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 1404 : fn new(
175 1404 : tenant_conf: pageserver_api::models::TenantConfig,
176 1404 : location: AttachedLocationConfig,
177 1404 : ) -> 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 1404 : let lsn_lease_deadline = if location.attach_mode == AttachmentMode::Single {
184 1404 : Some(
185 1404 : tokio::time::Instant::now()
186 1404 : + tenant_conf
187 1404 : .lsn_lease_length
188 1404 : .unwrap_or(LsnLease::DEFAULT_LENGTH),
189 1404 : )
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 1404 : Self {
197 1404 : tenant_conf,
198 1404 : location,
199 1404 : lsn_lease_deadline,
200 1404 : }
201 1404 : }
202 :
203 1404 : fn try_from(location_conf: LocationConf) -> anyhow::Result<Self> {
204 1404 : match &location_conf.mode {
205 1404 : LocationMode::Attached(attach_conf) => {
206 1404 : 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 1404 : }
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 1404 : fn from(mgr: harness::TestRedoManager) -> Self {
440 1404 : Self::Test(mgr)
441 1404 : }
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 1404 : async fn read_on_disk_heatmap(&self) -> Option<(HeatMapTenant, std::time::Instant)> {
1601 1404 : if !self.conf.load_previous_heatmap {
1602 0 : return None;
1603 1404 : }
1604 1404 :
1605 1404 : let on_disk_heatmap_path = self.conf.tenant_heatmap_path(&self.tenant_shard_id);
1606 1404 : 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 1404 : Err(err) => match err.kind() {
1615 1404 : std::io::ErrorKind::NotFound => None,
1616 : _ => {
1617 0 : error!("Unexpected IO error reading old heatmap: {err}");
1618 0 : None
1619 : }
1620 : },
1621 : }
1622 1404 : }
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 1404 : async fn attach(
1630 1404 : self: &Arc<TenantShard>,
1631 1404 : preload: Option<TenantPreload>,
1632 1404 : ctx: &RequestContext,
1633 1404 : ) -> anyhow::Result<()> {
1634 1404 : span::debug_assert_current_span_has_tenant_id();
1635 1404 :
1636 1404 : failpoint_support::sleep_millis_async!("before-attaching-tenant");
1637 :
1638 1404 : 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 1404 : let mut offloaded_timeline_ids = HashSet::new();
1645 1404 : let mut offloaded_timelines_list = Vec::new();
1646 1404 : 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 1368 : }
1655 : // Complete deletions for offloaded timeline id's from manifest.
1656 : // The manifest will be uploaded later in this function.
1657 1404 : offloaded_timelines_list
1658 1404 : .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 1404 : });
1668 1404 :
1669 1404 : let mut timelines_to_resume_deletions = vec![];
1670 1404 :
1671 1404 : let mut remote_index_and_client = HashMap::new();
1672 1404 : let mut timeline_ancestors = HashMap::new();
1673 1404 : let mut existent_timelines = HashSet::new();
1674 1440 : 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 1404 : 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 1404 : let sorted_timelines = tree_sort_timelines(timeline_ancestors, |m| m.ancestor_timeline())?;
1736 1440 : 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 1404 : 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 1404 : {
1812 1404 : let mut offloaded_timelines_accessor = self.timelines_offloaded.lock().unwrap();
1813 1404 : offloaded_timelines_accessor.extend(offloaded_timelines_list.into_iter());
1814 1404 : }
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 1404 : let mut guard = self.remote_tenant_manifest.lock().await;
1822 1404 : assert!(guard.is_none(), "tenant manifest set before preload"); // first populated here
1823 1404 : *guard = preload.tenant_manifest;
1824 1404 : }
1825 1404 : 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 1404 : self.clean_up_timelines(&existent_timelines)?;
1830 :
1831 1404 : self.gc_block.set_scanned(gc_blocks);
1832 1404 :
1833 1404 : fail::fail_point!("attach-before-activate", |_| {
1834 0 : anyhow::bail!("attach-before-activate");
1835 1404 : });
1836 1404 : failpoint_support::sleep_millis_async!("attach-before-activate-sleep", &self.cancel);
1837 :
1838 1404 : info!("Done");
1839 :
1840 1404 : Ok(())
1841 1404 : }
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 1404 : fn clean_up_timelines(&self, existent_timelines: &HashSet<TimelineId>) -> anyhow::Result<()> {
1847 1404 : let timelines_dir = self.conf.timelines_path(&self.tenant_shard_id);
1848 :
1849 1404 : let entries = match timelines_dir.read_dir_utf8() {
1850 1404 : 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 1452 : 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 1404 : Ok(())
1899 1404 : }
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 1404 : async fn load_timelines_metadata(
1960 1404 : self: &Arc<TenantShard>,
1961 1404 : timeline_ids: HashSet<TimelineId>,
1962 1404 : remote_storage: &GenericRemoteStorage,
1963 1404 : heatmap: Option<(HeatMapTenant, std::time::Instant)>,
1964 1404 : cancel: CancellationToken,
1965 1404 : ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
1966 1404 : let mut timeline_heatmaps = heatmap.map(|h| (h.0.into_timelines_index(), h.1));
1967 1404 :
1968 1404 : let mut part_downloads = JoinSet::new();
1969 1440 : 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 1404 : let mut timeline_preloads: HashMap<TimelineId, TimelinePreload> = HashMap::new();
1991 :
1992 : loop {
1993 1440 : tokio::select!(
1994 1440 : next = part_downloads.join_next() => {
1995 1440 : 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 1404 : break;
2002 : }
2003 : }
2004 : },
2005 1440 : _ = cancel.cancelled() => {
2006 0 : anyhow::bail!("Cancelled while waiting for remote index download")
2007 : }
2008 : )
2009 : }
2010 :
2011 1404 : Ok(timeline_preloads)
2012 1404 : }
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 1356 : pub(crate) async fn create_empty_timeline(
2441 1356 : self: &Arc<Self>,
2442 1356 : new_timeline_id: TimelineId,
2443 1356 : initdb_lsn: Lsn,
2444 1356 : pg_version: u32,
2445 1356 : ctx: &RequestContext,
2446 1356 : ) -> anyhow::Result<(UninitializedTimeline, RequestContext)> {
2447 1356 : anyhow::ensure!(
2448 1356 : self.is_active(),
2449 0 : "Cannot create empty timelines on inactive tenant"
2450 : );
2451 :
2452 : // Protect against concurrent attempts to use this TimelineId
2453 1356 : let create_guard = match self
2454 1356 : .start_creating_timeline(new_timeline_id, CreateTimelineIdempotency::FailWithConflict)
2455 1356 : .await?
2456 : {
2457 1344 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2458 : StartCreatingTimelineResult::Idempotent(_) => {
2459 0 : unreachable!("FailWithConflict implies we get an error instead")
2460 : }
2461 : };
2462 :
2463 1344 : let new_metadata = TimelineMetadata::new(
2464 1344 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2465 1344 : // make it valid, before calling finish_creation()
2466 1344 : Lsn(0),
2467 1344 : None,
2468 1344 : None,
2469 1344 : Lsn(0),
2470 1344 : initdb_lsn,
2471 1344 : initdb_lsn,
2472 1344 : pg_version,
2473 1344 : );
2474 1344 : self.prepare_new_timeline(
2475 1344 : new_timeline_id,
2476 1344 : &new_metadata,
2477 1344 : create_guard,
2478 1344 : initdb_lsn,
2479 1344 : None,
2480 1344 : None,
2481 1344 : ctx,
2482 1344 : )
2483 1344 : .await
2484 1356 : }
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 1296 : pub async fn create_test_timeline(
2493 1296 : self: &Arc<Self>,
2494 1296 : new_timeline_id: TimelineId,
2495 1296 : initdb_lsn: Lsn,
2496 1296 : pg_version: u32,
2497 1296 : ctx: &RequestContext,
2498 1296 : ) -> anyhow::Result<Arc<Timeline>> {
2499 1296 : let (uninit_tl, ctx) = self
2500 1296 : .create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2501 1296 : .await?;
2502 1296 : let tline = uninit_tl.raw_timeline().expect("we just created it");
2503 1296 : assert_eq!(tline.get_last_record_lsn(), Lsn(0));
2504 :
2505 : // Setup minimum keys required for the timeline to be usable.
2506 1296 : let mut modification = tline.begin_modification(initdb_lsn);
2507 1296 : modification
2508 1296 : .init_empty_test_timeline()
2509 1296 : .context("init_empty_test_timeline")?;
2510 1296 : modification
2511 1296 : .commit(&ctx)
2512 1296 : .await
2513 1296 : .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 1296 : tline.maybe_spawn_flush_loop();
2517 1296 : tline.freeze_and_flush().await.context("freeze_and_flush")?;
2518 :
2519 : // Make sure the freeze_and_flush reaches remote storage.
2520 1296 : tline.remote_client.wait_completion().await.unwrap();
2521 :
2522 1296 : let tl = uninit_tl.finish_creation().await?;
2523 : // The non-test code would call tl.activate() here.
2524 1296 : tl.set_state(TimelineState::Active);
2525 1296 : Ok(tl)
2526 1296 : }
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 10500 : pub fn current_state(&self) -> TenantState {
3303 10500 : self.state.borrow().clone()
3304 10500 : }
3305 :
3306 5928 : pub fn is_active(&self) -> bool {
3307 5928 : self.current_state() == TenantState::Active
3308 5928 : }
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 1416 : pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
3745 1416 : self.shard_identity.stripe_size
3746 1416 : }
3747 :
3748 0 : pub(crate) fn get_generation(&self) -> Generation {
3749 0 : self.generation
3750 0 : }
3751 :
3752 : /// This function partially shuts down the tenant (it shuts down the Timelines) and is fallible,
3753 : /// and can leave the tenant in a bad state if it fails. The caller is responsible for
3754 : /// resetting this tenant to a valid state if we fail.
3755 0 : pub(crate) async fn split_prepare(
3756 0 : &self,
3757 0 : child_shards: &Vec<TenantShardId>,
3758 0 : ) -> anyhow::Result<()> {
3759 0 : let (timelines, offloaded) = {
3760 0 : let timelines = self.timelines.lock().unwrap();
3761 0 : let offloaded = self.timelines_offloaded.lock().unwrap();
3762 0 : (timelines.clone(), offloaded.clone())
3763 0 : };
3764 0 : let timelines_iter = timelines
3765 0 : .values()
3766 0 : .map(TimelineOrOffloadedArcRef::<'_>::from)
3767 0 : .chain(
3768 0 : offloaded
3769 0 : .values()
3770 0 : .map(TimelineOrOffloadedArcRef::<'_>::from),
3771 0 : );
3772 0 : for timeline in timelines_iter {
3773 : // We do not block timeline creation/deletion during splits inside the pageserver: it is up to higher levels
3774 : // to ensure that they do not start a split if currently in the process of doing these.
3775 :
3776 0 : let timeline_id = timeline.timeline_id();
3777 :
3778 0 : if let TimelineOrOffloadedArcRef::Timeline(timeline) = timeline {
3779 : // Upload an index from the parent: this is partly to provide freshness for the
3780 : // child tenants that will copy it, and partly for general ease-of-debugging: there will
3781 : // always be a parent shard index in the same generation as we wrote the child shard index.
3782 0 : tracing::info!(%timeline_id, "Uploading index");
3783 0 : timeline
3784 0 : .remote_client
3785 0 : .schedule_index_upload_for_file_changes()?;
3786 0 : timeline.remote_client.wait_completion().await?;
3787 0 : }
3788 :
3789 0 : let remote_client = match timeline {
3790 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.remote_client.clone(),
3791 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => {
3792 0 : let remote_client = self
3793 0 : .build_timeline_client(offloaded.timeline_id, self.remote_storage.clone());
3794 0 : Arc::new(remote_client)
3795 : }
3796 : };
3797 :
3798 : // Shut down the timeline's remote client: this means that the indices we write
3799 : // for child shards will not be invalidated by the parent shard deleting layers.
3800 0 : tracing::info!(%timeline_id, "Shutting down remote storage client");
3801 0 : remote_client.shutdown().await;
3802 :
3803 : // Download methods can still be used after shutdown, as they don't flow through the remote client's
3804 : // queue. In principal the RemoteTimelineClient could provide this without downloading it, but this
3805 : // operation is rare, so it's simpler to just download it (and robustly guarantees that the index
3806 : // we use here really is the remotely persistent one).
3807 0 : tracing::info!(%timeline_id, "Downloading index_part from parent");
3808 0 : let result = remote_client
3809 0 : .download_index_file(&self.cancel)
3810 0 : .instrument(info_span!("download_index_file", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))
3811 0 : .await?;
3812 0 : let index_part = match result {
3813 : MaybeDeletedIndexPart::Deleted(_) => {
3814 0 : anyhow::bail!("Timeline deletion happened concurrently with split")
3815 : }
3816 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
3817 : };
3818 :
3819 : // A shard split may not take place while a timeline import is on-going
3820 : // for the tenant. Timeline imports run as part of each tenant shard
3821 : // and rely on the sharding scheme to split the work among pageservers.
3822 : // If we were to split in the middle of this process, we would have to
3823 : // either ensure that it's driven to completion on the old shard set
3824 : // or transfer it to the new shard set. It's technically possible, but complex.
3825 0 : match index_part.import_pgdata {
3826 0 : Some(ref import) if !import.is_done() => {
3827 0 : anyhow::bail!(
3828 0 : "Cannot split due to import with idempotency key: {:?}",
3829 0 : import.idempotency_key()
3830 0 : );
3831 : }
3832 0 : Some(_) | None => {
3833 0 : // fallthrough
3834 0 : }
3835 : }
3836 :
3837 0 : for child_shard in child_shards {
3838 0 : tracing::info!(%timeline_id, "Uploading index_part for child {}", child_shard.to_index());
3839 0 : upload_index_part(
3840 0 : &self.remote_storage,
3841 0 : child_shard,
3842 0 : &timeline_id,
3843 0 : self.generation,
3844 0 : &index_part,
3845 0 : &self.cancel,
3846 0 : )
3847 0 : .await?;
3848 : }
3849 : }
3850 :
3851 0 : let tenant_manifest = self.build_tenant_manifest();
3852 0 : for child_shard in child_shards {
3853 0 : tracing::info!(
3854 0 : "Uploading tenant manifest for child {}",
3855 0 : child_shard.to_index()
3856 : );
3857 0 : upload_tenant_manifest(
3858 0 : &self.remote_storage,
3859 0 : child_shard,
3860 0 : self.generation,
3861 0 : &tenant_manifest,
3862 0 : &self.cancel,
3863 0 : )
3864 0 : .await?;
3865 : }
3866 :
3867 0 : Ok(())
3868 0 : }
3869 :
3870 0 : pub(crate) fn get_sizes(&self) -> TopTenantShardItem {
3871 0 : let mut result = TopTenantShardItem {
3872 0 : id: self.tenant_shard_id,
3873 0 : resident_size: 0,
3874 0 : physical_size: 0,
3875 0 : max_logical_size: 0,
3876 0 : max_logical_size_per_shard: 0,
3877 0 : };
3878 :
3879 0 : for timeline in self.timelines.lock().unwrap().values() {
3880 0 : result.resident_size += timeline.metrics.resident_physical_size_gauge.get();
3881 0 :
3882 0 : result.physical_size += timeline
3883 0 : .remote_client
3884 0 : .metrics
3885 0 : .remote_physical_size_gauge
3886 0 : .get();
3887 0 : result.max_logical_size = std::cmp::max(
3888 0 : result.max_logical_size,
3889 0 : timeline.metrics.current_logical_size_gauge.get(),
3890 0 : );
3891 0 : }
3892 :
3893 0 : result.max_logical_size_per_shard = result
3894 0 : .max_logical_size
3895 0 : .div_ceil(self.tenant_shard_id.shard_count.count() as u64);
3896 0 :
3897 0 : result
3898 0 : }
3899 : }
3900 :
3901 : /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
3902 : /// perform a topological sort, so that the parent of each timeline comes
3903 : /// before the children.
3904 : /// E extracts the ancestor from T
3905 : /// This allows for T to be different. It can be TimelineMetadata, can be Timeline itself, etc.
3906 1404 : fn tree_sort_timelines<T, E>(
3907 1404 : timelines: HashMap<TimelineId, T>,
3908 1404 : extractor: E,
3909 1404 : ) -> anyhow::Result<Vec<(TimelineId, T)>>
3910 1404 : where
3911 1404 : E: Fn(&T) -> Option<TimelineId>,
3912 1404 : {
3913 1404 : let mut result = Vec::with_capacity(timelines.len());
3914 1404 :
3915 1404 : let mut now = Vec::with_capacity(timelines.len());
3916 1404 : // (ancestor, children)
3917 1404 : let mut later: HashMap<TimelineId, Vec<(TimelineId, T)>> =
3918 1404 : HashMap::with_capacity(timelines.len());
3919 :
3920 1440 : for (timeline_id, value) in timelines {
3921 36 : if let Some(ancestor_id) = extractor(&value) {
3922 12 : let children = later.entry(ancestor_id).or_default();
3923 12 : children.push((timeline_id, value));
3924 24 : } else {
3925 24 : now.push((timeline_id, value));
3926 24 : }
3927 : }
3928 :
3929 1440 : while let Some((timeline_id, metadata)) = now.pop() {
3930 36 : result.push((timeline_id, metadata));
3931 : // All children of this can be loaded now
3932 36 : if let Some(mut children) = later.remove(&timeline_id) {
3933 12 : now.append(&mut children);
3934 24 : }
3935 : }
3936 :
3937 : // All timelines should be visited now. Unless there were timelines with missing ancestors.
3938 1404 : if !later.is_empty() {
3939 0 : for (missing_id, orphan_ids) in later {
3940 0 : for (orphan_id, _) in orphan_ids {
3941 0 : error!(
3942 0 : "could not load timeline {orphan_id} because its ancestor timeline {missing_id} could not be loaded"
3943 : );
3944 : }
3945 : }
3946 0 : bail!("could not load tenant because some timelines are missing ancestors");
3947 1404 : }
3948 1404 :
3949 1404 : Ok(result)
3950 1404 : }
3951 :
3952 : enum ActivateTimelineArgs {
3953 : Yes {
3954 : broker_client: storage_broker::BrokerClientChannel,
3955 : },
3956 : No,
3957 : }
3958 :
3959 : impl TenantShard {
3960 0 : pub fn tenant_specific_overrides(&self) -> pageserver_api::models::TenantConfig {
3961 0 : self.tenant_conf.load().tenant_conf.clone()
3962 0 : }
3963 :
3964 0 : pub fn effective_config(&self) -> pageserver_api::config::TenantConfigToml {
3965 0 : self.tenant_specific_overrides()
3966 0 : .merge(self.conf.default_tenant_conf.clone())
3967 0 : }
3968 :
3969 0 : pub fn get_checkpoint_distance(&self) -> u64 {
3970 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3971 0 : tenant_conf
3972 0 : .checkpoint_distance
3973 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
3974 0 : }
3975 :
3976 0 : pub fn get_checkpoint_timeout(&self) -> Duration {
3977 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3978 0 : tenant_conf
3979 0 : .checkpoint_timeout
3980 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
3981 0 : }
3982 :
3983 0 : pub fn get_compaction_target_size(&self) -> u64 {
3984 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3985 0 : tenant_conf
3986 0 : .compaction_target_size
3987 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
3988 0 : }
3989 :
3990 0 : pub fn get_compaction_period(&self) -> Duration {
3991 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3992 0 : tenant_conf
3993 0 : .compaction_period
3994 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
3995 0 : }
3996 :
3997 0 : pub fn get_compaction_threshold(&self) -> usize {
3998 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3999 0 : tenant_conf
4000 0 : .compaction_threshold
4001 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
4002 0 : }
4003 :
4004 0 : pub fn get_rel_size_v2_enabled(&self) -> bool {
4005 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4006 0 : tenant_conf
4007 0 : .rel_size_v2_enabled
4008 0 : .unwrap_or(self.conf.default_tenant_conf.rel_size_v2_enabled)
4009 0 : }
4010 :
4011 0 : pub fn get_compaction_upper_limit(&self) -> usize {
4012 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4013 0 : tenant_conf
4014 0 : .compaction_upper_limit
4015 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_upper_limit)
4016 0 : }
4017 :
4018 0 : pub fn get_compaction_l0_first(&self) -> bool {
4019 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4020 0 : tenant_conf
4021 0 : .compaction_l0_first
4022 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_first)
4023 0 : }
4024 :
4025 24 : pub fn get_gc_horizon(&self) -> u64 {
4026 24 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4027 24 : tenant_conf
4028 24 : .gc_horizon
4029 24 : .unwrap_or(self.conf.default_tenant_conf.gc_horizon)
4030 24 : }
4031 :
4032 0 : pub fn get_gc_period(&self) -> Duration {
4033 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4034 0 : tenant_conf
4035 0 : .gc_period
4036 0 : .unwrap_or(self.conf.default_tenant_conf.gc_period)
4037 0 : }
4038 :
4039 0 : pub fn get_image_creation_threshold(&self) -> usize {
4040 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4041 0 : tenant_conf
4042 0 : .image_creation_threshold
4043 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
4044 0 : }
4045 :
4046 24 : pub fn get_pitr_interval(&self) -> Duration {
4047 24 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4048 24 : tenant_conf
4049 24 : .pitr_interval
4050 24 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
4051 24 : }
4052 :
4053 0 : pub fn get_min_resident_size_override(&self) -> Option<u64> {
4054 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4055 0 : tenant_conf
4056 0 : .min_resident_size_override
4057 0 : .or(self.conf.default_tenant_conf.min_resident_size_override)
4058 0 : }
4059 :
4060 0 : pub fn get_heatmap_period(&self) -> Option<Duration> {
4061 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4062 0 : let heatmap_period = tenant_conf
4063 0 : .heatmap_period
4064 0 : .unwrap_or(self.conf.default_tenant_conf.heatmap_period);
4065 0 : if heatmap_period.is_zero() {
4066 0 : None
4067 : } else {
4068 0 : Some(heatmap_period)
4069 : }
4070 0 : }
4071 :
4072 24 : pub fn get_lsn_lease_length(&self) -> Duration {
4073 24 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4074 24 : tenant_conf
4075 24 : .lsn_lease_length
4076 24 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
4077 24 : }
4078 :
4079 0 : pub fn get_timeline_offloading_enabled(&self) -> bool {
4080 0 : if self.conf.timeline_offloading {
4081 0 : return true;
4082 0 : }
4083 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4084 0 : tenant_conf
4085 0 : .timeline_offloading
4086 0 : .unwrap_or(self.conf.default_tenant_conf.timeline_offloading)
4087 0 : }
4088 :
4089 : /// Generate an up-to-date TenantManifest based on the state of this Tenant.
4090 1416 : fn build_tenant_manifest(&self) -> TenantManifest {
4091 1416 : // Collect the offloaded timelines, and sort them for deterministic output.
4092 1416 : let offloaded_timelines = self
4093 1416 : .timelines_offloaded
4094 1416 : .lock()
4095 1416 : .unwrap()
4096 1416 : .values()
4097 1416 : .map(|tli| tli.manifest())
4098 1416 : .sorted_by_key(|m| m.timeline_id)
4099 1416 : .collect_vec();
4100 1416 :
4101 1416 : TenantManifest {
4102 1416 : version: LATEST_TENANT_MANIFEST_VERSION,
4103 1416 : stripe_size: Some(self.get_shard_stripe_size()),
4104 1416 : offloaded_timelines,
4105 1416 : }
4106 1416 : }
4107 :
4108 0 : pub fn update_tenant_config<
4109 0 : F: Fn(
4110 0 : pageserver_api::models::TenantConfig,
4111 0 : ) -> anyhow::Result<pageserver_api::models::TenantConfig>,
4112 0 : >(
4113 0 : &self,
4114 0 : update: F,
4115 0 : ) -> anyhow::Result<pageserver_api::models::TenantConfig> {
4116 0 : // Use read-copy-update in order to avoid overwriting the location config
4117 0 : // state if this races with [`TenantShard::set_new_location_config`]. Note that
4118 0 : // this race is not possible if both request types come from the storage
4119 0 : // controller (as they should!) because an exclusive op lock is required
4120 0 : // on the storage controller side.
4121 0 :
4122 0 : self.tenant_conf
4123 0 : .try_rcu(|attached_conf| -> Result<_, anyhow::Error> {
4124 0 : Ok(Arc::new(AttachedTenantConf {
4125 0 : tenant_conf: update(attached_conf.tenant_conf.clone())?,
4126 0 : location: attached_conf.location,
4127 0 : lsn_lease_deadline: attached_conf.lsn_lease_deadline,
4128 : }))
4129 0 : })?;
4130 :
4131 0 : let updated = self.tenant_conf.load();
4132 0 :
4133 0 : self.tenant_conf_updated(&updated.tenant_conf);
4134 0 : // Don't hold self.timelines.lock() during the notifies.
4135 0 : // There's no risk of deadlock right now, but there could be if we consolidate
4136 0 : // mutexes in struct Timeline in the future.
4137 0 : let timelines = self.list_timelines();
4138 0 : for timeline in timelines {
4139 0 : timeline.tenant_conf_updated(&updated);
4140 0 : }
4141 :
4142 0 : Ok(updated.tenant_conf.clone())
4143 0 : }
4144 :
4145 0 : pub(crate) fn set_new_location_config(&self, new_conf: AttachedTenantConf) {
4146 0 : let new_tenant_conf = new_conf.tenant_conf.clone();
4147 0 :
4148 0 : self.tenant_conf.store(Arc::new(new_conf.clone()));
4149 0 :
4150 0 : self.tenant_conf_updated(&new_tenant_conf);
4151 0 : // Don't hold self.timelines.lock() during the notifies.
4152 0 : // There's no risk of deadlock right now, but there could be if we consolidate
4153 0 : // mutexes in struct Timeline in the future.
4154 0 : let timelines = self.list_timelines();
4155 0 : for timeline in timelines {
4156 0 : timeline.tenant_conf_updated(&new_conf);
4157 0 : }
4158 0 : }
4159 :
4160 1404 : fn get_pagestream_throttle_config(
4161 1404 : psconf: &'static PageServerConf,
4162 1404 : overrides: &pageserver_api::models::TenantConfig,
4163 1404 : ) -> throttle::Config {
4164 1404 : overrides
4165 1404 : .timeline_get_throttle
4166 1404 : .clone()
4167 1404 : .unwrap_or(psconf.default_tenant_conf.timeline_get_throttle.clone())
4168 1404 : }
4169 :
4170 0 : pub(crate) fn tenant_conf_updated(&self, new_conf: &pageserver_api::models::TenantConfig) {
4171 0 : let conf = Self::get_pagestream_throttle_config(self.conf, new_conf);
4172 0 : self.pagestream_throttle.reconfigure(conf)
4173 0 : }
4174 :
4175 : /// Helper function to create a new Timeline struct.
4176 : ///
4177 : /// The returned Timeline is in Loading state. The caller is responsible for
4178 : /// initializing any on-disk state, and for inserting the Timeline to the 'timelines'
4179 : /// map.
4180 : ///
4181 : /// `validate_ancestor == false` is used when a timeline is created for deletion
4182 : /// and we might not have the ancestor present anymore which is fine for to be
4183 : /// deleted timelines.
4184 : #[allow(clippy::too_many_arguments)]
4185 2796 : fn create_timeline_struct(
4186 2796 : &self,
4187 2796 : new_timeline_id: TimelineId,
4188 2796 : new_metadata: &TimelineMetadata,
4189 2796 : previous_heatmap: Option<PreviousHeatmap>,
4190 2796 : ancestor: Option<Arc<Timeline>>,
4191 2796 : resources: TimelineResources,
4192 2796 : cause: CreateTimelineCause,
4193 2796 : create_idempotency: CreateTimelineIdempotency,
4194 2796 : gc_compaction_state: Option<GcCompactionState>,
4195 2796 : rel_size_v2_status: Option<RelSizeMigration>,
4196 2796 : ctx: &RequestContext,
4197 2796 : ) -> anyhow::Result<(Arc<Timeline>, RequestContext)> {
4198 2796 : let state = match cause {
4199 : CreateTimelineCause::Load => {
4200 2796 : let ancestor_id = new_metadata.ancestor_timeline();
4201 2796 : anyhow::ensure!(
4202 2796 : ancestor_id == ancestor.as_ref().map(|t| t.timeline_id),
4203 0 : "Timeline's {new_timeline_id} ancestor {ancestor_id:?} was not found"
4204 : );
4205 2796 : TimelineState::Loading
4206 : }
4207 0 : CreateTimelineCause::Delete => TimelineState::Stopping,
4208 : };
4209 :
4210 2796 : let pg_version = new_metadata.pg_version();
4211 2796 :
4212 2796 : let timeline = Timeline::new(
4213 2796 : self.conf,
4214 2796 : Arc::clone(&self.tenant_conf),
4215 2796 : new_metadata,
4216 2796 : previous_heatmap,
4217 2796 : ancestor,
4218 2796 : new_timeline_id,
4219 2796 : self.tenant_shard_id,
4220 2796 : self.generation,
4221 2796 : self.shard_identity,
4222 2796 : self.walredo_mgr.clone(),
4223 2796 : resources,
4224 2796 : pg_version,
4225 2796 : state,
4226 2796 : self.attach_wal_lag_cooldown.clone(),
4227 2796 : create_idempotency,
4228 2796 : gc_compaction_state,
4229 2796 : rel_size_v2_status,
4230 2796 : self.cancel.child_token(),
4231 2796 : );
4232 2796 :
4233 2796 : let timeline_ctx = RequestContextBuilder::from(ctx)
4234 2796 : .scope(context::Scope::new_timeline(&timeline))
4235 2796 : .detached_child();
4236 2796 :
4237 2796 : Ok((timeline, timeline_ctx))
4238 2796 : }
4239 :
4240 : /// [`TenantShard::shutdown`] must be called before dropping the returned [`TenantShard`] object
4241 : /// to ensure proper cleanup of background tasks and metrics.
4242 : //
4243 : // Allow too_many_arguments because a constructor's argument list naturally grows with the
4244 : // number of attributes in the struct: breaking these out into a builder wouldn't be helpful.
4245 : #[allow(clippy::too_many_arguments)]
4246 1404 : fn new(
4247 1404 : state: TenantState,
4248 1404 : conf: &'static PageServerConf,
4249 1404 : attached_conf: AttachedTenantConf,
4250 1404 : shard_identity: ShardIdentity,
4251 1404 : walredo_mgr: Option<Arc<WalRedoManager>>,
4252 1404 : tenant_shard_id: TenantShardId,
4253 1404 : remote_storage: GenericRemoteStorage,
4254 1404 : deletion_queue_client: DeletionQueueClient,
4255 1404 : l0_flush_global_state: L0FlushGlobalState,
4256 1404 : ) -> TenantShard {
4257 1404 : assert!(!attached_conf.location.generation.is_none());
4258 :
4259 1404 : let (state, mut rx) = watch::channel(state);
4260 1404 :
4261 1404 : tokio::spawn(async move {
4262 1402 : // reflect tenant state in metrics:
4263 1402 : // - global per tenant state: TENANT_STATE_METRIC
4264 1402 : // - "set" of broken tenants: BROKEN_TENANTS_SET
4265 1402 : //
4266 1402 : // set of broken tenants should not have zero counts so that it remains accessible for
4267 1402 : // alerting.
4268 1402 :
4269 1402 : let tid = tenant_shard_id.to_string();
4270 1402 : let shard_id = tenant_shard_id.shard_slug().to_string();
4271 1402 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
4272 :
4273 2798 : fn inspect_state(state: &TenantState) -> ([&'static str; 1], bool) {
4274 2798 : ([state.into()], matches!(state, TenantState::Broken { .. }))
4275 2798 : }
4276 :
4277 1402 : let mut tuple = inspect_state(&rx.borrow_and_update());
4278 1402 :
4279 1402 : let is_broken = tuple.1;
4280 1402 : let mut counted_broken = if is_broken {
4281 : // add the id to the set right away, there should not be any updates on the channel
4282 : // after before tenant is removed, if ever
4283 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
4284 0 : true
4285 : } else {
4286 1402 : false
4287 : };
4288 :
4289 : loop {
4290 2798 : let labels = &tuple.0;
4291 2798 : let current = TENANT_STATE_METRIC.with_label_values(labels);
4292 2798 : current.inc();
4293 2798 :
4294 2798 : if rx.changed().await.is_err() {
4295 : // tenant has been dropped
4296 84 : current.dec();
4297 84 : drop(BROKEN_TENANTS_SET.remove_label_values(set_key));
4298 84 : break;
4299 1396 : }
4300 1396 :
4301 1396 : current.dec();
4302 1396 : tuple = inspect_state(&rx.borrow_and_update());
4303 1396 :
4304 1396 : let is_broken = tuple.1;
4305 1396 : if is_broken && !counted_broken {
4306 0 : counted_broken = true;
4307 0 : // insert the tenant_id (back) into the set while avoiding needless counter
4308 0 : // access
4309 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
4310 1396 : }
4311 : }
4312 1404 : });
4313 1404 :
4314 1404 : TenantShard {
4315 1404 : tenant_shard_id,
4316 1404 : shard_identity,
4317 1404 : generation: attached_conf.location.generation,
4318 1404 : conf,
4319 1404 : // using now here is good enough approximation to catch tenants with really long
4320 1404 : // activation times.
4321 1404 : constructed_at: Instant::now(),
4322 1404 : timelines: Mutex::new(HashMap::new()),
4323 1404 : timelines_creating: Mutex::new(HashSet::new()),
4324 1404 : timelines_offloaded: Mutex::new(HashMap::new()),
4325 1404 : remote_tenant_manifest: Default::default(),
4326 1404 : gc_cs: tokio::sync::Mutex::new(()),
4327 1404 : walredo_mgr,
4328 1404 : remote_storage,
4329 1404 : deletion_queue_client,
4330 1404 : state,
4331 1404 : cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
4332 1404 : cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
4333 1404 : eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
4334 1404 : compaction_circuit_breaker: std::sync::Mutex::new(CircuitBreaker::new(
4335 1404 : format!("compaction-{tenant_shard_id}"),
4336 1404 : 5,
4337 1404 : // Compaction can be a very expensive operation, and might leak disk space. It also ought
4338 1404 : // to be infallible, as long as remote storage is available. So if it repeatedly fails,
4339 1404 : // use an extremely long backoff.
4340 1404 : Some(Duration::from_secs(3600 * 24)),
4341 1404 : )),
4342 1404 : l0_compaction_trigger: Arc::new(Notify::new()),
4343 1404 : scheduled_compaction_tasks: Mutex::new(Default::default()),
4344 1404 : activate_now_sem: tokio::sync::Semaphore::new(0),
4345 1404 : attach_wal_lag_cooldown: Arc::new(std::sync::OnceLock::new()),
4346 1404 : cancel: CancellationToken::default(),
4347 1404 : gate: Gate::default(),
4348 1404 : pagestream_throttle: Arc::new(throttle::Throttle::new(
4349 1404 : TenantShard::get_pagestream_throttle_config(conf, &attached_conf.tenant_conf),
4350 1404 : )),
4351 1404 : pagestream_throttle_metrics: Arc::new(
4352 1404 : crate::metrics::tenant_throttling::Pagestream::new(&tenant_shard_id),
4353 1404 : ),
4354 1404 : tenant_conf: Arc::new(ArcSwap::from_pointee(attached_conf)),
4355 1404 : ongoing_timeline_detach: std::sync::Mutex::default(),
4356 1404 : gc_block: Default::default(),
4357 1404 : l0_flush_global_state,
4358 1404 : }
4359 1404 : }
4360 :
4361 : /// Locate and load config
4362 0 : pub(super) fn load_tenant_config(
4363 0 : conf: &'static PageServerConf,
4364 0 : tenant_shard_id: &TenantShardId,
4365 0 : ) -> Result<LocationConf, LoadConfigError> {
4366 0 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4367 0 :
4368 0 : info!("loading tenant configuration from {config_path}");
4369 :
4370 : // load and parse file
4371 0 : let config = fs::read_to_string(&config_path).map_err(|e| {
4372 0 : match e.kind() {
4373 : std::io::ErrorKind::NotFound => {
4374 : // The config should almost always exist for a tenant directory:
4375 : // - When attaching a tenant, the config is the first thing we write
4376 : // - When detaching a tenant, we atomically move the directory to a tmp location
4377 : // before deleting contents.
4378 : //
4379 : // The very rare edge case that can result in a missing config is if we crash during attach
4380 : // between creating directory and writing config. Callers should handle that as if the
4381 : // directory didn't exist.
4382 :
4383 0 : LoadConfigError::NotFound(config_path)
4384 : }
4385 : _ => {
4386 : // No IO errors except NotFound are acceptable here: other kinds of error indicate local storage or permissions issues
4387 : // that we cannot cleanly recover
4388 0 : crate::virtual_file::on_fatal_io_error(&e, "Reading tenant config file")
4389 : }
4390 : }
4391 0 : })?;
4392 :
4393 0 : Ok(toml_edit::de::from_str::<LocationConf>(&config)?)
4394 0 : }
4395 :
4396 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4397 : pub(super) async fn persist_tenant_config(
4398 : conf: &'static PageServerConf,
4399 : tenant_shard_id: &TenantShardId,
4400 : location_conf: &LocationConf,
4401 : ) -> std::io::Result<()> {
4402 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4403 :
4404 : Self::persist_tenant_config_at(tenant_shard_id, &config_path, location_conf).await
4405 : }
4406 :
4407 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4408 : pub(super) async fn persist_tenant_config_at(
4409 : tenant_shard_id: &TenantShardId,
4410 : config_path: &Utf8Path,
4411 : location_conf: &LocationConf,
4412 : ) -> std::io::Result<()> {
4413 : debug!("persisting tenantconf to {config_path}");
4414 :
4415 : let mut conf_content = r#"# This file contains a specific per-tenant's config.
4416 : # It is read in case of pageserver restart.
4417 : "#
4418 : .to_string();
4419 :
4420 0 : fail::fail_point!("tenant-config-before-write", |_| {
4421 0 : Err(std::io::Error::other("tenant-config-before-write"))
4422 0 : });
4423 :
4424 : // Convert the config to a toml file.
4425 : conf_content +=
4426 : &toml_edit::ser::to_string_pretty(&location_conf).expect("Config serialization failed");
4427 :
4428 : let temp_path = path_with_suffix_extension(config_path, TEMP_FILE_SUFFIX);
4429 :
4430 : let conf_content = conf_content.into_bytes();
4431 : VirtualFile::crashsafe_overwrite(config_path.to_owned(), temp_path, conf_content).await
4432 : }
4433 :
4434 : //
4435 : // How garbage collection works:
4436 : //
4437 : // +--bar------------->
4438 : // /
4439 : // +----+-----foo---------------->
4440 : // /
4441 : // ----main--+-------------------------->
4442 : // \
4443 : // +-----baz-------->
4444 : //
4445 : //
4446 : // 1. Grab 'gc_cs' mutex to prevent new timelines from being created while Timeline's
4447 : // `gc_infos` are being refreshed
4448 : // 2. Scan collected timelines, and on each timeline, make note of the
4449 : // all the points where other timelines have been branched off.
4450 : // We will refrain from removing page versions at those LSNs.
4451 : // 3. For each timeline, scan all layer files on the timeline.
4452 : // Remove all files for which a newer file exists and which
4453 : // don't cover any branch point LSNs.
4454 : //
4455 : // TODO:
4456 : // - if a relation has a non-incremental persistent layer on a child branch, then we
4457 : // don't need to keep that in the parent anymore. But currently
4458 : // we do.
4459 24 : async fn gc_iteration_internal(
4460 24 : &self,
4461 24 : target_timeline_id: Option<TimelineId>,
4462 24 : horizon: u64,
4463 24 : pitr: Duration,
4464 24 : cancel: &CancellationToken,
4465 24 : ctx: &RequestContext,
4466 24 : ) -> Result<GcResult, GcError> {
4467 24 : let mut totals: GcResult = Default::default();
4468 24 : let now = Instant::now();
4469 :
4470 24 : let gc_timelines = self
4471 24 : .refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4472 24 : .await?;
4473 :
4474 24 : failpoint_support::sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
4475 :
4476 : // If there is nothing to GC, we don't want any messages in the INFO log.
4477 24 : if !gc_timelines.is_empty() {
4478 24 : info!("{} timelines need GC", gc_timelines.len());
4479 : } else {
4480 0 : debug!("{} timelines need GC", gc_timelines.len());
4481 : }
4482 :
4483 : // Perform GC for each timeline.
4484 : //
4485 : // Note that we don't hold the `TenantShard::gc_cs` lock here because we don't want to delay the
4486 : // branch creation task, which requires the GC lock. A GC iteration can run concurrently
4487 : // with branch creation.
4488 : //
4489 : // See comments in [`TenantShard::branch_timeline`] for more information about why branch
4490 : // creation task can run concurrently with timeline's GC iteration.
4491 48 : for timeline in gc_timelines {
4492 24 : if cancel.is_cancelled() {
4493 : // We were requested to shut down. Stop and return with the progress we
4494 : // made.
4495 0 : break;
4496 24 : }
4497 24 : let result = match timeline.gc().await {
4498 : Err(GcError::TimelineCancelled) => {
4499 0 : if target_timeline_id.is_some() {
4500 : // If we were targetting this specific timeline, surface cancellation to caller
4501 0 : return Err(GcError::TimelineCancelled);
4502 : } else {
4503 : // A timeline may be shutting down independently of the tenant's lifecycle: we should
4504 : // skip past this and proceed to try GC on other timelines.
4505 0 : continue;
4506 : }
4507 : }
4508 24 : r => r?,
4509 : };
4510 24 : totals += result;
4511 : }
4512 :
4513 24 : totals.elapsed = now.elapsed();
4514 24 : Ok(totals)
4515 24 : }
4516 :
4517 : /// Refreshes the Timeline::gc_info for all timelines, returning the
4518 : /// vector of timelines which have [`Timeline::get_last_record_lsn`] past
4519 : /// [`TenantShard::get_gc_horizon`].
4520 : ///
4521 : /// This is usually executed as part of periodic gc, but can now be triggered more often.
4522 24 : pub(crate) async fn refresh_gc_info(
4523 24 : &self,
4524 24 : cancel: &CancellationToken,
4525 24 : ctx: &RequestContext,
4526 24 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4527 24 : // since this method can now be called at different rates than the configured gc loop, it
4528 24 : // might be that these configuration values get applied faster than what it was previously,
4529 24 : // since these were only read from the gc task.
4530 24 : let horizon = self.get_gc_horizon();
4531 24 : let pitr = self.get_pitr_interval();
4532 24 :
4533 24 : // refresh all timelines
4534 24 : let target_timeline_id = None;
4535 24 :
4536 24 : self.refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4537 24 : .await
4538 24 : }
4539 :
4540 : /// Populate all Timelines' `GcInfo` with information about their children. We do not set the
4541 : /// PITR cutoffs here, because that requires I/O: this is done later, before GC, by [`Self::refresh_gc_info_internal`]
4542 : ///
4543 : /// Subsequently, parent-child relationships are updated incrementally inside [`Timeline::new`] and [`Timeline::drop`].
4544 0 : fn initialize_gc_info(
4545 0 : &self,
4546 0 : timelines: &std::sync::MutexGuard<HashMap<TimelineId, Arc<Timeline>>>,
4547 0 : timelines_offloaded: &std::sync::MutexGuard<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
4548 0 : restrict_to_timeline: Option<TimelineId>,
4549 0 : ) {
4550 0 : if restrict_to_timeline.is_none() {
4551 : // This function must be called before activation: after activation timeline create/delete operations
4552 : // might happen, and this function is not safe to run concurrently with those.
4553 0 : assert!(!self.is_active());
4554 0 : }
4555 :
4556 : // Scan all timelines. For each timeline, remember the timeline ID and
4557 : // the branch point where it was created.
