Line data Source code
1 : //! Timeline repository implementation that keeps old data in layer files, and
2 : //! the recent changes in ephemeral files.
3 : //!
4 : //! See tenant/*_layer.rs files. The functions here are responsible for locating
5 : //! the correct layer for the get/put call, walking back the timeline branching
6 : //! history as needed.
7 : //!
8 : //! The files are stored in the .neon/tenants/<tenant_id>/timelines/<timeline_id>
9 : //! directory. See docs/pageserver-storage.md for how the files are managed.
10 : //! In addition to the layer files, there is a metadata file in the same
11 : //! directory that contains information about the timeline, in particular its
12 : //! parent timeline, and the last LSN that has been written to disk.
13 : //!
14 :
15 : use anyhow::{bail, Context};
16 : use arc_swap::ArcSwap;
17 : use camino::Utf8Path;
18 : use camino::Utf8PathBuf;
19 : use chrono::NaiveDateTime;
20 : use enumset::EnumSet;
21 : use futures::stream::FuturesUnordered;
22 : use futures::StreamExt;
23 : use pageserver_api::models;
24 : use pageserver_api::models::CompactInfoResponse;
25 : use pageserver_api::models::LsnLease;
26 : use pageserver_api::models::TimelineArchivalState;
27 : use pageserver_api::models::TimelineState;
28 : use pageserver_api::models::TopTenantShardItem;
29 : use pageserver_api::models::WalRedoManagerStatus;
30 : use pageserver_api::shard::ShardIdentity;
31 : use pageserver_api::shard::ShardStripeSize;
32 : use pageserver_api::shard::TenantShardId;
33 : use remote_storage::DownloadError;
34 : use remote_storage::GenericRemoteStorage;
35 : use remote_storage::TimeoutOrCancel;
36 : use remote_timeline_client::manifest::{
37 : OffloadedTimelineManifest, TenantManifest, LATEST_TENANT_MANIFEST_VERSION,
38 : };
39 : use remote_timeline_client::UploadQueueNotReadyError;
40 : use std::collections::BTreeMap;
41 : use std::fmt;
42 : use std::future::Future;
43 : use std::sync::atomic::AtomicBool;
44 : use std::sync::Weak;
45 : use std::time::SystemTime;
46 : use storage_broker::BrokerClientChannel;
47 : use timeline::compaction::GcCompactionQueue;
48 : use timeline::import_pgdata;
49 : use timeline::offload::offload_timeline;
50 : use timeline::offload::OffloadError;
51 : use timeline::CompactOptions;
52 : use timeline::ShutdownMode;
53 : use tokio::io::BufReader;
54 : use tokio::sync::watch;
55 : use tokio::task::JoinSet;
56 : use tokio_util::sync::CancellationToken;
57 : use tracing::*;
58 : use upload_queue::NotInitialized;
59 : use utils::backoff;
60 : use utils::circuit_breaker::CircuitBreaker;
61 : use utils::completion;
62 : use utils::crashsafe::path_with_suffix_extension;
63 : use utils::failpoint_support;
64 : use utils::fs_ext;
65 : use utils::pausable_failpoint;
66 : use utils::sync::gate::Gate;
67 : use utils::sync::gate::GateGuard;
68 : use utils::timeout::timeout_cancellable;
69 : use utils::timeout::TimeoutCancellableError;
70 : use utils::try_rcu::ArcSwapExt;
71 : use utils::zstd::create_zst_tarball;
72 : use utils::zstd::extract_zst_tarball;
73 :
74 : use self::config::AttachedLocationConfig;
75 : use self::config::AttachmentMode;
76 : use self::config::LocationConf;
77 : use self::config::TenantConf;
78 : use self::metadata::TimelineMetadata;
79 : use self::mgr::GetActiveTenantError;
80 : use self::mgr::GetTenantError;
81 : use self::remote_timeline_client::upload::{upload_index_part, upload_tenant_manifest};
82 : use self::remote_timeline_client::{RemoteTimelineClient, WaitCompletionError};
83 : use self::timeline::uninit::TimelineCreateGuard;
84 : use self::timeline::uninit::TimelineExclusionError;
85 : use self::timeline::uninit::UninitializedTimeline;
86 : use self::timeline::EvictionTaskTenantState;
87 : use self::timeline::GcCutoffs;
88 : use self::timeline::TimelineDeleteProgress;
89 : use self::timeline::TimelineResources;
90 : use self::timeline::WaitLsnError;
91 : use crate::config::PageServerConf;
92 : use crate::context::{DownloadBehavior, RequestContext};
93 : use crate::deletion_queue::DeletionQueueClient;
94 : use crate::deletion_queue::DeletionQueueError;
95 : use crate::import_datadir;
96 : use crate::is_uninit_mark;
97 : use crate::l0_flush::L0FlushGlobalState;
98 : use crate::metrics::TENANT;
99 : use crate::metrics::{
100 : remove_tenant_metrics, BROKEN_TENANTS_SET, CIRCUIT_BREAKERS_BROKEN, CIRCUIT_BREAKERS_UNBROKEN,
101 : TENANT_STATE_METRIC, TENANT_SYNTHETIC_SIZE_METRIC,
102 : };
103 : use crate::task_mgr;
104 : use crate::task_mgr::TaskKind;
105 : use crate::tenant::config::LocationMode;
106 : use crate::tenant::config::TenantConfOpt;
107 : use crate::tenant::gc_result::GcResult;
108 : pub use crate::tenant::remote_timeline_client::index::IndexPart;
109 : use crate::tenant::remote_timeline_client::remote_initdb_archive_path;
110 : use crate::tenant::remote_timeline_client::MaybeDeletedIndexPart;
111 : use crate::tenant::remote_timeline_client::INITDB_PATH;
112 : use crate::tenant::storage_layer::DeltaLayer;
113 : use crate::tenant::storage_layer::ImageLayer;
114 : use crate::walingest::WalLagCooldown;
115 : use crate::walredo;
116 : use crate::InitializationOrder;
117 : use std::collections::hash_map::Entry;
118 : use std::collections::HashMap;
119 : use std::collections::HashSet;
120 : use std::fmt::Debug;
121 : use std::fmt::Display;
122 : use std::fs;
123 : use std::fs::File;
124 : use std::sync::atomic::{AtomicU64, Ordering};
125 : use std::sync::Arc;
126 : use std::sync::Mutex;
127 : use std::time::{Duration, Instant};
128 :
129 : use crate::span;
130 : use crate::tenant::timeline::delete::DeleteTimelineFlow;
131 : use crate::tenant::timeline::uninit::cleanup_timeline_directory;
132 : use crate::virtual_file::VirtualFile;
133 : use crate::walredo::PostgresRedoManager;
134 : use crate::TEMP_FILE_SUFFIX;
135 : use once_cell::sync::Lazy;
136 : pub use pageserver_api::models::TenantState;
137 : use tokio::sync::Semaphore;
138 :
139 0 : static INIT_DB_SEMAPHORE: Lazy<Semaphore> = Lazy::new(|| Semaphore::new(8));
140 : use utils::{
141 : crashsafe,
142 : generation::Generation,
143 : id::TimelineId,
144 : lsn::{Lsn, RecordLsn},
145 : };
146 :
147 : pub mod blob_io;
148 : pub mod block_io;
149 : pub mod vectored_blob_io;
150 :
151 : pub mod disk_btree;
152 : pub(crate) mod ephemeral_file;
153 : pub mod layer_map;
154 :
155 : pub mod metadata;
156 : pub mod remote_timeline_client;
157 : pub mod storage_layer;
158 :
159 : pub mod checks;
160 : pub mod config;
161 : pub mod mgr;
162 : pub mod secondary;
163 : pub mod tasks;
164 : pub mod upload_queue;
165 :
166 : pub(crate) mod timeline;
167 :
168 : pub mod size;
169 :
170 : mod gc_block;
171 : mod gc_result;
172 : pub(crate) mod throttle;
173 :
174 : pub(crate) use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
175 : pub(crate) use timeline::{LogicalSizeCalculationCause, PageReconstructError, Timeline};
176 :
177 : // re-export for use in walreceiver
178 : pub use crate::tenant::timeline::WalReceiverInfo;
179 :
180 : /// The "tenants" part of `tenants/<tenant>/timelines...`
181 : pub const TENANTS_SEGMENT_NAME: &str = "tenants";
182 :
183 : /// Parts of the `.neon/tenants/<tenant_id>/timelines/<timeline_id>` directory prefix.
184 : pub const TIMELINES_SEGMENT_NAME: &str = "timelines";
185 :
186 : /// References to shared objects that are passed into each tenant, such
187 : /// as the shared remote storage client and process initialization state.
188 : #[derive(Clone)]
189 : pub struct TenantSharedResources {
190 : pub broker_client: storage_broker::BrokerClientChannel,
191 : pub remote_storage: GenericRemoteStorage,
192 : pub deletion_queue_client: DeletionQueueClient,
193 : pub l0_flush_global_state: L0FlushGlobalState,
194 : }
195 :
196 : /// A [`Tenant`] is really an _attached_ tenant. The configuration
197 : /// for an attached tenant is a subset of the [`LocationConf`], represented
198 : /// in this struct.
199 : #[derive(Clone)]
200 : pub(super) struct AttachedTenantConf {
201 : tenant_conf: TenantConfOpt,
202 : location: AttachedLocationConfig,
203 : /// The deadline before which we are blocked from GC so that
204 : /// leases have a chance to be renewed.
205 : lsn_lease_deadline: Option<tokio::time::Instant>,
206 : }
207 :
208 : impl AttachedTenantConf {
209 220 : fn new(tenant_conf: TenantConfOpt, location: AttachedLocationConfig) -> Self {
210 : // Sets a deadline before which we cannot proceed to GC due to lsn lease.
211 : //
212 : // We do this as the leases mapping are not persisted to disk. By delaying GC by lease
213 : // length, we guarantee that all the leases we granted before will have a chance to renew
214 : // when we run GC for the first time after restart / transition from AttachedMulti to AttachedSingle.
215 220 : let lsn_lease_deadline = if location.attach_mode == AttachmentMode::Single {
216 220 : Some(
217 220 : tokio::time::Instant::now()
218 220 : + tenant_conf
219 220 : .lsn_lease_length
220 220 : .unwrap_or(LsnLease::DEFAULT_LENGTH),
221 220 : )
222 : } else {
223 : // We don't use `lsn_lease_deadline` to delay GC in AttachedMulti and AttachedStale
224 : // because we don't do GC in these modes.
225 0 : None
226 : };
227 :
228 220 : Self {
229 220 : tenant_conf,
230 220 : location,
231 220 : lsn_lease_deadline,
232 220 : }
233 220 : }
234 :
235 220 : fn try_from(location_conf: LocationConf) -> anyhow::Result<Self> {
236 220 : match &location_conf.mode {
237 220 : LocationMode::Attached(attach_conf) => {
238 220 : Ok(Self::new(location_conf.tenant_conf, *attach_conf))
239 : }
240 : LocationMode::Secondary(_) => {
241 0 : anyhow::bail!("Attempted to construct AttachedTenantConf from a LocationConf in secondary mode")
242 : }
243 : }
244 220 : }
245 :
246 762 : fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
247 762 : self.lsn_lease_deadline
248 762 : .map(|d| tokio::time::Instant::now() < d)
249 762 : .unwrap_or(false)
250 762 : }
251 : }
252 : struct TimelinePreload {
253 : timeline_id: TimelineId,
254 : client: RemoteTimelineClient,
255 : index_part: Result<MaybeDeletedIndexPart, DownloadError>,
256 : }
257 :
258 : pub(crate) struct TenantPreload {
259 : tenant_manifest: TenantManifest,
260 : /// Map from timeline ID to a possible timeline preload. It is None iff the timeline is offloaded according to the manifest.
261 : timelines: HashMap<TimelineId, Option<TimelinePreload>>,
262 : }
263 :
264 : /// When we spawn a tenant, there is a special mode for tenant creation that
265 : /// avoids trying to read anything from remote storage.
266 : pub(crate) enum SpawnMode {
267 : /// Activate as soon as possible
268 : Eager,
269 : /// Lazy activation in the background, with the option to skip the queue if the need comes up
270 : Lazy,
271 : }
272 :
273 : ///
274 : /// Tenant consists of multiple timelines. Keep them in a hash table.
275 : ///
276 : pub struct Tenant {
277 : // Global pageserver config parameters
278 : pub conf: &'static PageServerConf,
279 :
280 : /// The value creation timestamp, used to measure activation delay, see:
281 : /// <https://github.com/neondatabase/neon/issues/4025>
282 : constructed_at: Instant,
283 :
284 : state: watch::Sender<TenantState>,
285 :
286 : // Overridden tenant-specific config parameters.
287 : // We keep TenantConfOpt sturct here to preserve the information
288 : // about parameters that are not set.
289 : // This is necessary to allow global config updates.
290 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
291 :
292 : tenant_shard_id: TenantShardId,
293 :
294 : // The detailed sharding information, beyond the number/count in tenant_shard_id
295 : shard_identity: ShardIdentity,
296 :
297 : /// The remote storage generation, used to protect S3 objects from split-brain.
298 : /// Does not change over the lifetime of the [`Tenant`] object.
299 : ///
300 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
301 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
302 : generation: Generation,
303 :
304 : timelines: Mutex<HashMap<TimelineId, Arc<Timeline>>>,
305 :
306 : /// During timeline creation, we first insert the TimelineId to the
307 : /// creating map, then `timelines`, then remove it from the creating map.
308 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
309 : timelines_creating: std::sync::Mutex<HashSet<TimelineId>>,
310 :
311 : /// Possibly offloaded and archived timelines
312 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
313 : timelines_offloaded: Mutex<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
314 :
315 : /// Serialize writes of the tenant manifest to remote storage. If there are concurrent operations
316 : /// affecting the manifest, such as timeline deletion and timeline offload, they must wait for
317 : /// each other (this could be optimized to coalesce writes if necessary).
318 : ///
319 : /// The contents of the Mutex are the last manifest we successfully uploaded
320 : tenant_manifest_upload: tokio::sync::Mutex<Option<TenantManifest>>,
321 :
322 : // This mutex prevents creation of new timelines during GC.
323 : // Adding yet another mutex (in addition to `timelines`) is needed because holding
324 : // `timelines` mutex during all GC iteration
325 : // may block for a long time `get_timeline`, `get_timelines_state`,... and other operations
326 : // with timelines, which in turn may cause dropping replication connection, expiration of wait_for_lsn
327 : // timeout...
328 : gc_cs: tokio::sync::Mutex<()>,
329 : walredo_mgr: Option<Arc<WalRedoManager>>,
330 :
331 : // provides access to timeline data sitting in the remote storage
332 : pub(crate) remote_storage: GenericRemoteStorage,
333 :
334 : // Access to global deletion queue for when this tenant wants to schedule a deletion
335 : deletion_queue_client: DeletionQueueClient,
336 :
337 : /// Cached logical sizes updated updated on each [`Tenant::gather_size_inputs`].
338 : cached_logical_sizes: tokio::sync::Mutex<HashMap<(TimelineId, Lsn), u64>>,
339 : cached_synthetic_tenant_size: Arc<AtomicU64>,
340 :
341 : eviction_task_tenant_state: tokio::sync::Mutex<EvictionTaskTenantState>,
342 :
343 : /// Track repeated failures to compact, so that we can back off.
344 : /// Overhead of mutex is acceptable because compaction is done with a multi-second period.
345 : compaction_circuit_breaker: std::sync::Mutex<CircuitBreaker>,
346 :
347 : /// Scheduled gc-compaction tasks.
348 : scheduled_compaction_tasks: std::sync::Mutex<HashMap<TimelineId, Arc<GcCompactionQueue>>>,
349 :
350 : /// If the tenant is in Activating state, notify this to encourage it
351 : /// to proceed to Active as soon as possible, rather than waiting for lazy
352 : /// background warmup.
353 : pub(crate) activate_now_sem: tokio::sync::Semaphore,
354 :
355 : /// Time it took for the tenant to activate. Zero if not active yet.
356 : attach_wal_lag_cooldown: Arc<std::sync::OnceLock<WalLagCooldown>>,
357 :
358 : // Cancellation token fires when we have entered shutdown(). This is a parent of
359 : // Timelines' cancellation token.
360 : pub(crate) cancel: CancellationToken,
361 :
362 : // Users of the Tenant such as the page service must take this Gate to avoid
363 : // trying to use a Tenant which is shutting down.
364 : pub(crate) gate: Gate,
365 :
366 : /// Throttle applied at the top of [`Timeline::get`].
367 : /// All [`Tenant::timelines`] of a given [`Tenant`] instance share the same [`throttle::Throttle`] instance.
368 : pub(crate) pagestream_throttle: Arc<throttle::Throttle>,
369 :
370 : pub(crate) pagestream_throttle_metrics: Arc<crate::metrics::tenant_throttling::Pagestream>,
371 :
372 : /// An ongoing timeline detach concurrency limiter.
