
9671 2 : (
9672 2 : get_key(2),
9673 2 : Lsn(0x30),
9674 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9675 2 : ),
9676 2 : (
9677 2 : get_key(3),
9678 2 : Lsn(0x28),
9679 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9680 2 : ),
9681 2 : (
9682 2 : get_key(3),
9683 2 : Lsn(0x30),
9684 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9685 2 : ),
9686 2 : (
9687 2 : get_key(3),
9688 2 : Lsn(0x40),
9689 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9690 2 : ),
9691 : ];
9692 2 : let delta2 = vec![
9693 2 : (
9694 2 : get_key(5),
9695 2 : Lsn(0x20),
9696 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9697 2 : ),
9698 2 : (
9699 2 : get_key(6),
9700 2 : Lsn(0x20),
9701 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9702 2 : ),
9703 : ];
9704 2 : let delta3 = vec![
9705 2 : (
9706 2 : get_key(8),
9707 2 : Lsn(0x48),
9708 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9709 2 : ),
9710 2 : (
9711 2 : get_key(9),
9712 2 : Lsn(0x48),
9713 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9714 2 : ),
9715 : ];
9716 :
9717 2 : let tline = if use_delta_bottom_layer {
9718 1 : tenant
9719 1 : .create_test_timeline_with_layers(
9720 1 : TIMELINE_ID,
9721 1 : Lsn(0x08),
9722 1 : DEFAULT_PG_VERSION,
9723 1 : &ctx,
9724 1 : Vec::new(), // in-memory layers
9725 1 : vec![
9726 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9727 1 : Lsn(0x08)..Lsn(0x10),
9728 1 : delta4,
9729 1 : ),
9730 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9731 1 : Lsn(0x20)..Lsn(0x48),
9732 1 : delta1,
9733 1 : ),
9734 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9735 1 : Lsn(0x20)..Lsn(0x48),
9736 1 : delta2,
9737 1 : ),
9738 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9739 1 : Lsn(0x48)..Lsn(0x50),
9740 1 : delta3,
9741 1 : ),
9742 1 : ], // delta layers
9743 1 : vec![], // image layers
9744 1 : Lsn(0x50),
9745 1 : )
9746 1 : .await?
9747 : } else {
9748 1 : tenant
9749 1 : .create_test_timeline_with_layers(
9750 1 : TIMELINE_ID,
9751 1 : Lsn(0x10),
9752 1 : DEFAULT_PG_VERSION,
9753 1 : &ctx,
9754 1 : Vec::new(), // in-memory layers
9755 1 : vec![
9756 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9757 1 : Lsn(0x10)..Lsn(0x48),
9758 1 : delta1,
9759 1 : ),
9760 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9761 1 : Lsn(0x10)..Lsn(0x48),
9762 1 : delta2,
9763 1 : ),
9764 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9765 1 : Lsn(0x48)..Lsn(0x50),
9766 1 : delta3,
9767 1 : ),
9768 1 : ], // delta layers
9769 1 : vec![(Lsn(0x10), img_layer)], // image layers
9770 1 : Lsn(0x50),
9771 1 : )
9772 1 : .await?
9773 : };
9774 : {
9775 2 : tline
9776 2 : .applied_gc_cutoff_lsn
9777 2 : .lock_for_write()
9778 2 : .store_and_unlock(Lsn(0x30))
9779 2 : .wait()
9780 2 : .await;
9781 : // Update GC info
9782 2 : let mut guard = tline.gc_info.write().unwrap();
9783 2 : *guard = GcInfo {
9784 2 : retain_lsns: vec![],
9785 2 : cutoffs: GcCutoffs {
9786 2 : time: Some(Lsn(0x30)),
9787 2 : space: Lsn(0x30),
9788 2 : },
9789 2 : leases: Default::default(),
9790 2 : within_ancestor_pitr: false,
9791 2 : };
9792 : }
9793 :
9794 2 : let expected_result = [
9795 2 : Bytes::from_static(b"value 0@0x10"),
9796 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9797 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9798 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9799 2 : Bytes::from_static(b"value 4@0x10"),
9800 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9801 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9802 2 : Bytes::from_static(b"value 7@0x10"),
9803 2 : Bytes::from_static(b"value 8@0x10@0x48"),
9804 2 : Bytes::from_static(b"value 9@0x10@0x48"),
9805 2 : ];
9806 :
9807 2 : let expected_result_at_gc_horizon = [
9808 2 : Bytes::from_static(b"value 0@0x10"),
9809 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9810 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9811 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
9812 2 : Bytes::from_static(b"value 4@0x10"),
9813 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9814 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9815 2 : Bytes::from_static(b"value 7@0x10"),
9816 2 : Bytes::from_static(b"value 8@0x10"),
9817 2 : Bytes::from_static(b"value 9@0x10"),
9818 2 : ];
9819 :
9820 22 : for idx in 0..10 {
9821 20 : assert_eq!(
9822 20 : tline
9823 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9824 20 : .await
9825 20 : .unwrap(),
9826 20 : &expected_result[idx]
9827 : );
9828 20 : assert_eq!(
9829 20 : tline
9830 20 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
9831 20 : .await
9832 20 : .unwrap(),
9833 20 : &expected_result_at_gc_horizon[idx]
9834 : );
9835 : }
9836 :
9837 2 : let cancel = CancellationToken::new();
9838 2 : tline
9839 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9840 2 : .await
9841 2 : .unwrap();
9842 :
9843 22 : for idx in 0..10 {
9844 20 : assert_eq!(
9845 20 : tline
9846 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9847 20 : .await
9848 20 : .unwrap(),
9849 20 : &expected_result[idx]
9850 : );
9851 20 : assert_eq!(
9852 20 : tline
9853 20 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
9854 20 : .await
9855 20 : .unwrap(),
9856 20 : &expected_result_at_gc_horizon[idx]
9857 : );
9858 : }
9859 :
9860 : // increase GC horizon and compact again
9861 : {
9862 2 : tline
9863 2 : .applied_gc_cutoff_lsn
9864 2 : .lock_for_write()
9865 2 : .store_and_unlock(Lsn(0x40))
9866 2 : .wait()
9867 2 : .await;
9868 : // Update GC info
9869 2 : let mut guard = tline.gc_info.write().unwrap();
9870 2 : guard.cutoffs.time = Some(Lsn(0x40));
9871 2 : guard.cutoffs.space = Lsn(0x40);
9872 : }
9873 2 : tline
9874 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9875 2 : .await
9876 2 : .unwrap();
9877 :
9878 2 : Ok(())
9879 2 : }
9880 :
9881 : #[cfg(feature = "testing")]
9882 : #[tokio::test]
9883 1 : async fn test_generate_key_retention() -> anyhow::Result<()> {
9884 1 : let harness = TenantHarness::create("test_generate_key_retention").await?;
9885 1 : let (tenant, ctx) = harness.load().await;
9886 1 : let tline = tenant
9887 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
9888 1 : .await?;
9889 1 : tline.force_advance_lsn(Lsn(0x70));
9890 1 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
9891 1 : let history = vec![
9892 1 : (
9893 1 : key,
9894 1 : Lsn(0x10),
9895 1 : Value::WalRecord(NeonWalRecord::wal_init("0x10")),
9896 1 : ),
9897 1 : (
9898 1 : key,
9899 1 : Lsn(0x20),
9900 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9901 1 : ),
9902 1 : (
9903 1 : key,
9904 1 : Lsn(0x30),
9905 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9906 1 : ),
9907 1 : (
9908 1 : key,
9909 1 : Lsn(0x40),
9910 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9911 1 : ),
9912 1 : (
9913 1 : key,
9914 1 : Lsn(0x50),
9915 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
9916 1 : ),
9917 1 : (
9918 1 : key,
9919 1 : Lsn(0x60),
9920 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9921 1 : ),
9922 1 : (
9923 1 : key,
9924 1 : Lsn(0x70),
9925 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9926 1 : ),
9927 1 : (
9928 1 : key,
9929 1 : Lsn(0x80),
9930 1 : Value::Image(Bytes::copy_from_slice(
9931 1 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9932 1 : )),
9933 1 : ),
9934 1 : (
9935 1 : key,
9936 1 : Lsn(0x90),
9937 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9938 1 : ),
9939 : ];
9940 1 : let res = tline
9941 1 : .generate_key_retention(
9942 1 : key,
9943 1 : &history,
9944 1 : Lsn(0x60),
9945 1 : &[Lsn(0x20), Lsn(0x40), Lsn(0x50)],
9946 1 : 3,
9947 1 : None,
9948 1 : true,
9949 1 : )
9950 1 : .await
9951 1 : .unwrap();
9952 1 : let expected_res = KeyHistoryRetention {
9953 1 : below_horizon: vec![
9954 1 : (
9955 1 : Lsn(0x20),
9956 1 : KeyLogAtLsn(vec![(
9957 1 : Lsn(0x20),
9958 1 : Value::Image(Bytes::from_static(b"0x10;0x20")),
9959 1 : )]),
9960 1 : ),
9961 1 : (
9962 1 : Lsn(0x40),
9963 1 : KeyLogAtLsn(vec![
9964 1 : (
9965 1 : Lsn(0x30),
9966 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9967 1 : ),
9968 1 : (
9969 1 : Lsn(0x40),
9970 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9971 1 : ),
9972 1 : ]),
9973 1 : ),
9974 1 : (
9975 1 : Lsn(0x50),
9976 1 : KeyLogAtLsn(vec![(
9977 1 : Lsn(0x50),
9978 1 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40;0x50")),
9979 1 : )]),
9980 1 : ),
9981 1 : (
9982 1 : Lsn(0x60),
9983 1 : KeyLogAtLsn(vec![(
9984 1 : Lsn(0x60),
9985 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9986 1 : )]),
9987 1 : ),
9988 1 : ],
9989 1 : above_horizon: KeyLogAtLsn(vec![
9990 1 : (
9991 1 : Lsn(0x70),
9992 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9993 1 : ),
9994 1 : (
9995 1 : Lsn(0x80),
9996 1 : Value::Image(Bytes::copy_from_slice(
9997 1 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9998 1 : )),
9999 1 : ),
10000 1 : (
10001 1 : Lsn(0x90),
10002 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
10003 1 : ),
10004 1 : ]),
10005 1 : };
10006 1 : assert_eq!(res, expected_res);
10007 :
10008 : // We expect GC-compaction to run with the original GC. This would create a situation that
10009 : // the original GC algorithm removes some delta layers b/c there are full image coverage,
10010 : // therefore causing some keys to have an incomplete history below the lowest retain LSN.
10011 : // For example, we have
10012 : // ```plain
10013 : // init delta @ 0x10, image @ 0x20, delta @ 0x30 (gc_horizon), image @ 0x40.
10014 : // ```
10015 : // Now the GC horizon moves up, and we have
10016 : // ```plain
10017 : // init delta @ 0x10, image @ 0x20, delta @ 0x30, image @ 0x40 (gc_horizon)
10018 : // ```
10019 : // The original GC algorithm kicks in, and removes delta @ 0x10, image @ 0x20.
10020 : // We will end up with
10021 : // ```plain
10022 : // delta @ 0x30, image @ 0x40 (gc_horizon)
10023 : // ```
10024 : // Now we run the GC-compaction, and this key does not have a full history.
10025 : // We should be able to handle this partial history and drop everything before the
10026 : // gc_horizon image.
10027 :
10028 1 : let history = vec![
10029 1 : (
10030 1 : key,
10031 1 : Lsn(0x20),
10032 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
10033 1 : ),
10034 1 : (
10035 1 : key,
10036 1 : Lsn(0x30),
10037 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
10038 1 : ),
10039 1 : (
10040 1 : key,
10041 1 : Lsn(0x40),
10042 1 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
10043 1 : ),
10044 1 : (
10045 1 : key,
10046 1 : Lsn(0x50),
10047 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
10048 1 : ),
10049 1 : (
10050 1 : key,
10051 1 : Lsn(0x60),
10052 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
10053 1 : ),
10054 1 : (
10055 1 : key,
10056 1 : Lsn(0x70),
10057 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
10058 1 : ),
10059 1 : (
10060 1 : key,
10061 1 : Lsn(0x80),
10062 1 : Value::Image(Bytes::copy_from_slice(
10063 1 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
10064 1 : )),
10065 1 : ),
10066 1 : (
10067 1 : key,
10068 1 : Lsn(0x90),
10069 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
10070 1 : ),
10071 : ];
10072 1 : let res = tline
10073 1 : .generate_key_retention(
10074 1 : key,
10075 1 : &history,
10076 1 : Lsn(0x60),
10077 1 : &[Lsn(0x40), Lsn(0x50)],
10078 1 : 3,
10079 1 : None,
10080 1 : true,
10081 1 : )
10082 1 : .await
10083 1 : .unwrap();
10084 1 : let expected_res = KeyHistoryRetention {
10085 1 : below_horizon: vec![
10086 1 : (
10087 1 : Lsn(0x40),
10088 1 : KeyLogAtLsn(vec![(
10089 1 : Lsn(0x40),
10090 1 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
10091 1 : )]),
10092 1 : ),
10093 1 : (
10094 1 : Lsn(0x50),
10095 1 : KeyLogAtLsn(vec![(
10096 1 : Lsn(0x50),
10097 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
10098 1 : )]),
10099 1 : ),
10100 1 : (
10101 1 : Lsn(0x60),
10102 1 : KeyLogAtLsn(vec![(
10103 1 : Lsn(0x60),
10104 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
10105 1 : )]),
10106 1 : ),
10107 1 : ],
10108 1 : above_horizon: KeyLogAtLsn(vec![
10109 1 : (
10110 1 : Lsn(0x70),
10111 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
10112 1 : ),
10113 1 : (
10114 1 : Lsn(0x80),
10115 1 : Value::Image(Bytes::copy_from_slice(
10116 1 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
10117 1 : )),
10118 1 : ),
10119 1 : (
10120 1 : Lsn(0x90),
10121 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
10122 1 : ),
10123 1 : ]),
10124 1 : };
10125 1 : assert_eq!(res, expected_res);
10126 :
10127 : // In case of branch compaction, the branch itself does not have the full history, and we need to provide
10128 : // the ancestor image in the test case.