4558 0 : let mut all_branchpoints: BTreeMap<TimelineId, Vec<(Lsn, TimelineId, MaybeOffloaded)>> =
4559 0 : BTreeMap::new();
4560 0 : timelines.iter().for_each(|(timeline_id, timeline_entry)| {
4561 0 : if let Some(ancestor_timeline_id) = &timeline_entry.get_ancestor_timeline_id() {
4562 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4563 0 : ancestor_children.push((
4564 0 : timeline_entry.get_ancestor_lsn(),
4565 0 : *timeline_id,
4566 0 : MaybeOffloaded::No,
4567 0 : ));
4568 0 : }
4569 0 : });
4570 0 : timelines_offloaded
4571 0 : .iter()
4572 0 : .for_each(|(timeline_id, timeline_entry)| {
4573 0 : let Some(ancestor_timeline_id) = &timeline_entry.ancestor_timeline_id else {
4574 0 : return;
4575 : };
4576 0 : let Some(retain_lsn) = timeline_entry.ancestor_retain_lsn else {
4577 0 : return;
4578 : };
4579 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4580 0 : ancestor_children.push((retain_lsn, *timeline_id, MaybeOffloaded::Yes));
4581 0 : });
4582 0 :
4583 0 : // The number of bytes we always keep, irrespective of PITR: this is a constant across timelines
4584 0 : let horizon = self.get_gc_horizon();
4585 :
4586 : // Populate each timeline's GcInfo with information about its child branches
4587 0 : let timelines_to_write = if let Some(timeline_id) = restrict_to_timeline {
4588 0 : itertools::Either::Left(timelines.get(&timeline_id).into_iter())
4589 : } else {
4590 0 : itertools::Either::Right(timelines.values())
4591 : };
4592 0 : for timeline in timelines_to_write {
4593 0 : let mut branchpoints: Vec<(Lsn, TimelineId, MaybeOffloaded)> = all_branchpoints
4594 0 : .remove(&timeline.timeline_id)
4595 0 : .unwrap_or_default();
4596 0 :
4597 0 : branchpoints.sort_by_key(|b| b.0);
4598 0 :
4599 0 : let mut target = timeline.gc_info.write().unwrap();
4600 0 :
4601 0 : target.retain_lsns = branchpoints;
4602 0 :
4603 0 : let space_cutoff = timeline
4604 0 : .get_last_record_lsn()
4605 0 : .checked_sub(horizon)
4606 0 : .unwrap_or(Lsn(0));
4607 0 :
4608 0 : target.cutoffs = GcCutoffs {
4609 0 : space: space_cutoff,
4610 0 : time: Lsn::INVALID,
4611 0 : };
4612 0 : }
4613 0 : }
4614 :
4615 48 : async fn refresh_gc_info_internal(
4616 48 : &self,
4617 48 : target_timeline_id: Option<TimelineId>,
4618 48 : horizon: u64,
4619 48 : pitr: Duration,
4620 48 : cancel: &CancellationToken,
4621 48 : ctx: &RequestContext,
4622 48 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4623 48 : // before taking the gc_cs lock, do the heavier weight finding of gc_cutoff points for
4624 48 : // currently visible timelines.
4625 48 : let timelines = self
4626 48 : .timelines
4627 48 : .lock()
4628 48 : .unwrap()
4629 48 : .values()
4630 120 : .filter(|tl| match target_timeline_id.as_ref() {
4631 24 : Some(target) => &tl.timeline_id == target,
4632 96 : None => true,
4633 120 : })
4634 48 : .cloned()
4635 48 : .collect::<Vec<_>>();
4636 48 :
4637 48 : if target_timeline_id.is_some() && timelines.is_empty() {
4638 : // We were to act on a particular timeline and it wasn't found
4639 0 : return Err(GcError::TimelineNotFound);
4640 48 : }
4641 48 :
4642 48 : let mut gc_cutoffs: HashMap<TimelineId, GcCutoffs> =
4643 48 : HashMap::with_capacity(timelines.len());
4644 48 :
4645 48 : // Ensures all timelines use the same start time when computing the time cutoff.
4646 48 : let now_ts_for_pitr_calc = SystemTime::now();
4647 120 : for timeline in timelines.iter() {
4648 120 : let ctx = &ctx.with_scope_timeline(timeline);
4649 120 : let cutoff = timeline
4650 120 : .get_last_record_lsn()
4651 120 : .checked_sub(horizon)
4652 120 : .unwrap_or(Lsn(0));
4653 :
4654 120 : let cutoffs = timeline
4655 120 : .find_gc_cutoffs(now_ts_for_pitr_calc, cutoff, pitr, cancel, ctx)
4656 120 : .await?;
4657 120 : let old = gc_cutoffs.insert(timeline.timeline_id, cutoffs);
4658 120 : assert!(old.is_none());
4659 : }
4660 :
4661 48 : if !self.is_active() || self.cancel.is_cancelled() {
4662 0 : return Err(GcError::TenantCancelled);
4663 48 : }
4664 :
4665 : // grab mutex to prevent new timelines from being created here; avoid doing long operations
4666 : // because that will stall branch creation.
4667 48 : let gc_cs = self.gc_cs.lock().await;
4668 :
4669 : // Ok, we now know all the branch points.
4670 : // Update the GC information for each timeline.
4671 48 : let mut gc_timelines = Vec::with_capacity(timelines.len());
4672 168 : for timeline in timelines {
4673 : // We filtered the timeline list above
4674 120 : if let Some(target_timeline_id) = target_timeline_id {
4675 24 : assert_eq!(target_timeline_id, timeline.timeline_id);
4676 96 : }
4677 :
4678 : {
4679 120 : let mut target = timeline.gc_info.write().unwrap();
4680 120 :
4681 120 : // Cull any expired leases
4682 120 : let now = SystemTime::now();
4683 120 : target.leases.retain(|_, lease| !lease.is_expired(&now));
4684 120 :
4685 120 : timeline
4686 120 : .metrics
4687 120 : .valid_lsn_lease_count_gauge
4688 120 : .set(target.leases.len() as u64);
4689 :
4690 : // Look up parent's PITR cutoff to update the child's knowledge of whether it is within parent's PITR
4691 120 : if let Some(ancestor_id) = timeline.get_ancestor_timeline_id() {
4692 72 : if let Some(ancestor_gc_cutoffs) = gc_cutoffs.get(&ancestor_id) {
4693 72 : target.within_ancestor_pitr =
4694 72 : timeline.get_ancestor_lsn() >= ancestor_gc_cutoffs.time;
4695 72 : }
4696 48 : }
4697 :
4698 : // Update metrics that depend on GC state
4699 120 : timeline
4700 120 : .metrics
4701 120 : .archival_size
4702 120 : .set(if target.within_ancestor_pitr {
4703 0 : timeline.metrics.current_logical_size_gauge.get()
4704 : } else {
4705 120 : 0
4706 : });
4707 120 : timeline.metrics.pitr_history_size.set(
4708 120 : timeline
4709 120 : .get_last_record_lsn()
4710 120 : .checked_sub(target.cutoffs.time)
4711 120 : .unwrap_or(Lsn(0))
4712 120 : .0,
4713 120 : );
4714 :
4715 : // Apply the cutoffs we found to the Timeline's GcInfo. Why might we _not_ have cutoffs for a timeline?
4716 : // - this timeline was created while we were finding cutoffs
4717 : // - lsn for timestamp search fails for this timeline repeatedly
4718 120 : if let Some(cutoffs) = gc_cutoffs.get(&timeline.timeline_id) {
4719 120 : let original_cutoffs = target.cutoffs.clone();
4720 120 : // GC cutoffs should never go back
4721 120 : target.cutoffs = GcCutoffs {
4722 120 : space: Lsn(cutoffs.space.0.max(original_cutoffs.space.0)),
4723 120 : time: Lsn(cutoffs.time.0.max(original_cutoffs.time.0)),
4724 120 : }
4725 0 : }
4726 : }
4727 :
4728 120 : gc_timelines.push(timeline);
4729 : }
4730 48 : drop(gc_cs);
4731 48 : Ok(gc_timelines)
4732 48 : }
4733 :
4734 : /// A substitute for `branch_timeline` for use in unit tests.
4735 : /// The returned timeline will have state value `Active` to make various `anyhow::ensure!()`
4736 : /// calls pass, but, we do not actually call `.activate()` under the hood. So, none of the
4737 : /// timeline background tasks are launched, except the flush loop.
4738 : #[cfg(test)]
4739 1428 : async fn branch_timeline_test(
4740 1428 : self: &Arc<Self>,
4741 1428 : src_timeline: &Arc<Timeline>,
4742 1428 : dst_id: TimelineId,
4743 1428 : ancestor_lsn: Option<Lsn>,
4744 1428 : ctx: &RequestContext,
4745 1428 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
4746 1428 : let tl = self
4747 1428 : .branch_timeline_impl(src_timeline, dst_id, ancestor_lsn, ctx)
4748 1428 : .await?
4749 1404 : .into_timeline_for_test();
4750 1404 : tl.set_state(TimelineState::Active);
4751 1404 : Ok(tl)
4752 1428 : }
4753 :
4754 : /// Helper for unit tests to branch a timeline with some pre-loaded states.
4755 : #[cfg(test)]
4756 : #[allow(clippy::too_many_arguments)]
4757 72 : pub async fn branch_timeline_test_with_layers(
4758 72 : self: &Arc<Self>,
4759 72 : src_timeline: &Arc<Timeline>,
4760 72 : dst_id: TimelineId,
4761 72 : ancestor_lsn: Option<Lsn>,
4762 72 : ctx: &RequestContext,
4763 72 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
4764 72 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
4765 72 : end_lsn: Lsn,
4766 72 : ) -> anyhow::Result<Arc<Timeline>> {
4767 : use checks::check_valid_layermap;
4768 : use itertools::Itertools;
4769 :
4770 72 : let tline = self
4771 72 : .branch_timeline_test(src_timeline, dst_id, ancestor_lsn, ctx)
4772 72 : .await?;
4773 72 : let ancestor_lsn = if let Some(ancestor_lsn) = ancestor_lsn {
4774 72 : ancestor_lsn
4775 : } else {
4776 0 : tline.get_last_record_lsn()
4777 : };
4778 72 : assert!(end_lsn >= ancestor_lsn);
4779 72 : tline.force_advance_lsn(end_lsn);
4780 108 : for deltas in delta_layer_desc {
4781 36 : tline
4782 36 : .force_create_delta_layer(deltas, Some(ancestor_lsn), ctx)
4783 36 : .await?;
4784 : }
4785 96 : for (lsn, images) in image_layer_desc {
4786 24 : tline
4787 24 : .force_create_image_layer(lsn, images, Some(ancestor_lsn), ctx)
4788 24 : .await?;
4789 : }
4790 72 : let layer_names = tline
4791 72 : .layers
4792 72 : .read()
4793 72 : .await
4794 72 : .layer_map()
4795 72 : .unwrap()
4796 72 : .iter_historic_layers()
4797 72 : .map(|layer| layer.layer_name())
4798 72 : .collect_vec();
4799 72 : if let Some(err) = check_valid_layermap(&layer_names) {
4800 0 : bail!("invalid layermap: {err}");
4801 72 : }
4802 72 : Ok(tline)
4803 72 : }
4804 :
4805 : /// Branch an existing timeline.
4806 0 : async fn branch_timeline(
4807 0 : self: &Arc<Self>,
4808 0 : src_timeline: &Arc<Timeline>,
4809 0 : dst_id: TimelineId,
4810 0 : start_lsn: Option<Lsn>,
4811 0 : ctx: &RequestContext,
4812 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4813 0 : self.branch_timeline_impl(src_timeline, dst_id, start_lsn, ctx)
4814 0 : .await
4815 0 : }
4816 :
4817 1428 : async fn branch_timeline_impl(
4818 1428 : self: &Arc<Self>,
4819 1428 : src_timeline: &Arc<Timeline>,
4820 1428 : dst_id: TimelineId,
4821 1428 : start_lsn: Option<Lsn>,
4822 1428 : ctx: &RequestContext,
4823 1428 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4824 1428 : let src_id = src_timeline.timeline_id;
4825 :
4826 : // We will validate our ancestor LSN in this function. Acquire the GC lock so that
4827 : // this check cannot race with GC, and the ancestor LSN is guaranteed to remain
4828 : // valid while we are creating the branch.
4829 1428 : let _gc_cs = self.gc_cs.lock().await;
4830 :
4831 : // If no start LSN is specified, we branch the new timeline from the source timeline's last record LSN
4832 1428 : let start_lsn = start_lsn.unwrap_or_else(|| {
4833 12 : let lsn = src_timeline.get_last_record_lsn();
4834 12 : info!("branching timeline {dst_id} from timeline {src_id} at last record LSN: {lsn}");
4835 12 : lsn
4836 1428 : });
4837 :
4838 : // we finally have determined the ancestor_start_lsn, so we can get claim exclusivity now
4839 1428 : let timeline_create_guard = match self
4840 1428 : .start_creating_timeline(
4841 1428 : dst_id,
4842 1428 : CreateTimelineIdempotency::Branch {
4843 1428 : ancestor_timeline_id: src_timeline.timeline_id,
4844 1428 : ancestor_start_lsn: start_lsn,
4845 1428 : },
4846 1428 : )
4847 1428 : .await?
4848 : {
4849 1428 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
4850 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
4851 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
4852 : }
4853 : };
4854 :
4855 : // Ensure that `start_lsn` is valid, i.e. the LSN is within the PITR
4856 : // horizon on the source timeline
4857 : //
4858 : // We check it against both the planned GC cutoff stored in 'gc_info',
4859 : // and the 'latest_gc_cutoff' of the last GC that was performed. The
4860 : // planned GC cutoff in 'gc_info' is normally larger than
4861 : // 'applied_gc_cutoff_lsn', but beware of corner cases like if you just
4862 : // changed the GC settings for the tenant to make the PITR window
4863 : // larger, but some of the data was already removed by an earlier GC
4864 : // iteration.
4865 :
4866 : // check against last actual 'latest_gc_cutoff' first
4867 1428 : let applied_gc_cutoff_lsn = src_timeline.get_applied_gc_cutoff_lsn();
4868 1428 : {
4869 1428 : let gc_info = src_timeline.gc_info.read().unwrap();
4870 1428 : let planned_cutoff = gc_info.min_cutoff();
4871 1428 : if gc_info.lsn_covered_by_lease(start_lsn) {
4872 0 : tracing::info!(
4873 0 : "skipping comparison of {start_lsn} with gc cutoff {} and planned gc cutoff {planned_cutoff} due to lsn lease",
4874 0 : *applied_gc_cutoff_lsn
4875 : );
4876 : } else {
4877 1428 : src_timeline
4878 1428 : .check_lsn_is_in_scope(start_lsn, &applied_gc_cutoff_lsn)
4879 1428 : .context(format!(
4880 1428 : "invalid branch start lsn: less than latest GC cutoff {}",
4881 1428 : *applied_gc_cutoff_lsn,
4882 1428 : ))
4883 1428 : .map_err(CreateTimelineError::AncestorLsn)?;
4884 :
4885 : // and then the planned GC cutoff
4886 1404 : if start_lsn < planned_cutoff {
4887 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
4888 0 : "invalid branch start lsn: less than planned GC cutoff {planned_cutoff}"
4889 0 : )));
4890 1404 : }
4891 : }
4892 : }
4893 :
4894 : //
4895 : // The branch point is valid, and we are still holding the 'gc_cs' lock
4896 : // so that GC cannot advance the GC cutoff until we are finished.
4897 : // Proceed with the branch creation.
4898 : //
4899 :
4900 : // Determine prev-LSN for the new timeline. We can only determine it if
4901 : // the timeline was branched at the current end of the source timeline.
4902 : let RecordLsn {
4903 1404 : last: src_last,
4904 1404 : prev: src_prev,
4905 1404 : } = src_timeline.get_last_record_rlsn();
4906 1404 : let dst_prev = if src_last == start_lsn {
4907 1296 : Some(src_prev)
4908 : } else {
4909 108 : None
4910 : };
4911 :
4912 : // Create the metadata file, noting the ancestor of the new timeline.
4913 : // There is initially no data in it, but all the read-calls know to look
4914 : // into the ancestor.
4915 1404 : let metadata = TimelineMetadata::new(
4916 1404 : start_lsn,
4917 1404 : dst_prev,
4918 1404 : Some(src_id),
4919 1404 : start_lsn,
4920 1404 : *src_timeline.applied_gc_cutoff_lsn.read(), // FIXME: should we hold onto this guard longer?
4921 1404 : src_timeline.initdb_lsn,
4922 1404 : src_timeline.pg_version,
4923 1404 : );
4924 :
4925 1404 : let (uninitialized_timeline, _timeline_ctx) = self
4926 1404 : .prepare_new_timeline(
4927 1404 : dst_id,
4928 1404 : &metadata,
4929 1404 : timeline_create_guard,
4930 1404 : start_lsn + 1,
4931 1404 : Some(Arc::clone(src_timeline)),
4932 1404 : Some(src_timeline.get_rel_size_v2_status()),
4933 1404 : ctx,
4934 1404 : )
4935 1404 : .await?;
4936 :
4937 1404 : let new_timeline = uninitialized_timeline.finish_creation().await?;
4938 :
4939 : // Root timeline gets its layers during creation and uploads them along with the metadata.
4940 : // A branch timeline though, when created, can get no writes for some time, hence won't get any layers created.
4941 : // We still need to upload its metadata eagerly: if other nodes `attach` the tenant and miss this timeline, their GC
4942 : // could get incorrect information and remove more layers, than needed.
4943 : // See also https://github.com/neondatabase/neon/issues/3865
4944 1404 : new_timeline
4945 1404 : .remote_client
4946 1404 : .schedule_index_upload_for_full_metadata_update(&metadata)
4947 1404 : .context("branch initial metadata upload")?;
4948 :
4949 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
4950 :
4951 1404 : Ok(CreateTimelineResult::Created(new_timeline))
4952 1428 : }
4953 :
4954 : /// For unit tests, make this visible so that other modules can directly create timelines
4955 : #[cfg(test)]
4956 : #[tracing::instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
4957 : pub(crate) async fn bootstrap_timeline_test(
4958 : self: &Arc<Self>,
4959 : timeline_id: TimelineId,
4960 : pg_version: u32,
4961 : load_existing_initdb: Option<TimelineId>,
4962 : ctx: &RequestContext,
4963 : ) -> anyhow::Result<Arc<Timeline>> {
4964 : self.bootstrap_timeline(timeline_id, pg_version, load_existing_initdb, ctx)
4965 : .await
4966 : .map_err(anyhow::Error::new)
4967 12 : .map(|r| r.into_timeline_for_test())
4968 : }
4969 :
4970 : /// Get exclusive access to the timeline ID for creation.
4971 : ///
4972 : /// Timeline-creating code paths must use this function before making changes
4973 : /// to in-memory or persistent state.
4974 : ///
4975 : /// The `state` parameter is a description of the timeline creation operation
4976 : /// we intend to perform.
4977 : /// If the timeline was already created in the meantime, we check whether this
4978 : /// request conflicts or is idempotent , based on `state`.
4979 2796 : async fn start_creating_timeline(
4980 2796 : self: &Arc<Self>,
4981 2796 : new_timeline_id: TimelineId,
4982 2796 : idempotency: CreateTimelineIdempotency,
4983 2796 : ) -> Result<StartCreatingTimelineResult, CreateTimelineError> {
4984 2796 : let allow_offloaded = false;
4985 2796 : match self.create_timeline_create_guard(new_timeline_id, idempotency, allow_offloaded) {
4986 2784 : Ok(create_guard) => {
4987 2784 : pausable_failpoint!("timeline-creation-after-uninit");
4988 2784 : Ok(StartCreatingTimelineResult::CreateGuard(create_guard))
4989 : }
4990 0 : Err(TimelineExclusionError::ShuttingDown) => Err(CreateTimelineError::ShuttingDown),
4991 : Err(TimelineExclusionError::AlreadyCreating) => {
4992 : // Creation is in progress, we cannot create it again, and we cannot
4993 : // check if this request matches the existing one, so caller must try
4994 : // again later.
4995 0 : Err(CreateTimelineError::AlreadyCreating)
4996 : }
4997 0 : Err(TimelineExclusionError::Other(e)) => Err(CreateTimelineError::Other(e)),
4998 : Err(TimelineExclusionError::AlreadyExists {
4999 0 : existing: TimelineOrOffloaded::Offloaded(_existing),
5000 0 : ..
5001 0 : }) => {
5002 0 : info!("timeline already exists but is offloaded");
5003 0 : Err(CreateTimelineError::Conflict)
5004 : }
5005 : Err(TimelineExclusionError::AlreadyExists {
5006 12 : existing: TimelineOrOffloaded::Timeline(existing),
5007 12 : arg,
5008 12 : }) => {
5009 12 : {
5010 12 : let existing = &existing.create_idempotency;
5011 12 : let _span = info_span!("idempotency_check", ?existing, ?arg).entered();
5012 12 : debug!("timeline already exists");
5013 :
5014 12 : match (existing, &arg) {
5015 : // FailWithConflict => no idempotency check
5016 : (CreateTimelineIdempotency::FailWithConflict, _)
5017 : | (_, CreateTimelineIdempotency::FailWithConflict) => {
5018 12 : warn!("timeline already exists, failing request");
5019 12 : return Err(CreateTimelineError::Conflict);
5020 : }
5021 : // Idempotent <=> CreateTimelineIdempotency is identical
5022 0 : (x, y) if x == y => {
5023 0 : info!(
5024 0 : "timeline already exists and idempotency matches, succeeding request"
5025 : );
5026 : // fallthrough
5027 : }
5028 : (_, _) => {
5029 0 : warn!("idempotency conflict, failing request");
5030 0 : return Err(CreateTimelineError::Conflict);
5031 : }
5032 : }
5033 : }
5034 :
5035 0 : Ok(StartCreatingTimelineResult::Idempotent(existing))
5036 : }
5037 : }
5038 2796 : }
5039 :
5040 0 : async fn upload_initdb(
5041 0 : &self,
5042 0 : timelines_path: &Utf8PathBuf,
5043 0 : pgdata_path: &Utf8PathBuf,
5044 0 : timeline_id: &TimelineId,
5045 0 : ) -> anyhow::Result<()> {
5046 0 : let temp_path = timelines_path.join(format!(
5047 0 : "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
5048 0 : ));
5049 0 :
5050 0 : scopeguard::defer! {
5051 0 : if let Err(e) = fs::remove_file(&temp_path) {
5052 0 : error!("Failed to remove temporary initdb archive '{temp_path}': {e}");
5053 0 : }
5054 0 : }
5055 :
5056 0 : let (pgdata_zstd, tar_zst_size) = create_zst_tarball(pgdata_path, &temp_path).await?;
5057 : const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
5058 0 : if tar_zst_size > INITDB_TAR_ZST_WARN_LIMIT {
5059 0 : warn!(
5060 0 : "compressed {temp_path} size of {tar_zst_size} is above limit {INITDB_TAR_ZST_WARN_LIMIT}."
5061 : );
5062 0 : }
5063 :
5064 0 : pausable_failpoint!("before-initdb-upload");
5065 :
5066 0 : backoff::retry(
5067 0 : || async {
5068 0 : self::remote_timeline_client::upload_initdb_dir(
5069 0 : &self.remote_storage,
5070 0 : &self.tenant_shard_id.tenant_id,
5071 0 : timeline_id,
5072 0 : pgdata_zstd.try_clone().await?,
5073 0 : tar_zst_size,
5074 0 : &self.cancel,
5075 0 : )
5076 0 : .await
5077 0 : },
5078 0 : |_| false,
5079 0 : 3,
5080 0 : u32::MAX,
5081 0 : "persist_initdb_tar_zst",
5082 0 : &self.cancel,
5083 0 : )
5084 0 : .await
5085 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
5086 0 : .and_then(|x| x)
5087 0 : }
5088 :
5089 : /// - run initdb to init temporary instance and get bootstrap data
5090 : /// - after initialization completes, tar up the temp dir and upload it to S3.
5091 12 : async fn bootstrap_timeline(
5092 12 : self: &Arc<Self>,
5093 12 : timeline_id: TimelineId,
5094 12 : pg_version: u32,
5095 12 : load_existing_initdb: Option<TimelineId>,
5096 12 : ctx: &RequestContext,
5097 12 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
5098 12 : let timeline_create_guard = match self
5099 12 : .start_creating_timeline(
5100 12 : timeline_id,
5101 12 : CreateTimelineIdempotency::Bootstrap { pg_version },
5102 12 : )
5103 12 : .await?
5104 : {
5105 12 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
5106 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
5107 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
5108 : }
5109 : };
5110 :
5111 : // create a `tenant/{tenant_id}/timelines/basebackup-{timeline_id}.{TEMP_FILE_SUFFIX}/`
5112 : // temporary directory for basebackup files for the given timeline.
5113 :
5114 12 : let timelines_path = self.conf.timelines_path(&self.tenant_shard_id);
5115 12 : let pgdata_path = path_with_suffix_extension(
5116 12 : timelines_path.join(format!("basebackup-{timeline_id}")),
5117 12 : TEMP_FILE_SUFFIX,
5118 12 : );
5119 12 :
5120 12 : // Remove whatever was left from the previous runs: safe because TimelineCreateGuard guarantees
5121 12 : // we won't race with other creations or existent timelines with the same path.
5122 12 : if pgdata_path.exists() {
5123 0 : fs::remove_dir_all(&pgdata_path).with_context(|| {
5124 0 : format!("Failed to remove already existing initdb directory: {pgdata_path}")
5125 0 : })?;
5126 0 : tracing::info!("removed previous attempt's temporary initdb directory '{pgdata_path}'");
5127 12 : }
5128 :
5129 : // this new directory is very temporary, set to remove it immediately after bootstrap, we don't need it
5130 12 : let pgdata_path_deferred = pgdata_path.clone();
5131 12 : scopeguard::defer! {
5132 12 : if let Err(e) = fs::remove_dir_all(&pgdata_path_deferred).or_else(fs_ext::ignore_not_found) {
5133 12 : // this is unlikely, but we will remove the directory on pageserver restart or another bootstrap call
5134 12 : error!("Failed to remove temporary initdb directory '{pgdata_path_deferred}': {e}");
5135 12 : } else {
5136 12 : tracing::info!("removed temporary initdb directory '{pgdata_path_deferred}'");
5137 12 : }
5138 12 : }
5139 12 : if let Some(existing_initdb_timeline_id) = load_existing_initdb {
5140 12 : if existing_initdb_timeline_id != timeline_id {
5141 0 : let source_path = &remote_initdb_archive_path(
5142 0 : &self.tenant_shard_id.tenant_id,
5143 0 : &existing_initdb_timeline_id,
5144 0 : );
5145 0 : let dest_path =
5146 0 : &remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &timeline_id);
5147 0 :
5148 0 : // if this fails, it will get retried by retried control plane requests
5149 0 : self.remote_storage
5150 0 : .copy_object(source_path, dest_path, &self.cancel)
5151 0 : .await
5152 0 : .context("copy initdb tar")?;
5153 12 : }
5154 12 : let (initdb_tar_zst_path, initdb_tar_zst) =
5155 12 : self::remote_timeline_client::download_initdb_tar_zst(
5156 12 : self.conf,
5157 12 : &self.remote_storage,
5158 12 : &self.tenant_shard_id,
5159 12 : &existing_initdb_timeline_id,
5160 12 : &self.cancel,
5161 12 : )
5162 12 : .await
5163 12 : .context("download initdb tar")?;
5164 :
5165 12 : scopeguard::defer! {
5166 12 : if let Err(e) = fs::remove_file(&initdb_tar_zst_path) {
5167 12 : error!("Failed to remove temporary initdb archive '{initdb_tar_zst_path}': {e}");
5168 12 : }
5169 12 : }
5170 12 :
5171 12 : let buf_read =
5172 12 : BufReader::with_capacity(remote_timeline_client::BUFFER_SIZE, initdb_tar_zst);
5173 12 : extract_zst_tarball(&pgdata_path, buf_read)
5174 12 : .await
5175 12 : .context("extract initdb tar")?;
5176 : } else {
5177 : // Init temporarily repo to get bootstrap data, this creates a directory in the `pgdata_path` path
5178 0 : run_initdb(self.conf, &pgdata_path, pg_version, &self.cancel)
5179 0 : .await
5180 0 : .context("run initdb")?;
5181 :
5182 : // Upload the created data dir to S3
5183 0 : if self.tenant_shard_id().is_shard_zero() {
5184 0 : self.upload_initdb(&timelines_path, &pgdata_path, &timeline_id)
5185 0 : .await?;
5186 0 : }
5187 : }
5188 12 : let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
5189 12 :
5190 12 : // Import the contents of the data directory at the initial checkpoint
5191 12 : // LSN, and any WAL after that.
5192 12 : // Initdb lsn will be equal to last_record_lsn which will be set after import.
5193 12 : // Because we know it upfront avoid having an option or dummy zero value by passing it to the metadata.
5194 12 : let new_metadata = TimelineMetadata::new(
5195 12 : Lsn(0),
5196 12 : None,
5197 12 : None,
5198 12 : Lsn(0),
5199 12 : pgdata_lsn,
5200 12 : pgdata_lsn,
5201 12 : pg_version,
5202 12 : );
5203 12 : let (mut raw_timeline, timeline_ctx) = self
5204 12 : .prepare_new_timeline(
5205 12 : timeline_id,
5206 12 : &new_metadata,
5207 12 : timeline_create_guard,
5208 12 : pgdata_lsn,
5209 12 : None,
5210 12 : None,
5211 12 : ctx,
5212 12 : )
5213 12 : .await?;
5214 :
5215 12 : let tenant_shard_id = raw_timeline.owning_tenant.tenant_shard_id;
5216 12 : raw_timeline
5217 12 : .write(|unfinished_timeline| async move {
5218 12 : import_datadir::import_timeline_from_postgres_datadir(
5219 12 : &unfinished_timeline,
5220 12 : &pgdata_path,
5221 12 : pgdata_lsn,
5222 12 : &timeline_ctx,
5223 12 : )
5224 12 : .await
5225 12 : .with_context(|| {
5226 0 : format!(
5227 0 : "Failed to import pgdatadir for timeline {tenant_shard_id}/{timeline_id}"
5228 0 : )
5229 12 : })?;
5230 :
5231 12 : fail::fail_point!("before-checkpoint-new-timeline", |_| {
5232 0 : Err(CreateTimelineError::Other(anyhow::anyhow!(
5233 0 : "failpoint before-checkpoint-new-timeline"
5234 0 : )))
5235 12 : });
5236 :
5237 12 : Ok(())
5238 24 : })
5239 12 : .await?;
5240 :
5241 : // All done!
5242 12 : let timeline = raw_timeline.finish_creation().await?;
5243 :
5244 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
5245 :
5246 12 : Ok(CreateTimelineResult::Created(timeline))
5247 12 : }
5248 :
5249 2760 : fn build_timeline_remote_client(&self, timeline_id: TimelineId) -> RemoteTimelineClient {
5250 2760 : RemoteTimelineClient::new(
5251 2760 : self.remote_storage.clone(),
5252 2760 : self.deletion_queue_client.clone(),
5253 2760 : self.conf,
5254 2760 : self.tenant_shard_id,
5255 2760 : timeline_id,
5256 2760 : self.generation,
5257 2760 : &self.tenant_conf.load().location,
5258 2760 : )
5259 2760 : }
5260 :
5261 : /// Builds required resources for a new timeline.
5262 2760 : fn build_timeline_resources(&self, timeline_id: TimelineId) -> TimelineResources {
5263 2760 : let remote_client = self.build_timeline_remote_client(timeline_id);
5264 2760 : self.get_timeline_resources_for(remote_client)
5265 2760 : }
5266 :
5267 : /// Builds timeline resources for the given remote client.
5268 2796 : fn get_timeline_resources_for(&self, remote_client: RemoteTimelineClient) -> TimelineResources {
5269 2796 : TimelineResources {
5270 2796 : remote_client,
5271 2796 : pagestream_throttle: self.pagestream_throttle.clone(),
5272 2796 : pagestream_throttle_metrics: self.pagestream_throttle_metrics.clone(),
5273 2796 : l0_compaction_trigger: self.l0_compaction_trigger.clone(),
5274 2796 : l0_flush_global_state: self.l0_flush_global_state.clone(),
5275 2796 : }
5276 2796 : }
5277 :
5278 : /// Creates intermediate timeline structure and its files.
5279 : ///
5280 : /// An empty layer map is initialized, and new data and WAL can be imported starting
5281 : /// at 'disk_consistent_lsn'. After any initial data has been imported, call
5282 : /// `finish_creation` to insert the Timeline into the timelines map.
5283 : #[allow(clippy::too_many_arguments)]
5284 2760 : async fn prepare_new_timeline<'a>(
5285 2760 : &'a self,
5286 2760 : new_timeline_id: TimelineId,
5287 2760 : new_metadata: &TimelineMetadata,
5288 2760 : create_guard: TimelineCreateGuard,
5289 2760 : start_lsn: Lsn,
5290 2760 : ancestor: Option<Arc<Timeline>>,
5291 2760 : rel_size_v2_status: Option<RelSizeMigration>,
5292 2760 : ctx: &RequestContext,
5293 2760 : ) -> anyhow::Result<(UninitializedTimeline<'a>, RequestContext)> {
5294 2760 : let tenant_shard_id = self.tenant_shard_id;
5295 2760 :
5296 2760 : let resources = self.build_timeline_resources(new_timeline_id);
5297 2760 : resources
5298 2760 : .remote_client
5299 2760 : .init_upload_queue_for_empty_remote(new_metadata, rel_size_v2_status.clone())?;
5300 :
5301 2760 : let (timeline_struct, timeline_ctx) = self
5302 2760 : .create_timeline_struct(
5303 2760 : new_timeline_id,
5304 2760 : new_metadata,
5305 2760 : None,
5306 2760 : ancestor,
5307 2760 : resources,
5308 2760 : CreateTimelineCause::Load,
5309 2760 : create_guard.idempotency.clone(),
5310 2760 : None,
5311 2760 : rel_size_v2_status,
5312 2760 : ctx,
5313 2760 : )
5314 2760 : .context("Failed to create timeline data structure")?;
5315 :
5316 2760 : timeline_struct.init_empty_layer_map(start_lsn);
5317 :
5318 2760 : if let Err(e) = self
5319 2760 : .create_timeline_files(&create_guard.timeline_path)
5320 2760 : .await
5321 : {
5322 0 : error!(
5323 0 : "Failed to create initial files for timeline {tenant_shard_id}/{new_timeline_id}, cleaning up: {e:?}"
5324 : );
5325 0 : cleanup_timeline_directory(create_guard);
5326 0 : return Err(e);
5327 2760 : }
5328 2760 :
5329 2760 : debug!(
5330 0 : "Successfully created initial files for timeline {tenant_shard_id}/{new_timeline_id}"
5331 : );
5332 :
5333 2760 : Ok((
5334 2760 : UninitializedTimeline::new(
5335 2760 : self,
5336 2760 : new_timeline_id,
5337 2760 : Some((timeline_struct, create_guard)),
5338 2760 : ),
5339 2760 : timeline_ctx,
5340 2760 : ))
5341 2760 : }
5342 :
5343 2760 : async fn create_timeline_files(&self, timeline_path: &Utf8Path) -> anyhow::Result<()> {
5344 2760 : crashsafe::create_dir(timeline_path).context("Failed to create timeline directory")?;
5345 :
5346 2760 : fail::fail_point!("after-timeline-dir-creation", |_| {
5347 0 : anyhow::bail!("failpoint after-timeline-dir-creation");
5348 2760 : });
5349 :
5350 2760 : Ok(())
5351 2760 : }
5352 :
5353 : /// Get a guard that provides exclusive access to the timeline directory, preventing
5354 : /// concurrent attempts to create the same timeline.
5355 : ///
5356 : /// The `allow_offloaded` parameter controls whether to tolerate the existence of
5357 : /// offloaded timelines or not.
5358 2796 : fn create_timeline_create_guard(
5359 2796 : self: &Arc<Self>,
5360 2796 : timeline_id: TimelineId,
5361 2796 : idempotency: CreateTimelineIdempotency,
5362 2796 : allow_offloaded: bool,
5363 2796 : ) -> Result<TimelineCreateGuard, TimelineExclusionError> {
5364 2796 : let tenant_shard_id = self.tenant_shard_id;
5365 2796 :
5366 2796 : let timeline_path = self.conf.timeline_path(&tenant_shard_id, &timeline_id);
5367 :
5368 2796 : let create_guard = TimelineCreateGuard::new(
5369 2796 : self,
5370 2796 : timeline_id,
5371 2796 : timeline_path.clone(),
5372 2796 : idempotency,
5373 2796 : allow_offloaded,
5374 2796 : )?;
5375 :
5376 : // At this stage, we have got exclusive access to in-memory state for this timeline ID
5377 : // for creation.
5378 : // A timeline directory should never exist on disk already:
5379 : // - a previous failed creation would have cleaned up after itself
5380 : // - a pageserver restart would clean up timeline directories that don't have valid remote state
5381 : //
5382 : // Therefore it is an unexpected internal error to encounter a timeline directory already existing here,
5383 : // this error may indicate a bug in cleanup on failed creations.
5384 2784 : if timeline_path.exists() {
5385 0 : return Err(TimelineExclusionError::Other(anyhow::anyhow!(
5386 0 : "Timeline directory already exists! This is a bug."
5387 0 : )));
5388 2784 : }
5389 2784 :
5390 2784 : Ok(create_guard)
5391 2796 : }
5392 :
5393 : /// Gathers inputs from all of the timelines to produce a sizing model input.
5394 : ///
5395 : /// Future is cancellation safe. Only one calculation can be running at once per tenant.
5396 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5397 : pub async fn gather_size_inputs(
5398 : &self,
5399 : // `max_retention_period` overrides the cutoff that is used to calculate the size
5400 : // (only if it is shorter than the real cutoff).
5401 : max_retention_period: Option<u64>,
5402 : cause: LogicalSizeCalculationCause,
5403 : cancel: &CancellationToken,
5404 : ctx: &RequestContext,
5405 : ) -> Result<size::ModelInputs, size::CalculateSyntheticSizeError> {
5406 : let logical_sizes_at_once = self
5407 : .conf
5408 : .concurrent_tenant_size_logical_size_queries
5409 : .inner();
5410 :
5411 : // TODO: Having a single mutex block concurrent reads is not great for performance.
5412 : //
5413 : // But the only case where we need to run multiple of these at once is when we
5414 : // request a size for a tenant manually via API, while another background calculation
5415 : // is in progress (which is not a common case).
5416 : //
5417 : // See more for on the issue #2748 condenced out of the initial PR review.
5418 : let mut shared_cache = tokio::select! {
5419 : locked = self.cached_logical_sizes.lock() => locked,
5420 : _ = cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5421 : _ = self.cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5422 : };
5423 :
5424 : size::gather_inputs(
5425 : self,
5426 : logical_sizes_at_once,
5427 : max_retention_period,
5428 : &mut shared_cache,
5429 : cause,
5430 : cancel,
5431 : ctx,
5432 : )
5433 : .await
5434 : }
5435 :
5436 : /// Calculate synthetic tenant size and cache the result.
5437 : /// This is periodically called by background worker.
5438 : /// result is cached in tenant struct
5439 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5440 : pub async fn calculate_synthetic_size(
5441 : &self,
5442 : cause: LogicalSizeCalculationCause,
5443 : cancel: &CancellationToken,
5444 : ctx: &RequestContext,
5445 : ) -> Result<u64, size::CalculateSyntheticSizeError> {
5446 : let inputs = self.gather_size_inputs(None, cause, cancel, ctx).await?;
5447 :
5448 : let size = inputs.calculate();
5449 :
5450 : self.set_cached_synthetic_size(size);
5451 :
5452 : Ok(size)
5453 : }
5454 :
5455 : /// Cache given synthetic size and update the metric value
5456 0 : pub fn set_cached_synthetic_size(&self, size: u64) {
5457 0 : self.cached_synthetic_tenant_size
5458 0 : .store(size, Ordering::Relaxed);
5459 0 :
5460 0 : // Only shard zero should be calculating synthetic sizes
5461 0 : debug_assert!(self.shard_identity.is_shard_zero());
5462 :
5463 0 : TENANT_SYNTHETIC_SIZE_METRIC
5464 0 : .get_metric_with_label_values(&[&self.tenant_shard_id.tenant_id.to_string()])
5465 0 : .unwrap()
5466 0 : .set(size);
5467 0 : }
5468 :
5469 0 : pub fn cached_synthetic_size(&self) -> u64 {
5470 0 : self.cached_synthetic_tenant_size.load(Ordering::Relaxed)
5471 0 : }
5472 :
5473 : /// Flush any in-progress layers, schedule uploads, and wait for uploads to complete.
5474 : ///
5475 : /// This function can take a long time: callers should wrap it in a timeout if calling
5476 : /// from an external API handler.
5477 : ///
5478 : /// Cancel-safety: cancelling this function may leave I/O running, but such I/O is
5479 : /// still bounded by tenant/timeline shutdown.
5480 : #[tracing::instrument(skip_all)]
5481 : pub(crate) async fn flush_remote(&self) -> anyhow::Result<()> {
5482 : let timelines = self.timelines.lock().unwrap().clone();
5483 :
5484 0 : async fn flush_timeline(_gate: GateGuard, timeline: Arc<Timeline>) -> anyhow::Result<()> {
5485 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Flushing...");
5486 0 : timeline.freeze_and_flush().await?;
5487 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Waiting for uploads...");
5488 0 : timeline.remote_client.wait_completion().await?;
5489 :
5490 0 : Ok(())
5491 0 : }
5492 :
5493 : // We do not use a JoinSet for these tasks, because we don't want them to be
5494 : // aborted when this function's future is cancelled: they should stay alive
5495 : // holding their GateGuard until they complete, to ensure their I/Os complete
5496 : // before Timeline shutdown completes.
5497 : let mut results = FuturesUnordered::new();
5498 :
5499 : for (_timeline_id, timeline) in timelines {
5500 : // Run each timeline's flush in a task holding the timeline's gate: this
5501 : // means that if this function's future is cancelled, the Timeline shutdown
5502 : // will still wait for any I/O in here to complete.
5503 : let Ok(gate) = timeline.gate.enter() else {
5504 : continue;
5505 : };
5506 0 : let jh = tokio::task::spawn(async move { flush_timeline(gate, timeline).await });
5507 : results.push(jh);
5508 : }
5509 :
5510 : while let Some(r) = results.next().await {
5511 : if let Err(e) = r {
5512 : if !e.is_cancelled() && !e.is_panic() {
5513 : tracing::error!("unexpected join error: {e:?}");
5514 : }
5515 : }
5516 : }
5517 :
5518 : // The flushes we did above were just writes, but the TenantShard might have had
5519 : // pending deletions as well from recent compaction/gc: we want to flush those
5520 : // as well. This requires flushing the global delete queue. This is cheap
5521 : // because it's typically a no-op.
5522 : match self.deletion_queue_client.flush_execute().await {
5523 : Ok(_) => {}
5524 : Err(DeletionQueueError::ShuttingDown) => {}
5525 : }
5526 :
5527 : Ok(())
5528 : }
5529 :
5530 0 : pub(crate) fn get_tenant_conf(&self) -> pageserver_api::models::TenantConfig {
5531 0 : self.tenant_conf.load().tenant_conf.clone()
5532 0 : }
5533 :
5534 : /// How much local storage would this tenant like to have? It can cope with
5535 : /// less than this (via eviction and on-demand downloads), but this function enables
5536 : /// the TenantShard to advertise how much storage it would prefer to have to provide fast I/O
5537 : /// by keeping important things on local disk.
5538 : ///
5539 : /// This is a heuristic, not a guarantee: tenants that are long-idle will actually use less
5540 : /// than they report here, due to layer eviction. Tenants with many active branches may
5541 : /// actually use more than they report here.
5542 0 : pub(crate) fn local_storage_wanted(&self) -> u64 {
5543 0 : let timelines = self.timelines.lock().unwrap();
5544 0 :
5545 0 : // Heuristic: we use the max() of the timelines' visible sizes, rather than the sum. This
5546 0 : // reflects the observation that on tenants with multiple large branches, typically only one
5547 0 : // of them is used actively enough to occupy space on disk.
5548 0 : timelines
5549 0 : .values()
5550 0 : .map(|t| t.metrics.visible_physical_size_gauge.get())
5551 0 : .max()
5552 0 : .unwrap_or(0)
5553 0 : }
5554 :
5555 : /// Builds a new tenant manifest, and uploads it if it differs from the last-known tenant
5556 : /// manifest in `Self::remote_tenant_manifest`.
5557 : ///
5558 : /// TODO: instead of requiring callers to remember to call `maybe_upload_tenant_manifest` after
5559 : /// changing any `TenantShard` state that's included in the manifest, consider making the manifest
5560 : /// the authoritative source of data with an API that automatically uploads on changes. Revisit
5561 : /// this when the manifest is more widely used and we have a better idea of the data model.
