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 remote_storage::{DownloadError, GenericRemoteStorage, TimeoutOrCancel};
42 : use remote_timeline_client::index::GcCompactionState;
43 : use remote_timeline_client::manifest::{
44 : LATEST_TENANT_MANIFEST_VERSION, OffloadedTimelineManifest, TenantManifest,
45 : };
46 : use remote_timeline_client::{
47 : FAILED_REMOTE_OP_RETRIES, FAILED_UPLOAD_WARN_THRESHOLD, UploadQueueNotReadyError,
48 : download_tenant_manifest,
49 : };
50 : use secondary::heatmap::{HeatMapTenant, HeatMapTimeline};
51 : use storage_broker::BrokerClientChannel;
52 : use timeline::compaction::{CompactionOutcome, GcCompactionQueue};
53 : use timeline::import_pgdata::ImportingTimeline;
54 : use timeline::offload::{OffloadError, offload_timeline};
55 : use timeline::{
56 : CompactFlags, CompactOptions, CompactionError, PreviousHeatmap, ShutdownMode, import_pgdata,
57 : };
58 : use tokio::io::BufReader;
59 : use tokio::sync::{Notify, Semaphore, watch};
60 : use tokio::task::JoinSet;
61 : use tokio_util::sync::CancellationToken;
62 : use tracing::*;
63 : use upload_queue::NotInitialized;
64 : use utils::circuit_breaker::CircuitBreaker;
65 : use utils::crashsafe::path_with_suffix_extension;
66 : use utils::sync::gate::{Gate, GateGuard};
67 : use utils::timeout::{TimeoutCancellableError, timeout_cancellable};
68 : use utils::try_rcu::ArcSwapExt;
69 : use utils::zstd::{create_zst_tarball, extract_zst_tarball};
70 : use utils::{backoff, completion, failpoint_support, fs_ext, pausable_failpoint};
71 :
72 : use self::config::{AttachedLocationConfig, AttachmentMode, LocationConf};
73 : use self::metadata::TimelineMetadata;
74 : use self::mgr::{GetActiveTenantError, GetTenantError};
75 : use self::remote_timeline_client::upload::{upload_index_part, upload_tenant_manifest};
76 : use self::remote_timeline_client::{RemoteTimelineClient, WaitCompletionError};
77 : use self::timeline::uninit::{TimelineCreateGuard, TimelineExclusionError, UninitializedTimeline};
78 : use self::timeline::{
79 : EvictionTaskTenantState, GcCutoffs, TimelineDeleteProgress, TimelineResources, WaitLsnError,
80 : };
81 : use crate::basebackup_cache::BasebackupPrepareSender;
82 : use crate::config::PageServerConf;
83 : use crate::context;
84 : use crate::context::RequestContextBuilder;
85 : use crate::context::{DownloadBehavior, RequestContext};
86 : use crate::deletion_queue::{DeletionQueueClient, DeletionQueueError};
87 : use crate::l0_flush::L0FlushGlobalState;
88 : use crate::metrics::{
89 : BROKEN_TENANTS_SET, CIRCUIT_BREAKERS_BROKEN, CIRCUIT_BREAKERS_UNBROKEN, CONCURRENT_INITDBS,
90 : INITDB_RUN_TIME, INITDB_SEMAPHORE_ACQUISITION_TIME, TENANT, TENANT_OFFLOADED_TIMELINES,
91 : TENANT_STATE_METRIC, TENANT_SYNTHETIC_SIZE_METRIC, remove_tenant_metrics,
92 : };
93 : use crate::task_mgr::TaskKind;
94 : use crate::tenant::config::LocationMode;
95 : use crate::tenant::gc_result::GcResult;
96 : pub use crate::tenant::remote_timeline_client::index::IndexPart;
97 : use crate::tenant::remote_timeline_client::{
98 : INITDB_PATH, MaybeDeletedIndexPart, remote_initdb_archive_path,
99 : };
100 : use crate::tenant::storage_layer::{DeltaLayer, ImageLayer};
101 : use crate::tenant::timeline::delete::DeleteTimelineFlow;
102 : use crate::tenant::timeline::uninit::cleanup_timeline_directory;
103 : use crate::virtual_file::VirtualFile;
104 : use crate::walingest::WalLagCooldown;
105 : use crate::walredo::{PostgresRedoManager, RedoAttemptType};
106 : use crate::{InitializationOrder, TEMP_FILE_SUFFIX, import_datadir, span, task_mgr, walredo};
107 :
108 0 : static INIT_DB_SEMAPHORE: Lazy<Semaphore> = Lazy::new(|| Semaphore::new(8));
109 : use utils::crashsafe;
110 : use utils::generation::Generation;
111 : use utils::id::TimelineId;
112 : use utils::lsn::{Lsn, RecordLsn};
113 :
114 : pub mod blob_io;
115 : pub mod block_io;
116 : pub mod vectored_blob_io;
117 :
118 : pub mod disk_btree;
119 : pub(crate) mod ephemeral_file;
120 : pub mod layer_map;
121 :
122 : pub mod metadata;
123 : pub mod remote_timeline_client;
124 : pub mod storage_layer;
125 :
126 : pub mod checks;
127 : pub mod config;
128 : pub mod mgr;
129 : pub mod secondary;
130 : pub mod tasks;
131 : pub mod upload_queue;
132 :
133 : pub(crate) mod timeline;
134 :
135 : pub mod size;
136 :
137 : mod gc_block;
138 : mod gc_result;
139 : pub(crate) mod throttle;
140 :
141 : pub(crate) use timeline::{LogicalSizeCalculationCause, PageReconstructError, Timeline};
142 :
143 : pub(crate) use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
144 : // re-export for use in walreceiver
145 : pub use crate::tenant::timeline::WalReceiverInfo;
146 :
147 : /// The "tenants" part of `tenants/<tenant>/timelines...`
148 : pub const TENANTS_SEGMENT_NAME: &str = "tenants";
149 :
150 : /// Parts of the `.neon/tenants/<tenant_id>/timelines/<timeline_id>` directory prefix.
151 : pub const TIMELINES_SEGMENT_NAME: &str = "timelines";
152 :
153 : /// References to shared objects that are passed into each tenant, such
154 : /// as the shared remote storage client and process initialization state.
155 : #[derive(Clone)]
156 : pub struct TenantSharedResources {
157 : pub broker_client: storage_broker::BrokerClientChannel,
158 : pub remote_storage: GenericRemoteStorage,
159 : pub deletion_queue_client: DeletionQueueClient,
160 : pub l0_flush_global_state: L0FlushGlobalState,
161 : pub basebackup_prepare_sender: BasebackupPrepareSender,
162 : }
163 :
164 : /// A [`TenantShard`] is really an _attached_ tenant. The configuration
165 : /// for an attached tenant is a subset of the [`LocationConf`], represented
166 : /// in this struct.
167 : #[derive(Clone)]
168 : pub(super) struct AttachedTenantConf {
169 : tenant_conf: pageserver_api::models::TenantConfig,
170 : location: AttachedLocationConfig,
171 : /// The deadline before which we are blocked from GC so that
172 : /// leases have a chance to be renewed.
173 : lsn_lease_deadline: Option<tokio::time::Instant>,
174 : }
175 :
176 : impl AttachedTenantConf {
177 117 : fn new(
178 117 : tenant_conf: pageserver_api::models::TenantConfig,
179 117 : location: AttachedLocationConfig,
180 117 : ) -> Self {
181 : // Sets a deadline before which we cannot proceed to GC due to lsn lease.
182 : //
183 : // We do this as the leases mapping are not persisted to disk. By delaying GC by lease
184 : // length, we guarantee that all the leases we granted before will have a chance to renew
185 : // when we run GC for the first time after restart / transition from AttachedMulti to AttachedSingle.
186 117 : let lsn_lease_deadline = if location.attach_mode == AttachmentMode::Single {
187 117 : Some(
188 117 : tokio::time::Instant::now()
189 117 : + tenant_conf
190 117 : .lsn_lease_length
191 117 : .unwrap_or(LsnLease::DEFAULT_LENGTH),
192 117 : )
193 : } else {
194 : // We don't use `lsn_lease_deadline` to delay GC in AttachedMulti and AttachedStale
195 : // because we don't do GC in these modes.
196 0 : None
197 : };
198 :
199 117 : Self {
200 117 : tenant_conf,
201 117 : location,
202 117 : lsn_lease_deadline,
203 117 : }
204 117 : }
205 :
206 117 : fn try_from(location_conf: LocationConf) -> anyhow::Result<Self> {
207 117 : match &location_conf.mode {
208 117 : LocationMode::Attached(attach_conf) => {
209 117 : Ok(Self::new(location_conf.tenant_conf, *attach_conf))
210 : }
211 : LocationMode::Secondary(_) => {
212 0 : anyhow::bail!(
213 0 : "Attempted to construct AttachedTenantConf from a LocationConf in secondary mode"
214 0 : )
215 : }
216 : }
217 117 : }
218 :
219 381 : fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
220 381 : self.lsn_lease_deadline
221 381 : .map(|d| tokio::time::Instant::now() < d)
222 381 : .unwrap_or(false)
223 381 : }
224 : }
225 : struct TimelinePreload {
226 : timeline_id: TimelineId,
227 : client: RemoteTimelineClient,
228 : index_part: Result<MaybeDeletedIndexPart, DownloadError>,
229 : previous_heatmap: Option<PreviousHeatmap>,
230 : }
231 :
232 : pub(crate) struct TenantPreload {
233 : /// The tenant manifest from remote storage, or None if no manifest was found.
234 : tenant_manifest: Option<TenantManifest>,
235 : /// Map from timeline ID to a possible timeline preload. It is None iff the timeline is offloaded according to the manifest.
236 : timelines: HashMap<TimelineId, Option<TimelinePreload>>,
237 : }
238 :
239 : /// When we spawn a tenant, there is a special mode for tenant creation that
240 : /// avoids trying to read anything from remote storage.
241 : pub(crate) enum SpawnMode {
242 : /// Activate as soon as possible
243 : Eager,
244 : /// Lazy activation in the background, with the option to skip the queue if the need comes up
245 : Lazy,
246 : }
247 :
248 : ///
249 : /// Tenant consists of multiple timelines. Keep them in a hash table.
250 : ///
251 : pub struct TenantShard {
252 : // Global pageserver config parameters
253 : pub conf: &'static PageServerConf,
254 :
255 : /// The value creation timestamp, used to measure activation delay, see:
256 : /// <https://github.com/neondatabase/neon/issues/4025>
257 : constructed_at: Instant,
258 :
259 : state: watch::Sender<TenantState>,
260 :
261 : // Overridden tenant-specific config parameters.
262 : // We keep pageserver_api::models::TenantConfig sturct here to preserve the information
263 : // about parameters that are not set.
264 : // This is necessary to allow global config updates.
265 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
266 :
267 : tenant_shard_id: TenantShardId,
268 :
269 : // The detailed sharding information, beyond the number/count in tenant_shard_id
270 : shard_identity: ShardIdentity,
271 :
272 : /// The remote storage generation, used to protect S3 objects from split-brain.
273 : /// Does not change over the lifetime of the [`TenantShard`] object.
274 : ///
275 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
276 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
277 : generation: Generation,
278 :
279 : timelines: Mutex<HashMap<TimelineId, Arc<Timeline>>>,
280 :
281 : /// During timeline creation, we first insert the TimelineId to the
282 : /// creating map, then `timelines`, then remove it from the creating map.
283 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
284 : timelines_creating: std::sync::Mutex<HashSet<TimelineId>>,
285 :
286 : /// Possibly offloaded and archived timelines
287 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
288 : timelines_offloaded: Mutex<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
289 :
290 : /// Tracks the timelines that are currently importing into this tenant shard.
291 : ///
292 : /// Note that importing timelines are also present in [`Self::timelines_creating`].
293 : /// Keep this in mind when ordering lock acquisition.
294 : ///
295 : /// Lifetime:
296 : /// * An imported timeline is created while scanning the bucket on tenant attach
297 : /// if the index part contains an `import_pgdata` entry and said field marks the import
298 : /// as in progress.
299 : /// * Imported timelines are removed when the storage controller calls the post timeline
300 : /// import activation endpoint.
301 : timelines_importing: std::sync::Mutex<HashMap<TimelineId, ImportingTimeline>>,
302 :
303 : /// The last tenant manifest known to be in remote storage. None if the manifest has not yet
304 : /// been either downloaded or uploaded. Always Some after tenant attach.
305 : ///
306 : /// Initially populated during tenant attach, updated via `maybe_upload_tenant_manifest`.
307 : ///
308 : /// Do not modify this directly. It is used to check whether a new manifest needs to be
309 : /// uploaded. The manifest is constructed in `build_tenant_manifest`, and uploaded via
310 : /// `maybe_upload_tenant_manifest`.
311 : remote_tenant_manifest: tokio::sync::Mutex<Option<TenantManifest>>,
312 :
313 : // This mutex prevents creation of new timelines during GC.
314 : // Adding yet another mutex (in addition to `timelines`) is needed because holding
315 : // `timelines` mutex during all GC iteration
316 : // may block for a long time `get_timeline`, `get_timelines_state`,... and other operations
317 : // with timelines, which in turn may cause dropping replication connection, expiration of wait_for_lsn
318 : // timeout...
319 : gc_cs: tokio::sync::Mutex<()>,
320 : walredo_mgr: Option<Arc<WalRedoManager>>,
321 :
322 : /// Provides access to timeline data sitting in the remote storage.
323 : pub(crate) remote_storage: GenericRemoteStorage,
324 :
325 : /// Access to global deletion queue for when this tenant wants to schedule a deletion.
326 : deletion_queue_client: DeletionQueueClient,
327 :
328 : /// A channel to send async requests to prepare a basebackup for the basebackup cache.
329 : basebackup_prepare_sender: BasebackupPrepareSender,
330 :
331 : /// Cached logical sizes updated updated on each [`TenantShard::gather_size_inputs`].
332 : cached_logical_sizes: tokio::sync::Mutex<HashMap<(TimelineId, Lsn), u64>>,
333 : cached_synthetic_tenant_size: Arc<AtomicU64>,
334 :
335 : eviction_task_tenant_state: tokio::sync::Mutex<EvictionTaskTenantState>,
336 :
337 : /// Track repeated failures to compact, so that we can back off.
338 : /// Overhead of mutex is acceptable because compaction is done with a multi-second period.
339 : compaction_circuit_breaker: std::sync::Mutex<CircuitBreaker>,
340 :
341 : /// Signals the tenant compaction loop that there is L0 compaction work to be done.
342 : pub(crate) l0_compaction_trigger: Arc<Notify>,
343 :
344 : /// Scheduled gc-compaction tasks.
345 : scheduled_compaction_tasks: std::sync::Mutex<HashMap<TimelineId, Arc<GcCompactionQueue>>>,
346 :
347 : /// If the tenant is in Activating state, notify this to encourage it
348 : /// to proceed to Active as soon as possible, rather than waiting for lazy
349 : /// background warmup.
350 : pub(crate) activate_now_sem: tokio::sync::Semaphore,
351 :
352 : /// Time it took for the tenant to activate. Zero if not active yet.
353 : attach_wal_lag_cooldown: Arc<std::sync::OnceLock<WalLagCooldown>>,
354 :
355 : // Cancellation token fires when we have entered shutdown(). This is a parent of
356 : // Timelines' cancellation token.
357 : pub(crate) cancel: CancellationToken,
358 :
359 : // Users of the TenantShard such as the page service must take this Gate to avoid
360 : // trying to use a TenantShard which is shutting down.
361 : pub(crate) gate: Gate,
362 :
363 : /// Throttle applied at the top of [`Timeline::get`].
364 : /// All [`TenantShard::timelines`] of a given [`TenantShard`] instance share the same [`throttle::Throttle`] instance.
365 : pub(crate) pagestream_throttle: Arc<throttle::Throttle>,
366 :
367 : pub(crate) pagestream_throttle_metrics: Arc<crate::metrics::tenant_throttling::Pagestream>,
368 :
369 : /// An ongoing timeline detach concurrency limiter.
370 : ///
371 : /// As a tenant will likely be restarted as part of timeline detach ancestor it makes no sense
372 : /// to have two running at the same time. A different one can be started if an earlier one
373 : /// has failed for whatever reason.
374 : ongoing_timeline_detach: std::sync::Mutex<Option<(TimelineId, utils::completion::Barrier)>>,
375 :
376 : /// `index_part.json` based gc blocking reason tracking.
377 : ///
378 : /// New gc iterations must start a new iteration by acquiring `GcBlock::start` before
379 : /// proceeding.
380 : pub(crate) gc_block: gc_block::GcBlock,
381 :
382 : l0_flush_global_state: L0FlushGlobalState,
383 : }
384 : impl std::fmt::Debug for TenantShard {
385 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
386 0 : write!(f, "{} ({})", self.tenant_shard_id, self.current_state())
387 0 : }
388 : }
389 :
390 : pub(crate) enum WalRedoManager {
391 : Prod(WalredoManagerId, PostgresRedoManager),
392 : #[cfg(test)]
393 : Test(harness::TestRedoManager),
394 : }
395 :
396 : #[derive(thiserror::Error, Debug)]
397 : #[error("pageserver is shutting down")]
398 : pub(crate) struct GlobalShutDown;
399 :
400 : impl WalRedoManager {
401 0 : pub(crate) fn new(mgr: PostgresRedoManager) -> Result<Arc<Self>, GlobalShutDown> {
402 0 : let id = WalredoManagerId::next();
403 0 : let arc = Arc::new(Self::Prod(id, mgr));
404 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
405 0 : match &mut *guard {
406 0 : Some(map) => {
407 0 : map.insert(id, Arc::downgrade(&arc));
408 0 : Ok(arc)
409 : }
410 0 : None => Err(GlobalShutDown),
411 : }
412 0 : }
413 : }
414 :
415 : impl Drop for WalRedoManager {
416 5 : fn drop(&mut self) {
417 5 : match self {
418 0 : Self::Prod(id, _) => {
419 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
420 0 : if let Some(map) = &mut *guard {
421 0 : map.remove(id).expect("new() registers, drop() unregisters");
422 0 : }
423 : }
424 : #[cfg(test)]
425 5 : Self::Test(_) => {
426 5 : // Not applicable to test redo manager
427 5 : }
428 : }
429 5 : }
430 : }
431 :
432 : /// Global registry of all walredo managers so that [`crate::shutdown_pageserver`] can shut down
433 : /// the walredo processes outside of the regular order.
434 : ///
435 : /// This is necessary to work around a systemd bug where it freezes if there are
436 : /// walredo processes left => <https://github.com/neondatabase/cloud/issues/11387>
437 : #[allow(clippy::type_complexity)]
438 : pub(crate) static WALREDO_MANAGERS: once_cell::sync::Lazy<
439 : Mutex<Option<HashMap<WalredoManagerId, Weak<WalRedoManager>>>>,
440 0 : > = once_cell::sync::Lazy::new(|| Mutex::new(Some(HashMap::new())));
441 : #[derive(PartialEq, Eq, Hash, Clone, Copy, Debug)]
442 : pub(crate) struct WalredoManagerId(u64);
443 : impl WalredoManagerId {
444 0 : pub fn next() -> Self {
445 : static NEXT: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(1);
446 0 : let id = NEXT.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
447 0 : if id == 0 {
448 0 : panic!(
449 0 : "WalredoManagerId::new() returned 0, indicating wraparound, risking it's no longer unique"
450 0 : );
451 0 : }
452 0 : Self(id)
453 0 : }
454 : }
455 :
456 : #[cfg(test)]
457 : impl From<harness::TestRedoManager> for WalRedoManager {
458 117 : fn from(mgr: harness::TestRedoManager) -> Self {
459 117 : Self::Test(mgr)
460 117 : }
461 : }
462 :
463 : impl WalRedoManager {
464 3 : pub(crate) async fn shutdown(&self) -> bool {
465 3 : match self {
466 0 : Self::Prod(_, mgr) => mgr.shutdown().await,
467 : #[cfg(test)]
468 : Self::Test(_) => {
469 : // Not applicable to test redo manager
470 3 : true
471 : }
472 : }
473 3 : }
474 :
475 0 : pub(crate) fn maybe_quiesce(&self, idle_timeout: Duration) {
476 0 : match self {
477 0 : Self::Prod(_, mgr) => mgr.maybe_quiesce(idle_timeout),
478 0 : #[cfg(test)]
479 0 : Self::Test(_) => {
480 0 : // Not applicable to test redo manager
481 0 : }
482 0 : }
483 0 : }
484 :
485 : /// # Cancel-Safety
486 : ///
487 : /// This method is cancellation-safe.
488 26774 : pub async fn request_redo(
489 26774 : &self,
490 26774 : key: pageserver_api::key::Key,
491 26774 : lsn: Lsn,
492 26774 : base_img: Option<(Lsn, bytes::Bytes)>,
493 26774 : records: Vec<(Lsn, pageserver_api::record::NeonWalRecord)>,
494 26774 : pg_version: u32,
495 26774 : redo_attempt_type: RedoAttemptType,
496 26774 : ) -> Result<bytes::Bytes, walredo::Error> {
497 26774 : match self {
498 0 : Self::Prod(_, mgr) => {
499 0 : mgr.request_redo(key, lsn, base_img, records, pg_version, redo_attempt_type)
500 0 : .await
501 : }
502 : #[cfg(test)]
503 26774 : Self::Test(mgr) => {
504 26774 : mgr.request_redo(key, lsn, base_img, records, pg_version, redo_attempt_type)
505 26774 : .await
506 : }
507 : }
508 26774 : }
509 :
510 0 : pub(crate) fn status(&self) -> Option<WalRedoManagerStatus> {
511 0 : match self {
512 0 : WalRedoManager::Prod(_, m) => Some(m.status()),
513 0 : #[cfg(test)]
514 0 : WalRedoManager::Test(_) => None,
515 0 : }
516 0 : }
517 : }
518 :
519 : /// A very lightweight memory representation of an offloaded timeline.
520 : ///
521 : /// We need to store the list of offloaded timelines so that we can perform operations on them,
522 : /// like unoffloading them, or (at a later date), decide to perform flattening.
523 : /// This type has a much smaller memory impact than [`Timeline`], and thus we can store many
524 : /// more offloaded timelines than we can manage ones that aren't.
525 : pub struct OffloadedTimeline {
526 : pub tenant_shard_id: TenantShardId,
527 : pub timeline_id: TimelineId,
528 : pub ancestor_timeline_id: Option<TimelineId>,
529 : /// Whether to retain the branch lsn at the ancestor or not
530 : pub ancestor_retain_lsn: Option<Lsn>,
531 :
532 : /// When the timeline was archived.
533 : ///
534 : /// Present for future flattening deliberations.
535 : pub archived_at: NaiveDateTime,
536 :
537 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
538 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
539 : pub delete_progress: TimelineDeleteProgress,
540 :
541 : /// Part of the `OffloadedTimeline` object's lifecycle: this needs to be set before we drop it
542 : pub deleted_from_ancestor: AtomicBool,
543 : }
544 :
545 : impl OffloadedTimeline {
546 : /// Obtains an offloaded timeline from a given timeline object.
547 : ///
548 : /// Returns `None` if the `archived_at` flag couldn't be obtained, i.e.
549 : /// the timeline is not in a stopped state.
550 : /// Panics if the timeline is not archived.
551 1 : fn from_timeline(timeline: &Timeline) -> Result<Self, UploadQueueNotReadyError> {
552 1 : let (ancestor_retain_lsn, ancestor_timeline_id) =
553 1 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
554 1 : let ancestor_lsn = timeline.get_ancestor_lsn();
555 1 : let ancestor_timeline_id = ancestor_timeline.timeline_id;
556 1 : let mut gc_info = ancestor_timeline.gc_info.write().unwrap();
557 1 : gc_info.insert_child(timeline.timeline_id, ancestor_lsn, MaybeOffloaded::Yes);
558 1 : (Some(ancestor_lsn), Some(ancestor_timeline_id))
559 : } else {
560 0 : (None, None)
561 : };
562 1 : let archived_at = timeline
563 1 : .remote_client
564 1 : .archived_at_stopped_queue()?
565 1 : .expect("must be called on an archived timeline");
566 1 : Ok(Self {
567 1 : tenant_shard_id: timeline.tenant_shard_id,
568 1 : timeline_id: timeline.timeline_id,
569 1 : ancestor_timeline_id,
570 1 : ancestor_retain_lsn,
571 1 : archived_at,
572 1 :
573 1 : delete_progress: timeline.delete_progress.clone(),
574 1 : deleted_from_ancestor: AtomicBool::new(false),
575 1 : })
576 1 : }
577 0 : fn from_manifest(tenant_shard_id: TenantShardId, manifest: &OffloadedTimelineManifest) -> Self {
578 0 : // We expect to reach this case in tenant loading, where the `retain_lsn` is populated in the parent's `gc_info`
579 0 : // by the `initialize_gc_info` function.
580 0 : let OffloadedTimelineManifest {
581 0 : timeline_id,
582 0 : ancestor_timeline_id,
583 0 : ancestor_retain_lsn,
584 0 : archived_at,
585 0 : } = *manifest;
586 0 : Self {
587 0 : tenant_shard_id,
588 0 : timeline_id,
589 0 : ancestor_timeline_id,
590 0 : ancestor_retain_lsn,
591 0 : archived_at,
592 0 : delete_progress: TimelineDeleteProgress::default(),
593 0 : deleted_from_ancestor: AtomicBool::new(false),
594 0 : }
595 0 : }
596 1 : fn manifest(&self) -> OffloadedTimelineManifest {
597 1 : let Self {
598 1 : timeline_id,
599 1 : ancestor_timeline_id,
600 1 : ancestor_retain_lsn,
601 1 : archived_at,
602 1 : ..
603 1 : } = self;
604 1 : OffloadedTimelineManifest {
605 1 : timeline_id: *timeline_id,
606 1 : ancestor_timeline_id: *ancestor_timeline_id,
607 1 : ancestor_retain_lsn: *ancestor_retain_lsn,
608 1 : archived_at: *archived_at,
609 1 : }
610 1 : }
611 : /// Delete this timeline's retain_lsn from its ancestor, if present in the given tenant
612 0 : fn delete_from_ancestor_with_timelines(
613 0 : &self,
614 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
615 0 : ) {
616 0 : if let (Some(_retain_lsn), Some(ancestor_timeline_id)) =
617 0 : (self.ancestor_retain_lsn, self.ancestor_timeline_id)
618 : {
619 0 : if let Some((_, ancestor_timeline)) = timelines
620 0 : .iter()
621 0 : .find(|(tid, _tl)| **tid == ancestor_timeline_id)
622 : {
623 0 : let removal_happened = ancestor_timeline
624 0 : .gc_info
625 0 : .write()
626 0 : .unwrap()
627 0 : .remove_child_offloaded(self.timeline_id);
628 0 : if !removal_happened {
629 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
630 0 : "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
631 0 : }
632 0 : }
633 0 : }
634 0 : self.deleted_from_ancestor.store(true, Ordering::Release);
635 0 : }
636 : /// Call [`Self::delete_from_ancestor_with_timelines`] instead if possible.
637 : ///
638 : /// As the entire tenant is being dropped, don't bother deregistering the `retain_lsn` from the ancestor.
639 1 : fn defuse_for_tenant_drop(&self) {
640 1 : self.deleted_from_ancestor.store(true, Ordering::Release);
641 1 : }
642 : }
643 :
644 : impl fmt::Debug for OffloadedTimeline {
645 0 : fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
646 0 : write!(f, "OffloadedTimeline<{}>", self.timeline_id)
647 0 : }
648 : }
649 :
650 : impl Drop for OffloadedTimeline {
651 1 : fn drop(&mut self) {
652 1 : if !self.deleted_from_ancestor.load(Ordering::Acquire) {
653 0 : tracing::warn!(
654 0 : "offloaded timeline {} was dropped without having cleaned it up at the ancestor",
655 : self.timeline_id
656 : );
657 1 : }
658 1 : }
659 : }
660 :
661 : #[derive(Copy, Clone, PartialEq, Eq, Hash, Debug)]
662 : pub enum MaybeOffloaded {
663 : Yes,
664 : No,
665 : }
666 :
667 : #[derive(Clone, Debug)]
668 : pub enum TimelineOrOffloaded {
669 : Timeline(Arc<Timeline>),
670 : Offloaded(Arc<OffloadedTimeline>),
671 : }
672 :
673 : impl TimelineOrOffloaded {
674 0 : pub fn arc_ref(&self) -> TimelineOrOffloadedArcRef<'_> {
675 0 : match self {
676 0 : TimelineOrOffloaded::Timeline(timeline) => {
677 0 : TimelineOrOffloadedArcRef::Timeline(timeline)
678 : }
679 0 : TimelineOrOffloaded::Offloaded(offloaded) => {
680 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded)
681 : }
682 : }
683 0 : }
684 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
685 0 : self.arc_ref().tenant_shard_id()
686 0 : }
687 0 : pub fn timeline_id(&self) -> TimelineId {
688 0 : self.arc_ref().timeline_id()
689 0 : }
690 1 : pub fn delete_progress(&self) -> &Arc<tokio::sync::Mutex<DeleteTimelineFlow>> {
691 1 : match self {
692 1 : TimelineOrOffloaded::Timeline(timeline) => &timeline.delete_progress,
693 0 : TimelineOrOffloaded::Offloaded(offloaded) => &offloaded.delete_progress,
694 : }
695 1 : }
696 0 : fn maybe_remote_client(&self) -> Option<Arc<RemoteTimelineClient>> {
697 0 : match self {
698 0 : TimelineOrOffloaded::Timeline(timeline) => Some(timeline.remote_client.clone()),
699 0 : TimelineOrOffloaded::Offloaded(_offloaded) => None,
700 : }
701 0 : }
702 : }
703 :
704 : pub enum TimelineOrOffloadedArcRef<'a> {
705 : Timeline(&'a Arc<Timeline>),
706 : Offloaded(&'a Arc<OffloadedTimeline>),
707 : }
708 :
709 : impl TimelineOrOffloadedArcRef<'_> {
710 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
711 0 : match self {
712 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.tenant_shard_id,
713 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.tenant_shard_id,
714 : }
715 0 : }
716 0 : pub fn timeline_id(&self) -> TimelineId {
717 0 : match self {
718 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.timeline_id,
719 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.timeline_id,
720 : }
721 0 : }
722 : }
723 :
724 : impl<'a> From<&'a Arc<Timeline>> for TimelineOrOffloadedArcRef<'a> {
725 0 : fn from(timeline: &'a Arc<Timeline>) -> Self {
726 0 : Self::Timeline(timeline)
727 0 : }
728 : }
729 :
730 : impl<'a> From<&'a Arc<OffloadedTimeline>> for TimelineOrOffloadedArcRef<'a> {
731 0 : fn from(timeline: &'a Arc<OffloadedTimeline>) -> Self {
732 0 : Self::Offloaded(timeline)
733 0 : }
734 : }
735 :
736 : #[derive(Debug, thiserror::Error, PartialEq, Eq)]
737 : pub enum GetTimelineError {
738 : #[error("Timeline is shutting down")]
739 : ShuttingDown,
740 : #[error("Timeline {tenant_id}/{timeline_id} is not active, state: {state:?}")]
741 : NotActive {
742 : tenant_id: TenantShardId,
743 : timeline_id: TimelineId,
744 : state: TimelineState,
745 : },
746 : #[error("Timeline {tenant_id}/{timeline_id} was not found")]
747 : NotFound {
748 : tenant_id: TenantShardId,
749 : timeline_id: TimelineId,
750 : },
751 : }
752 :
753 : #[derive(Debug, thiserror::Error)]
754 : pub enum LoadLocalTimelineError {
755 : #[error("FailedToLoad")]
756 : Load(#[source] anyhow::Error),
757 : #[error("FailedToResumeDeletion")]
758 : ResumeDeletion(#[source] anyhow::Error),
759 : }
760 :
761 : #[derive(thiserror::Error)]
762 : pub enum DeleteTimelineError {
763 : #[error("NotFound")]
764 : NotFound,
765 :
766 : #[error("HasChildren")]
767 : HasChildren(Vec<TimelineId>),
768 :
769 : #[error("Timeline deletion is already in progress")]
770 : AlreadyInProgress(Arc<tokio::sync::Mutex<DeleteTimelineFlow>>),
771 :
772 : #[error("Cancelled")]
773 : Cancelled,
774 :
775 : #[error(transparent)]
776 : Other(#[from] anyhow::Error),
777 : }
778 :
779 : impl Debug for DeleteTimelineError {
780 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
781 0 : match self {
782 0 : Self::NotFound => write!(f, "NotFound"),
783 0 : Self::HasChildren(c) => f.debug_tuple("HasChildren").field(c).finish(),
784 0 : Self::AlreadyInProgress(_) => f.debug_tuple("AlreadyInProgress").finish(),
785 0 : Self::Cancelled => f.debug_tuple("Cancelled").finish(),
786 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
787 : }
788 0 : }
789 : }
790 :
791 : #[derive(thiserror::Error)]
792 : pub enum TimelineArchivalError {
793 : #[error("NotFound")]
794 : NotFound,
795 :
796 : #[error("Timeout")]
797 : Timeout,
798 :
799 : #[error("Cancelled")]
800 : Cancelled,
801 :
802 : #[error("ancestor is archived: {}", .0)]
803 : HasArchivedParent(TimelineId),
804 :
805 : #[error("HasUnarchivedChildren")]
806 : HasUnarchivedChildren(Vec<TimelineId>),
807 :
808 : #[error("Timeline archival is already in progress")]
809 : AlreadyInProgress,
810 :
811 : #[error(transparent)]
812 : Other(anyhow::Error),
813 : }
814 :
815 : #[derive(thiserror::Error, Debug)]
816 : pub(crate) enum TenantManifestError {
817 : #[error("Remote storage error: {0}")]
818 : RemoteStorage(anyhow::Error),
819 :
820 : #[error("Cancelled")]
821 : Cancelled,
822 : }
823 :
824 : impl From<TenantManifestError> for TimelineArchivalError {
825 0 : fn from(e: TenantManifestError) -> Self {
826 0 : match e {
827 0 : TenantManifestError::RemoteStorage(e) => Self::Other(e),
828 0 : TenantManifestError::Cancelled => Self::Cancelled,
829 : }
830 0 : }
831 : }
832 :
833 : impl Debug for TimelineArchivalError {
834 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
835 0 : match self {
836 0 : Self::NotFound => write!(f, "NotFound"),
837 0 : Self::Timeout => write!(f, "Timeout"),
838 0 : Self::Cancelled => write!(f, "Cancelled"),
839 0 : Self::HasArchivedParent(p) => f.debug_tuple("HasArchivedParent").field(p).finish(),
840 0 : Self::HasUnarchivedChildren(c) => {
841 0 : f.debug_tuple("HasUnarchivedChildren").field(c).finish()
842 : }
843 0 : Self::AlreadyInProgress => f.debug_tuple("AlreadyInProgress").finish(),
844 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
845 : }
846 0 : }
847 : }
848 :
849 : pub enum SetStoppingError {
850 : AlreadyStopping(completion::Barrier),
851 : Broken,
852 : }
853 :
854 : impl Debug for SetStoppingError {
855 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
856 0 : match self {
857 0 : Self::AlreadyStopping(_) => f.debug_tuple("AlreadyStopping").finish(),
858 0 : Self::Broken => write!(f, "Broken"),
859 : }
860 0 : }
861 : }
862 :
863 : /// Arguments to [`TenantShard::create_timeline`].
864 : ///
865 : /// Not usable as an idempotency key for timeline creation because if [`CreateTimelineParamsBranch::ancestor_start_lsn`]
866 : /// is `None`, the result of the timeline create call is not deterministic.
867 : ///
868 : /// See [`CreateTimelineIdempotency`] for an idempotency key.
869 : #[derive(Debug)]
870 : pub(crate) enum CreateTimelineParams {
871 : Bootstrap(CreateTimelineParamsBootstrap),
872 : Branch(CreateTimelineParamsBranch),
873 : ImportPgdata(CreateTimelineParamsImportPgdata),
874 : }
875 :
876 : #[derive(Debug)]
877 : pub(crate) struct CreateTimelineParamsBootstrap {
878 : pub(crate) new_timeline_id: TimelineId,
879 : pub(crate) existing_initdb_timeline_id: Option<TimelineId>,
880 : pub(crate) pg_version: u32,
881 : }
882 :
883 : /// NB: See comment on [`CreateTimelineIdempotency::Branch`] for why there's no `pg_version` here.
884 : #[derive(Debug)]
885 : pub(crate) struct CreateTimelineParamsBranch {
886 : pub(crate) new_timeline_id: TimelineId,
887 : pub(crate) ancestor_timeline_id: TimelineId,
888 : pub(crate) ancestor_start_lsn: Option<Lsn>,
889 : }
890 :
891 : #[derive(Debug)]
892 : pub(crate) struct CreateTimelineParamsImportPgdata {
893 : pub(crate) new_timeline_id: TimelineId,
894 : pub(crate) location: import_pgdata::index_part_format::Location,
895 : pub(crate) idempotency_key: import_pgdata::index_part_format::IdempotencyKey,
896 : }
897 :
898 : /// What is used to determine idempotency of a [`TenantShard::create_timeline`] call in [`TenantShard::start_creating_timeline`] in [`TenantShard::start_creating_timeline`].
899 : ///
900 : /// Each [`Timeline`] object holds [`Self`] as an immutable property in [`Timeline::create_idempotency`].
901 : ///
902 : /// We lower timeline creation requests to [`Self`], and then use [`PartialEq::eq`] to compare [`Timeline::create_idempotency`] with the request.
903 : /// If they are equal, we return a reference to the existing timeline, otherwise it's an idempotency conflict.
904 : ///
905 : /// There is special treatment for [`Self::FailWithConflict`] to always return an idempotency conflict.
906 : /// It would be nice to have more advanced derive macros to make that special treatment declarative.
907 : ///
908 : /// Notes:
909 : /// - Unlike [`CreateTimelineParams`], ancestor LSN is fixed, so, branching will be at a deterministic LSN.
910 : /// - We make some trade-offs though, e.g., [`CreateTimelineParamsBootstrap::existing_initdb_timeline_id`]
911 : /// is not considered for idempotency. We can improve on this over time if we deem it necessary.
912 : ///
913 : #[derive(Debug, Clone, PartialEq, Eq)]
914 : pub(crate) enum CreateTimelineIdempotency {
915 : /// NB: special treatment, see comment in [`Self`].
916 : FailWithConflict,
917 : Bootstrap {
918 : pg_version: u32,
919 : },
920 : /// NB: branches always have the same `pg_version` as their ancestor.
921 : /// While [`pageserver_api::models::TimelineCreateRequestMode::Branch::pg_version`]
922 : /// exists as a field, and is set by cplane, it has always been ignored by pageserver when
923 : /// determining the child branch pg_version.
924 : Branch {
925 : ancestor_timeline_id: TimelineId,
926 : ancestor_start_lsn: Lsn,
927 : },
928 : ImportPgdata(CreatingTimelineIdempotencyImportPgdata),
929 : }
930 :
931 : #[derive(Debug, Clone, PartialEq, Eq)]
932 : pub(crate) struct CreatingTimelineIdempotencyImportPgdata {
933 : idempotency_key: import_pgdata::index_part_format::IdempotencyKey,
934 : }
935 :
936 : /// What is returned by [`TenantShard::start_creating_timeline`].
937 : #[must_use]
938 : enum StartCreatingTimelineResult {
939 : CreateGuard(TimelineCreateGuard),
940 : Idempotent(Arc<Timeline>),
941 : }
942 :
943 : #[allow(clippy::large_enum_variant, reason = "TODO")]
944 : enum TimelineInitAndSyncResult {
945 : ReadyToActivate,
946 : NeedsSpawnImportPgdata(TimelineInitAndSyncNeedsSpawnImportPgdata),
947 : }
948 :
949 : #[must_use]
950 : struct TimelineInitAndSyncNeedsSpawnImportPgdata {
951 : timeline: Arc<Timeline>,
952 : import_pgdata: import_pgdata::index_part_format::Root,
953 : guard: TimelineCreateGuard,
954 : }
955 :
956 : /// What is returned by [`TenantShard::create_timeline`].
957 : enum CreateTimelineResult {
958 : Created(Arc<Timeline>),
959 : Idempotent(Arc<Timeline>),
960 : /// IMPORTANT: This [`Arc<Timeline>`] object is not in [`TenantShard::timelines`] when
961 : /// we return this result, nor will this concrete object ever be added there.
962 : /// Cf method comment on [`TenantShard::create_timeline_import_pgdata`].
963 : ImportSpawned(Arc<Timeline>),
964 : }
965 :
966 : impl CreateTimelineResult {
967 0 : fn discriminant(&self) -> &'static str {
968 0 : match self {
969 0 : Self::Created(_) => "Created",
970 0 : Self::Idempotent(_) => "Idempotent",
971 0 : Self::ImportSpawned(_) => "ImportSpawned",
972 : }
973 0 : }
974 0 : fn timeline(&self) -> &Arc<Timeline> {
975 0 : match self {
976 0 : Self::Created(t) | Self::Idempotent(t) | Self::ImportSpawned(t) => t,
977 0 : }
978 0 : }
979 : /// Unit test timelines aren't activated, test has to do it if it needs to.
980 : #[cfg(test)]
981 118 : fn into_timeline_for_test(self) -> Arc<Timeline> {
982 118 : match self {
983 118 : Self::Created(t) | Self::Idempotent(t) | Self::ImportSpawned(t) => t,
984 118 : }
985 118 : }
986 : }
987 :
988 : #[derive(thiserror::Error, Debug)]
989 : pub enum CreateTimelineError {
990 : #[error("creation of timeline with the given ID is in progress")]
991 : AlreadyCreating,
992 : #[error("timeline already exists with different parameters")]
993 : Conflict,
994 : #[error(transparent)]
995 : AncestorLsn(anyhow::Error),
996 : #[error("ancestor timeline is not active")]
997 : AncestorNotActive,
998 : #[error("ancestor timeline is archived")]
999 : AncestorArchived,
1000 : #[error("tenant shutting down")]
1001 : ShuttingDown,
1002 : #[error(transparent)]
1003 : Other(#[from] anyhow::Error),
1004 : }
1005 :
1006 : #[derive(thiserror::Error, Debug)]
1007 : pub enum InitdbError {
1008 : #[error("Operation was cancelled")]
1009 : Cancelled,
1010 : #[error(transparent)]
1011 : Other(anyhow::Error),
1012 : #[error(transparent)]
1013 : Inner(postgres_initdb::Error),
1014 : }
1015 :
1016 : enum CreateTimelineCause {
1017 : Load,
1018 : Delete,
1019 : }
1020 :
1021 : #[allow(clippy::large_enum_variant, reason = "TODO")]
1022 : enum LoadTimelineCause {
1023 : Attach,
1024 : Unoffload,
1025 : }
1026 :
1027 : #[derive(thiserror::Error, Debug)]
1028 : pub(crate) enum GcError {
1029 : // The tenant is shutting down
1030 : #[error("tenant shutting down")]
1031 : TenantCancelled,
1032 :
1033 : // The tenant is shutting down
1034 : #[error("timeline shutting down")]
1035 : TimelineCancelled,
1036 :
1037 : // The tenant is in a state inelegible to run GC
1038 : #[error("not active")]
1039 : NotActive,
1040 :
1041 : // A requested GC cutoff LSN was invalid, for example it tried to move backwards
1042 : #[error("not active")]
1043 : BadLsn { why: String },
1044 :
1045 : // A remote storage error while scheduling updates after compaction
1046 : #[error(transparent)]
1047 : Remote(anyhow::Error),
1048 :
1049 : // An error reading while calculating GC cutoffs
1050 : #[error(transparent)]
1051 : GcCutoffs(PageReconstructError),
1052 :
1053 : // If GC was invoked for a particular timeline, this error means it didn't exist
1054 : #[error("timeline not found")]
1055 : TimelineNotFound,
1056 : }
1057 :
1058 : impl From<PageReconstructError> for GcError {
1059 0 : fn from(value: PageReconstructError) -> Self {
1060 0 : match value {
1061 0 : PageReconstructError::Cancelled => Self::TimelineCancelled,
1062 0 : other => Self::GcCutoffs(other),
1063 : }
1064 0 : }
1065 : }
1066 :
1067 : impl From<NotInitialized> for GcError {
1068 0 : fn from(value: NotInitialized) -> Self {
1069 0 : match value {
1070 0 : NotInitialized::Uninitialized => GcError::Remote(value.into()),
1071 0 : NotInitialized::Stopped | NotInitialized::ShuttingDown => GcError::TimelineCancelled,
1072 : }
1073 0 : }
1074 : }
1075 :
1076 : impl From<timeline::layer_manager::Shutdown> for GcError {
1077 0 : fn from(_: timeline::layer_manager::Shutdown) -> Self {
1078 0 : GcError::TimelineCancelled
1079 0 : }
1080 : }
1081 :
1082 : #[derive(thiserror::Error, Debug)]
1083 : pub(crate) enum LoadConfigError {
1084 : #[error("TOML deserialization error: '{0}'")]
1085 : DeserializeToml(#[from] toml_edit::de::Error),
1086 :
1087 : #[error("Config not found at {0}")]
1088 : NotFound(Utf8PathBuf),
1089 : }
1090 :
1091 : impl TenantShard {
1092 : /// Yet another helper for timeline initialization.
1093 : ///
1094 : /// - Initializes the Timeline struct and inserts it into the tenant's hash map
1095 : /// - Scans the local timeline directory for layer files and builds the layer map
1096 : /// - Downloads remote index file and adds remote files to the layer map
1097 : /// - Schedules remote upload tasks for any files that are present locally but missing from remote storage.
1098 : ///
1099 : /// If the operation fails, the timeline is left in the tenant's hash map in Broken state. On success,
1100 : /// it is marked as Active.
1101 : #[allow(clippy::too_many_arguments)]
1102 3 : async fn timeline_init_and_sync(
1103 3 : self: &Arc<Self>,
1104 3 : timeline_id: TimelineId,
1105 3 : resources: TimelineResources,
1106 3 : index_part: IndexPart,
1107 3 : metadata: TimelineMetadata,
1108 3 : previous_heatmap: Option<PreviousHeatmap>,
1109 3 : ancestor: Option<Arc<Timeline>>,
1110 3 : cause: LoadTimelineCause,
1111 3 : ctx: &RequestContext,
1112 3 : ) -> anyhow::Result<TimelineInitAndSyncResult> {
1113 3 : let tenant_id = self.tenant_shard_id;
1114 3 :
1115 3 : let import_pgdata = index_part.import_pgdata.clone();
1116 3 : let idempotency = match &import_pgdata {
1117 0 : Some(import_pgdata) => {
1118 0 : CreateTimelineIdempotency::ImportPgdata(CreatingTimelineIdempotencyImportPgdata {
1119 0 : idempotency_key: import_pgdata.idempotency_key().clone(),
1120 0 : })
1121 : }
1122 : None => {
1123 3 : if metadata.ancestor_timeline().is_none() {
1124 2 : CreateTimelineIdempotency::Bootstrap {
1125 2 : pg_version: metadata.pg_version(),
1126 2 : }
1127 : } else {
1128 1 : CreateTimelineIdempotency::Branch {
1129 1 : ancestor_timeline_id: metadata.ancestor_timeline().unwrap(),
1130 1 : ancestor_start_lsn: metadata.ancestor_lsn(),
1131 1 : }
1132 : }
1133 : }
1134 : };
1135 :
1136 3 : let (timeline, _timeline_ctx) = self.create_timeline_struct(
1137 3 : timeline_id,
1138 3 : &metadata,
1139 3 : previous_heatmap,
1140 3 : ancestor.clone(),
1141 3 : resources,
1142 3 : CreateTimelineCause::Load,
1143 3 : idempotency.clone(),
1144 3 : index_part.gc_compaction.clone(),
1145 3 : index_part.rel_size_migration.clone(),
1146 3 : ctx,
1147 3 : )?;
1148 3 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
1149 3 : anyhow::ensure!(
1150 3 : disk_consistent_lsn.is_valid(),
1151 0 : "Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn"
1152 : );
1153 3 : assert_eq!(
1154 3 : disk_consistent_lsn,
1155 3 : metadata.disk_consistent_lsn(),
1156 0 : "these are used interchangeably"
1157 : );
1158 :
1159 3 : timeline.remote_client.init_upload_queue(&index_part)?;
1160 :
1161 3 : timeline
1162 3 : .load_layer_map(disk_consistent_lsn, index_part)
1163 3 : .await
1164 3 : .with_context(|| {
1165 0 : format!("Failed to load layermap for timeline {tenant_id}/{timeline_id}")
1166 3 : })?;
1167 :
1168 : // When unarchiving, we've mostly likely lost the heatmap generated prior
1169 : // to the archival operation. To allow warming this timeline up, generate
1170 : // a previous heatmap which contains all visible layers in the layer map.
1171 : // This previous heatmap will be used whenever a fresh heatmap is generated
1172 : // for the timeline.
1173 3 : if self.conf.generate_unarchival_heatmap && matches!(cause, LoadTimelineCause::Unoffload) {
1174 0 : let mut tline_ending_at = Some((&timeline, timeline.get_last_record_lsn()));
1175 0 : while let Some((tline, end_lsn)) = tline_ending_at {
1176 0 : let unarchival_heatmap = tline.generate_unarchival_heatmap(end_lsn).await;
1177 : // Another unearchived timeline might have generated a heatmap for this ancestor.
1178 : // If the current branch point greater than the previous one use the the heatmap
1179 : // we just generated - it should include more layers.
1180 0 : if !tline.should_keep_previous_heatmap(end_lsn) {
1181 0 : tline
1182 0 : .previous_heatmap
1183 0 : .store(Some(Arc::new(unarchival_heatmap)));
1184 0 : } else {
1185 0 : tracing::info!("Previous heatmap preferred. Dropping unarchival heatmap.")
1186 : }
1187 :
1188 0 : match tline.ancestor_timeline() {
1189 0 : Some(ancestor) => {
1190 0 : if ancestor.update_layer_visibility().await.is_err() {
1191 : // Ancestor timeline is shutting down.
1192 0 : break;
1193 0 : }
1194 0 :
1195 0 : tline_ending_at = Some((ancestor, tline.get_ancestor_lsn()));
1196 : }
1197 0 : None => {
1198 0 : tline_ending_at = None;
1199 0 : }
1200 : }
1201 : }
1202 3 : }
1203 :
1204 0 : match import_pgdata {
1205 0 : Some(import_pgdata) if !import_pgdata.is_done() => {
1206 0 : let mut guard = self.timelines_creating.lock().unwrap();
1207 0 : if !guard.insert(timeline_id) {
1208 : // We should never try and load the same timeline twice during startup
1209 0 : unreachable!("Timeline {tenant_id}/{timeline_id} is already being created")
1210 0 : }
1211 0 : let timeline_create_guard = TimelineCreateGuard {
1212 0 : _tenant_gate_guard: self.gate.enter()?,
1213 0 : owning_tenant: self.clone(),
1214 0 : timeline_id,
1215 0 : idempotency,
1216 0 : // The users of this specific return value don't need the timline_path in there.
1217 0 : timeline_path: timeline
1218 0 : .conf
1219 0 : .timeline_path(&timeline.tenant_shard_id, &timeline.timeline_id),
1220 0 : };
1221 0 : Ok(TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
1222 0 : TimelineInitAndSyncNeedsSpawnImportPgdata {
1223 0 : timeline,
1224 0 : import_pgdata,
1225 0 : guard: timeline_create_guard,
1226 0 : },
1227 0 : ))
1228 : }
1229 : Some(_) | None => {
1230 : {
1231 3 : let mut timelines_accessor = self.timelines.lock().unwrap();
1232 3 : match timelines_accessor.entry(timeline_id) {
1233 : // We should never try and load the same timeline twice during startup
1234 : Entry::Occupied(_) => {
1235 0 : unreachable!(
1236 0 : "Timeline {tenant_id}/{timeline_id} already exists in the tenant map"
1237 0 : );
1238 : }
1239 3 : Entry::Vacant(v) => {
1240 3 : v.insert(Arc::clone(&timeline));
1241 3 : timeline.maybe_spawn_flush_loop();
1242 3 : }
1243 : }
1244 : }
1245 :
1246 : // Sanity check: a timeline should have some content.
1247 3 : anyhow::ensure!(
1248 3 : ancestor.is_some()
1249 2 : || timeline
1250 2 : .layers
1251 2 : .read()
1252 2 : .await
1253 2 : .layer_map()
1254 2 : .expect("currently loading, layer manager cannot be shutdown already")
1255 2 : .iter_historic_layers()
1256 2 : .next()
1257 2 : .is_some(),
1258 0 : "Timeline has no ancestor and no layer files"
1259 : );
1260 :
1261 3 : Ok(TimelineInitAndSyncResult::ReadyToActivate)
1262 : }
1263 : }
1264 3 : }
1265 :
1266 : /// Attach a tenant that's available in cloud storage.
1267 : ///
1268 : /// This returns quickly, after just creating the in-memory object
1269 : /// Tenant struct and launching a background task to download
1270 : /// the remote index files. On return, the tenant is most likely still in
1271 : /// Attaching state, and it will become Active once the background task
1272 : /// finishes. You can use wait_until_active() to wait for the task to
1273 : /// complete.
1274 : ///
1275 : #[allow(clippy::too_many_arguments)]
1276 0 : pub(crate) fn spawn(
1277 0 : conf: &'static PageServerConf,
1278 0 : tenant_shard_id: TenantShardId,
1279 0 : resources: TenantSharedResources,
1280 0 : attached_conf: AttachedTenantConf,
1281 0 : shard_identity: ShardIdentity,
1282 0 : init_order: Option<InitializationOrder>,
1283 0 : mode: SpawnMode,
1284 0 : ctx: &RequestContext,
1285 0 : ) -> Result<Arc<TenantShard>, GlobalShutDown> {
1286 0 : let wal_redo_manager =
1287 0 : WalRedoManager::new(PostgresRedoManager::new(conf, tenant_shard_id))?;
1288 :
1289 : let TenantSharedResources {
1290 0 : broker_client,
1291 0 : remote_storage,
1292 0 : deletion_queue_client,
1293 0 : l0_flush_global_state,
1294 0 : basebackup_prepare_sender,
1295 0 : } = resources;
1296 0 :
1297 0 : let attach_mode = attached_conf.location.attach_mode;
1298 0 : let generation = attached_conf.location.generation;
1299 0 :
1300 0 : let tenant = Arc::new(TenantShard::new(
1301 0 : TenantState::Attaching,
1302 0 : conf,
1303 0 : attached_conf,
1304 0 : shard_identity,
1305 0 : Some(wal_redo_manager),
1306 0 : tenant_shard_id,
1307 0 : remote_storage.clone(),
1308 0 : deletion_queue_client,
1309 0 : l0_flush_global_state,
1310 0 : basebackup_prepare_sender,
1311 0 : ));
1312 0 :
1313 0 : // The attach task will carry a GateGuard, so that shutdown() reliably waits for it to drop out if
1314 0 : // we shut down while attaching.
1315 0 : let attach_gate_guard = tenant
1316 0 : .gate
1317 0 : .enter()
1318 0 : .expect("We just created the TenantShard: nothing else can have shut it down yet");
1319 0 :
1320 0 : // Do all the hard work in the background
1321 0 : let tenant_clone = Arc::clone(&tenant);
1322 0 : let ctx = ctx.detached_child(TaskKind::Attach, DownloadBehavior::Warn);
1323 0 : task_mgr::spawn(
1324 0 : &tokio::runtime::Handle::current(),
1325 0 : TaskKind::Attach,
1326 0 : tenant_shard_id,
1327 0 : None,
1328 0 : "attach tenant",
1329 0 : async move {
1330 0 :
1331 0 : info!(
1332 : ?attach_mode,
1333 0 : "Attaching tenant"
1334 : );
1335 :
1336 0 : let _gate_guard = attach_gate_guard;
1337 0 :
1338 0 : // Is this tenant being spawned as part of process startup?
1339 0 : let starting_up = init_order.is_some();
1340 0 : scopeguard::defer! {
1341 0 : if starting_up {
1342 0 : TENANT.startup_complete.inc();
1343 0 : }
1344 0 : }
1345 :
1346 0 : fn make_broken_or_stopping(t: &TenantShard, err: anyhow::Error) {
1347 0 : t.state.send_modify(|state| match state {
1348 : // TODO: the old code alluded to DeleteTenantFlow sometimes setting
1349 : // TenantState::Stopping before we get here, but this may be outdated.
1350 : // Let's find out with a testing assertion. If this doesn't fire, and the
1351 : // logs don't show this happening in production, remove the Stopping cases.
1352 0 : TenantState::Stopping{..} if cfg!(any(test, feature = "testing")) => {
1353 0 : panic!("unexpected TenantState::Stopping during attach")
1354 : }
1355 : // If the tenant is cancelled, assume the error was caused by cancellation.
1356 0 : TenantState::Attaching if t.cancel.is_cancelled() => {
1357 0 : info!("attach cancelled, setting tenant state to Stopping: {err}");
1358 : // NB: progress None tells `set_stopping` that attach has cancelled.
1359 0 : *state = TenantState::Stopping { progress: None };
1360 : }
1361 : // According to the old code, DeleteTenantFlow may already have set this to
1362 : // Stopping. Retain its progress.
1363 : // TODO: there is no DeleteTenantFlow. Is this still needed? See above.
1364 0 : TenantState::Stopping { progress } if t.cancel.is_cancelled() => {
1365 0 : assert!(progress.is_some(), "concurrent attach cancellation");
1366 0 : info!("attach cancelled, already Stopping: {err}");
1367 : }
1368 : // Mark the tenant as broken.
1369 : TenantState::Attaching | TenantState::Stopping { .. } => {
1370 0 : error!("attach failed, setting tenant state to Broken (was {state}): {err:?}");
1371 0 : *state = TenantState::broken_from_reason(err.to_string())
1372 : }
1373 : // The attach task owns the tenant state until activated.
1374 0 : state => panic!("invalid tenant state {state} during attach: {err:?}"),
1375 0 : });
1376 0 : }
1377 :
1378 : // TODO: should also be rejecting tenant conf changes that violate this check.
1379 0 : if let Err(e) = crate::tenant::storage_layer::inmemory_layer::IndexEntry::validate_checkpoint_distance(tenant_clone.get_checkpoint_distance()) {
1380 0 : make_broken_or_stopping(&tenant_clone, anyhow::anyhow!(e));
1381 0 : return Ok(());
1382 0 : }
1383 0 :
1384 0 : let mut init_order = init_order;
1385 0 : // take the completion because initial tenant loading will complete when all of
1386 0 : // these tasks complete.
1387 0 : let _completion = init_order
1388 0 : .as_mut()
1389 0 : .and_then(|x| x.initial_tenant_load.take());
1390 0 : let remote_load_completion = init_order
1391 0 : .as_mut()
1392 0 : .and_then(|x| x.initial_tenant_load_remote.take());
1393 :
1394 : enum AttachType<'a> {
1395 : /// We are attaching this tenant lazily in the background.
1396 : Warmup {
1397 : _permit: tokio::sync::SemaphorePermit<'a>,
1398 : during_startup: bool
1399 : },
1400 : /// We are attaching this tenant as soon as we can, because for example an
1401 : /// endpoint tried to access it.
1402 : OnDemand,
1403 : /// During normal operations after startup, we are attaching a tenant, and
1404 : /// eager attach was requested.
1405 : Normal,
1406 : }
1407 :
1408 0 : let attach_type = if matches!(mode, SpawnMode::Lazy) {
1409 : // Before doing any I/O, wait for at least one of:
1410 : // - A client attempting to access to this tenant (on-demand loading)
1411 : // - A permit becoming available in the warmup semaphore (background warmup)
1412 :
1413 0 : tokio::select!(
1414 0 : permit = tenant_clone.activate_now_sem.acquire() => {
1415 0 : let _ = permit.expect("activate_now_sem is never closed");
1416 0 : tracing::info!("Activating tenant (on-demand)");
1417 0 : AttachType::OnDemand
1418 : },
1419 0 : permit = conf.concurrent_tenant_warmup.inner().acquire() => {
1420 0 : let _permit = permit.expect("concurrent_tenant_warmup semaphore is never closed");
1421 0 : tracing::info!("Activating tenant (warmup)");
1422 0 : AttachType::Warmup {
1423 0 : _permit,
1424 0 : during_startup: init_order.is_some()
1425 0 : }
1426 : }
1427 0 : _ = tenant_clone.cancel.cancelled() => {
1428 : // This is safe, but should be pretty rare: it is interesting if a tenant
1429 : // stayed in Activating for such a long time that shutdown found it in
1430 : // that state.
1431 0 : tracing::info!(state=%tenant_clone.current_state(), "Tenant shut down before activation");
1432 : // Set the tenant to Stopping to signal `set_stopping` that we're done.
1433 0 : make_broken_or_stopping(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"));
1434 0 : return Ok(());
1435 : },
1436 : )
1437 : } else {
1438 : // SpawnMode::{Create,Eager} always cause jumping ahead of the
1439 : // concurrent_tenant_warmup queue
1440 0 : AttachType::Normal
1441 : };
1442 :
1443 0 : let preload = match &mode {
1444 : SpawnMode::Eager | SpawnMode::Lazy => {
1445 0 : let _preload_timer = TENANT.preload.start_timer();
1446 0 : let res = tenant_clone
1447 0 : .preload(&remote_storage, task_mgr::shutdown_token())
1448 0 : .await;
1449 0 : match res {
1450 0 : Ok(p) => Some(p),
1451 0 : Err(e) => {
1452 0 : make_broken_or_stopping(&tenant_clone, anyhow::anyhow!(e));
1453 0 : return Ok(());
1454 : }
1455 : }
1456 : }
1457 :
1458 : };
1459 :
1460 : // Remote preload is complete.
1461 0 : drop(remote_load_completion);
1462 0 :
1463 0 :
1464 0 : // We will time the duration of the attach phase unless this is a creation (attach will do no work)
1465 0 : let attach_start = std::time::Instant::now();
1466 0 : let attached = {
1467 0 : let _attach_timer = Some(TENANT.attach.start_timer());
1468 0 : tenant_clone.attach(preload, &ctx).await
1469 : };
1470 0 : let attach_duration = attach_start.elapsed();
1471 0 : _ = tenant_clone.attach_wal_lag_cooldown.set(WalLagCooldown::new(attach_start, attach_duration));
1472 0 :
1473 0 : match attached {
1474 : Ok(()) => {
1475 0 : info!("attach finished, activating");
1476 0 : tenant_clone.activate(broker_client, None, &ctx);
1477 : }
1478 0 : Err(e) => make_broken_or_stopping(&tenant_clone, anyhow::anyhow!(e)),
1479 : }
1480 :
1481 : // If we are doing an opportunistic warmup attachment at startup, initialize
1482 : // logical size at the same time. This is better than starting a bunch of idle tenants
1483 : // with cold caches and then coming back later to initialize their logical sizes.