10129 :
10130 1 : let history = vec![
10131 1 : (
10132 1 : key,
10133 1 : Lsn(0x20),
10134 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
10135 1 : ),
10136 1 : (
10137 1 : key,
10138 1 : Lsn(0x30),
10139 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
10140 1 : ),
10141 1 : (
10142 1 : key,
10143 1 : Lsn(0x40),
10144 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
10145 1 : ),
10146 1 : (
10147 1 : key,
10148 1 : Lsn(0x70),
10149 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
10150 1 : ),
10151 : ];
10152 1 : let res = tline
10153 1 : .generate_key_retention(
10154 1 : key,
10155 1 : &history,
10156 1 : Lsn(0x60),
10157 1 : &[],
10158 1 : 3,
10159 1 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
10160 1 : true,
10161 1 : )
10162 1 : .await
10163 1 : .unwrap();
10164 1 : let expected_res = KeyHistoryRetention {
10165 1 : below_horizon: vec![(
10166 1 : Lsn(0x60),
10167 1 : KeyLogAtLsn(vec![(
10168 1 : Lsn(0x60),
10169 1 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")), // use the ancestor image to reconstruct the page
10170 1 : )]),
10171 1 : )],
10172 1 : above_horizon: KeyLogAtLsn(vec![(
10173 1 : Lsn(0x70),
10174 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
10175 1 : )]),
10176 1 : };
10177 1 : assert_eq!(res, expected_res);
10178 :
10179 1 : let history = vec![
10180 1 : (
10181 1 : key,
10182 1 : Lsn(0x20),
10183 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
10184 1 : ),
10185 1 : (
10186 1 : key,
10187 1 : Lsn(0x40),
10188 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
10189 1 : ),
10190 1 : (
10191 1 : key,
10192 1 : Lsn(0x60),
10193 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
10194 1 : ),
10195 1 : (
10196 1 : key,
10197 1 : Lsn(0x70),
10198 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
10199 1 : ),
10200 : ];
10201 1 : let res = tline
10202 1 : .generate_key_retention(
10203 1 : key,
10204 1 : &history,
10205 1 : Lsn(0x60),
10206 1 : &[Lsn(0x30)],
10207 1 : 3,
10208 1 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
10209 1 : true,
10210 1 : )
10211 1 : .await
10212 1 : .unwrap();
10213 1 : let expected_res = KeyHistoryRetention {
10214 1 : below_horizon: vec![
10215 1 : (
10216 1 : Lsn(0x30),
10217 1 : KeyLogAtLsn(vec![(
10218 1 : Lsn(0x20),
10219 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
10220 1 : )]),
10221 1 : ),
10222 1 : (
10223 1 : Lsn(0x60),
10224 1 : KeyLogAtLsn(vec![(
10225 1 : Lsn(0x60),
10226 1 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x40;0x60")),
10227 1 : )]),
10228 1 : ),
10229 1 : ],
10230 1 : above_horizon: KeyLogAtLsn(vec![(
10231 1 : Lsn(0x70),
10232 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
10233 1 : )]),
10234 1 : };
10235 1 : assert_eq!(res, expected_res);
10236 :
10237 2 : Ok(())
10238 1 : }
10239 :
10240 : #[cfg(feature = "testing")]
10241 : #[tokio::test]
10242 1 : async fn test_simple_bottom_most_compaction_with_retain_lsns() -> anyhow::Result<()> {
10243 1 : let harness =
10244 1 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns").await?;
10245 1 : let (tenant, ctx) = harness.load().await;
10246 :
10247 259 : fn get_key(id: u32) -> Key {
10248 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
10249 259 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
10250 259 : key.field6 = id;
10251 259 : key
10252 259 : }
10253 :
10254 1 : let img_layer = (0..10)
10255 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10256 1 : .collect_vec();
10257 :
10258 1 : let delta1 = vec![
10259 1 : (
10260 1 : get_key(1),
10261 1 : Lsn(0x20),
10262 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10263 1 : ),
10264 1 : (
10265 1 : get_key(2),
10266 1 : Lsn(0x30),
10267 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10268 1 : ),
10269 1 : (
10270 1 : get_key(3),
10271 1 : Lsn(0x28),
10272 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
10273 1 : ),
10274 1 : (
10275 1 : get_key(3),
10276 1 : Lsn(0x30),
10277 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10278 1 : ),
10279 1 : (
10280 1 : get_key(3),
10281 1 : Lsn(0x40),
10282 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
10283 1 : ),
10284 : ];
10285 1 : let delta2 = vec![
10286 1 : (
10287 1 : get_key(5),
10288 1 : Lsn(0x20),
10289 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10290 1 : ),
10291 1 : (
10292 1 : get_key(6),
10293 1 : Lsn(0x20),
10294 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10295 1 : ),
10296 : ];
10297 1 : let delta3 = vec![
10298 1 : (
10299 1 : get_key(8),
10300 1 : Lsn(0x48),
10301 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10302 1 : ),
10303 1 : (
10304 1 : get_key(9),
10305 1 : Lsn(0x48),
10306 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10307 1 : ),
10308 : ];
10309 :
10310 1 : let tline = tenant
10311 1 : .create_test_timeline_with_layers(
10312 1 : TIMELINE_ID,
10313 1 : Lsn(0x10),
10314 1 : DEFAULT_PG_VERSION,
10315 1 : &ctx,
10316 1 : Vec::new(), // in-memory layers
10317 1 : vec![
10318 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
10319 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
10320 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
10321 1 : ], // delta layers
10322 1 : vec![(Lsn(0x10), img_layer)], // image layers
10323 1 : Lsn(0x50),
10324 1 : )
10325 1 : .await?;
10326 : {
10327 1 : tline
10328 1 : .applied_gc_cutoff_lsn
10329 1 : .lock_for_write()
10330 1 : .store_and_unlock(Lsn(0x30))
10331 1 : .wait()
10332 1 : .await;
10333 : // Update GC info
10334 1 : let mut guard = tline.gc_info.write().unwrap();
10335 1 : *guard = GcInfo {
10336 1 : retain_lsns: vec![
10337 1 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
10338 1 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
10339 1 : ],
10340 1 : cutoffs: GcCutoffs {
10341 1 : time: Some(Lsn(0x30)),
10342 1 : space: Lsn(0x30),
10343 1 : },
10344 1 : leases: Default::default(),
10345 1 : within_ancestor_pitr: false,
10346 1 : };
10347 : }
10348 :
10349 1 : let expected_result = [
10350 1 : Bytes::from_static(b"value 0@0x10"),
10351 1 : Bytes::from_static(b"value 1@0x10@0x20"),
10352 1 : Bytes::from_static(b"value 2@0x10@0x30"),
10353 1 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
10354 1 : Bytes::from_static(b"value 4@0x10"),
10355 1 : Bytes::from_static(b"value 5@0x10@0x20"),
10356 1 : Bytes::from_static(b"value 6@0x10@0x20"),
10357 1 : Bytes::from_static(b"value 7@0x10"),
10358 1 : Bytes::from_static(b"value 8@0x10@0x48"),
10359 1 : Bytes::from_static(b"value 9@0x10@0x48"),
10360 1 : ];
10361 :
10362 1 : let expected_result_at_gc_horizon = [
10363 1 : Bytes::from_static(b"value 0@0x10"),
10364 1 : Bytes::from_static(b"value 1@0x10@0x20"),
10365 1 : Bytes::from_static(b"value 2@0x10@0x30"),
10366 1 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
10367 1 : Bytes::from_static(b"value 4@0x10"),
10368 1 : Bytes::from_static(b"value 5@0x10@0x20"),
10369 1 : Bytes::from_static(b"value 6@0x10@0x20"),
10370 1 : Bytes::from_static(b"value 7@0x10"),
10371 1 : Bytes::from_static(b"value 8@0x10"),
10372 1 : Bytes::from_static(b"value 9@0x10"),
10373 1 : ];
10374 :
10375 1 : let expected_result_at_lsn_20 = [
10376 1 : Bytes::from_static(b"value 0@0x10"),
10377 1 : Bytes::from_static(b"value 1@0x10@0x20"),
10378 1 : Bytes::from_static(b"value 2@0x10"),
10379 1 : Bytes::from_static(b"value 3@0x10"),
10380 1 : Bytes::from_static(b"value 4@0x10"),
10381 1 : Bytes::from_static(b"value 5@0x10@0x20"),
10382 1 : Bytes::from_static(b"value 6@0x10@0x20"),
10383 1 : Bytes::from_static(b"value 7@0x10"),
10384 1 : Bytes::from_static(b"value 8@0x10"),
10385 1 : Bytes::from_static(b"value 9@0x10"),
10386 1 : ];
10387 :
10388 1 : let expected_result_at_lsn_10 = [
10389 1 : Bytes::from_static(b"value 0@0x10"),
10390 1 : Bytes::from_static(b"value 1@0x10"),
10391 1 : Bytes::from_static(b"value 2@0x10"),
10392 1 : Bytes::from_static(b"value 3@0x10"),
10393 1 : Bytes::from_static(b"value 4@0x10"),
10394 1 : Bytes::from_static(b"value 5@0x10"),
10395 1 : Bytes::from_static(b"value 6@0x10"),
10396 1 : Bytes::from_static(b"value 7@0x10"),
10397 1 : Bytes::from_static(b"value 8@0x10"),
10398 1 : Bytes::from_static(b"value 9@0x10"),
10399 1 : ];
10400 :
10401 6 : let verify_result = || async {
10402 6 : let gc_horizon = {
10403 6 : let gc_info = tline.gc_info.read().unwrap();
10404 6 : gc_info.cutoffs.time.unwrap_or_default()
10405 : };
10406 66 : for idx in 0..10 {
10407 60 : assert_eq!(
10408 60 : tline
10409 60 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10410 60 : .await
10411 60 : .unwrap(),
10412 60 : &expected_result[idx]
10413 : );
10414 60 : assert_eq!(
10415 60 : tline
10416 60 : .get(get_key(idx as u32), gc_horizon, &ctx)
10417 60 : .await
10418 60 : .unwrap(),
10419 60 : &expected_result_at_gc_horizon[idx]
10420 : );
10421 60 : assert_eq!(
10422 60 : tline
10423 60 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
10424 60 : .await
10425 60 : .unwrap(),
10426 60 : &expected_result_at_lsn_20[idx]
10427 : );
10428 60 : assert_eq!(
10429 60 : tline
10430 60 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
10431 60 : .await
10432 60 : .unwrap(),
10433 60 : &expected_result_at_lsn_10[idx]
10434 : );
10435 : }
10436 12 : };
10437 :
10438 1 : verify_result().await;
10439 :
10440 1 : let cancel = CancellationToken::new();
10441 1 : let mut dryrun_flags = EnumSet::new();
10442 1 : dryrun_flags.insert(CompactFlags::DryRun);
10443 :
10444 1 : tline
10445 1 : .compact_with_gc(
10446 1 : &cancel,
10447 1 : CompactOptions {
10448 1 : flags: dryrun_flags,
10449 1 : ..Default::default()
10450 1 : },
10451 1 : &ctx,
10452 1 : )
10453 1 : .await
10454 1 : .unwrap();
10455 : // 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
10456 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
10457 1 : verify_result().await;
10458 :
10459 1 : tline
10460 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10461 1 : .await
10462 1 : .unwrap();
10463 1 : verify_result().await;
10464 :
10465 : // compact again
10466 1 : tline
10467 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10468 1 : .await
10469 1 : .unwrap();
10470 1 : verify_result().await;
10471 :
10472 : // increase GC horizon and compact again
10473 : {
10474 1 : tline
10475 1 : .applied_gc_cutoff_lsn
10476 1 : .lock_for_write()
10477 1 : .store_and_unlock(Lsn(0x38))
10478 1 : .wait()
10479 1 : .await;
10480 : // Update GC info
10481 1 : let mut guard = tline.gc_info.write().unwrap();
10482 1 : guard.cutoffs.time = Some(Lsn(0x38));
10483 1 : guard.cutoffs.space = Lsn(0x38);
10484 : }
10485 1 : tline
10486 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10487 1 : .await
10488 1 : .unwrap();
10489 1 : verify_result().await; // no wals between 0x30 and 0x38, so we should obtain the same result
10490 :
10491 : // not increasing the GC horizon and compact again
10492 1 : tline
10493 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10494 1 : .await
10495 1 : .unwrap();
10496 1 : verify_result().await;
10497 :
10498 2 : Ok(())
10499 1 : }
10500 :
10501 : #[cfg(feature = "testing")]
10502 : #[tokio::test]
10503 1 : async fn test_simple_bottom_most_compaction_with_retain_lsns_single_key() -> anyhow::Result<()>
10504 1 : {
10505 1 : let harness =
10506 1 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns_single_key")
10507 1 : .await?;
10508 1 : let (tenant, ctx) = harness.load().await;
10509 :
10510 176 : fn get_key(id: u32) -> Key {
10511 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
10512 176 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
10513 176 : key.field6 = id;
10514 176 : key
10515 176 : }
10516 :
10517 1 : let img_layer = (0..10)
10518 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10519 1 : .collect_vec();
10520 :
10521 1 : let delta1 = vec![
10522 1 : (
10523 1 : get_key(1),
10524 1 : Lsn(0x20),
10525 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10526 1 : ),
10527 1 : (
10528 1 : get_key(1),
10529 1 : Lsn(0x28),
10530 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
10531 1 : ),
10532 : ];
10533 1 : let delta2 = vec![
10534 1 : (
10535 1 : get_key(1),
10536 1 : Lsn(0x30),
10537 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10538 1 : ),
10539 1 : (
10540 1 : get_key(1),
10541 1 : Lsn(0x38),
10542 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
10543 1 : ),
10544 : ];
10545 1 : let delta3 = vec![
10546 1 : (
10547 1 : get_key(8),
10548 1 : Lsn(0x48),
10549 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10550 1 : ),
10551 1 : (
10552 1 : get_key(9),
10553 1 : Lsn(0x48),
10554 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10555 1 : ),
10556 : ];
10557 :
10558 1 : let tline = tenant
10559 1 : .create_test_timeline_with_layers(
10560 1 : TIMELINE_ID,
10561 1 : Lsn(0x10),
