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::offload::{OffloadError, offload_timeline};
54 : use timeline::{
55 : CompactFlags, CompactOptions, CompactionError, PreviousHeatmap, ShutdownMode, import_pgdata,
56 : };
57 : use tokio::io::BufReader;
58 : use tokio::sync::{Notify, Semaphore, watch};
59 : use tokio::task::JoinSet;
60 : use tokio_util::sync::CancellationToken;
61 : use tracing::*;
62 : use upload_queue::NotInitialized;
63 : use utils::circuit_breaker::CircuitBreaker;
64 : use utils::crashsafe::path_with_suffix_extension;
65 : use utils::sync::gate::{Gate, GateGuard};
66 : use utils::timeout::{TimeoutCancellableError, timeout_cancellable};
67 : use utils::try_rcu::ArcSwapExt;
68 : use utils::zstd::{create_zst_tarball, extract_zst_tarball};
69 : use utils::{backoff, completion, failpoint_support, fs_ext, pausable_failpoint};
70 :
71 : use self::config::{AttachedLocationConfig, AttachmentMode, LocationConf};
72 : use self::metadata::TimelineMetadata;
73 : use self::mgr::{GetActiveTenantError, GetTenantError};
74 : use self::remote_timeline_client::upload::{upload_index_part, upload_tenant_manifest};
75 : use self::remote_timeline_client::{RemoteTimelineClient, WaitCompletionError};
76 : use self::timeline::uninit::{TimelineCreateGuard, TimelineExclusionError, UninitializedTimeline};
77 : use self::timeline::{
78 : EvictionTaskTenantState, GcCutoffs, TimelineDeleteProgress, TimelineResources, WaitLsnError,
79 : };
80 : use crate::config::PageServerConf;
81 : use crate::context;
82 : use crate::context::RequestContextBuilder;
83 : use crate::context::{DownloadBehavior, RequestContext};
84 : use crate::deletion_queue::{DeletionQueueClient, DeletionQueueError};
85 : use crate::l0_flush::L0FlushGlobalState;
86 : use crate::metrics::{
87 : BROKEN_TENANTS_SET, CIRCUIT_BREAKERS_BROKEN, CIRCUIT_BREAKERS_UNBROKEN, CONCURRENT_INITDBS,
88 : INITDB_RUN_TIME, INITDB_SEMAPHORE_ACQUISITION_TIME, TENANT, TENANT_STATE_METRIC,
89 : TENANT_SYNTHETIC_SIZE_METRIC, remove_tenant_metrics,
90 : };
91 : use crate::task_mgr::TaskKind;
92 : use crate::tenant::config::LocationMode;
93 : use crate::tenant::gc_result::GcResult;
94 : pub use crate::tenant::remote_timeline_client::index::IndexPart;
95 : use crate::tenant::remote_timeline_client::{
96 : INITDB_PATH, MaybeDeletedIndexPart, remote_initdb_archive_path,
97 : };
98 : use crate::tenant::storage_layer::{DeltaLayer, ImageLayer};
99 : use crate::tenant::timeline::delete::DeleteTimelineFlow;
100 : use crate::tenant::timeline::uninit::cleanup_timeline_directory;
101 : use crate::virtual_file::VirtualFile;
102 : use crate::walingest::WalLagCooldown;
103 : use crate::walredo::{PostgresRedoManager, RedoAttemptType};
104 : use crate::{InitializationOrder, TEMP_FILE_SUFFIX, import_datadir, span, task_mgr, walredo};
105 :
106 0 : static INIT_DB_SEMAPHORE: Lazy<Semaphore> = Lazy::new(|| Semaphore::new(8));
107 : use utils::crashsafe;
108 : use utils::generation::Generation;
109 : use utils::id::TimelineId;
110 : use utils::lsn::{Lsn, RecordLsn};
111 :
112 : pub mod blob_io;
113 : pub mod block_io;
114 : pub mod vectored_blob_io;
115 :
116 : pub mod disk_btree;
117 : pub(crate) mod ephemeral_file;
118 : pub mod layer_map;
119 :
120 : pub mod metadata;
121 : pub mod remote_timeline_client;
122 : pub mod storage_layer;
123 :
124 : pub mod checks;
125 : pub mod config;
126 : pub mod mgr;
127 : pub mod secondary;
128 : pub mod tasks;
129 : pub mod upload_queue;
130 :
131 : pub(crate) mod timeline;
132 :
133 : pub mod size;
134 :
135 : mod gc_block;
136 : mod gc_result;
137 : pub(crate) mod throttle;
138 :
139 : pub(crate) use timeline::{LogicalSizeCalculationCause, PageReconstructError, Timeline};
140 :
141 : pub(crate) use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
142 : // re-export for use in walreceiver
143 : pub use crate::tenant::timeline::WalReceiverInfo;
144 :
145 : /// The "tenants" part of `tenants/<tenant>/timelines...`
146 : pub const TENANTS_SEGMENT_NAME: &str = "tenants";
147 :
148 : /// Parts of the `.neon/tenants/<tenant_id>/timelines/<timeline_id>` directory prefix.
149 : pub const TIMELINES_SEGMENT_NAME: &str = "timelines";
150 :
151 : /// References to shared objects that are passed into each tenant, such
152 : /// as the shared remote storage client and process initialization state.
153 : #[derive(Clone)]
154 : pub struct TenantSharedResources {
155 : pub broker_client: storage_broker::BrokerClientChannel,
156 : pub remote_storage: GenericRemoteStorage,
157 : pub deletion_queue_client: DeletionQueueClient,
158 : pub l0_flush_global_state: L0FlushGlobalState,
159 : }
160 :
161 : /// A [`TenantShard`] is really an _attached_ tenant. The configuration
162 : /// for an attached tenant is a subset of the [`LocationConf`], represented
163 : /// in this struct.
164 : #[derive(Clone)]
165 : pub(super) struct AttachedTenantConf {
166 : tenant_conf: pageserver_api::models::TenantConfig,
167 : location: AttachedLocationConfig,
168 : /// The deadline before which we are blocked from GC so that
169 : /// leases have a chance to be renewed.
170 : lsn_lease_deadline: Option<tokio::time::Instant>,
171 : }
172 :
173 : impl AttachedTenantConf {
174 1392 : fn new(
175 1392 : tenant_conf: pageserver_api::models::TenantConfig,
176 1392 : location: AttachedLocationConfig,
177 1392 : ) -> Self {
178 : // Sets a deadline before which we cannot proceed to GC due to lsn lease.
179 : //
180 : // We do this as the leases mapping are not persisted to disk. By delaying GC by lease
181 : // length, we guarantee that all the leases we granted before will have a chance to renew
182 : // when we run GC for the first time after restart / transition from AttachedMulti to AttachedSingle.
183 1392 : let lsn_lease_deadline = if location.attach_mode == AttachmentMode::Single {
184 1392 : Some(
185 1392 : tokio::time::Instant::now()
186 1392 : + tenant_conf
187 1392 : .lsn_lease_length
188 1392 : .unwrap_or(LsnLease::DEFAULT_LENGTH),
189 1392 : )
190 : } else {
191 : // We don't use `lsn_lease_deadline` to delay GC in AttachedMulti and AttachedStale
192 : // because we don't do GC in these modes.
193 0 : None
194 : };
195 :
196 1392 : Self {
197 1392 : tenant_conf,
198 1392 : location,
199 1392 : lsn_lease_deadline,
200 1392 : }
201 1392 : }
202 :
203 1392 : fn try_from(location_conf: LocationConf) -> anyhow::Result<Self> {
204 1392 : match &location_conf.mode {
205 1392 : LocationMode::Attached(attach_conf) => {
206 1392 : Ok(Self::new(location_conf.tenant_conf, *attach_conf))
207 : }
208 : LocationMode::Secondary(_) => {
209 0 : anyhow::bail!(
210 0 : "Attempted to construct AttachedTenantConf from a LocationConf in secondary mode"
211 0 : )
212 : }
213 : }
214 1392 : }
215 :
216 4572 : fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
217 4572 : self.lsn_lease_deadline
218 4572 : .map(|d| tokio::time::Instant::now() < d)
219 4572 : .unwrap_or(false)
220 4572 : }
221 : }
222 : struct TimelinePreload {
223 : timeline_id: TimelineId,
224 : client: RemoteTimelineClient,
225 : index_part: Result<MaybeDeletedIndexPart, DownloadError>,
226 : previous_heatmap: Option<PreviousHeatmap>,
227 : }
228 :
229 : pub(crate) struct TenantPreload {
230 : /// The tenant manifest from remote storage, or None if no manifest was found.
231 : tenant_manifest: Option<TenantManifest>,
232 : /// Map from timeline ID to a possible timeline preload. It is None iff the timeline is offloaded according to the manifest.
233 : timelines: HashMap<TimelineId, Option<TimelinePreload>>,
234 : }
235 :
236 : /// When we spawn a tenant, there is a special mode for tenant creation that
237 : /// avoids trying to read anything from remote storage.
238 : pub(crate) enum SpawnMode {
239 : /// Activate as soon as possible
240 : Eager,
241 : /// Lazy activation in the background, with the option to skip the queue if the need comes up
242 : Lazy,
243 : }
244 :
245 : ///
246 : /// Tenant consists of multiple timelines. Keep them in a hash table.
247 : ///
248 : pub struct TenantShard {
249 : // Global pageserver config parameters
250 : pub conf: &'static PageServerConf,
251 :
252 : /// The value creation timestamp, used to measure activation delay, see:
253 : /// <https://github.com/neondatabase/neon/issues/4025>
254 : constructed_at: Instant,
255 :
256 : state: watch::Sender<TenantState>,
257 :
258 : // Overridden tenant-specific config parameters.
259 : // We keep pageserver_api::models::TenantConfig sturct here to preserve the information
260 : // about parameters that are not set.
261 : // This is necessary to allow global config updates.
262 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
263 :
264 : tenant_shard_id: TenantShardId,
265 :
266 : // The detailed sharding information, beyond the number/count in tenant_shard_id
267 : shard_identity: ShardIdentity,
268 :
269 : /// The remote storage generation, used to protect S3 objects from split-brain.
270 : /// Does not change over the lifetime of the [`TenantShard`] object.
271 : ///
272 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
273 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
274 : generation: Generation,
275 :
276 : timelines: Mutex<HashMap<TimelineId, Arc<Timeline>>>,
277 :
278 : /// During timeline creation, we first insert the TimelineId to the
279 : /// creating map, then `timelines`, then remove it from the creating map.
280 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
281 : timelines_creating: std::sync::Mutex<HashSet<TimelineId>>,
282 :
283 : /// Possibly offloaded and archived timelines
284 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
285 : timelines_offloaded: Mutex<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
286 :
287 : /// The last tenant manifest known to be in remote storage. None if the manifest has not yet
288 : /// been either downloaded or uploaded. Always Some after tenant attach.
289 : ///
290 : /// Initially populated during tenant attach, updated via `maybe_upload_tenant_manifest`.
291 : ///
292 : /// Do not modify this directly. It is used to check whether a new manifest needs to be
293 : /// uploaded. The manifest is constructed in `build_tenant_manifest`, and uploaded via
294 : /// `maybe_upload_tenant_manifest`.
295 : remote_tenant_manifest: tokio::sync::Mutex<Option<TenantManifest>>,
296 :
297 : // This mutex prevents creation of new timelines during GC.
298 : // Adding yet another mutex (in addition to `timelines`) is needed because holding
299 : // `timelines` mutex during all GC iteration
300 : // may block for a long time `get_timeline`, `get_timelines_state`,... and other operations
301 : // with timelines, which in turn may cause dropping replication connection, expiration of wait_for_lsn
302 : // timeout...
303 : gc_cs: tokio::sync::Mutex<()>,
304 : walredo_mgr: Option<Arc<WalRedoManager>>,
305 :
306 : // provides access to timeline data sitting in the remote storage
307 : pub(crate) remote_storage: GenericRemoteStorage,
308 :
309 : // Access to global deletion queue for when this tenant wants to schedule a deletion
310 : deletion_queue_client: DeletionQueueClient,
311 :
312 : /// Cached logical sizes updated updated on each [`TenantShard::gather_size_inputs`].
313 : cached_logical_sizes: tokio::sync::Mutex<HashMap<(TimelineId, Lsn), u64>>,
314 : cached_synthetic_tenant_size: Arc<AtomicU64>,
315 :
316 : eviction_task_tenant_state: tokio::sync::Mutex<EvictionTaskTenantState>,
317 :
318 : /// Track repeated failures to compact, so that we can back off.
319 : /// Overhead of mutex is acceptable because compaction is done with a multi-second period.
320 : compaction_circuit_breaker: std::sync::Mutex<CircuitBreaker>,
321 :
322 : /// Signals the tenant compaction loop that there is L0 compaction work to be done.
323 : pub(crate) l0_compaction_trigger: Arc<Notify>,
324 :
325 : /// Scheduled gc-compaction tasks.
326 : scheduled_compaction_tasks: std::sync::Mutex<HashMap<TimelineId, Arc<GcCompactionQueue>>>,
327 :
328 : /// If the tenant is in Activating state, notify this to encourage it
329 : /// to proceed to Active as soon as possible, rather than waiting for lazy
330 : /// background warmup.
331 : pub(crate) activate_now_sem: tokio::sync::Semaphore,
332 :
333 : /// Time it took for the tenant to activate. Zero if not active yet.
334 : attach_wal_lag_cooldown: Arc<std::sync::OnceLock<WalLagCooldown>>,
335 :
336 : // Cancellation token fires when we have entered shutdown(). This is a parent of
337 : // Timelines' cancellation token.
338 : pub(crate) cancel: CancellationToken,
339 :
340 : // Users of the TenantShard such as the page service must take this Gate to avoid
341 : // trying to use a TenantShard which is shutting down.
342 : pub(crate) gate: Gate,
343 :
344 : /// Throttle applied at the top of [`Timeline::get`].
345 : /// All [`TenantShard::timelines`] of a given [`TenantShard`] instance share the same [`throttle::Throttle`] instance.
346 : pub(crate) pagestream_throttle: Arc<throttle::Throttle>,
347 :
348 : pub(crate) pagestream_throttle_metrics: Arc<crate::metrics::tenant_throttling::Pagestream>,
349 :
350 : /// An ongoing timeline detach concurrency limiter.
351 : ///
352 : /// As a tenant will likely be restarted as part of timeline detach ancestor it makes no sense
353 : /// to have two running at the same time. A different one can be started if an earlier one
354 : /// has failed for whatever reason.
355 : ongoing_timeline_detach: std::sync::Mutex<Option<(TimelineId, utils::completion::Barrier)>>,
356 :
357 : /// `index_part.json` based gc blocking reason tracking.
358 : ///
359 : /// New gc iterations must start a new iteration by acquiring `GcBlock::start` before
360 : /// proceeding.
361 : pub(crate) gc_block: gc_block::GcBlock,
362 :
363 : l0_flush_global_state: L0FlushGlobalState,
364 : }
365 : impl std::fmt::Debug for TenantShard {
366 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
367 0 : write!(f, "{} ({})", self.tenant_shard_id, self.current_state())
368 0 : }
369 : }
370 :
371 : pub(crate) enum WalRedoManager {
372 : Prod(WalredoManagerId, PostgresRedoManager),
373 : #[cfg(test)]
374 : Test(harness::TestRedoManager),
375 : }
376 :
377 : #[derive(thiserror::Error, Debug)]
378 : #[error("pageserver is shutting down")]
379 : pub(crate) struct GlobalShutDown;
380 :
381 : impl WalRedoManager {
382 0 : pub(crate) fn new(mgr: PostgresRedoManager) -> Result<Arc<Self>, GlobalShutDown> {
383 0 : let id = WalredoManagerId::next();
384 0 : let arc = Arc::new(Self::Prod(id, mgr));
385 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
386 0 : match &mut *guard {
387 0 : Some(map) => {
388 0 : map.insert(id, Arc::downgrade(&arc));
389 0 : Ok(arc)
390 : }
391 0 : None => Err(GlobalShutDown),
392 : }
393 0 : }
394 : }
395 :
396 : impl Drop for WalRedoManager {
397 60 : fn drop(&mut self) {
398 60 : match self {
399 0 : Self::Prod(id, _) => {
400 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
401 0 : if let Some(map) = &mut *guard {
402 0 : map.remove(id).expect("new() registers, drop() unregisters");
403 0 : }
404 : }
405 : #[cfg(test)]
406 60 : Self::Test(_) => {
407 60 : // Not applicable to test redo manager
408 60 : }
409 : }
410 60 : }
411 : }
412 :
413 : /// Global registry of all walredo managers so that [`crate::shutdown_pageserver`] can shut down
414 : /// the walredo processes outside of the regular order.
415 : ///
416 : /// This is necessary to work around a systemd bug where it freezes if there are
417 : /// walredo processes left => <https://github.com/neondatabase/cloud/issues/11387>
418 : #[allow(clippy::type_complexity)]
419 : pub(crate) static WALREDO_MANAGERS: once_cell::sync::Lazy<
420 : Mutex<Option<HashMap<WalredoManagerId, Weak<WalRedoManager>>>>,
421 0 : > = once_cell::sync::Lazy::new(|| Mutex::new(Some(HashMap::new())));
422 : #[derive(PartialEq, Eq, Hash, Clone, Copy, Debug)]
423 : pub(crate) struct WalredoManagerId(u64);
424 : impl WalredoManagerId {
425 0 : pub fn next() -> Self {
426 : static NEXT: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(1);
427 0 : let id = NEXT.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
428 0 : if id == 0 {
429 0 : panic!(
430 0 : "WalredoManagerId::new() returned 0, indicating wraparound, risking it's no longer unique"
431 0 : );
432 0 : }
433 0 : Self(id)
434 0 : }
435 : }
436 :
437 : #[cfg(test)]
438 : impl From<harness::TestRedoManager> for WalRedoManager {
439 1392 : fn from(mgr: harness::TestRedoManager) -> Self {
440 1392 : Self::Test(mgr)
441 1392 : }
442 : }
443 :
444 : impl WalRedoManager {
445 36 : pub(crate) async fn shutdown(&self) -> bool {
446 36 : match self {
447 0 : Self::Prod(_, mgr) => mgr.shutdown().await,
448 : #[cfg(test)]
449 : Self::Test(_) => {
450 : // Not applicable to test redo manager
451 36 : true
452 : }
453 : }
454 36 : }
455 :
456 0 : pub(crate) fn maybe_quiesce(&self, idle_timeout: Duration) {
457 0 : match self {
458 0 : Self::Prod(_, mgr) => mgr.maybe_quiesce(idle_timeout),
459 0 : #[cfg(test)]
460 0 : Self::Test(_) => {
461 0 : // Not applicable to test redo manager
462 0 : }
463 0 : }
464 0 : }
465 :
466 : /// # Cancel-Safety
467 : ///
468 : /// This method is cancellation-safe.
469 321288 : pub async fn request_redo(
470 321288 : &self,
471 321288 : key: pageserver_api::key::Key,
472 321288 : lsn: Lsn,
473 321288 : base_img: Option<(Lsn, bytes::Bytes)>,
474 321288 : records: Vec<(Lsn, pageserver_api::record::NeonWalRecord)>,
475 321288 : pg_version: u32,
476 321288 : redo_attempt_type: RedoAttemptType,
477 321288 : ) -> Result<bytes::Bytes, walredo::Error> {
478 321288 : match self {
479 0 : Self::Prod(_, mgr) => {
480 0 : mgr.request_redo(key, lsn, base_img, records, pg_version, redo_attempt_type)
481 0 : .await
482 : }
483 : #[cfg(test)]
484 321288 : Self::Test(mgr) => {
485 321288 : mgr.request_redo(key, lsn, base_img, records, pg_version, redo_attempt_type)
486 321288 : .await
487 : }
488 : }
489 321288 : }
490 :
491 0 : pub(crate) fn status(&self) -> Option<WalRedoManagerStatus> {
492 0 : match self {
493 0 : WalRedoManager::Prod(_, m) => Some(m.status()),
494 0 : #[cfg(test)]
495 0 : WalRedoManager::Test(_) => None,
496 0 : }
497 0 : }
498 : }
499 :
500 : /// A very lightweight memory representation of an offloaded timeline.
501 : ///
502 : /// We need to store the list of offloaded timelines so that we can perform operations on them,
503 : /// like unoffloading them, or (at a later date), decide to perform flattening.
504 : /// This type has a much smaller memory impact than [`Timeline`], and thus we can store many
505 : /// more offloaded timelines than we can manage ones that aren't.
506 : pub struct OffloadedTimeline {
507 : pub tenant_shard_id: TenantShardId,
508 : pub timeline_id: TimelineId,
509 : pub ancestor_timeline_id: Option<TimelineId>,
510 : /// Whether to retain the branch lsn at the ancestor or not
511 : pub ancestor_retain_lsn: Option<Lsn>,
512 :
513 : /// When the timeline was archived.
514 : ///
515 : /// Present for future flattening deliberations.
516 : pub archived_at: NaiveDateTime,
517 :
518 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
519 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
520 : pub delete_progress: TimelineDeleteProgress,
521 :
522 : /// Part of the `OffloadedTimeline` object's lifecycle: this needs to be set before we drop it
523 : pub deleted_from_ancestor: AtomicBool,
524 : }
525 :
526 : impl OffloadedTimeline {
527 : /// Obtains an offloaded timeline from a given timeline object.
528 : ///
529 : /// Returns `None` if the `archived_at` flag couldn't be obtained, i.e.
530 : /// the timeline is not in a stopped state.
531 : /// Panics if the timeline is not archived.
532 12 : fn from_timeline(timeline: &Timeline) -> Result<Self, UploadQueueNotReadyError> {
533 12 : let (ancestor_retain_lsn, ancestor_timeline_id) =
534 12 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
535 12 : let ancestor_lsn = timeline.get_ancestor_lsn();
536 12 : let ancestor_timeline_id = ancestor_timeline.timeline_id;
537 12 : let mut gc_info = ancestor_timeline.gc_info.write().unwrap();
538 12 : gc_info.insert_child(timeline.timeline_id, ancestor_lsn, MaybeOffloaded::Yes);
539 12 : (Some(ancestor_lsn), Some(ancestor_timeline_id))
540 : } else {
541 0 : (None, None)
542 : };
543 12 : let archived_at = timeline
544 12 : .remote_client
545 12 : .archived_at_stopped_queue()?
546 12 : .expect("must be called on an archived timeline");
547 12 : Ok(Self {
548 12 : tenant_shard_id: timeline.tenant_shard_id,
549 12 : timeline_id: timeline.timeline_id,
550 12 : ancestor_timeline_id,
551 12 : ancestor_retain_lsn,
552 12 : archived_at,
553 12 :
554 12 : delete_progress: timeline.delete_progress.clone(),
555 12 : deleted_from_ancestor: AtomicBool::new(false),
556 12 : })
557 12 : }
558 0 : fn from_manifest(tenant_shard_id: TenantShardId, manifest: &OffloadedTimelineManifest) -> Self {
559 0 : // We expect to reach this case in tenant loading, where the `retain_lsn` is populated in the parent's `gc_info`
560 0 : // by the `initialize_gc_info` function.
561 0 : let OffloadedTimelineManifest {
562 0 : timeline_id,
563 0 : ancestor_timeline_id,
564 0 : ancestor_retain_lsn,
565 0 : archived_at,
566 0 : } = *manifest;
567 0 : Self {
568 0 : tenant_shard_id,
569 0 : timeline_id,
570 0 : ancestor_timeline_id,
571 0 : ancestor_retain_lsn,
572 0 : archived_at,
573 0 : delete_progress: TimelineDeleteProgress::default(),
574 0 : deleted_from_ancestor: AtomicBool::new(false),
575 0 : }
576 0 : }
577 12 : fn manifest(&self) -> OffloadedTimelineManifest {
578 12 : let Self {
579 12 : timeline_id,
580 12 : ancestor_timeline_id,
581 12 : ancestor_retain_lsn,
582 12 : archived_at,
583 12 : ..
584 12 : } = self;
585 12 : OffloadedTimelineManifest {
586 12 : timeline_id: *timeline_id,
587 12 : ancestor_timeline_id: *ancestor_timeline_id,
588 12 : ancestor_retain_lsn: *ancestor_retain_lsn,
589 12 : archived_at: *archived_at,
590 12 : }
591 12 : }
592 : /// Delete this timeline's retain_lsn from its ancestor, if present in the given tenant
593 0 : fn delete_from_ancestor_with_timelines(
594 0 : &self,
595 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
596 0 : ) {
597 0 : if let (Some(_retain_lsn), Some(ancestor_timeline_id)) =
598 0 : (self.ancestor_retain_lsn, self.ancestor_timeline_id)
599 : {
600 0 : if let Some((_, ancestor_timeline)) = timelines
601 0 : .iter()
602 0 : .find(|(tid, _tl)| **tid == ancestor_timeline_id)
603 : {
604 0 : let removal_happened = ancestor_timeline
605 0 : .gc_info
606 0 : .write()
607 0 : .unwrap()
608 0 : .remove_child_offloaded(self.timeline_id);
609 0 : if !removal_happened {
610 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
611 0 : "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
612 0 : }
613 0 : }
614 0 : }
615 0 : self.deleted_from_ancestor.store(true, Ordering::Release);
616 0 : }
617 : /// Call [`Self::delete_from_ancestor_with_timelines`] instead if possible.
618 : ///
619 : /// As the entire tenant is being dropped, don't bother deregistering the `retain_lsn` from the ancestor.
620 12 : fn defuse_for_tenant_drop(&self) {
621 12 : self.deleted_from_ancestor.store(true, Ordering::Release);
622 12 : }
623 : }
624 :
625 : impl fmt::Debug for OffloadedTimeline {
626 0 : fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
627 0 : write!(f, "OffloadedTimeline<{}>", self.timeline_id)
628 0 : }
629 : }
630 :
631 : impl Drop for OffloadedTimeline {
632 12 : fn drop(&mut self) {
633 12 : if !self.deleted_from_ancestor.load(Ordering::Acquire) {
634 0 : tracing::warn!(
635 0 : "offloaded timeline {} was dropped without having cleaned it up at the ancestor",
636 : self.timeline_id
637 : );
638 12 : }
639 12 : }
640 : }
641 :
642 : #[derive(Copy, Clone, PartialEq, Eq, Hash, Debug)]
643 : pub enum MaybeOffloaded {
644 : Yes,
645 : No,
646 : }
647 :
648 : #[derive(Clone, Debug)]
649 : pub enum TimelineOrOffloaded {
650 : Timeline(Arc<Timeline>),
651 : Offloaded(Arc<OffloadedTimeline>),
652 : }
653 :
654 : impl TimelineOrOffloaded {
655 0 : pub fn arc_ref(&self) -> TimelineOrOffloadedArcRef<'_> {
656 0 : match self {
657 0 : TimelineOrOffloaded::Timeline(timeline) => {
658 0 : TimelineOrOffloadedArcRef::Timeline(timeline)
659 : }
660 0 : TimelineOrOffloaded::Offloaded(offloaded) => {
661 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded)
662 : }
663 : }
664 0 : }
665 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
666 0 : self.arc_ref().tenant_shard_id()
667 0 : }
668 0 : pub fn timeline_id(&self) -> TimelineId {
669 0 : self.arc_ref().timeline_id()
670 0 : }
671 12 : pub fn delete_progress(&self) -> &Arc<tokio::sync::Mutex<DeleteTimelineFlow>> {
672 12 : match self {
673 12 : TimelineOrOffloaded::Timeline(timeline) => &timeline.delete_progress,
674 0 : TimelineOrOffloaded::Offloaded(offloaded) => &offloaded.delete_progress,
675 : }
676 12 : }
677 0 : fn maybe_remote_client(&self) -> Option<Arc<RemoteTimelineClient>> {
678 0 : match self {
679 0 : TimelineOrOffloaded::Timeline(timeline) => Some(timeline.remote_client.clone()),
680 0 : TimelineOrOffloaded::Offloaded(_offloaded) => None,
681 : }
682 0 : }
683 : }
684 :
685 : pub enum TimelineOrOffloadedArcRef<'a> {
686 : Timeline(&'a Arc<Timeline>),
687 : Offloaded(&'a Arc<OffloadedTimeline>),
688 : }
689 :
690 : impl TimelineOrOffloadedArcRef<'_> {
691 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
692 0 : match self {
693 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.tenant_shard_id,
694 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.tenant_shard_id,
695 : }
696 0 : }
697 0 : pub fn timeline_id(&self) -> TimelineId {
698 0 : match self {
699 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.timeline_id,
700 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.timeline_id,
701 : }
702 0 : }
703 : }
704 :
705 : impl<'a> From<&'a Arc<Timeline>> for TimelineOrOffloadedArcRef<'a> {
706 0 : fn from(timeline: &'a Arc<Timeline>) -> Self {
707 0 : Self::Timeline(timeline)
708 0 : }
709 : }
710 :
711 : impl<'a> From<&'a Arc<OffloadedTimeline>> for TimelineOrOffloadedArcRef<'a> {
712 0 : fn from(timeline: &'a Arc<OffloadedTimeline>) -> Self {
713 0 : Self::Offloaded(timeline)
714 0 : }
715 : }
716 :
717 : #[derive(Debug, thiserror::Error, PartialEq, Eq)]
718 : pub enum GetTimelineError {
719 : #[error("Timeline is shutting down")]
720 : ShuttingDown,
721 : #[error("Timeline {tenant_id}/{timeline_id} is not active, state: {state:?}")]
722 : NotActive {
723 : tenant_id: TenantShardId,
724 : timeline_id: TimelineId,
725 : state: TimelineState,
726 : },
727 : #[error("Timeline {tenant_id}/{timeline_id} was not found")]
728 : NotFound {
729 : tenant_id: TenantShardId,
730 : timeline_id: TimelineId,
731 : },
732 : }
733 :
734 : #[derive(Debug, thiserror::Error)]
735 : pub enum LoadLocalTimelineError {
736 : #[error("FailedToLoad")]
737 : Load(#[source] anyhow::Error),
738 : #[error("FailedToResumeDeletion")]
739 : ResumeDeletion(#[source] anyhow::Error),
740 : }
741 :
742 : #[derive(thiserror::Error)]
743 : pub enum DeleteTimelineError {
744 : #[error("NotFound")]
745 : NotFound,
746 :
747 : #[error("HasChildren")]
748 : HasChildren(Vec<TimelineId>),
749 :
750 : #[error("Timeline deletion is already in progress")]
751 : AlreadyInProgress(Arc<tokio::sync::Mutex<DeleteTimelineFlow>>),
752 :
753 : #[error("Cancelled")]
754 : Cancelled,
755 :
756 : #[error(transparent)]
757 : Other(#[from] anyhow::Error),
758 : }
759 :
760 : impl Debug for DeleteTimelineError {
761 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
762 0 : match self {
763 0 : Self::NotFound => write!(f, "NotFound"),
764 0 : Self::HasChildren(c) => f.debug_tuple("HasChildren").field(c).finish(),
765 0 : Self::AlreadyInProgress(_) => f.debug_tuple("AlreadyInProgress").finish(),
766 0 : Self::Cancelled => f.debug_tuple("Cancelled").finish(),
767 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
768 : }
769 0 : }
770 : }
771 :
772 : #[derive(thiserror::Error)]
773 : pub enum TimelineArchivalError {
774 : #[error("NotFound")]
775 : NotFound,
776 :
777 : #[error("Timeout")]
778 : Timeout,
779 :
780 : #[error("Cancelled")]
781 : Cancelled,
782 :
783 : #[error("ancestor is archived: {}", .0)]
784 : HasArchivedParent(TimelineId),
785 :
786 : #[error("HasUnarchivedChildren")]
787 : HasUnarchivedChildren(Vec<TimelineId>),
788 :
789 : #[error("Timeline archival is already in progress")]
790 : AlreadyInProgress,
791 :
792 : #[error(transparent)]
793 : Other(anyhow::Error),
794 : }
795 :
796 : #[derive(thiserror::Error, Debug)]
797 : pub(crate) enum TenantManifestError {
798 : #[error("Remote storage error: {0}")]
799 : RemoteStorage(anyhow::Error),
800 :
801 : #[error("Cancelled")]
802 : Cancelled,
803 : }
804 :
805 : impl From<TenantManifestError> for TimelineArchivalError {
806 0 : fn from(e: TenantManifestError) -> Self {
807 0 : match e {
808 0 : TenantManifestError::RemoteStorage(e) => Self::Other(e),
809 0 : TenantManifestError::Cancelled => Self::Cancelled,
810 : }
811 0 : }
812 : }
813 :
814 : impl Debug for TimelineArchivalError {
815 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
816 0 : match self {
817 0 : Self::NotFound => write!(f, "NotFound"),
818 0 : Self::Timeout => write!(f, "Timeout"),
819 0 : Self::Cancelled => write!(f, "Cancelled"),
820 0 : Self::HasArchivedParent(p) => f.debug_tuple("HasArchivedParent").field(p).finish(),
821 0 : Self::HasUnarchivedChildren(c) => {
822 0 : f.debug_tuple("HasUnarchivedChildren").field(c).finish()
823 : }
824 0 : Self::AlreadyInProgress => f.debug_tuple("AlreadyInProgress").finish(),
825 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
826 : }
827 0 : }
828 : }
829 :
830 : pub enum SetStoppingError {
831 : AlreadyStopping(completion::Barrier),
832 : Broken,
833 : }
834 :
835 : impl Debug for SetStoppingError {
836 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
837 0 : match self {
838 0 : Self::AlreadyStopping(_) => f.debug_tuple("AlreadyStopping").finish(),
839 0 : Self::Broken => write!(f, "Broken"),
840 : }
841 0 : }
842 : }
843 :
844 : /// Arguments to [`TenantShard::create_timeline`].
845 : ///
846 : /// Not usable as an idempotency key for timeline creation because if [`CreateTimelineParamsBranch::ancestor_start_lsn`]
847 : /// is `None`, the result of the timeline create call is not deterministic.
848 : ///
849 : /// See [`CreateTimelineIdempotency`] for an idempotency key.
850 : #[derive(Debug)]
851 : pub(crate) enum CreateTimelineParams {
852 : Bootstrap(CreateTimelineParamsBootstrap),
853 : Branch(CreateTimelineParamsBranch),
854 : ImportPgdata(CreateTimelineParamsImportPgdata),
855 : }
856 :
857 : #[derive(Debug)]
858 : pub(crate) struct CreateTimelineParamsBootstrap {
859 : pub(crate) new_timeline_id: TimelineId,
860 : pub(crate) existing_initdb_timeline_id: Option<TimelineId>,
861 : pub(crate) pg_version: u32,
862 : }
863 :
864 : /// NB: See comment on [`CreateTimelineIdempotency::Branch`] for why there's no `pg_version` here.
865 : #[derive(Debug)]
866 : pub(crate) struct CreateTimelineParamsBranch {
867 : pub(crate) new_timeline_id: TimelineId,
868 : pub(crate) ancestor_timeline_id: TimelineId,
869 : pub(crate) ancestor_start_lsn: Option<Lsn>,
870 : }
871 :
872 : #[derive(Debug)]
873 : pub(crate) struct CreateTimelineParamsImportPgdata {
874 : pub(crate) new_timeline_id: TimelineId,
875 : pub(crate) location: import_pgdata::index_part_format::Location,
876 : pub(crate) idempotency_key: import_pgdata::index_part_format::IdempotencyKey,
877 : }
878 :
879 : /// What is used to determine idempotency of a [`TenantShard::create_timeline`] call in [`TenantShard::start_creating_timeline`] in [`TenantShard::start_creating_timeline`].
880 : ///
881 : /// Each [`Timeline`] object holds [`Self`] as an immutable property in [`Timeline::create_idempotency`].
882 : ///
883 : /// We lower timeline creation requests to [`Self`], and then use [`PartialEq::eq`] to compare [`Timeline::create_idempotency`] with the request.
884 : /// If they are equal, we return a reference to the existing timeline, otherwise it's an idempotency conflict.
885 : ///
886 : /// There is special treatment for [`Self::FailWithConflict`] to always return an idempotency conflict.
887 : /// It would be nice to have more advanced derive macros to make that special treatment declarative.
888 : ///
889 : /// Notes:
890 : /// - Unlike [`CreateTimelineParams`], ancestor LSN is fixed, so, branching will be at a deterministic LSN.
891 : /// - We make some trade-offs though, e.g., [`CreateTimelineParamsBootstrap::existing_initdb_timeline_id`]
892 : /// is not considered for idempotency. We can improve on this over time if we deem it necessary.
893 : ///
894 : #[derive(Debug, Clone, PartialEq, Eq)]
895 : pub(crate) enum CreateTimelineIdempotency {
896 : /// NB: special treatment, see comment in [`Self`].
897 : FailWithConflict,
898 : Bootstrap {
899 : pg_version: u32,
900 : },
901 : /// NB: branches always have the same `pg_version` as their ancestor.
902 : /// While [`pageserver_api::models::TimelineCreateRequestMode::Branch::pg_version`]
903 : /// exists as a field, and is set by cplane, it has always been ignored by pageserver when
904 : /// determining the child branch pg_version.
905 : Branch {
906 : ancestor_timeline_id: TimelineId,
907 : ancestor_start_lsn: Lsn,
908 : },
909 : ImportPgdata(CreatingTimelineIdempotencyImportPgdata),
910 : }
911 :
912 : #[derive(Debug, Clone, PartialEq, Eq)]
913 : pub(crate) struct CreatingTimelineIdempotencyImportPgdata {
914 : idempotency_key: import_pgdata::index_part_format::IdempotencyKey,
915 : }
916 :
917 : /// What is returned by [`TenantShard::start_creating_timeline`].
918 : #[must_use]
919 : enum StartCreatingTimelineResult {
920 : CreateGuard(TimelineCreateGuard),
921 : Idempotent(Arc<Timeline>),
922 : }
923 :
924 : #[allow(clippy::large_enum_variant, reason = "TODO")]
925 : enum TimelineInitAndSyncResult {
926 : ReadyToActivate(Arc<Timeline>),
927 : NeedsSpawnImportPgdata(TimelineInitAndSyncNeedsSpawnImportPgdata),
928 : }
929 :
930 : impl TimelineInitAndSyncResult {
931 0 : fn ready_to_activate(self) -> Option<Arc<Timeline>> {
932 0 : match self {
933 0 : Self::ReadyToActivate(timeline) => Some(timeline),
934 0 : _ => None,
935 : }
936 0 : }
937 : }
938 :
939 : #[must_use]
940 : struct TimelineInitAndSyncNeedsSpawnImportPgdata {
941 : timeline: Arc<Timeline>,
942 : import_pgdata: import_pgdata::index_part_format::Root,
943 : guard: TimelineCreateGuard,
944 : }
945 :
946 : /// What is returned by [`TenantShard::create_timeline`].
947 : enum CreateTimelineResult {
948 : Created(Arc<Timeline>),
949 : Idempotent(Arc<Timeline>),
950 : /// IMPORTANT: This [`Arc<Timeline>`] object is not in [`TenantShard::timelines`] when
951 : /// we return this result, nor will this concrete object ever be added there.
952 : /// Cf method comment on [`TenantShard::create_timeline_import_pgdata`].
953 : ImportSpawned(Arc<Timeline>),
954 : }
955 :
956 : impl CreateTimelineResult {
957 0 : fn discriminant(&self) -> &'static str {
958 0 : match self {
959 0 : Self::Created(_) => "Created",
960 0 : Self::Idempotent(_) => "Idempotent",
961 0 : Self::ImportSpawned(_) => "ImportSpawned",
962 : }
963 0 : }
964 0 : fn timeline(&self) -> &Arc<Timeline> {
965 0 : match self {
966 0 : Self::Created(t) | Self::Idempotent(t) | Self::ImportSpawned(t) => t,
967 0 : }
968 0 : }
969 : /// Unit test timelines aren't activated, test has to do it if it needs to.
970 : #[cfg(test)]
971 1416 : fn into_timeline_for_test(self) -> Arc<Timeline> {
972 1416 : match self {
973 1416 : Self::Created(t) | Self::Idempotent(t) | Self::ImportSpawned(t) => t,
974 1416 : }
975 1416 : }
976 : }
977 :
978 : #[derive(thiserror::Error, Debug)]
979 : pub enum CreateTimelineError {
980 : #[error("creation of timeline with the given ID is in progress")]
981 : AlreadyCreating,
982 : #[error("timeline already exists with different parameters")]
983 : Conflict,
984 : #[error(transparent)]
985 : AncestorLsn(anyhow::Error),
986 : #[error("ancestor timeline is not active")]
987 : AncestorNotActive,
988 : #[error("ancestor timeline is archived")]
989 : AncestorArchived,
990 : #[error("tenant shutting down")]
991 : ShuttingDown,
992 : #[error(transparent)]
993 : Other(#[from] anyhow::Error),
994 : }
995 :
996 : #[derive(thiserror::Error, Debug)]
997 : pub enum InitdbError {
998 : #[error("Operation was cancelled")]
999 : Cancelled,
1000 : #[error(transparent)]
1001 : Other(anyhow::Error),
1002 : #[error(transparent)]
1003 : Inner(postgres_initdb::Error),
1004 : }
1005 :
1006 : enum CreateTimelineCause {
1007 : Load,
1008 : Delete,
1009 : }
1010 :
1011 : #[allow(clippy::large_enum_variant, reason = "TODO")]
1012 : enum LoadTimelineCause {
1013 : Attach,
1014 : Unoffload,
1015 : ImportPgdata {
1016 : create_guard: TimelineCreateGuard,
1017 : activate: ActivateTimelineArgs,
1018 : },
1019 : }
1020 :
1021 : #[derive(thiserror::Error, Debug)]
1022 : pub(crate) enum GcError {
1023 : // The tenant is shutting down
1024 : #[error("tenant shutting down")]
1025 : TenantCancelled,
1026 :
1027 : // The tenant is shutting down
1028 : #[error("timeline shutting down")]
1029 : TimelineCancelled,
1030 :
1031 : // The tenant is in a state inelegible to run GC
1032 : #[error("not active")]
1033 : NotActive,
1034 :
1035 : // A requested GC cutoff LSN was invalid, for example it tried to move backwards
1036 : #[error("not active")]
1037 : BadLsn { why: String },
1038 :
1039 : // A remote storage error while scheduling updates after compaction
1040 : #[error(transparent)]
1041 : Remote(anyhow::Error),
1042 :
1043 : // An error reading while calculating GC cutoffs
1044 : #[error(transparent)]
1045 : GcCutoffs(PageReconstructError),
1046 :
1047 : // If GC was invoked for a particular timeline, this error means it didn't exist
1048 : #[error("timeline not found")]
1049 : TimelineNotFound,
1050 : }
1051 :
1052 : impl From<PageReconstructError> for GcError {
1053 0 : fn from(value: PageReconstructError) -> Self {
1054 0 : match value {
1055 0 : PageReconstructError::Cancelled => Self::TimelineCancelled,
1056 0 : other => Self::GcCutoffs(other),
1057 : }
1058 0 : }
1059 : }
1060 :
1061 : impl From<NotInitialized> for GcError {
1062 0 : fn from(value: NotInitialized) -> Self {
1063 0 : match value {
1064 0 : NotInitialized::Uninitialized => GcError::Remote(value.into()),
1065 0 : NotInitialized::Stopped | NotInitialized::ShuttingDown => GcError::TimelineCancelled,
1066 : }
1067 0 : }
1068 : }
1069 :
1070 : impl From<timeline::layer_manager::Shutdown> for GcError {
1071 0 : fn from(_: timeline::layer_manager::Shutdown) -> Self {
1072 0 : GcError::TimelineCancelled
1073 0 : }
1074 : }
1075 :
1076 : #[derive(thiserror::Error, Debug)]
1077 : pub(crate) enum LoadConfigError {
1078 : #[error("TOML deserialization error: '{0}'")]
1079 : DeserializeToml(#[from] toml_edit::de::Error),
1080 :
1081 : #[error("Config not found at {0}")]
1082 : NotFound(Utf8PathBuf),
1083 : }
1084 :
1085 : impl TenantShard {
1086 : /// Yet another helper for timeline initialization.
1087 : ///
1088 : /// - Initializes the Timeline struct and inserts it into the tenant's hash map
1089 : /// - Scans the local timeline directory for layer files and builds the layer map
1090 : /// - Downloads remote index file and adds remote files to the layer map
1091 : /// - Schedules remote upload tasks for any files that are present locally but missing from remote storage.
1092 : ///
1093 : /// If the operation fails, the timeline is left in the tenant's hash map in Broken state. On success,
1094 : /// it is marked as Active.
1095 : #[allow(clippy::too_many_arguments)]
1096 36 : async fn timeline_init_and_sync(
1097 36 : self: &Arc<Self>,
1098 36 : timeline_id: TimelineId,
1099 36 : resources: TimelineResources,
1100 36 : mut index_part: IndexPart,
1101 36 : metadata: TimelineMetadata,
1102 36 : previous_heatmap: Option<PreviousHeatmap>,
1103 36 : ancestor: Option<Arc<Timeline>>,
1104 36 : cause: LoadTimelineCause,
1105 36 : ctx: &RequestContext,
1106 36 : ) -> anyhow::Result<TimelineInitAndSyncResult> {
1107 36 : let tenant_id = self.tenant_shard_id;
1108 36 :
1109 36 : let import_pgdata = index_part.import_pgdata.take();
1110 36 : let idempotency = match &import_pgdata {
1111 0 : Some(import_pgdata) => {
1112 0 : CreateTimelineIdempotency::ImportPgdata(CreatingTimelineIdempotencyImportPgdata {
1113 0 : idempotency_key: import_pgdata.idempotency_key().clone(),
1114 0 : })
1115 : }
1116 : None => {
1117 36 : if metadata.ancestor_timeline().is_none() {
1118 24 : CreateTimelineIdempotency::Bootstrap {
1119 24 : pg_version: metadata.pg_version(),
1120 24 : }
1121 : } else {
1122 12 : CreateTimelineIdempotency::Branch {
1123 12 : ancestor_timeline_id: metadata.ancestor_timeline().unwrap(),
1124 12 : ancestor_start_lsn: metadata.ancestor_lsn(),
1125 12 : }
1126 : }
1127 : }
1128 : };
1129 :
1130 36 : let (timeline, timeline_ctx) = self.create_timeline_struct(
1131 36 : timeline_id,
1132 36 : &metadata,
1133 36 : previous_heatmap,
1134 36 : ancestor.clone(),
1135 36 : resources,
1136 36 : CreateTimelineCause::Load,
1137 36 : idempotency.clone(),
1138 36 : index_part.gc_compaction.clone(),
1139 36 : index_part.rel_size_migration.clone(),
1140 36 : ctx,
1141 36 : )?;
1142 36 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
1143 36 : anyhow::ensure!(
1144 36 : disk_consistent_lsn.is_valid(),
1145 0 : "Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn"
1146 : );
1147 36 : assert_eq!(
1148 36 : disk_consistent_lsn,
1149 36 : metadata.disk_consistent_lsn(),
1150 0 : "these are used interchangeably"
1151 : );
1152 :
1153 36 : timeline.remote_client.init_upload_queue(&index_part)?;
1154 :
1155 36 : timeline
1156 36 : .load_layer_map(disk_consistent_lsn, index_part)
1157 36 : .await
1158 36 : .with_context(|| {
1159 0 : format!("Failed to load layermap for timeline {tenant_id}/{timeline_id}")
1160 36 : })?;
1161 :
1162 : // When unarchiving, we've mostly likely lost the heatmap generated prior
1163 : // to the archival operation. To allow warming this timeline up, generate
1164 : // a previous heatmap which contains all visible layers in the layer map.
1165 : // This previous heatmap will be used whenever a fresh heatmap is generated
1166 : // for the timeline.
1167 36 : if self.conf.generate_unarchival_heatmap && matches!(cause, LoadTimelineCause::Unoffload) {
1168 0 : let mut tline_ending_at = Some((&timeline, timeline.get_last_record_lsn()));
1169 0 : while let Some((tline, end_lsn)) = tline_ending_at {
1170 0 : let unarchival_heatmap = tline.generate_unarchival_heatmap(end_lsn).await;
1171 : // Another unearchived timeline might have generated a heatmap for this ancestor.
1172 : // If the current branch point greater than the previous one use the the heatmap
1173 : // we just generated - it should include more layers.
1174 0 : if !tline.should_keep_previous_heatmap(end_lsn) {
1175 0 : tline
1176 0 : .previous_heatmap
1177 0 : .store(Some(Arc::new(unarchival_heatmap)));
1178 0 : } else {
1179 0 : tracing::info!("Previous heatmap preferred. Dropping unarchival heatmap.")
1180 : }
1181 :
1182 0 : match tline.ancestor_timeline() {
1183 0 : Some(ancestor) => {
1184 0 : if ancestor.update_layer_visibility().await.is_err() {
1185 : // Ancestor timeline is shutting down.
1186 0 : break;
1187 0 : }
1188 0 :
1189 0 : tline_ending_at = Some((ancestor, tline.get_ancestor_lsn()));
1190 : }
1191 0 : None => {
1192 0 : tline_ending_at = None;
1193 0 : }
1194 : }
1195 : }
1196 36 : }
1197 :
1198 0 : match import_pgdata {
1199 0 : Some(import_pgdata) if !import_pgdata.is_done() => {
1200 0 : match cause {
1201 0 : LoadTimelineCause::Attach | LoadTimelineCause::Unoffload => (),
1202 : LoadTimelineCause::ImportPgdata { .. } => {
1203 0 : unreachable!(
1204 0 : "ImportPgdata should not be reloading timeline import is done and persisted as such in s3"
1205 0 : )
1206 : }
1207 : }
1208 0 : let mut guard = self.timelines_creating.lock().unwrap();
1209 0 : if !guard.insert(timeline_id) {
1210 : // We should never try and load the same timeline twice during startup
1211 0 : unreachable!("Timeline {tenant_id}/{timeline_id} is already being created")
1212 0 : }
1213 0 : let timeline_create_guard = TimelineCreateGuard {
1214 0 : _tenant_gate_guard: self.gate.enter()?,
1215 0 : owning_tenant: self.clone(),
1216 0 : timeline_id,
1217 0 : idempotency,
1218 0 : // The users of this specific return value don't need the timline_path in there.
1219 0 : timeline_path: timeline
1220 0 : .conf
1221 0 : .timeline_path(&timeline.tenant_shard_id, &timeline.timeline_id),
1222 0 : };
1223 0 : Ok(TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
1224 0 : TimelineInitAndSyncNeedsSpawnImportPgdata {
1225 0 : timeline,
1226 0 : import_pgdata,
1227 0 : guard: timeline_create_guard,
1228 0 : },
1229 0 : ))
1230 : }
1231 : Some(_) | None => {
1232 : {
1233 36 : let mut timelines_accessor = self.timelines.lock().unwrap();
1234 36 : match timelines_accessor.entry(timeline_id) {
1235 : // We should never try and load the same timeline twice during startup
1236 : Entry::Occupied(_) => {
1237 0 : unreachable!(
1238 0 : "Timeline {tenant_id}/{timeline_id} already exists in the tenant map"
1239 0 : );
1240 : }
1241 36 : Entry::Vacant(v) => {
1242 36 : v.insert(Arc::clone(&timeline));
1243 36 : timeline.maybe_spawn_flush_loop();
1244 36 : }
1245 : }
1246 : }
1247 :
1248 : // Sanity check: a timeline should have some content.
1249 36 : anyhow::ensure!(
1250 36 : ancestor.is_some()
1251 24 : || timeline
1252 24 : .layers
1253 24 : .read()
1254 24 : .await
1255 24 : .layer_map()
1256 24 : .expect("currently loading, layer manager cannot be shutdown already")
1257 24 : .iter_historic_layers()
1258 24 : .next()
1259 24 : .is_some(),
1260 0 : "Timeline has no ancestor and no layer files"
1261 : );
1262 :
1263 36 : match cause {
1264 36 : LoadTimelineCause::Attach | LoadTimelineCause::Unoffload => (),
1265 : LoadTimelineCause::ImportPgdata {
1266 0 : create_guard,
1267 0 : activate,
1268 0 : } => {
1269 0 : // TODO: see the comment in the task code above how I'm not so certain
1270 0 : // it is safe to activate here because of concurrent shutdowns.
1271 0 : match activate {
1272 0 : ActivateTimelineArgs::Yes { broker_client } => {
1273 0 : info!("activating timeline after reload from pgdata import task");
1274 0 : timeline.activate(self.clone(), broker_client, None, &timeline_ctx);
1275 : }
1276 0 : ActivateTimelineArgs::No => (),
1277 : }
1278 0 : drop(create_guard);
1279 : }
1280 : }
1281 :
1282 36 : Ok(TimelineInitAndSyncResult::ReadyToActivate(timeline))
1283 : }
1284 : }
1285 36 : }
1286 :
1287 : /// Attach a tenant that's available in cloud storage.
1288 : ///
1289 : /// This returns quickly, after just creating the in-memory object
1290 : /// Tenant struct and launching a background task to download
1291 : /// the remote index files. On return, the tenant is most likely still in
1292 : /// Attaching state, and it will become Active once the background task
1293 : /// finishes. You can use wait_until_active() to wait for the task to
1294 : /// complete.
1295 : ///
1296 : #[allow(clippy::too_many_arguments)]
1297 0 : pub(crate) fn spawn(
1298 0 : conf: &'static PageServerConf,
1299 0 : tenant_shard_id: TenantShardId,
1300 0 : resources: TenantSharedResources,
1301 0 : attached_conf: AttachedTenantConf,
1302 0 : shard_identity: ShardIdentity,
1303 0 : init_order: Option<InitializationOrder>,
1304 0 : mode: SpawnMode,
1305 0 : ctx: &RequestContext,
1306 0 : ) -> Result<Arc<TenantShard>, GlobalShutDown> {
1307 0 : let wal_redo_manager =
1308 0 : WalRedoManager::new(PostgresRedoManager::new(conf, tenant_shard_id))?;
1309 :
1310 : let TenantSharedResources {
1311 0 : broker_client,
1312 0 : remote_storage,
1313 0 : deletion_queue_client,
1314 0 : l0_flush_global_state,
1315 0 : } = resources;
1316 0 :
1317 0 : let attach_mode = attached_conf.location.attach_mode;
1318 0 : let generation = attached_conf.location.generation;
1319 0 :
1320 0 : let tenant = Arc::new(TenantShard::new(
1321 0 : TenantState::Attaching,
1322 0 : conf,
1323 0 : attached_conf,
1324 0 : shard_identity,
1325 0 : Some(wal_redo_manager),
1326 0 : tenant_shard_id,
1327 0 : remote_storage.clone(),
1328 0 : deletion_queue_client,
1329 0 : l0_flush_global_state,
1330 0 : ));
1331 0 :
1332 0 : // The attach task will carry a GateGuard, so that shutdown() reliably waits for it to drop out if
1333 0 : // we shut down while attaching.
1334 0 : let attach_gate_guard = tenant
1335 0 : .gate
1336 0 : .enter()
1337 0 : .expect("We just created the TenantShard: nothing else can have shut it down yet");
1338 0 :
1339 0 : // Do all the hard work in the background
1340 0 : let tenant_clone = Arc::clone(&tenant);
1341 0 : let ctx = ctx.detached_child(TaskKind::Attach, DownloadBehavior::Warn);
1342 0 : task_mgr::spawn(
1343 0 : &tokio::runtime::Handle::current(),
1344 0 : TaskKind::Attach,
1345 0 : tenant_shard_id,
1346 0 : None,
1347 0 : "attach tenant",
1348 0 : async move {
1349 0 :
1350 0 : info!(
1351 : ?attach_mode,
1352 0 : "Attaching tenant"
1353 : );
1354 :
1355 0 : let _gate_guard = attach_gate_guard;
1356 0 :
1357 0 : // Is this tenant being spawned as part of process startup?
1358 0 : let starting_up = init_order.is_some();
1359 0 : scopeguard::defer! {
1360 0 : if starting_up {
1361 0 : TENANT.startup_complete.inc();
1362 0 : }
1363 0 : }
1364 :
1365 0 : fn make_broken_or_stopping(t: &TenantShard, err: anyhow::Error) {
1366 0 : t.state.send_modify(|state| match state {
1367 : // TODO: the old code alluded to DeleteTenantFlow sometimes setting
1368 : // TenantState::Stopping before we get here, but this may be outdated.
1369 : // Let's find out with a testing assertion. If this doesn't fire, and the
1370 : // logs don't show this happening in production, remove the Stopping cases.
1371 0 : TenantState::Stopping{..} if cfg!(any(test, feature = "testing")) => {
1372 0 : panic!("unexpected TenantState::Stopping during attach")
1373 : }
1374 : // If the tenant is cancelled, assume the error was caused by cancellation.
1375 0 : TenantState::Attaching if t.cancel.is_cancelled() => {
1376 0 : info!("attach cancelled, setting tenant state to Stopping: {err}");
1377 : // NB: progress None tells `set_stopping` that attach has cancelled.
1378 0 : *state = TenantState::Stopping { progress: None };
1379 : }
1380 : // According to the old code, DeleteTenantFlow may already have set this to
1381 : // Stopping. Retain its progress.
1382 : // TODO: there is no DeleteTenantFlow. Is this still needed? See above.
1383 0 : TenantState::Stopping { progress } if t.cancel.is_cancelled() => {
1384 0 : assert!(progress.is_some(), "concurrent attach cancellation");
1385 0 : info!("attach cancelled, already Stopping: {err}");
1386 : }
1387 : // Mark the tenant as broken.
1388 : TenantState::Attaching | TenantState::Stopping { .. } => {
1389 0 : error!("attach failed, setting tenant state to Broken (was {state}): {err:?}");
1390 0 : *state = TenantState::broken_from_reason(err.to_string())
1391 : }
1392 : // The attach task owns the tenant state until activated.
1393 0 : state => panic!("invalid tenant state {state} during attach: {err:?}"),
1394 0 : });
1395 0 : }
1396 :
1397 : // TODO: should also be rejecting tenant conf changes that violate this check.
1398 0 : if let Err(e) = crate::tenant::storage_layer::inmemory_layer::IndexEntry::validate_checkpoint_distance(tenant_clone.get_checkpoint_distance()) {
1399 0 : make_broken_or_stopping(&tenant_clone, anyhow::anyhow!(e));
1400 0 : return Ok(());
1401 0 : }
1402 0 :
1403 0 : let mut init_order = init_order;
1404 0 : // take the completion because initial tenant loading will complete when all of
1405 0 : // these tasks complete.
1406 0 : let _completion = init_order
1407 0 : .as_mut()
1408 0 : .and_then(|x| x.initial_tenant_load.take());
1409 0 : let remote_load_completion = init_order
1410 0 : .as_mut()
1411 0 : .and_then(|x| x.initial_tenant_load_remote.take());
1412 :
1413 : enum AttachType<'a> {
1414 : /// We are attaching this tenant lazily in the background.
1415 : Warmup {
1416 : _permit: tokio::sync::SemaphorePermit<'a>,
1417 : during_startup: bool
1418 : },
1419 : /// We are attaching this tenant as soon as we can, because for example an
1420 : /// endpoint tried to access it.
1421 : OnDemand,
1422 : /// During normal operations after startup, we are attaching a tenant, and
1423 : /// eager attach was requested.
1424 : Normal,
1425 : }
1426 :
1427 0 : let attach_type = if matches!(mode, SpawnMode::Lazy) {
1428 : // Before doing any I/O, wait for at least one of:
1429 : // - A client attempting to access to this tenant (on-demand loading)
1430 : // - A permit becoming available in the warmup semaphore (background warmup)
1431 :
1432 0 : tokio::select!(
1433 0 : permit = tenant_clone.activate_now_sem.acquire() => {
1434 0 : let _ = permit.expect("activate_now_sem is never closed");
1435 0 : tracing::info!("Activating tenant (on-demand)");
1436 0 : AttachType::OnDemand
1437 : },
1438 0 : permit = conf.concurrent_tenant_warmup.inner().acquire() => {
1439 0 : let _permit = permit.expect("concurrent_tenant_warmup semaphore is never closed");
1440 0 : tracing::info!("Activating tenant (warmup)");
1441 0 : AttachType::Warmup {
1442 0 : _permit,
1443 0 : during_startup: init_order.is_some()
1444 0 : }
1445 : }
1446 0 : _ = tenant_clone.cancel.cancelled() => {
1447 : // This is safe, but should be pretty rare: it is interesting if a tenant
1448 : // stayed in Activating for such a long time that shutdown found it in
1449 : // that state.
1450 0 : tracing::info!(state=%tenant_clone.current_state(), "Tenant shut down before activation");
1451 : // Set the tenant to Stopping to signal `set_stopping` that we're done.
1452 0 : make_broken_or_stopping(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"));
1453 0 : return Ok(());
1454 : },
1455 : )
1456 : } else {
1457 : // SpawnMode::{Create,Eager} always cause jumping ahead of the
1458 : // concurrent_tenant_warmup queue
1459 0 : AttachType::Normal
1460 : };
1461 :
1462 0 : let preload = match &mode {
1463 : SpawnMode::Eager | SpawnMode::Lazy => {
1464 0 : let _preload_timer = TENANT.preload.start_timer();
1465 0 : let res = tenant_clone
1466 0 : .preload(&remote_storage, task_mgr::shutdown_token())
1467 0 : .await;
1468 0 : match res {
1469 0 : Ok(p) => Some(p),
1470 0 : Err(e) => {
1471 0 : make_broken_or_stopping(&tenant_clone, anyhow::anyhow!(e));
1472 0 : return Ok(());
1473 : }
1474 : }
1475 : }
1476 :
1477 : };
1478 :
1479 : // Remote preload is complete.
