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 itertools::Itertools as _;
24 : use pageserver_api::models;
25 : use pageserver_api::models::CompactInfoResponse;
26 : use pageserver_api::models::LsnLease;
27 : use pageserver_api::models::TimelineArchivalState;
28 : use pageserver_api::models::TimelineState;
29 : use pageserver_api::models::TopTenantShardItem;
30 : use pageserver_api::models::WalRedoManagerStatus;
31 : use pageserver_api::shard::ShardIdentity;
32 : use pageserver_api::shard::ShardStripeSize;
33 : use pageserver_api::shard::TenantShardId;
34 : use remote_storage::DownloadError;
35 : use remote_storage::GenericRemoteStorage;
36 : use remote_storage::TimeoutOrCancel;
37 : use remote_timeline_client::manifest::{
38 : OffloadedTimelineManifest, TenantManifest, LATEST_TENANT_MANIFEST_VERSION,
39 : };
40 : use remote_timeline_client::UploadQueueNotReadyError;
41 : use remote_timeline_client::FAILED_REMOTE_OP_RETRIES;
42 : use remote_timeline_client::FAILED_UPLOAD_WARN_THRESHOLD;
43 : use secondary::heatmap::HeatMapTenant;
44 : use secondary::heatmap::HeatMapTimeline;
45 : use std::collections::BTreeMap;
46 : use std::fmt;
47 : use std::future::Future;
48 : use std::sync::atomic::AtomicBool;
49 : use std::sync::Weak;
50 : use std::time::SystemTime;
51 : use storage_broker::BrokerClientChannel;
52 : use timeline::compaction::CompactionOutcome;
53 : use timeline::compaction::GcCompactionQueue;
54 : use timeline::import_pgdata;
55 : use timeline::offload::offload_timeline;
56 : use timeline::offload::OffloadError;
57 : use timeline::CompactFlags;
58 : use timeline::CompactOptions;
59 : use timeline::CompactionError;
60 : use timeline::PreviousHeatmap;
61 : use timeline::ShutdownMode;
62 : use tokio::io::BufReader;
63 : use tokio::sync::watch;
64 : use tokio::sync::Notify;
65 : use tokio::task::JoinSet;
66 : use tokio_util::sync::CancellationToken;
67 : use tracing::*;
68 : use upload_queue::NotInitialized;
69 : use utils::backoff;
70 : use utils::circuit_breaker::CircuitBreaker;
71 : use utils::completion;
72 : use utils::crashsafe::path_with_suffix_extension;
73 : use utils::failpoint_support;
74 : use utils::fs_ext;
75 : use utils::pausable_failpoint;
76 : use utils::sync::gate::Gate;
77 : use utils::sync::gate::GateGuard;
78 : use utils::timeout::timeout_cancellable;
79 : use utils::timeout::TimeoutCancellableError;
80 : use utils::try_rcu::ArcSwapExt;
81 : use utils::zstd::create_zst_tarball;
82 : use utils::zstd::extract_zst_tarball;
83 :
84 : use self::config::AttachedLocationConfig;
85 : use self::config::AttachmentMode;
86 : use self::config::LocationConf;
87 : use self::config::TenantConf;
88 : use self::metadata::TimelineMetadata;
89 : use self::mgr::GetActiveTenantError;
90 : use self::mgr::GetTenantError;
91 : use self::remote_timeline_client::upload::{upload_index_part, upload_tenant_manifest};
92 : use self::remote_timeline_client::{RemoteTimelineClient, WaitCompletionError};
93 : use self::timeline::uninit::TimelineCreateGuard;
94 : use self::timeline::uninit::TimelineExclusionError;
95 : use self::timeline::uninit::UninitializedTimeline;
96 : use self::timeline::EvictionTaskTenantState;
97 : use self::timeline::GcCutoffs;
98 : use self::timeline::TimelineDeleteProgress;
99 : use self::timeline::TimelineResources;
100 : use self::timeline::WaitLsnError;
101 : use crate::config::PageServerConf;
102 : use crate::context::{DownloadBehavior, RequestContext};
103 : use crate::deletion_queue::DeletionQueueClient;
104 : use crate::deletion_queue::DeletionQueueError;
105 : use crate::import_datadir;
106 : use crate::l0_flush::L0FlushGlobalState;
107 : use crate::metrics::CONCURRENT_INITDBS;
108 : use crate::metrics::INITDB_RUN_TIME;
109 : use crate::metrics::INITDB_SEMAPHORE_ACQUISITION_TIME;
110 : use crate::metrics::TENANT;
111 : use crate::metrics::{
112 : remove_tenant_metrics, BROKEN_TENANTS_SET, CIRCUIT_BREAKERS_BROKEN, CIRCUIT_BREAKERS_UNBROKEN,
113 : TENANT_STATE_METRIC, TENANT_SYNTHETIC_SIZE_METRIC,
114 : };
115 : use crate::task_mgr;
116 : use crate::task_mgr::TaskKind;
117 : use crate::tenant::config::LocationMode;
118 : use crate::tenant::config::TenantConfOpt;
119 : use crate::tenant::gc_result::GcResult;
120 : pub use crate::tenant::remote_timeline_client::index::IndexPart;
121 : use crate::tenant::remote_timeline_client::remote_initdb_archive_path;
122 : use crate::tenant::remote_timeline_client::MaybeDeletedIndexPart;
123 : use crate::tenant::remote_timeline_client::INITDB_PATH;
124 : use crate::tenant::storage_layer::DeltaLayer;
125 : use crate::tenant::storage_layer::ImageLayer;
126 : use crate::walingest::WalLagCooldown;
127 : use crate::walredo;
128 : use crate::InitializationOrder;
129 : use std::collections::hash_map::Entry;
130 : use std::collections::HashMap;
131 : use std::collections::HashSet;
132 : use std::fmt::Debug;
133 : use std::fmt::Display;
134 : use std::fs;
135 : use std::fs::File;
136 : use std::sync::atomic::{AtomicU64, Ordering};
137 : use std::sync::Arc;
138 : use std::sync::Mutex;
139 : use std::time::{Duration, Instant};
140 :
141 : use crate::span;
142 : use crate::tenant::timeline::delete::DeleteTimelineFlow;
143 : use crate::tenant::timeline::uninit::cleanup_timeline_directory;
144 : use crate::virtual_file::VirtualFile;
145 : use crate::walredo::PostgresRedoManager;
146 : use crate::TEMP_FILE_SUFFIX;
147 : use once_cell::sync::Lazy;
148 : pub use pageserver_api::models::TenantState;
149 : use tokio::sync::Semaphore;
150 :
151 0 : static INIT_DB_SEMAPHORE: Lazy<Semaphore> = Lazy::new(|| Semaphore::new(8));
152 : use utils::{
153 : crashsafe,
154 : generation::Generation,
155 : id::TimelineId,
156 : lsn::{Lsn, RecordLsn},
157 : };
158 :
159 : pub mod blob_io;
160 : pub mod block_io;
161 : pub mod vectored_blob_io;
162 :
163 : pub mod disk_btree;
164 : pub(crate) mod ephemeral_file;
165 : pub mod layer_map;
166 :
167 : pub mod metadata;
168 : pub mod remote_timeline_client;
169 : pub mod storage_layer;
170 :
171 : pub mod checks;
172 : pub mod config;
173 : pub mod mgr;
174 : pub mod secondary;
175 : pub mod tasks;
176 : pub mod upload_queue;
177 :
178 : pub(crate) mod timeline;
179 :
180 : pub mod size;
181 :
182 : mod gc_block;
183 : mod gc_result;
184 : pub(crate) mod throttle;
185 :
186 : pub(crate) use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
187 : pub(crate) use timeline::{LogicalSizeCalculationCause, PageReconstructError, Timeline};
188 :
189 : // re-export for use in walreceiver
190 : pub use crate::tenant::timeline::WalReceiverInfo;
191 :
192 : /// The "tenants" part of `tenants/<tenant>/timelines...`
193 : pub const TENANTS_SEGMENT_NAME: &str = "tenants";
194 :
195 : /// Parts of the `.neon/tenants/<tenant_id>/timelines/<timeline_id>` directory prefix.
196 : pub const TIMELINES_SEGMENT_NAME: &str = "timelines";
197 :
198 : /// References to shared objects that are passed into each tenant, such
199 : /// as the shared remote storage client and process initialization state.
200 : #[derive(Clone)]
201 : pub struct TenantSharedResources {
202 : pub broker_client: storage_broker::BrokerClientChannel,
203 : pub remote_storage: GenericRemoteStorage,
204 : pub deletion_queue_client: DeletionQueueClient,
205 : pub l0_flush_global_state: L0FlushGlobalState,
206 : }
207 :
208 : /// A [`Tenant`] is really an _attached_ tenant. The configuration
209 : /// for an attached tenant is a subset of the [`LocationConf`], represented
210 : /// in this struct.
211 : #[derive(Clone)]
212 : pub(super) struct AttachedTenantConf {
213 : tenant_conf: TenantConfOpt,
214 : location: AttachedLocationConfig,
215 : /// The deadline before which we are blocked from GC so that
216 : /// leases have a chance to be renewed.
217 : lsn_lease_deadline: Option<tokio::time::Instant>,
218 : }
219 :
220 : impl AttachedTenantConf {
221 444 : fn new(tenant_conf: TenantConfOpt, location: AttachedLocationConfig) -> Self {
222 : // Sets a deadline before which we cannot proceed to GC due to lsn lease.
223 : //
224 : // We do this as the leases mapping are not persisted to disk. By delaying GC by lease
225 : // length, we guarantee that all the leases we granted before will have a chance to renew
226 : // when we run GC for the first time after restart / transition from AttachedMulti to AttachedSingle.
227 444 : let lsn_lease_deadline = if location.attach_mode == AttachmentMode::Single {
228 444 : Some(
229 444 : tokio::time::Instant::now()
230 444 : + tenant_conf
231 444 : .lsn_lease_length
232 444 : .unwrap_or(LsnLease::DEFAULT_LENGTH),
233 444 : )
234 : } else {
235 : // We don't use `lsn_lease_deadline` to delay GC in AttachedMulti and AttachedStale
236 : // because we don't do GC in these modes.
237 0 : None
238 : };
239 :
240 444 : Self {
241 444 : tenant_conf,
242 444 : location,
243 444 : lsn_lease_deadline,
244 444 : }
245 444 : }
246 :
247 444 : fn try_from(location_conf: LocationConf) -> anyhow::Result<Self> {
248 444 : match &location_conf.mode {
249 444 : LocationMode::Attached(attach_conf) => {
250 444 : Ok(Self::new(location_conf.tenant_conf, *attach_conf))
251 : }
252 : LocationMode::Secondary(_) => {
253 0 : anyhow::bail!("Attempted to construct AttachedTenantConf from a LocationConf in secondary mode")
254 : }
255 : }
256 444 : }
257 :
258 1524 : fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
259 1524 : self.lsn_lease_deadline
260 1524 : .map(|d| tokio::time::Instant::now() < d)
261 1524 : .unwrap_or(false)
262 1524 : }
263 : }
264 : struct TimelinePreload {
265 : timeline_id: TimelineId,
266 : client: RemoteTimelineClient,
267 : index_part: Result<MaybeDeletedIndexPart, DownloadError>,
268 : previous_heatmap: Option<PreviousHeatmap>,
269 : }
270 :
271 : pub(crate) struct TenantPreload {
272 : tenant_manifest: TenantManifest,
273 : /// Map from timeline ID to a possible timeline preload. It is None iff the timeline is offloaded according to the manifest.
274 : timelines: HashMap<TimelineId, Option<TimelinePreload>>,
275 : }
276 :
277 : /// When we spawn a tenant, there is a special mode for tenant creation that
278 : /// avoids trying to read anything from remote storage.
279 : pub(crate) enum SpawnMode {
280 : /// Activate as soon as possible
281 : Eager,
282 : /// Lazy activation in the background, with the option to skip the queue if the need comes up
283 : Lazy,
284 : }
285 :
286 : ///
287 : /// Tenant consists of multiple timelines. Keep them in a hash table.
288 : ///
289 : pub struct Tenant {
290 : // Global pageserver config parameters
291 : pub conf: &'static PageServerConf,
292 :
293 : /// The value creation timestamp, used to measure activation delay, see:
294 : /// <https://github.com/neondatabase/neon/issues/4025>
295 : constructed_at: Instant,
296 :
297 : state: watch::Sender<TenantState>,
298 :
299 : // Overridden tenant-specific config parameters.
300 : // We keep TenantConfOpt sturct here to preserve the information
301 : // about parameters that are not set.
302 : // This is necessary to allow global config updates.
303 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
304 :
305 : tenant_shard_id: TenantShardId,
306 :
307 : // The detailed sharding information, beyond the number/count in tenant_shard_id
308 : shard_identity: ShardIdentity,
309 :
310 : /// The remote storage generation, used to protect S3 objects from split-brain.
311 : /// Does not change over the lifetime of the [`Tenant`] object.
312 : ///
313 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
314 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
315 : generation: Generation,
316 :
317 : timelines: Mutex<HashMap<TimelineId, Arc<Timeline>>>,
318 :
319 : /// During timeline creation, we first insert the TimelineId to the
320 : /// creating map, then `timelines`, then remove it from the creating map.
321 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
322 : timelines_creating: std::sync::Mutex<HashSet<TimelineId>>,
323 :
324 : /// Possibly offloaded and archived timelines
325 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
326 : timelines_offloaded: Mutex<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
327 :
328 : /// Serialize writes of the tenant manifest to remote storage. If there are concurrent operations
329 : /// affecting the manifest, such as timeline deletion and timeline offload, they must wait for
330 : /// each other (this could be optimized to coalesce writes if necessary).
331 : ///
332 : /// The contents of the Mutex are the last manifest we successfully uploaded
333 : tenant_manifest_upload: tokio::sync::Mutex<Option<TenantManifest>>,
334 :
335 : // This mutex prevents creation of new timelines during GC.
336 : // Adding yet another mutex (in addition to `timelines`) is needed because holding
337 : // `timelines` mutex during all GC iteration
338 : // may block for a long time `get_timeline`, `get_timelines_state`,... and other operations
339 : // with timelines, which in turn may cause dropping replication connection, expiration of wait_for_lsn
340 : // timeout...
341 : gc_cs: tokio::sync::Mutex<()>,
342 : walredo_mgr: Option<Arc<WalRedoManager>>,
343 :
344 : // provides access to timeline data sitting in the remote storage
345 : pub(crate) remote_storage: GenericRemoteStorage,
346 :
347 : // Access to global deletion queue for when this tenant wants to schedule a deletion
348 : deletion_queue_client: DeletionQueueClient,
349 :
350 : /// Cached logical sizes updated updated on each [`Tenant::gather_size_inputs`].
351 : cached_logical_sizes: tokio::sync::Mutex<HashMap<(TimelineId, Lsn), u64>>,
352 : cached_synthetic_tenant_size: Arc<AtomicU64>,
353 :
354 : eviction_task_tenant_state: tokio::sync::Mutex<EvictionTaskTenantState>,
355 :
356 : /// Track repeated failures to compact, so that we can back off.
357 : /// Overhead of mutex is acceptable because compaction is done with a multi-second period.
358 : compaction_circuit_breaker: std::sync::Mutex<CircuitBreaker>,
359 :
360 : /// Signals the tenant compaction loop that there is L0 compaction work to be done.
361 : pub(crate) l0_compaction_trigger: Arc<Notify>,
362 :
363 : /// Scheduled gc-compaction tasks.
364 : scheduled_compaction_tasks: std::sync::Mutex<HashMap<TimelineId, Arc<GcCompactionQueue>>>,
365 :
366 : /// If the tenant is in Activating state, notify this to encourage it
367 : /// to proceed to Active as soon as possible, rather than waiting for lazy
368 : /// background warmup.
369 : pub(crate) activate_now_sem: tokio::sync::Semaphore,
370 :
371 : /// Time it took for the tenant to activate. Zero if not active yet.
372 : attach_wal_lag_cooldown: Arc<std::sync::OnceLock<WalLagCooldown>>,
373 :
374 : // Cancellation token fires when we have entered shutdown(). This is a parent of
375 : // Timelines' cancellation token.
376 : pub(crate) cancel: CancellationToken,
377 :
378 : // Users of the Tenant such as the page service must take this Gate to avoid
379 : // trying to use a Tenant which is shutting down.
380 : pub(crate) gate: Gate,
381 :
382 : /// Throttle applied at the top of [`Timeline::get`].
383 : /// All [`Tenant::timelines`] of a given [`Tenant`] instance share the same [`throttle::Throttle`] instance.
384 : pub(crate) pagestream_throttle: Arc<throttle::Throttle>,
385 :
386 : pub(crate) pagestream_throttle_metrics: Arc<crate::metrics::tenant_throttling::Pagestream>,
387 :
388 : /// An ongoing timeline detach concurrency limiter.
389 : ///
390 : /// As a tenant will likely be restarted as part of timeline detach ancestor it makes no sense
391 : /// to have two running at the same time. A different one can be started if an earlier one
392 : /// has failed for whatever reason.
393 : ongoing_timeline_detach: std::sync::Mutex<Option<(TimelineId, utils::completion::Barrier)>>,
394 :
395 : /// `index_part.json` based gc blocking reason tracking.
396 : ///
397 : /// New gc iterations must start a new iteration by acquiring `GcBlock::start` before
398 : /// proceeding.
399 : pub(crate) gc_block: gc_block::GcBlock,
400 :
401 : l0_flush_global_state: L0FlushGlobalState,
402 : }
403 : impl std::fmt::Debug for Tenant {
404 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
405 0 : write!(f, "{} ({})", self.tenant_shard_id, self.current_state())
406 0 : }
407 : }
408 :
409 : pub(crate) enum WalRedoManager {
410 : Prod(WalredoManagerId, PostgresRedoManager),
411 : #[cfg(test)]
412 : Test(harness::TestRedoManager),
413 : }
414 :
415 : #[derive(thiserror::Error, Debug)]
416 : #[error("pageserver is shutting down")]
417 : pub(crate) struct GlobalShutDown;
418 :
419 : impl WalRedoManager {
420 0 : pub(crate) fn new(mgr: PostgresRedoManager) -> Result<Arc<Self>, GlobalShutDown> {
421 0 : let id = WalredoManagerId::next();
422 0 : let arc = Arc::new(Self::Prod(id, mgr));
423 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
424 0 : match &mut *guard {
425 0 : Some(map) => {
426 0 : map.insert(id, Arc::downgrade(&arc));
427 0 : Ok(arc)
428 : }
429 0 : None => Err(GlobalShutDown),
430 : }
431 0 : }
432 : }
433 :
434 : impl Drop for WalRedoManager {
435 20 : fn drop(&mut self) {
436 20 : match self {
437 0 : Self::Prod(id, _) => {
438 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
439 0 : if let Some(map) = &mut *guard {
440 0 : map.remove(id).expect("new() registers, drop() unregisters");
441 0 : }
442 : }
443 : #[cfg(test)]
444 20 : Self::Test(_) => {
445 20 : // Not applicable to test redo manager
446 20 : }
447 : }
448 20 : }
449 : }
450 :
451 : /// Global registry of all walredo managers so that [`crate::shutdown_pageserver`] can shut down
452 : /// the walredo processes outside of the regular order.
453 : ///
454 : /// This is necessary to work around a systemd bug where it freezes if there are
455 : /// walredo processes left => <https://github.com/neondatabase/cloud/issues/11387>
456 : #[allow(clippy::type_complexity)]
457 : pub(crate) static WALREDO_MANAGERS: once_cell::sync::Lazy<
458 : Mutex<Option<HashMap<WalredoManagerId, Weak<WalRedoManager>>>>,
459 0 : > = once_cell::sync::Lazy::new(|| Mutex::new(Some(HashMap::new())));
460 : #[derive(PartialEq, Eq, Hash, Clone, Copy, Debug)]
461 : pub(crate) struct WalredoManagerId(u64);
462 : impl WalredoManagerId {
463 0 : pub fn next() -> Self {
464 : static NEXT: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(1);
465 0 : let id = NEXT.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
466 0 : if id == 0 {
467 0 : panic!("WalredoManagerId::new() returned 0, indicating wraparound, risking it's no longer unique");
468 0 : }
469 0 : Self(id)
470 0 : }
471 : }
472 :
473 : #[cfg(test)]
474 : impl From<harness::TestRedoManager> for WalRedoManager {
475 444 : fn from(mgr: harness::TestRedoManager) -> Self {
476 444 : Self::Test(mgr)
477 444 : }
478 : }
479 :
480 : impl WalRedoManager {
481 12 : pub(crate) async fn shutdown(&self) -> bool {
482 12 : match self {
483 0 : Self::Prod(_, mgr) => mgr.shutdown().await,
484 : #[cfg(test)]
485 : Self::Test(_) => {
486 : // Not applicable to test redo manager
487 12 : true
488 : }
489 : }
490 12 : }
491 :
492 0 : pub(crate) fn maybe_quiesce(&self, idle_timeout: Duration) {
493 0 : match self {
494 0 : Self::Prod(_, mgr) => mgr.maybe_quiesce(idle_timeout),
495 0 : #[cfg(test)]
496 0 : Self::Test(_) => {
497 0 : // Not applicable to test redo manager
498 0 : }
499 0 : }
500 0 : }
501 :
502 : /// # Cancel-Safety
503 : ///
504 : /// This method is cancellation-safe.
505 1636 : pub async fn request_redo(
506 1636 : &self,
507 1636 : key: pageserver_api::key::Key,
508 1636 : lsn: Lsn,
509 1636 : base_img: Option<(Lsn, bytes::Bytes)>,
510 1636 : records: Vec<(Lsn, pageserver_api::record::NeonWalRecord)>,
511 1636 : pg_version: u32,
512 1636 : ) -> Result<bytes::Bytes, walredo::Error> {
513 1636 : match self {
514 0 : Self::Prod(_, mgr) => {
515 0 : mgr.request_redo(key, lsn, base_img, records, pg_version)
516 0 : .await
517 : }
518 : #[cfg(test)]
519 1636 : Self::Test(mgr) => {
520 1636 : mgr.request_redo(key, lsn, base_img, records, pg_version)
521 1636 : .await
522 : }
523 : }
524 1636 : }
525 :
526 0 : pub(crate) fn status(&self) -> Option<WalRedoManagerStatus> {
527 0 : match self {
528 0 : WalRedoManager::Prod(_, m) => Some(m.status()),
529 0 : #[cfg(test)]
530 0 : WalRedoManager::Test(_) => None,
531 0 : }
532 0 : }
533 : }
534 :
535 : /// A very lightweight memory representation of an offloaded timeline.
536 : ///
537 : /// We need to store the list of offloaded timelines so that we can perform operations on them,
538 : /// like unoffloading them, or (at a later date), decide to perform flattening.
539 : /// This type has a much smaller memory impact than [`Timeline`], and thus we can store many
540 : /// more offloaded timelines than we can manage ones that aren't.
541 : pub struct OffloadedTimeline {
542 : pub tenant_shard_id: TenantShardId,
543 : pub timeline_id: TimelineId,
544 : pub ancestor_timeline_id: Option<TimelineId>,
545 : /// Whether to retain the branch lsn at the ancestor or not
546 : pub ancestor_retain_lsn: Option<Lsn>,
547 :
548 : /// When the timeline was archived.
549 : ///
550 : /// Present for future flattening deliberations.
551 : pub archived_at: NaiveDateTime,
552 :
553 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
554 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
555 : pub delete_progress: TimelineDeleteProgress,
556 :
557 : /// Part of the `OffloadedTimeline` object's lifecycle: this needs to be set before we drop it
558 : pub deleted_from_ancestor: AtomicBool,
559 : }
560 :
561 : impl OffloadedTimeline {
562 : /// Obtains an offloaded timeline from a given timeline object.
563 : ///
564 : /// Returns `None` if the `archived_at` flag couldn't be obtained, i.e.
565 : /// the timeline is not in a stopped state.
566 : /// Panics if the timeline is not archived.
567 4 : fn from_timeline(timeline: &Timeline) -> Result<Self, UploadQueueNotReadyError> {
568 4 : let (ancestor_retain_lsn, ancestor_timeline_id) =
569 4 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
570 4 : let ancestor_lsn = timeline.get_ancestor_lsn();
571 4 : let ancestor_timeline_id = ancestor_timeline.timeline_id;
572 4 : let mut gc_info = ancestor_timeline.gc_info.write().unwrap();
573 4 : gc_info.insert_child(timeline.timeline_id, ancestor_lsn, MaybeOffloaded::Yes);
574 4 : (Some(ancestor_lsn), Some(ancestor_timeline_id))
575 : } else {
576 0 : (None, None)
577 : };
578 4 : let archived_at = timeline
579 4 : .remote_client
580 4 : .archived_at_stopped_queue()?
581 4 : .expect("must be called on an archived timeline");
582 4 : Ok(Self {
583 4 : tenant_shard_id: timeline.tenant_shard_id,
584 4 : timeline_id: timeline.timeline_id,
585 4 : ancestor_timeline_id,
586 4 : ancestor_retain_lsn,
587 4 : archived_at,
588 4 :
589 4 : delete_progress: timeline.delete_progress.clone(),
590 4 : deleted_from_ancestor: AtomicBool::new(false),
591 4 : })
592 4 : }
593 0 : fn from_manifest(tenant_shard_id: TenantShardId, manifest: &OffloadedTimelineManifest) -> Self {
594 0 : // We expect to reach this case in tenant loading, where the `retain_lsn` is populated in the parent's `gc_info`
595 0 : // by the `initialize_gc_info` function.
596 0 : let OffloadedTimelineManifest {
597 0 : timeline_id,
598 0 : ancestor_timeline_id,
599 0 : ancestor_retain_lsn,
600 0 : archived_at,
601 0 : } = *manifest;
602 0 : Self {
603 0 : tenant_shard_id,
604 0 : timeline_id,
605 0 : ancestor_timeline_id,
606 0 : ancestor_retain_lsn,
607 0 : archived_at,
608 0 : delete_progress: TimelineDeleteProgress::default(),
609 0 : deleted_from_ancestor: AtomicBool::new(false),
610 0 : }
611 0 : }
612 4 : fn manifest(&self) -> OffloadedTimelineManifest {
613 4 : let Self {
614 4 : timeline_id,
615 4 : ancestor_timeline_id,
616 4 : ancestor_retain_lsn,
617 4 : archived_at,
618 4 : ..
619 4 : } = self;
620 4 : OffloadedTimelineManifest {
621 4 : timeline_id: *timeline_id,
622 4 : ancestor_timeline_id: *ancestor_timeline_id,
623 4 : ancestor_retain_lsn: *ancestor_retain_lsn,
624 4 : archived_at: *archived_at,
625 4 : }
626 4 : }
627 : /// Delete this timeline's retain_lsn from its ancestor, if present in the given tenant
628 0 : fn delete_from_ancestor_with_timelines(
629 0 : &self,
630 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
631 0 : ) {
632 0 : if let (Some(_retain_lsn), Some(ancestor_timeline_id)) =
633 0 : (self.ancestor_retain_lsn, self.ancestor_timeline_id)
634 : {
635 0 : if let Some((_, ancestor_timeline)) = timelines
636 0 : .iter()
637 0 : .find(|(tid, _tl)| **tid == ancestor_timeline_id)
638 : {
639 0 : let removal_happened = ancestor_timeline
640 0 : .gc_info
641 0 : .write()
642 0 : .unwrap()
643 0 : .remove_child_offloaded(self.timeline_id);
644 0 : if !removal_happened {
645 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
646 0 : "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
647 0 : }
648 0 : }
649 0 : }
650 0 : self.deleted_from_ancestor.store(true, Ordering::Release);
651 0 : }
652 : /// Call [`Self::delete_from_ancestor_with_timelines`] instead if possible.
653 : ///
654 : /// As the entire tenant is being dropped, don't bother deregistering the `retain_lsn` from the ancestor.
655 4 : fn defuse_for_tenant_drop(&self) {
656 4 : self.deleted_from_ancestor.store(true, Ordering::Release);
657 4 : }
658 : }
659 :
660 : impl fmt::Debug for OffloadedTimeline {
661 0 : fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
662 0 : write!(f, "OffloadedTimeline<{}>", self.timeline_id)
663 0 : }
664 : }
665 :
666 : impl Drop for OffloadedTimeline {
667 4 : fn drop(&mut self) {
668 4 : if !self.deleted_from_ancestor.load(Ordering::Acquire) {
669 0 : tracing::warn!(
670 0 : "offloaded timeline {} was dropped without having cleaned it up at the ancestor",
671 : self.timeline_id
672 : );
673 4 : }
674 4 : }
675 : }
676 :
677 : #[derive(Copy, Clone, PartialEq, Eq, Hash, Debug)]
678 : pub enum MaybeOffloaded {
679 : Yes,
680 : No,
681 : }
682 :
683 : #[derive(Clone, Debug)]
684 : pub enum TimelineOrOffloaded {
685 : Timeline(Arc<Timeline>),
686 : Offloaded(Arc<OffloadedTimeline>),
687 : }
688 :
689 : impl TimelineOrOffloaded {
690 0 : pub fn arc_ref(&self) -> TimelineOrOffloadedArcRef<'_> {
691 0 : match self {
692 0 : TimelineOrOffloaded::Timeline(timeline) => {
693 0 : TimelineOrOffloadedArcRef::Timeline(timeline)
694 : }
695 0 : TimelineOrOffloaded::Offloaded(offloaded) => {
696 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded)
697 : }
698 : }
699 0 : }
700 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
701 0 : self.arc_ref().tenant_shard_id()
702 0 : }
703 0 : pub fn timeline_id(&self) -> TimelineId {
704 0 : self.arc_ref().timeline_id()
705 0 : }
706 4 : pub fn delete_progress(&self) -> &Arc<tokio::sync::Mutex<DeleteTimelineFlow>> {
707 4 : match self {
708 4 : TimelineOrOffloaded::Timeline(timeline) => &timeline.delete_progress,
709 0 : TimelineOrOffloaded::Offloaded(offloaded) => &offloaded.delete_progress,
710 : }
711 4 : }
712 0 : fn maybe_remote_client(&self) -> Option<Arc<RemoteTimelineClient>> {
713 0 : match self {
714 0 : TimelineOrOffloaded::Timeline(timeline) => Some(timeline.remote_client.clone()),
715 0 : TimelineOrOffloaded::Offloaded(_offloaded) => None,
716 : }
717 0 : }
718 : }
719 :
720 : pub enum TimelineOrOffloadedArcRef<'a> {
721 : Timeline(&'a Arc<Timeline>),
722 : Offloaded(&'a Arc<OffloadedTimeline>),
723 : }
724 :
725 : impl TimelineOrOffloadedArcRef<'_> {
726 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
727 0 : match self {
728 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.tenant_shard_id,
729 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.tenant_shard_id,
730 : }
731 0 : }
732 0 : pub fn timeline_id(&self) -> TimelineId {
733 0 : match self {
734 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.timeline_id,
735 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.timeline_id,
736 : }
737 0 : }
738 : }
739 :
740 : impl<'a> From<&'a Arc<Timeline>> for TimelineOrOffloadedArcRef<'a> {
741 0 : fn from(timeline: &'a Arc<Timeline>) -> Self {
742 0 : Self::Timeline(timeline)
743 0 : }
744 : }
745 :
746 : impl<'a> From<&'a Arc<OffloadedTimeline>> for TimelineOrOffloadedArcRef<'a> {
747 0 : fn from(timeline: &'a Arc<OffloadedTimeline>) -> Self {
748 0 : Self::Offloaded(timeline)
749 0 : }
750 : }
751 :
752 : #[derive(Debug, thiserror::Error, PartialEq, Eq)]
753 : pub enum GetTimelineError {
754 : #[error("Timeline is shutting down")]
755 : ShuttingDown,
756 : #[error("Timeline {tenant_id}/{timeline_id} is not active, state: {state:?}")]
757 : NotActive {
758 : tenant_id: TenantShardId,
759 : timeline_id: TimelineId,
760 : state: TimelineState,
761 : },
762 : #[error("Timeline {tenant_id}/{timeline_id} was not found")]
763 : NotFound {
764 : tenant_id: TenantShardId,
765 : timeline_id: TimelineId,
766 : },
767 : }
768 :
769 : #[derive(Debug, thiserror::Error)]
770 : pub enum LoadLocalTimelineError {
771 : #[error("FailedToLoad")]
772 : Load(#[source] anyhow::Error),
773 : #[error("FailedToResumeDeletion")]
774 : ResumeDeletion(#[source] anyhow::Error),
775 : }
776 :
777 : #[derive(thiserror::Error)]
778 : pub enum DeleteTimelineError {
779 : #[error("NotFound")]
780 : NotFound,
781 :
782 : #[error("HasChildren")]
783 : HasChildren(Vec<TimelineId>),
784 :
785 : #[error("Timeline deletion is already in progress")]
786 : AlreadyInProgress(Arc<tokio::sync::Mutex<DeleteTimelineFlow>>),
787 :
788 : #[error("Cancelled")]
789 : Cancelled,
790 :
791 : #[error(transparent)]
792 : Other(#[from] anyhow::Error),
793 : }
794 :
795 : impl Debug for DeleteTimelineError {
796 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
797 0 : match self {
798 0 : Self::NotFound => write!(f, "NotFound"),
799 0 : Self::HasChildren(c) => f.debug_tuple("HasChildren").field(c).finish(),
800 0 : Self::AlreadyInProgress(_) => f.debug_tuple("AlreadyInProgress").finish(),
801 0 : Self::Cancelled => f.debug_tuple("Cancelled").finish(),
802 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
803 : }
804 0 : }
805 : }
806 :
807 : #[derive(thiserror::Error)]
808 : pub enum TimelineArchivalError {
809 : #[error("NotFound")]
810 : NotFound,
811 :
812 : #[error("Timeout")]
813 : Timeout,
814 :
815 : #[error("Cancelled")]
816 : Cancelled,
817 :
818 : #[error("ancestor is archived: {}", .0)]
819 : HasArchivedParent(TimelineId),
820 :
821 : #[error("HasUnarchivedChildren")]
822 : HasUnarchivedChildren(Vec<TimelineId>),
823 :
824 : #[error("Timeline archival is already in progress")]
825 : AlreadyInProgress,
826 :
827 : #[error(transparent)]
828 : Other(anyhow::Error),
829 : }
830 :
831 : #[derive(thiserror::Error, Debug)]
832 : pub(crate) enum TenantManifestError {
833 : #[error("Remote storage error: {0}")]
834 : RemoteStorage(anyhow::Error),
835 :
836 : #[error("Cancelled")]
837 : Cancelled,
838 : }
839 :
840 : impl From<TenantManifestError> for TimelineArchivalError {
841 0 : fn from(e: TenantManifestError) -> Self {
842 0 : match e {
843 0 : TenantManifestError::RemoteStorage(e) => Self::Other(e),
844 0 : TenantManifestError::Cancelled => Self::Cancelled,
845 : }
846 0 : }
847 : }
848 :
849 : impl Debug for TimelineArchivalError {
850 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
851 0 : match self {
852 0 : Self::NotFound => write!(f, "NotFound"),
853 0 : Self::Timeout => write!(f, "Timeout"),
854 0 : Self::Cancelled => write!(f, "Cancelled"),
855 0 : Self::HasArchivedParent(p) => f.debug_tuple("HasArchivedParent").field(p).finish(),
856 0 : Self::HasUnarchivedChildren(c) => {
857 0 : f.debug_tuple("HasUnarchivedChildren").field(c).finish()
858 : }
859 0 : Self::AlreadyInProgress => f.debug_tuple("AlreadyInProgress").finish(),
860 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
861 : }
862 0 : }
863 : }
864 :
865 : pub enum SetStoppingError {
866 : AlreadyStopping(completion::Barrier),
867 : Broken,
868 : }
869 :
870 : impl Debug for SetStoppingError {
871 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
872 0 : match self {
873 0 : Self::AlreadyStopping(_) => f.debug_tuple("AlreadyStopping").finish(),
874 0 : Self::Broken => write!(f, "Broken"),
875 : }
876 0 : }
877 : }
878 :
879 : /// Arguments to [`Tenant::create_timeline`].
880 : ///
881 : /// Not usable as an idempotency key for timeline creation because if [`CreateTimelineParamsBranch::ancestor_start_lsn`]
882 : /// is `None`, the result of the timeline create call is not deterministic.
883 : ///
884 : /// See [`CreateTimelineIdempotency`] for an idempotency key.
885 : #[derive(Debug)]
886 : pub(crate) enum CreateTimelineParams {
887 : Bootstrap(CreateTimelineParamsBootstrap),
888 : Branch(CreateTimelineParamsBranch),
889 : ImportPgdata(CreateTimelineParamsImportPgdata),
890 : }
891 :
892 : #[derive(Debug)]
893 : pub(crate) struct CreateTimelineParamsBootstrap {
894 : pub(crate) new_timeline_id: TimelineId,
895 : pub(crate) existing_initdb_timeline_id: Option<TimelineId>,
896 : pub(crate) pg_version: u32,
897 : }
898 :
899 : /// NB: See comment on [`CreateTimelineIdempotency::Branch`] for why there's no `pg_version` here.
900 : #[derive(Debug)]
901 : pub(crate) struct CreateTimelineParamsBranch {
902 : pub(crate) new_timeline_id: TimelineId,
903 : pub(crate) ancestor_timeline_id: TimelineId,
904 : pub(crate) ancestor_start_lsn: Option<Lsn>,
905 : }
906 :
907 : #[derive(Debug)]
908 : pub(crate) struct CreateTimelineParamsImportPgdata {
909 : pub(crate) new_timeline_id: TimelineId,
910 : pub(crate) location: import_pgdata::index_part_format::Location,
911 : pub(crate) idempotency_key: import_pgdata::index_part_format::IdempotencyKey,
912 : }
913 :
914 : /// What is used to determine idempotency of a [`Tenant::create_timeline`] call in [`Tenant::start_creating_timeline`] in [`Tenant::start_creating_timeline`].
915 : ///
916 : /// Each [`Timeline`] object holds [`Self`] as an immutable property in [`Timeline::create_idempotency`].
917 : ///
918 : /// We lower timeline creation requests to [`Self`], and then use [`PartialEq::eq`] to compare [`Timeline::create_idempotency`] with the request.
919 : /// If they are equal, we return a reference to the existing timeline, otherwise it's an idempotency conflict.
920 : ///
921 : /// There is special treatment for [`Self::FailWithConflict`] to always return an idempotency conflict.
922 : /// It would be nice to have more advanced derive macros to make that special treatment declarative.
923 : ///
924 : /// Notes:
925 : /// - Unlike [`CreateTimelineParams`], ancestor LSN is fixed, so, branching will be at a deterministic LSN.
926 : /// - We make some trade-offs though, e.g., [`CreateTimelineParamsBootstrap::existing_initdb_timeline_id`]
927 : /// is not considered for idempotency. We can improve on this over time if we deem it necessary.
928 : ///
929 : #[derive(Debug, Clone, PartialEq, Eq)]
930 : pub(crate) enum CreateTimelineIdempotency {
931 : /// NB: special treatment, see comment in [`Self`].
932 : FailWithConflict,
933 : Bootstrap {
934 : pg_version: u32,
935 : },
936 : /// NB: branches always have the same `pg_version` as their ancestor.
937 : /// While [`pageserver_api::models::TimelineCreateRequestMode::Branch::pg_version`]
938 : /// exists as a field, and is set by cplane, it has always been ignored by pageserver when
939 : /// determining the child branch pg_version.
940 : Branch {
941 : ancestor_timeline_id: TimelineId,
942 : ancestor_start_lsn: Lsn,
943 : },
944 : ImportPgdata(CreatingTimelineIdempotencyImportPgdata),
945 : }
946 :
947 : #[derive(Debug, Clone, PartialEq, Eq)]
948 : pub(crate) struct CreatingTimelineIdempotencyImportPgdata {
949 : idempotency_key: import_pgdata::index_part_format::IdempotencyKey,
950 : }
951 :
952 : /// What is returned by [`Tenant::start_creating_timeline`].
953 : #[must_use]
954 : enum StartCreatingTimelineResult {
955 : CreateGuard(TimelineCreateGuard),
956 : Idempotent(Arc<Timeline>),
957 : }
958 :
959 : enum TimelineInitAndSyncResult {
960 : ReadyToActivate(Arc<Timeline>),
961 : NeedsSpawnImportPgdata(TimelineInitAndSyncNeedsSpawnImportPgdata),
962 : }
963 :
964 : impl TimelineInitAndSyncResult {
965 0 : fn ready_to_activate(self) -> Option<Arc<Timeline>> {
966 0 : match self {
967 0 : Self::ReadyToActivate(timeline) => Some(timeline),
968 0 : _ => None,
969 : }
970 0 : }
971 : }
972 :
973 : #[must_use]
974 : struct TimelineInitAndSyncNeedsSpawnImportPgdata {
975 : timeline: Arc<Timeline>,
976 : import_pgdata: import_pgdata::index_part_format::Root,
977 : guard: TimelineCreateGuard,
978 : }
979 :
980 : /// What is returned by [`Tenant::create_timeline`].
981 : enum CreateTimelineResult {
982 : Created(Arc<Timeline>),
983 : Idempotent(Arc<Timeline>),
984 : /// IMPORTANT: This [`Arc<Timeline>`] object is not in [`Tenant::timelines`] when
985 : /// we return this result, nor will this concrete object ever be added there.
986 : /// Cf method comment on [`Tenant::create_timeline_import_pgdata`].
987 : ImportSpawned(Arc<Timeline>),
988 : }
989 :
990 : impl CreateTimelineResult {
991 0 : fn discriminant(&self) -> &'static str {
992 0 : match self {
993 0 : Self::Created(_) => "Created",
994 0 : Self::Idempotent(_) => "Idempotent",
995 0 : Self::ImportSpawned(_) => "ImportSpawned",
996 : }
997 0 : }
998 0 : fn timeline(&self) -> &Arc<Timeline> {
999 0 : match self {
1000 0 : Self::Created(t) | Self::Idempotent(t) | Self::ImportSpawned(t) => t,
1001 0 : }
1002 0 : }
1003 : /// Unit test timelines aren't activated, test has to do it if it needs to.
1004 : #[cfg(test)]
1005 460 : fn into_timeline_for_test(self) -> Arc<Timeline> {
1006 460 : match self {
1007 460 : Self::Created(t) | Self::Idempotent(t) | Self::ImportSpawned(t) => t,
1008 460 : }
1009 460 : }
1010 : }
1011 :
1012 : #[derive(thiserror::Error, Debug)]
1013 : pub enum CreateTimelineError {
1014 : #[error("creation of timeline with the given ID is in progress")]
1015 : AlreadyCreating,
1016 : #[error("timeline already exists with different parameters")]
1017 : Conflict,
1018 : #[error(transparent)]
1019 : AncestorLsn(anyhow::Error),
1020 : #[error("ancestor timeline is not active")]
1021 : AncestorNotActive,
1022 : #[error("ancestor timeline is archived")]
1023 : AncestorArchived,
1024 : #[error("tenant shutting down")]
1025 : ShuttingDown,
1026 : #[error(transparent)]
1027 : Other(#[from] anyhow::Error),
1028 : }
1029 :
1030 : #[derive(thiserror::Error, Debug)]
1031 : pub enum InitdbError {
1032 : #[error("Operation was cancelled")]
1033 : Cancelled,
1034 : #[error(transparent)]
1035 : Other(anyhow::Error),
1036 : #[error(transparent)]
1037 : Inner(postgres_initdb::Error),
1038 : }
1039 :
1040 : enum CreateTimelineCause {
1041 : Load,
1042 : Delete,
1043 : }
1044 :
1045 : enum LoadTimelineCause {
1046 : Attach,
1047 : Unoffload,
1048 : ImportPgdata {
1049 : create_guard: TimelineCreateGuard,
1050 : activate: ActivateTimelineArgs,
1051 : },
1052 : }
1053 :
1054 : #[derive(thiserror::Error, Debug)]
1055 : pub(crate) enum GcError {
1056 : // The tenant is shutting down
1057 : #[error("tenant shutting down")]
1058 : TenantCancelled,
1059 :
1060 : // The tenant is shutting down
1061 : #[error("timeline shutting down")]
1062 : TimelineCancelled,
1063 :
1064 : // The tenant is in a state inelegible to run GC
1065 : #[error("not active")]
1066 : NotActive,
1067 :
1068 : // A requested GC cutoff LSN was invalid, for example it tried to move backwards
1069 : #[error("not active")]
1070 : BadLsn { why: String },
1071 :
1072 : // A remote storage error while scheduling updates after compaction
1073 : #[error(transparent)]
1074 : Remote(anyhow::Error),
1075 :
1076 : // An error reading while calculating GC cutoffs
1077 : #[error(transparent)]
1078 : GcCutoffs(PageReconstructError),
1079 :
1080 : // If GC was invoked for a particular timeline, this error means it didn't exist
1081 : #[error("timeline not found")]
1082 : TimelineNotFound,
1083 : }
1084 :
1085 : impl From<PageReconstructError> for GcError {
1086 0 : fn from(value: PageReconstructError) -> Self {
1087 0 : match value {
1088 0 : PageReconstructError::Cancelled => Self::TimelineCancelled,
1089 0 : other => Self::GcCutoffs(other),
1090 : }
1091 0 : }
1092 : }
1093 :
1094 : impl From<NotInitialized> for GcError {
1095 0 : fn from(value: NotInitialized) -> Self {
1096 0 : match value {
1097 0 : NotInitialized::Uninitialized => GcError::Remote(value.into()),
1098 0 : NotInitialized::Stopped | NotInitialized::ShuttingDown => GcError::TimelineCancelled,
1099 : }
1100 0 : }
1101 : }
1102 :
1103 : impl From<timeline::layer_manager::Shutdown> for GcError {
1104 0 : fn from(_: timeline::layer_manager::Shutdown) -> Self {
1105 0 : GcError::TimelineCancelled
1106 0 : }
1107 : }
1108 :
1109 : #[derive(thiserror::Error, Debug)]
1110 : pub(crate) enum LoadConfigError {
1111 : #[error("TOML deserialization error: '{0}'")]
1112 : DeserializeToml(#[from] toml_edit::de::Error),
1113 :
1114 : #[error("Config not found at {0}")]
1115 : NotFound(Utf8PathBuf),
1116 : }
1117 :
1118 : impl Tenant {
1119 : /// Yet another helper for timeline initialization.
1120 : ///
1121 : /// - Initializes the Timeline struct and inserts it into the tenant's hash map
1122 : /// - Scans the local timeline directory for layer files and builds the layer map
1123 : /// - Downloads remote index file and adds remote files to the layer map
1124 : /// - Schedules remote upload tasks for any files that are present locally but missing from remote storage.
1125 : ///
1126 : /// If the operation fails, the timeline is left in the tenant's hash map in Broken state. On success,
1127 : /// it is marked as Active.
1128 : #[allow(clippy::too_many_arguments)]
1129 12 : async fn timeline_init_and_sync(
1130 12 : self: &Arc<Self>,
1131 12 : timeline_id: TimelineId,
1132 12 : resources: TimelineResources,
1133 12 : mut index_part: IndexPart,
1134 12 : metadata: TimelineMetadata,
1135 12 : previous_heatmap: Option<PreviousHeatmap>,
1136 12 : ancestor: Option<Arc<Timeline>>,
1137 12 : cause: LoadTimelineCause,
1138 12 : ctx: &RequestContext,
1139 12 : ) -> anyhow::Result<TimelineInitAndSyncResult> {
1140 12 : let tenant_id = self.tenant_shard_id;
1141 12 :
1142 12 : let import_pgdata = index_part.import_pgdata.take();
1143 12 : let idempotency = match &import_pgdata {
1144 0 : Some(import_pgdata) => {
1145 0 : CreateTimelineIdempotency::ImportPgdata(CreatingTimelineIdempotencyImportPgdata {
1146 0 : idempotency_key: import_pgdata.idempotency_key().clone(),
1147 0 : })
1148 : }
1149 : None => {
1150 12 : if metadata.ancestor_timeline().is_none() {
1151 8 : CreateTimelineIdempotency::Bootstrap {
1152 8 : pg_version: metadata.pg_version(),
1153 8 : }
1154 : } else {
1155 4 : CreateTimelineIdempotency::Branch {
1156 4 : ancestor_timeline_id: metadata.ancestor_timeline().unwrap(),
1157 4 : ancestor_start_lsn: metadata.ancestor_lsn(),
1158 4 : }
1159 : }
1160 : }
1161 : };
1162 :
1163 12 : let timeline = self.create_timeline_struct(
1164 12 : timeline_id,
1165 12 : &metadata,
1166 12 : previous_heatmap,
1167 12 : ancestor.clone(),
1168 12 : resources,
1169 12 : CreateTimelineCause::Load,
1170 12 : idempotency.clone(),
1171 12 : )?;
1172 12 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
1173 12 : anyhow::ensure!(
1174 12 : disk_consistent_lsn.is_valid(),
1175 0 : "Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn"
1176 : );
1177 12 : assert_eq!(
1178 12 : disk_consistent_lsn,
1179 12 : metadata.disk_consistent_lsn(),
1180 0 : "these are used interchangeably"
1181 : );
1182 :
1183 12 : timeline.remote_client.init_upload_queue(&index_part)?;
1184 :
1185 12 : timeline
1186 12 : .load_layer_map(disk_consistent_lsn, index_part)
1187 12 : .await
1188 12 : .with_context(|| {
1189 0 : format!("Failed to load layermap for timeline {tenant_id}/{timeline_id}")
1190 12 : })?;
1191 :
1192 0 : match import_pgdata {
1193 0 : Some(import_pgdata) if !import_pgdata.is_done() => {
1194 0 : match cause {
1195 0 : LoadTimelineCause::Attach | LoadTimelineCause::Unoffload => (),
1196 : LoadTimelineCause::ImportPgdata { .. } => {
1197 0 : unreachable!("ImportPgdata should not be reloading timeline import is done and persisted as such in s3")
1198 : }
1199 : }
1200 0 : let mut guard = self.timelines_creating.lock().unwrap();
1201 0 : if !guard.insert(timeline_id) {
1202 : // We should never try and load the same timeline twice during startup
1203 0 : unreachable!("Timeline {tenant_id}/{timeline_id} is already being created")
1204 0 : }
1205 0 : let timeline_create_guard = TimelineCreateGuard {
1206 0 : _tenant_gate_guard: self.gate.enter()?,
1207 0 : owning_tenant: self.clone(),
1208 0 : timeline_id,
1209 0 : idempotency,
1210 0 : // The users of this specific return value don't need the timline_path in there.
1211 0 : timeline_path: timeline
1212 0 : .conf
1213 0 : .timeline_path(&timeline.tenant_shard_id, &timeline.timeline_id),
1214 0 : };
1215 0 : Ok(TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
1216 0 : TimelineInitAndSyncNeedsSpawnImportPgdata {
1217 0 : timeline,
1218 0 : import_pgdata,
1219 0 : guard: timeline_create_guard,
1220 0 : },
1221 0 : ))
1222 : }
1223 : Some(_) | None => {
1224 : {
1225 12 : let mut timelines_accessor = self.timelines.lock().unwrap();
1226 12 : match timelines_accessor.entry(timeline_id) {
1227 : // We should never try and load the same timeline twice during startup
1228 : Entry::Occupied(_) => {
1229 0 : unreachable!(
1230 0 : "Timeline {tenant_id}/{timeline_id} already exists in the tenant map"
1231 0 : );
1232 : }
1233 12 : Entry::Vacant(v) => {
1234 12 : v.insert(Arc::clone(&timeline));
1235 12 : timeline.maybe_spawn_flush_loop();
1236 12 : }
1237 : }
1238 : }
1239 :
1240 : // Sanity check: a timeline should have some content.
1241 12 : anyhow::ensure!(
1242 12 : ancestor.is_some()
1243 8 : || timeline
1244 8 : .layers
1245 8 : .read()
1246 8 : .await
1247 8 : .layer_map()
1248 8 : .expect("currently loading, layer manager cannot be shutdown already")
1249 8 : .iter_historic_layers()
1250 8 : .next()
1251 8 : .is_some(),
1252 0 : "Timeline has no ancestor and no layer files"
1253 : );
1254 :
1255 12 : match cause {
1256 12 : LoadTimelineCause::Attach | LoadTimelineCause::Unoffload => (),
1257 : LoadTimelineCause::ImportPgdata {
1258 0 : create_guard,
1259 0 : activate,
1260 0 : } => {
1261 0 : // TODO: see the comment in the task code above how I'm not so certain
1262 0 : // it is safe to activate here because of concurrent shutdowns.
1263 0 : match activate {
1264 0 : ActivateTimelineArgs::Yes { broker_client } => {
1265 0 : info!("activating timeline after reload from pgdata import task");
1266 0 : timeline.activate(self.clone(), broker_client, None, ctx);
1267 : }
1268 0 : ActivateTimelineArgs::No => (),
1269 : }
1270 0 : drop(create_guard);
1271 : }
1272 : }
1273 :
1274 12 : Ok(TimelineInitAndSyncResult::ReadyToActivate(timeline))
1275 : }
1276 : }
1277 12 : }
1278 :
1279 : /// Attach a tenant that's available in cloud storage.
1280 : ///
1281 : /// This returns quickly, after just creating the in-memory object
1282 : /// Tenant struct and launching a background task to download
1283 : /// the remote index files. On return, the tenant is most likely still in
1284 : /// Attaching state, and it will become Active once the background task
1285 : /// finishes. You can use wait_until_active() to wait for the task to
1286 : /// complete.
1287 : ///
1288 : #[allow(clippy::too_many_arguments)]
1289 0 : pub(crate) fn spawn(
1290 0 : conf: &'static PageServerConf,
1291 0 : tenant_shard_id: TenantShardId,
1292 0 : resources: TenantSharedResources,
1293 0 : attached_conf: AttachedTenantConf,
1294 0 : shard_identity: ShardIdentity,
1295 0 : init_order: Option<InitializationOrder>,
1296 0 : mode: SpawnMode,
1297 0 : ctx: &RequestContext,
1298 0 : ) -> Result<Arc<Tenant>, GlobalShutDown> {
1299 0 : let wal_redo_manager =
1300 0 : WalRedoManager::new(PostgresRedoManager::new(conf, tenant_shard_id))?;
1301 :
1302 : let TenantSharedResources {
1303 0 : broker_client,
1304 0 : remote_storage,
1305 0 : deletion_queue_client,
1306 0 : l0_flush_global_state,
1307 0 : } = resources;
1308 0 :
1309 0 : let attach_mode = attached_conf.location.attach_mode;
1310 0 : let generation = attached_conf.location.generation;
1311 0 :
1312 0 : let tenant = Arc::new(Tenant::new(
1313 0 : TenantState::Attaching,
1314 0 : conf,
1315 0 : attached_conf,
1316 0 : shard_identity,
1317 0 : Some(wal_redo_manager),
1318 0 : tenant_shard_id,
1319 0 : remote_storage.clone(),
1320 0 : deletion_queue_client,
1321 0 : l0_flush_global_state,
1322 0 : ));
1323 0 :
1324 0 : // The attach task will carry a GateGuard, so that shutdown() reliably waits for it to drop out if
1325 0 : // we shut down while attaching.
1326 0 : let attach_gate_guard = tenant
1327 0 : .gate
1328 0 : .enter()
1329 0 : .expect("We just created the Tenant: nothing else can have shut it down yet");
1330 0 :
1331 0 : // Do all the hard work in the background
1332 0 : let tenant_clone = Arc::clone(&tenant);
1333 0 : let ctx = ctx.detached_child(TaskKind::Attach, DownloadBehavior::Warn);
1334 0 : task_mgr::spawn(
1335 0 : &tokio::runtime::Handle::current(),
1336 0 : TaskKind::Attach,
1337 0 : tenant_shard_id,
1338 0 : None,
1339 0 : "attach tenant",
1340 0 : async move {
1341 0 :
1342 0 : info!(
1343 : ?attach_mode,
1344 0 : "Attaching tenant"
1345 : );
1346 :
1347 0 : let _gate_guard = attach_gate_guard;
1348 0 :
1349 0 : // Is this tenant being spawned as part of process startup?
1350 0 : let starting_up = init_order.is_some();
1351 0 : scopeguard::defer! {
1352 0 : if starting_up {
1353 0 : TENANT.startup_complete.inc();
1354 0 : }
1355 0 : }
1356 :
1357 : // Ideally we should use Tenant::set_broken_no_wait, but it is not supposed to be used when tenant is in loading state.
1358 : enum BrokenVerbosity {
1359 : Error,
1360 : Info
1361 : }
1362 0 : let make_broken =
1363 0 : |t: &Tenant, err: anyhow::Error, verbosity: BrokenVerbosity| {
1364 0 : match verbosity {
1365 : BrokenVerbosity::Info => {
1366 0 : info!("attach cancelled, setting tenant state to Broken: {err}");
1367 : },
1368 : BrokenVerbosity::Error => {
1369 0 : error!("attach failed, setting tenant state to Broken: {err:?}");
1370 : }
1371 : }
1372 0 : t.state.send_modify(|state| {
1373 0 : // The Stopping case is for when we have passed control on to DeleteTenantFlow:
1374 0 : // if it errors, we will call make_broken when tenant is already in Stopping.