373 : ///
374 : /// As a tenant will likely be restarted as part of timeline detach ancestor it makes no sense
375 : /// to have two running at the same time. A different one can be started if an earlier one
376 : /// has failed for whatever reason.
377 : ongoing_timeline_detach: std::sync::Mutex<Option<(TimelineId, utils::completion::Barrier)>>,
378 :
379 : /// `index_part.json` based gc blocking reason tracking.
380 : ///
381 : /// New gc iterations must start a new iteration by acquiring `GcBlock::start` before
382 : /// proceeding.
383 : pub(crate) gc_block: gc_block::GcBlock,
384 :
385 : l0_flush_global_state: L0FlushGlobalState,
386 : }
387 : impl std::fmt::Debug for Tenant {
388 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
389 0 : write!(f, "{} ({})", self.tenant_shard_id, self.current_state())
390 0 : }
391 : }
392 :
393 : pub(crate) enum WalRedoManager {
394 : Prod(WalredoManagerId, PostgresRedoManager),
395 : #[cfg(test)]
396 : Test(harness::TestRedoManager),
397 : }
398 :
399 : #[derive(thiserror::Error, Debug)]
400 : #[error("pageserver is shutting down")]
401 : pub(crate) struct GlobalShutDown;
402 :
403 : impl WalRedoManager {
404 0 : pub(crate) fn new(mgr: PostgresRedoManager) -> Result<Arc<Self>, GlobalShutDown> {
405 0 : let id = WalredoManagerId::next();
406 0 : let arc = Arc::new(Self::Prod(id, mgr));
407 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
408 0 : match &mut *guard {
409 0 : Some(map) => {
410 0 : map.insert(id, Arc::downgrade(&arc));
411 0 : Ok(arc)
412 : }
413 0 : None => Err(GlobalShutDown),
414 : }
415 0 : }
416 : }
417 :
418 : impl Drop for WalRedoManager {
419 10 : fn drop(&mut self) {
420 10 : match self {
421 0 : Self::Prod(id, _) => {
422 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
423 0 : if let Some(map) = &mut *guard {
424 0 : map.remove(id).expect("new() registers, drop() unregisters");
425 0 : }
426 : }
427 : #[cfg(test)]
428 10 : Self::Test(_) => {
429 10 : // Not applicable to test redo manager
430 10 : }
431 : }
432 10 : }
433 : }
434 :
435 : /// Global registry of all walredo managers so that [`crate::shutdown_pageserver`] can shut down
436 : /// the walredo processes outside of the regular order.
437 : ///
438 : /// This is necessary to work around a systemd bug where it freezes if there are
439 : /// walredo processes left => <https://github.com/neondatabase/cloud/issues/11387>
440 : #[allow(clippy::type_complexity)]
441 : pub(crate) static WALREDO_MANAGERS: once_cell::sync::Lazy<
442 : Mutex<Option<HashMap<WalredoManagerId, Weak<WalRedoManager>>>>,
443 0 : > = once_cell::sync::Lazy::new(|| Mutex::new(Some(HashMap::new())));
444 : #[derive(PartialEq, Eq, Hash, Clone, Copy, Debug)]
445 : pub(crate) struct WalredoManagerId(u64);
446 : impl WalredoManagerId {
447 0 : pub fn next() -> Self {
448 : static NEXT: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(1);
449 0 : let id = NEXT.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
450 0 : if id == 0 {
451 0 : panic!("WalredoManagerId::new() returned 0, indicating wraparound, risking it's no longer unique");
452 0 : }
453 0 : Self(id)
454 0 : }
455 : }
456 :
457 : #[cfg(test)]
458 : impl From<harness::TestRedoManager> for WalRedoManager {
459 220 : fn from(mgr: harness::TestRedoManager) -> Self {
460 220 : Self::Test(mgr)
461 220 : }
462 : }
463 :
464 : impl WalRedoManager {
465 6 : pub(crate) async fn shutdown(&self) -> bool {
466 6 : match self {
467 0 : Self::Prod(_, mgr) => mgr.shutdown().await,
468 : #[cfg(test)]
469 : Self::Test(_) => {
470 : // Not applicable to test redo manager
471 6 : true
472 : }
473 : }
474 6 : }
475 :
476 0 : pub(crate) fn maybe_quiesce(&self, idle_timeout: Duration) {
477 0 : match self {
478 0 : Self::Prod(_, mgr) => mgr.maybe_quiesce(idle_timeout),
479 0 : #[cfg(test)]
480 0 : Self::Test(_) => {
481 0 : // Not applicable to test redo manager
482 0 : }
483 0 : }
484 0 : }
485 :
486 : /// # Cancel-Safety
487 : ///
488 : /// This method is cancellation-safe.
489 818 : pub async fn request_redo(
490 818 : &self,
491 818 : key: pageserver_api::key::Key,
492 818 : lsn: Lsn,
493 818 : base_img: Option<(Lsn, bytes::Bytes)>,
494 818 : records: Vec<(Lsn, pageserver_api::record::NeonWalRecord)>,
495 818 : pg_version: u32,
496 818 : ) -> Result<bytes::Bytes, walredo::Error> {
497 818 : match self {
498 0 : Self::Prod(_, mgr) => {
499 0 : mgr.request_redo(key, lsn, base_img, records, pg_version)
500 0 : .await
501 : }
502 : #[cfg(test)]
503 818 : Self::Test(mgr) => {
504 818 : mgr.request_redo(key, lsn, base_img, records, pg_version)
505 818 : .await
506 : }
507 : }
508 818 : }
509 :
510 0 : pub(crate) fn status(&self) -> Option<WalRedoManagerStatus> {
511 0 : match self {
512 0 : WalRedoManager::Prod(_, m) => Some(m.status()),
513 0 : #[cfg(test)]
514 0 : WalRedoManager::Test(_) => None,
515 0 : }
516 0 : }
517 : }
518 :
519 : /// A very lightweight memory representation of an offloaded timeline.
520 : ///
521 : /// We need to store the list of offloaded timelines so that we can perform operations on them,
522 : /// like unoffloading them, or (at a later date), decide to perform flattening.
523 : /// This type has a much smaller memory impact than [`Timeline`], and thus we can store many
524 : /// more offloaded timelines than we can manage ones that aren't.
525 : pub struct OffloadedTimeline {
526 : pub tenant_shard_id: TenantShardId,
527 : pub timeline_id: TimelineId,
528 : pub ancestor_timeline_id: Option<TimelineId>,
529 : /// Whether to retain the branch lsn at the ancestor or not
530 : pub ancestor_retain_lsn: Option<Lsn>,
531 :
532 : /// When the timeline was archived.
533 : ///
534 : /// Present for future flattening deliberations.
535 : pub archived_at: NaiveDateTime,
536 :
537 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
538 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
539 : pub delete_progress: TimelineDeleteProgress,
540 :
541 : /// Part of the `OffloadedTimeline` object's lifecycle: this needs to be set before we drop it
542 : pub deleted_from_ancestor: AtomicBool,
543 : }
544 :
545 : impl OffloadedTimeline {
546 : /// Obtains an offloaded timeline from a given timeline object.
547 : ///
548 : /// Returns `None` if the `archived_at` flag couldn't be obtained, i.e.
549 : /// the timeline is not in a stopped state.
550 : /// Panics if the timeline is not archived.
551 2 : fn from_timeline(timeline: &Timeline) -> Result<Self, UploadQueueNotReadyError> {
552 2 : let (ancestor_retain_lsn, ancestor_timeline_id) =
553 2 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
554 2 : let ancestor_lsn = timeline.get_ancestor_lsn();
555 2 : let ancestor_timeline_id = ancestor_timeline.timeline_id;
556 2 : let mut gc_info = ancestor_timeline.gc_info.write().unwrap();
557 2 : gc_info.insert_child(timeline.timeline_id, ancestor_lsn, MaybeOffloaded::Yes);
558 2 : (Some(ancestor_lsn), Some(ancestor_timeline_id))
559 : } else {
560 0 : (None, None)
561 : };
562 2 : let archived_at = timeline
563 2 : .remote_client
564 2 : .archived_at_stopped_queue()?
565 2 : .expect("must be called on an archived timeline");
566 2 : Ok(Self {
567 2 : tenant_shard_id: timeline.tenant_shard_id,
568 2 : timeline_id: timeline.timeline_id,
569 2 : ancestor_timeline_id,
570 2 : ancestor_retain_lsn,
571 2 : archived_at,
572 2 :
573 2 : delete_progress: timeline.delete_progress.clone(),
574 2 : deleted_from_ancestor: AtomicBool::new(false),
575 2 : })
576 2 : }
577 0 : fn from_manifest(tenant_shard_id: TenantShardId, manifest: &OffloadedTimelineManifest) -> Self {
578 0 : // We expect to reach this case in tenant loading, where the `retain_lsn` is populated in the parent's `gc_info`
579 0 : // by the `initialize_gc_info` function.
580 0 : let OffloadedTimelineManifest {
581 0 : timeline_id,
582 0 : ancestor_timeline_id,
583 0 : ancestor_retain_lsn,
584 0 : archived_at,
585 0 : } = *manifest;
586 0 : Self {
587 0 : tenant_shard_id,
588 0 : timeline_id,
589 0 : ancestor_timeline_id,
590 0 : ancestor_retain_lsn,
591 0 : archived_at,
592 0 : delete_progress: TimelineDeleteProgress::default(),
593 0 : deleted_from_ancestor: AtomicBool::new(false),
594 0 : }
595 0 : }
596 2 : fn manifest(&self) -> OffloadedTimelineManifest {
597 2 : let Self {
598 2 : timeline_id,
599 2 : ancestor_timeline_id,
600 2 : ancestor_retain_lsn,
601 2 : archived_at,
602 2 : ..
603 2 : } = self;
604 2 : OffloadedTimelineManifest {
605 2 : timeline_id: *timeline_id,
606 2 : ancestor_timeline_id: *ancestor_timeline_id,
607 2 : ancestor_retain_lsn: *ancestor_retain_lsn,
608 2 : archived_at: *archived_at,
609 2 : }
610 2 : }
611 : /// Delete this timeline's retain_lsn from its ancestor, if present in the given tenant
612 0 : fn delete_from_ancestor_with_timelines(
613 0 : &self,
614 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
615 0 : ) {
616 0 : if let (Some(_retain_lsn), Some(ancestor_timeline_id)) =
617 0 : (self.ancestor_retain_lsn, self.ancestor_timeline_id)
618 : {
619 0 : if let Some((_, ancestor_timeline)) = timelines
620 0 : .iter()
621 0 : .find(|(tid, _tl)| **tid == ancestor_timeline_id)
622 : {
623 0 : let removal_happened = ancestor_timeline
624 0 : .gc_info
625 0 : .write()
626 0 : .unwrap()
627 0 : .remove_child_offloaded(self.timeline_id);
628 0 : if !removal_happened {
629 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
630 0 : "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
631 0 : }
632 0 : }
633 0 : }
634 0 : self.deleted_from_ancestor.store(true, Ordering::Release);
635 0 : }
636 : /// Call [`Self::delete_from_ancestor_with_timelines`] instead if possible.
637 : ///
638 : /// As the entire tenant is being dropped, don't bother deregistering the `retain_lsn` from the ancestor.
639 2 : fn defuse_for_tenant_drop(&self) {
640 2 : self.deleted_from_ancestor.store(true, Ordering::Release);
641 2 : }
642 : }
643 :
644 : impl fmt::Debug for OffloadedTimeline {
645 0 : fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
646 0 : write!(f, "OffloadedTimeline<{}>", self.timeline_id)
647 0 : }
648 : }
649 :
650 : impl Drop for OffloadedTimeline {
651 2 : fn drop(&mut self) {
652 2 : if !self.deleted_from_ancestor.load(Ordering::Acquire) {
653 0 : tracing::warn!(
654 0 : "offloaded timeline {} was dropped without having cleaned it up at the ancestor",
655 : self.timeline_id
656 : );
657 2 : }
658 2 : }
659 : }
660 :
661 : #[derive(Copy, Clone, PartialEq, Eq, Hash, Debug)]
662 : pub enum MaybeOffloaded {
663 : Yes,
664 : No,
665 : }
666 :
667 : #[derive(Clone, Debug)]
668 : pub enum TimelineOrOffloaded {
669 : Timeline(Arc<Timeline>),
670 : Offloaded(Arc<OffloadedTimeline>),
671 : }
672 :
673 : impl TimelineOrOffloaded {
674 0 : pub fn arc_ref(&self) -> TimelineOrOffloadedArcRef<'_> {
675 0 : match self {
676 0 : TimelineOrOffloaded::Timeline(timeline) => {
677 0 : TimelineOrOffloadedArcRef::Timeline(timeline)
678 : }
679 0 : TimelineOrOffloaded::Offloaded(offloaded) => {
680 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded)
681 : }
682 : }
683 0 : }
684 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
685 0 : self.arc_ref().tenant_shard_id()
686 0 : }
687 0 : pub fn timeline_id(&self) -> TimelineId {
688 0 : self.arc_ref().timeline_id()
689 0 : }
690 2 : pub fn delete_progress(&self) -> &Arc<tokio::sync::Mutex<DeleteTimelineFlow>> {
691 2 : match self {
692 2 : TimelineOrOffloaded::Timeline(timeline) => &timeline.delete_progress,
693 0 : TimelineOrOffloaded::Offloaded(offloaded) => &offloaded.delete_progress,
694 : }
695 2 : }
696 0 : fn maybe_remote_client(&self) -> Option<Arc<RemoteTimelineClient>> {
697 0 : match self {
698 0 : TimelineOrOffloaded::Timeline(timeline) => Some(timeline.remote_client.clone()),
699 0 : TimelineOrOffloaded::Offloaded(_offloaded) => None,
700 : }
701 0 : }
702 : }
703 :
704 : pub enum TimelineOrOffloadedArcRef<'a> {
705 : Timeline(&'a Arc<Timeline>),
706 : Offloaded(&'a Arc<OffloadedTimeline>),
707 : }
708 :
709 : impl TimelineOrOffloadedArcRef<'_> {
710 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
711 0 : match self {
712 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.tenant_shard_id,
713 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.tenant_shard_id,
714 : }
715 0 : }
716 0 : pub fn timeline_id(&self) -> TimelineId {
717 0 : match self {
718 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.timeline_id,
719 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.timeline_id,
720 : }
721 0 : }
722 : }
723 :
724 : impl<'a> From<&'a Arc<Timeline>> for TimelineOrOffloadedArcRef<'a> {
725 0 : fn from(timeline: &'a Arc<Timeline>) -> Self {
726 0 : Self::Timeline(timeline)
727 0 : }
728 : }
729 :
730 : impl<'a> From<&'a Arc<OffloadedTimeline>> for TimelineOrOffloadedArcRef<'a> {
731 0 : fn from(timeline: &'a Arc<OffloadedTimeline>) -> Self {
732 0 : Self::Offloaded(timeline)
733 0 : }
734 : }
735 :
736 : #[derive(Debug, thiserror::Error, PartialEq, Eq)]
737 : pub enum GetTimelineError {
738 : #[error("Timeline is shutting down")]
739 : ShuttingDown,
740 : #[error("Timeline {tenant_id}/{timeline_id} is not active, state: {state:?}")]
741 : NotActive {
742 : tenant_id: TenantShardId,
743 : timeline_id: TimelineId,
744 : state: TimelineState,
745 : },
746 : #[error("Timeline {tenant_id}/{timeline_id} was not found")]
747 : NotFound {
748 : tenant_id: TenantShardId,
749 : timeline_id: TimelineId,
750 : },
751 : }
752 :
753 : #[derive(Debug, thiserror::Error)]
754 : pub enum LoadLocalTimelineError {
755 : #[error("FailedToLoad")]
756 : Load(#[source] anyhow::Error),
757 : #[error("FailedToResumeDeletion")]
758 : ResumeDeletion(#[source] anyhow::Error),
759 : }
760 :
761 : #[derive(thiserror::Error)]
762 : pub enum DeleteTimelineError {
763 : #[error("NotFound")]
764 : NotFound,
765 :
766 : #[error("HasChildren")]
767 : HasChildren(Vec<TimelineId>),
768 :
769 : #[error("Timeline deletion is already in progress")]
770 : AlreadyInProgress(Arc<tokio::sync::Mutex<DeleteTimelineFlow>>),
771 :
772 : #[error("Cancelled")]
773 : Cancelled,
774 :
775 : #[error(transparent)]
776 : Other(#[from] anyhow::Error),
777 : }
778 :
779 : impl Debug for DeleteTimelineError {
780 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
781 0 : match self {
782 0 : Self::NotFound => write!(f, "NotFound"),
783 0 : Self::HasChildren(c) => f.debug_tuple("HasChildren").field(c).finish(),
784 0 : Self::AlreadyInProgress(_) => f.debug_tuple("AlreadyInProgress").finish(),
785 0 : Self::Cancelled => f.debug_tuple("Cancelled").finish(),
786 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
787 : }
788 0 : }
789 : }
790 :
791 : #[derive(thiserror::Error)]
792 : pub enum TimelineArchivalError {
793 : #[error("NotFound")]
794 : NotFound,
795 :
796 : #[error("Timeout")]
797 : Timeout,
798 :
799 : #[error("Cancelled")]
800 : Cancelled,
801 :
802 : #[error("ancestor is archived: {}", .0)]
803 : HasArchivedParent(TimelineId),
804 :
805 : #[error("HasUnarchivedChildren")]
806 : HasUnarchivedChildren(Vec<TimelineId>),
807 :
808 : #[error("Timeline archival is already in progress")]
809 : AlreadyInProgress,
810 :
811 : #[error(transparent)]
812 : Other(anyhow::Error),
813 : }
814 :
815 : #[derive(thiserror::Error, Debug)]
816 : pub(crate) enum TenantManifestError {
817 : #[error("Remote storage error: {0}")]
818 : RemoteStorage(anyhow::Error),
819 :
820 : #[error("Cancelled")]
821 : Cancelled,
822 : }
823 :
824 : impl From<TenantManifestError> for TimelineArchivalError {
825 0 : fn from(e: TenantManifestError) -> Self {
826 0 : match e {
827 0 : TenantManifestError::RemoteStorage(e) => Self::Other(e),
828 0 : TenantManifestError::Cancelled => Self::Cancelled,
829 : }
830 0 : }
831 : }
832 :
833 : impl Debug for TimelineArchivalError {
834 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
835 0 : match self {
836 0 : Self::NotFound => write!(f, "NotFound"),
837 0 : Self::Timeout => write!(f, "Timeout"),
838 0 : Self::Cancelled => write!(f, "Cancelled"),
839 0 : Self::HasArchivedParent(p) => f.debug_tuple("HasArchivedParent").field(p).finish(),
840 0 : Self::HasUnarchivedChildren(c) => {
841 0 : f.debug_tuple("HasUnarchivedChildren").field(c).finish()
842 : }
843 0 : Self::AlreadyInProgress => f.debug_tuple("AlreadyInProgress").finish(),
844 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
845 : }
846 0 : }
847 : }
848 :
849 : pub enum SetStoppingError {
850 : AlreadyStopping(completion::Barrier),
851 : Broken,
852 : }
853 :
854 : impl Debug for SetStoppingError {
855 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
856 0 : match self {
857 0 : Self::AlreadyStopping(_) => f.debug_tuple("AlreadyStopping").finish(),
858 0 : Self::Broken => write!(f, "Broken"),
859 : }
860 0 : }
861 : }
862 :
863 : /// Arguments to [`Tenant::create_timeline`].
