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, LsnLease, 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 : pub(crate) use timeline::{LogicalSizeCalculationCause, PageReconstructError, Timeline};
146 :
147 : pub(crate) use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
148 : // re-export for use in walreceiver
149 : pub use crate::tenant::timeline::WalReceiverInfo;
150 :
151 : /// The "tenants" part of `tenants/<tenant>/timelines...`
152 : pub const TENANTS_SEGMENT_NAME: &str = "tenants";
153 :
154 : /// Parts of the `.neon/tenants/<tenant_id>/timelines/<timeline_id>` directory prefix.
155 : pub const TIMELINES_SEGMENT_NAME: &str = "timelines";
156 :
157 : /// References to shared objects that are passed into each tenant, such
158 : /// as the shared remote storage client and process initialization state.
159 : #[derive(Clone)]
160 : pub struct TenantSharedResources {
161 : pub broker_client: storage_broker::BrokerClientChannel,
162 : pub remote_storage: GenericRemoteStorage,
163 : pub deletion_queue_client: DeletionQueueClient,
164 : pub l0_flush_global_state: L0FlushGlobalState,
165 : pub basebackup_cache: Arc<BasebackupCache>,
166 : pub feature_resolver: FeatureResolver,
167 : }
168 :
169 : /// A [`TenantShard`] is really an _attached_ tenant. The configuration
170 : /// for an attached tenant is a subset of the [`LocationConf`], represented
171 : /// in this struct.
172 : #[derive(Clone)]
173 : pub(super) struct AttachedTenantConf {
174 : tenant_conf: pageserver_api::models::TenantConfig,
175 : location: AttachedLocationConfig,
176 : /// The deadline before which we are blocked from GC so that
177 : /// leases have a chance to be renewed.
178 : lsn_lease_deadline: Option<tokio::time::Instant>,
179 : }
180 :
181 : impl AttachedTenantConf {
182 118 : fn new(
183 118 : tenant_conf: pageserver_api::models::TenantConfig,
184 118 : location: AttachedLocationConfig,
185 118 : ) -> Self {
186 : // Sets a deadline before which we cannot proceed to GC due to lsn lease.
187 : //
188 : // We do this as the leases mapping are not persisted to disk. By delaying GC by lease
189 : // length, we guarantee that all the leases we granted before will have a chance to renew
190 : // when we run GC for the first time after restart / transition from AttachedMulti to AttachedSingle.
191 118 : let lsn_lease_deadline = if location.attach_mode == AttachmentMode::Single {
192 118 : Some(
193 118 : tokio::time::Instant::now()
194 118 : + tenant_conf
195 118 : .lsn_lease_length
196 118 : .unwrap_or(LsnLease::DEFAULT_LENGTH),
197 118 : )
198 : } else {
199 : // We don't use `lsn_lease_deadline` to delay GC in AttachedMulti and AttachedStale
200 : // because we don't do GC in these modes.
201 0 : None
202 : };
203 :
204 118 : Self {
205 118 : tenant_conf,
206 118 : location,
207 118 : lsn_lease_deadline,
208 118 : }
209 118 : }
210 :
211 118 : fn try_from(location_conf: LocationConf) -> anyhow::Result<Self> {
212 118 : match &location_conf.mode {
213 118 : LocationMode::Attached(attach_conf) => {
214 118 : Ok(Self::new(location_conf.tenant_conf, *attach_conf))
215 : }
216 : LocationMode::Secondary(_) => {
217 0 : anyhow::bail!(
218 0 : "Attempted to construct AttachedTenantConf from a LocationConf in secondary mode"
219 : )
220 : }
221 : }
222 118 : }
223 :
224 381 : fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
225 381 : self.lsn_lease_deadline
226 381 : .map(|d| tokio::time::Instant::now() < d)
227 381 : .unwrap_or(false)
228 381 : }
229 : }
230 : struct TimelinePreload {
231 : timeline_id: TimelineId,
232 : client: RemoteTimelineClient,
233 : index_part: Result<MaybeDeletedIndexPart, DownloadError>,
234 : previous_heatmap: Option<PreviousHeatmap>,
235 : }
236 :
237 : pub(crate) struct TenantPreload {
238 : /// The tenant manifest from remote storage, or None if no manifest was found.
239 : tenant_manifest: Option<TenantManifest>,
240 : /// Map from timeline ID to a possible timeline preload. It is None iff the timeline is offloaded according to the manifest.
241 : timelines: HashMap<TimelineId, Option<TimelinePreload>>,
242 : }
243 :
244 : /// When we spawn a tenant, there is a special mode for tenant creation that
245 : /// avoids trying to read anything from remote storage.
246 : pub(crate) enum SpawnMode {
247 : /// Activate as soon as possible
248 : Eager,
249 : /// Lazy activation in the background, with the option to skip the queue if the need comes up
250 : Lazy,
251 : }
252 :
253 : ///
254 : /// Tenant consists of multiple timelines. Keep them in a hash table.
255 : ///
256 : pub struct TenantShard {
257 : // Global pageserver config parameters
258 : pub conf: &'static PageServerConf,
259 :
260 : /// The value creation timestamp, used to measure activation delay, see:
261 : /// <https://github.com/neondatabase/neon/issues/4025>
262 : constructed_at: Instant,
263 :
264 : state: watch::Sender<TenantState>,
265 :
266 : // Overridden tenant-specific config parameters.
267 : // We keep pageserver_api::models::TenantConfig sturct here to preserve the information
268 : // about parameters that are not set.
269 : // This is necessary to allow global config updates.
270 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
271 :
272 : tenant_shard_id: TenantShardId,
273 :
274 : // The detailed sharding information, beyond the number/count in tenant_shard_id
275 : shard_identity: ShardIdentity,
276 :
277 : /// The remote storage generation, used to protect S3 objects from split-brain.
278 : /// Does not change over the lifetime of the [`TenantShard`] object.
279 : ///
280 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
281 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
282 : generation: Generation,
283 :
284 : timelines: Mutex<HashMap<TimelineId, Arc<Timeline>>>,
285 :
286 : /// During timeline creation, we first insert the TimelineId to the
287 : /// creating map, then `timelines`, then remove it from the creating map.
288 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
289 : timelines_creating: std::sync::Mutex<HashSet<TimelineId>>,
290 :
291 : /// Possibly offloaded and archived timelines
292 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
293 : timelines_offloaded: Mutex<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
294 :
295 : /// Tracks the timelines that are currently importing into this tenant shard.
296 : ///
297 : /// Note that importing timelines are also present in [`Self::timelines_creating`].
298 : /// Keep this in mind when ordering lock acquisition.
299 : ///
300 : /// Lifetime:
301 : /// * An imported timeline is created while scanning the bucket on tenant attach
302 : /// if the index part contains an `import_pgdata` entry and said field marks the import
303 : /// as in progress.
304 : /// * Imported timelines are removed when the storage controller calls the post timeline
305 : /// import activation endpoint.
306 : timelines_importing: std::sync::Mutex<HashMap<TimelineId, Arc<ImportingTimeline>>>,
307 :
308 : /// The last tenant manifest known to be in remote storage. None if the manifest has not yet
309 : /// been either downloaded or uploaded. Always Some after tenant attach.
310 : ///
311 : /// Initially populated during tenant attach, updated via `maybe_upload_tenant_manifest`.
312 : ///
313 : /// Do not modify this directly. It is used to check whether a new manifest needs to be
314 : /// uploaded. The manifest is constructed in `build_tenant_manifest`, and uploaded via
315 : /// `maybe_upload_tenant_manifest`.
316 : remote_tenant_manifest: tokio::sync::Mutex<Option<TenantManifest>>,
317 :
318 : // This mutex prevents creation of new timelines during GC.
319 : // Adding yet another mutex (in addition to `timelines`) is needed because holding
320 : // `timelines` mutex during all GC iteration
321 : // may block for a long time `get_timeline`, `get_timelines_state`,... and other operations
322 : // with timelines, which in turn may cause dropping replication connection, expiration of wait_for_lsn
323 : // timeout...
324 : gc_cs: tokio::sync::Mutex<()>,
325 : walredo_mgr: Option<Arc<WalRedoManager>>,
326 :
327 : /// Provides access to timeline data sitting in the remote storage.
328 : pub(crate) remote_storage: GenericRemoteStorage,
329 :
330 : /// Access to global deletion queue for when this tenant wants to schedule a deletion.
331 : deletion_queue_client: DeletionQueueClient,
332 :
333 : /// A channel to send async requests to prepare a basebackup for the basebackup cache.
334 : basebackup_cache: Arc<BasebackupCache>,
335 :
336 : /// Cached logical sizes updated updated on each [`TenantShard::gather_size_inputs`].
337 : cached_logical_sizes: tokio::sync::Mutex<HashMap<(TimelineId, Lsn), u64>>,
338 : cached_synthetic_tenant_size: Arc<AtomicU64>,
339 :
340 : eviction_task_tenant_state: tokio::sync::Mutex<EvictionTaskTenantState>,
341 :
342 : /// Track repeated failures to compact, so that we can back off.
343 : /// Overhead of mutex is acceptable because compaction is done with a multi-second period.
344 : compaction_circuit_breaker: std::sync::Mutex<CircuitBreaker>,
345 :
346 : /// Signals the tenant compaction loop that there is L0 compaction work to be done.
347 : pub(crate) l0_compaction_trigger: Arc<Notify>,
348 :
349 : /// Scheduled gc-compaction tasks.
350 : scheduled_compaction_tasks: std::sync::Mutex<HashMap<TimelineId, Arc<GcCompactionQueue>>>,
351 :
352 : /// If the tenant is in Activating state, notify this to encourage it
353 : /// to proceed to Active as soon as possible, rather than waiting for lazy
354 : /// background warmup.
355 : pub(crate) activate_now_sem: tokio::sync::Semaphore,
356 :
357 : /// Time it took for the tenant to activate. Zero if not active yet.
358 : attach_wal_lag_cooldown: Arc<std::sync::OnceLock<WalLagCooldown>>,
359 :
360 : // Cancellation token fires when we have entered shutdown(). This is a parent of
361 : // Timelines' cancellation token.
362 : pub(crate) cancel: CancellationToken,
363 :
364 : // Users of the TenantShard such as the page service must take this Gate to avoid
365 : // trying to use a TenantShard which is shutting down.
366 : pub(crate) gate: Gate,
367 :
368 : /// Throttle applied at the top of [`Timeline::get`].
369 : /// All [`TenantShard::timelines`] of a given [`TenantShard`] instance share the same [`throttle::Throttle`] instance.
370 : pub(crate) pagestream_throttle: Arc<throttle::Throttle>,
371 :
372 : pub(crate) pagestream_throttle_metrics: Arc<crate::metrics::tenant_throttling::Pagestream>,
373 :
374 : /// An ongoing timeline detach concurrency limiter.
375 : ///
376 : /// As a tenant will likely be restarted as part of timeline detach ancestor it makes no sense
377 : /// to have two running at the same time. A different one can be started if an earlier one
378 : /// has failed for whatever reason.
379 : ongoing_timeline_detach: std::sync::Mutex<Option<(TimelineId, utils::completion::Barrier)>>,
380 :
381 : /// `index_part.json` based gc blocking reason tracking.
382 : ///
383 : /// New gc iterations must start a new iteration by acquiring `GcBlock::start` before
384 : /// proceeding.
385 : pub(crate) gc_block: gc_block::GcBlock,
386 :
387 : l0_flush_global_state: L0FlushGlobalState,
388 :
389 : pub(crate) feature_resolver: TenantFeatureResolver,
390 : }
391 : impl std::fmt::Debug for TenantShard {
392 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
393 0 : write!(f, "{} ({})", self.tenant_shard_id, self.current_state())
394 0 : }
395 : }
396 :
397 : pub(crate) enum WalRedoManager {
398 : Prod(WalredoManagerId, PostgresRedoManager),
399 : #[cfg(test)]
400 : Test(harness::TestRedoManager),
401 : }
402 :
403 : #[derive(thiserror::Error, Debug)]
404 : #[error("pageserver is shutting down")]
405 : pub(crate) struct GlobalShutDown;
406 :
407 : impl WalRedoManager {
408 0 : pub(crate) fn new(mgr: PostgresRedoManager) -> Result<Arc<Self>, GlobalShutDown> {
409 0 : let id = WalredoManagerId::next();
410 0 : let arc = Arc::new(Self::Prod(id, mgr));
411 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
412 0 : match &mut *guard {
413 0 : Some(map) => {
414 0 : map.insert(id, Arc::downgrade(&arc));
415 0 : Ok(arc)
416 : }
417 0 : None => Err(GlobalShutDown),
418 : }
419 0 : }
420 : }
421 :
422 : impl Drop for WalRedoManager {
423 5 : fn drop(&mut self) {
424 5 : match self {
425 0 : Self::Prod(id, _) => {
426 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
427 0 : if let Some(map) = &mut *guard {
428 0 : map.remove(id).expect("new() registers, drop() unregisters");
429 0 : }
430 : }
431 : #[cfg(test)]
432 5 : Self::Test(_) => {
433 5 : // Not applicable to test redo manager
434 5 : }
435 : }
436 5 : }
437 : }
438 :
439 : /// Global registry of all walredo managers so that [`crate::shutdown_pageserver`] can shut down
440 : /// the walredo processes outside of the regular order.
441 : ///
442 : /// This is necessary to work around a systemd bug where it freezes if there are
443 : /// walredo processes left => <https://github.com/neondatabase/cloud/issues/11387>
444 : #[allow(clippy::type_complexity)]
445 : pub(crate) static WALREDO_MANAGERS: once_cell::sync::Lazy<
446 : Mutex<Option<HashMap<WalredoManagerId, Weak<WalRedoManager>>>>,
447 0 : > = once_cell::sync::Lazy::new(|| Mutex::new(Some(HashMap::new())));
448 : #[derive(PartialEq, Eq, Hash, Clone, Copy, Debug)]
449 : pub(crate) struct WalredoManagerId(u64);
450 : impl WalredoManagerId {
451 0 : pub fn next() -> Self {
452 : static NEXT: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(1);
453 0 : let id = NEXT.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
454 0 : if id == 0 {
455 0 : panic!(
456 0 : "WalredoManagerId::new() returned 0, indicating wraparound, risking it's no longer unique"
457 : );
458 0 : }
459 0 : Self(id)
460 0 : }
461 : }
462 :
463 : #[cfg(test)]
464 : impl From<harness::TestRedoManager> for WalRedoManager {
465 118 : fn from(mgr: harness::TestRedoManager) -> Self {
466 118 : Self::Test(mgr)
467 118 : }
468 : }
469 :
470 : impl WalRedoManager {
471 3 : pub(crate) async fn shutdown(&self) -> bool {
472 3 : match self {
473 0 : Self::Prod(_, mgr) => mgr.shutdown().await,
474 : #[cfg(test)]
475 : Self::Test(_) => {
476 : // Not applicable to test redo manager
477 3 : true
478 : }
479 : }
480 3 : }
481 :
482 0 : pub(crate) fn maybe_quiesce(&self, idle_timeout: Duration) {
483 0 : match self {
484 0 : Self::Prod(_, mgr) => mgr.maybe_quiesce(idle_timeout),
485 : #[cfg(test)]
486 0 : Self::Test(_) => {
487 0 : // Not applicable to test redo manager
488 0 : }
489 : }
490 0 : }
491 :
492 : /// # Cancel-Safety
493 : ///
494 : /// This method is cancellation-safe.
495 26774 : pub async fn request_redo(
496 26774 : &self,
497 26774 : key: pageserver_api::key::Key,
498 26774 : lsn: Lsn,
499 26774 : base_img: Option<(Lsn, bytes::Bytes)>,
500 26774 : records: Vec<(Lsn, wal_decoder::models::record::NeonWalRecord)>,
501 26774 : pg_version: PgMajorVersion,
502 26774 : redo_attempt_type: RedoAttemptType,
503 26774 : ) -> Result<bytes::Bytes, walredo::Error> {
504 26774 : match self {
505 0 : Self::Prod(_, mgr) => {
506 0 : mgr.request_redo(key, lsn, base_img, records, pg_version, redo_attempt_type)
507 0 : .await
508 : }
509 : #[cfg(test)]
510 26774 : Self::Test(mgr) => {
511 26774 : mgr.request_redo(key, lsn, base_img, records, pg_version, redo_attempt_type)
512 26774 : .await
513 : }
514 : }
515 26774 : }
516 :
517 0 : pub(crate) fn status(&self) -> Option<WalRedoManagerStatus> {
518 0 : match self {
519 0 : WalRedoManager::Prod(_, m) => Some(m.status()),
520 : #[cfg(test)]
521 0 : WalRedoManager::Test(_) => None,
522 : }
523 0 : }
524 : }
525 :
526 : /// A very lightweight memory representation of an offloaded timeline.
527 : ///
528 : /// We need to store the list of offloaded timelines so that we can perform operations on them,
529 : /// like unoffloading them, or (at a later date), decide to perform flattening.
530 : /// This type has a much smaller memory impact than [`Timeline`], and thus we can store many
531 : /// more offloaded timelines than we can manage ones that aren't.
532 : pub struct OffloadedTimeline {
533 : pub tenant_shard_id: TenantShardId,
534 : pub timeline_id: TimelineId,
535 : pub ancestor_timeline_id: Option<TimelineId>,
536 : /// Whether to retain the branch lsn at the ancestor or not
537 : pub ancestor_retain_lsn: Option<Lsn>,
538 :
539 : /// When the timeline was archived.
540 : ///
541 : /// Present for future flattening deliberations.
542 : pub archived_at: NaiveDateTime,
543 :
544 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
545 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
546 : pub delete_progress: TimelineDeleteProgress,
547 :
548 : /// Part of the `OffloadedTimeline` object's lifecycle: this needs to be set before we drop it
549 : pub deleted_from_ancestor: AtomicBool,
550 :
551 : _metrics_guard: OffloadedTimelineMetricsGuard,
552 : }
553 :
554 : /// Increases the offloaded timeline count metric when created, and decreases when dropped.
555 : struct OffloadedTimelineMetricsGuard;
556 :
557 : impl OffloadedTimelineMetricsGuard {
558 1 : fn new() -> Self {
559 1 : TIMELINE_STATE_METRIC
560 1 : .with_label_values(&["offloaded"])
561 1 : .inc();
562 1 : Self
563 1 : }
564 : }
565 :
566 : impl Drop for OffloadedTimelineMetricsGuard {
567 1 : fn drop(&mut self) {
568 1 : TIMELINE_STATE_METRIC
569 1 : .with_label_values(&["offloaded"])
570 1 : .dec();
571 1 : }
572 : }
573 :
574 : impl OffloadedTimeline {
575 : /// Obtains an offloaded timeline from a given timeline object.
576 : ///
577 : /// Returns `None` if the `archived_at` flag couldn't be obtained, i.e.
578 : /// the timeline is not in a stopped state.
579 : /// Panics if the timeline is not archived.
580 1 : fn from_timeline(timeline: &Timeline) -> Result<Self, UploadQueueNotReadyError> {
581 1 : let (ancestor_retain_lsn, ancestor_timeline_id) =
582 1 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
583 1 : let ancestor_lsn = timeline.get_ancestor_lsn();
584 1 : let ancestor_timeline_id = ancestor_timeline.timeline_id;
585 1 : let mut gc_info = ancestor_timeline.gc_info.write().unwrap();
586 1 : gc_info.insert_child(timeline.timeline_id, ancestor_lsn, MaybeOffloaded::Yes);
587 1 : (Some(ancestor_lsn), Some(ancestor_timeline_id))
588 : } else {
589 0 : (None, None)
590 : };
591 1 : let archived_at = timeline
592 1 : .remote_client
593 1 : .archived_at_stopped_queue()?
594 1 : .expect("must be called on an archived timeline");
595 1 : Ok(Self {
596 1 : tenant_shard_id: timeline.tenant_shard_id,
597 1 : timeline_id: timeline.timeline_id,
598 1 : ancestor_timeline_id,
599 1 : ancestor_retain_lsn,
600 1 : archived_at,
601 1 :
602 1 : delete_progress: timeline.delete_progress.clone(),
603 1 : deleted_from_ancestor: AtomicBool::new(false),
604 1 :
605 1 : _metrics_guard: OffloadedTimelineMetricsGuard::new(),
606 1 : })
607 1 : }
608 0 : fn from_manifest(tenant_shard_id: TenantShardId, manifest: &OffloadedTimelineManifest) -> Self {
609 : // We expect to reach this case in tenant loading, where the `retain_lsn` is populated in the parent's `gc_info`
610 : // by the `initialize_gc_info` function.
611 : let OffloadedTimelineManifest {
612 0 : timeline_id,
613 0 : ancestor_timeline_id,
614 0 : ancestor_retain_lsn,
615 0 : archived_at,
616 0 : } = *manifest;
617 0 : Self {
618 0 : tenant_shard_id,
619 0 : timeline_id,
620 0 : ancestor_timeline_id,
621 0 : ancestor_retain_lsn,
622 0 : archived_at,
623 0 : delete_progress: TimelineDeleteProgress::default(),
624 0 : deleted_from_ancestor: AtomicBool::new(false),
625 0 : _metrics_guard: OffloadedTimelineMetricsGuard::new(),
626 0 : }
627 0 : }
628 1 : fn manifest(&self) -> OffloadedTimelineManifest {
629 : let Self {
630 1 : timeline_id,
631 1 : ancestor_timeline_id,
632 1 : ancestor_retain_lsn,
633 1 : archived_at,
634 : ..
635 1 : } = self;
636 1 : OffloadedTimelineManifest {
637 1 : timeline_id: *timeline_id,
638 1 : ancestor_timeline_id: *ancestor_timeline_id,
639 1 : ancestor_retain_lsn: *ancestor_retain_lsn,
640 1 : archived_at: *archived_at,
641 1 : }
642 1 : }
643 : /// Delete this timeline's retain_lsn from its ancestor, if present in the given tenant
644 0 : fn delete_from_ancestor_with_timelines(
645 0 : &self,
646 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
647 0 : ) {
648 0 : if let (Some(_retain_lsn), Some(ancestor_timeline_id)) =
649 0 : (self.ancestor_retain_lsn, self.ancestor_timeline_id)
650 : {
651 0 : if let Some((_, ancestor_timeline)) = timelines
652 0 : .iter()
653 0 : .find(|(tid, _tl)| **tid == ancestor_timeline_id)
654 : {
655 0 : let removal_happened = ancestor_timeline
656 0 : .gc_info
657 0 : .write()
658 0 : .unwrap()
659 0 : .remove_child_offloaded(self.timeline_id);
660 0 : if !removal_happened {
661 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
662 0 : "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
663 0 : }
664 0 : }
665 0 : }
666 0 : self.deleted_from_ancestor.store(true, Ordering::Release);
667 0 : }
668 : /// Call [`Self::delete_from_ancestor_with_timelines`] instead if possible.
669 : ///
670 : /// As the entire tenant is being dropped, don't bother deregistering the `retain_lsn` from the ancestor.
671 1 : fn defuse_for_tenant_drop(&self) {
672 1 : self.deleted_from_ancestor.store(true, Ordering::Release);
673 1 : }
674 : }
675 :
676 : impl fmt::Debug for OffloadedTimeline {
677 0 : fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
678 0 : write!(f, "OffloadedTimeline<{}>", self.timeline_id)
679 0 : }
680 : }
681 :
682 : impl Drop for OffloadedTimeline {
683 1 : fn drop(&mut self) {
684 1 : if !self.deleted_from_ancestor.load(Ordering::Acquire) {
685 0 : tracing::warn!(
686 0 : "offloaded timeline {} was dropped without having cleaned it up at the ancestor",
687 : self.timeline_id
688 : );
689 1 : }
690 1 : }
691 : }
692 :
693 : #[derive(Copy, Clone, PartialEq, Eq, Hash, Debug)]
694 : pub enum MaybeOffloaded {
695 : Yes,
696 : No,
697 : }
698 :
699 : #[derive(Clone, Debug)]
700 : pub enum TimelineOrOffloaded {
701 : Timeline(Arc<Timeline>),
702 : Offloaded(Arc<OffloadedTimeline>),
703 : Importing(Arc<ImportingTimeline>),
704 : }
705 :
706 : impl TimelineOrOffloaded {
707 0 : pub fn arc_ref(&self) -> TimelineOrOffloadedArcRef<'_> {
708 0 : match self {
709 0 : TimelineOrOffloaded::Timeline(timeline) => {
710 0 : TimelineOrOffloadedArcRef::Timeline(timeline)
711 : }
712 0 : TimelineOrOffloaded::Offloaded(offloaded) => {
713 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded)
714 : }
715 0 : TimelineOrOffloaded::Importing(importing) => {
716 0 : TimelineOrOffloadedArcRef::Importing(importing)
717 : }
718 : }
719 0 : }
720 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
721 0 : self.arc_ref().tenant_shard_id()
722 0 : }
723 0 : pub fn timeline_id(&self) -> TimelineId {
724 0 : self.arc_ref().timeline_id()
725 0 : }
726 1 : pub fn delete_progress(&self) -> &Arc<tokio::sync::Mutex<DeleteTimelineFlow>> {
727 1 : match self {
728 1 : TimelineOrOffloaded::Timeline(timeline) => &timeline.delete_progress,
729 0 : TimelineOrOffloaded::Offloaded(offloaded) => &offloaded.delete_progress,
730 0 : TimelineOrOffloaded::Importing(importing) => &importing.delete_progress,
731 : }
732 1 : }
733 0 : fn maybe_remote_client(&self) -> Option<Arc<RemoteTimelineClient>> {
734 0 : match self {
735 0 : TimelineOrOffloaded::Timeline(timeline) => Some(timeline.remote_client.clone()),
736 0 : TimelineOrOffloaded::Offloaded(_offloaded) => None,
737 0 : TimelineOrOffloaded::Importing(importing) => {
738 0 : Some(importing.timeline.remote_client.clone())
739 : }
740 : }
741 0 : }
742 : }
743 :
744 : pub enum TimelineOrOffloadedArcRef<'a> {
745 : Timeline(&'a Arc<Timeline>),
746 : Offloaded(&'a Arc<OffloadedTimeline>),
747 : Importing(&'a Arc<ImportingTimeline>),
748 : }
749 :
750 : impl TimelineOrOffloadedArcRef<'_> {
751 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
752 0 : match self {
753 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.tenant_shard_id,
754 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.tenant_shard_id,
755 0 : TimelineOrOffloadedArcRef::Importing(importing) => importing.timeline.tenant_shard_id,
756 : }
757 0 : }
758 0 : pub fn timeline_id(&self) -> TimelineId {
759 0 : match self {
760 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.timeline_id,
761 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.timeline_id,
762 0 : TimelineOrOffloadedArcRef::Importing(importing) => importing.timeline.timeline_id,
763 : }
764 0 : }
765 : }
766 :
767 : impl<'a> From<&'a Arc<Timeline>> for TimelineOrOffloadedArcRef<'a> {
768 0 : fn from(timeline: &'a Arc<Timeline>) -> Self {
769 0 : Self::Timeline(timeline)
770 0 : }
771 : }
772 :
773 : impl<'a> From<&'a Arc<OffloadedTimeline>> for TimelineOrOffloadedArcRef<'a> {
774 0 : fn from(timeline: &'a Arc<OffloadedTimeline>) -> Self {
775 0 : Self::Offloaded(timeline)
776 0 : }
777 : }
778 :
779 : impl<'a> From<&'a Arc<ImportingTimeline>> for TimelineOrOffloadedArcRef<'a> {
780 0 : fn from(timeline: &'a Arc<ImportingTimeline>) -> Self {
781 0 : Self::Importing(timeline)
782 0 : }
783 : }
784 :
785 : #[derive(Debug, thiserror::Error, PartialEq, Eq)]
786 : pub enum GetTimelineError {
787 : #[error("Timeline is shutting down")]
788 : ShuttingDown,
789 : #[error("Timeline {tenant_id}/{timeline_id} is not active, state: {state:?}")]
790 : NotActive {
791 : tenant_id: TenantShardId,
792 : timeline_id: TimelineId,
793 : state: TimelineState,
794 : },
795 : #[error("Timeline {tenant_id}/{timeline_id} was not found")]
796 : NotFound {
797 : tenant_id: TenantShardId,
798 : timeline_id: TimelineId,
799 : },
800 : }
801 :
802 : #[derive(Debug, thiserror::Error)]
803 : pub enum LoadLocalTimelineError {
804 : #[error("FailedToLoad")]
805 : Load(#[source] anyhow::Error),
806 : #[error("FailedToResumeDeletion")]
807 : ResumeDeletion(#[source] anyhow::Error),
808 : }
809 :
810 : #[derive(thiserror::Error)]
811 : pub enum DeleteTimelineError {
812 : #[error("NotFound")]
813 : NotFound,
814 :
815 : #[error("HasChildren")]
816 : HasChildren(Vec<TimelineId>),
817 :
818 : #[error("Timeline deletion is already in progress")]
819 : AlreadyInProgress(Arc<tokio::sync::Mutex<DeleteTimelineFlow>>),
820 :
821 : #[error("Cancelled")]
822 : Cancelled,
823 :
824 : #[error(transparent)]
825 : Other(#[from] anyhow::Error),
826 : }
827 :
828 : impl Debug for DeleteTimelineError {
829 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
830 0 : match self {
831 0 : Self::NotFound => write!(f, "NotFound"),
832 0 : Self::HasChildren(c) => f.debug_tuple("HasChildren").field(c).finish(),
833 0 : Self::AlreadyInProgress(_) => f.debug_tuple("AlreadyInProgress").finish(),
834 0 : Self::Cancelled => f.debug_tuple("Cancelled").finish(),
835 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
836 : }
837 0 : }
838 : }
839 :
840 : #[derive(thiserror::Error)]
841 : pub enum TimelineArchivalError {
842 : #[error("NotFound")]
843 : NotFound,
844 :
845 : #[error("Timeout")]
846 : Timeout,
847 :
848 : #[error("Cancelled")]
849 : Cancelled,
850 :
851 : #[error("ancestor is archived: {}", .0)]
852 : HasArchivedParent(TimelineId),
853 :
854 : #[error("HasUnarchivedChildren")]
855 : HasUnarchivedChildren(Vec<TimelineId>),
856 :
857 : #[error("Timeline archival is already in progress")]
858 : AlreadyInProgress,
859 :
860 : #[error(transparent)]
861 : Other(anyhow::Error),
862 : }
863 :
864 : #[derive(thiserror::Error, Debug)]
865 : pub(crate) enum TenantManifestError {
866 : #[error("Remote storage error: {0}")]
867 : RemoteStorage(anyhow::Error),
868 :
869 : #[error("Cancelled")]
870 : Cancelled,
871 : }
872 :
873 : impl From<TenantManifestError> for TimelineArchivalError {
874 0 : fn from(e: TenantManifestError) -> Self {
875 0 : match e {
876 0 : TenantManifestError::RemoteStorage(e) => Self::Other(e),
877 0 : TenantManifestError::Cancelled => Self::Cancelled,
878 : }
879 0 : }
880 : }
881 :
882 : impl Debug for TimelineArchivalError {
883 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
884 0 : match self {
885 0 : Self::NotFound => write!(f, "NotFound"),
886 0 : Self::Timeout => write!(f, "Timeout"),
887 0 : Self::Cancelled => write!(f, "Cancelled"),
888 0 : Self::HasArchivedParent(p) => f.debug_tuple("HasArchivedParent").field(p).finish(),
889 0 : Self::HasUnarchivedChildren(c) => {
890 0 : f.debug_tuple("HasUnarchivedChildren").field(c).finish()
891 : }
892 0 : Self::AlreadyInProgress => f.debug_tuple("AlreadyInProgress").finish(),
893 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
894 : }
895 0 : }
896 : }
897 :
898 : pub enum SetStoppingError {
899 : AlreadyStopping(completion::Barrier),
900 : Broken,
901 : }
902 :
903 : impl Debug for SetStoppingError {
904 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
905 0 : match self {
906 0 : Self::AlreadyStopping(_) => f.debug_tuple("AlreadyStopping").finish(),
907 0 : Self::Broken => write!(f, "Broken"),
908 : }
909 0 : }
910 : }
911 :
912 : #[derive(thiserror::Error, Debug)]
913 : pub(crate) enum FinalizeTimelineImportError {
914 : #[error("Import task not done yet")]
915 : ImportTaskStillRunning,
916 : #[error("Shutting down")]
917 : ShuttingDown,
918 : }
919 :
920 : /// Arguments to [`TenantShard::create_timeline`].
921 : ///
922 : /// Not usable as an idempotency key for timeline creation because if [`CreateTimelineParamsBranch::ancestor_start_lsn`]
923 : /// is `None`, the result of the timeline create call is not deterministic.
924 : ///
925 : /// See [`CreateTimelineIdempotency`] for an idempotency key.
926 : #[derive(Debug)]
927 : pub(crate) enum CreateTimelineParams {
928 : Bootstrap(CreateTimelineParamsBootstrap),
929 : Branch(CreateTimelineParamsBranch),
930 : ImportPgdata(CreateTimelineParamsImportPgdata),
931 : }
932 :
933 : #[derive(Debug)]
934 : pub(crate) struct CreateTimelineParamsBootstrap {
935 : pub(crate) new_timeline_id: TimelineId,
936 : pub(crate) existing_initdb_timeline_id: Option<TimelineId>,
937 : pub(crate) pg_version: PgMajorVersion,
938 : }
939 :
940 : /// NB: See comment on [`CreateTimelineIdempotency::Branch`] for why there's no `pg_version` here.
941 : #[derive(Debug)]
942 : pub(crate) struct CreateTimelineParamsBranch {
943 : pub(crate) new_timeline_id: TimelineId,
944 : pub(crate) ancestor_timeline_id: TimelineId,
945 : pub(crate) ancestor_start_lsn: Option<Lsn>,
946 : }
947 :
948 : #[derive(Debug)]
949 : pub(crate) struct CreateTimelineParamsImportPgdata {
950 : pub(crate) new_timeline_id: TimelineId,
951 : pub(crate) location: import_pgdata::index_part_format::Location,
952 : pub(crate) idempotency_key: import_pgdata::index_part_format::IdempotencyKey,
953 : }
954 :
955 : /// What is used to determine idempotency of a [`TenantShard::create_timeline`] call in [`TenantShard::start_creating_timeline`] in [`TenantShard::start_creating_timeline`].
956 : ///
957 : /// Each [`Timeline`] object holds [`Self`] as an immutable property in [`Timeline::create_idempotency`].
958 : ///
959 : /// We lower timeline creation requests to [`Self`], and then use [`PartialEq::eq`] to compare [`Timeline::create_idempotency`] with the request.
960 : /// If they are equal, we return a reference to the existing timeline, otherwise it's an idempotency conflict.
961 : ///
962 : /// There is special treatment for [`Self::FailWithConflict`] to always return an idempotency conflict.
963 : /// It would be nice to have more advanced derive macros to make that special treatment declarative.
964 : ///
965 : /// Notes:
966 : /// - Unlike [`CreateTimelineParams`], ancestor LSN is fixed, so, branching will be at a deterministic LSN.
967 : /// - We make some trade-offs though, e.g., [`CreateTimelineParamsBootstrap::existing_initdb_timeline_id`]
968 : /// is not considered for idempotency. We can improve on this over time if we deem it necessary.
969 : ///
970 : #[derive(Debug, Clone, PartialEq, Eq)]
971 : pub(crate) enum CreateTimelineIdempotency {
972 : /// NB: special treatment, see comment in [`Self`].
973 : FailWithConflict,
974 : Bootstrap {
975 : pg_version: PgMajorVersion,
976 : },
977 : /// NB: branches always have the same `pg_version` as their ancestor.
978 : /// While [`pageserver_api::models::TimelineCreateRequestMode::Branch::pg_version`]
979 : /// exists as a field, and is set by cplane, it has always been ignored by pageserver when
980 : /// determining the child branch pg_version.
981 : Branch {
982 : ancestor_timeline_id: TimelineId,
983 : ancestor_start_lsn: Lsn,
984 : },
985 : ImportPgdata(CreatingTimelineIdempotencyImportPgdata),
986 : }
987 :
988 : #[derive(Debug, Clone, PartialEq, Eq)]
989 : pub(crate) struct CreatingTimelineIdempotencyImportPgdata {
990 : idempotency_key: import_pgdata::index_part_format::IdempotencyKey,
991 : }
992 :
993 : /// What is returned by [`TenantShard::start_creating_timeline`].
994 : #[must_use]
995 : enum StartCreatingTimelineResult {
996 : CreateGuard(TimelineCreateGuard),
997 : Idempotent(Arc<Timeline>),
998 : }
999 :
1000 : #[allow(clippy::large_enum_variant, reason = "TODO")]
1001 : enum TimelineInitAndSyncResult {
1002 : ReadyToActivate,
1003 : NeedsSpawnImportPgdata(TimelineInitAndSyncNeedsSpawnImportPgdata),
1004 : }
1005 :
1006 : #[must_use]
1007 : struct TimelineInitAndSyncNeedsSpawnImportPgdata {
1008 : timeline: Arc<Timeline>,
1009 : import_pgdata: import_pgdata::index_part_format::Root,
1010 : guard: TimelineCreateGuard,
1011 : }
1012 :
1013 : /// What is returned by [`TenantShard::create_timeline`].
1014 : enum CreateTimelineResult {
1015 : Created(Arc<Timeline>),
1016 : Idempotent(Arc<Timeline>),
1017 : /// IMPORTANT: This [`Arc<Timeline>`] object is not in [`TenantShard::timelines`] when
1018 : /// we return this result, nor will this concrete object ever be added there.
1019 : /// Cf method comment on [`TenantShard::create_timeline_import_pgdata`].
1020 : ImportSpawned(Arc<Timeline>),
1021 : }
1022 :
1023 : impl CreateTimelineResult {
1024 0 : fn discriminant(&self) -> &'static str {
1025 0 : match self {
1026 0 : Self::Created(_) => "Created",
1027 0 : Self::Idempotent(_) => "Idempotent",
1028 0 : Self::ImportSpawned(_) => "ImportSpawned",
1029 : }
1030 0 : }
1031 0 : fn timeline(&self) -> &Arc<Timeline> {
1032 0 : match self {
1033 0 : Self::Created(t) | Self::Idempotent(t) | Self::ImportSpawned(t) => t,
1034 : }
1035 0 : }
1036 : /// Unit test timelines aren't activated, test has to do it if it needs to.
1037 : #[cfg(test)]
1038 118 : fn into_timeline_for_test(self) -> Arc<Timeline> {
1039 118 : match self {
1040 118 : Self::Created(t) | Self::Idempotent(t) | Self::ImportSpawned(t) => t,
1041 : }
1042 118 : }
1043 : }
1044 :
1045 : #[derive(thiserror::Error, Debug)]
1046 : pub enum CreateTimelineError {
1047 : #[error("creation of timeline with the given ID is in progress")]
1048 : AlreadyCreating,
1049 : #[error("timeline already exists with different parameters")]
1050 : Conflict,
1051 : #[error(transparent)]
1052 : AncestorLsn(anyhow::Error),
1053 : #[error("ancestor timeline is not active")]
1054 : AncestorNotActive,
1055 : #[error("ancestor timeline is archived")]
1056 : AncestorArchived,
1057 : #[error("tenant shutting down")]
1058 : ShuttingDown,
1059 : #[error(transparent)]
1060 : Other(#[from] anyhow::Error),
1061 : }
1062 :
1063 : #[derive(thiserror::Error, Debug)]
1064 : pub enum InitdbError {
1065 : #[error("Operation was cancelled")]
1066 : Cancelled,
1067 : #[error(transparent)]
1068 : Other(anyhow::Error),
1069 : #[error(transparent)]
1070 : Inner(postgres_initdb::Error),
1071 : }
1072 :
1073 : enum CreateTimelineCause {
1074 : Load,
1075 : Delete,
1076 : }
1077 :
1078 : #[allow(clippy::large_enum_variant, reason = "TODO")]
1079 : enum LoadTimelineCause {
1080 : Attach,
1081 : Unoffload,
1082 : }
1083 :
1084 : #[derive(thiserror::Error, Debug)]
1085 : pub(crate) enum GcError {
1086 : // The tenant is shutting down
1087 : #[error("tenant shutting down")]
1088 : TenantCancelled,
1089 :
1090 : // The tenant is shutting down
1091 : #[error("timeline shutting down")]
1092 : TimelineCancelled,
1093 :
1094 : // The tenant is in a state inelegible to run GC
1095 : #[error("not active")]
1096 : NotActive,
1097 :
1098 : // A requested GC cutoff LSN was invalid, for example it tried to move backwards
1099 : #[error("not active")]
1100 : BadLsn { why: String },
1101 :
1102 : // A remote storage error while scheduling updates after compaction
1103 : #[error(transparent)]
1104 : Remote(anyhow::Error),
1105 :
1106 : // An error reading while calculating GC cutoffs
1107 : #[error(transparent)]
1108 : GcCutoffs(PageReconstructError),
1109 :
1110 : // If GC was invoked for a particular timeline, this error means it didn't exist
1111 : #[error("timeline not found")]
1112 : TimelineNotFound,
1113 : }
1114 :
1115 : impl From<PageReconstructError> for GcError {
1116 0 : fn from(value: PageReconstructError) -> Self {
1117 0 : match value {
1118 0 : PageReconstructError::Cancelled => Self::TimelineCancelled,
1119 0 : other => Self::GcCutoffs(other),
1120 : }
1121 0 : }
1122 : }
1123 :
1124 : impl From<NotInitialized> for GcError {
1125 0 : fn from(value: NotInitialized) -> Self {
1126 0 : match value {
1127 0 : NotInitialized::Uninitialized => GcError::Remote(value.into()),
1128 0 : NotInitialized::Stopped | NotInitialized::ShuttingDown => GcError::TimelineCancelled,
1129 : }
1130 0 : }
1131 : }
1132 :
1133 : impl From<timeline::layer_manager::Shutdown> for GcError {
1134 0 : fn from(_: timeline::layer_manager::Shutdown) -> Self {
1135 0 : GcError::TimelineCancelled
1136 0 : }
1137 : }
1138 :
1139 : #[derive(thiserror::Error, Debug)]
1140 : pub(crate) enum LoadConfigError {
1141 : #[error("TOML deserialization error: '{0}'")]
1142 : DeserializeToml(#[from] toml_edit::de::Error),
1143 :
1144 : #[error("Config not found at {0}")]
1145 : NotFound(Utf8PathBuf),
1146 : }
1147 :
1148 : impl TenantShard {
1149 : /// Yet another helper for timeline initialization.
1150 : ///
1151 : /// - Initializes the Timeline struct and inserts it into the tenant's hash map
1152 : /// - Scans the local timeline directory for layer files and builds the layer map
1153 : /// - Downloads remote index file and adds remote files to the layer map
1154 : /// - Schedules remote upload tasks for any files that are present locally but missing from remote storage.
1155 : ///
1156 : /// If the operation fails, the timeline is left in the tenant's hash map in Broken state. On success,
1157 : /// it is marked as Active.
1158 : #[allow(clippy::too_many_arguments)]
1159 3 : async fn timeline_init_and_sync(
1160 3 : self: &Arc<Self>,
1161 3 : timeline_id: TimelineId,
1162 3 : resources: TimelineResources,
1163 3 : index_part: IndexPart,
1164 3 : metadata: TimelineMetadata,
1165 3 : previous_heatmap: Option<PreviousHeatmap>,
1166 3 : ancestor: Option<Arc<Timeline>>,
1167 3 : cause: LoadTimelineCause,
1168 3 : ctx: &RequestContext,
1169 3 : ) -> anyhow::Result<TimelineInitAndSyncResult> {
1170 3 : let tenant_id = self.tenant_shard_id;
1171 :
1172 3 : let import_pgdata = index_part.import_pgdata.clone();
1173 3 : let idempotency = match &import_pgdata {
1174 0 : Some(import_pgdata) => {
1175 0 : CreateTimelineIdempotency::ImportPgdata(CreatingTimelineIdempotencyImportPgdata {
1176 0 : idempotency_key: import_pgdata.idempotency_key().clone(),
1177 0 : })
1178 : }
1179 : None => {
1180 3 : if metadata.ancestor_timeline().is_none() {
1181 2 : CreateTimelineIdempotency::Bootstrap {
1182 2 : pg_version: metadata.pg_version(),
1183 2 : }
1184 : } else {
1185 1 : CreateTimelineIdempotency::Branch {
1186 1 : ancestor_timeline_id: metadata.ancestor_timeline().unwrap(),
1187 1 : ancestor_start_lsn: metadata.ancestor_lsn(),
1188 1 : }
1189 : }
1190 : }
1191 : };
1192 :
1193 3 : let (timeline, _timeline_ctx) = self.create_timeline_struct(
1194 3 : timeline_id,
1195 3 : &metadata,
1196 3 : previous_heatmap,
1197 3 : ancestor.clone(),
1198 3 : resources,
1199 3 : CreateTimelineCause::Load,
1200 3 : idempotency.clone(),
1201 3 : index_part.gc_compaction.clone(),
1202 3 : index_part.rel_size_migration.clone(),
1203 3 : ctx,
1204 3 : )?;
1205 3 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
1206 :
1207 3 : if !disk_consistent_lsn.is_valid() {
1208 : // As opposed to normal timelines which get initialised with a disk consitent LSN
1209 : // via initdb, imported timelines start from 0. If the import task stops before
1210 : // it advances disk consitent LSN, allow it to resume.
1211 0 : let in_progress_import = import_pgdata
1212 0 : .as_ref()
1213 0 : .map(|import| !import.is_done())
1214 0 : .unwrap_or(false);
1215 0 : if !in_progress_import {
1216 0 : anyhow::bail!("Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn");
1217 0 : }
1218 3 : }
1219 :
1220 3 : assert_eq!(
1221 : disk_consistent_lsn,
1222 3 : metadata.disk_consistent_lsn(),
1223 0 : "these are used interchangeably"
1224 : );
1225 :
1226 3 : timeline.remote_client.init_upload_queue(&index_part)?;
1227 :
1228 3 : timeline
1229 3 : .load_layer_map(disk_consistent_lsn, index_part)
1230 3 : .await
1231 3 : .with_context(|| {
1232 0 : format!("Failed to load layermap for timeline {tenant_id}/{timeline_id}")
1233 0 : })?;
1234 :
1235 : // When unarchiving, we've mostly likely lost the heatmap generated prior
1236 : // to the archival operation. To allow warming this timeline up, generate
1237 : // a previous heatmap which contains all visible layers in the layer map.
1238 : // This previous heatmap will be used whenever a fresh heatmap is generated
1239 : // for the timeline.
1240 3 : if self.conf.generate_unarchival_heatmap && matches!(cause, LoadTimelineCause::Unoffload) {
1241 0 : let mut tline_ending_at = Some((&timeline, timeline.get_last_record_lsn()));
1242 0 : while let Some((tline, end_lsn)) = tline_ending_at {
1243 0 : let unarchival_heatmap = tline.generate_unarchival_heatmap(end_lsn).await;
1244 : // Another unearchived timeline might have generated a heatmap for this ancestor.
1245 : // If the current branch point greater than the previous one use the the heatmap
1246 : // we just generated - it should include more layers.
1247 0 : if !tline.should_keep_previous_heatmap(end_lsn) {
1248 0 : tline
1249 0 : .previous_heatmap
1250 0 : .store(Some(Arc::new(unarchival_heatmap)));
1251 0 : } else {
1252 0 : tracing::info!("Previous heatmap preferred. Dropping unarchival heatmap.")
1253 : }
1254 :
1255 0 : match tline.ancestor_timeline() {
1256 0 : Some(ancestor) => {
1257 0 : if ancestor.update_layer_visibility().await.is_err() {
1258 : // Ancestor timeline is shutting down.
1259 0 : break;
1260 0 : }
1261 :
1262 0 : tline_ending_at = Some((ancestor, tline.get_ancestor_lsn()));
1263 : }
1264 0 : None => {
1265 0 : tline_ending_at = None;
1266 0 : }
1267 : }
1268 : }
1269 3 : }
1270 :
1271 0 : match import_pgdata {
1272 0 : Some(import_pgdata) if !import_pgdata.is_done() => {
1273 0 : let mut guard = self.timelines_creating.lock().unwrap();
1274 0 : if !guard.insert(timeline_id) {
1275 : // We should never try and load the same timeline twice during startup
1276 0 : unreachable!("Timeline {tenant_id}/{timeline_id} is already being created")
1277 0 : }
1278 0 : let timeline_create_guard = TimelineCreateGuard {
1279 0 : _tenant_gate_guard: self.gate.enter()?,
1280 0 : owning_tenant: self.clone(),
1281 0 : timeline_id,
1282 0 : idempotency,
1283 : // The users of this specific return value don't need the timline_path in there.
1284 0 : timeline_path: timeline
1285 0 : .conf
1286 0 : .timeline_path(&timeline.tenant_shard_id, &timeline.timeline_id),
1287 : };
1288 0 : Ok(TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
1289 0 : TimelineInitAndSyncNeedsSpawnImportPgdata {
1290 0 : timeline,
1291 0 : import_pgdata,
1292 0 : guard: timeline_create_guard,
1293 0 : },
1294 0 : ))
1295 : }
1296 : Some(_) | None => {
1297 : {
1298 3 : let mut timelines_accessor = self.timelines.lock().unwrap();
1299 3 : match timelines_accessor.entry(timeline_id) {
1300 : // We should never try and load the same timeline twice during startup
1301 : Entry::Occupied(_) => {
1302 0 : unreachable!(
1303 : "Timeline {tenant_id}/{timeline_id} already exists in the tenant map"
1304 : );
1305 : }
1306 3 : Entry::Vacant(v) => {
1307 3 : v.insert(Arc::clone(&timeline));
1308 3 : timeline.maybe_spawn_flush_loop();
1309 3 : }
1310 : }
1311 : }
1312 :
1313 3 : if disk_consistent_lsn.is_valid() {
1314 : // Sanity check: a timeline should have some content.
1315 : // Exception: importing timelines might not yet have any
1316 3 : anyhow::ensure!(
1317 3 : ancestor.is_some()
1318 2 : || timeline
1319 2 : .layers
1320 2 : .read(LayerManagerLockHolder::LoadLayerMap)
1321 2 : .await
1322 2 : .layer_map()
1323 2 : .expect(
1324 2 : "currently loading, layer manager cannot be shutdown already"
1325 : )
1326 2 : .iter_historic_layers()
1327 2 : .next()
1328 2 : .is_some(),
1329 0 : "Timeline has no ancestor and no layer files"
1330 : );
1331 0 : }
1332 :
1333 3 : Ok(TimelineInitAndSyncResult::ReadyToActivate)
1334 : }
1335 : }
1336 3 : }
1337 :
1338 : /// Attach a tenant that's available in cloud storage.
1339 : ///
1340 : /// This returns quickly, after just creating the in-memory object
1341 : /// Tenant struct and launching a background task to download
1342 : /// the remote index files. On return, the tenant is most likely still in
1343 : /// Attaching state, and it will become Active once the background task
1344 : /// finishes. You can use wait_until_active() to wait for the task to
1345 : /// complete.
1346 : ///
1347 : #[allow(clippy::too_many_arguments)]
1348 0 : pub(crate) fn spawn(
1349 0 : conf: &'static PageServerConf,
1350 0 : tenant_shard_id: TenantShardId,
1351 0 : resources: TenantSharedResources,
1352 0 : attached_conf: AttachedTenantConf,
1353 0 : shard_identity: ShardIdentity,
1354 0 : init_order: Option<InitializationOrder>,
1355 0 : mode: SpawnMode,
1356 0 : ctx: &RequestContext,
1357 0 : ) -> Result<Arc<TenantShard>, GlobalShutDown> {
1358 0 : let wal_redo_manager =
1359 0 : WalRedoManager::new(PostgresRedoManager::new(conf, tenant_shard_id))?;
1360 :
1361 : let TenantSharedResources {
1362 0 : broker_client,
1363 0 : remote_storage,
1364 0 : deletion_queue_client,
1365 0 : l0_flush_global_state,
1366 0 : basebackup_cache,
1367 0 : feature_resolver,
1368 0 : } = resources;
1369 :
1370 0 : let attach_mode = attached_conf.location.attach_mode;
1371 0 : let generation = attached_conf.location.generation;
1372 :
1373 0 : let tenant = Arc::new(TenantShard::new(
1374 0 : TenantState::Attaching,
1375 0 : conf,
1376 0 : attached_conf,
1377 0 : shard_identity,
1378 0 : Some(wal_redo_manager),
1379 0 : tenant_shard_id,
1380 0 : remote_storage.clone(),
1381 0 : deletion_queue_client,
1382 0 : l0_flush_global_state,
1383 0 : basebackup_cache,
1384 0 : feature_resolver,
1385 : ));
1386 :
1387 : // The attach task will carry a GateGuard, so that shutdown() reliably waits for it to drop out if
1388 : // we shut down while attaching.
1389 0 : let attach_gate_guard = tenant
1390 0 : .gate
1391 0 : .enter()
1392 0 : .expect("We just created the TenantShard: nothing else can have shut it down yet");
1393 :
1394 : // Do all the hard work in the background
1395 0 : let tenant_clone = Arc::clone(&tenant);
1396 0 : let ctx = ctx.detached_child(TaskKind::Attach, DownloadBehavior::Warn);
1397 0 : task_mgr::spawn(
1398 0 : &tokio::runtime::Handle::current(),
1399 0 : TaskKind::Attach,
1400 0 : tenant_shard_id,
1401 0 : None,
1402 0 : "attach tenant",
1403 0 : async move {
1404 :
1405 0 : info!(
1406 : ?attach_mode,
1407 0 : "Attaching tenant"
1408 : );
1409 :
1410 0 : let _gate_guard = attach_gate_guard;
1411 :
1412 : // Is this tenant being spawned as part of process startup?
1413 0 : let starting_up = init_order.is_some();
1414 0 : scopeguard::defer! {
1415 : if starting_up {
1416 : TENANT.startup_complete.inc();
1417 : }
1418 : }
1419 :
1420 0 : fn make_broken_or_stopping(t: &TenantShard, err: anyhow::Error) {
1421 0 : t.state.send_modify(|state| match state {
1422 : // TODO: the old code alluded to DeleteTenantFlow sometimes setting
1423 : // TenantState::Stopping before we get here, but this may be outdated.
1424 : // Let's find out with a testing assertion. If this doesn't fire, and the
1425 : // logs don't show this happening in production, remove the Stopping cases.
1426 0 : TenantState::Stopping{..} if cfg!(any(test, feature = "testing")) => {
1427 0 : panic!("unexpected TenantState::Stopping during attach")
1428 : }
1429 : // If the tenant is cancelled, assume the error was caused by cancellation.
1430 0 : TenantState::Attaching if t.cancel.is_cancelled() => {
1431 0 : info!("attach cancelled, setting tenant state to Stopping: {err}");
1432 : // NB: progress None tells `set_stopping` that attach has cancelled.
1433 0 : *state = TenantState::Stopping { progress: None };
1434 : }
1435 : // According to the old code, DeleteTenantFlow may already have set this to
1436 : // Stopping. Retain its progress.
1437 : // TODO: there is no DeleteTenantFlow. Is this still needed? See above.
1438 0 : TenantState::Stopping { progress } if t.cancel.is_cancelled() => {
1439 0 : assert!(progress.is_some(), "concurrent attach cancellation");
1440 0 : info!("attach cancelled, already Stopping: {err}");
1441 : }
1442 : // Mark the tenant as broken.
1443 : TenantState::Attaching | TenantState::Stopping { .. } => {
1444 0 : error!("attach failed, setting tenant state to Broken (was {state}): {err:?}");
1445 0 : *state = TenantState::broken_from_reason(err.to_string())
1446 : }
1447 : // The attach task owns the tenant state until activated.
1448 0 : state => panic!("invalid tenant state {state} during attach: {err:?}"),
1449 0 : });
1450 0 : }
1451 :
1452 : // TODO: should also be rejecting tenant conf changes that violate this check.
1453 0 : if let Err(e) = crate::tenant::storage_layer::inmemory_layer::IndexEntry::validate_checkpoint_distance(tenant_clone.get_checkpoint_distance()) {
1454 0 : make_broken_or_stopping(&tenant_clone, anyhow::anyhow!(e));
1455 0 : return Ok(());
1456 0 : }
1457 :
1458 0 : let mut init_order = init_order;
1459 : // take the completion because initial tenant loading will complete when all of
1460 : // these tasks complete.
1461 0 : let _completion = init_order
1462 0 : .as_mut()
1463 0 : .and_then(|x| x.initial_tenant_load.take());
1464 0 : let remote_load_completion = init_order
1465 0 : .as_mut()
1466 0 : .and_then(|x| x.initial_tenant_load_remote.take());
1467 :
1468 : enum AttachType<'a> {
1469 : /// We are attaching this tenant lazily in the background.
1470 : Warmup {
1471 : _permit: tokio::sync::SemaphorePermit<'a>,
1472 : during_startup: bool
1473 : },
1474 : /// We are attaching this tenant as soon as we can, because for example an
1475 : /// endpoint tried to access it.
1476 : OnDemand,
1477 : /// During normal operations after startup, we are attaching a tenant, and
1478 : /// eager attach was requested.
1479 : Normal,
1480 : }
1481 :
1482 0 : let attach_type = if matches!(mode, SpawnMode::Lazy) {
1483 : // Before doing any I/O, wait for at least one of:
1484 : // - A client attempting to access to this tenant (on-demand loading)
1485 : // - A permit becoming available in the warmup semaphore (background warmup)
1486 :
1487 0 : tokio::select!(
1488 0 : permit = tenant_clone.activate_now_sem.acquire() => {
1489 0 : let _ = permit.expect("activate_now_sem is never closed");
1490 0 : tracing::info!("Activating tenant (on-demand)");
1491 0 : AttachType::OnDemand
1492 : },
1493 0 : permit = conf.concurrent_tenant_warmup.inner().acquire() => {
1494 0 : let _permit = permit.expect("concurrent_tenant_warmup semaphore is never closed");
1495 0 : tracing::info!("Activating tenant (warmup)");
1496 0 : AttachType::Warmup {
1497 0 : _permit,
1498 0 : during_startup: init_order.is_some()
1499 0 : }
1500 : }
1501 0 : _ = tenant_clone.cancel.cancelled() => {
1502 : // This is safe, but should be pretty rare: it is interesting if a tenant
1503 : // stayed in Activating for such a long time that shutdown found it in
1504 : // that state.
1505 0 : tracing::info!(state=%tenant_clone.current_state(), "Tenant shut down before activation");
1506 : // Set the tenant to Stopping to signal `set_stopping` that we're done.
1507 0 : make_broken_or_stopping(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"));
1508 0 : return Ok(());
1509 : },
1510 : )
1511 : } else {
1512 : // SpawnMode::{Create,Eager} always cause jumping ahead of the
1513 : // concurrent_tenant_warmup queue
1514 0 : AttachType::Normal
1515 : };
1516 :
1517 0 : let preload = match &mode {
1518 : SpawnMode::Eager | SpawnMode::Lazy => {
1519 0 : let _preload_timer = TENANT.preload.start_timer();
1520 0 : let res = tenant_clone
1521 0 : .preload(&remote_storage, task_mgr::shutdown_token())
1522 0 : .await;
1523 0 : match res {
1524 0 : Ok(p) => Some(p),
1525 0 : Err(e) => {
1526 0 : make_broken_or_stopping(&tenant_clone, anyhow::anyhow!(e));
1527 0 : return Ok(());
1528 : }
1529 : }
1530 : }
1531 :
1532 : };
1533 :
1534 : // Remote preload is complete.
1535 0 : drop(remote_load_completion);
1536 :
1537 :
1538 : // We will time the duration of the attach phase unless this is a creation (attach will do no work)
1539 0 : let attach_start = std::time::Instant::now();
1540 0 : let attached = {
1541 0 : let _attach_timer = Some(TENANT.attach.start_timer());
1542 0 : tenant_clone.attach(preload, &ctx).await
1543 : };
1544 0 : let attach_duration = attach_start.elapsed();
1545 0 : _ = tenant_clone.attach_wal_lag_cooldown.set(WalLagCooldown::new(attach_start, attach_duration));
1546 :
1547 0 : match attached {
1548 : Ok(()) => {
1549 0 : info!("attach finished, activating");
1550 0 : tenant_clone.activate(broker_client, None, &ctx);
1551 : }
1552 0 : Err(e) => make_broken_or_stopping(&tenant_clone, anyhow::anyhow!(e)),
1553 : }
1554 :
1555 : // If we are doing an opportunistic warmup attachment at startup, initialize
1556 : // logical size at the same time. This is better than starting a bunch of idle tenants
1557 : // with cold caches and then coming back later to initialize their logical sizes.
1558 : //
1559 : // It also prevents the warmup proccess competing with the concurrency limit on
1560 : // logical size calculations: if logical size calculation semaphore is saturated,
1561 : // then warmup will wait for that before proceeding to the next tenant.
1562 0 : if matches!(attach_type, AttachType::Warmup { during_startup: true, .. }) {
1563 0 : let mut futs: FuturesUnordered<_> = tenant_clone.timelines.lock().unwrap().values().cloned().map(|t| t.await_initial_logical_size()).collect();
1564 0 : tracing::info!("Waiting for initial logical sizes while warming up...");
1565 0 : while futs.next().await.is_some() {}
1566 0 : tracing::info!("Warm-up complete");
1567 0 : }
1568 :
1569 0 : Ok(())
1570 0 : }
1571 0 : .instrument(tracing::info_span!(parent: None, "attach", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), gen=?generation)),
1572 : );
1573 0 : Ok(tenant)
1574 0 : }
1575 :
1576 : #[instrument(skip_all)]
1577 : pub(crate) async fn preload(
1578 : self: &Arc<Self>,
1579 : remote_storage: &GenericRemoteStorage,
1580 : cancel: CancellationToken,
1581 : ) -> anyhow::Result<TenantPreload> {
1582 : span::debug_assert_current_span_has_tenant_id();
1583 : // Get list of remote timelines
1584 : // download index files for every tenant timeline
1585 : info!("listing remote timelines");
1586 : let (mut remote_timeline_ids, other_keys) = remote_timeline_client::list_remote_timelines(
1587 : remote_storage,
1588 : self.tenant_shard_id,
1589 : cancel.clone(),
1590 : )
1591 : .await?;
1592 :
1593 : let tenant_manifest = match download_tenant_manifest(
1594 : remote_storage,
1595 : &self.tenant_shard_id,
1596 : self.generation,
1597 : &cancel,
1598 : )
1599 : .await
1600 : {
1601 : Ok((tenant_manifest, _, _)) => Some(tenant_manifest),
1602 : Err(DownloadError::NotFound) => None,
1603 : Err(err) => return Err(err.into()),
1604 : };
1605 :
1606 : info!(
1607 : "found {} timelines ({} offloaded timelines)",
1608 : remote_timeline_ids.len(),
1609 : tenant_manifest
1610 : .as_ref()
1611 3 : .map(|m| m.offloaded_timelines.len())
1612 : .unwrap_or(0)
1613 : );
1614 :
1615 : for k in other_keys {
1616 : warn!("Unexpected non timeline key {k}");
1617 : }
1618 :
1619 : // Avoid downloading IndexPart of offloaded timelines.