1375 0 : assert!(
1376 0 : matches!(*state, TenantState::Attaching | TenantState::Stopping { .. }),
1377 0 : "the attach task owns the tenant state until activation is complete"
1378 : );
1379 :
1380 0 : *state = TenantState::broken_from_reason(err.to_string());
1381 0 : });
1382 0 : };
1383 :
1384 : // TODO: should also be rejecting tenant conf changes that violate this check.
1385 0 : if let Err(e) = crate::tenant::storage_layer::inmemory_layer::IndexEntry::validate_checkpoint_distance(tenant_clone.get_checkpoint_distance()) {
1386 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1387 0 : return Ok(());
1388 0 : }
1389 0 :
1390 0 : let mut init_order = init_order;
1391 0 : // take the completion because initial tenant loading will complete when all of
1392 0 : // these tasks complete.
1393 0 : let _completion = init_order
1394 0 : .as_mut()
1395 0 : .and_then(|x| x.initial_tenant_load.take());
1396 0 : let remote_load_completion = init_order
1397 0 : .as_mut()
1398 0 : .and_then(|x| x.initial_tenant_load_remote.take());
1399 :
1400 : enum AttachType<'a> {
1401 : /// We are attaching this tenant lazily in the background.
1402 : Warmup {
1403 : _permit: tokio::sync::SemaphorePermit<'a>,
1404 : during_startup: bool
1405 : },
1406 : /// We are attaching this tenant as soon as we can, because for example an
1407 : /// endpoint tried to access it.
1408 : OnDemand,
1409 : /// During normal operations after startup, we are attaching a tenant, and
1410 : /// eager attach was requested.
1411 : Normal,
1412 : }
1413 :
1414 0 : let attach_type = if matches!(mode, SpawnMode::Lazy) {
1415 : // Before doing any I/O, wait for at least one of:
1416 : // - A client attempting to access to this tenant (on-demand loading)
1417 : // - A permit becoming available in the warmup semaphore (background warmup)
1418 :
1419 0 : tokio::select!(
1420 0 : permit = tenant_clone.activate_now_sem.acquire() => {
1421 0 : let _ = permit.expect("activate_now_sem is never closed");
1422 0 : tracing::info!("Activating tenant (on-demand)");
1423 0 : AttachType::OnDemand
1424 : },
1425 0 : permit = conf.concurrent_tenant_warmup.inner().acquire() => {
1426 0 : let _permit = permit.expect("concurrent_tenant_warmup semaphore is never closed");
1427 0 : tracing::info!("Activating tenant (warmup)");
1428 0 : AttachType::Warmup {
1429 0 : _permit,
1430 0 : during_startup: init_order.is_some()
1431 0 : }
1432 : }
1433 0 : _ = tenant_clone.cancel.cancelled() => {
1434 : // This is safe, but should be pretty rare: it is interesting if a tenant
1435 : // stayed in Activating for such a long time that shutdown found it in
1436 : // that state.
1437 0 : tracing::info!(state=%tenant_clone.current_state(), "Tenant shut down before activation");
1438 : // Make the tenant broken so that set_stopping will not hang waiting for it to leave
1439 : // the Attaching state. This is an over-reaction (nothing really broke, the tenant is
1440 : // just shutting down), but ensures progress.
1441 0 : make_broken(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"), BrokenVerbosity::Info);
1442 0 : return Ok(());
1443 : },
1444 : )
1445 : } else {
1446 : // SpawnMode::{Create,Eager} always cause jumping ahead of the
1447 : // concurrent_tenant_warmup queue
1448 0 : AttachType::Normal
1449 : };
1450 :
1451 0 : let preload = match &mode {
1452 : SpawnMode::Eager | SpawnMode::Lazy => {
1453 0 : let _preload_timer = TENANT.preload.start_timer();
1454 0 : let res = tenant_clone
1455 0 : .preload(&remote_storage, task_mgr::shutdown_token())
1456 0 : .await;
1457 0 : match res {
1458 0 : Ok(p) => Some(p),
1459 0 : Err(e) => {
1460 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1461 0 : return Ok(());
1462 : }
1463 : }
1464 : }
1465 :
1466 : };
1467 :
1468 : // Remote preload is complete.
1469 0 : drop(remote_load_completion);
1470 0 :
1471 0 :
1472 0 : // We will time the duration of the attach phase unless this is a creation (attach will do no work)
1473 0 : let attach_start = std::time::Instant::now();
1474 0 : let attached = {
1475 0 : let _attach_timer = Some(TENANT.attach.start_timer());
1476 0 : tenant_clone.attach(preload, &ctx).await
1477 : };
1478 0 : let attach_duration = attach_start.elapsed();
1479 0 : _ = tenant_clone.attach_wal_lag_cooldown.set(WalLagCooldown::new(attach_start, attach_duration));
1480 0 :
1481 0 : match attached {
1482 : Ok(()) => {
1483 0 : info!("attach finished, activating");
1484 0 : tenant_clone.activate(broker_client, None, &ctx);
1485 : }
1486 0 : Err(e) => {
1487 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1488 0 : }
1489 : }
1490 :
1491 : // If we are doing an opportunistic warmup attachment at startup, initialize
1492 : // logical size at the same time. This is better than starting a bunch of idle tenants
1493 : // with cold caches and then coming back later to initialize their logical sizes.
1494 : //
1495 : // It also prevents the warmup proccess competing with the concurrency limit on
1496 : // logical size calculations: if logical size calculation semaphore is saturated,
1497 : // then warmup will wait for that before proceeding to the next tenant.
1498 0 : if matches!(attach_type, AttachType::Warmup { during_startup: true, .. }) {
1499 0 : let mut futs: FuturesUnordered<_> = tenant_clone.timelines.lock().unwrap().values().cloned().map(|t| t.await_initial_logical_size()).collect();
1500 0 : tracing::info!("Waiting for initial logical sizes while warming up...");
1501 0 : while futs.next().await.is_some() {}
1502 0 : tracing::info!("Warm-up complete");
1503 0 : }
1504 :
1505 0 : Ok(())
1506 0 : }
1507 0 : .instrument(tracing::info_span!(parent: None, "attach", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), gen=?generation)),
1508 : );
1509 0 : Ok(tenant)
1510 0 : }
1511 :
1512 : #[instrument(skip_all)]
1513 : pub(crate) async fn preload(
1514 : self: &Arc<Self>,
1515 : remote_storage: &GenericRemoteStorage,
1516 : cancel: CancellationToken,
1517 : ) -> anyhow::Result<TenantPreload> {
1518 : span::debug_assert_current_span_has_tenant_id();
1519 : // Get list of remote timelines
1520 : // download index files for every tenant timeline
1521 : info!("listing remote timelines");
1522 : let (mut remote_timeline_ids, other_keys) = remote_timeline_client::list_remote_timelines(
1523 : remote_storage,
1524 : self.tenant_shard_id,
1525 : cancel.clone(),
1526 : )
1527 : .await?;
1528 : let (offloaded_add, tenant_manifest) =
1529 : match remote_timeline_client::download_tenant_manifest(
1530 : remote_storage,
1531 : &self.tenant_shard_id,
1532 : self.generation,
1533 : &cancel,
1534 : )
1535 : .await
1536 : {
1537 : Ok((tenant_manifest, _generation, _manifest_mtime)) => (
1538 : format!("{} offloaded", tenant_manifest.offloaded_timelines.len()),
1539 : tenant_manifest,
1540 : ),
1541 : Err(DownloadError::NotFound) => {
1542 : ("no manifest".to_string(), TenantManifest::empty())
1543 : }
1544 : Err(e) => Err(e)?,
1545 : };
1546 :
1547 : info!(
1548 : "found {} timelines, and {offloaded_add}",
1549 : remote_timeline_ids.len()
1550 : );
1551 :
1552 : for k in other_keys {
1553 : warn!("Unexpected non timeline key {k}");
1554 : }
1555 :
1556 : // Avoid downloading IndexPart of offloaded timelines.
1557 : let mut offloaded_with_prefix = HashSet::new();
1558 : for offloaded in tenant_manifest.offloaded_timelines.iter() {
1559 : if remote_timeline_ids.remove(&offloaded.timeline_id) {
1560 : offloaded_with_prefix.insert(offloaded.timeline_id);
1561 : } else {
1562 : // We'll take care later of timelines in the manifest without a prefix
1563 : }
1564 : }
1565 :
1566 : // TODO(vlad): Could go to S3 if the secondary is freezing cold and hasn't even
1567 : // pulled the first heatmap. Not entirely necessary since the storage controller
1568 : // will kick the secondary in any case and cause a download.
1569 : let maybe_heatmap_at = self.read_on_disk_heatmap().await;
1570 :
1571 : let timelines = self
1572 : .load_timelines_metadata(
1573 : remote_timeline_ids,
1574 : remote_storage,
1575 : maybe_heatmap_at,
1576 : cancel,
1577 : )
1578 : .await?;
1579 :
1580 : Ok(TenantPreload {
1581 : tenant_manifest,
1582 : timelines: timelines
1583 : .into_iter()
1584 12 : .map(|(id, tl)| (id, Some(tl)))
1585 0 : .chain(offloaded_with_prefix.into_iter().map(|id| (id, None)))
1586 : .collect(),
1587 : })
1588 : }
1589 :
1590 444 : async fn read_on_disk_heatmap(&self) -> Option<(HeatMapTenant, std::time::Instant)> {
1591 444 : let on_disk_heatmap_path = self.conf.tenant_heatmap_path(&self.tenant_shard_id);
1592 444 : match tokio::fs::read_to_string(on_disk_heatmap_path).await {
1593 0 : Ok(heatmap) => match serde_json::from_str::<HeatMapTenant>(&heatmap) {
1594 0 : Ok(heatmap) => Some((heatmap, std::time::Instant::now())),
1595 0 : Err(err) => {
1596 0 : error!("Failed to deserialize old heatmap: {err}");
1597 0 : None
1598 : }
1599 : },
1600 444 : Err(err) => match err.kind() {
1601 444 : std::io::ErrorKind::NotFound => None,
1602 : _ => {
1603 0 : error!("Unexpected IO error reading old heatmap: {err}");
1604 0 : None
1605 : }
1606 : },
1607 : }
1608 444 : }
1609 :
1610 : ///
1611 : /// Background task that downloads all data for a tenant and brings it to Active state.
1612 : ///
1613 : /// No background tasks are started as part of this routine.
1614 : ///
1615 444 : async fn attach(
1616 444 : self: &Arc<Tenant>,
1617 444 : preload: Option<TenantPreload>,
1618 444 : ctx: &RequestContext,
1619 444 : ) -> anyhow::Result<()> {
1620 444 : span::debug_assert_current_span_has_tenant_id();
1621 444 :
1622 444 : failpoint_support::sleep_millis_async!("before-attaching-tenant");
1623 :
1624 444 : let Some(preload) = preload else {
1625 0 : anyhow::bail!("local-only deployment is no longer supported, https://github.com/neondatabase/neon/issues/5624");
1626 : };
1627 :
1628 444 : let mut offloaded_timeline_ids = HashSet::new();
1629 444 : let mut offloaded_timelines_list = Vec::new();
1630 444 : for timeline_manifest in preload.tenant_manifest.offloaded_timelines.iter() {
1631 0 : let timeline_id = timeline_manifest.timeline_id;
1632 0 : let offloaded_timeline =
1633 0 : OffloadedTimeline::from_manifest(self.tenant_shard_id, timeline_manifest);
1634 0 : offloaded_timelines_list.push((timeline_id, Arc::new(offloaded_timeline)));
1635 0 : offloaded_timeline_ids.insert(timeline_id);
1636 0 : }
1637 : // Complete deletions for offloaded timeline id's from manifest.
1638 : // The manifest will be uploaded later in this function.
1639 444 : offloaded_timelines_list
1640 444 : .retain(|(offloaded_id, offloaded)| {
1641 0 : // Existence of a timeline is finally determined by the existence of an index-part.json in remote storage.
1642 0 : // If there is dangling references in another location, they need to be cleaned up.
1643 0 : let delete = !preload.timelines.contains_key(offloaded_id);
1644 0 : if delete {
1645 0 : tracing::info!("Removing offloaded timeline {offloaded_id} from manifest as no remote prefix was found");
1646 0 : offloaded.defuse_for_tenant_drop();
1647 0 : }
1648 0 : !delete
1649 444 : });
1650 444 :
1651 444 : let mut timelines_to_resume_deletions = vec![];
1652 444 :
1653 444 : let mut remote_index_and_client = HashMap::new();
1654 444 : let mut timeline_ancestors = HashMap::new();
1655 444 : let mut existent_timelines = HashSet::new();
1656 456 : for (timeline_id, preload) in preload.timelines {
1657 12 : let Some(preload) = preload else { continue };
1658 : // This is an invariant of the `preload` function's API
1659 12 : assert!(!offloaded_timeline_ids.contains(&timeline_id));
1660 12 : let index_part = match preload.index_part {
1661 12 : Ok(i) => {
1662 12 : debug!("remote index part exists for timeline {timeline_id}");
1663 : // We found index_part on the remote, this is the standard case.
1664 12 : existent_timelines.insert(timeline_id);
1665 12 : i
1666 : }
1667 : Err(DownloadError::NotFound) => {
1668 : // There is no index_part on the remote. We only get here
1669 : // if there is some prefix for the timeline in the remote storage.
1670 : // This can e.g. be the initdb.tar.zst archive, maybe a
1671 : // remnant from a prior incomplete creation or deletion attempt.
1672 : // Delete the local directory as the deciding criterion for a
1673 : // timeline's existence is presence of index_part.
1674 0 : info!(%timeline_id, "index_part not found on remote");
1675 0 : continue;
1676 : }
1677 0 : Err(DownloadError::Fatal(why)) => {
1678 0 : // If, while loading one remote timeline, we saw an indication that our generation
1679 0 : // number is likely invalid, then we should not load the whole tenant.
1680 0 : error!(%timeline_id, "Fatal error loading timeline: {why}");
1681 0 : anyhow::bail!(why.to_string());
1682 : }
1683 0 : Err(e) => {
1684 0 : // Some (possibly ephemeral) error happened during index_part download.
1685 0 : // Pretend the timeline exists to not delete the timeline directory,
1686 0 : // as it might be a temporary issue and we don't want to re-download
1687 0 : // everything after it resolves.
1688 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1689 :
1690 0 : existent_timelines.insert(timeline_id);
1691 0 : continue;
1692 : }
1693 : };
1694 12 : match index_part {
1695 12 : MaybeDeletedIndexPart::IndexPart(index_part) => {
1696 12 : timeline_ancestors.insert(timeline_id, index_part.metadata.clone());
1697 12 : remote_index_and_client.insert(
1698 12 : timeline_id,
1699 12 : (index_part, preload.client, preload.previous_heatmap),
1700 12 : );
1701 12 : }
1702 0 : MaybeDeletedIndexPart::Deleted(index_part) => {
1703 0 : info!(
1704 0 : "timeline {} is deleted, picking to resume deletion",
1705 : timeline_id
1706 : );
1707 0 : timelines_to_resume_deletions.push((timeline_id, index_part, preload.client));
1708 : }
1709 : }
1710 : }
1711 :
1712 444 : let mut gc_blocks = HashMap::new();
1713 :
1714 : // For every timeline, download the metadata file, scan the local directory,
1715 : // and build a layer map that contains an entry for each remote and local
1716 : // layer file.
1717 444 : let sorted_timelines = tree_sort_timelines(timeline_ancestors, |m| m.ancestor_timeline())?;
1718 456 : for (timeline_id, remote_metadata) in sorted_timelines {
1719 12 : let (index_part, remote_client, previous_heatmap) = remote_index_and_client
1720 12 : .remove(&timeline_id)
1721 12 : .expect("just put it in above");
1722 :
1723 12 : if let Some(blocking) = index_part.gc_blocking.as_ref() {
1724 : // could just filter these away, but it helps while testing
1725 0 : anyhow::ensure!(
1726 0 : !blocking.reasons.is_empty(),
1727 0 : "index_part for {timeline_id} is malformed: it should not have gc blocking with zero reasons"
1728 : );
1729 0 : let prev = gc_blocks.insert(timeline_id, blocking.reasons);
1730 0 : assert!(prev.is_none());
1731 12 : }
1732 :
1733 : // TODO again handle early failure
1734 12 : let effect = self
1735 12 : .load_remote_timeline(
1736 12 : timeline_id,
1737 12 : index_part,
1738 12 : remote_metadata,
1739 12 : previous_heatmap,
1740 12 : self.get_timeline_resources_for(remote_client),
1741 12 : LoadTimelineCause::Attach,
1742 12 : ctx,
1743 12 : )
1744 12 : .await
1745 12 : .with_context(|| {
1746 0 : format!(
1747 0 : "failed to load remote timeline {} for tenant {}",
1748 0 : timeline_id, self.tenant_shard_id
1749 0 : )
1750 12 : })?;
1751 :
1752 12 : match effect {
1753 12 : TimelineInitAndSyncResult::ReadyToActivate(_) => {
1754 12 : // activation happens later, on Tenant::activate
1755 12 : }
1756 : TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
1757 : TimelineInitAndSyncNeedsSpawnImportPgdata {
1758 0 : timeline,
1759 0 : import_pgdata,
1760 0 : guard,
1761 0 : },
1762 0 : ) => {
1763 0 : tokio::task::spawn(self.clone().create_timeline_import_pgdata_task(
1764 0 : timeline,
1765 0 : import_pgdata,
1766 0 : ActivateTimelineArgs::No,
1767 0 : guard,
1768 0 : ));
1769 0 : }
1770 : }
1771 : }
1772 :
1773 : // Walk through deleted timelines, resume deletion
1774 444 : for (timeline_id, index_part, remote_timeline_client) in timelines_to_resume_deletions {
1775 0 : remote_timeline_client
1776 0 : .init_upload_queue_stopped_to_continue_deletion(&index_part)
1777 0 : .context("init queue stopped")
1778 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1779 :
1780 0 : DeleteTimelineFlow::resume_deletion(
1781 0 : Arc::clone(self),
1782 0 : timeline_id,
1783 0 : &index_part.metadata,
1784 0 : remote_timeline_client,
1785 0 : )
1786 0 : .instrument(tracing::info_span!("timeline_delete", %timeline_id))
1787 0 : .await
1788 0 : .context("resume_deletion")
1789 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1790 : }
1791 444 : let needs_manifest_upload =
1792 444 : offloaded_timelines_list.len() != preload.tenant_manifest.offloaded_timelines.len();
1793 444 : {
1794 444 : let mut offloaded_timelines_accessor = self.timelines_offloaded.lock().unwrap();
1795 444 : offloaded_timelines_accessor.extend(offloaded_timelines_list.into_iter());
1796 444 : }
1797 444 : if needs_manifest_upload {
1798 0 : self.store_tenant_manifest().await?;
1799 444 : }
1800 :
1801 : // The local filesystem contents are a cache of what's in the remote IndexPart;
1802 : // IndexPart is the source of truth.
1803 444 : self.clean_up_timelines(&existent_timelines)?;
1804 :
1805 444 : self.gc_block.set_scanned(gc_blocks);
1806 444 :
1807 444 : fail::fail_point!("attach-before-activate", |_| {
1808 0 : anyhow::bail!("attach-before-activate");
1809 444 : });
1810 444 : failpoint_support::sleep_millis_async!("attach-before-activate-sleep", &self.cancel);
1811 :
1812 444 : info!("Done");
1813 :
1814 444 : Ok(())
1815 444 : }
1816 :
1817 : /// Check for any local timeline directories that are temporary, or do not correspond to a
1818 : /// timeline that still exists: this can happen if we crashed during a deletion/creation, or
1819 : /// if a timeline was deleted while the tenant was attached to a different pageserver.
1820 444 : fn clean_up_timelines(&self, existent_timelines: &HashSet<TimelineId>) -> anyhow::Result<()> {
1821 444 : let timelines_dir = self.conf.timelines_path(&self.tenant_shard_id);
1822 :
1823 444 : let entries = match timelines_dir.read_dir_utf8() {
1824 444 : Ok(d) => d,
1825 0 : Err(e) => {
1826 0 : if e.kind() == std::io::ErrorKind::NotFound {
1827 0 : return Ok(());
1828 : } else {
1829 0 : return Err(e).context("list timelines directory for tenant");
1830 : }
1831 : }
1832 : };
1833 :
1834 460 : for entry in entries {
1835 16 : let entry = entry.context("read timeline dir entry")?;
1836 16 : let entry_path = entry.path();
1837 :
1838 16 : let purge = if crate::is_temporary(entry_path) {
1839 0 : true
1840 : } else {
1841 16 : match TimelineId::try_from(entry_path.file_name()) {
1842 16 : Ok(i) => {
1843 16 : // Purge if the timeline ID does not exist in remote storage: remote storage is the authority.
1844 16 : !existent_timelines.contains(&i)
1845 : }
1846 0 : Err(e) => {
1847 0 : tracing::warn!(
1848 0 : "Unparseable directory in timelines directory: {entry_path}, ignoring ({e})"
1849 : );
1850 : // Do not purge junk: if we don't recognize it, be cautious and leave it for a human.
1851 0 : false
1852 : }
1853 : }
1854 : };
1855 :
1856 16 : if purge {
1857 4 : tracing::info!("Purging stale timeline dentry {entry_path}");
1858 4 : if let Err(e) = match entry.file_type() {
1859 4 : Ok(t) => if t.is_dir() {
1860 4 : std::fs::remove_dir_all(entry_path)
1861 : } else {
1862 0 : std::fs::remove_file(entry_path)
1863 : }
1864 4 : .or_else(fs_ext::ignore_not_found),
1865 0 : Err(e) => Err(e),
1866 : } {
1867 0 : tracing::warn!("Failed to purge stale timeline dentry {entry_path}: {e}");
1868 4 : }
1869 12 : }
1870 : }
1871 :
1872 444 : Ok(())
1873 444 : }
1874 :
1875 : /// Get sum of all remote timelines sizes
1876 : ///
1877 : /// This function relies on the index_part instead of listing the remote storage
1878 0 : pub fn remote_size(&self) -> u64 {
1879 0 : let mut size = 0;
1880 :
1881 0 : for timeline in self.list_timelines() {
1882 0 : size += timeline.remote_client.get_remote_physical_size();
1883 0 : }
1884 :
1885 0 : size
1886 0 : }
1887 :
1888 : #[instrument(skip_all, fields(timeline_id=%timeline_id))]
1889 : #[allow(clippy::too_many_arguments)]
1890 : async fn load_remote_timeline(
1891 : self: &Arc<Self>,
1892 : timeline_id: TimelineId,
1893 : index_part: IndexPart,
1894 : remote_metadata: TimelineMetadata,
1895 : previous_heatmap: Option<PreviousHeatmap>,
1896 : resources: TimelineResources,
1897 : cause: LoadTimelineCause,
1898 : ctx: &RequestContext,
1899 : ) -> anyhow::Result<TimelineInitAndSyncResult> {
1900 : span::debug_assert_current_span_has_tenant_id();
1901 :
1902 : info!("downloading index file for timeline {}", timeline_id);
1903 : tokio::fs::create_dir_all(self.conf.timeline_path(&self.tenant_shard_id, &timeline_id))
1904 : .await
1905 : .context("Failed to create new timeline directory")?;
1906 :
1907 : let ancestor = if let Some(ancestor_id) = remote_metadata.ancestor_timeline() {
1908 : let timelines = self.timelines.lock().unwrap();
1909 : Some(Arc::clone(timelines.get(&ancestor_id).ok_or_else(
1910 0 : || {
1911 0 : anyhow::anyhow!(
1912 0 : "cannot find ancestor timeline {ancestor_id} for timeline {timeline_id}"
1913 0 : )
1914 0 : },
1915 : )?))
1916 : } else {
1917 : None
1918 : };
1919 :
1920 : self.timeline_init_and_sync(
1921 : timeline_id,
1922 : resources,
1923 : index_part,
1924 : remote_metadata,
1925 : previous_heatmap,
1926 : ancestor,
1927 : cause,
1928 : ctx,
1929 : )
1930 : .await
1931 : }
1932 :
1933 444 : async fn load_timelines_metadata(
1934 444 : self: &Arc<Tenant>,
1935 444 : timeline_ids: HashSet<TimelineId>,
1936 444 : remote_storage: &GenericRemoteStorage,
1937 444 : heatmap: Option<(HeatMapTenant, std::time::Instant)>,
1938 444 : cancel: CancellationToken,
1939 444 : ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
1940 444 : let mut timeline_heatmaps = heatmap.map(|h| (h.0.into_timelines_index(), h.1));
1941 444 :
1942 444 : let mut part_downloads = JoinSet::new();
1943 456 : for timeline_id in timeline_ids {
1944 12 : let cancel_clone = cancel.clone();
1945 12 :
1946 12 : let previous_timeline_heatmap = timeline_heatmaps.as_mut().and_then(|hs| {
1947 0 : hs.0.remove(&timeline_id).map(|h| PreviousHeatmap::Active {
1948 0 : heatmap: h,
1949 0 : read_at: hs.1,
1950 0 : })
1951 12 : });
1952 12 : part_downloads.spawn(
1953 12 : self.load_timeline_metadata(
1954 12 : timeline_id,
1955 12 : remote_storage.clone(),
1956 12 : previous_timeline_heatmap,
1957 12 : cancel_clone,
1958 12 : )
1959 12 : .instrument(info_span!("download_index_part", %timeline_id)),
1960 : );
1961 : }
1962 :
1963 444 : let mut timeline_preloads: HashMap<TimelineId, TimelinePreload> = HashMap::new();
1964 :
1965 : loop {
1966 456 : tokio::select!(
1967 456 : next = part_downloads.join_next() => {
1968 456 : match next {
1969 12 : Some(result) => {
1970 12 : let preload = result.context("join preload task")?;
1971 12 : timeline_preloads.insert(preload.timeline_id, preload);
1972 : },
1973 : None => {
1974 444 : break;
1975 : }
1976 : }
1977 : },
1978 456 : _ = cancel.cancelled() => {
1979 0 : anyhow::bail!("Cancelled while waiting for remote index download")
1980 : }
1981 : )
1982 : }
1983 :
1984 444 : Ok(timeline_preloads)
1985 444 : }
1986 :
1987 12 : fn build_timeline_client(
1988 12 : &self,
1989 12 : timeline_id: TimelineId,
1990 12 : remote_storage: GenericRemoteStorage,
1991 12 : ) -> RemoteTimelineClient {
1992 12 : RemoteTimelineClient::new(
1993 12 : remote_storage.clone(),
1994 12 : self.deletion_queue_client.clone(),
1995 12 : self.conf,
1996 12 : self.tenant_shard_id,
1997 12 : timeline_id,
1998 12 : self.generation,
1999 12 : &self.tenant_conf.load().location,
2000 12 : )
2001 12 : }
2002 :
2003 12 : fn load_timeline_metadata(
2004 12 : self: &Arc<Tenant>,
2005 12 : timeline_id: TimelineId,
2006 12 : remote_storage: GenericRemoteStorage,
2007 12 : previous_heatmap: Option<PreviousHeatmap>,
2008 12 : cancel: CancellationToken,
2009 12 : ) -> impl Future<Output = TimelinePreload> {
2010 12 : let client = self.build_timeline_client(timeline_id, remote_storage);
2011 12 : async move {
2012 12 : debug_assert_current_span_has_tenant_and_timeline_id();
2013 12 : debug!("starting index part download");
2014 :
2015 12 : let index_part = client.download_index_file(&cancel).await;
2016 :
2017 12 : debug!("finished index part download");
2018 :
2019 12 : TimelinePreload {
2020 12 : client,
2021 12 : timeline_id,
2022 12 : index_part,
2023 12 : previous_heatmap,
2024 12 : }
2025 12 : }
2026 12 : }
2027 :
2028 0 : fn check_to_be_archived_has_no_unarchived_children(
2029 0 : timeline_id: TimelineId,
2030 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
2031 0 : ) -> Result<(), TimelineArchivalError> {
2032 0 : let children: Vec<TimelineId> = timelines
2033 0 : .iter()
2034 0 : .filter_map(|(id, entry)| {
2035 0 : if entry.get_ancestor_timeline_id() != Some(timeline_id) {
2036 0 : return None;
2037 0 : }
2038 0 : if entry.is_archived() == Some(true) {
2039 0 : return None;
2040 0 : }
2041 0 : Some(*id)
2042 0 : })
2043 0 : .collect();
2044 0 :
2045 0 : if !children.is_empty() {
2046 0 : return Err(TimelineArchivalError::HasUnarchivedChildren(children));
2047 0 : }
2048 0 : Ok(())
2049 0 : }
2050 :
2051 0 : fn check_ancestor_of_to_be_unarchived_is_not_archived(
2052 0 : ancestor_timeline_id: TimelineId,
2053 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
2054 0 : offloaded_timelines: &std::sync::MutexGuard<
2055 0 : '_,
2056 0 : HashMap<TimelineId, Arc<OffloadedTimeline>>,
2057 0 : >,
2058 0 : ) -> Result<(), TimelineArchivalError> {
2059 0 : let has_archived_parent =
2060 0 : if let Some(ancestor_timeline) = timelines.get(&ancestor_timeline_id) {
2061 0 : ancestor_timeline.is_archived() == Some(true)
2062 0 : } else if offloaded_timelines.contains_key(&ancestor_timeline_id) {
2063 0 : true
2064 : } else {
2065 0 : error!("ancestor timeline {ancestor_timeline_id} not found");
2066 0 : if cfg!(debug_assertions) {
2067 0 : panic!("ancestor timeline {ancestor_timeline_id} not found");
2068 0 : }
2069 0 : return Err(TimelineArchivalError::NotFound);
2070 : };
2071 0 : if has_archived_parent {
2072 0 : return Err(TimelineArchivalError::HasArchivedParent(
2073 0 : ancestor_timeline_id,
2074 0 : ));
2075 0 : }
2076 0 : Ok(())
2077 0 : }
2078 :
2079 0 : fn check_to_be_unarchived_timeline_has_no_archived_parent(
2080 0 : timeline: &Arc<Timeline>,
2081 0 : ) -> Result<(), TimelineArchivalError> {
2082 0 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
2083 0 : if ancestor_timeline.is_archived() == Some(true) {
2084 0 : return Err(TimelineArchivalError::HasArchivedParent(
2085 0 : ancestor_timeline.timeline_id,
2086 0 : ));
2087 0 : }
2088 0 : }
2089 0 : Ok(())
2090 0 : }
2091 :
2092 : /// Loads the specified (offloaded) timeline from S3 and attaches it as a loaded timeline
2093 : ///
2094 : /// Counterpart to [`offload_timeline`].
2095 0 : async fn unoffload_timeline(
2096 0 : self: &Arc<Self>,
2097 0 : timeline_id: TimelineId,
2098 0 : broker_client: storage_broker::BrokerClientChannel,
2099 0 : ctx: RequestContext,
2100 0 : ) -> Result<Arc<Timeline>, TimelineArchivalError> {
2101 0 : info!("unoffloading timeline");
2102 :
2103 : // We activate the timeline below manually, so this must be called on an active tenant.
2104 : // We expect callers of this function to ensure this.
2105 0 : match self.current_state() {
2106 : TenantState::Activating { .. }
2107 : | TenantState::Attaching
2108 : | TenantState::Broken { .. } => {
2109 0 : panic!("Timeline expected to be active")
2110 : }
2111 0 : TenantState::Stopping { .. } => return Err(TimelineArchivalError::Cancelled),
2112 0 : TenantState::Active => {}
2113 0 : }
2114 0 : let cancel = self.cancel.clone();
2115 0 :
2116 0 : // Protect against concurrent attempts to use this TimelineId
2117 0 : // We don't care much about idempotency, as it's ensured a layer above.
2118 0 : let allow_offloaded = true;
2119 0 : let _create_guard = self
2120 0 : .create_timeline_create_guard(
2121 0 : timeline_id,
2122 0 : CreateTimelineIdempotency::FailWithConflict,
2123 0 : allow_offloaded,
2124 0 : )
2125 0 : .map_err(|err| match err {
2126 0 : TimelineExclusionError::AlreadyCreating => TimelineArchivalError::AlreadyInProgress,
2127 : TimelineExclusionError::AlreadyExists { .. } => {
2128 0 : TimelineArchivalError::Other(anyhow::anyhow!("Timeline already exists"))
2129 : }
2130 0 : TimelineExclusionError::Other(e) => TimelineArchivalError::Other(e),
2131 0 : TimelineExclusionError::ShuttingDown => TimelineArchivalError::Cancelled,
2132 0 : })?;
2133 :
2134 0 : let timeline_preload = self
2135 0 : .load_timeline_metadata(
2136 0 : timeline_id,
2137 0 : self.remote_storage.clone(),
2138 0 : None,
2139 0 : cancel.clone(),
2140 0 : )
2141 0 : .await;
2142 :
2143 0 : let index_part = match timeline_preload.index_part {
2144 0 : Ok(index_part) => {
2145 0 : debug!("remote index part exists for timeline {timeline_id}");
2146 0 : index_part
2147 : }
2148 : Err(DownloadError::NotFound) => {
2149 0 : error!(%timeline_id, "index_part not found on remote");
2150 0 : return Err(TimelineArchivalError::NotFound);
2151 : }
2152 0 : Err(DownloadError::Cancelled) => return Err(TimelineArchivalError::Cancelled),
2153 0 : Err(e) => {
2154 0 : // Some (possibly ephemeral) error happened during index_part download.
2155 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
2156 0 : return Err(TimelineArchivalError::Other(
2157 0 : anyhow::Error::new(e).context("downloading index_part from remote storage"),
2158 0 : ));
2159 : }
2160 : };
2161 0 : let index_part = match index_part {
2162 0 : MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
2163 0 : MaybeDeletedIndexPart::Deleted(_index_part) => {
2164 0 : info!("timeline is deleted according to index_part.json");
2165 0 : return Err(TimelineArchivalError::NotFound);
2166 : }
2167 : };
2168 0 : let remote_metadata = index_part.metadata.clone();
2169 0 : let timeline_resources = self.build_timeline_resources(timeline_id);
2170 0 : self.load_remote_timeline(
2171 0 : timeline_id,
2172 0 : index_part,
2173 0 : remote_metadata,
2174 0 : None,
2175 0 : timeline_resources,
2176 0 : LoadTimelineCause::Unoffload,
2177 0 : &ctx,
2178 0 : )
2179 0 : .await
2180 0 : .with_context(|| {
2181 0 : format!(
2182 0 : "failed to load remote timeline {} for tenant {}",
2183 0 : timeline_id, self.tenant_shard_id
2184 0 : )
2185 0 : })
2186 0 : .map_err(TimelineArchivalError::Other)?;
2187 :
2188 0 : let timeline = {
2189 0 : let timelines = self.timelines.lock().unwrap();
2190 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2191 0 : warn!("timeline not available directly after attach");
2192 : // This is not a panic because no locks are held between `load_remote_timeline`
2193 : // which puts the timeline into timelines, and our look into the timeline map.
2194 0 : return Err(TimelineArchivalError::Other(anyhow::anyhow!(
2195 0 : "timeline not available directly after attach"
2196 0 : )));
2197 : };
2198 0 : let mut offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2199 0 : match offloaded_timelines.remove(&timeline_id) {
2200 0 : Some(offloaded) => {
2201 0 : offloaded.delete_from_ancestor_with_timelines(&timelines);
2202 0 : }
2203 0 : None => warn!("timeline already removed from offloaded timelines"),
2204 : }
2205 :
2206 0 : self.initialize_gc_info(&timelines, &offloaded_timelines, Some(timeline_id));
2207 0 :
2208 0 : Arc::clone(timeline)
2209 0 : };
2210 0 :
2211 0 : // Upload new list of offloaded timelines to S3
2212 0 : self.store_tenant_manifest().await?;
2213 :
2214 : // Activate the timeline (if it makes sense)
2215 0 : if !(timeline.is_broken() || timeline.is_stopping()) {
2216 0 : let background_jobs_can_start = None;
2217 0 : timeline.activate(
2218 0 : self.clone(),
2219 0 : broker_client.clone(),
2220 0 : background_jobs_can_start,
2221 0 : &ctx,
2222 0 : );
2223 0 : }
2224 :
2225 0 : info!("timeline unoffloading complete");
2226 0 : Ok(timeline)
2227 0 : }
2228 :
2229 0 : pub(crate) async fn apply_timeline_archival_config(
2230 0 : self: &Arc<Self>,
2231 0 : timeline_id: TimelineId,
2232 0 : new_state: TimelineArchivalState,
2233 0 : broker_client: storage_broker::BrokerClientChannel,
2234 0 : ctx: RequestContext,
2235 0 : ) -> Result<(), TimelineArchivalError> {
2236 0 : info!("setting timeline archival config");
2237 : // First part: figure out what is needed to do, and do validation
2238 0 : let timeline_or_unarchive_offloaded = 'outer: {
2239 0 : let timelines = self.timelines.lock().unwrap();
2240 :
2241 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2242 0 : let offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2243 0 : let Some(offloaded) = offloaded_timelines.get(&timeline_id) else {
2244 0 : return Err(TimelineArchivalError::NotFound);
2245 : };
2246 0 : if new_state == TimelineArchivalState::Archived {
2247 : // It's offloaded already, so nothing to do
2248 0 : return Ok(());
2249 0 : }
2250 0 : if let Some(ancestor_timeline_id) = offloaded.ancestor_timeline_id {
2251 0 : Self::check_ancestor_of_to_be_unarchived_is_not_archived(
2252 0 : ancestor_timeline_id,
2253 0 : &timelines,
2254 0 : &offloaded_timelines,
2255 0 : )?;
2256 0 : }
2257 0 : break 'outer None;
2258 : };
2259 :
2260 : // Do some validation. We release the timelines lock below, so there is potential
2261 : // for race conditions: these checks are more present to prevent misunderstandings of
2262 : // the API's capabilities, instead of serving as the sole way to defend their invariants.
2263 0 : match new_state {
2264 : TimelineArchivalState::Unarchived => {
2265 0 : Self::check_to_be_unarchived_timeline_has_no_archived_parent(timeline)?
2266 : }
2267 : TimelineArchivalState::Archived => {
2268 0 : Self::check_to_be_archived_has_no_unarchived_children(timeline_id, &timelines)?
2269 : }
2270 : }
2271 0 : Some(Arc::clone(timeline))
2272 : };
2273 :
2274 : // Second part: unoffload timeline (if needed)
2275 0 : let timeline = if let Some(timeline) = timeline_or_unarchive_offloaded {
2276 0 : timeline
2277 : } else {
2278 : // Turn offloaded timeline into a non-offloaded one
2279 0 : self.unoffload_timeline(timeline_id, broker_client, ctx)
2280 0 : .await?
2281 : };
2282 :
2283 : // Third part: upload new timeline archival state and block until it is present in S3
2284 0 : let upload_needed = match timeline
2285 0 : .remote_client
2286 0 : .schedule_index_upload_for_timeline_archival_state(new_state)
2287 : {
2288 0 : Ok(upload_needed) => upload_needed,
2289 0 : Err(e) => {
2290 0 : if timeline.cancel.is_cancelled() {
2291 0 : return Err(TimelineArchivalError::Cancelled);
2292 : } else {
2293 0 : return Err(TimelineArchivalError::Other(e));
2294 : }
2295 : }
2296 : };
2297 :
2298 0 : if upload_needed {
2299 0 : info!("Uploading new state");
2300 : const MAX_WAIT: Duration = Duration::from_secs(10);
2301 0 : let Ok(v) =
2302 0 : tokio::time::timeout(MAX_WAIT, timeline.remote_client.wait_completion()).await
2303 : else {
2304 0 : tracing::warn!("reached timeout for waiting on upload queue");
2305 0 : return Err(TimelineArchivalError::Timeout);
2306 : };
2307 0 : v.map_err(|e| match e {
2308 0 : WaitCompletionError::NotInitialized(e) => {
2309 0 : TimelineArchivalError::Other(anyhow::anyhow!(e))
2310 : }
2311 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2312 0 : TimelineArchivalError::Cancelled
2313 : }
2314 0 : })?;
2315 0 : }
2316 0 : Ok(())
2317 0 : }
2318 :
2319 4 : pub fn get_offloaded_timeline(
2320 4 : &self,
2321 4 : timeline_id: TimelineId,
2322 4 : ) -> Result<Arc<OffloadedTimeline>, GetTimelineError> {
2323 4 : self.timelines_offloaded
2324 4 : .lock()
2325 4 : .unwrap()
2326 4 : .get(&timeline_id)
2327 4 : .map(Arc::clone)
2328 4 : .ok_or(GetTimelineError::NotFound {
2329 4 : tenant_id: self.tenant_shard_id,
2330 4 : timeline_id,
2331 4 : })
2332 4 : }
2333 :
2334 8 : pub(crate) fn tenant_shard_id(&self) -> TenantShardId {
2335 8 : self.tenant_shard_id
2336 8 : }
2337 :
2338 : /// Get Timeline handle for given Neon timeline ID.
2339 : /// This function is idempotent. It doesn't change internal state in any way.
2340 444 : pub fn get_timeline(
2341 444 : &self,
2342 444 : timeline_id: TimelineId,
2343 444 : active_only: bool,
2344 444 : ) -> Result<Arc<Timeline>, GetTimelineError> {
2345 444 : let timelines_accessor = self.timelines.lock().unwrap();
2346 444 : let timeline = timelines_accessor
2347 444 : .get(&timeline_id)
2348 444 : .ok_or(GetTimelineError::NotFound {
2349 444 : tenant_id: self.tenant_shard_id,
2350 444 : timeline_id,
2351 444 : })?;
2352 :
2353 440 : if active_only && !timeline.is_active() {
2354 0 : Err(GetTimelineError::NotActive {
2355 0 : tenant_id: self.tenant_shard_id,
2356 0 : timeline_id,
2357 0 : state: timeline.current_state(),
2358 0 : })
2359 : } else {
2360 440 : Ok(Arc::clone(timeline))
2361 : }
2362 444 : }
2363 :
2364 : /// Lists timelines the tenant contains.
2365 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2366 0 : pub fn list_timelines(&self) -> Vec<Arc<Timeline>> {
2367 0 : self.timelines
2368 0 : .lock()
2369 0 : .unwrap()
2370 0 : .values()
2371 0 : .map(Arc::clone)
2372 0 : .collect()
2373 0 : }
2374 :
2375 : /// Lists timelines the tenant manages, including offloaded ones.
2376 : ///
2377 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2378 0 : pub fn list_timelines_and_offloaded(
2379 0 : &self,
2380 0 : ) -> (Vec<Arc<Timeline>>, Vec<Arc<OffloadedTimeline>>) {
2381 0 : let timelines = self
2382 0 : .timelines
2383 0 : .lock()
2384 0 : .unwrap()
2385 0 : .values()
2386 0 : .map(Arc::clone)
2387 0 : .collect();
2388 0 : let offloaded = self
2389 0 : .timelines_offloaded
2390 0 : .lock()
2391 0 : .unwrap()
2392 0 : .values()
2393 0 : .map(Arc::clone)
2394 0 : .collect();
2395 0 : (timelines, offloaded)
2396 0 : }
2397 :
2398 0 : pub fn list_timeline_ids(&self) -> Vec<TimelineId> {
2399 0 : self.timelines.lock().unwrap().keys().cloned().collect()
2400 0 : }
2401 :
2402 : /// This is used by tests & import-from-basebackup.
2403 : ///
2404 : /// The returned [`UninitializedTimeline`] contains no data nor metadata and it is in
2405 : /// a state that will fail [`Tenant::load_remote_timeline`] because `disk_consistent_lsn=Lsn(0)`.
2406 : ///
2407 : /// The caller is responsible for getting the timeline into a state that will be accepted
2408 : /// by [`Tenant::load_remote_timeline`] / [`Tenant::attach`].
2409 : /// Then they may call [`UninitializedTimeline::finish_creation`] to add the timeline
2410 : /// to the [`Tenant::timelines`].
2411 : ///
2412 : /// Tests should use `Tenant::create_test_timeline` to set up the minimum required metadata keys.
2413 428 : pub(crate) async fn create_empty_timeline(
2414 428 : self: &Arc<Self>,
2415 428 : new_timeline_id: TimelineId,
2416 428 : initdb_lsn: Lsn,
2417 428 : pg_version: u32,
2418 428 : _ctx: &RequestContext,
2419 428 : ) -> anyhow::Result<UninitializedTimeline> {
2420 428 : anyhow::ensure!(
2421 428 : self.is_active(),
2422 0 : "Cannot create empty timelines on inactive tenant"
2423 : );
2424 :
2425 : // Protect against concurrent attempts to use this TimelineId
2426 428 : let create_guard = match self
2427 428 : .start_creating_timeline(new_timeline_id, CreateTimelineIdempotency::FailWithConflict)
2428 428 : .await?
2429 : {
2430 424 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2431 : StartCreatingTimelineResult::Idempotent(_) => {
2432 0 : unreachable!("FailWithConflict implies we get an error instead")
2433 : }
2434 : };
2435 :
2436 424 : let new_metadata = TimelineMetadata::new(
2437 424 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2438 424 : // make it valid, before calling finish_creation()
2439 424 : Lsn(0),
2440 424 : None,
2441 424 : None,
2442 424 : Lsn(0),
2443 424 : initdb_lsn,
2444 424 : initdb_lsn,
2445 424 : pg_version,
2446 424 : );
2447 424 : self.prepare_new_timeline(
2448 424 : new_timeline_id,
2449 424 : &new_metadata,
2450 424 : create_guard,
2451 424 : initdb_lsn,
2452 424 : None,
2453 424 : )
2454 424 : .await
2455 428 : }
2456 :
2457 : /// Helper for unit tests to create an empty timeline.
2458 : ///
2459 : /// The timeline is has state value `Active` but its background loops are not running.
2460 : // This makes the various functions which anyhow::ensure! for Active state work in tests.
2461 : // Our current tests don't need the background loops.
2462 : #[cfg(test)]
2463 408 : pub async fn create_test_timeline(
2464 408 : self: &Arc<Self>,
2465 408 : new_timeline_id: TimelineId,
2466 408 : initdb_lsn: Lsn,
2467 408 : pg_version: u32,
2468 408 : ctx: &RequestContext,
2469 408 : ) -> anyhow::Result<Arc<Timeline>> {
2470 408 : let uninit_tl = self
2471 408 : .create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2472 408 : .await?;
2473 408 : let tline = uninit_tl.raw_timeline().expect("we just created it");
2474 408 : assert_eq!(tline.get_last_record_lsn(), Lsn(0));
2475 :
2476 : // Setup minimum keys required for the timeline to be usable.
2477 408 : let mut modification = tline.begin_modification(initdb_lsn);
2478 408 : modification
2479 408 : .init_empty_test_timeline()
2480 408 : .context("init_empty_test_timeline")?;
2481 408 : modification
2482 408 : .commit(ctx)
2483 408 : .await
2484 408 : .context("commit init_empty_test_timeline modification")?;
2485 :
2486 : // Flush to disk so that uninit_tl's check for valid disk_consistent_lsn passes.
2487 408 : tline.maybe_spawn_flush_loop();
2488 408 : tline.freeze_and_flush().await.context("freeze_and_flush")?;
2489 :
2490 : // Make sure the freeze_and_flush reaches remote storage.
2491 408 : tline.remote_client.wait_completion().await.unwrap();
2492 :
2493 408 : let tl = uninit_tl.finish_creation().await?;
2494 : // The non-test code would call tl.activate() here.
2495 408 : tl.set_state(TimelineState::Active);
2496 408 : Ok(tl)
2497 408 : }
2498 :
2499 : /// Helper for unit tests to create a timeline with some pre-loaded states.
2500 : #[cfg(test)]
2501 : #[allow(clippy::too_many_arguments)]
2502 80 : pub async fn create_test_timeline_with_layers(
2503 80 : self: &Arc<Self>,
2504 80 : new_timeline_id: TimelineId,
2505 80 : initdb_lsn: Lsn,
2506 80 : pg_version: u32,
2507 80 : ctx: &RequestContext,
2508 80 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
2509 80 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
2510 80 : end_lsn: Lsn,
2511 80 : ) -> anyhow::Result<Arc<Timeline>> {
2512 : use checks::check_valid_layermap;
2513 : use itertools::Itertools;
2514 :
2515 80 : let tline = self
2516 80 : .create_test_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2517 80 : .await?;
2518 80 : tline.force_advance_lsn(end_lsn);
2519 252 : for deltas in delta_layer_desc {
2520 172 : tline
2521 172 : .force_create_delta_layer(deltas, Some(initdb_lsn), ctx)
2522 172 : .await?;
2523 : }
2524 192 : for (lsn, images) in image_layer_desc {
2525 112 : tline
2526 112 : .force_create_image_layer(lsn, images, Some(initdb_lsn), ctx)
2527 112 : .await?;
2528 : }
2529 80 : let layer_names = tline
2530 80 : .layers
2531 80 : .read()
2532 80 : .await
2533 80 : .layer_map()
2534 80 : .unwrap()
2535 80 : .iter_historic_layers()
2536 364 : .map(|layer| layer.layer_name())
2537 80 : .collect_vec();
2538 80 : if let Some(err) = check_valid_layermap(&layer_names) {
2539 0 : bail!("invalid layermap: {err}");
2540 80 : }
2541 80 : Ok(tline)
2542 80 : }
2543 :
2544 : /// Create a new timeline.
2545 : ///
2546 : /// Returns the new timeline ID and reference to its Timeline object.
2547 : ///
2548 : /// If the caller specified the timeline ID to use (`new_timeline_id`), and timeline with
2549 : /// the same timeline ID already exists, returns CreateTimelineError::AlreadyExists.
2550 : #[allow(clippy::too_many_arguments)]
2551 0 : pub(crate) async fn create_timeline(
2552 0 : self: &Arc<Tenant>,
2553 0 : params: CreateTimelineParams,
2554 0 : broker_client: storage_broker::BrokerClientChannel,
2555 0 : ctx: &RequestContext,
2556 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
2557 0 : if !self.is_active() {
2558 0 : if matches!(self.current_state(), TenantState::Stopping { .. }) {
2559 0 : return Err(CreateTimelineError::ShuttingDown);
2560 : } else {
2561 0 : return Err(CreateTimelineError::Other(anyhow::anyhow!(
2562 0 : "Cannot create timelines on inactive tenant"
2563 0 : )));
2564 : }
2565 0 : }
2566 :
2567 0 : let _gate = self
2568 0 : .gate
2569 0 : .enter()
2570 0 : .map_err(|_| CreateTimelineError::ShuttingDown)?;
2571 :
2572 0 : let result: CreateTimelineResult = match params {
2573 : CreateTimelineParams::Bootstrap(CreateTimelineParamsBootstrap {
2574 0 : new_timeline_id,
2575 0 : existing_initdb_timeline_id,
2576 0 : pg_version,
2577 0 : }) => {
2578 0 : self.bootstrap_timeline(
2579 0 : new_timeline_id,
2580 0 : pg_version,
2581 0 : existing_initdb_timeline_id,
2582 0 : ctx,
2583 0 : )
2584 0 : .await?
2585 : }
2586 : CreateTimelineParams::Branch(CreateTimelineParamsBranch {
2587 0 : new_timeline_id,
2588 0 : ancestor_timeline_id,
2589 0 : mut ancestor_start_lsn,
2590 : }) => {
2591 0 : let ancestor_timeline = self
2592 0 : .get_timeline(ancestor_timeline_id, false)
2593 0 : .context("Cannot branch off the timeline that's not present in pageserver")?;
2594 :
2595 : // instead of waiting around, just deny the request because ancestor is not yet
2596 : // ready for other purposes either.
2597 0 : if !ancestor_timeline.is_active() {
2598 0 : return Err(CreateTimelineError::AncestorNotActive);
2599 0 : }
2600 0 :
2601 0 : if ancestor_timeline.is_archived() == Some(true) {
2602 0 : info!("tried to branch archived timeline");
2603 0 : return Err(CreateTimelineError::AncestorArchived);
2604 0 : }
2605 :
2606 0 : if let Some(lsn) = ancestor_start_lsn.as_mut() {
2607 0 : *lsn = lsn.align();
2608 0 :
2609 0 : let ancestor_ancestor_lsn = ancestor_timeline.get_ancestor_lsn();
2610 0 : if ancestor_ancestor_lsn > *lsn {
2611 : // can we safely just branch from the ancestor instead?
2612 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
2613 0 : "invalid start lsn {} for ancestor timeline {}: less than timeline ancestor lsn {}",
2614 0 : lsn,
2615 0 : ancestor_timeline_id,
2616 0 : ancestor_ancestor_lsn,
2617 0 : )));
2618 0 : }
2619 0 :
2620 0 : // Wait for the WAL to arrive and be processed on the parent branch up
2621 0 : // to the requested branch point. The repository code itself doesn't
2622 0 : // require it, but if we start to receive WAL on the new timeline,
2623 0 : // decoding the new WAL might need to look up previous pages, relation
2624 0 : // sizes etc. and that would get confused if the previous page versions
2625 0 : // are not in the repository yet.
2626 0 : ancestor_timeline
2627 0 : .wait_lsn(
2628 0 : *lsn,
2629 0 : timeline::WaitLsnWaiter::Tenant,
2630 0 : timeline::WaitLsnTimeout::Default,
2631 0 : ctx,
2632 0 : )
2633 0 : .await
2634 0 : .map_err(|e| match e {
2635 0 : e @ (WaitLsnError::Timeout(_) | WaitLsnError::BadState { .. }) => {
2636 0 : CreateTimelineError::AncestorLsn(anyhow::anyhow!(e))
2637 : }
2638 0 : WaitLsnError::Shutdown => CreateTimelineError::ShuttingDown,
2639 0 : })?;
2640 0 : }
2641 :
2642 0 : self.branch_timeline(&ancestor_timeline, new_timeline_id, ancestor_start_lsn, ctx)
2643 0 : .await?
2644 : }
2645 0 : CreateTimelineParams::ImportPgdata(params) => {
2646 0 : self.create_timeline_import_pgdata(
2647 0 : params,
2648 0 : ActivateTimelineArgs::Yes {
2649 0 : broker_client: broker_client.clone(),
2650 0 : },
2651 0 : ctx,
2652 0 : )
2653 0 : .await?
2654 : }
2655 : };
2656 :
2657 : // At this point we have dropped our guard on [`Self::timelines_creating`], and
2658 : // the timeline is visible in [`Self::timelines`], but it is _not_ durable yet. We must
2659 : // not send a success to the caller until it is. The same applies to idempotent retries.
2660 : //
2661 : // TODO: the timeline is already visible in [`Self::timelines`]; a caller could incorrectly
2662 : // assume that, because they can see the timeline via API, that the creation is done and
2663 : // that it is durable. Ideally, we would keep the timeline hidden (in [`Self::timelines_creating`])
2664 : // until it is durable, e.g., by extending the time we hold the creation guard. This also
2665 : // interacts with UninitializedTimeline and is generally a bit tricky.
2666 : //
2667 : // To re-emphasize: the only correct way to create a timeline is to repeat calling the
2668 : // creation API until it returns success. Only then is durability guaranteed.
2669 0 : info!(creation_result=%result.discriminant(), "waiting for timeline to be durable");
2670 0 : result
2671 0 : .timeline()
2672 0 : .remote_client
2673 0 : .wait_completion()
2674 0 : .await
2675 0 : .map_err(|e| match e {
2676 : WaitCompletionError::NotInitialized(
2677 0 : e, // If the queue is already stopped, it's a shutdown error.
2678 0 : ) if e.is_stopping() => CreateTimelineError::ShuttingDown,
2679 : WaitCompletionError::NotInitialized(_) => {
2680 : // This is a bug: we should never try to wait for uploads before initializing the timeline
2681 0 : debug_assert!(false);
2682 0 : CreateTimelineError::Other(anyhow::anyhow!("timeline not initialized"))
2683 : }
2684 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2685 0 : CreateTimelineError::ShuttingDown
2686 : }
2687 0 : })?;
2688 :
2689 : // The creating task is responsible for activating the timeline.
2690 : // We do this after `wait_completion()` so that we don't spin up tasks that start
2691 : // doing stuff before the IndexPart is durable in S3, which is done by the previous section.