864 : ///
865 : /// Not usable as an idempotency key for timeline creation because if [`CreateTimelineParamsBranch::ancestor_start_lsn`]
866 : /// is `None`, the result of the timeline create call is not deterministic.
867 : ///
868 : /// See [`CreateTimelineIdempotency`] for an idempotency key.
869 : #[derive(Debug)]
870 : pub(crate) enum CreateTimelineParams {
871 : Bootstrap(CreateTimelineParamsBootstrap),
872 : Branch(CreateTimelineParamsBranch),
873 : ImportPgdata(CreateTimelineParamsImportPgdata),
874 : }
875 :
876 : #[derive(Debug)]
877 : pub(crate) struct CreateTimelineParamsBootstrap {
878 : pub(crate) new_timeline_id: TimelineId,
879 : pub(crate) existing_initdb_timeline_id: Option<TimelineId>,
880 : pub(crate) pg_version: u32,
881 : }
882 :
883 : /// NB: See comment on [`CreateTimelineIdempotency::Branch`] for why there's no `pg_version` here.
884 : #[derive(Debug)]
885 : pub(crate) struct CreateTimelineParamsBranch {
886 : pub(crate) new_timeline_id: TimelineId,
887 : pub(crate) ancestor_timeline_id: TimelineId,
888 : pub(crate) ancestor_start_lsn: Option<Lsn>,
889 : }
890 :
891 : #[derive(Debug)]
892 : pub(crate) struct CreateTimelineParamsImportPgdata {
893 : pub(crate) new_timeline_id: TimelineId,
894 : pub(crate) location: import_pgdata::index_part_format::Location,
895 : pub(crate) idempotency_key: import_pgdata::index_part_format::IdempotencyKey,
896 : }
897 :
898 : /// What is used to determine idempotency of a [`Tenant::create_timeline`] call in [`Tenant::start_creating_timeline`] in [`Tenant::start_creating_timeline`].
899 : ///
900 : /// Each [`Timeline`] object holds [`Self`] as an immutable property in [`Timeline::create_idempotency`].
901 : ///
902 : /// We lower timeline creation requests to [`Self`], and then use [`PartialEq::eq`] to compare [`Timeline::create_idempotency`] with the request.
903 : /// If they are equal, we return a reference to the existing timeline, otherwise it's an idempotency conflict.
904 : ///
905 : /// There is special treatment for [`Self::FailWithConflict`] to always return an idempotency conflict.
906 : /// It would be nice to have more advanced derive macros to make that special treatment declarative.
907 : ///
908 : /// Notes:
909 : /// - Unlike [`CreateTimelineParams`], ancestor LSN is fixed, so, branching will be at a deterministic LSN.
910 : /// - We make some trade-offs though, e.g., [`CreateTimelineParamsBootstrap::existing_initdb_timeline_id`]
911 : /// is not considered for idempotency. We can improve on this over time if we deem it necessary.
912 : ///
913 : #[derive(Debug, Clone, PartialEq, Eq)]
914 : pub(crate) enum CreateTimelineIdempotency {
915 : /// NB: special treatment, see comment in [`Self`].
916 : FailWithConflict,
917 : Bootstrap {
918 : pg_version: u32,
919 : },
920 : /// NB: branches always have the same `pg_version` as their ancestor.
921 : /// While [`pageserver_api::models::TimelineCreateRequestMode::Branch::pg_version`]
922 : /// exists as a field, and is set by cplane, it has always been ignored by pageserver when
923 : /// determining the child branch pg_version.
924 : Branch {
925 : ancestor_timeline_id: TimelineId,
926 : ancestor_start_lsn: Lsn,
927 : },
928 : ImportPgdata(CreatingTimelineIdempotencyImportPgdata),
929 : }
930 :
931 : #[derive(Debug, Clone, PartialEq, Eq)]
932 : pub(crate) struct CreatingTimelineIdempotencyImportPgdata {
933 : idempotency_key: import_pgdata::index_part_format::IdempotencyKey,
934 : }
935 :
936 : /// What is returned by [`Tenant::start_creating_timeline`].
937 : #[must_use]
938 : enum StartCreatingTimelineResult {
939 : CreateGuard(TimelineCreateGuard),
940 : Idempotent(Arc<Timeline>),
941 : }
942 :
943 : enum TimelineInitAndSyncResult {
944 : ReadyToActivate(Arc<Timeline>),
945 : NeedsSpawnImportPgdata(TimelineInitAndSyncNeedsSpawnImportPgdata),
946 : }
947 :
948 : impl TimelineInitAndSyncResult {
949 0 : fn ready_to_activate(self) -> Option<Arc<Timeline>> {
950 0 : match self {
951 0 : Self::ReadyToActivate(timeline) => Some(timeline),
952 0 : _ => None,
953 : }
954 0 : }
955 : }
956 :
957 : #[must_use]
958 : struct TimelineInitAndSyncNeedsSpawnImportPgdata {
959 : timeline: Arc<Timeline>,
960 : import_pgdata: import_pgdata::index_part_format::Root,
961 : guard: TimelineCreateGuard,
962 : }
963 :
964 : /// What is returned by [`Tenant::create_timeline`].
965 : enum CreateTimelineResult {
966 : Created(Arc<Timeline>),
967 : Idempotent(Arc<Timeline>),
968 : /// IMPORTANT: This [`Arc<Timeline>`] object is not in [`Tenant::timelines`] when
969 : /// we return this result, nor will this concrete object ever be added there.
970 : /// Cf method comment on [`Tenant::create_timeline_import_pgdata`].
971 : ImportSpawned(Arc<Timeline>),
972 : }
973 :
974 : impl CreateTimelineResult {
975 0 : fn discriminant(&self) -> &'static str {
976 0 : match self {
977 0 : Self::Created(_) => "Created",
978 0 : Self::Idempotent(_) => "Idempotent",
979 0 : Self::ImportSpawned(_) => "ImportSpawned",
980 : }
981 0 : }
982 0 : fn timeline(&self) -> &Arc<Timeline> {
983 0 : match self {
984 0 : Self::Created(t) | Self::Idempotent(t) | Self::ImportSpawned(t) => t,
985 0 : }
986 0 : }
987 : /// Unit test timelines aren't activated, test has to do it if it needs to.
988 : #[cfg(test)]
989 230 : fn into_timeline_for_test(self) -> Arc<Timeline> {
990 230 : match self {
991 230 : Self::Created(t) | Self::Idempotent(t) | Self::ImportSpawned(t) => t,
992 230 : }
993 230 : }
994 : }
995 :
996 : #[derive(thiserror::Error, Debug)]
997 : pub enum CreateTimelineError {
998 : #[error("creation of timeline with the given ID is in progress")]
999 : AlreadyCreating,
1000 : #[error("timeline already exists with different parameters")]
1001 : Conflict,
1002 : #[error(transparent)]
1003 : AncestorLsn(anyhow::Error),
1004 : #[error("ancestor timeline is not active")]
1005 : AncestorNotActive,
1006 : #[error("ancestor timeline is archived")]
1007 : AncestorArchived,
1008 : #[error("tenant shutting down")]
1009 : ShuttingDown,
1010 : #[error(transparent)]
1011 : Other(#[from] anyhow::Error),
1012 : }
1013 :
1014 : #[derive(thiserror::Error, Debug)]
1015 : pub enum InitdbError {
1016 : #[error("Operation was cancelled")]
1017 : Cancelled,
1018 : #[error(transparent)]
1019 : Other(anyhow::Error),
1020 : #[error(transparent)]
1021 : Inner(postgres_initdb::Error),
1022 : }
1023 :
1024 : enum CreateTimelineCause {
1025 : Load,
1026 : Delete,
1027 : }
1028 :
1029 : enum LoadTimelineCause {
1030 : Attach,
1031 : Unoffload,
1032 : ImportPgdata {
1033 : create_guard: TimelineCreateGuard,
1034 : activate: ActivateTimelineArgs,
1035 : },
1036 : }
1037 :
1038 : #[derive(thiserror::Error, Debug)]
1039 : pub(crate) enum GcError {
1040 : // The tenant is shutting down
1041 : #[error("tenant shutting down")]
1042 : TenantCancelled,
1043 :
1044 : // The tenant is shutting down
1045 : #[error("timeline shutting down")]
1046 : TimelineCancelled,
1047 :
1048 : // The tenant is in a state inelegible to run GC
1049 : #[error("not active")]
1050 : NotActive,
1051 :
1052 : // A requested GC cutoff LSN was invalid, for example it tried to move backwards
1053 : #[error("not active")]
1054 : BadLsn { why: String },
1055 :
1056 : // A remote storage error while scheduling updates after compaction
1057 : #[error(transparent)]
1058 : Remote(anyhow::Error),
1059 :
1060 : // An error reading while calculating GC cutoffs
1061 : #[error(transparent)]
1062 : GcCutoffs(PageReconstructError),
1063 :
1064 : // If GC was invoked for a particular timeline, this error means it didn't exist
1065 : #[error("timeline not found")]
1066 : TimelineNotFound,
1067 : }
1068 :
1069 : impl From<PageReconstructError> for GcError {
1070 0 : fn from(value: PageReconstructError) -> Self {
1071 0 : match value {
1072 0 : PageReconstructError::Cancelled => Self::TimelineCancelled,
1073 0 : other => Self::GcCutoffs(other),
1074 : }
1075 0 : }
1076 : }
1077 :
1078 : impl From<NotInitialized> for GcError {
1079 0 : fn from(value: NotInitialized) -> Self {
1080 0 : match value {
1081 0 : NotInitialized::Uninitialized => GcError::Remote(value.into()),
1082 0 : NotInitialized::Stopped | NotInitialized::ShuttingDown => GcError::TimelineCancelled,
1083 : }
1084 0 : }
1085 : }
1086 :
1087 : impl From<timeline::layer_manager::Shutdown> for GcError {
1088 0 : fn from(_: timeline::layer_manager::Shutdown) -> Self {
1089 0 : GcError::TimelineCancelled
1090 0 : }
1091 : }
1092 :
1093 : #[derive(thiserror::Error, Debug)]
1094 : pub(crate) enum LoadConfigError {
1095 : #[error("TOML deserialization error: '{0}'")]
1096 : DeserializeToml(#[from] toml_edit::de::Error),
1097 :
1098 : #[error("Config not found at {0}")]
1099 : NotFound(Utf8PathBuf),
1100 : }
1101 :
1102 : impl Tenant {
1103 : /// Yet another helper for timeline initialization.
1104 : ///
1105 : /// - Initializes the Timeline struct and inserts it into the tenant's hash map
1106 : /// - Scans the local timeline directory for layer files and builds the layer map
1107 : /// - Downloads remote index file and adds remote files to the layer map
1108 : /// - Schedules remote upload tasks for any files that are present locally but missing from remote storage.
1109 : ///
1110 : /// If the operation fails, the timeline is left in the tenant's hash map in Broken state. On success,
1111 : /// it is marked as Active.
1112 : #[allow(clippy::too_many_arguments)]
1113 6 : async fn timeline_init_and_sync(
1114 6 : self: &Arc<Self>,
1115 6 : timeline_id: TimelineId,
1116 6 : resources: TimelineResources,
1117 6 : mut index_part: IndexPart,
1118 6 : metadata: TimelineMetadata,
1119 6 : ancestor: Option<Arc<Timeline>>,
1120 6 : cause: LoadTimelineCause,
1121 6 : ctx: &RequestContext,
1122 6 : ) -> anyhow::Result<TimelineInitAndSyncResult> {
1123 6 : let tenant_id = self.tenant_shard_id;
1124 6 :
1125 6 : let import_pgdata = index_part.import_pgdata.take();
1126 6 : let idempotency = match &import_pgdata {
1127 0 : Some(import_pgdata) => {
1128 0 : CreateTimelineIdempotency::ImportPgdata(CreatingTimelineIdempotencyImportPgdata {
1129 0 : idempotency_key: import_pgdata.idempotency_key().clone(),
1130 0 : })
1131 : }
1132 : None => {
1133 6 : if metadata.ancestor_timeline().is_none() {
1134 4 : CreateTimelineIdempotency::Bootstrap {
1135 4 : pg_version: metadata.pg_version(),
1136 4 : }
1137 : } else {
1138 2 : CreateTimelineIdempotency::Branch {
1139 2 : ancestor_timeline_id: metadata.ancestor_timeline().unwrap(),
1140 2 : ancestor_start_lsn: metadata.ancestor_lsn(),
1141 2 : }
1142 : }
1143 : }
1144 : };
1145 :
1146 6 : let timeline = self.create_timeline_struct(
1147 6 : timeline_id,
1148 6 : &metadata,
1149 6 : ancestor.clone(),
1150 6 : resources,
1151 6 : CreateTimelineCause::Load,
1152 6 : idempotency.clone(),
1153 6 : )?;
1154 6 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
1155 6 : anyhow::ensure!(
1156 6 : disk_consistent_lsn.is_valid(),
1157 0 : "Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn"
1158 : );
1159 6 : assert_eq!(
1160 6 : disk_consistent_lsn,
1161 6 : metadata.disk_consistent_lsn(),
1162 0 : "these are used interchangeably"
1163 : );
1164 :
1165 6 : timeline.remote_client.init_upload_queue(&index_part)?;
1166 :
1167 6 : timeline
1168 6 : .load_layer_map(disk_consistent_lsn, index_part)
1169 6 : .await
1170 6 : .with_context(|| {
1171 0 : format!("Failed to load layermap for timeline {tenant_id}/{timeline_id}")
1172 6 : })?;
1173 :
1174 0 : match import_pgdata {
1175 0 : Some(import_pgdata) if !import_pgdata.is_done() => {
1176 0 : match cause {
1177 0 : LoadTimelineCause::Attach | LoadTimelineCause::Unoffload => (),
1178 : LoadTimelineCause::ImportPgdata { .. } => {
1179 0 : unreachable!("ImportPgdata should not be reloading timeline import is done and persisted as such in s3")
1180 : }
1181 : }
1182 0 : let mut guard = self.timelines_creating.lock().unwrap();
1183 0 : if !guard.insert(timeline_id) {
1184 : // We should never try and load the same timeline twice during startup
1185 0 : unreachable!("Timeline {tenant_id}/{timeline_id} is already being created")
1186 0 : }
1187 0 : let timeline_create_guard = TimelineCreateGuard {
1188 0 : _tenant_gate_guard: self.gate.enter()?,
1189 0 : owning_tenant: self.clone(),
1190 0 : timeline_id,
1191 0 : idempotency,
1192 0 : // The users of this specific return value don't need the timline_path in there.