5562 1416 : pub(crate) async fn maybe_upload_tenant_manifest(&self) -> Result<(), TenantManifestError> {
5563 : // Multiple tasks may call this function concurrently after mutating the TenantShard runtime
5564 : // state, affecting the manifest generated by `build_tenant_manifest`. We use an async mutex
5565 : // to serialize these callers. `eq_ignoring_version` acts as a slightly inefficient but
5566 : // simple coalescing mechanism.
5567 1416 : let mut guard = tokio::select! {
5568 1416 : guard = self.remote_tenant_manifest.lock() => guard,
5569 1416 : _ = self.cancel.cancelled() => return Err(TenantManifestError::Cancelled),
5570 : };
5571 :
5572 : // Build a new manifest.
5573 1416 : let manifest = self.build_tenant_manifest();
5574 :
5575 : // Check if the manifest has changed. We ignore the version number here, to avoid
5576 : // uploading every manifest on version number bumps.
5577 1416 : if let Some(old) = guard.as_ref() {
5578 48 : if manifest.eq_ignoring_version(old) {
5579 36 : return Ok(());
5580 12 : }
5581 1368 : }
5582 :
5583 : // Upload the manifest. Remote storage does no retries internally, so retry here.
5584 1380 : match backoff::retry(
5585 1380 : || async {
5586 1380 : upload_tenant_manifest(
5587 1380 : &self.remote_storage,
5588 1380 : &self.tenant_shard_id,
5589 1380 : self.generation,
5590 1380 : &manifest,
5591 1380 : &self.cancel,
5592 1380 : )
5593 1380 : .await
5594 2760 : },
5595 1380 : |_| self.cancel.is_cancelled(),
5596 1380 : FAILED_UPLOAD_WARN_THRESHOLD,
5597 1380 : FAILED_REMOTE_OP_RETRIES,
5598 1380 : "uploading tenant manifest",
5599 1380 : &self.cancel,
5600 1380 : )
5601 1380 : .await
5602 : {
5603 0 : None => Err(TenantManifestError::Cancelled),
5604 0 : Some(Err(_)) if self.cancel.is_cancelled() => Err(TenantManifestError::Cancelled),
5605 0 : Some(Err(e)) => Err(TenantManifestError::RemoteStorage(e)),
5606 : Some(Ok(_)) => {
5607 : // Store the successfully uploaded manifest, so that future callers can avoid
5608 : // re-uploading the same thing.
5609 1380 : *guard = Some(manifest);
5610 1380 :
5611 1380 : Ok(())
5612 : }
5613 : }
5614 1416 : }
5615 : }
5616 :
5617 : /// Create the cluster temporarily in 'initdbpath' directory inside the repository
5618 : /// to get bootstrap data for timeline initialization.
5619 0 : async fn run_initdb(
5620 0 : conf: &'static PageServerConf,
5621 0 : initdb_target_dir: &Utf8Path,
5622 0 : pg_version: u32,
5623 0 : cancel: &CancellationToken,
5624 0 : ) -> Result<(), InitdbError> {
5625 0 : let initdb_bin_path = conf
5626 0 : .pg_bin_dir(pg_version)
5627 0 : .map_err(InitdbError::Other)?
5628 0 : .join("initdb");
5629 0 : let initdb_lib_dir = conf.pg_lib_dir(pg_version).map_err(InitdbError::Other)?;
5630 0 : info!(
5631 0 : "running {} in {}, libdir: {}",
5632 : initdb_bin_path, initdb_target_dir, initdb_lib_dir,
5633 : );
5634 :
5635 0 : let _permit = {
5636 0 : let _timer = INITDB_SEMAPHORE_ACQUISITION_TIME.start_timer();
5637 0 : INIT_DB_SEMAPHORE.acquire().await
5638 : };
5639 :
5640 0 : CONCURRENT_INITDBS.inc();
5641 0 : scopeguard::defer! {
5642 0 : CONCURRENT_INITDBS.dec();
5643 0 : }
5644 0 :
5645 0 : let _timer = INITDB_RUN_TIME.start_timer();
5646 0 : let res = postgres_initdb::do_run_initdb(postgres_initdb::RunInitdbArgs {
5647 0 : superuser: &conf.superuser,
5648 0 : locale: &conf.locale,
5649 0 : initdb_bin: &initdb_bin_path,
5650 0 : pg_version,
5651 0 : library_search_path: &initdb_lib_dir,
5652 0 : pgdata: initdb_target_dir,
5653 0 : })
5654 0 : .await
5655 0 : .map_err(InitdbError::Inner);
5656 0 :
5657 0 : // This isn't true cancellation support, see above. Still return an error to
5658 0 : // excercise the cancellation code path.
5659 0 : if cancel.is_cancelled() {
5660 0 : return Err(InitdbError::Cancelled);
5661 0 : }
5662 0 :
5663 0 : res
5664 0 : }
5665 :
5666 : /// Dump contents of a layer file to stdout.
5667 0 : pub async fn dump_layerfile_from_path(
5668 0 : path: &Utf8Path,
5669 0 : verbose: bool,
5670 0 : ctx: &RequestContext,
5671 0 : ) -> anyhow::Result<()> {
5672 : use std::os::unix::fs::FileExt;
5673 :
5674 : // All layer files start with a two-byte "magic" value, to identify the kind of
5675 : // file.
5676 0 : let file = File::open(path)?;
5677 0 : let mut header_buf = [0u8; 2];
5678 0 : file.read_exact_at(&mut header_buf, 0)?;
5679 :
5680 0 : match u16::from_be_bytes(header_buf) {
5681 : crate::IMAGE_FILE_MAGIC => {
5682 0 : ImageLayer::new_for_path(path, file)?
5683 0 : .dump(verbose, ctx)
5684 0 : .await?
5685 : }
5686 : crate::DELTA_FILE_MAGIC => {
5687 0 : DeltaLayer::new_for_path(path, file)?
5688 0 : .dump(verbose, ctx)
5689 0 : .await?
5690 : }
5691 0 : magic => bail!("unrecognized magic identifier: {:?}", magic),
5692 : }
5693 :
5694 0 : Ok(())
5695 0 : }
5696 :
5697 : #[cfg(test)]
5698 : pub(crate) mod harness {
5699 : use bytes::{Bytes, BytesMut};
5700 : use hex_literal::hex;
5701 : use once_cell::sync::OnceCell;
5702 : use pageserver_api::key::Key;
5703 : use pageserver_api::models::ShardParameters;
5704 : use pageserver_api::record::NeonWalRecord;
5705 : use pageserver_api::shard::ShardIndex;
5706 : use utils::id::TenantId;
5707 : use utils::logging;
5708 :
5709 : use super::*;
5710 : use crate::deletion_queue::mock::MockDeletionQueue;
5711 : use crate::l0_flush::L0FlushConfig;
5712 : use crate::walredo::apply_neon;
5713 :
5714 : pub const TIMELINE_ID: TimelineId =
5715 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
5716 : pub const NEW_TIMELINE_ID: TimelineId =
5717 : TimelineId::from_array(hex!("AA223344556677881122334455667788"));
5718 :
5719 : /// Convenience function to create a page image with given string as the only content
5720 30172675 : pub fn test_img(s: &str) -> Bytes {
5721 30172675 : let mut buf = BytesMut::new();
5722 30172675 : buf.extend_from_slice(s.as_bytes());
5723 30172675 : buf.resize(64, 0);
5724 30172675 :
5725 30172675 : buf.freeze()
5726 30172675 : }
5727 :
5728 : pub struct TenantHarness {
5729 : pub conf: &'static PageServerConf,
5730 : pub tenant_conf: pageserver_api::models::TenantConfig,
5731 : pub tenant_shard_id: TenantShardId,
5732 : pub generation: Generation,
5733 : pub shard: ShardIndex,
5734 : pub remote_storage: GenericRemoteStorage,
5735 : pub remote_fs_dir: Utf8PathBuf,
5736 : pub deletion_queue: MockDeletionQueue,
5737 : }
5738 :
5739 : static LOG_HANDLE: OnceCell<()> = OnceCell::new();
5740 :
5741 1548 : pub(crate) fn setup_logging() {
5742 1548 : LOG_HANDLE.get_or_init(|| {
5743 1476 : logging::init(
5744 1476 : logging::LogFormat::Test,
5745 1476 : // enable it in case the tests exercise code paths that use
5746 1476 : // debug_assert_current_span_has_tenant_and_timeline_id
5747 1476 : logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
5748 1476 : logging::Output::Stdout,
5749 1476 : )
5750 1476 : .expect("Failed to init test logging");
5751 1548 : });
5752 1548 : }
5753 :
5754 : impl TenantHarness {
5755 1404 : pub async fn create_custom(
5756 1404 : test_name: &'static str,
5757 1404 : tenant_conf: pageserver_api::models::TenantConfig,
5758 1404 : tenant_id: TenantId,
5759 1404 : shard_identity: ShardIdentity,
5760 1404 : generation: Generation,
5761 1404 : ) -> anyhow::Result<Self> {
5762 1404 : setup_logging();
5763 1404 :
5764 1404 : let repo_dir = PageServerConf::test_repo_dir(test_name);
5765 1404 : let _ = fs::remove_dir_all(&repo_dir);
5766 1404 : fs::create_dir_all(&repo_dir)?;
5767 :
5768 1404 : let conf = PageServerConf::dummy_conf(repo_dir);
5769 1404 : // Make a static copy of the config. This can never be free'd, but that's
5770 1404 : // OK in a test.
5771 1404 : let conf: &'static PageServerConf = Box::leak(Box::new(conf));
5772 1404 :
5773 1404 : let shard = shard_identity.shard_index();
5774 1404 : let tenant_shard_id = TenantShardId {
5775 1404 : tenant_id,
5776 1404 : shard_number: shard.shard_number,
5777 1404 : shard_count: shard.shard_count,
5778 1404 : };
5779 1404 : fs::create_dir_all(conf.tenant_path(&tenant_shard_id))?;
5780 1404 : fs::create_dir_all(conf.timelines_path(&tenant_shard_id))?;
5781 :
5782 : use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
5783 1404 : let remote_fs_dir = conf.workdir.join("localfs");
5784 1404 : std::fs::create_dir_all(&remote_fs_dir).unwrap();
5785 1404 : let config = RemoteStorageConfig {
5786 1404 : storage: RemoteStorageKind::LocalFs {
5787 1404 : local_path: remote_fs_dir.clone(),
5788 1404 : },
5789 1404 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
5790 1404 : small_timeout: RemoteStorageConfig::DEFAULT_SMALL_TIMEOUT,
5791 1404 : };
5792 1404 : let remote_storage = GenericRemoteStorage::from_config(&config).await.unwrap();
5793 1404 : let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()));
5794 1404 :
5795 1404 : Ok(Self {
5796 1404 : conf,
5797 1404 : tenant_conf,
5798 1404 : tenant_shard_id,
5799 1404 : generation,
5800 1404 : shard,
5801 1404 : remote_storage,
5802 1404 : remote_fs_dir,
5803 1404 : deletion_queue,
5804 1404 : })
5805 1404 : }
5806 :
5807 1320 : pub async fn create(test_name: &'static str) -> anyhow::Result<Self> {
5808 1320 : // Disable automatic GC and compaction to make the unit tests more deterministic.
5809 1320 : // The tests perform them manually if needed.
5810 1320 : let tenant_conf = pageserver_api::models::TenantConfig {
5811 1320 : gc_period: Some(Duration::ZERO),
5812 1320 : compaction_period: Some(Duration::ZERO),
5813 1320 : ..Default::default()
5814 1320 : };
5815 1320 : let tenant_id = TenantId::generate();
5816 1320 : let shard = ShardIdentity::unsharded();
5817 1320 : Self::create_custom(
5818 1320 : test_name,
5819 1320 : tenant_conf,
5820 1320 : tenant_id,
5821 1320 : shard,
5822 1320 : Generation::new(0xdeadbeef),
5823 1320 : )
5824 1320 : .await
5825 1320 : }
5826 :
5827 120 : pub fn span(&self) -> tracing::Span {
5828 120 : info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
5829 120 : }
5830 :
5831 1404 : pub(crate) async fn load(&self) -> (Arc<TenantShard>, RequestContext) {
5832 1404 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error)
5833 1404 : .with_scope_unit_test();
5834 1404 : (
5835 1404 : self.do_try_load(&ctx)
5836 1404 : .await
5837 1404 : .expect("failed to load test tenant"),
5838 1404 : ctx,
5839 1404 : )
5840 1404 : }
5841 :
5842 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5843 : pub(crate) async fn do_try_load(
5844 : &self,
5845 : ctx: &RequestContext,
5846 : ) -> anyhow::Result<Arc<TenantShard>> {
5847 : let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
5848 :
5849 : let tenant = Arc::new(TenantShard::new(
5850 : TenantState::Attaching,
5851 : self.conf,
5852 : AttachedTenantConf::try_from(LocationConf::attached_single(
5853 : self.tenant_conf.clone(),
5854 : self.generation,
5855 : &ShardParameters::default(),
5856 : ))
5857 : .unwrap(),
5858 : // This is a legacy/test code path: sharding isn't supported here.
5859 : ShardIdentity::unsharded(),
5860 : Some(walredo_mgr),
5861 : self.tenant_shard_id,
5862 : self.remote_storage.clone(),
5863 : self.deletion_queue.new_client(),
5864 : // TODO: ideally we should run all unit tests with both configs
5865 : L0FlushGlobalState::new(L0FlushConfig::default()),
5866 : ));
5867 :
5868 : let preload = tenant
5869 : .preload(&self.remote_storage, CancellationToken::new())
5870 : .await?;
5871 : tenant.attach(Some(preload), ctx).await?;
5872 :
5873 : tenant.state.send_replace(TenantState::Active);
5874 : for timeline in tenant.timelines.lock().unwrap().values() {
5875 : timeline.set_state(TimelineState::Active);
5876 : }
5877 : Ok(tenant)
5878 : }
5879 :
5880 12 : pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
5881 12 : self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
5882 12 : }
5883 : }
5884 :
5885 : // Mock WAL redo manager that doesn't do much
5886 : pub(crate) struct TestRedoManager;
5887 :
5888 : impl TestRedoManager {
5889 : /// # Cancel-Safety
5890 : ///
5891 : /// This method is cancellation-safe.
5892 321288 : pub async fn request_redo(
5893 321288 : &self,
5894 321288 : key: Key,
5895 321288 : lsn: Lsn,
5896 321288 : base_img: Option<(Lsn, Bytes)>,
5897 321288 : records: Vec<(Lsn, NeonWalRecord)>,
5898 321288 : _pg_version: u32,
5899 321288 : _redo_attempt_type: RedoAttemptType,
5900 321288 : ) -> Result<Bytes, walredo::Error> {
5901 16842120 : let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
5902 321288 : if records_neon {
5903 : // For Neon wal records, we can decode without spawning postgres, so do so.
5904 321288 : let mut page = match (base_img, records.first()) {
5905 156348 : (Some((_lsn, img)), _) => {
5906 156348 : let mut page = BytesMut::new();
5907 156348 : page.extend_from_slice(&img);
5908 156348 : page
5909 : }
5910 164940 : (_, Some((_lsn, rec))) if rec.will_init() => BytesMut::new(),
5911 : _ => {
5912 0 : panic!("Neon WAL redo requires base image or will init record");
5913 : }
5914 : };
5915 :
5916 17163396 : for (record_lsn, record) in records {
5917 16842120 : apply_neon::apply_in_neon(&record, record_lsn, key, &mut page)?;
5918 : }
5919 321276 : Ok(page.freeze())
5920 : } else {
5921 : // We never spawn a postgres walredo process in unit tests: just log what we might have done.
5922 0 : let s = format!(
5923 0 : "redo for {} to get to {}, with {} and {} records",
5924 0 : key,
5925 0 : lsn,
5926 0 : if base_img.is_some() {
5927 0 : "base image"
5928 : } else {
5929 0 : "no base image"
5930 : },
5931 0 : records.len()
5932 0 : );
5933 0 : println!("{s}");
5934 0 :
5935 0 : Ok(test_img(&s))
5936 : }
5937 321288 : }
5938 : }
5939 : }
5940 :
5941 : #[cfg(test)]
5942 : mod tests {
5943 : use std::collections::{BTreeMap, BTreeSet};
5944 :
5945 : use bytes::{Bytes, BytesMut};
5946 : use hex_literal::hex;
5947 : use itertools::Itertools;
5948 : #[cfg(feature = "testing")]
5949 : use models::CompactLsnRange;
5950 : use pageserver_api::key::{
5951 : AUX_KEY_PREFIX, Key, NON_INHERITED_RANGE, RELATION_SIZE_PREFIX, repl_origin_key,
5952 : };
5953 : use pageserver_api::keyspace::KeySpace;
5954 : #[cfg(feature = "testing")]
5955 : use pageserver_api::keyspace::KeySpaceRandomAccum;
5956 : use pageserver_api::models::{CompactionAlgorithm, CompactionAlgorithmSettings};
5957 : #[cfg(feature = "testing")]
5958 : use pageserver_api::record::NeonWalRecord;
5959 : use pageserver_api::value::Value;
5960 : use pageserver_compaction::helpers::overlaps_with;
5961 : #[cfg(feature = "testing")]
5962 : use rand::SeedableRng;
5963 : #[cfg(feature = "testing")]
5964 : use rand::rngs::StdRng;
5965 : use rand::{Rng, thread_rng};
5966 : #[cfg(feature = "testing")]
5967 : use std::ops::Range;
5968 : use storage_layer::{IoConcurrency, PersistentLayerKey};
5969 : use tests::storage_layer::ValuesReconstructState;
5970 : use tests::timeline::{GetVectoredError, ShutdownMode};
5971 : #[cfg(feature = "testing")]
5972 : use timeline::GcInfo;
5973 : #[cfg(feature = "testing")]
5974 : use timeline::InMemoryLayerTestDesc;
5975 : #[cfg(feature = "testing")]
5976 : use timeline::compaction::{KeyHistoryRetention, KeyLogAtLsn};
5977 : use timeline::{CompactOptions, DeltaLayerTestDesc, VersionedKeySpaceQuery};
5978 : use utils::id::TenantId;
5979 :
5980 : use super::*;
5981 : use crate::DEFAULT_PG_VERSION;
5982 : use crate::keyspace::KeySpaceAccum;
5983 : use crate::tenant::harness::*;
5984 : use crate::tenant::timeline::CompactFlags;
5985 :
5986 : static TEST_KEY: Lazy<Key> =
5987 108 : Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));
5988 :
5989 : #[cfg(feature = "testing")]
5990 : struct TestTimelineSpecification {
5991 : start_lsn: Lsn,
5992 : last_record_lsn: Lsn,
5993 :
5994 : in_memory_layers_shape: Vec<(Range<Key>, Range<Lsn>)>,
5995 : delta_layers_shape: Vec<(Range<Key>, Range<Lsn>)>,
5996 : image_layers_shape: Vec<(Range<Key>, Lsn)>,
5997 :
5998 : gap_chance: u8,
5999 : will_init_chance: u8,
6000 : }
6001 :
6002 : #[cfg(feature = "testing")]
6003 : struct Storage {
6004 : storage: HashMap<(Key, Lsn), Value>,
6005 : start_lsn: Lsn,
6006 : }
6007 :
6008 : #[cfg(feature = "testing")]
6009 : impl Storage {
6010 384000 : fn get(&self, key: Key, lsn: Lsn) -> Bytes {
6011 : use bytes::BufMut;
6012 :
6013 384000 : let mut crnt_lsn = lsn;
6014 384000 : let mut got_base = false;
6015 384000 :
6016 384000 : let mut acc = Vec::new();
6017 :
6018 33982452 : while crnt_lsn >= self.start_lsn {
6019 33982452 : if let Some(value) = self.storage.get(&(key, crnt_lsn)) {
6020 17054064 : acc.push(value.clone());
6021 :
6022 16834572 : match value {
6023 16834572 : Value::WalRecord(NeonWalRecord::Test { will_init, .. }) => {
6024 16834572 : if *will_init {
6025 164508 : got_base = true;
6026 164508 : break;
6027 16670064 : }
6028 : }
6029 : Value::Image(_) => {
6030 219492 : got_base = true;
6031 219492 : break;
6032 : }
6033 0 : _ => unreachable!(),
6034 : }
6035 16928388 : }
6036 :
6037 33598452 : crnt_lsn = crnt_lsn.checked_sub(1u64).unwrap();
6038 : }
6039 :
6040 384000 : assert!(
6041 384000 : got_base,
6042 0 : "Input data was incorrect. No base image for {key}@{lsn}"
6043 : );
6044 :
6045 384000 : tracing::debug!("Wal redo depth for {key}@{lsn} is {}", acc.len());
6046 :
6047 384000 : let mut blob = BytesMut::new();
6048 17054064 : for value in acc.into_iter().rev() {
6049 16834572 : match value {
6050 16834572 : Value::WalRecord(NeonWalRecord::Test { append, .. }) => {
6051 16834572 : blob.extend_from_slice(append.as_bytes());
6052 16834572 : }
6053 219492 : Value::Image(img) => {
6054 219492 : blob.put(img);
6055 219492 : }
6056 0 : _ => unreachable!(),
6057 : }
6058 : }
6059 :
6060 384000 : blob.into()
6061 384000 : }
6062 : }
6063 :
6064 : #[cfg(feature = "testing")]
6065 : #[allow(clippy::too_many_arguments)]
6066 12 : async fn randomize_timeline(
6067 12 : tenant: &Arc<TenantShard>,
6068 12 : new_timeline_id: TimelineId,
6069 12 : pg_version: u32,
6070 12 : spec: TestTimelineSpecification,
6071 12 : random: &mut rand::rngs::StdRng,
6072 12 : ctx: &RequestContext,
6073 12 : ) -> anyhow::Result<(Arc<Timeline>, Storage, Vec<Lsn>)> {
6074 12 : let mut storage: HashMap<(Key, Lsn), Value> = HashMap::default();
6075 12 : let mut interesting_lsns = vec![spec.last_record_lsn];
6076 :
6077 24 : for (key_range, lsn_range) in spec.in_memory_layers_shape.iter() {
6078 24 : let mut lsn = lsn_range.start;
6079 2424 : while lsn < lsn_range.end {
6080 2400 : let mut key = key_range.start;
6081 252216 : while key < key_range.end {
6082 249816 : let gap = random.gen_range(1..=100) <= spec.gap_chance;
6083 249816 : let will_init = random.gen_range(1..=100) <= spec.will_init_chance;
6084 249816 :
6085 249816 : if gap {
6086 12216 : continue;
6087 237600 : }
6088 :
6089 237600 : let record = if will_init {
6090 2292 : Value::WalRecord(NeonWalRecord::wal_init(format!("[wil_init {key}@{lsn}]")))
6091 : } else {
6092 235308 : Value::WalRecord(NeonWalRecord::wal_append(format!("[delta {key}@{lsn}]")))
6093 : };
6094 :
6095 237600 : storage.insert((key, lsn), record);
6096 237600 :
6097 237600 : key = key.next();
6098 : }
6099 2400 : lsn = Lsn(lsn.0 + 1);
6100 : }
6101 :
6102 : // Stash some interesting LSN for future use
6103 72 : for offset in [0, 5, 100].iter() {
6104 72 : if *offset == 0 {
6105 24 : interesting_lsns.push(lsn_range.start);
6106 24 : } else {
6107 48 : let below = lsn_range.start.checked_sub(*offset);
6108 48 : match below {
6109 48 : Some(v) if v >= spec.start_lsn => {
6110 48 : interesting_lsns.push(v);
6111 48 : }
6112 0 : _ => {}
6113 : }
6114 :
6115 48 : let above = Lsn(lsn_range.start.0 + offset);
6116 48 : interesting_lsns.push(above);
6117 : }
6118 : }
6119 : }
6120 :
6121 36 : for (key_range, lsn_range) in spec.delta_layers_shape.iter() {
6122 36 : let mut lsn = lsn_range.start;
6123 3780 : while lsn < lsn_range.end {
6124 3744 : let mut key = key_range.start;
6125 133344 : while key < key_range.end {
6126 129600 : let gap = random.gen_range(1..=100) <= spec.gap_chance;
6127 129600 : let will_init = random.gen_range(1..=100) <= spec.will_init_chance;
6128 129600 :
6129 129600 : if gap {
6130 6048 : continue;
6131 123552 : }
6132 :
6133 123552 : let record = if will_init {
6134 1236 : Value::WalRecord(NeonWalRecord::wal_init(format!("[wil_init {key}@{lsn}]")))
6135 : } else {
6136 122316 : Value::WalRecord(NeonWalRecord::wal_append(format!("[delta {key}@{lsn}]")))
6137 : };
6138 :
6139 123552 : storage.insert((key, lsn), record);
6140 123552 :
6141 123552 : key = key.next();
6142 : }
6143 3744 : lsn = Lsn(lsn.0 + 1);
6144 : }
6145 :
6146 : // Stash some interesting LSN for future use
6147 108 : for offset in [0, 5, 100].iter() {
6148 108 : if *offset == 0 {
6149 36 : interesting_lsns.push(lsn_range.start);
6150 36 : } else {
6151 72 : let below = lsn_range.start.checked_sub(*offset);
6152 72 : match below {
6153 72 : Some(v) if v >= spec.start_lsn => {
6154 36 : interesting_lsns.push(v);
6155 36 : }
6156 36 : _ => {}
6157 : }
6158 :
6159 72 : let above = Lsn(lsn_range.start.0 + offset);
6160 72 : interesting_lsns.push(above);
6161 : }
6162 : }
6163 : }
6164 :
6165 36 : for (key_range, lsn) in spec.image_layers_shape.iter() {
6166 36 : let mut key = key_range.start;
6167 1704 : while key < key_range.end {
6168 1668 : let blob = Bytes::from(format!("[image {key}@{lsn}]"));
6169 1668 : let record = Value::Image(blob.clone());
6170 1668 : storage.insert((key, *lsn), record);
6171 1668 :
6172 1668 : key = key.next();
6173 1668 : }
6174 :
6175 : // Stash some interesting LSN for future use
6176 108 : for offset in [0, 5, 100].iter() {
6177 108 : if *offset == 0 {
6178 36 : interesting_lsns.push(*lsn);
6179 36 : } else {
6180 72 : let below = lsn.checked_sub(*offset);
6181 72 : match below {
6182 72 : Some(v) if v >= spec.start_lsn => {
6183 48 : interesting_lsns.push(v);
6184 48 : }
6185 24 : _ => {}
6186 : }
6187 :
6188 72 : let above = Lsn(lsn.0 + offset);
6189 72 : interesting_lsns.push(above);
6190 : }
6191 : }
6192 : }
6193 :
6194 12 : let in_memory_test_layers = {
6195 12 : let mut acc = Vec::new();
6196 :
6197 24 : for (key_range, lsn_range) in spec.in_memory_layers_shape.iter() {
6198 24 : let mut data = Vec::new();
6199 24 :
6200 24 : let mut lsn = lsn_range.start;
6201 2424 : while lsn < lsn_range.end {
6202 2400 : let mut key = key_range.start;
6203 240000 : while key < key_range.end {
6204 237600 : if let Some(record) = storage.get(&(key, lsn)) {
6205 237600 : data.push((key, lsn, record.clone()));
6206 237600 : }
6207 :
6208 237600 : key = key.next();
6209 : }
6210 2400 : lsn = Lsn(lsn.0 + 1);
6211 : }
6212 :
6213 24 : acc.push(InMemoryLayerTestDesc {
6214 24 : data,
6215 24 : lsn_range: lsn_range.clone(),
6216 24 : is_open: false,
6217 24 : })
6218 : }
6219 :
6220 12 : acc
6221 : };
6222 :
6223 12 : let delta_test_layers = {
6224 12 : let mut acc = Vec::new();
6225 :
6226 36 : for (key_range, lsn_range) in spec.delta_layers_shape.iter() {
6227 36 : let mut data = Vec::new();
6228 36 :
6229 36 : let mut lsn = lsn_range.start;
6230 3780 : while lsn < lsn_range.end {
6231 3744 : let mut key = key_range.start;
6232 127296 : while key < key_range.end {
6233 123552 : if let Some(record) = storage.get(&(key, lsn)) {
6234 123552 : data.push((key, lsn, record.clone()));
6235 123552 : }
6236 :
6237 123552 : key = key.next();
6238 : }
6239 3744 : lsn = Lsn(lsn.0 + 1);
6240 : }
6241 :
6242 36 : acc.push(DeltaLayerTestDesc {
6243 36 : data,
6244 36 : lsn_range: lsn_range.clone(),
6245 36 : key_range: key_range.clone(),
6246 36 : })
6247 : }
6248 :
6249 12 : acc
6250 : };
6251 :
6252 12 : let image_test_layers = {
6253 12 : let mut acc = Vec::new();
6254 :
6255 36 : for (key_range, lsn) in spec.image_layers_shape.iter() {
6256 36 : let mut data = Vec::new();
6257 36 :
6258 36 : let mut key = key_range.start;
6259 1704 : while key < key_range.end {
6260 1668 : if let Some(record) = storage.get(&(key, *lsn)) {
6261 1668 : let blob = match record {
6262 1668 : Value::Image(blob) => blob.clone(),
6263 0 : _ => unreachable!(),
6264 : };
6265 :
6266 1668 : data.push((key, blob));
6267 0 : }
6268 :
6269 1668 : key = key.next();
6270 : }
6271 :
6272 36 : acc.push((*lsn, data));
6273 : }
6274 :
6275 12 : acc
6276 : };
6277 :
6278 12 : let tline = tenant
6279 12 : .create_test_timeline_with_layers(
6280 12 : new_timeline_id,
6281 12 : spec.start_lsn,
6282 12 : pg_version,
6283 12 : ctx,
6284 12 : in_memory_test_layers,
6285 12 : delta_test_layers,
6286 12 : image_test_layers,
6287 12 : spec.last_record_lsn,
6288 12 : )
6289 12 : .await?;
6290 :
6291 12 : Ok((
6292 12 : tline,
6293 12 : Storage {
6294 12 : storage,
6295 12 : start_lsn: spec.start_lsn,
6296 12 : },
6297 12 : interesting_lsns,
6298 12 : ))
6299 12 : }
6300 :
6301 : #[tokio::test]
6302 12 : async fn test_basic() -> anyhow::Result<()> {
6303 12 : let (tenant, ctx) = TenantHarness::create("test_basic").await?.load().await;
6304 12 : let tline = tenant
6305 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6306 12 : .await?;
6307 12 :
6308 12 : let mut writer = tline.writer().await;
6309 12 : writer
6310 12 : .put(
6311 12 : *TEST_KEY,
6312 12 : Lsn(0x10),
6313 12 : &Value::Image(test_img("foo at 0x10")),
6314 12 : &ctx,
6315 12 : )
6316 12 : .await?;
6317 12 : writer.finish_write(Lsn(0x10));
6318 12 : drop(writer);
6319 12 :
6320 12 : let mut writer = tline.writer().await;
6321 12 : writer
6322 12 : .put(
6323 12 : *TEST_KEY,
6324 12 : Lsn(0x20),
6325 12 : &Value::Image(test_img("foo at 0x20")),
6326 12 : &ctx,
6327 12 : )
6328 12 : .await?;
6329 12 : writer.finish_write(Lsn(0x20));
6330 12 : drop(writer);
6331 12 :
6332 12 : assert_eq!(
6333 12 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
6334 12 : test_img("foo at 0x10")
6335 12 : );
6336 12 : assert_eq!(
6337 12 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
6338 12 : test_img("foo at 0x10")
6339 12 : );
6340 12 : assert_eq!(
6341 12 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
6342 12 : test_img("foo at 0x20")
6343 12 : );
6344 12 :
6345 12 : Ok(())
6346 12 : }
6347 :
6348 : #[tokio::test]
6349 12 : async fn no_duplicate_timelines() -> anyhow::Result<()> {
6350 12 : let (tenant, ctx) = TenantHarness::create("no_duplicate_timelines")
6351 12 : .await?
6352 12 : .load()
6353 12 : .await;
6354 12 : let _ = tenant
6355 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6356 12 : .await?;
6357 12 :
6358 12 : match tenant
6359 12 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6360 12 : .await
6361 12 : {
6362 12 : Ok(_) => panic!("duplicate timeline creation should fail"),
6363 12 : Err(e) => assert_eq!(
6364 12 : e.to_string(),
6365 12 : "timeline already exists with different parameters".to_string()
6366 12 : ),
6367 12 : }
6368 12 :
6369 12 : Ok(())
6370 12 : }
6371 :
6372 : /// Convenience function to create a page image with given string as the only content
6373 60 : pub fn test_value(s: &str) -> Value {
6374 60 : let mut buf = BytesMut::new();
6375 60 : buf.extend_from_slice(s.as_bytes());
6376 60 : Value::Image(buf.freeze())
6377 60 : }
6378 :
6379 : ///
6380 : /// Test branch creation
6381 : ///
6382 : #[tokio::test]
6383 12 : async fn test_branch() -> anyhow::Result<()> {
6384 12 : use std::str::from_utf8;
6385 12 :
6386 12 : let (tenant, ctx) = TenantHarness::create("test_branch").await?.load().await;
6387 12 : let tline = tenant
6388 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6389 12 : .await?;
6390 12 : let mut writer = tline.writer().await;
6391 12 :
6392 12 : #[allow(non_snake_case)]
6393 12 : let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
6394 12 : #[allow(non_snake_case)]
6395 12 : let TEST_KEY_B: Key = Key::from_hex("110000000033333333444444445500000002").unwrap();
6396 12 :
6397 12 : // Insert a value on the timeline
6398 12 : writer
6399 12 : .put(TEST_KEY_A, Lsn(0x20), &test_value("foo at 0x20"), &ctx)
6400 12 : .await?;
6401 12 : writer
6402 12 : .put(TEST_KEY_B, Lsn(0x20), &test_value("foobar at 0x20"), &ctx)
6403 12 : .await?;
6404 12 : writer.finish_write(Lsn(0x20));
6405 12 :
6406 12 : writer
6407 12 : .put(TEST_KEY_A, Lsn(0x30), &test_value("foo at 0x30"), &ctx)
6408 12 : .await?;
6409 12 : writer.finish_write(Lsn(0x30));
6410 12 : writer
6411 12 : .put(TEST_KEY_A, Lsn(0x40), &test_value("foo at 0x40"), &ctx)
6412 12 : .await?;
6413 12 : writer.finish_write(Lsn(0x40));
6414 12 :
6415 12 : //assert_current_logical_size(&tline, Lsn(0x40));
6416 12 :
6417 12 : // Branch the history, modify relation differently on the new timeline
6418 12 : tenant
6419 12 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x30)), &ctx)
6420 12 : .await?;
6421 12 : let newtline = tenant
6422 12 : .get_timeline(NEW_TIMELINE_ID, true)
6423 12 : .expect("Should have a local timeline");
6424 12 : let mut new_writer = newtline.writer().await;
6425 12 : new_writer
6426 12 : .put(TEST_KEY_A, Lsn(0x40), &test_value("bar at 0x40"), &ctx)
6427 12 : .await?;
6428 12 : new_writer.finish_write(Lsn(0x40));
6429 12 :
6430 12 : // Check page contents on both branches
6431 12 : assert_eq!(
6432 12 : from_utf8(&tline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
6433 12 : "foo at 0x40"
6434 12 : );
6435 12 : assert_eq!(
6436 12 : from_utf8(&newtline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
6437 12 : "bar at 0x40"
6438 12 : );
6439 12 : assert_eq!(
6440 12 : from_utf8(&newtline.get(TEST_KEY_B, Lsn(0x40), &ctx).await?)?,
6441 12 : "foobar at 0x20"
6442 12 : );
6443 12 :
6444 12 : //assert_current_logical_size(&tline, Lsn(0x40));
6445 12 :
6446 12 : Ok(())
6447 12 : }
6448 :
6449 120 : async fn make_some_layers(
6450 120 : tline: &Timeline,
6451 120 : start_lsn: Lsn,
6452 120 : ctx: &RequestContext,
6453 120 : ) -> anyhow::Result<()> {
6454 120 : let mut lsn = start_lsn;
6455 : {
6456 120 : let mut writer = tline.writer().await;
6457 : // Create a relation on the timeline
6458 120 : writer
6459 120 : .put(
6460 120 : *TEST_KEY,
6461 120 : lsn,
6462 120 : &Value::Image(test_img(&format!("foo at {}", lsn))),
6463 120 : ctx,
6464 120 : )
6465 120 : .await?;
6466 120 : writer.finish_write(lsn);
6467 120 : lsn += 0x10;
6468 120 : writer
6469 120 : .put(
6470 120 : *TEST_KEY,
6471 120 : lsn,
6472 120 : &Value::Image(test_img(&format!("foo at {}", lsn))),
6473 120 : ctx,
6474 120 : )
6475 120 : .await?;
6476 120 : writer.finish_write(lsn);
6477 120 : lsn += 0x10;
6478 120 : }
6479 120 : tline.freeze_and_flush().await?;
6480 : {
6481 120 : let mut writer = tline.writer().await;
6482 120 : writer
6483 120 : .put(
6484 120 : *TEST_KEY,
6485 120 : lsn,
6486 120 : &Value::Image(test_img(&format!("foo at {}", lsn))),
6487 120 : ctx,
6488 120 : )
6489 120 : .await?;
6490 120 : writer.finish_write(lsn);
6491 120 : lsn += 0x10;
6492 120 : writer
6493 120 : .put(
6494 120 : *TEST_KEY,
6495 120 : lsn,
6496 120 : &Value::Image(test_img(&format!("foo at {}", lsn))),
6497 120 : ctx,
6498 120 : )
6499 120 : .await?;
6500 120 : writer.finish_write(lsn);
6501 120 : }
6502 120 : tline.freeze_and_flush().await.map_err(|e| e.into())
6503 120 : }
6504 :
6505 : #[tokio::test(start_paused = true)]
6506 12 : async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
6507 12 : let (tenant, ctx) =
6508 12 : TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
6509 12 : .await?
6510 12 : .load()
6511 12 : .await;
6512 12 : // Advance to the lsn lease deadline so that GC is not blocked by
6513 12 : // initial transition into AttachedSingle.
6514 12 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
6515 12 : tokio::time::resume();
6516 12 : let tline = tenant
6517 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6518 12 : .await?;
6519 12 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6520 12 :
6521 12 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
6522 12 : // FIXME: this doesn't actually remove any layer currently, given how the flushing
6523 12 : // and compaction works. But it does set the 'cutoff' point so that the cross check
6524 12 : // below should fail.
6525 12 : tenant
6526 12 : .gc_iteration(
6527 12 : Some(TIMELINE_ID),
6528 12 : 0x10,
6529 12 : Duration::ZERO,
6530 12 : &CancellationToken::new(),
6531 12 : &ctx,
6532 12 : )
6533 12 : .await?;
6534 12 :
6535 12 : // try to branch at lsn 25, should fail because we already garbage collected the data
6536 12 : match tenant
6537 12 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
6538 12 : .await
6539 12 : {
6540 12 : Ok(_) => panic!("branching should have failed"),
6541 12 : Err(err) => {
6542 12 : let CreateTimelineError::AncestorLsn(err) = err else {
6543 12 : panic!("wrong error type")
6544 12 : };
6545 12 : assert!(err.to_string().contains("invalid branch start lsn"));
6546 12 : assert!(
6547 12 : err.source()
6548 12 : .unwrap()
6549 12 : .to_string()
6550 12 : .contains("we might've already garbage collected needed data")
6551 12 : )
6552 12 : }
6553 12 : }
6554 12 :
6555 12 : Ok(())
6556 12 : }
6557 :
6558 : #[tokio::test]
6559 12 : async fn test_prohibit_branch_creation_on_pre_initdb_lsn() -> anyhow::Result<()> {
6560 12 : let (tenant, ctx) =
6561 12 : TenantHarness::create("test_prohibit_branch_creation_on_pre_initdb_lsn")
6562 12 : .await?
6563 12 : .load()
6564 12 : .await;
6565 12 :
6566 12 : let tline = tenant
6567 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x50), DEFAULT_PG_VERSION, &ctx)
6568 12 : .await?;
6569 12 : // try to branch at lsn 0x25, should fail because initdb lsn is 0x50
6570 12 : match tenant
6571 12 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
6572 12 : .await
6573 12 : {
6574 12 : Ok(_) => panic!("branching should have failed"),
6575 12 : Err(err) => {
6576 12 : let CreateTimelineError::AncestorLsn(err) = err else {
6577 12 : panic!("wrong error type");
6578 12 : };
6579 12 : assert!(&err.to_string().contains("invalid branch start lsn"));
6580 12 : assert!(
6581 12 : &err.source()
6582 12 : .unwrap()
6583 12 : .to_string()
6584 12 : .contains("is earlier than latest GC cutoff")
6585 12 : );
6586 12 : }
6587 12 : }
6588 12 :
6589 12 : Ok(())
6590 12 : }
6591 :
6592 : /*
6593 : // FIXME: This currently fails to error out. Calling GC doesn't currently
6594 : // remove the old value, we'd need to work a little harder
6595 : #[tokio::test]
6596 : async fn test_prohibit_get_for_garbage_collected_data() -> anyhow::Result<()> {
6597 : let repo =
6598 : RepoHarness::create("test_prohibit_get_for_garbage_collected_data")?
6599 : .load();
6600 :
6601 : let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION)?;
6602 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6603 :
6604 : repo.gc_iteration(Some(TIMELINE_ID), 0x10, Duration::ZERO)?;
6605 : let applied_gc_cutoff_lsn = tline.get_applied_gc_cutoff_lsn();
6606 : assert!(*applied_gc_cutoff_lsn > Lsn(0x25));
6607 : match tline.get(*TEST_KEY, Lsn(0x25)) {
6608 : Ok(_) => panic!("request for page should have failed"),
6609 : Err(err) => assert!(err.to_string().contains("not found at")),
6610 : }
6611 : Ok(())
6612 : }
6613 : */
6614 :
6615 : #[tokio::test]
6616 12 : async fn test_get_branchpoints_from_an_inactive_timeline() -> anyhow::Result<()> {
6617 12 : let (tenant, ctx) =
6618 12 : TenantHarness::create("test_get_branchpoints_from_an_inactive_timeline")
6619 12 : .await?
6620 12 : .load()
6621 12 : .await;
6622 12 : let tline = tenant
6623 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6624 12 : .await?;
6625 12 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6626 12 :
6627 12 : tenant
6628 12 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6629 12 : .await?;
6630 12 : let newtline = tenant
6631 12 : .get_timeline(NEW_TIMELINE_ID, true)
6632 12 : .expect("Should have a local timeline");
6633 12 :
6634 12 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6635 12 :
6636 12 : tline.set_broken("test".to_owned());
6637 12 :
6638 12 : tenant
6639 12 : .gc_iteration(
6640 12 : Some(TIMELINE_ID),
6641 12 : 0x10,
6642 12 : Duration::ZERO,
6643 12 : &CancellationToken::new(),
6644 12 : &ctx,
6645 12 : )
6646 12 : .await?;
6647 12 :
6648 12 : // The branchpoints should contain all timelines, even ones marked
6649 12 : // as Broken.
6650 12 : {
6651 12 : let branchpoints = &tline.gc_info.read().unwrap().retain_lsns;
6652 12 : assert_eq!(branchpoints.len(), 1);
6653 12 : assert_eq!(
6654 12 : branchpoints[0],
6655 12 : (Lsn(0x40), NEW_TIMELINE_ID, MaybeOffloaded::No)
6656 12 : );
6657 12 : }
6658 12 :
6659 12 : // You can read the key from the child branch even though the parent is
6660 12 : // Broken, as long as you don't need to access data from the parent.
6661 12 : assert_eq!(
6662 12 : newtline.get(*TEST_KEY, Lsn(0x70), &ctx).await?,
6663 12 : test_img(&format!("foo at {}", Lsn(0x70)))
6664 12 : );
6665 12 :
6666 12 : // This needs to traverse to the parent, and fails.
6667 12 : let err = newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await.unwrap_err();
6668 12 : assert!(
6669 12 : err.to_string().starts_with(&format!(
6670 12 : "bad state on timeline {}: Broken",
6671 12 : tline.timeline_id
6672 12 : )),
6673 12 : "{err}"
6674 12 : );
6675 12 :
6676 12 : Ok(())
6677 12 : }
6678 :
6679 : #[tokio::test]
6680 12 : async fn test_retain_data_in_parent_which_is_needed_for_child() -> anyhow::Result<()> {
6681 12 : let (tenant, ctx) =
6682 12 : TenantHarness::create("test_retain_data_in_parent_which_is_needed_for_child")
6683 12 : .await?