1484 : //
1485 : // It also prevents the warmup proccess competing with the concurrency limit on
1486 : // logical size calculations: if logical size calculation semaphore is saturated,
1487 : // then warmup will wait for that before proceeding to the next tenant.
1488 0 : if matches!(attach_type, AttachType::Warmup { during_startup: true, .. }) {
1489 0 : let mut futs: FuturesUnordered<_> = tenant_clone.timelines.lock().unwrap().values().cloned().map(|t| t.await_initial_logical_size()).collect();
1490 0 : tracing::info!("Waiting for initial logical sizes while warming up...");
1491 0 : while futs.next().await.is_some() {}
1492 0 : tracing::info!("Warm-up complete");
1493 0 : }
1494 :
1495 0 : Ok(())
1496 0 : }
1497 0 : .instrument(tracing::info_span!(parent: None, "attach", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), gen=?generation)),
1498 : );
1499 0 : Ok(tenant)
1500 0 : }
1501 :
1502 : #[instrument(skip_all)]
1503 : pub(crate) async fn preload(
1504 : self: &Arc<Self>,
1505 : remote_storage: &GenericRemoteStorage,
1506 : cancel: CancellationToken,
1507 : ) -> anyhow::Result<TenantPreload> {
1508 : span::debug_assert_current_span_has_tenant_id();
1509 : // Get list of remote timelines
1510 : // download index files for every tenant timeline
1511 : info!("listing remote timelines");
1512 : let (mut remote_timeline_ids, other_keys) = remote_timeline_client::list_remote_timelines(
1513 : remote_storage,
1514 : self.tenant_shard_id,
1515 : cancel.clone(),
1516 : )
1517 : .await?;
1518 :
1519 : let tenant_manifest = match download_tenant_manifest(
1520 : remote_storage,
1521 : &self.tenant_shard_id,
1522 : self.generation,
1523 : &cancel,
1524 : )
1525 : .await
1526 : {
1527 : Ok((tenant_manifest, _, _)) => Some(tenant_manifest),
1528 : Err(DownloadError::NotFound) => None,
1529 : Err(err) => return Err(err.into()),
1530 : };
1531 :
1532 : info!(
1533 : "found {} timelines ({} offloaded timelines)",
1534 : remote_timeline_ids.len(),
1535 : tenant_manifest
1536 : .as_ref()
1537 3 : .map(|m| m.offloaded_timelines.len())
1538 : .unwrap_or(0)
1539 : );
1540 :
1541 : for k in other_keys {
1542 : warn!("Unexpected non timeline key {k}");
1543 : }
1544 :
1545 : // Avoid downloading IndexPart of offloaded timelines.
1546 : let mut offloaded_with_prefix = HashSet::new();
1547 : if let Some(tenant_manifest) = &tenant_manifest {
1548 : for offloaded in tenant_manifest.offloaded_timelines.iter() {
1549 : if remote_timeline_ids.remove(&offloaded.timeline_id) {
1550 : offloaded_with_prefix.insert(offloaded.timeline_id);
1551 : } else {
1552 : // We'll take care later of timelines in the manifest without a prefix
1553 : }
1554 : }
1555 : }
1556 :
1557 : // TODO(vlad): Could go to S3 if the secondary is freezing cold and hasn't even
1558 : // pulled the first heatmap. Not entirely necessary since the storage controller
1559 : // will kick the secondary in any case and cause a download.
1560 : let maybe_heatmap_at = self.read_on_disk_heatmap().await;
1561 :
1562 : let timelines = self
1563 : .load_timelines_metadata(
1564 : remote_timeline_ids,
1565 : remote_storage,
1566 : maybe_heatmap_at,
1567 : cancel,
1568 : )
1569 : .await?;
1570 :
1571 : Ok(TenantPreload {
1572 : tenant_manifest,
1573 : timelines: timelines
1574 : .into_iter()
1575 3 : .map(|(id, tl)| (id, Some(tl)))
1576 0 : .chain(offloaded_with_prefix.into_iter().map(|id| (id, None)))
1577 : .collect(),
1578 : })
1579 : }
1580 :
1581 117 : async fn read_on_disk_heatmap(&self) -> Option<(HeatMapTenant, std::time::Instant)> {
1582 117 : if !self.conf.load_previous_heatmap {
1583 0 : return None;
1584 117 : }
1585 117 :
1586 117 : let on_disk_heatmap_path = self.conf.tenant_heatmap_path(&self.tenant_shard_id);
1587 117 : match tokio::fs::read_to_string(on_disk_heatmap_path).await {
1588 0 : Ok(heatmap) => match serde_json::from_str::<HeatMapTenant>(&heatmap) {
1589 0 : Ok(heatmap) => Some((heatmap, std::time::Instant::now())),
1590 0 : Err(err) => {
1591 0 : error!("Failed to deserialize old heatmap: {err}");
1592 0 : None
1593 : }
1594 : },
1595 117 : Err(err) => match err.kind() {
1596 117 : std::io::ErrorKind::NotFound => None,
1597 : _ => {
1598 0 : error!("Unexpected IO error reading old heatmap: {err}");
1599 0 : None
1600 : }
1601 : },
1602 : }
1603 117 : }
1604 :
1605 : ///
1606 : /// Background task that downloads all data for a tenant and brings it to Active state.
1607 : ///
1608 : /// No background tasks are started as part of this routine.
1609 : ///
1610 117 : async fn attach(
1611 117 : self: &Arc<TenantShard>,
1612 117 : preload: Option<TenantPreload>,
1613 117 : ctx: &RequestContext,
1614 117 : ) -> anyhow::Result<()> {
1615 117 : span::debug_assert_current_span_has_tenant_id();
1616 117 :
1617 117 : failpoint_support::sleep_millis_async!("before-attaching-tenant");
1618 :
1619 117 : let Some(preload) = preload else {
1620 0 : anyhow::bail!(
1621 0 : "local-only deployment is no longer supported, https://github.com/neondatabase/neon/issues/5624"
1622 0 : );
1623 : };
1624 :
1625 117 : let mut offloaded_timeline_ids = HashSet::new();
1626 117 : let mut offloaded_timelines_list = Vec::new();
1627 117 : if let Some(tenant_manifest) = &preload.tenant_manifest {
1628 3 : for timeline_manifest in tenant_manifest.offloaded_timelines.iter() {
1629 0 : let timeline_id = timeline_manifest.timeline_id;
1630 0 : let offloaded_timeline =
1631 0 : OffloadedTimeline::from_manifest(self.tenant_shard_id, timeline_manifest);
1632 0 : offloaded_timelines_list.push((timeline_id, Arc::new(offloaded_timeline)));
1633 0 : offloaded_timeline_ids.insert(timeline_id);
1634 0 : }
1635 114 : }
1636 : // Complete deletions for offloaded timeline id's from manifest.
1637 : // The manifest will be uploaded later in this function.
1638 117 : offloaded_timelines_list
1639 117 : .retain(|(offloaded_id, offloaded)| {
1640 0 : // Existence of a timeline is finally determined by the existence of an index-part.json in remote storage.
1641 0 : // If there is dangling references in another location, they need to be cleaned up.
1642 0 : let delete = !preload.timelines.contains_key(offloaded_id);
1643 0 : if delete {
1644 0 : tracing::info!("Removing offloaded timeline {offloaded_id} from manifest as no remote prefix was found");
1645 0 : offloaded.defuse_for_tenant_drop();
1646 0 : }
1647 0 : !delete
1648 117 : });
1649 117 :
1650 117 : let mut timelines_to_resume_deletions = vec![];
1651 117 :
1652 117 : let mut remote_index_and_client = HashMap::new();
1653 117 : let mut timeline_ancestors = HashMap::new();
1654 117 : let mut existent_timelines = HashSet::new();
1655 120 : for (timeline_id, preload) in preload.timelines {
1656 3 : let Some(preload) = preload else { continue };
1657 : // This is an invariant of the `preload` function's API
1658 3 : assert!(!offloaded_timeline_ids.contains(&timeline_id));
1659 3 : let index_part = match preload.index_part {
1660 3 : Ok(i) => {
1661 3 : debug!("remote index part exists for timeline {timeline_id}");
1662 : // We found index_part on the remote, this is the standard case.
1663 3 : existent_timelines.insert(timeline_id);
1664 3 : i
1665 : }
1666 : Err(DownloadError::NotFound) => {
1667 : // There is no index_part on the remote. We only get here
1668 : // if there is some prefix for the timeline in the remote storage.
1669 : // This can e.g. be the initdb.tar.zst archive, maybe a
1670 : // remnant from a prior incomplete creation or deletion attempt.
1671 : // Delete the local directory as the deciding criterion for a
1672 : // timeline's existence is presence of index_part.
1673 0 : info!(%timeline_id, "index_part not found on remote");
1674 0 : continue;
1675 : }
1676 0 : Err(DownloadError::Fatal(why)) => {
1677 0 : // If, while loading one remote timeline, we saw an indication that our generation
1678 0 : // number is likely invalid, then we should not load the whole tenant.
1679 0 : error!(%timeline_id, "Fatal error loading timeline: {why}");
1680 0 : anyhow::bail!(why.to_string());
1681 : }
1682 0 : Err(e) => {
1683 0 : // Some (possibly ephemeral) error happened during index_part download.
1684 0 : // Pretend the timeline exists to not delete the timeline directory,
1685 0 : // as it might be a temporary issue and we don't want to re-download
1686 0 : // everything after it resolves.
1687 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1688 :
1689 0 : existent_timelines.insert(timeline_id);
1690 0 : continue;
1691 : }
1692 : };
1693 3 : match index_part {
1694 3 : MaybeDeletedIndexPart::IndexPart(index_part) => {
1695 3 : timeline_ancestors.insert(timeline_id, index_part.metadata.clone());
1696 3 : remote_index_and_client.insert(
1697 3 : timeline_id,
1698 3 : (index_part, preload.client, preload.previous_heatmap),
1699 3 : );
1700 3 : }
1701 0 : MaybeDeletedIndexPart::Deleted(index_part) => {
1702 0 : info!(
1703 0 : "timeline {} is deleted, picking to resume deletion",
1704 : timeline_id
1705 : );
1706 0 : timelines_to_resume_deletions.push((timeline_id, index_part, preload.client));
1707 : }
1708 : }
1709 : }
1710 :
1711 117 : let mut gc_blocks = HashMap::new();
1712 :
1713 : // For every timeline, download the metadata file, scan the local directory,
1714 : // and build a layer map that contains an entry for each remote and local
1715 : // layer file.
1716 117 : let sorted_timelines = tree_sort_timelines(timeline_ancestors, |m| m.ancestor_timeline())?;
1717 120 : for (timeline_id, remote_metadata) in sorted_timelines {
1718 3 : let (index_part, remote_client, previous_heatmap) = remote_index_and_client
1719 3 : .remove(&timeline_id)
1720 3 : .expect("just put it in above");
1721 :
1722 3 : if let Some(blocking) = index_part.gc_blocking.as_ref() {
1723 : // could just filter these away, but it helps while testing
1724 0 : anyhow::ensure!(
1725 0 : !blocking.reasons.is_empty(),
1726 0 : "index_part for {timeline_id} is malformed: it should not have gc blocking with zero reasons"
1727 : );
1728 0 : let prev = gc_blocks.insert(timeline_id, blocking.reasons);
1729 0 : assert!(prev.is_none());
1730 3 : }
1731 :
1732 : // TODO again handle early failure
1733 3 : let effect = self
1734 3 : .load_remote_timeline(
1735 3 : timeline_id,
1736 3 : index_part,
1737 3 : remote_metadata,
1738 3 : previous_heatmap,
1739 3 : self.get_timeline_resources_for(remote_client),
1740 3 : LoadTimelineCause::Attach,
1741 3 : ctx,
1742 3 : )
1743 3 : .await
1744 3 : .with_context(|| {
1745 0 : format!(
1746 0 : "failed to load remote timeline {} for tenant {}",
1747 0 : timeline_id, self.tenant_shard_id
1748 0 : )
1749 3 : })?;
1750 :
1751 3 : match effect {
1752 3 : TimelineInitAndSyncResult::ReadyToActivate => {
1753 3 : // activation happens later, on Tenant::activate
1754 3 : }
1755 : TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
1756 : TimelineInitAndSyncNeedsSpawnImportPgdata {
1757 0 : timeline,
1758 0 : import_pgdata,
1759 0 : guard,
1760 0 : },
1761 0 : ) => {
1762 0 : let timeline_id = timeline.timeline_id;
1763 0 : let import_task_handle =
1764 0 : tokio::task::spawn(self.clone().create_timeline_import_pgdata_task(
1765 0 : timeline.clone(),
1766 0 : import_pgdata,
1767 0 : guard,
1768 0 : ctx.detached_child(TaskKind::ImportPgdata, DownloadBehavior::Warn),
1769 0 : ));
1770 0 :
1771 0 : let prev = self.timelines_importing.lock().unwrap().insert(
1772 0 : timeline_id,
1773 0 : ImportingTimeline {
1774 0 : timeline: timeline.clone(),
1775 0 : import_task_handle,
1776 0 : },
1777 0 : );
1778 0 :
1779 0 : assert!(prev.is_none());
1780 : }
1781 : }
1782 : }
1783 :
1784 : // Walk through deleted timelines, resume deletion
1785 117 : for (timeline_id, index_part, remote_timeline_client) in timelines_to_resume_deletions {
1786 0 : remote_timeline_client
1787 0 : .init_upload_queue_stopped_to_continue_deletion(&index_part)
1788 0 : .context("init queue stopped")
1789 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1790 :
1791 0 : DeleteTimelineFlow::resume_deletion(
1792 0 : Arc::clone(self),
1793 0 : timeline_id,
1794 0 : &index_part.metadata,
1795 0 : remote_timeline_client,
1796 0 : ctx,
1797 0 : )
1798 0 : .instrument(tracing::info_span!("timeline_delete", %timeline_id))
1799 0 : .await
1800 0 : .context("resume_deletion")
1801 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1802 : }
1803 117 : {
1804 117 : let mut offloaded_timelines_accessor = self.timelines_offloaded.lock().unwrap();
1805 117 : offloaded_timelines_accessor.extend(offloaded_timelines_list.into_iter());
1806 117 : }
1807 :
1808 : // Stash the preloaded tenant manifest, and upload a new manifest if changed.
1809 : //
1810 : // NB: this must happen after the tenant is fully populated above. In particular the
1811 : // offloaded timelines, which are included in the manifest.
1812 : {
1813 117 : let mut guard = self.remote_tenant_manifest.lock().await;
1814 117 : assert!(guard.is_none(), "tenant manifest set before preload"); // first populated here
1815 117 : *guard = preload.tenant_manifest;
1816 117 : }
1817 117 : self.maybe_upload_tenant_manifest().await?;
1818 :
1819 : // The local filesystem contents are a cache of what's in the remote IndexPart;
1820 : // IndexPart is the source of truth.
1821 117 : self.clean_up_timelines(&existent_timelines)?;
1822 :
1823 117 : self.gc_block.set_scanned(gc_blocks);
1824 117 :
1825 117 : fail::fail_point!("attach-before-activate", |_| {
1826 0 : anyhow::bail!("attach-before-activate");
1827 117 : });
1828 117 : failpoint_support::sleep_millis_async!("attach-before-activate-sleep", &self.cancel);
1829 :
1830 117 : info!("Done");
1831 :
1832 117 : Ok(())
1833 117 : }
1834 :
1835 : /// Check for any local timeline directories that are temporary, or do not correspond to a
1836 : /// timeline that still exists: this can happen if we crashed during a deletion/creation, or
1837 : /// if a timeline was deleted while the tenant was attached to a different pageserver.
1838 117 : fn clean_up_timelines(&self, existent_timelines: &HashSet<TimelineId>) -> anyhow::Result<()> {
1839 117 : let timelines_dir = self.conf.timelines_path(&self.tenant_shard_id);
1840 :
1841 117 : let entries = match timelines_dir.read_dir_utf8() {
1842 117 : Ok(d) => d,
1843 0 : Err(e) => {
1844 0 : if e.kind() == std::io::ErrorKind::NotFound {
1845 0 : return Ok(());
1846 : } else {
1847 0 : return Err(e).context("list timelines directory for tenant");
1848 : }
1849 : }
1850 : };
1851 :
1852 121 : for entry in entries {
1853 4 : let entry = entry.context("read timeline dir entry")?;
1854 4 : let entry_path = entry.path();
1855 :
1856 4 : let purge = if crate::is_temporary(entry_path) {
1857 0 : true
1858 : } else {
1859 4 : match TimelineId::try_from(entry_path.file_name()) {
1860 4 : Ok(i) => {
1861 4 : // Purge if the timeline ID does not exist in remote storage: remote storage is the authority.
1862 4 : !existent_timelines.contains(&i)
1863 : }
1864 0 : Err(e) => {
1865 0 : tracing::warn!(
1866 0 : "Unparseable directory in timelines directory: {entry_path}, ignoring ({e})"
1867 : );
1868 : // Do not purge junk: if we don't recognize it, be cautious and leave it for a human.
1869 0 : false
1870 : }
1871 : }
1872 : };
1873 :
1874 4 : if purge {
1875 1 : tracing::info!("Purging stale timeline dentry {entry_path}");
1876 1 : if let Err(e) = match entry.file_type() {
1877 1 : Ok(t) => if t.is_dir() {
1878 1 : std::fs::remove_dir_all(entry_path)
1879 : } else {
1880 0 : std::fs::remove_file(entry_path)
1881 : }
1882 1 : .or_else(fs_ext::ignore_not_found),
1883 0 : Err(e) => Err(e),
1884 : } {
1885 0 : tracing::warn!("Failed to purge stale timeline dentry {entry_path}: {e}");
1886 1 : }
1887 3 : }
1888 : }
1889 :
1890 117 : Ok(())
1891 117 : }
1892 :
1893 : /// Get sum of all remote timelines sizes
1894 : ///
1895 : /// This function relies on the index_part instead of listing the remote storage
1896 0 : pub fn remote_size(&self) -> u64 {
1897 0 : let mut size = 0;
1898 :
1899 0 : for timeline in self.list_timelines() {
1900 0 : size += timeline.remote_client.get_remote_physical_size();
1901 0 : }
1902 :
1903 0 : size
1904 0 : }
1905 :
1906 : #[instrument(skip_all, fields(timeline_id=%timeline_id))]
1907 : #[allow(clippy::too_many_arguments)]
1908 : async fn load_remote_timeline(
1909 : self: &Arc<Self>,
1910 : timeline_id: TimelineId,
1911 : index_part: IndexPart,
1912 : remote_metadata: TimelineMetadata,
1913 : previous_heatmap: Option<PreviousHeatmap>,
1914 : resources: TimelineResources,
1915 : cause: LoadTimelineCause,
1916 : ctx: &RequestContext,
1917 : ) -> anyhow::Result<TimelineInitAndSyncResult> {
1918 : span::debug_assert_current_span_has_tenant_id();
1919 :
1920 : info!("downloading index file for timeline {}", timeline_id);
1921 : tokio::fs::create_dir_all(self.conf.timeline_path(&self.tenant_shard_id, &timeline_id))
1922 : .await
1923 : .context("Failed to create new timeline directory")?;
1924 :
1925 : let ancestor = if let Some(ancestor_id) = remote_metadata.ancestor_timeline() {
1926 : let timelines = self.timelines.lock().unwrap();
1927 : Some(Arc::clone(timelines.get(&ancestor_id).ok_or_else(
1928 0 : || {
1929 0 : anyhow::anyhow!(
1930 0 : "cannot find ancestor timeline {ancestor_id} for timeline {timeline_id}"
1931 0 : )
1932 0 : },
1933 : )?))
1934 : } else {
1935 : None
1936 : };
1937 :
1938 : self.timeline_init_and_sync(
1939 : timeline_id,
1940 : resources,
1941 : index_part,
1942 : remote_metadata,
1943 : previous_heatmap,
1944 : ancestor,
1945 : cause,
1946 : ctx,
1947 : )
1948 : .await
1949 : }
1950 :
1951 117 : async fn load_timelines_metadata(
1952 117 : self: &Arc<TenantShard>,
1953 117 : timeline_ids: HashSet<TimelineId>,
1954 117 : remote_storage: &GenericRemoteStorage,
1955 117 : heatmap: Option<(HeatMapTenant, std::time::Instant)>,
1956 117 : cancel: CancellationToken,
1957 117 : ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
1958 117 : let mut timeline_heatmaps = heatmap.map(|h| (h.0.into_timelines_index(), h.1));
1959 117 :
1960 117 : let mut part_downloads = JoinSet::new();
1961 120 : for timeline_id in timeline_ids {
1962 3 : let cancel_clone = cancel.clone();
1963 3 :
1964 3 : let previous_timeline_heatmap = timeline_heatmaps.as_mut().and_then(|hs| {
1965 0 : hs.0.remove(&timeline_id).map(|h| PreviousHeatmap::Active {
1966 0 : heatmap: h,
1967 0 : read_at: hs.1,
1968 0 : end_lsn: None,
1969 0 : })
1970 3 : });
1971 3 : part_downloads.spawn(
1972 3 : self.load_timeline_metadata(
1973 3 : timeline_id,
1974 3 : remote_storage.clone(),
1975 3 : previous_timeline_heatmap,
1976 3 : cancel_clone,
1977 3 : )
1978 3 : .instrument(info_span!("download_index_part", %timeline_id)),
1979 : );
1980 : }
1981 :
1982 117 : let mut timeline_preloads: HashMap<TimelineId, TimelinePreload> = HashMap::new();
1983 :
1984 : loop {
1985 120 : tokio::select!(
1986 120 : next = part_downloads.join_next() => {
1987 120 : match next {
1988 3 : Some(result) => {
1989 3 : let preload = result.context("join preload task")?;
1990 3 : timeline_preloads.insert(preload.timeline_id, preload);
1991 : },
1992 : None => {
1993 117 : break;
1994 : }
1995 : }
1996 : },
1997 120 : _ = cancel.cancelled() => {
1998 0 : anyhow::bail!("Cancelled while waiting for remote index download")
1999 : }
2000 : )
2001 : }
2002 :
2003 117 : Ok(timeline_preloads)
2004 117 : }
2005 :
2006 3 : fn build_timeline_client(
2007 3 : &self,
2008 3 : timeline_id: TimelineId,
2009 3 : remote_storage: GenericRemoteStorage,
2010 3 : ) -> RemoteTimelineClient {
2011 3 : RemoteTimelineClient::new(
2012 3 : remote_storage.clone(),
2013 3 : self.deletion_queue_client.clone(),
2014 3 : self.conf,
2015 3 : self.tenant_shard_id,
2016 3 : timeline_id,
2017 3 : self.generation,
2018 3 : &self.tenant_conf.load().location,
2019 3 : )
2020 3 : }
2021 :
2022 3 : fn load_timeline_metadata(
2023 3 : self: &Arc<TenantShard>,
2024 3 : timeline_id: TimelineId,
2025 3 : remote_storage: GenericRemoteStorage,
2026 3 : previous_heatmap: Option<PreviousHeatmap>,
2027 3 : cancel: CancellationToken,
2028 3 : ) -> impl Future<Output = TimelinePreload> + use<> {
2029 3 : let client = self.build_timeline_client(timeline_id, remote_storage);
2030 3 : async move {
2031 3 : debug_assert_current_span_has_tenant_and_timeline_id();
2032 3 : debug!("starting index part download");
2033 :
2034 3 : let index_part = client.download_index_file(&cancel).await;
2035 :
2036 3 : debug!("finished index part download");
2037 :
2038 3 : TimelinePreload {
2039 3 : client,
2040 3 : timeline_id,
2041 3 : index_part,
2042 3 : previous_heatmap,
2043 3 : }
2044 3 : }
2045 3 : }
2046 :
2047 0 : fn check_to_be_archived_has_no_unarchived_children(
2048 0 : timeline_id: TimelineId,
2049 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
2050 0 : ) -> Result<(), TimelineArchivalError> {
2051 0 : let children: Vec<TimelineId> = timelines
2052 0 : .iter()
2053 0 : .filter_map(|(id, entry)| {
2054 0 : if entry.get_ancestor_timeline_id() != Some(timeline_id) {
2055 0 : return None;
2056 0 : }
2057 0 : if entry.is_archived() == Some(true) {
2058 0 : return None;
2059 0 : }
2060 0 : Some(*id)
2061 0 : })
2062 0 : .collect();
2063 0 :
2064 0 : if !children.is_empty() {
2065 0 : return Err(TimelineArchivalError::HasUnarchivedChildren(children));
2066 0 : }
2067 0 : Ok(())
2068 0 : }
2069 :
2070 0 : fn check_ancestor_of_to_be_unarchived_is_not_archived(
2071 0 : ancestor_timeline_id: TimelineId,
2072 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
2073 0 : offloaded_timelines: &std::sync::MutexGuard<
2074 0 : '_,
2075 0 : HashMap<TimelineId, Arc<OffloadedTimeline>>,
2076 0 : >,
2077 0 : ) -> Result<(), TimelineArchivalError> {
2078 0 : let has_archived_parent =
2079 0 : if let Some(ancestor_timeline) = timelines.get(&ancestor_timeline_id) {
2080 0 : ancestor_timeline.is_archived() == Some(true)
2081 0 : } else if offloaded_timelines.contains_key(&ancestor_timeline_id) {
2082 0 : true
2083 : } else {
2084 0 : error!("ancestor timeline {ancestor_timeline_id} not found");
2085 0 : if cfg!(debug_assertions) {
2086 0 : panic!("ancestor timeline {ancestor_timeline_id} not found");
2087 0 : }
2088 0 : return Err(TimelineArchivalError::NotFound);
2089 : };
2090 0 : if has_archived_parent {
2091 0 : return Err(TimelineArchivalError::HasArchivedParent(
2092 0 : ancestor_timeline_id,
2093 0 : ));
2094 0 : }
2095 0 : Ok(())
2096 0 : }
2097 :
2098 0 : fn check_to_be_unarchived_timeline_has_no_archived_parent(
2099 0 : timeline: &Arc<Timeline>,
2100 0 : ) -> Result<(), TimelineArchivalError> {
2101 0 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
2102 0 : if ancestor_timeline.is_archived() == Some(true) {
2103 0 : return Err(TimelineArchivalError::HasArchivedParent(
2104 0 : ancestor_timeline.timeline_id,
2105 0 : ));
2106 0 : }
2107 0 : }
2108 0 : Ok(())
2109 0 : }
2110 :
2111 : /// Loads the specified (offloaded) timeline from S3 and attaches it as a loaded timeline
2112 : ///
2113 : /// Counterpart to [`offload_timeline`].
2114 0 : async fn unoffload_timeline(
2115 0 : self: &Arc<Self>,
2116 0 : timeline_id: TimelineId,
2117 0 : broker_client: storage_broker::BrokerClientChannel,
2118 0 : ctx: RequestContext,
2119 0 : ) -> Result<Arc<Timeline>, TimelineArchivalError> {
2120 0 : info!("unoffloading timeline");
2121 :
2122 : // We activate the timeline below manually, so this must be called on an active tenant.
2123 : // We expect callers of this function to ensure this.
2124 0 : match self.current_state() {
2125 : TenantState::Activating { .. }
2126 : | TenantState::Attaching
2127 : | TenantState::Broken { .. } => {
2128 0 : panic!("Timeline expected to be active")
2129 : }
2130 0 : TenantState::Stopping { .. } => return Err(TimelineArchivalError::Cancelled),
2131 0 : TenantState::Active => {}
2132 0 : }
2133 0 : let cancel = self.cancel.clone();
2134 0 :
2135 0 : // Protect against concurrent attempts to use this TimelineId
2136 0 : // We don't care much about idempotency, as it's ensured a layer above.
2137 0 : let allow_offloaded = true;
2138 0 : let _create_guard = self
2139 0 : .create_timeline_create_guard(
2140 0 : timeline_id,
2141 0 : CreateTimelineIdempotency::FailWithConflict,
2142 0 : allow_offloaded,
2143 0 : )
2144 0 : .map_err(|err| match err {
2145 0 : TimelineExclusionError::AlreadyCreating => TimelineArchivalError::AlreadyInProgress,
2146 : TimelineExclusionError::AlreadyExists { .. } => {
2147 0 : TimelineArchivalError::Other(anyhow::anyhow!("Timeline already exists"))
2148 : }
2149 0 : TimelineExclusionError::Other(e) => TimelineArchivalError::Other(e),
2150 0 : TimelineExclusionError::ShuttingDown => TimelineArchivalError::Cancelled,
2151 0 : })?;
2152 :
2153 0 : let timeline_preload = self
2154 0 : .load_timeline_metadata(
2155 0 : timeline_id,
2156 0 : self.remote_storage.clone(),
2157 0 : None,
2158 0 : cancel.clone(),
2159 0 : )
2160 0 : .await;
2161 :
2162 0 : let index_part = match timeline_preload.index_part {
2163 0 : Ok(index_part) => {
2164 0 : debug!("remote index part exists for timeline {timeline_id}");
2165 0 : index_part
2166 : }
2167 : Err(DownloadError::NotFound) => {
2168 0 : error!(%timeline_id, "index_part not found on remote");
2169 0 : return Err(TimelineArchivalError::NotFound);
2170 : }
2171 0 : Err(DownloadError::Cancelled) => return Err(TimelineArchivalError::Cancelled),
2172 0 : Err(e) => {
2173 0 : // Some (possibly ephemeral) error happened during index_part download.
2174 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
2175 0 : return Err(TimelineArchivalError::Other(
2176 0 : anyhow::Error::new(e).context("downloading index_part from remote storage"),
2177 0 : ));
2178 : }
2179 : };
2180 0 : let index_part = match index_part {
2181 0 : MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
2182 0 : MaybeDeletedIndexPart::Deleted(_index_part) => {
2183 0 : info!("timeline is deleted according to index_part.json");
2184 0 : return Err(TimelineArchivalError::NotFound);
2185 : }
2186 : };
2187 0 : let remote_metadata = index_part.metadata.clone();
2188 0 : let timeline_resources = self.build_timeline_resources(timeline_id);
2189 0 : self.load_remote_timeline(
2190 0 : timeline_id,
2191 0 : index_part,
2192 0 : remote_metadata,
2193 0 : None,
2194 0 : timeline_resources,
2195 0 : LoadTimelineCause::Unoffload,
2196 0 : &ctx,
2197 0 : )
2198 0 : .await
2199 0 : .with_context(|| {
2200 0 : format!(
2201 0 : "failed to load remote timeline {} for tenant {}",
2202 0 : timeline_id, self.tenant_shard_id
2203 0 : )
2204 0 : })
2205 0 : .map_err(TimelineArchivalError::Other)?;
2206 :
2207 0 : let timeline = {
2208 0 : let timelines = self.timelines.lock().unwrap();
2209 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2210 0 : warn!("timeline not available directly after attach");
2211 : // This is not a panic because no locks are held between `load_remote_timeline`
2212 : // which puts the timeline into timelines, and our look into the timeline map.
2213 0 : return Err(TimelineArchivalError::Other(anyhow::anyhow!(
2214 0 : "timeline not available directly after attach"
2215 0 : )));
2216 : };
2217 0 : let mut offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2218 0 : match offloaded_timelines.remove(&timeline_id) {
2219 0 : Some(offloaded) => {
2220 0 : offloaded.delete_from_ancestor_with_timelines(&timelines);
2221 0 : }
2222 0 : None => warn!("timeline already removed from offloaded timelines"),
2223 : }
2224 :
2225 0 : self.initialize_gc_info(&timelines, &offloaded_timelines, Some(timeline_id));
2226 0 :
2227 0 : Arc::clone(timeline)
2228 0 : };
2229 0 :
2230 0 : // Upload new list of offloaded timelines to S3
2231 0 : self.maybe_upload_tenant_manifest().await?;
2232 :
2233 : // Activate the timeline (if it makes sense)
2234 0 : if !(timeline.is_broken() || timeline.is_stopping()) {
2235 0 : let background_jobs_can_start = None;
2236 0 : timeline.activate(
2237 0 : self.clone(),
2238 0 : broker_client.clone(),
2239 0 : background_jobs_can_start,
2240 0 : &ctx.with_scope_timeline(&timeline),
2241 0 : );
2242 0 : }
2243 :
2244 0 : info!("timeline unoffloading complete");
2245 0 : Ok(timeline)
2246 0 : }
2247 :
2248 0 : pub(crate) async fn apply_timeline_archival_config(
2249 0 : self: &Arc<Self>,
2250 0 : timeline_id: TimelineId,
2251 0 : new_state: TimelineArchivalState,
2252 0 : broker_client: storage_broker::BrokerClientChannel,
2253 0 : ctx: RequestContext,
2254 0 : ) -> Result<(), TimelineArchivalError> {
2255 0 : info!("setting timeline archival config");
2256 : // First part: figure out what is needed to do, and do validation
2257 0 : let timeline_or_unarchive_offloaded = 'outer: {
2258 0 : let timelines = self.timelines.lock().unwrap();
2259 :
2260 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2261 0 : let offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2262 0 : let Some(offloaded) = offloaded_timelines.get(&timeline_id) else {
2263 0 : return Err(TimelineArchivalError::NotFound);
2264 : };
2265 0 : if new_state == TimelineArchivalState::Archived {
2266 : // It's offloaded already, so nothing to do
2267 0 : return Ok(());
2268 0 : }
2269 0 : if let Some(ancestor_timeline_id) = offloaded.ancestor_timeline_id {
2270 0 : Self::check_ancestor_of_to_be_unarchived_is_not_archived(
2271 0 : ancestor_timeline_id,
2272 0 : &timelines,
2273 0 : &offloaded_timelines,
2274 0 : )?;
2275 0 : }
2276 0 : break 'outer None;
2277 : };
2278 :
2279 : // Do some validation. We release the timelines lock below, so there is potential
2280 : // for race conditions: these checks are more present to prevent misunderstandings of
2281 : // the API's capabilities, instead of serving as the sole way to defend their invariants.
2282 0 : match new_state {
2283 : TimelineArchivalState::Unarchived => {
2284 0 : Self::check_to_be_unarchived_timeline_has_no_archived_parent(timeline)?
2285 : }
2286 : TimelineArchivalState::Archived => {
2287 0 : Self::check_to_be_archived_has_no_unarchived_children(timeline_id, &timelines)?
2288 : }
2289 : }
2290 0 : Some(Arc::clone(timeline))
2291 : };
2292 :
2293 : // Second part: unoffload timeline (if needed)
2294 0 : let timeline = if let Some(timeline) = timeline_or_unarchive_offloaded {
2295 0 : timeline
2296 : } else {
2297 : // Turn offloaded timeline into a non-offloaded one
2298 0 : self.unoffload_timeline(timeline_id, broker_client, ctx)
2299 0 : .await?
2300 : };
2301 :
2302 : // Third part: upload new timeline archival state and block until it is present in S3
2303 0 : let upload_needed = match timeline
2304 0 : .remote_client
2305 0 : .schedule_index_upload_for_timeline_archival_state(new_state)
2306 : {
2307 0 : Ok(upload_needed) => upload_needed,
2308 0 : Err(e) => {
2309 0 : if timeline.cancel.is_cancelled() {
2310 0 : return Err(TimelineArchivalError::Cancelled);
2311 : } else {
2312 0 : return Err(TimelineArchivalError::Other(e));
2313 : }
2314 : }
2315 : };
2316 :
2317 0 : if upload_needed {
2318 0 : info!("Uploading new state");
2319 : const MAX_WAIT: Duration = Duration::from_secs(10);
2320 0 : let Ok(v) =
2321 0 : tokio::time::timeout(MAX_WAIT, timeline.remote_client.wait_completion()).await
2322 : else {
2323 0 : tracing::warn!("reached timeout for waiting on upload queue");
2324 0 : return Err(TimelineArchivalError::Timeout);
2325 : };
2326 0 : v.map_err(|e| match e {
2327 0 : WaitCompletionError::NotInitialized(e) => {
2328 0 : TimelineArchivalError::Other(anyhow::anyhow!(e))
2329 : }
2330 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2331 0 : TimelineArchivalError::Cancelled
2332 : }
2333 0 : })?;
2334 0 : }
2335 0 : Ok(())
2336 0 : }
2337 :
2338 1 : pub fn get_offloaded_timeline(
2339 1 : &self,
2340 1 : timeline_id: TimelineId,
2341 1 : ) -> Result<Arc<OffloadedTimeline>, GetTimelineError> {
2342 1 : self.timelines_offloaded
2343 1 : .lock()
2344 1 : .unwrap()
2345 1 : .get(&timeline_id)
2346 1 : .map(Arc::clone)
2347 1 : .ok_or(GetTimelineError::NotFound {
2348 1 : tenant_id: self.tenant_shard_id,
2349 1 : timeline_id,
2350 1 : })
2351 1 : }
2352 :
2353 2 : pub(crate) fn tenant_shard_id(&self) -> TenantShardId {
2354 2 : self.tenant_shard_id
2355 2 : }
2356 :
2357 : /// Get Timeline handle for given Neon timeline ID.
2358 : /// This function is idempotent. It doesn't change internal state in any way.
2359 111 : pub fn get_timeline(
2360 111 : &self,
2361 111 : timeline_id: TimelineId,
2362 111 : active_only: bool,
2363 111 : ) -> Result<Arc<Timeline>, GetTimelineError> {
2364 111 : let timelines_accessor = self.timelines.lock().unwrap();
2365 111 : let timeline = timelines_accessor
2366 111 : .get(&timeline_id)
2367 111 : .ok_or(GetTimelineError::NotFound {
2368 111 : tenant_id: self.tenant_shard_id,
2369 111 : timeline_id,
2370 111 : })?;
2371 :
2372 110 : if active_only && !timeline.is_active() {
2373 0 : Err(GetTimelineError::NotActive {
2374 0 : tenant_id: self.tenant_shard_id,
2375 0 : timeline_id,
2376 0 : state: timeline.current_state(),
2377 0 : })
2378 : } else {
2379 110 : Ok(Arc::clone(timeline))
2380 : }
2381 111 : }
2382 :
2383 : /// Lists timelines the tenant contains.
2384 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2385 2 : pub fn list_timelines(&self) -> Vec<Arc<Timeline>> {
2386 2 : self.timelines
2387 2 : .lock()
2388 2 : .unwrap()
2389 2 : .values()
2390 2 : .map(Arc::clone)
2391 2 : .collect()
2392 2 : }
2393 :
2394 : /// Lists timelines the tenant manages, including offloaded ones.
2395 : ///
2396 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2397 0 : pub fn list_timelines_and_offloaded(
2398 0 : &self,
2399 0 : ) -> (Vec<Arc<Timeline>>, Vec<Arc<OffloadedTimeline>>) {
2400 0 : let timelines = self
2401 0 : .timelines
2402 0 : .lock()
2403 0 : .unwrap()
2404 0 : .values()
2405 0 : .map(Arc::clone)
2406 0 : .collect();
2407 0 : let offloaded = self
2408 0 : .timelines_offloaded
2409 0 : .lock()
2410 0 : .unwrap()
2411 0 : .values()
2412 0 : .map(Arc::clone)
2413 0 : .collect();
2414 0 : (timelines, offloaded)
2415 0 : }
2416 :
2417 0 : pub fn list_timeline_ids(&self) -> Vec<TimelineId> {
2418 0 : self.timelines.lock().unwrap().keys().cloned().collect()
2419 0 : }
2420 :
2421 : /// This is used by tests & import-from-basebackup.
2422 : ///
2423 : /// The returned [`UninitializedTimeline`] contains no data nor metadata and it is in
2424 : /// a state that will fail [`TenantShard::load_remote_timeline`] because `disk_consistent_lsn=Lsn(0)`.
2425 : ///
2426 : /// The caller is responsible for getting the timeline into a state that will be accepted
2427 : /// by [`TenantShard::load_remote_timeline`] / [`TenantShard::attach`].
2428 : /// Then they may call [`UninitializedTimeline::finish_creation`] to add the timeline
2429 : /// to the [`TenantShard::timelines`].
2430 : ///
2431 : /// Tests should use `TenantShard::create_test_timeline` to set up the minimum required metadata keys.
2432 113 : pub(crate) async fn create_empty_timeline(
2433 113 : self: &Arc<Self>,
2434 113 : new_timeline_id: TimelineId,
2435 113 : initdb_lsn: Lsn,
2436 113 : pg_version: u32,
2437 113 : ctx: &RequestContext,
2438 113 : ) -> anyhow::Result<(UninitializedTimeline, RequestContext)> {
2439 113 : anyhow::ensure!(
2440 113 : self.is_active(),
2441 0 : "Cannot create empty timelines on inactive tenant"
2442 : );
2443 :
2444 : // Protect against concurrent attempts to use this TimelineId
2445 113 : let create_guard = match self
2446 113 : .start_creating_timeline(new_timeline_id, CreateTimelineIdempotency::FailWithConflict)
2447 113 : .await?
2448 : {
2449 112 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2450 : StartCreatingTimelineResult::Idempotent(_) => {
2451 0 : unreachable!("FailWithConflict implies we get an error instead")
2452 : }
2453 : };
2454 :
2455 112 : let new_metadata = TimelineMetadata::new(
2456 112 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2457 112 : // make it valid, before calling finish_creation()
2458 112 : Lsn(0),
2459 112 : None,
2460 112 : None,
2461 112 : Lsn(0),
2462 112 : initdb_lsn,
2463 112 : initdb_lsn,
2464 112 : pg_version,
2465 112 : );
2466 112 : self.prepare_new_timeline(
2467 112 : new_timeline_id,
2468 112 : &new_metadata,
2469 112 : create_guard,
2470 112 : initdb_lsn,
2471 112 : None,
2472 112 : None,
2473 112 : ctx,
2474 112 : )
2475 112 : .await
2476 113 : }
2477 :
2478 : /// Helper for unit tests to create an empty timeline.
2479 : ///
2480 : /// The timeline is has state value `Active` but its background loops are not running.
2481 : // This makes the various functions which anyhow::ensure! for Active state work in tests.
2482 : // Our current tests don't need the background loops.
2483 : #[cfg(test)]
2484 108 : pub async fn create_test_timeline(
2485 108 : self: &Arc<Self>,
2486 108 : new_timeline_id: TimelineId,
2487 108 : initdb_lsn: Lsn,
2488 108 : pg_version: u32,
2489 108 : ctx: &RequestContext,
2490 108 : ) -> anyhow::Result<Arc<Timeline>> {
2491 108 : let (uninit_tl, ctx) = self
2492 108 : .create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2493 108 : .await?;
2494 108 : let tline = uninit_tl.raw_timeline().expect("we just created it");
2495 108 : assert_eq!(tline.get_last_record_lsn(), Lsn(0));
2496 :
2497 : // Setup minimum keys required for the timeline to be usable.
2498 108 : let mut modification = tline.begin_modification(initdb_lsn);
2499 108 : modification
2500 108 : .init_empty_test_timeline()
2501 108 : .context("init_empty_test_timeline")?;
2502 108 : modification
2503 108 : .commit(&ctx)
2504 108 : .await
2505 108 : .context("commit init_empty_test_timeline modification")?;
2506 :
2507 : // Flush to disk so that uninit_tl's check for valid disk_consistent_lsn passes.
2508 108 : tline.maybe_spawn_flush_loop();
2509 108 : tline.freeze_and_flush().await.context("freeze_and_flush")?;
2510 :
2511 : // Make sure the freeze_and_flush reaches remote storage.
2512 108 : tline.remote_client.wait_completion().await.unwrap();
2513 :
2514 108 : let tl = uninit_tl.finish_creation().await?;
2515 : // The non-test code would call tl.activate() here.
2516 108 : tl.set_state(TimelineState::Active);
2517 108 : Ok(tl)
2518 108 : }
2519 :
2520 : /// Helper for unit tests to create a timeline with some pre-loaded states.
2521 : #[cfg(test)]
2522 : #[allow(clippy::too_many_arguments)]
2523 24 : pub async fn create_test_timeline_with_layers(
2524 24 : self: &Arc<Self>,
2525 24 : new_timeline_id: TimelineId,
2526 24 : initdb_lsn: Lsn,
2527 24 : pg_version: u32,
2528 24 : ctx: &RequestContext,
2529 24 : in_memory_layer_desc: Vec<timeline::InMemoryLayerTestDesc>,
2530 24 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
2531 24 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
2532 24 : end_lsn: Lsn,
2533 24 : ) -> anyhow::Result<Arc<Timeline>> {
2534 : use checks::check_valid_layermap;
2535 : use itertools::Itertools;
2536 :
2537 24 : let tline = self
2538 24 : .create_test_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2539 24 : .await?;
2540 24 : tline.force_advance_lsn(end_lsn);
2541 71 : for deltas in delta_layer_desc {
2542 47 : tline
2543 47 : .force_create_delta_layer(deltas, Some(initdb_lsn), ctx)
2544 47 : .await?;
2545 : }
2546 58 : for (lsn, images) in image_layer_desc {
2547 34 : tline
2548 34 : .force_create_image_layer(lsn, images, Some(initdb_lsn), ctx)
2549 34 : .await?;
2550 : }
2551 28 : for in_memory in in_memory_layer_desc {
2552 4 : tline
2553 4 : .force_create_in_memory_layer(in_memory, Some(initdb_lsn), ctx)
2554 4 : .await?;
2555 : }
2556 24 : let layer_names = tline
2557 24 : .layers
2558 24 : .read()
2559 24 : .await
2560 24 : .layer_map()
2561 24 : .unwrap()
2562 24 : .iter_historic_layers()
2563 105 : .map(|layer| layer.layer_name())
2564 24 : .collect_vec();
2565 24 : if let Some(err) = check_valid_layermap(&layer_names) {
2566 0 : bail!("invalid layermap: {err}");
2567 24 : }
2568 24 : Ok(tline)
2569 24 : }
2570 :
2571 : /// Create a new timeline.
2572 : ///
2573 : /// Returns the new timeline ID and reference to its Timeline object.
2574 : ///
2575 : /// If the caller specified the timeline ID to use (`new_timeline_id`), and timeline with
2576 : /// the same timeline ID already exists, returns CreateTimelineError::AlreadyExists.
2577 : #[allow(clippy::too_many_arguments)]
2578 0 : pub(crate) async fn create_timeline(
2579 0 : self: &Arc<TenantShard>,
2580 0 : params: CreateTimelineParams,
2581 0 : broker_client: storage_broker::BrokerClientChannel,
2582 0 : ctx: &RequestContext,
2583 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
2584 0 : if !self.is_active() {
2585 0 : if matches!(self.current_state(), TenantState::Stopping { .. }) {
2586 0 : return Err(CreateTimelineError::ShuttingDown);
2587 : } else {
2588 0 : return Err(CreateTimelineError::Other(anyhow::anyhow!(
2589 0 : "Cannot create timelines on inactive tenant"
2590 0 : )));
2591 : }
2592 0 : }
2593 :
2594 0 : let _gate = self
2595 0 : .gate
2596 0 : .enter()
2597 0 : .map_err(|_| CreateTimelineError::ShuttingDown)?;
2598 :
2599 0 : let result: CreateTimelineResult = match params {
2600 : CreateTimelineParams::Bootstrap(CreateTimelineParamsBootstrap {
2601 0 : new_timeline_id,
2602 0 : existing_initdb_timeline_id,
2603 0 : pg_version,
2604 0 : }) => {
2605 0 : self.bootstrap_timeline(
2606 0 : new_timeline_id,
2607 0 : pg_version,
2608 0 : existing_initdb_timeline_id,
2609 0 : ctx,
2610 0 : )
2611 0 : .await?
2612 : }
2613 : CreateTimelineParams::Branch(CreateTimelineParamsBranch {
2614 0 : new_timeline_id,
2615 0 : ancestor_timeline_id,
2616 0 : mut ancestor_start_lsn,
2617 : }) => {
2618 0 : let ancestor_timeline = self
2619 0 : .get_timeline(ancestor_timeline_id, false)
2620 0 : .context("Cannot branch off the timeline that's not present in pageserver")?;
2621 :
2622 : // instead of waiting around, just deny the request because ancestor is not yet
2623 : // ready for other purposes either.
2624 0 : if !ancestor_timeline.is_active() {
2625 0 : return Err(CreateTimelineError::AncestorNotActive);
2626 0 : }
2627 0 :
2628 0 : if ancestor_timeline.is_archived() == Some(true) {
2629 0 : info!("tried to branch archived timeline");
2630 0 : return Err(CreateTimelineError::AncestorArchived);
2631 0 : }
2632 :
2633 0 : if let Some(lsn) = ancestor_start_lsn.as_mut() {
2634 0 : *lsn = lsn.align();
2635 0 :
2636 0 : let ancestor_ancestor_lsn = ancestor_timeline.get_ancestor_lsn();
2637 0 : if ancestor_ancestor_lsn > *lsn {
2638 : // can we safely just branch from the ancestor instead?
2639 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
2640 0 : "invalid start lsn {} for ancestor timeline {}: less than timeline ancestor lsn {}",
2641 0 : lsn,
2642 0 : ancestor_timeline_id,
2643 0 : ancestor_ancestor_lsn,
2644 0 : )));
2645 0 : }
2646 0 :
2647 0 : // Wait for the WAL to arrive and be processed on the parent branch up
2648 0 : // to the requested branch point. The repository code itself doesn't
2649 0 : // require it, but if we start to receive WAL on the new timeline,
2650 0 : // decoding the new WAL might need to look up previous pages, relation
2651 0 : // sizes etc. and that would get confused if the previous page versions
2652 0 : // are not in the repository yet.
2653 0 : ancestor_timeline
2654 0 : .wait_lsn(
2655 0 : *lsn,
2656 0 : timeline::WaitLsnWaiter::Tenant,
2657 0 : timeline::WaitLsnTimeout::Default,
2658 0 : ctx,
2659 0 : )
2660 0 : .await
2661 0 : .map_err(|e| match e {
2662 0 : e @ (WaitLsnError::Timeout(_) | WaitLsnError::BadState { .. }) => {
2663 0 : CreateTimelineError::AncestorLsn(anyhow::anyhow!(e))
2664 : }
2665 0 : WaitLsnError::Shutdown => CreateTimelineError::ShuttingDown,
2666 0 : })?;
2667 0 : }
2668 :
2669 0 : self.branch_timeline(&ancestor_timeline, new_timeline_id, ancestor_start_lsn, ctx)
2670 0 : .await?
2671 : }
2672 0 : CreateTimelineParams::ImportPgdata(params) => {
2673 0 : self.create_timeline_import_pgdata(params, ctx).await?
2674 : }
2675 : };
2676 :
2677 : // At this point we have dropped our guard on [`Self::timelines_creating`], and
2678 : // the timeline is visible in [`Self::timelines`], but it is _not_ durable yet. We must
2679 : // not send a success to the caller until it is. The same applies to idempotent retries.
2680 : //
2681 : // TODO: the timeline is already visible in [`Self::timelines`]; a caller could incorrectly
2682 : // assume that, because they can see the timeline via API, that the creation is done and
2683 : // that it is durable. Ideally, we would keep the timeline hidden (in [`Self::timelines_creating`])
2684 : // until it is durable, e.g., by extending the time we hold the creation guard. This also
2685 : // interacts with UninitializedTimeline and is generally a bit tricky.
2686 : //
2687 : // To re-emphasize: the only correct way to create a timeline is to repeat calling the
2688 : // creation API until it returns success. Only then is durability guaranteed.
2689 0 : info!(creation_result=%result.discriminant(), "waiting for timeline to be durable");
2690 0 : result
2691 0 : .timeline()
2692 0 : .remote_client
2693 0 : .wait_completion()
2694 0 : .await
2695 0 : .map_err(|e| match e {
2696 : WaitCompletionError::NotInitialized(
2697 0 : e, // If the queue is already stopped, it's a shutdown error.
2698 0 : ) if e.is_stopping() => CreateTimelineError::ShuttingDown,
2699 : WaitCompletionError::NotInitialized(_) => {
2700 : // This is a bug: we should never try to wait for uploads before initializing the timeline
2701 0 : debug_assert!(false);
2702 0 : CreateTimelineError::Other(anyhow::anyhow!("timeline not initialized"))
2703 : }
2704 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2705 0 : CreateTimelineError::ShuttingDown
2706 : }
2707 0 : })?;
2708 :
2709 : // The creating task is responsible for activating the timeline.
2710 : // We do this after `wait_completion()` so that we don't spin up tasks that start
2711 : // doing stuff before the IndexPart is durable in S3, which is done by the previous section.
2712 0 : let activated_timeline = match result {
2713 0 : CreateTimelineResult::Created(timeline) => {
2714 0 : timeline.activate(
2715 0 : self.clone(),
2716 0 : broker_client,
2717 0 : None,
2718 0 : &ctx.with_scope_timeline(&timeline),
2719 0 : );
2720 0 : timeline
2721 : }
2722 0 : CreateTimelineResult::Idempotent(timeline) => {
2723 0 : info!(
2724 0 : "request was deemed idempotent, activation will be done by the creating task"
2725 : );
2726 0 : timeline
2727 : }
2728 0 : CreateTimelineResult::ImportSpawned(timeline) => {
2729 0 : info!(
2730 0 : "import task spawned, timeline will become visible and activated once the import is done"
2731 : );
2732 0 : timeline
2733 : }
2734 : };
2735 :
2736 0 : Ok(activated_timeline)
2737 0 : }
2738 :
2739 : /// The returned [`Arc<Timeline>`] is NOT in the [`TenantShard::timelines`] map until the import
2740 : /// completes in the background. A DIFFERENT [`Arc<Timeline>`] will be inserted into the
2741 : /// [`TenantShard::timelines`] map when the import completes.
2742 : /// We only return an [`Arc<Timeline>`] here so the API handler can create a [`pageserver_api::models::TimelineInfo`]
2743 : /// for the response.
2744 0 : async fn create_timeline_import_pgdata(
2745 0 : self: &Arc<Self>,
2746 0 : params: CreateTimelineParamsImportPgdata,
2747 0 : ctx: &RequestContext,
2748 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
2749 0 : let CreateTimelineParamsImportPgdata {
2750 0 : new_timeline_id,
2751 0 : location,
2752 0 : idempotency_key,
2753 0 : } = params;
2754 0 :
2755 0 : let started_at = chrono::Utc::now().naive_utc();
2756 :
2757 : //
2758 : // There's probably a simpler way to upload an index part, but, remote_timeline_client
2759 : // is the canonical way we do it.
2760 : // - create an empty timeline in-memory
2761 : // - use its remote_timeline_client to do the upload
2762 : // - dispose of the uninit timeline
2763 : // - keep the creation guard alive
2764 :
2765 0 : let timeline_create_guard = match self
2766 0 : .start_creating_timeline(
2767 0 : new_timeline_id,
2768 0 : CreateTimelineIdempotency::ImportPgdata(CreatingTimelineIdempotencyImportPgdata {
2769 0 : idempotency_key: idempotency_key.clone(),
2770 0 : }),
2771 0 : )
2772 0 : .await?
2773 : {
2774 0 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2775 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
2776 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
2777 : }
2778 : };
2779 :
2780 0 : let (mut uninit_timeline, timeline_ctx) = {
2781 0 : let this = &self;
2782 0 : let initdb_lsn = Lsn(0);
2783 0 : async move {
2784 0 : let new_metadata = TimelineMetadata::new(
2785 0 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2786 0 : // make it valid, before calling finish_creation()
2787 0 : Lsn(0),
2788 0 : None,
2789 0 : None,
2790 0 : Lsn(0),
2791 0 : initdb_lsn,
2792 0 : initdb_lsn,
2793 0 : 15,
2794 0 : );
2795 0 : this.prepare_new_timeline(
2796 0 : new_timeline_id,
2797 0 : &new_metadata,
2798 0 : timeline_create_guard,
2799 0 : initdb_lsn,
2800 0 : None,
2801 0 : None,
2802 0 : ctx,
2803 0 : )
2804 0 : .await
2805 0 : }
2806 0 : }
2807 0 : .await?;
2808 :
2809 0 : let in_progress = import_pgdata::index_part_format::InProgress {
2810 0 : idempotency_key,
2811 0 : location,
2812 0 : started_at,
2813 0 : };
2814 0 : let index_part = import_pgdata::index_part_format::Root::V1(
2815 0 : import_pgdata::index_part_format::V1::InProgress(in_progress),
2816 0 : );
2817 0 : uninit_timeline
2818 0 : .raw_timeline()
2819 0 : .unwrap()
2820 0 : .remote_client
2821 0 : .schedule_index_upload_for_import_pgdata_state_update(Some(index_part.clone()))?;
2822 :
2823 : // wait_completion happens in caller
2824 :
2825 0 : let (timeline, timeline_create_guard) = uninit_timeline.finish_creation_myself();
2826 0 :
2827 0 : let import_task_handle = tokio::spawn(self.clone().create_timeline_import_pgdata_task(
2828 0 : timeline.clone(),
2829 0 : index_part,
2830 0 : timeline_create_guard,
2831 0 : timeline_ctx.detached_child(TaskKind::ImportPgdata, DownloadBehavior::Warn),
2832 0 : ));
2833 0 :
2834 0 : let prev = self.timelines_importing.lock().unwrap().insert(
2835 0 : timeline.timeline_id,
2836 0 : ImportingTimeline {
2837 0 : timeline: timeline.clone(),
2838 0 : import_task_handle,
2839 0 : },
2840 0 : );
2841 0 :
2842 0 : // Idempotency is enforced higher up the stack
2843 0 : assert!(prev.is_none());
2844 :
2845 : // NB: the timeline doesn't exist in self.timelines at this point
2846 0 : Ok(CreateTimelineResult::ImportSpawned(timeline))
2847 0 : }
2848 :
2849 : /// Finalize the import of a timeline on this shard by marking it complete in
2850 : /// the index part. If the import task hasn't finished yet, returns an error.
2851 : ///
2852 : /// This method is idempotent. If the import was finalized once, the next call
2853 : /// will be a no-op.
2854 0 : pub(crate) async fn finalize_importing_timeline(
2855 0 : &self,
2856 0 : timeline_id: TimelineId,
2857 0 : ) -> anyhow::Result<()> {
2858 0 : let timeline = {
2859 0 : let locked = self.timelines_importing.lock().unwrap();
2860 0 : match locked.get(&timeline_id) {
2861 0 : Some(importing_timeline) => {
2862 0 : if !importing_timeline.import_task_handle.is_finished() {
2863 0 : return Err(anyhow::anyhow!("Import task not done yet"));
2864 0 : }
2865 0 :
2866 0 : importing_timeline.timeline.clone()
2867 : }
2868 : None => {
2869 0 : return Ok(());
2870 : }
2871 : }
2872 : };
2873 :
2874 0 : timeline
2875 0 : .remote_client
2876 0 : .schedule_index_upload_for_import_pgdata_finalize()?;
2877 0 : timeline.remote_client.wait_completion().await?;
2878 :
2879 0 : self.timelines_importing
2880 0 : .lock()
2881 0 : .unwrap()
2882 0 : .remove(&timeline_id);
2883 0 :
2884 0 : Ok(())
2885 0 : }
2886 :
2887 : #[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))]
2888 : async fn create_timeline_import_pgdata_task(
2889 : self: Arc<TenantShard>,
2890 : timeline: Arc<Timeline>,
2891 : index_part: import_pgdata::index_part_format::Root,
2892 : timeline_create_guard: TimelineCreateGuard,
2893 : ctx: RequestContext,
2894 : ) {
2895 : debug_assert_current_span_has_tenant_and_timeline_id();
2896 : info!("starting");
2897 : scopeguard::defer! {info!("exiting")};
2898 :
2899 : let res = self
2900 : .create_timeline_import_pgdata_task_impl(
2901 : timeline,
2902 : index_part,
2903 : timeline_create_guard,
2904 : ctx,
2905 : )
2906 : .await;
2907 : if let Err(err) = &res {
2908 : error!(?err, "task failed");
2909 : // TODO sleep & retry, sensitive to tenant shutdown
2910 : // TODO: allow timeline deletion requests => should cancel the task
2911 : }
2912 : }
2913 :
2914 0 : async fn create_timeline_import_pgdata_task_impl(
2915 0 : self: Arc<TenantShard>,
2916 0 : timeline: Arc<Timeline>,
2917 0 : index_part: import_pgdata::index_part_format::Root,
2918 0 : _timeline_create_guard: TimelineCreateGuard,
2919 0 : ctx: RequestContext,
2920 0 : ) -> Result<(), anyhow::Error> {
2921 0 : info!("importing pgdata");
2922 0 : let ctx = ctx.with_scope_timeline(&timeline);
2923 0 : import_pgdata::doit(&timeline, index_part, &ctx, self.cancel.clone())
2924 0 : .await
2925 0 : .context("import")?;
2926 0 : info!("import done - waiting for activation");
2927 :
2928 0 : anyhow::Ok(())
2929 0 : }
2930 :
2931 0 : pub(crate) async fn delete_timeline(
2932 0 : self: Arc<Self>,
2933 0 : timeline_id: TimelineId,
2934 0 : ) -> Result<(), DeleteTimelineError> {
2935 0 : DeleteTimelineFlow::run(&self, timeline_id).await?;
2936 :
2937 0 : Ok(())
2938 0 : }
2939 :
2940 : /// perform one garbage collection iteration, removing old data files from disk.
2941 : /// this function is periodically called by gc task.
2942 : /// also it can be explicitly requested through page server api 'do_gc' command.
2943 : ///
2944 : /// `target_timeline_id` specifies the timeline to GC, or None for all.
2945 : ///
2946 : /// The `horizon` an `pitr` parameters determine how much WAL history needs to be retained.
2947 : /// Also known as the retention period, or the GC cutoff point. `horizon` specifies
2948 : /// the amount of history, as LSN difference from current latest LSN on each timeline.
2949 : /// `pitr` specifies the same as a time difference from the current time. The effective
2950 : /// GC cutoff point is determined conservatively by either `horizon` and `pitr`, whichever
2951 : /// requires more history to be retained.
2952 : //
2953 377 : pub(crate) async fn gc_iteration(
2954 377 : &self,
2955 377 : target_timeline_id: Option<TimelineId>,
2956 377 : horizon: u64,
2957 377 : pitr: Duration,
2958 377 : cancel: &CancellationToken,
2959 377 : ctx: &RequestContext,
2960 377 : ) -> Result<GcResult, GcError> {
2961 377 : // Don't start doing work during shutdown
2962 377 : if let TenantState::Stopping { .. } = self.current_state() {
2963 0 : return Ok(GcResult::default());
2964 377 : }
2965 377 :
2966 377 : // there is a global allowed_error for this
2967 377 : if !self.is_active() {
2968 0 : return Err(GcError::NotActive);
2969 377 : }
2970 377 :
2971 377 : {
2972 377 : let conf = self.tenant_conf.load();
2973 377 :
2974 377 : // If we may not delete layers, then simply skip GC. Even though a tenant
2975 377 : // in AttachedMulti state could do GC and just enqueue the blocked deletions,
2976 377 : // the only advantage to doing it is to perhaps shrink the LayerMap metadata
2977 377 : // a bit sooner than we would achieve by waiting for AttachedSingle status.