2692 0 : let activated_timeline = match result {
2693 0 : CreateTimelineResult::Created(timeline) => {
2694 0 : timeline.activate(self.clone(), broker_client, None, ctx);
2695 0 : timeline
2696 : }
2697 0 : CreateTimelineResult::Idempotent(timeline) => {
2698 0 : info!(
2699 0 : "request was deemed idempotent, activation will be done by the creating task"
2700 : );
2701 0 : timeline
2702 : }
2703 0 : CreateTimelineResult::ImportSpawned(timeline) => {
2704 0 : info!("import task spawned, timeline will become visible and activated once the import is done");
2705 0 : timeline
2706 : }
2707 : };
2708 :
2709 0 : Ok(activated_timeline)
2710 0 : }
2711 :
2712 : /// The returned [`Arc<Timeline>`] is NOT in the [`Tenant::timelines`] map until the import
2713 : /// completes in the background. A DIFFERENT [`Arc<Timeline>`] will be inserted into the
2714 : /// [`Tenant::timelines`] map when the import completes.
2715 : /// We only return an [`Arc<Timeline>`] here so the API handler can create a [`pageserver_api::models::TimelineInfo`]
2716 : /// for the response.
2717 0 : async fn create_timeline_import_pgdata(
2718 0 : self: &Arc<Tenant>,
2719 0 : params: CreateTimelineParamsImportPgdata,
2720 0 : activate: ActivateTimelineArgs,
2721 0 : ctx: &RequestContext,
2722 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
2723 0 : let CreateTimelineParamsImportPgdata {
2724 0 : new_timeline_id,
2725 0 : location,
2726 0 : idempotency_key,
2727 0 : } = params;
2728 0 :
2729 0 : let started_at = chrono::Utc::now().naive_utc();
2730 :
2731 : //
2732 : // There's probably a simpler way to upload an index part, but, remote_timeline_client
2733 : // is the canonical way we do it.
2734 : // - create an empty timeline in-memory
2735 : // - use its remote_timeline_client to do the upload
2736 : // - dispose of the uninit timeline
2737 : // - keep the creation guard alive
2738 :
2739 0 : let timeline_create_guard = match self
2740 0 : .start_creating_timeline(
2741 0 : new_timeline_id,
2742 0 : CreateTimelineIdempotency::ImportPgdata(CreatingTimelineIdempotencyImportPgdata {
2743 0 : idempotency_key: idempotency_key.clone(),
2744 0 : }),
2745 0 : )
2746 0 : .await?
2747 : {
2748 0 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2749 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
2750 0 : return Ok(CreateTimelineResult::Idempotent(timeline))
2751 : }
2752 : };
2753 :
2754 0 : let mut uninit_timeline = {
2755 0 : let this = &self;
2756 0 : let initdb_lsn = Lsn(0);
2757 0 : let _ctx = ctx;
2758 0 : async move {
2759 0 : let new_metadata = TimelineMetadata::new(
2760 0 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2761 0 : // make it valid, before calling finish_creation()
2762 0 : Lsn(0),
2763 0 : None,
2764 0 : None,
2765 0 : Lsn(0),
2766 0 : initdb_lsn,
2767 0 : initdb_lsn,
2768 0 : 15,
2769 0 : );
2770 0 : this.prepare_new_timeline(
2771 0 : new_timeline_id,
2772 0 : &new_metadata,
2773 0 : timeline_create_guard,
2774 0 : initdb_lsn,
2775 0 : None,
2776 0 : )
2777 0 : .await
2778 0 : }
2779 0 : }
2780 0 : .await?;
2781 :
2782 0 : let in_progress = import_pgdata::index_part_format::InProgress {
2783 0 : idempotency_key,
2784 0 : location,
2785 0 : started_at,
2786 0 : };
2787 0 : let index_part = import_pgdata::index_part_format::Root::V1(
2788 0 : import_pgdata::index_part_format::V1::InProgress(in_progress),
2789 0 : );
2790 0 : uninit_timeline
2791 0 : .raw_timeline()
2792 0 : .unwrap()
2793 0 : .remote_client
2794 0 : .schedule_index_upload_for_import_pgdata_state_update(Some(index_part.clone()))?;
2795 :
2796 : // wait_completion happens in caller
2797 :
2798 0 : let (timeline, timeline_create_guard) = uninit_timeline.finish_creation_myself();
2799 0 :
2800 0 : tokio::spawn(self.clone().create_timeline_import_pgdata_task(
2801 0 : timeline.clone(),
2802 0 : index_part,
2803 0 : activate,
2804 0 : timeline_create_guard,
2805 0 : ));
2806 0 :
2807 0 : // NB: the timeline doesn't exist in self.timelines at this point
2808 0 : Ok(CreateTimelineResult::ImportSpawned(timeline))
2809 0 : }
2810 :
2811 : #[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))]
2812 : async fn create_timeline_import_pgdata_task(
2813 : self: Arc<Tenant>,
2814 : timeline: Arc<Timeline>,
2815 : index_part: import_pgdata::index_part_format::Root,
2816 : activate: ActivateTimelineArgs,
2817 : timeline_create_guard: TimelineCreateGuard,
2818 : ) {
2819 : debug_assert_current_span_has_tenant_and_timeline_id();
2820 : info!("starting");
2821 : scopeguard::defer! {info!("exiting")};
2822 :
2823 : let res = self
2824 : .create_timeline_import_pgdata_task_impl(
2825 : timeline,
2826 : index_part,
2827 : activate,
2828 : timeline_create_guard,
2829 : )
2830 : .await;
2831 : if let Err(err) = &res {
2832 : error!(?err, "task failed");
2833 : // TODO sleep & retry, sensitive to tenant shutdown
2834 : // TODO: allow timeline deletion requests => should cancel the task
2835 : }
2836 : }
2837 :
2838 0 : async fn create_timeline_import_pgdata_task_impl(
2839 0 : self: Arc<Tenant>,
2840 0 : timeline: Arc<Timeline>,
2841 0 : index_part: import_pgdata::index_part_format::Root,
2842 0 : activate: ActivateTimelineArgs,
2843 0 : timeline_create_guard: TimelineCreateGuard,
2844 0 : ) -> Result<(), anyhow::Error> {
2845 0 : let ctx = RequestContext::new(TaskKind::ImportPgdata, DownloadBehavior::Warn);
2846 0 :
2847 0 : info!("importing pgdata");
2848 0 : import_pgdata::doit(&timeline, index_part, &ctx, self.cancel.clone())
2849 0 : .await
2850 0 : .context("import")?;
2851 0 : info!("import done");
2852 :
2853 : //
2854 : // Reload timeline from remote.
2855 : // This proves that the remote state is attachable, and it reuses the code.
2856 : //
2857 : // TODO: think about whether this is safe to do with concurrent Tenant::shutdown.
2858 : // timeline_create_guard hols the tenant gate open, so, shutdown cannot _complete_ until we exit.
2859 : // But our activate() call might launch new background tasks after Tenant::shutdown
2860 : // already went past shutting down the Tenant::timelines, which this timeline here is no part of.
2861 : // I think the same problem exists with the bootstrap & branch mgmt API tasks (tenant shutting
2862 : // down while bootstrapping/branching + activating), but, the race condition is much more likely
2863 : // to manifest because of the long runtime of this import task.
2864 :
2865 : // in theory this shouldn't even .await anything except for coop yield
2866 0 : info!("shutting down timeline");
2867 0 : timeline.shutdown(ShutdownMode::Hard).await;
2868 0 : info!("timeline shut down, reloading from remote");
2869 : // TODO: we can't do the following check because create_timeline_import_pgdata must return an Arc<Timeline>
2870 : // let Some(timeline) = Arc::into_inner(timeline) else {
2871 : // anyhow::bail!("implementation error: timeline that we shut down was still referenced from somewhere");
2872 : // };
2873 0 : let timeline_id = timeline.timeline_id;
2874 0 :
2875 0 : // load from object storage like Tenant::attach does
2876 0 : let resources = self.build_timeline_resources(timeline_id);
2877 0 : let index_part = resources
2878 0 : .remote_client
2879 0 : .download_index_file(&self.cancel)
2880 0 : .await?;
2881 0 : let index_part = match index_part {
2882 : MaybeDeletedIndexPart::Deleted(_) => {
2883 : // likely concurrent delete call, cplane should prevent this
2884 0 : anyhow::bail!("index part says deleted but we are not done creating yet, this should not happen but")
2885 : }
2886 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
2887 0 : };
2888 0 : let metadata = index_part.metadata.clone();
2889 0 : self
2890 0 : .load_remote_timeline(timeline_id, index_part, metadata, None, resources, LoadTimelineCause::ImportPgdata{
2891 0 : create_guard: timeline_create_guard, activate, }, &ctx)
2892 0 : .await?
2893 0 : .ready_to_activate()
2894 0 : .context("implementation error: reloaded timeline still needs import after import reported success")?;
2895 :
2896 0 : anyhow::Ok(())
2897 0 : }
2898 :
2899 0 : pub(crate) async fn delete_timeline(
2900 0 : self: Arc<Self>,
2901 0 : timeline_id: TimelineId,
2902 0 : ) -> Result<(), DeleteTimelineError> {
2903 0 : DeleteTimelineFlow::run(&self, timeline_id).await?;
2904 :
2905 0 : Ok(())
2906 0 : }
2907 :
2908 : /// perform one garbage collection iteration, removing old data files from disk.
2909 : /// this function is periodically called by gc task.
2910 : /// also it can be explicitly requested through page server api 'do_gc' command.
2911 : ///
2912 : /// `target_timeline_id` specifies the timeline to GC, or None for all.
2913 : ///
2914 : /// The `horizon` an `pitr` parameters determine how much WAL history needs to be retained.
2915 : /// Also known as the retention period, or the GC cutoff point. `horizon` specifies
2916 : /// the amount of history, as LSN difference from current latest LSN on each timeline.
2917 : /// `pitr` specifies the same as a time difference from the current time. The effective
2918 : /// GC cutoff point is determined conservatively by either `horizon` and `pitr`, whichever
2919 : /// requires more history to be retained.
2920 : //
2921 1508 : pub(crate) async fn gc_iteration(
2922 1508 : &self,
2923 1508 : target_timeline_id: Option<TimelineId>,
2924 1508 : horizon: u64,
2925 1508 : pitr: Duration,
2926 1508 : cancel: &CancellationToken,
2927 1508 : ctx: &RequestContext,
2928 1508 : ) -> Result<GcResult, GcError> {
2929 1508 : // Don't start doing work during shutdown
2930 1508 : if let TenantState::Stopping { .. } = self.current_state() {
2931 0 : return Ok(GcResult::default());
2932 1508 : }
2933 1508 :
2934 1508 : // there is a global allowed_error for this
2935 1508 : if !self.is_active() {
2936 0 : return Err(GcError::NotActive);
2937 1508 : }
2938 1508 :
2939 1508 : {
2940 1508 : let conf = self.tenant_conf.load();
2941 1508 :
2942 1508 : // If we may not delete layers, then simply skip GC. Even though a tenant
2943 1508 : // in AttachedMulti state could do GC and just enqueue the blocked deletions,
2944 1508 : // the only advantage to doing it is to perhaps shrink the LayerMap metadata
2945 1508 : // a bit sooner than we would achieve by waiting for AttachedSingle status.
2946 1508 : if !conf.location.may_delete_layers_hint() {
2947 0 : info!("Skipping GC in location state {:?}", conf.location);
2948 0 : return Ok(GcResult::default());
2949 1508 : }
2950 1508 :
2951 1508 : if conf.is_gc_blocked_by_lsn_lease_deadline() {
2952 1500 : info!("Skipping GC because lsn lease deadline is not reached");
2953 1500 : return Ok(GcResult::default());
2954 8 : }
2955 : }
2956 :
2957 8 : let _guard = match self.gc_block.start().await {
2958 8 : Ok(guard) => guard,
2959 0 : Err(reasons) => {
2960 0 : info!("Skipping GC: {reasons}");
2961 0 : return Ok(GcResult::default());
2962 : }
2963 : };
2964 :
2965 8 : self.gc_iteration_internal(target_timeline_id, horizon, pitr, cancel, ctx)
2966 8 : .await
2967 1508 : }
2968 :
2969 : /// Performs one compaction iteration. Called periodically from the compaction loop. Returns
2970 : /// whether another compaction is needed, if we still have pending work or if we yield for
2971 : /// immediate L0 compaction.
2972 : ///
2973 : /// Compaction can also be explicitly requested for a timeline via the HTTP API.
2974 0 : async fn compaction_iteration(
2975 0 : self: &Arc<Self>,
2976 0 : cancel: &CancellationToken,
2977 0 : ctx: &RequestContext,
2978 0 : ) -> Result<CompactionOutcome, CompactionError> {
2979 0 : // Don't compact inactive tenants.
2980 0 : if !self.is_active() {
2981 0 : return Ok(CompactionOutcome::Skipped);
2982 0 : }
2983 0 :
2984 0 : // Don't compact tenants that can't upload layers. We don't check `may_delete_layers_hint`,
2985 0 : // since we need to compact L0 even in AttachedMulti to bound read amplification.
2986 0 : let location = self.tenant_conf.load().location;
2987 0 : if !location.may_upload_layers_hint() {
2988 0 : info!("skipping compaction in location state {location:?}");
2989 0 : return Ok(CompactionOutcome::Skipped);
2990 0 : }
2991 0 :
2992 0 : // Don't compact if the circuit breaker is tripped.
2993 0 : if self.compaction_circuit_breaker.lock().unwrap().is_broken() {
2994 0 : info!("skipping compaction due to previous failures");
2995 0 : return Ok(CompactionOutcome::Skipped);
2996 0 : }
2997 0 :
2998 0 : // Collect all timelines to compact, along with offload instructions and L0 counts.
2999 0 : let mut compact: Vec<Arc<Timeline>> = Vec::new();
3000 0 : let mut offload: HashSet<TimelineId> = HashSet::new();
3001 0 : let mut l0_counts: HashMap<TimelineId, usize> = HashMap::new();
3002 0 :
3003 0 : {
3004 0 : let offload_enabled = self.get_timeline_offloading_enabled();
3005 0 : let timelines = self.timelines.lock().unwrap();
3006 0 : for (&timeline_id, timeline) in timelines.iter() {
3007 : // Skip inactive timelines.
3008 0 : if !timeline.is_active() {
3009 0 : continue;
3010 0 : }
3011 0 :
3012 0 : // Schedule the timeline for compaction.
3013 0 : compact.push(timeline.clone());
3014 :
3015 : // Schedule the timeline for offloading if eligible.
3016 0 : let can_offload = offload_enabled
3017 0 : && timeline.can_offload().0
3018 0 : && !timelines
3019 0 : .iter()
3020 0 : .any(|(_, tli)| tli.get_ancestor_timeline_id() == Some(timeline_id));
3021 0 : if can_offload {
3022 0 : offload.insert(timeline_id);
3023 0 : }
3024 : }
3025 : } // release timelines lock
3026 :
3027 0 : for timeline in &compact {
3028 : // Collect L0 counts. Can't await while holding lock above.
3029 0 : if let Ok(lm) = timeline.layers.read().await.layer_map() {
3030 0 : l0_counts.insert(timeline.timeline_id, lm.level0_deltas().len());
3031 0 : }
3032 : }
3033 :
3034 : // Pass 1: L0 compaction across all timelines, in order of L0 count. We prioritize this to
3035 : // bound read amplification.
3036 : //
3037 : // TODO: this may spin on one or more ingest-heavy timelines, starving out image/GC
3038 : // compaction and offloading. We leave that as a potential problem to solve later. Consider
3039 : // splitting L0 and image/GC compaction to separate background jobs.
3040 0 : if self.get_compaction_l0_first() {
3041 0 : let compaction_threshold = self.get_compaction_threshold();
3042 0 : let compact_l0 = compact
3043 0 : .iter()
3044 0 : .map(|tli| (tli, l0_counts.get(&tli.timeline_id).copied().unwrap_or(0)))
3045 0 : .filter(|&(_, l0)| l0 >= compaction_threshold)
3046 0 : .sorted_by_key(|&(_, l0)| l0)
3047 0 : .rev()
3048 0 : .map(|(tli, _)| tli.clone())
3049 0 : .collect_vec();
3050 0 :
3051 0 : let mut has_pending_l0 = false;
3052 0 : for timeline in compact_l0 {
3053 0 : let outcome = timeline
3054 0 : .compact(cancel, CompactFlags::OnlyL0Compaction.into(), ctx)
3055 0 : .instrument(info_span!("compact_timeline", timeline_id = %timeline.timeline_id))
3056 0 : .await
3057 0 : .inspect_err(|err| self.maybe_trip_compaction_breaker(err))?;
3058 0 : match outcome {
3059 0 : CompactionOutcome::Done => {}
3060 0 : CompactionOutcome::Skipped => {}
3061 0 : CompactionOutcome::Pending => has_pending_l0 = true,
3062 0 : CompactionOutcome::YieldForL0 => has_pending_l0 = true,
3063 : }
3064 : }
3065 0 : if has_pending_l0 {
3066 0 : return Ok(CompactionOutcome::YieldForL0); // do another pass
3067 0 : }
3068 0 : }
3069 :
3070 : // Pass 2: image compaction and timeline offloading. If any timelines have accumulated
3071 : // more L0 layers, they may also be compacted here.
3072 : //
3073 : // NB: image compaction may yield if there is pending L0 compaction.
3074 : //
3075 : // TODO: it will only yield if there is pending L0 compaction on the same timeline. If a
3076 : // different timeline needs compaction, it won't. It should check `l0_compaction_trigger`.
3077 : // We leave this for a later PR.
3078 : //
3079 : // TODO: consider ordering timelines by some priority, e.g. time since last full compaction,
3080 : // amount of L1 delta debt or garbage, offload-eligible timelines first, etc.
3081 0 : let mut has_pending = false;
3082 0 : for timeline in compact {
3083 0 : if !timeline.is_active() {
3084 0 : continue;
3085 0 : }
3086 :
3087 0 : let mut outcome = timeline
3088 0 : .compact(cancel, EnumSet::default(), ctx)
3089 0 : .instrument(info_span!("compact_timeline", timeline_id = %timeline.timeline_id))
3090 0 : .await
3091 0 : .inspect_err(|err| self.maybe_trip_compaction_breaker(err))?;
3092 :
3093 : // If we're done compacting, check the scheduled GC compaction queue for more work.
3094 0 : if outcome == CompactionOutcome::Done {
3095 0 : let queue = self
3096 0 : .scheduled_compaction_tasks
3097 0 : .lock()
3098 0 : .unwrap()
3099 0 : .get(&timeline.timeline_id)
3100 0 : .cloned();
3101 0 : if let Some(queue) = queue {
3102 0 : outcome = queue
3103 0 : .iteration(cancel, ctx, &self.gc_block, &timeline)
3104 0 : .instrument(
3105 0 : info_span!("gc_compact_timeline", timeline_id = %timeline.timeline_id),
3106 : )
3107 0 : .await?;
3108 0 : }
3109 0 : }
3110 :
3111 : // If we're done compacting, offload the timeline if requested.
3112 0 : if outcome == CompactionOutcome::Done && offload.contains(&timeline.timeline_id) {
3113 0 : pausable_failpoint!("before-timeline-auto-offload");
3114 0 : offload_timeline(self, &timeline)
3115 0 : .instrument(info_span!("offload_timeline", timeline_id = %timeline.timeline_id))
3116 0 : .await
3117 0 : .or_else(|err| match err {
3118 : // Ignore this, we likely raced with unarchival.
3119 0 : OffloadError::NotArchived => Ok(()),
3120 0 : err => Err(err),
3121 0 : })?;
3122 0 : }
3123 :
3124 0 : match outcome {
3125 0 : CompactionOutcome::Done => {}
3126 0 : CompactionOutcome::Skipped => {}
3127 0 : CompactionOutcome::Pending => has_pending = true,
3128 : // This mostly makes sense when the L0-only pass above is enabled, since there's
3129 : // otherwise no guarantee that we'll start with the timeline that has high L0.
3130 0 : CompactionOutcome::YieldForL0 => return Ok(CompactionOutcome::YieldForL0),
3131 : }
3132 : }
3133 :
3134 : // Success! Untrip the breaker if necessary.
3135 0 : self.compaction_circuit_breaker
3136 0 : .lock()
3137 0 : .unwrap()
3138 0 : .success(&CIRCUIT_BREAKERS_UNBROKEN);
3139 0 :
3140 0 : match has_pending {
3141 0 : true => Ok(CompactionOutcome::Pending),
3142 0 : false => Ok(CompactionOutcome::Done),
3143 : }
3144 0 : }
3145 :
3146 : /// Trips the compaction circuit breaker if appropriate.
3147 0 : pub(crate) fn maybe_trip_compaction_breaker(&self, err: &CompactionError) {
3148 0 : match err {
3149 0 : CompactionError::ShuttingDown => (),
3150 : // Offload failures don't trip the circuit breaker, since they're cheap to retry and
3151 : // shouldn't block compaction.
3152 0 : CompactionError::Offload(_) => {}
3153 0 : CompactionError::CollectKeySpaceError(err) => {
3154 0 : self.compaction_circuit_breaker
3155 0 : .lock()
3156 0 : .unwrap()
3157 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, err);
3158 0 : }
3159 0 : CompactionError::Other(err) => {
3160 0 : self.compaction_circuit_breaker
3161 0 : .lock()
3162 0 : .unwrap()
3163 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, err);
3164 0 : }
3165 : }
3166 0 : }
3167 :
3168 : /// Cancel scheduled compaction tasks
3169 0 : pub(crate) fn cancel_scheduled_compaction(&self, timeline_id: TimelineId) {
3170 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3171 0 : if let Some(q) = guard.get_mut(&timeline_id) {
3172 0 : q.cancel_scheduled();
3173 0 : }
3174 0 : }
3175 :
3176 0 : pub(crate) fn get_scheduled_compaction_tasks(
3177 0 : &self,
3178 0 : timeline_id: TimelineId,
3179 0 : ) -> Vec<CompactInfoResponse> {
3180 0 : let res = {
3181 0 : let guard = self.scheduled_compaction_tasks.lock().unwrap();
3182 0 : guard.get(&timeline_id).map(|q| q.remaining_jobs())
3183 : };
3184 0 : let Some((running, remaining)) = res else {
3185 0 : return Vec::new();
3186 : };
3187 0 : let mut result = Vec::new();
3188 0 : if let Some((id, running)) = running {
3189 0 : result.extend(running.into_compact_info_resp(id, true));
3190 0 : }
3191 0 : for (id, job) in remaining {
3192 0 : result.extend(job.into_compact_info_resp(id, false));
3193 0 : }
3194 0 : result
3195 0 : }
3196 :
3197 : /// Schedule a compaction task for a timeline.
3198 0 : pub(crate) async fn schedule_compaction(
3199 0 : &self,
3200 0 : timeline_id: TimelineId,
3201 0 : options: CompactOptions,
3202 0 : ) -> anyhow::Result<tokio::sync::oneshot::Receiver<()>> {
3203 0 : let (tx, rx) = tokio::sync::oneshot::channel();
3204 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3205 0 : let q = guard
3206 0 : .entry(timeline_id)
3207 0 : .or_insert_with(|| Arc::new(GcCompactionQueue::new()));
3208 0 : q.schedule_manual_compaction(options, Some(tx));
3209 0 : Ok(rx)
3210 0 : }
3211 :
3212 : /// Performs periodic housekeeping, via the tenant housekeeping background task.
3213 0 : async fn housekeeping(&self) {
3214 0 : // Call through to all timelines to freeze ephemeral layers as needed. This usually happens
3215 0 : // during ingest, but we don't want idle timelines to hold open layers for too long.
3216 0 : let timelines = self
3217 0 : .timelines
3218 0 : .lock()
3219 0 : .unwrap()
3220 0 : .values()
3221 0 : .filter(|tli| tli.is_active())
3222 0 : .cloned()
3223 0 : .collect_vec();
3224 :
3225 0 : for timeline in timelines {
3226 0 : timeline.maybe_freeze_ephemeral_layer().await;
3227 : }
3228 :
3229 : // Shut down walredo if idle.
3230 : const WALREDO_IDLE_TIMEOUT: Duration = Duration::from_secs(180);
3231 0 : if let Some(ref walredo_mgr) = self.walredo_mgr {
3232 0 : walredo_mgr.maybe_quiesce(WALREDO_IDLE_TIMEOUT);
3233 0 : }
3234 0 : }
3235 :
3236 0 : pub fn timeline_has_no_attached_children(&self, timeline_id: TimelineId) -> bool {
3237 0 : let timelines = self.timelines.lock().unwrap();
3238 0 : !timelines
3239 0 : .iter()
3240 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(timeline_id))
3241 0 : }
3242 :
3243 3468 : pub fn current_state(&self) -> TenantState {
3244 3468 : self.state.borrow().clone()
3245 3468 : }
3246 :
3247 1944 : pub fn is_active(&self) -> bool {
3248 1944 : self.current_state() == TenantState::Active
3249 1944 : }
3250 :
3251 0 : pub fn generation(&self) -> Generation {
3252 0 : self.generation
3253 0 : }
3254 :
3255 0 : pub(crate) fn wal_redo_manager_status(&self) -> Option<WalRedoManagerStatus> {
3256 0 : self.walredo_mgr.as_ref().and_then(|mgr| mgr.status())
3257 0 : }
3258 :
3259 : /// Changes tenant status to active, unless shutdown was already requested.
3260 : ///
3261 : /// `background_jobs_can_start` is an optional barrier set to a value during pageserver startup
3262 : /// to delay background jobs. Background jobs can be started right away when None is given.
3263 0 : fn activate(
3264 0 : self: &Arc<Self>,
3265 0 : broker_client: BrokerClientChannel,
3266 0 : background_jobs_can_start: Option<&completion::Barrier>,
3267 0 : ctx: &RequestContext,
3268 0 : ) {
3269 0 : span::debug_assert_current_span_has_tenant_id();
3270 0 :
3271 0 : let mut activating = false;
3272 0 : self.state.send_modify(|current_state| {
3273 : use pageserver_api::models::ActivatingFrom;
3274 0 : match &*current_state {
3275 : TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
3276 0 : panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {state:?}", state = current_state);
3277 : }
3278 0 : TenantState::Attaching => {
3279 0 : *current_state = TenantState::Activating(ActivatingFrom::Attaching);
3280 0 : }
3281 0 : }
3282 0 : debug!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), "Activating tenant");
3283 0 : activating = true;
3284 0 : // Continue outside the closure. We need to grab timelines.lock()
3285 0 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
3286 0 : });
3287 0 :
3288 0 : if activating {
3289 0 : let timelines_accessor = self.timelines.lock().unwrap();
3290 0 : let timelines_offloaded_accessor = self.timelines_offloaded.lock().unwrap();
3291 0 : let timelines_to_activate = timelines_accessor
3292 0 : .values()
3293 0 : .filter(|timeline| !(timeline.is_broken() || timeline.is_stopping()));
3294 0 :
3295 0 : // Before activation, populate each Timeline's GcInfo with information about its children
3296 0 : self.initialize_gc_info(&timelines_accessor, &timelines_offloaded_accessor, None);
3297 0 :
3298 0 : // Spawn gc and compaction loops. The loops will shut themselves
3299 0 : // down when they notice that the tenant is inactive.
3300 0 : tasks::start_background_loops(self, background_jobs_can_start);
3301 0 :
3302 0 : let mut activated_timelines = 0;
3303 :
3304 0 : for timeline in timelines_to_activate {
3305 0 : timeline.activate(
3306 0 : self.clone(),
3307 0 : broker_client.clone(),
3308 0 : background_jobs_can_start,
3309 0 : ctx,
3310 0 : );
3311 0 : activated_timelines += 1;
3312 0 : }
3313 :
3314 0 : self.state.send_modify(move |current_state| {
3315 0 : assert!(
3316 0 : matches!(current_state, TenantState::Activating(_)),
3317 0 : "set_stopping and set_broken wait for us to leave Activating state",
3318 : );
3319 0 : *current_state = TenantState::Active;
3320 0 :
3321 0 : let elapsed = self.constructed_at.elapsed();
3322 0 : let total_timelines = timelines_accessor.len();
3323 0 :
3324 0 : // log a lot of stuff, because some tenants sometimes suffer from user-visible
3325 0 : // times to activate. see https://github.com/neondatabase/neon/issues/4025
3326 0 : info!(
3327 0 : since_creation_millis = elapsed.as_millis(),
3328 0 : tenant_id = %self.tenant_shard_id.tenant_id,
3329 0 : shard_id = %self.tenant_shard_id.shard_slug(),
3330 0 : activated_timelines,
3331 0 : total_timelines,
3332 0 : post_state = <&'static str>::from(&*current_state),
3333 0 : "activation attempt finished"
3334 : );
3335 :
3336 0 : TENANT.activation.observe(elapsed.as_secs_f64());
3337 0 : });
3338 0 : }
3339 0 : }
3340 :
3341 : /// Shutdown the tenant and join all of the spawned tasks.
3342 : ///
3343 : /// The method caters for all use-cases:
3344 : /// - pageserver shutdown (freeze_and_flush == true)
3345 : /// - detach + ignore (freeze_and_flush == false)
3346 : ///
3347 : /// This will attempt to shutdown even if tenant is broken.
3348 : ///
3349 : /// `shutdown_progress` is a [`completion::Barrier`] for the shutdown initiated by this call.
3350 : /// If the tenant is already shutting down, we return a clone of the first shutdown call's
3351 : /// `Barrier` as an `Err`. This not-first caller can use the returned barrier to join with
3352 : /// the ongoing shutdown.
3353 12 : async fn shutdown(
3354 12 : &self,
3355 12 : shutdown_progress: completion::Barrier,
3356 12 : shutdown_mode: timeline::ShutdownMode,
3357 12 : ) -> Result<(), completion::Barrier> {
3358 12 : span::debug_assert_current_span_has_tenant_id();
3359 :
3360 : // Set tenant (and its timlines) to Stoppping state.
3361 : //
3362 : // Since we can only transition into Stopping state after activation is complete,
3363 : // run it in a JoinSet so all tenants have a chance to stop before we get SIGKILLed.
3364 : //
3365 : // Transitioning tenants to Stopping state has a couple of non-obvious side effects:
3366 : // 1. Lock out any new requests to the tenants.
3367 : // 2. Signal cancellation to WAL receivers (we wait on it below).
3368 : // 3. Signal cancellation for other tenant background loops.
3369 : // 4. ???
3370 : //
3371 : // The waiting for the cancellation is not done uniformly.
3372 : // We certainly wait for WAL receivers to shut down.
3373 : // That is necessary so that no new data comes in before the freeze_and_flush.
3374 : // But the tenant background loops are joined-on in our caller.
3375 : // It's mesed up.
3376 : // we just ignore the failure to stop
3377 :
3378 : // If we're still attaching, fire the cancellation token early to drop out: this
3379 : // will prevent us flushing, but ensures timely shutdown if some I/O during attach
3380 : // is very slow.
3381 12 : let shutdown_mode = if matches!(self.current_state(), TenantState::Attaching) {
3382 0 : self.cancel.cancel();
3383 0 :
3384 0 : // Having fired our cancellation token, do not try and flush timelines: their cancellation tokens
3385 0 : // are children of ours, so their flush loops will have shut down already
3386 0 : timeline::ShutdownMode::Hard
3387 : } else {
3388 12 : shutdown_mode
3389 : };
3390 :
3391 12 : match self.set_stopping(shutdown_progress, false, false).await {
3392 12 : Ok(()) => {}
3393 0 : Err(SetStoppingError::Broken) => {
3394 0 : // assume that this is acceptable
3395 0 : }
3396 0 : Err(SetStoppingError::AlreadyStopping(other)) => {
3397 0 : // give caller the option to wait for this this shutdown
3398 0 : info!("Tenant::shutdown: AlreadyStopping");
3399 0 : return Err(other);
3400 : }
3401 : };
3402 :
3403 12 : let mut js = tokio::task::JoinSet::new();
3404 12 : {
3405 12 : let timelines = self.timelines.lock().unwrap();
3406 12 : timelines.values().for_each(|timeline| {
3407 12 : let timeline = Arc::clone(timeline);
3408 12 : let timeline_id = timeline.timeline_id;
3409 12 : let span = tracing::info_span!("timeline_shutdown", %timeline_id, ?shutdown_mode);
3410 12 : js.spawn(async move { timeline.shutdown(shutdown_mode).instrument(span).await });
3411 12 : });
3412 12 : }
3413 12 : {
3414 12 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
3415 12 : timelines_offloaded.values().for_each(|timeline| {
3416 0 : timeline.defuse_for_tenant_drop();
3417 12 : });
3418 12 : }
3419 12 : // test_long_timeline_create_then_tenant_delete is leaning on this message
3420 12 : tracing::info!("Waiting for timelines...");
3421 24 : while let Some(res) = js.join_next().await {
3422 0 : match res {
3423 12 : Ok(()) => {}
3424 0 : Err(je) if je.is_cancelled() => unreachable!("no cancelling used"),
3425 0 : Err(je) if je.is_panic() => { /* logged already */ }
3426 0 : Err(je) => warn!("unexpected JoinError: {je:?}"),
3427 : }
3428 : }
3429 :
3430 12 : if let ShutdownMode::Reload = shutdown_mode {
3431 0 : tracing::info!("Flushing deletion queue");
3432 0 : if let Err(e) = self.deletion_queue_client.flush().await {
3433 0 : match e {
3434 0 : DeletionQueueError::ShuttingDown => {
3435 0 : // This is the only error we expect for now. In the future, if more error
3436 0 : // variants are added, we should handle them here.
3437 0 : }
3438 : }
3439 0 : }
3440 12 : }
3441 :
3442 : // We cancel the Tenant's cancellation token _after_ the timelines have all shut down. This permits
3443 : // them to continue to do work during their shutdown methods, e.g. flushing data.
3444 12 : tracing::debug!("Cancelling CancellationToken");
3445 12 : self.cancel.cancel();
3446 12 :
3447 12 : // shutdown all tenant and timeline tasks: gc, compaction, page service
3448 12 : // No new tasks will be started for this tenant because it's in `Stopping` state.
3449 12 : //
3450 12 : // this will additionally shutdown and await all timeline tasks.
3451 12 : tracing::debug!("Waiting for tasks...");
3452 12 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), None).await;
3453 :
3454 12 : if let Some(walredo_mgr) = self.walredo_mgr.as_ref() {
3455 12 : walredo_mgr.shutdown().await;
3456 0 : }
3457 :
3458 : // Wait for any in-flight operations to complete
3459 12 : self.gate.close().await;
3460 :
3461 12 : remove_tenant_metrics(&self.tenant_shard_id);
3462 12 :
3463 12 : Ok(())
3464 12 : }
3465 :
3466 : /// Change tenant status to Stopping, to mark that it is being shut down.
3467 : ///
3468 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
3469 : ///
3470 : /// This function is not cancel-safe!
3471 : ///
3472 : /// `allow_transition_from_loading` is needed for the special case of loading task deleting the tenant.
3473 : /// `allow_transition_from_attaching` is needed for the special case of attaching deleted tenant.
3474 12 : async fn set_stopping(
3475 12 : &self,
3476 12 : progress: completion::Barrier,
3477 12 : _allow_transition_from_loading: bool,
3478 12 : allow_transition_from_attaching: bool,
3479 12 : ) -> Result<(), SetStoppingError> {
3480 12 : let mut rx = self.state.subscribe();
3481 12 :
3482 12 : // cannot stop before we're done activating, so wait out until we're done activating
3483 12 : rx.wait_for(|state| match state {
3484 0 : TenantState::Attaching if allow_transition_from_attaching => true,
3485 : TenantState::Activating(_) | TenantState::Attaching => {
3486 0 : info!(
3487 0 : "waiting for {} to turn Active|Broken|Stopping",
3488 0 : <&'static str>::from(state)
3489 : );
3490 0 : false
3491 : }
3492 12 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3493 12 : })
3494 12 : .await
3495 12 : .expect("cannot drop self.state while on a &self method");
3496 12 :
3497 12 : // we now know we're done activating, let's see whether this task is the winner to transition into Stopping
3498 12 : let mut err = None;
3499 12 : let stopping = self.state.send_if_modified(|current_state| match current_state {
3500 : TenantState::Activating(_) => {
3501 0 : unreachable!("1we ensured above that we're done with activation, and, there is no re-activation")
3502 : }
3503 : TenantState::Attaching => {
3504 0 : if !allow_transition_from_attaching {
3505 0 : unreachable!("2we ensured above that we're done with activation, and, there is no re-activation")
3506 0 : };
3507 0 : *current_state = TenantState::Stopping { progress };
3508 0 : true
3509 : }
3510 : TenantState::Active => {
3511 : // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
3512 : // are created after the transition to Stopping. That's harmless, as the Timelines
3513 : // won't be accessible to anyone afterwards, because the Tenant is in Stopping state.
3514 12 : *current_state = TenantState::Stopping { progress };
3515 12 : // Continue stopping outside the closure. We need to grab timelines.lock()
3516 12 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
3517 12 : true
3518 : }
3519 0 : TenantState::Broken { reason, .. } => {
3520 0 : info!(
3521 0 : "Cannot set tenant to Stopping state, it is in Broken state due to: {reason}"
3522 : );
3523 0 : err = Some(SetStoppingError::Broken);
3524 0 : false
3525 : }
3526 0 : TenantState::Stopping { progress } => {
3527 0 : info!("Tenant is already in Stopping state");
3528 0 : err = Some(SetStoppingError::AlreadyStopping(progress.clone()));
3529 0 : false
3530 : }
3531 12 : });
3532 12 : match (stopping, err) {
3533 12 : (true, None) => {} // continue
3534 0 : (false, Some(err)) => return Err(err),
3535 0 : (true, Some(_)) => unreachable!(
3536 0 : "send_if_modified closure must error out if not transitioning to Stopping"
3537 0 : ),
3538 0 : (false, None) => unreachable!(
3539 0 : "send_if_modified closure must return true if transitioning to Stopping"
3540 0 : ),
3541 : }
3542 :
3543 12 : let timelines_accessor = self.timelines.lock().unwrap();
3544 12 : let not_broken_timelines = timelines_accessor
3545 12 : .values()
3546 12 : .filter(|timeline| !timeline.is_broken());
3547 24 : for timeline in not_broken_timelines {
3548 12 : timeline.set_state(TimelineState::Stopping);
3549 12 : }
3550 12 : Ok(())
3551 12 : }
3552 :
3553 : /// Method for tenant::mgr to transition us into Broken state in case of a late failure in
3554 : /// `remove_tenant_from_memory`
3555 : ///
3556 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
3557 : ///
3558 : /// In tests, we also use this to set tenants to Broken state on purpose.
3559 0 : pub(crate) async fn set_broken(&self, reason: String) {
3560 0 : let mut rx = self.state.subscribe();
3561 0 :
3562 0 : // The load & attach routines own the tenant state until it has reached `Active`.
3563 0 : // So, wait until it's done.
3564 0 : rx.wait_for(|state| match state {
3565 : TenantState::Activating(_) | TenantState::Attaching => {
3566 0 : info!(
3567 0 : "waiting for {} to turn Active|Broken|Stopping",
3568 0 : <&'static str>::from(state)
3569 : );
3570 0 : false
3571 : }
3572 0 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3573 0 : })
3574 0 : .await
3575 0 : .expect("cannot drop self.state while on a &self method");
3576 0 :
3577 0 : // we now know we're done activating, let's see whether this task is the winner to transition into Broken
3578 0 : self.set_broken_no_wait(reason)
3579 0 : }
3580 :
3581 0 : pub(crate) fn set_broken_no_wait(&self, reason: impl Display) {
3582 0 : let reason = reason.to_string();
3583 0 : self.state.send_modify(|current_state| {
3584 0 : match *current_state {
3585 : TenantState::Activating(_) | TenantState::Attaching => {
3586 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
3587 : }
3588 : TenantState::Active => {
3589 0 : if cfg!(feature = "testing") {
3590 0 : warn!("Changing Active tenant to Broken state, reason: {}", reason);
3591 0 : *current_state = TenantState::broken_from_reason(reason);
3592 : } else {
3593 0 : unreachable!("not allowed to call set_broken on Active tenants in non-testing builds")
3594 : }
3595 : }
3596 : TenantState::Broken { .. } => {
3597 0 : warn!("Tenant is already in Broken state");
3598 : }
3599 : // This is the only "expected" path, any other path is a bug.
3600 : TenantState::Stopping { .. } => {
3601 0 : warn!(
3602 0 : "Marking Stopping tenant as Broken state, reason: {}",
3603 : reason
3604 : );
3605 0 : *current_state = TenantState::broken_from_reason(reason);
3606 : }
3607 : }
3608 0 : });
3609 0 : }
3610 :
3611 0 : pub fn subscribe_for_state_updates(&self) -> watch::Receiver<TenantState> {
3612 0 : self.state.subscribe()
3613 0 : }
3614 :
3615 : /// The activate_now semaphore is initialized with zero units. As soon as
3616 : /// we add a unit, waiters will be able to acquire a unit and proceed.
3617 0 : pub(crate) fn activate_now(&self) {
3618 0 : self.activate_now_sem.add_permits(1);
3619 0 : }
3620 :
3621 0 : pub(crate) async fn wait_to_become_active(
3622 0 : &self,
3623 0 : timeout: Duration,
3624 0 : ) -> Result<(), GetActiveTenantError> {
3625 0 : let mut receiver = self.state.subscribe();
3626 : loop {
3627 0 : let current_state = receiver.borrow_and_update().clone();
3628 0 : match current_state {
3629 : TenantState::Attaching | TenantState::Activating(_) => {
3630 : // in these states, there's a chance that we can reach ::Active
3631 0 : self.activate_now();
3632 0 : match timeout_cancellable(timeout, &self.cancel, receiver.changed()).await {
3633 0 : Ok(r) => {
3634 0 : r.map_err(
3635 0 : |_e: tokio::sync::watch::error::RecvError|
3636 : // Tenant existed but was dropped: report it as non-existent
3637 0 : GetActiveTenantError::NotFound(GetTenantError::ShardNotFound(self.tenant_shard_id))
3638 0 : )?
3639 : }
3640 : Err(TimeoutCancellableError::Cancelled) => {
3641 0 : return Err(GetActiveTenantError::Cancelled);
3642 : }
3643 : Err(TimeoutCancellableError::Timeout) => {
3644 0 : return Err(GetActiveTenantError::WaitForActiveTimeout {
3645 0 : latest_state: Some(self.current_state()),
3646 0 : wait_time: timeout,
3647 0 : });
3648 : }
3649 : }
3650 : }
3651 : TenantState::Active { .. } => {
3652 0 : return Ok(());
3653 : }
3654 0 : TenantState::Broken { reason, .. } => {
3655 0 : // This is fatal, and reported distinctly from the general case of "will never be active" because
3656 0 : // it's logically a 500 to external API users (broken is always a bug).
3657 0 : return Err(GetActiveTenantError::Broken(reason));
3658 : }
3659 : TenantState::Stopping { .. } => {
3660 : // There's no chance the tenant can transition back into ::Active
3661 0 : return Err(GetActiveTenantError::WillNotBecomeActive(current_state));
3662 : }
3663 : }
3664 : }
3665 0 : }
3666 :
3667 0 : pub(crate) fn get_attach_mode(&self) -> AttachmentMode {
3668 0 : self.tenant_conf.load().location.attach_mode
3669 0 : }
3670 :
3671 : /// For API access: generate a LocationConfig equivalent to the one that would be used to
3672 : /// create a Tenant in the same state. Do not use this in hot paths: it's for relatively
3673 : /// rare external API calls, like a reconciliation at startup.
3674 0 : pub(crate) fn get_location_conf(&self) -> models::LocationConfig {
3675 0 : let conf = self.tenant_conf.load();
3676 :
3677 0 : let location_config_mode = match conf.location.attach_mode {
3678 0 : AttachmentMode::Single => models::LocationConfigMode::AttachedSingle,
3679 0 : AttachmentMode::Multi => models::LocationConfigMode::AttachedMulti,
3680 0 : AttachmentMode::Stale => models::LocationConfigMode::AttachedStale,
3681 : };
3682 :
3683 : // We have a pageserver TenantConf, we need the API-facing TenantConfig.
3684 0 : let tenant_config: models::TenantConfig = conf.tenant_conf.clone().into();
3685 0 :
3686 0 : models::LocationConfig {
3687 0 : mode: location_config_mode,
3688 0 : generation: self.generation.into(),
3689 0 : secondary_conf: None,
3690 0 : shard_number: self.shard_identity.number.0,
3691 0 : shard_count: self.shard_identity.count.literal(),
3692 0 : shard_stripe_size: self.shard_identity.stripe_size.0,
3693 0 : tenant_conf: tenant_config,
3694 0 : }
3695 0 : }
3696 :
3697 0 : pub(crate) fn get_tenant_shard_id(&self) -> &TenantShardId {
3698 0 : &self.tenant_shard_id
3699 0 : }
3700 :
3701 0 : pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
3702 0 : self.shard_identity.stripe_size
3703 0 : }
3704 :
3705 0 : pub(crate) fn get_generation(&self) -> Generation {
3706 0 : self.generation
3707 0 : }
3708 :
3709 : /// This function partially shuts down the tenant (it shuts down the Timelines) and is fallible,
3710 : /// and can leave the tenant in a bad state if it fails. The caller is responsible for
3711 : /// resetting this tenant to a valid state if we fail.
3712 0 : pub(crate) async fn split_prepare(
3713 0 : &self,
3714 0 : child_shards: &Vec<TenantShardId>,
3715 0 : ) -> anyhow::Result<()> {
3716 0 : let (timelines, offloaded) = {
3717 0 : let timelines = self.timelines.lock().unwrap();
3718 0 : let offloaded = self.timelines_offloaded.lock().unwrap();
3719 0 : (timelines.clone(), offloaded.clone())
3720 0 : };
3721 0 : let timelines_iter = timelines
3722 0 : .values()
3723 0 : .map(TimelineOrOffloadedArcRef::<'_>::from)
3724 0 : .chain(
3725 0 : offloaded
3726 0 : .values()
3727 0 : .map(TimelineOrOffloadedArcRef::<'_>::from),
3728 0 : );
3729 0 : for timeline in timelines_iter {
3730 : // We do not block timeline creation/deletion during splits inside the pageserver: it is up to higher levels
3731 : // to ensure that they do not start a split if currently in the process of doing these.
3732 :
3733 0 : let timeline_id = timeline.timeline_id();
3734 :
3735 0 : if let TimelineOrOffloadedArcRef::Timeline(timeline) = timeline {
3736 : // Upload an index from the parent: this is partly to provide freshness for the
3737 : // child tenants that will copy it, and partly for general ease-of-debugging: there will
3738 : // always be a parent shard index in the same generation as we wrote the child shard index.
3739 0 : tracing::info!(%timeline_id, "Uploading index");
3740 0 : timeline
3741 0 : .remote_client
3742 0 : .schedule_index_upload_for_file_changes()?;
3743 0 : timeline.remote_client.wait_completion().await?;
3744 0 : }
3745 :
3746 0 : let remote_client = match timeline {
3747 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.remote_client.clone(),
3748 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => {
3749 0 : let remote_client = self
3750 0 : .build_timeline_client(offloaded.timeline_id, self.remote_storage.clone());
3751 0 : Arc::new(remote_client)
3752 : }
3753 : };
3754 :
3755 : // Shut down the timeline's remote client: this means that the indices we write
3756 : // for child shards will not be invalidated by the parent shard deleting layers.
3757 0 : tracing::info!(%timeline_id, "Shutting down remote storage client");
3758 0 : remote_client.shutdown().await;
3759 :
3760 : // Download methods can still be used after shutdown, as they don't flow through the remote client's
3761 : // queue. In principal the RemoteTimelineClient could provide this without downloading it, but this
3762 : // operation is rare, so it's simpler to just download it (and robustly guarantees that the index
3763 : // we use here really is the remotely persistent one).
3764 0 : tracing::info!(%timeline_id, "Downloading index_part from parent");
3765 0 : let result = remote_client
3766 0 : .download_index_file(&self.cancel)
3767 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))
3768 0 : .await?;
3769 0 : let index_part = match result {
3770 : MaybeDeletedIndexPart::Deleted(_) => {
3771 0 : anyhow::bail!("Timeline deletion happened concurrently with split")
3772 : }
3773 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
3774 : };
3775 :
3776 0 : for child_shard in child_shards {
3777 0 : tracing::info!(%timeline_id, "Uploading index_part for child {}", child_shard.to_index());
3778 0 : upload_index_part(
3779 0 : &self.remote_storage,
3780 0 : child_shard,
3781 0 : &timeline_id,
3782 0 : self.generation,
3783 0 : &index_part,
3784 0 : &self.cancel,
3785 0 : )
3786 0 : .await?;
3787 : }
3788 : }
3789 :
3790 0 : let tenant_manifest = self.build_tenant_manifest();
3791 0 : for child_shard in child_shards {
3792 0 : tracing::info!(
3793 0 : "Uploading tenant manifest for child {}",
3794 0 : child_shard.to_index()
3795 : );
3796 0 : upload_tenant_manifest(
3797 0 : &self.remote_storage,
3798 0 : child_shard,
3799 0 : self.generation,
3800 0 : &tenant_manifest,
3801 0 : &self.cancel,
3802 0 : )
3803 0 : .await?;
3804 : }
3805 :
3806 0 : Ok(())
3807 0 : }
3808 :
3809 0 : pub(crate) fn get_sizes(&self) -> TopTenantShardItem {
3810 0 : let mut result = TopTenantShardItem {
3811 0 : id: self.tenant_shard_id,
3812 0 : resident_size: 0,
3813 0 : physical_size: 0,
3814 0 : max_logical_size: 0,
3815 0 : };
3816 :
3817 0 : for timeline in self.timelines.lock().unwrap().values() {
3818 0 : result.resident_size += timeline.metrics.resident_physical_size_gauge.get();
3819 0 :
3820 0 : result.physical_size += timeline
3821 0 : .remote_client
3822 0 : .metrics
3823 0 : .remote_physical_size_gauge
3824 0 : .get();
3825 0 : result.max_logical_size = std::cmp::max(
3826 0 : result.max_logical_size,
3827 0 : timeline.metrics.current_logical_size_gauge.get(),
3828 0 : );
3829 0 : }
3830 :
3831 0 : result
3832 0 : }
3833 : }
3834 :
3835 : /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
3836 : /// perform a topological sort, so that the parent of each timeline comes
3837 : /// before the children.
3838 : /// E extracts the ancestor from T
3839 : /// This allows for T to be different. It can be TimelineMetadata, can be Timeline itself, etc.
3840 444 : fn tree_sort_timelines<T, E>(
3841 444 : timelines: HashMap<TimelineId, T>,
3842 444 : extractor: E,
3843 444 : ) -> anyhow::Result<Vec<(TimelineId, T)>>
3844 444 : where
3845 444 : E: Fn(&T) -> Option<TimelineId>,
3846 444 : {
3847 444 : let mut result = Vec::with_capacity(timelines.len());
3848 444 :
3849 444 : let mut now = Vec::with_capacity(timelines.len());
3850 444 : // (ancestor, children)
3851 444 : let mut later: HashMap<TimelineId, Vec<(TimelineId, T)>> =
3852 444 : HashMap::with_capacity(timelines.len());
3853 :
3854 456 : for (timeline_id, value) in timelines {
3855 12 : if let Some(ancestor_id) = extractor(&value) {
3856 4 : let children = later.entry(ancestor_id).or_default();
3857 4 : children.push((timeline_id, value));
3858 8 : } else {
3859 8 : now.push((timeline_id, value));
3860 8 : }
3861 : }
3862 :
3863 456 : while let Some((timeline_id, metadata)) = now.pop() {
3864 12 : result.push((timeline_id, metadata));
3865 : // All children of this can be loaded now
3866 12 : if let Some(mut children) = later.remove(&timeline_id) {
3867 4 : now.append(&mut children);
3868 8 : }
3869 : }
3870 :
3871 : // All timelines should be visited now. Unless there were timelines with missing ancestors.
3872 444 : if !later.is_empty() {
3873 0 : for (missing_id, orphan_ids) in later {
3874 0 : for (orphan_id, _) in orphan_ids {
3875 0 : error!("could not load timeline {orphan_id} because its ancestor timeline {missing_id} could not be loaded");
3876 : }
3877 : }
3878 0 : bail!("could not load tenant because some timelines are missing ancestors");
3879 444 : }
3880 444 :
3881 444 : Ok(result)
3882 444 : }
3883 :
3884 : enum ActivateTimelineArgs {
3885 : Yes {
3886 : broker_client: storage_broker::BrokerClientChannel,
3887 : },
3888 : No,
3889 : }
3890 :
3891 : impl Tenant {
3892 0 : pub fn tenant_specific_overrides(&self) -> TenantConfOpt {
3893 0 : self.tenant_conf.load().tenant_conf.clone()
3894 0 : }
3895 :
3896 0 : pub fn effective_config(&self) -> TenantConf {
3897 0 : self.tenant_specific_overrides()
3898 0 : .merge(self.conf.default_tenant_conf.clone())
3899 0 : }
3900 :
3901 0 : pub fn get_checkpoint_distance(&self) -> u64 {
3902 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3903 0 : tenant_conf
3904 0 : .checkpoint_distance
3905 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
3906 0 : }
3907 :
3908 0 : pub fn get_checkpoint_timeout(&self) -> Duration {
3909 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3910 0 : tenant_conf
3911 0 : .checkpoint_timeout
3912 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
3913 0 : }
3914 :
3915 0 : pub fn get_compaction_target_size(&self) -> u64 {
3916 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3917 0 : tenant_conf
3918 0 : .compaction_target_size
3919 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
3920 0 : }
3921 :
3922 0 : pub fn get_compaction_period(&self) -> Duration {
3923 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3924 0 : tenant_conf
3925 0 : .compaction_period
3926 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
3927 0 : }
3928 :
3929 0 : pub fn get_compaction_threshold(&self) -> usize {
3930 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3931 0 : tenant_conf
3932 0 : .compaction_threshold
3933 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
3934 0 : }
3935 :
3936 0 : pub fn get_rel_size_v2_enabled(&self) -> bool {
3937 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3938 0 : tenant_conf
3939 0 : .rel_size_v2_enabled
3940 0 : .unwrap_or(self.conf.default_tenant_conf.rel_size_v2_enabled)
3941 0 : }
3942 :
3943 0 : pub fn get_compaction_upper_limit(&self) -> usize {
3944 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3945 0 : tenant_conf
3946 0 : .compaction_upper_limit
3947 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_upper_limit)
3948 0 : }
3949 :
3950 0 : pub fn get_compaction_l0_first(&self) -> bool {
3951 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3952 0 : tenant_conf
3953 0 : .compaction_l0_first
3954 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_first)
3955 0 : }
3956 :
3957 0 : pub fn get_gc_horizon(&self) -> u64 {
3958 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3959 0 : tenant_conf
3960 0 : .gc_horizon
3961 0 : .unwrap_or(self.conf.default_tenant_conf.gc_horizon)
3962 0 : }
3963 :
3964 0 : pub fn get_gc_period(&self) -> Duration {
3965 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3966 0 : tenant_conf
3967 0 : .gc_period
3968 0 : .unwrap_or(self.conf.default_tenant_conf.gc_period)
3969 0 : }
3970 :
3971 0 : pub fn get_image_creation_threshold(&self) -> usize {
3972 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3973 0 : tenant_conf
3974 0 : .image_creation_threshold
3975 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
3976 0 : }
3977 :
3978 0 : pub fn get_pitr_interval(&self) -> Duration {
3979 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3980 0 : tenant_conf
3981 0 : .pitr_interval
3982 0 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
3983 0 : }
3984 :
3985 0 : pub fn get_min_resident_size_override(&self) -> Option<u64> {
3986 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3987 0 : tenant_conf
3988 0 : .min_resident_size_override
3989 0 : .or(self.conf.default_tenant_conf.min_resident_size_override)
3990 0 : }
3991 :
3992 0 : pub fn get_heatmap_period(&self) -> Option<Duration> {
3993 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3994 0 : let heatmap_period = tenant_conf
3995 0 : .heatmap_period
3996 0 : .unwrap_or(self.conf.default_tenant_conf.heatmap_period);
3997 0 : if heatmap_period.is_zero() {
3998 0 : None
3999 : } else {
4000 0 : Some(heatmap_period)
4001 : }
4002 0 : }
4003 :
4004 8 : pub fn get_lsn_lease_length(&self) -> Duration {
4005 8 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4006 8 : tenant_conf
4007 8 : .lsn_lease_length
4008 8 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
4009 8 : }
4010 :
4011 0 : pub fn get_timeline_offloading_enabled(&self) -> bool {
4012 0 : if self.conf.timeline_offloading {
4013 0 : return true;
4014 0 : }
4015 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4016 0 : tenant_conf
4017 0 : .timeline_offloading
4018 0 : .unwrap_or(self.conf.default_tenant_conf.timeline_offloading)
4019 0 : }
4020 :
4021 : /// Generate an up-to-date TenantManifest based on the state of this Tenant.