6684 12 : .load()
6685 12 : .await;
6686 12 : let tline = tenant
6687 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6688 12 : .await?;
6689 12 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6690 12 :
6691 12 : tenant
6692 12 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6693 12 : .await?;
6694 12 : let newtline = tenant
6695 12 : .get_timeline(NEW_TIMELINE_ID, true)
6696 12 : .expect("Should have a local timeline");
6697 12 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
6698 12 : tenant
6699 12 : .gc_iteration(
6700 12 : Some(TIMELINE_ID),
6701 12 : 0x10,
6702 12 : Duration::ZERO,
6703 12 : &CancellationToken::new(),
6704 12 : &ctx,
6705 12 : )
6706 12 : .await?;
6707 12 : assert!(newtline.get(*TEST_KEY, Lsn(0x25), &ctx).await.is_ok());
6708 12 :
6709 12 : Ok(())
6710 12 : }
6711 : #[tokio::test]
6712 12 : async fn test_parent_keeps_data_forever_after_branching() -> anyhow::Result<()> {
6713 12 : let (tenant, ctx) = TenantHarness::create("test_parent_keeps_data_forever_after_branching")
6714 12 : .await?
6715 12 : .load()
6716 12 : .await;
6717 12 : let tline = tenant
6718 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6719 12 : .await?;
6720 12 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6721 12 :
6722 12 : tenant
6723 12 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6724 12 : .await?;
6725 12 : let newtline = tenant
6726 12 : .get_timeline(NEW_TIMELINE_ID, true)
6727 12 : .expect("Should have a local timeline");
6728 12 :
6729 12 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6730 12 :
6731 12 : // run gc on parent
6732 12 : tenant
6733 12 : .gc_iteration(
6734 12 : Some(TIMELINE_ID),
6735 12 : 0x10,
6736 12 : Duration::ZERO,
6737 12 : &CancellationToken::new(),
6738 12 : &ctx,
6739 12 : )
6740 12 : .await?;
6741 12 :
6742 12 : // Check that the data is still accessible on the branch.
6743 12 : assert_eq!(
6744 12 : newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
6745 12 : test_img(&format!("foo at {}", Lsn(0x40)))
6746 12 : );
6747 12 :
6748 12 : Ok(())
6749 12 : }
6750 :
6751 : #[tokio::test]
6752 12 : async fn timeline_load() -> anyhow::Result<()> {
6753 12 : const TEST_NAME: &str = "timeline_load";
6754 12 : let harness = TenantHarness::create(TEST_NAME).await?;
6755 12 : {
6756 12 : let (tenant, ctx) = harness.load().await;
6757 12 : let tline = tenant
6758 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x7000), DEFAULT_PG_VERSION, &ctx)
6759 12 : .await?;
6760 12 : make_some_layers(tline.as_ref(), Lsn(0x8000), &ctx).await?;
6761 12 : // so that all uploads finish & we can call harness.load() below again
6762 12 : tenant
6763 12 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
6764 12 : .instrument(harness.span())
6765 12 : .await
6766 12 : .ok()
6767 12 : .unwrap();
6768 12 : }
6769 12 :
6770 12 : let (tenant, _ctx) = harness.load().await;
6771 12 : tenant
6772 12 : .get_timeline(TIMELINE_ID, true)
6773 12 : .expect("cannot load timeline");
6774 12 :
6775 12 : Ok(())
6776 12 : }
6777 :
6778 : #[tokio::test]
6779 12 : async fn timeline_load_with_ancestor() -> anyhow::Result<()> {
6780 12 : const TEST_NAME: &str = "timeline_load_with_ancestor";
6781 12 : let harness = TenantHarness::create(TEST_NAME).await?;
6782 12 : // create two timelines
6783 12 : {
6784 12 : let (tenant, ctx) = harness.load().await;
6785 12 : let tline = tenant
6786 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6787 12 : .await?;
6788 12 :
6789 12 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6790 12 :
6791 12 : let child_tline = tenant
6792 12 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6793 12 : .await?;
6794 12 : child_tline.set_state(TimelineState::Active);
6795 12 :
6796 12 : let newtline = tenant
6797 12 : .get_timeline(NEW_TIMELINE_ID, true)
6798 12 : .expect("Should have a local timeline");
6799 12 :
6800 12 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6801 12 :
6802 12 : // so that all uploads finish & we can call harness.load() below again
6803 12 : tenant
6804 12 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
6805 12 : .instrument(harness.span())
6806 12 : .await
6807 12 : .ok()
6808 12 : .unwrap();
6809 12 : }
6810 12 :
6811 12 : // check that both of them are initially unloaded
6812 12 : let (tenant, _ctx) = harness.load().await;
6813 12 :
6814 12 : // check that both, child and ancestor are loaded
6815 12 : let _child_tline = tenant
6816 12 : .get_timeline(NEW_TIMELINE_ID, true)
6817 12 : .expect("cannot get child timeline loaded");
6818 12 :
6819 12 : let _ancestor_tline = tenant
6820 12 : .get_timeline(TIMELINE_ID, true)
6821 12 : .expect("cannot get ancestor timeline loaded");
6822 12 :
6823 12 : Ok(())
6824 12 : }
6825 :
6826 : #[tokio::test]
6827 12 : async fn delta_layer_dumping() -> anyhow::Result<()> {
6828 12 : use storage_layer::AsLayerDesc;
6829 12 : let (tenant, ctx) = TenantHarness::create("test_layer_dumping")
6830 12 : .await?
6831 12 : .load()
6832 12 : .await;
6833 12 : let tline = tenant
6834 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6835 12 : .await?;
6836 12 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6837 12 :
6838 12 : let layer_map = tline.layers.read().await;
6839 12 : let level0_deltas = layer_map
6840 12 : .layer_map()?
6841 12 : .level0_deltas()
6842 12 : .iter()
6843 24 : .map(|desc| layer_map.get_from_desc(desc))
6844 12 : .collect::<Vec<_>>();
6845 12 :
6846 12 : assert!(!level0_deltas.is_empty());
6847 12 :
6848 36 : for delta in level0_deltas {
6849 12 : // Ensure we are dumping a delta layer here
6850 24 : assert!(delta.layer_desc().is_delta);
6851 24 : delta.dump(true, &ctx).await.unwrap();
6852 12 : }
6853 12 :
6854 12 : Ok(())
6855 12 : }
6856 :
6857 : #[tokio::test]
6858 12 : async fn test_images() -> anyhow::Result<()> {
6859 12 : let (tenant, ctx) = TenantHarness::create("test_images").await?.load().await;
6860 12 : let tline = tenant
6861 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6862 12 : .await?;
6863 12 :
6864 12 : let mut writer = tline.writer().await;
6865 12 : writer
6866 12 : .put(
6867 12 : *TEST_KEY,
6868 12 : Lsn(0x10),
6869 12 : &Value::Image(test_img("foo at 0x10")),
6870 12 : &ctx,
6871 12 : )
6872 12 : .await?;
6873 12 : writer.finish_write(Lsn(0x10));
6874 12 : drop(writer);
6875 12 :
6876 12 : tline.freeze_and_flush().await?;
6877 12 : tline
6878 12 : .compact(&CancellationToken::new(), EnumSet::default(), &ctx)
6879 12 : .await?;
6880 12 :
6881 12 : let mut writer = tline.writer().await;
6882 12 : writer
6883 12 : .put(
6884 12 : *TEST_KEY,
6885 12 : Lsn(0x20),
6886 12 : &Value::Image(test_img("foo at 0x20")),
6887 12 : &ctx,
6888 12 : )
6889 12 : .await?;
6890 12 : writer.finish_write(Lsn(0x20));
6891 12 : drop(writer);
6892 12 :
6893 12 : tline.freeze_and_flush().await?;
6894 12 : tline
6895 12 : .compact(&CancellationToken::new(), EnumSet::default(), &ctx)
6896 12 : .await?;
6897 12 :
6898 12 : let mut writer = tline.writer().await;
6899 12 : writer
6900 12 : .put(
6901 12 : *TEST_KEY,
6902 12 : Lsn(0x30),
6903 12 : &Value::Image(test_img("foo at 0x30")),
6904 12 : &ctx,
6905 12 : )
6906 12 : .await?;
6907 12 : writer.finish_write(Lsn(0x30));
6908 12 : drop(writer);
6909 12 :
6910 12 : tline.freeze_and_flush().await?;
6911 12 : tline
6912 12 : .compact(&CancellationToken::new(), EnumSet::default(), &ctx)
6913 12 : .await?;
6914 12 :
6915 12 : let mut writer = tline.writer().await;
6916 12 : writer
6917 12 : .put(
6918 12 : *TEST_KEY,
6919 12 : Lsn(0x40),
6920 12 : &Value::Image(test_img("foo at 0x40")),
6921 12 : &ctx,
6922 12 : )
6923 12 : .await?;
6924 12 : writer.finish_write(Lsn(0x40));
6925 12 : drop(writer);
6926 12 :
6927 12 : tline.freeze_and_flush().await?;
6928 12 : tline
6929 12 : .compact(&CancellationToken::new(), EnumSet::default(), &ctx)
6930 12 : .await?;
6931 12 :
6932 12 : assert_eq!(
6933 12 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
6934 12 : test_img("foo at 0x10")
6935 12 : );
6936 12 : assert_eq!(
6937 12 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
6938 12 : test_img("foo at 0x10")
6939 12 : );
6940 12 : assert_eq!(
6941 12 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
6942 12 : test_img("foo at 0x20")
6943 12 : );
6944 12 : assert_eq!(
6945 12 : tline.get(*TEST_KEY, Lsn(0x30), &ctx).await?,
6946 12 : test_img("foo at 0x30")
6947 12 : );
6948 12 : assert_eq!(
6949 12 : tline.get(*TEST_KEY, Lsn(0x40), &ctx).await?,
6950 12 : test_img("foo at 0x40")
6951 12 : );
6952 12 :
6953 12 : Ok(())
6954 12 : }
6955 :
6956 24 : async fn bulk_insert_compact_gc(
6957 24 : tenant: &TenantShard,
6958 24 : timeline: &Arc<Timeline>,
6959 24 : ctx: &RequestContext,
6960 24 : lsn: Lsn,
6961 24 : repeat: usize,
6962 24 : key_count: usize,
6963 24 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
6964 24 : let compact = true;
6965 24 : bulk_insert_maybe_compact_gc(tenant, timeline, ctx, lsn, repeat, key_count, compact).await
6966 24 : }
6967 :
6968 48 : async fn bulk_insert_maybe_compact_gc(
6969 48 : tenant: &TenantShard,
6970 48 : timeline: &Arc<Timeline>,
6971 48 : ctx: &RequestContext,
6972 48 : mut lsn: Lsn,
6973 48 : repeat: usize,
6974 48 : key_count: usize,
6975 48 : compact: bool,
6976 48 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
6977 48 : let mut inserted: HashMap<Key, BTreeSet<Lsn>> = Default::default();
6978 48 :
6979 48 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6980 48 : let mut blknum = 0;
6981 48 :
6982 48 : // Enforce that key range is monotonously increasing
6983 48 : let mut keyspace = KeySpaceAccum::new();
6984 48 :
6985 48 : let cancel = CancellationToken::new();
6986 48 :
6987 48 : for _ in 0..repeat {
6988 2400 : for _ in 0..key_count {
6989 24000000 : test_key.field6 = blknum;
6990 24000000 : let mut writer = timeline.writer().await;
6991 24000000 : writer
6992 24000000 : .put(
6993 24000000 : test_key,
6994 24000000 : lsn,
6995 24000000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6996 24000000 : ctx,
6997 24000000 : )
6998 24000000 : .await?;
6999 24000000 : inserted.entry(test_key).or_default().insert(lsn);
7000 24000000 : writer.finish_write(lsn);
7001 24000000 : drop(writer);
7002 24000000 :
7003 24000000 : keyspace.add_key(test_key);
7004 24000000 :
7005 24000000 : lsn = Lsn(lsn.0 + 0x10);
7006 24000000 : blknum += 1;
7007 : }
7008 :
7009 2400 : timeline.freeze_and_flush().await?;
7010 2400 : if compact {
7011 : // this requires timeline to be &Arc<Timeline>
7012 1200 : timeline.compact(&cancel, EnumSet::default(), ctx).await?;
7013 1200 : }
7014 :
7015 : // this doesn't really need to use the timeline_id target, but it is closer to what it
7016 : // originally was.
7017 2400 : let res = tenant
7018 2400 : .gc_iteration(Some(timeline.timeline_id), 0, Duration::ZERO, &cancel, ctx)
7019 2400 : .await?;
7020 :
7021 2400 : assert_eq!(res.layers_removed, 0, "this never removes anything");
7022 : }
7023 :
7024 48 : Ok(inserted)
7025 48 : }
7026 :
7027 : //
7028 : // Insert 1000 key-value pairs with increasing keys, flush, compact, GC.
7029 : // Repeat 50 times.
7030 : //
7031 : #[tokio::test]
7032 12 : async fn test_bulk_insert() -> anyhow::Result<()> {
7033 12 : let harness = TenantHarness::create("test_bulk_insert").await?;
7034 12 : let (tenant, ctx) = harness.load().await;
7035 12 : let tline = tenant
7036 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
7037 12 : .await?;
7038 12 :
7039 12 : let lsn = Lsn(0x10);
7040 12 : bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
7041 12 :
7042 12 : Ok(())
7043 12 : }
7044 :
7045 : // Test the vectored get real implementation against a simple sequential implementation.
7046 : //
7047 : // The test generates a keyspace by repeatedly flushing the in-memory layer and compacting.
7048 : // Projected to 2D the key space looks like below. Lsn grows upwards on the Y axis and keys
7049 : // grow to the right on the X axis.
7050 : // [Delta]
7051 : // [Delta]
7052 : // [Delta]
7053 : // [Delta]
7054 : // ------------ Image ---------------
7055 : //
7056 : // After layer generation we pick the ranges to query as follows:
7057 : // 1. The beginning of each delta layer
7058 : // 2. At the seam between two adjacent delta layers
7059 : //
7060 : // There's one major downside to this test: delta layers only contains images,
7061 : // so the search can stop at the first delta layer and doesn't traverse any deeper.
7062 : #[tokio::test]
7063 12 : async fn test_get_vectored() -> anyhow::Result<()> {
7064 12 : let harness = TenantHarness::create("test_get_vectored").await?;
7065 12 : let (tenant, ctx) = harness.load().await;
7066 12 : let io_concurrency = IoConcurrency::spawn_for_test();
7067 12 : let tline = tenant
7068 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
7069 12 : .await?;
7070 12 :
7071 12 : let lsn = Lsn(0x10);
7072 12 : let inserted = bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
7073 12 :
7074 12 : let guard = tline.layers.read().await;
7075 12 : let lm = guard.layer_map()?;
7076 12 :
7077 12 : lm.dump(true, &ctx).await?;
7078 12 :
7079 12 : let mut reads = Vec::new();
7080 12 : let mut prev = None;
7081 72 : lm.iter_historic_layers().for_each(|desc| {
7082 72 : if !desc.is_delta() {
7083 12 : prev = Some(desc.clone());
7084 12 : return;
7085 60 : }
7086 60 :
7087 60 : let start = desc.key_range.start;
7088 60 : let end = desc
7089 60 : .key_range
7090 60 : .start
7091 60 : .add(Timeline::MAX_GET_VECTORED_KEYS.try_into().unwrap());
7092 60 : reads.push(KeySpace {
7093 60 : ranges: vec![start..end],
7094 60 : });
7095 12 :
7096 60 : if let Some(prev) = &prev {
7097 60 : if !prev.is_delta() {
7098 60 : return;
7099 12 : }
7100 0 :
7101 0 : let first_range = Key {
7102 0 : field6: prev.key_range.end.field6 - 4,
7103 0 : ..prev.key_range.end
7104 0 : }..prev.key_range.end;
7105 0 :
7106 0 : let second_range = desc.key_range.start..Key {
7107 0 : field6: desc.key_range.start.field6 + 4,
7108 0 : ..desc.key_range.start
7109 0 : };
7110 0 :
7111 0 : reads.push(KeySpace {
7112 0 : ranges: vec![first_range, second_range],
7113 0 : });
7114 12 : };
7115 12 :
7116 12 : prev = Some(desc.clone());
7117 72 : });
7118 12 :
7119 12 : drop(guard);
7120 12 :
7121 12 : // Pick a big LSN such that we query over all the changes.
7122 12 : let reads_lsn = Lsn(u64::MAX - 1);
7123 12 :
7124 72 : for read in reads {
7125 60 : info!("Doing vectored read on {:?}", read);
7126 12 :
7127 60 : let query = VersionedKeySpaceQuery::uniform(read.clone(), reads_lsn);
7128 12 :
7129 60 : let vectored_res = tline
7130 60 : .get_vectored_impl(
7131 60 : query,
7132 60 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7133 60 : &ctx,
7134 60 : )
7135 60 : .await;
7136 12 :
7137 60 : let mut expected_lsns: HashMap<Key, Lsn> = Default::default();
7138 60 : let mut expect_missing = false;
7139 60 : let mut key = read.start().unwrap();
7140 1980 : while key != read.end().unwrap() {
7141 1920 : if let Some(lsns) = inserted.get(&key) {
7142 1920 : let expected_lsn = lsns.iter().rfind(|lsn| **lsn <= reads_lsn);
7143 1920 : match expected_lsn {
7144 1920 : Some(lsn) => {
7145 1920 : expected_lsns.insert(key, *lsn);
7146 1920 : }
7147 12 : None => {
7148 12 : expect_missing = true;
7149 0 : break;
7150 12 : }
7151 12 : }
7152 12 : } else {
7153 12 : expect_missing = true;
7154 0 : break;
7155 12 : }
7156 12 :
7157 1920 : key = key.next();
7158 12 : }
7159 12 :
7160 60 : if expect_missing {
7161 12 : assert!(matches!(vectored_res, Err(GetVectoredError::MissingKey(_))));
7162 12 : } else {
7163 1920 : for (key, image) in vectored_res? {
7164 1920 : let expected_lsn = expected_lsns.get(&key).expect("determined above");
7165 1920 : let expected_image = test_img(&format!("{} at {}", key.field6, expected_lsn));
7166 1920 : assert_eq!(image?, expected_image);
7167 12 : }
7168 12 : }
7169 12 : }
7170 12 :
7171 12 : Ok(())
7172 12 : }
7173 :
7174 : #[tokio::test]
7175 12 : async fn test_get_vectored_aux_files() -> anyhow::Result<()> {
7176 12 : let harness = TenantHarness::create("test_get_vectored_aux_files").await?;
7177 12 :
7178 12 : let (tenant, ctx) = harness.load().await;
7179 12 : let io_concurrency = IoConcurrency::spawn_for_test();
7180 12 : let (tline, ctx) = tenant
7181 12 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
7182 12 : .await?;
7183 12 : let tline = tline.raw_timeline().unwrap();
7184 12 :
7185 12 : let mut modification = tline.begin_modification(Lsn(0x1000));
7186 12 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
7187 12 : modification.set_lsn(Lsn(0x1008))?;
7188 12 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
7189 12 : modification.commit(&ctx).await?;
7190 12 :
7191 12 : let child_timeline_id = TimelineId::generate();
7192 12 : tenant
7193 12 : .branch_timeline_test(
7194 12 : tline,
7195 12 : child_timeline_id,
7196 12 : Some(tline.get_last_record_lsn()),
7197 12 : &ctx,
7198 12 : )
7199 12 : .await?;
7200 12 :
7201 12 : let child_timeline = tenant
7202 12 : .get_timeline(child_timeline_id, true)
7203 12 : .expect("Should have the branched timeline");
7204 12 :
7205 12 : let aux_keyspace = KeySpace {
7206 12 : ranges: vec![NON_INHERITED_RANGE],
7207 12 : };
7208 12 : let read_lsn = child_timeline.get_last_record_lsn();
7209 12 :
7210 12 : let query = VersionedKeySpaceQuery::uniform(aux_keyspace.clone(), read_lsn);
7211 12 :
7212 12 : let vectored_res = child_timeline
7213 12 : .get_vectored_impl(
7214 12 : query,
7215 12 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7216 12 : &ctx,
7217 12 : )
7218 12 : .await;
7219 12 :
7220 12 : let images = vectored_res?;
7221 12 : assert!(images.is_empty());
7222 12 : Ok(())
7223 12 : }
7224 :
7225 : // Test that vectored get handles layer gaps correctly
7226 : // by advancing into the next ancestor timeline if required.
7227 : //
7228 : // The test generates timelines that look like the diagram below.
7229 : // We leave a gap in one of the L1 layers at `gap_at_key` (`/` in the diagram).
7230 : // The reconstruct data for that key lies in the ancestor timeline (`X` in the diagram).
7231 : //
7232 : // ```
7233 : //-------------------------------+
7234 : // ... |
7235 : // [ L1 ] |
7236 : // [ / L1 ] | Child Timeline
7237 : // ... |
7238 : // ------------------------------+
7239 : // [ X L1 ] | Parent Timeline
7240 : // ------------------------------+
7241 : // ```
7242 : #[tokio::test]
7243 12 : async fn test_get_vectored_key_gap() -> anyhow::Result<()> {
7244 12 : let tenant_conf = pageserver_api::models::TenantConfig {
7245 12 : // Make compaction deterministic
7246 12 : gc_period: Some(Duration::ZERO),
7247 12 : compaction_period: Some(Duration::ZERO),
7248 12 : // Encourage creation of L1 layers
7249 12 : checkpoint_distance: Some(16 * 1024),
7250 12 : compaction_target_size: Some(8 * 1024),
7251 12 : ..Default::default()
7252 12 : };
7253 12 :
7254 12 : let harness = TenantHarness::create_custom(
7255 12 : "test_get_vectored_key_gap",
7256 12 : tenant_conf,
7257 12 : TenantId::generate(),
7258 12 : ShardIdentity::unsharded(),
7259 12 : Generation::new(0xdeadbeef),
7260 12 : )
7261 12 : .await?;
7262 12 : let (tenant, ctx) = harness.load().await;
7263 12 : let io_concurrency = IoConcurrency::spawn_for_test();
7264 12 :
7265 12 : let mut current_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7266 12 : let gap_at_key = current_key.add(100);
7267 12 : let mut current_lsn = Lsn(0x10);
7268 12 :
7269 12 : const KEY_COUNT: usize = 10_000;
7270 12 :
7271 12 : let timeline_id = TimelineId::generate();
7272 12 : let current_timeline = tenant
7273 12 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
7274 12 : .await?;
7275 12 :
7276 12 : current_lsn += 0x100;
7277 12 :
7278 12 : let mut writer = current_timeline.writer().await;
7279 12 : writer
7280 12 : .put(
7281 12 : gap_at_key,
7282 12 : current_lsn,
7283 12 : &Value::Image(test_img(&format!("{} at {}", gap_at_key, current_lsn))),
7284 12 : &ctx,
7285 12 : )
7286 12 : .await?;
7287 12 : writer.finish_write(current_lsn);
7288 12 : drop(writer);
7289 12 :
7290 12 : let mut latest_lsns = HashMap::new();
7291 12 : latest_lsns.insert(gap_at_key, current_lsn);
7292 12 :
7293 12 : current_timeline.freeze_and_flush().await?;
7294 12 :
7295 12 : let child_timeline_id = TimelineId::generate();
7296 12 :
7297 12 : tenant
7298 12 : .branch_timeline_test(
7299 12 : ¤t_timeline,
7300 12 : child_timeline_id,
7301 12 : Some(current_lsn),
7302 12 : &ctx,
7303 12 : )
7304 12 : .await?;
7305 12 : let child_timeline = tenant
7306 12 : .get_timeline(child_timeline_id, true)
7307 12 : .expect("Should have the branched timeline");
7308 12 :
7309 120012 : for i in 0..KEY_COUNT {
7310 120000 : if current_key == gap_at_key {
7311 12 : current_key = current_key.next();
7312 12 : continue;
7313 119988 : }
7314 119988 :
7315 119988 : current_lsn += 0x10;
7316 12 :
7317 119988 : let mut writer = child_timeline.writer().await;
7318 119988 : writer
7319 119988 : .put(
7320 119988 : current_key,
7321 119988 : current_lsn,
7322 119988 : &Value::Image(test_img(&format!("{} at {}", current_key, current_lsn))),
7323 119988 : &ctx,
7324 119988 : )
7325 119988 : .await?;
7326 119988 : writer.finish_write(current_lsn);
7327 119988 : drop(writer);
7328 119988 :
7329 119988 : latest_lsns.insert(current_key, current_lsn);
7330 119988 : current_key = current_key.next();
7331 119988 :
7332 119988 : // Flush every now and then to encourage layer file creation.
7333 119988 : if i % 500 == 0 {
7334 240 : child_timeline.freeze_and_flush().await?;
7335 119748 : }
7336 12 : }
7337 12 :
7338 12 : child_timeline.freeze_and_flush().await?;
7339 12 : let mut flags = EnumSet::new();
7340 12 : flags.insert(CompactFlags::ForceRepartition);
7341 12 : child_timeline
7342 12 : .compact(&CancellationToken::new(), flags, &ctx)
7343 12 : .await?;
7344 12 :
7345 12 : let key_near_end = {
7346 12 : let mut tmp = current_key;
7347 12 : tmp.field6 -= 10;
7348 12 : tmp
7349 12 : };
7350 12 :
7351 12 : let key_near_gap = {
7352 12 : let mut tmp = gap_at_key;
7353 12 : tmp.field6 -= 10;
7354 12 : tmp
7355 12 : };
7356 12 :
7357 12 : let read = KeySpace {
7358 12 : ranges: vec![key_near_gap..gap_at_key.next(), key_near_end..current_key],
7359 12 : };
7360 12 :
7361 12 : let query = VersionedKeySpaceQuery::uniform(read.clone(), current_lsn);
7362 12 :
7363 12 : let results = child_timeline
7364 12 : .get_vectored_impl(
7365 12 : query,
7366 12 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7367 12 : &ctx,
7368 12 : )
7369 12 : .await?;
7370 12 :
7371 264 : for (key, img_res) in results {
7372 252 : let expected = test_img(&format!("{} at {}", key, latest_lsns[&key]));
7373 252 : assert_eq!(img_res?, expected);
7374 12 : }
7375 12 :
7376 12 : Ok(())
7377 12 : }
7378 :
7379 : // Test that vectored get descends into ancestor timelines correctly and
7380 : // does not return an image that's newer than requested.
7381 : //
7382 : // The diagram below ilustrates an interesting case. We have a parent timeline
7383 : // (top of the Lsn range) and a child timeline. The request key cannot be reconstructed
7384 : // from the child timeline, so the parent timeline must be visited. When advacing into
7385 : // the child timeline, the read path needs to remember what the requested Lsn was in
7386 : // order to avoid returning an image that's too new. The test below constructs such
7387 : // a timeline setup and does a few queries around the Lsn of each page image.
7388 : // ```
7389 : // LSN
7390 : // ^
7391 : // |
7392 : // |
7393 : // 500 | --------------------------------------> branch point
7394 : // 400 | X
7395 : // 300 | X
7396 : // 200 | --------------------------------------> requested lsn
7397 : // 100 | X
7398 : // |---------------------------------------> Key
7399 : // |
7400 : // ------> requested key
7401 : //
7402 : // Legend:
7403 : // * X - page images
7404 : // ```
7405 : #[tokio::test]
7406 12 : async fn test_get_vectored_ancestor_descent() -> anyhow::Result<()> {
7407 12 : let harness = TenantHarness::create("test_get_vectored_on_lsn_axis").await?;
7408 12 : let (tenant, ctx) = harness.load().await;
7409 12 : let io_concurrency = IoConcurrency::spawn_for_test();
7410 12 :
7411 12 : let start_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7412 12 : let end_key = start_key.add(1000);
7413 12 : let child_gap_at_key = start_key.add(500);
7414 12 : let mut parent_gap_lsns: BTreeMap<Lsn, String> = BTreeMap::new();
7415 12 :
7416 12 : let mut current_lsn = Lsn(0x10);
7417 12 :
7418 12 : let timeline_id = TimelineId::generate();
7419 12 : let parent_timeline = tenant
7420 12 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
7421 12 : .await?;
7422 12 :
7423 12 : current_lsn += 0x100;
7424 12 :
7425 48 : for _ in 0..3 {
7426 36 : let mut key = start_key;
7427 36036 : while key < end_key {
7428 36000 : current_lsn += 0x10;
7429 36000 :
7430 36000 : let image_value = format!("{} at {}", child_gap_at_key, current_lsn);
7431 12 :
7432 36000 : let mut writer = parent_timeline.writer().await;
7433 36000 : writer
7434 36000 : .put(
7435 36000 : key,
7436 36000 : current_lsn,
7437 36000 : &Value::Image(test_img(&image_value)),
7438 36000 : &ctx,
7439 36000 : )
7440 36000 : .await?;
7441 36000 : writer.finish_write(current_lsn);
7442 36000 :
7443 36000 : if key == child_gap_at_key {
7444 36 : parent_gap_lsns.insert(current_lsn, image_value);
7445 35964 : }
7446 12 :
7447 36000 : key = key.next();
7448 12 : }
7449 12 :
7450 36 : parent_timeline.freeze_and_flush().await?;
7451 12 : }
7452 12 :
7453 12 : let child_timeline_id = TimelineId::generate();
7454 12 :
7455 12 : let child_timeline = tenant
7456 12 : .branch_timeline_test(&parent_timeline, child_timeline_id, Some(current_lsn), &ctx)
7457 12 : .await?;
7458 12 :
7459 12 : let mut key = start_key;
7460 12012 : while key < end_key {
7461 12000 : if key == child_gap_at_key {
7462 12 : key = key.next();
7463 12 : continue;
7464 11988 : }
7465 11988 :
7466 11988 : current_lsn += 0x10;
7467 12 :
7468 11988 : let mut writer = child_timeline.writer().await;
7469 11988 : writer
7470 11988 : .put(
7471 11988 : key,
7472 11988 : current_lsn,
7473 11988 : &Value::Image(test_img(&format!("{} at {}", key, current_lsn))),
7474 11988 : &ctx,
7475 11988 : )
7476 11988 : .await?;
7477 11988 : writer.finish_write(current_lsn);
7478 11988 :
7479 11988 : key = key.next();
7480 12 : }
7481 12 :
7482 12 : child_timeline.freeze_and_flush().await?;
7483 12 :
7484 12 : let lsn_offsets: [i64; 5] = [-10, -1, 0, 1, 10];
7485 12 : let mut query_lsns = Vec::new();
7486 36 : for image_lsn in parent_gap_lsns.keys().rev() {
7487 216 : for offset in lsn_offsets {
7488 180 : query_lsns.push(Lsn(image_lsn
7489 180 : .0
7490 180 : .checked_add_signed(offset)
7491 180 : .expect("Shouldn't overflow")));
7492 180 : }
7493 12 : }
7494 12 :
7495 192 : for query_lsn in query_lsns {
7496 180 : let query = VersionedKeySpaceQuery::uniform(
7497 180 : KeySpace {
7498 180 : ranges: vec![child_gap_at_key..child_gap_at_key.next()],
7499 180 : },
7500 180 : query_lsn,
7501 180 : );
7502 12 :
7503 180 : let results = child_timeline
7504 180 : .get_vectored_impl(
7505 180 : query,
7506 180 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7507 180 : &ctx,
7508 180 : )
7509 180 : .await;
7510 12 :
7511 180 : let expected_item = parent_gap_lsns
7512 180 : .iter()
7513 180 : .rev()
7514 408 : .find(|(lsn, _)| **lsn <= query_lsn);
7515 180 :
7516 180 : info!(
7517 12 : "Doing vectored read at LSN {}. Expecting image to be: {:?}",
7518 12 : query_lsn, expected_item
7519 12 : );
7520 12 :
7521 180 : match expected_item {
7522 156 : Some((_, img_value)) => {
7523 156 : let key_results = results.expect("No vectored get error expected");
7524 156 : let key_result = &key_results[&child_gap_at_key];
7525 156 : let returned_img = key_result
7526 156 : .as_ref()
7527 156 : .expect("No page reconstruct error expected");
7528 156 :
7529 156 : info!(
7530 12 : "Vectored read at LSN {} returned image {}",
7531 0 : query_lsn,
7532 0 : std::str::from_utf8(returned_img)?
7533 12 : );
7534 156 : assert_eq!(*returned_img, test_img(img_value));
7535 12 : }
7536 12 : None => {
7537 24 : assert!(matches!(results, Err(GetVectoredError::MissingKey(_))));
7538 12 : }
7539 12 : }
7540 12 : }
7541 12 :
7542 12 : Ok(())
7543 12 : }
7544 :
7545 : #[tokio::test]
7546 12 : async fn test_random_updates() -> anyhow::Result<()> {
7547 12 : let names_algorithms = [
7548 12 : ("test_random_updates_legacy", CompactionAlgorithm::Legacy),
7549 12 : ("test_random_updates_tiered", CompactionAlgorithm::Tiered),
7550 12 : ];
7551 36 : for (name, algorithm) in names_algorithms {
7552 24 : test_random_updates_algorithm(name, algorithm).await?;
7553 12 : }
7554 12 : Ok(())
7555 12 : }
7556 :
7557 24 : async fn test_random_updates_algorithm(
7558 24 : name: &'static str,
7559 24 : compaction_algorithm: CompactionAlgorithm,
7560 24 : ) -> anyhow::Result<()> {
7561 24 : let mut harness = TenantHarness::create(name).await?;
7562 24 : harness.tenant_conf.compaction_algorithm = Some(CompactionAlgorithmSettings {
7563 24 : kind: compaction_algorithm,
7564 24 : });
7565 24 : let (tenant, ctx) = harness.load().await;
7566 24 : let tline = tenant
7567 24 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7568 24 : .await?;
7569 :
7570 : const NUM_KEYS: usize = 1000;
7571 24 : let cancel = CancellationToken::new();
7572 24 :
7573 24 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7574 24 : let mut test_key_end = test_key;
7575 24 : test_key_end.field6 = NUM_KEYS as u32;
7576 24 : tline.add_extra_test_dense_keyspace(KeySpace::single(test_key..test_key_end));
7577 24 :
7578 24 : let mut keyspace = KeySpaceAccum::new();
7579 24 :
7580 24 : // Track when each page was last modified. Used to assert that
7581 24 : // a read sees the latest page version.
7582 24 : let mut updated = [Lsn(0); NUM_KEYS];
7583 24 :
7584 24 : let mut lsn = Lsn(0x10);
7585 : #[allow(clippy::needless_range_loop)]
7586 24024 : for blknum in 0..NUM_KEYS {
7587 24000 : lsn = Lsn(lsn.0 + 0x10);
7588 24000 : test_key.field6 = blknum as u32;
7589 24000 : let mut writer = tline.writer().await;
7590 24000 : writer
7591 24000 : .put(
7592 24000 : test_key,
7593 24000 : lsn,
7594 24000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7595 24000 : &ctx,
7596 24000 : )
7597 24000 : .await?;
7598 24000 : writer.finish_write(lsn);
7599 24000 : updated[blknum] = lsn;
7600 24000 : drop(writer);
7601 24000 :
7602 24000 : keyspace.add_key(test_key);
7603 : }
7604 :
7605 1224 : for _ in 0..50 {
7606 1201200 : for _ in 0..NUM_KEYS {
7607 1200000 : lsn = Lsn(lsn.0 + 0x10);
7608 1200000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7609 1200000 : test_key.field6 = blknum as u32;
7610 1200000 : let mut writer = tline.writer().await;
7611 1200000 : writer
7612 1200000 : .put(
7613 1200000 : test_key,
7614 1200000 : lsn,
7615 1200000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7616 1200000 : &ctx,
7617 1200000 : )
7618 1200000 : .await?;
7619 1200000 : writer.finish_write(lsn);
7620 1200000 : drop(writer);
7621 1200000 : updated[blknum] = lsn;
7622 : }
7623 :
7624 : // Read all the blocks
7625 1200000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7626 1200000 : test_key.field6 = blknum as u32;
7627 1200000 : assert_eq!(
7628 1200000 : tline.get(test_key, lsn, &ctx).await?,
7629 1200000 : test_img(&format!("{} at {}", blknum, last_lsn))
7630 : );
7631 : }
7632 :
7633 : // Perform a cycle of flush, and GC
7634 1200 : tline.freeze_and_flush().await?;
7635 1200 : tenant
7636 1200 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7637 1200 : .await?;
7638 : }
7639 :
7640 24 : Ok(())
7641 24 : }
7642 :
7643 : #[tokio::test]
7644 12 : async fn test_traverse_branches() -> anyhow::Result<()> {
7645 12 : let (tenant, ctx) = TenantHarness::create("test_traverse_branches")
7646 12 : .await?
7647 12 : .load()
7648 12 : .await;
7649 12 : let mut tline = tenant
7650 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7651 12 : .await?;
7652 12 :
7653 12 : const NUM_KEYS: usize = 1000;
7654 12 :
7655 12 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7656 12 :
7657 12 : let mut keyspace = KeySpaceAccum::new();
7658 12 :
7659 12 : let cancel = CancellationToken::new();
7660 12 :
7661 12 : // Track when each page was last modified. Used to assert that
7662 12 : // a read sees the latest page version.
7663 12 : let mut updated = [Lsn(0); NUM_KEYS];
7664 12 :
7665 12 : let mut lsn = Lsn(0x10);
7666 12 : #[allow(clippy::needless_range_loop)]
7667 12012 : for blknum in 0..NUM_KEYS {
7668 12000 : lsn = Lsn(lsn.0 + 0x10);
7669 12000 : test_key.field6 = blknum as u32;
7670 12000 : let mut writer = tline.writer().await;
7671 12000 : writer
7672 12000 : .put(
7673 12000 : test_key,
7674 12000 : lsn,
7675 12000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7676 12000 : &ctx,
7677 12000 : )
7678 12000 : .await?;
7679 12000 : writer.finish_write(lsn);
7680 12000 : updated[blknum] = lsn;
7681 12000 : drop(writer);
7682 12000 :
7683 12000 : keyspace.add_key(test_key);
7684 12 : }
7685 12 :
7686 612 : for _ in 0..50 {
7687 600 : let new_tline_id = TimelineId::generate();
7688 600 : tenant
7689 600 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7690 600 : .await?;
7691 600 : tline = tenant
7692 600 : .get_timeline(new_tline_id, true)
7693 600 : .expect("Should have the branched timeline");
7694 12 :
7695 600600 : for _ in 0..NUM_KEYS {
7696 600000 : lsn = Lsn(lsn.0 + 0x10);
7697 600000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7698 600000 : test_key.field6 = blknum as u32;
7699 600000 : let mut writer = tline.writer().await;
7700 600000 : writer
7701 600000 : .put(
7702 600000 : test_key,
7703 600000 : lsn,
7704 600000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7705 600000 : &ctx,
7706 600000 : )
7707 600000 : .await?;
7708 600000 : println!("updating {} at {}", blknum, lsn);
7709 600000 : writer.finish_write(lsn);
7710 600000 : drop(writer);
7711 600000 : updated[blknum] = lsn;
7712 12 : }
7713 12 :
7714 12 : // Read all the blocks
7715 600000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7716 600000 : test_key.field6 = blknum as u32;
7717 600000 : assert_eq!(
7718 600000 : tline.get(test_key, lsn, &ctx).await?,
7719 600000 : test_img(&format!("{} at {}", blknum, last_lsn))
7720 12 : );
7721 12 : }
7722 12 :
7723 12 : // Perform a cycle of flush, compact, and GC
7724 600 : tline.freeze_and_flush().await?;
7725 600 : tline.compact(&cancel, EnumSet::default(), &ctx).await?;
7726 600 : tenant
7727 600 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7728 600 : .await?;
7729 12 : }
7730 12 :
7731 12 : Ok(())
7732 12 : }
7733 :
7734 : #[tokio::test]
7735 12 : async fn test_traverse_ancestors() -> anyhow::Result<()> {
7736 12 : let (tenant, ctx) = TenantHarness::create("test_traverse_ancestors")
7737 12 : .await?
7738 12 : .load()
7739 12 : .await;
7740 12 : let mut tline = tenant
7741 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7742 12 : .await?;
7743 12 :
7744 12 : const NUM_KEYS: usize = 100;
7745 12 : const NUM_TLINES: usize = 50;
7746 12 :
7747 12 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7748 12 : // Track page mutation lsns across different timelines.
7749 12 : let mut updated = [[Lsn(0); NUM_KEYS]; NUM_TLINES];
7750 12 :
7751 12 : let mut lsn = Lsn(0x10);
7752 12 :
7753 12 : #[allow(clippy::needless_range_loop)]
7754 612 : for idx in 0..NUM_TLINES {
7755 600 : let new_tline_id = TimelineId::generate();
7756 600 : tenant
7757 600 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7758 600 : .await?;
7759 600 : tline = tenant
7760 600 : .get_timeline(new_tline_id, true)
7761 600 : .expect("Should have the branched timeline");
7762 12 :
7763 60600 : for _ in 0..NUM_KEYS {
7764 60000 : lsn = Lsn(lsn.0 + 0x10);
7765 60000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7766 60000 : test_key.field6 = blknum as u32;
7767 60000 : let mut writer = tline.writer().await;
7768 60000 : writer
7769 60000 : .put(
7770 60000 : test_key,
7771 60000 : lsn,
7772 60000 : &Value::Image(test_img(&format!("{} {} at {}", idx, blknum, lsn))),
7773 60000 : &ctx,
7774 60000 : )
7775 60000 : .await?;
7776 60000 : println!("updating [{}][{}] at {}", idx, blknum, lsn);
7777 60000 : writer.finish_write(lsn);
7778 60000 : drop(writer);
7779 60000 : updated[idx][blknum] = lsn;
7780 12 : }
7781 12 : }
7782 12 :
7783 12 : // Read pages from leaf timeline across all ancestors.
7784 600 : for (idx, lsns) in updated.iter().enumerate() {
7785 60000 : for (blknum, lsn) in lsns.iter().enumerate() {
7786 12 : // Skip empty mutations.
7787 60000 : if lsn.0 == 0 {
7788 21989 : continue;
7789 38011 : }
7790 38011 : println!("checking [{idx}][{blknum}] at {lsn}");
7791 38011 : test_key.field6 = blknum as u32;
7792 38011 : assert_eq!(
7793 38011 : tline.get(test_key, *lsn, &ctx).await?,
7794 38011 : test_img(&format!("{idx} {blknum} at {lsn}"))
7795 12 : );
7796 12 : }
7797 12 : }
7798 12 : Ok(())
7799 12 : }
7800 :
7801 : #[tokio::test]
7802 12 : async fn test_write_at_initdb_lsn_takes_optimization_code_path() -> anyhow::Result<()> {
7803 12 : let (tenant, ctx) = TenantHarness::create("test_empty_test_timeline_is_usable")
7804 12 : .await?
7805 12 : .load()
7806 12 : .await;
7807 12 :
7808 12 : let initdb_lsn = Lsn(0x20);
7809 12 : let (utline, ctx) = tenant
7810 12 : .create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)
7811 12 : .await?;
7812 12 : let tline = utline.raw_timeline().unwrap();
7813 12 :
7814 12 : // Spawn flush loop now so that we can set the `expect_initdb_optimization`
7815 12 : tline.maybe_spawn_flush_loop();
7816 12 :
7817 12 : // Make sure the timeline has the minimum set of required keys for operation.
7818 12 : // The only operation you can always do on an empty timeline is to `put` new data.
7819 12 : // Except if you `put` at `initdb_lsn`.
7820 12 : // In that case, there's an optimization to directly create image layers instead of delta layers.
7821 12 : // It uses `repartition()`, which assumes some keys to be present.
7822 12 : // Let's make sure the test timeline can handle that case.
7823 12 : {
7824 12 : let mut state = tline.flush_loop_state.lock().unwrap();
7825 12 : assert_eq!(
7826 12 : timeline::FlushLoopState::Running {
7827 12 : expect_initdb_optimization: false,
7828 12 : initdb_optimization_count: 0,
7829 12 : },
7830 12 : *state
7831 12 : );
7832 12 : *state = timeline::FlushLoopState::Running {
7833 12 : expect_initdb_optimization: true,
7834 12 : initdb_optimization_count: 0,
7835 12 : };
7836 12 : }
7837 12 :
7838 12 : // Make writes at the initdb_lsn. When we flush it below, it should be handled by the optimization.
7839 12 : // As explained above, the optimization requires some keys to be present.
7840 12 : // As per `create_empty_timeline` documentation, use init_empty to set them.
