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 anyhow::{bail, Context};
16 : use arc_swap::ArcSwap;
17 : use camino::Utf8Path;
18 : use camino::Utf8PathBuf;
19 : use chrono::NaiveDateTime;
20 : use enumset::EnumSet;
21 : use futures::stream::FuturesUnordered;
22 : use futures::StreamExt;
23 : use pageserver_api::models;
24 : use pageserver_api::models::CompactInfoResponse;
25 : use pageserver_api::models::LsnLease;
26 : use pageserver_api::models::TimelineArchivalState;
27 : use pageserver_api::models::TimelineState;
28 : use pageserver_api::models::TopTenantShardItem;
29 : use pageserver_api::models::WalRedoManagerStatus;
30 : use pageserver_api::shard::ShardIdentity;
31 : use pageserver_api::shard::ShardStripeSize;
32 : use pageserver_api::shard::TenantShardId;
33 : use remote_storage::DownloadError;
34 : use remote_storage::GenericRemoteStorage;
35 : use remote_storage::TimeoutOrCancel;
36 : use remote_timeline_client::manifest::{
37 : OffloadedTimelineManifest, TenantManifest, LATEST_TENANT_MANIFEST_VERSION,
38 : };
39 : use remote_timeline_client::UploadQueueNotReadyError;
40 : use std::collections::BTreeMap;
41 : use std::fmt;
42 : use std::future::Future;
43 : use std::sync::atomic::AtomicBool;
44 : use std::sync::Weak;
45 : use std::time::SystemTime;
46 : use storage_broker::BrokerClientChannel;
47 : use timeline::compaction::GcCompactionQueue;
48 : use timeline::import_pgdata;
49 : use timeline::offload::offload_timeline;
50 : use timeline::offload::OffloadError;
51 : use timeline::CompactOptions;
52 : use timeline::ShutdownMode;
53 : use tokio::io::BufReader;
54 : use tokio::sync::watch;
55 : use tokio::task::JoinSet;
56 : use tokio_util::sync::CancellationToken;
57 : use tracing::*;
58 : use upload_queue::NotInitialized;
59 : use utils::backoff;
60 : use utils::circuit_breaker::CircuitBreaker;
61 : use utils::completion;
62 : use utils::crashsafe::path_with_suffix_extension;
63 : use utils::failpoint_support;
64 : use utils::fs_ext;
65 : use utils::pausable_failpoint;
66 : use utils::sync::gate::Gate;
67 : use utils::sync::gate::GateGuard;
68 : use utils::timeout::timeout_cancellable;
69 : use utils::timeout::TimeoutCancellableError;
70 : use utils::try_rcu::ArcSwapExt;
71 : use utils::zstd::create_zst_tarball;
72 : use utils::zstd::extract_zst_tarball;
73 :
74 : use self::config::AttachedLocationConfig;
75 : use self::config::AttachmentMode;
76 : use self::config::LocationConf;
77 : use self::config::TenantConf;
78 : use self::metadata::TimelineMetadata;
79 : use self::mgr::GetActiveTenantError;
80 : use self::mgr::GetTenantError;
81 : use self::remote_timeline_client::upload::{upload_index_part, upload_tenant_manifest};
82 : use self::remote_timeline_client::{RemoteTimelineClient, WaitCompletionError};
83 : use self::timeline::uninit::TimelineCreateGuard;
84 : use self::timeline::uninit::TimelineExclusionError;
85 : use self::timeline::uninit::UninitializedTimeline;
86 : use self::timeline::EvictionTaskTenantState;
87 : use self::timeline::GcCutoffs;
88 : use self::timeline::TimelineDeleteProgress;
89 : use self::timeline::TimelineResources;
90 : use self::timeline::WaitLsnError;
91 : use crate::config::PageServerConf;
92 : use crate::context::{DownloadBehavior, RequestContext};
93 : use crate::deletion_queue::DeletionQueueClient;
94 : use crate::deletion_queue::DeletionQueueError;
95 : use crate::import_datadir;
96 : use crate::is_uninit_mark;
97 : use crate::l0_flush::L0FlushGlobalState;
98 : use crate::metrics::CONCURRENT_INITDBS;
99 : use crate::metrics::INITDB_RUN_TIME;
100 : use crate::metrics::INITDB_SEMAPHORE_ACQUISITION_TIME;
101 : use crate::metrics::TENANT;
102 : use crate::metrics::{
103 : remove_tenant_metrics, BROKEN_TENANTS_SET, CIRCUIT_BREAKERS_BROKEN, CIRCUIT_BREAKERS_UNBROKEN,
104 : TENANT_STATE_METRIC, TENANT_SYNTHETIC_SIZE_METRIC,
105 : };
106 : use crate::task_mgr;
107 : use crate::task_mgr::TaskKind;
108 : use crate::tenant::config::LocationMode;
109 : use crate::tenant::config::TenantConfOpt;
110 : use crate::tenant::gc_result::GcResult;
111 : pub use crate::tenant::remote_timeline_client::index::IndexPart;
112 : use crate::tenant::remote_timeline_client::remote_initdb_archive_path;
113 : use crate::tenant::remote_timeline_client::MaybeDeletedIndexPart;
114 : use crate::tenant::remote_timeline_client::INITDB_PATH;
115 : use crate::tenant::storage_layer::DeltaLayer;
116 : use crate::tenant::storage_layer::ImageLayer;
117 : use crate::walingest::WalLagCooldown;
118 : use crate::walredo;
119 : use crate::InitializationOrder;
120 : use std::collections::hash_map::Entry;
121 : use std::collections::HashMap;
122 : use std::collections::HashSet;
123 : use std::fmt::Debug;
124 : use std::fmt::Display;
125 : use std::fs;
126 : use std::fs::File;
127 : use std::sync::atomic::{AtomicU64, Ordering};
128 : use std::sync::Arc;
129 : use std::sync::Mutex;
130 : use std::time::{Duration, Instant};
131 :
132 : use crate::span;
133 : use crate::tenant::timeline::delete::DeleteTimelineFlow;
134 : use crate::tenant::timeline::uninit::cleanup_timeline_directory;
135 : use crate::virtual_file::VirtualFile;
136 : use crate::walredo::PostgresRedoManager;
137 : use crate::TEMP_FILE_SUFFIX;
138 : use once_cell::sync::Lazy;
139 : pub use pageserver_api::models::TenantState;
140 : use tokio::sync::Semaphore;
141 :
142 0 : static INIT_DB_SEMAPHORE: Lazy<Semaphore> = Lazy::new(|| Semaphore::new(8));
143 : use utils::{
144 : crashsafe,
145 : generation::Generation,
146 : id::TimelineId,
147 : lsn::{Lsn, RecordLsn},
148 : };
149 :
150 : pub mod blob_io;
151 : pub mod block_io;
152 : pub mod vectored_blob_io;
153 :
154 : pub mod disk_btree;
155 : pub(crate) mod ephemeral_file;
156 : pub mod layer_map;
157 :
158 : pub mod metadata;
159 : pub mod remote_timeline_client;
160 : pub mod storage_layer;
161 :
162 : pub mod checks;
163 : pub mod config;
164 : pub mod mgr;
165 : pub mod secondary;
166 : pub mod tasks;
167 : pub mod upload_queue;
168 :
169 : pub(crate) mod timeline;
170 :
171 : pub mod size;
172 :
173 : mod gc_block;
174 : mod gc_result;
175 : pub(crate) mod throttle;
176 :
177 : pub(crate) use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
178 : pub(crate) use timeline::{LogicalSizeCalculationCause, PageReconstructError, Timeline};
179 :
180 : // re-export for use in walreceiver
181 : pub use crate::tenant::timeline::WalReceiverInfo;
182 :
183 : /// The "tenants" part of `tenants/<tenant>/timelines...`
184 : pub const TENANTS_SEGMENT_NAME: &str = "tenants";
185 :
186 : /// Parts of the `.neon/tenants/<tenant_id>/timelines/<timeline_id>` directory prefix.
187 : pub const TIMELINES_SEGMENT_NAME: &str = "timelines";
188 :
189 : /// References to shared objects that are passed into each tenant, such
190 : /// as the shared remote storage client and process initialization state.
191 : #[derive(Clone)]
192 : pub struct TenantSharedResources {
193 : pub broker_client: storage_broker::BrokerClientChannel,
194 : pub remote_storage: GenericRemoteStorage,
195 : pub deletion_queue_client: DeletionQueueClient,
196 : pub l0_flush_global_state: L0FlushGlobalState,
197 : }
198 :
199 : /// A [`Tenant`] is really an _attached_ tenant. The configuration
200 : /// for an attached tenant is a subset of the [`LocationConf`], represented
201 : /// in this struct.
202 : #[derive(Clone)]
203 : pub(super) struct AttachedTenantConf {
204 : tenant_conf: TenantConfOpt,
205 : location: AttachedLocationConfig,
206 : /// The deadline before which we are blocked from GC so that
207 : /// leases have a chance to be renewed.
208 : lsn_lease_deadline: Option<tokio::time::Instant>,
209 : }
210 :
211 : impl AttachedTenantConf {
212 440 : fn new(tenant_conf: TenantConfOpt, location: AttachedLocationConfig) -> Self {
213 : // Sets a deadline before which we cannot proceed to GC due to lsn lease.
214 : //
215 : // We do this as the leases mapping are not persisted to disk. By delaying GC by lease
216 : // length, we guarantee that all the leases we granted before will have a chance to renew
217 : // when we run GC for the first time after restart / transition from AttachedMulti to AttachedSingle.
218 440 : let lsn_lease_deadline = if location.attach_mode == AttachmentMode::Single {
219 440 : Some(
220 440 : tokio::time::Instant::now()
221 440 : + tenant_conf
222 440 : .lsn_lease_length
223 440 : .unwrap_or(LsnLease::DEFAULT_LENGTH),
224 440 : )
225 : } else {
226 : // We don't use `lsn_lease_deadline` to delay GC in AttachedMulti and AttachedStale
227 : // because we don't do GC in these modes.
228 0 : None
229 : };
230 :
231 440 : Self {
232 440 : tenant_conf,
233 440 : location,
234 440 : lsn_lease_deadline,
235 440 : }
236 440 : }
237 :
238 440 : fn try_from(location_conf: LocationConf) -> anyhow::Result<Self> {
239 440 : match &location_conf.mode {
240 440 : LocationMode::Attached(attach_conf) => {
241 440 : Ok(Self::new(location_conf.tenant_conf, *attach_conf))
242 : }
243 : LocationMode::Secondary(_) => {
244 0 : anyhow::bail!("Attempted to construct AttachedTenantConf from a LocationConf in secondary mode")
245 : }
246 : }
247 440 : }
248 :
249 1524 : fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
250 1524 : self.lsn_lease_deadline
251 1524 : .map(|d| tokio::time::Instant::now() < d)
252 1524 : .unwrap_or(false)
253 1524 : }
254 : }
255 : struct TimelinePreload {
256 : timeline_id: TimelineId,
257 : client: RemoteTimelineClient,
258 : index_part: Result<MaybeDeletedIndexPart, DownloadError>,
259 : }
260 :
261 : pub(crate) struct TenantPreload {
262 : tenant_manifest: TenantManifest,
263 : /// Map from timeline ID to a possible timeline preload. It is None iff the timeline is offloaded according to the manifest.
264 : timelines: HashMap<TimelineId, Option<TimelinePreload>>,
265 : }
266 :
267 : /// When we spawn a tenant, there is a special mode for tenant creation that
268 : /// avoids trying to read anything from remote storage.
269 : pub(crate) enum SpawnMode {
270 : /// Activate as soon as possible
271 : Eager,
272 : /// Lazy activation in the background, with the option to skip the queue if the need comes up
273 : Lazy,
274 : }
275 :
276 : ///
277 : /// Tenant consists of multiple timelines. Keep them in a hash table.
278 : ///
279 : pub struct Tenant {
280 : // Global pageserver config parameters
281 : pub conf: &'static PageServerConf,
282 :
283 : /// The value creation timestamp, used to measure activation delay, see:
284 : /// <https://github.com/neondatabase/neon/issues/4025>
285 : constructed_at: Instant,
286 :
287 : state: watch::Sender<TenantState>,
288 :
289 : // Overridden tenant-specific config parameters.
290 : // We keep TenantConfOpt sturct here to preserve the information
291 : // about parameters that are not set.
292 : // This is necessary to allow global config updates.
293 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
294 :
295 : tenant_shard_id: TenantShardId,
296 :
297 : // The detailed sharding information, beyond the number/count in tenant_shard_id
298 : shard_identity: ShardIdentity,
299 :
300 : /// The remote storage generation, used to protect S3 objects from split-brain.
301 : /// Does not change over the lifetime of the [`Tenant`] object.
302 : ///
303 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
304 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
305 : generation: Generation,
306 :
307 : timelines: Mutex<HashMap<TimelineId, Arc<Timeline>>>,
308 :
309 : /// During timeline creation, we first insert the TimelineId to the
310 : /// creating map, then `timelines`, then remove it from the creating map.
311 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
312 : timelines_creating: std::sync::Mutex<HashSet<TimelineId>>,
313 :
314 : /// Possibly offloaded and archived timelines
315 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
316 : timelines_offloaded: Mutex<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
317 :
318 : /// Serialize writes of the tenant manifest to remote storage. If there are concurrent operations
319 : /// affecting the manifest, such as timeline deletion and timeline offload, they must wait for
320 : /// each other (this could be optimized to coalesce writes if necessary).
321 : ///
322 : /// The contents of the Mutex are the last manifest we successfully uploaded
323 : tenant_manifest_upload: tokio::sync::Mutex<Option<TenantManifest>>,
324 :
325 : // This mutex prevents creation of new timelines during GC.
326 : // Adding yet another mutex (in addition to `timelines`) is needed because holding
327 : // `timelines` mutex during all GC iteration
328 : // may block for a long time `get_timeline`, `get_timelines_state`,... and other operations
329 : // with timelines, which in turn may cause dropping replication connection, expiration of wait_for_lsn
330 : // timeout...
331 : gc_cs: tokio::sync::Mutex<()>,
332 : walredo_mgr: Option<Arc<WalRedoManager>>,
333 :
334 : // provides access to timeline data sitting in the remote storage
335 : pub(crate) remote_storage: GenericRemoteStorage,
336 :
337 : // Access to global deletion queue for when this tenant wants to schedule a deletion
338 : deletion_queue_client: DeletionQueueClient,
339 :
340 : /// Cached logical sizes updated updated on each [`Tenant::gather_size_inputs`].
341 : cached_logical_sizes: tokio::sync::Mutex<HashMap<(TimelineId, Lsn), u64>>,
342 : cached_synthetic_tenant_size: Arc<AtomicU64>,
343 :
344 : eviction_task_tenant_state: tokio::sync::Mutex<EvictionTaskTenantState>,
345 :
346 : /// Track repeated failures to compact, so that we can back off.
347 : /// Overhead of mutex is acceptable because compaction is done with a multi-second period.
348 : compaction_circuit_breaker: std::sync::Mutex<CircuitBreaker>,
349 :
350 : /// Scheduled gc-compaction tasks.
351 : scheduled_compaction_tasks: std::sync::Mutex<HashMap<TimelineId, Arc<GcCompactionQueue>>>,
352 :
353 : /// If the tenant is in Activating state, notify this to encourage it
354 : /// to proceed to Active as soon as possible, rather than waiting for lazy
355 : /// background warmup.
356 : pub(crate) activate_now_sem: tokio::sync::Semaphore,
357 :
358 : /// Time it took for the tenant to activate. Zero if not active yet.
359 : attach_wal_lag_cooldown: Arc<std::sync::OnceLock<WalLagCooldown>>,
360 :
361 : // Cancellation token fires when we have entered shutdown(). This is a parent of
362 : // Timelines' cancellation token.
363 : pub(crate) cancel: CancellationToken,
364 :
365 : // Users of the Tenant such as the page service must take this Gate to avoid
366 : // trying to use a Tenant which is shutting down.
367 : pub(crate) gate: Gate,
368 :
369 : /// Throttle applied at the top of [`Timeline::get`].
370 : /// All [`Tenant::timelines`] of a given [`Tenant`] instance share the same [`throttle::Throttle`] instance.
371 : pub(crate) pagestream_throttle: Arc<throttle::Throttle>,
372 :
373 : pub(crate) pagestream_throttle_metrics: Arc<crate::metrics::tenant_throttling::Pagestream>,
374 :
375 : /// An ongoing timeline detach concurrency limiter.
376 : ///
377 : /// As a tenant will likely be restarted as part of timeline detach ancestor it makes no sense
378 : /// to have two running at the same time. A different one can be started if an earlier one
379 : /// has failed for whatever reason.
380 : ongoing_timeline_detach: std::sync::Mutex<Option<(TimelineId, utils::completion::Barrier)>>,
381 :
382 : /// `index_part.json` based gc blocking reason tracking.
383 : ///
384 : /// New gc iterations must start a new iteration by acquiring `GcBlock::start` before
385 : /// proceeding.
386 : pub(crate) gc_block: gc_block::GcBlock,
387 :
388 : l0_flush_global_state: L0FlushGlobalState,
389 : }
390 : impl std::fmt::Debug for Tenant {
391 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
392 0 : write!(f, "{} ({})", self.tenant_shard_id, self.current_state())
393 0 : }
394 : }
395 :
396 : pub(crate) enum WalRedoManager {
397 : Prod(WalredoManagerId, PostgresRedoManager),
398 : #[cfg(test)]
399 : Test(harness::TestRedoManager),
400 : }
401 :
402 : #[derive(thiserror::Error, Debug)]
403 : #[error("pageserver is shutting down")]
404 : pub(crate) struct GlobalShutDown;
405 :
406 : impl WalRedoManager {
407 0 : pub(crate) fn new(mgr: PostgresRedoManager) -> Result<Arc<Self>, GlobalShutDown> {
408 0 : let id = WalredoManagerId::next();
409 0 : let arc = Arc::new(Self::Prod(id, mgr));
410 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
411 0 : match &mut *guard {
412 0 : Some(map) => {
413 0 : map.insert(id, Arc::downgrade(&arc));
414 0 : Ok(arc)
415 : }
416 0 : None => Err(GlobalShutDown),
417 : }
418 0 : }
419 : }
420 :
421 : impl Drop for WalRedoManager {
422 20 : fn drop(&mut self) {
423 20 : match self {
424 0 : Self::Prod(id, _) => {
425 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
426 0 : if let Some(map) = &mut *guard {
427 0 : map.remove(id).expect("new() registers, drop() unregisters");
428 0 : }
429 : }
430 : #[cfg(test)]
431 20 : Self::Test(_) => {
432 20 : // Not applicable to test redo manager
433 20 : }
434 : }
435 20 : }
436 : }
437 :
438 : /// Global registry of all walredo managers so that [`crate::shutdown_pageserver`] can shut down
439 : /// the walredo processes outside of the regular order.
440 : ///
441 : /// This is necessary to work around a systemd bug where it freezes if there are
442 : /// walredo processes left => <https://github.com/neondatabase/cloud/issues/11387>
443 : #[allow(clippy::type_complexity)]
444 : pub(crate) static WALREDO_MANAGERS: once_cell::sync::Lazy<
445 : Mutex<Option<HashMap<WalredoManagerId, Weak<WalRedoManager>>>>,
446 0 : > = once_cell::sync::Lazy::new(|| Mutex::new(Some(HashMap::new())));
447 : #[derive(PartialEq, Eq, Hash, Clone, Copy, Debug)]
448 : pub(crate) struct WalredoManagerId(u64);
449 : impl WalredoManagerId {
450 0 : pub fn next() -> Self {
451 : static NEXT: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(1);
452 0 : let id = NEXT.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
453 0 : if id == 0 {
454 0 : panic!("WalredoManagerId::new() returned 0, indicating wraparound, risking it's no longer unique");
455 0 : }
456 0 : Self(id)
457 0 : }
458 : }
459 :
460 : #[cfg(test)]
461 : impl From<harness::TestRedoManager> for WalRedoManager {
462 440 : fn from(mgr: harness::TestRedoManager) -> Self {
463 440 : Self::Test(mgr)
464 440 : }
465 : }
466 :
467 : impl WalRedoManager {
468 12 : pub(crate) async fn shutdown(&self) -> bool {
469 12 : match self {
470 0 : Self::Prod(_, mgr) => mgr.shutdown().await,
471 : #[cfg(test)]
472 : Self::Test(_) => {
473 : // Not applicable to test redo manager
474 12 : true
475 : }
476 : }
477 12 : }
478 :
479 0 : pub(crate) fn maybe_quiesce(&self, idle_timeout: Duration) {
480 0 : match self {
481 0 : Self::Prod(_, mgr) => mgr.maybe_quiesce(idle_timeout),
482 0 : #[cfg(test)]
483 0 : Self::Test(_) => {
484 0 : // Not applicable to test redo manager
485 0 : }
486 0 : }
487 0 : }
488 :
489 : /// # Cancel-Safety
490 : ///
491 : /// This method is cancellation-safe.
492 1636 : pub async fn request_redo(
493 1636 : &self,
494 1636 : key: pageserver_api::key::Key,
495 1636 : lsn: Lsn,
496 1636 : base_img: Option<(Lsn, bytes::Bytes)>,
497 1636 : records: Vec<(Lsn, pageserver_api::record::NeonWalRecord)>,
498 1636 : pg_version: u32,
499 1636 : ) -> Result<bytes::Bytes, walredo::Error> {
500 1636 : match self {
501 0 : Self::Prod(_, mgr) => {
502 0 : mgr.request_redo(key, lsn, base_img, records, pg_version)
503 0 : .await
504 : }
505 : #[cfg(test)]
506 1636 : Self::Test(mgr) => {
507 1636 : mgr.request_redo(key, lsn, base_img, records, pg_version)
508 1636 : .await
509 : }
510 : }
511 1636 : }
512 :
513 0 : pub(crate) fn status(&self) -> Option<WalRedoManagerStatus> {
514 0 : match self {
515 0 : WalRedoManager::Prod(_, m) => Some(m.status()),
516 0 : #[cfg(test)]
517 0 : WalRedoManager::Test(_) => None,
518 0 : }
519 0 : }
520 : }
521 :
522 : /// A very lightweight memory representation of an offloaded timeline.
523 : ///
524 : /// We need to store the list of offloaded timelines so that we can perform operations on them,
525 : /// like unoffloading them, or (at a later date), decide to perform flattening.
526 : /// This type has a much smaller memory impact than [`Timeline`], and thus we can store many
527 : /// more offloaded timelines than we can manage ones that aren't.
528 : pub struct OffloadedTimeline {
529 : pub tenant_shard_id: TenantShardId,
530 : pub timeline_id: TimelineId,
531 : pub ancestor_timeline_id: Option<TimelineId>,
532 : /// Whether to retain the branch lsn at the ancestor or not
533 : pub ancestor_retain_lsn: Option<Lsn>,
534 :
535 : /// When the timeline was archived.
536 : ///
537 : /// Present for future flattening deliberations.
538 : pub archived_at: NaiveDateTime,
539 :
540 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
541 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
542 : pub delete_progress: TimelineDeleteProgress,
543 :
544 : /// Part of the `OffloadedTimeline` object's lifecycle: this needs to be set before we drop it
545 : pub deleted_from_ancestor: AtomicBool,
546 : }
547 :
548 : impl OffloadedTimeline {
549 : /// Obtains an offloaded timeline from a given timeline object.
550 : ///
551 : /// Returns `None` if the `archived_at` flag couldn't be obtained, i.e.
552 : /// the timeline is not in a stopped state.
553 : /// Panics if the timeline is not archived.
554 4 : fn from_timeline(timeline: &Timeline) -> Result<Self, UploadQueueNotReadyError> {
555 4 : let (ancestor_retain_lsn, ancestor_timeline_id) =
556 4 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
557 4 : let ancestor_lsn = timeline.get_ancestor_lsn();
558 4 : let ancestor_timeline_id = ancestor_timeline.timeline_id;
559 4 : let mut gc_info = ancestor_timeline.gc_info.write().unwrap();
560 4 : gc_info.insert_child(timeline.timeline_id, ancestor_lsn, MaybeOffloaded::Yes);
561 4 : (Some(ancestor_lsn), Some(ancestor_timeline_id))
562 : } else {
563 0 : (None, None)
564 : };
565 4 : let archived_at = timeline
566 4 : .remote_client
567 4 : .archived_at_stopped_queue()?
568 4 : .expect("must be called on an archived timeline");
569 4 : Ok(Self {
570 4 : tenant_shard_id: timeline.tenant_shard_id,
571 4 : timeline_id: timeline.timeline_id,
572 4 : ancestor_timeline_id,
573 4 : ancestor_retain_lsn,
574 4 : archived_at,
575 4 :
576 4 : delete_progress: timeline.delete_progress.clone(),
577 4 : deleted_from_ancestor: AtomicBool::new(false),
578 4 : })
579 4 : }
580 0 : fn from_manifest(tenant_shard_id: TenantShardId, manifest: &OffloadedTimelineManifest) -> Self {
581 0 : // We expect to reach this case in tenant loading, where the `retain_lsn` is populated in the parent's `gc_info`
582 0 : // by the `initialize_gc_info` function.
583 0 : let OffloadedTimelineManifest {
584 0 : timeline_id,
585 0 : ancestor_timeline_id,
586 0 : ancestor_retain_lsn,
587 0 : archived_at,
588 0 : } = *manifest;
589 0 : Self {
590 0 : tenant_shard_id,
591 0 : timeline_id,
592 0 : ancestor_timeline_id,
593 0 : ancestor_retain_lsn,
594 0 : archived_at,
595 0 : delete_progress: TimelineDeleteProgress::default(),
596 0 : deleted_from_ancestor: AtomicBool::new(false),
597 0 : }
598 0 : }
599 4 : fn manifest(&self) -> OffloadedTimelineManifest {
600 4 : let Self {
601 4 : timeline_id,
602 4 : ancestor_timeline_id,
603 4 : ancestor_retain_lsn,
604 4 : archived_at,
605 4 : ..
606 4 : } = self;
607 4 : OffloadedTimelineManifest {
608 4 : timeline_id: *timeline_id,
609 4 : ancestor_timeline_id: *ancestor_timeline_id,
610 4 : ancestor_retain_lsn: *ancestor_retain_lsn,
611 4 : archived_at: *archived_at,
612 4 : }
613 4 : }
614 : /// Delete this timeline's retain_lsn from its ancestor, if present in the given tenant
615 0 : fn delete_from_ancestor_with_timelines(
616 0 : &self,
617 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
618 0 : ) {
619 0 : if let (Some(_retain_lsn), Some(ancestor_timeline_id)) =
620 0 : (self.ancestor_retain_lsn, self.ancestor_timeline_id)
621 : {
622 0 : if let Some((_, ancestor_timeline)) = timelines
623 0 : .iter()
624 0 : .find(|(tid, _tl)| **tid == ancestor_timeline_id)
625 : {
626 0 : let removal_happened = ancestor_timeline
627 0 : .gc_info
628 0 : .write()
629 0 : .unwrap()
630 0 : .remove_child_offloaded(self.timeline_id);
631 0 : if !removal_happened {
632 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
633 0 : "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
634 0 : }
635 0 : }
636 0 : }
637 0 : self.deleted_from_ancestor.store(true, Ordering::Release);
638 0 : }
639 : /// Call [`Self::delete_from_ancestor_with_timelines`] instead if possible.
640 : ///
641 : /// As the entire tenant is being dropped, don't bother deregistering the `retain_lsn` from the ancestor.
642 4 : fn defuse_for_tenant_drop(&self) {
643 4 : self.deleted_from_ancestor.store(true, Ordering::Release);
644 4 : }
645 : }
646 :
647 : impl fmt::Debug for OffloadedTimeline {
648 0 : fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
649 0 : write!(f, "OffloadedTimeline<{}>", self.timeline_id)
650 0 : }
651 : }
652 :
653 : impl Drop for OffloadedTimeline {
654 4 : fn drop(&mut self) {
655 4 : if !self.deleted_from_ancestor.load(Ordering::Acquire) {
656 0 : tracing::warn!(
657 0 : "offloaded timeline {} was dropped without having cleaned it up at the ancestor",
658 : self.timeline_id
659 : );
660 4 : }
661 4 : }
662 : }
663 :
664 : #[derive(Copy, Clone, PartialEq, Eq, Hash, Debug)]
665 : pub enum MaybeOffloaded {
666 : Yes,
667 : No,
668 : }
669 :
670 : #[derive(Clone, Debug)]
671 : pub enum TimelineOrOffloaded {
672 : Timeline(Arc<Timeline>),
673 : Offloaded(Arc<OffloadedTimeline>),
674 : }
675 :
676 : impl TimelineOrOffloaded {
677 0 : pub fn arc_ref(&self) -> TimelineOrOffloadedArcRef<'_> {
678 0 : match self {
679 0 : TimelineOrOffloaded::Timeline(timeline) => {
680 0 : TimelineOrOffloadedArcRef::Timeline(timeline)
681 : }
682 0 : TimelineOrOffloaded::Offloaded(offloaded) => {
683 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded)
684 : }
685 : }
686 0 : }
687 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
688 0 : self.arc_ref().tenant_shard_id()
689 0 : }
690 0 : pub fn timeline_id(&self) -> TimelineId {
691 0 : self.arc_ref().timeline_id()
692 0 : }
693 4 : pub fn delete_progress(&self) -> &Arc<tokio::sync::Mutex<DeleteTimelineFlow>> {
694 4 : match self {
695 4 : TimelineOrOffloaded::Timeline(timeline) => &timeline.delete_progress,
696 0 : TimelineOrOffloaded::Offloaded(offloaded) => &offloaded.delete_progress,
697 : }
698 4 : }
699 0 : fn maybe_remote_client(&self) -> Option<Arc<RemoteTimelineClient>> {
700 0 : match self {
701 0 : TimelineOrOffloaded::Timeline(timeline) => Some(timeline.remote_client.clone()),
702 0 : TimelineOrOffloaded::Offloaded(_offloaded) => None,
703 : }
704 0 : }
705 : }
706 :
707 : pub enum TimelineOrOffloadedArcRef<'a> {
708 : Timeline(&'a Arc<Timeline>),
709 : Offloaded(&'a Arc<OffloadedTimeline>),
710 : }
711 :
712 : impl TimelineOrOffloadedArcRef<'_> {
713 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
714 0 : match self {
715 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.tenant_shard_id,
716 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.tenant_shard_id,
717 : }
718 0 : }
719 0 : pub fn timeline_id(&self) -> TimelineId {
720 0 : match self {
721 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.timeline_id,
722 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.timeline_id,
723 : }
724 0 : }
725 : }
726 :
727 : impl<'a> From<&'a Arc<Timeline>> for TimelineOrOffloadedArcRef<'a> {
728 0 : fn from(timeline: &'a Arc<Timeline>) -> Self {
729 0 : Self::Timeline(timeline)
730 0 : }
731 : }
732 :
733 : impl<'a> From<&'a Arc<OffloadedTimeline>> for TimelineOrOffloadedArcRef<'a> {
734 0 : fn from(timeline: &'a Arc<OffloadedTimeline>) -> Self {
735 0 : Self::Offloaded(timeline)
736 0 : }
737 : }
738 :
739 : #[derive(Debug, thiserror::Error, PartialEq, Eq)]
740 : pub enum GetTimelineError {
741 : #[error("Timeline is shutting down")]
742 : ShuttingDown,
743 : #[error("Timeline {tenant_id}/{timeline_id} is not active, state: {state:?}")]
744 : NotActive {
745 : tenant_id: TenantShardId,
746 : timeline_id: TimelineId,
747 : state: TimelineState,
748 : },
749 : #[error("Timeline {tenant_id}/{timeline_id} was not found")]
750 : NotFound {
751 : tenant_id: TenantShardId,
752 : timeline_id: TimelineId,
753 : },
754 : }
755 :
756 : #[derive(Debug, thiserror::Error)]
757 : pub enum LoadLocalTimelineError {
758 : #[error("FailedToLoad")]
759 : Load(#[source] anyhow::Error),
760 : #[error("FailedToResumeDeletion")]
761 : ResumeDeletion(#[source] anyhow::Error),
762 : }
763 :
764 : #[derive(thiserror::Error)]
765 : pub enum DeleteTimelineError {
766 : #[error("NotFound")]
767 : NotFound,
768 :
769 : #[error("HasChildren")]
770 : HasChildren(Vec<TimelineId>),
771 :
772 : #[error("Timeline deletion is already in progress")]
773 : AlreadyInProgress(Arc<tokio::sync::Mutex<DeleteTimelineFlow>>),
774 :
775 : #[error("Cancelled")]
776 : Cancelled,
777 :
778 : #[error(transparent)]
779 : Other(#[from] anyhow::Error),
780 : }
781 :
782 : impl Debug for DeleteTimelineError {
783 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
784 0 : match self {
785 0 : Self::NotFound => write!(f, "NotFound"),
786 0 : Self::HasChildren(c) => f.debug_tuple("HasChildren").field(c).finish(),
787 0 : Self::AlreadyInProgress(_) => f.debug_tuple("AlreadyInProgress").finish(),
788 0 : Self::Cancelled => f.debug_tuple("Cancelled").finish(),
789 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
790 : }
791 0 : }
792 : }
793 :
794 : #[derive(thiserror::Error)]
795 : pub enum TimelineArchivalError {
796 : #[error("NotFound")]
797 : NotFound,
798 :
799 : #[error("Timeout")]
800 : Timeout,
801 :
802 : #[error("Cancelled")]
803 : Cancelled,
804 :
805 : #[error("ancestor is archived: {}", .0)]
806 : HasArchivedParent(TimelineId),
807 :
808 : #[error("HasUnarchivedChildren")]
809 : HasUnarchivedChildren(Vec<TimelineId>),
810 :
811 : #[error("Timeline archival is already in progress")]
812 : AlreadyInProgress,
813 :
814 : #[error(transparent)]
815 : Other(anyhow::Error),
816 : }
817 :
818 : #[derive(thiserror::Error, Debug)]
819 : pub(crate) enum TenantManifestError {
820 : #[error("Remote storage error: {0}")]
821 : RemoteStorage(anyhow::Error),
822 :
823 : #[error("Cancelled")]
824 : Cancelled,
825 : }
826 :
827 : impl From<TenantManifestError> for TimelineArchivalError {
828 0 : fn from(e: TenantManifestError) -> Self {
829 0 : match e {
830 0 : TenantManifestError::RemoteStorage(e) => Self::Other(e),
831 0 : TenantManifestError::Cancelled => Self::Cancelled,
832 : }
833 0 : }
834 : }
835 :
836 : impl Debug for TimelineArchivalError {
837 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
838 0 : match self {
839 0 : Self::NotFound => write!(f, "NotFound"),
840 0 : Self::Timeout => write!(f, "Timeout"),
841 0 : Self::Cancelled => write!(f, "Cancelled"),
842 0 : Self::HasArchivedParent(p) => f.debug_tuple("HasArchivedParent").field(p).finish(),
843 0 : Self::HasUnarchivedChildren(c) => {
844 0 : f.debug_tuple("HasUnarchivedChildren").field(c).finish()
845 : }
846 0 : Self::AlreadyInProgress => f.debug_tuple("AlreadyInProgress").finish(),
847 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
848 : }
849 0 : }
850 : }
851 :
852 : pub enum SetStoppingError {
853 : AlreadyStopping(completion::Barrier),
854 : Broken,
855 : }
856 :
857 : impl Debug for SetStoppingError {
858 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
859 0 : match self {
860 0 : Self::AlreadyStopping(_) => f.debug_tuple("AlreadyStopping").finish(),
861 0 : Self::Broken => write!(f, "Broken"),
862 : }
863 0 : }
864 : }
865 :
866 : /// Arguments to [`Tenant::create_timeline`].
867 : ///
868 : /// Not usable as an idempotency key for timeline creation because if [`CreateTimelineParamsBranch::ancestor_start_lsn`]
869 : /// is `None`, the result of the timeline create call is not deterministic.
870 : ///
871 : /// See [`CreateTimelineIdempotency`] for an idempotency key.
872 : #[derive(Debug)]
873 : pub(crate) enum CreateTimelineParams {
874 : Bootstrap(CreateTimelineParamsBootstrap),
875 : Branch(CreateTimelineParamsBranch),
876 : ImportPgdata(CreateTimelineParamsImportPgdata),
877 : }
878 :
879 : #[derive(Debug)]
880 : pub(crate) struct CreateTimelineParamsBootstrap {
881 : pub(crate) new_timeline_id: TimelineId,
882 : pub(crate) existing_initdb_timeline_id: Option<TimelineId>,
883 : pub(crate) pg_version: u32,
884 : }
885 :
886 : /// NB: See comment on [`CreateTimelineIdempotency::Branch`] for why there's no `pg_version` here.
887 : #[derive(Debug)]
888 : pub(crate) struct CreateTimelineParamsBranch {
889 : pub(crate) new_timeline_id: TimelineId,
890 : pub(crate) ancestor_timeline_id: TimelineId,
891 : pub(crate) ancestor_start_lsn: Option<Lsn>,
892 : }
893 :
894 : #[derive(Debug)]
895 : pub(crate) struct CreateTimelineParamsImportPgdata {
896 : pub(crate) new_timeline_id: TimelineId,
897 : pub(crate) location: import_pgdata::index_part_format::Location,
898 : pub(crate) idempotency_key: import_pgdata::index_part_format::IdempotencyKey,
899 : }
900 :
901 : /// What is used to determine idempotency of a [`Tenant::create_timeline`] call in [`Tenant::start_creating_timeline`] in [`Tenant::start_creating_timeline`].
902 : ///
903 : /// Each [`Timeline`] object holds [`Self`] as an immutable property in [`Timeline::create_idempotency`].
904 : ///
905 : /// We lower timeline creation requests to [`Self`], and then use [`PartialEq::eq`] to compare [`Timeline::create_idempotency`] with the request.
906 : /// If they are equal, we return a reference to the existing timeline, otherwise it's an idempotency conflict.
907 : ///
908 : /// There is special treatment for [`Self::FailWithConflict`] to always return an idempotency conflict.
909 : /// It would be nice to have more advanced derive macros to make that special treatment declarative.
910 : ///
911 : /// Notes:
912 : /// - Unlike [`CreateTimelineParams`], ancestor LSN is fixed, so, branching will be at a deterministic LSN.
913 : /// - We make some trade-offs though, e.g., [`CreateTimelineParamsBootstrap::existing_initdb_timeline_id`]
914 : /// is not considered for idempotency. We can improve on this over time if we deem it necessary.
915 : ///
916 : #[derive(Debug, Clone, PartialEq, Eq)]
917 : pub(crate) enum CreateTimelineIdempotency {
918 : /// NB: special treatment, see comment in [`Self`].
919 : FailWithConflict,
920 : Bootstrap {
921 : pg_version: u32,
922 : },
923 : /// NB: branches always have the same `pg_version` as their ancestor.
924 : /// While [`pageserver_api::models::TimelineCreateRequestMode::Branch::pg_version`]
925 : /// exists as a field, and is set by cplane, it has always been ignored by pageserver when
926 : /// determining the child branch pg_version.
927 : Branch {
928 : ancestor_timeline_id: TimelineId,
929 : ancestor_start_lsn: Lsn,
930 : },
931 : ImportPgdata(CreatingTimelineIdempotencyImportPgdata),
932 : }
933 :
934 : #[derive(Debug, Clone, PartialEq, Eq)]
935 : pub(crate) struct CreatingTimelineIdempotencyImportPgdata {
936 : idempotency_key: import_pgdata::index_part_format::IdempotencyKey,
937 : }
938 :
939 : /// What is returned by [`Tenant::start_creating_timeline`].
940 : #[must_use]
941 : enum StartCreatingTimelineResult {
942 : CreateGuard(TimelineCreateGuard),
943 : Idempotent(Arc<Timeline>),
944 : }
945 :
946 : enum TimelineInitAndSyncResult {
947 : ReadyToActivate(Arc<Timeline>),
948 : NeedsSpawnImportPgdata(TimelineInitAndSyncNeedsSpawnImportPgdata),
949 : }
950 :
951 : impl TimelineInitAndSyncResult {
952 0 : fn ready_to_activate(self) -> Option<Arc<Timeline>> {
953 0 : match self {
954 0 : Self::ReadyToActivate(timeline) => Some(timeline),
955 0 : _ => None,
956 : }
957 0 : }
958 : }
959 :
960 : #[must_use]
961 : struct TimelineInitAndSyncNeedsSpawnImportPgdata {
962 : timeline: Arc<Timeline>,
963 : import_pgdata: import_pgdata::index_part_format::Root,
964 : guard: TimelineCreateGuard,
965 : }
966 :
967 : /// What is returned by [`Tenant::create_timeline`].
968 : enum CreateTimelineResult {
969 : Created(Arc<Timeline>),
970 : Idempotent(Arc<Timeline>),
971 : /// IMPORTANT: This [`Arc<Timeline>`] object is not in [`Tenant::timelines`] when
972 : /// we return this result, nor will this concrete object ever be added there.
973 : /// Cf method comment on [`Tenant::create_timeline_import_pgdata`].
974 : ImportSpawned(Arc<Timeline>),
975 : }
976 :
977 : impl CreateTimelineResult {
978 0 : fn discriminant(&self) -> &'static str {
979 0 : match self {
980 0 : Self::Created(_) => "Created",
981 0 : Self::Idempotent(_) => "Idempotent",
982 0 : Self::ImportSpawned(_) => "ImportSpawned",
983 : }
984 0 : }
985 0 : fn timeline(&self) -> &Arc<Timeline> {
986 0 : match self {
987 0 : Self::Created(t) | Self::Idempotent(t) | Self::ImportSpawned(t) => t,
988 0 : }
989 0 : }
990 : /// Unit test timelines aren't activated, test has to do it if it needs to.
991 : #[cfg(test)]
992 460 : fn into_timeline_for_test(self) -> Arc<Timeline> {
993 460 : match self {
994 460 : Self::Created(t) | Self::Idempotent(t) | Self::ImportSpawned(t) => t,
995 460 : }
996 460 : }
997 : }
998 :
999 : #[derive(thiserror::Error, Debug)]
1000 : pub enum CreateTimelineError {
1001 : #[error("creation of timeline with the given ID is in progress")]
1002 : AlreadyCreating,
1003 : #[error("timeline already exists with different parameters")]
1004 : Conflict,
1005 : #[error(transparent)]
1006 : AncestorLsn(anyhow::Error),
1007 : #[error("ancestor timeline is not active")]
1008 : AncestorNotActive,
1009 : #[error("ancestor timeline is archived")]
1010 : AncestorArchived,
1011 : #[error("tenant shutting down")]
1012 : ShuttingDown,
1013 : #[error(transparent)]
1014 : Other(#[from] anyhow::Error),
1015 : }
1016 :
1017 : #[derive(thiserror::Error, Debug)]
1018 : pub enum InitdbError {
1019 : #[error("Operation was cancelled")]
1020 : Cancelled,
1021 : #[error(transparent)]
1022 : Other(anyhow::Error),
1023 : #[error(transparent)]
1024 : Inner(postgres_initdb::Error),
1025 : }
1026 :
1027 : enum CreateTimelineCause {
1028 : Load,
1029 : Delete,
1030 : }
1031 :
1032 : enum LoadTimelineCause {
1033 : Attach,
1034 : Unoffload,
1035 : ImportPgdata {
1036 : create_guard: TimelineCreateGuard,
1037 : activate: ActivateTimelineArgs,
1038 : },
1039 : }
1040 :
1041 : #[derive(thiserror::Error, Debug)]
1042 : pub(crate) enum GcError {
1043 : // The tenant is shutting down
1044 : #[error("tenant shutting down")]
1045 : TenantCancelled,
1046 :
1047 : // The tenant is shutting down
1048 : #[error("timeline shutting down")]
1049 : TimelineCancelled,
1050 :
1051 : // The tenant is in a state inelegible to run GC
1052 : #[error("not active")]
1053 : NotActive,
1054 :
1055 : // A requested GC cutoff LSN was invalid, for example it tried to move backwards
1056 : #[error("not active")]
1057 : BadLsn { why: String },
1058 :
1059 : // A remote storage error while scheduling updates after compaction
1060 : #[error(transparent)]
1061 : Remote(anyhow::Error),
1062 :
1063 : // An error reading while calculating GC cutoffs
1064 : #[error(transparent)]
1065 : GcCutoffs(PageReconstructError),
1066 :
1067 : // If GC was invoked for a particular timeline, this error means it didn't exist
1068 : #[error("timeline not found")]
1069 : TimelineNotFound,
1070 : }
1071 :
1072 : impl From<PageReconstructError> for GcError {
1073 0 : fn from(value: PageReconstructError) -> Self {
1074 0 : match value {
1075 0 : PageReconstructError::Cancelled => Self::TimelineCancelled,
1076 0 : other => Self::GcCutoffs(other),
1077 : }
1078 0 : }
1079 : }
1080 :
1081 : impl From<NotInitialized> for GcError {
1082 0 : fn from(value: NotInitialized) -> Self {
1083 0 : match value {
1084 0 : NotInitialized::Uninitialized => GcError::Remote(value.into()),
1085 0 : NotInitialized::Stopped | NotInitialized::ShuttingDown => GcError::TimelineCancelled,
1086 : }
1087 0 : }
1088 : }
1089 :
1090 : impl From<timeline::layer_manager::Shutdown> for GcError {
1091 0 : fn from(_: timeline::layer_manager::Shutdown) -> Self {
1092 0 : GcError::TimelineCancelled
1093 0 : }
1094 : }
1095 :
1096 : #[derive(thiserror::Error, Debug)]
1097 : pub(crate) enum LoadConfigError {
1098 : #[error("TOML deserialization error: '{0}'")]
1099 : DeserializeToml(#[from] toml_edit::de::Error),
1100 :
1101 : #[error("Config not found at {0}")]
1102 : NotFound(Utf8PathBuf),
1103 : }
1104 :
1105 : impl Tenant {
1106 : /// Yet another helper for timeline initialization.
1107 : ///
1108 : /// - Initializes the Timeline struct and inserts it into the tenant's hash map
1109 : /// - Scans the local timeline directory for layer files and builds the layer map
1110 : /// - Downloads remote index file and adds remote files to the layer map
1111 : /// - Schedules remote upload tasks for any files that are present locally but missing from remote storage.
1112 : ///
1113 : /// If the operation fails, the timeline is left in the tenant's hash map in Broken state. On success,
1114 : /// it is marked as Active.
1115 : #[allow(clippy::too_many_arguments)]
1116 12 : async fn timeline_init_and_sync(
1117 12 : self: &Arc<Self>,
1118 12 : timeline_id: TimelineId,
1119 12 : resources: TimelineResources,
1120 12 : mut index_part: IndexPart,
1121 12 : metadata: TimelineMetadata,
1122 12 : ancestor: Option<Arc<Timeline>>,
1123 12 : cause: LoadTimelineCause,
1124 12 : ctx: &RequestContext,
1125 12 : ) -> anyhow::Result<TimelineInitAndSyncResult> {
1126 12 : let tenant_id = self.tenant_shard_id;
1127 12 :
1128 12 : let import_pgdata = index_part.import_pgdata.take();
1129 12 : let idempotency = match &import_pgdata {
1130 0 : Some(import_pgdata) => {
1131 0 : CreateTimelineIdempotency::ImportPgdata(CreatingTimelineIdempotencyImportPgdata {
1132 0 : idempotency_key: import_pgdata.idempotency_key().clone(),
1133 0 : })
1134 : }
1135 : None => {
1136 12 : if metadata.ancestor_timeline().is_none() {
1137 8 : CreateTimelineIdempotency::Bootstrap {
1138 8 : pg_version: metadata.pg_version(),
1139 8 : }
1140 : } else {
1141 4 : CreateTimelineIdempotency::Branch {
1142 4 : ancestor_timeline_id: metadata.ancestor_timeline().unwrap(),
1143 4 : ancestor_start_lsn: metadata.ancestor_lsn(),
1144 4 : }
1145 : }
1146 : }
1147 : };
1148 :
1149 12 : let timeline = self.create_timeline_struct(
1150 12 : timeline_id,
1151 12 : &metadata,
1152 12 : ancestor.clone(),
1153 12 : resources,
1154 12 : CreateTimelineCause::Load,
1155 12 : idempotency.clone(),
1156 12 : )?;
1157 12 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
1158 12 : anyhow::ensure!(
1159 12 : disk_consistent_lsn.is_valid(),
1160 0 : "Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn"
1161 : );
1162 12 : assert_eq!(
1163 12 : disk_consistent_lsn,
1164 12 : metadata.disk_consistent_lsn(),
1165 0 : "these are used interchangeably"
1166 : );
1167 :
1168 12 : timeline.remote_client.init_upload_queue(&index_part)?;
1169 :
1170 12 : timeline
1171 12 : .load_layer_map(disk_consistent_lsn, index_part)
1172 12 : .await
1173 12 : .with_context(|| {
1174 0 : format!("Failed to load layermap for timeline {tenant_id}/{timeline_id}")
1175 12 : })?;
1176 :
1177 0 : match import_pgdata {
1178 0 : Some(import_pgdata) if !import_pgdata.is_done() => {
1179 0 : match cause {
1180 0 : LoadTimelineCause::Attach | LoadTimelineCause::Unoffload => (),
1181 : LoadTimelineCause::ImportPgdata { .. } => {
1182 0 : unreachable!("ImportPgdata should not be reloading timeline import is done and persisted as such in s3")
1183 : }
1184 : }
1185 0 : let mut guard = self.timelines_creating.lock().unwrap();
1186 0 : if !guard.insert(timeline_id) {
1187 : // We should never try and load the same timeline twice during startup
1188 0 : unreachable!("Timeline {tenant_id}/{timeline_id} is already being created")
1189 0 : }
1190 0 : let timeline_create_guard = TimelineCreateGuard {
1191 0 : _tenant_gate_guard: self.gate.enter()?,
1192 0 : owning_tenant: self.clone(),
1193 0 : timeline_id,
1194 0 : idempotency,
1195 0 : // The users of this specific return value don't need the timline_path in there.
1196 0 : timeline_path: timeline
1197 0 : .conf
1198 0 : .timeline_path(&timeline.tenant_shard_id, &timeline.timeline_id),
1199 0 : };
1200 0 : Ok(TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
1201 0 : TimelineInitAndSyncNeedsSpawnImportPgdata {
1202 0 : timeline,
1203 0 : import_pgdata,
1204 0 : guard: timeline_create_guard,
1205 0 : },
1206 0 : ))
1207 : }
1208 : Some(_) | None => {
1209 : {
1210 12 : let mut timelines_accessor = self.timelines.lock().unwrap();
1211 12 : match timelines_accessor.entry(timeline_id) {
1212 : // We should never try and load the same timeline twice during startup
1213 : Entry::Occupied(_) => {
1214 0 : unreachable!(
1215 0 : "Timeline {tenant_id}/{timeline_id} already exists in the tenant map"
1216 0 : );
1217 : }
1218 12 : Entry::Vacant(v) => {
1219 12 : v.insert(Arc::clone(&timeline));
1220 12 : timeline.maybe_spawn_flush_loop();
1221 12 : }
1222 : }
1223 : }
1224 :
1225 : // Sanity check: a timeline should have some content.
1226 12 : anyhow::ensure!(
1227 12 : ancestor.is_some()
1228 8 : || timeline
1229 8 : .layers
1230 8 : .read()
1231 8 : .await
1232 8 : .layer_map()
1233 8 : .expect("currently loading, layer manager cannot be shutdown already")
1234 8 : .iter_historic_layers()
1235 8 : .next()
1236 8 : .is_some(),
1237 0 : "Timeline has no ancestor and no layer files"
1238 : );
1239 :
1240 12 : match cause {
1241 12 : LoadTimelineCause::Attach | LoadTimelineCause::Unoffload => (),
1242 : LoadTimelineCause::ImportPgdata {
1243 0 : create_guard,
1244 0 : activate,
1245 0 : } => {
1246 0 : // TODO: see the comment in the task code above how I'm not so certain
1247 0 : // it is safe to activate here because of concurrent shutdowns.
1248 0 : match activate {
1249 0 : ActivateTimelineArgs::Yes { broker_client } => {
1250 0 : info!("activating timeline after reload from pgdata import task");
1251 0 : timeline.activate(self.clone(), broker_client, None, ctx);
1252 : }
1253 0 : ActivateTimelineArgs::No => (),
1254 : }
1255 0 : drop(create_guard);
1256 : }
1257 : }
1258 :
1259 12 : Ok(TimelineInitAndSyncResult::ReadyToActivate(timeline))
1260 : }
1261 : }
1262 12 : }
1263 :
1264 : /// Attach a tenant that's available in cloud storage.
1265 : ///
1266 : /// This returns quickly, after just creating the in-memory object
1267 : /// Tenant struct and launching a background task to download
1268 : /// the remote index files. On return, the tenant is most likely still in
1269 : /// Attaching state, and it will become Active once the background task
1270 : /// finishes. You can use wait_until_active() to wait for the task to
1271 : /// complete.
1272 : ///
1273 : #[allow(clippy::too_many_arguments)]
1274 0 : pub(crate) fn spawn(
1275 0 : conf: &'static PageServerConf,
1276 0 : tenant_shard_id: TenantShardId,
1277 0 : resources: TenantSharedResources,
1278 0 : attached_conf: AttachedTenantConf,
1279 0 : shard_identity: ShardIdentity,
1280 0 : init_order: Option<InitializationOrder>,
1281 0 : mode: SpawnMode,
1282 0 : ctx: &RequestContext,
1283 0 : ) -> Result<Arc<Tenant>, GlobalShutDown> {
1284 0 : let wal_redo_manager =
1285 0 : WalRedoManager::new(PostgresRedoManager::new(conf, tenant_shard_id))?;
1286 :
1287 : let TenantSharedResources {
1288 0 : broker_client,
1289 0 : remote_storage,
1290 0 : deletion_queue_client,
1291 0 : l0_flush_global_state,
1292 0 : } = resources;
1293 0 :
1294 0 : let attach_mode = attached_conf.location.attach_mode;
1295 0 : let generation = attached_conf.location.generation;
1296 0 :
1297 0 : let tenant = Arc::new(Tenant::new(
1298 0 : TenantState::Attaching,
1299 0 : conf,
1300 0 : attached_conf,
1301 0 : shard_identity,
1302 0 : Some(wal_redo_manager),
1303 0 : tenant_shard_id,
1304 0 : remote_storage.clone(),
1305 0 : deletion_queue_client,
1306 0 : l0_flush_global_state,
1307 0 : ));
1308 0 :
1309 0 : // The attach task will carry a GateGuard, so that shutdown() reliably waits for it to drop out if
1310 0 : // we shut down while attaching.
1311 0 : let attach_gate_guard = tenant
1312 0 : .gate
1313 0 : .enter()
1314 0 : .expect("We just created the Tenant: nothing else can have shut it down yet");
1315 0 :
1316 0 : // Do all the hard work in the background
1317 0 : let tenant_clone = Arc::clone(&tenant);
1318 0 : let ctx = ctx.detached_child(TaskKind::Attach, DownloadBehavior::Warn);
1319 0 : task_mgr::spawn(
1320 0 : &tokio::runtime::Handle::current(),
1321 0 : TaskKind::Attach,
1322 0 : tenant_shard_id,
1323 0 : None,
1324 0 : "attach tenant",
1325 0 : async move {
1326 0 :
1327 0 : info!(
1328 : ?attach_mode,
1329 0 : "Attaching tenant"
1330 : );
1331 :
1332 0 : let _gate_guard = attach_gate_guard;
1333 0 :
1334 0 : // Is this tenant being spawned as part of process startup?
1335 0 : let starting_up = init_order.is_some();
1336 0 : scopeguard::defer! {
1337 0 : if starting_up {
1338 0 : TENANT.startup_complete.inc();
1339 0 : }
1340 0 : }
1341 :
1342 : // Ideally we should use Tenant::set_broken_no_wait, but it is not supposed to be used when tenant is in loading state.