10562 1 : DEFAULT_PG_VERSION,
10563 1 : &ctx,
10564 1 : Vec::new(), // in-memory layers
10565 1 : vec![
10566 1 : // delta1 and delta 2 only contain a single key but multiple updates
10567 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x30), delta1),
10568 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
10569 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x50), delta3),
10570 1 : ], // delta layers
10571 1 : vec![(Lsn(0x10), img_layer)], // image layers
10572 1 : Lsn(0x50),
10573 1 : )
10574 1 : .await?;
10575 : {
10576 1 : tline
10577 1 : .applied_gc_cutoff_lsn
10578 1 : .lock_for_write()
10579 1 : .store_and_unlock(Lsn(0x30))
10580 1 : .wait()
10581 1 : .await;
10582 : // Update GC info
10583 1 : let mut guard = tline.gc_info.write().unwrap();
10584 1 : *guard = GcInfo {
10585 1 : retain_lsns: vec![
10586 1 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
10587 1 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
10588 1 : ],
10589 1 : cutoffs: GcCutoffs {
10590 1 : time: Some(Lsn(0x30)),
10591 1 : space: Lsn(0x30),
10592 1 : },
10593 1 : leases: Default::default(),
10594 1 : within_ancestor_pitr: false,
10595 1 : };
10596 : }
10597 :
10598 1 : let expected_result = [
10599 1 : Bytes::from_static(b"value 0@0x10"),
10600 1 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
10601 1 : Bytes::from_static(b"value 2@0x10"),
10602 1 : Bytes::from_static(b"value 3@0x10"),
10603 1 : Bytes::from_static(b"value 4@0x10"),
10604 1 : Bytes::from_static(b"value 5@0x10"),
10605 1 : Bytes::from_static(b"value 6@0x10"),
10606 1 : Bytes::from_static(b"value 7@0x10"),
10607 1 : Bytes::from_static(b"value 8@0x10@0x48"),
10608 1 : Bytes::from_static(b"value 9@0x10@0x48"),
10609 1 : ];
10610 :
10611 1 : let expected_result_at_gc_horizon = [
10612 1 : Bytes::from_static(b"value 0@0x10"),
10613 1 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
10614 1 : Bytes::from_static(b"value 2@0x10"),
10615 1 : Bytes::from_static(b"value 3@0x10"),
10616 1 : Bytes::from_static(b"value 4@0x10"),
10617 1 : Bytes::from_static(b"value 5@0x10"),
10618 1 : Bytes::from_static(b"value 6@0x10"),
10619 1 : Bytes::from_static(b"value 7@0x10"),
10620 1 : Bytes::from_static(b"value 8@0x10"),
10621 1 : Bytes::from_static(b"value 9@0x10"),
10622 1 : ];
10623 :
10624 1 : let expected_result_at_lsn_20 = [
10625 1 : Bytes::from_static(b"value 0@0x10"),
10626 1 : Bytes::from_static(b"value 1@0x10@0x20"),
10627 1 : Bytes::from_static(b"value 2@0x10"),
10628 1 : Bytes::from_static(b"value 3@0x10"),
10629 1 : Bytes::from_static(b"value 4@0x10"),
10630 1 : Bytes::from_static(b"value 5@0x10"),
10631 1 : Bytes::from_static(b"value 6@0x10"),
10632 1 : Bytes::from_static(b"value 7@0x10"),
10633 1 : Bytes::from_static(b"value 8@0x10"),
10634 1 : Bytes::from_static(b"value 9@0x10"),
10635 1 : ];
10636 :
10637 1 : let expected_result_at_lsn_10 = [
10638 1 : Bytes::from_static(b"value 0@0x10"),
10639 1 : Bytes::from_static(b"value 1@0x10"),
10640 1 : Bytes::from_static(b"value 2@0x10"),
10641 1 : Bytes::from_static(b"value 3@0x10"),
10642 1 : Bytes::from_static(b"value 4@0x10"),
10643 1 : Bytes::from_static(b"value 5@0x10"),
10644 1 : Bytes::from_static(b"value 6@0x10"),
10645 1 : Bytes::from_static(b"value 7@0x10"),
10646 1 : Bytes::from_static(b"value 8@0x10"),
10647 1 : Bytes::from_static(b"value 9@0x10"),
10648 1 : ];
10649 :
10650 4 : let verify_result = || async {
10651 4 : let gc_horizon = {
10652 4 : let gc_info = tline.gc_info.read().unwrap();
10653 4 : gc_info.cutoffs.time.unwrap_or_default()
10654 : };
10655 44 : for idx in 0..10 {
10656 40 : assert_eq!(
10657 40 : tline
10658 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10659 40 : .await
10660 40 : .unwrap(),
10661 40 : &expected_result[idx]
10662 : );
10663 40 : assert_eq!(
10664 40 : tline
10665 40 : .get(get_key(idx as u32), gc_horizon, &ctx)
10666 40 : .await
10667 40 : .unwrap(),
10668 40 : &expected_result_at_gc_horizon[idx]
10669 : );
10670 40 : assert_eq!(
10671 40 : tline
10672 40 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
10673 40 : .await
10674 40 : .unwrap(),
10675 40 : &expected_result_at_lsn_20[idx]
10676 : );
10677 40 : assert_eq!(
10678 40 : tline
10679 40 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
10680 40 : .await
10681 40 : .unwrap(),
10682 40 : &expected_result_at_lsn_10[idx]
10683 : );
10684 : }
10685 8 : };
10686 :
10687 1 : verify_result().await;
10688 :
10689 1 : let cancel = CancellationToken::new();
10690 1 : let mut dryrun_flags = EnumSet::new();
10691 1 : dryrun_flags.insert(CompactFlags::DryRun);
10692 :
10693 1 : tline
10694 1 : .compact_with_gc(
10695 1 : &cancel,
10696 1 : CompactOptions {
10697 1 : flags: dryrun_flags,
10698 1 : ..Default::default()
10699 1 : },
10700 1 : &ctx,
10701 1 : )
10702 1 : .await
10703 1 : .unwrap();
10704 : // 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
10705 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
10706 1 : verify_result().await;
10707 :
10708 1 : tline
10709 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10710 1 : .await
10711 1 : .unwrap();
10712 1 : verify_result().await;
10713 :
10714 : // compact again
10715 1 : tline
10716 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10717 1 : .await
10718 1 : .unwrap();
10719 1 : verify_result().await;
10720 :
10721 2 : Ok(())
10722 1 : }
10723 :
10724 : #[cfg(feature = "testing")]
10725 : #[tokio::test]
10726 1 : async fn test_simple_bottom_most_compaction_on_branch() -> anyhow::Result<()> {
10727 : use models::CompactLsnRange;
10728 :
10729 1 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_on_branch").await?;
10730 1 : let (tenant, ctx) = harness.load().await;
10731 :
10732 83 : fn get_key(id: u32) -> Key {
10733 83 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
10734 83 : key.field6 = id;
10735 83 : key
10736 83 : }
10737 :
10738 1 : let img_layer = (0..10)
10739 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10740 1 : .collect_vec();
10741 :
10742 1 : let delta1 = vec![
10743 1 : (
10744 1 : get_key(1),
10745 1 : Lsn(0x20),
10746 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10747 1 : ),
10748 1 : (
10749 1 : get_key(2),
10750 1 : Lsn(0x30),
10751 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10752 1 : ),
10753 1 : (
10754 1 : get_key(3),
10755 1 : Lsn(0x28),
10756 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
10757 1 : ),
10758 1 : (
10759 1 : get_key(3),
10760 1 : Lsn(0x30),
10761 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10762 1 : ),
10763 1 : (
10764 1 : get_key(3),
10765 1 : Lsn(0x40),
10766 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
10767 1 : ),
10768 : ];
10769 1 : let delta2 = vec![
10770 1 : (
10771 1 : get_key(5),
10772 1 : Lsn(0x20),
10773 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10774 1 : ),
10775 1 : (
10776 1 : get_key(6),
10777 1 : Lsn(0x20),
10778 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10779 1 : ),
10780 : ];
10781 1 : let delta3 = vec![
10782 1 : (
10783 1 : get_key(8),
10784 1 : Lsn(0x48),
10785 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10786 1 : ),
10787 1 : (
10788 1 : get_key(9),
10789 1 : Lsn(0x48),
10790 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10791 1 : ),
10792 : ];
10793 :
10794 1 : let parent_tline = tenant
10795 1 : .create_test_timeline_with_layers(
10796 1 : TIMELINE_ID,
10797 1 : Lsn(0x10),
10798 1 : DEFAULT_PG_VERSION,
10799 1 : &ctx,
10800 1 : vec![], // in-memory layers
10801 1 : vec![], // delta layers
10802 1 : vec![(Lsn(0x18), img_layer)], // image layers
10803 1 : Lsn(0x18),
10804 1 : )
10805 1 : .await?;
10806 :
10807 1 : parent_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
10808 :
10809 1 : let branch_tline = tenant
10810 1 : .branch_timeline_test_with_layers(
10811 1 : &parent_tline,
10812 1 : NEW_TIMELINE_ID,
10813 1 : Some(Lsn(0x18)),
10814 1 : &ctx,
10815 1 : vec![
10816 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
10817 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
10818 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
10819 1 : ], // delta layers
10820 1 : vec![], // image layers
10821 1 : Lsn(0x50),
10822 1 : )
10823 1 : .await?;
10824 :
10825 1 : branch_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
10826 :
10827 : {
10828 1 : parent_tline
10829 1 : .applied_gc_cutoff_lsn
10830 1 : .lock_for_write()
10831 1 : .store_and_unlock(Lsn(0x10))
10832 1 : .wait()
10833 1 : .await;
10834 : // Update GC info
10835 1 : let mut guard = parent_tline.gc_info.write().unwrap();
10836 1 : *guard = GcInfo {
10837 1 : retain_lsns: vec![(Lsn(0x18), branch_tline.timeline_id, MaybeOffloaded::No)],
10838 1 : cutoffs: GcCutoffs {
10839 1 : time: Some(Lsn(0x10)),
10840 1 : space: Lsn(0x10),
10841 1 : },
10842 1 : leases: Default::default(),
10843 1 : within_ancestor_pitr: false,
10844 1 : };
10845 : }
10846 :
10847 : {
10848 1 : branch_tline
10849 1 : .applied_gc_cutoff_lsn
10850 1 : .lock_for_write()
10851 1 : .store_and_unlock(Lsn(0x50))
10852 1 : .wait()
10853 1 : .await;
10854 : // Update GC info
10855 1 : let mut guard = branch_tline.gc_info.write().unwrap();
10856 1 : *guard = GcInfo {
10857 1 : retain_lsns: vec![(Lsn(0x40), branch_tline.timeline_id, MaybeOffloaded::No)],
10858 1 : cutoffs: GcCutoffs {
10859 1 : time: Some(Lsn(0x50)),
10860 1 : space: Lsn(0x50),
10861 1 : },
10862 1 : leases: Default::default(),
10863 1 : within_ancestor_pitr: false,
10864 1 : };
10865 : }
10866 :
10867 1 : let expected_result_at_gc_horizon = [
10868 1 : Bytes::from_static(b"value 0@0x10"),
10869 1 : Bytes::from_static(b"value 1@0x10@0x20"),
10870 1 : Bytes::from_static(b"value 2@0x10@0x30"),
10871 1 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
10872 1 : Bytes::from_static(b"value 4@0x10"),
10873 1 : Bytes::from_static(b"value 5@0x10@0x20"),
10874 1 : Bytes::from_static(b"value 6@0x10@0x20"),
10875 1 : Bytes::from_static(b"value 7@0x10"),
10876 1 : Bytes::from_static(b"value 8@0x10@0x48"),
10877 1 : Bytes::from_static(b"value 9@0x10@0x48"),
10878 1 : ];
10879 :
10880 1 : let expected_result_at_lsn_40 = [
10881 1 : Bytes::from_static(b"value 0@0x10"),
10882 1 : Bytes::from_static(b"value 1@0x10@0x20"),
10883 1 : Bytes::from_static(b"value 2@0x10@0x30"),
10884 1 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
10885 1 : Bytes::from_static(b"value 4@0x10"),
10886 1 : Bytes::from_static(b"value 5@0x10@0x20"),
10887 1 : Bytes::from_static(b"value 6@0x10@0x20"),
10888 1 : Bytes::from_static(b"value 7@0x10"),
10889 1 : Bytes::from_static(b"value 8@0x10"),
10890 1 : Bytes::from_static(b"value 9@0x10"),
10891 1 : ];
10892 :
10893 3 : let verify_result = || async {
10894 33 : for idx in 0..10 {
10895 30 : assert_eq!(
10896 30 : branch_tline
10897 30 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10898 30 : .await
10899 30 : .unwrap(),
10900 30 : &expected_result_at_gc_horizon[idx]
10901 : );
10902 30 : assert_eq!(
10903 30 : branch_tline
10904 30 : .get(get_key(idx as u32), Lsn(0x40), &ctx)
10905 30 : .await
10906 30 : .unwrap(),
10907 30 : &expected_result_at_lsn_40[idx]
10908 : );
10909 : }
10910 6 : };
10911 :
10912 1 : verify_result().await;
10913 :
10914 1 : let cancel = CancellationToken::new();
10915 1 : branch_tline
10916 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10917 1 : .await
10918 1 : .unwrap();
10919 :
10920 1 : verify_result().await;
10921 :
10922 : // Piggyback a compaction with above_lsn. Ensure it works correctly when the specified LSN intersects with the layer files.
10923 : // Now we already have a single large delta layer, so the compaction min_layer_lsn should be the same as ancestor LSN (0x18).
10924 1 : branch_tline
10925 1 : .compact_with_gc(
10926 1 : &cancel,
10927 1 : CompactOptions {
10928 1 : compact_lsn_range: Some(CompactLsnRange::above(Lsn(0x40))),
10929 1 : ..Default::default()
10930 1 : },
10931 1 : &ctx,
10932 1 : )
10933 1 : .await
10934 1 : .unwrap();
10935 :
10936 1 : verify_result().await;
10937 :
10938 2 : Ok(())
10939 1 : }
10940 :
10941 : // Regression test for https://github.com/neondatabase/neon/issues/9012
10942 : // Create an image arrangement where we have to read at different LSN ranges
10943 : // from a delta layer. This is achieved by overlapping an image layer on top of
10944 : // a delta layer. Like so:
10945 : //
10946 : // A B
10947 : // +----------------+ -> delta_layer
10948 : // | | ^ lsn
10949 : // | =========|-> nested_image_layer |
10950 : // | C | |
10951 : // +----------------+ |
10952 : // ======== -> baseline_image_layer +-------> key
10953 : //
10954 : //
10955 : // When querying the key range [A, B) we need to read at different LSN ranges
10956 : // for [A, C) and [C, B). This test checks that the described edge case is handled correctly.