7841 12 : // This is what `create_test_timeline` does, by the way.
7842 12 : let mut modification = tline.begin_modification(initdb_lsn);
7843 12 : modification
7844 12 : .init_empty_test_timeline()
7845 12 : .context("init_empty_test_timeline")?;
7846 12 : modification
7847 12 : .commit(&ctx)
7848 12 : .await
7849 12 : .context("commit init_empty_test_timeline modification")?;
7850 12 :
7851 12 : // Do the flush. The flush code will check the expectations that we set above.
7852 12 : tline.freeze_and_flush().await?;
7853 12 :
7854 12 : // assert freeze_and_flush exercised the initdb optimization
7855 12 : {
7856 12 : let state = tline.flush_loop_state.lock().unwrap();
7857 12 : let timeline::FlushLoopState::Running {
7858 12 : expect_initdb_optimization,
7859 12 : initdb_optimization_count,
7860 12 : } = *state
7861 12 : else {
7862 12 : panic!("unexpected state: {:?}", *state);
7863 12 : };
7864 12 : assert!(expect_initdb_optimization);
7865 12 : assert!(initdb_optimization_count > 0);
7866 12 : }
7867 12 : Ok(())
7868 12 : }
7869 :
7870 : #[tokio::test]
7871 12 : async fn test_create_guard_crash() -> anyhow::Result<()> {
7872 12 : let name = "test_create_guard_crash";
7873 12 : let harness = TenantHarness::create(name).await?;
7874 12 : {
7875 12 : let (tenant, ctx) = harness.load().await;
7876 12 : let (tline, _ctx) = tenant
7877 12 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
7878 12 : .await?;
7879 12 : // Leave the timeline ID in [`TenantShard::timelines_creating`] to exclude attempting to create it again
7880 12 : let raw_tline = tline.raw_timeline().unwrap();
7881 12 : raw_tline
7882 12 : .shutdown(super::timeline::ShutdownMode::Hard)
7883 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))
7884 12 : .await;
7885 12 : std::mem::forget(tline);
7886 12 : }
7887 12 :
7888 12 : let (tenant, _) = harness.load().await;
7889 12 : match tenant.get_timeline(TIMELINE_ID, false) {
7890 12 : Ok(_) => panic!("timeline should've been removed during load"),
7891 12 : Err(e) => {
7892 12 : assert_eq!(
7893 12 : e,
7894 12 : GetTimelineError::NotFound {
7895 12 : tenant_id: tenant.tenant_shard_id,
7896 12 : timeline_id: TIMELINE_ID,
7897 12 : }
7898 12 : )
7899 12 : }
7900 12 : }
7901 12 :
7902 12 : assert!(
7903 12 : !harness
7904 12 : .conf
7905 12 : .timeline_path(&tenant.tenant_shard_id, &TIMELINE_ID)
7906 12 : .exists()
7907 12 : );
7908 12 :
7909 12 : Ok(())
7910 12 : }
7911 :
7912 : #[tokio::test]
7913 12 : async fn test_read_at_max_lsn() -> anyhow::Result<()> {
7914 12 : let names_algorithms = [
7915 12 : ("test_read_at_max_lsn_legacy", CompactionAlgorithm::Legacy),
7916 12 : ("test_read_at_max_lsn_tiered", CompactionAlgorithm::Tiered),
7917 12 : ];
7918 36 : for (name, algorithm) in names_algorithms {
7919 24 : test_read_at_max_lsn_algorithm(name, algorithm).await?;
7920 12 : }
7921 12 : Ok(())
7922 12 : }
7923 :
7924 24 : async fn test_read_at_max_lsn_algorithm(
7925 24 : name: &'static str,
7926 24 : compaction_algorithm: CompactionAlgorithm,
7927 24 : ) -> anyhow::Result<()> {
7928 24 : let mut harness = TenantHarness::create(name).await?;
7929 24 : harness.tenant_conf.compaction_algorithm = Some(CompactionAlgorithmSettings {
7930 24 : kind: compaction_algorithm,
7931 24 : });
7932 24 : let (tenant, ctx) = harness.load().await;
7933 24 : let tline = tenant
7934 24 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
7935 24 : .await?;
7936 :
7937 24 : let lsn = Lsn(0x10);
7938 24 : let compact = false;
7939 24 : bulk_insert_maybe_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000, compact).await?;
7940 :
7941 24 : let test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7942 24 : let read_lsn = Lsn(u64::MAX - 1);
7943 :
7944 24 : let result = tline.get(test_key, read_lsn, &ctx).await;
7945 24 : assert!(result.is_ok(), "result is not Ok: {}", result.unwrap_err());
7946 :
7947 24 : Ok(())
7948 24 : }
7949 :
7950 : #[tokio::test]
7951 12 : async fn test_metadata_scan() -> anyhow::Result<()> {
7952 12 : let harness = TenantHarness::create("test_metadata_scan").await?;
7953 12 : let (tenant, ctx) = harness.load().await;
7954 12 : let io_concurrency = IoConcurrency::spawn_for_test();
7955 12 : let tline = tenant
7956 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7957 12 : .await?;
7958 12 :
7959 12 : const NUM_KEYS: usize = 1000;
7960 12 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
7961 12 :
7962 12 : let cancel = CancellationToken::new();
7963 12 :
7964 12 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7965 12 : base_key.field1 = AUX_KEY_PREFIX;
7966 12 : let mut test_key = base_key;
7967 12 :
7968 12 : // Track when each page was last modified. Used to assert that
7969 12 : // a read sees the latest page version.
7970 12 : let mut updated = [Lsn(0); NUM_KEYS];
7971 12 :
7972 12 : let mut lsn = Lsn(0x10);
7973 12 : #[allow(clippy::needless_range_loop)]
7974 12012 : for blknum in 0..NUM_KEYS {
7975 12000 : lsn = Lsn(lsn.0 + 0x10);
7976 12000 : test_key.field6 = (blknum * STEP) as u32;
7977 12000 : let mut writer = tline.writer().await;
7978 12000 : writer
7979 12000 : .put(
7980 12000 : test_key,
7981 12000 : lsn,
7982 12000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7983 12000 : &ctx,
7984 12000 : )
7985 12000 : .await?;
7986 12000 : writer.finish_write(lsn);
7987 12000 : updated[blknum] = lsn;
7988 12000 : drop(writer);
7989 12 : }
7990 12 :
7991 12 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
7992 12 :
7993 144 : for iter in 0..=10 {
7994 12 : // Read all the blocks
7995 132000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7996 132000 : test_key.field6 = (blknum * STEP) as u32;
7997 132000 : assert_eq!(
7998 132000 : tline.get(test_key, lsn, &ctx).await?,
7999 132000 : test_img(&format!("{} at {}", blknum, last_lsn))
8000 12 : );
8001 12 : }
8002 12 :
8003 132 : let mut cnt = 0;
8004 132 : let query = VersionedKeySpaceQuery::uniform(keyspace.clone(), lsn);
8005 12 :
8006 132000 : for (key, value) in tline
8007 132 : .get_vectored_impl(
8008 132 : query,
8009 132 : &mut ValuesReconstructState::new(io_concurrency.clone()),
8010 132 : &ctx,
8011 132 : )
8012 132 : .await?
8013 12 : {
8014 132000 : let blknum = key.field6 as usize;
8015 132000 : let value = value?;
8016 132000 : assert!(blknum % STEP == 0);
8017 132000 : let blknum = blknum / STEP;
8018 132000 : assert_eq!(
8019 132000 : value,
8020 132000 : test_img(&format!("{} at {}", blknum, updated[blknum]))
8021 132000 : );
8022 132000 : cnt += 1;
8023 12 : }
8024 12 :
8025 132 : assert_eq!(cnt, NUM_KEYS);
8026 12 :
8027 132132 : for _ in 0..NUM_KEYS {
8028 132000 : lsn = Lsn(lsn.0 + 0x10);
8029 132000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
8030 132000 : test_key.field6 = (blknum * STEP) as u32;
8031 132000 : let mut writer = tline.writer().await;
8032 132000 : writer
8033 132000 : .put(
8034 132000 : test_key,
8035 132000 : lsn,
8036 132000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
8037 132000 : &ctx,
8038 132000 : )
8039 132000 : .await?;
8040 132000 : writer.finish_write(lsn);
8041 132000 : drop(writer);
8042 132000 : updated[blknum] = lsn;
8043 12 : }
8044 12 :
8045 12 : // Perform two cycles of flush, compact, and GC
8046 396 : for round in 0..2 {
8047 264 : tline.freeze_and_flush().await?;
8048 264 : tline
8049 264 : .compact(
8050 264 : &cancel,
8051 264 : if iter % 5 == 0 && round == 0 {
8052 36 : let mut flags = EnumSet::new();
8053 36 : flags.insert(CompactFlags::ForceImageLayerCreation);
8054 36 : flags.insert(CompactFlags::ForceRepartition);
8055 36 : flags
8056 12 : } else {
8057 228 : EnumSet::empty()
8058 12 : },
8059 264 : &ctx,
8060 264 : )
8061 264 : .await?;
8062 264 : tenant
8063 264 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
8064 264 : .await?;
8065 12 : }
8066 12 : }
8067 12 :
8068 12 : Ok(())
8069 12 : }
8070 :
8071 : #[tokio::test]
8072 12 : async fn test_metadata_compaction_trigger() -> anyhow::Result<()> {
8073 12 : let harness = TenantHarness::create("test_metadata_compaction_trigger").await?;
8074 12 : let (tenant, ctx) = harness.load().await;
8075 12 : let tline = tenant
8076 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
8077 12 : .await?;
8078 12 :
8079 12 : let cancel = CancellationToken::new();
8080 12 :
8081 12 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
8082 12 : base_key.field1 = AUX_KEY_PREFIX;
8083 12 : let test_key = base_key;
8084 12 : let mut lsn = Lsn(0x10);
8085 12 :
8086 252 : for _ in 0..20 {
8087 240 : lsn = Lsn(lsn.0 + 0x10);
8088 240 : let mut writer = tline.writer().await;
8089 240 : writer
8090 240 : .put(
8091 240 : test_key,
8092 240 : lsn,
8093 240 : &Value::Image(test_img(&format!("{} at {}", 0, lsn))),
8094 240 : &ctx,
8095 240 : )
8096 240 : .await?;
8097 240 : writer.finish_write(lsn);
8098 240 : drop(writer);
8099 240 : tline.freeze_and_flush().await?; // force create a delta layer
8100 12 : }
8101 12 :
8102 12 : let before_num_l0_delta_files =
8103 12 : tline.layers.read().await.layer_map()?.level0_deltas().len();
8104 12 :
8105 12 : tline.compact(&cancel, EnumSet::default(), &ctx).await?;
8106 12 :
8107 12 : let after_num_l0_delta_files = tline.layers.read().await.layer_map()?.level0_deltas().len();
8108 12 :
8109 12 : assert!(
8110 12 : after_num_l0_delta_files < before_num_l0_delta_files,
8111 12 : "after_num_l0_delta_files={after_num_l0_delta_files}, before_num_l0_delta_files={before_num_l0_delta_files}"
8112 12 : );
8113 12 :
8114 12 : assert_eq!(
8115 12 : tline.get(test_key, lsn, &ctx).await?,
8116 12 : test_img(&format!("{} at {}", 0, lsn))
8117 12 : );
8118 12 :
8119 12 : Ok(())
8120 12 : }
8121 :
8122 : #[tokio::test]
8123 12 : async fn test_aux_file_e2e() {
8124 12 : let harness = TenantHarness::create("test_aux_file_e2e").await.unwrap();
8125 12 :
8126 12 : let (tenant, ctx) = harness.load().await;
8127 12 : let io_concurrency = IoConcurrency::spawn_for_test();
8128 12 :
8129 12 : let mut lsn = Lsn(0x08);
8130 12 :
8131 12 : let tline: Arc<Timeline> = tenant
8132 12 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
8133 12 : .await
8134 12 : .unwrap();
8135 12 :
8136 12 : {
8137 12 : lsn += 8;
8138 12 : let mut modification = tline.begin_modification(lsn);
8139 12 : modification
8140 12 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
8141 12 : .await
8142 12 : .unwrap();
8143 12 : modification.commit(&ctx).await.unwrap();
8144 12 : }
8145 12 :
8146 12 : // we can read everything from the storage
8147 12 : let files = tline
8148 12 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
8149 12 : .await
8150 12 : .unwrap();
8151 12 : assert_eq!(
8152 12 : files.get("pg_logical/mappings/test1"),
8153 12 : Some(&bytes::Bytes::from_static(b"first"))
8154 12 : );
8155 12 :
8156 12 : {
8157 12 : lsn += 8;
8158 12 : let mut modification = tline.begin_modification(lsn);
8159 12 : modification
8160 12 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
8161 12 : .await
8162 12 : .unwrap();
8163 12 : modification.commit(&ctx).await.unwrap();
8164 12 : }
8165 12 :
8166 12 : let files = tline
8167 12 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
8168 12 : .await
8169 12 : .unwrap();
8170 12 : assert_eq!(
8171 12 : files.get("pg_logical/mappings/test2"),
8172 12 : Some(&bytes::Bytes::from_static(b"second"))
8173 12 : );
8174 12 :
8175 12 : let child = tenant
8176 12 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(lsn), &ctx)
8177 12 : .await
8178 12 : .unwrap();
8179 12 :
8180 12 : let files = child
8181 12 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
8182 12 : .await
8183 12 : .unwrap();
8184 12 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
8185 12 : assert_eq!(files.get("pg_logical/mappings/test2"), None);
8186 12 : }
8187 :
8188 : #[tokio::test]
8189 12 : async fn test_repl_origin_tombstones() {
8190 12 : let harness = TenantHarness::create("test_repl_origin_tombstones")
8191 12 : .await
8192 12 : .unwrap();
8193 12 :
8194 12 : let (tenant, ctx) = harness.load().await;
8195 12 : let io_concurrency = IoConcurrency::spawn_for_test();
8196 12 :
8197 12 : let mut lsn = Lsn(0x08);
8198 12 :
8199 12 : let tline: Arc<Timeline> = tenant
8200 12 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
8201 12 : .await
8202 12 : .unwrap();
8203 12 :
8204 12 : let repl_lsn = Lsn(0x10);
8205 12 : {
8206 12 : lsn += 8;
8207 12 : let mut modification = tline.begin_modification(lsn);
8208 12 : modification.put_for_unit_test(repl_origin_key(2), Value::Image(Bytes::new()));
8209 12 : modification.set_replorigin(1, repl_lsn).await.unwrap();
8210 12 : modification.commit(&ctx).await.unwrap();
8211 12 : }
8212 12 :
8213 12 : // we can read everything from the storage
8214 12 : let repl_origins = tline
8215 12 : .get_replorigins(lsn, &ctx, io_concurrency.clone())
8216 12 : .await
8217 12 : .unwrap();
8218 12 : assert_eq!(repl_origins.len(), 1);
8219 12 : assert_eq!(repl_origins[&1], lsn);
8220 12 :
8221 12 : {
8222 12 : lsn += 8;
8223 12 : let mut modification = tline.begin_modification(lsn);
8224 12 : modification.put_for_unit_test(
8225 12 : repl_origin_key(3),
8226 12 : Value::Image(Bytes::copy_from_slice(b"cannot_decode_this")),
8227 12 : );
8228 12 : modification.commit(&ctx).await.unwrap();
8229 12 : }
8230 12 : let result = tline
8231 12 : .get_replorigins(lsn, &ctx, io_concurrency.clone())
8232 12 : .await;
8233 12 : assert!(result.is_err());
8234 12 : }
8235 :
8236 : #[tokio::test]
8237 12 : async fn test_metadata_image_creation() -> anyhow::Result<()> {
8238 12 : let harness = TenantHarness::create("test_metadata_image_creation").await?;
8239 12 : let (tenant, ctx) = harness.load().await;
8240 12 : let io_concurrency = IoConcurrency::spawn_for_test();
8241 12 : let tline = tenant
8242 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
8243 12 : .await?;
8244 12 :
8245 12 : const NUM_KEYS: usize = 1000;
8246 12 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
8247 12 :
8248 12 : let cancel = CancellationToken::new();
8249 12 :
8250 12 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8251 12 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
8252 12 : let mut test_key = base_key;
8253 12 : let mut lsn = Lsn(0x10);
8254 12 :
8255 48 : async fn scan_with_statistics(
8256 48 : tline: &Timeline,
8257 48 : keyspace: &KeySpace,
8258 48 : lsn: Lsn,
8259 48 : ctx: &RequestContext,
8260 48 : io_concurrency: IoConcurrency,
8261 48 : ) -> anyhow::Result<(BTreeMap<Key, Result<Bytes, PageReconstructError>>, usize)> {
8262 48 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
8263 48 : let query = VersionedKeySpaceQuery::uniform(keyspace.clone(), lsn);
8264 48 : let res = tline
8265 48 : .get_vectored_impl(query, &mut reconstruct_state, ctx)
8266 48 : .await?;
8267 48 : Ok((res, reconstruct_state.get_delta_layers_visited() as usize))
8268 48 : }
8269 12 :
8270 12012 : for blknum in 0..NUM_KEYS {
8271 12000 : lsn = Lsn(lsn.0 + 0x10);
8272 12000 : test_key.field6 = (blknum * STEP) as u32;
8273 12000 : let mut writer = tline.writer().await;
8274 12000 : writer
8275 12000 : .put(
8276 12000 : test_key,
8277 12000 : lsn,
8278 12000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
8279 12000 : &ctx,
8280 12000 : )
8281 12000 : .await?;
8282 12000 : writer.finish_write(lsn);
8283 12000 : drop(writer);
8284 12 : }
8285 12 :
8286 12 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
8287 12 :
8288 132 : for iter in 1..=10 {
8289 120120 : for _ in 0..NUM_KEYS {
8290 120000 : lsn = Lsn(lsn.0 + 0x10);
8291 120000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
8292 120000 : test_key.field6 = (blknum * STEP) as u32;
8293 120000 : let mut writer = tline.writer().await;
8294 120000 : writer
8295 120000 : .put(
8296 120000 : test_key,
8297 120000 : lsn,
8298 120000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
8299 120000 : &ctx,
8300 120000 : )
8301 120000 : .await?;
8302 120000 : writer.finish_write(lsn);
8303 120000 : drop(writer);
8304 12 : }
8305 12 :
8306 120 : tline.freeze_and_flush().await?;
8307 12 :
8308 120 : if iter % 5 == 0 {
8309 24 : let (_, before_delta_file_accessed) =
8310 24 : scan_with_statistics(&tline, &keyspace, lsn, &ctx, io_concurrency.clone())
8311 24 : .await?;
8312 24 : tline
8313 24 : .compact(
8314 24 : &cancel,
8315 24 : {
8316 24 : let mut flags = EnumSet::new();
8317 24 : flags.insert(CompactFlags::ForceImageLayerCreation);
8318 24 : flags.insert(CompactFlags::ForceRepartition);
8319 24 : flags
8320 24 : },
8321 24 : &ctx,
8322 24 : )
8323 24 : .await?;
8324 24 : let (_, after_delta_file_accessed) =
8325 24 : scan_with_statistics(&tline, &keyspace, lsn, &ctx, io_concurrency.clone())
8326 24 : .await?;
8327 24 : assert!(
8328 24 : after_delta_file_accessed < before_delta_file_accessed,
8329 12 : "after_delta_file_accessed={after_delta_file_accessed}, before_delta_file_accessed={before_delta_file_accessed}"
8330 12 : );
8331 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.
8332 24 : assert!(
8333 24 : after_delta_file_accessed <= 2,
8334 12 : "after_delta_file_accessed={after_delta_file_accessed}"
8335 12 : );
8336 96 : }
8337 12 : }
8338 12 :
8339 12 : Ok(())
8340 12 : }
8341 :
8342 : #[tokio::test]
8343 12 : async fn test_vectored_missing_data_key_reads() -> anyhow::Result<()> {
8344 12 : let harness = TenantHarness::create("test_vectored_missing_data_key_reads").await?;
8345 12 : let (tenant, ctx) = harness.load().await;
8346 12 :
8347 12 : let base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
8348 12 : let base_key_child = Key::from_hex("000000000033333333444444445500000001").unwrap();
8349 12 : let base_key_nonexist = Key::from_hex("000000000033333333444444445500000002").unwrap();
8350 12 :
8351 12 : let tline = tenant
8352 12 : .create_test_timeline_with_layers(
8353 12 : TIMELINE_ID,
8354 12 : Lsn(0x10),
8355 12 : DEFAULT_PG_VERSION,
8356 12 : &ctx,
8357 12 : Vec::new(), // in-memory layers
8358 12 : Vec::new(), // delta layers
8359 12 : vec![(Lsn(0x20), vec![(base_key, test_img("data key 1"))])], // image layers
8360 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
8361 12 : )
8362 12 : .await?;
8363 12 : tline.add_extra_test_dense_keyspace(KeySpace::single(base_key..(base_key_nonexist.next())));
8364 12 :
8365 12 : let child = tenant
8366 12 : .branch_timeline_test_with_layers(
8367 12 : &tline,
8368 12 : NEW_TIMELINE_ID,
8369 12 : Some(Lsn(0x20)),
8370 12 : &ctx,
8371 12 : Vec::new(), // delta layers
8372 12 : vec![(Lsn(0x30), vec![(base_key_child, test_img("data key 2"))])], // image layers
8373 12 : Lsn(0x30),
8374 12 : )
8375 12 : .await
8376 12 : .unwrap();
8377 12 :
8378 12 : let lsn = Lsn(0x30);
8379 12 :
8380 12 : // test vectored get on parent timeline
8381 12 : assert_eq!(
8382 12 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
8383 12 : Some(test_img("data key 1"))
8384 12 : );
8385 12 : assert!(
8386 12 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx)
8387 12 : .await
8388 12 : .unwrap_err()
8389 12 : .is_missing_key_error()
8390 12 : );
8391 12 : assert!(
8392 12 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx)
8393 12 : .await
8394 12 : .unwrap_err()
8395 12 : .is_missing_key_error()
8396 12 : );
8397 12 :
8398 12 : // test vectored get on child timeline
8399 12 : assert_eq!(
8400 12 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
8401 12 : Some(test_img("data key 1"))
8402 12 : );
8403 12 : assert_eq!(
8404 12 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
8405 12 : Some(test_img("data key 2"))
8406 12 : );
8407 12 : assert!(
8408 12 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx)
8409 12 : .await
8410 12 : .unwrap_err()
8411 12 : .is_missing_key_error()
8412 12 : );
8413 12 :
8414 12 : Ok(())
8415 12 : }
8416 :
8417 : #[tokio::test]
8418 12 : async fn test_vectored_missing_metadata_key_reads() -> anyhow::Result<()> {
8419 12 : let harness = TenantHarness::create("test_vectored_missing_metadata_key_reads").await?;
8420 12 : let (tenant, ctx) = harness.load().await;
8421 12 : let io_concurrency = IoConcurrency::spawn_for_test();
8422 12 :
8423 12 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8424 12 : let base_key_child = Key::from_hex("620000000033333333444444445500000001").unwrap();
8425 12 : let base_key_nonexist = Key::from_hex("620000000033333333444444445500000002").unwrap();
8426 12 : let base_key_overwrite = Key::from_hex("620000000033333333444444445500000003").unwrap();
8427 12 :
8428 12 : let base_inherited_key = Key::from_hex("610000000033333333444444445500000000").unwrap();
8429 12 : let base_inherited_key_child =
8430 12 : Key::from_hex("610000000033333333444444445500000001").unwrap();
8431 12 : let base_inherited_key_nonexist =
8432 12 : Key::from_hex("610000000033333333444444445500000002").unwrap();
8433 12 : let base_inherited_key_overwrite =
8434 12 : Key::from_hex("610000000033333333444444445500000003").unwrap();
8435 12 :
8436 12 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
8437 12 : assert_eq!(base_inherited_key.field1, RELATION_SIZE_PREFIX);
8438 12 :
8439 12 : let tline = tenant
8440 12 : .create_test_timeline_with_layers(
8441 12 : TIMELINE_ID,
8442 12 : Lsn(0x10),
8443 12 : DEFAULT_PG_VERSION,
8444 12 : &ctx,
8445 12 : Vec::new(), // in-memory layers
8446 12 : Vec::new(), // delta layers
8447 12 : vec![(
8448 12 : Lsn(0x20),
8449 12 : vec![
8450 12 : (base_inherited_key, test_img("metadata inherited key 1")),
8451 12 : (
8452 12 : base_inherited_key_overwrite,
8453 12 : test_img("metadata key overwrite 1a"),
8454 12 : ),
8455 12 : (base_key, test_img("metadata key 1")),
8456 12 : (base_key_overwrite, test_img("metadata key overwrite 1b")),
8457 12 : ],
8458 12 : )], // image layers
8459 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
8460 12 : )
8461 12 : .await?;
8462 12 :
8463 12 : let child = tenant
8464 12 : .branch_timeline_test_with_layers(
8465 12 : &tline,
8466 12 : NEW_TIMELINE_ID,
8467 12 : Some(Lsn(0x20)),
8468 12 : &ctx,
8469 12 : Vec::new(), // delta layers
8470 12 : vec![(
8471 12 : Lsn(0x30),
8472 12 : vec![
8473 12 : (
8474 12 : base_inherited_key_child,
8475 12 : test_img("metadata inherited key 2"),
8476 12 : ),
8477 12 : (
8478 12 : base_inherited_key_overwrite,
8479 12 : test_img("metadata key overwrite 2a"),
8480 12 : ),
8481 12 : (base_key_child, test_img("metadata key 2")),
8482 12 : (base_key_overwrite, test_img("metadata key overwrite 2b")),
8483 12 : ],
8484 12 : )], // image layers
8485 12 : Lsn(0x30),
8486 12 : )
8487 12 : .await
8488 12 : .unwrap();
8489 12 :
8490 12 : let lsn = Lsn(0x30);
8491 12 :
8492 12 : // test vectored get on parent timeline
8493 12 : assert_eq!(
8494 12 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
8495 12 : Some(test_img("metadata key 1"))
8496 12 : );
8497 12 : assert_eq!(
8498 12 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx).await?,
8499 12 : None
8500 12 : );
8501 12 : assert_eq!(
8502 12 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx).await?,
8503 12 : None
8504 12 : );
8505 12 : assert_eq!(
8506 12 : get_vectored_impl_wrapper(&tline, base_key_overwrite, lsn, &ctx).await?,
8507 12 : Some(test_img("metadata key overwrite 1b"))
8508 12 : );
8509 12 : assert_eq!(
8510 12 : get_vectored_impl_wrapper(&tline, base_inherited_key, lsn, &ctx).await?,
8511 12 : Some(test_img("metadata inherited key 1"))
8512 12 : );
8513 12 : assert_eq!(
8514 12 : get_vectored_impl_wrapper(&tline, base_inherited_key_child, lsn, &ctx).await?,
8515 12 : None
8516 12 : );
8517 12 : assert_eq!(
8518 12 : get_vectored_impl_wrapper(&tline, base_inherited_key_nonexist, lsn, &ctx).await?,
8519 12 : None
8520 12 : );
8521 12 : assert_eq!(
8522 12 : get_vectored_impl_wrapper(&tline, base_inherited_key_overwrite, lsn, &ctx).await?,
8523 12 : Some(test_img("metadata key overwrite 1a"))
8524 12 : );
8525 12 :
8526 12 : // test vectored get on child timeline
8527 12 : assert_eq!(
8528 12 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
8529 12 : None
8530 12 : );
8531 12 : assert_eq!(
8532 12 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
8533 12 : Some(test_img("metadata key 2"))
8534 12 : );
8535 12 : assert_eq!(
8536 12 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx).await?,
8537 12 : None
8538 12 : );
8539 12 : assert_eq!(
8540 12 : get_vectored_impl_wrapper(&child, base_inherited_key, lsn, &ctx).await?,
8541 12 : Some(test_img("metadata inherited key 1"))
8542 12 : );
8543 12 : assert_eq!(
8544 12 : get_vectored_impl_wrapper(&child, base_inherited_key_child, lsn, &ctx).await?,
8545 12 : Some(test_img("metadata inherited key 2"))
8546 12 : );
8547 12 : assert_eq!(
8548 12 : get_vectored_impl_wrapper(&child, base_inherited_key_nonexist, lsn, &ctx).await?,
8549 12 : None
8550 12 : );
8551 12 : assert_eq!(
8552 12 : get_vectored_impl_wrapper(&child, base_key_overwrite, lsn, &ctx).await?,
8553 12 : Some(test_img("metadata key overwrite 2b"))
8554 12 : );
8555 12 : assert_eq!(
8556 12 : get_vectored_impl_wrapper(&child, base_inherited_key_overwrite, lsn, &ctx).await?,
8557 12 : Some(test_img("metadata key overwrite 2a"))
8558 12 : );
8559 12 :
8560 12 : // test vectored scan on parent timeline
8561 12 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency.clone());
8562 12 : let query =
8563 12 : VersionedKeySpaceQuery::uniform(KeySpace::single(Key::metadata_key_range()), lsn);
8564 12 : let res = tline
8565 12 : .get_vectored_impl(query, &mut reconstruct_state, &ctx)
8566 12 : .await?;
8567 12 :
8568 12 : assert_eq!(
8569 12 : res.into_iter()
8570 48 : .map(|(k, v)| (k, v.unwrap()))
8571 12 : .collect::<Vec<_>>(),
8572 12 : vec![
8573 12 : (base_inherited_key, test_img("metadata inherited key 1")),
8574 12 : (
8575 12 : base_inherited_key_overwrite,
8576 12 : test_img("metadata key overwrite 1a")
8577 12 : ),
8578 12 : (base_key, test_img("metadata key 1")),
8579 12 : (base_key_overwrite, test_img("metadata key overwrite 1b")),
8580 12 : ]
8581 12 : );
8582 12 :
8583 12 : // test vectored scan on child timeline
8584 12 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency.clone());
8585 12 : let query =
8586 12 : VersionedKeySpaceQuery::uniform(KeySpace::single(Key::metadata_key_range()), lsn);
8587 12 : let res = child
8588 12 : .get_vectored_impl(query, &mut reconstruct_state, &ctx)
8589 12 : .await?;
8590 12 :
8591 12 : assert_eq!(
8592 12 : res.into_iter()
8593 60 : .map(|(k, v)| (k, v.unwrap()))
8594 12 : .collect::<Vec<_>>(),
8595 12 : vec![
8596 12 : (base_inherited_key, test_img("metadata inherited key 1")),
8597 12 : (
8598 12 : base_inherited_key_child,
8599 12 : test_img("metadata inherited key 2")
8600 12 : ),
8601 12 : (
8602 12 : base_inherited_key_overwrite,
8603 12 : test_img("metadata key overwrite 2a")
8604 12 : ),
8605 12 : (base_key_child, test_img("metadata key 2")),
8606 12 : (base_key_overwrite, test_img("metadata key overwrite 2b")),
8607 12 : ]
8608 12 : );
8609 12 :
8610 12 : Ok(())
8611 12 : }
8612 :
8613 336 : async fn get_vectored_impl_wrapper(
8614 336 : tline: &Arc<Timeline>,
8615 336 : key: Key,
8616 336 : lsn: Lsn,
8617 336 : ctx: &RequestContext,
8618 336 : ) -> Result<Option<Bytes>, GetVectoredError> {
8619 336 : let io_concurrency =
8620 336 : IoConcurrency::spawn_from_conf(tline.conf, tline.gate.enter().unwrap());
8621 336 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
8622 336 : let query = VersionedKeySpaceQuery::uniform(KeySpace::single(key..key.next()), lsn);
8623 336 : let mut res = tline
8624 336 : .get_vectored_impl(query, &mut reconstruct_state, ctx)
8625 336 : .await?;
8626 300 : Ok(res.pop_last().map(|(k, v)| {
8627 192 : assert_eq!(k, key);
8628 192 : v.unwrap()
8629 300 : }))
8630 336 : }
8631 :
8632 : #[tokio::test]
8633 12 : async fn test_metadata_tombstone_reads() -> anyhow::Result<()> {
8634 12 : let harness = TenantHarness::create("test_metadata_tombstone_reads").await?;
8635 12 : let (tenant, ctx) = harness.load().await;
8636 12 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
8637 12 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8638 12 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8639 12 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
8640 12 :
8641 12 : // We emulate the situation that the compaction algorithm creates an image layer that removes the tombstones
8642 12 : // Lsn 0x30 key0, key3, no key1+key2
8643 12 : // Lsn 0x20 key1+key2 tomestones
8644 12 : // Lsn 0x10 key1 in image, key2 in delta
8645 12 : let tline = tenant
8646 12 : .create_test_timeline_with_layers(
8647 12 : TIMELINE_ID,
8648 12 : Lsn(0x10),
8649 12 : DEFAULT_PG_VERSION,
8650 12 : &ctx,
8651 12 : Vec::new(), // in-memory layers
8652 12 : // delta layers
8653 12 : vec![
8654 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
8655 12 : Lsn(0x10)..Lsn(0x20),
8656 12 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8657 12 : ),
8658 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
8659 12 : Lsn(0x20)..Lsn(0x30),
8660 12 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8661 12 : ),
8662 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
8663 12 : Lsn(0x20)..Lsn(0x30),
8664 12 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8665 12 : ),
8666 12 : ],
8667 12 : // image layers
8668 12 : vec![
8669 12 : (Lsn(0x10), vec![(key1, test_img("metadata key 1"))]),
8670 12 : (
8671 12 : Lsn(0x30),
8672 12 : vec![
8673 12 : (key0, test_img("metadata key 0")),
8674 12 : (key3, test_img("metadata key 3")),
8675 12 : ],
8676 12 : ),
8677 12 : ],
8678 12 : Lsn(0x30),
8679 12 : )
8680 12 : .await?;
8681 12 :
8682 12 : let lsn = Lsn(0x30);
8683 12 : let old_lsn = Lsn(0x20);
8684 12 :
8685 12 : assert_eq!(
8686 12 : get_vectored_impl_wrapper(&tline, key0, lsn, &ctx).await?,
8687 12 : Some(test_img("metadata key 0"))
8688 12 : );
8689 12 : assert_eq!(
8690 12 : get_vectored_impl_wrapper(&tline, key1, lsn, &ctx).await?,
8691 12 : None,
8692 12 : );
8693 12 : assert_eq!(
8694 12 : get_vectored_impl_wrapper(&tline, key2, lsn, &ctx).await?,
8695 12 : None,
8696 12 : );
8697 12 : assert_eq!(
8698 12 : get_vectored_impl_wrapper(&tline, key1, old_lsn, &ctx).await?,
8699 12 : Some(Bytes::new()),
8700 12 : );
8701 12 : assert_eq!(
8702 12 : get_vectored_impl_wrapper(&tline, key2, old_lsn, &ctx).await?,
8703 12 : Some(Bytes::new()),
8704 12 : );
8705 12 : assert_eq!(
8706 12 : get_vectored_impl_wrapper(&tline, key3, lsn, &ctx).await?,
8707 12 : Some(test_img("metadata key 3"))
8708 12 : );
8709 12 :
8710 12 : Ok(())
8711 12 : }
8712 :
8713 : #[tokio::test]
8714 12 : async fn test_metadata_tombstone_image_creation() {
8715 12 : let harness = TenantHarness::create("test_metadata_tombstone_image_creation")
8716 12 : .await
8717 12 : .unwrap();
8718 12 : let (tenant, ctx) = harness.load().await;
8719 12 : let io_concurrency = IoConcurrency::spawn_for_test();
8720 12 :
8721 12 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
8722 12 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8723 12 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8724 12 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
8725 12 :
8726 12 : let tline = tenant
8727 12 : .create_test_timeline_with_layers(
8728 12 : TIMELINE_ID,
8729 12 : Lsn(0x10),
8730 12 : DEFAULT_PG_VERSION,
8731 12 : &ctx,
8732 12 : Vec::new(), // in-memory layers
8733 12 : // delta layers
8734 12 : vec![
8735 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
8736 12 : Lsn(0x10)..Lsn(0x20),
8737 12 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8738 12 : ),
8739 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
8740 12 : Lsn(0x20)..Lsn(0x30),
8741 12 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8742 12 : ),
8743 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
8744 12 : Lsn(0x20)..Lsn(0x30),
8745 12 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8746 12 : ),
8747 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
8748 12 : Lsn(0x30)..Lsn(0x40),
8749 12 : vec![
8750 12 : (key0, Lsn(0x30), Value::Image(test_img("metadata key 0"))),
8751 12 : (key3, Lsn(0x30), Value::Image(test_img("metadata key 3"))),
8752 12 : ],
8753 12 : ),
8754 12 : ],
8755 12 : // image layers
8756 12 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
8757 12 : Lsn(0x40),
8758 12 : )
8759 12 : .await
8760 12 : .unwrap();
8761 12 :
8762 12 : let cancel = CancellationToken::new();
8763 12 :
8764 12 : tline
8765 12 : .compact(
8766 12 : &cancel,
8767 12 : {
8768 12 : let mut flags = EnumSet::new();
8769 12 : flags.insert(CompactFlags::ForceImageLayerCreation);
8770 12 : flags.insert(CompactFlags::ForceRepartition);
8771 12 : flags
8772 12 : },
8773 12 : &ctx,
8774 12 : )
8775 12 : .await
8776 12 : .unwrap();
8777 12 :
8778 12 : // Image layers are created at last_record_lsn
8779 12 : let images = tline
8780 12 : .inspect_image_layers(Lsn(0x40), &ctx, io_concurrency.clone())
8781 12 : .await
8782 12 : .unwrap()
8783 12 : .into_iter()
8784 108 : .filter(|(k, _)| k.is_metadata_key())
8785 12 : .collect::<Vec<_>>();
8786 12 : assert_eq!(images.len(), 2); // the image layer should only contain two existing keys, tombstones should be removed.
8787 12 : }
8788 :
8789 : #[tokio::test]
8790 12 : async fn test_metadata_tombstone_empty_image_creation() {
8791 12 : let harness = TenantHarness::create("test_metadata_tombstone_empty_image_creation")
8792 12 : .await
8793 12 : .unwrap();
8794 12 : let (tenant, ctx) = harness.load().await;
8795 12 : let io_concurrency = IoConcurrency::spawn_for_test();
8796 12 :
8797 12 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8798 12 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8799 12 :
8800 12 : let tline = tenant
8801 12 : .create_test_timeline_with_layers(
8802 12 : TIMELINE_ID,
8803 12 : Lsn(0x10),
8804 12 : DEFAULT_PG_VERSION,
8805 12 : &ctx,
8806 12 : Vec::new(), // in-memory layers
8807 12 : // delta layers
8808 12 : vec![
8809 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
8810 12 : Lsn(0x10)..Lsn(0x20),
8811 12 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8812 12 : ),
8813 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
8814 12 : Lsn(0x20)..Lsn(0x30),
8815 12 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8816 12 : ),
8817 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
8818 12 : Lsn(0x20)..Lsn(0x30),
8819 12 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8820 12 : ),
8821 12 : ],
8822 12 : // image layers
8823 12 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
8824 12 : Lsn(0x30),
8825 12 : )
8826 12 : .await
8827 12 : .unwrap();
8828 12 :
8829 12 : let cancel = CancellationToken::new();
8830 12 :
8831 12 : tline
8832 12 : .compact(
8833 12 : &cancel,
8834 12 : {
8835 12 : let mut flags = EnumSet::new();
8836 12 : flags.insert(CompactFlags::ForceImageLayerCreation);
8837 12 : flags.insert(CompactFlags::ForceRepartition);
8838 12 : flags
8839 12 : },
8840 12 : &ctx,
8841 12 : )
8842 12 : .await
8843 12 : .unwrap();
8844 12 :
8845 12 : // Image layers are created at last_record_lsn
8846 12 : let images = tline
8847 12 : .inspect_image_layers(Lsn(0x30), &ctx, io_concurrency.clone())
8848 12 : .await
8849 12 : .unwrap()
8850 12 : .into_iter()
8851 84 : .filter(|(k, _)| k.is_metadata_key())
8852 12 : .collect::<Vec<_>>();
8853 12 : assert_eq!(images.len(), 0); // the image layer should not contain tombstones, or it is not created
8854 12 : }
8855 :
8856 : #[tokio::test]
8857 12 : async fn test_simple_bottom_most_compaction_images() -> anyhow::Result<()> {
8858 12 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_images").await?;
8859 12 : let (tenant, ctx) = harness.load().await;
8860 12 : let io_concurrency = IoConcurrency::spawn_for_test();
8861 12 :
8862 612 : fn get_key(id: u32) -> Key {
8863 612 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8864 612 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8865 612 : key.field6 = id;
8866 612 : key
8867 612 : }
8868 12 :
8869 12 : // We create
8870 12 : // - one bottom-most image layer,
8871 12 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
8872 12 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
8873 12 : // - a delta layer D3 above the horizon.