1343 : enum BrokenVerbosity {
1344 : Error,
1345 : Info
1346 : }
1347 0 : let make_broken =
1348 0 : |t: &Tenant, err: anyhow::Error, verbosity: BrokenVerbosity| {
1349 0 : match verbosity {
1350 : BrokenVerbosity::Info => {
1351 0 : info!("attach cancelled, setting tenant state to Broken: {err}");
1352 : },
1353 : BrokenVerbosity::Error => {
1354 0 : error!("attach failed, setting tenant state to Broken: {err:?}");
1355 : }
1356 : }
1357 0 : t.state.send_modify(|state| {
1358 0 : // The Stopping case is for when we have passed control on to DeleteTenantFlow:
1359 0 : // if it errors, we will call make_broken when tenant is already in Stopping.
1360 0 : assert!(
1361 0 : matches!(*state, TenantState::Attaching | TenantState::Stopping { .. }),
1362 0 : "the attach task owns the tenant state until activation is complete"
1363 : );
1364 :
1365 0 : *state = TenantState::broken_from_reason(err.to_string());
1366 0 : });
1367 0 : };
1368 :
1369 : // TODO: should also be rejecting tenant conf changes that violate this check.
1370 0 : if let Err(e) = crate::tenant::storage_layer::inmemory_layer::IndexEntry::validate_checkpoint_distance(tenant_clone.get_checkpoint_distance()) {
1371 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1372 0 : return Ok(());
1373 0 : }
1374 0 :
1375 0 : let mut init_order = init_order;
1376 0 : // take the completion because initial tenant loading will complete when all of
1377 0 : // these tasks complete.
1378 0 : let _completion = init_order
1379 0 : .as_mut()
1380 0 : .and_then(|x| x.initial_tenant_load.take());
1381 0 : let remote_load_completion = init_order
1382 0 : .as_mut()
1383 0 : .and_then(|x| x.initial_tenant_load_remote.take());
1384 :
1385 : enum AttachType<'a> {
1386 : /// We are attaching this tenant lazily in the background.
1387 : Warmup {
1388 : _permit: tokio::sync::SemaphorePermit<'a>,
1389 : during_startup: bool
1390 : },
1391 : /// We are attaching this tenant as soon as we can, because for example an
1392 : /// endpoint tried to access it.
1393 : OnDemand,
1394 : /// During normal operations after startup, we are attaching a tenant, and
1395 : /// eager attach was requested.
1396 : Normal,
1397 : }
1398 :
1399 0 : let attach_type = if matches!(mode, SpawnMode::Lazy) {
1400 : // Before doing any I/O, wait for at least one of:
1401 : // - A client attempting to access to this tenant (on-demand loading)
1402 : // - A permit becoming available in the warmup semaphore (background warmup)
1403 :
1404 0 : tokio::select!(
1405 0 : permit = tenant_clone.activate_now_sem.acquire() => {
1406 0 : let _ = permit.expect("activate_now_sem is never closed");
1407 0 : tracing::info!("Activating tenant (on-demand)");
1408 0 : AttachType::OnDemand
1409 : },
1410 0 : permit = conf.concurrent_tenant_warmup.inner().acquire() => {
1411 0 : let _permit = permit.expect("concurrent_tenant_warmup semaphore is never closed");
1412 0 : tracing::info!("Activating tenant (warmup)");
1413 0 : AttachType::Warmup {
1414 0 : _permit,
1415 0 : during_startup: init_order.is_some()
1416 0 : }
1417 : }
1418 0 : _ = tenant_clone.cancel.cancelled() => {
1419 : // This is safe, but should be pretty rare: it is interesting if a tenant
1420 : // stayed in Activating for such a long time that shutdown found it in
1421 : // that state.
1422 0 : tracing::info!(state=%tenant_clone.current_state(), "Tenant shut down before activation");
1423 : // Make the tenant broken so that set_stopping will not hang waiting for it to leave
1424 : // the Attaching state. This is an over-reaction (nothing really broke, the tenant is
1425 : // just shutting down), but ensures progress.
1426 0 : make_broken(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"), BrokenVerbosity::Info);
1427 0 : return Ok(());
1428 : },
1429 : )
1430 : } else {
1431 : // SpawnMode::{Create,Eager} always cause jumping ahead of the
1432 : // concurrent_tenant_warmup queue
1433 0 : AttachType::Normal
1434 : };
1435 :
1436 0 : let preload = match &mode {
1437 : SpawnMode::Eager | SpawnMode::Lazy => {
1438 0 : let _preload_timer = TENANT.preload.start_timer();
1439 0 : let res = tenant_clone
1440 0 : .preload(&remote_storage, task_mgr::shutdown_token())
1441 0 : .await;
1442 0 : match res {
1443 0 : Ok(p) => Some(p),
1444 0 : Err(e) => {
1445 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1446 0 : return Ok(());
1447 : }
1448 : }
1449 : }
1450 :
1451 : };
1452 :
1453 : // Remote preload is complete.
1454 0 : drop(remote_load_completion);
1455 0 :
1456 0 :
1457 0 : // We will time the duration of the attach phase unless this is a creation (attach will do no work)
1458 0 : let attach_start = std::time::Instant::now();
1459 0 : let attached = {
1460 0 : let _attach_timer = Some(TENANT.attach.start_timer());
1461 0 : tenant_clone.attach(preload, &ctx).await
1462 : };
1463 0 : let attach_duration = attach_start.elapsed();
1464 0 : _ = tenant_clone.attach_wal_lag_cooldown.set(WalLagCooldown::new(attach_start, attach_duration));
1465 0 :
1466 0 : match attached {
1467 : Ok(()) => {
1468 0 : info!("attach finished, activating");
1469 0 : tenant_clone.activate(broker_client, None, &ctx);
1470 : }
1471 0 : Err(e) => {
1472 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1473 0 : }
1474 : }
1475 :
1476 : // If we are doing an opportunistic warmup attachment at startup, initialize
1477 : // logical size at the same time. This is better than starting a bunch of idle tenants
1478 : // with cold caches and then coming back later to initialize their logical sizes.
1479 : //
1480 : // It also prevents the warmup proccess competing with the concurrency limit on
1481 : // logical size calculations: if logical size calculation semaphore is saturated,
1482 : // then warmup will wait for that before proceeding to the next tenant.
1483 0 : if matches!(attach_type, AttachType::Warmup { during_startup: true, .. }) {
1484 0 : let mut futs: FuturesUnordered<_> = tenant_clone.timelines.lock().unwrap().values().cloned().map(|t| t.await_initial_logical_size()).collect();
1485 0 : tracing::info!("Waiting for initial logical sizes while warming up...");
1486 0 : while futs.next().await.is_some() {}
1487 0 : tracing::info!("Warm-up complete");
1488 0 : }
1489 :
1490 0 : Ok(())
1491 0 : }
1492 0 : .instrument(tracing::info_span!(parent: None, "attach", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), gen=?generation)),
1493 : );
1494 0 : Ok(tenant)
1495 0 : }
1496 :
1497 : #[instrument(skip_all)]
1498 : pub(crate) async fn preload(
1499 : self: &Arc<Self>,
1500 : remote_storage: &GenericRemoteStorage,
1501 : cancel: CancellationToken,
1502 : ) -> anyhow::Result<TenantPreload> {
1503 : span::debug_assert_current_span_has_tenant_id();
1504 : // Get list of remote timelines
1505 : // download index files for every tenant timeline
1506 : info!("listing remote timelines");
1507 : let (mut remote_timeline_ids, other_keys) = remote_timeline_client::list_remote_timelines(
1508 : remote_storage,
1509 : self.tenant_shard_id,
1510 : cancel.clone(),
1511 : )
1512 : .await?;
1513 : let (offloaded_add, tenant_manifest) =
1514 : match remote_timeline_client::download_tenant_manifest(
1515 : remote_storage,
1516 : &self.tenant_shard_id,
1517 : self.generation,
1518 : &cancel,
1519 : )
1520 : .await
1521 : {
1522 : Ok((tenant_manifest, _generation, _manifest_mtime)) => (
1523 : format!("{} offloaded", tenant_manifest.offloaded_timelines.len()),
1524 : tenant_manifest,
1525 : ),
1526 : Err(DownloadError::NotFound) => {
1527 : ("no manifest".to_string(), TenantManifest::empty())
1528 : }
1529 : Err(e) => Err(e)?,
1530 : };
1531 :
1532 : info!(
1533 : "found {} timelines, and {offloaded_add}",
1534 : remote_timeline_ids.len()
1535 : );
1536 :
1537 : for k in other_keys {
1538 : warn!("Unexpected non timeline key {k}");
1539 : }
1540 :
1541 : // Avoid downloading IndexPart of offloaded timelines.
1542 : let mut offloaded_with_prefix = HashSet::new();
1543 : for offloaded in tenant_manifest.offloaded_timelines.iter() {
1544 : if remote_timeline_ids.remove(&offloaded.timeline_id) {
1545 : offloaded_with_prefix.insert(offloaded.timeline_id);
1546 : } else {
1547 : // We'll take care later of timelines in the manifest without a prefix
1548 : }
1549 : }
1550 :
1551 : let timelines = self
1552 : .load_timelines_metadata(remote_timeline_ids, remote_storage, cancel)
1553 : .await?;
1554 :
1555 : Ok(TenantPreload {
1556 : tenant_manifest,
1557 : timelines: timelines
1558 : .into_iter()
1559 12 : .map(|(id, tl)| (id, Some(tl)))
1560 0 : .chain(offloaded_with_prefix.into_iter().map(|id| (id, None)))
1561 : .collect(),
1562 : })
1563 : }
1564 :
1565 : ///
1566 : /// Background task that downloads all data for a tenant and brings it to Active state.
1567 : ///
1568 : /// No background tasks are started as part of this routine.
1569 : ///
1570 440 : async fn attach(
1571 440 : self: &Arc<Tenant>,
1572 440 : preload: Option<TenantPreload>,
1573 440 : ctx: &RequestContext,
1574 440 : ) -> anyhow::Result<()> {
1575 440 : span::debug_assert_current_span_has_tenant_id();
1576 440 :
1577 440 : failpoint_support::sleep_millis_async!("before-attaching-tenant");
1578 :
1579 440 : let Some(preload) = preload else {
1580 0 : anyhow::bail!("local-only deployment is no longer supported, https://github.com/neondatabase/neon/issues/5624");
1581 : };
1582 :
1583 440 : let mut offloaded_timeline_ids = HashSet::new();
1584 440 : let mut offloaded_timelines_list = Vec::new();
1585 440 : for timeline_manifest in preload.tenant_manifest.offloaded_timelines.iter() {
1586 0 : let timeline_id = timeline_manifest.timeline_id;
1587 0 : let offloaded_timeline =
1588 0 : OffloadedTimeline::from_manifest(self.tenant_shard_id, timeline_manifest);
1589 0 : offloaded_timelines_list.push((timeline_id, Arc::new(offloaded_timeline)));
1590 0 : offloaded_timeline_ids.insert(timeline_id);
1591 0 : }
1592 : // Complete deletions for offloaded timeline id's from manifest.
1593 : // The manifest will be uploaded later in this function.
1594 440 : offloaded_timelines_list
1595 440 : .retain(|(offloaded_id, offloaded)| {
1596 0 : // Existence of a timeline is finally determined by the existence of an index-part.json in remote storage.
1597 0 : // If there is dangling references in another location, they need to be cleaned up.
1598 0 : let delete = !preload.timelines.contains_key(offloaded_id);
1599 0 : if delete {
1600 0 : tracing::info!("Removing offloaded timeline {offloaded_id} from manifest as no remote prefix was found");
1601 0 : offloaded.defuse_for_tenant_drop();
1602 0 : }
1603 0 : !delete
1604 440 : });
1605 440 :
1606 440 : let mut timelines_to_resume_deletions = vec![];
1607 440 :
1608 440 : let mut remote_index_and_client = HashMap::new();
1609 440 : let mut timeline_ancestors = HashMap::new();
1610 440 : let mut existent_timelines = HashSet::new();
1611 452 : for (timeline_id, preload) in preload.timelines {
1612 12 : let Some(preload) = preload else { continue };
1613 : // This is an invariant of the `preload` function's API
1614 12 : assert!(!offloaded_timeline_ids.contains(&timeline_id));
1615 12 : let index_part = match preload.index_part {
1616 12 : Ok(i) => {
1617 12 : debug!("remote index part exists for timeline {timeline_id}");
1618 : // We found index_part on the remote, this is the standard case.
1619 12 : existent_timelines.insert(timeline_id);
1620 12 : i
1621 : }
1622 : Err(DownloadError::NotFound) => {
1623 : // There is no index_part on the remote. We only get here
1624 : // if there is some prefix for the timeline in the remote storage.
1625 : // This can e.g. be the initdb.tar.zst archive, maybe a
1626 : // remnant from a prior incomplete creation or deletion attempt.
1627 : // Delete the local directory as the deciding criterion for a
1628 : // timeline's existence is presence of index_part.
1629 0 : info!(%timeline_id, "index_part not found on remote");
1630 0 : continue;
1631 : }
1632 0 : Err(DownloadError::Fatal(why)) => {
1633 0 : // If, while loading one remote timeline, we saw an indication that our generation
1634 0 : // number is likely invalid, then we should not load the whole tenant.
1635 0 : error!(%timeline_id, "Fatal error loading timeline: {why}");
1636 0 : anyhow::bail!(why.to_string());
1637 : }
1638 0 : Err(e) => {
1639 0 : // Some (possibly ephemeral) error happened during index_part download.
1640 0 : // Pretend the timeline exists to not delete the timeline directory,
1641 0 : // as it might be a temporary issue and we don't want to re-download
1642 0 : // everything after it resolves.
1643 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1644 :
1645 0 : existent_timelines.insert(timeline_id);
1646 0 : continue;
1647 : }
1648 : };
1649 12 : match index_part {
1650 12 : MaybeDeletedIndexPart::IndexPart(index_part) => {
1651 12 : timeline_ancestors.insert(timeline_id, index_part.metadata.clone());
1652 12 : remote_index_and_client.insert(timeline_id, (index_part, preload.client));
1653 12 : }
1654 0 : MaybeDeletedIndexPart::Deleted(index_part) => {
1655 0 : info!(
1656 0 : "timeline {} is deleted, picking to resume deletion",
1657 : timeline_id
1658 : );
1659 0 : timelines_to_resume_deletions.push((timeline_id, index_part, preload.client));
1660 : }
1661 : }
1662 : }
1663 :
1664 440 : let mut gc_blocks = HashMap::new();
1665 :
1666 : // For every timeline, download the metadata file, scan the local directory,
1667 : // and build a layer map that contains an entry for each remote and local
1668 : // layer file.
1669 440 : let sorted_timelines = tree_sort_timelines(timeline_ancestors, |m| m.ancestor_timeline())?;
1670 452 : for (timeline_id, remote_metadata) in sorted_timelines {
1671 12 : let (index_part, remote_client) = remote_index_and_client
1672 12 : .remove(&timeline_id)
1673 12 : .expect("just put it in above");
1674 :
1675 12 : if let Some(blocking) = index_part.gc_blocking.as_ref() {
1676 : // could just filter these away, but it helps while testing
1677 0 : anyhow::ensure!(
1678 0 : !blocking.reasons.is_empty(),
1679 0 : "index_part for {timeline_id} is malformed: it should not have gc blocking with zero reasons"
1680 : );
1681 0 : let prev = gc_blocks.insert(timeline_id, blocking.reasons);
1682 0 : assert!(prev.is_none());
1683 12 : }
1684 :
1685 : // TODO again handle early failure
1686 12 : let effect = self
1687 12 : .load_remote_timeline(
1688 12 : timeline_id,
1689 12 : index_part,
1690 12 : remote_metadata,
1691 12 : TimelineResources {
1692 12 : remote_client,
1693 12 : pagestream_throttle: self.pagestream_throttle.clone(),
1694 12 : pagestream_throttle_metrics: self.pagestream_throttle_metrics.clone(),
1695 12 : l0_flush_global_state: self.l0_flush_global_state.clone(),
1696 12 : },
1697 12 : LoadTimelineCause::Attach,
1698 12 : ctx,
1699 12 : )
1700 12 : .await
1701 12 : .with_context(|| {
1702 0 : format!(
1703 0 : "failed to load remote timeline {} for tenant {}",
1704 0 : timeline_id, self.tenant_shard_id
1705 0 : )
1706 12 : })?;
1707 :
1708 12 : match effect {
1709 12 : TimelineInitAndSyncResult::ReadyToActivate(_) => {
1710 12 : // activation happens later, on Tenant::activate
1711 12 : }
1712 : TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
1713 : TimelineInitAndSyncNeedsSpawnImportPgdata {
1714 0 : timeline,
1715 0 : import_pgdata,
1716 0 : guard,
1717 0 : },
1718 0 : ) => {
1719 0 : tokio::task::spawn(self.clone().create_timeline_import_pgdata_task(
1720 0 : timeline,
1721 0 : import_pgdata,
1722 0 : ActivateTimelineArgs::No,
1723 0 : guard,
1724 0 : ));
1725 0 : }
1726 : }
1727 : }
1728 :
1729 : // Walk through deleted timelines, resume deletion
1730 440 : for (timeline_id, index_part, remote_timeline_client) in timelines_to_resume_deletions {
1731 0 : remote_timeline_client
1732 0 : .init_upload_queue_stopped_to_continue_deletion(&index_part)
1733 0 : .context("init queue stopped")
1734 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1735 :
1736 0 : DeleteTimelineFlow::resume_deletion(
1737 0 : Arc::clone(self),
1738 0 : timeline_id,
1739 0 : &index_part.metadata,
1740 0 : remote_timeline_client,
1741 0 : )
1742 0 : .instrument(tracing::info_span!("timeline_delete", %timeline_id))
1743 0 : .await
1744 0 : .context("resume_deletion")
1745 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1746 : }
1747 440 : let needs_manifest_upload =
1748 440 : offloaded_timelines_list.len() != preload.tenant_manifest.offloaded_timelines.len();
1749 440 : {
1750 440 : let mut offloaded_timelines_accessor = self.timelines_offloaded.lock().unwrap();
1751 440 : offloaded_timelines_accessor.extend(offloaded_timelines_list.into_iter());
1752 440 : }
1753 440 : if needs_manifest_upload {
1754 0 : self.store_tenant_manifest().await?;
1755 440 : }
1756 :
1757 : // The local filesystem contents are a cache of what's in the remote IndexPart;
1758 : // IndexPart is the source of truth.
1759 440 : self.clean_up_timelines(&existent_timelines)?;
1760 :
1761 440 : self.gc_block.set_scanned(gc_blocks);
1762 440 :
1763 440 : fail::fail_point!("attach-before-activate", |_| {
1764 0 : anyhow::bail!("attach-before-activate");
1765 440 : });
1766 440 : failpoint_support::sleep_millis_async!("attach-before-activate-sleep", &self.cancel);
1767 :
1768 440 : info!("Done");
1769 :
1770 440 : Ok(())
1771 440 : }
1772 :
1773 : /// Check for any local timeline directories that are temporary, or do not correspond to a
1774 : /// timeline that still exists: this can happen if we crashed during a deletion/creation, or
1775 : /// if a timeline was deleted while the tenant was attached to a different pageserver.
1776 440 : fn clean_up_timelines(&self, existent_timelines: &HashSet<TimelineId>) -> anyhow::Result<()> {
1777 440 : let timelines_dir = self.conf.timelines_path(&self.tenant_shard_id);
1778 :
1779 440 : let entries = match timelines_dir.read_dir_utf8() {
1780 440 : Ok(d) => d,
1781 0 : Err(e) => {
1782 0 : if e.kind() == std::io::ErrorKind::NotFound {
1783 0 : return Ok(());
1784 : } else {
1785 0 : return Err(e).context("list timelines directory for tenant");
1786 : }
1787 : }
1788 : };
1789 :
1790 456 : for entry in entries {
1791 16 : let entry = entry.context("read timeline dir entry")?;
1792 16 : let entry_path = entry.path();
1793 :
1794 16 : let purge = if crate::is_temporary(entry_path)
1795 : // TODO: remove uninit mark code (https://github.com/neondatabase/neon/issues/5718)
1796 16 : || is_uninit_mark(entry_path)
1797 16 : || crate::is_delete_mark(entry_path)
1798 : {
1799 0 : true
1800 : } else {
1801 16 : match TimelineId::try_from(entry_path.file_name()) {
1802 16 : Ok(i) => {
1803 16 : // Purge if the timeline ID does not exist in remote storage: remote storage is the authority.
1804 16 : !existent_timelines.contains(&i)
1805 : }
1806 0 : Err(e) => {
1807 0 : tracing::warn!(
1808 0 : "Unparseable directory in timelines directory: {entry_path}, ignoring ({e})"
1809 : );
1810 : // Do not purge junk: if we don't recognize it, be cautious and leave it for a human.
1811 0 : false
1812 : }
1813 : }
1814 : };
1815 :
1816 16 : if purge {
1817 4 : tracing::info!("Purging stale timeline dentry {entry_path}");
1818 4 : if let Err(e) = match entry.file_type() {
1819 4 : Ok(t) => if t.is_dir() {
1820 4 : std::fs::remove_dir_all(entry_path)
1821 : } else {
1822 0 : std::fs::remove_file(entry_path)
1823 : }
1824 4 : .or_else(fs_ext::ignore_not_found),
1825 0 : Err(e) => Err(e),
1826 : } {
1827 0 : tracing::warn!("Failed to purge stale timeline dentry {entry_path}: {e}");
1828 4 : }
1829 12 : }
1830 : }
1831 :
1832 440 : Ok(())
1833 440 : }
1834 :
1835 : /// Get sum of all remote timelines sizes
1836 : ///
1837 : /// This function relies on the index_part instead of listing the remote storage
1838 0 : pub fn remote_size(&self) -> u64 {
1839 0 : let mut size = 0;
1840 :
1841 0 : for timeline in self.list_timelines() {
1842 0 : size += timeline.remote_client.get_remote_physical_size();
1843 0 : }
1844 :
1845 0 : size
1846 0 : }
1847 :
1848 : #[instrument(skip_all, fields(timeline_id=%timeline_id))]
1849 : async fn load_remote_timeline(
1850 : self: &Arc<Self>,
1851 : timeline_id: TimelineId,
1852 : index_part: IndexPart,
1853 : remote_metadata: TimelineMetadata,
1854 : resources: TimelineResources,
1855 : cause: LoadTimelineCause,
1856 : ctx: &RequestContext,
1857 : ) -> anyhow::Result<TimelineInitAndSyncResult> {
1858 : span::debug_assert_current_span_has_tenant_id();
1859 :
1860 : info!("downloading index file for timeline {}", timeline_id);
1861 : tokio::fs::create_dir_all(self.conf.timeline_path(&self.tenant_shard_id, &timeline_id))
1862 : .await
1863 : .context("Failed to create new timeline directory")?;
1864 :
1865 : let ancestor = if let Some(ancestor_id) = remote_metadata.ancestor_timeline() {
1866 : let timelines = self.timelines.lock().unwrap();
1867 : Some(Arc::clone(timelines.get(&ancestor_id).ok_or_else(
1868 0 : || {
1869 0 : anyhow::anyhow!(
1870 0 : "cannot find ancestor timeline {ancestor_id} for timeline {timeline_id}"
1871 0 : )
1872 0 : },
1873 : )?))
1874 : } else {
1875 : None
1876 : };
1877 :
1878 : self.timeline_init_and_sync(
1879 : timeline_id,
1880 : resources,
1881 : index_part,
1882 : remote_metadata,
1883 : ancestor,
1884 : cause,
1885 : ctx,
1886 : )
1887 : .await
1888 : }
1889 :
1890 440 : async fn load_timelines_metadata(
1891 440 : self: &Arc<Tenant>,
1892 440 : timeline_ids: HashSet<TimelineId>,
1893 440 : remote_storage: &GenericRemoteStorage,
1894 440 : cancel: CancellationToken,
1895 440 : ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
1896 440 : let mut part_downloads = JoinSet::new();
1897 452 : for timeline_id in timeline_ids {
1898 12 : let cancel_clone = cancel.clone();
1899 12 : part_downloads.spawn(
1900 12 : self.load_timeline_metadata(timeline_id, remote_storage.clone(), cancel_clone)
1901 12 : .instrument(info_span!("download_index_part", %timeline_id)),
1902 : );
1903 : }
1904 :
1905 440 : let mut timeline_preloads: HashMap<TimelineId, TimelinePreload> = HashMap::new();
1906 :
1907 : loop {
1908 452 : tokio::select!(
1909 452 : next = part_downloads.join_next() => {
1910 452 : match next {
1911 12 : Some(result) => {
1912 12 : let preload = result.context("join preload task")?;
1913 12 : timeline_preloads.insert(preload.timeline_id, preload);
1914 : },
1915 : None => {
1916 440 : break;
1917 : }
1918 : }
1919 : },
1920 452 : _ = cancel.cancelled() => {
1921 0 : anyhow::bail!("Cancelled while waiting for remote index download")
1922 : }
1923 : )
1924 : }
1925 :
1926 440 : Ok(timeline_preloads)
1927 440 : }
1928 :
1929 12 : fn build_timeline_client(
1930 12 : &self,
1931 12 : timeline_id: TimelineId,
1932 12 : remote_storage: GenericRemoteStorage,
1933 12 : ) -> RemoteTimelineClient {
1934 12 : RemoteTimelineClient::new(
1935 12 : remote_storage.clone(),
1936 12 : self.deletion_queue_client.clone(),
1937 12 : self.conf,
1938 12 : self.tenant_shard_id,
1939 12 : timeline_id,
1940 12 : self.generation,
1941 12 : &self.tenant_conf.load().location,
1942 12 : )
1943 12 : }
1944 :
1945 12 : fn load_timeline_metadata(
1946 12 : self: &Arc<Tenant>,
1947 12 : timeline_id: TimelineId,
1948 12 : remote_storage: GenericRemoteStorage,
1949 12 : cancel: CancellationToken,
1950 12 : ) -> impl Future<Output = TimelinePreload> {
1951 12 : let client = self.build_timeline_client(timeline_id, remote_storage);
1952 12 : async move {
1953 12 : debug_assert_current_span_has_tenant_and_timeline_id();
1954 12 : debug!("starting index part download");
1955 :
1956 12 : let index_part = client.download_index_file(&cancel).await;
1957 :
1958 12 : debug!("finished index part download");
1959 :
1960 12 : TimelinePreload {
1961 12 : client,
1962 12 : timeline_id,
1963 12 : index_part,
1964 12 : }
1965 12 : }
1966 12 : }
1967 :
1968 0 : fn check_to_be_archived_has_no_unarchived_children(
1969 0 : timeline_id: TimelineId,
1970 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
1971 0 : ) -> Result<(), TimelineArchivalError> {
1972 0 : let children: Vec<TimelineId> = timelines
1973 0 : .iter()
1974 0 : .filter_map(|(id, entry)| {
1975 0 : if entry.get_ancestor_timeline_id() != Some(timeline_id) {
1976 0 : return None;
1977 0 : }
1978 0 : if entry.is_archived() == Some(true) {
1979 0 : return None;
1980 0 : }
1981 0 : Some(*id)
1982 0 : })
1983 0 : .collect();
1984 0 :
1985 0 : if !children.is_empty() {
1986 0 : return Err(TimelineArchivalError::HasUnarchivedChildren(children));
1987 0 : }
1988 0 : Ok(())
1989 0 : }
1990 :
1991 0 : fn check_ancestor_of_to_be_unarchived_is_not_archived(
1992 0 : ancestor_timeline_id: TimelineId,
1993 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
1994 0 : offloaded_timelines: &std::sync::MutexGuard<
1995 0 : '_,
1996 0 : HashMap<TimelineId, Arc<OffloadedTimeline>>,
1997 0 : >,
1998 0 : ) -> Result<(), TimelineArchivalError> {
1999 0 : let has_archived_parent =
2000 0 : if let Some(ancestor_timeline) = timelines.get(&ancestor_timeline_id) {
2001 0 : ancestor_timeline.is_archived() == Some(true)
2002 0 : } else if offloaded_timelines.contains_key(&ancestor_timeline_id) {
2003 0 : true
2004 : } else {
2005 0 : error!("ancestor timeline {ancestor_timeline_id} not found");
2006 0 : if cfg!(debug_assertions) {
2007 0 : panic!("ancestor timeline {ancestor_timeline_id} not found");
2008 0 : }
2009 0 : return Err(TimelineArchivalError::NotFound);
2010 : };
2011 0 : if has_archived_parent {
2012 0 : return Err(TimelineArchivalError::HasArchivedParent(
2013 0 : ancestor_timeline_id,
2014 0 : ));
2015 0 : }
2016 0 : Ok(())
2017 0 : }
2018 :
2019 0 : fn check_to_be_unarchived_timeline_has_no_archived_parent(
2020 0 : timeline: &Arc<Timeline>,
2021 0 : ) -> Result<(), TimelineArchivalError> {
2022 0 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
2023 0 : if ancestor_timeline.is_archived() == Some(true) {
2024 0 : return Err(TimelineArchivalError::HasArchivedParent(
2025 0 : ancestor_timeline.timeline_id,
2026 0 : ));
2027 0 : }
2028 0 : }
2029 0 : Ok(())
2030 0 : }
2031 :
2032 : /// Loads the specified (offloaded) timeline from S3 and attaches it as a loaded timeline
2033 : ///
2034 : /// Counterpart to [`offload_timeline`].
2035 0 : async fn unoffload_timeline(
2036 0 : self: &Arc<Self>,
2037 0 : timeline_id: TimelineId,
2038 0 : broker_client: storage_broker::BrokerClientChannel,
2039 0 : ctx: RequestContext,
2040 0 : ) -> Result<Arc<Timeline>, TimelineArchivalError> {
2041 0 : info!("unoffloading timeline");
2042 :
2043 : // We activate the timeline below manually, so this must be called on an active tenant.
2044 : // We expect callers of this function to ensure this.
2045 0 : match self.current_state() {
2046 : TenantState::Activating { .. }
2047 : | TenantState::Attaching
2048 : | TenantState::Broken { .. } => {
2049 0 : panic!("Timeline expected to be active")
2050 : }
2051 0 : TenantState::Stopping { .. } => return Err(TimelineArchivalError::Cancelled),
2052 0 : TenantState::Active => {}
2053 0 : }
2054 0 : let cancel = self.cancel.clone();
2055 0 :
2056 0 : // Protect against concurrent attempts to use this TimelineId
2057 0 : // We don't care much about idempotency, as it's ensured a layer above.
2058 0 : let allow_offloaded = true;
2059 0 : let _create_guard = self
2060 0 : .create_timeline_create_guard(
2061 0 : timeline_id,
2062 0 : CreateTimelineIdempotency::FailWithConflict,
2063 0 : allow_offloaded,
2064 0 : )
2065 0 : .map_err(|err| match err {
2066 0 : TimelineExclusionError::AlreadyCreating => TimelineArchivalError::AlreadyInProgress,
2067 : TimelineExclusionError::AlreadyExists { .. } => {
2068 0 : TimelineArchivalError::Other(anyhow::anyhow!("Timeline already exists"))
2069 : }
2070 0 : TimelineExclusionError::Other(e) => TimelineArchivalError::Other(e),
2071 0 : TimelineExclusionError::ShuttingDown => TimelineArchivalError::Cancelled,
2072 0 : })?;
2073 :
2074 0 : let timeline_preload = self
2075 0 : .load_timeline_metadata(timeline_id, self.remote_storage.clone(), cancel.clone())
2076 0 : .await;
2077 :
2078 0 : let index_part = match timeline_preload.index_part {
2079 0 : Ok(index_part) => {
2080 0 : debug!("remote index part exists for timeline {timeline_id}");
2081 0 : index_part
2082 : }
2083 : Err(DownloadError::NotFound) => {
2084 0 : error!(%timeline_id, "index_part not found on remote");
2085 0 : return Err(TimelineArchivalError::NotFound);
2086 : }
2087 0 : Err(DownloadError::Cancelled) => return Err(TimelineArchivalError::Cancelled),
2088 0 : Err(e) => {
2089 0 : // Some (possibly ephemeral) error happened during index_part download.
2090 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
2091 0 : return Err(TimelineArchivalError::Other(
2092 0 : anyhow::Error::new(e).context("downloading index_part from remote storage"),
2093 0 : ));
2094 : }
2095 : };
2096 0 : let index_part = match index_part {
2097 0 : MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
2098 0 : MaybeDeletedIndexPart::Deleted(_index_part) => {
2099 0 : info!("timeline is deleted according to index_part.json");
2100 0 : return Err(TimelineArchivalError::NotFound);
2101 : }
2102 : };
2103 0 : let remote_metadata = index_part.metadata.clone();
2104 0 : let timeline_resources = self.build_timeline_resources(timeline_id);
2105 0 : self.load_remote_timeline(
2106 0 : timeline_id,
2107 0 : index_part,
2108 0 : remote_metadata,
2109 0 : timeline_resources,
2110 0 : LoadTimelineCause::Unoffload,
2111 0 : &ctx,
2112 0 : )
2113 0 : .await
2114 0 : .with_context(|| {
2115 0 : format!(
2116 0 : "failed to load remote timeline {} for tenant {}",
2117 0 : timeline_id, self.tenant_shard_id
2118 0 : )
2119 0 : })
2120 0 : .map_err(TimelineArchivalError::Other)?;
2121 :
2122 0 : let timeline = {
2123 0 : let timelines = self.timelines.lock().unwrap();
2124 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2125 0 : warn!("timeline not available directly after attach");
2126 : // This is not a panic because no locks are held between `load_remote_timeline`
2127 : // which puts the timeline into timelines, and our look into the timeline map.
2128 0 : return Err(TimelineArchivalError::Other(anyhow::anyhow!(
2129 0 : "timeline not available directly after attach"
2130 0 : )));
2131 : };
2132 0 : let mut offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2133 0 : match offloaded_timelines.remove(&timeline_id) {
2134 0 : Some(offloaded) => {
2135 0 : offloaded.delete_from_ancestor_with_timelines(&timelines);
2136 0 : }
2137 0 : None => warn!("timeline already removed from offloaded timelines"),
2138 : }
2139 :
2140 0 : self.initialize_gc_info(&timelines, &offloaded_timelines, Some(timeline_id));
2141 0 :
2142 0 : Arc::clone(timeline)
2143 0 : };
2144 0 :
2145 0 : // Upload new list of offloaded timelines to S3
2146 0 : self.store_tenant_manifest().await?;
2147 :
2148 : // Activate the timeline (if it makes sense)
2149 0 : if !(timeline.is_broken() || timeline.is_stopping()) {
2150 0 : let background_jobs_can_start = None;
2151 0 : timeline.activate(
2152 0 : self.clone(),
2153 0 : broker_client.clone(),
2154 0 : background_jobs_can_start,
2155 0 : &ctx,
2156 0 : );
2157 0 : }
2158 :
2159 0 : info!("timeline unoffloading complete");
2160 0 : Ok(timeline)
2161 0 : }
2162 :
2163 0 : pub(crate) async fn apply_timeline_archival_config(
2164 0 : self: &Arc<Self>,
2165 0 : timeline_id: TimelineId,
2166 0 : new_state: TimelineArchivalState,
2167 0 : broker_client: storage_broker::BrokerClientChannel,
2168 0 : ctx: RequestContext,
2169 0 : ) -> Result<(), TimelineArchivalError> {
2170 0 : info!("setting timeline archival config");
2171 : // First part: figure out what is needed to do, and do validation
2172 0 : let timeline_or_unarchive_offloaded = 'outer: {
2173 0 : let timelines = self.timelines.lock().unwrap();
2174 :
2175 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2176 0 : let offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2177 0 : let Some(offloaded) = offloaded_timelines.get(&timeline_id) else {
2178 0 : return Err(TimelineArchivalError::NotFound);
2179 : };
2180 0 : if new_state == TimelineArchivalState::Archived {
2181 : // It's offloaded already, so nothing to do
2182 0 : return Ok(());
2183 0 : }
2184 0 : if let Some(ancestor_timeline_id) = offloaded.ancestor_timeline_id {
2185 0 : Self::check_ancestor_of_to_be_unarchived_is_not_archived(
2186 0 : ancestor_timeline_id,
2187 0 : &timelines,
2188 0 : &offloaded_timelines,
2189 0 : )?;
2190 0 : }
2191 0 : break 'outer None;
2192 : };
2193 :
2194 : // Do some validation. We release the timelines lock below, so there is potential
2195 : // for race conditions: these checks are more present to prevent misunderstandings of
2196 : // the API's capabilities, instead of serving as the sole way to defend their invariants.
2197 0 : match new_state {
2198 : TimelineArchivalState::Unarchived => {
2199 0 : Self::check_to_be_unarchived_timeline_has_no_archived_parent(timeline)?
2200 : }
2201 : TimelineArchivalState::Archived => {
2202 0 : Self::check_to_be_archived_has_no_unarchived_children(timeline_id, &timelines)?
2203 : }
2204 : }
2205 0 : Some(Arc::clone(timeline))
2206 : };
2207 :
2208 : // Second part: unoffload timeline (if needed)
2209 0 : let timeline = if let Some(timeline) = timeline_or_unarchive_offloaded {
2210 0 : timeline
2211 : } else {
2212 : // Turn offloaded timeline into a non-offloaded one
2213 0 : self.unoffload_timeline(timeline_id, broker_client, ctx)
2214 0 : .await?
2215 : };
2216 :
2217 : // Third part: upload new timeline archival state and block until it is present in S3
2218 0 : let upload_needed = match timeline
2219 0 : .remote_client
2220 0 : .schedule_index_upload_for_timeline_archival_state(new_state)
2221 : {
2222 0 : Ok(upload_needed) => upload_needed,
2223 0 : Err(e) => {
2224 0 : if timeline.cancel.is_cancelled() {
2225 0 : return Err(TimelineArchivalError::Cancelled);
2226 : } else {
2227 0 : return Err(TimelineArchivalError::Other(e));
2228 : }
2229 : }
2230 : };
2231 :
2232 0 : if upload_needed {
2233 0 : info!("Uploading new state");
2234 : const MAX_WAIT: Duration = Duration::from_secs(10);
2235 0 : let Ok(v) =
2236 0 : tokio::time::timeout(MAX_WAIT, timeline.remote_client.wait_completion()).await
2237 : else {
2238 0 : tracing::warn!("reached timeout for waiting on upload queue");
2239 0 : return Err(TimelineArchivalError::Timeout);
2240 : };
2241 0 : v.map_err(|e| match e {
2242 0 : WaitCompletionError::NotInitialized(e) => {
2243 0 : TimelineArchivalError::Other(anyhow::anyhow!(e))
2244 : }
2245 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2246 0 : TimelineArchivalError::Cancelled
2247 : }
2248 0 : })?;
2249 0 : }
2250 0 : Ok(())
2251 0 : }
2252 :
2253 4 : pub fn get_offloaded_timeline(
2254 4 : &self,
2255 4 : timeline_id: TimelineId,
2256 4 : ) -> Result<Arc<OffloadedTimeline>, GetTimelineError> {
2257 4 : self.timelines_offloaded
2258 4 : .lock()
2259 4 : .unwrap()
2260 4 : .get(&timeline_id)
2261 4 : .map(Arc::clone)
2262 4 : .ok_or(GetTimelineError::NotFound {
2263 4 : tenant_id: self.tenant_shard_id,
2264 4 : timeline_id,
2265 4 : })
2266 4 : }
2267 :
2268 8 : pub(crate) fn tenant_shard_id(&self) -> TenantShardId {
2269 8 : self.tenant_shard_id
2270 8 : }
2271 :
2272 : /// Get Timeline handle for given Neon timeline ID.
2273 : /// This function is idempotent. It doesn't change internal state in any way.
2274 444 : pub fn get_timeline(
2275 444 : &self,
2276 444 : timeline_id: TimelineId,
2277 444 : active_only: bool,
2278 444 : ) -> Result<Arc<Timeline>, GetTimelineError> {
2279 444 : let timelines_accessor = self.timelines.lock().unwrap();
2280 444 : let timeline = timelines_accessor
2281 444 : .get(&timeline_id)
2282 444 : .ok_or(GetTimelineError::NotFound {
2283 444 : tenant_id: self.tenant_shard_id,
2284 444 : timeline_id,
2285 444 : })?;
2286 :
2287 440 : if active_only && !timeline.is_active() {
2288 0 : Err(GetTimelineError::NotActive {
2289 0 : tenant_id: self.tenant_shard_id,
2290 0 : timeline_id,
2291 0 : state: timeline.current_state(),
2292 0 : })
2293 : } else {
2294 440 : Ok(Arc::clone(timeline))
2295 : }
2296 444 : }
2297 :
2298 : /// Lists timelines the tenant contains.
2299 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2300 0 : pub fn list_timelines(&self) -> Vec<Arc<Timeline>> {
2301 0 : self.timelines
2302 0 : .lock()
2303 0 : .unwrap()
2304 0 : .values()
2305 0 : .map(Arc::clone)
2306 0 : .collect()
2307 0 : }
2308 :
2309 : /// Lists timelines the tenant manages, including offloaded ones.
2310 : ///
2311 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2312 0 : pub fn list_timelines_and_offloaded(
2313 0 : &self,
2314 0 : ) -> (Vec<Arc<Timeline>>, Vec<Arc<OffloadedTimeline>>) {
2315 0 : let timelines = self
2316 0 : .timelines
2317 0 : .lock()
2318 0 : .unwrap()
2319 0 : .values()
2320 0 : .map(Arc::clone)
2321 0 : .collect();
2322 0 : let offloaded = self
2323 0 : .timelines_offloaded
2324 0 : .lock()
2325 0 : .unwrap()
2326 0 : .values()
2327 0 : .map(Arc::clone)
2328 0 : .collect();
2329 0 : (timelines, offloaded)
2330 0 : }
2331 :
2332 0 : pub fn list_timeline_ids(&self) -> Vec<TimelineId> {
2333 0 : self.timelines.lock().unwrap().keys().cloned().collect()
2334 0 : }
2335 :
2336 : /// This is used by tests & import-from-basebackup.
2337 : ///
2338 : /// The returned [`UninitializedTimeline`] contains no data nor metadata and it is in
2339 : /// a state that will fail [`Tenant::load_remote_timeline`] because `disk_consistent_lsn=Lsn(0)`.
2340 : ///
2341 : /// The caller is responsible for getting the timeline into a state that will be accepted
2342 : /// by [`Tenant::load_remote_timeline`] / [`Tenant::attach`].
2343 : /// Then they may call [`UninitializedTimeline::finish_creation`] to add the timeline
2344 : /// to the [`Tenant::timelines`].
2345 : ///
2346 : /// Tests should use `Tenant::create_test_timeline` to set up the minimum required metadata keys.
2347 424 : pub(crate) async fn create_empty_timeline(
2348 424 : self: &Arc<Self>,
2349 424 : new_timeline_id: TimelineId,
2350 424 : initdb_lsn: Lsn,
2351 424 : pg_version: u32,
2352 424 : _ctx: &RequestContext,
2353 424 : ) -> anyhow::Result<UninitializedTimeline> {
2354 424 : anyhow::ensure!(
2355 424 : self.is_active(),
2356 0 : "Cannot create empty timelines on inactive tenant"
2357 : );
2358 :
2359 : // Protect against concurrent attempts to use this TimelineId
2360 424 : let create_guard = match self
2361 424 : .start_creating_timeline(new_timeline_id, CreateTimelineIdempotency::FailWithConflict)
2362 424 : .await?
2363 : {
2364 420 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2365 : StartCreatingTimelineResult::Idempotent(_) => {
2366 0 : unreachable!("FailWithConflict implies we get an error instead")
2367 : }
2368 : };
2369 :
2370 420 : let new_metadata = TimelineMetadata::new(
2371 420 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2372 420 : // make it valid, before calling finish_creation()
2373 420 : Lsn(0),
2374 420 : None,
2375 420 : None,
2376 420 : Lsn(0),
2377 420 : initdb_lsn,
2378 420 : initdb_lsn,
2379 420 : pg_version,
2380 420 : );
2381 420 : self.prepare_new_timeline(
2382 420 : new_timeline_id,
2383 420 : &new_metadata,
2384 420 : create_guard,
2385 420 : initdb_lsn,
2386 420 : None,
2387 420 : )
2388 420 : .await
2389 424 : }
2390 :
2391 : /// Helper for unit tests to create an empty timeline.
2392 : ///
2393 : /// The timeline is has state value `Active` but its background loops are not running.
2394 : // This makes the various functions which anyhow::ensure! for Active state work in tests.
2395 : // Our current tests don't need the background loops.
2396 : #[cfg(test)]
2397 404 : pub async fn create_test_timeline(
2398 404 : self: &Arc<Self>,
2399 404 : new_timeline_id: TimelineId,
2400 404 : initdb_lsn: Lsn,
2401 404 : pg_version: u32,
2402 404 : ctx: &RequestContext,
2403 404 : ) -> anyhow::Result<Arc<Timeline>> {
2404 404 : let uninit_tl = self
2405 404 : .create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2406 404 : .await?;
2407 404 : let tline = uninit_tl.raw_timeline().expect("we just created it");
2408 404 : assert_eq!(tline.get_last_record_lsn(), Lsn(0));
2409 :
2410 : // Setup minimum keys required for the timeline to be usable.
2411 404 : let mut modification = tline.begin_modification(initdb_lsn);
2412 404 : modification
2413 404 : .init_empty_test_timeline()
2414 404 : .context("init_empty_test_timeline")?;
2415 404 : modification
2416 404 : .commit(ctx)
2417 404 : .await
2418 404 : .context("commit init_empty_test_timeline modification")?;
2419 :
2420 : // Flush to disk so that uninit_tl's check for valid disk_consistent_lsn passes.
2421 404 : tline.maybe_spawn_flush_loop();
2422 404 : tline.freeze_and_flush().await.context("freeze_and_flush")?;
2423 :
2424 : // Make sure the freeze_and_flush reaches remote storage.
2425 404 : tline.remote_client.wait_completion().await.unwrap();
2426 :
2427 404 : let tl = uninit_tl.finish_creation()?;
2428 : // The non-test code would call tl.activate() here.
2429 404 : tl.set_state(TimelineState::Active);
2430 404 : Ok(tl)
2431 404 : }
2432 :
2433 : /// Helper for unit tests to create a timeline with some pre-loaded states.
2434 : #[cfg(test)]
2435 : #[allow(clippy::too_many_arguments)]
2436 76 : pub async fn create_test_timeline_with_layers(
2437 76 : self: &Arc<Self>,
2438 76 : new_timeline_id: TimelineId,
2439 76 : initdb_lsn: Lsn,
2440 76 : pg_version: u32,
2441 76 : ctx: &RequestContext,
2442 76 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
2443 76 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
2444 76 : end_lsn: Lsn,
2445 76 : ) -> anyhow::Result<Arc<Timeline>> {
2446 : use checks::check_valid_layermap;
2447 : use itertools::Itertools;
2448 :
2449 76 : let tline = self
2450 76 : .create_test_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2451 76 : .await?;
2452 76 : tline.force_advance_lsn(end_lsn);
2453 244 : for deltas in delta_layer_desc {
2454 168 : tline
2455 168 : .force_create_delta_layer(deltas, Some(initdb_lsn), ctx)
2456 168 : .await?;
2457 : }
2458 184 : for (lsn, images) in image_layer_desc {
2459 108 : tline
2460 108 : .force_create_image_layer(lsn, images, Some(initdb_lsn), ctx)
2461 108 : .await?;
2462 : }
2463 76 : let layer_names = tline
2464 76 : .layers
2465 76 : .read()
2466 76 : .await
2467 76 : .layer_map()
2468 76 : .unwrap()
2469 76 : .iter_historic_layers()
2470 352 : .map(|layer| layer.layer_name())
2471 76 : .collect_vec();
2472 76 : if let Some(err) = check_valid_layermap(&layer_names) {
2473 0 : bail!("invalid layermap: {err}");
2474 76 : }
2475 76 : Ok(tline)
2476 76 : }
2477 :
2478 : /// Create a new timeline.
2479 : ///
2480 : /// Returns the new timeline ID and reference to its Timeline object.
2481 : ///
2482 : /// If the caller specified the timeline ID to use (`new_timeline_id`), and timeline with
2483 : /// the same timeline ID already exists, returns CreateTimelineError::AlreadyExists.
2484 : #[allow(clippy::too_many_arguments)]
2485 0 : pub(crate) async fn create_timeline(
2486 0 : self: &Arc<Tenant>,
2487 0 : params: CreateTimelineParams,
2488 0 : broker_client: storage_broker::BrokerClientChannel,
2489 0 : ctx: &RequestContext,
2490 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
2491 0 : if !self.is_active() {
2492 0 : if matches!(self.current_state(), TenantState::Stopping { .. }) {
2493 0 : return Err(CreateTimelineError::ShuttingDown);
2494 : } else {
2495 0 : return Err(CreateTimelineError::Other(anyhow::anyhow!(
2496 0 : "Cannot create timelines on inactive tenant"
2497 0 : )));
2498 : }
2499 0 : }
2500 :
2501 0 : let _gate = self
2502 0 : .gate
2503 0 : .enter()
2504 0 : .map_err(|_| CreateTimelineError::ShuttingDown)?;
2505 :
2506 0 : let result: CreateTimelineResult = match params {
2507 : CreateTimelineParams::Bootstrap(CreateTimelineParamsBootstrap {
2508 0 : new_timeline_id,
2509 0 : existing_initdb_timeline_id,
2510 0 : pg_version,
2511 0 : }) => {
2512 0 : self.bootstrap_timeline(
2513 0 : new_timeline_id,
2514 0 : pg_version,
2515 0 : existing_initdb_timeline_id,
2516 0 : ctx,
2517 0 : )
2518 0 : .await?
2519 : }
2520 : CreateTimelineParams::Branch(CreateTimelineParamsBranch {
2521 0 : new_timeline_id,
2522 0 : ancestor_timeline_id,
2523 0 : mut ancestor_start_lsn,
2524 : }) => {
2525 0 : let ancestor_timeline = self
2526 0 : .get_timeline(ancestor_timeline_id, false)
2527 0 : .context("Cannot branch off the timeline that's not present in pageserver")?;
2528 :
2529 : // instead of waiting around, just deny the request because ancestor is not yet
2530 : // ready for other purposes either.
2531 0 : if !ancestor_timeline.is_active() {
2532 0 : return Err(CreateTimelineError::AncestorNotActive);
2533 0 : }
2534 0 :
2535 0 : if ancestor_timeline.is_archived() == Some(true) {
2536 0 : info!("tried to branch archived timeline");
2537 0 : return Err(CreateTimelineError::AncestorArchived);
2538 0 : }
2539 :
2540 0 : if let Some(lsn) = ancestor_start_lsn.as_mut() {
2541 0 : *lsn = lsn.align();
2542 0 :
2543 0 : let ancestor_ancestor_lsn = ancestor_timeline.get_ancestor_lsn();
2544 0 : if ancestor_ancestor_lsn > *lsn {
2545 : // can we safely just branch from the ancestor instead?
2546 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
2547 0 : "invalid start lsn {} for ancestor timeline {}: less than timeline ancestor lsn {}",
2548 0 : lsn,
2549 0 : ancestor_timeline_id,
2550 0 : ancestor_ancestor_lsn,
2551 0 : )));
2552 0 : }
2553 0 :
2554 0 : // Wait for the WAL to arrive and be processed on the parent branch up
2555 0 : // to the requested branch point. The repository code itself doesn't
2556 0 : // require it, but if we start to receive WAL on the new timeline,
2557 0 : // decoding the new WAL might need to look up previous pages, relation
2558 0 : // sizes etc. and that would get confused if the previous page versions
2559 0 : // are not in the repository yet.
2560 0 : ancestor_timeline
2561 0 : .wait_lsn(*lsn, timeline::WaitLsnWaiter::Tenant, ctx)
2562 0 : .await
2563 0 : .map_err(|e| match e {
2564 0 : e @ (WaitLsnError::Timeout(_) | WaitLsnError::BadState { .. }) => {
2565 0 : CreateTimelineError::AncestorLsn(anyhow::anyhow!(e))
2566 : }
2567 0 : WaitLsnError::Shutdown => CreateTimelineError::ShuttingDown,
2568 0 : })?;
2569 0 : }
2570 :
2571 0 : self.branch_timeline(&ancestor_timeline, new_timeline_id, ancestor_start_lsn, ctx)
2572 0 : .await?
2573 : }
2574 0 : CreateTimelineParams::ImportPgdata(params) => {
2575 0 : self.create_timeline_import_pgdata(
2576 0 : params,
2577 0 : ActivateTimelineArgs::Yes {
2578 0 : broker_client: broker_client.clone(),
2579 0 : },
2580 0 : ctx,
2581 0 : )
2582 0 : .await?
2583 : }
2584 : };
2585 :
2586 : // At this point we have dropped our guard on [`Self::timelines_creating`], and
2587 : // the timeline is visible in [`Self::timelines`], but it is _not_ durable yet. We must
2588 : // not send a success to the caller until it is. The same applies to idempotent retries.
2589 : //
2590 : // TODO: the timeline is already visible in [`Self::timelines`]; a caller could incorrectly
2591 : // assume that, because they can see the timeline via API, that the creation is done and
2592 : // that it is durable. Ideally, we would keep the timeline hidden (in [`Self::timelines_creating`])
2593 : // until it is durable, e.g., by extending the time we hold the creation guard. This also
2594 : // interacts with UninitializedTimeline and is generally a bit tricky.
2595 : //
2596 : // To re-emphasize: the only correct way to create a timeline is to repeat calling the
2597 : // creation API until it returns success. Only then is durability guaranteed.
2598 0 : info!(creation_result=%result.discriminant(), "waiting for timeline to be durable");
2599 0 : result
2600 0 : .timeline()
2601 0 : .remote_client
2602 0 : .wait_completion()
2603 0 : .await
2604 0 : .map_err(|e| match e {
2605 : WaitCompletionError::NotInitialized(
2606 0 : e, // If the queue is already stopped, it's a shutdown error.
2607 0 : ) if e.is_stopping() => CreateTimelineError::ShuttingDown,
2608 : WaitCompletionError::NotInitialized(_) => {
2609 : // This is a bug: we should never try to wait for uploads before initializing the timeline
2610 0 : debug_assert!(false);
2611 0 : CreateTimelineError::Other(anyhow::anyhow!("timeline not initialized"))
2612 : }
2613 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2614 0 : CreateTimelineError::ShuttingDown
2615 : }
2616 0 : })?;
2617 :
2618 : // The creating task is responsible for activating the timeline.
2619 : // We do this after `wait_completion()` so that we don't spin up tasks that start
2620 : // doing stuff before the IndexPart is durable in S3, which is done by the previous section.
2621 0 : let activated_timeline = match result {
2622 0 : CreateTimelineResult::Created(timeline) => {
2623 0 : timeline.activate(self.clone(), broker_client, None, ctx);
2624 0 : timeline
2625 : }
2626 0 : CreateTimelineResult::Idempotent(timeline) => {
2627 0 : info!(
2628 0 : "request was deemed idempotent, activation will be done by the creating task"
2629 : );
2630 0 : timeline
2631 : }
2632 0 : CreateTimelineResult::ImportSpawned(timeline) => {
2633 0 : info!("import task spawned, timeline will become visible and activated once the import is done");
2634 0 : timeline
2635 : }
2636 : };
2637 :
2638 0 : Ok(activated_timeline)
2639 0 : }
2640 :
2641 : /// The returned [`Arc<Timeline>`] is NOT in the [`Tenant::timelines`] map until the import
2642 : /// completes in the background. A DIFFERENT [`Arc<Timeline>`] will be inserted into the
2643 : /// [`Tenant::timelines`] map when the import completes.