1193 0 : timeline_path: timeline
1194 0 : .conf
1195 0 : .timeline_path(&timeline.tenant_shard_id, &timeline.timeline_id),
1196 0 : };
1197 0 : Ok(TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
1198 0 : TimelineInitAndSyncNeedsSpawnImportPgdata {
1199 0 : timeline,
1200 0 : import_pgdata,
1201 0 : guard: timeline_create_guard,
1202 0 : },
1203 0 : ))
1204 : }
1205 : Some(_) | None => {
1206 : {
1207 6 : let mut timelines_accessor = self.timelines.lock().unwrap();
1208 6 : match timelines_accessor.entry(timeline_id) {
1209 : // We should never try and load the same timeline twice during startup
1210 : Entry::Occupied(_) => {
1211 0 : unreachable!(
1212 0 : "Timeline {tenant_id}/{timeline_id} already exists in the tenant map"
1213 0 : );
1214 : }
1215 6 : Entry::Vacant(v) => {
1216 6 : v.insert(Arc::clone(&timeline));
1217 6 : timeline.maybe_spawn_flush_loop();
1218 6 : }
1219 : }
1220 : }
1221 :
1222 : // Sanity check: a timeline should have some content.
1223 6 : anyhow::ensure!(
1224 6 : ancestor.is_some()
1225 4 : || timeline
1226 4 : .layers
1227 4 : .read()
1228 4 : .await
1229 4 : .layer_map()
1230 4 : .expect("currently loading, layer manager cannot be shutdown already")
1231 4 : .iter_historic_layers()
1232 4 : .next()
1233 4 : .is_some(),
1234 0 : "Timeline has no ancestor and no layer files"
1235 : );
1236 :
1237 6 : match cause {
1238 6 : LoadTimelineCause::Attach | LoadTimelineCause::Unoffload => (),
1239 : LoadTimelineCause::ImportPgdata {
1240 0 : create_guard,
1241 0 : activate,
1242 0 : } => {
1243 0 : // TODO: see the comment in the task code above how I'm not so certain
1244 0 : // it is safe to activate here because of concurrent shutdowns.
1245 0 : match activate {
1246 0 : ActivateTimelineArgs::Yes { broker_client } => {
1247 0 : info!("activating timeline after reload from pgdata import task");
1248 0 : timeline.activate(self.clone(), broker_client, None, ctx);
1249 : }
1250 0 : ActivateTimelineArgs::No => (),
1251 : }
1252 0 : drop(create_guard);
1253 : }
1254 : }
1255 :
1256 6 : Ok(TimelineInitAndSyncResult::ReadyToActivate(timeline))
1257 : }
1258 : }
1259 6 : }
1260 :
1261 : /// Attach a tenant that's available in cloud storage.
1262 : ///
1263 : /// This returns quickly, after just creating the in-memory object
1264 : /// Tenant struct and launching a background task to download
1265 : /// the remote index files. On return, the tenant is most likely still in
1266 : /// Attaching state, and it will become Active once the background task
1267 : /// finishes. You can use wait_until_active() to wait for the task to
1268 : /// complete.
1269 : ///
1270 : #[allow(clippy::too_many_arguments)]
1271 0 : pub(crate) fn spawn(
1272 0 : conf: &'static PageServerConf,
1273 0 : tenant_shard_id: TenantShardId,
1274 0 : resources: TenantSharedResources,
1275 0 : attached_conf: AttachedTenantConf,
1276 0 : shard_identity: ShardIdentity,
1277 0 : init_order: Option<InitializationOrder>,
1278 0 : mode: SpawnMode,
1279 0 : ctx: &RequestContext,
1280 0 : ) -> Result<Arc<Tenant>, GlobalShutDown> {
1281 0 : let wal_redo_manager =
1282 0 : WalRedoManager::new(PostgresRedoManager::new(conf, tenant_shard_id))?;
1283 :
1284 : let TenantSharedResources {
1285 0 : broker_client,
1286 0 : remote_storage,
1287 0 : deletion_queue_client,
1288 0 : l0_flush_global_state,
1289 0 : } = resources;
1290 0 :
1291 0 : let attach_mode = attached_conf.location.attach_mode;
1292 0 : let generation = attached_conf.location.generation;
1293 0 :
1294 0 : let tenant = Arc::new(Tenant::new(
1295 0 : TenantState::Attaching,
1296 0 : conf,
1297 0 : attached_conf,
1298 0 : shard_identity,
1299 0 : Some(wal_redo_manager),
1300 0 : tenant_shard_id,
1301 0 : remote_storage.clone(),
1302 0 : deletion_queue_client,
1303 0 : l0_flush_global_state,
1304 0 : ));
1305 0 :
1306 0 : // The attach task will carry a GateGuard, so that shutdown() reliably waits for it to drop out if
1307 0 : // we shut down while attaching.
1308 0 : let attach_gate_guard = tenant
1309 0 : .gate
1310 0 : .enter()
1311 0 : .expect("We just created the Tenant: nothing else can have shut it down yet");
1312 0 :
1313 0 : // Do all the hard work in the background
1314 0 : let tenant_clone = Arc::clone(&tenant);
1315 0 : let ctx = ctx.detached_child(TaskKind::Attach, DownloadBehavior::Warn);
1316 0 : task_mgr::spawn(
1317 0 : &tokio::runtime::Handle::current(),
1318 0 : TaskKind::Attach,
1319 0 : tenant_shard_id,
1320 0 : None,
1321 0 : "attach tenant",
1322 0 : async move {
1323 0 :
1324 0 : info!(
1325 : ?attach_mode,
1326 0 : "Attaching tenant"
1327 : );
1328 :
1329 0 : let _gate_guard = attach_gate_guard;
1330 0 :
1331 0 : // Is this tenant being spawned as part of process startup?
1332 0 : let starting_up = init_order.is_some();
1333 0 : scopeguard::defer! {
1334 0 : if starting_up {
1335 0 : TENANT.startup_complete.inc();
1336 0 : }
1337 0 : }
1338 :
1339 : // Ideally we should use Tenant::set_broken_no_wait, but it is not supposed to be used when tenant is in loading state.
1340 : enum BrokenVerbosity {
1341 : Error,
1342 : Info
1343 : }
1344 0 : let make_broken =
1345 0 : |t: &Tenant, err: anyhow::Error, verbosity: BrokenVerbosity| {
1346 0 : match verbosity {
1347 : BrokenVerbosity::Info => {
1348 0 : info!("attach cancelled, setting tenant state to Broken: {err}");
1349 : },
1350 : BrokenVerbosity::Error => {
1351 0 : error!("attach failed, setting tenant state to Broken: {err:?}");
1352 : }
1353 : }
1354 0 : t.state.send_modify(|state| {
1355 0 : // The Stopping case is for when we have passed control on to DeleteTenantFlow:
1356 0 : // if it errors, we will call make_broken when tenant is already in Stopping.
1357 0 : assert!(
1358 0 : matches!(*state, TenantState::Attaching | TenantState::Stopping { .. }),
1359 0 : "the attach task owns the tenant state until activation is complete"
1360 : );
1361 :
1362 0 : *state = TenantState::broken_from_reason(err.to_string());
1363 0 : });
1364 0 : };
1365 :
1366 : // TODO: should also be rejecting tenant conf changes that violate this check.
1367 0 : if let Err(e) = crate::tenant::storage_layer::inmemory_layer::IndexEntry::validate_checkpoint_distance(tenant_clone.get_checkpoint_distance()) {
1368 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1369 0 : return Ok(());
1370 0 : }
1371 0 :
1372 0 : let mut init_order = init_order;
1373 0 : // take the completion because initial tenant loading will complete when all of
1374 0 : // these tasks complete.
1375 0 : let _completion = init_order
1376 0 : .as_mut()
1377 0 : .and_then(|x| x.initial_tenant_load.take());
1378 0 : let remote_load_completion = init_order
1379 0 : .as_mut()
1380 0 : .and_then(|x| x.initial_tenant_load_remote.take());
1381 :
1382 : enum AttachType<'a> {
1383 : /// We are attaching this tenant lazily in the background.
1384 : Warmup {
1385 : _permit: tokio::sync::SemaphorePermit<'a>,
1386 : during_startup: bool
1387 : },
1388 : /// We are attaching this tenant as soon as we can, because for example an
1389 : /// endpoint tried to access it.
1390 : OnDemand,
1391 : /// During normal operations after startup, we are attaching a tenant, and
1392 : /// eager attach was requested.
1393 : Normal,
1394 : }
1395 :
1396 0 : let attach_type = if matches!(mode, SpawnMode::Lazy) {
1397 : // Before doing any I/O, wait for at least one of:
1398 : // - A client attempting to access to this tenant (on-demand loading)
1399 : // - A permit becoming available in the warmup semaphore (background warmup)
1400 :
1401 0 : tokio::select!(
1402 0 : permit = tenant_clone.activate_now_sem.acquire() => {
1403 0 : let _ = permit.expect("activate_now_sem is never closed");
1404 0 : tracing::info!("Activating tenant (on-demand)");
1405 0 : AttachType::OnDemand
1406 : },
1407 0 : permit = conf.concurrent_tenant_warmup.inner().acquire() => {
1408 0 : let _permit = permit.expect("concurrent_tenant_warmup semaphore is never closed");
1409 0 : tracing::info!("Activating tenant (warmup)");
1410 0 : AttachType::Warmup {
1411 0 : _permit,
1412 0 : during_startup: init_order.is_some()
1413 0 : }
1414 : }
1415 0 : _ = tenant_clone.cancel.cancelled() => {
1416 : // This is safe, but should be pretty rare: it is interesting if a tenant
1417 : // stayed in Activating for such a long time that shutdown found it in
1418 : // that state.
1419 0 : tracing::info!(state=%tenant_clone.current_state(), "Tenant shut down before activation");
1420 : // Make the tenant broken so that set_stopping will not hang waiting for it to leave
1421 : // the Attaching state. This is an over-reaction (nothing really broke, the tenant is
1422 : // just shutting down), but ensures progress.
1423 0 : make_broken(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"), BrokenVerbosity::Info);
1424 0 : return Ok(());
1425 : },
1426 : )
1427 : } else {
1428 : // SpawnMode::{Create,Eager} always cause jumping ahead of the
1429 : // concurrent_tenant_warmup queue
1430 0 : AttachType::Normal
1431 : };
1432 :
1433 0 : let preload = match &mode {
1434 : SpawnMode::Eager | SpawnMode::Lazy => {
1435 0 : let _preload_timer = TENANT.preload.start_timer();
1436 0 : let res = tenant_clone
1437 0 : .preload(&remote_storage, task_mgr::shutdown_token())
1438 0 : .await;
1439 0 : match res {
1440 0 : Ok(p) => Some(p),
1441 0 : Err(e) => {
1442 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1443 0 : return Ok(());
1444 : }
1445 : }
1446 : }
1447 :
1448 : };
1449 :
1450 : // Remote preload is complete.
1451 0 : drop(remote_load_completion);
1452 0 :
1453 0 :
1454 0 : // We will time the duration of the attach phase unless this is a creation (attach will do no work)
1455 0 : let attach_start = std::time::Instant::now();
1456 0 : let attached = {
1457 0 : let _attach_timer = Some(TENANT.attach.start_timer());
1458 0 : tenant_clone.attach(preload, &ctx).await
1459 : };
1460 0 : let attach_duration = attach_start.elapsed();
1461 0 : _ = tenant_clone.attach_wal_lag_cooldown.set(WalLagCooldown::new(attach_start, attach_duration));
1462 0 :
1463 0 : match attached {
1464 : Ok(()) => {
1465 0 : info!("attach finished, activating");
1466 0 : tenant_clone.activate(broker_client, None, &ctx);
1467 : }
1468 0 : Err(e) => {
1469 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1470 0 : }
1471 : }
1472 :
1473 : // If we are doing an opportunistic warmup attachment at startup, initialize
1474 : // logical size at the same time. This is better than starting a bunch of idle tenants
1475 : // with cold caches and then coming back later to initialize their logical sizes.
1476 : //
1477 : // It also prevents the warmup proccess competing with the concurrency limit on
1478 : // logical size calculations: if logical size calculation semaphore is saturated,
1479 : // then warmup will wait for that before proceeding to the next tenant.
1480 0 : if matches!(attach_type, AttachType::Warmup { during_startup: true, .. }) {
1481 0 : let mut futs: FuturesUnordered<_> = tenant_clone.timelines.lock().unwrap().values().cloned().map(|t| t.await_initial_logical_size()).collect();
1482 0 : tracing::info!("Waiting for initial logical sizes while warming up...");
1483 0 : while futs.next().await.is_some() {}
1484 0 : tracing::info!("Warm-up complete");
1485 0 : }
1486 :
1487 0 : Ok(())
1488 0 : }
1489 0 : .instrument(tracing::info_span!(parent: None, "attach", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), gen=?generation)),
1490 : );
1491 0 : Ok(tenant)
1492 0 : }
1493 :
1494 : #[instrument(skip_all)]
1495 : pub(crate) async fn preload(
1496 : self: &Arc<Self>,
1497 : remote_storage: &GenericRemoteStorage,
1498 : cancel: CancellationToken,
1499 : ) -> anyhow::Result<TenantPreload> {
1500 : span::debug_assert_current_span_has_tenant_id();
1501 : // Get list of remote timelines
1502 : // download index files for every tenant timeline
1503 : info!("listing remote timelines");
1504 : let (mut remote_timeline_ids, other_keys) = remote_timeline_client::list_remote_timelines(
1505 : remote_storage,
1506 : self.tenant_shard_id,
1507 : cancel.clone(),
1508 : )
1509 : .await?;
1510 : let (offloaded_add, tenant_manifest) =
1511 : match remote_timeline_client::download_tenant_manifest(
1512 : remote_storage,
1513 : &self.tenant_shard_id,
1514 : self.generation,
1515 : &cancel,
1516 : )
1517 : .await
1518 : {
1519 : Ok((tenant_manifest, _generation, _manifest_mtime)) => (
1520 : format!("{} offloaded", tenant_manifest.offloaded_timelines.len()),
1521 : tenant_manifest,
1522 : ),
1523 : Err(DownloadError::NotFound) => {
1524 : ("no manifest".to_string(), TenantManifest::empty())
1525 : }
1526 : Err(e) => Err(e)?,
1527 : };
1528 :
1529 : info!(
1530 : "found {} timelines, and {offloaded_add}",
1531 : remote_timeline_ids.len()
1532 : );
1533 :
1534 : for k in other_keys {
1535 : warn!("Unexpected non timeline key {k}");
1536 : }
1537 :
1538 : // Avoid downloading IndexPart of offloaded timelines.
1539 : let mut offloaded_with_prefix = HashSet::new();
1540 : for offloaded in tenant_manifest.offloaded_timelines.iter() {
1541 : if remote_timeline_ids.remove(&offloaded.timeline_id) {
1542 : offloaded_with_prefix.insert(offloaded.timeline_id);
1543 : } else {
1544 : // We'll take care later of timelines in the manifest without a prefix
1545 : }
1546 : }
1547 :
1548 : let timelines = self
1549 : .load_timelines_metadata(remote_timeline_ids, remote_storage, cancel)
1550 : .await?;
1551 :
1552 : Ok(TenantPreload {
1553 : tenant_manifest,
1554 : timelines: timelines
1555 : .into_iter()
1556 6 : .map(|(id, tl)| (id, Some(tl)))
1557 0 : .chain(offloaded_with_prefix.into_iter().map(|id| (id, None)))
1558 : .collect(),
1559 : })
1560 : }
1561 :
1562 : ///
1563 : /// Background task that downloads all data for a tenant and brings it to Active state.
1564 : ///
1565 : /// No background tasks are started as part of this routine.