8874 12 : //
8875 12 : // | D3 |
8876 12 : // | D1 |
8877 12 : // -| |-- gc horizon -----------------
8878 12 : // | | | D2 |
8879 12 : // --------- img layer ------------------
8880 12 : //
8881 12 : // What we should expact from this compaction is:
8882 12 : // | D3 |
8883 12 : // | Part of D1 |
8884 12 : // --------- img layer with D1+D2 at GC horizon------------------
8885 12 :
8886 12 : // img layer at 0x10
8887 12 : let img_layer = (0..10)
8888 120 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8889 12 : .collect_vec();
8890 12 :
8891 12 : let delta1 = vec![
8892 12 : (
8893 12 : get_key(1),
8894 12 : Lsn(0x20),
8895 12 : Value::Image(Bytes::from("value 1@0x20")),
8896 12 : ),
8897 12 : (
8898 12 : get_key(2),
8899 12 : Lsn(0x30),
8900 12 : Value::Image(Bytes::from("value 2@0x30")),
8901 12 : ),
8902 12 : (
8903 12 : get_key(3),
8904 12 : Lsn(0x40),
8905 12 : Value::Image(Bytes::from("value 3@0x40")),
8906 12 : ),
8907 12 : ];
8908 12 : let delta2 = vec![
8909 12 : (
8910 12 : get_key(5),
8911 12 : Lsn(0x20),
8912 12 : Value::Image(Bytes::from("value 5@0x20")),
8913 12 : ),
8914 12 : (
8915 12 : get_key(6),
8916 12 : Lsn(0x20),
8917 12 : Value::Image(Bytes::from("value 6@0x20")),
8918 12 : ),
8919 12 : ];
8920 12 : let delta3 = vec![
8921 12 : (
8922 12 : get_key(8),
8923 12 : Lsn(0x48),
8924 12 : Value::Image(Bytes::from("value 8@0x48")),
8925 12 : ),
8926 12 : (
8927 12 : get_key(9),
8928 12 : Lsn(0x48),
8929 12 : Value::Image(Bytes::from("value 9@0x48")),
8930 12 : ),
8931 12 : ];
8932 12 :
8933 12 : let tline = tenant
8934 12 : .create_test_timeline_with_layers(
8935 12 : TIMELINE_ID,
8936 12 : Lsn(0x10),
8937 12 : DEFAULT_PG_VERSION,
8938 12 : &ctx,
8939 12 : Vec::new(), // in-memory layers
8940 12 : vec![
8941 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
8942 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
8943 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
8944 12 : ], // delta layers
8945 12 : vec![(Lsn(0x10), img_layer)], // image layers
8946 12 : Lsn(0x50),
8947 12 : )
8948 12 : .await?;
8949 12 : {
8950 12 : tline
8951 12 : .applied_gc_cutoff_lsn
8952 12 : .lock_for_write()
8953 12 : .store_and_unlock(Lsn(0x30))
8954 12 : .wait()
8955 12 : .await;
8956 12 : // Update GC info
8957 12 : let mut guard = tline.gc_info.write().unwrap();
8958 12 : guard.cutoffs.time = Lsn(0x30);
8959 12 : guard.cutoffs.space = Lsn(0x30);
8960 12 : }
8961 12 :
8962 12 : let expected_result = [
8963 12 : Bytes::from_static(b"value 0@0x10"),
8964 12 : Bytes::from_static(b"value 1@0x20"),
8965 12 : Bytes::from_static(b"value 2@0x30"),
8966 12 : Bytes::from_static(b"value 3@0x40"),
8967 12 : Bytes::from_static(b"value 4@0x10"),
8968 12 : Bytes::from_static(b"value 5@0x20"),
8969 12 : Bytes::from_static(b"value 6@0x20"),
8970 12 : Bytes::from_static(b"value 7@0x10"),
8971 12 : Bytes::from_static(b"value 8@0x48"),
8972 12 : Bytes::from_static(b"value 9@0x48"),
8973 12 : ];
8974 12 :
8975 120 : for (idx, expected) in expected_result.iter().enumerate() {
8976 120 : assert_eq!(
8977 120 : tline
8978 120 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8979 120 : .await
8980 120 : .unwrap(),
8981 12 : expected
8982 12 : );
8983 12 : }
8984 12 :
8985 12 : let cancel = CancellationToken::new();
8986 12 : tline
8987 12 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8988 12 : .await
8989 12 : .unwrap();
8990 12 :
8991 120 : for (idx, expected) in expected_result.iter().enumerate() {
8992 120 : assert_eq!(
8993 120 : tline
8994 120 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8995 120 : .await
8996 120 : .unwrap(),
8997 12 : expected
8998 12 : );
8999 12 : }
9000 12 :
9001 12 : // Check if the image layer at the GC horizon contains exactly what we want
9002 12 : let image_at_gc_horizon = tline
9003 12 : .inspect_image_layers(Lsn(0x30), &ctx, io_concurrency.clone())
9004 12 : .await
9005 12 : .unwrap()
9006 12 : .into_iter()
9007 204 : .filter(|(k, _)| k.is_metadata_key())
9008 12 : .collect::<Vec<_>>();
9009 12 :
9010 12 : assert_eq!(image_at_gc_horizon.len(), 10);
9011 12 : let expected_result = [
9012 12 : Bytes::from_static(b"value 0@0x10"),
9013 12 : Bytes::from_static(b"value 1@0x20"),
9014 12 : Bytes::from_static(b"value 2@0x30"),
9015 12 : Bytes::from_static(b"value 3@0x10"),
9016 12 : Bytes::from_static(b"value 4@0x10"),
9017 12 : Bytes::from_static(b"value 5@0x20"),
9018 12 : Bytes::from_static(b"value 6@0x20"),
9019 12 : Bytes::from_static(b"value 7@0x10"),
9020 12 : Bytes::from_static(b"value 8@0x10"),
9021 12 : Bytes::from_static(b"value 9@0x10"),
9022 12 : ];
9023 132 : for idx in 0..10 {
9024 120 : assert_eq!(
9025 120 : image_at_gc_horizon[idx],
9026 120 : (get_key(idx as u32), expected_result[idx].clone())
9027 120 : );
9028 12 : }
9029 12 :
9030 12 : // Check if old layers are removed / new layers have the expected LSN
9031 12 : let all_layers = inspect_and_sort(&tline, None).await;
9032 12 : assert_eq!(
9033 12 : all_layers,
9034 12 : vec![
9035 12 : // Image layer at GC horizon
9036 12 : PersistentLayerKey {
9037 12 : key_range: Key::MIN..Key::MAX,
9038 12 : lsn_range: Lsn(0x30)..Lsn(0x31),
9039 12 : is_delta: false
9040 12 : },
9041 12 : // The delta layer below the horizon
9042 12 : PersistentLayerKey {
9043 12 : key_range: get_key(3)..get_key(4),
9044 12 : lsn_range: Lsn(0x30)..Lsn(0x48),
9045 12 : is_delta: true
9046 12 : },
9047 12 : // The delta3 layer that should not be picked for the compaction
9048 12 : PersistentLayerKey {
9049 12 : key_range: get_key(8)..get_key(10),
9050 12 : lsn_range: Lsn(0x48)..Lsn(0x50),
9051 12 : is_delta: true
9052 12 : }
9053 12 : ]
9054 12 : );
9055 12 :
9056 12 : // increase GC horizon and compact again
9057 12 : {
9058 12 : tline
9059 12 : .applied_gc_cutoff_lsn
9060 12 : .lock_for_write()
9061 12 : .store_and_unlock(Lsn(0x40))
9062 12 : .wait()
9063 12 : .await;
9064 12 : // Update GC info
9065 12 : let mut guard = tline.gc_info.write().unwrap();
9066 12 : guard.cutoffs.time = Lsn(0x40);
9067 12 : guard.cutoffs.space = Lsn(0x40);
9068 12 : }
9069 12 : tline
9070 12 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9071 12 : .await
9072 12 : .unwrap();
9073 12 :
9074 12 : Ok(())
9075 12 : }
9076 :
9077 : #[cfg(feature = "testing")]
9078 : #[tokio::test]
9079 12 : async fn test_neon_test_record() -> anyhow::Result<()> {
9080 12 : let harness = TenantHarness::create("test_neon_test_record").await?;
9081 12 : let (tenant, ctx) = harness.load().await;
9082 12 :
9083 204 : fn get_key(id: u32) -> Key {
9084 204 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9085 204 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9086 204 : key.field6 = id;
9087 204 : key
9088 204 : }
9089 12 :
9090 12 : let delta1 = vec![
9091 12 : (
9092 12 : get_key(1),
9093 12 : Lsn(0x20),
9094 12 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
9095 12 : ),
9096 12 : (
9097 12 : get_key(1),
9098 12 : Lsn(0x30),
9099 12 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
9100 12 : ),
9101 12 : (get_key(2), Lsn(0x10), Value::Image("0x10".into())),
9102 12 : (
9103 12 : get_key(2),
9104 12 : Lsn(0x20),
9105 12 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
9106 12 : ),
9107 12 : (
9108 12 : get_key(2),
9109 12 : Lsn(0x30),
9110 12 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
9111 12 : ),
9112 12 : (get_key(3), Lsn(0x10), Value::Image("0x10".into())),
9113 12 : (
9114 12 : get_key(3),
9115 12 : Lsn(0x20),
9116 12 : Value::WalRecord(NeonWalRecord::wal_clear("c")),
9117 12 : ),
9118 12 : (get_key(4), Lsn(0x10), Value::Image("0x10".into())),
9119 12 : (
9120 12 : get_key(4),
9121 12 : Lsn(0x20),
9122 12 : Value::WalRecord(NeonWalRecord::wal_init("i")),
9123 12 : ),
9124 12 : (
9125 12 : get_key(4),
9126 12 : Lsn(0x30),
9127 12 : Value::WalRecord(NeonWalRecord::wal_append_conditional("j", "i")),
9128 12 : ),
9129 12 : (
9130 12 : get_key(5),
9131 12 : Lsn(0x20),
9132 12 : Value::WalRecord(NeonWalRecord::wal_init("1")),
9133 12 : ),
9134 12 : (
9135 12 : get_key(5),
9136 12 : Lsn(0x30),
9137 12 : Value::WalRecord(NeonWalRecord::wal_append_conditional("j", "2")),
9138 12 : ),
9139 12 : ];
9140 12 : let image1 = vec![(get_key(1), "0x10".into())];
9141 12 :
9142 12 : let tline = tenant
9143 12 : .create_test_timeline_with_layers(
9144 12 : TIMELINE_ID,
9145 12 : Lsn(0x10),
9146 12 : DEFAULT_PG_VERSION,
9147 12 : &ctx,
9148 12 : Vec::new(), // in-memory layers
9149 12 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
9150 12 : Lsn(0x10)..Lsn(0x40),
9151 12 : delta1,
9152 12 : )], // delta layers
9153 12 : vec![(Lsn(0x10), image1)], // image layers
9154 12 : Lsn(0x50),
9155 12 : )
9156 12 : .await?;
9157 12 :
9158 12 : assert_eq!(
9159 12 : tline.get(get_key(1), Lsn(0x50), &ctx).await?,
9160 12 : Bytes::from_static(b"0x10,0x20,0x30")
9161 12 : );
9162 12 : assert_eq!(
9163 12 : tline.get(get_key(2), Lsn(0x50), &ctx).await?,
9164 12 : Bytes::from_static(b"0x10,0x20,0x30")
9165 12 : );
9166 12 :
9167 12 : // Need to remove the limit of "Neon WAL redo requires base image".
9168 12 :
9169 12 : assert_eq!(
9170 12 : tline.get(get_key(3), Lsn(0x50), &ctx).await?,
9171 12 : Bytes::from_static(b"c")
9172 12 : );
9173 12 : assert_eq!(
9174 12 : tline.get(get_key(4), Lsn(0x50), &ctx).await?,
9175 12 : Bytes::from_static(b"ij")
9176 12 : );
9177 12 :
9178 12 : // Manual testing required: currently, read errors will panic the process in debug mode. So we
9179 12 : // cannot enable this assertion in the unit test.
9180 12 : // assert!(tline.get(get_key(5), Lsn(0x50), &ctx).await.is_err());
9181 12 :
9182 12 : Ok(())
9183 12 : }
9184 :
9185 : #[tokio::test(start_paused = true)]
9186 12 : async fn test_lsn_lease() -> anyhow::Result<()> {
9187 12 : let (tenant, ctx) = TenantHarness::create("test_lsn_lease")
9188 12 : .await
9189 12 : .unwrap()
9190 12 : .load()
9191 12 : .await;
9192 12 : // Advance to the lsn lease deadline so that GC is not blocked by
9193 12 : // initial transition into AttachedSingle.
9194 12 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
9195 12 : tokio::time::resume();
9196 12 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
9197 12 :
9198 12 : let end_lsn = Lsn(0x100);
9199 12 : let image_layers = (0x20..=0x90)
9200 12 : .step_by(0x10)
9201 96 : .map(|n| {
9202 96 : (
9203 96 : Lsn(n),
9204 96 : vec![(key, test_img(&format!("data key at {:x}", n)))],
9205 96 : )
9206 96 : })
9207 12 : .collect();
9208 12 :
9209 12 : let timeline = tenant
9210 12 : .create_test_timeline_with_layers(
9211 12 : TIMELINE_ID,
9212 12 : Lsn(0x10),
9213 12 : DEFAULT_PG_VERSION,
9214 12 : &ctx,
9215 12 : Vec::new(), // in-memory layers
9216 12 : Vec::new(),
9217 12 : image_layers,
9218 12 : end_lsn,
9219 12 : )
9220 12 : .await?;
9221 12 :
9222 12 : let leased_lsns = [0x30, 0x50, 0x70];
9223 12 : let mut leases = Vec::new();
9224 36 : leased_lsns.iter().for_each(|n| {
9225 36 : leases.push(
9226 36 : timeline
9227 36 : .init_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)
9228 36 : .expect("lease request should succeed"),
9229 36 : );
9230 36 : });
9231 12 :
9232 12 : let updated_lease_0 = timeline
9233 12 : .renew_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)
9234 12 : .expect("lease renewal should succeed");
9235 12 : assert_eq!(
9236 12 : updated_lease_0.valid_until, leases[0].valid_until,
9237 12 : " Renewing with shorter lease should not change the lease."
9238 12 : );
9239 12 :
9240 12 : let updated_lease_1 = timeline
9241 12 : .renew_lsn_lease(
9242 12 : Lsn(leased_lsns[1]),
9243 12 : timeline.get_lsn_lease_length() * 2,
9244 12 : &ctx,
9245 12 : )
9246 12 : .expect("lease renewal should succeed");
9247 12 : assert!(
9248 12 : updated_lease_1.valid_until > leases[1].valid_until,
9249 12 : "Renewing with a long lease should renew lease with later expiration time."
9250 12 : );
9251 12 :
9252 12 : // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
9253 12 : info!(
9254 12 : "applied_gc_cutoff_lsn: {}",
9255 0 : *timeline.get_applied_gc_cutoff_lsn()
9256 12 : );
9257 12 : timeline.force_set_disk_consistent_lsn(end_lsn);
9258 12 :
9259 12 : let res = tenant
9260 12 : .gc_iteration(
9261 12 : Some(TIMELINE_ID),
9262 12 : 0,
9263 12 : Duration::ZERO,
9264 12 : &CancellationToken::new(),
9265 12 : &ctx,
9266 12 : )
9267 12 : .await
9268 12 : .unwrap();
9269 12 :
9270 12 : // Keeping everything <= Lsn(0x80) b/c leases:
9271 12 : // 0/10: initdb layer
9272 12 : // (0/20..=0/70).step_by(0x10): image layers added when creating the timeline.
9273 12 : assert_eq!(res.layers_needed_by_leases, 7);
9274 12 : // Keeping 0/90 b/c it is the latest layer.
9275 12 : assert_eq!(res.layers_not_updated, 1);
9276 12 : // Removed 0/80.
9277 12 : assert_eq!(res.layers_removed, 1);
9278 12 :
9279 12 : // Make lease on a already GC-ed LSN.
9280 12 : // 0/80 does not have a valid lease + is below latest_gc_cutoff
9281 12 : assert!(Lsn(0x80) < *timeline.get_applied_gc_cutoff_lsn());
9282 12 : timeline
9283 12 : .init_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx)
9284 12 : .expect_err("lease request on GC-ed LSN should fail");
9285 12 :
9286 12 : // Should still be able to renew a currently valid lease
9287 12 : // Assumption: original lease to is still valid for 0/50.
9288 12 : // (use `Timeline::init_lsn_lease` for testing so it always does validation)
9289 12 : timeline
9290 12 : .init_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)
9291 12 : .expect("lease renewal with validation should succeed");
9292 12 :
9293 12 : Ok(())
9294 12 : }
9295 :
9296 : #[cfg(feature = "testing")]
9297 : #[tokio::test]
9298 12 : async fn test_simple_bottom_most_compaction_deltas_1() -> anyhow::Result<()> {
9299 12 : test_simple_bottom_most_compaction_deltas_helper(
9300 12 : "test_simple_bottom_most_compaction_deltas_1",
9301 12 : false,
9302 12 : )
9303 12 : .await
9304 12 : }
9305 :
9306 : #[cfg(feature = "testing")]
9307 : #[tokio::test]
9308 12 : async fn test_simple_bottom_most_compaction_deltas_2() -> anyhow::Result<()> {
9309 12 : test_simple_bottom_most_compaction_deltas_helper(
9310 12 : "test_simple_bottom_most_compaction_deltas_2",
9311 12 : true,
9312 12 : )
9313 12 : .await
9314 12 : }
9315 :
9316 : #[cfg(feature = "testing")]
9317 24 : async fn test_simple_bottom_most_compaction_deltas_helper(
9318 24 : test_name: &'static str,
9319 24 : use_delta_bottom_layer: bool,
9320 24 : ) -> anyhow::Result<()> {
9321 24 : let harness = TenantHarness::create(test_name).await?;
9322 24 : let (tenant, ctx) = harness.load().await;
9323 :
9324 1656 : fn get_key(id: u32) -> Key {
9325 1656 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9326 1656 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9327 1656 : key.field6 = id;
9328 1656 : key
9329 1656 : }
9330 :
9331 : // We create
9332 : // - one bottom-most image layer,
9333 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
9334 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
9335 : // - a delta layer D3 above the horizon.
9336 : //
9337 : // | D3 |
9338 : // | D1 |
9339 : // -| |-- gc horizon -----------------
9340 : // | | | D2 |
9341 : // --------- img layer ------------------
9342 : //
9343 : // What we should expact from this compaction is:
9344 : // | D3 |
9345 : // | Part of D1 |
9346 : // --------- img layer with D1+D2 at GC horizon------------------
9347 :
9348 : // img layer at 0x10
9349 24 : let img_layer = (0..10)
9350 240 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9351 24 : .collect_vec();
9352 24 : // or, delta layer at 0x10 if `use_delta_bottom_layer` is true
9353 24 : let delta4 = (0..10)
9354 240 : .map(|id| {
9355 240 : (
9356 240 : get_key(id),
9357 240 : Lsn(0x08),
9358 240 : Value::WalRecord(NeonWalRecord::wal_init(format!("value {id}@0x10"))),
9359 240 : )
9360 240 : })
9361 24 : .collect_vec();
9362 24 :
9363 24 : let delta1 = vec![
9364 24 : (
9365 24 : get_key(1),
9366 24 : Lsn(0x20),
9367 24 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9368 24 : ),
9369 24 : (
9370 24 : get_key(2),
9371 24 : Lsn(0x30),
9372 24 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9373 24 : ),
9374 24 : (
9375 24 : get_key(3),
9376 24 : Lsn(0x28),
9377 24 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9378 24 : ),
9379 24 : (
9380 24 : get_key(3),
9381 24 : Lsn(0x30),
9382 24 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9383 24 : ),
9384 24 : (
9385 24 : get_key(3),
9386 24 : Lsn(0x40),
9387 24 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9388 24 : ),
9389 24 : ];
9390 24 : let delta2 = vec![
9391 24 : (
9392 24 : get_key(5),
9393 24 : Lsn(0x20),
9394 24 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9395 24 : ),
9396 24 : (
9397 24 : get_key(6),
9398 24 : Lsn(0x20),
9399 24 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9400 24 : ),
9401 24 : ];
9402 24 : let delta3 = vec![
9403 24 : (
9404 24 : get_key(8),
9405 24 : Lsn(0x48),
9406 24 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9407 24 : ),
9408 24 : (
9409 24 : get_key(9),
9410 24 : Lsn(0x48),
9411 24 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9412 24 : ),
9413 24 : ];
9414 :
9415 24 : let tline = if use_delta_bottom_layer {
9416 12 : tenant
9417 12 : .create_test_timeline_with_layers(
9418 12 : TIMELINE_ID,
9419 12 : Lsn(0x08),
9420 12 : DEFAULT_PG_VERSION,
9421 12 : &ctx,
9422 12 : Vec::new(), // in-memory layers
9423 12 : vec![
9424 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
9425 12 : Lsn(0x08)..Lsn(0x10),
9426 12 : delta4,
9427 12 : ),
9428 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
9429 12 : Lsn(0x20)..Lsn(0x48),
9430 12 : delta1,
9431 12 : ),
9432 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
9433 12 : Lsn(0x20)..Lsn(0x48),
9434 12 : delta2,
9435 12 : ),
9436 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
9437 12 : Lsn(0x48)..Lsn(0x50),
9438 12 : delta3,
9439 12 : ),
9440 12 : ], // delta layers
9441 12 : vec![], // image layers
9442 12 : Lsn(0x50),
9443 12 : )
9444 12 : .await?
9445 : } else {
9446 12 : tenant
9447 12 : .create_test_timeline_with_layers(
9448 12 : TIMELINE_ID,
9449 12 : Lsn(0x10),
9450 12 : DEFAULT_PG_VERSION,
9451 12 : &ctx,
9452 12 : Vec::new(), // in-memory layers
9453 12 : vec![
9454 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
9455 12 : Lsn(0x10)..Lsn(0x48),
9456 12 : delta1,
9457 12 : ),
9458 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
9459 12 : Lsn(0x10)..Lsn(0x48),
9460 12 : delta2,
9461 12 : ),
9462 12 : DeltaLayerTestDesc::new_with_inferred_key_range(
9463 12 : Lsn(0x48)..Lsn(0x50),
9464 12 : delta3,
9465 12 : ),
9466 12 : ], // delta layers
9467 12 : vec![(Lsn(0x10), img_layer)], // image layers
9468 12 : Lsn(0x50),
9469 12 : )
9470 12 : .await?
9471 : };
9472 : {
9473 24 : tline
9474 24 : .applied_gc_cutoff_lsn
9475 24 : .lock_for_write()
9476 24 : .store_and_unlock(Lsn(0x30))
9477 24 : .wait()
9478 24 : .await;
9479 : // Update GC info
9480 24 : let mut guard = tline.gc_info.write().unwrap();
9481 24 : *guard = GcInfo {
9482 24 : retain_lsns: vec![],
9483 24 : cutoffs: GcCutoffs {
9484 24 : time: Lsn(0x30),
9485 24 : space: Lsn(0x30),
9486 24 : },
9487 24 : leases: Default::default(),
9488 24 : within_ancestor_pitr: false,
9489 24 : };
9490 24 : }
9491 24 :
9492 24 : let expected_result = [
9493 24 : Bytes::from_static(b"value 0@0x10"),
9494 24 : Bytes::from_static(b"value 1@0x10@0x20"),
9495 24 : Bytes::from_static(b"value 2@0x10@0x30"),
9496 24 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9497 24 : Bytes::from_static(b"value 4@0x10"),
9498 24 : Bytes::from_static(b"value 5@0x10@0x20"),
9499 24 : Bytes::from_static(b"value 6@0x10@0x20"),
9500 24 : Bytes::from_static(b"value 7@0x10"),
9501 24 : Bytes::from_static(b"value 8@0x10@0x48"),
9502 24 : Bytes::from_static(b"value 9@0x10@0x48"),
9503 24 : ];
9504 24 :
9505 24 : let expected_result_at_gc_horizon = [
9506 24 : Bytes::from_static(b"value 0@0x10"),
9507 24 : Bytes::from_static(b"value 1@0x10@0x20"),
9508 24 : Bytes::from_static(b"value 2@0x10@0x30"),
9509 24 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
9510 24 : Bytes::from_static(b"value 4@0x10"),
9511 24 : Bytes::from_static(b"value 5@0x10@0x20"),
9512 24 : Bytes::from_static(b"value 6@0x10@0x20"),
9513 24 : Bytes::from_static(b"value 7@0x10"),
9514 24 : Bytes::from_static(b"value 8@0x10"),
9515 24 : Bytes::from_static(b"value 9@0x10"),
9516 24 : ];
9517 :
9518 264 : for idx in 0..10 {
9519 240 : assert_eq!(
9520 240 : tline
9521 240 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9522 240 : .await
9523 240 : .unwrap(),
9524 240 : &expected_result[idx]
9525 : );
9526 240 : assert_eq!(
9527 240 : tline
9528 240 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
9529 240 : .await
9530 240 : .unwrap(),
9531 240 : &expected_result_at_gc_horizon[idx]
9532 : );
9533 : }
9534 :
9535 24 : let cancel = CancellationToken::new();
9536 24 : tline
9537 24 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9538 24 : .await
9539 24 : .unwrap();
9540 :
9541 264 : for idx in 0..10 {
9542 240 : assert_eq!(
9543 240 : tline
9544 240 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9545 240 : .await
9546 240 : .unwrap(),
9547 240 : &expected_result[idx]
9548 : );
9549 240 : assert_eq!(
9550 240 : tline
9551 240 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
9552 240 : .await
9553 240 : .unwrap(),
9554 240 : &expected_result_at_gc_horizon[idx]
9555 : );
9556 : }
9557 :
9558 : // increase GC horizon and compact again
9559 : {
9560 24 : tline
9561 24 : .applied_gc_cutoff_lsn
9562 24 : .lock_for_write()
9563 24 : .store_and_unlock(Lsn(0x40))
9564 24 : .wait()
9565 24 : .await;
9566 : // Update GC info
9567 24 : let mut guard = tline.gc_info.write().unwrap();
9568 24 : guard.cutoffs.time = Lsn(0x40);
9569 24 : guard.cutoffs.space = Lsn(0x40);
9570 24 : }
9571 24 : tline
9572 24 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9573 24 : .await
9574 24 : .unwrap();
9575 24 :
9576 24 : Ok(())
9577 24 : }
9578 :
9579 : #[cfg(feature = "testing")]
9580 : #[tokio::test]
9581 12 : async fn test_generate_key_retention() -> anyhow::Result<()> {
9582 12 : let harness = TenantHarness::create("test_generate_key_retention").await?;
9583 12 : let (tenant, ctx) = harness.load().await;
9584 12 : let tline = tenant
9585 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
9586 12 : .await?;
9587 12 : tline.force_advance_lsn(Lsn(0x70));
9588 12 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
9589 12 : let history = vec![
9590 12 : (
9591 12 : key,
9592 12 : Lsn(0x10),
9593 12 : Value::WalRecord(NeonWalRecord::wal_init("0x10")),
9594 12 : ),
9595 12 : (
9596 12 : key,
9597 12 : Lsn(0x20),
9598 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9599 12 : ),
9600 12 : (
9601 12 : key,
9602 12 : Lsn(0x30),
9603 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9604 12 : ),
9605 12 : (
9606 12 : key,
9607 12 : Lsn(0x40),
9608 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9609 12 : ),
9610 12 : (
9611 12 : key,
9612 12 : Lsn(0x50),
9613 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
9614 12 : ),
9615 12 : (
9616 12 : key,
9617 12 : Lsn(0x60),
9618 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9619 12 : ),
9620 12 : (
9621 12 : key,
9622 12 : Lsn(0x70),
9623 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9624 12 : ),
9625 12 : (
9626 12 : key,
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 : key,
9634 12 : Lsn(0x90),
9635 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9636 12 : ),
9637 12 : ];
9638 12 : let res = tline
9639 12 : .generate_key_retention(
9640 12 : key,
9641 12 : &history,
9642 12 : Lsn(0x60),
9643 12 : &[Lsn(0x20), Lsn(0x40), Lsn(0x50)],
9644 12 : 3,
9645 12 : None,
9646 12 : true,
9647 12 : )
9648 12 : .await
9649 12 : .unwrap();
9650 12 : let expected_res = KeyHistoryRetention {
9651 12 : below_horizon: vec![
9652 12 : (
9653 12 : Lsn(0x20),
9654 12 : KeyLogAtLsn(vec![(
9655 12 : Lsn(0x20),
9656 12 : Value::Image(Bytes::from_static(b"0x10;0x20")),
9657 12 : )]),
9658 12 : ),
9659 12 : (
9660 12 : Lsn(0x40),
9661 12 : KeyLogAtLsn(vec![
9662 12 : (
9663 12 : Lsn(0x30),
9664 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9665 12 : ),
9666 12 : (
9667 12 : Lsn(0x40),
9668 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9669 12 : ),
9670 12 : ]),
9671 12 : ),
9672 12 : (
9673 12 : Lsn(0x50),
9674 12 : KeyLogAtLsn(vec![(
9675 12 : Lsn(0x50),
9676 12 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40;0x50")),
9677 12 : )]),
9678 12 : ),
9679 12 : (
9680 12 : Lsn(0x60),
9681 12 : KeyLogAtLsn(vec![(
9682 12 : Lsn(0x60),
9683 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9684 12 : )]),
9685 12 : ),
9686 12 : ],
9687 12 : above_horizon: KeyLogAtLsn(vec![
9688 12 : (
9689 12 : Lsn(0x70),
9690 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9691 12 : ),
9692 12 : (
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 : Lsn(0x90),
9700 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9701 12 : ),
9702 12 : ]),
9703 12 : };
9704 12 : assert_eq!(res, expected_res);
9705 12 :
9706 12 : // We expect GC-compaction to run with the original GC. This would create a situation that
9707 12 : // the original GC algorithm removes some delta layers b/c there are full image coverage,
9708 12 : // therefore causing some keys to have an incomplete history below the lowest retain LSN.
9709 12 : // For example, we have
9710 12 : // ```plain
9711 12 : // init delta @ 0x10, image @ 0x20, delta @ 0x30 (gc_horizon), image @ 0x40.
9712 12 : // ```
9713 12 : // Now the GC horizon moves up, and we have
9714 12 : // ```plain
9715 12 : // init delta @ 0x10, image @ 0x20, delta @ 0x30, image @ 0x40 (gc_horizon)
9716 12 : // ```
9717 12 : // The original GC algorithm kicks in, and removes delta @ 0x10, image @ 0x20.
9718 12 : // We will end up with
9719 12 : // ```plain
9720 12 : // delta @ 0x30, image @ 0x40 (gc_horizon)
9721 12 : // ```
9722 12 : // Now we run the GC-compaction, and this key does not have a full history.
9723 12 : // We should be able to handle this partial history and drop everything before the
9724 12 : // gc_horizon image.
9725 12 :
9726 12 : let history = vec![
9727 12 : (
9728 12 : key,
9729 12 : Lsn(0x20),
9730 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9731 12 : ),
9732 12 : (
9733 12 : key,
9734 12 : Lsn(0x30),
9735 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9736 12 : ),
9737 12 : (
9738 12 : key,
9739 12 : Lsn(0x40),
9740 12 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
9741 12 : ),
9742 12 : (
9743 12 : key,
9744 12 : Lsn(0x50),
9745 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
9746 12 : ),
9747 12 : (
9748 12 : key,
9749 12 : Lsn(0x60),
9750 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9751 12 : ),
9752 12 : (
9753 12 : key,
9754 12 : Lsn(0x70),
9755 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9756 12 : ),
9757 12 : (
9758 12 : key,
9759 12 : Lsn(0x80),
9760 12 : Value::Image(Bytes::copy_from_slice(
9761 12 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9762 12 : )),
9763 12 : ),
9764 12 : (
9765 12 : key,
9766 12 : Lsn(0x90),
9767 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9768 12 : ),
9769 12 : ];
9770 12 : let res = tline
9771 12 : .generate_key_retention(
9772 12 : key,
9773 12 : &history,
9774 12 : Lsn(0x60),
9775 12 : &[Lsn(0x40), Lsn(0x50)],
9776 12 : 3,
9777 12 : None,
9778 12 : true,
9779 12 : )
9780 12 : .await
9781 12 : .unwrap();
9782 12 : let expected_res = KeyHistoryRetention {
9783 12 : below_horizon: vec![
9784 12 : (
9785 12 : Lsn(0x40),
9786 12 : KeyLogAtLsn(vec![(
9787 12 : Lsn(0x40),
9788 12 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
9789 12 : )]),
9790 12 : ),
9791 12 : (
9792 12 : Lsn(0x50),
9793 12 : KeyLogAtLsn(vec![(
9794 12 : Lsn(0x50),
9795 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
9796 12 : )]),
9797 12 : ),
9798 12 : (
9799 12 : Lsn(0x60),
9800 12 : KeyLogAtLsn(vec![(
9801 12 : Lsn(0x60),
9802 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9803 12 : )]),
9804 12 : ),
9805 12 : ],
9806 12 : above_horizon: KeyLogAtLsn(vec![
9807 12 : (
9808 12 : Lsn(0x70),
9809 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9810 12 : ),
9811 12 : (
9812 12 : Lsn(0x80),
9813 12 : Value::Image(Bytes::copy_from_slice(
9814 12 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9815 12 : )),
9816 12 : ),
9817 12 : (
9818 12 : Lsn(0x90),
9819 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9820 12 : ),
9821 12 : ]),
9822 12 : };
9823 12 : assert_eq!(res, expected_res);
9824 12 :
9825 12 : // In case of branch compaction, the branch itself does not have the full history, and we need to provide
9826 12 : // the ancestor image in the test case.
9827 12 :
9828 12 : let history = vec![
9829 12 : (
9830 12 : key,
9831 12 : Lsn(0x20),
9832 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9833 12 : ),
9834 12 : (
9835 12 : key,
9836 12 : Lsn(0x30),
9837 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9838 12 : ),
9839 12 : (
9840 12 : key,
9841 12 : Lsn(0x40),
9842 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9843 12 : ),
9844 12 : (
9845 12 : key,
9846 12 : Lsn(0x70),
9847 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9848 12 : ),
9849 12 : ];
9850 12 : let res = tline
9851 12 : .generate_key_retention(
9852 12 : key,
9853 12 : &history,
9854 12 : Lsn(0x60),
9855 12 : &[],
9856 12 : 3,
9857 12 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
9858 12 : true,
9859 12 : )
9860 12 : .await
9861 12 : .unwrap();
9862 12 : let expected_res = KeyHistoryRetention {
9863 12 : below_horizon: vec![(
9864 12 : Lsn(0x60),
9865 12 : KeyLogAtLsn(vec![(
9866 12 : Lsn(0x60),
9867 12 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")), // use the ancestor image to reconstruct the page
9868 12 : )]),
9869 12 : )],
9870 12 : above_horizon: KeyLogAtLsn(vec![(
9871 12 : Lsn(0x70),
9872 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9873 12 : )]),
9874 12 : };
9875 12 : assert_eq!(res, expected_res);
9876 12 :
9877 12 : let history = vec![
9878 12 : (
9879 12 : key,
9880 12 : Lsn(0x20),
9881 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9882 12 : ),
9883 12 : (
9884 12 : key,
9885 12 : Lsn(0x40),
9886 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9887 12 : ),
9888 12 : (
9889 12 : key,
9890 12 : Lsn(0x60),
9891 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9892 12 : ),
9893 12 : (
9894 12 : key,
9895 12 : Lsn(0x70),
9896 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9897 12 : ),
9898 12 : ];
9899 12 : let res = tline
9900 12 : .generate_key_retention(
9901 12 : key,
9902 12 : &history,
9903 12 : Lsn(0x60),
9904 12 : &[Lsn(0x30)],
9905 12 : 3,
9906 12 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
9907 12 : true,
9908 12 : )
9909 12 : .await
9910 12 : .unwrap();
9911 12 : let expected_res = KeyHistoryRetention {
9912 12 : below_horizon: vec![
9913 12 : (
9914 12 : Lsn(0x30),
9915 12 : KeyLogAtLsn(vec![(
9916 12 : Lsn(0x20),
9917 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9918 12 : )]),
9919 12 : ),
9920 12 : (
9921 12 : Lsn(0x60),
9922 12 : KeyLogAtLsn(vec![(
9923 12 : Lsn(0x60),
9924 12 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x40;0x60")),
9925 12 : )]),
9926 12 : ),
9927 12 : ],
9928 12 : above_horizon: KeyLogAtLsn(vec![(
9929 12 : Lsn(0x70),
9930 12 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9931 12 : )]),
9932 12 : };
9933 12 : assert_eq!(res, expected_res);
9934 12 :
9935 12 : Ok(())
9936 12 : }
9937 :
9938 : #[cfg(feature = "testing")]
9939 : #[tokio::test]
9940 12 : async fn test_simple_bottom_most_compaction_with_retain_lsns() -> anyhow::Result<()> {
9941 12 : let harness =
9942 12 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns").await?;
9943 12 : let (tenant, ctx) = harness.load().await;
9944 12 :
9945 3108 : fn get_key(id: u32) -> Key {
9946 3108 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9947 3108 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9948 3108 : key.field6 = id;
9949 3108 : key
9950 3108 : }
9951 12 :
9952 12 : let img_layer = (0..10)
9953 120 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9954 12 : .collect_vec();
9955 12 :
9956 12 : let delta1 = vec![
9957 12 : (
9958 12 : get_key(1),
9959 12 : Lsn(0x20),
9960 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9961 12 : ),
9962 12 : (
9963 12 : get_key(2),
9964 12 : Lsn(0x30),
9965 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9966 12 : ),
9967 12 : (
9968 12 : get_key(3),
9969 12 : Lsn(0x28),
9970 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9971 12 : ),
9972 12 : (
9973 12 : get_key(3),
9974 12 : Lsn(0x30),
9975 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9976 12 : ),
9977 12 : (
9978 12 : get_key(3),
9979 12 : Lsn(0x40),
9980 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9981 12 : ),
9982 12 : ];
9983 12 : let delta2 = vec![
9984 12 : (
9985 12 : get_key(5),
9986 12 : Lsn(0x20),
9987 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9988 12 : ),
9989 12 : (
9990 12 : get_key(6),
9991 12 : Lsn(0x20),
9992 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9993 12 : ),
9994 12 : ];
9995 12 : let delta3 = vec![
9996 12 : (
9997 12 : get_key(8),
9998 12 : Lsn(0x48),
9999 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10000 12 : ),
10001 12 : (
10002 12 : get_key(9),
10003 12 : Lsn(0x48),
10004 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10005 12 : ),
10006 12 : ];
10007 12 :
10008 12 : let tline = tenant
10009 12 : .create_test_timeline_with_layers(
10010 12 : TIMELINE_ID,
10011 12 : Lsn(0x10),
10012 12 : DEFAULT_PG_VERSION,
10013 12 : &ctx,
10014 12 : Vec::new(), // in-memory layers
10015 12 : vec![
10016 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
10017 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
10018 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
10019 12 : ], // delta layers
10020 12 : vec![(Lsn(0x10), img_layer)], // image layers
10021 12 : Lsn(0x50),
10022 12 : )
10023 12 : .await?;
10024 12 : {
10025 12 : tline
10026 12 : .applied_gc_cutoff_lsn
10027 12 : .lock_for_write()
10028 12 : .store_and_unlock(Lsn(0x30))
10029 12 : .wait()
10030 12 : .await;
10031 12 : // Update GC info
10032 12 : let mut guard = tline.gc_info.write().unwrap();
10033 12 : *guard = GcInfo {
10034 12 : retain_lsns: vec![
10035 12 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
10036 12 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
10037 12 : ],
10038 12 : cutoffs: GcCutoffs {
10039 12 : time: Lsn(0x30),
10040 12 : space: Lsn(0x30),
10041 12 : },
10042 12 : leases: Default::default(),
10043 12 : within_ancestor_pitr: false,
10044 12 : };
10045 12 : }
10046 12 :
10047 12 : let expected_result = [
10048 12 : Bytes::from_static(b"value 0@0x10"),
10049 12 : Bytes::from_static(b"value 1@0x10@0x20"),
10050 12 : Bytes::from_static(b"value 2@0x10@0x30"),
10051 12 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
10052 12 : Bytes::from_static(b"value 4@0x10"),
10053 12 : Bytes::from_static(b"value 5@0x10@0x20"),
10054 12 : Bytes::from_static(b"value 6@0x10@0x20"),
10055 12 : Bytes::from_static(b"value 7@0x10"),
10056 12 : Bytes::from_static(b"value 8@0x10@0x48"),
10057 12 : Bytes::from_static(b"value 9@0x10@0x48"),
10058 12 : ];
10059 12 :
10060 12 : let expected_result_at_gc_horizon = [
10061 12 : Bytes::from_static(b"value 0@0x10"),
10062 12 : Bytes::from_static(b"value 1@0x10@0x20"),
10063 12 : Bytes::from_static(b"value 2@0x10@0x30"),
10064 12 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
10065 12 : Bytes::from_static(b"value 4@0x10"),
10066 12 : Bytes::from_static(b"value 5@0x10@0x20"),
10067 12 : Bytes::from_static(b"value 6@0x10@0x20"),
10068 12 : Bytes::from_static(b"value 7@0x10"),
10069 12 : Bytes::from_static(b"value 8@0x10"),
10070 12 : Bytes::from_static(b"value 9@0x10"),
10071 12 : ];
10072 12 :
10073 12 : let expected_result_at_lsn_20 = [
10074 12 : Bytes::from_static(b"value 0@0x10"),
10075 12 : Bytes::from_static(b"value 1@0x10@0x20"),
10076 12 : Bytes::from_static(b"value 2@0x10"),
10077 12 : Bytes::from_static(b"value 3@0x10"),
10078 12 : Bytes::from_static(b"value 4@0x10"),
10079 12 : Bytes::from_static(b"value 5@0x10@0x20"),
10080 12 : Bytes::from_static(b"value 6@0x10@0x20"),
10081 12 : Bytes::from_static(b"value 7@0x10"),
10082 12 : Bytes::from_static(b"value 8@0x10"),
10083 12 : Bytes::from_static(b"value 9@0x10"),
10084 12 : ];
10085 12 :
10086 12 : let expected_result_at_lsn_10 = [
10087 12 : Bytes::from_static(b"value 0@0x10"),
10088 12 : Bytes::from_static(b"value 1@0x10"),
10089 12 : Bytes::from_static(b"value 2@0x10"),
10090 12 : Bytes::from_static(b"value 3@0x10"),
10091 12 : Bytes::from_static(b"value 4@0x10"),
10092 12 : Bytes::from_static(b"value 5@0x10"),
10093 12 : Bytes::from_static(b"value 6@0x10"),
10094 12 : Bytes::from_static(b"value 7@0x10"),
10095 12 : Bytes::from_static(b"value 8@0x10"),
10096 12 : Bytes::from_static(b"value 9@0x10"),
10097 12 : ];
10098 12 :
10099 72 : let verify_result = || async {
10100 72 : let gc_horizon = {
10101 72 : let gc_info = tline.gc_info.read().unwrap();
10102 72 : gc_info.cutoffs.time
10103 12 : };
10104 792 : for idx in 0..10 {
10105 720 : assert_eq!(
10106 720 : tline
10107 720 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10108 720 : .await
10109 720 : .unwrap(),
10110 720 : &expected_result[idx]
10111 12 : );
10112 720 : assert_eq!(
10113 720 : tline
10114 720 : .get(get_key(idx as u32), gc_horizon, &ctx)
10115 720 : .await
10116 720 : .unwrap(),
10117 720 : &expected_result_at_gc_horizon[idx]
10118 12 : );
10119 720 : assert_eq!(
10120 720 : tline
10121 720 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
10122 720 : .await
10123 720 : .unwrap(),
10124 720 : &expected_result_at_lsn_20[idx]
10125 12 : );
10126 720 : assert_eq!(
10127 720 : tline
10128 720 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
10129 720 : .await
10130 720 : .unwrap(),
10131 720 : &expected_result_at_lsn_10[idx]
10132 12 : );
10133 12 : }
10134 144 : };
10135 12 :
10136 12 : verify_result().await;
10137 12 :
10138 12 : let cancel = CancellationToken::new();
10139 12 : let mut dryrun_flags = EnumSet::new();
10140 12 : dryrun_flags.insert(CompactFlags::DryRun);
10141 12 :
10142 12 : tline
10143 12 : .compact_with_gc(
10144 12 : &cancel,
10145 12 : CompactOptions {
10146 12 : flags: dryrun_flags,
10147 12 : ..Default::default()
10148 12 : },
10149 12 : &ctx,
10150 12 : )
10151 12 : .await
10152 12 : .unwrap();
10153 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
10154 12 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
10155 12 : verify_result().await;
10156 12 :
10157 12 : tline
10158 12 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10159 12 : .await
10160 12 : .unwrap();
10161 12 : verify_result().await;
10162 12 :
10163 12 : // compact again
10164 12 : tline
10165 12 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10166 12 : .await
10167 12 : .unwrap();
10168 12 : verify_result().await;
10169 12 :
10170 12 : // increase GC horizon and compact again
10171 12 : {
10172 12 : tline
10173 12 : .applied_gc_cutoff_lsn
10174 12 : .lock_for_write()
10175 12 : .store_and_unlock(Lsn(0x38))
10176 12 : .wait()
10177 12 : .await;
10178 12 : // Update GC info
10179 12 : let mut guard = tline.gc_info.write().unwrap();
10180 12 : guard.cutoffs.time = Lsn(0x38);
10181 12 : guard.cutoffs.space = Lsn(0x38);
10182 12 : }
10183 12 : tline
10184 12 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10185 12 : .await
10186 12 : .unwrap();
10187 12 : verify_result().await; // no wals between 0x30 and 0x38, so we should obtain the same result
10188 12 :
10189 12 : // not increasing the GC horizon and compact again
10190 12 : tline
10191 12 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10192 12 : .await
10193 12 : .unwrap();
10194 12 : verify_result().await;
10195 12 :
10196 12 : Ok(())
10197 12 : }
10198 :
10199 : #[cfg(feature = "testing")]
10200 : #[tokio::test]
10201 12 : async fn test_simple_bottom_most_compaction_with_retain_lsns_single_key() -> anyhow::Result<()>
10202 12 : {
10203 12 : let harness =
10204 12 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns_single_key")
10205 12 : .await?;
10206 12 : let (tenant, ctx) = harness.load().await;
10207 12 :
10208 2112 : fn get_key(id: u32) -> Key {
10209 2112 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
10210 2112 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
10211 2112 : key.field6 = id;
10212 2112 : key
10213 2112 : }
10214 12 :
10215 12 : let img_layer = (0..10)
10216 120 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10217 12 : .collect_vec();
10218 12 :
10219 12 : let delta1 = vec![
10220 12 : (
10221 12 : get_key(1),
10222 12 : Lsn(0x20),
10223 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10224 12 : ),
10225 12 : (
10226 12 : get_key(1),
10227 12 : Lsn(0x28),
10228 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
10229 12 : ),
10230 12 : ];
10231 12 : let delta2 = vec![
10232 12 : (
10233 12 : get_key(1),
10234 12 : Lsn(0x30),
10235 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10236 12 : ),
10237 12 : (
10238 12 : get_key(1),
10239 12 : Lsn(0x38),
10240 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
10241 12 : ),
10242 12 : ];
10243 12 : let delta3 = vec![
10244 12 : (
10245 12 : get_key(8),
10246 12 : Lsn(0x48),
10247 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10248 12 : ),
10249 12 : (
10250 12 : get_key(9),
10251 12 : Lsn(0x48),
10252 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10253 12 : ),
10254 12 : ];
10255 12 :
10256 12 : let tline = tenant
10257 12 : .create_test_timeline_with_layers(
10258 12 : TIMELINE_ID,
10259 12 : Lsn(0x10),
10260 12 : DEFAULT_PG_VERSION,
10261 12 : &ctx,
10262 12 : Vec::new(), // in-memory layers
10263 12 : vec![
10264 12 : // delta1 and delta 2 only contain a single key but multiple updates
10265 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x30), delta1),
10266 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
10267 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x50), delta3),
10268 12 : ], // delta layers
10269 12 : vec![(Lsn(0x10), img_layer)], // image layers
10270 12 : Lsn(0x50),
10271 12 : )
10272 12 : .await?;
10273 12 : {
10274 12 : tline
10275 12 : .applied_gc_cutoff_lsn
10276 12 : .lock_for_write()
10277 12 : .store_and_unlock(Lsn(0x30))
10278 12 : .wait()
10279 12 : .await;
10280 12 : // Update GC info
10281 12 : let mut guard = tline.gc_info.write().unwrap();
10282 12 : *guard = GcInfo {
10283 12 : retain_lsns: vec![
10284 12 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
10285 12 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
10286 12 : ],
10287 12 : cutoffs: GcCutoffs {
10288 12 : time: Lsn(0x30),
10289 12 : space: Lsn(0x30),
10290 12 : },
10291 12 : leases: Default::default(),
10292 12 : within_ancestor_pitr: false,
10293 12 : };
10294 12 : }
10295 12 :
10296 12 : let expected_result = [
10297 12 : Bytes::from_static(b"value 0@0x10"),
10298 12 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
10299 12 : Bytes::from_static(b"value 2@0x10"),
10300 12 : Bytes::from_static(b"value 3@0x10"),
10301 12 : Bytes::from_static(b"value 4@0x10"),
10302 12 : Bytes::from_static(b"value 5@0x10"),
10303 12 : Bytes::from_static(b"value 6@0x10"),
10304 12 : Bytes::from_static(b"value 7@0x10"),
10305 12 : Bytes::from_static(b"value 8@0x10@0x48"),
10306 12 : Bytes::from_static(b"value 9@0x10@0x48"),
10307 12 : ];
10308 12 :
10309 12 : let expected_result_at_gc_horizon = [
10310 12 : Bytes::from_static(b"value 0@0x10"),
10311 12 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
10312 12 : Bytes::from_static(b"value 2@0x10"),
10313 12 : Bytes::from_static(b"value 3@0x10"),
10314 12 : Bytes::from_static(b"value 4@0x10"),
10315 12 : Bytes::from_static(b"value 5@0x10"),
10316 12 : Bytes::from_static(b"value 6@0x10"),
10317 12 : Bytes::from_static(b"value 7@0x10"),
10318 12 : Bytes::from_static(b"value 8@0x10"),
10319 12 : Bytes::from_static(b"value 9@0x10"),
10320 12 : ];
10321 12 :
10322 12 : let expected_result_at_lsn_20 = [
10323 12 : Bytes::from_static(b"value 0@0x10"),
10324 12 : Bytes::from_static(b"value 1@0x10@0x20"),
10325 12 : Bytes::from_static(b"value 2@0x10"),
10326 12 : Bytes::from_static(b"value 3@0x10"),
10327 12 : Bytes::from_static(b"value 4@0x10"),
10328 12 : Bytes::from_static(b"value 5@0x10"),
10329 12 : Bytes::from_static(b"value 6@0x10"),
10330 12 : Bytes::from_static(b"value 7@0x10"),
10331 12 : Bytes::from_static(b"value 8@0x10"),
10332 12 : Bytes::from_static(b"value 9@0x10"),
10333 12 : ];
10334 12 :
10335 12 : let expected_result_at_lsn_10 = [
10336 12 : Bytes::from_static(b"value 0@0x10"),
10337 12 : Bytes::from_static(b"value 1@0x10"),
10338 12 : Bytes::from_static(b"value 2@0x10"),
10339 12 : Bytes::from_static(b"value 3@0x10"),
10340 12 : Bytes::from_static(b"value 4@0x10"),
10341 12 : Bytes::from_static(b"value 5@0x10"),
10342 12 : Bytes::from_static(b"value 6@0x10"),
10343 12 : Bytes::from_static(b"value 7@0x10"),
10344 12 : Bytes::from_static(b"value 8@0x10"),
10345 12 : Bytes::from_static(b"value 9@0x10"),
10346 12 : ];
10347 12 :
10348 48 : let verify_result = || async {
10349 48 : let gc_horizon = {
10350 48 : let gc_info = tline.gc_info.read().unwrap();
10351 48 : gc_info.cutoffs.time
10352 12 : };
10353 528 : for idx in 0..10 {
10354 480 : assert_eq!(
10355 480 : tline
10356 480 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10357 480 : .await
10358 480 : .unwrap(),
10359 480 : &expected_result[idx]
10360 12 : );
10361 480 : assert_eq!(
10362 480 : tline
10363 480 : .get(get_key(idx as u32), gc_horizon, &ctx)
10364 480 : .await
10365 480 : .unwrap(),
10366 480 : &expected_result_at_gc_horizon[idx]
10367 12 : );
10368 480 : assert_eq!(
10369 480 : tline
10370 480 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
10371 480 : .await
10372 480 : .unwrap(),
10373 480 : &expected_result_at_lsn_20[idx]
10374 12 : );
10375 480 : assert_eq!(
10376 480 : tline
10377 480 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
10378 480 : .await
10379 480 : .unwrap(),
10380 480 : &expected_result_at_lsn_10[idx]
10381 12 : );
10382 12 : }
10383 96 : };
10384 12 :
10385 12 : verify_result().await;
10386 12 :
10387 12 : let cancel = CancellationToken::new();
10388 12 : let mut dryrun_flags = EnumSet::new();
10389 12 : dryrun_flags.insert(CompactFlags::DryRun);
10390 12 :
10391 12 : tline
10392 12 : .compact_with_gc(
10393 12 : &cancel,
10394 12 : CompactOptions {
10395 12 : flags: dryrun_flags,
10396 12 : ..Default::default()
10397 12 : },
10398 12 : &ctx,
10399 12 : )
10400 12 : .await
10401 12 : .unwrap();
10402 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
10403 12 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
10404 12 : verify_result().await;
10405 12 :
10406 12 : tline
10407 12 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10408 12 : .await
10409 12 : .unwrap();
10410 12 : verify_result().await;
10411 12 :
10412 12 : // compact again
10413 12 : tline
10414 12 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10415 12 : .await
10416 12 : .unwrap();
10417 12 : verify_result().await;
10418 12 :
10419 12 : Ok(())
10420 12 : }
10421 :
10422 : #[cfg(feature = "testing")]
10423 : #[tokio::test]
10424 12 : async fn test_simple_bottom_most_compaction_on_branch() -> anyhow::Result<()> {
10425 12 : use models::CompactLsnRange;
10426 12 :
10427 12 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_on_branch").await?;
10428 12 : let (tenant, ctx) = harness.load().await;
10429 12 :
10430 996 : fn get_key(id: u32) -> Key {
10431 996 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
10432 996 : key.field6 = id;
10433 996 : key
10434 996 : }
10435 12 :
10436 12 : let img_layer = (0..10)
10437 120 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10438 12 : .collect_vec();
10439 12 :
10440 12 : let delta1 = vec![
10441 12 : (
10442 12 : get_key(1),
10443 12 : Lsn(0x20),
10444 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10445 12 : ),
10446 12 : (
10447 12 : get_key(2),
10448 12 : Lsn(0x30),
10449 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10450 12 : ),
10451 12 : (
10452 12 : get_key(3),
10453 12 : Lsn(0x28),
10454 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
10455 12 : ),
10456 12 : (
10457 12 : get_key(3),
10458 12 : Lsn(0x30),
10459 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10460 12 : ),
10461 12 : (
10462 12 : get_key(3),
10463 12 : Lsn(0x40),
10464 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
10465 12 : ),
10466 12 : ];
10467 12 : let delta2 = vec![
10468 12 : (
10469 12 : get_key(5),
10470 12 : Lsn(0x20),
10471 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10472 12 : ),
10473 12 : (
10474 12 : get_key(6),
10475 12 : Lsn(0x20),
10476 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10477 12 : ),
10478 12 : ];
10479 12 : let delta3 = vec![
10480 12 : (
10481 12 : get_key(8),
10482 12 : Lsn(0x48),
10483 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10484 12 : ),
10485 12 : (
10486 12 : get_key(9),
10487 12 : Lsn(0x48),
10488 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10489 12 : ),
10490 12 : ];
10491 12 :
10492 12 : let parent_tline = tenant
10493 12 : .create_test_timeline_with_layers(
10494 12 : TIMELINE_ID,
10495 12 : Lsn(0x10),
10496 12 : DEFAULT_PG_VERSION,
10497 12 : &ctx,
10498 12 : vec![], // in-memory layers
10499 12 : vec![], // delta layers
10500 12 : vec![(Lsn(0x18), img_layer)], // image layers
10501 12 : Lsn(0x18),
10502 12 : )
10503 12 : .await?;
10504 12 :
10505 12 : parent_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
10506 12 :
10507 12 : let branch_tline = tenant
10508 12 : .branch_timeline_test_with_layers(
10509 12 : &parent_tline,
10510 12 : NEW_TIMELINE_ID,
10511 12 : Some(Lsn(0x18)),
10512 12 : &ctx,
10513 12 : vec![
10514 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
10515 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
10516 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
10517 12 : ], // delta layers
10518 12 : vec![], // image layers
10519 12 : Lsn(0x50),
10520 12 : )
10521 12 : .await?;
10522 12 :
10523 12 : branch_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
10524 12 :
10525 12 : {
10526 12 : parent_tline
10527 12 : .applied_gc_cutoff_lsn
10528 12 : .lock_for_write()
10529 12 : .store_and_unlock(Lsn(0x10))
10530 12 : .wait()
10531 12 : .await;
10532 12 : // Update GC info
10533 12 : let mut guard = parent_tline.gc_info.write().unwrap();
10534 12 : *guard = GcInfo {
10535 12 : retain_lsns: vec![(Lsn(0x18), branch_tline.timeline_id, MaybeOffloaded::No)],
10536 12 : cutoffs: GcCutoffs {
10537 12 : time: Lsn(0x10),
10538 12 : space: Lsn(0x10),
10539 12 : },
10540 12 : leases: Default::default(),
10541 12 : within_ancestor_pitr: false,
10542 12 : };
10543 12 : }
10544 12 :
10545 12 : {
10546 12 : branch_tline
10547 12 : .applied_gc_cutoff_lsn
10548 12 : .lock_for_write()
10549 12 : .store_and_unlock(Lsn(0x50))
10550 12 : .wait()
10551 12 : .await;
10552 12 : // Update GC info
10553 12 : let mut guard = branch_tline.gc_info.write().unwrap();
10554 12 : *guard = GcInfo {
10555 12 : retain_lsns: vec![(Lsn(0x40), branch_tline.timeline_id, MaybeOffloaded::No)],
10556 12 : cutoffs: GcCutoffs {
10557 12 : time: Lsn(0x50),
10558 12 : space: Lsn(0x50),
10559 12 : },
10560 12 : leases: Default::default(),
10561 12 : within_ancestor_pitr: false,
10562 12 : };
10563 12 : }
10564 12 :
10565 12 : let expected_result_at_gc_horizon = [
10566 12 : Bytes::from_static(b"value 0@0x10"),
10567 12 : Bytes::from_static(b"value 1@0x10@0x20"),
10568 12 : Bytes::from_static(b"value 2@0x10@0x30"),
10569 12 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
10570 12 : Bytes::from_static(b"value 4@0x10"),
10571 12 : Bytes::from_static(b"value 5@0x10@0x20"),
10572 12 : Bytes::from_static(b"value 6@0x10@0x20"),
10573 12 : Bytes::from_static(b"value 7@0x10"),
10574 12 : Bytes::from_static(b"value 8@0x10@0x48"),
10575 12 : Bytes::from_static(b"value 9@0x10@0x48"),
10576 12 : ];
10577 12 :
10578 12 : let expected_result_at_lsn_40 = [
10579 12 : Bytes::from_static(b"value 0@0x10"),
10580 12 : Bytes::from_static(b"value 1@0x10@0x20"),
10581 12 : Bytes::from_static(b"value 2@0x10@0x30"),
10582 12 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
10583 12 : Bytes::from_static(b"value 4@0x10"),
10584 12 : Bytes::from_static(b"value 5@0x10@0x20"),
10585 12 : Bytes::from_static(b"value 6@0x10@0x20"),
10586 12 : Bytes::from_static(b"value 7@0x10"),
10587 12 : Bytes::from_static(b"value 8@0x10"),
10588 12 : Bytes::from_static(b"value 9@0x10"),
10589 12 : ];
10590 12 :
10591 36 : let verify_result = || async {
10592 396 : for idx in 0..10 {
10593 360 : assert_eq!(
10594 360 : branch_tline
10595 360 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10596 360 : .await
10597 360 : .unwrap(),
10598 360 : &expected_result_at_gc_horizon[idx]
10599 12 : );
10600 360 : assert_eq!(
10601 360 : branch_tline
10602 360 : .get(get_key(idx as u32), Lsn(0x40), &ctx)
10603 360 : .await
10604 360 : .unwrap(),
10605 360 : &expected_result_at_lsn_40[idx]
10606 12 : );
10607 12 : }
10608 72 : };
10609 12 :
10610 12 : verify_result().await;
10611 12 :
10612 12 : let cancel = CancellationToken::new();
10613 12 : branch_tline
10614 12 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10615 12 : .await
10616 12 : .unwrap();
10617 12 :
10618 12 : verify_result().await;
10619 12 :
10620 12 : // Piggyback a compaction with above_lsn. Ensure it works correctly when the specified LSN intersects with the layer files.