2644 : /// We only return an [`Arc<Timeline>`] here so the API handler can create a [`pageserver_api::models::TimelineInfo`]
2645 : /// for the response.
2646 0 : async fn create_timeline_import_pgdata(
2647 0 : self: &Arc<Tenant>,
2648 0 : params: CreateTimelineParamsImportPgdata,
2649 0 : activate: ActivateTimelineArgs,
2650 0 : ctx: &RequestContext,
2651 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
2652 0 : let CreateTimelineParamsImportPgdata {
2653 0 : new_timeline_id,
2654 0 : location,
2655 0 : idempotency_key,
2656 0 : } = params;
2657 0 :
2658 0 : let started_at = chrono::Utc::now().naive_utc();
2659 :
2660 : //
2661 : // There's probably a simpler way to upload an index part, but, remote_timeline_client
2662 : // is the canonical way we do it.
2663 : // - create an empty timeline in-memory
2664 : // - use its remote_timeline_client to do the upload
2665 : // - dispose of the uninit timeline
2666 : // - keep the creation guard alive
2667 :
2668 0 : let timeline_create_guard = match self
2669 0 : .start_creating_timeline(
2670 0 : new_timeline_id,
2671 0 : CreateTimelineIdempotency::ImportPgdata(CreatingTimelineIdempotencyImportPgdata {
2672 0 : idempotency_key: idempotency_key.clone(),
2673 0 : }),
2674 0 : )
2675 0 : .await?
2676 : {
2677 0 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2678 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
2679 0 : return Ok(CreateTimelineResult::Idempotent(timeline))
2680 : }
2681 : };
2682 :
2683 0 : let mut uninit_timeline = {
2684 0 : let this = &self;
2685 0 : let initdb_lsn = Lsn(0);
2686 0 : let _ctx = ctx;
2687 0 : async move {
2688 0 : let new_metadata = TimelineMetadata::new(
2689 0 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2690 0 : // make it valid, before calling finish_creation()
2691 0 : Lsn(0),
2692 0 : None,
2693 0 : None,
2694 0 : Lsn(0),
2695 0 : initdb_lsn,
2696 0 : initdb_lsn,
2697 0 : 15,
2698 0 : );
2699 0 : this.prepare_new_timeline(
2700 0 : new_timeline_id,
2701 0 : &new_metadata,
2702 0 : timeline_create_guard,
2703 0 : initdb_lsn,
2704 0 : None,
2705 0 : )
2706 0 : .await
2707 0 : }
2708 0 : }
2709 0 : .await?;
2710 :
2711 0 : let in_progress = import_pgdata::index_part_format::InProgress {
2712 0 : idempotency_key,
2713 0 : location,
2714 0 : started_at,
2715 0 : };
2716 0 : let index_part = import_pgdata::index_part_format::Root::V1(
2717 0 : import_pgdata::index_part_format::V1::InProgress(in_progress),
2718 0 : );
2719 0 : uninit_timeline
2720 0 : .raw_timeline()
2721 0 : .unwrap()
2722 0 : .remote_client
2723 0 : .schedule_index_upload_for_import_pgdata_state_update(Some(index_part.clone()))?;
2724 :
2725 : // wait_completion happens in caller
2726 :
2727 0 : let (timeline, timeline_create_guard) = uninit_timeline.finish_creation_myself();
2728 0 :
2729 0 : tokio::spawn(self.clone().create_timeline_import_pgdata_task(
2730 0 : timeline.clone(),
2731 0 : index_part,
2732 0 : activate,
2733 0 : timeline_create_guard,
2734 0 : ));
2735 0 :
2736 0 : // NB: the timeline doesn't exist in self.timelines at this point
2737 0 : Ok(CreateTimelineResult::ImportSpawned(timeline))
2738 0 : }
2739 :
2740 : #[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))]
2741 : async fn create_timeline_import_pgdata_task(
2742 : self: Arc<Tenant>,
2743 : timeline: Arc<Timeline>,
2744 : index_part: import_pgdata::index_part_format::Root,
2745 : activate: ActivateTimelineArgs,
2746 : timeline_create_guard: TimelineCreateGuard,
2747 : ) {
2748 : debug_assert_current_span_has_tenant_and_timeline_id();
2749 : info!("starting");
2750 : scopeguard::defer! {info!("exiting")};
2751 :
2752 : let res = self
2753 : .create_timeline_import_pgdata_task_impl(
2754 : timeline,
2755 : index_part,
2756 : activate,
2757 : timeline_create_guard,
2758 : )
2759 : .await;
2760 : if let Err(err) = &res {
2761 : error!(?err, "task failed");
2762 : // TODO sleep & retry, sensitive to tenant shutdown
2763 : // TODO: allow timeline deletion requests => should cancel the task
2764 : }
2765 : }
2766 :
2767 0 : async fn create_timeline_import_pgdata_task_impl(
2768 0 : self: Arc<Tenant>,
2769 0 : timeline: Arc<Timeline>,
2770 0 : index_part: import_pgdata::index_part_format::Root,
2771 0 : activate: ActivateTimelineArgs,
2772 0 : timeline_create_guard: TimelineCreateGuard,
2773 0 : ) -> Result<(), anyhow::Error> {
2774 0 : let ctx = RequestContext::new(TaskKind::ImportPgdata, DownloadBehavior::Warn);
2775 0 :
2776 0 : info!("importing pgdata");
2777 0 : import_pgdata::doit(&timeline, index_part, &ctx, self.cancel.clone())
2778 0 : .await
2779 0 : .context("import")?;
2780 0 : info!("import done");
2781 :
2782 : //
2783 : // Reload timeline from remote.
2784 : // This proves that the remote state is attachable, and it reuses the code.
2785 : //
2786 : // TODO: think about whether this is safe to do with concurrent Tenant::shutdown.
2787 : // timeline_create_guard hols the tenant gate open, so, shutdown cannot _complete_ until we exit.
2788 : // But our activate() call might launch new background tasks after Tenant::shutdown
2789 : // already went past shutting down the Tenant::timelines, which this timeline here is no part of.
2790 : // I think the same problem exists with the bootstrap & branch mgmt API tasks (tenant shutting
2791 : // down while bootstrapping/branching + activating), but, the race condition is much more likely
2792 : // to manifest because of the long runtime of this import task.
2793 :
2794 : // in theory this shouldn't even .await anything except for coop yield
2795 0 : info!("shutting down timeline");
2796 0 : timeline.shutdown(ShutdownMode::Hard).await;
2797 0 : info!("timeline shut down, reloading from remote");
2798 : // TODO: we can't do the following check because create_timeline_import_pgdata must return an Arc<Timeline>
2799 : // let Some(timeline) = Arc::into_inner(timeline) else {
2800 : // anyhow::bail!("implementation error: timeline that we shut down was still referenced from somewhere");
2801 : // };
2802 0 : let timeline_id = timeline.timeline_id;
2803 0 :
2804 0 : // load from object storage like Tenant::attach does
2805 0 : let resources = self.build_timeline_resources(timeline_id);
2806 0 : let index_part = resources
2807 0 : .remote_client
2808 0 : .download_index_file(&self.cancel)
2809 0 : .await?;
2810 0 : let index_part = match index_part {
2811 : MaybeDeletedIndexPart::Deleted(_) => {
2812 : // likely concurrent delete call, cplane should prevent this
2813 0 : anyhow::bail!("index part says deleted but we are not done creating yet, this should not happen but")
2814 : }
2815 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
2816 0 : };
2817 0 : let metadata = index_part.metadata.clone();
2818 0 : self
2819 0 : .load_remote_timeline(timeline_id, index_part, metadata, resources, LoadTimelineCause::ImportPgdata{
2820 0 : create_guard: timeline_create_guard, activate, }, &ctx)
2821 0 : .await?
2822 0 : .ready_to_activate()
2823 0 : .context("implementation error: reloaded timeline still needs import after import reported success")?;
2824 :
2825 0 : anyhow::Ok(())
2826 0 : }
2827 :
2828 0 : pub(crate) async fn delete_timeline(
2829 0 : self: Arc<Self>,
2830 0 : timeline_id: TimelineId,
2831 0 : ) -> Result<(), DeleteTimelineError> {
2832 0 : DeleteTimelineFlow::run(&self, timeline_id).await?;
2833 :
2834 0 : Ok(())
2835 0 : }
2836 :
2837 : /// perform one garbage collection iteration, removing old data files from disk.
2838 : /// this function is periodically called by gc task.
2839 : /// also it can be explicitly requested through page server api 'do_gc' command.
2840 : ///
2841 : /// `target_timeline_id` specifies the timeline to GC, or None for all.
2842 : ///
2843 : /// The `horizon` an `pitr` parameters determine how much WAL history needs to be retained.
2844 : /// Also known as the retention period, or the GC cutoff point. `horizon` specifies
2845 : /// the amount of history, as LSN difference from current latest LSN on each timeline.
2846 : /// `pitr` specifies the same as a time difference from the current time. The effective
2847 : /// GC cutoff point is determined conservatively by either `horizon` and `pitr`, whichever
2848 : /// requires more history to be retained.
2849 : //
2850 1508 : pub(crate) async fn gc_iteration(
2851 1508 : &self,
2852 1508 : target_timeline_id: Option<TimelineId>,
2853 1508 : horizon: u64,
2854 1508 : pitr: Duration,
2855 1508 : cancel: &CancellationToken,
2856 1508 : ctx: &RequestContext,
2857 1508 : ) -> Result<GcResult, GcError> {
2858 1508 : // Don't start doing work during shutdown
2859 1508 : if let TenantState::Stopping { .. } = self.current_state() {
2860 0 : return Ok(GcResult::default());
2861 1508 : }
2862 1508 :
2863 1508 : // there is a global allowed_error for this
2864 1508 : if !self.is_active() {
2865 0 : return Err(GcError::NotActive);
2866 1508 : }
2867 1508 :
2868 1508 : {
2869 1508 : let conf = self.tenant_conf.load();
2870 1508 :
2871 1508 : // If we may not delete layers, then simply skip GC. Even though a tenant
2872 1508 : // in AttachedMulti state could do GC and just enqueue the blocked deletions,
2873 1508 : // the only advantage to doing it is to perhaps shrink the LayerMap metadata
2874 1508 : // a bit sooner than we would achieve by waiting for AttachedSingle status.
2875 1508 : if !conf.location.may_delete_layers_hint() {
2876 0 : info!("Skipping GC in location state {:?}", conf.location);
2877 0 : return Ok(GcResult::default());
2878 1508 : }
2879 1508 :
2880 1508 : if conf.is_gc_blocked_by_lsn_lease_deadline() {
2881 1500 : info!("Skipping GC because lsn lease deadline is not reached");
2882 1500 : return Ok(GcResult::default());
2883 8 : }
2884 : }
2885 :
2886 8 : let _guard = match self.gc_block.start().await {
2887 8 : Ok(guard) => guard,
2888 0 : Err(reasons) => {
2889 0 : info!("Skipping GC: {reasons}");
2890 0 : return Ok(GcResult::default());
2891 : }
2892 : };
2893 :
2894 8 : self.gc_iteration_internal(target_timeline_id, horizon, pitr, cancel, ctx)
2895 8 : .await
2896 1508 : }
2897 :
2898 : /// Perform one compaction iteration.
2899 : /// This function is periodically called by compactor task.
2900 : /// Also it can be explicitly requested per timeline through page server
2901 : /// api's 'compact' command.
2902 : ///
2903 : /// Returns whether we have pending compaction task.
2904 0 : async fn compaction_iteration(
2905 0 : self: &Arc<Self>,
2906 0 : cancel: &CancellationToken,
2907 0 : ctx: &RequestContext,
2908 0 : ) -> Result<bool, timeline::CompactionError> {
2909 0 : // Don't start doing work during shutdown, or when broken, we do not need those in the logs
2910 0 : if !self.is_active() {
2911 0 : return Ok(false);
2912 0 : }
2913 0 :
2914 0 : {
2915 0 : let conf = self.tenant_conf.load();
2916 0 :
2917 0 : // Note that compaction usually requires deletions, but we don't respect
2918 0 : // may_delete_layers_hint here: that is because tenants in AttachedMulti
2919 0 : // should proceed with compaction even if they can't do deletion, to avoid
2920 0 : // accumulating dangerously deep stacks of L0 layers. Deletions will be
2921 0 : // enqueued inside RemoteTimelineClient, and executed layer if/when we transition
2922 0 : // to AttachedSingle state.
2923 0 : if !conf.location.may_upload_layers_hint() {
2924 0 : info!("Skipping compaction in location state {:?}", conf.location);
2925 0 : return Ok(false);
2926 0 : }
2927 0 : }
2928 0 :
2929 0 : // Scan through the hashmap and collect a list of all the timelines,
2930 0 : // while holding the lock. Then drop the lock and actually perform the
2931 0 : // compactions. We don't want to block everything else while the
2932 0 : // compaction runs.
2933 0 : let timelines_to_compact_or_offload;
2934 0 : {
2935 0 : let timelines = self.timelines.lock().unwrap();
2936 0 : timelines_to_compact_or_offload = timelines
2937 0 : .iter()
2938 0 : .filter_map(|(timeline_id, timeline)| {
2939 0 : let (is_active, (can_offload, _)) =
2940 0 : (timeline.is_active(), timeline.can_offload());
2941 0 : let has_no_unoffloaded_children = {
2942 0 : !timelines
2943 0 : .iter()
2944 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(*timeline_id))
2945 : };
2946 0 : let config_allows_offload = self.conf.timeline_offloading
2947 0 : || self
2948 0 : .tenant_conf
2949 0 : .load()
2950 0 : .tenant_conf
2951 0 : .timeline_offloading
2952 0 : .unwrap_or_default();
2953 0 : let can_offload =
2954 0 : can_offload && has_no_unoffloaded_children && config_allows_offload;
2955 0 : if (is_active, can_offload) == (false, false) {
2956 0 : None
2957 : } else {
2958 0 : Some((*timeline_id, timeline.clone(), (is_active, can_offload)))
2959 : }
2960 0 : })
2961 0 : .collect::<Vec<_>>();
2962 0 : drop(timelines);
2963 0 : }
2964 0 :
2965 0 : // Before doing any I/O work, check our circuit breaker
2966 0 : if self.compaction_circuit_breaker.lock().unwrap().is_broken() {
2967 0 : info!("Skipping compaction due to previous failures");
2968 0 : return Ok(false);
2969 0 : }
2970 0 :
2971 0 : let mut has_pending_task = false;
2972 :
2973 0 : for (timeline_id, timeline, (can_compact, can_offload)) in &timelines_to_compact_or_offload
2974 : {
2975 : // pending_task_left == None: cannot compact, maybe still pending tasks
2976 : // pending_task_left == Some(true): compaction task left
2977 : // pending_task_left == Some(false): no compaction task left
2978 0 : let pending_task_left = if *can_compact {
2979 0 : let has_pending_l0_compaction_task = timeline
2980 0 : .compact(cancel, EnumSet::empty(), ctx)
2981 0 : .instrument(info_span!("compact_timeline", %timeline_id))
2982 0 : .await
2983 0 : .inspect_err(|e| match e {
2984 0 : timeline::CompactionError::ShuttingDown => (),
2985 0 : timeline::CompactionError::Offload(_) => {
2986 0 : // Failures to offload timelines do not trip the circuit breaker, because
2987 0 : // they do not do lots of writes the way compaction itself does: it is cheap
2988 0 : // to retry, and it would be bad to stop all compaction because of an issue with offloading.
2989 0 : }
2990 0 : timeline::CompactionError::Other(e) => {
2991 0 : self.compaction_circuit_breaker
2992 0 : .lock()
2993 0 : .unwrap()
2994 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, e);
2995 0 : }
2996 0 : })?;
2997 0 : if has_pending_l0_compaction_task {
2998 0 : Some(true)
2999 : } else {
3000 0 : let queue = {
3001 0 : let guard = self.scheduled_compaction_tasks.lock().unwrap();
3002 0 : guard.get(timeline_id).cloned()
3003 : };
3004 0 : if let Some(queue) = queue {
3005 0 : let has_pending_tasks = queue
3006 0 : .iteration(cancel, ctx, &self.gc_block, timeline)
3007 0 : .await?;
3008 0 : Some(has_pending_tasks)
3009 : } else {
3010 0 : Some(false)
3011 : }
3012 : }
3013 : } else {
3014 0 : None
3015 : };
3016 0 : has_pending_task |= pending_task_left.unwrap_or(false);
3017 0 : if pending_task_left == Some(false) && *can_offload {
3018 0 : pausable_failpoint!("before-timeline-auto-offload");
3019 0 : match offload_timeline(self, timeline)
3020 0 : .instrument(info_span!("offload_timeline", %timeline_id))
3021 0 : .await
3022 : {
3023 : Err(OffloadError::NotArchived) => {
3024 : // Ignore this, we likely raced with unarchival
3025 0 : Ok(())
3026 : }
3027 0 : other => other,
3028 0 : }?;
3029 0 : }
3030 : }
3031 :
3032 0 : self.compaction_circuit_breaker
3033 0 : .lock()
3034 0 : .unwrap()
3035 0 : .success(&CIRCUIT_BREAKERS_UNBROKEN);
3036 0 :
3037 0 : Ok(has_pending_task)
3038 0 : }
3039 :
3040 : /// Cancel scheduled compaction tasks
3041 0 : pub(crate) fn cancel_scheduled_compaction(&self, timeline_id: TimelineId) {
3042 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3043 0 : if let Some(q) = guard.get_mut(&timeline_id) {
3044 0 : q.cancel_scheduled();
3045 0 : }
3046 0 : }
3047 :
3048 0 : pub(crate) fn get_scheduled_compaction_tasks(
3049 0 : &self,
3050 0 : timeline_id: TimelineId,
3051 0 : ) -> Vec<CompactInfoResponse> {
3052 0 : let res = {
3053 0 : let guard = self.scheduled_compaction_tasks.lock().unwrap();
3054 0 : guard.get(&timeline_id).map(|q| q.remaining_jobs())
3055 : };
3056 0 : let Some((running, remaining)) = res else {
3057 0 : return Vec::new();
3058 : };
3059 0 : let mut result = Vec::new();
3060 0 : if let Some((id, running)) = running {
3061 0 : result.extend(running.into_compact_info_resp(id, true));
3062 0 : }
3063 0 : for (id, job) in remaining {
3064 0 : result.extend(job.into_compact_info_resp(id, false));
3065 0 : }
3066 0 : result
3067 0 : }
3068 :
3069 : /// Schedule a compaction task for a timeline.
3070 0 : pub(crate) async fn schedule_compaction(
3071 0 : &self,
3072 0 : timeline_id: TimelineId,
3073 0 : options: CompactOptions,
3074 0 : ) -> anyhow::Result<tokio::sync::oneshot::Receiver<()>> {
3075 0 : let (tx, rx) = tokio::sync::oneshot::channel();
3076 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3077 0 : let q = guard
3078 0 : .entry(timeline_id)
3079 0 : .or_insert_with(|| Arc::new(GcCompactionQueue::new()));
3080 0 : q.schedule_manual_compaction(options, Some(tx));
3081 0 : Ok(rx)
3082 0 : }
3083 :
3084 : // Call through to all timelines to freeze ephemeral layers if needed. Usually
3085 : // this happens during ingest: this background housekeeping is for freezing layers
3086 : // that are open but haven't been written to for some time.
3087 0 : async fn ingest_housekeeping(&self) {
3088 0 : // Scan through the hashmap and collect a list of all the timelines,
3089 0 : // while holding the lock. Then drop the lock and actually perform the
3090 0 : // compactions. We don't want to block everything else while the
3091 0 : // compaction runs.
3092 0 : let timelines = {
3093 0 : self.timelines
3094 0 : .lock()
3095 0 : .unwrap()
3096 0 : .values()
3097 0 : .filter_map(|timeline| {
3098 0 : if timeline.is_active() {
3099 0 : Some(timeline.clone())
3100 : } else {
3101 0 : None
3102 : }
3103 0 : })
3104 0 : .collect::<Vec<_>>()
3105 : };
3106 :
3107 0 : for timeline in &timelines {
3108 0 : timeline.maybe_freeze_ephemeral_layer().await;
3109 : }
3110 0 : }
3111 :
3112 0 : pub fn timeline_has_no_attached_children(&self, timeline_id: TimelineId) -> bool {
3113 0 : let timelines = self.timelines.lock().unwrap();
3114 0 : !timelines
3115 0 : .iter()
3116 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(timeline_id))
3117 0 : }
3118 :
3119 3464 : pub fn current_state(&self) -> TenantState {
3120 3464 : self.state.borrow().clone()
3121 3464 : }
3122 :
3123 1940 : pub fn is_active(&self) -> bool {
3124 1940 : self.current_state() == TenantState::Active
3125 1940 : }
3126 :
3127 0 : pub fn generation(&self) -> Generation {
3128 0 : self.generation
3129 0 : }
3130 :
3131 0 : pub(crate) fn wal_redo_manager_status(&self) -> Option<WalRedoManagerStatus> {
3132 0 : self.walredo_mgr.as_ref().and_then(|mgr| mgr.status())
3133 0 : }
3134 :
3135 : /// Changes tenant status to active, unless shutdown was already requested.
3136 : ///
3137 : /// `background_jobs_can_start` is an optional barrier set to a value during pageserver startup
3138 : /// to delay background jobs. Background jobs can be started right away when None is given.
3139 0 : fn activate(
3140 0 : self: &Arc<Self>,
3141 0 : broker_client: BrokerClientChannel,
3142 0 : background_jobs_can_start: Option<&completion::Barrier>,
3143 0 : ctx: &RequestContext,
3144 0 : ) {
3145 0 : span::debug_assert_current_span_has_tenant_id();
3146 0 :
3147 0 : let mut activating = false;
3148 0 : self.state.send_modify(|current_state| {
3149 : use pageserver_api::models::ActivatingFrom;
3150 0 : match &*current_state {
3151 : TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
3152 0 : panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {state:?}", state = current_state);
3153 : }
3154 0 : TenantState::Attaching => {
3155 0 : *current_state = TenantState::Activating(ActivatingFrom::Attaching);
3156 0 : }
3157 0 : }
3158 0 : debug!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), "Activating tenant");
3159 0 : activating = true;
3160 0 : // Continue outside the closure. We need to grab timelines.lock()
3161 0 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
3162 0 : });
3163 0 :
3164 0 : if activating {
3165 0 : let timelines_accessor = self.timelines.lock().unwrap();
3166 0 : let timelines_offloaded_accessor = self.timelines_offloaded.lock().unwrap();
3167 0 : let timelines_to_activate = timelines_accessor
3168 0 : .values()
3169 0 : .filter(|timeline| !(timeline.is_broken() || timeline.is_stopping()));
3170 0 :
3171 0 : // Before activation, populate each Timeline's GcInfo with information about its children
3172 0 : self.initialize_gc_info(&timelines_accessor, &timelines_offloaded_accessor, None);
3173 0 :
3174 0 : // Spawn gc and compaction loops. The loops will shut themselves
3175 0 : // down when they notice that the tenant is inactive.
3176 0 : tasks::start_background_loops(self, background_jobs_can_start);
3177 0 :
3178 0 : let mut activated_timelines = 0;
3179 :
3180 0 : for timeline in timelines_to_activate {
3181 0 : timeline.activate(
3182 0 : self.clone(),
3183 0 : broker_client.clone(),
3184 0 : background_jobs_can_start,
3185 0 : ctx,
3186 0 : );
3187 0 : activated_timelines += 1;
3188 0 : }
3189 :
3190 0 : self.state.send_modify(move |current_state| {
3191 0 : assert!(
3192 0 : matches!(current_state, TenantState::Activating(_)),
3193 0 : "set_stopping and set_broken wait for us to leave Activating state",
3194 : );
3195 0 : *current_state = TenantState::Active;
3196 0 :
3197 0 : let elapsed = self.constructed_at.elapsed();
3198 0 : let total_timelines = timelines_accessor.len();
3199 0 :
3200 0 : // log a lot of stuff, because some tenants sometimes suffer from user-visible
3201 0 : // times to activate. see https://github.com/neondatabase/neon/issues/4025
3202 0 : info!(
3203 0 : since_creation_millis = elapsed.as_millis(),
3204 0 : tenant_id = %self.tenant_shard_id.tenant_id,
3205 0 : shard_id = %self.tenant_shard_id.shard_slug(),
3206 0 : activated_timelines,
3207 0 : total_timelines,
3208 0 : post_state = <&'static str>::from(&*current_state),
3209 0 : "activation attempt finished"
3210 : );
3211 :
3212 0 : TENANT.activation.observe(elapsed.as_secs_f64());
3213 0 : });
3214 0 : }
3215 0 : }
3216 :
3217 : /// Shutdown the tenant and join all of the spawned tasks.
3218 : ///
3219 : /// The method caters for all use-cases:
3220 : /// - pageserver shutdown (freeze_and_flush == true)
3221 : /// - detach + ignore (freeze_and_flush == false)
3222 : ///
3223 : /// This will attempt to shutdown even if tenant is broken.
3224 : ///
3225 : /// `shutdown_progress` is a [`completion::Barrier`] for the shutdown initiated by this call.
3226 : /// If the tenant is already shutting down, we return a clone of the first shutdown call's
3227 : /// `Barrier` as an `Err`. This not-first caller can use the returned barrier to join with
3228 : /// the ongoing shutdown.
3229 12 : async fn shutdown(
3230 12 : &self,
3231 12 : shutdown_progress: completion::Barrier,
3232 12 : shutdown_mode: timeline::ShutdownMode,
3233 12 : ) -> Result<(), completion::Barrier> {
3234 12 : span::debug_assert_current_span_has_tenant_id();
3235 :
3236 : // Set tenant (and its timlines) to Stoppping state.
3237 : //
3238 : // Since we can only transition into Stopping state after activation is complete,
3239 : // run it in a JoinSet so all tenants have a chance to stop before we get SIGKILLed.
3240 : //
3241 : // Transitioning tenants to Stopping state has a couple of non-obvious side effects:
3242 : // 1. Lock out any new requests to the tenants.
3243 : // 2. Signal cancellation to WAL receivers (we wait on it below).
3244 : // 3. Signal cancellation for other tenant background loops.
3245 : // 4. ???
3246 : //
3247 : // The waiting for the cancellation is not done uniformly.
3248 : // We certainly wait for WAL receivers to shut down.
3249 : // That is necessary so that no new data comes in before the freeze_and_flush.
3250 : // But the tenant background loops are joined-on in our caller.
3251 : // It's mesed up.
3252 : // we just ignore the failure to stop
3253 :
3254 : // If we're still attaching, fire the cancellation token early to drop out: this
3255 : // will prevent us flushing, but ensures timely shutdown if some I/O during attach
3256 : // is very slow.
3257 12 : let shutdown_mode = if matches!(self.current_state(), TenantState::Attaching) {
3258 0 : self.cancel.cancel();
3259 0 :
3260 0 : // Having fired our cancellation token, do not try and flush timelines: their cancellation tokens
3261 0 : // are children of ours, so their flush loops will have shut down already
3262 0 : timeline::ShutdownMode::Hard
3263 : } else {
3264 12 : shutdown_mode
3265 : };
3266 :
3267 12 : match self.set_stopping(shutdown_progress, false, false).await {
3268 12 : Ok(()) => {}
3269 0 : Err(SetStoppingError::Broken) => {
3270 0 : // assume that this is acceptable
3271 0 : }
3272 0 : Err(SetStoppingError::AlreadyStopping(other)) => {
3273 0 : // give caller the option to wait for this this shutdown
3274 0 : info!("Tenant::shutdown: AlreadyStopping");
3275 0 : return Err(other);
3276 : }
3277 : };
3278 :
3279 12 : let mut js = tokio::task::JoinSet::new();
3280 12 : {
3281 12 : let timelines = self.timelines.lock().unwrap();
3282 12 : timelines.values().for_each(|timeline| {
3283 12 : let timeline = Arc::clone(timeline);
3284 12 : let timeline_id = timeline.timeline_id;
3285 12 : let span = tracing::info_span!("timeline_shutdown", %timeline_id, ?shutdown_mode);
3286 12 : js.spawn(async move { timeline.shutdown(shutdown_mode).instrument(span).await });
3287 12 : });
3288 12 : }
3289 12 : {
3290 12 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
3291 12 : timelines_offloaded.values().for_each(|timeline| {
3292 0 : timeline.defuse_for_tenant_drop();
3293 12 : });
3294 12 : }
3295 12 : // test_long_timeline_create_then_tenant_delete is leaning on this message
3296 12 : tracing::info!("Waiting for timelines...");
3297 24 : while let Some(res) = js.join_next().await {
3298 0 : match res {
3299 12 : Ok(()) => {}
3300 0 : Err(je) if je.is_cancelled() => unreachable!("no cancelling used"),
3301 0 : Err(je) if je.is_panic() => { /* logged already */ }
3302 0 : Err(je) => warn!("unexpected JoinError: {je:?}"),
3303 : }
3304 : }
3305 :
3306 12 : if let ShutdownMode::Reload = shutdown_mode {
3307 0 : tracing::info!("Flushing deletion queue");
3308 0 : if let Err(e) = self.deletion_queue_client.flush().await {
3309 0 : match e {
3310 0 : DeletionQueueError::ShuttingDown => {
3311 0 : // This is the only error we expect for now. In the future, if more error
3312 0 : // variants are added, we should handle them here.
3313 0 : }
3314 : }
3315 0 : }
3316 12 : }
3317 :
3318 : // We cancel the Tenant's cancellation token _after_ the timelines have all shut down. This permits
3319 : // them to continue to do work during their shutdown methods, e.g. flushing data.
3320 12 : tracing::debug!("Cancelling CancellationToken");
3321 12 : self.cancel.cancel();
3322 12 :
3323 12 : // shutdown all tenant and timeline tasks: gc, compaction, page service
3324 12 : // No new tasks will be started for this tenant because it's in `Stopping` state.
3325 12 : //
3326 12 : // this will additionally shutdown and await all timeline tasks.
3327 12 : tracing::debug!("Waiting for tasks...");
3328 12 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), None).await;
3329 :
3330 12 : if let Some(walredo_mgr) = self.walredo_mgr.as_ref() {
3331 12 : walredo_mgr.shutdown().await;
3332 0 : }
3333 :
3334 : // Wait for any in-flight operations to complete
3335 12 : self.gate.close().await;
3336 :
3337 12 : remove_tenant_metrics(&self.tenant_shard_id);
3338 12 :
3339 12 : Ok(())
3340 12 : }
3341 :
3342 : /// Change tenant status to Stopping, to mark that it is being shut down.
3343 : ///
3344 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
3345 : ///
3346 : /// This function is not cancel-safe!
3347 : ///
3348 : /// `allow_transition_from_loading` is needed for the special case of loading task deleting the tenant.
3349 : /// `allow_transition_from_attaching` is needed for the special case of attaching deleted tenant.
3350 12 : async fn set_stopping(
3351 12 : &self,
3352 12 : progress: completion::Barrier,
3353 12 : _allow_transition_from_loading: bool,
3354 12 : allow_transition_from_attaching: bool,
3355 12 : ) -> Result<(), SetStoppingError> {
3356 12 : let mut rx = self.state.subscribe();
3357 12 :
3358 12 : // cannot stop before we're done activating, so wait out until we're done activating
3359 12 : rx.wait_for(|state| match state {
3360 0 : TenantState::Attaching if allow_transition_from_attaching => true,
3361 : TenantState::Activating(_) | TenantState::Attaching => {
3362 0 : info!(
3363 0 : "waiting for {} to turn Active|Broken|Stopping",
3364 0 : <&'static str>::from(state)
3365 : );
3366 0 : false
3367 : }
3368 12 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3369 12 : })
3370 12 : .await
3371 12 : .expect("cannot drop self.state while on a &self method");
3372 12 :
3373 12 : // we now know we're done activating, let's see whether this task is the winner to transition into Stopping
3374 12 : let mut err = None;
3375 12 : let stopping = self.state.send_if_modified(|current_state| match current_state {
3376 : TenantState::Activating(_) => {
3377 0 : unreachable!("1we ensured above that we're done with activation, and, there is no re-activation")
3378 : }
3379 : TenantState::Attaching => {
3380 0 : if !allow_transition_from_attaching {
3381 0 : unreachable!("2we ensured above that we're done with activation, and, there is no re-activation")
3382 0 : };
3383 0 : *current_state = TenantState::Stopping { progress };
3384 0 : true
3385 : }
3386 : TenantState::Active => {
3387 : // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
3388 : // are created after the transition to Stopping. That's harmless, as the Timelines
3389 : // won't be accessible to anyone afterwards, because the Tenant is in Stopping state.
3390 12 : *current_state = TenantState::Stopping { progress };
3391 12 : // Continue stopping outside the closure. We need to grab timelines.lock()
3392 12 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
3393 12 : true
3394 : }
3395 0 : TenantState::Broken { reason, .. } => {
3396 0 : info!(
3397 0 : "Cannot set tenant to Stopping state, it is in Broken state due to: {reason}"
3398 : );
3399 0 : err = Some(SetStoppingError::Broken);
3400 0 : false
3401 : }
3402 0 : TenantState::Stopping { progress } => {
3403 0 : info!("Tenant is already in Stopping state");
3404 0 : err = Some(SetStoppingError::AlreadyStopping(progress.clone()));
3405 0 : false
3406 : }
3407 12 : });
3408 12 : match (stopping, err) {
3409 12 : (true, None) => {} // continue
3410 0 : (false, Some(err)) => return Err(err),
3411 0 : (true, Some(_)) => unreachable!(
3412 0 : "send_if_modified closure must error out if not transitioning to Stopping"
3413 0 : ),
3414 0 : (false, None) => unreachable!(
3415 0 : "send_if_modified closure must return true if transitioning to Stopping"
3416 0 : ),
3417 : }
3418 :
3419 12 : let timelines_accessor = self.timelines.lock().unwrap();
3420 12 : let not_broken_timelines = timelines_accessor
3421 12 : .values()
3422 12 : .filter(|timeline| !timeline.is_broken());
3423 24 : for timeline in not_broken_timelines {
3424 12 : timeline.set_state(TimelineState::Stopping);
3425 12 : }
3426 12 : Ok(())
3427 12 : }
3428 :
3429 : /// Method for tenant::mgr to transition us into Broken state in case of a late failure in
3430 : /// `remove_tenant_from_memory`
3431 : ///
3432 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
3433 : ///
3434 : /// In tests, we also use this to set tenants to Broken state on purpose.
3435 0 : pub(crate) async fn set_broken(&self, reason: String) {
3436 0 : let mut rx = self.state.subscribe();
3437 0 :
3438 0 : // The load & attach routines own the tenant state until it has reached `Active`.
3439 0 : // So, wait until it's done.
3440 0 : rx.wait_for(|state| match state {
3441 : TenantState::Activating(_) | TenantState::Attaching => {
3442 0 : info!(
3443 0 : "waiting for {} to turn Active|Broken|Stopping",
3444 0 : <&'static str>::from(state)
3445 : );
3446 0 : false
3447 : }
3448 0 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3449 0 : })
3450 0 : .await
3451 0 : .expect("cannot drop self.state while on a &self method");
3452 0 :
3453 0 : // we now know we're done activating, let's see whether this task is the winner to transition into Broken
3454 0 : self.set_broken_no_wait(reason)
3455 0 : }
3456 :
3457 0 : pub(crate) fn set_broken_no_wait(&self, reason: impl Display) {
3458 0 : let reason = reason.to_string();
3459 0 : self.state.send_modify(|current_state| {
3460 0 : match *current_state {
3461 : TenantState::Activating(_) | TenantState::Attaching => {
3462 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
3463 : }
3464 : TenantState::Active => {
3465 0 : if cfg!(feature = "testing") {
3466 0 : warn!("Changing Active tenant to Broken state, reason: {}", reason);
3467 0 : *current_state = TenantState::broken_from_reason(reason);
3468 : } else {
3469 0 : unreachable!("not allowed to call set_broken on Active tenants in non-testing builds")
3470 : }
3471 : }
3472 : TenantState::Broken { .. } => {
3473 0 : warn!("Tenant is already in Broken state");
3474 : }
3475 : // This is the only "expected" path, any other path is a bug.
3476 : TenantState::Stopping { .. } => {
3477 0 : warn!(
3478 0 : "Marking Stopping tenant as Broken state, reason: {}",
3479 : reason
3480 : );
3481 0 : *current_state = TenantState::broken_from_reason(reason);
3482 : }
3483 : }
3484 0 : });
3485 0 : }
3486 :
3487 0 : pub fn subscribe_for_state_updates(&self) -> watch::Receiver<TenantState> {
3488 0 : self.state.subscribe()
3489 0 : }
3490 :
3491 : /// The activate_now semaphore is initialized with zero units. As soon as
3492 : /// we add a unit, waiters will be able to acquire a unit and proceed.
3493 0 : pub(crate) fn activate_now(&self) {
3494 0 : self.activate_now_sem.add_permits(1);
3495 0 : }
3496 :
3497 0 : pub(crate) async fn wait_to_become_active(
3498 0 : &self,
3499 0 : timeout: Duration,
3500 0 : ) -> Result<(), GetActiveTenantError> {
3501 0 : let mut receiver = self.state.subscribe();
3502 : loop {
3503 0 : let current_state = receiver.borrow_and_update().clone();
3504 0 : match current_state {
3505 : TenantState::Attaching | TenantState::Activating(_) => {
3506 : // in these states, there's a chance that we can reach ::Active
3507 0 : self.activate_now();
3508 0 : match timeout_cancellable(timeout, &self.cancel, receiver.changed()).await {
3509 0 : Ok(r) => {
3510 0 : r.map_err(
3511 0 : |_e: tokio::sync::watch::error::RecvError|
3512 : // Tenant existed but was dropped: report it as non-existent
3513 0 : GetActiveTenantError::NotFound(GetTenantError::ShardNotFound(self.tenant_shard_id))
3514 0 : )?
3515 : }
3516 : Err(TimeoutCancellableError::Cancelled) => {
3517 0 : return Err(GetActiveTenantError::Cancelled);
3518 : }
3519 : Err(TimeoutCancellableError::Timeout) => {
3520 0 : return Err(GetActiveTenantError::WaitForActiveTimeout {
3521 0 : latest_state: Some(self.current_state()),
3522 0 : wait_time: timeout,
3523 0 : });
3524 : }
3525 : }
3526 : }
3527 : TenantState::Active { .. } => {
3528 0 : return Ok(());
3529 : }
3530 0 : TenantState::Broken { reason, .. } => {
3531 0 : // This is fatal, and reported distinctly from the general case of "will never be active" because
3532 0 : // it's logically a 500 to external API users (broken is always a bug).
3533 0 : return Err(GetActiveTenantError::Broken(reason));
3534 : }
3535 : TenantState::Stopping { .. } => {
3536 : // There's no chance the tenant can transition back into ::Active
3537 0 : return Err(GetActiveTenantError::WillNotBecomeActive(current_state));
3538 : }
3539 : }
3540 : }
3541 0 : }
3542 :
3543 0 : pub(crate) fn get_attach_mode(&self) -> AttachmentMode {
3544 0 : self.tenant_conf.load().location.attach_mode
3545 0 : }
3546 :
3547 : /// For API access: generate a LocationConfig equivalent to the one that would be used to
3548 : /// create a Tenant in the same state. Do not use this in hot paths: it's for relatively
3549 : /// rare external API calls, like a reconciliation at startup.
3550 0 : pub(crate) fn get_location_conf(&self) -> models::LocationConfig {
3551 0 : let conf = self.tenant_conf.load();
3552 :
3553 0 : let location_config_mode = match conf.location.attach_mode {
3554 0 : AttachmentMode::Single => models::LocationConfigMode::AttachedSingle,
3555 0 : AttachmentMode::Multi => models::LocationConfigMode::AttachedMulti,
3556 0 : AttachmentMode::Stale => models::LocationConfigMode::AttachedStale,
3557 : };
3558 :
3559 : // We have a pageserver TenantConf, we need the API-facing TenantConfig.
3560 0 : let tenant_config: models::TenantConfig = conf.tenant_conf.clone().into();
3561 0 :
3562 0 : models::LocationConfig {
3563 0 : mode: location_config_mode,
3564 0 : generation: self.generation.into(),
3565 0 : secondary_conf: None,
3566 0 : shard_number: self.shard_identity.number.0,
3567 0 : shard_count: self.shard_identity.count.literal(),
3568 0 : shard_stripe_size: self.shard_identity.stripe_size.0,
3569 0 : tenant_conf: tenant_config,
3570 0 : }
3571 0 : }
3572 :
3573 0 : pub(crate) fn get_tenant_shard_id(&self) -> &TenantShardId {
3574 0 : &self.tenant_shard_id
3575 0 : }
3576 :
3577 0 : pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
3578 0 : self.shard_identity.stripe_size
3579 0 : }
3580 :
3581 0 : pub(crate) fn get_generation(&self) -> Generation {
3582 0 : self.generation
3583 0 : }
3584 :
3585 : /// This function partially shuts down the tenant (it shuts down the Timelines) and is fallible,
3586 : /// and can leave the tenant in a bad state if it fails. The caller is responsible for
3587 : /// resetting this tenant to a valid state if we fail.
3588 0 : pub(crate) async fn split_prepare(
3589 0 : &self,
3590 0 : child_shards: &Vec<TenantShardId>,
3591 0 : ) -> anyhow::Result<()> {
3592 0 : let (timelines, offloaded) = {
3593 0 : let timelines = self.timelines.lock().unwrap();
3594 0 : let offloaded = self.timelines_offloaded.lock().unwrap();
3595 0 : (timelines.clone(), offloaded.clone())
3596 0 : };
3597 0 : let timelines_iter = timelines
3598 0 : .values()
3599 0 : .map(TimelineOrOffloadedArcRef::<'_>::from)
3600 0 : .chain(
3601 0 : offloaded
3602 0 : .values()
3603 0 : .map(TimelineOrOffloadedArcRef::<'_>::from),
3604 0 : );
3605 0 : for timeline in timelines_iter {
3606 : // We do not block timeline creation/deletion during splits inside the pageserver: it is up to higher levels
3607 : // to ensure that they do not start a split if currently in the process of doing these.
3608 :
3609 0 : let timeline_id = timeline.timeline_id();
3610 :
3611 0 : if let TimelineOrOffloadedArcRef::Timeline(timeline) = timeline {
3612 : // Upload an index from the parent: this is partly to provide freshness for the
3613 : // child tenants that will copy it, and partly for general ease-of-debugging: there will
3614 : // always be a parent shard index in the same generation as we wrote the child shard index.
3615 0 : tracing::info!(%timeline_id, "Uploading index");
3616 0 : timeline
3617 0 : .remote_client
3618 0 : .schedule_index_upload_for_file_changes()?;
3619 0 : timeline.remote_client.wait_completion().await?;
3620 0 : }
3621 :
3622 0 : let remote_client = match timeline {
3623 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.remote_client.clone(),
3624 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => {
3625 0 : let remote_client = self
3626 0 : .build_timeline_client(offloaded.timeline_id, self.remote_storage.clone());
3627 0 : Arc::new(remote_client)
3628 : }
3629 : };
3630 :
3631 : // Shut down the timeline's remote client: this means that the indices we write
3632 : // for child shards will not be invalidated by the parent shard deleting layers.
3633 0 : tracing::info!(%timeline_id, "Shutting down remote storage client");
3634 0 : remote_client.shutdown().await;
3635 :
3636 : // Download methods can still be used after shutdown, as they don't flow through the remote client's
3637 : // queue. In principal the RemoteTimelineClient could provide this without downloading it, but this
3638 : // operation is rare, so it's simpler to just download it (and robustly guarantees that the index
3639 : // we use here really is the remotely persistent one).
3640 0 : tracing::info!(%timeline_id, "Downloading index_part from parent");
3641 0 : let result = remote_client
3642 0 : .download_index_file(&self.cancel)
3643 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))
3644 0 : .await?;
3645 0 : let index_part = match result {
3646 : MaybeDeletedIndexPart::Deleted(_) => {
3647 0 : anyhow::bail!("Timeline deletion happened concurrently with split")
3648 : }
3649 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
3650 : };
3651 :
3652 0 : for child_shard in child_shards {
3653 0 : tracing::info!(%timeline_id, "Uploading index_part for child {}", child_shard.to_index());
3654 0 : upload_index_part(
3655 0 : &self.remote_storage,
3656 0 : child_shard,
3657 0 : &timeline_id,
3658 0 : self.generation,
3659 0 : &index_part,
3660 0 : &self.cancel,
3661 0 : )
3662 0 : .await?;
3663 : }
3664 : }
3665 :
3666 0 : let tenant_manifest = self.build_tenant_manifest();
3667 0 : for child_shard in child_shards {
3668 0 : tracing::info!(
3669 0 : "Uploading tenant manifest for child {}",
3670 0 : child_shard.to_index()
3671 : );
3672 0 : upload_tenant_manifest(
3673 0 : &self.remote_storage,
3674 0 : child_shard,
3675 0 : self.generation,
3676 0 : &tenant_manifest,
3677 0 : &self.cancel,
3678 0 : )
3679 0 : .await?;
3680 : }
3681 :
3682 0 : Ok(())
3683 0 : }
3684 :
3685 0 : pub(crate) fn get_sizes(&self) -> TopTenantShardItem {
3686 0 : let mut result = TopTenantShardItem {
3687 0 : id: self.tenant_shard_id,
3688 0 : resident_size: 0,
3689 0 : physical_size: 0,
3690 0 : max_logical_size: 0,
3691 0 : };
3692 :
3693 0 : for timeline in self.timelines.lock().unwrap().values() {
3694 0 : result.resident_size += timeline.metrics.resident_physical_size_gauge.get();
3695 0 :
3696 0 : result.physical_size += timeline
3697 0 : .remote_client
3698 0 : .metrics
3699 0 : .remote_physical_size_gauge
3700 0 : .get();
3701 0 : result.max_logical_size = std::cmp::max(
3702 0 : result.max_logical_size,
3703 0 : timeline.metrics.current_logical_size_gauge.get(),
3704 0 : );
3705 0 : }
3706 :
3707 0 : result
3708 0 : }
3709 : }
3710 :
3711 : /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
3712 : /// perform a topological sort, so that the parent of each timeline comes
3713 : /// before the children.
3714 : /// E extracts the ancestor from T
3715 : /// This allows for T to be different. It can be TimelineMetadata, can be Timeline itself, etc.
3716 440 : fn tree_sort_timelines<T, E>(
3717 440 : timelines: HashMap<TimelineId, T>,
3718 440 : extractor: E,
3719 440 : ) -> anyhow::Result<Vec<(TimelineId, T)>>
3720 440 : where
3721 440 : E: Fn(&T) -> Option<TimelineId>,
3722 440 : {
3723 440 : let mut result = Vec::with_capacity(timelines.len());
3724 440 :
3725 440 : let mut now = Vec::with_capacity(timelines.len());
3726 440 : // (ancestor, children)
3727 440 : let mut later: HashMap<TimelineId, Vec<(TimelineId, T)>> =
3728 440 : HashMap::with_capacity(timelines.len());
3729 :
3730 452 : for (timeline_id, value) in timelines {
3731 12 : if let Some(ancestor_id) = extractor(&value) {
3732 4 : let children = later.entry(ancestor_id).or_default();
3733 4 : children.push((timeline_id, value));
3734 8 : } else {
3735 8 : now.push((timeline_id, value));
3736 8 : }
3737 : }
3738 :
3739 452 : while let Some((timeline_id, metadata)) = now.pop() {
3740 12 : result.push((timeline_id, metadata));
3741 : // All children of this can be loaded now
3742 12 : if let Some(mut children) = later.remove(&timeline_id) {
3743 4 : now.append(&mut children);
3744 8 : }
3745 : }
3746 :
3747 : // All timelines should be visited now. Unless there were timelines with missing ancestors.
3748 440 : if !later.is_empty() {
3749 0 : for (missing_id, orphan_ids) in later {
3750 0 : for (orphan_id, _) in orphan_ids {
3751 0 : error!("could not load timeline {orphan_id} because its ancestor timeline {missing_id} could not be loaded");
3752 : }
3753 : }
3754 0 : bail!("could not load tenant because some timelines are missing ancestors");
3755 440 : }
3756 440 :
3757 440 : Ok(result)
3758 440 : }
3759 :
3760 : enum ActivateTimelineArgs {
3761 : Yes {
3762 : broker_client: storage_broker::BrokerClientChannel,
3763 : },
3764 : No,
3765 : }
3766 :
3767 : impl Tenant {
3768 0 : pub fn tenant_specific_overrides(&self) -> TenantConfOpt {
3769 0 : self.tenant_conf.load().tenant_conf.clone()
3770 0 : }
3771 :
3772 0 : pub fn effective_config(&self) -> TenantConf {
3773 0 : self.tenant_specific_overrides()
3774 0 : .merge(self.conf.default_tenant_conf.clone())
3775 0 : }
3776 :
3777 0 : pub fn get_checkpoint_distance(&self) -> u64 {
3778 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3779 0 : tenant_conf
3780 0 : .checkpoint_distance
3781 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
3782 0 : }
3783 :
3784 0 : pub fn get_checkpoint_timeout(&self) -> Duration {
3785 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3786 0 : tenant_conf
3787 0 : .checkpoint_timeout
3788 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
3789 0 : }
3790 :
3791 0 : pub fn get_compaction_target_size(&self) -> u64 {
3792 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3793 0 : tenant_conf
3794 0 : .compaction_target_size
3795 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
3796 0 : }
3797 :
3798 0 : pub fn get_compaction_period(&self) -> Duration {
3799 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3800 0 : tenant_conf
3801 0 : .compaction_period
3802 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
3803 0 : }
3804 :
3805 0 : pub fn get_compaction_threshold(&self) -> usize {
3806 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3807 0 : tenant_conf
3808 0 : .compaction_threshold
3809 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
3810 0 : }
3811 :
3812 0 : pub fn get_gc_horizon(&self) -> u64 {
3813 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3814 0 : tenant_conf
3815 0 : .gc_horizon
3816 0 : .unwrap_or(self.conf.default_tenant_conf.gc_horizon)
3817 0 : }
3818 :
3819 0 : pub fn get_gc_period(&self) -> Duration {
3820 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3821 0 : tenant_conf
3822 0 : .gc_period
3823 0 : .unwrap_or(self.conf.default_tenant_conf.gc_period)
3824 0 : }
3825 :
3826 0 : pub fn get_image_creation_threshold(&self) -> usize {
3827 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3828 0 : tenant_conf
3829 0 : .image_creation_threshold
3830 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
3831 0 : }
3832 :
3833 0 : pub fn get_pitr_interval(&self) -> Duration {
3834 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3835 0 : tenant_conf
3836 0 : .pitr_interval
3837 0 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
3838 0 : }
3839 :
3840 0 : pub fn get_min_resident_size_override(&self) -> Option<u64> {
3841 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3842 0 : tenant_conf
3843 0 : .min_resident_size_override
3844 0 : .or(self.conf.default_tenant_conf.min_resident_size_override)
3845 0 : }
3846 :
3847 0 : pub fn get_heatmap_period(&self) -> Option<Duration> {
3848 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3849 0 : let heatmap_period = tenant_conf
3850 0 : .heatmap_period
3851 0 : .unwrap_or(self.conf.default_tenant_conf.heatmap_period);
3852 0 : if heatmap_period.is_zero() {
3853 0 : None
3854 : } else {
3855 0 : Some(heatmap_period)
3856 : }
3857 0 : }
3858 :
3859 8 : pub fn get_lsn_lease_length(&self) -> Duration {
3860 8 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3861 8 : tenant_conf
3862 8 : .lsn_lease_length
3863 8 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
3864 8 : }
3865 :
3866 : /// Generate an up-to-date TenantManifest based on the state of this Tenant.
3867 4 : fn build_tenant_manifest(&self) -> TenantManifest {
3868 4 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
3869 4 :
3870 4 : let mut timeline_manifests = timelines_offloaded
3871 4 : .iter()
3872 4 : .map(|(_timeline_id, offloaded)| offloaded.manifest())
3873 4 : .collect::<Vec<_>>();
3874 4 : // Sort the manifests so that our output is deterministic
3875 4 : timeline_manifests.sort_by_key(|timeline_manifest| timeline_manifest.timeline_id);
3876 4 :
3877 4 : TenantManifest {
3878 4 : version: LATEST_TENANT_MANIFEST_VERSION,
3879 4 : offloaded_timelines: timeline_manifests,
3880 4 : }
3881 4 : }
3882 :
3883 0 : pub fn update_tenant_config<F: Fn(TenantConfOpt) -> anyhow::Result<TenantConfOpt>>(
3884 0 : &self,
3885 0 : update: F,
3886 0 : ) -> anyhow::Result<TenantConfOpt> {
3887 0 : // Use read-copy-update in order to avoid overwriting the location config
3888 0 : // state if this races with [`Tenant::set_new_location_config`]. Note that
3889 0 : // this race is not possible if both request types come from the storage
3890 0 : // controller (as they should!) because an exclusive op lock is required
3891 0 : // on the storage controller side.
3892 0 :
3893 0 : self.tenant_conf
3894 0 : .try_rcu(|attached_conf| -> Result<_, anyhow::Error> {
3895 0 : Ok(Arc::new(AttachedTenantConf {
3896 0 : tenant_conf: update(attached_conf.tenant_conf.clone())?,
3897 0 : location: attached_conf.location,
3898 0 : lsn_lease_deadline: attached_conf.lsn_lease_deadline,
3899 : }))
3900 0 : })?;
3901 :
3902 0 : let updated = self.tenant_conf.load();
3903 0 :
3904 0 : self.tenant_conf_updated(&updated.tenant_conf);
3905 0 : // Don't hold self.timelines.lock() during the notifies.
3906 0 : // There's no risk of deadlock right now, but there could be if we consolidate
3907 0 : // mutexes in struct Timeline in the future.
3908 0 : let timelines = self.list_timelines();
3909 0 : for timeline in timelines {
3910 0 : timeline.tenant_conf_updated(&updated);
3911 0 : }
3912 :
3913 0 : Ok(updated.tenant_conf.clone())
3914 0 : }
3915 :
3916 0 : pub(crate) fn set_new_location_config(&self, new_conf: AttachedTenantConf) {
3917 0 : let new_tenant_conf = new_conf.tenant_conf.clone();
3918 0 :
3919 0 : self.tenant_conf.store(Arc::new(new_conf.clone()));
3920 0 :
3921 0 : self.tenant_conf_updated(&new_tenant_conf);
3922 0 : // Don't hold self.timelines.lock() during the notifies.
3923 0 : // There's no risk of deadlock right now, but there could be if we consolidate
3924 0 : // mutexes in struct Timeline in the future.
3925 0 : let timelines = self.list_timelines();
3926 0 : for timeline in timelines {
3927 0 : timeline.tenant_conf_updated(&new_conf);
3928 0 : }
3929 0 : }
3930 :
3931 440 : fn get_pagestream_throttle_config(
3932 440 : psconf: &'static PageServerConf,
3933 440 : overrides: &TenantConfOpt,
3934 440 : ) -> throttle::Config {
3935 440 : overrides
3936 440 : .timeline_get_throttle
3937 440 : .clone()
3938 440 : .unwrap_or(psconf.default_tenant_conf.timeline_get_throttle.clone())
3939 440 : }
3940 :
3941 0 : pub(crate) fn tenant_conf_updated(&self, new_conf: &TenantConfOpt) {
3942 0 : let conf = Self::get_pagestream_throttle_config(self.conf, new_conf);
3943 0 : self.pagestream_throttle.reconfigure(conf)
3944 0 : }
3945 :
3946 : /// Helper function to create a new Timeline struct.
3947 : ///
3948 : /// The returned Timeline is in Loading state. The caller is responsible for
3949 : /// initializing any on-disk state, and for inserting the Timeline to the 'timelines'
3950 : /// map.