1620 : let mut offloaded_with_prefix = HashSet::new();
1621 : if let Some(tenant_manifest) = &tenant_manifest {
1622 : for offloaded in tenant_manifest.offloaded_timelines.iter() {
1623 : if remote_timeline_ids.remove(&offloaded.timeline_id) {
1624 : offloaded_with_prefix.insert(offloaded.timeline_id);
1625 : } else {
1626 : // We'll take care later of timelines in the manifest without a prefix
1627 : }
1628 : }
1629 : }
1630 :
1631 : // TODO(vlad): Could go to S3 if the secondary is freezing cold and hasn't even
1632 : // pulled the first heatmap. Not entirely necessary since the storage controller
1633 : // will kick the secondary in any case and cause a download.
1634 : let maybe_heatmap_at = self.read_on_disk_heatmap().await;
1635 :
1636 : let timelines = self
1637 : .load_timelines_metadata(
1638 : remote_timeline_ids,
1639 : remote_storage,
1640 : maybe_heatmap_at,
1641 : cancel,
1642 : )
1643 : .await?;
1644 :
1645 : Ok(TenantPreload {
1646 : tenant_manifest,
1647 : timelines: timelines
1648 : .into_iter()
1649 3 : .map(|(id, tl)| (id, Some(tl)))
1650 0 : .chain(offloaded_with_prefix.into_iter().map(|id| (id, None)))
1651 : .collect(),
1652 : })
1653 : }
1654 :
1655 118 : async fn read_on_disk_heatmap(&self) -> Option<(HeatMapTenant, std::time::Instant)> {
1656 118 : if !self.conf.load_previous_heatmap {
1657 0 : return None;
1658 118 : }
1659 :
1660 118 : let on_disk_heatmap_path = self.conf.tenant_heatmap_path(&self.tenant_shard_id);
1661 118 : match tokio::fs::read_to_string(on_disk_heatmap_path).await {
1662 0 : Ok(heatmap) => match serde_json::from_str::<HeatMapTenant>(&heatmap) {
1663 0 : Ok(heatmap) => Some((heatmap, std::time::Instant::now())),
1664 0 : Err(err) => {
1665 0 : error!("Failed to deserialize old heatmap: {err}");
1666 0 : None
1667 : }
1668 : },
1669 118 : Err(err) => match err.kind() {
1670 118 : std::io::ErrorKind::NotFound => None,
1671 : _ => {
1672 0 : error!("Unexpected IO error reading old heatmap: {err}");
1673 0 : None
1674 : }
1675 : },
1676 : }
1677 118 : }
1678 :
1679 : ///
1680 : /// Background task that downloads all data for a tenant and brings it to Active state.
1681 : ///
1682 : /// No background tasks are started as part of this routine.
1683 : ///
1684 118 : async fn attach(
1685 118 : self: &Arc<TenantShard>,
1686 118 : preload: Option<TenantPreload>,
1687 118 : ctx: &RequestContext,
1688 118 : ) -> anyhow::Result<()> {
1689 118 : span::debug_assert_current_span_has_tenant_id();
1690 :
1691 118 : failpoint_support::sleep_millis_async!("before-attaching-tenant");
1692 :
1693 118 : let Some(preload) = preload else {
1694 0 : anyhow::bail!(
1695 0 : "local-only deployment is no longer supported, https://github.com/neondatabase/neon/issues/5624"
1696 : );
1697 : };
1698 :
1699 118 : let mut offloaded_timeline_ids = HashSet::new();
1700 118 : let mut offloaded_timelines_list = Vec::new();
1701 118 : if let Some(tenant_manifest) = &preload.tenant_manifest {
1702 3 : for timeline_manifest in tenant_manifest.offloaded_timelines.iter() {
1703 0 : let timeline_id = timeline_manifest.timeline_id;
1704 0 : let offloaded_timeline =
1705 0 : OffloadedTimeline::from_manifest(self.tenant_shard_id, timeline_manifest);
1706 0 : offloaded_timelines_list.push((timeline_id, Arc::new(offloaded_timeline)));
1707 0 : offloaded_timeline_ids.insert(timeline_id);
1708 0 : }
1709 115 : }
1710 : // Complete deletions for offloaded timeline id's from manifest.
1711 : // The manifest will be uploaded later in this function.
1712 118 : offloaded_timelines_list
1713 118 : .retain(|(offloaded_id, offloaded)| {
1714 : // Existence of a timeline is finally determined by the existence of an index-part.json in remote storage.
1715 : // If there is dangling references in another location, they need to be cleaned up.
1716 0 : let delete = !preload.timelines.contains_key(offloaded_id);
1717 0 : if delete {
1718 0 : tracing::info!("Removing offloaded timeline {offloaded_id} from manifest as no remote prefix was found");
1719 0 : offloaded.defuse_for_tenant_drop();
1720 0 : }
1721 0 : !delete
1722 0 : });
1723 :
1724 118 : let mut timelines_to_resume_deletions = vec![];
1725 :
1726 118 : let mut remote_index_and_client = HashMap::new();
1727 118 : let mut timeline_ancestors = HashMap::new();
1728 118 : let mut existent_timelines = HashSet::new();
1729 121 : for (timeline_id, preload) in preload.timelines {
1730 3 : let Some(preload) = preload else { continue };
1731 : // This is an invariant of the `preload` function's API
1732 3 : assert!(!offloaded_timeline_ids.contains(&timeline_id));
1733 3 : let index_part = match preload.index_part {
1734 3 : Ok(i) => {
1735 3 : debug!("remote index part exists for timeline {timeline_id}");
1736 : // We found index_part on the remote, this is the standard case.
1737 3 : existent_timelines.insert(timeline_id);
1738 3 : i
1739 : }
1740 : Err(DownloadError::NotFound) => {
1741 : // There is no index_part on the remote. We only get here
1742 : // if there is some prefix for the timeline in the remote storage.
1743 : // This can e.g. be the initdb.tar.zst archive, maybe a
1744 : // remnant from a prior incomplete creation or deletion attempt.
1745 : // Delete the local directory as the deciding criterion for a
1746 : // timeline's existence is presence of index_part.
1747 0 : info!(%timeline_id, "index_part not found on remote");
1748 0 : continue;
1749 : }
1750 0 : Err(DownloadError::Fatal(why)) => {
1751 : // If, while loading one remote timeline, we saw an indication that our generation
1752 : // number is likely invalid, then we should not load the whole tenant.
1753 0 : error!(%timeline_id, "Fatal error loading timeline: {why}");
1754 0 : anyhow::bail!(why.to_string());
1755 : }
1756 0 : Err(e) => {
1757 : // Some (possibly ephemeral) error happened during index_part download.
1758 : // Pretend the timeline exists to not delete the timeline directory,
1759 : // as it might be a temporary issue and we don't want to re-download
1760 : // everything after it resolves.
1761 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1762 :
1763 0 : existent_timelines.insert(timeline_id);
1764 0 : continue;
1765 : }
1766 : };
1767 3 : match index_part {
1768 3 : MaybeDeletedIndexPart::IndexPart(index_part) => {
1769 3 : timeline_ancestors.insert(timeline_id, index_part.metadata.clone());
1770 3 : remote_index_and_client.insert(
1771 3 : timeline_id,
1772 3 : (index_part, preload.client, preload.previous_heatmap),
1773 3 : );
1774 3 : }
1775 0 : MaybeDeletedIndexPart::Deleted(index_part) => {
1776 0 : info!(
1777 0 : "timeline {} is deleted, picking to resume deletion",
1778 : timeline_id
1779 : );
1780 0 : timelines_to_resume_deletions.push((timeline_id, index_part, preload.client));
1781 : }
1782 : }
1783 : }
1784 :
1785 118 : let mut gc_blocks = HashMap::new();
1786 :
1787 : // For every timeline, download the metadata file, scan the local directory,
1788 : // and build a layer map that contains an entry for each remote and local
1789 : // layer file.
1790 118 : let sorted_timelines = tree_sort_timelines(timeline_ancestors, |m| m.ancestor_timeline())?;
1791 121 : for (timeline_id, remote_metadata) in sorted_timelines {
1792 3 : let (index_part, remote_client, previous_heatmap) = remote_index_and_client
1793 3 : .remove(&timeline_id)
1794 3 : .expect("just put it in above");
1795 :
1796 3 : if let Some(blocking) = index_part.gc_blocking.as_ref() {
1797 : // could just filter these away, but it helps while testing
1798 0 : anyhow::ensure!(
1799 0 : !blocking.reasons.is_empty(),
1800 0 : "index_part for {timeline_id} is malformed: it should not have gc blocking with zero reasons"
1801 : );
1802 0 : let prev = gc_blocks.insert(timeline_id, blocking.reasons);
1803 0 : assert!(prev.is_none());
1804 3 : }
1805 :
1806 : // TODO again handle early failure
1807 3 : let effect = self
1808 3 : .load_remote_timeline(
1809 3 : timeline_id,
1810 3 : index_part,
1811 3 : remote_metadata,
1812 3 : previous_heatmap,
1813 3 : self.get_timeline_resources_for(remote_client),
1814 3 : LoadTimelineCause::Attach,
1815 3 : ctx,
1816 3 : )
1817 3 : .await
1818 3 : .with_context(|| {
1819 0 : format!(
1820 0 : "failed to load remote timeline {} for tenant {}",
1821 0 : timeline_id, self.tenant_shard_id
1822 : )
1823 0 : })?;
1824 :
1825 3 : match effect {
1826 3 : TimelineInitAndSyncResult::ReadyToActivate => {
1827 3 : // activation happens later, on Tenant::activate
1828 3 : }
1829 : TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
1830 : TimelineInitAndSyncNeedsSpawnImportPgdata {
1831 0 : timeline,
1832 0 : import_pgdata,
1833 0 : guard,
1834 : },
1835 : ) => {
1836 0 : let timeline_id = timeline.timeline_id;
1837 0 : let import_task_gate = Gate::default();
1838 0 : let import_task_guard = import_task_gate.enter().unwrap();
1839 0 : let import_task_handle =
1840 0 : tokio::task::spawn(self.clone().create_timeline_import_pgdata_task(
1841 0 : timeline.clone(),
1842 0 : import_pgdata,
1843 0 : guard,
1844 0 : import_task_guard,
1845 0 : ctx.detached_child(TaskKind::ImportPgdata, DownloadBehavior::Warn),
1846 : ));
1847 :
1848 0 : let prev = self.timelines_importing.lock().unwrap().insert(
1849 0 : timeline_id,
1850 0 : Arc::new(ImportingTimeline {
1851 0 : timeline: timeline.clone(),
1852 0 : import_task_handle,
1853 0 : import_task_gate,
1854 0 : delete_progress: TimelineDeleteProgress::default(),
1855 0 : }),
1856 0 : );
1857 :
1858 0 : assert!(prev.is_none());
1859 : }
1860 : }
1861 : }
1862 :
1863 : // At this point we've initialized all timelines and are tracking them.
1864 : // Now compute the layer visibility for all (not offloaded) timelines.
1865 118 : let compute_visiblity_for = {
1866 118 : let timelines_accessor = self.timelines.lock().unwrap();
1867 118 : let mut timelines_offloaded_accessor = self.timelines_offloaded.lock().unwrap();
1868 :
1869 118 : timelines_offloaded_accessor.extend(offloaded_timelines_list.into_iter());
1870 :
1871 : // Before activation, populate each Timeline's GcInfo with information about its children
1872 118 : self.initialize_gc_info(&timelines_accessor, &timelines_offloaded_accessor, None);
1873 :
1874 118 : timelines_accessor.values().cloned().collect::<Vec<_>>()
1875 : };
1876 :
1877 121 : for tl in compute_visiblity_for {
1878 3 : tl.update_layer_visibility().await.with_context(|| {
1879 0 : format!(
1880 0 : "failed initial timeline visibility computation {} for tenant {}",
1881 0 : tl.timeline_id, self.tenant_shard_id
1882 : )
1883 0 : })?;
1884 : }
1885 :
1886 : // Walk through deleted timelines, resume deletion
1887 118 : for (timeline_id, index_part, remote_timeline_client) in timelines_to_resume_deletions {
1888 0 : remote_timeline_client
1889 0 : .init_upload_queue_stopped_to_continue_deletion(&index_part)
1890 0 : .context("init queue stopped")
1891 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1892 :
1893 0 : DeleteTimelineFlow::resume_deletion(
1894 0 : Arc::clone(self),
1895 0 : timeline_id,
1896 0 : &index_part.metadata,
1897 0 : remote_timeline_client,
1898 0 : ctx,
1899 : )
1900 0 : .instrument(tracing::info_span!("timeline_delete", %timeline_id))
1901 0 : .await
1902 0 : .context("resume_deletion")
1903 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1904 : }
1905 :
1906 : // Stash the preloaded tenant manifest, and upload a new manifest if changed.
1907 : //
1908 : // NB: this must happen after the tenant is fully populated above. In particular the
1909 : // offloaded timelines, which are included in the manifest.
1910 : {
1911 118 : let mut guard = self.remote_tenant_manifest.lock().await;
1912 118 : assert!(guard.is_none(), "tenant manifest set before preload"); // first populated here
1913 118 : *guard = preload.tenant_manifest;
1914 : }
1915 118 : self.maybe_upload_tenant_manifest().await?;
1916 :
1917 : // The local filesystem contents are a cache of what's in the remote IndexPart;
1918 : // IndexPart is the source of truth.
1919 118 : self.clean_up_timelines(&existent_timelines)?;
1920 :
1921 118 : self.gc_block.set_scanned(gc_blocks);
1922 :
1923 118 : fail::fail_point!("attach-before-activate", |_| {
1924 0 : anyhow::bail!("attach-before-activate");
1925 0 : });
1926 118 : failpoint_support::sleep_millis_async!("attach-before-activate-sleep", &self.cancel);
1927 :
1928 118 : info!("Done");
1929 :
1930 118 : Ok(())
1931 118 : }
1932 :
1933 : /// Check for any local timeline directories that are temporary, or do not correspond to a
1934 : /// timeline that still exists: this can happen if we crashed during a deletion/creation, or
1935 : /// if a timeline was deleted while the tenant was attached to a different pageserver.
1936 118 : fn clean_up_timelines(&self, existent_timelines: &HashSet<TimelineId>) -> anyhow::Result<()> {
1937 118 : let timelines_dir = self.conf.timelines_path(&self.tenant_shard_id);
1938 :
1939 118 : let entries = match timelines_dir.read_dir_utf8() {
1940 118 : Ok(d) => d,
1941 0 : Err(e) => {
1942 0 : if e.kind() == std::io::ErrorKind::NotFound {
1943 0 : return Ok(());
1944 : } else {
1945 0 : return Err(e).context("list timelines directory for tenant");
1946 : }
1947 : }
1948 : };
1949 :
1950 122 : for entry in entries {
1951 4 : let entry = entry.context("read timeline dir entry")?;
1952 4 : let entry_path = entry.path();
1953 :
1954 4 : let purge = if crate::is_temporary(entry_path) {
1955 0 : true
1956 : } else {
1957 4 : match TimelineId::try_from(entry_path.file_name()) {
1958 4 : Ok(i) => {
1959 : // Purge if the timeline ID does not exist in remote storage: remote storage is the authority.
1960 4 : !existent_timelines.contains(&i)
1961 : }
1962 0 : Err(e) => {
1963 0 : tracing::warn!(
1964 0 : "Unparseable directory in timelines directory: {entry_path}, ignoring ({e})"
1965 : );
1966 : // Do not purge junk: if we don't recognize it, be cautious and leave it for a human.
1967 0 : false
1968 : }
1969 : }
1970 : };
1971 :
1972 4 : if purge {
1973 1 : tracing::info!("Purging stale timeline dentry {entry_path}");
1974 1 : if let Err(e) = match entry.file_type() {
1975 1 : Ok(t) => if t.is_dir() {
1976 1 : std::fs::remove_dir_all(entry_path)
1977 : } else {
1978 0 : std::fs::remove_file(entry_path)
1979 : }
1980 1 : .or_else(fs_ext::ignore_not_found),
1981 0 : Err(e) => Err(e),
1982 : } {
1983 0 : tracing::warn!("Failed to purge stale timeline dentry {entry_path}: {e}");
1984 1 : }
1985 3 : }
1986 : }
1987 :
1988 118 : Ok(())
1989 118 : }
1990 :
1991 : /// Get sum of all remote timelines sizes
1992 : ///
1993 : /// This function relies on the index_part instead of listing the remote storage
1994 0 : pub fn remote_size(&self) -> u64 {
1995 0 : let mut size = 0;
1996 :
1997 0 : for timeline in self.list_timelines() {
1998 0 : size += timeline.remote_client.get_remote_physical_size();
1999 0 : }
2000 :
2001 0 : size
2002 0 : }
2003 :
2004 : #[instrument(skip_all, fields(timeline_id=%timeline_id))]
2005 : #[allow(clippy::too_many_arguments)]
2006 : async fn load_remote_timeline(
2007 : self: &Arc<Self>,
2008 : timeline_id: TimelineId,
2009 : index_part: IndexPart,
2010 : remote_metadata: TimelineMetadata,
2011 : previous_heatmap: Option<PreviousHeatmap>,
2012 : resources: TimelineResources,
2013 : cause: LoadTimelineCause,
2014 : ctx: &RequestContext,
2015 : ) -> anyhow::Result<TimelineInitAndSyncResult> {
2016 : span::debug_assert_current_span_has_tenant_id();
2017 :
2018 : info!("downloading index file for timeline {}", timeline_id);
2019 : tokio::fs::create_dir_all(self.conf.timeline_path(&self.tenant_shard_id, &timeline_id))
2020 : .await
2021 : .context("Failed to create new timeline directory")?;
2022 :
2023 : let ancestor = if let Some(ancestor_id) = remote_metadata.ancestor_timeline() {
2024 : let timelines = self.timelines.lock().unwrap();
2025 : Some(Arc::clone(timelines.get(&ancestor_id).ok_or_else(
2026 0 : || {
2027 0 : anyhow::anyhow!(
2028 0 : "cannot find ancestor timeline {ancestor_id} for timeline {timeline_id}"
2029 : )
2030 0 : },
2031 : )?))
2032 : } else {
2033 : None
2034 : };
2035 :
2036 : self.timeline_init_and_sync(
2037 : timeline_id,
2038 : resources,
2039 : index_part,
2040 : remote_metadata,
2041 : previous_heatmap,
2042 : ancestor,
2043 : cause,
2044 : ctx,
2045 : )
2046 : .await
2047 : }
2048 :
2049 118 : async fn load_timelines_metadata(
2050 118 : self: &Arc<TenantShard>,
2051 118 : timeline_ids: HashSet<TimelineId>,
2052 118 : remote_storage: &GenericRemoteStorage,
2053 118 : heatmap: Option<(HeatMapTenant, std::time::Instant)>,
2054 118 : cancel: CancellationToken,
2055 118 : ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
2056 118 : let mut timeline_heatmaps = heatmap.map(|h| (h.0.into_timelines_index(), h.1));
2057 :
2058 118 : let mut part_downloads = JoinSet::new();
2059 121 : for timeline_id in timeline_ids {
2060 3 : let cancel_clone = cancel.clone();
2061 :
2062 3 : let previous_timeline_heatmap = timeline_heatmaps.as_mut().and_then(|hs| {
2063 0 : hs.0.remove(&timeline_id).map(|h| PreviousHeatmap::Active {
2064 0 : heatmap: h,
2065 0 : read_at: hs.1,
2066 0 : end_lsn: None,
2067 0 : })
2068 0 : });
2069 3 : part_downloads.spawn(
2070 3 : self.load_timeline_metadata(
2071 3 : timeline_id,
2072 3 : remote_storage.clone(),
2073 3 : previous_timeline_heatmap,
2074 3 : cancel_clone,
2075 : )
2076 3 : .instrument(info_span!("download_index_part", %timeline_id)),
2077 : );
2078 : }
2079 :
2080 118 : let mut timeline_preloads: HashMap<TimelineId, TimelinePreload> = HashMap::new();
2081 :
2082 : loop {
2083 121 : tokio::select!(
2084 121 : next = part_downloads.join_next() => {
2085 121 : match next {
2086 3 : Some(result) => {
2087 3 : let preload = result.context("join preload task")?;
2088 3 : timeline_preloads.insert(preload.timeline_id, preload);
2089 : },
2090 : None => {
2091 118 : break;
2092 : }
2093 : }
2094 : },
2095 121 : _ = cancel.cancelled() => {
2096 0 : anyhow::bail!("Cancelled while waiting for remote index download")
2097 : }
2098 : )
2099 : }
2100 :
2101 118 : Ok(timeline_preloads)
2102 118 : }
2103 :
2104 3 : fn build_timeline_client(
2105 3 : &self,
2106 3 : timeline_id: TimelineId,
2107 3 : remote_storage: GenericRemoteStorage,
2108 3 : ) -> RemoteTimelineClient {
2109 3 : RemoteTimelineClient::new(
2110 3 : remote_storage.clone(),
2111 3 : self.deletion_queue_client.clone(),
2112 3 : self.conf,
2113 3 : self.tenant_shard_id,
2114 3 : timeline_id,
2115 3 : self.generation,
2116 3 : &self.tenant_conf.load().location,
2117 : )
2118 3 : }
2119 :
2120 3 : fn load_timeline_metadata(
2121 3 : self: &Arc<TenantShard>,
2122 3 : timeline_id: TimelineId,
2123 3 : remote_storage: GenericRemoteStorage,
2124 3 : previous_heatmap: Option<PreviousHeatmap>,
2125 3 : cancel: CancellationToken,
2126 3 : ) -> impl Future<Output = TimelinePreload> + use<> {
2127 3 : let client = self.build_timeline_client(timeline_id, remote_storage);
2128 3 : async move {
2129 3 : debug_assert_current_span_has_tenant_and_timeline_id();
2130 3 : debug!("starting index part download");
2131 :
2132 3 : let index_part = client.download_index_file(&cancel).await;
2133 :
2134 3 : debug!("finished index part download");
2135 :
2136 3 : TimelinePreload {
2137 3 : client,
2138 3 : timeline_id,
2139 3 : index_part,
2140 3 : previous_heatmap,
2141 3 : }
2142 3 : }
2143 3 : }
2144 :
2145 0 : fn check_to_be_archived_has_no_unarchived_children(
2146 0 : timeline_id: TimelineId,
2147 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
2148 0 : ) -> Result<(), TimelineArchivalError> {
2149 0 : let children: Vec<TimelineId> = timelines
2150 0 : .iter()
2151 0 : .filter_map(|(id, entry)| {
2152 0 : if entry.get_ancestor_timeline_id() != Some(timeline_id) {
2153 0 : return None;
2154 0 : }
2155 0 : if entry.is_archived() == Some(true) {
2156 0 : return None;
2157 0 : }
2158 0 : Some(*id)
2159 0 : })
2160 0 : .collect();
2161 :
2162 0 : if !children.is_empty() {
2163 0 : return Err(TimelineArchivalError::HasUnarchivedChildren(children));
2164 0 : }
2165 0 : Ok(())
2166 0 : }
2167 :
2168 0 : fn check_ancestor_of_to_be_unarchived_is_not_archived(
2169 0 : ancestor_timeline_id: TimelineId,
2170 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
2171 0 : offloaded_timelines: &std::sync::MutexGuard<
2172 0 : '_,
2173 0 : HashMap<TimelineId, Arc<OffloadedTimeline>>,
2174 0 : >,
2175 0 : ) -> Result<(), TimelineArchivalError> {
2176 0 : let has_archived_parent =
2177 0 : if let Some(ancestor_timeline) = timelines.get(&ancestor_timeline_id) {
2178 0 : ancestor_timeline.is_archived() == Some(true)
2179 0 : } else if offloaded_timelines.contains_key(&ancestor_timeline_id) {
2180 0 : true
2181 : } else {
2182 0 : error!("ancestor timeline {ancestor_timeline_id} not found");
2183 0 : if cfg!(debug_assertions) {
2184 0 : panic!("ancestor timeline {ancestor_timeline_id} not found");
2185 0 : }
2186 0 : return Err(TimelineArchivalError::NotFound);
2187 : };
2188 0 : if has_archived_parent {
2189 0 : return Err(TimelineArchivalError::HasArchivedParent(
2190 0 : ancestor_timeline_id,
2191 0 : ));
2192 0 : }
2193 0 : Ok(())
2194 0 : }
2195 :
2196 0 : fn check_to_be_unarchived_timeline_has_no_archived_parent(
2197 0 : timeline: &Arc<Timeline>,
2198 0 : ) -> Result<(), TimelineArchivalError> {
2199 0 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
2200 0 : if ancestor_timeline.is_archived() == Some(true) {
2201 0 : return Err(TimelineArchivalError::HasArchivedParent(
2202 0 : ancestor_timeline.timeline_id,
2203 0 : ));
2204 0 : }
2205 0 : }
2206 0 : Ok(())
2207 0 : }
2208 :
2209 : /// Loads the specified (offloaded) timeline from S3 and attaches it as a loaded timeline
2210 : ///
2211 : /// Counterpart to [`offload_timeline`].
2212 0 : async fn unoffload_timeline(
2213 0 : self: &Arc<Self>,
2214 0 : timeline_id: TimelineId,
2215 0 : broker_client: storage_broker::BrokerClientChannel,
2216 0 : ctx: RequestContext,
2217 0 : ) -> Result<Arc<Timeline>, TimelineArchivalError> {
2218 0 : info!("unoffloading timeline");
2219 :
2220 : // We activate the timeline below manually, so this must be called on an active tenant.
2221 : // We expect callers of this function to ensure this.
2222 0 : match self.current_state() {
2223 : TenantState::Activating { .. }
2224 : | TenantState::Attaching
2225 : | TenantState::Broken { .. } => {
2226 0 : panic!("Timeline expected to be active")
2227 : }
2228 0 : TenantState::Stopping { .. } => return Err(TimelineArchivalError::Cancelled),
2229 0 : TenantState::Active => {}
2230 : }
2231 0 : let cancel = self.cancel.clone();
2232 :
2233 : // Protect against concurrent attempts to use this TimelineId
2234 : // We don't care much about idempotency, as it's ensured a layer above.
2235 0 : let allow_offloaded = true;
2236 0 : let _create_guard = self
2237 0 : .create_timeline_create_guard(
2238 0 : timeline_id,
2239 0 : CreateTimelineIdempotency::FailWithConflict,
2240 0 : allow_offloaded,
2241 : )
2242 0 : .map_err(|err| match err {
2243 0 : TimelineExclusionError::AlreadyCreating => TimelineArchivalError::AlreadyInProgress,
2244 : TimelineExclusionError::AlreadyExists { .. } => {
2245 0 : TimelineArchivalError::Other(anyhow::anyhow!("Timeline already exists"))
2246 : }
2247 0 : TimelineExclusionError::Other(e) => TimelineArchivalError::Other(e),
2248 0 : TimelineExclusionError::ShuttingDown => TimelineArchivalError::Cancelled,
2249 0 : })?;
2250 :
2251 0 : let timeline_preload = self
2252 0 : .load_timeline_metadata(
2253 0 : timeline_id,
2254 0 : self.remote_storage.clone(),
2255 0 : None,
2256 0 : cancel.clone(),
2257 0 : )
2258 0 : .await;
2259 :
2260 0 : let index_part = match timeline_preload.index_part {
2261 0 : Ok(index_part) => {
2262 0 : debug!("remote index part exists for timeline {timeline_id}");
2263 0 : index_part
2264 : }
2265 : Err(DownloadError::NotFound) => {
2266 0 : error!(%timeline_id, "index_part not found on remote");
2267 0 : return Err(TimelineArchivalError::NotFound);
2268 : }
2269 0 : Err(DownloadError::Cancelled) => return Err(TimelineArchivalError::Cancelled),
2270 0 : Err(e) => {
2271 : // Some (possibly ephemeral) error happened during index_part download.
2272 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
2273 0 : return Err(TimelineArchivalError::Other(
2274 0 : anyhow::Error::new(e).context("downloading index_part from remote storage"),
2275 0 : ));
2276 : }
2277 : };
2278 0 : let index_part = match index_part {
2279 0 : MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
2280 0 : MaybeDeletedIndexPart::Deleted(_index_part) => {
2281 0 : info!("timeline is deleted according to index_part.json");
2282 0 : return Err(TimelineArchivalError::NotFound);
2283 : }
2284 : };
2285 0 : let remote_metadata = index_part.metadata.clone();
2286 0 : let timeline_resources = self.build_timeline_resources(timeline_id);
2287 0 : self.load_remote_timeline(
2288 0 : timeline_id,
2289 0 : index_part,
2290 0 : remote_metadata,
2291 0 : None,
2292 0 : timeline_resources,
2293 0 : LoadTimelineCause::Unoffload,
2294 0 : &ctx,
2295 0 : )
2296 0 : .await
2297 0 : .with_context(|| {
2298 0 : format!(
2299 0 : "failed to load remote timeline {} for tenant {}",
2300 0 : timeline_id, self.tenant_shard_id
2301 : )
2302 0 : })
2303 0 : .map_err(TimelineArchivalError::Other)?;
2304 :
2305 0 : let timeline = {
2306 0 : let timelines = self.timelines.lock().unwrap();
2307 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2308 0 : warn!("timeline not available directly after attach");
2309 : // This is not a panic because no locks are held between `load_remote_timeline`
2310 : // which puts the timeline into timelines, and our look into the timeline map.
2311 0 : return Err(TimelineArchivalError::Other(anyhow::anyhow!(
2312 0 : "timeline not available directly after attach"
2313 0 : )));
2314 : };
2315 0 : let mut offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2316 0 : match offloaded_timelines.remove(&timeline_id) {
2317 0 : Some(offloaded) => {
2318 0 : offloaded.delete_from_ancestor_with_timelines(&timelines);
2319 0 : }
2320 0 : None => warn!("timeline already removed from offloaded timelines"),
2321 : }
2322 :
2323 0 : self.initialize_gc_info(&timelines, &offloaded_timelines, Some(timeline_id));
2324 :
2325 0 : Arc::clone(timeline)
2326 : };
2327 :
2328 : // Upload new list of offloaded timelines to S3
2329 0 : self.maybe_upload_tenant_manifest().await?;
2330 :
2331 : // Activate the timeline (if it makes sense)
2332 0 : if !(timeline.is_broken() || timeline.is_stopping()) {
2333 0 : let background_jobs_can_start = None;
2334 0 : timeline.activate(
2335 0 : self.clone(),
2336 0 : broker_client.clone(),
2337 0 : background_jobs_can_start,
2338 0 : &ctx.with_scope_timeline(&timeline),
2339 0 : );
2340 0 : }
2341 :
2342 0 : info!("timeline unoffloading complete");
2343 0 : Ok(timeline)
2344 0 : }
2345 :
2346 0 : pub(crate) async fn apply_timeline_archival_config(
2347 0 : self: &Arc<Self>,
2348 0 : timeline_id: TimelineId,
2349 0 : new_state: TimelineArchivalState,
2350 0 : broker_client: storage_broker::BrokerClientChannel,
2351 0 : ctx: RequestContext,
2352 0 : ) -> Result<(), TimelineArchivalError> {
2353 0 : info!("setting timeline archival config");
2354 : // First part: figure out what is needed to do, and do validation
2355 0 : let timeline_or_unarchive_offloaded = 'outer: {
2356 0 : let timelines = self.timelines.lock().unwrap();
2357 :
2358 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2359 0 : let offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2360 0 : let Some(offloaded) = offloaded_timelines.get(&timeline_id) else {
2361 0 : return Err(TimelineArchivalError::NotFound);
2362 : };
2363 0 : if new_state == TimelineArchivalState::Archived {
2364 : // It's offloaded already, so nothing to do
2365 0 : return Ok(());
2366 0 : }
2367 0 : if let Some(ancestor_timeline_id) = offloaded.ancestor_timeline_id {
2368 0 : Self::check_ancestor_of_to_be_unarchived_is_not_archived(
2369 0 : ancestor_timeline_id,
2370 0 : &timelines,
2371 0 : &offloaded_timelines,
2372 0 : )?;
2373 0 : }
2374 0 : break 'outer None;
2375 : };
2376 :
2377 : // Do some validation. We release the timelines lock below, so there is potential
2378 : // for race conditions: these checks are more present to prevent misunderstandings of
2379 : // the API's capabilities, instead of serving as the sole way to defend their invariants.
2380 0 : match new_state {
2381 : TimelineArchivalState::Unarchived => {
2382 0 : Self::check_to_be_unarchived_timeline_has_no_archived_parent(timeline)?
2383 : }
2384 : TimelineArchivalState::Archived => {
2385 0 : Self::check_to_be_archived_has_no_unarchived_children(timeline_id, &timelines)?
2386 : }
2387 : }
2388 0 : Some(Arc::clone(timeline))
2389 : };
2390 :
2391 : // Second part: unoffload timeline (if needed)
2392 0 : let timeline = if let Some(timeline) = timeline_or_unarchive_offloaded {
2393 0 : timeline
2394 : } else {
2395 : // Turn offloaded timeline into a non-offloaded one
2396 0 : self.unoffload_timeline(timeline_id, broker_client, ctx)
2397 0 : .await?
2398 : };
2399 :
2400 : // Third part: upload new timeline archival state and block until it is present in S3
2401 0 : let upload_needed = match timeline
2402 0 : .remote_client
2403 0 : .schedule_index_upload_for_timeline_archival_state(new_state)
2404 : {
2405 0 : Ok(upload_needed) => upload_needed,
2406 0 : Err(e) => {
2407 0 : if timeline.cancel.is_cancelled() {
2408 0 : return Err(TimelineArchivalError::Cancelled);
2409 : } else {
2410 0 : return Err(TimelineArchivalError::Other(e));
2411 : }
2412 : }
2413 : };
2414 :
2415 0 : if upload_needed {
2416 0 : info!("Uploading new state");
2417 : const MAX_WAIT: Duration = Duration::from_secs(10);
2418 0 : let Ok(v) =
2419 0 : tokio::time::timeout(MAX_WAIT, timeline.remote_client.wait_completion()).await
2420 : else {
2421 0 : tracing::warn!("reached timeout for waiting on upload queue");
2422 0 : return Err(TimelineArchivalError::Timeout);
2423 : };
2424 0 : v.map_err(|e| match e {
2425 0 : WaitCompletionError::NotInitialized(e) => {
2426 0 : TimelineArchivalError::Other(anyhow::anyhow!(e))
2427 : }
2428 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2429 0 : TimelineArchivalError::Cancelled
2430 : }
2431 0 : })?;
2432 0 : }
2433 0 : Ok(())
2434 0 : }
2435 :
2436 1 : pub fn get_offloaded_timeline(
2437 1 : &self,
2438 1 : timeline_id: TimelineId,
2439 1 : ) -> Result<Arc<OffloadedTimeline>, GetTimelineError> {
2440 1 : self.timelines_offloaded
2441 1 : .lock()
2442 1 : .unwrap()
2443 1 : .get(&timeline_id)
2444 1 : .map(Arc::clone)
2445 1 : .ok_or(GetTimelineError::NotFound {
2446 1 : tenant_id: self.tenant_shard_id,
2447 1 : timeline_id,
2448 1 : })
2449 1 : }
2450 :
2451 2 : pub(crate) fn tenant_shard_id(&self) -> TenantShardId {
2452 2 : self.tenant_shard_id
2453 2 : }
2454 :
2455 : /// Get Timeline handle for given Neon timeline ID.
2456 : /// This function is idempotent. It doesn't change internal state in any way.
2457 111 : pub fn get_timeline(
2458 111 : &self,
2459 111 : timeline_id: TimelineId,
2460 111 : active_only: bool,
2461 111 : ) -> Result<Arc<Timeline>, GetTimelineError> {
2462 111 : let timelines_accessor = self.timelines.lock().unwrap();
2463 111 : let timeline = timelines_accessor
2464 111 : .get(&timeline_id)
2465 111 : .ok_or(GetTimelineError::NotFound {
2466 111 : tenant_id: self.tenant_shard_id,
2467 111 : timeline_id,
2468 111 : })?;
2469 :
2470 110 : if active_only && !timeline.is_active() {
2471 0 : Err(GetTimelineError::NotActive {
2472 0 : tenant_id: self.tenant_shard_id,
2473 0 : timeline_id,
2474 0 : state: timeline.current_state(),
2475 0 : })
2476 : } else {
2477 110 : Ok(Arc::clone(timeline))
2478 : }
2479 111 : }
2480 :
2481 : /// Lists timelines the tenant contains.
2482 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2483 2 : pub fn list_timelines(&self) -> Vec<Arc<Timeline>> {
2484 2 : self.timelines
2485 2 : .lock()
2486 2 : .unwrap()
2487 2 : .values()
2488 2 : .map(Arc::clone)
2489 2 : .collect()
2490 2 : }
2491 :
2492 : /// Lists timelines the tenant contains.
2493 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2494 0 : pub fn list_importing_timelines(&self) -> Vec<Arc<ImportingTimeline>> {
2495 0 : self.timelines_importing
2496 0 : .lock()
2497 0 : .unwrap()
2498 0 : .values()
2499 0 : .map(Arc::clone)
2500 0 : .collect()
2501 0 : }
2502 :
2503 : /// Lists timelines the tenant manages, including offloaded ones.
2504 : ///
2505 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2506 0 : pub fn list_timelines_and_offloaded(
2507 0 : &self,
2508 0 : ) -> (Vec<Arc<Timeline>>, Vec<Arc<OffloadedTimeline>>) {
2509 0 : let timelines = self
2510 0 : .timelines
2511 0 : .lock()
2512 0 : .unwrap()
2513 0 : .values()
2514 0 : .map(Arc::clone)
2515 0 : .collect();
2516 0 : let offloaded = self
2517 0 : .timelines_offloaded
2518 0 : .lock()
2519 0 : .unwrap()
2520 0 : .values()
2521 0 : .map(Arc::clone)
2522 0 : .collect();
2523 0 : (timelines, offloaded)
2524 0 : }
2525 :
2526 0 : pub fn list_timeline_ids(&self) -> Vec<TimelineId> {
2527 0 : self.timelines.lock().unwrap().keys().cloned().collect()
2528 0 : }
2529 :
2530 : /// This is used by tests & import-from-basebackup.
2531 : ///
2532 : /// The returned [`UninitializedTimeline`] contains no data nor metadata and it is in
2533 : /// a state that will fail [`TenantShard::load_remote_timeline`] because `disk_consistent_lsn=Lsn(0)`.
2534 : ///
2535 : /// The caller is responsible for getting the timeline into a state that will be accepted
2536 : /// by [`TenantShard::load_remote_timeline`] / [`TenantShard::attach`].
2537 : /// Then they may call [`UninitializedTimeline::finish_creation`] to add the timeline
2538 : /// to the [`TenantShard::timelines`].
2539 : ///
2540 : /// Tests should use `TenantShard::create_test_timeline` to set up the minimum required metadata keys.
2541 114 : pub(crate) async fn create_empty_timeline(
2542 114 : self: &Arc<Self>,
2543 114 : new_timeline_id: TimelineId,
2544 114 : initdb_lsn: Lsn,
2545 114 : pg_version: PgMajorVersion,
2546 114 : ctx: &RequestContext,
2547 114 : ) -> anyhow::Result<(UninitializedTimeline, RequestContext)> {
2548 114 : anyhow::ensure!(
2549 114 : self.is_active(),
2550 0 : "Cannot create empty timelines on inactive tenant"
2551 : );
2552 :
2553 : // Protect against concurrent attempts to use this TimelineId
2554 114 : let create_guard = match self
2555 114 : .start_creating_timeline(new_timeline_id, CreateTimelineIdempotency::FailWithConflict)
2556 114 : .await?
2557 : {
2558 113 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2559 : StartCreatingTimelineResult::Idempotent(_) => {
2560 0 : unreachable!("FailWithConflict implies we get an error instead")
2561 : }
2562 : };
2563 :
2564 113 : let new_metadata = TimelineMetadata::new(
2565 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2566 : // make it valid, before calling finish_creation()
2567 113 : Lsn(0),
2568 113 : None,
2569 113 : None,
2570 113 : Lsn(0),
2571 113 : initdb_lsn,
2572 113 : initdb_lsn,
2573 113 : pg_version,
2574 : );
2575 113 : self.prepare_new_timeline(
2576 113 : new_timeline_id,
2577 113 : &new_metadata,
2578 113 : create_guard,
2579 113 : initdb_lsn,
2580 113 : None,
2581 113 : None,
2582 113 : ctx,
2583 113 : )
2584 113 : .await
2585 114 : }
2586 :
2587 : /// Helper for unit tests to create an empty timeline.
2588 : ///
2589 : /// The timeline is has state value `Active` but its background loops are not running.
2590 : // This makes the various functions which anyhow::ensure! for Active state work in tests.
2591 : // Our current tests don't need the background loops.
2592 : #[cfg(test)]
2593 109 : pub async fn create_test_timeline(
2594 109 : self: &Arc<Self>,
2595 109 : new_timeline_id: TimelineId,
2596 109 : initdb_lsn: Lsn,
2597 109 : pg_version: PgMajorVersion,
2598 109 : ctx: &RequestContext,
2599 109 : ) -> anyhow::Result<Arc<Timeline>> {
2600 109 : let (uninit_tl, ctx) = self
2601 109 : .create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2602 109 : .await?;
2603 109 : let tline = uninit_tl.raw_timeline().expect("we just created it");
2604 109 : assert_eq!(tline.get_last_record_lsn(), Lsn(0));
2605 :
2606 : // Setup minimum keys required for the timeline to be usable.
2607 109 : let mut modification = tline.begin_modification(initdb_lsn);
2608 109 : modification
2609 109 : .init_empty_test_timeline()
2610 109 : .context("init_empty_test_timeline")?;
2611 109 : modification
2612 109 : .commit(&ctx)
2613 109 : .await
2614 109 : .context("commit init_empty_test_timeline modification")?;
2615 :
2616 : // Flush to disk so that uninit_tl's check for valid disk_consistent_lsn passes.
2617 109 : tline.maybe_spawn_flush_loop();
2618 109 : tline.freeze_and_flush().await.context("freeze_and_flush")?;
2619 :
2620 : // Make sure the freeze_and_flush reaches remote storage.
2621 109 : tline.remote_client.wait_completion().await.unwrap();
2622 :
2623 109 : let tl = uninit_tl.finish_creation().await?;
2624 : // The non-test code would call tl.activate() here.
2625 109 : tl.set_state(TimelineState::Active);
2626 109 : Ok(tl)
2627 109 : }
2628 :
2629 : /// Helper for unit tests to create a timeline with some pre-loaded states.
2630 : #[cfg(test)]
2631 : #[allow(clippy::too_many_arguments)]
2632 24 : pub async fn create_test_timeline_with_layers(
2633 24 : self: &Arc<Self>,
2634 24 : new_timeline_id: TimelineId,
2635 24 : initdb_lsn: Lsn,
2636 24 : pg_version: PgMajorVersion,
2637 24 : ctx: &RequestContext,
2638 24 : in_memory_layer_desc: Vec<timeline::InMemoryLayerTestDesc>,
2639 24 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
2640 24 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
2641 24 : end_lsn: Lsn,
2642 24 : ) -> anyhow::Result<Arc<Timeline>> {
2643 : use checks::check_valid_layermap;
2644 : use itertools::Itertools;
2645 :
2646 24 : let tline = self
2647 24 : .create_test_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2648 24 : .await?;
2649 24 : tline.force_advance_lsn(end_lsn);
2650 71 : for deltas in delta_layer_desc {
2651 47 : tline
2652 47 : .force_create_delta_layer(deltas, Some(initdb_lsn), ctx)
2653 47 : .await?;
2654 : }
2655 58 : for (lsn, images) in image_layer_desc {
2656 34 : tline
2657 34 : .force_create_image_layer(lsn, images, Some(initdb_lsn), ctx)
2658 34 : .await?;
2659 : }
2660 28 : for in_memory in in_memory_layer_desc {
2661 4 : tline
2662 4 : .force_create_in_memory_layer(in_memory, Some(initdb_lsn), ctx)
2663 4 : .await?;
2664 : }
2665 24 : let layer_names = tline
2666 24 : .layers
2667 24 : .read(LayerManagerLockHolder::Testing)
2668 24 : .await
2669 24 : .layer_map()
2670 24 : .unwrap()
2671 24 : .iter_historic_layers()
2672 105 : .map(|layer| layer.layer_name())
2673 24 : .collect_vec();
2674 24 : if let Some(err) = check_valid_layermap(&layer_names) {
2675 0 : bail!("invalid layermap: {err}");
2676 24 : }
2677 24 : Ok(tline)
2678 24 : }
2679 :
2680 : /// Create a new timeline.
2681 : ///
2682 : /// Returns the new timeline ID and reference to its Timeline object.
2683 : ///
2684 : /// If the caller specified the timeline ID to use (`new_timeline_id`), and timeline with
2685 : /// the same timeline ID already exists, returns CreateTimelineError::AlreadyExists.
2686 : #[allow(clippy::too_many_arguments)]
2687 0 : pub(crate) async fn create_timeline(
2688 0 : self: &Arc<TenantShard>,
2689 0 : params: CreateTimelineParams,
2690 0 : broker_client: storage_broker::BrokerClientChannel,
2691 0 : ctx: &RequestContext,
2692 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
2693 0 : if !self.is_active() {
2694 0 : if matches!(self.current_state(), TenantState::Stopping { .. }) {
2695 0 : return Err(CreateTimelineError::ShuttingDown);
2696 : } else {
2697 0 : return Err(CreateTimelineError::Other(anyhow::anyhow!(
2698 0 : "Cannot create timelines on inactive tenant"
2699 0 : )));
2700 : }
2701 0 : }
2702 :
2703 0 : let _gate = self
2704 0 : .gate
2705 0 : .enter()
2706 0 : .map_err(|_| CreateTimelineError::ShuttingDown)?;
2707 :
2708 0 : let result: CreateTimelineResult = match params {
2709 : CreateTimelineParams::Bootstrap(CreateTimelineParamsBootstrap {
2710 0 : new_timeline_id,
2711 0 : existing_initdb_timeline_id,
2712 0 : pg_version,
2713 : }) => {
2714 0 : self.bootstrap_timeline(
2715 0 : new_timeline_id,
2716 0 : pg_version,
2717 0 : existing_initdb_timeline_id,
2718 0 : ctx,
2719 0 : )
2720 0 : .await?
2721 : }
2722 : CreateTimelineParams::Branch(CreateTimelineParamsBranch {
2723 0 : new_timeline_id,
2724 0 : ancestor_timeline_id,
2725 0 : mut ancestor_start_lsn,
2726 : }) => {
2727 0 : let ancestor_timeline = self
2728 0 : .get_timeline(ancestor_timeline_id, false)
2729 0 : .context("Cannot branch off the timeline that's not present in pageserver")?;
2730 :
2731 : // instead of waiting around, just deny the request because ancestor is not yet
2732 : // ready for other purposes either.
2733 0 : if !ancestor_timeline.is_active() {
2734 0 : return Err(CreateTimelineError::AncestorNotActive);
2735 0 : }
2736 :
2737 0 : if ancestor_timeline.is_archived() == Some(true) {
2738 0 : info!("tried to branch archived timeline");
2739 0 : return Err(CreateTimelineError::AncestorArchived);
2740 0 : }
2741 :
2742 0 : if let Some(lsn) = ancestor_start_lsn.as_mut() {
2743 0 : *lsn = lsn.align();
2744 :
2745 0 : let ancestor_ancestor_lsn = ancestor_timeline.get_ancestor_lsn();
2746 0 : if ancestor_ancestor_lsn > *lsn {
2747 : // can we safely just branch from the ancestor instead?
2748 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
2749 0 : "invalid start lsn {} for ancestor timeline {}: less than timeline ancestor lsn {}",
2750 0 : lsn,
2751 0 : ancestor_timeline_id,
2752 0 : ancestor_ancestor_lsn,
2753 0 : )));
2754 0 : }
2755 :
2756 : // Wait for the WAL to arrive and be processed on the parent branch up
2757 : // to the requested branch point. The repository code itself doesn't
2758 : // require it, but if we start to receive WAL on the new timeline,
2759 : // decoding the new WAL might need to look up previous pages, relation
2760 : // sizes etc. and that would get confused if the previous page versions
2761 : // are not in the repository yet.
2762 0 : ancestor_timeline
2763 0 : .wait_lsn(
2764 0 : *lsn,
2765 0 : timeline::WaitLsnWaiter::Tenant,
2766 0 : timeline::WaitLsnTimeout::Default,
2767 0 : ctx,
2768 0 : )
2769 0 : .await
2770 0 : .map_err(|e| match e {
2771 0 : e @ (WaitLsnError::Timeout(_) | WaitLsnError::BadState { .. }) => {
2772 0 : CreateTimelineError::AncestorLsn(anyhow::anyhow!(e))
2773 : }
2774 0 : WaitLsnError::Shutdown => CreateTimelineError::ShuttingDown,
2775 0 : })?;
2776 0 : }
2777 :
2778 0 : self.branch_timeline(&ancestor_timeline, new_timeline_id, ancestor_start_lsn, ctx)
2779 0 : .await?
2780 : }
2781 0 : CreateTimelineParams::ImportPgdata(params) => {
2782 0 : self.create_timeline_import_pgdata(params, ctx).await?
2783 : }
2784 : };
2785 :
2786 : // At this point we have dropped our guard on [`Self::timelines_creating`], and
2787 : // the timeline is visible in [`Self::timelines`], but it is _not_ durable yet. We must
2788 : // not send a success to the caller until it is. The same applies to idempotent retries.
2789 : //
2790 : // TODO: the timeline is already visible in [`Self::timelines`]; a caller could incorrectly
2791 : // assume that, because they can see the timeline via API, that the creation is done and
2792 : // that it is durable. Ideally, we would keep the timeline hidden (in [`Self::timelines_creating`])
2793 : // until it is durable, e.g., by extending the time we hold the creation guard. This also
2794 : // interacts with UninitializedTimeline and is generally a bit tricky.
2795 : //
2796 : // To re-emphasize: the only correct way to create a timeline is to repeat calling the
2797 : // creation API until it returns success. Only then is durability guaranteed.
2798 0 : info!(creation_result=%result.discriminant(), "waiting for timeline to be durable");
2799 0 : result
2800 0 : .timeline()
2801 0 : .remote_client
2802 0 : .wait_completion()
2803 0 : .await
2804 0 : .map_err(|e| match e {
2805 : WaitCompletionError::NotInitialized(
2806 0 : e, // If the queue is already stopped, it's a shutdown error.
2807 0 : ) if e.is_stopping() => CreateTimelineError::ShuttingDown,
2808 : WaitCompletionError::NotInitialized(_) => {
2809 : // This is a bug: we should never try to wait for uploads before initializing the timeline
2810 0 : debug_assert!(false);
2811 0 : CreateTimelineError::Other(anyhow::anyhow!("timeline not initialized"))
2812 : }
2813 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2814 0 : CreateTimelineError::ShuttingDown
2815 : }
2816 0 : })?;
2817 :
2818 : // The creating task is responsible for activating the timeline.
2819 : // We do this after `wait_completion()` so that we don't spin up tasks that start
2820 : // doing stuff before the IndexPart is durable in S3, which is done by the previous section.
2821 0 : let activated_timeline = match result {
2822 0 : CreateTimelineResult::Created(timeline) => {
2823 0 : timeline.activate(
2824 0 : self.clone(),
2825 0 : broker_client,
2826 0 : None,
2827 0 : &ctx.with_scope_timeline(&timeline),
2828 : );
2829 0 : timeline
2830 : }
2831 0 : CreateTimelineResult::Idempotent(timeline) => {
2832 0 : info!(
2833 0 : "request was deemed idempotent, activation will be done by the creating task"
2834 : );
2835 0 : timeline
2836 : }
2837 0 : CreateTimelineResult::ImportSpawned(timeline) => {
2838 0 : info!(
2839 0 : "import task spawned, timeline will become visible and activated once the import is done"
2840 : );
2841 0 : timeline
2842 : }
2843 : };
2844 :
2845 0 : Ok(activated_timeline)
2846 0 : }
2847 :
2848 : /// The returned [`Arc<Timeline>`] is NOT in the [`TenantShard::timelines`] map until the import
2849 : /// completes in the background. A DIFFERENT [`Arc<Timeline>`] will be inserted into the
2850 : /// [`TenantShard::timelines`] map when the import completes.
2851 : /// We only return an [`Arc<Timeline>`] here so the API handler can create a [`pageserver_api::models::TimelineInfo`]
2852 : /// for the response.
2853 0 : async fn create_timeline_import_pgdata(
2854 0 : self: &Arc<Self>,
2855 0 : params: CreateTimelineParamsImportPgdata,
2856 0 : ctx: &RequestContext,
2857 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
2858 : let CreateTimelineParamsImportPgdata {
2859 0 : new_timeline_id,
2860 0 : location,
2861 0 : idempotency_key,
2862 0 : } = params;
2863 :
2864 0 : let started_at = chrono::Utc::now().naive_utc();
2865 :
2866 : //
2867 : // There's probably a simpler way to upload an index part, but, remote_timeline_client
2868 : // is the canonical way we do it.
2869 : // - create an empty timeline in-memory
2870 : // - use its remote_timeline_client to do the upload
2871 : // - dispose of the uninit timeline
2872 : // - keep the creation guard alive
2873 :
2874 0 : let timeline_create_guard = match self
2875 0 : .start_creating_timeline(
2876 0 : new_timeline_id,
2877 0 : CreateTimelineIdempotency::ImportPgdata(CreatingTimelineIdempotencyImportPgdata {
2878 0 : idempotency_key: idempotency_key.clone(),
2879 0 : }),
2880 0 : )
2881 0 : .await?
2882 : {
2883 0 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2884 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
2885 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
2886 : }
2887 : };
2888 :
2889 0 : let (mut uninit_timeline, timeline_ctx) = {
2890 0 : let this = &self;
2891 0 : let initdb_lsn = Lsn(0);
2892 0 : async move {
2893 0 : let new_metadata = TimelineMetadata::new(
2894 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2895 : // make it valid, before calling finish_creation()
2896 0 : Lsn(0),
2897 0 : None,
2898 0 : None,
2899 0 : Lsn(0),
2900 0 : initdb_lsn,
2901 0 : initdb_lsn,
2902 0 : PgMajorVersion::PG15,
2903 : );
2904 0 : this.prepare_new_timeline(
2905 0 : new_timeline_id,
2906 0 : &new_metadata,
2907 0 : timeline_create_guard,
2908 0 : initdb_lsn,
2909 0 : None,
2910 0 : None,
2911 0 : ctx,
2912 0 : )
2913 0 : .await
2914 0 : }
2915 : }
2916 0 : .await?;
2917 :
2918 0 : let in_progress = import_pgdata::index_part_format::InProgress {
2919 0 : idempotency_key,
2920 0 : location,
2921 0 : started_at,
2922 0 : };
2923 0 : let index_part = import_pgdata::index_part_format::Root::V1(
2924 0 : import_pgdata::index_part_format::V1::InProgress(in_progress),
2925 0 : );
2926 0 : uninit_timeline
2927 0 : .raw_timeline()
2928 0 : .unwrap()
2929 0 : .remote_client
2930 0 : .schedule_index_upload_for_import_pgdata_state_update(Some(index_part.clone()))?;
2931 :
2932 : // wait_completion happens in caller
2933 :
2934 0 : let (timeline, timeline_create_guard) = uninit_timeline.finish_creation_myself();
2935 :
2936 0 : let import_task_gate = Gate::default();
2937 0 : let import_task_guard = import_task_gate.enter().unwrap();
2938 :
2939 0 : let import_task_handle = tokio::spawn(self.clone().create_timeline_import_pgdata_task(
2940 0 : timeline.clone(),
2941 0 : index_part,
2942 0 : timeline_create_guard,
2943 0 : import_task_guard,
2944 0 : timeline_ctx.detached_child(TaskKind::ImportPgdata, DownloadBehavior::Warn),
2945 : ));
2946 :
2947 0 : let prev = self.timelines_importing.lock().unwrap().insert(
2948 0 : timeline.timeline_id,
2949 0 : Arc::new(ImportingTimeline {
2950 0 : timeline: timeline.clone(),
2951 0 : import_task_handle,
2952 0 : import_task_gate,
2953 0 : delete_progress: TimelineDeleteProgress::default(),
2954 0 : }),
2955 0 : );
2956 :
2957 : // Idempotency is enforced higher up the stack
2958 0 : assert!(prev.is_none());
2959 :
2960 : // NB: the timeline doesn't exist in self.timelines at this point
2961 0 : Ok(CreateTimelineResult::ImportSpawned(timeline))
2962 0 : }
2963 :
2964 : /// Finalize the import of a timeline on this shard by marking it complete in
2965 : /// the index part. If the import task hasn't finished yet, returns an error.
2966 : ///
2967 : /// This method is idempotent. If the import was finalized once, the next call
2968 : /// will be a no-op.
2969 0 : pub(crate) async fn finalize_importing_timeline(
2970 0 : &self,
2971 0 : timeline_id: TimelineId,
2972 0 : ) -> Result<(), FinalizeTimelineImportError> {
2973 0 : let timeline = {
2974 0 : let locked = self.timelines_importing.lock().unwrap();
2975 0 : match locked.get(&timeline_id) {
2976 0 : Some(importing_timeline) => {
2977 0 : if !importing_timeline.import_task_handle.is_finished() {
2978 0 : return Err(FinalizeTimelineImportError::ImportTaskStillRunning);
2979 0 : }
2980 :
2981 0 : importing_timeline.timeline.clone()
2982 : }
2983 : None => {
2984 0 : return Ok(());
2985 : }
2986 : }
2987 : };
2988 :
2989 0 : timeline
2990 0 : .remote_client
2991 0 : .schedule_index_upload_for_import_pgdata_finalize()
2992 0 : .map_err(|_err| FinalizeTimelineImportError::ShuttingDown)?;
2993 0 : timeline
2994 0 : .remote_client
2995 0 : .wait_completion()
2996 0 : .await
2997 0 : .map_err(|_err| FinalizeTimelineImportError::ShuttingDown)?;
2998 :
2999 0 : self.timelines_importing
3000 0 : .lock()
3001 0 : .unwrap()
3002 0 : .remove(&timeline_id);
3003 :
3004 0 : Ok(())
3005 0 : }
3006 :
3007 : #[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))]
3008 : async fn create_timeline_import_pgdata_task(
3009 : self: Arc<TenantShard>,
3010 : timeline: Arc<Timeline>,
3011 : index_part: import_pgdata::index_part_format::Root,
3012 : timeline_create_guard: TimelineCreateGuard,
3013 : _import_task_guard: GateGuard,
3014 : ctx: RequestContext,
3015 : ) {
3016 : debug_assert_current_span_has_tenant_and_timeline_id();
3017 : info!("starting");
3018 : scopeguard::defer! {info!("exiting")};
3019 :
3020 : let res = self
3021 : .create_timeline_import_pgdata_task_impl(
3022 : timeline,
3023 : index_part,
3024 : timeline_create_guard,
3025 : ctx,
3026 : )
3027 : .await;
3028 : if let Err(err) = &res {
3029 : error!(?err, "task failed");
3030 : // TODO sleep & retry, sensitive to tenant shutdown
3031 : // TODO: allow timeline deletion requests => should cancel the task
3032 : }
3033 : }
3034 :
3035 0 : async fn create_timeline_import_pgdata_task_impl(
3036 0 : self: Arc<TenantShard>,
3037 0 : timeline: Arc<Timeline>,
3038 0 : index_part: import_pgdata::index_part_format::Root,
3039 0 : _timeline_create_guard: TimelineCreateGuard,
3040 0 : ctx: RequestContext,
3041 0 : ) -> Result<(), anyhow::Error> {
3042 0 : info!("importing pgdata");
3043 0 : let ctx = ctx.with_scope_timeline(&timeline);
3044 0 : import_pgdata::doit(&timeline, index_part, &ctx, self.cancel.clone())
3045 0 : .await
3046 0 : .context("import")?;
3047 0 : info!("import done - waiting for activation");
3048 :
3049 0 : anyhow::Ok(())
3050 0 : }
3051 :
3052 0 : pub(crate) async fn delete_timeline(
3053 0 : self: Arc<Self>,
3054 0 : timeline_id: TimelineId,
3055 0 : ) -> Result<(), DeleteTimelineError> {
3056 0 : DeleteTimelineFlow::run(&self, timeline_id).await?;
3057 :
3058 0 : Ok(())
3059 0 : }
3060 :
3061 : /// perform one garbage collection iteration, removing old data files from disk.
3062 : /// this function is periodically called by gc task.
3063 : /// also it can be explicitly requested through page server api 'do_gc' command.
3064 : ///
3065 : /// `target_timeline_id` specifies the timeline to GC, or None for all.
3066 : ///
3067 : /// The `horizon` an `pitr` parameters determine how much WAL history needs to be retained.
3068 : /// Also known as the retention period, or the GC cutoff point. `horizon` specifies
3069 : /// the amount of history, as LSN difference from current latest LSN on each timeline.
3070 : /// `pitr` specifies the same as a time difference from the current time. The effective
3071 : /// GC cutoff point is determined conservatively by either `horizon` and `pitr`, whichever
3072 : /// requires more history to be retained.
3073 : //
3074 377 : pub(crate) async fn gc_iteration(
3075 377 : &self,
3076 377 : target_timeline_id: Option<TimelineId>,
3077 377 : horizon: u64,
3078 377 : pitr: Duration,
3079 377 : cancel: &CancellationToken,
3080 377 : ctx: &RequestContext,
3081 377 : ) -> Result<GcResult, GcError> {
3082 : // Don't start doing work during shutdown
3083 377 : if let TenantState::Stopping { .. } = self.current_state() {
3084 0 : return Ok(GcResult::default());
3085 377 : }
3086 :
3087 : // there is a global allowed_error for this
3088 377 : if !self.is_active() {
3089 0 : return Err(GcError::NotActive);
3090 377 : }
3091 :
3092 : {
3093 377 : let conf = self.tenant_conf.load();
3094 :
3095 : // If we may not delete layers, then simply skip GC. Even though a tenant
3096 : // in AttachedMulti state could do GC and just enqueue the blocked deletions,
3097 : // the only advantage to doing it is to perhaps shrink the LayerMap metadata
3098 : // a bit sooner than we would achieve by waiting for AttachedSingle status.