1480 0 : drop(remote_load_completion);
1481 0 :
1482 0 :
1483 0 : // We will time the duration of the attach phase unless this is a creation (attach will do no work)
1484 0 : let attach_start = std::time::Instant::now();
1485 0 : let attached = {
1486 0 : let _attach_timer = Some(TENANT.attach.start_timer());
1487 0 : tenant_clone.attach(preload, &ctx).await
1488 : };
1489 0 : let attach_duration = attach_start.elapsed();
1490 0 : _ = tenant_clone.attach_wal_lag_cooldown.set(WalLagCooldown::new(attach_start, attach_duration));
1491 0 :
1492 0 : match attached {
1493 : Ok(()) => {
1494 0 : info!("attach finished, activating");
1495 0 : tenant_clone.activate(broker_client, None, &ctx);
1496 : }
1497 0 : Err(e) => make_broken_or_stopping(&tenant_clone, anyhow::anyhow!(e)),
1498 : }
1499 :
1500 : // If we are doing an opportunistic warmup attachment at startup, initialize
1501 : // logical size at the same time. This is better than starting a bunch of idle tenants
1502 : // with cold caches and then coming back later to initialize their logical sizes.
1503 : //
1504 : // It also prevents the warmup proccess competing with the concurrency limit on
1505 : // logical size calculations: if logical size calculation semaphore is saturated,
1506 : // then warmup will wait for that before proceeding to the next tenant.
1507 0 : if matches!(attach_type, AttachType::Warmup { during_startup: true, .. }) {
1508 0 : let mut futs: FuturesUnordered<_> = tenant_clone.timelines.lock().unwrap().values().cloned().map(|t| t.await_initial_logical_size()).collect();
1509 0 : tracing::info!("Waiting for initial logical sizes while warming up...");
1510 0 : while futs.next().await.is_some() {}
1511 0 : tracing::info!("Warm-up complete");
1512 0 : }
1513 :
1514 0 : Ok(())
1515 0 : }
1516 0 : .instrument(tracing::info_span!(parent: None, "attach", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), gen=?generation)),
1517 : );
1518 0 : Ok(tenant)
1519 0 : }
1520 :
1521 : #[instrument(skip_all)]
1522 : pub(crate) async fn preload(
1523 : self: &Arc<Self>,
1524 : remote_storage: &GenericRemoteStorage,
1525 : cancel: CancellationToken,
1526 : ) -> anyhow::Result<TenantPreload> {
1527 : span::debug_assert_current_span_has_tenant_id();
1528 : // Get list of remote timelines
1529 : // download index files for every tenant timeline
1530 : info!("listing remote timelines");
1531 : let (mut remote_timeline_ids, other_keys) = remote_timeline_client::list_remote_timelines(
1532 : remote_storage,
1533 : self.tenant_shard_id,
1534 : cancel.clone(),
1535 : )
1536 : .await?;
1537 :
1538 : let tenant_manifest = match download_tenant_manifest(
1539 : remote_storage,
1540 : &self.tenant_shard_id,
1541 : self.generation,
1542 : &cancel,
1543 : )
1544 : .await
1545 : {
1546 : Ok((tenant_manifest, _, _)) => Some(tenant_manifest),
1547 : Err(DownloadError::NotFound) => None,
1548 : Err(err) => return Err(err.into()),
1549 : };
1550 :
1551 : info!(
1552 : "found {} timelines ({} offloaded timelines)",
1553 : remote_timeline_ids.len(),
1554 : tenant_manifest
1555 : .as_ref()
1556 36 : .map(|m| m.offloaded_timelines.len())
1557 : .unwrap_or(0)
1558 : );
1559 :
1560 : for k in other_keys {
1561 : warn!("Unexpected non timeline key {k}");
1562 : }
1563 :
1564 : // Avoid downloading IndexPart of offloaded timelines.
1565 : let mut offloaded_with_prefix = HashSet::new();
1566 : if let Some(tenant_manifest) = &tenant_manifest {
1567 : for offloaded in tenant_manifest.offloaded_timelines.iter() {
1568 : if remote_timeline_ids.remove(&offloaded.timeline_id) {
1569 : offloaded_with_prefix.insert(offloaded.timeline_id);
1570 : } else {
1571 : // We'll take care later of timelines in the manifest without a prefix
1572 : }
1573 : }
1574 : }
1575 :
1576 : // TODO(vlad): Could go to S3 if the secondary is freezing cold and hasn't even
1577 : // pulled the first heatmap. Not entirely necessary since the storage controller
1578 : // will kick the secondary in any case and cause a download.
1579 : let maybe_heatmap_at = self.read_on_disk_heatmap().await;
1580 :
1581 : let timelines = self
1582 : .load_timelines_metadata(
1583 : remote_timeline_ids,
1584 : remote_storage,
1585 : maybe_heatmap_at,
1586 : cancel,
1587 : )
1588 : .await?;
1589 :
1590 : Ok(TenantPreload {
1591 : tenant_manifest,
1592 : timelines: timelines
1593 : .into_iter()
1594 36 : .map(|(id, tl)| (id, Some(tl)))
1595 0 : .chain(offloaded_with_prefix.into_iter().map(|id| (id, None)))
1596 : .collect(),
1597 : })
1598 : }
1599 :
1600 1392 : async fn read_on_disk_heatmap(&self) -> Option<(HeatMapTenant, std::time::Instant)> {
1601 1392 : if !self.conf.load_previous_heatmap {
1602 0 : return None;
1603 1392 : }
1604 1392 :
1605 1392 : let on_disk_heatmap_path = self.conf.tenant_heatmap_path(&self.tenant_shard_id);
1606 1392 : match tokio::fs::read_to_string(on_disk_heatmap_path).await {
1607 0 : Ok(heatmap) => match serde_json::from_str::<HeatMapTenant>(&heatmap) {
1608 0 : Ok(heatmap) => Some((heatmap, std::time::Instant::now())),
1609 0 : Err(err) => {
1610 0 : error!("Failed to deserialize old heatmap: {err}");
1611 0 : None
1612 : }
1613 : },
1614 1392 : Err(err) => match err.kind() {
1615 1392 : std::io::ErrorKind::NotFound => None,
1616 : _ => {
1617 0 : error!("Unexpected IO error reading old heatmap: {err}");
1618 0 : None
1619 : }
1620 : },
1621 : }
1622 1392 : }
1623 :
1624 : ///
1625 : /// Background task that downloads all data for a tenant and brings it to Active state.
1626 : ///
1627 : /// No background tasks are started as part of this routine.
1628 : ///
1629 1392 : async fn attach(
1630 1392 : self: &Arc<TenantShard>,
1631 1392 : preload: Option<TenantPreload>,
1632 1392 : ctx: &RequestContext,
1633 1392 : ) -> anyhow::Result<()> {
1634 1392 : span::debug_assert_current_span_has_tenant_id();
1635 1392 :
1636 1392 : failpoint_support::sleep_millis_async!("before-attaching-tenant");
1637 :
1638 1392 : let Some(preload) = preload else {
1639 0 : anyhow::bail!(
1640 0 : "local-only deployment is no longer supported, https://github.com/neondatabase/neon/issues/5624"
1641 0 : );
1642 : };
1643 :
1644 1392 : let mut offloaded_timeline_ids = HashSet::new();
1645 1392 : let mut offloaded_timelines_list = Vec::new();
1646 1392 : if let Some(tenant_manifest) = &preload.tenant_manifest {
1647 36 : for timeline_manifest in tenant_manifest.offloaded_timelines.iter() {
1648 0 : let timeline_id = timeline_manifest.timeline_id;
1649 0 : let offloaded_timeline =
1650 0 : OffloadedTimeline::from_manifest(self.tenant_shard_id, timeline_manifest);
1651 0 : offloaded_timelines_list.push((timeline_id, Arc::new(offloaded_timeline)));
1652 0 : offloaded_timeline_ids.insert(timeline_id);
1653 0 : }
1654 1356 : }
1655 : // Complete deletions for offloaded timeline id's from manifest.
1656 : // The manifest will be uploaded later in this function.
1657 1392 : offloaded_timelines_list
1658 1392 : .retain(|(offloaded_id, offloaded)| {
1659 0 : // Existence of a timeline is finally determined by the existence of an index-part.json in remote storage.
1660 0 : // If there is dangling references in another location, they need to be cleaned up.
1661 0 : let delete = !preload.timelines.contains_key(offloaded_id);
1662 0 : if delete {
1663 0 : tracing::info!("Removing offloaded timeline {offloaded_id} from manifest as no remote prefix was found");
1664 0 : offloaded.defuse_for_tenant_drop();
1665 0 : }
1666 0 : !delete
1667 1392 : });
1668 1392 :
1669 1392 : let mut timelines_to_resume_deletions = vec![];
1670 1392 :
1671 1392 : let mut remote_index_and_client = HashMap::new();
1672 1392 : let mut timeline_ancestors = HashMap::new();
1673 1392 : let mut existent_timelines = HashSet::new();
1674 1428 : for (timeline_id, preload) in preload.timelines {
1675 36 : let Some(preload) = preload else { continue };
1676 : // This is an invariant of the `preload` function's API
1677 36 : assert!(!offloaded_timeline_ids.contains(&timeline_id));
1678 36 : let index_part = match preload.index_part {
1679 36 : Ok(i) => {
1680 36 : debug!("remote index part exists for timeline {timeline_id}");
1681 : // We found index_part on the remote, this is the standard case.
1682 36 : existent_timelines.insert(timeline_id);
1683 36 : i
1684 : }
1685 : Err(DownloadError::NotFound) => {
1686 : // There is no index_part on the remote. We only get here
1687 : // if there is some prefix for the timeline in the remote storage.
1688 : // This can e.g. be the initdb.tar.zst archive, maybe a
1689 : // remnant from a prior incomplete creation or deletion attempt.
1690 : // Delete the local directory as the deciding criterion for a
1691 : // timeline's existence is presence of index_part.
1692 0 : info!(%timeline_id, "index_part not found on remote");
1693 0 : continue;
1694 : }
1695 0 : Err(DownloadError::Fatal(why)) => {
1696 0 : // If, while loading one remote timeline, we saw an indication that our generation
1697 0 : // number is likely invalid, then we should not load the whole tenant.
1698 0 : error!(%timeline_id, "Fatal error loading timeline: {why}");
1699 0 : anyhow::bail!(why.to_string());
1700 : }
1701 0 : Err(e) => {
1702 0 : // Some (possibly ephemeral) error happened during index_part download.
1703 0 : // Pretend the timeline exists to not delete the timeline directory,
1704 0 : // as it might be a temporary issue and we don't want to re-download
1705 0 : // everything after it resolves.
1706 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1707 :
1708 0 : existent_timelines.insert(timeline_id);
1709 0 : continue;
1710 : }
1711 : };
1712 36 : match index_part {
1713 36 : MaybeDeletedIndexPart::IndexPart(index_part) => {
1714 36 : timeline_ancestors.insert(timeline_id, index_part.metadata.clone());
1715 36 : remote_index_and_client.insert(
1716 36 : timeline_id,
1717 36 : (index_part, preload.client, preload.previous_heatmap),
1718 36 : );
1719 36 : }
1720 0 : MaybeDeletedIndexPart::Deleted(index_part) => {
1721 0 : info!(
1722 0 : "timeline {} is deleted, picking to resume deletion",
1723 : timeline_id
1724 : );
1725 0 : timelines_to_resume_deletions.push((timeline_id, index_part, preload.client));
1726 : }
1727 : }
1728 : }
1729 :
1730 1392 : let mut gc_blocks = HashMap::new();
1731 :
1732 : // For every timeline, download the metadata file, scan the local directory,
1733 : // and build a layer map that contains an entry for each remote and local
1734 : // layer file.
1735 1392 : let sorted_timelines = tree_sort_timelines(timeline_ancestors, |m| m.ancestor_timeline())?;
1736 1428 : for (timeline_id, remote_metadata) in sorted_timelines {
1737 36 : let (index_part, remote_client, previous_heatmap) = remote_index_and_client
1738 36 : .remove(&timeline_id)
1739 36 : .expect("just put it in above");
1740 :
1741 36 : if let Some(blocking) = index_part.gc_blocking.as_ref() {
1742 : // could just filter these away, but it helps while testing
1743 0 : anyhow::ensure!(
1744 0 : !blocking.reasons.is_empty(),
1745 0 : "index_part for {timeline_id} is malformed: it should not have gc blocking with zero reasons"
1746 : );
1747 0 : let prev = gc_blocks.insert(timeline_id, blocking.reasons);
1748 0 : assert!(prev.is_none());
1749 36 : }
1750 :
1751 : // TODO again handle early failure
1752 36 : let effect = self
1753 36 : .load_remote_timeline(
1754 36 : timeline_id,
1755 36 : index_part,
1756 36 : remote_metadata,
1757 36 : previous_heatmap,
1758 36 : self.get_timeline_resources_for(remote_client),
1759 36 : LoadTimelineCause::Attach,
1760 36 : ctx,
1761 36 : )
1762 36 : .await
1763 36 : .with_context(|| {
1764 0 : format!(
1765 0 : "failed to load remote timeline {} for tenant {}",
1766 0 : timeline_id, self.tenant_shard_id
1767 0 : )
1768 36 : })?;
1769 :
1770 36 : match effect {
1771 36 : TimelineInitAndSyncResult::ReadyToActivate(_) => {
1772 36 : // activation happens later, on Tenant::activate
1773 36 : }
1774 : TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
1775 : TimelineInitAndSyncNeedsSpawnImportPgdata {
1776 0 : timeline,
1777 0 : import_pgdata,
1778 0 : guard,
1779 0 : },
1780 0 : ) => {
1781 0 : tokio::task::spawn(self.clone().create_timeline_import_pgdata_task(
1782 0 : timeline,
1783 0 : import_pgdata,
1784 0 : ActivateTimelineArgs::No,
1785 0 : guard,
1786 0 : ctx.detached_child(TaskKind::ImportPgdata, DownloadBehavior::Warn),
1787 0 : ));
1788 0 : }
1789 : }
1790 : }
1791 :
1792 : // Walk through deleted timelines, resume deletion
1793 1392 : for (timeline_id, index_part, remote_timeline_client) in timelines_to_resume_deletions {
1794 0 : remote_timeline_client
1795 0 : .init_upload_queue_stopped_to_continue_deletion(&index_part)
1796 0 : .context("init queue stopped")
1797 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1798 :
1799 0 : DeleteTimelineFlow::resume_deletion(
1800 0 : Arc::clone(self),
1801 0 : timeline_id,
1802 0 : &index_part.metadata,
1803 0 : remote_timeline_client,
1804 0 : ctx,
1805 0 : )
1806 0 : .instrument(tracing::info_span!("timeline_delete", %timeline_id))
1807 0 : .await
1808 0 : .context("resume_deletion")
1809 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1810 : }
1811 1392 : {
1812 1392 : let mut offloaded_timelines_accessor = self.timelines_offloaded.lock().unwrap();
1813 1392 : offloaded_timelines_accessor.extend(offloaded_timelines_list.into_iter());
1814 1392 : }
1815 :
1816 : // Stash the preloaded tenant manifest, and upload a new manifest if changed.
1817 : //
1818 : // NB: this must happen after the tenant is fully populated above. In particular the
1819 : // offloaded timelines, which are included in the manifest.
1820 : {
1821 1392 : let mut guard = self.remote_tenant_manifest.lock().await;
1822 1392 : assert!(guard.is_none(), "tenant manifest set before preload"); // first populated here
1823 1392 : *guard = preload.tenant_manifest;
1824 1392 : }
1825 1392 : self.maybe_upload_tenant_manifest().await?;
1826 :
1827 : // The local filesystem contents are a cache of what's in the remote IndexPart;
1828 : // IndexPart is the source of truth.
1829 1392 : self.clean_up_timelines(&existent_timelines)?;
1830 :
1831 1392 : self.gc_block.set_scanned(gc_blocks);
1832 1392 :
1833 1392 : fail::fail_point!("attach-before-activate", |_| {
1834 0 : anyhow::bail!("attach-before-activate");
1835 1392 : });
1836 1392 : failpoint_support::sleep_millis_async!("attach-before-activate-sleep", &self.cancel);
1837 :
1838 1392 : info!("Done");
1839 :
1840 1392 : Ok(())
1841 1392 : }
1842 :
1843 : /// Check for any local timeline directories that are temporary, or do not correspond to a
1844 : /// timeline that still exists: this can happen if we crashed during a deletion/creation, or
1845 : /// if a timeline was deleted while the tenant was attached to a different pageserver.
1846 1392 : fn clean_up_timelines(&self, existent_timelines: &HashSet<TimelineId>) -> anyhow::Result<()> {
1847 1392 : let timelines_dir = self.conf.timelines_path(&self.tenant_shard_id);
1848 :
1849 1392 : let entries = match timelines_dir.read_dir_utf8() {
1850 1392 : Ok(d) => d,
1851 0 : Err(e) => {
1852 0 : if e.kind() == std::io::ErrorKind::NotFound {
1853 0 : return Ok(());
1854 : } else {
1855 0 : return Err(e).context("list timelines directory for tenant");
1856 : }
1857 : }
1858 : };
1859 :
1860 1440 : for entry in entries {
1861 48 : let entry = entry.context("read timeline dir entry")?;
1862 48 : let entry_path = entry.path();
1863 :
1864 48 : let purge = if crate::is_temporary(entry_path) {
1865 0 : true
1866 : } else {
1867 48 : match TimelineId::try_from(entry_path.file_name()) {
1868 48 : Ok(i) => {
1869 48 : // Purge if the timeline ID does not exist in remote storage: remote storage is the authority.
1870 48 : !existent_timelines.contains(&i)
1871 : }
1872 0 : Err(e) => {
1873 0 : tracing::warn!(
1874 0 : "Unparseable directory in timelines directory: {entry_path}, ignoring ({e})"
1875 : );
1876 : // Do not purge junk: if we don't recognize it, be cautious and leave it for a human.
1877 0 : false
1878 : }
1879 : }
1880 : };
1881 :
1882 48 : if purge {
1883 12 : tracing::info!("Purging stale timeline dentry {entry_path}");
1884 12 : if let Err(e) = match entry.file_type() {
1885 12 : Ok(t) => if t.is_dir() {
1886 12 : std::fs::remove_dir_all(entry_path)
1887 : } else {
1888 0 : std::fs::remove_file(entry_path)
1889 : }
1890 12 : .or_else(fs_ext::ignore_not_found),
1891 0 : Err(e) => Err(e),
1892 : } {
1893 0 : tracing::warn!("Failed to purge stale timeline dentry {entry_path}: {e}");
1894 12 : }
1895 36 : }
1896 : }
1897 :
1898 1392 : Ok(())
1899 1392 : }
1900 :
1901 : /// Get sum of all remote timelines sizes
1902 : ///
1903 : /// This function relies on the index_part instead of listing the remote storage
1904 0 : pub fn remote_size(&self) -> u64 {
1905 0 : let mut size = 0;
1906 :
1907 0 : for timeline in self.list_timelines() {
1908 0 : size += timeline.remote_client.get_remote_physical_size();
1909 0 : }
1910 :
1911 0 : size
1912 0 : }
1913 :
1914 : #[instrument(skip_all, fields(timeline_id=%timeline_id))]
1915 : #[allow(clippy::too_many_arguments)]
1916 : async fn load_remote_timeline(
1917 : self: &Arc<Self>,
1918 : timeline_id: TimelineId,
1919 : index_part: IndexPart,
1920 : remote_metadata: TimelineMetadata,
1921 : previous_heatmap: Option<PreviousHeatmap>,
1922 : resources: TimelineResources,
1923 : cause: LoadTimelineCause,
1924 : ctx: &RequestContext,
1925 : ) -> anyhow::Result<TimelineInitAndSyncResult> {
1926 : span::debug_assert_current_span_has_tenant_id();
1927 :
1928 : info!("downloading index file for timeline {}", timeline_id);
1929 : tokio::fs::create_dir_all(self.conf.timeline_path(&self.tenant_shard_id, &timeline_id))
1930 : .await
1931 : .context("Failed to create new timeline directory")?;
1932 :
1933 : let ancestor = if let Some(ancestor_id) = remote_metadata.ancestor_timeline() {
1934 : let timelines = self.timelines.lock().unwrap();
1935 : Some(Arc::clone(timelines.get(&ancestor_id).ok_or_else(
1936 0 : || {
1937 0 : anyhow::anyhow!(
1938 0 : "cannot find ancestor timeline {ancestor_id} for timeline {timeline_id}"
1939 0 : )
1940 0 : },
1941 : )?))
1942 : } else {
1943 : None
1944 : };
1945 :
1946 : self.timeline_init_and_sync(
1947 : timeline_id,
1948 : resources,
1949 : index_part,
1950 : remote_metadata,
1951 : previous_heatmap,
1952 : ancestor,
1953 : cause,
1954 : ctx,
1955 : )
1956 : .await
1957 : }
1958 :
1959 1392 : async fn load_timelines_metadata(
1960 1392 : self: &Arc<TenantShard>,
1961 1392 : timeline_ids: HashSet<TimelineId>,
1962 1392 : remote_storage: &GenericRemoteStorage,
1963 1392 : heatmap: Option<(HeatMapTenant, std::time::Instant)>,
1964 1392 : cancel: CancellationToken,
1965 1392 : ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
1966 1392 : let mut timeline_heatmaps = heatmap.map(|h| (h.0.into_timelines_index(), h.1));
1967 1392 :
1968 1392 : let mut part_downloads = JoinSet::new();
1969 1428 : for timeline_id in timeline_ids {
1970 36 : let cancel_clone = cancel.clone();
1971 36 :
1972 36 : let previous_timeline_heatmap = timeline_heatmaps.as_mut().and_then(|hs| {
1973 0 : hs.0.remove(&timeline_id).map(|h| PreviousHeatmap::Active {
1974 0 : heatmap: h,
1975 0 : read_at: hs.1,
1976 0 : end_lsn: None,
1977 0 : })
1978 36 : });
1979 36 : part_downloads.spawn(
1980 36 : self.load_timeline_metadata(
1981 36 : timeline_id,
1982 36 : remote_storage.clone(),
1983 36 : previous_timeline_heatmap,
1984 36 : cancel_clone,
1985 36 : )
1986 36 : .instrument(info_span!("download_index_part", %timeline_id)),
1987 : );
1988 : }
1989 :
1990 1392 : let mut timeline_preloads: HashMap<TimelineId, TimelinePreload> = HashMap::new();
1991 :
1992 : loop {
1993 1428 : tokio::select!(
1994 1428 : next = part_downloads.join_next() => {
1995 1428 : match next {
1996 36 : Some(result) => {
1997 36 : let preload = result.context("join preload task")?;
1998 36 : timeline_preloads.insert(preload.timeline_id, preload);
1999 : },
2000 : None => {
2001 1392 : break;
2002 : }
2003 : }
2004 : },
2005 1428 : _ = cancel.cancelled() => {
2006 0 : anyhow::bail!("Cancelled while waiting for remote index download")
2007 : }
2008 : )
2009 : }
2010 :
2011 1392 : Ok(timeline_preloads)
2012 1392 : }
2013 :
2014 36 : fn build_timeline_client(
2015 36 : &self,
2016 36 : timeline_id: TimelineId,
2017 36 : remote_storage: GenericRemoteStorage,
2018 36 : ) -> RemoteTimelineClient {
2019 36 : RemoteTimelineClient::new(
2020 36 : remote_storage.clone(),
2021 36 : self.deletion_queue_client.clone(),
2022 36 : self.conf,
2023 36 : self.tenant_shard_id,
2024 36 : timeline_id,
2025 36 : self.generation,
2026 36 : &self.tenant_conf.load().location,
2027 36 : )
2028 36 : }
2029 :
2030 36 : fn load_timeline_metadata(
2031 36 : self: &Arc<TenantShard>,
2032 36 : timeline_id: TimelineId,
2033 36 : remote_storage: GenericRemoteStorage,
2034 36 : previous_heatmap: Option<PreviousHeatmap>,
2035 36 : cancel: CancellationToken,
2036 36 : ) -> impl Future<Output = TimelinePreload> + use<> {
2037 36 : let client = self.build_timeline_client(timeline_id, remote_storage);
2038 36 : async move {
2039 36 : debug_assert_current_span_has_tenant_and_timeline_id();
2040 36 : debug!("starting index part download");
2041 :
2042 36 : let index_part = client.download_index_file(&cancel).await;
2043 :
2044 36 : debug!("finished index part download");
2045 :
2046 36 : TimelinePreload {
2047 36 : client,
2048 36 : timeline_id,
2049 36 : index_part,
2050 36 : previous_heatmap,
2051 36 : }
2052 36 : }
2053 36 : }
2054 :
2055 0 : fn check_to_be_archived_has_no_unarchived_children(
2056 0 : timeline_id: TimelineId,
2057 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
2058 0 : ) -> Result<(), TimelineArchivalError> {
2059 0 : let children: Vec<TimelineId> = timelines
2060 0 : .iter()
2061 0 : .filter_map(|(id, entry)| {
2062 0 : if entry.get_ancestor_timeline_id() != Some(timeline_id) {
2063 0 : return None;
2064 0 : }
2065 0 : if entry.is_archived() == Some(true) {
2066 0 : return None;
2067 0 : }
2068 0 : Some(*id)
2069 0 : })
2070 0 : .collect();
2071 0 :
2072 0 : if !children.is_empty() {
2073 0 : return Err(TimelineArchivalError::HasUnarchivedChildren(children));
2074 0 : }
2075 0 : Ok(())
2076 0 : }
2077 :
2078 0 : fn check_ancestor_of_to_be_unarchived_is_not_archived(
2079 0 : ancestor_timeline_id: TimelineId,
2080 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
2081 0 : offloaded_timelines: &std::sync::MutexGuard<
2082 0 : '_,
2083 0 : HashMap<TimelineId, Arc<OffloadedTimeline>>,
2084 0 : >,
2085 0 : ) -> Result<(), TimelineArchivalError> {
2086 0 : let has_archived_parent =
2087 0 : if let Some(ancestor_timeline) = timelines.get(&ancestor_timeline_id) {
2088 0 : ancestor_timeline.is_archived() == Some(true)
2089 0 : } else if offloaded_timelines.contains_key(&ancestor_timeline_id) {
2090 0 : true
2091 : } else {
2092 0 : error!("ancestor timeline {ancestor_timeline_id} not found");
2093 0 : if cfg!(debug_assertions) {
2094 0 : panic!("ancestor timeline {ancestor_timeline_id} not found");
2095 0 : }
2096 0 : return Err(TimelineArchivalError::NotFound);
2097 : };
2098 0 : if has_archived_parent {
2099 0 : return Err(TimelineArchivalError::HasArchivedParent(
2100 0 : ancestor_timeline_id,
2101 0 : ));
2102 0 : }
2103 0 : Ok(())
2104 0 : }
2105 :
2106 0 : fn check_to_be_unarchived_timeline_has_no_archived_parent(
2107 0 : timeline: &Arc<Timeline>,
2108 0 : ) -> Result<(), TimelineArchivalError> {
2109 0 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
2110 0 : if ancestor_timeline.is_archived() == Some(true) {
2111 0 : return Err(TimelineArchivalError::HasArchivedParent(
2112 0 : ancestor_timeline.timeline_id,
2113 0 : ));
2114 0 : }
2115 0 : }
2116 0 : Ok(())
2117 0 : }
2118 :
2119 : /// Loads the specified (offloaded) timeline from S3 and attaches it as a loaded timeline
2120 : ///
2121 : /// Counterpart to [`offload_timeline`].
2122 0 : async fn unoffload_timeline(
2123 0 : self: &Arc<Self>,
2124 0 : timeline_id: TimelineId,
2125 0 : broker_client: storage_broker::BrokerClientChannel,
2126 0 : ctx: RequestContext,
2127 0 : ) -> Result<Arc<Timeline>, TimelineArchivalError> {
2128 0 : info!("unoffloading timeline");
2129 :
2130 : // We activate the timeline below manually, so this must be called on an active tenant.
2131 : // We expect callers of this function to ensure this.
2132 0 : match self.current_state() {
2133 : TenantState::Activating { .. }
2134 : | TenantState::Attaching
2135 : | TenantState::Broken { .. } => {
2136 0 : panic!("Timeline expected to be active")
2137 : }
2138 0 : TenantState::Stopping { .. } => return Err(TimelineArchivalError::Cancelled),
2139 0 : TenantState::Active => {}
2140 0 : }
2141 0 : let cancel = self.cancel.clone();
2142 0 :
2143 0 : // Protect against concurrent attempts to use this TimelineId
2144 0 : // We don't care much about idempotency, as it's ensured a layer above.
2145 0 : let allow_offloaded = true;
2146 0 : let _create_guard = self
2147 0 : .create_timeline_create_guard(
2148 0 : timeline_id,
2149 0 : CreateTimelineIdempotency::FailWithConflict,
2150 0 : allow_offloaded,
2151 0 : )
2152 0 : .map_err(|err| match err {
2153 0 : TimelineExclusionError::AlreadyCreating => TimelineArchivalError::AlreadyInProgress,
2154 : TimelineExclusionError::AlreadyExists { .. } => {
2155 0 : TimelineArchivalError::Other(anyhow::anyhow!("Timeline already exists"))
2156 : }
2157 0 : TimelineExclusionError::Other(e) => TimelineArchivalError::Other(e),
2158 0 : TimelineExclusionError::ShuttingDown => TimelineArchivalError::Cancelled,
2159 0 : })?;
2160 :
2161 0 : let timeline_preload = self
2162 0 : .load_timeline_metadata(
2163 0 : timeline_id,
2164 0 : self.remote_storage.clone(),
2165 0 : None,
2166 0 : cancel.clone(),
2167 0 : )
2168 0 : .await;
2169 :
2170 0 : let index_part = match timeline_preload.index_part {
2171 0 : Ok(index_part) => {
2172 0 : debug!("remote index part exists for timeline {timeline_id}");
2173 0 : index_part
2174 : }
2175 : Err(DownloadError::NotFound) => {
2176 0 : error!(%timeline_id, "index_part not found on remote");
2177 0 : return Err(TimelineArchivalError::NotFound);
2178 : }
2179 0 : Err(DownloadError::Cancelled) => return Err(TimelineArchivalError::Cancelled),
2180 0 : Err(e) => {
2181 0 : // Some (possibly ephemeral) error happened during index_part download.
2182 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
2183 0 : return Err(TimelineArchivalError::Other(
2184 0 : anyhow::Error::new(e).context("downloading index_part from remote storage"),
2185 0 : ));
2186 : }
2187 : };
2188 0 : let index_part = match index_part {
2189 0 : MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
2190 0 : MaybeDeletedIndexPart::Deleted(_index_part) => {
2191 0 : info!("timeline is deleted according to index_part.json");
2192 0 : return Err(TimelineArchivalError::NotFound);
2193 : }
2194 : };
2195 0 : let remote_metadata = index_part.metadata.clone();
2196 0 : let timeline_resources = self.build_timeline_resources(timeline_id);
2197 0 : self.load_remote_timeline(
2198 0 : timeline_id,
2199 0 : index_part,
2200 0 : remote_metadata,
2201 0 : None,
2202 0 : timeline_resources,
2203 0 : LoadTimelineCause::Unoffload,
2204 0 : &ctx,
2205 0 : )
2206 0 : .await
2207 0 : .with_context(|| {
2208 0 : format!(
2209 0 : "failed to load remote timeline {} for tenant {}",
2210 0 : timeline_id, self.tenant_shard_id
2211 0 : )
2212 0 : })
2213 0 : .map_err(TimelineArchivalError::Other)?;
2214 :
2215 0 : let timeline = {
2216 0 : let timelines = self.timelines.lock().unwrap();
2217 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2218 0 : warn!("timeline not available directly after attach");
2219 : // This is not a panic because no locks are held between `load_remote_timeline`
2220 : // which puts the timeline into timelines, and our look into the timeline map.
2221 0 : return Err(TimelineArchivalError::Other(anyhow::anyhow!(
2222 0 : "timeline not available directly after attach"
2223 0 : )));
2224 : };
2225 0 : let mut offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2226 0 : match offloaded_timelines.remove(&timeline_id) {
2227 0 : Some(offloaded) => {
2228 0 : offloaded.delete_from_ancestor_with_timelines(&timelines);
2229 0 : }
2230 0 : None => warn!("timeline already removed from offloaded timelines"),
2231 : }
2232 :
2233 0 : self.initialize_gc_info(&timelines, &offloaded_timelines, Some(timeline_id));
2234 0 :
2235 0 : Arc::clone(timeline)
2236 0 : };
2237 0 :
2238 0 : // Upload new list of offloaded timelines to S3
2239 0 : self.maybe_upload_tenant_manifest().await?;
2240 :
2241 : // Activate the timeline (if it makes sense)
2242 0 : if !(timeline.is_broken() || timeline.is_stopping()) {
2243 0 : let background_jobs_can_start = None;
2244 0 : timeline.activate(
2245 0 : self.clone(),
2246 0 : broker_client.clone(),
2247 0 : background_jobs_can_start,
2248 0 : &ctx.with_scope_timeline(&timeline),
2249 0 : );
2250 0 : }
2251 :
2252 0 : info!("timeline unoffloading complete");
2253 0 : Ok(timeline)
2254 0 : }
2255 :
2256 0 : pub(crate) async fn apply_timeline_archival_config(
2257 0 : self: &Arc<Self>,
2258 0 : timeline_id: TimelineId,
2259 0 : new_state: TimelineArchivalState,
2260 0 : broker_client: storage_broker::BrokerClientChannel,
2261 0 : ctx: RequestContext,
2262 0 : ) -> Result<(), TimelineArchivalError> {
2263 0 : info!("setting timeline archival config");
2264 : // First part: figure out what is needed to do, and do validation
2265 0 : let timeline_or_unarchive_offloaded = 'outer: {
2266 0 : let timelines = self.timelines.lock().unwrap();
2267 :
2268 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2269 0 : let offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2270 0 : let Some(offloaded) = offloaded_timelines.get(&timeline_id) else {
2271 0 : return Err(TimelineArchivalError::NotFound);
2272 : };
2273 0 : if new_state == TimelineArchivalState::Archived {
2274 : // It's offloaded already, so nothing to do
2275 0 : return Ok(());
2276 0 : }
2277 0 : if let Some(ancestor_timeline_id) = offloaded.ancestor_timeline_id {
2278 0 : Self::check_ancestor_of_to_be_unarchived_is_not_archived(
2279 0 : ancestor_timeline_id,
2280 0 : &timelines,
2281 0 : &offloaded_timelines,
2282 0 : )?;
2283 0 : }
2284 0 : break 'outer None;
2285 : };
2286 :
2287 : // Do some validation. We release the timelines lock below, so there is potential
2288 : // for race conditions: these checks are more present to prevent misunderstandings of
2289 : // the API's capabilities, instead of serving as the sole way to defend their invariants.
2290 0 : match new_state {
2291 : TimelineArchivalState::Unarchived => {
2292 0 : Self::check_to_be_unarchived_timeline_has_no_archived_parent(timeline)?
2293 : }
2294 : TimelineArchivalState::Archived => {
2295 0 : Self::check_to_be_archived_has_no_unarchived_children(timeline_id, &timelines)?
2296 : }
2297 : }
2298 0 : Some(Arc::clone(timeline))
2299 : };
2300 :
2301 : // Second part: unoffload timeline (if needed)
2302 0 : let timeline = if let Some(timeline) = timeline_or_unarchive_offloaded {
2303 0 : timeline
2304 : } else {
2305 : // Turn offloaded timeline into a non-offloaded one
2306 0 : self.unoffload_timeline(timeline_id, broker_client, ctx)
2307 0 : .await?
2308 : };
2309 :
2310 : // Third part: upload new timeline archival state and block until it is present in S3
2311 0 : let upload_needed = match timeline
2312 0 : .remote_client
2313 0 : .schedule_index_upload_for_timeline_archival_state(new_state)
2314 : {
2315 0 : Ok(upload_needed) => upload_needed,
2316 0 : Err(e) => {
2317 0 : if timeline.cancel.is_cancelled() {
2318 0 : return Err(TimelineArchivalError::Cancelled);
2319 : } else {
2320 0 : return Err(TimelineArchivalError::Other(e));
2321 : }
2322 : }
2323 : };
2324 :
2325 0 : if upload_needed {
2326 0 : info!("Uploading new state");
2327 : const MAX_WAIT: Duration = Duration::from_secs(10);
2328 0 : let Ok(v) =
2329 0 : tokio::time::timeout(MAX_WAIT, timeline.remote_client.wait_completion()).await
2330 : else {
2331 0 : tracing::warn!("reached timeout for waiting on upload queue");
2332 0 : return Err(TimelineArchivalError::Timeout);
2333 : };
2334 0 : v.map_err(|e| match e {
2335 0 : WaitCompletionError::NotInitialized(e) => {
2336 0 : TimelineArchivalError::Other(anyhow::anyhow!(e))
2337 : }
2338 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2339 0 : TimelineArchivalError::Cancelled
2340 : }
2341 0 : })?;
2342 0 : }
2343 0 : Ok(())
2344 0 : }
2345 :
2346 12 : pub fn get_offloaded_timeline(
2347 12 : &self,
2348 12 : timeline_id: TimelineId,
2349 12 : ) -> Result<Arc<OffloadedTimeline>, GetTimelineError> {
2350 12 : self.timelines_offloaded
2351 12 : .lock()
2352 12 : .unwrap()
2353 12 : .get(&timeline_id)
2354 12 : .map(Arc::clone)
2355 12 : .ok_or(GetTimelineError::NotFound {
2356 12 : tenant_id: self.tenant_shard_id,
2357 12 : timeline_id,
2358 12 : })
2359 12 : }
2360 :
2361 24 : pub(crate) fn tenant_shard_id(&self) -> TenantShardId {
2362 24 : self.tenant_shard_id
2363 24 : }
2364 :
2365 : /// Get Timeline handle for given Neon timeline ID.
2366 : /// This function is idempotent. It doesn't change internal state in any way.
2367 1332 : pub fn get_timeline(
2368 1332 : &self,
2369 1332 : timeline_id: TimelineId,
2370 1332 : active_only: bool,
2371 1332 : ) -> Result<Arc<Timeline>, GetTimelineError> {
2372 1332 : let timelines_accessor = self.timelines.lock().unwrap();
2373 1332 : let timeline = timelines_accessor
2374 1332 : .get(&timeline_id)
2375 1332 : .ok_or(GetTimelineError::NotFound {
2376 1332 : tenant_id: self.tenant_shard_id,
2377 1332 : timeline_id,
2378 1332 : })?;
2379 :
2380 1320 : if active_only && !timeline.is_active() {
2381 0 : Err(GetTimelineError::NotActive {
2382 0 : tenant_id: self.tenant_shard_id,
2383 0 : timeline_id,
2384 0 : state: timeline.current_state(),
2385 0 : })
2386 : } else {
2387 1320 : Ok(Arc::clone(timeline))
2388 : }
2389 1332 : }
2390 :
2391 : /// Lists timelines the tenant contains.
2392 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2393 24 : pub fn list_timelines(&self) -> Vec<Arc<Timeline>> {
2394 24 : self.timelines
2395 24 : .lock()
2396 24 : .unwrap()
2397 24 : .values()
2398 24 : .map(Arc::clone)
2399 24 : .collect()
2400 24 : }
2401 :
2402 : /// Lists timelines the tenant manages, including offloaded ones.
2403 : ///
2404 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2405 0 : pub fn list_timelines_and_offloaded(
2406 0 : &self,
2407 0 : ) -> (Vec<Arc<Timeline>>, Vec<Arc<OffloadedTimeline>>) {
2408 0 : let timelines = self
2409 0 : .timelines
2410 0 : .lock()
2411 0 : .unwrap()
2412 0 : .values()
2413 0 : .map(Arc::clone)
2414 0 : .collect();
2415 0 : let offloaded = self
2416 0 : .timelines_offloaded
2417 0 : .lock()
2418 0 : .unwrap()
2419 0 : .values()
2420 0 : .map(Arc::clone)
2421 0 : .collect();
2422 0 : (timelines, offloaded)
2423 0 : }
2424 :
2425 0 : pub fn list_timeline_ids(&self) -> Vec<TimelineId> {
2426 0 : self.timelines.lock().unwrap().keys().cloned().collect()
2427 0 : }
2428 :
2429 : /// This is used by tests & import-from-basebackup.
2430 : ///
2431 : /// The returned [`UninitializedTimeline`] contains no data nor metadata and it is in
2432 : /// a state that will fail [`TenantShard::load_remote_timeline`] because `disk_consistent_lsn=Lsn(0)`.
2433 : ///
2434 : /// The caller is responsible for getting the timeline into a state that will be accepted
2435 : /// by [`TenantShard::load_remote_timeline`] / [`TenantShard::attach`].
2436 : /// Then they may call [`UninitializedTimeline::finish_creation`] to add the timeline
2437 : /// to the [`TenantShard::timelines`].
2438 : ///
2439 : /// Tests should use `TenantShard::create_test_timeline` to set up the minimum required metadata keys.
2440 1344 : pub(crate) async fn create_empty_timeline(
2441 1344 : self: &Arc<Self>,
2442 1344 : new_timeline_id: TimelineId,
2443 1344 : initdb_lsn: Lsn,
2444 1344 : pg_version: u32,
2445 1344 : ctx: &RequestContext,
2446 1344 : ) -> anyhow::Result<(UninitializedTimeline, RequestContext)> {
2447 1344 : anyhow::ensure!(
2448 1344 : self.is_active(),
2449 0 : "Cannot create empty timelines on inactive tenant"
2450 : );
2451 :
2452 : // Protect against concurrent attempts to use this TimelineId
2453 1344 : let create_guard = match self
2454 1344 : .start_creating_timeline(new_timeline_id, CreateTimelineIdempotency::FailWithConflict)
2455 1344 : .await?
2456 : {
2457 1332 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2458 : StartCreatingTimelineResult::Idempotent(_) => {
2459 0 : unreachable!("FailWithConflict implies we get an error instead")
2460 : }
2461 : };
2462 :
2463 1332 : let new_metadata = TimelineMetadata::new(
2464 1332 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2465 1332 : // make it valid, before calling finish_creation()
2466 1332 : Lsn(0),
2467 1332 : None,
2468 1332 : None,
2469 1332 : Lsn(0),
2470 1332 : initdb_lsn,
2471 1332 : initdb_lsn,
2472 1332 : pg_version,
2473 1332 : );
2474 1332 : self.prepare_new_timeline(
2475 1332 : new_timeline_id,
2476 1332 : &new_metadata,
2477 1332 : create_guard,
2478 1332 : initdb_lsn,
2479 1332 : None,
2480 1332 : None,
2481 1332 : ctx,
2482 1332 : )
2483 1332 : .await
2484 1344 : }
2485 :
2486 : /// Helper for unit tests to create an empty timeline.
2487 : ///
2488 : /// The timeline is has state value `Active` but its background loops are not running.
2489 : // This makes the various functions which anyhow::ensure! for Active state work in tests.
2490 : // Our current tests don't need the background loops.
2491 : #[cfg(test)]
2492 1284 : pub async fn create_test_timeline(
2493 1284 : self: &Arc<Self>,
2494 1284 : new_timeline_id: TimelineId,
2495 1284 : initdb_lsn: Lsn,
2496 1284 : pg_version: u32,
2497 1284 : ctx: &RequestContext,
2498 1284 : ) -> anyhow::Result<Arc<Timeline>> {
2499 1284 : let (uninit_tl, ctx) = self
2500 1284 : .create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2501 1284 : .await?;
2502 1284 : let tline = uninit_tl.raw_timeline().expect("we just created it");
2503 1284 : assert_eq!(tline.get_last_record_lsn(), Lsn(0));
2504 :
2505 : // Setup minimum keys required for the timeline to be usable.
2506 1284 : let mut modification = tline.begin_modification(initdb_lsn);
2507 1284 : modification
2508 1284 : .init_empty_test_timeline()
2509 1284 : .context("init_empty_test_timeline")?;
2510 1284 : modification
2511 1284 : .commit(&ctx)
2512 1284 : .await
2513 1284 : .context("commit init_empty_test_timeline modification")?;
2514 :
2515 : // Flush to disk so that uninit_tl's check for valid disk_consistent_lsn passes.
2516 1284 : tline.maybe_spawn_flush_loop();
2517 1284 : tline.freeze_and_flush().await.context("freeze_and_flush")?;
2518 :
2519 : // Make sure the freeze_and_flush reaches remote storage.
2520 1284 : tline.remote_client.wait_completion().await.unwrap();
2521 :
2522 1284 : let tl = uninit_tl.finish_creation().await?;
2523 : // The non-test code would call tl.activate() here.
2524 1284 : tl.set_state(TimelineState::Active);
2525 1284 : Ok(tl)
2526 1284 : }
2527 :
2528 : /// Helper for unit tests to create a timeline with some pre-loaded states.
2529 : #[cfg(test)]
2530 : #[allow(clippy::too_many_arguments)]
2531 288 : pub async fn create_test_timeline_with_layers(
2532 288 : self: &Arc<Self>,
2533 288 : new_timeline_id: TimelineId,
2534 288 : initdb_lsn: Lsn,
2535 288 : pg_version: u32,
2536 288 : ctx: &RequestContext,
2537 288 : in_memory_layer_desc: Vec<timeline::InMemoryLayerTestDesc>,
2538 288 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
2539 288 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
2540 288 : end_lsn: Lsn,
2541 288 : ) -> anyhow::Result<Arc<Timeline>> {
2542 : use checks::check_valid_layermap;
2543 : use itertools::Itertools;
2544 :
2545 288 : let tline = self
2546 288 : .create_test_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2547 288 : .await?;
2548 288 : tline.force_advance_lsn(end_lsn);
2549 852 : for deltas in delta_layer_desc {
2550 564 : tline
2551 564 : .force_create_delta_layer(deltas, Some(initdb_lsn), ctx)
2552 564 : .await?;
2553 : }
2554 696 : for (lsn, images) in image_layer_desc {
2555 408 : tline
2556 408 : .force_create_image_layer(lsn, images, Some(initdb_lsn), ctx)
2557 408 : .await?;
2558 : }
2559 336 : for in_memory in in_memory_layer_desc {
2560 48 : tline
2561 48 : .force_create_in_memory_layer(in_memory, Some(initdb_lsn), ctx)
2562 48 : .await?;
2563 : }
2564 288 : let layer_names = tline
2565 288 : .layers
2566 288 : .read()
2567 288 : .await
2568 288 : .layer_map()
2569 288 : .unwrap()
2570 288 : .iter_historic_layers()
2571 1260 : .map(|layer| layer.layer_name())
2572 288 : .collect_vec();
2573 288 : if let Some(err) = check_valid_layermap(&layer_names) {
2574 0 : bail!("invalid layermap: {err}");
2575 288 : }
2576 288 : Ok(tline)
2577 288 : }
2578 :
2579 : /// Create a new timeline.
2580 : ///
2581 : /// Returns the new timeline ID and reference to its Timeline object.
2582 : ///
2583 : /// If the caller specified the timeline ID to use (`new_timeline_id`), and timeline with
2584 : /// the same timeline ID already exists, returns CreateTimelineError::AlreadyExists.
2585 : #[allow(clippy::too_many_arguments)]
2586 0 : pub(crate) async fn create_timeline(
2587 0 : self: &Arc<TenantShard>,
2588 0 : params: CreateTimelineParams,
2589 0 : broker_client: storage_broker::BrokerClientChannel,
2590 0 : ctx: &RequestContext,
2591 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
2592 0 : if !self.is_active() {
2593 0 : if matches!(self.current_state(), TenantState::Stopping { .. }) {
2594 0 : return Err(CreateTimelineError::ShuttingDown);
2595 : } else {
2596 0 : return Err(CreateTimelineError::Other(anyhow::anyhow!(
2597 0 : "Cannot create timelines on inactive tenant"
2598 0 : )));
2599 : }
2600 0 : }
2601 :
2602 0 : let _gate = self
2603 0 : .gate
2604 0 : .enter()
2605 0 : .map_err(|_| CreateTimelineError::ShuttingDown)?;
2606 :
2607 0 : let result: CreateTimelineResult = match params {
2608 : CreateTimelineParams::Bootstrap(CreateTimelineParamsBootstrap {
2609 0 : new_timeline_id,
2610 0 : existing_initdb_timeline_id,
2611 0 : pg_version,
2612 0 : }) => {
2613 0 : self.bootstrap_timeline(
2614 0 : new_timeline_id,
2615 0 : pg_version,
2616 0 : existing_initdb_timeline_id,
2617 0 : ctx,
2618 0 : )
2619 0 : .await?
2620 : }
2621 : CreateTimelineParams::Branch(CreateTimelineParamsBranch {
2622 0 : new_timeline_id,
2623 0 : ancestor_timeline_id,
2624 0 : mut ancestor_start_lsn,
2625 : }) => {
2626 0 : let ancestor_timeline = self
2627 0 : .get_timeline(ancestor_timeline_id, false)
2628 0 : .context("Cannot branch off the timeline that's not present in pageserver")?;
2629 :
2630 : // instead of waiting around, just deny the request because ancestor is not yet
2631 : // ready for other purposes either.
2632 0 : if !ancestor_timeline.is_active() {
2633 0 : return Err(CreateTimelineError::AncestorNotActive);
2634 0 : }
2635 0 :
2636 0 : if ancestor_timeline.is_archived() == Some(true) {
2637 0 : info!("tried to branch archived timeline");
2638 0 : return Err(CreateTimelineError::AncestorArchived);
2639 0 : }
2640 :
2641 0 : if let Some(lsn) = ancestor_start_lsn.as_mut() {
2642 0 : *lsn = lsn.align();
2643 0 :
2644 0 : let ancestor_ancestor_lsn = ancestor_timeline.get_ancestor_lsn();
2645 0 : if ancestor_ancestor_lsn > *lsn {
2646 : // can we safely just branch from the ancestor instead?
2647 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
2648 0 : "invalid start lsn {} for ancestor timeline {}: less than timeline ancestor lsn {}",
2649 0 : lsn,
2650 0 : ancestor_timeline_id,
2651 0 : ancestor_ancestor_lsn,
2652 0 : )));
2653 0 : }
2654 0 :
2655 0 : // Wait for the WAL to arrive and be processed on the parent branch up
2656 0 : // to the requested branch point. The repository code itself doesn't
2657 0 : // require it, but if we start to receive WAL on the new timeline,
2658 0 : // decoding the new WAL might need to look up previous pages, relation
2659 0 : // sizes etc. and that would get confused if the previous page versions
2660 0 : // are not in the repository yet.
2661 0 : ancestor_timeline
2662 0 : .wait_lsn(
2663 0 : *lsn,
2664 0 : timeline::WaitLsnWaiter::Tenant,
2665 0 : timeline::WaitLsnTimeout::Default,
2666 0 : ctx,
2667 0 : )
2668 0 : .await
2669 0 : .map_err(|e| match e {
2670 0 : e @ (WaitLsnError::Timeout(_) | WaitLsnError::BadState { .. }) => {
2671 0 : CreateTimelineError::AncestorLsn(anyhow::anyhow!(e))
2672 : }
2673 0 : WaitLsnError::Shutdown => CreateTimelineError::ShuttingDown,
2674 0 : })?;
2675 0 : }
2676 :
2677 0 : self.branch_timeline(&ancestor_timeline, new_timeline_id, ancestor_start_lsn, ctx)
2678 0 : .await?
2679 : }
2680 0 : CreateTimelineParams::ImportPgdata(params) => {
2681 0 : self.create_timeline_import_pgdata(
2682 0 : params,
2683 0 : ActivateTimelineArgs::Yes {
2684 0 : broker_client: broker_client.clone(),
2685 0 : },
2686 0 : ctx,
2687 0 : )
2688 0 : .await?
2689 : }
2690 : };
2691 :
2692 : // At this point we have dropped our guard on [`Self::timelines_creating`], and
2693 : // the timeline is visible in [`Self::timelines`], but it is _not_ durable yet. We must
2694 : // not send a success to the caller until it is. The same applies to idempotent retries.
2695 : //
2696 : // TODO: the timeline is already visible in [`Self::timelines`]; a caller could incorrectly
2697 : // assume that, because they can see the timeline via API, that the creation is done and
2698 : // that it is durable. Ideally, we would keep the timeline hidden (in [`Self::timelines_creating`])
2699 : // until it is durable, e.g., by extending the time we hold the creation guard. This also
2700 : // interacts with UninitializedTimeline and is generally a bit tricky.
2701 : //
2702 : // To re-emphasize: the only correct way to create a timeline is to repeat calling the
2703 : // creation API until it returns success. Only then is durability guaranteed.
2704 0 : info!(creation_result=%result.discriminant(), "waiting for timeline to be durable");
2705 0 : result
2706 0 : .timeline()
2707 0 : .remote_client
2708 0 : .wait_completion()
2709 0 : .await
2710 0 : .map_err(|e| match e {
2711 : WaitCompletionError::NotInitialized(
2712 0 : e, // If the queue is already stopped, it's a shutdown error.
2713 0 : ) if e.is_stopping() => CreateTimelineError::ShuttingDown,
2714 : WaitCompletionError::NotInitialized(_) => {
2715 : // This is a bug: we should never try to wait for uploads before initializing the timeline
2716 0 : debug_assert!(false);
2717 0 : CreateTimelineError::Other(anyhow::anyhow!("timeline not initialized"))
2718 : }
2719 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2720 0 : CreateTimelineError::ShuttingDown
2721 : }
2722 0 : })?;
2723 :
2724 : // The creating task is responsible for activating the timeline.
2725 : // We do this after `wait_completion()` so that we don't spin up tasks that start
2726 : // doing stuff before the IndexPart is durable in S3, which is done by the previous section.
2727 0 : let activated_timeline = match result {
2728 0 : CreateTimelineResult::Created(timeline) => {
2729 0 : timeline.activate(
2730 0 : self.clone(),
2731 0 : broker_client,
2732 0 : None,
2733 0 : &ctx.with_scope_timeline(&timeline),
2734 0 : );
2735 0 : timeline
2736 : }
2737 0 : CreateTimelineResult::Idempotent(timeline) => {
2738 0 : info!(
2739 0 : "request was deemed idempotent, activation will be done by the creating task"
2740 : );
2741 0 : timeline
2742 : }
2743 0 : CreateTimelineResult::ImportSpawned(timeline) => {
2744 0 : info!(
2745 0 : "import task spawned, timeline will become visible and activated once the import is done"
2746 : );
2747 0 : timeline
2748 : }
2749 : };
2750 :
2751 0 : Ok(activated_timeline)
2752 0 : }
2753 :
2754 : /// The returned [`Arc<Timeline>`] is NOT in the [`TenantShard::timelines`] map until the import
2755 : /// completes in the background. A DIFFERENT [`Arc<Timeline>`] will be inserted into the
2756 : /// [`TenantShard::timelines`] map when the import completes.
2757 : /// We only return an [`Arc<Timeline>`] here so the API handler can create a [`pageserver_api::models::TimelineInfo`]
2758 : /// for the response.
2759 0 : async fn create_timeline_import_pgdata(
2760 0 : self: &Arc<Self>,
2761 0 : params: CreateTimelineParamsImportPgdata,
2762 0 : activate: ActivateTimelineArgs,
2763 0 : ctx: &RequestContext,
2764 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
2765 0 : let CreateTimelineParamsImportPgdata {
2766 0 : new_timeline_id,
2767 0 : location,
2768 0 : idempotency_key,
2769 0 : } = params;
2770 0 :
2771 0 : let started_at = chrono::Utc::now().naive_utc();
2772 :
2773 : //
2774 : // There's probably a simpler way to upload an index part, but, remote_timeline_client
2775 : // is the canonical way we do it.
2776 : // - create an empty timeline in-memory
2777 : // - use its remote_timeline_client to do the upload
2778 : // - dispose of the uninit timeline
2779 : // - keep the creation guard alive
2780 :
2781 0 : let timeline_create_guard = match self
2782 0 : .start_creating_timeline(
2783 0 : new_timeline_id,
2784 0 : CreateTimelineIdempotency::ImportPgdata(CreatingTimelineIdempotencyImportPgdata {
2785 0 : idempotency_key: idempotency_key.clone(),
2786 0 : }),
2787 0 : )
2788 0 : .await?
2789 : {
2790 0 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2791 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
2792 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
2793 : }
2794 : };
2795 :
2796 0 : let (mut uninit_timeline, timeline_ctx) = {
2797 0 : let this = &self;
2798 0 : let initdb_lsn = Lsn(0);
2799 0 : async move {
2800 0 : let new_metadata = TimelineMetadata::new(
2801 0 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2802 0 : // make it valid, before calling finish_creation()
2803 0 : Lsn(0),
2804 0 : None,
2805 0 : None,
2806 0 : Lsn(0),
2807 0 : initdb_lsn,
2808 0 : initdb_lsn,
2809 0 : 15,
2810 0 : );
2811 0 : this.prepare_new_timeline(
2812 0 : new_timeline_id,
2813 0 : &new_metadata,
2814 0 : timeline_create_guard,
2815 0 : initdb_lsn,
2816 0 : None,
2817 0 : None,
2818 0 : ctx,
2819 0 : )
2820 0 : .await
2821 0 : }
2822 0 : }
2823 0 : .await?;
2824 :
2825 0 : let in_progress = import_pgdata::index_part_format::InProgress {
2826 0 : idempotency_key,
2827 0 : location,
2828 0 : started_at,
2829 0 : };
2830 0 : let index_part = import_pgdata::index_part_format::Root::V1(
2831 0 : import_pgdata::index_part_format::V1::InProgress(in_progress),
2832 0 : );
2833 0 : uninit_timeline
2834 0 : .raw_timeline()
2835 0 : .unwrap()
2836 0 : .remote_client
2837 0 : .schedule_index_upload_for_import_pgdata_state_update(Some(index_part.clone()))?;
2838 :
2839 : // wait_completion happens in caller
2840 :
2841 0 : let (timeline, timeline_create_guard) = uninit_timeline.finish_creation_myself();
2842 0 :
2843 0 : tokio::spawn(self.clone().create_timeline_import_pgdata_task(
2844 0 : timeline.clone(),
2845 0 : index_part,
2846 0 : activate,
2847 0 : timeline_create_guard,
2848 0 : timeline_ctx.detached_child(TaskKind::ImportPgdata, DownloadBehavior::Warn),
2849 0 : ));
2850 0 :
2851 0 : // NB: the timeline doesn't exist in self.timelines at this point
2852 0 : Ok(CreateTimelineResult::ImportSpawned(timeline))
2853 0 : }
2854 :
2855 : #[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))]
2856 : async fn create_timeline_import_pgdata_task(
2857 : self: Arc<TenantShard>,
2858 : timeline: Arc<Timeline>,
2859 : index_part: import_pgdata::index_part_format::Root,
2860 : activate: ActivateTimelineArgs,
2861 : timeline_create_guard: TimelineCreateGuard,
2862 : ctx: RequestContext,
2863 : ) {
2864 : debug_assert_current_span_has_tenant_and_timeline_id();
2865 : info!("starting");
2866 : scopeguard::defer! {info!("exiting")};
2867 :
2868 : let res = self
2869 : .create_timeline_import_pgdata_task_impl(
2870 : timeline,
2871 : index_part,
2872 : activate,
2873 : timeline_create_guard,
2874 : ctx,
2875 : )
2876 : .await;
2877 : if let Err(err) = &res {
2878 : error!(?err, "task failed");
2879 : // TODO sleep & retry, sensitive to tenant shutdown
2880 : // TODO: allow timeline deletion requests => should cancel the task
2881 : }
2882 : }
2883 :
2884 0 : async fn create_timeline_import_pgdata_task_impl(
2885 0 : self: Arc<TenantShard>,
2886 0 : timeline: Arc<Timeline>,
2887 0 : index_part: import_pgdata::index_part_format::Root,
2888 0 : activate: ActivateTimelineArgs,
2889 0 : timeline_create_guard: TimelineCreateGuard,
2890 0 : ctx: RequestContext,
2891 0 : ) -> Result<(), anyhow::Error> {
2892 0 : info!("importing pgdata");
2893 0 : import_pgdata::doit(&timeline, index_part, &ctx, self.cancel.clone())
2894 0 : .await
2895 0 : .context("import")?;
2896 0 : info!("import done");
2897 :
2898 : //
2899 : // Reload timeline from remote.
2900 : // This proves that the remote state is attachable, and it reuses the code.
2901 : //
2902 : // TODO: think about whether this is safe to do with concurrent TenantShard::shutdown.
2903 : // timeline_create_guard hols the tenant gate open, so, shutdown cannot _complete_ until we exit.
2904 : // But our activate() call might launch new background tasks after TenantShard::shutdown
2905 : // already went past shutting down the TenantShard::timelines, which this timeline here is no part of.
2906 : // I think the same problem exists with the bootstrap & branch mgmt API tasks (tenant shutting
2907 : // down while bootstrapping/branching + activating), but, the race condition is much more likely
2908 : // to manifest because of the long runtime of this import task.