10621 12 : // Now we already have a single large delta layer, so the compaction min_layer_lsn should be the same as ancestor LSN (0x18).
10622 12 : branch_tline
10623 12 : .compact_with_gc(
10624 12 : &cancel,
10625 12 : CompactOptions {
10626 12 : compact_lsn_range: Some(CompactLsnRange::above(Lsn(0x40))),
10627 12 : ..Default::default()
10628 12 : },
10629 12 : &ctx,
10630 12 : )
10631 12 : .await
10632 12 : .unwrap();
10633 12 :
10634 12 : verify_result().await;
10635 12 :
10636 12 : Ok(())
10637 12 : }
10638 :
10639 : // Regression test for https://github.com/neondatabase/neon/issues/9012
10640 : // Create an image arrangement where we have to read at different LSN ranges
10641 : // from a delta layer. This is achieved by overlapping an image layer on top of
10642 : // a delta layer. Like so:
10643 : //
10644 : // A B
10645 : // +----------------+ -> delta_layer
10646 : // | | ^ lsn
10647 : // | =========|-> nested_image_layer |
10648 : // | C | |
10649 : // +----------------+ |
10650 : // ======== -> baseline_image_layer +-------> key
10651 : //
10652 : //
10653 : // When querying the key range [A, B) we need to read at different LSN ranges
10654 : // for [A, C) and [C, B). This test checks that the described edge case is handled correctly.
10655 : #[cfg(feature = "testing")]
10656 : #[tokio::test]
10657 12 : async fn test_vectored_read_with_nested_image_layer() -> anyhow::Result<()> {
10658 12 : let harness = TenantHarness::create("test_vectored_read_with_nested_image_layer").await?;
10659 12 : let (tenant, ctx) = harness.load().await;
10660 12 :
10661 12 : let will_init_keys = [2, 6];
10662 264 : fn get_key(id: u32) -> Key {
10663 264 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
10664 264 : key.field6 = id;
10665 264 : key
10666 264 : }
10667 12 :
10668 12 : let mut expected_key_values = HashMap::new();
10669 12 :
10670 12 : let baseline_image_layer_lsn = Lsn(0x10);
10671 12 : let mut baseline_img_layer = Vec::new();
10672 72 : for i in 0..5 {
10673 60 : let key = get_key(i);
10674 60 : let value = format!("value {i}@{baseline_image_layer_lsn}");
10675 60 :
10676 60 : let removed = expected_key_values.insert(key, value.clone());
10677 60 : assert!(removed.is_none());
10678 12 :
10679 60 : baseline_img_layer.push((key, Bytes::from(value)));
10680 12 : }
10681 12 :
10682 12 : let nested_image_layer_lsn = Lsn(0x50);
10683 12 : let mut nested_img_layer = Vec::new();
10684 72 : for i in 5..10 {
10685 60 : let key = get_key(i);
10686 60 : let value = format!("value {i}@{nested_image_layer_lsn}");
10687 60 :
10688 60 : let removed = expected_key_values.insert(key, value.clone());
10689 60 : assert!(removed.is_none());
10690 12 :
10691 60 : nested_img_layer.push((key, Bytes::from(value)));
10692 12 : }
10693 12 :
10694 12 : let mut delta_layer_spec = Vec::default();
10695 12 : let delta_layer_start_lsn = Lsn(0x20);
10696 12 : let mut delta_layer_end_lsn = delta_layer_start_lsn;
10697 12 :
10698 132 : for i in 0..10 {
10699 120 : let key = get_key(i);
10700 120 : let key_in_nested = nested_img_layer
10701 120 : .iter()
10702 480 : .any(|(key_with_img, _)| *key_with_img == key);
10703 120 : let lsn = {
10704 120 : if key_in_nested {
10705 60 : Lsn(nested_image_layer_lsn.0 + 0x10)
10706 12 : } else {
10707 60 : delta_layer_start_lsn
10708 12 : }
10709 12 : };
10710 12 :
10711 120 : let will_init = will_init_keys.contains(&i);
10712 120 : if will_init {
10713 24 : delta_layer_spec.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init(""))));
10714 24 :
10715 24 : expected_key_values.insert(key, "".to_string());
10716 96 : } else {
10717 96 : let delta = format!("@{lsn}");
10718 96 : delta_layer_spec.push((
10719 96 : key,
10720 96 : lsn,
10721 96 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
10722 96 : ));
10723 96 :
10724 96 : expected_key_values
10725 96 : .get_mut(&key)
10726 96 : .expect("An image exists for each key")
10727 96 : .push_str(delta.as_str());
10728 96 : }
10729 120 : delta_layer_end_lsn = std::cmp::max(delta_layer_start_lsn, lsn);
10730 12 : }
10731 12 :
10732 12 : delta_layer_end_lsn = Lsn(delta_layer_end_lsn.0 + 1);
10733 12 :
10734 12 : assert!(
10735 12 : nested_image_layer_lsn > delta_layer_start_lsn
10736 12 : && nested_image_layer_lsn < delta_layer_end_lsn
10737 12 : );
10738 12 :
10739 12 : let tline = tenant
10740 12 : .create_test_timeline_with_layers(
10741 12 : TIMELINE_ID,
10742 12 : baseline_image_layer_lsn,
10743 12 : DEFAULT_PG_VERSION,
10744 12 : &ctx,
10745 12 : vec![], // in-memory layers
10746 12 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
10747 12 : delta_layer_start_lsn..delta_layer_end_lsn,
10748 12 : delta_layer_spec,
10749 12 : )], // delta layers
10750 12 : vec![
10751 12 : (baseline_image_layer_lsn, baseline_img_layer),
10752 12 : (nested_image_layer_lsn, nested_img_layer),
10753 12 : ], // image layers
10754 12 : delta_layer_end_lsn,
10755 12 : )
10756 12 : .await?;
10757 12 :
10758 12 : let query = VersionedKeySpaceQuery::uniform(
10759 12 : KeySpace::single(get_key(0)..get_key(10)),
10760 12 : delta_layer_end_lsn,
10761 12 : );
10762 12 :
10763 12 : let results = tline
10764 12 : .get_vectored(query, IoConcurrency::sequential(), &ctx)
10765 12 : .await
10766 12 : .expect("No vectored errors");
10767 132 : for (key, res) in results {
10768 120 : let value = res.expect("No key errors");
10769 120 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
10770 120 : assert_eq!(value, Bytes::from(expected_value));
10771 12 : }
10772 12 :
10773 12 : Ok(())
10774 12 : }
10775 :
10776 : #[cfg(feature = "testing")]
10777 : #[tokio::test]
10778 12 : async fn test_vectored_read_with_image_layer_inside_inmem() -> anyhow::Result<()> {
10779 12 : let harness =
10780 12 : TenantHarness::create("test_vectored_read_with_image_layer_inside_inmem").await?;
10781 12 : let (tenant, ctx) = harness.load().await;
10782 12 :
10783 12 : let will_init_keys = [2, 6];
10784 384 : fn get_key(id: u32) -> Key {
10785 384 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
10786 384 : key.field6 = id;
10787 384 : key
10788 384 : }
10789 12 :
10790 12 : let mut expected_key_values = HashMap::new();
10791 12 :
10792 12 : let baseline_image_layer_lsn = Lsn(0x10);
10793 12 : let mut baseline_img_layer = Vec::new();
10794 72 : for i in 0..5 {
10795 60 : let key = get_key(i);
10796 60 : let value = format!("value {i}@{baseline_image_layer_lsn}");
10797 60 :
10798 60 : let removed = expected_key_values.insert(key, value.clone());
10799 60 : assert!(removed.is_none());
10800 12 :
10801 60 : baseline_img_layer.push((key, Bytes::from(value)));
10802 12 : }
10803 12 :
10804 12 : let nested_image_layer_lsn = Lsn(0x50);
10805 12 : let mut nested_img_layer = Vec::new();
10806 72 : for i in 5..10 {
10807 60 : let key = get_key(i);
10808 60 : let value = format!("value {i}@{nested_image_layer_lsn}");
10809 60 :
10810 60 : let removed = expected_key_values.insert(key, value.clone());
10811 60 : assert!(removed.is_none());
10812 12 :
10813 60 : nested_img_layer.push((key, Bytes::from(value)));
10814 12 : }
10815 12 :
10816 12 : let frozen_layer = {
10817 12 : let lsn_range = Lsn(0x40)..Lsn(0x60);
10818 12 : let mut data = Vec::new();
10819 132 : for i in 0..10 {
10820 120 : let key = get_key(i);
10821 120 : let key_in_nested = nested_img_layer
10822 120 : .iter()
10823 480 : .any(|(key_with_img, _)| *key_with_img == key);
10824 120 : let lsn = {
10825 120 : if key_in_nested {
10826 60 : Lsn(nested_image_layer_lsn.0 + 5)
10827 12 : } else {
10828 60 : lsn_range.start
10829 12 : }
10830 12 : };
10831 12 :
10832 120 : let will_init = will_init_keys.contains(&i);
10833 120 : if will_init {
10834 24 : data.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init(""))));
10835 24 :
10836 24 : expected_key_values.insert(key, "".to_string());
10837 96 : } else {
10838 96 : let delta = format!("@{lsn}");
10839 96 : data.push((
10840 96 : key,
10841 96 : lsn,
10842 96 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
10843 96 : ));
10844 96 :
10845 96 : expected_key_values
10846 96 : .get_mut(&key)
10847 96 : .expect("An image exists for each key")
10848 96 : .push_str(delta.as_str());
10849 96 : }
10850 12 : }
10851 12 :
10852 12 : InMemoryLayerTestDesc {
10853 12 : lsn_range,
10854 12 : is_open: false,
10855 12 : data,
10856 12 : }
10857 12 : };
10858 12 :
10859 12 : let (open_layer, last_record_lsn) = {
10860 12 : let start_lsn = Lsn(0x70);
10861 12 : let mut data = Vec::new();
10862 12 : let mut end_lsn = Lsn(0);
10863 132 : for i in 0..10 {
10864 120 : let key = get_key(i);
10865 120 : let lsn = Lsn(start_lsn.0 + i as u64);
10866 120 : let delta = format!("@{lsn}");
10867 120 : data.push((
10868 120 : key,
10869 120 : lsn,
10870 120 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
10871 120 : ));
10872 120 :
10873 120 : expected_key_values
10874 120 : .get_mut(&key)
10875 120 : .expect("An image exists for each key")
10876 120 : .push_str(delta.as_str());
10877 120 :
10878 120 : end_lsn = std::cmp::max(end_lsn, lsn);
10879 120 : }
10880 12 :
10881 12 : (
10882 12 : InMemoryLayerTestDesc {
10883 12 : lsn_range: start_lsn..Lsn::MAX,
10884 12 : is_open: true,
10885 12 : data,
10886 12 : },
10887 12 : end_lsn,
10888 12 : )
10889 12 : };
10890 12 :
10891 12 : assert!(
10892 12 : nested_image_layer_lsn > frozen_layer.lsn_range.start
10893 12 : && nested_image_layer_lsn < frozen_layer.lsn_range.end
10894 12 : );
10895 12 :
10896 12 : let tline = tenant
10897 12 : .create_test_timeline_with_layers(
10898 12 : TIMELINE_ID,
10899 12 : baseline_image_layer_lsn,
10900 12 : DEFAULT_PG_VERSION,
10901 12 : &ctx,
10902 12 : vec![open_layer, frozen_layer], // in-memory layers
10903 12 : Vec::new(), // delta layers
10904 12 : vec![
10905 12 : (baseline_image_layer_lsn, baseline_img_layer),
10906 12 : (nested_image_layer_lsn, nested_img_layer),
10907 12 : ], // image layers
10908 12 : last_record_lsn,
10909 12 : )
10910 12 : .await?;
10911 12 :
10912 12 : let query = VersionedKeySpaceQuery::uniform(
10913 12 : KeySpace::single(get_key(0)..get_key(10)),
10914 12 : last_record_lsn,
10915 12 : );
10916 12 :
10917 12 : let results = tline
10918 12 : .get_vectored(query, IoConcurrency::sequential(), &ctx)
10919 12 : .await
10920 12 : .expect("No vectored errors");
10921 132 : for (key, res) in results {
10922 120 : let value = res.expect("No key errors");
10923 120 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
10924 120 : assert_eq!(value, Bytes::from(expected_value.clone()));
10925 12 :
10926 120 : tracing::info!("key={key} value={expected_value}");
10927 12 : }
10928 12 :
10929 12 : Ok(())
10930 12 : }
10931 :
10932 : // A randomized read path test. Generates a layer map according to a deterministic
10933 : // specification. Fills the (key, LSN) space in random manner and then performs
10934 : // random scattered queries validating the results against in-memory storage.
10935 : //
10936 : // See this internal Notion page for a diagram of the layer map:
10937 : // https://www.notion.so/neondatabase/Read-Path-Unit-Testing-Fuzzing-1d1f189e0047806c8e5cd37781b0a350?pvs=4
10938 : //
10939 : // A fuzzing mode is also supported. In this mode, the test will use a random
10940 : // seed instead of a hardcoded one. Use it in conjunction with `cargo stress`
10941 : // to run multiple instances in parallel:
10942 : //
10943 : // $ RUST_BACKTRACE=1 RUST_LOG=INFO \
10944 : // cargo stress --package=pageserver --features=testing,fuzz-read-path --release -- test_read_path
10945 : #[cfg(feature = "testing")]
10946 : #[tokio::test]
10947 12 : async fn test_read_path() -> anyhow::Result<()> {
10948 12 : use rand::seq::SliceRandom;
10949 12 :
10950 12 : let seed = if cfg!(feature = "fuzz-read-path") {
10951 12 : let seed: u64 = thread_rng().r#gen();
10952 0 : seed
10953 12 : } else {
10954 12 : // Use a hard-coded seed when not in fuzzing mode.
10955 12 : // Note that with the current approach results are not reproducible
10956 12 : // accross platforms and Rust releases.
10957 12 : const SEED: u64 = 0;
10958 12 : SEED
10959 12 : };
10960 12 :
10961 12 : let mut random = StdRng::seed_from_u64(seed);
10962 12 :
10963 12 : let (queries, will_init_chance, gap_chance) = if cfg!(feature = "fuzz-read-path") {
10964 12 : const QUERIES: u64 = 5000;
10965 12 : let will_init_chance: u8 = random.gen_range(0..=10);
10966 0 : let gap_chance: u8 = random.gen_range(0..=50);
10967 0 :
10968 0 : (QUERIES, will_init_chance, gap_chance)
10969 12 : } else {
10970 12 : const QUERIES: u64 = 1000;
10971 12 : const WILL_INIT_CHANCE: u8 = 1;
10972 12 : const GAP_CHANCE: u8 = 5;
10973 12 :
10974 12 : (QUERIES, WILL_INIT_CHANCE, GAP_CHANCE)
10975 12 : };
10976 12 :
10977 12 : let harness = TenantHarness::create("test_read_path").await?;
10978 12 : let (tenant, ctx) = harness.load().await;
10979 12 :
10980 12 : tracing::info!("Using random seed: {seed}");
10981 12 : tracing::info!(%will_init_chance, %gap_chance, "Fill params");
10982 12 :
10983 12 : // Define the layer map shape. Note that this part is not randomized.
10984 12 :
10985 12 : const KEY_DIMENSION_SIZE: u32 = 99;
10986 12 : let start_key = Key::from_hex("110000000033333333444444445500000000").unwrap();
10987 12 : let end_key = start_key.add(KEY_DIMENSION_SIZE);
10988 12 : let total_key_range = start_key..end_key;
10989 12 : let total_key_range_size = end_key.to_i128() - start_key.to_i128();
10990 12 : let total_start_lsn = Lsn(104);
10991 12 : let last_record_lsn = Lsn(504);
10992 12 :
10993 12 : assert!(total_key_range_size % 3 == 0);
10994 12 :
10995 12 : let in_memory_layers_shape = vec![
10996 12 : (total_key_range.clone(), Lsn(304)..Lsn(400)),
10997 12 : (total_key_range.clone(), Lsn(400)..last_record_lsn),
10998 12 : ];
10999 12 :
11000 12 : let delta_layers_shape = vec![
11001 12 : (
11002 12 : start_key..(start_key.add((total_key_range_size / 3) as u32)),
11003 12 : Lsn(200)..Lsn(304),
11004 12 : ),
11005 12 : (
11006 12 : (start_key.add((total_key_range_size / 3) as u32))
11007 12 : ..(start_key.add((total_key_range_size * 2 / 3) as u32)),
11008 12 : Lsn(200)..Lsn(304),
11009 12 : ),
11010 12 : (
11011 12 : (start_key.add((total_key_range_size * 2 / 3) as u32))
11012 12 : ..(start_key.add(total_key_range_size as u32)),
11013 12 : Lsn(200)..Lsn(304),
11014 12 : ),
11015 12 : ];
11016 12 :
11017 12 : let image_layers_shape = vec![
11018 12 : (
11019 12 : start_key.add((total_key_range_size * 2 / 3 - 10) as u32)
11020 12 : ..start_key.add((total_key_range_size * 2 / 3 + 10) as u32),
11021 12 : Lsn(456),
11022 12 : ),
11023 12 : (
11024 12 : start_key.add((total_key_range_size / 3 - 10) as u32)
11025 12 : ..start_key.add((total_key_range_size / 3 + 10) as u32),
11026 12 : Lsn(256),
11027 12 : ),
11028 12 : (total_key_range.clone(), total_start_lsn),
11029 12 : ];
11030 12 :
11031 12 : let specification = TestTimelineSpecification {
11032 12 : start_lsn: total_start_lsn,
11033 12 : last_record_lsn,
11034 12 : in_memory_layers_shape,
11035 12 : delta_layers_shape,
11036 12 : image_layers_shape,
11037 12 : gap_chance,
11038 12 : will_init_chance,
11039 12 : };
11040 12 :
11041 12 : // Create and randomly fill in the layers according to the specification
11042 12 : let (tline, storage, interesting_lsns) = randomize_timeline(
11043 12 : &tenant,
11044 12 : TIMELINE_ID,
11045 12 : DEFAULT_PG_VERSION,
11046 12 : specification,
11047 12 : &mut random,
11048 12 : &ctx,
11049 12 : )
11050 12 : .await?;
11051 12 :
11052 12 : // Now generate queries based on the interesting lsns that we've collected.
11053 12 : //
11054 12 : // While there's still room in the query, pick and interesting LSN and a random
11055 12 : // key. Then roll the dice to see if the next key should also be included in
11056 12 : // the query. When the roll fails, break the "batch" and pick another point in the
11057 12 : // (key, LSN) space.
11058 12 :
11059 12 : const PICK_NEXT_CHANCE: u8 = 50;
11060 12 : for _ in 0..queries {
11061 12000 : let query = {
11062 12000 : let mut keyspaces_at_lsn: HashMap<Lsn, KeySpaceRandomAccum> = HashMap::default();
11063 12000 : let mut used_keys: HashSet<Key> = HashSet::default();
11064 12 :
11065 270432 : while used_keys.len() < Timeline::MAX_GET_VECTORED_KEYS as usize {
11066 258432 : let selected_lsn = interesting_lsns.choose(&mut random).expect("not empty");
11067 258432 : let mut selected_key = start_key.add(random.gen_range(0..KEY_DIMENSION_SIZE));
11068 12 :
11069 451368 : while used_keys.len() < Timeline::MAX_GET_VECTORED_KEYS as usize {
11070 445116 : if used_keys.contains(&selected_key)
11071 385848 : || selected_key >= start_key.add(KEY_DIMENSION_SIZE)
11072 12 : {
11073 61116 : break;
11074 384000 : }
11075 384000 :
11076 384000 : keyspaces_at_lsn
11077 384000 : .entry(*selected_lsn)
11078 384000 : .or_default()
11079 384000 : .add_key(selected_key);
11080 384000 : used_keys.insert(selected_key);
11081 384000 :
11082 384000 : let pick_next = random.gen_range(0..=100) <= PICK_NEXT_CHANCE;
11083 384000 : if pick_next {
11084 192936 : selected_key = selected_key.next();
11085 192936 : } else {
11086 191064 : break;
11087 12 : }
11088 12 : }
11089 12 : }
11090 12 :
11091 12000 : VersionedKeySpaceQuery::scattered(
11092 12000 : keyspaces_at_lsn
11093 12000 : .into_iter()
11094 143004 : .map(|(lsn, acc)| (lsn, acc.to_keyspace()))
11095 12000 : .collect(),
11096 12000 : )
11097 12 : };
11098 12 :
11099 12 : // Run the query and validate the results
11100 12 :
11101 12000 : let results = tline
11102 12000 : .get_vectored(query.clone(), IoConcurrency::Sequential, &ctx)
11103 12000 : .await;
11104 12 :
11105 12000 : let blobs = match results {
11106 12000 : Ok(ok) => ok,
11107 12 : Err(err) => {
11108 0 : panic!("seed={seed} Error returned for query {query}: {err}");
11109 12 : }
11110 12 : };
11111 12 :
11112 384000 : for (key, key_res) in blobs.into_iter() {
11113 384000 : match key_res {
11114 384000 : Ok(blob) => {
11115 384000 : let requested_at_lsn = query.map_key_to_lsn(&key);
11116 384000 : let expected = storage.get(key, requested_at_lsn);
11117 384000 :
11118 384000 : if blob != expected {
11119 12 : tracing::error!(
11120 12 : "seed={seed} Mismatch for {key}@{requested_at_lsn} from query: {query}"
11121 12 : );
11122 384000 : }
11123 12 :
11124 384000 : assert_eq!(blob, expected);
11125 12 : }
11126 12 : Err(err) => {
11127 0 : let requested_at_lsn = query.map_key_to_lsn(&key);
11128 0 :
11129 0 : panic!(
11130 0 : "seed={seed} Error returned for {key}@{requested_at_lsn} from query {query}: {err}"
11131 0 : );
11132 12 : }
11133 12 : }
11134 12 : }
11135 12 : }
11136 12 :
11137 12 : Ok(())
11138 12 : }
11139 :
11140 1284 : fn sort_layer_key(k1: &PersistentLayerKey, k2: &PersistentLayerKey) -> std::cmp::Ordering {
11141 1284 : (
11142 1284 : k1.is_delta,
11143 1284 : k1.key_range.start,
11144 1284 : k1.key_range.end,
11145 1284 : k1.lsn_range.start,
11146 1284 : k1.lsn_range.end,
11147 1284 : )
11148 1284 : .cmp(&(
11149 1284 : k2.is_delta,
11150 1284 : k2.key_range.start,
11151 1284 : k2.key_range.end,
11152 1284 : k2.lsn_range.start,
11153 1284 : k2.lsn_range.end,
11154 1284 : ))
11155 1284 : }
11156 :
11157 144 : async fn inspect_and_sort(
11158 144 : tline: &Arc<Timeline>,
11159 144 : filter: Option<std::ops::Range<Key>>,
11160 144 : ) -> Vec<PersistentLayerKey> {
11161 144 : let mut all_layers = tline.inspect_historic_layers().await.unwrap();
11162 144 : if let Some(filter) = filter {
11163 648 : all_layers.retain(|layer| overlaps_with(&layer.key_range, &filter));
11164 132 : }
11165 144 : all_layers.sort_by(sort_layer_key);
11166 144 : all_layers
11167 144 : }
11168 :
11169 : #[cfg(feature = "testing")]
11170 132 : fn check_layer_map_key_eq(
11171 132 : mut left: Vec<PersistentLayerKey>,
11172 132 : mut right: Vec<PersistentLayerKey>,
11173 132 : ) {
11174 132 : left.sort_by(sort_layer_key);
11175 132 : right.sort_by(sort_layer_key);
11176 132 : if left != right {
11177 0 : eprintln!("---LEFT---");
11178 0 : for left in left.iter() {
11179 0 : eprintln!("{}", left);
11180 0 : }
11181 0 : eprintln!("---RIGHT---");
11182 0 : for right in right.iter() {
11183 0 : eprintln!("{}", right);
11184 0 : }
11185 0 : assert_eq!(left, right);
11186 132 : }
11187 132 : }
11188 :
11189 : #[cfg(feature = "testing")]
11190 : #[tokio::test]
11191 12 : async fn test_simple_partial_bottom_most_compaction() -> anyhow::Result<()> {
11192 12 : let harness = TenantHarness::create("test_simple_partial_bottom_most_compaction").await?;
11193 12 : let (tenant, ctx) = harness.load().await;
11194 12 :
11195 1092 : fn get_key(id: u32) -> Key {
11196 1092 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
11197 1092 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
11198 1092 : key.field6 = id;
11199 1092 : key
11200 1092 : }
11201 12 :
11202 12 : // img layer at 0x10
11203 12 : let img_layer = (0..10)
11204 120 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
11205 12 : .collect_vec();
11206 12 :
11207 12 : let delta1 = vec![
11208 12 : (
11209 12 : get_key(1),
11210 12 : Lsn(0x20),
11211 12 : Value::Image(Bytes::from("value 1@0x20")),
11212 12 : ),
11213 12 : (
11214 12 : get_key(2),
11215 12 : Lsn(0x30),
11216 12 : Value::Image(Bytes::from("value 2@0x30")),
11217 12 : ),
11218 12 : (
11219 12 : get_key(3),
11220 12 : Lsn(0x40),
11221 12 : Value::Image(Bytes::from("value 3@0x40")),
11222 12 : ),
11223 12 : ];
11224 12 : let delta2 = vec![
11225 12 : (
11226 12 : get_key(5),
11227 12 : Lsn(0x20),
11228 12 : Value::Image(Bytes::from("value 5@0x20")),
11229 12 : ),
11230 12 : (
11231 12 : get_key(6),
11232 12 : Lsn(0x20),
11233 12 : Value::Image(Bytes::from("value 6@0x20")),
11234 12 : ),
11235 12 : ];
11236 12 : let delta3 = vec![
11237 12 : (
11238 12 : get_key(8),
11239 12 : Lsn(0x48),
11240 12 : Value::Image(Bytes::from("value 8@0x48")),
11241 12 : ),
11242 12 : (
11243 12 : get_key(9),
11244 12 : Lsn(0x48),
11245 12 : Value::Image(Bytes::from("value 9@0x48")),
11246 12 : ),
11247 12 : ];
11248 12 :
11249 12 : let tline = tenant
11250 12 : .create_test_timeline_with_layers(
11251 12 : TIMELINE_ID,
11252 12 : Lsn(0x10),
11253 12 : DEFAULT_PG_VERSION,
11254 12 : &ctx,
11255 12 : vec![], // in-memory layers
11256 12 : vec![
11257 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
11258 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
11259 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
11260 12 : ], // delta layers
11261 12 : vec![(Lsn(0x10), img_layer)], // image layers
11262 12 : Lsn(0x50),
11263 12 : )
11264 12 : .await?;
11265 12 :
11266 12 : {
11267 12 : tline
11268 12 : .applied_gc_cutoff_lsn
11269 12 : .lock_for_write()
11270 12 : .store_and_unlock(Lsn(0x30))
11271 12 : .wait()
11272 12 : .await;
11273 12 : // Update GC info
11274 12 : let mut guard = tline.gc_info.write().unwrap();
11275 12 : *guard = GcInfo {
11276 12 : retain_lsns: vec![(Lsn(0x20), tline.timeline_id, MaybeOffloaded::No)],
11277 12 : cutoffs: GcCutoffs {
11278 12 : time: Lsn(0x30),
11279 12 : space: Lsn(0x30),
11280 12 : },
11281 12 : leases: Default::default(),
11282 12 : within_ancestor_pitr: false,
11283 12 : };
11284 12 : }
11285 12 :
11286 12 : let cancel = CancellationToken::new();
11287 12 :
11288 12 : // Do a partial compaction on key range 0..2
11289 12 : tline
11290 12 : .compact_with_gc(
11291 12 : &cancel,
11292 12 : CompactOptions {
11293 12 : flags: EnumSet::new(),
11294 12 : compact_key_range: Some((get_key(0)..get_key(2)).into()),
11295 12 : ..Default::default()
11296 12 : },
11297 12 : &ctx,
11298 12 : )
11299 12 : .await
11300 12 : .unwrap();
11301 12 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11302 12 : check_layer_map_key_eq(
11303 12 : all_layers,
11304 12 : vec![
11305 12 : // newly-generated image layer for the partial compaction range 0-2
11306 12 : PersistentLayerKey {
11307 12 : key_range: get_key(0)..get_key(2),
11308 12 : lsn_range: Lsn(0x20)..Lsn(0x21),
11309 12 : is_delta: false,
11310 12 : },
11311 12 : PersistentLayerKey {
11312 12 : key_range: get_key(0)..get_key(10),
11313 12 : lsn_range: Lsn(0x10)..Lsn(0x11),
11314 12 : is_delta: false,
11315 12 : },
11316 12 : // delta1 is split and the second part is rewritten
11317 12 : PersistentLayerKey {
11318 12 : key_range: get_key(2)..get_key(4),
11319 12 : lsn_range: Lsn(0x20)..Lsn(0x48),
11320 12 : is_delta: true,
11321 12 : },
11322 12 : PersistentLayerKey {
11323 12 : key_range: get_key(5)..get_key(7),
11324 12 : lsn_range: Lsn(0x20)..Lsn(0x48),
11325 12 : is_delta: true,
11326 12 : },
11327 12 : PersistentLayerKey {
11328 12 : key_range: get_key(8)..get_key(10),
11329 12 : lsn_range: Lsn(0x48)..Lsn(0x50),
11330 12 : is_delta: true,
11331 12 : },
11332 12 : ],
11333 12 : );
11334 12 :
11335 12 : // Do a partial compaction on key range 2..4
11336 12 : tline
11337 12 : .compact_with_gc(
11338 12 : &cancel,
11339 12 : CompactOptions {
11340 12 : flags: EnumSet::new(),
11341 12 : compact_key_range: Some((get_key(2)..get_key(4)).into()),
11342 12 : ..Default::default()
11343 12 : },
11344 12 : &ctx,
11345 12 : )
11346 12 : .await
11347 12 : .unwrap();
11348 12 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11349 12 : check_layer_map_key_eq(
11350 12 : all_layers,
11351 12 : vec![
11352 12 : PersistentLayerKey {
11353 12 : key_range: get_key(0)..get_key(2),
11354 12 : lsn_range: Lsn(0x20)..Lsn(0x21),
11355 12 : is_delta: false,
11356 12 : },
11357 12 : PersistentLayerKey {
11358 12 : key_range: get_key(0)..get_key(10),
11359 12 : lsn_range: Lsn(0x10)..Lsn(0x11),
11360 12 : is_delta: false,
11361 12 : },
11362 12 : // image layer generated for the compaction range 2-4
11363 12 : PersistentLayerKey {
11364 12 : key_range: get_key(2)..get_key(4),
11365 12 : lsn_range: Lsn(0x20)..Lsn(0x21),
11366 12 : is_delta: false,
11367 12 : },
11368 12 : // we have key2/key3 above the retain_lsn, so we still need this delta layer
11369 12 : PersistentLayerKey {
11370 12 : key_range: get_key(2)..get_key(4),
11371 12 : lsn_range: Lsn(0x20)..Lsn(0x48),
11372 12 : is_delta: true,
11373 12 : },
11374 12 : PersistentLayerKey {
11375 12 : key_range: get_key(5)..get_key(7),
11376 12 : lsn_range: Lsn(0x20)..Lsn(0x48),
11377 12 : is_delta: true,
11378 12 : },
11379 12 : PersistentLayerKey {
11380 12 : key_range: get_key(8)..get_key(10),
11381 12 : lsn_range: Lsn(0x48)..Lsn(0x50),
11382 12 : is_delta: true,
11383 12 : },
11384 12 : ],
11385 12 : );
11386 12 :
11387 12 : // Do a partial compaction on key range 4..9
11388 12 : tline
11389 12 : .compact_with_gc(
11390 12 : &cancel,
11391 12 : CompactOptions {
11392 12 : flags: EnumSet::new(),
11393 12 : compact_key_range: Some((get_key(4)..get_key(9)).into()),
11394 12 : ..Default::default()
11395 12 : },
11396 12 : &ctx,
11397 12 : )
11398 12 : .await
11399 12 : .unwrap();
11400 12 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11401 12 : check_layer_map_key_eq(
11402 12 : all_layers,
11403 12 : vec![
11404 12 : PersistentLayerKey {
11405 12 : key_range: get_key(0)..get_key(2),
11406 12 : lsn_range: Lsn(0x20)..Lsn(0x21),
11407 12 : is_delta: false,
11408 12 : },
11409 12 : PersistentLayerKey {
11410 12 : key_range: get_key(0)..get_key(10),
11411 12 : lsn_range: Lsn(0x10)..Lsn(0x11),
11412 12 : is_delta: false,
11413 12 : },
11414 12 : PersistentLayerKey {
11415 12 : key_range: get_key(2)..get_key(4),
11416 12 : lsn_range: Lsn(0x20)..Lsn(0x21),
11417 12 : is_delta: false,
11418 12 : },
11419 12 : PersistentLayerKey {
11420 12 : key_range: get_key(2)..get_key(4),
11421 12 : lsn_range: Lsn(0x20)..Lsn(0x48),
11422 12 : is_delta: true,
11423 12 : },
11424 12 : // image layer generated for this compaction range
11425 12 : PersistentLayerKey {
11426 12 : key_range: get_key(4)..get_key(9),
11427 12 : lsn_range: Lsn(0x20)..Lsn(0x21),
11428 12 : is_delta: false,
11429 12 : },
11430 12 : PersistentLayerKey {
11431 12 : key_range: get_key(8)..get_key(10),
11432 12 : lsn_range: Lsn(0x48)..Lsn(0x50),
11433 12 : is_delta: true,
11434 12 : },
11435 12 : ],
11436 12 : );
11437 12 :
11438 12 : // Do a partial compaction on key range 9..10
11439 12 : tline
11440 12 : .compact_with_gc(
11441 12 : &cancel,
11442 12 : CompactOptions {
11443 12 : flags: EnumSet::new(),
11444 12 : compact_key_range: Some((get_key(9)..get_key(10)).into()),
11445 12 : ..Default::default()
11446 12 : },
11447 12 : &ctx,
11448 12 : )
11449 12 : .await
11450 12 : .unwrap();
11451 12 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11452 12 : check_layer_map_key_eq(
11453 12 : all_layers,
11454 12 : vec![
11455 12 : PersistentLayerKey {
11456 12 : key_range: get_key(0)..get_key(2),
11457 12 : lsn_range: Lsn(0x20)..Lsn(0x21),
11458 12 : is_delta: false,
11459 12 : },
11460 12 : PersistentLayerKey {
11461 12 : key_range: get_key(0)..get_key(10),
11462 12 : lsn_range: Lsn(0x10)..Lsn(0x11),
11463 12 : is_delta: false,
11464 12 : },
11465 12 : PersistentLayerKey {
11466 12 : key_range: get_key(2)..get_key(4),
11467 12 : lsn_range: Lsn(0x20)..Lsn(0x21),
11468 12 : is_delta: false,
11469 12 : },
11470 12 : PersistentLayerKey {
11471 12 : key_range: get_key(2)..get_key(4),
11472 12 : lsn_range: Lsn(0x20)..Lsn(0x48),
11473 12 : is_delta: true,
11474 12 : },
11475 12 : PersistentLayerKey {
11476 12 : key_range: get_key(4)..get_key(9),
11477 12 : lsn_range: Lsn(0x20)..Lsn(0x21),
11478 12 : is_delta: false,
11479 12 : },
11480 12 : // image layer generated for the compaction range
11481 12 : PersistentLayerKey {
11482 12 : key_range: get_key(9)..get_key(10),
11483 12 : lsn_range: Lsn(0x20)..Lsn(0x21),
11484 12 : is_delta: false,
11485 12 : },
11486 12 : PersistentLayerKey {
11487 12 : key_range: get_key(8)..get_key(10),
11488 12 : lsn_range: Lsn(0x48)..Lsn(0x50),
11489 12 : is_delta: true,
11490 12 : },
11491 12 : ],
11492 12 : );
11493 12 :
11494 12 : // Do a partial compaction on key range 0..10, all image layers below LSN 20 can be replaced with new ones.