3099 377 : if !conf.location.may_delete_layers_hint() {
3100 0 : info!("Skipping GC in location state {:?}", conf.location);
3101 0 : return Ok(GcResult::default());
3102 377 : }
3103 :
3104 377 : if conf.is_gc_blocked_by_lsn_lease_deadline() {
3105 375 : info!("Skipping GC because lsn lease deadline is not reached");
3106 375 : return Ok(GcResult::default());
3107 2 : }
3108 : }
3109 :
3110 2 : let _guard = match self.gc_block.start().await {
3111 2 : Ok(guard) => guard,
3112 0 : Err(reasons) => {
3113 0 : info!("Skipping GC: {reasons}");
3114 0 : return Ok(GcResult::default());
3115 : }
3116 : };
3117 :
3118 2 : self.gc_iteration_internal(target_timeline_id, horizon, pitr, cancel, ctx)
3119 2 : .await
3120 377 : }
3121 :
3122 : /// Performs one compaction iteration. Called periodically from the compaction loop. Returns
3123 : /// whether another compaction is needed, if we still have pending work or if we yield for
3124 : /// immediate L0 compaction.
3125 : ///
3126 : /// Compaction can also be explicitly requested for a timeline via the HTTP API.
3127 0 : async fn compaction_iteration(
3128 0 : self: &Arc<Self>,
3129 0 : cancel: &CancellationToken,
3130 0 : ctx: &RequestContext,
3131 0 : ) -> Result<CompactionOutcome, CompactionError> {
3132 : // Don't compact inactive tenants.
3133 0 : if !self.is_active() {
3134 0 : return Ok(CompactionOutcome::Skipped);
3135 0 : }
3136 :
3137 : // Don't compact tenants that can't upload layers. We don't check `may_delete_layers_hint`,
3138 : // since we need to compact L0 even in AttachedMulti to bound read amplification.
3139 0 : let location = self.tenant_conf.load().location;
3140 0 : if !location.may_upload_layers_hint() {
3141 0 : info!("skipping compaction in location state {location:?}");
3142 0 : return Ok(CompactionOutcome::Skipped);
3143 0 : }
3144 :
3145 : // Don't compact if the circuit breaker is tripped.
3146 0 : if self.compaction_circuit_breaker.lock().unwrap().is_broken() {
3147 0 : info!("skipping compaction due to previous failures");
3148 0 : return Ok(CompactionOutcome::Skipped);
3149 0 : }
3150 :
3151 : // Collect all timelines to compact, along with offload instructions and L0 counts.
3152 0 : let mut compact: Vec<Arc<Timeline>> = Vec::new();
3153 0 : let mut offload: HashSet<TimelineId> = HashSet::new();
3154 0 : let mut l0_counts: HashMap<TimelineId, usize> = HashMap::new();
3155 :
3156 : {
3157 0 : let offload_enabled = self.get_timeline_offloading_enabled();
3158 0 : let timelines = self.timelines.lock().unwrap();
3159 0 : for (&timeline_id, timeline) in timelines.iter() {
3160 : // Skip inactive timelines.
3161 0 : if !timeline.is_active() {
3162 0 : continue;
3163 0 : }
3164 :
3165 : // Schedule the timeline for compaction.
3166 0 : compact.push(timeline.clone());
3167 :
3168 : // Schedule the timeline for offloading if eligible.
3169 0 : let can_offload = offload_enabled
3170 0 : && timeline.can_offload().0
3171 0 : && !timelines
3172 0 : .iter()
3173 0 : .any(|(_, tli)| tli.get_ancestor_timeline_id() == Some(timeline_id));
3174 0 : if can_offload {
3175 0 : offload.insert(timeline_id);
3176 0 : }
3177 : }
3178 : } // release timelines lock
3179 :
3180 0 : for timeline in &compact {
3181 : // Collect L0 counts. Can't await while holding lock above.
3182 0 : if let Ok(lm) = timeline
3183 0 : .layers
3184 0 : .read(LayerManagerLockHolder::Compaction)
3185 0 : .await
3186 0 : .layer_map()
3187 0 : {
3188 0 : l0_counts.insert(timeline.timeline_id, lm.level0_deltas().len());
3189 0 : }
3190 : }
3191 :
3192 : // Pass 1: L0 compaction across all timelines, in order of L0 count. We prioritize this to
3193 : // bound read amplification.
3194 : //
3195 : // TODO: this may spin on one or more ingest-heavy timelines, starving out image/GC
3196 : // compaction and offloading. We leave that as a potential problem to solve later. Consider
3197 : // splitting L0 and image/GC compaction to separate background jobs.
3198 0 : if self.get_compaction_l0_first() {
3199 0 : let compaction_threshold = self.get_compaction_threshold();
3200 0 : let compact_l0 = compact
3201 0 : .iter()
3202 0 : .map(|tli| (tli, l0_counts.get(&tli.timeline_id).copied().unwrap_or(0)))
3203 0 : .filter(|&(_, l0)| l0 >= compaction_threshold)
3204 0 : .sorted_by_key(|&(_, l0)| l0)
3205 0 : .rev()
3206 0 : .map(|(tli, _)| tli.clone())
3207 0 : .collect_vec();
3208 :
3209 0 : let mut has_pending_l0 = false;
3210 0 : for timeline in compact_l0 {
3211 0 : let ctx = &ctx.with_scope_timeline(&timeline);
3212 : // NB: don't set CompactFlags::YieldForL0, since this is an L0-only compaction pass.
3213 0 : let outcome = timeline
3214 0 : .compact(cancel, CompactFlags::OnlyL0Compaction.into(), ctx)
3215 0 : .instrument(info_span!("compact_timeline", timeline_id = %timeline.timeline_id))
3216 0 : .await
3217 0 : .inspect_err(|err| self.maybe_trip_compaction_breaker(err))?;
3218 0 : match outcome {
3219 0 : CompactionOutcome::Done => {}
3220 0 : CompactionOutcome::Skipped => {}
3221 0 : CompactionOutcome::Pending => has_pending_l0 = true,
3222 0 : CompactionOutcome::YieldForL0 => has_pending_l0 = true,
3223 : }
3224 : }
3225 0 : if has_pending_l0 {
3226 0 : return Ok(CompactionOutcome::YieldForL0); // do another pass
3227 0 : }
3228 0 : }
3229 :
3230 : // Pass 2: image compaction and timeline offloading. If any timelines have accumulated more
3231 : // L0 layers, they may also be compacted here. Image compaction will yield if there is
3232 : // pending L0 compaction on any tenant timeline.
3233 : //
3234 : // TODO: consider ordering timelines by some priority, e.g. time since last full compaction,
3235 : // amount of L1 delta debt or garbage, offload-eligible timelines first, etc.
3236 0 : let mut has_pending = false;
3237 0 : for timeline in compact {
3238 0 : if !timeline.is_active() {
3239 0 : continue;
3240 0 : }
3241 0 : let ctx = &ctx.with_scope_timeline(&timeline);
3242 :
3243 : // Yield for L0 if the separate L0 pass is enabled (otherwise there's no point).
3244 0 : let mut flags = EnumSet::default();
3245 0 : if self.get_compaction_l0_first() {
3246 0 : flags |= CompactFlags::YieldForL0;
3247 0 : }
3248 :
3249 0 : let mut outcome = timeline
3250 0 : .compact(cancel, flags, ctx)
3251 0 : .instrument(info_span!("compact_timeline", timeline_id = %timeline.timeline_id))
3252 0 : .await
3253 0 : .inspect_err(|err| self.maybe_trip_compaction_breaker(err))?;
3254 :
3255 : // If we're done compacting, check the scheduled GC compaction queue for more work.
3256 0 : if outcome == CompactionOutcome::Done {
3257 0 : let queue = {
3258 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3259 0 : guard
3260 0 : .entry(timeline.timeline_id)
3261 0 : .or_insert_with(|| Arc::new(GcCompactionQueue::new()))
3262 0 : .clone()
3263 : };
3264 0 : let gc_compaction_strategy = self
3265 0 : .feature_resolver
3266 0 : .evaluate_multivariate("gc-comapction-strategy")
3267 0 : .ok();
3268 0 : let span = if let Some(gc_compaction_strategy) = gc_compaction_strategy {
3269 0 : info_span!("gc_compact_timeline", timeline_id = %timeline.timeline_id, strategy = %gc_compaction_strategy)
3270 : } else {
3271 0 : info_span!("gc_compact_timeline", timeline_id = %timeline.timeline_id)
3272 : };
3273 0 : outcome = queue
3274 0 : .iteration(cancel, ctx, &self.gc_block, &timeline)
3275 0 : .instrument(span)
3276 0 : .await?;
3277 0 : }
3278 :
3279 : // If we're done compacting, offload the timeline if requested.
3280 0 : if outcome == CompactionOutcome::Done && offload.contains(&timeline.timeline_id) {
3281 0 : pausable_failpoint!("before-timeline-auto-offload");
3282 0 : offload_timeline(self, &timeline)
3283 0 : .instrument(info_span!("offload_timeline", timeline_id = %timeline.timeline_id))
3284 0 : .await
3285 0 : .or_else(|err| match err {
3286 : // Ignore this, we likely raced with unarchival.
3287 0 : OffloadError::NotArchived => Ok(()),
3288 0 : OffloadError::AlreadyInProgress => Ok(()),
3289 0 : err => Err(err),
3290 0 : })?;
3291 0 : }
3292 :
3293 0 : match outcome {
3294 0 : CompactionOutcome::Done => {}
3295 0 : CompactionOutcome::Skipped => {}
3296 0 : CompactionOutcome::Pending => has_pending = true,
3297 : // This mostly makes sense when the L0-only pass above is enabled, since there's
3298 : // otherwise no guarantee that we'll start with the timeline that has high L0.
3299 0 : CompactionOutcome::YieldForL0 => return Ok(CompactionOutcome::YieldForL0),
3300 : }
3301 : }
3302 :
3303 : // Success! Untrip the breaker if necessary.
3304 0 : self.compaction_circuit_breaker
3305 0 : .lock()
3306 0 : .unwrap()
3307 0 : .success(&CIRCUIT_BREAKERS_UNBROKEN);
3308 :
3309 0 : match has_pending {
3310 0 : true => Ok(CompactionOutcome::Pending),
3311 0 : false => Ok(CompactionOutcome::Done),
3312 : }
3313 0 : }
3314 :
3315 : /// Trips the compaction circuit breaker if appropriate.
3316 0 : pub(crate) fn maybe_trip_compaction_breaker(&self, err: &CompactionError) {
3317 0 : match err {
3318 0 : err if err.is_cancel() => {}
3319 0 : CompactionError::ShuttingDown => (),
3320 : // Offload failures don't trip the circuit breaker, since they're cheap to retry and
3321 : // shouldn't block compaction.
3322 0 : CompactionError::Offload(_) => {}
3323 0 : CompactionError::CollectKeySpaceError(err) => {
3324 0 : // CollectKeySpaceError::Cancelled and PageRead::Cancelled are handled in `err.is_cancel` branch.
3325 0 : self.compaction_circuit_breaker
3326 0 : .lock()
3327 0 : .unwrap()
3328 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, err);
3329 0 : }
3330 0 : CompactionError::Other(err) => {
3331 0 : self.compaction_circuit_breaker
3332 0 : .lock()
3333 0 : .unwrap()
3334 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, err);
3335 0 : }
3336 0 : CompactionError::AlreadyRunning(_) => {}
3337 : }
3338 0 : }
3339 :
3340 : /// Cancel scheduled compaction tasks
3341 0 : pub(crate) fn cancel_scheduled_compaction(&self, timeline_id: TimelineId) {
3342 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3343 0 : if let Some(q) = guard.get_mut(&timeline_id) {
3344 0 : q.cancel_scheduled();
3345 0 : }
3346 0 : }
3347 :
3348 0 : pub(crate) fn get_scheduled_compaction_tasks(
3349 0 : &self,
3350 0 : timeline_id: TimelineId,
3351 0 : ) -> Vec<CompactInfoResponse> {
3352 0 : let res = {
3353 0 : let guard = self.scheduled_compaction_tasks.lock().unwrap();
3354 0 : guard.get(&timeline_id).map(|q| q.remaining_jobs())
3355 : };
3356 0 : let Some((running, remaining)) = res else {
3357 0 : return Vec::new();
3358 : };
3359 0 : let mut result = Vec::new();
3360 0 : if let Some((id, running)) = running {
3361 0 : result.extend(running.into_compact_info_resp(id, true));
3362 0 : }
3363 0 : for (id, job) in remaining {
3364 0 : result.extend(job.into_compact_info_resp(id, false));
3365 0 : }
3366 0 : result
3367 0 : }
3368 :
3369 : /// Schedule a compaction task for a timeline.
3370 0 : pub(crate) async fn schedule_compaction(
3371 0 : &self,
3372 0 : timeline_id: TimelineId,
3373 0 : options: CompactOptions,
3374 0 : ) -> anyhow::Result<tokio::sync::oneshot::Receiver<()>> {
3375 0 : let (tx, rx) = tokio::sync::oneshot::channel();
3376 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3377 0 : let q = guard
3378 0 : .entry(timeline_id)
3379 0 : .or_insert_with(|| Arc::new(GcCompactionQueue::new()));
3380 0 : q.schedule_manual_compaction(options, Some(tx));
3381 0 : Ok(rx)
3382 0 : }
3383 :
3384 : /// Performs periodic housekeeping, via the tenant housekeeping background task.
3385 0 : async fn housekeeping(&self) {
3386 : // Call through to all timelines to freeze ephemeral layers as needed. This usually happens
3387 : // during ingest, but we don't want idle timelines to hold open layers for too long.
3388 : //
3389 : // We don't do this if the tenant can't upload layers (i.e. it's in stale attachment mode).
3390 : // We don't run compaction in this case either, and don't want to keep flushing tiny L0
3391 : // layers that won't be compacted down.
3392 0 : if self.tenant_conf.load().location.may_upload_layers_hint() {
3393 0 : let timelines = self
3394 0 : .timelines
3395 0 : .lock()
3396 0 : .unwrap()
3397 0 : .values()
3398 0 : .filter(|tli| tli.is_active())
3399 0 : .cloned()
3400 0 : .collect_vec();
3401 :
3402 0 : for timeline in timelines {
3403 0 : timeline.maybe_freeze_ephemeral_layer().await;
3404 : }
3405 0 : }
3406 :
3407 : // Shut down walredo if idle.
3408 : const WALREDO_IDLE_TIMEOUT: Duration = Duration::from_secs(180);
3409 0 : if let Some(ref walredo_mgr) = self.walredo_mgr {
3410 0 : walredo_mgr.maybe_quiesce(WALREDO_IDLE_TIMEOUT);
3411 0 : }
3412 :
3413 : // Update the feature resolver with the latest tenant-spcific data.
3414 0 : self.feature_resolver.update_cached_tenant_properties(self);
3415 0 : }
3416 :
3417 0 : pub fn timeline_has_no_attached_children(&self, timeline_id: TimelineId) -> bool {
3418 0 : let timelines = self.timelines.lock().unwrap();
3419 0 : !timelines
3420 0 : .iter()
3421 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(timeline_id))
3422 0 : }
3423 :
3424 994 : pub fn current_state(&self) -> TenantState {
3425 994 : self.state.borrow().clone()
3426 994 : }
3427 :
3428 613 : pub fn is_active(&self) -> bool {
3429 613 : self.current_state() == TenantState::Active
3430 613 : }
3431 :
3432 0 : pub fn generation(&self) -> Generation {
3433 0 : self.generation
3434 0 : }
3435 :
3436 0 : pub(crate) fn wal_redo_manager_status(&self) -> Option<WalRedoManagerStatus> {
3437 0 : self.walredo_mgr.as_ref().and_then(|mgr| mgr.status())
3438 0 : }
3439 :
3440 : /// Changes tenant status to active, unless shutdown was already requested.
3441 : ///
3442 : /// `background_jobs_can_start` is an optional barrier set to a value during pageserver startup
3443 : /// to delay background jobs. Background jobs can be started right away when None is given.
3444 0 : fn activate(
3445 0 : self: &Arc<Self>,
3446 0 : broker_client: BrokerClientChannel,
3447 0 : background_jobs_can_start: Option<&completion::Barrier>,
3448 0 : ctx: &RequestContext,
3449 0 : ) {
3450 0 : span::debug_assert_current_span_has_tenant_id();
3451 :
3452 0 : let mut activating = false;
3453 0 : self.state.send_modify(|current_state| {
3454 : use pageserver_api::models::ActivatingFrom;
3455 0 : match &*current_state {
3456 : TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
3457 0 : panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {current_state:?}");
3458 : }
3459 0 : TenantState::Attaching => {
3460 0 : *current_state = TenantState::Activating(ActivatingFrom::Attaching);
3461 0 : }
3462 : }
3463 0 : debug!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), "Activating tenant");
3464 0 : activating = true;
3465 : // Continue outside the closure. We need to grab timelines.lock()
3466 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
3467 0 : });
3468 :
3469 0 : if activating {
3470 0 : let timelines_accessor = self.timelines.lock().unwrap();
3471 0 : let timelines_offloaded_accessor = self.timelines_offloaded.lock().unwrap();
3472 0 : let timelines_to_activate = timelines_accessor
3473 0 : .values()
3474 0 : .filter(|timeline| !(timeline.is_broken() || timeline.is_stopping()));
3475 :
3476 : // Spawn gc and compaction loops. The loops will shut themselves
3477 : // down when they notice that the tenant is inactive.
3478 0 : tasks::start_background_loops(self, background_jobs_can_start);
3479 :
3480 0 : let mut activated_timelines = 0;
3481 :
3482 0 : for timeline in timelines_to_activate {
3483 0 : timeline.activate(
3484 0 : self.clone(),
3485 0 : broker_client.clone(),
3486 0 : background_jobs_can_start,
3487 0 : &ctx.with_scope_timeline(timeline),
3488 0 : );
3489 0 : activated_timelines += 1;
3490 0 : }
3491 :
3492 0 : let tid = self.tenant_shard_id.tenant_id.to_string();
3493 0 : let shard_id = self.tenant_shard_id.shard_slug().to_string();
3494 0 : let offloaded_timeline_count = timelines_offloaded_accessor.len();
3495 0 : TENANT_OFFLOADED_TIMELINES
3496 0 : .with_label_values(&[&tid, &shard_id])
3497 0 : .set(offloaded_timeline_count as u64);
3498 :
3499 0 : self.state.send_modify(move |current_state| {
3500 0 : assert!(
3501 0 : matches!(current_state, TenantState::Activating(_)),
3502 0 : "set_stopping and set_broken wait for us to leave Activating state",
3503 : );
3504 0 : *current_state = TenantState::Active;
3505 :
3506 0 : let elapsed = self.constructed_at.elapsed();
3507 0 : let total_timelines = timelines_accessor.len();
3508 :
3509 : // log a lot of stuff, because some tenants sometimes suffer from user-visible
3510 : // times to activate. see https://github.com/neondatabase/neon/issues/4025
3511 0 : info!(
3512 0 : since_creation_millis = elapsed.as_millis(),
3513 0 : tenant_id = %self.tenant_shard_id.tenant_id,
3514 0 : shard_id = %self.tenant_shard_id.shard_slug(),
3515 : activated_timelines,
3516 : total_timelines,
3517 0 : post_state = <&'static str>::from(&*current_state),
3518 0 : "activation attempt finished"
3519 : );
3520 :
3521 0 : TENANT.activation.observe(elapsed.as_secs_f64());
3522 0 : });
3523 0 : }
3524 0 : }
3525 :
3526 : /// Shutdown the tenant and join all of the spawned tasks.
3527 : ///
3528 : /// The method caters for all use-cases:
3529 : /// - pageserver shutdown (freeze_and_flush == true)
3530 : /// - detach + ignore (freeze_and_flush == false)
3531 : ///
3532 : /// This will attempt to shutdown even if tenant is broken.
3533 : ///
3534 : /// `shutdown_progress` is a [`completion::Barrier`] for the shutdown initiated by this call.
3535 : /// If the tenant is already shutting down, we return a clone of the first shutdown call's
3536 : /// `Barrier` as an `Err`. This not-first caller can use the returned barrier to join with
3537 : /// the ongoing shutdown.
3538 3 : async fn shutdown(
3539 3 : &self,
3540 3 : shutdown_progress: completion::Barrier,
3541 3 : shutdown_mode: timeline::ShutdownMode,
3542 3 : ) -> Result<(), completion::Barrier> {
3543 3 : span::debug_assert_current_span_has_tenant_id();
3544 :
3545 : // Set tenant (and its timlines) to Stoppping state.
3546 : //
3547 : // Since we can only transition into Stopping state after activation is complete,
3548 : // run it in a JoinSet so all tenants have a chance to stop before we get SIGKILLed.
3549 : //
3550 : // Transitioning tenants to Stopping state has a couple of non-obvious side effects:
3551 : // 1. Lock out any new requests to the tenants.
3552 : // 2. Signal cancellation to WAL receivers (we wait on it below).
3553 : // 3. Signal cancellation for other tenant background loops.
3554 : // 4. ???
3555 : //
3556 : // The waiting for the cancellation is not done uniformly.
3557 : // We certainly wait for WAL receivers to shut down.
3558 : // That is necessary so that no new data comes in before the freeze_and_flush.
3559 : // But the tenant background loops are joined-on in our caller.
3560 : // It's mesed up.
3561 : // we just ignore the failure to stop
3562 :
3563 : // If we're still attaching, fire the cancellation token early to drop out: this
3564 : // will prevent us flushing, but ensures timely shutdown if some I/O during attach
3565 : // is very slow.
3566 3 : let shutdown_mode = if matches!(self.current_state(), TenantState::Attaching) {
3567 0 : self.cancel.cancel();
3568 :
3569 : // Having fired our cancellation token, do not try and flush timelines: their cancellation tokens
3570 : // are children of ours, so their flush loops will have shut down already
3571 0 : timeline::ShutdownMode::Hard
3572 : } else {
3573 3 : shutdown_mode
3574 : };
3575 :
3576 3 : match self.set_stopping(shutdown_progress).await {
3577 3 : Ok(()) => {}
3578 0 : Err(SetStoppingError::Broken) => {
3579 0 : // assume that this is acceptable
3580 0 : }
3581 0 : Err(SetStoppingError::AlreadyStopping(other)) => {
3582 : // give caller the option to wait for this this shutdown
3583 0 : info!("Tenant::shutdown: AlreadyStopping");
3584 0 : return Err(other);
3585 : }
3586 : };
3587 :
3588 3 : let mut js = tokio::task::JoinSet::new();
3589 : {
3590 3 : let timelines = self.timelines.lock().unwrap();
3591 3 : timelines.values().for_each(|timeline| {
3592 3 : let timeline = Arc::clone(timeline);
3593 3 : let timeline_id = timeline.timeline_id;
3594 3 : let span = tracing::info_span!("timeline_shutdown", %timeline_id, ?shutdown_mode);
3595 3 : js.spawn(async move { timeline.shutdown(shutdown_mode).instrument(span).await });
3596 3 : });
3597 : }
3598 : {
3599 3 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
3600 3 : timelines_offloaded.values().for_each(|timeline| {
3601 0 : timeline.defuse_for_tenant_drop();
3602 0 : });
3603 : }
3604 : {
3605 3 : let mut timelines_importing = self.timelines_importing.lock().unwrap();
3606 3 : timelines_importing
3607 3 : .drain()
3608 3 : .for_each(|(timeline_id, importing_timeline)| {
3609 0 : let span = tracing::info_span!("importing_timeline_shutdown", %timeline_id);
3610 0 : js.spawn(async move { importing_timeline.shutdown().instrument(span).await });
3611 0 : });
3612 : }
3613 : // test_long_timeline_create_then_tenant_delete is leaning on this message
3614 3 : tracing::info!("Waiting for timelines...");
3615 6 : while let Some(res) = js.join_next().await {
3616 0 : match res {
3617 3 : Ok(()) => {}
3618 0 : Err(je) if je.is_cancelled() => unreachable!("no cancelling used"),
3619 0 : Err(je) if je.is_panic() => { /* logged already */ }
3620 0 : Err(je) => warn!("unexpected JoinError: {je:?}"),
3621 : }
3622 : }
3623 :
3624 3 : if let ShutdownMode::Reload = shutdown_mode {
3625 0 : tracing::info!("Flushing deletion queue");
3626 0 : if let Err(e) = self.deletion_queue_client.flush().await {
3627 0 : match e {
3628 0 : DeletionQueueError::ShuttingDown => {
3629 0 : // This is the only error we expect for now. In the future, if more error
3630 0 : // variants are added, we should handle them here.
3631 0 : }
3632 : }
3633 0 : }
3634 3 : }
3635 :
3636 : // We cancel the Tenant's cancellation token _after_ the timelines have all shut down. This permits
3637 : // them to continue to do work during their shutdown methods, e.g. flushing data.
3638 3 : tracing::debug!("Cancelling CancellationToken");
3639 3 : self.cancel.cancel();
3640 :
3641 : // shutdown all tenant and timeline tasks: gc, compaction, page service
3642 : // No new tasks will be started for this tenant because it's in `Stopping` state.
3643 : //
3644 : // this will additionally shutdown and await all timeline tasks.
3645 3 : tracing::debug!("Waiting for tasks...");
3646 3 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), None).await;
3647 :
3648 3 : if let Some(walredo_mgr) = self.walredo_mgr.as_ref() {
3649 3 : walredo_mgr.shutdown().await;
3650 0 : }
3651 :
3652 : // Wait for any in-flight operations to complete
3653 3 : self.gate.close().await;
3654 :
3655 3 : remove_tenant_metrics(&self.tenant_shard_id);
3656 :
3657 3 : Ok(())
3658 3 : }
3659 :
3660 : /// Change tenant status to Stopping, to mark that it is being shut down.
3661 : ///
3662 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
3663 : ///
3664 : /// This function is not cancel-safe!
3665 3 : async fn set_stopping(&self, progress: completion::Barrier) -> Result<(), SetStoppingError> {
3666 3 : let mut rx = self.state.subscribe();
3667 :
3668 : // cannot stop before we're done activating, so wait out until we're done activating
3669 3 : rx.wait_for(|state| match state {
3670 : TenantState::Activating(_) | TenantState::Attaching => {
3671 0 : info!("waiting for {state} to turn Active|Broken|Stopping");
3672 0 : false
3673 : }
3674 3 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3675 3 : })
3676 3 : .await
3677 3 : .expect("cannot drop self.state while on a &self method");
3678 :
3679 : // we now know we're done activating, let's see whether this task is the winner to transition into Stopping
3680 3 : let mut err = None;
3681 3 : let stopping = self.state.send_if_modified(|current_state| match current_state {
3682 : TenantState::Activating(_) | TenantState::Attaching => {
3683 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
3684 : }
3685 : TenantState::Active => {
3686 : // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
3687 : // are created after the transition to Stopping. That's harmless, as the Timelines
3688 : // won't be accessible to anyone afterwards, because the Tenant is in Stopping state.
3689 3 : *current_state = TenantState::Stopping { progress: Some(progress) };
3690 : // Continue stopping outside the closure. We need to grab timelines.lock()
3691 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
3692 3 : true
3693 : }
3694 : TenantState::Stopping { progress: None } => {
3695 : // An attach was cancelled, and the attach transitioned the tenant from Attaching to
3696 : // Stopping(None) to let us know it exited. Register our progress and continue.
3697 0 : *current_state = TenantState::Stopping { progress: Some(progress) };
3698 0 : true
3699 : }
3700 0 : TenantState::Broken { reason, .. } => {
3701 0 : info!(
3702 0 : "Cannot set tenant to Stopping state, it is in Broken state due to: {reason}"
3703 : );
3704 0 : err = Some(SetStoppingError::Broken);
3705 0 : false
3706 : }
3707 0 : TenantState::Stopping { progress: Some(progress) } => {
3708 0 : info!("Tenant is already in Stopping state");
3709 0 : err = Some(SetStoppingError::AlreadyStopping(progress.clone()));
3710 0 : false
3711 : }
3712 3 : });
3713 3 : match (stopping, err) {
3714 3 : (true, None) => {} // continue
3715 0 : (false, Some(err)) => return Err(err),
3716 0 : (true, Some(_)) => unreachable!(
3717 : "send_if_modified closure must error out if not transitioning to Stopping"
3718 : ),
3719 0 : (false, None) => unreachable!(
3720 : "send_if_modified closure must return true if transitioning to Stopping"
3721 : ),
3722 : }
3723 :
3724 3 : let timelines_accessor = self.timelines.lock().unwrap();
3725 3 : let not_broken_timelines = timelines_accessor
3726 3 : .values()
3727 3 : .filter(|timeline| !timeline.is_broken());
3728 6 : for timeline in not_broken_timelines {
3729 3 : timeline.set_state(TimelineState::Stopping);
3730 3 : }
3731 3 : Ok(())
3732 3 : }
3733 :
3734 : /// Method for tenant::mgr to transition us into Broken state in case of a late failure in
3735 : /// `remove_tenant_from_memory`
3736 : ///
3737 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
3738 : ///
3739 : /// In tests, we also use this to set tenants to Broken state on purpose.
3740 0 : pub(crate) async fn set_broken(&self, reason: String) {
3741 0 : let mut rx = self.state.subscribe();
3742 :
3743 : // The load & attach routines own the tenant state until it has reached `Active`.
3744 : // So, wait until it's done.
3745 0 : rx.wait_for(|state| match state {
3746 : TenantState::Activating(_) | TenantState::Attaching => {
3747 0 : info!(
3748 0 : "waiting for {} to turn Active|Broken|Stopping",
3749 0 : <&'static str>::from(state)
3750 : );
3751 0 : false
3752 : }
3753 0 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3754 0 : })
3755 0 : .await
3756 0 : .expect("cannot drop self.state while on a &self method");
3757 :
3758 : // we now know we're done activating, let's see whether this task is the winner to transition into Broken
3759 0 : self.set_broken_no_wait(reason)
3760 0 : }
3761 :
3762 0 : pub(crate) fn set_broken_no_wait(&self, reason: impl Display) {
3763 0 : let reason = reason.to_string();
3764 0 : self.state.send_modify(|current_state| {
3765 0 : match *current_state {
3766 : TenantState::Activating(_) | TenantState::Attaching => {
3767 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
3768 : }
3769 : TenantState::Active => {
3770 0 : if cfg!(feature = "testing") {
3771 0 : warn!("Changing Active tenant to Broken state, reason: {}", reason);
3772 0 : *current_state = TenantState::broken_from_reason(reason);
3773 : } else {
3774 0 : unreachable!("not allowed to call set_broken on Active tenants in non-testing builds")
3775 : }
3776 : }
3777 : TenantState::Broken { .. } => {
3778 0 : warn!("Tenant is already in Broken state");
3779 : }
3780 : // This is the only "expected" path, any other path is a bug.
3781 : TenantState::Stopping { .. } => {
3782 0 : warn!(
3783 0 : "Marking Stopping tenant as Broken state, reason: {}",
3784 : reason
3785 : );
3786 0 : *current_state = TenantState::broken_from_reason(reason);
3787 : }
3788 : }
3789 0 : });
3790 0 : }
3791 :
3792 0 : pub fn subscribe_for_state_updates(&self) -> watch::Receiver<TenantState> {
3793 0 : self.state.subscribe()
3794 0 : }
3795 :
3796 : /// The activate_now semaphore is initialized with zero units. As soon as
3797 : /// we add a unit, waiters will be able to acquire a unit and proceed.
3798 0 : pub(crate) fn activate_now(&self) {
3799 0 : self.activate_now_sem.add_permits(1);
3800 0 : }
3801 :
3802 0 : pub(crate) async fn wait_to_become_active(
3803 0 : &self,
3804 0 : timeout: Duration,
3805 0 : ) -> Result<(), GetActiveTenantError> {
3806 0 : let mut receiver = self.state.subscribe();
3807 : loop {
3808 0 : let current_state = receiver.borrow_and_update().clone();
3809 0 : match current_state {
3810 : TenantState::Attaching | TenantState::Activating(_) => {
3811 : // in these states, there's a chance that we can reach ::Active
3812 0 : self.activate_now();
3813 0 : match timeout_cancellable(timeout, &self.cancel, receiver.changed()).await {
3814 0 : Ok(r) => {
3815 0 : r.map_err(
3816 : |_e: tokio::sync::watch::error::RecvError|
3817 : // Tenant existed but was dropped: report it as non-existent
3818 0 : GetActiveTenantError::NotFound(GetTenantError::ShardNotFound(self.tenant_shard_id))
3819 0 : )?
3820 : }
3821 : Err(TimeoutCancellableError::Cancelled) => {
3822 0 : return Err(GetActiveTenantError::Cancelled);
3823 : }
3824 : Err(TimeoutCancellableError::Timeout) => {
3825 0 : return Err(GetActiveTenantError::WaitForActiveTimeout {
3826 0 : latest_state: Some(self.current_state()),
3827 0 : wait_time: timeout,
3828 0 : });
3829 : }
3830 : }
3831 : }
3832 : TenantState::Active => {
3833 0 : return Ok(());
3834 : }
3835 0 : TenantState::Broken { reason, .. } => {
3836 : // This is fatal, and reported distinctly from the general case of "will never be active" because
3837 : // it's logically a 500 to external API users (broken is always a bug).
3838 0 : return Err(GetActiveTenantError::Broken(reason));
3839 : }
3840 : TenantState::Stopping { .. } => {
3841 : // There's no chance the tenant can transition back into ::Active
3842 0 : return Err(GetActiveTenantError::WillNotBecomeActive(current_state));
3843 : }
3844 : }
3845 : }
3846 0 : }
3847 :
3848 0 : pub(crate) fn get_attach_mode(&self) -> AttachmentMode {
3849 0 : self.tenant_conf.load().location.attach_mode
3850 0 : }
3851 :
3852 : /// For API access: generate a LocationConfig equivalent to the one that would be used to
3853 : /// create a Tenant in the same state. Do not use this in hot paths: it's for relatively
3854 : /// rare external API calls, like a reconciliation at startup.
3855 0 : pub(crate) fn get_location_conf(&self) -> models::LocationConfig {
3856 0 : let attached_tenant_conf = self.tenant_conf.load();
3857 :
3858 0 : let location_config_mode = match attached_tenant_conf.location.attach_mode {
3859 0 : AttachmentMode::Single => models::LocationConfigMode::AttachedSingle,
3860 0 : AttachmentMode::Multi => models::LocationConfigMode::AttachedMulti,
3861 0 : AttachmentMode::Stale => models::LocationConfigMode::AttachedStale,
3862 : };
3863 :
3864 0 : models::LocationConfig {
3865 0 : mode: location_config_mode,
3866 0 : generation: self.generation.into(),
3867 0 : secondary_conf: None,
3868 0 : shard_number: self.shard_identity.number.0,
3869 0 : shard_count: self.shard_identity.count.literal(),
3870 0 : shard_stripe_size: self.shard_identity.stripe_size.0,
3871 0 : tenant_conf: attached_tenant_conf.tenant_conf.clone(),
3872 0 : }
3873 0 : }
3874 :
3875 0 : pub(crate) fn get_tenant_shard_id(&self) -> &TenantShardId {
3876 0 : &self.tenant_shard_id
3877 0 : }
3878 :
3879 0 : pub(crate) fn get_shard_identity(&self) -> ShardIdentity {
3880 0 : self.shard_identity
3881 0 : }
3882 :
3883 119 : pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
3884 119 : self.shard_identity.stripe_size
3885 119 : }
3886 :
3887 0 : pub(crate) fn get_generation(&self) -> Generation {
3888 0 : self.generation
3889 0 : }
3890 :
3891 : /// This function partially shuts down the tenant (it shuts down the Timelines) and is fallible,
3892 : /// and can leave the tenant in a bad state if it fails. The caller is responsible for
3893 : /// resetting this tenant to a valid state if we fail.
3894 0 : pub(crate) async fn split_prepare(
3895 0 : &self,
3896 0 : child_shards: &Vec<TenantShardId>,
3897 0 : ) -> anyhow::Result<()> {
3898 0 : let (timelines, offloaded) = {
3899 0 : let timelines = self.timelines.lock().unwrap();
3900 0 : let offloaded = self.timelines_offloaded.lock().unwrap();
3901 0 : (timelines.clone(), offloaded.clone())
3902 0 : };
3903 0 : let timelines_iter = timelines
3904 0 : .values()
3905 0 : .map(TimelineOrOffloadedArcRef::<'_>::from)
3906 0 : .chain(
3907 0 : offloaded
3908 0 : .values()
3909 0 : .map(TimelineOrOffloadedArcRef::<'_>::from),
3910 : );
3911 0 : for timeline in timelines_iter {
3912 : // We do not block timeline creation/deletion during splits inside the pageserver: it is up to higher levels
3913 : // to ensure that they do not start a split if currently in the process of doing these.
3914 :
3915 0 : let timeline_id = timeline.timeline_id();
3916 :
3917 0 : if let TimelineOrOffloadedArcRef::Timeline(timeline) = timeline {
3918 : // Upload an index from the parent: this is partly to provide freshness for the
3919 : // child tenants that will copy it, and partly for general ease-of-debugging: there will
3920 : // always be a parent shard index in the same generation as we wrote the child shard index.
3921 0 : tracing::info!(%timeline_id, "Uploading index");
3922 0 : timeline
3923 0 : .remote_client
3924 0 : .schedule_index_upload_for_file_changes()?;
3925 0 : timeline.remote_client.wait_completion().await?;
3926 0 : }
3927 :
3928 0 : let remote_client = match timeline {
3929 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.remote_client.clone(),
3930 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => {
3931 0 : let remote_client = self
3932 0 : .build_timeline_client(offloaded.timeline_id, self.remote_storage.clone());
3933 0 : Arc::new(remote_client)
3934 : }
3935 : TimelineOrOffloadedArcRef::Importing(_) => {
3936 0 : unreachable!("Importing timelines are not included in the iterator")
3937 : }
3938 : };
3939 :
3940 : // Shut down the timeline's remote client: this means that the indices we write
3941 : // for child shards will not be invalidated by the parent shard deleting layers.
3942 0 : tracing::info!(%timeline_id, "Shutting down remote storage client");
3943 0 : remote_client.shutdown().await;
3944 :
3945 : // Download methods can still be used after shutdown, as they don't flow through the remote client's
3946 : // queue. In principal the RemoteTimelineClient could provide this without downloading it, but this
3947 : // operation is rare, so it's simpler to just download it (and robustly guarantees that the index
3948 : // we use here really is the remotely persistent one).
3949 0 : tracing::info!(%timeline_id, "Downloading index_part from parent");
3950 0 : let result = remote_client
3951 0 : .download_index_file(&self.cancel)
3952 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))
3953 0 : .await?;
3954 0 : let index_part = match result {
3955 : MaybeDeletedIndexPart::Deleted(_) => {
3956 0 : anyhow::bail!("Timeline deletion happened concurrently with split")
3957 : }
3958 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
3959 : };
3960 :
3961 : // A shard split may not take place while a timeline import is on-going
3962 : // for the tenant. Timeline imports run as part of each tenant shard
3963 : // and rely on the sharding scheme to split the work among pageservers.
3964 : // If we were to split in the middle of this process, we would have to
3965 : // either ensure that it's driven to completion on the old shard set
3966 : // or transfer it to the new shard set. It's technically possible, but complex.
3967 0 : match index_part.import_pgdata {
3968 0 : Some(ref import) if !import.is_done() => {
3969 0 : anyhow::bail!(
3970 0 : "Cannot split due to import with idempotency key: {:?}",
3971 0 : import.idempotency_key()
3972 : );
3973 : }
3974 0 : Some(_) | None => {
3975 0 : // fallthrough
3976 0 : }
3977 : }
3978 :
3979 0 : for child_shard in child_shards {
3980 0 : tracing::info!(%timeline_id, "Uploading index_part for child {}", child_shard.to_index());
3981 0 : upload_index_part(
3982 0 : &self.remote_storage,
3983 0 : child_shard,
3984 0 : &timeline_id,
3985 0 : self.generation,
3986 0 : &index_part,
3987 0 : &self.cancel,
3988 0 : )
3989 0 : .await?;
3990 : }
3991 : }
3992 :
3993 0 : let tenant_manifest = self.build_tenant_manifest();
3994 0 : for child_shard in child_shards {
3995 0 : tracing::info!(
3996 0 : "Uploading tenant manifest for child {}",
3997 0 : child_shard.to_index()
3998 : );
3999 0 : upload_tenant_manifest(
4000 0 : &self.remote_storage,
4001 0 : child_shard,
4002 0 : self.generation,
4003 0 : &tenant_manifest,
4004 0 : &self.cancel,
4005 0 : )
4006 0 : .await?;
4007 : }
4008 :
4009 0 : Ok(())
4010 0 : }
4011 :
4012 0 : pub(crate) fn get_sizes(&self) -> TopTenantShardItem {
4013 0 : let mut result = TopTenantShardItem {
4014 0 : id: self.tenant_shard_id,
4015 0 : resident_size: 0,
4016 0 : physical_size: 0,
4017 0 : max_logical_size: 0,
4018 0 : max_logical_size_per_shard: 0,
4019 0 : };
4020 :
4021 0 : for timeline in self.timelines.lock().unwrap().values() {
4022 0 : result.resident_size += timeline.metrics.resident_physical_size_gauge.get();
4023 0 :
4024 0 : result.physical_size += timeline
4025 0 : .remote_client
4026 0 : .metrics
4027 0 : .remote_physical_size_gauge
4028 0 : .get();
4029 0 : result.max_logical_size = std::cmp::max(
4030 0 : result.max_logical_size,
4031 0 : timeline.metrics.current_logical_size_gauge.get(),
4032 0 : );
4033 0 : }
4034 :
4035 0 : result.max_logical_size_per_shard = result
4036 0 : .max_logical_size
4037 0 : .div_ceil(self.tenant_shard_id.shard_count.count() as u64);
4038 :
4039 0 : result
4040 0 : }
4041 : }
4042 :
4043 : /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
4044 : /// perform a topological sort, so that the parent of each timeline comes
4045 : /// before the children.
4046 : /// E extracts the ancestor from T
4047 : /// This allows for T to be different. It can be TimelineMetadata, can be Timeline itself, etc.
4048 118 : fn tree_sort_timelines<T, E>(
4049 118 : timelines: HashMap<TimelineId, T>,
4050 118 : extractor: E,
4051 118 : ) -> anyhow::Result<Vec<(TimelineId, T)>>
4052 118 : where
4053 118 : E: Fn(&T) -> Option<TimelineId>,
4054 : {
4055 118 : let mut result = Vec::with_capacity(timelines.len());
4056 :
4057 118 : let mut now = Vec::with_capacity(timelines.len());
4058 : // (ancestor, children)
4059 118 : let mut later: HashMap<TimelineId, Vec<(TimelineId, T)>> =
4060 118 : HashMap::with_capacity(timelines.len());
4061 :
4062 121 : for (timeline_id, value) in timelines {
4063 3 : if let Some(ancestor_id) = extractor(&value) {
4064 1 : let children = later.entry(ancestor_id).or_default();
4065 1 : children.push((timeline_id, value));
4066 2 : } else {
4067 2 : now.push((timeline_id, value));
4068 2 : }
4069 : }
4070 :
4071 121 : while let Some((timeline_id, metadata)) = now.pop() {
4072 3 : result.push((timeline_id, metadata));
4073 : // All children of this can be loaded now
4074 3 : if let Some(mut children) = later.remove(&timeline_id) {
4075 1 : now.append(&mut children);
4076 2 : }
4077 : }
4078 :
4079 : // All timelines should be visited now. Unless there were timelines with missing ancestors.
4080 118 : if !later.is_empty() {
4081 0 : for (missing_id, orphan_ids) in later {
4082 0 : for (orphan_id, _) in orphan_ids {
4083 0 : error!(
4084 0 : "could not load timeline {orphan_id} because its ancestor timeline {missing_id} could not be loaded"
4085 : );
4086 : }
4087 : }
4088 0 : bail!("could not load tenant because some timelines are missing ancestors");
4089 118 : }
4090 :
4091 118 : Ok(result)
4092 118 : }
4093 :
4094 : impl TenantShard {
4095 0 : pub fn tenant_specific_overrides(&self) -> pageserver_api::models::TenantConfig {
4096 0 : self.tenant_conf.load().tenant_conf.clone()
4097 0 : }
4098 :
4099 0 : pub fn effective_config(&self) -> pageserver_api::config::TenantConfigToml {
4100 0 : self.tenant_specific_overrides()
4101 0 : .merge(self.conf.default_tenant_conf.clone())
4102 0 : }
4103 :
4104 0 : pub fn get_checkpoint_distance(&self) -> u64 {
4105 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4106 0 : tenant_conf
4107 0 : .checkpoint_distance
4108 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
4109 0 : }
4110 :
4111 0 : pub fn get_checkpoint_timeout(&self) -> Duration {
4112 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4113 0 : tenant_conf
4114 0 : .checkpoint_timeout
4115 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
4116 0 : }
4117 :
4118 0 : pub fn get_compaction_target_size(&self) -> u64 {
4119 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4120 0 : tenant_conf
4121 0 : .compaction_target_size
4122 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
4123 0 : }
4124 :
4125 0 : pub fn get_compaction_period(&self) -> Duration {
4126 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4127 0 : tenant_conf
4128 0 : .compaction_period
4129 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
4130 0 : }
4131 :
4132 0 : pub fn get_compaction_threshold(&self) -> usize {
4133 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4134 0 : tenant_conf
4135 0 : .compaction_threshold
4136 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
4137 0 : }
4138 :
4139 0 : pub fn get_rel_size_v2_enabled(&self) -> bool {
4140 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4141 0 : tenant_conf
4142 0 : .rel_size_v2_enabled
4143 0 : .unwrap_or(self.conf.default_tenant_conf.rel_size_v2_enabled)
4144 0 : }
4145 :
4146 0 : pub fn get_compaction_upper_limit(&self) -> usize {
4147 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4148 0 : tenant_conf
4149 0 : .compaction_upper_limit
4150 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_upper_limit)
4151 0 : }
4152 :
4153 0 : pub fn get_compaction_l0_first(&self) -> bool {
4154 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4155 0 : tenant_conf
4156 0 : .compaction_l0_first
4157 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_first)
4158 0 : }
4159 :
4160 120 : pub fn get_gc_horizon(&self) -> u64 {
4161 120 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4162 120 : tenant_conf
4163 120 : .gc_horizon
4164 120 : .unwrap_or(self.conf.default_tenant_conf.gc_horizon)
4165 120 : }
4166 :
4167 0 : pub fn get_gc_period(&self) -> Duration {
4168 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4169 0 : tenant_conf
4170 0 : .gc_period
4171 0 : .unwrap_or(self.conf.default_tenant_conf.gc_period)
4172 0 : }
4173 :
4174 0 : pub fn get_image_creation_threshold(&self) -> usize {
4175 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4176 0 : tenant_conf
4177 0 : .image_creation_threshold
4178 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
4179 0 : }
4180 :
4181 2 : pub fn get_pitr_interval(&self) -> Duration {
4182 2 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4183 2 : tenant_conf
4184 2 : .pitr_interval
4185 2 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
4186 2 : }
4187 :
4188 0 : pub fn get_min_resident_size_override(&self) -> Option<u64> {
4189 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4190 0 : tenant_conf
4191 0 : .min_resident_size_override
4192 0 : .or(self.conf.default_tenant_conf.min_resident_size_override)
4193 0 : }
4194 :
4195 0 : pub fn get_heatmap_period(&self) -> Option<Duration> {
4196 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4197 0 : let heatmap_period = tenant_conf
4198 0 : .heatmap_period
4199 0 : .unwrap_or(self.conf.default_tenant_conf.heatmap_period);
4200 0 : if heatmap_period.is_zero() {
4201 0 : None
4202 : } else {
4203 0 : Some(heatmap_period)
4204 : }
4205 0 : }
4206 :
4207 2 : pub fn get_lsn_lease_length(&self) -> Duration {
4208 2 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4209 2 : tenant_conf
4210 2 : .lsn_lease_length
4211 2 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
4212 2 : }
4213 :
4214 0 : pub fn get_timeline_offloading_enabled(&self) -> bool {
4215 0 : if self.conf.timeline_offloading {
4216 0 : return true;
4217 0 : }
4218 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4219 0 : tenant_conf
4220 0 : .timeline_offloading
4221 0 : .unwrap_or(self.conf.default_tenant_conf.timeline_offloading)
4222 0 : }
4223 :
4224 : /// Generate an up-to-date TenantManifest based on the state of this Tenant.
4225 119 : fn build_tenant_manifest(&self) -> TenantManifest {
4226 : // Collect the offloaded timelines, and sort them for deterministic output.
4227 119 : let offloaded_timelines = self
4228 119 : .timelines_offloaded
4229 119 : .lock()
4230 119 : .unwrap()
4231 119 : .values()
4232 119 : .map(|tli| tli.manifest())
4233 119 : .sorted_by_key(|m| m.timeline_id)
4234 119 : .collect_vec();
4235 :
4236 119 : TenantManifest {
4237 119 : version: LATEST_TENANT_MANIFEST_VERSION,
4238 119 : stripe_size: Some(self.get_shard_stripe_size()),
4239 119 : offloaded_timelines,
4240 119 : }
4241 119 : }
4242 :
4243 0 : pub fn update_tenant_config<
4244 0 : F: Fn(
4245 0 : pageserver_api::models::TenantConfig,
4246 0 : ) -> anyhow::Result<pageserver_api::models::TenantConfig>,
4247 0 : >(
4248 0 : &self,
4249 0 : update: F,
4250 0 : ) -> anyhow::Result<pageserver_api::models::TenantConfig> {
4251 : // Use read-copy-update in order to avoid overwriting the location config
4252 : // state if this races with [`TenantShard::set_new_location_config`]. Note that
4253 : // this race is not possible if both request types come from the storage
4254 : // controller (as they should!) because an exclusive op lock is required
4255 : // on the storage controller side.
4256 :
4257 0 : self.tenant_conf
4258 0 : .try_rcu(|attached_conf| -> Result<_, anyhow::Error> {
4259 0 : Ok(Arc::new(AttachedTenantConf {
4260 0 : tenant_conf: update(attached_conf.tenant_conf.clone())?,
4261 0 : location: attached_conf.location,
4262 0 : lsn_lease_deadline: attached_conf.lsn_lease_deadline,
4263 : }))
4264 0 : })?;
4265 :
4266 0 : let updated = self.tenant_conf.load();
4267 :
4268 0 : self.tenant_conf_updated(&updated.tenant_conf);
4269 : // Don't hold self.timelines.lock() during the notifies.
4270 : // There's no risk of deadlock right now, but there could be if we consolidate
4271 : // mutexes in struct Timeline in the future.
4272 0 : let timelines = self.list_timelines();
4273 0 : for timeline in timelines {
4274 0 : timeline.tenant_conf_updated(&updated);
4275 0 : }
4276 :
4277 0 : Ok(updated.tenant_conf.clone())
4278 0 : }
4279 :
4280 0 : pub(crate) fn set_new_location_config(&self, new_conf: AttachedTenantConf) {
4281 0 : let new_tenant_conf = new_conf.tenant_conf.clone();
4282 :
4283 0 : self.tenant_conf.store(Arc::new(new_conf.clone()));
4284 :
4285 0 : self.tenant_conf_updated(&new_tenant_conf);
4286 : // Don't hold self.timelines.lock() during the notifies.
4287 : // There's no risk of deadlock right now, but there could be if we consolidate
4288 : // mutexes in struct Timeline in the future.
4289 0 : let timelines = self.list_timelines();
4290 0 : for timeline in timelines {
4291 0 : timeline.tenant_conf_updated(&new_conf);
4292 0 : }
4293 0 : }
4294 :
4295 118 : fn get_pagestream_throttle_config(
4296 118 : psconf: &'static PageServerConf,
4297 118 : overrides: &pageserver_api::models::TenantConfig,
4298 118 : ) -> throttle::Config {
4299 118 : overrides
4300 118 : .timeline_get_throttle
4301 118 : .clone()
4302 118 : .unwrap_or(psconf.default_tenant_conf.timeline_get_throttle.clone())
4303 118 : }
4304 :
4305 0 : pub(crate) fn tenant_conf_updated(&self, new_conf: &pageserver_api::models::TenantConfig) {
4306 0 : let conf = Self::get_pagestream_throttle_config(self.conf, new_conf);
4307 0 : self.pagestream_throttle.reconfigure(conf)
4308 0 : }
4309 :
4310 : /// Helper function to create a new Timeline struct.
4311 : ///
4312 : /// The returned Timeline is in Loading state. The caller is responsible for
4313 : /// initializing any on-disk state, and for inserting the Timeline to the 'timelines'
4314 : /// map.
4315 : ///
4316 : /// `validate_ancestor == false` is used when a timeline is created for deletion
4317 : /// and we might not have the ancestor present anymore which is fine for to be
4318 : /// deleted timelines.
4319 : #[allow(clippy::too_many_arguments)]
4320 234 : fn create_timeline_struct(
4321 234 : &self,
4322 234 : new_timeline_id: TimelineId,
4323 234 : new_metadata: &TimelineMetadata,
4324 234 : previous_heatmap: Option<PreviousHeatmap>,
4325 234 : ancestor: Option<Arc<Timeline>>,
4326 234 : resources: TimelineResources,
4327 234 : cause: CreateTimelineCause,
4328 234 : create_idempotency: CreateTimelineIdempotency,
4329 234 : gc_compaction_state: Option<GcCompactionState>,
4330 234 : rel_size_v2_status: Option<RelSizeMigration>,
4331 234 : ctx: &RequestContext,
4332 234 : ) -> anyhow::Result<(Arc<Timeline>, RequestContext)> {
4333 234 : let state = match cause {
4334 : CreateTimelineCause::Load => {
4335 234 : let ancestor_id = new_metadata.ancestor_timeline();
4336 234 : anyhow::ensure!(
4337 234 : ancestor_id == ancestor.as_ref().map(|t| t.timeline_id),
4338 0 : "Timeline's {new_timeline_id} ancestor {ancestor_id:?} was not found"
4339 : );
4340 234 : TimelineState::Loading
4341 : }
4342 0 : CreateTimelineCause::Delete => TimelineState::Stopping,
4343 : };
4344 :
4345 234 : let pg_version = new_metadata.pg_version();
4346 :
4347 234 : let timeline = Timeline::new(
4348 234 : self.conf,
4349 234 : Arc::clone(&self.tenant_conf),
4350 234 : new_metadata,
4351 234 : previous_heatmap,
4352 234 : ancestor,
4353 234 : new_timeline_id,
4354 234 : self.tenant_shard_id,
4355 234 : self.generation,
4356 234 : self.shard_identity,
4357 234 : self.walredo_mgr.clone(),
4358 234 : resources,
4359 234 : pg_version,
4360 234 : state,
4361 234 : self.attach_wal_lag_cooldown.clone(),
4362 234 : create_idempotency,
4363 234 : gc_compaction_state,
4364 234 : rel_size_v2_status,
4365 234 : self.cancel.child_token(),
4366 : );
4367 :
4368 234 : let timeline_ctx = RequestContextBuilder::from(ctx)
4369 234 : .scope(context::Scope::new_timeline(&timeline))
4370 234 : .detached_child();
4371 :
4372 234 : Ok((timeline, timeline_ctx))
4373 234 : }
4374 :
4375 : /// [`TenantShard::shutdown`] must be called before dropping the returned [`TenantShard`] object
4376 : /// to ensure proper cleanup of background tasks and metrics.
4377 : //
4378 : // Allow too_many_arguments because a constructor's argument list naturally grows with the
4379 : // number of attributes in the struct: breaking these out into a builder wouldn't be helpful.
4380 : #[allow(clippy::too_many_arguments)]
4381 118 : fn new(
4382 118 : state: TenantState,
4383 118 : conf: &'static PageServerConf,
4384 118 : attached_conf: AttachedTenantConf,
4385 118 : shard_identity: ShardIdentity,
4386 118 : walredo_mgr: Option<Arc<WalRedoManager>>,
4387 118 : tenant_shard_id: TenantShardId,
4388 118 : remote_storage: GenericRemoteStorage,
4389 118 : deletion_queue_client: DeletionQueueClient,
4390 118 : l0_flush_global_state: L0FlushGlobalState,
4391 118 : basebackup_cache: Arc<BasebackupCache>,
4392 118 : feature_resolver: FeatureResolver,
4393 118 : ) -> TenantShard {
4394 118 : assert!(!attached_conf.location.generation.is_none());
4395 :
4396 118 : let (state, mut rx) = watch::channel(state);
4397 :
4398 118 : tokio::spawn(async move {
4399 : // reflect tenant state in metrics:
4400 : // - global per tenant state: TENANT_STATE_METRIC
4401 : // - "set" of broken tenants: BROKEN_TENANTS_SET
4402 : //
4403 : // set of broken tenants should not have zero counts so that it remains accessible for
4404 : // alerting.
4405 :
4406 118 : let tid = tenant_shard_id.to_string();
4407 118 : let shard_id = tenant_shard_id.shard_slug().to_string();
4408 118 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
4409 :
4410 236 : fn inspect_state(state: &TenantState) -> ([&'static str; 1], bool) {
4411 236 : ([state.into()], matches!(state, TenantState::Broken { .. }))
4412 236 : }
4413 :
4414 118 : let mut tuple = inspect_state(&rx.borrow_and_update());
4415 :
4416 118 : let is_broken = tuple.1;
4417 118 : let mut counted_broken = if is_broken {
4418 : // add the id to the set right away, there should not be any updates on the channel
4419 : // after before tenant is removed, if ever
4420 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
4421 0 : true
4422 : } else {
4423 118 : false
4424 : };
4425 :
4426 : loop {
4427 236 : let labels = &tuple.0;
4428 236 : let current = TENANT_STATE_METRIC.with_label_values(labels);
4429 236 : current.inc();
4430 :
4431 236 : if rx.changed().await.is_err() {
4432 : // tenant has been dropped
4433 7 : current.dec();
4434 7 : drop(BROKEN_TENANTS_SET.remove_label_values(set_key));
4435 7 : break;
4436 118 : }
4437 :
4438 118 : current.dec();
4439 118 : tuple = inspect_state(&rx.borrow_and_update());
4440 :
4441 118 : let is_broken = tuple.1;
4442 118 : if is_broken && !counted_broken {
4443 0 : counted_broken = true;
4444 0 : // insert the tenant_id (back) into the set while avoiding needless counter
4445 0 : // access
4446 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
4447 118 : }
4448 : }
4449 7 : });
4450 :
4451 118 : TenantShard {
4452 118 : tenant_shard_id,
4453 118 : shard_identity,
4454 118 : generation: attached_conf.location.generation,
4455 118 : conf,
4456 118 : // using now here is good enough approximation to catch tenants with really long
4457 118 : // activation times.
4458 118 : constructed_at: Instant::now(),
4459 118 : timelines: Mutex::new(HashMap::new()),
4460 118 : timelines_creating: Mutex::new(HashSet::new()),
4461 118 : timelines_offloaded: Mutex::new(HashMap::new()),
4462 118 : timelines_importing: Mutex::new(HashMap::new()),
4463 118 : remote_tenant_manifest: Default::default(),
4464 118 : gc_cs: tokio::sync::Mutex::new(()),
4465 118 : walredo_mgr,
4466 118 : remote_storage,
4467 118 : deletion_queue_client,
4468 118 : state,
4469 118 : cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
4470 118 : cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
4471 118 : eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
4472 118 : compaction_circuit_breaker: std::sync::Mutex::new(CircuitBreaker::new(
4473 118 : format!("compaction-{tenant_shard_id}"),
4474 118 : 5,
4475 118 : // Compaction can be a very expensive operation, and might leak disk space. It also ought
4476 118 : // to be infallible, as long as remote storage is available. So if it repeatedly fails,
4477 118 : // use an extremely long backoff.
4478 118 : Some(Duration::from_secs(3600 * 24)),
4479 118 : )),
4480 118 : l0_compaction_trigger: Arc::new(Notify::new()),
4481 118 : scheduled_compaction_tasks: Mutex::new(Default::default()),
4482 118 : activate_now_sem: tokio::sync::Semaphore::new(0),
4483 118 : attach_wal_lag_cooldown: Arc::new(std::sync::OnceLock::new()),
4484 118 : cancel: CancellationToken::default(),
4485 118 : gate: Gate::default(),
4486 118 : pagestream_throttle: Arc::new(throttle::Throttle::new(
4487 118 : TenantShard::get_pagestream_throttle_config(conf, &attached_conf.tenant_conf),
4488 118 : )),
4489 118 : pagestream_throttle_metrics: Arc::new(
4490 118 : crate::metrics::tenant_throttling::Pagestream::new(&tenant_shard_id),
4491 118 : ),
4492 118 : tenant_conf: Arc::new(ArcSwap::from_pointee(attached_conf)),
4493 118 : ongoing_timeline_detach: std::sync::Mutex::default(),
4494 118 : gc_block: Default::default(),
4495 118 : l0_flush_global_state,
4496 118 : basebackup_cache,
4497 118 : feature_resolver: TenantFeatureResolver::new(
4498 118 : feature_resolver,
4499 118 : tenant_shard_id.tenant_id,
4500 118 : ),
4501 118 : }
4502 118 : }
4503 :
4504 : /// Locate and load config
4505 0 : pub(super) fn load_tenant_config(
4506 0 : conf: &'static PageServerConf,
4507 0 : tenant_shard_id: &TenantShardId,
4508 0 : ) -> Result<LocationConf, LoadConfigError> {
4509 0 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4510 :
4511 0 : info!("loading tenant configuration from {config_path}");
4512 :
4513 : // load and parse file
4514 0 : let config = fs::read_to_string(&config_path).map_err(|e| {
4515 0 : match e.kind() {
4516 : std::io::ErrorKind::NotFound => {
4517 : // The config should almost always exist for a tenant directory:
4518 : // - When attaching a tenant, the config is the first thing we write
4519 : // - When detaching a tenant, we atomically move the directory to a tmp location
4520 : // before deleting contents.
4521 : //
4522 : // The very rare edge case that can result in a missing config is if we crash during attach
4523 : // between creating directory and writing config. Callers should handle that as if the
4524 : // directory didn't exist.
4525 :
4526 0 : LoadConfigError::NotFound(config_path)
4527 : }
4528 : _ => {
4529 : // No IO errors except NotFound are acceptable here: other kinds of error indicate local storage or permissions issues
4530 : // that we cannot cleanly recover
4531 0 : crate::virtual_file::on_fatal_io_error(&e, "Reading tenant config file")
4532 : }
4533 : }
4534 0 : })?;
4535 :
4536 0 : Ok(toml_edit::de::from_str::<LocationConf>(&config)?)
4537 0 : }
4538 :
4539 : /// Stores a tenant location config to disk.