4022 4 : fn build_tenant_manifest(&self) -> TenantManifest {
4023 4 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
4024 4 :
4025 4 : let mut timeline_manifests = timelines_offloaded
4026 4 : .iter()
4027 4 : .map(|(_timeline_id, offloaded)| offloaded.manifest())
4028 4 : .collect::<Vec<_>>();
4029 4 : // Sort the manifests so that our output is deterministic
4030 4 : timeline_manifests.sort_by_key(|timeline_manifest| timeline_manifest.timeline_id);
4031 4 :
4032 4 : TenantManifest {
4033 4 : version: LATEST_TENANT_MANIFEST_VERSION,
4034 4 : offloaded_timelines: timeline_manifests,
4035 4 : }
4036 4 : }
4037 :
4038 0 : pub fn update_tenant_config<F: Fn(TenantConfOpt) -> anyhow::Result<TenantConfOpt>>(
4039 0 : &self,
4040 0 : update: F,
4041 0 : ) -> anyhow::Result<TenantConfOpt> {
4042 0 : // Use read-copy-update in order to avoid overwriting the location config
4043 0 : // state if this races with [`Tenant::set_new_location_config`]. Note that
4044 0 : // this race is not possible if both request types come from the storage
4045 0 : // controller (as they should!) because an exclusive op lock is required
4046 0 : // on the storage controller side.
4047 0 :
4048 0 : self.tenant_conf
4049 0 : .try_rcu(|attached_conf| -> Result<_, anyhow::Error> {
4050 0 : Ok(Arc::new(AttachedTenantConf {
4051 0 : tenant_conf: update(attached_conf.tenant_conf.clone())?,
4052 0 : location: attached_conf.location,
4053 0 : lsn_lease_deadline: attached_conf.lsn_lease_deadline,
4054 : }))
4055 0 : })?;
4056 :
4057 0 : let updated = self.tenant_conf.load();
4058 0 :
4059 0 : self.tenant_conf_updated(&updated.tenant_conf);
4060 0 : // Don't hold self.timelines.lock() during the notifies.
4061 0 : // There's no risk of deadlock right now, but there could be if we consolidate
4062 0 : // mutexes in struct Timeline in the future.
4063 0 : let timelines = self.list_timelines();
4064 0 : for timeline in timelines {
4065 0 : timeline.tenant_conf_updated(&updated);
4066 0 : }
4067 :
4068 0 : Ok(updated.tenant_conf.clone())
4069 0 : }
4070 :
4071 0 : pub(crate) fn set_new_location_config(&self, new_conf: AttachedTenantConf) {
4072 0 : let new_tenant_conf = new_conf.tenant_conf.clone();
4073 0 :
4074 0 : self.tenant_conf.store(Arc::new(new_conf.clone()));
4075 0 :
4076 0 : self.tenant_conf_updated(&new_tenant_conf);
4077 0 : // Don't hold self.timelines.lock() during the notifies.
4078 0 : // There's no risk of deadlock right now, but there could be if we consolidate
4079 0 : // mutexes in struct Timeline in the future.
4080 0 : let timelines = self.list_timelines();
4081 0 : for timeline in timelines {
4082 0 : timeline.tenant_conf_updated(&new_conf);
4083 0 : }
4084 0 : }
4085 :
4086 444 : fn get_pagestream_throttle_config(
4087 444 : psconf: &'static PageServerConf,
4088 444 : overrides: &TenantConfOpt,
4089 444 : ) -> throttle::Config {
4090 444 : overrides
4091 444 : .timeline_get_throttle
4092 444 : .clone()
4093 444 : .unwrap_or(psconf.default_tenant_conf.timeline_get_throttle.clone())
4094 444 : }
4095 :
4096 0 : pub(crate) fn tenant_conf_updated(&self, new_conf: &TenantConfOpt) {
4097 0 : let conf = Self::get_pagestream_throttle_config(self.conf, new_conf);
4098 0 : self.pagestream_throttle.reconfigure(conf)
4099 0 : }
4100 :
4101 : /// Helper function to create a new Timeline struct.
4102 : ///
4103 : /// The returned Timeline is in Loading state. The caller is responsible for
4104 : /// initializing any on-disk state, and for inserting the Timeline to the 'timelines'
4105 : /// map.
4106 : ///
4107 : /// `validate_ancestor == false` is used when a timeline is created for deletion
4108 : /// and we might not have the ancestor present anymore which is fine for to be
4109 : /// deleted timelines.
4110 : #[allow(clippy::too_many_arguments)]
4111 896 : fn create_timeline_struct(
4112 896 : &self,
4113 896 : new_timeline_id: TimelineId,
4114 896 : new_metadata: &TimelineMetadata,
4115 896 : previous_heatmap: Option<PreviousHeatmap>,
4116 896 : ancestor: Option<Arc<Timeline>>,
4117 896 : resources: TimelineResources,
4118 896 : cause: CreateTimelineCause,
4119 896 : create_idempotency: CreateTimelineIdempotency,
4120 896 : ) -> anyhow::Result<Arc<Timeline>> {
4121 896 : let state = match cause {
4122 : CreateTimelineCause::Load => {
4123 896 : let ancestor_id = new_metadata.ancestor_timeline();
4124 896 : anyhow::ensure!(
4125 896 : ancestor_id == ancestor.as_ref().map(|t| t.timeline_id),
4126 0 : "Timeline's {new_timeline_id} ancestor {ancestor_id:?} was not found"
4127 : );
4128 896 : TimelineState::Loading
4129 : }
4130 0 : CreateTimelineCause::Delete => TimelineState::Stopping,
4131 : };
4132 :
4133 896 : let pg_version = new_metadata.pg_version();
4134 896 :
4135 896 : let timeline = Timeline::new(
4136 896 : self.conf,
4137 896 : Arc::clone(&self.tenant_conf),
4138 896 : new_metadata,
4139 896 : previous_heatmap,
4140 896 : ancestor,
4141 896 : new_timeline_id,
4142 896 : self.tenant_shard_id,
4143 896 : self.generation,
4144 896 : self.shard_identity,
4145 896 : self.walredo_mgr.clone(),
4146 896 : resources,
4147 896 : pg_version,
4148 896 : state,
4149 896 : self.attach_wal_lag_cooldown.clone(),
4150 896 : create_idempotency,
4151 896 : self.cancel.child_token(),
4152 896 : );
4153 896 :
4154 896 : Ok(timeline)
4155 896 : }
4156 :
4157 : /// [`Tenant::shutdown`] must be called before dropping the returned [`Tenant`] object
4158 : /// to ensure proper cleanup of background tasks and metrics.
4159 : //
4160 : // Allow too_many_arguments because a constructor's argument list naturally grows with the
4161 : // number of attributes in the struct: breaking these out into a builder wouldn't be helpful.
4162 : #[allow(clippy::too_many_arguments)]
4163 444 : fn new(
4164 444 : state: TenantState,
4165 444 : conf: &'static PageServerConf,
4166 444 : attached_conf: AttachedTenantConf,
4167 444 : shard_identity: ShardIdentity,
4168 444 : walredo_mgr: Option<Arc<WalRedoManager>>,
4169 444 : tenant_shard_id: TenantShardId,
4170 444 : remote_storage: GenericRemoteStorage,
4171 444 : deletion_queue_client: DeletionQueueClient,
4172 444 : l0_flush_global_state: L0FlushGlobalState,
4173 444 : ) -> Tenant {
4174 444 : debug_assert!(
4175 444 : !attached_conf.location.generation.is_none() || conf.control_plane_api.is_none()
4176 : );
4177 :
4178 444 : let (state, mut rx) = watch::channel(state);
4179 444 :
4180 444 : tokio::spawn(async move {
4181 444 : // reflect tenant state in metrics:
4182 444 : // - global per tenant state: TENANT_STATE_METRIC
4183 444 : // - "set" of broken tenants: BROKEN_TENANTS_SET
4184 444 : //
4185 444 : // set of broken tenants should not have zero counts so that it remains accessible for
4186 444 : // alerting.
4187 444 :
4188 444 : let tid = tenant_shard_id.to_string();
4189 444 : let shard_id = tenant_shard_id.shard_slug().to_string();
4190 444 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
4191 :
4192 888 : fn inspect_state(state: &TenantState) -> ([&'static str; 1], bool) {
4193 888 : ([state.into()], matches!(state, TenantState::Broken { .. }))
4194 888 : }
4195 :
4196 444 : let mut tuple = inspect_state(&rx.borrow_and_update());
4197 444 :
4198 444 : let is_broken = tuple.1;
4199 444 : let mut counted_broken = if is_broken {
4200 : // add the id to the set right away, there should not be any updates on the channel
4201 : // after before tenant is removed, if ever
4202 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
4203 0 : true
4204 : } else {
4205 444 : false
4206 : };
4207 :
4208 : loop {
4209 888 : let labels = &tuple.0;
4210 888 : let current = TENANT_STATE_METRIC.with_label_values(labels);
4211 888 : current.inc();
4212 888 :
4213 888 : if rx.changed().await.is_err() {
4214 : // tenant has been dropped
4215 28 : current.dec();
4216 28 : drop(BROKEN_TENANTS_SET.remove_label_values(set_key));
4217 28 : break;
4218 444 : }
4219 444 :
4220 444 : current.dec();
4221 444 : tuple = inspect_state(&rx.borrow_and_update());
4222 444 :
4223 444 : let is_broken = tuple.1;
4224 444 : if is_broken && !counted_broken {
4225 0 : counted_broken = true;
4226 0 : // insert the tenant_id (back) into the set while avoiding needless counter
4227 0 : // access
4228 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
4229 444 : }
4230 : }
4231 444 : });
4232 444 :
4233 444 : Tenant {
4234 444 : tenant_shard_id,
4235 444 : shard_identity,
4236 444 : generation: attached_conf.location.generation,
4237 444 : conf,
4238 444 : // using now here is good enough approximation to catch tenants with really long
4239 444 : // activation times.
4240 444 : constructed_at: Instant::now(),
4241 444 : timelines: Mutex::new(HashMap::new()),
4242 444 : timelines_creating: Mutex::new(HashSet::new()),
4243 444 : timelines_offloaded: Mutex::new(HashMap::new()),
4244 444 : tenant_manifest_upload: Default::default(),
4245 444 : gc_cs: tokio::sync::Mutex::new(()),
4246 444 : walredo_mgr,
4247 444 : remote_storage,
4248 444 : deletion_queue_client,
4249 444 : state,
4250 444 : cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
4251 444 : cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
4252 444 : eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
4253 444 : compaction_circuit_breaker: std::sync::Mutex::new(CircuitBreaker::new(
4254 444 : format!("compaction-{tenant_shard_id}"),
4255 444 : 5,
4256 444 : // Compaction can be a very expensive operation, and might leak disk space. It also ought
4257 444 : // to be infallible, as long as remote storage is available. So if it repeatedly fails,
4258 444 : // use an extremely long backoff.
4259 444 : Some(Duration::from_secs(3600 * 24)),
4260 444 : )),
4261 444 : l0_compaction_trigger: Arc::new(Notify::new()),
4262 444 : scheduled_compaction_tasks: Mutex::new(Default::default()),
4263 444 : activate_now_sem: tokio::sync::Semaphore::new(0),
4264 444 : attach_wal_lag_cooldown: Arc::new(std::sync::OnceLock::new()),
4265 444 : cancel: CancellationToken::default(),
4266 444 : gate: Gate::default(),
4267 444 : pagestream_throttle: Arc::new(throttle::Throttle::new(
4268 444 : Tenant::get_pagestream_throttle_config(conf, &attached_conf.tenant_conf),
4269 444 : )),
4270 444 : pagestream_throttle_metrics: Arc::new(
4271 444 : crate::metrics::tenant_throttling::Pagestream::new(&tenant_shard_id),
4272 444 : ),
4273 444 : tenant_conf: Arc::new(ArcSwap::from_pointee(attached_conf)),
4274 444 : ongoing_timeline_detach: std::sync::Mutex::default(),
4275 444 : gc_block: Default::default(),
4276 444 : l0_flush_global_state,
4277 444 : }
4278 444 : }
4279 :
4280 : /// Locate and load config
4281 0 : pub(super) fn load_tenant_config(
4282 0 : conf: &'static PageServerConf,
4283 0 : tenant_shard_id: &TenantShardId,
4284 0 : ) -> Result<LocationConf, LoadConfigError> {
4285 0 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4286 0 :
4287 0 : info!("loading tenant configuration from {config_path}");
4288 :
4289 : // load and parse file
4290 0 : let config = fs::read_to_string(&config_path).map_err(|e| {
4291 0 : match e.kind() {
4292 : std::io::ErrorKind::NotFound => {
4293 : // The config should almost always exist for a tenant directory:
4294 : // - When attaching a tenant, the config is the first thing we write
4295 : // - When detaching a tenant, we atomically move the directory to a tmp location
4296 : // before deleting contents.
4297 : //
4298 : // The very rare edge case that can result in a missing config is if we crash during attach
4299 : // between creating directory and writing config. Callers should handle that as if the
4300 : // directory didn't exist.
4301 :
4302 0 : LoadConfigError::NotFound(config_path)
4303 : }
4304 : _ => {
4305 : // No IO errors except NotFound are acceptable here: other kinds of error indicate local storage or permissions issues
4306 : // that we cannot cleanly recover
4307 0 : crate::virtual_file::on_fatal_io_error(&e, "Reading tenant config file")
4308 : }
4309 : }
4310 0 : })?;
4311 :
4312 0 : Ok(toml_edit::de::from_str::<LocationConf>(&config)?)
4313 0 : }
4314 :
4315 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4316 : pub(super) async fn persist_tenant_config(
4317 : conf: &'static PageServerConf,
4318 : tenant_shard_id: &TenantShardId,
4319 : location_conf: &LocationConf,
4320 : ) -> std::io::Result<()> {
4321 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4322 :
4323 : Self::persist_tenant_config_at(tenant_shard_id, &config_path, location_conf).await
4324 : }
4325 :
4326 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4327 : pub(super) async fn persist_tenant_config_at(
4328 : tenant_shard_id: &TenantShardId,
4329 : config_path: &Utf8Path,
4330 : location_conf: &LocationConf,
4331 : ) -> std::io::Result<()> {
4332 : debug!("persisting tenantconf to {config_path}");
4333 :
4334 : let mut conf_content = r#"# This file contains a specific per-tenant's config.
4335 : # It is read in case of pageserver restart.
4336 : "#
4337 : .to_string();
4338 :
4339 0 : fail::fail_point!("tenant-config-before-write", |_| {
4340 0 : Err(std::io::Error::new(
4341 0 : std::io::ErrorKind::Other,
4342 0 : "tenant-config-before-write",
4343 0 : ))
4344 0 : });
4345 :
4346 : // Convert the config to a toml file.
4347 : conf_content +=
4348 : &toml_edit::ser::to_string_pretty(&location_conf).expect("Config serialization failed");
4349 :
4350 : let temp_path = path_with_suffix_extension(config_path, TEMP_FILE_SUFFIX);
4351 :
4352 : let conf_content = conf_content.into_bytes();
4353 : VirtualFile::crashsafe_overwrite(config_path.to_owned(), temp_path, conf_content).await
4354 : }
4355 :
4356 : //
4357 : // How garbage collection works:
4358 : //
4359 : // +--bar------------->
4360 : // /
4361 : // +----+-----foo---------------->
4362 : // /
4363 : // ----main--+-------------------------->
4364 : // \
4365 : // +-----baz-------->
4366 : //
4367 : //
4368 : // 1. Grab 'gc_cs' mutex to prevent new timelines from being created while Timeline's
4369 : // `gc_infos` are being refreshed
4370 : // 2. Scan collected timelines, and on each timeline, make note of the
4371 : // all the points where other timelines have been branched off.
4372 : // We will refrain from removing page versions at those LSNs.
4373 : // 3. For each timeline, scan all layer files on the timeline.
4374 : // Remove all files for which a newer file exists and which
4375 : // don't cover any branch point LSNs.
4376 : //
4377 : // TODO:
4378 : // - if a relation has a non-incremental persistent layer on a child branch, then we
4379 : // don't need to keep that in the parent anymore. But currently
4380 : // we do.
4381 8 : async fn gc_iteration_internal(
4382 8 : &self,
4383 8 : target_timeline_id: Option<TimelineId>,
4384 8 : horizon: u64,
4385 8 : pitr: Duration,
4386 8 : cancel: &CancellationToken,
4387 8 : ctx: &RequestContext,
4388 8 : ) -> Result<GcResult, GcError> {
4389 8 : let mut totals: GcResult = Default::default();
4390 8 : let now = Instant::now();
4391 :
4392 8 : let gc_timelines = self
4393 8 : .refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4394 8 : .await?;
4395 :
4396 8 : failpoint_support::sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
4397 :
4398 : // If there is nothing to GC, we don't want any messages in the INFO log.
4399 8 : if !gc_timelines.is_empty() {
4400 8 : info!("{} timelines need GC", gc_timelines.len());
4401 : } else {
4402 0 : debug!("{} timelines need GC", gc_timelines.len());
4403 : }
4404 :
4405 : // Perform GC for each timeline.
4406 : //
4407 : // Note that we don't hold the `Tenant::gc_cs` lock here because we don't want to delay the
4408 : // branch creation task, which requires the GC lock. A GC iteration can run concurrently
4409 : // with branch creation.
4410 : //
4411 : // See comments in [`Tenant::branch_timeline`] for more information about why branch
4412 : // creation task can run concurrently with timeline's GC iteration.
4413 16 : for timeline in gc_timelines {
4414 8 : if cancel.is_cancelled() {
4415 : // We were requested to shut down. Stop and return with the progress we
4416 : // made.
4417 0 : break;
4418 8 : }
4419 8 : let result = match timeline.gc().await {
4420 : Err(GcError::TimelineCancelled) => {
4421 0 : if target_timeline_id.is_some() {
4422 : // If we were targetting this specific timeline, surface cancellation to caller
4423 0 : return Err(GcError::TimelineCancelled);
4424 : } else {
4425 : // A timeline may be shutting down independently of the tenant's lifecycle: we should
4426 : // skip past this and proceed to try GC on other timelines.
4427 0 : continue;
4428 : }
4429 : }
4430 8 : r => r?,
4431 : };
4432 8 : totals += result;
4433 : }
4434 :
4435 8 : totals.elapsed = now.elapsed();
4436 8 : Ok(totals)
4437 8 : }
4438 :
4439 : /// Refreshes the Timeline::gc_info for all timelines, returning the
4440 : /// vector of timelines which have [`Timeline::get_last_record_lsn`] past
4441 : /// [`Tenant::get_gc_horizon`].
4442 : ///
4443 : /// This is usually executed as part of periodic gc, but can now be triggered more often.
4444 0 : pub(crate) async fn refresh_gc_info(
4445 0 : &self,
4446 0 : cancel: &CancellationToken,
4447 0 : ctx: &RequestContext,
4448 0 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4449 0 : // since this method can now be called at different rates than the configured gc loop, it
4450 0 : // might be that these configuration values get applied faster than what it was previously,
4451 0 : // since these were only read from the gc task.
4452 0 : let horizon = self.get_gc_horizon();
4453 0 : let pitr = self.get_pitr_interval();
4454 0 :
4455 0 : // refresh all timelines
4456 0 : let target_timeline_id = None;
4457 0 :
4458 0 : self.refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4459 0 : .await
4460 0 : }
4461 :
4462 : /// Populate all Timelines' `GcInfo` with information about their children. We do not set the
4463 : /// PITR cutoffs here, because that requires I/O: this is done later, before GC, by [`Self::refresh_gc_info_internal`]
4464 : ///
4465 : /// Subsequently, parent-child relationships are updated incrementally inside [`Timeline::new`] and [`Timeline::drop`].
4466 0 : fn initialize_gc_info(
4467 0 : &self,
4468 0 : timelines: &std::sync::MutexGuard<HashMap<TimelineId, Arc<Timeline>>>,
4469 0 : timelines_offloaded: &std::sync::MutexGuard<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
4470 0 : restrict_to_timeline: Option<TimelineId>,
4471 0 : ) {
4472 0 : if restrict_to_timeline.is_none() {
4473 : // This function must be called before activation: after activation timeline create/delete operations
4474 : // might happen, and this function is not safe to run concurrently with those.
4475 0 : assert!(!self.is_active());
4476 0 : }
4477 :
4478 : // Scan all timelines. For each timeline, remember the timeline ID and
4479 : // the branch point where it was created.
4480 0 : let mut all_branchpoints: BTreeMap<TimelineId, Vec<(Lsn, TimelineId, MaybeOffloaded)>> =
4481 0 : BTreeMap::new();
4482 0 : timelines.iter().for_each(|(timeline_id, timeline_entry)| {
4483 0 : if let Some(ancestor_timeline_id) = &timeline_entry.get_ancestor_timeline_id() {
4484 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4485 0 : ancestor_children.push((
4486 0 : timeline_entry.get_ancestor_lsn(),
4487 0 : *timeline_id,
4488 0 : MaybeOffloaded::No,
4489 0 : ));
4490 0 : }
4491 0 : });
4492 0 : timelines_offloaded
4493 0 : .iter()
4494 0 : .for_each(|(timeline_id, timeline_entry)| {
4495 0 : let Some(ancestor_timeline_id) = &timeline_entry.ancestor_timeline_id else {
4496 0 : return;
4497 : };
4498 0 : let Some(retain_lsn) = timeline_entry.ancestor_retain_lsn else {
4499 0 : return;
4500 : };
4501 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4502 0 : ancestor_children.push((retain_lsn, *timeline_id, MaybeOffloaded::Yes));
4503 0 : });
4504 0 :
4505 0 : // The number of bytes we always keep, irrespective of PITR: this is a constant across timelines
4506 0 : let horizon = self.get_gc_horizon();
4507 :
4508 : // Populate each timeline's GcInfo with information about its child branches
4509 0 : let timelines_to_write = if let Some(timeline_id) = restrict_to_timeline {
4510 0 : itertools::Either::Left(timelines.get(&timeline_id).into_iter())
4511 : } else {
4512 0 : itertools::Either::Right(timelines.values())
4513 : };
4514 0 : for timeline in timelines_to_write {
4515 0 : let mut branchpoints: Vec<(Lsn, TimelineId, MaybeOffloaded)> = all_branchpoints
4516 0 : .remove(&timeline.timeline_id)
4517 0 : .unwrap_or_default();
4518 0 :
4519 0 : branchpoints.sort_by_key(|b| b.0);
4520 0 :
4521 0 : let mut target = timeline.gc_info.write().unwrap();
4522 0 :
4523 0 : target.retain_lsns = branchpoints;
4524 0 :
4525 0 : let space_cutoff = timeline
4526 0 : .get_last_record_lsn()
4527 0 : .checked_sub(horizon)
4528 0 : .unwrap_or(Lsn(0));
4529 0 :
4530 0 : target.cutoffs = GcCutoffs {
4531 0 : space: space_cutoff,
4532 0 : time: Lsn::INVALID,
4533 0 : };
4534 0 : }
4535 0 : }
4536 :
4537 8 : async fn refresh_gc_info_internal(
4538 8 : &self,
4539 8 : target_timeline_id: Option<TimelineId>,
4540 8 : horizon: u64,
4541 8 : pitr: Duration,
4542 8 : cancel: &CancellationToken,
4543 8 : ctx: &RequestContext,
4544 8 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4545 8 : // before taking the gc_cs lock, do the heavier weight finding of gc_cutoff points for
4546 8 : // currently visible timelines.
4547 8 : let timelines = self
4548 8 : .timelines
4549 8 : .lock()
4550 8 : .unwrap()
4551 8 : .values()
4552 8 : .filter(|tl| match target_timeline_id.as_ref() {
4553 8 : Some(target) => &tl.timeline_id == target,
4554 0 : None => true,
4555 8 : })
4556 8 : .cloned()
4557 8 : .collect::<Vec<_>>();
4558 8 :
4559 8 : if target_timeline_id.is_some() && timelines.is_empty() {
4560 : // We were to act on a particular timeline and it wasn't found
4561 0 : return Err(GcError::TimelineNotFound);
4562 8 : }
4563 8 :
4564 8 : let mut gc_cutoffs: HashMap<TimelineId, GcCutoffs> =
4565 8 : HashMap::with_capacity(timelines.len());
4566 8 :
4567 8 : // Ensures all timelines use the same start time when computing the time cutoff.
4568 8 : let now_ts_for_pitr_calc = SystemTime::now();
4569 8 : for timeline in timelines.iter() {
4570 8 : let cutoff = timeline
4571 8 : .get_last_record_lsn()
4572 8 : .checked_sub(horizon)
4573 8 : .unwrap_or(Lsn(0));
4574 :
4575 8 : let cutoffs = timeline
4576 8 : .find_gc_cutoffs(now_ts_for_pitr_calc, cutoff, pitr, cancel, ctx)
4577 8 : .await?;
4578 8 : let old = gc_cutoffs.insert(timeline.timeline_id, cutoffs);
4579 8 : assert!(old.is_none());
4580 : }
4581 :
4582 8 : if !self.is_active() || self.cancel.is_cancelled() {
4583 0 : return Err(GcError::TenantCancelled);
4584 8 : }
4585 :
4586 : // grab mutex to prevent new timelines from being created here; avoid doing long operations
4587 : // because that will stall branch creation.
4588 8 : let gc_cs = self.gc_cs.lock().await;
4589 :
4590 : // Ok, we now know all the branch points.
4591 : // Update the GC information for each timeline.
4592 8 : let mut gc_timelines = Vec::with_capacity(timelines.len());
4593 16 : for timeline in timelines {
4594 : // We filtered the timeline list above
4595 8 : if let Some(target_timeline_id) = target_timeline_id {
4596 8 : assert_eq!(target_timeline_id, timeline.timeline_id);
4597 0 : }
4598 :
4599 : {
4600 8 : let mut target = timeline.gc_info.write().unwrap();
4601 8 :
4602 8 : // Cull any expired leases
4603 8 : let now = SystemTime::now();
4604 12 : target.leases.retain(|_, lease| !lease.is_expired(&now));
4605 8 :
4606 8 : timeline
4607 8 : .metrics
4608 8 : .valid_lsn_lease_count_gauge
4609 8 : .set(target.leases.len() as u64);
4610 :
4611 : // Look up parent's PITR cutoff to update the child's knowledge of whether it is within parent's PITR
4612 8 : if let Some(ancestor_id) = timeline.get_ancestor_timeline_id() {
4613 0 : if let Some(ancestor_gc_cutoffs) = gc_cutoffs.get(&ancestor_id) {
4614 0 : target.within_ancestor_pitr =
4615 0 : timeline.get_ancestor_lsn() >= ancestor_gc_cutoffs.time;
4616 0 : }
4617 8 : }
4618 :
4619 : // Update metrics that depend on GC state
4620 8 : timeline
4621 8 : .metrics
4622 8 : .archival_size
4623 8 : .set(if target.within_ancestor_pitr {
4624 0 : timeline.metrics.current_logical_size_gauge.get()
4625 : } else {
4626 8 : 0
4627 : });
4628 8 : timeline.metrics.pitr_history_size.set(
4629 8 : timeline
4630 8 : .get_last_record_lsn()
4631 8 : .checked_sub(target.cutoffs.time)
4632 8 : .unwrap_or(Lsn(0))
4633 8 : .0,
4634 8 : );
4635 :
4636 : // Apply the cutoffs we found to the Timeline's GcInfo. Why might we _not_ have cutoffs for a timeline?
4637 : // - this timeline was created while we were finding cutoffs
4638 : // - lsn for timestamp search fails for this timeline repeatedly
4639 8 : if let Some(cutoffs) = gc_cutoffs.get(&timeline.timeline_id) {
4640 8 : let original_cutoffs = target.cutoffs.clone();
4641 8 : // GC cutoffs should never go back
4642 8 : target.cutoffs = GcCutoffs {
4643 8 : space: Lsn(cutoffs.space.0.max(original_cutoffs.space.0)),
4644 8 : time: Lsn(cutoffs.time.0.max(original_cutoffs.time.0)),
4645 8 : }
4646 0 : }
4647 : }
4648 :
4649 8 : gc_timelines.push(timeline);
4650 : }
4651 8 : drop(gc_cs);
4652 8 : Ok(gc_timelines)
4653 8 : }
4654 :
4655 : /// A substitute for `branch_timeline` for use in unit tests.
4656 : /// The returned timeline will have state value `Active` to make various `anyhow::ensure!()`
4657 : /// calls pass, but, we do not actually call `.activate()` under the hood. So, none of the
4658 : /// timeline background tasks are launched, except the flush loop.
4659 : #[cfg(test)]
4660 464 : async fn branch_timeline_test(
4661 464 : self: &Arc<Self>,
4662 464 : src_timeline: &Arc<Timeline>,
4663 464 : dst_id: TimelineId,
4664 464 : ancestor_lsn: Option<Lsn>,
4665 464 : ctx: &RequestContext,
4666 464 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
4667 464 : let tl = self
4668 464 : .branch_timeline_impl(src_timeline, dst_id, ancestor_lsn, ctx)
4669 464 : .await?
4670 456 : .into_timeline_for_test();
4671 456 : tl.set_state(TimelineState::Active);
4672 456 : Ok(tl)
4673 464 : }
4674 :
4675 : /// Helper for unit tests to branch a timeline with some pre-loaded states.
4676 : #[cfg(test)]
4677 : #[allow(clippy::too_many_arguments)]
4678 12 : pub async fn branch_timeline_test_with_layers(
4679 12 : self: &Arc<Self>,
4680 12 : src_timeline: &Arc<Timeline>,
4681 12 : dst_id: TimelineId,
4682 12 : ancestor_lsn: Option<Lsn>,
4683 12 : ctx: &RequestContext,
4684 12 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
4685 12 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
4686 12 : end_lsn: Lsn,
4687 12 : ) -> anyhow::Result<Arc<Timeline>> {
4688 : use checks::check_valid_layermap;
4689 : use itertools::Itertools;
4690 :
4691 12 : let tline = self
4692 12 : .branch_timeline_test(src_timeline, dst_id, ancestor_lsn, ctx)
4693 12 : .await?;
4694 12 : let ancestor_lsn = if let Some(ancestor_lsn) = ancestor_lsn {
4695 12 : ancestor_lsn
4696 : } else {
4697 0 : tline.get_last_record_lsn()
4698 : };
4699 12 : assert!(end_lsn >= ancestor_lsn);
4700 12 : tline.force_advance_lsn(end_lsn);
4701 24 : for deltas in delta_layer_desc {
4702 12 : tline
4703 12 : .force_create_delta_layer(deltas, Some(ancestor_lsn), ctx)
4704 12 : .await?;
4705 : }
4706 20 : for (lsn, images) in image_layer_desc {
4707 8 : tline
4708 8 : .force_create_image_layer(lsn, images, Some(ancestor_lsn), ctx)
4709 8 : .await?;
4710 : }
4711 12 : let layer_names = tline
4712 12 : .layers
4713 12 : .read()
4714 12 : .await
4715 12 : .layer_map()
4716 12 : .unwrap()
4717 12 : .iter_historic_layers()
4718 20 : .map(|layer| layer.layer_name())
4719 12 : .collect_vec();
4720 12 : if let Some(err) = check_valid_layermap(&layer_names) {
4721 0 : bail!("invalid layermap: {err}");
4722 12 : }
4723 12 : Ok(tline)
4724 12 : }
4725 :
4726 : /// Branch an existing timeline.
4727 0 : async fn branch_timeline(
4728 0 : self: &Arc<Self>,
4729 0 : src_timeline: &Arc<Timeline>,
4730 0 : dst_id: TimelineId,
4731 0 : start_lsn: Option<Lsn>,
4732 0 : ctx: &RequestContext,
4733 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4734 0 : self.branch_timeline_impl(src_timeline, dst_id, start_lsn, ctx)
4735 0 : .await
4736 0 : }
4737 :
4738 464 : async fn branch_timeline_impl(
4739 464 : self: &Arc<Self>,
4740 464 : src_timeline: &Arc<Timeline>,
4741 464 : dst_id: TimelineId,
4742 464 : start_lsn: Option<Lsn>,
4743 464 : _ctx: &RequestContext,
4744 464 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4745 464 : let src_id = src_timeline.timeline_id;
4746 :
4747 : // We will validate our ancestor LSN in this function. Acquire the GC lock so that
4748 : // this check cannot race with GC, and the ancestor LSN is guaranteed to remain
4749 : // valid while we are creating the branch.
4750 464 : let _gc_cs = self.gc_cs.lock().await;
4751 :
4752 : // If no start LSN is specified, we branch the new timeline from the source timeline's last record LSN
4753 464 : let start_lsn = start_lsn.unwrap_or_else(|| {
4754 4 : let lsn = src_timeline.get_last_record_lsn();
4755 4 : info!("branching timeline {dst_id} from timeline {src_id} at last record LSN: {lsn}");
4756 4 : lsn
4757 464 : });
4758 :
4759 : // we finally have determined the ancestor_start_lsn, so we can get claim exclusivity now
4760 464 : let timeline_create_guard = match self
4761 464 : .start_creating_timeline(
4762 464 : dst_id,
4763 464 : CreateTimelineIdempotency::Branch {
4764 464 : ancestor_timeline_id: src_timeline.timeline_id,
4765 464 : ancestor_start_lsn: start_lsn,
4766 464 : },
4767 464 : )
4768 464 : .await?
4769 : {
4770 464 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
4771 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
4772 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
4773 : }
4774 : };
4775 :
4776 : // Ensure that `start_lsn` is valid, i.e. the LSN is within the PITR
4777 : // horizon on the source timeline
4778 : //
4779 : // We check it against both the planned GC cutoff stored in 'gc_info',
4780 : // and the 'latest_gc_cutoff' of the last GC that was performed. The
4781 : // planned GC cutoff in 'gc_info' is normally larger than
4782 : // 'applied_gc_cutoff_lsn', but beware of corner cases like if you just
4783 : // changed the GC settings for the tenant to make the PITR window
4784 : // larger, but some of the data was already removed by an earlier GC
4785 : // iteration.
4786 :
4787 : // check against last actual 'latest_gc_cutoff' first
4788 464 : let applied_gc_cutoff_lsn = src_timeline.get_applied_gc_cutoff_lsn();
4789 464 : {
4790 464 : let gc_info = src_timeline.gc_info.read().unwrap();
4791 464 : let planned_cutoff = gc_info.min_cutoff();
4792 464 : if gc_info.lsn_covered_by_lease(start_lsn) {
4793 0 : tracing::info!("skipping comparison of {start_lsn} with gc cutoff {} and planned gc cutoff {planned_cutoff} due to lsn lease", *applied_gc_cutoff_lsn);
4794 : } else {
4795 464 : src_timeline
4796 464 : .check_lsn_is_in_scope(start_lsn, &applied_gc_cutoff_lsn)
4797 464 : .context(format!(
4798 464 : "invalid branch start lsn: less than latest GC cutoff {}",
4799 464 : *applied_gc_cutoff_lsn,
4800 464 : ))
4801 464 : .map_err(CreateTimelineError::AncestorLsn)?;
4802 :
4803 : // and then the planned GC cutoff
4804 456 : if start_lsn < planned_cutoff {
4805 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
4806 0 : "invalid branch start lsn: less than planned GC cutoff {planned_cutoff}"
4807 0 : )));
4808 456 : }
4809 : }
4810 : }
4811 :
4812 : //
4813 : // The branch point is valid, and we are still holding the 'gc_cs' lock
4814 : // so that GC cannot advance the GC cutoff until we are finished.
4815 : // Proceed with the branch creation.
4816 : //
4817 :
4818 : // Determine prev-LSN for the new timeline. We can only determine it if
4819 : // the timeline was branched at the current end of the source timeline.
4820 : let RecordLsn {
4821 456 : last: src_last,
4822 456 : prev: src_prev,
4823 456 : } = src_timeline.get_last_record_rlsn();
4824 456 : let dst_prev = if src_last == start_lsn {
4825 432 : Some(src_prev)
4826 : } else {
4827 24 : None
4828 : };
4829 :
4830 : // Create the metadata file, noting the ancestor of the new timeline.
4831 : // There is initially no data in it, but all the read-calls know to look
4832 : // into the ancestor.
4833 456 : let metadata = TimelineMetadata::new(
4834 456 : start_lsn,
4835 456 : dst_prev,
4836 456 : Some(src_id),
4837 456 : start_lsn,
4838 456 : *src_timeline.applied_gc_cutoff_lsn.read(), // FIXME: should we hold onto this guard longer?
4839 456 : src_timeline.initdb_lsn,
4840 456 : src_timeline.pg_version,
4841 456 : );
4842 :
4843 456 : let uninitialized_timeline = self
4844 456 : .prepare_new_timeline(
4845 456 : dst_id,
4846 456 : &metadata,
4847 456 : timeline_create_guard,
4848 456 : start_lsn + 1,
4849 456 : Some(Arc::clone(src_timeline)),
4850 456 : )
4851 456 : .await?;
4852 :
4853 456 : let new_timeline = uninitialized_timeline.finish_creation().await?;
4854 :
4855 : // Root timeline gets its layers during creation and uploads them along with the metadata.
4856 : // A branch timeline though, when created, can get no writes for some time, hence won't get any layers created.
4857 : // We still need to upload its metadata eagerly: if other nodes `attach` the tenant and miss this timeline, their GC
4858 : // could get incorrect information and remove more layers, than needed.
4859 : // See also https://github.com/neondatabase/neon/issues/3865
4860 456 : new_timeline
4861 456 : .remote_client
4862 456 : .schedule_index_upload_for_full_metadata_update(&metadata)
4863 456 : .context("branch initial metadata upload")?;
4864 :
4865 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
4866 :
4867 456 : Ok(CreateTimelineResult::Created(new_timeline))
4868 464 : }
4869 :
4870 : /// For unit tests, make this visible so that other modules can directly create timelines
4871 : #[cfg(test)]
4872 : #[tracing::instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
4873 : pub(crate) async fn bootstrap_timeline_test(
4874 : self: &Arc<Self>,
4875 : timeline_id: TimelineId,
4876 : pg_version: u32,
4877 : load_existing_initdb: Option<TimelineId>,
4878 : ctx: &RequestContext,
4879 : ) -> anyhow::Result<Arc<Timeline>> {
4880 : self.bootstrap_timeline(timeline_id, pg_version, load_existing_initdb, ctx)
4881 : .await
4882 : .map_err(anyhow::Error::new)
4883 4 : .map(|r| r.into_timeline_for_test())
4884 : }
4885 :
4886 : /// Get exclusive access to the timeline ID for creation.
4887 : ///
4888 : /// Timeline-creating code paths must use this function before making changes
4889 : /// to in-memory or persistent state.
4890 : ///
4891 : /// The `state` parameter is a description of the timeline creation operation
4892 : /// we intend to perform.
4893 : /// If the timeline was already created in the meantime, we check whether this
4894 : /// request conflicts or is idempotent , based on `state`.
4895 896 : async fn start_creating_timeline(
4896 896 : self: &Arc<Self>,
4897 896 : new_timeline_id: TimelineId,
4898 896 : idempotency: CreateTimelineIdempotency,
4899 896 : ) -> Result<StartCreatingTimelineResult, CreateTimelineError> {
4900 896 : let allow_offloaded = false;
4901 896 : match self.create_timeline_create_guard(new_timeline_id, idempotency, allow_offloaded) {
4902 892 : Ok(create_guard) => {
4903 892 : pausable_failpoint!("timeline-creation-after-uninit");
4904 892 : Ok(StartCreatingTimelineResult::CreateGuard(create_guard))
4905 : }
4906 0 : Err(TimelineExclusionError::ShuttingDown) => Err(CreateTimelineError::ShuttingDown),
4907 : Err(TimelineExclusionError::AlreadyCreating) => {
4908 : // Creation is in progress, we cannot create it again, and we cannot
4909 : // check if this request matches the existing one, so caller must try
4910 : // again later.
4911 0 : Err(CreateTimelineError::AlreadyCreating)
4912 : }
4913 0 : Err(TimelineExclusionError::Other(e)) => Err(CreateTimelineError::Other(e)),
4914 : Err(TimelineExclusionError::AlreadyExists {
4915 0 : existing: TimelineOrOffloaded::Offloaded(_existing),
4916 0 : ..
4917 0 : }) => {
4918 0 : info!("timeline already exists but is offloaded");
4919 0 : Err(CreateTimelineError::Conflict)
4920 : }
4921 : Err(TimelineExclusionError::AlreadyExists {
4922 4 : existing: TimelineOrOffloaded::Timeline(existing),
4923 4 : arg,
4924 4 : }) => {
4925 4 : {
4926 4 : let existing = &existing.create_idempotency;
4927 4 : let _span = info_span!("idempotency_check", ?existing, ?arg).entered();
4928 4 : debug!("timeline already exists");
4929 :
4930 4 : match (existing, &arg) {
4931 : // FailWithConflict => no idempotency check
4932 : (CreateTimelineIdempotency::FailWithConflict, _)
4933 : | (_, CreateTimelineIdempotency::FailWithConflict) => {
4934 4 : warn!("timeline already exists, failing request");
4935 4 : return Err(CreateTimelineError::Conflict);
4936 : }
4937 : // Idempotent <=> CreateTimelineIdempotency is identical
4938 0 : (x, y) if x == y => {
4939 0 : info!("timeline already exists and idempotency matches, succeeding request");
4940 : // fallthrough
4941 : }
4942 : (_, _) => {
4943 0 : warn!("idempotency conflict, failing request");
4944 0 : return Err(CreateTimelineError::Conflict);
4945 : }
4946 : }
4947 : }
4948 :
4949 0 : Ok(StartCreatingTimelineResult::Idempotent(existing))
4950 : }
4951 : }
4952 896 : }
4953 :
4954 0 : async fn upload_initdb(
4955 0 : &self,
4956 0 : timelines_path: &Utf8PathBuf,
4957 0 : pgdata_path: &Utf8PathBuf,
4958 0 : timeline_id: &TimelineId,
4959 0 : ) -> anyhow::Result<()> {
4960 0 : let temp_path = timelines_path.join(format!(
4961 0 : "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
4962 0 : ));
4963 0 :
4964 0 : scopeguard::defer! {
4965 0 : if let Err(e) = fs::remove_file(&temp_path) {
4966 0 : error!("Failed to remove temporary initdb archive '{temp_path}': {e}");
4967 0 : }
4968 0 : }
4969 :
4970 0 : let (pgdata_zstd, tar_zst_size) = create_zst_tarball(pgdata_path, &temp_path).await?;
4971 : const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
4972 0 : if tar_zst_size > INITDB_TAR_ZST_WARN_LIMIT {
4973 0 : warn!(
4974 0 : "compressed {temp_path} size of {tar_zst_size} is above limit {INITDB_TAR_ZST_WARN_LIMIT}."
4975 : );
4976 0 : }
4977 :
4978 0 : pausable_failpoint!("before-initdb-upload");
4979 :
4980 0 : backoff::retry(
4981 0 : || async {
4982 0 : self::remote_timeline_client::upload_initdb_dir(
4983 0 : &self.remote_storage,
4984 0 : &self.tenant_shard_id.tenant_id,
4985 0 : timeline_id,
4986 0 : pgdata_zstd.try_clone().await?,
4987 0 : tar_zst_size,
4988 0 : &self.cancel,
4989 0 : )
4990 0 : .await
4991 0 : },
4992 0 : |_| false,
4993 0 : 3,
4994 0 : u32::MAX,
4995 0 : "persist_initdb_tar_zst",
4996 0 : &self.cancel,
4997 0 : )
4998 0 : .await
4999 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
5000 0 : .and_then(|x| x)
5001 0 : }
5002 :
5003 : /// - run initdb to init temporary instance and get bootstrap data
5004 : /// - after initialization completes, tar up the temp dir and upload it to S3.
5005 4 : async fn bootstrap_timeline(
5006 4 : self: &Arc<Self>,
5007 4 : timeline_id: TimelineId,
5008 4 : pg_version: u32,
5009 4 : load_existing_initdb: Option<TimelineId>,
5010 4 : ctx: &RequestContext,
5011 4 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
5012 4 : let timeline_create_guard = match self
5013 4 : .start_creating_timeline(
5014 4 : timeline_id,
5015 4 : CreateTimelineIdempotency::Bootstrap { pg_version },
5016 4 : )
5017 4 : .await?
5018 : {
5019 4 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
5020 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
5021 0 : return Ok(CreateTimelineResult::Idempotent(timeline))
5022 : }
5023 : };
5024 :
5025 : // create a `tenant/{tenant_id}/timelines/basebackup-{timeline_id}.{TEMP_FILE_SUFFIX}/`
5026 : // temporary directory for basebackup files for the given timeline.
5027 :
5028 4 : let timelines_path = self.conf.timelines_path(&self.tenant_shard_id);
5029 4 : let pgdata_path = path_with_suffix_extension(
5030 4 : timelines_path.join(format!("basebackup-{timeline_id}")),
5031 4 : TEMP_FILE_SUFFIX,
5032 4 : );
5033 4 :
5034 4 : // Remove whatever was left from the previous runs: safe because TimelineCreateGuard guarantees
5035 4 : // we won't race with other creations or existent timelines with the same path.
5036 4 : if pgdata_path.exists() {
5037 0 : fs::remove_dir_all(&pgdata_path).with_context(|| {
5038 0 : format!("Failed to remove already existing initdb directory: {pgdata_path}")
5039 0 : })?;
5040 4 : }
5041 :
5042 : // this new directory is very temporary, set to remove it immediately after bootstrap, we don't need it
5043 4 : let pgdata_path_deferred = pgdata_path.clone();
5044 4 : scopeguard::defer! {
5045 4 : if let Err(e) = fs::remove_dir_all(&pgdata_path_deferred) {
5046 4 : // this is unlikely, but we will remove the directory on pageserver restart or another bootstrap call
5047 4 : error!("Failed to remove temporary initdb directory '{pgdata_path_deferred}': {e}");
5048 4 : }
5049 4 : }
5050 4 : if let Some(existing_initdb_timeline_id) = load_existing_initdb {
5051 4 : if existing_initdb_timeline_id != timeline_id {
5052 0 : let source_path = &remote_initdb_archive_path(
5053 0 : &self.tenant_shard_id.tenant_id,
5054 0 : &existing_initdb_timeline_id,
5055 0 : );
5056 0 : let dest_path =
5057 0 : &remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &timeline_id);
5058 0 :
5059 0 : // if this fails, it will get retried by retried control plane requests
5060 0 : self.remote_storage
5061 0 : .copy_object(source_path, dest_path, &self.cancel)
5062 0 : .await
5063 0 : .context("copy initdb tar")?;
5064 4 : }
5065 4 : let (initdb_tar_zst_path, initdb_tar_zst) =
5066 4 : self::remote_timeline_client::download_initdb_tar_zst(
5067 4 : self.conf,
5068 4 : &self.remote_storage,
5069 4 : &self.tenant_shard_id,
5070 4 : &existing_initdb_timeline_id,
5071 4 : &self.cancel,
5072 4 : )
5073 4 : .await
5074 4 : .context("download initdb tar")?;
5075 :
5076 4 : scopeguard::defer! {
5077 4 : if let Err(e) = fs::remove_file(&initdb_tar_zst_path) {
5078 4 : error!("Failed to remove temporary initdb archive '{initdb_tar_zst_path}': {e}");
5079 4 : }
5080 4 : }
5081 4 :
5082 4 : let buf_read =
5083 4 : BufReader::with_capacity(remote_timeline_client::BUFFER_SIZE, initdb_tar_zst);
5084 4 : extract_zst_tarball(&pgdata_path, buf_read)
5085 4 : .await
5086 4 : .context("extract initdb tar")?;
5087 : } else {
5088 : // Init temporarily repo to get bootstrap data, this creates a directory in the `pgdata_path` path
5089 0 : run_initdb(self.conf, &pgdata_path, pg_version, &self.cancel)
5090 0 : .await
5091 0 : .context("run initdb")?;
5092 :
5093 : // Upload the created data dir to S3
5094 0 : if self.tenant_shard_id().is_shard_zero() {
5095 0 : self.upload_initdb(&timelines_path, &pgdata_path, &timeline_id)
5096 0 : .await?;
5097 0 : }
5098 : }
5099 4 : let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
5100 4 :
5101 4 : // Import the contents of the data directory at the initial checkpoint
5102 4 : // LSN, and any WAL after that.
5103 4 : // Initdb lsn will be equal to last_record_lsn which will be set after import.
5104 4 : // Because we know it upfront avoid having an option or dummy zero value by passing it to the metadata.
5105 4 : let new_metadata = TimelineMetadata::new(
5106 4 : Lsn(0),
5107 4 : None,
5108 4 : None,
5109 4 : Lsn(0),
5110 4 : pgdata_lsn,
5111 4 : pgdata_lsn,
5112 4 : pg_version,
5113 4 : );
5114 4 : let mut raw_timeline = self
5115 4 : .prepare_new_timeline(
5116 4 : timeline_id,
5117 4 : &new_metadata,
5118 4 : timeline_create_guard,
5119 4 : pgdata_lsn,
5120 4 : None,
5121 4 : )
5122 4 : .await?;
5123 :
5124 4 : let tenant_shard_id = raw_timeline.owning_tenant.tenant_shard_id;
5125 4 : raw_timeline
5126 4 : .write(|unfinished_timeline| async move {
5127 4 : import_datadir::import_timeline_from_postgres_datadir(
5128 4 : &unfinished_timeline,
5129 4 : &pgdata_path,
5130 4 : pgdata_lsn,
5131 4 : ctx,
5132 4 : )
5133 4 : .await
5134 4 : .with_context(|| {
5135 0 : format!(
5136 0 : "Failed to import pgdatadir for timeline {tenant_shard_id}/{timeline_id}"
5137 0 : )
5138 4 : })?;
5139 :
5140 4 : fail::fail_point!("before-checkpoint-new-timeline", |_| {
5141 0 : Err(CreateTimelineError::Other(anyhow::anyhow!(
5142 0 : "failpoint before-checkpoint-new-timeline"
5143 0 : )))
5144 4 : });
5145 :
5146 4 : Ok(())
5147 8 : })
5148 4 : .await?;
5149 :
5150 : // All done!
5151 4 : let timeline = raw_timeline.finish_creation().await?;
5152 :
5153 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
5154 :
5155 4 : Ok(CreateTimelineResult::Created(timeline))
5156 4 : }
5157 :
5158 884 : fn build_timeline_remote_client(&self, timeline_id: TimelineId) -> RemoteTimelineClient {
5159 884 : RemoteTimelineClient::new(
5160 884 : self.remote_storage.clone(),
5161 884 : self.deletion_queue_client.clone(),
5162 884 : self.conf,
5163 884 : self.tenant_shard_id,
5164 884 : timeline_id,
5165 884 : self.generation,
5166 884 : &self.tenant_conf.load().location,
5167 884 : )
5168 884 : }
5169 :
5170 : /// Builds required resources for a new timeline.
5171 884 : fn build_timeline_resources(&self, timeline_id: TimelineId) -> TimelineResources {
5172 884 : let remote_client = self.build_timeline_remote_client(timeline_id);
5173 884 : self.get_timeline_resources_for(remote_client)
5174 884 : }
5175 :
5176 : /// Builds timeline resources for the given remote client.
5177 896 : fn get_timeline_resources_for(&self, remote_client: RemoteTimelineClient) -> TimelineResources {
5178 896 : TimelineResources {
5179 896 : remote_client,
5180 896 : pagestream_throttle: self.pagestream_throttle.clone(),
5181 896 : pagestream_throttle_metrics: self.pagestream_throttle_metrics.clone(),
5182 896 : l0_compaction_trigger: self.l0_compaction_trigger.clone(),
5183 896 : l0_flush_global_state: self.l0_flush_global_state.clone(),
5184 896 : }
5185 896 : }
5186 :
5187 : /// Creates intermediate timeline structure and its files.
5188 : ///
5189 : /// An empty layer map is initialized, and new data and WAL can be imported starting
5190 : /// at 'disk_consistent_lsn'. After any initial data has been imported, call
5191 : /// `finish_creation` to insert the Timeline into the timelines map.
5192 884 : async fn prepare_new_timeline<'a>(
5193 884 : &'a self,
5194 884 : new_timeline_id: TimelineId,
5195 884 : new_metadata: &TimelineMetadata,
5196 884 : create_guard: TimelineCreateGuard,
5197 884 : start_lsn: Lsn,
5198 884 : ancestor: Option<Arc<Timeline>>,
5199 884 : ) -> anyhow::Result<UninitializedTimeline<'a>> {
5200 884 : let tenant_shard_id = self.tenant_shard_id;
5201 884 :
5202 884 : let resources = self.build_timeline_resources(new_timeline_id);
5203 884 : resources
5204 884 : .remote_client
5205 884 : .init_upload_queue_for_empty_remote(new_metadata)?;
5206 :
5207 884 : let timeline_struct = self
5208 884 : .create_timeline_struct(
5209 884 : new_timeline_id,
5210 884 : new_metadata,
5211 884 : None,
5212 884 : ancestor,
5213 884 : resources,
5214 884 : CreateTimelineCause::Load,
5215 884 : create_guard.idempotency.clone(),
5216 884 : )
5217 884 : .context("Failed to create timeline data structure")?;
5218 :
5219 884 : timeline_struct.init_empty_layer_map(start_lsn);
5220 :
5221 884 : if let Err(e) = self
5222 884 : .create_timeline_files(&create_guard.timeline_path)
5223 884 : .await
5224 : {
5225 0 : error!("Failed to create initial files for timeline {tenant_shard_id}/{new_timeline_id}, cleaning up: {e:?}");
5226 0 : cleanup_timeline_directory(create_guard);
5227 0 : return Err(e);
5228 884 : }
5229 884 :
5230 884 : debug!(
5231 0 : "Successfully created initial files for timeline {tenant_shard_id}/{new_timeline_id}"
5232 : );
5233 :
5234 884 : Ok(UninitializedTimeline::new(
5235 884 : self,
5236 884 : new_timeline_id,
5237 884 : Some((timeline_struct, create_guard)),
5238 884 : ))
5239 884 : }
5240 :
5241 884 : async fn create_timeline_files(&self, timeline_path: &Utf8Path) -> anyhow::Result<()> {
5242 884 : crashsafe::create_dir(timeline_path).context("Failed to create timeline directory")?;
5243 :
5244 884 : fail::fail_point!("after-timeline-dir-creation", |_| {
5245 0 : anyhow::bail!("failpoint after-timeline-dir-creation");
5246 884 : });
5247 :
5248 884 : Ok(())
5249 884 : }
5250 :
5251 : /// Get a guard that provides exclusive access to the timeline directory, preventing
5252 : /// concurrent attempts to create the same timeline.
5253 : ///
5254 : /// The `allow_offloaded` parameter controls whether to tolerate the existence of
5255 : /// offloaded timelines or not.
5256 896 : fn create_timeline_create_guard(
5257 896 : self: &Arc<Self>,
5258 896 : timeline_id: TimelineId,
5259 896 : idempotency: CreateTimelineIdempotency,
5260 896 : allow_offloaded: bool,
5261 896 : ) -> Result<TimelineCreateGuard, TimelineExclusionError> {
5262 896 : let tenant_shard_id = self.tenant_shard_id;
5263 896 :
5264 896 : let timeline_path = self.conf.timeline_path(&tenant_shard_id, &timeline_id);
5265 :
5266 896 : let create_guard = TimelineCreateGuard::new(
5267 896 : self,
5268 896 : timeline_id,
5269 896 : timeline_path.clone(),
5270 896 : idempotency,
5271 896 : allow_offloaded,
5272 896 : )?;
5273 :
5274 : // At this stage, we have got exclusive access to in-memory state for this timeline ID
5275 : // for creation.
5276 : // A timeline directory should never exist on disk already:
5277 : // - a previous failed creation would have cleaned up after itself
5278 : // - a pageserver restart would clean up timeline directories that don't have valid remote state
5279 : //
5280 : // Therefore it is an unexpected internal error to encounter a timeline directory already existing here,
5281 : // this error may indicate a bug in cleanup on failed creations.
5282 892 : if timeline_path.exists() {
5283 0 : return Err(TimelineExclusionError::Other(anyhow::anyhow!(
5284 0 : "Timeline directory already exists! This is a bug."
5285 0 : )));
5286 892 : }
5287 892 :
5288 892 : Ok(create_guard)
5289 896 : }
5290 :
5291 : /// Gathers inputs from all of the timelines to produce a sizing model input.
5292 : ///
5293 : /// Future is cancellation safe. Only one calculation can be running at once per tenant.
5294 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5295 : pub async fn gather_size_inputs(
5296 : &self,
5297 : // `max_retention_period` overrides the cutoff that is used to calculate the size
5298 : // (only if it is shorter than the real cutoff).
5299 : max_retention_period: Option<u64>,
5300 : cause: LogicalSizeCalculationCause,
5301 : cancel: &CancellationToken,
5302 : ctx: &RequestContext,
5303 : ) -> Result<size::ModelInputs, size::CalculateSyntheticSizeError> {
5304 : let logical_sizes_at_once = self
5305 : .conf
5306 : .concurrent_tenant_size_logical_size_queries
5307 : .inner();
5308 :
5309 : // TODO: Having a single mutex block concurrent reads is not great for performance.