2909 :
2910 : // in theory this shouldn't even .await anything except for coop yield
2911 0 : info!("shutting down timeline");
2912 0 : timeline.shutdown(ShutdownMode::Hard).await;
2913 0 : info!("timeline shut down, reloading from remote");
2914 : // TODO: we can't do the following check because create_timeline_import_pgdata must return an Arc<Timeline>
2915 : // let Some(timeline) = Arc::into_inner(timeline) else {
2916 : // anyhow::bail!("implementation error: timeline that we shut down was still referenced from somewhere");
2917 : // };
2918 0 : let timeline_id = timeline.timeline_id;
2919 0 :
2920 0 : // load from object storage like TenantShard::attach does
2921 0 : let resources = self.build_timeline_resources(timeline_id);
2922 0 : let index_part = resources
2923 0 : .remote_client
2924 0 : .download_index_file(&self.cancel)
2925 0 : .await?;
2926 0 : let index_part = match index_part {
2927 : MaybeDeletedIndexPart::Deleted(_) => {
2928 : // likely concurrent delete call, cplane should prevent this
2929 0 : anyhow::bail!(
2930 0 : "index part says deleted but we are not done creating yet, this should not happen but"
2931 0 : )
2932 : }
2933 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
2934 0 : };
2935 0 : let metadata = index_part.metadata.clone();
2936 0 : self
2937 0 : .load_remote_timeline(timeline_id, index_part, metadata, None, resources, LoadTimelineCause::ImportPgdata{
2938 0 : create_guard: timeline_create_guard, activate, }, &ctx)
2939 0 : .await?
2940 0 : .ready_to_activate()
2941 0 : .context("implementation error: reloaded timeline still needs import after import reported success")?;
2942 :
2943 0 : anyhow::Ok(())
2944 0 : }
2945 :
2946 0 : pub(crate) async fn delete_timeline(
2947 0 : self: Arc<Self>,
2948 0 : timeline_id: TimelineId,
2949 0 : ) -> Result<(), DeleteTimelineError> {
2950 0 : DeleteTimelineFlow::run(&self, timeline_id).await?;
2951 :
2952 0 : Ok(())
2953 0 : }
2954 :
2955 : /// perform one garbage collection iteration, removing old data files from disk.
2956 : /// this function is periodically called by gc task.
2957 : /// also it can be explicitly requested through page server api 'do_gc' command.
2958 : ///
2959 : /// `target_timeline_id` specifies the timeline to GC, or None for all.
2960 : ///
2961 : /// The `horizon` an `pitr` parameters determine how much WAL history needs to be retained.
2962 : /// Also known as the retention period, or the GC cutoff point. `horizon` specifies
2963 : /// the amount of history, as LSN difference from current latest LSN on each timeline.
2964 : /// `pitr` specifies the same as a time difference from the current time. The effective
2965 : /// GC cutoff point is determined conservatively by either `horizon` and `pitr`, whichever
2966 : /// requires more history to be retained.
2967 : //
2968 4524 : pub(crate) async fn gc_iteration(
2969 4524 : &self,
2970 4524 : target_timeline_id: Option<TimelineId>,
2971 4524 : horizon: u64,
2972 4524 : pitr: Duration,
2973 4524 : cancel: &CancellationToken,
2974 4524 : ctx: &RequestContext,
2975 4524 : ) -> Result<GcResult, GcError> {
2976 4524 : // Don't start doing work during shutdown
2977 4524 : if let TenantState::Stopping { .. } = self.current_state() {
2978 0 : return Ok(GcResult::default());
2979 4524 : }
2980 4524 :
2981 4524 : // there is a global allowed_error for this
2982 4524 : if !self.is_active() {
2983 0 : return Err(GcError::NotActive);
2984 4524 : }
2985 4524 :
2986 4524 : {
2987 4524 : let conf = self.tenant_conf.load();
2988 4524 :
2989 4524 : // If we may not delete layers, then simply skip GC. Even though a tenant
2990 4524 : // in AttachedMulti state could do GC and just enqueue the blocked deletions,
2991 4524 : // the only advantage to doing it is to perhaps shrink the LayerMap metadata
2992 4524 : // a bit sooner than we would achieve by waiting for AttachedSingle status.
2993 4524 : if !conf.location.may_delete_layers_hint() {
2994 0 : info!("Skipping GC in location state {:?}", conf.location);
2995 0 : return Ok(GcResult::default());
2996 4524 : }
2997 4524 :
2998 4524 : if conf.is_gc_blocked_by_lsn_lease_deadline() {
2999 4500 : info!("Skipping GC because lsn lease deadline is not reached");
3000 4500 : return Ok(GcResult::default());
3001 24 : }
3002 : }
3003 :
3004 24 : let _guard = match self.gc_block.start().await {
3005 24 : Ok(guard) => guard,
3006 0 : Err(reasons) => {
3007 0 : info!("Skipping GC: {reasons}");
3008 0 : return Ok(GcResult::default());
3009 : }
3010 : };
3011 :
3012 24 : self.gc_iteration_internal(target_timeline_id, horizon, pitr, cancel, ctx)
3013 24 : .await
3014 4524 : }
3015 :
3016 : /// Performs one compaction iteration. Called periodically from the compaction loop. Returns
3017 : /// whether another compaction is needed, if we still have pending work or if we yield for
3018 : /// immediate L0 compaction.
3019 : ///
3020 : /// Compaction can also be explicitly requested for a timeline via the HTTP API.
3021 0 : async fn compaction_iteration(
3022 0 : self: &Arc<Self>,
3023 0 : cancel: &CancellationToken,
3024 0 : ctx: &RequestContext,
3025 0 : ) -> Result<CompactionOutcome, CompactionError> {
3026 0 : // Don't compact inactive tenants.
3027 0 : if !self.is_active() {
3028 0 : return Ok(CompactionOutcome::Skipped);
3029 0 : }
3030 0 :
3031 0 : // Don't compact tenants that can't upload layers. We don't check `may_delete_layers_hint`,
3032 0 : // since we need to compact L0 even in AttachedMulti to bound read amplification.
3033 0 : let location = self.tenant_conf.load().location;
3034 0 : if !location.may_upload_layers_hint() {
3035 0 : info!("skipping compaction in location state {location:?}");
3036 0 : return Ok(CompactionOutcome::Skipped);
3037 0 : }
3038 0 :
3039 0 : // Don't compact if the circuit breaker is tripped.
3040 0 : if self.compaction_circuit_breaker.lock().unwrap().is_broken() {
3041 0 : info!("skipping compaction due to previous failures");
3042 0 : return Ok(CompactionOutcome::Skipped);
3043 0 : }
3044 0 :
3045 0 : // Collect all timelines to compact, along with offload instructions and L0 counts.
3046 0 : let mut compact: Vec<Arc<Timeline>> = Vec::new();
3047 0 : let mut offload: HashSet<TimelineId> = HashSet::new();
3048 0 : let mut l0_counts: HashMap<TimelineId, usize> = HashMap::new();
3049 0 :
3050 0 : {
3051 0 : let offload_enabled = self.get_timeline_offloading_enabled();
3052 0 : let timelines = self.timelines.lock().unwrap();
3053 0 : for (&timeline_id, timeline) in timelines.iter() {
3054 : // Skip inactive timelines.
3055 0 : if !timeline.is_active() {
3056 0 : continue;
3057 0 : }
3058 0 :
3059 0 : // Schedule the timeline for compaction.
3060 0 : compact.push(timeline.clone());
3061 :
3062 : // Schedule the timeline for offloading if eligible.
3063 0 : let can_offload = offload_enabled
3064 0 : && timeline.can_offload().0
3065 0 : && !timelines
3066 0 : .iter()
3067 0 : .any(|(_, tli)| tli.get_ancestor_timeline_id() == Some(timeline_id));
3068 0 : if can_offload {
3069 0 : offload.insert(timeline_id);
3070 0 : }
3071 : }
3072 : } // release timelines lock
3073 :
3074 0 : for timeline in &compact {
3075 : // Collect L0 counts. Can't await while holding lock above.
3076 0 : if let Ok(lm) = timeline.layers.read().await.layer_map() {
3077 0 : l0_counts.insert(timeline.timeline_id, lm.level0_deltas().len());
3078 0 : }
3079 : }
3080 :
3081 : // Pass 1: L0 compaction across all timelines, in order of L0 count. We prioritize this to
3082 : // bound read amplification.
3083 : //
3084 : // TODO: this may spin on one or more ingest-heavy timelines, starving out image/GC
3085 : // compaction and offloading. We leave that as a potential problem to solve later. Consider
3086 : // splitting L0 and image/GC compaction to separate background jobs.
3087 0 : if self.get_compaction_l0_first() {
3088 0 : let compaction_threshold = self.get_compaction_threshold();
3089 0 : let compact_l0 = compact
3090 0 : .iter()
3091 0 : .map(|tli| (tli, l0_counts.get(&tli.timeline_id).copied().unwrap_or(0)))
3092 0 : .filter(|&(_, l0)| l0 >= compaction_threshold)
3093 0 : .sorted_by_key(|&(_, l0)| l0)
3094 0 : .rev()
3095 0 : .map(|(tli, _)| tli.clone())
3096 0 : .collect_vec();
3097 0 :
3098 0 : let mut has_pending_l0 = false;
3099 0 : for timeline in compact_l0 {
3100 0 : let ctx = &ctx.with_scope_timeline(&timeline);
3101 : // NB: don't set CompactFlags::YieldForL0, since this is an L0-only compaction pass.
3102 0 : let outcome = timeline
3103 0 : .compact(cancel, CompactFlags::OnlyL0Compaction.into(), ctx)
3104 0 : .instrument(info_span!("compact_timeline", timeline_id = %timeline.timeline_id))
3105 0 : .await
3106 0 : .inspect_err(|err| self.maybe_trip_compaction_breaker(err))?;
3107 0 : match outcome {
3108 0 : CompactionOutcome::Done => {}
3109 0 : CompactionOutcome::Skipped => {}
3110 0 : CompactionOutcome::Pending => has_pending_l0 = true,
3111 0 : CompactionOutcome::YieldForL0 => has_pending_l0 = true,
3112 : }
3113 : }
3114 0 : if has_pending_l0 {
3115 0 : return Ok(CompactionOutcome::YieldForL0); // do another pass
3116 0 : }
3117 0 : }
3118 :
3119 : // Pass 2: image compaction and timeline offloading. If any timelines have accumulated more
3120 : // L0 layers, they may also be compacted here. Image compaction will yield if there is
3121 : // pending L0 compaction on any tenant timeline.
3122 : //
3123 : // TODO: consider ordering timelines by some priority, e.g. time since last full compaction,
3124 : // amount of L1 delta debt or garbage, offload-eligible timelines first, etc.
3125 0 : let mut has_pending = false;
3126 0 : for timeline in compact {
3127 0 : if !timeline.is_active() {
3128 0 : continue;
3129 0 : }
3130 0 : let ctx = &ctx.with_scope_timeline(&timeline);
3131 0 :
3132 0 : // Yield for L0 if the separate L0 pass is enabled (otherwise there's no point).
3133 0 : let mut flags = EnumSet::default();
3134 0 : if self.get_compaction_l0_first() {
3135 0 : flags |= CompactFlags::YieldForL0;
3136 0 : }
3137 :
3138 0 : let mut outcome = timeline
3139 0 : .compact(cancel, flags, ctx)
3140 0 : .instrument(info_span!("compact_timeline", timeline_id = %timeline.timeline_id))
3141 0 : .await
3142 0 : .inspect_err(|err| self.maybe_trip_compaction_breaker(err))?;
3143 :
3144 : // If we're done compacting, check the scheduled GC compaction queue for more work.
3145 0 : if outcome == CompactionOutcome::Done {
3146 0 : let queue = {
3147 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3148 0 : guard
3149 0 : .entry(timeline.timeline_id)
3150 0 : .or_insert_with(|| Arc::new(GcCompactionQueue::new()))
3151 0 : .clone()
3152 0 : };
3153 0 : outcome = queue
3154 0 : .iteration(cancel, ctx, &self.gc_block, &timeline)
3155 0 : .instrument(
3156 0 : info_span!("gc_compact_timeline", timeline_id = %timeline.timeline_id),
3157 : )
3158 0 : .await?;
3159 0 : }
3160 :
3161 : // If we're done compacting, offload the timeline if requested.
3162 0 : if outcome == CompactionOutcome::Done && offload.contains(&timeline.timeline_id) {
3163 0 : pausable_failpoint!("before-timeline-auto-offload");
3164 0 : offload_timeline(self, &timeline)
3165 0 : .instrument(info_span!("offload_timeline", timeline_id = %timeline.timeline_id))
3166 0 : .await
3167 0 : .or_else(|err| match err {
3168 : // Ignore this, we likely raced with unarchival.
3169 0 : OffloadError::NotArchived => Ok(()),
3170 0 : err => Err(err),
3171 0 : })?;
3172 0 : }
3173 :
3174 0 : match outcome {
3175 0 : CompactionOutcome::Done => {}
3176 0 : CompactionOutcome::Skipped => {}
3177 0 : CompactionOutcome::Pending => has_pending = true,
3178 : // This mostly makes sense when the L0-only pass above is enabled, since there's
3179 : // otherwise no guarantee that we'll start with the timeline that has high L0.
3180 0 : CompactionOutcome::YieldForL0 => return Ok(CompactionOutcome::YieldForL0),
3181 : }
3182 : }
3183 :
3184 : // Success! Untrip the breaker if necessary.
3185 0 : self.compaction_circuit_breaker
3186 0 : .lock()
3187 0 : .unwrap()
3188 0 : .success(&CIRCUIT_BREAKERS_UNBROKEN);
3189 0 :
3190 0 : match has_pending {
3191 0 : true => Ok(CompactionOutcome::Pending),
3192 0 : false => Ok(CompactionOutcome::Done),
3193 : }
3194 0 : }
3195 :
3196 : /// Trips the compaction circuit breaker if appropriate.
3197 0 : pub(crate) fn maybe_trip_compaction_breaker(&self, err: &CompactionError) {
3198 0 : match err {
3199 0 : err if err.is_cancel() => {}
3200 0 : CompactionError::ShuttingDown => (),
3201 : // Offload failures don't trip the circuit breaker, since they're cheap to retry and
3202 : // shouldn't block compaction.
3203 0 : CompactionError::Offload(_) => {}
3204 0 : CompactionError::CollectKeySpaceError(err) => {
3205 0 : // CollectKeySpaceError::Cancelled and PageRead::Cancelled are handled in `err.is_cancel` branch.
3206 0 : self.compaction_circuit_breaker
3207 0 : .lock()
3208 0 : .unwrap()
3209 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, err);
3210 0 : }
3211 0 : CompactionError::Other(err) => {
3212 0 : self.compaction_circuit_breaker
3213 0 : .lock()
3214 0 : .unwrap()
3215 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, err);
3216 0 : }
3217 0 : CompactionError::AlreadyRunning(_) => {}
3218 : }
3219 0 : }
3220 :
3221 : /// Cancel scheduled compaction tasks
3222 0 : pub(crate) fn cancel_scheduled_compaction(&self, timeline_id: TimelineId) {
3223 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3224 0 : if let Some(q) = guard.get_mut(&timeline_id) {
3225 0 : q.cancel_scheduled();
3226 0 : }
3227 0 : }
3228 :
3229 0 : pub(crate) fn get_scheduled_compaction_tasks(
3230 0 : &self,
3231 0 : timeline_id: TimelineId,
3232 0 : ) -> Vec<CompactInfoResponse> {
3233 0 : let res = {
3234 0 : let guard = self.scheduled_compaction_tasks.lock().unwrap();
3235 0 : guard.get(&timeline_id).map(|q| q.remaining_jobs())
3236 : };
3237 0 : let Some((running, remaining)) = res else {
3238 0 : return Vec::new();
3239 : };
3240 0 : let mut result = Vec::new();
3241 0 : if let Some((id, running)) = running {
3242 0 : result.extend(running.into_compact_info_resp(id, true));
3243 0 : }
3244 0 : for (id, job) in remaining {
3245 0 : result.extend(job.into_compact_info_resp(id, false));
3246 0 : }
3247 0 : result
3248 0 : }
3249 :
3250 : /// Schedule a compaction task for a timeline.
3251 0 : pub(crate) async fn schedule_compaction(
3252 0 : &self,
3253 0 : timeline_id: TimelineId,
3254 0 : options: CompactOptions,
3255 0 : ) -> anyhow::Result<tokio::sync::oneshot::Receiver<()>> {
3256 0 : let (tx, rx) = tokio::sync::oneshot::channel();
3257 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3258 0 : let q = guard
3259 0 : .entry(timeline_id)
3260 0 : .or_insert_with(|| Arc::new(GcCompactionQueue::new()));
3261 0 : q.schedule_manual_compaction(options, Some(tx));
3262 0 : Ok(rx)
3263 0 : }
3264 :
3265 : /// Performs periodic housekeeping, via the tenant housekeeping background task.
3266 0 : async fn housekeeping(&self) {
3267 0 : // Call through to all timelines to freeze ephemeral layers as needed. This usually happens
3268 0 : // during ingest, but we don't want idle timelines to hold open layers for too long.
3269 0 : //
3270 0 : // We don't do this if the tenant can't upload layers (i.e. it's in stale attachment mode).
3271 0 : // We don't run compaction in this case either, and don't want to keep flushing tiny L0
3272 0 : // layers that won't be compacted down.
3273 0 : if self.tenant_conf.load().location.may_upload_layers_hint() {
3274 0 : let timelines = self
3275 0 : .timelines
3276 0 : .lock()
3277 0 : .unwrap()
3278 0 : .values()
3279 0 : .filter(|tli| tli.is_active())
3280 0 : .cloned()
3281 0 : .collect_vec();
3282 :
3283 0 : for timeline in timelines {
3284 0 : timeline.maybe_freeze_ephemeral_layer().await;
3285 : }
3286 0 : }
3287 :
3288 : // Shut down walredo if idle.
3289 : const WALREDO_IDLE_TIMEOUT: Duration = Duration::from_secs(180);
3290 0 : if let Some(ref walredo_mgr) = self.walredo_mgr {
3291 0 : walredo_mgr.maybe_quiesce(WALREDO_IDLE_TIMEOUT);
3292 0 : }
3293 0 : }
3294 :
3295 0 : pub fn timeline_has_no_attached_children(&self, timeline_id: TimelineId) -> bool {
3296 0 : let timelines = self.timelines.lock().unwrap();
3297 0 : !timelines
3298 0 : .iter()
3299 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(timeline_id))
3300 0 : }
3301 :
3302 10488 : pub fn current_state(&self) -> TenantState {
3303 10488 : self.state.borrow().clone()
3304 10488 : }
3305 :
3306 5916 : pub fn is_active(&self) -> bool {
3307 5916 : self.current_state() == TenantState::Active
3308 5916 : }
3309 :
3310 0 : pub fn generation(&self) -> Generation {
3311 0 : self.generation
3312 0 : }
3313 :
3314 0 : pub(crate) fn wal_redo_manager_status(&self) -> Option<WalRedoManagerStatus> {
3315 0 : self.walredo_mgr.as_ref().and_then(|mgr| mgr.status())
3316 0 : }
3317 :
3318 : /// Changes tenant status to active, unless shutdown was already requested.
3319 : ///
3320 : /// `background_jobs_can_start` is an optional barrier set to a value during pageserver startup
3321 : /// to delay background jobs. Background jobs can be started right away when None is given.
3322 0 : fn activate(
3323 0 : self: &Arc<Self>,
3324 0 : broker_client: BrokerClientChannel,
3325 0 : background_jobs_can_start: Option<&completion::Barrier>,
3326 0 : ctx: &RequestContext,
3327 0 : ) {
3328 0 : span::debug_assert_current_span_has_tenant_id();
3329 0 :
3330 0 : let mut activating = false;
3331 0 : self.state.send_modify(|current_state| {
3332 : use pageserver_api::models::ActivatingFrom;
3333 0 : match &*current_state {
3334 : TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
3335 0 : panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {state:?}", state = current_state);
3336 : }
3337 0 : TenantState::Attaching => {
3338 0 : *current_state = TenantState::Activating(ActivatingFrom::Attaching);
3339 0 : }
3340 0 : }
3341 0 : debug!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), "Activating tenant");
3342 0 : activating = true;
3343 0 : // Continue outside the closure. We need to grab timelines.lock()
3344 0 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
3345 0 : });
3346 0 :
3347 0 : if activating {
3348 0 : let timelines_accessor = self.timelines.lock().unwrap();
3349 0 : let timelines_offloaded_accessor = self.timelines_offloaded.lock().unwrap();
3350 0 : let timelines_to_activate = timelines_accessor
3351 0 : .values()
3352 0 : .filter(|timeline| !(timeline.is_broken() || timeline.is_stopping()));
3353 0 :
3354 0 : // Before activation, populate each Timeline's GcInfo with information about its children
3355 0 : self.initialize_gc_info(&timelines_accessor, &timelines_offloaded_accessor, None);
3356 0 :
3357 0 : // Spawn gc and compaction loops. The loops will shut themselves
3358 0 : // down when they notice that the tenant is inactive.
3359 0 : tasks::start_background_loops(self, background_jobs_can_start);
3360 0 :
3361 0 : let mut activated_timelines = 0;
3362 :
3363 0 : for timeline in timelines_to_activate {
3364 0 : timeline.activate(
3365 0 : self.clone(),
3366 0 : broker_client.clone(),
3367 0 : background_jobs_can_start,
3368 0 : &ctx.with_scope_timeline(timeline),
3369 0 : );
3370 0 : activated_timelines += 1;
3371 0 : }
3372 :
3373 0 : self.state.send_modify(move |current_state| {
3374 0 : assert!(
3375 0 : matches!(current_state, TenantState::Activating(_)),
3376 0 : "set_stopping and set_broken wait for us to leave Activating state",
3377 : );
3378 0 : *current_state = TenantState::Active;
3379 0 :
3380 0 : let elapsed = self.constructed_at.elapsed();
3381 0 : let total_timelines = timelines_accessor.len();
3382 0 :
3383 0 : // log a lot of stuff, because some tenants sometimes suffer from user-visible
3384 0 : // times to activate. see https://github.com/neondatabase/neon/issues/4025
3385 0 : info!(
3386 0 : since_creation_millis = elapsed.as_millis(),
3387 0 : tenant_id = %self.tenant_shard_id.tenant_id,
3388 0 : shard_id = %self.tenant_shard_id.shard_slug(),
3389 0 : activated_timelines,
3390 0 : total_timelines,
3391 0 : post_state = <&'static str>::from(&*current_state),
3392 0 : "activation attempt finished"
3393 : );
3394 :
3395 0 : TENANT.activation.observe(elapsed.as_secs_f64());
3396 0 : });
3397 0 : }
3398 0 : }
3399 :
3400 : /// Shutdown the tenant and join all of the spawned tasks.
3401 : ///
3402 : /// The method caters for all use-cases:
3403 : /// - pageserver shutdown (freeze_and_flush == true)
3404 : /// - detach + ignore (freeze_and_flush == false)
3405 : ///
3406 : /// This will attempt to shutdown even if tenant is broken.
3407 : ///
3408 : /// `shutdown_progress` is a [`completion::Barrier`] for the shutdown initiated by this call.
3409 : /// If the tenant is already shutting down, we return a clone of the first shutdown call's
3410 : /// `Barrier` as an `Err`. This not-first caller can use the returned barrier to join with
3411 : /// the ongoing shutdown.
3412 36 : async fn shutdown(
3413 36 : &self,
3414 36 : shutdown_progress: completion::Barrier,
3415 36 : shutdown_mode: timeline::ShutdownMode,
3416 36 : ) -> Result<(), completion::Barrier> {
3417 36 : span::debug_assert_current_span_has_tenant_id();
3418 :
3419 : // Set tenant (and its timlines) to Stoppping state.
3420 : //
3421 : // Since we can only transition into Stopping state after activation is complete,
3422 : // run it in a JoinSet so all tenants have a chance to stop before we get SIGKILLed.
3423 : //
3424 : // Transitioning tenants to Stopping state has a couple of non-obvious side effects:
3425 : // 1. Lock out any new requests to the tenants.
3426 : // 2. Signal cancellation to WAL receivers (we wait on it below).
3427 : // 3. Signal cancellation for other tenant background loops.
3428 : // 4. ???
3429 : //
3430 : // The waiting for the cancellation is not done uniformly.
3431 : // We certainly wait for WAL receivers to shut down.
3432 : // That is necessary so that no new data comes in before the freeze_and_flush.
3433 : // But the tenant background loops are joined-on in our caller.
3434 : // It's mesed up.
3435 : // we just ignore the failure to stop
3436 :
3437 : // If we're still attaching, fire the cancellation token early to drop out: this
3438 : // will prevent us flushing, but ensures timely shutdown if some I/O during attach
3439 : // is very slow.
3440 36 : let shutdown_mode = if matches!(self.current_state(), TenantState::Attaching) {
3441 0 : self.cancel.cancel();
3442 0 :
3443 0 : // Having fired our cancellation token, do not try and flush timelines: their cancellation tokens
3444 0 : // are children of ours, so their flush loops will have shut down already
3445 0 : timeline::ShutdownMode::Hard
3446 : } else {
3447 36 : shutdown_mode
3448 : };
3449 :
3450 36 : match self.set_stopping(shutdown_progress).await {
3451 36 : Ok(()) => {}
3452 0 : Err(SetStoppingError::Broken) => {
3453 0 : // assume that this is acceptable
3454 0 : }
3455 0 : Err(SetStoppingError::AlreadyStopping(other)) => {
3456 0 : // give caller the option to wait for this this shutdown
3457 0 : info!("Tenant::shutdown: AlreadyStopping");
3458 0 : return Err(other);
3459 : }
3460 : };
3461 :
3462 36 : let mut js = tokio::task::JoinSet::new();
3463 36 : {
3464 36 : let timelines = self.timelines.lock().unwrap();
3465 36 : timelines.values().for_each(|timeline| {
3466 36 : let timeline = Arc::clone(timeline);
3467 36 : let timeline_id = timeline.timeline_id;
3468 36 : let span = tracing::info_span!("timeline_shutdown", %timeline_id, ?shutdown_mode);
3469 36 : js.spawn(async move { timeline.shutdown(shutdown_mode).instrument(span).await });
3470 36 : });
3471 36 : }
3472 36 : {
3473 36 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
3474 36 : timelines_offloaded.values().for_each(|timeline| {
3475 0 : timeline.defuse_for_tenant_drop();
3476 36 : });
3477 36 : }
3478 36 : // test_long_timeline_create_then_tenant_delete is leaning on this message
3479 36 : tracing::info!("Waiting for timelines...");
3480 72 : while let Some(res) = js.join_next().await {
3481 0 : match res {
3482 36 : 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 36 : 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 36 : }
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 36 : tracing::debug!("Cancelling CancellationToken");
3504 36 : self.cancel.cancel();
3505 36 :
3506 36 : // shutdown all tenant and timeline tasks: gc, compaction, page service
3507 36 : // No new tasks will be started for this tenant because it's in `Stopping` state.
3508 36 : //
3509 36 : // this will additionally shutdown and await all timeline tasks.
3510 36 : tracing::debug!("Waiting for tasks...");
3511 36 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), None).await;
3512 :
3513 36 : if let Some(walredo_mgr) = self.walredo_mgr.as_ref() {
3514 36 : walredo_mgr.shutdown().await;
3515 0 : }
3516 :
3517 : // Wait for any in-flight operations to complete
3518 36 : self.gate.close().await;
3519 :
3520 36 : remove_tenant_metrics(&self.tenant_shard_id);
3521 36 :
3522 36 : Ok(())
3523 36 : }
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 36 : async fn set_stopping(&self, progress: completion::Barrier) -> Result<(), SetStoppingError> {
3531 36 : let mut rx = self.state.subscribe();
3532 36 :
3533 36 : // cannot stop before we're done activating, so wait out until we're done activating
3534 36 : 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 36 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3540 36 : })
3541 36 : .await
3542 36 : .expect("cannot drop self.state while on a &self method");
3543 36 :
3544 36 : // we now know we're done activating, let's see whether this task is the winner to transition into Stopping
3545 36 : let mut err = None;
3546 36 : 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 36 : *current_state = TenantState::Stopping { progress: Some(progress) };
3555 36 : // Continue stopping outside the closure. We need to grab timelines.lock()
3556 36 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
3557 36 : 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 36 : });
3578 36 : match (stopping, err) {
3579 36 : (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 36 : let timelines_accessor = self.timelines.lock().unwrap();
3590 36 : let not_broken_timelines = timelines_accessor
3591 36 : .values()
3592 36 : .filter(|timeline| !timeline.is_broken());
3593 72 : for timeline in not_broken_timelines {
3594 36 : timeline.set_state(TimelineState::Stopping);
3595 36 : }
3596 36 : Ok(())
3597 36 : }
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 1404 : pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
3745 1404 : self.shard_identity.stripe_size
3746 1404 : }
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 0 : for child_shard in child_shards {
3820 0 : tracing::info!(%timeline_id, "Uploading index_part for child {}", child_shard.to_index());
3821 0 : upload_index_part(
3822 0 : &self.remote_storage,
3823 0 : child_shard,
3824 0 : &timeline_id,
3825 0 : self.generation,
3826 0 : &index_part,
3827 0 : &self.cancel,
3828 0 : )
3829 0 : .await?;
3830 : }
3831 : }
3832 :
3833 0 : let tenant_manifest = self.build_tenant_manifest();
3834 0 : for child_shard in child_shards {
3835 0 : tracing::info!(
3836 0 : "Uploading tenant manifest for child {}",
3837 0 : child_shard.to_index()
3838 : );
3839 0 : upload_tenant_manifest(
3840 0 : &self.remote_storage,
3841 0 : child_shard,
3842 0 : self.generation,
3843 0 : &tenant_manifest,
3844 0 : &self.cancel,
3845 0 : )
3846 0 : .await?;
3847 : }
3848 :
3849 0 : Ok(())
3850 0 : }
3851 :
3852 0 : pub(crate) fn get_sizes(&self) -> TopTenantShardItem {
3853 0 : let mut result = TopTenantShardItem {
3854 0 : id: self.tenant_shard_id,
3855 0 : resident_size: 0,
3856 0 : physical_size: 0,
3857 0 : max_logical_size: 0,
3858 0 : max_logical_size_per_shard: 0,
3859 0 : };
3860 :
3861 0 : for timeline in self.timelines.lock().unwrap().values() {
3862 0 : result.resident_size += timeline.metrics.resident_physical_size_gauge.get();
3863 0 :
3864 0 : result.physical_size += timeline
3865 0 : .remote_client
3866 0 : .metrics
3867 0 : .remote_physical_size_gauge
3868 0 : .get();
3869 0 : result.max_logical_size = std::cmp::max(
3870 0 : result.max_logical_size,
3871 0 : timeline.metrics.current_logical_size_gauge.get(),
3872 0 : );
3873 0 : }
3874 :
3875 0 : result.max_logical_size_per_shard = result
3876 0 : .max_logical_size
3877 0 : .div_ceil(self.tenant_shard_id.shard_count.count() as u64);
3878 0 :
3879 0 : result
3880 0 : }
3881 : }
3882 :
3883 : /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
3884 : /// perform a topological sort, so that the parent of each timeline comes
3885 : /// before the children.
3886 : /// E extracts the ancestor from T
3887 : /// This allows for T to be different. It can be TimelineMetadata, can be Timeline itself, etc.
3888 1392 : fn tree_sort_timelines<T, E>(
3889 1392 : timelines: HashMap<TimelineId, T>,
3890 1392 : extractor: E,
3891 1392 : ) -> anyhow::Result<Vec<(TimelineId, T)>>
3892 1392 : where
3893 1392 : E: Fn(&T) -> Option<TimelineId>,
3894 1392 : {
3895 1392 : let mut result = Vec::with_capacity(timelines.len());
3896 1392 :
3897 1392 : let mut now = Vec::with_capacity(timelines.len());
3898 1392 : // (ancestor, children)
3899 1392 : let mut later: HashMap<TimelineId, Vec<(TimelineId, T)>> =
3900 1392 : HashMap::with_capacity(timelines.len());
3901 :
3902 1428 : for (timeline_id, value) in timelines {
3903 36 : if let Some(ancestor_id) = extractor(&value) {
3904 12 : let children = later.entry(ancestor_id).or_default();
3905 12 : children.push((timeline_id, value));
3906 24 : } else {
3907 24 : now.push((timeline_id, value));
3908 24 : }
3909 : }
3910 :
3911 1428 : while let Some((timeline_id, metadata)) = now.pop() {
3912 36 : result.push((timeline_id, metadata));
3913 : // All children of this can be loaded now
3914 36 : if let Some(mut children) = later.remove(&timeline_id) {
3915 12 : now.append(&mut children);
3916 24 : }
3917 : }
3918 :
3919 : // All timelines should be visited now. Unless there were timelines with missing ancestors.
3920 1392 : if !later.is_empty() {
3921 0 : for (missing_id, orphan_ids) in later {
3922 0 : for (orphan_id, _) in orphan_ids {
3923 0 : error!(
3924 0 : "could not load timeline {orphan_id} because its ancestor timeline {missing_id} could not be loaded"
3925 : );
3926 : }
3927 : }
3928 0 : bail!("could not load tenant because some timelines are missing ancestors");
3929 1392 : }
3930 1392 :
3931 1392 : Ok(result)
3932 1392 : }
3933 :
3934 : enum ActivateTimelineArgs {
3935 : Yes {
3936 : broker_client: storage_broker::BrokerClientChannel,
3937 : },
3938 : No,
3939 : }
3940 :
3941 : impl TenantShard {
3942 0 : pub fn tenant_specific_overrides(&self) -> pageserver_api::models::TenantConfig {
3943 0 : self.tenant_conf.load().tenant_conf.clone()
3944 0 : }
3945 :
3946 0 : pub fn effective_config(&self) -> pageserver_api::config::TenantConfigToml {
3947 0 : self.tenant_specific_overrides()
3948 0 : .merge(self.conf.default_tenant_conf.clone())
3949 0 : }
3950 :
3951 0 : pub fn get_checkpoint_distance(&self) -> u64 {
3952 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3953 0 : tenant_conf
3954 0 : .checkpoint_distance
3955 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
3956 0 : }
3957 :
3958 0 : pub fn get_checkpoint_timeout(&self) -> Duration {
3959 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3960 0 : tenant_conf
3961 0 : .checkpoint_timeout
3962 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
3963 0 : }
3964 :
3965 0 : pub fn get_compaction_target_size(&self) -> u64 {
3966 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3967 0 : tenant_conf
3968 0 : .compaction_target_size
3969 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
3970 0 : }
3971 :
3972 0 : pub fn get_compaction_period(&self) -> Duration {
3973 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3974 0 : tenant_conf
3975 0 : .compaction_period
3976 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
3977 0 : }
3978 :
3979 0 : pub fn get_compaction_threshold(&self) -> usize {
3980 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3981 0 : tenant_conf
3982 0 : .compaction_threshold
3983 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
3984 0 : }
3985 :
3986 0 : pub fn get_rel_size_v2_enabled(&self) -> bool {
3987 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3988 0 : tenant_conf
3989 0 : .rel_size_v2_enabled
3990 0 : .unwrap_or(self.conf.default_tenant_conf.rel_size_v2_enabled)
3991 0 : }
3992 :
3993 0 : pub fn get_compaction_upper_limit(&self) -> usize {
3994 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3995 0 : tenant_conf
3996 0 : .compaction_upper_limit
3997 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_upper_limit)
3998 0 : }
3999 :
4000 0 : pub fn get_compaction_l0_first(&self) -> bool {
4001 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4002 0 : tenant_conf
4003 0 : .compaction_l0_first
4004 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_first)
4005 0 : }
4006 :
4007 24 : pub fn get_gc_horizon(&self) -> u64 {
4008 24 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4009 24 : tenant_conf
4010 24 : .gc_horizon
4011 24 : .unwrap_or(self.conf.default_tenant_conf.gc_horizon)
4012 24 : }
4013 :
4014 0 : pub fn get_gc_period(&self) -> Duration {
4015 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4016 0 : tenant_conf
4017 0 : .gc_period
4018 0 : .unwrap_or(self.conf.default_tenant_conf.gc_period)
4019 0 : }
4020 :
4021 0 : pub fn get_image_creation_threshold(&self) -> usize {
4022 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4023 0 : tenant_conf
4024 0 : .image_creation_threshold
4025 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
4026 0 : }
4027 :
4028 24 : pub fn get_pitr_interval(&self) -> Duration {
4029 24 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4030 24 : tenant_conf
4031 24 : .pitr_interval
4032 24 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
4033 24 : }
4034 :
4035 0 : pub fn get_min_resident_size_override(&self) -> Option<u64> {
4036 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4037 0 : tenant_conf
4038 0 : .min_resident_size_override
4039 0 : .or(self.conf.default_tenant_conf.min_resident_size_override)
4040 0 : }
4041 :
4042 0 : pub fn get_heatmap_period(&self) -> Option<Duration> {
4043 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4044 0 : let heatmap_period = tenant_conf
4045 0 : .heatmap_period
4046 0 : .unwrap_or(self.conf.default_tenant_conf.heatmap_period);
4047 0 : if heatmap_period.is_zero() {
4048 0 : None
4049 : } else {
4050 0 : Some(heatmap_period)
4051 : }
4052 0 : }
4053 :
4054 24 : pub fn get_lsn_lease_length(&self) -> Duration {
4055 24 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4056 24 : tenant_conf
4057 24 : .lsn_lease_length
4058 24 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
4059 24 : }
4060 :
4061 0 : pub fn get_timeline_offloading_enabled(&self) -> bool {
4062 0 : if self.conf.timeline_offloading {
4063 0 : return true;
4064 0 : }
4065 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4066 0 : tenant_conf
4067 0 : .timeline_offloading
4068 0 : .unwrap_or(self.conf.default_tenant_conf.timeline_offloading)
4069 0 : }
4070 :
4071 : /// Generate an up-to-date TenantManifest based on the state of this Tenant.
4072 1404 : fn build_tenant_manifest(&self) -> TenantManifest {
4073 1404 : // Collect the offloaded timelines, and sort them for deterministic output.
4074 1404 : let offloaded_timelines = self
4075 1404 : .timelines_offloaded
4076 1404 : .lock()
4077 1404 : .unwrap()
4078 1404 : .values()
4079 1404 : .map(|tli| tli.manifest())
4080 1404 : .sorted_by_key(|m| m.timeline_id)
4081 1404 : .collect_vec();
4082 1404 :
4083 1404 : TenantManifest {
4084 1404 : version: LATEST_TENANT_MANIFEST_VERSION,
4085 1404 : stripe_size: Some(self.get_shard_stripe_size()),
4086 1404 : offloaded_timelines,
4087 1404 : }
4088 1404 : }
4089 :
4090 0 : pub fn update_tenant_config<
4091 0 : F: Fn(
4092 0 : pageserver_api::models::TenantConfig,
4093 0 : ) -> anyhow::Result<pageserver_api::models::TenantConfig>,
4094 0 : >(
4095 0 : &self,
4096 0 : update: F,
4097 0 : ) -> anyhow::Result<pageserver_api::models::TenantConfig> {
4098 0 : // Use read-copy-update in order to avoid overwriting the location config
4099 0 : // state if this races with [`TenantShard::set_new_location_config`]. Note that
4100 0 : // this race is not possible if both request types come from the storage
4101 0 : // controller (as they should!) because an exclusive op lock is required
4102 0 : // on the storage controller side.
4103 0 :
4104 0 : self.tenant_conf
4105 0 : .try_rcu(|attached_conf| -> Result<_, anyhow::Error> {
4106 0 : Ok(Arc::new(AttachedTenantConf {
4107 0 : tenant_conf: update(attached_conf.tenant_conf.clone())?,
4108 0 : location: attached_conf.location,
4109 0 : lsn_lease_deadline: attached_conf.lsn_lease_deadline,
4110 : }))
4111 0 : })?;
4112 :
4113 0 : let updated = self.tenant_conf.load();
4114 0 :
4115 0 : self.tenant_conf_updated(&updated.tenant_conf);
4116 0 : // Don't hold self.timelines.lock() during the notifies.
4117 0 : // There's no risk of deadlock right now, but there could be if we consolidate
4118 0 : // mutexes in struct Timeline in the future.
4119 0 : let timelines = self.list_timelines();
4120 0 : for timeline in timelines {
4121 0 : timeline.tenant_conf_updated(&updated);
4122 0 : }
4123 :
4124 0 : Ok(updated.tenant_conf.clone())
4125 0 : }
4126 :
4127 0 : pub(crate) fn set_new_location_config(&self, new_conf: AttachedTenantConf) {
4128 0 : let new_tenant_conf = new_conf.tenant_conf.clone();
4129 0 :
4130 0 : self.tenant_conf.store(Arc::new(new_conf.clone()));
4131 0 :
4132 0 : self.tenant_conf_updated(&new_tenant_conf);
4133 0 : // Don't hold self.timelines.lock() during the notifies.
4134 0 : // There's no risk of deadlock right now, but there could be if we consolidate
4135 0 : // mutexes in struct Timeline in the future.
4136 0 : let timelines = self.list_timelines();
4137 0 : for timeline in timelines {
4138 0 : timeline.tenant_conf_updated(&new_conf);
4139 0 : }
4140 0 : }
4141 :
4142 1392 : fn get_pagestream_throttle_config(
4143 1392 : psconf: &'static PageServerConf,
4144 1392 : overrides: &pageserver_api::models::TenantConfig,
4145 1392 : ) -> throttle::Config {
4146 1392 : overrides
4147 1392 : .timeline_get_throttle
4148 1392 : .clone()
4149 1392 : .unwrap_or(psconf.default_tenant_conf.timeline_get_throttle.clone())
4150 1392 : }
4151 :
4152 0 : pub(crate) fn tenant_conf_updated(&self, new_conf: &pageserver_api::models::TenantConfig) {
4153 0 : let conf = Self::get_pagestream_throttle_config(self.conf, new_conf);
4154 0 : self.pagestream_throttle.reconfigure(conf)
4155 0 : }
4156 :
4157 : /// Helper function to create a new Timeline struct.
4158 : ///
4159 : /// The returned Timeline is in Loading state. The caller is responsible for
4160 : /// initializing any on-disk state, and for inserting the Timeline to the 'timelines'
4161 : /// map.
4162 : ///
4163 : /// `validate_ancestor == false` is used when a timeline is created for deletion
4164 : /// and we might not have the ancestor present anymore which is fine for to be
4165 : /// deleted timelines.
4166 : #[allow(clippy::too_many_arguments)]
4167 2784 : fn create_timeline_struct(
4168 2784 : &self,
4169 2784 : new_timeline_id: TimelineId,
4170 2784 : new_metadata: &TimelineMetadata,
4171 2784 : previous_heatmap: Option<PreviousHeatmap>,
4172 2784 : ancestor: Option<Arc<Timeline>>,
4173 2784 : resources: TimelineResources,
4174 2784 : cause: CreateTimelineCause,
4175 2784 : create_idempotency: CreateTimelineIdempotency,
4176 2784 : gc_compaction_state: Option<GcCompactionState>,
4177 2784 : rel_size_v2_status: Option<RelSizeMigration>,
4178 2784 : ctx: &RequestContext,
4179 2784 : ) -> anyhow::Result<(Arc<Timeline>, RequestContext)> {
4180 2784 : let state = match cause {
4181 : CreateTimelineCause::Load => {
4182 2784 : let ancestor_id = new_metadata.ancestor_timeline();
4183 2784 : anyhow::ensure!(
4184 2784 : ancestor_id == ancestor.as_ref().map(|t| t.timeline_id),
4185 0 : "Timeline's {new_timeline_id} ancestor {ancestor_id:?} was not found"
4186 : );
4187 2784 : TimelineState::Loading
4188 : }
4189 0 : CreateTimelineCause::Delete => TimelineState::Stopping,
4190 : };
4191 :
4192 2784 : let pg_version = new_metadata.pg_version();
4193 2784 :
4194 2784 : let timeline = Timeline::new(
4195 2784 : self.conf,
4196 2784 : Arc::clone(&self.tenant_conf),
4197 2784 : new_metadata,
4198 2784 : previous_heatmap,
4199 2784 : ancestor,
4200 2784 : new_timeline_id,
4201 2784 : self.tenant_shard_id,
4202 2784 : self.generation,
4203 2784 : self.shard_identity,
4204 2784 : self.walredo_mgr.clone(),
4205 2784 : resources,
4206 2784 : pg_version,
4207 2784 : state,
4208 2784 : self.attach_wal_lag_cooldown.clone(),
4209 2784 : create_idempotency,
4210 2784 : gc_compaction_state,
4211 2784 : rel_size_v2_status,
4212 2784 : self.cancel.child_token(),
4213 2784 : );
4214 2784 :
4215 2784 : let timeline_ctx = RequestContextBuilder::from(ctx)
4216 2784 : .scope(context::Scope::new_timeline(&timeline))
4217 2784 : .detached_child();
4218 2784 :
4219 2784 : Ok((timeline, timeline_ctx))
4220 2784 : }
4221 :
4222 : /// [`TenantShard::shutdown`] must be called before dropping the returned [`TenantShard`] object
4223 : /// to ensure proper cleanup of background tasks and metrics.
4224 : //
4225 : // Allow too_many_arguments because a constructor's argument list naturally grows with the
4226 : // number of attributes in the struct: breaking these out into a builder wouldn't be helpful.
4227 : #[allow(clippy::too_many_arguments)]
4228 1392 : fn new(
4229 1392 : state: TenantState,
4230 1392 : conf: &'static PageServerConf,
4231 1392 : attached_conf: AttachedTenantConf,
4232 1392 : shard_identity: ShardIdentity,
4233 1392 : walredo_mgr: Option<Arc<WalRedoManager>>,
4234 1392 : tenant_shard_id: TenantShardId,
4235 1392 : remote_storage: GenericRemoteStorage,
4236 1392 : deletion_queue_client: DeletionQueueClient,
4237 1392 : l0_flush_global_state: L0FlushGlobalState,
4238 1392 : ) -> TenantShard {
4239 1392 : debug_assert!(
4240 1392 : !attached_conf.location.generation.is_none() || conf.control_plane_api.is_none()
4241 : );
4242 :
4243 1392 : let (state, mut rx) = watch::channel(state);
4244 1392 :
4245 1392 : tokio::spawn(async move {
4246 1387 : // reflect tenant state in metrics:
4247 1387 : // - global per tenant state: TENANT_STATE_METRIC
4248 1387 : // - "set" of broken tenants: BROKEN_TENANTS_SET
4249 1387 : //
4250 1387 : // set of broken tenants should not have zero counts so that it remains accessible for
4251 1387 : // alerting.
4252 1387 :
4253 1387 : let tid = tenant_shard_id.to_string();
4254 1387 : let shard_id = tenant_shard_id.shard_slug().to_string();
4255 1387 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
4256 :
4257 2774 : fn inspect_state(state: &TenantState) -> ([&'static str; 1], bool) {
4258 2774 : ([state.into()], matches!(state, TenantState::Broken { .. }))
4259 2774 : }
4260 :
4261 1387 : let mut tuple = inspect_state(&rx.borrow_and_update());
4262 1387 :
4263 1387 : let is_broken = tuple.1;
4264 1387 : let mut counted_broken = if is_broken {
4265 : // add the id to the set right away, there should not be any updates on the channel
4266 : // after before tenant is removed, if ever
4267 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
4268 0 : true
4269 : } else {
4270 1387 : false
4271 : };
4272 :
4273 : loop {
4274 2774 : let labels = &tuple.0;
4275 2774 : let current = TENANT_STATE_METRIC.with_label_values(labels);
4276 2774 : current.inc();
4277 2774 :
4278 2774 : if rx.changed().await.is_err() {
4279 : // tenant has been dropped
4280 84 : current.dec();
4281 84 : drop(BROKEN_TENANTS_SET.remove_label_values(set_key));
4282 84 : break;
4283 1387 : }
4284 1387 :
4285 1387 : current.dec();
4286 1387 : tuple = inspect_state(&rx.borrow_and_update());
4287 1387 :
4288 1387 : let is_broken = tuple.1;
4289 1387 : if is_broken && !counted_broken {
4290 0 : counted_broken = true;
4291 0 : // insert the tenant_id (back) into the set while avoiding needless counter
4292 0 : // access
4293 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
4294 1387 : }
4295 : }
4296 1392 : });
4297 1392 :
4298 1392 : TenantShard {
4299 1392 : tenant_shard_id,
4300 1392 : shard_identity,
4301 1392 : generation: attached_conf.location.generation,
4302 1392 : conf,
4303 1392 : // using now here is good enough approximation to catch tenants with really long
4304 1392 : // activation times.
4305 1392 : constructed_at: Instant::now(),
4306 1392 : timelines: Mutex::new(HashMap::new()),
4307 1392 : timelines_creating: Mutex::new(HashSet::new()),
4308 1392 : timelines_offloaded: Mutex::new(HashMap::new()),
4309 1392 : remote_tenant_manifest: Default::default(),
4310 1392 : gc_cs: tokio::sync::Mutex::new(()),
4311 1392 : walredo_mgr,
4312 1392 : remote_storage,
4313 1392 : deletion_queue_client,
4314 1392 : state,
4315 1392 : cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
4316 1392 : cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
4317 1392 : eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
4318 1392 : compaction_circuit_breaker: std::sync::Mutex::new(CircuitBreaker::new(
4319 1392 : format!("compaction-{tenant_shard_id}"),
4320 1392 : 5,
4321 1392 : // Compaction can be a very expensive operation, and might leak disk space. It also ought
4322 1392 : // to be infallible, as long as remote storage is available. So if it repeatedly fails,
4323 1392 : // use an extremely long backoff.
4324 1392 : Some(Duration::from_secs(3600 * 24)),
4325 1392 : )),
4326 1392 : l0_compaction_trigger: Arc::new(Notify::new()),
4327 1392 : scheduled_compaction_tasks: Mutex::new(Default::default()),
4328 1392 : activate_now_sem: tokio::sync::Semaphore::new(0),
4329 1392 : attach_wal_lag_cooldown: Arc::new(std::sync::OnceLock::new()),
4330 1392 : cancel: CancellationToken::default(),
4331 1392 : gate: Gate::default(),
4332 1392 : pagestream_throttle: Arc::new(throttle::Throttle::new(
4333 1392 : TenantShard::get_pagestream_throttle_config(conf, &attached_conf.tenant_conf),
4334 1392 : )),
4335 1392 : pagestream_throttle_metrics: Arc::new(
4336 1392 : crate::metrics::tenant_throttling::Pagestream::new(&tenant_shard_id),
4337 1392 : ),
4338 1392 : tenant_conf: Arc::new(ArcSwap::from_pointee(attached_conf)),
4339 1392 : ongoing_timeline_detach: std::sync::Mutex::default(),
4340 1392 : gc_block: Default::default(),
4341 1392 : l0_flush_global_state,
4342 1392 : }
4343 1392 : }
4344 :
4345 : /// Locate and load config
4346 0 : pub(super) fn load_tenant_config(
4347 0 : conf: &'static PageServerConf,
4348 0 : tenant_shard_id: &TenantShardId,
4349 0 : ) -> Result<LocationConf, LoadConfigError> {
4350 0 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4351 0 :
4352 0 : info!("loading tenant configuration from {config_path}");
4353 :
4354 : // load and parse file
4355 0 : let config = fs::read_to_string(&config_path).map_err(|e| {
4356 0 : match e.kind() {
4357 : std::io::ErrorKind::NotFound => {
4358 : // The config should almost always exist for a tenant directory:
4359 : // - When attaching a tenant, the config is the first thing we write
4360 : // - When detaching a tenant, we atomically move the directory to a tmp location
4361 : // before deleting contents.
4362 : //
4363 : // The very rare edge case that can result in a missing config is if we crash during attach
4364 : // between creating directory and writing config. Callers should handle that as if the
4365 : // directory didn't exist.
4366 :
4367 0 : LoadConfigError::NotFound(config_path)
4368 : }
4369 : _ => {
4370 : // No IO errors except NotFound are acceptable here: other kinds of error indicate local storage or permissions issues
4371 : // that we cannot cleanly recover
4372 0 : crate::virtual_file::on_fatal_io_error(&e, "Reading tenant config file")
4373 : }
4374 : }
4375 0 : })?;
4376 :
4377 0 : Ok(toml_edit::de::from_str::<LocationConf>(&config)?)
4378 0 : }
4379 :
4380 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4381 : pub(super) async fn persist_tenant_config(
4382 : conf: &'static PageServerConf,
4383 : tenant_shard_id: &TenantShardId,
4384 : location_conf: &LocationConf,
4385 : ) -> std::io::Result<()> {
4386 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4387 :
4388 : Self::persist_tenant_config_at(tenant_shard_id, &config_path, location_conf).await
4389 : }
4390 :
4391 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4392 : pub(super) async fn persist_tenant_config_at(
4393 : tenant_shard_id: &TenantShardId,
4394 : config_path: &Utf8Path,
4395 : location_conf: &LocationConf,
4396 : ) -> std::io::Result<()> {
4397 : debug!("persisting tenantconf to {config_path}");
4398 :
4399 : let mut conf_content = r#"# This file contains a specific per-tenant's config.
4400 : # It is read in case of pageserver restart.
4401 : "#
4402 : .to_string();
4403 :
4404 0 : fail::fail_point!("tenant-config-before-write", |_| {
4405 0 : Err(std::io::Error::other("tenant-config-before-write"))
4406 0 : });
4407 :
4408 : // Convert the config to a toml file.
4409 : conf_content +=
4410 : &toml_edit::ser::to_string_pretty(&location_conf).expect("Config serialization failed");
4411 :
4412 : let temp_path = path_with_suffix_extension(config_path, TEMP_FILE_SUFFIX);
4413 :
4414 : let conf_content = conf_content.into_bytes();
4415 : VirtualFile::crashsafe_overwrite(config_path.to_owned(), temp_path, conf_content).await
4416 : }
4417 :
4418 : //
4419 : // How garbage collection works:
4420 : //
4421 : // +--bar------------->
4422 : // /
4423 : // +----+-----foo---------------->
4424 : // /
4425 : // ----main--+-------------------------->
4426 : // \
4427 : // +-----baz-------->
4428 : //
4429 : //
4430 : // 1. Grab 'gc_cs' mutex to prevent new timelines from being created while Timeline's
4431 : // `gc_infos` are being refreshed
4432 : // 2. Scan collected timelines, and on each timeline, make note of the
4433 : // all the points where other timelines have been branched off.
4434 : // We will refrain from removing page versions at those LSNs.
4435 : // 3. For each timeline, scan all layer files on the timeline.
4436 : // Remove all files for which a newer file exists and which
4437 : // don't cover any branch point LSNs.
4438 : //
4439 : // TODO:
4440 : // - if a relation has a non-incremental persistent layer on a child branch, then we
4441 : // don't need to keep that in the parent anymore. But currently
4442 : // we do.
4443 24 : async fn gc_iteration_internal(
4444 24 : &self,
4445 24 : target_timeline_id: Option<TimelineId>,
4446 24 : horizon: u64,
4447 24 : pitr: Duration,
4448 24 : cancel: &CancellationToken,
4449 24 : ctx: &RequestContext,
4450 24 : ) -> Result<GcResult, GcError> {
4451 24 : let mut totals: GcResult = Default::default();
4452 24 : let now = Instant::now();
4453 :
4454 24 : let gc_timelines = self
4455 24 : .refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4456 24 : .await?;
4457 :
4458 24 : failpoint_support::sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
4459 :
4460 : // If there is nothing to GC, we don't want any messages in the INFO log.
4461 24 : if !gc_timelines.is_empty() {
4462 24 : info!("{} timelines need GC", gc_timelines.len());
4463 : } else {
4464 0 : debug!("{} timelines need GC", gc_timelines.len());
4465 : }
4466 :
4467 : // Perform GC for each timeline.
4468 : //
4469 : // Note that we don't hold the `TenantShard::gc_cs` lock here because we don't want to delay the
4470 : // branch creation task, which requires the GC lock. A GC iteration can run concurrently
4471 : // with branch creation.
4472 : //
4473 : // See comments in [`TenantShard::branch_timeline`] for more information about why branch
4474 : // creation task can run concurrently with timeline's GC iteration.
4475 48 : for timeline in gc_timelines {
4476 24 : if cancel.is_cancelled() {
4477 : // We were requested to shut down. Stop and return with the progress we
4478 : // made.
4479 0 : break;
4480 24 : }
4481 24 : let result = match timeline.gc().await {
4482 : Err(GcError::TimelineCancelled) => {
4483 0 : if target_timeline_id.is_some() {
4484 : // If we were targetting this specific timeline, surface cancellation to caller
4485 0 : return Err(GcError::TimelineCancelled);
4486 : } else {
4487 : // A timeline may be shutting down independently of the tenant's lifecycle: we should
4488 : // skip past this and proceed to try GC on other timelines.
4489 0 : continue;
4490 : }
4491 : }
4492 24 : r => r?,
4493 : };
4494 24 : totals += result;
4495 : }
4496 :
4497 24 : totals.elapsed = now.elapsed();
4498 24 : Ok(totals)
4499 24 : }
4500 :
4501 : /// Refreshes the Timeline::gc_info for all timelines, returning the
4502 : /// vector of timelines which have [`Timeline::get_last_record_lsn`] past
4503 : /// [`TenantShard::get_gc_horizon`].
4504 : ///
4505 : /// This is usually executed as part of periodic gc, but can now be triggered more often.
4506 24 : pub(crate) async fn refresh_gc_info(
4507 24 : &self,
4508 24 : cancel: &CancellationToken,
4509 24 : ctx: &RequestContext,
4510 24 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4511 24 : // since this method can now be called at different rates than the configured gc loop, it
4512 24 : // might be that these configuration values get applied faster than what it was previously,
4513 24 : // since these were only read from the gc task.
4514 24 : let horizon = self.get_gc_horizon();
4515 24 : let pitr = self.get_pitr_interval();
4516 24 :
4517 24 : // refresh all timelines
4518 24 : let target_timeline_id = None;
4519 24 :
4520 24 : self.refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4521 24 : .await
4522 24 : }
4523 :
4524 : /// Populate all Timelines' `GcInfo` with information about their children. We do not set the
4525 : /// PITR cutoffs here, because that requires I/O: this is done later, before GC, by [`Self::refresh_gc_info_internal`]
4526 : ///
4527 : /// Subsequently, parent-child relationships are updated incrementally inside [`Timeline::new`] and [`Timeline::drop`].
4528 0 : fn initialize_gc_info(
4529 0 : &self,
4530 0 : timelines: &std::sync::MutexGuard<HashMap<TimelineId, Arc<Timeline>>>,
4531 0 : timelines_offloaded: &std::sync::MutexGuard<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
4532 0 : restrict_to_timeline: Option<TimelineId>,
4533 0 : ) {
4534 0 : if restrict_to_timeline.is_none() {
4535 : // This function must be called before activation: after activation timeline create/delete operations
4536 : // might happen, and this function is not safe to run concurrently with those.
4537 0 : assert!(!self.is_active());
4538 0 : }
4539 :
4540 : // Scan all timelines. For each timeline, remember the timeline ID and
4541 : // the branch point where it was created.
4542 0 : let mut all_branchpoints: BTreeMap<TimelineId, Vec<(Lsn, TimelineId, MaybeOffloaded)>> =
4543 0 : BTreeMap::new();
4544 0 : timelines.iter().for_each(|(timeline_id, timeline_entry)| {
4545 0 : if let Some(ancestor_timeline_id) = &timeline_entry.get_ancestor_timeline_id() {
4546 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4547 0 : ancestor_children.push((
4548 0 : timeline_entry.get_ancestor_lsn(),
4549 0 : *timeline_id,
4550 0 : MaybeOffloaded::No,
4551 0 : ));
4552 0 : }
4553 0 : });
4554 0 : timelines_offloaded
4555 0 : .iter()
4556 0 : .for_each(|(timeline_id, timeline_entry)| {
4557 0 : let Some(ancestor_timeline_id) = &timeline_entry.ancestor_timeline_id else {
4558 0 : return;
4559 : };
4560 0 : let Some(retain_lsn) = timeline_entry.ancestor_retain_lsn else {
4561 0 : return;
4562 : };
4563 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4564 0 : ancestor_children.push((retain_lsn, *timeline_id, MaybeOffloaded::Yes));
4565 0 : });
4566 0 :
4567 0 : // The number of bytes we always keep, irrespective of PITR: this is a constant across timelines
4568 0 : let horizon = self.get_gc_horizon();
4569 :
4570 : // Populate each timeline's GcInfo with information about its child branches
4571 0 : let timelines_to_write = if let Some(timeline_id) = restrict_to_timeline {
4572 0 : itertools::Either::Left(timelines.get(&timeline_id).into_iter())
4573 : } else {
4574 0 : itertools::Either::Right(timelines.values())
4575 : };
4576 0 : for timeline in timelines_to_write {
4577 0 : let mut branchpoints: Vec<(Lsn, TimelineId, MaybeOffloaded)> = all_branchpoints
4578 0 : .remove(&timeline.timeline_id)
4579 0 : .unwrap_or_default();
4580 0 :
4581 0 : branchpoints.sort_by_key(|b| b.0);
4582 0 :
4583 0 : let mut target = timeline.gc_info.write().unwrap();
4584 0 :
4585 0 : target.retain_lsns = branchpoints;
4586 0 :
4587 0 : let space_cutoff = timeline
4588 0 : .get_last_record_lsn()
4589 0 : .checked_sub(horizon)
4590 0 : .unwrap_or(Lsn(0));
4591 0 :
4592 0 : target.cutoffs = GcCutoffs {
4593 0 : space: space_cutoff,
4594 0 : time: Lsn::INVALID,
4595 0 : };
4596 0 : }
4597 0 : }
4598 :
4599 48 : async fn refresh_gc_info_internal(
4600 48 : &self,
4601 48 : target_timeline_id: Option<TimelineId>,
4602 48 : horizon: u64,
4603 48 : pitr: Duration,
4604 48 : cancel: &CancellationToken,
4605 48 : ctx: &RequestContext,
4606 48 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4607 48 : // before taking the gc_cs lock, do the heavier weight finding of gc_cutoff points for
4608 48 : // currently visible timelines.