4540 : ///
4541 : /// NB: make sure to call `ShardIdentity::assert_equal` before persisting a new config, to avoid
4542 : /// changes to shard parameters that may result in data corruption.
4543 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4544 : pub(super) async fn persist_tenant_config(
4545 : conf: &'static PageServerConf,
4546 : tenant_shard_id: &TenantShardId,
4547 : location_conf: &LocationConf,
4548 : ) -> std::io::Result<()> {
4549 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4550 :
4551 : Self::persist_tenant_config_at(tenant_shard_id, &config_path, location_conf).await
4552 : }
4553 :
4554 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4555 : pub(super) async fn persist_tenant_config_at(
4556 : tenant_shard_id: &TenantShardId,
4557 : config_path: &Utf8Path,
4558 : location_conf: &LocationConf,
4559 : ) -> std::io::Result<()> {
4560 : debug!("persisting tenantconf to {config_path}");
4561 :
4562 : let mut conf_content = r#"# This file contains a specific per-tenant's config.
4563 : # It is read in case of pageserver restart.
4564 : "#
4565 : .to_string();
4566 :
4567 0 : fail::fail_point!("tenant-config-before-write", |_| {
4568 0 : Err(std::io::Error::other("tenant-config-before-write"))
4569 0 : });
4570 :
4571 : // Convert the config to a toml file.
4572 : conf_content +=
4573 : &toml_edit::ser::to_string_pretty(&location_conf).expect("Config serialization failed");
4574 :
4575 : let temp_path = path_with_suffix_extension(config_path, TEMP_FILE_SUFFIX);
4576 :
4577 : let conf_content = conf_content.into_bytes();
4578 : VirtualFile::crashsafe_overwrite(config_path.to_owned(), temp_path, conf_content).await
4579 : }
4580 :
4581 : //
4582 : // How garbage collection works:
4583 : //
4584 : // +--bar------------->
4585 : // /
4586 : // +----+-----foo---------------->
4587 : // /
4588 : // ----main--+-------------------------->
4589 : // \
4590 : // +-----baz-------->
4591 : //
4592 : //
4593 : // 1. Grab 'gc_cs' mutex to prevent new timelines from being created while Timeline's
4594 : // `gc_infos` are being refreshed
4595 : // 2. Scan collected timelines, and on each timeline, make note of the
4596 : // all the points where other timelines have been branched off.
4597 : // We will refrain from removing page versions at those LSNs.
4598 : // 3. For each timeline, scan all layer files on the timeline.
4599 : // Remove all files for which a newer file exists and which
4600 : // don't cover any branch point LSNs.
4601 : //
4602 : // TODO:
4603 : // - if a relation has a non-incremental persistent layer on a child branch, then we
4604 : // don't need to keep that in the parent anymore. But currently
4605 : // we do.
4606 2 : async fn gc_iteration_internal(
4607 2 : &self,
4608 2 : target_timeline_id: Option<TimelineId>,
4609 2 : horizon: u64,
4610 2 : pitr: Duration,
4611 2 : cancel: &CancellationToken,
4612 2 : ctx: &RequestContext,
4613 2 : ) -> Result<GcResult, GcError> {
4614 2 : let mut totals: GcResult = Default::default();
4615 2 : let now = Instant::now();
4616 :
4617 2 : let gc_timelines = self
4618 2 : .refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4619 2 : .await?;
4620 :
4621 2 : failpoint_support::sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
4622 :
4623 : // If there is nothing to GC, we don't want any messages in the INFO log.
4624 2 : if !gc_timelines.is_empty() {
4625 2 : info!("{} timelines need GC", gc_timelines.len());
4626 : } else {
4627 0 : debug!("{} timelines need GC", gc_timelines.len());
4628 : }
4629 :
4630 : // Perform GC for each timeline.
4631 : //
4632 : // Note that we don't hold the `TenantShard::gc_cs` lock here because we don't want to delay the
4633 : // branch creation task, which requires the GC lock. A GC iteration can run concurrently
4634 : // with branch creation.
4635 : //
4636 : // See comments in [`TenantShard::branch_timeline`] for more information about why branch
4637 : // creation task can run concurrently with timeline's GC iteration.
4638 4 : for timeline in gc_timelines {
4639 2 : if cancel.is_cancelled() {
4640 : // We were requested to shut down. Stop and return with the progress we
4641 : // made.
4642 0 : break;
4643 2 : }
4644 2 : let result = match timeline.gc().await {
4645 : Err(GcError::TimelineCancelled) => {
4646 0 : if target_timeline_id.is_some() {
4647 : // If we were targetting this specific timeline, surface cancellation to caller
4648 0 : return Err(GcError::TimelineCancelled);
4649 : } else {
4650 : // A timeline may be shutting down independently of the tenant's lifecycle: we should
4651 : // skip past this and proceed to try GC on other timelines.
4652 0 : continue;
4653 : }
4654 : }
4655 2 : r => r?,
4656 : };
4657 2 : totals += result;
4658 : }
4659 :
4660 2 : totals.elapsed = now.elapsed();
4661 2 : Ok(totals)
4662 2 : }
4663 :
4664 : /// Refreshes the Timeline::gc_info for all timelines, returning the
4665 : /// vector of timelines which have [`Timeline::get_last_record_lsn`] past
4666 : /// [`TenantShard::get_gc_horizon`].
4667 : ///
4668 : /// This is usually executed as part of periodic gc, but can now be triggered more often.
4669 2 : pub(crate) async fn refresh_gc_info(
4670 2 : &self,
4671 2 : cancel: &CancellationToken,
4672 2 : ctx: &RequestContext,
4673 2 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4674 : // since this method can now be called at different rates than the configured gc loop, it
4675 : // might be that these configuration values get applied faster than what it was previously,
4676 : // since these were only read from the gc task.
4677 2 : let horizon = self.get_gc_horizon();
4678 2 : let pitr = self.get_pitr_interval();
4679 :
4680 : // refresh all timelines
4681 2 : let target_timeline_id = None;
4682 :
4683 2 : self.refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4684 2 : .await
4685 2 : }
4686 :
4687 : /// Populate all Timelines' `GcInfo` with information about their children. We do not set the
4688 : /// PITR cutoffs here, because that requires I/O: this is done later, before GC, by [`Self::refresh_gc_info_internal`]
4689 : ///
4690 : /// Subsequently, parent-child relationships are updated incrementally inside [`Timeline::new`] and [`Timeline::drop`].
4691 118 : fn initialize_gc_info(
4692 118 : &self,
4693 118 : timelines: &std::sync::MutexGuard<HashMap<TimelineId, Arc<Timeline>>>,
4694 118 : timelines_offloaded: &std::sync::MutexGuard<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
4695 118 : restrict_to_timeline: Option<TimelineId>,
4696 118 : ) {
4697 118 : if restrict_to_timeline.is_none() {
4698 : // This function must be called before activation: after activation timeline create/delete operations
4699 : // might happen, and this function is not safe to run concurrently with those.
4700 118 : assert!(!self.is_active());
4701 0 : }
4702 :
4703 : // Scan all timelines. For each timeline, remember the timeline ID and
4704 : // the branch point where it was created.
4705 118 : let mut all_branchpoints: BTreeMap<TimelineId, Vec<(Lsn, TimelineId, MaybeOffloaded)>> =
4706 118 : BTreeMap::new();
4707 118 : timelines.iter().for_each(|(timeline_id, timeline_entry)| {
4708 3 : if let Some(ancestor_timeline_id) = &timeline_entry.get_ancestor_timeline_id() {
4709 1 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4710 1 : ancestor_children.push((
4711 1 : timeline_entry.get_ancestor_lsn(),
4712 1 : *timeline_id,
4713 1 : MaybeOffloaded::No,
4714 1 : ));
4715 2 : }
4716 3 : });
4717 118 : timelines_offloaded
4718 118 : .iter()
4719 118 : .for_each(|(timeline_id, timeline_entry)| {
4720 0 : let Some(ancestor_timeline_id) = &timeline_entry.ancestor_timeline_id else {
4721 0 : return;
4722 : };
4723 0 : let Some(retain_lsn) = timeline_entry.ancestor_retain_lsn else {
4724 0 : return;
4725 : };
4726 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4727 0 : ancestor_children.push((retain_lsn, *timeline_id, MaybeOffloaded::Yes));
4728 0 : });
4729 :
4730 : // The number of bytes we always keep, irrespective of PITR: this is a constant across timelines
4731 118 : let horizon = self.get_gc_horizon();
4732 :
4733 : // Populate each timeline's GcInfo with information about its child branches
4734 118 : let timelines_to_write = if let Some(timeline_id) = restrict_to_timeline {
4735 0 : itertools::Either::Left(timelines.get(&timeline_id).into_iter())
4736 : } else {
4737 118 : itertools::Either::Right(timelines.values())
4738 : };
4739 121 : for timeline in timelines_to_write {
4740 3 : let mut branchpoints: Vec<(Lsn, TimelineId, MaybeOffloaded)> = all_branchpoints
4741 3 : .remove(&timeline.timeline_id)
4742 3 : .unwrap_or_default();
4743 :
4744 3 : branchpoints.sort_by_key(|b| b.0);
4745 :
4746 3 : let mut target = timeline.gc_info.write().unwrap();
4747 :
4748 3 : target.retain_lsns = branchpoints;
4749 :
4750 3 : let space_cutoff = timeline
4751 3 : .get_last_record_lsn()
4752 3 : .checked_sub(horizon)
4753 3 : .unwrap_or(Lsn(0));
4754 :
4755 3 : target.cutoffs = GcCutoffs {
4756 3 : space: space_cutoff,
4757 3 : time: None,
4758 3 : };
4759 : }
4760 118 : }
4761 :
4762 4 : async fn refresh_gc_info_internal(
4763 4 : &self,
4764 4 : target_timeline_id: Option<TimelineId>,
4765 4 : horizon: u64,
4766 4 : pitr: Duration,
4767 4 : cancel: &CancellationToken,
4768 4 : ctx: &RequestContext,
4769 4 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4770 : // before taking the gc_cs lock, do the heavier weight finding of gc_cutoff points for
4771 : // currently visible timelines.
4772 4 : let timelines = self
4773 4 : .timelines
4774 4 : .lock()
4775 4 : .unwrap()
4776 4 : .values()
4777 10 : .filter(|tl| match target_timeline_id.as_ref() {
4778 2 : Some(target) => &tl.timeline_id == target,
4779 8 : None => true,
4780 10 : })
4781 4 : .cloned()
4782 4 : .collect::<Vec<_>>();
4783 :
4784 4 : if target_timeline_id.is_some() && timelines.is_empty() {
4785 : // We were to act on a particular timeline and it wasn't found
4786 0 : return Err(GcError::TimelineNotFound);
4787 4 : }
4788 :
4789 4 : let mut gc_cutoffs: HashMap<TimelineId, GcCutoffs> =
4790 4 : HashMap::with_capacity(timelines.len());
4791 :
4792 : // Ensures all timelines use the same start time when computing the time cutoff.
4793 4 : let now_ts_for_pitr_calc = SystemTime::now();
4794 10 : for timeline in timelines.iter() {
4795 10 : let ctx = &ctx.with_scope_timeline(timeline);
4796 10 : let cutoff = timeline
4797 10 : .get_last_record_lsn()
4798 10 : .checked_sub(horizon)
4799 10 : .unwrap_or(Lsn(0));
4800 :
4801 10 : let cutoffs = timeline
4802 10 : .find_gc_cutoffs(now_ts_for_pitr_calc, cutoff, pitr, cancel, ctx)
4803 10 : .await?;
4804 10 : let old = gc_cutoffs.insert(timeline.timeline_id, cutoffs);
4805 10 : assert!(old.is_none());
4806 : }
4807 :
4808 4 : if !self.is_active() || self.cancel.is_cancelled() {
4809 0 : return Err(GcError::TenantCancelled);
4810 4 : }
4811 :
4812 : // grab mutex to prevent new timelines from being created here; avoid doing long operations
4813 : // because that will stall branch creation.
4814 4 : let gc_cs = self.gc_cs.lock().await;
4815 :
4816 : // Ok, we now know all the branch points.
4817 : // Update the GC information for each timeline.
4818 4 : let mut gc_timelines = Vec::with_capacity(timelines.len());
4819 14 : for timeline in timelines {
4820 : // We filtered the timeline list above
4821 10 : if let Some(target_timeline_id) = target_timeline_id {
4822 2 : assert_eq!(target_timeline_id, timeline.timeline_id);
4823 8 : }
4824 :
4825 : {
4826 10 : let mut target = timeline.gc_info.write().unwrap();
4827 :
4828 : // Cull any expired leases
4829 10 : let now = SystemTime::now();
4830 10 : target.leases.retain(|_, lease| !lease.is_expired(&now));
4831 :
4832 10 : timeline
4833 10 : .metrics
4834 10 : .valid_lsn_lease_count_gauge
4835 10 : .set(target.leases.len() as u64);
4836 :
4837 : // Look up parent's PITR cutoff to update the child's knowledge of whether it is within parent's PITR
4838 10 : if let Some(ancestor_id) = timeline.get_ancestor_timeline_id() {
4839 6 : if let Some(ancestor_gc_cutoffs) = gc_cutoffs.get(&ancestor_id) {
4840 6 : target.within_ancestor_pitr =
4841 6 : Some(timeline.get_ancestor_lsn()) >= ancestor_gc_cutoffs.time;
4842 6 : }
4843 4 : }
4844 :
4845 : // Update metrics that depend on GC state
4846 10 : timeline
4847 10 : .metrics
4848 10 : .archival_size
4849 10 : .set(if target.within_ancestor_pitr {
4850 0 : timeline.metrics.current_logical_size_gauge.get()
4851 : } else {
4852 10 : 0
4853 : });
4854 10 : if let Some(time_cutoff) = target.cutoffs.time {
4855 4 : timeline.metrics.pitr_history_size.set(
4856 4 : timeline
4857 4 : .get_last_record_lsn()
4858 4 : .checked_sub(time_cutoff)
4859 4 : .unwrap_or_default()
4860 4 : .0,
4861 4 : );
4862 6 : }
4863 :
4864 : // Apply the cutoffs we found to the Timeline's GcInfo. Why might we _not_ have cutoffs for a timeline?
4865 : // - this timeline was created while we were finding cutoffs
4866 : // - lsn for timestamp search fails for this timeline repeatedly
4867 10 : if let Some(cutoffs) = gc_cutoffs.get(&timeline.timeline_id) {
4868 10 : let original_cutoffs = target.cutoffs.clone();
4869 : // GC cutoffs should never go back
4870 10 : target.cutoffs = GcCutoffs {
4871 10 : space: cutoffs.space.max(original_cutoffs.space),
4872 10 : time: cutoffs.time.max(original_cutoffs.time),
4873 10 : }
4874 0 : }
4875 : }
4876 :
4877 10 : gc_timelines.push(timeline);
4878 : }
4879 4 : drop(gc_cs);
4880 4 : Ok(gc_timelines)
4881 4 : }
4882 :
4883 : /// A substitute for `branch_timeline` for use in unit tests.
4884 : /// The returned timeline will have state value `Active` to make various `anyhow::ensure!()`
4885 : /// calls pass, but, we do not actually call `.activate()` under the hood. So, none of the
4886 : /// timeline background tasks are launched, except the flush loop.
4887 : #[cfg(test)]
4888 119 : async fn branch_timeline_test(
4889 119 : self: &Arc<Self>,
4890 119 : src_timeline: &Arc<Timeline>,
4891 119 : dst_id: TimelineId,
4892 119 : ancestor_lsn: Option<Lsn>,
4893 119 : ctx: &RequestContext,
4894 119 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
4895 119 : let tl = self
4896 119 : .branch_timeline_impl(src_timeline, dst_id, ancestor_lsn, ctx)
4897 119 : .await?
4898 117 : .into_timeline_for_test();
4899 117 : tl.set_state(TimelineState::Active);
4900 117 : Ok(tl)
4901 119 : }
4902 :
4903 : /// Helper for unit tests to branch a timeline with some pre-loaded states.
4904 : #[cfg(test)]
4905 : #[allow(clippy::too_many_arguments)]
4906 6 : pub async fn branch_timeline_test_with_layers(
4907 6 : self: &Arc<Self>,
4908 6 : src_timeline: &Arc<Timeline>,
4909 6 : dst_id: TimelineId,
4910 6 : ancestor_lsn: Option<Lsn>,
4911 6 : ctx: &RequestContext,
4912 6 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
4913 6 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
4914 6 : end_lsn: Lsn,
4915 6 : ) -> anyhow::Result<Arc<Timeline>> {
4916 : use checks::check_valid_layermap;
4917 : use itertools::Itertools;
4918 :
4919 6 : let tline = self
4920 6 : .branch_timeline_test(src_timeline, dst_id, ancestor_lsn, ctx)
4921 6 : .await?;
4922 6 : let ancestor_lsn = if let Some(ancestor_lsn) = ancestor_lsn {
4923 6 : ancestor_lsn
4924 : } else {
4925 0 : tline.get_last_record_lsn()
4926 : };
4927 6 : assert!(end_lsn >= ancestor_lsn);
4928 6 : tline.force_advance_lsn(end_lsn);
4929 9 : for deltas in delta_layer_desc {
4930 3 : tline
4931 3 : .force_create_delta_layer(deltas, Some(ancestor_lsn), ctx)
4932 3 : .await?;
4933 : }
4934 8 : for (lsn, images) in image_layer_desc {
4935 2 : tline
4936 2 : .force_create_image_layer(lsn, images, Some(ancestor_lsn), ctx)
4937 2 : .await?;
4938 : }
4939 6 : let layer_names = tline
4940 6 : .layers
4941 6 : .read(LayerManagerLockHolder::Testing)
4942 6 : .await
4943 6 : .layer_map()
4944 6 : .unwrap()
4945 6 : .iter_historic_layers()
4946 6 : .map(|layer| layer.layer_name())
4947 6 : .collect_vec();
4948 6 : if let Some(err) = check_valid_layermap(&layer_names) {
4949 0 : bail!("invalid layermap: {err}");
4950 6 : }
4951 6 : Ok(tline)
4952 6 : }
4953 :
4954 : /// Branch an existing timeline.
4955 0 : async fn branch_timeline(
4956 0 : self: &Arc<Self>,
4957 0 : src_timeline: &Arc<Timeline>,
4958 0 : dst_id: TimelineId,
4959 0 : start_lsn: Option<Lsn>,
4960 0 : ctx: &RequestContext,
4961 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4962 0 : self.branch_timeline_impl(src_timeline, dst_id, start_lsn, ctx)
4963 0 : .await
4964 0 : }
4965 :
4966 119 : async fn branch_timeline_impl(
4967 119 : self: &Arc<Self>,
4968 119 : src_timeline: &Arc<Timeline>,
4969 119 : dst_id: TimelineId,
4970 119 : start_lsn: Option<Lsn>,
4971 119 : ctx: &RequestContext,
4972 119 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4973 119 : let src_id = src_timeline.timeline_id;
4974 :
4975 : // We will validate our ancestor LSN in this function. Acquire the GC lock so that
4976 : // this check cannot race with GC, and the ancestor LSN is guaranteed to remain
4977 : // valid while we are creating the branch.
4978 119 : let _gc_cs = self.gc_cs.lock().await;
4979 :
4980 : // If no start LSN is specified, we branch the new timeline from the source timeline's last record LSN
4981 119 : let start_lsn = start_lsn.unwrap_or_else(|| {
4982 1 : let lsn = src_timeline.get_last_record_lsn();
4983 1 : info!("branching timeline {dst_id} from timeline {src_id} at last record LSN: {lsn}");
4984 1 : lsn
4985 1 : });
4986 :
4987 : // we finally have determined the ancestor_start_lsn, so we can get claim exclusivity now
4988 119 : let timeline_create_guard = match self
4989 119 : .start_creating_timeline(
4990 119 : dst_id,
4991 119 : CreateTimelineIdempotency::Branch {
4992 119 : ancestor_timeline_id: src_timeline.timeline_id,
4993 119 : ancestor_start_lsn: start_lsn,
4994 119 : },
4995 119 : )
4996 119 : .await?
4997 : {
4998 119 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
4999 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
5000 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
5001 : }
5002 : };
5003 :
5004 : // Ensure that `start_lsn` is valid, i.e. the LSN is within the PITR
5005 : // horizon on the source timeline
5006 : //
5007 : // We check it against both the planned GC cutoff stored in 'gc_info',
5008 : // and the 'latest_gc_cutoff' of the last GC that was performed. The
5009 : // planned GC cutoff in 'gc_info' is normally larger than
5010 : // 'applied_gc_cutoff_lsn', but beware of corner cases like if you just
5011 : // changed the GC settings for the tenant to make the PITR window
5012 : // larger, but some of the data was already removed by an earlier GC
5013 : // iteration.
5014 :
5015 : // check against last actual 'latest_gc_cutoff' first
5016 119 : let applied_gc_cutoff_lsn = src_timeline.get_applied_gc_cutoff_lsn();
5017 : {
5018 119 : let gc_info = src_timeline.gc_info.read().unwrap();
5019 119 : let planned_cutoff = gc_info.min_cutoff();
5020 119 : if gc_info.lsn_covered_by_lease(start_lsn) {
5021 0 : tracing::info!(
5022 0 : "skipping comparison of {start_lsn} with gc cutoff {} and planned gc cutoff {planned_cutoff} due to lsn lease",
5023 0 : *applied_gc_cutoff_lsn
5024 : );
5025 : } else {
5026 119 : src_timeline
5027 119 : .check_lsn_is_in_scope(start_lsn, &applied_gc_cutoff_lsn)
5028 119 : .context(format!(
5029 119 : "invalid branch start lsn: less than latest GC cutoff {}",
5030 119 : *applied_gc_cutoff_lsn,
5031 : ))
5032 119 : .map_err(CreateTimelineError::AncestorLsn)?;
5033 :
5034 : // and then the planned GC cutoff
5035 117 : if start_lsn < planned_cutoff {
5036 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
5037 0 : "invalid branch start lsn: less than planned GC cutoff {planned_cutoff}"
5038 0 : )));
5039 117 : }
5040 : }
5041 : }
5042 :
5043 : //
5044 : // The branch point is valid, and we are still holding the 'gc_cs' lock
5045 : // so that GC cannot advance the GC cutoff until we are finished.
5046 : // Proceed with the branch creation.
5047 : //
5048 :
5049 : // Determine prev-LSN for the new timeline. We can only determine it if
5050 : // the timeline was branched at the current end of the source timeline.
5051 : let RecordLsn {
5052 117 : last: src_last,
5053 117 : prev: src_prev,
5054 117 : } = src_timeline.get_last_record_rlsn();
5055 117 : let dst_prev = if src_last == start_lsn {
5056 108 : Some(src_prev)
5057 : } else {
5058 9 : None
5059 : };
5060 :
5061 : // Create the metadata file, noting the ancestor of the new timeline.
5062 : // There is initially no data in it, but all the read-calls know to look
5063 : // into the ancestor.
5064 117 : let metadata = TimelineMetadata::new(
5065 117 : start_lsn,
5066 117 : dst_prev,
5067 117 : Some(src_id),
5068 117 : start_lsn,
5069 117 : *src_timeline.applied_gc_cutoff_lsn.read(), // FIXME: should we hold onto this guard longer?
5070 117 : src_timeline.initdb_lsn,
5071 117 : src_timeline.pg_version,
5072 : );
5073 :
5074 117 : let (uninitialized_timeline, _timeline_ctx) = self
5075 117 : .prepare_new_timeline(
5076 117 : dst_id,
5077 117 : &metadata,
5078 117 : timeline_create_guard,
5079 117 : start_lsn + 1,
5080 117 : Some(Arc::clone(src_timeline)),
5081 117 : Some(src_timeline.get_rel_size_v2_status()),
5082 117 : ctx,
5083 117 : )
5084 117 : .await?;
5085 :
5086 117 : let new_timeline = uninitialized_timeline.finish_creation().await?;
5087 :
5088 : // Root timeline gets its layers during creation and uploads them along with the metadata.
5089 : // A branch timeline though, when created, can get no writes for some time, hence won't get any layers created.
5090 : // We still need to upload its metadata eagerly: if other nodes `attach` the tenant and miss this timeline, their GC
5091 : // could get incorrect information and remove more layers, than needed.
5092 : // See also https://github.com/neondatabase/neon/issues/3865
5093 117 : new_timeline
5094 117 : .remote_client
5095 117 : .schedule_index_upload_for_full_metadata_update(&metadata)
5096 117 : .context("branch initial metadata upload")?;
5097 :
5098 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
5099 :
5100 117 : Ok(CreateTimelineResult::Created(new_timeline))
5101 119 : }
5102 :
5103 : /// For unit tests, make this visible so that other modules can directly create timelines
5104 : #[cfg(test)]
5105 : #[tracing::instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
5106 : pub(crate) async fn bootstrap_timeline_test(
5107 : self: &Arc<Self>,
5108 : timeline_id: TimelineId,
5109 : pg_version: PgMajorVersion,
5110 : load_existing_initdb: Option<TimelineId>,
5111 : ctx: &RequestContext,
5112 : ) -> anyhow::Result<Arc<Timeline>> {
5113 : self.bootstrap_timeline(timeline_id, pg_version, load_existing_initdb, ctx)
5114 : .await
5115 : .map_err(anyhow::Error::new)
5116 1 : .map(|r| r.into_timeline_for_test())
5117 : }
5118 :
5119 : /// Get exclusive access to the timeline ID for creation.
5120 : ///
5121 : /// Timeline-creating code paths must use this function before making changes
5122 : /// to in-memory or persistent state.
5123 : ///
5124 : /// The `state` parameter is a description of the timeline creation operation
5125 : /// we intend to perform.
5126 : /// If the timeline was already created in the meantime, we check whether this
5127 : /// request conflicts or is idempotent , based on `state`.
5128 234 : async fn start_creating_timeline(
5129 234 : self: &Arc<Self>,
5130 234 : new_timeline_id: TimelineId,
5131 234 : idempotency: CreateTimelineIdempotency,
5132 234 : ) -> Result<StartCreatingTimelineResult, CreateTimelineError> {
5133 234 : let allow_offloaded = false;
5134 234 : match self.create_timeline_create_guard(new_timeline_id, idempotency, allow_offloaded) {
5135 233 : Ok(create_guard) => {
5136 233 : pausable_failpoint!("timeline-creation-after-uninit");
5137 233 : Ok(StartCreatingTimelineResult::CreateGuard(create_guard))
5138 : }
5139 0 : Err(TimelineExclusionError::ShuttingDown) => Err(CreateTimelineError::ShuttingDown),
5140 : Err(TimelineExclusionError::AlreadyCreating) => {
5141 : // Creation is in progress, we cannot create it again, and we cannot
5142 : // check if this request matches the existing one, so caller must try
5143 : // again later.
5144 0 : Err(CreateTimelineError::AlreadyCreating)
5145 : }
5146 0 : Err(TimelineExclusionError::Other(e)) => Err(CreateTimelineError::Other(e)),
5147 : Err(TimelineExclusionError::AlreadyExists {
5148 0 : existing: TimelineOrOffloaded::Offloaded(_existing),
5149 : ..
5150 : }) => {
5151 0 : info!("timeline already exists but is offloaded");
5152 0 : Err(CreateTimelineError::Conflict)
5153 : }
5154 : Err(TimelineExclusionError::AlreadyExists {
5155 0 : existing: TimelineOrOffloaded::Importing(_existing),
5156 : ..
5157 : }) => {
5158 : // If there's a timeline already importing, then we would hit
5159 : // the [`TimelineExclusionError::AlreadyCreating`] branch above.
5160 0 : unreachable!("Importing timelines hold the creation guard")
5161 : }
5162 : Err(TimelineExclusionError::AlreadyExists {
5163 1 : existing: TimelineOrOffloaded::Timeline(existing),
5164 1 : arg,
5165 : }) => {
5166 : {
5167 1 : let existing = &existing.create_idempotency;
5168 1 : let _span = info_span!("idempotency_check", ?existing, ?arg).entered();
5169 1 : debug!("timeline already exists");
5170 :
5171 1 : match (existing, &arg) {
5172 : // FailWithConflict => no idempotency check
5173 : (CreateTimelineIdempotency::FailWithConflict, _)
5174 : | (_, CreateTimelineIdempotency::FailWithConflict) => {
5175 1 : warn!("timeline already exists, failing request");
5176 1 : return Err(CreateTimelineError::Conflict);
5177 : }
5178 : // Idempotent <=> CreateTimelineIdempotency is identical
5179 0 : (x, y) if x == y => {
5180 0 : info!(
5181 0 : "timeline already exists and idempotency matches, succeeding request"
5182 : );
5183 : // fallthrough
5184 : }
5185 : (_, _) => {
5186 0 : warn!("idempotency conflict, failing request");
5187 0 : return Err(CreateTimelineError::Conflict);
5188 : }
5189 : }
5190 : }
5191 :
5192 0 : Ok(StartCreatingTimelineResult::Idempotent(existing))
5193 : }
5194 : }
5195 234 : }
5196 :
5197 0 : async fn upload_initdb(
5198 0 : &self,
5199 0 : timelines_path: &Utf8PathBuf,
5200 0 : pgdata_path: &Utf8PathBuf,
5201 0 : timeline_id: &TimelineId,
5202 0 : ) -> anyhow::Result<()> {
5203 0 : let temp_path = timelines_path.join(format!(
5204 0 : "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
5205 0 : ));
5206 :
5207 0 : scopeguard::defer! {
5208 : if let Err(e) = fs::remove_file(&temp_path) {
5209 : error!("Failed to remove temporary initdb archive '{temp_path}': {e}");
5210 : }
5211 : }
5212 :
5213 0 : let (pgdata_zstd, tar_zst_size) = create_zst_tarball(pgdata_path, &temp_path).await?;
5214 : const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
5215 0 : if tar_zst_size > INITDB_TAR_ZST_WARN_LIMIT {
5216 0 : warn!(
5217 0 : "compressed {temp_path} size of {tar_zst_size} is above limit {INITDB_TAR_ZST_WARN_LIMIT}."
5218 : );
5219 0 : }
5220 :
5221 0 : pausable_failpoint!("before-initdb-upload");
5222 :
5223 0 : backoff::retry(
5224 0 : || async {
5225 0 : self::remote_timeline_client::upload_initdb_dir(
5226 0 : &self.remote_storage,
5227 0 : &self.tenant_shard_id.tenant_id,
5228 0 : timeline_id,
5229 0 : pgdata_zstd.try_clone().await?,
5230 0 : tar_zst_size,
5231 0 : &self.cancel,
5232 : )
5233 0 : .await
5234 0 : },
5235 : |_| false,
5236 : 3,
5237 : u32::MAX,
5238 0 : "persist_initdb_tar_zst",
5239 0 : &self.cancel,
5240 : )
5241 0 : .await
5242 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
5243 0 : .and_then(|x| x)
5244 0 : }
5245 :
5246 : /// - run initdb to init temporary instance and get bootstrap data
5247 : /// - after initialization completes, tar up the temp dir and upload it to S3.
5248 1 : async fn bootstrap_timeline(
5249 1 : self: &Arc<Self>,
5250 1 : timeline_id: TimelineId,
5251 1 : pg_version: PgMajorVersion,
5252 1 : load_existing_initdb: Option<TimelineId>,
5253 1 : ctx: &RequestContext,
5254 1 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
5255 1 : let timeline_create_guard = match self
5256 1 : .start_creating_timeline(
5257 1 : timeline_id,
5258 1 : CreateTimelineIdempotency::Bootstrap { pg_version },
5259 1 : )
5260 1 : .await?
5261 : {
5262 1 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
5263 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
5264 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
5265 : }
5266 : };
5267 :
5268 : // create a `tenant/{tenant_id}/timelines/basebackup-{timeline_id}.{TEMP_FILE_SUFFIX}/`
5269 : // temporary directory for basebackup files for the given timeline.
5270 :
5271 1 : let timelines_path = self.conf.timelines_path(&self.tenant_shard_id);
5272 1 : let pgdata_path = path_with_suffix_extension(
5273 1 : timelines_path.join(format!("basebackup-{timeline_id}")),
5274 1 : TEMP_FILE_SUFFIX,
5275 : );
5276 :
5277 : // Remove whatever was left from the previous runs: safe because TimelineCreateGuard guarantees
5278 : // we won't race with other creations or existent timelines with the same path.
5279 1 : if pgdata_path.exists() {
5280 0 : fs::remove_dir_all(&pgdata_path).with_context(|| {
5281 0 : format!("Failed to remove already existing initdb directory: {pgdata_path}")
5282 0 : })?;
5283 0 : tracing::info!("removed previous attempt's temporary initdb directory '{pgdata_path}'");
5284 1 : }
5285 :
5286 : // this new directory is very temporary, set to remove it immediately after bootstrap, we don't need it
5287 1 : let pgdata_path_deferred = pgdata_path.clone();
5288 1 : scopeguard::defer! {
5289 : if let Err(e) = fs::remove_dir_all(&pgdata_path_deferred).or_else(fs_ext::ignore_not_found) {
5290 : // this is unlikely, but we will remove the directory on pageserver restart or another bootstrap call
5291 : error!("Failed to remove temporary initdb directory '{pgdata_path_deferred}': {e}");
5292 : } else {
5293 : tracing::info!("removed temporary initdb directory '{pgdata_path_deferred}'");
5294 : }
5295 : }
5296 1 : if let Some(existing_initdb_timeline_id) = load_existing_initdb {
5297 1 : if existing_initdb_timeline_id != timeline_id {
5298 0 : let source_path = &remote_initdb_archive_path(
5299 0 : &self.tenant_shard_id.tenant_id,
5300 0 : &existing_initdb_timeline_id,
5301 0 : );
5302 0 : let dest_path =
5303 0 : &remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &timeline_id);
5304 :
5305 : // if this fails, it will get retried by retried control plane requests
5306 0 : self.remote_storage
5307 0 : .copy_object(source_path, dest_path, &self.cancel)
5308 0 : .await
5309 0 : .context("copy initdb tar")?;
5310 1 : }
5311 1 : let (initdb_tar_zst_path, initdb_tar_zst) =
5312 1 : self::remote_timeline_client::download_initdb_tar_zst(
5313 1 : self.conf,
5314 1 : &self.remote_storage,
5315 1 : &self.tenant_shard_id,
5316 1 : &existing_initdb_timeline_id,
5317 1 : &self.cancel,
5318 1 : )
5319 1 : .await
5320 1 : .context("download initdb tar")?;
5321 :
5322 1 : scopeguard::defer! {
5323 : if let Err(e) = fs::remove_file(&initdb_tar_zst_path) {
5324 : error!("Failed to remove temporary initdb archive '{initdb_tar_zst_path}': {e}");
5325 : }
5326 : }
5327 :
5328 1 : let buf_read =
5329 1 : BufReader::with_capacity(remote_timeline_client::BUFFER_SIZE, initdb_tar_zst);
5330 1 : extract_zst_tarball(&pgdata_path, buf_read)
5331 1 : .await
5332 1 : .context("extract initdb tar")?;
5333 : } else {
5334 : // Init temporarily repo to get bootstrap data, this creates a directory in the `pgdata_path` path
5335 0 : run_initdb(self.conf, &pgdata_path, pg_version, &self.cancel)
5336 0 : .await
5337 0 : .context("run initdb")?;
5338 :
5339 : // Upload the created data dir to S3
5340 0 : if self.tenant_shard_id().is_shard_zero() {
5341 0 : self.upload_initdb(&timelines_path, &pgdata_path, &timeline_id)
5342 0 : .await?;
5343 0 : }
5344 : }
5345 1 : let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
5346 :
5347 : // Import the contents of the data directory at the initial checkpoint
5348 : // LSN, and any WAL after that.
5349 : // Initdb lsn will be equal to last_record_lsn which will be set after import.
5350 : // Because we know it upfront avoid having an option or dummy zero value by passing it to the metadata.
5351 1 : let new_metadata = TimelineMetadata::new(
5352 1 : Lsn(0),
5353 1 : None,
5354 1 : None,
5355 1 : Lsn(0),
5356 1 : pgdata_lsn,
5357 1 : pgdata_lsn,
5358 1 : pg_version,
5359 : );
5360 1 : let (mut raw_timeline, timeline_ctx) = self
5361 1 : .prepare_new_timeline(
5362 1 : timeline_id,
5363 1 : &new_metadata,
5364 1 : timeline_create_guard,
5365 1 : pgdata_lsn,
5366 1 : None,
5367 1 : None,
5368 1 : ctx,
5369 1 : )
5370 1 : .await?;
5371 :
5372 1 : let tenant_shard_id = raw_timeline.owning_tenant.tenant_shard_id;
5373 1 : raw_timeline
5374 1 : .write(|unfinished_timeline| async move {
5375 1 : import_datadir::import_timeline_from_postgres_datadir(
5376 1 : &unfinished_timeline,
5377 1 : &pgdata_path,
5378 1 : pgdata_lsn,
5379 1 : &timeline_ctx,
5380 1 : )
5381 1 : .await
5382 1 : .with_context(|| {
5383 0 : format!(
5384 0 : "Failed to import pgdatadir for timeline {tenant_shard_id}/{timeline_id}"
5385 : )
5386 0 : })?;
5387 :
5388 1 : fail::fail_point!("before-checkpoint-new-timeline", |_| {
5389 0 : Err(CreateTimelineError::Other(anyhow::anyhow!(
5390 0 : "failpoint before-checkpoint-new-timeline"
5391 0 : )))
5392 0 : });
5393 :
5394 1 : Ok(())
5395 2 : })
5396 1 : .await?;
5397 :
5398 : // All done!
5399 1 : let timeline = raw_timeline.finish_creation().await?;
5400 :
5401 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
5402 :
5403 1 : Ok(CreateTimelineResult::Created(timeline))
5404 1 : }
5405 :
5406 231 : fn build_timeline_remote_client(&self, timeline_id: TimelineId) -> RemoteTimelineClient {
5407 231 : RemoteTimelineClient::new(
5408 231 : self.remote_storage.clone(),
5409 231 : self.deletion_queue_client.clone(),
5410 231 : self.conf,
5411 231 : self.tenant_shard_id,
5412 231 : timeline_id,
5413 231 : self.generation,
5414 231 : &self.tenant_conf.load().location,
5415 : )
5416 231 : }
5417 :
5418 : /// Builds required resources for a new timeline.
5419 231 : fn build_timeline_resources(&self, timeline_id: TimelineId) -> TimelineResources {
5420 231 : let remote_client = self.build_timeline_remote_client(timeline_id);
5421 231 : self.get_timeline_resources_for(remote_client)
5422 231 : }
5423 :
5424 : /// Builds timeline resources for the given remote client.
5425 234 : fn get_timeline_resources_for(&self, remote_client: RemoteTimelineClient) -> TimelineResources {
5426 234 : TimelineResources {
5427 234 : remote_client,
5428 234 : pagestream_throttle: self.pagestream_throttle.clone(),
5429 234 : pagestream_throttle_metrics: self.pagestream_throttle_metrics.clone(),
5430 234 : l0_compaction_trigger: self.l0_compaction_trigger.clone(),
5431 234 : l0_flush_global_state: self.l0_flush_global_state.clone(),
5432 234 : basebackup_cache: self.basebackup_cache.clone(),
5433 234 : feature_resolver: self.feature_resolver.clone(),
5434 234 : }
5435 234 : }
5436 :
5437 : /// Creates intermediate timeline structure and its files.
5438 : ///
5439 : /// An empty layer map is initialized, and new data and WAL can be imported starting
5440 : /// at 'disk_consistent_lsn'. After any initial data has been imported, call
5441 : /// `finish_creation` to insert the Timeline into the timelines map.
5442 : #[allow(clippy::too_many_arguments)]
5443 231 : async fn prepare_new_timeline<'a>(
5444 231 : &'a self,
5445 231 : new_timeline_id: TimelineId,
5446 231 : new_metadata: &TimelineMetadata,
5447 231 : create_guard: TimelineCreateGuard,
5448 231 : start_lsn: Lsn,
5449 231 : ancestor: Option<Arc<Timeline>>,
5450 231 : rel_size_v2_status: Option<RelSizeMigration>,
5451 231 : ctx: &RequestContext,
5452 231 : ) -> anyhow::Result<(UninitializedTimeline<'a>, RequestContext)> {
5453 231 : let tenant_shard_id = self.tenant_shard_id;
5454 :
5455 231 : let resources = self.build_timeline_resources(new_timeline_id);
5456 231 : resources
5457 231 : .remote_client
5458 231 : .init_upload_queue_for_empty_remote(new_metadata, rel_size_v2_status.clone())?;
5459 :
5460 231 : let (timeline_struct, timeline_ctx) = self
5461 231 : .create_timeline_struct(
5462 231 : new_timeline_id,
5463 231 : new_metadata,
5464 231 : None,
5465 231 : ancestor,
5466 231 : resources,
5467 231 : CreateTimelineCause::Load,
5468 231 : create_guard.idempotency.clone(),
5469 231 : None,
5470 231 : rel_size_v2_status,
5471 231 : ctx,
5472 : )
5473 231 : .context("Failed to create timeline data structure")?;
5474 :
5475 231 : timeline_struct.init_empty_layer_map(start_lsn);
5476 :
5477 231 : if let Err(e) = self
5478 231 : .create_timeline_files(&create_guard.timeline_path)
5479 231 : .await
5480 : {
5481 0 : error!(
5482 0 : "Failed to create initial files for timeline {tenant_shard_id}/{new_timeline_id}, cleaning up: {e:?}"
5483 : );
5484 0 : cleanup_timeline_directory(create_guard);
5485 0 : return Err(e);
5486 231 : }
5487 :
5488 231 : debug!(
5489 0 : "Successfully created initial files for timeline {tenant_shard_id}/{new_timeline_id}"
5490 : );
5491 :
5492 231 : Ok((
5493 231 : UninitializedTimeline::new(
5494 231 : self,
5495 231 : new_timeline_id,
5496 231 : Some((timeline_struct, create_guard)),
5497 231 : ),
5498 231 : timeline_ctx,
5499 231 : ))
5500 231 : }
5501 :
5502 231 : async fn create_timeline_files(&self, timeline_path: &Utf8Path) -> anyhow::Result<()> {
5503 231 : crashsafe::create_dir(timeline_path).context("Failed to create timeline directory")?;
5504 :
5505 231 : fail::fail_point!("after-timeline-dir-creation", |_| {
5506 0 : anyhow::bail!("failpoint after-timeline-dir-creation");
5507 0 : });
5508 :
5509 231 : Ok(())
5510 231 : }
5511 :
5512 : /// Get a guard that provides exclusive access to the timeline directory, preventing
5513 : /// concurrent attempts to create the same timeline.
5514 : ///
5515 : /// The `allow_offloaded` parameter controls whether to tolerate the existence of
5516 : /// offloaded timelines or not.
5517 234 : fn create_timeline_create_guard(
5518 234 : self: &Arc<Self>,
5519 234 : timeline_id: TimelineId,
5520 234 : idempotency: CreateTimelineIdempotency,
5521 234 : allow_offloaded: bool,
5522 234 : ) -> Result<TimelineCreateGuard, TimelineExclusionError> {
5523 234 : let tenant_shard_id = self.tenant_shard_id;
5524 :
5525 234 : let timeline_path = self.conf.timeline_path(&tenant_shard_id, &timeline_id);
5526 :
5527 234 : let create_guard = TimelineCreateGuard::new(
5528 234 : self,
5529 234 : timeline_id,
5530 234 : timeline_path.clone(),
5531 234 : idempotency,
5532 234 : allow_offloaded,
5533 1 : )?;
5534 :
5535 : // At this stage, we have got exclusive access to in-memory state for this timeline ID
5536 : // for creation.
5537 : // A timeline directory should never exist on disk already:
5538 : // - a previous failed creation would have cleaned up after itself
5539 : // - a pageserver restart would clean up timeline directories that don't have valid remote state
5540 : //
5541 : // Therefore it is an unexpected internal error to encounter a timeline directory already existing here,
5542 : // this error may indicate a bug in cleanup on failed creations.
5543 233 : if timeline_path.exists() {
5544 0 : return Err(TimelineExclusionError::Other(anyhow::anyhow!(
5545 0 : "Timeline directory already exists! This is a bug."
5546 0 : )));
5547 233 : }
5548 :
5549 233 : Ok(create_guard)
5550 234 : }
5551 :
5552 : /// Gathers inputs from all of the timelines to produce a sizing model input.
5553 : ///
5554 : /// Future is cancellation safe. Only one calculation can be running at once per tenant.
5555 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5556 : pub async fn gather_size_inputs(
5557 : &self,
5558 : // `max_retention_period` overrides the cutoff that is used to calculate the size
5559 : // (only if it is shorter than the real cutoff).
5560 : max_retention_period: Option<u64>,
5561 : cause: LogicalSizeCalculationCause,
5562 : cancel: &CancellationToken,
5563 : ctx: &RequestContext,
5564 : ) -> Result<size::ModelInputs, size::CalculateSyntheticSizeError> {
5565 : let logical_sizes_at_once = self
5566 : .conf
5567 : .concurrent_tenant_size_logical_size_queries
5568 : .inner();
5569 :
5570 : // TODO: Having a single mutex block concurrent reads is not great for performance.
5571 : //
5572 : // But the only case where we need to run multiple of these at once is when we
5573 : // request a size for a tenant manually via API, while another background calculation
5574 : // is in progress (which is not a common case).
5575 : //
5576 : // See more for on the issue #2748 condenced out of the initial PR review.
5577 : let mut shared_cache = tokio::select! {
5578 : locked = self.cached_logical_sizes.lock() => locked,
5579 : _ = cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5580 : _ = self.cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5581 : };
5582 :
5583 : size::gather_inputs(
5584 : self,
5585 : logical_sizes_at_once,
5586 : max_retention_period,
5587 : &mut shared_cache,
5588 : cause,
5589 : cancel,
5590 : ctx,
5591 : )
5592 : .await
5593 : }
5594 :
5595 : /// Calculate synthetic tenant size and cache the result.
5596 : /// This is periodically called by background worker.
5597 : /// result is cached in tenant struct
5598 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5599 : pub async fn calculate_synthetic_size(
5600 : &self,
5601 : cause: LogicalSizeCalculationCause,
5602 : cancel: &CancellationToken,
5603 : ctx: &RequestContext,
5604 : ) -> Result<u64, size::CalculateSyntheticSizeError> {
5605 : let inputs = self.gather_size_inputs(None, cause, cancel, ctx).await?;
5606 :
5607 : let size = inputs.calculate();
5608 :
5609 : self.set_cached_synthetic_size(size);
5610 :
5611 : Ok(size)
5612 : }
5613 :
5614 : /// Cache given synthetic size and update the metric value
5615 0 : pub fn set_cached_synthetic_size(&self, size: u64) {
5616 0 : self.cached_synthetic_tenant_size
5617 0 : .store(size, Ordering::Relaxed);
5618 :
5619 : // Only shard zero should be calculating synthetic sizes
5620 0 : debug_assert!(self.shard_identity.is_shard_zero());
5621 :
5622 0 : TENANT_SYNTHETIC_SIZE_METRIC
5623 0 : .get_metric_with_label_values(&[&self.tenant_shard_id.tenant_id.to_string()])
5624 0 : .unwrap()
5625 0 : .set(size);
5626 0 : }
5627 :
5628 0 : pub fn cached_synthetic_size(&self) -> u64 {
5629 0 : self.cached_synthetic_tenant_size.load(Ordering::Relaxed)
5630 0 : }
5631 :
5632 : /// Flush any in-progress layers, schedule uploads, and wait for uploads to complete.
5633 : ///
5634 : /// This function can take a long time: callers should wrap it in a timeout if calling
5635 : /// from an external API handler.
5636 : ///
5637 : /// Cancel-safety: cancelling this function may leave I/O running, but such I/O is
5638 : /// still bounded by tenant/timeline shutdown.
5639 : #[tracing::instrument(skip_all)]
5640 : pub(crate) async fn flush_remote(&self) -> anyhow::Result<()> {
5641 : let timelines = self.timelines.lock().unwrap().clone();
5642 :
5643 0 : async fn flush_timeline(_gate: GateGuard, timeline: Arc<Timeline>) -> anyhow::Result<()> {
5644 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Flushing...");
5645 0 : timeline.freeze_and_flush().await?;
5646 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Waiting for uploads...");
5647 0 : timeline.remote_client.wait_completion().await?;
5648 :
5649 0 : Ok(())
5650 0 : }
5651 :
5652 : // We do not use a JoinSet for these tasks, because we don't want them to be
5653 : // aborted when this function's future is cancelled: they should stay alive
5654 : // holding their GateGuard until they complete, to ensure their I/Os complete
5655 : // before Timeline shutdown completes.
5656 : let mut results = FuturesUnordered::new();
5657 :
5658 : for (_timeline_id, timeline) in timelines {
5659 : // Run each timeline's flush in a task holding the timeline's gate: this
5660 : // means that if this function's future is cancelled, the Timeline shutdown
5661 : // will still wait for any I/O in here to complete.
5662 : let Ok(gate) = timeline.gate.enter() else {
5663 : continue;
5664 : };
5665 0 : let jh = tokio::task::spawn(async move { flush_timeline(gate, timeline).await });
5666 : results.push(jh);
5667 : }
5668 :
5669 : while let Some(r) = results.next().await {
5670 : if let Err(e) = r {
5671 : if !e.is_cancelled() && !e.is_panic() {
5672 : tracing::error!("unexpected join error: {e:?}");
5673 : }
5674 : }
5675 : }
5676 :
5677 : // The flushes we did above were just writes, but the TenantShard might have had
5678 : // pending deletions as well from recent compaction/gc: we want to flush those
5679 : // as well. This requires flushing the global delete queue. This is cheap
5680 : // because it's typically a no-op.
5681 : match self.deletion_queue_client.flush_execute().await {
5682 : Ok(_) => {}
5683 : Err(DeletionQueueError::ShuttingDown) => {}
5684 : }
5685 :
5686 : Ok(())
5687 : }
5688 :
5689 0 : pub(crate) fn get_tenant_conf(&self) -> pageserver_api::models::TenantConfig {
5690 0 : self.tenant_conf.load().tenant_conf.clone()
5691 0 : }
5692 :
5693 : /// How much local storage would this tenant like to have? It can cope with
5694 : /// less than this (via eviction and on-demand downloads), but this function enables
5695 : /// the TenantShard to advertise how much storage it would prefer to have to provide fast I/O
5696 : /// by keeping important things on local disk.
5697 : ///
5698 : /// This is a heuristic, not a guarantee: tenants that are long-idle will actually use less
5699 : /// than they report here, due to layer eviction. Tenants with many active branches may
5700 : /// actually use more than they report here.
5701 0 : pub(crate) fn local_storage_wanted(&self) -> u64 {
5702 0 : let timelines = self.timelines.lock().unwrap();
5703 :
5704 : // Heuristic: we use the max() of the timelines' visible sizes, rather than the sum. This
5705 : // reflects the observation that on tenants with multiple large branches, typically only one
5706 : // of them is used actively enough to occupy space on disk.
5707 0 : timelines
5708 0 : .values()
5709 0 : .map(|t| t.metrics.visible_physical_size_gauge.get())
5710 0 : .max()
5711 0 : .unwrap_or(0)
5712 0 : }
5713 :
5714 : /// Builds a new tenant manifest, and uploads it if it differs from the last-known tenant
5715 : /// manifest in `Self::remote_tenant_manifest`.
5716 : ///
5717 : /// TODO: instead of requiring callers to remember to call `maybe_upload_tenant_manifest` after
5718 : /// changing any `TenantShard` state that's included in the manifest, consider making the manifest
5719 : /// the authoritative source of data with an API that automatically uploads on changes. Revisit
5720 : /// this when the manifest is more widely used and we have a better idea of the data model.
5721 119 : pub(crate) async fn maybe_upload_tenant_manifest(&self) -> Result<(), TenantManifestError> {
5722 : // Multiple tasks may call this function concurrently after mutating the TenantShard runtime
5723 : // state, affecting the manifest generated by `build_tenant_manifest`. We use an async mutex
5724 : // to serialize these callers. `eq_ignoring_version` acts as a slightly inefficient but
5725 : // simple coalescing mechanism.
5726 119 : let mut guard = tokio::select! {
5727 119 : guard = self.remote_tenant_manifest.lock() => guard,
5728 119 : _ = self.cancel.cancelled() => return Err(TenantManifestError::Cancelled),
5729 : };
5730 :
5731 : // Build a new manifest.
5732 119 : let manifest = self.build_tenant_manifest();
5733 :
5734 : // Check if the manifest has changed. We ignore the version number here, to avoid
5735 : // uploading every manifest on version number bumps.
5736 119 : if let Some(old) = guard.as_ref() {
5737 4 : if manifest.eq_ignoring_version(old) {
5738 3 : return Ok(());
5739 1 : }
5740 115 : }
5741 :
5742 : // Update metrics
5743 116 : let tid = self.tenant_shard_id.to_string();
5744 116 : let shard_id = self.tenant_shard_id.shard_slug().to_string();
5745 116 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
5746 116 : TENANT_OFFLOADED_TIMELINES
5747 116 : .with_label_values(set_key)
5748 116 : .set(manifest.offloaded_timelines.len() as u64);
5749 :
5750 : // Upload the manifest. Remote storage does no retries internally, so retry here.
5751 116 : match backoff::retry(
5752 116 : || async {
5753 116 : upload_tenant_manifest(
5754 116 : &self.remote_storage,
5755 116 : &self.tenant_shard_id,
5756 116 : self.generation,
5757 116 : &manifest,
5758 116 : &self.cancel,
5759 116 : )
5760 116 : .await
5761 232 : },
5762 0 : |_| self.cancel.is_cancelled(),
5763 : FAILED_UPLOAD_WARN_THRESHOLD,
5764 : FAILED_REMOTE_OP_RETRIES,
5765 116 : "uploading tenant manifest",
5766 116 : &self.cancel,
5767 : )
5768 116 : .await
5769 : {
5770 0 : None => Err(TenantManifestError::Cancelled),
5771 0 : Some(Err(_)) if self.cancel.is_cancelled() => Err(TenantManifestError::Cancelled),
5772 0 : Some(Err(e)) => Err(TenantManifestError::RemoteStorage(e)),
5773 : Some(Ok(_)) => {
5774 : // Store the successfully uploaded manifest, so that future callers can avoid
5775 : // re-uploading the same thing.
5776 116 : *guard = Some(manifest);
5777 :
5778 116 : Ok(())
5779 : }
5780 : }
5781 119 : }
5782 : }
5783 :
5784 : /// Create the cluster temporarily in 'initdbpath' directory inside the repository
5785 : /// to get bootstrap data for timeline initialization.
5786 0 : async fn run_initdb(
5787 0 : conf: &'static PageServerConf,
5788 0 : initdb_target_dir: &Utf8Path,
5789 0 : pg_version: PgMajorVersion,
5790 0 : cancel: &CancellationToken,
5791 0 : ) -> Result<(), InitdbError> {
5792 0 : let initdb_bin_path = conf
5793 0 : .pg_bin_dir(pg_version)
5794 0 : .map_err(InitdbError::Other)?
5795 0 : .join("initdb");
5796 0 : let initdb_lib_dir = conf.pg_lib_dir(pg_version).map_err(InitdbError::Other)?;
5797 0 : info!(
5798 0 : "running {} in {}, libdir: {}",
5799 : initdb_bin_path, initdb_target_dir, initdb_lib_dir,
5800 : );
5801 :
5802 0 : let _permit = {
5803 0 : let _timer = INITDB_SEMAPHORE_ACQUISITION_TIME.start_timer();
5804 0 : INIT_DB_SEMAPHORE.acquire().await
5805 : };
5806 :
5807 0 : CONCURRENT_INITDBS.inc();
5808 0 : scopeguard::defer! {
5809 : CONCURRENT_INITDBS.dec();
5810 : }
5811 :
5812 0 : let _timer = INITDB_RUN_TIME.start_timer();
5813 0 : let res = postgres_initdb::do_run_initdb(postgres_initdb::RunInitdbArgs {
5814 0 : superuser: &conf.superuser,
5815 0 : locale: &conf.locale,
5816 0 : initdb_bin: &initdb_bin_path,
5817 0 : pg_version,
5818 0 : library_search_path: &initdb_lib_dir,
5819 0 : pgdata: initdb_target_dir,
5820 0 : })
5821 0 : .await
5822 0 : .map_err(InitdbError::Inner);
5823 :
5824 : // This isn't true cancellation support, see above. Still return an error to
5825 : // excercise the cancellation code path.
5826 0 : if cancel.is_cancelled() {
5827 0 : return Err(InitdbError::Cancelled);
5828 0 : }
5829 :
5830 0 : res
5831 0 : }
5832 :
5833 : /// Dump contents of a layer file to stdout.
5834 0 : pub async fn dump_layerfile_from_path(
5835 0 : path: &Utf8Path,
5836 0 : verbose: bool,
5837 0 : ctx: &RequestContext,
5838 0 : ) -> anyhow::Result<()> {
5839 : use std::os::unix::fs::FileExt;
5840 :
5841 : // All layer files start with a two-byte "magic" value, to identify the kind of
5842 : // file.
5843 0 : let file = File::open(path)?;
5844 0 : let mut header_buf = [0u8; 2];
5845 0 : file.read_exact_at(&mut header_buf, 0)?;
5846 :
5847 0 : match u16::from_be_bytes(header_buf) {
5848 : crate::IMAGE_FILE_MAGIC => {
5849 0 : ImageLayer::new_for_path(path, file)?
5850 0 : .dump(verbose, ctx)
5851 0 : .await?
5852 : }
5853 : crate::DELTA_FILE_MAGIC => {
5854 0 : DeltaLayer::new_for_path(path, file)?
5855 0 : .dump(verbose, ctx)
5856 0 : .await?
5857 : }
5858 0 : magic => bail!("unrecognized magic identifier: {:?}", magic),
5859 : }
5860 :
5861 0 : Ok(())
5862 0 : }
5863 :
5864 : #[cfg(test)]
5865 : pub(crate) mod harness {
5866 : use bytes::{Bytes, BytesMut};
5867 : use hex_literal::hex;
5868 : use once_cell::sync::OnceCell;
5869 : use pageserver_api::key::Key;
5870 : use pageserver_api::models::ShardParameters;
5871 : use pageserver_api::shard::ShardIndex;
5872 : use utils::id::TenantId;
5873 : use utils::logging;
5874 : use wal_decoder::models::record::NeonWalRecord;
5875 :
5876 : use super::*;
5877 : use crate::deletion_queue::mock::MockDeletionQueue;
5878 : use crate::l0_flush::L0FlushConfig;
5879 : use crate::walredo::apply_neon;
5880 :
5881 : pub const TIMELINE_ID: TimelineId =
5882 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
5883 : pub const NEW_TIMELINE_ID: TimelineId =
5884 : TimelineId::from_array(hex!("AA223344556677881122334455667788"));
5885 :
5886 : /// Convenience function to create a page image with given string as the only content
5887 2514347 : pub fn test_img(s: &str) -> Bytes {
5888 2514347 : let mut buf = BytesMut::new();
5889 2514347 : buf.extend_from_slice(s.as_bytes());
5890 2514347 : buf.resize(64, 0);
5891 :
5892 2514347 : buf.freeze()
5893 2514347 : }
5894 :
5895 : pub struct TenantHarness {
5896 : pub conf: &'static PageServerConf,
5897 : pub tenant_conf: pageserver_api::models::TenantConfig,
5898 : pub tenant_shard_id: TenantShardId,
5899 : pub shard_identity: ShardIdentity,
5900 : pub generation: Generation,
5901 : pub shard: ShardIndex,
5902 : pub remote_storage: GenericRemoteStorage,
5903 : pub remote_fs_dir: Utf8PathBuf,
5904 : pub deletion_queue: MockDeletionQueue,
5905 : }
5906 :
5907 : static LOG_HANDLE: OnceCell<()> = OnceCell::new();
5908 :
5909 130 : pub(crate) fn setup_logging() {
5910 130 : LOG_HANDLE.get_or_init(|| {
5911 124 : logging::init(
5912 124 : logging::LogFormat::Test,
5913 : // enable it in case the tests exercise code paths that use
5914 : // debug_assert_current_span_has_tenant_and_timeline_id
5915 124 : logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
5916 124 : logging::Output::Stdout,
5917 : )
5918 124 : .expect("Failed to init test logging");
5919 124 : });
5920 130 : }
5921 :
5922 : impl TenantHarness {
5923 118 : pub async fn create_custom(
5924 118 : test_name: &'static str,
5925 118 : tenant_conf: pageserver_api::models::TenantConfig,
5926 118 : tenant_id: TenantId,
5927 118 : shard_identity: ShardIdentity,
5928 118 : generation: Generation,
5929 118 : ) -> anyhow::Result<Self> {
5930 118 : setup_logging();
5931 :
5932 118 : let repo_dir = PageServerConf::test_repo_dir(test_name);
5933 118 : let _ = fs::remove_dir_all(&repo_dir);
5934 118 : fs::create_dir_all(&repo_dir)?;
5935 :
5936 118 : let conf = PageServerConf::dummy_conf(repo_dir);
5937 : // Make a static copy of the config. This can never be free'd, but that's
5938 : // OK in a test.
5939 118 : let conf: &'static PageServerConf = Box::leak(Box::new(conf));
5940 :
5941 118 : let shard = shard_identity.shard_index();
5942 118 : let tenant_shard_id = TenantShardId {
5943 118 : tenant_id,
5944 118 : shard_number: shard.shard_number,
5945 118 : shard_count: shard.shard_count,
5946 118 : };
5947 118 : fs::create_dir_all(conf.tenant_path(&tenant_shard_id))?;
5948 118 : fs::create_dir_all(conf.timelines_path(&tenant_shard_id))?;
5949 :
5950 : use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
5951 118 : let remote_fs_dir = conf.workdir.join("localfs");
5952 118 : std::fs::create_dir_all(&remote_fs_dir).unwrap();
5953 118 : let config = RemoteStorageConfig {
5954 118 : storage: RemoteStorageKind::LocalFs {
5955 118 : local_path: remote_fs_dir.clone(),
5956 118 : },
5957 118 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
5958 118 : small_timeout: RemoteStorageConfig::DEFAULT_SMALL_TIMEOUT,
5959 118 : };
5960 118 : let remote_storage = GenericRemoteStorage::from_config(&config).await.unwrap();
5961 118 : let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()));
5962 :
5963 118 : Ok(Self {
5964 118 : conf,
5965 118 : tenant_conf,
5966 118 : tenant_shard_id,
5967 118 : shard_identity,
5968 118 : generation,
5969 118 : shard,
5970 118 : remote_storage,
5971 118 : remote_fs_dir,
5972 118 : deletion_queue,
5973 118 : })
5974 118 : }
5975 :
5976 110 : pub async fn create(test_name: &'static str) -> anyhow::Result<Self> {
5977 : // Disable automatic GC and compaction to make the unit tests more deterministic.