3951 : ///
3952 : /// `validate_ancestor == false` is used when a timeline is created for deletion
3953 : /// and we might not have the ancestor present anymore which is fine for to be
3954 : /// deleted timelines.
3955 : #[allow(clippy::too_many_arguments)]
3956 892 : fn create_timeline_struct(
3957 892 : &self,
3958 892 : new_timeline_id: TimelineId,
3959 892 : new_metadata: &TimelineMetadata,
3960 892 : ancestor: Option<Arc<Timeline>>,
3961 892 : resources: TimelineResources,
3962 892 : cause: CreateTimelineCause,
3963 892 : create_idempotency: CreateTimelineIdempotency,
3964 892 : ) -> anyhow::Result<Arc<Timeline>> {
3965 892 : let state = match cause {
3966 : CreateTimelineCause::Load => {
3967 892 : let ancestor_id = new_metadata.ancestor_timeline();
3968 892 : anyhow::ensure!(
3969 892 : ancestor_id == ancestor.as_ref().map(|t| t.timeline_id),
3970 0 : "Timeline's {new_timeline_id} ancestor {ancestor_id:?} was not found"
3971 : );
3972 892 : TimelineState::Loading
3973 : }
3974 0 : CreateTimelineCause::Delete => TimelineState::Stopping,
3975 : };
3976 :
3977 892 : let pg_version = new_metadata.pg_version();
3978 892 :
3979 892 : let timeline = Timeline::new(
3980 892 : self.conf,
3981 892 : Arc::clone(&self.tenant_conf),
3982 892 : new_metadata,
3983 892 : ancestor,
3984 892 : new_timeline_id,
3985 892 : self.tenant_shard_id,
3986 892 : self.generation,
3987 892 : self.shard_identity,
3988 892 : self.walredo_mgr.clone(),
3989 892 : resources,
3990 892 : pg_version,
3991 892 : state,
3992 892 : self.attach_wal_lag_cooldown.clone(),
3993 892 : create_idempotency,
3994 892 : self.cancel.child_token(),
3995 892 : );
3996 892 :
3997 892 : Ok(timeline)
3998 892 : }
3999 :
4000 : /// [`Tenant::shutdown`] must be called before dropping the returned [`Tenant`] object
4001 : /// to ensure proper cleanup of background tasks and metrics.
4002 : //
4003 : // Allow too_many_arguments because a constructor's argument list naturally grows with the
4004 : // number of attributes in the struct: breaking these out into a builder wouldn't be helpful.
4005 : #[allow(clippy::too_many_arguments)]
4006 440 : fn new(
4007 440 : state: TenantState,
4008 440 : conf: &'static PageServerConf,
4009 440 : attached_conf: AttachedTenantConf,
4010 440 : shard_identity: ShardIdentity,
4011 440 : walredo_mgr: Option<Arc<WalRedoManager>>,
4012 440 : tenant_shard_id: TenantShardId,
4013 440 : remote_storage: GenericRemoteStorage,
4014 440 : deletion_queue_client: DeletionQueueClient,
4015 440 : l0_flush_global_state: L0FlushGlobalState,
4016 440 : ) -> Tenant {
4017 440 : debug_assert!(
4018 440 : !attached_conf.location.generation.is_none() || conf.control_plane_api.is_none()
4019 : );
4020 :
4021 440 : let (state, mut rx) = watch::channel(state);
4022 440 :
4023 440 : tokio::spawn(async move {
4024 439 : // reflect tenant state in metrics:
4025 439 : // - global per tenant state: TENANT_STATE_METRIC
4026 439 : // - "set" of broken tenants: BROKEN_TENANTS_SET
4027 439 : //
4028 439 : // set of broken tenants should not have zero counts so that it remains accessible for
4029 439 : // alerting.
4030 439 :
4031 439 : let tid = tenant_shard_id.to_string();
4032 439 : let shard_id = tenant_shard_id.shard_slug().to_string();
4033 439 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
4034 :
4035 878 : fn inspect_state(state: &TenantState) -> ([&'static str; 1], bool) {
4036 878 : ([state.into()], matches!(state, TenantState::Broken { .. }))
4037 878 : }
4038 :
4039 439 : let mut tuple = inspect_state(&rx.borrow_and_update());
4040 439 :
4041 439 : let is_broken = tuple.1;
4042 439 : let mut counted_broken = if is_broken {
4043 : // add the id to the set right away, there should not be any updates on the channel
4044 : // after before tenant is removed, if ever
4045 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
4046 0 : true
4047 : } else {
4048 439 : false
4049 : };
4050 :
4051 : loop {
4052 878 : let labels = &tuple.0;
4053 878 : let current = TENANT_STATE_METRIC.with_label_values(labels);
4054 878 : current.inc();
4055 878 :
4056 878 : if rx.changed().await.is_err() {
4057 : // tenant has been dropped
4058 28 : current.dec();
4059 28 : drop(BROKEN_TENANTS_SET.remove_label_values(set_key));
4060 28 : break;
4061 439 : }
4062 439 :
4063 439 : current.dec();
4064 439 : tuple = inspect_state(&rx.borrow_and_update());
4065 439 :
4066 439 : let is_broken = tuple.1;
4067 439 : if is_broken && !counted_broken {
4068 0 : counted_broken = true;
4069 0 : // insert the tenant_id (back) into the set while avoiding needless counter
4070 0 : // access
4071 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
4072 439 : }
4073 : }
4074 440 : });
4075 440 :
4076 440 : Tenant {
4077 440 : tenant_shard_id,
4078 440 : shard_identity,
4079 440 : generation: attached_conf.location.generation,
4080 440 : conf,
4081 440 : // using now here is good enough approximation to catch tenants with really long
4082 440 : // activation times.
4083 440 : constructed_at: Instant::now(),
4084 440 : timelines: Mutex::new(HashMap::new()),
4085 440 : timelines_creating: Mutex::new(HashSet::new()),
4086 440 : timelines_offloaded: Mutex::new(HashMap::new()),
4087 440 : tenant_manifest_upload: Default::default(),
4088 440 : gc_cs: tokio::sync::Mutex::new(()),
4089 440 : walredo_mgr,
4090 440 : remote_storage,
4091 440 : deletion_queue_client,
4092 440 : state,
4093 440 : cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
4094 440 : cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
4095 440 : eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
4096 440 : compaction_circuit_breaker: std::sync::Mutex::new(CircuitBreaker::new(
4097 440 : format!("compaction-{tenant_shard_id}"),
4098 440 : 5,
4099 440 : // Compaction can be a very expensive operation, and might leak disk space. It also ought
4100 440 : // to be infallible, as long as remote storage is available. So if it repeatedly fails,
4101 440 : // use an extremely long backoff.
4102 440 : Some(Duration::from_secs(3600 * 24)),
4103 440 : )),
4104 440 : scheduled_compaction_tasks: Mutex::new(Default::default()),
4105 440 : activate_now_sem: tokio::sync::Semaphore::new(0),
4106 440 : attach_wal_lag_cooldown: Arc::new(std::sync::OnceLock::new()),
4107 440 : cancel: CancellationToken::default(),
4108 440 : gate: Gate::default(),
4109 440 : pagestream_throttle: Arc::new(throttle::Throttle::new(
4110 440 : Tenant::get_pagestream_throttle_config(conf, &attached_conf.tenant_conf),
4111 440 : )),
4112 440 : pagestream_throttle_metrics: Arc::new(
4113 440 : crate::metrics::tenant_throttling::Pagestream::new(&tenant_shard_id),
4114 440 : ),
4115 440 : tenant_conf: Arc::new(ArcSwap::from_pointee(attached_conf)),
4116 440 : ongoing_timeline_detach: std::sync::Mutex::default(),
4117 440 : gc_block: Default::default(),
4118 440 : l0_flush_global_state,
4119 440 : }
4120 440 : }
4121 :
4122 : /// Locate and load config
4123 0 : pub(super) fn load_tenant_config(
4124 0 : conf: &'static PageServerConf,
4125 0 : tenant_shard_id: &TenantShardId,
4126 0 : ) -> Result<LocationConf, LoadConfigError> {
4127 0 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4128 0 :
4129 0 : info!("loading tenant configuration from {config_path}");
4130 :
4131 : // load and parse file
4132 0 : let config = fs::read_to_string(&config_path).map_err(|e| {
4133 0 : match e.kind() {
4134 : std::io::ErrorKind::NotFound => {
4135 : // The config should almost always exist for a tenant directory:
4136 : // - When attaching a tenant, the config is the first thing we write
4137 : // - When detaching a tenant, we atomically move the directory to a tmp location
4138 : // before deleting contents.
4139 : //
4140 : // The very rare edge case that can result in a missing config is if we crash during attach
4141 : // between creating directory and writing config. Callers should handle that as if the
4142 : // directory didn't exist.
4143 :
4144 0 : LoadConfigError::NotFound(config_path)
4145 : }
4146 : _ => {
4147 : // No IO errors except NotFound are acceptable here: other kinds of error indicate local storage or permissions issues
4148 : // that we cannot cleanly recover
4149 0 : crate::virtual_file::on_fatal_io_error(&e, "Reading tenant config file")
4150 : }
4151 : }
4152 0 : })?;
4153 :
4154 0 : Ok(toml_edit::de::from_str::<LocationConf>(&config)?)
4155 0 : }
4156 :
4157 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4158 : pub(super) async fn persist_tenant_config(
4159 : conf: &'static PageServerConf,
4160 : tenant_shard_id: &TenantShardId,
4161 : location_conf: &LocationConf,
4162 : ) -> std::io::Result<()> {
4163 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4164 :
4165 : Self::persist_tenant_config_at(tenant_shard_id, &config_path, location_conf).await
4166 : }
4167 :
4168 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4169 : pub(super) async fn persist_tenant_config_at(
4170 : tenant_shard_id: &TenantShardId,
4171 : config_path: &Utf8Path,
4172 : location_conf: &LocationConf,
4173 : ) -> std::io::Result<()> {
4174 : debug!("persisting tenantconf to {config_path}");
4175 :
4176 : let mut conf_content = r#"# This file contains a specific per-tenant's config.
4177 : # It is read in case of pageserver restart.
4178 : "#
4179 : .to_string();
4180 :
4181 0 : fail::fail_point!("tenant-config-before-write", |_| {
4182 0 : Err(std::io::Error::new(
4183 0 : std::io::ErrorKind::Other,
4184 0 : "tenant-config-before-write",
4185 0 : ))
4186 0 : });
4187 :
4188 : // Convert the config to a toml file.
4189 : conf_content +=
4190 : &toml_edit::ser::to_string_pretty(&location_conf).expect("Config serialization failed");
4191 :
4192 : let temp_path = path_with_suffix_extension(config_path, TEMP_FILE_SUFFIX);
4193 :
4194 : let conf_content = conf_content.into_bytes();
4195 : VirtualFile::crashsafe_overwrite(config_path.to_owned(), temp_path, conf_content).await
4196 : }
4197 :
4198 : //
4199 : // How garbage collection works:
4200 : //
4201 : // +--bar------------->
4202 : // /
4203 : // +----+-----foo---------------->
4204 : // /
4205 : // ----main--+-------------------------->
4206 : // \
4207 : // +-----baz-------->
4208 : //
4209 : //
4210 : // 1. Grab 'gc_cs' mutex to prevent new timelines from being created while Timeline's
4211 : // `gc_infos` are being refreshed
4212 : // 2. Scan collected timelines, and on each timeline, make note of the
4213 : // all the points where other timelines have been branched off.
4214 : // We will refrain from removing page versions at those LSNs.
4215 : // 3. For each timeline, scan all layer files on the timeline.
4216 : // Remove all files for which a newer file exists and which
4217 : // don't cover any branch point LSNs.
4218 : //
4219 : // TODO:
4220 : // - if a relation has a non-incremental persistent layer on a child branch, then we
4221 : // don't need to keep that in the parent anymore. But currently
4222 : // we do.
4223 8 : async fn gc_iteration_internal(
4224 8 : &self,
4225 8 : target_timeline_id: Option<TimelineId>,
4226 8 : horizon: u64,
4227 8 : pitr: Duration,
4228 8 : cancel: &CancellationToken,
4229 8 : ctx: &RequestContext,
4230 8 : ) -> Result<GcResult, GcError> {
4231 8 : let mut totals: GcResult = Default::default();
4232 8 : let now = Instant::now();
4233 :
4234 8 : let gc_timelines = self
4235 8 : .refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4236 8 : .await?;
4237 :
4238 8 : failpoint_support::sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
4239 :
4240 : // If there is nothing to GC, we don't want any messages in the INFO log.
4241 8 : if !gc_timelines.is_empty() {
4242 8 : info!("{} timelines need GC", gc_timelines.len());
4243 : } else {
4244 0 : debug!("{} timelines need GC", gc_timelines.len());
4245 : }
4246 :
4247 : // Perform GC for each timeline.
4248 : //
4249 : // Note that we don't hold the `Tenant::gc_cs` lock here because we don't want to delay the
4250 : // branch creation task, which requires the GC lock. A GC iteration can run concurrently
4251 : // with branch creation.
4252 : //
4253 : // See comments in [`Tenant::branch_timeline`] for more information about why branch
4254 : // creation task can run concurrently with timeline's GC iteration.
4255 16 : for timeline in gc_timelines {
4256 8 : if cancel.is_cancelled() {
4257 : // We were requested to shut down. Stop and return with the progress we
4258 : // made.
4259 0 : break;
4260 8 : }
4261 8 : let result = match timeline.gc().await {
4262 : Err(GcError::TimelineCancelled) => {
4263 0 : if target_timeline_id.is_some() {
4264 : // If we were targetting this specific timeline, surface cancellation to caller
4265 0 : return Err(GcError::TimelineCancelled);
4266 : } else {
4267 : // A timeline may be shutting down independently of the tenant's lifecycle: we should
4268 : // skip past this and proceed to try GC on other timelines.
4269 0 : continue;
4270 : }
4271 : }
4272 8 : r => r?,
4273 : };
4274 8 : totals += result;
4275 : }
4276 :
4277 8 : totals.elapsed = now.elapsed();
4278 8 : Ok(totals)
4279 8 : }
4280 :
4281 : /// Refreshes the Timeline::gc_info for all timelines, returning the
4282 : /// vector of timelines which have [`Timeline::get_last_record_lsn`] past
4283 : /// [`Tenant::get_gc_horizon`].
4284 : ///
4285 : /// This is usually executed as part of periodic gc, but can now be triggered more often.
4286 0 : pub(crate) async fn refresh_gc_info(
4287 0 : &self,
4288 0 : cancel: &CancellationToken,
4289 0 : ctx: &RequestContext,
4290 0 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4291 0 : // since this method can now be called at different rates than the configured gc loop, it
4292 0 : // might be that these configuration values get applied faster than what it was previously,
4293 0 : // since these were only read from the gc task.
4294 0 : let horizon = self.get_gc_horizon();
4295 0 : let pitr = self.get_pitr_interval();
4296 0 :
4297 0 : // refresh all timelines
4298 0 : let target_timeline_id = None;
4299 0 :
4300 0 : self.refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4301 0 : .await
4302 0 : }
4303 :
4304 : /// Populate all Timelines' `GcInfo` with information about their children. We do not set the
4305 : /// PITR cutoffs here, because that requires I/O: this is done later, before GC, by [`Self::refresh_gc_info_internal`]
4306 : ///
4307 : /// Subsequently, parent-child relationships are updated incrementally inside [`Timeline::new`] and [`Timeline::drop`].
4308 0 : fn initialize_gc_info(
4309 0 : &self,
4310 0 : timelines: &std::sync::MutexGuard<HashMap<TimelineId, Arc<Timeline>>>,
4311 0 : timelines_offloaded: &std::sync::MutexGuard<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
4312 0 : restrict_to_timeline: Option<TimelineId>,
4313 0 : ) {
4314 0 : if restrict_to_timeline.is_none() {
4315 : // This function must be called before activation: after activation timeline create/delete operations
4316 : // might happen, and this function is not safe to run concurrently with those.
4317 0 : assert!(!self.is_active());
4318 0 : }
4319 :
4320 : // Scan all timelines. For each timeline, remember the timeline ID and
4321 : // the branch point where it was created.
4322 0 : let mut all_branchpoints: BTreeMap<TimelineId, Vec<(Lsn, TimelineId, MaybeOffloaded)>> =
4323 0 : BTreeMap::new();
4324 0 : timelines.iter().for_each(|(timeline_id, timeline_entry)| {
4325 0 : if let Some(ancestor_timeline_id) = &timeline_entry.get_ancestor_timeline_id() {
4326 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4327 0 : ancestor_children.push((
4328 0 : timeline_entry.get_ancestor_lsn(),
4329 0 : *timeline_id,
4330 0 : MaybeOffloaded::No,
4331 0 : ));
4332 0 : }
4333 0 : });
4334 0 : timelines_offloaded
4335 0 : .iter()
4336 0 : .for_each(|(timeline_id, timeline_entry)| {
4337 0 : let Some(ancestor_timeline_id) = &timeline_entry.ancestor_timeline_id else {
4338 0 : return;
4339 : };
4340 0 : let Some(retain_lsn) = timeline_entry.ancestor_retain_lsn else {
4341 0 : return;
4342 : };
4343 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4344 0 : ancestor_children.push((retain_lsn, *timeline_id, MaybeOffloaded::Yes));
4345 0 : });
4346 0 :
4347 0 : // The number of bytes we always keep, irrespective of PITR: this is a constant across timelines
4348 0 : let horizon = self.get_gc_horizon();
4349 :
4350 : // Populate each timeline's GcInfo with information about its child branches
4351 0 : let timelines_to_write = if let Some(timeline_id) = restrict_to_timeline {
4352 0 : itertools::Either::Left(timelines.get(&timeline_id).into_iter())
4353 : } else {
4354 0 : itertools::Either::Right(timelines.values())
4355 : };
4356 0 : for timeline in timelines_to_write {
4357 0 : let mut branchpoints: Vec<(Lsn, TimelineId, MaybeOffloaded)> = all_branchpoints
4358 0 : .remove(&timeline.timeline_id)
4359 0 : .unwrap_or_default();
4360 0 :
4361 0 : branchpoints.sort_by_key(|b| b.0);
4362 0 :
4363 0 : let mut target = timeline.gc_info.write().unwrap();
4364 0 :
4365 0 : target.retain_lsns = branchpoints;
4366 0 :
4367 0 : let space_cutoff = timeline
4368 0 : .get_last_record_lsn()
4369 0 : .checked_sub(horizon)
4370 0 : .unwrap_or(Lsn(0));
4371 0 :
4372 0 : target.cutoffs = GcCutoffs {
4373 0 : space: space_cutoff,
4374 0 : time: Lsn::INVALID,
4375 0 : };
4376 0 : }
4377 0 : }
4378 :
4379 8 : async fn refresh_gc_info_internal(
4380 8 : &self,
4381 8 : target_timeline_id: Option<TimelineId>,
4382 8 : horizon: u64,
4383 8 : pitr: Duration,
4384 8 : cancel: &CancellationToken,
4385 8 : ctx: &RequestContext,
4386 8 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4387 8 : // before taking the gc_cs lock, do the heavier weight finding of gc_cutoff points for
4388 8 : // currently visible timelines.
4389 8 : let timelines = self
4390 8 : .timelines
4391 8 : .lock()
4392 8 : .unwrap()
4393 8 : .values()
4394 8 : .filter(|tl| match target_timeline_id.as_ref() {
4395 8 : Some(target) => &tl.timeline_id == target,
4396 0 : None => true,
4397 8 : })
4398 8 : .cloned()
4399 8 : .collect::<Vec<_>>();
4400 8 :
4401 8 : if target_timeline_id.is_some() && timelines.is_empty() {
4402 : // We were to act on a particular timeline and it wasn't found
4403 0 : return Err(GcError::TimelineNotFound);
4404 8 : }
4405 8 :
4406 8 : let mut gc_cutoffs: HashMap<TimelineId, GcCutoffs> =
4407 8 : HashMap::with_capacity(timelines.len());
4408 8 :
4409 8 : // Ensures all timelines use the same start time when computing the time cutoff.
4410 8 : let now_ts_for_pitr_calc = SystemTime::now();
4411 8 : for timeline in timelines.iter() {
4412 8 : let cutoff = timeline
4413 8 : .get_last_record_lsn()
4414 8 : .checked_sub(horizon)
4415 8 : .unwrap_or(Lsn(0));
4416 :
4417 8 : let cutoffs = timeline
4418 8 : .find_gc_cutoffs(now_ts_for_pitr_calc, cutoff, pitr, cancel, ctx)
4419 8 : .await?;
4420 8 : let old = gc_cutoffs.insert(timeline.timeline_id, cutoffs);
4421 8 : assert!(old.is_none());
4422 : }
4423 :
4424 8 : if !self.is_active() || self.cancel.is_cancelled() {
4425 0 : return Err(GcError::TenantCancelled);
4426 8 : }
4427 :
4428 : // grab mutex to prevent new timelines from being created here; avoid doing long operations
4429 : // because that will stall branch creation.
4430 8 : let gc_cs = self.gc_cs.lock().await;
4431 :
4432 : // Ok, we now know all the branch points.
4433 : // Update the GC information for each timeline.
4434 8 : let mut gc_timelines = Vec::with_capacity(timelines.len());
4435 16 : for timeline in timelines {
4436 : // We filtered the timeline list above
4437 8 : if let Some(target_timeline_id) = target_timeline_id {
4438 8 : assert_eq!(target_timeline_id, timeline.timeline_id);
4439 0 : }
4440 :
4441 : {
4442 8 : let mut target = timeline.gc_info.write().unwrap();
4443 8 :
4444 8 : // Cull any expired leases
4445 8 : let now = SystemTime::now();
4446 12 : target.leases.retain(|_, lease| !lease.is_expired(&now));
4447 8 :
4448 8 : timeline
4449 8 : .metrics
4450 8 : .valid_lsn_lease_count_gauge
4451 8 : .set(target.leases.len() as u64);
4452 :
4453 : // Look up parent's PITR cutoff to update the child's knowledge of whether it is within parent's PITR
4454 8 : if let Some(ancestor_id) = timeline.get_ancestor_timeline_id() {
4455 0 : if let Some(ancestor_gc_cutoffs) = gc_cutoffs.get(&ancestor_id) {
4456 0 : target.within_ancestor_pitr =
4457 0 : timeline.get_ancestor_lsn() >= ancestor_gc_cutoffs.time;
4458 0 : }
4459 8 : }
4460 :
4461 : // Update metrics that depend on GC state
4462 8 : timeline
4463 8 : .metrics
4464 8 : .archival_size
4465 8 : .set(if target.within_ancestor_pitr {
4466 0 : timeline.metrics.current_logical_size_gauge.get()
4467 : } else {
4468 8 : 0
4469 : });
4470 8 : timeline.metrics.pitr_history_size.set(
4471 8 : timeline
4472 8 : .get_last_record_lsn()
4473 8 : .checked_sub(target.cutoffs.time)
4474 8 : .unwrap_or(Lsn(0))
4475 8 : .0,
4476 8 : );
4477 :
4478 : // Apply the cutoffs we found to the Timeline's GcInfo. Why might we _not_ have cutoffs for a timeline?
4479 : // - this timeline was created while we were finding cutoffs
4480 : // - lsn for timestamp search fails for this timeline repeatedly
4481 8 : if let Some(cutoffs) = gc_cutoffs.get(&timeline.timeline_id) {
4482 8 : let original_cutoffs = target.cutoffs.clone();
4483 8 : // GC cutoffs should never go back
4484 8 : target.cutoffs = GcCutoffs {
4485 8 : space: Lsn(cutoffs.space.0.max(original_cutoffs.space.0)),
4486 8 : time: Lsn(cutoffs.time.0.max(original_cutoffs.time.0)),
4487 8 : }
4488 0 : }
4489 : }
4490 :
4491 8 : gc_timelines.push(timeline);
4492 : }
4493 8 : drop(gc_cs);
4494 8 : Ok(gc_timelines)
4495 8 : }
4496 :
4497 : /// A substitute for `branch_timeline` for use in unit tests.
4498 : /// The returned timeline will have state value `Active` to make various `anyhow::ensure!()`
4499 : /// calls pass, but, we do not actually call `.activate()` under the hood. So, none of the
4500 : /// timeline background tasks are launched, except the flush loop.
4501 : #[cfg(test)]
4502 464 : async fn branch_timeline_test(
4503 464 : self: &Arc<Self>,
4504 464 : src_timeline: &Arc<Timeline>,
4505 464 : dst_id: TimelineId,
4506 464 : ancestor_lsn: Option<Lsn>,
4507 464 : ctx: &RequestContext,
4508 464 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
4509 464 : let tl = self
4510 464 : .branch_timeline_impl(src_timeline, dst_id, ancestor_lsn, ctx)
4511 464 : .await?
4512 456 : .into_timeline_for_test();
4513 456 : tl.set_state(TimelineState::Active);
4514 456 : Ok(tl)
4515 464 : }
4516 :
4517 : /// Helper for unit tests to branch a timeline with some pre-loaded states.
4518 : #[cfg(test)]
4519 : #[allow(clippy::too_many_arguments)]
4520 12 : pub async fn branch_timeline_test_with_layers(
4521 12 : self: &Arc<Self>,
4522 12 : src_timeline: &Arc<Timeline>,
4523 12 : dst_id: TimelineId,
4524 12 : ancestor_lsn: Option<Lsn>,
4525 12 : ctx: &RequestContext,
4526 12 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
4527 12 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
4528 12 : end_lsn: Lsn,
4529 12 : ) -> anyhow::Result<Arc<Timeline>> {
4530 : use checks::check_valid_layermap;
4531 : use itertools::Itertools;
4532 :
4533 12 : let tline = self
4534 12 : .branch_timeline_test(src_timeline, dst_id, ancestor_lsn, ctx)
4535 12 : .await?;
4536 12 : let ancestor_lsn = if let Some(ancestor_lsn) = ancestor_lsn {
4537 12 : ancestor_lsn
4538 : } else {
4539 0 : tline.get_last_record_lsn()
4540 : };
4541 12 : assert!(end_lsn >= ancestor_lsn);
4542 12 : tline.force_advance_lsn(end_lsn);
4543 24 : for deltas in delta_layer_desc {
4544 12 : tline
4545 12 : .force_create_delta_layer(deltas, Some(ancestor_lsn), ctx)
4546 12 : .await?;
4547 : }
4548 20 : for (lsn, images) in image_layer_desc {
4549 8 : tline
4550 8 : .force_create_image_layer(lsn, images, Some(ancestor_lsn), ctx)
4551 8 : .await?;
4552 : }
4553 12 : let layer_names = tline
4554 12 : .layers
4555 12 : .read()
4556 12 : .await
4557 12 : .layer_map()
4558 12 : .unwrap()
4559 12 : .iter_historic_layers()
4560 20 : .map(|layer| layer.layer_name())
4561 12 : .collect_vec();
4562 12 : if let Some(err) = check_valid_layermap(&layer_names) {
4563 0 : bail!("invalid layermap: {err}");
4564 12 : }
4565 12 : Ok(tline)
4566 12 : }
4567 :
4568 : /// Branch an existing timeline.
4569 0 : async fn branch_timeline(
4570 0 : self: &Arc<Self>,
4571 0 : src_timeline: &Arc<Timeline>,
4572 0 : dst_id: TimelineId,
4573 0 : start_lsn: Option<Lsn>,
4574 0 : ctx: &RequestContext,
4575 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4576 0 : self.branch_timeline_impl(src_timeline, dst_id, start_lsn, ctx)
4577 0 : .await
4578 0 : }
4579 :
4580 464 : async fn branch_timeline_impl(
4581 464 : self: &Arc<Self>,
4582 464 : src_timeline: &Arc<Timeline>,
4583 464 : dst_id: TimelineId,
4584 464 : start_lsn: Option<Lsn>,
4585 464 : _ctx: &RequestContext,
4586 464 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4587 464 : let src_id = src_timeline.timeline_id;
4588 :
4589 : // We will validate our ancestor LSN in this function. Acquire the GC lock so that
4590 : // this check cannot race with GC, and the ancestor LSN is guaranteed to remain
4591 : // valid while we are creating the branch.
4592 464 : let _gc_cs = self.gc_cs.lock().await;
4593 :
4594 : // If no start LSN is specified, we branch the new timeline from the source timeline's last record LSN
4595 464 : let start_lsn = start_lsn.unwrap_or_else(|| {
4596 4 : let lsn = src_timeline.get_last_record_lsn();
4597 4 : info!("branching timeline {dst_id} from timeline {src_id} at last record LSN: {lsn}");
4598 4 : lsn
4599 464 : });
4600 :
4601 : // we finally have determined the ancestor_start_lsn, so we can get claim exclusivity now
4602 464 : let timeline_create_guard = match self
4603 464 : .start_creating_timeline(
4604 464 : dst_id,
4605 464 : CreateTimelineIdempotency::Branch {
4606 464 : ancestor_timeline_id: src_timeline.timeline_id,
4607 464 : ancestor_start_lsn: start_lsn,
4608 464 : },
4609 464 : )
4610 464 : .await?
4611 : {
4612 464 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
4613 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
4614 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
4615 : }
4616 : };
4617 :
4618 : // Ensure that `start_lsn` is valid, i.e. the LSN is within the PITR
4619 : // horizon on the source timeline
4620 : //
4621 : // We check it against both the planned GC cutoff stored in 'gc_info',
4622 : // and the 'latest_gc_cutoff' of the last GC that was performed. The
4623 : // planned GC cutoff in 'gc_info' is normally larger than
4624 : // 'latest_gc_cutoff_lsn', but beware of corner cases like if you just
4625 : // changed the GC settings for the tenant to make the PITR window
4626 : // larger, but some of the data was already removed by an earlier GC
4627 : // iteration.
4628 :
4629 : // check against last actual 'latest_gc_cutoff' first
4630 464 : let latest_gc_cutoff_lsn = src_timeline.get_latest_gc_cutoff_lsn();
4631 464 : src_timeline
4632 464 : .check_lsn_is_in_scope(start_lsn, &latest_gc_cutoff_lsn)
4633 464 : .context(format!(
4634 464 : "invalid branch start lsn: less than latest GC cutoff {}",
4635 464 : *latest_gc_cutoff_lsn,
4636 464 : ))
4637 464 : .map_err(CreateTimelineError::AncestorLsn)?;
4638 :
4639 : // and then the planned GC cutoff
4640 : {
4641 456 : let gc_info = src_timeline.gc_info.read().unwrap();
4642 456 : let cutoff = gc_info.min_cutoff();
4643 456 : if start_lsn < cutoff {
4644 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
4645 0 : "invalid branch start lsn: less than planned GC cutoff {cutoff}"
4646 0 : )));
4647 456 : }
4648 456 : }
4649 456 :
4650 456 : //
4651 456 : // The branch point is valid, and we are still holding the 'gc_cs' lock
4652 456 : // so that GC cannot advance the GC cutoff until we are finished.
4653 456 : // Proceed with the branch creation.
4654 456 : //
4655 456 :
4656 456 : // Determine prev-LSN for the new timeline. We can only determine it if
4657 456 : // the timeline was branched at the current end of the source timeline.
4658 456 : let RecordLsn {
4659 456 : last: src_last,
4660 456 : prev: src_prev,
4661 456 : } = src_timeline.get_last_record_rlsn();
4662 456 : let dst_prev = if src_last == start_lsn {
4663 432 : Some(src_prev)
4664 : } else {
4665 24 : None
4666 : };
4667 :
4668 : // Create the metadata file, noting the ancestor of the new timeline.
4669 : // There is initially no data in it, but all the read-calls know to look
4670 : // into the ancestor.
4671 456 : let metadata = TimelineMetadata::new(
4672 456 : start_lsn,
4673 456 : dst_prev,
4674 456 : Some(src_id),
4675 456 : start_lsn,
4676 456 : *src_timeline.latest_gc_cutoff_lsn.read(), // FIXME: should we hold onto this guard longer?
4677 456 : src_timeline.initdb_lsn,
4678 456 : src_timeline.pg_version,
4679 456 : );
4680 :
4681 456 : let uninitialized_timeline = self
4682 456 : .prepare_new_timeline(
4683 456 : dst_id,
4684 456 : &metadata,
4685 456 : timeline_create_guard,
4686 456 : start_lsn + 1,
4687 456 : Some(Arc::clone(src_timeline)),
4688 456 : )
4689 456 : .await?;
4690 :
4691 456 : let new_timeline = uninitialized_timeline.finish_creation()?;
4692 :
4693 : // Root timeline gets its layers during creation and uploads them along with the metadata.
4694 : // A branch timeline though, when created, can get no writes for some time, hence won't get any layers created.
4695 : // We still need to upload its metadata eagerly: if other nodes `attach` the tenant and miss this timeline, their GC
4696 : // could get incorrect information and remove more layers, than needed.
4697 : // See also https://github.com/neondatabase/neon/issues/3865
4698 456 : new_timeline
4699 456 : .remote_client
4700 456 : .schedule_index_upload_for_full_metadata_update(&metadata)
4701 456 : .context("branch initial metadata upload")?;
4702 :
4703 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
4704 :
4705 456 : Ok(CreateTimelineResult::Created(new_timeline))
4706 464 : }
4707 :
4708 : /// For unit tests, make this visible so that other modules can directly create timelines
4709 : #[cfg(test)]
4710 : #[tracing::instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
4711 : pub(crate) async fn bootstrap_timeline_test(
4712 : self: &Arc<Self>,
4713 : timeline_id: TimelineId,
4714 : pg_version: u32,
4715 : load_existing_initdb: Option<TimelineId>,
4716 : ctx: &RequestContext,
4717 : ) -> anyhow::Result<Arc<Timeline>> {
4718 : self.bootstrap_timeline(timeline_id, pg_version, load_existing_initdb, ctx)
4719 : .await
4720 : .map_err(anyhow::Error::new)
4721 4 : .map(|r| r.into_timeline_for_test())
4722 : }
4723 :
4724 : /// Get exclusive access to the timeline ID for creation.
4725 : ///
4726 : /// Timeline-creating code paths must use this function before making changes
4727 : /// to in-memory or persistent state.
4728 : ///
4729 : /// The `state` parameter is a description of the timeline creation operation
4730 : /// we intend to perform.
4731 : /// If the timeline was already created in the meantime, we check whether this
4732 : /// request conflicts or is idempotent , based on `state`.
4733 892 : async fn start_creating_timeline(
4734 892 : self: &Arc<Self>,
4735 892 : new_timeline_id: TimelineId,
4736 892 : idempotency: CreateTimelineIdempotency,
4737 892 : ) -> Result<StartCreatingTimelineResult, CreateTimelineError> {
4738 892 : let allow_offloaded = false;
4739 892 : match self.create_timeline_create_guard(new_timeline_id, idempotency, allow_offloaded) {
4740 888 : Ok(create_guard) => {
4741 888 : pausable_failpoint!("timeline-creation-after-uninit");
4742 888 : Ok(StartCreatingTimelineResult::CreateGuard(create_guard))
4743 : }
4744 0 : Err(TimelineExclusionError::ShuttingDown) => Err(CreateTimelineError::ShuttingDown),
4745 : Err(TimelineExclusionError::AlreadyCreating) => {
4746 : // Creation is in progress, we cannot create it again, and we cannot
4747 : // check if this request matches the existing one, so caller must try
4748 : // again later.
4749 0 : Err(CreateTimelineError::AlreadyCreating)
4750 : }
4751 0 : Err(TimelineExclusionError::Other(e)) => Err(CreateTimelineError::Other(e)),
4752 : Err(TimelineExclusionError::AlreadyExists {
4753 0 : existing: TimelineOrOffloaded::Offloaded(_existing),
4754 0 : ..
4755 0 : }) => {
4756 0 : info!("timeline already exists but is offloaded");
4757 0 : Err(CreateTimelineError::Conflict)
4758 : }
4759 : Err(TimelineExclusionError::AlreadyExists {
4760 4 : existing: TimelineOrOffloaded::Timeline(existing),
4761 4 : arg,
4762 4 : }) => {
4763 4 : {
4764 4 : let existing = &existing.create_idempotency;
4765 4 : let _span = info_span!("idempotency_check", ?existing, ?arg).entered();
4766 4 : debug!("timeline already exists");
4767 :
4768 4 : match (existing, &arg) {
4769 : // FailWithConflict => no idempotency check
4770 : (CreateTimelineIdempotency::FailWithConflict, _)
4771 : | (_, CreateTimelineIdempotency::FailWithConflict) => {
4772 4 : warn!("timeline already exists, failing request");
4773 4 : return Err(CreateTimelineError::Conflict);
4774 : }
4775 : // Idempotent <=> CreateTimelineIdempotency is identical
4776 0 : (x, y) if x == y => {
4777 0 : info!("timeline already exists and idempotency matches, succeeding request");
4778 : // fallthrough
4779 : }
4780 : (_, _) => {
4781 0 : warn!("idempotency conflict, failing request");
4782 0 : return Err(CreateTimelineError::Conflict);
4783 : }
4784 : }
4785 : }
4786 :
4787 0 : Ok(StartCreatingTimelineResult::Idempotent(existing))
4788 : }
4789 : }
4790 892 : }
4791 :
4792 0 : async fn upload_initdb(
4793 0 : &self,
4794 0 : timelines_path: &Utf8PathBuf,
4795 0 : pgdata_path: &Utf8PathBuf,
4796 0 : timeline_id: &TimelineId,
4797 0 : ) -> anyhow::Result<()> {
4798 0 : let temp_path = timelines_path.join(format!(
4799 0 : "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
4800 0 : ));
4801 0 :
4802 0 : scopeguard::defer! {
4803 0 : if let Err(e) = fs::remove_file(&temp_path) {
4804 0 : error!("Failed to remove temporary initdb archive '{temp_path}': {e}");
4805 0 : }
4806 0 : }
4807 :
4808 0 : let (pgdata_zstd, tar_zst_size) = create_zst_tarball(pgdata_path, &temp_path).await?;
4809 : const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
4810 0 : if tar_zst_size > INITDB_TAR_ZST_WARN_LIMIT {
4811 0 : warn!(
4812 0 : "compressed {temp_path} size of {tar_zst_size} is above limit {INITDB_TAR_ZST_WARN_LIMIT}."
4813 : );
4814 0 : }
4815 :
4816 0 : pausable_failpoint!("before-initdb-upload");
4817 :
4818 0 : backoff::retry(
4819 0 : || async {
4820 0 : self::remote_timeline_client::upload_initdb_dir(
4821 0 : &self.remote_storage,
4822 0 : &self.tenant_shard_id.tenant_id,
4823 0 : timeline_id,
4824 0 : pgdata_zstd.try_clone().await?,
4825 0 : tar_zst_size,
4826 0 : &self.cancel,
4827 0 : )
4828 0 : .await
4829 0 : },
4830 0 : |_| false,
4831 0 : 3,
4832 0 : u32::MAX,
4833 0 : "persist_initdb_tar_zst",
4834 0 : &self.cancel,
4835 0 : )
4836 0 : .await
4837 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
4838 0 : .and_then(|x| x)
4839 0 : }
4840 :
4841 : /// - run initdb to init temporary instance and get bootstrap data
4842 : /// - after initialization completes, tar up the temp dir and upload it to S3.
4843 4 : async fn bootstrap_timeline(
4844 4 : self: &Arc<Self>,
4845 4 : timeline_id: TimelineId,
4846 4 : pg_version: u32,
4847 4 : load_existing_initdb: Option<TimelineId>,
4848 4 : ctx: &RequestContext,
4849 4 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4850 4 : let timeline_create_guard = match self
4851 4 : .start_creating_timeline(
4852 4 : timeline_id,
4853 4 : CreateTimelineIdempotency::Bootstrap { pg_version },
4854 4 : )
4855 4 : .await?
4856 : {
4857 4 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
4858 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
4859 0 : return Ok(CreateTimelineResult::Idempotent(timeline))
4860 : }
4861 : };
4862 :
4863 : // create a `tenant/{tenant_id}/timelines/basebackup-{timeline_id}.{TEMP_FILE_SUFFIX}/`
4864 : // temporary directory for basebackup files for the given timeline.
4865 :
4866 4 : let timelines_path = self.conf.timelines_path(&self.tenant_shard_id);
4867 4 : let pgdata_path = path_with_suffix_extension(
4868 4 : timelines_path.join(format!("basebackup-{timeline_id}")),
4869 4 : TEMP_FILE_SUFFIX,
4870 4 : );
4871 4 :
4872 4 : // Remove whatever was left from the previous runs: safe because TimelineCreateGuard guarantees
4873 4 : // we won't race with other creations or existent timelines with the same path.
4874 4 : if pgdata_path.exists() {
4875 0 : fs::remove_dir_all(&pgdata_path).with_context(|| {
4876 0 : format!("Failed to remove already existing initdb directory: {pgdata_path}")
4877 0 : })?;
4878 4 : }
4879 :
4880 : // this new directory is very temporary, set to remove it immediately after bootstrap, we don't need it
4881 4 : scopeguard::defer! {
4882 4 : if let Err(e) = fs::remove_dir_all(&pgdata_path) {
4883 4 : // this is unlikely, but we will remove the directory on pageserver restart or another bootstrap call
4884 4 : error!("Failed to remove temporary initdb directory '{pgdata_path}': {e}");
4885 4 : }
4886 4 : }
4887 4 : if let Some(existing_initdb_timeline_id) = load_existing_initdb {
4888 4 : if existing_initdb_timeline_id != timeline_id {
4889 0 : let source_path = &remote_initdb_archive_path(
4890 0 : &self.tenant_shard_id.tenant_id,
4891 0 : &existing_initdb_timeline_id,
4892 0 : );
4893 0 : let dest_path =
4894 0 : &remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &timeline_id);
4895 0 :
4896 0 : // if this fails, it will get retried by retried control plane requests
4897 0 : self.remote_storage
4898 0 : .copy_object(source_path, dest_path, &self.cancel)
4899 0 : .await
4900 0 : .context("copy initdb tar")?;
4901 4 : }
4902 4 : let (initdb_tar_zst_path, initdb_tar_zst) =
4903 4 : self::remote_timeline_client::download_initdb_tar_zst(
4904 4 : self.conf,
4905 4 : &self.remote_storage,
4906 4 : &self.tenant_shard_id,
4907 4 : &existing_initdb_timeline_id,
4908 4 : &self.cancel,
4909 4 : )
4910 4 : .await
4911 4 : .context("download initdb tar")?;
4912 :
4913 4 : scopeguard::defer! {
4914 4 : if let Err(e) = fs::remove_file(&initdb_tar_zst_path) {
4915 4 : error!("Failed to remove temporary initdb archive '{initdb_tar_zst_path}': {e}");
4916 4 : }
4917 4 : }
4918 4 :
4919 4 : let buf_read =
4920 4 : BufReader::with_capacity(remote_timeline_client::BUFFER_SIZE, initdb_tar_zst);
4921 4 : extract_zst_tarball(&pgdata_path, buf_read)
4922 4 : .await
4923 4 : .context("extract initdb tar")?;
4924 : } else {
4925 : // Init temporarily repo to get bootstrap data, this creates a directory in the `pgdata_path` path
4926 0 : run_initdb(self.conf, &pgdata_path, pg_version, &self.cancel)
4927 0 : .await
4928 0 : .context("run initdb")?;
4929 :
4930 : // Upload the created data dir to S3
4931 0 : if self.tenant_shard_id().is_shard_zero() {
4932 0 : self.upload_initdb(&timelines_path, &pgdata_path, &timeline_id)
4933 0 : .await?;
4934 0 : }
4935 : }
4936 4 : let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
4937 4 :
4938 4 : // Import the contents of the data directory at the initial checkpoint
4939 4 : // LSN, and any WAL after that.
4940 4 : // Initdb lsn will be equal to last_record_lsn which will be set after import.
4941 4 : // Because we know it upfront avoid having an option or dummy zero value by passing it to the metadata.
4942 4 : let new_metadata = TimelineMetadata::new(
4943 4 : Lsn(0),
4944 4 : None,
4945 4 : None,
4946 4 : Lsn(0),
4947 4 : pgdata_lsn,
4948 4 : pgdata_lsn,
4949 4 : pg_version,
4950 4 : );
4951 4 : let raw_timeline = self
4952 4 : .prepare_new_timeline(
4953 4 : timeline_id,
4954 4 : &new_metadata,
4955 4 : timeline_create_guard,
4956 4 : pgdata_lsn,
4957 4 : None,
4958 4 : )
4959 4 : .await?;
4960 :
4961 4 : let tenant_shard_id = raw_timeline.owning_tenant.tenant_shard_id;
4962 4 : let unfinished_timeline = raw_timeline.raw_timeline()?;
4963 :
4964 : // Flush the new layer files to disk, before we make the timeline as available to
4965 : // the outside world.
4966 : //
4967 : // Flush loop needs to be spawned in order to be able to flush.
4968 4 : unfinished_timeline.maybe_spawn_flush_loop();
4969 4 :
4970 4 : import_datadir::import_timeline_from_postgres_datadir(
4971 4 : unfinished_timeline,
4972 4 : &pgdata_path,
4973 4 : pgdata_lsn,
4974 4 : ctx,
4975 4 : )
4976 4 : .await
4977 4 : .with_context(|| {
4978 0 : format!("Failed to import pgdatadir for timeline {tenant_shard_id}/{timeline_id}")
4979 4 : })?;
4980 :
4981 4 : fail::fail_point!("before-checkpoint-new-timeline", |_| {
4982 0 : Err(CreateTimelineError::Other(anyhow::anyhow!(
4983 0 : "failpoint before-checkpoint-new-timeline"
4984 0 : )))
4985 4 : });
4986 :
4987 4 : unfinished_timeline
4988 4 : .freeze_and_flush()
4989 4 : .await
4990 4 : .with_context(|| {
4991 0 : format!(
4992 0 : "Failed to flush after pgdatadir import for timeline {tenant_shard_id}/{timeline_id}"
4993 0 : )
4994 4 : })?;
4995 :
4996 : // All done!
4997 4 : let timeline = raw_timeline.finish_creation()?;
4998 :
4999 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
5000 :
5001 4 : Ok(CreateTimelineResult::Created(timeline))
5002 4 : }
5003 :
5004 880 : fn build_timeline_remote_client(&self, timeline_id: TimelineId) -> RemoteTimelineClient {
5005 880 : RemoteTimelineClient::new(
5006 880 : self.remote_storage.clone(),
5007 880 : self.deletion_queue_client.clone(),
5008 880 : self.conf,
5009 880 : self.tenant_shard_id,
5010 880 : timeline_id,
5011 880 : self.generation,
5012 880 : &self.tenant_conf.load().location,
5013 880 : )
5014 880 : }
5015 :
5016 : /// Call this before constructing a timeline, to build its required structures
5017 880 : fn build_timeline_resources(&self, timeline_id: TimelineId) -> TimelineResources {
5018 880 : TimelineResources {
5019 880 : remote_client: self.build_timeline_remote_client(timeline_id),
5020 880 : pagestream_throttle: self.pagestream_throttle.clone(),
5021 880 : pagestream_throttle_metrics: self.pagestream_throttle_metrics.clone(),
5022 880 : l0_flush_global_state: self.l0_flush_global_state.clone(),
5023 880 : }
5024 880 : }
5025 :
5026 : /// Creates intermediate timeline structure and its files.
5027 : ///
5028 : /// An empty layer map is initialized, and new data and WAL can be imported starting
5029 : /// at 'disk_consistent_lsn'. After any initial data has been imported, call
5030 : /// `finish_creation` to insert the Timeline into the timelines map.
5031 880 : async fn prepare_new_timeline<'a>(
5032 880 : &'a self,
5033 880 : new_timeline_id: TimelineId,
5034 880 : new_metadata: &TimelineMetadata,
5035 880 : create_guard: TimelineCreateGuard,
5036 880 : start_lsn: Lsn,
5037 880 : ancestor: Option<Arc<Timeline>>,
5038 880 : ) -> anyhow::Result<UninitializedTimeline<'a>> {
5039 880 : let tenant_shard_id = self.tenant_shard_id;
5040 880 :
5041 880 : let resources = self.build_timeline_resources(new_timeline_id);
5042 880 : resources
5043 880 : .remote_client
5044 880 : .init_upload_queue_for_empty_remote(new_metadata)?;
5045 :
5046 880 : let timeline_struct = self
5047 880 : .create_timeline_struct(
5048 880 : new_timeline_id,
5049 880 : new_metadata,
5050 880 : ancestor,
5051 880 : resources,
5052 880 : CreateTimelineCause::Load,
5053 880 : create_guard.idempotency.clone(),
5054 880 : )
5055 880 : .context("Failed to create timeline data structure")?;
5056 :
5057 880 : timeline_struct.init_empty_layer_map(start_lsn);
5058 :
5059 880 : if let Err(e) = self
5060 880 : .create_timeline_files(&create_guard.timeline_path)
5061 880 : .await
5062 : {
5063 0 : error!("Failed to create initial files for timeline {tenant_shard_id}/{new_timeline_id}, cleaning up: {e:?}");
5064 0 : cleanup_timeline_directory(create_guard);
5065 0 : return Err(e);
5066 880 : }
5067 880 :
5068 880 : debug!(
5069 0 : "Successfully created initial files for timeline {tenant_shard_id}/{new_timeline_id}"
5070 : );
5071 :
5072 880 : Ok(UninitializedTimeline::new(
5073 880 : self,
5074 880 : new_timeline_id,
5075 880 : Some((timeline_struct, create_guard)),
5076 880 : ))
5077 880 : }
5078 :
5079 880 : async fn create_timeline_files(&self, timeline_path: &Utf8Path) -> anyhow::Result<()> {
5080 880 : crashsafe::create_dir(timeline_path).context("Failed to create timeline directory")?;
5081 :
5082 880 : fail::fail_point!("after-timeline-dir-creation", |_| {
5083 0 : anyhow::bail!("failpoint after-timeline-dir-creation");
5084 880 : });
5085 :
5086 880 : Ok(())
5087 880 : }
5088 :
5089 : /// Get a guard that provides exclusive access to the timeline directory, preventing
5090 : /// concurrent attempts to create the same timeline.
5091 : ///
5092 : /// The `allow_offloaded` parameter controls whether to tolerate the existence of
5093 : /// offloaded timelines or not.
5094 892 : fn create_timeline_create_guard(
5095 892 : self: &Arc<Self>,
5096 892 : timeline_id: TimelineId,
5097 892 : idempotency: CreateTimelineIdempotency,
5098 892 : allow_offloaded: bool,
5099 892 : ) -> Result<TimelineCreateGuard, TimelineExclusionError> {
5100 892 : let tenant_shard_id = self.tenant_shard_id;
5101 892 :
5102 892 : let timeline_path = self.conf.timeline_path(&tenant_shard_id, &timeline_id);
5103 :
5104 892 : let create_guard = TimelineCreateGuard::new(
5105 892 : self,
5106 892 : timeline_id,
5107 892 : timeline_path.clone(),
5108 892 : idempotency,
5109 892 : allow_offloaded,
5110 892 : )?;
5111 :
5112 : // At this stage, we have got exclusive access to in-memory state for this timeline ID
5113 : // for creation.
5114 : // A timeline directory should never exist on disk already:
5115 : // - a previous failed creation would have cleaned up after itself
5116 : // - a pageserver restart would clean up timeline directories that don't have valid remote state
5117 : //
5118 : // Therefore it is an unexpected internal error to encounter a timeline directory already existing here,
5119 : // this error may indicate a bug in cleanup on failed creations.
5120 888 : if timeline_path.exists() {
5121 0 : return Err(TimelineExclusionError::Other(anyhow::anyhow!(
5122 0 : "Timeline directory already exists! This is a bug."
5123 0 : )));
5124 888 : }
5125 888 :
5126 888 : Ok(create_guard)
5127 892 : }
5128 :
5129 : /// Gathers inputs from all of the timelines to produce a sizing model input.
5130 : ///
5131 : /// Future is cancellation safe. Only one calculation can be running at once per tenant.
5132 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5133 : pub async fn gather_size_inputs(
5134 : &self,
5135 : // `max_retention_period` overrides the cutoff that is used to calculate the size
5136 : // (only if it is shorter than the real cutoff).
5137 : max_retention_period: Option<u64>,
5138 : cause: LogicalSizeCalculationCause,
5139 : cancel: &CancellationToken,
5140 : ctx: &RequestContext,
5141 : ) -> Result<size::ModelInputs, size::CalculateSyntheticSizeError> {
5142 : let logical_sizes_at_once = self
5143 : .conf
5144 : .concurrent_tenant_size_logical_size_queries
5145 : .inner();
5146 :
5147 : // TODO: Having a single mutex block concurrent reads is not great for performance.
5148 : //
5149 : // But the only case where we need to run multiple of these at once is when we
5150 : // request a size for a tenant manually via API, while another background calculation
5151 : // is in progress (which is not a common case).
5152 : //
5153 : // See more for on the issue #2748 condenced out of the initial PR review.
5154 : let mut shared_cache = tokio::select! {
5155 : locked = self.cached_logical_sizes.lock() => locked,
5156 : _ = cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5157 : _ = self.cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5158 : };
5159 :
5160 : size::gather_inputs(
5161 : self,
5162 : logical_sizes_at_once,
5163 : max_retention_period,
5164 : &mut shared_cache,
5165 : cause,
5166 : cancel,
5167 : ctx,
5168 : )
5169 : .await
5170 : }
5171 :
5172 : /// Calculate synthetic tenant size and cache the result.
5173 : /// This is periodically called by background worker.
5174 : /// result is cached in tenant struct
5175 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5176 : pub async fn calculate_synthetic_size(
5177 : &self,
5178 : cause: LogicalSizeCalculationCause,
5179 : cancel: &CancellationToken,
5180 : ctx: &RequestContext,
5181 : ) -> Result<u64, size::CalculateSyntheticSizeError> {
5182 : let inputs = self.gather_size_inputs(None, cause, cancel, ctx).await?;
5183 :
5184 : let size = inputs.calculate();
5185 :
5186 : self.set_cached_synthetic_size(size);
5187 :
5188 : Ok(size)
5189 : }
5190 :
5191 : /// Cache given synthetic size and update the metric value
5192 0 : pub fn set_cached_synthetic_size(&self, size: u64) {
5193 0 : self.cached_synthetic_tenant_size
5194 0 : .store(size, Ordering::Relaxed);
5195 0 :
5196 0 : // Only shard zero should be calculating synthetic sizes
5197 0 : debug_assert!(self.shard_identity.is_shard_zero());
5198 :
5199 0 : TENANT_SYNTHETIC_SIZE_METRIC
5200 0 : .get_metric_with_label_values(&[&self.tenant_shard_id.tenant_id.to_string()])
5201 0 : .unwrap()
5202 0 : .set(size);
5203 0 : }
5204 :
5205 0 : pub fn cached_synthetic_size(&self) -> u64 {
5206 0 : self.cached_synthetic_tenant_size.load(Ordering::Relaxed)
5207 0 : }
5208 :
5209 : /// Flush any in-progress layers, schedule uploads, and wait for uploads to complete.
5210 : ///
5211 : /// This function can take a long time: callers should wrap it in a timeout if calling
5212 : /// from an external API handler.
5213 : ///
5214 : /// Cancel-safety: cancelling this function may leave I/O running, but such I/O is
5215 : /// still bounded by tenant/timeline shutdown.
5216 : #[tracing::instrument(skip_all)]
5217 : pub(crate) async fn flush_remote(&self) -> anyhow::Result<()> {
5218 : let timelines = self.timelines.lock().unwrap().clone();
5219 :
5220 0 : async fn flush_timeline(_gate: GateGuard, timeline: Arc<Timeline>) -> anyhow::Result<()> {
5221 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Flushing...");
5222 0 : timeline.freeze_and_flush().await?;
5223 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Waiting for uploads...");
5224 0 : timeline.remote_client.wait_completion().await?;
5225 :
5226 0 : Ok(())
5227 0 : }
5228 :
5229 : // We do not use a JoinSet for these tasks, because we don't want them to be
5230 : // aborted when this function's future is cancelled: they should stay alive
5231 : // holding their GateGuard until they complete, to ensure their I/Os complete
5232 : // before Timeline shutdown completes.