1566 : ///
1567 220 : async fn attach(
1568 220 : self: &Arc<Tenant>,
1569 220 : preload: Option<TenantPreload>,
1570 220 : ctx: &RequestContext,
1571 220 : ) -> anyhow::Result<()> {
1572 220 : span::debug_assert_current_span_has_tenant_id();
1573 220 :
1574 220 : failpoint_support::sleep_millis_async!("before-attaching-tenant");
1575 :
1576 220 : let Some(preload) = preload else {
1577 0 : anyhow::bail!("local-only deployment is no longer supported, https://github.com/neondatabase/neon/issues/5624");
1578 : };
1579 :
1580 220 : let mut offloaded_timeline_ids = HashSet::new();
1581 220 : let mut offloaded_timelines_list = Vec::new();
1582 220 : for timeline_manifest in preload.tenant_manifest.offloaded_timelines.iter() {
1583 0 : let timeline_id = timeline_manifest.timeline_id;
1584 0 : let offloaded_timeline =
1585 0 : OffloadedTimeline::from_manifest(self.tenant_shard_id, timeline_manifest);
1586 0 : offloaded_timelines_list.push((timeline_id, Arc::new(offloaded_timeline)));
1587 0 : offloaded_timeline_ids.insert(timeline_id);
1588 0 : }
1589 : // Complete deletions for offloaded timeline id's from manifest.
1590 : // The manifest will be uploaded later in this function.
1591 220 : offloaded_timelines_list
1592 220 : .retain(|(offloaded_id, offloaded)| {
1593 0 : // Existence of a timeline is finally determined by the existence of an index-part.json in remote storage.
1594 0 : // If there is dangling references in another location, they need to be cleaned up.
1595 0 : let delete = !preload.timelines.contains_key(offloaded_id);
1596 0 : if delete {
1597 0 : tracing::info!("Removing offloaded timeline {offloaded_id} from manifest as no remote prefix was found");
1598 0 : offloaded.defuse_for_tenant_drop();
1599 0 : }
1600 0 : !delete
1601 220 : });
1602 220 :
1603 220 : let mut timelines_to_resume_deletions = vec![];
1604 220 :
1605 220 : let mut remote_index_and_client = HashMap::new();
1606 220 : let mut timeline_ancestors = HashMap::new();
1607 220 : let mut existent_timelines = HashSet::new();
1608 226 : for (timeline_id, preload) in preload.timelines {
1609 6 : let Some(preload) = preload else { continue };
1610 : // This is an invariant of the `preload` function's API
1611 6 : assert!(!offloaded_timeline_ids.contains(&timeline_id));
1612 6 : let index_part = match preload.index_part {
1613 6 : Ok(i) => {
1614 6 : debug!("remote index part exists for timeline {timeline_id}");
1615 : // We found index_part on the remote, this is the standard case.
1616 6 : existent_timelines.insert(timeline_id);
1617 6 : i
1618 : }
1619 : Err(DownloadError::NotFound) => {
1620 : // There is no index_part on the remote. We only get here
1621 : // if there is some prefix for the timeline in the remote storage.
1622 : // This can e.g. be the initdb.tar.zst archive, maybe a
1623 : // remnant from a prior incomplete creation or deletion attempt.
1624 : // Delete the local directory as the deciding criterion for a
1625 : // timeline's existence is presence of index_part.
1626 0 : info!(%timeline_id, "index_part not found on remote");
1627 0 : continue;
1628 : }
1629 0 : Err(DownloadError::Fatal(why)) => {
1630 0 : // If, while loading one remote timeline, we saw an indication that our generation
1631 0 : // number is likely invalid, then we should not load the whole tenant.
1632 0 : error!(%timeline_id, "Fatal error loading timeline: {why}");
1633 0 : anyhow::bail!(why.to_string());
1634 : }
1635 0 : Err(e) => {
1636 0 : // Some (possibly ephemeral) error happened during index_part download.
1637 0 : // Pretend the timeline exists to not delete the timeline directory,
1638 0 : // as it might be a temporary issue and we don't want to re-download
1639 0 : // everything after it resolves.
1640 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1641 :
1642 0 : existent_timelines.insert(timeline_id);
1643 0 : continue;
1644 : }
1645 : };
1646 6 : match index_part {
1647 6 : MaybeDeletedIndexPart::IndexPart(index_part) => {
1648 6 : timeline_ancestors.insert(timeline_id, index_part.metadata.clone());
1649 6 : remote_index_and_client.insert(timeline_id, (index_part, preload.client));
1650 6 : }
1651 0 : MaybeDeletedIndexPart::Deleted(index_part) => {
1652 0 : info!(
1653 0 : "timeline {} is deleted, picking to resume deletion",
1654 : timeline_id
1655 : );
1656 0 : timelines_to_resume_deletions.push((timeline_id, index_part, preload.client));
1657 : }
1658 : }
1659 : }
1660 :
1661 220 : let mut gc_blocks = HashMap::new();
1662 :
1663 : // For every timeline, download the metadata file, scan the local directory,
1664 : // and build a layer map that contains an entry for each remote and local
1665 : // layer file.
1666 220 : let sorted_timelines = tree_sort_timelines(timeline_ancestors, |m| m.ancestor_timeline())?;
1667 226 : for (timeline_id, remote_metadata) in sorted_timelines {
1668 6 : let (index_part, remote_client) = remote_index_and_client
1669 6 : .remove(&timeline_id)
1670 6 : .expect("just put it in above");
1671 :
1672 6 : if let Some(blocking) = index_part.gc_blocking.as_ref() {
1673 : // could just filter these away, but it helps while testing
1674 0 : anyhow::ensure!(
1675 0 : !blocking.reasons.is_empty(),
1676 0 : "index_part for {timeline_id} is malformed: it should not have gc blocking with zero reasons"
1677 : );
1678 0 : let prev = gc_blocks.insert(timeline_id, blocking.reasons);
1679 0 : assert!(prev.is_none());
1680 6 : }
1681 :
1682 : // TODO again handle early failure
1683 6 : let effect = self
1684 6 : .load_remote_timeline(
1685 6 : timeline_id,
1686 6 : index_part,
1687 6 : remote_metadata,
1688 6 : TimelineResources {
1689 6 : remote_client,
1690 6 : pagestream_throttle: self.pagestream_throttle.clone(),
1691 6 : pagestream_throttle_metrics: self.pagestream_throttle_metrics.clone(),
1692 6 : l0_flush_global_state: self.l0_flush_global_state.clone(),
1693 6 : },
1694 6 : LoadTimelineCause::Attach,
1695 6 : ctx,
1696 6 : )
1697 6 : .await
1698 6 : .with_context(|| {
1699 0 : format!(
1700 0 : "failed to load remote timeline {} for tenant {}",
1701 0 : timeline_id, self.tenant_shard_id
1702 0 : )
1703 6 : })?;
1704 :
1705 6 : match effect {
1706 6 : TimelineInitAndSyncResult::ReadyToActivate(_) => {
1707 6 : // activation happens later, on Tenant::activate
1708 6 : }
1709 : TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
1710 : TimelineInitAndSyncNeedsSpawnImportPgdata {
1711 0 : timeline,
1712 0 : import_pgdata,
1713 0 : guard,
1714 0 : },
1715 0 : ) => {
1716 0 : tokio::task::spawn(self.clone().create_timeline_import_pgdata_task(
1717 0 : timeline,
1718 0 : import_pgdata,
1719 0 : ActivateTimelineArgs::No,
1720 0 : guard,
1721 0 : ));
1722 0 : }
1723 : }
1724 : }
1725 :
1726 : // Walk through deleted timelines, resume deletion
1727 220 : for (timeline_id, index_part, remote_timeline_client) in timelines_to_resume_deletions {
1728 0 : remote_timeline_client
1729 0 : .init_upload_queue_stopped_to_continue_deletion(&index_part)
1730 0 : .context("init queue stopped")
1731 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1732 :
1733 0 : DeleteTimelineFlow::resume_deletion(
1734 0 : Arc::clone(self),
1735 0 : timeline_id,
1736 0 : &index_part.metadata,
1737 0 : remote_timeline_client,
1738 0 : )
1739 0 : .instrument(tracing::info_span!("timeline_delete", %timeline_id))
1740 0 : .await
1741 0 : .context("resume_deletion")
1742 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1743 : }
1744 220 : let needs_manifest_upload =
1745 220 : offloaded_timelines_list.len() != preload.tenant_manifest.offloaded_timelines.len();
1746 220 : {
1747 220 : let mut offloaded_timelines_accessor = self.timelines_offloaded.lock().unwrap();
1748 220 : offloaded_timelines_accessor.extend(offloaded_timelines_list.into_iter());
1749 220 : }
1750 220 : if needs_manifest_upload {
1751 0 : self.store_tenant_manifest().await?;
1752 220 : }
1753 :
1754 : // The local filesystem contents are a cache of what's in the remote IndexPart;
1755 : // IndexPart is the source of truth.
1756 220 : self.clean_up_timelines(&existent_timelines)?;
1757 :
1758 220 : self.gc_block.set_scanned(gc_blocks);
1759 220 :
1760 220 : fail::fail_point!("attach-before-activate", |_| {
1761 0 : anyhow::bail!("attach-before-activate");
1762 220 : });
1763 220 : failpoint_support::sleep_millis_async!("attach-before-activate-sleep", &self.cancel);
1764 :
1765 220 : info!("Done");
1766 :
1767 220 : Ok(())
1768 220 : }
1769 :
1770 : /// Check for any local timeline directories that are temporary, or do not correspond to a
1771 : /// timeline that still exists: this can happen if we crashed during a deletion/creation, or
1772 : /// if a timeline was deleted while the tenant was attached to a different pageserver.
1773 220 : fn clean_up_timelines(&self, existent_timelines: &HashSet<TimelineId>) -> anyhow::Result<()> {
1774 220 : let timelines_dir = self.conf.timelines_path(&self.tenant_shard_id);
1775 :
1776 220 : let entries = match timelines_dir.read_dir_utf8() {
1777 220 : Ok(d) => d,
1778 0 : Err(e) => {
1779 0 : if e.kind() == std::io::ErrorKind::NotFound {
1780 0 : return Ok(());
1781 : } else {
1782 0 : return Err(e).context("list timelines directory for tenant");
1783 : }
1784 : }
1785 : };
1786 :
1787 228 : for entry in entries {
1788 8 : let entry = entry.context("read timeline dir entry")?;
1789 8 : let entry_path = entry.path();
1790 :
1791 8 : let purge = if crate::is_temporary(entry_path)
1792 : // TODO: remove uninit mark code (https://github.com/neondatabase/neon/issues/5718)
1793 8 : || is_uninit_mark(entry_path)
1794 8 : || crate::is_delete_mark(entry_path)
1795 : {
1796 0 : true
1797 : } else {
1798 8 : match TimelineId::try_from(entry_path.file_name()) {
1799 8 : Ok(i) => {
1800 8 : // Purge if the timeline ID does not exist in remote storage: remote storage is the authority.
1801 8 : !existent_timelines.contains(&i)
1802 : }
1803 0 : Err(e) => {
1804 0 : tracing::warn!(
1805 0 : "Unparseable directory in timelines directory: {entry_path}, ignoring ({e})"
1806 : );
1807 : // Do not purge junk: if we don't recognize it, be cautious and leave it for a human.
1808 0 : false
1809 : }
1810 : }
1811 : };
1812 :
1813 8 : if purge {
1814 2 : tracing::info!("Purging stale timeline dentry {entry_path}");
1815 2 : if let Err(e) = match entry.file_type() {
1816 2 : Ok(t) => if t.is_dir() {
1817 2 : std::fs::remove_dir_all(entry_path)
1818 : } else {
1819 0 : std::fs::remove_file(entry_path)
1820 : }
1821 2 : .or_else(fs_ext::ignore_not_found),
1822 0 : Err(e) => Err(e),
1823 : } {
1824 0 : tracing::warn!("Failed to purge stale timeline dentry {entry_path}: {e}");
1825 2 : }
1826 6 : }
1827 : }
1828 :
1829 220 : Ok(())
1830 220 : }
1831 :
1832 : /// Get sum of all remote timelines sizes
1833 : ///
1834 : /// This function relies on the index_part instead of listing the remote storage
1835 0 : pub fn remote_size(&self) -> u64 {
1836 0 : let mut size = 0;
1837 :
1838 0 : for timeline in self.list_timelines() {
1839 0 : size += timeline.remote_client.get_remote_physical_size();
1840 0 : }
1841 :
1842 0 : size
1843 0 : }
1844 :
1845 : #[instrument(skip_all, fields(timeline_id=%timeline_id))]
1846 : async fn load_remote_timeline(
1847 : self: &Arc<Self>,
1848 : timeline_id: TimelineId,
1849 : index_part: IndexPart,
1850 : remote_metadata: TimelineMetadata,
1851 : resources: TimelineResources,
1852 : cause: LoadTimelineCause,
1853 : ctx: &RequestContext,
1854 : ) -> anyhow::Result<TimelineInitAndSyncResult> {
1855 : span::debug_assert_current_span_has_tenant_id();
1856 :
1857 : info!("downloading index file for timeline {}", timeline_id);
1858 : tokio::fs::create_dir_all(self.conf.timeline_path(&self.tenant_shard_id, &timeline_id))
1859 : .await
1860 : .context("Failed to create new timeline directory")?;
1861 :
1862 : let ancestor = if let Some(ancestor_id) = remote_metadata.ancestor_timeline() {
1863 : let timelines = self.timelines.lock().unwrap();
1864 : Some(Arc::clone(timelines.get(&ancestor_id).ok_or_else(
1865 0 : || {
1866 0 : anyhow::anyhow!(
1867 0 : "cannot find ancestor timeline {ancestor_id} for timeline {timeline_id}"
1868 0 : )
1869 0 : },
1870 : )?))
1871 : } else {
1872 : None
1873 : };
1874 :
1875 : self.timeline_init_and_sync(
1876 : timeline_id,
1877 : resources,
1878 : index_part,
1879 : remote_metadata,
1880 : ancestor,
1881 : cause,
1882 : ctx,
1883 : )
1884 : .await
1885 : }
1886 :
1887 220 : async fn load_timelines_metadata(
1888 220 : self: &Arc<Tenant>,
1889 220 : timeline_ids: HashSet<TimelineId>,
1890 220 : remote_storage: &GenericRemoteStorage,
1891 220 : cancel: CancellationToken,
1892 220 : ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
1893 220 : let mut part_downloads = JoinSet::new();
1894 226 : for timeline_id in timeline_ids {
1895 6 : let cancel_clone = cancel.clone();
1896 6 : part_downloads.spawn(
1897 6 : self.load_timeline_metadata(timeline_id, remote_storage.clone(), cancel_clone)
1898 6 : .instrument(info_span!("download_index_part", %timeline_id)),
1899 : );
1900 : }
1901 :
1902 220 : let mut timeline_preloads: HashMap<TimelineId, TimelinePreload> = HashMap::new();
1903 :
1904 : loop {
1905 226 : tokio::select!(
1906 226 : next = part_downloads.join_next() => {
1907 226 : match next {
1908 6 : Some(result) => {
1909 6 : let preload = result.context("join preload task")?;
1910 6 : timeline_preloads.insert(preload.timeline_id, preload);
1911 : },
1912 : None => {
1913 220 : break;
1914 : }
1915 : }
1916 : },
1917 226 : _ = cancel.cancelled() => {
1918 0 : anyhow::bail!("Cancelled while waiting for remote index download")
1919 : }
1920 : )
1921 : }
1922 :
1923 220 : Ok(timeline_preloads)
1924 220 : }
1925 :
1926 6 : fn build_timeline_client(
1927 6 : &self,
1928 6 : timeline_id: TimelineId,
1929 6 : remote_storage: GenericRemoteStorage,
1930 6 : ) -> RemoteTimelineClient {
1931 6 : RemoteTimelineClient::new(
1932 6 : remote_storage.clone(),
1933 6 : self.deletion_queue_client.clone(),
1934 6 : self.conf,
1935 6 : self.tenant_shard_id,
1936 6 : timeline_id,
1937 6 : self.generation,
1938 6 : &self.tenant_conf.load().location,
1939 6 : )
1940 6 : }
1941 :
1942 6 : fn load_timeline_metadata(
1943 6 : self: &Arc<Tenant>,
1944 6 : timeline_id: TimelineId,
1945 6 : remote_storage: GenericRemoteStorage,
1946 6 : cancel: CancellationToken,
1947 6 : ) -> impl Future<Output = TimelinePreload> {
1948 6 : let client = self.build_timeline_client(timeline_id, remote_storage);
1949 6 : async move {
1950 6 : debug_assert_current_span_has_tenant_and_timeline_id();
1951 6 : debug!("starting index part download");
1952 :
1953 6 : let index_part = client.download_index_file(&cancel).await;
1954 :
1955 6 : debug!("finished index part download");
1956 :
1957 6 : TimelinePreload {
1958 6 : client,
1959 6 : timeline_id,
1960 6 : index_part,
1961 6 : }
1962 6 : }
1963 6 : }
1964 :
1965 0 : fn check_to_be_archived_has_no_unarchived_children(
1966 0 : timeline_id: TimelineId,
1967 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
1968 0 : ) -> Result<(), TimelineArchivalError> {
1969 0 : let children: Vec<TimelineId> = timelines
1970 0 : .iter()
1971 0 : .filter_map(|(id, entry)| {
1972 0 : if entry.get_ancestor_timeline_id() != Some(timeline_id) {
1973 0 : return None;
1974 0 : }
1975 0 : if entry.is_archived() == Some(true) {
1976 0 : return None;
1977 0 : }
1978 0 : Some(*id)
1979 0 : })
1980 0 : .collect();
1981 0 :
1982 0 : if !children.is_empty() {
1983 0 : return Err(TimelineArchivalError::HasUnarchivedChildren(children));
1984 0 : }
1985 0 : Ok(())
1986 0 : }
1987 :
1988 0 : fn check_ancestor_of_to_be_unarchived_is_not_archived(
1989 0 : ancestor_timeline_id: TimelineId,
1990 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
1991 0 : offloaded_timelines: &std::sync::MutexGuard<
1992 0 : '_,
1993 0 : HashMap<TimelineId, Arc<OffloadedTimeline>>,
1994 0 : >,
1995 0 : ) -> Result<(), TimelineArchivalError> {
1996 0 : let has_archived_parent =
1997 0 : if let Some(ancestor_timeline) = timelines.get(&ancestor_timeline_id) {
1998 0 : ancestor_timeline.is_archived() == Some(true)
1999 0 : } else if offloaded_timelines.contains_key(&ancestor_timeline_id) {
2000 0 : true
2001 : } else {
2002 0 : error!("ancestor timeline {ancestor_timeline_id} not found");
2003 0 : if cfg!(debug_assertions) {
2004 0 : panic!("ancestor timeline {ancestor_timeline_id} not found");
2005 0 : }
2006 0 : return Err(TimelineArchivalError::NotFound);
2007 : };
2008 0 : if has_archived_parent {
2009 0 : return Err(TimelineArchivalError::HasArchivedParent(
2010 0 : ancestor_timeline_id,
2011 0 : ));
2012 0 : }
2013 0 : Ok(())
2014 0 : }
2015 :
2016 0 : fn check_to_be_unarchived_timeline_has_no_archived_parent(
2017 0 : timeline: &Arc<Timeline>,
2018 0 : ) -> Result<(), TimelineArchivalError> {
2019 0 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
2020 0 : if ancestor_timeline.is_archived() == Some(true) {
2021 0 : return Err(TimelineArchivalError::HasArchivedParent(
2022 0 : ancestor_timeline.timeline_id,
2023 0 : ));
2024 0 : }
2025 0 : }
2026 0 : Ok(())
2027 0 : }
2028 :
2029 : /// Loads the specified (offloaded) timeline from S3 and attaches it as a loaded timeline
2030 : ///
2031 : /// Counterpart to [`offload_timeline`].