5310 : //
5311 : // But the only case where we need to run multiple of these at once is when we
5312 : // request a size for a tenant manually via API, while another background calculation
5313 : // is in progress (which is not a common case).
5314 : //
5315 : // See more for on the issue #2748 condenced out of the initial PR review.
5316 : let mut shared_cache = tokio::select! {
5317 : locked = self.cached_logical_sizes.lock() => locked,
5318 : _ = cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5319 : _ = self.cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5320 : };
5321 :
5322 : size::gather_inputs(
5323 : self,
5324 : logical_sizes_at_once,
5325 : max_retention_period,
5326 : &mut shared_cache,
5327 : cause,
5328 : cancel,
5329 : ctx,
5330 : )
5331 : .await
5332 : }
5333 :
5334 : /// Calculate synthetic tenant size and cache the result.
5335 : /// This is periodically called by background worker.
5336 : /// result is cached in tenant struct
5337 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5338 : pub async fn calculate_synthetic_size(
5339 : &self,
5340 : cause: LogicalSizeCalculationCause,
5341 : cancel: &CancellationToken,
5342 : ctx: &RequestContext,
5343 : ) -> Result<u64, size::CalculateSyntheticSizeError> {
5344 : let inputs = self.gather_size_inputs(None, cause, cancel, ctx).await?;
5345 :
5346 : let size = inputs.calculate();
5347 :
5348 : self.set_cached_synthetic_size(size);
5349 :
5350 : Ok(size)
5351 : }
5352 :
5353 : /// Cache given synthetic size and update the metric value
5354 0 : pub fn set_cached_synthetic_size(&self, size: u64) {
5355 0 : self.cached_synthetic_tenant_size
5356 0 : .store(size, Ordering::Relaxed);
5357 0 :
5358 0 : // Only shard zero should be calculating synthetic sizes
5359 0 : debug_assert!(self.shard_identity.is_shard_zero());
5360 :
5361 0 : TENANT_SYNTHETIC_SIZE_METRIC
5362 0 : .get_metric_with_label_values(&[&self.tenant_shard_id.tenant_id.to_string()])
5363 0 : .unwrap()
5364 0 : .set(size);
5365 0 : }
5366 :
5367 0 : pub fn cached_synthetic_size(&self) -> u64 {
5368 0 : self.cached_synthetic_tenant_size.load(Ordering::Relaxed)
5369 0 : }
5370 :
5371 : /// Flush any in-progress layers, schedule uploads, and wait for uploads to complete.
5372 : ///
5373 : /// This function can take a long time: callers should wrap it in a timeout if calling
5374 : /// from an external API handler.
5375 : ///
5376 : /// Cancel-safety: cancelling this function may leave I/O running, but such I/O is
5377 : /// still bounded by tenant/timeline shutdown.
5378 : #[tracing::instrument(skip_all)]
5379 : pub(crate) async fn flush_remote(&self) -> anyhow::Result<()> {
5380 : let timelines = self.timelines.lock().unwrap().clone();
5381 :
5382 0 : async fn flush_timeline(_gate: GateGuard, timeline: Arc<Timeline>) -> anyhow::Result<()> {
5383 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Flushing...");
5384 0 : timeline.freeze_and_flush().await?;
5385 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Waiting for uploads...");
5386 0 : timeline.remote_client.wait_completion().await?;
5387 :
5388 0 : Ok(())
5389 0 : }
5390 :
5391 : // We do not use a JoinSet for these tasks, because we don't want them to be
5392 : // aborted when this function's future is cancelled: they should stay alive
5393 : // holding their GateGuard until they complete, to ensure their I/Os complete
5394 : // before Timeline shutdown completes.
5395 : let mut results = FuturesUnordered::new();
5396 :
5397 : for (_timeline_id, timeline) in timelines {
5398 : // Run each timeline's flush in a task holding the timeline's gate: this
5399 : // means that if this function's future is cancelled, the Timeline shutdown
5400 : // will still wait for any I/O in here to complete.
5401 : let Ok(gate) = timeline.gate.enter() else {
5402 : continue;
5403 : };
5404 0 : let jh = tokio::task::spawn(async move { flush_timeline(gate, timeline).await });
5405 : results.push(jh);
5406 : }
5407 :
5408 : while let Some(r) = results.next().await {
5409 : if let Err(e) = r {
5410 : if !e.is_cancelled() && !e.is_panic() {
5411 : tracing::error!("unexpected join error: {e:?}");
5412 : }
5413 : }
5414 : }
5415 :
5416 : // The flushes we did above were just writes, but the Tenant might have had
5417 : // pending deletions as well from recent compaction/gc: we want to flush those
5418 : // as well. This requires flushing the global delete queue. This is cheap
5419 : // because it's typically a no-op.
5420 : match self.deletion_queue_client.flush_execute().await {
5421 : Ok(_) => {}
5422 : Err(DeletionQueueError::ShuttingDown) => {}
5423 : }
5424 :
5425 : Ok(())
5426 : }
5427 :
5428 0 : pub(crate) fn get_tenant_conf(&self) -> TenantConfOpt {
5429 0 : self.tenant_conf.load().tenant_conf.clone()
5430 0 : }
5431 :
5432 : /// How much local storage would this tenant like to have? It can cope with
5433 : /// less than this (via eviction and on-demand downloads), but this function enables
5434 : /// the Tenant to advertise how much storage it would prefer to have to provide fast I/O
5435 : /// by keeping important things on local disk.
5436 : ///
5437 : /// This is a heuristic, not a guarantee: tenants that are long-idle will actually use less
5438 : /// than they report here, due to layer eviction. Tenants with many active branches may
5439 : /// actually use more than they report here.
5440 0 : pub(crate) fn local_storage_wanted(&self) -> u64 {
5441 0 : let timelines = self.timelines.lock().unwrap();
5442 0 :
5443 0 : // Heuristic: we use the max() of the timelines' visible sizes, rather than the sum. This
5444 0 : // reflects the observation that on tenants with multiple large branches, typically only one
5445 0 : // of them is used actively enough to occupy space on disk.
5446 0 : timelines
5447 0 : .values()
5448 0 : .map(|t| t.metrics.visible_physical_size_gauge.get())
5449 0 : .max()
5450 0 : .unwrap_or(0)
5451 0 : }
5452 :
5453 : /// Serialize and write the latest TenantManifest to remote storage.
5454 4 : pub(crate) async fn store_tenant_manifest(&self) -> Result<(), TenantManifestError> {
5455 : // Only one manifest write may be done at at time, and the contents of the manifest
5456 : // must be loaded while holding this lock. This makes it safe to call this function
5457 : // from anywhere without worrying about colliding updates.
5458 4 : let mut guard = tokio::select! {
5459 4 : g = self.tenant_manifest_upload.lock() => {
5460 4 : g
5461 : },
5462 4 : _ = self.cancel.cancelled() => {
5463 0 : return Err(TenantManifestError::Cancelled);
5464 : }
5465 : };
5466 :
5467 4 : let manifest = self.build_tenant_manifest();
5468 4 : if Some(&manifest) == (*guard).as_ref() {
5469 : // Optimisation: skip uploads that don't change anything.
5470 0 : return Ok(());
5471 4 : }
5472 4 :
5473 4 : // Remote storage does no retries internally, so wrap it
5474 4 : match backoff::retry(
5475 4 : || async {
5476 4 : upload_tenant_manifest(
5477 4 : &self.remote_storage,
5478 4 : &self.tenant_shard_id,
5479 4 : self.generation,
5480 4 : &manifest,
5481 4 : &self.cancel,
5482 4 : )
5483 4 : .await
5484 8 : },
5485 4 : |_e| self.cancel.is_cancelled(),
5486 4 : FAILED_UPLOAD_WARN_THRESHOLD,
5487 4 : FAILED_REMOTE_OP_RETRIES,
5488 4 : "uploading tenant manifest",
5489 4 : &self.cancel,
5490 4 : )
5491 4 : .await
5492 : {
5493 0 : None => Err(TenantManifestError::Cancelled),
5494 0 : Some(Err(_)) if self.cancel.is_cancelled() => Err(TenantManifestError::Cancelled),
5495 0 : Some(Err(e)) => Err(TenantManifestError::RemoteStorage(e)),
5496 : Some(Ok(_)) => {
5497 : // Store the successfully uploaded manifest, so that future callers can avoid
5498 : // re-uploading the same thing.
5499 4 : *guard = Some(manifest);
5500 4 :
5501 4 : Ok(())
5502 : }
5503 : }
5504 4 : }
5505 : }
5506 :
5507 : /// Create the cluster temporarily in 'initdbpath' directory inside the repository
5508 : /// to get bootstrap data for timeline initialization.
5509 0 : async fn run_initdb(
5510 0 : conf: &'static PageServerConf,
5511 0 : initdb_target_dir: &Utf8Path,
5512 0 : pg_version: u32,
5513 0 : cancel: &CancellationToken,
5514 0 : ) -> Result<(), InitdbError> {
5515 0 : let initdb_bin_path = conf
5516 0 : .pg_bin_dir(pg_version)
5517 0 : .map_err(InitdbError::Other)?
5518 0 : .join("initdb");
5519 0 : let initdb_lib_dir = conf.pg_lib_dir(pg_version).map_err(InitdbError::Other)?;
5520 0 : info!(
5521 0 : "running {} in {}, libdir: {}",
5522 : initdb_bin_path, initdb_target_dir, initdb_lib_dir,
5523 : );
5524 :
5525 0 : let _permit = {
5526 0 : let _timer = INITDB_SEMAPHORE_ACQUISITION_TIME.start_timer();
5527 0 : INIT_DB_SEMAPHORE.acquire().await
5528 : };
5529 :
5530 0 : CONCURRENT_INITDBS.inc();
5531 0 : scopeguard::defer! {
5532 0 : CONCURRENT_INITDBS.dec();
5533 0 : }
5534 0 :
5535 0 : let _timer = INITDB_RUN_TIME.start_timer();
5536 0 : let res = postgres_initdb::do_run_initdb(postgres_initdb::RunInitdbArgs {
5537 0 : superuser: &conf.superuser,
5538 0 : locale: &conf.locale,
5539 0 : initdb_bin: &initdb_bin_path,
5540 0 : pg_version,
5541 0 : library_search_path: &initdb_lib_dir,
5542 0 : pgdata: initdb_target_dir,
5543 0 : })
5544 0 : .await
5545 0 : .map_err(InitdbError::Inner);
5546 0 :
5547 0 : // This isn't true cancellation support, see above. Still return an error to
5548 0 : // excercise the cancellation code path.
5549 0 : if cancel.is_cancelled() {
5550 0 : return Err(InitdbError::Cancelled);
5551 0 : }
5552 0 :
5553 0 : res
5554 0 : }
5555 :
5556 : /// Dump contents of a layer file to stdout.
5557 0 : pub async fn dump_layerfile_from_path(
5558 0 : path: &Utf8Path,
5559 0 : verbose: bool,
5560 0 : ctx: &RequestContext,
5561 0 : ) -> anyhow::Result<()> {
5562 : use std::os::unix::fs::FileExt;
5563 :
5564 : // All layer files start with a two-byte "magic" value, to identify the kind of
5565 : // file.
5566 0 : let file = File::open(path)?;
5567 0 : let mut header_buf = [0u8; 2];
5568 0 : file.read_exact_at(&mut header_buf, 0)?;
5569 :
5570 0 : match u16::from_be_bytes(header_buf) {
5571 : crate::IMAGE_FILE_MAGIC => {
5572 0 : ImageLayer::new_for_path(path, file)?
5573 0 : .dump(verbose, ctx)
5574 0 : .await?
5575 : }
5576 : crate::DELTA_FILE_MAGIC => {
5577 0 : DeltaLayer::new_for_path(path, file)?
5578 0 : .dump(verbose, ctx)
5579 0 : .await?
5580 : }
5581 0 : magic => bail!("unrecognized magic identifier: {:?}", magic),
5582 : }
5583 :
5584 0 : Ok(())
5585 0 : }
5586 :
5587 : #[cfg(test)]
5588 : pub(crate) mod harness {
5589 : use bytes::{Bytes, BytesMut};
5590 : use once_cell::sync::OnceCell;
5591 : use pageserver_api::models::ShardParameters;
5592 : use pageserver_api::shard::ShardIndex;
5593 : use utils::logging;
5594 :
5595 : use crate::deletion_queue::mock::MockDeletionQueue;
5596 : use crate::l0_flush::L0FlushConfig;
5597 : use crate::walredo::apply_neon;
5598 : use pageserver_api::key::Key;
5599 : use pageserver_api::record::NeonWalRecord;
5600 :
5601 : use super::*;
5602 : use hex_literal::hex;
5603 : use utils::id::TenantId;
5604 :
5605 : pub const TIMELINE_ID: TimelineId =
5606 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
5607 : pub const NEW_TIMELINE_ID: TimelineId =
5608 : TimelineId::from_array(hex!("AA223344556677881122334455667788"));
5609 :
5610 : /// Convenience function to create a page image with given string as the only content
5611 10057571 : pub fn test_img(s: &str) -> Bytes {
5612 10057571 : let mut buf = BytesMut::new();
5613 10057571 : buf.extend_from_slice(s.as_bytes());
5614 10057571 : buf.resize(64, 0);
5615 10057571 :
5616 10057571 : buf.freeze()
5617 10057571 : }
5618 :
5619 : impl From<TenantConf> for TenantConfOpt {
5620 444 : fn from(tenant_conf: TenantConf) -> Self {
5621 444 : Self {
5622 444 : checkpoint_distance: Some(tenant_conf.checkpoint_distance),
5623 444 : checkpoint_timeout: Some(tenant_conf.checkpoint_timeout),
5624 444 : compaction_target_size: Some(tenant_conf.compaction_target_size),
5625 444 : compaction_period: Some(tenant_conf.compaction_period),
5626 444 : compaction_threshold: Some(tenant_conf.compaction_threshold),
5627 444 : compaction_upper_limit: Some(tenant_conf.compaction_upper_limit),
5628 444 : compaction_algorithm: Some(tenant_conf.compaction_algorithm),
5629 444 : compaction_l0_first: Some(tenant_conf.compaction_l0_first),
5630 444 : compaction_l0_semaphore: Some(tenant_conf.compaction_l0_semaphore),
5631 444 : l0_flush_delay_threshold: tenant_conf.l0_flush_delay_threshold,
5632 444 : l0_flush_stall_threshold: tenant_conf.l0_flush_stall_threshold,
5633 444 : l0_flush_wait_upload: Some(tenant_conf.l0_flush_wait_upload),
5634 444 : gc_horizon: Some(tenant_conf.gc_horizon),
5635 444 : gc_period: Some(tenant_conf.gc_period),
5636 444 : image_creation_threshold: Some(tenant_conf.image_creation_threshold),
5637 444 : pitr_interval: Some(tenant_conf.pitr_interval),
5638 444 : walreceiver_connect_timeout: Some(tenant_conf.walreceiver_connect_timeout),
5639 444 : lagging_wal_timeout: Some(tenant_conf.lagging_wal_timeout),
5640 444 : max_lsn_wal_lag: Some(tenant_conf.max_lsn_wal_lag),
5641 444 : eviction_policy: Some(tenant_conf.eviction_policy),
5642 444 : min_resident_size_override: tenant_conf.min_resident_size_override,
5643 444 : evictions_low_residence_duration_metric_threshold: Some(
5644 444 : tenant_conf.evictions_low_residence_duration_metric_threshold,
5645 444 : ),
5646 444 : heatmap_period: Some(tenant_conf.heatmap_period),
5647 444 : lazy_slru_download: Some(tenant_conf.lazy_slru_download),
5648 444 : timeline_get_throttle: Some(tenant_conf.timeline_get_throttle),
5649 444 : image_layer_creation_check_threshold: Some(
5650 444 : tenant_conf.image_layer_creation_check_threshold,
5651 444 : ),
5652 444 : image_creation_preempt_threshold: Some(
5653 444 : tenant_conf.image_creation_preempt_threshold,
5654 444 : ),
5655 444 : lsn_lease_length: Some(tenant_conf.lsn_lease_length),
5656 444 : lsn_lease_length_for_ts: Some(tenant_conf.lsn_lease_length_for_ts),
5657 444 : timeline_offloading: Some(tenant_conf.timeline_offloading),
5658 444 : wal_receiver_protocol_override: tenant_conf.wal_receiver_protocol_override,
5659 444 : rel_size_v2_enabled: Some(tenant_conf.rel_size_v2_enabled),
5660 444 : gc_compaction_enabled: Some(tenant_conf.gc_compaction_enabled),
5661 444 : gc_compaction_initial_threshold_kb: Some(
5662 444 : tenant_conf.gc_compaction_initial_threshold_kb,
5663 444 : ),
5664 444 : gc_compaction_ratio_percent: Some(tenant_conf.gc_compaction_ratio_percent),
5665 444 : }
5666 444 : }
5667 : }
5668 :
5669 : pub struct TenantHarness {
5670 : pub conf: &'static PageServerConf,
5671 : pub tenant_conf: TenantConf,
5672 : pub tenant_shard_id: TenantShardId,
5673 : pub generation: Generation,
5674 : pub shard: ShardIndex,
5675 : pub remote_storage: GenericRemoteStorage,
5676 : pub remote_fs_dir: Utf8PathBuf,
5677 : pub deletion_queue: MockDeletionQueue,
5678 : }
5679 :
5680 : static LOG_HANDLE: OnceCell<()> = OnceCell::new();
5681 :
5682 492 : pub(crate) fn setup_logging() {
5683 492 : LOG_HANDLE.get_or_init(|| {
5684 468 : logging::init(
5685 468 : logging::LogFormat::Test,
5686 468 : // enable it in case the tests exercise code paths that use
5687 468 : // debug_assert_current_span_has_tenant_and_timeline_id
5688 468 : logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
5689 468 : logging::Output::Stdout,
5690 468 : )
5691 468 : .expect("Failed to init test logging")
5692 492 : });
5693 492 : }
5694 :
5695 : impl TenantHarness {
5696 444 : pub async fn create_custom(
5697 444 : test_name: &'static str,
5698 444 : tenant_conf: TenantConf,
5699 444 : tenant_id: TenantId,
5700 444 : shard_identity: ShardIdentity,
5701 444 : generation: Generation,
5702 444 : ) -> anyhow::Result<Self> {
5703 444 : setup_logging();
5704 444 :
5705 444 : let repo_dir = PageServerConf::test_repo_dir(test_name);
5706 444 : let _ = fs::remove_dir_all(&repo_dir);
5707 444 : fs::create_dir_all(&repo_dir)?;
5708 :
5709 444 : let conf = PageServerConf::dummy_conf(repo_dir);
5710 444 : // Make a static copy of the config. This can never be free'd, but that's
5711 444 : // OK in a test.
5712 444 : let conf: &'static PageServerConf = Box::leak(Box::new(conf));
5713 444 :
5714 444 : let shard = shard_identity.shard_index();
5715 444 : let tenant_shard_id = TenantShardId {
5716 444 : tenant_id,
5717 444 : shard_number: shard.shard_number,
5718 444 : shard_count: shard.shard_count,
5719 444 : };
5720 444 : fs::create_dir_all(conf.tenant_path(&tenant_shard_id))?;
5721 444 : fs::create_dir_all(conf.timelines_path(&tenant_shard_id))?;
5722 :
5723 : use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
5724 444 : let remote_fs_dir = conf.workdir.join("localfs");
5725 444 : std::fs::create_dir_all(&remote_fs_dir).unwrap();
5726 444 : let config = RemoteStorageConfig {
5727 444 : storage: RemoteStorageKind::LocalFs {
5728 444 : local_path: remote_fs_dir.clone(),
5729 444 : },
5730 444 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
5731 444 : small_timeout: RemoteStorageConfig::DEFAULT_SMALL_TIMEOUT,
5732 444 : };
5733 444 : let remote_storage = GenericRemoteStorage::from_config(&config).await.unwrap();
5734 444 : let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()));
5735 444 :
5736 444 : Ok(Self {
5737 444 : conf,
5738 444 : tenant_conf,
5739 444 : tenant_shard_id,
5740 444 : generation,
5741 444 : shard,
5742 444 : remote_storage,
5743 444 : remote_fs_dir,
5744 444 : deletion_queue,
5745 444 : })
5746 444 : }
5747 :
5748 420 : pub async fn create(test_name: &'static str) -> anyhow::Result<Self> {
5749 420 : // Disable automatic GC and compaction to make the unit tests more deterministic.
5750 420 : // The tests perform them manually if needed.
5751 420 : let tenant_conf = TenantConf {
5752 420 : gc_period: Duration::ZERO,
5753 420 : compaction_period: Duration::ZERO,
5754 420 : ..TenantConf::default()
5755 420 : };
5756 420 : let tenant_id = TenantId::generate();
5757 420 : let shard = ShardIdentity::unsharded();
5758 420 : Self::create_custom(
5759 420 : test_name,
5760 420 : tenant_conf,
5761 420 : tenant_id,
5762 420 : shard,
5763 420 : Generation::new(0xdeadbeef),
5764 420 : )
5765 420 : .await
5766 420 : }
5767 :
5768 40 : pub fn span(&self) -> tracing::Span {
5769 40 : info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
5770 40 : }
5771 :
5772 444 : pub(crate) async fn load(&self) -> (Arc<Tenant>, RequestContext) {
5773 444 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
5774 444 : (
5775 444 : self.do_try_load(&ctx)
5776 444 : .await
5777 444 : .expect("failed to load test tenant"),
5778 444 : ctx,
5779 444 : )
5780 444 : }
5781 :
5782 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5783 : pub(crate) async fn do_try_load(
5784 : &self,
5785 : ctx: &RequestContext,
5786 : ) -> anyhow::Result<Arc<Tenant>> {
5787 : let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
5788 :
5789 : let tenant = Arc::new(Tenant::new(
5790 : TenantState::Attaching,
5791 : self.conf,
5792 : AttachedTenantConf::try_from(LocationConf::attached_single(
5793 : TenantConfOpt::from(self.tenant_conf.clone()),
5794 : self.generation,
5795 : &ShardParameters::default(),
5796 : ))
5797 : .unwrap(),
5798 : // This is a legacy/test code path: sharding isn't supported here.
5799 : ShardIdentity::unsharded(),
5800 : Some(walredo_mgr),
5801 : self.tenant_shard_id,
5802 : self.remote_storage.clone(),
5803 : self.deletion_queue.new_client(),
5804 : // TODO: ideally we should run all unit tests with both configs
5805 : L0FlushGlobalState::new(L0FlushConfig::default()),
5806 : ));
5807 :
5808 : let preload = tenant
5809 : .preload(&self.remote_storage, CancellationToken::new())
5810 : .await?;
5811 : tenant.attach(Some(preload), ctx).await?;
5812 :
5813 : tenant.state.send_replace(TenantState::Active);
5814 : for timeline in tenant.timelines.lock().unwrap().values() {
5815 : timeline.set_state(TimelineState::Active);
5816 : }
5817 : Ok(tenant)
5818 : }
5819 :
5820 4 : pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
5821 4 : self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
5822 4 : }
5823 : }
5824 :
5825 : // Mock WAL redo manager that doesn't do much
5826 : pub(crate) struct TestRedoManager;
5827 :
5828 : impl TestRedoManager {
5829 : /// # Cancel-Safety
5830 : ///
5831 : /// This method is cancellation-safe.
5832 1636 : pub async fn request_redo(
5833 1636 : &self,
5834 1636 : key: Key,
5835 1636 : lsn: Lsn,
5836 1636 : base_img: Option<(Lsn, Bytes)>,
5837 1636 : records: Vec<(Lsn, NeonWalRecord)>,
5838 1636 : _pg_version: u32,
5839 1636 : ) -> Result<Bytes, walredo::Error> {
5840 2392 : let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
5841 1636 : if records_neon {
5842 : // For Neon wal records, we can decode without spawning postgres, so do so.
5843 1636 : let mut page = match (base_img, records.first()) {
5844 1504 : (Some((_lsn, img)), _) => {
5845 1504 : let mut page = BytesMut::new();
5846 1504 : page.extend_from_slice(&img);
5847 1504 : page
5848 : }
5849 132 : (_, Some((_lsn, rec))) if rec.will_init() => BytesMut::new(),
5850 : _ => {
5851 0 : panic!("Neon WAL redo requires base image or will init record");
5852 : }
5853 : };
5854 :
5855 4028 : for (record_lsn, record) in records {
5856 2392 : apply_neon::apply_in_neon(&record, record_lsn, key, &mut page)?;
5857 : }
5858 1636 : Ok(page.freeze())
5859 : } else {
5860 : // We never spawn a postgres walredo process in unit tests: just log what we might have done.
5861 0 : let s = format!(
5862 0 : "redo for {} to get to {}, with {} and {} records",
5863 0 : key,
5864 0 : lsn,
5865 0 : if base_img.is_some() {
5866 0 : "base image"
5867 : } else {
5868 0 : "no base image"
5869 : },
5870 0 : records.len()
5871 0 : );
5872 0 : println!("{s}");
5873 0 :
5874 0 : Ok(test_img(&s))
5875 : }
5876 1636 : }
5877 : }
5878 : }
5879 :
5880 : #[cfg(test)]
5881 : mod tests {
5882 : use std::collections::{BTreeMap, BTreeSet};
5883 :
5884 : use super::*;
5885 : use crate::keyspace::KeySpaceAccum;
5886 : use crate::tenant::harness::*;
5887 : use crate::tenant::timeline::CompactFlags;
5888 : use crate::DEFAULT_PG_VERSION;
5889 : use bytes::{Bytes, BytesMut};
5890 : use hex_literal::hex;
5891 : use itertools::Itertools;
5892 : use pageserver_api::key::{Key, AUX_KEY_PREFIX, NON_INHERITED_RANGE, RELATION_SIZE_PREFIX};
5893 : use pageserver_api::keyspace::KeySpace;
5894 : use pageserver_api::models::{CompactionAlgorithm, CompactionAlgorithmSettings};
5895 : use pageserver_api::value::Value;
5896 : use pageserver_compaction::helpers::overlaps_with;
5897 : use rand::{thread_rng, Rng};
5898 : use storage_layer::{IoConcurrency, PersistentLayerKey};
5899 : use tests::storage_layer::ValuesReconstructState;
5900 : use tests::timeline::{GetVectoredError, ShutdownMode};
5901 : use timeline::{CompactOptions, DeltaLayerTestDesc};
5902 : use utils::id::TenantId;
5903 :
5904 : #[cfg(feature = "testing")]
5905 : use models::CompactLsnRange;
5906 : #[cfg(feature = "testing")]
5907 : use pageserver_api::record::NeonWalRecord;
5908 : #[cfg(feature = "testing")]
5909 : use timeline::compaction::{KeyHistoryRetention, KeyLogAtLsn};
5910 : #[cfg(feature = "testing")]
5911 : use timeline::GcInfo;
5912 :
5913 : static TEST_KEY: Lazy<Key> =
5914 36 : Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));
5915 :
5916 : #[tokio::test]
5917 4 : async fn test_basic() -> anyhow::Result<()> {
5918 4 : let (tenant, ctx) = TenantHarness::create("test_basic").await?.load().await;
5919 4 : let tline = tenant
5920 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5921 4 : .await?;
5922 4 :
5923 4 : let mut writer = tline.writer().await;
5924 4 : writer
5925 4 : .put(
5926 4 : *TEST_KEY,
5927 4 : Lsn(0x10),
5928 4 : &Value::Image(test_img("foo at 0x10")),
5929 4 : &ctx,
5930 4 : )
5931 4 : .await?;
5932 4 : writer.finish_write(Lsn(0x10));
5933 4 : drop(writer);
5934 4 :
5935 4 : let mut writer = tline.writer().await;
5936 4 : writer
5937 4 : .put(
5938 4 : *TEST_KEY,
5939 4 : Lsn(0x20),
5940 4 : &Value::Image(test_img("foo at 0x20")),
5941 4 : &ctx,
5942 4 : )
5943 4 : .await?;
5944 4 : writer.finish_write(Lsn(0x20));
5945 4 : drop(writer);
5946 4 :
5947 4 : assert_eq!(
5948 4 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
5949 4 : test_img("foo at 0x10")
5950 4 : );
5951 4 : assert_eq!(
5952 4 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
5953 4 : test_img("foo at 0x10")
5954 4 : );
5955 4 : assert_eq!(
5956 4 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
5957 4 : test_img("foo at 0x20")
5958 4 : );
5959 4 :
5960 4 : Ok(())
5961 4 : }
5962 :
5963 : #[tokio::test]
5964 4 : async fn no_duplicate_timelines() -> anyhow::Result<()> {
5965 4 : let (tenant, ctx) = TenantHarness::create("no_duplicate_timelines")
5966 4 : .await?
5967 4 : .load()
5968 4 : .await;
5969 4 : let _ = tenant
5970 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5971 4 : .await?;
5972 4 :
5973 4 : match tenant
5974 4 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5975 4 : .await
5976 4 : {
5977 4 : Ok(_) => panic!("duplicate timeline creation should fail"),
5978 4 : Err(e) => assert_eq!(
5979 4 : e.to_string(),
5980 4 : "timeline already exists with different parameters".to_string()
5981 4 : ),
5982 4 : }
5983 4 :
5984 4 : Ok(())
5985 4 : }
5986 :
5987 : /// Convenience function to create a page image with given string as the only content
5988 20 : pub fn test_value(s: &str) -> Value {
5989 20 : let mut buf = BytesMut::new();
5990 20 : buf.extend_from_slice(s.as_bytes());
5991 20 : Value::Image(buf.freeze())
5992 20 : }
5993 :
5994 : ///
5995 : /// Test branch creation
5996 : ///
5997 : #[tokio::test]
5998 4 : async fn test_branch() -> anyhow::Result<()> {
5999 4 : use std::str::from_utf8;
6000 4 :
6001 4 : let (tenant, ctx) = TenantHarness::create("test_branch").await?.load().await;
6002 4 : let tline = tenant
6003 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6004 4 : .await?;
6005 4 : let mut writer = tline.writer().await;
6006 4 :
6007 4 : #[allow(non_snake_case)]
6008 4 : let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
6009 4 : #[allow(non_snake_case)]
6010 4 : let TEST_KEY_B: Key = Key::from_hex("110000000033333333444444445500000002").unwrap();
6011 4 :
6012 4 : // Insert a value on the timeline
6013 4 : writer
6014 4 : .put(TEST_KEY_A, Lsn(0x20), &test_value("foo at 0x20"), &ctx)
6015 4 : .await?;
6016 4 : writer
6017 4 : .put(TEST_KEY_B, Lsn(0x20), &test_value("foobar at 0x20"), &ctx)
6018 4 : .await?;
6019 4 : writer.finish_write(Lsn(0x20));
6020 4 :
6021 4 : writer
6022 4 : .put(TEST_KEY_A, Lsn(0x30), &test_value("foo at 0x30"), &ctx)
6023 4 : .await?;
6024 4 : writer.finish_write(Lsn(0x30));
6025 4 : writer
6026 4 : .put(TEST_KEY_A, Lsn(0x40), &test_value("foo at 0x40"), &ctx)
6027 4 : .await?;
6028 4 : writer.finish_write(Lsn(0x40));
6029 4 :
6030 4 : //assert_current_logical_size(&tline, Lsn(0x40));
6031 4 :
6032 4 : // Branch the history, modify relation differently on the new timeline
6033 4 : tenant
6034 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x30)), &ctx)
6035 4 : .await?;
6036 4 : let newtline = tenant
6037 4 : .get_timeline(NEW_TIMELINE_ID, true)
6038 4 : .expect("Should have a local timeline");
6039 4 : let mut new_writer = newtline.writer().await;
6040 4 : new_writer
6041 4 : .put(TEST_KEY_A, Lsn(0x40), &test_value("bar at 0x40"), &ctx)
6042 4 : .await?;
6043 4 : new_writer.finish_write(Lsn(0x40));
6044 4 :
6045 4 : // Check page contents on both branches
6046 4 : assert_eq!(
6047 4 : from_utf8(&tline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
6048 4 : "foo at 0x40"
6049 4 : );
6050 4 : assert_eq!(
6051 4 : from_utf8(&newtline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
6052 4 : "bar at 0x40"
6053 4 : );
6054 4 : assert_eq!(
6055 4 : from_utf8(&newtline.get(TEST_KEY_B, Lsn(0x40), &ctx).await?)?,
6056 4 : "foobar at 0x20"
6057 4 : );
6058 4 :
6059 4 : //assert_current_logical_size(&tline, Lsn(0x40));
6060 4 :
6061 4 : Ok(())
6062 4 : }
6063 :
6064 40 : async fn make_some_layers(
6065 40 : tline: &Timeline,
6066 40 : start_lsn: Lsn,
6067 40 : ctx: &RequestContext,
6068 40 : ) -> anyhow::Result<()> {
6069 40 : let mut lsn = start_lsn;
6070 : {
6071 40 : let mut writer = tline.writer().await;
6072 : // Create a relation on the timeline
6073 40 : writer
6074 40 : .put(
6075 40 : *TEST_KEY,
6076 40 : lsn,
6077 40 : &Value::Image(test_img(&format!("foo at {}", lsn))),
6078 40 : ctx,
6079 40 : )
6080 40 : .await?;
6081 40 : writer.finish_write(lsn);
6082 40 : lsn += 0x10;
6083 40 : writer
6084 40 : .put(
6085 40 : *TEST_KEY,
6086 40 : lsn,
6087 40 : &Value::Image(test_img(&format!("foo at {}", lsn))),
6088 40 : ctx,
6089 40 : )
6090 40 : .await?;
6091 40 : writer.finish_write(lsn);
6092 40 : lsn += 0x10;
6093 40 : }
6094 40 : tline.freeze_and_flush().await?;
6095 : {
6096 40 : let mut writer = tline.writer().await;
6097 40 : writer
6098 40 : .put(
6099 40 : *TEST_KEY,
6100 40 : lsn,
6101 40 : &Value::Image(test_img(&format!("foo at {}", lsn))),
6102 40 : ctx,
6103 40 : )
6104 40 : .await?;
6105 40 : writer.finish_write(lsn);
6106 40 : lsn += 0x10;
6107 40 : writer
6108 40 : .put(
6109 40 : *TEST_KEY,
6110 40 : lsn,
6111 40 : &Value::Image(test_img(&format!("foo at {}", lsn))),
6112 40 : ctx,
6113 40 : )
6114 40 : .await?;
6115 40 : writer.finish_write(lsn);
6116 40 : }
6117 40 : tline.freeze_and_flush().await.map_err(|e| e.into())
6118 40 : }
6119 :
6120 : #[tokio::test(start_paused = true)]
6121 4 : async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
6122 4 : let (tenant, ctx) =
6123 4 : TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
6124 4 : .await?
6125 4 : .load()
6126 4 : .await;
6127 4 : // Advance to the lsn lease deadline so that GC is not blocked by
6128 4 : // initial transition into AttachedSingle.
6129 4 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
6130 4 : tokio::time::resume();
6131 4 : let tline = tenant
6132 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6133 4 : .await?;
6134 4 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6135 4 :
6136 4 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
6137 4 : // FIXME: this doesn't actually remove any layer currently, given how the flushing
6138 4 : // and compaction works. But it does set the 'cutoff' point so that the cross check
6139 4 : // below should fail.
6140 4 : tenant
6141 4 : .gc_iteration(
6142 4 : Some(TIMELINE_ID),
6143 4 : 0x10,
6144 4 : Duration::ZERO,
6145 4 : &CancellationToken::new(),
6146 4 : &ctx,
6147 4 : )
6148 4 : .await?;
6149 4 :
6150 4 : // try to branch at lsn 25, should fail because we already garbage collected the data
6151 4 : match tenant
6152 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
6153 4 : .await
6154 4 : {
6155 4 : Ok(_) => panic!("branching should have failed"),
6156 4 : Err(err) => {
6157 4 : let CreateTimelineError::AncestorLsn(err) = err else {
6158 4 : panic!("wrong error type")
6159 4 : };
6160 4 : assert!(err.to_string().contains("invalid branch start lsn"));
6161 4 : assert!(err
6162 4 : .source()
6163 4 : .unwrap()
6164 4 : .to_string()
6165 4 : .contains("we might've already garbage collected needed data"))
6166 4 : }
6167 4 : }
6168 4 :
6169 4 : Ok(())
6170 4 : }
6171 :
6172 : #[tokio::test]
6173 4 : async fn test_prohibit_branch_creation_on_pre_initdb_lsn() -> anyhow::Result<()> {
6174 4 : let (tenant, ctx) =
6175 4 : TenantHarness::create("test_prohibit_branch_creation_on_pre_initdb_lsn")
6176 4 : .await?
6177 4 : .load()
6178 4 : .await;
6179 4 :
6180 4 : let tline = tenant
6181 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x50), DEFAULT_PG_VERSION, &ctx)
6182 4 : .await?;
6183 4 : // try to branch at lsn 0x25, should fail because initdb lsn is 0x50
6184 4 : match tenant
6185 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
6186 4 : .await
6187 4 : {
6188 4 : Ok(_) => panic!("branching should have failed"),
6189 4 : Err(err) => {
6190 4 : let CreateTimelineError::AncestorLsn(err) = err else {
6191 4 : panic!("wrong error type");
6192 4 : };
6193 4 : assert!(&err.to_string().contains("invalid branch start lsn"));
6194 4 : assert!(&err
6195 4 : .source()
6196 4 : .unwrap()
6197 4 : .to_string()
6198 4 : .contains("is earlier than latest GC cutoff"));
6199 4 : }
6200 4 : }
6201 4 :
6202 4 : Ok(())
6203 4 : }
6204 :
6205 : /*
6206 : // FIXME: This currently fails to error out. Calling GC doesn't currently
6207 : // remove the old value, we'd need to work a little harder
6208 : #[tokio::test]
6209 : async fn test_prohibit_get_for_garbage_collected_data() -> anyhow::Result<()> {
6210 : let repo =
6211 : RepoHarness::create("test_prohibit_get_for_garbage_collected_data")?
6212 : .load();
6213 :
6214 : let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION)?;
6215 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6216 :
6217 : repo.gc_iteration(Some(TIMELINE_ID), 0x10, Duration::ZERO)?;
6218 : let applied_gc_cutoff_lsn = tline.get_applied_gc_cutoff_lsn();
6219 : assert!(*applied_gc_cutoff_lsn > Lsn(0x25));
6220 : match tline.get(*TEST_KEY, Lsn(0x25)) {
6221 : Ok(_) => panic!("request for page should have failed"),
6222 : Err(err) => assert!(err.to_string().contains("not found at")),
6223 : }
6224 : Ok(())
6225 : }
6226 : */
6227 :
6228 : #[tokio::test]
6229 4 : async fn test_get_branchpoints_from_an_inactive_timeline() -> anyhow::Result<()> {
6230 4 : let (tenant, ctx) =
6231 4 : TenantHarness::create("test_get_branchpoints_from_an_inactive_timeline")
6232 4 : .await?
6233 4 : .load()
6234 4 : .await;
6235 4 : let tline = tenant
6236 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6237 4 : .await?;
6238 4 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6239 4 :
6240 4 : tenant
6241 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6242 4 : .await?;
6243 4 : let newtline = tenant
6244 4 : .get_timeline(NEW_TIMELINE_ID, true)
6245 4 : .expect("Should have a local timeline");
6246 4 :
6247 4 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6248 4 :
6249 4 : tline.set_broken("test".to_owned());
6250 4 :
6251 4 : tenant
6252 4 : .gc_iteration(
6253 4 : Some(TIMELINE_ID),
6254 4 : 0x10,
6255 4 : Duration::ZERO,
6256 4 : &CancellationToken::new(),
6257 4 : &ctx,
6258 4 : )
6259 4 : .await?;
6260 4 :
6261 4 : // The branchpoints should contain all timelines, even ones marked
6262 4 : // as Broken.
6263 4 : {
6264 4 : let branchpoints = &tline.gc_info.read().unwrap().retain_lsns;
6265 4 : assert_eq!(branchpoints.len(), 1);
6266 4 : assert_eq!(
6267 4 : branchpoints[0],
6268 4 : (Lsn(0x40), NEW_TIMELINE_ID, MaybeOffloaded::No)
6269 4 : );
6270 4 : }
6271 4 :
6272 4 : // You can read the key from the child branch even though the parent is
6273 4 : // Broken, as long as you don't need to access data from the parent.
6274 4 : assert_eq!(
6275 4 : newtline.get(*TEST_KEY, Lsn(0x70), &ctx).await?,
6276 4 : test_img(&format!("foo at {}", Lsn(0x70)))
6277 4 : );
6278 4 :
6279 4 : // This needs to traverse to the parent, and fails.
6280 4 : let err = newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await.unwrap_err();
6281 4 : assert!(
6282 4 : err.to_string().starts_with(&format!(
6283 4 : "bad state on timeline {}: Broken",
6284 4 : tline.timeline_id
6285 4 : )),
6286 4 : "{err}"
6287 4 : );
6288 4 :
6289 4 : Ok(())
6290 4 : }
6291 :
6292 : #[tokio::test]
6293 4 : async fn test_retain_data_in_parent_which_is_needed_for_child() -> anyhow::Result<()> {
6294 4 : let (tenant, ctx) =
6295 4 : TenantHarness::create("test_retain_data_in_parent_which_is_needed_for_child")
6296 4 : .await?
6297 4 : .load()
6298 4 : .await;
6299 4 : let tline = tenant
6300 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6301 4 : .await?;
6302 4 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6303 4 :
6304 4 : tenant
6305 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6306 4 : .await?;
6307 4 : let newtline = tenant
6308 4 : .get_timeline(NEW_TIMELINE_ID, true)
6309 4 : .expect("Should have a local timeline");
6310 4 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
6311 4 : tenant
6312 4 : .gc_iteration(
6313 4 : Some(TIMELINE_ID),
6314 4 : 0x10,
6315 4 : Duration::ZERO,
6316 4 : &CancellationToken::new(),
6317 4 : &ctx,
6318 4 : )
6319 4 : .await?;
6320 4 : assert!(newtline.get(*TEST_KEY, Lsn(0x25), &ctx).await.is_ok());
6321 4 :
6322 4 : Ok(())
6323 4 : }
6324 : #[tokio::test]
6325 4 : async fn test_parent_keeps_data_forever_after_branching() -> anyhow::Result<()> {
6326 4 : let (tenant, ctx) = TenantHarness::create("test_parent_keeps_data_forever_after_branching")
6327 4 : .await?
6328 4 : .load()
6329 4 : .await;
6330 4 : let tline = tenant
6331 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6332 4 : .await?;
6333 4 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6334 4 :
6335 4 : tenant
6336 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6337 4 : .await?;
6338 4 : let newtline = tenant
6339 4 : .get_timeline(NEW_TIMELINE_ID, true)
6340 4 : .expect("Should have a local timeline");
6341 4 :
6342 4 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6343 4 :
6344 4 : // run gc on parent
6345 4 : tenant
6346 4 : .gc_iteration(
6347 4 : Some(TIMELINE_ID),
6348 4 : 0x10,
6349 4 : Duration::ZERO,
6350 4 : &CancellationToken::new(),
6351 4 : &ctx,
6352 4 : )
6353 4 : .await?;
6354 4 :
6355 4 : // Check that the data is still accessible on the branch.
6356 4 : assert_eq!(
6357 4 : newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
6358 4 : test_img(&format!("foo at {}", Lsn(0x40)))
6359 4 : );
6360 4 :
6361 4 : Ok(())
6362 4 : }
6363 :
6364 : #[tokio::test]
6365 4 : async fn timeline_load() -> anyhow::Result<()> {
6366 4 : const TEST_NAME: &str = "timeline_load";
6367 4 : let harness = TenantHarness::create(TEST_NAME).await?;
6368 4 : {
6369 4 : let (tenant, ctx) = harness.load().await;
6370 4 : let tline = tenant
6371 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x7000), DEFAULT_PG_VERSION, &ctx)
6372 4 : .await?;
6373 4 : make_some_layers(tline.as_ref(), Lsn(0x8000), &ctx).await?;
6374 4 : // so that all uploads finish & we can call harness.load() below again
6375 4 : tenant
6376 4 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
6377 4 : .instrument(harness.span())
6378 4 : .await
6379 4 : .ok()
6380 4 : .unwrap();
6381 4 : }
6382 4 :
6383 4 : let (tenant, _ctx) = harness.load().await;
6384 4 : tenant
6385 4 : .get_timeline(TIMELINE_ID, true)
6386 4 : .expect("cannot load timeline");
6387 4 :
6388 4 : Ok(())
6389 4 : }
6390 :
6391 : #[tokio::test]
6392 4 : async fn timeline_load_with_ancestor() -> anyhow::Result<()> {
6393 4 : const TEST_NAME: &str = "timeline_load_with_ancestor";
6394 4 : let harness = TenantHarness::create(TEST_NAME).await?;
6395 4 : // create two timelines
6396 4 : {
6397 4 : let (tenant, ctx) = harness.load().await;
6398 4 : let tline = tenant
6399 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6400 4 : .await?;
6401 4 :
6402 4 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6403 4 :
6404 4 : let child_tline = tenant
6405 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6406 4 : .await?;
6407 4 : child_tline.set_state(TimelineState::Active);
6408 4 :
6409 4 : let newtline = tenant
6410 4 : .get_timeline(NEW_TIMELINE_ID, true)
6411 4 : .expect("Should have a local timeline");
6412 4 :
6413 4 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6414 4 :
6415 4 : // so that all uploads finish & we can call harness.load() below again
6416 4 : tenant
6417 4 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
6418 4 : .instrument(harness.span())
6419 4 : .await
6420 4 : .ok()
6421 4 : .unwrap();
6422 4 : }
6423 4 :
6424 4 : // check that both of them are initially unloaded
6425 4 : let (tenant, _ctx) = harness.load().await;
6426 4 :
6427 4 : // check that both, child and ancestor are loaded
6428 4 : let _child_tline = tenant
6429 4 : .get_timeline(NEW_TIMELINE_ID, true)
6430 4 : .expect("cannot get child timeline loaded");
6431 4 :
6432 4 : let _ancestor_tline = tenant
6433 4 : .get_timeline(TIMELINE_ID, true)
6434 4 : .expect("cannot get ancestor timeline loaded");
6435 4 :
6436 4 : Ok(())
6437 4 : }
6438 :
6439 : #[tokio::test]
6440 4 : async fn delta_layer_dumping() -> anyhow::Result<()> {
6441 4 : use storage_layer::AsLayerDesc;
6442 4 : let (tenant, ctx) = TenantHarness::create("test_layer_dumping")
6443 4 : .await?
6444 4 : .load()
6445 4 : .await;
6446 4 : let tline = tenant
6447 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6448 4 : .await?;
6449 4 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6450 4 :
6451 4 : let layer_map = tline.layers.read().await;
6452 4 : let level0_deltas = layer_map
6453 4 : .layer_map()?
6454 4 : .level0_deltas()
6455 4 : .iter()
6456 8 : .map(|desc| layer_map.get_from_desc(desc))
6457 4 : .collect::<Vec<_>>();
6458 4 :
6459 4 : assert!(!level0_deltas.is_empty());
6460 4 :
6461 12 : for delta in level0_deltas {
6462 4 : // Ensure we are dumping a delta layer here
6463 8 : assert!(delta.layer_desc().is_delta);
6464 8 : delta.dump(true, &ctx).await.unwrap();
6465 4 : }
6466 4 :
6467 4 : Ok(())
6468 4 : }
6469 :
6470 : #[tokio::test]
6471 4 : async fn test_images() -> anyhow::Result<()> {
6472 4 : let (tenant, ctx) = TenantHarness::create("test_images").await?.load().await;
6473 4 : let tline = tenant
6474 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6475 4 : .await?;
6476 4 :
6477 4 : let mut writer = tline.writer().await;
6478 4 : writer
6479 4 : .put(
6480 4 : *TEST_KEY,
6481 4 : Lsn(0x10),
6482 4 : &Value::Image(test_img("foo at 0x10")),
6483 4 : &ctx,
6484 4 : )
6485 4 : .await?;
6486 4 : writer.finish_write(Lsn(0x10));
6487 4 : drop(writer);
6488 4 :
6489 4 : tline.freeze_and_flush().await?;
6490 4 : tline
6491 4 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6492 4 : .await?;
6493 4 :
6494 4 : let mut writer = tline.writer().await;
6495 4 : writer
6496 4 : .put(
6497 4 : *TEST_KEY,
6498 4 : Lsn(0x20),
6499 4 : &Value::Image(test_img("foo at 0x20")),
6500 4 : &ctx,
6501 4 : )
6502 4 : .await?;
6503 4 : writer.finish_write(Lsn(0x20));
6504 4 : drop(writer);
6505 4 :
6506 4 : tline.freeze_and_flush().await?;
6507 4 : tline
6508 4 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6509 4 : .await?;
6510 4 :
6511 4 : let mut writer = tline.writer().await;
6512 4 : writer
6513 4 : .put(
6514 4 : *TEST_KEY,
6515 4 : Lsn(0x30),
6516 4 : &Value::Image(test_img("foo at 0x30")),
6517 4 : &ctx,
6518 4 : )
6519 4 : .await?;
6520 4 : writer.finish_write(Lsn(0x30));
6521 4 : drop(writer);
6522 4 :
6523 4 : tline.freeze_and_flush().await?;
6524 4 : tline
6525 4 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6526 4 : .await?;
6527 4 :
6528 4 : let mut writer = tline.writer().await;
6529 4 : writer
6530 4 : .put(
6531 4 : *TEST_KEY,
6532 4 : Lsn(0x40),
6533 4 : &Value::Image(test_img("foo at 0x40")),
6534 4 : &ctx,
6535 4 : )
6536 4 : .await?;
6537 4 : writer.finish_write(Lsn(0x40));
6538 4 : drop(writer);
6539 4 :
6540 4 : tline.freeze_and_flush().await?;
6541 4 : tline
6542 4 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6543 4 : .await?;
6544 4 :
6545 4 : assert_eq!(
6546 4 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
6547 4 : test_img("foo at 0x10")
6548 4 : );
6549 4 : assert_eq!(
6550 4 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
6551 4 : test_img("foo at 0x10")
6552 4 : );
6553 4 : assert_eq!(
6554 4 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
6555 4 : test_img("foo at 0x20")
6556 4 : );
6557 4 : assert_eq!(
6558 4 : tline.get(*TEST_KEY, Lsn(0x30), &ctx).await?,
6559 4 : test_img("foo at 0x30")
6560 4 : );
6561 4 : assert_eq!(
6562 4 : tline.get(*TEST_KEY, Lsn(0x40), &ctx).await?,
6563 4 : test_img("foo at 0x40")
6564 4 : );
6565 4 :
6566 4 : Ok(())
6567 4 : }
6568 :
6569 8 : async fn bulk_insert_compact_gc(
6570 8 : tenant: &Tenant,
6571 8 : timeline: &Arc<Timeline>,
6572 8 : ctx: &RequestContext,
6573 8 : lsn: Lsn,
6574 8 : repeat: usize,
6575 8 : key_count: usize,
6576 8 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
6577 8 : let compact = true;
6578 8 : bulk_insert_maybe_compact_gc(tenant, timeline, ctx, lsn, repeat, key_count, compact).await
6579 8 : }
6580 :
6581 16 : async fn bulk_insert_maybe_compact_gc(
6582 16 : tenant: &Tenant,
6583 16 : timeline: &Arc<Timeline>,
6584 16 : ctx: &RequestContext,
6585 16 : mut lsn: Lsn,
6586 16 : repeat: usize,
6587 16 : key_count: usize,
6588 16 : compact: bool,
6589 16 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
6590 16 : let mut inserted: HashMap<Key, BTreeSet<Lsn>> = Default::default();
6591 16 :
6592 16 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6593 16 : let mut blknum = 0;
6594 16 :
6595 16 : // Enforce that key range is monotonously increasing
6596 16 : let mut keyspace = KeySpaceAccum::new();
6597 16 :
6598 16 : let cancel = CancellationToken::new();
6599 16 :
6600 16 : for _ in 0..repeat {
6601 800 : for _ in 0..key_count {
6602 8000000 : test_key.field6 = blknum;
6603 8000000 : let mut writer = timeline.writer().await;
6604 8000000 : writer
6605 8000000 : .put(
6606 8000000 : test_key,
6607 8000000 : lsn,
6608 8000000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6609 8000000 : ctx,
6610 8000000 : )
6611 8000000 : .await?;
6612 8000000 : inserted.entry(test_key).or_default().insert(lsn);
6613 8000000 : writer.finish_write(lsn);
6614 8000000 : drop(writer);
6615 8000000 :
6616 8000000 : keyspace.add_key(test_key);
6617 8000000 :
6618 8000000 : lsn = Lsn(lsn.0 + 0x10);
6619 8000000 : blknum += 1;
6620 : }
6621 :
6622 800 : timeline.freeze_and_flush().await?;
6623 800 : if compact {
6624 : // this requires timeline to be &Arc<Timeline>
6625 400 : timeline.compact(&cancel, EnumSet::empty(), ctx).await?;
6626 400 : }
6627 :
6628 : // this doesn't really need to use the timeline_id target, but it is closer to what it
6629 : // originally was.
6630 800 : let res = tenant
6631 800 : .gc_iteration(Some(timeline.timeline_id), 0, Duration::ZERO, &cancel, ctx)
6632 800 : .await?;
6633 :
6634 800 : assert_eq!(res.layers_removed, 0, "this never removes anything");
6635 : }
6636 :
6637 16 : Ok(inserted)
6638 16 : }
6639 :
6640 : //
6641 : // Insert 1000 key-value pairs with increasing keys, flush, compact, GC.
6642 : // Repeat 50 times.
6643 : //
6644 : #[tokio::test]
6645 4 : async fn test_bulk_insert() -> anyhow::Result<()> {
6646 4 : let harness = TenantHarness::create("test_bulk_insert").await?;
6647 4 : let (tenant, ctx) = harness.load().await;
6648 4 : let tline = tenant
6649 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6650 4 : .await?;
6651 4 :
6652 4 : let lsn = Lsn(0x10);
6653 4 : bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
6654 4 :
6655 4 : Ok(())
6656 4 : }
6657 :
6658 : // Test the vectored get real implementation against a simple sequential implementation.
6659 : //
6660 : // The test generates a keyspace by repeatedly flushing the in-memory layer and compacting.
6661 : // Projected to 2D the key space looks like below. Lsn grows upwards on the Y axis and keys
6662 : // grow to the right on the X axis.
6663 : // [Delta]
6664 : // [Delta]
6665 : // [Delta]
6666 : // [Delta]
6667 : // ------------ Image ---------------
6668 : //
6669 : // After layer generation we pick the ranges to query as follows:
6670 : // 1. The beginning of each delta layer
6671 : // 2. At the seam between two adjacent delta layers
6672 : //
6673 : // There's one major downside to this test: delta layers only contains images,
6674 : // so the search can stop at the first delta layer and doesn't traverse any deeper.
6675 : #[tokio::test]
6676 4 : async fn test_get_vectored() -> anyhow::Result<()> {
6677 4 : let harness = TenantHarness::create("test_get_vectored").await?;
6678 4 : let (tenant, ctx) = harness.load().await;
6679 4 : let io_concurrency = IoConcurrency::spawn_for_test();
6680 4 : let tline = tenant
6681 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6682 4 : .await?;
6683 4 :
6684 4 : let lsn = Lsn(0x10);
6685 4 : let inserted = bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
6686 4 :
6687 4 : let guard = tline.layers.read().await;
6688 4 : let lm = guard.layer_map()?;
6689 4 :
6690 4 : lm.dump(true, &ctx).await?;
6691 4 :
6692 4 : let mut reads = Vec::new();
6693 4 : let mut prev = None;
6694 24 : lm.iter_historic_layers().for_each(|desc| {
6695 24 : if !desc.is_delta() {
6696 4 : prev = Some(desc.clone());
6697 4 : return;
6698 20 : }
6699 20 :
6700 20 : let start = desc.key_range.start;
6701 20 : let end = desc
6702 20 : .key_range
6703 20 : .start
6704 20 : .add(Timeline::MAX_GET_VECTORED_KEYS.try_into().unwrap());
6705 20 : reads.push(KeySpace {
6706 20 : ranges: vec![start..end],
6707 20 : });
6708 4 :
6709 20 : if let Some(prev) = &prev {
6710 20 : if !prev.is_delta() {
6711 20 : return;
6712 4 : }
6713 0 :
6714 0 : let first_range = Key {
6715 0 : field6: prev.key_range.end.field6 - 4,
6716 0 : ..prev.key_range.end
6717 0 : }..prev.key_range.end;
6718 0 :
6719 0 : let second_range = desc.key_range.start..Key {
6720 0 : field6: desc.key_range.start.field6 + 4,
6721 0 : ..desc.key_range.start
6722 0 : };
6723 0 :
6724 0 : reads.push(KeySpace {
6725 0 : ranges: vec![first_range, second_range],
6726 0 : });
6727 4 : };
6728 4 :
6729 4 : prev = Some(desc.clone());
6730 24 : });
6731 4 :
6732 4 : drop(guard);
6733 4 :
6734 4 : // Pick a big LSN such that we query over all the changes.