5233 : let mut results = FuturesUnordered::new();
5234 :
5235 : for (_timeline_id, timeline) in timelines {
5236 : // Run each timeline's flush in a task holding the timeline's gate: this
5237 : // means that if this function's future is cancelled, the Timeline shutdown
5238 : // will still wait for any I/O in here to complete.
5239 : let Ok(gate) = timeline.gate.enter() else {
5240 : continue;
5241 : };
5242 0 : let jh = tokio::task::spawn(async move { flush_timeline(gate, timeline).await });
5243 : results.push(jh);
5244 : }
5245 :
5246 : while let Some(r) = results.next().await {
5247 : if let Err(e) = r {
5248 : if !e.is_cancelled() && !e.is_panic() {
5249 : tracing::error!("unexpected join error: {e:?}");
5250 : }
5251 : }
5252 : }
5253 :
5254 : // The flushes we did above were just writes, but the Tenant might have had
5255 : // pending deletions as well from recent compaction/gc: we want to flush those
5256 : // as well. This requires flushing the global delete queue. This is cheap
5257 : // because it's typically a no-op.
5258 : match self.deletion_queue_client.flush_execute().await {
5259 : Ok(_) => {}
5260 : Err(DeletionQueueError::ShuttingDown) => {}
5261 : }
5262 :
5263 : Ok(())
5264 : }
5265 :
5266 0 : pub(crate) fn get_tenant_conf(&self) -> TenantConfOpt {
5267 0 : self.tenant_conf.load().tenant_conf.clone()
5268 0 : }
5269 :
5270 : /// How much local storage would this tenant like to have? It can cope with
5271 : /// less than this (via eviction and on-demand downloads), but this function enables
5272 : /// the Tenant to advertise how much storage it would prefer to have to provide fast I/O
5273 : /// by keeping important things on local disk.
5274 : ///
5275 : /// This is a heuristic, not a guarantee: tenants that are long-idle will actually use less
5276 : /// than they report here, due to layer eviction. Tenants with many active branches may
5277 : /// actually use more than they report here.
5278 0 : pub(crate) fn local_storage_wanted(&self) -> u64 {
5279 0 : let timelines = self.timelines.lock().unwrap();
5280 0 :
5281 0 : // Heuristic: we use the max() of the timelines' visible sizes, rather than the sum. This
5282 0 : // reflects the observation that on tenants with multiple large branches, typically only one
5283 0 : // of them is used actively enough to occupy space on disk.
5284 0 : timelines
5285 0 : .values()
5286 0 : .map(|t| t.metrics.visible_physical_size_gauge.get())
5287 0 : .max()
5288 0 : .unwrap_or(0)
5289 0 : }
5290 :
5291 : /// Serialize and write the latest TenantManifest to remote storage.
5292 4 : pub(crate) async fn store_tenant_manifest(&self) -> Result<(), TenantManifestError> {
5293 : // Only one manifest write may be done at at time, and the contents of the manifest
5294 : // must be loaded while holding this lock. This makes it safe to call this function
5295 : // from anywhere without worrying about colliding updates.
5296 4 : let mut guard = tokio::select! {
5297 4 : g = self.tenant_manifest_upload.lock() => {
5298 4 : g
5299 : },
5300 4 : _ = self.cancel.cancelled() => {
5301 0 : return Err(TenantManifestError::Cancelled);
5302 : }
5303 : };
5304 :
5305 4 : let manifest = self.build_tenant_manifest();
5306 4 : if Some(&manifest) == (*guard).as_ref() {
5307 : // Optimisation: skip uploads that don't change anything.
5308 0 : return Ok(());
5309 4 : }
5310 4 :
5311 4 : upload_tenant_manifest(
5312 4 : &self.remote_storage,
5313 4 : &self.tenant_shard_id,
5314 4 : self.generation,
5315 4 : &manifest,
5316 4 : &self.cancel,
5317 4 : )
5318 4 : .await
5319 4 : .map_err(|e| {
5320 0 : if self.cancel.is_cancelled() {
5321 0 : TenantManifestError::Cancelled
5322 : } else {
5323 0 : TenantManifestError::RemoteStorage(e)
5324 : }
5325 4 : })?;
5326 :
5327 : // Store the successfully uploaded manifest, so that future callers can avoid
5328 : // re-uploading the same thing.
5329 4 : *guard = Some(manifest);
5330 4 :
5331 4 : Ok(())
5332 4 : }
5333 : }
5334 :
5335 : /// Create the cluster temporarily in 'initdbpath' directory inside the repository
5336 : /// to get bootstrap data for timeline initialization.
5337 0 : async fn run_initdb(
5338 0 : conf: &'static PageServerConf,
5339 0 : initdb_target_dir: &Utf8Path,
5340 0 : pg_version: u32,
5341 0 : cancel: &CancellationToken,
5342 0 : ) -> Result<(), InitdbError> {
5343 0 : let initdb_bin_path = conf
5344 0 : .pg_bin_dir(pg_version)
5345 0 : .map_err(InitdbError::Other)?
5346 0 : .join("initdb");
5347 0 : let initdb_lib_dir = conf.pg_lib_dir(pg_version).map_err(InitdbError::Other)?;
5348 0 : info!(
5349 0 : "running {} in {}, libdir: {}",
5350 : initdb_bin_path, initdb_target_dir, initdb_lib_dir,
5351 : );
5352 :
5353 0 : let _permit = {
5354 0 : let _timer = INITDB_SEMAPHORE_ACQUISITION_TIME.start_timer();
5355 0 : INIT_DB_SEMAPHORE.acquire().await
5356 : };
5357 :
5358 0 : CONCURRENT_INITDBS.inc();
5359 0 : scopeguard::defer! {
5360 0 : CONCURRENT_INITDBS.dec();
5361 0 : }
5362 0 :
5363 0 : let _timer = INITDB_RUN_TIME.start_timer();
5364 0 : let res = postgres_initdb::do_run_initdb(postgres_initdb::RunInitdbArgs {
5365 0 : superuser: &conf.superuser,
5366 0 : locale: &conf.locale,
5367 0 : initdb_bin: &initdb_bin_path,
5368 0 : pg_version,
5369 0 : library_search_path: &initdb_lib_dir,
5370 0 : pgdata: initdb_target_dir,
5371 0 : })
5372 0 : .await
5373 0 : .map_err(InitdbError::Inner);
5374 0 :
5375 0 : // This isn't true cancellation support, see above. Still return an error to
5376 0 : // excercise the cancellation code path.
5377 0 : if cancel.is_cancelled() {
5378 0 : return Err(InitdbError::Cancelled);
5379 0 : }
5380 0 :
5381 0 : res
5382 0 : }
5383 :
5384 : /// Dump contents of a layer file to stdout.
5385 0 : pub async fn dump_layerfile_from_path(
5386 0 : path: &Utf8Path,
5387 0 : verbose: bool,
5388 0 : ctx: &RequestContext,
5389 0 : ) -> anyhow::Result<()> {
5390 : use std::os::unix::fs::FileExt;
5391 :
5392 : // All layer files start with a two-byte "magic" value, to identify the kind of
5393 : // file.
5394 0 : let file = File::open(path)?;
5395 0 : let mut header_buf = [0u8; 2];
5396 0 : file.read_exact_at(&mut header_buf, 0)?;
5397 :
5398 0 : match u16::from_be_bytes(header_buf) {
5399 : crate::IMAGE_FILE_MAGIC => {
5400 0 : ImageLayer::new_for_path(path, file)?
5401 0 : .dump(verbose, ctx)
5402 0 : .await?
5403 : }
5404 : crate::DELTA_FILE_MAGIC => {
5405 0 : DeltaLayer::new_for_path(path, file)?
5406 0 : .dump(verbose, ctx)
5407 0 : .await?
5408 : }
5409 0 : magic => bail!("unrecognized magic identifier: {:?}", magic),
5410 : }
5411 :
5412 0 : Ok(())
5413 0 : }
5414 :
5415 : #[cfg(test)]
5416 : pub(crate) mod harness {
5417 : use bytes::{Bytes, BytesMut};
5418 : use once_cell::sync::OnceCell;
5419 : use pageserver_api::models::ShardParameters;
5420 : use pageserver_api::shard::ShardIndex;
5421 : use utils::logging;
5422 :
5423 : use crate::deletion_queue::mock::MockDeletionQueue;
5424 : use crate::l0_flush::L0FlushConfig;
5425 : use crate::walredo::apply_neon;
5426 : use pageserver_api::key::Key;
5427 : use pageserver_api::record::NeonWalRecord;
5428 :
5429 : use super::*;
5430 : use hex_literal::hex;
5431 : use utils::id::TenantId;
5432 :
5433 : pub const TIMELINE_ID: TimelineId =
5434 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
5435 : pub const NEW_TIMELINE_ID: TimelineId =
5436 : TimelineId::from_array(hex!("AA223344556677881122334455667788"));
5437 :
5438 : /// Convenience function to create a page image with given string as the only content
5439 10057565 : pub fn test_img(s: &str) -> Bytes {
5440 10057565 : let mut buf = BytesMut::new();
5441 10057565 : buf.extend_from_slice(s.as_bytes());
5442 10057565 : buf.resize(64, 0);
5443 10057565 :
5444 10057565 : buf.freeze()
5445 10057565 : }
5446 :
5447 : impl From<TenantConf> for TenantConfOpt {
5448 440 : fn from(tenant_conf: TenantConf) -> Self {
5449 440 : Self {
5450 440 : checkpoint_distance: Some(tenant_conf.checkpoint_distance),
5451 440 : checkpoint_timeout: Some(tenant_conf.checkpoint_timeout),
5452 440 : compaction_target_size: Some(tenant_conf.compaction_target_size),
5453 440 : compaction_period: Some(tenant_conf.compaction_period),
5454 440 : compaction_threshold: Some(tenant_conf.compaction_threshold),
5455 440 : compaction_algorithm: Some(tenant_conf.compaction_algorithm),
5456 440 : l0_flush_delay_threshold: tenant_conf.l0_flush_delay_threshold,
5457 440 : l0_flush_stall_threshold: tenant_conf.l0_flush_stall_threshold,
5458 440 : gc_horizon: Some(tenant_conf.gc_horizon),
5459 440 : gc_period: Some(tenant_conf.gc_period),
5460 440 : image_creation_threshold: Some(tenant_conf.image_creation_threshold),
5461 440 : pitr_interval: Some(tenant_conf.pitr_interval),
5462 440 : walreceiver_connect_timeout: Some(tenant_conf.walreceiver_connect_timeout),
5463 440 : lagging_wal_timeout: Some(tenant_conf.lagging_wal_timeout),
5464 440 : max_lsn_wal_lag: Some(tenant_conf.max_lsn_wal_lag),
5465 440 : eviction_policy: Some(tenant_conf.eviction_policy),
5466 440 : min_resident_size_override: tenant_conf.min_resident_size_override,
5467 440 : evictions_low_residence_duration_metric_threshold: Some(
5468 440 : tenant_conf.evictions_low_residence_duration_metric_threshold,
5469 440 : ),
5470 440 : heatmap_period: Some(tenant_conf.heatmap_period),
5471 440 : lazy_slru_download: Some(tenant_conf.lazy_slru_download),
5472 440 : timeline_get_throttle: Some(tenant_conf.timeline_get_throttle),
5473 440 : image_layer_creation_check_threshold: Some(
5474 440 : tenant_conf.image_layer_creation_check_threshold,
5475 440 : ),
5476 440 : lsn_lease_length: Some(tenant_conf.lsn_lease_length),
5477 440 : lsn_lease_length_for_ts: Some(tenant_conf.lsn_lease_length_for_ts),
5478 440 : timeline_offloading: Some(tenant_conf.timeline_offloading),
5479 440 : wal_receiver_protocol_override: tenant_conf.wal_receiver_protocol_override,
5480 440 : rel_size_v2_enabled: tenant_conf.rel_size_v2_enabled,
5481 440 : gc_compaction_enabled: Some(tenant_conf.gc_compaction_enabled),
5482 440 : gc_compaction_initial_threshold_kb: Some(
5483 440 : tenant_conf.gc_compaction_initial_threshold_kb,
5484 440 : ),
5485 440 : gc_compaction_ratio_percent: Some(tenant_conf.gc_compaction_ratio_percent),
5486 440 : }
5487 440 : }
5488 : }
5489 :
5490 : pub struct TenantHarness {
5491 : pub conf: &'static PageServerConf,
5492 : pub tenant_conf: TenantConf,
5493 : pub tenant_shard_id: TenantShardId,
5494 : pub generation: Generation,
5495 : pub shard: ShardIndex,
5496 : pub remote_storage: GenericRemoteStorage,
5497 : pub remote_fs_dir: Utf8PathBuf,
5498 : pub deletion_queue: MockDeletionQueue,
5499 : }
5500 :
5501 : static LOG_HANDLE: OnceCell<()> = OnceCell::new();
5502 :
5503 488 : pub(crate) fn setup_logging() {
5504 488 : LOG_HANDLE.get_or_init(|| {
5505 464 : logging::init(
5506 464 : logging::LogFormat::Test,
5507 464 : // enable it in case the tests exercise code paths that use
5508 464 : // debug_assert_current_span_has_tenant_and_timeline_id
5509 464 : logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
5510 464 : logging::Output::Stdout,
5511 464 : )
5512 464 : .expect("Failed to init test logging")
5513 488 : });
5514 488 : }
5515 :
5516 : impl TenantHarness {
5517 440 : pub async fn create_custom(
5518 440 : test_name: &'static str,
5519 440 : tenant_conf: TenantConf,
5520 440 : tenant_id: TenantId,
5521 440 : shard_identity: ShardIdentity,
5522 440 : generation: Generation,
5523 440 : ) -> anyhow::Result<Self> {
5524 440 : setup_logging();
5525 440 :
5526 440 : let repo_dir = PageServerConf::test_repo_dir(test_name);
5527 440 : let _ = fs::remove_dir_all(&repo_dir);
5528 440 : fs::create_dir_all(&repo_dir)?;
5529 :
5530 440 : let conf = PageServerConf::dummy_conf(repo_dir);
5531 440 : // Make a static copy of the config. This can never be free'd, but that's
5532 440 : // OK in a test.
5533 440 : let conf: &'static PageServerConf = Box::leak(Box::new(conf));
5534 440 :
5535 440 : let shard = shard_identity.shard_index();
5536 440 : let tenant_shard_id = TenantShardId {
5537 440 : tenant_id,
5538 440 : shard_number: shard.shard_number,
5539 440 : shard_count: shard.shard_count,
5540 440 : };
5541 440 : fs::create_dir_all(conf.tenant_path(&tenant_shard_id))?;
5542 440 : fs::create_dir_all(conf.timelines_path(&tenant_shard_id))?;
5543 :
5544 : use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
5545 440 : let remote_fs_dir = conf.workdir.join("localfs");
5546 440 : std::fs::create_dir_all(&remote_fs_dir).unwrap();
5547 440 : let config = RemoteStorageConfig {
5548 440 : storage: RemoteStorageKind::LocalFs {
5549 440 : local_path: remote_fs_dir.clone(),
5550 440 : },
5551 440 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
5552 440 : small_timeout: RemoteStorageConfig::DEFAULT_SMALL_TIMEOUT,
5553 440 : };
5554 440 : let remote_storage = GenericRemoteStorage::from_config(&config).await.unwrap();
5555 440 : let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()));
5556 440 :
5557 440 : Ok(Self {
5558 440 : conf,
5559 440 : tenant_conf,
5560 440 : tenant_shard_id,
5561 440 : generation,
5562 440 : shard,
5563 440 : remote_storage,
5564 440 : remote_fs_dir,
5565 440 : deletion_queue,
5566 440 : })
5567 440 : }
5568 :
5569 416 : pub async fn create(test_name: &'static str) -> anyhow::Result<Self> {
5570 416 : // Disable automatic GC and compaction to make the unit tests more deterministic.
5571 416 : // The tests perform them manually if needed.
5572 416 : let tenant_conf = TenantConf {
5573 416 : gc_period: Duration::ZERO,
5574 416 : compaction_period: Duration::ZERO,
5575 416 : ..TenantConf::default()
5576 416 : };
5577 416 : let tenant_id = TenantId::generate();
5578 416 : let shard = ShardIdentity::unsharded();
5579 416 : Self::create_custom(
5580 416 : test_name,
5581 416 : tenant_conf,
5582 416 : tenant_id,
5583 416 : shard,
5584 416 : Generation::new(0xdeadbeef),
5585 416 : )
5586 416 : .await
5587 416 : }
5588 :
5589 40 : pub fn span(&self) -> tracing::Span {
5590 40 : info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
5591 40 : }
5592 :
5593 440 : pub(crate) async fn load(&self) -> (Arc<Tenant>, RequestContext) {
5594 440 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
5595 440 : (
5596 440 : self.do_try_load(&ctx)
5597 440 : .await
5598 440 : .expect("failed to load test tenant"),
5599 440 : ctx,
5600 440 : )
5601 440 : }
5602 :
5603 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5604 : pub(crate) async fn do_try_load(
5605 : &self,
5606 : ctx: &RequestContext,
5607 : ) -> anyhow::Result<Arc<Tenant>> {
5608 : let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
5609 :
5610 : let tenant = Arc::new(Tenant::new(
5611 : TenantState::Attaching,
5612 : self.conf,
5613 : AttachedTenantConf::try_from(LocationConf::attached_single(
5614 : TenantConfOpt::from(self.tenant_conf.clone()),
5615 : self.generation,
5616 : &ShardParameters::default(),
5617 : ))
5618 : .unwrap(),
5619 : // This is a legacy/test code path: sharding isn't supported here.
5620 : ShardIdentity::unsharded(),
5621 : Some(walredo_mgr),
5622 : self.tenant_shard_id,
5623 : self.remote_storage.clone(),
5624 : self.deletion_queue.new_client(),
5625 : // TODO: ideally we should run all unit tests with both configs
5626 : L0FlushGlobalState::new(L0FlushConfig::default()),
5627 : ));
5628 :
5629 : let preload = tenant
5630 : .preload(&self.remote_storage, CancellationToken::new())
5631 : .await?;
5632 : tenant.attach(Some(preload), ctx).await?;
5633 :
5634 : tenant.state.send_replace(TenantState::Active);
5635 : for timeline in tenant.timelines.lock().unwrap().values() {
5636 : timeline.set_state(TimelineState::Active);
5637 : }
5638 : Ok(tenant)
5639 : }
5640 :
5641 4 : pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
5642 4 : self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
5643 4 : }
5644 : }
5645 :
5646 : // Mock WAL redo manager that doesn't do much
5647 : pub(crate) struct TestRedoManager;
5648 :
5649 : impl TestRedoManager {
5650 : /// # Cancel-Safety
5651 : ///
5652 : /// This method is cancellation-safe.
5653 1636 : pub async fn request_redo(
5654 1636 : &self,
5655 1636 : key: Key,
5656 1636 : lsn: Lsn,
5657 1636 : base_img: Option<(Lsn, Bytes)>,
5658 1636 : records: Vec<(Lsn, NeonWalRecord)>,
5659 1636 : _pg_version: u32,
5660 1636 : ) -> Result<Bytes, walredo::Error> {
5661 2392 : let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
5662 1636 : if records_neon {
5663 : // For Neon wal records, we can decode without spawning postgres, so do so.
5664 1636 : let mut page = match (base_img, records.first()) {
5665 1504 : (Some((_lsn, img)), _) => {
5666 1504 : let mut page = BytesMut::new();
5667 1504 : page.extend_from_slice(&img);
5668 1504 : page
5669 : }
5670 132 : (_, Some((_lsn, rec))) if rec.will_init() => BytesMut::new(),
5671 : _ => {
5672 0 : panic!("Neon WAL redo requires base image or will init record");
5673 : }
5674 : };
5675 :
5676 4028 : for (record_lsn, record) in records {
5677 2392 : apply_neon::apply_in_neon(&record, record_lsn, key, &mut page)?;
5678 : }
5679 1636 : Ok(page.freeze())
5680 : } else {
5681 : // We never spawn a postgres walredo process in unit tests: just log what we might have done.
5682 0 : let s = format!(
5683 0 : "redo for {} to get to {}, with {} and {} records",
5684 0 : key,
5685 0 : lsn,
5686 0 : if base_img.is_some() {
5687 0 : "base image"
5688 : } else {
5689 0 : "no base image"
5690 : },
5691 0 : records.len()
5692 0 : );
5693 0 : println!("{s}");
5694 0 :
5695 0 : Ok(test_img(&s))
5696 : }
5697 1636 : }
5698 : }
5699 : }
5700 :
5701 : #[cfg(test)]
5702 : mod tests {
5703 : use std::collections::{BTreeMap, BTreeSet};
5704 :
5705 : use super::*;
5706 : use crate::keyspace::KeySpaceAccum;
5707 : use crate::tenant::harness::*;
5708 : use crate::tenant::timeline::CompactFlags;
5709 : use crate::DEFAULT_PG_VERSION;
5710 : use bytes::{Bytes, BytesMut};
5711 : use hex_literal::hex;
5712 : use itertools::Itertools;
5713 : use pageserver_api::key::{Key, AUX_KEY_PREFIX, NON_INHERITED_RANGE, RELATION_SIZE_PREFIX};
5714 : use pageserver_api::keyspace::KeySpace;
5715 : use pageserver_api::models::{CompactionAlgorithm, CompactionAlgorithmSettings};
5716 : use pageserver_api::value::Value;
5717 : use pageserver_compaction::helpers::overlaps_with;
5718 : use rand::{thread_rng, Rng};
5719 : use storage_layer::{IoConcurrency, PersistentLayerKey};
5720 : use tests::storage_layer::ValuesReconstructState;
5721 : use tests::timeline::{GetVectoredError, ShutdownMode};
5722 : use timeline::{CompactOptions, DeltaLayerTestDesc};
5723 : use utils::id::TenantId;
5724 :
5725 : #[cfg(feature = "testing")]
5726 : use models::CompactLsnRange;
5727 : #[cfg(feature = "testing")]
5728 : use pageserver_api::record::NeonWalRecord;
5729 : #[cfg(feature = "testing")]
5730 : use timeline::compaction::{KeyHistoryRetention, KeyLogAtLsn};
5731 : #[cfg(feature = "testing")]
5732 : use timeline::GcInfo;
5733 :
5734 : static TEST_KEY: Lazy<Key> =
5735 36 : Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));
5736 :
5737 : #[tokio::test]
5738 4 : async fn test_basic() -> anyhow::Result<()> {
5739 4 : let (tenant, ctx) = TenantHarness::create("test_basic").await?.load().await;
5740 4 : let tline = tenant
5741 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5742 4 : .await?;
5743 4 :
5744 4 : let mut writer = tline.writer().await;
5745 4 : writer
5746 4 : .put(
5747 4 : *TEST_KEY,
5748 4 : Lsn(0x10),
5749 4 : &Value::Image(test_img("foo at 0x10")),
5750 4 : &ctx,
5751 4 : )
5752 4 : .await?;
5753 4 : writer.finish_write(Lsn(0x10));
5754 4 : drop(writer);
5755 4 :
5756 4 : let mut writer = tline.writer().await;
5757 4 : writer
5758 4 : .put(
5759 4 : *TEST_KEY,
5760 4 : Lsn(0x20),
5761 4 : &Value::Image(test_img("foo at 0x20")),
5762 4 : &ctx,
5763 4 : )
5764 4 : .await?;
5765 4 : writer.finish_write(Lsn(0x20));
5766 4 : drop(writer);
5767 4 :
5768 4 : assert_eq!(
5769 4 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
5770 4 : test_img("foo at 0x10")
5771 4 : );
5772 4 : assert_eq!(
5773 4 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
5774 4 : test_img("foo at 0x10")
5775 4 : );
5776 4 : assert_eq!(
5777 4 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
5778 4 : test_img("foo at 0x20")
5779 4 : );
5780 4 :
5781 4 : Ok(())
5782 4 : }
5783 :
5784 : #[tokio::test]
5785 4 : async fn no_duplicate_timelines() -> anyhow::Result<()> {
5786 4 : let (tenant, ctx) = TenantHarness::create("no_duplicate_timelines")
5787 4 : .await?
5788 4 : .load()
5789 4 : .await;
5790 4 : let _ = tenant
5791 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5792 4 : .await?;
5793 4 :
5794 4 : match tenant
5795 4 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5796 4 : .await
5797 4 : {
5798 4 : Ok(_) => panic!("duplicate timeline creation should fail"),
5799 4 : Err(e) => assert_eq!(
5800 4 : e.to_string(),
5801 4 : "timeline already exists with different parameters".to_string()
5802 4 : ),
5803 4 : }
5804 4 :
5805 4 : Ok(())
5806 4 : }
5807 :
5808 : /// Convenience function to create a page image with given string as the only content
5809 20 : pub fn test_value(s: &str) -> Value {
5810 20 : let mut buf = BytesMut::new();
5811 20 : buf.extend_from_slice(s.as_bytes());
5812 20 : Value::Image(buf.freeze())
5813 20 : }
5814 :
5815 : ///
5816 : /// Test branch creation
5817 : ///
5818 : #[tokio::test]
5819 4 : async fn test_branch() -> anyhow::Result<()> {
5820 4 : use std::str::from_utf8;
5821 4 :
5822 4 : let (tenant, ctx) = TenantHarness::create("test_branch").await?.load().await;
5823 4 : let tline = tenant
5824 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5825 4 : .await?;
5826 4 : let mut writer = tline.writer().await;
5827 4 :
5828 4 : #[allow(non_snake_case)]
5829 4 : let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
5830 4 : #[allow(non_snake_case)]
5831 4 : let TEST_KEY_B: Key = Key::from_hex("110000000033333333444444445500000002").unwrap();
5832 4 :
5833 4 : // Insert a value on the timeline
5834 4 : writer
5835 4 : .put(TEST_KEY_A, Lsn(0x20), &test_value("foo at 0x20"), &ctx)
5836 4 : .await?;
5837 4 : writer
5838 4 : .put(TEST_KEY_B, Lsn(0x20), &test_value("foobar at 0x20"), &ctx)
5839 4 : .await?;
5840 4 : writer.finish_write(Lsn(0x20));
5841 4 :
5842 4 : writer
5843 4 : .put(TEST_KEY_A, Lsn(0x30), &test_value("foo at 0x30"), &ctx)
5844 4 : .await?;
5845 4 : writer.finish_write(Lsn(0x30));
5846 4 : writer
5847 4 : .put(TEST_KEY_A, Lsn(0x40), &test_value("foo at 0x40"), &ctx)
5848 4 : .await?;
5849 4 : writer.finish_write(Lsn(0x40));
5850 4 :
5851 4 : //assert_current_logical_size(&tline, Lsn(0x40));
5852 4 :
5853 4 : // Branch the history, modify relation differently on the new timeline
5854 4 : tenant
5855 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x30)), &ctx)
5856 4 : .await?;
5857 4 : let newtline = tenant
5858 4 : .get_timeline(NEW_TIMELINE_ID, true)
5859 4 : .expect("Should have a local timeline");
5860 4 : let mut new_writer = newtline.writer().await;
5861 4 : new_writer
5862 4 : .put(TEST_KEY_A, Lsn(0x40), &test_value("bar at 0x40"), &ctx)
5863 4 : .await?;
5864 4 : new_writer.finish_write(Lsn(0x40));
5865 4 :
5866 4 : // Check page contents on both branches
5867 4 : assert_eq!(
5868 4 : from_utf8(&tline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
5869 4 : "foo at 0x40"
5870 4 : );
5871 4 : assert_eq!(
5872 4 : from_utf8(&newtline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
5873 4 : "bar at 0x40"
5874 4 : );
5875 4 : assert_eq!(
5876 4 : from_utf8(&newtline.get(TEST_KEY_B, Lsn(0x40), &ctx).await?)?,
5877 4 : "foobar at 0x20"
5878 4 : );
5879 4 :
5880 4 : //assert_current_logical_size(&tline, Lsn(0x40));
5881 4 :
5882 4 : Ok(())
5883 4 : }
5884 :
5885 40 : async fn make_some_layers(
5886 40 : tline: &Timeline,
5887 40 : start_lsn: Lsn,
5888 40 : ctx: &RequestContext,
5889 40 : ) -> anyhow::Result<()> {
5890 40 : let mut lsn = start_lsn;
5891 : {
5892 40 : let mut writer = tline.writer().await;
5893 : // Create a relation on the timeline
5894 40 : writer
5895 40 : .put(
5896 40 : *TEST_KEY,
5897 40 : lsn,
5898 40 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5899 40 : ctx,
5900 40 : )
5901 40 : .await?;
5902 40 : writer.finish_write(lsn);
5903 40 : lsn += 0x10;
5904 40 : writer
5905 40 : .put(
5906 40 : *TEST_KEY,
5907 40 : lsn,
5908 40 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5909 40 : ctx,
5910 40 : )
5911 40 : .await?;
5912 40 : writer.finish_write(lsn);
5913 40 : lsn += 0x10;
5914 40 : }
5915 40 : tline.freeze_and_flush().await?;
5916 : {
5917 40 : let mut writer = tline.writer().await;
5918 40 : writer
5919 40 : .put(
5920 40 : *TEST_KEY,
5921 40 : lsn,
5922 40 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5923 40 : ctx,
5924 40 : )
5925 40 : .await?;
5926 40 : writer.finish_write(lsn);
5927 40 : lsn += 0x10;
5928 40 : writer
5929 40 : .put(
5930 40 : *TEST_KEY,
5931 40 : lsn,
5932 40 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5933 40 : ctx,
5934 40 : )
5935 40 : .await?;
5936 40 : writer.finish_write(lsn);
5937 40 : }
5938 40 : tline.freeze_and_flush().await.map_err(|e| e.into())
5939 40 : }
5940 :
5941 : #[tokio::test(start_paused = true)]
5942 4 : async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
5943 4 : let (tenant, ctx) =
5944 4 : TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
5945 4 : .await?
5946 4 : .load()
5947 4 : .await;
5948 4 : // Advance to the lsn lease deadline so that GC is not blocked by
5949 4 : // initial transition into AttachedSingle.
5950 4 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
5951 4 : tokio::time::resume();
5952 4 : let tline = tenant
5953 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5954 4 : .await?;
5955 4 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5956 4 :
5957 4 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
5958 4 : // FIXME: this doesn't actually remove any layer currently, given how the flushing
5959 4 : // and compaction works. But it does set the 'cutoff' point so that the cross check
5960 4 : // below should fail.
5961 4 : tenant
5962 4 : .gc_iteration(
5963 4 : Some(TIMELINE_ID),
5964 4 : 0x10,
5965 4 : Duration::ZERO,
5966 4 : &CancellationToken::new(),
5967 4 : &ctx,
5968 4 : )
5969 4 : .await?;
5970 4 :
5971 4 : // try to branch at lsn 25, should fail because we already garbage collected the data
5972 4 : match tenant
5973 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
5974 4 : .await
5975 4 : {
5976 4 : Ok(_) => panic!("branching should have failed"),
5977 4 : Err(err) => {
5978 4 : let CreateTimelineError::AncestorLsn(err) = err else {
5979 4 : panic!("wrong error type")
5980 4 : };
5981 4 : assert!(err.to_string().contains("invalid branch start lsn"));
5982 4 : assert!(err
5983 4 : .source()
5984 4 : .unwrap()
5985 4 : .to_string()
5986 4 : .contains("we might've already garbage collected needed data"))
5987 4 : }
5988 4 : }
5989 4 :
5990 4 : Ok(())
5991 4 : }
5992 :
5993 : #[tokio::test]
5994 4 : async fn test_prohibit_branch_creation_on_pre_initdb_lsn() -> anyhow::Result<()> {
5995 4 : let (tenant, ctx) =
5996 4 : TenantHarness::create("test_prohibit_branch_creation_on_pre_initdb_lsn")
5997 4 : .await?
5998 4 : .load()
5999 4 : .await;
6000 4 :
6001 4 : let tline = tenant
6002 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x50), DEFAULT_PG_VERSION, &ctx)
6003 4 : .await?;
6004 4 : // try to branch at lsn 0x25, should fail because initdb lsn is 0x50
6005 4 : match tenant
6006 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
6007 4 : .await
6008 4 : {
6009 4 : Ok(_) => panic!("branching should have failed"),
6010 4 : Err(err) => {
6011 4 : let CreateTimelineError::AncestorLsn(err) = err else {
6012 4 : panic!("wrong error type");
6013 4 : };
6014 4 : assert!(&err.to_string().contains("invalid branch start lsn"));
6015 4 : assert!(&err
6016 4 : .source()
6017 4 : .unwrap()
6018 4 : .to_string()
6019 4 : .contains("is earlier than latest GC cutoff"));
6020 4 : }
6021 4 : }
6022 4 :
6023 4 : Ok(())
6024 4 : }
6025 :
6026 : /*
6027 : // FIXME: This currently fails to error out. Calling GC doesn't currently
6028 : // remove the old value, we'd need to work a little harder
6029 : #[tokio::test]
6030 : async fn test_prohibit_get_for_garbage_collected_data() -> anyhow::Result<()> {
6031 : let repo =
6032 : RepoHarness::create("test_prohibit_get_for_garbage_collected_data")?
6033 : .load();
6034 :
6035 : let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION)?;
6036 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6037 :
6038 : repo.gc_iteration(Some(TIMELINE_ID), 0x10, Duration::ZERO)?;
6039 : let latest_gc_cutoff_lsn = tline.get_latest_gc_cutoff_lsn();
6040 : assert!(*latest_gc_cutoff_lsn > Lsn(0x25));
6041 : match tline.get(*TEST_KEY, Lsn(0x25)) {
6042 : Ok(_) => panic!("request for page should have failed"),
6043 : Err(err) => assert!(err.to_string().contains("not found at")),
6044 : }
6045 : Ok(())
6046 : }
6047 : */
6048 :
6049 : #[tokio::test]
6050 4 : async fn test_get_branchpoints_from_an_inactive_timeline() -> anyhow::Result<()> {
6051 4 : let (tenant, ctx) =
6052 4 : TenantHarness::create("test_get_branchpoints_from_an_inactive_timeline")
6053 4 : .await?
6054 4 : .load()
6055 4 : .await;
6056 4 : let tline = tenant
6057 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6058 4 : .await?;
6059 4 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6060 4 :
6061 4 : tenant
6062 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6063 4 : .await?;
6064 4 : let newtline = tenant
6065 4 : .get_timeline(NEW_TIMELINE_ID, true)
6066 4 : .expect("Should have a local timeline");
6067 4 :
6068 4 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6069 4 :
6070 4 : tline.set_broken("test".to_owned());
6071 4 :
6072 4 : tenant
6073 4 : .gc_iteration(
6074 4 : Some(TIMELINE_ID),
6075 4 : 0x10,
6076 4 : Duration::ZERO,
6077 4 : &CancellationToken::new(),
6078 4 : &ctx,
6079 4 : )
6080 4 : .await?;
6081 4 :
6082 4 : // The branchpoints should contain all timelines, even ones marked
6083 4 : // as Broken.
6084 4 : {
6085 4 : let branchpoints = &tline.gc_info.read().unwrap().retain_lsns;
6086 4 : assert_eq!(branchpoints.len(), 1);
6087 4 : assert_eq!(
6088 4 : branchpoints[0],
6089 4 : (Lsn(0x40), NEW_TIMELINE_ID, MaybeOffloaded::No)
6090 4 : );
6091 4 : }
6092 4 :
6093 4 : // You can read the key from the child branch even though the parent is
6094 4 : // Broken, as long as you don't need to access data from the parent.
6095 4 : assert_eq!(
6096 4 : newtline.get(*TEST_KEY, Lsn(0x70), &ctx).await?,
6097 4 : test_img(&format!("foo at {}", Lsn(0x70)))
6098 4 : );
6099 4 :
6100 4 : // This needs to traverse to the parent, and fails.
6101 4 : let err = newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await.unwrap_err();
6102 4 : assert!(
6103 4 : err.to_string().starts_with(&format!(
6104 4 : "bad state on timeline {}: Broken",
6105 4 : tline.timeline_id
6106 4 : )),
6107 4 : "{err}"
6108 4 : );
6109 4 :
6110 4 : Ok(())
6111 4 : }
6112 :
6113 : #[tokio::test]
6114 4 : async fn test_retain_data_in_parent_which_is_needed_for_child() -> anyhow::Result<()> {
6115 4 : let (tenant, ctx) =
6116 4 : TenantHarness::create("test_retain_data_in_parent_which_is_needed_for_child")
6117 4 : .await?
6118 4 : .load()
6119 4 : .await;
6120 4 : let tline = tenant
6121 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6122 4 : .await?;
6123 4 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6124 4 :
6125 4 : tenant
6126 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6127 4 : .await?;
6128 4 : let newtline = tenant
6129 4 : .get_timeline(NEW_TIMELINE_ID, true)
6130 4 : .expect("Should have a local timeline");
6131 4 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
6132 4 : tenant
6133 4 : .gc_iteration(
6134 4 : Some(TIMELINE_ID),
6135 4 : 0x10,
6136 4 : Duration::ZERO,
6137 4 : &CancellationToken::new(),
6138 4 : &ctx,
6139 4 : )
6140 4 : .await?;
6141 4 : assert!(newtline.get(*TEST_KEY, Lsn(0x25), &ctx).await.is_ok());
6142 4 :
6143 4 : Ok(())
6144 4 : }
6145 : #[tokio::test]
6146 4 : async fn test_parent_keeps_data_forever_after_branching() -> anyhow::Result<()> {
6147 4 : let (tenant, ctx) = TenantHarness::create("test_parent_keeps_data_forever_after_branching")
6148 4 : .await?
6149 4 : .load()
6150 4 : .await;
6151 4 : let tline = tenant
6152 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6153 4 : .await?;
6154 4 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6155 4 :
6156 4 : tenant
6157 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6158 4 : .await?;
6159 4 : let newtline = tenant
6160 4 : .get_timeline(NEW_TIMELINE_ID, true)
6161 4 : .expect("Should have a local timeline");
6162 4 :
6163 4 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6164 4 :
6165 4 : // run gc on parent
6166 4 : tenant
6167 4 : .gc_iteration(
6168 4 : Some(TIMELINE_ID),
6169 4 : 0x10,
6170 4 : Duration::ZERO,
6171 4 : &CancellationToken::new(),
6172 4 : &ctx,
6173 4 : )
6174 4 : .await?;
6175 4 :
6176 4 : // Check that the data is still accessible on the branch.
6177 4 : assert_eq!(
6178 4 : newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
6179 4 : test_img(&format!("foo at {}", Lsn(0x40)))
6180 4 : );
6181 4 :
6182 4 : Ok(())
6183 4 : }
6184 :
6185 : #[tokio::test]
6186 4 : async fn timeline_load() -> anyhow::Result<()> {
6187 4 : const TEST_NAME: &str = "timeline_load";
6188 4 : let harness = TenantHarness::create(TEST_NAME).await?;
6189 4 : {
6190 4 : let (tenant, ctx) = harness.load().await;
6191 4 : let tline = tenant
6192 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x7000), DEFAULT_PG_VERSION, &ctx)
6193 4 : .await?;
6194 4 : make_some_layers(tline.as_ref(), Lsn(0x8000), &ctx).await?;
6195 4 : // so that all uploads finish & we can call harness.load() below again
6196 4 : tenant
6197 4 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
6198 4 : .instrument(harness.span())
6199 4 : .await
6200 4 : .ok()
6201 4 : .unwrap();
6202 4 : }
6203 4 :
6204 4 : let (tenant, _ctx) = harness.load().await;
6205 4 : tenant
6206 4 : .get_timeline(TIMELINE_ID, true)
6207 4 : .expect("cannot load timeline");
6208 4 :
6209 4 : Ok(())
6210 4 : }
6211 :
6212 : #[tokio::test]
6213 4 : async fn timeline_load_with_ancestor() -> anyhow::Result<()> {
6214 4 : const TEST_NAME: &str = "timeline_load_with_ancestor";
6215 4 : let harness = TenantHarness::create(TEST_NAME).await?;
6216 4 : // create two timelines
6217 4 : {
6218 4 : let (tenant, ctx) = harness.load().await;
6219 4 : let tline = tenant
6220 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6221 4 : .await?;
6222 4 :
6223 4 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6224 4 :
6225 4 : let child_tline = tenant
6226 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6227 4 : .await?;
6228 4 : child_tline.set_state(TimelineState::Active);
6229 4 :
6230 4 : let newtline = tenant
6231 4 : .get_timeline(NEW_TIMELINE_ID, true)
6232 4 : .expect("Should have a local timeline");
6233 4 :
6234 4 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6235 4 :
6236 4 : // so that all uploads finish & we can call harness.load() below again
6237 4 : tenant
6238 4 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
6239 4 : .instrument(harness.span())
6240 4 : .await
6241 4 : .ok()
6242 4 : .unwrap();
6243 4 : }
6244 4 :
6245 4 : // check that both of them are initially unloaded
6246 4 : let (tenant, _ctx) = harness.load().await;
6247 4 :
6248 4 : // check that both, child and ancestor are loaded
6249 4 : let _child_tline = tenant
6250 4 : .get_timeline(NEW_TIMELINE_ID, true)
6251 4 : .expect("cannot get child timeline loaded");
6252 4 :
6253 4 : let _ancestor_tline = tenant
6254 4 : .get_timeline(TIMELINE_ID, true)
6255 4 : .expect("cannot get ancestor timeline loaded");
6256 4 :
6257 4 : Ok(())
6258 4 : }
6259 :
6260 : #[tokio::test]
6261 4 : async fn delta_layer_dumping() -> anyhow::Result<()> {
6262 4 : use storage_layer::AsLayerDesc;
6263 4 : let (tenant, ctx) = TenantHarness::create("test_layer_dumping")
6264 4 : .await?
6265 4 : .load()
6266 4 : .await;
6267 4 : let tline = tenant
6268 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6269 4 : .await?;
6270 4 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6271 4 :
6272 4 : let layer_map = tline.layers.read().await;
6273 4 : let level0_deltas = layer_map
6274 4 : .layer_map()?
6275 4 : .level0_deltas()
6276 4 : .iter()
6277 8 : .map(|desc| layer_map.get_from_desc(desc))
6278 4 : .collect::<Vec<_>>();
6279 4 :
6280 4 : assert!(!level0_deltas.is_empty());
6281 4 :
6282 12 : for delta in level0_deltas {
6283 4 : // Ensure we are dumping a delta layer here
6284 8 : assert!(delta.layer_desc().is_delta);
6285 8 : delta.dump(true, &ctx).await.unwrap();
6286 4 : }
6287 4 :
6288 4 : Ok(())
6289 4 : }
6290 :
6291 : #[tokio::test]
6292 4 : async fn test_images() -> anyhow::Result<()> {
6293 4 : let (tenant, ctx) = TenantHarness::create("test_images").await?.load().await;
6294 4 : let tline = tenant
6295 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6296 4 : .await?;
6297 4 :
6298 4 : let mut writer = tline.writer().await;
6299 4 : writer
6300 4 : .put(
6301 4 : *TEST_KEY,
6302 4 : Lsn(0x10),
6303 4 : &Value::Image(test_img("foo at 0x10")),
6304 4 : &ctx,
6305 4 : )
6306 4 : .await?;
6307 4 : writer.finish_write(Lsn(0x10));
6308 4 : drop(writer);
6309 4 :
6310 4 : tline.freeze_and_flush().await?;
6311 4 : tline
6312 4 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6313 4 : .await?;
6314 4 :
6315 4 : let mut writer = tline.writer().await;
6316 4 : writer
6317 4 : .put(
6318 4 : *TEST_KEY,
6319 4 : Lsn(0x20),
6320 4 : &Value::Image(test_img("foo at 0x20")),
6321 4 : &ctx,
6322 4 : )
6323 4 : .await?;
6324 4 : writer.finish_write(Lsn(0x20));
6325 4 : drop(writer);
6326 4 :
6327 4 : tline.freeze_and_flush().await?;
6328 4 : tline
6329 4 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6330 4 : .await?;
6331 4 :
6332 4 : let mut writer = tline.writer().await;
6333 4 : writer
6334 4 : .put(
6335 4 : *TEST_KEY,
6336 4 : Lsn(0x30),
6337 4 : &Value::Image(test_img("foo at 0x30")),
6338 4 : &ctx,
6339 4 : )
6340 4 : .await?;
6341 4 : writer.finish_write(Lsn(0x30));
6342 4 : drop(writer);
6343 4 :
6344 4 : tline.freeze_and_flush().await?;
6345 4 : tline
6346 4 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6347 4 : .await?;
6348 4 :
6349 4 : let mut writer = tline.writer().await;
6350 4 : writer
6351 4 : .put(
6352 4 : *TEST_KEY,
6353 4 : Lsn(0x40),
6354 4 : &Value::Image(test_img("foo at 0x40")),
6355 4 : &ctx,
6356 4 : )
6357 4 : .await?;
6358 4 : writer.finish_write(Lsn(0x40));
6359 4 : drop(writer);
6360 4 :
6361 4 : tline.freeze_and_flush().await?;
6362 4 : tline
6363 4 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6364 4 : .await?;
6365 4 :
6366 4 : assert_eq!(
6367 4 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
6368 4 : test_img("foo at 0x10")
6369 4 : );
6370 4 : assert_eq!(
6371 4 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
6372 4 : test_img("foo at 0x10")
6373 4 : );
6374 4 : assert_eq!(
6375 4 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
6376 4 : test_img("foo at 0x20")
6377 4 : );
6378 4 : assert_eq!(
6379 4 : tline.get(*TEST_KEY, Lsn(0x30), &ctx).await?,
6380 4 : test_img("foo at 0x30")
6381 4 : );
6382 4 : assert_eq!(
6383 4 : tline.get(*TEST_KEY, Lsn(0x40), &ctx).await?,
6384 4 : test_img("foo at 0x40")
6385 4 : );
6386 4 :
6387 4 : Ok(())
6388 4 : }
6389 :
6390 8 : async fn bulk_insert_compact_gc(
6391 8 : tenant: &Tenant,
6392 8 : timeline: &Arc<Timeline>,
6393 8 : ctx: &RequestContext,
6394 8 : lsn: Lsn,
6395 8 : repeat: usize,
6396 8 : key_count: usize,
6397 8 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
6398 8 : let compact = true;
6399 8 : bulk_insert_maybe_compact_gc(tenant, timeline, ctx, lsn, repeat, key_count, compact).await
6400 8 : }
6401 :
6402 16 : async fn bulk_insert_maybe_compact_gc(
6403 16 : tenant: &Tenant,
6404 16 : timeline: &Arc<Timeline>,
6405 16 : ctx: &RequestContext,
6406 16 : mut lsn: Lsn,
6407 16 : repeat: usize,
6408 16 : key_count: usize,
6409 16 : compact: bool,
6410 16 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
6411 16 : let mut inserted: HashMap<Key, BTreeSet<Lsn>> = Default::default();
6412 16 :
6413 16 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6414 16 : let mut blknum = 0;
6415 16 :
6416 16 : // Enforce that key range is monotonously increasing
6417 16 : let mut keyspace = KeySpaceAccum::new();
6418 16 :
6419 16 : let cancel = CancellationToken::new();
6420 16 :
6421 16 : for _ in 0..repeat {
6422 800 : for _ in 0..key_count {
6423 8000000 : test_key.field6 = blknum;
6424 8000000 : let mut writer = timeline.writer().await;
6425 8000000 : writer
6426 8000000 : .put(
6427 8000000 : test_key,
6428 8000000 : lsn,
6429 8000000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6430 8000000 : ctx,
6431 8000000 : )
6432 8000000 : .await?;
6433 8000000 : inserted.entry(test_key).or_default().insert(lsn);
6434 8000000 : writer.finish_write(lsn);
6435 8000000 : drop(writer);
6436 8000000 :
6437 8000000 : keyspace.add_key(test_key);
6438 8000000 :
6439 8000000 : lsn = Lsn(lsn.0 + 0x10);
6440 8000000 : blknum += 1;
6441 : }
6442 :
6443 800 : timeline.freeze_and_flush().await?;
6444 800 : if compact {
6445 : // this requires timeline to be &Arc<Timeline>
6446 400 : timeline.compact(&cancel, EnumSet::empty(), ctx).await?;
6447 400 : }
6448 :
6449 : // this doesn't really need to use the timeline_id target, but it is closer to what it
6450 : // originally was.
6451 800 : let res = tenant
6452 800 : .gc_iteration(Some(timeline.timeline_id), 0, Duration::ZERO, &cancel, ctx)
6453 800 : .await?;
6454 :
6455 800 : assert_eq!(res.layers_removed, 0, "this never removes anything");
6456 : }
6457 :
6458 16 : Ok(inserted)
6459 16 : }
6460 :
6461 : //
6462 : // Insert 1000 key-value pairs with increasing keys, flush, compact, GC.
6463 : // Repeat 50 times.
6464 : //
6465 : #[tokio::test]
6466 4 : async fn test_bulk_insert() -> anyhow::Result<()> {
6467 4 : let harness = TenantHarness::create("test_bulk_insert").await?;
6468 4 : let (tenant, ctx) = harness.load().await;
6469 4 : let tline = tenant
6470 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6471 4 : .await?;
6472 4 :
6473 4 : let lsn = Lsn(0x10);
6474 4 : bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
6475 4 :
6476 4 : Ok(())
6477 4 : }
6478 :
6479 : // Test the vectored get real implementation against a simple sequential implementation.
6480 : //
6481 : // The test generates a keyspace by repeatedly flushing the in-memory layer and compacting.
6482 : // Projected to 2D the key space looks like below. Lsn grows upwards on the Y axis and keys
6483 : // grow to the right on the X axis.
6484 : // [Delta]
6485 : // [Delta]
6486 : // [Delta]
6487 : // [Delta]
6488 : // ------------ Image ---------------
6489 : //
6490 : // After layer generation we pick the ranges to query as follows:
6491 : // 1. The beginning of each delta layer
6492 : // 2. At the seam between two adjacent delta layers
6493 : //
6494 : // There's one major downside to this test: delta layers only contains images,
6495 : // so the search can stop at the first delta layer and doesn't traverse any deeper.
6496 : #[tokio::test]
6497 4 : async fn test_get_vectored() -> anyhow::Result<()> {
6498 4 : let harness = TenantHarness::create("test_get_vectored").await?;
6499 4 : let (tenant, ctx) = harness.load().await;
6500 4 : let io_concurrency = IoConcurrency::spawn_for_test();
6501 4 : let tline = tenant
6502 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6503 4 : .await?;
6504 4 :
6505 4 : let lsn = Lsn(0x10);
6506 4 : let inserted = bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
6507 4 :
6508 4 : let guard = tline.layers.read().await;
6509 4 : let lm = guard.layer_map()?;
6510 4 :
6511 4 : lm.dump(true, &ctx).await?;
6512 4 :
6513 4 : let mut reads = Vec::new();
6514 4 : let mut prev = None;
6515 24 : lm.iter_historic_layers().for_each(|desc| {
6516 24 : if !desc.is_delta() {
6517 4 : prev = Some(desc.clone());
6518 4 : return;
6519 20 : }
6520 20 :
6521 20 : let start = desc.key_range.start;
6522 20 : let end = desc
6523 20 : .key_range
6524 20 : .start
6525 20 : .add(Timeline::MAX_GET_VECTORED_KEYS.try_into().unwrap());
6526 20 : reads.push(KeySpace {
6527 20 : ranges: vec![start..end],
6528 20 : });
6529 4 :
6530 20 : if let Some(prev) = &prev {
6531 20 : if !prev.is_delta() {
6532 20 : return;
6533 4 : }
6534 0 :
6535 0 : let first_range = Key {
6536 0 : field6: prev.key_range.end.field6 - 4,
6537 0 : ..prev.key_range.end
6538 0 : }..prev.key_range.end;
6539 0 :
6540 0 : let second_range = desc.key_range.start..Key {
6541 0 : field6: desc.key_range.start.field6 + 4,
6542 0 : ..desc.key_range.start
6543 0 : };
6544 0 :
6545 0 : reads.push(KeySpace {
6546 0 : ranges: vec![first_range, second_range],
6547 0 : });
6548 4 : };
6549 4 :
6550 4 : prev = Some(desc.clone());
6551 24 : });
6552 4 :
6553 4 : drop(guard);
6554 4 :
6555 4 : // Pick a big LSN such that we query over all the changes.
6556 4 : let reads_lsn = Lsn(u64::MAX - 1);
6557 4 :
6558 24 : for read in reads {
6559 20 : info!("Doing vectored read on {:?}", read);
6560 4 :
6561 20 : let vectored_res = tline
6562 20 : .get_vectored_impl(
6563 20 : read.clone(),
6564 20 : reads_lsn,
6565 20 : &mut ValuesReconstructState::new(io_concurrency.clone()),
6566 20 : &ctx,
6567 20 : )
6568 20 : .await;
6569 4 :
6570 20 : let mut expected_lsns: HashMap<Key, Lsn> = Default::default();
6571 20 : let mut expect_missing = false;
6572 20 : let mut key = read.start().unwrap();
6573 660 : while key != read.end().unwrap() {
6574 640 : if let Some(lsns) = inserted.get(&key) {
6575 640 : let expected_lsn = lsns.iter().rfind(|lsn| **lsn <= reads_lsn);
6576 640 : match expected_lsn {
6577 640 : Some(lsn) => {
6578 640 : expected_lsns.insert(key, *lsn);
6579 640 : }
6580 4 : None => {
6581 4 : expect_missing = true;
6582 0 : break;
6583 4 : }
6584 4 : }
6585 4 : } else {
6586 4 : expect_missing = true;
6587 0 : break;
6588 4 : }
6589 4 :
6590 640 : key = key.next();
6591 4 : }
6592 4 :
6593 20 : if expect_missing {
6594 4 : assert!(matches!(vectored_res, Err(GetVectoredError::MissingKey(_))));
6595 4 : } else {
6596 640 : for (key, image) in vectored_res? {
6597 640 : let expected_lsn = expected_lsns.get(&key).expect("determined above");
6598 640 : let expected_image = test_img(&format!("{} at {}", key.field6, expected_lsn));
6599 640 : assert_eq!(image?, expected_image);
6600 4 : }
6601 4 : }
6602 4 : }
6603 4 :
6604 4 : Ok(())
6605 4 : }
6606 :
6607 : #[tokio::test]
6608 4 : async fn test_get_vectored_aux_files() -> anyhow::Result<()> {
6609 4 : let harness = TenantHarness::create("test_get_vectored_aux_files").await?;
6610 4 :
6611 4 : let (tenant, ctx) = harness.load().await;
6612 4 : let io_concurrency = IoConcurrency::spawn_for_test();
6613 4 : let tline = tenant
6614 4 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
6615 4 : .await?;
6616 4 : let tline = tline.raw_timeline().unwrap();
6617 4 :
6618 4 : let mut modification = tline.begin_modification(Lsn(0x1000));
6619 4 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
6620 4 : modification.set_lsn(Lsn(0x1008))?;
6621 4 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
6622 4 : modification.commit(&ctx).await?;
6623 4 :
6624 4 : let child_timeline_id = TimelineId::generate();
6625 4 : tenant
6626 4 : .branch_timeline_test(
6627 4 : tline,
6628 4 : child_timeline_id,
6629 4 : Some(tline.get_last_record_lsn()),
6630 4 : &ctx,
6631 4 : )
6632 4 : .await?;
6633 4 :
6634 4 : let child_timeline = tenant
6635 4 : .get_timeline(child_timeline_id, true)
6636 4 : .expect("Should have the branched timeline");
6637 4 :
6638 4 : let aux_keyspace = KeySpace {
6639 4 : ranges: vec![NON_INHERITED_RANGE],
6640 4 : };
6641 4 : let read_lsn = child_timeline.get_last_record_lsn();
6642 4 :
6643 4 : let vectored_res = child_timeline
6644 4 : .get_vectored_impl(
6645 4 : aux_keyspace.clone(),
6646 4 : read_lsn,
6647 4 : &mut ValuesReconstructState::new(io_concurrency.clone()),
6648 4 : &ctx,
6649 4 : )
6650 4 : .await;
6651 4 :
6652 4 : let images = vectored_res?;
6653 4 : assert!(images.is_empty());
6654 4 : Ok(())
6655 4 : }
6656 :
6657 : // Test that vectored get handles layer gaps correctly
6658 : // by advancing into the next ancestor timeline if required.