2032 0 : async fn unoffload_timeline(
2033 0 : self: &Arc<Self>,
2034 0 : timeline_id: TimelineId,
2035 0 : broker_client: storage_broker::BrokerClientChannel,
2036 0 : ctx: RequestContext,
2037 0 : ) -> Result<Arc<Timeline>, TimelineArchivalError> {
2038 0 : info!("unoffloading timeline");
2039 :
2040 : // We activate the timeline below manually, so this must be called on an active tenant.
2041 : // We expect callers of this function to ensure this.
2042 0 : match self.current_state() {
2043 : TenantState::Activating { .. }
2044 : | TenantState::Attaching
2045 : | TenantState::Broken { .. } => {
2046 0 : panic!("Timeline expected to be active")
2047 : }
2048 0 : TenantState::Stopping { .. } => return Err(TimelineArchivalError::Cancelled),
2049 0 : TenantState::Active => {}
2050 0 : }
2051 0 : let cancel = self.cancel.clone();
2052 0 :
2053 0 : // Protect against concurrent attempts to use this TimelineId
2054 0 : // We don't care much about idempotency, as it's ensured a layer above.
2055 0 : let allow_offloaded = true;
2056 0 : let _create_guard = self
2057 0 : .create_timeline_create_guard(
2058 0 : timeline_id,
2059 0 : CreateTimelineIdempotency::FailWithConflict,
2060 0 : allow_offloaded,
2061 0 : )
2062 0 : .map_err(|err| match err {
2063 0 : TimelineExclusionError::AlreadyCreating => TimelineArchivalError::AlreadyInProgress,
2064 : TimelineExclusionError::AlreadyExists { .. } => {
2065 0 : TimelineArchivalError::Other(anyhow::anyhow!("Timeline already exists"))
2066 : }
2067 0 : TimelineExclusionError::Other(e) => TimelineArchivalError::Other(e),
2068 0 : TimelineExclusionError::ShuttingDown => TimelineArchivalError::Cancelled,
2069 0 : })?;
2070 :
2071 0 : let timeline_preload = self
2072 0 : .load_timeline_metadata(timeline_id, self.remote_storage.clone(), cancel.clone())
2073 0 : .await;
2074 :
2075 0 : let index_part = match timeline_preload.index_part {
2076 0 : Ok(index_part) => {
2077 0 : debug!("remote index part exists for timeline {timeline_id}");
2078 0 : index_part
2079 : }
2080 : Err(DownloadError::NotFound) => {
2081 0 : error!(%timeline_id, "index_part not found on remote");
2082 0 : return Err(TimelineArchivalError::NotFound);
2083 : }
2084 0 : Err(DownloadError::Cancelled) => return Err(TimelineArchivalError::Cancelled),
2085 0 : Err(e) => {
2086 0 : // Some (possibly ephemeral) error happened during index_part download.
2087 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
2088 0 : return Err(TimelineArchivalError::Other(
2089 0 : anyhow::Error::new(e).context("downloading index_part from remote storage"),
2090 0 : ));
2091 : }
2092 : };
2093 0 : let index_part = match index_part {
2094 0 : MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
2095 0 : MaybeDeletedIndexPart::Deleted(_index_part) => {
2096 0 : info!("timeline is deleted according to index_part.json");
2097 0 : return Err(TimelineArchivalError::NotFound);
2098 : }
2099 : };
2100 0 : let remote_metadata = index_part.metadata.clone();
2101 0 : let timeline_resources = self.build_timeline_resources(timeline_id);
2102 0 : self.load_remote_timeline(
2103 0 : timeline_id,
2104 0 : index_part,
2105 0 : remote_metadata,
2106 0 : timeline_resources,
2107 0 : LoadTimelineCause::Unoffload,
2108 0 : &ctx,
2109 0 : )
2110 0 : .await
2111 0 : .with_context(|| {
2112 0 : format!(
2113 0 : "failed to load remote timeline {} for tenant {}",
2114 0 : timeline_id, self.tenant_shard_id
2115 0 : )
2116 0 : })
2117 0 : .map_err(TimelineArchivalError::Other)?;
2118 :
2119 0 : let timeline = {
2120 0 : let timelines = self.timelines.lock().unwrap();
2121 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2122 0 : warn!("timeline not available directly after attach");
2123 : // This is not a panic because no locks are held between `load_remote_timeline`
2124 : // which puts the timeline into timelines, and our look into the timeline map.
2125 0 : return Err(TimelineArchivalError::Other(anyhow::anyhow!(
2126 0 : "timeline not available directly after attach"
2127 0 : )));
2128 : };
2129 0 : let mut offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2130 0 : match offloaded_timelines.remove(&timeline_id) {
2131 0 : Some(offloaded) => {
2132 0 : offloaded.delete_from_ancestor_with_timelines(&timelines);
2133 0 : }
2134 0 : None => warn!("timeline already removed from offloaded timelines"),
2135 : }
2136 :
2137 0 : self.initialize_gc_info(&timelines, &offloaded_timelines, Some(timeline_id));
2138 0 :
2139 0 : Arc::clone(timeline)
2140 0 : };
2141 0 :
2142 0 : // Upload new list of offloaded timelines to S3
2143 0 : self.store_tenant_manifest().await?;
2144 :
2145 : // Activate the timeline (if it makes sense)
2146 0 : if !(timeline.is_broken() || timeline.is_stopping()) {
2147 0 : let background_jobs_can_start = None;
2148 0 : timeline.activate(
2149 0 : self.clone(),
2150 0 : broker_client.clone(),
2151 0 : background_jobs_can_start,
2152 0 : &ctx,
2153 0 : );
2154 0 : }
2155 :
2156 0 : info!("timeline unoffloading complete");
2157 0 : Ok(timeline)
2158 0 : }
2159 :
2160 0 : pub(crate) async fn apply_timeline_archival_config(
2161 0 : self: &Arc<Self>,
2162 0 : timeline_id: TimelineId,
2163 0 : new_state: TimelineArchivalState,
2164 0 : broker_client: storage_broker::BrokerClientChannel,
2165 0 : ctx: RequestContext,
2166 0 : ) -> Result<(), TimelineArchivalError> {
2167 0 : info!("setting timeline archival config");
2168 : // First part: figure out what is needed to do, and do validation
2169 0 : let timeline_or_unarchive_offloaded = 'outer: {
2170 0 : let timelines = self.timelines.lock().unwrap();
2171 :
2172 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2173 0 : let offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2174 0 : let Some(offloaded) = offloaded_timelines.get(&timeline_id) else {
2175 0 : return Err(TimelineArchivalError::NotFound);
2176 : };
2177 0 : if new_state == TimelineArchivalState::Archived {
2178 : // It's offloaded already, so nothing to do
2179 0 : return Ok(());
2180 0 : }
2181 0 : if let Some(ancestor_timeline_id) = offloaded.ancestor_timeline_id {
2182 0 : Self::check_ancestor_of_to_be_unarchived_is_not_archived(
2183 0 : ancestor_timeline_id,
2184 0 : &timelines,
2185 0 : &offloaded_timelines,
2186 0 : )?;
2187 0 : }
2188 0 : break 'outer None;
2189 : };
2190 :
2191 : // Do some validation. We release the timelines lock below, so there is potential
2192 : // for race conditions: these checks are more present to prevent misunderstandings of
2193 : // the API's capabilities, instead of serving as the sole way to defend their invariants.
2194 0 : match new_state {
2195 : TimelineArchivalState::Unarchived => {
2196 0 : Self::check_to_be_unarchived_timeline_has_no_archived_parent(timeline)?
2197 : }
2198 : TimelineArchivalState::Archived => {
2199 0 : Self::check_to_be_archived_has_no_unarchived_children(timeline_id, &timelines)?
2200 : }
2201 : }
2202 0 : Some(Arc::clone(timeline))
2203 : };
2204 :
2205 : // Second part: unoffload timeline (if needed)
2206 0 : let timeline = if let Some(timeline) = timeline_or_unarchive_offloaded {
2207 0 : timeline
2208 : } else {
2209 : // Turn offloaded timeline into a non-offloaded one
2210 0 : self.unoffload_timeline(timeline_id, broker_client, ctx)
2211 0 : .await?
2212 : };
2213 :
2214 : // Third part: upload new timeline archival state and block until it is present in S3
2215 0 : let upload_needed = match timeline
2216 0 : .remote_client
2217 0 : .schedule_index_upload_for_timeline_archival_state(new_state)
2218 : {
2219 0 : Ok(upload_needed) => upload_needed,
2220 0 : Err(e) => {
2221 0 : if timeline.cancel.is_cancelled() {
2222 0 : return Err(TimelineArchivalError::Cancelled);
2223 : } else {
2224 0 : return Err(TimelineArchivalError::Other(e));
2225 : }
2226 : }
2227 : };
2228 :
2229 0 : if upload_needed {
2230 0 : info!("Uploading new state");
2231 : const MAX_WAIT: Duration = Duration::from_secs(10);
2232 0 : let Ok(v) =
2233 0 : tokio::time::timeout(MAX_WAIT, timeline.remote_client.wait_completion()).await
2234 : else {
2235 0 : tracing::warn!("reached timeout for waiting on upload queue");
2236 0 : return Err(TimelineArchivalError::Timeout);
2237 : };
2238 0 : v.map_err(|e| match e {
2239 0 : WaitCompletionError::NotInitialized(e) => {
2240 0 : TimelineArchivalError::Other(anyhow::anyhow!(e))
2241 : }
2242 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2243 0 : TimelineArchivalError::Cancelled
2244 : }
2245 0 : })?;
2246 0 : }
2247 0 : Ok(())
2248 0 : }
2249 :
2250 2 : pub fn get_offloaded_timeline(
2251 2 : &self,
2252 2 : timeline_id: TimelineId,
2253 2 : ) -> Result<Arc<OffloadedTimeline>, GetTimelineError> {
2254 2 : self.timelines_offloaded
2255 2 : .lock()
2256 2 : .unwrap()
2257 2 : .get(&timeline_id)
2258 2 : .map(Arc::clone)
2259 2 : .ok_or(GetTimelineError::NotFound {
2260 2 : tenant_id: self.tenant_shard_id,
2261 2 : timeline_id,
2262 2 : })
2263 2 : }
2264 :
2265 4 : pub(crate) fn tenant_shard_id(&self) -> TenantShardId {
2266 4 : self.tenant_shard_id
2267 4 : }
2268 :
2269 : /// Get Timeline handle for given Neon timeline ID.
2270 : /// This function is idempotent. It doesn't change internal state in any way.
2271 222 : pub fn get_timeline(
2272 222 : &self,
2273 222 : timeline_id: TimelineId,
2274 222 : active_only: bool,
2275 222 : ) -> Result<Arc<Timeline>, GetTimelineError> {
2276 222 : let timelines_accessor = self.timelines.lock().unwrap();
2277 222 : let timeline = timelines_accessor
2278 222 : .get(&timeline_id)
2279 222 : .ok_or(GetTimelineError::NotFound {
2280 222 : tenant_id: self.tenant_shard_id,
2281 222 : timeline_id,
2282 222 : })?;
2283 :
2284 220 : if active_only && !timeline.is_active() {
2285 0 : Err(GetTimelineError::NotActive {
2286 0 : tenant_id: self.tenant_shard_id,
2287 0 : timeline_id,
2288 0 : state: timeline.current_state(),
2289 0 : })
2290 : } else {
2291 220 : Ok(Arc::clone(timeline))
2292 : }
2293 222 : }
2294 :
2295 : /// Lists timelines the tenant contains.
2296 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2297 0 : pub fn list_timelines(&self) -> Vec<Arc<Timeline>> {
2298 0 : self.timelines
2299 0 : .lock()
2300 0 : .unwrap()
2301 0 : .values()
2302 0 : .map(Arc::clone)
2303 0 : .collect()
2304 0 : }
2305 :
2306 : /// Lists timelines the tenant manages, including offloaded ones.
2307 : ///
2308 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2309 0 : pub fn list_timelines_and_offloaded(
2310 0 : &self,
2311 0 : ) -> (Vec<Arc<Timeline>>, Vec<Arc<OffloadedTimeline>>) {
2312 0 : let timelines = self
2313 0 : .timelines
2314 0 : .lock()
2315 0 : .unwrap()
2316 0 : .values()
2317 0 : .map(Arc::clone)
2318 0 : .collect();
2319 0 : let offloaded = self
2320 0 : .timelines_offloaded
2321 0 : .lock()
2322 0 : .unwrap()
2323 0 : .values()
2324 0 : .map(Arc::clone)
2325 0 : .collect();
2326 0 : (timelines, offloaded)
2327 0 : }
2328 :
2329 0 : pub fn list_timeline_ids(&self) -> Vec<TimelineId> {
2330 0 : self.timelines.lock().unwrap().keys().cloned().collect()
2331 0 : }
2332 :
2333 : /// This is used by tests & import-from-basebackup.
2334 : ///
2335 : /// The returned [`UninitializedTimeline`] contains no data nor metadata and it is in
2336 : /// a state that will fail [`Tenant::load_remote_timeline`] because `disk_consistent_lsn=Lsn(0)`.
2337 : ///
2338 : /// The caller is responsible for getting the timeline into a state that will be accepted
2339 : /// by [`Tenant::load_remote_timeline`] / [`Tenant::attach`].
2340 : /// Then they may call [`UninitializedTimeline::finish_creation`] to add the timeline
2341 : /// to the [`Tenant::timelines`].
2342 : ///
2343 : /// Tests should use `Tenant::create_test_timeline` to set up the minimum required metadata keys.
2344 212 : pub(crate) async fn create_empty_timeline(
2345 212 : self: &Arc<Self>,
2346 212 : new_timeline_id: TimelineId,
2347 212 : initdb_lsn: Lsn,
2348 212 : pg_version: u32,
2349 212 : _ctx: &RequestContext,
2350 212 : ) -> anyhow::Result<UninitializedTimeline> {
2351 212 : anyhow::ensure!(
2352 212 : self.is_active(),
2353 0 : "Cannot create empty timelines on inactive tenant"
2354 : );
2355 :
2356 : // Protect against concurrent attempts to use this TimelineId
2357 212 : let create_guard = match self
2358 212 : .start_creating_timeline(new_timeline_id, CreateTimelineIdempotency::FailWithConflict)
2359 212 : .await?