5978 : // The tests perform them manually if needed.
5979 110 : let tenant_conf = pageserver_api::models::TenantConfig {
5980 110 : gc_period: Some(Duration::ZERO),
5981 110 : compaction_period: Some(Duration::ZERO),
5982 110 : ..Default::default()
5983 110 : };
5984 110 : let tenant_id = TenantId::generate();
5985 110 : let shard = ShardIdentity::unsharded();
5986 110 : Self::create_custom(
5987 110 : test_name,
5988 110 : tenant_conf,
5989 110 : tenant_id,
5990 110 : shard,
5991 110 : Generation::new(0xdeadbeef),
5992 110 : )
5993 110 : .await
5994 110 : }
5995 :
5996 10 : pub fn span(&self) -> tracing::Span {
5997 10 : info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
5998 10 : }
5999 :
6000 118 : pub(crate) async fn load(&self) -> (Arc<TenantShard>, RequestContext) {
6001 118 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error)
6002 118 : .with_scope_unit_test();
6003 : (
6004 118 : self.do_try_load(&ctx)
6005 118 : .await
6006 118 : .expect("failed to load test tenant"),
6007 118 : ctx,
6008 : )
6009 118 : }
6010 :
6011 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
6012 : pub(crate) async fn do_try_load(
6013 : &self,
6014 : ctx: &RequestContext,
6015 : ) -> anyhow::Result<Arc<TenantShard>> {
6016 : let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
6017 :
6018 : let (basebackup_cache, _) = BasebackupCache::new(Utf8PathBuf::new(), None);
6019 :
6020 : let tenant = Arc::new(TenantShard::new(
6021 : TenantState::Attaching,
6022 : self.conf,
6023 : AttachedTenantConf::try_from(LocationConf::attached_single(
6024 : self.tenant_conf.clone(),
6025 : self.generation,
6026 : ShardParameters::default(),
6027 : ))
6028 : .unwrap(),
6029 : self.shard_identity,
6030 : Some(walredo_mgr),
6031 : self.tenant_shard_id,
6032 : self.remote_storage.clone(),
6033 : self.deletion_queue.new_client(),
6034 : // TODO: ideally we should run all unit tests with both configs
6035 : L0FlushGlobalState::new(L0FlushConfig::default()),
6036 : basebackup_cache,
6037 : FeatureResolver::new_disabled(),
6038 : ));
6039 :
6040 : let preload = tenant
6041 : .preload(&self.remote_storage, CancellationToken::new())
6042 : .await?;
6043 : tenant.attach(Some(preload), ctx).await?;
6044 :
6045 : tenant.state.send_replace(TenantState::Active);
6046 : for timeline in tenant.timelines.lock().unwrap().values() {
6047 : timeline.set_state(TimelineState::Active);
6048 : }
6049 : Ok(tenant)
6050 : }
6051 :
6052 1 : pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
6053 1 : self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
6054 1 : }
6055 : }
6056 :
6057 : // Mock WAL redo manager that doesn't do much
6058 : pub(crate) struct TestRedoManager;
6059 :
6060 : impl TestRedoManager {
6061 : /// # Cancel-Safety
6062 : ///
6063 : /// This method is cancellation-safe.
6064 26774 : pub async fn request_redo(
6065 26774 : &self,
6066 26774 : key: Key,
6067 26774 : lsn: Lsn,
6068 26774 : base_img: Option<(Lsn, Bytes)>,
6069 26774 : records: Vec<(Lsn, NeonWalRecord)>,
6070 26774 : _pg_version: PgMajorVersion,
6071 26774 : _redo_attempt_type: RedoAttemptType,
6072 26774 : ) -> Result<Bytes, walredo::Error> {
6073 1403510 : let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
6074 26774 : if records_neon {
6075 : // For Neon wal records, we can decode without spawning postgres, so do so.
6076 26774 : let mut page = match (base_img, records.first()) {
6077 13029 : (Some((_lsn, img)), _) => {
6078 13029 : let mut page = BytesMut::new();
6079 13029 : page.extend_from_slice(&img);
6080 13029 : page
6081 : }
6082 13745 : (_, Some((_lsn, rec))) if rec.will_init() => BytesMut::new(),
6083 : _ => {
6084 0 : panic!("Neon WAL redo requires base image or will init record");
6085 : }
6086 : };
6087 :
6088 1430283 : for (record_lsn, record) in records {
6089 1403510 : apply_neon::apply_in_neon(&record, record_lsn, key, &mut page)?;
6090 : }
6091 26773 : Ok(page.freeze())
6092 : } else {
6093 : // We never spawn a postgres walredo process in unit tests: just log what we might have done.
6094 0 : let s = format!(
6095 0 : "redo for {} to get to {}, with {} and {} records",
6096 : key,
6097 : lsn,
6098 0 : if base_img.is_some() {
6099 0 : "base image"
6100 : } else {
6101 0 : "no base image"
6102 : },
6103 0 : records.len()
6104 : );
6105 0 : println!("{s}");
6106 :
6107 0 : Ok(test_img(&s))
6108 : }
6109 26774 : }
6110 : }
6111 : }
6112 :
6113 : #[cfg(test)]
6114 : mod tests {
6115 : use std::collections::{BTreeMap, BTreeSet};
6116 :
6117 : use bytes::{Bytes, BytesMut};
6118 : use hex_literal::hex;
6119 : use itertools::Itertools;
6120 : #[cfg(feature = "testing")]
6121 : use models::CompactLsnRange;
6122 : use pageserver_api::key::{
6123 : AUX_KEY_PREFIX, Key, NON_INHERITED_RANGE, RELATION_SIZE_PREFIX, repl_origin_key,
6124 : };
6125 : use pageserver_api::keyspace::KeySpace;
6126 : #[cfg(feature = "testing")]
6127 : use pageserver_api::keyspace::KeySpaceRandomAccum;
6128 : use pageserver_api::models::{CompactionAlgorithm, CompactionAlgorithmSettings};
6129 : use pageserver_compaction::helpers::overlaps_with;
6130 : #[cfg(feature = "testing")]
6131 : use rand::SeedableRng;
6132 : #[cfg(feature = "testing")]
6133 : use rand::rngs::StdRng;
6134 : use rand::{Rng, thread_rng};
6135 : #[cfg(feature = "testing")]
6136 : use std::ops::Range;
6137 : use storage_layer::{IoConcurrency, PersistentLayerKey};
6138 : use tests::storage_layer::ValuesReconstructState;
6139 : use tests::timeline::{GetVectoredError, ShutdownMode};
6140 : #[cfg(feature = "testing")]
6141 : use timeline::GcInfo;
6142 : #[cfg(feature = "testing")]
6143 : use timeline::InMemoryLayerTestDesc;
6144 : #[cfg(feature = "testing")]
6145 : use timeline::compaction::{KeyHistoryRetention, KeyLogAtLsn};
6146 : use timeline::{CompactOptions, DeltaLayerTestDesc, VersionedKeySpaceQuery};
6147 : use utils::id::TenantId;
6148 : use utils::shard::{ShardCount, ShardNumber};
6149 : #[cfg(feature = "testing")]
6150 : use wal_decoder::models::record::NeonWalRecord;
6151 : use wal_decoder::models::value::Value;
6152 :
6153 : use super::*;
6154 : use crate::DEFAULT_PG_VERSION;
6155 : use crate::keyspace::KeySpaceAccum;
6156 : use crate::tenant::harness::*;
6157 : use crate::tenant::timeline::CompactFlags;
6158 :
6159 : static TEST_KEY: Lazy<Key> =
6160 10 : Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));
6161 :
6162 : #[cfg(feature = "testing")]
6163 : struct TestTimelineSpecification {
6164 : start_lsn: Lsn,
6165 : last_record_lsn: Lsn,
6166 :
6167 : in_memory_layers_shape: Vec<(Range<Key>, Range<Lsn>)>,
6168 : delta_layers_shape: Vec<(Range<Key>, Range<Lsn>)>,
6169 : image_layers_shape: Vec<(Range<Key>, Lsn)>,
6170 :
6171 : gap_chance: u8,
6172 : will_init_chance: u8,
6173 : }
6174 :
6175 : #[cfg(feature = "testing")]
6176 : struct Storage {
6177 : storage: HashMap<(Key, Lsn), Value>,
6178 : start_lsn: Lsn,
6179 : }
6180 :
6181 : #[cfg(feature = "testing")]
6182 : impl Storage {
6183 32000 : fn get(&self, key: Key, lsn: Lsn) -> Bytes {
6184 : use bytes::BufMut;
6185 :
6186 32000 : let mut crnt_lsn = lsn;
6187 32000 : let mut got_base = false;
6188 :
6189 32000 : let mut acc = Vec::new();
6190 :
6191 2831871 : while crnt_lsn >= self.start_lsn {
6192 2831871 : if let Some(value) = self.storage.get(&(key, crnt_lsn)) {
6193 1421172 : acc.push(value.clone());
6194 :
6195 1402881 : match value {
6196 1402881 : Value::WalRecord(NeonWalRecord::Test { will_init, .. }) => {
6197 1402881 : if *will_init {
6198 13709 : got_base = true;
6199 13709 : break;
6200 1389172 : }
6201 : }
6202 : Value::Image(_) => {
6203 18291 : got_base = true;
6204 18291 : break;
6205 : }
6206 0 : _ => unreachable!(),
6207 : }
6208 1410699 : }
6209 :
6210 2799871 : crnt_lsn = crnt_lsn.checked_sub(1u64).unwrap();
6211 : }
6212 :
6213 32000 : assert!(
6214 32000 : got_base,
6215 0 : "Input data was incorrect. No base image for {key}@{lsn}"
6216 : );
6217 :
6218 32000 : tracing::debug!("Wal redo depth for {key}@{lsn} is {}", acc.len());
6219 :
6220 32000 : let mut blob = BytesMut::new();
6221 1421172 : for value in acc.into_iter().rev() {
6222 1402881 : match value {
6223 1402881 : Value::WalRecord(NeonWalRecord::Test { append, .. }) => {
6224 1402881 : blob.extend_from_slice(append.as_bytes());
6225 1402881 : }
6226 18291 : Value::Image(img) => {
6227 18291 : blob.put(img);
6228 18291 : }
6229 0 : _ => unreachable!(),
6230 : }
6231 : }
6232 :
6233 32000 : blob.into()
6234 32000 : }
6235 : }
6236 :
6237 : #[cfg(feature = "testing")]
6238 : #[allow(clippy::too_many_arguments)]
6239 1 : async fn randomize_timeline(
6240 1 : tenant: &Arc<TenantShard>,
6241 1 : new_timeline_id: TimelineId,
6242 1 : pg_version: PgMajorVersion,
6243 1 : spec: TestTimelineSpecification,
6244 1 : random: &mut rand::rngs::StdRng,
6245 1 : ctx: &RequestContext,
6246 1 : ) -> anyhow::Result<(Arc<Timeline>, Storage, Vec<Lsn>)> {
6247 1 : let mut storage: HashMap<(Key, Lsn), Value> = HashMap::default();
6248 1 : let mut interesting_lsns = vec![spec.last_record_lsn];
6249 :
6250 2 : for (key_range, lsn_range) in spec.in_memory_layers_shape.iter() {
6251 2 : let mut lsn = lsn_range.start;
6252 202 : while lsn < lsn_range.end {
6253 200 : let mut key = key_range.start;
6254 21018 : while key < key_range.end {
6255 20818 : let gap = random.gen_range(1..=100) <= spec.gap_chance;
6256 20818 : let will_init = random.gen_range(1..=100) <= spec.will_init_chance;
6257 :
6258 20818 : if gap {
6259 1018 : continue;
6260 19800 : }
6261 :
6262 19800 : let record = if will_init {
6263 191 : Value::WalRecord(NeonWalRecord::wal_init(format!("[wil_init {key}@{lsn}]")))
6264 : } else {
6265 19609 : Value::WalRecord(NeonWalRecord::wal_append(format!("[delta {key}@{lsn}]")))
6266 : };
6267 :
6268 19800 : storage.insert((key, lsn), record);
6269 :
6270 19800 : key = key.next();
6271 : }
6272 200 : lsn = Lsn(lsn.0 + 1);
6273 : }
6274 :
6275 : // Stash some interesting LSN for future use
6276 6 : for offset in [0, 5, 100].iter() {
6277 6 : if *offset == 0 {
6278 2 : interesting_lsns.push(lsn_range.start);
6279 2 : } else {
6280 4 : let below = lsn_range.start.checked_sub(*offset);
6281 4 : match below {
6282 4 : Some(v) if v >= spec.start_lsn => {
6283 4 : interesting_lsns.push(v);
6284 4 : }
6285 0 : _ => {}
6286 : }
6287 :
6288 4 : let above = Lsn(lsn_range.start.0 + offset);
6289 4 : interesting_lsns.push(above);
6290 : }
6291 : }
6292 : }
6293 :
6294 3 : for (key_range, lsn_range) in spec.delta_layers_shape.iter() {
6295 3 : let mut lsn = lsn_range.start;
6296 315 : while lsn < lsn_range.end {
6297 312 : let mut key = key_range.start;
6298 11112 : while key < key_range.end {
6299 10800 : let gap = random.gen_range(1..=100) <= spec.gap_chance;
6300 10800 : let will_init = random.gen_range(1..=100) <= spec.will_init_chance;
6301 :
6302 10800 : if gap {
6303 504 : continue;
6304 10296 : }
6305 :
6306 10296 : let record = if will_init {
6307 103 : Value::WalRecord(NeonWalRecord::wal_init(format!("[wil_init {key}@{lsn}]")))
6308 : } else {
6309 10193 : Value::WalRecord(NeonWalRecord::wal_append(format!("[delta {key}@{lsn}]")))
6310 : };
6311 :
6312 10296 : storage.insert((key, lsn), record);
6313 :
6314 10296 : key = key.next();
6315 : }
6316 312 : lsn = Lsn(lsn.0 + 1);
6317 : }
6318 :
6319 : // Stash some interesting LSN for future use
6320 9 : for offset in [0, 5, 100].iter() {
6321 9 : if *offset == 0 {
6322 3 : interesting_lsns.push(lsn_range.start);
6323 3 : } else {
6324 6 : let below = lsn_range.start.checked_sub(*offset);
6325 6 : match below {
6326 6 : Some(v) if v >= spec.start_lsn => {
6327 3 : interesting_lsns.push(v);
6328 3 : }
6329 3 : _ => {}
6330 : }
6331 :
6332 6 : let above = Lsn(lsn_range.start.0 + offset);
6333 6 : interesting_lsns.push(above);
6334 : }
6335 : }
6336 : }
6337 :
6338 3 : for (key_range, lsn) in spec.image_layers_shape.iter() {
6339 3 : let mut key = key_range.start;
6340 142 : while key < key_range.end {
6341 139 : let blob = Bytes::from(format!("[image {key}@{lsn}]"));
6342 139 : let record = Value::Image(blob.clone());
6343 139 : storage.insert((key, *lsn), record);
6344 139 :
6345 139 : key = key.next();
6346 139 : }
6347 :
6348 : // Stash some interesting LSN for future use
6349 9 : for offset in [0, 5, 100].iter() {
6350 9 : if *offset == 0 {
6351 3 : interesting_lsns.push(*lsn);
6352 3 : } else {
6353 6 : let below = lsn.checked_sub(*offset);
6354 6 : match below {
6355 6 : Some(v) if v >= spec.start_lsn => {
6356 4 : interesting_lsns.push(v);
6357 4 : }
6358 2 : _ => {}
6359 : }
6360 :
6361 6 : let above = Lsn(lsn.0 + offset);
6362 6 : interesting_lsns.push(above);
6363 : }
6364 : }
6365 : }
6366 :
6367 1 : let in_memory_test_layers = {
6368 1 : let mut acc = Vec::new();
6369 :
6370 2 : for (key_range, lsn_range) in spec.in_memory_layers_shape.iter() {
6371 2 : let mut data = Vec::new();
6372 :
6373 2 : let mut lsn = lsn_range.start;
6374 202 : while lsn < lsn_range.end {
6375 200 : let mut key = key_range.start;
6376 20000 : while key < key_range.end {
6377 19800 : if let Some(record) = storage.get(&(key, lsn)) {
6378 19800 : data.push((key, lsn, record.clone()));
6379 19800 : }
6380 :
6381 19800 : key = key.next();
6382 : }
6383 200 : lsn = Lsn(lsn.0 + 1);
6384 : }
6385 :
6386 2 : acc.push(InMemoryLayerTestDesc {
6387 2 : data,
6388 2 : lsn_range: lsn_range.clone(),
6389 2 : is_open: false,
6390 2 : })
6391 : }
6392 :
6393 1 : acc
6394 : };
6395 :
6396 1 : let delta_test_layers = {
6397 1 : let mut acc = Vec::new();
6398 :
6399 3 : for (key_range, lsn_range) in spec.delta_layers_shape.iter() {
6400 3 : let mut data = Vec::new();
6401 :
6402 3 : let mut lsn = lsn_range.start;
6403 315 : while lsn < lsn_range.end {
6404 312 : let mut key = key_range.start;
6405 10608 : while key < key_range.end {
6406 10296 : if let Some(record) = storage.get(&(key, lsn)) {
6407 10296 : data.push((key, lsn, record.clone()));
6408 10296 : }
6409 :
6410 10296 : key = key.next();
6411 : }
6412 312 : lsn = Lsn(lsn.0 + 1);
6413 : }
6414 :
6415 3 : acc.push(DeltaLayerTestDesc {
6416 3 : data,
6417 3 : lsn_range: lsn_range.clone(),
6418 3 : key_range: key_range.clone(),
6419 3 : })
6420 : }
6421 :
6422 1 : acc
6423 : };
6424 :
6425 1 : let image_test_layers = {
6426 1 : let mut acc = Vec::new();
6427 :
6428 3 : for (key_range, lsn) in spec.image_layers_shape.iter() {
6429 3 : let mut data = Vec::new();
6430 :
6431 3 : let mut key = key_range.start;
6432 142 : while key < key_range.end {
6433 139 : if let Some(record) = storage.get(&(key, *lsn)) {
6434 139 : let blob = match record {
6435 139 : Value::Image(blob) => blob.clone(),
6436 0 : _ => unreachable!(),
6437 : };
6438 :
6439 139 : data.push((key, blob));
6440 0 : }
6441 :
6442 139 : key = key.next();
6443 : }
6444 :
6445 3 : acc.push((*lsn, data));
6446 : }
6447 :
6448 1 : acc
6449 : };
6450 :
6451 1 : let tline = tenant
6452 1 : .create_test_timeline_with_layers(
6453 1 : new_timeline_id,
6454 1 : spec.start_lsn,
6455 1 : pg_version,
6456 1 : ctx,
6457 1 : in_memory_test_layers,
6458 1 : delta_test_layers,
6459 1 : image_test_layers,
6460 1 : spec.last_record_lsn,
6461 1 : )
6462 1 : .await?;
6463 :
6464 1 : Ok((
6465 1 : tline,
6466 1 : Storage {
6467 1 : storage,
6468 1 : start_lsn: spec.start_lsn,
6469 1 : },
6470 1 : interesting_lsns,
6471 1 : ))
6472 1 : }
6473 :
6474 : #[tokio::test]
6475 1 : async fn test_basic() -> anyhow::Result<()> {
6476 1 : let (tenant, ctx) = TenantHarness::create("test_basic").await?.load().await;
6477 1 : let tline = tenant
6478 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6479 1 : .await?;
6480 :
6481 1 : let mut writer = tline.writer().await;
6482 1 : writer
6483 1 : .put(
6484 1 : *TEST_KEY,
6485 1 : Lsn(0x10),
6486 1 : &Value::Image(test_img("foo at 0x10")),
6487 1 : &ctx,
6488 1 : )
6489 1 : .await?;
6490 1 : writer.finish_write(Lsn(0x10));
6491 1 : drop(writer);
6492 :
6493 1 : let mut writer = tline.writer().await;
6494 1 : writer
6495 1 : .put(
6496 1 : *TEST_KEY,
6497 1 : Lsn(0x20),
6498 1 : &Value::Image(test_img("foo at 0x20")),
6499 1 : &ctx,
6500 1 : )
6501 1 : .await?;
6502 1 : writer.finish_write(Lsn(0x20));
6503 1 : drop(writer);
6504 :
6505 1 : assert_eq!(
6506 1 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
6507 1 : test_img("foo at 0x10")
6508 : );
6509 1 : assert_eq!(
6510 1 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
6511 1 : test_img("foo at 0x10")
6512 : );
6513 1 : assert_eq!(
6514 1 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
6515 1 : test_img("foo at 0x20")
6516 : );
6517 :
6518 2 : Ok(())
6519 1 : }
6520 :
6521 : #[tokio::test]
6522 1 : async fn no_duplicate_timelines() -> anyhow::Result<()> {
6523 1 : let (tenant, ctx) = TenantHarness::create("no_duplicate_timelines")
6524 1 : .await?
6525 1 : .load()
6526 1 : .await;
6527 1 : let _ = tenant
6528 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6529 1 : .await?;
6530 :
6531 1 : match tenant
6532 1 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6533 1 : .await
6534 1 : {
6535 1 : Ok(_) => panic!("duplicate timeline creation should fail"),
6536 1 : Err(e) => assert_eq!(
6537 1 : e.to_string(),
6538 1 : "timeline already exists with different parameters".to_string()
6539 1 : ),
6540 1 : }
6541 1 :
6542 1 : Ok(())
6543 1 : }
6544 :
6545 : /// Convenience function to create a page image with given string as the only content
6546 5 : pub fn test_value(s: &str) -> Value {
6547 5 : let mut buf = BytesMut::new();
6548 5 : buf.extend_from_slice(s.as_bytes());
6549 5 : Value::Image(buf.freeze())
6550 5 : }
6551 :
6552 : ///
6553 : /// Test branch creation
6554 : ///
6555 : #[tokio::test]
6556 1 : async fn test_branch() -> anyhow::Result<()> {
6557 : use std::str::from_utf8;
6558 :
6559 1 : let (tenant, ctx) = TenantHarness::create("test_branch").await?.load().await;
6560 1 : let tline = tenant
6561 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6562 1 : .await?;
6563 1 : let mut writer = tline.writer().await;
6564 :
6565 : #[allow(non_snake_case)]
6566 1 : let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
6567 : #[allow(non_snake_case)]
6568 1 : let TEST_KEY_B: Key = Key::from_hex("110000000033333333444444445500000002").unwrap();
6569 :
6570 : // Insert a value on the timeline
6571 1 : writer
6572 1 : .put(TEST_KEY_A, Lsn(0x20), &test_value("foo at 0x20"), &ctx)
6573 1 : .await?;
6574 1 : writer
6575 1 : .put(TEST_KEY_B, Lsn(0x20), &test_value("foobar at 0x20"), &ctx)
6576 1 : .await?;
6577 1 : writer.finish_write(Lsn(0x20));
6578 :
6579 1 : writer
6580 1 : .put(TEST_KEY_A, Lsn(0x30), &test_value("foo at 0x30"), &ctx)
6581 1 : .await?;
6582 1 : writer.finish_write(Lsn(0x30));
6583 1 : writer
6584 1 : .put(TEST_KEY_A, Lsn(0x40), &test_value("foo at 0x40"), &ctx)
6585 1 : .await?;
6586 1 : writer.finish_write(Lsn(0x40));
6587 :
6588 : //assert_current_logical_size(&tline, Lsn(0x40));
6589 :
6590 : // Branch the history, modify relation differently on the new timeline
6591 1 : tenant
6592 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x30)), &ctx)
6593 1 : .await?;
6594 1 : let newtline = tenant
6595 1 : .get_timeline(NEW_TIMELINE_ID, true)
6596 1 : .expect("Should have a local timeline");
6597 1 : let mut new_writer = newtline.writer().await;
6598 1 : new_writer
6599 1 : .put(TEST_KEY_A, Lsn(0x40), &test_value("bar at 0x40"), &ctx)
6600 1 : .await?;
6601 1 : new_writer.finish_write(Lsn(0x40));
6602 :
6603 : // Check page contents on both branches
6604 1 : assert_eq!(
6605 1 : from_utf8(&tline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
6606 : "foo at 0x40"
6607 : );
6608 1 : assert_eq!(
6609 1 : from_utf8(&newtline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
6610 : "bar at 0x40"
6611 : );
6612 1 : assert_eq!(
6613 1 : from_utf8(&newtline.get(TEST_KEY_B, Lsn(0x40), &ctx).await?)?,
6614 : "foobar at 0x20"
6615 : );
6616 :
6617 : //assert_current_logical_size(&tline, Lsn(0x40));
6618 :
6619 2 : Ok(())
6620 1 : }
6621 :
6622 10 : async fn make_some_layers(
6623 10 : tline: &Timeline,
6624 10 : start_lsn: Lsn,
6625 10 : ctx: &RequestContext,
6626 10 : ) -> anyhow::Result<()> {
6627 10 : let mut lsn = start_lsn;
6628 : {
6629 10 : let mut writer = tline.writer().await;
6630 : // Create a relation on the timeline
6631 10 : writer
6632 10 : .put(
6633 10 : *TEST_KEY,
6634 10 : lsn,
6635 10 : &Value::Image(test_img(&format!("foo at {lsn}"))),
6636 10 : ctx,
6637 10 : )
6638 10 : .await?;
6639 10 : writer.finish_write(lsn);
6640 10 : lsn += 0x10;
6641 10 : writer
6642 10 : .put(
6643 10 : *TEST_KEY,
6644 10 : lsn,
6645 10 : &Value::Image(test_img(&format!("foo at {lsn}"))),
6646 10 : ctx,
6647 10 : )
6648 10 : .await?;
6649 10 : writer.finish_write(lsn);
6650 10 : lsn += 0x10;
6651 : }
6652 10 : tline.freeze_and_flush().await?;
6653 : {
6654 10 : let mut writer = tline.writer().await;
6655 10 : writer
6656 10 : .put(
6657 10 : *TEST_KEY,
6658 10 : lsn,
6659 10 : &Value::Image(test_img(&format!("foo at {lsn}"))),
6660 10 : ctx,
6661 10 : )
6662 10 : .await?;
6663 10 : writer.finish_write(lsn);
6664 10 : lsn += 0x10;
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 : }
6675 10 : tline.freeze_and_flush().await.map_err(|e| e.into())
6676 10 : }
6677 :
6678 : #[tokio::test(start_paused = true)]
6679 1 : async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
6680 1 : let (tenant, ctx) =
6681 1 : TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
6682 1 : .await?
6683 1 : .load()
6684 1 : .await;
6685 : // Advance to the lsn lease deadline so that GC is not blocked by
6686 : // initial transition into AttachedSingle.
6687 1 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
6688 1 : tokio::time::resume();
6689 1 : let tline = tenant
6690 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6691 1 : .await?;
6692 1 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6693 :
6694 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
6695 : // FIXME: this doesn't actually remove any layer currently, given how the flushing
6696 : // and compaction works. But it does set the 'cutoff' point so that the cross check
6697 : // below should fail.
6698 1 : tenant
6699 1 : .gc_iteration(
6700 1 : Some(TIMELINE_ID),
6701 1 : 0x10,
6702 1 : Duration::ZERO,
6703 1 : &CancellationToken::new(),
6704 1 : &ctx,
6705 1 : )
6706 1 : .await?;
6707 :
6708 : // try to branch at lsn 25, should fail because we already garbage collected the data
6709 1 : match tenant
6710 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
6711 1 : .await
6712 1 : {
6713 1 : Ok(_) => panic!("branching should have failed"),
6714 1 : Err(err) => {
6715 1 : let CreateTimelineError::AncestorLsn(err) = err else {
6716 1 : panic!("wrong error type")
6717 1 : };
6718 1 : assert!(err.to_string().contains("invalid branch start lsn"));
6719 1 : assert!(
6720 1 : err.source()
6721 1 : .unwrap()
6722 1 : .to_string()
6723 1 : .contains("we might've already garbage collected needed data")
6724 1 : )
6725 1 : }
6726 1 : }
6727 1 :
6728 1 : Ok(())
6729 1 : }
6730 :
6731 : #[tokio::test]
6732 1 : async fn test_prohibit_branch_creation_on_pre_initdb_lsn() -> anyhow::Result<()> {
6733 1 : let (tenant, ctx) =
6734 1 : TenantHarness::create("test_prohibit_branch_creation_on_pre_initdb_lsn")
6735 1 : .await?
6736 1 : .load()
6737 1 : .await;
6738 :
6739 1 : let tline = tenant
6740 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x50), DEFAULT_PG_VERSION, &ctx)
6741 1 : .await?;
6742 : // try to branch at lsn 0x25, should fail because initdb lsn is 0x50
6743 1 : match tenant
6744 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
6745 1 : .await
6746 1 : {
6747 1 : Ok(_) => panic!("branching should have failed"),
6748 1 : Err(err) => {
6749 1 : let CreateTimelineError::AncestorLsn(err) = err else {
6750 1 : panic!("wrong error type");
6751 1 : };
6752 1 : assert!(&err.to_string().contains("invalid branch start lsn"));
6753 1 : assert!(
6754 1 : &err.source()
6755 1 : .unwrap()
6756 1 : .to_string()
6757 1 : .contains("is earlier than latest GC cutoff")
6758 1 : );
6759 1 : }
6760 1 : }
6761 1 :
6762 1 : Ok(())
6763 1 : }
6764 :
6765 : /*
6766 : // FIXME: This currently fails to error out. Calling GC doesn't currently
6767 : // remove the old value, we'd need to work a little harder
6768 : #[tokio::test]
6769 : async fn test_prohibit_get_for_garbage_collected_data() -> anyhow::Result<()> {
6770 : let repo =
6771 : RepoHarness::create("test_prohibit_get_for_garbage_collected_data")?
6772 : .load();
6773 :
6774 : let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION)?;
6775 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6776 :
6777 : repo.gc_iteration(Some(TIMELINE_ID), 0x10, Duration::ZERO)?;
6778 : let applied_gc_cutoff_lsn = tline.get_applied_gc_cutoff_lsn();
6779 : assert!(*applied_gc_cutoff_lsn > Lsn(0x25));
6780 : match tline.get(*TEST_KEY, Lsn(0x25)) {
6781 : Ok(_) => panic!("request for page should have failed"),
6782 : Err(err) => assert!(err.to_string().contains("not found at")),
6783 : }
6784 : Ok(())
6785 : }
6786 : */
6787 :
6788 : #[tokio::test]
6789 1 : async fn test_get_branchpoints_from_an_inactive_timeline() -> anyhow::Result<()> {
6790 1 : let (tenant, ctx) =
6791 1 : TenantHarness::create("test_get_branchpoints_from_an_inactive_timeline")
6792 1 : .await?
6793 1 : .load()
6794 1 : .await;
6795 1 : let tline = tenant
6796 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6797 1 : .await?;
6798 1 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6799 :
6800 1 : tenant
6801 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6802 1 : .await?;
6803 1 : let newtline = tenant
6804 1 : .get_timeline(NEW_TIMELINE_ID, true)
6805 1 : .expect("Should have a local timeline");
6806 :
6807 1 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6808 :
6809 1 : tline.set_broken("test".to_owned());
6810 :
6811 1 : tenant
6812 1 : .gc_iteration(
6813 1 : Some(TIMELINE_ID),
6814 1 : 0x10,
6815 1 : Duration::ZERO,
6816 1 : &CancellationToken::new(),
6817 1 : &ctx,
6818 1 : )
6819 1 : .await?;
6820 :
6821 : // The branchpoints should contain all timelines, even ones marked
6822 : // as Broken.
6823 : {
6824 1 : let branchpoints = &tline.gc_info.read().unwrap().retain_lsns;
6825 1 : assert_eq!(branchpoints.len(), 1);
6826 1 : assert_eq!(
6827 1 : branchpoints[0],
6828 : (Lsn(0x40), NEW_TIMELINE_ID, MaybeOffloaded::No)
6829 : );
6830 : }
6831 :
6832 : // You can read the key from the child branch even though the parent is
6833 : // Broken, as long as you don't need to access data from the parent.
6834 1 : assert_eq!(
6835 1 : newtline.get(*TEST_KEY, Lsn(0x70), &ctx).await?,
6836 1 : test_img(&format!("foo at {}", Lsn(0x70)))
6837 : );
6838 :
6839 : // This needs to traverse to the parent, and fails.
6840 1 : let err = newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await.unwrap_err();
6841 1 : assert!(
6842 1 : err.to_string().starts_with(&format!(
6843 1 : "bad state on timeline {}: Broken",
6844 1 : tline.timeline_id
6845 1 : )),
6846 0 : "{err}"
6847 : );
6848 :
6849 2 : Ok(())
6850 1 : }
6851 :
6852 : #[tokio::test]
6853 1 : async fn test_retain_data_in_parent_which_is_needed_for_child() -> anyhow::Result<()> {
6854 1 : let (tenant, ctx) =
6855 1 : TenantHarness::create("test_retain_data_in_parent_which_is_needed_for_child")
6856 1 : .await?
6857 1 : .load()
6858 1 : .await;
6859 1 : let tline = tenant
6860 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6861 1 : .await?;
6862 1 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6863 :
6864 1 : tenant
6865 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6866 1 : .await?;
6867 1 : let newtline = tenant
6868 1 : .get_timeline(NEW_TIMELINE_ID, true)
6869 1 : .expect("Should have a local timeline");
6870 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
6871 1 : tenant
6872 1 : .gc_iteration(
6873 1 : Some(TIMELINE_ID),
6874 1 : 0x10,
6875 1 : Duration::ZERO,
6876 1 : &CancellationToken::new(),
6877 1 : &ctx,
6878 1 : )
6879 1 : .await?;
6880 1 : assert!(newtline.get(*TEST_KEY, Lsn(0x25), &ctx).await.is_ok());
6881 :
6882 2 : Ok(())
6883 1 : }
6884 : #[tokio::test]
6885 1 : async fn test_parent_keeps_data_forever_after_branching() -> anyhow::Result<()> {
6886 1 : let (tenant, ctx) = TenantHarness::create("test_parent_keeps_data_forever_after_branching")
6887 1 : .await?
6888 1 : .load()
6889 1 : .await;
6890 1 : let tline = tenant
6891 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6892 1 : .await?;
6893 1 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6894 :
6895 1 : tenant
6896 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6897 1 : .await?;
6898 1 : let newtline = tenant
6899 1 : .get_timeline(NEW_TIMELINE_ID, true)
6900 1 : .expect("Should have a local timeline");
6901 :
6902 1 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6903 :
6904 : // run gc on parent
6905 1 : tenant
6906 1 : .gc_iteration(
6907 1 : Some(TIMELINE_ID),
6908 1 : 0x10,
6909 1 : Duration::ZERO,
6910 1 : &CancellationToken::new(),
6911 1 : &ctx,
6912 1 : )
6913 1 : .await?;
6914 :
6915 : // Check that the data is still accessible on the branch.
6916 1 : assert_eq!(
6917 1 : newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
6918 1 : test_img(&format!("foo at {}", Lsn(0x40)))
6919 : );
6920 :
6921 2 : Ok(())
6922 1 : }
6923 :
6924 : #[tokio::test]
6925 1 : async fn timeline_load() -> anyhow::Result<()> {
6926 : const TEST_NAME: &str = "timeline_load";
6927 1 : let harness = TenantHarness::create(TEST_NAME).await?;
6928 : {
6929 1 : let (tenant, ctx) = harness.load().await;
6930 1 : let tline = tenant
6931 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x7000), DEFAULT_PG_VERSION, &ctx)
6932 1 : .await?;
6933 1 : make_some_layers(tline.as_ref(), Lsn(0x8000), &ctx).await?;
6934 : // so that all uploads finish & we can call harness.load() below again
6935 1 : tenant
6936 1 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
6937 1 : .instrument(harness.span())
6938 1 : .await
6939 1 : .ok()
6940 1 : .unwrap();
6941 : }
6942 :
6943 1 : let (tenant, _ctx) = harness.load().await;
6944 1 : tenant
6945 1 : .get_timeline(TIMELINE_ID, true)
6946 1 : .expect("cannot load timeline");
6947 :
6948 2 : Ok(())
6949 1 : }
6950 :
6951 : #[tokio::test]
6952 1 : async fn timeline_load_with_ancestor() -> anyhow::Result<()> {
6953 : const TEST_NAME: &str = "timeline_load_with_ancestor";
6954 1 : let harness = TenantHarness::create(TEST_NAME).await?;
6955 : // create two timelines
6956 : {
6957 1 : let (tenant, ctx) = harness.load().await;
6958 1 : let tline = tenant
6959 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6960 1 : .await?;
6961 :
6962 1 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6963 :
6964 1 : let child_tline = tenant
6965 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6966 1 : .await?;
6967 1 : child_tline.set_state(TimelineState::Active);
6968 :
6969 1 : let newtline = tenant
6970 1 : .get_timeline(NEW_TIMELINE_ID, true)
6971 1 : .expect("Should have a local timeline");
6972 :
6973 1 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6974 :
6975 : // so that all uploads finish & we can call harness.load() below again
6976 1 : tenant
6977 1 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
6978 1 : .instrument(harness.span())
6979 1 : .await
6980 1 : .ok()
6981 1 : .unwrap();
6982 : }
6983 :
6984 : // check that both of them are initially unloaded
6985 1 : let (tenant, _ctx) = harness.load().await;
6986 :
6987 : // check that both, child and ancestor are loaded
6988 1 : let _child_tline = tenant
6989 1 : .get_timeline(NEW_TIMELINE_ID, true)
6990 1 : .expect("cannot get child timeline loaded");
6991 :
6992 1 : let _ancestor_tline = tenant
6993 1 : .get_timeline(TIMELINE_ID, true)
6994 1 : .expect("cannot get ancestor timeline loaded");
6995 :
6996 2 : Ok(())
6997 1 : }
6998 :
6999 : #[tokio::test]
7000 1 : async fn delta_layer_dumping() -> anyhow::Result<()> {
7001 : use storage_layer::AsLayerDesc;
7002 1 : let (tenant, ctx) = TenantHarness::create("test_layer_dumping")
7003 1 : .await?
7004 1 : .load()
7005 1 : .await;
7006 1 : let tline = tenant
7007 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7008 1 : .await?;
7009 1 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
7010 :
7011 1 : let layer_map = tline.layers.read(LayerManagerLockHolder::Testing).await;
7012 1 : let level0_deltas = layer_map
7013 1 : .layer_map()?
7014 1 : .level0_deltas()
7015 1 : .iter()
7016 2 : .map(|desc| layer_map.get_from_desc(desc))
7017 1 : .collect::<Vec<_>>();
7018 :
7019 1 : assert!(!level0_deltas.is_empty());
7020 :
7021 3 : for delta in level0_deltas {
7022 1 : // Ensure we are dumping a delta layer here
7023 2 : assert!(delta.layer_desc().is_delta);
7024 2 : delta.dump(true, &ctx).await.unwrap();
7025 1 : }
7026 1 :
7027 1 : Ok(())
7028 1 : }
7029 :
7030 : #[tokio::test]
7031 1 : async fn test_images() -> anyhow::Result<()> {
7032 1 : let (tenant, ctx) = TenantHarness::create("test_images").await?.load().await;
7033 1 : let tline = tenant
7034 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
7035 1 : .await?;
7036 :
7037 1 : let mut writer = tline.writer().await;
7038 1 : writer
7039 1 : .put(
7040 1 : *TEST_KEY,
7041 1 : Lsn(0x10),
7042 1 : &Value::Image(test_img("foo at 0x10")),
7043 1 : &ctx,
7044 1 : )
7045 1 : .await?;
7046 1 : writer.finish_write(Lsn(0x10));
7047 1 : drop(writer);
7048 :
7049 1 : tline.freeze_and_flush().await?;
7050 1 : tline
7051 1 : .compact(&CancellationToken::new(), EnumSet::default(), &ctx)
7052 1 : .await?;
7053 :
7054 1 : let mut writer = tline.writer().await;
7055 1 : writer
7056 1 : .put(
7057 1 : *TEST_KEY,
7058 1 : Lsn(0x20),
7059 1 : &Value::Image(test_img("foo at 0x20")),
7060 1 : &ctx,
7061 1 : )
7062 1 : .await?;
7063 1 : writer.finish_write(Lsn(0x20));
7064 1 : drop(writer);
7065 :
7066 1 : tline.freeze_and_flush().await?;
7067 1 : tline
7068 1 : .compact(&CancellationToken::new(), EnumSet::default(), &ctx)
7069 1 : .await?;
7070 :
7071 1 : let mut writer = tline.writer().await;
7072 1 : writer
7073 1 : .put(
7074 1 : *TEST_KEY,
7075 1 : Lsn(0x30),
7076 1 : &Value::Image(test_img("foo at 0x30")),
7077 1 : &ctx,
7078 1 : )
7079 1 : .await?;
7080 1 : writer.finish_write(Lsn(0x30));
7081 1 : drop(writer);
7082 :
7083 1 : tline.freeze_and_flush().await?;
7084 1 : tline
7085 1 : .compact(&CancellationToken::new(), EnumSet::default(), &ctx)
7086 1 : .await?;
7087 :
7088 1 : let mut writer = tline.writer().await;
7089 1 : writer
7090 1 : .put(
7091 1 : *TEST_KEY,
7092 1 : Lsn(0x40),
7093 1 : &Value::Image(test_img("foo at 0x40")),
7094 1 : &ctx,
7095 1 : )
7096 1 : .await?;
7097 1 : writer.finish_write(Lsn(0x40));
7098 1 : drop(writer);
7099 :
7100 1 : tline.freeze_and_flush().await?;
7101 1 : tline
7102 1 : .compact(&CancellationToken::new(), EnumSet::default(), &ctx)
7103 1 : .await?;
7104 :
7105 1 : assert_eq!(
7106 1 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
7107 1 : test_img("foo at 0x10")
7108 : );
7109 1 : assert_eq!(
7110 1 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
7111 1 : test_img("foo at 0x10")
7112 : );
7113 1 : assert_eq!(
7114 1 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
7115 1 : test_img("foo at 0x20")
7116 : );
7117 1 : assert_eq!(
7118 1 : tline.get(*TEST_KEY, Lsn(0x30), &ctx).await?,
7119 1 : test_img("foo at 0x30")
7120 : );
7121 1 : assert_eq!(
7122 1 : tline.get(*TEST_KEY, Lsn(0x40), &ctx).await?,
7123 1 : test_img("foo at 0x40")
7124 : );
7125 :
7126 2 : Ok(())
7127 1 : }
7128 :
7129 2 : async fn bulk_insert_compact_gc(
7130 2 : tenant: &TenantShard,
7131 2 : timeline: &Arc<Timeline>,
7132 2 : ctx: &RequestContext,
7133 2 : lsn: Lsn,
7134 2 : repeat: usize,
7135 2 : key_count: usize,
7136 2 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
7137 2 : let compact = true;
7138 2 : bulk_insert_maybe_compact_gc(tenant, timeline, ctx, lsn, repeat, key_count, compact).await
7139 2 : }
7140 :
7141 4 : async fn bulk_insert_maybe_compact_gc(
7142 4 : tenant: &TenantShard,
7143 4 : timeline: &Arc<Timeline>,
7144 4 : ctx: &RequestContext,
7145 4 : mut lsn: Lsn,
7146 4 : repeat: usize,
7147 4 : key_count: usize,
7148 4 : compact: bool,
7149 4 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
7150 4 : let mut inserted: HashMap<Key, BTreeSet<Lsn>> = Default::default();
7151 :
7152 4 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7153 4 : let mut blknum = 0;
7154 :
7155 : // Enforce that key range is monotonously increasing
7156 4 : let mut keyspace = KeySpaceAccum::new();
7157 :
7158 4 : let cancel = CancellationToken::new();
7159 :
7160 4 : for _ in 0..repeat {
7161 200 : for _ in 0..key_count {
7162 2000000 : test_key.field6 = blknum;
7163 2000000 : let mut writer = timeline.writer().await;
7164 2000000 : writer
7165 2000000 : .put(
7166 2000000 : test_key,
7167 2000000 : lsn,
7168 2000000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
7169 2000000 : ctx,
7170 2000000 : )
7171 2000000 : .await?;
7172 2000000 : inserted.entry(test_key).or_default().insert(lsn);
7173 2000000 : writer.finish_write(lsn);
7174 2000000 : drop(writer);
7175 :
7176 2000000 : keyspace.add_key(test_key);
7177 :
7178 2000000 : lsn = Lsn(lsn.0 + 0x10);
7179 2000000 : blknum += 1;
7180 : }
7181 :
7182 200 : timeline.freeze_and_flush().await?;
7183 200 : if compact {
7184 : // this requires timeline to be &Arc<Timeline>
7185 100 : timeline.compact(&cancel, EnumSet::default(), ctx).await?;
7186 100 : }
7187 :
7188 : // this doesn't really need to use the timeline_id target, but it is closer to what it
7189 : // originally was.
7190 200 : let res = tenant
7191 200 : .gc_iteration(Some(timeline.timeline_id), 0, Duration::ZERO, &cancel, ctx)
7192 200 : .await?;
7193 :
7194 200 : assert_eq!(res.layers_removed, 0, "this never removes anything");
7195 : }
7196 :
7197 4 : Ok(inserted)
7198 4 : }
7199 :
7200 : //
7201 : // Insert 1000 key-value pairs with increasing keys, flush, compact, GC.
7202 : // Repeat 50 times.
7203 : //
7204 : #[tokio::test]
7205 1 : async fn test_bulk_insert() -> anyhow::Result<()> {
7206 1 : let harness = TenantHarness::create("test_bulk_insert").await?;
7207 1 : let (tenant, ctx) = harness.load().await;
7208 1 : let tline = tenant
7209 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
7210 1 : .await?;
7211 :
7212 1 : let lsn = Lsn(0x10);
7213 1 : bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
7214 :
7215 2 : Ok(())
7216 1 : }
7217 :
7218 : // Test the vectored get real implementation against a simple sequential implementation.
7219 : //
7220 : // The test generates a keyspace by repeatedly flushing the in-memory layer and compacting.
7221 : // Projected to 2D the key space looks like below. Lsn grows upwards on the Y axis and keys
7222 : // grow to the right on the X axis.
7223 : // [Delta]
7224 : // [Delta]
7225 : // [Delta]
7226 : // [Delta]
7227 : // ------------ Image ---------------
7228 : //
7229 : // After layer generation we pick the ranges to query as follows:
7230 : // 1. The beginning of each delta layer
7231 : // 2. At the seam between two adjacent delta layers
7232 : //
7233 : // There's one major downside to this test: delta layers only contains images,
7234 : // so the search can stop at the first delta layer and doesn't traverse any deeper.
7235 : #[tokio::test]
7236 1 : async fn test_get_vectored() -> anyhow::Result<()> {
7237 1 : let harness = TenantHarness::create("test_get_vectored").await?;
7238 1 : let (tenant, ctx) = harness.load().await;
7239 1 : let io_concurrency = IoConcurrency::spawn_for_test();
7240 1 : let tline = tenant
7241 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
7242 1 : .await?;
7243 :
7244 1 : let lsn = Lsn(0x10);
7245 1 : let inserted = bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
7246 :
7247 1 : let guard = tline.layers.read(LayerManagerLockHolder::Testing).await;
7248 1 : let lm = guard.layer_map()?;
7249 :
7250 1 : lm.dump(true, &ctx).await?;
7251 :
7252 1 : let mut reads = Vec::new();
7253 1 : let mut prev = None;
7254 6 : lm.iter_historic_layers().for_each(|desc| {
7255 6 : if !desc.is_delta() {
7256 1 : prev = Some(desc.clone());
7257 1 : return;
7258 5 : }
7259 :
7260 5 : let start = desc.key_range.start;
7261 5 : let end = desc
7262 5 : .key_range
7263 5 : .start
7264 5 : .add(tenant.conf.max_get_vectored_keys.get() as u32);
7265 5 : reads.push(KeySpace {
7266 5 : ranges: vec![start..end],
7267 5 : });
7268 :
7269 5 : if let Some(prev) = &prev {
7270 5 : if !prev.is_delta() {
7271 5 : return;
7272 0 : }
7273 :
7274 0 : let first_range = Key {
7275 0 : field6: prev.key_range.end.field6 - 4,
7276 0 : ..prev.key_range.end
7277 0 : }..prev.key_range.end;
7278 :
7279 0 : let second_range = desc.key_range.start..Key {
7280 0 : field6: desc.key_range.start.field6 + 4,
7281 0 : ..desc.key_range.start
7282 0 : };
7283 :
7284 0 : reads.push(KeySpace {
7285 0 : ranges: vec![first_range, second_range],
7286 0 : });
7287 0 : };
7288 :
7289 0 : prev = Some(desc.clone());
7290 6 : });
7291 :
7292 1 : drop(guard);
7293 :
7294 : // Pick a big LSN such that we query over all the changes.
7295 1 : let reads_lsn = Lsn(u64::MAX - 1);
7296 :
7297 6 : for read in reads {
7298 5 : info!("Doing vectored read on {:?}", read);
7299 1 :
7300 5 : let query = VersionedKeySpaceQuery::uniform(read.clone(), reads_lsn);
7301 1 :
7302 5 : let vectored_res = tline
7303 5 : .get_vectored_impl(
7304 5 : query,
7305 5 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7306 5 : &ctx,
7307 5 : )
7308 5 : .await;
7309 1 :
7310 5 : let mut expected_lsns: HashMap<Key, Lsn> = Default::default();
7311 5 : let mut expect_missing = false;
7312 5 : let mut key = read.start().unwrap();
7313 165 : while key != read.end().unwrap() {
7314 160 : if let Some(lsns) = inserted.get(&key) {
7315 160 : let expected_lsn = lsns.iter().rfind(|lsn| **lsn <= reads_lsn);
7316 160 : match expected_lsn {
7317 160 : Some(lsn) => {
7318 160 : expected_lsns.insert(key, *lsn);
7319 160 : }
7320 1 : None => {
7321 1 : expect_missing = true;
7322 1 : break;
7323 1 : }
7324 1 : }
7325 1 : } else {
7326 1 : expect_missing = true;
7327 1 : break;
7328 1 : }
7329 1 :
7330 160 : key = key.next();
7331 1 : }
7332 1 :
7333 5 : if expect_missing {
7334 1 : assert!(matches!(vectored_res, Err(GetVectoredError::MissingKey(_))));
7335 1 : } else {
7336 160 : for (key, image) in vectored_res? {
7337 160 : let expected_lsn = expected_lsns.get(&key).expect("determined above");
7338 160 : let expected_image = test_img(&format!("{} at {}", key.field6, expected_lsn));
7339 160 : assert_eq!(image?, expected_image);
7340 1 : }
7341 1 : }
7342 1 : }
7343 1 :
7344 1 : Ok(())
7345 1 : }
7346 :
7347 : #[tokio::test]
7348 1 : async fn test_get_vectored_aux_files() -> anyhow::Result<()> {
7349 1 : let harness = TenantHarness::create("test_get_vectored_aux_files").await?;
7350 :
7351 1 : let (tenant, ctx) = harness.load().await;
7352 1 : let io_concurrency = IoConcurrency::spawn_for_test();
7353 1 : let (tline, ctx) = tenant
7354 1 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
7355 1 : .await?;
7356 1 : let tline = tline.raw_timeline().unwrap();
7357 :
7358 1 : let mut modification = tline.begin_modification(Lsn(0x1000));
7359 1 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
7360 1 : modification.set_lsn(Lsn(0x1008))?;
7361 1 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
7362 1 : modification.commit(&ctx).await?;
7363 :
7364 1 : let child_timeline_id = TimelineId::generate();
7365 1 : tenant
7366 1 : .branch_timeline_test(
7367 1 : tline,
7368 1 : child_timeline_id,
7369 1 : Some(tline.get_last_record_lsn()),
7370 1 : &ctx,
7371 1 : )
7372 1 : .await?;
7373 :
7374 1 : let child_timeline = tenant
7375 1 : .get_timeline(child_timeline_id, true)
7376 1 : .expect("Should have the branched timeline");
7377 :
7378 1 : let aux_keyspace = KeySpace {
7379 1 : ranges: vec![NON_INHERITED_RANGE],
7380 1 : };
7381 1 : let read_lsn = child_timeline.get_last_record_lsn();
7382 :
7383 1 : let query = VersionedKeySpaceQuery::uniform(aux_keyspace.clone(), read_lsn);
7384 :
7385 1 : let vectored_res = child_timeline
7386 1 : .get_vectored_impl(
7387 1 : query,
7388 1 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7389 1 : &ctx,
7390 1 : )
7391 1 : .await;
7392 :
7393 1 : let images = vectored_res?;
7394 1 : assert!(images.is_empty());
7395 2 : Ok(())
7396 1 : }
7397 :
7398 : // Test that vectored get handles layer gaps correctly
7399 : // by advancing into the next ancestor timeline if required.
7400 : //
7401 : // The test generates timelines that look like the diagram below.
7402 : // We leave a gap in one of the L1 layers at `gap_at_key` (`/` in the diagram).
7403 : // The reconstruct data for that key lies in the ancestor timeline (`X` in the diagram).
7404 : //
7405 : // ```
7406 : //-------------------------------+
7407 : // ... |
7408 : // [ L1 ] |
7409 : // [ / L1 ] | Child Timeline
7410 : // ... |
7411 : // ------------------------------+
7412 : // [ X L1 ] | Parent Timeline
7413 : // ------------------------------+
7414 : // ```
7415 : #[tokio::test]
7416 1 : async fn test_get_vectored_key_gap() -> anyhow::Result<()> {
7417 1 : let tenant_conf = pageserver_api::models::TenantConfig {
7418 1 : // Make compaction deterministic
7419 1 : gc_period: Some(Duration::ZERO),
7420 1 : compaction_period: Some(Duration::ZERO),
7421 1 : // Encourage creation of L1 layers
7422 1 : checkpoint_distance: Some(16 * 1024),
7423 1 : compaction_target_size: Some(8 * 1024),
7424 1 : ..Default::default()
7425 1 : };
7426 :
7427 1 : let harness = TenantHarness::create_custom(
7428 1 : "test_get_vectored_key_gap",
7429 1 : tenant_conf,
7430 1 : TenantId::generate(),
7431 1 : ShardIdentity::unsharded(),
7432 1 : Generation::new(0xdeadbeef),
7433 1 : )
7434 1 : .await?;
7435 1 : let (tenant, ctx) = harness.load().await;
7436 1 : let io_concurrency = IoConcurrency::spawn_for_test();
7437 :
7438 1 : let mut current_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7439 1 : let gap_at_key = current_key.add(100);
7440 1 : let mut current_lsn = Lsn(0x10);
7441 :
7442 : const KEY_COUNT: usize = 10_000;
7443 :
7444 1 : let timeline_id = TimelineId::generate();
7445 1 : let current_timeline = tenant
7446 1 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
7447 1 : .await?;
7448 :
7449 1 : current_lsn += 0x100;
7450 :
7451 1 : let mut writer = current_timeline.writer().await;
7452 1 : writer
7453 1 : .put(
7454 1 : gap_at_key,
7455 1 : current_lsn,
7456 1 : &Value::Image(test_img(&format!("{gap_at_key} at {current_lsn}"))),
7457 1 : &ctx,
7458 1 : )
7459 1 : .await?;
7460 1 : writer.finish_write(current_lsn);
7461 1 : drop(writer);
7462 :
7463 1 : let mut latest_lsns = HashMap::new();
7464 1 : latest_lsns.insert(gap_at_key, current_lsn);
7465 :
7466 1 : current_timeline.freeze_and_flush().await?;
7467 :
7468 1 : let child_timeline_id = TimelineId::generate();
7469 :
7470 1 : tenant
7471 1 : .branch_timeline_test(
7472 1 : ¤t_timeline,
7473 1 : child_timeline_id,
7474 1 : Some(current_lsn),
7475 1 : &ctx,
7476 1 : )
7477 1 : .await?;
7478 1 : let child_timeline = tenant
7479 1 : .get_timeline(child_timeline_id, true)
7480 1 : .expect("Should have the branched timeline");
7481 :
7482 10001 : for i in 0..KEY_COUNT {
7483 10000 : if current_key == gap_at_key {
7484 1 : current_key = current_key.next();
7485 1 : continue;
7486 9999 : }
7487 :
7488 9999 : current_lsn += 0x10;
7489 :
7490 9999 : let mut writer = child_timeline.writer().await;
7491 9999 : writer
7492 9999 : .put(
7493 9999 : current_key,
7494 9999 : current_lsn,
7495 9999 : &Value::Image(test_img(&format!("{current_key} at {current_lsn}"))),
7496 9999 : &ctx,
7497 9999 : )
7498 9999 : .await?;
7499 9999 : writer.finish_write(current_lsn);
7500 9999 : drop(writer);
7501 :
7502 9999 : latest_lsns.insert(current_key, current_lsn);
7503 9999 : current_key = current_key.next();
7504 :
7505 : // Flush every now and then to encourage layer file creation.
7506 9999 : if i % 500 == 0 {
7507 20 : child_timeline.freeze_and_flush().await?;
7508 9979 : }
7509 : }
7510 :
7511 1 : child_timeline.freeze_and_flush().await?;
7512 1 : let mut flags = EnumSet::new();
7513 1 : flags.insert(CompactFlags::ForceRepartition);
7514 1 : child_timeline
7515 1 : .compact(&CancellationToken::new(), flags, &ctx)
7516 1 : .await?;
7517 :
7518 1 : let key_near_end = {
7519 1 : let mut tmp = current_key;
7520 1 : tmp.field6 -= 10;
7521 1 : tmp
7522 : };
7523 :
7524 1 : let key_near_gap = {
7525 1 : let mut tmp = gap_at_key;
7526 1 : tmp.field6 -= 10;
7527 1 : tmp
7528 : };
7529 :
7530 1 : let read = KeySpace {
7531 1 : ranges: vec![key_near_gap..gap_at_key.next(), key_near_end..current_key],
7532 1 : };
7533 :
7534 1 : let query = VersionedKeySpaceQuery::uniform(read.clone(), current_lsn);
7535 :
7536 1 : let results = child_timeline
7537 1 : .get_vectored_impl(
7538 1 : query,
7539 1 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7540 1 : &ctx,
7541 1 : )
7542 1 : .await?;
7543 :
7544 22 : for (key, img_res) in results {
7545 21 : let expected = test_img(&format!("{} at {}", key, latest_lsns[&key]));
7546 21 : assert_eq!(img_res?, expected);
7547 1 : }
7548 1 :
7549 1 : Ok(())
7550 1 : }
7551 :
7552 : // Test that vectored get descends into ancestor timelines correctly and
7553 : // does not return an image that's newer than requested.
7554 : //
7555 : // The diagram below ilustrates an interesting case. We have a parent timeline
7556 : // (top of the Lsn range) and a child timeline. The request key cannot be reconstructed
7557 : // from the child timeline, so the parent timeline must be visited. When advacing into
7558 : // the child timeline, the read path needs to remember what the requested Lsn was in
7559 : // order to avoid returning an image that's too new. The test below constructs such
7560 : // a timeline setup and does a few queries around the Lsn of each page image.
7561 : // ```
7562 : // LSN
7563 : // ^
7564 : // |
7565 : // |
7566 : // 500 | --------------------------------------> branch point
7567 : // 400 | X
7568 : // 300 | X
7569 : // 200 | --------------------------------------> requested lsn
7570 : // 100 | X
7571 : // |---------------------------------------> Key
7572 : // |
7573 : // ------> requested key
7574 : //
7575 : // Legend:
7576 : // * X - page images
7577 : // ```
7578 : #[tokio::test]
7579 1 : async fn test_get_vectored_ancestor_descent() -> anyhow::Result<()> {
7580 1 : let harness = TenantHarness::create("test_get_vectored_on_lsn_axis").await?;
7581 1 : let (tenant, ctx) = harness.load().await;
7582 1 : let io_concurrency = IoConcurrency::spawn_for_test();
7583 :
7584 1 : let start_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7585 1 : let end_key = start_key.add(1000);
7586 1 : let child_gap_at_key = start_key.add(500);
7587 1 : let mut parent_gap_lsns: BTreeMap<Lsn, String> = BTreeMap::new();
7588 :
7589 1 : let mut current_lsn = Lsn(0x10);
7590 :
7591 1 : let timeline_id = TimelineId::generate();
7592 1 : let parent_timeline = tenant
7593 1 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
7594 1 : .await?;
7595 :
7596 1 : current_lsn += 0x100;
7597 :
7598 4 : for _ in 0..3 {
7599 3 : let mut key = start_key;
7600 3003 : while key < end_key {
7601 3000 : current_lsn += 0x10;
7602 :
7603 3000 : let image_value = format!("{child_gap_at_key} at {current_lsn}");
7604 :
7605 3000 : let mut writer = parent_timeline.writer().await;
7606 3000 : writer
7607 3000 : .put(
7608 3000 : key,
7609 3000 : current_lsn,
7610 3000 : &Value::Image(test_img(&image_value)),
7611 3000 : &ctx,
7612 3000 : )
7613 3000 : .await?;
7614 3000 : writer.finish_write(current_lsn);
7615 :
7616 3000 : if key == child_gap_at_key {
7617 3 : parent_gap_lsns.insert(current_lsn, image_value);
7618 2997 : }
7619 :
7620 3000 : key = key.next();
7621 : }
7622 :
7623 3 : parent_timeline.freeze_and_flush().await?;
7624 : }
7625 :
7626 1 : let child_timeline_id = TimelineId::generate();
7627 :
7628 1 : let child_timeline = tenant
7629 1 : .branch_timeline_test(&parent_timeline, child_timeline_id, Some(current_lsn), &ctx)
7630 1 : .await?;
7631 :
7632 1 : let mut key = start_key;
7633 1001 : while key < end_key {
7634 1000 : if key == child_gap_at_key {
7635 1 : key = key.next();
7636 1 : continue;
7637 999 : }
7638 :
7639 999 : current_lsn += 0x10;
7640 :
7641 999 : let mut writer = child_timeline.writer().await;
7642 999 : writer
7643 999 : .put(
7644 999 : key,
7645 999 : current_lsn,
7646 999 : &Value::Image(test_img(&format!("{key} at {current_lsn}"))),
7647 999 : &ctx,
7648 999 : )
7649 999 : .await?;
7650 999 : writer.finish_write(current_lsn);
7651 :
7652 999 : key = key.next();
7653 : }
7654 :
7655 1 : child_timeline.freeze_and_flush().await?;
7656 :
7657 1 : let lsn_offsets: [i64; 5] = [-10, -1, 0, 1, 10];
7658 1 : let mut query_lsns = Vec::new();
7659 3 : for image_lsn in parent_gap_lsns.keys().rev() {
7660 18 : for offset in lsn_offsets {
7661 15 : query_lsns.push(Lsn(image_lsn
7662 15 : .0
7663 15 : .checked_add_signed(offset)
7664 15 : .expect("Shouldn't overflow")));
7665 15 : }
7666 1 : }
7667 1 :
7668 16 : for query_lsn in query_lsns {
7669 15 : let query = VersionedKeySpaceQuery::uniform(
7670 15 : KeySpace {
7671 15 : ranges: vec![child_gap_at_key..child_gap_at_key.next()],
7672 15 : },
7673 15 : query_lsn,
7674 1 : );
7675 1 :
7676 15 : let results = child_timeline
7677 15 : .get_vectored_impl(
7678 15 : query,
7679 15 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7680 15 : &ctx,
7681 15 : )
7682 15 : .await;
7683 1 :
7684 15 : let expected_item = parent_gap_lsns
7685 15 : .iter()
7686 15 : .rev()
7687 34 : .find(|(lsn, _)| **lsn <= query_lsn);
7688 1 :
7689 15 : info!(
7690 1 : "Doing vectored read at LSN {}. Expecting image to be: {:?}",
7691 1 : query_lsn, expected_item
7692 1 : );
7693 1 :
7694 15 : match expected_item {
7695 13 : Some((_, img_value)) => {
7696 13 : let key_results = results.expect("No vectored get error expected");
7697 13 : let key_result = &key_results[&child_gap_at_key];
7698 13 : let returned_img = key_result
7699 13 : .as_ref()
7700 13 : .expect("No page reconstruct error expected");
7701 1 :
7702 13 : info!(
7703 1 : "Vectored read at LSN {} returned image {}",
7704 1 : query_lsn,
7705 1 : std::str::from_utf8(returned_img)?