6659 : //
6660 : // The test generates timelines that look like the diagram below.
6661 : // We leave a gap in one of the L1 layers at `gap_at_key` (`/` in the diagram).
6662 : // The reconstruct data for that key lies in the ancestor timeline (`X` in the diagram).
6663 : //
6664 : // ```
6665 : //-------------------------------+
6666 : // ... |
6667 : // [ L1 ] |
6668 : // [ / L1 ] | Child Timeline
6669 : // ... |
6670 : // ------------------------------+
6671 : // [ X L1 ] | Parent Timeline
6672 : // ------------------------------+
6673 : // ```
6674 : #[tokio::test]
6675 4 : async fn test_get_vectored_key_gap() -> anyhow::Result<()> {
6676 4 : let tenant_conf = TenantConf {
6677 4 : // Make compaction deterministic
6678 4 : gc_period: Duration::ZERO,
6679 4 : compaction_period: Duration::ZERO,
6680 4 : // Encourage creation of L1 layers
6681 4 : checkpoint_distance: 16 * 1024,
6682 4 : compaction_target_size: 8 * 1024,
6683 4 : ..TenantConf::default()
6684 4 : };
6685 4 :
6686 4 : let harness = TenantHarness::create_custom(
6687 4 : "test_get_vectored_key_gap",
6688 4 : tenant_conf,
6689 4 : TenantId::generate(),
6690 4 : ShardIdentity::unsharded(),
6691 4 : Generation::new(0xdeadbeef),
6692 4 : )
6693 4 : .await?;
6694 4 : let (tenant, ctx) = harness.load().await;
6695 4 : let io_concurrency = IoConcurrency::spawn_for_test();
6696 4 :
6697 4 : let mut current_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6698 4 : let gap_at_key = current_key.add(100);
6699 4 : let mut current_lsn = Lsn(0x10);
6700 4 :
6701 4 : const KEY_COUNT: usize = 10_000;
6702 4 :
6703 4 : let timeline_id = TimelineId::generate();
6704 4 : let current_timeline = tenant
6705 4 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
6706 4 : .await?;
6707 4 :
6708 4 : current_lsn += 0x100;
6709 4 :
6710 4 : let mut writer = current_timeline.writer().await;
6711 4 : writer
6712 4 : .put(
6713 4 : gap_at_key,
6714 4 : current_lsn,
6715 4 : &Value::Image(test_img(&format!("{} at {}", gap_at_key, current_lsn))),
6716 4 : &ctx,
6717 4 : )
6718 4 : .await?;
6719 4 : writer.finish_write(current_lsn);
6720 4 : drop(writer);
6721 4 :
6722 4 : let mut latest_lsns = HashMap::new();
6723 4 : latest_lsns.insert(gap_at_key, current_lsn);
6724 4 :
6725 4 : current_timeline.freeze_and_flush().await?;
6726 4 :
6727 4 : let child_timeline_id = TimelineId::generate();
6728 4 :
6729 4 : tenant
6730 4 : .branch_timeline_test(
6731 4 : ¤t_timeline,
6732 4 : child_timeline_id,
6733 4 : Some(current_lsn),
6734 4 : &ctx,
6735 4 : )
6736 4 : .await?;
6737 4 : let child_timeline = tenant
6738 4 : .get_timeline(child_timeline_id, true)
6739 4 : .expect("Should have the branched timeline");
6740 4 :
6741 40004 : for i in 0..KEY_COUNT {
6742 40000 : if current_key == gap_at_key {
6743 4 : current_key = current_key.next();
6744 4 : continue;
6745 39996 : }
6746 39996 :
6747 39996 : current_lsn += 0x10;
6748 4 :
6749 39996 : let mut writer = child_timeline.writer().await;
6750 39996 : writer
6751 39996 : .put(
6752 39996 : current_key,
6753 39996 : current_lsn,
6754 39996 : &Value::Image(test_img(&format!("{} at {}", current_key, current_lsn))),
6755 39996 : &ctx,
6756 39996 : )
6757 39996 : .await?;
6758 39996 : writer.finish_write(current_lsn);
6759 39996 : drop(writer);
6760 39996 :
6761 39996 : latest_lsns.insert(current_key, current_lsn);
6762 39996 : current_key = current_key.next();
6763 39996 :
6764 39996 : // Flush every now and then to encourage layer file creation.
6765 39996 : if i % 500 == 0 {
6766 80 : child_timeline.freeze_and_flush().await?;
6767 39916 : }
6768 4 : }
6769 4 :
6770 4 : child_timeline.freeze_and_flush().await?;
6771 4 : let mut flags = EnumSet::new();
6772 4 : flags.insert(CompactFlags::ForceRepartition);
6773 4 : child_timeline
6774 4 : .compact(&CancellationToken::new(), flags, &ctx)
6775 4 : .await?;
6776 4 :
6777 4 : let key_near_end = {
6778 4 : let mut tmp = current_key;
6779 4 : tmp.field6 -= 10;
6780 4 : tmp
6781 4 : };
6782 4 :
6783 4 : let key_near_gap = {
6784 4 : let mut tmp = gap_at_key;
6785 4 : tmp.field6 -= 10;
6786 4 : tmp
6787 4 : };
6788 4 :
6789 4 : let read = KeySpace {
6790 4 : ranges: vec![key_near_gap..gap_at_key.next(), key_near_end..current_key],
6791 4 : };
6792 4 : let results = child_timeline
6793 4 : .get_vectored_impl(
6794 4 : read.clone(),
6795 4 : current_lsn,
6796 4 : &mut ValuesReconstructState::new(io_concurrency.clone()),
6797 4 : &ctx,
6798 4 : )
6799 4 : .await?;
6800 4 :
6801 88 : for (key, img_res) in results {
6802 84 : let expected = test_img(&format!("{} at {}", key, latest_lsns[&key]));
6803 84 : assert_eq!(img_res?, expected);
6804 4 : }
6805 4 :
6806 4 : Ok(())
6807 4 : }
6808 :
6809 : // Test that vectored get descends into ancestor timelines correctly and
6810 : // does not return an image that's newer than requested.
6811 : //
6812 : // The diagram below ilustrates an interesting case. We have a parent timeline
6813 : // (top of the Lsn range) and a child timeline. The request key cannot be reconstructed
6814 : // from the child timeline, so the parent timeline must be visited. When advacing into
6815 : // the child timeline, the read path needs to remember what the requested Lsn was in
6816 : // order to avoid returning an image that's too new. The test below constructs such
6817 : // a timeline setup and does a few queries around the Lsn of each page image.
6818 : // ```
6819 : // LSN
6820 : // ^
6821 : // |
6822 : // |
6823 : // 500 | --------------------------------------> branch point
6824 : // 400 | X
6825 : // 300 | X
6826 : // 200 | --------------------------------------> requested lsn
6827 : // 100 | X
6828 : // |---------------------------------------> Key
6829 : // |
6830 : // ------> requested key
6831 : //
6832 : // Legend:
6833 : // * X - page images
6834 : // ```
6835 : #[tokio::test]
6836 4 : async fn test_get_vectored_ancestor_descent() -> anyhow::Result<()> {
6837 4 : let harness = TenantHarness::create("test_get_vectored_on_lsn_axis").await?;
6838 4 : let (tenant, ctx) = harness.load().await;
6839 4 : let io_concurrency = IoConcurrency::spawn_for_test();
6840 4 :
6841 4 : let start_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6842 4 : let end_key = start_key.add(1000);
6843 4 : let child_gap_at_key = start_key.add(500);
6844 4 : let mut parent_gap_lsns: BTreeMap<Lsn, String> = BTreeMap::new();
6845 4 :
6846 4 : let mut current_lsn = Lsn(0x10);
6847 4 :
6848 4 : let timeline_id = TimelineId::generate();
6849 4 : let parent_timeline = tenant
6850 4 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
6851 4 : .await?;
6852 4 :
6853 4 : current_lsn += 0x100;
6854 4 :
6855 16 : for _ in 0..3 {
6856 12 : let mut key = start_key;
6857 12012 : while key < end_key {
6858 12000 : current_lsn += 0x10;
6859 12000 :
6860 12000 : let image_value = format!("{} at {}", child_gap_at_key, current_lsn);
6861 4 :
6862 12000 : let mut writer = parent_timeline.writer().await;
6863 12000 : writer
6864 12000 : .put(
6865 12000 : key,
6866 12000 : current_lsn,
6867 12000 : &Value::Image(test_img(&image_value)),
6868 12000 : &ctx,
6869 12000 : )
6870 12000 : .await?;
6871 12000 : writer.finish_write(current_lsn);
6872 12000 :
6873 12000 : if key == child_gap_at_key {
6874 12 : parent_gap_lsns.insert(current_lsn, image_value);
6875 11988 : }
6876 4 :
6877 12000 : key = key.next();
6878 4 : }
6879 4 :
6880 12 : parent_timeline.freeze_and_flush().await?;
6881 4 : }
6882 4 :
6883 4 : let child_timeline_id = TimelineId::generate();
6884 4 :
6885 4 : let child_timeline = tenant
6886 4 : .branch_timeline_test(&parent_timeline, child_timeline_id, Some(current_lsn), &ctx)
6887 4 : .await?;
6888 4 :
6889 4 : let mut key = start_key;
6890 4004 : while key < end_key {
6891 4000 : if key == child_gap_at_key {
6892 4 : key = key.next();
6893 4 : continue;
6894 3996 : }
6895 3996 :
6896 3996 : current_lsn += 0x10;
6897 4 :
6898 3996 : let mut writer = child_timeline.writer().await;
6899 3996 : writer
6900 3996 : .put(
6901 3996 : key,
6902 3996 : current_lsn,
6903 3996 : &Value::Image(test_img(&format!("{} at {}", key, current_lsn))),
6904 3996 : &ctx,
6905 3996 : )
6906 3996 : .await?;
6907 3996 : writer.finish_write(current_lsn);
6908 3996 :
6909 3996 : key = key.next();
6910 4 : }
6911 4 :
6912 4 : child_timeline.freeze_and_flush().await?;
6913 4 :
6914 4 : let lsn_offsets: [i64; 5] = [-10, -1, 0, 1, 10];
6915 4 : let mut query_lsns = Vec::new();
6916 12 : for image_lsn in parent_gap_lsns.keys().rev() {
6917 72 : for offset in lsn_offsets {
6918 60 : query_lsns.push(Lsn(image_lsn
6919 60 : .0
6920 60 : .checked_add_signed(offset)
6921 60 : .expect("Shouldn't overflow")));
6922 60 : }
6923 4 : }
6924 4 :
6925 64 : for query_lsn in query_lsns {
6926 60 : let results = child_timeline
6927 60 : .get_vectored_impl(
6928 60 : KeySpace {
6929 60 : ranges: vec![child_gap_at_key..child_gap_at_key.next()],
6930 60 : },
6931 60 : query_lsn,
6932 60 : &mut ValuesReconstructState::new(io_concurrency.clone()),
6933 60 : &ctx,
6934 60 : )
6935 60 : .await;
6936 4 :
6937 60 : let expected_item = parent_gap_lsns
6938 60 : .iter()
6939 60 : .rev()
6940 136 : .find(|(lsn, _)| **lsn <= query_lsn);
6941 60 :
6942 60 : info!(
6943 4 : "Doing vectored read at LSN {}. Expecting image to be: {:?}",
6944 4 : query_lsn, expected_item
6945 4 : );
6946 4 :
6947 60 : match expected_item {
6948 52 : Some((_, img_value)) => {
6949 52 : let key_results = results.expect("No vectored get error expected");
6950 52 : let key_result = &key_results[&child_gap_at_key];
6951 52 : let returned_img = key_result
6952 52 : .as_ref()
6953 52 : .expect("No page reconstruct error expected");
6954 52 :
6955 52 : info!(
6956 4 : "Vectored read at LSN {} returned image {}",
6957 0 : query_lsn,
6958 0 : std::str::from_utf8(returned_img)?
6959 4 : );
6960 52 : assert_eq!(*returned_img, test_img(img_value));
6961 4 : }
6962 4 : None => {
6963 8 : assert!(matches!(results, Err(GetVectoredError::MissingKey(_))));
6964 4 : }
6965 4 : }
6966 4 : }
6967 4 :
6968 4 : Ok(())
6969 4 : }
6970 :
6971 : #[tokio::test]
6972 4 : async fn test_random_updates() -> anyhow::Result<()> {
6973 4 : let names_algorithms = [
6974 4 : ("test_random_updates_legacy", CompactionAlgorithm::Legacy),
6975 4 : ("test_random_updates_tiered", CompactionAlgorithm::Tiered),
6976 4 : ];
6977 12 : for (name, algorithm) in names_algorithms {
6978 8 : test_random_updates_algorithm(name, algorithm).await?;
6979 4 : }
6980 4 : Ok(())
6981 4 : }
6982 :
6983 8 : async fn test_random_updates_algorithm(
6984 8 : name: &'static str,
6985 8 : compaction_algorithm: CompactionAlgorithm,
6986 8 : ) -> anyhow::Result<()> {
6987 8 : let mut harness = TenantHarness::create(name).await?;
6988 8 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
6989 8 : kind: compaction_algorithm,
6990 8 : };
6991 8 : let (tenant, ctx) = harness.load().await;
6992 8 : let tline = tenant
6993 8 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6994 8 : .await?;
6995 :
6996 : const NUM_KEYS: usize = 1000;
6997 8 : let cancel = CancellationToken::new();
6998 8 :
6999 8 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7000 8 : let mut test_key_end = test_key;
7001 8 : test_key_end.field6 = NUM_KEYS as u32;
7002 8 : tline.add_extra_test_dense_keyspace(KeySpace::single(test_key..test_key_end));
7003 8 :
7004 8 : let mut keyspace = KeySpaceAccum::new();
7005 8 :
7006 8 : // Track when each page was last modified. Used to assert that
7007 8 : // a read sees the latest page version.
7008 8 : let mut updated = [Lsn(0); NUM_KEYS];
7009 8 :
7010 8 : let mut lsn = Lsn(0x10);
7011 : #[allow(clippy::needless_range_loop)]
7012 8008 : for blknum in 0..NUM_KEYS {
7013 8000 : lsn = Lsn(lsn.0 + 0x10);
7014 8000 : test_key.field6 = blknum as u32;
7015 8000 : let mut writer = tline.writer().await;
7016 8000 : writer
7017 8000 : .put(
7018 8000 : test_key,
7019 8000 : lsn,
7020 8000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7021 8000 : &ctx,
7022 8000 : )
7023 8000 : .await?;
7024 8000 : writer.finish_write(lsn);
7025 8000 : updated[blknum] = lsn;
7026 8000 : drop(writer);
7027 8000 :
7028 8000 : keyspace.add_key(test_key);
7029 : }
7030 :
7031 408 : for _ in 0..50 {
7032 400400 : for _ in 0..NUM_KEYS {
7033 400000 : lsn = Lsn(lsn.0 + 0x10);
7034 400000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7035 400000 : test_key.field6 = blknum as u32;
7036 400000 : let mut writer = tline.writer().await;
7037 400000 : writer
7038 400000 : .put(
7039 400000 : test_key,
7040 400000 : lsn,
7041 400000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7042 400000 : &ctx,
7043 400000 : )
7044 400000 : .await?;
7045 400000 : writer.finish_write(lsn);
7046 400000 : drop(writer);
7047 400000 : updated[blknum] = lsn;
7048 : }
7049 :
7050 : // Read all the blocks
7051 400000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7052 400000 : test_key.field6 = blknum as u32;
7053 400000 : assert_eq!(
7054 400000 : tline.get(test_key, lsn, &ctx).await?,
7055 400000 : test_img(&format!("{} at {}", blknum, last_lsn))
7056 : );
7057 : }
7058 :
7059 : // Perform a cycle of flush, and GC
7060 400 : tline.freeze_and_flush().await?;
7061 400 : tenant
7062 400 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7063 400 : .await?;
7064 : }
7065 :
7066 8 : Ok(())
7067 8 : }
7068 :
7069 : #[tokio::test]
7070 4 : async fn test_traverse_branches() -> anyhow::Result<()> {
7071 4 : let (tenant, ctx) = TenantHarness::create("test_traverse_branches")
7072 4 : .await?
7073 4 : .load()
7074 4 : .await;
7075 4 : let mut tline = tenant
7076 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7077 4 : .await?;
7078 4 :
7079 4 : const NUM_KEYS: usize = 1000;
7080 4 :
7081 4 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7082 4 :
7083 4 : let mut keyspace = KeySpaceAccum::new();
7084 4 :
7085 4 : let cancel = CancellationToken::new();
7086 4 :
7087 4 : // Track when each page was last modified. Used to assert that
7088 4 : // a read sees the latest page version.
7089 4 : let mut updated = [Lsn(0); NUM_KEYS];
7090 4 :
7091 4 : let mut lsn = Lsn(0x10);
7092 4 : #[allow(clippy::needless_range_loop)]
7093 4004 : for blknum in 0..NUM_KEYS {
7094 4000 : lsn = Lsn(lsn.0 + 0x10);
7095 4000 : test_key.field6 = blknum as u32;
7096 4000 : let mut writer = tline.writer().await;
7097 4000 : writer
7098 4000 : .put(
7099 4000 : test_key,
7100 4000 : lsn,
7101 4000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7102 4000 : &ctx,
7103 4000 : )
7104 4000 : .await?;
7105 4000 : writer.finish_write(lsn);
7106 4000 : updated[blknum] = lsn;
7107 4000 : drop(writer);
7108 4000 :
7109 4000 : keyspace.add_key(test_key);
7110 4 : }
7111 4 :
7112 204 : for _ in 0..50 {
7113 200 : let new_tline_id = TimelineId::generate();
7114 200 : tenant
7115 200 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7116 200 : .await?;
7117 200 : tline = tenant
7118 200 : .get_timeline(new_tline_id, true)
7119 200 : .expect("Should have the branched timeline");
7120 4 :
7121 200200 : for _ in 0..NUM_KEYS {
7122 200000 : lsn = Lsn(lsn.0 + 0x10);
7123 200000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7124 200000 : test_key.field6 = blknum as u32;
7125 200000 : let mut writer = tline.writer().await;
7126 200000 : writer
7127 200000 : .put(
7128 200000 : test_key,
7129 200000 : lsn,
7130 200000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7131 200000 : &ctx,
7132 200000 : )
7133 200000 : .await?;
7134 200000 : println!("updating {} at {}", blknum, lsn);
7135 200000 : writer.finish_write(lsn);
7136 200000 : drop(writer);
7137 200000 : updated[blknum] = lsn;
7138 4 : }
7139 4 :
7140 4 : // Read all the blocks
7141 200000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7142 200000 : test_key.field6 = blknum as u32;
7143 200000 : assert_eq!(
7144 200000 : tline.get(test_key, lsn, &ctx).await?,
7145 200000 : test_img(&format!("{} at {}", blknum, last_lsn))
7146 4 : );
7147 4 : }
7148 4 :
7149 4 : // Perform a cycle of flush, compact, and GC
7150 200 : tline.freeze_and_flush().await?;
7151 200 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
7152 200 : tenant
7153 200 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7154 200 : .await?;
7155 4 : }
7156 4 :
7157 4 : Ok(())
7158 4 : }
7159 :
7160 : #[tokio::test]
7161 4 : async fn test_traverse_ancestors() -> anyhow::Result<()> {
7162 4 : let (tenant, ctx) = TenantHarness::create("test_traverse_ancestors")
7163 4 : .await?
7164 4 : .load()
7165 4 : .await;
7166 4 : let mut tline = tenant
7167 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7168 4 : .await?;
7169 4 :
7170 4 : const NUM_KEYS: usize = 100;
7171 4 : const NUM_TLINES: usize = 50;
7172 4 :
7173 4 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7174 4 : // Track page mutation lsns across different timelines.
7175 4 : let mut updated = [[Lsn(0); NUM_KEYS]; NUM_TLINES];
7176 4 :
7177 4 : let mut lsn = Lsn(0x10);
7178 4 :
7179 4 : #[allow(clippy::needless_range_loop)]
7180 204 : for idx in 0..NUM_TLINES {
7181 200 : let new_tline_id = TimelineId::generate();
7182 200 : tenant
7183 200 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7184 200 : .await?;
7185 200 : tline = tenant
7186 200 : .get_timeline(new_tline_id, true)
7187 200 : .expect("Should have the branched timeline");
7188 4 :
7189 20200 : for _ in 0..NUM_KEYS {
7190 20000 : lsn = Lsn(lsn.0 + 0x10);
7191 20000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7192 20000 : test_key.field6 = blknum as u32;
7193 20000 : let mut writer = tline.writer().await;
7194 20000 : writer
7195 20000 : .put(
7196 20000 : test_key,
7197 20000 : lsn,
7198 20000 : &Value::Image(test_img(&format!("{} {} at {}", idx, blknum, lsn))),
7199 20000 : &ctx,
7200 20000 : )
7201 20000 : .await?;
7202 20000 : println!("updating [{}][{}] at {}", idx, blknum, lsn);
7203 20000 : writer.finish_write(lsn);
7204 20000 : drop(writer);
7205 20000 : updated[idx][blknum] = lsn;
7206 4 : }
7207 4 : }
7208 4 :
7209 4 : // Read pages from leaf timeline across all ancestors.
7210 200 : for (idx, lsns) in updated.iter().enumerate() {
7211 20000 : for (blknum, lsn) in lsns.iter().enumerate() {
7212 4 : // Skip empty mutations.
7213 20000 : if lsn.0 == 0 {
7214 7315 : continue;
7215 12685 : }
7216 12685 : println!("checking [{idx}][{blknum}] at {lsn}");
7217 12685 : test_key.field6 = blknum as u32;
7218 12685 : assert_eq!(
7219 12685 : tline.get(test_key, *lsn, &ctx).await?,
7220 12685 : test_img(&format!("{idx} {blknum} at {lsn}"))
7221 4 : );
7222 4 : }
7223 4 : }
7224 4 : Ok(())
7225 4 : }
7226 :
7227 : #[tokio::test]
7228 4 : async fn test_write_at_initdb_lsn_takes_optimization_code_path() -> anyhow::Result<()> {
7229 4 : let (tenant, ctx) = TenantHarness::create("test_empty_test_timeline_is_usable")
7230 4 : .await?
7231 4 : .load()
7232 4 : .await;
7233 4 :
7234 4 : let initdb_lsn = Lsn(0x20);
7235 4 : let utline = tenant
7236 4 : .create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)
7237 4 : .await?;
7238 4 : let tline = utline.raw_timeline().unwrap();
7239 4 :
7240 4 : // Spawn flush loop now so that we can set the `expect_initdb_optimization`
7241 4 : tline.maybe_spawn_flush_loop();
7242 4 :
7243 4 : // Make sure the timeline has the minimum set of required keys for operation.
7244 4 : // The only operation you can always do on an empty timeline is to `put` new data.
7245 4 : // Except if you `put` at `initdb_lsn`.
7246 4 : // In that case, there's an optimization to directly create image layers instead of delta layers.
7247 4 : // It uses `repartition()`, which assumes some keys to be present.
7248 4 : // Let's make sure the test timeline can handle that case.
7249 4 : {
7250 4 : let mut state = tline.flush_loop_state.lock().unwrap();
7251 4 : assert_eq!(
7252 4 : timeline::FlushLoopState::Running {
7253 4 : expect_initdb_optimization: false,
7254 4 : initdb_optimization_count: 0,
7255 4 : },
7256 4 : *state
7257 4 : );
7258 4 : *state = timeline::FlushLoopState::Running {
7259 4 : expect_initdb_optimization: true,
7260 4 : initdb_optimization_count: 0,
7261 4 : };
7262 4 : }
7263 4 :
7264 4 : // Make writes at the initdb_lsn. When we flush it below, it should be handled by the optimization.
7265 4 : // As explained above, the optimization requires some keys to be present.
7266 4 : // As per `create_empty_timeline` documentation, use init_empty to set them.
7267 4 : // This is what `create_test_timeline` does, by the way.
7268 4 : let mut modification = tline.begin_modification(initdb_lsn);
7269 4 : modification
7270 4 : .init_empty_test_timeline()
7271 4 : .context("init_empty_test_timeline")?;
7272 4 : modification
7273 4 : .commit(&ctx)
7274 4 : .await
7275 4 : .context("commit init_empty_test_timeline modification")?;
7276 4 :
7277 4 : // Do the flush. The flush code will check the expectations that we set above.
7278 4 : tline.freeze_and_flush().await?;
7279 4 :
7280 4 : // assert freeze_and_flush exercised the initdb optimization
7281 4 : {
7282 4 : let state = tline.flush_loop_state.lock().unwrap();
7283 4 : let timeline::FlushLoopState::Running {
7284 4 : expect_initdb_optimization,
7285 4 : initdb_optimization_count,
7286 4 : } = *state
7287 4 : else {
7288 4 : panic!("unexpected state: {:?}", *state);
7289 4 : };
7290 4 : assert!(expect_initdb_optimization);
7291 4 : assert!(initdb_optimization_count > 0);
7292 4 : }
7293 4 : Ok(())
7294 4 : }
7295 :
7296 : #[tokio::test]
7297 4 : async fn test_create_guard_crash() -> anyhow::Result<()> {
7298 4 : let name = "test_create_guard_crash";
7299 4 : let harness = TenantHarness::create(name).await?;
7300 4 : {
7301 4 : let (tenant, ctx) = harness.load().await;
7302 4 : let tline = tenant
7303 4 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
7304 4 : .await?;
7305 4 : // Leave the timeline ID in [`Tenant::timelines_creating`] to exclude attempting to create it again
7306 4 : let raw_tline = tline.raw_timeline().unwrap();
7307 4 : raw_tline
7308 4 : .shutdown(super::timeline::ShutdownMode::Hard)
7309 4 : .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))
7310 4 : .await;
7311 4 : std::mem::forget(tline);
7312 4 : }
7313 4 :
7314 4 : let (tenant, _) = harness.load().await;
7315 4 : match tenant.get_timeline(TIMELINE_ID, false) {
7316 4 : Ok(_) => panic!("timeline should've been removed during load"),
7317 4 : Err(e) => {
7318 4 : assert_eq!(
7319 4 : e,
7320 4 : GetTimelineError::NotFound {
7321 4 : tenant_id: tenant.tenant_shard_id,
7322 4 : timeline_id: TIMELINE_ID,
7323 4 : }
7324 4 : )
7325 4 : }
7326 4 : }
7327 4 :
7328 4 : assert!(!harness
7329 4 : .conf
7330 4 : .timeline_path(&tenant.tenant_shard_id, &TIMELINE_ID)
7331 4 : .exists());
7332 4 :
7333 4 : Ok(())
7334 4 : }
7335 :
7336 : #[tokio::test]
7337 4 : async fn test_read_at_max_lsn() -> anyhow::Result<()> {
7338 4 : let names_algorithms = [
7339 4 : ("test_read_at_max_lsn_legacy", CompactionAlgorithm::Legacy),
7340 4 : ("test_read_at_max_lsn_tiered", CompactionAlgorithm::Tiered),
7341 4 : ];
7342 12 : for (name, algorithm) in names_algorithms {
7343 8 : test_read_at_max_lsn_algorithm(name, algorithm).await?;
7344 4 : }
7345 4 : Ok(())
7346 4 : }
7347 :
7348 8 : async fn test_read_at_max_lsn_algorithm(
7349 8 : name: &'static str,
7350 8 : compaction_algorithm: CompactionAlgorithm,
7351 8 : ) -> anyhow::Result<()> {
7352 8 : let mut harness = TenantHarness::create(name).await?;
7353 8 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
7354 8 : kind: compaction_algorithm,
7355 8 : };
7356 8 : let (tenant, ctx) = harness.load().await;
7357 8 : let tline = tenant
7358 8 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
7359 8 : .await?;
7360 :
7361 8 : let lsn = Lsn(0x10);
7362 8 : let compact = false;
7363 8 : bulk_insert_maybe_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000, compact).await?;
7364 :
7365 8 : let test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7366 8 : let read_lsn = Lsn(u64::MAX - 1);
7367 :
7368 8 : let result = tline.get(test_key, read_lsn, &ctx).await;
7369 8 : assert!(result.is_ok(), "result is not Ok: {}", result.unwrap_err());
7370 :
7371 8 : Ok(())
7372 8 : }
7373 :
7374 : #[tokio::test]
7375 4 : async fn test_metadata_scan() -> anyhow::Result<()> {
7376 4 : let harness = TenantHarness::create("test_metadata_scan").await?;
7377 4 : let (tenant, ctx) = harness.load().await;
7378 4 : let io_concurrency = IoConcurrency::spawn_for_test();
7379 4 : let tline = tenant
7380 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7381 4 : .await?;
7382 4 :
7383 4 : const NUM_KEYS: usize = 1000;
7384 4 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
7385 4 :
7386 4 : let cancel = CancellationToken::new();
7387 4 :
7388 4 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7389 4 : base_key.field1 = AUX_KEY_PREFIX;
7390 4 : let mut test_key = base_key;
7391 4 :
7392 4 : // Track when each page was last modified. Used to assert that
7393 4 : // a read sees the latest page version.
7394 4 : let mut updated = [Lsn(0); NUM_KEYS];
7395 4 :
7396 4 : let mut lsn = Lsn(0x10);
7397 4 : #[allow(clippy::needless_range_loop)]
7398 4004 : for blknum in 0..NUM_KEYS {
7399 4000 : lsn = Lsn(lsn.0 + 0x10);
7400 4000 : test_key.field6 = (blknum * STEP) as u32;
7401 4000 : let mut writer = tline.writer().await;
7402 4000 : writer
7403 4000 : .put(
7404 4000 : test_key,
7405 4000 : lsn,
7406 4000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7407 4000 : &ctx,
7408 4000 : )
7409 4000 : .await?;
7410 4000 : writer.finish_write(lsn);
7411 4000 : updated[blknum] = lsn;
7412 4000 : drop(writer);
7413 4 : }
7414 4 :
7415 4 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
7416 4 :
7417 48 : for iter in 0..=10 {
7418 4 : // Read all the blocks
7419 44000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7420 44000 : test_key.field6 = (blknum * STEP) as u32;
7421 44000 : assert_eq!(
7422 44000 : tline.get(test_key, lsn, &ctx).await?,
7423 44000 : test_img(&format!("{} at {}", blknum, last_lsn))
7424 4 : );
7425 4 : }
7426 4 :
7427 44 : let mut cnt = 0;
7428 44000 : for (key, value) in tline
7429 44 : .get_vectored_impl(
7430 44 : keyspace.clone(),
7431 44 : lsn,
7432 44 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7433 44 : &ctx,
7434 44 : )
7435 44 : .await?
7436 4 : {
7437 44000 : let blknum = key.field6 as usize;
7438 44000 : let value = value?;
7439 44000 : assert!(blknum % STEP == 0);
7440 44000 : let blknum = blknum / STEP;
7441 44000 : assert_eq!(
7442 44000 : value,
7443 44000 : test_img(&format!("{} at {}", blknum, updated[blknum]))
7444 44000 : );
7445 44000 : cnt += 1;
7446 4 : }
7447 4 :
7448 44 : assert_eq!(cnt, NUM_KEYS);
7449 4 :
7450 44044 : for _ in 0..NUM_KEYS {
7451 44000 : lsn = Lsn(lsn.0 + 0x10);
7452 44000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7453 44000 : test_key.field6 = (blknum * STEP) as u32;
7454 44000 : let mut writer = tline.writer().await;
7455 44000 : writer
7456 44000 : .put(
7457 44000 : test_key,
7458 44000 : lsn,
7459 44000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7460 44000 : &ctx,
7461 44000 : )
7462 44000 : .await?;
7463 44000 : writer.finish_write(lsn);
7464 44000 : drop(writer);
7465 44000 : updated[blknum] = lsn;
7466 4 : }
7467 4 :
7468 4 : // Perform two cycles of flush, compact, and GC
7469 132 : for round in 0..2 {
7470 88 : tline.freeze_and_flush().await?;
7471 88 : tline
7472 88 : .compact(
7473 88 : &cancel,
7474 88 : if iter % 5 == 0 && round == 0 {
7475 12 : let mut flags = EnumSet::new();
7476 12 : flags.insert(CompactFlags::ForceImageLayerCreation);
7477 12 : flags.insert(CompactFlags::ForceRepartition);
7478 12 : flags
7479 4 : } else {
7480 76 : EnumSet::empty()
7481 4 : },
7482 88 : &ctx,
7483 88 : )
7484 88 : .await?;
7485 88 : tenant
7486 88 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7487 88 : .await?;
7488 4 : }
7489 4 : }
7490 4 :
7491 4 : Ok(())
7492 4 : }
7493 :
7494 : #[tokio::test]
7495 4 : async fn test_metadata_compaction_trigger() -> anyhow::Result<()> {
7496 4 : let harness = TenantHarness::create("test_metadata_compaction_trigger").await?;
7497 4 : let (tenant, ctx) = harness.load().await;
7498 4 : let tline = tenant
7499 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7500 4 : .await?;
7501 4 :
7502 4 : let cancel = CancellationToken::new();
7503 4 :
7504 4 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7505 4 : base_key.field1 = AUX_KEY_PREFIX;
7506 4 : let test_key = base_key;
7507 4 : let mut lsn = Lsn(0x10);
7508 4 :
7509 84 : for _ in 0..20 {
7510 80 : lsn = Lsn(lsn.0 + 0x10);
7511 80 : let mut writer = tline.writer().await;
7512 80 : writer
7513 80 : .put(
7514 80 : test_key,
7515 80 : lsn,
7516 80 : &Value::Image(test_img(&format!("{} at {}", 0, lsn))),
7517 80 : &ctx,
7518 80 : )
7519 80 : .await?;
7520 80 : writer.finish_write(lsn);
7521 80 : drop(writer);
7522 80 : tline.freeze_and_flush().await?; // force create a delta layer
7523 4 : }
7524 4 :
7525 4 : let before_num_l0_delta_files =
7526 4 : tline.layers.read().await.layer_map()?.level0_deltas().len();
7527 4 :
7528 4 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
7529 4 :
7530 4 : let after_num_l0_delta_files = tline.layers.read().await.layer_map()?.level0_deltas().len();
7531 4 :
7532 4 : assert!(after_num_l0_delta_files < before_num_l0_delta_files, "after_num_l0_delta_files={after_num_l0_delta_files}, before_num_l0_delta_files={before_num_l0_delta_files}");
7533 4 :
7534 4 : assert_eq!(
7535 4 : tline.get(test_key, lsn, &ctx).await?,
7536 4 : test_img(&format!("{} at {}", 0, lsn))
7537 4 : );
7538 4 :
7539 4 : Ok(())
7540 4 : }
7541 :
7542 : #[tokio::test]
7543 4 : async fn test_aux_file_e2e() {
7544 4 : let harness = TenantHarness::create("test_aux_file_e2e").await.unwrap();
7545 4 :
7546 4 : let (tenant, ctx) = harness.load().await;
7547 4 : let io_concurrency = IoConcurrency::spawn_for_test();
7548 4 :
7549 4 : let mut lsn = Lsn(0x08);
7550 4 :
7551 4 : let tline: Arc<Timeline> = tenant
7552 4 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
7553 4 : .await
7554 4 : .unwrap();
7555 4 :
7556 4 : {
7557 4 : lsn += 8;
7558 4 : let mut modification = tline.begin_modification(lsn);
7559 4 : modification
7560 4 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
7561 4 : .await
7562 4 : .unwrap();
7563 4 : modification.commit(&ctx).await.unwrap();
7564 4 : }
7565 4 :
7566 4 : // we can read everything from the storage
7567 4 : let files = tline
7568 4 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
7569 4 : .await
7570 4 : .unwrap();
7571 4 : assert_eq!(
7572 4 : files.get("pg_logical/mappings/test1"),
7573 4 : Some(&bytes::Bytes::from_static(b"first"))
7574 4 : );
7575 4 :
7576 4 : {
7577 4 : lsn += 8;
7578 4 : let mut modification = tline.begin_modification(lsn);
7579 4 : modification
7580 4 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
7581 4 : .await
7582 4 : .unwrap();
7583 4 : modification.commit(&ctx).await.unwrap();
7584 4 : }
7585 4 :
7586 4 : let files = tline
7587 4 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
7588 4 : .await
7589 4 : .unwrap();
7590 4 : assert_eq!(
7591 4 : files.get("pg_logical/mappings/test2"),
7592 4 : Some(&bytes::Bytes::from_static(b"second"))
7593 4 : );
7594 4 :
7595 4 : let child = tenant
7596 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(lsn), &ctx)
7597 4 : .await
7598 4 : .unwrap();
7599 4 :
7600 4 : let files = child
7601 4 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
7602 4 : .await
7603 4 : .unwrap();
7604 4 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
7605 4 : assert_eq!(files.get("pg_logical/mappings/test2"), None);
7606 4 : }
7607 :
7608 : #[tokio::test]
7609 4 : async fn test_metadata_image_creation() -> anyhow::Result<()> {
7610 4 : let harness = TenantHarness::create("test_metadata_image_creation").await?;
7611 4 : let (tenant, ctx) = harness.load().await;
7612 4 : let io_concurrency = IoConcurrency::spawn_for_test();
7613 4 : let tline = tenant
7614 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7615 4 : .await?;
7616 4 :
7617 4 : const NUM_KEYS: usize = 1000;
7618 4 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
7619 4 :
7620 4 : let cancel = CancellationToken::new();
7621 4 :
7622 4 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7623 4 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
7624 4 : let mut test_key = base_key;
7625 4 : let mut lsn = Lsn(0x10);
7626 4 :
7627 16 : async fn scan_with_statistics(
7628 16 : tline: &Timeline,
7629 16 : keyspace: &KeySpace,
7630 16 : lsn: Lsn,
7631 16 : ctx: &RequestContext,
7632 16 : io_concurrency: IoConcurrency,
7633 16 : ) -> anyhow::Result<(BTreeMap<Key, Result<Bytes, PageReconstructError>>, usize)> {
7634 16 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
7635 16 : let res = tline
7636 16 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
7637 16 : .await?;
7638 16 : Ok((res, reconstruct_state.get_delta_layers_visited() as usize))
7639 16 : }
7640 4 :
7641 4 : #[allow(clippy::needless_range_loop)]
7642 4004 : for blknum in 0..NUM_KEYS {
7643 4000 : lsn = Lsn(lsn.0 + 0x10);
7644 4000 : test_key.field6 = (blknum * STEP) as u32;
7645 4000 : let mut writer = tline.writer().await;
7646 4000 : writer
7647 4000 : .put(
7648 4000 : test_key,
7649 4000 : lsn,
7650 4000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7651 4000 : &ctx,
7652 4000 : )
7653 4000 : .await?;
7654 4000 : writer.finish_write(lsn);
7655 4000 : drop(writer);
7656 4 : }
7657 4 :
7658 4 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
7659 4 :
7660 44 : for iter in 1..=10 {
7661 40040 : for _ in 0..NUM_KEYS {
7662 40000 : lsn = Lsn(lsn.0 + 0x10);
7663 40000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7664 40000 : test_key.field6 = (blknum * STEP) as u32;
7665 40000 : let mut writer = tline.writer().await;
7666 40000 : writer
7667 40000 : .put(
7668 40000 : test_key,
7669 40000 : lsn,
7670 40000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7671 40000 : &ctx,
7672 40000 : )
7673 40000 : .await?;
7674 40000 : writer.finish_write(lsn);
7675 40000 : drop(writer);
7676 4 : }
7677 4 :
7678 40 : tline.freeze_and_flush().await?;
7679 4 :
7680 40 : if iter % 5 == 0 {
7681 8 : let (_, before_delta_file_accessed) =
7682 8 : scan_with_statistics(&tline, &keyspace, lsn, &ctx, io_concurrency.clone())
7683 8 : .await?;
7684 8 : tline
7685 8 : .compact(
7686 8 : &cancel,
7687 8 : {
7688 8 : let mut flags = EnumSet::new();
7689 8 : flags.insert(CompactFlags::ForceImageLayerCreation);
7690 8 : flags.insert(CompactFlags::ForceRepartition);
7691 8 : flags
7692 8 : },
7693 8 : &ctx,
7694 8 : )
7695 8 : .await?;
7696 8 : let (_, after_delta_file_accessed) =
7697 8 : scan_with_statistics(&tline, &keyspace, lsn, &ctx, io_concurrency.clone())
7698 8 : .await?;
7699 8 : assert!(after_delta_file_accessed < before_delta_file_accessed, "after_delta_file_accessed={after_delta_file_accessed}, before_delta_file_accessed={before_delta_file_accessed}");
7700 4 : // 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.
7701 8 : assert!(
7702 8 : after_delta_file_accessed <= 2,
7703 4 : "after_delta_file_accessed={after_delta_file_accessed}"
7704 4 : );
7705 32 : }
7706 4 : }
7707 4 :
7708 4 : Ok(())
7709 4 : }
7710 :
7711 : #[tokio::test]
7712 4 : async fn test_vectored_missing_data_key_reads() -> anyhow::Result<()> {
7713 4 : let harness = TenantHarness::create("test_vectored_missing_data_key_reads").await?;
7714 4 : let (tenant, ctx) = harness.load().await;
7715 4 :
7716 4 : let base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7717 4 : let base_key_child = Key::from_hex("000000000033333333444444445500000001").unwrap();
7718 4 : let base_key_nonexist = Key::from_hex("000000000033333333444444445500000002").unwrap();
7719 4 :
7720 4 : let tline = tenant
7721 4 : .create_test_timeline_with_layers(
7722 4 : TIMELINE_ID,
7723 4 : Lsn(0x10),
7724 4 : DEFAULT_PG_VERSION,
7725 4 : &ctx,
7726 4 : Vec::new(), // delta layers
7727 4 : vec![(Lsn(0x20), vec![(base_key, test_img("data key 1"))])], // image layers
7728 4 : 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
7729 4 : )
7730 4 : .await?;
7731 4 : tline.add_extra_test_dense_keyspace(KeySpace::single(base_key..(base_key_nonexist.next())));
7732 4 :
7733 4 : let child = tenant
7734 4 : .branch_timeline_test_with_layers(
7735 4 : &tline,
7736 4 : NEW_TIMELINE_ID,
7737 4 : Some(Lsn(0x20)),
7738 4 : &ctx,
7739 4 : Vec::new(), // delta layers
7740 4 : vec![(Lsn(0x30), vec![(base_key_child, test_img("data key 2"))])], // image layers
7741 4 : Lsn(0x30),
7742 4 : )
7743 4 : .await
7744 4 : .unwrap();
7745 4 :
7746 4 : let lsn = Lsn(0x30);
7747 4 :
7748 4 : // test vectored get on parent timeline
7749 4 : assert_eq!(
7750 4 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
7751 4 : Some(test_img("data key 1"))
7752 4 : );
7753 4 : assert!(get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx)
7754 4 : .await
7755 4 : .unwrap_err()
7756 4 : .is_missing_key_error());
7757 4 : assert!(
7758 4 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx)
7759 4 : .await
7760 4 : .unwrap_err()
7761 4 : .is_missing_key_error()
7762 4 : );
7763 4 :
7764 4 : // test vectored get on child timeline
7765 4 : assert_eq!(
7766 4 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
7767 4 : Some(test_img("data key 1"))
7768 4 : );
7769 4 : assert_eq!(
7770 4 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
7771 4 : Some(test_img("data key 2"))
7772 4 : );
7773 4 : assert!(
7774 4 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx)
7775 4 : .await
7776 4 : .unwrap_err()
7777 4 : .is_missing_key_error()
7778 4 : );
7779 4 :
7780 4 : Ok(())
7781 4 : }
7782 :
7783 : #[tokio::test]
7784 4 : async fn test_vectored_missing_metadata_key_reads() -> anyhow::Result<()> {
7785 4 : let harness = TenantHarness::create("test_vectored_missing_metadata_key_reads").await?;
7786 4 : let (tenant, ctx) = harness.load().await;
7787 4 : let io_concurrency = IoConcurrency::spawn_for_test();
7788 4 :
7789 4 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7790 4 : let base_key_child = Key::from_hex("620000000033333333444444445500000001").unwrap();
7791 4 : let base_key_nonexist = Key::from_hex("620000000033333333444444445500000002").unwrap();
7792 4 : let base_key_overwrite = Key::from_hex("620000000033333333444444445500000003").unwrap();
7793 4 :
7794 4 : let base_inherited_key = Key::from_hex("610000000033333333444444445500000000").unwrap();
7795 4 : let base_inherited_key_child =
7796 4 : Key::from_hex("610000000033333333444444445500000001").unwrap();
7797 4 : let base_inherited_key_nonexist =
7798 4 : Key::from_hex("610000000033333333444444445500000002").unwrap();
7799 4 : let base_inherited_key_overwrite =
7800 4 : Key::from_hex("610000000033333333444444445500000003").unwrap();
7801 4 :
7802 4 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
7803 4 : assert_eq!(base_inherited_key.field1, RELATION_SIZE_PREFIX);
7804 4 :
7805 4 : let tline = tenant
7806 4 : .create_test_timeline_with_layers(
7807 4 : TIMELINE_ID,
7808 4 : Lsn(0x10),
7809 4 : DEFAULT_PG_VERSION,
7810 4 : &ctx,
7811 4 : Vec::new(), // delta layers
7812 4 : vec![(
7813 4 : Lsn(0x20),
7814 4 : vec![
7815 4 : (base_inherited_key, test_img("metadata inherited key 1")),
7816 4 : (
7817 4 : base_inherited_key_overwrite,
7818 4 : test_img("metadata key overwrite 1a"),
7819 4 : ),
7820 4 : (base_key, test_img("metadata key 1")),
7821 4 : (base_key_overwrite, test_img("metadata key overwrite 1b")),
7822 4 : ],
7823 4 : )], // image layers
7824 4 : 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
7825 4 : )
7826 4 : .await?;
7827 4 :
7828 4 : let child = tenant
7829 4 : .branch_timeline_test_with_layers(
7830 4 : &tline,
7831 4 : NEW_TIMELINE_ID,
7832 4 : Some(Lsn(0x20)),
7833 4 : &ctx,
7834 4 : Vec::new(), // delta layers
7835 4 : vec![(
7836 4 : Lsn(0x30),
7837 4 : vec![
7838 4 : (
7839 4 : base_inherited_key_child,
7840 4 : test_img("metadata inherited key 2"),
7841 4 : ),
7842 4 : (
7843 4 : base_inherited_key_overwrite,
7844 4 : test_img("metadata key overwrite 2a"),
7845 4 : ),
7846 4 : (base_key_child, test_img("metadata key 2")),
7847 4 : (base_key_overwrite, test_img("metadata key overwrite 2b")),
7848 4 : ],
7849 4 : )], // image layers
7850 4 : Lsn(0x30),
7851 4 : )
7852 4 : .await
7853 4 : .unwrap();
7854 4 :
7855 4 : let lsn = Lsn(0x30);
7856 4 :
7857 4 : // test vectored get on parent timeline
7858 4 : assert_eq!(
7859 4 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
7860 4 : Some(test_img("metadata key 1"))
7861 4 : );
7862 4 : assert_eq!(
7863 4 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx).await?,
7864 4 : None
7865 4 : );
7866 4 : assert_eq!(
7867 4 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx).await?,
7868 4 : None
7869 4 : );
7870 4 : assert_eq!(
7871 4 : get_vectored_impl_wrapper(&tline, base_key_overwrite, lsn, &ctx).await?,
7872 4 : Some(test_img("metadata key overwrite 1b"))
7873 4 : );
7874 4 : assert_eq!(
7875 4 : get_vectored_impl_wrapper(&tline, base_inherited_key, lsn, &ctx).await?,
7876 4 : Some(test_img("metadata inherited key 1"))
7877 4 : );
7878 4 : assert_eq!(
7879 4 : get_vectored_impl_wrapper(&tline, base_inherited_key_child, lsn, &ctx).await?,
7880 4 : None
7881 4 : );
7882 4 : assert_eq!(
7883 4 : get_vectored_impl_wrapper(&tline, base_inherited_key_nonexist, lsn, &ctx).await?,
7884 4 : None
7885 4 : );
7886 4 : assert_eq!(
7887 4 : get_vectored_impl_wrapper(&tline, base_inherited_key_overwrite, lsn, &ctx).await?,
7888 4 : Some(test_img("metadata key overwrite 1a"))
7889 4 : );
7890 4 :
7891 4 : // test vectored get on child timeline
7892 4 : assert_eq!(
7893 4 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
7894 4 : None
7895 4 : );
7896 4 : assert_eq!(
7897 4 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
7898 4 : Some(test_img("metadata key 2"))
7899 4 : );
7900 4 : assert_eq!(
7901 4 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx).await?,
7902 4 : None
7903 4 : );
7904 4 : assert_eq!(
7905 4 : get_vectored_impl_wrapper(&child, base_inherited_key, lsn, &ctx).await?,
7906 4 : Some(test_img("metadata inherited key 1"))
7907 4 : );
7908 4 : assert_eq!(
7909 4 : get_vectored_impl_wrapper(&child, base_inherited_key_child, lsn, &ctx).await?,
7910 4 : Some(test_img("metadata inherited key 2"))
7911 4 : );
7912 4 : assert_eq!(
7913 4 : get_vectored_impl_wrapper(&child, base_inherited_key_nonexist, lsn, &ctx).await?,
7914 4 : None
7915 4 : );
7916 4 : assert_eq!(
7917 4 : get_vectored_impl_wrapper(&child, base_key_overwrite, lsn, &ctx).await?,
7918 4 : Some(test_img("metadata key overwrite 2b"))
7919 4 : );
7920 4 : assert_eq!(
7921 4 : get_vectored_impl_wrapper(&child, base_inherited_key_overwrite, lsn, &ctx).await?,
7922 4 : Some(test_img("metadata key overwrite 2a"))
7923 4 : );
7924 4 :
7925 4 : // test vectored scan on parent timeline
7926 4 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency.clone());
7927 4 : let res = tline
7928 4 : .get_vectored_impl(
7929 4 : KeySpace::single(Key::metadata_key_range()),
7930 4 : lsn,
7931 4 : &mut reconstruct_state,
7932 4 : &ctx,
7933 4 : )
7934 4 : .await?;
7935 4 :
7936 4 : assert_eq!(
7937 4 : res.into_iter()
7938 16 : .map(|(k, v)| (k, v.unwrap()))
7939 4 : .collect::<Vec<_>>(),
7940 4 : vec![
7941 4 : (base_inherited_key, test_img("metadata inherited key 1")),
7942 4 : (
7943 4 : base_inherited_key_overwrite,
7944 4 : test_img("metadata key overwrite 1a")
7945 4 : ),
7946 4 : (base_key, test_img("metadata key 1")),
7947 4 : (base_key_overwrite, test_img("metadata key overwrite 1b")),
7948 4 : ]
7949 4 : );
7950 4 :
7951 4 : // test vectored scan on child timeline
7952 4 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency.clone());
7953 4 : let res = child
7954 4 : .get_vectored_impl(
7955 4 : KeySpace::single(Key::metadata_key_range()),
7956 4 : lsn,
7957 4 : &mut reconstruct_state,
7958 4 : &ctx,
7959 4 : )
7960 4 : .await?;
7961 4 :
7962 4 : assert_eq!(
7963 4 : res.into_iter()
7964 20 : .map(|(k, v)| (k, v.unwrap()))
7965 4 : .collect::<Vec<_>>(),
7966 4 : vec![
7967 4 : (base_inherited_key, test_img("metadata inherited key 1")),
7968 4 : (
7969 4 : base_inherited_key_child,
7970 4 : test_img("metadata inherited key 2")
7971 4 : ),
7972 4 : (
7973 4 : base_inherited_key_overwrite,
7974 4 : test_img("metadata key overwrite 2a")
7975 4 : ),
7976 4 : (base_key_child, test_img("metadata key 2")),
7977 4 : (base_key_overwrite, test_img("metadata key overwrite 2b")),
7978 4 : ]
7979 4 : );
7980 4 :
7981 4 : Ok(())
7982 4 : }
7983 :
7984 112 : async fn get_vectored_impl_wrapper(
7985 112 : tline: &Arc<Timeline>,
7986 112 : key: Key,
7987 112 : lsn: Lsn,
7988 112 : ctx: &RequestContext,
7989 112 : ) -> Result<Option<Bytes>, GetVectoredError> {
7990 112 : let io_concurrency =
7991 112 : IoConcurrency::spawn_from_conf(tline.conf, tline.gate.enter().unwrap());
7992 112 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
7993 112 : let mut res = tline
7994 112 : .get_vectored_impl(
7995 112 : KeySpace::single(key..key.next()),
7996 112 : lsn,
7997 112 : &mut reconstruct_state,
7998 112 : ctx,
7999 112 : )
8000 112 : .await?;
8001 100 : Ok(res.pop_last().map(|(k, v)| {
8002 64 : assert_eq!(k, key);
8003 64 : v.unwrap()
8004 100 : }))
8005 112 : }
8006 :
8007 : #[tokio::test]
8008 4 : async fn test_metadata_tombstone_reads() -> anyhow::Result<()> {
8009 4 : let harness = TenantHarness::create("test_metadata_tombstone_reads").await?;
8010 4 : let (tenant, ctx) = harness.load().await;
8011 4 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
8012 4 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8013 4 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8014 4 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
8015 4 :
8016 4 : // We emulate the situation that the compaction algorithm creates an image layer that removes the tombstones
8017 4 : // Lsn 0x30 key0, key3, no key1+key2
8018 4 : // Lsn 0x20 key1+key2 tomestones
8019 4 : // Lsn 0x10 key1 in image, key2 in delta
8020 4 : let tline = tenant
8021 4 : .create_test_timeline_with_layers(
8022 4 : TIMELINE_ID,
8023 4 : Lsn(0x10),
8024 4 : DEFAULT_PG_VERSION,
8025 4 : &ctx,
8026 4 : // delta layers
8027 4 : vec![
8028 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8029 4 : Lsn(0x10)..Lsn(0x20),
8030 4 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8031 4 : ),
8032 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8033 4 : Lsn(0x20)..Lsn(0x30),
8034 4 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8035 4 : ),
8036 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8037 4 : Lsn(0x20)..Lsn(0x30),
8038 4 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8039 4 : ),
8040 4 : ],
8041 4 : // image layers
8042 4 : vec![
8043 4 : (Lsn(0x10), vec![(key1, test_img("metadata key 1"))]),
8044 4 : (
8045 4 : Lsn(0x30),
8046 4 : vec![
8047 4 : (key0, test_img("metadata key 0")),
8048 4 : (key3, test_img("metadata key 3")),
8049 4 : ],
8050 4 : ),
8051 4 : ],
8052 4 : Lsn(0x30),
8053 4 : )
8054 4 : .await?;
8055 4 :
8056 4 : let lsn = Lsn(0x30);
8057 4 : let old_lsn = Lsn(0x20);
8058 4 :
8059 4 : assert_eq!(
8060 4 : get_vectored_impl_wrapper(&tline, key0, lsn, &ctx).await?,
8061 4 : Some(test_img("metadata key 0"))
8062 4 : );
8063 4 : assert_eq!(
8064 4 : get_vectored_impl_wrapper(&tline, key1, lsn, &ctx).await?,
8065 4 : None,
8066 4 : );
8067 4 : assert_eq!(
8068 4 : get_vectored_impl_wrapper(&tline, key2, lsn, &ctx).await?,
8069 4 : None,
8070 4 : );
8071 4 : assert_eq!(
8072 4 : get_vectored_impl_wrapper(&tline, key1, old_lsn, &ctx).await?,
8073 4 : Some(Bytes::new()),
8074 4 : );
8075 4 : assert_eq!(
8076 4 : get_vectored_impl_wrapper(&tline, key2, old_lsn, &ctx).await?,
8077 4 : Some(Bytes::new()),
8078 4 : );
8079 4 : assert_eq!(
8080 4 : get_vectored_impl_wrapper(&tline, key3, lsn, &ctx).await?,
8081 4 : Some(test_img("metadata key 3"))
8082 4 : );
8083 4 :
8084 4 : Ok(())
8085 4 : }
8086 :
8087 : #[tokio::test]
8088 4 : async fn test_metadata_tombstone_image_creation() {
8089 4 : let harness = TenantHarness::create("test_metadata_tombstone_image_creation")
8090 4 : .await
8091 4 : .unwrap();
8092 4 : let (tenant, ctx) = harness.load().await;
8093 4 : let io_concurrency = IoConcurrency::spawn_for_test();
8094 4 :
8095 4 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
8096 4 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8097 4 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8098 4 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
8099 4 :
8100 4 : let tline = tenant
8101 4 : .create_test_timeline_with_layers(
8102 4 : TIMELINE_ID,
8103 4 : Lsn(0x10),
8104 4 : DEFAULT_PG_VERSION,
8105 4 : &ctx,
8106 4 : // delta layers
8107 4 : vec![
8108 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8109 4 : Lsn(0x10)..Lsn(0x20),
8110 4 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8111 4 : ),
8112 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8113 4 : Lsn(0x20)..Lsn(0x30),
8114 4 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8115 4 : ),
8116 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8117 4 : Lsn(0x20)..Lsn(0x30),
8118 4 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8119 4 : ),
8120 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8121 4 : Lsn(0x30)..Lsn(0x40),
8122 4 : vec![
8123 4 : (key0, Lsn(0x30), Value::Image(test_img("metadata key 0"))),
8124 4 : (key3, Lsn(0x30), Value::Image(test_img("metadata key 3"))),
8125 4 : ],
8126 4 : ),
8127 4 : ],
8128 4 : // image layers
8129 4 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
8130 4 : Lsn(0x40),
8131 4 : )
8132 4 : .await
8133 4 : .unwrap();
8134 4 :
8135 4 : let cancel = CancellationToken::new();
8136 4 :
8137 4 : tline
8138 4 : .compact(
8139 4 : &cancel,
8140 4 : {
8141 4 : let mut flags = EnumSet::new();
8142 4 : flags.insert(CompactFlags::ForceImageLayerCreation);
8143 4 : flags.insert(CompactFlags::ForceRepartition);
8144 4 : flags
8145 4 : },
8146 4 : &ctx,
8147 4 : )
8148 4 : .await
8149 4 : .unwrap();
8150 4 :
8151 4 : // Image layers are created at last_record_lsn
8152 4 : let images = tline
8153 4 : .inspect_image_layers(Lsn(0x40), &ctx, io_concurrency.clone())
8154 4 : .await
8155 4 : .unwrap()
8156 4 : .into_iter()
8157 36 : .filter(|(k, _)| k.is_metadata_key())
8158 4 : .collect::<Vec<_>>();
8159 4 : assert_eq!(images.len(), 2); // the image layer should only contain two existing keys, tombstones should be removed.