11495 12 : tline
11496 12 : .compact_with_gc(
11497 12 : &cancel,
11498 12 : CompactOptions {
11499 12 : flags: EnumSet::new(),
11500 12 : compact_key_range: Some((get_key(0)..get_key(10)).into()),
11501 12 : ..Default::default()
11502 12 : },
11503 12 : &ctx,
11504 12 : )
11505 12 : .await
11506 12 : .unwrap();
11507 12 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11508 12 : check_layer_map_key_eq(
11509 12 : all_layers,
11510 12 : vec![
11511 12 : // aha, we removed all unnecessary image/delta layers and got a very clean layer map!
11512 12 : PersistentLayerKey {
11513 12 : key_range: get_key(0)..get_key(10),
11514 12 : lsn_range: Lsn(0x20)..Lsn(0x21),
11515 12 : is_delta: false,
11516 12 : },
11517 12 : PersistentLayerKey {
11518 12 : key_range: get_key(2)..get_key(4),
11519 12 : lsn_range: Lsn(0x20)..Lsn(0x48),
11520 12 : is_delta: true,
11521 12 : },
11522 12 : PersistentLayerKey {
11523 12 : key_range: get_key(8)..get_key(10),
11524 12 : lsn_range: Lsn(0x48)..Lsn(0x50),
11525 12 : is_delta: true,
11526 12 : },
11527 12 : ],
11528 12 : );
11529 12 : Ok(())
11530 12 : }
11531 :
11532 : #[cfg(feature = "testing")]
11533 : #[tokio::test]
11534 12 : async fn test_timeline_offload_retain_lsn() -> anyhow::Result<()> {
11535 12 : let harness = TenantHarness::create("test_timeline_offload_retain_lsn")
11536 12 : .await
11537 12 : .unwrap();
11538 12 : let (tenant, ctx) = harness.load().await;
11539 12 : let tline_parent = tenant
11540 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
11541 12 : .await
11542 12 : .unwrap();
11543 12 : let tline_child = tenant
11544 12 : .branch_timeline_test(&tline_parent, NEW_TIMELINE_ID, Some(Lsn(0x20)), &ctx)
11545 12 : .await
11546 12 : .unwrap();
11547 12 : {
11548 12 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
11549 12 : assert_eq!(
11550 12 : gc_info_parent.retain_lsns,
11551 12 : vec![(Lsn(0x20), tline_child.timeline_id, MaybeOffloaded::No)]
11552 12 : );
11553 12 : }
11554 12 : // We have to directly call the remote_client instead of using the archive function to avoid constructing broker client...
11555 12 : tline_child
11556 12 : .remote_client
11557 12 : .schedule_index_upload_for_timeline_archival_state(TimelineArchivalState::Archived)
11558 12 : .unwrap();
11559 12 : tline_child.remote_client.wait_completion().await.unwrap();
11560 12 : offload_timeline(&tenant, &tline_child)
11561 12 : .instrument(tracing::info_span!(parent: None, "offload_test", tenant_id=%"test", shard_id=%"test", timeline_id=%"test"))
11562 12 : .await.unwrap();
11563 12 : let child_timeline_id = tline_child.timeline_id;
11564 12 : Arc::try_unwrap(tline_child).unwrap();
11565 12 :
11566 12 : {
11567 12 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
11568 12 : assert_eq!(
11569 12 : gc_info_parent.retain_lsns,
11570 12 : vec![(Lsn(0x20), child_timeline_id, MaybeOffloaded::Yes)]
11571 12 : );
11572 12 : }
11573 12 :
11574 12 : tenant
11575 12 : .get_offloaded_timeline(child_timeline_id)
11576 12 : .unwrap()
11577 12 : .defuse_for_tenant_drop();
11578 12 :
11579 12 : Ok(())
11580 12 : }
11581 :
11582 : #[cfg(feature = "testing")]
11583 : #[tokio::test]
11584 12 : async fn test_simple_bottom_most_compaction_above_lsn() -> anyhow::Result<()> {
11585 12 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_above_lsn").await?;
11586 12 : let (tenant, ctx) = harness.load().await;
11587 12 :
11588 1776 : fn get_key(id: u32) -> Key {
11589 1776 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
11590 1776 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
11591 1776 : key.field6 = id;
11592 1776 : key
11593 1776 : }
11594 12 :
11595 12 : let img_layer = (0..10)
11596 120 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
11597 12 : .collect_vec();
11598 12 :
11599 12 : let delta1 = vec![(
11600 12 : get_key(1),
11601 12 : Lsn(0x20),
11602 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
11603 12 : )];
11604 12 : let delta4 = vec![(
11605 12 : get_key(1),
11606 12 : Lsn(0x28),
11607 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
11608 12 : )];
11609 12 : let delta2 = vec![
11610 12 : (
11611 12 : get_key(1),
11612 12 : Lsn(0x30),
11613 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
11614 12 : ),
11615 12 : (
11616 12 : get_key(1),
11617 12 : Lsn(0x38),
11618 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
11619 12 : ),
11620 12 : ];
11621 12 : let delta3 = vec![
11622 12 : (
11623 12 : get_key(8),
11624 12 : Lsn(0x48),
11625 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
11626 12 : ),
11627 12 : (
11628 12 : get_key(9),
11629 12 : Lsn(0x48),
11630 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
11631 12 : ),
11632 12 : ];
11633 12 :
11634 12 : let tline = tenant
11635 12 : .create_test_timeline_with_layers(
11636 12 : TIMELINE_ID,
11637 12 : Lsn(0x10),
11638 12 : DEFAULT_PG_VERSION,
11639 12 : &ctx,
11640 12 : vec![], // in-memory layers
11641 12 : vec![
11642 12 : // delta1/2/4 only contain a single key but multiple updates
11643 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x28), delta1),
11644 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
11645 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x28)..Lsn(0x30), delta4),
11646 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta3),
11647 12 : ], // delta layers
11648 12 : vec![(Lsn(0x10), img_layer)], // image layers
11649 12 : Lsn(0x50),
11650 12 : )
11651 12 : .await?;
11652 12 : {
11653 12 : tline
11654 12 : .applied_gc_cutoff_lsn
11655 12 : .lock_for_write()
11656 12 : .store_and_unlock(Lsn(0x30))
11657 12 : .wait()
11658 12 : .await;
11659 12 : // Update GC info
11660 12 : let mut guard = tline.gc_info.write().unwrap();
11661 12 : *guard = GcInfo {
11662 12 : retain_lsns: vec![
11663 12 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
11664 12 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
11665 12 : ],
11666 12 : cutoffs: GcCutoffs {
11667 12 : time: Lsn(0x30),
11668 12 : space: Lsn(0x30),
11669 12 : },
11670 12 : leases: Default::default(),
11671 12 : within_ancestor_pitr: false,
11672 12 : };
11673 12 : }
11674 12 :
11675 12 : let expected_result = [
11676 12 : Bytes::from_static(b"value 0@0x10"),
11677 12 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
11678 12 : Bytes::from_static(b"value 2@0x10"),
11679 12 : Bytes::from_static(b"value 3@0x10"),
11680 12 : Bytes::from_static(b"value 4@0x10"),
11681 12 : Bytes::from_static(b"value 5@0x10"),
11682 12 : Bytes::from_static(b"value 6@0x10"),
11683 12 : Bytes::from_static(b"value 7@0x10"),
11684 12 : Bytes::from_static(b"value 8@0x10@0x48"),
11685 12 : Bytes::from_static(b"value 9@0x10@0x48"),
11686 12 : ];
11687 12 :
11688 12 : let expected_result_at_gc_horizon = [
11689 12 : Bytes::from_static(b"value 0@0x10"),
11690 12 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
11691 12 : Bytes::from_static(b"value 2@0x10"),
11692 12 : Bytes::from_static(b"value 3@0x10"),
11693 12 : Bytes::from_static(b"value 4@0x10"),
11694 12 : Bytes::from_static(b"value 5@0x10"),
11695 12 : Bytes::from_static(b"value 6@0x10"),
11696 12 : Bytes::from_static(b"value 7@0x10"),
11697 12 : Bytes::from_static(b"value 8@0x10"),
11698 12 : Bytes::from_static(b"value 9@0x10"),
11699 12 : ];
11700 12 :
11701 12 : let expected_result_at_lsn_20 = [
11702 12 : Bytes::from_static(b"value 0@0x10"),
11703 12 : Bytes::from_static(b"value 1@0x10@0x20"),
11704 12 : Bytes::from_static(b"value 2@0x10"),
11705 12 : Bytes::from_static(b"value 3@0x10"),
11706 12 : Bytes::from_static(b"value 4@0x10"),
11707 12 : Bytes::from_static(b"value 5@0x10"),
11708 12 : Bytes::from_static(b"value 6@0x10"),
11709 12 : Bytes::from_static(b"value 7@0x10"),
11710 12 : Bytes::from_static(b"value 8@0x10"),
11711 12 : Bytes::from_static(b"value 9@0x10"),
11712 12 : ];
11713 12 :
11714 12 : let expected_result_at_lsn_10 = [
11715 12 : Bytes::from_static(b"value 0@0x10"),
11716 12 : Bytes::from_static(b"value 1@0x10"),
11717 12 : Bytes::from_static(b"value 2@0x10"),
11718 12 : Bytes::from_static(b"value 3@0x10"),
11719 12 : Bytes::from_static(b"value 4@0x10"),
11720 12 : Bytes::from_static(b"value 5@0x10"),
11721 12 : Bytes::from_static(b"value 6@0x10"),
11722 12 : Bytes::from_static(b"value 7@0x10"),
11723 12 : Bytes::from_static(b"value 8@0x10"),
11724 12 : Bytes::from_static(b"value 9@0x10"),
11725 12 : ];
11726 12 :
11727 36 : let verify_result = || async {
11728 36 : let gc_horizon = {
11729 36 : let gc_info = tline.gc_info.read().unwrap();
11730 36 : gc_info.cutoffs.time
11731 12 : };
11732 396 : for idx in 0..10 {
11733 360 : assert_eq!(
11734 360 : tline
11735 360 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
11736 360 : .await
11737 360 : .unwrap(),
11738 360 : &expected_result[idx]
11739 12 : );
11740 360 : assert_eq!(
11741 360 : tline
11742 360 : .get(get_key(idx as u32), gc_horizon, &ctx)
11743 360 : .await
11744 360 : .unwrap(),
11745 360 : &expected_result_at_gc_horizon[idx]
11746 12 : );
11747 360 : assert_eq!(
11748 360 : tline
11749 360 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
11750 360 : .await
11751 360 : .unwrap(),
11752 360 : &expected_result_at_lsn_20[idx]
11753 12 : );
11754 360 : assert_eq!(
11755 360 : tline
11756 360 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
11757 360 : .await
11758 360 : .unwrap(),
11759 360 : &expected_result_at_lsn_10[idx]
11760 12 : );
11761 12 : }
11762 72 : };
11763 12 :
11764 12 : verify_result().await;
11765 12 :
11766 12 : let cancel = CancellationToken::new();
11767 12 : tline
11768 12 : .compact_with_gc(
11769 12 : &cancel,
11770 12 : CompactOptions {
11771 12 : compact_lsn_range: Some(CompactLsnRange::above(Lsn(0x28))),
11772 12 : ..Default::default()
11773 12 : },
11774 12 : &ctx,
11775 12 : )
11776 12 : .await
11777 12 : .unwrap();
11778 12 : verify_result().await;
11779 12 :
11780 12 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11781 12 : check_layer_map_key_eq(
11782 12 : all_layers,
11783 12 : vec![
11784 12 : // The original image layer, not compacted
11785 12 : PersistentLayerKey {
11786 12 : key_range: get_key(0)..get_key(10),
11787 12 : lsn_range: Lsn(0x10)..Lsn(0x11),
11788 12 : is_delta: false,
11789 12 : },
11790 12 : // Delta layer below the specified above_lsn not compacted
11791 12 : PersistentLayerKey {
11792 12 : key_range: get_key(1)..get_key(2),
11793 12 : lsn_range: Lsn(0x20)..Lsn(0x28),
11794 12 : is_delta: true,
11795 12 : },
11796 12 : // Delta layer compacted above the LSN
11797 12 : PersistentLayerKey {
11798 12 : key_range: get_key(1)..get_key(10),
11799 12 : lsn_range: Lsn(0x28)..Lsn(0x50),
11800 12 : is_delta: true,
11801 12 : },
11802 12 : ],
11803 12 : );
11804 12 :
11805 12 : // compact again
11806 12 : tline
11807 12 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
11808 12 : .await
11809 12 : .unwrap();
11810 12 : verify_result().await;
11811 12 :
11812 12 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11813 12 : check_layer_map_key_eq(
11814 12 : all_layers,
11815 12 : vec![
11816 12 : // The compacted image layer (full key range)
11817 12 : PersistentLayerKey {
11818 12 : key_range: Key::MIN..Key::MAX,
11819 12 : lsn_range: Lsn(0x10)..Lsn(0x11),
11820 12 : is_delta: false,
11821 12 : },
11822 12 : // All other data in the delta layer
11823 12 : PersistentLayerKey {
11824 12 : key_range: get_key(1)..get_key(10),
11825 12 : lsn_range: Lsn(0x10)..Lsn(0x50),
11826 12 : is_delta: true,
11827 12 : },
11828 12 : ],
11829 12 : );
11830 12 :
11831 12 : Ok(())
11832 12 : }
11833 :
11834 : #[cfg(feature = "testing")]
11835 : #[tokio::test]
11836 12 : async fn test_simple_bottom_most_compaction_rectangle() -> anyhow::Result<()> {
11837 12 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_rectangle").await?;
11838 12 : let (tenant, ctx) = harness.load().await;
11839 12 :
11840 3048 : fn get_key(id: u32) -> Key {
11841 3048 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
11842 3048 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
11843 3048 : key.field6 = id;
11844 3048 : key
11845 3048 : }
11846 12 :
11847 12 : let img_layer = (0..10)
11848 120 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
11849 12 : .collect_vec();
11850 12 :
11851 12 : let delta1 = vec![(
11852 12 : get_key(1),
11853 12 : Lsn(0x20),
11854 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
11855 12 : )];
11856 12 : let delta4 = vec![(
11857 12 : get_key(1),
11858 12 : Lsn(0x28),
11859 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
11860 12 : )];
11861 12 : let delta2 = vec![
11862 12 : (
11863 12 : get_key(1),
11864 12 : Lsn(0x30),
11865 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
11866 12 : ),
11867 12 : (
11868 12 : get_key(1),
11869 12 : Lsn(0x38),
11870 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
11871 12 : ),
11872 12 : ];
11873 12 : let delta3 = vec![
11874 12 : (
11875 12 : get_key(8),
11876 12 : Lsn(0x48),
11877 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
11878 12 : ),
11879 12 : (
11880 12 : get_key(9),
11881 12 : Lsn(0x48),
11882 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
11883 12 : ),
11884 12 : ];
11885 12 :
11886 12 : let tline = tenant
11887 12 : .create_test_timeline_with_layers(
11888 12 : TIMELINE_ID,
11889 12 : Lsn(0x10),
11890 12 : DEFAULT_PG_VERSION,
11891 12 : &ctx,
11892 12 : vec![], // in-memory layers
11893 12 : vec![
11894 12 : // delta1/2/4 only contain a single key but multiple updates
11895 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x28), delta1),
11896 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
11897 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x28)..Lsn(0x30), delta4),
11898 12 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta3),
11899 12 : ], // delta layers
11900 12 : vec![(Lsn(0x10), img_layer)], // image layers
11901 12 : Lsn(0x50),
11902 12 : )
11903 12 : .await?;
11904 12 : {
11905 12 : tline
11906 12 : .applied_gc_cutoff_lsn
11907 12 : .lock_for_write()
11908 12 : .store_and_unlock(Lsn(0x30))
11909 12 : .wait()
11910 12 : .await;
11911 12 : // Update GC info
11912 12 : let mut guard = tline.gc_info.write().unwrap();
11913 12 : *guard = GcInfo {
11914 12 : retain_lsns: vec![
11915 12 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
11916 12 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
11917 12 : ],
11918 12 : cutoffs: GcCutoffs {
11919 12 : time: Lsn(0x30),
11920 12 : space: Lsn(0x30),
11921 12 : },
11922 12 : leases: Default::default(),
11923 12 : within_ancestor_pitr: false,
11924 12 : };
11925 12 : }
11926 12 :
11927 12 : let expected_result = [
11928 12 : Bytes::from_static(b"value 0@0x10"),
11929 12 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
11930 12 : Bytes::from_static(b"value 2@0x10"),
11931 12 : Bytes::from_static(b"value 3@0x10"),
11932 12 : Bytes::from_static(b"value 4@0x10"),
11933 12 : Bytes::from_static(b"value 5@0x10"),
11934 12 : Bytes::from_static(b"value 6@0x10"),
11935 12 : Bytes::from_static(b"value 7@0x10"),
11936 12 : Bytes::from_static(b"value 8@0x10@0x48"),
11937 12 : Bytes::from_static(b"value 9@0x10@0x48"),
11938 12 : ];
11939 12 :
11940 12 : let expected_result_at_gc_horizon = [
11941 12 : Bytes::from_static(b"value 0@0x10"),
11942 12 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
11943 12 : Bytes::from_static(b"value 2@0x10"),
11944 12 : Bytes::from_static(b"value 3@0x10"),
11945 12 : Bytes::from_static(b"value 4@0x10"),
11946 12 : Bytes::from_static(b"value 5@0x10"),
11947 12 : Bytes::from_static(b"value 6@0x10"),
11948 12 : Bytes::from_static(b"value 7@0x10"),
11949 12 : Bytes::from_static(b"value 8@0x10"),
11950 12 : Bytes::from_static(b"value 9@0x10"),
11951 12 : ];
11952 12 :
11953 12 : let expected_result_at_lsn_20 = [
11954 12 : Bytes::from_static(b"value 0@0x10"),
11955 12 : Bytes::from_static(b"value 1@0x10@0x20"),
11956 12 : Bytes::from_static(b"value 2@0x10"),
11957 12 : Bytes::from_static(b"value 3@0x10"),
11958 12 : Bytes::from_static(b"value 4@0x10"),
11959 12 : Bytes::from_static(b"value 5@0x10"),
11960 12 : Bytes::from_static(b"value 6@0x10"),
11961 12 : Bytes::from_static(b"value 7@0x10"),
11962 12 : Bytes::from_static(b"value 8@0x10"),
11963 12 : Bytes::from_static(b"value 9@0x10"),
11964 12 : ];
11965 12 :
11966 12 : let expected_result_at_lsn_10 = [
11967 12 : Bytes::from_static(b"value 0@0x10"),
11968 12 : Bytes::from_static(b"value 1@0x10"),
11969 12 : Bytes::from_static(b"value 2@0x10"),
11970 12 : Bytes::from_static(b"value 3@0x10"),
11971 12 : Bytes::from_static(b"value 4@0x10"),
11972 12 : Bytes::from_static(b"value 5@0x10"),
11973 12 : Bytes::from_static(b"value 6@0x10"),
11974 12 : Bytes::from_static(b"value 7@0x10"),
11975 12 : Bytes::from_static(b"value 8@0x10"),
11976 12 : Bytes::from_static(b"value 9@0x10"),
11977 12 : ];
11978 12 :
11979 60 : let verify_result = || async {
11980 60 : let gc_horizon = {
11981 60 : let gc_info = tline.gc_info.read().unwrap();
11982 60 : gc_info.cutoffs.time
11983 12 : };
11984 660 : for idx in 0..10 {
11985 600 : assert_eq!(
11986 600 : tline
11987 600 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
11988 600 : .await
11989 600 : .unwrap(),
11990 600 : &expected_result[idx]
11991 12 : );
11992 600 : assert_eq!(
11993 600 : tline
11994 600 : .get(get_key(idx as u32), gc_horizon, &ctx)
11995 600 : .await
11996 600 : .unwrap(),
11997 600 : &expected_result_at_gc_horizon[idx]
11998 12 : );
11999 600 : assert_eq!(
12000 600 : tline
12001 600 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
12002 600 : .await
12003 600 : .unwrap(),
12004 600 : &expected_result_at_lsn_20[idx]
12005 12 : );
12006 600 : assert_eq!(
12007 600 : tline
12008 600 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
12009 600 : .await
12010 600 : .unwrap(),
12011 600 : &expected_result_at_lsn_10[idx]
12012 12 : );
12013 12 : }
12014 120 : };
12015 12 :
12016 12 : verify_result().await;
12017 12 :
12018 12 : let cancel = CancellationToken::new();
12019 12 :
12020 12 : tline
12021 12 : .compact_with_gc(
12022 12 : &cancel,
12023 12 : CompactOptions {
12024 12 : compact_key_range: Some((get_key(0)..get_key(2)).into()),
12025 12 : compact_lsn_range: Some((Lsn(0x20)..Lsn(0x28)).into()),
12026 12 : ..Default::default()
12027 12 : },
12028 12 : &ctx,
12029 12 : )
12030 12 : .await
12031 12 : .unwrap();
12032 12 : verify_result().await;
12033 12 :
12034 12 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12035 12 : check_layer_map_key_eq(
12036 12 : all_layers,
12037 12 : vec![
12038 12 : // The original image layer, not compacted
12039 12 : PersistentLayerKey {
12040 12 : key_range: get_key(0)..get_key(10),
12041 12 : lsn_range: Lsn(0x10)..Lsn(0x11),
12042 12 : is_delta: false,
12043 12 : },
12044 12 : // According the selection logic, we select all layers with start key <= 0x28, so we would merge the layer 0x20-0x28 and
12045 12 : // the layer 0x28-0x30 into one.
12046 12 : PersistentLayerKey {
12047 12 : key_range: get_key(1)..get_key(2),
12048 12 : lsn_range: Lsn(0x20)..Lsn(0x30),
12049 12 : is_delta: true,
12050 12 : },
12051 12 : // Above the upper bound and untouched
12052 12 : PersistentLayerKey {
12053 12 : key_range: get_key(1)..get_key(2),
12054 12 : lsn_range: Lsn(0x30)..Lsn(0x50),
12055 12 : is_delta: true,
12056 12 : },
12057 12 : // This layer is untouched
12058 12 : PersistentLayerKey {
12059 12 : key_range: get_key(8)..get_key(10),
12060 12 : lsn_range: Lsn(0x30)..Lsn(0x50),
12061 12 : is_delta: true,
12062 12 : },
12063 12 : ],
12064 12 : );
12065 12 :
12066 12 : tline
12067 12 : .compact_with_gc(
12068 12 : &cancel,
12069 12 : CompactOptions {
12070 12 : compact_key_range: Some((get_key(3)..get_key(8)).into()),
12071 12 : compact_lsn_range: Some((Lsn(0x28)..Lsn(0x40)).into()),
12072 12 : ..Default::default()
12073 12 : },
12074 12 : &ctx,
12075 12 : )
12076 12 : .await
12077 12 : .unwrap();
12078 12 : verify_result().await;
12079 12 :
12080 12 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12081 12 : check_layer_map_key_eq(
12082 12 : all_layers,
12083 12 : vec![
12084 12 : // The original image layer, not compacted
12085 12 : PersistentLayerKey {
12086 12 : key_range: get_key(0)..get_key(10),
12087 12 : lsn_range: Lsn(0x10)..Lsn(0x11),
12088 12 : is_delta: false,
12089 12 : },
12090 12 : // Not in the compaction key range, uncompacted
12091 12 : PersistentLayerKey {
12092 12 : key_range: get_key(1)..get_key(2),
12093 12 : lsn_range: Lsn(0x20)..Lsn(0x30),
12094 12 : is_delta: true,
12095 12 : },
12096 12 : // Not in the compaction key range, uncompacted but need rewrite because the delta layer overlaps with the range
12097 12 : PersistentLayerKey {
12098 12 : key_range: get_key(1)..get_key(2),
12099 12 : lsn_range: Lsn(0x30)..Lsn(0x50),
12100 12 : is_delta: true,
12101 12 : },
12102 12 : // Note that when we specify the LSN upper bound to be 0x40, the compaction algorithm will not try to cut the layer
12103 12 : // horizontally in half. Instead, it will include all LSNs that overlap with 0x40. So the real max_lsn of the compaction
12104 12 : // becomes 0x50.
12105 12 : PersistentLayerKey {
12106 12 : key_range: get_key(8)..get_key(10),
12107 12 : lsn_range: Lsn(0x30)..Lsn(0x50),
12108 12 : is_delta: true,
12109 12 : },
12110 12 : ],
12111 12 : );
12112 12 :
12113 12 : // compact again
12114 12 : tline
12115 12 : .compact_with_gc(
12116 12 : &cancel,
12117 12 : CompactOptions {
12118 12 : compact_key_range: Some((get_key(0)..get_key(5)).into()),
12119 12 : compact_lsn_range: Some((Lsn(0x20)..Lsn(0x50)).into()),
12120 12 : ..Default::default()
12121 12 : },
12122 12 : &ctx,
12123 12 : )
12124 12 : .await
12125 12 : .unwrap();
12126 12 : verify_result().await;
12127 12 :
12128 12 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12129 12 : check_layer_map_key_eq(
12130 12 : all_layers,
12131 12 : vec![
12132 12 : // The original image layer, not compacted
12133 12 : PersistentLayerKey {
12134 12 : key_range: get_key(0)..get_key(10),
12135 12 : lsn_range: Lsn(0x10)..Lsn(0x11),
12136 12 : is_delta: false,
12137 12 : },
12138 12 : // The range gets compacted
12139 12 : PersistentLayerKey {
12140 12 : key_range: get_key(1)..get_key(2),
12141 12 : lsn_range: Lsn(0x20)..Lsn(0x50),
12142 12 : is_delta: true,
12143 12 : },
12144 12 : // Not touched during this iteration of compaction
12145 12 : PersistentLayerKey {
12146 12 : key_range: get_key(8)..get_key(10),
12147 12 : lsn_range: Lsn(0x30)..Lsn(0x50),
12148 12 : is_delta: true,
12149 12 : },
12150 12 : ],
12151 12 : );
12152 12 :
12153 12 : // final full compaction
12154 12 : tline
12155 12 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
12156 12 : .await
12157 12 : .unwrap();
12158 12 : verify_result().await;
12159 12 :
12160 12 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12161 12 : check_layer_map_key_eq(
12162 12 : all_layers,
12163 12 : vec![
12164 12 : // The compacted image layer (full key range)
12165 12 : PersistentLayerKey {
12166 12 : key_range: Key::MIN..Key::MAX,
12167 12 : lsn_range: Lsn(0x10)..Lsn(0x11),
12168 12 : is_delta: false,
12169 12 : },
12170 12 : // All other data in the delta layer
12171 12 : PersistentLayerKey {
12172 12 : key_range: get_key(1)..get_key(10),
12173 12 : lsn_range: Lsn(0x10)..Lsn(0x50),
12174 12 : is_delta: true,
12175 12 : },
12176 12 : ],
12177 12 : );
12178 12 :
12179 12 : Ok(())
12180 12 : }
12181 :
12182 : #[cfg(feature = "testing")]
12183 : #[tokio::test]
12184 12 : async fn test_bottom_most_compation_redo_failure() -> anyhow::Result<()> {
12185 12 : let harness = TenantHarness::create("test_bottom_most_compation_redo_failure").await?;
12186 12 : let (tenant, ctx) = harness.load().await;
12187 12 :
12188 156 : fn get_key(id: u32) -> Key {
12189 156 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
12190 156 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
12191 156 : key.field6 = id;
12192 156 : key
12193 156 : }
12194 12 :
12195 12 : let img_layer = (0..10)
12196 120 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
12197 12 : .collect_vec();
12198 12 :
12199 12 : let delta1 = vec![
12200 12 : (
12201 12 : get_key(1),
12202 12 : Lsn(0x20),
12203 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
12204 12 : ),
12205 12 : (
12206 12 : get_key(1),
12207 12 : Lsn(0x24),
12208 12 : Value::WalRecord(NeonWalRecord::wal_append("@0x24")),
12209 12 : ),
12210 12 : (
12211 12 : get_key(1),
12212 12 : Lsn(0x28),
12213 12 : // This record will fail to redo
12214 12 : Value::WalRecord(NeonWalRecord::wal_append_conditional("@0x28", "???")),
12215 12 : ),
12216 12 : ];
12217 12 :
12218 12 : let tline = tenant
12219 12 : .create_test_timeline_with_layers(
12220 12 : TIMELINE_ID,
12221 12 : Lsn(0x10),
12222 12 : DEFAULT_PG_VERSION,
12223 12 : &ctx,
12224 12 : vec![], // in-memory layers
12225 12 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
12226 12 : Lsn(0x20)..Lsn(0x30),
12227 12 : delta1,
12228 12 : )], // delta layers
12229 12 : vec![(Lsn(0x10), img_layer)], // image layers
12230 12 : Lsn(0x50),
12231 12 : )
12232 12 : .await?;
12233 12 : {
12234 12 : tline
12235 12 : .applied_gc_cutoff_lsn
12236 12 : .lock_for_write()
12237 12 : .store_and_unlock(Lsn(0x30))
12238 12 : .wait()
12239 12 : .await;
12240 12 : // Update GC info
12241 12 : let mut guard = tline.gc_info.write().unwrap();
12242 12 : *guard = GcInfo {
12243 12 : retain_lsns: vec![],
12244 12 : cutoffs: GcCutoffs {
12245 12 : time: Lsn(0x30),
12246 12 : space: Lsn(0x30),
12247 12 : },
12248 12 : leases: Default::default(),
12249 12 : within_ancestor_pitr: false,
12250 12 : };
12251 12 : }
12252 12 :
12253 12 : let cancel = CancellationToken::new();
12254 12 :
12255 12 : // Compaction will fail, but should not fire any critical error.
12256 12 : // Gc-compaction currently cannot figure out what keys are not in the keyspace during the compaction
12257 12 : // process. It will always try to redo the logs it reads and if it doesn't work, fail the entire
12258 12 : // compaction job. Tracked in <https://github.com/neondatabase/neon/issues/10395>.
12259 12 : let res = tline
12260 12 : .compact_with_gc(
12261 12 : &cancel,
12262 12 : CompactOptions {
12263 12 : compact_key_range: None,
12264 12 : compact_lsn_range: None,
12265 12 : ..Default::default()
12266 12 : },
12267 12 : &ctx,
12268 12 : )
12269 12 : .await;
12270 12 : assert!(res.is_err());
12271 12 :
12272 12 : Ok(())
12273 12 : }
12274 :
12275 : #[cfg(feature = "testing")]
12276 : #[tokio::test]
12277 12 : async fn test_synthetic_size_calculation_with_invisible_branches() -> anyhow::Result<()> {
12278 12 : use pageserver_api::models::TimelineVisibilityState;
12279 12 :
12280 12 : use crate::tenant::size::gather_inputs;
12281 12 :
12282 12 : let tenant_conf = pageserver_api::models::TenantConfig {
12283 12 : // Ensure that we don't compute gc_cutoffs (which needs reading the layer files)
12284 12 : pitr_interval: Some(Duration::ZERO),
12285 12 : ..Default::default()
12286 12 : };
12287 12 : let harness = TenantHarness::create_custom(
12288 12 : "test_synthetic_size_calculation_with_invisible_branches",
12289 12 : tenant_conf,
12290 12 : TenantId::generate(),
12291 12 : ShardIdentity::unsharded(),
12292 12 : Generation::new(0xdeadbeef),
12293 12 : )
12294 12 : .await?;
12295 12 : let (tenant, ctx) = harness.load().await;
12296 12 : let main_tline = tenant
12297 12 : .create_test_timeline_with_layers(
12298 12 : TIMELINE_ID,
12299 12 : Lsn(0x10),
12300 12 : DEFAULT_PG_VERSION,
12301 12 : &ctx,
12302 12 : vec![],
12303 12 : vec![],
12304 12 : vec![],
12305 12 : Lsn(0x100),
12306 12 : )
12307 12 : .await?;
12308 12 :
12309 12 : let snapshot1 = TimelineId::from_array(hex!("11223344556677881122334455667790"));
12310 12 : tenant
12311 12 : .branch_timeline_test_with_layers(
12312 12 : &main_tline,
12313 12 : snapshot1,
12314 12 : Some(Lsn(0x20)),
12315 12 : &ctx,
12316 12 : vec![],
12317 12 : vec![],
12318 12 : Lsn(0x50),
12319 12 : )
12320 12 : .await?;
12321 12 : let snapshot2 = TimelineId::from_array(hex!("11223344556677881122334455667791"));
12322 12 : tenant
12323 12 : .branch_timeline_test_with_layers(
12324 12 : &main_tline,
12325 12 : snapshot2,
12326 12 : Some(Lsn(0x30)),
12327 12 : &ctx,
12328 12 : vec![],
12329 12 : vec![],
12330 12 : Lsn(0x50),
12331 12 : )
12332 12 : .await?;
12333 12 : let snapshot3 = TimelineId::from_array(hex!("11223344556677881122334455667792"));
12334 12 : tenant
12335 12 : .branch_timeline_test_with_layers(
12336 12 : &main_tline,
12337 12 : snapshot3,
12338 12 : Some(Lsn(0x40)),
12339 12 : &ctx,
12340 12 : vec![],
12341 12 : vec![],
12342 12 : Lsn(0x50),
12343 12 : )
12344 12 : .await?;
12345 12 : let limit = Arc::new(Semaphore::new(1));
12346 12 : let max_retention_period = None;
12347 12 : let mut logical_size_cache = HashMap::new();
12348 12 : let cause = LogicalSizeCalculationCause::EvictionTaskImitation;
12349 12 : let cancel = CancellationToken::new();
12350 12 :
12351 12 : let inputs = gather_inputs(
12352 12 : &tenant,
12353 12 : &limit,
12354 12 : max_retention_period,
12355 12 : &mut logical_size_cache,
12356 12 : cause,
12357 12 : &cancel,
12358 12 : &ctx,
12359 12 : )
12360 12 : .instrument(info_span!(
12361 12 : "gather_inputs",
12362 12 : tenant_id = "unknown",
12363 12 : shard_id = "unknown",
12364 12 : ))
12365 12 : .await?;
12366 12 : use crate::tenant::size::{LsnKind, ModelInputs, SegmentMeta};
12367 12 : use LsnKind::*;
12368 12 : use tenant_size_model::Segment;
12369 12 : let ModelInputs { mut segments, .. } = inputs;
12370 180 : segments.retain(|s| s.timeline_id == TIMELINE_ID);
12371 72 : for segment in segments.iter_mut() {
12372 72 : segment.segment.parent = None; // We don't care about the parent for the test
12373 72 : segment.segment.size = None; // We don't care about the size for the test
12374 72 : }
12375 12 : assert_eq!(
12376 12 : segments,
12377 12 : [
12378 12 : SegmentMeta {
12379 12 : segment: Segment {
12380 12 : parent: None,
12381 12 : lsn: 0x10,
12382 12 : size: None,
12383 12 : needed: false,
12384 12 : },
12385 12 : timeline_id: TIMELINE_ID,
12386 12 : kind: BranchStart,
12387 12 : },
12388 12 : SegmentMeta {
12389 12 : segment: Segment {
12390 12 : parent: None,
12391 12 : lsn: 0x20,
12392 12 : size: None,
12393 12 : needed: false,
12394 12 : },
12395 12 : timeline_id: TIMELINE_ID,
12396 12 : kind: BranchPoint,
12397 12 : },
12398 12 : SegmentMeta {
12399 12 : segment: Segment {
12400 12 : parent: None,
12401 12 : lsn: 0x30,
12402 12 : size: None,
12403 12 : needed: false,
12404 12 : },
12405 12 : timeline_id: TIMELINE_ID,
12406 12 : kind: BranchPoint,
12407 12 : },
12408 12 : SegmentMeta {
12409 12 : segment: Segment {
12410 12 : parent: None,
12411 12 : lsn: 0x40,
12412 12 : size: None,
12413 12 : needed: false,
12414 12 : },
12415 12 : timeline_id: TIMELINE_ID,
12416 12 : kind: BranchPoint,
12417 12 : },
12418 12 : SegmentMeta {
12419 12 : segment: Segment {
12420 12 : parent: None,
12421 12 : lsn: 0x100,
12422 12 : size: None,
12423 12 : needed: false,
12424 12 : },
12425 12 : timeline_id: TIMELINE_ID,
12426 12 : kind: GcCutOff,
12427 12 : }, // we need to retain everything above the last branch point
12428 12 : SegmentMeta {
12429 12 : segment: Segment {
12430 12 : parent: None,
12431 12 : lsn: 0x100,
12432 12 : size: None,
12433 12 : needed: true,
12434 12 : },
12435 12 : timeline_id: TIMELINE_ID,
12436 12 : kind: BranchEnd,
12437 12 : },
12438 12 : ]
12439 12 : );
12440 12 :
12441 12 : main_tline
12442 12 : .remote_client
12443 12 : .schedule_index_upload_for_timeline_invisible_state(
12444 12 : TimelineVisibilityState::Invisible,
12445 12 : )?;
12446 12 : main_tline.remote_client.wait_completion().await?;
12447 12 : let inputs = gather_inputs(
12448 12 : &tenant,
12449 12 : &limit,
12450 12 : max_retention_period,
12451 12 : &mut logical_size_cache,
12452 12 : cause,
12453 12 : &cancel,
12454 12 : &ctx,
12455 12 : )
12456 12 : .instrument(info_span!(
12457 12 : "gather_inputs",
12458 12 : tenant_id = "unknown",
12459 12 : shard_id = "unknown",
12460 12 : ))
12461 12 : .await?;
12462 12 : let ModelInputs { mut segments, .. } = inputs;
12463 168 : segments.retain(|s| s.timeline_id == TIMELINE_ID);
12464 60 : for segment in segments.iter_mut() {
12465 60 : segment.segment.parent = None; // We don't care about the parent for the test
12466 60 : segment.segment.size = None; // We don't care about the size for the test
12467 60 : }
12468 12 : assert_eq!(
12469 12 : segments,
12470 12 : [
12471 12 : SegmentMeta {
12472 12 : segment: Segment {
12473 12 : parent: None,
12474 12 : lsn: 0x10,
12475 12 : size: None,
12476 12 : needed: false,
12477 12 : },
12478 12 : timeline_id: TIMELINE_ID,
12479 12 : kind: BranchStart,
12480 12 : },
12481 12 : SegmentMeta {
12482 12 : segment: Segment {
12483 12 : parent: None,
12484 12 : lsn: 0x20,
12485 12 : size: None,
12486 12 : needed: false,
12487 12 : },
12488 12 : timeline_id: TIMELINE_ID,
12489 12 : kind: BranchPoint,
12490 12 : },
12491 12 : SegmentMeta {
12492 12 : segment: Segment {
12493 12 : parent: None,
12494 12 : lsn: 0x30,
12495 12 : size: None,
12496 12 : needed: false,
12497 12 : },
12498 12 : timeline_id: TIMELINE_ID,
12499 12 : kind: BranchPoint,
12500 12 : },
12501 12 : SegmentMeta {
12502 12 : segment: Segment {
12503 12 : parent: None,
12504 12 : lsn: 0x40,
12505 12 : size: None,
12506 12 : needed: false,
12507 12 : },
12508 12 : timeline_id: TIMELINE_ID,
12509 12 : kind: BranchPoint,
12510 12 : },
12511 12 : SegmentMeta {
12512 12 : segment: Segment {
12513 12 : parent: None,
12514 12 : lsn: 0x40, // Branch end LSN == last branch point LSN
12515 12 : size: None,
12516 12 : needed: true,
12517 12 : },
12518 12 : timeline_id: TIMELINE_ID,
12519 12 : kind: BranchEnd,
12520 12 : },
12521 12 : ]
12522 12 : );
12523 12 : Ok(())
12524 12 : }
12525 : }
|