4609 48 : let timelines = self
4610 48 : .timelines
4611 48 : .lock()
4612 48 : .unwrap()
4613 48 : .values()
4614 120 : .filter(|tl| match target_timeline_id.as_ref() {
4615 24 : Some(target) => &tl.timeline_id == target,
4616 96 : None => true,
4617 120 : })
4618 48 : .cloned()
4619 48 : .collect::<Vec<_>>();
4620 48 :
4621 48 : if target_timeline_id.is_some() && timelines.is_empty() {
4622 : // We were to act on a particular timeline and it wasn't found
4623 0 : return Err(GcError::TimelineNotFound);
4624 48 : }
4625 48 :
4626 48 : let mut gc_cutoffs: HashMap<TimelineId, GcCutoffs> =
4627 48 : HashMap::with_capacity(timelines.len());
4628 48 :
4629 48 : // Ensures all timelines use the same start time when computing the time cutoff.
4630 48 : let now_ts_for_pitr_calc = SystemTime::now();
4631 120 : for timeline in timelines.iter() {
4632 120 : let ctx = &ctx.with_scope_timeline(timeline);
4633 120 : let cutoff = timeline
4634 120 : .get_last_record_lsn()
4635 120 : .checked_sub(horizon)
4636 120 : .unwrap_or(Lsn(0));
4637 :
4638 120 : let cutoffs = timeline
4639 120 : .find_gc_cutoffs(now_ts_for_pitr_calc, cutoff, pitr, cancel, ctx)
4640 120 : .await?;
4641 120 : let old = gc_cutoffs.insert(timeline.timeline_id, cutoffs);
4642 120 : assert!(old.is_none());
4643 : }
4644 :
4645 48 : if !self.is_active() || self.cancel.is_cancelled() {
4646 0 : return Err(GcError::TenantCancelled);
4647 48 : }
4648 :
4649 : // grab mutex to prevent new timelines from being created here; avoid doing long operations
4650 : // because that will stall branch creation.
4651 48 : let gc_cs = self.gc_cs.lock().await;
4652 :
4653 : // Ok, we now know all the branch points.
4654 : // Update the GC information for each timeline.
4655 48 : let mut gc_timelines = Vec::with_capacity(timelines.len());
4656 168 : for timeline in timelines {
4657 : // We filtered the timeline list above
4658 120 : if let Some(target_timeline_id) = target_timeline_id {
4659 24 : assert_eq!(target_timeline_id, timeline.timeline_id);
4660 96 : }
4661 :
4662 : {
4663 120 : let mut target = timeline.gc_info.write().unwrap();
4664 120 :
4665 120 : // Cull any expired leases
4666 120 : let now = SystemTime::now();
4667 120 : target.leases.retain(|_, lease| !lease.is_expired(&now));
4668 120 :
4669 120 : timeline
4670 120 : .metrics
4671 120 : .valid_lsn_lease_count_gauge
4672 120 : .set(target.leases.len() as u64);
4673 :
4674 : // Look up parent's PITR cutoff to update the child's knowledge of whether it is within parent's PITR
4675 120 : if let Some(ancestor_id) = timeline.get_ancestor_timeline_id() {
4676 72 : if let Some(ancestor_gc_cutoffs) = gc_cutoffs.get(&ancestor_id) {
4677 72 : target.within_ancestor_pitr =
4678 72 : timeline.get_ancestor_lsn() >= ancestor_gc_cutoffs.time;
4679 72 : }
4680 48 : }
4681 :
4682 : // Update metrics that depend on GC state
4683 120 : timeline
4684 120 : .metrics
4685 120 : .archival_size
4686 120 : .set(if target.within_ancestor_pitr {
4687 0 : timeline.metrics.current_logical_size_gauge.get()
4688 : } else {
4689 120 : 0
4690 : });
4691 120 : timeline.metrics.pitr_history_size.set(
4692 120 : timeline
4693 120 : .get_last_record_lsn()
4694 120 : .checked_sub(target.cutoffs.time)
4695 120 : .unwrap_or(Lsn(0))
4696 120 : .0,
4697 120 : );
4698 :
4699 : // Apply the cutoffs we found to the Timeline's GcInfo. Why might we _not_ have cutoffs for a timeline?
4700 : // - this timeline was created while we were finding cutoffs
4701 : // - lsn for timestamp search fails for this timeline repeatedly
4702 120 : if let Some(cutoffs) = gc_cutoffs.get(&timeline.timeline_id) {
4703 120 : let original_cutoffs = target.cutoffs.clone();
4704 120 : // GC cutoffs should never go back
4705 120 : target.cutoffs = GcCutoffs {
4706 120 : space: Lsn(cutoffs.space.0.max(original_cutoffs.space.0)),
4707 120 : time: Lsn(cutoffs.time.0.max(original_cutoffs.time.0)),
4708 120 : }
4709 0 : }
4710 : }
4711 :
4712 120 : gc_timelines.push(timeline);
4713 : }
4714 48 : drop(gc_cs);
4715 48 : Ok(gc_timelines)
4716 48 : }
4717 :
4718 : /// A substitute for `branch_timeline` for use in unit tests.
4719 : /// The returned timeline will have state value `Active` to make various `anyhow::ensure!()`
4720 : /// calls pass, but, we do not actually call `.activate()` under the hood. So, none of the
4721 : /// timeline background tasks are launched, except the flush loop.
4722 : #[cfg(test)]
4723 1428 : async fn branch_timeline_test(
4724 1428 : self: &Arc<Self>,
4725 1428 : src_timeline: &Arc<Timeline>,
4726 1428 : dst_id: TimelineId,
4727 1428 : ancestor_lsn: Option<Lsn>,
4728 1428 : ctx: &RequestContext,
4729 1428 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
4730 1428 : let tl = self
4731 1428 : .branch_timeline_impl(src_timeline, dst_id, ancestor_lsn, ctx)
4732 1428 : .await?
4733 1404 : .into_timeline_for_test();
4734 1404 : tl.set_state(TimelineState::Active);
4735 1404 : Ok(tl)
4736 1428 : }
4737 :
4738 : /// Helper for unit tests to branch a timeline with some pre-loaded states.
4739 : #[cfg(test)]
4740 : #[allow(clippy::too_many_arguments)]
4741 72 : pub async fn branch_timeline_test_with_layers(
4742 72 : self: &Arc<Self>,
4743 72 : src_timeline: &Arc<Timeline>,
4744 72 : dst_id: TimelineId,
4745 72 : ancestor_lsn: Option<Lsn>,
4746 72 : ctx: &RequestContext,
4747 72 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
4748 72 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
4749 72 : end_lsn: Lsn,
4750 72 : ) -> anyhow::Result<Arc<Timeline>> {
4751 : use checks::check_valid_layermap;
4752 : use itertools::Itertools;
4753 :
4754 72 : let tline = self
4755 72 : .branch_timeline_test(src_timeline, dst_id, ancestor_lsn, ctx)
4756 72 : .await?;
4757 72 : let ancestor_lsn = if let Some(ancestor_lsn) = ancestor_lsn {
4758 72 : ancestor_lsn
4759 : } else {
4760 0 : tline.get_last_record_lsn()
4761 : };
4762 72 : assert!(end_lsn >= ancestor_lsn);
4763 72 : tline.force_advance_lsn(end_lsn);
4764 108 : for deltas in delta_layer_desc {
4765 36 : tline
4766 36 : .force_create_delta_layer(deltas, Some(ancestor_lsn), ctx)
4767 36 : .await?;
4768 : }
4769 96 : for (lsn, images) in image_layer_desc {
4770 24 : tline
4771 24 : .force_create_image_layer(lsn, images, Some(ancestor_lsn), ctx)
4772 24 : .await?;
4773 : }
4774 72 : let layer_names = tline
4775 72 : .layers
4776 72 : .read()
4777 72 : .await
4778 72 : .layer_map()
4779 72 : .unwrap()
4780 72 : .iter_historic_layers()
4781 72 : .map(|layer| layer.layer_name())
4782 72 : .collect_vec();
4783 72 : if let Some(err) = check_valid_layermap(&layer_names) {
4784 0 : bail!("invalid layermap: {err}");
4785 72 : }
4786 72 : Ok(tline)
4787 72 : }
4788 :
4789 : /// Branch an existing timeline.
4790 0 : async fn branch_timeline(
4791 0 : self: &Arc<Self>,
4792 0 : src_timeline: &Arc<Timeline>,
4793 0 : dst_id: TimelineId,
4794 0 : start_lsn: Option<Lsn>,
4795 0 : ctx: &RequestContext,
4796 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4797 0 : self.branch_timeline_impl(src_timeline, dst_id, start_lsn, ctx)
4798 0 : .await
4799 0 : }
4800 :
4801 1428 : async fn branch_timeline_impl(
4802 1428 : self: &Arc<Self>,
4803 1428 : src_timeline: &Arc<Timeline>,
4804 1428 : dst_id: TimelineId,
4805 1428 : start_lsn: Option<Lsn>,
4806 1428 : ctx: &RequestContext,
4807 1428 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4808 1428 : let src_id = src_timeline.timeline_id;
4809 :
4810 : // We will validate our ancestor LSN in this function. Acquire the GC lock so that
4811 : // this check cannot race with GC, and the ancestor LSN is guaranteed to remain
4812 : // valid while we are creating the branch.
4813 1428 : let _gc_cs = self.gc_cs.lock().await;
4814 :
4815 : // If no start LSN is specified, we branch the new timeline from the source timeline's last record LSN
4816 1428 : let start_lsn = start_lsn.unwrap_or_else(|| {
4817 12 : let lsn = src_timeline.get_last_record_lsn();
4818 12 : info!("branching timeline {dst_id} from timeline {src_id} at last record LSN: {lsn}");
4819 12 : lsn
4820 1428 : });
4821 :
4822 : // we finally have determined the ancestor_start_lsn, so we can get claim exclusivity now
4823 1428 : let timeline_create_guard = match self
4824 1428 : .start_creating_timeline(
4825 1428 : dst_id,
4826 1428 : CreateTimelineIdempotency::Branch {
4827 1428 : ancestor_timeline_id: src_timeline.timeline_id,
4828 1428 : ancestor_start_lsn: start_lsn,
4829 1428 : },
4830 1428 : )
4831 1428 : .await?
4832 : {
4833 1428 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
4834 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
4835 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
4836 : }
4837 : };
4838 :
4839 : // Ensure that `start_lsn` is valid, i.e. the LSN is within the PITR
4840 : // horizon on the source timeline
4841 : //
4842 : // We check it against both the planned GC cutoff stored in 'gc_info',
4843 : // and the 'latest_gc_cutoff' of the last GC that was performed. The
4844 : // planned GC cutoff in 'gc_info' is normally larger than
4845 : // 'applied_gc_cutoff_lsn', but beware of corner cases like if you just
4846 : // changed the GC settings for the tenant to make the PITR window
4847 : // larger, but some of the data was already removed by an earlier GC
4848 : // iteration.
4849 :
4850 : // check against last actual 'latest_gc_cutoff' first
4851 1428 : let applied_gc_cutoff_lsn = src_timeline.get_applied_gc_cutoff_lsn();
4852 1428 : {
4853 1428 : let gc_info = src_timeline.gc_info.read().unwrap();
4854 1428 : let planned_cutoff = gc_info.min_cutoff();
4855 1428 : if gc_info.lsn_covered_by_lease(start_lsn) {
4856 0 : tracing::info!(
4857 0 : "skipping comparison of {start_lsn} with gc cutoff {} and planned gc cutoff {planned_cutoff} due to lsn lease",
4858 0 : *applied_gc_cutoff_lsn
4859 : );
4860 : } else {
4861 1428 : src_timeline
4862 1428 : .check_lsn_is_in_scope(start_lsn, &applied_gc_cutoff_lsn)
4863 1428 : .context(format!(
4864 1428 : "invalid branch start lsn: less than latest GC cutoff {}",
4865 1428 : *applied_gc_cutoff_lsn,
4866 1428 : ))
4867 1428 : .map_err(CreateTimelineError::AncestorLsn)?;
4868 :
4869 : // and then the planned GC cutoff
4870 1404 : if start_lsn < planned_cutoff {
4871 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
4872 0 : "invalid branch start lsn: less than planned GC cutoff {planned_cutoff}"
4873 0 : )));
4874 1404 : }
4875 : }
4876 : }
4877 :
4878 : //
4879 : // The branch point is valid, and we are still holding the 'gc_cs' lock
4880 : // so that GC cannot advance the GC cutoff until we are finished.
4881 : // Proceed with the branch creation.
4882 : //
4883 :
4884 : // Determine prev-LSN for the new timeline. We can only determine it if
4885 : // the timeline was branched at the current end of the source timeline.
4886 : let RecordLsn {
4887 1404 : last: src_last,
4888 1404 : prev: src_prev,
4889 1404 : } = src_timeline.get_last_record_rlsn();
4890 1404 : let dst_prev = if src_last == start_lsn {
4891 1296 : Some(src_prev)
4892 : } else {
4893 108 : None
4894 : };
4895 :
4896 : // Create the metadata file, noting the ancestor of the new timeline.
4897 : // There is initially no data in it, but all the read-calls know to look
4898 : // into the ancestor.
4899 1404 : let metadata = TimelineMetadata::new(
4900 1404 : start_lsn,
4901 1404 : dst_prev,
4902 1404 : Some(src_id),
4903 1404 : start_lsn,
4904 1404 : *src_timeline.applied_gc_cutoff_lsn.read(), // FIXME: should we hold onto this guard longer?
4905 1404 : src_timeline.initdb_lsn,
4906 1404 : src_timeline.pg_version,
4907 1404 : );
4908 :
4909 1404 : let (uninitialized_timeline, _timeline_ctx) = self
4910 1404 : .prepare_new_timeline(
4911 1404 : dst_id,
4912 1404 : &metadata,
4913 1404 : timeline_create_guard,
4914 1404 : start_lsn + 1,
4915 1404 : Some(Arc::clone(src_timeline)),
4916 1404 : Some(src_timeline.get_rel_size_v2_status()),
4917 1404 : ctx,
4918 1404 : )
4919 1404 : .await?;
4920 :
4921 1404 : let new_timeline = uninitialized_timeline.finish_creation().await?;
4922 :
4923 : // Root timeline gets its layers during creation and uploads them along with the metadata.
4924 : // A branch timeline though, when created, can get no writes for some time, hence won't get any layers created.
4925 : // We still need to upload its metadata eagerly: if other nodes `attach` the tenant and miss this timeline, their GC
4926 : // could get incorrect information and remove more layers, than needed.
4927 : // See also https://github.com/neondatabase/neon/issues/3865
4928 1404 : new_timeline
4929 1404 : .remote_client
4930 1404 : .schedule_index_upload_for_full_metadata_update(&metadata)
4931 1404 : .context("branch initial metadata upload")?;
4932 :
4933 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
4934 :
4935 1404 : Ok(CreateTimelineResult::Created(new_timeline))
4936 1428 : }
4937 :
4938 : /// For unit tests, make this visible so that other modules can directly create timelines
4939 : #[cfg(test)]
4940 : #[tracing::instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
4941 : pub(crate) async fn bootstrap_timeline_test(
4942 : self: &Arc<Self>,
4943 : timeline_id: TimelineId,
4944 : pg_version: u32,
4945 : load_existing_initdb: Option<TimelineId>,
4946 : ctx: &RequestContext,
4947 : ) -> anyhow::Result<Arc<Timeline>> {
4948 : self.bootstrap_timeline(timeline_id, pg_version, load_existing_initdb, ctx)
4949 : .await
4950 : .map_err(anyhow::Error::new)
4951 12 : .map(|r| r.into_timeline_for_test())
4952 : }
4953 :
4954 : /// Get exclusive access to the timeline ID for creation.
4955 : ///
4956 : /// Timeline-creating code paths must use this function before making changes
4957 : /// to in-memory or persistent state.
4958 : ///
4959 : /// The `state` parameter is a description of the timeline creation operation
4960 : /// we intend to perform.
4961 : /// If the timeline was already created in the meantime, we check whether this
4962 : /// request conflicts or is idempotent , based on `state`.
4963 2784 : async fn start_creating_timeline(
4964 2784 : self: &Arc<Self>,
4965 2784 : new_timeline_id: TimelineId,
4966 2784 : idempotency: CreateTimelineIdempotency,
4967 2784 : ) -> Result<StartCreatingTimelineResult, CreateTimelineError> {
4968 2784 : let allow_offloaded = false;
4969 2784 : match self.create_timeline_create_guard(new_timeline_id, idempotency, allow_offloaded) {
4970 2772 : Ok(create_guard) => {
4971 2772 : pausable_failpoint!("timeline-creation-after-uninit");
4972 2772 : Ok(StartCreatingTimelineResult::CreateGuard(create_guard))
4973 : }
4974 0 : Err(TimelineExclusionError::ShuttingDown) => Err(CreateTimelineError::ShuttingDown),
4975 : Err(TimelineExclusionError::AlreadyCreating) => {
4976 : // Creation is in progress, we cannot create it again, and we cannot
4977 : // check if this request matches the existing one, so caller must try
4978 : // again later.
4979 0 : Err(CreateTimelineError::AlreadyCreating)
4980 : }
4981 0 : Err(TimelineExclusionError::Other(e)) => Err(CreateTimelineError::Other(e)),
4982 : Err(TimelineExclusionError::AlreadyExists {
4983 0 : existing: TimelineOrOffloaded::Offloaded(_existing),
4984 0 : ..
4985 0 : }) => {
4986 0 : info!("timeline already exists but is offloaded");
4987 0 : Err(CreateTimelineError::Conflict)
4988 : }
4989 : Err(TimelineExclusionError::AlreadyExists {
4990 12 : existing: TimelineOrOffloaded::Timeline(existing),
4991 12 : arg,
4992 12 : }) => {
4993 12 : {
4994 12 : let existing = &existing.create_idempotency;
4995 12 : let _span = info_span!("idempotency_check", ?existing, ?arg).entered();
4996 12 : debug!("timeline already exists");
4997 :
4998 12 : match (existing, &arg) {
4999 : // FailWithConflict => no idempotency check
5000 : (CreateTimelineIdempotency::FailWithConflict, _)
5001 : | (_, CreateTimelineIdempotency::FailWithConflict) => {
5002 12 : warn!("timeline already exists, failing request");
5003 12 : return Err(CreateTimelineError::Conflict);
5004 : }
5005 : // Idempotent <=> CreateTimelineIdempotency is identical
5006 0 : (x, y) if x == y => {
5007 0 : info!(
5008 0 : "timeline already exists and idempotency matches, succeeding request"
5009 : );
5010 : // fallthrough
5011 : }
5012 : (_, _) => {
5013 0 : warn!("idempotency conflict, failing request");
5014 0 : return Err(CreateTimelineError::Conflict);
5015 : }
5016 : }
5017 : }
5018 :
5019 0 : Ok(StartCreatingTimelineResult::Idempotent(existing))
5020 : }
5021 : }
5022 2784 : }
5023 :
5024 0 : async fn upload_initdb(
5025 0 : &self,
5026 0 : timelines_path: &Utf8PathBuf,
5027 0 : pgdata_path: &Utf8PathBuf,
5028 0 : timeline_id: &TimelineId,
5029 0 : ) -> anyhow::Result<()> {
5030 0 : let temp_path = timelines_path.join(format!(
5031 0 : "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
5032 0 : ));
5033 0 :
5034 0 : scopeguard::defer! {
5035 0 : if let Err(e) = fs::remove_file(&temp_path) {
5036 0 : error!("Failed to remove temporary initdb archive '{temp_path}': {e}");
5037 0 : }
5038 0 : }
5039 :
5040 0 : let (pgdata_zstd, tar_zst_size) = create_zst_tarball(pgdata_path, &temp_path).await?;
5041 : const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
5042 0 : if tar_zst_size > INITDB_TAR_ZST_WARN_LIMIT {
5043 0 : warn!(
5044 0 : "compressed {temp_path} size of {tar_zst_size} is above limit {INITDB_TAR_ZST_WARN_LIMIT}."
5045 : );
5046 0 : }
5047 :
5048 0 : pausable_failpoint!("before-initdb-upload");
5049 :
5050 0 : backoff::retry(
5051 0 : || async {
5052 0 : self::remote_timeline_client::upload_initdb_dir(
5053 0 : &self.remote_storage,
5054 0 : &self.tenant_shard_id.tenant_id,
5055 0 : timeline_id,
5056 0 : pgdata_zstd.try_clone().await?,
5057 0 : tar_zst_size,
5058 0 : &self.cancel,
5059 0 : )
5060 0 : .await
5061 0 : },
5062 0 : |_| false,
5063 0 : 3,
5064 0 : u32::MAX,
5065 0 : "persist_initdb_tar_zst",
5066 0 : &self.cancel,
5067 0 : )
5068 0 : .await
5069 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
5070 0 : .and_then(|x| x)
5071 0 : }
5072 :
5073 : /// - run initdb to init temporary instance and get bootstrap data
5074 : /// - after initialization completes, tar up the temp dir and upload it to S3.
5075 12 : async fn bootstrap_timeline(
5076 12 : self: &Arc<Self>,
5077 12 : timeline_id: TimelineId,
5078 12 : pg_version: u32,
5079 12 : load_existing_initdb: Option<TimelineId>,
5080 12 : ctx: &RequestContext,
5081 12 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
5082 12 : let timeline_create_guard = match self
5083 12 : .start_creating_timeline(
5084 12 : timeline_id,
5085 12 : CreateTimelineIdempotency::Bootstrap { pg_version },
5086 12 : )
5087 12 : .await?
5088 : {
5089 12 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
5090 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
5091 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
5092 : }
5093 : };
5094 :
5095 : // create a `tenant/{tenant_id}/timelines/basebackup-{timeline_id}.{TEMP_FILE_SUFFIX}/`
5096 : // temporary directory for basebackup files for the given timeline.
5097 :
5098 12 : let timelines_path = self.conf.timelines_path(&self.tenant_shard_id);
5099 12 : let pgdata_path = path_with_suffix_extension(
5100 12 : timelines_path.join(format!("basebackup-{timeline_id}")),
5101 12 : TEMP_FILE_SUFFIX,
5102 12 : );
5103 12 :
5104 12 : // Remove whatever was left from the previous runs: safe because TimelineCreateGuard guarantees
5105 12 : // we won't race with other creations or existent timelines with the same path.
5106 12 : if pgdata_path.exists() {
5107 0 : fs::remove_dir_all(&pgdata_path).with_context(|| {
5108 0 : format!("Failed to remove already existing initdb directory: {pgdata_path}")
5109 0 : })?;
5110 0 : tracing::info!("removed previous attempt's temporary initdb directory '{pgdata_path}'");
5111 12 : }
5112 :
5113 : // this new directory is very temporary, set to remove it immediately after bootstrap, we don't need it
5114 12 : let pgdata_path_deferred = pgdata_path.clone();
5115 12 : scopeguard::defer! {
5116 12 : if let Err(e) = fs::remove_dir_all(&pgdata_path_deferred).or_else(fs_ext::ignore_not_found) {
5117 12 : // this is unlikely, but we will remove the directory on pageserver restart or another bootstrap call
5118 12 : error!("Failed to remove temporary initdb directory '{pgdata_path_deferred}': {e}");
5119 12 : } else {
5120 12 : tracing::info!("removed temporary initdb directory '{pgdata_path_deferred}'");
5121 12 : }
5122 12 : }
5123 12 : if let Some(existing_initdb_timeline_id) = load_existing_initdb {
5124 12 : if existing_initdb_timeline_id != timeline_id {
5125 0 : let source_path = &remote_initdb_archive_path(
5126 0 : &self.tenant_shard_id.tenant_id,
5127 0 : &existing_initdb_timeline_id,
5128 0 : );
5129 0 : let dest_path =
5130 0 : &remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &timeline_id);
5131 0 :
5132 0 : // if this fails, it will get retried by retried control plane requests
5133 0 : self.remote_storage
5134 0 : .copy_object(source_path, dest_path, &self.cancel)
5135 0 : .await
5136 0 : .context("copy initdb tar")?;
5137 12 : }
5138 12 : let (initdb_tar_zst_path, initdb_tar_zst) =
5139 12 : self::remote_timeline_client::download_initdb_tar_zst(
5140 12 : self.conf,
5141 12 : &self.remote_storage,
5142 12 : &self.tenant_shard_id,
5143 12 : &existing_initdb_timeline_id,
5144 12 : &self.cancel,
5145 12 : )
5146 12 : .await
5147 12 : .context("download initdb tar")?;
5148 :
5149 12 : scopeguard::defer! {
5150 12 : if let Err(e) = fs::remove_file(&initdb_tar_zst_path) {
5151 12 : error!("Failed to remove temporary initdb archive '{initdb_tar_zst_path}': {e}");
5152 12 : }
5153 12 : }
5154 12 :
5155 12 : let buf_read =
5156 12 : BufReader::with_capacity(remote_timeline_client::BUFFER_SIZE, initdb_tar_zst);
5157 12 : extract_zst_tarball(&pgdata_path, buf_read)
5158 12 : .await
5159 12 : .context("extract initdb tar")?;
5160 : } else {
5161 : // Init temporarily repo to get bootstrap data, this creates a directory in the `pgdata_path` path
5162 0 : run_initdb(self.conf, &pgdata_path, pg_version, &self.cancel)
5163 0 : .await
5164 0 : .context("run initdb")?;
5165 :
5166 : // Upload the created data dir to S3
5167 0 : if self.tenant_shard_id().is_shard_zero() {
5168 0 : self.upload_initdb(&timelines_path, &pgdata_path, &timeline_id)
5169 0 : .await?;
5170 0 : }
5171 : }
5172 12 : let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
5173 12 :
5174 12 : // Import the contents of the data directory at the initial checkpoint
5175 12 : // LSN, and any WAL after that.
5176 12 : // Initdb lsn will be equal to last_record_lsn which will be set after import.
5177 12 : // Because we know it upfront avoid having an option or dummy zero value by passing it to the metadata.
5178 12 : let new_metadata = TimelineMetadata::new(
5179 12 : Lsn(0),
5180 12 : None,
5181 12 : None,
5182 12 : Lsn(0),
5183 12 : pgdata_lsn,
5184 12 : pgdata_lsn,
5185 12 : pg_version,
5186 12 : );
5187 12 : let (mut raw_timeline, timeline_ctx) = self
5188 12 : .prepare_new_timeline(
5189 12 : timeline_id,
5190 12 : &new_metadata,
5191 12 : timeline_create_guard,
5192 12 : pgdata_lsn,
5193 12 : None,
5194 12 : None,
5195 12 : ctx,
5196 12 : )
5197 12 : .await?;
5198 :
5199 12 : let tenant_shard_id = raw_timeline.owning_tenant.tenant_shard_id;
5200 12 : raw_timeline
5201 12 : .write(|unfinished_timeline| async move {
5202 12 : import_datadir::import_timeline_from_postgres_datadir(
5203 12 : &unfinished_timeline,
5204 12 : &pgdata_path,
5205 12 : pgdata_lsn,
5206 12 : &timeline_ctx,
5207 12 : )
5208 12 : .await
5209 12 : .with_context(|| {
5210 0 : format!(
5211 0 : "Failed to import pgdatadir for timeline {tenant_shard_id}/{timeline_id}"
5212 0 : )
5213 12 : })?;
5214 :
5215 12 : fail::fail_point!("before-checkpoint-new-timeline", |_| {
5216 0 : Err(CreateTimelineError::Other(anyhow::anyhow!(
5217 0 : "failpoint before-checkpoint-new-timeline"
5218 0 : )))
5219 12 : });
5220 :
5221 12 : Ok(())
5222 24 : })
5223 12 : .await?;
5224 :
5225 : // All done!
5226 12 : let timeline = raw_timeline.finish_creation().await?;
5227 :
5228 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
5229 :
5230 12 : Ok(CreateTimelineResult::Created(timeline))
5231 12 : }
5232 :
5233 2748 : fn build_timeline_remote_client(&self, timeline_id: TimelineId) -> RemoteTimelineClient {
5234 2748 : RemoteTimelineClient::new(
5235 2748 : self.remote_storage.clone(),
5236 2748 : self.deletion_queue_client.clone(),
5237 2748 : self.conf,
5238 2748 : self.tenant_shard_id,
5239 2748 : timeline_id,
5240 2748 : self.generation,
5241 2748 : &self.tenant_conf.load().location,
5242 2748 : )
5243 2748 : }
5244 :
5245 : /// Builds required resources for a new timeline.
5246 2748 : fn build_timeline_resources(&self, timeline_id: TimelineId) -> TimelineResources {
5247 2748 : let remote_client = self.build_timeline_remote_client(timeline_id);
5248 2748 : self.get_timeline_resources_for(remote_client)
5249 2748 : }
5250 :
5251 : /// Builds timeline resources for the given remote client.
5252 2784 : fn get_timeline_resources_for(&self, remote_client: RemoteTimelineClient) -> TimelineResources {
5253 2784 : TimelineResources {
5254 2784 : remote_client,
5255 2784 : pagestream_throttle: self.pagestream_throttle.clone(),
5256 2784 : pagestream_throttle_metrics: self.pagestream_throttle_metrics.clone(),
5257 2784 : l0_compaction_trigger: self.l0_compaction_trigger.clone(),
5258 2784 : l0_flush_global_state: self.l0_flush_global_state.clone(),
5259 2784 : }
5260 2784 : }
5261 :
5262 : /// Creates intermediate timeline structure and its files.
5263 : ///
5264 : /// An empty layer map is initialized, and new data and WAL can be imported starting
5265 : /// at 'disk_consistent_lsn'. After any initial data has been imported, call
5266 : /// `finish_creation` to insert the Timeline into the timelines map.
5267 : #[allow(clippy::too_many_arguments)]
5268 2748 : async fn prepare_new_timeline<'a>(
5269 2748 : &'a self,
5270 2748 : new_timeline_id: TimelineId,
5271 2748 : new_metadata: &TimelineMetadata,
5272 2748 : create_guard: TimelineCreateGuard,
5273 2748 : start_lsn: Lsn,
5274 2748 : ancestor: Option<Arc<Timeline>>,
5275 2748 : rel_size_v2_status: Option<RelSizeMigration>,
5276 2748 : ctx: &RequestContext,
5277 2748 : ) -> anyhow::Result<(UninitializedTimeline<'a>, RequestContext)> {
5278 2748 : let tenant_shard_id = self.tenant_shard_id;
5279 2748 :
5280 2748 : let resources = self.build_timeline_resources(new_timeline_id);
5281 2748 : resources
5282 2748 : .remote_client
5283 2748 : .init_upload_queue_for_empty_remote(new_metadata, rel_size_v2_status.clone())?;
5284 :
5285 2748 : let (timeline_struct, timeline_ctx) = self
5286 2748 : .create_timeline_struct(
5287 2748 : new_timeline_id,
5288 2748 : new_metadata,
5289 2748 : None,
5290 2748 : ancestor,
5291 2748 : resources,
5292 2748 : CreateTimelineCause::Load,
5293 2748 : create_guard.idempotency.clone(),
5294 2748 : None,
5295 2748 : rel_size_v2_status,
5296 2748 : ctx,
5297 2748 : )
5298 2748 : .context("Failed to create timeline data structure")?;
5299 :
5300 2748 : timeline_struct.init_empty_layer_map(start_lsn);
5301 :
5302 2748 : if let Err(e) = self
5303 2748 : .create_timeline_files(&create_guard.timeline_path)
5304 2748 : .await
5305 : {
5306 0 : error!(
5307 0 : "Failed to create initial files for timeline {tenant_shard_id}/{new_timeline_id}, cleaning up: {e:?}"
5308 : );
5309 0 : cleanup_timeline_directory(create_guard);
5310 0 : return Err(e);
5311 2748 : }
5312 2748 :
5313 2748 : debug!(
5314 0 : "Successfully created initial files for timeline {tenant_shard_id}/{new_timeline_id}"
5315 : );
5316 :
5317 2748 : Ok((
5318 2748 : UninitializedTimeline::new(
5319 2748 : self,
5320 2748 : new_timeline_id,
5321 2748 : Some((timeline_struct, create_guard)),
5322 2748 : ),
5323 2748 : timeline_ctx,
5324 2748 : ))
5325 2748 : }
5326 :
5327 2748 : async fn create_timeline_files(&self, timeline_path: &Utf8Path) -> anyhow::Result<()> {
5328 2748 : crashsafe::create_dir(timeline_path).context("Failed to create timeline directory")?;
5329 :
5330 2748 : fail::fail_point!("after-timeline-dir-creation", |_| {
5331 0 : anyhow::bail!("failpoint after-timeline-dir-creation");
5332 2748 : });
5333 :
5334 2748 : Ok(())
5335 2748 : }
5336 :
5337 : /// Get a guard that provides exclusive access to the timeline directory, preventing
5338 : /// concurrent attempts to create the same timeline.
5339 : ///
5340 : /// The `allow_offloaded` parameter controls whether to tolerate the existence of
5341 : /// offloaded timelines or not.
5342 2784 : fn create_timeline_create_guard(
5343 2784 : self: &Arc<Self>,
5344 2784 : timeline_id: TimelineId,
5345 2784 : idempotency: CreateTimelineIdempotency,
5346 2784 : allow_offloaded: bool,
5347 2784 : ) -> Result<TimelineCreateGuard, TimelineExclusionError> {
5348 2784 : let tenant_shard_id = self.tenant_shard_id;
5349 2784 :
5350 2784 : let timeline_path = self.conf.timeline_path(&tenant_shard_id, &timeline_id);
5351 :
5352 2784 : let create_guard = TimelineCreateGuard::new(
5353 2784 : self,
5354 2784 : timeline_id,
5355 2784 : timeline_path.clone(),
5356 2784 : idempotency,
5357 2784 : allow_offloaded,
5358 2784 : )?;
5359 :
5360 : // At this stage, we have got exclusive access to in-memory state for this timeline ID
5361 : // for creation.
5362 : // A timeline directory should never exist on disk already:
5363 : // - a previous failed creation would have cleaned up after itself
5364 : // - a pageserver restart would clean up timeline directories that don't have valid remote state
5365 : //
5366 : // Therefore it is an unexpected internal error to encounter a timeline directory already existing here,
5367 : // this error may indicate a bug in cleanup on failed creations.
5368 2772 : if timeline_path.exists() {
5369 0 : return Err(TimelineExclusionError::Other(anyhow::anyhow!(
5370 0 : "Timeline directory already exists! This is a bug."
5371 0 : )));
5372 2772 : }
5373 2772 :
5374 2772 : Ok(create_guard)
5375 2784 : }
5376 :
5377 : /// Gathers inputs from all of the timelines to produce a sizing model input.
5378 : ///
5379 : /// Future is cancellation safe. Only one calculation can be running at once per tenant.
5380 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5381 : pub async fn gather_size_inputs(
5382 : &self,
5383 : // `max_retention_period` overrides the cutoff that is used to calculate the size
5384 : // (only if it is shorter than the real cutoff).
5385 : max_retention_period: Option<u64>,
5386 : cause: LogicalSizeCalculationCause,
5387 : cancel: &CancellationToken,
5388 : ctx: &RequestContext,
5389 : ) -> Result<size::ModelInputs, size::CalculateSyntheticSizeError> {
5390 : let logical_sizes_at_once = self
5391 : .conf
5392 : .concurrent_tenant_size_logical_size_queries
5393 : .inner();
5394 :
5395 : // TODO: Having a single mutex block concurrent reads is not great for performance.
5396 : //
5397 : // But the only case where we need to run multiple of these at once is when we
5398 : // request a size for a tenant manually via API, while another background calculation
5399 : // is in progress (which is not a common case).
5400 : //
5401 : // See more for on the issue #2748 condenced out of the initial PR review.
5402 : let mut shared_cache = tokio::select! {
5403 : locked = self.cached_logical_sizes.lock() => locked,
5404 : _ = cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5405 : _ = self.cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5406 : };
5407 :
5408 : size::gather_inputs(
5409 : self,
5410 : logical_sizes_at_once,
5411 : max_retention_period,
5412 : &mut shared_cache,
5413 : cause,
5414 : cancel,
5415 : ctx,
5416 : )
5417 : .await
5418 : }
5419 :
5420 : /// Calculate synthetic tenant size and cache the result.
5421 : /// This is periodically called by background worker.
5422 : /// result is cached in tenant struct
5423 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5424 : pub async fn calculate_synthetic_size(
5425 : &self,
5426 : cause: LogicalSizeCalculationCause,
5427 : cancel: &CancellationToken,
5428 : ctx: &RequestContext,
5429 : ) -> Result<u64, size::CalculateSyntheticSizeError> {
5430 : let inputs = self.gather_size_inputs(None, cause, cancel, ctx).await?;
5431 :
5432 : let size = inputs.calculate();
5433 :
5434 : self.set_cached_synthetic_size(size);
5435 :
5436 : Ok(size)
5437 : }
5438 :
5439 : /// Cache given synthetic size and update the metric value
5440 0 : pub fn set_cached_synthetic_size(&self, size: u64) {
5441 0 : self.cached_synthetic_tenant_size
5442 0 : .store(size, Ordering::Relaxed);
5443 0 :
5444 0 : // Only shard zero should be calculating synthetic sizes
5445 0 : debug_assert!(self.shard_identity.is_shard_zero());
5446 :
5447 0 : TENANT_SYNTHETIC_SIZE_METRIC
5448 0 : .get_metric_with_label_values(&[&self.tenant_shard_id.tenant_id.to_string()])
5449 0 : .unwrap()
5450 0 : .set(size);
5451 0 : }
5452 :
5453 0 : pub fn cached_synthetic_size(&self) -> u64 {
5454 0 : self.cached_synthetic_tenant_size.load(Ordering::Relaxed)
5455 0 : }
5456 :
5457 : /// Flush any in-progress layers, schedule uploads, and wait for uploads to complete.
5458 : ///
5459 : /// This function can take a long time: callers should wrap it in a timeout if calling
5460 : /// from an external API handler.
5461 : ///
5462 : /// Cancel-safety: cancelling this function may leave I/O running, but such I/O is
5463 : /// still bounded by tenant/timeline shutdown.
5464 : #[tracing::instrument(skip_all)]
5465 : pub(crate) async fn flush_remote(&self) -> anyhow::Result<()> {
5466 : let timelines = self.timelines.lock().unwrap().clone();
5467 :
5468 0 : async fn flush_timeline(_gate: GateGuard, timeline: Arc<Timeline>) -> anyhow::Result<()> {
5469 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Flushing...");
5470 0 : timeline.freeze_and_flush().await?;
5471 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Waiting for uploads...");
5472 0 : timeline.remote_client.wait_completion().await?;
5473 :
5474 0 : Ok(())
5475 0 : }
5476 :
5477 : // We do not use a JoinSet for these tasks, because we don't want them to be
5478 : // aborted when this function's future is cancelled: they should stay alive
5479 : // holding their GateGuard until they complete, to ensure their I/Os complete
5480 : // before Timeline shutdown completes.
5481 : let mut results = FuturesUnordered::new();
5482 :
5483 : for (_timeline_id, timeline) in timelines {
5484 : // Run each timeline's flush in a task holding the timeline's gate: this
5485 : // means that if this function's future is cancelled, the Timeline shutdown
5486 : // will still wait for any I/O in here to complete.
5487 : let Ok(gate) = timeline.gate.enter() else {
5488 : continue;
5489 : };
5490 0 : let jh = tokio::task::spawn(async move { flush_timeline(gate, timeline).await });
5491 : results.push(jh);
5492 : }
5493 :
5494 : while let Some(r) = results.next().await {
5495 : if let Err(e) = r {
5496 : if !e.is_cancelled() && !e.is_panic() {
5497 : tracing::error!("unexpected join error: {e:?}");
5498 : }
5499 : }
5500 : }
5501 :
5502 : // The flushes we did above were just writes, but the TenantShard might have had
5503 : // pending deletions as well from recent compaction/gc: we want to flush those
5504 : // as well. This requires flushing the global delete queue. This is cheap
5505 : // because it's typically a no-op.
5506 : match self.deletion_queue_client.flush_execute().await {
5507 : Ok(_) => {}
5508 : Err(DeletionQueueError::ShuttingDown) => {}
5509 : }
5510 :
5511 : Ok(())
5512 : }
5513 :
5514 0 : pub(crate) fn get_tenant_conf(&self) -> pageserver_api::models::TenantConfig {
5515 0 : self.tenant_conf.load().tenant_conf.clone()
5516 0 : }
5517 :
5518 : /// How much local storage would this tenant like to have? It can cope with
5519 : /// less than this (via eviction and on-demand downloads), but this function enables
5520 : /// the TenantShard to advertise how much storage it would prefer to have to provide fast I/O
5521 : /// by keeping important things on local disk.
5522 : ///
5523 : /// This is a heuristic, not a guarantee: tenants that are long-idle will actually use less
5524 : /// than they report here, due to layer eviction. Tenants with many active branches may
5525 : /// actually use more than they report here.
5526 0 : pub(crate) fn local_storage_wanted(&self) -> u64 {
5527 0 : let timelines = self.timelines.lock().unwrap();
5528 0 :
5529 0 : // Heuristic: we use the max() of the timelines' visible sizes, rather than the sum. This
5530 0 : // reflects the observation that on tenants with multiple large branches, typically only one
5531 0 : // of them is used actively enough to occupy space on disk.
5532 0 : timelines
5533 0 : .values()
5534 0 : .map(|t| t.metrics.visible_physical_size_gauge.get())
5535 0 : .max()
5536 0 : .unwrap_or(0)
5537 0 : }
5538 :
5539 : /// Builds a new tenant manifest, and uploads it if it differs from the last-known tenant
5540 : /// manifest in `Self::remote_tenant_manifest`.
5541 : ///
5542 : /// TODO: instead of requiring callers to remember to call `maybe_upload_tenant_manifest` after
5543 : /// changing any `TenantShard` state that's included in the manifest, consider making the manifest
5544 : /// the authoritative source of data with an API that automatically uploads on changes. Revisit
5545 : /// this when the manifest is more widely used and we have a better idea of the data model.
5546 1404 : pub(crate) async fn maybe_upload_tenant_manifest(&self) -> Result<(), TenantManifestError> {
5547 : // Multiple tasks may call this function concurrently after mutating the TenantShard runtime
5548 : // state, affecting the manifest generated by `build_tenant_manifest`. We use an async mutex
5549 : // to serialize these callers. `eq_ignoring_version` acts as a slightly inefficient but
5550 : // simple coalescing mechanism.
5551 1404 : let mut guard = tokio::select! {
5552 1404 : guard = self.remote_tenant_manifest.lock() => guard,
5553 1404 : _ = self.cancel.cancelled() => return Err(TenantManifestError::Cancelled),
5554 : };
5555 :
5556 : // Build a new manifest.
5557 1404 : let manifest = self.build_tenant_manifest();
5558 :
5559 : // Check if the manifest has changed. We ignore the version number here, to avoid
5560 : // uploading every manifest on version number bumps.
5561 1404 : if let Some(old) = guard.as_ref() {
5562 48 : if manifest.eq_ignoring_version(old) {
5563 36 : return Ok(());
5564 12 : }
5565 1356 : }
5566 :
5567 : // Upload the manifest. Remote storage does no retries internally, so retry here.
5568 1368 : match backoff::retry(
5569 1368 : || async {
5570 1368 : upload_tenant_manifest(
5571 1368 : &self.remote_storage,
5572 1368 : &self.tenant_shard_id,
5573 1368 : self.generation,
5574 1368 : &manifest,
5575 1368 : &self.cancel,
5576 1368 : )
5577 1368 : .await
5578 2736 : },
5579 1368 : |_| self.cancel.is_cancelled(),
5580 1368 : FAILED_UPLOAD_WARN_THRESHOLD,
5581 1368 : FAILED_REMOTE_OP_RETRIES,
5582 1368 : "uploading tenant manifest",
5583 1368 : &self.cancel,
5584 1368 : )
5585 1368 : .await
5586 : {
5587 0 : None => Err(TenantManifestError::Cancelled),
5588 0 : Some(Err(_)) if self.cancel.is_cancelled() => Err(TenantManifestError::Cancelled),
5589 0 : Some(Err(e)) => Err(TenantManifestError::RemoteStorage(e)),
5590 : Some(Ok(_)) => {
5591 : // Store the successfully uploaded manifest, so that future callers can avoid
5592 : // re-uploading the same thing.
5593 1368 : *guard = Some(manifest);
5594 1368 :
5595 1368 : Ok(())
5596 : }
5597 : }
5598 1404 : }
5599 : }
5600 :
5601 : /// Create the cluster temporarily in 'initdbpath' directory inside the repository
5602 : /// to get bootstrap data for timeline initialization.
5603 0 : async fn run_initdb(
5604 0 : conf: &'static PageServerConf,
5605 0 : initdb_target_dir: &Utf8Path,
5606 0 : pg_version: u32,
5607 0 : cancel: &CancellationToken,
5608 0 : ) -> Result<(), InitdbError> {
5609 0 : let initdb_bin_path = conf
5610 0 : .pg_bin_dir(pg_version)
5611 0 : .map_err(InitdbError::Other)?
5612 0 : .join("initdb");
5613 0 : let initdb_lib_dir = conf.pg_lib_dir(pg_version).map_err(InitdbError::Other)?;
5614 0 : info!(
5615 0 : "running {} in {}, libdir: {}",
5616 : initdb_bin_path, initdb_target_dir, initdb_lib_dir,
5617 : );
5618 :
5619 0 : let _permit = {
5620 0 : let _timer = INITDB_SEMAPHORE_ACQUISITION_TIME.start_timer();
5621 0 : INIT_DB_SEMAPHORE.acquire().await
5622 : };
5623 :
5624 0 : CONCURRENT_INITDBS.inc();
5625 0 : scopeguard::defer! {
5626 0 : CONCURRENT_INITDBS.dec();
5627 0 : }
5628 0 :
5629 0 : let _timer = INITDB_RUN_TIME.start_timer();
5630 0 : let res = postgres_initdb::do_run_initdb(postgres_initdb::RunInitdbArgs {
5631 0 : superuser: &conf.superuser,
5632 0 : locale: &conf.locale,
5633 0 : initdb_bin: &initdb_bin_path,
5634 0 : pg_version,
5635 0 : library_search_path: &initdb_lib_dir,
5636 0 : pgdata: initdb_target_dir,
5637 0 : })
5638 0 : .await
5639 0 : .map_err(InitdbError::Inner);
5640 0 :
5641 0 : // This isn't true cancellation support, see above. Still return an error to
5642 0 : // excercise the cancellation code path.
5643 0 : if cancel.is_cancelled() {
5644 0 : return Err(InitdbError::Cancelled);
5645 0 : }
5646 0 :
5647 0 : res
5648 0 : }
5649 :
5650 : /// Dump contents of a layer file to stdout.
5651 0 : pub async fn dump_layerfile_from_path(
5652 0 : path: &Utf8Path,
5653 0 : verbose: bool,
5654 0 : ctx: &RequestContext,
5655 0 : ) -> anyhow::Result<()> {
5656 : use std::os::unix::fs::FileExt;
5657 :
5658 : // All layer files start with a two-byte "magic" value, to identify the kind of
5659 : // file.
5660 0 : let file = File::open(path)?;
5661 0 : let mut header_buf = [0u8; 2];
5662 0 : file.read_exact_at(&mut header_buf, 0)?;
5663 :
5664 0 : match u16::from_be_bytes(header_buf) {
5665 : crate::IMAGE_FILE_MAGIC => {
5666 0 : ImageLayer::new_for_path(path, file)?
5667 0 : .dump(verbose, ctx)
5668 0 : .await?
5669 : }
5670 : crate::DELTA_FILE_MAGIC => {
5671 0 : DeltaLayer::new_for_path(path, file)?
5672 0 : .dump(verbose, ctx)
5673 0 : .await?
5674 : }
5675 0 : magic => bail!("unrecognized magic identifier: {:?}", magic),
5676 : }
5677 :
5678 0 : Ok(())
5679 0 : }
5680 :
5681 : #[cfg(test)]
5682 : pub(crate) mod harness {
5683 : use bytes::{Bytes, BytesMut};
5684 : use hex_literal::hex;
5685 : use once_cell::sync::OnceCell;
5686 : use pageserver_api::key::Key;
5687 : use pageserver_api::models::ShardParameters;
5688 : use pageserver_api::record::NeonWalRecord;
5689 : use pageserver_api::shard::ShardIndex;
5690 : use utils::id::TenantId;
5691 : use utils::logging;
5692 :
5693 : use super::*;
5694 : use crate::deletion_queue::mock::MockDeletionQueue;
5695 : use crate::l0_flush::L0FlushConfig;
5696 : use crate::walredo::apply_neon;
5697 :
5698 : pub const TIMELINE_ID: TimelineId =
5699 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
5700 : pub const NEW_TIMELINE_ID: TimelineId =
5701 : TimelineId::from_array(hex!("AA223344556677881122334455667788"));
5702 :
5703 : /// Convenience function to create a page image with given string as the only content
5704 30172700 : pub fn test_img(s: &str) -> Bytes {
5705 30172700 : let mut buf = BytesMut::new();
5706 30172700 : buf.extend_from_slice(s.as_bytes());
5707 30172700 : buf.resize(64, 0);
5708 30172700 :
5709 30172700 : buf.freeze()
5710 30172700 : }
5711 :
5712 : pub struct TenantHarness {
5713 : pub conf: &'static PageServerConf,
5714 : pub tenant_conf: pageserver_api::models::TenantConfig,
5715 : pub tenant_shard_id: TenantShardId,
5716 : pub generation: Generation,
5717 : pub shard: ShardIndex,
5718 : pub remote_storage: GenericRemoteStorage,
5719 : pub remote_fs_dir: Utf8PathBuf,
5720 : pub deletion_queue: MockDeletionQueue,
5721 : }
5722 :
5723 : static LOG_HANDLE: OnceCell<()> = OnceCell::new();
5724 :
5725 1536 : pub(crate) fn setup_logging() {
5726 1536 : LOG_HANDLE.get_or_init(|| {
5727 1464 : logging::init(
5728 1464 : logging::LogFormat::Test,
5729 1464 : // enable it in case the tests exercise code paths that use
5730 1464 : // debug_assert_current_span_has_tenant_and_timeline_id
5731 1464 : logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
5732 1464 : logging::Output::Stdout,
5733 1464 : )
5734 1464 : .expect("Failed to init test logging");
5735 1536 : });
5736 1536 : }
5737 :
5738 : impl TenantHarness {
5739 1392 : pub async fn create_custom(
5740 1392 : test_name: &'static str,
5741 1392 : tenant_conf: pageserver_api::models::TenantConfig,
5742 1392 : tenant_id: TenantId,
5743 1392 : shard_identity: ShardIdentity,
5744 1392 : generation: Generation,
5745 1392 : ) -> anyhow::Result<Self> {
5746 1392 : setup_logging();
5747 1392 :
5748 1392 : let repo_dir = PageServerConf::test_repo_dir(test_name);
5749 1392 : let _ = fs::remove_dir_all(&repo_dir);
5750 1392 : fs::create_dir_all(&repo_dir)?;
5751 :
5752 1392 : let conf = PageServerConf::dummy_conf(repo_dir);
5753 1392 : // Make a static copy of the config. This can never be free'd, but that's
5754 1392 : // OK in a test.
5755 1392 : let conf: &'static PageServerConf = Box::leak(Box::new(conf));
5756 1392 :
5757 1392 : let shard = shard_identity.shard_index();
5758 1392 : let tenant_shard_id = TenantShardId {
5759 1392 : tenant_id,
5760 1392 : shard_number: shard.shard_number,
5761 1392 : shard_count: shard.shard_count,
5762 1392 : };
5763 1392 : fs::create_dir_all(conf.tenant_path(&tenant_shard_id))?;
5764 1392 : fs::create_dir_all(conf.timelines_path(&tenant_shard_id))?;
5765 :
5766 : use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
5767 1392 : let remote_fs_dir = conf.workdir.join("localfs");
5768 1392 : std::fs::create_dir_all(&remote_fs_dir).unwrap();
5769 1392 : let config = RemoteStorageConfig {
5770 1392 : storage: RemoteStorageKind::LocalFs {
5771 1392 : local_path: remote_fs_dir.clone(),
5772 1392 : },
5773 1392 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
5774 1392 : small_timeout: RemoteStorageConfig::DEFAULT_SMALL_TIMEOUT,
5775 1392 : };
5776 1392 : let remote_storage = GenericRemoteStorage::from_config(&config).await.unwrap();
5777 1392 : let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()));
5778 1392 :
5779 1392 : Ok(Self {
5780 1392 : conf,
5781 1392 : tenant_conf,
5782 1392 : tenant_shard_id,
5783 1392 : generation,
5784 1392 : shard,
5785 1392 : remote_storage,
5786 1392 : remote_fs_dir,
5787 1392 : deletion_queue,
5788 1392 : })
5789 1392 : }
5790 :
5791 1308 : pub async fn create(test_name: &'static str) -> anyhow::Result<Self> {
5792 1308 : // Disable automatic GC and compaction to make the unit tests more deterministic.
5793 1308 : // The tests perform them manually if needed.
5794 1308 : let tenant_conf = pageserver_api::models::TenantConfig {
5795 1308 : gc_period: Some(Duration::ZERO),
5796 1308 : compaction_period: Some(Duration::ZERO),
5797 1308 : ..Default::default()
5798 1308 : };
5799 1308 : let tenant_id = TenantId::generate();
5800 1308 : let shard = ShardIdentity::unsharded();
5801 1308 : Self::create_custom(
5802 1308 : test_name,
5803 1308 : tenant_conf,
5804 1308 : tenant_id,
5805 1308 : shard,
5806 1308 : Generation::new(0xdeadbeef),
5807 1308 : )
5808 1308 : .await
5809 1308 : }
5810 :
5811 120 : pub fn span(&self) -> tracing::Span {
5812 120 : info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
5813 120 : }
5814 :
5815 1392 : pub(crate) async fn load(&self) -> (Arc<TenantShard>, RequestContext) {
5816 1392 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error)
5817 1392 : .with_scope_unit_test();
5818 1392 : (
5819 1392 : self.do_try_load(&ctx)
5820 1392 : .await
5821 1392 : .expect("failed to load test tenant"),
5822 1392 : ctx,
5823 1392 : )
5824 1392 : }
5825 :
5826 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5827 : pub(crate) async fn do_try_load(
5828 : &self,
5829 : ctx: &RequestContext,
5830 : ) -> anyhow::Result<Arc<TenantShard>> {
5831 : let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
5832 :
5833 : let tenant = Arc::new(TenantShard::new(
5834 : TenantState::Attaching,
5835 : self.conf,
5836 : AttachedTenantConf::try_from(LocationConf::attached_single(
5837 : self.tenant_conf.clone(),
5838 : self.generation,
5839 : &ShardParameters::default(),
5840 : ))
5841 : .unwrap(),
5842 : // This is a legacy/test code path: sharding isn't supported here.
5843 : ShardIdentity::unsharded(),
5844 : Some(walredo_mgr),
5845 : self.tenant_shard_id,
5846 : self.remote_storage.clone(),
5847 : self.deletion_queue.new_client(),
5848 : // TODO: ideally we should run all unit tests with both configs
5849 : L0FlushGlobalState::new(L0FlushConfig::default()),
5850 : ));
5851 :
5852 : let preload = tenant
5853 : .preload(&self.remote_storage, CancellationToken::new())
5854 : .await?;
5855 : tenant.attach(Some(preload), ctx).await?;
5856 :
5857 : tenant.state.send_replace(TenantState::Active);
5858 : for timeline in tenant.timelines.lock().unwrap().values() {
5859 : timeline.set_state(TimelineState::Active);
5860 : }
5861 : Ok(tenant)
5862 : }
5863 :
5864 12 : pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
5865 12 : self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
5866 12 : }
5867 : }
5868 :
5869 : // Mock WAL redo manager that doesn't do much
5870 : pub(crate) struct TestRedoManager;
5871 :
5872 : impl TestRedoManager {
5873 : /// # Cancel-Safety
5874 : ///
5875 : /// This method is cancellation-safe.
5876 321288 : pub async fn request_redo(
5877 321288 : &self,
5878 321288 : key: Key,
5879 321288 : lsn: Lsn,
5880 321288 : base_img: Option<(Lsn, Bytes)>,
5881 321288 : records: Vec<(Lsn, NeonWalRecord)>,
5882 321288 : _pg_version: u32,
5883 321288 : _redo_attempt_type: RedoAttemptType,
5884 321288 : ) -> Result<Bytes, walredo::Error> {
5885 16842120 : let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
5886 321288 : if records_neon {
5887 : // For Neon wal records, we can decode without spawning postgres, so do so.
5888 321288 : let mut page = match (base_img, records.first()) {
5889 156348 : (Some((_lsn, img)), _) => {
5890 156348 : let mut page = BytesMut::new();
5891 156348 : page.extend_from_slice(&img);
5892 156348 : page
5893 : }
5894 164940 : (_, Some((_lsn, rec))) if rec.will_init() => BytesMut::new(),
5895 : _ => {
5896 0 : panic!("Neon WAL redo requires base image or will init record");
5897 : }
5898 : };
5899 :
5900 17163396 : for (record_lsn, record) in records {
5901 16842120 : apply_neon::apply_in_neon(&record, record_lsn, key, &mut page)?;
5902 : }
5903 321276 : Ok(page.freeze())
5904 : } else {
5905 : // We never spawn a postgres walredo process in unit tests: just log what we might have done.