2360 : {
2361 210 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2362 : StartCreatingTimelineResult::Idempotent(_) => {
2363 0 : unreachable!("FailWithConflict implies we get an error instead")
2364 : }
2365 : };
2366 :
2367 210 : let new_metadata = TimelineMetadata::new(
2368 210 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2369 210 : // make it valid, before calling finish_creation()
2370 210 : Lsn(0),
2371 210 : None,
2372 210 : None,
2373 210 : Lsn(0),
2374 210 : initdb_lsn,
2375 210 : initdb_lsn,
2376 210 : pg_version,
2377 210 : );
2378 210 : self.prepare_new_timeline(
2379 210 : new_timeline_id,
2380 210 : &new_metadata,
2381 210 : create_guard,
2382 210 : initdb_lsn,
2383 210 : None,
2384 210 : )
2385 210 : .await
2386 212 : }
2387 :
2388 : /// Helper for unit tests to create an empty timeline.
2389 : ///
2390 : /// The timeline is has state value `Active` but its background loops are not running.
2391 : // This makes the various functions which anyhow::ensure! for Active state work in tests.
2392 : // Our current tests don't need the background loops.
2393 : #[cfg(test)]
2394 202 : pub async fn create_test_timeline(
2395 202 : self: &Arc<Self>,
2396 202 : new_timeline_id: TimelineId,
2397 202 : initdb_lsn: Lsn,
2398 202 : pg_version: u32,
2399 202 : ctx: &RequestContext,
2400 202 : ) -> anyhow::Result<Arc<Timeline>> {
2401 202 : let uninit_tl = self
2402 202 : .create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2403 202 : .await?;
2404 202 : let tline = uninit_tl.raw_timeline().expect("we just created it");
2405 202 : assert_eq!(tline.get_last_record_lsn(), Lsn(0));
2406 :
2407 : // Setup minimum keys required for the timeline to be usable.
2408 202 : let mut modification = tline.begin_modification(initdb_lsn);
2409 202 : modification
2410 202 : .init_empty_test_timeline()
2411 202 : .context("init_empty_test_timeline")?;
2412 202 : modification
2413 202 : .commit(ctx)
2414 202 : .await
2415 202 : .context("commit init_empty_test_timeline modification")?;
2416 :
2417 : // Flush to disk so that uninit_tl's check for valid disk_consistent_lsn passes.
2418 202 : tline.maybe_spawn_flush_loop();
2419 202 : tline.freeze_and_flush().await.context("freeze_and_flush")?;
2420 :
2421 : // Make sure the freeze_and_flush reaches remote storage.
2422 202 : tline.remote_client.wait_completion().await.unwrap();
2423 :
2424 202 : let tl = uninit_tl.finish_creation()?;
2425 : // The non-test code would call tl.activate() here.
2426 202 : tl.set_state(TimelineState::Active);
2427 202 : Ok(tl)
2428 202 : }
2429 :
2430 : /// Helper for unit tests to create a timeline with some pre-loaded states.
2431 : #[cfg(test)]
2432 : #[allow(clippy::too_many_arguments)]
2433 36 : pub async fn create_test_timeline_with_layers(
2434 36 : self: &Arc<Self>,
2435 36 : new_timeline_id: TimelineId,
2436 36 : initdb_lsn: Lsn,
2437 36 : pg_version: u32,
2438 36 : ctx: &RequestContext,
2439 36 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
2440 36 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
2441 36 : end_lsn: Lsn,
2442 36 : ) -> anyhow::Result<Arc<Timeline>> {
2443 : use checks::check_valid_layermap;
2444 : use itertools::Itertools;
2445 :
2446 36 : let tline = self
2447 36 : .create_test_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2448 36 : .await?;
2449 36 : tline.force_advance_lsn(end_lsn);
2450 120 : for deltas in delta_layer_desc {
2451 84 : tline
2452 84 : .force_create_delta_layer(deltas, Some(initdb_lsn), ctx)
2453 84 : .await?;
2454 : }
2455 88 : for (lsn, images) in image_layer_desc {
2456 52 : tline
2457 52 : .force_create_image_layer(lsn, images, Some(initdb_lsn), ctx)
2458 52 : .await?;
2459 : }
2460 36 : let layer_names = tline
2461 36 : .layers
2462 36 : .read()
2463 36 : .await
2464 36 : .layer_map()
2465 36 : .unwrap()
2466 36 : .iter_historic_layers()
2467 172 : .map(|layer| layer.layer_name())
2468 36 : .collect_vec();
2469 36 : if let Some(err) = check_valid_layermap(&layer_names) {
2470 0 : bail!("invalid layermap: {err}");
2471 36 : }
2472 36 : Ok(tline)
2473 36 : }
2474 :
2475 : /// Create a new timeline.
2476 : ///
2477 : /// Returns the new timeline ID and reference to its Timeline object.
2478 : ///
2479 : /// If the caller specified the timeline ID to use (`new_timeline_id`), and timeline with
2480 : /// the same timeline ID already exists, returns CreateTimelineError::AlreadyExists.
2481 : #[allow(clippy::too_many_arguments)]
2482 0 : pub(crate) async fn create_timeline(
2483 0 : self: &Arc<Tenant>,
2484 0 : params: CreateTimelineParams,
2485 0 : broker_client: storage_broker::BrokerClientChannel,
2486 0 : ctx: &RequestContext,
2487 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
2488 0 : if !self.is_active() {
2489 0 : if matches!(self.current_state(), TenantState::Stopping { .. }) {
2490 0 : return Err(CreateTimelineError::ShuttingDown);
2491 : } else {
2492 0 : return Err(CreateTimelineError::Other(anyhow::anyhow!(
2493 0 : "Cannot create timelines on inactive tenant"
2494 0 : )));
2495 : }
2496 0 : }
2497 :
2498 0 : let _gate = self
2499 0 : .gate
2500 0 : .enter()
2501 0 : .map_err(|_| CreateTimelineError::ShuttingDown)?;
2502 :
2503 0 : let result: CreateTimelineResult = match params {
2504 : CreateTimelineParams::Bootstrap(CreateTimelineParamsBootstrap {
2505 0 : new_timeline_id,
2506 0 : existing_initdb_timeline_id,
2507 0 : pg_version,
2508 0 : }) => {
2509 0 : self.bootstrap_timeline(
2510 0 : new_timeline_id,
2511 0 : pg_version,
2512 0 : existing_initdb_timeline_id,
2513 0 : ctx,
2514 0 : )
2515 0 : .await?
2516 : }
2517 : CreateTimelineParams::Branch(CreateTimelineParamsBranch {
2518 0 : new_timeline_id,
2519 0 : ancestor_timeline_id,
2520 0 : mut ancestor_start_lsn,
2521 : }) => {
2522 0 : let ancestor_timeline = self
2523 0 : .get_timeline(ancestor_timeline_id, false)
2524 0 : .context("Cannot branch off the timeline that's not present in pageserver")?;
2525 :
2526 : // instead of waiting around, just deny the request because ancestor is not yet
2527 : // ready for other purposes either.
2528 0 : if !ancestor_timeline.is_active() {
2529 0 : return Err(CreateTimelineError::AncestorNotActive);
2530 0 : }
2531 0 :
2532 0 : if ancestor_timeline.is_archived() == Some(true) {
2533 0 : info!("tried to branch archived timeline");
2534 0 : return Err(CreateTimelineError::AncestorArchived);
2535 0 : }
2536 :
2537 0 : if let Some(lsn) = ancestor_start_lsn.as_mut() {
2538 0 : *lsn = lsn.align();
2539 0 :
2540 0 : let ancestor_ancestor_lsn = ancestor_timeline.get_ancestor_lsn();
2541 0 : if ancestor_ancestor_lsn > *lsn {
2542 : // can we safely just branch from the ancestor instead?
2543 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
2544 0 : "invalid start lsn {} for ancestor timeline {}: less than timeline ancestor lsn {}",
2545 0 : lsn,
2546 0 : ancestor_timeline_id,
2547 0 : ancestor_ancestor_lsn,
2548 0 : )));
2549 0 : }
2550 0 :
2551 0 : // Wait for the WAL to arrive and be processed on the parent branch up
2552 0 : // to the requested branch point. The repository code itself doesn't
2553 0 : // require it, but if we start to receive WAL on the new timeline,
2554 0 : // decoding the new WAL might need to look up previous pages, relation
2555 0 : // sizes etc. and that would get confused if the previous page versions
2556 0 : // are not in the repository yet.
2557 0 : ancestor_timeline
2558 0 : .wait_lsn(*lsn, timeline::WaitLsnWaiter::Tenant, ctx)
2559 0 : .await
2560 0 : .map_err(|e| match e {
2561 0 : e @ (WaitLsnError::Timeout(_) | WaitLsnError::BadState { .. }) => {
2562 0 : CreateTimelineError::AncestorLsn(anyhow::anyhow!(e))
2563 : }
2564 0 : WaitLsnError::Shutdown => CreateTimelineError::ShuttingDown,
2565 0 : })?;
2566 0 : }
2567 :
2568 0 : self.branch_timeline(&ancestor_timeline, new_timeline_id, ancestor_start_lsn, ctx)
2569 0 : .await?
2570 : }
2571 0 : CreateTimelineParams::ImportPgdata(params) => {
2572 0 : self.create_timeline_import_pgdata(
2573 0 : params,
2574 0 : ActivateTimelineArgs::Yes {
2575 0 : broker_client: broker_client.clone(),
2576 0 : },
2577 0 : ctx,
2578 0 : )
2579 0 : .await?
2580 : }
2581 : };
2582 :
2583 : // At this point we have dropped our guard on [`Self::timelines_creating`], and
2584 : // the timeline is visible in [`Self::timelines`], but it is _not_ durable yet. We must
2585 : // not send a success to the caller until it is. The same applies to idempotent retries.
2586 : //
2587 : // TODO: the timeline is already visible in [`Self::timelines`]; a caller could incorrectly
2588 : // assume that, because they can see the timeline via API, that the creation is done and
2589 : // that it is durable. Ideally, we would keep the timeline hidden (in [`Self::timelines_creating`])
2590 : // until it is durable, e.g., by extending the time we hold the creation guard. This also
2591 : // interacts with UninitializedTimeline and is generally a bit tricky.
2592 : //
2593 : // To re-emphasize: the only correct way to create a timeline is to repeat calling the
2594 : // creation API until it returns success. Only then is durability guaranteed.
2595 0 : info!(creation_result=%result.discriminant(), "waiting for timeline to be durable");
2596 0 : result
2597 0 : .timeline()
2598 0 : .remote_client
2599 0 : .wait_completion()
2600 0 : .await
2601 0 : .map_err(|e| match e {
2602 : WaitCompletionError::NotInitialized(
2603 0 : e, // If the queue is already stopped, it's a shutdown error.
2604 0 : ) if e.is_stopping() => CreateTimelineError::ShuttingDown,
2605 : WaitCompletionError::NotInitialized(_) => {
2606 : // This is a bug: we should never try to wait for uploads before initializing the timeline
2607 0 : debug_assert!(false);
2608 0 : CreateTimelineError::Other(anyhow::anyhow!("timeline not initialized"))
2609 : }
2610 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2611 0 : CreateTimelineError::ShuttingDown
2612 : }
2613 0 : })?;
2614 :
2615 : // The creating task is responsible for activating the timeline.
2616 : // We do this after `wait_completion()` so that we don't spin up tasks that start
2617 : // doing stuff before the IndexPart is durable in S3, which is done by the previous section.
2618 0 : let activated_timeline = match result {
2619 0 : CreateTimelineResult::Created(timeline) => {
2620 0 : timeline.activate(self.clone(), broker_client, None, ctx);
2621 0 : timeline
2622 : }
2623 0 : CreateTimelineResult::Idempotent(timeline) => {
2624 0 : info!(
2625 0 : "request was deemed idempotent, activation will be done by the creating task"
2626 : );
2627 0 : timeline
2628 : }
2629 0 : CreateTimelineResult::ImportSpawned(timeline) => {
2630 0 : info!("import task spawned, timeline will become visible and activated once the import is done");
2631 0 : timeline
2632 : }
2633 : };
2634 :
2635 0 : Ok(activated_timeline)
2636 0 : }
2637 :
2638 : /// The returned [`Arc<Timeline>`] is NOT in the [`Tenant::timelines`] map until the import
2639 : /// completes in the background. A DIFFERENT [`Arc<Timeline>`] will be inserted into the
2640 : /// [`Tenant::timelines`] map when the import completes.
2641 : /// We only return an [`Arc<Timeline>`] here so the API handler can create a [`pageserver_api::models::TimelineInfo`]
2642 : /// for the response.
2643 0 : async fn create_timeline_import_pgdata(
2644 0 : self: &Arc<Tenant>,
2645 0 : params: CreateTimelineParamsImportPgdata,
2646 0 : activate: ActivateTimelineArgs,
2647 0 : ctx: &RequestContext,
2648 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
2649 0 : let CreateTimelineParamsImportPgdata {
2650 0 : new_timeline_id,
2651 0 : location,
2652 0 : idempotency_key,
2653 0 : } = params;
2654 0 :
2655 0 : let started_at = chrono::Utc::now().naive_utc();
2656 :
2657 : //
2658 : // There's probably a simpler way to upload an index part, but, remote_timeline_client
2659 : // is the canonical way we do it.
2660 : // - create an empty timeline in-memory
2661 : // - use its remote_timeline_client to do the upload
2662 : // - dispose of the uninit timeline
2663 : // - keep the creation guard alive
2664 :
2665 0 : let timeline_create_guard = match self
2666 0 : .start_creating_timeline(
2667 0 : new_timeline_id,
2668 0 : CreateTimelineIdempotency::ImportPgdata(CreatingTimelineIdempotencyImportPgdata {
2669 0 : idempotency_key: idempotency_key.clone(),
2670 0 : }),
2671 0 : )
2672 0 : .await?
2673 : {
2674 0 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2675 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
2676 0 : return Ok(CreateTimelineResult::Idempotent(timeline))
2677 : }
2678 : };
2679 :
2680 0 : let mut uninit_timeline = {
2681 0 : let this = &self;
2682 0 : let initdb_lsn = Lsn(0);
2683 0 : let _ctx = ctx;
2684 0 : async move {
2685 0 : let new_metadata = TimelineMetadata::new(
2686 0 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2687 0 : // make it valid, before calling finish_creation()
2688 0 : Lsn(0),
2689 0 : None,
2690 0 : None,
2691 0 : Lsn(0),
2692 0 : initdb_lsn,
2693 0 : initdb_lsn,
2694 0 : 15,
2695 0 : );
2696 0 : this.prepare_new_timeline(
2697 0 : new_timeline_id,
2698 0 : &new_metadata,
2699 0 : timeline_create_guard,
2700 0 : initdb_lsn,
2701 0 : None,
2702 0 : )
2703 0 : .await
2704 0 : }
2705 0 : }
2706 0 : .await?;
2707 :
2708 0 : let in_progress = import_pgdata::index_part_format::InProgress {
2709 0 : idempotency_key,
2710 0 : location,
2711 0 : started_at,
2712 0 : };
2713 0 : let index_part = import_pgdata::index_part_format::Root::V1(
2714 0 : import_pgdata::index_part_format::V1::InProgress(in_progress),
2715 0 : );
2716 0 : uninit_timeline
2717 0 : .raw_timeline()
2718 0 : .unwrap()
2719 0 : .remote_client
2720 0 : .schedule_index_upload_for_import_pgdata_state_update(Some(index_part.clone()))?;
2721 :
2722 : // wait_completion happens in caller
2723 :
2724 0 : let (timeline, timeline_create_guard) = uninit_timeline.finish_creation_myself();
2725 0 :
2726 0 : tokio::spawn(self.clone().create_timeline_import_pgdata_task(
2727 0 : timeline.clone(),
2728 0 : index_part,
2729 0 : activate,
2730 0 : timeline_create_guard,
2731 0 : ));
2732 0 :
2733 0 : // NB: the timeline doesn't exist in self.timelines at this point
2734 0 : Ok(CreateTimelineResult::ImportSpawned(timeline))
2735 0 : }
2736 :
2737 : #[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))]
2738 : async fn create_timeline_import_pgdata_task(
2739 : self: Arc<Tenant>,
2740 : timeline: Arc<Timeline>,
2741 : index_part: import_pgdata::index_part_format::Root,
2742 : activate: ActivateTimelineArgs,
2743 : timeline_create_guard: TimelineCreateGuard,
2744 : ) {
2745 : debug_assert_current_span_has_tenant_and_timeline_id();
2746 : info!("starting");
2747 : scopeguard::defer! {info!("exiting")};
2748 :
2749 : let res = self
2750 : .create_timeline_import_pgdata_task_impl(
2751 : timeline,
2752 : index_part,
2753 : activate,
2754 : timeline_create_guard,
2755 : )
2756 : .await;
2757 : if let Err(err) = &res {
2758 : error!(?err, "task failed");
2759 : // TODO sleep & retry, sensitive to tenant shutdown
2760 : // TODO: allow timeline deletion requests => should cancel the task
2761 : }
2762 : }
2763 :
2764 0 : async fn create_timeline_import_pgdata_task_impl(
2765 0 : self: Arc<Tenant>,
2766 0 : timeline: Arc<Timeline>,
2767 0 : index_part: import_pgdata::index_part_format::Root,
2768 0 : activate: ActivateTimelineArgs,
2769 0 : timeline_create_guard: TimelineCreateGuard,
2770 0 : ) -> Result<(), anyhow::Error> {
2771 0 : let ctx = RequestContext::new(TaskKind::ImportPgdata, DownloadBehavior::Warn);
2772 0 :
2773 0 : info!("importing pgdata");
2774 0 : import_pgdata::doit(&timeline, index_part, &ctx, self.cancel.clone())
2775 0 : .await
2776 0 : .context("import")?;
2777 0 : info!("import done");
2778 :
2779 : //
2780 : // Reload timeline from remote.