2978 377 : if !conf.location.may_delete_layers_hint() {
2979 0 : info!("Skipping GC in location state {:?}", conf.location);
2980 0 : return Ok(GcResult::default());
2981 377 : }
2982 377 :
2983 377 : if conf.is_gc_blocked_by_lsn_lease_deadline() {
2984 375 : info!("Skipping GC because lsn lease deadline is not reached");
2985 375 : return Ok(GcResult::default());
2986 2 : }
2987 : }
2988 :
2989 2 : let _guard = match self.gc_block.start().await {
2990 2 : Ok(guard) => guard,
2991 0 : Err(reasons) => {
2992 0 : info!("Skipping GC: {reasons}");
2993 0 : return Ok(GcResult::default());
2994 : }
2995 : };
2996 :
2997 2 : self.gc_iteration_internal(target_timeline_id, horizon, pitr, cancel, ctx)
2998 2 : .await
2999 377 : }
3000 :
3001 : /// Performs one compaction iteration. Called periodically from the compaction loop. Returns
3002 : /// whether another compaction is needed, if we still have pending work or if we yield for
3003 : /// immediate L0 compaction.
3004 : ///
3005 : /// Compaction can also be explicitly requested for a timeline via the HTTP API.
3006 0 : async fn compaction_iteration(
3007 0 : self: &Arc<Self>,
3008 0 : cancel: &CancellationToken,
3009 0 : ctx: &RequestContext,
3010 0 : ) -> Result<CompactionOutcome, CompactionError> {
3011 0 : // Don't compact inactive tenants.
3012 0 : if !self.is_active() {
3013 0 : return Ok(CompactionOutcome::Skipped);
3014 0 : }
3015 0 :
3016 0 : // Don't compact tenants that can't upload layers. We don't check `may_delete_layers_hint`,
3017 0 : // since we need to compact L0 even in AttachedMulti to bound read amplification.
3018 0 : let location = self.tenant_conf.load().location;
3019 0 : if !location.may_upload_layers_hint() {
3020 0 : info!("skipping compaction in location state {location:?}");
3021 0 : return Ok(CompactionOutcome::Skipped);
3022 0 : }
3023 0 :
3024 0 : // Don't compact if the circuit breaker is tripped.
3025 0 : if self.compaction_circuit_breaker.lock().unwrap().is_broken() {
3026 0 : info!("skipping compaction due to previous failures");
3027 0 : return Ok(CompactionOutcome::Skipped);
3028 0 : }
3029 0 :
3030 0 : // Collect all timelines to compact, along with offload instructions and L0 counts.
3031 0 : let mut compact: Vec<Arc<Timeline>> = Vec::new();
3032 0 : let mut offload: HashSet<TimelineId> = HashSet::new();
3033 0 : let mut l0_counts: HashMap<TimelineId, usize> = HashMap::new();
3034 0 :
3035 0 : {
3036 0 : let offload_enabled = self.get_timeline_offloading_enabled();
3037 0 : let timelines = self.timelines.lock().unwrap();
3038 0 : for (&timeline_id, timeline) in timelines.iter() {
3039 : // Skip inactive timelines.
3040 0 : if !timeline.is_active() {
3041 0 : continue;
3042 0 : }
3043 0 :
3044 0 : // Schedule the timeline for compaction.
3045 0 : compact.push(timeline.clone());
3046 :
3047 : // Schedule the timeline for offloading if eligible.
3048 0 : let can_offload = offload_enabled
3049 0 : && timeline.can_offload().0
3050 0 : && !timelines
3051 0 : .iter()
3052 0 : .any(|(_, tli)| tli.get_ancestor_timeline_id() == Some(timeline_id));
3053 0 : if can_offload {
3054 0 : offload.insert(timeline_id);
3055 0 : }
3056 : }
3057 : } // release timelines lock
3058 :
3059 0 : for timeline in &compact {
3060 : // Collect L0 counts. Can't await while holding lock above.
3061 0 : if let Ok(lm) = timeline.layers.read().await.layer_map() {
3062 0 : l0_counts.insert(timeline.timeline_id, lm.level0_deltas().len());
3063 0 : }
3064 : }
3065 :
3066 : // Pass 1: L0 compaction across all timelines, in order of L0 count. We prioritize this to
3067 : // bound read amplification.
3068 : //
3069 : // TODO: this may spin on one or more ingest-heavy timelines, starving out image/GC
3070 : // compaction and offloading. We leave that as a potential problem to solve later. Consider
3071 : // splitting L0 and image/GC compaction to separate background jobs.
3072 0 : if self.get_compaction_l0_first() {
3073 0 : let compaction_threshold = self.get_compaction_threshold();
3074 0 : let compact_l0 = compact
3075 0 : .iter()
3076 0 : .map(|tli| (tli, l0_counts.get(&tli.timeline_id).copied().unwrap_or(0)))
3077 0 : .filter(|&(_, l0)| l0 >= compaction_threshold)
3078 0 : .sorted_by_key(|&(_, l0)| l0)
3079 0 : .rev()
3080 0 : .map(|(tli, _)| tli.clone())
3081 0 : .collect_vec();
3082 0 :
3083 0 : let mut has_pending_l0 = false;
3084 0 : for timeline in compact_l0 {
3085 0 : let ctx = &ctx.with_scope_timeline(&timeline);
3086 : // NB: don't set CompactFlags::YieldForL0, since this is an L0-only compaction pass.
3087 0 : let outcome = timeline
3088 0 : .compact(cancel, CompactFlags::OnlyL0Compaction.into(), ctx)
3089 0 : .instrument(info_span!("compact_timeline", timeline_id = %timeline.timeline_id))
3090 0 : .await
3091 0 : .inspect_err(|err| self.maybe_trip_compaction_breaker(err))?;
3092 0 : match outcome {
3093 0 : CompactionOutcome::Done => {}
3094 0 : CompactionOutcome::Skipped => {}
3095 0 : CompactionOutcome::Pending => has_pending_l0 = true,
3096 0 : CompactionOutcome::YieldForL0 => has_pending_l0 = true,
3097 : }
3098 : }
3099 0 : if has_pending_l0 {
3100 0 : return Ok(CompactionOutcome::YieldForL0); // do another pass
3101 0 : }
3102 0 : }
3103 :
3104 : // Pass 2: image compaction and timeline offloading. If any timelines have accumulated more
3105 : // L0 layers, they may also be compacted here. Image compaction will yield if there is
3106 : // pending L0 compaction on any tenant timeline.
3107 : //
3108 : // TODO: consider ordering timelines by some priority, e.g. time since last full compaction,
3109 : // amount of L1 delta debt or garbage, offload-eligible timelines first, etc.
3110 0 : let mut has_pending = false;
3111 0 : for timeline in compact {
3112 0 : if !timeline.is_active() {
3113 0 : continue;
3114 0 : }
3115 0 : let ctx = &ctx.with_scope_timeline(&timeline);
3116 0 :
3117 0 : // Yield for L0 if the separate L0 pass is enabled (otherwise there's no point).
3118 0 : let mut flags = EnumSet::default();
3119 0 : if self.get_compaction_l0_first() {
3120 0 : flags |= CompactFlags::YieldForL0;
3121 0 : }
3122 :
3123 0 : let mut outcome = timeline
3124 0 : .compact(cancel, flags, ctx)
3125 0 : .instrument(info_span!("compact_timeline", timeline_id = %timeline.timeline_id))
3126 0 : .await
3127 0 : .inspect_err(|err| self.maybe_trip_compaction_breaker(err))?;
3128 :
3129 : // If we're done compacting, check the scheduled GC compaction queue for more work.
3130 0 : if outcome == CompactionOutcome::Done {
3131 0 : let queue = {
3132 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3133 0 : guard
3134 0 : .entry(timeline.timeline_id)
3135 0 : .or_insert_with(|| Arc::new(GcCompactionQueue::new()))
3136 0 : .clone()
3137 0 : };
3138 0 : outcome = queue
3139 0 : .iteration(cancel, ctx, &self.gc_block, &timeline)
3140 0 : .instrument(
3141 0 : info_span!("gc_compact_timeline", timeline_id = %timeline.timeline_id),
3142 : )
3143 0 : .await?;
3144 0 : }
3145 :
3146 : // If we're done compacting, offload the timeline if requested.
3147 0 : if outcome == CompactionOutcome::Done && offload.contains(&timeline.timeline_id) {
3148 0 : pausable_failpoint!("before-timeline-auto-offload");
3149 0 : offload_timeline(self, &timeline)
3150 0 : .instrument(info_span!("offload_timeline", timeline_id = %timeline.timeline_id))
3151 0 : .await
3152 0 : .or_else(|err| match err {
3153 : // Ignore this, we likely raced with unarchival.
3154 0 : OffloadError::NotArchived => Ok(()),
3155 0 : err => Err(err),
3156 0 : })?;
3157 0 : }
3158 :
3159 0 : match outcome {
3160 0 : CompactionOutcome::Done => {}
3161 0 : CompactionOutcome::Skipped => {}
3162 0 : CompactionOutcome::Pending => has_pending = true,
3163 : // This mostly makes sense when the L0-only pass above is enabled, since there's
3164 : // otherwise no guarantee that we'll start with the timeline that has high L0.
3165 0 : CompactionOutcome::YieldForL0 => return Ok(CompactionOutcome::YieldForL0),
3166 : }
3167 : }
3168 :
3169 : // Success! Untrip the breaker if necessary.
3170 0 : self.compaction_circuit_breaker
3171 0 : .lock()
3172 0 : .unwrap()
3173 0 : .success(&CIRCUIT_BREAKERS_UNBROKEN);
3174 0 :
3175 0 : match has_pending {
3176 0 : true => Ok(CompactionOutcome::Pending),
3177 0 : false => Ok(CompactionOutcome::Done),
3178 : }
3179 0 : }
3180 :
3181 : /// Trips the compaction circuit breaker if appropriate.
3182 0 : pub(crate) fn maybe_trip_compaction_breaker(&self, err: &CompactionError) {
3183 0 : match err {
3184 0 : err if err.is_cancel() => {}
3185 0 : CompactionError::ShuttingDown => (),
3186 : // Offload failures don't trip the circuit breaker, since they're cheap to retry and
3187 : // shouldn't block compaction.
3188 0 : CompactionError::Offload(_) => {}
3189 0 : CompactionError::CollectKeySpaceError(err) => {
3190 0 : // CollectKeySpaceError::Cancelled and PageRead::Cancelled are handled in `err.is_cancel` branch.
3191 0 : self.compaction_circuit_breaker
3192 0 : .lock()
3193 0 : .unwrap()
3194 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, err);
3195 0 : }
3196 0 : CompactionError::Other(err) => {
3197 0 : self.compaction_circuit_breaker
3198 0 : .lock()
3199 0 : .unwrap()
3200 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, err);
3201 0 : }
3202 0 : CompactionError::AlreadyRunning(_) => {}
3203 : }
3204 0 : }
3205 :
3206 : /// Cancel scheduled compaction tasks
3207 0 : pub(crate) fn cancel_scheduled_compaction(&self, timeline_id: TimelineId) {
3208 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3209 0 : if let Some(q) = guard.get_mut(&timeline_id) {
3210 0 : q.cancel_scheduled();
3211 0 : }
3212 0 : }
3213 :
3214 0 : pub(crate) fn get_scheduled_compaction_tasks(
3215 0 : &self,
3216 0 : timeline_id: TimelineId,
3217 0 : ) -> Vec<CompactInfoResponse> {
3218 0 : let res = {
3219 0 : let guard = self.scheduled_compaction_tasks.lock().unwrap();
3220 0 : guard.get(&timeline_id).map(|q| q.remaining_jobs())
3221 : };
3222 0 : let Some((running, remaining)) = res else {
3223 0 : return Vec::new();
3224 : };
3225 0 : let mut result = Vec::new();
3226 0 : if let Some((id, running)) = running {
3227 0 : result.extend(running.into_compact_info_resp(id, true));
3228 0 : }
3229 0 : for (id, job) in remaining {
3230 0 : result.extend(job.into_compact_info_resp(id, false));
3231 0 : }
3232 0 : result
3233 0 : }
3234 :
3235 : /// Schedule a compaction task for a timeline.
3236 0 : pub(crate) async fn schedule_compaction(
3237 0 : &self,
3238 0 : timeline_id: TimelineId,
3239 0 : options: CompactOptions,
3240 0 : ) -> anyhow::Result<tokio::sync::oneshot::Receiver<()>> {
3241 0 : let (tx, rx) = tokio::sync::oneshot::channel();
3242 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3243 0 : let q = guard
3244 0 : .entry(timeline_id)
3245 0 : .or_insert_with(|| Arc::new(GcCompactionQueue::new()));
3246 0 : q.schedule_manual_compaction(options, Some(tx));
3247 0 : Ok(rx)
3248 0 : }
3249 :
3250 : /// Performs periodic housekeeping, via the tenant housekeeping background task.
3251 0 : async fn housekeeping(&self) {
3252 0 : // Call through to all timelines to freeze ephemeral layers as needed. This usually happens
3253 0 : // during ingest, but we don't want idle timelines to hold open layers for too long.
3254 0 : //
3255 0 : // We don't do this if the tenant can't upload layers (i.e. it's in stale attachment mode).
3256 0 : // We don't run compaction in this case either, and don't want to keep flushing tiny L0
3257 0 : // layers that won't be compacted down.
3258 0 : if self.tenant_conf.load().location.may_upload_layers_hint() {
3259 0 : let timelines = self
3260 0 : .timelines
3261 0 : .lock()
3262 0 : .unwrap()
3263 0 : .values()
3264 0 : .filter(|tli| tli.is_active())
3265 0 : .cloned()
3266 0 : .collect_vec();
3267 :
3268 0 : for timeline in timelines {
3269 0 : timeline.maybe_freeze_ephemeral_layer().await;
3270 : }
3271 0 : }
3272 :
3273 : // Shut down walredo if idle.
3274 : const WALREDO_IDLE_TIMEOUT: Duration = Duration::from_secs(180);
3275 0 : if let Some(ref walredo_mgr) = self.walredo_mgr {
3276 0 : walredo_mgr.maybe_quiesce(WALREDO_IDLE_TIMEOUT);
3277 0 : }
3278 0 : }
3279 :
3280 0 : pub fn timeline_has_no_attached_children(&self, timeline_id: TimelineId) -> bool {
3281 0 : let timelines = self.timelines.lock().unwrap();
3282 0 : !timelines
3283 0 : .iter()
3284 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(timeline_id))
3285 0 : }
3286 :
3287 875 : pub fn current_state(&self) -> TenantState {
3288 875 : self.state.borrow().clone()
3289 875 : }
3290 :
3291 494 : pub fn is_active(&self) -> bool {
3292 494 : self.current_state() == TenantState::Active
3293 494 : }
3294 :
3295 0 : pub fn generation(&self) -> Generation {
3296 0 : self.generation
3297 0 : }
3298 :
3299 0 : pub(crate) fn wal_redo_manager_status(&self) -> Option<WalRedoManagerStatus> {
3300 0 : self.walredo_mgr.as_ref().and_then(|mgr| mgr.status())
3301 0 : }
3302 :
3303 : /// Changes tenant status to active, unless shutdown was already requested.
3304 : ///
3305 : /// `background_jobs_can_start` is an optional barrier set to a value during pageserver startup
3306 : /// to delay background jobs. Background jobs can be started right away when None is given.
3307 0 : fn activate(
3308 0 : self: &Arc<Self>,
3309 0 : broker_client: BrokerClientChannel,
3310 0 : background_jobs_can_start: Option<&completion::Barrier>,
3311 0 : ctx: &RequestContext,
3312 0 : ) {
3313 0 : span::debug_assert_current_span_has_tenant_id();
3314 0 :
3315 0 : let mut activating = false;
3316 0 : self.state.send_modify(|current_state| {
3317 : use pageserver_api::models::ActivatingFrom;
3318 0 : match &*current_state {
3319 : TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
3320 0 : panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {state:?}", state = current_state);
3321 : }
3322 0 : TenantState::Attaching => {
3323 0 : *current_state = TenantState::Activating(ActivatingFrom::Attaching);
3324 0 : }
3325 0 : }
3326 0 : debug!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), "Activating tenant");
3327 0 : activating = true;
3328 0 : // Continue outside the closure. We need to grab timelines.lock()
3329 0 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
3330 0 : });
3331 0 :
3332 0 : if activating {
3333 0 : let timelines_accessor = self.timelines.lock().unwrap();
3334 0 : let timelines_offloaded_accessor = self.timelines_offloaded.lock().unwrap();
3335 0 : let timelines_to_activate = timelines_accessor
3336 0 : .values()
3337 0 : .filter(|timeline| !(timeline.is_broken() || timeline.is_stopping()));
3338 0 :
3339 0 : // Before activation, populate each Timeline's GcInfo with information about its children
3340 0 : self.initialize_gc_info(&timelines_accessor, &timelines_offloaded_accessor, None);
3341 0 :
3342 0 : // Spawn gc and compaction loops. The loops will shut themselves
3343 0 : // down when they notice that the tenant is inactive.
3344 0 : tasks::start_background_loops(self, background_jobs_can_start);
3345 0 :
3346 0 : let mut activated_timelines = 0;
3347 :
3348 0 : for timeline in timelines_to_activate {
3349 0 : timeline.activate(
3350 0 : self.clone(),
3351 0 : broker_client.clone(),
3352 0 : background_jobs_can_start,
3353 0 : &ctx.with_scope_timeline(timeline),
3354 0 : );
3355 0 : activated_timelines += 1;
3356 0 : }
3357 :
3358 0 : let tid = self.tenant_shard_id.tenant_id.to_string();
3359 0 : let shard_id = self.tenant_shard_id.shard_slug().to_string();
3360 0 : let offloaded_timeline_count = timelines_offloaded_accessor.len();
3361 0 : TENANT_OFFLOADED_TIMELINES
3362 0 : .with_label_values(&[&tid, &shard_id])
3363 0 : .set(offloaded_timeline_count as u64);
3364 0 :
3365 0 : self.state.send_modify(move |current_state| {
3366 0 : assert!(
3367 0 : matches!(current_state, TenantState::Activating(_)),
3368 0 : "set_stopping and set_broken wait for us to leave Activating state",
3369 : );
3370 0 : *current_state = TenantState::Active;
3371 0 :
3372 0 : let elapsed = self.constructed_at.elapsed();
3373 0 : let total_timelines = timelines_accessor.len();
3374 0 :
3375 0 : // log a lot of stuff, because some tenants sometimes suffer from user-visible
3376 0 : // times to activate. see https://github.com/neondatabase/neon/issues/4025
3377 0 : info!(
3378 0 : since_creation_millis = elapsed.as_millis(),
3379 0 : tenant_id = %self.tenant_shard_id.tenant_id,
3380 0 : shard_id = %self.tenant_shard_id.shard_slug(),
3381 0 : activated_timelines,
3382 0 : total_timelines,
3383 0 : post_state = <&'static str>::from(&*current_state),
3384 0 : "activation attempt finished"
3385 : );
3386 :
3387 0 : TENANT.activation.observe(elapsed.as_secs_f64());
3388 0 : });
3389 0 : }
3390 0 : }
3391 :
3392 : /// Shutdown the tenant and join all of the spawned tasks.
3393 : ///
3394 : /// The method caters for all use-cases:
3395 : /// - pageserver shutdown (freeze_and_flush == true)
3396 : /// - detach + ignore (freeze_and_flush == false)
3397 : ///
3398 : /// This will attempt to shutdown even if tenant is broken.
3399 : ///
3400 : /// `shutdown_progress` is a [`completion::Barrier`] for the shutdown initiated by this call.
3401 : /// If the tenant is already shutting down, we return a clone of the first shutdown call's
3402 : /// `Barrier` as an `Err`. This not-first caller can use the returned barrier to join with
3403 : /// the ongoing shutdown.
3404 3 : async fn shutdown(
3405 3 : &self,
3406 3 : shutdown_progress: completion::Barrier,
3407 3 : shutdown_mode: timeline::ShutdownMode,
3408 3 : ) -> Result<(), completion::Barrier> {
3409 3 : span::debug_assert_current_span_has_tenant_id();
3410 :
3411 : // Set tenant (and its timlines) to Stoppping state.
3412 : //
3413 : // Since we can only transition into Stopping state after activation is complete,
3414 : // run it in a JoinSet so all tenants have a chance to stop before we get SIGKILLed.
3415 : //
3416 : // Transitioning tenants to Stopping state has a couple of non-obvious side effects:
3417 : // 1. Lock out any new requests to the tenants.
3418 : // 2. Signal cancellation to WAL receivers (we wait on it below).
3419 : // 3. Signal cancellation for other tenant background loops.
3420 : // 4. ???
3421 : //
3422 : // The waiting for the cancellation is not done uniformly.
3423 : // We certainly wait for WAL receivers to shut down.
3424 : // That is necessary so that no new data comes in before the freeze_and_flush.
3425 : // But the tenant background loops are joined-on in our caller.
3426 : // It's mesed up.
3427 : // we just ignore the failure to stop
3428 :
3429 : // If we're still attaching, fire the cancellation token early to drop out: this
3430 : // will prevent us flushing, but ensures timely shutdown if some I/O during attach
3431 : // is very slow.
3432 3 : let shutdown_mode = if matches!(self.current_state(), TenantState::Attaching) {
3433 0 : self.cancel.cancel();
3434 0 :
3435 0 : // Having fired our cancellation token, do not try and flush timelines: their cancellation tokens
3436 0 : // are children of ours, so their flush loops will have shut down already
3437 0 : timeline::ShutdownMode::Hard
3438 : } else {
3439 3 : shutdown_mode
3440 : };
3441 :
3442 3 : match self.set_stopping(shutdown_progress).await {
3443 3 : Ok(()) => {}
3444 0 : Err(SetStoppingError::Broken) => {
3445 0 : // assume that this is acceptable
3446 0 : }
3447 0 : Err(SetStoppingError::AlreadyStopping(other)) => {
3448 0 : // give caller the option to wait for this this shutdown
3449 0 : info!("Tenant::shutdown: AlreadyStopping");
3450 0 : return Err(other);
3451 : }
3452 : };
3453 :
3454 3 : let mut js = tokio::task::JoinSet::new();
3455 3 : {
3456 3 : let timelines = self.timelines.lock().unwrap();
3457 3 : timelines.values().for_each(|timeline| {
3458 3 : let timeline = Arc::clone(timeline);
3459 3 : let timeline_id = timeline.timeline_id;
3460 3 : let span = tracing::info_span!("timeline_shutdown", %timeline_id, ?shutdown_mode);
3461 3 : js.spawn(async move { timeline.shutdown(shutdown_mode).instrument(span).await });
3462 3 : });
3463 3 : }
3464 3 : {
3465 3 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
3466 3 : timelines_offloaded.values().for_each(|timeline| {
3467 0 : timeline.defuse_for_tenant_drop();
3468 3 : });
3469 3 : }
3470 3 : {
3471 3 : let mut timelines_importing = self.timelines_importing.lock().unwrap();
3472 3 : timelines_importing
3473 3 : .drain()
3474 3 : .for_each(|(_timeline_id, importing_timeline)| {
3475 0 : importing_timeline.shutdown();
3476 3 : });
3477 3 : }
3478 3 : // test_long_timeline_create_then_tenant_delete is leaning on this message
3479 3 : tracing::info!("Waiting for timelines...");
3480 6 : while let Some(res) = js.join_next().await {
3481 0 : match res {
3482 3 : Ok(()) => {}
3483 0 : Err(je) if je.is_cancelled() => unreachable!("no cancelling used"),
3484 0 : Err(je) if je.is_panic() => { /* logged already */ }
3485 0 : Err(je) => warn!("unexpected JoinError: {je:?}"),
3486 : }
3487 : }
3488 :
3489 3 : if let ShutdownMode::Reload = shutdown_mode {
3490 0 : tracing::info!("Flushing deletion queue");
3491 0 : if let Err(e) = self.deletion_queue_client.flush().await {
3492 0 : match e {
3493 0 : DeletionQueueError::ShuttingDown => {
3494 0 : // This is the only error we expect for now. In the future, if more error
3495 0 : // variants are added, we should handle them here.
3496 0 : }
3497 : }
3498 0 : }
3499 3 : }
3500 :
3501 : // We cancel the Tenant's cancellation token _after_ the timelines have all shut down. This permits
3502 : // them to continue to do work during their shutdown methods, e.g. flushing data.
3503 3 : tracing::debug!("Cancelling CancellationToken");
3504 3 : self.cancel.cancel();
3505 3 :
3506 3 : // shutdown all tenant and timeline tasks: gc, compaction, page service
3507 3 : // No new tasks will be started for this tenant because it's in `Stopping` state.
3508 3 : //
3509 3 : // this will additionally shutdown and await all timeline tasks.
3510 3 : tracing::debug!("Waiting for tasks...");
3511 3 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), None).await;
3512 :
3513 3 : if let Some(walredo_mgr) = self.walredo_mgr.as_ref() {
3514 3 : walredo_mgr.shutdown().await;
3515 0 : }
3516 :
3517 : // Wait for any in-flight operations to complete
3518 3 : self.gate.close().await;
3519 :
3520 3 : remove_tenant_metrics(&self.tenant_shard_id);
3521 3 :
3522 3 : Ok(())
3523 3 : }
3524 :
3525 : /// Change tenant status to Stopping, to mark that it is being shut down.
3526 : ///
3527 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
3528 : ///
3529 : /// This function is not cancel-safe!
3530 3 : async fn set_stopping(&self, progress: completion::Barrier) -> Result<(), SetStoppingError> {
3531 3 : let mut rx = self.state.subscribe();
3532 3 :
3533 3 : // cannot stop before we're done activating, so wait out until we're done activating
3534 3 : rx.wait_for(|state| match state {
3535 : TenantState::Activating(_) | TenantState::Attaching => {
3536 0 : info!("waiting for {state} to turn Active|Broken|Stopping");
3537 0 : false
3538 : }
3539 3 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3540 3 : })
3541 3 : .await
3542 3 : .expect("cannot drop self.state while on a &self method");
3543 3 :
3544 3 : // we now know we're done activating, let's see whether this task is the winner to transition into Stopping
3545 3 : let mut err = None;
3546 3 : let stopping = self.state.send_if_modified(|current_state| match current_state {
3547 : TenantState::Activating(_) | TenantState::Attaching => {
3548 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
3549 : }
3550 : TenantState::Active => {
3551 : // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
3552 : // are created after the transition to Stopping. That's harmless, as the Timelines
3553 : // won't be accessible to anyone afterwards, because the Tenant is in Stopping state.
3554 3 : *current_state = TenantState::Stopping { progress: Some(progress) };
3555 3 : // Continue stopping outside the closure. We need to grab timelines.lock()
3556 3 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
3557 3 : true
3558 : }
3559 : TenantState::Stopping { progress: None } => {
3560 : // An attach was cancelled, and the attach transitioned the tenant from Attaching to
3561 : // Stopping(None) to let us know it exited. Register our progress and continue.
3562 0 : *current_state = TenantState::Stopping { progress: Some(progress) };
3563 0 : true
3564 : }
3565 0 : TenantState::Broken { reason, .. } => {
3566 0 : info!(
3567 0 : "Cannot set tenant to Stopping state, it is in Broken state due to: {reason}"
3568 : );
3569 0 : err = Some(SetStoppingError::Broken);
3570 0 : false
3571 : }
3572 0 : TenantState::Stopping { progress: Some(progress) } => {
3573 0 : info!("Tenant is already in Stopping state");
3574 0 : err = Some(SetStoppingError::AlreadyStopping(progress.clone()));
3575 0 : false
3576 : }
3577 3 : });
3578 3 : match (stopping, err) {
3579 3 : (true, None) => {} // continue
3580 0 : (false, Some(err)) => return Err(err),
3581 0 : (true, Some(_)) => unreachable!(
3582 0 : "send_if_modified closure must error out if not transitioning to Stopping"
3583 0 : ),
3584 0 : (false, None) => unreachable!(
3585 0 : "send_if_modified closure must return true if transitioning to Stopping"
3586 0 : ),
3587 : }
3588 :
3589 3 : let timelines_accessor = self.timelines.lock().unwrap();
3590 3 : let not_broken_timelines = timelines_accessor
3591 3 : .values()
3592 3 : .filter(|timeline| !timeline.is_broken());
3593 6 : for timeline in not_broken_timelines {
3594 3 : timeline.set_state(TimelineState::Stopping);
3595 3 : }
3596 3 : Ok(())
3597 3 : }
3598 :
3599 : /// Method for tenant::mgr to transition us into Broken state in case of a late failure in
3600 : /// `remove_tenant_from_memory`
3601 : ///
3602 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
3603 : ///
3604 : /// In tests, we also use this to set tenants to Broken state on purpose.
3605 0 : pub(crate) async fn set_broken(&self, reason: String) {
3606 0 : let mut rx = self.state.subscribe();
3607 0 :
3608 0 : // The load & attach routines own the tenant state until it has reached `Active`.
3609 0 : // So, wait until it's done.
3610 0 : rx.wait_for(|state| match state {
3611 : TenantState::Activating(_) | TenantState::Attaching => {
3612 0 : info!(
3613 0 : "waiting for {} to turn Active|Broken|Stopping",
3614 0 : <&'static str>::from(state)
3615 : );
3616 0 : false
3617 : }
3618 0 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3619 0 : })
3620 0 : .await
3621 0 : .expect("cannot drop self.state while on a &self method");
3622 0 :
3623 0 : // we now know we're done activating, let's see whether this task is the winner to transition into Broken
3624 0 : self.set_broken_no_wait(reason)
3625 0 : }
3626 :
3627 0 : pub(crate) fn set_broken_no_wait(&self, reason: impl Display) {
3628 0 : let reason = reason.to_string();
3629 0 : self.state.send_modify(|current_state| {
3630 0 : match *current_state {
3631 : TenantState::Activating(_) | TenantState::Attaching => {
3632 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
3633 : }
3634 : TenantState::Active => {
3635 0 : if cfg!(feature = "testing") {
3636 0 : warn!("Changing Active tenant to Broken state, reason: {}", reason);
3637 0 : *current_state = TenantState::broken_from_reason(reason);
3638 : } else {
3639 0 : unreachable!("not allowed to call set_broken on Active tenants in non-testing builds")
3640 : }
3641 : }
3642 : TenantState::Broken { .. } => {
3643 0 : warn!("Tenant is already in Broken state");
3644 : }
3645 : // This is the only "expected" path, any other path is a bug.
3646 : TenantState::Stopping { .. } => {
3647 0 : warn!(
3648 0 : "Marking Stopping tenant as Broken state, reason: {}",
3649 : reason
3650 : );
3651 0 : *current_state = TenantState::broken_from_reason(reason);
3652 : }
3653 : }
3654 0 : });
3655 0 : }
3656 :
3657 0 : pub fn subscribe_for_state_updates(&self) -> watch::Receiver<TenantState> {
3658 0 : self.state.subscribe()
3659 0 : }
3660 :
3661 : /// The activate_now semaphore is initialized with zero units. As soon as
3662 : /// we add a unit, waiters will be able to acquire a unit and proceed.
3663 0 : pub(crate) fn activate_now(&self) {
3664 0 : self.activate_now_sem.add_permits(1);
3665 0 : }
3666 :
3667 0 : pub(crate) async fn wait_to_become_active(
3668 0 : &self,
3669 0 : timeout: Duration,
3670 0 : ) -> Result<(), GetActiveTenantError> {
3671 0 : let mut receiver = self.state.subscribe();
3672 : loop {
3673 0 : let current_state = receiver.borrow_and_update().clone();
3674 0 : match current_state {
3675 : TenantState::Attaching | TenantState::Activating(_) => {
3676 : // in these states, there's a chance that we can reach ::Active
3677 0 : self.activate_now();
3678 0 : match timeout_cancellable(timeout, &self.cancel, receiver.changed()).await {
3679 0 : Ok(r) => {
3680 0 : r.map_err(
3681 0 : |_e: tokio::sync::watch::error::RecvError|
3682 : // Tenant existed but was dropped: report it as non-existent
3683 0 : GetActiveTenantError::NotFound(GetTenantError::ShardNotFound(self.tenant_shard_id))
3684 0 : )?
3685 : }
3686 : Err(TimeoutCancellableError::Cancelled) => {
3687 0 : return Err(GetActiveTenantError::Cancelled);
3688 : }
3689 : Err(TimeoutCancellableError::Timeout) => {
3690 0 : return Err(GetActiveTenantError::WaitForActiveTimeout {
3691 0 : latest_state: Some(self.current_state()),
3692 0 : wait_time: timeout,
3693 0 : });
3694 : }
3695 : }
3696 : }
3697 : TenantState::Active => {
3698 0 : return Ok(());
3699 : }
3700 0 : TenantState::Broken { reason, .. } => {
3701 0 : // This is fatal, and reported distinctly from the general case of "will never be active" because
3702 0 : // it's logically a 500 to external API users (broken is always a bug).
3703 0 : return Err(GetActiveTenantError::Broken(reason));
3704 : }
3705 : TenantState::Stopping { .. } => {
3706 : // There's no chance the tenant can transition back into ::Active
3707 0 : return Err(GetActiveTenantError::WillNotBecomeActive(current_state));
3708 : }
3709 : }
3710 : }
3711 0 : }
3712 :
3713 0 : pub(crate) fn get_attach_mode(&self) -> AttachmentMode {
3714 0 : self.tenant_conf.load().location.attach_mode
3715 0 : }
3716 :
3717 : /// For API access: generate a LocationConfig equivalent to the one that would be used to
3718 : /// create a Tenant in the same state. Do not use this in hot paths: it's for relatively
3719 : /// rare external API calls, like a reconciliation at startup.
3720 0 : pub(crate) fn get_location_conf(&self) -> models::LocationConfig {
3721 0 : let attached_tenant_conf = self.tenant_conf.load();
3722 :
3723 0 : let location_config_mode = match attached_tenant_conf.location.attach_mode {
3724 0 : AttachmentMode::Single => models::LocationConfigMode::AttachedSingle,
3725 0 : AttachmentMode::Multi => models::LocationConfigMode::AttachedMulti,
3726 0 : AttachmentMode::Stale => models::LocationConfigMode::AttachedStale,
3727 : };
3728 :
3729 0 : models::LocationConfig {
3730 0 : mode: location_config_mode,
3731 0 : generation: self.generation.into(),
3732 0 : secondary_conf: None,
3733 0 : shard_number: self.shard_identity.number.0,
3734 0 : shard_count: self.shard_identity.count.literal(),
3735 0 : shard_stripe_size: self.shard_identity.stripe_size.0,
3736 0 : tenant_conf: attached_tenant_conf.tenant_conf.clone(),
3737 0 : }
3738 0 : }
3739 :
3740 0 : pub(crate) fn get_tenant_shard_id(&self) -> &TenantShardId {
3741 0 : &self.tenant_shard_id
3742 0 : }
3743 :
3744 118 : pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
3745 118 : self.shard_identity.stripe_size
3746 118 : }
3747 :
3748 0 : pub(crate) fn get_generation(&self) -> Generation {
3749 0 : self.generation
3750 0 : }
3751 :
3752 : /// This function partially shuts down the tenant (it shuts down the Timelines) and is fallible,
3753 : /// and can leave the tenant in a bad state if it fails. The caller is responsible for
3754 : /// resetting this tenant to a valid state if we fail.
3755 0 : pub(crate) async fn split_prepare(
3756 0 : &self,
3757 0 : child_shards: &Vec<TenantShardId>,
3758 0 : ) -> anyhow::Result<()> {
3759 0 : let (timelines, offloaded) = {
3760 0 : let timelines = self.timelines.lock().unwrap();
3761 0 : let offloaded = self.timelines_offloaded.lock().unwrap();
3762 0 : (timelines.clone(), offloaded.clone())
3763 0 : };
3764 0 : let timelines_iter = timelines
3765 0 : .values()
3766 0 : .map(TimelineOrOffloadedArcRef::<'_>::from)
3767 0 : .chain(
3768 0 : offloaded
3769 0 : .values()
3770 0 : .map(TimelineOrOffloadedArcRef::<'_>::from),
3771 0 : );
3772 0 : for timeline in timelines_iter {
3773 : // We do not block timeline creation/deletion during splits inside the pageserver: it is up to higher levels
3774 : // to ensure that they do not start a split if currently in the process of doing these.
3775 :
3776 0 : let timeline_id = timeline.timeline_id();
3777 :
3778 0 : if let TimelineOrOffloadedArcRef::Timeline(timeline) = timeline {
3779 : // Upload an index from the parent: this is partly to provide freshness for the
3780 : // child tenants that will copy it, and partly for general ease-of-debugging: there will
3781 : // always be a parent shard index in the same generation as we wrote the child shard index.
3782 0 : tracing::info!(%timeline_id, "Uploading index");
3783 0 : timeline
3784 0 : .remote_client
3785 0 : .schedule_index_upload_for_file_changes()?;
3786 0 : timeline.remote_client.wait_completion().await?;
3787 0 : }
3788 :
3789 0 : let remote_client = match timeline {
3790 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.remote_client.clone(),
3791 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => {
3792 0 : let remote_client = self
3793 0 : .build_timeline_client(offloaded.timeline_id, self.remote_storage.clone());
3794 0 : Arc::new(remote_client)
3795 : }
3796 : };
3797 :
3798 : // Shut down the timeline's remote client: this means that the indices we write
3799 : // for child shards will not be invalidated by the parent shard deleting layers.
3800 0 : tracing::info!(%timeline_id, "Shutting down remote storage client");
3801 0 : remote_client.shutdown().await;
3802 :
3803 : // Download methods can still be used after shutdown, as they don't flow through the remote client's
3804 : // queue. In principal the RemoteTimelineClient could provide this without downloading it, but this
3805 : // operation is rare, so it's simpler to just download it (and robustly guarantees that the index
3806 : // we use here really is the remotely persistent one).
3807 0 : tracing::info!(%timeline_id, "Downloading index_part from parent");
3808 0 : let result = remote_client
3809 0 : .download_index_file(&self.cancel)
3810 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))
3811 0 : .await?;
3812 0 : let index_part = match result {
3813 : MaybeDeletedIndexPart::Deleted(_) => {
3814 0 : anyhow::bail!("Timeline deletion happened concurrently with split")
3815 : }
3816 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
3817 : };
3818 :
3819 : // A shard split may not take place while a timeline import is on-going
3820 : // for the tenant. Timeline imports run as part of each tenant shard
3821 : // and rely on the sharding scheme to split the work among pageservers.
3822 : // If we were to split in the middle of this process, we would have to
3823 : // either ensure that it's driven to completion on the old shard set
3824 : // or transfer it to the new shard set. It's technically possible, but complex.
3825 0 : match index_part.import_pgdata {
3826 0 : Some(ref import) if !import.is_done() => {
3827 0 : anyhow::bail!(
3828 0 : "Cannot split due to import with idempotency key: {:?}",
3829 0 : import.idempotency_key()
3830 0 : );
3831 : }
3832 0 : Some(_) | None => {
3833 0 : // fallthrough
3834 0 : }
3835 : }
3836 :
3837 0 : for child_shard in child_shards {
3838 0 : tracing::info!(%timeline_id, "Uploading index_part for child {}", child_shard.to_index());
3839 0 : upload_index_part(
3840 0 : &self.remote_storage,
3841 0 : child_shard,
3842 0 : &timeline_id,
3843 0 : self.generation,
3844 0 : &index_part,
3845 0 : &self.cancel,
3846 0 : )
3847 0 : .await?;
3848 : }
3849 : }
3850 :
3851 0 : let tenant_manifest = self.build_tenant_manifest();
3852 0 : for child_shard in child_shards {
3853 0 : tracing::info!(
3854 0 : "Uploading tenant manifest for child {}",
3855 0 : child_shard.to_index()
3856 : );
3857 0 : upload_tenant_manifest(
3858 0 : &self.remote_storage,
3859 0 : child_shard,
3860 0 : self.generation,
3861 0 : &tenant_manifest,
3862 0 : &self.cancel,
3863 0 : )
3864 0 : .await?;
3865 : }
3866 :
3867 0 : Ok(())
3868 0 : }
3869 :
3870 0 : pub(crate) fn get_sizes(&self) -> TopTenantShardItem {
3871 0 : let mut result = TopTenantShardItem {
3872 0 : id: self.tenant_shard_id,
3873 0 : resident_size: 0,
3874 0 : physical_size: 0,
3875 0 : max_logical_size: 0,
3876 0 : max_logical_size_per_shard: 0,
3877 0 : };
3878 :
3879 0 : for timeline in self.timelines.lock().unwrap().values() {
3880 0 : result.resident_size += timeline.metrics.resident_physical_size_gauge.get();
3881 0 :
3882 0 : result.physical_size += timeline
3883 0 : .remote_client
3884 0 : .metrics
3885 0 : .remote_physical_size_gauge
3886 0 : .get();
3887 0 : result.max_logical_size = std::cmp::max(
3888 0 : result.max_logical_size,
3889 0 : timeline.metrics.current_logical_size_gauge.get(),
3890 0 : );
3891 0 : }
3892 :
3893 0 : result.max_logical_size_per_shard = result
3894 0 : .max_logical_size
3895 0 : .div_ceil(self.tenant_shard_id.shard_count.count() as u64);
3896 0 :
3897 0 : result
3898 0 : }
3899 : }
3900 :
3901 : /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
3902 : /// perform a topological sort, so that the parent of each timeline comes
3903 : /// before the children.
3904 : /// E extracts the ancestor from T
3905 : /// This allows for T to be different. It can be TimelineMetadata, can be Timeline itself, etc.
3906 117 : fn tree_sort_timelines<T, E>(
3907 117 : timelines: HashMap<TimelineId, T>,
3908 117 : extractor: E,
3909 117 : ) -> anyhow::Result<Vec<(TimelineId, T)>>
3910 117 : where
3911 117 : E: Fn(&T) -> Option<TimelineId>,
3912 117 : {
3913 117 : let mut result = Vec::with_capacity(timelines.len());
3914 117 :
3915 117 : let mut now = Vec::with_capacity(timelines.len());
3916 117 : // (ancestor, children)
3917 117 : let mut later: HashMap<TimelineId, Vec<(TimelineId, T)>> =
3918 117 : HashMap::with_capacity(timelines.len());
3919 :
3920 120 : for (timeline_id, value) in timelines {
3921 3 : if let Some(ancestor_id) = extractor(&value) {
3922 1 : let children = later.entry(ancestor_id).or_default();
3923 1 : children.push((timeline_id, value));
3924 2 : } else {
3925 2 : now.push((timeline_id, value));
3926 2 : }
3927 : }
3928 :
3929 120 : while let Some((timeline_id, metadata)) = now.pop() {
3930 3 : result.push((timeline_id, metadata));
3931 : // All children of this can be loaded now
3932 3 : if let Some(mut children) = later.remove(&timeline_id) {
3933 1 : now.append(&mut children);
3934 2 : }
3935 : }
3936 :
3937 : // All timelines should be visited now. Unless there were timelines with missing ancestors.
3938 117 : if !later.is_empty() {
3939 0 : for (missing_id, orphan_ids) in later {
3940 0 : for (orphan_id, _) in orphan_ids {
3941 0 : error!(
3942 0 : "could not load timeline {orphan_id} because its ancestor timeline {missing_id} could not be loaded"
3943 : );
3944 : }
3945 : }
3946 0 : bail!("could not load tenant because some timelines are missing ancestors");
3947 117 : }
3948 117 :
3949 117 : Ok(result)
3950 117 : }
3951 :
3952 : impl TenantShard {
3953 0 : pub fn tenant_specific_overrides(&self) -> pageserver_api::models::TenantConfig {
3954 0 : self.tenant_conf.load().tenant_conf.clone()
3955 0 : }
3956 :
3957 0 : pub fn effective_config(&self) -> pageserver_api::config::TenantConfigToml {
3958 0 : self.tenant_specific_overrides()
3959 0 : .merge(self.conf.default_tenant_conf.clone())
3960 0 : }
3961 :
3962 0 : pub fn get_checkpoint_distance(&self) -> u64 {
3963 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3964 0 : tenant_conf
3965 0 : .checkpoint_distance
3966 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
3967 0 : }
3968 :
3969 0 : pub fn get_checkpoint_timeout(&self) -> Duration {
3970 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3971 0 : tenant_conf
3972 0 : .checkpoint_timeout
3973 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
3974 0 : }
3975 :
3976 0 : pub fn get_compaction_target_size(&self) -> u64 {
3977 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3978 0 : tenant_conf
3979 0 : .compaction_target_size
3980 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
3981 0 : }
3982 :
3983 0 : pub fn get_compaction_period(&self) -> Duration {
3984 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3985 0 : tenant_conf
3986 0 : .compaction_period
3987 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
3988 0 : }
3989 :
3990 0 : pub fn get_compaction_threshold(&self) -> usize {
3991 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3992 0 : tenant_conf
3993 0 : .compaction_threshold
3994 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
3995 0 : }
3996 :
3997 0 : pub fn get_rel_size_v2_enabled(&self) -> bool {
3998 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3999 0 : tenant_conf
4000 0 : .rel_size_v2_enabled
4001 0 : .unwrap_or(self.conf.default_tenant_conf.rel_size_v2_enabled)
4002 0 : }
4003 :
4004 0 : pub fn get_compaction_upper_limit(&self) -> usize {
4005 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4006 0 : tenant_conf
4007 0 : .compaction_upper_limit
4008 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_upper_limit)
4009 0 : }
4010 :
4011 0 : pub fn get_compaction_l0_first(&self) -> bool {
4012 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4013 0 : tenant_conf
4014 0 : .compaction_l0_first
4015 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_first)
4016 0 : }
4017 :
4018 2 : pub fn get_gc_horizon(&self) -> u64 {
4019 2 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4020 2 : tenant_conf
4021 2 : .gc_horizon
4022 2 : .unwrap_or(self.conf.default_tenant_conf.gc_horizon)
4023 2 : }
4024 :
4025 0 : pub fn get_gc_period(&self) -> Duration {
4026 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4027 0 : tenant_conf
4028 0 : .gc_period
4029 0 : .unwrap_or(self.conf.default_tenant_conf.gc_period)
4030 0 : }
4031 :
4032 0 : pub fn get_image_creation_threshold(&self) -> usize {
4033 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4034 0 : tenant_conf
4035 0 : .image_creation_threshold
4036 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
4037 0 : }
4038 :
4039 2 : pub fn get_pitr_interval(&self) -> Duration {
4040 2 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4041 2 : tenant_conf
4042 2 : .pitr_interval
4043 2 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
4044 2 : }
4045 :
4046 0 : pub fn get_min_resident_size_override(&self) -> Option<u64> {
4047 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4048 0 : tenant_conf
4049 0 : .min_resident_size_override
4050 0 : .or(self.conf.default_tenant_conf.min_resident_size_override)
4051 0 : }
4052 :
4053 0 : pub fn get_heatmap_period(&self) -> Option<Duration> {
4054 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4055 0 : let heatmap_period = tenant_conf
4056 0 : .heatmap_period
4057 0 : .unwrap_or(self.conf.default_tenant_conf.heatmap_period);
4058 0 : if heatmap_period.is_zero() {
4059 0 : None
4060 : } else {
4061 0 : Some(heatmap_period)
4062 : }
4063 0 : }
4064 :
4065 2 : pub fn get_lsn_lease_length(&self) -> Duration {
4066 2 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4067 2 : tenant_conf
4068 2 : .lsn_lease_length
4069 2 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
4070 2 : }
4071 :
4072 0 : pub fn get_timeline_offloading_enabled(&self) -> bool {
4073 0 : if self.conf.timeline_offloading {
4074 0 : return true;
4075 0 : }
4076 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4077 0 : tenant_conf
4078 0 : .timeline_offloading
4079 0 : .unwrap_or(self.conf.default_tenant_conf.timeline_offloading)
4080 0 : }
4081 :
4082 : /// Generate an up-to-date TenantManifest based on the state of this Tenant.
4083 118 : fn build_tenant_manifest(&self) -> TenantManifest {
4084 118 : // Collect the offloaded timelines, and sort them for deterministic output.
4085 118 : let offloaded_timelines = self
4086 118 : .timelines_offloaded
4087 118 : .lock()
4088 118 : .unwrap()
4089 118 : .values()
4090 118 : .map(|tli| tli.manifest())
4091 118 : .sorted_by_key(|m| m.timeline_id)
4092 118 : .collect_vec();
4093 118 :
4094 118 : TenantManifest {
4095 118 : version: LATEST_TENANT_MANIFEST_VERSION,
4096 118 : stripe_size: Some(self.get_shard_stripe_size()),
4097 118 : offloaded_timelines,
4098 118 : }
4099 118 : }
4100 :
4101 0 : pub fn update_tenant_config<
4102 0 : F: Fn(
4103 0 : pageserver_api::models::TenantConfig,
4104 0 : ) -> anyhow::Result<pageserver_api::models::TenantConfig>,
4105 0 : >(
4106 0 : &self,
4107 0 : update: F,
4108 0 : ) -> anyhow::Result<pageserver_api::models::TenantConfig> {
4109 0 : // Use read-copy-update in order to avoid overwriting the location config
4110 0 : // state if this races with [`TenantShard::set_new_location_config`]. Note that
4111 0 : // this race is not possible if both request types come from the storage
4112 0 : // controller (as they should!) because an exclusive op lock is required
4113 0 : // on the storage controller side.
4114 0 :
4115 0 : self.tenant_conf
4116 0 : .try_rcu(|attached_conf| -> Result<_, anyhow::Error> {
4117 0 : Ok(Arc::new(AttachedTenantConf {
4118 0 : tenant_conf: update(attached_conf.tenant_conf.clone())?,
4119 0 : location: attached_conf.location,
4120 0 : lsn_lease_deadline: attached_conf.lsn_lease_deadline,
4121 : }))
4122 0 : })?;
4123 :
4124 0 : let updated = self.tenant_conf.load();
4125 0 :
4126 0 : self.tenant_conf_updated(&updated.tenant_conf);
4127 0 : // Don't hold self.timelines.lock() during the notifies.
4128 0 : // There's no risk of deadlock right now, but there could be if we consolidate
4129 0 : // mutexes in struct Timeline in the future.
4130 0 : let timelines = self.list_timelines();
4131 0 : for timeline in timelines {
4132 0 : timeline.tenant_conf_updated(&updated);
4133 0 : }
4134 :
4135 0 : Ok(updated.tenant_conf.clone())
4136 0 : }
4137 :
4138 0 : pub(crate) fn set_new_location_config(&self, new_conf: AttachedTenantConf) {
4139 0 : let new_tenant_conf = new_conf.tenant_conf.clone();
4140 0 :
4141 0 : self.tenant_conf.store(Arc::new(new_conf.clone()));
4142 0 :
4143 0 : self.tenant_conf_updated(&new_tenant_conf);
4144 0 : // Don't hold self.timelines.lock() during the notifies.
4145 0 : // There's no risk of deadlock right now, but there could be if we consolidate
4146 0 : // mutexes in struct Timeline in the future.
4147 0 : let timelines = self.list_timelines();
4148 0 : for timeline in timelines {
4149 0 : timeline.tenant_conf_updated(&new_conf);
4150 0 : }
4151 0 : }
4152 :
4153 117 : fn get_pagestream_throttle_config(
4154 117 : psconf: &'static PageServerConf,
4155 117 : overrides: &pageserver_api::models::TenantConfig,
4156 117 : ) -> throttle::Config {
4157 117 : overrides
4158 117 : .timeline_get_throttle
4159 117 : .clone()
4160 117 : .unwrap_or(psconf.default_tenant_conf.timeline_get_throttle.clone())
4161 117 : }
4162 :
4163 0 : pub(crate) fn tenant_conf_updated(&self, new_conf: &pageserver_api::models::TenantConfig) {
4164 0 : let conf = Self::get_pagestream_throttle_config(self.conf, new_conf);
4165 0 : self.pagestream_throttle.reconfigure(conf)
4166 0 : }
4167 :
4168 : /// Helper function to create a new Timeline struct.
4169 : ///
4170 : /// The returned Timeline is in Loading state. The caller is responsible for
4171 : /// initializing any on-disk state, and for inserting the Timeline to the 'timelines'
4172 : /// map.
4173 : ///
4174 : /// `validate_ancestor == false` is used when a timeline is created for deletion
4175 : /// and we might not have the ancestor present anymore which is fine for to be
4176 : /// deleted timelines.
4177 : #[allow(clippy::too_many_arguments)]
4178 233 : fn create_timeline_struct(
4179 233 : &self,
4180 233 : new_timeline_id: TimelineId,
4181 233 : new_metadata: &TimelineMetadata,
4182 233 : previous_heatmap: Option<PreviousHeatmap>,
4183 233 : ancestor: Option<Arc<Timeline>>,
4184 233 : resources: TimelineResources,
4185 233 : cause: CreateTimelineCause,
4186 233 : create_idempotency: CreateTimelineIdempotency,
4187 233 : gc_compaction_state: Option<GcCompactionState>,
4188 233 : rel_size_v2_status: Option<RelSizeMigration>,
4189 233 : ctx: &RequestContext,
4190 233 : ) -> anyhow::Result<(Arc<Timeline>, RequestContext)> {
4191 233 : let state = match cause {
4192 : CreateTimelineCause::Load => {
4193 233 : let ancestor_id = new_metadata.ancestor_timeline();
4194 233 : anyhow::ensure!(
4195 233 : ancestor_id == ancestor.as_ref().map(|t| t.timeline_id),
4196 0 : "Timeline's {new_timeline_id} ancestor {ancestor_id:?} was not found"
4197 : );
4198 233 : TimelineState::Loading
4199 : }
4200 0 : CreateTimelineCause::Delete => TimelineState::Stopping,
4201 : };
4202 :
4203 233 : let pg_version = new_metadata.pg_version();
4204 233 :
4205 233 : let timeline = Timeline::new(
4206 233 : self.conf,
4207 233 : Arc::clone(&self.tenant_conf),
4208 233 : new_metadata,
4209 233 : previous_heatmap,
4210 233 : ancestor,
4211 233 : new_timeline_id,
4212 233 : self.tenant_shard_id,
4213 233 : self.generation,
4214 233 : self.shard_identity,
4215 233 : self.walredo_mgr.clone(),
4216 233 : resources,
4217 233 : pg_version,
4218 233 : state,
4219 233 : self.attach_wal_lag_cooldown.clone(),
4220 233 : create_idempotency,
4221 233 : gc_compaction_state,
4222 233 : rel_size_v2_status,
4223 233 : self.cancel.child_token(),
4224 233 : );
4225 233 :
4226 233 : let timeline_ctx = RequestContextBuilder::from(ctx)
4227 233 : .scope(context::Scope::new_timeline(&timeline))
4228 233 : .detached_child();
4229 233 :
4230 233 : Ok((timeline, timeline_ctx))
4231 233 : }
4232 :
4233 : /// [`TenantShard::shutdown`] must be called before dropping the returned [`TenantShard`] object
4234 : /// to ensure proper cleanup of background tasks and metrics.
4235 : //
4236 : // Allow too_many_arguments because a constructor's argument list naturally grows with the
4237 : // number of attributes in the struct: breaking these out into a builder wouldn't be helpful.
4238 : #[allow(clippy::too_many_arguments)]
4239 117 : fn new(
4240 117 : state: TenantState,
4241 117 : conf: &'static PageServerConf,
4242 117 : attached_conf: AttachedTenantConf,
4243 117 : shard_identity: ShardIdentity,
4244 117 : walredo_mgr: Option<Arc<WalRedoManager>>,
4245 117 : tenant_shard_id: TenantShardId,
4246 117 : remote_storage: GenericRemoteStorage,
4247 117 : deletion_queue_client: DeletionQueueClient,
4248 117 : l0_flush_global_state: L0FlushGlobalState,
4249 117 : basebackup_prepare_sender: BasebackupPrepareSender,
4250 117 : ) -> TenantShard {
4251 117 : assert!(!attached_conf.location.generation.is_none());
4252 :
4253 117 : let (state, mut rx) = watch::channel(state);
4254 117 :
4255 117 : tokio::spawn(async move {
4256 117 : // reflect tenant state in metrics:
4257 117 : // - global per tenant state: TENANT_STATE_METRIC
4258 117 : // - "set" of broken tenants: BROKEN_TENANTS_SET
4259 117 : //
4260 117 : // set of broken tenants should not have zero counts so that it remains accessible for
4261 117 : // alerting.
4262 117 :
4263 117 : let tid = tenant_shard_id.to_string();
4264 117 : let shard_id = tenant_shard_id.shard_slug().to_string();
4265 117 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
4266 :
4267 233 : fn inspect_state(state: &TenantState) -> ([&'static str; 1], bool) {
4268 233 : ([state.into()], matches!(state, TenantState::Broken { .. }))
4269 233 : }
4270 :
4271 117 : let mut tuple = inspect_state(&rx.borrow_and_update());
4272 117 :
4273 117 : let is_broken = tuple.1;
4274 117 : let mut counted_broken = if is_broken {
4275 : // add the id to the set right away, there should not be any updates on the channel
4276 : // after before tenant is removed, if ever
4277 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
4278 0 : true
4279 : } else {
4280 117 : false
4281 : };
4282 :
4283 : loop {
4284 233 : let labels = &tuple.0;
4285 233 : let current = TENANT_STATE_METRIC.with_label_values(labels);
4286 233 : current.inc();
4287 233 :
4288 233 : if rx.changed().await.is_err() {
4289 : // tenant has been dropped
4290 7 : current.dec();
4291 7 : drop(BROKEN_TENANTS_SET.remove_label_values(set_key));
4292 7 : break;
4293 116 : }
4294 116 :
4295 116 : current.dec();
4296 116 : tuple = inspect_state(&rx.borrow_and_update());
4297 116 :
4298 116 : let is_broken = tuple.1;
4299 116 : if is_broken && !counted_broken {
4300 0 : counted_broken = true;
4301 0 : // insert the tenant_id (back) into the set while avoiding needless counter
4302 0 : // access
4303 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
4304 116 : }
4305 : }
4306 117 : });
4307 117 :
4308 117 : TenantShard {
4309 117 : tenant_shard_id,
4310 117 : shard_identity,
4311 117 : generation: attached_conf.location.generation,
4312 117 : conf,
4313 117 : // using now here is good enough approximation to catch tenants with really long
4314 117 : // activation times.
4315 117 : constructed_at: Instant::now(),
4316 117 : timelines: Mutex::new(HashMap::new()),
4317 117 : timelines_creating: Mutex::new(HashSet::new()),
4318 117 : timelines_offloaded: Mutex::new(HashMap::new()),
4319 117 : timelines_importing: Mutex::new(HashMap::new()),
4320 117 : remote_tenant_manifest: Default::default(),
4321 117 : gc_cs: tokio::sync::Mutex::new(()),
4322 117 : walredo_mgr,
4323 117 : remote_storage,
4324 117 : deletion_queue_client,
4325 117 : state,
4326 117 : cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
4327 117 : cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
4328 117 : eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
4329 117 : compaction_circuit_breaker: std::sync::Mutex::new(CircuitBreaker::new(
4330 117 : format!("compaction-{tenant_shard_id}"),
4331 117 : 5,
4332 117 : // Compaction can be a very expensive operation, and might leak disk space. It also ought
4333 117 : // to be infallible, as long as remote storage is available. So if it repeatedly fails,
4334 117 : // use an extremely long backoff.
4335 117 : Some(Duration::from_secs(3600 * 24)),
4336 117 : )),
4337 117 : l0_compaction_trigger: Arc::new(Notify::new()),
4338 117 : scheduled_compaction_tasks: Mutex::new(Default::default()),
4339 117 : activate_now_sem: tokio::sync::Semaphore::new(0),
4340 117 : attach_wal_lag_cooldown: Arc::new(std::sync::OnceLock::new()),
4341 117 : cancel: CancellationToken::default(),
4342 117 : gate: Gate::default(),
4343 117 : pagestream_throttle: Arc::new(throttle::Throttle::new(
4344 117 : TenantShard::get_pagestream_throttle_config(conf, &attached_conf.tenant_conf),
4345 117 : )),
4346 117 : pagestream_throttle_metrics: Arc::new(
4347 117 : crate::metrics::tenant_throttling::Pagestream::new(&tenant_shard_id),
4348 117 : ),
4349 117 : tenant_conf: Arc::new(ArcSwap::from_pointee(attached_conf)),
4350 117 : ongoing_timeline_detach: std::sync::Mutex::default(),
4351 117 : gc_block: Default::default(),
4352 117 : l0_flush_global_state,
4353 117 : basebackup_prepare_sender,
4354 117 : }
4355 117 : }
4356 :
4357 : /// Locate and load config
4358 0 : pub(super) fn load_tenant_config(
4359 0 : conf: &'static PageServerConf,
4360 0 : tenant_shard_id: &TenantShardId,
4361 0 : ) -> Result<LocationConf, LoadConfigError> {
4362 0 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4363 0 :
4364 0 : info!("loading tenant configuration from {config_path}");
4365 :
4366 : // load and parse file
4367 0 : let config = fs::read_to_string(&config_path).map_err(|e| {
4368 0 : match e.kind() {
4369 : std::io::ErrorKind::NotFound => {
4370 : // The config should almost always exist for a tenant directory:
4371 : // - When attaching a tenant, the config is the first thing we write
4372 : // - When detaching a tenant, we atomically move the directory to a tmp location
4373 : // before deleting contents.
4374 : //
4375 : // The very rare edge case that can result in a missing config is if we crash during attach
4376 : // between creating directory and writing config. Callers should handle that as if the
4377 : // directory didn't exist.
4378 :
4379 0 : LoadConfigError::NotFound(config_path)
4380 : }
4381 : _ => {
4382 : // No IO errors except NotFound are acceptable here: other kinds of error indicate local storage or permissions issues
4383 : // that we cannot cleanly recover
4384 0 : crate::virtual_file::on_fatal_io_error(&e, "Reading tenant config file")
4385 : }
4386 : }
4387 0 : })?;
4388 :
4389 0 : Ok(toml_edit::de::from_str::<LocationConf>(&config)?)
4390 0 : }
4391 :
4392 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4393 : pub(super) async fn persist_tenant_config(
4394 : conf: &'static PageServerConf,
4395 : tenant_shard_id: &TenantShardId,
4396 : location_conf: &LocationConf,
4397 : ) -> std::io::Result<()> {
4398 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4399 :
4400 : Self::persist_tenant_config_at(tenant_shard_id, &config_path, location_conf).await
4401 : }
4402 :
4403 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4404 : pub(super) async fn persist_tenant_config_at(
4405 : tenant_shard_id: &TenantShardId,
4406 : config_path: &Utf8Path,
4407 : location_conf: &LocationConf,
4408 : ) -> std::io::Result<()> {
4409 : debug!("persisting tenantconf to {config_path}");
4410 :
4411 : let mut conf_content = r#"# This file contains a specific per-tenant's config.