7706 1 : );
7707 13 : assert_eq!(*returned_img, test_img(img_value));
7708 1 : }
7709 1 : None => {
7710 2 : assert!(matches!(results, Err(GetVectoredError::MissingKey(_))));
7711 1 : }
7712 1 : }
7713 1 : }
7714 1 :
7715 1 : Ok(())
7716 1 : }
7717 :
7718 : #[tokio::test]
7719 1 : async fn test_random_updates() -> anyhow::Result<()> {
7720 1 : let names_algorithms = [
7721 1 : ("test_random_updates_legacy", CompactionAlgorithm::Legacy),
7722 1 : ("test_random_updates_tiered", CompactionAlgorithm::Tiered),
7723 1 : ];
7724 3 : for (name, algorithm) in names_algorithms {
7725 2 : test_random_updates_algorithm(name, algorithm).await?;
7726 1 : }
7727 1 : Ok(())
7728 1 : }
7729 :
7730 2 : async fn test_random_updates_algorithm(
7731 2 : name: &'static str,
7732 2 : compaction_algorithm: CompactionAlgorithm,
7733 2 : ) -> anyhow::Result<()> {
7734 2 : let mut harness = TenantHarness::create(name).await?;
7735 2 : harness.tenant_conf.compaction_algorithm = Some(CompactionAlgorithmSettings {
7736 2 : kind: compaction_algorithm,
7737 2 : });
7738 2 : let (tenant, ctx) = harness.load().await;
7739 2 : let tline = tenant
7740 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7741 2 : .await?;
7742 :
7743 : const NUM_KEYS: usize = 1000;
7744 2 : let cancel = CancellationToken::new();
7745 :
7746 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7747 2 : let mut test_key_end = test_key;
7748 2 : test_key_end.field6 = NUM_KEYS as u32;
7749 2 : tline.add_extra_test_dense_keyspace(KeySpace::single(test_key..test_key_end));
7750 :
7751 2 : let mut keyspace = KeySpaceAccum::new();
7752 :
7753 : // Track when each page was last modified. Used to assert that
7754 : // a read sees the latest page version.
7755 2 : let mut updated = [Lsn(0); NUM_KEYS];
7756 :
7757 2 : let mut lsn = Lsn(0x10);
7758 : #[allow(clippy::needless_range_loop)]
7759 2002 : for blknum in 0..NUM_KEYS {
7760 2000 : lsn = Lsn(lsn.0 + 0x10);
7761 2000 : test_key.field6 = blknum as u32;
7762 2000 : let mut writer = tline.writer().await;
7763 2000 : writer
7764 2000 : .put(
7765 2000 : test_key,
7766 2000 : lsn,
7767 2000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
7768 2000 : &ctx,
7769 2000 : )
7770 2000 : .await?;
7771 2000 : writer.finish_write(lsn);
7772 2000 : updated[blknum] = lsn;
7773 2000 : drop(writer);
7774 :
7775 2000 : keyspace.add_key(test_key);
7776 : }
7777 :
7778 102 : for _ in 0..50 {
7779 100100 : for _ in 0..NUM_KEYS {
7780 100000 : lsn = Lsn(lsn.0 + 0x10);
7781 100000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7782 100000 : test_key.field6 = blknum as u32;
7783 100000 : let mut writer = tline.writer().await;
7784 100000 : writer
7785 100000 : .put(
7786 100000 : test_key,
7787 100000 : lsn,
7788 100000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
7789 100000 : &ctx,
7790 100000 : )
7791 100000 : .await?;
7792 100000 : writer.finish_write(lsn);
7793 100000 : drop(writer);
7794 100000 : updated[blknum] = lsn;
7795 : }
7796 :
7797 : // Read all the blocks
7798 100000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7799 100000 : test_key.field6 = blknum as u32;
7800 100000 : assert_eq!(
7801 100000 : tline.get(test_key, lsn, &ctx).await?,
7802 100000 : test_img(&format!("{blknum} at {last_lsn}"))
7803 : );
7804 : }
7805 :
7806 : // Perform a cycle of flush, and GC
7807 100 : tline.freeze_and_flush().await?;
7808 100 : tenant
7809 100 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7810 100 : .await?;
7811 : }
7812 :
7813 2 : Ok(())
7814 2 : }
7815 :
7816 : #[tokio::test]
7817 1 : async fn test_traverse_branches() -> anyhow::Result<()> {
7818 1 : let (tenant, ctx) = TenantHarness::create("test_traverse_branches")
7819 1 : .await?
7820 1 : .load()
7821 1 : .await;
7822 1 : let mut tline = tenant
7823 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7824 1 : .await?;
7825 :
7826 : const NUM_KEYS: usize = 1000;
7827 :
7828 1 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7829 :
7830 1 : let mut keyspace = KeySpaceAccum::new();
7831 :
7832 1 : let cancel = CancellationToken::new();
7833 :
7834 : // Track when each page was last modified. Used to assert that
7835 : // a read sees the latest page version.
7836 1 : let mut updated = [Lsn(0); NUM_KEYS];
7837 :
7838 1 : let mut lsn = Lsn(0x10);
7839 1 : #[allow(clippy::needless_range_loop)]
7840 1001 : for blknum in 0..NUM_KEYS {
7841 1000 : lsn = Lsn(lsn.0 + 0x10);
7842 1000 : test_key.field6 = blknum as u32;
7843 1000 : let mut writer = tline.writer().await;
7844 1000 : writer
7845 1000 : .put(
7846 1000 : test_key,
7847 1000 : lsn,
7848 1000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
7849 1000 : &ctx,
7850 1000 : )
7851 1000 : .await?;
7852 1000 : writer.finish_write(lsn);
7853 1000 : updated[blknum] = lsn;
7854 1000 : drop(writer);
7855 1 :
7856 1000 : keyspace.add_key(test_key);
7857 1 : }
7858 1 :
7859 51 : for _ in 0..50 {
7860 50 : let new_tline_id = TimelineId::generate();
7861 50 : tenant
7862 50 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7863 50 : .await?;
7864 50 : tline = tenant
7865 50 : .get_timeline(new_tline_id, true)
7866 50 : .expect("Should have the branched timeline");
7867 1 :
7868 50050 : for _ in 0..NUM_KEYS {
7869 50000 : lsn = Lsn(lsn.0 + 0x10);
7870 50000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7871 50000 : test_key.field6 = blknum as u32;
7872 50000 : let mut writer = tline.writer().await;
7873 50000 : writer
7874 50000 : .put(
7875 50000 : test_key,
7876 50000 : lsn,
7877 50000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
7878 50000 : &ctx,
7879 50000 : )
7880 50000 : .await?;
7881 50000 : println!("updating {blknum} at {lsn}");
7882 50000 : writer.finish_write(lsn);
7883 50000 : drop(writer);
7884 50000 : updated[blknum] = lsn;
7885 1 : }
7886 1 :
7887 1 : // Read all the blocks
7888 50000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7889 50000 : test_key.field6 = blknum as u32;
7890 50000 : assert_eq!(
7891 50000 : tline.get(test_key, lsn, &ctx).await?,
7892 50000 : test_img(&format!("{blknum} at {last_lsn}"))
7893 1 : );
7894 1 : }
7895 1 :
7896 1 : // Perform a cycle of flush, compact, and GC
7897 50 : tline.freeze_and_flush().await?;
7898 50 : tline.compact(&cancel, EnumSet::default(), &ctx).await?;
7899 50 : tenant
7900 50 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7901 50 : .await?;
7902 1 : }
7903 1 :
7904 1 : Ok(())
7905 1 : }
7906 :
7907 : #[tokio::test]
7908 1 : async fn test_traverse_ancestors() -> anyhow::Result<()> {
7909 1 : let (tenant, ctx) = TenantHarness::create("test_traverse_ancestors")
7910 1 : .await?
7911 1 : .load()
7912 1 : .await;
7913 1 : let mut tline = tenant
7914 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7915 1 : .await?;
7916 :
7917 : const NUM_KEYS: usize = 100;
7918 : const NUM_TLINES: usize = 50;
7919 :
7920 1 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7921 : // Track page mutation lsns across different timelines.
7922 1 : let mut updated = [[Lsn(0); NUM_KEYS]; NUM_TLINES];
7923 :
7924 1 : let mut lsn = Lsn(0x10);
7925 :
7926 1 : #[allow(clippy::needless_range_loop)]
7927 51 : for idx in 0..NUM_TLINES {
7928 50 : let new_tline_id = TimelineId::generate();
7929 50 : tenant
7930 50 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7931 50 : .await?;
7932 50 : tline = tenant
7933 50 : .get_timeline(new_tline_id, true)
7934 50 : .expect("Should have the branched timeline");
7935 1 :
7936 5050 : for _ in 0..NUM_KEYS {
7937 5000 : lsn = Lsn(lsn.0 + 0x10);
7938 5000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7939 5000 : test_key.field6 = blknum as u32;
7940 5000 : let mut writer = tline.writer().await;
7941 5000 : writer
7942 5000 : .put(
7943 5000 : test_key,
7944 5000 : lsn,
7945 5000 : &Value::Image(test_img(&format!("{idx} {blknum} at {lsn}"))),
7946 5000 : &ctx,
7947 5000 : )
7948 5000 : .await?;
7949 5000 : println!("updating [{idx}][{blknum}] at {lsn}");
7950 5000 : writer.finish_write(lsn);
7951 5000 : drop(writer);
7952 5000 : updated[idx][blknum] = lsn;
7953 1 : }
7954 1 : }
7955 1 :
7956 1 : // Read pages from leaf timeline across all ancestors.
7957 50 : for (idx, lsns) in updated.iter().enumerate() {
7958 5000 : for (blknum, lsn) in lsns.iter().enumerate() {
7959 1 : // Skip empty mutations.
7960 5000 : if lsn.0 == 0 {
7961 1877 : continue;
7962 3123 : }
7963 3123 : println!("checking [{idx}][{blknum}] at {lsn}");
7964 3123 : test_key.field6 = blknum as u32;
7965 3123 : assert_eq!(
7966 3123 : tline.get(test_key, *lsn, &ctx).await?,
7967 3123 : test_img(&format!("{idx} {blknum} at {lsn}"))
7968 1 : );
7969 1 : }
7970 1 : }
7971 1 : Ok(())
7972 1 : }
7973 :
7974 : #[tokio::test]
7975 1 : async fn test_write_at_initdb_lsn_takes_optimization_code_path() -> anyhow::Result<()> {
7976 1 : let (tenant, ctx) = TenantHarness::create("test_empty_test_timeline_is_usable")
7977 1 : .await?
7978 1 : .load()
7979 1 : .await;
7980 :
7981 1 : let initdb_lsn = Lsn(0x20);
7982 1 : let (utline, ctx) = tenant
7983 1 : .create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)
7984 1 : .await?;
7985 1 : let tline = utline.raw_timeline().unwrap();
7986 :
7987 : // Spawn flush loop now so that we can set the `expect_initdb_optimization`
7988 1 : tline.maybe_spawn_flush_loop();
7989 :
7990 : // Make sure the timeline has the minimum set of required keys for operation.
7991 : // The only operation you can always do on an empty timeline is to `put` new data.
7992 : // Except if you `put` at `initdb_lsn`.
7993 : // In that case, there's an optimization to directly create image layers instead of delta layers.
7994 : // It uses `repartition()`, which assumes some keys to be present.
7995 : // Let's make sure the test timeline can handle that case.
7996 : {
7997 1 : let mut state = tline.flush_loop_state.lock().unwrap();
7998 1 : assert_eq!(
7999 : timeline::FlushLoopState::Running {
8000 : expect_initdb_optimization: false,
8001 : initdb_optimization_count: 0,
8002 : },
8003 1 : *state
8004 : );
8005 1 : *state = timeline::FlushLoopState::Running {
8006 1 : expect_initdb_optimization: true,
8007 1 : initdb_optimization_count: 0,
8008 1 : };
8009 : }
8010 :
8011 : // Make writes at the initdb_lsn. When we flush it below, it should be handled by the optimization.
8012 : // As explained above, the optimization requires some keys to be present.
8013 : // As per `create_empty_timeline` documentation, use init_empty to set them.
8014 : // This is what `create_test_timeline` does, by the way.
8015 1 : let mut modification = tline.begin_modification(initdb_lsn);
8016 1 : modification
8017 1 : .init_empty_test_timeline()
8018 1 : .context("init_empty_test_timeline")?;
8019 1 : modification
8020 1 : .commit(&ctx)
8021 1 : .await
8022 1 : .context("commit init_empty_test_timeline modification")?;
8023 :
8024 : // Do the flush. The flush code will check the expectations that we set above.
8025 1 : tline.freeze_and_flush().await?;
8026 :
8027 : // assert freeze_and_flush exercised the initdb optimization
8028 1 : {
8029 1 : let state = tline.flush_loop_state.lock().unwrap();
8030 1 : let timeline::FlushLoopState::Running {
8031 1 : expect_initdb_optimization,
8032 1 : initdb_optimization_count,
8033 1 : } = *state
8034 1 : else {
8035 1 : panic!("unexpected state: {:?}", *state);
8036 1 : };
8037 1 : assert!(expect_initdb_optimization);
8038 1 : assert!(initdb_optimization_count > 0);
8039 1 : }
8040 1 : Ok(())
8041 1 : }
8042 :
8043 : #[tokio::test]
8044 1 : async fn test_create_guard_crash() -> anyhow::Result<()> {
8045 1 : let name = "test_create_guard_crash";
8046 1 : let harness = TenantHarness::create(name).await?;
8047 : {
8048 1 : let (tenant, ctx) = harness.load().await;
8049 1 : let (tline, _ctx) = tenant
8050 1 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
8051 1 : .await?;
8052 : // Leave the timeline ID in [`TenantShard::timelines_creating`] to exclude attempting to create it again
8053 1 : let raw_tline = tline.raw_timeline().unwrap();
8054 1 : raw_tline
8055 1 : .shutdown(super::timeline::ShutdownMode::Hard)
8056 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))
8057 1 : .await;
8058 1 : std::mem::forget(tline);
8059 : }
8060 :
8061 1 : let (tenant, _) = harness.load().await;
8062 1 : match tenant.get_timeline(TIMELINE_ID, false) {
8063 0 : Ok(_) => panic!("timeline should've been removed during load"),
8064 1 : Err(e) => {
8065 1 : assert_eq!(
8066 : e,
8067 1 : GetTimelineError::NotFound {
8068 1 : tenant_id: tenant.tenant_shard_id,
8069 1 : timeline_id: TIMELINE_ID,
8070 1 : }
8071 : )
8072 : }
8073 : }
8074 :
8075 1 : assert!(
8076 1 : !harness
8077 1 : .conf
8078 1 : .timeline_path(&tenant.tenant_shard_id, &TIMELINE_ID)
8079 1 : .exists()
8080 : );
8081 :
8082 2 : Ok(())
8083 1 : }
8084 :
8085 : #[tokio::test]
8086 1 : async fn test_read_at_max_lsn() -> anyhow::Result<()> {
8087 1 : let names_algorithms = [
8088 1 : ("test_read_at_max_lsn_legacy", CompactionAlgorithm::Legacy),
8089 1 : ("test_read_at_max_lsn_tiered", CompactionAlgorithm::Tiered),
8090 1 : ];
8091 3 : for (name, algorithm) in names_algorithms {
8092 2 : test_read_at_max_lsn_algorithm(name, algorithm).await?;
8093 1 : }
8094 1 : Ok(())
8095 1 : }
8096 :
8097 2 : async fn test_read_at_max_lsn_algorithm(
8098 2 : name: &'static str,
8099 2 : compaction_algorithm: CompactionAlgorithm,
8100 2 : ) -> anyhow::Result<()> {
8101 2 : let mut harness = TenantHarness::create(name).await?;
8102 2 : harness.tenant_conf.compaction_algorithm = Some(CompactionAlgorithmSettings {
8103 2 : kind: compaction_algorithm,
8104 2 : });
8105 2 : let (tenant, ctx) = harness.load().await;
8106 2 : let tline = tenant
8107 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
8108 2 : .await?;
8109 :
8110 2 : let lsn = Lsn(0x10);
8111 2 : let compact = false;
8112 2 : bulk_insert_maybe_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000, compact).await?;
8113 :
8114 2 : let test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
8115 2 : let read_lsn = Lsn(u64::MAX - 1);
8116 :
8117 2 : let result = tline.get(test_key, read_lsn, &ctx).await;
8118 2 : assert!(result.is_ok(), "result is not Ok: {}", result.unwrap_err());
8119 :
8120 2 : Ok(())
8121 2 : }
8122 :
8123 : #[tokio::test]
8124 1 : async fn test_metadata_scan() -> anyhow::Result<()> {
8125 1 : let harness = TenantHarness::create("test_metadata_scan").await?;
8126 1 : let (tenant, ctx) = harness.load().await;
8127 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8128 1 : let tline = tenant
8129 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
8130 1 : .await?;
8131 :
8132 : const NUM_KEYS: usize = 1000;
8133 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
8134 :
8135 1 : let cancel = CancellationToken::new();
8136 :
8137 1 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
8138 1 : base_key.field1 = AUX_KEY_PREFIX;
8139 1 : let mut test_key = base_key;
8140 :
8141 : // Track when each page was last modified. Used to assert that
8142 : // a read sees the latest page version.
8143 1 : let mut updated = [Lsn(0); NUM_KEYS];
8144 :
8145 1 : let mut lsn = Lsn(0x10);
8146 : #[allow(clippy::needless_range_loop)]
8147 1001 : for blknum in 0..NUM_KEYS {
8148 1000 : lsn = Lsn(lsn.0 + 0x10);
8149 1000 : test_key.field6 = (blknum * STEP) as u32;
8150 1000 : let mut writer = tline.writer().await;
8151 1000 : writer
8152 1000 : .put(
8153 1000 : test_key,
8154 1000 : lsn,
8155 1000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
8156 1000 : &ctx,
8157 1000 : )
8158 1000 : .await?;
8159 1000 : writer.finish_write(lsn);
8160 1000 : updated[blknum] = lsn;
8161 1000 : drop(writer);
8162 : }
8163 :
8164 1 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
8165 :
8166 12 : for iter in 0..=10 {
8167 1 : // Read all the blocks
8168 11000 : for (blknum, last_lsn) in updated.iter().enumerate() {
8169 11000 : test_key.field6 = (blknum * STEP) as u32;
8170 11000 : assert_eq!(
8171 11000 : tline.get(test_key, lsn, &ctx).await?,
8172 11000 : test_img(&format!("{blknum} at {last_lsn}"))
8173 1 : );
8174 1 : }
8175 1 :
8176 11 : let mut cnt = 0;
8177 11 : let query = VersionedKeySpaceQuery::uniform(keyspace.clone(), lsn);
8178 1 :
8179 11000 : for (key, value) in tline
8180 11 : .get_vectored_impl(
8181 11 : query,
8182 11 : &mut ValuesReconstructState::new(io_concurrency.clone()),
8183 11 : &ctx,
8184 11 : )
8185 11 : .await?
8186 1 : {
8187 11000 : let blknum = key.field6 as usize;
8188 11000 : let value = value?;
8189 11000 : assert!(blknum % STEP == 0);
8190 11000 : let blknum = blknum / STEP;
8191 11000 : assert_eq!(
8192 1 : value,
8193 11000 : test_img(&format!("{} at {}", blknum, updated[blknum]))
8194 1 : );
8195 11000 : cnt += 1;
8196 1 : }
8197 1 :
8198 11 : assert_eq!(cnt, NUM_KEYS);
8199 1 :
8200 11011 : for _ in 0..NUM_KEYS {
8201 11000 : lsn = Lsn(lsn.0 + 0x10);
8202 11000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
8203 11000 : test_key.field6 = (blknum * STEP) as u32;
8204 11000 : let mut writer = tline.writer().await;
8205 11000 : writer
8206 11000 : .put(
8207 11000 : test_key,
8208 11000 : lsn,
8209 11000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
8210 11000 : &ctx,
8211 11000 : )
8212 11000 : .await?;
8213 11000 : writer.finish_write(lsn);
8214 11000 : drop(writer);
8215 11000 : updated[blknum] = lsn;
8216 1 : }
8217 1 :
8218 1 : // Perform two cycles of flush, compact, and GC
8219 33 : for round in 0..2 {
8220 22 : tline.freeze_and_flush().await?;
8221 22 : tline
8222 22 : .compact(
8223 22 : &cancel,
8224 22 : if iter % 5 == 0 && round == 0 {
8225 3 : let mut flags = EnumSet::new();
8226 3 : flags.insert(CompactFlags::ForceImageLayerCreation);
8227 3 : flags.insert(CompactFlags::ForceRepartition);
8228 3 : flags
8229 1 : } else {
8230 19 : EnumSet::empty()
8231 1 : },
8232 22 : &ctx,
8233 1 : )
8234 22 : .await?;
8235 22 : tenant
8236 22 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
8237 22 : .await?;
8238 1 : }
8239 1 : }
8240 1 :
8241 1 : Ok(())
8242 1 : }
8243 :
8244 : #[tokio::test]
8245 1 : async fn test_metadata_compaction_trigger() -> anyhow::Result<()> {
8246 1 : let harness = TenantHarness::create("test_metadata_compaction_trigger").await?;
8247 1 : let (tenant, ctx) = harness.load().await;
8248 1 : let tline = tenant
8249 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
8250 1 : .await?;
8251 :
8252 1 : let cancel = CancellationToken::new();
8253 :
8254 1 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
8255 1 : base_key.field1 = AUX_KEY_PREFIX;
8256 1 : let test_key = base_key;
8257 1 : let mut lsn = Lsn(0x10);
8258 :
8259 21 : for _ in 0..20 {
8260 20 : lsn = Lsn(lsn.0 + 0x10);
8261 20 : let mut writer = tline.writer().await;
8262 20 : writer
8263 20 : .put(
8264 20 : test_key,
8265 20 : lsn,
8266 20 : &Value::Image(test_img(&format!("{} at {}", 0, lsn))),
8267 20 : &ctx,
8268 20 : )
8269 20 : .await?;
8270 20 : writer.finish_write(lsn);
8271 20 : drop(writer);
8272 20 : tline.freeze_and_flush().await?; // force create a delta layer
8273 : }
8274 :
8275 1 : let before_num_l0_delta_files = tline
8276 1 : .layers
8277 1 : .read(LayerManagerLockHolder::Testing)
8278 1 : .await
8279 1 : .layer_map()?
8280 1 : .level0_deltas()
8281 1 : .len();
8282 :
8283 1 : tline.compact(&cancel, EnumSet::default(), &ctx).await?;
8284 :
8285 1 : let after_num_l0_delta_files = tline
8286 1 : .layers
8287 1 : .read(LayerManagerLockHolder::Testing)
8288 1 : .await
8289 1 : .layer_map()?
8290 1 : .level0_deltas()
8291 1 : .len();
8292 :
8293 1 : assert!(
8294 1 : after_num_l0_delta_files < before_num_l0_delta_files,
8295 0 : "after_num_l0_delta_files={after_num_l0_delta_files}, before_num_l0_delta_files={before_num_l0_delta_files}"
8296 : );
8297 :
8298 1 : assert_eq!(
8299 1 : tline.get(test_key, lsn, &ctx).await?,
8300 1 : test_img(&format!("{} at {}", 0, lsn))
8301 : );
8302 :
8303 2 : Ok(())
8304 1 : }
8305 :
8306 : #[tokio::test]
8307 1 : async fn test_aux_file_e2e() {
8308 1 : let harness = TenantHarness::create("test_aux_file_e2e").await.unwrap();
8309 :
8310 1 : let (tenant, ctx) = harness.load().await;
8311 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8312 :
8313 1 : let mut lsn = Lsn(0x08);
8314 :
8315 1 : let tline: Arc<Timeline> = tenant
8316 1 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
8317 1 : .await
8318 1 : .unwrap();
8319 :
8320 : {
8321 1 : lsn += 8;
8322 1 : let mut modification = tline.begin_modification(lsn);
8323 1 : modification
8324 1 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
8325 1 : .await
8326 1 : .unwrap();
8327 1 : modification.commit(&ctx).await.unwrap();
8328 : }
8329 :
8330 : // we can read everything from the storage
8331 1 : let files = tline
8332 1 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
8333 1 : .await
8334 1 : .unwrap();
8335 1 : assert_eq!(
8336 1 : files.get("pg_logical/mappings/test1"),
8337 1 : Some(&bytes::Bytes::from_static(b"first"))
8338 : );
8339 :
8340 : {
8341 1 : lsn += 8;
8342 1 : let mut modification = tline.begin_modification(lsn);
8343 1 : modification
8344 1 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
8345 1 : .await
8346 1 : .unwrap();
8347 1 : modification.commit(&ctx).await.unwrap();
8348 : }
8349 :
8350 1 : let files = tline
8351 1 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
8352 1 : .await
8353 1 : .unwrap();
8354 1 : assert_eq!(
8355 1 : files.get("pg_logical/mappings/test2"),
8356 1 : Some(&bytes::Bytes::from_static(b"second"))
8357 : );
8358 :
8359 1 : let child = tenant
8360 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(lsn), &ctx)
8361 1 : .await
8362 1 : .unwrap();
8363 :
8364 1 : let files = child
8365 1 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
8366 1 : .await
8367 1 : .unwrap();
8368 1 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
8369 1 : assert_eq!(files.get("pg_logical/mappings/test2"), None);
8370 1 : }
8371 :
8372 : #[tokio::test]
8373 1 : async fn test_repl_origin_tombstones() {
8374 1 : let harness = TenantHarness::create("test_repl_origin_tombstones")
8375 1 : .await
8376 1 : .unwrap();
8377 :
8378 1 : let (tenant, ctx) = harness.load().await;
8379 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8380 :
8381 1 : let mut lsn = Lsn(0x08);
8382 :
8383 1 : let tline: Arc<Timeline> = tenant
8384 1 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
8385 1 : .await
8386 1 : .unwrap();
8387 :
8388 1 : let repl_lsn = Lsn(0x10);
8389 : {
8390 1 : lsn += 8;
8391 1 : let mut modification = tline.begin_modification(lsn);
8392 1 : modification.put_for_unit_test(repl_origin_key(2), Value::Image(Bytes::new()));
8393 1 : modification.set_replorigin(1, repl_lsn).await.unwrap();
8394 1 : modification.commit(&ctx).await.unwrap();
8395 : }
8396 :
8397 : // we can read everything from the storage
8398 1 : let repl_origins = tline
8399 1 : .get_replorigins(lsn, &ctx, io_concurrency.clone())
8400 1 : .await
8401 1 : .unwrap();
8402 1 : assert_eq!(repl_origins.len(), 1);
8403 1 : assert_eq!(repl_origins[&1], lsn);
8404 :
8405 : {
8406 1 : lsn += 8;
8407 1 : let mut modification = tline.begin_modification(lsn);
8408 1 : modification.put_for_unit_test(
8409 1 : repl_origin_key(3),
8410 1 : Value::Image(Bytes::copy_from_slice(b"cannot_decode_this")),
8411 : );
8412 1 : modification.commit(&ctx).await.unwrap();
8413 : }
8414 1 : let result = tline
8415 1 : .get_replorigins(lsn, &ctx, io_concurrency.clone())
8416 1 : .await;
8417 1 : assert!(result.is_err());
8418 1 : }
8419 :
8420 : #[tokio::test]
8421 1 : async fn test_metadata_image_creation() -> anyhow::Result<()> {
8422 1 : let harness = TenantHarness::create("test_metadata_image_creation").await?;
8423 1 : let (tenant, ctx) = harness.load().await;
8424 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8425 1 : let tline = tenant
8426 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
8427 1 : .await?;
8428 :
8429 : const NUM_KEYS: usize = 1000;
8430 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
8431 :
8432 1 : let cancel = CancellationToken::new();
8433 :
8434 1 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8435 1 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
8436 1 : let mut test_key = base_key;
8437 1 : let mut lsn = Lsn(0x10);
8438 :
8439 4 : async fn scan_with_statistics(
8440 4 : tline: &Timeline,
8441 4 : keyspace: &KeySpace,
8442 4 : lsn: Lsn,
8443 4 : ctx: &RequestContext,
8444 4 : io_concurrency: IoConcurrency,
8445 4 : ) -> anyhow::Result<(BTreeMap<Key, Result<Bytes, PageReconstructError>>, usize)> {
8446 4 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
8447 4 : let query = VersionedKeySpaceQuery::uniform(keyspace.clone(), lsn);
8448 4 : let res = tline
8449 4 : .get_vectored_impl(query, &mut reconstruct_state, ctx)
8450 4 : .await?;
8451 4 : Ok((res, reconstruct_state.get_delta_layers_visited() as usize))
8452 4 : }
8453 :
8454 1001 : for blknum in 0..NUM_KEYS {
8455 1000 : lsn = Lsn(lsn.0 + 0x10);
8456 1000 : test_key.field6 = (blknum * STEP) as u32;
8457 1000 : let mut writer = tline.writer().await;
8458 1000 : writer
8459 1000 : .put(
8460 1000 : test_key,
8461 1000 : lsn,
8462 1000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
8463 1000 : &ctx,
8464 1000 : )
8465 1000 : .await?;
8466 1000 : writer.finish_write(lsn);
8467 1000 : drop(writer);
8468 : }
8469 :
8470 1 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
8471 :
8472 11 : for iter in 1..=10 {
8473 10010 : for _ in 0..NUM_KEYS {
8474 10000 : lsn = Lsn(lsn.0 + 0x10);
8475 10000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
8476 10000 : test_key.field6 = (blknum * STEP) as u32;
8477 10000 : let mut writer = tline.writer().await;
8478 10000 : writer
8479 10000 : .put(
8480 10000 : test_key,
8481 10000 : lsn,
8482 10000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
8483 10000 : &ctx,
8484 10000 : )
8485 10000 : .await?;
8486 10000 : writer.finish_write(lsn);
8487 10000 : drop(writer);
8488 1 : }
8489 1 :
8490 10 : tline.freeze_and_flush().await?;
8491 1 : // Force layers to L1
8492 10 : tline
8493 10 : .compact(
8494 10 : &cancel,
8495 10 : {
8496 10 : let mut flags = EnumSet::new();
8497 10 : flags.insert(CompactFlags::ForceL0Compaction);
8498 10 : flags
8499 10 : },
8500 10 : &ctx,
8501 10 : )
8502 10 : .await?;
8503 1 :
8504 10 : if iter % 5 == 0 {
8505 2 : let scan_lsn = Lsn(lsn.0 + 1);
8506 2 : info!("scanning at {}", scan_lsn);
8507 2 : let (_, before_delta_file_accessed) =
8508 2 : scan_with_statistics(&tline, &keyspace, scan_lsn, &ctx, io_concurrency.clone())
8509 2 : .await?;
8510 2 : tline
8511 2 : .compact(
8512 2 : &cancel,
8513 2 : {
8514 2 : let mut flags = EnumSet::new();
8515 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
8516 2 : flags.insert(CompactFlags::ForceRepartition);
8517 2 : flags.insert(CompactFlags::ForceL0Compaction);
8518 2 : flags
8519 2 : },
8520 2 : &ctx,
8521 2 : )
8522 2 : .await?;
8523 2 : let (_, after_delta_file_accessed) =
8524 2 : scan_with_statistics(&tline, &keyspace, scan_lsn, &ctx, io_concurrency.clone())
8525 2 : .await?;
8526 2 : assert!(
8527 2 : after_delta_file_accessed < before_delta_file_accessed,
8528 1 : "after_delta_file_accessed={after_delta_file_accessed}, before_delta_file_accessed={before_delta_file_accessed}"
8529 1 : );
8530 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.
8531 2 : assert!(
8532 2 : after_delta_file_accessed <= 2,
8533 1 : "after_delta_file_accessed={after_delta_file_accessed}"
8534 1 : );
8535 8 : }
8536 1 : }
8537 1 :
8538 1 : Ok(())
8539 1 : }
8540 :
8541 : #[tokio::test]
8542 1 : async fn test_vectored_missing_data_key_reads() -> anyhow::Result<()> {
8543 1 : let harness = TenantHarness::create("test_vectored_missing_data_key_reads").await?;
8544 1 : let (tenant, ctx) = harness.load().await;
8545 :
8546 1 : let base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
8547 1 : let base_key_child = Key::from_hex("000000000033333333444444445500000001").unwrap();
8548 1 : let base_key_nonexist = Key::from_hex("000000000033333333444444445500000002").unwrap();
8549 :
8550 1 : let tline = tenant
8551 1 : .create_test_timeline_with_layers(
8552 1 : TIMELINE_ID,
8553 1 : Lsn(0x10),
8554 1 : DEFAULT_PG_VERSION,
8555 1 : &ctx,
8556 1 : Vec::new(), // in-memory layers
8557 1 : Vec::new(), // delta layers
8558 1 : vec![(Lsn(0x20), vec![(base_key, test_img("data key 1"))])], // image layers
8559 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
8560 1 : )
8561 1 : .await?;
8562 1 : tline.add_extra_test_dense_keyspace(KeySpace::single(base_key..(base_key_nonexist.next())));
8563 :
8564 1 : let child = tenant
8565 1 : .branch_timeline_test_with_layers(
8566 1 : &tline,
8567 1 : NEW_TIMELINE_ID,
8568 1 : Some(Lsn(0x20)),
8569 1 : &ctx,
8570 1 : Vec::new(), // delta layers
8571 1 : vec![(Lsn(0x30), vec![(base_key_child, test_img("data key 2"))])], // image layers
8572 1 : Lsn(0x30),
8573 1 : )
8574 1 : .await
8575 1 : .unwrap();
8576 :
8577 1 : let lsn = Lsn(0x30);
8578 :
8579 : // test vectored get on parent timeline
8580 1 : assert_eq!(
8581 1 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
8582 1 : Some(test_img("data key 1"))
8583 : );
8584 1 : assert!(
8585 1 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx)
8586 1 : .await
8587 1 : .unwrap_err()
8588 1 : .is_missing_key_error()
8589 : );
8590 1 : assert!(
8591 1 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx)
8592 1 : .await
8593 1 : .unwrap_err()
8594 1 : .is_missing_key_error()
8595 : );
8596 :
8597 : // test vectored get on child timeline
8598 1 : assert_eq!(
8599 1 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
8600 1 : Some(test_img("data key 1"))
8601 : );
8602 1 : assert_eq!(
8603 1 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
8604 1 : Some(test_img("data key 2"))
8605 : );
8606 1 : assert!(
8607 1 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx)
8608 1 : .await
8609 1 : .unwrap_err()
8610 1 : .is_missing_key_error()
8611 : );
8612 :
8613 2 : Ok(())
8614 1 : }
8615 :
8616 : #[tokio::test]
8617 1 : async fn test_vectored_missing_metadata_key_reads() -> anyhow::Result<()> {
8618 1 : let harness = TenantHarness::create("test_vectored_missing_metadata_key_reads").await?;
8619 1 : let (tenant, ctx) = harness.load().await;
8620 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8621 :
8622 1 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8623 1 : let base_key_child = Key::from_hex("620000000033333333444444445500000001").unwrap();
8624 1 : let base_key_nonexist = Key::from_hex("620000000033333333444444445500000002").unwrap();
8625 1 : let base_key_overwrite = Key::from_hex("620000000033333333444444445500000003").unwrap();
8626 :
8627 1 : let base_inherited_key = Key::from_hex("610000000033333333444444445500000000").unwrap();
8628 1 : let base_inherited_key_child =
8629 1 : Key::from_hex("610000000033333333444444445500000001").unwrap();
8630 1 : let base_inherited_key_nonexist =
8631 1 : Key::from_hex("610000000033333333444444445500000002").unwrap();
8632 1 : let base_inherited_key_overwrite =
8633 1 : Key::from_hex("610000000033333333444444445500000003").unwrap();
8634 :
8635 1 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
8636 1 : assert_eq!(base_inherited_key.field1, RELATION_SIZE_PREFIX);
8637 :
8638 1 : let tline = tenant
8639 1 : .create_test_timeline_with_layers(
8640 1 : TIMELINE_ID,
8641 1 : Lsn(0x10),
8642 1 : DEFAULT_PG_VERSION,
8643 1 : &ctx,
8644 1 : Vec::new(), // in-memory layers
8645 1 : Vec::new(), // delta layers
8646 1 : vec![(
8647 1 : Lsn(0x20),
8648 1 : vec![
8649 1 : (base_inherited_key, test_img("metadata inherited key 1")),
8650 1 : (
8651 1 : base_inherited_key_overwrite,
8652 1 : test_img("metadata key overwrite 1a"),
8653 1 : ),
8654 1 : (base_key, test_img("metadata key 1")),
8655 1 : (base_key_overwrite, test_img("metadata key overwrite 1b")),
8656 1 : ],
8657 1 : )], // image layers
8658 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
8659 1 : )
8660 1 : .await?;
8661 :
8662 1 : let child = tenant
8663 1 : .branch_timeline_test_with_layers(
8664 1 : &tline,
8665 1 : NEW_TIMELINE_ID,
8666 1 : Some(Lsn(0x20)),
8667 1 : &ctx,
8668 1 : Vec::new(), // delta layers
8669 1 : vec![(
8670 1 : Lsn(0x30),
8671 1 : vec![
8672 1 : (
8673 1 : base_inherited_key_child,
8674 1 : test_img("metadata inherited key 2"),
8675 1 : ),
8676 1 : (
8677 1 : base_inherited_key_overwrite,
8678 1 : test_img("metadata key overwrite 2a"),
8679 1 : ),
8680 1 : (base_key_child, test_img("metadata key 2")),
8681 1 : (base_key_overwrite, test_img("metadata key overwrite 2b")),
8682 1 : ],
8683 1 : )], // image layers
8684 1 : Lsn(0x30),
8685 1 : )
8686 1 : .await
8687 1 : .unwrap();
8688 :
8689 1 : let lsn = Lsn(0x30);
8690 :
8691 : // test vectored get on parent timeline
8692 1 : assert_eq!(
8693 1 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
8694 1 : Some(test_img("metadata key 1"))
8695 : );
8696 1 : assert_eq!(
8697 1 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx).await?,
8698 : None
8699 : );
8700 1 : assert_eq!(
8701 1 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx).await?,
8702 : None
8703 : );
8704 1 : assert_eq!(
8705 1 : get_vectored_impl_wrapper(&tline, base_key_overwrite, lsn, &ctx).await?,
8706 1 : Some(test_img("metadata key overwrite 1b"))
8707 : );
8708 1 : assert_eq!(
8709 1 : get_vectored_impl_wrapper(&tline, base_inherited_key, lsn, &ctx).await?,
8710 1 : Some(test_img("metadata inherited key 1"))
8711 : );
8712 1 : assert_eq!(
8713 1 : get_vectored_impl_wrapper(&tline, base_inherited_key_child, lsn, &ctx).await?,
8714 : None
8715 : );
8716 1 : assert_eq!(
8717 1 : get_vectored_impl_wrapper(&tline, base_inherited_key_nonexist, lsn, &ctx).await?,
8718 : None
8719 : );
8720 1 : assert_eq!(
8721 1 : get_vectored_impl_wrapper(&tline, base_inherited_key_overwrite, lsn, &ctx).await?,
8722 1 : Some(test_img("metadata key overwrite 1a"))
8723 : );
8724 :
8725 : // test vectored get on child timeline
8726 1 : assert_eq!(
8727 1 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
8728 : None
8729 : );
8730 1 : assert_eq!(
8731 1 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
8732 1 : Some(test_img("metadata key 2"))
8733 : );
8734 1 : assert_eq!(
8735 1 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx).await?,
8736 : None
8737 : );
8738 1 : assert_eq!(
8739 1 : get_vectored_impl_wrapper(&child, 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(&child, base_inherited_key_child, lsn, &ctx).await?,
8744 1 : Some(test_img("metadata inherited key 2"))
8745 : );
8746 1 : assert_eq!(
8747 1 : get_vectored_impl_wrapper(&child, base_inherited_key_nonexist, lsn, &ctx).await?,
8748 : None
8749 : );
8750 1 : assert_eq!(
8751 1 : get_vectored_impl_wrapper(&child, base_key_overwrite, lsn, &ctx).await?,
8752 1 : Some(test_img("metadata key overwrite 2b"))
8753 : );
8754 1 : assert_eq!(
8755 1 : get_vectored_impl_wrapper(&child, base_inherited_key_overwrite, lsn, &ctx).await?,
8756 1 : Some(test_img("metadata key overwrite 2a"))
8757 : );
8758 :
8759 : // test vectored scan on parent timeline
8760 1 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency.clone());
8761 1 : let query =
8762 1 : VersionedKeySpaceQuery::uniform(KeySpace::single(Key::metadata_key_range()), lsn);
8763 1 : let res = tline
8764 1 : .get_vectored_impl(query, &mut reconstruct_state, &ctx)
8765 1 : .await?;
8766 :
8767 1 : assert_eq!(
8768 1 : res.into_iter()
8769 4 : .map(|(k, v)| (k, v.unwrap()))
8770 1 : .collect::<Vec<_>>(),
8771 1 : vec![
8772 1 : (base_inherited_key, test_img("metadata inherited key 1")),
8773 1 : (
8774 1 : base_inherited_key_overwrite,
8775 1 : test_img("metadata key overwrite 1a")
8776 1 : ),
8777 1 : (base_key, test_img("metadata key 1")),
8778 1 : (base_key_overwrite, test_img("metadata key overwrite 1b")),
8779 : ]
8780 : );
8781 :
8782 : // test vectored scan on child timeline
8783 1 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency.clone());
8784 1 : let query =
8785 1 : VersionedKeySpaceQuery::uniform(KeySpace::single(Key::metadata_key_range()), lsn);
8786 1 : let res = child
8787 1 : .get_vectored_impl(query, &mut reconstruct_state, &ctx)
8788 1 : .await?;
8789 :
8790 1 : assert_eq!(
8791 1 : res.into_iter()
8792 5 : .map(|(k, v)| (k, v.unwrap()))
8793 1 : .collect::<Vec<_>>(),
8794 1 : vec![
8795 1 : (base_inherited_key, test_img("metadata inherited key 1")),
8796 1 : (
8797 1 : base_inherited_key_child,
8798 1 : test_img("metadata inherited key 2")
8799 1 : ),
8800 1 : (
8801 1 : base_inherited_key_overwrite,
8802 1 : test_img("metadata key overwrite 2a")
8803 1 : ),
8804 1 : (base_key_child, test_img("metadata key 2")),
8805 1 : (base_key_overwrite, test_img("metadata key overwrite 2b")),
8806 : ]
8807 : );
8808 :
8809 2 : Ok(())
8810 1 : }
8811 :
8812 28 : async fn get_vectored_impl_wrapper(
8813 28 : tline: &Arc<Timeline>,
8814 28 : key: Key,
8815 28 : lsn: Lsn,
8816 28 : ctx: &RequestContext,
8817 28 : ) -> Result<Option<Bytes>, GetVectoredError> {
8818 28 : let io_concurrency = IoConcurrency::spawn_from_conf(
8819 28 : tline.conf.get_vectored_concurrent_io,
8820 28 : tline.gate.enter().unwrap(),
8821 : );
8822 28 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
8823 28 : let query = VersionedKeySpaceQuery::uniform(KeySpace::single(key..key.next()), lsn);
8824 28 : let mut res = tline
8825 28 : .get_vectored_impl(query, &mut reconstruct_state, ctx)
8826 28 : .await?;
8827 25 : Ok(res.pop_last().map(|(k, v)| {
8828 16 : assert_eq!(k, key);
8829 16 : v.unwrap()
8830 16 : }))
8831 28 : }
8832 :
8833 : #[tokio::test]
8834 1 : async fn test_metadata_tombstone_reads() -> anyhow::Result<()> {
8835 1 : let harness = TenantHarness::create("test_metadata_tombstone_reads").await?;
8836 1 : let (tenant, ctx) = harness.load().await;
8837 1 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
8838 1 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8839 1 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8840 1 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
8841 :
8842 : // We emulate the situation that the compaction algorithm creates an image layer that removes the tombstones
8843 : // Lsn 0x30 key0, key3, no key1+key2
8844 : // Lsn 0x20 key1+key2 tomestones
8845 : // Lsn 0x10 key1 in image, key2 in delta
8846 1 : let tline = tenant
8847 1 : .create_test_timeline_with_layers(
8848 1 : TIMELINE_ID,
8849 1 : Lsn(0x10),
8850 1 : DEFAULT_PG_VERSION,
8851 1 : &ctx,
8852 1 : Vec::new(), // in-memory layers
8853 1 : // delta layers
8854 1 : vec![
8855 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8856 1 : Lsn(0x10)..Lsn(0x20),
8857 1 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8858 1 : ),
8859 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8860 1 : Lsn(0x20)..Lsn(0x30),
8861 1 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8862 1 : ),
8863 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8864 1 : Lsn(0x20)..Lsn(0x30),
8865 1 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8866 1 : ),
8867 1 : ],
8868 1 : // image layers
8869 1 : vec![
8870 1 : (Lsn(0x10), vec![(key1, test_img("metadata key 1"))]),
8871 1 : (
8872 1 : Lsn(0x30),
8873 1 : vec![
8874 1 : (key0, test_img("metadata key 0")),
8875 1 : (key3, test_img("metadata key 3")),
8876 1 : ],
8877 1 : ),
8878 1 : ],
8879 1 : Lsn(0x30),
8880 1 : )
8881 1 : .await?;
8882 :
8883 1 : let lsn = Lsn(0x30);
8884 1 : let old_lsn = Lsn(0x20);
8885 :
8886 1 : assert_eq!(
8887 1 : get_vectored_impl_wrapper(&tline, key0, lsn, &ctx).await?,
8888 1 : Some(test_img("metadata key 0"))
8889 : );
8890 1 : assert_eq!(
8891 1 : get_vectored_impl_wrapper(&tline, key1, lsn, &ctx).await?,
8892 : None,
8893 : );
8894 1 : assert_eq!(
8895 1 : get_vectored_impl_wrapper(&tline, key2, lsn, &ctx).await?,
8896 : None,
8897 : );
8898 1 : assert_eq!(
8899 1 : get_vectored_impl_wrapper(&tline, key1, old_lsn, &ctx).await?,
8900 1 : Some(Bytes::new()),
8901 : );
8902 1 : assert_eq!(
8903 1 : get_vectored_impl_wrapper(&tline, key2, old_lsn, &ctx).await?,
8904 1 : Some(Bytes::new()),
8905 : );
8906 1 : assert_eq!(
8907 1 : get_vectored_impl_wrapper(&tline, key3, lsn, &ctx).await?,
8908 1 : Some(test_img("metadata key 3"))
8909 : );
8910 :
8911 2 : Ok(())
8912 1 : }
8913 :
8914 : #[tokio::test]
8915 1 : async fn test_metadata_tombstone_image_creation() {
8916 1 : let harness = TenantHarness::create("test_metadata_tombstone_image_creation")
8917 1 : .await
8918 1 : .unwrap();
8919 1 : let (tenant, ctx) = harness.load().await;
8920 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8921 :
8922 1 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
8923 1 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8924 1 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8925 1 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
8926 :
8927 1 : let tline = tenant
8928 1 : .create_test_timeline_with_layers(
8929 1 : TIMELINE_ID,
8930 1 : Lsn(0x10),
8931 1 : DEFAULT_PG_VERSION,
8932 1 : &ctx,
8933 1 : Vec::new(), // in-memory layers
8934 1 : // delta layers
8935 1 : vec![
8936 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8937 1 : Lsn(0x10)..Lsn(0x20),
8938 1 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8939 1 : ),
8940 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8941 1 : Lsn(0x20)..Lsn(0x30),
8942 1 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8943 1 : ),
8944 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8945 1 : Lsn(0x20)..Lsn(0x30),
8946 1 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8947 1 : ),
8948 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8949 1 : Lsn(0x30)..Lsn(0x40),
8950 1 : vec![
8951 1 : (key0, Lsn(0x30), Value::Image(test_img("metadata key 0"))),
8952 1 : (key3, Lsn(0x30), Value::Image(test_img("metadata key 3"))),
8953 1 : ],
8954 1 : ),
8955 1 : ],
8956 1 : // image layers
8957 1 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
8958 1 : Lsn(0x40),
8959 1 : )
8960 1 : .await
8961 1 : .unwrap();
8962 :
8963 1 : let cancel = CancellationToken::new();
8964 :
8965 : // Image layer creation happens on the disk_consistent_lsn so we need to force set it now.
8966 1 : tline.force_set_disk_consistent_lsn(Lsn(0x40));
8967 1 : tline
8968 1 : .compact(
8969 1 : &cancel,
8970 1 : {
8971 1 : let mut flags = EnumSet::new();
8972 1 : flags.insert(CompactFlags::ForceImageLayerCreation);
8973 1 : flags.insert(CompactFlags::ForceRepartition);
8974 1 : flags
8975 1 : },
8976 1 : &ctx,
8977 1 : )
8978 1 : .await
8979 1 : .unwrap();
8980 : // Image layers are created at repartition LSN
8981 1 : let images = tline
8982 1 : .inspect_image_layers(Lsn(0x40), &ctx, io_concurrency.clone())
8983 1 : .await
8984 1 : .unwrap()
8985 1 : .into_iter()
8986 9 : .filter(|(k, _)| k.is_metadata_key())
8987 1 : .collect::<Vec<_>>();
8988 1 : assert_eq!(images.len(), 2); // the image layer should only contain two existing keys, tombstones should be removed.
8989 1 : }
8990 :
8991 : #[tokio::test]
8992 1 : async fn test_metadata_tombstone_empty_image_creation() {
8993 1 : let harness = TenantHarness::create("test_metadata_tombstone_empty_image_creation")
8994 1 : .await
8995 1 : .unwrap();
8996 1 : let (tenant, ctx) = harness.load().await;
8997 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8998 :
8999 1 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
9000 1 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
9001 :
9002 1 : let tline = tenant
9003 1 : .create_test_timeline_with_layers(
9004 1 : TIMELINE_ID,
9005 1 : Lsn(0x10),
9006 1 : DEFAULT_PG_VERSION,
9007 1 : &ctx,
9008 1 : Vec::new(), // in-memory layers
9009 1 : // delta layers
9010 1 : vec![
9011 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9012 1 : Lsn(0x10)..Lsn(0x20),
9013 1 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
9014 1 : ),
9015 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9016 1 : Lsn(0x20)..Lsn(0x30),
9017 1 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
9018 1 : ),
9019 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9020 1 : Lsn(0x20)..Lsn(0x30),
9021 1 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
9022 1 : ),
9023 1 : ],
9024 1 : // image layers
9025 1 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
9026 1 : Lsn(0x30),
9027 1 : )
9028 1 : .await
9029 1 : .unwrap();
9030 :
9031 1 : let cancel = CancellationToken::new();
9032 :
9033 1 : tline
9034 1 : .compact(
9035 1 : &cancel,
9036 1 : {
9037 1 : let mut flags = EnumSet::new();
9038 1 : flags.insert(CompactFlags::ForceImageLayerCreation);
9039 1 : flags.insert(CompactFlags::ForceRepartition);
9040 1 : flags
9041 1 : },
9042 1 : &ctx,
9043 1 : )
9044 1 : .await
9045 1 : .unwrap();
9046 :
9047 : // Image layers are created at last_record_lsn
9048 1 : let images = tline
9049 1 : .inspect_image_layers(Lsn(0x30), &ctx, io_concurrency.clone())
9050 1 : .await
9051 1 : .unwrap()
9052 1 : .into_iter()
9053 7 : .filter(|(k, _)| k.is_metadata_key())
9054 1 : .collect::<Vec<_>>();
9055 1 : assert_eq!(images.len(), 0); // the image layer should not contain tombstones, or it is not created
9056 1 : }
9057 :
9058 : #[tokio::test]
9059 1 : async fn test_simple_bottom_most_compaction_images() -> anyhow::Result<()> {
9060 1 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_images").await?;
9061 1 : let (tenant, ctx) = harness.load().await;
9062 1 : let io_concurrency = IoConcurrency::spawn_for_test();
9063 :
9064 51 : fn get_key(id: u32) -> Key {
9065 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9066 51 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9067 51 : key.field6 = id;
9068 51 : key
9069 51 : }
9070 :
9071 : // We create
9072 : // - one bottom-most image layer,
9073 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
9074 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
9075 : // - a delta layer D3 above the horizon.
9076 : //
9077 : // | D3 |
9078 : // | D1 |
9079 : // -| |-- gc horizon -----------------
9080 : // | | | D2 |
9081 : // --------- img layer ------------------
9082 : //
9083 : // What we should expact from this compaction is:
9084 : // | D3 |
9085 : // | Part of D1 |
9086 : // --------- img layer with D1+D2 at GC horizon------------------
9087 :
9088 : // img layer at 0x10
9089 1 : let img_layer = (0..10)
9090 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9091 1 : .collect_vec();
9092 :
9093 1 : let delta1 = vec![
9094 1 : (
9095 1 : get_key(1),
9096 1 : Lsn(0x20),
9097 1 : Value::Image(Bytes::from("value 1@0x20")),
9098 1 : ),
9099 1 : (
9100 1 : get_key(2),
9101 1 : Lsn(0x30),
9102 1 : Value::Image(Bytes::from("value 2@0x30")),
9103 1 : ),
9104 1 : (
9105 1 : get_key(3),
9106 1 : Lsn(0x40),
9107 1 : Value::Image(Bytes::from("value 3@0x40")),
9108 1 : ),
9109 : ];
9110 1 : let delta2 = vec![
9111 1 : (
9112 1 : get_key(5),
9113 1 : Lsn(0x20),
9114 1 : Value::Image(Bytes::from("value 5@0x20")),
9115 1 : ),
9116 1 : (
9117 1 : get_key(6),
9118 1 : Lsn(0x20),
9119 1 : Value::Image(Bytes::from("value 6@0x20")),
9120 1 : ),
9121 : ];
9122 1 : let delta3 = vec![
9123 1 : (
9124 1 : get_key(8),
9125 1 : Lsn(0x48),
9126 1 : Value::Image(Bytes::from("value 8@0x48")),
9127 1 : ),
9128 1 : (
9129 1 : get_key(9),
9130 1 : Lsn(0x48),
9131 1 : Value::Image(Bytes::from("value 9@0x48")),
9132 1 : ),
9133 : ];
9134 :
9135 1 : let tline = tenant
9136 1 : .create_test_timeline_with_layers(
9137 1 : TIMELINE_ID,
9138 1 : Lsn(0x10),
9139 1 : DEFAULT_PG_VERSION,
9140 1 : &ctx,
9141 1 : Vec::new(), // in-memory layers
9142 1 : vec![
9143 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
9144 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
9145 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
9146 1 : ], // delta layers
9147 1 : vec![(Lsn(0x10), img_layer)], // image layers
9148 1 : Lsn(0x50),
9149 1 : )
9150 1 : .await?;
9151 : {
9152 1 : tline
9153 1 : .applied_gc_cutoff_lsn
9154 1 : .lock_for_write()
9155 1 : .store_and_unlock(Lsn(0x30))
9156 1 : .wait()
9157 1 : .await;
9158 : // Update GC info
9159 1 : let mut guard = tline.gc_info.write().unwrap();
9160 1 : guard.cutoffs.time = Some(Lsn(0x30));
9161 1 : guard.cutoffs.space = Lsn(0x30);
9162 : }
9163 :
9164 1 : let expected_result = [
9165 1 : Bytes::from_static(b"value 0@0x10"),
9166 1 : Bytes::from_static(b"value 1@0x20"),
9167 1 : Bytes::from_static(b"value 2@0x30"),
9168 1 : Bytes::from_static(b"value 3@0x40"),
9169 1 : Bytes::from_static(b"value 4@0x10"),
9170 1 : Bytes::from_static(b"value 5@0x20"),
9171 1 : Bytes::from_static(b"value 6@0x20"),
9172 1 : Bytes::from_static(b"value 7@0x10"),
9173 1 : Bytes::from_static(b"value 8@0x48"),
9174 1 : Bytes::from_static(b"value 9@0x48"),
9175 1 : ];
9176 :
9177 10 : for (idx, expected) in expected_result.iter().enumerate() {
9178 10 : assert_eq!(
9179 10 : tline
9180 10 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9181 10 : .await
9182 10 : .unwrap(),
9183 : expected
9184 : );
9185 : }
9186 :
9187 1 : let cancel = CancellationToken::new();
9188 1 : tline
9189 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9190 1 : .await
9191 1 : .unwrap();
9192 :
9193 10 : for (idx, expected) in expected_result.iter().enumerate() {
9194 10 : assert_eq!(
9195 10 : tline
9196 10 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9197 10 : .await
9198 10 : .unwrap(),
9199 : expected
9200 : );
9201 : }
9202 :
9203 : // Check if the image layer at the GC horizon contains exactly what we want
9204 1 : let image_at_gc_horizon = tline
9205 1 : .inspect_image_layers(Lsn(0x30), &ctx, io_concurrency.clone())
9206 1 : .await
9207 1 : .unwrap()
9208 1 : .into_iter()
9209 17 : .filter(|(k, _)| k.is_metadata_key())
9210 1 : .collect::<Vec<_>>();
9211 :
9212 1 : assert_eq!(image_at_gc_horizon.len(), 10);
9213 1 : let expected_result = [
9214 1 : Bytes::from_static(b"value 0@0x10"),
9215 1 : Bytes::from_static(b"value 1@0x20"),
9216 1 : Bytes::from_static(b"value 2@0x30"),
9217 1 : Bytes::from_static(b"value 3@0x10"),
9218 1 : Bytes::from_static(b"value 4@0x10"),
9219 1 : Bytes::from_static(b"value 5@0x20"),
9220 1 : Bytes::from_static(b"value 6@0x20"),
9221 1 : Bytes::from_static(b"value 7@0x10"),
9222 1 : Bytes::from_static(b"value 8@0x10"),
9223 1 : Bytes::from_static(b"value 9@0x10"),
9224 1 : ];
9225 11 : for idx in 0..10 {
9226 10 : assert_eq!(
9227 10 : image_at_gc_horizon[idx],
9228 10 : (get_key(idx as u32), expected_result[idx].clone())
9229 : );
9230 : }
9231 :
9232 : // Check if old layers are removed / new layers have the expected LSN
9233 1 : let all_layers = inspect_and_sort(&tline, None).await;
9234 1 : assert_eq!(
9235 : all_layers,
9236 1 : vec![
9237 : // Image layer at GC horizon
9238 1 : PersistentLayerKey {
9239 1 : key_range: Key::MIN..Key::MAX,
9240 1 : lsn_range: Lsn(0x30)..Lsn(0x31),
9241 1 : is_delta: false
9242 1 : },
9243 : // The delta layer below the horizon
9244 1 : PersistentLayerKey {
9245 1 : key_range: get_key(3)..get_key(4),
9246 1 : lsn_range: Lsn(0x30)..Lsn(0x48),
9247 1 : is_delta: true
9248 1 : },
9249 : // The delta3 layer that should not be picked for the compaction
9250 1 : PersistentLayerKey {
9251 1 : key_range: get_key(8)..get_key(10),
9252 1 : lsn_range: Lsn(0x48)..Lsn(0x50),
9253 1 : is_delta: true
9254 1 : }
9255 : ]
9256 : );
9257 :
9258 : // increase GC horizon and compact again
9259 : {
9260 1 : tline
9261 1 : .applied_gc_cutoff_lsn
9262 1 : .lock_for_write()
9263 1 : .store_and_unlock(Lsn(0x40))
9264 1 : .wait()
9265 1 : .await;
9266 : // Update GC info
9267 1 : let mut guard = tline.gc_info.write().unwrap();
9268 1 : guard.cutoffs.time = Some(Lsn(0x40));
9269 1 : guard.cutoffs.space = Lsn(0x40);
9270 : }
9271 1 : tline
9272 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9273 1 : .await
9274 1 : .unwrap();
9275 :
9276 2 : Ok(())
9277 1 : }
9278 :
9279 : #[cfg(feature = "testing")]
9280 : #[tokio::test]
9281 1 : async fn test_neon_test_record() -> anyhow::Result<()> {
9282 1 : let harness = TenantHarness::create("test_neon_test_record").await?;
9283 1 : let (tenant, ctx) = harness.load().await;
9284 :
9285 17 : fn get_key(id: u32) -> Key {
9286 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9287 17 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9288 17 : key.field6 = id;
9289 17 : key
9290 17 : }
9291 :
9292 1 : let delta1 = vec![
9293 1 : (
9294 1 : get_key(1),
9295 1 : Lsn(0x20),
9296 1 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
9297 1 : ),
9298 1 : (
9299 1 : get_key(1),
9300 1 : Lsn(0x30),
9301 1 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
9302 1 : ),
9303 1 : (get_key(2), Lsn(0x10), Value::Image("0x10".into())),
9304 1 : (
9305 1 : get_key(2),
9306 1 : Lsn(0x20),
9307 1 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
9308 1 : ),
9309 1 : (
9310 1 : get_key(2),
9311 1 : Lsn(0x30),
9312 1 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
9313 1 : ),
9314 1 : (get_key(3), Lsn(0x10), Value::Image("0x10".into())),
9315 1 : (
9316 1 : get_key(3),
9317 1 : Lsn(0x20),
9318 1 : Value::WalRecord(NeonWalRecord::wal_clear("c")),
9319 1 : ),
9320 1 : (get_key(4), Lsn(0x10), Value::Image("0x10".into())),
9321 1 : (
9322 1 : get_key(4),
9323 1 : Lsn(0x20),
9324 1 : Value::WalRecord(NeonWalRecord::wal_init("i")),
9325 1 : ),
9326 1 : (
9327 1 : get_key(4),
9328 1 : Lsn(0x30),
9329 1 : Value::WalRecord(NeonWalRecord::wal_append_conditional("j", "i")),
9330 1 : ),
9331 1 : (
9332 1 : get_key(5),
9333 1 : Lsn(0x20),
9334 1 : Value::WalRecord(NeonWalRecord::wal_init("1")),
9335 1 : ),
9336 1 : (
9337 1 : get_key(5),
9338 1 : Lsn(0x30),
9339 1 : Value::WalRecord(NeonWalRecord::wal_append_conditional("j", "2")),
9340 1 : ),
9341 : ];
9342 1 : let image1 = vec![(get_key(1), "0x10".into())];
9343 :
9344 1 : let tline = tenant
9345 1 : .create_test_timeline_with_layers(
9346 1 : TIMELINE_ID,
9347 1 : Lsn(0x10),
9348 1 : DEFAULT_PG_VERSION,
9349 1 : &ctx,
9350 1 : Vec::new(), // in-memory layers
9351 1 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
9352 1 : Lsn(0x10)..Lsn(0x40),
9353 1 : delta1,
9354 1 : )], // delta layers
9355 1 : vec![(Lsn(0x10), image1)], // image layers
9356 1 : Lsn(0x50),
9357 1 : )
9358 1 : .await?;
9359 :
9360 1 : assert_eq!(
9361 1 : tline.get(get_key(1), Lsn(0x50), &ctx).await?,
9362 1 : Bytes::from_static(b"0x10,0x20,0x30")
9363 : );
9364 1 : assert_eq!(
9365 1 : tline.get(get_key(2), Lsn(0x50), &ctx).await?,
9366 1 : Bytes::from_static(b"0x10,0x20,0x30")
9367 : );
9368 :
9369 : // Need to remove the limit of "Neon WAL redo requires base image".