5906 0 : let s = format!(
5907 0 : "redo for {} to get to {}, with {} and {} records",
5908 0 : key,
5909 0 : lsn,
5910 0 : if base_img.is_some() {
5911 0 : "base image"
5912 : } else {
5913 0 : "no base image"
5914 : },
5915 0 : records.len()
5916 0 : );
5917 0 : println!("{s}");
5918 0 :
5919 0 : Ok(test_img(&s))
5920 : }
5921 321288 : }
5922 : }
5923 : }
5924 :
5925 : #[cfg(test)]
5926 : mod tests {
5927 : use std::collections::{BTreeMap, BTreeSet};
5928 :
5929 : use bytes::{Bytes, BytesMut};
5930 : use hex_literal::hex;
5931 : use itertools::Itertools;
5932 : #[cfg(feature = "testing")]
5933 : use models::CompactLsnRange;
5934 : use pageserver_api::key::{AUX_KEY_PREFIX, Key, NON_INHERITED_RANGE, RELATION_SIZE_PREFIX};
5935 : use pageserver_api::keyspace::KeySpace;
5936 : #[cfg(feature = "testing")]
5937 : use pageserver_api::keyspace::KeySpaceRandomAccum;
5938 : use pageserver_api::models::{CompactionAlgorithm, CompactionAlgorithmSettings};
5939 : #[cfg(feature = "testing")]
5940 : use pageserver_api::record::NeonWalRecord;
5941 : use pageserver_api::value::Value;
5942 : use pageserver_compaction::helpers::overlaps_with;
5943 : #[cfg(feature = "testing")]
5944 : use rand::SeedableRng;
5945 : #[cfg(feature = "testing")]
5946 : use rand::rngs::StdRng;
5947 : use rand::{Rng, thread_rng};
5948 : #[cfg(feature = "testing")]
5949 : use std::ops::Range;
5950 : use storage_layer::{IoConcurrency, PersistentLayerKey};
5951 : use tests::storage_layer::ValuesReconstructState;
5952 : use tests::timeline::{GetVectoredError, ShutdownMode};
5953 : #[cfg(feature = "testing")]
5954 : use timeline::GcInfo;
5955 : #[cfg(feature = "testing")]
5956 : use timeline::InMemoryLayerTestDesc;
5957 : #[cfg(feature = "testing")]
5958 : use timeline::compaction::{KeyHistoryRetention, KeyLogAtLsn};
5959 : use timeline::{CompactOptions, DeltaLayerTestDesc, VersionedKeySpaceQuery};
5960 : use utils::id::TenantId;
5961 :
5962 : use super::*;
5963 : use crate::DEFAULT_PG_VERSION;
5964 : use crate::keyspace::KeySpaceAccum;
5965 : use crate::tenant::harness::*;
5966 : use crate::tenant::timeline::CompactFlags;
5967 :
5968 : static TEST_KEY: Lazy<Key> =
5969 108 : Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));
5970 :
5971 : #[cfg(feature = "testing")]
5972 : struct TestTimelineSpecification {
5973 : start_lsn: Lsn,
5974 : last_record_lsn: Lsn,
5975 :
5976 : in_memory_layers_shape: Vec<(Range<Key>, Range<Lsn>)>,
5977 : delta_layers_shape: Vec<(Range<Key>, Range<Lsn>)>,
5978 : image_layers_shape: Vec<(Range<Key>, Lsn)>,
5979 :
5980 : gap_chance: u8,
5981 : will_init_chance: u8,
5982 : }
5983 :
5984 : #[cfg(feature = "testing")]
5985 : struct Storage {
5986 : storage: HashMap<(Key, Lsn), Value>,
5987 : start_lsn: Lsn,
5988 : }
5989 :
5990 : #[cfg(feature = "testing")]
5991 : impl Storage {
5992 384000 : fn get(&self, key: Key, lsn: Lsn) -> Bytes {
5993 : use bytes::BufMut;
5994 :
5995 384000 : let mut crnt_lsn = lsn;
5996 384000 : let mut got_base = false;
5997 384000 :
5998 384000 : let mut acc = Vec::new();
5999 :
6000 33982452 : while crnt_lsn >= self.start_lsn {
6001 33982452 : if let Some(value) = self.storage.get(&(key, crnt_lsn)) {
6002 17054064 : acc.push(value.clone());
6003 :
6004 16834572 : match value {
6005 16834572 : Value::WalRecord(NeonWalRecord::Test { will_init, .. }) => {
6006 16834572 : if *will_init {
6007 164508 : got_base = true;
6008 164508 : break;
6009 16670064 : }
6010 : }
6011 : Value::Image(_) => {
6012 219492 : got_base = true;
6013 219492 : break;
6014 : }
6015 0 : _ => unreachable!(),
6016 : }
6017 16928388 : }
6018 :
6019 33598452 : crnt_lsn = crnt_lsn.checked_sub(1u64).unwrap();
6020 : }
6021 :
6022 384000 : assert!(
6023 384000 : got_base,
6024 0 : "Input data was incorrect. No base image for {key}@{lsn}"
6025 : );
6026 :
6027 384000 : tracing::debug!("Wal redo depth for {key}@{lsn} is {}", acc.len());
6028 :
6029 384000 : let mut blob = BytesMut::new();
6030 17054064 : for value in acc.into_iter().rev() {
6031 16834572 : match value {
6032 16834572 : Value::WalRecord(NeonWalRecord::Test { append, .. }) => {
6033 16834572 : blob.extend_from_slice(append.as_bytes());
6034 16834572 : }
6035 219492 : Value::Image(img) => {
6036 219492 : blob.put(img);
6037 219492 : }
6038 0 : _ => unreachable!(),
6039 : }
6040 : }
6041 :
6042 384000 : blob.into()
6043 384000 : }
6044 : }
6045 :
6046 : #[cfg(feature = "testing")]
6047 : #[allow(clippy::too_many_arguments)]
6048 12 : async fn randomize_timeline(
6049 12 : tenant: &Arc<TenantShard>,
6050 12 : new_timeline_id: TimelineId,
6051 12 : pg_version: u32,
6052 12 : spec: TestTimelineSpecification,
6053 12 : random: &mut rand::rngs::StdRng,
6054 12 : ctx: &RequestContext,
6055 12 : ) -> anyhow::Result<(Arc<Timeline>, Storage, Vec<Lsn>)> {
6056 12 : let mut storage: HashMap<(Key, Lsn), Value> = HashMap::default();
6057 12 : let mut interesting_lsns = vec![spec.last_record_lsn];
6058 :
6059 24 : for (key_range, lsn_range) in spec.in_memory_layers_shape.iter() {
6060 24 : let mut lsn = lsn_range.start;
6061 2424 : while lsn < lsn_range.end {
6062 2400 : let mut key = key_range.start;
6063 252216 : while key < key_range.end {
6064 249816 : let gap = random.gen_range(1..=100) <= spec.gap_chance;
6065 249816 : let will_init = random.gen_range(1..=100) <= spec.will_init_chance;
6066 249816 :
6067 249816 : if gap {
6068 12216 : continue;
6069 237600 : }
6070 :
6071 237600 : let record = if will_init {
6072 2292 : Value::WalRecord(NeonWalRecord::wal_init(format!("[wil_init {key}@{lsn}]")))
6073 : } else {
6074 235308 : Value::WalRecord(NeonWalRecord::wal_append(format!("[delta {key}@{lsn}]")))
6075 : };
6076 :
6077 237600 : storage.insert((key, lsn), record);
6078 237600 :
6079 237600 : key = key.next();
6080 : }
6081 2400 : lsn = Lsn(lsn.0 + 1);
6082 : }
6083 :
6084 : // Stash some interesting LSN for future use
6085 72 : for offset in [0, 5, 100].iter() {
6086 72 : if *offset == 0 {
6087 24 : interesting_lsns.push(lsn_range.start);
6088 24 : } else {
6089 48 : let below = lsn_range.start.checked_sub(*offset);
6090 48 : match below {
6091 48 : Some(v) if v >= spec.start_lsn => {
6092 48 : interesting_lsns.push(v);
6093 48 : }
6094 0 : _ => {}
6095 : }
6096 :
6097 48 : let above = Lsn(lsn_range.start.0 + offset);
6098 48 : interesting_lsns.push(above);
6099 : }
6100 : }
6101 : }
6102 :
6103 36 : for (key_range, lsn_range) in spec.delta_layers_shape.iter() {
6104 36 : let mut lsn = lsn_range.start;
6105 3780 : while lsn < lsn_range.end {
6106 3744 : let mut key = key_range.start;
6107 133344 : while key < key_range.end {
6108 129600 : let gap = random.gen_range(1..=100) <= spec.gap_chance;
6109 129600 : let will_init = random.gen_range(1..=100) <= spec.will_init_chance;
6110 129600 :
6111 129600 : if gap {
6112 6048 : continue;
6113 123552 : }
6114 :
6115 123552 : let record = if will_init {
6116 1236 : Value::WalRecord(NeonWalRecord::wal_init(format!("[wil_init {key}@{lsn}]")))
6117 : } else {
6118 122316 : Value::WalRecord(NeonWalRecord::wal_append(format!("[delta {key}@{lsn}]")))
6119 : };
6120 :
6121 123552 : storage.insert((key, lsn), record);
6122 123552 :
6123 123552 : key = key.next();
6124 : }
6125 3744 : lsn = Lsn(lsn.0 + 1);
6126 : }
6127 :
6128 : // Stash some interesting LSN for future use
6129 108 : for offset in [0, 5, 100].iter() {
6130 108 : if *offset == 0 {
6131 36 : interesting_lsns.push(lsn_range.start);
6132 36 : } else {
6133 72 : let below = lsn_range.start.checked_sub(*offset);
6134 72 : match below {
6135 72 : Some(v) if v >= spec.start_lsn => {
6136 36 : interesting_lsns.push(v);
6137 36 : }
6138 36 : _ => {}
6139 : }
6140 :
6141 72 : let above = Lsn(lsn_range.start.0 + offset);
6142 72 : interesting_lsns.push(above);
6143 : }
6144 : }
6145 : }
6146 :
6147 36 : for (key_range, lsn) in spec.image_layers_shape.iter() {
6148 36 : let mut key = key_range.start;
6149 1704 : while key < key_range.end {
6150 1668 : let blob = Bytes::from(format!("[image {key}@{lsn}]"));
6151 1668 : let record = Value::Image(blob.clone());
6152 1668 : storage.insert((key, *lsn), record);
6153 1668 :
6154 1668 : key = key.next();
6155 1668 : }
6156 :
6157 : // Stash some interesting LSN for future use
6158 108 : for offset in [0, 5, 100].iter() {
6159 108 : if *offset == 0 {
6160 36 : interesting_lsns.push(*lsn);
6161 36 : } else {
6162 72 : let below = lsn.checked_sub(*offset);
6163 72 : match below {
6164 72 : Some(v) if v >= spec.start_lsn => {
6165 48 : interesting_lsns.push(v);
6166 48 : }
6167 24 : _ => {}
6168 : }
6169 :
6170 72 : let above = Lsn(lsn.0 + offset);
6171 72 : interesting_lsns.push(above);
6172 : }
6173 : }
6174 : }
6175 :
6176 12 : let in_memory_test_layers = {
6177 12 : let mut acc = Vec::new();
6178 :
6179 24 : for (key_range, lsn_range) in spec.in_memory_layers_shape.iter() {
6180 24 : let mut data = Vec::new();
6181 24 :
6182 24 : let mut lsn = lsn_range.start;
6183 2424 : while lsn < lsn_range.end {
6184 2400 : let mut key = key_range.start;
6185 240000 : while key < key_range.end {
6186 237600 : if let Some(record) = storage.get(&(key, lsn)) {
6187 237600 : data.push((key, lsn, record.clone()));
6188 237600 : }
6189 :
6190 237600 : key = key.next();
6191 : }
6192 2400 : lsn = Lsn(lsn.0 + 1);
6193 : }
6194 :
6195 24 : acc.push(InMemoryLayerTestDesc {
6196 24 : data,
6197 24 : lsn_range: lsn_range.clone(),
6198 24 : is_open: false,
6199 24 : })
6200 : }
6201 :
6202 12 : acc
6203 : };
6204 :
6205 12 : let delta_test_layers = {
6206 12 : let mut acc = Vec::new();
6207 :
6208 36 : for (key_range, lsn_range) in spec.delta_layers_shape.iter() {
6209 36 : let mut data = Vec::new();
6210 36 :
6211 36 : let mut lsn = lsn_range.start;
6212 3780 : while lsn < lsn_range.end {
6213 3744 : let mut key = key_range.start;
6214 127296 : while key < key_range.end {
6215 123552 : if let Some(record) = storage.get(&(key, lsn)) {
6216 123552 : data.push((key, lsn, record.clone()));
6217 123552 : }
6218 :
6219 123552 : key = key.next();
6220 : }
6221 3744 : lsn = Lsn(lsn.0 + 1);
6222 : }
6223 :
6224 36 : acc.push(DeltaLayerTestDesc {
6225 36 : data,
6226 36 : lsn_range: lsn_range.clone(),
6227 36 : key_range: key_range.clone(),
6228 36 : })
6229 : }
6230 :
6231 12 : acc
6232 : };
6233 :
6234 12 : let image_test_layers = {
6235 12 : let mut acc = Vec::new();
6236 :
6237 36 : for (key_range, lsn) in spec.image_layers_shape.iter() {
6238 36 : let mut data = Vec::new();
6239 36 :
6240 36 : let mut key = key_range.start;
6241 1704 : while key < key_range.end {
6242 1668 : if let Some(record) = storage.get(&(key, *lsn)) {
6243 1668 : let blob = match record {
6244 1668 : Value::Image(blob) => blob.clone(),
6245 0 : _ => unreachable!(),
6246 : };
6247 :
6248 1668 : data.push((key, blob));
6249 0 : }
6250 :
6251 1668 : key = key.next();
6252 : }
6253 :
6254 36 : acc.push((*lsn, data));
6255 : }
6256 :
6257 12 : acc
6258 : };
6259 :
6260 12 : let tline = tenant
6261 12 : .create_test_timeline_with_layers(
6262 12 : new_timeline_id,
6263 12 : spec.start_lsn,
6264 12 : pg_version,
6265 12 : ctx,
6266 12 : in_memory_test_layers,
6267 12 : delta_test_layers,
6268 12 : image_test_layers,
6269 12 : spec.last_record_lsn,
6270 12 : )
6271 12 : .await?;
6272 :
6273 12 : Ok((
6274 12 : tline,
6275 12 : Storage {
6276 12 : storage,
6277 12 : start_lsn: spec.start_lsn,
6278 12 : },
6279 12 : interesting_lsns,
6280 12 : ))
6281 12 : }
6282 :
6283 : #[tokio::test]
6284 12 : async fn test_basic() -> anyhow::Result<()> {
6285 12 : let (tenant, ctx) = TenantHarness::create("test_basic").await?.load().await;
6286 12 : let tline = tenant
6287 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6288 12 : .await?;
6289 12 :
6290 12 : let mut writer = tline.writer().await;
6291 12 : writer
6292 12 : .put(
6293 12 : *TEST_KEY,
6294 12 : Lsn(0x10),
6295 12 : &Value::Image(test_img("foo at 0x10")),
6296 12 : &ctx,
6297 12 : )
6298 12 : .await?;
6299 12 : writer.finish_write(Lsn(0x10));
6300 12 : drop(writer);
6301 12 :
6302 12 : let mut writer = tline.writer().await;
6303 12 : writer
6304 12 : .put(
6305 12 : *TEST_KEY,
6306 12 : Lsn(0x20),
6307 12 : &Value::Image(test_img("foo at 0x20")),
6308 12 : &ctx,
6309 12 : )
6310 12 : .await?;
6311 12 : writer.finish_write(Lsn(0x20));
6312 12 : drop(writer);
6313 12 :
6314 12 : assert_eq!(
6315 12 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
6316 12 : test_img("foo at 0x10")
6317 12 : );
6318 12 : assert_eq!(
6319 12 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
6320 12 : test_img("foo at 0x10")
6321 12 : );
6322 12 : assert_eq!(
6323 12 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
6324 12 : test_img("foo at 0x20")
6325 12 : );
6326 12 :
6327 12 : Ok(())
6328 12 : }
6329 :
6330 : #[tokio::test]
6331 12 : async fn no_duplicate_timelines() -> anyhow::Result<()> {
6332 12 : let (tenant, ctx) = TenantHarness::create("no_duplicate_timelines")
6333 12 : .await?
6334 12 : .load()
6335 12 : .await;
6336 12 : let _ = tenant
6337 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6338 12 : .await?;
6339 12 :
6340 12 : match tenant
6341 12 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6342 12 : .await
6343 12 : {
6344 12 : Ok(_) => panic!("duplicate timeline creation should fail"),
6345 12 : Err(e) => assert_eq!(
6346 12 : e.to_string(),
6347 12 : "timeline already exists with different parameters".to_string()
6348 12 : ),
6349 12 : }
6350 12 :
6351 12 : Ok(())
6352 12 : }
6353 :
6354 : /// Convenience function to create a page image with given string as the only content
6355 60 : pub fn test_value(s: &str) -> Value {
6356 60 : let mut buf = BytesMut::new();
6357 60 : buf.extend_from_slice(s.as_bytes());
6358 60 : Value::Image(buf.freeze())
6359 60 : }
6360 :
6361 : ///
6362 : /// Test branch creation
6363 : ///
6364 : #[tokio::test]
6365 12 : async fn test_branch() -> anyhow::Result<()> {
6366 12 : use std::str::from_utf8;
6367 12 :
6368 12 : let (tenant, ctx) = TenantHarness::create("test_branch").await?.load().await;
6369 12 : let tline = tenant
6370 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6371 12 : .await?;
6372 12 : let mut writer = tline.writer().await;
6373 12 :
6374 12 : #[allow(non_snake_case)]
6375 12 : let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
6376 12 : #[allow(non_snake_case)]
6377 12 : let TEST_KEY_B: Key = Key::from_hex("110000000033333333444444445500000002").unwrap();
6378 12 :
6379 12 : // Insert a value on the timeline
6380 12 : writer
6381 12 : .put(TEST_KEY_A, Lsn(0x20), &test_value("foo at 0x20"), &ctx)
6382 12 : .await?;
6383 12 : writer
6384 12 : .put(TEST_KEY_B, Lsn(0x20), &test_value("foobar at 0x20"), &ctx)
6385 12 : .await?;
6386 12 : writer.finish_write(Lsn(0x20));
6387 12 :
6388 12 : writer
6389 12 : .put(TEST_KEY_A, Lsn(0x30), &test_value("foo at 0x30"), &ctx)
6390 12 : .await?;
6391 12 : writer.finish_write(Lsn(0x30));
6392 12 : writer
6393 12 : .put(TEST_KEY_A, Lsn(0x40), &test_value("foo at 0x40"), &ctx)
6394 12 : .await?;
6395 12 : writer.finish_write(Lsn(0x40));
6396 12 :
6397 12 : //assert_current_logical_size(&tline, Lsn(0x40));
6398 12 :
6399 12 : // Branch the history, modify relation differently on the new timeline
6400 12 : tenant
6401 12 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x30)), &ctx)
6402 12 : .await?;
6403 12 : let newtline = tenant
6404 12 : .get_timeline(NEW_TIMELINE_ID, true)
6405 12 : .expect("Should have a local timeline");
6406 12 : let mut new_writer = newtline.writer().await;
6407 12 : new_writer
6408 12 : .put(TEST_KEY_A, Lsn(0x40), &test_value("bar at 0x40"), &ctx)
6409 12 : .await?;
6410 12 : new_writer.finish_write(Lsn(0x40));
6411 12 :
6412 12 : // Check page contents on both branches
6413 12 : assert_eq!(
6414 12 : from_utf8(&tline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
6415 12 : "foo at 0x40"
6416 12 : );
6417 12 : assert_eq!(
6418 12 : from_utf8(&newtline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
6419 12 : "bar at 0x40"
6420 12 : );
6421 12 : assert_eq!(
6422 12 : from_utf8(&newtline.get(TEST_KEY_B, Lsn(0x40), &ctx).await?)?,
6423 12 : "foobar at 0x20"
6424 12 : );
6425 12 :
6426 12 : //assert_current_logical_size(&tline, Lsn(0x40));
6427 12 :
6428 12 : Ok(())
6429 12 : }
6430 :
6431 120 : async fn make_some_layers(
6432 120 : tline: &Timeline,
6433 120 : start_lsn: Lsn,
6434 120 : ctx: &RequestContext,
6435 120 : ) -> anyhow::Result<()> {
6436 120 : let mut lsn = start_lsn;
6437 : {
6438 120 : let mut writer = tline.writer().await;
6439 : // Create a relation on the timeline
6440 120 : writer
6441 120 : .put(
6442 120 : *TEST_KEY,
6443 120 : lsn,
6444 120 : &Value::Image(test_img(&format!("foo at {}", lsn))),
6445 120 : ctx,
6446 120 : )
6447 120 : .await?;
6448 120 : writer.finish_write(lsn);
6449 120 : lsn += 0x10;
6450 120 : writer
6451 120 : .put(
6452 120 : *TEST_KEY,
6453 120 : lsn,
6454 120 : &Value::Image(test_img(&format!("foo at {}", lsn))),
6455 120 : ctx,
6456 120 : )
6457 120 : .await?;
6458 120 : writer.finish_write(lsn);
6459 120 : lsn += 0x10;
6460 120 : }
6461 120 : tline.freeze_and_flush().await?;
6462 : {
6463 120 : let mut writer = tline.writer().await;
6464 120 : writer
6465 120 : .put(
6466 120 : *TEST_KEY,
6467 120 : lsn,
6468 120 : &Value::Image(test_img(&format!("foo at {}", lsn))),
6469 120 : ctx,
6470 120 : )
6471 120 : .await?;
6472 120 : writer.finish_write(lsn);
6473 120 : lsn += 0x10;
6474 120 : writer
6475 120 : .put(
6476 120 : *TEST_KEY,
6477 120 : lsn,
6478 120 : &Value::Image(test_img(&format!("foo at {}", lsn))),
6479 120 : ctx,
6480 120 : )
6481 120 : .await?;
6482 120 : writer.finish_write(lsn);
6483 120 : }
6484 120 : tline.freeze_and_flush().await.map_err(|e| e.into())
6485 120 : }
6486 :
6487 : #[tokio::test(start_paused = true)]
6488 12 : async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
6489 12 : let (tenant, ctx) =
6490 12 : TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
6491 12 : .await?
6492 12 : .load()
6493 12 : .await;
6494 12 : // Advance to the lsn lease deadline so that GC is not blocked by
6495 12 : // initial transition into AttachedSingle.
6496 12 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
6497 12 : tokio::time::resume();
6498 12 : let tline = tenant
6499 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6500 12 : .await?;
6501 12 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6502 12 :
6503 12 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
6504 12 : // FIXME: this doesn't actually remove any layer currently, given how the flushing
6505 12 : // and compaction works. But it does set the 'cutoff' point so that the cross check
6506 12 : // below should fail.
6507 12 : tenant
6508 12 : .gc_iteration(
6509 12 : Some(TIMELINE_ID),
6510 12 : 0x10,
6511 12 : Duration::ZERO,
6512 12 : &CancellationToken::new(),
6513 12 : &ctx,
6514 12 : )
6515 12 : .await?;
6516 12 :
6517 12 : // try to branch at lsn 25, should fail because we already garbage collected the data
6518 12 : match tenant
6519 12 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
6520 12 : .await
6521 12 : {
6522 12 : Ok(_) => panic!("branching should have failed"),
6523 12 : Err(err) => {
6524 12 : let CreateTimelineError::AncestorLsn(err) = err else {
6525 12 : panic!("wrong error type")
6526 12 : };
6527 12 : assert!(err.to_string().contains("invalid branch start lsn"));
6528 12 : assert!(
6529 12 : err.source()
6530 12 : .unwrap()
6531 12 : .to_string()
6532 12 : .contains("we might've already garbage collected needed data")
6533 12 : )
6534 12 : }
6535 12 : }
6536 12 :
6537 12 : Ok(())
6538 12 : }
6539 :
6540 : #[tokio::test]
6541 12 : async fn test_prohibit_branch_creation_on_pre_initdb_lsn() -> anyhow::Result<()> {
6542 12 : let (tenant, ctx) =
6543 12 : TenantHarness::create("test_prohibit_branch_creation_on_pre_initdb_lsn")
6544 12 : .await?
6545 12 : .load()
6546 12 : .await;
6547 12 :
6548 12 : let tline = tenant
6549 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x50), DEFAULT_PG_VERSION, &ctx)
6550 12 : .await?;
6551 12 : // try to branch at lsn 0x25, should fail because initdb lsn is 0x50
6552 12 : match tenant
6553 12 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
6554 12 : .await
6555 12 : {
6556 12 : Ok(_) => panic!("branching should have failed"),
6557 12 : Err(err) => {
6558 12 : let CreateTimelineError::AncestorLsn(err) = err else {
6559 12 : panic!("wrong error type");
6560 12 : };
6561 12 : assert!(&err.to_string().contains("invalid branch start lsn"));
6562 12 : assert!(
6563 12 : &err.source()
6564 12 : .unwrap()
6565 12 : .to_string()
6566 12 : .contains("is earlier than latest GC cutoff")
6567 12 : );
6568 12 : }
6569 12 : }
6570 12 :
6571 12 : Ok(())
6572 12 : }
6573 :
6574 : /*
6575 : // FIXME: This currently fails to error out. Calling GC doesn't currently
6576 : // remove the old value, we'd need to work a little harder
6577 : #[tokio::test]
6578 : async fn test_prohibit_get_for_garbage_collected_data() -> anyhow::Result<()> {
6579 : let repo =
6580 : RepoHarness::create("test_prohibit_get_for_garbage_collected_data")?
6581 : .load();
6582 :
6583 : let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION)?;
6584 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6585 :
6586 : repo.gc_iteration(Some(TIMELINE_ID), 0x10, Duration::ZERO)?;
6587 : let applied_gc_cutoff_lsn = tline.get_applied_gc_cutoff_lsn();
6588 : assert!(*applied_gc_cutoff_lsn > Lsn(0x25));
6589 : match tline.get(*TEST_KEY, Lsn(0x25)) {
6590 : Ok(_) => panic!("request for page should have failed"),
6591 : Err(err) => assert!(err.to_string().contains("not found at")),
6592 : }
6593 : Ok(())
6594 : }
6595 : */
6596 :
6597 : #[tokio::test]
6598 12 : async fn test_get_branchpoints_from_an_inactive_timeline() -> anyhow::Result<()> {
6599 12 : let (tenant, ctx) =
6600 12 : TenantHarness::create("test_get_branchpoints_from_an_inactive_timeline")
6601 12 : .await?
6602 12 : .load()
6603 12 : .await;
6604 12 : let tline = tenant
6605 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6606 12 : .await?;
6607 12 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6608 12 :
6609 12 : tenant
6610 12 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6611 12 : .await?;
6612 12 : let newtline = tenant
6613 12 : .get_timeline(NEW_TIMELINE_ID, true)
6614 12 : .expect("Should have a local timeline");
6615 12 :
6616 12 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6617 12 :
6618 12 : tline.set_broken("test".to_owned());
6619 12 :
6620 12 : tenant
6621 12 : .gc_iteration(
6622 12 : Some(TIMELINE_ID),
6623 12 : 0x10,
6624 12 : Duration::ZERO,
6625 12 : &CancellationToken::new(),
6626 12 : &ctx,
6627 12 : )
6628 12 : .await?;
6629 12 :
6630 12 : // The branchpoints should contain all timelines, even ones marked
6631 12 : // as Broken.
6632 12 : {
6633 12 : let branchpoints = &tline.gc_info.read().unwrap().retain_lsns;
6634 12 : assert_eq!(branchpoints.len(), 1);
6635 12 : assert_eq!(
6636 12 : branchpoints[0],
6637 12 : (Lsn(0x40), NEW_TIMELINE_ID, MaybeOffloaded::No)
6638 12 : );
6639 12 : }
6640 12 :
6641 12 : // You can read the key from the child branch even though the parent is
6642 12 : // Broken, as long as you don't need to access data from the parent.
6643 12 : assert_eq!(
6644 12 : newtline.get(*TEST_KEY, Lsn(0x70), &ctx).await?,
6645 12 : test_img(&format!("foo at {}", Lsn(0x70)))
6646 12 : );
6647 12 :
6648 12 : // This needs to traverse to the parent, and fails.
6649 12 : let err = newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await.unwrap_err();
6650 12 : assert!(
6651 12 : err.to_string().starts_with(&format!(
6652 12 : "bad state on timeline {}: Broken",
6653 12 : tline.timeline_id
6654 12 : )),
6655 12 : "{err}"
6656 12 : );
6657 12 :
6658 12 : Ok(())
6659 12 : }
6660 :
6661 : #[tokio::test]
6662 12 : async fn test_retain_data_in_parent_which_is_needed_for_child() -> anyhow::Result<()> {
6663 12 : let (tenant, ctx) =
6664 12 : TenantHarness::create("test_retain_data_in_parent_which_is_needed_for_child")
6665 12 : .await?
6666 12 : .load()
6667 12 : .await;
6668 12 : let tline = tenant
6669 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6670 12 : .await?;
6671 12 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6672 12 :
6673 12 : tenant
6674 12 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6675 12 : .await?;
6676 12 : let newtline = tenant
6677 12 : .get_timeline(NEW_TIMELINE_ID, true)
6678 12 : .expect("Should have a local timeline");
6679 12 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
6680 12 : tenant
6681 12 : .gc_iteration(
6682 12 : Some(TIMELINE_ID),
6683 12 : 0x10,
6684 12 : Duration::ZERO,
6685 12 : &CancellationToken::new(),
6686 12 : &ctx,
6687 12 : )
6688 12 : .await?;
6689 12 : assert!(newtline.get(*TEST_KEY, Lsn(0x25), &ctx).await.is_ok());
6690 12 :
6691 12 : Ok(())
6692 12 : }
6693 : #[tokio::test]
6694 12 : async fn test_parent_keeps_data_forever_after_branching() -> anyhow::Result<()> {
6695 12 : let (tenant, ctx) = TenantHarness::create("test_parent_keeps_data_forever_after_branching")
6696 12 : .await?
6697 12 : .load()
6698 12 : .await;
6699 12 : let tline = tenant
6700 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6701 12 : .await?;
6702 12 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6703 12 :
6704 12 : tenant
6705 12 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6706 12 : .await?;
6707 12 : let newtline = tenant
6708 12 : .get_timeline(NEW_TIMELINE_ID, true)
6709 12 : .expect("Should have a local timeline");
6710 12 :
6711 12 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6712 12 :
6713 12 : // run gc on parent
6714 12 : tenant
6715 12 : .gc_iteration(
6716 12 : Some(TIMELINE_ID),
6717 12 : 0x10,
6718 12 : Duration::ZERO,
6719 12 : &CancellationToken::new(),
6720 12 : &ctx,
6721 12 : )
6722 12 : .await?;
6723 12 :
6724 12 : // Check that the data is still accessible on the branch.
6725 12 : assert_eq!(
6726 12 : newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
6727 12 : test_img(&format!("foo at {}", Lsn(0x40)))
6728 12 : );
6729 12 :
6730 12 : Ok(())
6731 12 : }
6732 :
6733 : #[tokio::test]
6734 12 : async fn timeline_load() -> anyhow::Result<()> {
6735 12 : const TEST_NAME: &str = "timeline_load";
6736 12 : let harness = TenantHarness::create(TEST_NAME).await?;
6737 12 : {
6738 12 : let (tenant, ctx) = harness.load().await;
6739 12 : let tline = tenant
6740 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x7000), DEFAULT_PG_VERSION, &ctx)
6741 12 : .await?;
6742 12 : make_some_layers(tline.as_ref(), Lsn(0x8000), &ctx).await?;
6743 12 : // so that all uploads finish & we can call harness.load() below again
6744 12 : tenant
6745 12 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
6746 12 : .instrument(harness.span())
6747 12 : .await
6748 12 : .ok()
6749 12 : .unwrap();
6750 12 : }
6751 12 :
6752 12 : let (tenant, _ctx) = harness.load().await;
6753 12 : tenant
6754 12 : .get_timeline(TIMELINE_ID, true)
6755 12 : .expect("cannot load timeline");
6756 12 :
6757 12 : Ok(())
6758 12 : }
6759 :
6760 : #[tokio::test]
6761 12 : async fn timeline_load_with_ancestor() -> anyhow::Result<()> {
6762 12 : const TEST_NAME: &str = "timeline_load_with_ancestor";
6763 12 : let harness = TenantHarness::create(TEST_NAME).await?;
6764 12 : // create two timelines
6765 12 : {
6766 12 : let (tenant, ctx) = harness.load().await;
6767 12 : let tline = tenant
6768 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6769 12 : .await?;
6770 12 :
6771 12 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6772 12 :
6773 12 : let child_tline = tenant
6774 12 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6775 12 : .await?;
6776 12 : child_tline.set_state(TimelineState::Active);
6777 12 :
6778 12 : let newtline = tenant
6779 12 : .get_timeline(NEW_TIMELINE_ID, true)
6780 12 : .expect("Should have a local timeline");
6781 12 :
6782 12 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6783 12 :
6784 12 : // so that all uploads finish & we can call harness.load() below again
6785 12 : tenant
6786 12 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
6787 12 : .instrument(harness.span())
6788 12 : .await
6789 12 : .ok()
6790 12 : .unwrap();
6791 12 : }
6792 12 :
6793 12 : // check that both of them are initially unloaded
6794 12 : let (tenant, _ctx) = harness.load().await;
6795 12 :
6796 12 : // check that both, child and ancestor are loaded
6797 12 : let _child_tline = tenant
6798 12 : .get_timeline(NEW_TIMELINE_ID, true)
6799 12 : .expect("cannot get child timeline loaded");
6800 12 :
6801 12 : let _ancestor_tline = tenant
6802 12 : .get_timeline(TIMELINE_ID, true)
6803 12 : .expect("cannot get ancestor timeline loaded");
6804 12 :
6805 12 : Ok(())
6806 12 : }
6807 :
6808 : #[tokio::test]
6809 12 : async fn delta_layer_dumping() -> anyhow::Result<()> {
6810 12 : use storage_layer::AsLayerDesc;
6811 12 : let (tenant, ctx) = TenantHarness::create("test_layer_dumping")
6812 12 : .await?
6813 12 : .load()
6814 12 : .await;
6815 12 : let tline = tenant
6816 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6817 12 : .await?;
6818 12 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6819 12 :
6820 12 : let layer_map = tline.layers.read().await;
6821 12 : let level0_deltas = layer_map
6822 12 : .layer_map()?
6823 12 : .level0_deltas()
6824 12 : .iter()
6825 24 : .map(|desc| layer_map.get_from_desc(desc))
6826 12 : .collect::<Vec<_>>();
6827 12 :
6828 12 : assert!(!level0_deltas.is_empty());
6829 12 :
6830 36 : for delta in level0_deltas {
6831 12 : // Ensure we are dumping a delta layer here
6832 24 : assert!(delta.layer_desc().is_delta);
6833 24 : delta.dump(true, &ctx).await.unwrap();
6834 12 : }
6835 12 :
6836 12 : Ok(())
6837 12 : }
6838 :
6839 : #[tokio::test]
6840 12 : async fn test_images() -> anyhow::Result<()> {
6841 12 : let (tenant, ctx) = TenantHarness::create("test_images").await?.load().await;
6842 12 : let tline = tenant
6843 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6844 12 : .await?;
6845 12 :
6846 12 : let mut writer = tline.writer().await;
6847 12 : writer
6848 12 : .put(
6849 12 : *TEST_KEY,
6850 12 : Lsn(0x10),
6851 12 : &Value::Image(test_img("foo at 0x10")),
6852 12 : &ctx,
6853 12 : )
6854 12 : .await?;
6855 12 : writer.finish_write(Lsn(0x10));
6856 12 : drop(writer);
6857 12 :
6858 12 : tline.freeze_and_flush().await?;
6859 12 : tline
6860 12 : .compact(&CancellationToken::new(), EnumSet::default(), &ctx)
6861 12 : .await?;
6862 12 :
6863 12 : let mut writer = tline.writer().await;
6864 12 : writer
6865 12 : .put(
6866 12 : *TEST_KEY,
6867 12 : Lsn(0x20),
6868 12 : &Value::Image(test_img("foo at 0x20")),
6869 12 : &ctx,
6870 12 : )
6871 12 : .await?;
6872 12 : writer.finish_write(Lsn(0x20));
6873 12 : drop(writer);
6874 12 :
6875 12 : tline.freeze_and_flush().await?;
6876 12 : tline
6877 12 : .compact(&CancellationToken::new(), EnumSet::default(), &ctx)
6878 12 : .await?;
6879 12 :
6880 12 : let mut writer = tline.writer().await;
6881 12 : writer
6882 12 : .put(
6883 12 : *TEST_KEY,
6884 12 : Lsn(0x30),
6885 12 : &Value::Image(test_img("foo at 0x30")),
6886 12 : &ctx,
6887 12 : )
6888 12 : .await?;
6889 12 : writer.finish_write(Lsn(0x30));
6890 12 : drop(writer);
6891 12 :
6892 12 : tline.freeze_and_flush().await?;
6893 12 : tline
6894 12 : .compact(&CancellationToken::new(), EnumSet::default(), &ctx)
6895 12 : .await?;
6896 12 :
6897 12 : let mut writer = tline.writer().await;
6898 12 : writer
6899 12 : .put(
6900 12 : *TEST_KEY,
6901 12 : Lsn(0x40),
6902 12 : &Value::Image(test_img("foo at 0x40")),
6903 12 : &ctx,
6904 12 : )
6905 12 : .await?;
6906 12 : writer.finish_write(Lsn(0x40));
6907 12 : drop(writer);
6908 12 :
6909 12 : tline.freeze_and_flush().await?;
6910 12 : tline
6911 12 : .compact(&CancellationToken::new(), EnumSet::default(), &ctx)
6912 12 : .await?;
6913 12 :
6914 12 : assert_eq!(
6915 12 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
6916 12 : test_img("foo at 0x10")
6917 12 : );
6918 12 : assert_eq!(
6919 12 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
6920 12 : test_img("foo at 0x10")
6921 12 : );
6922 12 : assert_eq!(
6923 12 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
6924 12 : test_img("foo at 0x20")
6925 12 : );
6926 12 : assert_eq!(
6927 12 : tline.get(*TEST_KEY, Lsn(0x30), &ctx).await?,
6928 12 : test_img("foo at 0x30")
6929 12 : );
6930 12 : assert_eq!(
6931 12 : tline.get(*TEST_KEY, Lsn(0x40), &ctx).await?,
6932 12 : test_img("foo at 0x40")
6933 12 : );
6934 12 :
6935 12 : Ok(())
6936 12 : }
6937 :
6938 24 : async fn bulk_insert_compact_gc(
6939 24 : tenant: &TenantShard,
6940 24 : timeline: &Arc<Timeline>,
6941 24 : ctx: &RequestContext,
6942 24 : lsn: Lsn,
6943 24 : repeat: usize,
6944 24 : key_count: usize,
6945 24 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
6946 24 : let compact = true;
6947 24 : bulk_insert_maybe_compact_gc(tenant, timeline, ctx, lsn, repeat, key_count, compact).await
6948 24 : }
6949 :
6950 48 : async fn bulk_insert_maybe_compact_gc(
6951 48 : tenant: &TenantShard,
6952 48 : timeline: &Arc<Timeline>,
6953 48 : ctx: &RequestContext,
6954 48 : mut lsn: Lsn,
6955 48 : repeat: usize,
6956 48 : key_count: usize,
6957 48 : compact: bool,
6958 48 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
6959 48 : let mut inserted: HashMap<Key, BTreeSet<Lsn>> = Default::default();
6960 48 :
6961 48 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6962 48 : let mut blknum = 0;
6963 48 :
6964 48 : // Enforce that key range is monotonously increasing
6965 48 : let mut keyspace = KeySpaceAccum::new();
6966 48 :
6967 48 : let cancel = CancellationToken::new();
6968 48 :
6969 48 : for _ in 0..repeat {
6970 2400 : for _ in 0..key_count {
6971 24000000 : test_key.field6 = blknum;
6972 24000000 : let mut writer = timeline.writer().await;
6973 24000000 : writer
6974 24000000 : .put(
6975 24000000 : test_key,
6976 24000000 : lsn,
6977 24000000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6978 24000000 : ctx,
6979 24000000 : )
6980 24000000 : .await?;
6981 24000000 : inserted.entry(test_key).or_default().insert(lsn);
6982 24000000 : writer.finish_write(lsn);
6983 24000000 : drop(writer);
6984 24000000 :
6985 24000000 : keyspace.add_key(test_key);
6986 24000000 :
6987 24000000 : lsn = Lsn(lsn.0 + 0x10);
6988 24000000 : blknum += 1;
6989 : }
6990 :
6991 2400 : timeline.freeze_and_flush().await?;
6992 2400 : if compact {
6993 : // this requires timeline to be &Arc<Timeline>
6994 1200 : timeline.compact(&cancel, EnumSet::default(), ctx).await?;
6995 1200 : }
6996 :
6997 : // this doesn't really need to use the timeline_id target, but it is closer to what it
6998 : // originally was.
6999 2400 : let res = tenant
7000 2400 : .gc_iteration(Some(timeline.timeline_id), 0, Duration::ZERO, &cancel, ctx)
7001 2400 : .await?;
7002 :
7003 2400 : assert_eq!(res.layers_removed, 0, "this never removes anything");
7004 : }
7005 :
7006 48 : Ok(inserted)
7007 48 : }
7008 :
7009 : //
7010 : // Insert 1000 key-value pairs with increasing keys, flush, compact, GC.
7011 : // Repeat 50 times.
7012 : //
7013 : #[tokio::test]
7014 12 : async fn test_bulk_insert() -> anyhow::Result<()> {
7015 12 : let harness = TenantHarness::create("test_bulk_insert").await?;
7016 12 : let (tenant, ctx) = harness.load().await;
7017 12 : let tline = tenant
7018 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
7019 12 : .await?;
7020 12 :
7021 12 : let lsn = Lsn(0x10);
7022 12 : bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
7023 12 :
7024 12 : Ok(())
7025 12 : }
7026 :
7027 : // Test the vectored get real implementation against a simple sequential implementation.
7028 : //
7029 : // The test generates a keyspace by repeatedly flushing the in-memory layer and compacting.
7030 : // Projected to 2D the key space looks like below. Lsn grows upwards on the Y axis and keys
7031 : // grow to the right on the X axis.
7032 : // [Delta]
7033 : // [Delta]
7034 : // [Delta]
7035 : // [Delta]
7036 : // ------------ Image ---------------
7037 : //
7038 : // After layer generation we pick the ranges to query as follows:
7039 : // 1. The beginning of each delta layer
7040 : // 2. At the seam between two adjacent delta layers
7041 : //
7042 : // There's one major downside to this test: delta layers only contains images,
7043 : // so the search can stop at the first delta layer and doesn't traverse any deeper.
7044 : #[tokio::test]
7045 12 : async fn test_get_vectored() -> anyhow::Result<()> {
7046 12 : let harness = TenantHarness::create("test_get_vectored").await?;
7047 12 : let (tenant, ctx) = harness.load().await;
7048 12 : let io_concurrency = IoConcurrency::spawn_for_test();
7049 12 : let tline = tenant
7050 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
7051 12 : .await?;
7052 12 :
7053 12 : let lsn = Lsn(0x10);
7054 12 : let inserted = bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
7055 12 :
7056 12 : let guard = tline.layers.read().await;
7057 12 : let lm = guard.layer_map()?;
7058 12 :
7059 12 : lm.dump(true, &ctx).await?;
7060 12 :
7061 12 : let mut reads = Vec::new();
7062 12 : let mut prev = None;
7063 72 : lm.iter_historic_layers().for_each(|desc| {
7064 72 : if !desc.is_delta() {
7065 12 : prev = Some(desc.clone());
7066 12 : return;
7067 60 : }
7068 60 :
7069 60 : let start = desc.key_range.start;
7070 60 : let end = desc
7071 60 : .key_range
7072 60 : .start
7073 60 : .add(Timeline::MAX_GET_VECTORED_KEYS.try_into().unwrap());
7074 60 : reads.push(KeySpace {
7075 60 : ranges: vec![start..end],
7076 60 : });
7077 12 :
7078 60 : if let Some(prev) = &prev {
7079 60 : if !prev.is_delta() {
7080 60 : return;
7081 12 : }
7082 0 :
7083 0 : let first_range = Key {
7084 0 : field6: prev.key_range.end.field6 - 4,
7085 0 : ..prev.key_range.end
7086 0 : }..prev.key_range.end;
7087 0 :
7088 0 : let second_range = desc.key_range.start..Key {
7089 0 : field6: desc.key_range.start.field6 + 4,
7090 0 : ..desc.key_range.start
7091 0 : };
7092 0 :
7093 0 : reads.push(KeySpace {
7094 0 : ranges: vec![first_range, second_range],
7095 0 : });
7096 12 : };
7097 12 :
7098 12 : prev = Some(desc.clone());
7099 72 : });
7100 12 :
7101 12 : drop(guard);
7102 12 :
7103 12 : // Pick a big LSN such that we query over all the changes.
7104 12 : let reads_lsn = Lsn(u64::MAX - 1);
7105 12 :
7106 72 : for read in reads {
7107 60 : info!("Doing vectored read on {:?}", read);
7108 12 :
7109 60 : let query = VersionedKeySpaceQuery::uniform(read.clone(), reads_lsn);
7110 12 :
7111 60 : let vectored_res = tline
7112 60 : .get_vectored_impl(
7113 60 : query,
7114 60 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7115 60 : &ctx,
7116 60 : )
7117 60 : .await;
7118 12 :
7119 60 : let mut expected_lsns: HashMap<Key, Lsn> = Default::default();
7120 60 : let mut expect_missing = false;
7121 60 : let mut key = read.start().unwrap();
7122 1980 : while key != read.end().unwrap() {
7123 1920 : if let Some(lsns) = inserted.get(&key) {
7124 1920 : let expected_lsn = lsns.iter().rfind(|lsn| **lsn <= reads_lsn);
7125 1920 : match expected_lsn {
7126 1920 : Some(lsn) => {
7127 1920 : expected_lsns.insert(key, *lsn);
7128 1920 : }
7129 12 : None => {
7130 12 : expect_missing = true;
7131 0 : break;
7132 12 : }
7133 12 : }
7134 12 : } else {
7135 12 : expect_missing = true;
7136 0 : break;
7137 12 : }
7138 12 :
7139 1920 : key = key.next();
7140 12 : }
7141 12 :
7142 60 : if expect_missing {
7143 12 : assert!(matches!(vectored_res, Err(GetVectoredError::MissingKey(_))));
7144 12 : } else {
7145 1920 : for (key, image) in vectored_res? {
7146 1920 : let expected_lsn = expected_lsns.get(&key).expect("determined above");
7147 1920 : let expected_image = test_img(&format!("{} at {}", key.field6, expected_lsn));
7148 1920 : assert_eq!(image?, expected_image);
7149 12 : }
7150 12 : }
7151 12 : }
7152 12 :
7153 12 : Ok(())
7154 12 : }
7155 :
7156 : #[tokio::test]
7157 12 : async fn test_get_vectored_aux_files() -> anyhow::Result<()> {
7158 12 : let harness = TenantHarness::create("test_get_vectored_aux_files").await?;
7159 12 :
7160 12 : let (tenant, ctx) = harness.load().await;
7161 12 : let io_concurrency = IoConcurrency::spawn_for_test();
7162 12 : let (tline, ctx) = tenant
7163 12 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
7164 12 : .await?;
7165 12 : let tline = tline.raw_timeline().unwrap();
7166 12 :
7167 12 : let mut modification = tline.begin_modification(Lsn(0x1000));
7168 12 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
7169 12 : modification.set_lsn(Lsn(0x1008))?;
7170 12 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
7171 12 : modification.commit(&ctx).await?;
7172 12 :
7173 12 : let child_timeline_id = TimelineId::generate();
7174 12 : tenant
7175 12 : .branch_timeline_test(
7176 12 : tline,
7177 12 : child_timeline_id,
7178 12 : Some(tline.get_last_record_lsn()),
7179 12 : &ctx,
7180 12 : )
7181 12 : .await?;
7182 12 :
7183 12 : let child_timeline = tenant
7184 12 : .get_timeline(child_timeline_id, true)
7185 12 : .expect("Should have the branched timeline");
7186 12 :
7187 12 : let aux_keyspace = KeySpace {
7188 12 : ranges: vec![NON_INHERITED_RANGE],
7189 12 : };
7190 12 : let read_lsn = child_timeline.get_last_record_lsn();
7191 12 :
7192 12 : let query = VersionedKeySpaceQuery::uniform(aux_keyspace.clone(), read_lsn);
7193 12 :
7194 12 : let vectored_res = child_timeline
7195 12 : .get_vectored_impl(
7196 12 : query,
7197 12 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7198 12 : &ctx,
7199 12 : )
7200 12 : .await;
7201 12 :
7202 12 : let images = vectored_res?;
7203 12 : assert!(images.is_empty());
7204 12 : Ok(())
7205 12 : }
7206 :
7207 : // Test that vectored get handles layer gaps correctly
7208 : // by advancing into the next ancestor timeline if required.
7209 : //
7210 : // The test generates timelines that look like the diagram below.
7211 : // We leave a gap in one of the L1 layers at `gap_at_key` (`/` in the diagram).
7212 : // The reconstruct data for that key lies in the ancestor timeline (`X` in the diagram).
7213 : //
7214 : // ```
7215 : //-------------------------------+
7216 : // ... |
7217 : // [ L1 ] |
7218 : // [ / L1 ] | Child Timeline
7219 : // ... |
7220 : // ------------------------------+
7221 : // [ X L1 ] | Parent Timeline
7222 : // ------------------------------+
7223 : // ```
7224 : #[tokio::test]
7225 12 : async fn test_get_vectored_key_gap() -> anyhow::Result<()> {
7226 12 : let tenant_conf = pageserver_api::models::TenantConfig {
7227 12 : // Make compaction deterministic
7228 12 : gc_period: Some(Duration::ZERO),
7229 12 : compaction_period: Some(Duration::ZERO),
7230 12 : // Encourage creation of L1 layers
7231 12 : checkpoint_distance: Some(16 * 1024),
7232 12 : compaction_target_size: Some(8 * 1024),
7233 12 : ..Default::default()
7234 12 : };
7235 12 :
7236 12 : let harness = TenantHarness::create_custom(
7237 12 : "test_get_vectored_key_gap",
7238 12 : tenant_conf,
7239 12 : TenantId::generate(),
7240 12 : ShardIdentity::unsharded(),
7241 12 : Generation::new(0xdeadbeef),
7242 12 : )
7243 12 : .await?;
7244 12 : let (tenant, ctx) = harness.load().await;
7245 12 : let io_concurrency = IoConcurrency::spawn_for_test();
7246 12 :
7247 12 : let mut current_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7248 12 : let gap_at_key = current_key.add(100);
7249 12 : let mut current_lsn = Lsn(0x10);
7250 12 :
7251 12 : const KEY_COUNT: usize = 10_000;
7252 12 :
7253 12 : let timeline_id = TimelineId::generate();
7254 12 : let current_timeline = tenant
7255 12 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
7256 12 : .await?;
7257 12 :
7258 12 : current_lsn += 0x100;
7259 12 :
7260 12 : let mut writer = current_timeline.writer().await;
7261 12 : writer
7262 12 : .put(
7263 12 : gap_at_key,
7264 12 : current_lsn,
7265 12 : &Value::Image(test_img(&format!("{} at {}", gap_at_key, current_lsn))),
7266 12 : &ctx,
7267 12 : )
7268 12 : .await?;
7269 12 : writer.finish_write(current_lsn);
7270 12 : drop(writer);
7271 12 :
7272 12 : let mut latest_lsns = HashMap::new();
7273 12 : latest_lsns.insert(gap_at_key, current_lsn);
7274 12 :
7275 12 : current_timeline.freeze_and_flush().await?;
7276 12 :
7277 12 : let child_timeline_id = TimelineId::generate();
7278 12 :
7279 12 : tenant
7280 12 : .branch_timeline_test(
7281 12 : ¤t_timeline,
7282 12 : child_timeline_id,
7283 12 : Some(current_lsn),
7284 12 : &ctx,
7285 12 : )
7286 12 : .await?;
7287 12 : let child_timeline = tenant
7288 12 : .get_timeline(child_timeline_id, true)
7289 12 : .expect("Should have the branched timeline");
7290 12 :
7291 120012 : for i in 0..KEY_COUNT {
7292 120000 : if current_key == gap_at_key {
7293 12 : current_key = current_key.next();
7294 12 : continue;
7295 119988 : }
7296 119988 :
7297 119988 : current_lsn += 0x10;
7298 12 :
7299 119988 : let mut writer = child_timeline.writer().await;
7300 119988 : writer
7301 119988 : .put(
7302 119988 : current_key,
7303 119988 : current_lsn,
7304 119988 : &Value::Image(test_img(&format!("{} at {}", current_key, current_lsn))),
7305 119988 : &ctx,
7306 119988 : )
7307 119988 : .await?;
7308 119988 : writer.finish_write(current_lsn);
7309 119988 : drop(writer);
7310 119988 :
7311 119988 : latest_lsns.insert(current_key, current_lsn);
7312 119988 : current_key = current_key.next();
7313 119988 :
7314 119988 : // Flush every now and then to encourage layer file creation.
7315 119988 : if i % 500 == 0 {
7316 240 : child_timeline.freeze_and_flush().await?;
7317 119748 : }
7318 12 : }
7319 12 :
7320 12 : child_timeline.freeze_and_flush().await?;
7321 12 : let mut flags = EnumSet::new();
7322 12 : flags.insert(CompactFlags::ForceRepartition);
7323 12 : child_timeline
7324 12 : .compact(&CancellationToken::new(), flags, &ctx)
7325 12 : .await?;
7326 12 :
7327 12 : let key_near_end = {
7328 12 : let mut tmp = current_key;
7329 12 : tmp.field6 -= 10;
7330 12 : tmp
7331 12 : };
7332 12 :
7333 12 : let key_near_gap = {
7334 12 : let mut tmp = gap_at_key;
7335 12 : tmp.field6 -= 10;
7336 12 : tmp
7337 12 : };
7338 12 :
7339 12 : let read = KeySpace {
7340 12 : ranges: vec![key_near_gap..gap_at_key.next(), key_near_end..current_key],
7341 12 : };
7342 12 :
7343 12 : let query = VersionedKeySpaceQuery::uniform(read.clone(), current_lsn);
7344 12 :
7345 12 : let results = child_timeline
7346 12 : .get_vectored_impl(
7347 12 : query,
7348 12 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7349 12 : &ctx,
7350 12 : )
7351 12 : .await?;
7352 12 :
7353 264 : for (key, img_res) in results {
7354 252 : let expected = test_img(&format!("{} at {}", key, latest_lsns[&key]));
7355 252 : assert_eq!(img_res?, expected);
7356 12 : }
7357 12 :
7358 12 : Ok(())
7359 12 : }
7360 :
7361 : // Test that vectored get descends into ancestor timelines correctly and
7362 : // does not return an image that's newer than requested.
7363 : //
7364 : // The diagram below ilustrates an interesting case. We have a parent timeline
7365 : // (top of the Lsn range) and a child timeline. The request key cannot be reconstructed
7366 : // from the child timeline, so the parent timeline must be visited. When advacing into
7367 : // the child timeline, the read path needs to remember what the requested Lsn was in
7368 : // order to avoid returning an image that's too new. The test below constructs such
7369 : // a timeline setup and does a few queries around the Lsn of each page image.
7370 : // ```
7371 : // LSN
7372 : // ^
7373 : // |
7374 : // |
7375 : // 500 | --------------------------------------> branch point
7376 : // 400 | X
7377 : // 300 | X
7378 : // 200 | --------------------------------------> requested lsn
7379 : // 100 | X
7380 : // |---------------------------------------> Key
7381 : // |
7382 : // ------> requested key
7383 : //
7384 : // Legend:
7385 : // * X - page images
7386 : // ```
7387 : #[tokio::test]
7388 12 : async fn test_get_vectored_ancestor_descent() -> anyhow::Result<()> {
7389 12 : let harness = TenantHarness::create("test_get_vectored_on_lsn_axis").await?;
7390 12 : let (tenant, ctx) = harness.load().await;
7391 12 : let io_concurrency = IoConcurrency::spawn_for_test();
7392 12 :
7393 12 : let start_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7394 12 : let end_key = start_key.add(1000);
7395 12 : let child_gap_at_key = start_key.add(500);
7396 12 : let mut parent_gap_lsns: BTreeMap<Lsn, String> = BTreeMap::new();
7397 12 :
7398 12 : let mut current_lsn = Lsn(0x10);
7399 12 :
7400 12 : let timeline_id = TimelineId::generate();
7401 12 : let parent_timeline = tenant
7402 12 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
7403 12 : .await?;
7404 12 :
7405 12 : current_lsn += 0x100;
7406 12 :
7407 48 : for _ in 0..3 {
7408 36 : let mut key = start_key;
7409 36036 : while key < end_key {
7410 36000 : current_lsn += 0x10;
7411 36000 :
7412 36000 : let image_value = format!("{} at {}", child_gap_at_key, current_lsn);
7413 12 :
7414 36000 : let mut writer = parent_timeline.writer().await;
7415 36000 : writer
7416 36000 : .put(
7417 36000 : key,
7418 36000 : current_lsn,
7419 36000 : &Value::Image(test_img(&image_value)),
7420 36000 : &ctx,
7421 36000 : )
7422 36000 : .await?;
7423 36000 : writer.finish_write(current_lsn);
7424 36000 :
7425 36000 : if key == child_gap_at_key {
7426 36 : parent_gap_lsns.insert(current_lsn, image_value);
7427 35964 : }
7428 12 :
7429 36000 : key = key.next();
7430 12 : }
7431 12 :
7432 36 : parent_timeline.freeze_and_flush().await?;
7433 12 : }
7434 12 :
7435 12 : let child_timeline_id = TimelineId::generate();
7436 12 :
7437 12 : let child_timeline = tenant
7438 12 : .branch_timeline_test(&parent_timeline, child_timeline_id, Some(current_lsn), &ctx)
7439 12 : .await?;
7440 12 :
7441 12 : let mut key = start_key;
7442 12012 : while key < end_key {
7443 12000 : if key == child_gap_at_key {
7444 12 : key = key.next();
7445 12 : continue;
7446 11988 : }
7447 11988 :
7448 11988 : current_lsn += 0x10;
7449 12 :
7450 11988 : let mut writer = child_timeline.writer().await;
7451 11988 : writer
7452 11988 : .put(
7453 11988 : key,
7454 11988 : current_lsn,
7455 11988 : &Value::Image(test_img(&format!("{} at {}", key, current_lsn))),
7456 11988 : &ctx,
7457 11988 : )
7458 11988 : .await?;
7459 11988 : writer.finish_write(current_lsn);
7460 11988 :
7461 11988 : key = key.next();
7462 12 : }
7463 12 :
7464 12 : child_timeline.freeze_and_flush().await?;
7465 12 :
7466 12 : let lsn_offsets: [i64; 5] = [-10, -1, 0, 1, 10];
7467 12 : let mut query_lsns = Vec::new();
7468 36 : for image_lsn in parent_gap_lsns.keys().rev() {
7469 216 : for offset in lsn_offsets {
7470 180 : query_lsns.push(Lsn(image_lsn
7471 180 : .0
7472 180 : .checked_add_signed(offset)
7473 180 : .expect("Shouldn't overflow")));
7474 180 : }
7475 12 : }
7476 12 :
7477 192 : for query_lsn in query_lsns {
7478 180 : let query = VersionedKeySpaceQuery::uniform(
7479 180 : KeySpace {
7480 180 : ranges: vec![child_gap_at_key..child_gap_at_key.next()],
7481 180 : },
7482 180 : query_lsn,
7483 180 : );
7484 12 :
7485 180 : let results = child_timeline
7486 180 : .get_vectored_impl(
7487 180 : query,
7488 180 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7489 180 : &ctx,
7490 180 : )
7491 180 : .await;
7492 12 :
7493 180 : let expected_item = parent_gap_lsns
7494 180 : .iter()
7495 180 : .rev()
7496 408 : .find(|(lsn, _)| **lsn <= query_lsn);
7497 180 :
7498 180 : info!(
7499 12 : "Doing vectored read at LSN {}. Expecting image to be: {:?}",
7500 12 : query_lsn, expected_item
7501 12 : );
7502 12 :
7503 180 : match expected_item {
7504 156 : Some((_, img_value)) => {
7505 156 : let key_results = results.expect("No vectored get error expected");
7506 156 : let key_result = &key_results[&child_gap_at_key];
7507 156 : let returned_img = key_result
7508 156 : .as_ref()
7509 156 : .expect("No page reconstruct error expected");
7510 156 :
7511 156 : info!(
7512 12 : "Vectored read at LSN {} returned image {}",
7513 0 : query_lsn,
7514 0 : std::str::from_utf8(returned_img)?
7515 12 : );
7516 156 : assert_eq!(*returned_img, test_img(img_value));
7517 12 : }
7518 12 : None => {
7519 24 : assert!(matches!(results, Err(GetVectoredError::MissingKey(_))));
7520 12 : }
7521 12 : }
7522 12 : }
7523 12 :
7524 12 : Ok(())
7525 12 : }
7526 :
7527 : #[tokio::test]
7528 12 : async fn test_random_updates() -> anyhow::Result<()> {
7529 12 : let names_algorithms = [
7530 12 : ("test_random_updates_legacy", CompactionAlgorithm::Legacy),
7531 12 : ("test_random_updates_tiered", CompactionAlgorithm::Tiered),
7532 12 : ];
7533 36 : for (name, algorithm) in names_algorithms {
7534 24 : test_random_updates_algorithm(name, algorithm).await?;
7535 12 : }
7536 12 : Ok(())
7537 12 : }
7538 :
7539 24 : async fn test_random_updates_algorithm(
7540 24 : name: &'static str,
7541 24 : compaction_algorithm: CompactionAlgorithm,
7542 24 : ) -> anyhow::Result<()> {
7543 24 : let mut harness = TenantHarness::create(name).await?;
7544 24 : harness.tenant_conf.compaction_algorithm = Some(CompactionAlgorithmSettings {
7545 24 : kind: compaction_algorithm,
7546 24 : });
7547 24 : let (tenant, ctx) = harness.load().await;
7548 24 : let tline = tenant
7549 24 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7550 24 : .await?;
7551 :
7552 : const NUM_KEYS: usize = 1000;
7553 24 : let cancel = CancellationToken::new();
7554 24 :
7555 24 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7556 24 : let mut test_key_end = test_key;
7557 24 : test_key_end.field6 = NUM_KEYS as u32;
7558 24 : tline.add_extra_test_dense_keyspace(KeySpace::single(test_key..test_key_end));
7559 24 :
7560 24 : let mut keyspace = KeySpaceAccum::new();
7561 24 :
7562 24 : // Track when each page was last modified. Used to assert that
7563 24 : // a read sees the latest page version.
7564 24 : let mut updated = [Lsn(0); NUM_KEYS];
7565 24 :
7566 24 : let mut lsn = Lsn(0x10);
7567 : #[allow(clippy::needless_range_loop)]
7568 24024 : for blknum in 0..NUM_KEYS {
7569 24000 : lsn = Lsn(lsn.0 + 0x10);
7570 24000 : test_key.field6 = blknum as u32;
7571 24000 : let mut writer = tline.writer().await;
7572 24000 : writer
7573 24000 : .put(
7574 24000 : test_key,
7575 24000 : lsn,
7576 24000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7577 24000 : &ctx,
7578 24000 : )
7579 24000 : .await?;
7580 24000 : writer.finish_write(lsn);
7581 24000 : updated[blknum] = lsn;
7582 24000 : drop(writer);
7583 24000 :
7584 24000 : keyspace.add_key(test_key);
7585 : }
7586 :
7587 1224 : for _ in 0..50 {
7588 1201200 : for _ in 0..NUM_KEYS {
7589 1200000 : lsn = Lsn(lsn.0 + 0x10);
7590 1200000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7591 1200000 : test_key.field6 = blknum as u32;
7592 1200000 : let mut writer = tline.writer().await;
7593 1200000 : writer
7594 1200000 : .put(
7595 1200000 : test_key,
7596 1200000 : lsn,
7597 1200000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7598 1200000 : &ctx,
7599 1200000 : )
7600 1200000 : .await?;
7601 1200000 : writer.finish_write(lsn);
7602 1200000 : drop(writer);
7603 1200000 : updated[blknum] = lsn;
7604 : }
7605 :
7606 : // Read all the blocks
7607 1200000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7608 1200000 : test_key.field6 = blknum as u32;
7609 1200000 : assert_eq!(
7610 1200000 : tline.get(test_key, lsn, &ctx).await?,
7611 1200000 : test_img(&format!("{} at {}", blknum, last_lsn))
7612 : );
7613 : }
7614 :
7615 : // Perform a cycle of flush, and GC
7616 1200 : tline.freeze_and_flush().await?;
7617 1200 : tenant
7618 1200 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7619 1200 : .await?;
7620 : }
7621 :
7622 24 : Ok(())
7623 24 : }
7624 :
7625 : #[tokio::test]
7626 12 : async fn test_traverse_branches() -> anyhow::Result<()> {
7627 12 : let (tenant, ctx) = TenantHarness::create("test_traverse_branches")
7628 12 : .await?