4412 : # It is read in case of pageserver restart.
4413 : "#
4414 : .to_string();
4415 :
4416 0 : fail::fail_point!("tenant-config-before-write", |_| {
4417 0 : Err(std::io::Error::other("tenant-config-before-write"))
4418 0 : });
4419 :
4420 : // Convert the config to a toml file.
4421 : conf_content +=
4422 : &toml_edit::ser::to_string_pretty(&location_conf).expect("Config serialization failed");
4423 :
4424 : let temp_path = path_with_suffix_extension(config_path, TEMP_FILE_SUFFIX);
4425 :
4426 : let conf_content = conf_content.into_bytes();
4427 : VirtualFile::crashsafe_overwrite(config_path.to_owned(), temp_path, conf_content).await
4428 : }
4429 :
4430 : //
4431 : // How garbage collection works:
4432 : //
4433 : // +--bar------------->
4434 : // /
4435 : // +----+-----foo---------------->
4436 : // /
4437 : // ----main--+-------------------------->
4438 : // \
4439 : // +-----baz-------->
4440 : //
4441 : //
4442 : // 1. Grab 'gc_cs' mutex to prevent new timelines from being created while Timeline's
4443 : // `gc_infos` are being refreshed
4444 : // 2. Scan collected timelines, and on each timeline, make note of the
4445 : // all the points where other timelines have been branched off.
4446 : // We will refrain from removing page versions at those LSNs.
4447 : // 3. For each timeline, scan all layer files on the timeline.
4448 : // Remove all files for which a newer file exists and which
4449 : // don't cover any branch point LSNs.
4450 : //
4451 : // TODO:
4452 : // - if a relation has a non-incremental persistent layer on a child branch, then we
4453 : // don't need to keep that in the parent anymore. But currently
4454 : // we do.
4455 2 : async fn gc_iteration_internal(
4456 2 : &self,
4457 2 : target_timeline_id: Option<TimelineId>,
4458 2 : horizon: u64,
4459 2 : pitr: Duration,
4460 2 : cancel: &CancellationToken,
4461 2 : ctx: &RequestContext,
4462 2 : ) -> Result<GcResult, GcError> {
4463 2 : let mut totals: GcResult = Default::default();
4464 2 : let now = Instant::now();
4465 :
4466 2 : let gc_timelines = self
4467 2 : .refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4468 2 : .await?;
4469 :
4470 2 : failpoint_support::sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
4471 :
4472 : // If there is nothing to GC, we don't want any messages in the INFO log.
4473 2 : if !gc_timelines.is_empty() {
4474 2 : info!("{} timelines need GC", gc_timelines.len());
4475 : } else {
4476 0 : debug!("{} timelines need GC", gc_timelines.len());
4477 : }
4478 :
4479 : // Perform GC for each timeline.
4480 : //
4481 : // Note that we don't hold the `TenantShard::gc_cs` lock here because we don't want to delay the
4482 : // branch creation task, which requires the GC lock. A GC iteration can run concurrently
4483 : // with branch creation.
4484 : //
4485 : // See comments in [`TenantShard::branch_timeline`] for more information about why branch
4486 : // creation task can run concurrently with timeline's GC iteration.
4487 4 : for timeline in gc_timelines {
4488 2 : if cancel.is_cancelled() {
4489 : // We were requested to shut down. Stop and return with the progress we
4490 : // made.
4491 0 : break;
4492 2 : }
4493 2 : let result = match timeline.gc().await {
4494 : Err(GcError::TimelineCancelled) => {
4495 0 : if target_timeline_id.is_some() {
4496 : // If we were targetting this specific timeline, surface cancellation to caller
4497 0 : return Err(GcError::TimelineCancelled);
4498 : } else {
4499 : // A timeline may be shutting down independently of the tenant's lifecycle: we should
4500 : // skip past this and proceed to try GC on other timelines.
4501 0 : continue;
4502 : }
4503 : }
4504 2 : r => r?,
4505 : };
4506 2 : totals += result;
4507 : }
4508 :
4509 2 : totals.elapsed = now.elapsed();
4510 2 : Ok(totals)
4511 2 : }
4512 :
4513 : /// Refreshes the Timeline::gc_info for all timelines, returning the
4514 : /// vector of timelines which have [`Timeline::get_last_record_lsn`] past
4515 : /// [`TenantShard::get_gc_horizon`].
4516 : ///
4517 : /// This is usually executed as part of periodic gc, but can now be triggered more often.
4518 2 : pub(crate) async fn refresh_gc_info(
4519 2 : &self,
4520 2 : cancel: &CancellationToken,
4521 2 : ctx: &RequestContext,
4522 2 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4523 2 : // since this method can now be called at different rates than the configured gc loop, it
4524 2 : // might be that these configuration values get applied faster than what it was previously,
4525 2 : // since these were only read from the gc task.
4526 2 : let horizon = self.get_gc_horizon();
4527 2 : let pitr = self.get_pitr_interval();
4528 2 :
4529 2 : // refresh all timelines
4530 2 : let target_timeline_id = None;
4531 2 :
4532 2 : self.refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4533 2 : .await
4534 2 : }
4535 :
4536 : /// Populate all Timelines' `GcInfo` with information about their children. We do not set the
4537 : /// PITR cutoffs here, because that requires I/O: this is done later, before GC, by [`Self::refresh_gc_info_internal`]
4538 : ///
4539 : /// Subsequently, parent-child relationships are updated incrementally inside [`Timeline::new`] and [`Timeline::drop`].
4540 0 : fn initialize_gc_info(
4541 0 : &self,
4542 0 : timelines: &std::sync::MutexGuard<HashMap<TimelineId, Arc<Timeline>>>,
4543 0 : timelines_offloaded: &std::sync::MutexGuard<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
4544 0 : restrict_to_timeline: Option<TimelineId>,
4545 0 : ) {
4546 0 : if restrict_to_timeline.is_none() {
4547 : // This function must be called before activation: after activation timeline create/delete operations
4548 : // might happen, and this function is not safe to run concurrently with those.
4549 0 : assert!(!self.is_active());
4550 0 : }
4551 :
4552 : // Scan all timelines. For each timeline, remember the timeline ID and
4553 : // the branch point where it was created.
4554 0 : let mut all_branchpoints: BTreeMap<TimelineId, Vec<(Lsn, TimelineId, MaybeOffloaded)>> =
4555 0 : BTreeMap::new();
4556 0 : timelines.iter().for_each(|(timeline_id, timeline_entry)| {
4557 0 : if let Some(ancestor_timeline_id) = &timeline_entry.get_ancestor_timeline_id() {
4558 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4559 0 : ancestor_children.push((
4560 0 : timeline_entry.get_ancestor_lsn(),
4561 0 : *timeline_id,
4562 0 : MaybeOffloaded::No,
4563 0 : ));
4564 0 : }
4565 0 : });
4566 0 : timelines_offloaded
4567 0 : .iter()
4568 0 : .for_each(|(timeline_id, timeline_entry)| {
4569 0 : let Some(ancestor_timeline_id) = &timeline_entry.ancestor_timeline_id else {
4570 0 : return;
4571 : };
4572 0 : let Some(retain_lsn) = timeline_entry.ancestor_retain_lsn else {
4573 0 : return;
4574 : };
4575 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4576 0 : ancestor_children.push((retain_lsn, *timeline_id, MaybeOffloaded::Yes));
4577 0 : });
4578 0 :
4579 0 : // The number of bytes we always keep, irrespective of PITR: this is a constant across timelines
4580 0 : let horizon = self.get_gc_horizon();
4581 :
4582 : // Populate each timeline's GcInfo with information about its child branches
4583 0 : let timelines_to_write = if let Some(timeline_id) = restrict_to_timeline {
4584 0 : itertools::Either::Left(timelines.get(&timeline_id).into_iter())
4585 : } else {
4586 0 : itertools::Either::Right(timelines.values())
4587 : };
4588 0 : for timeline in timelines_to_write {
4589 0 : let mut branchpoints: Vec<(Lsn, TimelineId, MaybeOffloaded)> = all_branchpoints
4590 0 : .remove(&timeline.timeline_id)
4591 0 : .unwrap_or_default();
4592 0 :
4593 0 : branchpoints.sort_by_key(|b| b.0);
4594 0 :
4595 0 : let mut target = timeline.gc_info.write().unwrap();
4596 0 :
4597 0 : target.retain_lsns = branchpoints;
4598 0 :
4599 0 : let space_cutoff = timeline
4600 0 : .get_last_record_lsn()
4601 0 : .checked_sub(horizon)
4602 0 : .unwrap_or(Lsn(0));
4603 0 :
4604 0 : target.cutoffs = GcCutoffs {
4605 0 : space: space_cutoff,
4606 0 : time: None,
4607 0 : };
4608 0 : }
4609 0 : }
4610 :
4611 4 : async fn refresh_gc_info_internal(
4612 4 : &self,
4613 4 : target_timeline_id: Option<TimelineId>,
4614 4 : horizon: u64,
4615 4 : pitr: Duration,
4616 4 : cancel: &CancellationToken,
4617 4 : ctx: &RequestContext,
4618 4 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4619 4 : // before taking the gc_cs lock, do the heavier weight finding of gc_cutoff points for
4620 4 : // currently visible timelines.
4621 4 : let timelines = self
4622 4 : .timelines
4623 4 : .lock()
4624 4 : .unwrap()
4625 4 : .values()
4626 10 : .filter(|tl| match target_timeline_id.as_ref() {
4627 2 : Some(target) => &tl.timeline_id == target,
4628 8 : None => true,
4629 10 : })
4630 4 : .cloned()
4631 4 : .collect::<Vec<_>>();
4632 4 :
4633 4 : if target_timeline_id.is_some() && timelines.is_empty() {
4634 : // We were to act on a particular timeline and it wasn't found
4635 0 : return Err(GcError::TimelineNotFound);
4636 4 : }
4637 4 :
4638 4 : let mut gc_cutoffs: HashMap<TimelineId, GcCutoffs> =
4639 4 : HashMap::with_capacity(timelines.len());
4640 4 :
4641 4 : // Ensures all timelines use the same start time when computing the time cutoff.
4642 4 : let now_ts_for_pitr_calc = SystemTime::now();
4643 10 : for timeline in timelines.iter() {
4644 10 : let ctx = &ctx.with_scope_timeline(timeline);
4645 10 : let cutoff = timeline
4646 10 : .get_last_record_lsn()
4647 10 : .checked_sub(horizon)
4648 10 : .unwrap_or(Lsn(0));
4649 :
4650 10 : let cutoffs = timeline
4651 10 : .find_gc_cutoffs(now_ts_for_pitr_calc, cutoff, pitr, cancel, ctx)
4652 10 : .await?;
4653 10 : let old = gc_cutoffs.insert(timeline.timeline_id, cutoffs);
4654 10 : assert!(old.is_none());
4655 : }
4656 :
4657 4 : if !self.is_active() || self.cancel.is_cancelled() {
4658 0 : return Err(GcError::TenantCancelled);
4659 4 : }
4660 :
4661 : // grab mutex to prevent new timelines from being created here; avoid doing long operations
4662 : // because that will stall branch creation.
4663 4 : let gc_cs = self.gc_cs.lock().await;
4664 :
4665 : // Ok, we now know all the branch points.
4666 : // Update the GC information for each timeline.
4667 4 : let mut gc_timelines = Vec::with_capacity(timelines.len());
4668 14 : for timeline in timelines {
4669 : // We filtered the timeline list above
4670 10 : if let Some(target_timeline_id) = target_timeline_id {
4671 2 : assert_eq!(target_timeline_id, timeline.timeline_id);
4672 8 : }
4673 :
4674 : {
4675 10 : let mut target = timeline.gc_info.write().unwrap();
4676 10 :
4677 10 : // Cull any expired leases
4678 10 : let now = SystemTime::now();
4679 10 : target.leases.retain(|_, lease| !lease.is_expired(&now));
4680 10 :
4681 10 : timeline
4682 10 : .metrics
4683 10 : .valid_lsn_lease_count_gauge
4684 10 : .set(target.leases.len() as u64);
4685 :
4686 : // Look up parent's PITR cutoff to update the child's knowledge of whether it is within parent's PITR
4687 10 : if let Some(ancestor_id) = timeline.get_ancestor_timeline_id() {
4688 6 : if let Some(ancestor_gc_cutoffs) = gc_cutoffs.get(&ancestor_id) {
4689 6 : target.within_ancestor_pitr =
4690 6 : Some(timeline.get_ancestor_lsn()) >= ancestor_gc_cutoffs.time;
4691 6 : }
4692 4 : }
4693 :
4694 : // Update metrics that depend on GC state
4695 10 : timeline
4696 10 : .metrics
4697 10 : .archival_size
4698 10 : .set(if target.within_ancestor_pitr {
4699 0 : timeline.metrics.current_logical_size_gauge.get()
4700 : } else {
4701 10 : 0
4702 : });
4703 10 : if let Some(time_cutoff) = target.cutoffs.time {
4704 4 : timeline.metrics.pitr_history_size.set(
4705 4 : timeline
4706 4 : .get_last_record_lsn()
4707 4 : .checked_sub(time_cutoff)
4708 4 : .unwrap_or_default()
4709 4 : .0,
4710 4 : );
4711 6 : }
4712 :
4713 : // Apply the cutoffs we found to the Timeline's GcInfo. Why might we _not_ have cutoffs for a timeline?
4714 : // - this timeline was created while we were finding cutoffs
4715 : // - lsn for timestamp search fails for this timeline repeatedly
4716 10 : if let Some(cutoffs) = gc_cutoffs.get(&timeline.timeline_id) {
4717 10 : let original_cutoffs = target.cutoffs.clone();
4718 10 : // GC cutoffs should never go back
4719 10 : target.cutoffs = GcCutoffs {
4720 10 : space: cutoffs.space.max(original_cutoffs.space),
4721 10 : time: cutoffs.time.max(original_cutoffs.time),
4722 10 : }
4723 0 : }
4724 : }
4725 :
4726 10 : gc_timelines.push(timeline);
4727 : }
4728 4 : drop(gc_cs);
4729 4 : Ok(gc_timelines)
4730 4 : }
4731 :
4732 : /// A substitute for `branch_timeline` for use in unit tests.
4733 : /// The returned timeline will have state value `Active` to make various `anyhow::ensure!()`
4734 : /// calls pass, but, we do not actually call `.activate()` under the hood. So, none of the
4735 : /// timeline background tasks are launched, except the flush loop.
4736 : #[cfg(test)]
4737 119 : async fn branch_timeline_test(
4738 119 : self: &Arc<Self>,
4739 119 : src_timeline: &Arc<Timeline>,
4740 119 : dst_id: TimelineId,
4741 119 : ancestor_lsn: Option<Lsn>,
4742 119 : ctx: &RequestContext,
4743 119 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
4744 119 : let tl = self
4745 119 : .branch_timeline_impl(src_timeline, dst_id, ancestor_lsn, ctx)
4746 119 : .await?
4747 117 : .into_timeline_for_test();
4748 117 : tl.set_state(TimelineState::Active);
4749 117 : Ok(tl)
4750 119 : }
4751 :
4752 : /// Helper for unit tests to branch a timeline with some pre-loaded states.
4753 : #[cfg(test)]
4754 : #[allow(clippy::too_many_arguments)]
4755 6 : pub async fn branch_timeline_test_with_layers(
4756 6 : self: &Arc<Self>,
4757 6 : src_timeline: &Arc<Timeline>,
4758 6 : dst_id: TimelineId,
4759 6 : ancestor_lsn: Option<Lsn>,
4760 6 : ctx: &RequestContext,
4761 6 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
4762 6 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
4763 6 : end_lsn: Lsn,
4764 6 : ) -> anyhow::Result<Arc<Timeline>> {
4765 : use checks::check_valid_layermap;
4766 : use itertools::Itertools;
4767 :
4768 6 : let tline = self
4769 6 : .branch_timeline_test(src_timeline, dst_id, ancestor_lsn, ctx)
4770 6 : .await?;
4771 6 : let ancestor_lsn = if let Some(ancestor_lsn) = ancestor_lsn {
4772 6 : ancestor_lsn
4773 : } else {
4774 0 : tline.get_last_record_lsn()
4775 : };
4776 6 : assert!(end_lsn >= ancestor_lsn);
4777 6 : tline.force_advance_lsn(end_lsn);
4778 9 : for deltas in delta_layer_desc {
4779 3 : tline
4780 3 : .force_create_delta_layer(deltas, Some(ancestor_lsn), ctx)
4781 3 : .await?;
4782 : }
4783 8 : for (lsn, images) in image_layer_desc {
4784 2 : tline
4785 2 : .force_create_image_layer(lsn, images, Some(ancestor_lsn), ctx)
4786 2 : .await?;
4787 : }
4788 6 : let layer_names = tline
4789 6 : .layers
4790 6 : .read()
4791 6 : .await
4792 6 : .layer_map()
4793 6 : .unwrap()
4794 6 : .iter_historic_layers()
4795 6 : .map(|layer| layer.layer_name())
4796 6 : .collect_vec();
4797 6 : if let Some(err) = check_valid_layermap(&layer_names) {
4798 0 : bail!("invalid layermap: {err}");
4799 6 : }
4800 6 : Ok(tline)
4801 6 : }
4802 :
4803 : /// Branch an existing timeline.
4804 0 : async fn branch_timeline(
4805 0 : self: &Arc<Self>,
4806 0 : src_timeline: &Arc<Timeline>,
4807 0 : dst_id: TimelineId,
4808 0 : start_lsn: Option<Lsn>,
4809 0 : ctx: &RequestContext,
4810 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4811 0 : self.branch_timeline_impl(src_timeline, dst_id, start_lsn, ctx)
4812 0 : .await
4813 0 : }
4814 :
4815 119 : async fn branch_timeline_impl(
4816 119 : self: &Arc<Self>,
4817 119 : src_timeline: &Arc<Timeline>,
4818 119 : dst_id: TimelineId,
4819 119 : start_lsn: Option<Lsn>,
4820 119 : ctx: &RequestContext,
4821 119 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4822 119 : let src_id = src_timeline.timeline_id;
4823 :
4824 : // We will validate our ancestor LSN in this function. Acquire the GC lock so that
4825 : // this check cannot race with GC, and the ancestor LSN is guaranteed to remain
4826 : // valid while we are creating the branch.
4827 119 : let _gc_cs = self.gc_cs.lock().await;
4828 :
4829 : // If no start LSN is specified, we branch the new timeline from the source timeline's last record LSN
4830 119 : let start_lsn = start_lsn.unwrap_or_else(|| {
4831 1 : let lsn = src_timeline.get_last_record_lsn();
4832 1 : info!("branching timeline {dst_id} from timeline {src_id} at last record LSN: {lsn}");
4833 1 : lsn
4834 119 : });
4835 :
4836 : // we finally have determined the ancestor_start_lsn, so we can get claim exclusivity now
4837 119 : let timeline_create_guard = match self
4838 119 : .start_creating_timeline(
4839 119 : dst_id,
4840 119 : CreateTimelineIdempotency::Branch {
4841 119 : ancestor_timeline_id: src_timeline.timeline_id,
4842 119 : ancestor_start_lsn: start_lsn,
4843 119 : },
4844 119 : )
4845 119 : .await?
4846 : {
4847 119 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
4848 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
4849 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
4850 : }
4851 : };
4852 :
4853 : // Ensure that `start_lsn` is valid, i.e. the LSN is within the PITR
4854 : // horizon on the source timeline
4855 : //
4856 : // We check it against both the planned GC cutoff stored in 'gc_info',
4857 : // and the 'latest_gc_cutoff' of the last GC that was performed. The
4858 : // planned GC cutoff in 'gc_info' is normally larger than
4859 : // 'applied_gc_cutoff_lsn', but beware of corner cases like if you just
4860 : // changed the GC settings for the tenant to make the PITR window
4861 : // larger, but some of the data was already removed by an earlier GC
4862 : // iteration.
4863 :
4864 : // check against last actual 'latest_gc_cutoff' first
4865 119 : let applied_gc_cutoff_lsn = src_timeline.get_applied_gc_cutoff_lsn();
4866 119 : {
4867 119 : let gc_info = src_timeline.gc_info.read().unwrap();
4868 119 : let planned_cutoff = gc_info.min_cutoff();
4869 119 : if gc_info.lsn_covered_by_lease(start_lsn) {
4870 0 : tracing::info!(
4871 0 : "skipping comparison of {start_lsn} with gc cutoff {} and planned gc cutoff {planned_cutoff} due to lsn lease",
4872 0 : *applied_gc_cutoff_lsn
4873 : );
4874 : } else {
4875 119 : src_timeline
4876 119 : .check_lsn_is_in_scope(start_lsn, &applied_gc_cutoff_lsn)
4877 119 : .context(format!(
4878 119 : "invalid branch start lsn: less than latest GC cutoff {}",
4879 119 : *applied_gc_cutoff_lsn,
4880 119 : ))
4881 119 : .map_err(CreateTimelineError::AncestorLsn)?;
4882 :
4883 : // and then the planned GC cutoff
4884 117 : if start_lsn < planned_cutoff {
4885 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
4886 0 : "invalid branch start lsn: less than planned GC cutoff {planned_cutoff}"
4887 0 : )));
4888 117 : }
4889 : }
4890 : }
4891 :
4892 : //
4893 : // The branch point is valid, and we are still holding the 'gc_cs' lock
4894 : // so that GC cannot advance the GC cutoff until we are finished.
4895 : // Proceed with the branch creation.
4896 : //
4897 :
4898 : // Determine prev-LSN for the new timeline. We can only determine it if
4899 : // the timeline was branched at the current end of the source timeline.
4900 : let RecordLsn {
4901 117 : last: src_last,
4902 117 : prev: src_prev,
4903 117 : } = src_timeline.get_last_record_rlsn();
4904 117 : let dst_prev = if src_last == start_lsn {
4905 108 : Some(src_prev)
4906 : } else {
4907 9 : None
4908 : };
4909 :
4910 : // Create the metadata file, noting the ancestor of the new timeline.
4911 : // There is initially no data in it, but all the read-calls know to look
4912 : // into the ancestor.
4913 117 : let metadata = TimelineMetadata::new(
4914 117 : start_lsn,
4915 117 : dst_prev,
4916 117 : Some(src_id),
4917 117 : start_lsn,
4918 117 : *src_timeline.applied_gc_cutoff_lsn.read(), // FIXME: should we hold onto this guard longer?
4919 117 : src_timeline.initdb_lsn,
4920 117 : src_timeline.pg_version,
4921 117 : );
4922 :
4923 117 : let (uninitialized_timeline, _timeline_ctx) = self
4924 117 : .prepare_new_timeline(
4925 117 : dst_id,
4926 117 : &metadata,
4927 117 : timeline_create_guard,
4928 117 : start_lsn + 1,
4929 117 : Some(Arc::clone(src_timeline)),
4930 117 : Some(src_timeline.get_rel_size_v2_status()),
4931 117 : ctx,
4932 117 : )
4933 117 : .await?;
4934 :
4935 117 : let new_timeline = uninitialized_timeline.finish_creation().await?;
4936 :
4937 : // Root timeline gets its layers during creation and uploads them along with the metadata.
4938 : // A branch timeline though, when created, can get no writes for some time, hence won't get any layers created.
4939 : // We still need to upload its metadata eagerly: if other nodes `attach` the tenant and miss this timeline, their GC
4940 : // could get incorrect information and remove more layers, than needed.
4941 : // See also https://github.com/neondatabase/neon/issues/3865
4942 117 : new_timeline
4943 117 : .remote_client
4944 117 : .schedule_index_upload_for_full_metadata_update(&metadata)
4945 117 : .context("branch initial metadata upload")?;
4946 :
4947 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
4948 :
4949 117 : Ok(CreateTimelineResult::Created(new_timeline))
4950 119 : }
4951 :
4952 : /// For unit tests, make this visible so that other modules can directly create timelines
4953 : #[cfg(test)]
4954 : #[tracing::instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
4955 : pub(crate) async fn bootstrap_timeline_test(
4956 : self: &Arc<Self>,
4957 : timeline_id: TimelineId,
4958 : pg_version: u32,
4959 : load_existing_initdb: Option<TimelineId>,
4960 : ctx: &RequestContext,
4961 : ) -> anyhow::Result<Arc<Timeline>> {
4962 : self.bootstrap_timeline(timeline_id, pg_version, load_existing_initdb, ctx)
4963 : .await
4964 : .map_err(anyhow::Error::new)
4965 1 : .map(|r| r.into_timeline_for_test())
4966 : }
4967 :
4968 : /// Get exclusive access to the timeline ID for creation.
4969 : ///
4970 : /// Timeline-creating code paths must use this function before making changes
4971 : /// to in-memory or persistent state.
4972 : ///
4973 : /// The `state` parameter is a description of the timeline creation operation
4974 : /// we intend to perform.
4975 : /// If the timeline was already created in the meantime, we check whether this
4976 : /// request conflicts or is idempotent , based on `state`.
4977 233 : async fn start_creating_timeline(
4978 233 : self: &Arc<Self>,
4979 233 : new_timeline_id: TimelineId,
4980 233 : idempotency: CreateTimelineIdempotency,
4981 233 : ) -> Result<StartCreatingTimelineResult, CreateTimelineError> {
4982 233 : let allow_offloaded = false;
4983 233 : match self.create_timeline_create_guard(new_timeline_id, idempotency, allow_offloaded) {
4984 232 : Ok(create_guard) => {
4985 232 : pausable_failpoint!("timeline-creation-after-uninit");
4986 232 : Ok(StartCreatingTimelineResult::CreateGuard(create_guard))
4987 : }
4988 0 : Err(TimelineExclusionError::ShuttingDown) => Err(CreateTimelineError::ShuttingDown),
4989 : Err(TimelineExclusionError::AlreadyCreating) => {
4990 : // Creation is in progress, we cannot create it again, and we cannot
4991 : // check if this request matches the existing one, so caller must try
4992 : // again later.
4993 0 : Err(CreateTimelineError::AlreadyCreating)
4994 : }
4995 0 : Err(TimelineExclusionError::Other(e)) => Err(CreateTimelineError::Other(e)),
4996 : Err(TimelineExclusionError::AlreadyExists {
4997 0 : existing: TimelineOrOffloaded::Offloaded(_existing),
4998 0 : ..
4999 0 : }) => {
5000 0 : info!("timeline already exists but is offloaded");
5001 0 : Err(CreateTimelineError::Conflict)
5002 : }
5003 : Err(TimelineExclusionError::AlreadyExists {
5004 1 : existing: TimelineOrOffloaded::Timeline(existing),
5005 1 : arg,
5006 1 : }) => {
5007 1 : {
5008 1 : let existing = &existing.create_idempotency;
5009 1 : let _span = info_span!("idempotency_check", ?existing, ?arg).entered();
5010 1 : debug!("timeline already exists");
5011 :
5012 1 : match (existing, &arg) {
5013 : // FailWithConflict => no idempotency check
5014 : (CreateTimelineIdempotency::FailWithConflict, _)
5015 : | (_, CreateTimelineIdempotency::FailWithConflict) => {
5016 1 : warn!("timeline already exists, failing request");
5017 1 : return Err(CreateTimelineError::Conflict);
5018 : }
5019 : // Idempotent <=> CreateTimelineIdempotency is identical
5020 0 : (x, y) if x == y => {
5021 0 : info!(
5022 0 : "timeline already exists and idempotency matches, succeeding request"
5023 : );
5024 : // fallthrough
5025 : }
5026 : (_, _) => {
5027 0 : warn!("idempotency conflict, failing request");
5028 0 : return Err(CreateTimelineError::Conflict);
5029 : }
5030 : }
5031 : }
5032 :
5033 0 : Ok(StartCreatingTimelineResult::Idempotent(existing))
5034 : }
5035 : }
5036 233 : }
5037 :
5038 0 : async fn upload_initdb(
5039 0 : &self,
5040 0 : timelines_path: &Utf8PathBuf,
5041 0 : pgdata_path: &Utf8PathBuf,
5042 0 : timeline_id: &TimelineId,
5043 0 : ) -> anyhow::Result<()> {
5044 0 : let temp_path = timelines_path.join(format!(
5045 0 : "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
5046 0 : ));
5047 0 :
5048 0 : scopeguard::defer! {
5049 0 : if let Err(e) = fs::remove_file(&temp_path) {
5050 0 : error!("Failed to remove temporary initdb archive '{temp_path}': {e}");
5051 0 : }
5052 0 : }
5053 :
5054 0 : let (pgdata_zstd, tar_zst_size) = create_zst_tarball(pgdata_path, &temp_path).await?;
5055 : const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
5056 0 : if tar_zst_size > INITDB_TAR_ZST_WARN_LIMIT {
5057 0 : warn!(
5058 0 : "compressed {temp_path} size of {tar_zst_size} is above limit {INITDB_TAR_ZST_WARN_LIMIT}."
5059 : );
5060 0 : }
5061 :
5062 0 : pausable_failpoint!("before-initdb-upload");
5063 :
5064 0 : backoff::retry(
5065 0 : || async {
5066 0 : self::remote_timeline_client::upload_initdb_dir(
5067 0 : &self.remote_storage,
5068 0 : &self.tenant_shard_id.tenant_id,
5069 0 : timeline_id,
5070 0 : pgdata_zstd.try_clone().await?,
5071 0 : tar_zst_size,
5072 0 : &self.cancel,
5073 0 : )
5074 0 : .await
5075 0 : },
5076 0 : |_| false,
5077 0 : 3,
5078 0 : u32::MAX,
5079 0 : "persist_initdb_tar_zst",
5080 0 : &self.cancel,
5081 0 : )
5082 0 : .await
5083 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
5084 0 : .and_then(|x| x)
5085 0 : }
5086 :
5087 : /// - run initdb to init temporary instance and get bootstrap data
5088 : /// - after initialization completes, tar up the temp dir and upload it to S3.
5089 1 : async fn bootstrap_timeline(
5090 1 : self: &Arc<Self>,
5091 1 : timeline_id: TimelineId,
5092 1 : pg_version: u32,
5093 1 : load_existing_initdb: Option<TimelineId>,
5094 1 : ctx: &RequestContext,
5095 1 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
5096 1 : let timeline_create_guard = match self
5097 1 : .start_creating_timeline(
5098 1 : timeline_id,
5099 1 : CreateTimelineIdempotency::Bootstrap { pg_version },
5100 1 : )
5101 1 : .await?
5102 : {
5103 1 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
5104 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
5105 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
5106 : }
5107 : };
5108 :
5109 : // create a `tenant/{tenant_id}/timelines/basebackup-{timeline_id}.{TEMP_FILE_SUFFIX}/`
5110 : // temporary directory for basebackup files for the given timeline.
5111 :
5112 1 : let timelines_path = self.conf.timelines_path(&self.tenant_shard_id);
5113 1 : let pgdata_path = path_with_suffix_extension(
5114 1 : timelines_path.join(format!("basebackup-{timeline_id}")),
5115 1 : TEMP_FILE_SUFFIX,
5116 1 : );
5117 1 :
5118 1 : // Remove whatever was left from the previous runs: safe because TimelineCreateGuard guarantees
5119 1 : // we won't race with other creations or existent timelines with the same path.
5120 1 : if pgdata_path.exists() {
5121 0 : fs::remove_dir_all(&pgdata_path).with_context(|| {
5122 0 : format!("Failed to remove already existing initdb directory: {pgdata_path}")
5123 0 : })?;
5124 0 : tracing::info!("removed previous attempt's temporary initdb directory '{pgdata_path}'");
5125 1 : }
5126 :
5127 : // this new directory is very temporary, set to remove it immediately after bootstrap, we don't need it
5128 1 : let pgdata_path_deferred = pgdata_path.clone();
5129 1 : scopeguard::defer! {
5130 1 : if let Err(e) = fs::remove_dir_all(&pgdata_path_deferred).or_else(fs_ext::ignore_not_found) {
5131 1 : // this is unlikely, but we will remove the directory on pageserver restart or another bootstrap call
5132 1 : error!("Failed to remove temporary initdb directory '{pgdata_path_deferred}': {e}");
5133 1 : } else {
5134 1 : tracing::info!("removed temporary initdb directory '{pgdata_path_deferred}'");
5135 1 : }
5136 1 : }
5137 1 : if let Some(existing_initdb_timeline_id) = load_existing_initdb {
5138 1 : if existing_initdb_timeline_id != timeline_id {
5139 0 : let source_path = &remote_initdb_archive_path(
5140 0 : &self.tenant_shard_id.tenant_id,
5141 0 : &existing_initdb_timeline_id,
5142 0 : );
5143 0 : let dest_path =
5144 0 : &remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &timeline_id);
5145 0 :
5146 0 : // if this fails, it will get retried by retried control plane requests
5147 0 : self.remote_storage
5148 0 : .copy_object(source_path, dest_path, &self.cancel)
5149 0 : .await
5150 0 : .context("copy initdb tar")?;
5151 1 : }
5152 1 : let (initdb_tar_zst_path, initdb_tar_zst) =
5153 1 : self::remote_timeline_client::download_initdb_tar_zst(
5154 1 : self.conf,
5155 1 : &self.remote_storage,
5156 1 : &self.tenant_shard_id,
5157 1 : &existing_initdb_timeline_id,
5158 1 : &self.cancel,
5159 1 : )
5160 1 : .await
5161 1 : .context("download initdb tar")?;
5162 :
5163 1 : scopeguard::defer! {
5164 1 : if let Err(e) = fs::remove_file(&initdb_tar_zst_path) {
5165 1 : error!("Failed to remove temporary initdb archive '{initdb_tar_zst_path}': {e}");
5166 1 : }
5167 1 : }
5168 1 :
5169 1 : let buf_read =
5170 1 : BufReader::with_capacity(remote_timeline_client::BUFFER_SIZE, initdb_tar_zst);
5171 1 : extract_zst_tarball(&pgdata_path, buf_read)
5172 1 : .await
5173 1 : .context("extract initdb tar")?;
5174 : } else {
5175 : // Init temporarily repo to get bootstrap data, this creates a directory in the `pgdata_path` path
5176 0 : run_initdb(self.conf, &pgdata_path, pg_version, &self.cancel)
5177 0 : .await
5178 0 : .context("run initdb")?;
5179 :
5180 : // Upload the created data dir to S3
5181 0 : if self.tenant_shard_id().is_shard_zero() {
5182 0 : self.upload_initdb(&timelines_path, &pgdata_path, &timeline_id)
5183 0 : .await?;
5184 0 : }
5185 : }
5186 1 : let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
5187 1 :
5188 1 : // Import the contents of the data directory at the initial checkpoint
5189 1 : // LSN, and any WAL after that.
5190 1 : // Initdb lsn will be equal to last_record_lsn which will be set after import.
5191 1 : // Because we know it upfront avoid having an option or dummy zero value by passing it to the metadata.
5192 1 : let new_metadata = TimelineMetadata::new(
5193 1 : Lsn(0),
5194 1 : None,
5195 1 : None,
5196 1 : Lsn(0),
5197 1 : pgdata_lsn,
5198 1 : pgdata_lsn,
5199 1 : pg_version,
5200 1 : );
5201 1 : let (mut raw_timeline, timeline_ctx) = self
5202 1 : .prepare_new_timeline(
5203 1 : timeline_id,
5204 1 : &new_metadata,
5205 1 : timeline_create_guard,
5206 1 : pgdata_lsn,
5207 1 : None,
5208 1 : None,
5209 1 : ctx,
5210 1 : )
5211 1 : .await?;
5212 :
5213 1 : let tenant_shard_id = raw_timeline.owning_tenant.tenant_shard_id;
5214 1 : raw_timeline
5215 1 : .write(|unfinished_timeline| async move {
5216 1 : import_datadir::import_timeline_from_postgres_datadir(
5217 1 : &unfinished_timeline,
5218 1 : &pgdata_path,
5219 1 : pgdata_lsn,
5220 1 : &timeline_ctx,
5221 1 : )
5222 1 : .await
5223 1 : .with_context(|| {
5224 0 : format!(
5225 0 : "Failed to import pgdatadir for timeline {tenant_shard_id}/{timeline_id}"
5226 0 : )
5227 1 : })?;
5228 :
5229 1 : fail::fail_point!("before-checkpoint-new-timeline", |_| {
5230 0 : Err(CreateTimelineError::Other(anyhow::anyhow!(
5231 0 : "failpoint before-checkpoint-new-timeline"
5232 0 : )))
5233 1 : });
5234 :
5235 1 : Ok(())
5236 2 : })
5237 1 : .await?;
5238 :
5239 : // All done!
5240 1 : let timeline = raw_timeline.finish_creation().await?;
5241 :
5242 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
5243 :
5244 1 : Ok(CreateTimelineResult::Created(timeline))
5245 1 : }
5246 :
5247 230 : fn build_timeline_remote_client(&self, timeline_id: TimelineId) -> RemoteTimelineClient {
5248 230 : RemoteTimelineClient::new(
5249 230 : self.remote_storage.clone(),
5250 230 : self.deletion_queue_client.clone(),
5251 230 : self.conf,
5252 230 : self.tenant_shard_id,
5253 230 : timeline_id,
5254 230 : self.generation,
5255 230 : &self.tenant_conf.load().location,
5256 230 : )
5257 230 : }
5258 :
5259 : /// Builds required resources for a new timeline.
5260 230 : fn build_timeline_resources(&self, timeline_id: TimelineId) -> TimelineResources {
5261 230 : let remote_client = self.build_timeline_remote_client(timeline_id);
5262 230 : self.get_timeline_resources_for(remote_client)
5263 230 : }
5264 :
5265 : /// Builds timeline resources for the given remote client.
5266 233 : fn get_timeline_resources_for(&self, remote_client: RemoteTimelineClient) -> TimelineResources {
5267 233 : TimelineResources {
5268 233 : remote_client,
5269 233 : pagestream_throttle: self.pagestream_throttle.clone(),
5270 233 : pagestream_throttle_metrics: self.pagestream_throttle_metrics.clone(),
5271 233 : l0_compaction_trigger: self.l0_compaction_trigger.clone(),
5272 233 : l0_flush_global_state: self.l0_flush_global_state.clone(),
5273 233 : basebackup_prepare_sender: self.basebackup_prepare_sender.clone(),
5274 233 : }
5275 233 : }
5276 :
5277 : /// Creates intermediate timeline structure and its files.
5278 : ///
5279 : /// An empty layer map is initialized, and new data and WAL can be imported starting
5280 : /// at 'disk_consistent_lsn'. After any initial data has been imported, call
5281 : /// `finish_creation` to insert the Timeline into the timelines map.
5282 : #[allow(clippy::too_many_arguments)]
5283 230 : async fn prepare_new_timeline<'a>(
5284 230 : &'a self,
5285 230 : new_timeline_id: TimelineId,
5286 230 : new_metadata: &TimelineMetadata,
5287 230 : create_guard: TimelineCreateGuard,
5288 230 : start_lsn: Lsn,
5289 230 : ancestor: Option<Arc<Timeline>>,
5290 230 : rel_size_v2_status: Option<RelSizeMigration>,
5291 230 : ctx: &RequestContext,
5292 230 : ) -> anyhow::Result<(UninitializedTimeline<'a>, RequestContext)> {
5293 230 : let tenant_shard_id = self.tenant_shard_id;
5294 230 :
5295 230 : let resources = self.build_timeline_resources(new_timeline_id);
5296 230 : resources
5297 230 : .remote_client
5298 230 : .init_upload_queue_for_empty_remote(new_metadata, rel_size_v2_status.clone())?;
5299 :
5300 230 : let (timeline_struct, timeline_ctx) = self
5301 230 : .create_timeline_struct(
5302 230 : new_timeline_id,
5303 230 : new_metadata,
5304 230 : None,
5305 230 : ancestor,
5306 230 : resources,
5307 230 : CreateTimelineCause::Load,
5308 230 : create_guard.idempotency.clone(),
5309 230 : None,
5310 230 : rel_size_v2_status,
5311 230 : ctx,
5312 230 : )
5313 230 : .context("Failed to create timeline data structure")?;
5314 :
5315 230 : timeline_struct.init_empty_layer_map(start_lsn);
5316 :
5317 230 : if let Err(e) = self
5318 230 : .create_timeline_files(&create_guard.timeline_path)
5319 230 : .await
5320 : {
5321 0 : error!(
5322 0 : "Failed to create initial files for timeline {tenant_shard_id}/{new_timeline_id}, cleaning up: {e:?}"
5323 : );
5324 0 : cleanup_timeline_directory(create_guard);
5325 0 : return Err(e);
5326 230 : }
5327 230 :
5328 230 : debug!(
5329 0 : "Successfully created initial files for timeline {tenant_shard_id}/{new_timeline_id}"
5330 : );
5331 :
5332 230 : Ok((
5333 230 : UninitializedTimeline::new(
5334 230 : self,
5335 230 : new_timeline_id,
5336 230 : Some((timeline_struct, create_guard)),
5337 230 : ),
5338 230 : timeline_ctx,
5339 230 : ))
5340 230 : }
5341 :
5342 230 : async fn create_timeline_files(&self, timeline_path: &Utf8Path) -> anyhow::Result<()> {
5343 230 : crashsafe::create_dir(timeline_path).context("Failed to create timeline directory")?;
5344 :
5345 230 : fail::fail_point!("after-timeline-dir-creation", |_| {
5346 0 : anyhow::bail!("failpoint after-timeline-dir-creation");
5347 230 : });
5348 :
5349 230 : Ok(())
5350 230 : }
5351 :
5352 : /// Get a guard that provides exclusive access to the timeline directory, preventing
5353 : /// concurrent attempts to create the same timeline.
5354 : ///
5355 : /// The `allow_offloaded` parameter controls whether to tolerate the existence of
5356 : /// offloaded timelines or not.
5357 233 : fn create_timeline_create_guard(
5358 233 : self: &Arc<Self>,
5359 233 : timeline_id: TimelineId,
5360 233 : idempotency: CreateTimelineIdempotency,
5361 233 : allow_offloaded: bool,
5362 233 : ) -> Result<TimelineCreateGuard, TimelineExclusionError> {
5363 233 : let tenant_shard_id = self.tenant_shard_id;
5364 233 :
5365 233 : let timeline_path = self.conf.timeline_path(&tenant_shard_id, &timeline_id);
5366 :
5367 233 : let create_guard = TimelineCreateGuard::new(
5368 233 : self,
5369 233 : timeline_id,
5370 233 : timeline_path.clone(),
5371 233 : idempotency,
5372 233 : allow_offloaded,
5373 233 : )?;
5374 :
5375 : // At this stage, we have got exclusive access to in-memory state for this timeline ID
5376 : // for creation.
5377 : // A timeline directory should never exist on disk already:
5378 : // - a previous failed creation would have cleaned up after itself
5379 : // - a pageserver restart would clean up timeline directories that don't have valid remote state
5380 : //
5381 : // Therefore it is an unexpected internal error to encounter a timeline directory already existing here,
5382 : // this error may indicate a bug in cleanup on failed creations.
5383 232 : if timeline_path.exists() {
5384 0 : return Err(TimelineExclusionError::Other(anyhow::anyhow!(
5385 0 : "Timeline directory already exists! This is a bug."
5386 0 : )));
5387 232 : }
5388 232 :
5389 232 : Ok(create_guard)
5390 233 : }
5391 :
5392 : /// Gathers inputs from all of the timelines to produce a sizing model input.
5393 : ///
5394 : /// Future is cancellation safe. Only one calculation can be running at once per tenant.
5395 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5396 : pub async fn gather_size_inputs(
5397 : &self,
5398 : // `max_retention_period` overrides the cutoff that is used to calculate the size
5399 : // (only if it is shorter than the real cutoff).
5400 : max_retention_period: Option<u64>,
5401 : cause: LogicalSizeCalculationCause,
5402 : cancel: &CancellationToken,
5403 : ctx: &RequestContext,
5404 : ) -> Result<size::ModelInputs, size::CalculateSyntheticSizeError> {
5405 : let logical_sizes_at_once = self
5406 : .conf
5407 : .concurrent_tenant_size_logical_size_queries
5408 : .inner();
5409 :
5410 : // TODO: Having a single mutex block concurrent reads is not great for performance.
5411 : //
5412 : // But the only case where we need to run multiple of these at once is when we
5413 : // request a size for a tenant manually via API, while another background calculation
5414 : // is in progress (which is not a common case).
5415 : //
5416 : // See more for on the issue #2748 condenced out of the initial PR review.
5417 : let mut shared_cache = tokio::select! {
5418 : locked = self.cached_logical_sizes.lock() => locked,
5419 : _ = cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5420 : _ = self.cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5421 : };
5422 :
5423 : size::gather_inputs(
5424 : self,
5425 : logical_sizes_at_once,
5426 : max_retention_period,
5427 : &mut shared_cache,
5428 : cause,
5429 : cancel,
5430 : ctx,
5431 : )
5432 : .await
5433 : }
5434 :
5435 : /// Calculate synthetic tenant size and cache the result.
5436 : /// This is periodically called by background worker.
5437 : /// result is cached in tenant struct
5438 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5439 : pub async fn calculate_synthetic_size(
5440 : &self,
5441 : cause: LogicalSizeCalculationCause,
5442 : cancel: &CancellationToken,
5443 : ctx: &RequestContext,
5444 : ) -> Result<u64, size::CalculateSyntheticSizeError> {
5445 : let inputs = self.gather_size_inputs(None, cause, cancel, ctx).await?;
5446 :
5447 : let size = inputs.calculate();
5448 :
5449 : self.set_cached_synthetic_size(size);
5450 :
5451 : Ok(size)
5452 : }
5453 :
5454 : /// Cache given synthetic size and update the metric value
5455 0 : pub fn set_cached_synthetic_size(&self, size: u64) {
5456 0 : self.cached_synthetic_tenant_size
5457 0 : .store(size, Ordering::Relaxed);
5458 0 :
5459 0 : // Only shard zero should be calculating synthetic sizes
5460 0 : debug_assert!(self.shard_identity.is_shard_zero());
5461 :
5462 0 : TENANT_SYNTHETIC_SIZE_METRIC
5463 0 : .get_metric_with_label_values(&[&self.tenant_shard_id.tenant_id.to_string()])
5464 0 : .unwrap()
5465 0 : .set(size);
5466 0 : }
5467 :
5468 0 : pub fn cached_synthetic_size(&self) -> u64 {
5469 0 : self.cached_synthetic_tenant_size.load(Ordering::Relaxed)
5470 0 : }
5471 :
5472 : /// Flush any in-progress layers, schedule uploads, and wait for uploads to complete.
5473 : ///
5474 : /// This function can take a long time: callers should wrap it in a timeout if calling
5475 : /// from an external API handler.
5476 : ///
5477 : /// Cancel-safety: cancelling this function may leave I/O running, but such I/O is
5478 : /// still bounded by tenant/timeline shutdown.
5479 : #[tracing::instrument(skip_all)]
5480 : pub(crate) async fn flush_remote(&self) -> anyhow::Result<()> {
5481 : let timelines = self.timelines.lock().unwrap().clone();
5482 :
5483 0 : async fn flush_timeline(_gate: GateGuard, timeline: Arc<Timeline>) -> anyhow::Result<()> {
5484 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Flushing...");
5485 0 : timeline.freeze_and_flush().await?;
5486 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Waiting for uploads...");
5487 0 : timeline.remote_client.wait_completion().await?;
5488 :
5489 0 : Ok(())
5490 0 : }
5491 :
5492 : // We do not use a JoinSet for these tasks, because we don't want them to be
5493 : // aborted when this function's future is cancelled: they should stay alive
5494 : // holding their GateGuard until they complete, to ensure their I/Os complete
5495 : // before Timeline shutdown completes.
5496 : let mut results = FuturesUnordered::new();
5497 :
5498 : for (_timeline_id, timeline) in timelines {
5499 : // Run each timeline's flush in a task holding the timeline's gate: this
5500 : // means that if this function's future is cancelled, the Timeline shutdown
5501 : // will still wait for any I/O in here to complete.
5502 : let Ok(gate) = timeline.gate.enter() else {
5503 : continue;
5504 : };
5505 0 : let jh = tokio::task::spawn(async move { flush_timeline(gate, timeline).await });
5506 : results.push(jh);
5507 : }
5508 :
5509 : while let Some(r) = results.next().await {
5510 : if let Err(e) = r {
5511 : if !e.is_cancelled() && !e.is_panic() {
5512 : tracing::error!("unexpected join error: {e:?}");
5513 : }
5514 : }
5515 : }
5516 :
5517 : // The flushes we did above were just writes, but the TenantShard might have had
5518 : // pending deletions as well from recent compaction/gc: we want to flush those
5519 : // as well. This requires flushing the global delete queue. This is cheap
5520 : // because it's typically a no-op.
5521 : match self.deletion_queue_client.flush_execute().await {
5522 : Ok(_) => {}
5523 : Err(DeletionQueueError::ShuttingDown) => {}
5524 : }
5525 :
5526 : Ok(())
5527 : }
5528 :
5529 0 : pub(crate) fn get_tenant_conf(&self) -> pageserver_api::models::TenantConfig {
5530 0 : self.tenant_conf.load().tenant_conf.clone()
5531 0 : }
5532 :
5533 : /// How much local storage would this tenant like to have? It can cope with
5534 : /// less than this (via eviction and on-demand downloads), but this function enables
5535 : /// the TenantShard to advertise how much storage it would prefer to have to provide fast I/O
5536 : /// by keeping important things on local disk.
5537 : ///
5538 : /// This is a heuristic, not a guarantee: tenants that are long-idle will actually use less
5539 : /// than they report here, due to layer eviction. Tenants with many active branches may
5540 : /// actually use more than they report here.
5541 0 : pub(crate) fn local_storage_wanted(&self) -> u64 {
5542 0 : let timelines = self.timelines.lock().unwrap();
5543 0 :
5544 0 : // Heuristic: we use the max() of the timelines' visible sizes, rather than the sum. This
5545 0 : // reflects the observation that on tenants with multiple large branches, typically only one
5546 0 : // of them is used actively enough to occupy space on disk.
5547 0 : timelines
5548 0 : .values()
5549 0 : .map(|t| t.metrics.visible_physical_size_gauge.get())
5550 0 : .max()
5551 0 : .unwrap_or(0)
5552 0 : }
5553 :
5554 : /// Builds a new tenant manifest, and uploads it if it differs from the last-known tenant
5555 : /// manifest in `Self::remote_tenant_manifest`.
5556 : ///
5557 : /// TODO: instead of requiring callers to remember to call `maybe_upload_tenant_manifest` after
5558 : /// changing any `TenantShard` state that's included in the manifest, consider making the manifest
5559 : /// the authoritative source of data with an API that automatically uploads on changes. Revisit
5560 : /// this when the manifest is more widely used and we have a better idea of the data model.
5561 118 : pub(crate) async fn maybe_upload_tenant_manifest(&self) -> Result<(), TenantManifestError> {
5562 : // Multiple tasks may call this function concurrently after mutating the TenantShard runtime
5563 : // state, affecting the manifest generated by `build_tenant_manifest`. We use an async mutex
5564 : // to serialize these callers. `eq_ignoring_version` acts as a slightly inefficient but
5565 : // simple coalescing mechanism.
5566 118 : let mut guard = tokio::select! {
5567 118 : guard = self.remote_tenant_manifest.lock() => guard,
5568 118 : _ = self.cancel.cancelled() => return Err(TenantManifestError::Cancelled),
5569 : };
5570 :
5571 : // Build a new manifest.
5572 118 : let manifest = self.build_tenant_manifest();
5573 :
5574 : // Check if the manifest has changed. We ignore the version number here, to avoid
5575 : // uploading every manifest on version number bumps.
5576 118 : if let Some(old) = guard.as_ref() {
5577 4 : if manifest.eq_ignoring_version(old) {
5578 3 : return Ok(());
5579 1 : }
5580 114 : }
5581 :
5582 : // Update metrics
5583 115 : let tid = self.tenant_shard_id.to_string();
5584 115 : let shard_id = self.tenant_shard_id.shard_slug().to_string();
5585 115 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
5586 115 : TENANT_OFFLOADED_TIMELINES
5587 115 : .with_label_values(set_key)
5588 115 : .set(manifest.offloaded_timelines.len() as u64);
5589 115 :
5590 115 : // Upload the manifest. Remote storage does no retries internally, so retry here.
5591 115 : match backoff::retry(
5592 115 : || async {
5593 115 : upload_tenant_manifest(
5594 115 : &self.remote_storage,
5595 115 : &self.tenant_shard_id,
5596 115 : self.generation,
5597 115 : &manifest,
5598 115 : &self.cancel,
5599 115 : )
5600 115 : .await
5601 230 : },
5602 115 : |_| self.cancel.is_cancelled(),
5603 115 : FAILED_UPLOAD_WARN_THRESHOLD,
5604 115 : FAILED_REMOTE_OP_RETRIES,
5605 115 : "uploading tenant manifest",
5606 115 : &self.cancel,
5607 115 : )
5608 115 : .await
5609 : {
5610 0 : None => Err(TenantManifestError::Cancelled),
5611 0 : Some(Err(_)) if self.cancel.is_cancelled() => Err(TenantManifestError::Cancelled),
5612 0 : Some(Err(e)) => Err(TenantManifestError::RemoteStorage(e)),
5613 : Some(Ok(_)) => {
5614 : // Store the successfully uploaded manifest, so that future callers can avoid
5615 : // re-uploading the same thing.
5616 115 : *guard = Some(manifest);
5617 115 :
5618 115 : Ok(())
5619 : }
5620 : }
5621 118 : }
5622 : }
5623 :
5624 : /// Create the cluster temporarily in 'initdbpath' directory inside the repository
5625 : /// to get bootstrap data for timeline initialization.
5626 0 : async fn run_initdb(
5627 0 : conf: &'static PageServerConf,
5628 0 : initdb_target_dir: &Utf8Path,
5629 0 : pg_version: u32,
5630 0 : cancel: &CancellationToken,
5631 0 : ) -> Result<(), InitdbError> {
5632 0 : let initdb_bin_path = conf
5633 0 : .pg_bin_dir(pg_version)
5634 0 : .map_err(InitdbError::Other)?
5635 0 : .join("initdb");
5636 0 : let initdb_lib_dir = conf.pg_lib_dir(pg_version).map_err(InitdbError::Other)?;
5637 0 : info!(
5638 0 : "running {} in {}, libdir: {}",
5639 : initdb_bin_path, initdb_target_dir, initdb_lib_dir,
5640 : );
5641 :
5642 0 : let _permit = {
5643 0 : let _timer = INITDB_SEMAPHORE_ACQUISITION_TIME.start_timer();
5644 0 : INIT_DB_SEMAPHORE.acquire().await
5645 : };
5646 :
5647 0 : CONCURRENT_INITDBS.inc();
5648 0 : scopeguard::defer! {
5649 0 : CONCURRENT_INITDBS.dec();
5650 0 : }
5651 0 :
5652 0 : let _timer = INITDB_RUN_TIME.start_timer();
5653 0 : let res = postgres_initdb::do_run_initdb(postgres_initdb::RunInitdbArgs {
5654 0 : superuser: &conf.superuser,
5655 0 : locale: &conf.locale,
5656 0 : initdb_bin: &initdb_bin_path,
5657 0 : pg_version,
5658 0 : library_search_path: &initdb_lib_dir,
5659 0 : pgdata: initdb_target_dir,
5660 0 : })
5661 0 : .await
5662 0 : .map_err(InitdbError::Inner);
5663 0 :
5664 0 : // This isn't true cancellation support, see above. Still return an error to
5665 0 : // excercise the cancellation code path.
5666 0 : if cancel.is_cancelled() {
5667 0 : return Err(InitdbError::Cancelled);
5668 0 : }
5669 0 :
5670 0 : res
5671 0 : }
5672 :
5673 : /// Dump contents of a layer file to stdout.
5674 0 : pub async fn dump_layerfile_from_path(
5675 0 : path: &Utf8Path,
5676 0 : verbose: bool,
5677 0 : ctx: &RequestContext,
5678 0 : ) -> anyhow::Result<()> {
5679 : use std::os::unix::fs::FileExt;
5680 :
5681 : // All layer files start with a two-byte "magic" value, to identify the kind of
5682 : // file.
5683 0 : let file = File::open(path)?;
5684 0 : let mut header_buf = [0u8; 2];
5685 0 : file.read_exact_at(&mut header_buf, 0)?;
5686 :
5687 0 : match u16::from_be_bytes(header_buf) {
5688 : crate::IMAGE_FILE_MAGIC => {
5689 0 : ImageLayer::new_for_path(path, file)?
5690 0 : .dump(verbose, ctx)
5691 0 : .await?
5692 : }
5693 : crate::DELTA_FILE_MAGIC => {
5694 0 : DeltaLayer::new_for_path(path, file)?
5695 0 : .dump(verbose, ctx)
5696 0 : .await?
5697 : }
5698 0 : magic => bail!("unrecognized magic identifier: {:?}", magic),
5699 : }
5700 :
5701 0 : Ok(())
5702 0 : }
5703 :
5704 : #[cfg(test)]
5705 : pub(crate) mod harness {
5706 : use bytes::{Bytes, BytesMut};
5707 : use hex_literal::hex;
5708 : use once_cell::sync::OnceCell;
5709 : use pageserver_api::key::Key;
5710 : use pageserver_api::models::ShardParameters;
5711 : use pageserver_api::record::NeonWalRecord;
5712 : use pageserver_api::shard::ShardIndex;
5713 : use utils::id::TenantId;
5714 : use utils::logging;
5715 :
5716 : use super::*;
5717 : use crate::deletion_queue::mock::MockDeletionQueue;
5718 : use crate::l0_flush::L0FlushConfig;
5719 : use crate::walredo::apply_neon;
5720 :
5721 : pub const TIMELINE_ID: TimelineId =
5722 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
5723 : pub const NEW_TIMELINE_ID: TimelineId =
5724 : TimelineId::from_array(hex!("AA223344556677881122334455667788"));
5725 :
5726 : /// Convenience function to create a page image with given string as the only content
5727 2514391 : pub fn test_img(s: &str) -> Bytes {
5728 2514391 : let mut buf = BytesMut::new();
5729 2514391 : buf.extend_from_slice(s.as_bytes());
5730 2514391 : buf.resize(64, 0);
5731 2514391 :
5732 2514391 : buf.freeze()
5733 2514391 : }
5734 :
5735 : pub struct TenantHarness {
5736 : pub conf: &'static PageServerConf,
5737 : pub tenant_conf: pageserver_api::models::TenantConfig,
5738 : pub tenant_shard_id: TenantShardId,
5739 : pub generation: Generation,
5740 : pub shard: ShardIndex,
5741 : pub remote_storage: GenericRemoteStorage,
5742 : pub remote_fs_dir: Utf8PathBuf,
5743 : pub deletion_queue: MockDeletionQueue,
5744 : }
5745 :
5746 : static LOG_HANDLE: OnceCell<()> = OnceCell::new();
5747 :
5748 129 : pub(crate) fn setup_logging() {
5749 129 : LOG_HANDLE.get_or_init(|| {
5750 123 : logging::init(
5751 123 : logging::LogFormat::Test,
5752 123 : // enable it in case the tests exercise code paths that use
5753 123 : // debug_assert_current_span_has_tenant_and_timeline_id
5754 123 : logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
5755 123 : logging::Output::Stdout,
5756 123 : )
5757 123 : .expect("Failed to init test logging");
5758 129 : });
5759 129 : }
5760 :
5761 : impl TenantHarness {
5762 117 : pub async fn create_custom(
5763 117 : test_name: &'static str,
5764 117 : tenant_conf: pageserver_api::models::TenantConfig,
5765 117 : tenant_id: TenantId,
5766 117 : shard_identity: ShardIdentity,
5767 117 : generation: Generation,
5768 117 : ) -> anyhow::Result<Self> {
5769 117 : setup_logging();
5770 117 :
5771 117 : let repo_dir = PageServerConf::test_repo_dir(test_name);
5772 117 : let _ = fs::remove_dir_all(&repo_dir);
5773 117 : fs::create_dir_all(&repo_dir)?;
5774 :
5775 117 : let conf = PageServerConf::dummy_conf(repo_dir);
5776 117 : // Make a static copy of the config. This can never be free'd, but that's
5777 117 : // OK in a test.
5778 117 : let conf: &'static PageServerConf = Box::leak(Box::new(conf));
5779 117 :
5780 117 : let shard = shard_identity.shard_index();
5781 117 : let tenant_shard_id = TenantShardId {
5782 117 : tenant_id,
5783 117 : shard_number: shard.shard_number,
5784 117 : shard_count: shard.shard_count,
5785 117 : };
5786 117 : fs::create_dir_all(conf.tenant_path(&tenant_shard_id))?;
5787 117 : fs::create_dir_all(conf.timelines_path(&tenant_shard_id))?;
5788 :
5789 : use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
5790 117 : let remote_fs_dir = conf.workdir.join("localfs");
5791 117 : std::fs::create_dir_all(&remote_fs_dir).unwrap();
5792 117 : let config = RemoteStorageConfig {
5793 117 : storage: RemoteStorageKind::LocalFs {
5794 117 : local_path: remote_fs_dir.clone(),
5795 117 : },
5796 117 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
5797 117 : small_timeout: RemoteStorageConfig::DEFAULT_SMALL_TIMEOUT,
5798 117 : };
5799 117 : let remote_storage = GenericRemoteStorage::from_config(&config).await.unwrap();
5800 117 : let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()));
5801 117 :
5802 117 : Ok(Self {
5803 117 : conf,
5804 117 : tenant_conf,
5805 117 : tenant_shard_id,
5806 117 : generation,
5807 117 : shard,
5808 117 : remote_storage,
5809 117 : remote_fs_dir,
5810 117 : deletion_queue,
5811 117 : })
5812 117 : }
5813 :
5814 110 : pub async fn create(test_name: &'static str) -> anyhow::Result<Self> {
5815 110 : // Disable automatic GC and compaction to make the unit tests more deterministic.
5816 110 : // The tests perform them manually if needed.