10957 : #[cfg(feature = "testing")]
10958 : #[tokio::test]
10959 1 : async fn test_vectored_read_with_nested_image_layer() -> anyhow::Result<()> {
10960 1 : let harness = TenantHarness::create("test_vectored_read_with_nested_image_layer").await?;
10961 1 : let (tenant, ctx) = harness.load().await;
10962 :
10963 1 : let will_init_keys = [2, 6];
10964 22 : fn get_key(id: u32) -> Key {
10965 22 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
10966 22 : key.field6 = id;
10967 22 : key
10968 22 : }
10969 :
10970 1 : let mut expected_key_values = HashMap::new();
10971 :
10972 1 : let baseline_image_layer_lsn = Lsn(0x10);
10973 1 : let mut baseline_img_layer = Vec::new();
10974 6 : for i in 0..5 {
10975 5 : let key = get_key(i);
10976 5 : let value = format!("value {i}@{baseline_image_layer_lsn}");
10977 :
10978 5 : let removed = expected_key_values.insert(key, value.clone());
10979 5 : assert!(removed.is_none());
10980 :
10981 5 : baseline_img_layer.push((key, Bytes::from(value)));
10982 : }
10983 :
10984 1 : let nested_image_layer_lsn = Lsn(0x50);
10985 1 : let mut nested_img_layer = Vec::new();
10986 6 : for i in 5..10 {
10987 5 : let key = get_key(i);
10988 5 : let value = format!("value {i}@{nested_image_layer_lsn}");
10989 :
10990 5 : let removed = expected_key_values.insert(key, value.clone());
10991 5 : assert!(removed.is_none());
10992 :
10993 5 : nested_img_layer.push((key, Bytes::from(value)));
10994 : }
10995 :
10996 1 : let mut delta_layer_spec = Vec::default();
10997 1 : let delta_layer_start_lsn = Lsn(0x20);
10998 1 : let mut delta_layer_end_lsn = delta_layer_start_lsn;
10999 :
11000 11 : for i in 0..10 {
11001 10 : let key = get_key(i);
11002 10 : let key_in_nested = nested_img_layer
11003 10 : .iter()
11004 40 : .any(|(key_with_img, _)| *key_with_img == key);
11005 10 : let lsn = {
11006 10 : if key_in_nested {
11007 5 : Lsn(nested_image_layer_lsn.0 + 0x10)
11008 : } else {
11009 5 : delta_layer_start_lsn
11010 : }
11011 : };
11012 :
11013 10 : let will_init = will_init_keys.contains(&i);
11014 10 : if will_init {
11015 2 : delta_layer_spec.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init(""))));
11016 2 :
11017 2 : expected_key_values.insert(key, "".to_string());
11018 8 : } else {
11019 8 : let delta = format!("@{lsn}");
11020 8 : delta_layer_spec.push((
11021 8 : key,
11022 8 : lsn,
11023 8 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
11024 8 : ));
11025 8 :
11026 8 : expected_key_values
11027 8 : .get_mut(&key)
11028 8 : .expect("An image exists for each key")
11029 8 : .push_str(delta.as_str());
11030 8 : }
11031 10 : delta_layer_end_lsn = std::cmp::max(delta_layer_start_lsn, lsn);
11032 : }
11033 :
11034 1 : delta_layer_end_lsn = Lsn(delta_layer_end_lsn.0 + 1);
11035 :
11036 1 : assert!(
11037 1 : nested_image_layer_lsn > delta_layer_start_lsn
11038 1 : && nested_image_layer_lsn < delta_layer_end_lsn
11039 : );
11040 :
11041 1 : let tline = tenant
11042 1 : .create_test_timeline_with_layers(
11043 1 : TIMELINE_ID,
11044 1 : baseline_image_layer_lsn,
11045 1 : DEFAULT_PG_VERSION,
11046 1 : &ctx,
11047 1 : vec![], // in-memory layers
11048 1 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
11049 1 : delta_layer_start_lsn..delta_layer_end_lsn,
11050 1 : delta_layer_spec,
11051 1 : )], // delta layers
11052 1 : vec![
11053 1 : (baseline_image_layer_lsn, baseline_img_layer),
11054 1 : (nested_image_layer_lsn, nested_img_layer),
11055 1 : ], // image layers
11056 1 : delta_layer_end_lsn,
11057 1 : )
11058 1 : .await?;
11059 :
11060 1 : let query = VersionedKeySpaceQuery::uniform(
11061 1 : KeySpace::single(get_key(0)..get_key(10)),
11062 1 : delta_layer_end_lsn,
11063 : );
11064 :
11065 1 : let results = tline
11066 1 : .get_vectored(query, IoConcurrency::sequential(), &ctx)
11067 1 : .await
11068 1 : .expect("No vectored errors");
11069 11 : for (key, res) in results {
11070 10 : let value = res.expect("No key errors");
11071 10 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
11072 10 : assert_eq!(value, Bytes::from(expected_value));
11073 1 : }
11074 1 :
11075 1 : Ok(())
11076 1 : }
11077 :
11078 : #[cfg(feature = "testing")]
11079 : #[tokio::test]
11080 1 : async fn test_vectored_read_with_image_layer_inside_inmem() -> anyhow::Result<()> {
11081 1 : let harness =
11082 1 : TenantHarness::create("test_vectored_read_with_image_layer_inside_inmem").await?;
11083 1 : let (tenant, ctx) = harness.load().await;
11084 :
11085 1 : let will_init_keys = [2, 6];
11086 32 : fn get_key(id: u32) -> Key {
11087 32 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
11088 32 : key.field6 = id;
11089 32 : key
11090 32 : }
11091 :
11092 1 : let mut expected_key_values = HashMap::new();
11093 :
11094 1 : let baseline_image_layer_lsn = Lsn(0x10);
11095 1 : let mut baseline_img_layer = Vec::new();
11096 6 : for i in 0..5 {
11097 5 : let key = get_key(i);
11098 5 : let value = format!("value {i}@{baseline_image_layer_lsn}");
11099 :
11100 5 : let removed = expected_key_values.insert(key, value.clone());
11101 5 : assert!(removed.is_none());
11102 :
11103 5 : baseline_img_layer.push((key, Bytes::from(value)));
11104 : }
11105 :
11106 1 : let nested_image_layer_lsn = Lsn(0x50);
11107 1 : let mut nested_img_layer = Vec::new();
11108 6 : for i in 5..10 {
11109 5 : let key = get_key(i);
11110 5 : let value = format!("value {i}@{nested_image_layer_lsn}");
11111 :
11112 5 : let removed = expected_key_values.insert(key, value.clone());
11113 5 : assert!(removed.is_none());
11114 :
11115 5 : nested_img_layer.push((key, Bytes::from(value)));
11116 : }
11117 :
11118 1 : let frozen_layer = {
11119 1 : let lsn_range = Lsn(0x40)..Lsn(0x60);
11120 1 : let mut data = Vec::new();
11121 11 : for i in 0..10 {
11122 10 : let key = get_key(i);
11123 10 : let key_in_nested = nested_img_layer
11124 10 : .iter()
11125 40 : .any(|(key_with_img, _)| *key_with_img == key);
11126 10 : let lsn = {
11127 10 : if key_in_nested {
11128 5 : Lsn(nested_image_layer_lsn.0 + 5)
11129 : } else {
11130 5 : lsn_range.start
11131 : }
11132 : };
11133 :
11134 10 : let will_init = will_init_keys.contains(&i);
11135 10 : if will_init {
11136 2 : data.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init(""))));
11137 2 :
11138 2 : expected_key_values.insert(key, "".to_string());
11139 8 : } else {
11140 8 : let delta = format!("@{lsn}");
11141 8 : data.push((
11142 8 : key,
11143 8 : lsn,
11144 8 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
11145 8 : ));
11146 8 :
11147 8 : expected_key_values
11148 8 : .get_mut(&key)
11149 8 : .expect("An image exists for each key")
11150 8 : .push_str(delta.as_str());
11151 8 : }
11152 : }
11153 :
11154 1 : InMemoryLayerTestDesc {
11155 1 : lsn_range,
11156 1 : is_open: false,
11157 1 : data,
11158 1 : }
11159 : };
11160 :
11161 1 : let (open_layer, last_record_lsn) = {
11162 1 : let start_lsn = Lsn(0x70);
11163 1 : let mut data = Vec::new();
11164 1 : let mut end_lsn = Lsn(0);
11165 11 : for i in 0..10 {
11166 10 : let key = get_key(i);
11167 10 : let lsn = Lsn(start_lsn.0 + i as u64);
11168 10 : let delta = format!("@{lsn}");
11169 10 : data.push((
11170 10 : key,
11171 10 : lsn,
11172 10 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
11173 10 : ));
11174 10 :
11175 10 : expected_key_values
11176 10 : .get_mut(&key)
11177 10 : .expect("An image exists for each key")
11178 10 : .push_str(delta.as_str());
11179 10 :
11180 10 : end_lsn = std::cmp::max(end_lsn, lsn);
11181 10 : }
11182 :
11183 1 : (
11184 1 : InMemoryLayerTestDesc {
11185 1 : lsn_range: start_lsn..Lsn::MAX,
11186 1 : is_open: true,
11187 1 : data,
11188 1 : },
11189 1 : end_lsn,
11190 1 : )
11191 : };
11192 :
11193 1 : assert!(
11194 1 : nested_image_layer_lsn > frozen_layer.lsn_range.start
11195 1 : && nested_image_layer_lsn < frozen_layer.lsn_range.end
11196 : );
11197 :
11198 1 : let tline = tenant
11199 1 : .create_test_timeline_with_layers(
11200 1 : TIMELINE_ID,
11201 1 : baseline_image_layer_lsn,
11202 1 : DEFAULT_PG_VERSION,
11203 1 : &ctx,
11204 1 : vec![open_layer, frozen_layer], // in-memory layers
11205 1 : Vec::new(), // delta layers
11206 1 : vec![
11207 1 : (baseline_image_layer_lsn, baseline_img_layer),
11208 1 : (nested_image_layer_lsn, nested_img_layer),
11209 1 : ], // image layers
11210 1 : last_record_lsn,
11211 1 : )
11212 1 : .await?;
11213 :
11214 1 : let query = VersionedKeySpaceQuery::uniform(
11215 1 : KeySpace::single(get_key(0)..get_key(10)),
11216 1 : last_record_lsn,
11217 : );
11218 :
11219 1 : let results = tline
11220 1 : .get_vectored(query, IoConcurrency::sequential(), &ctx)
11221 1 : .await
11222 1 : .expect("No vectored errors");
11223 11 : for (key, res) in results {
11224 10 : let value = res.expect("No key errors");
11225 10 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
11226 10 : assert_eq!(value, Bytes::from(expected_value.clone()));
11227 1 :
11228 10 : tracing::info!("key={key} value={expected_value}");
11229 1 : }
11230 1 :
11231 1 : Ok(())
11232 1 : }
11233 :
11234 : // A randomized read path test. Generates a layer map according to a deterministic
11235 : // specification. Fills the (key, LSN) space in random manner and then performs
11236 : // random scattered queries validating the results against in-memory storage.
11237 : //
11238 : // See this internal Notion page for a diagram of the layer map:
11239 : // https://www.notion.so/neondatabase/Read-Path-Unit-Testing-Fuzzing-1d1f189e0047806c8e5cd37781b0a350?pvs=4
11240 : //
11241 : // A fuzzing mode is also supported. In this mode, the test will use a random
11242 : // seed instead of a hardcoded one. Use it in conjunction with `cargo stress`
11243 : // to run multiple instances in parallel:
11244 : //
11245 : // $ RUST_BACKTRACE=1 RUST_LOG=INFO \
11246 : // cargo stress --package=pageserver --features=testing,fuzz-read-path --release -- test_read_path
11247 : #[cfg(feature = "testing")]
11248 : #[tokio::test]
11249 1 : async fn test_read_path() -> anyhow::Result<()> {
11250 : use rand::seq::SliceRandom;
11251 :
11252 1 : let seed = if cfg!(feature = "fuzz-read-path") {
11253 0 : let seed: u64 = thread_rng().r#gen();
11254 0 : seed
11255 : } else {
11256 : // Use a hard-coded seed when not in fuzzing mode.
11257 : // Note that with the current approach results are not reproducible
11258 : // accross platforms and Rust releases.
11259 : const SEED: u64 = 0;
11260 1 : SEED
11261 : };
11262 :
11263 1 : let mut random = StdRng::seed_from_u64(seed);
11264 :
11265 1 : let (queries, will_init_chance, gap_chance) = if cfg!(feature = "fuzz-read-path") {
11266 : const QUERIES: u64 = 5000;
11267 0 : let will_init_chance: u8 = random.gen_range(0..=10);
11268 0 : let gap_chance: u8 = random.gen_range(0..=50);
11269 :
11270 0 : (QUERIES, will_init_chance, gap_chance)
11271 : } else {
11272 : const QUERIES: u64 = 1000;
11273 : const WILL_INIT_CHANCE: u8 = 1;
11274 : const GAP_CHANCE: u8 = 5;
11275 :
11276 1 : (QUERIES, WILL_INIT_CHANCE, GAP_CHANCE)
11277 : };
11278 :
11279 1 : let harness = TenantHarness::create("test_read_path").await?;
11280 1 : let (tenant, ctx) = harness.load().await;
11281 :
11282 1 : tracing::info!("Using random seed: {seed}");
11283 1 : tracing::info!(%will_init_chance, %gap_chance, "Fill params");
11284 :
11285 : // Define the layer map shape. Note that this part is not randomized.