7629 12 : .load()
7630 12 : .await;
7631 12 : let mut tline = tenant
7632 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7633 12 : .await?;
7634 12 :
7635 12 : const NUM_KEYS: usize = 1000;
7636 12 :
7637 12 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7638 12 :
7639 12 : let mut keyspace = KeySpaceAccum::new();
7640 12 :
7641 12 : let cancel = CancellationToken::new();
7642 12 :
7643 12 : // Track when each page was last modified. Used to assert that
7644 12 : // a read sees the latest page version.
7645 12 : let mut updated = [Lsn(0); NUM_KEYS];
7646 12 :
7647 12 : let mut lsn = Lsn(0x10);
7648 12 : #[allow(clippy::needless_range_loop)]
7649 12012 : for blknum in 0..NUM_KEYS {
7650 12000 : lsn = Lsn(lsn.0 + 0x10);
7651 12000 : test_key.field6 = blknum as u32;
7652 12000 : let mut writer = tline.writer().await;
7653 12000 : writer
7654 12000 : .put(
7655 12000 : test_key,
7656 12000 : lsn,
7657 12000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7658 12000 : &ctx,
7659 12000 : )
7660 12000 : .await?;
7661 12000 : writer.finish_write(lsn);
7662 12000 : updated[blknum] = lsn;
7663 12000 : drop(writer);
7664 12000 :
7665 12000 : keyspace.add_key(test_key);
7666 12 : }
7667 12 :
7668 612 : for _ in 0..50 {
7669 600 : let new_tline_id = TimelineId::generate();
7670 600 : tenant
7671 600 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7672 600 : .await?;
7673 600 : tline = tenant
7674 600 : .get_timeline(new_tline_id, true)
7675 600 : .expect("Should have the branched timeline");
7676 12 :
7677 600600 : for _ in 0..NUM_KEYS {
7678 600000 : lsn = Lsn(lsn.0 + 0x10);
7679 600000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7680 600000 : test_key.field6 = blknum as u32;
7681 600000 : let mut writer = tline.writer().await;
7682 600000 : writer
7683 600000 : .put(
7684 600000 : test_key,
7685 600000 : lsn,
7686 600000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7687 600000 : &ctx,
7688 600000 : )
7689 600000 : .await?;
7690 600000 : println!("updating {} at {}", blknum, lsn);
7691 600000 : writer.finish_write(lsn);
7692 600000 : drop(writer);
7693 600000 : updated[blknum] = lsn;
7694 12 : }
7695 12 :
7696 12 : // Read all the blocks
7697 600000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7698 600000 : test_key.field6 = blknum as u32;
7699 600000 : assert_eq!(
7700 600000 : tline.get(test_key, lsn, &ctx).await?,
7701 600000 : test_img(&format!("{} at {}", blknum, last_lsn))
7702 12 : );
7703 12 : }
7704 12 :
7705 12 : // Perform a cycle of flush, compact, and GC
7706 600 : tline.freeze_and_flush().await?;
7707 600 : tline.compact(&cancel, EnumSet::default(), &ctx).await?;
7708 600 : tenant
7709 600 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7710 600 : .await?;
7711 12 : }
7712 12 :
7713 12 : Ok(())
7714 12 : }
7715 :
7716 : #[tokio::test]
7717 12 : async fn test_traverse_ancestors() -> anyhow::Result<()> {
7718 12 : let (tenant, ctx) = TenantHarness::create("test_traverse_ancestors")
7719 12 : .await?
7720 12 : .load()
7721 12 : .await;
7722 12 : let mut tline = tenant
7723 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7724 12 : .await?;
7725 12 :
7726 12 : const NUM_KEYS: usize = 100;
7727 12 : const NUM_TLINES: usize = 50;
7728 12 :
7729 12 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7730 12 : // Track page mutation lsns across different timelines.
7731 12 : let mut updated = [[Lsn(0); NUM_KEYS]; NUM_TLINES];
7732 12 :
7733 12 : let mut lsn = Lsn(0x10);
7734 12 :
7735 12 : #[allow(clippy::needless_range_loop)]
7736 612 : for idx in 0..NUM_TLINES {
7737 600 : let new_tline_id = TimelineId::generate();
7738 600 : tenant
7739 600 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7740 600 : .await?;
7741 600 : tline = tenant
7742 600 : .get_timeline(new_tline_id, true)
7743 600 : .expect("Should have the branched timeline");
7744 12 :
7745 60600 : for _ in 0..NUM_KEYS {
7746 60000 : lsn = Lsn(lsn.0 + 0x10);
7747 60000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7748 60000 : test_key.field6 = blknum as u32;
7749 60000 : let mut writer = tline.writer().await;
7750 60000 : writer
7751 60000 : .put(
7752 60000 : test_key,
7753 60000 : lsn,
7754 60000 : &Value::Image(test_img(&format!("{} {} at {}", idx, blknum, lsn))),
7755 60000 : &ctx,
7756 60000 : )
7757 60000 : .await?;
7758 60000 : println!("updating [{}][{}] at {}", idx, blknum, lsn);
7759 60000 : writer.finish_write(lsn);
7760 60000 : drop(writer);
7761 60000 : updated[idx][blknum] = lsn;
7762 12 : }
7763 12 : }
7764 12 :
7765 12 : // Read pages from leaf timeline across all ancestors.
7766 600 : for (idx, lsns) in updated.iter().enumerate() {
7767 60000 : for (blknum, lsn) in lsns.iter().enumerate() {
7768 12 : // Skip empty mutations.
7769 60000 : if lsn.0 == 0 {
7770 21964 : continue;
7771 38036 : }
7772 38036 : println!("checking [{idx}][{blknum}] at {lsn}");
7773 38036 : test_key.field6 = blknum as u32;
7774 38036 : assert_eq!(
7775 38036 : tline.get(test_key, *lsn, &ctx).await?,
7776 38036 : test_img(&format!("{idx} {blknum} at {lsn}"))
7777 12 : );
7778 12 : }
7779 12 : }
7780 12 : Ok(())
7781 12 : }
7782 :
7783 : #[tokio::test]
7784 12 : async fn test_write_at_initdb_lsn_takes_optimization_code_path() -> anyhow::Result<()> {
7785 12 : let (tenant, ctx) = TenantHarness::create("test_empty_test_timeline_is_usable")
7786 12 : .await?
7787 12 : .load()
7788 12 : .await;
7789 12 :
7790 12 : let initdb_lsn = Lsn(0x20);
7791 12 : let (utline, ctx) = tenant
7792 12 : .create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)
7793 12 : .await?;
7794 12 : let tline = utline.raw_timeline().unwrap();
7795 12 :
7796 12 : // Spawn flush loop now so that we can set the `expect_initdb_optimization`
7797 12 : tline.maybe_spawn_flush_loop();
7798 12 :
7799 12 : // Make sure the timeline has the minimum set of required keys for operation.
7800 12 : // The only operation you can always do on an empty timeline is to `put` new data.
7801 12 : // Except if you `put` at `initdb_lsn`.
7802 12 : // In that case, there's an optimization to directly create image layers instead of delta layers.
7803 12 : // It uses `repartition()`, which assumes some keys to be present.
7804 12 : // Let's make sure the test timeline can handle that case.
7805 12 : {
7806 12 : let mut state = tline.flush_loop_state.lock().unwrap();
7807 12 : assert_eq!(
7808 12 : timeline::FlushLoopState::Running {
7809 12 : expect_initdb_optimization: false,
7810 12 : initdb_optimization_count: 0,
7811 12 : },
7812 12 : *state
7813 12 : );
7814 12 : *state = timeline::FlushLoopState::Running {
7815 12 : expect_initdb_optimization: true,
7816 12 : initdb_optimization_count: 0,
7817 12 : };
7818 12 : }
7819 12 :
7820 12 : // Make writes at the initdb_lsn. When we flush it below, it should be handled by the optimization.
7821 12 : // As explained above, the optimization requires some keys to be present.
7822 12 : // As per `create_empty_timeline` documentation, use init_empty to set them.
7823 12 : // This is what `create_test_timeline` does, by the way.
7824 12 : let mut modification = tline.begin_modification(initdb_lsn);
7825 12 : modification
7826 12 : .init_empty_test_timeline()
7827 12 : .context("init_empty_test_timeline")?;
7828 12 : modification
7829 12 : .commit(&ctx)
7830 12 : .await
7831 12 : .context("commit init_empty_test_timeline modification")?;
7832 12 :
7833 12 : // Do the flush. The flush code will check the expectations that we set above.
7834 12 : tline.freeze_and_flush().await?;
7835 12 :
7836 12 : // assert freeze_and_flush exercised the initdb optimization
7837 12 : {
7838 12 : let state = tline.flush_loop_state.lock().unwrap();
7839 12 : let timeline::FlushLoopState::Running {
7840 12 : expect_initdb_optimization,
7841 12 : initdb_optimization_count,
7842 12 : } = *state
7843 12 : else {
7844 12 : panic!("unexpected state: {:?}", *state);
7845 12 : };
7846 12 : assert!(expect_initdb_optimization);
7847 12 : assert!(initdb_optimization_count > 0);
7848 12 : }
7849 12 : Ok(())
7850 12 : }
7851 :
7852 : #[tokio::test]
7853 12 : async fn test_create_guard_crash() -> anyhow::Result<()> {
7854 12 : let name = "test_create_guard_crash";
7855 12 : let harness = TenantHarness::create(name).await?;
7856 12 : {
7857 12 : let (tenant, ctx) = harness.load().await;
7858 12 : let (tline, _ctx) = tenant
7859 12 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
7860 12 : .await?;
7861 12 : // Leave the timeline ID in [`TenantShard::timelines_creating`] to exclude attempting to create it again
7862 12 : let raw_tline = tline.raw_timeline().unwrap();
7863 12 : raw_tline
7864 12 : .shutdown(super::timeline::ShutdownMode::Hard)
7865 12 : .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))
7866 12 : .await;
7867 12 : std::mem::forget(tline);
7868 12 : }
7869 12 :
7870 12 : let (tenant, _) = harness.load().await;
7871 12 : match tenant.get_timeline(TIMELINE_ID, false) {
7872 12 : Ok(_) => panic!("timeline should've been removed during load"),
7873 12 : Err(e) => {
7874 12 : assert_eq!(
7875 12 : e,
7876 12 : GetTimelineError::NotFound {
7877 12 : tenant_id: tenant.tenant_shard_id,
7878 12 : timeline_id: TIMELINE_ID,
7879 12 : }
7880 12 : )
7881 12 : }
7882 12 : }
7883 12 :
7884 12 : assert!(
7885 12 : !harness
7886 12 : .conf
7887 12 : .timeline_path(&tenant.tenant_shard_id, &TIMELINE_ID)
7888 12 : .exists()
7889 12 : );
7890 12 :
7891 12 : Ok(())
7892 12 : }
7893 :
7894 : #[tokio::test]
7895 12 : async fn test_read_at_max_lsn() -> anyhow::Result<()> {
7896 12 : let names_algorithms = [
7897 12 : ("test_read_at_max_lsn_legacy", CompactionAlgorithm::Legacy),
7898 12 : ("test_read_at_max_lsn_tiered", CompactionAlgorithm::Tiered),
7899 12 : ];
7900 36 : for (name, algorithm) in names_algorithms {
7901 24 : test_read_at_max_lsn_algorithm(name, algorithm).await?;
7902 12 : }
7903 12 : Ok(())
7904 12 : }
7905 :
7906 24 : async fn test_read_at_max_lsn_algorithm(
7907 24 : name: &'static str,
7908 24 : compaction_algorithm: CompactionAlgorithm,
7909 24 : ) -> anyhow::Result<()> {
7910 24 : let mut harness = TenantHarness::create(name).await?;
7911 24 : harness.tenant_conf.compaction_algorithm = Some(CompactionAlgorithmSettings {
7912 24 : kind: compaction_algorithm,
7913 24 : });
7914 24 : let (tenant, ctx) = harness.load().await;
7915 24 : let tline = tenant
7916 24 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
7917 24 : .await?;
7918 :
7919 24 : let lsn = Lsn(0x10);
7920 24 : let compact = false;
7921 24 : bulk_insert_maybe_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000, compact).await?;
7922 :
7923 24 : let test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7924 24 : let read_lsn = Lsn(u64::MAX - 1);
7925 :
7926 24 : let result = tline.get(test_key, read_lsn, &ctx).await;
7927 24 : assert!(result.is_ok(), "result is not Ok: {}", result.unwrap_err());
7928 :
7929 24 : Ok(())
7930 24 : }
7931 :
7932 : #[tokio::test]
7933 12 : async fn test_metadata_scan() -> anyhow::Result<()> {
7934 12 : let harness = TenantHarness::create("test_metadata_scan").await?;
7935 12 : let (tenant, ctx) = harness.load().await;
7936 12 : let io_concurrency = IoConcurrency::spawn_for_test();
7937 12 : let tline = tenant
7938 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7939 12 : .await?;
7940 12 :
7941 12 : const NUM_KEYS: usize = 1000;
7942 12 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
7943 12 :
7944 12 : let cancel = CancellationToken::new();
7945 12 :
7946 12 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7947 12 : base_key.field1 = AUX_KEY_PREFIX;
7948 12 : let mut test_key = base_key;
7949 12 :
7950 12 : // Track when each page was last modified. Used to assert that
7951 12 : // a read sees the latest page version.
7952 12 : let mut updated = [Lsn(0); NUM_KEYS];
7953 12 :
7954 12 : let mut lsn = Lsn(0x10);
7955 12 : #[allow(clippy::needless_range_loop)]
7956 12012 : for blknum in 0..NUM_KEYS {
7957 12000 : lsn = Lsn(lsn.0 + 0x10);
7958 12000 : test_key.field6 = (blknum * STEP) as u32;
7959 12000 : let mut writer = tline.writer().await;
7960 12000 : writer
7961 12000 : .put(
7962 12000 : test_key,
7963 12000 : lsn,
7964 12000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7965 12000 : &ctx,
7966 12000 : )
7967 12000 : .await?;
7968 12000 : writer.finish_write(lsn);
7969 12000 : updated[blknum] = lsn;
7970 12000 : drop(writer);
7971 12 : }
7972 12 :
7973 12 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
7974 12 :
7975 144 : for iter in 0..=10 {
7976 12 : // Read all the blocks
7977 132000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7978 132000 : test_key.field6 = (blknum * STEP) as u32;
7979 132000 : assert_eq!(
7980 132000 : tline.get(test_key, lsn, &ctx).await?,
7981 132000 : test_img(&format!("{} at {}", blknum, last_lsn))
7982 12 : );
7983 12 : }
7984 12 :
7985 132 : let mut cnt = 0;
7986 132 : let query = VersionedKeySpaceQuery::uniform(keyspace.clone(), lsn);
7987 12 :
7988 132000 : for (key, value) in tline
7989 132 : .get_vectored_impl(
7990 132 : query,
7991 132 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7992 132 : &ctx,
7993 132 : )
7994 132 : .await?
7995 12 : {
7996 132000 : let blknum = key.field6 as usize;
7997 132000 : let value = value?;
7998 132000 : assert!(blknum % STEP == 0);
7999 132000 : let blknum = blknum / STEP;
8000 132000 : assert_eq!(
8001 132000 : value,
8002 132000 : test_img(&format!("{} at {}", blknum, updated[blknum]))
8003 132000 : );
8004 132000 : cnt += 1;
8005 12 : }
8006 12 :
8007 132 : assert_eq!(cnt, NUM_KEYS);
8008 12 :
8009 132132 : for _ in 0..NUM_KEYS {
8010 132000 : lsn = Lsn(lsn.0 + 0x10);
8011 132000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
8012 132000 : test_key.field6 = (blknum * STEP) as u32;
8013 132000 : let mut writer = tline.writer().await;
8014 132000 : writer
8015 132000 : .put(
8016 132000 : test_key,
8017 132000 : lsn,
8018 132000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
8019 132000 : &ctx,
8020 132000 : )
8021 132000 : .await?;
8022 132000 : writer.finish_write(lsn);
8023 132000 : drop(writer);
8024 132000 : updated[blknum] = lsn;
8025 12 : }
8026 12 :
8027 12 : // Perform two cycles of flush, compact, and GC
8028 396 : for round in 0..2 {
8029 264 : tline.freeze_and_flush().await?;
8030 264 : tline
8031 264 : .compact(
8032 264 : &cancel,
8033 264 : if iter % 5 == 0 && round == 0 {
8034 36 : let mut flags = EnumSet::new();
8035 36 : flags.insert(CompactFlags::ForceImageLayerCreation);
8036 36 : flags.insert(CompactFlags::ForceRepartition);
8037 36 : flags
8038 12 : } else {
8039 228 : EnumSet::empty()
8040 12 : },
8041 264 : &ctx,
8042 264 : )
8043 264 : .await?;
8044 264 : tenant
8045 264 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
8046 264 : .await?;
8047 12 : }
8048 12 : }
8049 12 :
8050 12 : Ok(())
8051 12 : }
8052 :
8053 : #[tokio::test]
8054 12 : async fn test_metadata_compaction_trigger() -> anyhow::Result<()> {
8055 12 : let harness = TenantHarness::create("test_metadata_compaction_trigger").await?;
8056 12 : let (tenant, ctx) = harness.load().await;
8057 12 : let tline = tenant
8058 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
8059 12 : .await?;
8060 12 :
8061 12 : let cancel = CancellationToken::new();
8062 12 :
8063 12 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
8064 12 : base_key.field1 = AUX_KEY_PREFIX;
8065 12 : let test_key = base_key;
8066 12 : let mut lsn = Lsn(0x10);
8067 12 :
8068 252 : for _ in 0..20 {
8069 240 : lsn = Lsn(lsn.0 + 0x10);
8070 240 : let mut writer = tline.writer().await;
8071 240 : writer
8072 240 : .put(
8073 240 : test_key,
8074 240 : lsn,
8075 240 : &Value::Image(test_img(&format!("{} at {}", 0, lsn))),
8076 240 : &ctx,
8077 240 : )
8078 240 : .await?;
8079 240 : writer.finish_write(lsn);
8080 240 : drop(writer);
8081 240 : tline.freeze_and_flush().await?; // force create a delta layer
8082 12 : }
8083 12 :
8084 12 : let before_num_l0_delta_files =
8085 12 : tline.layers.read().await.layer_map()?.level0_deltas().len();
8086 12 :
8087 12 : tline.compact(&cancel, EnumSet::default(), &ctx).await?;
8088 12 :
8089 12 : let after_num_l0_delta_files = tline.layers.read().await.layer_map()?.level0_deltas().len();
8090 12 :
8091 12 : assert!(
8092 12 : after_num_l0_delta_files < before_num_l0_delta_files,
8093 12 : "after_num_l0_delta_files={after_num_l0_delta_files}, before_num_l0_delta_files={before_num_l0_delta_files}"
8094 12 : );
8095 12 :
8096 12 : assert_eq!(
8097 12 : tline.get(test_key, lsn, &ctx).await?,
8098 12 : test_img(&format!("{} at {}", 0, lsn))
8099 12 : );
8100 12 :
8101 12 : Ok(())
8102 12 : }
8103 :
8104 : #[tokio::test]
8105 12 : async fn test_aux_file_e2e() {
8106 12 : let harness = TenantHarness::create("test_aux_file_e2e").await.unwrap();
8107 12 :
8108 12 : let (tenant, ctx) = harness.load().await;
8109 12 : let io_concurrency = IoConcurrency::spawn_for_test();
8110 12 :
8111 12 : let mut lsn = Lsn(0x08);
8112 12 :
8113 12 : let tline: Arc<Timeline> = tenant
8114 12 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
8115 12 : .await
8116 12 : .unwrap();
8117 12 :
8118 12 : {
8119 12 : lsn += 8;
8120 12 : let mut modification = tline.begin_modification(lsn);
8121 12 : modification
8122 12 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
8123 12 : .await
8124 12 : .unwrap();
8125 12 : modification.commit(&ctx).await.unwrap();
8126 12 : }
8127 12 :
8128 12 : // we can read everything from the storage
8129 12 : let files = tline
8130 12 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
8131 12 : .await
8132 12 : .unwrap();
8133 12 : assert_eq!(
8134 12 : files.get("pg_logical/mappings/test1"),
8135 12 : Some(&bytes::Bytes::from_static(b"first"))
8136 12 : );
8137 12 :
8138 12 : {
8139 12 : lsn += 8;
8140 12 : let mut modification = tline.begin_modification(lsn);
8141 12 : modification
8142 12 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
8143 12 : .await
8144 12 : .unwrap();
8145 12 : modification.commit(&ctx).await.unwrap();
8146 12 : }
8147 12 :
8148 12 : let files = tline
8149 12 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
8150 12 : .await
8151 12 : .unwrap();
8152 12 : assert_eq!(
8153 12 : files.get("pg_logical/mappings/test2"),
8154 12 : Some(&bytes::Bytes::from_static(b"second"))
8155 12 : );
8156 12 :
8157 12 : let child = tenant
8158 12 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(lsn), &ctx)
8159 12 : .await
8160 12 : .unwrap();
8161 12 :
8162 12 : let files = child
8163 12 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
8164 12 : .await
8165 12 : .unwrap();
8166 12 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
8167 12 : assert_eq!(files.get("pg_logical/mappings/test2"), None);
8168 12 : }
8169 :
8170 : #[tokio::test]
8171 12 : async fn test_metadata_image_creation() -> anyhow::Result<()> {
8172 12 : let harness = TenantHarness::create("test_metadata_image_creation").await?;
8173 12 : let (tenant, ctx) = harness.load().await;
8174 12 : let io_concurrency = IoConcurrency::spawn_for_test();
8175 12 : let tline = tenant
8176 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
8177 12 : .await?;
8178 12 :
8179 12 : const NUM_KEYS: usize = 1000;
8180 12 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
8181 12 :
8182 12 : let cancel = CancellationToken::new();
8183 12 :
8184 12 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8185 12 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
8186 12 : let mut test_key = base_key;
8187 12 : let mut lsn = Lsn(0x10);
8188 12 :
8189 48 : async fn scan_with_statistics(
8190 48 : tline: &Timeline,
8191 48 : keyspace: &KeySpace,
8192 48 : lsn: Lsn,
8193 48 : ctx: &RequestContext,
8194 48 : io_concurrency: IoConcurrency,
8195 48 : ) -> anyhow::Result<(BTreeMap<Key, Result<Bytes, PageReconstructError>>, usize)> {
8196 48 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
8197 48 : let query = VersionedKeySpaceQuery::uniform(keyspace.clone(), lsn);
8198 48 : let res = tline
8199 48 : .get_vectored_impl(query, &mut reconstruct_state, ctx)
8200 48 : .await?;
8201 48 : Ok((res, reconstruct_state.get_delta_layers_visited() as usize))
8202 48 : }
8203 12 :
8204 12012 : for blknum in 0..NUM_KEYS {
8205 12000 : lsn = Lsn(lsn.0 + 0x10);
8206 12000 : test_key.field6 = (blknum * STEP) as u32;
8207 12000 : let mut writer = tline.writer().await;
8208 12000 : writer
8209 12000 : .put(
8210 12000 : test_key,
8211 12000 : lsn,
8212 12000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
8213 12000 : &ctx,
8214 12000 : )
8215 12000 : .await?;
8216 12000 : writer.finish_write(lsn);
8217 12000 : drop(writer);
8218 12 : }
8219 12 :
8220 12 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
8221 12 :
8222 132 : for iter in 1..=10 {
8223 120120 : for _ in 0..NUM_KEYS {
8224 120000 : lsn = Lsn(lsn.0 + 0x10);
8225 120000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
8226 120000 : test_key.field6 = (blknum * STEP) as u32;
8227 120000 : let mut writer = tline.writer().await;
8228 120000 : writer
8229 120000 : .put(
8230 120000 : test_key,
8231 120000 : lsn,
8232 120000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
8233 120000 : &ctx,
8234 120000 : )
8235 120000 : .await?;
8236 120000 : writer.finish_write(lsn);
8237 120000 : drop(writer);
8238 12 : }
8239 12 :
8240 120 : tline.freeze_and_flush().await?;
8241 12 :
8242 120 : if iter % 5 == 0 {
8243 24 : let (_, before_delta_file_accessed) =
8244 24 : scan_with_statistics(&tline, &keyspace, lsn, &ctx, io_concurrency.clone())
8245 24 : .await?;
8246 24 : tline
8247 24 : .compact(
8248 24 : &cancel,
8249 24 : {
8250 24 : let mut flags = EnumSet::new();
8251 24 : flags.insert(CompactFlags::ForceImageLayerCreation);
8252 24 : flags.insert(CompactFlags::ForceRepartition);
8253 24 : flags
8254 24 : },
8255 24 : &ctx,
8256 24 : )
8257 24 : .await?;
8258 24 : let (_, after_delta_file_accessed) =
8259 24 : scan_with_statistics(&tline, &keyspace, lsn, &ctx, io_concurrency.clone())
8260 24 : .await?;
8261 24 : assert!(
8262 24 : after_delta_file_accessed < before_delta_file_accessed,
8263 12 : "after_delta_file_accessed={after_delta_file_accessed}, before_delta_file_accessed={before_delta_file_accessed}"
8264 12 : );
8265 12 : // 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.
8266 24 : assert!(
8267 24 : after_delta_file_accessed <= 2,
8268 12 : "after_delta_file_accessed={after_delta_file_accessed}"
8269 12 : );
8270 96 : }
8271 12 : }
8272 12 :
8273 12 : Ok(())
8274 12 : }
8275 :
8276 : #[tokio::test]
8277 12 : async fn test_vectored_missing_data_key_reads() -> anyhow::Result<()> {
8278 12 : let harness = TenantHarness::create("test_vectored_missing_data_key_reads").await?;
8279 12 : let (tenant, ctx) = harness.load().await;
8280 12 :
8281 12 : let base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
8282 12 : let base_key_child = Key::from_hex("000000000033333333444444445500000001").unwrap();
8283 12 : let base_key_nonexist = Key::from_hex("000000000033333333444444445500000002").unwrap();
8284 12 :
8285 12 : let tline = tenant
8286 12 : .create_test_timeline_with_layers(
8287 12 : TIMELINE_ID,
8288 12 : Lsn(0x10),
8289 12 : DEFAULT_PG_VERSION,
8290 12 : &ctx,
8291 12 : Vec::new(), // in-memory layers
8292 12 : Vec::new(), // delta layers
8293 12 : vec![(Lsn(0x20), vec![(base_key, test_img("data key 1"))])], // image layers
8294 12 : 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
8295 12 : )
8296 12 : .await?;
8297 12 : tline.add_extra_test_dense_keyspace(KeySpace::single(base_key..(base_key_nonexist.next())));
8298 12 :
8299 12 : let child = tenant
8300 12 : .branch_timeline_test_with_layers(
8301 12 : &tline,
8302 12 : NEW_TIMELINE_ID,
8303 12 : Some(Lsn(0x20)),
8304 12 : &ctx,
8305 12 : Vec::new(), // delta layers
8306 12 : vec![(Lsn(0x30), vec![(base_key_child, test_img("data key 2"))])], // image layers
8307 12 : Lsn(0x30),
8308 12 : )
8309 12 : .await
8310 12 : .unwrap();
8311 12 :
8312 12 : let lsn = Lsn(0x30);
8313 12 :
8314 12 : // test vectored get on parent timeline
8315 12 : assert_eq!(
8316 12 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
8317 12 : Some(test_img("data key 1"))
8318 12 : );
8319 12 : assert!(
8320 12 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx)
8321 12 : .await
8322 12 : .unwrap_err()
8323 12 : .is_missing_key_error()
8324 12 : );
8325 12 : assert!(
8326 12 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx)
8327 12 : .await
8328 12 : .unwrap_err()
8329 12 : .is_missing_key_error()
8330 12 : );
8331 12 :
8332 12 : // test vectored get on child timeline
8333 12 : assert_eq!(
8334 12 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
8335 12 : Some(test_img("data key 1"))
8336 12 : );
8337 12 : assert_eq!(
8338 12 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
8339 12 : Some(test_img("data key 2"))
8340 12 : );
8341 12 : assert!(
8342 12 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx)
8343 12 : .await
8344 12 : .unwrap_err()
8345 12 : .is_missing_key_error()
8346 12 : );
8347 12 :
8348 12 : Ok(())
8349 12 : }
8350 :
8351 : #[tokio::test]
8352 12 : async fn test_vectored_missing_metadata_key_reads() -> anyhow::Result<()> {
8353 12 : let harness = TenantHarness::create("test_vectored_missing_metadata_key_reads").await?;
8354 12 : let (tenant, ctx) = harness.load().await;
8355 12 : let io_concurrency = IoConcurrency::spawn_for_test();
8356 12 :
8357 12 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8358 12 : let base_key_child = Key::from_hex("620000000033333333444444445500000001").unwrap();
8359 12 : let base_key_nonexist = Key::from_hex("620000000033333333444444445500000002").unwrap();
8360 12 : let base_key_overwrite = Key::from_hex("620000000033333333444444445500000003").unwrap();
8361 12 :
8362 12 : let base_inherited_key = Key::from_hex("610000000033333333444444445500000000").unwrap();
8363 12 : let base_inherited_key_child =
8364 12 : Key::from_hex("610000000033333333444444445500000001").unwrap();
8365 12 : let base_inherited_key_nonexist =
8366 12 : Key::from_hex("610000000033333333444444445500000002").unwrap();
8367 12 : let base_inherited_key_overwrite =
8368 12 : Key::from_hex("610000000033333333444444445500000003").unwrap();
8369 12 :
8370 12 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
8371 12 : assert_eq!(base_inherited_key.field1, RELATION_SIZE_PREFIX);
8372 12 :
8373 12 : let tline = tenant
8374 12 : .create_test_timeline_with_layers(
8375 12 : TIMELINE_ID,
8376 12 : Lsn(0x10),
8377 12 : DEFAULT_PG_VERSION,
8378 12 : &ctx,
8379 12 : Vec::new(), // in-memory layers
8380 12 : Vec::new(), // delta layers
8381 12 : vec![(
8382 12 : Lsn(0x20),
8383 12 : vec![
8384 12 : (base_inherited_key, test_img("metadata inherited key 1")),
8385 12 : (
8386 12 : base_inherited_key_overwrite,
8387 12 : test_img("metadata key overwrite 1a"),
8388 12 : ),
8389 12 : (base_key, test_img("metadata key 1")),
8390 12 : (base_key_overwrite, test_img("metadata key overwrite 1b")),
8391 12 : ],
8392 12 : )], // image layers
8393 12 : 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
8394 12 : )
8395 12 : .await?;
8396 12 :
8397 12 : let child = tenant
8398 12 : .branch_timeline_test_with_layers(
8399 12 : &tline,
8400 12 : NEW_TIMELINE_ID,
8401 12 : Some(Lsn(0x20)),
8402 12 : &ctx,
8403 12 : Vec::new(), // delta layers
8404 12 : vec![(
8405 12 : Lsn(0x30),
8406 12 : vec![
8407 12 : (
8408 12 : base_inherited_key_child,
8409 12 : test_img("metadata inherited key 2"),
8410 12 : ),
8411 12 : (
8412 12 : base_inherited_key_overwrite,
8413 12 : test_img("metadata key overwrite 2a"),
8414 12 : ),
8415 12 : (base_key_child, test_img("metadata key 2")),
8416 12 : (base_key_overwrite, test_img("metadata key overwrite 2b")),
8417 12 : ],
8418 12 : )], // image layers
8419 12 : Lsn(0x30),
8420 12 : )
8421 12 : .await
8422 12 : .unwrap();
8423 12 :
8424 12 : let lsn = Lsn(0x30);
8425 12 :
8426 12 : // test vectored get on parent timeline
8427 12 : assert_eq!(
8428 12 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
8429 12 : Some(test_img("metadata key 1"))
8430 12 : );
8431 12 : assert_eq!(
8432 12 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx).await?,
8433 12 : None
8434 12 : );
8435 12 : assert_eq!(
8436 12 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx).await?,
8437 12 : None
8438 12 : );
8439 12 : assert_eq!(
8440 12 : get_vectored_impl_wrapper(&tline, base_key_overwrite, lsn, &ctx).await?,
8441 12 : Some(test_img("metadata key overwrite 1b"))
8442 12 : );
8443 12 : assert_eq!(
8444 12 : get_vectored_impl_wrapper(&tline, base_inherited_key, lsn, &ctx).await?,
8445 12 : Some(test_img("metadata inherited key 1"))
8446 12 : );
8447 12 : assert_eq!(
8448 12 : get_vectored_impl_wrapper(&tline, base_inherited_key_child, lsn, &ctx).await?,
8449 12 : None
8450 12 : );
8451 12 : assert_eq!(
8452 12 : get_vectored_impl_wrapper(&tline, base_inherited_key_nonexist, lsn, &ctx).await?,
8453 12 : None
8454 12 : );
8455 12 : assert_eq!(
8456 12 : get_vectored_impl_wrapper(&tline, base_inherited_key_overwrite, lsn, &ctx).await?,
8457 12 : Some(test_img("metadata key overwrite 1a"))
8458 12 : );
8459 12 :
8460 12 : // test vectored get on child timeline
8461 12 : assert_eq!(
8462 12 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
8463 12 : None
8464 12 : );
8465 12 : assert_eq!(
8466 12 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
8467 12 : Some(test_img("metadata key 2"))
8468 12 : );
8469 12 : assert_eq!(
8470 12 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx).await?,
8471 12 : None
8472 12 : );
8473 12 : assert_eq!(
8474 12 : get_vectored_impl_wrapper(&child, base_inherited_key, lsn, &ctx).await?,
8475 12 : Some(test_img("metadata inherited key 1"))
8476 12 : );
8477 12 : assert_eq!(
8478 12 : get_vectored_impl_wrapper(&child, base_inherited_key_child, lsn, &ctx).await?,
8479 12 : Some(test_img("metadata inherited key 2"))
8480 12 : );
8481 12 : assert_eq!(
8482 12 : get_vectored_impl_wrapper(&child, base_inherited_key_nonexist, lsn, &ctx).await?,
8483 12 : None
8484 12 : );
8485 12 : assert_eq!(
8486 12 : get_vectored_impl_wrapper(&child, base_key_overwrite, lsn, &ctx).await?,
8487 12 : Some(test_img("metadata key overwrite 2b"))
8488 12 : );
8489 12 : assert_eq!(
8490 12 : get_vectored_impl_wrapper(&child, base_inherited_key_overwrite, lsn, &ctx).await?,
8491 12 : Some(test_img("metadata key overwrite 2a"))
8492 12 : );
8493 12 :
8494 12 : // test vectored scan on parent timeline
8495 12 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency.clone());
8496 12 : let query =
8497 12 : VersionedKeySpaceQuery::uniform(KeySpace::single(Key::metadata_key_range()), lsn);
8498 12 : let res = tline
8499 12 : .get_vectored_impl(query, &mut reconstruct_state, &ctx)
8500 12 : .await?;
8501 12 :
8502 12 : assert_eq!(
8503 12 : res.into_iter()
8504 48 : .map(|(k, v)| (k, v.unwrap()))
8505 12 : .collect::<Vec<_>>(),
8506 12 : vec![
8507 12 : (base_inherited_key, test_img("metadata inherited key 1")),
8508 12 : (
8509 12 : base_inherited_key_overwrite,
8510 12 : test_img("metadata key overwrite 1a")
8511 12 : ),
8512 12 : (base_key, test_img("metadata key 1")),
8513 12 : (base_key_overwrite, test_img("metadata key overwrite 1b")),
8514 12 : ]
8515 12 : );
8516 12 :
8517 12 : // test vectored scan on child timeline
8518 12 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency.clone());
8519 12 : let query =
8520 12 : VersionedKeySpaceQuery::uniform(KeySpace::single(Key::metadata_key_range()), lsn);
8521 12 : let res = child
8522 12 : .get_vectored_impl(query, &mut reconstruct_state, &ctx)
8523 12 : .await?;
8524 12 :
8525 12 : assert_eq!(
8526 12 : res.into_iter()
8527 60 : .map(|(k, v)| (k, v.unwrap()))
8528 12 : .collect::<Vec<_>>(),
8529 12 : vec![
8530 12 : (base_inherited_key, test_img("metadata inherited key 1")),
8531 12 : (
8532 12 : base_inherited_key_child,
8533 12 : test_img("metadata inherited key 2")
8534 12 : ),
8535 12 : (
8536 12 : base_inherited_key_overwrite,
8537 12 : test_img("metadata key overwrite 2a")
8538 12 : ),
8539 12 : (base_key_child, test_img("metadata key 2")),
8540 12 : (base_key_overwrite, test_img("metadata key overwrite 2b")),
8541 12 : ]
8542 12 : );
8543 12 :
8544 12 : Ok(())
8545 12 : }
8546 :
8547 336 : async fn get_vectored_impl_wrapper(
8548 336 : tline: &Arc<Timeline>,
8549 336 : key: Key,
8550 336 : lsn: Lsn,
8551 336 : ctx: &RequestContext,
8552 336 : ) -> Result<Option<Bytes>, GetVectoredError> {
8553 336 : let io_concurrency =
8554 336 : IoConcurrency::spawn_from_conf(tline.conf, tline.gate.enter().unwrap());
8555 336 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
8556 336 : let query = VersionedKeySpaceQuery::uniform(KeySpace::single(key..key.next()), lsn);
8557 336 : let mut res = tline
8558 336 : .get_vectored_impl(query, &mut reconstruct_state, ctx)
8559 336 : .await?;
8560 300 : Ok(res.pop_last().map(|(k, v)| {
8561 192 : assert_eq!(k, key);
8562 192 : v.unwrap()
8563 300 : }))
8564 336 : }
8565 :
8566 : #[tokio::test]
8567 12 : async fn test_metadata_tombstone_reads() -> anyhow::Result<()> {
8568 12 : let harness = TenantHarness::create("test_metadata_tombstone_reads").await?;
8569 12 : let (tenant, ctx) = harness.load().await;
8570 12 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
8571 12 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8572 12 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8573 12 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
8574 12 :
8575 12 : // We emulate the situation that the compaction algorithm creates an image layer that removes the tombstones
8576 12 : // Lsn 0x30 key0, key3, no key1+key2
8577 12 : // Lsn 0x20 key1+key2 tomestones
8578 12 : // Lsn 0x10 key1 in image, key2 in delta
8579 12 : let tline = tenant
8580 12 : .create_test_timeline_with_layers(
8581 12 : TIMELINE_ID,
8582 12 : Lsn(0x10),
8583 12 : DEFAULT_PG_VERSION,
8584 12 : &ctx,
8585 12 : Vec::new(), // in-memory layers
8586 12 : // delta layers
8587 12 : vec![
8588 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
8589 12 : Lsn(0x10)..Lsn(0x20),
8590 12 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8591 12 : ),
8592 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
8593 12 : Lsn(0x20)..Lsn(0x30),
8594 12 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8595 12 : ),
8596 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
8597 12 : Lsn(0x20)..Lsn(0x30),
8598 12 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8599 12 : ),
8600 12 : ],
8601 12 : // image layers
8602 12 : vec![
8603 12 : (Lsn(0x10), vec![(key1, test_img("metadata key 1"))]),
8604 12 : (
8605 12 : Lsn(0x30),
8606 12 : vec![
8607 12 : (key0, test_img("metadata key 0")),
8608 12 : (key3, test_img("metadata key 3")),
8609 12 : ],
8610 12 : ),
8611 12 : ],
8612 12 : Lsn(0x30),
8613 12 : )
8614 12 : .await?;
8615 12 :
8616 12 : let lsn = Lsn(0x30);
8617 12 : let old_lsn = Lsn(0x20);
8618 12 :
8619 12 : assert_eq!(
8620 12 : get_vectored_impl_wrapper(&tline, key0, lsn, &ctx).await?,
8621 12 : Some(test_img("metadata key 0"))
8622 12 : );
8623 12 : assert_eq!(
8624 12 : get_vectored_impl_wrapper(&tline, key1, lsn, &ctx).await?,
8625 12 : None,
8626 12 : );
8627 12 : assert_eq!(
8628 12 : get_vectored_impl_wrapper(&tline, key2, lsn, &ctx).await?,
8629 12 : None,
8630 12 : );
8631 12 : assert_eq!(
8632 12 : get_vectored_impl_wrapper(&tline, key1, old_lsn, &ctx).await?,
8633 12 : Some(Bytes::new()),
8634 12 : );
8635 12 : assert_eq!(
8636 12 : get_vectored_impl_wrapper(&tline, key2, old_lsn, &ctx).await?,
8637 12 : Some(Bytes::new()),
8638 12 : );
8639 12 : assert_eq!(
8640 12 : get_vectored_impl_wrapper(&tline, key3, lsn, &ctx).await?,
8641 12 : Some(test_img("metadata key 3"))
8642 12 : );
8643 12 :
8644 12 : Ok(())
8645 12 : }
8646 :
8647 : #[tokio::test]
8648 12 : async fn test_metadata_tombstone_image_creation() {
8649 12 : let harness = TenantHarness::create("test_metadata_tombstone_image_creation")
8650 12 : .await
8651 12 : .unwrap();
8652 12 : let (tenant, ctx) = harness.load().await;
8653 12 : let io_concurrency = IoConcurrency::spawn_for_test();
8654 12 :
8655 12 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
8656 12 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8657 12 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8658 12 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
8659 12 :
8660 12 : let tline = tenant
8661 12 : .create_test_timeline_with_layers(
8662 12 : TIMELINE_ID,
8663 12 : Lsn(0x10),
8664 12 : DEFAULT_PG_VERSION,
8665 12 : &ctx,
8666 12 : Vec::new(), // in-memory layers
8667 12 : // delta layers
8668 12 : vec![
8669 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
8670 12 : Lsn(0x10)..Lsn(0x20),
8671 12 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8672 12 : ),
8673 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
8674 12 : Lsn(0x20)..Lsn(0x30),
8675 12 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8676 12 : ),
8677 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
8678 12 : Lsn(0x20)..Lsn(0x30),
8679 12 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8680 12 : ),
8681 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
8682 12 : Lsn(0x30)..Lsn(0x40),
8683 12 : vec![
8684 12 : (key0, Lsn(0x30), Value::Image(test_img("metadata key 0"))),
8685 12 : (key3, Lsn(0x30), Value::Image(test_img("metadata key 3"))),
8686 12 : ],
8687 12 : ),
8688 12 : ],
8689 12 : // image layers
8690 12 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
8691 12 : Lsn(0x40),
8692 12 : )
8693 12 : .await
8694 12 : .unwrap();
8695 12 :
8696 12 : let cancel = CancellationToken::new();
8697 12 :
8698 12 : tline
8699 12 : .compact(
8700 12 : &cancel,
8701 12 : {
8702 12 : let mut flags = EnumSet::new();
8703 12 : flags.insert(CompactFlags::ForceImageLayerCreation);
8704 12 : flags.insert(CompactFlags::ForceRepartition);
8705 12 : flags
8706 12 : },
8707 12 : &ctx,
8708 12 : )
8709 12 : .await
8710 12 : .unwrap();
8711 12 :
8712 12 : // Image layers are created at last_record_lsn
8713 12 : let images = tline
8714 12 : .inspect_image_layers(Lsn(0x40), &ctx, io_concurrency.clone())
8715 12 : .await
8716 12 : .unwrap()
8717 12 : .into_iter()
8718 108 : .filter(|(k, _)| k.is_metadata_key())
8719 12 : .collect::<Vec<_>>();
8720 12 : assert_eq!(images.len(), 2); // the image layer should only contain two existing keys, tombstones should be removed.
8721 12 : }
8722 :
8723 : #[tokio::test]
8724 12 : async fn test_metadata_tombstone_empty_image_creation() {
8725 12 : let harness = TenantHarness::create("test_metadata_tombstone_empty_image_creation")
8726 12 : .await
8727 12 : .unwrap();
8728 12 : let (tenant, ctx) = harness.load().await;
8729 12 : let io_concurrency = IoConcurrency::spawn_for_test();
8730 12 :
8731 12 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8732 12 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8733 12 :
8734 12 : let tline = tenant
8735 12 : .create_test_timeline_with_layers(
8736 12 : TIMELINE_ID,
8737 12 : Lsn(0x10),
8738 12 : DEFAULT_PG_VERSION,
8739 12 : &ctx,
8740 12 : Vec::new(), // in-memory layers
8741 12 : // delta layers
8742 12 : vec![
8743 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
8744 12 : Lsn(0x10)..Lsn(0x20),
8745 12 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8746 12 : ),
8747 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
8748 12 : Lsn(0x20)..Lsn(0x30),
8749 12 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8750 12 : ),
8751 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
8752 12 : Lsn(0x20)..Lsn(0x30),
8753 12 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8754 12 : ),
8755 12 : ],
8756 12 : // image layers
8757 12 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
8758 12 : Lsn(0x30),
8759 12 : )
8760 12 : .await
8761 12 : .unwrap();
8762 12 :
8763 12 : let cancel = CancellationToken::new();
8764 12 :
8765 12 : tline
8766 12 : .compact(
8767 12 : &cancel,
8768 12 : {
8769 12 : let mut flags = EnumSet::new();
8770 12 : flags.insert(CompactFlags::ForceImageLayerCreation);
8771 12 : flags.insert(CompactFlags::ForceRepartition);
8772 12 : flags
8773 12 : },
8774 12 : &ctx,
8775 12 : )
8776 12 : .await
8777 12 : .unwrap();
8778 12 :
8779 12 : // Image layers are created at last_record_lsn
8780 12 : let images = tline
8781 12 : .inspect_image_layers(Lsn(0x30), &ctx, io_concurrency.clone())
8782 12 : .await
8783 12 : .unwrap()
8784 12 : .into_iter()
8785 84 : .filter(|(k, _)| k.is_metadata_key())
8786 12 : .collect::<Vec<_>>();
8787 12 : assert_eq!(images.len(), 0); // the image layer should not contain tombstones, or it is not created
8788 12 : }
8789 :
8790 : #[tokio::test]
8791 12 : async fn test_simple_bottom_most_compaction_images() -> anyhow::Result<()> {
8792 12 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_images").await?;
8793 12 : let (tenant, ctx) = harness.load().await;
8794 12 : let io_concurrency = IoConcurrency::spawn_for_test();
8795 12 :
8796 612 : fn get_key(id: u32) -> Key {
8797 612 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8798 612 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8799 612 : key.field6 = id;
8800 612 : key
8801 612 : }
8802 12 :
8803 12 : // We create
8804 12 : // - one bottom-most image layer,
8805 12 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
8806 12 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
8807 12 : // - a delta layer D3 above the horizon.
8808 12 : //
8809 12 : // | D3 |
8810 12 : // | D1 |
8811 12 : // -| |-- gc horizon -----------------
8812 12 : // | | | D2 |
8813 12 : // --------- img layer ------------------
8814 12 : //
8815 12 : // What we should expact from this compaction is:
8816 12 : // | D3 |
8817 12 : // | Part of D1 |
8818 12 : // --------- img layer with D1+D2 at GC horizon------------------
8819 12 :
8820 12 : // img layer at 0x10
8821 12 : let img_layer = (0..10)
8822 120 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8823 12 : .collect_vec();
8824 12 :
8825 12 : let delta1 = vec![
8826 12 : (
8827 12 : get_key(1),
8828 12 : Lsn(0x20),
8829 12 : Value::Image(Bytes::from("value 1@0x20")),
8830 12 : ),
8831 12 : (
8832 12 : get_key(2),
8833 12 : Lsn(0x30),
8834 12 : Value::Image(Bytes::from("value 2@0x30")),
8835 12 : ),
8836 12 : (
8837 12 : get_key(3),
8838 12 : Lsn(0x40),
8839 12 : Value::Image(Bytes::from("value 3@0x40")),
8840 12 : ),
8841 12 : ];
8842 12 : let delta2 = vec![
8843 12 : (
8844 12 : get_key(5),
8845 12 : Lsn(0x20),
8846 12 : Value::Image(Bytes::from("value 5@0x20")),
8847 12 : ),
8848 12 : (
8849 12 : get_key(6),
8850 12 : Lsn(0x20),
8851 12 : Value::Image(Bytes::from("value 6@0x20")),
8852 12 : ),
8853 12 : ];
8854 12 : let delta3 = vec![
8855 12 : (
8856 12 : get_key(8),
8857 12 : Lsn(0x48),
8858 12 : Value::Image(Bytes::from("value 8@0x48")),
8859 12 : ),
8860 12 : (
8861 12 : get_key(9),
8862 12 : Lsn(0x48),
8863 12 : Value::Image(Bytes::from("value 9@0x48")),
8864 12 : ),
8865 12 : ];
8866 12 :
8867 12 : let tline = tenant
8868 12 : .create_test_timeline_with_layers(
8869 12 : TIMELINE_ID,
8870 12 : Lsn(0x10),
8871 12 : DEFAULT_PG_VERSION,
8872 12 : &ctx,
8873 12 : Vec::new(), // in-memory layers
8874 12 : vec![
8875 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
8876 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
8877 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
8878 12 : ], // delta layers
8879 12 : vec![(Lsn(0x10), img_layer)], // image layers
8880 12 : Lsn(0x50),
8881 12 : )
8882 12 : .await?;
8883 12 : {
8884 12 : tline
8885 12 : .applied_gc_cutoff_lsn
8886 12 : .lock_for_write()
8887 12 : .store_and_unlock(Lsn(0x30))
8888 12 : .wait()
8889 12 : .await;
8890 12 : // Update GC info
8891 12 : let mut guard = tline.gc_info.write().unwrap();
8892 12 : guard.cutoffs.time = Lsn(0x30);
8893 12 : guard.cutoffs.space = Lsn(0x30);
8894 12 : }
8895 12 :
8896 12 : let expected_result = [
8897 12 : Bytes::from_static(b"value 0@0x10"),
8898 12 : Bytes::from_static(b"value 1@0x20"),
8899 12 : Bytes::from_static(b"value 2@0x30"),
8900 12 : Bytes::from_static(b"value 3@0x40"),
8901 12 : Bytes::from_static(b"value 4@0x10"),
8902 12 : Bytes::from_static(b"value 5@0x20"),
8903 12 : Bytes::from_static(b"value 6@0x20"),
8904 12 : Bytes::from_static(b"value 7@0x10"),
8905 12 : Bytes::from_static(b"value 8@0x48"),
8906 12 : Bytes::from_static(b"value 9@0x48"),
8907 12 : ];
8908 12 :
8909 120 : for (idx, expected) in expected_result.iter().enumerate() {
8910 120 : assert_eq!(
8911 120 : tline
8912 120 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8913 120 : .await
8914 120 : .unwrap(),
8915 12 : expected
8916 12 : );
8917 12 : }
8918 12 :
8919 12 : let cancel = CancellationToken::new();
8920 12 : tline
8921 12 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8922 12 : .await
8923 12 : .unwrap();
8924 12 :
8925 120 : for (idx, expected) in expected_result.iter().enumerate() {
8926 120 : assert_eq!(
8927 120 : tline
8928 120 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8929 120 : .await
8930 120 : .unwrap(),
8931 12 : expected
8932 12 : );
8933 12 : }
8934 12 :
8935 12 : // Check if the image layer at the GC horizon contains exactly what we want
8936 12 : let image_at_gc_horizon = tline
8937 12 : .inspect_image_layers(Lsn(0x30), &ctx, io_concurrency.clone())
8938 12 : .await
8939 12 : .unwrap()
8940 12 : .into_iter()
8941 204 : .filter(|(k, _)| k.is_metadata_key())
8942 12 : .collect::<Vec<_>>();
8943 12 :
8944 12 : assert_eq!(image_at_gc_horizon.len(), 10);
8945 12 : let expected_result = [
8946 12 : Bytes::from_static(b"value 0@0x10"),
8947 12 : Bytes::from_static(b"value 1@0x20"),
8948 12 : Bytes::from_static(b"value 2@0x30"),
8949 12 : Bytes::from_static(b"value 3@0x10"),
8950 12 : Bytes::from_static(b"value 4@0x10"),
8951 12 : Bytes::from_static(b"value 5@0x20"),
8952 12 : Bytes::from_static(b"value 6@0x20"),
8953 12 : Bytes::from_static(b"value 7@0x10"),
8954 12 : Bytes::from_static(b"value 8@0x10"),
8955 12 : Bytes::from_static(b"value 9@0x10"),
8956 12 : ];
8957 132 : for idx in 0..10 {
8958 120 : assert_eq!(
8959 120 : image_at_gc_horizon[idx],
8960 120 : (get_key(idx as u32), expected_result[idx].clone())
8961 120 : );
8962 12 : }
8963 12 :
8964 12 : // Check if old layers are removed / new layers have the expected LSN
8965 12 : let all_layers = inspect_and_sort(&tline, None).await;
8966 12 : assert_eq!(
8967 12 : all_layers,
8968 12 : vec![
8969 12 : // Image layer at GC horizon
8970 12 : PersistentLayerKey {
8971 12 : key_range: Key::MIN..Key::MAX,
8972 12 : lsn_range: Lsn(0x30)..Lsn(0x31),
8973 12 : is_delta: false
8974 12 : },
8975 12 : // The delta layer below the horizon
8976 12 : PersistentLayerKey {
8977 12 : key_range: get_key(3)..get_key(4),
8978 12 : lsn_range: Lsn(0x30)..Lsn(0x48),
8979 12 : is_delta: true
8980 12 : },
8981 12 : // The delta3 layer that should not be picked for the compaction
8982 12 : PersistentLayerKey {
8983 12 : key_range: get_key(8)..get_key(10),
8984 12 : lsn_range: Lsn(0x48)..Lsn(0x50),
8985 12 : is_delta: true
8986 12 : }
8987 12 : ]
8988 12 : );
8989 12 :
8990 12 : // increase GC horizon and compact again
8991 12 : {
8992 12 : tline
8993 12 : .applied_gc_cutoff_lsn
8994 12 : .lock_for_write()
8995 12 : .store_and_unlock(Lsn(0x40))
8996 12 : .wait()
8997 12 : .await;
8998 12 : // Update GC info
8999 12 : let mut guard = tline.gc_info.write().unwrap();
9000 12 : guard.cutoffs.time = Lsn(0x40);
9001 12 : guard.cutoffs.space = Lsn(0x40);
9002 12 : }
9003 12 : tline
9004 12 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9005 12 : .await
9006 12 : .unwrap();
9007 12 :
9008 12 : Ok(())
9009 12 : }
9010 :
9011 : #[cfg(feature = "testing")]
9012 : #[tokio::test]
9013 12 : async fn test_neon_test_record() -> anyhow::Result<()> {
9014 12 : let harness = TenantHarness::create("test_neon_test_record").await?;
9015 12 : let (tenant, ctx) = harness.load().await;
9016 12 :
9017 204 : fn get_key(id: u32) -> Key {
9018 204 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9019 204 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9020 204 : key.field6 = id;
9021 204 : key
9022 204 : }
9023 12 :
9024 12 : let delta1 = vec![
9025 12 : (
9026 12 : get_key(1),
9027 12 : Lsn(0x20),
9028 12 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
9029 12 : ),
9030 12 : (
9031 12 : get_key(1),
9032 12 : Lsn(0x30),
9033 12 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
9034 12 : ),
9035 12 : (get_key(2), Lsn(0x10), Value::Image("0x10".into())),
9036 12 : (
9037 12 : get_key(2),
9038 12 : Lsn(0x20),
9039 12 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
9040 12 : ),
9041 12 : (
9042 12 : get_key(2),
9043 12 : Lsn(0x30),
9044 12 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
9045 12 : ),
9046 12 : (get_key(3), Lsn(0x10), Value::Image("0x10".into())),
9047 12 : (
9048 12 : get_key(3),
9049 12 : Lsn(0x20),
9050 12 : Value::WalRecord(NeonWalRecord::wal_clear("c")),
9051 12 : ),
9052 12 : (get_key(4), Lsn(0x10), Value::Image("0x10".into())),
9053 12 : (
9054 12 : get_key(4),
9055 12 : Lsn(0x20),
9056 12 : Value::WalRecord(NeonWalRecord::wal_init("i")),
9057 12 : ),
9058 12 : (
9059 12 : get_key(4),
9060 12 : Lsn(0x30),
9061 12 : Value::WalRecord(NeonWalRecord::wal_append_conditional("j", "i")),
9062 12 : ),
9063 12 : (
9064 12 : get_key(5),
9065 12 : Lsn(0x20),
9066 12 : Value::WalRecord(NeonWalRecord::wal_init("1")),
9067 12 : ),
9068 12 : (
9069 12 : get_key(5),
9070 12 : Lsn(0x30),
9071 12 : Value::WalRecord(NeonWalRecord::wal_append_conditional("j", "2")),
9072 12 : ),
9073 12 : ];
9074 12 : let image1 = vec![(get_key(1), "0x10".into())];
9075 12 :
9076 12 : let tline = tenant
9077 12 : .create_test_timeline_with_layers(
9078 12 : TIMELINE_ID,
9079 12 : Lsn(0x10),
9080 12 : DEFAULT_PG_VERSION,
9081 12 : &ctx,
9082 12 : Vec::new(), // in-memory layers
9083 12 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
9084 12 : Lsn(0x10)..Lsn(0x40),
9085 12 : delta1,
9086 12 : )], // delta layers
9087 12 : vec![(Lsn(0x10), image1)], // image layers
9088 12 : Lsn(0x50),
9089 12 : )
9090 12 : .await?;
9091 12 :
9092 12 : assert_eq!(
9093 12 : tline.get(get_key(1), Lsn(0x50), &ctx).await?,
9094 12 : Bytes::from_static(b"0x10,0x20,0x30")
9095 12 : );
9096 12 : assert_eq!(
9097 12 : tline.get(get_key(2), Lsn(0x50), &ctx).await?,
9098 12 : Bytes::from_static(b"0x10,0x20,0x30")
9099 12 : );
9100 12 :
9101 12 : // Need to remove the limit of "Neon WAL redo requires base image".
9102 12 :
9103 12 : assert_eq!(
9104 12 : tline.get(get_key(3), Lsn(0x50), &ctx).await?,
9105 12 : Bytes::from_static(b"c")
9106 12 : );
9107 12 : assert_eq!(
9108 12 : tline.get(get_key(4), Lsn(0x50), &ctx).await?,
9109 12 : Bytes::from_static(b"ij")
9110 12 : );
9111 12 :
9112 12 : // Manual testing required: currently, read errors will panic the process in debug mode. So we
9113 12 : // cannot enable this assertion in the unit test.
9114 12 : // assert!(tline.get(get_key(5), Lsn(0x50), &ctx).await.is_err());
9115 12 :
9116 12 : Ok(())
9117 12 : }
9118 :
9119 : #[tokio::test(start_paused = true)]
9120 12 : async fn test_lsn_lease() -> anyhow::Result<()> {
9121 12 : let (tenant, ctx) = TenantHarness::create("test_lsn_lease")
9122 12 : .await
9123 12 : .unwrap()
9124 12 : .load()
9125 12 : .await;
9126 12 : // Advance to the lsn lease deadline so that GC is not blocked by
9127 12 : // initial transition into AttachedSingle.
9128 12 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
9129 12 : tokio::time::resume();
9130 12 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
9131 12 :
9132 12 : let end_lsn = Lsn(0x100);
9133 12 : let image_layers = (0x20..=0x90)
9134 12 : .step_by(0x10)
9135 96 : .map(|n| {
9136 96 : (
9137 96 : Lsn(n),
9138 96 : vec![(key, test_img(&format!("data key at {:x}", n)))],
9139 96 : )
9140 96 : })
9141 12 : .collect();
9142 12 :
9143 12 : let timeline = tenant
9144 12 : .create_test_timeline_with_layers(
9145 12 : TIMELINE_ID,
9146 12 : Lsn(0x10),
9147 12 : DEFAULT_PG_VERSION,
9148 12 : &ctx,
9149 12 : Vec::new(), // in-memory layers
9150 12 : Vec::new(),
9151 12 : image_layers,
9152 12 : end_lsn,
9153 12 : )
9154 12 : .await?;
9155 12 :
9156 12 : let leased_lsns = [0x30, 0x50, 0x70];
9157 12 : let mut leases = Vec::new();
9158 36 : leased_lsns.iter().for_each(|n| {
9159 36 : leases.push(
9160 36 : timeline
9161 36 : .init_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)
9162 36 : .expect("lease request should succeed"),
9163 36 : );
9164 36 : });
9165 12 :
9166 12 : let updated_lease_0 = timeline
9167 12 : .renew_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)
9168 12 : .expect("lease renewal should succeed");
9169 12 : assert_eq!(
9170 12 : updated_lease_0.valid_until, leases[0].valid_until,
9171 12 : " Renewing with shorter lease should not change the lease."