6735 4 : let reads_lsn = Lsn(u64::MAX - 1);
6736 4 :
6737 24 : for read in reads {
6738 20 : info!("Doing vectored read on {:?}", read);
6739 4 :
6740 20 : let vectored_res = tline
6741 20 : .get_vectored_impl(
6742 20 : read.clone(),
6743 20 : reads_lsn,
6744 20 : &mut ValuesReconstructState::new(io_concurrency.clone()),
6745 20 : &ctx,
6746 20 : )
6747 20 : .await;
6748 4 :
6749 20 : let mut expected_lsns: HashMap<Key, Lsn> = Default::default();
6750 20 : let mut expect_missing = false;
6751 20 : let mut key = read.start().unwrap();
6752 660 : while key != read.end().unwrap() {
6753 640 : if let Some(lsns) = inserted.get(&key) {
6754 640 : let expected_lsn = lsns.iter().rfind(|lsn| **lsn <= reads_lsn);
6755 640 : match expected_lsn {
6756 640 : Some(lsn) => {
6757 640 : expected_lsns.insert(key, *lsn);
6758 640 : }
6759 4 : None => {
6760 4 : expect_missing = true;
6761 0 : break;
6762 4 : }
6763 4 : }
6764 4 : } else {
6765 4 : expect_missing = true;
6766 0 : break;
6767 4 : }
6768 4 :
6769 640 : key = key.next();
6770 4 : }
6771 4 :
6772 20 : if expect_missing {
6773 4 : assert!(matches!(vectored_res, Err(GetVectoredError::MissingKey(_))));
6774 4 : } else {
6775 640 : for (key, image) in vectored_res? {
6776 640 : let expected_lsn = expected_lsns.get(&key).expect("determined above");
6777 640 : let expected_image = test_img(&format!("{} at {}", key.field6, expected_lsn));
6778 640 : assert_eq!(image?, expected_image);
6779 4 : }
6780 4 : }
6781 4 : }
6782 4 :
6783 4 : Ok(())
6784 4 : }
6785 :
6786 : #[tokio::test]
6787 4 : async fn test_get_vectored_aux_files() -> anyhow::Result<()> {
6788 4 : let harness = TenantHarness::create("test_get_vectored_aux_files").await?;
6789 4 :
6790 4 : let (tenant, ctx) = harness.load().await;
6791 4 : let io_concurrency = IoConcurrency::spawn_for_test();
6792 4 : let tline = tenant
6793 4 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
6794 4 : .await?;
6795 4 : let tline = tline.raw_timeline().unwrap();
6796 4 :
6797 4 : let mut modification = tline.begin_modification(Lsn(0x1000));
6798 4 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
6799 4 : modification.set_lsn(Lsn(0x1008))?;
6800 4 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
6801 4 : modification.commit(&ctx).await?;
6802 4 :
6803 4 : let child_timeline_id = TimelineId::generate();
6804 4 : tenant
6805 4 : .branch_timeline_test(
6806 4 : tline,
6807 4 : child_timeline_id,
6808 4 : Some(tline.get_last_record_lsn()),
6809 4 : &ctx,
6810 4 : )
6811 4 : .await?;
6812 4 :
6813 4 : let child_timeline = tenant
6814 4 : .get_timeline(child_timeline_id, true)
6815 4 : .expect("Should have the branched timeline");
6816 4 :
6817 4 : let aux_keyspace = KeySpace {
6818 4 : ranges: vec![NON_INHERITED_RANGE],
6819 4 : };
6820 4 : let read_lsn = child_timeline.get_last_record_lsn();
6821 4 :
6822 4 : let vectored_res = child_timeline
6823 4 : .get_vectored_impl(
6824 4 : aux_keyspace.clone(),
6825 4 : read_lsn,
6826 4 : &mut ValuesReconstructState::new(io_concurrency.clone()),
6827 4 : &ctx,
6828 4 : )
6829 4 : .await;
6830 4 :
6831 4 : let images = vectored_res?;
6832 4 : assert!(images.is_empty());
6833 4 : Ok(())
6834 4 : }
6835 :
6836 : // Test that vectored get handles layer gaps correctly
6837 : // by advancing into the next ancestor timeline if required.
6838 : //
6839 : // The test generates timelines that look like the diagram below.
6840 : // We leave a gap in one of the L1 layers at `gap_at_key` (`/` in the diagram).
6841 : // The reconstruct data for that key lies in the ancestor timeline (`X` in the diagram).
6842 : //
6843 : // ```
6844 : //-------------------------------+
6845 : // ... |
6846 : // [ L1 ] |
6847 : // [ / L1 ] | Child Timeline
6848 : // ... |
6849 : // ------------------------------+
6850 : // [ X L1 ] | Parent Timeline
6851 : // ------------------------------+
6852 : // ```
6853 : #[tokio::test]
6854 4 : async fn test_get_vectored_key_gap() -> anyhow::Result<()> {
6855 4 : let tenant_conf = TenantConf {
6856 4 : // Make compaction deterministic
6857 4 : gc_period: Duration::ZERO,
6858 4 : compaction_period: Duration::ZERO,
6859 4 : // Encourage creation of L1 layers
6860 4 : checkpoint_distance: 16 * 1024,
6861 4 : compaction_target_size: 8 * 1024,
6862 4 : ..TenantConf::default()
6863 4 : };
6864 4 :
6865 4 : let harness = TenantHarness::create_custom(
6866 4 : "test_get_vectored_key_gap",
6867 4 : tenant_conf,
6868 4 : TenantId::generate(),
6869 4 : ShardIdentity::unsharded(),
6870 4 : Generation::new(0xdeadbeef),
6871 4 : )
6872 4 : .await?;
6873 4 : let (tenant, ctx) = harness.load().await;
6874 4 : let io_concurrency = IoConcurrency::spawn_for_test();
6875 4 :
6876 4 : let mut current_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6877 4 : let gap_at_key = current_key.add(100);
6878 4 : let mut current_lsn = Lsn(0x10);
6879 4 :
6880 4 : const KEY_COUNT: usize = 10_000;
6881 4 :
6882 4 : let timeline_id = TimelineId::generate();
6883 4 : let current_timeline = tenant
6884 4 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
6885 4 : .await?;
6886 4 :
6887 4 : current_lsn += 0x100;
6888 4 :
6889 4 : let mut writer = current_timeline.writer().await;
6890 4 : writer
6891 4 : .put(
6892 4 : gap_at_key,
6893 4 : current_lsn,
6894 4 : &Value::Image(test_img(&format!("{} at {}", gap_at_key, current_lsn))),
6895 4 : &ctx,
6896 4 : )
6897 4 : .await?;
6898 4 : writer.finish_write(current_lsn);
6899 4 : drop(writer);
6900 4 :
6901 4 : let mut latest_lsns = HashMap::new();
6902 4 : latest_lsns.insert(gap_at_key, current_lsn);
6903 4 :
6904 4 : current_timeline.freeze_and_flush().await?;
6905 4 :
6906 4 : let child_timeline_id = TimelineId::generate();
6907 4 :
6908 4 : tenant
6909 4 : .branch_timeline_test(
6910 4 : ¤t_timeline,
6911 4 : child_timeline_id,
6912 4 : Some(current_lsn),
6913 4 : &ctx,
6914 4 : )
6915 4 : .await?;
6916 4 : let child_timeline = tenant
6917 4 : .get_timeline(child_timeline_id, true)
6918 4 : .expect("Should have the branched timeline");
6919 4 :
6920 40004 : for i in 0..KEY_COUNT {
6921 40000 : if current_key == gap_at_key {
6922 4 : current_key = current_key.next();
6923 4 : continue;
6924 39996 : }
6925 39996 :
6926 39996 : current_lsn += 0x10;
6927 4 :
6928 39996 : let mut writer = child_timeline.writer().await;
6929 39996 : writer
6930 39996 : .put(
6931 39996 : current_key,
6932 39996 : current_lsn,
6933 39996 : &Value::Image(test_img(&format!("{} at {}", current_key, current_lsn))),
6934 39996 : &ctx,
6935 39996 : )
6936 39996 : .await?;
6937 39996 : writer.finish_write(current_lsn);
6938 39996 : drop(writer);
6939 39996 :
6940 39996 : latest_lsns.insert(current_key, current_lsn);
6941 39996 : current_key = current_key.next();
6942 39996 :
6943 39996 : // Flush every now and then to encourage layer file creation.
6944 39996 : if i % 500 == 0 {
6945 80 : child_timeline.freeze_and_flush().await?;
6946 39916 : }
6947 4 : }
6948 4 :
6949 4 : child_timeline.freeze_and_flush().await?;
6950 4 : let mut flags = EnumSet::new();
6951 4 : flags.insert(CompactFlags::ForceRepartition);
6952 4 : child_timeline
6953 4 : .compact(&CancellationToken::new(), flags, &ctx)
6954 4 : .await?;
6955 4 :
6956 4 : let key_near_end = {
6957 4 : let mut tmp = current_key;
6958 4 : tmp.field6 -= 10;
6959 4 : tmp
6960 4 : };
6961 4 :
6962 4 : let key_near_gap = {
6963 4 : let mut tmp = gap_at_key;
6964 4 : tmp.field6 -= 10;
6965 4 : tmp
6966 4 : };
6967 4 :
6968 4 : let read = KeySpace {
6969 4 : ranges: vec![key_near_gap..gap_at_key.next(), key_near_end..current_key],
6970 4 : };
6971 4 : let results = child_timeline
6972 4 : .get_vectored_impl(
6973 4 : read.clone(),
6974 4 : current_lsn,
6975 4 : &mut ValuesReconstructState::new(io_concurrency.clone()),
6976 4 : &ctx,
6977 4 : )
6978 4 : .await?;
6979 4 :
6980 88 : for (key, img_res) in results {
6981 84 : let expected = test_img(&format!("{} at {}", key, latest_lsns[&key]));
6982 84 : assert_eq!(img_res?, expected);
6983 4 : }
6984 4 :
6985 4 : Ok(())
6986 4 : }
6987 :
6988 : // Test that vectored get descends into ancestor timelines correctly and
6989 : // does not return an image that's newer than requested.
6990 : //
6991 : // The diagram below ilustrates an interesting case. We have a parent timeline
6992 : // (top of the Lsn range) and a child timeline. The request key cannot be reconstructed
6993 : // from the child timeline, so the parent timeline must be visited. When advacing into
6994 : // the child timeline, the read path needs to remember what the requested Lsn was in
6995 : // order to avoid returning an image that's too new. The test below constructs such
6996 : // a timeline setup and does a few queries around the Lsn of each page image.
6997 : // ```
6998 : // LSN
6999 : // ^
7000 : // |
7001 : // |
7002 : // 500 | --------------------------------------> branch point
7003 : // 400 | X
7004 : // 300 | X
7005 : // 200 | --------------------------------------> requested lsn
7006 : // 100 | X
7007 : // |---------------------------------------> Key
7008 : // |
7009 : // ------> requested key
7010 : //
7011 : // Legend:
7012 : // * X - page images
7013 : // ```
7014 : #[tokio::test]
7015 4 : async fn test_get_vectored_ancestor_descent() -> anyhow::Result<()> {
7016 4 : let harness = TenantHarness::create("test_get_vectored_on_lsn_axis").await?;
7017 4 : let (tenant, ctx) = harness.load().await;
7018 4 : let io_concurrency = IoConcurrency::spawn_for_test();
7019 4 :
7020 4 : let start_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7021 4 : let end_key = start_key.add(1000);
7022 4 : let child_gap_at_key = start_key.add(500);
7023 4 : let mut parent_gap_lsns: BTreeMap<Lsn, String> = BTreeMap::new();
7024 4 :
7025 4 : let mut current_lsn = Lsn(0x10);
7026 4 :
7027 4 : let timeline_id = TimelineId::generate();
7028 4 : let parent_timeline = tenant
7029 4 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
7030 4 : .await?;
7031 4 :
7032 4 : current_lsn += 0x100;
7033 4 :
7034 16 : for _ in 0..3 {
7035 12 : let mut key = start_key;
7036 12012 : while key < end_key {
7037 12000 : current_lsn += 0x10;
7038 12000 :
7039 12000 : let image_value = format!("{} at {}", child_gap_at_key, current_lsn);
7040 4 :
7041 12000 : let mut writer = parent_timeline.writer().await;
7042 12000 : writer
7043 12000 : .put(
7044 12000 : key,
7045 12000 : current_lsn,
7046 12000 : &Value::Image(test_img(&image_value)),
7047 12000 : &ctx,
7048 12000 : )
7049 12000 : .await?;
7050 12000 : writer.finish_write(current_lsn);
7051 12000 :
7052 12000 : if key == child_gap_at_key {
7053 12 : parent_gap_lsns.insert(current_lsn, image_value);
7054 11988 : }
7055 4 :
7056 12000 : key = key.next();
7057 4 : }
7058 4 :
7059 12 : parent_timeline.freeze_and_flush().await?;
7060 4 : }
7061 4 :
7062 4 : let child_timeline_id = TimelineId::generate();
7063 4 :
7064 4 : let child_timeline = tenant
7065 4 : .branch_timeline_test(&parent_timeline, child_timeline_id, Some(current_lsn), &ctx)
7066 4 : .await?;
7067 4 :
7068 4 : let mut key = start_key;
7069 4004 : while key < end_key {
7070 4000 : if key == child_gap_at_key {
7071 4 : key = key.next();
7072 4 : continue;
7073 3996 : }
7074 3996 :
7075 3996 : current_lsn += 0x10;
7076 4 :
7077 3996 : let mut writer = child_timeline.writer().await;
7078 3996 : writer
7079 3996 : .put(
7080 3996 : key,
7081 3996 : current_lsn,
7082 3996 : &Value::Image(test_img(&format!("{} at {}", key, current_lsn))),
7083 3996 : &ctx,
7084 3996 : )
7085 3996 : .await?;
7086 3996 : writer.finish_write(current_lsn);
7087 3996 :
7088 3996 : key = key.next();
7089 4 : }
7090 4 :
7091 4 : child_timeline.freeze_and_flush().await?;
7092 4 :
7093 4 : let lsn_offsets: [i64; 5] = [-10, -1, 0, 1, 10];
7094 4 : let mut query_lsns = Vec::new();
7095 12 : for image_lsn in parent_gap_lsns.keys().rev() {
7096 72 : for offset in lsn_offsets {
7097 60 : query_lsns.push(Lsn(image_lsn
7098 60 : .0
7099 60 : .checked_add_signed(offset)
7100 60 : .expect("Shouldn't overflow")));
7101 60 : }
7102 4 : }
7103 4 :
7104 64 : for query_lsn in query_lsns {
7105 60 : let results = child_timeline
7106 60 : .get_vectored_impl(
7107 60 : KeySpace {
7108 60 : ranges: vec![child_gap_at_key..child_gap_at_key.next()],
7109 60 : },
7110 60 : query_lsn,
7111 60 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7112 60 : &ctx,
7113 60 : )
7114 60 : .await;
7115 4 :
7116 60 : let expected_item = parent_gap_lsns
7117 60 : .iter()
7118 60 : .rev()
7119 136 : .find(|(lsn, _)| **lsn <= query_lsn);
7120 60 :
7121 60 : info!(
7122 4 : "Doing vectored read at LSN {}. Expecting image to be: {:?}",
7123 4 : query_lsn, expected_item
7124 4 : );
7125 4 :
7126 60 : match expected_item {
7127 52 : Some((_, img_value)) => {
7128 52 : let key_results = results.expect("No vectored get error expected");
7129 52 : let key_result = &key_results[&child_gap_at_key];
7130 52 : let returned_img = key_result
7131 52 : .as_ref()
7132 52 : .expect("No page reconstruct error expected");
7133 52 :
7134 52 : info!(
7135 4 : "Vectored read at LSN {} returned image {}",
7136 0 : query_lsn,
7137 0 : std::str::from_utf8(returned_img)?
7138 4 : );
7139 52 : assert_eq!(*returned_img, test_img(img_value));
7140 4 : }
7141 4 : None => {
7142 8 : assert!(matches!(results, Err(GetVectoredError::MissingKey(_))));
7143 4 : }
7144 4 : }
7145 4 : }
7146 4 :
7147 4 : Ok(())
7148 4 : }
7149 :
7150 : #[tokio::test]
7151 4 : async fn test_random_updates() -> anyhow::Result<()> {
7152 4 : let names_algorithms = [
7153 4 : ("test_random_updates_legacy", CompactionAlgorithm::Legacy),
7154 4 : ("test_random_updates_tiered", CompactionAlgorithm::Tiered),
7155 4 : ];
7156 12 : for (name, algorithm) in names_algorithms {
7157 8 : test_random_updates_algorithm(name, algorithm).await?;
7158 4 : }
7159 4 : Ok(())
7160 4 : }
7161 :
7162 8 : async fn test_random_updates_algorithm(
7163 8 : name: &'static str,
7164 8 : compaction_algorithm: CompactionAlgorithm,
7165 8 : ) -> anyhow::Result<()> {
7166 8 : let mut harness = TenantHarness::create(name).await?;
7167 8 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
7168 8 : kind: compaction_algorithm,
7169 8 : };
7170 8 : let (tenant, ctx) = harness.load().await;
7171 8 : let tline = tenant
7172 8 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7173 8 : .await?;
7174 :
7175 : const NUM_KEYS: usize = 1000;
7176 8 : let cancel = CancellationToken::new();
7177 8 :
7178 8 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7179 8 : let mut test_key_end = test_key;
7180 8 : test_key_end.field6 = NUM_KEYS as u32;
7181 8 : tline.add_extra_test_dense_keyspace(KeySpace::single(test_key..test_key_end));
7182 8 :
7183 8 : let mut keyspace = KeySpaceAccum::new();
7184 8 :
7185 8 : // Track when each page was last modified. Used to assert that
7186 8 : // a read sees the latest page version.
7187 8 : let mut updated = [Lsn(0); NUM_KEYS];
7188 8 :
7189 8 : let mut lsn = Lsn(0x10);
7190 : #[allow(clippy::needless_range_loop)]
7191 8008 : for blknum in 0..NUM_KEYS {
7192 8000 : lsn = Lsn(lsn.0 + 0x10);
7193 8000 : test_key.field6 = blknum as u32;
7194 8000 : let mut writer = tline.writer().await;
7195 8000 : writer
7196 8000 : .put(
7197 8000 : test_key,
7198 8000 : lsn,
7199 8000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7200 8000 : &ctx,
7201 8000 : )
7202 8000 : .await?;
7203 8000 : writer.finish_write(lsn);
7204 8000 : updated[blknum] = lsn;
7205 8000 : drop(writer);
7206 8000 :
7207 8000 : keyspace.add_key(test_key);
7208 : }
7209 :
7210 408 : for _ in 0..50 {
7211 400400 : for _ in 0..NUM_KEYS {
7212 400000 : lsn = Lsn(lsn.0 + 0x10);
7213 400000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7214 400000 : test_key.field6 = blknum as u32;
7215 400000 : let mut writer = tline.writer().await;
7216 400000 : writer
7217 400000 : .put(
7218 400000 : test_key,
7219 400000 : lsn,
7220 400000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7221 400000 : &ctx,
7222 400000 : )
7223 400000 : .await?;
7224 400000 : writer.finish_write(lsn);
7225 400000 : drop(writer);
7226 400000 : updated[blknum] = lsn;
7227 : }
7228 :
7229 : // Read all the blocks
7230 400000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7231 400000 : test_key.field6 = blknum as u32;
7232 400000 : assert_eq!(
7233 400000 : tline.get(test_key, lsn, &ctx).await?,
7234 400000 : test_img(&format!("{} at {}", blknum, last_lsn))
7235 : );
7236 : }
7237 :
7238 : // Perform a cycle of flush, and GC
7239 400 : tline.freeze_and_flush().await?;
7240 400 : tenant
7241 400 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7242 400 : .await?;
7243 : }
7244 :
7245 8 : Ok(())
7246 8 : }
7247 :
7248 : #[tokio::test]
7249 4 : async fn test_traverse_branches() -> anyhow::Result<()> {
7250 4 : let (tenant, ctx) = TenantHarness::create("test_traverse_branches")
7251 4 : .await?
7252 4 : .load()
7253 4 : .await;
7254 4 : let mut tline = tenant
7255 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7256 4 : .await?;
7257 4 :
7258 4 : const NUM_KEYS: usize = 1000;
7259 4 :
7260 4 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7261 4 :
7262 4 : let mut keyspace = KeySpaceAccum::new();
7263 4 :
7264 4 : let cancel = CancellationToken::new();
7265 4 :
7266 4 : // Track when each page was last modified. Used to assert that
7267 4 : // a read sees the latest page version.
7268 4 : let mut updated = [Lsn(0); NUM_KEYS];
7269 4 :
7270 4 : let mut lsn = Lsn(0x10);
7271 4 : #[allow(clippy::needless_range_loop)]
7272 4004 : for blknum in 0..NUM_KEYS {
7273 4000 : lsn = Lsn(lsn.0 + 0x10);
7274 4000 : test_key.field6 = blknum as u32;
7275 4000 : let mut writer = tline.writer().await;
7276 4000 : writer
7277 4000 : .put(
7278 4000 : test_key,
7279 4000 : lsn,
7280 4000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7281 4000 : &ctx,
7282 4000 : )
7283 4000 : .await?;
7284 4000 : writer.finish_write(lsn);
7285 4000 : updated[blknum] = lsn;
7286 4000 : drop(writer);
7287 4000 :
7288 4000 : keyspace.add_key(test_key);
7289 4 : }
7290 4 :
7291 204 : for _ in 0..50 {
7292 200 : let new_tline_id = TimelineId::generate();
7293 200 : tenant
7294 200 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7295 200 : .await?;
7296 200 : tline = tenant
7297 200 : .get_timeline(new_tline_id, true)
7298 200 : .expect("Should have the branched timeline");
7299 4 :
7300 200200 : for _ in 0..NUM_KEYS {
7301 200000 : lsn = Lsn(lsn.0 + 0x10);
7302 200000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7303 200000 : test_key.field6 = blknum as u32;
7304 200000 : let mut writer = tline.writer().await;
7305 200000 : writer
7306 200000 : .put(
7307 200000 : test_key,
7308 200000 : lsn,
7309 200000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7310 200000 : &ctx,
7311 200000 : )
7312 200000 : .await?;
7313 200000 : println!("updating {} at {}", blknum, lsn);
7314 200000 : writer.finish_write(lsn);
7315 200000 : drop(writer);
7316 200000 : updated[blknum] = lsn;
7317 4 : }
7318 4 :
7319 4 : // Read all the blocks
7320 200000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7321 200000 : test_key.field6 = blknum as u32;
7322 200000 : assert_eq!(
7323 200000 : tline.get(test_key, lsn, &ctx).await?,
7324 200000 : test_img(&format!("{} at {}", blknum, last_lsn))
7325 4 : );
7326 4 : }
7327 4 :
7328 4 : // Perform a cycle of flush, compact, and GC
7329 200 : tline.freeze_and_flush().await?;
7330 200 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
7331 200 : tenant
7332 200 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7333 200 : .await?;
7334 4 : }
7335 4 :
7336 4 : Ok(())
7337 4 : }
7338 :
7339 : #[tokio::test]
7340 4 : async fn test_traverse_ancestors() -> anyhow::Result<()> {
7341 4 : let (tenant, ctx) = TenantHarness::create("test_traverse_ancestors")
7342 4 : .await?
7343 4 : .load()
7344 4 : .await;
7345 4 : let mut tline = tenant
7346 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7347 4 : .await?;
7348 4 :
7349 4 : const NUM_KEYS: usize = 100;
7350 4 : const NUM_TLINES: usize = 50;
7351 4 :
7352 4 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7353 4 : // Track page mutation lsns across different timelines.
7354 4 : let mut updated = [[Lsn(0); NUM_KEYS]; NUM_TLINES];
7355 4 :
7356 4 : let mut lsn = Lsn(0x10);
7357 4 :
7358 4 : #[allow(clippy::needless_range_loop)]
7359 204 : for idx in 0..NUM_TLINES {
7360 200 : let new_tline_id = TimelineId::generate();
7361 200 : tenant
7362 200 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7363 200 : .await?;
7364 200 : tline = tenant
7365 200 : .get_timeline(new_tline_id, true)
7366 200 : .expect("Should have the branched timeline");
7367 4 :
7368 20200 : for _ in 0..NUM_KEYS {
7369 20000 : lsn = Lsn(lsn.0 + 0x10);
7370 20000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7371 20000 : test_key.field6 = blknum as u32;
7372 20000 : let mut writer = tline.writer().await;
7373 20000 : writer
7374 20000 : .put(
7375 20000 : test_key,
7376 20000 : lsn,
7377 20000 : &Value::Image(test_img(&format!("{} {} at {}", idx, blknum, lsn))),
7378 20000 : &ctx,
7379 20000 : )
7380 20000 : .await?;
7381 20000 : println!("updating [{}][{}] at {}", idx, blknum, lsn);
7382 20000 : writer.finish_write(lsn);
7383 20000 : drop(writer);
7384 20000 : updated[idx][blknum] = lsn;
7385 4 : }
7386 4 : }
7387 4 :
7388 4 : // Read pages from leaf timeline across all ancestors.
7389 200 : for (idx, lsns) in updated.iter().enumerate() {
7390 20000 : for (blknum, lsn) in lsns.iter().enumerate() {
7391 4 : // Skip empty mutations.
7392 20000 : if lsn.0 == 0 {
7393 7317 : continue;
7394 12683 : }
7395 12683 : println!("checking [{idx}][{blknum}] at {lsn}");
7396 12683 : test_key.field6 = blknum as u32;
7397 12683 : assert_eq!(
7398 12683 : tline.get(test_key, *lsn, &ctx).await?,
7399 12683 : test_img(&format!("{idx} {blknum} at {lsn}"))
7400 4 : );
7401 4 : }
7402 4 : }
7403 4 : Ok(())
7404 4 : }
7405 :
7406 : #[tokio::test]
7407 4 : async fn test_write_at_initdb_lsn_takes_optimization_code_path() -> anyhow::Result<()> {
7408 4 : let (tenant, ctx) = TenantHarness::create("test_empty_test_timeline_is_usable")
7409 4 : .await?
7410 4 : .load()
7411 4 : .await;
7412 4 :
7413 4 : let initdb_lsn = Lsn(0x20);
7414 4 : let utline = tenant
7415 4 : .create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)
7416 4 : .await?;
7417 4 : let tline = utline.raw_timeline().unwrap();
7418 4 :
7419 4 : // Spawn flush loop now so that we can set the `expect_initdb_optimization`
7420 4 : tline.maybe_spawn_flush_loop();
7421 4 :
7422 4 : // Make sure the timeline has the minimum set of required keys for operation.
7423 4 : // The only operation you can always do on an empty timeline is to `put` new data.
7424 4 : // Except if you `put` at `initdb_lsn`.
7425 4 : // In that case, there's an optimization to directly create image layers instead of delta layers.
7426 4 : // It uses `repartition()`, which assumes some keys to be present.
7427 4 : // Let's make sure the test timeline can handle that case.
7428 4 : {
7429 4 : let mut state = tline.flush_loop_state.lock().unwrap();
7430 4 : assert_eq!(
7431 4 : timeline::FlushLoopState::Running {
7432 4 : expect_initdb_optimization: false,
7433 4 : initdb_optimization_count: 0,
7434 4 : },
7435 4 : *state
7436 4 : );
7437 4 : *state = timeline::FlushLoopState::Running {
7438 4 : expect_initdb_optimization: true,
7439 4 : initdb_optimization_count: 0,
7440 4 : };
7441 4 : }
7442 4 :
7443 4 : // Make writes at the initdb_lsn. When we flush it below, it should be handled by the optimization.
7444 4 : // As explained above, the optimization requires some keys to be present.
7445 4 : // As per `create_empty_timeline` documentation, use init_empty to set them.
7446 4 : // This is what `create_test_timeline` does, by the way.
7447 4 : let mut modification = tline.begin_modification(initdb_lsn);
7448 4 : modification
7449 4 : .init_empty_test_timeline()
7450 4 : .context("init_empty_test_timeline")?;
7451 4 : modification
7452 4 : .commit(&ctx)
7453 4 : .await
7454 4 : .context("commit init_empty_test_timeline modification")?;
7455 4 :
7456 4 : // Do the flush. The flush code will check the expectations that we set above.
7457 4 : tline.freeze_and_flush().await?;
7458 4 :
7459 4 : // assert freeze_and_flush exercised the initdb optimization
7460 4 : {
7461 4 : let state = tline.flush_loop_state.lock().unwrap();
7462 4 : let timeline::FlushLoopState::Running {
7463 4 : expect_initdb_optimization,
7464 4 : initdb_optimization_count,
7465 4 : } = *state
7466 4 : else {
7467 4 : panic!("unexpected state: {:?}", *state);
7468 4 : };
7469 4 : assert!(expect_initdb_optimization);
7470 4 : assert!(initdb_optimization_count > 0);
7471 4 : }
7472 4 : Ok(())
7473 4 : }
7474 :
7475 : #[tokio::test]
7476 4 : async fn test_create_guard_crash() -> anyhow::Result<()> {
7477 4 : let name = "test_create_guard_crash";
7478 4 : let harness = TenantHarness::create(name).await?;
7479 4 : {
7480 4 : let (tenant, ctx) = harness.load().await;
7481 4 : let tline = tenant
7482 4 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
7483 4 : .await?;
7484 4 : // Leave the timeline ID in [`Tenant::timelines_creating`] to exclude attempting to create it again
7485 4 : let raw_tline = tline.raw_timeline().unwrap();
7486 4 : raw_tline
7487 4 : .shutdown(super::timeline::ShutdownMode::Hard)
7488 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))
7489 4 : .await;
7490 4 : std::mem::forget(tline);
7491 4 : }
7492 4 :
7493 4 : let (tenant, _) = harness.load().await;
7494 4 : match tenant.get_timeline(TIMELINE_ID, false) {
7495 4 : Ok(_) => panic!("timeline should've been removed during load"),
7496 4 : Err(e) => {
7497 4 : assert_eq!(
7498 4 : e,
7499 4 : GetTimelineError::NotFound {
7500 4 : tenant_id: tenant.tenant_shard_id,
7501 4 : timeline_id: TIMELINE_ID,
7502 4 : }
7503 4 : )
7504 4 : }
7505 4 : }
7506 4 :
7507 4 : assert!(!harness
7508 4 : .conf
7509 4 : .timeline_path(&tenant.tenant_shard_id, &TIMELINE_ID)
7510 4 : .exists());
7511 4 :
7512 4 : Ok(())
7513 4 : }
7514 :
7515 : #[tokio::test]
7516 4 : async fn test_read_at_max_lsn() -> anyhow::Result<()> {
7517 4 : let names_algorithms = [
7518 4 : ("test_read_at_max_lsn_legacy", CompactionAlgorithm::Legacy),
7519 4 : ("test_read_at_max_lsn_tiered", CompactionAlgorithm::Tiered),
7520 4 : ];
7521 12 : for (name, algorithm) in names_algorithms {
7522 8 : test_read_at_max_lsn_algorithm(name, algorithm).await?;
7523 4 : }
7524 4 : Ok(())
7525 4 : }
7526 :
7527 8 : async fn test_read_at_max_lsn_algorithm(
7528 8 : name: &'static str,
7529 8 : compaction_algorithm: CompactionAlgorithm,
7530 8 : ) -> anyhow::Result<()> {
7531 8 : let mut harness = TenantHarness::create(name).await?;
7532 8 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
7533 8 : kind: compaction_algorithm,
7534 8 : };
7535 8 : let (tenant, ctx) = harness.load().await;
7536 8 : let tline = tenant
7537 8 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
7538 8 : .await?;
7539 :
7540 8 : let lsn = Lsn(0x10);
7541 8 : let compact = false;
7542 8 : bulk_insert_maybe_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000, compact).await?;
7543 :
7544 8 : let test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7545 8 : let read_lsn = Lsn(u64::MAX - 1);
7546 :
7547 8 : let result = tline.get(test_key, read_lsn, &ctx).await;
7548 8 : assert!(result.is_ok(), "result is not Ok: {}", result.unwrap_err());
7549 :
7550 8 : Ok(())
7551 8 : }
7552 :
7553 : #[tokio::test]
7554 4 : async fn test_metadata_scan() -> anyhow::Result<()> {
7555 4 : let harness = TenantHarness::create("test_metadata_scan").await?;
7556 4 : let (tenant, ctx) = harness.load().await;
7557 4 : let io_concurrency = IoConcurrency::spawn_for_test();
7558 4 : let tline = tenant
7559 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7560 4 : .await?;
7561 4 :
7562 4 : const NUM_KEYS: usize = 1000;
7563 4 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
7564 4 :
7565 4 : let cancel = CancellationToken::new();
7566 4 :
7567 4 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7568 4 : base_key.field1 = AUX_KEY_PREFIX;
7569 4 : let mut test_key = base_key;
7570 4 :
7571 4 : // Track when each page was last modified. Used to assert that
7572 4 : // a read sees the latest page version.
7573 4 : let mut updated = [Lsn(0); NUM_KEYS];
7574 4 :
7575 4 : let mut lsn = Lsn(0x10);
7576 4 : #[allow(clippy::needless_range_loop)]
7577 4004 : for blknum in 0..NUM_KEYS {
7578 4000 : lsn = Lsn(lsn.0 + 0x10);
7579 4000 : test_key.field6 = (blknum * STEP) as u32;
7580 4000 : let mut writer = tline.writer().await;
7581 4000 : writer
7582 4000 : .put(
7583 4000 : test_key,
7584 4000 : lsn,
7585 4000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7586 4000 : &ctx,
7587 4000 : )
7588 4000 : .await?;
7589 4000 : writer.finish_write(lsn);
7590 4000 : updated[blknum] = lsn;
7591 4000 : drop(writer);
7592 4 : }
7593 4 :
7594 4 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
7595 4 :
7596 48 : for iter in 0..=10 {
7597 4 : // Read all the blocks
7598 44000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7599 44000 : test_key.field6 = (blknum * STEP) as u32;
7600 44000 : assert_eq!(
7601 44000 : tline.get(test_key, lsn, &ctx).await?,
7602 44000 : test_img(&format!("{} at {}", blknum, last_lsn))
7603 4 : );
7604 4 : }
7605 4 :
7606 44 : let mut cnt = 0;
7607 44000 : for (key, value) in tline
7608 44 : .get_vectored_impl(
7609 44 : keyspace.clone(),
7610 44 : lsn,
7611 44 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7612 44 : &ctx,
7613 44 : )
7614 44 : .await?
7615 4 : {
7616 44000 : let blknum = key.field6 as usize;
7617 44000 : let value = value?;
7618 44000 : assert!(blknum % STEP == 0);
7619 44000 : let blknum = blknum / STEP;
7620 44000 : assert_eq!(
7621 44000 : value,
7622 44000 : test_img(&format!("{} at {}", blknum, updated[blknum]))
7623 44000 : );
7624 44000 : cnt += 1;
7625 4 : }
7626 4 :
7627 44 : assert_eq!(cnt, NUM_KEYS);
7628 4 :
7629 44044 : for _ in 0..NUM_KEYS {
7630 44000 : lsn = Lsn(lsn.0 + 0x10);
7631 44000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7632 44000 : test_key.field6 = (blknum * STEP) as u32;
7633 44000 : let mut writer = tline.writer().await;
7634 44000 : writer
7635 44000 : .put(
7636 44000 : test_key,
7637 44000 : lsn,
7638 44000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7639 44000 : &ctx,
7640 44000 : )
7641 44000 : .await?;
7642 44000 : writer.finish_write(lsn);
7643 44000 : drop(writer);
7644 44000 : updated[blknum] = lsn;
7645 4 : }
7646 4 :
7647 4 : // Perform two cycles of flush, compact, and GC
7648 132 : for round in 0..2 {
7649 88 : tline.freeze_and_flush().await?;
7650 88 : tline
7651 88 : .compact(
7652 88 : &cancel,
7653 88 : if iter % 5 == 0 && round == 0 {
7654 12 : let mut flags = EnumSet::new();
7655 12 : flags.insert(CompactFlags::ForceImageLayerCreation);
7656 12 : flags.insert(CompactFlags::ForceRepartition);
7657 12 : flags
7658 4 : } else {
7659 76 : EnumSet::empty()
7660 4 : },
7661 88 : &ctx,
7662 88 : )
7663 88 : .await?;
7664 88 : tenant
7665 88 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7666 88 : .await?;
7667 4 : }
7668 4 : }
7669 4 :
7670 4 : Ok(())
7671 4 : }
7672 :
7673 : #[tokio::test]
7674 4 : async fn test_metadata_compaction_trigger() -> anyhow::Result<()> {
7675 4 : let harness = TenantHarness::create("test_metadata_compaction_trigger").await?;
7676 4 : let (tenant, ctx) = harness.load().await;
7677 4 : let tline = tenant
7678 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7679 4 : .await?;
7680 4 :
7681 4 : let cancel = CancellationToken::new();
7682 4 :
7683 4 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7684 4 : base_key.field1 = AUX_KEY_PREFIX;
7685 4 : let test_key = base_key;
7686 4 : let mut lsn = Lsn(0x10);
7687 4 :
7688 84 : for _ in 0..20 {
7689 80 : lsn = Lsn(lsn.0 + 0x10);
7690 80 : let mut writer = tline.writer().await;
7691 80 : writer
7692 80 : .put(
7693 80 : test_key,
7694 80 : lsn,
7695 80 : &Value::Image(test_img(&format!("{} at {}", 0, lsn))),
7696 80 : &ctx,
7697 80 : )
7698 80 : .await?;
7699 80 : writer.finish_write(lsn);
7700 80 : drop(writer);
7701 80 : tline.freeze_and_flush().await?; // force create a delta layer
7702 4 : }
7703 4 :
7704 4 : let before_num_l0_delta_files =
7705 4 : tline.layers.read().await.layer_map()?.level0_deltas().len();
7706 4 :
7707 4 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
7708 4 :
7709 4 : let after_num_l0_delta_files = tline.layers.read().await.layer_map()?.level0_deltas().len();
7710 4 :
7711 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}");
7712 4 :
7713 4 : assert_eq!(
7714 4 : tline.get(test_key, lsn, &ctx).await?,
7715 4 : test_img(&format!("{} at {}", 0, lsn))
7716 4 : );
7717 4 :
7718 4 : Ok(())
7719 4 : }
7720 :
7721 : #[tokio::test]
7722 4 : async fn test_aux_file_e2e() {
7723 4 : let harness = TenantHarness::create("test_aux_file_e2e").await.unwrap();
7724 4 :
7725 4 : let (tenant, ctx) = harness.load().await;
7726 4 : let io_concurrency = IoConcurrency::spawn_for_test();
7727 4 :
7728 4 : let mut lsn = Lsn(0x08);
7729 4 :
7730 4 : let tline: Arc<Timeline> = tenant
7731 4 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
7732 4 : .await
7733 4 : .unwrap();
7734 4 :
7735 4 : {
7736 4 : lsn += 8;
7737 4 : let mut modification = tline.begin_modification(lsn);
7738 4 : modification
7739 4 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
7740 4 : .await
7741 4 : .unwrap();
7742 4 : modification.commit(&ctx).await.unwrap();
7743 4 : }
7744 4 :
7745 4 : // we can read everything from the storage
7746 4 : let files = tline
7747 4 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
7748 4 : .await
7749 4 : .unwrap();
7750 4 : assert_eq!(
7751 4 : files.get("pg_logical/mappings/test1"),
7752 4 : Some(&bytes::Bytes::from_static(b"first"))
7753 4 : );
7754 4 :
7755 4 : {
7756 4 : lsn += 8;
7757 4 : let mut modification = tline.begin_modification(lsn);
7758 4 : modification
7759 4 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
7760 4 : .await
7761 4 : .unwrap();
7762 4 : modification.commit(&ctx).await.unwrap();
7763 4 : }
7764 4 :
7765 4 : let files = tline
7766 4 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
7767 4 : .await
7768 4 : .unwrap();
7769 4 : assert_eq!(
7770 4 : files.get("pg_logical/mappings/test2"),
7771 4 : Some(&bytes::Bytes::from_static(b"second"))
7772 4 : );
7773 4 :
7774 4 : let child = tenant
7775 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(lsn), &ctx)
7776 4 : .await
7777 4 : .unwrap();
7778 4 :
7779 4 : let files = child
7780 4 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
7781 4 : .await
7782 4 : .unwrap();
7783 4 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
7784 4 : assert_eq!(files.get("pg_logical/mappings/test2"), None);
7785 4 : }
7786 :
7787 : #[tokio::test]
7788 4 : async fn test_metadata_image_creation() -> anyhow::Result<()> {
7789 4 : let harness = TenantHarness::create("test_metadata_image_creation").await?;
7790 4 : let (tenant, ctx) = harness.load().await;
7791 4 : let io_concurrency = IoConcurrency::spawn_for_test();
7792 4 : let tline = tenant
7793 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7794 4 : .await?;
7795 4 :
7796 4 : const NUM_KEYS: usize = 1000;
7797 4 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
7798 4 :
7799 4 : let cancel = CancellationToken::new();
7800 4 :
7801 4 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7802 4 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
7803 4 : let mut test_key = base_key;
7804 4 : let mut lsn = Lsn(0x10);
7805 4 :
7806 16 : async fn scan_with_statistics(
7807 16 : tline: &Timeline,
7808 16 : keyspace: &KeySpace,
7809 16 : lsn: Lsn,
7810 16 : ctx: &RequestContext,
7811 16 : io_concurrency: IoConcurrency,
7812 16 : ) -> anyhow::Result<(BTreeMap<Key, Result<Bytes, PageReconstructError>>, usize)> {
7813 16 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
7814 16 : let res = tline
7815 16 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
7816 16 : .await?;
7817 16 : Ok((res, reconstruct_state.get_delta_layers_visited() as usize))
7818 16 : }
7819 4 :
7820 4 : #[allow(clippy::needless_range_loop)]
7821 4004 : for blknum in 0..NUM_KEYS {
7822 4000 : lsn = Lsn(lsn.0 + 0x10);
7823 4000 : test_key.field6 = (blknum * STEP) as u32;
7824 4000 : let mut writer = tline.writer().await;
7825 4000 : writer
7826 4000 : .put(
7827 4000 : test_key,
7828 4000 : lsn,
7829 4000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7830 4000 : &ctx,
7831 4000 : )
7832 4000 : .await?;
7833 4000 : writer.finish_write(lsn);
7834 4000 : drop(writer);
7835 4 : }
7836 4 :
7837 4 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
7838 4 :
7839 44 : for iter in 1..=10 {
7840 40040 : for _ in 0..NUM_KEYS {
7841 40000 : lsn = Lsn(lsn.0 + 0x10);
7842 40000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7843 40000 : test_key.field6 = (blknum * STEP) as u32;
7844 40000 : let mut writer = tline.writer().await;
7845 40000 : writer
7846 40000 : .put(
7847 40000 : test_key,
7848 40000 : lsn,
7849 40000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7850 40000 : &ctx,
7851 40000 : )
7852 40000 : .await?;
7853 40000 : writer.finish_write(lsn);
7854 40000 : drop(writer);
7855 4 : }
7856 4 :
7857 40 : tline.freeze_and_flush().await?;
7858 4 :
7859 40 : if iter % 5 == 0 {
7860 8 : let (_, before_delta_file_accessed) =
7861 8 : scan_with_statistics(&tline, &keyspace, lsn, &ctx, io_concurrency.clone())
7862 8 : .await?;
7863 8 : tline
7864 8 : .compact(
7865 8 : &cancel,
7866 8 : {
7867 8 : let mut flags = EnumSet::new();
7868 8 : flags.insert(CompactFlags::ForceImageLayerCreation);
7869 8 : flags.insert(CompactFlags::ForceRepartition);
7870 8 : flags
7871 8 : },
7872 8 : &ctx,
7873 8 : )
7874 8 : .await?;
7875 8 : let (_, after_delta_file_accessed) =
7876 8 : scan_with_statistics(&tline, &keyspace, lsn, &ctx, io_concurrency.clone())
7877 8 : .await?;
7878 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}");
7879 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.
7880 8 : assert!(
7881 8 : after_delta_file_accessed <= 2,
7882 4 : "after_delta_file_accessed={after_delta_file_accessed}"
7883 4 : );
7884 32 : }
7885 4 : }
7886 4 :
7887 4 : Ok(())
7888 4 : }
7889 :
7890 : #[tokio::test]
7891 4 : async fn test_vectored_missing_data_key_reads() -> anyhow::Result<()> {
7892 4 : let harness = TenantHarness::create("test_vectored_missing_data_key_reads").await?;
7893 4 : let (tenant, ctx) = harness.load().await;
7894 4 :
7895 4 : let base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7896 4 : let base_key_child = Key::from_hex("000000000033333333444444445500000001").unwrap();
7897 4 : let base_key_nonexist = Key::from_hex("000000000033333333444444445500000002").unwrap();
7898 4 :
7899 4 : let tline = tenant
7900 4 : .create_test_timeline_with_layers(
7901 4 : TIMELINE_ID,
7902 4 : Lsn(0x10),
7903 4 : DEFAULT_PG_VERSION,
7904 4 : &ctx,
7905 4 : Vec::new(), // delta layers
7906 4 : vec![(Lsn(0x20), vec![(base_key, test_img("data key 1"))])], // image layers
7907 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
7908 4 : )
7909 4 : .await?;
7910 4 : tline.add_extra_test_dense_keyspace(KeySpace::single(base_key..(base_key_nonexist.next())));
7911 4 :
7912 4 : let child = tenant
7913 4 : .branch_timeline_test_with_layers(
7914 4 : &tline,
7915 4 : NEW_TIMELINE_ID,
7916 4 : Some(Lsn(0x20)),
7917 4 : &ctx,
7918 4 : Vec::new(), // delta layers
7919 4 : vec![(Lsn(0x30), vec![(base_key_child, test_img("data key 2"))])], // image layers
7920 4 : Lsn(0x30),
7921 4 : )
7922 4 : .await
7923 4 : .unwrap();
7924 4 :
7925 4 : let lsn = Lsn(0x30);
7926 4 :
7927 4 : // test vectored get on parent timeline
7928 4 : assert_eq!(
7929 4 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
7930 4 : Some(test_img("data key 1"))
7931 4 : );
7932 4 : assert!(get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx)
7933 4 : .await
7934 4 : .unwrap_err()
7935 4 : .is_missing_key_error());
7936 4 : assert!(
7937 4 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx)
7938 4 : .await
7939 4 : .unwrap_err()
7940 4 : .is_missing_key_error()
7941 4 : );
7942 4 :
7943 4 : // test vectored get on child timeline
7944 4 : assert_eq!(
7945 4 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
7946 4 : Some(test_img("data key 1"))
7947 4 : );
7948 4 : assert_eq!(
7949 4 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
7950 4 : Some(test_img("data key 2"))
7951 4 : );
7952 4 : assert!(
7953 4 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx)
7954 4 : .await
7955 4 : .unwrap_err()
7956 4 : .is_missing_key_error()
7957 4 : );
7958 4 :
7959 4 : Ok(())
7960 4 : }
7961 :
7962 : #[tokio::test]
7963 4 : async fn test_vectored_missing_metadata_key_reads() -> anyhow::Result<()> {
7964 4 : let harness = TenantHarness::create("test_vectored_missing_metadata_key_reads").await?;
7965 4 : let (tenant, ctx) = harness.load().await;
7966 4 : let io_concurrency = IoConcurrency::spawn_for_test();
7967 4 :
7968 4 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7969 4 : let base_key_child = Key::from_hex("620000000033333333444444445500000001").unwrap();
7970 4 : let base_key_nonexist = Key::from_hex("620000000033333333444444445500000002").unwrap();
7971 4 : let base_key_overwrite = Key::from_hex("620000000033333333444444445500000003").unwrap();
7972 4 :
7973 4 : let base_inherited_key = Key::from_hex("610000000033333333444444445500000000").unwrap();
7974 4 : let base_inherited_key_child =
7975 4 : Key::from_hex("610000000033333333444444445500000001").unwrap();
7976 4 : let base_inherited_key_nonexist =
7977 4 : Key::from_hex("610000000033333333444444445500000002").unwrap();
7978 4 : let base_inherited_key_overwrite =
7979 4 : Key::from_hex("610000000033333333444444445500000003").unwrap();
7980 4 :
7981 4 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
7982 4 : assert_eq!(base_inherited_key.field1, RELATION_SIZE_PREFIX);
7983 4 :
7984 4 : let tline = tenant
7985 4 : .create_test_timeline_with_layers(
7986 4 : TIMELINE_ID,
7987 4 : Lsn(0x10),
7988 4 : DEFAULT_PG_VERSION,
7989 4 : &ctx,
7990 4 : Vec::new(), // delta layers
7991 4 : vec![(
7992 4 : Lsn(0x20),
7993 4 : vec![
7994 4 : (base_inherited_key, test_img("metadata inherited key 1")),
7995 4 : (
7996 4 : base_inherited_key_overwrite,
7997 4 : test_img("metadata key overwrite 1a"),
7998 4 : ),
7999 4 : (base_key, test_img("metadata key 1")),
8000 4 : (base_key_overwrite, test_img("metadata key overwrite 1b")),
8001 4 : ],
8002 4 : )], // image layers
8003 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
8004 4 : )
8005 4 : .await?;
8006 4 :
8007 4 : let child = tenant
8008 4 : .branch_timeline_test_with_layers(
8009 4 : &tline,
8010 4 : NEW_TIMELINE_ID,
8011 4 : Some(Lsn(0x20)),
8012 4 : &ctx,
8013 4 : Vec::new(), // delta layers
8014 4 : vec![(
8015 4 : Lsn(0x30),
8016 4 : vec![
8017 4 : (
8018 4 : base_inherited_key_child,
8019 4 : test_img("metadata inherited key 2"),
8020 4 : ),
8021 4 : (
8022 4 : base_inherited_key_overwrite,
8023 4 : test_img("metadata key overwrite 2a"),
8024 4 : ),
8025 4 : (base_key_child, test_img("metadata key 2")),
8026 4 : (base_key_overwrite, test_img("metadata key overwrite 2b")),
8027 4 : ],
8028 4 : )], // image layers
8029 4 : Lsn(0x30),
8030 4 : )
8031 4 : .await
8032 4 : .unwrap();
8033 4 :
8034 4 : let lsn = Lsn(0x30);
8035 4 :
8036 4 : // test vectored get on parent timeline
8037 4 : assert_eq!(
8038 4 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
8039 4 : Some(test_img("metadata key 1"))
8040 4 : );
8041 4 : assert_eq!(
8042 4 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx).await?,
8043 4 : None
8044 4 : );
8045 4 : assert_eq!(
8046 4 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx).await?,
8047 4 : None
8048 4 : );
8049 4 : assert_eq!(
8050 4 : get_vectored_impl_wrapper(&tline, base_key_overwrite, lsn, &ctx).await?,
8051 4 : Some(test_img("metadata key overwrite 1b"))
8052 4 : );
8053 4 : assert_eq!(
8054 4 : get_vectored_impl_wrapper(&tline, base_inherited_key, lsn, &ctx).await?,
8055 4 : Some(test_img("metadata inherited key 1"))
8056 4 : );
8057 4 : assert_eq!(
8058 4 : get_vectored_impl_wrapper(&tline, base_inherited_key_child, lsn, &ctx).await?,
8059 4 : None
8060 4 : );
8061 4 : assert_eq!(
8062 4 : get_vectored_impl_wrapper(&tline, base_inherited_key_nonexist, lsn, &ctx).await?,
8063 4 : None
8064 4 : );
8065 4 : assert_eq!(
8066 4 : get_vectored_impl_wrapper(&tline, base_inherited_key_overwrite, lsn, &ctx).await?,
8067 4 : Some(test_img("metadata key overwrite 1a"))
8068 4 : );
8069 4 :
8070 4 : // test vectored get on child timeline
8071 4 : assert_eq!(
8072 4 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
8073 4 : None
8074 4 : );
8075 4 : assert_eq!(
8076 4 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
8077 4 : Some(test_img("metadata key 2"))
8078 4 : );
8079 4 : assert_eq!(
8080 4 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx).await?,
8081 4 : None
8082 4 : );
8083 4 : assert_eq!(
8084 4 : get_vectored_impl_wrapper(&child, base_inherited_key, lsn, &ctx).await?,
8085 4 : Some(test_img("metadata inherited key 1"))
8086 4 : );
8087 4 : assert_eq!(
8088 4 : get_vectored_impl_wrapper(&child, base_inherited_key_child, lsn, &ctx).await?,
8089 4 : Some(test_img("metadata inherited key 2"))
8090 4 : );
8091 4 : assert_eq!(
8092 4 : get_vectored_impl_wrapper(&child, base_inherited_key_nonexist, lsn, &ctx).await?,
8093 4 : None
8094 4 : );
8095 4 : assert_eq!(
8096 4 : get_vectored_impl_wrapper(&child, base_key_overwrite, lsn, &ctx).await?,
8097 4 : Some(test_img("metadata key overwrite 2b"))
8098 4 : );
8099 4 : assert_eq!(
8100 4 : get_vectored_impl_wrapper(&child, base_inherited_key_overwrite, lsn, &ctx).await?,
8101 4 : Some(test_img("metadata key overwrite 2a"))
8102 4 : );
8103 4 :
8104 4 : // test vectored scan on parent timeline
8105 4 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency.clone());
8106 4 : let res = tline
8107 4 : .get_vectored_impl(
8108 4 : KeySpace::single(Key::metadata_key_range()),
8109 4 : lsn,
8110 4 : &mut reconstruct_state,
8111 4 : &ctx,
8112 4 : )
8113 4 : .await?;
8114 4 :
8115 4 : assert_eq!(
8116 4 : res.into_iter()
8117 16 : .map(|(k, v)| (k, v.unwrap()))
8118 4 : .collect::<Vec<_>>(),
8119 4 : vec![
8120 4 : (base_inherited_key, test_img("metadata inherited key 1")),
8121 4 : (
8122 4 : base_inherited_key_overwrite,
8123 4 : test_img("metadata key overwrite 1a")
8124 4 : ),
8125 4 : (base_key, test_img("metadata key 1")),
8126 4 : (base_key_overwrite, test_img("metadata key overwrite 1b")),
8127 4 : ]
8128 4 : );
8129 4 :
8130 4 : // test vectored scan on child timeline
8131 4 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency.clone());
8132 4 : let res = child
8133 4 : .get_vectored_impl(
8134 4 : KeySpace::single(Key::metadata_key_range()),
8135 4 : lsn,
8136 4 : &mut reconstruct_state,
8137 4 : &ctx,
8138 4 : )
8139 4 : .await?;
8140 4 :
8141 4 : assert_eq!(
8142 4 : res.into_iter()
8143 20 : .map(|(k, v)| (k, v.unwrap()))
8144 4 : .collect::<Vec<_>>(),
8145 4 : vec![
8146 4 : (base_inherited_key, test_img("metadata inherited key 1")),
8147 4 : (
8148 4 : base_inherited_key_child,
8149 4 : test_img("metadata inherited key 2")
8150 4 : ),
8151 4 : (
8152 4 : base_inherited_key_overwrite,
8153 4 : test_img("metadata key overwrite 2a")
8154 4 : ),
8155 4 : (base_key_child, test_img("metadata key 2")),
8156 4 : (base_key_overwrite, test_img("metadata key overwrite 2b")),
8157 4 : ]
8158 4 : );
8159 4 :
8160 4 : Ok(())
8161 4 : }
8162 :
8163 112 : async fn get_vectored_impl_wrapper(
8164 112 : tline: &Arc<Timeline>,
8165 112 : key: Key,
8166 112 : lsn: Lsn,
8167 112 : ctx: &RequestContext,
8168 112 : ) -> Result<Option<Bytes>, GetVectoredError> {
8169 112 : let io_concurrency =
8170 112 : IoConcurrency::spawn_from_conf(tline.conf, tline.gate.enter().unwrap());
8171 112 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
8172 112 : let mut res = tline
8173 112 : .get_vectored_impl(
8174 112 : KeySpace::single(key..key.next()),
8175 112 : lsn,
8176 112 : &mut reconstruct_state,
8177 112 : ctx,
8178 112 : )
8179 112 : .await?;
8180 100 : Ok(res.pop_last().map(|(k, v)| {
8181 64 : assert_eq!(k, key);
8182 64 : v.unwrap()
8183 100 : }))
8184 112 : }
8185 :
8186 : #[tokio::test]
8187 4 : async fn test_metadata_tombstone_reads() -> anyhow::Result<()> {
8188 4 : let harness = TenantHarness::create("test_metadata_tombstone_reads").await?;
8189 4 : let (tenant, ctx) = harness.load().await;
8190 4 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
8191 4 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8192 4 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8193 4 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
8194 4 :
8195 4 : // We emulate the situation that the compaction algorithm creates an image layer that removes the tombstones
8196 4 : // Lsn 0x30 key0, key3, no key1+key2
8197 4 : // Lsn 0x20 key1+key2 tomestones
8198 4 : // Lsn 0x10 key1 in image, key2 in delta
8199 4 : let tline = tenant
8200 4 : .create_test_timeline_with_layers(
8201 4 : TIMELINE_ID,
8202 4 : Lsn(0x10),
8203 4 : DEFAULT_PG_VERSION,
8204 4 : &ctx,
8205 4 : // delta layers
8206 4 : vec![
8207 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8208 4 : Lsn(0x10)..Lsn(0x20),
8209 4 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8210 4 : ),
8211 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8212 4 : Lsn(0x20)..Lsn(0x30),
8213 4 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8214 4 : ),
8215 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8216 4 : Lsn(0x20)..Lsn(0x30),
8217 4 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8218 4 : ),
8219 4 : ],
8220 4 : // image layers
8221 4 : vec![
8222 4 : (Lsn(0x10), vec![(key1, test_img("metadata key 1"))]),
8223 4 : (
8224 4 : Lsn(0x30),
8225 4 : vec![
8226 4 : (key0, test_img("metadata key 0")),
8227 4 : (key3, test_img("metadata key 3")),
8228 4 : ],
8229 4 : ),
8230 4 : ],
8231 4 : Lsn(0x30),
8232 4 : )
8233 4 : .await?;
8234 4 :
8235 4 : let lsn = Lsn(0x30);
8236 4 : let old_lsn = Lsn(0x20);
8237 4 :
8238 4 : assert_eq!(
8239 4 : get_vectored_impl_wrapper(&tline, key0, lsn, &ctx).await?,
8240 4 : Some(test_img("metadata key 0"))
8241 4 : );
8242 4 : assert_eq!(
8243 4 : get_vectored_impl_wrapper(&tline, key1, lsn, &ctx).await?,
8244 4 : None,
8245 4 : );
8246 4 : assert_eq!(
8247 4 : get_vectored_impl_wrapper(&tline, key2, lsn, &ctx).await?,
8248 4 : None,
8249 4 : );
8250 4 : assert_eq!(
8251 4 : get_vectored_impl_wrapper(&tline, key1, old_lsn, &ctx).await?,
8252 4 : Some(Bytes::new()),
8253 4 : );
8254 4 : assert_eq!(
8255 4 : get_vectored_impl_wrapper(&tline, key2, old_lsn, &ctx).await?,
8256 4 : Some(Bytes::new()),
8257 4 : );
8258 4 : assert_eq!(
8259 4 : get_vectored_impl_wrapper(&tline, key3, lsn, &ctx).await?,
8260 4 : Some(test_img("metadata key 3"))
8261 4 : );
8262 4 :
8263 4 : Ok(())
8264 4 : }
8265 :
8266 : #[tokio::test]
8267 4 : async fn test_metadata_tombstone_image_creation() {
8268 4 : let harness = TenantHarness::create("test_metadata_tombstone_image_creation")
8269 4 : .await
8270 4 : .unwrap();
8271 4 : let (tenant, ctx) = harness.load().await;
8272 4 : let io_concurrency = IoConcurrency::spawn_for_test();
8273 4 :
8274 4 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
8275 4 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8276 4 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8277 4 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
8278 4 :
8279 4 : let tline = tenant
8280 4 : .create_test_timeline_with_layers(
8281 4 : TIMELINE_ID,
8282 4 : Lsn(0x10),
8283 4 : DEFAULT_PG_VERSION,
8284 4 : &ctx,
8285 4 : // delta layers
8286 4 : vec![
8287 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8288 4 : Lsn(0x10)..Lsn(0x20),
8289 4 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8290 4 : ),
8291 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8292 4 : Lsn(0x20)..Lsn(0x30),
8293 4 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8294 4 : ),
8295 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8296 4 : Lsn(0x20)..Lsn(0x30),
8297 4 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8298 4 : ),
8299 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8300 4 : Lsn(0x30)..Lsn(0x40),
8301 4 : vec![
8302 4 : (key0, Lsn(0x30), Value::Image(test_img("metadata key 0"))),
8303 4 : (key3, Lsn(0x30), Value::Image(test_img("metadata key 3"))),
8304 4 : ],
8305 4 : ),
8306 4 : ],
8307 4 : // image layers
8308 4 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
8309 4 : Lsn(0x40),
8310 4 : )
8311 4 : .await
8312 4 : .unwrap();
8313 4 :
8314 4 : let cancel = CancellationToken::new();
8315 4 :
8316 4 : tline
8317 4 : .compact(
8318 4 : &cancel,
8319 4 : {
8320 4 : let mut flags = EnumSet::new();
8321 4 : flags.insert(CompactFlags::ForceImageLayerCreation);
8322 4 : flags.insert(CompactFlags::ForceRepartition);
8323 4 : flags
8324 4 : },
8325 4 : &ctx,
8326 4 : )
8327 4 : .await
8328 4 : .unwrap();
8329 4 :
8330 4 : // Image layers are created at last_record_lsn
8331 4 : let images = tline
8332 4 : .inspect_image_layers(Lsn(0x40), &ctx, io_concurrency.clone())
8333 4 : .await
8334 4 : .unwrap()
8335 4 : .into_iter()
8336 36 : .filter(|(k, _)| k.is_metadata_key())
8337 4 : .collect::<Vec<_>>();
8338 4 : assert_eq!(images.len(), 2); // the image layer should only contain two existing keys, tombstones should be removed.