8160 4 : }
8161 :
8162 : #[tokio::test]
8163 4 : async fn test_metadata_tombstone_empty_image_creation() {
8164 4 : let harness = TenantHarness::create("test_metadata_tombstone_empty_image_creation")
8165 4 : .await
8166 4 : .unwrap();
8167 4 : let (tenant, ctx) = harness.load().await;
8168 4 : let io_concurrency = IoConcurrency::spawn_for_test();
8169 4 :
8170 4 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8171 4 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8172 4 :
8173 4 : let tline = tenant
8174 4 : .create_test_timeline_with_layers(
8175 4 : TIMELINE_ID,
8176 4 : Lsn(0x10),
8177 4 : DEFAULT_PG_VERSION,
8178 4 : &ctx,
8179 4 : // delta layers
8180 4 : vec![
8181 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8182 4 : Lsn(0x10)..Lsn(0x20),
8183 4 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8184 4 : ),
8185 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8186 4 : Lsn(0x20)..Lsn(0x30),
8187 4 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8188 4 : ),
8189 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8190 4 : Lsn(0x20)..Lsn(0x30),
8191 4 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8192 4 : ),
8193 4 : ],
8194 4 : // image layers
8195 4 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
8196 4 : Lsn(0x30),
8197 4 : )
8198 4 : .await
8199 4 : .unwrap();
8200 4 :
8201 4 : let cancel = CancellationToken::new();
8202 4 :
8203 4 : tline
8204 4 : .compact(
8205 4 : &cancel,
8206 4 : {
8207 4 : let mut flags = EnumSet::new();
8208 4 : flags.insert(CompactFlags::ForceImageLayerCreation);
8209 4 : flags.insert(CompactFlags::ForceRepartition);
8210 4 : flags
8211 4 : },
8212 4 : &ctx,
8213 4 : )
8214 4 : .await
8215 4 : .unwrap();
8216 4 :
8217 4 : // Image layers are created at last_record_lsn
8218 4 : let images = tline
8219 4 : .inspect_image_layers(Lsn(0x30), &ctx, io_concurrency.clone())
8220 4 : .await
8221 4 : .unwrap()
8222 4 : .into_iter()
8223 28 : .filter(|(k, _)| k.is_metadata_key())
8224 4 : .collect::<Vec<_>>();
8225 4 : assert_eq!(images.len(), 0); // the image layer should not contain tombstones, or it is not created
8226 4 : }
8227 :
8228 : #[tokio::test]
8229 4 : async fn test_simple_bottom_most_compaction_images() -> anyhow::Result<()> {
8230 4 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_images").await?;
8231 4 : let (tenant, ctx) = harness.load().await;
8232 4 : let io_concurrency = IoConcurrency::spawn_for_test();
8233 4 :
8234 204 : fn get_key(id: u32) -> Key {
8235 204 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8236 204 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8237 204 : key.field6 = id;
8238 204 : key
8239 204 : }
8240 4 :
8241 4 : // We create
8242 4 : // - one bottom-most image layer,
8243 4 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
8244 4 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
8245 4 : // - a delta layer D3 above the horizon.
8246 4 : //
8247 4 : // | D3 |
8248 4 : // | D1 |
8249 4 : // -| |-- gc horizon -----------------
8250 4 : // | | | D2 |
8251 4 : // --------- img layer ------------------
8252 4 : //
8253 4 : // What we should expact from this compaction is:
8254 4 : // | D3 |
8255 4 : // | Part of D1 |
8256 4 : // --------- img layer with D1+D2 at GC horizon------------------
8257 4 :
8258 4 : // img layer at 0x10
8259 4 : let img_layer = (0..10)
8260 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8261 4 : .collect_vec();
8262 4 :
8263 4 : let delta1 = vec![
8264 4 : (
8265 4 : get_key(1),
8266 4 : Lsn(0x20),
8267 4 : Value::Image(Bytes::from("value 1@0x20")),
8268 4 : ),
8269 4 : (
8270 4 : get_key(2),
8271 4 : Lsn(0x30),
8272 4 : Value::Image(Bytes::from("value 2@0x30")),
8273 4 : ),
8274 4 : (
8275 4 : get_key(3),
8276 4 : Lsn(0x40),
8277 4 : Value::Image(Bytes::from("value 3@0x40")),
8278 4 : ),
8279 4 : ];
8280 4 : let delta2 = vec![
8281 4 : (
8282 4 : get_key(5),
8283 4 : Lsn(0x20),
8284 4 : Value::Image(Bytes::from("value 5@0x20")),
8285 4 : ),
8286 4 : (
8287 4 : get_key(6),
8288 4 : Lsn(0x20),
8289 4 : Value::Image(Bytes::from("value 6@0x20")),
8290 4 : ),
8291 4 : ];
8292 4 : let delta3 = vec![
8293 4 : (
8294 4 : get_key(8),
8295 4 : Lsn(0x48),
8296 4 : Value::Image(Bytes::from("value 8@0x48")),
8297 4 : ),
8298 4 : (
8299 4 : get_key(9),
8300 4 : Lsn(0x48),
8301 4 : Value::Image(Bytes::from("value 9@0x48")),
8302 4 : ),
8303 4 : ];
8304 4 :
8305 4 : let tline = tenant
8306 4 : .create_test_timeline_with_layers(
8307 4 : TIMELINE_ID,
8308 4 : Lsn(0x10),
8309 4 : DEFAULT_PG_VERSION,
8310 4 : &ctx,
8311 4 : vec![
8312 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
8313 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
8314 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
8315 4 : ], // delta layers
8316 4 : vec![(Lsn(0x10), img_layer)], // image layers
8317 4 : Lsn(0x50),
8318 4 : )
8319 4 : .await?;
8320 4 : {
8321 4 : tline
8322 4 : .latest_gc_cutoff_lsn
8323 4 : .lock_for_write()
8324 4 : .store_and_unlock(Lsn(0x30))
8325 4 : .wait()
8326 4 : .await;
8327 4 : // Update GC info
8328 4 : let mut guard = tline.gc_info.write().unwrap();
8329 4 : guard.cutoffs.time = Lsn(0x30);
8330 4 : guard.cutoffs.space = Lsn(0x30);
8331 4 : }
8332 4 :
8333 4 : let expected_result = [
8334 4 : Bytes::from_static(b"value 0@0x10"),
8335 4 : Bytes::from_static(b"value 1@0x20"),
8336 4 : Bytes::from_static(b"value 2@0x30"),
8337 4 : Bytes::from_static(b"value 3@0x40"),
8338 4 : Bytes::from_static(b"value 4@0x10"),
8339 4 : Bytes::from_static(b"value 5@0x20"),
8340 4 : Bytes::from_static(b"value 6@0x20"),
8341 4 : Bytes::from_static(b"value 7@0x10"),
8342 4 : Bytes::from_static(b"value 8@0x48"),
8343 4 : Bytes::from_static(b"value 9@0x48"),
8344 4 : ];
8345 4 :
8346 40 : for (idx, expected) in expected_result.iter().enumerate() {
8347 40 : assert_eq!(
8348 40 : tline
8349 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8350 40 : .await
8351 40 : .unwrap(),
8352 4 : expected
8353 4 : );
8354 4 : }
8355 4 :
8356 4 : let cancel = CancellationToken::new();
8357 4 : tline
8358 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8359 4 : .await
8360 4 : .unwrap();
8361 4 :
8362 40 : for (idx, expected) in expected_result.iter().enumerate() {
8363 40 : assert_eq!(
8364 40 : tline
8365 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8366 40 : .await
8367 40 : .unwrap(),
8368 4 : expected
8369 4 : );
8370 4 : }
8371 4 :
8372 4 : // Check if the image layer at the GC horizon contains exactly what we want
8373 4 : let image_at_gc_horizon = tline
8374 4 : .inspect_image_layers(Lsn(0x30), &ctx, io_concurrency.clone())
8375 4 : .await
8376 4 : .unwrap()
8377 4 : .into_iter()
8378 68 : .filter(|(k, _)| k.is_metadata_key())
8379 4 : .collect::<Vec<_>>();
8380 4 :
8381 4 : assert_eq!(image_at_gc_horizon.len(), 10);
8382 4 : let expected_result = [
8383 4 : Bytes::from_static(b"value 0@0x10"),
8384 4 : Bytes::from_static(b"value 1@0x20"),
8385 4 : Bytes::from_static(b"value 2@0x30"),
8386 4 : Bytes::from_static(b"value 3@0x10"),
8387 4 : Bytes::from_static(b"value 4@0x10"),
8388 4 : Bytes::from_static(b"value 5@0x20"),
8389 4 : Bytes::from_static(b"value 6@0x20"),
8390 4 : Bytes::from_static(b"value 7@0x10"),
8391 4 : Bytes::from_static(b"value 8@0x10"),
8392 4 : Bytes::from_static(b"value 9@0x10"),
8393 4 : ];
8394 44 : for idx in 0..10 {
8395 40 : assert_eq!(
8396 40 : image_at_gc_horizon[idx],
8397 40 : (get_key(idx as u32), expected_result[idx].clone())
8398 40 : );
8399 4 : }
8400 4 :
8401 4 : // Check if old layers are removed / new layers have the expected LSN
8402 4 : let all_layers = inspect_and_sort(&tline, None).await;
8403 4 : assert_eq!(
8404 4 : all_layers,
8405 4 : vec![
8406 4 : // Image layer at GC horizon
8407 4 : PersistentLayerKey {
8408 4 : key_range: Key::MIN..Key::MAX,
8409 4 : lsn_range: Lsn(0x30)..Lsn(0x31),
8410 4 : is_delta: false
8411 4 : },
8412 4 : // The delta layer below the horizon
8413 4 : PersistentLayerKey {
8414 4 : key_range: get_key(3)..get_key(4),
8415 4 : lsn_range: Lsn(0x30)..Lsn(0x48),
8416 4 : is_delta: true
8417 4 : },
8418 4 : // The delta3 layer that should not be picked for the compaction
8419 4 : PersistentLayerKey {
8420 4 : key_range: get_key(8)..get_key(10),
8421 4 : lsn_range: Lsn(0x48)..Lsn(0x50),
8422 4 : is_delta: true
8423 4 : }
8424 4 : ]
8425 4 : );
8426 4 :
8427 4 : // increase GC horizon and compact again
8428 4 : {
8429 4 : tline
8430 4 : .latest_gc_cutoff_lsn
8431 4 : .lock_for_write()
8432 4 : .store_and_unlock(Lsn(0x40))
8433 4 : .wait()
8434 4 : .await;
8435 4 : // Update GC info
8436 4 : let mut guard = tline.gc_info.write().unwrap();
8437 4 : guard.cutoffs.time = Lsn(0x40);
8438 4 : guard.cutoffs.space = Lsn(0x40);
8439 4 : }
8440 4 : tline
8441 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8442 4 : .await
8443 4 : .unwrap();
8444 4 :
8445 4 : Ok(())
8446 4 : }
8447 :
8448 : #[cfg(feature = "testing")]
8449 : #[tokio::test]
8450 4 : async fn test_neon_test_record() -> anyhow::Result<()> {
8451 4 : let harness = TenantHarness::create("test_neon_test_record").await?;
8452 4 : let (tenant, ctx) = harness.load().await;
8453 4 :
8454 48 : fn get_key(id: u32) -> Key {
8455 48 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8456 48 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8457 48 : key.field6 = id;
8458 48 : key
8459 48 : }
8460 4 :
8461 4 : let delta1 = vec![
8462 4 : (
8463 4 : get_key(1),
8464 4 : Lsn(0x20),
8465 4 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
8466 4 : ),
8467 4 : (
8468 4 : get_key(1),
8469 4 : Lsn(0x30),
8470 4 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
8471 4 : ),
8472 4 : (get_key(2), Lsn(0x10), Value::Image("0x10".into())),
8473 4 : (
8474 4 : get_key(2),
8475 4 : Lsn(0x20),
8476 4 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
8477 4 : ),
8478 4 : (
8479 4 : get_key(2),
8480 4 : Lsn(0x30),
8481 4 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
8482 4 : ),
8483 4 : (get_key(3), Lsn(0x10), Value::Image("0x10".into())),
8484 4 : (
8485 4 : get_key(3),
8486 4 : Lsn(0x20),
8487 4 : Value::WalRecord(NeonWalRecord::wal_clear("c")),
8488 4 : ),
8489 4 : (get_key(4), Lsn(0x10), Value::Image("0x10".into())),
8490 4 : (
8491 4 : get_key(4),
8492 4 : Lsn(0x20),
8493 4 : Value::WalRecord(NeonWalRecord::wal_init("i")),
8494 4 : ),
8495 4 : ];
8496 4 : let image1 = vec![(get_key(1), "0x10".into())];
8497 4 :
8498 4 : let tline = tenant
8499 4 : .create_test_timeline_with_layers(
8500 4 : TIMELINE_ID,
8501 4 : Lsn(0x10),
8502 4 : DEFAULT_PG_VERSION,
8503 4 : &ctx,
8504 4 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
8505 4 : Lsn(0x10)..Lsn(0x40),
8506 4 : delta1,
8507 4 : )], // delta layers
8508 4 : vec![(Lsn(0x10), image1)], // image layers
8509 4 : Lsn(0x50),
8510 4 : )
8511 4 : .await?;
8512 4 :
8513 4 : assert_eq!(
8514 4 : tline.get(get_key(1), Lsn(0x50), &ctx).await?,
8515 4 : Bytes::from_static(b"0x10,0x20,0x30")
8516 4 : );
8517 4 : assert_eq!(
8518 4 : tline.get(get_key(2), Lsn(0x50), &ctx).await?,
8519 4 : Bytes::from_static(b"0x10,0x20,0x30")
8520 4 : );
8521 4 :
8522 4 : // Need to remove the limit of "Neon WAL redo requires base image".
8523 4 :
8524 4 : // assert_eq!(tline.get(get_key(3), Lsn(0x50), &ctx).await?, Bytes::new());
8525 4 : // assert_eq!(tline.get(get_key(4), Lsn(0x50), &ctx).await?, Bytes::new());
8526 4 :
8527 4 : Ok(())
8528 4 : }
8529 :
8530 : #[tokio::test(start_paused = true)]
8531 4 : async fn test_lsn_lease() -> anyhow::Result<()> {
8532 4 : let (tenant, ctx) = TenantHarness::create("test_lsn_lease")
8533 4 : .await
8534 4 : .unwrap()
8535 4 : .load()
8536 4 : .await;
8537 4 : // Advance to the lsn lease deadline so that GC is not blocked by
8538 4 : // initial transition into AttachedSingle.
8539 4 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
8540 4 : tokio::time::resume();
8541 4 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
8542 4 :
8543 4 : let end_lsn = Lsn(0x100);
8544 4 : let image_layers = (0x20..=0x90)
8545 4 : .step_by(0x10)
8546 32 : .map(|n| {
8547 32 : (
8548 32 : Lsn(n),
8549 32 : vec![(key, test_img(&format!("data key at {:x}", n)))],
8550 32 : )
8551 32 : })
8552 4 : .collect();
8553 4 :
8554 4 : let timeline = tenant
8555 4 : .create_test_timeline_with_layers(
8556 4 : TIMELINE_ID,
8557 4 : Lsn(0x10),
8558 4 : DEFAULT_PG_VERSION,
8559 4 : &ctx,
8560 4 : Vec::new(),
8561 4 : image_layers,
8562 4 : end_lsn,
8563 4 : )
8564 4 : .await?;
8565 4 :
8566 4 : let leased_lsns = [0x30, 0x50, 0x70];
8567 4 : let mut leases = Vec::new();
8568 12 : leased_lsns.iter().for_each(|n| {
8569 12 : leases.push(
8570 12 : timeline
8571 12 : .init_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)
8572 12 : .expect("lease request should succeed"),
8573 12 : );
8574 12 : });
8575 4 :
8576 4 : let updated_lease_0 = timeline
8577 4 : .renew_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)
8578 4 : .expect("lease renewal should succeed");
8579 4 : assert_eq!(
8580 4 : updated_lease_0.valid_until, leases[0].valid_until,
8581 4 : " Renewing with shorter lease should not change the lease."
8582 4 : );
8583 4 :
8584 4 : let updated_lease_1 = timeline
8585 4 : .renew_lsn_lease(
8586 4 : Lsn(leased_lsns[1]),
8587 4 : timeline.get_lsn_lease_length() * 2,
8588 4 : &ctx,
8589 4 : )
8590 4 : .expect("lease renewal should succeed");
8591 4 : assert!(
8592 4 : updated_lease_1.valid_until > leases[1].valid_until,
8593 4 : "Renewing with a long lease should renew lease with later expiration time."
8594 4 : );
8595 4 :
8596 4 : // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
8597 4 : info!(
8598 4 : "latest_gc_cutoff_lsn: {}",
8599 0 : *timeline.get_latest_gc_cutoff_lsn()
8600 4 : );
8601 4 : timeline.force_set_disk_consistent_lsn(end_lsn);
8602 4 :
8603 4 : let res = tenant
8604 4 : .gc_iteration(
8605 4 : Some(TIMELINE_ID),
8606 4 : 0,
8607 4 : Duration::ZERO,
8608 4 : &CancellationToken::new(),
8609 4 : &ctx,
8610 4 : )
8611 4 : .await
8612 4 : .unwrap();
8613 4 :
8614 4 : // Keeping everything <= Lsn(0x80) b/c leases:
8615 4 : // 0/10: initdb layer
8616 4 : // (0/20..=0/70).step_by(0x10): image layers added when creating the timeline.
8617 4 : assert_eq!(res.layers_needed_by_leases, 7);
8618 4 : // Keeping 0/90 b/c it is the latest layer.
8619 4 : assert_eq!(res.layers_not_updated, 1);
8620 4 : // Removed 0/80.
8621 4 : assert_eq!(res.layers_removed, 1);
8622 4 :
8623 4 : // Make lease on a already GC-ed LSN.
8624 4 : // 0/80 does not have a valid lease + is below latest_gc_cutoff
8625 4 : assert!(Lsn(0x80) < *timeline.get_latest_gc_cutoff_lsn());
8626 4 : timeline
8627 4 : .init_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx)
8628 4 : .expect_err("lease request on GC-ed LSN should fail");
8629 4 :
8630 4 : // Should still be able to renew a currently valid lease
8631 4 : // Assumption: original lease to is still valid for 0/50.
8632 4 : // (use `Timeline::init_lsn_lease` for testing so it always does validation)
8633 4 : timeline
8634 4 : .init_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)
8635 4 : .expect("lease renewal with validation should succeed");
8636 4 :
8637 4 : Ok(())
8638 4 : }
8639 :
8640 : #[cfg(feature = "testing")]
8641 : #[tokio::test]
8642 4 : async fn test_simple_bottom_most_compaction_deltas_1() -> anyhow::Result<()> {
8643 4 : test_simple_bottom_most_compaction_deltas_helper(
8644 4 : "test_simple_bottom_most_compaction_deltas_1",
8645 4 : false,
8646 4 : )
8647 4 : .await
8648 4 : }
8649 :
8650 : #[cfg(feature = "testing")]
8651 : #[tokio::test]
8652 4 : async fn test_simple_bottom_most_compaction_deltas_2() -> anyhow::Result<()> {
8653 4 : test_simple_bottom_most_compaction_deltas_helper(
8654 4 : "test_simple_bottom_most_compaction_deltas_2",
8655 4 : true,
8656 4 : )
8657 4 : .await
8658 4 : }
8659 :
8660 : #[cfg(feature = "testing")]
8661 8 : async fn test_simple_bottom_most_compaction_deltas_helper(
8662 8 : test_name: &'static str,
8663 8 : use_delta_bottom_layer: bool,
8664 8 : ) -> anyhow::Result<()> {
8665 8 : let harness = TenantHarness::create(test_name).await?;
8666 8 : let (tenant, ctx) = harness.load().await;
8667 :
8668 552 : fn get_key(id: u32) -> Key {
8669 552 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8670 552 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8671 552 : key.field6 = id;
8672 552 : key
8673 552 : }
8674 :
8675 : // We create
8676 : // - one bottom-most image layer,
8677 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
8678 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
8679 : // - a delta layer D3 above the horizon.
8680 : //
8681 : // | D3 |
8682 : // | D1 |
8683 : // -| |-- gc horizon -----------------
8684 : // | | | D2 |
8685 : // --------- img layer ------------------
8686 : //
8687 : // What we should expact from this compaction is:
8688 : // | D3 |
8689 : // | Part of D1 |
8690 : // --------- img layer with D1+D2 at GC horizon------------------
8691 :
8692 : // img layer at 0x10
8693 8 : let img_layer = (0..10)
8694 80 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8695 8 : .collect_vec();
8696 8 : // or, delta layer at 0x10 if `use_delta_bottom_layer` is true
8697 8 : let delta4 = (0..10)
8698 80 : .map(|id| {
8699 80 : (
8700 80 : get_key(id),
8701 80 : Lsn(0x08),
8702 80 : Value::WalRecord(NeonWalRecord::wal_init(format!("value {id}@0x10"))),
8703 80 : )
8704 80 : })
8705 8 : .collect_vec();
8706 8 :
8707 8 : let delta1 = vec![
8708 8 : (
8709 8 : get_key(1),
8710 8 : Lsn(0x20),
8711 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8712 8 : ),
8713 8 : (
8714 8 : get_key(2),
8715 8 : Lsn(0x30),
8716 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8717 8 : ),
8718 8 : (
8719 8 : get_key(3),
8720 8 : Lsn(0x28),
8721 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8722 8 : ),
8723 8 : (
8724 8 : get_key(3),
8725 8 : Lsn(0x30),
8726 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8727 8 : ),
8728 8 : (
8729 8 : get_key(3),
8730 8 : Lsn(0x40),
8731 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
8732 8 : ),
8733 8 : ];
8734 8 : let delta2 = vec![
8735 8 : (
8736 8 : get_key(5),
8737 8 : Lsn(0x20),
8738 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8739 8 : ),
8740 8 : (
8741 8 : get_key(6),
8742 8 : Lsn(0x20),
8743 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8744 8 : ),
8745 8 : ];
8746 8 : let delta3 = vec![
8747 8 : (
8748 8 : get_key(8),
8749 8 : Lsn(0x48),
8750 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8751 8 : ),
8752 8 : (
8753 8 : get_key(9),
8754 8 : Lsn(0x48),
8755 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8756 8 : ),
8757 8 : ];
8758 :
8759 8 : let tline = if use_delta_bottom_layer {
8760 4 : tenant
8761 4 : .create_test_timeline_with_layers(
8762 4 : TIMELINE_ID,
8763 4 : Lsn(0x08),
8764 4 : DEFAULT_PG_VERSION,
8765 4 : &ctx,
8766 4 : vec![
8767 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8768 4 : Lsn(0x08)..Lsn(0x10),
8769 4 : delta4,
8770 4 : ),
8771 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8772 4 : Lsn(0x20)..Lsn(0x48),
8773 4 : delta1,
8774 4 : ),
8775 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8776 4 : Lsn(0x20)..Lsn(0x48),
8777 4 : delta2,
8778 4 : ),
8779 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8780 4 : Lsn(0x48)..Lsn(0x50),
8781 4 : delta3,
8782 4 : ),
8783 4 : ], // delta layers
8784 4 : vec![], // image layers
8785 4 : Lsn(0x50),
8786 4 : )
8787 4 : .await?
8788 : } else {
8789 4 : tenant
8790 4 : .create_test_timeline_with_layers(
8791 4 : TIMELINE_ID,
8792 4 : Lsn(0x10),
8793 4 : DEFAULT_PG_VERSION,
8794 4 : &ctx,
8795 4 : vec![
8796 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8797 4 : Lsn(0x10)..Lsn(0x48),
8798 4 : delta1,
8799 4 : ),
8800 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8801 4 : Lsn(0x10)..Lsn(0x48),
8802 4 : delta2,
8803 4 : ),
8804 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8805 4 : Lsn(0x48)..Lsn(0x50),
8806 4 : delta3,
8807 4 : ),
8808 4 : ], // delta layers
8809 4 : vec![(Lsn(0x10), img_layer)], // image layers
8810 4 : Lsn(0x50),
8811 4 : )
8812 4 : .await?
8813 : };
8814 : {
8815 8 : tline
8816 8 : .latest_gc_cutoff_lsn
8817 8 : .lock_for_write()
8818 8 : .store_and_unlock(Lsn(0x30))
8819 8 : .wait()
8820 8 : .await;
8821 : // Update GC info
8822 8 : let mut guard = tline.gc_info.write().unwrap();
8823 8 : *guard = GcInfo {
8824 8 : retain_lsns: vec![],
8825 8 : cutoffs: GcCutoffs {
8826 8 : time: Lsn(0x30),
8827 8 : space: Lsn(0x30),
8828 8 : },
8829 8 : leases: Default::default(),
8830 8 : within_ancestor_pitr: false,
8831 8 : };
8832 8 : }
8833 8 :
8834 8 : let expected_result = [
8835 8 : Bytes::from_static(b"value 0@0x10"),
8836 8 : Bytes::from_static(b"value 1@0x10@0x20"),
8837 8 : Bytes::from_static(b"value 2@0x10@0x30"),
8838 8 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
8839 8 : Bytes::from_static(b"value 4@0x10"),
8840 8 : Bytes::from_static(b"value 5@0x10@0x20"),
8841 8 : Bytes::from_static(b"value 6@0x10@0x20"),
8842 8 : Bytes::from_static(b"value 7@0x10"),
8843 8 : Bytes::from_static(b"value 8@0x10@0x48"),
8844 8 : Bytes::from_static(b"value 9@0x10@0x48"),
8845 8 : ];
8846 8 :
8847 8 : let expected_result_at_gc_horizon = [
8848 8 : Bytes::from_static(b"value 0@0x10"),
8849 8 : Bytes::from_static(b"value 1@0x10@0x20"),
8850 8 : Bytes::from_static(b"value 2@0x10@0x30"),
8851 8 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
8852 8 : Bytes::from_static(b"value 4@0x10"),
8853 8 : Bytes::from_static(b"value 5@0x10@0x20"),
8854 8 : Bytes::from_static(b"value 6@0x10@0x20"),
8855 8 : Bytes::from_static(b"value 7@0x10"),
8856 8 : Bytes::from_static(b"value 8@0x10"),
8857 8 : Bytes::from_static(b"value 9@0x10"),
8858 8 : ];
8859 :
8860 88 : for idx in 0..10 {
8861 80 : assert_eq!(
8862 80 : tline
8863 80 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8864 80 : .await
8865 80 : .unwrap(),
8866 80 : &expected_result[idx]
8867 : );
8868 80 : assert_eq!(
8869 80 : tline
8870 80 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
8871 80 : .await
8872 80 : .unwrap(),
8873 80 : &expected_result_at_gc_horizon[idx]
8874 : );
8875 : }
8876 :
8877 8 : let cancel = CancellationToken::new();
8878 8 : tline
8879 8 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8880 8 : .await
8881 8 : .unwrap();
8882 :
8883 88 : for idx in 0..10 {
8884 80 : assert_eq!(
8885 80 : tline
8886 80 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8887 80 : .await
8888 80 : .unwrap(),
8889 80 : &expected_result[idx]
8890 : );
8891 80 : assert_eq!(
8892 80 : tline
8893 80 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
8894 80 : .await
8895 80 : .unwrap(),
8896 80 : &expected_result_at_gc_horizon[idx]
8897 : );
8898 : }
8899 :
8900 : // increase GC horizon and compact again
8901 : {
8902 8 : tline
8903 8 : .latest_gc_cutoff_lsn
8904 8 : .lock_for_write()
8905 8 : .store_and_unlock(Lsn(0x40))
8906 8 : .wait()
8907 8 : .await;
8908 : // Update GC info
8909 8 : let mut guard = tline.gc_info.write().unwrap();
8910 8 : guard.cutoffs.time = Lsn(0x40);
8911 8 : guard.cutoffs.space = Lsn(0x40);
8912 8 : }
8913 8 : tline
8914 8 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8915 8 : .await
8916 8 : .unwrap();
8917 8 :
8918 8 : Ok(())
8919 8 : }
8920 :
8921 : #[cfg(feature = "testing")]
8922 : #[tokio::test]
8923 4 : async fn test_generate_key_retention() -> anyhow::Result<()> {
8924 4 : let harness = TenantHarness::create("test_generate_key_retention").await?;
8925 4 : let (tenant, ctx) = harness.load().await;
8926 4 : let tline = tenant
8927 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
8928 4 : .await?;
8929 4 : tline.force_advance_lsn(Lsn(0x70));
8930 4 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
8931 4 : let history = vec![
8932 4 : (
8933 4 : key,
8934 4 : Lsn(0x10),
8935 4 : Value::WalRecord(NeonWalRecord::wal_init("0x10")),
8936 4 : ),
8937 4 : (
8938 4 : key,
8939 4 : Lsn(0x20),
8940 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8941 4 : ),
8942 4 : (
8943 4 : key,
8944 4 : Lsn(0x30),
8945 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8946 4 : ),
8947 4 : (
8948 4 : key,
8949 4 : Lsn(0x40),
8950 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8951 4 : ),
8952 4 : (
8953 4 : key,
8954 4 : Lsn(0x50),
8955 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
8956 4 : ),
8957 4 : (
8958 4 : key,
8959 4 : Lsn(0x60),
8960 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8961 4 : ),
8962 4 : (
8963 4 : key,
8964 4 : Lsn(0x70),
8965 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8966 4 : ),
8967 4 : (
8968 4 : key,
8969 4 : Lsn(0x80),
8970 4 : Value::Image(Bytes::copy_from_slice(
8971 4 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8972 4 : )),
8973 4 : ),
8974 4 : (
8975 4 : key,
8976 4 : Lsn(0x90),
8977 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8978 4 : ),
8979 4 : ];
8980 4 : let res = tline
8981 4 : .generate_key_retention(
8982 4 : key,
8983 4 : &history,
8984 4 : Lsn(0x60),
8985 4 : &[Lsn(0x20), Lsn(0x40), Lsn(0x50)],
8986 4 : 3,
8987 4 : None,
8988 4 : )
8989 4 : .await
8990 4 : .unwrap();
8991 4 : let expected_res = KeyHistoryRetention {
8992 4 : below_horizon: vec![
8993 4 : (
8994 4 : Lsn(0x20),
8995 4 : KeyLogAtLsn(vec![(
8996 4 : Lsn(0x20),
8997 4 : Value::Image(Bytes::from_static(b"0x10;0x20")),
8998 4 : )]),
8999 4 : ),
9000 4 : (
9001 4 : Lsn(0x40),
9002 4 : KeyLogAtLsn(vec![
9003 4 : (
9004 4 : Lsn(0x30),
9005 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9006 4 : ),
9007 4 : (
9008 4 : Lsn(0x40),
9009 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9010 4 : ),
9011 4 : ]),
9012 4 : ),
9013 4 : (
9014 4 : Lsn(0x50),
9015 4 : KeyLogAtLsn(vec![(
9016 4 : Lsn(0x50),
9017 4 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40;0x50")),
9018 4 : )]),
9019 4 : ),
9020 4 : (
9021 4 : Lsn(0x60),
9022 4 : KeyLogAtLsn(vec![(
9023 4 : Lsn(0x60),
9024 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9025 4 : )]),
9026 4 : ),
9027 4 : ],
9028 4 : above_horizon: KeyLogAtLsn(vec![
9029 4 : (
9030 4 : Lsn(0x70),
9031 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9032 4 : ),
9033 4 : (
9034 4 : Lsn(0x80),
9035 4 : Value::Image(Bytes::copy_from_slice(
9036 4 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9037 4 : )),
9038 4 : ),
9039 4 : (
9040 4 : Lsn(0x90),
9041 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9042 4 : ),
9043 4 : ]),
9044 4 : };
9045 4 : assert_eq!(res, expected_res);
9046 4 :
9047 4 : // We expect GC-compaction to run with the original GC. This would create a situation that
9048 4 : // the original GC algorithm removes some delta layers b/c there are full image coverage,
9049 4 : // therefore causing some keys to have an incomplete history below the lowest retain LSN.
9050 4 : // For example, we have
9051 4 : // ```plain
9052 4 : // init delta @ 0x10, image @ 0x20, delta @ 0x30 (gc_horizon), image @ 0x40.
9053 4 : // ```
9054 4 : // Now the GC horizon moves up, and we have
9055 4 : // ```plain
9056 4 : // init delta @ 0x10, image @ 0x20, delta @ 0x30, image @ 0x40 (gc_horizon)
9057 4 : // ```
9058 4 : // The original GC algorithm kicks in, and removes delta @ 0x10, image @ 0x20.
9059 4 : // We will end up with
9060 4 : // ```plain
9061 4 : // delta @ 0x30, image @ 0x40 (gc_horizon)
9062 4 : // ```
9063 4 : // Now we run the GC-compaction, and this key does not have a full history.
9064 4 : // We should be able to handle this partial history and drop everything before the
9065 4 : // gc_horizon image.
9066 4 :
9067 4 : let history = vec![
9068 4 : (
9069 4 : key,
9070 4 : Lsn(0x20),
9071 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9072 4 : ),
9073 4 : (
9074 4 : key,
9075 4 : Lsn(0x30),
9076 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9077 4 : ),
9078 4 : (
9079 4 : key,
9080 4 : Lsn(0x40),
9081 4 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
9082 4 : ),
9083 4 : (
9084 4 : key,
9085 4 : Lsn(0x50),
9086 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
9087 4 : ),
9088 4 : (
9089 4 : key,
9090 4 : Lsn(0x60),
9091 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9092 4 : ),
9093 4 : (
9094 4 : key,
9095 4 : Lsn(0x70),
9096 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9097 4 : ),
9098 4 : (
9099 4 : key,
9100 4 : Lsn(0x80),
9101 4 : Value::Image(Bytes::copy_from_slice(
9102 4 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9103 4 : )),
9104 4 : ),
9105 4 : (
9106 4 : key,
9107 4 : Lsn(0x90),
9108 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9109 4 : ),
9110 4 : ];
9111 4 : let res = tline
9112 4 : .generate_key_retention(key, &history, Lsn(0x60), &[Lsn(0x40), Lsn(0x50)], 3, None)
9113 4 : .await
9114 4 : .unwrap();
9115 4 : let expected_res = KeyHistoryRetention {
9116 4 : below_horizon: vec![
9117 4 : (
9118 4 : Lsn(0x40),
9119 4 : KeyLogAtLsn(vec![(
9120 4 : Lsn(0x40),
9121 4 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
9122 4 : )]),
9123 4 : ),
9124 4 : (
9125 4 : Lsn(0x50),
9126 4 : KeyLogAtLsn(vec![(
9127 4 : Lsn(0x50),
9128 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
9129 4 : )]),
9130 4 : ),
9131 4 : (
9132 4 : Lsn(0x60),
9133 4 : KeyLogAtLsn(vec![(
9134 4 : Lsn(0x60),
9135 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9136 4 : )]),
9137 4 : ),
9138 4 : ],
9139 4 : above_horizon: KeyLogAtLsn(vec![
9140 4 : (
9141 4 : Lsn(0x70),
9142 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9143 4 : ),
9144 4 : (
9145 4 : Lsn(0x80),
9146 4 : Value::Image(Bytes::copy_from_slice(
9147 4 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9148 4 : )),
9149 4 : ),
9150 4 : (
9151 4 : Lsn(0x90),
9152 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9153 4 : ),
9154 4 : ]),
9155 4 : };
9156 4 : assert_eq!(res, expected_res);
9157 4 :
9158 4 : // In case of branch compaction, the branch itself does not have the full history, and we need to provide
9159 4 : // the ancestor image in the test case.
9160 4 :
9161 4 : let history = vec![
9162 4 : (
9163 4 : key,
9164 4 : Lsn(0x20),
9165 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9166 4 : ),
9167 4 : (
9168 4 : key,
9169 4 : Lsn(0x30),
9170 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9171 4 : ),
9172 4 : (
9173 4 : key,
9174 4 : Lsn(0x40),
9175 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9176 4 : ),
9177 4 : (
9178 4 : key,
9179 4 : Lsn(0x70),
9180 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9181 4 : ),
9182 4 : ];
9183 4 : let res = tline
9184 4 : .generate_key_retention(
9185 4 : key,
9186 4 : &history,
9187 4 : Lsn(0x60),
9188 4 : &[],
9189 4 : 3,
9190 4 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
9191 4 : )
9192 4 : .await
9193 4 : .unwrap();
9194 4 : let expected_res = KeyHistoryRetention {
9195 4 : below_horizon: vec![(
9196 4 : Lsn(0x60),
9197 4 : KeyLogAtLsn(vec![(
9198 4 : Lsn(0x60),
9199 4 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")), // use the ancestor image to reconstruct the page
9200 4 : )]),
9201 4 : )],
9202 4 : above_horizon: KeyLogAtLsn(vec![(
9203 4 : Lsn(0x70),
9204 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9205 4 : )]),
9206 4 : };
9207 4 : assert_eq!(res, expected_res);
9208 4 :
9209 4 : let history = vec![
9210 4 : (
9211 4 : key,
9212 4 : Lsn(0x20),
9213 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9214 4 : ),
9215 4 : (
9216 4 : key,
9217 4 : Lsn(0x40),
9218 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9219 4 : ),
9220 4 : (
9221 4 : key,
9222 4 : Lsn(0x60),
9223 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9224 4 : ),
9225 4 : (
9226 4 : key,
9227 4 : Lsn(0x70),
9228 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9229 4 : ),
9230 4 : ];
9231 4 : let res = tline
9232 4 : .generate_key_retention(
9233 4 : key,
9234 4 : &history,
9235 4 : Lsn(0x60),
9236 4 : &[Lsn(0x30)],
9237 4 : 3,
9238 4 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
9239 4 : )
9240 4 : .await
9241 4 : .unwrap();
9242 4 : let expected_res = KeyHistoryRetention {
9243 4 : below_horizon: vec![
9244 4 : (
9245 4 : Lsn(0x30),
9246 4 : KeyLogAtLsn(vec![(
9247 4 : Lsn(0x20),
9248 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9249 4 : )]),
9250 4 : ),
9251 4 : (
9252 4 : Lsn(0x60),
9253 4 : KeyLogAtLsn(vec![(
9254 4 : Lsn(0x60),
9255 4 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x40;0x60")),
9256 4 : )]),
9257 4 : ),
9258 4 : ],
9259 4 : above_horizon: KeyLogAtLsn(vec![(
9260 4 : Lsn(0x70),
9261 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9262 4 : )]),
9263 4 : };
9264 4 : assert_eq!(res, expected_res);
9265 4 :
9266 4 : Ok(())
9267 4 : }
9268 :
9269 : #[cfg(feature = "testing")]
9270 : #[tokio::test]
9271 4 : async fn test_simple_bottom_most_compaction_with_retain_lsns() -> anyhow::Result<()> {
9272 4 : let harness =
9273 4 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns").await?;
9274 4 : let (tenant, ctx) = harness.load().await;
9275 4 :
9276 1036 : fn get_key(id: u32) -> Key {
9277 1036 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9278 1036 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9279 1036 : key.field6 = id;
9280 1036 : key
9281 1036 : }
9282 4 :
9283 4 : let img_layer = (0..10)
9284 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9285 4 : .collect_vec();
9286 4 :
9287 4 : let delta1 = vec![
9288 4 : (
9289 4 : get_key(1),
9290 4 : Lsn(0x20),
9291 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9292 4 : ),
9293 4 : (
9294 4 : get_key(2),
9295 4 : Lsn(0x30),
9296 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9297 4 : ),
9298 4 : (
9299 4 : get_key(3),
9300 4 : Lsn(0x28),
9301 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9302 4 : ),
9303 4 : (
9304 4 : get_key(3),
9305 4 : Lsn(0x30),
9306 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9307 4 : ),
9308 4 : (
9309 4 : get_key(3),
9310 4 : Lsn(0x40),
9311 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9312 4 : ),
9313 4 : ];
9314 4 : let delta2 = vec![
9315 4 : (
9316 4 : get_key(5),
9317 4 : Lsn(0x20),
9318 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9319 4 : ),
9320 4 : (
9321 4 : get_key(6),
9322 4 : Lsn(0x20),
9323 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9324 4 : ),
9325 4 : ];
9326 4 : let delta3 = vec![
9327 4 : (
9328 4 : get_key(8),
9329 4 : Lsn(0x48),
9330 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9331 4 : ),
9332 4 : (
9333 4 : get_key(9),
9334 4 : Lsn(0x48),
9335 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9336 4 : ),
9337 4 : ];
9338 4 :
9339 4 : let tline = tenant
9340 4 : .create_test_timeline_with_layers(
9341 4 : TIMELINE_ID,
9342 4 : Lsn(0x10),
9343 4 : DEFAULT_PG_VERSION,
9344 4 : &ctx,
9345 4 : vec![
9346 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
9347 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
9348 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
9349 4 : ], // delta layers
9350 4 : vec![(Lsn(0x10), img_layer)], // image layers
9351 4 : Lsn(0x50),
9352 4 : )
9353 4 : .await?;
9354 4 : {
9355 4 : tline
9356 4 : .latest_gc_cutoff_lsn
9357 4 : .lock_for_write()
9358 4 : .store_and_unlock(Lsn(0x30))
9359 4 : .wait()
9360 4 : .await;
9361 4 : // Update GC info
9362 4 : let mut guard = tline.gc_info.write().unwrap();
9363 4 : *guard = GcInfo {
9364 4 : retain_lsns: vec![
9365 4 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
9366 4 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
9367 4 : ],
9368 4 : cutoffs: GcCutoffs {
9369 4 : time: Lsn(0x30),
9370 4 : space: Lsn(0x30),
9371 4 : },
9372 4 : leases: Default::default(),
9373 4 : within_ancestor_pitr: false,
9374 4 : };
9375 4 : }
9376 4 :
9377 4 : let expected_result = [
9378 4 : Bytes::from_static(b"value 0@0x10"),
9379 4 : Bytes::from_static(b"value 1@0x10@0x20"),
9380 4 : Bytes::from_static(b"value 2@0x10@0x30"),
9381 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9382 4 : Bytes::from_static(b"value 4@0x10"),
9383 4 : Bytes::from_static(b"value 5@0x10@0x20"),
9384 4 : Bytes::from_static(b"value 6@0x10@0x20"),
9385 4 : Bytes::from_static(b"value 7@0x10"),
9386 4 : Bytes::from_static(b"value 8@0x10@0x48"),
9387 4 : Bytes::from_static(b"value 9@0x10@0x48"),
9388 4 : ];
9389 4 :
9390 4 : let expected_result_at_gc_horizon = [
9391 4 : Bytes::from_static(b"value 0@0x10"),
9392 4 : Bytes::from_static(b"value 1@0x10@0x20"),
9393 4 : Bytes::from_static(b"value 2@0x10@0x30"),
9394 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
9395 4 : Bytes::from_static(b"value 4@0x10"),
9396 4 : Bytes::from_static(b"value 5@0x10@0x20"),
9397 4 : Bytes::from_static(b"value 6@0x10@0x20"),
9398 4 : Bytes::from_static(b"value 7@0x10"),
9399 4 : Bytes::from_static(b"value 8@0x10"),
9400 4 : Bytes::from_static(b"value 9@0x10"),
9401 4 : ];
9402 4 :
9403 4 : let expected_result_at_lsn_20 = [
9404 4 : Bytes::from_static(b"value 0@0x10"),
9405 4 : Bytes::from_static(b"value 1@0x10@0x20"),
9406 4 : Bytes::from_static(b"value 2@0x10"),
9407 4 : Bytes::from_static(b"value 3@0x10"),
9408 4 : Bytes::from_static(b"value 4@0x10"),
9409 4 : Bytes::from_static(b"value 5@0x10@0x20"),
9410 4 : Bytes::from_static(b"value 6@0x10@0x20"),
9411 4 : Bytes::from_static(b"value 7@0x10"),
9412 4 : Bytes::from_static(b"value 8@0x10"),
9413 4 : Bytes::from_static(b"value 9@0x10"),
9414 4 : ];
9415 4 :
9416 4 : let expected_result_at_lsn_10 = [
9417 4 : Bytes::from_static(b"value 0@0x10"),
9418 4 : Bytes::from_static(b"value 1@0x10"),
9419 4 : Bytes::from_static(b"value 2@0x10"),
9420 4 : Bytes::from_static(b"value 3@0x10"),
9421 4 : Bytes::from_static(b"value 4@0x10"),
9422 4 : Bytes::from_static(b"value 5@0x10"),
9423 4 : Bytes::from_static(b"value 6@0x10"),
9424 4 : Bytes::from_static(b"value 7@0x10"),
9425 4 : Bytes::from_static(b"value 8@0x10"),
9426 4 : Bytes::from_static(b"value 9@0x10"),
9427 4 : ];
9428 4 :
9429 24 : let verify_result = || async {
9430 24 : let gc_horizon = {
9431 24 : let gc_info = tline.gc_info.read().unwrap();
9432 24 : gc_info.cutoffs.time
9433 4 : };
9434 264 : for idx in 0..10 {
9435 240 : assert_eq!(
9436 240 : tline
9437 240 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9438 240 : .await
9439 240 : .unwrap(),
9440 240 : &expected_result[idx]
9441 4 : );
9442 240 : assert_eq!(
9443 240 : tline
9444 240 : .get(get_key(idx as u32), gc_horizon, &ctx)
9445 240 : .await
9446 240 : .unwrap(),
9447 240 : &expected_result_at_gc_horizon[idx]
9448 4 : );
9449 240 : assert_eq!(
9450 240 : tline
9451 240 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
9452 240 : .await
9453 240 : .unwrap(),
9454 240 : &expected_result_at_lsn_20[idx]
9455 4 : );
9456 240 : assert_eq!(
9457 240 : tline
9458 240 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
9459 240 : .await
9460 240 : .unwrap(),
9461 240 : &expected_result_at_lsn_10[idx]
9462 4 : );
9463 4 : }
9464 48 : };
9465 4 :
9466 4 : verify_result().await;
9467 4 :
9468 4 : let cancel = CancellationToken::new();
9469 4 : let mut dryrun_flags = EnumSet::new();
9470 4 : dryrun_flags.insert(CompactFlags::DryRun);
9471 4 :
9472 4 : tline
9473 4 : .compact_with_gc(
9474 4 : &cancel,
9475 4 : CompactOptions {
9476 4 : flags: dryrun_flags,
9477 4 : ..Default::default()
9478 4 : },
9479 4 : &ctx,
9480 4 : )
9481 4 : .await
9482 4 : .unwrap();
9483 4 : // 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
9484 4 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
9485 4 : verify_result().await;
9486 4 :
9487 4 : tline
9488 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9489 4 : .await
9490 4 : .unwrap();
9491 4 : verify_result().await;
9492 4 :
9493 4 : // compact again
9494 4 : tline
9495 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9496 4 : .await
9497 4 : .unwrap();
9498 4 : verify_result().await;
9499 4 :
9500 4 : // increase GC horizon and compact again
9501 4 : {
9502 4 : tline
9503 4 : .latest_gc_cutoff_lsn
9504 4 : .lock_for_write()
9505 4 : .store_and_unlock(Lsn(0x38))
9506 4 : .wait()
9507 4 : .await;
9508 4 : // Update GC info
9509 4 : let mut guard = tline.gc_info.write().unwrap();
9510 4 : guard.cutoffs.time = Lsn(0x38);
9511 4 : guard.cutoffs.space = Lsn(0x38);
9512 4 : }
9513 4 : tline
9514 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9515 4 : .await
9516 4 : .unwrap();
9517 4 : verify_result().await; // no wals between 0x30 and 0x38, so we should obtain the same result
9518 4 :
9519 4 : // not increasing the GC horizon and compact again
9520 4 : tline
9521 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9522 4 : .await
9523 4 : .unwrap();
9524 4 : verify_result().await;
9525 4 :
9526 4 : Ok(())
9527 4 : }
9528 :
9529 : #[cfg(feature = "testing")]
9530 : #[tokio::test]
9531 4 : async fn test_simple_bottom_most_compaction_with_retain_lsns_single_key() -> anyhow::Result<()>
9532 4 : {
9533 4 : let harness =
9534 4 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns_single_key")
9535 4 : .await?;
9536 4 : let (tenant, ctx) = harness.load().await;
9537 4 :
9538 704 : fn get_key(id: u32) -> Key {
9539 704 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9540 704 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9541 704 : key.field6 = id;
9542 704 : key
9543 704 : }
9544 4 :
9545 4 : let img_layer = (0..10)
9546 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9547 4 : .collect_vec();
9548 4 :
9549 4 : let delta1 = vec![
9550 4 : (
9551 4 : get_key(1),
9552 4 : Lsn(0x20),
9553 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9554 4 : ),
9555 4 : (
9556 4 : get_key(1),
9557 4 : Lsn(0x28),
9558 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9559 4 : ),
9560 4 : ];
9561 4 : let delta2 = vec![
9562 4 : (
9563 4 : get_key(1),
9564 4 : Lsn(0x30),
9565 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9566 4 : ),
9567 4 : (
9568 4 : get_key(1),
9569 4 : Lsn(0x38),
9570 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
9571 4 : ),
9572 4 : ];
9573 4 : let delta3 = vec![
9574 4 : (
9575 4 : get_key(8),
9576 4 : Lsn(0x48),
9577 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9578 4 : ),
9579 4 : (
9580 4 : get_key(9),
9581 4 : Lsn(0x48),
9582 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9583 4 : ),
9584 4 : ];
9585 4 :
9586 4 : let tline = tenant
9587 4 : .create_test_timeline_with_layers(
9588 4 : TIMELINE_ID,
9589 4 : Lsn(0x10),
9590 4 : DEFAULT_PG_VERSION,
9591 4 : &ctx,
9592 4 : vec![
9593 4 : // delta1 and delta 2 only contain a single key but multiple updates
9594 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x30), delta1),
9595 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
9596 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x50), delta3),
9597 4 : ], // delta layers
9598 4 : vec![(Lsn(0x10), img_layer)], // image layers
9599 4 : Lsn(0x50),
9600 4 : )
9601 4 : .await?;
9602 4 : {
9603 4 : tline
9604 4 : .latest_gc_cutoff_lsn
9605 4 : .lock_for_write()
9606 4 : .store_and_unlock(Lsn(0x30))
9607 4 : .wait()
9608 4 : .await;
9609 4 : // Update GC info
9610 4 : let mut guard = tline.gc_info.write().unwrap();
9611 4 : *guard = GcInfo {
9612 4 : retain_lsns: vec![
9613 4 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
9614 4 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
9615 4 : ],
9616 4 : cutoffs: GcCutoffs {
9617 4 : time: Lsn(0x30),
9618 4 : space: Lsn(0x30),
9619 4 : },
9620 4 : leases: Default::default(),
9621 4 : within_ancestor_pitr: false,
9622 4 : };
9623 4 : }
9624 4 :
9625 4 : let expected_result = [
9626 4 : Bytes::from_static(b"value 0@0x10"),
9627 4 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
9628 4 : Bytes::from_static(b"value 2@0x10"),
9629 4 : Bytes::from_static(b"value 3@0x10"),
9630 4 : Bytes::from_static(b"value 4@0x10"),
9631 4 : Bytes::from_static(b"value 5@0x10"),
9632 4 : Bytes::from_static(b"value 6@0x10"),
9633 4 : Bytes::from_static(b"value 7@0x10"),
9634 4 : Bytes::from_static(b"value 8@0x10@0x48"),
9635 4 : Bytes::from_static(b"value 9@0x10@0x48"),
9636 4 : ];
9637 4 :
9638 4 : let expected_result_at_gc_horizon = [
9639 4 : Bytes::from_static(b"value 0@0x10"),
9640 4 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
9641 4 : Bytes::from_static(b"value 2@0x10"),
9642 4 : Bytes::from_static(b"value 3@0x10"),
9643 4 : Bytes::from_static(b"value 4@0x10"),
9644 4 : Bytes::from_static(b"value 5@0x10"),
9645 4 : Bytes::from_static(b"value 6@0x10"),
9646 4 : Bytes::from_static(b"value 7@0x10"),
9647 4 : Bytes::from_static(b"value 8@0x10"),
9648 4 : Bytes::from_static(b"value 9@0x10"),
9649 4 : ];
9650 4 :
9651 4 : let expected_result_at_lsn_20 = [
9652 4 : Bytes::from_static(b"value 0@0x10"),
9653 4 : Bytes::from_static(b"value 1@0x10@0x20"),
9654 4 : Bytes::from_static(b"value 2@0x10"),
9655 4 : Bytes::from_static(b"value 3@0x10"),
9656 4 : Bytes::from_static(b"value 4@0x10"),
9657 4 : Bytes::from_static(b"value 5@0x10"),
9658 4 : Bytes::from_static(b"value 6@0x10"),
9659 4 : Bytes::from_static(b"value 7@0x10"),
9660 4 : Bytes::from_static(b"value 8@0x10"),
9661 4 : Bytes::from_static(b"value 9@0x10"),
9662 4 : ];
9663 4 :
9664 4 : let expected_result_at_lsn_10 = [
9665 4 : Bytes::from_static(b"value 0@0x10"),
9666 4 : Bytes::from_static(b"value 1@0x10"),
9667 4 : Bytes::from_static(b"value 2@0x10"),
9668 4 : Bytes::from_static(b"value 3@0x10"),
9669 4 : Bytes::from_static(b"value 4@0x10"),
9670 4 : Bytes::from_static(b"value 5@0x10"),
9671 4 : Bytes::from_static(b"value 6@0x10"),
9672 4 : Bytes::from_static(b"value 7@0x10"),
9673 4 : Bytes::from_static(b"value 8@0x10"),
9674 4 : Bytes::from_static(b"value 9@0x10"),
9675 4 : ];
9676 4 :
9677 16 : let verify_result = || async {
9678 16 : let gc_horizon = {
9679 16 : let gc_info = tline.gc_info.read().unwrap();
9680 16 : gc_info.cutoffs.time
9681 4 : };
9682 176 : for idx in 0..10 {
9683 160 : assert_eq!(
9684 160 : tline
9685 160 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9686 160 : .await
9687 160 : .unwrap(),
9688 160 : &expected_result[idx]
9689 4 : );
9690 160 : assert_eq!(
9691 160 : tline
9692 160 : .get(get_key(idx as u32), gc_horizon, &ctx)
9693 160 : .await
9694 160 : .unwrap(),
9695 160 : &expected_result_at_gc_horizon[idx]
9696 4 : );
9697 160 : assert_eq!(
9698 160 : tline
9699 160 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
9700 160 : .await
9701 160 : .unwrap(),
9702 160 : &expected_result_at_lsn_20[idx]
9703 4 : );
9704 160 : assert_eq!(
9705 160 : tline
9706 160 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
9707 160 : .await
9708 160 : .unwrap(),
9709 160 : &expected_result_at_lsn_10[idx]
9710 4 : );
9711 4 : }
9712 32 : };
9713 4 :
9714 4 : verify_result().await;
9715 4 :
9716 4 : let cancel = CancellationToken::new();
9717 4 : let mut dryrun_flags = EnumSet::new();
9718 4 : dryrun_flags.insert(CompactFlags::DryRun);
9719 4 :
9720 4 : tline
9721 4 : .compact_with_gc(
9722 4 : &cancel,
9723 4 : CompactOptions {
9724 4 : flags: dryrun_flags,
9725 4 : ..Default::default()
9726 4 : },
9727 4 : &ctx,
9728 4 : )
9729 4 : .await
9730 4 : .unwrap();
9731 4 : // 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
9732 4 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
9733 4 : verify_result().await;
9734 4 :
9735 4 : tline
9736 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9737 4 : .await
9738 4 : .unwrap();
9739 4 : verify_result().await;
9740 4 :
9741 4 : // compact again
9742 4 : tline
9743 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9744 4 : .await
9745 4 : .unwrap();
9746 4 : verify_result().await;
9747 4 :
9748 4 : Ok(())
9749 4 : }
9750 :
9751 : #[cfg(feature = "testing")]
9752 : #[tokio::test]
9753 4 : async fn test_simple_bottom_most_compaction_on_branch() -> anyhow::Result<()> {
9754 4 : use models::CompactLsnRange;
9755 4 :
9756 4 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_on_branch").await?;
9757 4 : let (tenant, ctx) = harness.load().await;
9758 4 :
9759 332 : fn get_key(id: u32) -> Key {
9760 332 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
9761 332 : key.field6 = id;
9762 332 : key
9763 332 : }
9764 4 :
9765 4 : let img_layer = (0..10)
9766 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9767 4 : .collect_vec();
9768 4 :
9769 4 : let delta1 = vec![
9770 4 : (
9771 4 : get_key(1),
9772 4 : Lsn(0x20),
9773 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9774 4 : ),
9775 4 : (
9776 4 : get_key(2),
9777 4 : Lsn(0x30),
9778 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9779 4 : ),
9780 4 : (
9781 4 : get_key(3),
9782 4 : Lsn(0x28),
9783 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9784 4 : ),
9785 4 : (
9786 4 : get_key(3),
9787 4 : Lsn(0x30),
9788 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9789 4 : ),
9790 4 : (
9791 4 : get_key(3),
9792 4 : Lsn(0x40),
9793 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9794 4 : ),
9795 4 : ];
9796 4 : let delta2 = vec![
9797 4 : (
9798 4 : get_key(5),
9799 4 : Lsn(0x20),
9800 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9801 4 : ),
9802 4 : (
9803 4 : get_key(6),
9804 4 : Lsn(0x20),
9805 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9806 4 : ),
9807 4 : ];
9808 4 : let delta3 = vec![
9809 4 : (
9810 4 : get_key(8),
9811 4 : Lsn(0x48),
9812 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9813 4 : ),
9814 4 : (
9815 4 : get_key(9),
9816 4 : Lsn(0x48),
9817 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9818 4 : ),
9819 4 : ];
9820 4 :
9821 4 : let parent_tline = tenant
9822 4 : .create_test_timeline_with_layers(
9823 4 : TIMELINE_ID,
9824 4 : Lsn(0x10),
9825 4 : DEFAULT_PG_VERSION,
9826 4 : &ctx,
9827 4 : vec![], // delta layers
9828 4 : vec![(Lsn(0x18), img_layer)], // image layers
9829 4 : Lsn(0x18),
9830 4 : )
9831 4 : .await?;
9832 4 :
9833 4 : parent_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
9834 4 :
9835 4 : let branch_tline = tenant
9836 4 : .branch_timeline_test_with_layers(
9837 4 : &parent_tline,
9838 4 : NEW_TIMELINE_ID,
9839 4 : Some(Lsn(0x18)),
9840 4 : &ctx,
9841 4 : vec![
9842 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
9843 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
9844 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
9845 4 : ], // delta layers
9846 4 : vec![], // image layers
9847 4 : Lsn(0x50),
9848 4 : )
9849 4 : .await?;
9850 4 :
9851 4 : branch_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
9852 4 :
9853 4 : {
9854 4 : parent_tline
9855 4 : .latest_gc_cutoff_lsn
9856 4 : .lock_for_write()
9857 4 : .store_and_unlock(Lsn(0x10))
9858 4 : .wait()
9859 4 : .await;
9860 4 : // Update GC info
9861 4 : let mut guard = parent_tline.gc_info.write().unwrap();
9862 4 : *guard = GcInfo {
9863 4 : retain_lsns: vec![(Lsn(0x18), branch_tline.timeline_id, MaybeOffloaded::No)],
9864 4 : cutoffs: GcCutoffs {
9865 4 : time: Lsn(0x10),
9866 4 : space: Lsn(0x10),
9867 4 : },
9868 4 : leases: Default::default(),
9869 4 : within_ancestor_pitr: false,
9870 4 : };
9871 4 : }
9872 4 :
9873 4 : {
9874 4 : branch_tline
9875 4 : .latest_gc_cutoff_lsn
9876 4 : .lock_for_write()
9877 4 : .store_and_unlock(Lsn(0x50))
9878 4 : .wait()
9879 4 : .await;
9880 4 : // Update GC info
9881 4 : let mut guard = branch_tline.gc_info.write().unwrap();
9882 4 : *guard = GcInfo {
9883 4 : retain_lsns: vec![(Lsn(0x40), branch_tline.timeline_id, MaybeOffloaded::No)],
9884 4 : cutoffs: GcCutoffs {
9885 4 : time: Lsn(0x50),
9886 4 : space: Lsn(0x50),
9887 4 : },
9888 4 : leases: Default::default(),
9889 4 : within_ancestor_pitr: false,
9890 4 : };
9891 4 : }
9892 4 :
9893 4 : let expected_result_at_gc_horizon = [
9894 4 : Bytes::from_static(b"value 0@0x10"),
9895 4 : Bytes::from_static(b"value 1@0x10@0x20"),
9896 4 : Bytes::from_static(b"value 2@0x10@0x30"),
9897 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9898 4 : Bytes::from_static(b"value 4@0x10"),
9899 4 : Bytes::from_static(b"value 5@0x10@0x20"),
9900 4 : Bytes::from_static(b"value 6@0x10@0x20"),
9901 4 : Bytes::from_static(b"value 7@0x10"),
9902 4 : Bytes::from_static(b"value 8@0x10@0x48"),
9903 4 : Bytes::from_static(b"value 9@0x10@0x48"),
9904 4 : ];
9905 4 :
9906 4 : let expected_result_at_lsn_40 = [
9907 4 : Bytes::from_static(b"value 0@0x10"),
9908 4 : Bytes::from_static(b"value 1@0x10@0x20"),
9909 4 : Bytes::from_static(b"value 2@0x10@0x30"),
9910 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9911 4 : Bytes::from_static(b"value 4@0x10"),
9912 4 : Bytes::from_static(b"value 5@0x10@0x20"),
9913 4 : Bytes::from_static(b"value 6@0x10@0x20"),
9914 4 : Bytes::from_static(b"value 7@0x10"),
9915 4 : Bytes::from_static(b"value 8@0x10"),
9916 4 : Bytes::from_static(b"value 9@0x10"),
9917 4 : ];
9918 4 :
9919 12 : let verify_result = || async {
9920 132 : for idx in 0..10 {
9921 120 : assert_eq!(
9922 120 : branch_tline
9923 120 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9924 120 : .await
9925 120 : .unwrap(),
9926 120 : &expected_result_at_gc_horizon[idx]
9927 4 : );
9928 120 : assert_eq!(
9929 120 : branch_tline
9930 120 : .get(get_key(idx as u32), Lsn(0x40), &ctx)
9931 120 : .await
9932 120 : .unwrap(),
9933 120 : &expected_result_at_lsn_40[idx]
9934 4 : );
9935 4 : }
9936 24 : };
9937 4 :
9938 4 : verify_result().await;
9939 4 :
9940 4 : let cancel = CancellationToken::new();
9941 4 : branch_tline
9942 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9943 4 : .await
9944 4 : .unwrap();
9945 4 :
9946 4 : verify_result().await;
9947 4 :
9948 4 : // Piggyback a compaction with above_lsn. Ensure it works correctly when the specified LSN intersects with the layer files.