9172 12 : );
9173 12 :
9174 12 : let updated_lease_1 = timeline
9175 12 : .renew_lsn_lease(
9176 12 : Lsn(leased_lsns[1]),
9177 12 : timeline.get_lsn_lease_length() * 2,
9178 12 : &ctx,
9179 12 : )
9180 12 : .expect("lease renewal should succeed");
9181 12 : assert!(
9182 12 : updated_lease_1.valid_until > leases[1].valid_until,
9183 12 : "Renewing with a long lease should renew lease with later expiration time."
9184 12 : );
9185 12 :
9186 12 : // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
9187 12 : info!(
9188 12 : "applied_gc_cutoff_lsn: {}",
9189 0 : *timeline.get_applied_gc_cutoff_lsn()
9190 12 : );
9191 12 : timeline.force_set_disk_consistent_lsn(end_lsn);
9192 12 :
9193 12 : let res = tenant
9194 12 : .gc_iteration(
9195 12 : Some(TIMELINE_ID),
9196 12 : 0,
9197 12 : Duration::ZERO,
9198 12 : &CancellationToken::new(),
9199 12 : &ctx,
9200 12 : )
9201 12 : .await
9202 12 : .unwrap();
9203 12 :
9204 12 : // Keeping everything <= Lsn(0x80) b/c leases:
9205 12 : // 0/10: initdb layer
9206 12 : // (0/20..=0/70).step_by(0x10): image layers added when creating the timeline.
9207 12 : assert_eq!(res.layers_needed_by_leases, 7);
9208 12 : // Keeping 0/90 b/c it is the latest layer.
9209 12 : assert_eq!(res.layers_not_updated, 1);
9210 12 : // Removed 0/80.
9211 12 : assert_eq!(res.layers_removed, 1);
9212 12 :
9213 12 : // Make lease on a already GC-ed LSN.
9214 12 : // 0/80 does not have a valid lease + is below latest_gc_cutoff
9215 12 : assert!(Lsn(0x80) < *timeline.get_applied_gc_cutoff_lsn());
9216 12 : timeline
9217 12 : .init_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx)
9218 12 : .expect_err("lease request on GC-ed LSN should fail");
9219 12 :
9220 12 : // Should still be able to renew a currently valid lease
9221 12 : // Assumption: original lease to is still valid for 0/50.
9222 12 : // (use `Timeline::init_lsn_lease` for testing so it always does validation)
9223 12 : timeline
9224 12 : .init_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)
9225 12 : .expect("lease renewal with validation should succeed");
9226 12 :
9227 12 : Ok(())
9228 12 : }
9229 :
9230 : #[cfg(feature = "testing")]
9231 : #[tokio::test]
9232 12 : async fn test_simple_bottom_most_compaction_deltas_1() -> anyhow::Result<()> {
9233 12 : test_simple_bottom_most_compaction_deltas_helper(
9234 12 : "test_simple_bottom_most_compaction_deltas_1",
9235 12 : false,
9236 12 : )
9237 12 : .await
9238 12 : }
9239 :
9240 : #[cfg(feature = "testing")]
9241 : #[tokio::test]
9242 12 : async fn test_simple_bottom_most_compaction_deltas_2() -> anyhow::Result<()> {
9243 12 : test_simple_bottom_most_compaction_deltas_helper(
9244 12 : "test_simple_bottom_most_compaction_deltas_2",
9245 12 : true,
9246 12 : )
9247 12 : .await
9248 12 : }
9249 :
9250 : #[cfg(feature = "testing")]
9251 24 : async fn test_simple_bottom_most_compaction_deltas_helper(
9252 24 : test_name: &'static str,
9253 24 : use_delta_bottom_layer: bool,
9254 24 : ) -> anyhow::Result<()> {
9255 24 : let harness = TenantHarness::create(test_name).await?;
9256 24 : let (tenant, ctx) = harness.load().await;
9257 :
9258 1656 : fn get_key(id: u32) -> Key {
9259 1656 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9260 1656 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9261 1656 : key.field6 = id;
9262 1656 : key
9263 1656 : }
9264 :
9265 : // We create
9266 : // - one bottom-most image layer,
9267 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
9268 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
9269 : // - a delta layer D3 above the horizon.
9270 : //
9271 : // | D3 |
9272 : // | D1 |
9273 : // -| |-- gc horizon -----------------
9274 : // | | | D2 |
9275 : // --------- img layer ------------------
9276 : //
9277 : // What we should expact from this compaction is:
9278 : // | D3 |
9279 : // | Part of D1 |
9280 : // --------- img layer with D1+D2 at GC horizon------------------
9281 :
9282 : // img layer at 0x10
9283 24 : let img_layer = (0..10)
9284 240 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9285 24 : .collect_vec();
9286 24 : // or, delta layer at 0x10 if `use_delta_bottom_layer` is true
9287 24 : let delta4 = (0..10)
9288 240 : .map(|id| {
9289 240 : (
9290 240 : get_key(id),
9291 240 : Lsn(0x08),
9292 240 : Value::WalRecord(NeonWalRecord::wal_init(format!("value {id}@0x10"))),
9293 240 : )
9294 240 : })
9295 24 : .collect_vec();
9296 24 :
9297 24 : let delta1 = vec![
9298 24 : (
9299 24 : get_key(1),
9300 24 : Lsn(0x20),
9301 24 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9302 24 : ),
9303 24 : (
9304 24 : get_key(2),
9305 24 : Lsn(0x30),
9306 24 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9307 24 : ),
9308 24 : (
9309 24 : get_key(3),
9310 24 : Lsn(0x28),
9311 24 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9312 24 : ),
9313 24 : (
9314 24 : get_key(3),
9315 24 : Lsn(0x30),
9316 24 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9317 24 : ),
9318 24 : (
9319 24 : get_key(3),
9320 24 : Lsn(0x40),
9321 24 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9322 24 : ),
9323 24 : ];
9324 24 : let delta2 = vec![
9325 24 : (
9326 24 : get_key(5),
9327 24 : Lsn(0x20),
9328 24 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9329 24 : ),
9330 24 : (
9331 24 : get_key(6),
9332 24 : Lsn(0x20),
9333 24 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9334 24 : ),
9335 24 : ];
9336 24 : let delta3 = vec![
9337 24 : (
9338 24 : get_key(8),
9339 24 : Lsn(0x48),
9340 24 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9341 24 : ),
9342 24 : (
9343 24 : get_key(9),
9344 24 : Lsn(0x48),
9345 24 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9346 24 : ),
9347 24 : ];
9348 :
9349 24 : let tline = if use_delta_bottom_layer {
9350 12 : tenant
9351 12 : .create_test_timeline_with_layers(
9352 12 : TIMELINE_ID,
9353 12 : Lsn(0x08),
9354 12 : DEFAULT_PG_VERSION,
9355 12 : &ctx,
9356 12 : Vec::new(), // in-memory layers
9357 12 : vec![
9358 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
9359 12 : Lsn(0x08)..Lsn(0x10),
9360 12 : delta4,
9361 12 : ),
9362 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
9363 12 : Lsn(0x20)..Lsn(0x48),
9364 12 : delta1,
9365 12 : ),
9366 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
9367 12 : Lsn(0x20)..Lsn(0x48),
9368 12 : delta2,
9369 12 : ),
9370 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
9371 12 : Lsn(0x48)..Lsn(0x50),
9372 12 : delta3,
9373 12 : ),
9374 12 : ], // delta layers
9375 12 : vec![], // image layers
9376 12 : Lsn(0x50),
9377 12 : )
9378 12 : .await?
9379 : } else {
9380 12 : tenant
9381 12 : .create_test_timeline_with_layers(
9382 12 : TIMELINE_ID,
9383 12 : Lsn(0x10),
9384 12 : DEFAULT_PG_VERSION,
9385 12 : &ctx,
9386 12 : Vec::new(), // in-memory layers
9387 12 : vec![
9388 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
9389 12 : Lsn(0x10)..Lsn(0x48),
9390 12 : delta1,
9391 12 : ),
9392 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
9393 12 : Lsn(0x10)..Lsn(0x48),
9394 12 : delta2,
9395 12 : ),
9396 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
9397 12 : Lsn(0x48)..Lsn(0x50),
9398 12 : delta3,
9399 12 : ),
9400 12 : ], // delta layers
9401 12 : vec![(Lsn(0x10), img_layer)], // image layers
9402 12 : Lsn(0x50),
9403 12 : )
9404 12 : .await?
9405 : };
9406 : {
9407 24 : tline
9408 24 : .applied_gc_cutoff_lsn
9409 24 : .lock_for_write()
9410 24 : .store_and_unlock(Lsn(0x30))
9411 24 : .wait()
9412 24 : .await;
9413 : // Update GC info
9414 24 : let mut guard = tline.gc_info.write().unwrap();
9415 24 : *guard = GcInfo {
9416 24 : retain_lsns: vec![],
9417 24 : cutoffs: GcCutoffs {
9418 24 : time: Lsn(0x30),
9419 24 : space: Lsn(0x30),
9420 24 : },
9421 24 : leases: Default::default(),
9422 24 : within_ancestor_pitr: false,
9423 24 : };
9424 24 : }
9425 24 :
9426 24 : let expected_result = [
9427 24 : Bytes::from_static(b"value 0@0x10"),
9428 24 : Bytes::from_static(b"value 1@0x10@0x20"),
9429 24 : Bytes::from_static(b"value 2@0x10@0x30"),
9430 24 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9431 24 : Bytes::from_static(b"value 4@0x10"),
9432 24 : Bytes::from_static(b"value 5@0x10@0x20"),
9433 24 : Bytes::from_static(b"value 6@0x10@0x20"),
9434 24 : Bytes::from_static(b"value 7@0x10"),
9435 24 : Bytes::from_static(b"value 8@0x10@0x48"),
9436 24 : Bytes::from_static(b"value 9@0x10@0x48"),
9437 24 : ];
9438 24 :
9439 24 : let expected_result_at_gc_horizon = [
9440 24 : Bytes::from_static(b"value 0@0x10"),
9441 24 : Bytes::from_static(b"value 1@0x10@0x20"),
9442 24 : Bytes::from_static(b"value 2@0x10@0x30"),
9443 24 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
9444 24 : Bytes::from_static(b"value 4@0x10"),
9445 24 : Bytes::from_static(b"value 5@0x10@0x20"),
9446 24 : Bytes::from_static(b"value 6@0x10@0x20"),
9447 24 : Bytes::from_static(b"value 7@0x10"),
9448 24 : Bytes::from_static(b"value 8@0x10"),
9449 24 : Bytes::from_static(b"value 9@0x10"),
9450 24 : ];
9451 :
9452 264 : for idx in 0..10 {
9453 240 : assert_eq!(
9454 240 : tline
9455 240 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9456 240 : .await
9457 240 : .unwrap(),
9458 240 : &expected_result[idx]
9459 : );
9460 240 : assert_eq!(
9461 240 : tline
9462 240 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
9463 240 : .await
9464 240 : .unwrap(),
9465 240 : &expected_result_at_gc_horizon[idx]
9466 : );
9467 : }
9468 :
9469 24 : let cancel = CancellationToken::new();
9470 24 : tline
9471 24 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9472 24 : .await
9473 24 : .unwrap();
9474 :
9475 264 : for idx in 0..10 {
9476 240 : assert_eq!(
9477 240 : tline
9478 240 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9479 240 : .await
9480 240 : .unwrap(),
9481 240 : &expected_result[idx]
9482 : );
9483 240 : assert_eq!(
9484 240 : tline
9485 240 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
9486 240 : .await
9487 240 : .unwrap(),
9488 240 : &expected_result_at_gc_horizon[idx]
9489 : );
9490 : }
9491 :
9492 : // increase GC horizon and compact again
9493 : {
9494 24 : tline
9495 24 : .applied_gc_cutoff_lsn
9496 24 : .lock_for_write()
9497 24 : .store_and_unlock(Lsn(0x40))
9498 24 : .wait()
9499 24 : .await;
9500 : // Update GC info
9501 24 : let mut guard = tline.gc_info.write().unwrap();
9502 24 : guard.cutoffs.time = Lsn(0x40);
9503 24 : guard.cutoffs.space = Lsn(0x40);
9504 24 : }
9505 24 : tline
9506 24 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9507 24 : .await
9508 24 : .unwrap();
9509 24 :
9510 24 : Ok(())
9511 24 : }
9512 :
9513 : #[cfg(feature = "testing")]
9514 : #[tokio::test]
9515 12 : async fn test_generate_key_retention() -> anyhow::Result<()> {
9516 12 : let harness = TenantHarness::create("test_generate_key_retention").await?;
9517 12 : let (tenant, ctx) = harness.load().await;
9518 12 : let tline = tenant
9519 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
9520 12 : .await?;
9521 12 : tline.force_advance_lsn(Lsn(0x70));
9522 12 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
9523 12 : let history = vec![
9524 12 : (
9525 12 : key,
9526 12 : Lsn(0x10),
9527 12 : Value::WalRecord(NeonWalRecord::wal_init("0x10")),
9528 12 : ),
9529 12 : (
9530 12 : key,
9531 12 : Lsn(0x20),
9532 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9533 12 : ),
9534 12 : (
9535 12 : key,
9536 12 : Lsn(0x30),
9537 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9538 12 : ),
9539 12 : (
9540 12 : key,
9541 12 : Lsn(0x40),
9542 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9543 12 : ),
9544 12 : (
9545 12 : key,
9546 12 : Lsn(0x50),
9547 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
9548 12 : ),
9549 12 : (
9550 12 : key,
9551 12 : Lsn(0x60),
9552 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9553 12 : ),
9554 12 : (
9555 12 : key,
9556 12 : Lsn(0x70),
9557 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9558 12 : ),
9559 12 : (
9560 12 : key,
9561 12 : Lsn(0x80),
9562 12 : Value::Image(Bytes::copy_from_slice(
9563 12 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9564 12 : )),
9565 12 : ),
9566 12 : (
9567 12 : key,
9568 12 : Lsn(0x90),
9569 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9570 12 : ),
9571 12 : ];
9572 12 : let res = tline
9573 12 : .generate_key_retention(
9574 12 : key,
9575 12 : &history,
9576 12 : Lsn(0x60),
9577 12 : &[Lsn(0x20), Lsn(0x40), Lsn(0x50)],
9578 12 : 3,
9579 12 : None,
9580 12 : true,
9581 12 : )
9582 12 : .await
9583 12 : .unwrap();
9584 12 : let expected_res = KeyHistoryRetention {
9585 12 : below_horizon: vec![
9586 12 : (
9587 12 : Lsn(0x20),
9588 12 : KeyLogAtLsn(vec![(
9589 12 : Lsn(0x20),
9590 12 : Value::Image(Bytes::from_static(b"0x10;0x20")),
9591 12 : )]),
9592 12 : ),
9593 12 : (
9594 12 : Lsn(0x40),
9595 12 : KeyLogAtLsn(vec![
9596 12 : (
9597 12 : Lsn(0x30),
9598 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9599 12 : ),
9600 12 : (
9601 12 : Lsn(0x40),
9602 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9603 12 : ),
9604 12 : ]),
9605 12 : ),
9606 12 : (
9607 12 : Lsn(0x50),
9608 12 : KeyLogAtLsn(vec![(
9609 12 : Lsn(0x50),
9610 12 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40;0x50")),
9611 12 : )]),
9612 12 : ),
9613 12 : (
9614 12 : Lsn(0x60),
9615 12 : KeyLogAtLsn(vec![(
9616 12 : Lsn(0x60),
9617 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9618 12 : )]),
9619 12 : ),
9620 12 : ],
9621 12 : above_horizon: KeyLogAtLsn(vec![
9622 12 : (
9623 12 : Lsn(0x70),
9624 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9625 12 : ),
9626 12 : (
9627 12 : Lsn(0x80),
9628 12 : Value::Image(Bytes::copy_from_slice(
9629 12 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9630 12 : )),
9631 12 : ),
9632 12 : (
9633 12 : Lsn(0x90),
9634 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9635 12 : ),
9636 12 : ]),
9637 12 : };
9638 12 : assert_eq!(res, expected_res);
9639 12 :
9640 12 : // We expect GC-compaction to run with the original GC. This would create a situation that
9641 12 : // the original GC algorithm removes some delta layers b/c there are full image coverage,
9642 12 : // therefore causing some keys to have an incomplete history below the lowest retain LSN.
9643 12 : // For example, we have
9644 12 : // ```plain
9645 12 : // init delta @ 0x10, image @ 0x20, delta @ 0x30 (gc_horizon), image @ 0x40.
9646 12 : // ```
9647 12 : // Now the GC horizon moves up, and we have
9648 12 : // ```plain
9649 12 : // init delta @ 0x10, image @ 0x20, delta @ 0x30, image @ 0x40 (gc_horizon)
9650 12 : // ```
9651 12 : // The original GC algorithm kicks in, and removes delta @ 0x10, image @ 0x20.
9652 12 : // We will end up with
9653 12 : // ```plain
9654 12 : // delta @ 0x30, image @ 0x40 (gc_horizon)
9655 12 : // ```
9656 12 : // Now we run the GC-compaction, and this key does not have a full history.
9657 12 : // We should be able to handle this partial history and drop everything before the
9658 12 : // gc_horizon image.
9659 12 :
9660 12 : let history = vec![
9661 12 : (
9662 12 : key,
9663 12 : Lsn(0x20),
9664 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9665 12 : ),
9666 12 : (
9667 12 : key,
9668 12 : Lsn(0x30),
9669 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9670 12 : ),
9671 12 : (
9672 12 : key,
9673 12 : Lsn(0x40),
9674 12 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
9675 12 : ),
9676 12 : (
9677 12 : key,
9678 12 : Lsn(0x50),
9679 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
9680 12 : ),
9681 12 : (
9682 12 : key,
9683 12 : Lsn(0x60),
9684 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9685 12 : ),
9686 12 : (
9687 12 : key,
9688 12 : Lsn(0x70),
9689 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9690 12 : ),
9691 12 : (
9692 12 : key,
9693 12 : Lsn(0x80),
9694 12 : Value::Image(Bytes::copy_from_slice(
9695 12 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9696 12 : )),
9697 12 : ),
9698 12 : (
9699 12 : key,
9700 12 : Lsn(0x90),
9701 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9702 12 : ),
9703 12 : ];
9704 12 : let res = tline
9705 12 : .generate_key_retention(
9706 12 : key,
9707 12 : &history,
9708 12 : Lsn(0x60),
9709 12 : &[Lsn(0x40), Lsn(0x50)],
9710 12 : 3,
9711 12 : None,
9712 12 : true,
9713 12 : )
9714 12 : .await
9715 12 : .unwrap();
9716 12 : let expected_res = KeyHistoryRetention {
9717 12 : below_horizon: vec![
9718 12 : (
9719 12 : Lsn(0x40),
9720 12 : KeyLogAtLsn(vec![(
9721 12 : Lsn(0x40),
9722 12 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
9723 12 : )]),
9724 12 : ),
9725 12 : (
9726 12 : Lsn(0x50),
9727 12 : KeyLogAtLsn(vec![(
9728 12 : Lsn(0x50),
9729 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
9730 12 : )]),
9731 12 : ),
9732 12 : (
9733 12 : Lsn(0x60),
9734 12 : KeyLogAtLsn(vec![(
9735 12 : Lsn(0x60),
9736 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9737 12 : )]),
9738 12 : ),
9739 12 : ],
9740 12 : above_horizon: KeyLogAtLsn(vec![
9741 12 : (
9742 12 : Lsn(0x70),
9743 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9744 12 : ),
9745 12 : (
9746 12 : Lsn(0x80),
9747 12 : Value::Image(Bytes::copy_from_slice(
9748 12 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9749 12 : )),
9750 12 : ),
9751 12 : (
9752 12 : Lsn(0x90),
9753 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9754 12 : ),
9755 12 : ]),
9756 12 : };
9757 12 : assert_eq!(res, expected_res);
9758 12 :
9759 12 : // In case of branch compaction, the branch itself does not have the full history, and we need to provide
9760 12 : // the ancestor image in the test case.
9761 12 :
9762 12 : let history = vec![
9763 12 : (
9764 12 : key,
9765 12 : Lsn(0x20),
9766 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9767 12 : ),
9768 12 : (
9769 12 : key,
9770 12 : Lsn(0x30),
9771 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9772 12 : ),
9773 12 : (
9774 12 : key,
9775 12 : Lsn(0x40),
9776 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9777 12 : ),
9778 12 : (
9779 12 : key,
9780 12 : Lsn(0x70),
9781 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9782 12 : ),
9783 12 : ];
9784 12 : let res = tline
9785 12 : .generate_key_retention(
9786 12 : key,
9787 12 : &history,
9788 12 : Lsn(0x60),
9789 12 : &[],
9790 12 : 3,
9791 12 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
9792 12 : true,
9793 12 : )
9794 12 : .await
9795 12 : .unwrap();
9796 12 : let expected_res = KeyHistoryRetention {
9797 12 : below_horizon: vec![(
9798 12 : Lsn(0x60),
9799 12 : KeyLogAtLsn(vec![(
9800 12 : Lsn(0x60),
9801 12 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")), // use the ancestor image to reconstruct the page
9802 12 : )]),
9803 12 : )],
9804 12 : above_horizon: KeyLogAtLsn(vec![(
9805 12 : Lsn(0x70),
9806 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9807 12 : )]),
9808 12 : };
9809 12 : assert_eq!(res, expected_res);
9810 12 :
9811 12 : let history = vec![
9812 12 : (
9813 12 : key,
9814 12 : Lsn(0x20),
9815 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9816 12 : ),
9817 12 : (
9818 12 : key,
9819 12 : Lsn(0x40),
9820 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9821 12 : ),
9822 12 : (
9823 12 : key,
9824 12 : Lsn(0x60),
9825 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9826 12 : ),
9827 12 : (
9828 12 : key,
9829 12 : Lsn(0x70),
9830 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9831 12 : ),
9832 12 : ];
9833 12 : let res = tline
9834 12 : .generate_key_retention(
9835 12 : key,
9836 12 : &history,
9837 12 : Lsn(0x60),
9838 12 : &[Lsn(0x30)],
9839 12 : 3,
9840 12 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
9841 12 : true,
9842 12 : )
9843 12 : .await
9844 12 : .unwrap();
9845 12 : let expected_res = KeyHistoryRetention {
9846 12 : below_horizon: vec![
9847 12 : (
9848 12 : Lsn(0x30),
9849 12 : KeyLogAtLsn(vec![(
9850 12 : Lsn(0x20),
9851 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9852 12 : )]),
9853 12 : ),
9854 12 : (
9855 12 : Lsn(0x60),
9856 12 : KeyLogAtLsn(vec![(
9857 12 : Lsn(0x60),
9858 12 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x40;0x60")),
9859 12 : )]),
9860 12 : ),
9861 12 : ],
9862 12 : above_horizon: KeyLogAtLsn(vec![(
9863 12 : Lsn(0x70),
9864 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9865 12 : )]),
9866 12 : };
9867 12 : assert_eq!(res, expected_res);
9868 12 :
9869 12 : Ok(())
9870 12 : }
9871 :
9872 : #[cfg(feature = "testing")]
9873 : #[tokio::test]
9874 12 : async fn test_simple_bottom_most_compaction_with_retain_lsns() -> anyhow::Result<()> {
9875 12 : let harness =
9876 12 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns").await?;
9877 12 : let (tenant, ctx) = harness.load().await;
9878 12 :
9879 3108 : fn get_key(id: u32) -> Key {
9880 3108 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9881 3108 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9882 3108 : key.field6 = id;
9883 3108 : key
9884 3108 : }
9885 12 :
9886 12 : let img_layer = (0..10)
9887 120 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9888 12 : .collect_vec();
9889 12 :
9890 12 : let delta1 = vec![
9891 12 : (
9892 12 : get_key(1),
9893 12 : Lsn(0x20),
9894 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9895 12 : ),
9896 12 : (
9897 12 : get_key(2),
9898 12 : Lsn(0x30),
9899 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9900 12 : ),
9901 12 : (
9902 12 : get_key(3),
9903 12 : Lsn(0x28),
9904 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9905 12 : ),
9906 12 : (
9907 12 : get_key(3),
9908 12 : Lsn(0x30),
9909 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9910 12 : ),
9911 12 : (
9912 12 : get_key(3),
9913 12 : Lsn(0x40),
9914 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9915 12 : ),
9916 12 : ];
9917 12 : let delta2 = vec![
9918 12 : (
9919 12 : get_key(5),
9920 12 : Lsn(0x20),
9921 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9922 12 : ),
9923 12 : (
9924 12 : get_key(6),
9925 12 : Lsn(0x20),
9926 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9927 12 : ),
9928 12 : ];
9929 12 : let delta3 = vec![
9930 12 : (
9931 12 : get_key(8),
9932 12 : Lsn(0x48),
9933 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9934 12 : ),
9935 12 : (
9936 12 : get_key(9),
9937 12 : Lsn(0x48),
9938 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9939 12 : ),
9940 12 : ];
9941 12 :
9942 12 : let tline = tenant
9943 12 : .create_test_timeline_with_layers(
9944 12 : TIMELINE_ID,
9945 12 : Lsn(0x10),
9946 12 : DEFAULT_PG_VERSION,
9947 12 : &ctx,
9948 12 : Vec::new(), // in-memory layers
9949 12 : vec![
9950 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
9951 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
9952 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
9953 12 : ], // delta layers
9954 12 : vec![(Lsn(0x10), img_layer)], // image layers
9955 12 : Lsn(0x50),
9956 12 : )
9957 12 : .await?;
9958 12 : {
9959 12 : tline
9960 12 : .applied_gc_cutoff_lsn
9961 12 : .lock_for_write()
9962 12 : .store_and_unlock(Lsn(0x30))
9963 12 : .wait()
9964 12 : .await;
9965 12 : // Update GC info
9966 12 : let mut guard = tline.gc_info.write().unwrap();
9967 12 : *guard = GcInfo {
9968 12 : retain_lsns: vec![
9969 12 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
9970 12 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
9971 12 : ],
9972 12 : cutoffs: GcCutoffs {
9973 12 : time: Lsn(0x30),
9974 12 : space: Lsn(0x30),
9975 12 : },
9976 12 : leases: Default::default(),
9977 12 : within_ancestor_pitr: false,
9978 12 : };
9979 12 : }
9980 12 :
9981 12 : let expected_result = [
9982 12 : Bytes::from_static(b"value 0@0x10"),
9983 12 : Bytes::from_static(b"value 1@0x10@0x20"),
9984 12 : Bytes::from_static(b"value 2@0x10@0x30"),
9985 12 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9986 12 : Bytes::from_static(b"value 4@0x10"),
9987 12 : Bytes::from_static(b"value 5@0x10@0x20"),
9988 12 : Bytes::from_static(b"value 6@0x10@0x20"),
9989 12 : Bytes::from_static(b"value 7@0x10"),
9990 12 : Bytes::from_static(b"value 8@0x10@0x48"),
9991 12 : Bytes::from_static(b"value 9@0x10@0x48"),
9992 12 : ];
9993 12 :
9994 12 : let expected_result_at_gc_horizon = [
9995 12 : Bytes::from_static(b"value 0@0x10"),
9996 12 : Bytes::from_static(b"value 1@0x10@0x20"),
9997 12 : Bytes::from_static(b"value 2@0x10@0x30"),
9998 12 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
9999 12 : Bytes::from_static(b"value 4@0x10"),
10000 12 : Bytes::from_static(b"value 5@0x10@0x20"),
10001 12 : Bytes::from_static(b"value 6@0x10@0x20"),
10002 12 : Bytes::from_static(b"value 7@0x10"),
10003 12 : Bytes::from_static(b"value 8@0x10"),
10004 12 : Bytes::from_static(b"value 9@0x10"),
10005 12 : ];
10006 12 :
10007 12 : let expected_result_at_lsn_20 = [
10008 12 : Bytes::from_static(b"value 0@0x10"),
10009 12 : Bytes::from_static(b"value 1@0x10@0x20"),
10010 12 : Bytes::from_static(b"value 2@0x10"),
10011 12 : Bytes::from_static(b"value 3@0x10"),
10012 12 : Bytes::from_static(b"value 4@0x10"),
10013 12 : Bytes::from_static(b"value 5@0x10@0x20"),
10014 12 : Bytes::from_static(b"value 6@0x10@0x20"),
10015 12 : Bytes::from_static(b"value 7@0x10"),
10016 12 : Bytes::from_static(b"value 8@0x10"),
10017 12 : Bytes::from_static(b"value 9@0x10"),
10018 12 : ];
10019 12 :
10020 12 : let expected_result_at_lsn_10 = [
10021 12 : Bytes::from_static(b"value 0@0x10"),
10022 12 : Bytes::from_static(b"value 1@0x10"),
10023 12 : Bytes::from_static(b"value 2@0x10"),
10024 12 : Bytes::from_static(b"value 3@0x10"),
10025 12 : Bytes::from_static(b"value 4@0x10"),
10026 12 : Bytes::from_static(b"value 5@0x10"),
10027 12 : Bytes::from_static(b"value 6@0x10"),
10028 12 : Bytes::from_static(b"value 7@0x10"),
10029 12 : Bytes::from_static(b"value 8@0x10"),
10030 12 : Bytes::from_static(b"value 9@0x10"),
10031 12 : ];
10032 12 :
10033 72 : let verify_result = || async {
10034 72 : let gc_horizon = {
10035 72 : let gc_info = tline.gc_info.read().unwrap();
10036 72 : gc_info.cutoffs.time
10037 12 : };
10038 792 : for idx in 0..10 {
10039 720 : assert_eq!(
10040 720 : tline
10041 720 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10042 720 : .await
10043 720 : .unwrap(),
10044 720 : &expected_result[idx]
10045 12 : );
10046 720 : assert_eq!(
10047 720 : tline
10048 720 : .get(get_key(idx as u32), gc_horizon, &ctx)
10049 720 : .await
10050 720 : .unwrap(),
10051 720 : &expected_result_at_gc_horizon[idx]
10052 12 : );
10053 720 : assert_eq!(
10054 720 : tline
10055 720 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
10056 720 : .await
10057 720 : .unwrap(),
10058 720 : &expected_result_at_lsn_20[idx]
10059 12 : );
10060 720 : assert_eq!(
10061 720 : tline
10062 720 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
10063 720 : .await
10064 720 : .unwrap(),
10065 720 : &expected_result_at_lsn_10[idx]
10066 12 : );
10067 12 : }
10068 144 : };
10069 12 :
10070 12 : verify_result().await;
10071 12 :
10072 12 : let cancel = CancellationToken::new();
10073 12 : let mut dryrun_flags = EnumSet::new();
10074 12 : dryrun_flags.insert(CompactFlags::DryRun);
10075 12 :
10076 12 : tline
10077 12 : .compact_with_gc(
10078 12 : &cancel,
10079 12 : CompactOptions {
10080 12 : flags: dryrun_flags,
10081 12 : ..Default::default()
10082 12 : },
10083 12 : &ctx,
10084 12 : )
10085 12 : .await
10086 12 : .unwrap();
10087 12 : // 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
10088 12 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
10089 12 : verify_result().await;
10090 12 :
10091 12 : tline
10092 12 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10093 12 : .await
10094 12 : .unwrap();
10095 12 : verify_result().await;
10096 12 :
10097 12 : // compact again
10098 12 : tline
10099 12 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10100 12 : .await
10101 12 : .unwrap();
10102 12 : verify_result().await;
10103 12 :
10104 12 : // increase GC horizon and compact again
10105 12 : {
10106 12 : tline
10107 12 : .applied_gc_cutoff_lsn
10108 12 : .lock_for_write()
10109 12 : .store_and_unlock(Lsn(0x38))
10110 12 : .wait()
10111 12 : .await;
10112 12 : // Update GC info
10113 12 : let mut guard = tline.gc_info.write().unwrap();
10114 12 : guard.cutoffs.time = Lsn(0x38);
10115 12 : guard.cutoffs.space = Lsn(0x38);
10116 12 : }
10117 12 : tline
10118 12 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10119 12 : .await
10120 12 : .unwrap();
10121 12 : verify_result().await; // no wals between 0x30 and 0x38, so we should obtain the same result
10122 12 :
10123 12 : // not increasing the GC horizon and compact again
10124 12 : tline
10125 12 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10126 12 : .await
10127 12 : .unwrap();
10128 12 : verify_result().await;
10129 12 :
10130 12 : Ok(())
10131 12 : }
10132 :
10133 : #[cfg(feature = "testing")]
10134 : #[tokio::test]
10135 12 : async fn test_simple_bottom_most_compaction_with_retain_lsns_single_key() -> anyhow::Result<()>
10136 12 : {
10137 12 : let harness =
10138 12 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns_single_key")
10139 12 : .await?;
10140 12 : let (tenant, ctx) = harness.load().await;
10141 12 :
10142 2112 : fn get_key(id: u32) -> Key {
10143 2112 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
10144 2112 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
10145 2112 : key.field6 = id;
10146 2112 : key
10147 2112 : }
10148 12 :
10149 12 : let img_layer = (0..10)
10150 120 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10151 12 : .collect_vec();
10152 12 :
10153 12 : let delta1 = vec![
10154 12 : (
10155 12 : get_key(1),
10156 12 : Lsn(0x20),
10157 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10158 12 : ),
10159 12 : (
10160 12 : get_key(1),
10161 12 : Lsn(0x28),
10162 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
10163 12 : ),
10164 12 : ];
10165 12 : let delta2 = vec![
10166 12 : (
10167 12 : get_key(1),
10168 12 : Lsn(0x30),
10169 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10170 12 : ),
10171 12 : (
10172 12 : get_key(1),
10173 12 : Lsn(0x38),
10174 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
10175 12 : ),
10176 12 : ];
10177 12 : let delta3 = vec![
10178 12 : (
10179 12 : get_key(8),
10180 12 : Lsn(0x48),
10181 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10182 12 : ),
10183 12 : (
10184 12 : get_key(9),
10185 12 : Lsn(0x48),
10186 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10187 12 : ),
10188 12 : ];
10189 12 :
10190 12 : let tline = tenant
10191 12 : .create_test_timeline_with_layers(
10192 12 : TIMELINE_ID,
10193 12 : Lsn(0x10),
10194 12 : DEFAULT_PG_VERSION,
10195 12 : &ctx,
10196 12 : Vec::new(), // in-memory layers
10197 12 : vec![
10198 12 : // delta1 and delta 2 only contain a single key but multiple updates
10199 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x30), delta1),
10200 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
10201 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x50), delta3),
10202 12 : ], // delta layers
10203 12 : vec![(Lsn(0x10), img_layer)], // image layers
10204 12 : Lsn(0x50),
10205 12 : )
10206 12 : .await?;
10207 12 : {
10208 12 : tline
10209 12 : .applied_gc_cutoff_lsn
10210 12 : .lock_for_write()
10211 12 : .store_and_unlock(Lsn(0x30))
10212 12 : .wait()
10213 12 : .await;
10214 12 : // Update GC info
10215 12 : let mut guard = tline.gc_info.write().unwrap();
10216 12 : *guard = GcInfo {
10217 12 : retain_lsns: vec![
10218 12 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
10219 12 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
10220 12 : ],
10221 12 : cutoffs: GcCutoffs {
10222 12 : time: Lsn(0x30),
10223 12 : space: Lsn(0x30),
10224 12 : },
10225 12 : leases: Default::default(),
10226 12 : within_ancestor_pitr: false,
10227 12 : };
10228 12 : }
10229 12 :
10230 12 : let expected_result = [
10231 12 : Bytes::from_static(b"value 0@0x10"),
10232 12 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
10233 12 : Bytes::from_static(b"value 2@0x10"),
10234 12 : Bytes::from_static(b"value 3@0x10"),
10235 12 : Bytes::from_static(b"value 4@0x10"),
10236 12 : Bytes::from_static(b"value 5@0x10"),
10237 12 : Bytes::from_static(b"value 6@0x10"),
10238 12 : Bytes::from_static(b"value 7@0x10"),
10239 12 : Bytes::from_static(b"value 8@0x10@0x48"),
10240 12 : Bytes::from_static(b"value 9@0x10@0x48"),
10241 12 : ];
10242 12 :
10243 12 : let expected_result_at_gc_horizon = [
10244 12 : Bytes::from_static(b"value 0@0x10"),
10245 12 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
10246 12 : Bytes::from_static(b"value 2@0x10"),
10247 12 : Bytes::from_static(b"value 3@0x10"),
10248 12 : Bytes::from_static(b"value 4@0x10"),
10249 12 : Bytes::from_static(b"value 5@0x10"),
10250 12 : Bytes::from_static(b"value 6@0x10"),
10251 12 : Bytes::from_static(b"value 7@0x10"),
10252 12 : Bytes::from_static(b"value 8@0x10"),
10253 12 : Bytes::from_static(b"value 9@0x10"),
10254 12 : ];
10255 12 :
10256 12 : let expected_result_at_lsn_20 = [
10257 12 : Bytes::from_static(b"value 0@0x10"),
10258 12 : Bytes::from_static(b"value 1@0x10@0x20"),
10259 12 : Bytes::from_static(b"value 2@0x10"),
10260 12 : Bytes::from_static(b"value 3@0x10"),
10261 12 : Bytes::from_static(b"value 4@0x10"),
10262 12 : Bytes::from_static(b"value 5@0x10"),
10263 12 : Bytes::from_static(b"value 6@0x10"),
10264 12 : Bytes::from_static(b"value 7@0x10"),
10265 12 : Bytes::from_static(b"value 8@0x10"),
10266 12 : Bytes::from_static(b"value 9@0x10"),
10267 12 : ];
10268 12 :
10269 12 : let expected_result_at_lsn_10 = [
10270 12 : Bytes::from_static(b"value 0@0x10"),
10271 12 : Bytes::from_static(b"value 1@0x10"),
10272 12 : Bytes::from_static(b"value 2@0x10"),
10273 12 : Bytes::from_static(b"value 3@0x10"),
10274 12 : Bytes::from_static(b"value 4@0x10"),
10275 12 : Bytes::from_static(b"value 5@0x10"),
10276 12 : Bytes::from_static(b"value 6@0x10"),
10277 12 : Bytes::from_static(b"value 7@0x10"),
10278 12 : Bytes::from_static(b"value 8@0x10"),
10279 12 : Bytes::from_static(b"value 9@0x10"),
10280 12 : ];
10281 12 :
10282 48 : let verify_result = || async {
10283 48 : let gc_horizon = {
10284 48 : let gc_info = tline.gc_info.read().unwrap();
10285 48 : gc_info.cutoffs.time
10286 12 : };
10287 528 : for idx in 0..10 {
10288 480 : assert_eq!(
10289 480 : tline
10290 480 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10291 480 : .await
10292 480 : .unwrap(),
10293 480 : &expected_result[idx]
10294 12 : );
10295 480 : assert_eq!(
10296 480 : tline
10297 480 : .get(get_key(idx as u32), gc_horizon, &ctx)
10298 480 : .await
10299 480 : .unwrap(),
10300 480 : &expected_result_at_gc_horizon[idx]
10301 12 : );
10302 480 : assert_eq!(
10303 480 : tline
10304 480 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
10305 480 : .await
10306 480 : .unwrap(),
10307 480 : &expected_result_at_lsn_20[idx]
10308 12 : );
10309 480 : assert_eq!(
10310 480 : tline
10311 480 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
10312 480 : .await
10313 480 : .unwrap(),
10314 480 : &expected_result_at_lsn_10[idx]
10315 12 : );
10316 12 : }
10317 96 : };
10318 12 :
10319 12 : verify_result().await;
10320 12 :
10321 12 : let cancel = CancellationToken::new();
10322 12 : let mut dryrun_flags = EnumSet::new();
10323 12 : dryrun_flags.insert(CompactFlags::DryRun);
10324 12 :
10325 12 : tline
10326 12 : .compact_with_gc(
10327 12 : &cancel,
10328 12 : CompactOptions {
10329 12 : flags: dryrun_flags,
10330 12 : ..Default::default()
10331 12 : },
10332 12 : &ctx,
10333 12 : )
10334 12 : .await
10335 12 : .unwrap();
10336 12 : // 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
10337 12 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
10338 12 : verify_result().await;
10339 12 :
10340 12 : tline
10341 12 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10342 12 : .await
10343 12 : .unwrap();
10344 12 : verify_result().await;
10345 12 :
10346 12 : // compact again
10347 12 : tline
10348 12 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10349 12 : .await
10350 12 : .unwrap();
10351 12 : verify_result().await;
10352 12 :
10353 12 : Ok(())
10354 12 : }
10355 :
10356 : #[cfg(feature = "testing")]
10357 : #[tokio::test]
10358 12 : async fn test_simple_bottom_most_compaction_on_branch() -> anyhow::Result<()> {
10359 12 : use models::CompactLsnRange;
10360 12 :
10361 12 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_on_branch").await?;
10362 12 : let (tenant, ctx) = harness.load().await;
10363 12 :
10364 996 : fn get_key(id: u32) -> Key {
10365 996 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
10366 996 : key.field6 = id;
10367 996 : key
10368 996 : }
10369 12 :
10370 12 : let img_layer = (0..10)
10371 120 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10372 12 : .collect_vec();
10373 12 :
10374 12 : let delta1 = vec![
10375 12 : (
10376 12 : get_key(1),
10377 12 : Lsn(0x20),
10378 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10379 12 : ),
10380 12 : (
10381 12 : get_key(2),
10382 12 : Lsn(0x30),
10383 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10384 12 : ),
10385 12 : (
10386 12 : get_key(3),
10387 12 : Lsn(0x28),
10388 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
10389 12 : ),
10390 12 : (
10391 12 : get_key(3),
10392 12 : Lsn(0x30),
10393 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10394 12 : ),
10395 12 : (
10396 12 : get_key(3),
10397 12 : Lsn(0x40),
10398 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
10399 12 : ),
10400 12 : ];
10401 12 : let delta2 = vec![
10402 12 : (
10403 12 : get_key(5),
10404 12 : Lsn(0x20),
10405 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10406 12 : ),
10407 12 : (
10408 12 : get_key(6),
10409 12 : Lsn(0x20),
10410 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10411 12 : ),
10412 12 : ];
10413 12 : let delta3 = vec![
10414 12 : (
10415 12 : get_key(8),
10416 12 : Lsn(0x48),
10417 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10418 12 : ),
10419 12 : (
10420 12 : get_key(9),
10421 12 : Lsn(0x48),
10422 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10423 12 : ),
10424 12 : ];
10425 12 :
10426 12 : let parent_tline = tenant
10427 12 : .create_test_timeline_with_layers(
10428 12 : TIMELINE_ID,
10429 12 : Lsn(0x10),
10430 12 : DEFAULT_PG_VERSION,
10431 12 : &ctx,
10432 12 : vec![], // in-memory layers
10433 12 : vec![], // delta layers
10434 12 : vec![(Lsn(0x18), img_layer)], // image layers
10435 12 : Lsn(0x18),
10436 12 : )
10437 12 : .await?;
10438 12 :
10439 12 : parent_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
10440 12 :
10441 12 : let branch_tline = tenant
10442 12 : .branch_timeline_test_with_layers(
10443 12 : &parent_tline,
10444 12 : NEW_TIMELINE_ID,
10445 12 : Some(Lsn(0x18)),
10446 12 : &ctx,
10447 12 : vec![
10448 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
10449 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
10450 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
10451 12 : ], // delta layers
10452 12 : vec![], // image layers
10453 12 : Lsn(0x50),
10454 12 : )
10455 12 : .await?;
10456 12 :
10457 12 : branch_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
10458 12 :
10459 12 : {
10460 12 : parent_tline
10461 12 : .applied_gc_cutoff_lsn
10462 12 : .lock_for_write()
10463 12 : .store_and_unlock(Lsn(0x10))
10464 12 : .wait()
10465 12 : .await;
10466 12 : // Update GC info
10467 12 : let mut guard = parent_tline.gc_info.write().unwrap();
10468 12 : *guard = GcInfo {
10469 12 : retain_lsns: vec![(Lsn(0x18), branch_tline.timeline_id, MaybeOffloaded::No)],
10470 12 : cutoffs: GcCutoffs {
10471 12 : time: Lsn(0x10),
10472 12 : space: Lsn(0x10),
10473 12 : },
10474 12 : leases: Default::default(),
10475 12 : within_ancestor_pitr: false,
10476 12 : };
10477 12 : }
10478 12 :
10479 12 : {
10480 12 : branch_tline
10481 12 : .applied_gc_cutoff_lsn
10482 12 : .lock_for_write()
10483 12 : .store_and_unlock(Lsn(0x50))
10484 12 : .wait()
10485 12 : .await;
10486 12 : // Update GC info
10487 12 : let mut guard = branch_tline.gc_info.write().unwrap();
10488 12 : *guard = GcInfo {
10489 12 : retain_lsns: vec![(Lsn(0x40), branch_tline.timeline_id, MaybeOffloaded::No)],
10490 12 : cutoffs: GcCutoffs {
10491 12 : time: Lsn(0x50),
10492 12 : space: Lsn(0x50),
10493 12 : },
10494 12 : leases: Default::default(),
10495 12 : within_ancestor_pitr: false,
10496 12 : };
10497 12 : }
10498 12 :
10499 12 : let expected_result_at_gc_horizon = [
10500 12 : Bytes::from_static(b"value 0@0x10"),
10501 12 : Bytes::from_static(b"value 1@0x10@0x20"),
10502 12 : Bytes::from_static(b"value 2@0x10@0x30"),
10503 12 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
10504 12 : Bytes::from_static(b"value 4@0x10"),
10505 12 : Bytes::from_static(b"value 5@0x10@0x20"),
10506 12 : Bytes::from_static(b"value 6@0x10@0x20"),
10507 12 : Bytes::from_static(b"value 7@0x10"),
10508 12 : Bytes::from_static(b"value 8@0x10@0x48"),
10509 12 : Bytes::from_static(b"value 9@0x10@0x48"),
10510 12 : ];
10511 12 :
10512 12 : let expected_result_at_lsn_40 = [
10513 12 : Bytes::from_static(b"value 0@0x10"),
10514 12 : Bytes::from_static(b"value 1@0x10@0x20"),
10515 12 : Bytes::from_static(b"value 2@0x10@0x30"),
10516 12 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
10517 12 : Bytes::from_static(b"value 4@0x10"),
10518 12 : Bytes::from_static(b"value 5@0x10@0x20"),
10519 12 : Bytes::from_static(b"value 6@0x10@0x20"),
10520 12 : Bytes::from_static(b"value 7@0x10"),
10521 12 : Bytes::from_static(b"value 8@0x10"),
10522 12 : Bytes::from_static(b"value 9@0x10"),
10523 12 : ];
10524 12 :
10525 36 : let verify_result = || async {
10526 396 : for idx in 0..10 {
10527 360 : assert_eq!(
10528 360 : branch_tline
10529 360 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10530 360 : .await
10531 360 : .unwrap(),
10532 360 : &expected_result_at_gc_horizon[idx]
10533 12 : );
10534 360 : assert_eq!(
10535 360 : branch_tline
10536 360 : .get(get_key(idx as u32), Lsn(0x40), &ctx)
10537 360 : .await
10538 360 : .unwrap(),
10539 360 : &expected_result_at_lsn_40[idx]
10540 12 : );
10541 12 : }
10542 72 : };
10543 12 :
10544 12 : verify_result().await;
10545 12 :
10546 12 : let cancel = CancellationToken::new();
10547 12 : branch_tline
10548 12 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10549 12 : .await
10550 12 : .unwrap();
10551 12 :
10552 12 : verify_result().await;
10553 12 :
10554 12 : // Piggyback a compaction with above_lsn. Ensure it works correctly when the specified LSN intersects with the layer files.
10555 12 : // Now we already have a single large delta layer, so the compaction min_layer_lsn should be the same as ancestor LSN (0x18).
10556 12 : branch_tline
10557 12 : .compact_with_gc(
10558 12 : &cancel,
10559 12 : CompactOptions {
10560 12 : compact_lsn_range: Some(CompactLsnRange::above(Lsn(0x40))),
10561 12 : ..Default::default()
10562 12 : },
10563 12 : &ctx,
10564 12 : )
10565 12 : .await
10566 12 : .unwrap();
10567 12 :
10568 12 : verify_result().await;
10569 12 :
10570 12 : Ok(())
10571 12 : }
10572 :
10573 : // Regression test for https://github.com/neondatabase/neon/issues/9012
10574 : // Create an image arrangement where we have to read at different LSN ranges
10575 : // from a delta layer. This is achieved by overlapping an image layer on top of
10576 : // a delta layer. Like so:
10577 : //
10578 : // A B
10579 : // +----------------+ -> delta_layer
10580 : // | | ^ lsn
10581 : // | =========|-> nested_image_layer |
10582 : // | C | |
10583 : // +----------------+ |
10584 : // ======== -> baseline_image_layer +-------> key
10585 : //
10586 : //
10587 : // When querying the key range [A, B) we need to read at different LSN ranges
10588 : // for [A, C) and [C, B). This test checks that the described edge case is handled correctly.
10589 : #[cfg(feature = "testing")]
10590 : #[tokio::test]
10591 12 : async fn test_vectored_read_with_nested_image_layer() -> anyhow::Result<()> {
10592 12 : let harness = TenantHarness::create("test_vectored_read_with_nested_image_layer").await?;
10593 12 : let (tenant, ctx) = harness.load().await;
10594 12 :
10595 12 : let will_init_keys = [2, 6];
10596 264 : fn get_key(id: u32) -> Key {
10597 264 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
10598 264 : key.field6 = id;
10599 264 : key
10600 264 : }
10601 12 :
10602 12 : let mut expected_key_values = HashMap::new();
10603 12 :
10604 12 : let baseline_image_layer_lsn = Lsn(0x10);
10605 12 : let mut baseline_img_layer = Vec::new();
10606 72 : for i in 0..5 {
10607 60 : let key = get_key(i);
10608 60 : let value = format!("value {i}@{baseline_image_layer_lsn}");
10609 60 :
10610 60 : let removed = expected_key_values.insert(key, value.clone());
10611 60 : assert!(removed.is_none());
10612 12 :
10613 60 : baseline_img_layer.push((key, Bytes::from(value)));
10614 12 : }
10615 12 :
10616 12 : let nested_image_layer_lsn = Lsn(0x50);
10617 12 : let mut nested_img_layer = Vec::new();
10618 72 : for i in 5..10 {
10619 60 : let key = get_key(i);
10620 60 : let value = format!("value {i}@{nested_image_layer_lsn}");
10621 60 :
10622 60 : let removed = expected_key_values.insert(key, value.clone());
10623 60 : assert!(removed.is_none());
10624 12 :
10625 60 : nested_img_layer.push((key, Bytes::from(value)));
10626 12 : }
10627 12 :
10628 12 : let mut delta_layer_spec = Vec::default();
10629 12 : let delta_layer_start_lsn = Lsn(0x20);
10630 12 : let mut delta_layer_end_lsn = delta_layer_start_lsn;
10631 12 :
10632 132 : for i in 0..10 {
10633 120 : let key = get_key(i);
10634 120 : let key_in_nested = nested_img_layer
10635 120 : .iter()
10636 480 : .any(|(key_with_img, _)| *key_with_img == key);
10637 120 : let lsn = {
10638 120 : if key_in_nested {
10639 60 : Lsn(nested_image_layer_lsn.0 + 0x10)
10640 12 : } else {
10641 60 : delta_layer_start_lsn
10642 12 : }
10643 12 : };
10644 12 :
10645 120 : let will_init = will_init_keys.contains(&i);
10646 120 : if will_init {
10647 24 : delta_layer_spec.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init(""))));
10648 24 :
10649 24 : expected_key_values.insert(key, "".to_string());
10650 96 : } else {
10651 96 : let delta = format!("@{lsn}");
10652 96 : delta_layer_spec.push((
10653 96 : key,
10654 96 : lsn,
10655 96 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
10656 96 : ));
10657 96 :
10658 96 : expected_key_values
10659 96 : .get_mut(&key)
10660 96 : .expect("An image exists for each key")
10661 96 : .push_str(delta.as_str());
10662 96 : }
10663 120 : delta_layer_end_lsn = std::cmp::max(delta_layer_start_lsn, lsn);
10664 12 : }
10665 12 :
10666 12 : delta_layer_end_lsn = Lsn(delta_layer_end_lsn.0 + 1);
10667 12 :
10668 12 : assert!(
10669 12 : nested_image_layer_lsn > delta_layer_start_lsn
10670 12 : && nested_image_layer_lsn < delta_layer_end_lsn
10671 12 : );
10672 12 :
10673 12 : let tline = tenant
10674 12 : .create_test_timeline_with_layers(
10675 12 : TIMELINE_ID,
10676 12 : baseline_image_layer_lsn,
10677 12 : DEFAULT_PG_VERSION,
10678 12 : &ctx,
10679 12 : vec![], // in-memory layers
10680 12 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
10681 12 : delta_layer_start_lsn..delta_layer_end_lsn,
10682 12 : delta_layer_spec,
10683 12 : )], // delta layers
10684 12 : vec![
10685 12 : (baseline_image_layer_lsn, baseline_img_layer),
10686 12 : (nested_image_layer_lsn, nested_img_layer),
10687 12 : ], // image layers
10688 12 : delta_layer_end_lsn,
10689 12 : )
10690 12 : .await?;
10691 12 :
10692 12 : let query = VersionedKeySpaceQuery::uniform(
10693 12 : KeySpace::single(get_key(0)..get_key(10)),
10694 12 : delta_layer_end_lsn,
10695 12 : );
10696 12 :
10697 12 : let results = tline
10698 12 : .get_vectored(query, IoConcurrency::sequential(), &ctx)
10699 12 : .await
10700 12 : .expect("No vectored errors");
10701 132 : for (key, res) in results {
10702 120 : let value = res.expect("No key errors");
10703 120 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
10704 120 : assert_eq!(value, Bytes::from(expected_value));
10705 12 : }
10706 12 :
10707 12 : Ok(())
10708 12 : }
10709 :
10710 : #[cfg(feature = "testing")]
10711 : #[tokio::test]
10712 12 : async fn test_vectored_read_with_image_layer_inside_inmem() -> anyhow::Result<()> {
10713 12 : let harness =
10714 12 : TenantHarness::create("test_vectored_read_with_image_layer_inside_inmem").await?;
10715 12 : let (tenant, ctx) = harness.load().await;
10716 12 :
10717 12 : let will_init_keys = [2, 6];
10718 384 : fn get_key(id: u32) -> Key {
10719 384 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
10720 384 : key.field6 = id;
10721 384 : key
10722 384 : }
10723 12 :
10724 12 : let mut expected_key_values = HashMap::new();
10725 12 :
10726 12 : let baseline_image_layer_lsn = Lsn(0x10);
10727 12 : let mut baseline_img_layer = Vec::new();
10728 72 : for i in 0..5 {
10729 60 : let key = get_key(i);
10730 60 : let value = format!("value {i}@{baseline_image_layer_lsn}");
10731 60 :
10732 60 : let removed = expected_key_values.insert(key, value.clone());
10733 60 : assert!(removed.is_none());
10734 12 :
10735 60 : baseline_img_layer.push((key, Bytes::from(value)));
10736 12 : }
10737 12 :
10738 12 : let nested_image_layer_lsn = Lsn(0x50);
10739 12 : let mut nested_img_layer = Vec::new();
10740 72 : for i in 5..10 {
10741 60 : let key = get_key(i);
10742 60 : let value = format!("value {i}@{nested_image_layer_lsn}");
10743 60 :
10744 60 : let removed = expected_key_values.insert(key, value.clone());
10745 60 : assert!(removed.is_none());
10746 12 :
10747 60 : nested_img_layer.push((key, Bytes::from(value)));
10748 12 : }
10749 12 :
10750 12 : let frozen_layer = {
10751 12 : let lsn_range = Lsn(0x40)..Lsn(0x60);
10752 12 : let mut data = Vec::new();
10753 132 : for i in 0..10 {
10754 120 : let key = get_key(i);
10755 120 : let key_in_nested = nested_img_layer
10756 120 : .iter()
10757 480 : .any(|(key_with_img, _)| *key_with_img == key);
10758 120 : let lsn = {
10759 120 : if key_in_nested {
10760 60 : Lsn(nested_image_layer_lsn.0 + 5)
10761 12 : } else {
10762 60 : lsn_range.start
10763 12 : }
10764 12 : };
10765 12 :
10766 120 : let will_init = will_init_keys.contains(&i);
10767 120 : if will_init {
10768 24 : data.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init(""))));
10769 24 :
10770 24 : expected_key_values.insert(key, "".to_string());
10771 96 : } else {
10772 96 : let delta = format!("@{lsn}");
10773 96 : data.push((
10774 96 : key,
10775 96 : lsn,
10776 96 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
10777 96 : ));
10778 96 :
10779 96 : expected_key_values
10780 96 : .get_mut(&key)
10781 96 : .expect("An image exists for each key")
10782 96 : .push_str(delta.as_str());
10783 96 : }
10784 12 : }
10785 12 :
10786 12 : InMemoryLayerTestDesc {
10787 12 : lsn_range,
10788 12 : is_open: false,
10789 12 : data,
10790 12 : }
10791 12 : };
10792 12 :
10793 12 : let (open_layer, last_record_lsn) = {
10794 12 : let start_lsn = Lsn(0x70);
10795 12 : let mut data = Vec::new();
10796 12 : let mut end_lsn = Lsn(0);
10797 132 : for i in 0..10 {
10798 120 : let key = get_key(i);
10799 120 : let lsn = Lsn(start_lsn.0 + i as u64);
10800 120 : let delta = format!("@{lsn}");
10801 120 : data.push((
10802 120 : key,
10803 120 : lsn,
10804 120 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
10805 120 : ));
10806 120 :
10807 120 : expected_key_values
10808 120 : .get_mut(&key)
10809 120 : .expect("An image exists for each key")
10810 120 : .push_str(delta.as_str());
10811 120 :
10812 120 : end_lsn = std::cmp::max(end_lsn, lsn);
10813 120 : }
10814 12 :
10815 12 : (
10816 12 : InMemoryLayerTestDesc {
10817 12 : lsn_range: start_lsn..Lsn::MAX,
10818 12 : is_open: true,
10819 12 : data,
10820 12 : },
10821 12 : end_lsn,
10822 12 : )
10823 12 : };
10824 12 :
10825 12 : assert!(
10826 12 : nested_image_layer_lsn > frozen_layer.lsn_range.start
10827 12 : && nested_image_layer_lsn < frozen_layer.lsn_range.end
10828 12 : );
10829 12 :
10830 12 : let tline = tenant
10831 12 : .create_test_timeline_with_layers(
10832 12 : TIMELINE_ID,
10833 12 : baseline_image_layer_lsn,
10834 12 : DEFAULT_PG_VERSION,
10835 12 : &ctx,
10836 12 : vec![open_layer, frozen_layer], // in-memory layers
10837 12 : Vec::new(), // delta layers
10838 12 : vec![
10839 12 : (baseline_image_layer_lsn, baseline_img_layer),
10840 12 : (nested_image_layer_lsn, nested_img_layer),
10841 12 : ], // image layers
10842 12 : last_record_lsn,
10843 12 : )
10844 12 : .await?;
10845 12 :
10846 12 : let query = VersionedKeySpaceQuery::uniform(
10847 12 : KeySpace::single(get_key(0)..get_key(10)),
10848 12 : last_record_lsn,
10849 12 : );
10850 12 :
10851 12 : let results = tline
10852 12 : .get_vectored(query, IoConcurrency::sequential(), &ctx)
10853 12 : .await
10854 12 : .expect("No vectored errors");
10855 132 : for (key, res) in results {
10856 120 : let value = res.expect("No key errors");
10857 120 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
10858 120 : assert_eq!(value, Bytes::from(expected_value.clone()));
10859 12 :
10860 120 : tracing::info!("key={key} value={expected_value}");
10861 12 : }
10862 12 :
10863 12 : Ok(())
10864 12 : }
10865 :
10866 : // A randomized read path test. Generates a layer map according to a deterministic
10867 : // specification. Fills the (key, LSN) space in random manner and then performs
10868 : // random scattered queries validating the results against in-memory storage.
10869 : //
10870 : // See this internal Notion page for a diagram of the layer map:
10871 : // https://www.notion.so/neondatabase/Read-Path-Unit-Testing-Fuzzing-1d1f189e0047806c8e5cd37781b0a350?pvs=4
10872 : //
10873 : // A fuzzing mode is also supported. In this mode, the test will use a random
10874 : // seed instead of a hardcoded one. Use it in conjunction with `cargo stress`
10875 : // to run multiple instances in parallel:
10876 : //
10877 : // $ RUST_BACKTRACE=1 RUST_LOG=INFO \
10878 : // cargo stress --package=pageserver --features=testing,fuzz-read-path --release -- test_read_path
10879 : #[cfg(feature = "testing")]
10880 : #[tokio::test]
10881 12 : async fn test_read_path() -> anyhow::Result<()> {
10882 12 : use rand::seq::SliceRandom;
10883 12 :
10884 12 : let seed = if cfg!(feature = "fuzz-read-path") {
10885 12 : let seed: u64 = thread_rng().r#gen();
10886 0 : seed
10887 12 : } else {
10888 12 : // Use a hard-coded seed when not in fuzzing mode.
10889 12 : // Note that with the current approach results are not reproducible
10890 12 : // accross platforms and Rust releases.