8339 4 : }
8340 :
8341 : #[tokio::test]
8342 4 : async fn test_metadata_tombstone_empty_image_creation() {
8343 4 : let harness = TenantHarness::create("test_metadata_tombstone_empty_image_creation")
8344 4 : .await
8345 4 : .unwrap();
8346 4 : let (tenant, ctx) = harness.load().await;
8347 4 : let io_concurrency = IoConcurrency::spawn_for_test();
8348 4 :
8349 4 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8350 4 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8351 4 :
8352 4 : let tline = tenant
8353 4 : .create_test_timeline_with_layers(
8354 4 : TIMELINE_ID,
8355 4 : Lsn(0x10),
8356 4 : DEFAULT_PG_VERSION,
8357 4 : &ctx,
8358 4 : // delta layers
8359 4 : vec![
8360 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8361 4 : Lsn(0x10)..Lsn(0x20),
8362 4 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8363 4 : ),
8364 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8365 4 : Lsn(0x20)..Lsn(0x30),
8366 4 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8367 4 : ),
8368 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8369 4 : Lsn(0x20)..Lsn(0x30),
8370 4 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8371 4 : ),
8372 4 : ],
8373 4 : // image layers
8374 4 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
8375 4 : Lsn(0x30),
8376 4 : )
8377 4 : .await
8378 4 : .unwrap();
8379 4 :
8380 4 : let cancel = CancellationToken::new();
8381 4 :
8382 4 : tline
8383 4 : .compact(
8384 4 : &cancel,
8385 4 : {
8386 4 : let mut flags = EnumSet::new();
8387 4 : flags.insert(CompactFlags::ForceImageLayerCreation);
8388 4 : flags.insert(CompactFlags::ForceRepartition);
8389 4 : flags
8390 4 : },
8391 4 : &ctx,
8392 4 : )
8393 4 : .await
8394 4 : .unwrap();
8395 4 :
8396 4 : // Image layers are created at last_record_lsn
8397 4 : let images = tline
8398 4 : .inspect_image_layers(Lsn(0x30), &ctx, io_concurrency.clone())
8399 4 : .await
8400 4 : .unwrap()
8401 4 : .into_iter()
8402 28 : .filter(|(k, _)| k.is_metadata_key())
8403 4 : .collect::<Vec<_>>();
8404 4 : assert_eq!(images.len(), 0); // the image layer should not contain tombstones, or it is not created
8405 4 : }
8406 :
8407 : #[tokio::test]
8408 4 : async fn test_simple_bottom_most_compaction_images() -> anyhow::Result<()> {
8409 4 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_images").await?;
8410 4 : let (tenant, ctx) = harness.load().await;
8411 4 : let io_concurrency = IoConcurrency::spawn_for_test();
8412 4 :
8413 204 : fn get_key(id: u32) -> Key {
8414 204 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8415 204 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8416 204 : key.field6 = id;
8417 204 : key
8418 204 : }
8419 4 :
8420 4 : // We create
8421 4 : // - one bottom-most image layer,
8422 4 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
8423 4 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
8424 4 : // - a delta layer D3 above the horizon.
8425 4 : //
8426 4 : // | D3 |
8427 4 : // | D1 |
8428 4 : // -| |-- gc horizon -----------------
8429 4 : // | | | D2 |
8430 4 : // --------- img layer ------------------
8431 4 : //
8432 4 : // What we should expact from this compaction is:
8433 4 : // | D3 |
8434 4 : // | Part of D1 |
8435 4 : // --------- img layer with D1+D2 at GC horizon------------------
8436 4 :
8437 4 : // img layer at 0x10
8438 4 : let img_layer = (0..10)
8439 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8440 4 : .collect_vec();
8441 4 :
8442 4 : let delta1 = vec![
8443 4 : (
8444 4 : get_key(1),
8445 4 : Lsn(0x20),
8446 4 : Value::Image(Bytes::from("value 1@0x20")),
8447 4 : ),
8448 4 : (
8449 4 : get_key(2),
8450 4 : Lsn(0x30),
8451 4 : Value::Image(Bytes::from("value 2@0x30")),
8452 4 : ),
8453 4 : (
8454 4 : get_key(3),
8455 4 : Lsn(0x40),
8456 4 : Value::Image(Bytes::from("value 3@0x40")),
8457 4 : ),
8458 4 : ];
8459 4 : let delta2 = vec![
8460 4 : (
8461 4 : get_key(5),
8462 4 : Lsn(0x20),
8463 4 : Value::Image(Bytes::from("value 5@0x20")),
8464 4 : ),
8465 4 : (
8466 4 : get_key(6),
8467 4 : Lsn(0x20),
8468 4 : Value::Image(Bytes::from("value 6@0x20")),
8469 4 : ),
8470 4 : ];
8471 4 : let delta3 = vec![
8472 4 : (
8473 4 : get_key(8),
8474 4 : Lsn(0x48),
8475 4 : Value::Image(Bytes::from("value 8@0x48")),
8476 4 : ),
8477 4 : (
8478 4 : get_key(9),
8479 4 : Lsn(0x48),
8480 4 : Value::Image(Bytes::from("value 9@0x48")),
8481 4 : ),
8482 4 : ];
8483 4 :
8484 4 : let tline = tenant
8485 4 : .create_test_timeline_with_layers(
8486 4 : TIMELINE_ID,
8487 4 : Lsn(0x10),
8488 4 : DEFAULT_PG_VERSION,
8489 4 : &ctx,
8490 4 : vec![
8491 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
8492 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
8493 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
8494 4 : ], // delta layers
8495 4 : vec![(Lsn(0x10), img_layer)], // image layers
8496 4 : Lsn(0x50),
8497 4 : )
8498 4 : .await?;
8499 4 : {
8500 4 : tline
8501 4 : .applied_gc_cutoff_lsn
8502 4 : .lock_for_write()
8503 4 : .store_and_unlock(Lsn(0x30))
8504 4 : .wait()
8505 4 : .await;
8506 4 : // Update GC info
8507 4 : let mut guard = tline.gc_info.write().unwrap();
8508 4 : guard.cutoffs.time = Lsn(0x30);
8509 4 : guard.cutoffs.space = Lsn(0x30);
8510 4 : }
8511 4 :
8512 4 : let expected_result = [
8513 4 : Bytes::from_static(b"value 0@0x10"),
8514 4 : Bytes::from_static(b"value 1@0x20"),
8515 4 : Bytes::from_static(b"value 2@0x30"),
8516 4 : Bytes::from_static(b"value 3@0x40"),
8517 4 : Bytes::from_static(b"value 4@0x10"),
8518 4 : Bytes::from_static(b"value 5@0x20"),
8519 4 : Bytes::from_static(b"value 6@0x20"),
8520 4 : Bytes::from_static(b"value 7@0x10"),
8521 4 : Bytes::from_static(b"value 8@0x48"),
8522 4 : Bytes::from_static(b"value 9@0x48"),
8523 4 : ];
8524 4 :
8525 40 : for (idx, expected) in expected_result.iter().enumerate() {
8526 40 : assert_eq!(
8527 40 : tline
8528 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8529 40 : .await
8530 40 : .unwrap(),
8531 4 : expected
8532 4 : );
8533 4 : }
8534 4 :
8535 4 : let cancel = CancellationToken::new();
8536 4 : tline
8537 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8538 4 : .await
8539 4 : .unwrap();
8540 4 :
8541 40 : for (idx, expected) in expected_result.iter().enumerate() {
8542 40 : assert_eq!(
8543 40 : tline
8544 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8545 40 : .await
8546 40 : .unwrap(),
8547 4 : expected
8548 4 : );
8549 4 : }
8550 4 :
8551 4 : // Check if the image layer at the GC horizon contains exactly what we want
8552 4 : let image_at_gc_horizon = tline
8553 4 : .inspect_image_layers(Lsn(0x30), &ctx, io_concurrency.clone())
8554 4 : .await
8555 4 : .unwrap()
8556 4 : .into_iter()
8557 68 : .filter(|(k, _)| k.is_metadata_key())
8558 4 : .collect::<Vec<_>>();
8559 4 :
8560 4 : assert_eq!(image_at_gc_horizon.len(), 10);
8561 4 : let expected_result = [
8562 4 : Bytes::from_static(b"value 0@0x10"),
8563 4 : Bytes::from_static(b"value 1@0x20"),
8564 4 : Bytes::from_static(b"value 2@0x30"),
8565 4 : Bytes::from_static(b"value 3@0x10"),
8566 4 : Bytes::from_static(b"value 4@0x10"),
8567 4 : Bytes::from_static(b"value 5@0x20"),
8568 4 : Bytes::from_static(b"value 6@0x20"),
8569 4 : Bytes::from_static(b"value 7@0x10"),
8570 4 : Bytes::from_static(b"value 8@0x10"),
8571 4 : Bytes::from_static(b"value 9@0x10"),
8572 4 : ];
8573 44 : for idx in 0..10 {
8574 40 : assert_eq!(
8575 40 : image_at_gc_horizon[idx],
8576 40 : (get_key(idx as u32), expected_result[idx].clone())
8577 40 : );
8578 4 : }
8579 4 :
8580 4 : // Check if old layers are removed / new layers have the expected LSN
8581 4 : let all_layers = inspect_and_sort(&tline, None).await;
8582 4 : assert_eq!(
8583 4 : all_layers,
8584 4 : vec![
8585 4 : // Image layer at GC horizon
8586 4 : PersistentLayerKey {
8587 4 : key_range: Key::MIN..Key::MAX,
8588 4 : lsn_range: Lsn(0x30)..Lsn(0x31),
8589 4 : is_delta: false
8590 4 : },
8591 4 : // The delta layer below the horizon
8592 4 : PersistentLayerKey {
8593 4 : key_range: get_key(3)..get_key(4),
8594 4 : lsn_range: Lsn(0x30)..Lsn(0x48),
8595 4 : is_delta: true
8596 4 : },
8597 4 : // The delta3 layer that should not be picked for the compaction
8598 4 : PersistentLayerKey {
8599 4 : key_range: get_key(8)..get_key(10),
8600 4 : lsn_range: Lsn(0x48)..Lsn(0x50),
8601 4 : is_delta: true
8602 4 : }
8603 4 : ]
8604 4 : );
8605 4 :
8606 4 : // increase GC horizon and compact again
8607 4 : {
8608 4 : tline
8609 4 : .applied_gc_cutoff_lsn
8610 4 : .lock_for_write()
8611 4 : .store_and_unlock(Lsn(0x40))
8612 4 : .wait()
8613 4 : .await;
8614 4 : // Update GC info
8615 4 : let mut guard = tline.gc_info.write().unwrap();
8616 4 : guard.cutoffs.time = Lsn(0x40);
8617 4 : guard.cutoffs.space = Lsn(0x40);
8618 4 : }
8619 4 : tline
8620 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8621 4 : .await
8622 4 : .unwrap();
8623 4 :
8624 4 : Ok(())
8625 4 : }
8626 :
8627 : #[cfg(feature = "testing")]
8628 : #[tokio::test]
8629 4 : async fn test_neon_test_record() -> anyhow::Result<()> {
8630 4 : let harness = TenantHarness::create("test_neon_test_record").await?;
8631 4 : let (tenant, ctx) = harness.load().await;
8632 4 :
8633 48 : fn get_key(id: u32) -> Key {
8634 48 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8635 48 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8636 48 : key.field6 = id;
8637 48 : key
8638 48 : }
8639 4 :
8640 4 : let delta1 = vec![
8641 4 : (
8642 4 : get_key(1),
8643 4 : Lsn(0x20),
8644 4 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
8645 4 : ),
8646 4 : (
8647 4 : get_key(1),
8648 4 : Lsn(0x30),
8649 4 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
8650 4 : ),
8651 4 : (get_key(2), Lsn(0x10), Value::Image("0x10".into())),
8652 4 : (
8653 4 : get_key(2),
8654 4 : Lsn(0x20),
8655 4 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
8656 4 : ),
8657 4 : (
8658 4 : get_key(2),
8659 4 : Lsn(0x30),
8660 4 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
8661 4 : ),
8662 4 : (get_key(3), Lsn(0x10), Value::Image("0x10".into())),
8663 4 : (
8664 4 : get_key(3),
8665 4 : Lsn(0x20),
8666 4 : Value::WalRecord(NeonWalRecord::wal_clear("c")),
8667 4 : ),
8668 4 : (get_key(4), Lsn(0x10), Value::Image("0x10".into())),
8669 4 : (
8670 4 : get_key(4),
8671 4 : Lsn(0x20),
8672 4 : Value::WalRecord(NeonWalRecord::wal_init("i")),
8673 4 : ),
8674 4 : ];
8675 4 : let image1 = vec![(get_key(1), "0x10".into())];
8676 4 :
8677 4 : let tline = tenant
8678 4 : .create_test_timeline_with_layers(
8679 4 : TIMELINE_ID,
8680 4 : Lsn(0x10),
8681 4 : DEFAULT_PG_VERSION,
8682 4 : &ctx,
8683 4 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
8684 4 : Lsn(0x10)..Lsn(0x40),
8685 4 : delta1,
8686 4 : )], // delta layers
8687 4 : vec![(Lsn(0x10), image1)], // image layers
8688 4 : Lsn(0x50),
8689 4 : )
8690 4 : .await?;
8691 4 :
8692 4 : assert_eq!(
8693 4 : tline.get(get_key(1), Lsn(0x50), &ctx).await?,
8694 4 : Bytes::from_static(b"0x10,0x20,0x30")
8695 4 : );
8696 4 : assert_eq!(
8697 4 : tline.get(get_key(2), Lsn(0x50), &ctx).await?,
8698 4 : Bytes::from_static(b"0x10,0x20,0x30")
8699 4 : );
8700 4 :
8701 4 : // Need to remove the limit of "Neon WAL redo requires base image".
8702 4 :
8703 4 : // assert_eq!(tline.get(get_key(3), Lsn(0x50), &ctx).await?, Bytes::new());
8704 4 : // assert_eq!(tline.get(get_key(4), Lsn(0x50), &ctx).await?, Bytes::new());
8705 4 :
8706 4 : Ok(())
8707 4 : }
8708 :
8709 : #[tokio::test(start_paused = true)]
8710 4 : async fn test_lsn_lease() -> anyhow::Result<()> {
8711 4 : let (tenant, ctx) = TenantHarness::create("test_lsn_lease")
8712 4 : .await
8713 4 : .unwrap()
8714 4 : .load()
8715 4 : .await;
8716 4 : // Advance to the lsn lease deadline so that GC is not blocked by
8717 4 : // initial transition into AttachedSingle.
8718 4 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
8719 4 : tokio::time::resume();
8720 4 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
8721 4 :
8722 4 : let end_lsn = Lsn(0x100);
8723 4 : let image_layers = (0x20..=0x90)
8724 4 : .step_by(0x10)
8725 32 : .map(|n| {
8726 32 : (
8727 32 : Lsn(n),
8728 32 : vec![(key, test_img(&format!("data key at {:x}", n)))],
8729 32 : )
8730 32 : })
8731 4 : .collect();
8732 4 :
8733 4 : let timeline = tenant
8734 4 : .create_test_timeline_with_layers(
8735 4 : TIMELINE_ID,
8736 4 : Lsn(0x10),
8737 4 : DEFAULT_PG_VERSION,
8738 4 : &ctx,
8739 4 : Vec::new(),
8740 4 : image_layers,
8741 4 : end_lsn,
8742 4 : )
8743 4 : .await?;
8744 4 :
8745 4 : let leased_lsns = [0x30, 0x50, 0x70];
8746 4 : let mut leases = Vec::new();
8747 12 : leased_lsns.iter().for_each(|n| {
8748 12 : leases.push(
8749 12 : timeline
8750 12 : .init_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)
8751 12 : .expect("lease request should succeed"),
8752 12 : );
8753 12 : });
8754 4 :
8755 4 : let updated_lease_0 = timeline
8756 4 : .renew_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)
8757 4 : .expect("lease renewal should succeed");
8758 4 : assert_eq!(
8759 4 : updated_lease_0.valid_until, leases[0].valid_until,
8760 4 : " Renewing with shorter lease should not change the lease."
8761 4 : );
8762 4 :
8763 4 : let updated_lease_1 = timeline
8764 4 : .renew_lsn_lease(
8765 4 : Lsn(leased_lsns[1]),
8766 4 : timeline.get_lsn_lease_length() * 2,
8767 4 : &ctx,
8768 4 : )
8769 4 : .expect("lease renewal should succeed");
8770 4 : assert!(
8771 4 : updated_lease_1.valid_until > leases[1].valid_until,
8772 4 : "Renewing with a long lease should renew lease with later expiration time."
8773 4 : );
8774 4 :
8775 4 : // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
8776 4 : info!(
8777 4 : "applied_gc_cutoff_lsn: {}",
8778 0 : *timeline.get_applied_gc_cutoff_lsn()
8779 4 : );
8780 4 : timeline.force_set_disk_consistent_lsn(end_lsn);
8781 4 :
8782 4 : let res = tenant
8783 4 : .gc_iteration(
8784 4 : Some(TIMELINE_ID),
8785 4 : 0,
8786 4 : Duration::ZERO,
8787 4 : &CancellationToken::new(),
8788 4 : &ctx,
8789 4 : )
8790 4 : .await
8791 4 : .unwrap();
8792 4 :
8793 4 : // Keeping everything <= Lsn(0x80) b/c leases:
8794 4 : // 0/10: initdb layer
8795 4 : // (0/20..=0/70).step_by(0x10): image layers added when creating the timeline.
8796 4 : assert_eq!(res.layers_needed_by_leases, 7);
8797 4 : // Keeping 0/90 b/c it is the latest layer.
8798 4 : assert_eq!(res.layers_not_updated, 1);
8799 4 : // Removed 0/80.
8800 4 : assert_eq!(res.layers_removed, 1);
8801 4 :
8802 4 : // Make lease on a already GC-ed LSN.
8803 4 : // 0/80 does not have a valid lease + is below latest_gc_cutoff
8804 4 : assert!(Lsn(0x80) < *timeline.get_applied_gc_cutoff_lsn());
8805 4 : timeline
8806 4 : .init_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx)
8807 4 : .expect_err("lease request on GC-ed LSN should fail");
8808 4 :
8809 4 : // Should still be able to renew a currently valid lease
8810 4 : // Assumption: original lease to is still valid for 0/50.
8811 4 : // (use `Timeline::init_lsn_lease` for testing so it always does validation)
8812 4 : timeline
8813 4 : .init_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)
8814 4 : .expect("lease renewal with validation should succeed");
8815 4 :
8816 4 : Ok(())
8817 4 : }
8818 :
8819 : #[cfg(feature = "testing")]
8820 : #[tokio::test]
8821 4 : async fn test_simple_bottom_most_compaction_deltas_1() -> anyhow::Result<()> {
8822 4 : test_simple_bottom_most_compaction_deltas_helper(
8823 4 : "test_simple_bottom_most_compaction_deltas_1",
8824 4 : false,
8825 4 : )
8826 4 : .await
8827 4 : }
8828 :
8829 : #[cfg(feature = "testing")]
8830 : #[tokio::test]
8831 4 : async fn test_simple_bottom_most_compaction_deltas_2() -> anyhow::Result<()> {
8832 4 : test_simple_bottom_most_compaction_deltas_helper(
8833 4 : "test_simple_bottom_most_compaction_deltas_2",
8834 4 : true,
8835 4 : )
8836 4 : .await
8837 4 : }
8838 :
8839 : #[cfg(feature = "testing")]
8840 8 : async fn test_simple_bottom_most_compaction_deltas_helper(
8841 8 : test_name: &'static str,
8842 8 : use_delta_bottom_layer: bool,
8843 8 : ) -> anyhow::Result<()> {
8844 8 : let harness = TenantHarness::create(test_name).await?;
8845 8 : let (tenant, ctx) = harness.load().await;
8846 :
8847 552 : fn get_key(id: u32) -> Key {
8848 552 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8849 552 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8850 552 : key.field6 = id;
8851 552 : key
8852 552 : }
8853 :
8854 : // We create
8855 : // - one bottom-most image layer,
8856 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
8857 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
8858 : // - a delta layer D3 above the horizon.
8859 : //
8860 : // | D3 |
8861 : // | D1 |
8862 : // -| |-- gc horizon -----------------
8863 : // | | | D2 |
8864 : // --------- img layer ------------------
8865 : //
8866 : // What we should expact from this compaction is:
8867 : // | D3 |
8868 : // | Part of D1 |
8869 : // --------- img layer with D1+D2 at GC horizon------------------
8870 :
8871 : // img layer at 0x10
8872 8 : let img_layer = (0..10)
8873 80 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8874 8 : .collect_vec();
8875 8 : // or, delta layer at 0x10 if `use_delta_bottom_layer` is true
8876 8 : let delta4 = (0..10)
8877 80 : .map(|id| {
8878 80 : (
8879 80 : get_key(id),
8880 80 : Lsn(0x08),
8881 80 : Value::WalRecord(NeonWalRecord::wal_init(format!("value {id}@0x10"))),
8882 80 : )
8883 80 : })
8884 8 : .collect_vec();
8885 8 :
8886 8 : let delta1 = vec![
8887 8 : (
8888 8 : get_key(1),
8889 8 : Lsn(0x20),
8890 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8891 8 : ),
8892 8 : (
8893 8 : get_key(2),
8894 8 : Lsn(0x30),
8895 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8896 8 : ),
8897 8 : (
8898 8 : get_key(3),
8899 8 : Lsn(0x28),
8900 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8901 8 : ),
8902 8 : (
8903 8 : get_key(3),
8904 8 : Lsn(0x30),
8905 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8906 8 : ),
8907 8 : (
8908 8 : get_key(3),
8909 8 : Lsn(0x40),
8910 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
8911 8 : ),
8912 8 : ];
8913 8 : let delta2 = vec![
8914 8 : (
8915 8 : get_key(5),
8916 8 : Lsn(0x20),
8917 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8918 8 : ),
8919 8 : (
8920 8 : get_key(6),
8921 8 : Lsn(0x20),
8922 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8923 8 : ),
8924 8 : ];
8925 8 : let delta3 = vec![
8926 8 : (
8927 8 : get_key(8),
8928 8 : Lsn(0x48),
8929 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8930 8 : ),
8931 8 : (
8932 8 : get_key(9),
8933 8 : Lsn(0x48),
8934 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8935 8 : ),
8936 8 : ];
8937 :
8938 8 : let tline = if use_delta_bottom_layer {
8939 4 : tenant
8940 4 : .create_test_timeline_with_layers(
8941 4 : TIMELINE_ID,
8942 4 : Lsn(0x08),
8943 4 : DEFAULT_PG_VERSION,
8944 4 : &ctx,
8945 4 : vec![
8946 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8947 4 : Lsn(0x08)..Lsn(0x10),
8948 4 : delta4,
8949 4 : ),
8950 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8951 4 : Lsn(0x20)..Lsn(0x48),
8952 4 : delta1,
8953 4 : ),
8954 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8955 4 : Lsn(0x20)..Lsn(0x48),
8956 4 : delta2,
8957 4 : ),
8958 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8959 4 : Lsn(0x48)..Lsn(0x50),
8960 4 : delta3,
8961 4 : ),
8962 4 : ], // delta layers
8963 4 : vec![], // image layers
8964 4 : Lsn(0x50),
8965 4 : )
8966 4 : .await?
8967 : } else {
8968 4 : tenant
8969 4 : .create_test_timeline_with_layers(
8970 4 : TIMELINE_ID,
8971 4 : Lsn(0x10),
8972 4 : DEFAULT_PG_VERSION,
8973 4 : &ctx,
8974 4 : vec![
8975 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8976 4 : Lsn(0x10)..Lsn(0x48),
8977 4 : delta1,
8978 4 : ),
8979 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8980 4 : Lsn(0x10)..Lsn(0x48),
8981 4 : delta2,
8982 4 : ),
8983 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8984 4 : Lsn(0x48)..Lsn(0x50),
8985 4 : delta3,
8986 4 : ),
8987 4 : ], // delta layers
8988 4 : vec![(Lsn(0x10), img_layer)], // image layers
8989 4 : Lsn(0x50),
8990 4 : )
8991 4 : .await?
8992 : };
8993 : {
8994 8 : tline
8995 8 : .applied_gc_cutoff_lsn
8996 8 : .lock_for_write()
8997 8 : .store_and_unlock(Lsn(0x30))
8998 8 : .wait()
8999 8 : .await;
9000 : // Update GC info
9001 8 : let mut guard = tline.gc_info.write().unwrap();
9002 8 : *guard = GcInfo {
9003 8 : retain_lsns: vec![],
9004 8 : cutoffs: GcCutoffs {
9005 8 : time: Lsn(0x30),
9006 8 : space: Lsn(0x30),
9007 8 : },
9008 8 : leases: Default::default(),
9009 8 : within_ancestor_pitr: false,
9010 8 : };
9011 8 : }
9012 8 :
9013 8 : let expected_result = [
9014 8 : Bytes::from_static(b"value 0@0x10"),
9015 8 : Bytes::from_static(b"value 1@0x10@0x20"),
9016 8 : Bytes::from_static(b"value 2@0x10@0x30"),
9017 8 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9018 8 : Bytes::from_static(b"value 4@0x10"),
9019 8 : Bytes::from_static(b"value 5@0x10@0x20"),
9020 8 : Bytes::from_static(b"value 6@0x10@0x20"),
9021 8 : Bytes::from_static(b"value 7@0x10"),
9022 8 : Bytes::from_static(b"value 8@0x10@0x48"),
9023 8 : Bytes::from_static(b"value 9@0x10@0x48"),
9024 8 : ];
9025 8 :
9026 8 : let expected_result_at_gc_horizon = [
9027 8 : Bytes::from_static(b"value 0@0x10"),
9028 8 : Bytes::from_static(b"value 1@0x10@0x20"),
9029 8 : Bytes::from_static(b"value 2@0x10@0x30"),
9030 8 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
9031 8 : Bytes::from_static(b"value 4@0x10"),
9032 8 : Bytes::from_static(b"value 5@0x10@0x20"),
9033 8 : Bytes::from_static(b"value 6@0x10@0x20"),
9034 8 : Bytes::from_static(b"value 7@0x10"),
9035 8 : Bytes::from_static(b"value 8@0x10"),
9036 8 : Bytes::from_static(b"value 9@0x10"),
9037 8 : ];
9038 :
9039 88 : for idx in 0..10 {
9040 80 : assert_eq!(
9041 80 : tline
9042 80 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9043 80 : .await
9044 80 : .unwrap(),
9045 80 : &expected_result[idx]
9046 : );
9047 80 : assert_eq!(
9048 80 : tline
9049 80 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
9050 80 : .await
9051 80 : .unwrap(),
9052 80 : &expected_result_at_gc_horizon[idx]
9053 : );
9054 : }
9055 :
9056 8 : let cancel = CancellationToken::new();
9057 8 : tline
9058 8 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9059 8 : .await
9060 8 : .unwrap();
9061 :
9062 88 : for idx in 0..10 {
9063 80 : assert_eq!(
9064 80 : tline
9065 80 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9066 80 : .await
9067 80 : .unwrap(),
9068 80 : &expected_result[idx]
9069 : );
9070 80 : assert_eq!(
9071 80 : tline
9072 80 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
9073 80 : .await
9074 80 : .unwrap(),
9075 80 : &expected_result_at_gc_horizon[idx]
9076 : );
9077 : }
9078 :
9079 : // increase GC horizon and compact again
9080 : {
9081 8 : tline
9082 8 : .applied_gc_cutoff_lsn
9083 8 : .lock_for_write()
9084 8 : .store_and_unlock(Lsn(0x40))
9085 8 : .wait()
9086 8 : .await;
9087 : // Update GC info
9088 8 : let mut guard = tline.gc_info.write().unwrap();
9089 8 : guard.cutoffs.time = Lsn(0x40);
9090 8 : guard.cutoffs.space = Lsn(0x40);
9091 8 : }
9092 8 : tline
9093 8 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9094 8 : .await
9095 8 : .unwrap();
9096 8 :
9097 8 : Ok(())
9098 8 : }
9099 :
9100 : #[cfg(feature = "testing")]
9101 : #[tokio::test]
9102 4 : async fn test_generate_key_retention() -> anyhow::Result<()> {
9103 4 : let harness = TenantHarness::create("test_generate_key_retention").await?;
9104 4 : let (tenant, ctx) = harness.load().await;
9105 4 : let tline = tenant
9106 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
9107 4 : .await?;
9108 4 : tline.force_advance_lsn(Lsn(0x70));
9109 4 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
9110 4 : let history = vec![
9111 4 : (
9112 4 : key,
9113 4 : Lsn(0x10),
9114 4 : Value::WalRecord(NeonWalRecord::wal_init("0x10")),
9115 4 : ),
9116 4 : (
9117 4 : key,
9118 4 : Lsn(0x20),
9119 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9120 4 : ),
9121 4 : (
9122 4 : key,
9123 4 : Lsn(0x30),
9124 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9125 4 : ),
9126 4 : (
9127 4 : key,
9128 4 : Lsn(0x40),
9129 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9130 4 : ),
9131 4 : (
9132 4 : key,
9133 4 : Lsn(0x50),
9134 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
9135 4 : ),
9136 4 : (
9137 4 : key,
9138 4 : Lsn(0x60),
9139 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9140 4 : ),
9141 4 : (
9142 4 : key,
9143 4 : Lsn(0x70),
9144 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9145 4 : ),
9146 4 : (
9147 4 : key,
9148 4 : Lsn(0x80),
9149 4 : Value::Image(Bytes::copy_from_slice(
9150 4 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9151 4 : )),
9152 4 : ),
9153 4 : (
9154 4 : key,
9155 4 : Lsn(0x90),
9156 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9157 4 : ),
9158 4 : ];
9159 4 : let res = tline
9160 4 : .generate_key_retention(
9161 4 : key,
9162 4 : &history,
9163 4 : Lsn(0x60),
9164 4 : &[Lsn(0x20), Lsn(0x40), Lsn(0x50)],
9165 4 : 3,
9166 4 : None,
9167 4 : )
9168 4 : .await
9169 4 : .unwrap();
9170 4 : let expected_res = KeyHistoryRetention {
9171 4 : below_horizon: vec![
9172 4 : (
9173 4 : Lsn(0x20),
9174 4 : KeyLogAtLsn(vec![(
9175 4 : Lsn(0x20),
9176 4 : Value::Image(Bytes::from_static(b"0x10;0x20")),
9177 4 : )]),
9178 4 : ),
9179 4 : (
9180 4 : Lsn(0x40),
9181 4 : KeyLogAtLsn(vec![
9182 4 : (
9183 4 : Lsn(0x30),
9184 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9185 4 : ),
9186 4 : (
9187 4 : Lsn(0x40),
9188 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9189 4 : ),
9190 4 : ]),
9191 4 : ),
9192 4 : (
9193 4 : Lsn(0x50),
9194 4 : KeyLogAtLsn(vec![(
9195 4 : Lsn(0x50),
9196 4 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40;0x50")),
9197 4 : )]),
9198 4 : ),
9199 4 : (
9200 4 : Lsn(0x60),
9201 4 : KeyLogAtLsn(vec![(
9202 4 : Lsn(0x60),
9203 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9204 4 : )]),
9205 4 : ),
9206 4 : ],
9207 4 : above_horizon: KeyLogAtLsn(vec![
9208 4 : (
9209 4 : Lsn(0x70),
9210 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9211 4 : ),
9212 4 : (
9213 4 : Lsn(0x80),
9214 4 : Value::Image(Bytes::copy_from_slice(
9215 4 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9216 4 : )),
9217 4 : ),
9218 4 : (
9219 4 : Lsn(0x90),
9220 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9221 4 : ),
9222 4 : ]),
9223 4 : };
9224 4 : assert_eq!(res, expected_res);
9225 4 :
9226 4 : // We expect GC-compaction to run with the original GC. This would create a situation that
9227 4 : // the original GC algorithm removes some delta layers b/c there are full image coverage,
9228 4 : // therefore causing some keys to have an incomplete history below the lowest retain LSN.
9229 4 : // For example, we have
9230 4 : // ```plain
9231 4 : // init delta @ 0x10, image @ 0x20, delta @ 0x30 (gc_horizon), image @ 0x40.
9232 4 : // ```
9233 4 : // Now the GC horizon moves up, and we have
9234 4 : // ```plain
9235 4 : // init delta @ 0x10, image @ 0x20, delta @ 0x30, image @ 0x40 (gc_horizon)
9236 4 : // ```
9237 4 : // The original GC algorithm kicks in, and removes delta @ 0x10, image @ 0x20.
9238 4 : // We will end up with
9239 4 : // ```plain
9240 4 : // delta @ 0x30, image @ 0x40 (gc_horizon)
9241 4 : // ```
9242 4 : // Now we run the GC-compaction, and this key does not have a full history.
9243 4 : // We should be able to handle this partial history and drop everything before the
9244 4 : // gc_horizon image.
9245 4 :
9246 4 : let history = vec![
9247 4 : (
9248 4 : key,
9249 4 : Lsn(0x20),
9250 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9251 4 : ),
9252 4 : (
9253 4 : key,
9254 4 : Lsn(0x30),
9255 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9256 4 : ),
9257 4 : (
9258 4 : key,
9259 4 : Lsn(0x40),
9260 4 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
9261 4 : ),
9262 4 : (
9263 4 : key,
9264 4 : Lsn(0x50),
9265 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
9266 4 : ),
9267 4 : (
9268 4 : key,
9269 4 : Lsn(0x60),
9270 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9271 4 : ),
9272 4 : (
9273 4 : key,
9274 4 : Lsn(0x70),
9275 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9276 4 : ),
9277 4 : (
9278 4 : key,
9279 4 : Lsn(0x80),
9280 4 : Value::Image(Bytes::copy_from_slice(
9281 4 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9282 4 : )),
9283 4 : ),
9284 4 : (
9285 4 : key,
9286 4 : Lsn(0x90),
9287 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9288 4 : ),
9289 4 : ];
9290 4 : let res = tline
9291 4 : .generate_key_retention(key, &history, Lsn(0x60), &[Lsn(0x40), Lsn(0x50)], 3, None)
9292 4 : .await
9293 4 : .unwrap();
9294 4 : let expected_res = KeyHistoryRetention {
9295 4 : below_horizon: vec![
9296 4 : (
9297 4 : Lsn(0x40),
9298 4 : KeyLogAtLsn(vec![(
9299 4 : Lsn(0x40),
9300 4 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
9301 4 : )]),
9302 4 : ),
9303 4 : (
9304 4 : Lsn(0x50),
9305 4 : KeyLogAtLsn(vec![(
9306 4 : Lsn(0x50),
9307 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
9308 4 : )]),
9309 4 : ),
9310 4 : (
9311 4 : Lsn(0x60),
9312 4 : KeyLogAtLsn(vec![(
9313 4 : Lsn(0x60),
9314 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9315 4 : )]),
9316 4 : ),
9317 4 : ],
9318 4 : above_horizon: KeyLogAtLsn(vec![
9319 4 : (
9320 4 : Lsn(0x70),
9321 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9322 4 : ),
9323 4 : (
9324 4 : Lsn(0x80),
9325 4 : Value::Image(Bytes::copy_from_slice(
9326 4 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9327 4 : )),
9328 4 : ),
9329 4 : (
9330 4 : Lsn(0x90),
9331 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9332 4 : ),
9333 4 : ]),
9334 4 : };
9335 4 : assert_eq!(res, expected_res);
9336 4 :
9337 4 : // In case of branch compaction, the branch itself does not have the full history, and we need to provide
9338 4 : // the ancestor image in the test case.
9339 4 :
9340 4 : let history = vec![
9341 4 : (
9342 4 : key,
9343 4 : Lsn(0x20),
9344 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9345 4 : ),
9346 4 : (
9347 4 : key,
9348 4 : Lsn(0x30),
9349 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9350 4 : ),
9351 4 : (
9352 4 : key,
9353 4 : Lsn(0x40),
9354 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9355 4 : ),
9356 4 : (
9357 4 : key,
9358 4 : Lsn(0x70),
9359 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9360 4 : ),
9361 4 : ];
9362 4 : let res = tline
9363 4 : .generate_key_retention(
9364 4 : key,
9365 4 : &history,
9366 4 : Lsn(0x60),
9367 4 : &[],
9368 4 : 3,
9369 4 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
9370 4 : )
9371 4 : .await
9372 4 : .unwrap();
9373 4 : let expected_res = KeyHistoryRetention {
9374 4 : below_horizon: vec![(
9375 4 : Lsn(0x60),
9376 4 : KeyLogAtLsn(vec![(
9377 4 : Lsn(0x60),
9378 4 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")), // use the ancestor image to reconstruct the page
9379 4 : )]),
9380 4 : )],
9381 4 : above_horizon: KeyLogAtLsn(vec![(
9382 4 : Lsn(0x70),
9383 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9384 4 : )]),
9385 4 : };
9386 4 : assert_eq!(res, expected_res);
9387 4 :
9388 4 : let history = vec![
9389 4 : (
9390 4 : key,
9391 4 : Lsn(0x20),
9392 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9393 4 : ),
9394 4 : (
9395 4 : key,
9396 4 : Lsn(0x40),
9397 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9398 4 : ),
9399 4 : (
9400 4 : key,
9401 4 : Lsn(0x60),
9402 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9403 4 : ),
9404 4 : (
9405 4 : key,
9406 4 : Lsn(0x70),
9407 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9408 4 : ),
9409 4 : ];
9410 4 : let res = tline
9411 4 : .generate_key_retention(
9412 4 : key,
9413 4 : &history,
9414 4 : Lsn(0x60),
9415 4 : &[Lsn(0x30)],
9416 4 : 3,
9417 4 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
9418 4 : )
9419 4 : .await
9420 4 : .unwrap();
9421 4 : let expected_res = KeyHistoryRetention {
9422 4 : below_horizon: vec![
9423 4 : (
9424 4 : Lsn(0x30),
9425 4 : KeyLogAtLsn(vec![(
9426 4 : Lsn(0x20),
9427 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9428 4 : )]),
9429 4 : ),
9430 4 : (
9431 4 : Lsn(0x60),
9432 4 : KeyLogAtLsn(vec![(
9433 4 : Lsn(0x60),
9434 4 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x40;0x60")),
9435 4 : )]),
9436 4 : ),
9437 4 : ],
9438 4 : above_horizon: KeyLogAtLsn(vec![(
9439 4 : Lsn(0x70),
9440 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9441 4 : )]),
9442 4 : };
9443 4 : assert_eq!(res, expected_res);
9444 4 :
9445 4 : Ok(())
9446 4 : }
9447 :
9448 : #[cfg(feature = "testing")]
9449 : #[tokio::test]
9450 4 : async fn test_simple_bottom_most_compaction_with_retain_lsns() -> anyhow::Result<()> {
9451 4 : let harness =
9452 4 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns").await?;
9453 4 : let (tenant, ctx) = harness.load().await;
9454 4 :
9455 1036 : fn get_key(id: u32) -> Key {
9456 1036 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9457 1036 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9458 1036 : key.field6 = id;
9459 1036 : key
9460 1036 : }
9461 4 :
9462 4 : let img_layer = (0..10)
9463 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9464 4 : .collect_vec();
9465 4 :
9466 4 : let delta1 = vec![
9467 4 : (
9468 4 : get_key(1),
9469 4 : Lsn(0x20),
9470 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9471 4 : ),
9472 4 : (
9473 4 : get_key(2),
9474 4 : Lsn(0x30),
9475 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9476 4 : ),
9477 4 : (
9478 4 : get_key(3),
9479 4 : Lsn(0x28),
9480 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9481 4 : ),
9482 4 : (
9483 4 : get_key(3),
9484 4 : Lsn(0x30),
9485 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9486 4 : ),
9487 4 : (
9488 4 : get_key(3),
9489 4 : Lsn(0x40),
9490 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9491 4 : ),
9492 4 : ];
9493 4 : let delta2 = vec![
9494 4 : (
9495 4 : get_key(5),
9496 4 : Lsn(0x20),
9497 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9498 4 : ),
9499 4 : (
9500 4 : get_key(6),
9501 4 : Lsn(0x20),
9502 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9503 4 : ),
9504 4 : ];
9505 4 : let delta3 = vec![
9506 4 : (
9507 4 : get_key(8),
9508 4 : Lsn(0x48),
9509 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9510 4 : ),
9511 4 : (
9512 4 : get_key(9),
9513 4 : Lsn(0x48),
9514 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9515 4 : ),
9516 4 : ];
9517 4 :
9518 4 : let tline = tenant
9519 4 : .create_test_timeline_with_layers(
9520 4 : TIMELINE_ID,
9521 4 : Lsn(0x10),
9522 4 : DEFAULT_PG_VERSION,
9523 4 : &ctx,
9524 4 : vec![
9525 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
9526 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
9527 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
9528 4 : ], // delta layers
9529 4 : vec![(Lsn(0x10), img_layer)], // image layers
9530 4 : Lsn(0x50),
9531 4 : )
9532 4 : .await?;
9533 4 : {
9534 4 : tline
9535 4 : .applied_gc_cutoff_lsn
9536 4 : .lock_for_write()
9537 4 : .store_and_unlock(Lsn(0x30))
9538 4 : .wait()
9539 4 : .await;
9540 4 : // Update GC info
9541 4 : let mut guard = tline.gc_info.write().unwrap();
9542 4 : *guard = GcInfo {
9543 4 : retain_lsns: vec![
9544 4 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
9545 4 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
9546 4 : ],
9547 4 : cutoffs: GcCutoffs {
9548 4 : time: Lsn(0x30),
9549 4 : space: Lsn(0x30),
9550 4 : },
9551 4 : leases: Default::default(),
9552 4 : within_ancestor_pitr: false,
9553 4 : };
9554 4 : }
9555 4 :
9556 4 : let expected_result = [
9557 4 : Bytes::from_static(b"value 0@0x10"),
9558 4 : Bytes::from_static(b"value 1@0x10@0x20"),
9559 4 : Bytes::from_static(b"value 2@0x10@0x30"),
9560 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9561 4 : Bytes::from_static(b"value 4@0x10"),
9562 4 : Bytes::from_static(b"value 5@0x10@0x20"),
9563 4 : Bytes::from_static(b"value 6@0x10@0x20"),
9564 4 : Bytes::from_static(b"value 7@0x10"),
9565 4 : Bytes::from_static(b"value 8@0x10@0x48"),
9566 4 : Bytes::from_static(b"value 9@0x10@0x48"),
9567 4 : ];
9568 4 :
9569 4 : let expected_result_at_gc_horizon = [
9570 4 : Bytes::from_static(b"value 0@0x10"),
9571 4 : Bytes::from_static(b"value 1@0x10@0x20"),
9572 4 : Bytes::from_static(b"value 2@0x10@0x30"),
9573 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
9574 4 : Bytes::from_static(b"value 4@0x10"),
9575 4 : Bytes::from_static(b"value 5@0x10@0x20"),
9576 4 : Bytes::from_static(b"value 6@0x10@0x20"),
9577 4 : Bytes::from_static(b"value 7@0x10"),
9578 4 : Bytes::from_static(b"value 8@0x10"),
9579 4 : Bytes::from_static(b"value 9@0x10"),
9580 4 : ];
9581 4 :
9582 4 : let expected_result_at_lsn_20 = [
9583 4 : Bytes::from_static(b"value 0@0x10"),
9584 4 : Bytes::from_static(b"value 1@0x10@0x20"),
9585 4 : Bytes::from_static(b"value 2@0x10"),
9586 4 : Bytes::from_static(b"value 3@0x10"),
9587 4 : Bytes::from_static(b"value 4@0x10"),
9588 4 : Bytes::from_static(b"value 5@0x10@0x20"),
9589 4 : Bytes::from_static(b"value 6@0x10@0x20"),
9590 4 : Bytes::from_static(b"value 7@0x10"),
9591 4 : Bytes::from_static(b"value 8@0x10"),
9592 4 : Bytes::from_static(b"value 9@0x10"),
9593 4 : ];
9594 4 :
9595 4 : let expected_result_at_lsn_10 = [
9596 4 : Bytes::from_static(b"value 0@0x10"),
9597 4 : Bytes::from_static(b"value 1@0x10"),
9598 4 : Bytes::from_static(b"value 2@0x10"),
9599 4 : Bytes::from_static(b"value 3@0x10"),
9600 4 : Bytes::from_static(b"value 4@0x10"),
9601 4 : Bytes::from_static(b"value 5@0x10"),
9602 4 : Bytes::from_static(b"value 6@0x10"),
9603 4 : Bytes::from_static(b"value 7@0x10"),
9604 4 : Bytes::from_static(b"value 8@0x10"),
9605 4 : Bytes::from_static(b"value 9@0x10"),
9606 4 : ];
9607 4 :
9608 24 : let verify_result = || async {
9609 24 : let gc_horizon = {
9610 24 : let gc_info = tline.gc_info.read().unwrap();
9611 24 : gc_info.cutoffs.time
9612 4 : };
9613 264 : for idx in 0..10 {
9614 240 : assert_eq!(
9615 240 : tline
9616 240 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9617 240 : .await
9618 240 : .unwrap(),
9619 240 : &expected_result[idx]
9620 4 : );
9621 240 : assert_eq!(
9622 240 : tline
9623 240 : .get(get_key(idx as u32), gc_horizon, &ctx)
9624 240 : .await
9625 240 : .unwrap(),
9626 240 : &expected_result_at_gc_horizon[idx]
9627 4 : );
9628 240 : assert_eq!(
9629 240 : tline
9630 240 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
9631 240 : .await
9632 240 : .unwrap(),
9633 240 : &expected_result_at_lsn_20[idx]
9634 4 : );
9635 240 : assert_eq!(
9636 240 : tline
9637 240 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
9638 240 : .await
9639 240 : .unwrap(),
9640 240 : &expected_result_at_lsn_10[idx]
9641 4 : );
9642 4 : }
9643 48 : };
9644 4 :
9645 4 : verify_result().await;
9646 4 :
9647 4 : let cancel = CancellationToken::new();
9648 4 : let mut dryrun_flags = EnumSet::new();
9649 4 : dryrun_flags.insert(CompactFlags::DryRun);
9650 4 :
9651 4 : tline
9652 4 : .compact_with_gc(
9653 4 : &cancel,
9654 4 : CompactOptions {
9655 4 : flags: dryrun_flags,
9656 4 : ..Default::default()
9657 4 : },
9658 4 : &ctx,
9659 4 : )
9660 4 : .await
9661 4 : .unwrap();
9662 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
9663 4 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
9664 4 : verify_result().await;
9665 4 :
9666 4 : tline
9667 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9668 4 : .await
9669 4 : .unwrap();
9670 4 : verify_result().await;
9671 4 :
9672 4 : // compact again
9673 4 : tline
9674 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9675 4 : .await
9676 4 : .unwrap();
9677 4 : verify_result().await;
9678 4 :
9679 4 : // increase GC horizon and compact again
9680 4 : {
9681 4 : tline
9682 4 : .applied_gc_cutoff_lsn
9683 4 : .lock_for_write()
9684 4 : .store_and_unlock(Lsn(0x38))
9685 4 : .wait()
9686 4 : .await;
9687 4 : // Update GC info
9688 4 : let mut guard = tline.gc_info.write().unwrap();
9689 4 : guard.cutoffs.time = Lsn(0x38);
9690 4 : guard.cutoffs.space = Lsn(0x38);
9691 4 : }
9692 4 : tline
9693 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9694 4 : .await
9695 4 : .unwrap();
9696 4 : verify_result().await; // no wals between 0x30 and 0x38, so we should obtain the same result
9697 4 :
9698 4 : // not increasing the GC horizon and compact again
9699 4 : tline
9700 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9701 4 : .await
9702 4 : .unwrap();
9703 4 : verify_result().await;
9704 4 :
9705 4 : Ok(())
9706 4 : }
9707 :
9708 : #[cfg(feature = "testing")]
9709 : #[tokio::test]
9710 4 : async fn test_simple_bottom_most_compaction_with_retain_lsns_single_key() -> anyhow::Result<()>
9711 4 : {
9712 4 : let harness =
9713 4 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns_single_key")
9714 4 : .await?;
9715 4 : let (tenant, ctx) = harness.load().await;
9716 4 :
9717 704 : fn get_key(id: u32) -> Key {
9718 704 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9719 704 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9720 704 : key.field6 = id;
9721 704 : key
9722 704 : }
9723 4 :
9724 4 : let img_layer = (0..10)
9725 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9726 4 : .collect_vec();
9727 4 :
9728 4 : let delta1 = vec![
9729 4 : (
9730 4 : get_key(1),
9731 4 : Lsn(0x20),
9732 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9733 4 : ),
9734 4 : (
9735 4 : get_key(1),
9736 4 : Lsn(0x28),
9737 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9738 4 : ),
9739 4 : ];
9740 4 : let delta2 = vec![
9741 4 : (
9742 4 : get_key(1),
9743 4 : Lsn(0x30),
9744 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9745 4 : ),
9746 4 : (
9747 4 : get_key(1),
9748 4 : Lsn(0x38),
9749 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
9750 4 : ),
9751 4 : ];
9752 4 : let delta3 = vec![
9753 4 : (
9754 4 : get_key(8),
9755 4 : Lsn(0x48),
9756 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9757 4 : ),
9758 4 : (
9759 4 : get_key(9),
9760 4 : Lsn(0x48),
9761 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9762 4 : ),
9763 4 : ];
9764 4 :
9765 4 : let tline = tenant
9766 4 : .create_test_timeline_with_layers(
9767 4 : TIMELINE_ID,
9768 4 : Lsn(0x10),
9769 4 : DEFAULT_PG_VERSION,
9770 4 : &ctx,
9771 4 : vec![
9772 4 : // delta1 and delta 2 only contain a single key but multiple updates
9773 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x30), delta1),
9774 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
9775 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x50), delta3),
9776 4 : ], // delta layers
9777 4 : vec![(Lsn(0x10), img_layer)], // image layers
9778 4 : Lsn(0x50),
9779 4 : )
9780 4 : .await?;
9781 4 : {
9782 4 : tline
9783 4 : .applied_gc_cutoff_lsn
9784 4 : .lock_for_write()
9785 4 : .store_and_unlock(Lsn(0x30))
9786 4 : .wait()
9787 4 : .await;
9788 4 : // Update GC info
9789 4 : let mut guard = tline.gc_info.write().unwrap();
9790 4 : *guard = GcInfo {
9791 4 : retain_lsns: vec![
9792 4 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
9793 4 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
9794 4 : ],
9795 4 : cutoffs: GcCutoffs {
9796 4 : time: Lsn(0x30),
9797 4 : space: Lsn(0x30),
9798 4 : },
9799 4 : leases: Default::default(),
9800 4 : within_ancestor_pitr: false,
9801 4 : };
9802 4 : }
9803 4 :
9804 4 : let expected_result = [
9805 4 : Bytes::from_static(b"value 0@0x10"),
9806 4 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
9807 4 : Bytes::from_static(b"value 2@0x10"),
9808 4 : Bytes::from_static(b"value 3@0x10"),
9809 4 : Bytes::from_static(b"value 4@0x10"),
9810 4 : Bytes::from_static(b"value 5@0x10"),
9811 4 : Bytes::from_static(b"value 6@0x10"),
9812 4 : Bytes::from_static(b"value 7@0x10"),
9813 4 : Bytes::from_static(b"value 8@0x10@0x48"),
9814 4 : Bytes::from_static(b"value 9@0x10@0x48"),
9815 4 : ];
9816 4 :
9817 4 : let expected_result_at_gc_horizon = [
9818 4 : Bytes::from_static(b"value 0@0x10"),
9819 4 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
9820 4 : Bytes::from_static(b"value 2@0x10"),
9821 4 : Bytes::from_static(b"value 3@0x10"),
9822 4 : Bytes::from_static(b"value 4@0x10"),
9823 4 : Bytes::from_static(b"value 5@0x10"),
9824 4 : Bytes::from_static(b"value 6@0x10"),
9825 4 : Bytes::from_static(b"value 7@0x10"),
9826 4 : Bytes::from_static(b"value 8@0x10"),
9827 4 : Bytes::from_static(b"value 9@0x10"),
9828 4 : ];
9829 4 :
9830 4 : let expected_result_at_lsn_20 = [
9831 4 : Bytes::from_static(b"value 0@0x10"),
9832 4 : Bytes::from_static(b"value 1@0x10@0x20"),
9833 4 : Bytes::from_static(b"value 2@0x10"),
9834 4 : Bytes::from_static(b"value 3@0x10"),
9835 4 : Bytes::from_static(b"value 4@0x10"),
9836 4 : Bytes::from_static(b"value 5@0x10"),
9837 4 : Bytes::from_static(b"value 6@0x10"),
9838 4 : Bytes::from_static(b"value 7@0x10"),
9839 4 : Bytes::from_static(b"value 8@0x10"),
9840 4 : Bytes::from_static(b"value 9@0x10"),
9841 4 : ];
9842 4 :
9843 4 : let expected_result_at_lsn_10 = [
9844 4 : Bytes::from_static(b"value 0@0x10"),
9845 4 : Bytes::from_static(b"value 1@0x10"),
9846 4 : Bytes::from_static(b"value 2@0x10"),
9847 4 : Bytes::from_static(b"value 3@0x10"),
9848 4 : Bytes::from_static(b"value 4@0x10"),
9849 4 : Bytes::from_static(b"value 5@0x10"),
9850 4 : Bytes::from_static(b"value 6@0x10"),
9851 4 : Bytes::from_static(b"value 7@0x10"),
9852 4 : Bytes::from_static(b"value 8@0x10"),
9853 4 : Bytes::from_static(b"value 9@0x10"),
9854 4 : ];
9855 4 :
9856 16 : let verify_result = || async {
9857 16 : let gc_horizon = {
9858 16 : let gc_info = tline.gc_info.read().unwrap();
9859 16 : gc_info.cutoffs.time
9860 4 : };
9861 176 : for idx in 0..10 {
9862 160 : assert_eq!(
9863 160 : tline
9864 160 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9865 160 : .await
9866 160 : .unwrap(),
9867 160 : &expected_result[idx]
9868 4 : );
9869 160 : assert_eq!(
9870 160 : tline
9871 160 : .get(get_key(idx as u32), gc_horizon, &ctx)
9872 160 : .await
9873 160 : .unwrap(),
9874 160 : &expected_result_at_gc_horizon[idx]
9875 4 : );
9876 160 : assert_eq!(
9877 160 : tline
9878 160 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
9879 160 : .await
9880 160 : .unwrap(),
9881 160 : &expected_result_at_lsn_20[idx]
9882 4 : );
9883 160 : assert_eq!(
9884 160 : tline
9885 160 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
9886 160 : .await
9887 160 : .unwrap(),
9888 160 : &expected_result_at_lsn_10[idx]
9889 4 : );
9890 4 : }
9891 32 : };
9892 4 :
9893 4 : verify_result().await;
9894 4 :
9895 4 : let cancel = CancellationToken::new();
9896 4 : let mut dryrun_flags = EnumSet::new();
9897 4 : dryrun_flags.insert(CompactFlags::DryRun);
9898 4 :
9899 4 : tline
9900 4 : .compact_with_gc(
9901 4 : &cancel,
9902 4 : CompactOptions {
9903 4 : flags: dryrun_flags,
9904 4 : ..Default::default()
9905 4 : },
9906 4 : &ctx,
9907 4 : )
9908 4 : .await
9909 4 : .unwrap();
9910 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
9911 4 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
9912 4 : verify_result().await;
9913 4 :
9914 4 : tline
9915 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9916 4 : .await
9917 4 : .unwrap();
9918 4 : verify_result().await;
9919 4 :
9920 4 : // compact again
9921 4 : tline
9922 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9923 4 : .await
9924 4 : .unwrap();
9925 4 : verify_result().await;
9926 4 :
9927 4 : Ok(())
9928 4 : }
9929 :
9930 : #[cfg(feature = "testing")]
9931 : #[tokio::test]
9932 4 : async fn test_simple_bottom_most_compaction_on_branch() -> anyhow::Result<()> {
9933 4 : use models::CompactLsnRange;
9934 4 :
9935 4 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_on_branch").await?;
9936 4 : let (tenant, ctx) = harness.load().await;
9937 4 :
9938 332 : fn get_key(id: u32) -> Key {
9939 332 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
9940 332 : key.field6 = id;
9941 332 : key
9942 332 : }
9943 4 :
9944 4 : let img_layer = (0..10)
9945 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9946 4 : .collect_vec();
9947 4 :
9948 4 : let delta1 = vec![
9949 4 : (
9950 4 : get_key(1),
9951 4 : Lsn(0x20),
9952 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9953 4 : ),
9954 4 : (
9955 4 : get_key(2),
9956 4 : Lsn(0x30),
9957 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9958 4 : ),
9959 4 : (
9960 4 : get_key(3),
9961 4 : Lsn(0x28),
9962 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9963 4 : ),
9964 4 : (
9965 4 : get_key(3),
9966 4 : Lsn(0x30),
9967 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9968 4 : ),
9969 4 : (
9970 4 : get_key(3),
9971 4 : Lsn(0x40),
9972 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9973 4 : ),
9974 4 : ];
9975 4 : let delta2 = vec![
9976 4 : (
9977 4 : get_key(5),
9978 4 : Lsn(0x20),
9979 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9980 4 : ),
9981 4 : (
9982 4 : get_key(6),
9983 4 : Lsn(0x20),
9984 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9985 4 : ),
9986 4 : ];
9987 4 : let delta3 = vec![
9988 4 : (
9989 4 : get_key(8),
9990 4 : Lsn(0x48),
9991 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9992 4 : ),
9993 4 : (
9994 4 : get_key(9),
9995 4 : Lsn(0x48),
9996 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9997 4 : ),
9998 4 : ];
9999 4 :
10000 4 : let parent_tline = tenant
10001 4 : .create_test_timeline_with_layers(
10002 4 : TIMELINE_ID,
10003 4 : Lsn(0x10),
10004 4 : DEFAULT_PG_VERSION,
10005 4 : &ctx,
10006 4 : vec![], // delta layers
10007 4 : vec![(Lsn(0x18), img_layer)], // image layers
10008 4 : Lsn(0x18),
10009 4 : )
10010 4 : .await?;
10011 4 :
10012 4 : parent_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
10013 4 :
10014 4 : let branch_tline = tenant
10015 4 : .branch_timeline_test_with_layers(
10016 4 : &parent_tline,
10017 4 : NEW_TIMELINE_ID,
10018 4 : Some(Lsn(0x18)),
10019 4 : &ctx,
10020 4 : vec![
10021 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
10022 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
10023 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
10024 4 : ], // delta layers
10025 4 : vec![], // image layers
10026 4 : Lsn(0x50),
10027 4 : )
10028 4 : .await?;
10029 4 :
10030 4 : branch_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
10031 4 :
10032 4 : {
10033 4 : parent_tline
10034 4 : .applied_gc_cutoff_lsn
10035 4 : .lock_for_write()
10036 4 : .store_and_unlock(Lsn(0x10))
10037 4 : .wait()
10038 4 : .await;
10039 4 : // Update GC info
10040 4 : let mut guard = parent_tline.gc_info.write().unwrap();
10041 4 : *guard = GcInfo {
10042 4 : retain_lsns: vec![(Lsn(0x18), branch_tline.timeline_id, MaybeOffloaded::No)],
10043 4 : cutoffs: GcCutoffs {
10044 4 : time: Lsn(0x10),
10045 4 : space: Lsn(0x10),
10046 4 : },
10047 4 : leases: Default::default(),
10048 4 : within_ancestor_pitr: false,
10049 4 : };
10050 4 : }
10051 4 :
10052 4 : {
10053 4 : branch_tline
10054 4 : .applied_gc_cutoff_lsn
10055 4 : .lock_for_write()
10056 4 : .store_and_unlock(Lsn(0x50))
10057 4 : .wait()
10058 4 : .await;
10059 4 : // Update GC info
10060 4 : let mut guard = branch_tline.gc_info.write().unwrap();
10061 4 : *guard = GcInfo {
10062 4 : retain_lsns: vec![(Lsn(0x40), branch_tline.timeline_id, MaybeOffloaded::No)],
10063 4 : cutoffs: GcCutoffs {
10064 4 : time: Lsn(0x50),
10065 4 : space: Lsn(0x50),
10066 4 : },
10067 4 : leases: Default::default(),
10068 4 : within_ancestor_pitr: false,
10069 4 : };
10070 4 : }
10071 4 :
10072 4 : let expected_result_at_gc_horizon = [
10073 4 : Bytes::from_static(b"value 0@0x10"),
10074 4 : Bytes::from_static(b"value 1@0x10@0x20"),
10075 4 : Bytes::from_static(b"value 2@0x10@0x30"),
10076 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
10077 4 : Bytes::from_static(b"value 4@0x10"),
10078 4 : Bytes::from_static(b"value 5@0x10@0x20"),
10079 4 : Bytes::from_static(b"value 6@0x10@0x20"),
10080 4 : Bytes::from_static(b"value 7@0x10"),
10081 4 : Bytes::from_static(b"value 8@0x10@0x48"),
10082 4 : Bytes::from_static(b"value 9@0x10@0x48"),
10083 4 : ];
10084 4 :
10085 4 : let expected_result_at_lsn_40 = [
10086 4 : Bytes::from_static(b"value 0@0x10"),
10087 4 : Bytes::from_static(b"value 1@0x10@0x20"),
10088 4 : Bytes::from_static(b"value 2@0x10@0x30"),
10089 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
10090 4 : Bytes::from_static(b"value 4@0x10"),
10091 4 : Bytes::from_static(b"value 5@0x10@0x20"),
10092 4 : Bytes::from_static(b"value 6@0x10@0x20"),
10093 4 : Bytes::from_static(b"value 7@0x10"),
10094 4 : Bytes::from_static(b"value 8@0x10"),
10095 4 : Bytes::from_static(b"value 9@0x10"),
10096 4 : ];
10097 4 :
10098 12 : let verify_result = || async {
10099 132 : for idx in 0..10 {
10100 120 : assert_eq!(
10101 120 : branch_tline
10102 120 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10103 120 : .await
10104 120 : .unwrap(),
10105 120 : &expected_result_at_gc_horizon[idx]
10106 4 : );
10107 120 : assert_eq!(
10108 120 : branch_tline
10109 120 : .get(get_key(idx as u32), Lsn(0x40), &ctx)
10110 120 : .await
10111 120 : .unwrap(),
10112 120 : &expected_result_at_lsn_40[idx]
10113 4 : );
10114 4 : }
10115 24 : };
10116 4 :
10117 4 : verify_result().await;
10118 4 :
10119 4 : let cancel = CancellationToken::new();
10120 4 : branch_tline
10121 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10122 4 : .await
10123 4 : .unwrap();
10124 4 :
10125 4 : verify_result().await;
10126 4 :
10127 4 : // Piggyback a compaction with above_lsn. Ensure it works correctly when the specified LSN intersects with the layer files.