11286 :
11287 : const KEY_DIMENSION_SIZE: u32 = 99;
11288 1 : let start_key = Key::from_hex("110000000033333333444444445500000000").unwrap();
11289 1 : let end_key = start_key.add(KEY_DIMENSION_SIZE);
11290 1 : let total_key_range = start_key..end_key;
11291 1 : let total_key_range_size = end_key.to_i128() - start_key.to_i128();
11292 1 : let total_start_lsn = Lsn(104);
11293 1 : let last_record_lsn = Lsn(504);
11294 :
11295 1 : assert!(total_key_range_size % 3 == 0);
11296 :
11297 1 : let in_memory_layers_shape = vec![
11298 1 : (total_key_range.clone(), Lsn(304)..Lsn(400)),
11299 1 : (total_key_range.clone(), Lsn(400)..last_record_lsn),
11300 : ];
11301 :
11302 1 : let delta_layers_shape = vec![
11303 1 : (
11304 1 : start_key..(start_key.add((total_key_range_size / 3) as u32)),
11305 1 : Lsn(200)..Lsn(304),
11306 1 : ),
11307 1 : (
11308 1 : (start_key.add((total_key_range_size / 3) as u32))
11309 1 : ..(start_key.add((total_key_range_size * 2 / 3) as u32)),
11310 1 : Lsn(200)..Lsn(304),
11311 1 : ),
11312 1 : (
11313 1 : (start_key.add((total_key_range_size * 2 / 3) as u32))
11314 1 : ..(start_key.add(total_key_range_size as u32)),
11315 1 : Lsn(200)..Lsn(304),
11316 1 : ),
11317 : ];
11318 :
11319 1 : let image_layers_shape = vec![
11320 1 : (
11321 1 : start_key.add((total_key_range_size * 2 / 3 - 10) as u32)
11322 1 : ..start_key.add((total_key_range_size * 2 / 3 + 10) as u32),
11323 1 : Lsn(456),
11324 1 : ),
11325 1 : (
11326 1 : start_key.add((total_key_range_size / 3 - 10) as u32)
11327 1 : ..start_key.add((total_key_range_size / 3 + 10) as u32),
11328 1 : Lsn(256),
11329 1 : ),
11330 1 : (total_key_range.clone(), total_start_lsn),
11331 : ];
11332 :
11333 1 : let specification = TestTimelineSpecification {
11334 1 : start_lsn: total_start_lsn,
11335 1 : last_record_lsn,
11336 1 : in_memory_layers_shape,
11337 1 : delta_layers_shape,
11338 1 : image_layers_shape,
11339 1 : gap_chance,
11340 1 : will_init_chance,
11341 1 : };
11342 :
11343 : // Create and randomly fill in the layers according to the specification
11344 1 : let (tline, storage, interesting_lsns) = randomize_timeline(
11345 1 : &tenant,
11346 1 : TIMELINE_ID,
11347 1 : DEFAULT_PG_VERSION,
11348 1 : specification,
11349 1 : &mut random,
11350 1 : &ctx,
11351 1 : )
11352 1 : .await?;
11353 :
11354 : // Now generate queries based on the interesting lsns that we've collected.
11355 : //
11356 : // While there's still room in the query, pick and interesting LSN and a random
11357 : // key. Then roll the dice to see if the next key should also be included in
11358 : // the query. When the roll fails, break the "batch" and pick another point in the
11359 : // (key, LSN) space.
11360 :
11361 : const PICK_NEXT_CHANCE: u8 = 50;
11362 1 : for _ in 0..queries {
11363 1000 : let query = {
11364 1000 : let mut keyspaces_at_lsn: HashMap<Lsn, KeySpaceRandomAccum> = HashMap::default();
11365 1000 : let mut used_keys: HashSet<Key> = HashSet::default();
11366 1 :
11367 22536 : while used_keys.len() < tenant.conf.max_get_vectored_keys.get() {
11368 21536 : let selected_lsn = interesting_lsns.choose(&mut random).expect("not empty");
11369 21536 : let mut selected_key = start_key.add(random.gen_range(0..KEY_DIMENSION_SIZE));
11370 1 :
11371 37614 : while used_keys.len() < tenant.conf.max_get_vectored_keys.get() {
11372 37093 : if used_keys.contains(&selected_key)
11373 32154 : || selected_key >= start_key.add(KEY_DIMENSION_SIZE)
11374 1 : {
11375 5093 : break;
11376 32000 : }
11377 1 :
11378 32000 : keyspaces_at_lsn
11379 32000 : .entry(*selected_lsn)
11380 32000 : .or_default()
11381 32000 : .add_key(selected_key);
11382 32000 : used_keys.insert(selected_key);
11383 1 :
11384 32000 : let pick_next = random.gen_range(0..=100) <= PICK_NEXT_CHANCE;
11385 32000 : if pick_next {
11386 16078 : selected_key = selected_key.next();
11387 16078 : } else {
11388 15922 : break;
11389 1 : }
11390 1 : }
11391 1 : }
11392 1 :
11393 1000 : VersionedKeySpaceQuery::scattered(
11394 1000 : keyspaces_at_lsn
11395 1000 : .into_iter()
11396 11917 : .map(|(lsn, acc)| (lsn, acc.to_keyspace()))
11397 1000 : .collect(),
11398 1 : )
11399 1 : };
11400 1 :
11401 1 : // Run the query and validate the results
11402 1 :
11403 1000 : let results = tline
11404 1000 : .get_vectored(query.clone(), IoConcurrency::Sequential, &ctx)
11405 1000 : .await;
11406 1 :
11407 1000 : let blobs = match results {
11408 1000 : Ok(ok) => ok,
11409 1 : Err(err) => {
11410 1 : panic!("seed={seed} Error returned for query {query}: {err}");
11411 1 : }
11412 1 : };
11413 1 :
11414 32000 : for (key, key_res) in blobs.into_iter() {
11415 32000 : match key_res {
11416 32000 : Ok(blob) => {
11417 32000 : let requested_at_lsn = query.map_key_to_lsn(&key);
11418 32000 : let expected = storage.get(key, requested_at_lsn);
11419 1 :
11420 32000 : if blob != expected {
11421 1 : tracing::error!(
11422 1 : "seed={seed} Mismatch for {key}@{requested_at_lsn} from query: {query}"
11423 1 : );
11424 32000 : }
11425 1 :
11426 32000 : assert_eq!(blob, expected);
11427 1 : }
11428 1 : Err(err) => {
11429 1 : let requested_at_lsn = query.map_key_to_lsn(&key);
11430 1 :
11431 1 : panic!(
11432 1 : "seed={seed} Error returned for {key}@{requested_at_lsn} from query {query}: {err}"
11433 1 : );
11434 1 : }
11435 1 : }
11436 1 : }
11437 1 : }
11438 1 :
11439 1 : Ok(())
11440 1 : }
11441 :
11442 107 : fn sort_layer_key(k1: &PersistentLayerKey, k2: &PersistentLayerKey) -> std::cmp::Ordering {
11443 107 : (
11444 107 : k1.is_delta,
11445 107 : k1.key_range.start,
11446 107 : k1.key_range.end,
11447 107 : k1.lsn_range.start,
11448 107 : k1.lsn_range.end,
11449 107 : )
11450 107 : .cmp(&(
11451 107 : k2.is_delta,
11452 107 : k2.key_range.start,
11453 107 : k2.key_range.end,
11454 107 : k2.lsn_range.start,
11455 107 : k2.lsn_range.end,
11456 107 : ))
11457 107 : }
11458 :
11459 12 : async fn inspect_and_sort(
11460 12 : tline: &Arc<Timeline>,
11461 12 : filter: Option<std::ops::Range<Key>>,
11462 12 : ) -> Vec<PersistentLayerKey> {
11463 12 : let mut all_layers = tline.inspect_historic_layers().await.unwrap();
11464 12 : if let Some(filter) = filter {
11465 54 : all_layers.retain(|layer| overlaps_with(&layer.key_range, &filter));
11466 1 : }
11467 12 : all_layers.sort_by(sort_layer_key);
11468 12 : all_layers
11469 12 : }
11470 :
11471 : #[cfg(feature = "testing")]
11472 11 : fn check_layer_map_key_eq(
11473 11 : mut left: Vec<PersistentLayerKey>,
11474 11 : mut right: Vec<PersistentLayerKey>,
11475 11 : ) {
11476 11 : left.sort_by(sort_layer_key);
11477 11 : right.sort_by(sort_layer_key);
11478 11 : if left != right {
11479 0 : eprintln!("---LEFT---");
11480 0 : for left in left.iter() {
11481 0 : eprintln!("{left}");
11482 0 : }
11483 0 : eprintln!("---RIGHT---");
11484 0 : for right in right.iter() {
11485 0 : eprintln!("{right}");
11486 0 : }
11487 0 : assert_eq!(left, right);
11488 11 : }
11489 11 : }
11490 :
11491 : #[cfg(feature = "testing")]
11492 : #[tokio::test]
11493 1 : async fn test_simple_partial_bottom_most_compaction() -> anyhow::Result<()> {
11494 1 : let harness = TenantHarness::create("test_simple_partial_bottom_most_compaction").await?;
11495 1 : let (tenant, ctx) = harness.load().await;
11496 :
11497 91 : fn get_key(id: u32) -> Key {
11498 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
11499 91 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
11500 91 : key.field6 = id;
11501 91 : key
11502 91 : }
11503 :
11504 : // img layer at 0x10
11505 1 : let img_layer = (0..10)
11506 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
11507 1 : .collect_vec();
11508 :
11509 1 : let delta1 = vec![
11510 1 : (
11511 1 : get_key(1),
11512 1 : Lsn(0x20),
11513 1 : Value::Image(Bytes::from("value 1@0x20")),
11514 1 : ),
11515 1 : (
11516 1 : get_key(2),
11517 1 : Lsn(0x30),
11518 1 : Value::Image(Bytes::from("value 2@0x30")),
11519 1 : ),
11520 1 : (
11521 1 : get_key(3),
11522 1 : Lsn(0x40),
11523 1 : Value::Image(Bytes::from("value 3@0x40")),
11524 1 : ),
11525 : ];
11526 1 : let delta2 = vec![
11527 1 : (
11528 1 : get_key(5),
11529 1 : Lsn(0x20),
11530 1 : Value::Image(Bytes::from("value 5@0x20")),
11531 1 : ),
11532 1 : (
11533 1 : get_key(6),
11534 1 : Lsn(0x20),
11535 1 : Value::Image(Bytes::from("value 6@0x20")),
11536 1 : ),
11537 : ];
11538 1 : let delta3 = vec![
11539 1 : (
11540 1 : get_key(8),
11541 1 : Lsn(0x48),
11542 1 : Value::Image(Bytes::from("value 8@0x48")),
11543 1 : ),
11544 1 : (
11545 1 : get_key(9),
11546 1 : Lsn(0x48),
11547 1 : Value::Image(Bytes::from("value 9@0x48")),
11548 1 : ),
11549 : ];
11550 :
11551 1 : let tline = tenant
11552 1 : .create_test_timeline_with_layers(
11553 1 : TIMELINE_ID,
11554 1 : Lsn(0x10),
11555 1 : DEFAULT_PG_VERSION,
11556 1 : &ctx,
11557 1 : vec![], // in-memory layers
11558 1 : vec![
11559 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
11560 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
11561 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
11562 1 : ], // delta layers
11563 1 : vec![(Lsn(0x10), img_layer)], // image layers
11564 1 : Lsn(0x50),
11565 1 : )
11566 1 : .await?;
11567 :
11568 : {
11569 1 : tline
11570 1 : .applied_gc_cutoff_lsn
11571 1 : .lock_for_write()
11572 1 : .store_and_unlock(Lsn(0x30))
11573 1 : .wait()
11574 1 : .await;
11575 : // Update GC info
11576 1 : let mut guard = tline.gc_info.write().unwrap();
11577 1 : *guard = GcInfo {
11578 1 : retain_lsns: vec![(Lsn(0x20), tline.timeline_id, MaybeOffloaded::No)],
11579 1 : cutoffs: GcCutoffs {
11580 1 : time: Some(Lsn(0x30)),
11581 1 : space: Lsn(0x30),
11582 1 : },
11583 1 : leases: Default::default(),
11584 1 : within_ancestor_pitr: false,
11585 1 : };
11586 : }
11587 :
11588 1 : let cancel = CancellationToken::new();
11589 :
11590 : // Do a partial compaction on key range 0..2
11591 1 : tline
11592 1 : .compact_with_gc(
11593 1 : &cancel,
11594 1 : CompactOptions {
11595 1 : flags: EnumSet::new(),
11596 1 : compact_key_range: Some((get_key(0)..get_key(2)).into()),
11597 1 : ..Default::default()
11598 1 : },
11599 1 : &ctx,
11600 1 : )
11601 1 : .await
11602 1 : .unwrap();
11603 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11604 1 : check_layer_map_key_eq(
11605 1 : all_layers,
11606 1 : vec![
11607 : // newly-generated image layer for the partial compaction range 0-2
11608 1 : PersistentLayerKey {
11609 1 : key_range: get_key(0)..get_key(2),
11610 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11611 1 : is_delta: false,
11612 1 : },
11613 1 : PersistentLayerKey {
11614 1 : key_range: get_key(0)..get_key(10),
11615 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
11616 1 : is_delta: false,
11617 1 : },
11618 : // delta1 is split and the second part is rewritten
11619 1 : PersistentLayerKey {
11620 1 : key_range: get_key(2)..get_key(4),
11621 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11622 1 : is_delta: true,
11623 1 : },
11624 1 : PersistentLayerKey {
11625 1 : key_range: get_key(5)..get_key(7),
11626 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11627 1 : is_delta: true,
11628 1 : },
11629 1 : PersistentLayerKey {
11630 1 : key_range: get_key(8)..get_key(10),
11631 1 : lsn_range: Lsn(0x48)..Lsn(0x50),
11632 1 : is_delta: true,
11633 1 : },
11634 : ],
11635 : );
11636 :
11637 : // Do a partial compaction on key range 2..4
11638 1 : tline
11639 1 : .compact_with_gc(
11640 1 : &cancel,
11641 1 : CompactOptions {
11642 1 : flags: EnumSet::new(),
11643 1 : compact_key_range: Some((get_key(2)..get_key(4)).into()),
11644 1 : ..Default::default()
11645 1 : },
11646 1 : &ctx,
11647 1 : )
11648 1 : .await
11649 1 : .unwrap();
11650 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11651 1 : check_layer_map_key_eq(
11652 1 : all_layers,
11653 1 : vec![
11654 1 : PersistentLayerKey {
11655 1 : key_range: get_key(0)..get_key(2),
11656 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11657 1 : is_delta: false,
11658 1 : },
11659 1 : PersistentLayerKey {
11660 1 : key_range: get_key(0)..get_key(10),
11661 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
11662 1 : is_delta: false,
11663 1 : },
11664 : // image layer generated for the compaction range 2-4
11665 1 : PersistentLayerKey {
11666 1 : key_range: get_key(2)..get_key(4),
11667 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11668 1 : is_delta: false,
11669 1 : },
11670 : // we have key2/key3 above the retain_lsn, so we still need this delta layer
11671 1 : PersistentLayerKey {
11672 1 : key_range: get_key(2)..get_key(4),
11673 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11674 1 : is_delta: true,
11675 1 : },
11676 1 : PersistentLayerKey {
11677 1 : key_range: get_key(5)..get_key(7),
11678 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11679 1 : is_delta: true,
11680 1 : },
11681 1 : PersistentLayerKey {
11682 1 : key_range: get_key(8)..get_key(10),
11683 1 : lsn_range: Lsn(0x48)..Lsn(0x50),
11684 1 : is_delta: true,
11685 1 : },
11686 : ],
11687 : );
11688 :
11689 : // Do a partial compaction on key range 4..9
11690 1 : tline
11691 1 : .compact_with_gc(
11692 1 : &cancel,
11693 1 : CompactOptions {
11694 1 : flags: EnumSet::new(),
11695 1 : compact_key_range: Some((get_key(4)..get_key(9)).into()),
11696 1 : ..Default::default()
11697 1 : },
11698 1 : &ctx,
11699 1 : )
11700 1 : .await
11701 1 : .unwrap();
11702 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11703 1 : check_layer_map_key_eq(
11704 1 : all_layers,
11705 1 : vec![
11706 1 : PersistentLayerKey {
11707 1 : key_range: get_key(0)..get_key(2),
11708 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11709 1 : is_delta: false,
11710 1 : },
11711 1 : PersistentLayerKey {
11712 1 : key_range: get_key(0)..get_key(10),
11713 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
11714 1 : is_delta: false,
11715 1 : },
11716 1 : PersistentLayerKey {
11717 1 : key_range: get_key(2)..get_key(4),
11718 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11719 1 : is_delta: false,
11720 1 : },
11721 1 : PersistentLayerKey {
11722 1 : key_range: get_key(2)..get_key(4),
11723 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11724 1 : is_delta: true,
11725 1 : },
11726 : // image layer generated for this compaction range
11727 1 : PersistentLayerKey {
11728 1 : key_range: get_key(4)..get_key(9),
11729 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11730 1 : is_delta: false,
11731 1 : },
11732 1 : PersistentLayerKey {
11733 1 : key_range: get_key(8)..get_key(10),
11734 1 : lsn_range: Lsn(0x48)..Lsn(0x50),
11735 1 : is_delta: true,
11736 1 : },
11737 : ],
11738 : );
11739 :
11740 : // Do a partial compaction on key range 9..10
11741 1 : tline
11742 1 : .compact_with_gc(
11743 1 : &cancel,
11744 1 : CompactOptions {
11745 1 : flags: EnumSet::new(),
11746 1 : compact_key_range: Some((get_key(9)..get_key(10)).into()),
11747 1 : ..Default::default()
11748 1 : },
11749 1 : &ctx,
11750 1 : )
11751 1 : .await
11752 1 : .unwrap();
11753 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11754 1 : check_layer_map_key_eq(
11755 1 : all_layers,
11756 1 : vec![
11757 1 : PersistentLayerKey {
11758 1 : key_range: get_key(0)..get_key(2),
11759 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11760 1 : is_delta: false,
11761 1 : },
11762 1 : PersistentLayerKey {
11763 1 : key_range: get_key(0)..get_key(10),
11764 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
11765 1 : is_delta: false,
11766 1 : },
11767 1 : PersistentLayerKey {
11768 1 : key_range: get_key(2)..get_key(4),
11769 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11770 1 : is_delta: false,
11771 1 : },
11772 1 : PersistentLayerKey {
11773 1 : key_range: get_key(2)..get_key(4),
11774 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11775 1 : is_delta: true,
11776 1 : },
11777 1 : PersistentLayerKey {
11778 1 : key_range: get_key(4)..get_key(9),
11779 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11780 1 : is_delta: false,
11781 1 : },
11782 : // image layer generated for the compaction range
11783 1 : PersistentLayerKey {
11784 1 : key_range: get_key(9)..get_key(10),
11785 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11786 1 : is_delta: false,
11787 1 : },
11788 1 : PersistentLayerKey {
11789 1 : key_range: get_key(8)..get_key(10),
11790 1 : lsn_range: Lsn(0x48)..Lsn(0x50),
11791 1 : is_delta: true,
11792 1 : },
11793 : ],
11794 : );
11795 :
11796 : // Do a partial compaction on key range 0..10, all image layers below LSN 20 can be replaced with new ones.