9949 4 : // Now we already have a single large delta layer, so the compaction min_layer_lsn should be the same as ancestor LSN (0x18).
9950 4 : branch_tline
9951 4 : .compact_with_gc(
9952 4 : &cancel,
9953 4 : CompactOptions {
9954 4 : compact_lsn_range: Some(CompactLsnRange::above(Lsn(0x40))),
9955 4 : ..Default::default()
9956 4 : },
9957 4 : &ctx,
9958 4 : )
9959 4 : .await
9960 4 : .unwrap();
9961 4 :
9962 4 : verify_result().await;
9963 4 :
9964 4 : Ok(())
9965 4 : }
9966 :
9967 : // Regression test for https://github.com/neondatabase/neon/issues/9012
9968 : // Create an image arrangement where we have to read at different LSN ranges
9969 : // from a delta layer. This is achieved by overlapping an image layer on top of
9970 : // a delta layer. Like so:
9971 : //
9972 : // A B
9973 : // +----------------+ -> delta_layer
9974 : // | | ^ lsn
9975 : // | =========|-> nested_image_layer |
9976 : // | C | |
9977 : // +----------------+ |
9978 : // ======== -> baseline_image_layer +-------> key
9979 : //
9980 : //
9981 : // When querying the key range [A, B) we need to read at different LSN ranges
9982 : // for [A, C) and [C, B). This test checks that the described edge case is handled correctly.
9983 : #[cfg(feature = "testing")]
9984 : #[tokio::test]
9985 4 : async fn test_vectored_read_with_nested_image_layer() -> anyhow::Result<()> {
9986 4 : let harness = TenantHarness::create("test_vectored_read_with_nested_image_layer").await?;
9987 4 : let (tenant, ctx) = harness.load().await;
9988 4 :
9989 4 : let will_init_keys = [2, 6];
9990 88 : fn get_key(id: u32) -> Key {
9991 88 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
9992 88 : key.field6 = id;
9993 88 : key
9994 88 : }
9995 4 :
9996 4 : let mut expected_key_values = HashMap::new();
9997 4 :
9998 4 : let baseline_image_layer_lsn = Lsn(0x10);
9999 4 : let mut baseline_img_layer = Vec::new();
10000 24 : for i in 0..5 {
10001 20 : let key = get_key(i);
10002 20 : let value = format!("value {i}@{baseline_image_layer_lsn}");
10003 20 :
10004 20 : let removed = expected_key_values.insert(key, value.clone());
10005 20 : assert!(removed.is_none());
10006 4 :
10007 20 : baseline_img_layer.push((key, Bytes::from(value)));
10008 4 : }
10009 4 :
10010 4 : let nested_image_layer_lsn = Lsn(0x50);
10011 4 : let mut nested_img_layer = Vec::new();
10012 24 : for i in 5..10 {
10013 20 : let key = get_key(i);
10014 20 : let value = format!("value {i}@{nested_image_layer_lsn}");
10015 20 :
10016 20 : let removed = expected_key_values.insert(key, value.clone());
10017 20 : assert!(removed.is_none());
10018 4 :
10019 20 : nested_img_layer.push((key, Bytes::from(value)));
10020 4 : }
10021 4 :
10022 4 : let mut delta_layer_spec = Vec::default();
10023 4 : let delta_layer_start_lsn = Lsn(0x20);
10024 4 : let mut delta_layer_end_lsn = delta_layer_start_lsn;
10025 4 :
10026 44 : for i in 0..10 {
10027 40 : let key = get_key(i);
10028 40 : let key_in_nested = nested_img_layer
10029 40 : .iter()
10030 160 : .any(|(key_with_img, _)| *key_with_img == key);
10031 40 : let lsn = {
10032 40 : if key_in_nested {
10033 20 : Lsn(nested_image_layer_lsn.0 + 0x10)
10034 4 : } else {
10035 20 : delta_layer_start_lsn
10036 4 : }
10037 4 : };
10038 4 :
10039 40 : let will_init = will_init_keys.contains(&i);
10040 40 : if will_init {
10041 8 : delta_layer_spec.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init(""))));
10042 8 :
10043 8 : expected_key_values.insert(key, "".to_string());
10044 32 : } else {
10045 32 : let delta = format!("@{lsn}");
10046 32 : delta_layer_spec.push((
10047 32 : key,
10048 32 : lsn,
10049 32 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
10050 32 : ));
10051 32 :
10052 32 : expected_key_values
10053 32 : .get_mut(&key)
10054 32 : .expect("An image exists for each key")
10055 32 : .push_str(delta.as_str());
10056 32 : }
10057 40 : delta_layer_end_lsn = std::cmp::max(delta_layer_start_lsn, lsn);
10058 4 : }
10059 4 :
10060 4 : delta_layer_end_lsn = Lsn(delta_layer_end_lsn.0 + 1);
10061 4 :
10062 4 : assert!(
10063 4 : nested_image_layer_lsn > delta_layer_start_lsn
10064 4 : && nested_image_layer_lsn < delta_layer_end_lsn
10065 4 : );
10066 4 :
10067 4 : let tline = tenant
10068 4 : .create_test_timeline_with_layers(
10069 4 : TIMELINE_ID,
10070 4 : baseline_image_layer_lsn,
10071 4 : DEFAULT_PG_VERSION,
10072 4 : &ctx,
10073 4 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
10074 4 : delta_layer_start_lsn..delta_layer_end_lsn,
10075 4 : delta_layer_spec,
10076 4 : )], // delta layers
10077 4 : vec![
10078 4 : (baseline_image_layer_lsn, baseline_img_layer),
10079 4 : (nested_image_layer_lsn, nested_img_layer),
10080 4 : ], // image layers
10081 4 : delta_layer_end_lsn,
10082 4 : )
10083 4 : .await?;
10084 4 :
10085 4 : let keyspace = KeySpace::single(get_key(0)..get_key(10));
10086 4 : let results = tline
10087 4 : .get_vectored(
10088 4 : keyspace,
10089 4 : delta_layer_end_lsn,
10090 4 : IoConcurrency::sequential(),
10091 4 : &ctx,
10092 4 : )
10093 4 : .await
10094 4 : .expect("No vectored errors");
10095 44 : for (key, res) in results {
10096 40 : let value = res.expect("No key errors");
10097 40 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
10098 40 : assert_eq!(value, Bytes::from(expected_value));
10099 4 : }
10100 4 :
10101 4 : Ok(())
10102 4 : }
10103 :
10104 428 : fn sort_layer_key(k1: &PersistentLayerKey, k2: &PersistentLayerKey) -> std::cmp::Ordering {
10105 428 : (
10106 428 : k1.is_delta,
10107 428 : k1.key_range.start,
10108 428 : k1.key_range.end,
10109 428 : k1.lsn_range.start,
10110 428 : k1.lsn_range.end,
10111 428 : )
10112 428 : .cmp(&(
10113 428 : k2.is_delta,
10114 428 : k2.key_range.start,
10115 428 : k2.key_range.end,
10116 428 : k2.lsn_range.start,
10117 428 : k2.lsn_range.end,
10118 428 : ))
10119 428 : }
10120 :
10121 48 : async fn inspect_and_sort(
10122 48 : tline: &Arc<Timeline>,
10123 48 : filter: Option<std::ops::Range<Key>>,
10124 48 : ) -> Vec<PersistentLayerKey> {
10125 48 : let mut all_layers = tline.inspect_historic_layers().await.unwrap();
10126 48 : if let Some(filter) = filter {
10127 216 : all_layers.retain(|layer| overlaps_with(&layer.key_range, &filter));
10128 44 : }
10129 48 : all_layers.sort_by(sort_layer_key);
10130 48 : all_layers
10131 48 : }
10132 :
10133 : #[cfg(feature = "testing")]
10134 44 : fn check_layer_map_key_eq(
10135 44 : mut left: Vec<PersistentLayerKey>,
10136 44 : mut right: Vec<PersistentLayerKey>,
10137 44 : ) {
10138 44 : left.sort_by(sort_layer_key);
10139 44 : right.sort_by(sort_layer_key);
10140 44 : if left != right {
10141 0 : eprintln!("---LEFT---");
10142 0 : for left in left.iter() {
10143 0 : eprintln!("{}", left);
10144 0 : }
10145 0 : eprintln!("---RIGHT---");
10146 0 : for right in right.iter() {
10147 0 : eprintln!("{}", right);
10148 0 : }
10149 0 : assert_eq!(left, right);
10150 44 : }
10151 44 : }
10152 :
10153 : #[cfg(feature = "testing")]
10154 : #[tokio::test]
10155 4 : async fn test_simple_partial_bottom_most_compaction() -> anyhow::Result<()> {
10156 4 : let harness = TenantHarness::create("test_simple_partial_bottom_most_compaction").await?;
10157 4 : let (tenant, ctx) = harness.load().await;
10158 4 :
10159 364 : fn get_key(id: u32) -> Key {
10160 364 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
10161 364 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
10162 364 : key.field6 = id;
10163 364 : key
10164 364 : }
10165 4 :
10166 4 : // img layer at 0x10
10167 4 : let img_layer = (0..10)
10168 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10169 4 : .collect_vec();
10170 4 :
10171 4 : let delta1 = vec![
10172 4 : (
10173 4 : get_key(1),
10174 4 : Lsn(0x20),
10175 4 : Value::Image(Bytes::from("value 1@0x20")),
10176 4 : ),
10177 4 : (
10178 4 : get_key(2),
10179 4 : Lsn(0x30),
10180 4 : Value::Image(Bytes::from("value 2@0x30")),
10181 4 : ),
10182 4 : (
10183 4 : get_key(3),
10184 4 : Lsn(0x40),
10185 4 : Value::Image(Bytes::from("value 3@0x40")),
10186 4 : ),
10187 4 : ];
10188 4 : let delta2 = vec![
10189 4 : (
10190 4 : get_key(5),
10191 4 : Lsn(0x20),
10192 4 : Value::Image(Bytes::from("value 5@0x20")),
10193 4 : ),
10194 4 : (
10195 4 : get_key(6),
10196 4 : Lsn(0x20),
10197 4 : Value::Image(Bytes::from("value 6@0x20")),
10198 4 : ),
10199 4 : ];
10200 4 : let delta3 = vec![
10201 4 : (
10202 4 : get_key(8),
10203 4 : Lsn(0x48),
10204 4 : Value::Image(Bytes::from("value 8@0x48")),
10205 4 : ),
10206 4 : (
10207 4 : get_key(9),
10208 4 : Lsn(0x48),
10209 4 : Value::Image(Bytes::from("value 9@0x48")),
10210 4 : ),
10211 4 : ];
10212 4 :
10213 4 : let tline = tenant
10214 4 : .create_test_timeline_with_layers(
10215 4 : TIMELINE_ID,
10216 4 : Lsn(0x10),
10217 4 : DEFAULT_PG_VERSION,
10218 4 : &ctx,
10219 4 : vec![
10220 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
10221 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
10222 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
10223 4 : ], // delta layers
10224 4 : vec![(Lsn(0x10), img_layer)], // image layers
10225 4 : Lsn(0x50),
10226 4 : )
10227 4 : .await?;
10228 4 :
10229 4 : {
10230 4 : tline
10231 4 : .latest_gc_cutoff_lsn
10232 4 : .lock_for_write()
10233 4 : .store_and_unlock(Lsn(0x30))
10234 4 : .wait()
10235 4 : .await;
10236 4 : // Update GC info
10237 4 : let mut guard = tline.gc_info.write().unwrap();
10238 4 : *guard = GcInfo {
10239 4 : retain_lsns: vec![(Lsn(0x20), tline.timeline_id, MaybeOffloaded::No)],
10240 4 : cutoffs: GcCutoffs {
10241 4 : time: Lsn(0x30),
10242 4 : space: Lsn(0x30),
10243 4 : },
10244 4 : leases: Default::default(),
10245 4 : within_ancestor_pitr: false,
10246 4 : };
10247 4 : }
10248 4 :
10249 4 : let cancel = CancellationToken::new();
10250 4 :
10251 4 : // Do a partial compaction on key range 0..2
10252 4 : tline
10253 4 : .compact_with_gc(
10254 4 : &cancel,
10255 4 : CompactOptions {
10256 4 : flags: EnumSet::new(),
10257 4 : compact_key_range: Some((get_key(0)..get_key(2)).into()),
10258 4 : ..Default::default()
10259 4 : },
10260 4 : &ctx,
10261 4 : )
10262 4 : .await
10263 4 : .unwrap();
10264 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10265 4 : check_layer_map_key_eq(
10266 4 : all_layers,
10267 4 : vec![
10268 4 : // newly-generated image layer for the partial compaction range 0-2
10269 4 : PersistentLayerKey {
10270 4 : key_range: get_key(0)..get_key(2),
10271 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10272 4 : is_delta: false,
10273 4 : },
10274 4 : PersistentLayerKey {
10275 4 : key_range: get_key(0)..get_key(10),
10276 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
10277 4 : is_delta: false,
10278 4 : },
10279 4 : // delta1 is split and the second part is rewritten
10280 4 : PersistentLayerKey {
10281 4 : key_range: get_key(2)..get_key(4),
10282 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10283 4 : is_delta: true,
10284 4 : },
10285 4 : PersistentLayerKey {
10286 4 : key_range: get_key(5)..get_key(7),
10287 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10288 4 : is_delta: true,
10289 4 : },
10290 4 : PersistentLayerKey {
10291 4 : key_range: get_key(8)..get_key(10),
10292 4 : lsn_range: Lsn(0x48)..Lsn(0x50),
10293 4 : is_delta: true,
10294 4 : },
10295 4 : ],
10296 4 : );
10297 4 :
10298 4 : // Do a partial compaction on key range 2..4
10299 4 : tline
10300 4 : .compact_with_gc(
10301 4 : &cancel,
10302 4 : CompactOptions {
10303 4 : flags: EnumSet::new(),
10304 4 : compact_key_range: Some((get_key(2)..get_key(4)).into()),
10305 4 : ..Default::default()
10306 4 : },
10307 4 : &ctx,
10308 4 : )
10309 4 : .await
10310 4 : .unwrap();
10311 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10312 4 : check_layer_map_key_eq(
10313 4 : all_layers,
10314 4 : vec![
10315 4 : PersistentLayerKey {
10316 4 : key_range: get_key(0)..get_key(2),
10317 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10318 4 : is_delta: false,
10319 4 : },
10320 4 : PersistentLayerKey {
10321 4 : key_range: get_key(0)..get_key(10),
10322 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
10323 4 : is_delta: false,
10324 4 : },
10325 4 : // image layer generated for the compaction range 2-4
10326 4 : PersistentLayerKey {
10327 4 : key_range: get_key(2)..get_key(4),
10328 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10329 4 : is_delta: false,
10330 4 : },
10331 4 : // we have key2/key3 above the retain_lsn, so we still need this delta layer
10332 4 : PersistentLayerKey {
10333 4 : key_range: get_key(2)..get_key(4),
10334 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10335 4 : is_delta: true,
10336 4 : },
10337 4 : PersistentLayerKey {
10338 4 : key_range: get_key(5)..get_key(7),
10339 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10340 4 : is_delta: true,
10341 4 : },
10342 4 : PersistentLayerKey {
10343 4 : key_range: get_key(8)..get_key(10),
10344 4 : lsn_range: Lsn(0x48)..Lsn(0x50),
10345 4 : is_delta: true,
10346 4 : },
10347 4 : ],
10348 4 : );
10349 4 :
10350 4 : // Do a partial compaction on key range 4..9
10351 4 : tline
10352 4 : .compact_with_gc(
10353 4 : &cancel,
10354 4 : CompactOptions {
10355 4 : flags: EnumSet::new(),
10356 4 : compact_key_range: Some((get_key(4)..get_key(9)).into()),
10357 4 : ..Default::default()
10358 4 : },
10359 4 : &ctx,
10360 4 : )
10361 4 : .await
10362 4 : .unwrap();
10363 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10364 4 : check_layer_map_key_eq(
10365 4 : all_layers,
10366 4 : vec![
10367 4 : PersistentLayerKey {
10368 4 : key_range: get_key(0)..get_key(2),
10369 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10370 4 : is_delta: false,
10371 4 : },
10372 4 : PersistentLayerKey {
10373 4 : key_range: get_key(0)..get_key(10),
10374 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
10375 4 : is_delta: false,
10376 4 : },
10377 4 : PersistentLayerKey {
10378 4 : key_range: get_key(2)..get_key(4),
10379 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10380 4 : is_delta: false,
10381 4 : },
10382 4 : PersistentLayerKey {
10383 4 : key_range: get_key(2)..get_key(4),
10384 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10385 4 : is_delta: true,
10386 4 : },
10387 4 : // image layer generated for this compaction range
10388 4 : PersistentLayerKey {
10389 4 : key_range: get_key(4)..get_key(9),
10390 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10391 4 : is_delta: false,
10392 4 : },
10393 4 : PersistentLayerKey {
10394 4 : key_range: get_key(8)..get_key(10),
10395 4 : lsn_range: Lsn(0x48)..Lsn(0x50),
10396 4 : is_delta: true,
10397 4 : },
10398 4 : ],
10399 4 : );
10400 4 :
10401 4 : // Do a partial compaction on key range 9..10
10402 4 : tline
10403 4 : .compact_with_gc(
10404 4 : &cancel,
10405 4 : CompactOptions {
10406 4 : flags: EnumSet::new(),
10407 4 : compact_key_range: Some((get_key(9)..get_key(10)).into()),
10408 4 : ..Default::default()
10409 4 : },
10410 4 : &ctx,
10411 4 : )
10412 4 : .await
10413 4 : .unwrap();
10414 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10415 4 : check_layer_map_key_eq(
10416 4 : all_layers,
10417 4 : vec![
10418 4 : PersistentLayerKey {
10419 4 : key_range: get_key(0)..get_key(2),
10420 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10421 4 : is_delta: false,
10422 4 : },
10423 4 : PersistentLayerKey {
10424 4 : key_range: get_key(0)..get_key(10),
10425 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
10426 4 : is_delta: false,
10427 4 : },
10428 4 : PersistentLayerKey {
10429 4 : key_range: get_key(2)..get_key(4),
10430 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10431 4 : is_delta: false,
10432 4 : },
10433 4 : PersistentLayerKey {
10434 4 : key_range: get_key(2)..get_key(4),
10435 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10436 4 : is_delta: true,
10437 4 : },
10438 4 : PersistentLayerKey {
10439 4 : key_range: get_key(4)..get_key(9),
10440 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10441 4 : is_delta: false,
10442 4 : },
10443 4 : // image layer generated for the compaction range
10444 4 : PersistentLayerKey {
10445 4 : key_range: get_key(9)..get_key(10),
10446 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10447 4 : is_delta: false,
10448 4 : },
10449 4 : PersistentLayerKey {
10450 4 : key_range: get_key(8)..get_key(10),
10451 4 : lsn_range: Lsn(0x48)..Lsn(0x50),
10452 4 : is_delta: true,
10453 4 : },
10454 4 : ],
10455 4 : );
10456 4 :
10457 4 : // Do a partial compaction on key range 0..10, all image layers below LSN 20 can be replaced with new ones.
10458 4 : tline
10459 4 : .compact_with_gc(
10460 4 : &cancel,
10461 4 : CompactOptions {
10462 4 : flags: EnumSet::new(),
10463 4 : compact_key_range: Some((get_key(0)..get_key(10)).into()),
10464 4 : ..Default::default()
10465 4 : },
10466 4 : &ctx,
10467 4 : )
10468 4 : .await
10469 4 : .unwrap();
10470 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10471 4 : check_layer_map_key_eq(
10472 4 : all_layers,
10473 4 : vec![
10474 4 : // aha, we removed all unnecessary image/delta layers and got a very clean layer map!
10475 4 : PersistentLayerKey {
10476 4 : key_range: get_key(0)..get_key(10),
10477 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10478 4 : is_delta: false,
10479 4 : },
10480 4 : PersistentLayerKey {
10481 4 : key_range: get_key(2)..get_key(4),
10482 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10483 4 : is_delta: true,
10484 4 : },
10485 4 : PersistentLayerKey {
10486 4 : key_range: get_key(8)..get_key(10),
10487 4 : lsn_range: Lsn(0x48)..Lsn(0x50),
10488 4 : is_delta: true,
10489 4 : },
10490 4 : ],
10491 4 : );
10492 4 : Ok(())
10493 4 : }
10494 :
10495 : #[cfg(feature = "testing")]
10496 : #[tokio::test]
10497 4 : async fn test_timeline_offload_retain_lsn() -> anyhow::Result<()> {
10498 4 : let harness = TenantHarness::create("test_timeline_offload_retain_lsn")
10499 4 : .await
10500 4 : .unwrap();
10501 4 : let (tenant, ctx) = harness.load().await;
10502 4 : let tline_parent = tenant
10503 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
10504 4 : .await
10505 4 : .unwrap();
10506 4 : let tline_child = tenant
10507 4 : .branch_timeline_test(&tline_parent, NEW_TIMELINE_ID, Some(Lsn(0x20)), &ctx)
10508 4 : .await
10509 4 : .unwrap();
10510 4 : {
10511 4 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
10512 4 : assert_eq!(
10513 4 : gc_info_parent.retain_lsns,
10514 4 : vec![(Lsn(0x20), tline_child.timeline_id, MaybeOffloaded::No)]
10515 4 : );
10516 4 : }
10517 4 : // We have to directly call the remote_client instead of using the archive function to avoid constructing broker client...
10518 4 : tline_child
10519 4 : .remote_client
10520 4 : .schedule_index_upload_for_timeline_archival_state(TimelineArchivalState::Archived)
10521 4 : .unwrap();
10522 4 : tline_child.remote_client.wait_completion().await.unwrap();
10523 4 : offload_timeline(&tenant, &tline_child)
10524 4 : .instrument(tracing::info_span!(parent: None, "offload_test", tenant_id=%"test", shard_id=%"test", timeline_id=%"test"))
10525 4 : .await.unwrap();
10526 4 : let child_timeline_id = tline_child.timeline_id;
10527 4 : Arc::try_unwrap(tline_child).unwrap();
10528 4 :
10529 4 : {
10530 4 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
10531 4 : assert_eq!(
10532 4 : gc_info_parent.retain_lsns,
10533 4 : vec![(Lsn(0x20), child_timeline_id, MaybeOffloaded::Yes)]
10534 4 : );
10535 4 : }
10536 4 :
10537 4 : tenant
10538 4 : .get_offloaded_timeline(child_timeline_id)
10539 4 : .unwrap()
10540 4 : .defuse_for_tenant_drop();
10541 4 :
10542 4 : Ok(())
10543 4 : }
10544 :
10545 : #[cfg(feature = "testing")]
10546 : #[tokio::test]
10547 4 : async fn test_simple_bottom_most_compaction_above_lsn() -> anyhow::Result<()> {
10548 4 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_above_lsn").await?;
10549 4 : let (tenant, ctx) = harness.load().await;
10550 4 :
10551 592 : fn get_key(id: u32) -> Key {
10552 592 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
10553 592 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
10554 592 : key.field6 = id;
10555 592 : key
10556 592 : }
10557 4 :
10558 4 : let img_layer = (0..10)
10559 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10560 4 : .collect_vec();
10561 4 :
10562 4 : let delta1 = vec![(
10563 4 : get_key(1),
10564 4 : Lsn(0x20),
10565 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10566 4 : )];
10567 4 : let delta4 = vec![(
10568 4 : get_key(1),
10569 4 : Lsn(0x28),
10570 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
10571 4 : )];
10572 4 : let delta2 = vec![
10573 4 : (
10574 4 : get_key(1),
10575 4 : Lsn(0x30),
10576 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10577 4 : ),
10578 4 : (
10579 4 : get_key(1),
10580 4 : Lsn(0x38),
10581 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
10582 4 : ),
10583 4 : ];
10584 4 : let delta3 = vec![
10585 4 : (
10586 4 : get_key(8),
10587 4 : Lsn(0x48),
10588 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10589 4 : ),
10590 4 : (
10591 4 : get_key(9),
10592 4 : Lsn(0x48),
10593 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10594 4 : ),
10595 4 : ];
10596 4 :
10597 4 : let tline = tenant
10598 4 : .create_test_timeline_with_layers(
10599 4 : TIMELINE_ID,
10600 4 : Lsn(0x10),
10601 4 : DEFAULT_PG_VERSION,
10602 4 : &ctx,
10603 4 : vec![
10604 4 : // delta1/2/4 only contain a single key but multiple updates
10605 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x28), delta1),
10606 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
10607 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x28)..Lsn(0x30), delta4),
10608 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta3),
10609 4 : ], // delta layers
10610 4 : vec![(Lsn(0x10), img_layer)], // image layers
10611 4 : Lsn(0x50),
10612 4 : )
10613 4 : .await?;
10614 4 : {
10615 4 : tline
10616 4 : .latest_gc_cutoff_lsn
10617 4 : .lock_for_write()
10618 4 : .store_and_unlock(Lsn(0x30))
10619 4 : .wait()
10620 4 : .await;
10621 4 : // Update GC info
10622 4 : let mut guard = tline.gc_info.write().unwrap();
10623 4 : *guard = GcInfo {
10624 4 : retain_lsns: vec![
10625 4 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
10626 4 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
10627 4 : ],
10628 4 : cutoffs: GcCutoffs {
10629 4 : time: Lsn(0x30),
10630 4 : space: Lsn(0x30),
10631 4 : },
10632 4 : leases: Default::default(),
10633 4 : within_ancestor_pitr: false,
10634 4 : };
10635 4 : }
10636 4 :
10637 4 : let expected_result = [
10638 4 : Bytes::from_static(b"value 0@0x10"),
10639 4 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
10640 4 : Bytes::from_static(b"value 2@0x10"),
10641 4 : Bytes::from_static(b"value 3@0x10"),
10642 4 : Bytes::from_static(b"value 4@0x10"),
10643 4 : Bytes::from_static(b"value 5@0x10"),
10644 4 : Bytes::from_static(b"value 6@0x10"),
10645 4 : Bytes::from_static(b"value 7@0x10"),
10646 4 : Bytes::from_static(b"value 8@0x10@0x48"),
10647 4 : Bytes::from_static(b"value 9@0x10@0x48"),
10648 4 : ];
10649 4 :
10650 4 : let expected_result_at_gc_horizon = [
10651 4 : Bytes::from_static(b"value 0@0x10"),
10652 4 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
10653 4 : Bytes::from_static(b"value 2@0x10"),
10654 4 : Bytes::from_static(b"value 3@0x10"),
10655 4 : Bytes::from_static(b"value 4@0x10"),
10656 4 : Bytes::from_static(b"value 5@0x10"),
10657 4 : Bytes::from_static(b"value 6@0x10"),
10658 4 : Bytes::from_static(b"value 7@0x10"),
10659 4 : Bytes::from_static(b"value 8@0x10"),
10660 4 : Bytes::from_static(b"value 9@0x10"),
10661 4 : ];
10662 4 :
10663 4 : let expected_result_at_lsn_20 = [
10664 4 : Bytes::from_static(b"value 0@0x10"),
10665 4 : Bytes::from_static(b"value 1@0x10@0x20"),
10666 4 : Bytes::from_static(b"value 2@0x10"),
10667 4 : Bytes::from_static(b"value 3@0x10"),
10668 4 : Bytes::from_static(b"value 4@0x10"),
10669 4 : Bytes::from_static(b"value 5@0x10"),
10670 4 : Bytes::from_static(b"value 6@0x10"),
10671 4 : Bytes::from_static(b"value 7@0x10"),
10672 4 : Bytes::from_static(b"value 8@0x10"),
10673 4 : Bytes::from_static(b"value 9@0x10"),
10674 4 : ];
10675 4 :
10676 4 : let expected_result_at_lsn_10 = [
10677 4 : Bytes::from_static(b"value 0@0x10"),
10678 4 : Bytes::from_static(b"value 1@0x10"),
10679 4 : Bytes::from_static(b"value 2@0x10"),
10680 4 : Bytes::from_static(b"value 3@0x10"),
10681 4 : Bytes::from_static(b"value 4@0x10"),
10682 4 : Bytes::from_static(b"value 5@0x10"),
10683 4 : Bytes::from_static(b"value 6@0x10"),
10684 4 : Bytes::from_static(b"value 7@0x10"),
10685 4 : Bytes::from_static(b"value 8@0x10"),
10686 4 : Bytes::from_static(b"value 9@0x10"),
10687 4 : ];
10688 4 :
10689 12 : let verify_result = || async {
10690 12 : let gc_horizon = {
10691 12 : let gc_info = tline.gc_info.read().unwrap();
10692 12 : gc_info.cutoffs.time
10693 4 : };
10694 132 : for idx in 0..10 {
10695 120 : assert_eq!(
10696 120 : tline
10697 120 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10698 120 : .await
10699 120 : .unwrap(),
10700 120 : &expected_result[idx]
10701 4 : );
10702 120 : assert_eq!(
10703 120 : tline
10704 120 : .get(get_key(idx as u32), gc_horizon, &ctx)
10705 120 : .await
10706 120 : .unwrap(),
10707 120 : &expected_result_at_gc_horizon[idx]
10708 4 : );
10709 120 : assert_eq!(
10710 120 : tline
10711 120 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
10712 120 : .await
10713 120 : .unwrap(),
10714 120 : &expected_result_at_lsn_20[idx]
10715 4 : );
10716 120 : assert_eq!(
10717 120 : tline
10718 120 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
10719 120 : .await
10720 120 : .unwrap(),
10721 120 : &expected_result_at_lsn_10[idx]
10722 4 : );
10723 4 : }
10724 24 : };
10725 4 :
10726 4 : verify_result().await;
10727 4 :
10728 4 : let cancel = CancellationToken::new();
10729 4 : tline
10730 4 : .compact_with_gc(
10731 4 : &cancel,
10732 4 : CompactOptions {
10733 4 : compact_lsn_range: Some(CompactLsnRange::above(Lsn(0x28))),
10734 4 : ..Default::default()
10735 4 : },
10736 4 : &ctx,
10737 4 : )
10738 4 : .await
10739 4 : .unwrap();
10740 4 : verify_result().await;
10741 4 :
10742 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10743 4 : check_layer_map_key_eq(
10744 4 : all_layers,
10745 4 : vec![
10746 4 : // The original image layer, not compacted
10747 4 : PersistentLayerKey {
10748 4 : key_range: get_key(0)..get_key(10),
10749 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
10750 4 : is_delta: false,
10751 4 : },
10752 4 : // Delta layer below the specified above_lsn not compacted
10753 4 : PersistentLayerKey {
10754 4 : key_range: get_key(1)..get_key(2),
10755 4 : lsn_range: Lsn(0x20)..Lsn(0x28),
10756 4 : is_delta: true,
10757 4 : },
10758 4 : // Delta layer compacted above the LSN
10759 4 : PersistentLayerKey {
10760 4 : key_range: get_key(1)..get_key(10),
10761 4 : lsn_range: Lsn(0x28)..Lsn(0x50),
10762 4 : is_delta: true,
10763 4 : },
10764 4 : ],
10765 4 : );
10766 4 :
10767 4 : // compact again
10768 4 : tline
10769 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10770 4 : .await
10771 4 : .unwrap();
10772 4 : verify_result().await;
10773 4 :
10774 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10775 4 : check_layer_map_key_eq(
10776 4 : all_layers,
10777 4 : vec![
10778 4 : // The compacted image layer (full key range)
10779 4 : PersistentLayerKey {
10780 4 : key_range: Key::MIN..Key::MAX,
10781 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
10782 4 : is_delta: false,
10783 4 : },
10784 4 : // All other data in the delta layer
10785 4 : PersistentLayerKey {
10786 4 : key_range: get_key(1)..get_key(10),
10787 4 : lsn_range: Lsn(0x10)..Lsn(0x50),
10788 4 : is_delta: true,
10789 4 : },
10790 4 : ],
10791 4 : );
10792 4 :
10793 4 : Ok(())
10794 4 : }
10795 :
10796 : #[cfg(feature = "testing")]
10797 : #[tokio::test]
10798 4 : async fn test_simple_bottom_most_compaction_rectangle() -> anyhow::Result<()> {
10799 4 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_rectangle").await?;
10800 4 : let (tenant, ctx) = harness.load().await;
10801 4 :
10802 1016 : fn get_key(id: u32) -> Key {
10803 1016 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
10804 1016 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
10805 1016 : key.field6 = id;
10806 1016 : key
10807 1016 : }
10808 4 :
10809 4 : let img_layer = (0..10)
10810 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10811 4 : .collect_vec();
10812 4 :
10813 4 : let delta1 = vec![(
10814 4 : get_key(1),
10815 4 : Lsn(0x20),
10816 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10817 4 : )];
10818 4 : let delta4 = vec![(
10819 4 : get_key(1),
10820 4 : Lsn(0x28),
10821 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
10822 4 : )];
10823 4 : let delta2 = vec![
10824 4 : (
10825 4 : get_key(1),
10826 4 : Lsn(0x30),
10827 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10828 4 : ),
10829 4 : (
10830 4 : get_key(1),
10831 4 : Lsn(0x38),
10832 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
10833 4 : ),
10834 4 : ];
10835 4 : let delta3 = vec![
10836 4 : (
10837 4 : get_key(8),
10838 4 : Lsn(0x48),
10839 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10840 4 : ),
10841 4 : (
10842 4 : get_key(9),
10843 4 : Lsn(0x48),
10844 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10845 4 : ),
10846 4 : ];
10847 4 :
10848 4 : let tline = tenant
10849 4 : .create_test_timeline_with_layers(
10850 4 : TIMELINE_ID,
10851 4 : Lsn(0x10),
10852 4 : DEFAULT_PG_VERSION,
10853 4 : &ctx,
10854 4 : vec![
10855 4 : // delta1/2/4 only contain a single key but multiple updates
10856 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x28), delta1),
10857 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
10858 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x28)..Lsn(0x30), delta4),
10859 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta3),
10860 4 : ], // delta layers
10861 4 : vec![(Lsn(0x10), img_layer)], // image layers
10862 4 : Lsn(0x50),
10863 4 : )
10864 4 : .await?;
10865 4 : {
10866 4 : tline
10867 4 : .latest_gc_cutoff_lsn
10868 4 : .lock_for_write()
10869 4 : .store_and_unlock(Lsn(0x30))
10870 4 : .wait()
10871 4 : .await;
10872 4 : // Update GC info
10873 4 : let mut guard = tline.gc_info.write().unwrap();
10874 4 : *guard = GcInfo {
10875 4 : retain_lsns: vec![
10876 4 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
10877 4 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
10878 4 : ],
10879 4 : cutoffs: GcCutoffs {
10880 4 : time: Lsn(0x30),
10881 4 : space: Lsn(0x30),
10882 4 : },
10883 4 : leases: Default::default(),
10884 4 : within_ancestor_pitr: false,
10885 4 : };
10886 4 : }
10887 4 :
10888 4 : let expected_result = [
10889 4 : Bytes::from_static(b"value 0@0x10"),
10890 4 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
10891 4 : Bytes::from_static(b"value 2@0x10"),
10892 4 : Bytes::from_static(b"value 3@0x10"),
10893 4 : Bytes::from_static(b"value 4@0x10"),
10894 4 : Bytes::from_static(b"value 5@0x10"),
10895 4 : Bytes::from_static(b"value 6@0x10"),
10896 4 : Bytes::from_static(b"value 7@0x10"),
10897 4 : Bytes::from_static(b"value 8@0x10@0x48"),
10898 4 : Bytes::from_static(b"value 9@0x10@0x48"),
10899 4 : ];
10900 4 :
10901 4 : let expected_result_at_gc_horizon = [
10902 4 : Bytes::from_static(b"value 0@0x10"),
10903 4 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
10904 4 : Bytes::from_static(b"value 2@0x10"),
10905 4 : Bytes::from_static(b"value 3@0x10"),
10906 4 : Bytes::from_static(b"value 4@0x10"),
10907 4 : Bytes::from_static(b"value 5@0x10"),
10908 4 : Bytes::from_static(b"value 6@0x10"),
10909 4 : Bytes::from_static(b"value 7@0x10"),
10910 4 : Bytes::from_static(b"value 8@0x10"),
10911 4 : Bytes::from_static(b"value 9@0x10"),
10912 4 : ];
10913 4 :
10914 4 : let expected_result_at_lsn_20 = [
10915 4 : Bytes::from_static(b"value 0@0x10"),
10916 4 : Bytes::from_static(b"value 1@0x10@0x20"),
10917 4 : Bytes::from_static(b"value 2@0x10"),
10918 4 : Bytes::from_static(b"value 3@0x10"),
10919 4 : Bytes::from_static(b"value 4@0x10"),
10920 4 : Bytes::from_static(b"value 5@0x10"),
10921 4 : Bytes::from_static(b"value 6@0x10"),
10922 4 : Bytes::from_static(b"value 7@0x10"),
10923 4 : Bytes::from_static(b"value 8@0x10"),
10924 4 : Bytes::from_static(b"value 9@0x10"),
10925 4 : ];
10926 4 :
10927 4 : let expected_result_at_lsn_10 = [
10928 4 : Bytes::from_static(b"value 0@0x10"),
10929 4 : Bytes::from_static(b"value 1@0x10"),
10930 4 : Bytes::from_static(b"value 2@0x10"),
10931 4 : Bytes::from_static(b"value 3@0x10"),
10932 4 : Bytes::from_static(b"value 4@0x10"),
10933 4 : Bytes::from_static(b"value 5@0x10"),
10934 4 : Bytes::from_static(b"value 6@0x10"),
10935 4 : Bytes::from_static(b"value 7@0x10"),
10936 4 : Bytes::from_static(b"value 8@0x10"),
10937 4 : Bytes::from_static(b"value 9@0x10"),
10938 4 : ];
10939 4 :
10940 20 : let verify_result = || async {
10941 20 : let gc_horizon = {
10942 20 : let gc_info = tline.gc_info.read().unwrap();
10943 20 : gc_info.cutoffs.time
10944 4 : };
10945 220 : for idx in 0..10 {
10946 200 : assert_eq!(
10947 200 : tline
10948 200 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10949 200 : .await
10950 200 : .unwrap(),
10951 200 : &expected_result[idx]
10952 4 : );
10953 200 : assert_eq!(
10954 200 : tline
10955 200 : .get(get_key(idx as u32), gc_horizon, &ctx)
10956 200 : .await
10957 200 : .unwrap(),
10958 200 : &expected_result_at_gc_horizon[idx]
10959 4 : );
10960 200 : assert_eq!(
10961 200 : tline
10962 200 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
10963 200 : .await
10964 200 : .unwrap(),
10965 200 : &expected_result_at_lsn_20[idx]
10966 4 : );
10967 200 : assert_eq!(
10968 200 : tline
10969 200 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
10970 200 : .await
10971 200 : .unwrap(),
10972 200 : &expected_result_at_lsn_10[idx]
10973 4 : );
10974 4 : }
10975 40 : };
10976 4 :
10977 4 : verify_result().await;
10978 4 :
10979 4 : let cancel = CancellationToken::new();
10980 4 :
10981 4 : tline
10982 4 : .compact_with_gc(
10983 4 : &cancel,
10984 4 : CompactOptions {
10985 4 : compact_key_range: Some((get_key(0)..get_key(2)).into()),
10986 4 : compact_lsn_range: Some((Lsn(0x20)..Lsn(0x28)).into()),
10987 4 : ..Default::default()
10988 4 : },
10989 4 : &ctx,
10990 4 : )
10991 4 : .await
10992 4 : .unwrap();
10993 4 : verify_result().await;
10994 4 :
10995 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10996 4 : check_layer_map_key_eq(
10997 4 : all_layers,
10998 4 : vec![
10999 4 : // The original image layer, not compacted
11000 4 : PersistentLayerKey {
11001 4 : key_range: get_key(0)..get_key(10),
11002 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
11003 4 : is_delta: false,
11004 4 : },
11005 4 : // According the selection logic, we select all layers with start key <= 0x28, so we would merge the layer 0x20-0x28 and
11006 4 : // the layer 0x28-0x30 into one.
11007 4 : PersistentLayerKey {
11008 4 : key_range: get_key(1)..get_key(2),
11009 4 : lsn_range: Lsn(0x20)..Lsn(0x30),
11010 4 : is_delta: true,
11011 4 : },
11012 4 : // Above the upper bound and untouched
11013 4 : PersistentLayerKey {
11014 4 : key_range: get_key(1)..get_key(2),
11015 4 : lsn_range: Lsn(0x30)..Lsn(0x50),
11016 4 : is_delta: true,
11017 4 : },
11018 4 : // This layer is untouched
11019 4 : PersistentLayerKey {
11020 4 : key_range: get_key(8)..get_key(10),
11021 4 : lsn_range: Lsn(0x30)..Lsn(0x50),
11022 4 : is_delta: true,
11023 4 : },
11024 4 : ],
11025 4 : );
11026 4 :
11027 4 : tline
11028 4 : .compact_with_gc(
11029 4 : &cancel,
11030 4 : CompactOptions {
11031 4 : compact_key_range: Some((get_key(3)..get_key(8)).into()),
11032 4 : compact_lsn_range: Some((Lsn(0x28)..Lsn(0x40)).into()),
11033 4 : ..Default::default()
11034 4 : },
11035 4 : &ctx,
11036 4 : )
11037 4 : .await
11038 4 : .unwrap();
11039 4 : verify_result().await;
11040 4 :
11041 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11042 4 : check_layer_map_key_eq(
11043 4 : all_layers,
11044 4 : vec![
11045 4 : // The original image layer, not compacted
11046 4 : PersistentLayerKey {
11047 4 : key_range: get_key(0)..get_key(10),
11048 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
11049 4 : is_delta: false,
11050 4 : },
11051 4 : // Not in the compaction key range, uncompacted
11052 4 : PersistentLayerKey {
11053 4 : key_range: get_key(1)..get_key(2),
11054 4 : lsn_range: Lsn(0x20)..Lsn(0x30),
11055 4 : is_delta: true,
11056 4 : },
11057 4 : // Not in the compaction key range, uncompacted but need rewrite because the delta layer overlaps with the range
11058 4 : PersistentLayerKey {
11059 4 : key_range: get_key(1)..get_key(2),
11060 4 : lsn_range: Lsn(0x30)..Lsn(0x50),
11061 4 : is_delta: true,
11062 4 : },
11063 4 : // Note that when we specify the LSN upper bound to be 0x40, the compaction algorithm will not try to cut the layer
11064 4 : // horizontally in half. Instead, it will include all LSNs that overlap with 0x40. So the real max_lsn of the compaction
11065 4 : // becomes 0x50.
11066 4 : PersistentLayerKey {
11067 4 : key_range: get_key(8)..get_key(10),
11068 4 : lsn_range: Lsn(0x30)..Lsn(0x50),
11069 4 : is_delta: true,
11070 4 : },
11071 4 : ],
11072 4 : );
11073 4 :
11074 4 : // compact again
11075 4 : tline
11076 4 : .compact_with_gc(
11077 4 : &cancel,
11078 4 : CompactOptions {
11079 4 : compact_key_range: Some((get_key(0)..get_key(5)).into()),
11080 4 : compact_lsn_range: Some((Lsn(0x20)..Lsn(0x50)).into()),
11081 4 : ..Default::default()
11082 4 : },
11083 4 : &ctx,
11084 4 : )
11085 4 : .await
11086 4 : .unwrap();
11087 4 : verify_result().await;
11088 4 :
11089 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11090 4 : check_layer_map_key_eq(
11091 4 : all_layers,
11092 4 : vec![
11093 4 : // The original image layer, not compacted
11094 4 : PersistentLayerKey {
11095 4 : key_range: get_key(0)..get_key(10),
11096 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
11097 4 : is_delta: false,
11098 4 : },
11099 4 : // The range gets compacted
11100 4 : PersistentLayerKey {
11101 4 : key_range: get_key(1)..get_key(2),
11102 4 : lsn_range: Lsn(0x20)..Lsn(0x50),
11103 4 : is_delta: true,
11104 4 : },
11105 4 : // Not touched during this iteration of compaction
11106 4 : PersistentLayerKey {
11107 4 : key_range: get_key(8)..get_key(10),
11108 4 : lsn_range: Lsn(0x30)..Lsn(0x50),
11109 4 : is_delta: true,
11110 4 : },
11111 4 : ],
11112 4 : );
11113 4 :
11114 4 : // final full compaction
11115 4 : tline
11116 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
11117 4 : .await
11118 4 : .unwrap();
11119 4 : verify_result().await;
11120 4 :
11121 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11122 4 : check_layer_map_key_eq(
11123 4 : all_layers,
11124 4 : vec![
11125 4 : // The compacted image layer (full key range)
11126 4 : PersistentLayerKey {
11127 4 : key_range: Key::MIN..Key::MAX,
11128 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
11129 4 : is_delta: false,
11130 4 : },
11131 4 : // All other data in the delta layer
11132 4 : PersistentLayerKey {
11133 4 : key_range: get_key(1)..get_key(10),
11134 4 : lsn_range: Lsn(0x10)..Lsn(0x50),
11135 4 : is_delta: true,
11136 4 : },
11137 4 : ],
11138 4 : );
11139 4 :
11140 4 : Ok(())
11141 4 : }
11142 : }
|