9370 :
9371 1 : assert_eq!(
9372 1 : tline.get(get_key(3), Lsn(0x50), &ctx).await?,
9373 1 : Bytes::from_static(b"c")
9374 : );
9375 1 : assert_eq!(
9376 1 : tline.get(get_key(4), Lsn(0x50), &ctx).await?,
9377 1 : Bytes::from_static(b"ij")
9378 : );
9379 :
9380 : // Manual testing required: currently, read errors will panic the process in debug mode. So we
9381 : // cannot enable this assertion in the unit test.
9382 : // assert!(tline.get(get_key(5), Lsn(0x50), &ctx).await.is_err());
9383 :
9384 2 : Ok(())
9385 1 : }
9386 :
9387 : #[tokio::test(start_paused = true)]
9388 1 : async fn test_lsn_lease() -> anyhow::Result<()> {
9389 1 : let (tenant, ctx) = TenantHarness::create("test_lsn_lease")
9390 1 : .await
9391 1 : .unwrap()
9392 1 : .load()
9393 1 : .await;
9394 : // Advance to the lsn lease deadline so that GC is not blocked by
9395 : // initial transition into AttachedSingle.
9396 1 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
9397 1 : tokio::time::resume();
9398 1 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
9399 :
9400 1 : let end_lsn = Lsn(0x100);
9401 1 : let image_layers = (0x20..=0x90)
9402 1 : .step_by(0x10)
9403 8 : .map(|n| (Lsn(n), vec![(key, test_img(&format!("data key at {n:x}")))]))
9404 1 : .collect();
9405 :
9406 1 : let timeline = tenant
9407 1 : .create_test_timeline_with_layers(
9408 1 : TIMELINE_ID,
9409 1 : Lsn(0x10),
9410 1 : DEFAULT_PG_VERSION,
9411 1 : &ctx,
9412 1 : Vec::new(), // in-memory layers
9413 1 : Vec::new(),
9414 1 : image_layers,
9415 1 : end_lsn,
9416 1 : )
9417 1 : .await?;
9418 :
9419 1 : let leased_lsns = [0x30, 0x50, 0x70];
9420 1 : let mut leases = Vec::new();
9421 3 : leased_lsns.iter().for_each(|n| {
9422 3 : leases.push(
9423 3 : timeline
9424 3 : .init_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)
9425 3 : .expect("lease request should succeed"),
9426 : );
9427 3 : });
9428 :
9429 1 : let updated_lease_0 = timeline
9430 1 : .renew_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)
9431 1 : .expect("lease renewal should succeed");
9432 1 : assert_eq!(
9433 1 : updated_lease_0.valid_until, leases[0].valid_until,
9434 0 : " Renewing with shorter lease should not change the lease."
9435 : );
9436 :
9437 1 : let updated_lease_1 = timeline
9438 1 : .renew_lsn_lease(
9439 1 : Lsn(leased_lsns[1]),
9440 1 : timeline.get_lsn_lease_length() * 2,
9441 1 : &ctx,
9442 1 : )
9443 1 : .expect("lease renewal should succeed");
9444 1 : assert!(
9445 1 : updated_lease_1.valid_until > leases[1].valid_until,
9446 0 : "Renewing with a long lease should renew lease with later expiration time."
9447 : );
9448 :
9449 : // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
9450 1 : info!(
9451 0 : "applied_gc_cutoff_lsn: {}",
9452 0 : *timeline.get_applied_gc_cutoff_lsn()
9453 : );
9454 1 : timeline.force_set_disk_consistent_lsn(end_lsn);
9455 :
9456 1 : let res = tenant
9457 1 : .gc_iteration(
9458 1 : Some(TIMELINE_ID),
9459 1 : 0,
9460 1 : Duration::ZERO,
9461 1 : &CancellationToken::new(),
9462 1 : &ctx,
9463 1 : )
9464 1 : .await
9465 1 : .unwrap();
9466 :
9467 : // Keeping everything <= Lsn(0x80) b/c leases:
9468 : // 0/10: initdb layer
9469 : // (0/20..=0/70).step_by(0x10): image layers added when creating the timeline.
9470 1 : assert_eq!(res.layers_needed_by_leases, 7);
9471 : // Keeping 0/90 b/c it is the latest layer.
9472 1 : assert_eq!(res.layers_not_updated, 1);
9473 : // Removed 0/80.
9474 1 : assert_eq!(res.layers_removed, 1);
9475 :
9476 : // Make lease on a already GC-ed LSN.
9477 : // 0/80 does not have a valid lease + is below latest_gc_cutoff
9478 1 : assert!(Lsn(0x80) < *timeline.get_applied_gc_cutoff_lsn());
9479 1 : timeline
9480 1 : .init_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx)
9481 1 : .expect_err("lease request on GC-ed LSN should fail");
9482 :
9483 : // Should still be able to renew a currently valid lease
9484 : // Assumption: original lease to is still valid for 0/50.
9485 : // (use `Timeline::init_lsn_lease` for testing so it always does validation)
9486 1 : timeline
9487 1 : .init_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)
9488 1 : .expect("lease renewal with validation should succeed");
9489 :
9490 2 : Ok(())
9491 1 : }
9492 :
9493 : #[tokio::test]
9494 1 : async fn test_failed_flush_should_not_update_disk_consistent_lsn() -> anyhow::Result<()> {
9495 : //
9496 : // Setup
9497 : //
9498 1 : let harness = TenantHarness::create_custom(
9499 1 : "test_failed_flush_should_not_upload_disk_consistent_lsn",
9500 1 : pageserver_api::models::TenantConfig::default(),
9501 1 : TenantId::generate(),
9502 1 : ShardIdentity::new(ShardNumber(0), ShardCount(4), ShardStripeSize(128)).unwrap(),
9503 1 : Generation::new(1),
9504 1 : )
9505 1 : .await?;
9506 1 : let (tenant, ctx) = harness.load().await;
9507 :
9508 1 : let timeline = tenant
9509 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
9510 1 : .await?;
9511 1 : assert_eq!(timeline.get_shard_identity().count, ShardCount(4));
9512 1 : let mut writer = timeline.writer().await;
9513 1 : writer
9514 1 : .put(
9515 1 : *TEST_KEY,
9516 1 : Lsn(0x20),
9517 1 : &Value::Image(test_img("foo at 0x20")),
9518 1 : &ctx,
9519 1 : )
9520 1 : .await?;
9521 1 : writer.finish_write(Lsn(0x20));
9522 1 : drop(writer);
9523 1 : timeline.freeze_and_flush().await.unwrap();
9524 :
9525 1 : timeline.remote_client.wait_completion().await.unwrap();
9526 1 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
9527 1 : let remote_consistent_lsn = timeline.get_remote_consistent_lsn_projected();
9528 1 : assert_eq!(Some(disk_consistent_lsn), remote_consistent_lsn);
9529 :
9530 : //
9531 : // Test
9532 : //
9533 :
9534 1 : let mut writer = timeline.writer().await;
9535 1 : writer
9536 1 : .put(
9537 1 : *TEST_KEY,
9538 1 : Lsn(0x30),
9539 1 : &Value::Image(test_img("foo at 0x30")),
9540 1 : &ctx,
9541 1 : )
9542 1 : .await?;
9543 1 : writer.finish_write(Lsn(0x30));
9544 1 : drop(writer);
9545 :
9546 1 : fail::cfg(
9547 : "flush-layer-before-update-remote-consistent-lsn",
9548 1 : "return()",
9549 : )
9550 1 : .unwrap();
9551 :
9552 1 : let flush_res = timeline.freeze_and_flush().await;
9553 : // if flush failed, the disk/remote consistent LSN should not be updated
9554 1 : assert!(flush_res.is_err());
9555 1 : assert_eq!(disk_consistent_lsn, timeline.get_disk_consistent_lsn());
9556 1 : assert_eq!(
9557 : remote_consistent_lsn,
9558 1 : timeline.get_remote_consistent_lsn_projected()
9559 : );
9560 :
9561 2 : Ok(())
9562 1 : }
9563 :
9564 : #[cfg(feature = "testing")]
9565 : #[tokio::test]
9566 1 : async fn test_simple_bottom_most_compaction_deltas_1() -> anyhow::Result<()> {
9567 2 : test_simple_bottom_most_compaction_deltas_helper(
9568 2 : "test_simple_bottom_most_compaction_deltas_1",
9569 2 : false,
9570 2 : )
9571 2 : .await
9572 1 : }
9573 :
9574 : #[cfg(feature = "testing")]
9575 : #[tokio::test]
9576 1 : async fn test_simple_bottom_most_compaction_deltas_2() -> anyhow::Result<()> {
9577 2 : test_simple_bottom_most_compaction_deltas_helper(
9578 2 : "test_simple_bottom_most_compaction_deltas_2",
9579 2 : true,
9580 2 : )
9581 2 : .await
9582 1 : }
9583 :
9584 : #[cfg(feature = "testing")]
9585 2 : async fn test_simple_bottom_most_compaction_deltas_helper(
9586 2 : test_name: &'static str,
9587 2 : use_delta_bottom_layer: bool,
9588 2 : ) -> anyhow::Result<()> {
9589 2 : let harness = TenantHarness::create(test_name).await?;
9590 2 : let (tenant, ctx) = harness.load().await;
9591 :
9592 138 : fn get_key(id: u32) -> Key {
9593 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9594 138 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9595 138 : key.field6 = id;
9596 138 : key
9597 138 : }
9598 :
9599 : // We create
9600 : // - one bottom-most image layer,
9601 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
9602 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
9603 : // - a delta layer D3 above the horizon.
9604 : //
9605 : // | D3 |
9606 : // | D1 |
9607 : // -| |-- gc horizon -----------------
9608 : // | | | D2 |
9609 : // --------- img layer ------------------
9610 : //
9611 : // What we should expact from this compaction is:
9612 : // | D3 |
9613 : // | Part of D1 |
9614 : // --------- img layer with D1+D2 at GC horizon------------------
9615 :
9616 : // img layer at 0x10
9617 2 : let img_layer = (0..10)
9618 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9619 2 : .collect_vec();
9620 : // or, delta layer at 0x10 if `use_delta_bottom_layer` is true
9621 2 : let delta4 = (0..10)
9622 20 : .map(|id| {
9623 20 : (
9624 20 : get_key(id),
9625 20 : Lsn(0x08),
9626 20 : Value::WalRecord(NeonWalRecord::wal_init(format!("value {id}@0x10"))),
9627 20 : )
9628 20 : })
9629 2 : .collect_vec();
9630 :
9631 2 : let delta1 = vec![
9632 2 : (
9633 2 : get_key(1),
9634 2 : Lsn(0x20),
9635 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9636 2 : ),
9637 2 : (
9638 2 : get_key(2),
9639 2 : Lsn(0x30),
9640 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9641 2 : ),
9642 2 : (
9643 2 : get_key(3),
9644 2 : Lsn(0x28),
9645 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9646 2 : ),
9647 2 : (
9648 2 : get_key(3),
9649 2 : Lsn(0x30),
9650 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9651 2 : ),
9652 2 : (
9653 2 : get_key(3),
9654 2 : Lsn(0x40),
9655 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9656 2 : ),
9657 : ];
9658 2 : let delta2 = vec![
9659 2 : (
9660 2 : get_key(5),
9661 2 : Lsn(0x20),
9662 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9663 2 : ),
9664 2 : (
9665 2 : get_key(6),
9666 2 : Lsn(0x20),
9667 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9668 2 : ),
9669 : ];
9670 2 : let delta3 = vec![
9671 2 : (
9672 2 : get_key(8),
9673 2 : Lsn(0x48),
9674 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9675 2 : ),
9676 2 : (
9677 2 : get_key(9),
9678 2 : Lsn(0x48),
9679 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9680 2 : ),
9681 : ];
9682 :
9683 2 : let tline = if use_delta_bottom_layer {
9684 1 : tenant
9685 1 : .create_test_timeline_with_layers(
9686 1 : TIMELINE_ID,
9687 1 : Lsn(0x08),
9688 1 : DEFAULT_PG_VERSION,
9689 1 : &ctx,
9690 1 : Vec::new(), // in-memory layers
9691 1 : vec![
9692 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9693 1 : Lsn(0x08)..Lsn(0x10),
9694 1 : delta4,
9695 1 : ),
9696 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9697 1 : Lsn(0x20)..Lsn(0x48),
9698 1 : delta1,
9699 1 : ),
9700 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9701 1 : Lsn(0x20)..Lsn(0x48),
9702 1 : delta2,
9703 1 : ),
9704 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9705 1 : Lsn(0x48)..Lsn(0x50),
9706 1 : delta3,
9707 1 : ),
9708 1 : ], // delta layers
9709 1 : vec![], // image layers
9710 1 : Lsn(0x50),
9711 1 : )
9712 1 : .await?
9713 : } else {
9714 1 : tenant
9715 1 : .create_test_timeline_with_layers(
9716 1 : TIMELINE_ID,
9717 1 : Lsn(0x10),
9718 1 : DEFAULT_PG_VERSION,
9719 1 : &ctx,
9720 1 : Vec::new(), // in-memory layers
9721 1 : vec![
9722 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9723 1 : Lsn(0x10)..Lsn(0x48),
9724 1 : delta1,
9725 1 : ),
9726 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9727 1 : Lsn(0x10)..Lsn(0x48),
9728 1 : delta2,
9729 1 : ),
9730 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9731 1 : Lsn(0x48)..Lsn(0x50),
9732 1 : delta3,
9733 1 : ),
9734 1 : ], // delta layers
9735 1 : vec![(Lsn(0x10), img_layer)], // image layers
9736 1 : Lsn(0x50),
9737 1 : )
9738 1 : .await?
9739 : };
9740 : {
9741 2 : tline
9742 2 : .applied_gc_cutoff_lsn
9743 2 : .lock_for_write()
9744 2 : .store_and_unlock(Lsn(0x30))
9745 2 : .wait()
9746 2 : .await;
9747 : // Update GC info
9748 2 : let mut guard = tline.gc_info.write().unwrap();
9749 2 : *guard = GcInfo {
9750 2 : retain_lsns: vec![],
9751 2 : cutoffs: GcCutoffs {
9752 2 : time: Some(Lsn(0x30)),
9753 2 : space: Lsn(0x30),
9754 2 : },
9755 2 : leases: Default::default(),
9756 2 : within_ancestor_pitr: false,
9757 2 : };
9758 : }
9759 :
9760 2 : let expected_result = [
9761 2 : Bytes::from_static(b"value 0@0x10"),
9762 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9763 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9764 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9765 2 : Bytes::from_static(b"value 4@0x10"),
9766 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9767 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9768 2 : Bytes::from_static(b"value 7@0x10"),
9769 2 : Bytes::from_static(b"value 8@0x10@0x48"),
9770 2 : Bytes::from_static(b"value 9@0x10@0x48"),
9771 2 : ];
9772 :
9773 2 : let expected_result_at_gc_horizon = [
9774 2 : Bytes::from_static(b"value 0@0x10"),
9775 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9776 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9777 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
9778 2 : Bytes::from_static(b"value 4@0x10"),
9779 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9780 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9781 2 : Bytes::from_static(b"value 7@0x10"),
9782 2 : Bytes::from_static(b"value 8@0x10"),
9783 2 : Bytes::from_static(b"value 9@0x10"),
9784 2 : ];
9785 :
9786 22 : for idx in 0..10 {
9787 20 : assert_eq!(
9788 20 : tline
9789 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9790 20 : .await
9791 20 : .unwrap(),
9792 20 : &expected_result[idx]
9793 : );
9794 20 : assert_eq!(
9795 20 : tline
9796 20 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
9797 20 : .await
9798 20 : .unwrap(),
9799 20 : &expected_result_at_gc_horizon[idx]
9800 : );
9801 : }
9802 :
9803 2 : let cancel = CancellationToken::new();
9804 2 : tline
9805 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9806 2 : .await
9807 2 : .unwrap();
9808 :
9809 22 : for idx in 0..10 {
9810 20 : assert_eq!(
9811 20 : tline
9812 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9813 20 : .await
9814 20 : .unwrap(),
9815 20 : &expected_result[idx]
9816 : );
9817 20 : assert_eq!(
9818 20 : tline
9819 20 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
9820 20 : .await
9821 20 : .unwrap(),
9822 20 : &expected_result_at_gc_horizon[idx]
9823 : );
9824 : }
9825 :
9826 : // increase GC horizon and compact again
9827 : {
9828 2 : tline
9829 2 : .applied_gc_cutoff_lsn
9830 2 : .lock_for_write()
9831 2 : .store_and_unlock(Lsn(0x40))
9832 2 : .wait()
9833 2 : .await;
9834 : // Update GC info
9835 2 : let mut guard = tline.gc_info.write().unwrap();
9836 2 : guard.cutoffs.time = Some(Lsn(0x40));
9837 2 : guard.cutoffs.space = Lsn(0x40);
9838 : }
9839 2 : tline
9840 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9841 2 : .await
9842 2 : .unwrap();
9843 :
9844 2 : Ok(())
9845 2 : }
9846 :
9847 : #[cfg(feature = "testing")]
9848 : #[tokio::test]
9849 1 : async fn test_generate_key_retention() -> anyhow::Result<()> {
9850 1 : let harness = TenantHarness::create("test_generate_key_retention").await?;
9851 1 : let (tenant, ctx) = harness.load().await;
9852 1 : let tline = tenant
9853 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
9854 1 : .await?;
9855 1 : tline.force_advance_lsn(Lsn(0x70));
9856 1 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
9857 1 : let history = vec![
9858 1 : (
9859 1 : key,
9860 1 : Lsn(0x10),
9861 1 : Value::WalRecord(NeonWalRecord::wal_init("0x10")),
9862 1 : ),
9863 1 : (
9864 1 : key,
9865 1 : Lsn(0x20),
9866 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9867 1 : ),
9868 1 : (
9869 1 : key,
9870 1 : Lsn(0x30),
9871 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9872 1 : ),
9873 1 : (
9874 1 : key,
9875 1 : Lsn(0x40),
9876 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9877 1 : ),
9878 1 : (
9879 1 : key,
9880 1 : Lsn(0x50),
9881 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
9882 1 : ),
9883 1 : (
9884 1 : key,
9885 1 : Lsn(0x60),
9886 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9887 1 : ),
9888 1 : (
9889 1 : key,
9890 1 : Lsn(0x70),
9891 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9892 1 : ),
9893 1 : (
9894 1 : key,
9895 1 : Lsn(0x80),
9896 1 : Value::Image(Bytes::copy_from_slice(
9897 1 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9898 1 : )),
9899 1 : ),
9900 1 : (
9901 1 : key,
9902 1 : Lsn(0x90),
9903 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9904 1 : ),
9905 : ];
9906 1 : let res = tline
9907 1 : .generate_key_retention(
9908 1 : key,
9909 1 : &history,
9910 1 : Lsn(0x60),
9911 1 : &[Lsn(0x20), Lsn(0x40), Lsn(0x50)],
9912 1 : 3,
9913 1 : None,
9914 1 : true,
9915 1 : )
9916 1 : .await
9917 1 : .unwrap();
9918 1 : let expected_res = KeyHistoryRetention {
9919 1 : below_horizon: vec![
9920 1 : (
9921 1 : Lsn(0x20),
9922 1 : KeyLogAtLsn(vec![(
9923 1 : Lsn(0x20),
9924 1 : Value::Image(Bytes::from_static(b"0x10;0x20")),
9925 1 : )]),
9926 1 : ),
9927 1 : (
9928 1 : Lsn(0x40),
9929 1 : KeyLogAtLsn(vec![
9930 1 : (
9931 1 : Lsn(0x30),
9932 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9933 1 : ),
9934 1 : (
9935 1 : Lsn(0x40),
9936 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9937 1 : ),
9938 1 : ]),
9939 1 : ),
9940 1 : (
9941 1 : Lsn(0x50),
9942 1 : KeyLogAtLsn(vec![(
9943 1 : Lsn(0x50),
9944 1 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40;0x50")),
9945 1 : )]),
9946 1 : ),
9947 1 : (
9948 1 : Lsn(0x60),
9949 1 : KeyLogAtLsn(vec![(
9950 1 : Lsn(0x60),
9951 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9952 1 : )]),
9953 1 : ),
9954 1 : ],
9955 1 : above_horizon: KeyLogAtLsn(vec![
9956 1 : (
9957 1 : Lsn(0x70),
9958 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9959 1 : ),
9960 1 : (
9961 1 : Lsn(0x80),
9962 1 : Value::Image(Bytes::copy_from_slice(
9963 1 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9964 1 : )),
9965 1 : ),
9966 1 : (
9967 1 : Lsn(0x90),
9968 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9969 1 : ),
9970 1 : ]),
9971 1 : };
9972 1 : assert_eq!(res, expected_res);
9973 :
9974 : // We expect GC-compaction to run with the original GC. This would create a situation that
9975 : // the original GC algorithm removes some delta layers b/c there are full image coverage,
9976 : // therefore causing some keys to have an incomplete history below the lowest retain LSN.
9977 : // For example, we have
9978 : // ```plain
9979 : // init delta @ 0x10, image @ 0x20, delta @ 0x30 (gc_horizon), image @ 0x40.
9980 : // ```
9981 : // Now the GC horizon moves up, and we have
9982 : // ```plain
9983 : // init delta @ 0x10, image @ 0x20, delta @ 0x30, image @ 0x40 (gc_horizon)
9984 : // ```
9985 : // The original GC algorithm kicks in, and removes delta @ 0x10, image @ 0x20.
9986 : // We will end up with
9987 : // ```plain
9988 : // delta @ 0x30, image @ 0x40 (gc_horizon)
9989 : // ```
9990 : // Now we run the GC-compaction, and this key does not have a full history.
9991 : // We should be able to handle this partial history and drop everything before the
9992 : // gc_horizon image.
9993 :
9994 1 : let history = vec![
9995 1 : (
9996 1 : key,
9997 1 : Lsn(0x20),
9998 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9999 1 : ),
10000 1 : (
10001 1 : key,
10002 1 : Lsn(0x30),
10003 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
10004 1 : ),
10005 1 : (
10006 1 : key,
10007 1 : Lsn(0x40),
10008 1 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
10009 1 : ),
10010 1 : (
10011 1 : key,
10012 1 : Lsn(0x50),
10013 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
10014 1 : ),
10015 1 : (
10016 1 : key,
10017 1 : Lsn(0x60),
10018 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
10019 1 : ),
10020 1 : (
10021 1 : key,
10022 1 : Lsn(0x70),
10023 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
10024 1 : ),
10025 1 : (
10026 1 : key,
10027 1 : Lsn(0x80),
10028 1 : Value::Image(Bytes::copy_from_slice(
10029 1 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
10030 1 : )),
10031 1 : ),
10032 1 : (
10033 1 : key,
10034 1 : Lsn(0x90),
10035 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
10036 1 : ),
10037 : ];
10038 1 : let res = tline
10039 1 : .generate_key_retention(
10040 1 : key,
10041 1 : &history,
10042 1 : Lsn(0x60),
10043 1 : &[Lsn(0x40), Lsn(0x50)],
10044 1 : 3,
10045 1 : None,
10046 1 : true,
10047 1 : )
10048 1 : .await
10049 1 : .unwrap();
10050 1 : let expected_res = KeyHistoryRetention {
10051 1 : below_horizon: vec![
10052 1 : (
10053 1 : Lsn(0x40),
10054 1 : KeyLogAtLsn(vec![(
10055 1 : Lsn(0x40),
10056 1 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
10057 1 : )]),
10058 1 : ),
10059 1 : (
10060 1 : Lsn(0x50),
10061 1 : KeyLogAtLsn(vec![(
10062 1 : Lsn(0x50),
10063 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
10064 1 : )]),
10065 1 : ),
10066 1 : (
10067 1 : Lsn(0x60),
10068 1 : KeyLogAtLsn(vec![(
10069 1 : Lsn(0x60),
10070 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
10071 1 : )]),
10072 1 : ),
10073 1 : ],
10074 1 : above_horizon: KeyLogAtLsn(vec![
10075 1 : (
10076 1 : Lsn(0x70),
10077 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
10078 1 : ),
10079 1 : (
10080 1 : Lsn(0x80),
10081 1 : Value::Image(Bytes::copy_from_slice(
10082 1 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
10083 1 : )),
10084 1 : ),
10085 1 : (
10086 1 : Lsn(0x90),
10087 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
10088 1 : ),
10089 1 : ]),
10090 1 : };
10091 1 : assert_eq!(res, expected_res);
10092 :
10093 : // In case of branch compaction, the branch itself does not have the full history, and we need to provide
10094 : // the ancestor image in the test case.
10095 :
10096 1 : let history = vec![
10097 1 : (
10098 1 : key,
10099 1 : Lsn(0x20),
10100 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
10101 1 : ),
10102 1 : (
10103 1 : key,
10104 1 : Lsn(0x30),
10105 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
10106 1 : ),
10107 1 : (
10108 1 : key,
10109 1 : Lsn(0x40),
10110 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
10111 1 : ),
10112 1 : (
10113 1 : key,
10114 1 : Lsn(0x70),
10115 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
10116 1 : ),
10117 : ];
10118 1 : let res = tline
10119 1 : .generate_key_retention(
10120 1 : key,
10121 1 : &history,
10122 1 : Lsn(0x60),
10123 1 : &[],
10124 1 : 3,
10125 1 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
10126 1 : true,
10127 1 : )
10128 1 : .await
10129 1 : .unwrap();
10130 1 : let expected_res = KeyHistoryRetention {
10131 1 : below_horizon: vec![(
10132 1 : Lsn(0x60),
10133 1 : KeyLogAtLsn(vec![(
10134 1 : Lsn(0x60),
10135 1 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")), // use the ancestor image to reconstruct the page
10136 1 : )]),
10137 1 : )],
10138 1 : above_horizon: KeyLogAtLsn(vec![(
10139 1 : Lsn(0x70),
10140 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
10141 1 : )]),
10142 1 : };
10143 1 : assert_eq!(res, expected_res);
10144 :
10145 1 : let history = vec![
10146 1 : (
10147 1 : key,
10148 1 : Lsn(0x20),
10149 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
10150 1 : ),
10151 1 : (
10152 1 : key,
10153 1 : Lsn(0x40),
10154 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
10155 1 : ),
10156 1 : (
10157 1 : key,
10158 1 : Lsn(0x60),
10159 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
10160 1 : ),
10161 1 : (
10162 1 : key,
10163 1 : Lsn(0x70),
10164 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
10165 1 : ),
10166 : ];
10167 1 : let res = tline
10168 1 : .generate_key_retention(
10169 1 : key,
10170 1 : &history,
10171 1 : Lsn(0x60),
10172 1 : &[Lsn(0x30)],
10173 1 : 3,
10174 1 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
10175 1 : true,
10176 1 : )
10177 1 : .await
10178 1 : .unwrap();
10179 1 : let expected_res = KeyHistoryRetention {
10180 1 : below_horizon: vec![
10181 1 : (
10182 1 : Lsn(0x30),
10183 1 : KeyLogAtLsn(vec![(
10184 1 : Lsn(0x20),
10185 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
10186 1 : )]),
10187 1 : ),
10188 1 : (
10189 1 : Lsn(0x60),
10190 1 : KeyLogAtLsn(vec![(
10191 1 : Lsn(0x60),
10192 1 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x40;0x60")),
10193 1 : )]),
10194 1 : ),
10195 1 : ],
10196 1 : above_horizon: KeyLogAtLsn(vec![(
10197 1 : Lsn(0x70),
10198 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
10199 1 : )]),
10200 1 : };
10201 1 : assert_eq!(res, expected_res);
10202 :
10203 2 : Ok(())
10204 1 : }
10205 :
10206 : #[cfg(feature = "testing")]
10207 : #[tokio::test]
10208 1 : async fn test_simple_bottom_most_compaction_with_retain_lsns() -> anyhow::Result<()> {
10209 1 : let harness =
10210 1 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns").await?;
10211 1 : let (tenant, ctx) = harness.load().await;
10212 :
10213 259 : fn get_key(id: u32) -> Key {
10214 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
10215 259 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
10216 259 : key.field6 = id;
10217 259 : key
10218 259 : }
10219 :
10220 1 : let img_layer = (0..10)
10221 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10222 1 : .collect_vec();
10223 :
10224 1 : let delta1 = vec![
10225 1 : (
10226 1 : get_key(1),
10227 1 : Lsn(0x20),
10228 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10229 1 : ),
10230 1 : (
10231 1 : get_key(2),
10232 1 : Lsn(0x30),
10233 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10234 1 : ),
10235 1 : (
10236 1 : get_key(3),
10237 1 : Lsn(0x28),
10238 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
10239 1 : ),
10240 1 : (
10241 1 : get_key(3),
10242 1 : Lsn(0x30),
10243 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10244 1 : ),
10245 1 : (
10246 1 : get_key(3),
10247 1 : Lsn(0x40),
10248 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
10249 1 : ),
10250 : ];
10251 1 : let delta2 = vec![
10252 1 : (
10253 1 : get_key(5),
10254 1 : Lsn(0x20),
10255 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10256 1 : ),
10257 1 : (
10258 1 : get_key(6),
10259 1 : Lsn(0x20),
10260 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10261 1 : ),
10262 : ];
10263 1 : let delta3 = vec![
10264 1 : (
10265 1 : get_key(8),
10266 1 : Lsn(0x48),
10267 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10268 1 : ),
10269 1 : (
10270 1 : get_key(9),
10271 1 : Lsn(0x48),
10272 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10273 1 : ),
10274 : ];
10275 :
10276 1 : let tline = tenant
10277 1 : .create_test_timeline_with_layers(
10278 1 : TIMELINE_ID,
10279 1 : Lsn(0x10),
10280 1 : DEFAULT_PG_VERSION,
10281 1 : &ctx,
10282 1 : Vec::new(), // in-memory layers
10283 1 : vec![
10284 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
10285 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
10286 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
10287 1 : ], // delta layers
10288 1 : vec![(Lsn(0x10), img_layer)], // image layers
10289 1 : Lsn(0x50),
10290 1 : )
10291 1 : .await?;
10292 : {
10293 1 : tline
10294 1 : .applied_gc_cutoff_lsn
10295 1 : .lock_for_write()
10296 1 : .store_and_unlock(Lsn(0x30))
10297 1 : .wait()
10298 1 : .await;
10299 : // Update GC info
10300 1 : let mut guard = tline.gc_info.write().unwrap();
10301 1 : *guard = GcInfo {
10302 1 : retain_lsns: vec![
10303 1 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
10304 1 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
10305 1 : ],
10306 1 : cutoffs: GcCutoffs {
10307 1 : time: Some(Lsn(0x30)),
10308 1 : space: Lsn(0x30),
10309 1 : },
10310 1 : leases: Default::default(),
10311 1 : within_ancestor_pitr: false,
10312 1 : };
10313 : }
10314 :
10315 1 : let expected_result = [
10316 1 : Bytes::from_static(b"value 0@0x10"),
10317 1 : Bytes::from_static(b"value 1@0x10@0x20"),
10318 1 : Bytes::from_static(b"value 2@0x10@0x30"),
10319 1 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
10320 1 : Bytes::from_static(b"value 4@0x10"),
10321 1 : Bytes::from_static(b"value 5@0x10@0x20"),
10322 1 : Bytes::from_static(b"value 6@0x10@0x20"),
10323 1 : Bytes::from_static(b"value 7@0x10"),
10324 1 : Bytes::from_static(b"value 8@0x10@0x48"),
10325 1 : Bytes::from_static(b"value 9@0x10@0x48"),
10326 1 : ];
10327 :
10328 1 : let expected_result_at_gc_horizon = [
10329 1 : Bytes::from_static(b"value 0@0x10"),
10330 1 : Bytes::from_static(b"value 1@0x10@0x20"),
10331 1 : Bytes::from_static(b"value 2@0x10@0x30"),
10332 1 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
10333 1 : Bytes::from_static(b"value 4@0x10"),
10334 1 : Bytes::from_static(b"value 5@0x10@0x20"),
10335 1 : Bytes::from_static(b"value 6@0x10@0x20"),
10336 1 : Bytes::from_static(b"value 7@0x10"),
10337 1 : Bytes::from_static(b"value 8@0x10"),
10338 1 : Bytes::from_static(b"value 9@0x10"),
10339 1 : ];
10340 :
10341 1 : let expected_result_at_lsn_20 = [
10342 1 : Bytes::from_static(b"value 0@0x10"),
10343 1 : Bytes::from_static(b"value 1@0x10@0x20"),
10344 1 : Bytes::from_static(b"value 2@0x10"),
10345 1 : Bytes::from_static(b"value 3@0x10"),
10346 1 : Bytes::from_static(b"value 4@0x10"),
10347 1 : Bytes::from_static(b"value 5@0x10@0x20"),
10348 1 : Bytes::from_static(b"value 6@0x10@0x20"),
10349 1 : Bytes::from_static(b"value 7@0x10"),
10350 1 : Bytes::from_static(b"value 8@0x10"),
10351 1 : Bytes::from_static(b"value 9@0x10"),
10352 1 : ];
10353 :
10354 1 : let expected_result_at_lsn_10 = [
10355 1 : Bytes::from_static(b"value 0@0x10"),
10356 1 : Bytes::from_static(b"value 1@0x10"),
10357 1 : Bytes::from_static(b"value 2@0x10"),
10358 1 : Bytes::from_static(b"value 3@0x10"),
10359 1 : Bytes::from_static(b"value 4@0x10"),
10360 1 : Bytes::from_static(b"value 5@0x10"),
10361 1 : Bytes::from_static(b"value 6@0x10"),
10362 1 : Bytes::from_static(b"value 7@0x10"),
10363 1 : Bytes::from_static(b"value 8@0x10"),
10364 1 : Bytes::from_static(b"value 9@0x10"),
10365 1 : ];
10366 :
10367 6 : let verify_result = || async {
10368 6 : let gc_horizon = {
10369 6 : let gc_info = tline.gc_info.read().unwrap();
10370 6 : gc_info.cutoffs.time.unwrap_or_default()
10371 : };
10372 66 : for idx in 0..10 {
10373 60 : assert_eq!(
10374 60 : tline
10375 60 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10376 60 : .await
10377 60 : .unwrap(),
10378 60 : &expected_result[idx]
10379 : );
10380 60 : assert_eq!(
10381 60 : tline
10382 60 : .get(get_key(idx as u32), gc_horizon, &ctx)
10383 60 : .await
10384 60 : .unwrap(),
10385 60 : &expected_result_at_gc_horizon[idx]
10386 : );
10387 60 : assert_eq!(
10388 60 : tline
10389 60 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
10390 60 : .await
10391 60 : .unwrap(),
10392 60 : &expected_result_at_lsn_20[idx]
10393 : );
10394 60 : assert_eq!(
10395 60 : tline
10396 60 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
10397 60 : .await
10398 60 : .unwrap(),
10399 60 : &expected_result_at_lsn_10[idx]
10400 : );
10401 : }
10402 12 : };
10403 :
10404 1 : verify_result().await;
10405 :
10406 1 : let cancel = CancellationToken::new();
10407 1 : let mut dryrun_flags = EnumSet::new();
10408 1 : dryrun_flags.insert(CompactFlags::DryRun);
10409 :
10410 1 : tline
10411 1 : .compact_with_gc(
10412 1 : &cancel,
10413 1 : CompactOptions {
10414 1 : flags: dryrun_flags,
10415 1 : ..Default::default()
10416 1 : },
10417 1 : &ctx,
10418 1 : )
10419 1 : .await
10420 1 : .unwrap();
10421 : // 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
10422 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
10423 1 : verify_result().await;
10424 :
10425 1 : tline
10426 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10427 1 : .await
10428 1 : .unwrap();
10429 1 : verify_result().await;
10430 :
10431 : // compact again
10432 1 : tline
10433 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10434 1 : .await
10435 1 : .unwrap();
10436 1 : verify_result().await;
10437 :
10438 : // increase GC horizon and compact again
10439 : {
10440 1 : tline
10441 1 : .applied_gc_cutoff_lsn
10442 1 : .lock_for_write()
10443 1 : .store_and_unlock(Lsn(0x38))
10444 1 : .wait()
10445 1 : .await;
10446 : // Update GC info
10447 1 : let mut guard = tline.gc_info.write().unwrap();
10448 1 : guard.cutoffs.time = Some(Lsn(0x38));
10449 1 : guard.cutoffs.space = Lsn(0x38);
10450 : }
10451 1 : tline
10452 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10453 1 : .await
10454 1 : .unwrap();
10455 1 : verify_result().await; // no wals between 0x30 and 0x38, so we should obtain the same result
10456 :
10457 : // not increasing the GC horizon and compact again
10458 1 : tline
10459 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10460 1 : .await
10461 1 : .unwrap();
10462 1 : verify_result().await;
10463 :
10464 2 : Ok(())
10465 1 : }
10466 :
10467 : #[cfg(feature = "testing")]
10468 : #[tokio::test]
10469 1 : async fn test_simple_bottom_most_compaction_with_retain_lsns_single_key() -> anyhow::Result<()>
10470 1 : {
10471 1 : let harness =
10472 1 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns_single_key")
10473 1 : .await?;
10474 1 : let (tenant, ctx) = harness.load().await;
10475 :
10476 176 : fn get_key(id: u32) -> Key {
10477 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
10478 176 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
10479 176 : key.field6 = id;
10480 176 : key
10481 176 : }
10482 :
10483 1 : let img_layer = (0..10)
10484 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10485 1 : .collect_vec();
10486 :
10487 1 : let delta1 = vec![
10488 1 : (
10489 1 : get_key(1),
10490 1 : Lsn(0x20),
10491 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10492 1 : ),
10493 1 : (
10494 1 : get_key(1),
10495 1 : Lsn(0x28),
10496 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
10497 1 : ),
10498 : ];
10499 1 : let delta2 = vec![
10500 1 : (
10501 1 : get_key(1),
10502 1 : Lsn(0x30),
10503 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10504 1 : ),
10505 1 : (
10506 1 : get_key(1),
10507 1 : Lsn(0x38),
10508 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
10509 1 : ),
10510 : ];
10511 1 : let delta3 = vec![
10512 1 : (
10513 1 : get_key(8),
10514 1 : Lsn(0x48),
10515 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10516 1 : ),
10517 1 : (
10518 1 : get_key(9),
10519 1 : Lsn(0x48),
10520 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10521 1 : ),
10522 : ];
10523 :
10524 1 : let tline = tenant
10525 1 : .create_test_timeline_with_layers(
10526 1 : TIMELINE_ID,
10527 1 : Lsn(0x10),
10528 1 : DEFAULT_PG_VERSION,
10529 1 : &ctx,
10530 1 : Vec::new(), // in-memory layers
10531 1 : vec![
10532 1 : // delta1 and delta 2 only contain a single key but multiple updates
10533 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x30), delta1),
10534 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
10535 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x50), delta3),
10536 1 : ], // delta layers
10537 1 : vec![(Lsn(0x10), img_layer)], // image layers
10538 1 : Lsn(0x50),
10539 1 : )
10540 1 : .await?;
10541 : {
10542 1 : tline
10543 1 : .applied_gc_cutoff_lsn
10544 1 : .lock_for_write()
10545 1 : .store_and_unlock(Lsn(0x30))
10546 1 : .wait()
10547 1 : .await;
10548 : // Update GC info
10549 1 : let mut guard = tline.gc_info.write().unwrap();
10550 1 : *guard = GcInfo {
10551 1 : retain_lsns: vec![
10552 1 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
10553 1 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
10554 1 : ],
10555 1 : cutoffs: GcCutoffs {
10556 1 : time: Some(Lsn(0x30)),
10557 1 : space: Lsn(0x30),
10558 1 : },
10559 1 : leases: Default::default(),
10560 1 : within_ancestor_pitr: false,
10561 1 : };
10562 : }
10563 :
10564 1 : let expected_result = [
10565 1 : Bytes::from_static(b"value 0@0x10"),
10566 1 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
10567 1 : Bytes::from_static(b"value 2@0x10"),
10568 1 : Bytes::from_static(b"value 3@0x10"),
10569 1 : Bytes::from_static(b"value 4@0x10"),
10570 1 : Bytes::from_static(b"value 5@0x10"),
10571 1 : Bytes::from_static(b"value 6@0x10"),
10572 1 : Bytes::from_static(b"value 7@0x10"),
10573 1 : Bytes::from_static(b"value 8@0x10@0x48"),
10574 1 : Bytes::from_static(b"value 9@0x10@0x48"),
10575 1 : ];
10576 :
10577 1 : let expected_result_at_gc_horizon = [
10578 1 : Bytes::from_static(b"value 0@0x10"),
10579 1 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
10580 1 : Bytes::from_static(b"value 2@0x10"),
10581 1 : Bytes::from_static(b"value 3@0x10"),
10582 1 : Bytes::from_static(b"value 4@0x10"),
10583 1 : Bytes::from_static(b"value 5@0x10"),
10584 1 : Bytes::from_static(b"value 6@0x10"),
10585 1 : Bytes::from_static(b"value 7@0x10"),
10586 1 : Bytes::from_static(b"value 8@0x10"),
10587 1 : Bytes::from_static(b"value 9@0x10"),
10588 1 : ];
10589 :
10590 1 : let expected_result_at_lsn_20 = [
10591 1 : Bytes::from_static(b"value 0@0x10"),
10592 1 : Bytes::from_static(b"value 1@0x10@0x20"),
10593 1 : Bytes::from_static(b"value 2@0x10"),
10594 1 : Bytes::from_static(b"value 3@0x10"),
10595 1 : Bytes::from_static(b"value 4@0x10"),
10596 1 : Bytes::from_static(b"value 5@0x10"),
10597 1 : Bytes::from_static(b"value 6@0x10"),
10598 1 : Bytes::from_static(b"value 7@0x10"),
10599 1 : Bytes::from_static(b"value 8@0x10"),
10600 1 : Bytes::from_static(b"value 9@0x10"),
10601 1 : ];
10602 :
10603 1 : let expected_result_at_lsn_10 = [
10604 1 : Bytes::from_static(b"value 0@0x10"),
10605 1 : Bytes::from_static(b"value 1@0x10"),
10606 1 : Bytes::from_static(b"value 2@0x10"),
10607 1 : Bytes::from_static(b"value 3@0x10"),
10608 1 : Bytes::from_static(b"value 4@0x10"),
10609 1 : Bytes::from_static(b"value 5@0x10"),
10610 1 : Bytes::from_static(b"value 6@0x10"),
10611 1 : Bytes::from_static(b"value 7@0x10"),
10612 1 : Bytes::from_static(b"value 8@0x10"),
10613 1 : Bytes::from_static(b"value 9@0x10"),
10614 1 : ];
10615 :
10616 4 : let verify_result = || async {
10617 4 : let gc_horizon = {
10618 4 : let gc_info = tline.gc_info.read().unwrap();
10619 4 : gc_info.cutoffs.time.unwrap_or_default()
10620 : };
10621 44 : for idx in 0..10 {
10622 40 : assert_eq!(
10623 40 : tline
10624 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10625 40 : .await
10626 40 : .unwrap(),
10627 40 : &expected_result[idx]
10628 : );
10629 40 : assert_eq!(
10630 40 : tline
10631 40 : .get(get_key(idx as u32), gc_horizon, &ctx)
10632 40 : .await
10633 40 : .unwrap(),
10634 40 : &expected_result_at_gc_horizon[idx]
10635 : );
10636 40 : assert_eq!(
10637 40 : tline
10638 40 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
10639 40 : .await
10640 40 : .unwrap(),
10641 40 : &expected_result_at_lsn_20[idx]
10642 : );
10643 40 : assert_eq!(
10644 40 : tline
10645 40 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
10646 40 : .await
10647 40 : .unwrap(),
10648 40 : &expected_result_at_lsn_10[idx]
10649 : );
10650 : }
10651 8 : };
10652 :
10653 1 : verify_result().await;
10654 :
10655 1 : let cancel = CancellationToken::new();
10656 1 : let mut dryrun_flags = EnumSet::new();
10657 1 : dryrun_flags.insert(CompactFlags::DryRun);
10658 :
10659 1 : tline
10660 1 : .compact_with_gc(
10661 1 : &cancel,
10662 1 : CompactOptions {
10663 1 : flags: dryrun_flags,
10664 1 : ..Default::default()
10665 1 : },
10666 1 : &ctx,
10667 1 : )
10668 1 : .await
10669 1 : .unwrap();
10670 : // 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
10671 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
10672 1 : verify_result().await;
10673 :
10674 1 : tline
10675 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10676 1 : .await
10677 1 : .unwrap();
10678 1 : verify_result().await;
10679 :
10680 : // compact again
10681 1 : tline
10682 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10683 1 : .await
10684 1 : .unwrap();
10685 1 : verify_result().await;
10686 :
10687 2 : Ok(())
10688 1 : }
10689 :
10690 : #[cfg(feature = "testing")]
10691 : #[tokio::test]
10692 1 : async fn test_simple_bottom_most_compaction_on_branch() -> anyhow::Result<()> {
10693 : use models::CompactLsnRange;
10694 :
10695 1 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_on_branch").await?;
10696 1 : let (tenant, ctx) = harness.load().await;
10697 :
10698 83 : fn get_key(id: u32) -> Key {
10699 83 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
10700 83 : key.field6 = id;
10701 83 : key
10702 83 : }
10703 :
10704 1 : let img_layer = (0..10)
10705 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10706 1 : .collect_vec();
10707 :
10708 1 : let delta1 = vec![
10709 1 : (
10710 1 : get_key(1),
10711 1 : Lsn(0x20),
10712 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10713 1 : ),
10714 1 : (
10715 1 : get_key(2),
10716 1 : Lsn(0x30),
10717 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10718 1 : ),
10719 1 : (
10720 1 : get_key(3),
10721 1 : Lsn(0x28),
10722 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
10723 1 : ),
10724 1 : (
10725 1 : get_key(3),
10726 1 : Lsn(0x30),
10727 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10728 1 : ),
10729 1 : (
10730 1 : get_key(3),
10731 1 : Lsn(0x40),
10732 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
10733 1 : ),
10734 : ];
10735 1 : let delta2 = vec![
10736 1 : (
10737 1 : get_key(5),
10738 1 : Lsn(0x20),
10739 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10740 1 : ),
10741 1 : (
10742 1 : get_key(6),
10743 1 : Lsn(0x20),
10744 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10745 1 : ),
10746 : ];
10747 1 : let delta3 = vec![
10748 1 : (
10749 1 : get_key(8),
10750 1 : Lsn(0x48),
10751 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10752 1 : ),
10753 1 : (
10754 1 : get_key(9),
10755 1 : Lsn(0x48),
10756 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10757 1 : ),
10758 : ];
10759 :
10760 1 : let parent_tline = tenant
10761 1 : .create_test_timeline_with_layers(
10762 1 : TIMELINE_ID,
10763 1 : Lsn(0x10),
10764 1 : DEFAULT_PG_VERSION,
10765 1 : &ctx,
10766 1 : vec![], // in-memory layers
10767 1 : vec![], // delta layers
10768 1 : vec![(Lsn(0x18), img_layer)], // image layers
10769 1 : Lsn(0x18),
10770 1 : )
10771 1 : .await?;
10772 :
10773 1 : parent_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
10774 :
10775 1 : let branch_tline = tenant
10776 1 : .branch_timeline_test_with_layers(
10777 1 : &parent_tline,
10778 1 : NEW_TIMELINE_ID,
10779 1 : Some(Lsn(0x18)),
10780 1 : &ctx,
10781 1 : vec![
10782 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
10783 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
10784 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
10785 1 : ], // delta layers
10786 1 : vec![], // image layers
10787 1 : Lsn(0x50),
10788 1 : )
10789 1 : .await?;
10790 :
10791 1 : branch_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
10792 :
10793 : {
10794 1 : parent_tline
10795 1 : .applied_gc_cutoff_lsn
10796 1 : .lock_for_write()
10797 1 : .store_and_unlock(Lsn(0x10))
10798 1 : .wait()
10799 1 : .await;
10800 : // Update GC info
10801 1 : let mut guard = parent_tline.gc_info.write().unwrap();
10802 1 : *guard = GcInfo {
10803 1 : retain_lsns: vec![(Lsn(0x18), branch_tline.timeline_id, MaybeOffloaded::No)],
10804 1 : cutoffs: GcCutoffs {
10805 1 : time: Some(Lsn(0x10)),
10806 1 : space: Lsn(0x10),
10807 1 : },
10808 1 : leases: Default::default(),
10809 1 : within_ancestor_pitr: false,
10810 1 : };
10811 : }
10812 :
10813 : {
10814 1 : branch_tline
10815 1 : .applied_gc_cutoff_lsn
10816 1 : .lock_for_write()
10817 1 : .store_and_unlock(Lsn(0x50))
10818 1 : .wait()
10819 1 : .await;
10820 : // Update GC info
10821 1 : let mut guard = branch_tline.gc_info.write().unwrap();
10822 1 : *guard = GcInfo {
10823 1 : retain_lsns: vec![(Lsn(0x40), branch_tline.timeline_id, MaybeOffloaded::No)],
10824 1 : cutoffs: GcCutoffs {
10825 1 : time: Some(Lsn(0x50)),
10826 1 : space: Lsn(0x50),
10827 1 : },
10828 1 : leases: Default::default(),
10829 1 : within_ancestor_pitr: false,
10830 1 : };
10831 : }
10832 :
10833 1 : let expected_result_at_gc_horizon = [
10834 1 : Bytes::from_static(b"value 0@0x10"),
10835 1 : Bytes::from_static(b"value 1@0x10@0x20"),
10836 1 : Bytes::from_static(b"value 2@0x10@0x30"),
10837 1 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
10838 1 : Bytes::from_static(b"value 4@0x10"),
10839 1 : Bytes::from_static(b"value 5@0x10@0x20"),
10840 1 : Bytes::from_static(b"value 6@0x10@0x20"),
10841 1 : Bytes::from_static(b"value 7@0x10"),
10842 1 : Bytes::from_static(b"value 8@0x10@0x48"),
10843 1 : Bytes::from_static(b"value 9@0x10@0x48"),
10844 1 : ];
10845 :
10846 1 : let expected_result_at_lsn_40 = [
10847 1 : Bytes::from_static(b"value 0@0x10"),
10848 1 : Bytes::from_static(b"value 1@0x10@0x20"),
10849 1 : Bytes::from_static(b"value 2@0x10@0x30"),
10850 1 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
10851 1 : Bytes::from_static(b"value 4@0x10"),
10852 1 : Bytes::from_static(b"value 5@0x10@0x20"),
10853 1 : Bytes::from_static(b"value 6@0x10@0x20"),
10854 1 : Bytes::from_static(b"value 7@0x10"),
10855 1 : Bytes::from_static(b"value 8@0x10"),
10856 1 : Bytes::from_static(b"value 9@0x10"),
10857 1 : ];
10858 :
10859 3 : let verify_result = || async {
10860 33 : for idx in 0..10 {
10861 30 : assert_eq!(
10862 30 : branch_tline
10863 30 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10864 30 : .await
10865 30 : .unwrap(),
10866 30 : &expected_result_at_gc_horizon[idx]
10867 : );
10868 30 : assert_eq!(
10869 30 : branch_tline
10870 30 : .get(get_key(idx as u32), Lsn(0x40), &ctx)
10871 30 : .await
10872 30 : .unwrap(),
10873 30 : &expected_result_at_lsn_40[idx]
10874 : );
10875 : }
10876 6 : };
10877 :
10878 1 : verify_result().await;
10879 :
10880 1 : let cancel = CancellationToken::new();
10881 1 : branch_tline
10882 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10883 1 : .await
10884 1 : .unwrap();
10885 :
10886 1 : verify_result().await;
10887 :
10888 : // Piggyback a compaction with above_lsn. Ensure it works correctly when the specified LSN intersects with the layer files.
10889 : // Now we already have a single large delta layer, so the compaction min_layer_lsn should be the same as ancestor LSN (0x18).
10890 1 : branch_tline
10891 1 : .compact_with_gc(
10892 1 : &cancel,
10893 1 : CompactOptions {
10894 1 : compact_lsn_range: Some(CompactLsnRange::above(Lsn(0x40))),
10895 1 : ..Default::default()
10896 1 : },
10897 1 : &ctx,
10898 1 : )
10899 1 : .await
10900 1 : .unwrap();
10901 :
10902 1 : verify_result().await;
10903 :
10904 2 : Ok(())
10905 1 : }
10906 :
10907 : // Regression test for https://github.com/neondatabase/neon/issues/9012
10908 : // Create an image arrangement where we have to read at different LSN ranges
10909 : // from a delta layer. This is achieved by overlapping an image layer on top of
10910 : // a delta layer. Like so:
10911 : //
10912 : // A B
10913 : // +----------------+ -> delta_layer
10914 : // | | ^ lsn
10915 : // | =========|-> nested_image_layer |
10916 : // | C | |
10917 : // +----------------+ |
10918 : // ======== -> baseline_image_layer +-------> key
10919 : //
10920 : //
10921 : // When querying the key range [A, B) we need to read at different LSN ranges
10922 : // for [A, C) and [C, B). This test checks that the described edge case is handled correctly.
10923 : #[cfg(feature = "testing")]
10924 : #[tokio::test]
10925 1 : async fn test_vectored_read_with_nested_image_layer() -> anyhow::Result<()> {
10926 1 : let harness = TenantHarness::create("test_vectored_read_with_nested_image_layer").await?;
10927 1 : let (tenant, ctx) = harness.load().await;
10928 :
10929 1 : let will_init_keys = [2, 6];
10930 22 : fn get_key(id: u32) -> Key {
10931 22 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
10932 22 : key.field6 = id;
10933 22 : key
10934 22 : }
10935 :
10936 1 : let mut expected_key_values = HashMap::new();
10937 :
10938 1 : let baseline_image_layer_lsn = Lsn(0x10);
10939 1 : let mut baseline_img_layer = Vec::new();
10940 6 : for i in 0..5 {
10941 5 : let key = get_key(i);
10942 5 : let value = format!("value {i}@{baseline_image_layer_lsn}");
10943 :
10944 5 : let removed = expected_key_values.insert(key, value.clone());
10945 5 : assert!(removed.is_none());
10946 :
10947 5 : baseline_img_layer.push((key, Bytes::from(value)));
10948 : }
10949 :
10950 1 : let nested_image_layer_lsn = Lsn(0x50);
10951 1 : let mut nested_img_layer = Vec::new();
10952 6 : for i in 5..10 {
10953 5 : let key = get_key(i);
10954 5 : let value = format!("value {i}@{nested_image_layer_lsn}");
10955 :
10956 5 : let removed = expected_key_values.insert(key, value.clone());
10957 5 : assert!(removed.is_none());
10958 :
10959 5 : nested_img_layer.push((key, Bytes::from(value)));
10960 : }
10961 :
10962 1 : let mut delta_layer_spec = Vec::default();
10963 1 : let delta_layer_start_lsn = Lsn(0x20);
10964 1 : let mut delta_layer_end_lsn = delta_layer_start_lsn;
10965 :
10966 11 : for i in 0..10 {
10967 10 : let key = get_key(i);
10968 10 : let key_in_nested = nested_img_layer
10969 10 : .iter()
10970 40 : .any(|(key_with_img, _)| *key_with_img == key);
10971 10 : let lsn = {
10972 10 : if key_in_nested {
10973 5 : Lsn(nested_image_layer_lsn.0 + 0x10)
10974 : } else {
10975 5 : delta_layer_start_lsn
10976 : }
10977 : };
10978 :
10979 10 : let will_init = will_init_keys.contains(&i);
10980 10 : if will_init {
10981 2 : delta_layer_spec.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init(""))));
10982 2 :
10983 2 : expected_key_values.insert(key, "".to_string());
10984 8 : } else {
10985 8 : let delta = format!("@{lsn}");
10986 8 : delta_layer_spec.push((
10987 8 : key,
10988 8 : lsn,
10989 8 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
10990 8 : ));
10991 8 :
10992 8 : expected_key_values
10993 8 : .get_mut(&key)
10994 8 : .expect("An image exists for each key")
10995 8 : .push_str(delta.as_str());
10996 8 : }
10997 10 : delta_layer_end_lsn = std::cmp::max(delta_layer_start_lsn, lsn);
10998 : }
10999 :
11000 1 : delta_layer_end_lsn = Lsn(delta_layer_end_lsn.0 + 1);
11001 :
11002 1 : assert!(
11003 1 : nested_image_layer_lsn > delta_layer_start_lsn
11004 1 : && nested_image_layer_lsn < delta_layer_end_lsn
11005 : );
11006 :
11007 1 : let tline = tenant
11008 1 : .create_test_timeline_with_layers(
11009 1 : TIMELINE_ID,
11010 1 : baseline_image_layer_lsn,
11011 1 : DEFAULT_PG_VERSION,
11012 1 : &ctx,
11013 1 : vec![], // in-memory layers
11014 1 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
11015 1 : delta_layer_start_lsn..delta_layer_end_lsn,
11016 1 : delta_layer_spec,
11017 1 : )], // delta layers
11018 1 : vec![
11019 1 : (baseline_image_layer_lsn, baseline_img_layer),
11020 1 : (nested_image_layer_lsn, nested_img_layer),
11021 1 : ], // image layers
11022 1 : delta_layer_end_lsn,
11023 1 : )
11024 1 : .await?;
11025 :
11026 1 : let query = VersionedKeySpaceQuery::uniform(
11027 1 : KeySpace::single(get_key(0)..get_key(10)),
11028 1 : delta_layer_end_lsn,
11029 : );
11030 :
11031 1 : let results = tline
11032 1 : .get_vectored(query, IoConcurrency::sequential(), &ctx)
11033 1 : .await
11034 1 : .expect("No vectored errors");
11035 11 : for (key, res) in results {
11036 10 : let value = res.expect("No key errors");
11037 10 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
11038 10 : assert_eq!(value, Bytes::from(expected_value));
11039 1 : }
11040 1 :
11041 1 : Ok(())
11042 1 : }
11043 :
11044 : #[cfg(feature = "testing")]
11045 : #[tokio::test]
11046 1 : async fn test_vectored_read_with_image_layer_inside_inmem() -> anyhow::Result<()> {
11047 1 : let harness =
11048 1 : TenantHarness::create("test_vectored_read_with_image_layer_inside_inmem").await?;
11049 1 : let (tenant, ctx) = harness.load().await;
11050 :
11051 1 : let will_init_keys = [2, 6];
11052 32 : fn get_key(id: u32) -> Key {
11053 32 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
11054 32 : key.field6 = id;
11055 32 : key
11056 32 : }
11057 :
11058 1 : let mut expected_key_values = HashMap::new();
11059 :
11060 1 : let baseline_image_layer_lsn = Lsn(0x10);
11061 1 : let mut baseline_img_layer = Vec::new();
11062 6 : for i in 0..5 {
11063 5 : let key = get_key(i);
11064 5 : let value = format!("value {i}@{baseline_image_layer_lsn}");
11065 :
11066 5 : let removed = expected_key_values.insert(key, value.clone());
11067 5 : assert!(removed.is_none());
11068 :
11069 5 : baseline_img_layer.push((key, Bytes::from(value)));
11070 : }
11071 :
11072 1 : let nested_image_layer_lsn = Lsn(0x50);
11073 1 : let mut nested_img_layer = Vec::new();
11074 6 : for i in 5..10 {
11075 5 : let key = get_key(i);
11076 5 : let value = format!("value {i}@{nested_image_layer_lsn}");
11077 :
11078 5 : let removed = expected_key_values.insert(key, value.clone());
11079 5 : assert!(removed.is_none());
11080 :
11081 5 : nested_img_layer.push((key, Bytes::from(value)));
11082 : }
11083 :
11084 1 : let frozen_layer = {
11085 1 : let lsn_range = Lsn(0x40)..Lsn(0x60);
11086 1 : let mut data = Vec::new();
11087 11 : for i in 0..10 {
11088 10 : let key = get_key(i);
11089 10 : let key_in_nested = nested_img_layer
11090 10 : .iter()
11091 40 : .any(|(key_with_img, _)| *key_with_img == key);
11092 10 : let lsn = {
11093 10 : if key_in_nested {
11094 5 : Lsn(nested_image_layer_lsn.0 + 5)
11095 : } else {
11096 5 : lsn_range.start
11097 : }
11098 : };
11099 :
11100 10 : let will_init = will_init_keys.contains(&i);
11101 10 : if will_init {
11102 2 : data.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init(""))));
11103 2 :
11104 2 : expected_key_values.insert(key, "".to_string());
11105 8 : } else {
11106 8 : let delta = format!("@{lsn}");
11107 8 : data.push((
11108 8 : key,
11109 8 : lsn,
11110 8 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
11111 8 : ));
11112 8 :
11113 8 : expected_key_values
11114 8 : .get_mut(&key)
11115 8 : .expect("An image exists for each key")
11116 8 : .push_str(delta.as_str());
11117 8 : }
11118 : }
11119 :
11120 1 : InMemoryLayerTestDesc {
11121 1 : lsn_range,
11122 1 : is_open: false,
11123 1 : data,
11124 1 : }
11125 : };
11126 :
11127 1 : let (open_layer, last_record_lsn) = {
11128 1 : let start_lsn = Lsn(0x70);
11129 1 : let mut data = Vec::new();
11130 1 : let mut end_lsn = Lsn(0);
11131 11 : for i in 0..10 {
11132 10 : let key = get_key(i);
11133 10 : let lsn = Lsn(start_lsn.0 + i as u64);
11134 10 : let delta = format!("@{lsn}");
11135 10 : data.push((
11136 10 : key,
11137 10 : lsn,
11138 10 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
11139 10 : ));
11140 10 :
11141 10 : expected_key_values
11142 10 : .get_mut(&key)
11143 10 : .expect("An image exists for each key")
11144 10 : .push_str(delta.as_str());
11145 10 :
11146 10 : end_lsn = std::cmp::max(end_lsn, lsn);
11147 10 : }
11148 :
11149 1 : (
11150 1 : InMemoryLayerTestDesc {
11151 1 : lsn_range: start_lsn..Lsn::MAX,
11152 1 : is_open: true,
11153 1 : data,
11154 1 : },
11155 1 : end_lsn,
11156 1 : )
11157 : };
11158 :
11159 1 : assert!(
11160 1 : nested_image_layer_lsn > frozen_layer.lsn_range.start
11161 1 : && nested_image_layer_lsn < frozen_layer.lsn_range.end
11162 : );
11163 :
11164 1 : let tline = tenant
11165 1 : .create_test_timeline_with_layers(
11166 1 : TIMELINE_ID,
11167 1 : baseline_image_layer_lsn,
11168 1 : DEFAULT_PG_VERSION,
11169 1 : &ctx,
11170 1 : vec![open_layer, frozen_layer], // in-memory layers
11171 1 : Vec::new(), // delta layers
11172 1 : vec![
11173 1 : (baseline_image_layer_lsn, baseline_img_layer),
11174 1 : (nested_image_layer_lsn, nested_img_layer),
11175 1 : ], // image layers
11176 1 : last_record_lsn,
11177 1 : )
11178 1 : .await?;
11179 :
11180 1 : let query = VersionedKeySpaceQuery::uniform(
11181 1 : KeySpace::single(get_key(0)..get_key(10)),
11182 1 : last_record_lsn,
11183 : );
11184 :
11185 1 : let results = tline
11186 1 : .get_vectored(query, IoConcurrency::sequential(), &ctx)
11187 1 : .await
11188 1 : .expect("No vectored errors");
11189 11 : for (key, res) in results {
11190 10 : let value = res.expect("No key errors");
11191 10 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
11192 10 : assert_eq!(value, Bytes::from(expected_value.clone()));
11193 1 :
11194 10 : tracing::info!("key={key} value={expected_value}");
11195 1 : }
11196 1 :
11197 1 : Ok(())
11198 1 : }
11199 :
11200 : // A randomized read path test. Generates a layer map according to a deterministic
11201 : // specification. Fills the (key, LSN) space in random manner and then performs
11202 : // random scattered queries validating the results against in-memory storage.