11797 1 : tline
11798 1 : .compact_with_gc(
11799 1 : &cancel,
11800 1 : CompactOptions {
11801 1 : flags: EnumSet::new(),
11802 1 : compact_key_range: Some((get_key(0)..get_key(10)).into()),
11803 1 : ..Default::default()
11804 1 : },
11805 1 : &ctx,
11806 1 : )
11807 1 : .await
11808 1 : .unwrap();
11809 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11810 1 : check_layer_map_key_eq(
11811 1 : all_layers,
11812 1 : vec![
11813 : // aha, we removed all unnecessary image/delta layers and got a very clean layer map!
11814 1 : PersistentLayerKey {
11815 1 : key_range: get_key(0)..get_key(10),
11816 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11817 1 : is_delta: false,
11818 1 : },
11819 1 : PersistentLayerKey {
11820 1 : key_range: get_key(2)..get_key(4),
11821 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11822 1 : is_delta: true,
11823 1 : },
11824 1 : PersistentLayerKey {
11825 1 : key_range: get_key(8)..get_key(10),
11826 1 : lsn_range: Lsn(0x48)..Lsn(0x50),
11827 1 : is_delta: true,
11828 1 : },
11829 : ],
11830 : );
11831 2 : Ok(())
11832 1 : }
11833 :
11834 : #[cfg(feature = "testing")]
11835 : #[tokio::test]
11836 1 : async fn test_timeline_offload_retain_lsn() -> anyhow::Result<()> {
11837 1 : let harness = TenantHarness::create("test_timeline_offload_retain_lsn")
11838 1 : .await
11839 1 : .unwrap();
11840 1 : let (tenant, ctx) = harness.load().await;
11841 1 : let tline_parent = tenant
11842 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
11843 1 : .await
11844 1 : .unwrap();
11845 1 : let tline_child = tenant
11846 1 : .branch_timeline_test(&tline_parent, NEW_TIMELINE_ID, Some(Lsn(0x20)), &ctx)
11847 1 : .await
11848 1 : .unwrap();
11849 : {
11850 1 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
11851 1 : assert_eq!(
11852 1 : gc_info_parent.retain_lsns,
11853 1 : vec![(Lsn(0x20), tline_child.timeline_id, MaybeOffloaded::No)]
11854 : );
11855 : }
11856 : // We have to directly call the remote_client instead of using the archive function to avoid constructing broker client...
11857 1 : tline_child
11858 1 : .remote_client
11859 1 : .schedule_index_upload_for_timeline_archival_state(TimelineArchivalState::Archived)
11860 1 : .unwrap();
11861 1 : tline_child.remote_client.wait_completion().await.unwrap();
11862 1 : offload_timeline(&tenant, &tline_child)
11863 1 : .instrument(tracing::info_span!(parent: None, "offload_test", tenant_id=%"test", shard_id=%"test", timeline_id=%"test"))
11864 1 : .await.unwrap();
11865 1 : let child_timeline_id = tline_child.timeline_id;
11866 1 : Arc::try_unwrap(tline_child).unwrap();
11867 :
11868 : {
11869 1 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
11870 1 : assert_eq!(
11871 1 : gc_info_parent.retain_lsns,
11872 1 : vec![(Lsn(0x20), child_timeline_id, MaybeOffloaded::Yes)]
11873 : );
11874 : }
11875 :
11876 1 : tenant
11877 1 : .get_offloaded_timeline(child_timeline_id)
11878 1 : .unwrap()
11879 1 : .defuse_for_tenant_drop();
11880 :
11881 2 : Ok(())
11882 1 : }
11883 :
11884 : #[cfg(feature = "testing")]
11885 : #[tokio::test]
11886 1 : async fn test_simple_bottom_most_compaction_above_lsn() -> anyhow::Result<()> {
11887 1 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_above_lsn").await?;
11888 1 : let (tenant, ctx) = harness.load().await;
11889 :
11890 148 : fn get_key(id: u32) -> Key {
11891 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
11892 148 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
11893 148 : key.field6 = id;
11894 148 : key
11895 148 : }
11896 :
11897 1 : let img_layer = (0..10)
11898 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
11899 1 : .collect_vec();
11900 :
11901 1 : let delta1 = vec![(
11902 1 : get_key(1),
11903 1 : Lsn(0x20),
11904 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
11905 1 : )];
11906 1 : let delta4 = vec![(
11907 1 : get_key(1),
11908 1 : Lsn(0x28),
11909 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
11910 1 : )];
11911 1 : let delta2 = vec![
11912 1 : (
11913 1 : get_key(1),
11914 1 : Lsn(0x30),
11915 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
11916 1 : ),
11917 1 : (
11918 1 : get_key(1),
11919 1 : Lsn(0x38),
11920 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
11921 1 : ),
11922 : ];
11923 1 : let delta3 = vec![
11924 1 : (
11925 1 : get_key(8),
11926 1 : Lsn(0x48),
11927 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
11928 1 : ),
11929 1 : (
11930 1 : get_key(9),
11931 1 : Lsn(0x48),
11932 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
11933 1 : ),
11934 : ];
11935 :
11936 1 : let tline = tenant
11937 1 : .create_test_timeline_with_layers(
11938 1 : TIMELINE_ID,
11939 1 : Lsn(0x10),
11940 1 : DEFAULT_PG_VERSION,
11941 1 : &ctx,
11942 1 : vec![], // in-memory layers
11943 1 : vec![
11944 1 : // delta1/2/4 only contain a single key but multiple updates
11945 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x28), delta1),
11946 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
11947 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x28)..Lsn(0x30), delta4),
11948 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta3),
11949 1 : ], // delta layers
11950 1 : vec![(Lsn(0x10), img_layer)], // image layers
11951 1 : Lsn(0x50),
11952 1 : )
11953 1 : .await?;
11954 : {
11955 1 : tline
11956 1 : .applied_gc_cutoff_lsn
11957 1 : .lock_for_write()
11958 1 : .store_and_unlock(Lsn(0x30))
11959 1 : .wait()
11960 1 : .await;
11961 : // Update GC info
11962 1 : let mut guard = tline.gc_info.write().unwrap();
11963 1 : *guard = GcInfo {
11964 1 : retain_lsns: vec![
11965 1 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
11966 1 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
11967 1 : ],
11968 1 : cutoffs: GcCutoffs {
11969 1 : time: Some(Lsn(0x30)),
11970 1 : space: Lsn(0x30),
11971 1 : },
11972 1 : leases: Default::default(),
11973 1 : within_ancestor_pitr: false,
11974 1 : };
11975 : }
11976 :
11977 1 : let expected_result = [
11978 1 : Bytes::from_static(b"value 0@0x10"),
11979 1 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
11980 1 : Bytes::from_static(b"value 2@0x10"),
11981 1 : Bytes::from_static(b"value 3@0x10"),
11982 1 : Bytes::from_static(b"value 4@0x10"),
11983 1 : Bytes::from_static(b"value 5@0x10"),
11984 1 : Bytes::from_static(b"value 6@0x10"),
11985 1 : Bytes::from_static(b"value 7@0x10"),
11986 1 : Bytes::from_static(b"value 8@0x10@0x48"),
11987 1 : Bytes::from_static(b"value 9@0x10@0x48"),
11988 1 : ];
11989 :
11990 1 : let expected_result_at_gc_horizon = [
11991 1 : Bytes::from_static(b"value 0@0x10"),
11992 1 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
11993 1 : Bytes::from_static(b"value 2@0x10"),
11994 1 : Bytes::from_static(b"value 3@0x10"),
11995 1 : Bytes::from_static(b"value 4@0x10"),
11996 1 : Bytes::from_static(b"value 5@0x10"),
11997 1 : Bytes::from_static(b"value 6@0x10"),
11998 1 : Bytes::from_static(b"value 7@0x10"),
11999 1 : Bytes::from_static(b"value 8@0x10"),
12000 1 : Bytes::from_static(b"value 9@0x10"),
12001 1 : ];
12002 :
12003 1 : let expected_result_at_lsn_20 = [
12004 1 : Bytes::from_static(b"value 0@0x10"),
12005 1 : Bytes::from_static(b"value 1@0x10@0x20"),
12006 1 : Bytes::from_static(b"value 2@0x10"),
12007 1 : Bytes::from_static(b"value 3@0x10"),
12008 1 : Bytes::from_static(b"value 4@0x10"),
12009 1 : Bytes::from_static(b"value 5@0x10"),
12010 1 : Bytes::from_static(b"value 6@0x10"),
12011 1 : Bytes::from_static(b"value 7@0x10"),
12012 1 : Bytes::from_static(b"value 8@0x10"),
12013 1 : Bytes::from_static(b"value 9@0x10"),
12014 1 : ];
12015 :
12016 1 : let expected_result_at_lsn_10 = [
12017 1 : Bytes::from_static(b"value 0@0x10"),
12018 1 : Bytes::from_static(b"value 1@0x10"),
12019 1 : Bytes::from_static(b"value 2@0x10"),
12020 1 : Bytes::from_static(b"value 3@0x10"),
12021 1 : Bytes::from_static(b"value 4@0x10"),
12022 1 : Bytes::from_static(b"value 5@0x10"),
12023 1 : Bytes::from_static(b"value 6@0x10"),
12024 1 : Bytes::from_static(b"value 7@0x10"),
12025 1 : Bytes::from_static(b"value 8@0x10"),
12026 1 : Bytes::from_static(b"value 9@0x10"),
12027 1 : ];
12028 :
12029 3 : let verify_result = || async {
12030 3 : let gc_horizon = {
12031 3 : let gc_info = tline.gc_info.read().unwrap();
12032 3 : gc_info.cutoffs.time.unwrap_or_default()
12033 : };
12034 33 : for idx in 0..10 {
12035 30 : assert_eq!(
12036 30 : tline
12037 30 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
12038 30 : .await
12039 30 : .unwrap(),
12040 30 : &expected_result[idx]
12041 : );
12042 30 : assert_eq!(
12043 30 : tline
12044 30 : .get(get_key(idx as u32), gc_horizon, &ctx)
12045 30 : .await
12046 30 : .unwrap(),
12047 30 : &expected_result_at_gc_horizon[idx]
12048 : );
12049 30 : assert_eq!(
12050 30 : tline
12051 30 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
12052 30 : .await
12053 30 : .unwrap(),
12054 30 : &expected_result_at_lsn_20[idx]
12055 : );
12056 30 : assert_eq!(
12057 30 : tline
12058 30 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
12059 30 : .await
12060 30 : .unwrap(),
12061 30 : &expected_result_at_lsn_10[idx]
12062 : );
12063 : }
12064 6 : };
12065 :
12066 1 : verify_result().await;
12067 :
12068 1 : let cancel = CancellationToken::new();
12069 1 : tline
12070 1 : .compact_with_gc(
12071 1 : &cancel,
12072 1 : CompactOptions {
12073 1 : compact_lsn_range: Some(CompactLsnRange::above(Lsn(0x28))),
12074 1 : ..Default::default()
12075 1 : },
12076 1 : &ctx,
12077 1 : )
12078 1 : .await
12079 1 : .unwrap();
12080 1 : verify_result().await;
12081 :
12082 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12083 1 : check_layer_map_key_eq(
12084 1 : all_layers,
12085 1 : vec![
12086 : // The original image layer, not compacted
12087 1 : PersistentLayerKey {
12088 1 : key_range: get_key(0)..get_key(10),
12089 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
12090 1 : is_delta: false,
12091 1 : },
12092 : // Delta layer below the specified above_lsn not compacted
12093 1 : PersistentLayerKey {
12094 1 : key_range: get_key(1)..get_key(2),
12095 1 : lsn_range: Lsn(0x20)..Lsn(0x28),
12096 1 : is_delta: true,
12097 1 : },
12098 : // Delta layer compacted above the LSN
12099 1 : PersistentLayerKey {