5817 110 : let tenant_conf = pageserver_api::models::TenantConfig {
5818 110 : gc_period: Some(Duration::ZERO),
5819 110 : compaction_period: Some(Duration::ZERO),
5820 110 : ..Default::default()
5821 110 : };
5822 110 : let tenant_id = TenantId::generate();
5823 110 : let shard = ShardIdentity::unsharded();
5824 110 : Self::create_custom(
5825 110 : test_name,
5826 110 : tenant_conf,
5827 110 : tenant_id,
5828 110 : shard,
5829 110 : Generation::new(0xdeadbeef),
5830 110 : )
5831 110 : .await
5832 110 : }
5833 :
5834 10 : pub fn span(&self) -> tracing::Span {
5835 10 : info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
5836 10 : }
5837 :
5838 117 : pub(crate) async fn load(&self) -> (Arc<TenantShard>, RequestContext) {
5839 117 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error)
5840 117 : .with_scope_unit_test();
5841 117 : (
5842 117 : self.do_try_load(&ctx)
5843 117 : .await
5844 117 : .expect("failed to load test tenant"),
5845 117 : ctx,
5846 117 : )
5847 117 : }
5848 :
5849 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5850 : pub(crate) async fn do_try_load(
5851 : &self,
5852 : ctx: &RequestContext,
5853 : ) -> anyhow::Result<Arc<TenantShard>> {
5854 : let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
5855 :
5856 : let (basebackup_requst_sender, _) = tokio::sync::mpsc::unbounded_channel();
5857 :
5858 : let tenant = Arc::new(TenantShard::new(
5859 : TenantState::Attaching,
5860 : self.conf,
5861 : AttachedTenantConf::try_from(LocationConf::attached_single(
5862 : self.tenant_conf.clone(),
5863 : self.generation,
5864 : &ShardParameters::default(),
5865 : ))
5866 : .unwrap(),
5867 : // This is a legacy/test code path: sharding isn't supported here.
5868 : ShardIdentity::unsharded(),
5869 : Some(walredo_mgr),
5870 : self.tenant_shard_id,
5871 : self.remote_storage.clone(),
5872 : self.deletion_queue.new_client(),
5873 : // TODO: ideally we should run all unit tests with both configs
5874 : L0FlushGlobalState::new(L0FlushConfig::default()),
5875 : basebackup_requst_sender,
5876 : ));
5877 :
5878 : let preload = tenant
5879 : .preload(&self.remote_storage, CancellationToken::new())
5880 : .await?;
5881 : tenant.attach(Some(preload), ctx).await?;
5882 :
5883 : tenant.state.send_replace(TenantState::Active);
5884 : for timeline in tenant.timelines.lock().unwrap().values() {
5885 : timeline.set_state(TimelineState::Active);
5886 : }
5887 : Ok(tenant)
5888 : }
5889 :
5890 1 : pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
5891 1 : self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
5892 1 : }
5893 : }
5894 :
5895 : // Mock WAL redo manager that doesn't do much
5896 : pub(crate) struct TestRedoManager;
5897 :
5898 : impl TestRedoManager {
5899 : /// # Cancel-Safety
5900 : ///
5901 : /// This method is cancellation-safe.
5902 26774 : pub async fn request_redo(
5903 26774 : &self,
5904 26774 : key: Key,
5905 26774 : lsn: Lsn,
5906 26774 : base_img: Option<(Lsn, Bytes)>,
5907 26774 : records: Vec<(Lsn, NeonWalRecord)>,
5908 26774 : _pg_version: u32,
5909 26774 : _redo_attempt_type: RedoAttemptType,
5910 26774 : ) -> Result<Bytes, walredo::Error> {
5911 1403510 : let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
5912 26774 : if records_neon {
5913 : // For Neon wal records, we can decode without spawning postgres, so do so.
5914 26774 : let mut page = match (base_img, records.first()) {
5915 13029 : (Some((_lsn, img)), _) => {
5916 13029 : let mut page = BytesMut::new();
5917 13029 : page.extend_from_slice(&img);
5918 13029 : page
5919 : }
5920 13745 : (_, Some((_lsn, rec))) if rec.will_init() => BytesMut::new(),
5921 : _ => {
5922 0 : panic!("Neon WAL redo requires base image or will init record");
5923 : }
5924 : };
5925 :
5926 1430283 : for (record_lsn, record) in records {
5927 1403510 : apply_neon::apply_in_neon(&record, record_lsn, key, &mut page)?;
5928 : }
5929 26773 : Ok(page.freeze())
5930 : } else {
5931 : // We never spawn a postgres walredo process in unit tests: just log what we might have done.
5932 0 : let s = format!(
5933 0 : "redo for {} to get to {}, with {} and {} records",
5934 0 : key,
5935 0 : lsn,
5936 0 : if base_img.is_some() {
5937 0 : "base image"
5938 : } else {
5939 0 : "no base image"
5940 : },
5941 0 : records.len()
5942 0 : );
5943 0 : println!("{s}");
5944 0 :
5945 0 : Ok(test_img(&s))
5946 : }
5947 26774 : }
5948 : }
5949 : }
5950 :
5951 : #[cfg(test)]
5952 : mod tests {
5953 : use std::collections::{BTreeMap, BTreeSet};
5954 :
5955 : use bytes::{Bytes, BytesMut};
5956 : use hex_literal::hex;
5957 : use itertools::Itertools;
5958 : #[cfg(feature = "testing")]
5959 : use models::CompactLsnRange;
5960 : use pageserver_api::key::{
5961 : AUX_KEY_PREFIX, Key, NON_INHERITED_RANGE, RELATION_SIZE_PREFIX, repl_origin_key,
5962 : };
5963 : use pageserver_api::keyspace::KeySpace;
5964 : #[cfg(feature = "testing")]
5965 : use pageserver_api::keyspace::KeySpaceRandomAccum;
5966 : use pageserver_api::models::{CompactionAlgorithm, CompactionAlgorithmSettings};
5967 : #[cfg(feature = "testing")]
5968 : use pageserver_api::record::NeonWalRecord;
5969 : use pageserver_api::value::Value;
5970 : use pageserver_compaction::helpers::overlaps_with;
5971 : #[cfg(feature = "testing")]
5972 : use rand::SeedableRng;
5973 : #[cfg(feature = "testing")]
5974 : use rand::rngs::StdRng;
5975 : use rand::{Rng, thread_rng};
5976 : #[cfg(feature = "testing")]
5977 : use std::ops::Range;
5978 : use storage_layer::{IoConcurrency, PersistentLayerKey};
5979 : use tests::storage_layer::ValuesReconstructState;
5980 : use tests::timeline::{GetVectoredError, ShutdownMode};
5981 : #[cfg(feature = "testing")]
5982 : use timeline::GcInfo;
5983 : #[cfg(feature = "testing")]
5984 : use timeline::InMemoryLayerTestDesc;
5985 : #[cfg(feature = "testing")]
5986 : use timeline::compaction::{KeyHistoryRetention, KeyLogAtLsn};
5987 : use timeline::{CompactOptions, DeltaLayerTestDesc, VersionedKeySpaceQuery};
5988 : use utils::id::TenantId;
5989 :
5990 : use super::*;
5991 : use crate::DEFAULT_PG_VERSION;
5992 : use crate::keyspace::KeySpaceAccum;
5993 : use crate::tenant::harness::*;
5994 : use crate::tenant::timeline::CompactFlags;
5995 :
5996 : static TEST_KEY: Lazy<Key> =
5997 9 : Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));
5998 :
5999 : #[cfg(feature = "testing")]
6000 : struct TestTimelineSpecification {
6001 : start_lsn: Lsn,
6002 : last_record_lsn: Lsn,
6003 :
6004 : in_memory_layers_shape: Vec<(Range<Key>, Range<Lsn>)>,
6005 : delta_layers_shape: Vec<(Range<Key>, Range<Lsn>)>,
6006 : image_layers_shape: Vec<(Range<Key>, Lsn)>,
6007 :
6008 : gap_chance: u8,
6009 : will_init_chance: u8,
6010 : }
6011 :
6012 : #[cfg(feature = "testing")]
6013 : struct Storage {
6014 : storage: HashMap<(Key, Lsn), Value>,
6015 : start_lsn: Lsn,
6016 : }
6017 :
6018 : #[cfg(feature = "testing")]
6019 : impl Storage {
6020 32000 : fn get(&self, key: Key, lsn: Lsn) -> Bytes {
6021 : use bytes::BufMut;
6022 :
6023 32000 : let mut crnt_lsn = lsn;
6024 32000 : let mut got_base = false;
6025 32000 :
6026 32000 : let mut acc = Vec::new();
6027 :
6028 2831871 : while crnt_lsn >= self.start_lsn {
6029 2831871 : if let Some(value) = self.storage.get(&(key, crnt_lsn)) {
6030 1421172 : acc.push(value.clone());
6031 :
6032 1402881 : match value {
6033 1402881 : Value::WalRecord(NeonWalRecord::Test { will_init, .. }) => {
6034 1402881 : if *will_init {
6035 13709 : got_base = true;
6036 13709 : break;
6037 1389172 : }
6038 : }
6039 : Value::Image(_) => {
6040 18291 : got_base = true;
6041 18291 : break;
6042 : }
6043 0 : _ => unreachable!(),
6044 : }
6045 1410699 : }
6046 :
6047 2799871 : crnt_lsn = crnt_lsn.checked_sub(1u64).unwrap();
6048 : }
6049 :
6050 32000 : assert!(
6051 32000 : got_base,
6052 0 : "Input data was incorrect. No base image for {key}@{lsn}"
6053 : );
6054 :
6055 32000 : tracing::debug!("Wal redo depth for {key}@{lsn} is {}", acc.len());
6056 :
6057 32000 : let mut blob = BytesMut::new();
6058 1421172 : for value in acc.into_iter().rev() {
6059 1402881 : match value {
6060 1402881 : Value::WalRecord(NeonWalRecord::Test { append, .. }) => {
6061 1402881 : blob.extend_from_slice(append.as_bytes());
6062 1402881 : }
6063 18291 : Value::Image(img) => {
6064 18291 : blob.put(img);
6065 18291 : }
6066 0 : _ => unreachable!(),
6067 : }
6068 : }
6069 :
6070 32000 : blob.into()
6071 32000 : }
6072 : }
6073 :
6074 : #[cfg(feature = "testing")]
6075 : #[allow(clippy::too_many_arguments)]
6076 1 : async fn randomize_timeline(
6077 1 : tenant: &Arc<TenantShard>,
6078 1 : new_timeline_id: TimelineId,
6079 1 : pg_version: u32,
6080 1 : spec: TestTimelineSpecification,
6081 1 : random: &mut rand::rngs::StdRng,
6082 1 : ctx: &RequestContext,
6083 1 : ) -> anyhow::Result<(Arc<Timeline>, Storage, Vec<Lsn>)> {
6084 1 : let mut storage: HashMap<(Key, Lsn), Value> = HashMap::default();
6085 1 : let mut interesting_lsns = vec![spec.last_record_lsn];
6086 :
6087 2 : for (key_range, lsn_range) in spec.in_memory_layers_shape.iter() {
6088 2 : let mut lsn = lsn_range.start;
6089 202 : while lsn < lsn_range.end {
6090 200 : let mut key = key_range.start;
6091 21018 : while key < key_range.end {
6092 20818 : let gap = random.gen_range(1..=100) <= spec.gap_chance;
6093 20818 : let will_init = random.gen_range(1..=100) <= spec.will_init_chance;
6094 20818 :
6095 20818 : if gap {
6096 1018 : continue;
6097 19800 : }
6098 :
6099 19800 : let record = if will_init {
6100 191 : Value::WalRecord(NeonWalRecord::wal_init(format!("[wil_init {key}@{lsn}]")))
6101 : } else {
6102 19609 : Value::WalRecord(NeonWalRecord::wal_append(format!("[delta {key}@{lsn}]")))
6103 : };
6104 :
6105 19800 : storage.insert((key, lsn), record);
6106 19800 :
6107 19800 : key = key.next();
6108 : }
6109 200 : lsn = Lsn(lsn.0 + 1);
6110 : }
6111 :
6112 : // Stash some interesting LSN for future use
6113 6 : for offset in [0, 5, 100].iter() {
6114 6 : if *offset == 0 {
6115 2 : interesting_lsns.push(lsn_range.start);
6116 2 : } else {
6117 4 : let below = lsn_range.start.checked_sub(*offset);
6118 4 : match below {
6119 4 : Some(v) if v >= spec.start_lsn => {
6120 4 : interesting_lsns.push(v);
6121 4 : }
6122 0 : _ => {}
6123 : }
6124 :
6125 4 : let above = Lsn(lsn_range.start.0 + offset);
6126 4 : interesting_lsns.push(above);
6127 : }
6128 : }
6129 : }
6130 :
6131 3 : for (key_range, lsn_range) in spec.delta_layers_shape.iter() {
6132 3 : let mut lsn = lsn_range.start;
6133 315 : while lsn < lsn_range.end {
6134 312 : let mut key = key_range.start;
6135 11112 : while key < key_range.end {
6136 10800 : let gap = random.gen_range(1..=100) <= spec.gap_chance;
6137 10800 : let will_init = random.gen_range(1..=100) <= spec.will_init_chance;
6138 10800 :
6139 10800 : if gap {
6140 504 : continue;
6141 10296 : }
6142 :
6143 10296 : let record = if will_init {
6144 103 : Value::WalRecord(NeonWalRecord::wal_init(format!("[wil_init {key}@{lsn}]")))
6145 : } else {
6146 10193 : Value::WalRecord(NeonWalRecord::wal_append(format!("[delta {key}@{lsn}]")))
6147 : };
6148 :
6149 10296 : storage.insert((key, lsn), record);
6150 10296 :
6151 10296 : key = key.next();
6152 : }
6153 312 : lsn = Lsn(lsn.0 + 1);
6154 : }
6155 :
6156 : // Stash some interesting LSN for future use
6157 9 : for offset in [0, 5, 100].iter() {
6158 9 : if *offset == 0 {
6159 3 : interesting_lsns.push(lsn_range.start);
6160 3 : } else {
6161 6 : let below = lsn_range.start.checked_sub(*offset);
6162 6 : match below {
6163 6 : Some(v) if v >= spec.start_lsn => {
6164 3 : interesting_lsns.push(v);
6165 3 : }
6166 3 : _ => {}
6167 : }
6168 :
6169 6 : let above = Lsn(lsn_range.start.0 + offset);
6170 6 : interesting_lsns.push(above);
6171 : }
6172 : }
6173 : }
6174 :
6175 3 : for (key_range, lsn) in spec.image_layers_shape.iter() {
6176 3 : let mut key = key_range.start;
6177 142 : while key < key_range.end {
6178 139 : let blob = Bytes::from(format!("[image {key}@{lsn}]"));
6179 139 : let record = Value::Image(blob.clone());
6180 139 : storage.insert((key, *lsn), record);
6181 139 :
6182 139 : key = key.next();
6183 139 : }
6184 :
6185 : // Stash some interesting LSN for future use
6186 9 : for offset in [0, 5, 100].iter() {
6187 9 : if *offset == 0 {
6188 3 : interesting_lsns.push(*lsn);
6189 3 : } else {
6190 6 : let below = lsn.checked_sub(*offset);
6191 6 : match below {
6192 6 : Some(v) if v >= spec.start_lsn => {
6193 4 : interesting_lsns.push(v);
6194 4 : }
6195 2 : _ => {}
6196 : }
6197 :
6198 6 : let above = Lsn(lsn.0 + offset);
6199 6 : interesting_lsns.push(above);
6200 : }
6201 : }
6202 : }
6203 :
6204 1 : let in_memory_test_layers = {
6205 1 : let mut acc = Vec::new();
6206 :
6207 2 : for (key_range, lsn_range) in spec.in_memory_layers_shape.iter() {
6208 2 : let mut data = Vec::new();
6209 2 :
6210 2 : let mut lsn = lsn_range.start;
6211 202 : while lsn < lsn_range.end {
6212 200 : let mut key = key_range.start;
6213 20000 : while key < key_range.end {
6214 19800 : if let Some(record) = storage.get(&(key, lsn)) {
6215 19800 : data.push((key, lsn, record.clone()));
6216 19800 : }
6217 :
6218 19800 : key = key.next();
6219 : }
6220 200 : lsn = Lsn(lsn.0 + 1);
6221 : }
6222 :
6223 2 : acc.push(InMemoryLayerTestDesc {
6224 2 : data,
6225 2 : lsn_range: lsn_range.clone(),
6226 2 : is_open: false,
6227 2 : })
6228 : }
6229 :
6230 1 : acc
6231 : };
6232 :
6233 1 : let delta_test_layers = {
6234 1 : let mut acc = Vec::new();
6235 :
6236 3 : for (key_range, lsn_range) in spec.delta_layers_shape.iter() {
6237 3 : let mut data = Vec::new();
6238 3 :
6239 3 : let mut lsn = lsn_range.start;
6240 315 : while lsn < lsn_range.end {
6241 312 : let mut key = key_range.start;
6242 10608 : while key < key_range.end {
6243 10296 : if let Some(record) = storage.get(&(key, lsn)) {
6244 10296 : data.push((key, lsn, record.clone()));
6245 10296 : }
6246 :
6247 10296 : key = key.next();
6248 : }
6249 312 : lsn = Lsn(lsn.0 + 1);
6250 : }
6251 :
6252 3 : acc.push(DeltaLayerTestDesc {
6253 3 : data,
6254 3 : lsn_range: lsn_range.clone(),
6255 3 : key_range: key_range.clone(),
6256 3 : })
6257 : }
6258 :
6259 1 : acc
6260 : };
6261 :
6262 1 : let image_test_layers = {
6263 1 : let mut acc = Vec::new();
6264 :
6265 3 : for (key_range, lsn) in spec.image_layers_shape.iter() {
6266 3 : let mut data = Vec::new();
6267 3 :
6268 3 : let mut key = key_range.start;
6269 142 : while key < key_range.end {
6270 139 : if let Some(record) = storage.get(&(key, *lsn)) {
6271 139 : let blob = match record {
6272 139 : Value::Image(blob) => blob.clone(),
6273 0 : _ => unreachable!(),
6274 : };
6275 :
6276 139 : data.push((key, blob));
6277 0 : }
6278 :
6279 139 : key = key.next();
6280 : }
6281 :
6282 3 : acc.push((*lsn, data));
6283 : }
6284 :
6285 1 : acc
6286 : };
6287 :
6288 1 : let tline = tenant
6289 1 : .create_test_timeline_with_layers(
6290 1 : new_timeline_id,
6291 1 : spec.start_lsn,
6292 1 : pg_version,
6293 1 : ctx,
6294 1 : in_memory_test_layers,
6295 1 : delta_test_layers,
6296 1 : image_test_layers,
6297 1 : spec.last_record_lsn,
6298 1 : )
6299 1 : .await?;
6300 :
6301 1 : Ok((
6302 1 : tline,
6303 1 : Storage {
6304 1 : storage,
6305 1 : start_lsn: spec.start_lsn,
6306 1 : },
6307 1 : interesting_lsns,
6308 1 : ))
6309 1 : }
6310 :
6311 : #[tokio::test]
6312 1 : async fn test_basic() -> anyhow::Result<()> {
6313 1 : let (tenant, ctx) = TenantHarness::create("test_basic").await?.load().await;
6314 1 : let tline = tenant
6315 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6316 1 : .await?;
6317 1 :
6318 1 : let mut writer = tline.writer().await;
6319 1 : writer
6320 1 : .put(
6321 1 : *TEST_KEY,
6322 1 : Lsn(0x10),
6323 1 : &Value::Image(test_img("foo at 0x10")),
6324 1 : &ctx,
6325 1 : )
6326 1 : .await?;
6327 1 : writer.finish_write(Lsn(0x10));
6328 1 : drop(writer);
6329 1 :
6330 1 : let mut writer = tline.writer().await;
6331 1 : writer
6332 1 : .put(
6333 1 : *TEST_KEY,
6334 1 : Lsn(0x20),
6335 1 : &Value::Image(test_img("foo at 0x20")),
6336 1 : &ctx,
6337 1 : )
6338 1 : .await?;
6339 1 : writer.finish_write(Lsn(0x20));
6340 1 : drop(writer);
6341 1 :
6342 1 : assert_eq!(
6343 1 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
6344 1 : test_img("foo at 0x10")
6345 1 : );
6346 1 : assert_eq!(
6347 1 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
6348 1 : test_img("foo at 0x10")
6349 1 : );
6350 1 : assert_eq!(
6351 1 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
6352 1 : test_img("foo at 0x20")
6353 1 : );
6354 1 :
6355 1 : Ok(())
6356 1 : }
6357 :
6358 : #[tokio::test]
6359 1 : async fn no_duplicate_timelines() -> anyhow::Result<()> {
6360 1 : let (tenant, ctx) = TenantHarness::create("no_duplicate_timelines")
6361 1 : .await?
6362 1 : .load()
6363 1 : .await;
6364 1 : let _ = tenant
6365 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6366 1 : .await?;
6367 1 :
6368 1 : match tenant
6369 1 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6370 1 : .await
6371 1 : {
6372 1 : Ok(_) => panic!("duplicate timeline creation should fail"),
6373 1 : Err(e) => assert_eq!(
6374 1 : e.to_string(),
6375 1 : "timeline already exists with different parameters".to_string()
6376 1 : ),
6377 1 : }
6378 1 :
6379 1 : Ok(())
6380 1 : }
6381 :
6382 : /// Convenience function to create a page image with given string as the only content
6383 5 : pub fn test_value(s: &str) -> Value {
6384 5 : let mut buf = BytesMut::new();
6385 5 : buf.extend_from_slice(s.as_bytes());
6386 5 : Value::Image(buf.freeze())
6387 5 : }
6388 :
6389 : ///
6390 : /// Test branch creation
6391 : ///
6392 : #[tokio::test]
6393 1 : async fn test_branch() -> anyhow::Result<()> {
6394 1 : use std::str::from_utf8;
6395 1 :
6396 1 : let (tenant, ctx) = TenantHarness::create("test_branch").await?.load().await;
6397 1 : let tline = tenant
6398 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6399 1 : .await?;
6400 1 : let mut writer = tline.writer().await;
6401 1 :
6402 1 : #[allow(non_snake_case)]
6403 1 : let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
6404 1 : #[allow(non_snake_case)]
6405 1 : let TEST_KEY_B: Key = Key::from_hex("110000000033333333444444445500000002").unwrap();
6406 1 :
6407 1 : // Insert a value on the timeline
6408 1 : writer
6409 1 : .put(TEST_KEY_A, Lsn(0x20), &test_value("foo at 0x20"), &ctx)
6410 1 : .await?;
6411 1 : writer
6412 1 : .put(TEST_KEY_B, Lsn(0x20), &test_value("foobar at 0x20"), &ctx)
6413 1 : .await?;
6414 1 : writer.finish_write(Lsn(0x20));
6415 1 :
6416 1 : writer
6417 1 : .put(TEST_KEY_A, Lsn(0x30), &test_value("foo at 0x30"), &ctx)
6418 1 : .await?;
6419 1 : writer.finish_write(Lsn(0x30));
6420 1 : writer
6421 1 : .put(TEST_KEY_A, Lsn(0x40), &test_value("foo at 0x40"), &ctx)
6422 1 : .await?;
6423 1 : writer.finish_write(Lsn(0x40));
6424 1 :
6425 1 : //assert_current_logical_size(&tline, Lsn(0x40));
6426 1 :
6427 1 : // Branch the history, modify relation differently on the new timeline
6428 1 : tenant
6429 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x30)), &ctx)
6430 1 : .await?;
6431 1 : let newtline = tenant
6432 1 : .get_timeline(NEW_TIMELINE_ID, true)
6433 1 : .expect("Should have a local timeline");
6434 1 : let mut new_writer = newtline.writer().await;
6435 1 : new_writer
6436 1 : .put(TEST_KEY_A, Lsn(0x40), &test_value("bar at 0x40"), &ctx)
6437 1 : .await?;
6438 1 : new_writer.finish_write(Lsn(0x40));
6439 1 :
6440 1 : // Check page contents on both branches
6441 1 : assert_eq!(
6442 1 : from_utf8(&tline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
6443 1 : "foo at 0x40"
6444 1 : );
6445 1 : assert_eq!(
6446 1 : from_utf8(&newtline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
6447 1 : "bar at 0x40"
6448 1 : );
6449 1 : assert_eq!(
6450 1 : from_utf8(&newtline.get(TEST_KEY_B, Lsn(0x40), &ctx).await?)?,
6451 1 : "foobar at 0x20"
6452 1 : );
6453 1 :
6454 1 : //assert_current_logical_size(&tline, Lsn(0x40));
6455 1 :
6456 1 : Ok(())
6457 1 : }
6458 :
6459 10 : async fn make_some_layers(
6460 10 : tline: &Timeline,
6461 10 : start_lsn: Lsn,
6462 10 : ctx: &RequestContext,
6463 10 : ) -> anyhow::Result<()> {
6464 10 : let mut lsn = start_lsn;
6465 : {
6466 10 : let mut writer = tline.writer().await;
6467 : // Create a relation on the timeline
6468 10 : writer
6469 10 : .put(
6470 10 : *TEST_KEY,
6471 10 : lsn,
6472 10 : &Value::Image(test_img(&format!("foo at {}", lsn))),
6473 10 : ctx,
6474 10 : )
6475 10 : .await?;
6476 10 : writer.finish_write(lsn);
6477 10 : lsn += 0x10;
6478 10 : writer
6479 10 : .put(
6480 10 : *TEST_KEY,
6481 10 : lsn,
6482 10 : &Value::Image(test_img(&format!("foo at {}", lsn))),
6483 10 : ctx,
6484 10 : )
6485 10 : .await?;
6486 10 : writer.finish_write(lsn);
6487 10 : lsn += 0x10;
6488 10 : }
6489 10 : tline.freeze_and_flush().await?;
6490 : {
6491 10 : let mut writer = tline.writer().await;
6492 10 : writer
6493 10 : .put(
6494 10 : *TEST_KEY,
6495 10 : lsn,
6496 10 : &Value::Image(test_img(&format!("foo at {}", lsn))),
6497 10 : ctx,
6498 10 : )
6499 10 : .await?;
6500 10 : writer.finish_write(lsn);
6501 10 : lsn += 0x10;
6502 10 : writer
6503 10 : .put(
6504 10 : *TEST_KEY,
6505 10 : lsn,
6506 10 : &Value::Image(test_img(&format!("foo at {}", lsn))),
6507 10 : ctx,
6508 10 : )
6509 10 : .await?;
6510 10 : writer.finish_write(lsn);
6511 10 : }
6512 10 : tline.freeze_and_flush().await.map_err(|e| e.into())
6513 10 : }
6514 :
6515 : #[tokio::test(start_paused = true)]
6516 1 : async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
6517 1 : let (tenant, ctx) =
6518 1 : TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
6519 1 : .await?
6520 1 : .load()
6521 1 : .await;
6522 1 : // Advance to the lsn lease deadline so that GC is not blocked by
6523 1 : // initial transition into AttachedSingle.
6524 1 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
6525 1 : tokio::time::resume();
6526 1 : let tline = tenant
6527 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6528 1 : .await?;
6529 1 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6530 1 :
6531 1 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
6532 1 : // FIXME: this doesn't actually remove any layer currently, given how the flushing
6533 1 : // and compaction works. But it does set the 'cutoff' point so that the cross check
6534 1 : // below should fail.
6535 1 : tenant
6536 1 : .gc_iteration(
6537 1 : Some(TIMELINE_ID),
6538 1 : 0x10,
6539 1 : Duration::ZERO,
6540 1 : &CancellationToken::new(),
6541 1 : &ctx,
6542 1 : )
6543 1 : .await?;
6544 1 :
6545 1 : // try to branch at lsn 25, should fail because we already garbage collected the data
6546 1 : match tenant
6547 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
6548 1 : .await
6549 1 : {
6550 1 : Ok(_) => panic!("branching should have failed"),
6551 1 : Err(err) => {
6552 1 : let CreateTimelineError::AncestorLsn(err) = err else {
6553 1 : panic!("wrong error type")
6554 1 : };
6555 1 : assert!(err.to_string().contains("invalid branch start lsn"));
6556 1 : assert!(
6557 1 : err.source()
6558 1 : .unwrap()
6559 1 : .to_string()
6560 1 : .contains("we might've already garbage collected needed data")
6561 1 : )
6562 1 : }
6563 1 : }
6564 1 :
6565 1 : Ok(())
6566 1 : }
6567 :
6568 : #[tokio::test]
6569 1 : async fn test_prohibit_branch_creation_on_pre_initdb_lsn() -> anyhow::Result<()> {
6570 1 : let (tenant, ctx) =
6571 1 : TenantHarness::create("test_prohibit_branch_creation_on_pre_initdb_lsn")
6572 1 : .await?
6573 1 : .load()
6574 1 : .await;
6575 1 :
6576 1 : let tline = tenant
6577 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x50), DEFAULT_PG_VERSION, &ctx)
6578 1 : .await?;
6579 1 : // try to branch at lsn 0x25, should fail because initdb lsn is 0x50
6580 1 : match tenant
6581 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
6582 1 : .await
6583 1 : {
6584 1 : Ok(_) => panic!("branching should have failed"),
6585 1 : Err(err) => {
6586 1 : let CreateTimelineError::AncestorLsn(err) = err else {
6587 1 : panic!("wrong error type");
6588 1 : };
6589 1 : assert!(&err.to_string().contains("invalid branch start lsn"));
6590 1 : assert!(
6591 1 : &err.source()
6592 1 : .unwrap()
6593 1 : .to_string()
6594 1 : .contains("is earlier than latest GC cutoff")
6595 1 : );
6596 1 : }
6597 1 : }
6598 1 :
6599 1 : Ok(())
6600 1 : }
6601 :
6602 : /*
6603 : // FIXME: This currently fails to error out. Calling GC doesn't currently
6604 : // remove the old value, we'd need to work a little harder
6605 : #[tokio::test]
6606 : async fn test_prohibit_get_for_garbage_collected_data() -> anyhow::Result<()> {
6607 : let repo =
6608 : RepoHarness::create("test_prohibit_get_for_garbage_collected_data")?
6609 : .load();
6610 :
6611 : let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION)?;
6612 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6613 :
6614 : repo.gc_iteration(Some(TIMELINE_ID), 0x10, Duration::ZERO)?;
6615 : let applied_gc_cutoff_lsn = tline.get_applied_gc_cutoff_lsn();
6616 : assert!(*applied_gc_cutoff_lsn > Lsn(0x25));
6617 : match tline.get(*TEST_KEY, Lsn(0x25)) {
6618 : Ok(_) => panic!("request for page should have failed"),
6619 : Err(err) => assert!(err.to_string().contains("not found at")),
6620 : }
6621 : Ok(())
6622 : }
6623 : */
6624 :
6625 : #[tokio::test]
6626 1 : async fn test_get_branchpoints_from_an_inactive_timeline() -> anyhow::Result<()> {
6627 1 : let (tenant, ctx) =
6628 1 : TenantHarness::create("test_get_branchpoints_from_an_inactive_timeline")
6629 1 : .await?
6630 1 : .load()
6631 1 : .await;
6632 1 : let tline = tenant
6633 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6634 1 : .await?;
6635 1 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6636 1 :
6637 1 : tenant
6638 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6639 1 : .await?;
6640 1 : let newtline = tenant
6641 1 : .get_timeline(NEW_TIMELINE_ID, true)
6642 1 : .expect("Should have a local timeline");
6643 1 :
6644 1 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6645 1 :
6646 1 : tline.set_broken("test".to_owned());
6647 1 :
6648 1 : tenant
6649 1 : .gc_iteration(
6650 1 : Some(TIMELINE_ID),
6651 1 : 0x10,
6652 1 : Duration::ZERO,
6653 1 : &CancellationToken::new(),
6654 1 : &ctx,
6655 1 : )
6656 1 : .await?;
6657 1 :
6658 1 : // The branchpoints should contain all timelines, even ones marked
6659 1 : // as Broken.
6660 1 : {
6661 1 : let branchpoints = &tline.gc_info.read().unwrap().retain_lsns;
6662 1 : assert_eq!(branchpoints.len(), 1);
6663 1 : assert_eq!(
6664 1 : branchpoints[0],
6665 1 : (Lsn(0x40), NEW_TIMELINE_ID, MaybeOffloaded::No)
6666 1 : );
6667 1 : }
6668 1 :
6669 1 : // You can read the key from the child branch even though the parent is
6670 1 : // Broken, as long as you don't need to access data from the parent.
6671 1 : assert_eq!(
6672 1 : newtline.get(*TEST_KEY, Lsn(0x70), &ctx).await?,
6673 1 : test_img(&format!("foo at {}", Lsn(0x70)))
6674 1 : );
6675 1 :
6676 1 : // This needs to traverse to the parent, and fails.
6677 1 : let err = newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await.unwrap_err();
6678 1 : assert!(
6679 1 : err.to_string().starts_with(&format!(
6680 1 : "bad state on timeline {}: Broken",
6681 1 : tline.timeline_id
6682 1 : )),
6683 1 : "{err}"
6684 1 : );
6685 1 :
6686 1 : Ok(())
6687 1 : }
6688 :
6689 : #[tokio::test]
6690 1 : async fn test_retain_data_in_parent_which_is_needed_for_child() -> anyhow::Result<()> {
6691 1 : let (tenant, ctx) =
6692 1 : TenantHarness::create("test_retain_data_in_parent_which_is_needed_for_child")
6693 1 : .await?
6694 1 : .load()
6695 1 : .await;
6696 1 : let tline = tenant
6697 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6698 1 : .await?;
6699 1 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6700 1 :
6701 1 : tenant
6702 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6703 1 : .await?;
6704 1 : let newtline = tenant
6705 1 : .get_timeline(NEW_TIMELINE_ID, true)
6706 1 : .expect("Should have a local timeline");
6707 1 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
6708 1 : tenant
6709 1 : .gc_iteration(
6710 1 : Some(TIMELINE_ID),
6711 1 : 0x10,
6712 1 : Duration::ZERO,
6713 1 : &CancellationToken::new(),
6714 1 : &ctx,
6715 1 : )
6716 1 : .await?;
6717 1 : assert!(newtline.get(*TEST_KEY, Lsn(0x25), &ctx).await.is_ok());
6718 1 :
6719 1 : Ok(())
6720 1 : }
6721 : #[tokio::test]
6722 1 : async fn test_parent_keeps_data_forever_after_branching() -> anyhow::Result<()> {
6723 1 : let (tenant, ctx) = TenantHarness::create("test_parent_keeps_data_forever_after_branching")
6724 1 : .await?
6725 1 : .load()
6726 1 : .await;
6727 1 : let tline = tenant
6728 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6729 1 : .await?;
6730 1 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6731 1 :
6732 1 : tenant
6733 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6734 1 : .await?;
6735 1 : let newtline = tenant
6736 1 : .get_timeline(NEW_TIMELINE_ID, true)
6737 1 : .expect("Should have a local timeline");
6738 1 :
6739 1 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6740 1 :
6741 1 : // run gc on parent
6742 1 : tenant
6743 1 : .gc_iteration(
6744 1 : Some(TIMELINE_ID),
6745 1 : 0x10,
6746 1 : Duration::ZERO,
6747 1 : &CancellationToken::new(),
6748 1 : &ctx,
6749 1 : )
6750 1 : .await?;
6751 1 :
6752 1 : // Check that the data is still accessible on the branch.
6753 1 : assert_eq!(
6754 1 : newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
6755 1 : test_img(&format!("foo at {}", Lsn(0x40)))
6756 1 : );
6757 1 :
6758 1 : Ok(())
6759 1 : }
6760 :
6761 : #[tokio::test]
6762 1 : async fn timeline_load() -> anyhow::Result<()> {
6763 1 : const TEST_NAME: &str = "timeline_load";
6764 1 : let harness = TenantHarness::create(TEST_NAME).await?;
6765 1 : {
6766 1 : let (tenant, ctx) = harness.load().await;
6767 1 : let tline = tenant
6768 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x7000), DEFAULT_PG_VERSION, &ctx)
6769 1 : .await?;
6770 1 : make_some_layers(tline.as_ref(), Lsn(0x8000), &ctx).await?;
6771 1 : // so that all uploads finish & we can call harness.load() below again
6772 1 : tenant
6773 1 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
6774 1 : .instrument(harness.span())
6775 1 : .await
6776 1 : .ok()
6777 1 : .unwrap();
6778 1 : }
6779 1 :
6780 1 : let (tenant, _ctx) = harness.load().await;
6781 1 : tenant
6782 1 : .get_timeline(TIMELINE_ID, true)
6783 1 : .expect("cannot load timeline");
6784 1 :
6785 1 : Ok(())
6786 1 : }
6787 :
6788 : #[tokio::test]
6789 1 : async fn timeline_load_with_ancestor() -> anyhow::Result<()> {
6790 1 : const TEST_NAME: &str = "timeline_load_with_ancestor";
6791 1 : let harness = TenantHarness::create(TEST_NAME).await?;
6792 1 : // create two timelines
6793 1 : {
6794 1 : let (tenant, ctx) = harness.load().await;
6795 1 : let tline = tenant
6796 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6797 1 : .await?;
6798 1 :
6799 1 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6800 1 :
6801 1 : let child_tline = tenant
6802 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6803 1 : .await?;
6804 1 : child_tline.set_state(TimelineState::Active);
6805 1 :
6806 1 : let newtline = tenant
6807 1 : .get_timeline(NEW_TIMELINE_ID, true)
6808 1 : .expect("Should have a local timeline");
6809 1 :
6810 1 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6811 1 :
6812 1 : // so that all uploads finish & we can call harness.load() below again
6813 1 : tenant
6814 1 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
6815 1 : .instrument(harness.span())
6816 1 : .await
6817 1 : .ok()
6818 1 : .unwrap();
6819 1 : }
6820 1 :
6821 1 : // check that both of them are initially unloaded
6822 1 : let (tenant, _ctx) = harness.load().await;
6823 1 :
6824 1 : // check that both, child and ancestor are loaded
6825 1 : let _child_tline = tenant
6826 1 : .get_timeline(NEW_TIMELINE_ID, true)
6827 1 : .expect("cannot get child timeline loaded");
6828 1 :
6829 1 : let _ancestor_tline = tenant
6830 1 : .get_timeline(TIMELINE_ID, true)
6831 1 : .expect("cannot get ancestor timeline loaded");
6832 1 :
6833 1 : Ok(())
6834 1 : }
6835 :
6836 : #[tokio::test]
6837 1 : async fn delta_layer_dumping() -> anyhow::Result<()> {
6838 1 : use storage_layer::AsLayerDesc;
6839 1 : let (tenant, ctx) = TenantHarness::create("test_layer_dumping")
6840 1 : .await?
6841 1 : .load()
6842 1 : .await;
6843 1 : let tline = tenant
6844 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6845 1 : .await?;
6846 1 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6847 1 :
6848 1 : let layer_map = tline.layers.read().await;
6849 1 : let level0_deltas = layer_map
6850 1 : .layer_map()?
6851 1 : .level0_deltas()
6852 1 : .iter()
6853 2 : .map(|desc| layer_map.get_from_desc(desc))
6854 1 : .collect::<Vec<_>>();
6855 1 :
6856 1 : assert!(!level0_deltas.is_empty());
6857 1 :
6858 3 : for delta in level0_deltas {
6859 1 : // Ensure we are dumping a delta layer here
6860 2 : assert!(delta.layer_desc().is_delta);
6861 2 : delta.dump(true, &ctx).await.unwrap();
6862 1 : }
6863 1 :
6864 1 : Ok(())
6865 1 : }
6866 :
6867 : #[tokio::test]
6868 1 : async fn test_images() -> anyhow::Result<()> {
6869 1 : let (tenant, ctx) = TenantHarness::create("test_images").await?.load().await;
6870 1 : let tline = tenant
6871 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6872 1 : .await?;
6873 1 :
6874 1 : let mut writer = tline.writer().await;
6875 1 : writer
6876 1 : .put(
6877 1 : *TEST_KEY,
6878 1 : Lsn(0x10),
6879 1 : &Value::Image(test_img("foo at 0x10")),
6880 1 : &ctx,
6881 1 : )
6882 1 : .await?;
6883 1 : writer.finish_write(Lsn(0x10));
6884 1 : drop(writer);
6885 1 :
6886 1 : tline.freeze_and_flush().await?;
6887 1 : tline
6888 1 : .compact(&CancellationToken::new(), EnumSet::default(), &ctx)
6889 1 : .await?;
6890 1 :
6891 1 : let mut writer = tline.writer().await;
6892 1 : writer
6893 1 : .put(
6894 1 : *TEST_KEY,
6895 1 : Lsn(0x20),
6896 1 : &Value::Image(test_img("foo at 0x20")),
6897 1 : &ctx,
6898 1 : )
6899 1 : .await?;
6900 1 : writer.finish_write(Lsn(0x20));
6901 1 : drop(writer);
6902 1 :
6903 1 : tline.freeze_and_flush().await?;
6904 1 : tline
6905 1 : .compact(&CancellationToken::new(), EnumSet::default(), &ctx)
6906 1 : .await?;
6907 1 :
6908 1 : let mut writer = tline.writer().await;
6909 1 : writer
6910 1 : .put(
6911 1 : *TEST_KEY,
6912 1 : Lsn(0x30),
6913 1 : &Value::Image(test_img("foo at 0x30")),
6914 1 : &ctx,
6915 1 : )
6916 1 : .await?;
6917 1 : writer.finish_write(Lsn(0x30));
6918 1 : drop(writer);
6919 1 :
6920 1 : tline.freeze_and_flush().await?;
6921 1 : tline
6922 1 : .compact(&CancellationToken::new(), EnumSet::default(), &ctx)
6923 1 : .await?;
6924 1 :
6925 1 : let mut writer = tline.writer().await;
6926 1 : writer
6927 1 : .put(
6928 1 : *TEST_KEY,
6929 1 : Lsn(0x40),
6930 1 : &Value::Image(test_img("foo at 0x40")),
6931 1 : &ctx,
6932 1 : )
6933 1 : .await?;
6934 1 : writer.finish_write(Lsn(0x40));
6935 1 : drop(writer);
6936 1 :
6937 1 : tline.freeze_and_flush().await?;
6938 1 : tline
6939 1 : .compact(&CancellationToken::new(), EnumSet::default(), &ctx)
6940 1 : .await?;
6941 1 :
6942 1 : assert_eq!(
6943 1 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
6944 1 : test_img("foo at 0x10")
6945 1 : );
6946 1 : assert_eq!(
6947 1 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
6948 1 : test_img("foo at 0x10")
6949 1 : );
6950 1 : assert_eq!(
6951 1 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
6952 1 : test_img("foo at 0x20")
6953 1 : );
6954 1 : assert_eq!(
6955 1 : tline.get(*TEST_KEY, Lsn(0x30), &ctx).await?,
6956 1 : test_img("foo at 0x30")
6957 1 : );
6958 1 : assert_eq!(
6959 1 : tline.get(*TEST_KEY, Lsn(0x40), &ctx).await?,
6960 1 : test_img("foo at 0x40")
6961 1 : );
6962 1 :
6963 1 : Ok(())
6964 1 : }
6965 :
6966 2 : async fn bulk_insert_compact_gc(
6967 2 : tenant: &TenantShard,
6968 2 : timeline: &Arc<Timeline>,
6969 2 : ctx: &RequestContext,
6970 2 : lsn: Lsn,
6971 2 : repeat: usize,
6972 2 : key_count: usize,
6973 2 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
6974 2 : let compact = true;
6975 2 : bulk_insert_maybe_compact_gc(tenant, timeline, ctx, lsn, repeat, key_count, compact).await
6976 2 : }
6977 :
6978 4 : async fn bulk_insert_maybe_compact_gc(
6979 4 : tenant: &TenantShard,
6980 4 : timeline: &Arc<Timeline>,
6981 4 : ctx: &RequestContext,
6982 4 : mut lsn: Lsn,
6983 4 : repeat: usize,
6984 4 : key_count: usize,
6985 4 : compact: bool,
6986 4 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
6987 4 : let mut inserted: HashMap<Key, BTreeSet<Lsn>> = Default::default();
6988 4 :
6989 4 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6990 4 : let mut blknum = 0;
6991 4 :
6992 4 : // Enforce that key range is monotonously increasing
6993 4 : let mut keyspace = KeySpaceAccum::new();
6994 4 :
6995 4 : let cancel = CancellationToken::new();
6996 4 :
6997 4 : for _ in 0..repeat {
6998 200 : for _ in 0..key_count {
6999 2000000 : test_key.field6 = blknum;
7000 2000000 : let mut writer = timeline.writer().await;
7001 2000000 : writer
7002 2000000 : .put(
7003 2000000 : test_key,
7004 2000000 : lsn,
7005 2000000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7006 2000000 : ctx,
7007 2000000 : )
7008 2000000 : .await?;
7009 2000000 : inserted.entry(test_key).or_default().insert(lsn);
7010 2000000 : writer.finish_write(lsn);
7011 2000000 : drop(writer);
7012 2000000 :
7013 2000000 : keyspace.add_key(test_key);
7014 2000000 :
7015 2000000 : lsn = Lsn(lsn.0 + 0x10);
7016 2000000 : blknum += 1;
7017 : }
7018 :
7019 200 : timeline.freeze_and_flush().await?;
7020 200 : if compact {
7021 : // this requires timeline to be &Arc<Timeline>
7022 100 : timeline.compact(&cancel, EnumSet::default(), ctx).await?;
7023 100 : }
7024 :
7025 : // this doesn't really need to use the timeline_id target, but it is closer to what it
7026 : // originally was.
7027 200 : let res = tenant
7028 200 : .gc_iteration(Some(timeline.timeline_id), 0, Duration::ZERO, &cancel, ctx)
7029 200 : .await?;
7030 :
7031 200 : assert_eq!(res.layers_removed, 0, "this never removes anything");
7032 : }
7033 :
7034 4 : Ok(inserted)
7035 4 : }
7036 :
7037 : //
7038 : // Insert 1000 key-value pairs with increasing keys, flush, compact, GC.
7039 : // Repeat 50 times.
7040 : //
7041 : #[tokio::test]
7042 1 : async fn test_bulk_insert() -> anyhow::Result<()> {
7043 1 : let harness = TenantHarness::create("test_bulk_insert").await?;
7044 1 : let (tenant, ctx) = harness.load().await;
7045 1 : let tline = tenant
7046 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
7047 1 : .await?;
7048 1 :
7049 1 : let lsn = Lsn(0x10);
7050 1 : bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
7051 1 :
7052 1 : Ok(())
7053 1 : }
7054 :
7055 : // Test the vectored get real implementation against a simple sequential implementation.
7056 : //
7057 : // The test generates a keyspace by repeatedly flushing the in-memory layer and compacting.
7058 : // Projected to 2D the key space looks like below. Lsn grows upwards on the Y axis and keys
7059 : // grow to the right on the X axis.
7060 : // [Delta]
7061 : // [Delta]
7062 : // [Delta]
7063 : // [Delta]
7064 : // ------------ Image ---------------
7065 : //
7066 : // After layer generation we pick the ranges to query as follows:
7067 : // 1. The beginning of each delta layer
7068 : // 2. At the seam between two adjacent delta layers
7069 : //
7070 : // There's one major downside to this test: delta layers only contains images,
7071 : // so the search can stop at the first delta layer and doesn't traverse any deeper.
7072 : #[tokio::test]
7073 1 : async fn test_get_vectored() -> anyhow::Result<()> {
7074 1 : let harness = TenantHarness::create("test_get_vectored").await?;
7075 1 : let (tenant, ctx) = harness.load().await;
7076 1 : let io_concurrency = IoConcurrency::spawn_for_test();
7077 1 : let tline = tenant
7078 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
7079 1 : .await?;
7080 1 :
7081 1 : let lsn = Lsn(0x10);
7082 1 : let inserted = bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
7083 1 :
7084 1 : let guard = tline.layers.read().await;
7085 1 : let lm = guard.layer_map()?;
7086 1 :
7087 1 : lm.dump(true, &ctx).await?;
7088 1 :
7089 1 : let mut reads = Vec::new();
7090 1 : let mut prev = None;
7091 6 : lm.iter_historic_layers().for_each(|desc| {
7092 6 : if !desc.is_delta() {
7093 1 : prev = Some(desc.clone());
7094 1 : return;
7095 5 : }
7096 5 :
7097 5 : let start = desc.key_range.start;
7098 5 : let end = desc
7099 5 : .key_range
7100 5 : .start
7101 5 : .add(Timeline::MAX_GET_VECTORED_KEYS.try_into().unwrap());
7102 5 : reads.push(KeySpace {
7103 5 : ranges: vec![start..end],
7104 5 : });
7105 1 :
7106 5 : if let Some(prev) = &prev {
7107 5 : if !prev.is_delta() {
7108 5 : return;
7109 1 : }
7110 0 :
7111 0 : let first_range = Key {
7112 0 : field6: prev.key_range.end.field6 - 4,
7113 0 : ..prev.key_range.end
7114 0 : }..prev.key_range.end;
7115 0 :
7116 0 : let second_range = desc.key_range.start..Key {
7117 0 : field6: desc.key_range.start.field6 + 4,
7118 0 : ..desc.key_range.start
7119 0 : };
7120 0 :
7121 0 : reads.push(KeySpace {
7122 0 : ranges: vec![first_range, second_range],
7123 0 : });
7124 1 : };
7125 1 :
7126 1 : prev = Some(desc.clone());
7127 6 : });
7128 1 :
7129 1 : drop(guard);
7130 1 :
7131 1 : // Pick a big LSN such that we query over all the changes.
7132 1 : let reads_lsn = Lsn(u64::MAX - 1);
7133 1 :
7134 6 : for read in reads {
7135 5 : info!("Doing vectored read on {:?}", read);
7136 1 :
7137 5 : let query = VersionedKeySpaceQuery::uniform(read.clone(), reads_lsn);
7138 1 :
7139 5 : let vectored_res = tline
7140 5 : .get_vectored_impl(
7141 5 : query,
7142 5 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7143 5 : &ctx,
7144 5 : )
7145 5 : .await;
7146 1 :
7147 5 : let mut expected_lsns: HashMap<Key, Lsn> = Default::default();
7148 5 : let mut expect_missing = false;
7149 5 : let mut key = read.start().unwrap();
7150 165 : while key != read.end().unwrap() {
7151 160 : if let Some(lsns) = inserted.get(&key) {
7152 160 : let expected_lsn = lsns.iter().rfind(|lsn| **lsn <= reads_lsn);
7153 160 : match expected_lsn {
7154 160 : Some(lsn) => {
7155 160 : expected_lsns.insert(key, *lsn);
7156 160 : }
7157 1 : None => {
7158 1 : expect_missing = true;
7159 0 : break;
7160 1 : }
7161 1 : }
7162 1 : } else {
7163 1 : expect_missing = true;
7164 0 : break;
7165 1 : }
7166 1 :
7167 160 : key = key.next();
7168 1 : }
7169 1 :
7170 5 : if expect_missing {
7171 1 : assert!(matches!(vectored_res, Err(GetVectoredError::MissingKey(_))));
7172 1 : } else {
7173 160 : for (key, image) in vectored_res? {
7174 160 : let expected_lsn = expected_lsns.get(&key).expect("determined above");
7175 160 : let expected_image = test_img(&format!("{} at {}", key.field6, expected_lsn));
7176 160 : assert_eq!(image?, expected_image);
7177 1 : }
7178 1 : }
7179 1 : }
7180 1 :
7181 1 : Ok(())
7182 1 : }
7183 :
7184 : #[tokio::test]
7185 1 : async fn test_get_vectored_aux_files() -> anyhow::Result<()> {
7186 1 : let harness = TenantHarness::create("test_get_vectored_aux_files").await?;
7187 1 :
7188 1 : let (tenant, ctx) = harness.load().await;
7189 1 : let io_concurrency = IoConcurrency::spawn_for_test();
7190 1 : let (tline, ctx) = tenant
7191 1 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
7192 1 : .await?;
7193 1 : let tline = tline.raw_timeline().unwrap();
7194 1 :
7195 1 : let mut modification = tline.begin_modification(Lsn(0x1000));
7196 1 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
7197 1 : modification.set_lsn(Lsn(0x1008))?;
7198 1 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
7199 1 : modification.commit(&ctx).await?;
7200 1 :
7201 1 : let child_timeline_id = TimelineId::generate();
7202 1 : tenant
7203 1 : .branch_timeline_test(
7204 1 : tline,
7205 1 : child_timeline_id,
7206 1 : Some(tline.get_last_record_lsn()),
7207 1 : &ctx,
7208 1 : )
7209 1 : .await?;
7210 1 :
7211 1 : let child_timeline = tenant
7212 1 : .get_timeline(child_timeline_id, true)
7213 1 : .expect("Should have the branched timeline");
7214 1 :
7215 1 : let aux_keyspace = KeySpace {
7216 1 : ranges: vec![NON_INHERITED_RANGE],
7217 1 : };
7218 1 : let read_lsn = child_timeline.get_last_record_lsn();
7219 1 :
7220 1 : let query = VersionedKeySpaceQuery::uniform(aux_keyspace.clone(), read_lsn);
7221 1 :
7222 1 : let vectored_res = child_timeline
7223 1 : .get_vectored_impl(
7224 1 : query,
7225 1 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7226 1 : &ctx,
7227 1 : )
7228 1 : .await;
7229 1 :
7230 1 : let images = vectored_res?;
7231 1 : assert!(images.is_empty());
7232 1 : Ok(())
7233 1 : }
7234 :
7235 : // Test that vectored get handles layer gaps correctly
7236 : // by advancing into the next ancestor timeline if required.
7237 : //
7238 : // The test generates timelines that look like the diagram below.
7239 : // We leave a gap in one of the L1 layers at `gap_at_key` (`/` in the diagram).
7240 : // The reconstruct data for that key lies in the ancestor timeline (`X` in the diagram).
7241 : //
7242 : // ```
7243 : //-------------------------------+
7244 : // ... |
7245 : // [ L1 ] |
7246 : // [ / L1 ] | Child Timeline
7247 : // ... |
7248 : // ------------------------------+
7249 : // [ X L1 ] | Parent Timeline
7250 : // ------------------------------+
7251 : // ```
7252 : #[tokio::test]
7253 1 : async fn test_get_vectored_key_gap() -> anyhow::Result<()> {
7254 1 : let tenant_conf = pageserver_api::models::TenantConfig {
7255 1 : // Make compaction deterministic
7256 1 : gc_period: Some(Duration::ZERO),
7257 1 : compaction_period: Some(Duration::ZERO),
7258 1 : // Encourage creation of L1 layers
7259 1 : checkpoint_distance: Some(16 * 1024),
7260 1 : compaction_target_size: Some(8 * 1024),
7261 1 : ..Default::default()
7262 1 : };
7263 1 :
7264 1 : let harness = TenantHarness::create_custom(
7265 1 : "test_get_vectored_key_gap",
7266 1 : tenant_conf,
7267 1 : TenantId::generate(),
7268 1 : ShardIdentity::unsharded(),
7269 1 : Generation::new(0xdeadbeef),
7270 1 : )
7271 1 : .await?;
7272 1 : let (tenant, ctx) = harness.load().await;
7273 1 : let io_concurrency = IoConcurrency::spawn_for_test();
7274 1 :
7275 1 : let mut current_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7276 1 : let gap_at_key = current_key.add(100);
7277 1 : let mut current_lsn = Lsn(0x10);
7278 1 :
7279 1 : const KEY_COUNT: usize = 10_000;
7280 1 :
7281 1 : let timeline_id = TimelineId::generate();
7282 1 : let current_timeline = tenant
7283 1 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
7284 1 : .await?;
7285 1 :
7286 1 : current_lsn += 0x100;
7287 1 :
7288 1 : let mut writer = current_timeline.writer().await;
7289 1 : writer
7290 1 : .put(
7291 1 : gap_at_key,
7292 1 : current_lsn,
7293 1 : &Value::Image(test_img(&format!("{} at {}", gap_at_key, current_lsn))),
7294 1 : &ctx,
7295 1 : )
7296 1 : .await?;
7297 1 : writer.finish_write(current_lsn);
7298 1 : drop(writer);
7299 1 :
7300 1 : let mut latest_lsns = HashMap::new();
7301 1 : latest_lsns.insert(gap_at_key, current_lsn);
7302 1 :
7303 1 : current_timeline.freeze_and_flush().await?;
7304 1 :
7305 1 : let child_timeline_id = TimelineId::generate();
7306 1 :
7307 1 : tenant
7308 1 : .branch_timeline_test(
7309 1 : ¤t_timeline,
7310 1 : child_timeline_id,
7311 1 : Some(current_lsn),
7312 1 : &ctx,
7313 1 : )
7314 1 : .await?;
7315 1 : let child_timeline = tenant
7316 1 : .get_timeline(child_timeline_id, true)
7317 1 : .expect("Should have the branched timeline");
7318 1 :
7319 10001 : for i in 0..KEY_COUNT {
7320 10000 : if current_key == gap_at_key {
7321 1 : current_key = current_key.next();
7322 1 : continue;
7323 9999 : }
7324 9999 :
7325 9999 : current_lsn += 0x10;
7326 1 :
7327 9999 : let mut writer = child_timeline.writer().await;
7328 9999 : writer
7329 9999 : .put(
7330 9999 : current_key,
7331 9999 : current_lsn,
7332 9999 : &Value::Image(test_img(&format!("{} at {}", current_key, current_lsn))),
7333 9999 : &ctx,
7334 9999 : )
7335 9999 : .await?;
7336 9999 : writer.finish_write(current_lsn);
7337 9999 : drop(writer);
7338 9999 :
7339 9999 : latest_lsns.insert(current_key, current_lsn);
7340 9999 : current_key = current_key.next();
7341 9999 :
7342 9999 : // Flush every now and then to encourage layer file creation.
7343 9999 : if i % 500 == 0 {
7344 20 : child_timeline.freeze_and_flush().await?;
7345 9979 : }
7346 1 : }
7347 1 :
7348 1 : child_timeline.freeze_and_flush().await?;
7349 1 : let mut flags = EnumSet::new();
7350 1 : flags.insert(CompactFlags::ForceRepartition);
7351 1 : child_timeline
7352 1 : .compact(&CancellationToken::new(), flags, &ctx)
7353 1 : .await?;
7354 1 :
7355 1 : let key_near_end = {
7356 1 : let mut tmp = current_key;
7357 1 : tmp.field6 -= 10;
7358 1 : tmp
7359 1 : };
7360 1 :
7361 1 : let key_near_gap = {
7362 1 : let mut tmp = gap_at_key;
7363 1 : tmp.field6 -= 10;
7364 1 : tmp
7365 1 : };
7366 1 :
7367 1 : let read = KeySpace {
7368 1 : ranges: vec![key_near_gap..gap_at_key.next(), key_near_end..current_key],
7369 1 : };
7370 1 :
7371 1 : let query = VersionedKeySpaceQuery::uniform(read.clone(), current_lsn);
7372 1 :
7373 1 : let results = child_timeline
7374 1 : .get_vectored_impl(
7375 1 : query,
7376 1 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7377 1 : &ctx,
7378 1 : )
7379 1 : .await?;
7380 1 :
7381 22 : for (key, img_res) in results {
7382 21 : let expected = test_img(&format!("{} at {}", key, latest_lsns[&key]));
7383 21 : assert_eq!(img_res?, expected);
7384 1 : }
7385 1 :
7386 1 : Ok(())
7387 1 : }
7388 :
7389 : // Test that vectored get descends into ancestor timelines correctly and
7390 : // does not return an image that's newer than requested.
7391 : //
7392 : // The diagram below ilustrates an interesting case. We have a parent timeline
7393 : // (top of the Lsn range) and a child timeline. The request key cannot be reconstructed
7394 : // from the child timeline, so the parent timeline must be visited. When advacing into
7395 : // the child timeline, the read path needs to remember what the requested Lsn was in
7396 : // order to avoid returning an image that's too new. The test below constructs such
7397 : // a timeline setup and does a few queries around the Lsn of each page image.
7398 : // ```
7399 : // LSN
7400 : // ^
7401 : // |
7402 : // |
7403 : // 500 | --------------------------------------> branch point
7404 : // 400 | X
7405 : // 300 | X
7406 : // 200 | --------------------------------------> requested lsn
7407 : // 100 | X
7408 : // |---------------------------------------> Key
7409 : // |
7410 : // ------> requested key
7411 : //
7412 : // Legend:
7413 : // * X - page images
7414 : // ```
7415 : #[tokio::test]
7416 1 : async fn test_get_vectored_ancestor_descent() -> anyhow::Result<()> {
7417 1 : let harness = TenantHarness::create("test_get_vectored_on_lsn_axis").await?;
7418 1 : let (tenant, ctx) = harness.load().await;
7419 1 : let io_concurrency = IoConcurrency::spawn_for_test();
7420 1 :
7421 1 : let start_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7422 1 : let end_key = start_key.add(1000);
7423 1 : let child_gap_at_key = start_key.add(500);
7424 1 : let mut parent_gap_lsns: BTreeMap<Lsn, String> = BTreeMap::new();
7425 1 :
7426 1 : let mut current_lsn = Lsn(0x10);
7427 1 :
7428 1 : let timeline_id = TimelineId::generate();
7429 1 : let parent_timeline = tenant
7430 1 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
7431 1 : .await?;
7432 1 :
7433 1 : current_lsn += 0x100;
7434 1 :
7435 4 : for _ in 0..3 {
7436 3 : let mut key = start_key;
7437 3003 : while key < end_key {
7438 3000 : current_lsn += 0x10;
7439 3000 :
7440 3000 : let image_value = format!("{} at {}", child_gap_at_key, current_lsn);
7441 1 :
7442 3000 : let mut writer = parent_timeline.writer().await;
7443 3000 : writer
7444 3000 : .put(
7445 3000 : key,
7446 3000 : current_lsn,
7447 3000 : &Value::Image(test_img(&image_value)),
7448 3000 : &ctx,
7449 3000 : )
7450 3000 : .await?;
7451 3000 : writer.finish_write(current_lsn);
7452 3000 :
7453 3000 : if key == child_gap_at_key {
7454 3 : parent_gap_lsns.insert(current_lsn, image_value);
7455 2997 : }
7456 1 :
7457 3000 : key = key.next();
7458 1 : }
7459 1 :
7460 3 : parent_timeline.freeze_and_flush().await?;
7461 1 : }
7462 1 :
7463 1 : let child_timeline_id = TimelineId::generate();
7464 1 :
7465 1 : let child_timeline = tenant
7466 1 : .branch_timeline_test(&parent_timeline, child_timeline_id, Some(current_lsn), &ctx)
7467 1 : .await?;
7468 1 :
7469 1 : let mut key = start_key;
7470 1001 : while key < end_key {
7471 1000 : if key == child_gap_at_key {
7472 1 : key = key.next();
7473 1 : continue;
7474 999 : }
7475 999 :
7476 999 : current_lsn += 0x10;
7477 1 :
7478 999 : let mut writer = child_timeline.writer().await;
7479 999 : writer
7480 999 : .put(
7481 999 : key,
7482 999 : current_lsn,
7483 999 : &Value::Image(test_img(&format!("{} at {}", key, current_lsn))),
7484 999 : &ctx,
7485 999 : )
7486 999 : .await?;
7487 999 : writer.finish_write(current_lsn);
7488 999 :
7489 999 : key = key.next();
7490 1 : }
7491 1 :
7492 1 : child_timeline.freeze_and_flush().await?;
7493 1 :
7494 1 : let lsn_offsets: [i64; 5] = [-10, -1, 0, 1, 10];
7495 1 : let mut query_lsns = Vec::new();
7496 3 : for image_lsn in parent_gap_lsns.keys().rev() {
7497 18 : for offset in lsn_offsets {
7498 15 : query_lsns.push(Lsn(image_lsn
7499 15 : .0
7500 15 : .checked_add_signed(offset)
7501 15 : .expect("Shouldn't overflow")));
7502 15 : }
7503 1 : }
7504 1 :
7505 16 : for query_lsn in query_lsns {
7506 15 : let query = VersionedKeySpaceQuery::uniform(
7507 15 : KeySpace {
7508 15 : ranges: vec![child_gap_at_key..child_gap_at_key.next()],
7509 15 : },
7510 15 : query_lsn,
7511 15 : );
7512 1 :
7513 15 : let results = child_timeline
7514 15 : .get_vectored_impl(
7515 15 : query,
7516 15 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7517 15 : &ctx,
7518 15 : )
7519 15 : .await;
7520 1 :
7521 15 : let expected_item = parent_gap_lsns
7522 15 : .iter()
7523 15 : .rev()
7524 34 : .find(|(lsn, _)| **lsn <= query_lsn);
7525 15 :
7526 15 : info!(
7527 1 : "Doing vectored read at LSN {}. Expecting image to be: {:?}",
7528 1 : query_lsn, expected_item
7529 1 : );
7530 1 :
7531 15 : match expected_item {
7532 13 : Some((_, img_value)) => {
7533 13 : let key_results = results.expect("No vectored get error expected");
7534 13 : let key_result = &key_results[&child_gap_at_key];
7535 13 : let returned_img = key_result
7536 13 : .as_ref()
7537 13 : .expect("No page reconstruct error expected");
7538 13 :
7539 13 : info!(
7540 1 : "Vectored read at LSN {} returned image {}",
7541 0 : query_lsn,
7542 0 : std::str::from_utf8(returned_img)?