11203 : //
11204 : // See this internal Notion page for a diagram of the layer map:
11205 : // https://www.notion.so/neondatabase/Read-Path-Unit-Testing-Fuzzing-1d1f189e0047806c8e5cd37781b0a350?pvs=4
11206 : //
11207 : // A fuzzing mode is also supported. In this mode, the test will use a random
11208 : // seed instead of a hardcoded one. Use it in conjunction with `cargo stress`
11209 : // to run multiple instances in parallel:
11210 : //
11211 : // $ RUST_BACKTRACE=1 RUST_LOG=INFO \
11212 : // cargo stress --package=pageserver --features=testing,fuzz-read-path --release -- test_read_path
11213 : #[cfg(feature = "testing")]
11214 : #[tokio::test]
11215 1 : async fn test_read_path() -> anyhow::Result<()> {
11216 : use rand::seq::SliceRandom;
11217 :
11218 1 : let seed = if cfg!(feature = "fuzz-read-path") {
11219 0 : let seed: u64 = thread_rng().r#gen();
11220 0 : seed
11221 : } else {
11222 : // Use a hard-coded seed when not in fuzzing mode.
11223 : // Note that with the current approach results are not reproducible
11224 : // accross platforms and Rust releases.
11225 : const SEED: u64 = 0;
11226 1 : SEED
11227 : };
11228 :
11229 1 : let mut random = StdRng::seed_from_u64(seed);
11230 :
11231 1 : let (queries, will_init_chance, gap_chance) = if cfg!(feature = "fuzz-read-path") {
11232 : const QUERIES: u64 = 5000;
11233 0 : let will_init_chance: u8 = random.gen_range(0..=10);
11234 0 : let gap_chance: u8 = random.gen_range(0..=50);
11235 :
11236 0 : (QUERIES, will_init_chance, gap_chance)
11237 : } else {
11238 : const QUERIES: u64 = 1000;
11239 : const WILL_INIT_CHANCE: u8 = 1;
11240 : const GAP_CHANCE: u8 = 5;
11241 :
11242 1 : (QUERIES, WILL_INIT_CHANCE, GAP_CHANCE)
11243 : };
11244 :
11245 1 : let harness = TenantHarness::create("test_read_path").await?;
11246 1 : let (tenant, ctx) = harness.load().await;
11247 :
11248 1 : tracing::info!("Using random seed: {seed}");
11249 1 : tracing::info!(%will_init_chance, %gap_chance, "Fill params");
11250 :
11251 : // Define the layer map shape. Note that this part is not randomized.
11252 :
11253 : const KEY_DIMENSION_SIZE: u32 = 99;
11254 1 : let start_key = Key::from_hex("110000000033333333444444445500000000").unwrap();
11255 1 : let end_key = start_key.add(KEY_DIMENSION_SIZE);
11256 1 : let total_key_range = start_key..end_key;
11257 1 : let total_key_range_size = end_key.to_i128() - start_key.to_i128();
11258 1 : let total_start_lsn = Lsn(104);
11259 1 : let last_record_lsn = Lsn(504);
11260 :
11261 1 : assert!(total_key_range_size % 3 == 0);
11262 :
11263 1 : let in_memory_layers_shape = vec![
11264 1 : (total_key_range.clone(), Lsn(304)..Lsn(400)),
11265 1 : (total_key_range.clone(), Lsn(400)..last_record_lsn),
11266 : ];
11267 :
11268 1 : let delta_layers_shape = vec![
11269 1 : (
11270 1 : start_key..(start_key.add((total_key_range_size / 3) as u32)),
11271 1 : Lsn(200)..Lsn(304),
11272 1 : ),
11273 1 : (
11274 1 : (start_key.add((total_key_range_size / 3) as u32))
11275 1 : ..(start_key.add((total_key_range_size * 2 / 3) as u32)),
11276 1 : Lsn(200)..Lsn(304),
11277 1 : ),
11278 1 : (
11279 1 : (start_key.add((total_key_range_size * 2 / 3) as u32))
11280 1 : ..(start_key.add(total_key_range_size as u32)),
11281 1 : Lsn(200)..Lsn(304),
11282 1 : ),
11283 : ];
11284 :
11285 1 : let image_layers_shape = vec![
11286 1 : (
11287 1 : start_key.add((total_key_range_size * 2 / 3 - 10) as u32)
11288 1 : ..start_key.add((total_key_range_size * 2 / 3 + 10) as u32),
11289 1 : Lsn(456),
11290 1 : ),
11291 1 : (
11292 1 : start_key.add((total_key_range_size / 3 - 10) as u32)
11293 1 : ..start_key.add((total_key_range_size / 3 + 10) as u32),
11294 1 : Lsn(256),
11295 1 : ),
11296 1 : (total_key_range.clone(), total_start_lsn),
11297 : ];
11298 :
11299 1 : let specification = TestTimelineSpecification {
11300 1 : start_lsn: total_start_lsn,
11301 1 : last_record_lsn,
11302 1 : in_memory_layers_shape,
11303 1 : delta_layers_shape,
11304 1 : image_layers_shape,
11305 1 : gap_chance,
11306 1 : will_init_chance,
11307 1 : };
11308 :
11309 : // Create and randomly fill in the layers according to the specification
11310 1 : let (tline, storage, interesting_lsns) = randomize_timeline(
11311 1 : &tenant,
11312 1 : TIMELINE_ID,
11313 1 : DEFAULT_PG_VERSION,
11314 1 : specification,
11315 1 : &mut random,
11316 1 : &ctx,
11317 1 : )
11318 1 : .await?;
11319 :
11320 : // Now generate queries based on the interesting lsns that we've collected.
11321 : //
11322 : // While there's still room in the query, pick and interesting LSN and a random
11323 : // key. Then roll the dice to see if the next key should also be included in
11324 : // the query. When the roll fails, break the "batch" and pick another point in the
11325 : // (key, LSN) space.
11326 :
11327 : const PICK_NEXT_CHANCE: u8 = 50;
11328 1 : for _ in 0..queries {
11329 1000 : let query = {
11330 1000 : let mut keyspaces_at_lsn: HashMap<Lsn, KeySpaceRandomAccum> = HashMap::default();
11331 1000 : let mut used_keys: HashSet<Key> = HashSet::default();
11332 1 :
11333 22536 : while used_keys.len() < tenant.conf.max_get_vectored_keys.get() {
11334 21536 : let selected_lsn = interesting_lsns.choose(&mut random).expect("not empty");
11335 21536 : let mut selected_key = start_key.add(random.gen_range(0..KEY_DIMENSION_SIZE));
11336 1 :
11337 37614 : while used_keys.len() < tenant.conf.max_get_vectored_keys.get() {
11338 37093 : if used_keys.contains(&selected_key)
11339 32154 : || selected_key >= start_key.add(KEY_DIMENSION_SIZE)
11340 1 : {
11341 5093 : break;
11342 32000 : }
11343 1 :
11344 32000 : keyspaces_at_lsn
11345 32000 : .entry(*selected_lsn)
11346 32000 : .or_default()
11347 32000 : .add_key(selected_key);
11348 32000 : used_keys.insert(selected_key);
11349 1 :
11350 32000 : let pick_next = random.gen_range(0..=100) <= PICK_NEXT_CHANCE;
11351 32000 : if pick_next {
11352 16078 : selected_key = selected_key.next();
11353 16078 : } else {
11354 15922 : break;
11355 1 : }
11356 1 : }
11357 1 : }
11358 1 :
11359 1000 : VersionedKeySpaceQuery::scattered(
11360 1000 : keyspaces_at_lsn
11361 1000 : .into_iter()
11362 11917 : .map(|(lsn, acc)| (lsn, acc.to_keyspace()))
11363 1000 : .collect(),
11364 1 : )
11365 1 : };
11366 1 :
11367 1 : // Run the query and validate the results
11368 1 :
11369 1000 : let results = tline
11370 1000 : .get_vectored(query.clone(), IoConcurrency::Sequential, &ctx)
11371 1000 : .await;
11372 1 :
11373 1000 : let blobs = match results {
11374 1000 : Ok(ok) => ok,
11375 1 : Err(err) => {
11376 1 : panic!("seed={seed} Error returned for query {query}: {err}");
11377 1 : }
11378 1 : };
11379 1 :
11380 32000 : for (key, key_res) in blobs.into_iter() {
11381 32000 : match key_res {
11382 32000 : Ok(blob) => {
11383 32000 : let requested_at_lsn = query.map_key_to_lsn(&key);
11384 32000 : let expected = storage.get(key, requested_at_lsn);
11385 1 :
11386 32000 : if blob != expected {
11387 1 : tracing::error!(
11388 1 : "seed={seed} Mismatch for {key}@{requested_at_lsn} from query: {query}"
11389 1 : );
11390 32000 : }
11391 1 :
11392 32000 : assert_eq!(blob, expected);
11393 1 : }
11394 1 : Err(err) => {
11395 1 : let requested_at_lsn = query.map_key_to_lsn(&key);
11396 1 :
11397 1 : panic!(
11398 1 : "seed={seed} Error returned for {key}@{requested_at_lsn} from query {query}: {err}"
11399 1 : );
11400 1 : }
11401 1 : }
11402 1 : }
11403 1 : }
11404 1 :
11405 1 : Ok(())
11406 1 : }
11407 :
11408 107 : fn sort_layer_key(k1: &PersistentLayerKey, k2: &PersistentLayerKey) -> std::cmp::Ordering {
11409 107 : (
11410 107 : k1.is_delta,
11411 107 : k1.key_range.start,
11412 107 : k1.key_range.end,
11413 107 : k1.lsn_range.start,
11414 107 : k1.lsn_range.end,
11415 107 : )
11416 107 : .cmp(&(
11417 107 : k2.is_delta,
11418 107 : k2.key_range.start,
11419 107 : k2.key_range.end,
11420 107 : k2.lsn_range.start,
11421 107 : k2.lsn_range.end,
11422 107 : ))
11423 107 : }
11424 :
11425 12 : async fn inspect_and_sort(
11426 12 : tline: &Arc<Timeline>,
11427 12 : filter: Option<std::ops::Range<Key>>,
11428 12 : ) -> Vec<PersistentLayerKey> {
11429 12 : let mut all_layers = tline.inspect_historic_layers().await.unwrap();
11430 12 : if let Some(filter) = filter {
11431 54 : all_layers.retain(|layer| overlaps_with(&layer.key_range, &filter));
11432 1 : }
11433 12 : all_layers.sort_by(sort_layer_key);
11434 12 : all_layers
11435 12 : }
11436 :
11437 : #[cfg(feature = "testing")]
11438 11 : fn check_layer_map_key_eq(
11439 11 : mut left: Vec<PersistentLayerKey>,
11440 11 : mut right: Vec<PersistentLayerKey>,
11441 11 : ) {
11442 11 : left.sort_by(sort_layer_key);
11443 11 : right.sort_by(sort_layer_key);
11444 11 : if left != right {
11445 0 : eprintln!("---LEFT---");
11446 0 : for left in left.iter() {
11447 0 : eprintln!("{left}");
11448 0 : }
11449 0 : eprintln!("---RIGHT---");
11450 0 : for right in right.iter() {
11451 0 : eprintln!("{right}");
11452 0 : }
11453 0 : assert_eq!(left, right);
11454 11 : }
11455 11 : }
11456 :
11457 : #[cfg(feature = "testing")]
11458 : #[tokio::test]
11459 1 : async fn test_simple_partial_bottom_most_compaction() -> anyhow::Result<()> {
11460 1 : let harness = TenantHarness::create("test_simple_partial_bottom_most_compaction").await?;
11461 1 : let (tenant, ctx) = harness.load().await;
11462 :
11463 91 : fn get_key(id: u32) -> Key {
11464 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
11465 91 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
11466 91 : key.field6 = id;
11467 91 : key
11468 91 : }
11469 :
11470 : // img layer at 0x10
11471 1 : let img_layer = (0..10)
11472 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
11473 1 : .collect_vec();
11474 :
11475 1 : let delta1 = vec![
11476 1 : (
11477 1 : get_key(1),
11478 1 : Lsn(0x20),
11479 1 : Value::Image(Bytes::from("value 1@0x20")),
11480 1 : ),
11481 1 : (
11482 1 : get_key(2),
11483 1 : Lsn(0x30),
11484 1 : Value::Image(Bytes::from("value 2@0x30")),
11485 1 : ),
11486 1 : (
11487 1 : get_key(3),
11488 1 : Lsn(0x40),
11489 1 : Value::Image(Bytes::from("value 3@0x40")),
11490 1 : ),
11491 : ];
11492 1 : let delta2 = vec![
11493 1 : (
11494 1 : get_key(5),
11495 1 : Lsn(0x20),
11496 1 : Value::Image(Bytes::from("value 5@0x20")),
11497 1 : ),
11498 1 : (
11499 1 : get_key(6),
11500 1 : Lsn(0x20),
11501 1 : Value::Image(Bytes::from("value 6@0x20")),
11502 1 : ),
11503 : ];
11504 1 : let delta3 = vec![
11505 1 : (
11506 1 : get_key(8),
11507 1 : Lsn(0x48),
11508 1 : Value::Image(Bytes::from("value 8@0x48")),
11509 1 : ),
11510 1 : (
11511 1 : get_key(9),
11512 1 : Lsn(0x48),
11513 1 : Value::Image(Bytes::from("value 9@0x48")),
11514 1 : ),
11515 : ];
11516 :
11517 1 : let tline = tenant
11518 1 : .create_test_timeline_with_layers(
11519 1 : TIMELINE_ID,
11520 1 : Lsn(0x10),
11521 1 : DEFAULT_PG_VERSION,
11522 1 : &ctx,
11523 1 : vec![], // in-memory layers
11524 1 : vec![
11525 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
11526 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
11527 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
11528 1 : ], // delta layers
11529 1 : vec![(Lsn(0x10), img_layer)], // image layers
11530 1 : Lsn(0x50),
11531 1 : )
11532 1 : .await?;
11533 :
11534 : {
11535 1 : tline
11536 1 : .applied_gc_cutoff_lsn
11537 1 : .lock_for_write()
11538 1 : .store_and_unlock(Lsn(0x30))
11539 1 : .wait()
11540 1 : .await;
11541 : // Update GC info
11542 1 : let mut guard = tline.gc_info.write().unwrap();
11543 1 : *guard = GcInfo {
11544 1 : retain_lsns: vec![(Lsn(0x20), tline.timeline_id, MaybeOffloaded::No)],
11545 1 : cutoffs: GcCutoffs {
11546 1 : time: Some(Lsn(0x30)),
11547 1 : space: Lsn(0x30),
11548 1 : },
11549 1 : leases: Default::default(),
11550 1 : within_ancestor_pitr: false,
11551 1 : };
11552 : }
11553 :
11554 1 : let cancel = CancellationToken::new();
11555 :
11556 : // Do a partial compaction on key range 0..2
11557 1 : tline
11558 1 : .compact_with_gc(
11559 1 : &cancel,
11560 1 : CompactOptions {
11561 1 : flags: EnumSet::new(),
11562 1 : compact_key_range: Some((get_key(0)..get_key(2)).into()),
11563 1 : ..Default::default()
11564 1 : },
11565 1 : &ctx,
11566 1 : )
11567 1 : .await
11568 1 : .unwrap();
11569 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11570 1 : check_layer_map_key_eq(
11571 1 : all_layers,
11572 1 : vec![
11573 : // newly-generated image layer for the partial compaction range 0-2
11574 1 : PersistentLayerKey {
11575 1 : key_range: get_key(0)..get_key(2),
11576 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11577 1 : is_delta: false,
11578 1 : },
11579 1 : PersistentLayerKey {
11580 1 : key_range: get_key(0)..get_key(10),
11581 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
11582 1 : is_delta: false,
11583 1 : },
11584 : // delta1 is split and the second part is rewritten
11585 1 : PersistentLayerKey {
11586 1 : key_range: get_key(2)..get_key(4),
11587 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11588 1 : is_delta: true,
11589 1 : },
11590 1 : PersistentLayerKey {
11591 1 : key_range: get_key(5)..get_key(7),
11592 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11593 1 : is_delta: true,
11594 1 : },
11595 1 : PersistentLayerKey {
11596 1 : key_range: get_key(8)..get_key(10),
11597 1 : lsn_range: Lsn(0x48)..Lsn(0x50),
11598 1 : is_delta: true,
11599 1 : },
11600 : ],
11601 : );
11602 :
11603 : // Do a partial compaction on key range 2..4
11604 1 : tline
11605 1 : .compact_with_gc(
11606 1 : &cancel,
11607 1 : CompactOptions {
11608 1 : flags: EnumSet::new(),
11609 1 : compact_key_range: Some((get_key(2)..get_key(4)).into()),
11610 1 : ..Default::default()
11611 1 : },
11612 1 : &ctx,
11613 1 : )
11614 1 : .await
11615 1 : .unwrap();
11616 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11617 1 : check_layer_map_key_eq(
11618 1 : all_layers,
11619 1 : vec![
11620 1 : PersistentLayerKey {
11621 1 : key_range: get_key(0)..get_key(2),
11622 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11623 1 : is_delta: false,
11624 1 : },
11625 1 : PersistentLayerKey {
11626 1 : key_range: get_key(0)..get_key(10),
11627 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
11628 1 : is_delta: false,
11629 1 : },
11630 : // image layer generated for the compaction range 2-4
11631 1 : PersistentLayerKey {
11632 1 : key_range: get_key(2)..get_key(4),
11633 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11634 1 : is_delta: false,
11635 1 : },
11636 : // we have key2/key3 above the retain_lsn, so we still need this delta layer
11637 1 : PersistentLayerKey {
11638 1 : key_range: get_key(2)..get_key(4),
11639 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11640 1 : is_delta: true,
11641 1 : },
11642 1 : PersistentLayerKey {
11643 1 : key_range: get_key(5)..get_key(7),
11644 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11645 1 : is_delta: true,
11646 1 : },
11647 1 : PersistentLayerKey {
11648 1 : key_range: get_key(8)..get_key(10),
11649 1 : lsn_range: Lsn(0x48)..Lsn(0x50),
11650 1 : is_delta: true,
11651 1 : },
11652 : ],
11653 : );
11654 :
11655 : // Do a partial compaction on key range 4..9
11656 1 : tline
11657 1 : .compact_with_gc(
11658 1 : &cancel,
11659 1 : CompactOptions {
11660 1 : flags: EnumSet::new(),
11661 1 : compact_key_range: Some((get_key(4)..get_key(9)).into()),
11662 1 : ..Default::default()
11663 1 : },
11664 1 : &ctx,
11665 1 : )
11666 1 : .await
11667 1 : .unwrap();
11668 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11669 1 : check_layer_map_key_eq(
11670 1 : all_layers,
11671 1 : vec![
11672 1 : PersistentLayerKey {
11673 1 : key_range: get_key(0)..get_key(2),
11674 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11675 1 : is_delta: false,
11676 1 : },
11677 1 : PersistentLayerKey {
11678 1 : key_range: get_key(0)..get_key(10),
11679 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
11680 1 : is_delta: false,
11681 1 : },
11682 1 : PersistentLayerKey {
11683 1 : key_range: get_key(2)..get_key(4),
11684 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11685 1 : is_delta: false,
11686 1 : },
11687 1 : PersistentLayerKey {
11688 1 : key_range: get_key(2)..get_key(4),
11689 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11690 1 : is_delta: true,
11691 1 : },
11692 : // image layer generated for this compaction range
11693 1 : PersistentLayerKey {
11694 1 : key_range: get_key(4)..get_key(9),
11695 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11696 1 : is_delta: false,
11697 1 : },
11698 1 : PersistentLayerKey {
11699 1 : key_range: get_key(8)..get_key(10),
11700 1 : lsn_range: Lsn(0x48)..Lsn(0x50),
11701 1 : is_delta: true,
11702 1 : },
11703 : ],
11704 : );
11705 :
11706 : // Do a partial compaction on key range 9..10
11707 1 : tline
11708 1 : .compact_with_gc(
11709 1 : &cancel,
11710 1 : CompactOptions {
11711 1 : flags: EnumSet::new(),
11712 1 : compact_key_range: Some((get_key(9)..get_key(10)).into()),
11713 1 : ..Default::default()
11714 1 : },
11715 1 : &ctx,
11716 1 : )
11717 1 : .await
11718 1 : .unwrap();
11719 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11720 1 : check_layer_map_key_eq(
11721 1 : all_layers,
11722 1 : vec![
11723 1 : PersistentLayerKey {
11724 1 : key_range: get_key(0)..get_key(2),
11725 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11726 1 : is_delta: false,
11727 1 : },
11728 1 : PersistentLayerKey {
11729 1 : key_range: get_key(0)..get_key(10),
11730 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
11731 1 : is_delta: false,
11732 1 : },
11733 1 : PersistentLayerKey {
11734 1 : key_range: get_key(2)..get_key(4),
11735 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11736 1 : is_delta: false,
11737 1 : },
11738 1 : PersistentLayerKey {
11739 1 : key_range: get_key(2)..get_key(4),
11740 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11741 1 : is_delta: true,
11742 1 : },
11743 1 : PersistentLayerKey {
11744 1 : key_range: get_key(4)..get_key(9),
11745 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11746 1 : is_delta: false,
11747 1 : },
11748 : // image layer generated for the compaction range
11749 1 : PersistentLayerKey {
11750 1 : key_range: get_key(9)..get_key(10),
11751 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11752 1 : is_delta: false,
11753 1 : },
11754 1 : PersistentLayerKey {
11755 1 : key_range: get_key(8)..get_key(10),
11756 1 : lsn_range: Lsn(0x48)..Lsn(0x50),
11757 1 : is_delta: true,
11758 1 : },
11759 : ],
11760 : );
11761 :
11762 : // Do a partial compaction on key range 0..10, all image layers below LSN 20 can be replaced with new ones.
11763 1 : tline
11764 1 : .compact_with_gc(
11765 1 : &cancel,
11766 1 : CompactOptions {
11767 1 : flags: EnumSet::new(),
11768 1 : compact_key_range: Some((get_key(0)..get_key(10)).into()),
11769 1 : ..Default::default()
11770 1 : },
11771 1 : &ctx,
11772 1 : )
11773 1 : .await
11774 1 : .unwrap();
11775 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11776 1 : check_layer_map_key_eq(
11777 1 : all_layers,
11778 1 : vec![
11779 : // aha, we removed all unnecessary image/delta layers and got a very clean layer map!
11780 1 : PersistentLayerKey {
11781 1 : key_range: get_key(0)..get_key(10),
11782 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11783 1 : is_delta: false,
11784 1 : },
11785 1 : PersistentLayerKey {
11786 1 : key_range: get_key(2)..get_key(4),
11787 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11788 1 : is_delta: true,
11789 1 : },
11790 1 : PersistentLayerKey {
11791 1 : key_range: get_key(8)..get_key(10),
11792 1 : lsn_range: Lsn(0x48)..Lsn(0x50),
11793 1 : is_delta: true,
11794 1 : },
11795 : ],
11796 : );
11797 2 : Ok(())
11798 1 : }
11799 :
11800 : #[cfg(feature = "testing")]
11801 : #[tokio::test]
11802 1 : async fn test_timeline_offload_retain_lsn() -> anyhow::Result<()> {
11803 1 : let harness = TenantHarness::create("test_timeline_offload_retain_lsn")
11804 1 : .await
11805 1 : .unwrap();
11806 1 : let (tenant, ctx) = harness.load().await;
11807 1 : let tline_parent = tenant
11808 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
11809 1 : .await
11810 1 : .unwrap();
11811 1 : let tline_child = tenant
11812 1 : .branch_timeline_test(&tline_parent, NEW_TIMELINE_ID, Some(Lsn(0x20)), &ctx)
11813 1 : .await
11814 1 : .unwrap();
11815 : {
11816 1 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
11817 1 : assert_eq!(
11818 1 : gc_info_parent.retain_lsns,
11819 1 : vec![(Lsn(0x20), tline_child.timeline_id, MaybeOffloaded::No)]
11820 : );
11821 : }
11822 : // We have to directly call the remote_client instead of using the archive function to avoid constructing broker client...
11823 1 : tline_child
11824 1 : .remote_client
11825 1 : .schedule_index_upload_for_timeline_archival_state(TimelineArchivalState::Archived)
11826 1 : .unwrap();
11827 1 : tline_child.remote_client.wait_completion().await.unwrap();
11828 1 : offload_timeline(&tenant, &tline_child)
11829 1 : .instrument(tracing::info_span!(parent: None, "offload_test", tenant_id=%"test", shard_id=%"test", timeline_id=%"test"))
11830 1 : .await.unwrap();
11831 1 : let child_timeline_id = tline_child.timeline_id;
11832 1 : Arc::try_unwrap(tline_child).unwrap();
11833 :
11834 : {
11835 1 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
11836 1 : assert_eq!(
11837 1 : gc_info_parent.retain_lsns,
11838 1 : vec![(Lsn(0x20), child_timeline_id, MaybeOffloaded::Yes)]
11839 : );
11840 : }
11841 :
11842 1 : tenant
11843 1 : .get_offloaded_timeline(child_timeline_id)
11844 1 : .unwrap()
11845 1 : .defuse_for_tenant_drop();
11846 :
11847 2 : Ok(())
11848 1 : }
11849 :
11850 : #[cfg(feature = "testing")]
11851 : #[tokio::test]
11852 1 : async fn test_simple_bottom_most_compaction_above_lsn() -> anyhow::Result<()> {
11853 1 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_above_lsn").await?;
11854 1 : let (tenant, ctx) = harness.load().await;
11855 :
11856 148 : fn get_key(id: u32) -> Key {
11857 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
11858 148 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
11859 148 : key.field6 = id;
11860 148 : key
11861 148 : }
11862 :
11863 1 : let img_layer = (0..10)
11864 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
11865 1 : .collect_vec();
11866 :
11867 1 : let delta1 = vec![(
11868 1 : get_key(1),
11869 1 : Lsn(0x20),
11870 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
11871 1 : )];
11872 1 : let delta4 = vec![(
11873 1 : get_key(1),
11874 1 : Lsn(0x28),
11875 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
11876 1 : )];
11877 1 : let delta2 = vec![
11878 1 : (
11879 1 : get_key(1),
11880 1 : Lsn(0x30),
11881 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
11882 1 : ),
11883 1 : (
11884 1 : get_key(1),
11885 1 : Lsn(0x38),
11886 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
11887 1 : ),
11888 : ];
11889 1 : let delta3 = vec![
11890 1 : (
11891 1 : get_key(8),
11892 1 : Lsn(0x48),
11893 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
11894 1 : ),
11895 1 : (
11896 1 : get_key(9),
11897 1 : Lsn(0x48),
11898 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
11899 1 : ),
11900 : ];
11901 :
11902 1 : let tline = tenant
11903 1 : .create_test_timeline_with_layers(
11904 1 : TIMELINE_ID,
11905 1 : Lsn(0x10),
11906 1 : DEFAULT_PG_VERSION,
11907 1 : &ctx,
11908 1 : vec![], // in-memory layers
11909 1 : vec![
11910 1 : // delta1/2/4 only contain a single key but multiple updates
11911 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x28), delta1),
11912 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
11913 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x28)..Lsn(0x30), delta4),
11914 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta3),
11915 1 : ], // delta layers
11916 1 : vec![(Lsn(0x10), img_layer)], // image layers
11917 1 : Lsn(0x50),
11918 1 : )
11919 1 : .await?;
11920 : {
11921 1 : tline
11922 1 : .applied_gc_cutoff_lsn
11923 1 : .lock_for_write()
11924 1 : .store_and_unlock(Lsn(0x30))
11925 1 : .wait()
11926 1 : .await;
11927 : // Update GC info
11928 1 : let mut guard = tline.gc_info.write().unwrap();
11929 1 : *guard = GcInfo {
11930 1 : retain_lsns: vec![
11931 1 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
11932 1 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
11933 1 : ],
11934 1 : cutoffs: GcCutoffs {
11935 1 : time: Some(Lsn(0x30)),
11936 1 : space: Lsn(0x30),
11937 1 : },
11938 1 : leases: Default::default(),
11939 1 : within_ancestor_pitr: false,
11940 1 : };
11941 : }
11942 :
11943 1 : let expected_result = [
11944 1 : Bytes::from_static(b"value 0@0x10"),
11945 1 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
11946 1 : Bytes::from_static(b"value 2@0x10"),
11947 1 : Bytes::from_static(b"value 3@0x10"),
11948 1 : Bytes::from_static(b"value 4@0x10"),
11949 1 : Bytes::from_static(b"value 5@0x10"),
11950 1 : Bytes::from_static(b"value 6@0x10"),
11951 1 : Bytes::from_static(b"value 7@0x10"),
11952 1 : Bytes::from_static(b"value 8@0x10@0x48"),
11953 1 : Bytes::from_static(b"value 9@0x10@0x48"),
11954 1 : ];
11955 :
11956 1 : let expected_result_at_gc_horizon = [
11957 1 : Bytes::from_static(b"value 0@0x10"),
11958 1 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
11959 1 : Bytes::from_static(b"value 2@0x10"),
11960 1 : Bytes::from_static(b"value 3@0x10"),
11961 1 : Bytes::from_static(b"value 4@0x10"),
11962 1 : Bytes::from_static(b"value 5@0x10"),
11963 1 : Bytes::from_static(b"value 6@0x10"),
11964 1 : Bytes::from_static(b"value 7@0x10"),
11965 1 : Bytes::from_static(b"value 8@0x10"),
11966 1 : Bytes::from_static(b"value 9@0x10"),
11967 1 : ];
11968 :
11969 1 : let expected_result_at_lsn_20 = [
11970 1 : Bytes::from_static(b"value 0@0x10"),
11971 1 : Bytes::from_static(b"value 1@0x10@0x20"),
11972 1 : Bytes::from_static(b"value 2@0x10"),
11973 1 : Bytes::from_static(b"value 3@0x10"),
11974 1 : Bytes::from_static(b"value 4@0x10"),
11975 1 : Bytes::from_static(b"value 5@0x10"),
11976 1 : Bytes::from_static(b"value 6@0x10"),
11977 1 : Bytes::from_static(b"value 7@0x10"),
11978 1 : Bytes::from_static(b"value 8@0x10"),
11979 1 : Bytes::from_static(b"value 9@0x10"),
11980 1 : ];
11981 :
11982 1 : let expected_result_at_lsn_10 = [
11983 1 : Bytes::from_static(b"value 0@0x10"),
11984 1 : Bytes::from_static(b"value 1@0x10"),
11985 1 : Bytes::from_static(b"value 2@0x10"),
11986 1 : Bytes::from_static(b"value 3@0x10"),
11987 1 : Bytes::from_static(b"value 4@0x10"),
11988 1 : Bytes::from_static(b"value 5@0x10"),
11989 1 : Bytes::from_static(b"value 6@0x10"),
11990 1 : Bytes::from_static(b"value 7@0x10"),
11991 1 : Bytes::from_static(b"value 8@0x10"),
11992 1 : Bytes::from_static(b"value 9@0x10"),
11993 1 : ];
11994 :
11995 3 : let verify_result = || async {
11996 3 : let gc_horizon = {
11997 3 : let gc_info = tline.gc_info.read().unwrap();
11998 3 : gc_info.cutoffs.time.unwrap_or_default()
11999 : };
12000 33 : for idx in 0..10 {
12001 30 : assert_eq!(
12002 30 : tline
12003 30 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
12004 30 : .await
12005 30 : .unwrap(),
12006 30 : &expected_result[idx]
12007 : );
12008 30 : assert_eq!(
12009 30 : tline
12010 30 : .get(get_key(idx as u32), gc_horizon, &ctx)
12011 30 : .await
12012 30 : .unwrap(),
12013 30 : &expected_result_at_gc_horizon[idx]
12014 : );
12015 30 : assert_eq!(
12016 30 : tline
12017 30 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
12018 30 : .await
12019 30 : .unwrap(),
12020 30 : &expected_result_at_lsn_20[idx]
12021 : );
12022 30 : assert_eq!(
12023 30 : tline
12024 30 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
12025 30 : .await
12026 30 : .unwrap(),
12027 30 : &expected_result_at_lsn_10[idx]
12028 : );
12029 : }
12030 6 : };
12031 :
12032 1 : verify_result().await;
12033 :
12034 1 : let cancel = CancellationToken::new();
12035 1 : tline
12036 1 : .compact_with_gc(
12037 1 : &cancel,
12038 1 : CompactOptions {
12039 1 : compact_lsn_range: Some(CompactLsnRange::above(Lsn(0x28))),
12040 1 : ..Default::default()
12041 1 : },
12042 1 : &ctx,
12043 1 : )
12044 1 : .await
12045 1 : .unwrap();
12046 1 : verify_result().await;
12047 :
12048 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12049 1 : check_layer_map_key_eq(
12050 1 : all_layers,
12051 1 : vec![
12052 : // The original image layer, not compacted
12053 1 : PersistentLayerKey {
12054 1 : key_range: get_key(0)..get_key(10),
12055 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
12056 1 : is_delta: false,
12057 1 : },
12058 : // Delta layer below the specified above_lsn not compacted
12059 1 : PersistentLayerKey {
12060 1 : key_range: get_key(1)..get_key(2),
12061 1 : lsn_range: Lsn(0x20)..Lsn(0x28),
12062 1 : is_delta: true,
12063 1 : },
12064 : // Delta layer compacted above the LSN
12065 1 : PersistentLayerKey {
12066 1 : key_range: get_key(1)..get_key(10),
12067 1 : lsn_range: Lsn(0x28)..Lsn(0x50),
12068 1 : is_delta: true,
12069 1 : },
12070 : ],
12071 : );
12072 :
12073 : // compact again
12074 1 : tline
12075 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
12076 1 : .await
12077 1 : .unwrap();
12078 1 : verify_result().await;
12079 :
12080 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12081 1 : check_layer_map_key_eq(
12082 1 : all_layers,
12083 1 : vec![
12084 : // The compacted image layer (full key range)
12085 1 : PersistentLayerKey {
12086 1 : key_range: Key::MIN..Key::MAX,
12087 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
12088 1 : is_delta: false,
12089 1 : },
12090 : // All other data in the delta layer
12091 1 : PersistentLayerKey {
12092 1 : key_range: get_key(1)..get_key(10),
12093 1 : lsn_range: Lsn(0x10)..Lsn(0x50),
12094 1 : is_delta: true,
12095 1 : },
12096 : ],
12097 : );
12098 :
12099 2 : Ok(())
12100 1 : }
12101 :
12102 : #[cfg(feature = "testing")]
12103 : #[tokio::test]
12104 1 : async fn test_simple_bottom_most_compaction_rectangle() -> anyhow::Result<()> {
12105 1 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_rectangle").await?;
12106 1 : let (tenant, ctx) = harness.load().await;
12107 :
12108 254 : fn get_key(id: u32) -> Key {
12109 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
12110 254 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
12111 254 : key.field6 = id;
12112 254 : key
12113 254 : }
12114 :
12115 1 : let img_layer = (0..10)
12116 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
12117 1 : .collect_vec();
12118 :
12119 1 : let delta1 = vec![(
12120 1 : get_key(1),
12121 1 : Lsn(0x20),
12122 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
12123 1 : )];
12124 1 : let delta4 = vec![(
12125 1 : get_key(1),
12126 1 : Lsn(0x28),
12127 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
12128 1 : )];
12129 1 : let delta2 = vec![
12130 1 : (
12131 1 : get_key(1),
12132 1 : Lsn(0x30),
12133 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
12134 1 : ),
12135 1 : (
12136 1 : get_key(1),
12137 1 : Lsn(0x38),
12138 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
12139 1 : ),
12140 : ];
12141 1 : let delta3 = vec![
12142 1 : (
12143 1 : get_key(8),
12144 1 : Lsn(0x48),
12145 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
12146 1 : ),
12147 1 : (
12148 1 : get_key(9),
12149 1 : Lsn(0x48),
12150 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
12151 1 : ),
12152 : ];
12153 :
12154 1 : let tline = tenant
12155 1 : .create_test_timeline_with_layers(
12156 1 : TIMELINE_ID,
12157 1 : Lsn(0x10),
12158 1 : DEFAULT_PG_VERSION,
12159 1 : &ctx,
12160 1 : vec![], // in-memory layers
12161 1 : vec![
12162 1 : // delta1/2/4 only contain a single key but multiple updates
12163 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x28), delta1),
12164 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
12165 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x28)..Lsn(0x30), delta4),
12166 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta3),
12167 1 : ], // delta layers
12168 1 : vec![(Lsn(0x10), img_layer)], // image layers
12169 1 : Lsn(0x50),
12170 1 : )
12171 1 : .await?;
12172 : {
12173 1 : tline
12174 1 : .applied_gc_cutoff_lsn
12175 1 : .lock_for_write()
12176 1 : .store_and_unlock(Lsn(0x30))
12177 1 : .wait()
12178 1 : .await;
12179 : // Update GC info
12180 1 : let mut guard = tline.gc_info.write().unwrap();
12181 1 : *guard = GcInfo {
12182 1 : retain_lsns: vec![
12183 1 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
12184 1 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
12185 1 : ],
12186 1 : cutoffs: GcCutoffs {
12187 1 : time: Some(Lsn(0x30)),
12188 1 : space: Lsn(0x30),
12189 1 : },
12190 1 : leases: Default::default(),
12191 1 : within_ancestor_pitr: false,
12192 1 : };
12193 : }
12194 :
12195 1 : let expected_result = [
12196 1 : Bytes::from_static(b"value 0@0x10"),
12197 1 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
12198 1 : Bytes::from_static(b"value 2@0x10"),
12199 1 : Bytes::from_static(b"value 3@0x10"),
12200 1 : Bytes::from_static(b"value 4@0x10"),
12201 1 : Bytes::from_static(b"value 5@0x10"),
12202 1 : Bytes::from_static(b"value 6@0x10"),
12203 1 : Bytes::from_static(b"value 7@0x10"),
12204 1 : Bytes::from_static(b"value 8@0x10@0x48"),
12205 1 : Bytes::from_static(b"value 9@0x10@0x48"),
12206 1 : ];
12207 :
12208 1 : let expected_result_at_gc_horizon = [
12209 1 : Bytes::from_static(b"value 0@0x10"),
12210 1 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
12211 1 : Bytes::from_static(b"value 2@0x10"),
12212 1 : Bytes::from_static(b"value 3@0x10"),
12213 1 : Bytes::from_static(b"value 4@0x10"),
12214 1 : Bytes::from_static(b"value 5@0x10"),
12215 1 : Bytes::from_static(b"value 6@0x10"),
12216 1 : Bytes::from_static(b"value 7@0x10"),
12217 1 : Bytes::from_static(b"value 8@0x10"),
12218 1 : Bytes::from_static(b"value 9@0x10"),
12219 1 : ];
12220 :
12221 1 : let expected_result_at_lsn_20 = [
12222 1 : Bytes::from_static(b"value 0@0x10"),
12223 1 : Bytes::from_static(b"value 1@0x10@0x20"),
12224 1 : Bytes::from_static(b"value 2@0x10"),
12225 1 : Bytes::from_static(b"value 3@0x10"),
12226 1 : Bytes::from_static(b"value 4@0x10"),
12227 1 : Bytes::from_static(b"value 5@0x10"),
12228 1 : Bytes::from_static(b"value 6@0x10"),
12229 1 : Bytes::from_static(b"value 7@0x10"),
12230 1 : Bytes::from_static(b"value 8@0x10"),
12231 1 : Bytes::from_static(b"value 9@0x10"),
12232 1 : ];
12233 :
12234 1 : let expected_result_at_lsn_10 = [
12235 1 : Bytes::from_static(b"value 0@0x10"),
12236 1 : Bytes::from_static(b"value 1@0x10"),
12237 1 : Bytes::from_static(b"value 2@0x10"),
12238 1 : Bytes::from_static(b"value 3@0x10"),
12239 1 : Bytes::from_static(b"value 4@0x10"),
12240 1 : Bytes::from_static(b"value 5@0x10"),
12241 1 : Bytes::from_static(b"value 6@0x10"),
12242 1 : Bytes::from_static(b"value 7@0x10"),
12243 1 : Bytes::from_static(b"value 8@0x10"),
12244 1 : Bytes::from_static(b"value 9@0x10"),
12245 1 : ];
12246 :
12247 5 : let verify_result = || async {
12248 5 : let gc_horizon = {
12249 5 : let gc_info = tline.gc_info.read().unwrap();
12250 5 : gc_info.cutoffs.time.unwrap_or_default()
12251 : };
12252 55 : for idx in 0..10 {
12253 50 : assert_eq!(
12254 50 : tline
12255 50 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
12256 50 : .await
12257 50 : .unwrap(),
12258 50 : &expected_result[idx]
12259 : );
12260 50 : assert_eq!(
12261 50 : tline
12262 50 : .get(get_key(idx as u32), gc_horizon, &ctx)
12263 50 : .await
12264 50 : .unwrap(),
12265 50 : &expected_result_at_gc_horizon[idx]
12266 : );
12267 50 : assert_eq!(
12268 50 : tline
12269 50 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
12270 50 : .await
12271 50 : .unwrap(),
12272 50 : &expected_result_at_lsn_20[idx]
12273 : );
12274 50 : assert_eq!(
12275 50 : tline
12276 50 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
12277 50 : .await
12278 50 : .unwrap(),
12279 50 : &expected_result_at_lsn_10[idx]
12280 : );
12281 : }
12282 10 : };
12283 :
12284 1 : verify_result().await;
12285 :
12286 1 : let cancel = CancellationToken::new();
12287 :
12288 1 : tline
12289 1 : .compact_with_gc(
12290 1 : &cancel,
12291 1 : CompactOptions {
12292 1 : compact_key_range: Some((get_key(0)..get_key(2)).into()),
12293 1 : compact_lsn_range: Some((Lsn(0x20)..Lsn(0x28)).into()),
12294 1 : ..Default::default()
12295 1 : },
12296 1 : &ctx,
12297 1 : )
12298 1 : .await
12299 1 : .unwrap();
12300 1 : verify_result().await;
12301 :
12302 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12303 1 : check_layer_map_key_eq(
12304 1 : all_layers,
12305 1 : vec![
12306 : // The original image layer, not compacted
12307 1 : PersistentLayerKey {
12308 1 : key_range: get_key(0)..get_key(10),
12309 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
12310 1 : is_delta: false,
12311 1 : },
12312 : // According the selection logic, we select all layers with start key <= 0x28, so we would merge the layer 0x20-0x28 and
12313 : // the layer 0x28-0x30 into one.
12314 1 : PersistentLayerKey {
12315 1 : key_range: get_key(1)..get_key(2),
12316 1 : lsn_range: Lsn(0x20)..Lsn(0x30),
12317 1 : is_delta: true,
12318 1 : },
12319 : // Above the upper bound and untouched
12320 1 : PersistentLayerKey {
12321 1 : key_range: get_key(1)..get_key(2),
12322 1 : lsn_range: Lsn(0x30)..Lsn(0x50),
12323 1 : is_delta: true,
12324 1 : },
12325 : // This layer is untouched
12326 1 : PersistentLayerKey {
12327 1 : key_range: get_key(8)..get_key(10),
12328 1 : lsn_range: Lsn(0x30)..Lsn(0x50),
12329 1 : is_delta: true,
12330 1 : },
12331 : ],
12332 : );
12333 :
12334 1 : tline
12335 1 : .compact_with_gc(
12336 1 : &cancel,
12337 1 : CompactOptions {
12338 1 : compact_key_range: Some((get_key(3)..get_key(8)).into()),
12339 1 : compact_lsn_range: Some((Lsn(0x28)..Lsn(0x40)).into()),
12340 1 : ..Default::default()
12341 1 : },
12342 1 : &ctx,
12343 1 : )
12344 1 : .await
12345 1 : .unwrap();
12346 1 : verify_result().await;
12347 :
12348 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12349 1 : check_layer_map_key_eq(
12350 1 : all_layers,
12351 1 : vec![
12352 : // The original image layer, not compacted
12353 1 : PersistentLayerKey {
12354 1 : key_range: get_key(0)..get_key(10),
12355 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
12356 1 : is_delta: false,
12357 1 : },
12358 : // Not in the compaction key range, uncompacted
12359 1 : PersistentLayerKey {
12360 1 : key_range: get_key(1)..get_key(2),
12361 1 : lsn_range: Lsn(0x20)..Lsn(0x30),
12362 1 : is_delta: true,
12363 1 : },
12364 : // Not in the compaction key range, uncompacted but need rewrite because the delta layer overlaps with the range
12365 1 : PersistentLayerKey {
12366 1 : key_range: get_key(1)..get_key(2),
12367 1 : lsn_range: Lsn(0x30)..Lsn(0x50),
12368 1 : is_delta: true,
12369 1 : },
12370 : // Note that when we specify the LSN upper bound to be 0x40, the compaction algorithm will not try to cut the layer
12371 : // horizontally in half. Instead, it will include all LSNs that overlap with 0x40. So the real max_lsn of the compaction
12372 : // becomes 0x50.
12373 1 : PersistentLayerKey {
12374 1 : key_range: get_key(8)..get_key(10),
12375 1 : lsn_range: Lsn(0x30)..Lsn(0x50),
12376 1 : is_delta: true,
12377 1 : },
12378 : ],
12379 : );
12380 :
12381 : // compact again
12382 1 : tline
12383 1 : .compact_with_gc(
12384 1 : &cancel,
12385 1 : CompactOptions {
12386 1 : compact_key_range: Some((get_key(0)..get_key(5)).into()),
12387 1 : compact_lsn_range: Some((Lsn(0x20)..Lsn(0x50)).into()),
12388 1 : ..Default::default()
12389 1 : },
12390 1 : &ctx,
12391 1 : )
12392 1 : .await
12393 1 : .unwrap();
12394 1 : verify_result().await;
12395 :
12396 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12397 1 : check_layer_map_key_eq(
12398 1 : all_layers,
12399 1 : vec![
12400 : // The original image layer, not compacted
12401 1 : PersistentLayerKey {
12402 1 : key_range: get_key(0)..get_key(10),
12403 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
12404 1 : is_delta: false,
12405 1 : },
12406 : // The range gets compacted
12407 1 : PersistentLayerKey {
12408 1 : key_range: get_key(1)..get_key(2),
12409 1 : lsn_range: Lsn(0x20)..Lsn(0x50),
12410 1 : is_delta: true,
12411 1 : },
12412 : // Not touched during this iteration of compaction
12413 1 : PersistentLayerKey {
12414 1 : key_range: get_key(8)..get_key(10),
12415 1 : lsn_range: Lsn(0x30)..Lsn(0x50),
12416 1 : is_delta: true,
12417 1 : },
12418 : ],
12419 : );
12420 :
12421 : // final full compaction
12422 1 : tline
12423 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
12424 1 : .await
12425 1 : .unwrap();
12426 1 : verify_result().await;
12427 :
12428 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12429 1 : check_layer_map_key_eq(
12430 1 : all_layers,
12431 1 : vec![
12432 : // The compacted image layer (full key range)
12433 1 : PersistentLayerKey {
12434 1 : key_range: Key::MIN..Key::MAX,
12435 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
12436 1 : is_delta: false,
12437 1 : },
12438 : // All other data in the delta layer
12439 1 : PersistentLayerKey {
12440 1 : key_range: get_key(1)..get_key(10),
12441 1 : lsn_range: Lsn(0x10)..Lsn(0x50),
12442 1 : is_delta: true,
12443 1 : },
12444 : ],
12445 : );
12446 :
12447 2 : Ok(())
12448 1 : }
12449 :
12450 : #[cfg(feature = "testing")]
12451 : #[tokio::test]
12452 1 : async fn test_bottom_most_compation_redo_failure() -> anyhow::Result<()> {
12453 1 : let harness = TenantHarness::create("test_bottom_most_compation_redo_failure").await?;
12454 1 : let (tenant, ctx) = harness.load().await;
12455 :
12456 13 : fn get_key(id: u32) -> Key {
12457 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
12458 13 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
12459 13 : key.field6 = id;
12460 13 : key
12461 13 : }
12462 :
12463 1 : let img_layer = (0..10)
12464 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
12465 1 : .collect_vec();
12466 :
12467 1 : let delta1 = vec![
12468 1 : (
12469 1 : get_key(1),
12470 1 : Lsn(0x20),
12471 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
12472 1 : ),
12473 1 : (
12474 1 : get_key(1),
12475 1 : Lsn(0x24),
12476 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x24")),
12477 1 : ),
12478 1 : (
12479 1 : get_key(1),
12480 1 : Lsn(0x28),
12481 1 : // This record will fail to redo
12482 1 : Value::WalRecord(NeonWalRecord::wal_append_conditional("@0x28", "???")),
12483 1 : ),
12484 : ];
12485 :
12486 1 : let tline = tenant
12487 1 : .create_test_timeline_with_layers(
12488 1 : TIMELINE_ID,
12489 1 : Lsn(0x10),
12490 1 : DEFAULT_PG_VERSION,
12491 1 : &ctx,
12492 1 : vec![], // in-memory layers
12493 1 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
12494 1 : Lsn(0x20)..Lsn(0x30),
12495 1 : delta1,
12496 1 : )], // delta layers
12497 1 : vec![(Lsn(0x10), img_layer)], // image layers
12498 1 : Lsn(0x50),
12499 1 : )
12500 1 : .await?;
12501 : {
12502 1 : tline
12503 1 : .applied_gc_cutoff_lsn
12504 1 : .lock_for_write()
12505 1 : .store_and_unlock(Lsn(0x30))
12506 1 : .wait()
12507 1 : .await;
12508 : // Update GC info
12509 1 : let mut guard = tline.gc_info.write().unwrap();
12510 1 : *guard = GcInfo {
12511 1 : retain_lsns: vec![],
12512 1 : cutoffs: GcCutoffs {
12513 1 : time: Some(Lsn(0x30)),
12514 1 : space: Lsn(0x30),
12515 1 : },
12516 1 : leases: Default::default(),
12517 1 : within_ancestor_pitr: false,
12518 1 : };
12519 : }
12520 :
12521 1 : let cancel = CancellationToken::new();
12522 :
12523 : // Compaction will fail, but should not fire any critical error.
12524 : // Gc-compaction currently cannot figure out what keys are not in the keyspace during the compaction
12525 : // process. It will always try to redo the logs it reads and if it doesn't work, fail the entire
12526 : // compaction job. Tracked in <https://github.com/neondatabase/neon/issues/10395>.
12527 1 : let res = tline
12528 1 : .compact_with_gc(
12529 1 : &cancel,
12530 1 : CompactOptions {
12531 1 : compact_key_range: None,
12532 1 : compact_lsn_range: None,
12533 1 : ..Default::default()
12534 1 : },
12535 1 : &ctx,
12536 1 : )
12537 1 : .await;
12538 1 : assert!(res.is_err());
12539 :
12540 2 : Ok(())
12541 1 : }
12542 :
12543 : #[cfg(feature = "testing")]
12544 : #[tokio::test]
12545 1 : async fn test_synthetic_size_calculation_with_invisible_branches() -> anyhow::Result<()> {
12546 : use pageserver_api::models::TimelineVisibilityState;
12547 :
12548 : use crate::tenant::size::gather_inputs;
12549 :
12550 1 : let tenant_conf = pageserver_api::models::TenantConfig {
12551 1 : // Ensure that we don't compute gc_cutoffs (which needs reading the layer files)
12552 1 : pitr_interval: Some(Duration::ZERO),
12553 1 : ..Default::default()
12554 1 : };
12555 1 : let harness = TenantHarness::create_custom(
12556 1 : "test_synthetic_size_calculation_with_invisible_branches",
12557 1 : tenant_conf,
12558 1 : TenantId::generate(),
12559 1 : ShardIdentity::unsharded(),
12560 1 : Generation::new(0xdeadbeef),
12561 1 : )
12562 1 : .await?;
12563 1 : let (tenant, ctx) = harness.load().await;
12564 1 : let main_tline = tenant
12565 1 : .create_test_timeline_with_layers(
12566 1 : TIMELINE_ID,
12567 1 : Lsn(0x10),
12568 1 : DEFAULT_PG_VERSION,
12569 1 : &ctx,
12570 1 : vec![],
12571 1 : vec![],
12572 1 : vec![],
12573 1 : Lsn(0x100),
12574 1 : )
12575 1 : .await?;
12576 :
12577 1 : let snapshot1 = TimelineId::from_array(hex!("11223344556677881122334455667790"));
12578 1 : tenant
12579 1 : .branch_timeline_test_with_layers(
12580 1 : &main_tline,
12581 1 : snapshot1,
12582 1 : Some(Lsn(0x20)),
12583 1 : &ctx,
12584 1 : vec![],
12585 1 : vec![],
12586 1 : Lsn(0x50),
12587 1 : )
12588 1 : .await?;
12589 1 : let snapshot2 = TimelineId::from_array(hex!("11223344556677881122334455667791"));
12590 1 : tenant
12591 1 : .branch_timeline_test_with_layers(
12592 1 : &main_tline,
12593 1 : snapshot2,
12594 1 : Some(Lsn(0x30)),
12595 1 : &ctx,
12596 1 : vec![],
12597 1 : vec![],
12598 1 : Lsn(0x50),
12599 1 : )
12600 1 : .await?;
12601 1 : let snapshot3 = TimelineId::from_array(hex!("11223344556677881122334455667792"));
12602 1 : tenant
12603 1 : .branch_timeline_test_with_layers(
12604 1 : &main_tline,
12605 1 : snapshot3,
12606 1 : Some(Lsn(0x40)),
12607 1 : &ctx,
12608 1 : vec![],
12609 1 : vec![],
12610 1 : Lsn(0x50),
12611 1 : )
12612 1 : .await?;
12613 1 : let limit = Arc::new(Semaphore::new(1));
12614 1 : let max_retention_period = None;
12615 1 : let mut logical_size_cache = HashMap::new();
12616 1 : let cause = LogicalSizeCalculationCause::EvictionTaskImitation;
12617 1 : let cancel = CancellationToken::new();
12618 :
12619 1 : let inputs = gather_inputs(
12620 1 : &tenant,
12621 1 : &limit,
12622 1 : max_retention_period,
12623 1 : &mut logical_size_cache,
12624 1 : cause,
12625 1 : &cancel,
12626 1 : &ctx,
12627 : )
12628 1 : .instrument(info_span!(
12629 : "gather_inputs",
12630 : tenant_id = "unknown",
12631 : shard_id = "unknown",
12632 : ))
12633 1 : .await?;
12634 : use crate::tenant::size::{LsnKind, ModelInputs, SegmentMeta};
12635 : use LsnKind::*;
12636 : use tenant_size_model::Segment;
12637 1 : let ModelInputs { mut segments, .. } = inputs;
12638 15 : segments.retain(|s| s.timeline_id == TIMELINE_ID);
12639 6 : for segment in segments.iter_mut() {
12640 6 : segment.segment.parent = None; // We don't care about the parent for the test
12641 6 : segment.segment.size = None; // We don't care about the size for the test
12642 6 : }
12643 1 : assert_eq!(
12644 : segments,
12645 : [
12646 : SegmentMeta {
12647 : segment: Segment {
12648 : parent: None,
12649 : lsn: 0x10,
12650 : size: None,
12651 : needed: false,
12652 : },
12653 : timeline_id: TIMELINE_ID,
12654 : kind: BranchStart,
12655 : },
12656 : SegmentMeta {
12657 : segment: Segment {
12658 : parent: None,
12659 : lsn: 0x20,
12660 : size: None,
12661 : needed: false,
12662 : },
12663 : timeline_id: TIMELINE_ID,
12664 : kind: BranchPoint,
12665 : },
12666 : SegmentMeta {
12667 : segment: Segment {
12668 : parent: None,
12669 : lsn: 0x30,
12670 : size: None,
12671 : needed: false,
12672 : },
12673 : timeline_id: TIMELINE_ID,
12674 : kind: BranchPoint,
12675 : },
12676 : SegmentMeta {
12677 : segment: Segment {
12678 : parent: None,
12679 : lsn: 0x40,
12680 : size: None,
12681 : needed: false,
12682 : },
12683 : timeline_id: TIMELINE_ID,
12684 : kind: BranchPoint,
12685 : },
12686 : SegmentMeta {
12687 : segment: Segment {
12688 : parent: None,
12689 : lsn: 0x100,
12690 : size: None,
12691 : needed: false,
12692 : },
12693 : timeline_id: TIMELINE_ID,
12694 : kind: GcCutOff,
12695 : }, // we need to retain everything above the last branch point
12696 : SegmentMeta {
12697 : segment: Segment {
12698 : parent: None,
12699 : lsn: 0x100,
12700 : size: None,
12701 : needed: true,
12702 : },
12703 : timeline_id: TIMELINE_ID,
12704 : kind: BranchEnd,
12705 : },
12706 : ]
12707 : );
12708 :
12709 1 : main_tline
12710 1 : .remote_client
12711 1 : .schedule_index_upload_for_timeline_invisible_state(
12712 1 : TimelineVisibilityState::Invisible,
12713 0 : )?;
12714 1 : main_tline.remote_client.wait_completion().await?;
12715 1 : let inputs = gather_inputs(
12716 1 : &tenant,
12717 1 : &limit,
12718 1 : max_retention_period,
12719 1 : &mut logical_size_cache,
12720 1 : cause,
12721 1 : &cancel,
12722 1 : &ctx,
12723 : )
12724 1 : .instrument(info_span!(
12725 : "gather_inputs",
12726 : tenant_id = "unknown",
12727 : shard_id = "unknown",
12728 : ))
12729 1 : .await?;
12730 1 : let ModelInputs { mut segments, .. } = inputs;
12731 14 : segments.retain(|s| s.timeline_id == TIMELINE_ID);
12732 5 : for segment in segments.iter_mut() {
12733 5 : segment.segment.parent = None; // We don't care about the parent for the test
12734 5 : segment.segment.size = None; // We don't care about the size for the test
12735 5 : }
12736 1 : assert_eq!(
12737 : segments,
12738 : [
12739 : SegmentMeta {
12740 : segment: Segment {
12741 : parent: None,
12742 : lsn: 0x10,
12743 : size: None,
12744 : needed: false,
12745 : },
12746 : timeline_id: TIMELINE_ID,
12747 : kind: BranchStart,
12748 : },
12749 : SegmentMeta {
12750 : segment: Segment {
12751 : parent: None,
12752 : lsn: 0x20,
12753 : size: None,
12754 : needed: false,
12755 : },
12756 : timeline_id: TIMELINE_ID,
12757 : kind: BranchPoint,
12758 : },
12759 : SegmentMeta {
12760 : segment: Segment {
12761 : parent: None,
12762 : lsn: 0x30,
12763 : size: None,
12764 : needed: false,
12765 : },
12766 : timeline_id: TIMELINE_ID,
12767 : kind: BranchPoint,
12768 : },
12769 : SegmentMeta {
12770 : segment: Segment {
12771 : parent: None,
12772 : lsn: 0x40,
12773 : size: None,
12774 : needed: false,
12775 : },
12776 : timeline_id: TIMELINE_ID,
12777 : kind: BranchPoint,
12778 : },
12779 : SegmentMeta {
12780 : segment: Segment {
12781 : parent: None,
12782 : lsn: 0x40, // Branch end LSN == last branch point LSN
12783 : size: None,
12784 : needed: true,
12785 : },
12786 : timeline_id: TIMELINE_ID,
12787 : kind: BranchEnd,
12788 : },
12789 : ]
12790 : );
12791 2 : Ok(())
12792 1 : }
12793 : }
|