12100 1 : key_range: get_key(1)..get_key(10),
12101 1 : lsn_range: Lsn(0x28)..Lsn(0x50),
12102 1 : is_delta: true,
12103 1 : },
12104 : ],
12105 : );
12106 :
12107 : // compact again
12108 1 : tline
12109 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
12110 1 : .await
12111 1 : .unwrap();
12112 1 : verify_result().await;
12113 :
12114 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12115 1 : check_layer_map_key_eq(
12116 1 : all_layers,
12117 1 : vec![
12118 : // The compacted image layer (full key range)
12119 1 : PersistentLayerKey {
12120 1 : key_range: Key::MIN..Key::MAX,
12121 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
12122 1 : is_delta: false,
12123 1 : },
12124 : // All other data in the delta layer
12125 1 : PersistentLayerKey {
12126 1 : key_range: get_key(1)..get_key(10),
12127 1 : lsn_range: Lsn(0x10)..Lsn(0x50),
12128 1 : is_delta: true,
12129 1 : },
12130 : ],
12131 : );
12132 :
12133 2 : Ok(())
12134 1 : }
12135 :
12136 : #[cfg(feature = "testing")]
12137 : #[tokio::test]
12138 1 : async fn test_simple_bottom_most_compaction_rectangle() -> anyhow::Result<()> {
12139 1 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_rectangle").await?;
12140 1 : let (tenant, ctx) = harness.load().await;
12141 :
12142 254 : fn get_key(id: u32) -> Key {
12143 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
12144 254 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
12145 254 : key.field6 = id;
12146 254 : key
12147 254 : }
12148 :
12149 1 : let img_layer = (0..10)
12150 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
12151 1 : .collect_vec();
12152 :
12153 1 : let delta1 = vec![(
12154 1 : get_key(1),
12155 1 : Lsn(0x20),
12156 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
12157 1 : )];
12158 1 : let delta4 = vec![(
12159 1 : get_key(1),
12160 1 : Lsn(0x28),
12161 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
12162 1 : )];
12163 1 : let delta2 = vec![
12164 1 : (
12165 1 : get_key(1),
12166 1 : Lsn(0x30),
12167 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
12168 1 : ),
12169 1 : (
12170 1 : get_key(1),
12171 1 : Lsn(0x38),
12172 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
12173 1 : ),
12174 : ];
12175 1 : let delta3 = vec![
12176 1 : (
12177 1 : get_key(8),
12178 1 : Lsn(0x48),
12179 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
12180 1 : ),
12181 1 : (
12182 1 : get_key(9),
12183 1 : Lsn(0x48),
12184 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
12185 1 : ),
12186 : ];
12187 :
12188 1 : let tline = tenant
12189 1 : .create_test_timeline_with_layers(
12190 1 : TIMELINE_ID,
12191 1 : Lsn(0x10),
12192 1 : DEFAULT_PG_VERSION,
12193 1 : &ctx,
12194 1 : vec![], // in-memory layers
12195 1 : vec![
12196 1 : // delta1/2/4 only contain a single key but multiple updates
12197 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x28), delta1),
12198 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
12199 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x28)..Lsn(0x30), delta4),
12200 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta3),
12201 1 : ], // delta layers
12202 1 : vec![(Lsn(0x10), img_layer)], // image layers
12203 1 : Lsn(0x50),
12204 1 : )
12205 1 : .await?;
12206 : {
12207 1 : tline
12208 1 : .applied_gc_cutoff_lsn
12209 1 : .lock_for_write()
12210 1 : .store_and_unlock(Lsn(0x30))
12211 1 : .wait()
12212 1 : .await;
12213 : // Update GC info
12214 1 : let mut guard = tline.gc_info.write().unwrap();
12215 1 : *guard = GcInfo {
12216 1 : retain_lsns: vec![
12217 1 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
12218 1 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
12219 1 : ],
12220 1 : cutoffs: GcCutoffs {
12221 1 : time: Some(Lsn(0x30)),
12222 1 : space: Lsn(0x30),
12223 1 : },
12224 1 : leases: Default::default(),
12225 1 : within_ancestor_pitr: false,
12226 1 : };
12227 : }
12228 :
12229 1 : let expected_result = [
12230 1 : Bytes::from_static(b"value 0@0x10"),
12231 1 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
12232 1 : Bytes::from_static(b"value 2@0x10"),
12233 1 : Bytes::from_static(b"value 3@0x10"),
12234 1 : Bytes::from_static(b"value 4@0x10"),
12235 1 : Bytes::from_static(b"value 5@0x10"),
12236 1 : Bytes::from_static(b"value 6@0x10"),
12237 1 : Bytes::from_static(b"value 7@0x10"),
12238 1 : Bytes::from_static(b"value 8@0x10@0x48"),
12239 1 : Bytes::from_static(b"value 9@0x10@0x48"),
12240 1 : ];
12241 :
12242 1 : let expected_result_at_gc_horizon = [
12243 1 : Bytes::from_static(b"value 0@0x10"),
12244 1 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
12245 1 : Bytes::from_static(b"value 2@0x10"),
12246 1 : Bytes::from_static(b"value 3@0x10"),
12247 1 : Bytes::from_static(b"value 4@0x10"),
12248 1 : Bytes::from_static(b"value 5@0x10"),
12249 1 : Bytes::from_static(b"value 6@0x10"),
12250 1 : Bytes::from_static(b"value 7@0x10"),
12251 1 : Bytes::from_static(b"value 8@0x10"),
12252 1 : Bytes::from_static(b"value 9@0x10"),
12253 1 : ];
12254 :
12255 1 : let expected_result_at_lsn_20 = [
12256 1 : Bytes::from_static(b"value 0@0x10"),
12257 1 : Bytes::from_static(b"value 1@0x10@0x20"),
12258 1 : Bytes::from_static(b"value 2@0x10"),
12259 1 : Bytes::from_static(b"value 3@0x10"),
12260 1 : Bytes::from_static(b"value 4@0x10"),
12261 1 : Bytes::from_static(b"value 5@0x10"),
12262 1 : Bytes::from_static(b"value 6@0x10"),
12263 1 : Bytes::from_static(b"value 7@0x10"),
12264 1 : Bytes::from_static(b"value 8@0x10"),
12265 1 : Bytes::from_static(b"value 9@0x10"),
12266 1 : ];
12267 :
12268 1 : let expected_result_at_lsn_10 = [
12269 1 : Bytes::from_static(b"value 0@0x10"),
12270 1 : Bytes::from_static(b"value 1@0x10"),
12271 1 : Bytes::from_static(b"value 2@0x10"),
12272 1 : Bytes::from_static(b"value 3@0x10"),
12273 1 : Bytes::from_static(b"value 4@0x10"),
12274 1 : Bytes::from_static(b"value 5@0x10"),
12275 1 : Bytes::from_static(b"value 6@0x10"),
12276 1 : Bytes::from_static(b"value 7@0x10"),
12277 1 : Bytes::from_static(b"value 8@0x10"),
12278 1 : Bytes::from_static(b"value 9@0x10"),
12279 1 : ];
12280 :
12281 5 : let verify_result = || async {
12282 5 : let gc_horizon = {
12283 5 : let gc_info = tline.gc_info.read().unwrap();
12284 5 : gc_info.cutoffs.time.unwrap_or_default()
12285 : };
12286 55 : for idx in 0..10 {
12287 50 : assert_eq!(
12288 50 : tline
12289 50 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
12290 50 : .await
12291 50 : .unwrap(),
12292 50 : &expected_result[idx]
12293 : );
12294 50 : assert_eq!(
12295 50 : tline
12296 50 : .get(get_key(idx as u32), gc_horizon, &ctx)
12297 50 : .await
12298 50 : .unwrap(),
12299 50 : &expected_result_at_gc_horizon[idx]
12300 : );
12301 50 : assert_eq!(
12302 50 : tline
12303 50 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
12304 50 : .await
12305 50 : .unwrap(),
12306 50 : &expected_result_at_lsn_20[idx]
12307 : );
12308 50 : assert_eq!(
12309 50 : tline
12310 50 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
12311 50 : .await
12312 50 : .unwrap(),
12313 50 : &expected_result_at_lsn_10[idx]
12314 : );
12315 : }
12316 10 : };
12317 :
12318 1 : verify_result().await;
12319 :
12320 1 : let cancel = CancellationToken::new();
12321 :
12322 1 : tline
12323 1 : .compact_with_gc(
12324 1 : &cancel,
12325 1 : CompactOptions {
12326 1 : compact_key_range: Some((get_key(0)..get_key(2)).into()),
12327 1 : compact_lsn_range: Some((Lsn(0x20)..Lsn(0x28)).into()),
12328 1 : ..Default::default()
12329 1 : },
12330 1 : &ctx,
12331 1 : )
12332 1 : .await
12333 1 : .unwrap();
12334 1 : verify_result().await;
12335 :
12336 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12337 1 : check_layer_map_key_eq(
12338 1 : all_layers,
12339 1 : vec![
12340 : // The original image layer, not compacted
12341 1 : PersistentLayerKey {
12342 1 : key_range: get_key(0)..get_key(10),
12343 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
12344 1 : is_delta: false,
12345 1 : },
12346 : // According the selection logic, we select all layers with start key <= 0x28, so we would merge the layer 0x20-0x28 and
12347 : // the layer 0x28-0x30 into one.
12348 1 : PersistentLayerKey {
12349 1 : key_range: get_key(1)..get_key(2),
12350 1 : lsn_range: Lsn(0x20)..Lsn(0x30),
12351 1 : is_delta: true,
12352 1 : },
12353 : // Above the upper bound and untouched
12354 1 : PersistentLayerKey {
12355 1 : key_range: get_key(1)..get_key(2),
12356 1 : lsn_range: Lsn(0x30)..Lsn(0x50),
12357 1 : is_delta: true,
12358 1 : },
12359 : // This layer is untouched
12360 1 : PersistentLayerKey {
12361 1 : key_range: get_key(8)..get_key(10),
12362 1 : lsn_range: Lsn(0x30)..Lsn(0x50),
12363 1 : is_delta: true,
12364 1 : },
12365 : ],
12366 : );
12367 :
12368 1 : tline
12369 1 : .compact_with_gc(
12370 1 : &cancel,
12371 1 : CompactOptions {
12372 1 : compact_key_range: Some((get_key(3)..get_key(8)).into()),
12373 1 : compact_lsn_range: Some((Lsn(0x28)..Lsn(0x40)).into()),
12374 1 : ..Default::default()
12375 1 : },
12376 1 : &ctx,
12377 1 : )
12378 1 : .await
12379 1 : .unwrap();
12380 1 : verify_result().await;
12381 :
12382 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12383 1 : check_layer_map_key_eq(
12384 1 : all_layers,
12385 1 : vec![
12386 : // The original image layer, not compacted
12387 1 : PersistentLayerKey {
12388 1 : key_range: get_key(0)..get_key(10),
12389 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
12390 1 : is_delta: false,
12391 1 : },
12392 : // Not in the compaction key range, uncompacted
12393 1 : PersistentLayerKey {
12394 1 : key_range: get_key(1)..get_key(2),
12395 1 : lsn_range: Lsn(0x20)..Lsn(0x30),
12396 1 : is_delta: true,
12397 1 : },
12398 : // Not in the compaction key range, uncompacted but need rewrite because the delta layer overlaps with the range
12399 1 : PersistentLayerKey {
12400 1 : key_range: get_key(1)..get_key(2),
12401 1 : lsn_range: Lsn(0x30)..Lsn(0x50),
12402 1 : is_delta: true,
12403 1 : },
12404 : // Note that when we specify the LSN upper bound to be 0x40, the compaction algorithm will not try to cut the layer
12405 : // horizontally in half. Instead, it will include all LSNs that overlap with 0x40. So the real max_lsn of the compaction
12406 : // becomes 0x50.