7543 1 : );
7544 13 : assert_eq!(*returned_img, test_img(img_value));
7545 1 : }
7546 1 : None => {
7547 2 : assert!(matches!(results, Err(GetVectoredError::MissingKey(_))));
7548 1 : }
7549 1 : }
7550 1 : }
7551 1 :
7552 1 : Ok(())
7553 1 : }
7554 :
7555 : #[tokio::test]
7556 1 : async fn test_random_updates() -> anyhow::Result<()> {
7557 1 : let names_algorithms = [
7558 1 : ("test_random_updates_legacy", CompactionAlgorithm::Legacy),
7559 1 : ("test_random_updates_tiered", CompactionAlgorithm::Tiered),
7560 1 : ];
7561 3 : for (name, algorithm) in names_algorithms {
7562 2 : test_random_updates_algorithm(name, algorithm).await?;
7563 1 : }
7564 1 : Ok(())
7565 1 : }
7566 :
7567 2 : async fn test_random_updates_algorithm(
7568 2 : name: &'static str,
7569 2 : compaction_algorithm: CompactionAlgorithm,
7570 2 : ) -> anyhow::Result<()> {
7571 2 : let mut harness = TenantHarness::create(name).await?;
7572 2 : harness.tenant_conf.compaction_algorithm = Some(CompactionAlgorithmSettings {
7573 2 : kind: compaction_algorithm,
7574 2 : });
7575 2 : let (tenant, ctx) = harness.load().await;
7576 2 : let tline = tenant
7577 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7578 2 : .await?;
7579 :
7580 : const NUM_KEYS: usize = 1000;
7581 2 : let cancel = CancellationToken::new();
7582 2 :
7583 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7584 2 : let mut test_key_end = test_key;
7585 2 : test_key_end.field6 = NUM_KEYS as u32;
7586 2 : tline.add_extra_test_dense_keyspace(KeySpace::single(test_key..test_key_end));
7587 2 :
7588 2 : let mut keyspace = KeySpaceAccum::new();
7589 2 :
7590 2 : // Track when each page was last modified. Used to assert that
7591 2 : // a read sees the latest page version.
7592 2 : let mut updated = [Lsn(0); NUM_KEYS];
7593 2 :
7594 2 : let mut lsn = Lsn(0x10);
7595 : #[allow(clippy::needless_range_loop)]
7596 2002 : for blknum in 0..NUM_KEYS {
7597 2000 : lsn = Lsn(lsn.0 + 0x10);
7598 2000 : test_key.field6 = blknum as u32;
7599 2000 : let mut writer = tline.writer().await;
7600 2000 : writer
7601 2000 : .put(
7602 2000 : test_key,
7603 2000 : lsn,
7604 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7605 2000 : &ctx,
7606 2000 : )
7607 2000 : .await?;
7608 2000 : writer.finish_write(lsn);
7609 2000 : updated[blknum] = lsn;
7610 2000 : drop(writer);
7611 2000 :
7612 2000 : keyspace.add_key(test_key);
7613 : }
7614 :
7615 102 : for _ in 0..50 {
7616 100100 : for _ in 0..NUM_KEYS {
7617 100000 : lsn = Lsn(lsn.0 + 0x10);
7618 100000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7619 100000 : test_key.field6 = blknum as u32;
7620 100000 : let mut writer = tline.writer().await;
7621 100000 : writer
7622 100000 : .put(
7623 100000 : test_key,
7624 100000 : lsn,
7625 100000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7626 100000 : &ctx,
7627 100000 : )
7628 100000 : .await?;
7629 100000 : writer.finish_write(lsn);
7630 100000 : drop(writer);
7631 100000 : updated[blknum] = lsn;
7632 : }
7633 :
7634 : // Read all the blocks
7635 100000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7636 100000 : test_key.field6 = blknum as u32;
7637 100000 : assert_eq!(
7638 100000 : tline.get(test_key, lsn, &ctx).await?,
7639 100000 : test_img(&format!("{} at {}", blknum, last_lsn))
7640 : );
7641 : }
7642 :
7643 : // Perform a cycle of flush, and GC
7644 100 : tline.freeze_and_flush().await?;
7645 100 : tenant
7646 100 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7647 100 : .await?;
7648 : }
7649 :
7650 2 : Ok(())
7651 2 : }
7652 :
7653 : #[tokio::test]
7654 1 : async fn test_traverse_branches() -> anyhow::Result<()> {
7655 1 : let (tenant, ctx) = TenantHarness::create("test_traverse_branches")
7656 1 : .await?
7657 1 : .load()
7658 1 : .await;
7659 1 : let mut tline = tenant
7660 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7661 1 : .await?;
7662 1 :
7663 1 : const NUM_KEYS: usize = 1000;
7664 1 :
7665 1 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7666 1 :
7667 1 : let mut keyspace = KeySpaceAccum::new();
7668 1 :
7669 1 : let cancel = CancellationToken::new();
7670 1 :
7671 1 : // Track when each page was last modified. Used to assert that
7672 1 : // a read sees the latest page version.
7673 1 : let mut updated = [Lsn(0); NUM_KEYS];
7674 1 :
7675 1 : let mut lsn = Lsn(0x10);
7676 1 : #[allow(clippy::needless_range_loop)]
7677 1001 : for blknum in 0..NUM_KEYS {
7678 1000 : lsn = Lsn(lsn.0 + 0x10);
7679 1000 : test_key.field6 = blknum as u32;
7680 1000 : let mut writer = tline.writer().await;
7681 1000 : writer
7682 1000 : .put(
7683 1000 : test_key,
7684 1000 : lsn,
7685 1000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7686 1000 : &ctx,
7687 1000 : )
7688 1000 : .await?;
7689 1000 : writer.finish_write(lsn);
7690 1000 : updated[blknum] = lsn;
7691 1000 : drop(writer);
7692 1000 :
7693 1000 : keyspace.add_key(test_key);
7694 1 : }
7695 1 :
7696 51 : for _ in 0..50 {
7697 50 : let new_tline_id = TimelineId::generate();
7698 50 : tenant
7699 50 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7700 50 : .await?;
7701 50 : tline = tenant
7702 50 : .get_timeline(new_tline_id, true)
7703 50 : .expect("Should have the branched timeline");
7704 1 :
7705 50050 : for _ in 0..NUM_KEYS {
7706 50000 : lsn = Lsn(lsn.0 + 0x10);
7707 50000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7708 50000 : test_key.field6 = blknum as u32;
7709 50000 : let mut writer = tline.writer().await;
7710 50000 : writer
7711 50000 : .put(
7712 50000 : test_key,
7713 50000 : lsn,
7714 50000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7715 50000 : &ctx,
7716 50000 : )
7717 50000 : .await?;
7718 50000 : println!("updating {} at {}", blknum, lsn);
7719 50000 : writer.finish_write(lsn);
7720 50000 : drop(writer);
7721 50000 : updated[blknum] = lsn;
7722 1 : }
7723 1 :
7724 1 : // Read all the blocks
7725 50000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7726 50000 : test_key.field6 = blknum as u32;
7727 50000 : assert_eq!(
7728 50000 : tline.get(test_key, lsn, &ctx).await?,
7729 50000 : test_img(&format!("{} at {}", blknum, last_lsn))
7730 1 : );
7731 1 : }
7732 1 :
7733 1 : // Perform a cycle of flush, compact, and GC
7734 50 : tline.freeze_and_flush().await?;
7735 50 : tline.compact(&cancel, EnumSet::default(), &ctx).await?;
7736 50 : tenant
7737 50 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7738 50 : .await?;
7739 1 : }
7740 1 :
7741 1 : Ok(())
7742 1 : }
7743 :
7744 : #[tokio::test]
7745 1 : async fn test_traverse_ancestors() -> anyhow::Result<()> {
7746 1 : let (tenant, ctx) = TenantHarness::create("test_traverse_ancestors")
7747 1 : .await?
7748 1 : .load()
7749 1 : .await;
7750 1 : let mut tline = tenant
7751 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7752 1 : .await?;
7753 1 :
7754 1 : const NUM_KEYS: usize = 100;
7755 1 : const NUM_TLINES: usize = 50;
7756 1 :
7757 1 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7758 1 : // Track page mutation lsns across different timelines.
7759 1 : let mut updated = [[Lsn(0); NUM_KEYS]; NUM_TLINES];
7760 1 :
7761 1 : let mut lsn = Lsn(0x10);
7762 1 :
7763 1 : #[allow(clippy::needless_range_loop)]
7764 51 : for idx in 0..NUM_TLINES {
7765 50 : let new_tline_id = TimelineId::generate();
7766 50 : tenant
7767 50 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7768 50 : .await?;
7769 50 : tline = tenant
7770 50 : .get_timeline(new_tline_id, true)
7771 50 : .expect("Should have the branched timeline");
7772 1 :
7773 5050 : for _ in 0..NUM_KEYS {
7774 5000 : lsn = Lsn(lsn.0 + 0x10);
7775 5000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7776 5000 : test_key.field6 = blknum as u32;
7777 5000 : let mut writer = tline.writer().await;
7778 5000 : writer
7779 5000 : .put(
7780 5000 : test_key,
7781 5000 : lsn,
7782 5000 : &Value::Image(test_img(&format!("{} {} at {}", idx, blknum, lsn))),
7783 5000 : &ctx,
7784 5000 : )
7785 5000 : .await?;
7786 5000 : println!("updating [{}][{}] at {}", idx, blknum, lsn);
7787 5000 : writer.finish_write(lsn);
7788 5000 : drop(writer);
7789 5000 : updated[idx][blknum] = lsn;
7790 1 : }
7791 1 : }
7792 1 :
7793 1 : // Read pages from leaf timeline across all ancestors.
7794 50 : for (idx, lsns) in updated.iter().enumerate() {
7795 5000 : for (blknum, lsn) in lsns.iter().enumerate() {
7796 1 : // Skip empty mutations.
7797 5000 : if lsn.0 == 0 {
7798 1831 : continue;
7799 3169 : }
7800 3169 : println!("checking [{idx}][{blknum}] at {lsn}");
7801 3169 : test_key.field6 = blknum as u32;
7802 3169 : assert_eq!(
7803 3169 : tline.get(test_key, *lsn, &ctx).await?,
7804 3169 : test_img(&format!("{idx} {blknum} at {lsn}"))
7805 1 : );
7806 1 : }
7807 1 : }
7808 1 : Ok(())
7809 1 : }
7810 :
7811 : #[tokio::test]
7812 1 : async fn test_write_at_initdb_lsn_takes_optimization_code_path() -> anyhow::Result<()> {
7813 1 : let (tenant, ctx) = TenantHarness::create("test_empty_test_timeline_is_usable")
7814 1 : .await?
7815 1 : .load()
7816 1 : .await;
7817 1 :
7818 1 : let initdb_lsn = Lsn(0x20);
7819 1 : let (utline, ctx) = tenant
7820 1 : .create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)
7821 1 : .await?;
7822 1 : let tline = utline.raw_timeline().unwrap();
7823 1 :
7824 1 : // Spawn flush loop now so that we can set the `expect_initdb_optimization`
7825 1 : tline.maybe_spawn_flush_loop();
7826 1 :
7827 1 : // Make sure the timeline has the minimum set of required keys for operation.
7828 1 : // The only operation you can always do on an empty timeline is to `put` new data.
7829 1 : // Except if you `put` at `initdb_lsn`.
7830 1 : // In that case, there's an optimization to directly create image layers instead of delta layers.
7831 1 : // It uses `repartition()`, which assumes some keys to be present.
7832 1 : // Let's make sure the test timeline can handle that case.
7833 1 : {
7834 1 : let mut state = tline.flush_loop_state.lock().unwrap();
7835 1 : assert_eq!(
7836 1 : timeline::FlushLoopState::Running {
7837 1 : expect_initdb_optimization: false,
7838 1 : initdb_optimization_count: 0,
7839 1 : },
7840 1 : *state
7841 1 : );
7842 1 : *state = timeline::FlushLoopState::Running {
7843 1 : expect_initdb_optimization: true,
7844 1 : initdb_optimization_count: 0,
7845 1 : };
7846 1 : }
7847 1 :
7848 1 : // Make writes at the initdb_lsn. When we flush it below, it should be handled by the optimization.
7849 1 : // As explained above, the optimization requires some keys to be present.
7850 1 : // As per `create_empty_timeline` documentation, use init_empty to set them.
7851 1 : // This is what `create_test_timeline` does, by the way.
7852 1 : let mut modification = tline.begin_modification(initdb_lsn);
7853 1 : modification
7854 1 : .init_empty_test_timeline()
7855 1 : .context("init_empty_test_timeline")?;
7856 1 : modification
7857 1 : .commit(&ctx)
7858 1 : .await
7859 1 : .context("commit init_empty_test_timeline modification")?;
7860 1 :
7861 1 : // Do the flush. The flush code will check the expectations that we set above.
7862 1 : tline.freeze_and_flush().await?;
7863 1 :
7864 1 : // assert freeze_and_flush exercised the initdb optimization
7865 1 : {
7866 1 : let state = tline.flush_loop_state.lock().unwrap();
7867 1 : let timeline::FlushLoopState::Running {
7868 1 : expect_initdb_optimization,
7869 1 : initdb_optimization_count,
7870 1 : } = *state
7871 1 : else {
7872 1 : panic!("unexpected state: {:?}", *state);
7873 1 : };
7874 1 : assert!(expect_initdb_optimization);
7875 1 : assert!(initdb_optimization_count > 0);
7876 1 : }
7877 1 : Ok(())
7878 1 : }
7879 :
7880 : #[tokio::test]
7881 1 : async fn test_create_guard_crash() -> anyhow::Result<()> {
7882 1 : let name = "test_create_guard_crash";
7883 1 : let harness = TenantHarness::create(name).await?;
7884 1 : {
7885 1 : let (tenant, ctx) = harness.load().await;
7886 1 : let (tline, _ctx) = tenant
7887 1 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
7888 1 : .await?;
7889 1 : // Leave the timeline ID in [`TenantShard::timelines_creating`] to exclude attempting to create it again
7890 1 : let raw_tline = tline.raw_timeline().unwrap();
7891 1 : raw_tline
7892 1 : .shutdown(super::timeline::ShutdownMode::Hard)
7893 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))
7894 1 : .await;
7895 1 : std::mem::forget(tline);
7896 1 : }
7897 1 :
7898 1 : let (tenant, _) = harness.load().await;
7899 1 : match tenant.get_timeline(TIMELINE_ID, false) {
7900 1 : Ok(_) => panic!("timeline should've been removed during load"),
7901 1 : Err(e) => {
7902 1 : assert_eq!(
7903 1 : e,
7904 1 : GetTimelineError::NotFound {
7905 1 : tenant_id: tenant.tenant_shard_id,
7906 1 : timeline_id: TIMELINE_ID,
7907 1 : }
7908 1 : )
7909 1 : }
7910 1 : }
7911 1 :
7912 1 : assert!(
7913 1 : !harness
7914 1 : .conf
7915 1 : .timeline_path(&tenant.tenant_shard_id, &TIMELINE_ID)
7916 1 : .exists()
7917 1 : );
7918 1 :
7919 1 : Ok(())
7920 1 : }
7921 :
7922 : #[tokio::test]
7923 1 : async fn test_read_at_max_lsn() -> anyhow::Result<()> {
7924 1 : let names_algorithms = [
7925 1 : ("test_read_at_max_lsn_legacy", CompactionAlgorithm::Legacy),
7926 1 : ("test_read_at_max_lsn_tiered", CompactionAlgorithm::Tiered),
7927 1 : ];
7928 3 : for (name, algorithm) in names_algorithms {
7929 2 : test_read_at_max_lsn_algorithm(name, algorithm).await?;
7930 1 : }
7931 1 : Ok(())
7932 1 : }
7933 :
7934 2 : async fn test_read_at_max_lsn_algorithm(
7935 2 : name: &'static str,
7936 2 : compaction_algorithm: CompactionAlgorithm,
7937 2 : ) -> anyhow::Result<()> {
7938 2 : let mut harness = TenantHarness::create(name).await?;
7939 2 : harness.tenant_conf.compaction_algorithm = Some(CompactionAlgorithmSettings {
7940 2 : kind: compaction_algorithm,
7941 2 : });
7942 2 : let (tenant, ctx) = harness.load().await;
7943 2 : let tline = tenant
7944 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
7945 2 : .await?;
7946 :
7947 2 : let lsn = Lsn(0x10);
7948 2 : let compact = false;
7949 2 : bulk_insert_maybe_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000, compact).await?;
7950 :
7951 2 : let test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7952 2 : let read_lsn = Lsn(u64::MAX - 1);
7953 :
7954 2 : let result = tline.get(test_key, read_lsn, &ctx).await;
7955 2 : assert!(result.is_ok(), "result is not Ok: {}", result.unwrap_err());
7956 :
7957 2 : Ok(())
7958 2 : }
7959 :
7960 : #[tokio::test]
7961 1 : async fn test_metadata_scan() -> anyhow::Result<()> {
7962 1 : let harness = TenantHarness::create("test_metadata_scan").await?;
7963 1 : let (tenant, ctx) = harness.load().await;
7964 1 : let io_concurrency = IoConcurrency::spawn_for_test();
7965 1 : let tline = tenant
7966 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7967 1 : .await?;
7968 1 :
7969 1 : const NUM_KEYS: usize = 1000;
7970 1 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
7971 1 :
7972 1 : let cancel = CancellationToken::new();
7973 1 :
7974 1 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7975 1 : base_key.field1 = AUX_KEY_PREFIX;
7976 1 : let mut test_key = base_key;
7977 1 :
7978 1 : // Track when each page was last modified. Used to assert that
7979 1 : // a read sees the latest page version.
7980 1 : let mut updated = [Lsn(0); NUM_KEYS];
7981 1 :
7982 1 : let mut lsn = Lsn(0x10);
7983 1 : #[allow(clippy::needless_range_loop)]
7984 1001 : for blknum in 0..NUM_KEYS {
7985 1000 : lsn = Lsn(lsn.0 + 0x10);
7986 1000 : test_key.field6 = (blknum * STEP) as u32;
7987 1000 : let mut writer = tline.writer().await;
7988 1000 : writer
7989 1000 : .put(
7990 1000 : test_key,
7991 1000 : lsn,
7992 1000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7993 1000 : &ctx,
7994 1000 : )
7995 1000 : .await?;
7996 1000 : writer.finish_write(lsn);
7997 1000 : updated[blknum] = lsn;
7998 1000 : drop(writer);
7999 1 : }
8000 1 :
8001 1 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
8002 1 :
8003 12 : for iter in 0..=10 {
8004 1 : // Read all the blocks
8005 11000 : for (blknum, last_lsn) in updated.iter().enumerate() {
8006 11000 : test_key.field6 = (blknum * STEP) as u32;
8007 11000 : assert_eq!(
8008 11000 : tline.get(test_key, lsn, &ctx).await?,
8009 11000 : test_img(&format!("{} at {}", blknum, last_lsn))
8010 1 : );
8011 1 : }
8012 1 :
8013 11 : let mut cnt = 0;
8014 11 : let query = VersionedKeySpaceQuery::uniform(keyspace.clone(), lsn);
8015 1 :
8016 11000 : for (key, value) in tline
8017 11 : .get_vectored_impl(
8018 11 : query,
8019 11 : &mut ValuesReconstructState::new(io_concurrency.clone()),
8020 11 : &ctx,
8021 11 : )
8022 11 : .await?
8023 1 : {
8024 11000 : let blknum = key.field6 as usize;
8025 11000 : let value = value?;
8026 11000 : assert!(blknum % STEP == 0);
8027 11000 : let blknum = blknum / STEP;
8028 11000 : assert_eq!(
8029 11000 : value,
8030 11000 : test_img(&format!("{} at {}", blknum, updated[blknum]))
8031 11000 : );
8032 11000 : cnt += 1;
8033 1 : }
8034 1 :
8035 11 : assert_eq!(cnt, NUM_KEYS);
8036 1 :
8037 11011 : for _ in 0..NUM_KEYS {
8038 11000 : lsn = Lsn(lsn.0 + 0x10);
8039 11000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
8040 11000 : test_key.field6 = (blknum * STEP) as u32;
8041 11000 : let mut writer = tline.writer().await;
8042 11000 : writer
8043 11000 : .put(
8044 11000 : test_key,
8045 11000 : lsn,
8046 11000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
8047 11000 : &ctx,
8048 11000 : )
8049 11000 : .await?;
8050 11000 : writer.finish_write(lsn);
8051 11000 : drop(writer);
8052 11000 : updated[blknum] = lsn;
8053 1 : }
8054 1 :
8055 1 : // Perform two cycles of flush, compact, and GC
8056 33 : for round in 0..2 {
8057 22 : tline.freeze_and_flush().await?;
8058 22 : tline
8059 22 : .compact(
8060 22 : &cancel,
8061 22 : if iter % 5 == 0 && round == 0 {
8062 3 : let mut flags = EnumSet::new();
8063 3 : flags.insert(CompactFlags::ForceImageLayerCreation);
8064 3 : flags.insert(CompactFlags::ForceRepartition);
8065 3 : flags
8066 1 : } else {
8067 19 : EnumSet::empty()
8068 1 : },
8069 22 : &ctx,
8070 22 : )
8071 22 : .await?;
8072 22 : tenant
8073 22 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
8074 22 : .await?;
8075 1 : }
8076 1 : }
8077 1 :
8078 1 : Ok(())
8079 1 : }
8080 :
8081 : #[tokio::test]
8082 1 : async fn test_metadata_compaction_trigger() -> anyhow::Result<()> {
8083 1 : let harness = TenantHarness::create("test_metadata_compaction_trigger").await?;
8084 1 : let (tenant, ctx) = harness.load().await;
8085 1 : let tline = tenant
8086 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
8087 1 : .await?;
8088 1 :
8089 1 : let cancel = CancellationToken::new();
8090 1 :
8091 1 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
8092 1 : base_key.field1 = AUX_KEY_PREFIX;
8093 1 : let test_key = base_key;
8094 1 : let mut lsn = Lsn(0x10);
8095 1 :
8096 21 : for _ in 0..20 {
8097 20 : lsn = Lsn(lsn.0 + 0x10);
8098 20 : let mut writer = tline.writer().await;
8099 20 : writer
8100 20 : .put(
8101 20 : test_key,
8102 20 : lsn,
8103 20 : &Value::Image(test_img(&format!("{} at {}", 0, lsn))),
8104 20 : &ctx,
8105 20 : )
8106 20 : .await?;
8107 20 : writer.finish_write(lsn);
8108 20 : drop(writer);
8109 20 : tline.freeze_and_flush().await?; // force create a delta layer
8110 1 : }
8111 1 :
8112 1 : let before_num_l0_delta_files =
8113 1 : tline.layers.read().await.layer_map()?.level0_deltas().len();
8114 1 :
8115 1 : tline.compact(&cancel, EnumSet::default(), &ctx).await?;
8116 1 :
8117 1 : let after_num_l0_delta_files = tline.layers.read().await.layer_map()?.level0_deltas().len();
8118 1 :
8119 1 : assert!(
8120 1 : after_num_l0_delta_files < before_num_l0_delta_files,
8121 1 : "after_num_l0_delta_files={after_num_l0_delta_files}, before_num_l0_delta_files={before_num_l0_delta_files}"
8122 1 : );
8123 1 :
8124 1 : assert_eq!(
8125 1 : tline.get(test_key, lsn, &ctx).await?,
8126 1 : test_img(&format!("{} at {}", 0, lsn))
8127 1 : );
8128 1 :
8129 1 : Ok(())
8130 1 : }
8131 :
8132 : #[tokio::test]
8133 1 : async fn test_aux_file_e2e() {
8134 1 : let harness = TenantHarness::create("test_aux_file_e2e").await.unwrap();
8135 1 :
8136 1 : let (tenant, ctx) = harness.load().await;
8137 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8138 1 :
8139 1 : let mut lsn = Lsn(0x08);
8140 1 :
8141 1 : let tline: Arc<Timeline> = tenant
8142 1 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
8143 1 : .await
8144 1 : .unwrap();
8145 1 :
8146 1 : {
8147 1 : lsn += 8;
8148 1 : let mut modification = tline.begin_modification(lsn);
8149 1 : modification
8150 1 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
8151 1 : .await
8152 1 : .unwrap();
8153 1 : modification.commit(&ctx).await.unwrap();
8154 1 : }
8155 1 :
8156 1 : // we can read everything from the storage
8157 1 : let files = tline
8158 1 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
8159 1 : .await
8160 1 : .unwrap();
8161 1 : assert_eq!(
8162 1 : files.get("pg_logical/mappings/test1"),
8163 1 : Some(&bytes::Bytes::from_static(b"first"))
8164 1 : );
8165 1 :
8166 1 : {
8167 1 : lsn += 8;
8168 1 : let mut modification = tline.begin_modification(lsn);
8169 1 : modification
8170 1 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
8171 1 : .await
8172 1 : .unwrap();
8173 1 : modification.commit(&ctx).await.unwrap();
8174 1 : }
8175 1 :
8176 1 : let files = tline
8177 1 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
8178 1 : .await
8179 1 : .unwrap();
8180 1 : assert_eq!(
8181 1 : files.get("pg_logical/mappings/test2"),
8182 1 : Some(&bytes::Bytes::from_static(b"second"))
8183 1 : );
8184 1 :
8185 1 : let child = tenant
8186 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(lsn), &ctx)
8187 1 : .await
8188 1 : .unwrap();
8189 1 :
8190 1 : let files = child
8191 1 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
8192 1 : .await
8193 1 : .unwrap();
8194 1 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
8195 1 : assert_eq!(files.get("pg_logical/mappings/test2"), None);
8196 1 : }
8197 :
8198 : #[tokio::test]
8199 1 : async fn test_repl_origin_tombstones() {
8200 1 : let harness = TenantHarness::create("test_repl_origin_tombstones")
8201 1 : .await
8202 1 : .unwrap();
8203 1 :
8204 1 : let (tenant, ctx) = harness.load().await;
8205 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8206 1 :
8207 1 : let mut lsn = Lsn(0x08);
8208 1 :
8209 1 : let tline: Arc<Timeline> = tenant
8210 1 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
8211 1 : .await
8212 1 : .unwrap();
8213 1 :
8214 1 : let repl_lsn = Lsn(0x10);
8215 1 : {
8216 1 : lsn += 8;
8217 1 : let mut modification = tline.begin_modification(lsn);
8218 1 : modification.put_for_unit_test(repl_origin_key(2), Value::Image(Bytes::new()));
8219 1 : modification.set_replorigin(1, repl_lsn).await.unwrap();
8220 1 : modification.commit(&ctx).await.unwrap();
8221 1 : }
8222 1 :
8223 1 : // we can read everything from the storage
8224 1 : let repl_origins = tline
8225 1 : .get_replorigins(lsn, &ctx, io_concurrency.clone())
8226 1 : .await
8227 1 : .unwrap();
8228 1 : assert_eq!(repl_origins.len(), 1);
8229 1 : assert_eq!(repl_origins[&1], lsn);
8230 1 :
8231 1 : {
8232 1 : lsn += 8;
8233 1 : let mut modification = tline.begin_modification(lsn);
8234 1 : modification.put_for_unit_test(
8235 1 : repl_origin_key(3),
8236 1 : Value::Image(Bytes::copy_from_slice(b"cannot_decode_this")),
8237 1 : );
8238 1 : modification.commit(&ctx).await.unwrap();
8239 1 : }
8240 1 : let result = tline
8241 1 : .get_replorigins(lsn, &ctx, io_concurrency.clone())
8242 1 : .await;
8243 1 : assert!(result.is_err());
8244 1 : }
8245 :
8246 : #[tokio::test]
8247 1 : async fn test_metadata_image_creation() -> anyhow::Result<()> {
8248 1 : let harness = TenantHarness::create("test_metadata_image_creation").await?;
8249 1 : let (tenant, ctx) = harness.load().await;
8250 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8251 1 : let tline = tenant
8252 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
8253 1 : .await?;
8254 1 :
8255 1 : const NUM_KEYS: usize = 1000;
8256 1 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
8257 1 :
8258 1 : let cancel = CancellationToken::new();
8259 1 :
8260 1 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8261 1 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
8262 1 : let mut test_key = base_key;
8263 1 : let mut lsn = Lsn(0x10);
8264 1 :
8265 4 : async fn scan_with_statistics(
8266 4 : tline: &Timeline,
8267 4 : keyspace: &KeySpace,
8268 4 : lsn: Lsn,
8269 4 : ctx: &RequestContext,
8270 4 : io_concurrency: IoConcurrency,
8271 4 : ) -> anyhow::Result<(BTreeMap<Key, Result<Bytes, PageReconstructError>>, usize)> {
8272 4 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
8273 4 : let query = VersionedKeySpaceQuery::uniform(keyspace.clone(), lsn);
8274 4 : let res = tline
8275 4 : .get_vectored_impl(query, &mut reconstruct_state, ctx)
8276 4 : .await?;
8277 4 : Ok((res, reconstruct_state.get_delta_layers_visited() as usize))
8278 4 : }
8279 1 :
8280 1001 : for blknum in 0..NUM_KEYS {
8281 1000 : lsn = Lsn(lsn.0 + 0x10);
8282 1000 : test_key.field6 = (blknum * STEP) as u32;
8283 1000 : let mut writer = tline.writer().await;
8284 1000 : writer
8285 1000 : .put(
8286 1000 : test_key,
8287 1000 : lsn,
8288 1000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
8289 1000 : &ctx,
8290 1000 : )
8291 1000 : .await?;
8292 1000 : writer.finish_write(lsn);
8293 1000 : drop(writer);
8294 1 : }
8295 1 :
8296 1 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
8297 1 :
8298 11 : for iter in 1..=10 {
8299 10010 : for _ in 0..NUM_KEYS {
8300 10000 : lsn = Lsn(lsn.0 + 0x10);
8301 10000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
8302 10000 : test_key.field6 = (blknum * STEP) as u32;
8303 10000 : let mut writer = tline.writer().await;
8304 10000 : writer
8305 10000 : .put(
8306 10000 : test_key,
8307 10000 : lsn,
8308 10000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
8309 10000 : &ctx,
8310 10000 : )
8311 10000 : .await?;
8312 10000 : writer.finish_write(lsn);
8313 10000 : drop(writer);
8314 1 : }
8315 1 :
8316 10 : tline.freeze_and_flush().await?;
8317 1 :
8318 10 : if iter % 5 == 0 {
8319 2 : let (_, before_delta_file_accessed) =
8320 2 : scan_with_statistics(&tline, &keyspace, lsn, &ctx, io_concurrency.clone())
8321 2 : .await?;
8322 2 : tline
8323 2 : .compact(
8324 2 : &cancel,
8325 2 : {
8326 2 : let mut flags = EnumSet::new();
8327 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
8328 2 : flags.insert(CompactFlags::ForceRepartition);
8329 2 : flags
8330 2 : },
8331 2 : &ctx,
8332 2 : )
8333 2 : .await?;
8334 2 : let (_, after_delta_file_accessed) =
8335 2 : scan_with_statistics(&tline, &keyspace, lsn, &ctx, io_concurrency.clone())
8336 2 : .await?;
8337 2 : assert!(
8338 2 : after_delta_file_accessed < before_delta_file_accessed,
8339 1 : "after_delta_file_accessed={after_delta_file_accessed}, before_delta_file_accessed={before_delta_file_accessed}"
8340 1 : );
8341 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.
8342 2 : assert!(
8343 2 : after_delta_file_accessed <= 2,
8344 1 : "after_delta_file_accessed={after_delta_file_accessed}"
8345 1 : );
8346 8 : }
8347 1 : }
8348 1 :
8349 1 : Ok(())
8350 1 : }
8351 :
8352 : #[tokio::test]
8353 1 : async fn test_vectored_missing_data_key_reads() -> anyhow::Result<()> {
8354 1 : let harness = TenantHarness::create("test_vectored_missing_data_key_reads").await?;
8355 1 : let (tenant, ctx) = harness.load().await;
8356 1 :
8357 1 : let base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
8358 1 : let base_key_child = Key::from_hex("000000000033333333444444445500000001").unwrap();
8359 1 : let base_key_nonexist = Key::from_hex("000000000033333333444444445500000002").unwrap();
8360 1 :
8361 1 : let tline = tenant
8362 1 : .create_test_timeline_with_layers(
8363 1 : TIMELINE_ID,
8364 1 : Lsn(0x10),
8365 1 : DEFAULT_PG_VERSION,
8366 1 : &ctx,
8367 1 : Vec::new(), // in-memory layers
8368 1 : Vec::new(), // delta layers
8369 1 : vec![(Lsn(0x20), vec![(base_key, test_img("data key 1"))])], // image layers
8370 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
8371 1 : )
8372 1 : .await?;
8373 1 : tline.add_extra_test_dense_keyspace(KeySpace::single(base_key..(base_key_nonexist.next())));
8374 1 :
8375 1 : let child = tenant
8376 1 : .branch_timeline_test_with_layers(
8377 1 : &tline,
8378 1 : NEW_TIMELINE_ID,
8379 1 : Some(Lsn(0x20)),
8380 1 : &ctx,
8381 1 : Vec::new(), // delta layers
8382 1 : vec![(Lsn(0x30), vec![(base_key_child, test_img("data key 2"))])], // image layers
8383 1 : Lsn(0x30),
8384 1 : )
8385 1 : .await
8386 1 : .unwrap();
8387 1 :
8388 1 : let lsn = Lsn(0x30);
8389 1 :
8390 1 : // test vectored get on parent timeline
8391 1 : assert_eq!(
8392 1 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
8393 1 : Some(test_img("data key 1"))
8394 1 : );
8395 1 : assert!(
8396 1 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx)
8397 1 : .await
8398 1 : .unwrap_err()
8399 1 : .is_missing_key_error()
8400 1 : );
8401 1 : assert!(
8402 1 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx)
8403 1 : .await
8404 1 : .unwrap_err()
8405 1 : .is_missing_key_error()
8406 1 : );
8407 1 :
8408 1 : // test vectored get on child timeline
8409 1 : assert_eq!(
8410 1 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
8411 1 : Some(test_img("data key 1"))
8412 1 : );
8413 1 : assert_eq!(
8414 1 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
8415 1 : Some(test_img("data key 2"))
8416 1 : );
8417 1 : assert!(
8418 1 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx)
8419 1 : .await
8420 1 : .unwrap_err()
8421 1 : .is_missing_key_error()
8422 1 : );
8423 1 :
8424 1 : Ok(())
8425 1 : }
8426 :
8427 : #[tokio::test]
8428 1 : async fn test_vectored_missing_metadata_key_reads() -> anyhow::Result<()> {
8429 1 : let harness = TenantHarness::create("test_vectored_missing_metadata_key_reads").await?;
8430 1 : let (tenant, ctx) = harness.load().await;
8431 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8432 1 :
8433 1 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8434 1 : let base_key_child = Key::from_hex("620000000033333333444444445500000001").unwrap();
8435 1 : let base_key_nonexist = Key::from_hex("620000000033333333444444445500000002").unwrap();
8436 1 : let base_key_overwrite = Key::from_hex("620000000033333333444444445500000003").unwrap();
8437 1 :
8438 1 : let base_inherited_key = Key::from_hex("610000000033333333444444445500000000").unwrap();
8439 1 : let base_inherited_key_child =
8440 1 : Key::from_hex("610000000033333333444444445500000001").unwrap();
8441 1 : let base_inherited_key_nonexist =
8442 1 : Key::from_hex("610000000033333333444444445500000002").unwrap();
8443 1 : let base_inherited_key_overwrite =
8444 1 : Key::from_hex("610000000033333333444444445500000003").unwrap();
8445 1 :
8446 1 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
8447 1 : assert_eq!(base_inherited_key.field1, RELATION_SIZE_PREFIX);
8448 1 :
8449 1 : let tline = tenant
8450 1 : .create_test_timeline_with_layers(
8451 1 : TIMELINE_ID,
8452 1 : Lsn(0x10),
8453 1 : DEFAULT_PG_VERSION,
8454 1 : &ctx,
8455 1 : Vec::new(), // in-memory layers
8456 1 : Vec::new(), // delta layers
8457 1 : vec![(
8458 1 : Lsn(0x20),
8459 1 : vec![
8460 1 : (base_inherited_key, test_img("metadata inherited key 1")),
8461 1 : (
8462 1 : base_inherited_key_overwrite,
8463 1 : test_img("metadata key overwrite 1a"),
8464 1 : ),
8465 1 : (base_key, test_img("metadata key 1")),
8466 1 : (base_key_overwrite, test_img("metadata key overwrite 1b")),
8467 1 : ],
8468 1 : )], // image layers
8469 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
8470 1 : )
8471 1 : .await?;
8472 1 :
8473 1 : let child = tenant
8474 1 : .branch_timeline_test_with_layers(
8475 1 : &tline,
8476 1 : NEW_TIMELINE_ID,
8477 1 : Some(Lsn(0x20)),
8478 1 : &ctx,
8479 1 : Vec::new(), // delta layers
8480 1 : vec![(
8481 1 : Lsn(0x30),
8482 1 : vec![
8483 1 : (
8484 1 : base_inherited_key_child,
8485 1 : test_img("metadata inherited key 2"),
8486 1 : ),
8487 1 : (
8488 1 : base_inherited_key_overwrite,
8489 1 : test_img("metadata key overwrite 2a"),
8490 1 : ),
8491 1 : (base_key_child, test_img("metadata key 2")),
8492 1 : (base_key_overwrite, test_img("metadata key overwrite 2b")),
8493 1 : ],
8494 1 : )], // image layers
8495 1 : Lsn(0x30),
8496 1 : )
8497 1 : .await
8498 1 : .unwrap();
8499 1 :
8500 1 : let lsn = Lsn(0x30);
8501 1 :
8502 1 : // test vectored get on parent timeline
8503 1 : assert_eq!(
8504 1 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
8505 1 : Some(test_img("metadata key 1"))
8506 1 : );
8507 1 : assert_eq!(
8508 1 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx).await?,
8509 1 : None
8510 1 : );
8511 1 : assert_eq!(
8512 1 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx).await?,
8513 1 : None
8514 1 : );
8515 1 : assert_eq!(
8516 1 : get_vectored_impl_wrapper(&tline, base_key_overwrite, lsn, &ctx).await?,
8517 1 : Some(test_img("metadata key overwrite 1b"))
8518 1 : );
8519 1 : assert_eq!(
8520 1 : get_vectored_impl_wrapper(&tline, base_inherited_key, lsn, &ctx).await?,
8521 1 : Some(test_img("metadata inherited key 1"))
8522 1 : );
8523 1 : assert_eq!(
8524 1 : get_vectored_impl_wrapper(&tline, base_inherited_key_child, lsn, &ctx).await?,
8525 1 : None
8526 1 : );
8527 1 : assert_eq!(
8528 1 : get_vectored_impl_wrapper(&tline, base_inherited_key_nonexist, lsn, &ctx).await?,
8529 1 : None
8530 1 : );
8531 1 : assert_eq!(
8532 1 : get_vectored_impl_wrapper(&tline, base_inherited_key_overwrite, lsn, &ctx).await?,
8533 1 : Some(test_img("metadata key overwrite 1a"))
8534 1 : );
8535 1 :
8536 1 : // test vectored get on child timeline
8537 1 : assert_eq!(
8538 1 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
8539 1 : None
8540 1 : );
8541 1 : assert_eq!(
8542 1 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
8543 1 : Some(test_img("metadata key 2"))
8544 1 : );
8545 1 : assert_eq!(
8546 1 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx).await?,
8547 1 : None
8548 1 : );
8549 1 : assert_eq!(
8550 1 : get_vectored_impl_wrapper(&child, base_inherited_key, lsn, &ctx).await?,
8551 1 : Some(test_img("metadata inherited key 1"))
8552 1 : );
8553 1 : assert_eq!(
8554 1 : get_vectored_impl_wrapper(&child, base_inherited_key_child, lsn, &ctx).await?,
8555 1 : Some(test_img("metadata inherited key 2"))
8556 1 : );
8557 1 : assert_eq!(
8558 1 : get_vectored_impl_wrapper(&child, base_inherited_key_nonexist, lsn, &ctx).await?,
8559 1 : None
8560 1 : );
8561 1 : assert_eq!(
8562 1 : get_vectored_impl_wrapper(&child, base_key_overwrite, lsn, &ctx).await?,
8563 1 : Some(test_img("metadata key overwrite 2b"))
8564 1 : );
8565 1 : assert_eq!(
8566 1 : get_vectored_impl_wrapper(&child, base_inherited_key_overwrite, lsn, &ctx).await?,
8567 1 : Some(test_img("metadata key overwrite 2a"))
8568 1 : );
8569 1 :
8570 1 : // test vectored scan on parent timeline
8571 1 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency.clone());
8572 1 : let query =
8573 1 : VersionedKeySpaceQuery::uniform(KeySpace::single(Key::metadata_key_range()), lsn);
8574 1 : let res = tline
8575 1 : .get_vectored_impl(query, &mut reconstruct_state, &ctx)
8576 1 : .await?;
8577 1 :
8578 1 : assert_eq!(
8579 1 : res.into_iter()
8580 4 : .map(|(k, v)| (k, v.unwrap()))
8581 1 : .collect::<Vec<_>>(),
8582 1 : vec![
8583 1 : (base_inherited_key, test_img("metadata inherited key 1")),
8584 1 : (
8585 1 : base_inherited_key_overwrite,
8586 1 : test_img("metadata key overwrite 1a")
8587 1 : ),
8588 1 : (base_key, test_img("metadata key 1")),
8589 1 : (base_key_overwrite, test_img("metadata key overwrite 1b")),
8590 1 : ]
8591 1 : );
8592 1 :
8593 1 : // test vectored scan on child timeline
8594 1 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency.clone());
8595 1 : let query =
8596 1 : VersionedKeySpaceQuery::uniform(KeySpace::single(Key::metadata_key_range()), lsn);
8597 1 : let res = child
8598 1 : .get_vectored_impl(query, &mut reconstruct_state, &ctx)
8599 1 : .await?;
8600 1 :
8601 1 : assert_eq!(
8602 1 : res.into_iter()
8603 5 : .map(|(k, v)| (k, v.unwrap()))
8604 1 : .collect::<Vec<_>>(),
8605 1 : vec![
8606 1 : (base_inherited_key, test_img("metadata inherited key 1")),
8607 1 : (
8608 1 : base_inherited_key_child,
8609 1 : test_img("metadata inherited key 2")
8610 1 : ),
8611 1 : (
8612 1 : base_inherited_key_overwrite,
8613 1 : test_img("metadata key overwrite 2a")
8614 1 : ),
8615 1 : (base_key_child, test_img("metadata key 2")),
8616 1 : (base_key_overwrite, test_img("metadata key overwrite 2b")),
8617 1 : ]
8618 1 : );
8619 1 :
8620 1 : Ok(())
8621 1 : }
8622 :
8623 28 : async fn get_vectored_impl_wrapper(
8624 28 : tline: &Arc<Timeline>,
8625 28 : key: Key,
8626 28 : lsn: Lsn,
8627 28 : ctx: &RequestContext,
8628 28 : ) -> Result<Option<Bytes>, GetVectoredError> {
8629 28 : let io_concurrency = IoConcurrency::spawn_from_conf(
8630 28 : tline.conf.get_vectored_concurrent_io,
8631 28 : tline.gate.enter().unwrap(),
8632 28 : );
8633 28 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
8634 28 : let query = VersionedKeySpaceQuery::uniform(KeySpace::single(key..key.next()), lsn);
8635 28 : let mut res = tline
8636 28 : .get_vectored_impl(query, &mut reconstruct_state, ctx)
8637 28 : .await?;
8638 25 : Ok(res.pop_last().map(|(k, v)| {
8639 16 : assert_eq!(k, key);
8640 16 : v.unwrap()
8641 25 : }))
8642 28 : }
8643 :
8644 : #[tokio::test]
8645 1 : async fn test_metadata_tombstone_reads() -> anyhow::Result<()> {
8646 1 : let harness = TenantHarness::create("test_metadata_tombstone_reads").await?;
8647 1 : let (tenant, ctx) = harness.load().await;
8648 1 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
8649 1 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8650 1 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8651 1 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
8652 1 :
8653 1 : // We emulate the situation that the compaction algorithm creates an image layer that removes the tombstones
8654 1 : // Lsn 0x30 key0, key3, no key1+key2
8655 1 : // Lsn 0x20 key1+key2 tomestones
8656 1 : // Lsn 0x10 key1 in image, key2 in delta
8657 1 : let tline = tenant
8658 1 : .create_test_timeline_with_layers(
8659 1 : TIMELINE_ID,
8660 1 : Lsn(0x10),
8661 1 : DEFAULT_PG_VERSION,
8662 1 : &ctx,
8663 1 : Vec::new(), // in-memory layers
8664 1 : // delta layers
8665 1 : vec![
8666 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8667 1 : Lsn(0x10)..Lsn(0x20),
8668 1 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8669 1 : ),
8670 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8671 1 : Lsn(0x20)..Lsn(0x30),
8672 1 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8673 1 : ),
8674 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8675 1 : Lsn(0x20)..Lsn(0x30),
8676 1 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8677 1 : ),
8678 1 : ],
8679 1 : // image layers
8680 1 : vec![
8681 1 : (Lsn(0x10), vec![(key1, test_img("metadata key 1"))]),
8682 1 : (
8683 1 : Lsn(0x30),
8684 1 : vec![
8685 1 : (key0, test_img("metadata key 0")),
8686 1 : (key3, test_img("metadata key 3")),
8687 1 : ],
8688 1 : ),
8689 1 : ],
8690 1 : Lsn(0x30),
8691 1 : )
8692 1 : .await?;
8693 1 :
8694 1 : let lsn = Lsn(0x30);
8695 1 : let old_lsn = Lsn(0x20);
8696 1 :
8697 1 : assert_eq!(
8698 1 : get_vectored_impl_wrapper(&tline, key0, lsn, &ctx).await?,
8699 1 : Some(test_img("metadata key 0"))
8700 1 : );
8701 1 : assert_eq!(
8702 1 : get_vectored_impl_wrapper(&tline, key1, lsn, &ctx).await?,
8703 1 : None,
8704 1 : );
8705 1 : assert_eq!(
8706 1 : get_vectored_impl_wrapper(&tline, key2, lsn, &ctx).await?,
8707 1 : None,
8708 1 : );
8709 1 : assert_eq!(
8710 1 : get_vectored_impl_wrapper(&tline, key1, old_lsn, &ctx).await?,
8711 1 : Some(Bytes::new()),
8712 1 : );
8713 1 : assert_eq!(
8714 1 : get_vectored_impl_wrapper(&tline, key2, old_lsn, &ctx).await?,
8715 1 : Some(Bytes::new()),
8716 1 : );
8717 1 : assert_eq!(
8718 1 : get_vectored_impl_wrapper(&tline, key3, lsn, &ctx).await?,
8719 1 : Some(test_img("metadata key 3"))
8720 1 : );
8721 1 :
8722 1 : Ok(())
8723 1 : }
8724 :
8725 : #[tokio::test]
8726 1 : async fn test_metadata_tombstone_image_creation() {
8727 1 : let harness = TenantHarness::create("test_metadata_tombstone_image_creation")
8728 1 : .await
8729 1 : .unwrap();
8730 1 : let (tenant, ctx) = harness.load().await;
8731 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8732 1 :
8733 1 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
8734 1 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8735 1 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8736 1 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
8737 1 :
8738 1 : let tline = tenant
8739 1 : .create_test_timeline_with_layers(
8740 1 : TIMELINE_ID,
8741 1 : Lsn(0x10),
8742 1 : DEFAULT_PG_VERSION,
8743 1 : &ctx,
8744 1 : Vec::new(), // in-memory layers
8745 1 : // delta layers
8746 1 : vec![
8747 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8748 1 : Lsn(0x10)..Lsn(0x20),
8749 1 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8750 1 : ),
8751 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8752 1 : Lsn(0x20)..Lsn(0x30),
8753 1 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8754 1 : ),
8755 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8756 1 : Lsn(0x20)..Lsn(0x30),
8757 1 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8758 1 : ),
8759 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8760 1 : Lsn(0x30)..Lsn(0x40),
8761 1 : vec![
8762 1 : (key0, Lsn(0x30), Value::Image(test_img("metadata key 0"))),
8763 1 : (key3, Lsn(0x30), Value::Image(test_img("metadata key 3"))),
8764 1 : ],
8765 1 : ),
8766 1 : ],
8767 1 : // image layers
8768 1 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
8769 1 : Lsn(0x40),
8770 1 : )
8771 1 : .await
8772 1 : .unwrap();
8773 1 :
8774 1 : let cancel = CancellationToken::new();
8775 1 :
8776 1 : tline
8777 1 : .compact(
8778 1 : &cancel,
8779 1 : {
8780 1 : let mut flags = EnumSet::new();
8781 1 : flags.insert(CompactFlags::ForceImageLayerCreation);
8782 1 : flags.insert(CompactFlags::ForceRepartition);
8783 1 : flags
8784 1 : },
8785 1 : &ctx,
8786 1 : )
8787 1 : .await
8788 1 : .unwrap();
8789 1 :
8790 1 : // Image layers are created at last_record_lsn
8791 1 : let images = tline
8792 1 : .inspect_image_layers(Lsn(0x40), &ctx, io_concurrency.clone())
8793 1 : .await
8794 1 : .unwrap()
8795 1 : .into_iter()
8796 9 : .filter(|(k, _)| k.is_metadata_key())
8797 1 : .collect::<Vec<_>>();
8798 1 : assert_eq!(images.len(), 2); // the image layer should only contain two existing keys, tombstones should be removed.
8799 1 : }
8800 :
8801 : #[tokio::test]
8802 1 : async fn test_metadata_tombstone_empty_image_creation() {
8803 1 : let harness = TenantHarness::create("test_metadata_tombstone_empty_image_creation")
8804 1 : .await
8805 1 : .unwrap();
8806 1 : let (tenant, ctx) = harness.load().await;
8807 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8808 1 :
8809 1 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8810 1 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8811 1 :
8812 1 : let tline = tenant
8813 1 : .create_test_timeline_with_layers(
8814 1 : TIMELINE_ID,
8815 1 : Lsn(0x10),
8816 1 : DEFAULT_PG_VERSION,
8817 1 : &ctx,
8818 1 : Vec::new(), // in-memory layers
8819 1 : // delta layers
8820 1 : vec![
8821 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8822 1 : Lsn(0x10)..Lsn(0x20),
8823 1 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8824 1 : ),
8825 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8826 1 : Lsn(0x20)..Lsn(0x30),
8827 1 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8828 1 : ),
8829 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8830 1 : Lsn(0x20)..Lsn(0x30),
8831 1 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8832 1 : ),
8833 1 : ],
8834 1 : // image layers
8835 1 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
8836 1 : Lsn(0x30),
8837 1 : )
8838 1 : .await
8839 1 : .unwrap();
8840 1 :
8841 1 : let cancel = CancellationToken::new();
8842 1 :
8843 1 : tline
8844 1 : .compact(
8845 1 : &cancel,
8846 1 : {
8847 1 : let mut flags = EnumSet::new();
8848 1 : flags.insert(CompactFlags::ForceImageLayerCreation);
8849 1 : flags.insert(CompactFlags::ForceRepartition);
8850 1 : flags
8851 1 : },
8852 1 : &ctx,
8853 1 : )
8854 1 : .await
8855 1 : .unwrap();
8856 1 :
8857 1 : // Image layers are created at last_record_lsn
8858 1 : let images = tline
8859 1 : .inspect_image_layers(Lsn(0x30), &ctx, io_concurrency.clone())
8860 1 : .await
8861 1 : .unwrap()
8862 1 : .into_iter()
8863 7 : .filter(|(k, _)| k.is_metadata_key())
8864 1 : .collect::<Vec<_>>();
8865 1 : assert_eq!(images.len(), 0); // the image layer should not contain tombstones, or it is not created
8866 1 : }
8867 :
8868 : #[tokio::test]
8869 1 : async fn test_simple_bottom_most_compaction_images() -> anyhow::Result<()> {
8870 1 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_images").await?;
8871 1 : let (tenant, ctx) = harness.load().await;
8872 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8873 1 :
8874 51 : fn get_key(id: u32) -> Key {
8875 51 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8876 51 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8877 51 : key.field6 = id;
8878 51 : key
8879 51 : }
8880 1 :
8881 1 : // We create
8882 1 : // - one bottom-most image layer,
8883 1 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
8884 1 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
8885 1 : // - a delta layer D3 above the horizon.
8886 1 : //
8887 1 : // | D3 |
8888 1 : // | D1 |
8889 1 : // -| |-- gc horizon -----------------
8890 1 : // | | | D2 |
8891 1 : // --------- img layer ------------------
8892 1 : //
8893 1 : // What we should expact from this compaction is:
8894 1 : // | D3 |
8895 1 : // | Part of D1 |
8896 1 : // --------- img layer with D1+D2 at GC horizon------------------
8897 1 :
8898 1 : // img layer at 0x10
8899 1 : let img_layer = (0..10)
8900 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8901 1 : .collect_vec();
8902 1 :
8903 1 : let delta1 = vec![
8904 1 : (
8905 1 : get_key(1),
8906 1 : Lsn(0x20),
8907 1 : Value::Image(Bytes::from("value 1@0x20")),
8908 1 : ),
8909 1 : (
8910 1 : get_key(2),
8911 1 : Lsn(0x30),
8912 1 : Value::Image(Bytes::from("value 2@0x30")),
8913 1 : ),
8914 1 : (
8915 1 : get_key(3),
8916 1 : Lsn(0x40),
8917 1 : Value::Image(Bytes::from("value 3@0x40")),
8918 1 : ),
8919 1 : ];
8920 1 : let delta2 = vec![
8921 1 : (
8922 1 : get_key(5),
8923 1 : Lsn(0x20),
8924 1 : Value::Image(Bytes::from("value 5@0x20")),
8925 1 : ),
8926 1 : (
8927 1 : get_key(6),
8928 1 : Lsn(0x20),
8929 1 : Value::Image(Bytes::from("value 6@0x20")),
8930 1 : ),
8931 1 : ];
8932 1 : let delta3 = vec![
8933 1 : (
8934 1 : get_key(8),
8935 1 : Lsn(0x48),
8936 1 : Value::Image(Bytes::from("value 8@0x48")),
8937 1 : ),
8938 1 : (
8939 1 : get_key(9),
8940 1 : Lsn(0x48),
8941 1 : Value::Image(Bytes::from("value 9@0x48")),
8942 1 : ),
8943 1 : ];
8944 1 :
8945 1 : let tline = tenant
8946 1 : .create_test_timeline_with_layers(
8947 1 : TIMELINE_ID,
8948 1 : Lsn(0x10),
8949 1 : DEFAULT_PG_VERSION,
8950 1 : &ctx,
8951 1 : Vec::new(), // in-memory layers
8952 1 : vec![
8953 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
8954 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
8955 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
8956 1 : ], // delta layers
8957 1 : vec![(Lsn(0x10), img_layer)], // image layers
8958 1 : Lsn(0x50),
8959 1 : )
8960 1 : .await?;
8961 1 : {
8962 1 : tline
8963 1 : .applied_gc_cutoff_lsn
8964 1 : .lock_for_write()
8965 1 : .store_and_unlock(Lsn(0x30))
8966 1 : .wait()
8967 1 : .await;
8968 1 : // Update GC info
8969 1 : let mut guard = tline.gc_info.write().unwrap();
8970 1 : guard.cutoffs.time = Some(Lsn(0x30));
8971 1 : guard.cutoffs.space = Lsn(0x30);
8972 1 : }
8973 1 :
8974 1 : let expected_result = [
8975 1 : Bytes::from_static(b"value 0@0x10"),
8976 1 : Bytes::from_static(b"value 1@0x20"),
8977 1 : Bytes::from_static(b"value 2@0x30"),
8978 1 : Bytes::from_static(b"value 3@0x40"),
8979 1 : Bytes::from_static(b"value 4@0x10"),
8980 1 : Bytes::from_static(b"value 5@0x20"),
8981 1 : Bytes::from_static(b"value 6@0x20"),
8982 1 : Bytes::from_static(b"value 7@0x10"),
8983 1 : Bytes::from_static(b"value 8@0x48"),
8984 1 : Bytes::from_static(b"value 9@0x48"),
8985 1 : ];
8986 1 :
8987 10 : for (idx, expected) in expected_result.iter().enumerate() {
8988 10 : assert_eq!(
8989 10 : tline
8990 10 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8991 10 : .await
8992 10 : .unwrap(),
8993 1 : expected
8994 1 : );
8995 1 : }
8996 1 :
8997 1 : let cancel = CancellationToken::new();
8998 1 : tline
8999 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9000 1 : .await
9001 1 : .unwrap();
9002 1 :
9003 10 : for (idx, expected) in expected_result.iter().enumerate() {
9004 10 : assert_eq!(
9005 10 : tline
9006 10 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9007 10 : .await
9008 10 : .unwrap(),
9009 1 : expected
9010 1 : );
9011 1 : }
9012 1 :
9013 1 : // Check if the image layer at the GC horizon contains exactly what we want
9014 1 : let image_at_gc_horizon = tline
9015 1 : .inspect_image_layers(Lsn(0x30), &ctx, io_concurrency.clone())
9016 1 : .await
9017 1 : .unwrap()
9018 1 : .into_iter()
9019 17 : .filter(|(k, _)| k.is_metadata_key())
9020 1 : .collect::<Vec<_>>();
9021 1 :
9022 1 : assert_eq!(image_at_gc_horizon.len(), 10);
9023 1 : let expected_result = [
9024 1 : Bytes::from_static(b"value 0@0x10"),
9025 1 : Bytes::from_static(b"value 1@0x20"),
9026 1 : Bytes::from_static(b"value 2@0x30"),
9027 1 : Bytes::from_static(b"value 3@0x10"),
9028 1 : Bytes::from_static(b"value 4@0x10"),
9029 1 : Bytes::from_static(b"value 5@0x20"),
9030 1 : Bytes::from_static(b"value 6@0x20"),
9031 1 : Bytes::from_static(b"value 7@0x10"),
9032 1 : Bytes::from_static(b"value 8@0x10"),
9033 1 : Bytes::from_static(b"value 9@0x10"),
9034 1 : ];
9035 11 : for idx in 0..10 {
9036 10 : assert_eq!(
9037 10 : image_at_gc_horizon[idx],
9038 10 : (get_key(idx as u32), expected_result[idx].clone())
9039 10 : );
9040 1 : }
9041 1 :
9042 1 : // Check if old layers are removed / new layers have the expected LSN
9043 1 : let all_layers = inspect_and_sort(&tline, None).await;
9044 1 : assert_eq!(
9045 1 : all_layers,
9046 1 : vec![
9047 1 : // Image layer at GC horizon
9048 1 : PersistentLayerKey {
9049 1 : key_range: Key::MIN..Key::MAX,
9050 1 : lsn_range: Lsn(0x30)..Lsn(0x31),
9051 1 : is_delta: false
9052 1 : },
9053 1 : // The delta layer below the horizon
9054 1 : PersistentLayerKey {
9055 1 : key_range: get_key(3)..get_key(4),
9056 1 : lsn_range: Lsn(0x30)..Lsn(0x48),
9057 1 : is_delta: true
9058 1 : },
9059 1 : // The delta3 layer that should not be picked for the compaction
9060 1 : PersistentLayerKey {
9061 1 : key_range: get_key(8)..get_key(10),
9062 1 : lsn_range: Lsn(0x48)..Lsn(0x50),
9063 1 : is_delta: true
9064 1 : }
9065 1 : ]
9066 1 : );
9067 1 :
9068 1 : // increase GC horizon and compact again
9069 1 : {
9070 1 : tline
9071 1 : .applied_gc_cutoff_lsn
9072 1 : .lock_for_write()
9073 1 : .store_and_unlock(Lsn(0x40))
9074 1 : .wait()
9075 1 : .await;
9076 1 : // Update GC info
9077 1 : let mut guard = tline.gc_info.write().unwrap();
9078 1 : guard.cutoffs.time = Some(Lsn(0x40));
9079 1 : guard.cutoffs.space = Lsn(0x40);
9080 1 : }
9081 1 : tline
9082 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9083 1 : .await
9084 1 : .unwrap();
9085 1 :
9086 1 : Ok(())
9087 1 : }
9088 :
9089 : #[cfg(feature = "testing")]
9090 : #[tokio::test]
9091 1 : async fn test_neon_test_record() -> anyhow::Result<()> {
9092 1 : let harness = TenantHarness::create("test_neon_test_record").await?;
9093 1 : let (tenant, ctx) = harness.load().await;
9094 1 :
9095 17 : fn get_key(id: u32) -> Key {
9096 17 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9097 17 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9098 17 : key.field6 = id;
9099 17 : key
9100 17 : }
9101 1 :
9102 1 : let delta1 = vec![
9103 1 : (
9104 1 : get_key(1),
9105 1 : Lsn(0x20),
9106 1 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
9107 1 : ),
9108 1 : (
9109 1 : get_key(1),
9110 1 : Lsn(0x30),
9111 1 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
9112 1 : ),
9113 1 : (get_key(2), Lsn(0x10), Value::Image("0x10".into())),
9114 1 : (
9115 1 : get_key(2),
9116 1 : Lsn(0x20),
9117 1 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
9118 1 : ),
9119 1 : (
9120 1 : get_key(2),
9121 1 : Lsn(0x30),
9122 1 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
9123 1 : ),
9124 1 : (get_key(3), Lsn(0x10), Value::Image("0x10".into())),
9125 1 : (
9126 1 : get_key(3),
9127 1 : Lsn(0x20),
9128 1 : Value::WalRecord(NeonWalRecord::wal_clear("c")),
9129 1 : ),
9130 1 : (get_key(4), Lsn(0x10), Value::Image("0x10".into())),
9131 1 : (
9132 1 : get_key(4),
9133 1 : Lsn(0x20),
9134 1 : Value::WalRecord(NeonWalRecord::wal_init("i")),
9135 1 : ),
9136 1 : (
9137 1 : get_key(4),
9138 1 : Lsn(0x30),
9139 1 : Value::WalRecord(NeonWalRecord::wal_append_conditional("j", "i")),
9140 1 : ),
9141 1 : (
9142 1 : get_key(5),
9143 1 : Lsn(0x20),
9144 1 : Value::WalRecord(NeonWalRecord::wal_init("1")),
9145 1 : ),
9146 1 : (
9147 1 : get_key(5),
9148 1 : Lsn(0x30),
9149 1 : Value::WalRecord(NeonWalRecord::wal_append_conditional("j", "2")),
9150 1 : ),
9151 1 : ];
9152 1 : let image1 = vec![(get_key(1), "0x10".into())];
9153 1 :
9154 1 : let tline = tenant
9155 1 : .create_test_timeline_with_layers(
9156 1 : TIMELINE_ID,
9157 1 : Lsn(0x10),
9158 1 : DEFAULT_PG_VERSION,
9159 1 : &ctx,
9160 1 : Vec::new(), // in-memory layers
9161 1 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
9162 1 : Lsn(0x10)..Lsn(0x40),
9163 1 : delta1,
9164 1 : )], // delta layers
9165 1 : vec![(Lsn(0x10), image1)], // image layers
9166 1 : Lsn(0x50),
9167 1 : )
9168 1 : .await?;
9169 1 :
9170 1 : assert_eq!(
9171 1 : tline.get(get_key(1), Lsn(0x50), &ctx).await?,
9172 1 : Bytes::from_static(b"0x10,0x20,0x30")
9173 1 : );
9174 1 : assert_eq!(
9175 1 : tline.get(get_key(2), Lsn(0x50), &ctx).await?,
9176 1 : Bytes::from_static(b"0x10,0x20,0x30")
9177 1 : );
9178 1 :
9179 1 : // Need to remove the limit of "Neon WAL redo requires base image".