10891 12 : const SEED: u64 = 0;
10892 12 : SEED
10893 12 : };
10894 12 :
10895 12 : let mut random = StdRng::seed_from_u64(seed);
10896 12 :
10897 12 : let (queries, will_init_chance, gap_chance) = if cfg!(feature = "fuzz-read-path") {
10898 12 : const QUERIES: u64 = 5000;
10899 12 : let will_init_chance: u8 = random.gen_range(0..=10);
10900 0 : let gap_chance: u8 = random.gen_range(0..=50);
10901 0 :
10902 0 : (QUERIES, will_init_chance, gap_chance)
10903 12 : } else {
10904 12 : const QUERIES: u64 = 1000;
10905 12 : const WILL_INIT_CHANCE: u8 = 1;
10906 12 : const GAP_CHANCE: u8 = 5;
10907 12 :
10908 12 : (QUERIES, WILL_INIT_CHANCE, GAP_CHANCE)
10909 12 : };
10910 12 :
10911 12 : let harness = TenantHarness::create("test_read_path").await?;
10912 12 : let (tenant, ctx) = harness.load().await;
10913 12 :
10914 12 : tracing::info!("Using random seed: {seed}");
10915 12 : tracing::info!(%will_init_chance, %gap_chance, "Fill params");
10916 12 :
10917 12 : // Define the layer map shape. Note that this part is not randomized.
10918 12 :
10919 12 : const KEY_DIMENSION_SIZE: u32 = 99;
10920 12 : let start_key = Key::from_hex("110000000033333333444444445500000000").unwrap();
10921 12 : let end_key = start_key.add(KEY_DIMENSION_SIZE);
10922 12 : let total_key_range = start_key..end_key;
10923 12 : let total_key_range_size = end_key.to_i128() - start_key.to_i128();
10924 12 : let total_start_lsn = Lsn(104);
10925 12 : let last_record_lsn = Lsn(504);
10926 12 :
10927 12 : assert!(total_key_range_size % 3 == 0);
10928 12 :
10929 12 : let in_memory_layers_shape = vec![
10930 12 : (total_key_range.clone(), Lsn(304)..Lsn(400)),
10931 12 : (total_key_range.clone(), Lsn(400)..last_record_lsn),
10932 12 : ];
10933 12 :
10934 12 : let delta_layers_shape = vec![
10935 12 : (
10936 12 : start_key..(start_key.add((total_key_range_size / 3) as u32)),
10937 12 : Lsn(200)..Lsn(304),
10938 12 : ),
10939 12 : (
10940 12 : (start_key.add((total_key_range_size / 3) as u32))
10941 12 : ..(start_key.add((total_key_range_size * 2 / 3) as u32)),
10942 12 : Lsn(200)..Lsn(304),
10943 12 : ),
10944 12 : (
10945 12 : (start_key.add((total_key_range_size * 2 / 3) as u32))
10946 12 : ..(start_key.add(total_key_range_size as u32)),
10947 12 : Lsn(200)..Lsn(304),
10948 12 : ),
10949 12 : ];
10950 12 :
10951 12 : let image_layers_shape = vec![
10952 12 : (
10953 12 : start_key.add((total_key_range_size * 2 / 3 - 10) as u32)
10954 12 : ..start_key.add((total_key_range_size * 2 / 3 + 10) as u32),
10955 12 : Lsn(456),
10956 12 : ),
10957 12 : (
10958 12 : start_key.add((total_key_range_size / 3 - 10) as u32)
10959 12 : ..start_key.add((total_key_range_size / 3 + 10) as u32),
10960 12 : Lsn(256),
10961 12 : ),
10962 12 : (total_key_range.clone(), total_start_lsn),
10963 12 : ];
10964 12 :
10965 12 : let specification = TestTimelineSpecification {
10966 12 : start_lsn: total_start_lsn,
10967 12 : last_record_lsn,
10968 12 : in_memory_layers_shape,
10969 12 : delta_layers_shape,
10970 12 : image_layers_shape,
10971 12 : gap_chance,
10972 12 : will_init_chance,
10973 12 : };
10974 12 :
10975 12 : // Create and randomly fill in the layers according to the specification
10976 12 : let (tline, storage, interesting_lsns) = randomize_timeline(
10977 12 : &tenant,
10978 12 : TIMELINE_ID,
10979 12 : DEFAULT_PG_VERSION,
10980 12 : specification,
10981 12 : &mut random,
10982 12 : &ctx,
10983 12 : )
10984 12 : .await?;
10985 12 :
10986 12 : // Now generate queries based on the interesting lsns that we've collected.
10987 12 : //
10988 12 : // While there's still room in the query, pick and interesting LSN and a random
10989 12 : // key. Then roll the dice to see if the next key should also be included in
10990 12 : // the query. When the roll fails, break the "batch" and pick another point in the
10991 12 : // (key, LSN) space.
10992 12 :
10993 12 : const PICK_NEXT_CHANCE: u8 = 50;
10994 12 : for _ in 0..queries {
10995 12000 : let query = {
10996 12000 : let mut keyspaces_at_lsn: HashMap<Lsn, KeySpaceRandomAccum> = HashMap::default();
10997 12000 : let mut used_keys: HashSet<Key> = HashSet::default();
10998 12 :
10999 270432 : while used_keys.len() < Timeline::MAX_GET_VECTORED_KEYS as usize {
11000 258432 : let selected_lsn = interesting_lsns.choose(&mut random).expect("not empty");
11001 258432 : let mut selected_key = start_key.add(random.gen_range(0..KEY_DIMENSION_SIZE));
11002 12 :
11003 451368 : while used_keys.len() < Timeline::MAX_GET_VECTORED_KEYS as usize {
11004 445116 : if used_keys.contains(&selected_key)
11005 385848 : || selected_key >= start_key.add(KEY_DIMENSION_SIZE)
11006 12 : {
11007 61116 : break;
11008 384000 : }
11009 384000 :
11010 384000 : keyspaces_at_lsn
11011 384000 : .entry(*selected_lsn)
11012 384000 : .or_default()
11013 384000 : .add_key(selected_key);
11014 384000 : used_keys.insert(selected_key);
11015 384000 :
11016 384000 : let pick_next = random.gen_range(0..=100) <= PICK_NEXT_CHANCE;
11017 384000 : if pick_next {
11018 192936 : selected_key = selected_key.next();
11019 192936 : } else {
11020 191064 : break;
11021 12 : }
11022 12 : }
11023 12 : }
11024 12 :
11025 12000 : VersionedKeySpaceQuery::scattered(
11026 12000 : keyspaces_at_lsn
11027 12000 : .into_iter()
11028 143004 : .map(|(lsn, acc)| (lsn, acc.to_keyspace()))
11029 12000 : .collect(),
11030 12000 : )
11031 12 : };
11032 12 :
11033 12 : // Run the query and validate the results
11034 12 :
11035 12000 : let results = tline
11036 12000 : .get_vectored(query.clone(), IoConcurrency::Sequential, &ctx)
11037 12000 : .await;
11038 12 :
11039 12000 : let blobs = match results {
11040 12000 : Ok(ok) => ok,
11041 12 : Err(err) => {
11042 0 : panic!("seed={seed} Error returned for query {query}: {err}");
11043 12 : }
11044 12 : };
11045 12 :
11046 384000 : for (key, key_res) in blobs.into_iter() {
11047 384000 : match key_res {
11048 384000 : Ok(blob) => {
11049 384000 : let requested_at_lsn = query.map_key_to_lsn(&key);
11050 384000 : let expected = storage.get(key, requested_at_lsn);
11051 384000 :
11052 384000 : if blob != expected {
11053 12 : tracing::error!(
11054 12 : "seed={seed} Mismatch for {key}@{requested_at_lsn} from query: {query}"
11055 12 : );
11056 384000 : }
11057 12 :
11058 384000 : assert_eq!(blob, expected);
11059 12 : }
11060 12 : Err(err) => {
11061 0 : let requested_at_lsn = query.map_key_to_lsn(&key);
11062 0 :
11063 0 : panic!(
11064 0 : "seed={seed} Error returned for {key}@{requested_at_lsn} from query {query}: {err}"
11065 0 : );
11066 12 : }
11067 12 : }
11068 12 : }
11069 12 : }
11070 12 :
11071 12 : Ok(())
11072 12 : }
11073 :
11074 1284 : fn sort_layer_key(k1: &PersistentLayerKey, k2: &PersistentLayerKey) -> std::cmp::Ordering {
11075 1284 : (
11076 1284 : k1.is_delta,
11077 1284 : k1.key_range.start,
11078 1284 : k1.key_range.end,
11079 1284 : k1.lsn_range.start,
11080 1284 : k1.lsn_range.end,
11081 1284 : )
11082 1284 : .cmp(&(
11083 1284 : k2.is_delta,
11084 1284 : k2.key_range.start,
11085 1284 : k2.key_range.end,
11086 1284 : k2.lsn_range.start,
11087 1284 : k2.lsn_range.end,
11088 1284 : ))
11089 1284 : }
11090 :
11091 144 : async fn inspect_and_sort(
11092 144 : tline: &Arc<Timeline>,
11093 144 : filter: Option<std::ops::Range<Key>>,
11094 144 : ) -> Vec<PersistentLayerKey> {
11095 144 : let mut all_layers = tline.inspect_historic_layers().await.unwrap();
11096 144 : if let Some(filter) = filter {
11097 648 : all_layers.retain(|layer| overlaps_with(&layer.key_range, &filter));
11098 132 : }
11099 144 : all_layers.sort_by(sort_layer_key);
11100 144 : all_layers
11101 144 : }
11102 :
11103 : #[cfg(feature = "testing")]
11104 132 : fn check_layer_map_key_eq(
11105 132 : mut left: Vec<PersistentLayerKey>,
11106 132 : mut right: Vec<PersistentLayerKey>,
11107 132 : ) {
11108 132 : left.sort_by(sort_layer_key);
11109 132 : right.sort_by(sort_layer_key);
11110 132 : if left != right {
11111 0 : eprintln!("---LEFT---");
11112 0 : for left in left.iter() {
11113 0 : eprintln!("{}", left);
11114 0 : }
11115 0 : eprintln!("---RIGHT---");
11116 0 : for right in right.iter() {
11117 0 : eprintln!("{}", right);
11118 0 : }
11119 0 : assert_eq!(left, right);
11120 132 : }
11121 132 : }
11122 :
11123 : #[cfg(feature = "testing")]
11124 : #[tokio::test]
11125 12 : async fn test_simple_partial_bottom_most_compaction() -> anyhow::Result<()> {
11126 12 : let harness = TenantHarness::create("test_simple_partial_bottom_most_compaction").await?;
11127 12 : let (tenant, ctx) = harness.load().await;
11128 12 :
11129 1092 : fn get_key(id: u32) -> Key {
11130 1092 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
11131 1092 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
11132 1092 : key.field6 = id;
11133 1092 : key
11134 1092 : }
11135 12 :
11136 12 : // img layer at 0x10
11137 12 : let img_layer = (0..10)
11138 120 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
11139 12 : .collect_vec();
11140 12 :
11141 12 : let delta1 = vec![
11142 12 : (
11143 12 : get_key(1),
11144 12 : Lsn(0x20),
11145 12 : Value::Image(Bytes::from("value 1@0x20")),
11146 12 : ),
11147 12 : (
11148 12 : get_key(2),
11149 12 : Lsn(0x30),
11150 12 : Value::Image(Bytes::from("value 2@0x30")),
11151 12 : ),
11152 12 : (
11153 12 : get_key(3),
11154 12 : Lsn(0x40),
11155 12 : Value::Image(Bytes::from("value 3@0x40")),
11156 12 : ),
11157 12 : ];
11158 12 : let delta2 = vec![
11159 12 : (
11160 12 : get_key(5),
11161 12 : Lsn(0x20),
11162 12 : Value::Image(Bytes::from("value 5@0x20")),
11163 12 : ),
11164 12 : (
11165 12 : get_key(6),
11166 12 : Lsn(0x20),
11167 12 : Value::Image(Bytes::from("value 6@0x20")),
11168 12 : ),
11169 12 : ];
11170 12 : let delta3 = vec![
11171 12 : (
11172 12 : get_key(8),
11173 12 : Lsn(0x48),
11174 12 : Value::Image(Bytes::from("value 8@0x48")),
11175 12 : ),
11176 12 : (
11177 12 : get_key(9),
11178 12 : Lsn(0x48),
11179 12 : Value::Image(Bytes::from("value 9@0x48")),
11180 12 : ),
11181 12 : ];
11182 12 :
11183 12 : let tline = tenant
11184 12 : .create_test_timeline_with_layers(
11185 12 : TIMELINE_ID,
11186 12 : Lsn(0x10),
11187 12 : DEFAULT_PG_VERSION,
11188 12 : &ctx,
11189 12 : vec![], // in-memory layers
11190 12 : vec![
11191 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
11192 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
11193 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
11194 12 : ], // delta layers
11195 12 : vec![(Lsn(0x10), img_layer)], // image layers
11196 12 : Lsn(0x50),
11197 12 : )
11198 12 : .await?;
11199 12 :
11200 12 : {
11201 12 : tline
11202 12 : .applied_gc_cutoff_lsn
11203 12 : .lock_for_write()
11204 12 : .store_and_unlock(Lsn(0x30))
11205 12 : .wait()
11206 12 : .await;
11207 12 : // Update GC info
11208 12 : let mut guard = tline.gc_info.write().unwrap();
11209 12 : *guard = GcInfo {
11210 12 : retain_lsns: vec![(Lsn(0x20), tline.timeline_id, MaybeOffloaded::No)],
11211 12 : cutoffs: GcCutoffs {
11212 12 : time: Lsn(0x30),
11213 12 : space: Lsn(0x30),
11214 12 : },
11215 12 : leases: Default::default(),
11216 12 : within_ancestor_pitr: false,
11217 12 : };
11218 12 : }
11219 12 :
11220 12 : let cancel = CancellationToken::new();
11221 12 :
11222 12 : // Do a partial compaction on key range 0..2
11223 12 : tline
11224 12 : .compact_with_gc(
11225 12 : &cancel,
11226 12 : CompactOptions {
11227 12 : flags: EnumSet::new(),
11228 12 : compact_key_range: Some((get_key(0)..get_key(2)).into()),
11229 12 : ..Default::default()
11230 12 : },
11231 12 : &ctx,
11232 12 : )
11233 12 : .await
11234 12 : .unwrap();
11235 12 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11236 12 : check_layer_map_key_eq(
11237 12 : all_layers,
11238 12 : vec![
11239 12 : // newly-generated image layer for the partial compaction range 0-2
11240 12 : PersistentLayerKey {
11241 12 : key_range: get_key(0)..get_key(2),
11242 12 : lsn_range: Lsn(0x20)..Lsn(0x21),
11243 12 : is_delta: false,
11244 12 : },
11245 12 : PersistentLayerKey {
11246 12 : key_range: get_key(0)..get_key(10),
11247 12 : lsn_range: Lsn(0x10)..Lsn(0x11),
11248 12 : is_delta: false,
11249 12 : },
11250 12 : // delta1 is split and the second part is rewritten
11251 12 : PersistentLayerKey {
11252 12 : key_range: get_key(2)..get_key(4),
11253 12 : lsn_range: Lsn(0x20)..Lsn(0x48),
11254 12 : is_delta: true,
11255 12 : },
11256 12 : PersistentLayerKey {
11257 12 : key_range: get_key(5)..get_key(7),
11258 12 : lsn_range: Lsn(0x20)..Lsn(0x48),
11259 12 : is_delta: true,
11260 12 : },
11261 12 : PersistentLayerKey {
11262 12 : key_range: get_key(8)..get_key(10),
11263 12 : lsn_range: Lsn(0x48)..Lsn(0x50),
11264 12 : is_delta: true,
11265 12 : },
11266 12 : ],
11267 12 : );
11268 12 :
11269 12 : // Do a partial compaction on key range 2..4
11270 12 : tline
11271 12 : .compact_with_gc(
11272 12 : &cancel,
11273 12 : CompactOptions {
11274 12 : flags: EnumSet::new(),
11275 12 : compact_key_range: Some((get_key(2)..get_key(4)).into()),
11276 12 : ..Default::default()
11277 12 : },
11278 12 : &ctx,
11279 12 : )
11280 12 : .await
11281 12 : .unwrap();
11282 12 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11283 12 : check_layer_map_key_eq(
11284 12 : all_layers,
11285 12 : vec![
11286 12 : PersistentLayerKey {
11287 12 : key_range: get_key(0)..get_key(2),
11288 12 : lsn_range: Lsn(0x20)..Lsn(0x21),
11289 12 : is_delta: false,
11290 12 : },
11291 12 : PersistentLayerKey {
11292 12 : key_range: get_key(0)..get_key(10),
11293 12 : lsn_range: Lsn(0x10)..Lsn(0x11),
11294 12 : is_delta: false,
11295 12 : },
11296 12 : // image layer generated for the compaction range 2-4
11297 12 : PersistentLayerKey {
11298 12 : key_range: get_key(2)..get_key(4),
11299 12 : lsn_range: Lsn(0x20)..Lsn(0x21),
11300 12 : is_delta: false,
11301 12 : },
11302 12 : // we have key2/key3 above the retain_lsn, so we still need this delta layer
11303 12 : PersistentLayerKey {
11304 12 : key_range: get_key(2)..get_key(4),
11305 12 : lsn_range: Lsn(0x20)..Lsn(0x48),
11306 12 : is_delta: true,
11307 12 : },
11308 12 : PersistentLayerKey {
11309 12 : key_range: get_key(5)..get_key(7),
11310 12 : lsn_range: Lsn(0x20)..Lsn(0x48),
11311 12 : is_delta: true,
11312 12 : },
11313 12 : PersistentLayerKey {
11314 12 : key_range: get_key(8)..get_key(10),
11315 12 : lsn_range: Lsn(0x48)..Lsn(0x50),
11316 12 : is_delta: true,
11317 12 : },
11318 12 : ],
11319 12 : );
11320 12 :
11321 12 : // Do a partial compaction on key range 4..9
11322 12 : tline
11323 12 : .compact_with_gc(
11324 12 : &cancel,
11325 12 : CompactOptions {
11326 12 : flags: EnumSet::new(),
11327 12 : compact_key_range: Some((get_key(4)..get_key(9)).into()),
11328 12 : ..Default::default()
11329 12 : },
11330 12 : &ctx,
11331 12 : )
11332 12 : .await
11333 12 : .unwrap();
11334 12 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11335 12 : check_layer_map_key_eq(
11336 12 : all_layers,
11337 12 : vec![
11338 12 : PersistentLayerKey {
11339 12 : key_range: get_key(0)..get_key(2),
11340 12 : lsn_range: Lsn(0x20)..Lsn(0x21),
11341 12 : is_delta: false,
11342 12 : },
11343 12 : PersistentLayerKey {
11344 12 : key_range: get_key(0)..get_key(10),
11345 12 : lsn_range: Lsn(0x10)..Lsn(0x11),
11346 12 : is_delta: false,
11347 12 : },
11348 12 : PersistentLayerKey {
11349 12 : key_range: get_key(2)..get_key(4),
11350 12 : lsn_range: Lsn(0x20)..Lsn(0x21),
11351 12 : is_delta: false,
11352 12 : },
11353 12 : PersistentLayerKey {
11354 12 : key_range: get_key(2)..get_key(4),
11355 12 : lsn_range: Lsn(0x20)..Lsn(0x48),
11356 12 : is_delta: true,
11357 12 : },
11358 12 : // image layer generated for this compaction range
11359 12 : PersistentLayerKey {
11360 12 : key_range: get_key(4)..get_key(9),
11361 12 : lsn_range: Lsn(0x20)..Lsn(0x21),
11362 12 : is_delta: false,
11363 12 : },
11364 12 : PersistentLayerKey {
11365 12 : key_range: get_key(8)..get_key(10),
11366 12 : lsn_range: Lsn(0x48)..Lsn(0x50),
11367 12 : is_delta: true,
11368 12 : },
11369 12 : ],
11370 12 : );
11371 12 :
11372 12 : // Do a partial compaction on key range 9..10
11373 12 : tline
11374 12 : .compact_with_gc(
11375 12 : &cancel,
11376 12 : CompactOptions {
11377 12 : flags: EnumSet::new(),
11378 12 : compact_key_range: Some((get_key(9)..get_key(10)).into()),
11379 12 : ..Default::default()
11380 12 : },
11381 12 : &ctx,
11382 12 : )
11383 12 : .await
11384 12 : .unwrap();
11385 12 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11386 12 : check_layer_map_key_eq(
11387 12 : all_layers,
11388 12 : vec![
11389 12 : PersistentLayerKey {
11390 12 : key_range: get_key(0)..get_key(2),
11391 12 : lsn_range: Lsn(0x20)..Lsn(0x21),
11392 12 : is_delta: false,
11393 12 : },
11394 12 : PersistentLayerKey {
11395 12 : key_range: get_key(0)..get_key(10),
11396 12 : lsn_range: Lsn(0x10)..Lsn(0x11),
11397 12 : is_delta: false,
11398 12 : },
11399 12 : PersistentLayerKey {
11400 12 : key_range: get_key(2)..get_key(4),
11401 12 : lsn_range: Lsn(0x20)..Lsn(0x21),
11402 12 : is_delta: false,
11403 12 : },
11404 12 : PersistentLayerKey {
11405 12 : key_range: get_key(2)..get_key(4),
11406 12 : lsn_range: Lsn(0x20)..Lsn(0x48),
11407 12 : is_delta: true,
11408 12 : },
11409 12 : PersistentLayerKey {
11410 12 : key_range: get_key(4)..get_key(9),
11411 12 : lsn_range: Lsn(0x20)..Lsn(0x21),
11412 12 : is_delta: false,
11413 12 : },
11414 12 : // image layer generated for the compaction range
11415 12 : PersistentLayerKey {
11416 12 : key_range: get_key(9)..get_key(10),
11417 12 : lsn_range: Lsn(0x20)..Lsn(0x21),
11418 12 : is_delta: false,
11419 12 : },
11420 12 : PersistentLayerKey {
11421 12 : key_range: get_key(8)..get_key(10),
11422 12 : lsn_range: Lsn(0x48)..Lsn(0x50),
11423 12 : is_delta: true,
11424 12 : },
11425 12 : ],
11426 12 : );
11427 12 :
11428 12 : // Do a partial compaction on key range 0..10, all image layers below LSN 20 can be replaced with new ones.
11429 12 : tline
11430 12 : .compact_with_gc(
11431 12 : &cancel,
11432 12 : CompactOptions {
11433 12 : flags: EnumSet::new(),
11434 12 : compact_key_range: Some((get_key(0)..get_key(10)).into()),
11435 12 : ..Default::default()
11436 12 : },
11437 12 : &ctx,
11438 12 : )
11439 12 : .await
11440 12 : .unwrap();
11441 12 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11442 12 : check_layer_map_key_eq(
11443 12 : all_layers,
11444 12 : vec![
11445 12 : // aha, we removed all unnecessary image/delta layers and got a very clean layer map!
11446 12 : PersistentLayerKey {
11447 12 : key_range: get_key(0)..get_key(10),
11448 12 : lsn_range: Lsn(0x20)..Lsn(0x21),
11449 12 : is_delta: false,
11450 12 : },
11451 12 : PersistentLayerKey {
11452 12 : key_range: get_key(2)..get_key(4),
11453 12 : lsn_range: Lsn(0x20)..Lsn(0x48),
11454 12 : is_delta: true,
11455 12 : },
11456 12 : PersistentLayerKey {
11457 12 : key_range: get_key(8)..get_key(10),
11458 12 : lsn_range: Lsn(0x48)..Lsn(0x50),
11459 12 : is_delta: true,
11460 12 : },
11461 12 : ],
11462 12 : );
11463 12 : Ok(())
11464 12 : }
11465 :
11466 : #[cfg(feature = "testing")]
11467 : #[tokio::test]
11468 12 : async fn test_timeline_offload_retain_lsn() -> anyhow::Result<()> {
11469 12 : let harness = TenantHarness::create("test_timeline_offload_retain_lsn")
11470 12 : .await
11471 12 : .unwrap();
11472 12 : let (tenant, ctx) = harness.load().await;
11473 12 : let tline_parent = tenant
11474 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
11475 12 : .await
11476 12 : .unwrap();
11477 12 : let tline_child = tenant
11478 12 : .branch_timeline_test(&tline_parent, NEW_TIMELINE_ID, Some(Lsn(0x20)), &ctx)
11479 12 : .await
11480 12 : .unwrap();
11481 12 : {
11482 12 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
11483 12 : assert_eq!(
11484 12 : gc_info_parent.retain_lsns,
11485 12 : vec![(Lsn(0x20), tline_child.timeline_id, MaybeOffloaded::No)]
11486 12 : );
11487 12 : }
11488 12 : // We have to directly call the remote_client instead of using the archive function to avoid constructing broker client...
11489 12 : tline_child
11490 12 : .remote_client
11491 12 : .schedule_index_upload_for_timeline_archival_state(TimelineArchivalState::Archived)
11492 12 : .unwrap();
11493 12 : tline_child.remote_client.wait_completion().await.unwrap();
11494 12 : offload_timeline(&tenant, &tline_child)
11495 12 : .instrument(tracing::info_span!(parent: None, "offload_test", tenant_id=%"test", shard_id=%"test", timeline_id=%"test"))
11496 12 : .await.unwrap();
11497 12 : let child_timeline_id = tline_child.timeline_id;
11498 12 : Arc::try_unwrap(tline_child).unwrap();
11499 12 :
11500 12 : {
11501 12 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
11502 12 : assert_eq!(
11503 12 : gc_info_parent.retain_lsns,
11504 12 : vec![(Lsn(0x20), child_timeline_id, MaybeOffloaded::Yes)]
11505 12 : );
11506 12 : }
11507 12 :
11508 12 : tenant
11509 12 : .get_offloaded_timeline(child_timeline_id)
11510 12 : .unwrap()
11511 12 : .defuse_for_tenant_drop();
11512 12 :
11513 12 : Ok(())
11514 12 : }
11515 :
11516 : #[cfg(feature = "testing")]
11517 : #[tokio::test]
11518 12 : async fn test_simple_bottom_most_compaction_above_lsn() -> anyhow::Result<()> {
11519 12 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_above_lsn").await?;
11520 12 : let (tenant, ctx) = harness.load().await;
11521 12 :
11522 1776 : fn get_key(id: u32) -> Key {
11523 1776 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
11524 1776 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
11525 1776 : key.field6 = id;
11526 1776 : key
11527 1776 : }
11528 12 :
11529 12 : let img_layer = (0..10)
11530 120 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
11531 12 : .collect_vec();
11532 12 :
11533 12 : let delta1 = vec![(
11534 12 : get_key(1),
11535 12 : Lsn(0x20),
11536 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
11537 12 : )];
11538 12 : let delta4 = vec![(
11539 12 : get_key(1),
11540 12 : Lsn(0x28),
11541 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
11542 12 : )];
11543 12 : let delta2 = vec![
11544 12 : (
11545 12 : get_key(1),
11546 12 : Lsn(0x30),
11547 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
11548 12 : ),
11549 12 : (
11550 12 : get_key(1),
11551 12 : Lsn(0x38),
11552 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
11553 12 : ),
11554 12 : ];
11555 12 : let delta3 = vec![
11556 12 : (
11557 12 : get_key(8),
11558 12 : Lsn(0x48),
11559 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
11560 12 : ),
11561 12 : (
11562 12 : get_key(9),
11563 12 : Lsn(0x48),
11564 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
11565 12 : ),
11566 12 : ];
11567 12 :
11568 12 : let tline = tenant
11569 12 : .create_test_timeline_with_layers(
11570 12 : TIMELINE_ID,
11571 12 : Lsn(0x10),
11572 12 : DEFAULT_PG_VERSION,
11573 12 : &ctx,
11574 12 : vec![], // in-memory layers
11575 12 : vec![
11576 12 : // delta1/2/4 only contain a single key but multiple updates
11577 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x28), delta1),
11578 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
11579 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x28)..Lsn(0x30), delta4),
11580 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta3),
11581 12 : ], // delta layers
11582 12 : vec![(Lsn(0x10), img_layer)], // image layers
11583 12 : Lsn(0x50),
11584 12 : )
11585 12 : .await?;
11586 12 : {
11587 12 : tline
11588 12 : .applied_gc_cutoff_lsn
11589 12 : .lock_for_write()
11590 12 : .store_and_unlock(Lsn(0x30))
11591 12 : .wait()
11592 12 : .await;
11593 12 : // Update GC info
11594 12 : let mut guard = tline.gc_info.write().unwrap();
11595 12 : *guard = GcInfo {
11596 12 : retain_lsns: vec![
11597 12 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
11598 12 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
11599 12 : ],
11600 12 : cutoffs: GcCutoffs {
11601 12 : time: Lsn(0x30),
11602 12 : space: Lsn(0x30),
11603 12 : },
11604 12 : leases: Default::default(),
11605 12 : within_ancestor_pitr: false,
11606 12 : };
11607 12 : }
11608 12 :
11609 12 : let expected_result = [
11610 12 : Bytes::from_static(b"value 0@0x10"),
11611 12 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
11612 12 : Bytes::from_static(b"value 2@0x10"),
11613 12 : Bytes::from_static(b"value 3@0x10"),
11614 12 : Bytes::from_static(b"value 4@0x10"),
11615 12 : Bytes::from_static(b"value 5@0x10"),
11616 12 : Bytes::from_static(b"value 6@0x10"),
11617 12 : Bytes::from_static(b"value 7@0x10"),
11618 12 : Bytes::from_static(b"value 8@0x10@0x48"),
11619 12 : Bytes::from_static(b"value 9@0x10@0x48"),
11620 12 : ];
11621 12 :
11622 12 : let expected_result_at_gc_horizon = [
11623 12 : Bytes::from_static(b"value 0@0x10"),
11624 12 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
11625 12 : Bytes::from_static(b"value 2@0x10"),
11626 12 : Bytes::from_static(b"value 3@0x10"),
11627 12 : Bytes::from_static(b"value 4@0x10"),
11628 12 : Bytes::from_static(b"value 5@0x10"),
11629 12 : Bytes::from_static(b"value 6@0x10"),
11630 12 : Bytes::from_static(b"value 7@0x10"),
11631 12 : Bytes::from_static(b"value 8@0x10"),
11632 12 : Bytes::from_static(b"value 9@0x10"),
11633 12 : ];
11634 12 :
11635 12 : let expected_result_at_lsn_20 = [
11636 12 : Bytes::from_static(b"value 0@0x10"),
11637 12 : Bytes::from_static(b"value 1@0x10@0x20"),
11638 12 : Bytes::from_static(b"value 2@0x10"),
11639 12 : Bytes::from_static(b"value 3@0x10"),
11640 12 : Bytes::from_static(b"value 4@0x10"),
11641 12 : Bytes::from_static(b"value 5@0x10"),
11642 12 : Bytes::from_static(b"value 6@0x10"),
11643 12 : Bytes::from_static(b"value 7@0x10"),
11644 12 : Bytes::from_static(b"value 8@0x10"),
11645 12 : Bytes::from_static(b"value 9@0x10"),
11646 12 : ];
11647 12 :
11648 12 : let expected_result_at_lsn_10 = [
11649 12 : Bytes::from_static(b"value 0@0x10"),
11650 12 : Bytes::from_static(b"value 1@0x10"),
11651 12 : Bytes::from_static(b"value 2@0x10"),
11652 12 : Bytes::from_static(b"value 3@0x10"),
11653 12 : Bytes::from_static(b"value 4@0x10"),
11654 12 : Bytes::from_static(b"value 5@0x10"),
11655 12 : Bytes::from_static(b"value 6@0x10"),
11656 12 : Bytes::from_static(b"value 7@0x10"),
11657 12 : Bytes::from_static(b"value 8@0x10"),
11658 12 : Bytes::from_static(b"value 9@0x10"),
11659 12 : ];
11660 12 :
11661 36 : let verify_result = || async {
11662 36 : let gc_horizon = {
11663 36 : let gc_info = tline.gc_info.read().unwrap();
11664 36 : gc_info.cutoffs.time
11665 12 : };
11666 396 : for idx in 0..10 {
11667 360 : assert_eq!(
11668 360 : tline
11669 360 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
11670 360 : .await
11671 360 : .unwrap(),
11672 360 : &expected_result[idx]
11673 12 : );
11674 360 : assert_eq!(
11675 360 : tline
11676 360 : .get(get_key(idx as u32), gc_horizon, &ctx)
11677 360 : .await
11678 360 : .unwrap(),
11679 360 : &expected_result_at_gc_horizon[idx]
11680 12 : );
11681 360 : assert_eq!(
11682 360 : tline
11683 360 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
11684 360 : .await
11685 360 : .unwrap(),
11686 360 : &expected_result_at_lsn_20[idx]
11687 12 : );
11688 360 : assert_eq!(
11689 360 : tline
11690 360 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
11691 360 : .await
11692 360 : .unwrap(),
11693 360 : &expected_result_at_lsn_10[idx]
11694 12 : );
11695 12 : }
11696 72 : };
11697 12 :
11698 12 : verify_result().await;
11699 12 :
11700 12 : let cancel = CancellationToken::new();
11701 12 : tline
11702 12 : .compact_with_gc(
11703 12 : &cancel,
11704 12 : CompactOptions {
11705 12 : compact_lsn_range: Some(CompactLsnRange::above(Lsn(0x28))),
11706 12 : ..Default::default()
11707 12 : },
11708 12 : &ctx,
11709 12 : )
11710 12 : .await
11711 12 : .unwrap();
11712 12 : verify_result().await;
11713 12 :
11714 12 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11715 12 : check_layer_map_key_eq(
11716 12 : all_layers,
11717 12 : vec![
11718 12 : // The original image layer, not compacted
11719 12 : PersistentLayerKey {
11720 12 : key_range: get_key(0)..get_key(10),
11721 12 : lsn_range: Lsn(0x10)..Lsn(0x11),
11722 12 : is_delta: false,
11723 12 : },
11724 12 : // Delta layer below the specified above_lsn not compacted
11725 12 : PersistentLayerKey {
11726 12 : key_range: get_key(1)..get_key(2),
11727 12 : lsn_range: Lsn(0x20)..Lsn(0x28),
11728 12 : is_delta: true,
11729 12 : },
11730 12 : // Delta layer compacted above the LSN
11731 12 : PersistentLayerKey {
11732 12 : key_range: get_key(1)..get_key(10),
11733 12 : lsn_range: Lsn(0x28)..Lsn(0x50),
11734 12 : is_delta: true,
11735 12 : },
11736 12 : ],
11737 12 : );
11738 12 :
11739 12 : // compact again
11740 12 : tline
11741 12 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
11742 12 : .await
11743 12 : .unwrap();
11744 12 : verify_result().await;
11745 12 :
11746 12 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11747 12 : check_layer_map_key_eq(
11748 12 : all_layers,
11749 12 : vec![
11750 12 : // The compacted image layer (full key range)
11751 12 : PersistentLayerKey {
11752 12 : key_range: Key::MIN..Key::MAX,
11753 12 : lsn_range: Lsn(0x10)..Lsn(0x11),
11754 12 : is_delta: false,
11755 12 : },
11756 12 : // All other data in the delta layer
11757 12 : PersistentLayerKey {
11758 12 : key_range: get_key(1)..get_key(10),
11759 12 : lsn_range: Lsn(0x10)..Lsn(0x50),
11760 12 : is_delta: true,
11761 12 : },
11762 12 : ],
11763 12 : );
11764 12 :
11765 12 : Ok(())
11766 12 : }
11767 :
11768 : #[cfg(feature = "testing")]
11769 : #[tokio::test]
11770 12 : async fn test_simple_bottom_most_compaction_rectangle() -> anyhow::Result<()> {
11771 12 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_rectangle").await?;
11772 12 : let (tenant, ctx) = harness.load().await;
11773 12 :
11774 3048 : fn get_key(id: u32) -> Key {
11775 3048 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
11776 3048 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
11777 3048 : key.field6 = id;
11778 3048 : key
11779 3048 : }
11780 12 :
11781 12 : let img_layer = (0..10)
11782 120 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
11783 12 : .collect_vec();
11784 12 :
11785 12 : let delta1 = vec![(
11786 12 : get_key(1),
11787 12 : Lsn(0x20),
11788 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
11789 12 : )];
11790 12 : let delta4 = vec![(
11791 12 : get_key(1),
11792 12 : Lsn(0x28),
11793 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
11794 12 : )];
11795 12 : let delta2 = vec![
11796 12 : (
11797 12 : get_key(1),
11798 12 : Lsn(0x30),
11799 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
11800 12 : ),
11801 12 : (
11802 12 : get_key(1),
11803 12 : Lsn(0x38),
11804 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
11805 12 : ),
11806 12 : ];
11807 12 : let delta3 = vec![
11808 12 : (
11809 12 : get_key(8),
11810 12 : Lsn(0x48),
11811 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
11812 12 : ),
11813 12 : (
11814 12 : get_key(9),
11815 12 : Lsn(0x48),
11816 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
11817 12 : ),
11818 12 : ];
11819 12 :
11820 12 : let tline = tenant
11821 12 : .create_test_timeline_with_layers(
11822 12 : TIMELINE_ID,
11823 12 : Lsn(0x10),
11824 12 : DEFAULT_PG_VERSION,
11825 12 : &ctx,
11826 12 : vec![], // in-memory layers
11827 12 : vec![
11828 12 : // delta1/2/4 only contain a single key but multiple updates
11829 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x28), delta1),
11830 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
11831 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x28)..Lsn(0x30), delta4),
11832 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta3),
11833 12 : ], // delta layers
11834 12 : vec![(Lsn(0x10), img_layer)], // image layers
11835 12 : Lsn(0x50),
11836 12 : )
11837 12 : .await?;
11838 12 : {
11839 12 : tline
11840 12 : .applied_gc_cutoff_lsn
11841 12 : .lock_for_write()
11842 12 : .store_and_unlock(Lsn(0x30))
11843 12 : .wait()
11844 12 : .await;
11845 12 : // Update GC info
11846 12 : let mut guard = tline.gc_info.write().unwrap();
11847 12 : *guard = GcInfo {
11848 12 : retain_lsns: vec![
11849 12 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
11850 12 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
11851 12 : ],
11852 12 : cutoffs: GcCutoffs {
11853 12 : time: Lsn(0x30),
11854 12 : space: Lsn(0x30),
11855 12 : },
11856 12 : leases: Default::default(),
11857 12 : within_ancestor_pitr: false,
11858 12 : };
11859 12 : }
11860 12 :
11861 12 : let expected_result = [
11862 12 : Bytes::from_static(b"value 0@0x10"),
11863 12 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
11864 12 : Bytes::from_static(b"value 2@0x10"),
11865 12 : Bytes::from_static(b"value 3@0x10"),
11866 12 : Bytes::from_static(b"value 4@0x10"),
11867 12 : Bytes::from_static(b"value 5@0x10"),
11868 12 : Bytes::from_static(b"value 6@0x10"),
11869 12 : Bytes::from_static(b"value 7@0x10"),
11870 12 : Bytes::from_static(b"value 8@0x10@0x48"),
11871 12 : Bytes::from_static(b"value 9@0x10@0x48"),
11872 12 : ];
11873 12 :
11874 12 : let expected_result_at_gc_horizon = [
11875 12 : Bytes::from_static(b"value 0@0x10"),
11876 12 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
11877 12 : Bytes::from_static(b"value 2@0x10"),
11878 12 : Bytes::from_static(b"value 3@0x10"),
11879 12 : Bytes::from_static(b"value 4@0x10"),
11880 12 : Bytes::from_static(b"value 5@0x10"),
11881 12 : Bytes::from_static(b"value 6@0x10"),
11882 12 : Bytes::from_static(b"value 7@0x10"),
11883 12 : Bytes::from_static(b"value 8@0x10"),
11884 12 : Bytes::from_static(b"value 9@0x10"),
11885 12 : ];
11886 12 :
11887 12 : let expected_result_at_lsn_20 = [
11888 12 : Bytes::from_static(b"value 0@0x10"),
11889 12 : Bytes::from_static(b"value 1@0x10@0x20"),
11890 12 : Bytes::from_static(b"value 2@0x10"),
11891 12 : Bytes::from_static(b"value 3@0x10"),
11892 12 : Bytes::from_static(b"value 4@0x10"),
11893 12 : Bytes::from_static(b"value 5@0x10"),
11894 12 : Bytes::from_static(b"value 6@0x10"),
11895 12 : Bytes::from_static(b"value 7@0x10"),
11896 12 : Bytes::from_static(b"value 8@0x10"),
11897 12 : Bytes::from_static(b"value 9@0x10"),
11898 12 : ];
11899 12 :
11900 12 : let expected_result_at_lsn_10 = [
11901 12 : Bytes::from_static(b"value 0@0x10"),
11902 12 : Bytes::from_static(b"value 1@0x10"),
11903 12 : Bytes::from_static(b"value 2@0x10"),
11904 12 : Bytes::from_static(b"value 3@0x10"),
11905 12 : Bytes::from_static(b"value 4@0x10"),
11906 12 : Bytes::from_static(b"value 5@0x10"),
11907 12 : Bytes::from_static(b"value 6@0x10"),
11908 12 : Bytes::from_static(b"value 7@0x10"),
11909 12 : Bytes::from_static(b"value 8@0x10"),
11910 12 : Bytes::from_static(b"value 9@0x10"),
11911 12 : ];
11912 12 :
11913 60 : let verify_result = || async {
11914 60 : let gc_horizon = {
11915 60 : let gc_info = tline.gc_info.read().unwrap();
11916 60 : gc_info.cutoffs.time
11917 12 : };
11918 660 : for idx in 0..10 {
11919 600 : assert_eq!(
11920 600 : tline
11921 600 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
11922 600 : .await
11923 600 : .unwrap(),
11924 600 : &expected_result[idx]
11925 12 : );
11926 600 : assert_eq!(
11927 600 : tline
11928 600 : .get(get_key(idx as u32), gc_horizon, &ctx)
11929 600 : .await
11930 600 : .unwrap(),
11931 600 : &expected_result_at_gc_horizon[idx]
11932 12 : );
11933 600 : assert_eq!(
11934 600 : tline
11935 600 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
11936 600 : .await
11937 600 : .unwrap(),
11938 600 : &expected_result_at_lsn_20[idx]
11939 12 : );
11940 600 : assert_eq!(
11941 600 : tline
11942 600 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
11943 600 : .await
11944 600 : .unwrap(),
11945 600 : &expected_result_at_lsn_10[idx]
11946 12 : );
11947 12 : }
11948 120 : };
11949 12 :
11950 12 : verify_result().await;
11951 12 :
11952 12 : let cancel = CancellationToken::new();
11953 12 :
11954 12 : tline
11955 12 : .compact_with_gc(
11956 12 : &cancel,
11957 12 : CompactOptions {
11958 12 : compact_key_range: Some((get_key(0)..get_key(2)).into()),
11959 12 : compact_lsn_range: Some((Lsn(0x20)..Lsn(0x28)).into()),
11960 12 : ..Default::default()
11961 12 : },
11962 12 : &ctx,
11963 12 : )
11964 12 : .await
11965 12 : .unwrap();
11966 12 : verify_result().await;
11967 12 :
11968 12 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11969 12 : check_layer_map_key_eq(
11970 12 : all_layers,
11971 12 : vec![
11972 12 : // The original image layer, not compacted
11973 12 : PersistentLayerKey {
11974 12 : key_range: get_key(0)..get_key(10),
11975 12 : lsn_range: Lsn(0x10)..Lsn(0x11),
11976 12 : is_delta: false,
11977 12 : },
11978 12 : // According the selection logic, we select all layers with start key <= 0x28, so we would merge the layer 0x20-0x28 and
11979 12 : // the layer 0x28-0x30 into one.
11980 12 : PersistentLayerKey {
11981 12 : key_range: get_key(1)..get_key(2),
11982 12 : lsn_range: Lsn(0x20)..Lsn(0x30),
11983 12 : is_delta: true,
11984 12 : },
11985 12 : // Above the upper bound and untouched
11986 12 : PersistentLayerKey {
11987 12 : key_range: get_key(1)..get_key(2),
11988 12 : lsn_range: Lsn(0x30)..Lsn(0x50),
11989 12 : is_delta: true,
11990 12 : },
11991 12 : // This layer is untouched
11992 12 : PersistentLayerKey {
11993 12 : key_range: get_key(8)..get_key(10),
11994 12 : lsn_range: Lsn(0x30)..Lsn(0x50),
11995 12 : is_delta: true,
11996 12 : },
11997 12 : ],
11998 12 : );
11999 12 :
12000 12 : tline
12001 12 : .compact_with_gc(
12002 12 : &cancel,
12003 12 : CompactOptions {
12004 12 : compact_key_range: Some((get_key(3)..get_key(8)).into()),
12005 12 : compact_lsn_range: Some((Lsn(0x28)..Lsn(0x40)).into()),
12006 12 : ..Default::default()
12007 12 : },
12008 12 : &ctx,
12009 12 : )
12010 12 : .await
12011 12 : .unwrap();
12012 12 : verify_result().await;
12013 12 :
12014 12 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12015 12 : check_layer_map_key_eq(
12016 12 : all_layers,
12017 12 : vec![
12018 12 : // The original image layer, not compacted
12019 12 : PersistentLayerKey {
12020 12 : key_range: get_key(0)..get_key(10),
12021 12 : lsn_range: Lsn(0x10)..Lsn(0x11),
12022 12 : is_delta: false,
12023 12 : },
12024 12 : // Not in the compaction key range, uncompacted
12025 12 : PersistentLayerKey {
12026 12 : key_range: get_key(1)..get_key(2),
12027 12 : lsn_range: Lsn(0x20)..Lsn(0x30),
12028 12 : is_delta: true,
12029 12 : },
12030 12 : // Not in the compaction key range, uncompacted but need rewrite because the delta layer overlaps with the range
12031 12 : PersistentLayerKey {
12032 12 : key_range: get_key(1)..get_key(2),
12033 12 : lsn_range: Lsn(0x30)..Lsn(0x50),
12034 12 : is_delta: true,
12035 12 : },
12036 12 : // Note that when we specify the LSN upper bound to be 0x40, the compaction algorithm will not try to cut the layer
12037 12 : // horizontally in half. Instead, it will include all LSNs that overlap with 0x40. So the real max_lsn of the compaction
12038 12 : // becomes 0x50.
12039 12 : PersistentLayerKey {
12040 12 : key_range: get_key(8)..get_key(10),
12041 12 : lsn_range: Lsn(0x30)..Lsn(0x50),
12042 12 : is_delta: true,
12043 12 : },
12044 12 : ],
12045 12 : );
12046 12 :
12047 12 : // compact again
12048 12 : tline
12049 12 : .compact_with_gc(
12050 12 : &cancel,
12051 12 : CompactOptions {
12052 12 : compact_key_range: Some((get_key(0)..get_key(5)).into()),
12053 12 : compact_lsn_range: Some((Lsn(0x20)..Lsn(0x50)).into()),
12054 12 : ..Default::default()
12055 12 : },
12056 12 : &ctx,
12057 12 : )
12058 12 : .await
12059 12 : .unwrap();
12060 12 : verify_result().await;
12061 12 :
12062 12 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12063 12 : check_layer_map_key_eq(
12064 12 : all_layers,
12065 12 : vec![
12066 12 : // The original image layer, not compacted
12067 12 : PersistentLayerKey {
12068 12 : key_range: get_key(0)..get_key(10),
12069 12 : lsn_range: Lsn(0x10)..Lsn(0x11),
12070 12 : is_delta: false,
12071 12 : },
12072 12 : // The range gets compacted
12073 12 : PersistentLayerKey {
12074 12 : key_range: get_key(1)..get_key(2),
12075 12 : lsn_range: Lsn(0x20)..Lsn(0x50),
12076 12 : is_delta: true,
12077 12 : },
12078 12 : // Not touched during this iteration of compaction
12079 12 : PersistentLayerKey {
12080 12 : key_range: get_key(8)..get_key(10),
12081 12 : lsn_range: Lsn(0x30)..Lsn(0x50),
12082 12 : is_delta: true,
12083 12 : },
12084 12 : ],
12085 12 : );
12086 12 :
12087 12 : // final full compaction
12088 12 : tline
12089 12 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
12090 12 : .await
12091 12 : .unwrap();
12092 12 : verify_result().await;
12093 12 :
12094 12 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12095 12 : check_layer_map_key_eq(
12096 12 : all_layers,
12097 12 : vec![
12098 12 : // The compacted image layer (full key range)
12099 12 : PersistentLayerKey {
12100 12 : key_range: Key::MIN..Key::MAX,
12101 12 : lsn_range: Lsn(0x10)..Lsn(0x11),
12102 12 : is_delta: false,
12103 12 : },
12104 12 : // All other data in the delta layer
12105 12 : PersistentLayerKey {
12106 12 : key_range: get_key(1)..get_key(10),
12107 12 : lsn_range: Lsn(0x10)..Lsn(0x50),
12108 12 : is_delta: true,
12109 12 : },
12110 12 : ],
12111 12 : );
12112 12 :
12113 12 : Ok(())
12114 12 : }
12115 :
12116 : #[cfg(feature = "testing")]
12117 : #[tokio::test]
12118 12 : async fn test_bottom_most_compation_redo_failure() -> anyhow::Result<()> {
12119 12 : let harness = TenantHarness::create("test_bottom_most_compation_redo_failure").await?;
12120 12 : let (tenant, ctx) = harness.load().await;
12121 12 :
12122 156 : fn get_key(id: u32) -> Key {
12123 156 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
12124 156 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
12125 156 : key.field6 = id;
12126 156 : key
12127 156 : }
12128 12 :
12129 12 : let img_layer = (0..10)
12130 120 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
12131 12 : .collect_vec();
12132 12 :
12133 12 : let delta1 = vec![
12134 12 : (
12135 12 : get_key(1),
12136 12 : Lsn(0x20),
12137 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
12138 12 : ),
12139 12 : (
12140 12 : get_key(1),
12141 12 : Lsn(0x24),
12142 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x24")),
12143 12 : ),
12144 12 : (
12145 12 : get_key(1),
12146 12 : Lsn(0x28),
12147 12 : // This record will fail to redo
12148 12 : Value::WalRecord(NeonWalRecord::wal_append_conditional("@0x28", "???")),
12149 12 : ),
12150 12 : ];
12151 12 :
12152 12 : let tline = tenant
12153 12 : .create_test_timeline_with_layers(
12154 12 : TIMELINE_ID,
12155 12 : Lsn(0x10),
12156 12 : DEFAULT_PG_VERSION,
12157 12 : &ctx,
12158 12 : vec![], // in-memory layers
12159 12 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
12160 12 : Lsn(0x20)..Lsn(0x30),
12161 12 : delta1,
12162 12 : )], // delta layers
12163 12 : vec![(Lsn(0x10), img_layer)], // image layers
12164 12 : Lsn(0x50),
12165 12 : )
12166 12 : .await?;
12167 12 : {
12168 12 : tline
12169 12 : .applied_gc_cutoff_lsn
12170 12 : .lock_for_write()
12171 12 : .store_and_unlock(Lsn(0x30))
12172 12 : .wait()
12173 12 : .await;
12174 12 : // Update GC info
12175 12 : let mut guard = tline.gc_info.write().unwrap();
12176 12 : *guard = GcInfo {
12177 12 : retain_lsns: vec![],
12178 12 : cutoffs: GcCutoffs {
12179 12 : time: Lsn(0x30),
12180 12 : space: Lsn(0x30),
12181 12 : },
12182 12 : leases: Default::default(),
12183 12 : within_ancestor_pitr: false,
12184 12 : };
12185 12 : }
12186 12 :
12187 12 : let cancel = CancellationToken::new();
12188 12 :
12189 12 : // Compaction will fail, but should not fire any critical error.
12190 12 : // Gc-compaction currently cannot figure out what keys are not in the keyspace during the compaction
12191 12 : // process. It will always try to redo the logs it reads and if it doesn't work, fail the entire
12192 12 : // compaction job. Tracked in <https://github.com/neondatabase/neon/issues/10395>.
12193 12 : let res = tline
12194 12 : .compact_with_gc(
12195 12 : &cancel,
12196 12 : CompactOptions {
12197 12 : compact_key_range: None,
12198 12 : compact_lsn_range: None,
12199 12 : ..Default::default()
12200 12 : },
12201 12 : &ctx,
12202 12 : )
12203 12 : .await;
12204 12 : assert!(res.is_err());
12205 12 :
12206 12 : Ok(())
12207 12 : }
12208 :
12209 : #[cfg(feature = "testing")]
12210 : #[tokio::test]
12211 12 : async fn test_synthetic_size_calculation_with_invisible_branches() -> anyhow::Result<()> {
12212 12 : use pageserver_api::models::TimelineVisibilityState;
12213 12 :
12214 12 : use crate::tenant::size::gather_inputs;
12215 12 :
12216 12 : let tenant_conf = pageserver_api::models::TenantConfig {
12217 12 : // Ensure that we don't compute gc_cutoffs (which needs reading the layer files)
12218 12 : pitr_interval: Some(Duration::ZERO),
12219 12 : ..Default::default()
12220 12 : };
12221 12 : let harness = TenantHarness::create_custom(
12222 12 : "test_synthetic_size_calculation_with_invisible_branches",
12223 12 : tenant_conf,
12224 12 : TenantId::generate(),
12225 12 : ShardIdentity::unsharded(),
12226 12 : Generation::new(0xdeadbeef),
12227 12 : )
12228 12 : .await?;
12229 12 : let (tenant, ctx) = harness.load().await;
12230 12 : let main_tline = tenant
12231 12 : .create_test_timeline_with_layers(
12232 12 : TIMELINE_ID,
12233 12 : Lsn(0x10),
12234 12 : DEFAULT_PG_VERSION,
12235 12 : &ctx,
12236 12 : vec![],
12237 12 : vec![],
12238 12 : vec![],
12239 12 : Lsn(0x100),
12240 12 : )
12241 12 : .await?;
12242 12 :
12243 12 : let snapshot1 = TimelineId::from_array(hex!("11223344556677881122334455667790"));
12244 12 : tenant
12245 12 : .branch_timeline_test_with_layers(
12246 12 : &main_tline,
12247 12 : snapshot1,
12248 12 : Some(Lsn(0x20)),
12249 12 : &ctx,
12250 12 : vec![],
12251 12 : vec![],
12252 12 : Lsn(0x50),
12253 12 : )
12254 12 : .await?;
12255 12 : let snapshot2 = TimelineId::from_array(hex!("11223344556677881122334455667791"));
12256 12 : tenant
12257 12 : .branch_timeline_test_with_layers(
12258 12 : &main_tline,
12259 12 : snapshot2,
12260 12 : Some(Lsn(0x30)),
12261 12 : &ctx,
12262 12 : vec![],
12263 12 : vec![],
12264 12 : Lsn(0x50),
12265 12 : )
12266 12 : .await?;
12267 12 : let snapshot3 = TimelineId::from_array(hex!("11223344556677881122334455667792"));
12268 12 : tenant
12269 12 : .branch_timeline_test_with_layers(
12270 12 : &main_tline,
12271 12 : snapshot3,
12272 12 : Some(Lsn(0x40)),
12273 12 : &ctx,
12274 12 : vec![],
12275 12 : vec![],
12276 12 : Lsn(0x50),
12277 12 : )
12278 12 : .await?;
12279 12 : let limit = Arc::new(Semaphore::new(1));
12280 12 : let max_retention_period = None;
12281 12 : let mut logical_size_cache = HashMap::new();
12282 12 : let cause = LogicalSizeCalculationCause::EvictionTaskImitation;
12283 12 : let cancel = CancellationToken::new();
12284 12 :
12285 12 : let inputs = gather_inputs(
12286 12 : &tenant,
12287 12 : &limit,
12288 12 : max_retention_period,
12289 12 : &mut logical_size_cache,
12290 12 : cause,
12291 12 : &cancel,
12292 12 : &ctx,
12293 12 : )
12294 12 : .instrument(info_span!(
12295 12 : "gather_inputs",
12296 12 : tenant_id = "unknown",
12297 12 : shard_id = "unknown",
12298 12 : ))
12299 12 : .await?;
12300 12 : use crate::tenant::size::{LsnKind, ModelInputs, SegmentMeta};
12301 12 : use LsnKind::*;
12302 12 : use tenant_size_model::Segment;
12303 12 : let ModelInputs { mut segments, .. } = inputs;
12304 180 : segments.retain(|s| s.timeline_id == TIMELINE_ID);
12305 72 : for segment in segments.iter_mut() {
12306 72 : segment.segment.parent = None; // We don't care about the parent for the test
12307 72 : segment.segment.size = None; // We don't care about the size for the test
12308 72 : }
12309 12 : assert_eq!(
12310 12 : segments,
12311 12 : [
12312 12 : SegmentMeta {
12313 12 : segment: Segment {
12314 12 : parent: None,
12315 12 : lsn: 0x10,
12316 12 : size: None,
12317 12 : needed: false,
12318 12 : },
12319 12 : timeline_id: TIMELINE_ID,
12320 12 : kind: BranchStart,
12321 12 : },
12322 12 : SegmentMeta {
12323 12 : segment: Segment {
12324 12 : parent: None,
12325 12 : lsn: 0x20,
12326 12 : size: None,
12327 12 : needed: false,
12328 12 : },
12329 12 : timeline_id: TIMELINE_ID,
12330 12 : kind: BranchPoint,
12331 12 : },
12332 12 : SegmentMeta {
12333 12 : segment: Segment {
12334 12 : parent: None,
12335 12 : lsn: 0x30,
12336 12 : size: None,
12337 12 : needed: false,
12338 12 : },
12339 12 : timeline_id: TIMELINE_ID,
12340 12 : kind: BranchPoint,
12341 12 : },
12342 12 : SegmentMeta {
12343 12 : segment: Segment {
12344 12 : parent: None,
12345 12 : lsn: 0x40,
12346 12 : size: None,
12347 12 : needed: false,
12348 12 : },
12349 12 : timeline_id: TIMELINE_ID,
12350 12 : kind: BranchPoint,
12351 12 : },
12352 12 : SegmentMeta {
12353 12 : segment: Segment {
12354 12 : parent: None,
12355 12 : lsn: 0x100,
12356 12 : size: None,
12357 12 : needed: false,
12358 12 : },
12359 12 : timeline_id: TIMELINE_ID,
12360 12 : kind: GcCutOff,
12361 12 : }, // we need to retain everything above the last branch point
12362 12 : SegmentMeta {
12363 12 : segment: Segment {
12364 12 : parent: None,
12365 12 : lsn: 0x100,
12366 12 : size: None,
12367 12 : needed: true,
12368 12 : },
12369 12 : timeline_id: TIMELINE_ID,
12370 12 : kind: BranchEnd,
12371 12 : },
12372 12 : ]
12373 12 : );
12374 12 :
12375 12 : main_tline
12376 12 : .remote_client
12377 12 : .schedule_index_upload_for_timeline_invisible_state(
12378 12 : TimelineVisibilityState::Invisible,
12379 12 : )?;
12380 12 : main_tline.remote_client.wait_completion().await?;
12381 12 : let inputs = gather_inputs(
12382 12 : &tenant,
12383 12 : &limit,
12384 12 : max_retention_period,
12385 12 : &mut logical_size_cache,
12386 12 : cause,
12387 12 : &cancel,
12388 12 : &ctx,
12389 12 : )
12390 12 : .instrument(info_span!(
12391 12 : "gather_inputs",
12392 12 : tenant_id = "unknown",
12393 12 : shard_id = "unknown",
12394 12 : ))
12395 12 : .await?;
12396 12 : let ModelInputs { mut segments, .. } = inputs;
12397 168 : segments.retain(|s| s.timeline_id == TIMELINE_ID);
12398 60 : for segment in segments.iter_mut() {
12399 60 : segment.segment.parent = None; // We don't care about the parent for the test
12400 60 : segment.segment.size = None; // We don't care about the size for the test
12401 60 : }
12402 12 : assert_eq!(
12403 12 : segments,
12404 12 : [
12405 12 : SegmentMeta {
12406 12 : segment: Segment {
12407 12 : parent: None,
12408 12 : lsn: 0x10,
12409 12 : size: None,
12410 12 : needed: false,
12411 12 : },
12412 12 : timeline_id: TIMELINE_ID,
12413 12 : kind: BranchStart,
12414 12 : },
12415 12 : SegmentMeta {
12416 12 : segment: Segment {
12417 12 : parent: None,
12418 12 : lsn: 0x20,
12419 12 : size: None,
12420 12 : needed: false,
12421 12 : },
12422 12 : timeline_id: TIMELINE_ID,
12423 12 : kind: BranchPoint,
12424 12 : },
12425 12 : SegmentMeta {
12426 12 : segment: Segment {
12427 12 : parent: None,
12428 12 : lsn: 0x30,
12429 12 : size: None,
12430 12 : needed: false,
12431 12 : },
12432 12 : timeline_id: TIMELINE_ID,
12433 12 : kind: BranchPoint,
12434 12 : },
12435 12 : SegmentMeta {
12436 12 : segment: Segment {
12437 12 : parent: None,
12438 12 : lsn: 0x40,
12439 12 : size: None,
12440 12 : needed: false,
12441 12 : },
12442 12 : timeline_id: TIMELINE_ID,
12443 12 : kind: BranchPoint,
12444 12 : },
12445 12 : SegmentMeta {
12446 12 : segment: Segment {
12447 12 : parent: None,
12448 12 : lsn: 0x40, // Branch end LSN == last branch point LSN
12449 12 : size: None,
12450 12 : needed: true,
12451 12 : },
12452 12 : timeline_id: TIMELINE_ID,
12453 12 : kind: BranchEnd,
12454 12 : },
12455 12 : ]
12456 12 : );
12457 12 : Ok(())
12458 12 : }
12459 : }
|