10128 4 : // Now we already have a single large delta layer, so the compaction min_layer_lsn should be the same as ancestor LSN (0x18).
10129 4 : branch_tline
10130 4 : .compact_with_gc(
10131 4 : &cancel,
10132 4 : CompactOptions {
10133 4 : compact_lsn_range: Some(CompactLsnRange::above(Lsn(0x40))),
10134 4 : ..Default::default()
10135 4 : },
10136 4 : &ctx,
10137 4 : )
10138 4 : .await
10139 4 : .unwrap();
10140 4 :
10141 4 : verify_result().await;
10142 4 :
10143 4 : Ok(())
10144 4 : }
10145 :
10146 : // Regression test for https://github.com/neondatabase/neon/issues/9012
10147 : // Create an image arrangement where we have to read at different LSN ranges
10148 : // from a delta layer. This is achieved by overlapping an image layer on top of
10149 : // a delta layer. Like so:
10150 : //
10151 : // A B
10152 : // +----------------+ -> delta_layer
10153 : // | | ^ lsn
10154 : // | =========|-> nested_image_layer |
10155 : // | C | |
10156 : // +----------------+ |
10157 : // ======== -> baseline_image_layer +-------> key
10158 : //
10159 : //
10160 : // When querying the key range [A, B) we need to read at different LSN ranges
10161 : // for [A, C) and [C, B). This test checks that the described edge case is handled correctly.
10162 : #[cfg(feature = "testing")]
10163 : #[tokio::test]
10164 4 : async fn test_vectored_read_with_nested_image_layer() -> anyhow::Result<()> {
10165 4 : let harness = TenantHarness::create("test_vectored_read_with_nested_image_layer").await?;
10166 4 : let (tenant, ctx) = harness.load().await;
10167 4 :
10168 4 : let will_init_keys = [2, 6];
10169 88 : fn get_key(id: u32) -> Key {
10170 88 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
10171 88 : key.field6 = id;
10172 88 : key
10173 88 : }
10174 4 :
10175 4 : let mut expected_key_values = HashMap::new();
10176 4 :
10177 4 : let baseline_image_layer_lsn = Lsn(0x10);
10178 4 : let mut baseline_img_layer = Vec::new();
10179 24 : for i in 0..5 {
10180 20 : let key = get_key(i);
10181 20 : let value = format!("value {i}@{baseline_image_layer_lsn}");
10182 20 :
10183 20 : let removed = expected_key_values.insert(key, value.clone());
10184 20 : assert!(removed.is_none());
10185 4 :
10186 20 : baseline_img_layer.push((key, Bytes::from(value)));
10187 4 : }
10188 4 :
10189 4 : let nested_image_layer_lsn = Lsn(0x50);
10190 4 : let mut nested_img_layer = Vec::new();
10191 24 : for i in 5..10 {
10192 20 : let key = get_key(i);
10193 20 : let value = format!("value {i}@{nested_image_layer_lsn}");
10194 20 :
10195 20 : let removed = expected_key_values.insert(key, value.clone());
10196 20 : assert!(removed.is_none());
10197 4 :
10198 20 : nested_img_layer.push((key, Bytes::from(value)));
10199 4 : }
10200 4 :
10201 4 : let mut delta_layer_spec = Vec::default();
10202 4 : let delta_layer_start_lsn = Lsn(0x20);
10203 4 : let mut delta_layer_end_lsn = delta_layer_start_lsn;
10204 4 :
10205 44 : for i in 0..10 {
10206 40 : let key = get_key(i);
10207 40 : let key_in_nested = nested_img_layer
10208 40 : .iter()
10209 160 : .any(|(key_with_img, _)| *key_with_img == key);
10210 40 : let lsn = {
10211 40 : if key_in_nested {
10212 20 : Lsn(nested_image_layer_lsn.0 + 0x10)
10213 4 : } else {
10214 20 : delta_layer_start_lsn
10215 4 : }
10216 4 : };
10217 4 :
10218 40 : let will_init = will_init_keys.contains(&i);
10219 40 : if will_init {
10220 8 : delta_layer_spec.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init(""))));
10221 8 :
10222 8 : expected_key_values.insert(key, "".to_string());
10223 32 : } else {
10224 32 : let delta = format!("@{lsn}");
10225 32 : delta_layer_spec.push((
10226 32 : key,
10227 32 : lsn,
10228 32 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
10229 32 : ));
10230 32 :
10231 32 : expected_key_values
10232 32 : .get_mut(&key)
10233 32 : .expect("An image exists for each key")
10234 32 : .push_str(delta.as_str());
10235 32 : }
10236 40 : delta_layer_end_lsn = std::cmp::max(delta_layer_start_lsn, lsn);
10237 4 : }
10238 4 :
10239 4 : delta_layer_end_lsn = Lsn(delta_layer_end_lsn.0 + 1);
10240 4 :
10241 4 : assert!(
10242 4 : nested_image_layer_lsn > delta_layer_start_lsn
10243 4 : && nested_image_layer_lsn < delta_layer_end_lsn
10244 4 : );
10245 4 :
10246 4 : let tline = tenant
10247 4 : .create_test_timeline_with_layers(
10248 4 : TIMELINE_ID,
10249 4 : baseline_image_layer_lsn,
10250 4 : DEFAULT_PG_VERSION,
10251 4 : &ctx,
10252 4 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
10253 4 : delta_layer_start_lsn..delta_layer_end_lsn,
10254 4 : delta_layer_spec,
10255 4 : )], // delta layers
10256 4 : vec![
10257 4 : (baseline_image_layer_lsn, baseline_img_layer),
10258 4 : (nested_image_layer_lsn, nested_img_layer),
10259 4 : ], // image layers
10260 4 : delta_layer_end_lsn,
10261 4 : )
10262 4 : .await?;
10263 4 :
10264 4 : let keyspace = KeySpace::single(get_key(0)..get_key(10));
10265 4 : let results = tline
10266 4 : .get_vectored(
10267 4 : keyspace,
10268 4 : delta_layer_end_lsn,
10269 4 : IoConcurrency::sequential(),
10270 4 : &ctx,
10271 4 : )
10272 4 : .await
10273 4 : .expect("No vectored errors");
10274 44 : for (key, res) in results {
10275 40 : let value = res.expect("No key errors");
10276 40 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
10277 40 : assert_eq!(value, Bytes::from(expected_value));
10278 4 : }
10279 4 :
10280 4 : Ok(())
10281 4 : }
10282 :
10283 428 : fn sort_layer_key(k1: &PersistentLayerKey, k2: &PersistentLayerKey) -> std::cmp::Ordering {
10284 428 : (
10285 428 : k1.is_delta,
10286 428 : k1.key_range.start,
10287 428 : k1.key_range.end,
10288 428 : k1.lsn_range.start,
10289 428 : k1.lsn_range.end,
10290 428 : )
10291 428 : .cmp(&(
10292 428 : k2.is_delta,
10293 428 : k2.key_range.start,
10294 428 : k2.key_range.end,
10295 428 : k2.lsn_range.start,
10296 428 : k2.lsn_range.end,
10297 428 : ))
10298 428 : }
10299 :
10300 48 : async fn inspect_and_sort(
10301 48 : tline: &Arc<Timeline>,
10302 48 : filter: Option<std::ops::Range<Key>>,
10303 48 : ) -> Vec<PersistentLayerKey> {
10304 48 : let mut all_layers = tline.inspect_historic_layers().await.unwrap();
10305 48 : if let Some(filter) = filter {
10306 216 : all_layers.retain(|layer| overlaps_with(&layer.key_range, &filter));
10307 44 : }
10308 48 : all_layers.sort_by(sort_layer_key);
10309 48 : all_layers
10310 48 : }
10311 :
10312 : #[cfg(feature = "testing")]
10313 44 : fn check_layer_map_key_eq(
10314 44 : mut left: Vec<PersistentLayerKey>,
10315 44 : mut right: Vec<PersistentLayerKey>,
10316 44 : ) {
10317 44 : left.sort_by(sort_layer_key);
10318 44 : right.sort_by(sort_layer_key);
10319 44 : if left != right {
10320 0 : eprintln!("---LEFT---");
10321 0 : for left in left.iter() {
10322 0 : eprintln!("{}", left);
10323 0 : }
10324 0 : eprintln!("---RIGHT---");
10325 0 : for right in right.iter() {
10326 0 : eprintln!("{}", right);
10327 0 : }
10328 0 : assert_eq!(left, right);
10329 44 : }
10330 44 : }
10331 :
10332 : #[cfg(feature = "testing")]
10333 : #[tokio::test]
10334 4 : async fn test_simple_partial_bottom_most_compaction() -> anyhow::Result<()> {
10335 4 : let harness = TenantHarness::create("test_simple_partial_bottom_most_compaction").await?;
10336 4 : let (tenant, ctx) = harness.load().await;
10337 4 :
10338 364 : fn get_key(id: u32) -> Key {
10339 364 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
10340 364 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
10341 364 : key.field6 = id;
10342 364 : key
10343 364 : }
10344 4 :
10345 4 : // img layer at 0x10
10346 4 : let img_layer = (0..10)
10347 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10348 4 : .collect_vec();
10349 4 :
10350 4 : let delta1 = vec![
10351 4 : (
10352 4 : get_key(1),
10353 4 : Lsn(0x20),
10354 4 : Value::Image(Bytes::from("value 1@0x20")),
10355 4 : ),
10356 4 : (
10357 4 : get_key(2),
10358 4 : Lsn(0x30),
10359 4 : Value::Image(Bytes::from("value 2@0x30")),
10360 4 : ),
10361 4 : (
10362 4 : get_key(3),
10363 4 : Lsn(0x40),
10364 4 : Value::Image(Bytes::from("value 3@0x40")),
10365 4 : ),
10366 4 : ];
10367 4 : let delta2 = vec![
10368 4 : (
10369 4 : get_key(5),
10370 4 : Lsn(0x20),
10371 4 : Value::Image(Bytes::from("value 5@0x20")),
10372 4 : ),
10373 4 : (
10374 4 : get_key(6),
10375 4 : Lsn(0x20),
10376 4 : Value::Image(Bytes::from("value 6@0x20")),
10377 4 : ),
10378 4 : ];
10379 4 : let delta3 = vec![
10380 4 : (
10381 4 : get_key(8),
10382 4 : Lsn(0x48),
10383 4 : Value::Image(Bytes::from("value 8@0x48")),
10384 4 : ),
10385 4 : (
10386 4 : get_key(9),
10387 4 : Lsn(0x48),
10388 4 : Value::Image(Bytes::from("value 9@0x48")),
10389 4 : ),
10390 4 : ];
10391 4 :
10392 4 : let tline = tenant
10393 4 : .create_test_timeline_with_layers(
10394 4 : TIMELINE_ID,
10395 4 : Lsn(0x10),
10396 4 : DEFAULT_PG_VERSION,
10397 4 : &ctx,
10398 4 : vec![
10399 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
10400 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
10401 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
10402 4 : ], // delta layers
10403 4 : vec![(Lsn(0x10), img_layer)], // image layers
10404 4 : Lsn(0x50),
10405 4 : )
10406 4 : .await?;
10407 4 :
10408 4 : {
10409 4 : tline
10410 4 : .applied_gc_cutoff_lsn
10411 4 : .lock_for_write()
10412 4 : .store_and_unlock(Lsn(0x30))
10413 4 : .wait()
10414 4 : .await;
10415 4 : // Update GC info
10416 4 : let mut guard = tline.gc_info.write().unwrap();
10417 4 : *guard = GcInfo {
10418 4 : retain_lsns: vec![(Lsn(0x20), tline.timeline_id, MaybeOffloaded::No)],
10419 4 : cutoffs: GcCutoffs {
10420 4 : time: Lsn(0x30),
10421 4 : space: Lsn(0x30),
10422 4 : },
10423 4 : leases: Default::default(),
10424 4 : within_ancestor_pitr: false,
10425 4 : };
10426 4 : }
10427 4 :
10428 4 : let cancel = CancellationToken::new();
10429 4 :
10430 4 : // Do a partial compaction on key range 0..2
10431 4 : tline
10432 4 : .compact_with_gc(
10433 4 : &cancel,
10434 4 : CompactOptions {
10435 4 : flags: EnumSet::new(),
10436 4 : compact_key_range: Some((get_key(0)..get_key(2)).into()),
10437 4 : ..Default::default()
10438 4 : },
10439 4 : &ctx,
10440 4 : )
10441 4 : .await
10442 4 : .unwrap();
10443 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10444 4 : check_layer_map_key_eq(
10445 4 : all_layers,
10446 4 : vec![
10447 4 : // newly-generated image layer for the partial compaction range 0-2
10448 4 : PersistentLayerKey {
10449 4 : key_range: get_key(0)..get_key(2),
10450 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10451 4 : is_delta: false,
10452 4 : },
10453 4 : PersistentLayerKey {
10454 4 : key_range: get_key(0)..get_key(10),
10455 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
10456 4 : is_delta: false,
10457 4 : },
10458 4 : // delta1 is split and the second part is rewritten
10459 4 : PersistentLayerKey {
10460 4 : key_range: get_key(2)..get_key(4),
10461 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10462 4 : is_delta: true,
10463 4 : },
10464 4 : PersistentLayerKey {
10465 4 : key_range: get_key(5)..get_key(7),
10466 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10467 4 : is_delta: true,
10468 4 : },
10469 4 : PersistentLayerKey {
10470 4 : key_range: get_key(8)..get_key(10),
10471 4 : lsn_range: Lsn(0x48)..Lsn(0x50),
10472 4 : is_delta: true,
10473 4 : },
10474 4 : ],
10475 4 : );
10476 4 :
10477 4 : // Do a partial compaction on key range 2..4
10478 4 : tline
10479 4 : .compact_with_gc(
10480 4 : &cancel,
10481 4 : CompactOptions {
10482 4 : flags: EnumSet::new(),
10483 4 : compact_key_range: Some((get_key(2)..get_key(4)).into()),
10484 4 : ..Default::default()
10485 4 : },
10486 4 : &ctx,
10487 4 : )
10488 4 : .await
10489 4 : .unwrap();
10490 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10491 4 : check_layer_map_key_eq(
10492 4 : all_layers,
10493 4 : vec![
10494 4 : PersistentLayerKey {
10495 4 : key_range: get_key(0)..get_key(2),
10496 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10497 4 : is_delta: false,
10498 4 : },
10499 4 : PersistentLayerKey {
10500 4 : key_range: get_key(0)..get_key(10),
10501 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
10502 4 : is_delta: false,
10503 4 : },
10504 4 : // image layer generated for the compaction range 2-4
10505 4 : PersistentLayerKey {
10506 4 : key_range: get_key(2)..get_key(4),
10507 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10508 4 : is_delta: false,
10509 4 : },
10510 4 : // we have key2/key3 above the retain_lsn, so we still need this delta layer
10511 4 : PersistentLayerKey {
10512 4 : key_range: get_key(2)..get_key(4),
10513 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10514 4 : is_delta: true,
10515 4 : },
10516 4 : PersistentLayerKey {
10517 4 : key_range: get_key(5)..get_key(7),
10518 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10519 4 : is_delta: true,
10520 4 : },
10521 4 : PersistentLayerKey {
10522 4 : key_range: get_key(8)..get_key(10),
10523 4 : lsn_range: Lsn(0x48)..Lsn(0x50),
10524 4 : is_delta: true,
10525 4 : },
10526 4 : ],
10527 4 : );
10528 4 :
10529 4 : // Do a partial compaction on key range 4..9
10530 4 : tline
10531 4 : .compact_with_gc(
10532 4 : &cancel,
10533 4 : CompactOptions {
10534 4 : flags: EnumSet::new(),
10535 4 : compact_key_range: Some((get_key(4)..get_key(9)).into()),
10536 4 : ..Default::default()
10537 4 : },
10538 4 : &ctx,
10539 4 : )
10540 4 : .await
10541 4 : .unwrap();
10542 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10543 4 : check_layer_map_key_eq(
10544 4 : all_layers,
10545 4 : vec![
10546 4 : PersistentLayerKey {
10547 4 : key_range: get_key(0)..get_key(2),
10548 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10549 4 : is_delta: false,
10550 4 : },
10551 4 : PersistentLayerKey {
10552 4 : key_range: get_key(0)..get_key(10),
10553 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
10554 4 : is_delta: false,
10555 4 : },
10556 4 : PersistentLayerKey {
10557 4 : key_range: get_key(2)..get_key(4),
10558 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10559 4 : is_delta: false,
10560 4 : },
10561 4 : PersistentLayerKey {
10562 4 : key_range: get_key(2)..get_key(4),
10563 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10564 4 : is_delta: true,
10565 4 : },
10566 4 : // image layer generated for this compaction range
10567 4 : PersistentLayerKey {
10568 4 : key_range: get_key(4)..get_key(9),
10569 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10570 4 : is_delta: false,
10571 4 : },
10572 4 : PersistentLayerKey {
10573 4 : key_range: get_key(8)..get_key(10),
10574 4 : lsn_range: Lsn(0x48)..Lsn(0x50),
10575 4 : is_delta: true,
10576 4 : },
10577 4 : ],
10578 4 : );
10579 4 :
10580 4 : // Do a partial compaction on key range 9..10
10581 4 : tline
10582 4 : .compact_with_gc(
10583 4 : &cancel,
10584 4 : CompactOptions {
10585 4 : flags: EnumSet::new(),
10586 4 : compact_key_range: Some((get_key(9)..get_key(10)).into()),
10587 4 : ..Default::default()
10588 4 : },
10589 4 : &ctx,
10590 4 : )
10591 4 : .await
10592 4 : .unwrap();
10593 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10594 4 : check_layer_map_key_eq(
10595 4 : all_layers,
10596 4 : vec![
10597 4 : PersistentLayerKey {
10598 4 : key_range: get_key(0)..get_key(2),
10599 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10600 4 : is_delta: false,
10601 4 : },
10602 4 : PersistentLayerKey {
10603 4 : key_range: get_key(0)..get_key(10),
10604 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
10605 4 : is_delta: false,
10606 4 : },
10607 4 : PersistentLayerKey {
10608 4 : key_range: get_key(2)..get_key(4),
10609 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10610 4 : is_delta: false,
10611 4 : },
10612 4 : PersistentLayerKey {
10613 4 : key_range: get_key(2)..get_key(4),
10614 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10615 4 : is_delta: true,
10616 4 : },
10617 4 : PersistentLayerKey {
10618 4 : key_range: get_key(4)..get_key(9),
10619 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10620 4 : is_delta: false,
10621 4 : },
10622 4 : // image layer generated for the compaction range
10623 4 : PersistentLayerKey {
10624 4 : key_range: get_key(9)..get_key(10),
10625 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10626 4 : is_delta: false,
10627 4 : },
10628 4 : PersistentLayerKey {
10629 4 : key_range: get_key(8)..get_key(10),
10630 4 : lsn_range: Lsn(0x48)..Lsn(0x50),
10631 4 : is_delta: true,
10632 4 : },
10633 4 : ],
10634 4 : );
10635 4 :
10636 4 : // Do a partial compaction on key range 0..10, all image layers below LSN 20 can be replaced with new ones.
10637 4 : tline
10638 4 : .compact_with_gc(
10639 4 : &cancel,
10640 4 : CompactOptions {
10641 4 : flags: EnumSet::new(),
10642 4 : compact_key_range: Some((get_key(0)..get_key(10)).into()),
10643 4 : ..Default::default()
10644 4 : },
10645 4 : &ctx,
10646 4 : )
10647 4 : .await
10648 4 : .unwrap();
10649 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10650 4 : check_layer_map_key_eq(
10651 4 : all_layers,
10652 4 : vec![
10653 4 : // aha, we removed all unnecessary image/delta layers and got a very clean layer map!
10654 4 : PersistentLayerKey {
10655 4 : key_range: get_key(0)..get_key(10),
10656 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10657 4 : is_delta: false,
10658 4 : },
10659 4 : PersistentLayerKey {
10660 4 : key_range: get_key(2)..get_key(4),
10661 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10662 4 : is_delta: true,
10663 4 : },
10664 4 : PersistentLayerKey {
10665 4 : key_range: get_key(8)..get_key(10),
10666 4 : lsn_range: Lsn(0x48)..Lsn(0x50),
10667 4 : is_delta: true,
10668 4 : },
10669 4 : ],
10670 4 : );
10671 4 : Ok(())
10672 4 : }
10673 :
10674 : #[cfg(feature = "testing")]
10675 : #[tokio::test]
10676 4 : async fn test_timeline_offload_retain_lsn() -> anyhow::Result<()> {
10677 4 : let harness = TenantHarness::create("test_timeline_offload_retain_lsn")
10678 4 : .await
10679 4 : .unwrap();
10680 4 : let (tenant, ctx) = harness.load().await;
10681 4 : let tline_parent = tenant
10682 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
10683 4 : .await
10684 4 : .unwrap();
10685 4 : let tline_child = tenant
10686 4 : .branch_timeline_test(&tline_parent, NEW_TIMELINE_ID, Some(Lsn(0x20)), &ctx)
10687 4 : .await
10688 4 : .unwrap();
10689 4 : {
10690 4 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
10691 4 : assert_eq!(
10692 4 : gc_info_parent.retain_lsns,
10693 4 : vec![(Lsn(0x20), tline_child.timeline_id, MaybeOffloaded::No)]
10694 4 : );
10695 4 : }
10696 4 : // We have to directly call the remote_client instead of using the archive function to avoid constructing broker client...
10697 4 : tline_child
10698 4 : .remote_client
10699 4 : .schedule_index_upload_for_timeline_archival_state(TimelineArchivalState::Archived)
10700 4 : .unwrap();
10701 4 : tline_child.remote_client.wait_completion().await.unwrap();
10702 4 : offload_timeline(&tenant, &tline_child)
10703 4 : .instrument(tracing::info_span!(parent: None, "offload_test", tenant_id=%"test", shard_id=%"test", timeline_id=%"test"))
10704 4 : .await.unwrap();
10705 4 : let child_timeline_id = tline_child.timeline_id;
10706 4 : Arc::try_unwrap(tline_child).unwrap();
10707 4 :
10708 4 : {
10709 4 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
10710 4 : assert_eq!(
10711 4 : gc_info_parent.retain_lsns,
10712 4 : vec![(Lsn(0x20), child_timeline_id, MaybeOffloaded::Yes)]
10713 4 : );
10714 4 : }
10715 4 :
10716 4 : tenant
10717 4 : .get_offloaded_timeline(child_timeline_id)
10718 4 : .unwrap()
10719 4 : .defuse_for_tenant_drop();
10720 4 :
10721 4 : Ok(())
10722 4 : }
10723 :
10724 : #[cfg(feature = "testing")]
10725 : #[tokio::test]
10726 4 : async fn test_simple_bottom_most_compaction_above_lsn() -> anyhow::Result<()> {
10727 4 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_above_lsn").await?;
10728 4 : let (tenant, ctx) = harness.load().await;
10729 4 :
10730 592 : fn get_key(id: u32) -> Key {
10731 592 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
10732 592 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
10733 592 : key.field6 = id;
10734 592 : key
10735 592 : }
10736 4 :
10737 4 : let img_layer = (0..10)
10738 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10739 4 : .collect_vec();
10740 4 :
10741 4 : let delta1 = vec![(
10742 4 : get_key(1),
10743 4 : Lsn(0x20),
10744 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10745 4 : )];
10746 4 : let delta4 = vec![(
10747 4 : get_key(1),
10748 4 : Lsn(0x28),
10749 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
10750 4 : )];
10751 4 : let delta2 = vec![
10752 4 : (
10753 4 : get_key(1),
10754 4 : Lsn(0x30),
10755 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10756 4 : ),
10757 4 : (
10758 4 : get_key(1),
10759 4 : Lsn(0x38),
10760 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
10761 4 : ),
10762 4 : ];
10763 4 : let delta3 = vec![
10764 4 : (
10765 4 : get_key(8),
10766 4 : Lsn(0x48),
10767 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10768 4 : ),
10769 4 : (
10770 4 : get_key(9),
10771 4 : Lsn(0x48),
10772 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10773 4 : ),
10774 4 : ];
10775 4 :
10776 4 : let tline = tenant
10777 4 : .create_test_timeline_with_layers(
10778 4 : TIMELINE_ID,
10779 4 : Lsn(0x10),
10780 4 : DEFAULT_PG_VERSION,
10781 4 : &ctx,
10782 4 : vec![
10783 4 : // delta1/2/4 only contain a single key but multiple updates
10784 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x28), delta1),
10785 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
10786 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x28)..Lsn(0x30), delta4),
10787 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta3),
10788 4 : ], // delta layers
10789 4 : vec![(Lsn(0x10), img_layer)], // image layers
10790 4 : Lsn(0x50),
10791 4 : )
10792 4 : .await?;
10793 4 : {
10794 4 : tline
10795 4 : .applied_gc_cutoff_lsn
10796 4 : .lock_for_write()
10797 4 : .store_and_unlock(Lsn(0x30))
10798 4 : .wait()
10799 4 : .await;
10800 4 : // Update GC info
10801 4 : let mut guard = tline.gc_info.write().unwrap();
10802 4 : *guard = GcInfo {
10803 4 : retain_lsns: vec![
10804 4 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
10805 4 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
10806 4 : ],
10807 4 : cutoffs: GcCutoffs {
10808 4 : time: Lsn(0x30),
10809 4 : space: Lsn(0x30),
10810 4 : },
10811 4 : leases: Default::default(),
10812 4 : within_ancestor_pitr: false,
10813 4 : };
10814 4 : }
10815 4 :
10816 4 : let expected_result = [
10817 4 : Bytes::from_static(b"value 0@0x10"),
10818 4 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
10819 4 : Bytes::from_static(b"value 2@0x10"),
10820 4 : Bytes::from_static(b"value 3@0x10"),
10821 4 : Bytes::from_static(b"value 4@0x10"),
10822 4 : Bytes::from_static(b"value 5@0x10"),
10823 4 : Bytes::from_static(b"value 6@0x10"),
10824 4 : Bytes::from_static(b"value 7@0x10"),
10825 4 : Bytes::from_static(b"value 8@0x10@0x48"),
10826 4 : Bytes::from_static(b"value 9@0x10@0x48"),
10827 4 : ];
10828 4 :
10829 4 : let expected_result_at_gc_horizon = [
10830 4 : Bytes::from_static(b"value 0@0x10"),
10831 4 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
10832 4 : Bytes::from_static(b"value 2@0x10"),
10833 4 : Bytes::from_static(b"value 3@0x10"),
10834 4 : Bytes::from_static(b"value 4@0x10"),
10835 4 : Bytes::from_static(b"value 5@0x10"),
10836 4 : Bytes::from_static(b"value 6@0x10"),
10837 4 : Bytes::from_static(b"value 7@0x10"),
10838 4 : Bytes::from_static(b"value 8@0x10"),
10839 4 : Bytes::from_static(b"value 9@0x10"),
10840 4 : ];
10841 4 :
10842 4 : let expected_result_at_lsn_20 = [
10843 4 : Bytes::from_static(b"value 0@0x10"),
10844 4 : Bytes::from_static(b"value 1@0x10@0x20"),
10845 4 : Bytes::from_static(b"value 2@0x10"),
10846 4 : Bytes::from_static(b"value 3@0x10"),
10847 4 : Bytes::from_static(b"value 4@0x10"),
10848 4 : Bytes::from_static(b"value 5@0x10"),
10849 4 : Bytes::from_static(b"value 6@0x10"),
10850 4 : Bytes::from_static(b"value 7@0x10"),
10851 4 : Bytes::from_static(b"value 8@0x10"),
10852 4 : Bytes::from_static(b"value 9@0x10"),
10853 4 : ];
10854 4 :
10855 4 : let expected_result_at_lsn_10 = [
10856 4 : Bytes::from_static(b"value 0@0x10"),
10857 4 : Bytes::from_static(b"value 1@0x10"),
10858 4 : Bytes::from_static(b"value 2@0x10"),
10859 4 : Bytes::from_static(b"value 3@0x10"),
10860 4 : Bytes::from_static(b"value 4@0x10"),
10861 4 : Bytes::from_static(b"value 5@0x10"),
10862 4 : Bytes::from_static(b"value 6@0x10"),
10863 4 : Bytes::from_static(b"value 7@0x10"),
10864 4 : Bytes::from_static(b"value 8@0x10"),
10865 4 : Bytes::from_static(b"value 9@0x10"),
10866 4 : ];
10867 4 :
10868 12 : let verify_result = || async {
10869 12 : let gc_horizon = {
10870 12 : let gc_info = tline.gc_info.read().unwrap();
10871 12 : gc_info.cutoffs.time
10872 4 : };
10873 132 : for idx in 0..10 {
10874 120 : assert_eq!(
10875 120 : tline
10876 120 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10877 120 : .await
10878 120 : .unwrap(),
10879 120 : &expected_result[idx]
10880 4 : );
10881 120 : assert_eq!(
10882 120 : tline
10883 120 : .get(get_key(idx as u32), gc_horizon, &ctx)
10884 120 : .await
10885 120 : .unwrap(),
10886 120 : &expected_result_at_gc_horizon[idx]
10887 4 : );
10888 120 : assert_eq!(
10889 120 : tline
10890 120 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
10891 120 : .await
10892 120 : .unwrap(),
10893 120 : &expected_result_at_lsn_20[idx]
10894 4 : );
10895 120 : assert_eq!(
10896 120 : tline
10897 120 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
10898 120 : .await
10899 120 : .unwrap(),
10900 120 : &expected_result_at_lsn_10[idx]
10901 4 : );
10902 4 : }
10903 24 : };
10904 4 :
10905 4 : verify_result().await;
10906 4 :
10907 4 : let cancel = CancellationToken::new();
10908 4 : tline
10909 4 : .compact_with_gc(
10910 4 : &cancel,
10911 4 : CompactOptions {
10912 4 : compact_lsn_range: Some(CompactLsnRange::above(Lsn(0x28))),
10913 4 : ..Default::default()
10914 4 : },
10915 4 : &ctx,
10916 4 : )
10917 4 : .await
10918 4 : .unwrap();
10919 4 : verify_result().await;
10920 4 :
10921 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10922 4 : check_layer_map_key_eq(
10923 4 : all_layers,
10924 4 : vec![
10925 4 : // The original image layer, not compacted
10926 4 : PersistentLayerKey {
10927 4 : key_range: get_key(0)..get_key(10),
10928 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
10929 4 : is_delta: false,
10930 4 : },
10931 4 : // Delta layer below the specified above_lsn not compacted
10932 4 : PersistentLayerKey {
10933 4 : key_range: get_key(1)..get_key(2),
10934 4 : lsn_range: Lsn(0x20)..Lsn(0x28),
10935 4 : is_delta: true,
10936 4 : },
10937 4 : // Delta layer compacted above the LSN
10938 4 : PersistentLayerKey {
10939 4 : key_range: get_key(1)..get_key(10),
10940 4 : lsn_range: Lsn(0x28)..Lsn(0x50),
10941 4 : is_delta: true,
10942 4 : },
10943 4 : ],
10944 4 : );
10945 4 :
10946 4 : // compact again
10947 4 : tline
10948 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10949 4 : .await
10950 4 : .unwrap();
10951 4 : verify_result().await;
10952 4 :
10953 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10954 4 : check_layer_map_key_eq(
10955 4 : all_layers,
10956 4 : vec![
10957 4 : // The compacted image layer (full key range)
10958 4 : PersistentLayerKey {
10959 4 : key_range: Key::MIN..Key::MAX,
10960 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
10961 4 : is_delta: false,
10962 4 : },
10963 4 : // All other data in the delta layer
10964 4 : PersistentLayerKey {
10965 4 : key_range: get_key(1)..get_key(10),
10966 4 : lsn_range: Lsn(0x10)..Lsn(0x50),
10967 4 : is_delta: true,
10968 4 : },
10969 4 : ],
10970 4 : );
10971 4 :
10972 4 : Ok(())
10973 4 : }
10974 :
10975 : #[cfg(feature = "testing")]
10976 : #[tokio::test]
10977 4 : async fn test_simple_bottom_most_compaction_rectangle() -> anyhow::Result<()> {
10978 4 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_rectangle").await?;
10979 4 : let (tenant, ctx) = harness.load().await;
10980 4 :
10981 1016 : fn get_key(id: u32) -> Key {
10982 1016 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
10983 1016 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
10984 1016 : key.field6 = id;
10985 1016 : key
10986 1016 : }
10987 4 :
10988 4 : let img_layer = (0..10)
10989 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10990 4 : .collect_vec();
10991 4 :
10992 4 : let delta1 = vec![(
10993 4 : get_key(1),
10994 4 : Lsn(0x20),
10995 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10996 4 : )];
10997 4 : let delta4 = vec![(
10998 4 : get_key(1),
10999 4 : Lsn(0x28),
11000 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
11001 4 : )];
11002 4 : let delta2 = vec![
11003 4 : (
11004 4 : get_key(1),
11005 4 : Lsn(0x30),
11006 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
11007 4 : ),
11008 4 : (
11009 4 : get_key(1),
11010 4 : Lsn(0x38),
11011 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
11012 4 : ),
11013 4 : ];
11014 4 : let delta3 = vec![
11015 4 : (
11016 4 : get_key(8),
11017 4 : Lsn(0x48),
11018 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
11019 4 : ),
11020 4 : (
11021 4 : get_key(9),
11022 4 : Lsn(0x48),
11023 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
11024 4 : ),
11025 4 : ];
11026 4 :
11027 4 : let tline = tenant
11028 4 : .create_test_timeline_with_layers(
11029 4 : TIMELINE_ID,
11030 4 : Lsn(0x10),
11031 4 : DEFAULT_PG_VERSION,
11032 4 : &ctx,
11033 4 : vec![
11034 4 : // delta1/2/4 only contain a single key but multiple updates
11035 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x28), delta1),
11036 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
11037 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x28)..Lsn(0x30), delta4),
11038 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta3),
11039 4 : ], // delta layers
11040 4 : vec![(Lsn(0x10), img_layer)], // image layers
11041 4 : Lsn(0x50),
11042 4 : )
11043 4 : .await?;
11044 4 : {
11045 4 : tline
11046 4 : .applied_gc_cutoff_lsn
11047 4 : .lock_for_write()
11048 4 : .store_and_unlock(Lsn(0x30))
11049 4 : .wait()
11050 4 : .await;
11051 4 : // Update GC info
11052 4 : let mut guard = tline.gc_info.write().unwrap();
11053 4 : *guard = GcInfo {
11054 4 : retain_lsns: vec![
11055 4 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
11056 4 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
11057 4 : ],
11058 4 : cutoffs: GcCutoffs {
11059 4 : time: Lsn(0x30),
11060 4 : space: Lsn(0x30),
11061 4 : },
11062 4 : leases: Default::default(),
11063 4 : within_ancestor_pitr: false,
11064 4 : };
11065 4 : }
11066 4 :
11067 4 : let expected_result = [
11068 4 : Bytes::from_static(b"value 0@0x10"),
11069 4 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
11070 4 : Bytes::from_static(b"value 2@0x10"),
11071 4 : Bytes::from_static(b"value 3@0x10"),
11072 4 : Bytes::from_static(b"value 4@0x10"),
11073 4 : Bytes::from_static(b"value 5@0x10"),
11074 4 : Bytes::from_static(b"value 6@0x10"),
11075 4 : Bytes::from_static(b"value 7@0x10"),
11076 4 : Bytes::from_static(b"value 8@0x10@0x48"),
11077 4 : Bytes::from_static(b"value 9@0x10@0x48"),
11078 4 : ];
11079 4 :
11080 4 : let expected_result_at_gc_horizon = [
11081 4 : Bytes::from_static(b"value 0@0x10"),
11082 4 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
11083 4 : Bytes::from_static(b"value 2@0x10"),
11084 4 : Bytes::from_static(b"value 3@0x10"),
11085 4 : Bytes::from_static(b"value 4@0x10"),
11086 4 : Bytes::from_static(b"value 5@0x10"),
11087 4 : Bytes::from_static(b"value 6@0x10"),
11088 4 : Bytes::from_static(b"value 7@0x10"),
11089 4 : Bytes::from_static(b"value 8@0x10"),
11090 4 : Bytes::from_static(b"value 9@0x10"),
11091 4 : ];
11092 4 :
11093 4 : let expected_result_at_lsn_20 = [
11094 4 : Bytes::from_static(b"value 0@0x10"),
11095 4 : Bytes::from_static(b"value 1@0x10@0x20"),
11096 4 : Bytes::from_static(b"value 2@0x10"),
11097 4 : Bytes::from_static(b"value 3@0x10"),
11098 4 : Bytes::from_static(b"value 4@0x10"),
11099 4 : Bytes::from_static(b"value 5@0x10"),
11100 4 : Bytes::from_static(b"value 6@0x10"),
11101 4 : Bytes::from_static(b"value 7@0x10"),
11102 4 : Bytes::from_static(b"value 8@0x10"),
11103 4 : Bytes::from_static(b"value 9@0x10"),
11104 4 : ];
11105 4 :
11106 4 : let expected_result_at_lsn_10 = [
11107 4 : Bytes::from_static(b"value 0@0x10"),
11108 4 : Bytes::from_static(b"value 1@0x10"),
11109 4 : Bytes::from_static(b"value 2@0x10"),
11110 4 : Bytes::from_static(b"value 3@0x10"),
11111 4 : Bytes::from_static(b"value 4@0x10"),
11112 4 : Bytes::from_static(b"value 5@0x10"),
11113 4 : Bytes::from_static(b"value 6@0x10"),
11114 4 : Bytes::from_static(b"value 7@0x10"),
11115 4 : Bytes::from_static(b"value 8@0x10"),
11116 4 : Bytes::from_static(b"value 9@0x10"),
11117 4 : ];
11118 4 :
11119 20 : let verify_result = || async {
11120 20 : let gc_horizon = {
11121 20 : let gc_info = tline.gc_info.read().unwrap();
11122 20 : gc_info.cutoffs.time
11123 4 : };
11124 220 : for idx in 0..10 {
11125 200 : assert_eq!(
11126 200 : tline
11127 200 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
11128 200 : .await
11129 200 : .unwrap(),
11130 200 : &expected_result[idx]
11131 4 : );
11132 200 : assert_eq!(
11133 200 : tline
11134 200 : .get(get_key(idx as u32), gc_horizon, &ctx)
11135 200 : .await
11136 200 : .unwrap(),
11137 200 : &expected_result_at_gc_horizon[idx]
11138 4 : );
11139 200 : assert_eq!(
11140 200 : tline
11141 200 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
11142 200 : .await
11143 200 : .unwrap(),
11144 200 : &expected_result_at_lsn_20[idx]
11145 4 : );
11146 200 : assert_eq!(
11147 200 : tline
11148 200 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
11149 200 : .await
11150 200 : .unwrap(),
11151 200 : &expected_result_at_lsn_10[idx]
11152 4 : );
11153 4 : }
11154 40 : };
11155 4 :
11156 4 : verify_result().await;
11157 4 :
11158 4 : let cancel = CancellationToken::new();
11159 4 :
11160 4 : tline
11161 4 : .compact_with_gc(
11162 4 : &cancel,
11163 4 : CompactOptions {
11164 4 : compact_key_range: Some((get_key(0)..get_key(2)).into()),
11165 4 : compact_lsn_range: Some((Lsn(0x20)..Lsn(0x28)).into()),
11166 4 : ..Default::default()
11167 4 : },
11168 4 : &ctx,
11169 4 : )
11170 4 : .await
11171 4 : .unwrap();
11172 4 : verify_result().await;
11173 4 :
11174 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11175 4 : check_layer_map_key_eq(
11176 4 : all_layers,
11177 4 : vec![
11178 4 : // The original image layer, not compacted
11179 4 : PersistentLayerKey {
11180 4 : key_range: get_key(0)..get_key(10),
11181 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
11182 4 : is_delta: false,
11183 4 : },
11184 4 : // According the selection logic, we select all layers with start key <= 0x28, so we would merge the layer 0x20-0x28 and
11185 4 : // the layer 0x28-0x30 into one.
11186 4 : PersistentLayerKey {
11187 4 : key_range: get_key(1)..get_key(2),
11188 4 : lsn_range: Lsn(0x20)..Lsn(0x30),
11189 4 : is_delta: true,
11190 4 : },
11191 4 : // Above the upper bound and untouched
11192 4 : PersistentLayerKey {
11193 4 : key_range: get_key(1)..get_key(2),
11194 4 : lsn_range: Lsn(0x30)..Lsn(0x50),
11195 4 : is_delta: true,
11196 4 : },
11197 4 : // This layer is untouched
11198 4 : PersistentLayerKey {
11199 4 : key_range: get_key(8)..get_key(10),
11200 4 : lsn_range: Lsn(0x30)..Lsn(0x50),
11201 4 : is_delta: true,
11202 4 : },
11203 4 : ],
11204 4 : );
11205 4 :
11206 4 : tline
11207 4 : .compact_with_gc(
11208 4 : &cancel,
11209 4 : CompactOptions {
11210 4 : compact_key_range: Some((get_key(3)..get_key(8)).into()),
11211 4 : compact_lsn_range: Some((Lsn(0x28)..Lsn(0x40)).into()),
11212 4 : ..Default::default()
11213 4 : },
11214 4 : &ctx,
11215 4 : )
11216 4 : .await
11217 4 : .unwrap();
11218 4 : verify_result().await;
11219 4 :
11220 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11221 4 : check_layer_map_key_eq(
11222 4 : all_layers,
11223 4 : vec![
11224 4 : // The original image layer, not compacted
11225 4 : PersistentLayerKey {
11226 4 : key_range: get_key(0)..get_key(10),
11227 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
11228 4 : is_delta: false,
11229 4 : },
11230 4 : // Not in the compaction key range, uncompacted
11231 4 : PersistentLayerKey {
11232 4 : key_range: get_key(1)..get_key(2),
11233 4 : lsn_range: Lsn(0x20)..Lsn(0x30),
11234 4 : is_delta: true,
11235 4 : },
11236 4 : // Not in the compaction key range, uncompacted but need rewrite because the delta layer overlaps with the range
11237 4 : PersistentLayerKey {
11238 4 : key_range: get_key(1)..get_key(2),
11239 4 : lsn_range: Lsn(0x30)..Lsn(0x50),
11240 4 : is_delta: true,
11241 4 : },
11242 4 : // Note that when we specify the LSN upper bound to be 0x40, the compaction algorithm will not try to cut the layer
11243 4 : // horizontally in half. Instead, it will include all LSNs that overlap with 0x40. So the real max_lsn of the compaction
11244 4 : // becomes 0x50.
11245 4 : PersistentLayerKey {
11246 4 : key_range: get_key(8)..get_key(10),
11247 4 : lsn_range: Lsn(0x30)..Lsn(0x50),
11248 4 : is_delta: true,
11249 4 : },
11250 4 : ],
11251 4 : );
11252 4 :
11253 4 : // compact again
11254 4 : tline
11255 4 : .compact_with_gc(
11256 4 : &cancel,
11257 4 : CompactOptions {
11258 4 : compact_key_range: Some((get_key(0)..get_key(5)).into()),
11259 4 : compact_lsn_range: Some((Lsn(0x20)..Lsn(0x50)).into()),
11260 4 : ..Default::default()
11261 4 : },
11262 4 : &ctx,
11263 4 : )
11264 4 : .await
11265 4 : .unwrap();
11266 4 : verify_result().await;
11267 4 :
11268 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11269 4 : check_layer_map_key_eq(
11270 4 : all_layers,
11271 4 : vec![
11272 4 : // The original image layer, not compacted
11273 4 : PersistentLayerKey {
11274 4 : key_range: get_key(0)..get_key(10),
11275 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
11276 4 : is_delta: false,
11277 4 : },
11278 4 : // The range gets compacted
11279 4 : PersistentLayerKey {
11280 4 : key_range: get_key(1)..get_key(2),
11281 4 : lsn_range: Lsn(0x20)..Lsn(0x50),
11282 4 : is_delta: true,
11283 4 : },
11284 4 : // Not touched during this iteration of compaction
11285 4 : PersistentLayerKey {
11286 4 : key_range: get_key(8)..get_key(10),
11287 4 : lsn_range: Lsn(0x30)..Lsn(0x50),
11288 4 : is_delta: true,
11289 4 : },
11290 4 : ],
11291 4 : );
11292 4 :
11293 4 : // final full compaction
11294 4 : tline
11295 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
11296 4 : .await
11297 4 : .unwrap();
11298 4 : verify_result().await;
11299 4 :
11300 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11301 4 : check_layer_map_key_eq(
11302 4 : all_layers,
11303 4 : vec![
11304 4 : // The compacted image layer (full key range)
11305 4 : PersistentLayerKey {
11306 4 : key_range: Key::MIN..Key::MAX,
11307 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
11308 4 : is_delta: false,
11309 4 : },
11310 4 : // All other data in the delta layer
11311 4 : PersistentLayerKey {
11312 4 : key_range: get_key(1)..get_key(10),
11313 4 : lsn_range: Lsn(0x10)..Lsn(0x50),
11314 4 : is_delta: true,
11315 4 : },
11316 4 : ],
11317 4 : );
11318 4 :
11319 4 : Ok(())
11320 4 : }
11321 : }
|