12407 1 : PersistentLayerKey {
12408 1 : key_range: get_key(8)..get_key(10),
12409 1 : lsn_range: Lsn(0x30)..Lsn(0x50),
12410 1 : is_delta: true,
12411 1 : },
12412 : ],
12413 : );
12414 :
12415 : // compact again
12416 1 : tline
12417 1 : .compact_with_gc(
12418 1 : &cancel,
12419 1 : CompactOptions {
12420 1 : compact_key_range: Some((get_key(0)..get_key(5)).into()),
12421 1 : compact_lsn_range: Some((Lsn(0x20)..Lsn(0x50)).into()),
12422 1 : ..Default::default()
12423 1 : },
12424 1 : &ctx,
12425 1 : )
12426 1 : .await
12427 1 : .unwrap();
12428 1 : verify_result().await;
12429 :
12430 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12431 1 : check_layer_map_key_eq(
12432 1 : all_layers,
12433 1 : vec![
12434 : // The original image layer, not compacted
12435 1 : PersistentLayerKey {
12436 1 : key_range: get_key(0)..get_key(10),
12437 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
12438 1 : is_delta: false,
12439 1 : },
12440 : // The range gets compacted
12441 1 : PersistentLayerKey {
12442 1 : key_range: get_key(1)..get_key(2),
12443 1 : lsn_range: Lsn(0x20)..Lsn(0x50),
12444 1 : is_delta: true,
12445 1 : },
12446 : // Not touched during this iteration of compaction
12447 1 : PersistentLayerKey {
12448 1 : key_range: get_key(8)..get_key(10),
12449 1 : lsn_range: Lsn(0x30)..Lsn(0x50),
12450 1 : is_delta: true,
12451 1 : },
12452 : ],
12453 : );
12454 :
12455 : // final full compaction
12456 1 : tline
12457 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
12458 1 : .await
12459 1 : .unwrap();
12460 1 : verify_result().await;
12461 :
12462 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12463 1 : check_layer_map_key_eq(
12464 1 : all_layers,
12465 1 : vec![
12466 : // The compacted image layer (full key range)
12467 1 : PersistentLayerKey {
12468 1 : key_range: Key::MIN..Key::MAX,
12469 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
12470 1 : is_delta: false,
12471 1 : },
12472 : // All other data in the delta layer
12473 1 : PersistentLayerKey {
12474 1 : key_range: get_key(1)..get_key(10),
12475 1 : lsn_range: Lsn(0x10)..Lsn(0x50),
12476 1 : is_delta: true,
12477 1 : },
12478 : ],
12479 : );
12480 :
12481 2 : Ok(())
12482 1 : }
12483 :
12484 : #[cfg(feature = "testing")]
12485 : #[tokio::test]
12486 1 : async fn test_bottom_most_compation_redo_failure() -> anyhow::Result<()> {
12487 1 : let harness = TenantHarness::create("test_bottom_most_compation_redo_failure").await?;
12488 1 : let (tenant, ctx) = harness.load().await;
12489 :
12490 13 : fn get_key(id: u32) -> Key {
12491 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
12492 13 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
12493 13 : key.field6 = id;
12494 13 : key
12495 13 : }
12496 :
12497 1 : let img_layer = (0..10)
12498 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
12499 1 : .collect_vec();
12500 :
12501 1 : let delta1 = vec![
12502 1 : (
12503 1 : get_key(1),
12504 1 : Lsn(0x20),
12505 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
12506 1 : ),
12507 1 : (
12508 1 : get_key(1),
12509 1 : Lsn(0x24),
12510 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x24")),
12511 1 : ),
12512 1 : (
12513 1 : get_key(1),
12514 1 : Lsn(0x28),
12515 1 : // This record will fail to redo
12516 1 : Value::WalRecord(NeonWalRecord::wal_append_conditional("@0x28", "???")),
12517 1 : ),
12518 : ];
12519 :
12520 1 : let tline = tenant
12521 1 : .create_test_timeline_with_layers(
12522 1 : TIMELINE_ID,
12523 1 : Lsn(0x10),
12524 1 : DEFAULT_PG_VERSION,
12525 1 : &ctx,
12526 1 : vec![], // in-memory layers
12527 1 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
12528 1 : Lsn(0x20)..Lsn(0x30),
12529 1 : delta1,
12530 1 : )], // delta layers
12531 1 : vec![(Lsn(0x10), img_layer)], // image layers
12532 1 : Lsn(0x50),
12533 1 : )
12534 1 : .await?;
12535 : {
12536 1 : tline
12537 1 : .applied_gc_cutoff_lsn
12538 1 : .lock_for_write()
12539 1 : .store_and_unlock(Lsn(0x30))
12540 1 : .wait()
12541 1 : .await;
12542 : // Update GC info
12543 1 : let mut guard = tline.gc_info.write().unwrap();
12544 1 : *guard = GcInfo {
12545 1 : retain_lsns: vec![],
12546 1 : cutoffs: GcCutoffs {
12547 1 : time: Some(Lsn(0x30)),
12548 1 : space: Lsn(0x30),
12549 1 : },
12550 1 : leases: Default::default(),
12551 1 : within_ancestor_pitr: false,
12552 1 : };
12553 : }
12554 :
12555 1 : let cancel = CancellationToken::new();
12556 :
12557 : // Compaction will fail, but should not fire any critical error.
12558 : // Gc-compaction currently cannot figure out what keys are not in the keyspace during the compaction
12559 : // process. It will always try to redo the logs it reads and if it doesn't work, fail the entire
12560 : // compaction job. Tracked in <https://github.com/neondatabase/neon/issues/10395>.
12561 1 : let res = tline
12562 1 : .compact_with_gc(
12563 1 : &cancel,
12564 1 : CompactOptions {
12565 1 : compact_key_range: None,
12566 1 : compact_lsn_range: None,
12567 1 : ..Default::default()
12568 1 : },
12569 1 : &ctx,
12570 1 : )
12571 1 : .await;
12572 1 : assert!(res.is_err());
12573 :
12574 2 : Ok(())
12575 1 : }
12576 :
12577 : #[cfg(feature = "testing")]
12578 : #[tokio::test]
12579 1 : async fn test_synthetic_size_calculation_with_invisible_branches() -> anyhow::Result<()> {
12580 : use pageserver_api::models::TimelineVisibilityState;
12581 :
12582 : use crate::tenant::size::gather_inputs;
12583 :
12584 1 : let tenant_conf = pageserver_api::models::TenantConfig {
12585 1 : // Ensure that we don't compute gc_cutoffs (which needs reading the layer files)
12586 1 : pitr_interval: Some(Duration::ZERO),
12587 1 : ..Default::default()
12588 1 : };
12589 1 : let harness = TenantHarness::create_custom(
12590 1 : "test_synthetic_size_calculation_with_invisible_branches",
12591 1 : tenant_conf,
12592 1 : TenantId::generate(),
12593 1 : ShardIdentity::unsharded(),
12594 1 : Generation::new(0xdeadbeef),
12595 1 : )
12596 1 : .await?;
12597 1 : let (tenant, ctx) = harness.load().await;
12598 1 : let main_tline = tenant
12599 1 : .create_test_timeline_with_layers(
12600 1 : TIMELINE_ID,
12601 1 : Lsn(0x10),
12602 1 : DEFAULT_PG_VERSION,
12603 1 : &ctx,
12604 1 : vec![],
12605 1 : vec![],
12606 1 : vec![],
12607 1 : Lsn(0x100),
12608 1 : )
12609 1 : .await?;
12610 :
12611 1 : let snapshot1 = TimelineId::from_array(hex!("11223344556677881122334455667790"));
12612 1 : tenant
12613 1 : .branch_timeline_test_with_layers(
12614 1 : &main_tline,
12615 1 : snapshot1,
12616 1 : Some(Lsn(0x20)),
12617 1 : &ctx,
12618 1 : vec![],
12619 1 : vec![],
12620 1 : Lsn(0x50),
12621 1 : )
12622 1 : .await?;
12623 1 : let snapshot2 = TimelineId::from_array(hex!("11223344556677881122334455667791"));
12624 1 : tenant
12625 1 : .branch_timeline_test_with_layers(
12626 1 : &main_tline,
12627 1 : snapshot2,
12628 1 : Some(Lsn(0x30)),
12629 1 : &ctx,
12630 1 : vec![],
12631 1 : vec![],
12632 1 : Lsn(0x50),
12633 1 : )
12634 1 : .await?;
12635 1 : let snapshot3 = TimelineId::from_array(hex!("11223344556677881122334455667792"));
12636 1 : tenant
12637 1 : .branch_timeline_test_with_layers(
12638 1 : &main_tline,
12639 1 : snapshot3,
12640 1 : Some(Lsn(0x40)),
12641 1 : &ctx,
12642 1 : vec![],
12643 1 : vec![],
12644 1 : Lsn(0x50),
12645 1 : )
12646 1 : .await?;
12647 1 : let limit = Arc::new(Semaphore::new(1));
12648 1 : let max_retention_period = None;
12649 1 : let mut logical_size_cache = HashMap::new();
12650 1 : let cause = LogicalSizeCalculationCause::EvictionTaskImitation;
12651 1 : let cancel = CancellationToken::new();
12652 :
12653 1 : let inputs = gather_inputs(
12654 1 : &tenant,
12655 1 : &limit,
12656 1 : max_retention_period,
12657 1 : &mut logical_size_cache,
12658 1 : cause,
12659 1 : &cancel,
12660 1 : &ctx,
12661 : )
12662 1 : .instrument(info_span!(
12663 : "gather_inputs",
12664 : tenant_id = "unknown",
12665 : shard_id = "unknown",
12666 : ))
12667 1 : .await?;
12668 : use crate::tenant::size::{LsnKind, ModelInputs, SegmentMeta};
12669 : use LsnKind::*;
12670 : use tenant_size_model::Segment;
12671 1 : let ModelInputs { mut segments, .. } = inputs;
12672 15 : segments.retain(|s| s.timeline_id == TIMELINE_ID);
12673 6 : for segment in segments.iter_mut() {
12674 6 : segment.segment.parent = None; // We don't care about the parent for the test
12675 6 : segment.segment.size = None; // We don't care about the size for the test
12676 6 : }
12677 1 : assert_eq!(
12678 : segments,
12679 : [
12680 : SegmentMeta {
12681 : segment: Segment {
12682 : parent: None,
12683 : lsn: 0x10,
12684 : size: None,
12685 : needed: false,
12686 : },
12687 : timeline_id: TIMELINE_ID,
12688 : kind: BranchStart,
12689 : },
12690 : SegmentMeta {
12691 : segment: Segment {
12692 : parent: None,
12693 : lsn: 0x20,
12694 : size: None,
12695 : needed: false,
12696 : },
12697 : timeline_id: TIMELINE_ID,
12698 : kind: BranchPoint,
12699 : },
12700 : SegmentMeta {
12701 : segment: Segment {
12702 : parent: None,
12703 : lsn: 0x30,
12704 : size: None,
12705 : needed: false,
12706 : },
12707 : timeline_id: TIMELINE_ID,
12708 : kind: BranchPoint,
12709 : },
12710 : SegmentMeta {
12711 : segment: Segment {
12712 : parent: None,
12713 : lsn: 0x40,
12714 : size: None,
12715 : needed: false,
12716 : },
12717 : timeline_id: TIMELINE_ID,
12718 : kind: BranchPoint,
12719 : },
12720 : SegmentMeta {
12721 : segment: Segment {
12722 : parent: None,
12723 : lsn: 0x100,
12724 : size: None,
12725 : needed: false,
12726 : },
12727 : timeline_id: TIMELINE_ID,
12728 : kind: GcCutOff,
12729 : }, // we need to retain everything above the last branch point
12730 : SegmentMeta {
12731 : segment: Segment {
12732 : parent: None,
12733 : lsn: 0x100,
12734 : size: None,
12735 : needed: true,
12736 : },
12737 : timeline_id: TIMELINE_ID,
12738 : kind: BranchEnd,
12739 : },
12740 : ]
12741 : );
12742 :
12743 1 : main_tline
12744 1 : .remote_client
12745 1 : .schedule_index_upload_for_timeline_invisible_state(
12746 1 : TimelineVisibilityState::Invisible,
12747 0 : )?;
12748 1 : main_tline.remote_client.wait_completion().await?;
12749 1 : let inputs = gather_inputs(
12750 1 : &tenant,
12751 1 : &limit,
12752 1 : max_retention_period,
12753 1 : &mut logical_size_cache,
12754 1 : cause,
12755 1 : &cancel,
12756 1 : &ctx,
12757 : )
12758 1 : .instrument(info_span!(
12759 : "gather_inputs",
12760 : tenant_id = "unknown",
12761 : shard_id = "unknown",
12762 : ))
12763 1 : .await?;
12764 1 : let ModelInputs { mut segments, .. } = inputs;
12765 14 : segments.retain(|s| s.timeline_id == TIMELINE_ID);
12766 5 : for segment in segments.iter_mut() {
12767 5 : segment.segment.parent = None; // We don't care about the parent for the test
12768 5 : segment.segment.size = None; // We don't care about the size for the test
12769 5 : }
12770 1 : assert_eq!(
12771 : segments,
12772 : [
12773 : SegmentMeta {
12774 : segment: Segment {
12775 : parent: None,
12776 : lsn: 0x10,
12777 : size: None,
12778 : needed: false,
12779 : },
12780 : timeline_id: TIMELINE_ID,
12781 : kind: BranchStart,
12782 : },
12783 : SegmentMeta {
12784 : segment: Segment {
12785 : parent: None,
12786 : lsn: 0x20,
12787 : size: None,
12788 : needed: false,
12789 : },
12790 : timeline_id: TIMELINE_ID,
12791 : kind: BranchPoint,
12792 : },
12793 : SegmentMeta {
12794 : segment: Segment {
12795 : parent: None,
12796 : lsn: 0x30,
12797 : size: None,
12798 : needed: false,
12799 : },
12800 : timeline_id: TIMELINE_ID,
12801 : kind: BranchPoint,
12802 : },
12803 : SegmentMeta {
12804 : segment: Segment {
12805 : parent: None,
12806 : lsn: 0x40,
12807 : size: None,
12808 : needed: false,
12809 : },
12810 : timeline_id: TIMELINE_ID,
12811 : kind: BranchPoint,
12812 : },
12813 : SegmentMeta {
12814 : segment: Segment {
12815 : parent: None,
12816 : lsn: 0x40, // Branch end LSN == last branch point LSN
12817 : size: None,
12818 : needed: true,
12819 : },
12820 : timeline_id: TIMELINE_ID,
12821 : kind: BranchEnd,
12822 : },
12823 : ]
12824 : );
12825 :
12826 2 : Ok(())
12827 1 : }
12828 :
12829 : #[tokio::test]
12830 1 : async fn test_get_force_image_creation_lsn() -> anyhow::Result<()> {
12831 1 : let tenant_conf = pageserver_api::models::TenantConfig {
12832 1 : pitr_interval: Some(Duration::from_secs(7 * 3600)),
12833 1 : image_layer_force_creation_period: Some(Duration::from_secs(3600)),
12834 1 : ..Default::default()
12835 1 : };
12836 :
12837 1 : let tenant_id = TenantId::generate();
12838 :
12839 1 : let harness = TenantHarness::create_custom(
12840 1 : "test_get_force_image_creation_lsn",
12841 1 : tenant_conf,
12842 1 : tenant_id,
12843 1 : ShardIdentity::unsharded(),
12844 1 : Generation::new(1),
12845 1 : )
12846 1 : .await?;
12847 1 : let (tenant, ctx) = harness.load().await;
12848 1 : let timeline = tenant
12849 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
12850 1 : .await?;
12851 1 : timeline.gc_info.write().unwrap().cutoffs.time = Some(Lsn(100));
12852 : {
12853 1 : let writer = timeline.writer().await;
12854 1 : writer.finish_write(Lsn(5000));
12855 : }
12856 :
12857 1 : let image_creation_lsn = timeline.get_force_image_creation_lsn().unwrap();
12858 1 : assert_eq!(image_creation_lsn, Lsn(4300));
12859 2 : Ok(())
12860 1 : }
12861 : }
|