9180 1 :
9181 1 : assert_eq!(
9182 1 : tline.get(get_key(3), Lsn(0x50), &ctx).await?,
9183 1 : Bytes::from_static(b"c")
9184 1 : );
9185 1 : assert_eq!(
9186 1 : tline.get(get_key(4), Lsn(0x50), &ctx).await?,
9187 1 : Bytes::from_static(b"ij")
9188 1 : );
9189 1 :
9190 1 : // Manual testing required: currently, read errors will panic the process in debug mode. So we
9191 1 : // cannot enable this assertion in the unit test.
9192 1 : // assert!(tline.get(get_key(5), Lsn(0x50), &ctx).await.is_err());
9193 1 :
9194 1 : Ok(())
9195 1 : }
9196 :
9197 : #[tokio::test(start_paused = true)]
9198 1 : async fn test_lsn_lease() -> anyhow::Result<()> {
9199 1 : let (tenant, ctx) = TenantHarness::create("test_lsn_lease")
9200 1 : .await
9201 1 : .unwrap()
9202 1 : .load()
9203 1 : .await;
9204 1 : // Advance to the lsn lease deadline so that GC is not blocked by
9205 1 : // initial transition into AttachedSingle.
9206 1 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
9207 1 : tokio::time::resume();
9208 1 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
9209 1 :
9210 1 : let end_lsn = Lsn(0x100);
9211 1 : let image_layers = (0x20..=0x90)
9212 1 : .step_by(0x10)
9213 8 : .map(|n| {
9214 8 : (
9215 8 : Lsn(n),
9216 8 : vec![(key, test_img(&format!("data key at {:x}", n)))],
9217 8 : )
9218 8 : })
9219 1 : .collect();
9220 1 :
9221 1 : let timeline = tenant
9222 1 : .create_test_timeline_with_layers(
9223 1 : TIMELINE_ID,
9224 1 : Lsn(0x10),
9225 1 : DEFAULT_PG_VERSION,
9226 1 : &ctx,
9227 1 : Vec::new(), // in-memory layers
9228 1 : Vec::new(),
9229 1 : image_layers,
9230 1 : end_lsn,
9231 1 : )
9232 1 : .await?;
9233 1 :
9234 1 : let leased_lsns = [0x30, 0x50, 0x70];
9235 1 : let mut leases = Vec::new();
9236 3 : leased_lsns.iter().for_each(|n| {
9237 3 : leases.push(
9238 3 : timeline
9239 3 : .init_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)
9240 3 : .expect("lease request should succeed"),
9241 3 : );
9242 3 : });
9243 1 :
9244 1 : let updated_lease_0 = timeline
9245 1 : .renew_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)
9246 1 : .expect("lease renewal should succeed");
9247 1 : assert_eq!(
9248 1 : updated_lease_0.valid_until, leases[0].valid_until,
9249 1 : " Renewing with shorter lease should not change the lease."
9250 1 : );
9251 1 :
9252 1 : let updated_lease_1 = timeline
9253 1 : .renew_lsn_lease(
9254 1 : Lsn(leased_lsns[1]),
9255 1 : timeline.get_lsn_lease_length() * 2,
9256 1 : &ctx,
9257 1 : )
9258 1 : .expect("lease renewal should succeed");
9259 1 : assert!(
9260 1 : updated_lease_1.valid_until > leases[1].valid_until,
9261 1 : "Renewing with a long lease should renew lease with later expiration time."
9262 1 : );
9263 1 :
9264 1 : // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
9265 1 : info!(
9266 1 : "applied_gc_cutoff_lsn: {}",
9267 0 : *timeline.get_applied_gc_cutoff_lsn()
9268 1 : );
9269 1 : timeline.force_set_disk_consistent_lsn(end_lsn);
9270 1 :
9271 1 : let res = tenant
9272 1 : .gc_iteration(
9273 1 : Some(TIMELINE_ID),
9274 1 : 0,
9275 1 : Duration::ZERO,
9276 1 : &CancellationToken::new(),
9277 1 : &ctx,
9278 1 : )
9279 1 : .await
9280 1 : .unwrap();
9281 1 :
9282 1 : // Keeping everything <= Lsn(0x80) b/c leases:
9283 1 : // 0/10: initdb layer
9284 1 : // (0/20..=0/70).step_by(0x10): image layers added when creating the timeline.
9285 1 : assert_eq!(res.layers_needed_by_leases, 7);
9286 1 : // Keeping 0/90 b/c it is the latest layer.
9287 1 : assert_eq!(res.layers_not_updated, 1);
9288 1 : // Removed 0/80.
9289 1 : assert_eq!(res.layers_removed, 1);
9290 1 :
9291 1 : // Make lease on a already GC-ed LSN.
9292 1 : // 0/80 does not have a valid lease + is below latest_gc_cutoff
9293 1 : assert!(Lsn(0x80) < *timeline.get_applied_gc_cutoff_lsn());
9294 1 : timeline
9295 1 : .init_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx)
9296 1 : .expect_err("lease request on GC-ed LSN should fail");
9297 1 :
9298 1 : // Should still be able to renew a currently valid lease
9299 1 : // Assumption: original lease to is still valid for 0/50.
9300 1 : // (use `Timeline::init_lsn_lease` for testing so it always does validation)
9301 1 : timeline
9302 1 : .init_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)
9303 1 : .expect("lease renewal with validation should succeed");
9304 1 :
9305 1 : Ok(())
9306 1 : }
9307 :
9308 : #[cfg(feature = "testing")]
9309 : #[tokio::test]
9310 1 : async fn test_simple_bottom_most_compaction_deltas_1() -> anyhow::Result<()> {
9311 1 : test_simple_bottom_most_compaction_deltas_helper(
9312 1 : "test_simple_bottom_most_compaction_deltas_1",
9313 1 : false,
9314 1 : )
9315 1 : .await
9316 1 : }
9317 :
9318 : #[cfg(feature = "testing")]
9319 : #[tokio::test]
9320 1 : async fn test_simple_bottom_most_compaction_deltas_2() -> anyhow::Result<()> {
9321 1 : test_simple_bottom_most_compaction_deltas_helper(
9322 1 : "test_simple_bottom_most_compaction_deltas_2",
9323 1 : true,
9324 1 : )
9325 1 : .await
9326 1 : }
9327 :
9328 : #[cfg(feature = "testing")]
9329 2 : async fn test_simple_bottom_most_compaction_deltas_helper(
9330 2 : test_name: &'static str,
9331 2 : use_delta_bottom_layer: bool,
9332 2 : ) -> anyhow::Result<()> {
9333 2 : let harness = TenantHarness::create(test_name).await?;
9334 2 : let (tenant, ctx) = harness.load().await;
9335 :
9336 138 : fn get_key(id: u32) -> Key {
9337 138 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9338 138 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9339 138 : key.field6 = id;
9340 138 : key
9341 138 : }
9342 :
9343 : // We create
9344 : // - one bottom-most image layer,
9345 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
9346 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
9347 : // - a delta layer D3 above the horizon.
9348 : //
9349 : // | D3 |
9350 : // | D1 |
9351 : // -| |-- gc horizon -----------------
9352 : // | | | D2 |
9353 : // --------- img layer ------------------
9354 : //
9355 : // What we should expact from this compaction is:
9356 : // | D3 |
9357 : // | Part of D1 |
9358 : // --------- img layer with D1+D2 at GC horizon------------------
9359 :
9360 : // img layer at 0x10
9361 2 : let img_layer = (0..10)
9362 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9363 2 : .collect_vec();
9364 2 : // or, delta layer at 0x10 if `use_delta_bottom_layer` is true
9365 2 : let delta4 = (0..10)
9366 20 : .map(|id| {
9367 20 : (
9368 20 : get_key(id),
9369 20 : Lsn(0x08),
9370 20 : Value::WalRecord(NeonWalRecord::wal_init(format!("value {id}@0x10"))),
9371 20 : )
9372 20 : })
9373 2 : .collect_vec();
9374 2 :
9375 2 : let delta1 = vec![
9376 2 : (
9377 2 : get_key(1),
9378 2 : Lsn(0x20),
9379 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9380 2 : ),
9381 2 : (
9382 2 : get_key(2),
9383 2 : Lsn(0x30),
9384 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9385 2 : ),
9386 2 : (
9387 2 : get_key(3),
9388 2 : Lsn(0x28),
9389 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9390 2 : ),
9391 2 : (
9392 2 : get_key(3),
9393 2 : Lsn(0x30),
9394 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9395 2 : ),
9396 2 : (
9397 2 : get_key(3),
9398 2 : Lsn(0x40),
9399 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9400 2 : ),
9401 2 : ];
9402 2 : let delta2 = vec![
9403 2 : (
9404 2 : get_key(5),
9405 2 : Lsn(0x20),
9406 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9407 2 : ),
9408 2 : (
9409 2 : get_key(6),
9410 2 : Lsn(0x20),
9411 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9412 2 : ),
9413 2 : ];
9414 2 : let delta3 = vec![
9415 2 : (
9416 2 : get_key(8),
9417 2 : Lsn(0x48),
9418 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9419 2 : ),
9420 2 : (
9421 2 : get_key(9),
9422 2 : Lsn(0x48),
9423 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9424 2 : ),
9425 2 : ];
9426 :
9427 2 : let tline = if use_delta_bottom_layer {
9428 1 : tenant
9429 1 : .create_test_timeline_with_layers(
9430 1 : TIMELINE_ID,
9431 1 : Lsn(0x08),
9432 1 : DEFAULT_PG_VERSION,
9433 1 : &ctx,
9434 1 : Vec::new(), // in-memory layers
9435 1 : vec![
9436 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9437 1 : Lsn(0x08)..Lsn(0x10),
9438 1 : delta4,
9439 1 : ),
9440 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9441 1 : Lsn(0x20)..Lsn(0x48),
9442 1 : delta1,
9443 1 : ),
9444 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9445 1 : Lsn(0x20)..Lsn(0x48),
9446 1 : delta2,
9447 1 : ),
9448 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9449 1 : Lsn(0x48)..Lsn(0x50),
9450 1 : delta3,
9451 1 : ),
9452 1 : ], // delta layers
9453 1 : vec![], // image layers
9454 1 : Lsn(0x50),
9455 1 : )
9456 1 : .await?
9457 : } else {
9458 1 : tenant
9459 1 : .create_test_timeline_with_layers(
9460 1 : TIMELINE_ID,
9461 1 : Lsn(0x10),
9462 1 : DEFAULT_PG_VERSION,
9463 1 : &ctx,
9464 1 : Vec::new(), // in-memory layers
9465 1 : vec![
9466 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9467 1 : Lsn(0x10)..Lsn(0x48),
9468 1 : delta1,
9469 1 : ),
9470 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9471 1 : Lsn(0x10)..Lsn(0x48),
9472 1 : delta2,
9473 1 : ),
9474 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9475 1 : Lsn(0x48)..Lsn(0x50),
9476 1 : delta3,
9477 1 : ),
9478 1 : ], // delta layers
9479 1 : vec![(Lsn(0x10), img_layer)], // image layers
9480 1 : Lsn(0x50),
9481 1 : )
9482 1 : .await?
9483 : };
9484 : {
9485 2 : tline
9486 2 : .applied_gc_cutoff_lsn
9487 2 : .lock_for_write()
9488 2 : .store_and_unlock(Lsn(0x30))
9489 2 : .wait()
9490 2 : .await;
9491 : // Update GC info
9492 2 : let mut guard = tline.gc_info.write().unwrap();
9493 2 : *guard = GcInfo {
9494 2 : retain_lsns: vec![],
9495 2 : cutoffs: GcCutoffs {
9496 2 : time: Some(Lsn(0x30)),
9497 2 : space: Lsn(0x30),
9498 2 : },
9499 2 : leases: Default::default(),
9500 2 : within_ancestor_pitr: false,
9501 2 : };
9502 2 : }
9503 2 :
9504 2 : let expected_result = [
9505 2 : Bytes::from_static(b"value 0@0x10"),
9506 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9507 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9508 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9509 2 : Bytes::from_static(b"value 4@0x10"),
9510 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9511 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9512 2 : Bytes::from_static(b"value 7@0x10"),
9513 2 : Bytes::from_static(b"value 8@0x10@0x48"),
9514 2 : Bytes::from_static(b"value 9@0x10@0x48"),
9515 2 : ];
9516 2 :
9517 2 : let expected_result_at_gc_horizon = [
9518 2 : Bytes::from_static(b"value 0@0x10"),
9519 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9520 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9521 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
9522 2 : Bytes::from_static(b"value 4@0x10"),
9523 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9524 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9525 2 : Bytes::from_static(b"value 7@0x10"),
9526 2 : Bytes::from_static(b"value 8@0x10"),
9527 2 : Bytes::from_static(b"value 9@0x10"),
9528 2 : ];
9529 :
9530 22 : for idx in 0..10 {
9531 20 : assert_eq!(
9532 20 : tline
9533 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9534 20 : .await
9535 20 : .unwrap(),
9536 20 : &expected_result[idx]
9537 : );
9538 20 : assert_eq!(
9539 20 : tline
9540 20 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
9541 20 : .await
9542 20 : .unwrap(),
9543 20 : &expected_result_at_gc_horizon[idx]
9544 : );
9545 : }
9546 :
9547 2 : let cancel = CancellationToken::new();
9548 2 : tline
9549 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9550 2 : .await
9551 2 : .unwrap();
9552 :
9553 22 : for idx in 0..10 {
9554 20 : assert_eq!(
9555 20 : tline
9556 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9557 20 : .await
9558 20 : .unwrap(),
9559 20 : &expected_result[idx]
9560 : );
9561 20 : assert_eq!(
9562 20 : tline
9563 20 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
9564 20 : .await
9565 20 : .unwrap(),
9566 20 : &expected_result_at_gc_horizon[idx]
9567 : );
9568 : }
9569 :
9570 : // increase GC horizon and compact again
9571 : {
9572 2 : tline
9573 2 : .applied_gc_cutoff_lsn
9574 2 : .lock_for_write()
9575 2 : .store_and_unlock(Lsn(0x40))
9576 2 : .wait()
9577 2 : .await;
9578 : // Update GC info
9579 2 : let mut guard = tline.gc_info.write().unwrap();
9580 2 : guard.cutoffs.time = Some(Lsn(0x40));
9581 2 : guard.cutoffs.space = Lsn(0x40);
9582 2 : }
9583 2 : tline
9584 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9585 2 : .await
9586 2 : .unwrap();
9587 2 :
9588 2 : Ok(())
9589 2 : }
9590 :
9591 : #[cfg(feature = "testing")]
9592 : #[tokio::test]
9593 1 : async fn test_generate_key_retention() -> anyhow::Result<()> {
9594 1 : let harness = TenantHarness::create("test_generate_key_retention").await?;
9595 1 : let (tenant, ctx) = harness.load().await;
9596 1 : let tline = tenant
9597 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
9598 1 : .await?;
9599 1 : tline.force_advance_lsn(Lsn(0x70));
9600 1 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
9601 1 : let history = vec![
9602 1 : (
9603 1 : key,
9604 1 : Lsn(0x10),
9605 1 : Value::WalRecord(NeonWalRecord::wal_init("0x10")),
9606 1 : ),
9607 1 : (
9608 1 : key,
9609 1 : Lsn(0x20),
9610 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9611 1 : ),
9612 1 : (
9613 1 : key,
9614 1 : Lsn(0x30),
9615 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9616 1 : ),
9617 1 : (
9618 1 : key,
9619 1 : Lsn(0x40),
9620 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9621 1 : ),
9622 1 : (
9623 1 : key,
9624 1 : Lsn(0x50),
9625 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
9626 1 : ),
9627 1 : (
9628 1 : key,
9629 1 : Lsn(0x60),
9630 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9631 1 : ),
9632 1 : (
9633 1 : key,
9634 1 : Lsn(0x70),
9635 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9636 1 : ),
9637 1 : (
9638 1 : key,
9639 1 : Lsn(0x80),
9640 1 : Value::Image(Bytes::copy_from_slice(
9641 1 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9642 1 : )),
9643 1 : ),
9644 1 : (
9645 1 : key,
9646 1 : Lsn(0x90),
9647 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9648 1 : ),
9649 1 : ];
9650 1 : let res = tline
9651 1 : .generate_key_retention(
9652 1 : key,
9653 1 : &history,
9654 1 : Lsn(0x60),
9655 1 : &[Lsn(0x20), Lsn(0x40), Lsn(0x50)],
9656 1 : 3,
9657 1 : None,
9658 1 : true,
9659 1 : )
9660 1 : .await
9661 1 : .unwrap();
9662 1 : let expected_res = KeyHistoryRetention {
9663 1 : below_horizon: vec![
9664 1 : (
9665 1 : Lsn(0x20),
9666 1 : KeyLogAtLsn(vec![(
9667 1 : Lsn(0x20),
9668 1 : Value::Image(Bytes::from_static(b"0x10;0x20")),
9669 1 : )]),
9670 1 : ),
9671 1 : (
9672 1 : Lsn(0x40),
9673 1 : KeyLogAtLsn(vec![
9674 1 : (
9675 1 : Lsn(0x30),
9676 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9677 1 : ),
9678 1 : (
9679 1 : Lsn(0x40),
9680 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9681 1 : ),
9682 1 : ]),
9683 1 : ),
9684 1 : (
9685 1 : Lsn(0x50),
9686 1 : KeyLogAtLsn(vec![(
9687 1 : Lsn(0x50),
9688 1 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40;0x50")),
9689 1 : )]),
9690 1 : ),
9691 1 : (
9692 1 : Lsn(0x60),
9693 1 : KeyLogAtLsn(vec![(
9694 1 : Lsn(0x60),
9695 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9696 1 : )]),
9697 1 : ),
9698 1 : ],
9699 1 : above_horizon: KeyLogAtLsn(vec![
9700 1 : (
9701 1 : Lsn(0x70),
9702 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9703 1 : ),
9704 1 : (
9705 1 : Lsn(0x80),
9706 1 : Value::Image(Bytes::copy_from_slice(
9707 1 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9708 1 : )),
9709 1 : ),
9710 1 : (
9711 1 : Lsn(0x90),
9712 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9713 1 : ),
9714 1 : ]),
9715 1 : };
9716 1 : assert_eq!(res, expected_res);
9717 1 :
9718 1 : // We expect GC-compaction to run with the original GC. This would create a situation that
9719 1 : // the original GC algorithm removes some delta layers b/c there are full image coverage,
9720 1 : // therefore causing some keys to have an incomplete history below the lowest retain LSN.
9721 1 : // For example, we have
9722 1 : // ```plain
9723 1 : // init delta @ 0x10, image @ 0x20, delta @ 0x30 (gc_horizon), image @ 0x40.
9724 1 : // ```
9725 1 : // Now the GC horizon moves up, and we have
9726 1 : // ```plain
9727 1 : // init delta @ 0x10, image @ 0x20, delta @ 0x30, image @ 0x40 (gc_horizon)
9728 1 : // ```
9729 1 : // The original GC algorithm kicks in, and removes delta @ 0x10, image @ 0x20.
9730 1 : // We will end up with
9731 1 : // ```plain
9732 1 : // delta @ 0x30, image @ 0x40 (gc_horizon)
9733 1 : // ```
9734 1 : // Now we run the GC-compaction, and this key does not have a full history.
9735 1 : // We should be able to handle this partial history and drop everything before the
9736 1 : // gc_horizon image.
9737 1 :
9738 1 : let history = vec![
9739 1 : (
9740 1 : key,
9741 1 : Lsn(0x20),
9742 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9743 1 : ),
9744 1 : (
9745 1 : key,
9746 1 : Lsn(0x30),
9747 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9748 1 : ),
9749 1 : (
9750 1 : key,
9751 1 : Lsn(0x40),
9752 1 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
9753 1 : ),
9754 1 : (
9755 1 : key,
9756 1 : Lsn(0x50),
9757 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
9758 1 : ),
9759 1 : (
9760 1 : key,
9761 1 : Lsn(0x60),
9762 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9763 1 : ),
9764 1 : (
9765 1 : key,
9766 1 : Lsn(0x70),
9767 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9768 1 : ),
9769 1 : (
9770 1 : key,
9771 1 : Lsn(0x80),
9772 1 : Value::Image(Bytes::copy_from_slice(
9773 1 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9774 1 : )),
9775 1 : ),
9776 1 : (
9777 1 : key,
9778 1 : Lsn(0x90),
9779 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9780 1 : ),
9781 1 : ];
9782 1 : let res = tline
9783 1 : .generate_key_retention(
9784 1 : key,
9785 1 : &history,
9786 1 : Lsn(0x60),
9787 1 : &[Lsn(0x40), Lsn(0x50)],
9788 1 : 3,
9789 1 : None,
9790 1 : true,
9791 1 : )
9792 1 : .await
9793 1 : .unwrap();
9794 1 : let expected_res = KeyHistoryRetention {
9795 1 : below_horizon: vec![
9796 1 : (
9797 1 : Lsn(0x40),
9798 1 : KeyLogAtLsn(vec![(
9799 1 : Lsn(0x40),
9800 1 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
9801 1 : )]),
9802 1 : ),
9803 1 : (
9804 1 : Lsn(0x50),
9805 1 : KeyLogAtLsn(vec![(
9806 1 : Lsn(0x50),
9807 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
9808 1 : )]),
9809 1 : ),
9810 1 : (
9811 1 : Lsn(0x60),
9812 1 : KeyLogAtLsn(vec![(
9813 1 : Lsn(0x60),
9814 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9815 1 : )]),
9816 1 : ),
9817 1 : ],
9818 1 : above_horizon: KeyLogAtLsn(vec![
9819 1 : (
9820 1 : Lsn(0x70),
9821 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9822 1 : ),
9823 1 : (
9824 1 : Lsn(0x80),
9825 1 : Value::Image(Bytes::copy_from_slice(
9826 1 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9827 1 : )),
9828 1 : ),
9829 1 : (
9830 1 : Lsn(0x90),
9831 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9832 1 : ),
9833 1 : ]),
9834 1 : };
9835 1 : assert_eq!(res, expected_res);
9836 1 :
9837 1 : // In case of branch compaction, the branch itself does not have the full history, and we need to provide
9838 1 : // the ancestor image in the test case.
9839 1 :
9840 1 : let history = vec![
9841 1 : (
9842 1 : key,
9843 1 : Lsn(0x20),
9844 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9845 1 : ),
9846 1 : (
9847 1 : key,
9848 1 : Lsn(0x30),
9849 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9850 1 : ),
9851 1 : (
9852 1 : key,
9853 1 : Lsn(0x40),
9854 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9855 1 : ),
9856 1 : (
9857 1 : key,
9858 1 : Lsn(0x70),
9859 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9860 1 : ),
9861 1 : ];
9862 1 : let res = tline
9863 1 : .generate_key_retention(
9864 1 : key,
9865 1 : &history,
9866 1 : Lsn(0x60),
9867 1 : &[],
9868 1 : 3,
9869 1 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
9870 1 : true,
9871 1 : )
9872 1 : .await
9873 1 : .unwrap();
9874 1 : let expected_res = KeyHistoryRetention {
9875 1 : below_horizon: vec![(
9876 1 : Lsn(0x60),
9877 1 : KeyLogAtLsn(vec![(
9878 1 : Lsn(0x60),
9879 1 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")), // use the ancestor image to reconstruct the page
9880 1 : )]),
9881 1 : )],
9882 1 : above_horizon: KeyLogAtLsn(vec![(
9883 1 : Lsn(0x70),
9884 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9885 1 : )]),
9886 1 : };
9887 1 : assert_eq!(res, expected_res);
9888 1 :
9889 1 : let history = vec![
9890 1 : (
9891 1 : key,
9892 1 : Lsn(0x20),
9893 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9894 1 : ),
9895 1 : (
9896 1 : key,
9897 1 : Lsn(0x40),
9898 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9899 1 : ),
9900 1 : (
9901 1 : key,
9902 1 : Lsn(0x60),
9903 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9904 1 : ),
9905 1 : (
9906 1 : key,
9907 1 : Lsn(0x70),
9908 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9909 1 : ),
9910 1 : ];
9911 1 : let res = tline
9912 1 : .generate_key_retention(
9913 1 : key,
9914 1 : &history,
9915 1 : Lsn(0x60),
9916 1 : &[Lsn(0x30)],
9917 1 : 3,
9918 1 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
9919 1 : true,
9920 1 : )
9921 1 : .await
9922 1 : .unwrap();
9923 1 : let expected_res = KeyHistoryRetention {
9924 1 : below_horizon: vec![
9925 1 : (
9926 1 : Lsn(0x30),
9927 1 : KeyLogAtLsn(vec![(
9928 1 : Lsn(0x20),
9929 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9930 1 : )]),
9931 1 : ),
9932 1 : (
9933 1 : Lsn(0x60),
9934 1 : KeyLogAtLsn(vec![(
9935 1 : Lsn(0x60),
9936 1 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x40;0x60")),
9937 1 : )]),
9938 1 : ),
9939 1 : ],
9940 1 : above_horizon: KeyLogAtLsn(vec![(
9941 1 : Lsn(0x70),
9942 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9943 1 : )]),
9944 1 : };
9945 1 : assert_eq!(res, expected_res);
9946 1 :
9947 1 : Ok(())
9948 1 : }
9949 :
9950 : #[cfg(feature = "testing")]
9951 : #[tokio::test]
9952 1 : async fn test_simple_bottom_most_compaction_with_retain_lsns() -> anyhow::Result<()> {
9953 1 : let harness =
9954 1 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns").await?;
9955 1 : let (tenant, ctx) = harness.load().await;
9956 1 :
9957 259 : fn get_key(id: u32) -> Key {
9958 259 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9959 259 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9960 259 : key.field6 = id;
9961 259 : key
9962 259 : }
9963 1 :
9964 1 : let img_layer = (0..10)
9965 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9966 1 : .collect_vec();
9967 1 :
9968 1 : let delta1 = vec![
9969 1 : (
9970 1 : get_key(1),
9971 1 : Lsn(0x20),
9972 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9973 1 : ),
9974 1 : (
9975 1 : get_key(2),
9976 1 : Lsn(0x30),
9977 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9978 1 : ),
9979 1 : (
9980 1 : get_key(3),
9981 1 : Lsn(0x28),
9982 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9983 1 : ),
9984 1 : (
9985 1 : get_key(3),
9986 1 : Lsn(0x30),
9987 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9988 1 : ),
9989 1 : (
9990 1 : get_key(3),
9991 1 : Lsn(0x40),
9992 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9993 1 : ),
9994 1 : ];
9995 1 : let delta2 = vec![
9996 1 : (
9997 1 : get_key(5),
9998 1 : Lsn(0x20),
9999 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10000 1 : ),
10001 1 : (
10002 1 : get_key(6),
10003 1 : Lsn(0x20),
10004 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10005 1 : ),
10006 1 : ];
10007 1 : let delta3 = vec![
10008 1 : (
10009 1 : get_key(8),
10010 1 : Lsn(0x48),
10011 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10012 1 : ),
10013 1 : (
10014 1 : get_key(9),
10015 1 : Lsn(0x48),
10016 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10017 1 : ),
10018 1 : ];
10019 1 :
10020 1 : let tline = tenant
10021 1 : .create_test_timeline_with_layers(
10022 1 : TIMELINE_ID,
10023 1 : Lsn(0x10),
10024 1 : DEFAULT_PG_VERSION,
10025 1 : &ctx,
10026 1 : Vec::new(), // in-memory layers
10027 1 : vec![
10028 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
10029 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
10030 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
10031 1 : ], // delta layers
10032 1 : vec![(Lsn(0x10), img_layer)], // image layers
10033 1 : Lsn(0x50),
10034 1 : )
10035 1 : .await?;
10036 1 : {
10037 1 : tline
10038 1 : .applied_gc_cutoff_lsn
10039 1 : .lock_for_write()
10040 1 : .store_and_unlock(Lsn(0x30))
10041 1 : .wait()
10042 1 : .await;
10043 1 : // Update GC info
10044 1 : let mut guard = tline.gc_info.write().unwrap();
10045 1 : *guard = GcInfo {
10046 1 : retain_lsns: vec![
10047 1 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
10048 1 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
10049 1 : ],
10050 1 : cutoffs: GcCutoffs {
10051 1 : time: Some(Lsn(0x30)),
10052 1 : space: Lsn(0x30),
10053 1 : },
10054 1 : leases: Default::default(),
10055 1 : within_ancestor_pitr: false,
10056 1 : };
10057 1 : }
10058 1 :
10059 1 : let expected_result = [
10060 1 : Bytes::from_static(b"value 0@0x10"),
10061 1 : Bytes::from_static(b"value 1@0x10@0x20"),
10062 1 : Bytes::from_static(b"value 2@0x10@0x30"),
10063 1 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
10064 1 : Bytes::from_static(b"value 4@0x10"),
10065 1 : Bytes::from_static(b"value 5@0x10@0x20"),
10066 1 : Bytes::from_static(b"value 6@0x10@0x20"),
10067 1 : Bytes::from_static(b"value 7@0x10"),
10068 1 : Bytes::from_static(b"value 8@0x10@0x48"),
10069 1 : Bytes::from_static(b"value 9@0x10@0x48"),
10070 1 : ];
10071 1 :
10072 1 : let expected_result_at_gc_horizon = [
10073 1 : Bytes::from_static(b"value 0@0x10"),
10074 1 : Bytes::from_static(b"value 1@0x10@0x20"),
10075 1 : Bytes::from_static(b"value 2@0x10@0x30"),
10076 1 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
10077 1 : Bytes::from_static(b"value 4@0x10"),
10078 1 : Bytes::from_static(b"value 5@0x10@0x20"),
10079 1 : Bytes::from_static(b"value 6@0x10@0x20"),
10080 1 : Bytes::from_static(b"value 7@0x10"),
10081 1 : Bytes::from_static(b"value 8@0x10"),
10082 1 : Bytes::from_static(b"value 9@0x10"),
10083 1 : ];
10084 1 :
10085 1 : let expected_result_at_lsn_20 = [
10086 1 : Bytes::from_static(b"value 0@0x10"),
10087 1 : Bytes::from_static(b"value 1@0x10@0x20"),
10088 1 : Bytes::from_static(b"value 2@0x10"),
10089 1 : Bytes::from_static(b"value 3@0x10"),
10090 1 : Bytes::from_static(b"value 4@0x10"),
10091 1 : Bytes::from_static(b"value 5@0x10@0x20"),
10092 1 : Bytes::from_static(b"value 6@0x10@0x20"),
10093 1 : Bytes::from_static(b"value 7@0x10"),
10094 1 : Bytes::from_static(b"value 8@0x10"),
10095 1 : Bytes::from_static(b"value 9@0x10"),
10096 1 : ];
10097 1 :
10098 1 : let expected_result_at_lsn_10 = [
10099 1 : Bytes::from_static(b"value 0@0x10"),
10100 1 : Bytes::from_static(b"value 1@0x10"),
10101 1 : Bytes::from_static(b"value 2@0x10"),
10102 1 : Bytes::from_static(b"value 3@0x10"),
10103 1 : Bytes::from_static(b"value 4@0x10"),
10104 1 : Bytes::from_static(b"value 5@0x10"),
10105 1 : Bytes::from_static(b"value 6@0x10"),
10106 1 : Bytes::from_static(b"value 7@0x10"),
10107 1 : Bytes::from_static(b"value 8@0x10"),
10108 1 : Bytes::from_static(b"value 9@0x10"),
10109 1 : ];
10110 1 :
10111 6 : let verify_result = || async {
10112 6 : let gc_horizon = {
10113 6 : let gc_info = tline.gc_info.read().unwrap();
10114 6 : gc_info.cutoffs.time.unwrap_or_default()
10115 1 : };
10116 66 : for idx in 0..10 {
10117 60 : assert_eq!(
10118 60 : tline
10119 60 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10120 60 : .await
10121 60 : .unwrap(),
10122 60 : &expected_result[idx]
10123 1 : );
10124 60 : assert_eq!(
10125 60 : tline
10126 60 : .get(get_key(idx as u32), gc_horizon, &ctx)
10127 60 : .await
10128 60 : .unwrap(),
10129 60 : &expected_result_at_gc_horizon[idx]
10130 1 : );
10131 60 : assert_eq!(
10132 60 : tline
10133 60 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
10134 60 : .await
10135 60 : .unwrap(),
10136 60 : &expected_result_at_lsn_20[idx]
10137 1 : );
10138 60 : assert_eq!(
10139 60 : tline
10140 60 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
10141 60 : .await
10142 60 : .unwrap(),
10143 60 : &expected_result_at_lsn_10[idx]
10144 1 : );
10145 1 : }
10146 12 : };
10147 1 :
10148 1 : verify_result().await;
10149 1 :
10150 1 : let cancel = CancellationToken::new();
10151 1 : let mut dryrun_flags = EnumSet::new();
10152 1 : dryrun_flags.insert(CompactFlags::DryRun);
10153 1 :
10154 1 : tline
10155 1 : .compact_with_gc(
10156 1 : &cancel,
10157 1 : CompactOptions {
10158 1 : flags: dryrun_flags,
10159 1 : ..Default::default()
10160 1 : },
10161 1 : &ctx,
10162 1 : )
10163 1 : .await
10164 1 : .unwrap();
10165 1 : // 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
10166 1 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
10167 1 : verify_result().await;
10168 1 :
10169 1 : tline
10170 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10171 1 : .await
10172 1 : .unwrap();
10173 1 : verify_result().await;
10174 1 :
10175 1 : // compact again
10176 1 : tline
10177 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10178 1 : .await
10179 1 : .unwrap();
10180 1 : verify_result().await;
10181 1 :
10182 1 : // increase GC horizon and compact again
10183 1 : {
10184 1 : tline
10185 1 : .applied_gc_cutoff_lsn
10186 1 : .lock_for_write()
10187 1 : .store_and_unlock(Lsn(0x38))
10188 1 : .wait()
10189 1 : .await;
10190 1 : // Update GC info
10191 1 : let mut guard = tline.gc_info.write().unwrap();
10192 1 : guard.cutoffs.time = Some(Lsn(0x38));
10193 1 : guard.cutoffs.space = Lsn(0x38);
10194 1 : }
10195 1 : tline
10196 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10197 1 : .await
10198 1 : .unwrap();
10199 1 : verify_result().await; // no wals between 0x30 and 0x38, so we should obtain the same result
10200 1 :
10201 1 : // not increasing the GC horizon and compact again
10202 1 : tline
10203 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10204 1 : .await
10205 1 : .unwrap();
10206 1 : verify_result().await;
10207 1 :
10208 1 : Ok(())
10209 1 : }
10210 :
10211 : #[cfg(feature = "testing")]
10212 : #[tokio::test]
10213 1 : async fn test_simple_bottom_most_compaction_with_retain_lsns_single_key() -> anyhow::Result<()>
10214 1 : {
10215 1 : let harness =
10216 1 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns_single_key")
10217 1 : .await?;
10218 1 : let (tenant, ctx) = harness.load().await;
10219 1 :
10220 176 : fn get_key(id: u32) -> Key {
10221 176 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
10222 176 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
10223 176 : key.field6 = id;
10224 176 : key
10225 176 : }
10226 1 :
10227 1 : let img_layer = (0..10)
10228 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10229 1 : .collect_vec();
10230 1 :
10231 1 : let delta1 = vec![
10232 1 : (
10233 1 : get_key(1),
10234 1 : Lsn(0x20),
10235 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10236 1 : ),
10237 1 : (
10238 1 : get_key(1),
10239 1 : Lsn(0x28),
10240 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
10241 1 : ),
10242 1 : ];
10243 1 : let delta2 = vec![
10244 1 : (
10245 1 : get_key(1),
10246 1 : Lsn(0x30),
10247 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10248 1 : ),
10249 1 : (
10250 1 : get_key(1),
10251 1 : Lsn(0x38),
10252 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
10253 1 : ),
10254 1 : ];
10255 1 : let delta3 = vec![
10256 1 : (
10257 1 : get_key(8),
10258 1 : Lsn(0x48),
10259 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10260 1 : ),
10261 1 : (
10262 1 : get_key(9),
10263 1 : Lsn(0x48),
10264 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10265 1 : ),
10266 1 : ];
10267 1 :
10268 1 : let tline = tenant
10269 1 : .create_test_timeline_with_layers(
10270 1 : TIMELINE_ID,
10271 1 : Lsn(0x10),
10272 1 : DEFAULT_PG_VERSION,
10273 1 : &ctx,
10274 1 : Vec::new(), // in-memory layers
10275 1 : vec![
10276 1 : // delta1 and delta 2 only contain a single key but multiple updates
10277 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x30), delta1),
10278 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
10279 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x50), delta3),
10280 1 : ], // delta layers
10281 1 : vec![(Lsn(0x10), img_layer)], // image layers
10282 1 : Lsn(0x50),
10283 1 : )
10284 1 : .await?;
10285 1 : {
10286 1 : tline
10287 1 : .applied_gc_cutoff_lsn
10288 1 : .lock_for_write()
10289 1 : .store_and_unlock(Lsn(0x30))
10290 1 : .wait()
10291 1 : .await;
10292 1 : // Update GC info
10293 1 : let mut guard = tline.gc_info.write().unwrap();
10294 1 : *guard = GcInfo {
10295 1 : retain_lsns: vec![
10296 1 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
10297 1 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
10298 1 : ],
10299 1 : cutoffs: GcCutoffs {
10300 1 : time: Some(Lsn(0x30)),
10301 1 : space: Lsn(0x30),
10302 1 : },
10303 1 : leases: Default::default(),
10304 1 : within_ancestor_pitr: false,
10305 1 : };
10306 1 : }
10307 1 :
10308 1 : let expected_result = [
10309 1 : Bytes::from_static(b"value 0@0x10"),
10310 1 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
10311 1 : Bytes::from_static(b"value 2@0x10"),
10312 1 : Bytes::from_static(b"value 3@0x10"),
10313 1 : Bytes::from_static(b"value 4@0x10"),
10314 1 : Bytes::from_static(b"value 5@0x10"),
10315 1 : Bytes::from_static(b"value 6@0x10"),
10316 1 : Bytes::from_static(b"value 7@0x10"),
10317 1 : Bytes::from_static(b"value 8@0x10@0x48"),
10318 1 : Bytes::from_static(b"value 9@0x10@0x48"),
10319 1 : ];
10320 1 :
10321 1 : let expected_result_at_gc_horizon = [
10322 1 : Bytes::from_static(b"value 0@0x10"),
10323 1 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
10324 1 : Bytes::from_static(b"value 2@0x10"),
10325 1 : Bytes::from_static(b"value 3@0x10"),
10326 1 : Bytes::from_static(b"value 4@0x10"),
10327 1 : Bytes::from_static(b"value 5@0x10"),
10328 1 : Bytes::from_static(b"value 6@0x10"),
10329 1 : Bytes::from_static(b"value 7@0x10"),
10330 1 : Bytes::from_static(b"value 8@0x10"),
10331 1 : Bytes::from_static(b"value 9@0x10"),
10332 1 : ];
10333 1 :
10334 1 : let expected_result_at_lsn_20 = [
10335 1 : Bytes::from_static(b"value 0@0x10"),
10336 1 : Bytes::from_static(b"value 1@0x10@0x20"),
10337 1 : Bytes::from_static(b"value 2@0x10"),
10338 1 : Bytes::from_static(b"value 3@0x10"),
10339 1 : Bytes::from_static(b"value 4@0x10"),
10340 1 : Bytes::from_static(b"value 5@0x10"),
10341 1 : Bytes::from_static(b"value 6@0x10"),
10342 1 : Bytes::from_static(b"value 7@0x10"),
10343 1 : Bytes::from_static(b"value 8@0x10"),
10344 1 : Bytes::from_static(b"value 9@0x10"),
10345 1 : ];
10346 1 :
10347 1 : let expected_result_at_lsn_10 = [
10348 1 : Bytes::from_static(b"value 0@0x10"),
10349 1 : Bytes::from_static(b"value 1@0x10"),
10350 1 : Bytes::from_static(b"value 2@0x10"),
10351 1 : Bytes::from_static(b"value 3@0x10"),
10352 1 : Bytes::from_static(b"value 4@0x10"),
10353 1 : Bytes::from_static(b"value 5@0x10"),
10354 1 : Bytes::from_static(b"value 6@0x10"),
10355 1 : Bytes::from_static(b"value 7@0x10"),
10356 1 : Bytes::from_static(b"value 8@0x10"),
10357 1 : Bytes::from_static(b"value 9@0x10"),
10358 1 : ];
10359 1 :
10360 4 : let verify_result = || async {
10361 4 : let gc_horizon = {
10362 4 : let gc_info = tline.gc_info.read().unwrap();
10363 4 : gc_info.cutoffs.time.unwrap_or_default()
10364 1 : };
10365 44 : for idx in 0..10 {
10366 40 : assert_eq!(
10367 40 : tline
10368 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10369 40 : .await
10370 40 : .unwrap(),
10371 40 : &expected_result[idx]
10372 1 : );
10373 40 : assert_eq!(
10374 40 : tline
10375 40 : .get(get_key(idx as u32), gc_horizon, &ctx)
10376 40 : .await
10377 40 : .unwrap(),
10378 40 : &expected_result_at_gc_horizon[idx]
10379 1 : );
10380 40 : assert_eq!(
10381 40 : tline
10382 40 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
10383 40 : .await
10384 40 : .unwrap(),
10385 40 : &expected_result_at_lsn_20[idx]
10386 1 : );
10387 40 : assert_eq!(
10388 40 : tline
10389 40 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
10390 40 : .await
10391 40 : .unwrap(),
10392 40 : &expected_result_at_lsn_10[idx]
10393 1 : );
10394 1 : }
10395 8 : };
10396 1 :
10397 1 : verify_result().await;
10398 1 :
10399 1 : let cancel = CancellationToken::new();
10400 1 : let mut dryrun_flags = EnumSet::new();
10401 1 : dryrun_flags.insert(CompactFlags::DryRun);
10402 1 :
10403 1 : tline
10404 1 : .compact_with_gc(
10405 1 : &cancel,
10406 1 : CompactOptions {
10407 1 : flags: dryrun_flags,
10408 1 : ..Default::default()
10409 1 : },
10410 1 : &ctx,
10411 1 : )
10412 1 : .await
10413 1 : .unwrap();
10414 1 : // 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
10415 1 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
10416 1 : verify_result().await;
10417 1 :
10418 1 : tline
10419 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10420 1 : .await
10421 1 : .unwrap();
10422 1 : verify_result().await;
10423 1 :
10424 1 : // compact again
10425 1 : tline
10426 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10427 1 : .await
10428 1 : .unwrap();
10429 1 : verify_result().await;
10430 1 :
10431 1 : Ok(())
10432 1 : }
10433 :
10434 : #[cfg(feature = "testing")]
10435 : #[tokio::test]
10436 1 : async fn test_simple_bottom_most_compaction_on_branch() -> anyhow::Result<()> {
10437 1 : use models::CompactLsnRange;
10438 1 :
10439 1 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_on_branch").await?;
10440 1 : let (tenant, ctx) = harness.load().await;
10441 1 :
10442 83 : fn get_key(id: u32) -> Key {
10443 83 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
10444 83 : key.field6 = id;
10445 83 : key
10446 83 : }
10447 1 :
10448 1 : let img_layer = (0..10)
10449 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10450 1 : .collect_vec();
10451 1 :
10452 1 : let delta1 = vec![
10453 1 : (
10454 1 : get_key(1),
10455 1 : Lsn(0x20),
10456 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10457 1 : ),
10458 1 : (
10459 1 : get_key(2),
10460 1 : Lsn(0x30),
10461 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10462 1 : ),
10463 1 : (
10464 1 : get_key(3),
10465 1 : Lsn(0x28),
10466 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
10467 1 : ),
10468 1 : (
10469 1 : get_key(3),
10470 1 : Lsn(0x30),
10471 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10472 1 : ),
10473 1 : (
10474 1 : get_key(3),
10475 1 : Lsn(0x40),
10476 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
10477 1 : ),
10478 1 : ];
10479 1 : let delta2 = vec![
10480 1 : (
10481 1 : get_key(5),
10482 1 : Lsn(0x20),
10483 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10484 1 : ),
10485 1 : (
10486 1 : get_key(6),
10487 1 : Lsn(0x20),
10488 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10489 1 : ),
10490 1 : ];
10491 1 : let delta3 = vec![
10492 1 : (
10493 1 : get_key(8),
10494 1 : Lsn(0x48),
10495 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10496 1 : ),
10497 1 : (
10498 1 : get_key(9),
10499 1 : Lsn(0x48),
10500 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10501 1 : ),
10502 1 : ];
10503 1 :
10504 1 : let parent_tline = tenant
10505 1 : .create_test_timeline_with_layers(
10506 1 : TIMELINE_ID,
10507 1 : Lsn(0x10),
10508 1 : DEFAULT_PG_VERSION,
10509 1 : &ctx,
10510 1 : vec![], // in-memory layers
10511 1 : vec![], // delta layers
10512 1 : vec![(Lsn(0x18), img_layer)], // image layers
10513 1 : Lsn(0x18),
10514 1 : )
10515 1 : .await?;
10516 1 :
10517 1 : parent_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
10518 1 :
10519 1 : let branch_tline = tenant
10520 1 : .branch_timeline_test_with_layers(
10521 1 : &parent_tline,
10522 1 : NEW_TIMELINE_ID,
10523 1 : Some(Lsn(0x18)),
10524 1 : &ctx,
10525 1 : vec![
10526 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
10527 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
10528 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
10529 1 : ], // delta layers
10530 1 : vec![], // image layers
10531 1 : Lsn(0x50),
10532 1 : )
10533 1 : .await?;
10534 1 :
10535 1 : branch_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
10536 1 :
10537 1 : {
10538 1 : parent_tline
10539 1 : .applied_gc_cutoff_lsn
10540 1 : .lock_for_write()
10541 1 : .store_and_unlock(Lsn(0x10))
10542 1 : .wait()
10543 1 : .await;
10544 1 : // Update GC info
10545 1 : let mut guard = parent_tline.gc_info.write().unwrap();
10546 1 : *guard = GcInfo {
10547 1 : retain_lsns: vec![(Lsn(0x18), branch_tline.timeline_id, MaybeOffloaded::No)],
10548 1 : cutoffs: GcCutoffs {
10549 1 : time: Some(Lsn(0x10)),
10550 1 : space: Lsn(0x10),
10551 1 : },
10552 1 : leases: Default::default(),
10553 1 : within_ancestor_pitr: false,
10554 1 : };
10555 1 : }
10556 1 :
10557 1 : {
10558 1 : branch_tline
10559 1 : .applied_gc_cutoff_lsn
10560 1 : .lock_for_write()
10561 1 : .store_and_unlock(Lsn(0x50))
10562 1 : .wait()
10563 1 : .await;
10564 1 : // Update GC info
10565 1 : let mut guard = branch_tline.gc_info.write().unwrap();
10566 1 : *guard = GcInfo {
10567 1 : retain_lsns: vec![(Lsn(0x40), branch_tline.timeline_id, MaybeOffloaded::No)],
10568 1 : cutoffs: GcCutoffs {
10569 1 : time: Some(Lsn(0x50)),
10570 1 : space: Lsn(0x50),
10571 1 : },
10572 1 : leases: Default::default(),
10573 1 : within_ancestor_pitr: false,
10574 1 : };
10575 1 : }
10576 1 :
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"),
10580 1 : Bytes::from_static(b"value 2@0x10@0x30"),
10581 1 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
10582 1 : Bytes::from_static(b"value 4@0x10"),
10583 1 : Bytes::from_static(b"value 5@0x10@0x20"),
10584 1 : Bytes::from_static(b"value 6@0x10@0x20"),
10585 1 : Bytes::from_static(b"value 7@0x10"),
10586 1 : Bytes::from_static(b"value 8@0x10@0x48"),
10587 1 : Bytes::from_static(b"value 9@0x10@0x48"),
10588 1 : ];
10589 1 :
10590 1 : let expected_result_at_lsn_40 = [
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@0x30"),
10594 1 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
10595 1 : Bytes::from_static(b"value 4@0x10"),
10596 1 : Bytes::from_static(b"value 5@0x10@0x20"),
10597 1 : Bytes::from_static(b"value 6@0x10@0x20"),
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 1 :
10603 3 : let verify_result = || async {
10604 33 : for idx in 0..10 {
10605 30 : assert_eq!(
10606 30 : branch_tline
10607 30 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10608 30 : .await
10609 30 : .unwrap(),
10610 30 : &expected_result_at_gc_horizon[idx]
10611 1 : );
10612 30 : assert_eq!(
10613 30 : branch_tline
10614 30 : .get(get_key(idx as u32), Lsn(0x40), &ctx)
10615 30 : .await
10616 30 : .unwrap(),
10617 30 : &expected_result_at_lsn_40[idx]
10618 1 : );
10619 1 : }
10620 6 : };
10621 1 :
10622 1 : verify_result().await;
10623 1 :
10624 1 : let cancel = CancellationToken::new();
10625 1 : branch_tline
10626 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10627 1 : .await
10628 1 : .unwrap();
10629 1 :
10630 1 : verify_result().await;
10631 1 :
10632 1 : // Piggyback a compaction with above_lsn. Ensure it works correctly when the specified LSN intersects with the layer files.
10633 1 : // Now we already have a single large delta layer, so the compaction min_layer_lsn should be the same as ancestor LSN (0x18).
10634 1 : branch_tline
10635 1 : .compact_with_gc(
10636 1 : &cancel,
10637 1 : CompactOptions {
10638 1 : compact_lsn_range: Some(CompactLsnRange::above(Lsn(0x40))),
10639 1 : ..Default::default()
10640 1 : },
10641 1 : &ctx,
10642 1 : )
10643 1 : .await
10644 1 : .unwrap();
10645 1 :
10646 1 : verify_result().await;
10647 1 :
10648 1 : Ok(())
10649 1 : }
10650 :
10651 : // Regression test for https://github.com/neondatabase/neon/issues/9012
10652 : // Create an image arrangement where we have to read at different LSN ranges
10653 : // from a delta layer. This is achieved by overlapping an image layer on top of
10654 : // a delta layer. Like so:
10655 : //
10656 : // A B
10657 : // +----------------+ -> delta_layer
10658 : // | | ^ lsn
10659 : // | =========|-> nested_image_layer |
10660 : // | C | |
10661 : // +----------------+ |
10662 : // ======== -> baseline_image_layer +-------> key
10663 : //
10664 : //
10665 : // When querying the key range [A, B) we need to read at different LSN ranges
10666 : // for [A, C) and [C, B). This test checks that the described edge case is handled correctly.
10667 : #[cfg(feature = "testing")]
10668 : #[tokio::test]
10669 1 : async fn test_vectored_read_with_nested_image_layer() -> anyhow::Result<()> {
10670 1 : let harness = TenantHarness::create("test_vectored_read_with_nested_image_layer").await?;
10671 1 : let (tenant, ctx) = harness.load().await;
10672 1 :
10673 1 : let will_init_keys = [2, 6];
10674 22 : fn get_key(id: u32) -> Key {
10675 22 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
10676 22 : key.field6 = id;
10677 22 : key
10678 22 : }
10679 1 :
10680 1 : let mut expected_key_values = HashMap::new();
10681 1 :
10682 1 : let baseline_image_layer_lsn = Lsn(0x10);
10683 1 : let mut baseline_img_layer = Vec::new();
10684 6 : for i in 0..5 {
10685 5 : let key = get_key(i);
10686 5 : let value = format!("value {i}@{baseline_image_layer_lsn}");
10687 5 :
10688 5 : let removed = expected_key_values.insert(key, value.clone());
10689 5 : assert!(removed.is_none());
10690 1 :
10691 5 : baseline_img_layer.push((key, Bytes::from(value)));
10692 1 : }
10693 1 :
10694 1 : let nested_image_layer_lsn = Lsn(0x50);
10695 1 : let mut nested_img_layer = Vec::new();
10696 6 : for i in 5..10 {
10697 5 : let key = get_key(i);
10698 5 : let value = format!("value {i}@{nested_image_layer_lsn}");
10699 5 :
10700 5 : let removed = expected_key_values.insert(key, value.clone());
10701 5 : assert!(removed.is_none());
10702 1 :
10703 5 : nested_img_layer.push((key, Bytes::from(value)));
10704 1 : }
10705 1 :
10706 1 : let mut delta_layer_spec = Vec::default();
10707 1 : let delta_layer_start_lsn = Lsn(0x20);
10708 1 : let mut delta_layer_end_lsn = delta_layer_start_lsn;
10709 1 :
10710 11 : for i in 0..10 {
10711 10 : let key = get_key(i);
10712 10 : let key_in_nested = nested_img_layer
10713 10 : .iter()
10714 40 : .any(|(key_with_img, _)| *key_with_img == key);
10715 10 : let lsn = {
10716 10 : if key_in_nested {
10717 5 : Lsn(nested_image_layer_lsn.0 + 0x10)
10718 1 : } else {
10719 5 : delta_layer_start_lsn
10720 1 : }
10721 1 : };
10722 1 :
10723 10 : let will_init = will_init_keys.contains(&i);
10724 10 : if will_init {
10725 2 : delta_layer_spec.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init(""))));
10726 2 :
10727 2 : expected_key_values.insert(key, "".to_string());
10728 8 : } else {
10729 8 : let delta = format!("@{lsn}");
10730 8 : delta_layer_spec.push((
10731 8 : key,
10732 8 : lsn,
10733 8 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
10734 8 : ));
10735 8 :
10736 8 : expected_key_values
10737 8 : .get_mut(&key)
10738 8 : .expect("An image exists for each key")
10739 8 : .push_str(delta.as_str());
10740 8 : }
10741 10 : delta_layer_end_lsn = std::cmp::max(delta_layer_start_lsn, lsn);
10742 1 : }
10743 1 :
10744 1 : delta_layer_end_lsn = Lsn(delta_layer_end_lsn.0 + 1);
10745 1 :
10746 1 : assert!(
10747 1 : nested_image_layer_lsn > delta_layer_start_lsn
10748 1 : && nested_image_layer_lsn < delta_layer_end_lsn
10749 1 : );
10750 1 :
10751 1 : let tline = tenant
10752 1 : .create_test_timeline_with_layers(
10753 1 : TIMELINE_ID,
10754 1 : baseline_image_layer_lsn,
10755 1 : DEFAULT_PG_VERSION,
10756 1 : &ctx,
10757 1 : vec![], // in-memory layers
10758 1 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
10759 1 : delta_layer_start_lsn..delta_layer_end_lsn,
10760 1 : delta_layer_spec,
10761 1 : )], // delta layers
10762 1 : vec![
10763 1 : (baseline_image_layer_lsn, baseline_img_layer),
10764 1 : (nested_image_layer_lsn, nested_img_layer),
10765 1 : ], // image layers
10766 1 : delta_layer_end_lsn,
10767 1 : )
10768 1 : .await?;
10769 1 :
10770 1 : let query = VersionedKeySpaceQuery::uniform(
10771 1 : KeySpace::single(get_key(0)..get_key(10)),
10772 1 : delta_layer_end_lsn,
10773 1 : );
10774 1 :
10775 1 : let results = tline
10776 1 : .get_vectored(query, IoConcurrency::sequential(), &ctx)
10777 1 : .await
10778 1 : .expect("No vectored errors");
10779 11 : for (key, res) in results {
10780 10 : let value = res.expect("No key errors");
10781 10 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
10782 10 : assert_eq!(value, Bytes::from(expected_value));
10783 1 : }
10784 1 :
10785 1 : Ok(())
10786 1 : }
10787 :
10788 : #[cfg(feature = "testing")]
10789 : #[tokio::test]
10790 1 : async fn test_vectored_read_with_image_layer_inside_inmem() -> anyhow::Result<()> {
10791 1 : let harness =
10792 1 : TenantHarness::create("test_vectored_read_with_image_layer_inside_inmem").await?;
10793 1 : let (tenant, ctx) = harness.load().await;
10794 1 :
10795 1 : let will_init_keys = [2, 6];
10796 32 : fn get_key(id: u32) -> Key {
10797 32 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
10798 32 : key.field6 = id;
10799 32 : key
10800 32 : }
10801 1 :
10802 1 : let mut expected_key_values = HashMap::new();
10803 1 :
10804 1 : let baseline_image_layer_lsn = Lsn(0x10);
10805 1 : let mut baseline_img_layer = Vec::new();
10806 6 : for i in 0..5 {
10807 5 : let key = get_key(i);
10808 5 : let value = format!("value {i}@{baseline_image_layer_lsn}");
10809 5 :
10810 5 : let removed = expected_key_values.insert(key, value.clone());
10811 5 : assert!(removed.is_none());
10812 1 :
10813 5 : baseline_img_layer.push((key, Bytes::from(value)));
10814 1 : }
10815 1 :
10816 1 : let nested_image_layer_lsn = Lsn(0x50);
10817 1 : let mut nested_img_layer = Vec::new();
10818 6 : for i in 5..10 {
10819 5 : let key = get_key(i);
10820 5 : let value = format!("value {i}@{nested_image_layer_lsn}");
10821 5 :
10822 5 : let removed = expected_key_values.insert(key, value.clone());
10823 5 : assert!(removed.is_none());
10824 1 :
10825 5 : nested_img_layer.push((key, Bytes::from(value)));
10826 1 : }
10827 1 :
10828 1 : let frozen_layer = {
10829 1 : let lsn_range = Lsn(0x40)..Lsn(0x60);
10830 1 : let mut data = Vec::new();
10831 11 : for i in 0..10 {
10832 10 : let key = get_key(i);
10833 10 : let key_in_nested = nested_img_layer
10834 10 : .iter()
10835 40 : .any(|(key_with_img, _)| *key_with_img == key);
10836 10 : let lsn = {
10837 10 : if key_in_nested {
10838 5 : Lsn(nested_image_layer_lsn.0 + 5)
10839 1 : } else {
10840 5 : lsn_range.start
10841 1 : }
10842 1 : };
10843 1 :
10844 10 : let will_init = will_init_keys.contains(&i);
10845 10 : if will_init {
10846 2 : data.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init(""))));
10847 2 :
10848 2 : expected_key_values.insert(key, "".to_string());
10849 8 : } else {
10850 8 : let delta = format!("@{lsn}");
10851 8 : data.push((
10852 8 : key,
10853 8 : lsn,
10854 8 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
10855 8 : ));
10856 8 :
10857 8 : expected_key_values
10858 8 : .get_mut(&key)
10859 8 : .expect("An image exists for each key")
10860 8 : .push_str(delta.as_str());
10861 8 : }
10862 1 : }
10863 1 :
10864 1 : InMemoryLayerTestDesc {
10865 1 : lsn_range,
10866 1 : is_open: false,
10867 1 : data,
10868 1 : }
10869 1 : };
10870 1 :
10871 1 : let (open_layer, last_record_lsn) = {
10872 1 : let start_lsn = Lsn(0x70);
10873 1 : let mut data = Vec::new();
10874 1 : let mut end_lsn = Lsn(0);
10875 11 : for i in 0..10 {
10876 10 : let key = get_key(i);
10877 10 : let lsn = Lsn(start_lsn.0 + i as u64);
10878 10 : let delta = format!("@{lsn}");
10879 10 : data.push((
10880 10 : key,
10881 10 : lsn,
10882 10 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
10883 10 : ));
10884 10 :
10885 10 : expected_key_values
10886 10 : .get_mut(&key)
10887 10 : .expect("An image exists for each key")
10888 10 : .push_str(delta.as_str());
10889 10 :
10890 10 : end_lsn = std::cmp::max(end_lsn, lsn);
10891 10 : }
10892 1 :
10893 1 : (
10894 1 : InMemoryLayerTestDesc {
10895 1 : lsn_range: start_lsn..Lsn::MAX,
10896 1 : is_open: true,
10897 1 : data,
10898 1 : },
10899 1 : end_lsn,
10900 1 : )
10901 1 : };
10902 1 :
10903 1 : assert!(
10904 1 : nested_image_layer_lsn > frozen_layer.lsn_range.start
10905 1 : && nested_image_layer_lsn < frozen_layer.lsn_range.end
10906 1 : );
10907 1 :
10908 1 : let tline = tenant
10909 1 : .create_test_timeline_with_layers(
10910 1 : TIMELINE_ID,
10911 1 : baseline_image_layer_lsn,
10912 1 : DEFAULT_PG_VERSION,
10913 1 : &ctx,
10914 1 : vec![open_layer, frozen_layer], // in-memory layers
10915 1 : Vec::new(), // delta layers
10916 1 : vec![
10917 1 : (baseline_image_layer_lsn, baseline_img_layer),
10918 1 : (nested_image_layer_lsn, nested_img_layer),
10919 1 : ], // image layers
10920 1 : last_record_lsn,
10921 1 : )
10922 1 : .await?;
10923 1 :
10924 1 : let query = VersionedKeySpaceQuery::uniform(
10925 1 : KeySpace::single(get_key(0)..get_key(10)),
10926 1 : last_record_lsn,
10927 1 : );
10928 1 :
10929 1 : let results = tline
10930 1 : .get_vectored(query, IoConcurrency::sequential(), &ctx)
10931 1 : .await
10932 1 : .expect("No vectored errors");
10933 11 : for (key, res) in results {
10934 10 : let value = res.expect("No key errors");
10935 10 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
10936 10 : assert_eq!(value, Bytes::from(expected_value.clone()));
10937 1 :
10938 10 : tracing::info!("key={key} value={expected_value}");
10939 1 : }
10940 1 :
10941 1 : Ok(())
10942 1 : }
10943 :
10944 : // A randomized read path test. Generates a layer map according to a deterministic
10945 : // specification. Fills the (key, LSN) space in random manner and then performs
10946 : // random scattered queries validating the results against in-memory storage.
10947 : //
10948 : // See this internal Notion page for a diagram of the layer map:
10949 : // https://www.notion.so/neondatabase/Read-Path-Unit-Testing-Fuzzing-1d1f189e0047806c8e5cd37781b0a350?pvs=4
10950 : //
10951 : // A fuzzing mode is also supported. In this mode, the test will use a random
10952 : // seed instead of a hardcoded one. Use it in conjunction with `cargo stress`
10953 : // to run multiple instances in parallel:
10954 : //
10955 : // $ RUST_BACKTRACE=1 RUST_LOG=INFO \
10956 : // cargo stress --package=pageserver --features=testing,fuzz-read-path --release -- test_read_path
10957 : #[cfg(feature = "testing")]
10958 : #[tokio::test]
10959 1 : async fn test_read_path() -> anyhow::Result<()> {
10960 1 : use rand::seq::SliceRandom;
10961 1 :
10962 1 : let seed = if cfg!(feature = "fuzz-read-path") {
10963 1 : let seed: u64 = thread_rng().r#gen();
10964 0 : seed
10965 1 : } else {
10966 1 : // Use a hard-coded seed when not in fuzzing mode.
10967 1 : // Note that with the current approach results are not reproducible
10968 1 : // accross platforms and Rust releases.
10969 1 : const SEED: u64 = 0;
10970 1 : SEED
10971 1 : };
10972 1 :
10973 1 : let mut random = StdRng::seed_from_u64(seed);
10974 1 :
10975 1 : let (queries, will_init_chance, gap_chance) = if cfg!(feature = "fuzz-read-path") {
10976 1 : const QUERIES: u64 = 5000;
10977 1 : let will_init_chance: u8 = random.gen_range(0..=10);
10978 0 : let gap_chance: u8 = random.gen_range(0..=50);
10979 0 :
10980 0 : (QUERIES, will_init_chance, gap_chance)
10981 1 : } else {
10982 1 : const QUERIES: u64 = 1000;
10983 1 : const WILL_INIT_CHANCE: u8 = 1;
10984 1 : const GAP_CHANCE: u8 = 5;
10985 1 :
10986 1 : (QUERIES, WILL_INIT_CHANCE, GAP_CHANCE)
10987 1 : };
10988 1 :
10989 1 : let harness = TenantHarness::create("test_read_path").await?;
10990 1 : let (tenant, ctx) = harness.load().await;
10991 1 :
10992 1 : tracing::info!("Using random seed: {seed}");
10993 1 : tracing::info!(%will_init_chance, %gap_chance, "Fill params");
10994 1 :
10995 1 : // Define the layer map shape. Note that this part is not randomized.
10996 1 :
10997 1 : const KEY_DIMENSION_SIZE: u32 = 99;
10998 1 : let start_key = Key::from_hex("110000000033333333444444445500000000").unwrap();
10999 1 : let end_key = start_key.add(KEY_DIMENSION_SIZE);
11000 1 : let total_key_range = start_key..end_key;
11001 1 : let total_key_range_size = end_key.to_i128() - start_key.to_i128();
11002 1 : let total_start_lsn = Lsn(104);
11003 1 : let last_record_lsn = Lsn(504);
11004 1 :
11005 1 : assert!(total_key_range_size % 3 == 0);
11006 1 :
11007 1 : let in_memory_layers_shape = vec![
11008 1 : (total_key_range.clone(), Lsn(304)..Lsn(400)),
11009 1 : (total_key_range.clone(), Lsn(400)..last_record_lsn),
11010 1 : ];
11011 1 :
11012 1 : let delta_layers_shape = vec![
11013 1 : (
11014 1 : start_key..(start_key.add((total_key_range_size / 3) as u32)),
11015 1 : Lsn(200)..Lsn(304),
11016 1 : ),
11017 1 : (
11018 1 : (start_key.add((total_key_range_size / 3) as u32))
11019 1 : ..(start_key.add((total_key_range_size * 2 / 3) as u32)),
11020 1 : Lsn(200)..Lsn(304),
11021 1 : ),
11022 1 : (
11023 1 : (start_key.add((total_key_range_size * 2 / 3) as u32))
11024 1 : ..(start_key.add(total_key_range_size as u32)),
11025 1 : Lsn(200)..Lsn(304),
11026 1 : ),
11027 1 : ];
11028 1 :
11029 1 : let image_layers_shape = vec![
11030 1 : (
11031 1 : start_key.add((total_key_range_size * 2 / 3 - 10) as u32)
11032 1 : ..start_key.add((total_key_range_size * 2 / 3 + 10) as u32),
11033 1 : Lsn(456),
11034 1 : ),
11035 1 : (
11036 1 : start_key.add((total_key_range_size / 3 - 10) as u32)
11037 1 : ..start_key.add((total_key_range_size / 3 + 10) as u32),
11038 1 : Lsn(256),
11039 1 : ),
11040 1 : (total_key_range.clone(), total_start_lsn),
11041 1 : ];
11042 1 :
11043 1 : let specification = TestTimelineSpecification {
11044 1 : start_lsn: total_start_lsn,
11045 1 : last_record_lsn,
11046 1 : in_memory_layers_shape,
11047 1 : delta_layers_shape,
11048 1 : image_layers_shape,
11049 1 : gap_chance,
11050 1 : will_init_chance,
11051 1 : };
11052 1 :
11053 1 : // Create and randomly fill in the layers according to the specification
11054 1 : let (tline, storage, interesting_lsns) = randomize_timeline(
11055 1 : &tenant,
11056 1 : TIMELINE_ID,
11057 1 : DEFAULT_PG_VERSION,
11058 1 : specification,
11059 1 : &mut random,
11060 1 : &ctx,
11061 1 : )
11062 1 : .await?;
11063 1 :
11064 1 : // Now generate queries based on the interesting lsns that we've collected.
11065 1 : //
11066 1 : // While there's still room in the query, pick and interesting LSN and a random
11067 1 : // key. Then roll the dice to see if the next key should also be included in
11068 1 : // the query. When the roll fails, break the "batch" and pick another point in the
11069 1 : // (key, LSN) space.
11070 1 :
11071 1 : const PICK_NEXT_CHANCE: u8 = 50;
11072 1 : for _ in 0..queries {
11073 1000 : let query = {
11074 1000 : let mut keyspaces_at_lsn: HashMap<Lsn, KeySpaceRandomAccum> = HashMap::default();
11075 1000 : let mut used_keys: HashSet<Key> = HashSet::default();
11076 1 :
11077 22536 : while used_keys.len() < Timeline::MAX_GET_VECTORED_KEYS as usize {
11078 21536 : let selected_lsn = interesting_lsns.choose(&mut random).expect("not empty");
11079 21536 : let mut selected_key = start_key.add(random.gen_range(0..KEY_DIMENSION_SIZE));
11080 1 :
11081 37614 : while used_keys.len() < Timeline::MAX_GET_VECTORED_KEYS as usize {
11082 37093 : if used_keys.contains(&selected_key)
11083 32154 : || selected_key >= start_key.add(KEY_DIMENSION_SIZE)
11084 1 : {
11085 5093 : break;
11086 32000 : }
11087 32000 :
11088 32000 : keyspaces_at_lsn
11089 32000 : .entry(*selected_lsn)
11090 32000 : .or_default()
11091 32000 : .add_key(selected_key);
11092 32000 : used_keys.insert(selected_key);
11093 32000 :
11094 32000 : let pick_next = random.gen_range(0..=100) <= PICK_NEXT_CHANCE;
11095 32000 : if pick_next {
11096 16078 : selected_key = selected_key.next();
11097 16078 : } else {
11098 15922 : break;
11099 1 : }
11100 1 : }
11101 1 : }
11102 1 :
11103 1000 : VersionedKeySpaceQuery::scattered(
11104 1000 : keyspaces_at_lsn
11105 1000 : .into_iter()
11106 11917 : .map(|(lsn, acc)| (lsn, acc.to_keyspace()))
11107 1000 : .collect(),
11108 1000 : )
11109 1 : };
11110 1 :
11111 1 : // Run the query and validate the results
11112 1 :
11113 1000 : let results = tline
11114 1000 : .get_vectored(query.clone(), IoConcurrency::Sequential, &ctx)
11115 1000 : .await;
11116 1 :
11117 1000 : let blobs = match results {
11118 1000 : Ok(ok) => ok,
11119 1 : Err(err) => {
11120 0 : panic!("seed={seed} Error returned for query {query}: {err}");
11121 1 : }
11122 1 : };
11123 1 :
11124 32000 : for (key, key_res) in blobs.into_iter() {
11125 32000 : match key_res {
11126 32000 : Ok(blob) => {
11127 32000 : let requested_at_lsn = query.map_key_to_lsn(&key);
11128 32000 : let expected = storage.get(key, requested_at_lsn);
11129 32000 :
11130 32000 : if blob != expected {
11131 1 : tracing::error!(
11132 1 : "seed={seed} Mismatch for {key}@{requested_at_lsn} from query: {query}"
11133 1 : );
11134 32000 : }
11135 1 :
11136 32000 : assert_eq!(blob, expected);
11137 1 : }
11138 1 : Err(err) => {
11139 0 : let requested_at_lsn = query.map_key_to_lsn(&key);
11140 0 :
11141 0 : panic!(
11142 0 : "seed={seed} Error returned for {key}@{requested_at_lsn} from query {query}: {err}"
11143 0 : );
11144 1 : }
11145 1 : }
11146 1 : }
11147 1 : }
11148 1 :
11149 1 : Ok(())
11150 1 : }
11151 :
11152 107 : fn sort_layer_key(k1: &PersistentLayerKey, k2: &PersistentLayerKey) -> std::cmp::Ordering {
11153 107 : (
11154 107 : k1.is_delta,
11155 107 : k1.key_range.start,
11156 107 : k1.key_range.end,
11157 107 : k1.lsn_range.start,
11158 107 : k1.lsn_range.end,
11159 107 : )
11160 107 : .cmp(&(
11161 107 : k2.is_delta,
11162 107 : k2.key_range.start,
11163 107 : k2.key_range.end,
11164 107 : k2.lsn_range.start,
11165 107 : k2.lsn_range.end,
11166 107 : ))
11167 107 : }
11168 :
11169 12 : async fn inspect_and_sort(
11170 12 : tline: &Arc<Timeline>,
11171 12 : filter: Option<std::ops::Range<Key>>,
11172 12 : ) -> Vec<PersistentLayerKey> {
11173 12 : let mut all_layers = tline.inspect_historic_layers().await.unwrap();
11174 12 : if let Some(filter) = filter {
11175 54 : all_layers.retain(|layer| overlaps_with(&layer.key_range, &filter));
11176 11 : }
11177 12 : all_layers.sort_by(sort_layer_key);
11178 12 : all_layers
11179 12 : }
11180 :
11181 : #[cfg(feature = "testing")]
11182 11 : fn check_layer_map_key_eq(
11183 11 : mut left: Vec<PersistentLayerKey>,
11184 11 : mut right: Vec<PersistentLayerKey>,
11185 11 : ) {
11186 11 : left.sort_by(sort_layer_key);
11187 11 : right.sort_by(sort_layer_key);
11188 11 : if left != right {
11189 0 : eprintln!("---LEFT---");
11190 0 : for left in left.iter() {
11191 0 : eprintln!("{}", left);
11192 0 : }
11193 0 : eprintln!("---RIGHT---");
11194 0 : for right in right.iter() {
11195 0 : eprintln!("{}", right);
11196 0 : }
11197 0 : assert_eq!(left, right);
11198 11 : }
11199 11 : }
11200 :
11201 : #[cfg(feature = "testing")]
11202 : #[tokio::test]
11203 1 : async fn test_simple_partial_bottom_most_compaction() -> anyhow::Result<()> {
11204 1 : let harness = TenantHarness::create("test_simple_partial_bottom_most_compaction").await?;
11205 1 : let (tenant, ctx) = harness.load().await;
11206 1 :
11207 91 : fn get_key(id: u32) -> Key {
11208 91 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
11209 91 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
11210 91 : key.field6 = id;
11211 91 : key
11212 91 : }
11213 1 :
11214 1 : // img layer at 0x10
11215 1 : let img_layer = (0..10)
11216 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
11217 1 : .collect_vec();
11218 1 :
11219 1 : let delta1 = vec![
11220 1 : (
11221 1 : get_key(1),
11222 1 : Lsn(0x20),
11223 1 : Value::Image(Bytes::from("value 1@0x20")),
11224 1 : ),
11225 1 : (
11226 1 : get_key(2),
11227 1 : Lsn(0x30),
11228 1 : Value::Image(Bytes::from("value 2@0x30")),
11229 1 : ),
11230 1 : (
11231 1 : get_key(3),
11232 1 : Lsn(0x40),
11233 1 : Value::Image(Bytes::from("value 3@0x40")),
11234 1 : ),
11235 1 : ];
11236 1 : let delta2 = vec![
11237 1 : (
11238 1 : get_key(5),
11239 1 : Lsn(0x20),
11240 1 : Value::Image(Bytes::from("value 5@0x20")),
11241 1 : ),
11242 1 : (
11243 1 : get_key(6),
11244 1 : Lsn(0x20),
11245 1 : Value::Image(Bytes::from("value 6@0x20")),
11246 1 : ),
11247 1 : ];
11248 1 : let delta3 = vec![
11249 1 : (
11250 1 : get_key(8),
11251 1 : Lsn(0x48),
11252 1 : Value::Image(Bytes::from("value 8@0x48")),
11253 1 : ),
11254 1 : (
11255 1 : get_key(9),
11256 1 : Lsn(0x48),
11257 1 : Value::Image(Bytes::from("value 9@0x48")),
11258 1 : ),
11259 1 : ];
11260 1 :
11261 1 : let tline = tenant
11262 1 : .create_test_timeline_with_layers(
11263 1 : TIMELINE_ID,
11264 1 : Lsn(0x10),
11265 1 : DEFAULT_PG_VERSION,
11266 1 : &ctx,
11267 1 : vec![], // in-memory layers
11268 1 : vec![
11269 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
11270 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
11271 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
11272 1 : ], // delta layers
11273 1 : vec![(Lsn(0x10), img_layer)], // image layers
11274 1 : Lsn(0x50),
11275 1 : )
11276 1 : .await?;
11277 1 :
11278 1 : {
11279 1 : tline
11280 1 : .applied_gc_cutoff_lsn
11281 1 : .lock_for_write()
11282 1 : .store_and_unlock(Lsn(0x30))
11283 1 : .wait()
11284 1 : .await;
11285 1 : // Update GC info
11286 1 : let mut guard = tline.gc_info.write().unwrap();
11287 1 : *guard = GcInfo {
11288 1 : retain_lsns: vec![(Lsn(0x20), tline.timeline_id, MaybeOffloaded::No)],
11289 1 : cutoffs: GcCutoffs {
11290 1 : time: Some(Lsn(0x30)),
11291 1 : space: Lsn(0x30),
11292 1 : },
11293 1 : leases: Default::default(),
11294 1 : within_ancestor_pitr: false,
11295 1 : };
11296 1 : }
11297 1 :
11298 1 : let cancel = CancellationToken::new();
11299 1 :
11300 1 : // Do a partial compaction on key range 0..2
11301 1 : tline
11302 1 : .compact_with_gc(
11303 1 : &cancel,
11304 1 : CompactOptions {
11305 1 : flags: EnumSet::new(),
11306 1 : compact_key_range: Some((get_key(0)..get_key(2)).into()),
11307 1 : ..Default::default()
11308 1 : },
11309 1 : &ctx,
11310 1 : )
11311 1 : .await
11312 1 : .unwrap();
11313 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11314 1 : check_layer_map_key_eq(
11315 1 : all_layers,
11316 1 : vec![
11317 1 : // newly-generated image layer for the partial compaction range 0-2
11318 1 : PersistentLayerKey {
11319 1 : key_range: get_key(0)..get_key(2),
11320 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11321 1 : is_delta: false,
11322 1 : },
11323 1 : PersistentLayerKey {
11324 1 : key_range: get_key(0)..get_key(10),
11325 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
11326 1 : is_delta: false,
11327 1 : },
11328 1 : // delta1 is split and the second part is rewritten
11329 1 : PersistentLayerKey {
11330 1 : key_range: get_key(2)..get_key(4),
11331 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11332 1 : is_delta: true,
11333 1 : },
11334 1 : PersistentLayerKey {
11335 1 : key_range: get_key(5)..get_key(7),
11336 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11337 1 : is_delta: true,
11338 1 : },
11339 1 : PersistentLayerKey {
11340 1 : key_range: get_key(8)..get_key(10),
11341 1 : lsn_range: Lsn(0x48)..Lsn(0x50),
11342 1 : is_delta: true,
11343 1 : },
11344 1 : ],
11345 1 : );
11346 1 :
11347 1 : // Do a partial compaction on key range 2..4
11348 1 : tline
11349 1 : .compact_with_gc(
11350 1 : &cancel,
11351 1 : CompactOptions {
11352 1 : flags: EnumSet::new(),
11353 1 : compact_key_range: Some((get_key(2)..get_key(4)).into()),
11354 1 : ..Default::default()
11355 1 : },
11356 1 : &ctx,
11357 1 : )
11358 1 : .await
11359 1 : .unwrap();
11360 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11361 1 : check_layer_map_key_eq(
11362 1 : all_layers,
11363 1 : vec![
11364 1 : PersistentLayerKey {
11365 1 : key_range: get_key(0)..get_key(2),
11366 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11367 1 : is_delta: false,
11368 1 : },
11369 1 : PersistentLayerKey {
11370 1 : key_range: get_key(0)..get_key(10),
11371 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
11372 1 : is_delta: false,
11373 1 : },
11374 1 : // image layer generated for the compaction range 2-4
11375 1 : PersistentLayerKey {
11376 1 : key_range: get_key(2)..get_key(4),
11377 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11378 1 : is_delta: false,
11379 1 : },
11380 1 : // we have key2/key3 above the retain_lsn, so we still need this delta layer
11381 1 : PersistentLayerKey {
11382 1 : key_range: get_key(2)..get_key(4),
11383 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11384 1 : is_delta: true,
11385 1 : },
11386 1 : PersistentLayerKey {
11387 1 : key_range: get_key(5)..get_key(7),
11388 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11389 1 : is_delta: true,
11390 1 : },
11391 1 : PersistentLayerKey {
11392 1 : key_range: get_key(8)..get_key(10),
11393 1 : lsn_range: Lsn(0x48)..Lsn(0x50),
11394 1 : is_delta: true,
11395 1 : },
11396 1 : ],
11397 1 : );
11398 1 :
11399 1 : // Do a partial compaction on key range 4..9
11400 1 : tline
11401 1 : .compact_with_gc(
11402 1 : &cancel,
11403 1 : CompactOptions {
11404 1 : flags: EnumSet::new(),
11405 1 : compact_key_range: Some((get_key(4)..get_key(9)).into()),
11406 1 : ..Default::default()
11407 1 : },
11408 1 : &ctx,
11409 1 : )
11410 1 : .await
11411 1 : .unwrap();
11412 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11413 1 : check_layer_map_key_eq(
11414 1 : all_layers,
11415 1 : vec![
11416 1 : PersistentLayerKey {
11417 1 : key_range: get_key(0)..get_key(2),
11418 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11419 1 : is_delta: false,
11420 1 : },
11421 1 : PersistentLayerKey {
11422 1 : key_range: get_key(0)..get_key(10),
11423 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
11424 1 : is_delta: false,
11425 1 : },
11426 1 : PersistentLayerKey {
11427 1 : key_range: get_key(2)..get_key(4),
11428 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11429 1 : is_delta: false,
11430 1 : },
11431 1 : PersistentLayerKey {
11432 1 : key_range: get_key(2)..get_key(4),
11433 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11434 1 : is_delta: true,
11435 1 : },
11436 1 : // image layer generated for this compaction range
11437 1 : PersistentLayerKey {
11438 1 : key_range: get_key(4)..get_key(9),
11439 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11440 1 : is_delta: false,
11441 1 : },
11442 1 : PersistentLayerKey {
11443 1 : key_range: get_key(8)..get_key(10),
11444 1 : lsn_range: Lsn(0x48)..Lsn(0x50),
11445 1 : is_delta: true,
11446 1 : },
11447 1 : ],
11448 1 : );
11449 1 :
11450 1 : // Do a partial compaction on key range 9..10
11451 1 : tline
11452 1 : .compact_with_gc(
11453 1 : &cancel,
11454 1 : CompactOptions {
11455 1 : flags: EnumSet::new(),
11456 1 : compact_key_range: Some((get_key(9)..get_key(10)).into()),
11457 1 : ..Default::default()
11458 1 : },
11459 1 : &ctx,
11460 1 : )
11461 1 : .await
11462 1 : .unwrap();
11463 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11464 1 : check_layer_map_key_eq(
11465 1 : all_layers,
11466 1 : vec![
11467 1 : PersistentLayerKey {
11468 1 : key_range: get_key(0)..get_key(2),
11469 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11470 1 : is_delta: false,
11471 1 : },
11472 1 : PersistentLayerKey {
11473 1 : key_range: get_key(0)..get_key(10),
11474 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
11475 1 : is_delta: false,
11476 1 : },
11477 1 : PersistentLayerKey {
11478 1 : key_range: get_key(2)..get_key(4),
11479 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11480 1 : is_delta: false,
11481 1 : },
11482 1 : PersistentLayerKey {
11483 1 : key_range: get_key(2)..get_key(4),
11484 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11485 1 : is_delta: true,
11486 1 : },
11487 1 : PersistentLayerKey {
11488 1 : key_range: get_key(4)..get_key(9),
11489 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11490 1 : is_delta: false,
11491 1 : },
11492 1 : // image layer generated for the compaction range
11493 1 : PersistentLayerKey {
11494 1 : key_range: get_key(9)..get_key(10),
11495 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11496 1 : is_delta: false,
11497 1 : },
11498 1 : PersistentLayerKey {
11499 1 : key_range: get_key(8)..get_key(10),
11500 1 : lsn_range: Lsn(0x48)..Lsn(0x50),
11501 1 : is_delta: true,
11502 1 : },
11503 1 : ],
11504 1 : );
11505 1 :
11506 1 : // Do a partial compaction on key range 0..10, all image layers below LSN 20 can be replaced with new ones.
11507 1 : tline
11508 1 : .compact_with_gc(
11509 1 : &cancel,
11510 1 : CompactOptions {
11511 1 : flags: EnumSet::new(),
11512 1 : compact_key_range: Some((get_key(0)..get_key(10)).into()),
11513 1 : ..Default::default()
11514 1 : },
11515 1 : &ctx,
11516 1 : )
11517 1 : .await
11518 1 : .unwrap();
11519 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11520 1 : check_layer_map_key_eq(
11521 1 : all_layers,
11522 1 : vec![
11523 1 : // aha, we removed all unnecessary image/delta layers and got a very clean layer map!
11524 1 : PersistentLayerKey {
11525 1 : key_range: get_key(0)..get_key(10),
11526 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11527 1 : is_delta: false,
11528 1 : },
11529 1 : PersistentLayerKey {
11530 1 : key_range: get_key(2)..get_key(4),
11531 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11532 1 : is_delta: true,
11533 1 : },
11534 1 : PersistentLayerKey {
11535 1 : key_range: get_key(8)..get_key(10),
11536 1 : lsn_range: Lsn(0x48)..Lsn(0x50),
11537 1 : is_delta: true,
11538 1 : },
11539 1 : ],
11540 1 : );
11541 1 : Ok(())
11542 1 : }
11543 :
11544 : #[cfg(feature = "testing")]
11545 : #[tokio::test]
11546 1 : async fn test_timeline_offload_retain_lsn() -> anyhow::Result<()> {
11547 1 : let harness = TenantHarness::create("test_timeline_offload_retain_lsn")
11548 1 : .await
11549 1 : .unwrap();
11550 1 : let (tenant, ctx) = harness.load().await;
11551 1 : let tline_parent = tenant
11552 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
11553 1 : .await
11554 1 : .unwrap();
11555 1 : let tline_child = tenant
11556 1 : .branch_timeline_test(&tline_parent, NEW_TIMELINE_ID, Some(Lsn(0x20)), &ctx)
11557 1 : .await
11558 1 : .unwrap();
11559 1 : {
11560 1 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
11561 1 : assert_eq!(
11562 1 : gc_info_parent.retain_lsns,
11563 1 : vec![(Lsn(0x20), tline_child.timeline_id, MaybeOffloaded::No)]
11564 1 : );
11565 1 : }
11566 1 : // We have to directly call the remote_client instead of using the archive function to avoid constructing broker client...
11567 1 : tline_child
11568 1 : .remote_client
11569 1 : .schedule_index_upload_for_timeline_archival_state(TimelineArchivalState::Archived)
11570 1 : .unwrap();
11571 1 : tline_child.remote_client.wait_completion().await.unwrap();
11572 1 : offload_timeline(&tenant, &tline_child)
11573 1 : .instrument(tracing::info_span!(parent: None, "offload_test", tenant_id=%"test", shard_id=%"test", timeline_id=%"test"))
11574 1 : .await.unwrap();
11575 1 : let child_timeline_id = tline_child.timeline_id;
11576 1 : Arc::try_unwrap(tline_child).unwrap();
11577 1 :
11578 1 : {
11579 1 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
11580 1 : assert_eq!(
11581 1 : gc_info_parent.retain_lsns,
11582 1 : vec![(Lsn(0x20), child_timeline_id, MaybeOffloaded::Yes)]
11583 1 : );
11584 1 : }
11585 1 :
11586 1 : tenant
11587 1 : .get_offloaded_timeline(child_timeline_id)
11588 1 : .unwrap()
11589 1 : .defuse_for_tenant_drop();
11590 1 :
11591 1 : Ok(())
11592 1 : }
11593 :
11594 : #[cfg(feature = "testing")]
11595 : #[tokio::test]
11596 1 : async fn test_simple_bottom_most_compaction_above_lsn() -> anyhow::Result<()> {
11597 1 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_above_lsn").await?;
11598 1 : let (tenant, ctx) = harness.load().await;
11599 1 :
11600 148 : fn get_key(id: u32) -> Key {
11601 148 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
11602 148 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
11603 148 : key.field6 = id;
11604 148 : key
11605 148 : }
11606 1 :
11607 1 : let img_layer = (0..10)
11608 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
11609 1 : .collect_vec();
11610 1 :
11611 1 : let delta1 = vec![(
11612 1 : get_key(1),
11613 1 : Lsn(0x20),
11614 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
11615 1 : )];
11616 1 : let delta4 = vec![(
11617 1 : get_key(1),
11618 1 : Lsn(0x28),
11619 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
11620 1 : )];
11621 1 : let delta2 = vec![
11622 1 : (
11623 1 : get_key(1),
11624 1 : Lsn(0x30),
11625 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
11626 1 : ),
11627 1 : (
11628 1 : get_key(1),
11629 1 : Lsn(0x38),
11630 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
11631 1 : ),
11632 1 : ];
11633 1 : let delta3 = vec![
11634 1 : (
11635 1 : get_key(8),
11636 1 : Lsn(0x48),
11637 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
11638 1 : ),
11639 1 : (
11640 1 : get_key(9),
11641 1 : Lsn(0x48),
11642 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
11643 1 : ),
11644 1 : ];
11645 1 :
11646 1 : let tline = tenant
11647 1 : .create_test_timeline_with_layers(
11648 1 : TIMELINE_ID,
11649 1 : Lsn(0x10),
11650 1 : DEFAULT_PG_VERSION,
11651 1 : &ctx,
11652 1 : vec![], // in-memory layers
11653 1 : vec![
11654 1 : // delta1/2/4 only contain a single key but multiple updates
11655 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x28), delta1),
11656 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
11657 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x28)..Lsn(0x30), delta4),
11658 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta3),
11659 1 : ], // delta layers
11660 1 : vec![(Lsn(0x10), img_layer)], // image layers
11661 1 : Lsn(0x50),
11662 1 : )
11663 1 : .await?;
11664 1 : {
11665 1 : tline
11666 1 : .applied_gc_cutoff_lsn
11667 1 : .lock_for_write()
11668 1 : .store_and_unlock(Lsn(0x30))
11669 1 : .wait()
11670 1 : .await;
11671 1 : // Update GC info
11672 1 : let mut guard = tline.gc_info.write().unwrap();
11673 1 : *guard = GcInfo {
11674 1 : retain_lsns: vec![
11675 1 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
11676 1 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
11677 1 : ],
11678 1 : cutoffs: GcCutoffs {
11679 1 : time: Some(Lsn(0x30)),
11680 1 : space: Lsn(0x30),
11681 1 : },
11682 1 : leases: Default::default(),
11683 1 : within_ancestor_pitr: false,
11684 1 : };
11685 1 : }
11686 1 :
11687 1 : let expected_result = [
11688 1 : Bytes::from_static(b"value 0@0x10"),
11689 1 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
11690 1 : Bytes::from_static(b"value 2@0x10"),
11691 1 : Bytes::from_static(b"value 3@0x10"),
11692 1 : Bytes::from_static(b"value 4@0x10"),
11693 1 : Bytes::from_static(b"value 5@0x10"),
11694 1 : Bytes::from_static(b"value 6@0x10"),
11695 1 : Bytes::from_static(b"value 7@0x10"),
11696 1 : Bytes::from_static(b"value 8@0x10@0x48"),
11697 1 : Bytes::from_static(b"value 9@0x10@0x48"),
11698 1 : ];
11699 1 :
11700 1 : let expected_result_at_gc_horizon = [
11701 1 : Bytes::from_static(b"value 0@0x10"),
11702 1 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
11703 1 : Bytes::from_static(b"value 2@0x10"),
11704 1 : Bytes::from_static(b"value 3@0x10"),
11705 1 : Bytes::from_static(b"value 4@0x10"),
11706 1 : Bytes::from_static(b"value 5@0x10"),
11707 1 : Bytes::from_static(b"value 6@0x10"),
11708 1 : Bytes::from_static(b"value 7@0x10"),
11709 1 : Bytes::from_static(b"value 8@0x10"),
11710 1 : Bytes::from_static(b"value 9@0x10"),
11711 1 : ];
11712 1 :
11713 1 : let expected_result_at_lsn_20 = [
11714 1 : Bytes::from_static(b"value 0@0x10"),
11715 1 : Bytes::from_static(b"value 1@0x10@0x20"),
11716 1 : Bytes::from_static(b"value 2@0x10"),
11717 1 : Bytes::from_static(b"value 3@0x10"),
11718 1 : Bytes::from_static(b"value 4@0x10"),
11719 1 : Bytes::from_static(b"value 5@0x10"),
11720 1 : Bytes::from_static(b"value 6@0x10"),
11721 1 : Bytes::from_static(b"value 7@0x10"),
11722 1 : Bytes::from_static(b"value 8@0x10"),
11723 1 : Bytes::from_static(b"value 9@0x10"),
11724 1 : ];
11725 1 :
11726 1 : let expected_result_at_lsn_10 = [
11727 1 : Bytes::from_static(b"value 0@0x10"),
11728 1 : Bytes::from_static(b"value 1@0x10"),
11729 1 : Bytes::from_static(b"value 2@0x10"),
11730 1 : Bytes::from_static(b"value 3@0x10"),
11731 1 : Bytes::from_static(b"value 4@0x10"),
11732 1 : Bytes::from_static(b"value 5@0x10"),
11733 1 : Bytes::from_static(b"value 6@0x10"),
11734 1 : Bytes::from_static(b"value 7@0x10"),
11735 1 : Bytes::from_static(b"value 8@0x10"),
11736 1 : Bytes::from_static(b"value 9@0x10"),
11737 1 : ];
11738 1 :
11739 3 : let verify_result = || async {
11740 3 : let gc_horizon = {
11741 3 : let gc_info = tline.gc_info.read().unwrap();
11742 3 : gc_info.cutoffs.time.unwrap_or_default()
11743 1 : };
11744 33 : for idx in 0..10 {
11745 30 : assert_eq!(
11746 30 : tline
11747 30 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
11748 30 : .await
11749 30 : .unwrap(),
11750 30 : &expected_result[idx]
11751 1 : );
11752 30 : assert_eq!(
11753 30 : tline
11754 30 : .get(get_key(idx as u32), gc_horizon, &ctx)
11755 30 : .await
11756 30 : .unwrap(),
11757 30 : &expected_result_at_gc_horizon[idx]
11758 1 : );
11759 30 : assert_eq!(
11760 30 : tline
11761 30 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
11762 30 : .await
11763 30 : .unwrap(),
11764 30 : &expected_result_at_lsn_20[idx]
11765 1 : );
11766 30 : assert_eq!(
11767 30 : tline
11768 30 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
11769 30 : .await
11770 30 : .unwrap(),
11771 30 : &expected_result_at_lsn_10[idx]
11772 1 : );
11773 1 : }
11774 6 : };
11775 1 :
11776 1 : verify_result().await;
11777 1 :
11778 1 : let cancel = CancellationToken::new();
11779 1 : tline
11780 1 : .compact_with_gc(
11781 1 : &cancel,
11782 1 : CompactOptions {
11783 1 : compact_lsn_range: Some(CompactLsnRange::above(Lsn(0x28))),
11784 1 : ..Default::default()
11785 1 : },
11786 1 : &ctx,
11787 1 : )
11788 1 : .await
11789 1 : .unwrap();
11790 1 : verify_result().await;
11791 1 :
11792 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11793 1 : check_layer_map_key_eq(
11794 1 : all_layers,
11795 1 : vec![
11796 1 : // The original image layer, not compacted
11797 1 : PersistentLayerKey {
11798 1 : key_range: get_key(0)..get_key(10),
11799 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
11800 1 : is_delta: false,
11801 1 : },
11802 1 : // Delta layer below the specified above_lsn not compacted
11803 1 : PersistentLayerKey {
11804 1 : key_range: get_key(1)..get_key(2),
11805 1 : lsn_range: Lsn(0x20)..Lsn(0x28),
11806 1 : is_delta: true,
11807 1 : },
11808 1 : // Delta layer compacted above the LSN
11809 1 : PersistentLayerKey {
11810 1 : key_range: get_key(1)..get_key(10),
11811 1 : lsn_range: Lsn(0x28)..Lsn(0x50),
11812 1 : is_delta: true,
11813 1 : },
11814 1 : ],
11815 1 : );
11816 1 :
11817 1 : // compact again
11818 1 : tline
11819 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
11820 1 : .await
11821 1 : .unwrap();
11822 1 : verify_result().await;
11823 1 :
11824 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11825 1 : check_layer_map_key_eq(
11826 1 : all_layers,
11827 1 : vec![
11828 1 : // The compacted image layer (full key range)
11829 1 : PersistentLayerKey {
11830 1 : key_range: Key::MIN..Key::MAX,
11831 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
11832 1 : is_delta: false,
11833 1 : },
11834 1 : // All other data in the delta layer
11835 1 : PersistentLayerKey {
11836 1 : key_range: get_key(1)..get_key(10),
11837 1 : lsn_range: Lsn(0x10)..Lsn(0x50),
11838 1 : is_delta: true,
11839 1 : },
11840 1 : ],
11841 1 : );
11842 1 :
11843 1 : Ok(())
11844 1 : }
11845 :
11846 : #[cfg(feature = "testing")]
11847 : #[tokio::test]
11848 1 : async fn test_simple_bottom_most_compaction_rectangle() -> anyhow::Result<()> {
11849 1 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_rectangle").await?;
11850 1 : let (tenant, ctx) = harness.load().await;
11851 1 :
11852 254 : fn get_key(id: u32) -> Key {
11853 254 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
11854 254 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
11855 254 : key.field6 = id;
11856 254 : key
11857 254 : }
11858 1 :
11859 1 : let img_layer = (0..10)
11860 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
11861 1 : .collect_vec();
11862 1 :
11863 1 : let delta1 = vec![(
11864 1 : get_key(1),
11865 1 : Lsn(0x20),
11866 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
11867 1 : )];
11868 1 : let delta4 = vec![(
11869 1 : get_key(1),
11870 1 : Lsn(0x28),
11871 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
11872 1 : )];
11873 1 : let delta2 = vec![
11874 1 : (
11875 1 : get_key(1),
11876 1 : Lsn(0x30),
11877 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
11878 1 : ),
11879 1 : (
11880 1 : get_key(1),
11881 1 : Lsn(0x38),
11882 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
11883 1 : ),
11884 1 : ];
11885 1 : let delta3 = vec![
11886 1 : (
11887 1 : get_key(8),
11888 1 : Lsn(0x48),
11889 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
11890 1 : ),
11891 1 : (
11892 1 : get_key(9),
11893 1 : Lsn(0x48),
11894 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
11895 1 : ),
11896 1 : ];
11897 1 :
11898 1 : let tline = tenant
11899 1 : .create_test_timeline_with_layers(
11900 1 : TIMELINE_ID,
11901 1 : Lsn(0x10),
11902 1 : DEFAULT_PG_VERSION,
11903 1 : &ctx,
11904 1 : vec![], // in-memory layers
11905 1 : vec![
11906 1 : // delta1/2/4 only contain a single key but multiple updates
11907 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x28), delta1),
11908 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
11909 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x28)..Lsn(0x30), delta4),
11910 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta3),
11911 1 : ], // delta layers
11912 1 : vec![(Lsn(0x10), img_layer)], // image layers
11913 1 : Lsn(0x50),
11914 1 : )
11915 1 : .await?;
11916 1 : {
11917 1 : tline
11918 1 : .applied_gc_cutoff_lsn
11919 1 : .lock_for_write()
11920 1 : .store_and_unlock(Lsn(0x30))
11921 1 : .wait()
11922 1 : .await;
11923 1 : // Update GC info
11924 1 : let mut guard = tline.gc_info.write().unwrap();
11925 1 : *guard = GcInfo {
11926 1 : retain_lsns: vec![
11927 1 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
11928 1 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
11929 1 : ],
11930 1 : cutoffs: GcCutoffs {
11931 1 : time: Some(Lsn(0x30)),
11932 1 : space: Lsn(0x30),
11933 1 : },
11934 1 : leases: Default::default(),
11935 1 : within_ancestor_pitr: false,
11936 1 : };
11937 1 : }
11938 1 :
11939 1 : let expected_result = [
11940 1 : Bytes::from_static(b"value 0@0x10"),
11941 1 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
11942 1 : Bytes::from_static(b"value 2@0x10"),
11943 1 : Bytes::from_static(b"value 3@0x10"),
11944 1 : Bytes::from_static(b"value 4@0x10"),
11945 1 : Bytes::from_static(b"value 5@0x10"),
11946 1 : Bytes::from_static(b"value 6@0x10"),
11947 1 : Bytes::from_static(b"value 7@0x10"),
11948 1 : Bytes::from_static(b"value 8@0x10@0x48"),
11949 1 : Bytes::from_static(b"value 9@0x10@0x48"),
11950 1 : ];
11951 1 :
11952 1 : let expected_result_at_gc_horizon = [
11953 1 : Bytes::from_static(b"value 0@0x10"),
11954 1 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
11955 1 : Bytes::from_static(b"value 2@0x10"),
11956 1 : Bytes::from_static(b"value 3@0x10"),
11957 1 : Bytes::from_static(b"value 4@0x10"),
11958 1 : Bytes::from_static(b"value 5@0x10"),
11959 1 : Bytes::from_static(b"value 6@0x10"),
11960 1 : Bytes::from_static(b"value 7@0x10"),
11961 1 : Bytes::from_static(b"value 8@0x10"),
11962 1 : Bytes::from_static(b"value 9@0x10"),
11963 1 : ];
11964 1 :
11965 1 : let expected_result_at_lsn_20 = [
11966 1 : Bytes::from_static(b"value 0@0x10"),
11967 1 : Bytes::from_static(b"value 1@0x10@0x20"),
11968 1 : Bytes::from_static(b"value 2@0x10"),
11969 1 : Bytes::from_static(b"value 3@0x10"),
11970 1 : Bytes::from_static(b"value 4@0x10"),
11971 1 : Bytes::from_static(b"value 5@0x10"),
11972 1 : Bytes::from_static(b"value 6@0x10"),
11973 1 : Bytes::from_static(b"value 7@0x10"),
11974 1 : Bytes::from_static(b"value 8@0x10"),
11975 1 : Bytes::from_static(b"value 9@0x10"),
11976 1 : ];
11977 1 :
11978 1 : let expected_result_at_lsn_10 = [
11979 1 : Bytes::from_static(b"value 0@0x10"),
11980 1 : Bytes::from_static(b"value 1@0x10"),
11981 1 : Bytes::from_static(b"value 2@0x10"),
11982 1 : Bytes::from_static(b"value 3@0x10"),
11983 1 : Bytes::from_static(b"value 4@0x10"),
11984 1 : Bytes::from_static(b"value 5@0x10"),
11985 1 : Bytes::from_static(b"value 6@0x10"),
11986 1 : Bytes::from_static(b"value 7@0x10"),
11987 1 : Bytes::from_static(b"value 8@0x10"),
11988 1 : Bytes::from_static(b"value 9@0x10"),
11989 1 : ];
11990 1 :
11991 5 : let verify_result = || async {
11992 5 : let gc_horizon = {
11993 5 : let gc_info = tline.gc_info.read().unwrap();
11994 5 : gc_info.cutoffs.time.unwrap_or_default()
11995 1 : };
11996 55 : for idx in 0..10 {
11997 50 : assert_eq!(
11998 50 : tline
11999 50 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
12000 50 : .await
12001 50 : .unwrap(),
12002 50 : &expected_result[idx]
12003 1 : );
12004 50 : assert_eq!(
12005 50 : tline
12006 50 : .get(get_key(idx as u32), gc_horizon, &ctx)
12007 50 : .await
12008 50 : .unwrap(),
12009 50 : &expected_result_at_gc_horizon[idx]
12010 1 : );
12011 50 : assert_eq!(
12012 50 : tline
12013 50 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
12014 50 : .await
12015 50 : .unwrap(),
12016 50 : &expected_result_at_lsn_20[idx]
12017 1 : );
12018 50 : assert_eq!(
12019 50 : tline
12020 50 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
12021 50 : .await
12022 50 : .unwrap(),
12023 50 : &expected_result_at_lsn_10[idx]
12024 1 : );
12025 1 : }
12026 10 : };
12027 1 :
12028 1 : verify_result().await;
12029 1 :
12030 1 : let cancel = CancellationToken::new();
12031 1 :
12032 1 : tline
12033 1 : .compact_with_gc(
12034 1 : &cancel,
12035 1 : CompactOptions {
12036 1 : compact_key_range: Some((get_key(0)..get_key(2)).into()),
12037 1 : compact_lsn_range: Some((Lsn(0x20)..Lsn(0x28)).into()),
12038 1 : ..Default::default()
12039 1 : },
12040 1 : &ctx,
12041 1 : )
12042 1 : .await
12043 1 : .unwrap();
12044 1 : verify_result().await;
12045 1 :
12046 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12047 1 : check_layer_map_key_eq(
12048 1 : all_layers,
12049 1 : vec![
12050 1 : // The original image layer, not compacted
12051 1 : PersistentLayerKey {
12052 1 : key_range: get_key(0)..get_key(10),
12053 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
12054 1 : is_delta: false,
12055 1 : },
12056 1 : // According the selection logic, we select all layers with start key <= 0x28, so we would merge the layer 0x20-0x28 and
12057 1 : // the layer 0x28-0x30 into one.
12058 1 : PersistentLayerKey {
12059 1 : key_range: get_key(1)..get_key(2),
12060 1 : lsn_range: Lsn(0x20)..Lsn(0x30),
12061 1 : is_delta: true,
12062 1 : },
12063 1 : // Above the upper bound and untouched
12064 1 : PersistentLayerKey {
12065 1 : key_range: get_key(1)..get_key(2),
12066 1 : lsn_range: Lsn(0x30)..Lsn(0x50),
12067 1 : is_delta: true,
12068 1 : },
12069 1 : // This layer is untouched
12070 1 : PersistentLayerKey {
12071 1 : key_range: get_key(8)..get_key(10),
12072 1 : lsn_range: Lsn(0x30)..Lsn(0x50),
12073 1 : is_delta: true,
12074 1 : },
12075 1 : ],
12076 1 : );
12077 1 :
12078 1 : tline
12079 1 : .compact_with_gc(
12080 1 : &cancel,
12081 1 : CompactOptions {
12082 1 : compact_key_range: Some((get_key(3)..get_key(8)).into()),
12083 1 : compact_lsn_range: Some((Lsn(0x28)..Lsn(0x40)).into()),
12084 1 : ..Default::default()
12085 1 : },
12086 1 : &ctx,
12087 1 : )
12088 1 : .await
12089 1 : .unwrap();
12090 1 : verify_result().await;
12091 1 :
12092 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12093 1 : check_layer_map_key_eq(
12094 1 : all_layers,
12095 1 : vec![
12096 1 : // The original image layer, not compacted
12097 1 : PersistentLayerKey {
12098 1 : key_range: get_key(0)..get_key(10),
12099 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
12100 1 : is_delta: false,
12101 1 : },
12102 1 : // Not in the compaction key range, uncompacted
12103 1 : PersistentLayerKey {
12104 1 : key_range: get_key(1)..get_key(2),
12105 1 : lsn_range: Lsn(0x20)..Lsn(0x30),
12106 1 : is_delta: true,
12107 1 : },
12108 1 : // Not in the compaction key range, uncompacted but need rewrite because the delta layer overlaps with the range
12109 1 : PersistentLayerKey {
12110 1 : key_range: get_key(1)..get_key(2),
12111 1 : lsn_range: Lsn(0x30)..Lsn(0x50),
12112 1 : is_delta: true,
12113 1 : },
12114 1 : // Note that when we specify the LSN upper bound to be 0x40, the compaction algorithm will not try to cut the layer
12115 1 : // horizontally in half. Instead, it will include all LSNs that overlap with 0x40. So the real max_lsn of the compaction
12116 1 : // becomes 0x50.
12117 1 : PersistentLayerKey {
12118 1 : key_range: get_key(8)..get_key(10),
12119 1 : lsn_range: Lsn(0x30)..Lsn(0x50),
12120 1 : is_delta: true,
12121 1 : },
12122 1 : ],
12123 1 : );
12124 1 :
12125 1 : // compact again
12126 1 : tline
12127 1 : .compact_with_gc(
12128 1 : &cancel,
12129 1 : CompactOptions {
12130 1 : compact_key_range: Some((get_key(0)..get_key(5)).into()),
12131 1 : compact_lsn_range: Some((Lsn(0x20)..Lsn(0x50)).into()),
12132 1 : ..Default::default()
12133 1 : },
12134 1 : &ctx,
12135 1 : )
12136 1 : .await
12137 1 : .unwrap();
12138 1 : verify_result().await;
12139 1 :
12140 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12141 1 : check_layer_map_key_eq(
12142 1 : all_layers,
12143 1 : vec![
12144 1 : // The original image layer, not compacted
12145 1 : PersistentLayerKey {
12146 1 : key_range: get_key(0)..get_key(10),
12147 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
12148 1 : is_delta: false,
12149 1 : },
12150 1 : // The range gets compacted
12151 1 : PersistentLayerKey {
12152 1 : key_range: get_key(1)..get_key(2),
12153 1 : lsn_range: Lsn(0x20)..Lsn(0x50),
12154 1 : is_delta: true,
12155 1 : },
12156 1 : // Not touched during this iteration of compaction
12157 1 : PersistentLayerKey {
12158 1 : key_range: get_key(8)..get_key(10),
12159 1 : lsn_range: Lsn(0x30)..Lsn(0x50),
12160 1 : is_delta: true,
12161 1 : },
12162 1 : ],
12163 1 : );
12164 1 :
12165 1 : // final full compaction
12166 1 : tline
12167 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
12168 1 : .await
12169 1 : .unwrap();
12170 1 : verify_result().await;
12171 1 :
12172 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12173 1 : check_layer_map_key_eq(
12174 1 : all_layers,
12175 1 : vec![
12176 1 : // The compacted image layer (full key range)
12177 1 : PersistentLayerKey {
12178 1 : key_range: Key::MIN..Key::MAX,
12179 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
12180 1 : is_delta: false,
12181 1 : },
12182 1 : // All other data in the delta layer
12183 1 : PersistentLayerKey {
12184 1 : key_range: get_key(1)..get_key(10),
12185 1 : lsn_range: Lsn(0x10)..Lsn(0x50),
12186 1 : is_delta: true,
12187 1 : },
12188 1 : ],
12189 1 : );
12190 1 :
12191 1 : Ok(())
12192 1 : }
12193 :
12194 : #[cfg(feature = "testing")]
12195 : #[tokio::test]
12196 1 : async fn test_bottom_most_compation_redo_failure() -> anyhow::Result<()> {
12197 1 : let harness = TenantHarness::create("test_bottom_most_compation_redo_failure").await?;
12198 1 : let (tenant, ctx) = harness.load().await;
12199 1 :
12200 13 : fn get_key(id: u32) -> Key {
12201 13 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
12202 13 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
12203 13 : key.field6 = id;
12204 13 : key
12205 13 : }
12206 1 :
12207 1 : let img_layer = (0..10)
12208 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
12209 1 : .collect_vec();
12210 1 :
12211 1 : let delta1 = vec![
12212 1 : (
12213 1 : get_key(1),
12214 1 : Lsn(0x20),
12215 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
12216 1 : ),
12217 1 : (
12218 1 : get_key(1),
12219 1 : Lsn(0x24),
12220 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x24")),
12221 1 : ),
12222 1 : (
12223 1 : get_key(1),
12224 1 : Lsn(0x28),
12225 1 : // This record will fail to redo
12226 1 : Value::WalRecord(NeonWalRecord::wal_append_conditional("@0x28", "???")),
12227 1 : ),
12228 1 : ];
12229 1 :
12230 1 : let tline = tenant
12231 1 : .create_test_timeline_with_layers(
12232 1 : TIMELINE_ID,
12233 1 : Lsn(0x10),
12234 1 : DEFAULT_PG_VERSION,
12235 1 : &ctx,
12236 1 : vec![], // in-memory layers
12237 1 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
12238 1 : Lsn(0x20)..Lsn(0x30),
12239 1 : delta1,
12240 1 : )], // delta layers
12241 1 : vec![(Lsn(0x10), img_layer)], // image layers
12242 1 : Lsn(0x50),
12243 1 : )
12244 1 : .await?;
12245 1 : {
12246 1 : tline
12247 1 : .applied_gc_cutoff_lsn
12248 1 : .lock_for_write()
12249 1 : .store_and_unlock(Lsn(0x30))
12250 1 : .wait()
12251 1 : .await;
12252 1 : // Update GC info
12253 1 : let mut guard = tline.gc_info.write().unwrap();
12254 1 : *guard = GcInfo {
12255 1 : retain_lsns: vec![],
12256 1 : cutoffs: GcCutoffs {
12257 1 : time: Some(Lsn(0x30)),
12258 1 : space: Lsn(0x30),
12259 1 : },
12260 1 : leases: Default::default(),
12261 1 : within_ancestor_pitr: false,
12262 1 : };
12263 1 : }
12264 1 :
12265 1 : let cancel = CancellationToken::new();
12266 1 :
12267 1 : // Compaction will fail, but should not fire any critical error.
12268 1 : // Gc-compaction currently cannot figure out what keys are not in the keyspace during the compaction
12269 1 : // process. It will always try to redo the logs it reads and if it doesn't work, fail the entire
12270 1 : // compaction job. Tracked in <https://github.com/neondatabase/neon/issues/10395>.
12271 1 : let res = tline
12272 1 : .compact_with_gc(
12273 1 : &cancel,
12274 1 : CompactOptions {
12275 1 : compact_key_range: None,
12276 1 : compact_lsn_range: None,
12277 1 : ..Default::default()
12278 1 : },
12279 1 : &ctx,
12280 1 : )
12281 1 : .await;
12282 1 : assert!(res.is_err());
12283 1 :
12284 1 : Ok(())
12285 1 : }
12286 :
12287 : #[cfg(feature = "testing")]
12288 : #[tokio::test]
12289 1 : async fn test_synthetic_size_calculation_with_invisible_branches() -> anyhow::Result<()> {
12290 1 : use pageserver_api::models::TimelineVisibilityState;
12291 1 :
12292 1 : use crate::tenant::size::gather_inputs;
12293 1 :
12294 1 : let tenant_conf = pageserver_api::models::TenantConfig {
12295 1 : // Ensure that we don't compute gc_cutoffs (which needs reading the layer files)
12296 1 : pitr_interval: Some(Duration::ZERO),
12297 1 : ..Default::default()
12298 1 : };
12299 1 : let harness = TenantHarness::create_custom(
12300 1 : "test_synthetic_size_calculation_with_invisible_branches",
12301 1 : tenant_conf,
12302 1 : TenantId::generate(),
12303 1 : ShardIdentity::unsharded(),
12304 1 : Generation::new(0xdeadbeef),
12305 1 : )
12306 1 : .await?;
12307 1 : let (tenant, ctx) = harness.load().await;
12308 1 : let main_tline = tenant
12309 1 : .create_test_timeline_with_layers(
12310 1 : TIMELINE_ID,
12311 1 : Lsn(0x10),
12312 1 : DEFAULT_PG_VERSION,
12313 1 : &ctx,
12314 1 : vec![],
12315 1 : vec![],
12316 1 : vec![],
12317 1 : Lsn(0x100),
12318 1 : )
12319 1 : .await?;
12320 1 :
12321 1 : let snapshot1 = TimelineId::from_array(hex!("11223344556677881122334455667790"));
12322 1 : tenant
12323 1 : .branch_timeline_test_with_layers(
12324 1 : &main_tline,
12325 1 : snapshot1,
12326 1 : Some(Lsn(0x20)),
12327 1 : &ctx,
12328 1 : vec![],
12329 1 : vec![],
12330 1 : Lsn(0x50),
12331 1 : )
12332 1 : .await?;
12333 1 : let snapshot2 = TimelineId::from_array(hex!("11223344556677881122334455667791"));
12334 1 : tenant
12335 1 : .branch_timeline_test_with_layers(
12336 1 : &main_tline,
12337 1 : snapshot2,
12338 1 : Some(Lsn(0x30)),
12339 1 : &ctx,
12340 1 : vec![],
12341 1 : vec![],
12342 1 : Lsn(0x50),
12343 1 : )
12344 1 : .await?;
12345 1 : let snapshot3 = TimelineId::from_array(hex!("11223344556677881122334455667792"));
12346 1 : tenant
12347 1 : .branch_timeline_test_with_layers(
12348 1 : &main_tline,
12349 1 : snapshot3,
12350 1 : Some(Lsn(0x40)),
12351 1 : &ctx,
12352 1 : vec![],
12353 1 : vec![],
12354 1 : Lsn(0x50),
12355 1 : )
12356 1 : .await?;
12357 1 : let limit = Arc::new(Semaphore::new(1));
12358 1 : let max_retention_period = None;
12359 1 : let mut logical_size_cache = HashMap::new();
12360 1 : let cause = LogicalSizeCalculationCause::EvictionTaskImitation;
12361 1 : let cancel = CancellationToken::new();
12362 1 :
12363 1 : let inputs = gather_inputs(
12364 1 : &tenant,
12365 1 : &limit,
12366 1 : max_retention_period,
12367 1 : &mut logical_size_cache,
12368 1 : cause,
12369 1 : &cancel,
12370 1 : &ctx,
12371 1 : )
12372 1 : .instrument(info_span!(
12373 1 : "gather_inputs",
12374 1 : tenant_id = "unknown",
12375 1 : shard_id = "unknown",
12376 1 : ))
12377 1 : .await?;
12378 1 : use crate::tenant::size::{LsnKind, ModelInputs, SegmentMeta};
12379 1 : use LsnKind::*;
12380 1 : use tenant_size_model::Segment;
12381 1 : let ModelInputs { mut segments, .. } = inputs;
12382 15 : segments.retain(|s| s.timeline_id == TIMELINE_ID);
12383 6 : for segment in segments.iter_mut() {
12384 6 : segment.segment.parent = None; // We don't care about the parent for the test
12385 6 : segment.segment.size = None; // We don't care about the size for the test
12386 6 : }
12387 1 : assert_eq!(
12388 1 : segments,
12389 1 : [
12390 1 : SegmentMeta {
12391 1 : segment: Segment {
12392 1 : parent: None,
12393 1 : lsn: 0x10,
12394 1 : size: None,
12395 1 : needed: false,
12396 1 : },
12397 1 : timeline_id: TIMELINE_ID,
12398 1 : kind: BranchStart,
12399 1 : },
12400 1 : SegmentMeta {
12401 1 : segment: Segment {
12402 1 : parent: None,
12403 1 : lsn: 0x20,
12404 1 : size: None,
12405 1 : needed: false,
12406 1 : },
12407 1 : timeline_id: TIMELINE_ID,
12408 1 : kind: BranchPoint,
12409 1 : },
12410 1 : SegmentMeta {
12411 1 : segment: Segment {
12412 1 : parent: None,
12413 1 : lsn: 0x30,
12414 1 : size: None,
12415 1 : needed: false,
12416 1 : },
12417 1 : timeline_id: TIMELINE_ID,
12418 1 : kind: BranchPoint,
12419 1 : },
12420 1 : SegmentMeta {
12421 1 : segment: Segment {
12422 1 : parent: None,
12423 1 : lsn: 0x40,
12424 1 : size: None,
12425 1 : needed: false,
12426 1 : },
12427 1 : timeline_id: TIMELINE_ID,
12428 1 : kind: BranchPoint,
12429 1 : },
12430 1 : SegmentMeta {
12431 1 : segment: Segment {
12432 1 : parent: None,
12433 1 : lsn: 0x100,
12434 1 : size: None,
12435 1 : needed: false,
12436 1 : },
12437 1 : timeline_id: TIMELINE_ID,
12438 1 : kind: GcCutOff,
12439 1 : }, // we need to retain everything above the last branch point
12440 1 : SegmentMeta {
12441 1 : segment: Segment {
12442 1 : parent: None,
12443 1 : lsn: 0x100,
12444 1 : size: None,
12445 1 : needed: true,
12446 1 : },
12447 1 : timeline_id: TIMELINE_ID,
12448 1 : kind: BranchEnd,
12449 1 : },
12450 1 : ]
12451 1 : );
12452 1 :
12453 1 : main_tline
12454 1 : .remote_client
12455 1 : .schedule_index_upload_for_timeline_invisible_state(
12456 1 : TimelineVisibilityState::Invisible,
12457 1 : )?;
12458 1 : main_tline.remote_client.wait_completion().await?;
12459 1 : let inputs = gather_inputs(
12460 1 : &tenant,
12461 1 : &limit,
12462 1 : max_retention_period,
12463 1 : &mut logical_size_cache,
12464 1 : cause,
12465 1 : &cancel,
12466 1 : &ctx,
12467 1 : )
12468 1 : .instrument(info_span!(
12469 1 : "gather_inputs",
12470 1 : tenant_id = "unknown",
12471 1 : shard_id = "unknown",
12472 1 : ))
12473 1 : .await?;
12474 1 : let ModelInputs { mut segments, .. } = inputs;
12475 14 : segments.retain(|s| s.timeline_id == TIMELINE_ID);
12476 5 : for segment in segments.iter_mut() {
12477 5 : segment.segment.parent = None; // We don't care about the parent for the test
12478 5 : segment.segment.size = None; // We don't care about the size for the test
12479 5 : }
12480 1 : assert_eq!(
12481 1 : segments,
12482 1 : [
12483 1 : SegmentMeta {
12484 1 : segment: Segment {
12485 1 : parent: None,
12486 1 : lsn: 0x10,
12487 1 : size: None,
12488 1 : needed: false,
12489 1 : },
12490 1 : timeline_id: TIMELINE_ID,
12491 1 : kind: BranchStart,
12492 1 : },
12493 1 : SegmentMeta {
12494 1 : segment: Segment {
12495 1 : parent: None,
12496 1 : lsn: 0x20,
12497 1 : size: None,
12498 1 : needed: false,
12499 1 : },
12500 1 : timeline_id: TIMELINE_ID,
12501 1 : kind: BranchPoint,
12502 1 : },
12503 1 : SegmentMeta {
12504 1 : segment: Segment {
12505 1 : parent: None,
12506 1 : lsn: 0x30,
12507 1 : size: None,
12508 1 : needed: false,
12509 1 : },
12510 1 : timeline_id: TIMELINE_ID,
12511 1 : kind: BranchPoint,
12512 1 : },
12513 1 : SegmentMeta {
12514 1 : segment: Segment {
12515 1 : parent: None,
12516 1 : lsn: 0x40,
12517 1 : size: None,
12518 1 : needed: false,
12519 1 : },
12520 1 : timeline_id: TIMELINE_ID,
12521 1 : kind: BranchPoint,
12522 1 : },
12523 1 : SegmentMeta {
12524 1 : segment: Segment {
12525 1 : parent: None,
12526 1 : lsn: 0x40, // Branch end LSN == last branch point LSN
12527 1 : size: None,
12528 1 : needed: true,
12529 1 : },
12530 1 : timeline_id: TIMELINE_ID,
12531 1 : kind: BranchEnd,
12532 1 : },
12533 1 : ]
12534 1 : );
12535 1 : Ok(())
12536 1 : }
12537 : }
|