2781 : // This proves that the remote state is attachable, and it reuses the code.
2782 : //
2783 : // TODO: think about whether this is safe to do with concurrent Tenant::shutdown.
2784 : // timeline_create_guard hols the tenant gate open, so, shutdown cannot _complete_ until we exit.
2785 : // But our activate() call might launch new background tasks after Tenant::shutdown
2786 : // already went past shutting down the Tenant::timelines, which this timeline here is no part of.
2787 : // I think the same problem exists with the bootstrap & branch mgmt API tasks (tenant shutting
2788 : // down while bootstrapping/branching + activating), but, the race condition is much more likely
2789 : // to manifest because of the long runtime of this import task.
2790 :
2791 : // in theory this shouldn't even .await anything except for coop yield
2792 0 : info!("shutting down timeline");
2793 0 : timeline.shutdown(ShutdownMode::Hard).await;
2794 0 : info!("timeline shut down, reloading from remote");
2795 : // TODO: we can't do the following check because create_timeline_import_pgdata must return an Arc<Timeline>
2796 : // let Some(timeline) = Arc::into_inner(timeline) else {
2797 : // anyhow::bail!("implementation error: timeline that we shut down was still referenced from somewhere");
2798 : // };
2799 0 : let timeline_id = timeline.timeline_id;
2800 0 :
2801 0 : // load from object storage like Tenant::attach does
2802 0 : let resources = self.build_timeline_resources(timeline_id);
2803 0 : let index_part = resources
2804 0 : .remote_client
2805 0 : .download_index_file(&self.cancel)
2806 0 : .await?;
2807 0 : let index_part = match index_part {
2808 : MaybeDeletedIndexPart::Deleted(_) => {
2809 : // likely concurrent delete call, cplane should prevent this
2810 0 : anyhow::bail!("index part says deleted but we are not done creating yet, this should not happen but")
2811 : }
2812 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
2813 0 : };
2814 0 : let metadata = index_part.metadata.clone();
2815 0 : self
2816 0 : .load_remote_timeline(timeline_id, index_part, metadata, resources, LoadTimelineCause::ImportPgdata{
2817 0 : create_guard: timeline_create_guard, activate, }, &ctx)
2818 0 : .await?
2819 0 : .ready_to_activate()
2820 0 : .context("implementation error: reloaded timeline still needs import after import reported success")?;
2821 :
2822 0 : anyhow::Ok(())
2823 0 : }
2824 :
2825 0 : pub(crate) async fn delete_timeline(
2826 0 : self: Arc<Self>,
2827 0 : timeline_id: TimelineId,
2828 0 : ) -> Result<(), DeleteTimelineError> {
2829 0 : DeleteTimelineFlow::run(&self, timeline_id).await?;
2830 :
2831 0 : Ok(())
2832 0 : }
2833 :
2834 : /// perform one garbage collection iteration, removing old data files from disk.
2835 : /// this function is periodically called by gc task.
2836 : /// also it can be explicitly requested through page server api 'do_gc' command.
2837 : ///
2838 : /// `target_timeline_id` specifies the timeline to GC, or None for all.
2839 : ///
2840 : /// The `horizon` an `pitr` parameters determine how much WAL history needs to be retained.
2841 : /// Also known as the retention period, or the GC cutoff point. `horizon` specifies
2842 : /// the amount of history, as LSN difference from current latest LSN on each timeline.
2843 : /// `pitr` specifies the same as a time difference from the current time. The effective
2844 : /// GC cutoff point is determined conservatively by either `horizon` and `pitr`, whichever
2845 : /// requires more history to be retained.
2846 : //
2847 754 : pub(crate) async fn gc_iteration(
2848 754 : &self,
2849 754 : target_timeline_id: Option<TimelineId>,
2850 754 : horizon: u64,
2851 754 : pitr: Duration,
2852 754 : cancel: &CancellationToken,
2853 754 : ctx: &RequestContext,
2854 754 : ) -> Result<GcResult, GcError> {
2855 754 : // Don't start doing work during shutdown
2856 754 : if let TenantState::Stopping { .. } = self.current_state() {
2857 0 : return Ok(GcResult::default());
2858 754 : }
2859 754 :
2860 754 : // there is a global allowed_error for this
2861 754 : if !self.is_active() {
2862 0 : return Err(GcError::NotActive);
2863 754 : }
2864 754 :
2865 754 : {
2866 754 : let conf = self.tenant_conf.load();
2867 754 :
2868 754 : // If we may not delete layers, then simply skip GC. Even though a tenant
2869 754 : // in AttachedMulti state could do GC and just enqueue the blocked deletions,
2870 754 : // the only advantage to doing it is to perhaps shrink the LayerMap metadata
2871 754 : // a bit sooner than we would achieve by waiting for AttachedSingle status.
2872 754 : if !conf.location.may_delete_layers_hint() {
2873 0 : info!("Skipping GC in location state {:?}", conf.location);
2874 0 : return Ok(GcResult::default());
2875 754 : }
2876 754 :
2877 754 : if conf.is_gc_blocked_by_lsn_lease_deadline() {
2878 750 : info!("Skipping GC because lsn lease deadline is not reached");
2879 750 : return Ok(GcResult::default());
2880 4 : }
2881 : }
2882 :
2883 4 : let _guard = match self.gc_block.start().await {
2884 4 : Ok(guard) => guard,
2885 0 : Err(reasons) => {
2886 0 : info!("Skipping GC: {reasons}");
2887 0 : return Ok(GcResult::default());
2888 : }
2889 : };
2890 :
2891 4 : self.gc_iteration_internal(target_timeline_id, horizon, pitr, cancel, ctx)
2892 4 : .await
2893 754 : }
2894 :
2895 : /// Perform one compaction iteration.
2896 : /// This function is periodically called by compactor task.
2897 : /// Also it can be explicitly requested per timeline through page server
2898 : /// api's 'compact' command.
2899 : ///
2900 : /// Returns whether we have pending compaction task.
2901 0 : async fn compaction_iteration(
2902 0 : self: &Arc<Self>,
2903 0 : cancel: &CancellationToken,
2904 0 : ctx: &RequestContext,
2905 0 : ) -> Result<bool, timeline::CompactionError> {
2906 0 : // Don't start doing work during shutdown, or when broken, we do not need those in the logs
2907 0 : if !self.is_active() {
2908 0 : return Ok(false);
2909 0 : }
2910 0 :
2911 0 : {
2912 0 : let conf = self.tenant_conf.load();
2913 0 :
2914 0 : // Note that compaction usually requires deletions, but we don't respect
2915 0 : // may_delete_layers_hint here: that is because tenants in AttachedMulti
2916 0 : // should proceed with compaction even if they can't do deletion, to avoid
2917 0 : // accumulating dangerously deep stacks of L0 layers. Deletions will be
2918 0 : // enqueued inside RemoteTimelineClient, and executed layer if/when we transition
2919 0 : // to AttachedSingle state.
2920 0 : if !conf.location.may_upload_layers_hint() {
2921 0 : info!("Skipping compaction in location state {:?}", conf.location);
2922 0 : return Ok(false);
2923 0 : }
2924 0 : }
2925 0 :
2926 0 : // Scan through the hashmap and collect a list of all the timelines,
2927 0 : // while holding the lock. Then drop the lock and actually perform the
2928 0 : // compactions. We don't want to block everything else while the
2929 0 : // compaction runs.
2930 0 : let timelines_to_compact_or_offload;
2931 0 : {
2932 0 : let timelines = self.timelines.lock().unwrap();
2933 0 : timelines_to_compact_or_offload = timelines
2934 0 : .iter()
2935 0 : .filter_map(|(timeline_id, timeline)| {
2936 0 : let (is_active, (can_offload, _)) =
2937 0 : (timeline.is_active(), timeline.can_offload());
2938 0 : let has_no_unoffloaded_children = {
2939 0 : !timelines
2940 0 : .iter()
2941 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(*timeline_id))
2942 : };
2943 0 : let config_allows_offload = self.conf.timeline_offloading
2944 0 : || self
2945 0 : .tenant_conf
2946 0 : .load()
2947 0 : .tenant_conf
2948 0 : .timeline_offloading
2949 0 : .unwrap_or_default();
2950 0 : let can_offload =
2951 0 : can_offload && has_no_unoffloaded_children && config_allows_offload;
2952 0 : if (is_active, can_offload) == (false, false) {
2953 0 : None
2954 : } else {
2955 0 : Some((*timeline_id, timeline.clone(), (is_active, can_offload)))
2956 : }
2957 0 : })
2958 0 : .collect::<Vec<_>>();
2959 0 : drop(timelines);
2960 0 : }
2961 0 :
2962 0 : // Before doing any I/O work, check our circuit breaker
2963 0 : if self.compaction_circuit_breaker.lock().unwrap().is_broken() {
2964 0 : info!("Skipping compaction due to previous failures");
2965 0 : return Ok(false);
2966 0 : }
2967 0 :
2968 0 : let mut has_pending_task = false;
2969 :
2970 0 : for (timeline_id, timeline, (can_compact, can_offload)) in &timelines_to_compact_or_offload
2971 : {
2972 : // pending_task_left == None: cannot compact, maybe still pending tasks
2973 : // pending_task_left == Some(true): compaction task left
2974 : // pending_task_left == Some(false): no compaction task left
2975 0 : let pending_task_left = if *can_compact {
2976 0 : let has_pending_l0_compaction_task = timeline
2977 0 : .compact(cancel, EnumSet::empty(), ctx)
2978 0 : .instrument(info_span!("compact_timeline", %timeline_id))
2979 0 : .await
2980 0 : .inspect_err(|e| match e {
2981 0 : timeline::CompactionError::ShuttingDown => (),
2982 0 : timeline::CompactionError::Offload(_) => {
2983 0 : // Failures to offload timelines do not trip the circuit breaker, because
2984 0 : // they do not do lots of writes the way compaction itself does: it is cheap
2985 0 : // to retry, and it would be bad to stop all compaction because of an issue with offloading.
2986 0 : }
2987 0 : timeline::CompactionError::Other(e) => {
2988 0 : self.compaction_circuit_breaker
2989 0 : .lock()
2990 0 : .unwrap()
2991 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, e);
2992 0 : }
2993 0 : })?;
2994 0 : if has_pending_l0_compaction_task {
2995 0 : Some(true)
2996 : } else {
2997 0 : let queue = {
2998 0 : let guard = self.scheduled_compaction_tasks.lock().unwrap();
2999 0 : guard.get(timeline_id).cloned()
3000 : };
3001 0 : if let Some(queue) = queue {
3002 0 : let has_pending_tasks = queue
3003 0 : .iteration(cancel, ctx, &self.gc_block, timeline)
3004 0 : .await?;
3005 0 : Some(has_pending_tasks)
3006 : } else {
3007 0 : Some(false)
3008 : }
3009 : }
3010 : } else {
3011 0 : None
3012 : };
3013 0 : has_pending_task |= pending_task_left.unwrap_or(false);
3014 0 : if pending_task_left == Some(false) && *can_offload {
3015 0 : pausable_failpoint!("before-timeline-auto-offload");
3016 0 : match offload_timeline(self, timeline)
3017 0 : .instrument(info_span!("offload_timeline", %timeline_id))
3018 0 : .await
3019 : {
3020 : Err(OffloadError::NotArchived) => {
3021 : // Ignore this, we likely raced with unarchival
3022 0 : Ok(())
3023 : }
3024 0 : other => other,
3025 0 : }?;
3026 0 : }
3027 : }
3028 :
3029 0 : self.compaction_circuit_breaker
3030 0 : .lock()
3031 0 : .unwrap()
3032 0 : .success(&CIRCUIT_BREAKERS_UNBROKEN);
3033 0 :
3034 0 : Ok(has_pending_task)
3035 0 : }
3036 :
3037 : /// Cancel scheduled compaction tasks
3038 0 : pub(crate) fn cancel_scheduled_compaction(&self, timeline_id: TimelineId) {
3039 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3040 0 : if let Some(q) = guard.get_mut(&timeline_id) {
3041 0 : q.cancel_scheduled();
3042 0 : }
3043 0 : }
3044 :
3045 0 : pub(crate) fn get_scheduled_compaction_tasks(
3046 0 : &self,
3047 0 : timeline_id: TimelineId,
3048 0 : ) -> Vec<CompactInfoResponse> {
3049 0 : let res = {
3050 0 : let guard = self.scheduled_compaction_tasks.lock().unwrap();
3051 0 : guard.get(&timeline_id).map(|q| q.remaining_jobs())
3052 : };
3053 0 : let Some((running, remaining)) = res else {
3054 0 : return Vec::new();
3055 : };
3056 0 : let mut result = Vec::new();
3057 0 : if let Some((id, running)) = running {
3058 0 : result.extend(running.into_compact_info_resp(id, true));
3059 0 : }
3060 0 : for (id, job) in remaining {
3061 0 : result.extend(job.into_compact_info_resp(id, false));
3062 0 : }
3063 0 : result
3064 0 : }
3065 :
3066 : /// Schedule a compaction task for a timeline.
3067 0 : pub(crate) async fn schedule_compaction(
3068 0 : &self,
3069 0 : timeline_id: TimelineId,
3070 0 : options: CompactOptions,
3071 0 : ) -> anyhow::Result<tokio::sync::oneshot::Receiver<()>> {
3072 0 : let (tx, rx) = tokio::sync::oneshot::channel();
3073 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3074 0 : let q = guard
3075 0 : .entry(timeline_id)
3076 0 : .or_insert_with(|| Arc::new(GcCompactionQueue::new()));
3077 0 : q.schedule_manual_compaction(options, Some(tx));
3078 0 : Ok(rx)
3079 0 : }
3080 :
3081 : // Call through to all timelines to freeze ephemeral layers if needed. Usually
3082 : // this happens during ingest: this background housekeeping is for freezing layers
3083 : // that are open but haven't been written to for some time.
3084 0 : async fn ingest_housekeeping(&self) {
3085 0 : // Scan through the hashmap and collect a list of all the timelines,
3086 0 : // while holding the lock. Then drop the lock and actually perform the
3087 0 : // compactions. We don't want to block everything else while the
3088 0 : // compaction runs.
3089 0 : let timelines = {
3090 0 : self.timelines
3091 0 : .lock()
3092 0 : .unwrap()
3093 0 : .values()
3094 0 : .filter_map(|timeline| {
3095 0 : if timeline.is_active() {
3096 0 : Some(timeline.clone())
3097 : } else {
3098 0 : None
3099 : }
3100 0 : })
3101 0 : .collect::<Vec<_>>()
3102 : };
3103 :
3104 0 : for timeline in &timelines {
3105 0 : timeline.maybe_freeze_ephemeral_layer().await;
3106 : }
3107 0 : }
3108 :
3109 0 : pub fn timeline_has_no_attached_children(&self, timeline_id: TimelineId) -> bool {
3110 0 : let timelines = self.timelines.lock().unwrap();
3111 0 : !timelines
3112 0 : .iter()
3113 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(timeline_id))
3114 0 : }
3115 :
3116 1732 : pub fn current_state(&self) -> TenantState {
3117 1732 : self.state.borrow().clone()
3118 1732 : }
3119 :
3120 970 : pub fn is_active(&self) -> bool {
3121 970 : self.current_state() == TenantState::Active
3122 970 : }
3123 :
3124 0 : pub fn generation(&self) -> Generation {
3125 0 : self.generation
3126 0 : }
3127 :
3128 0 : pub(crate) fn wal_redo_manager_status(&self) -> Option<WalRedoManagerStatus> {
3129 0 : self.walredo_mgr.as_ref().and_then(|mgr| mgr.status())
3130 0 : }
3131 :
3132 : /// Changes tenant status to active, unless shutdown was already requested.
3133 : ///
3134 : /// `background_jobs_can_start` is an optional barrier set to a value during pageserver startup
3135 : /// to delay background jobs. Background jobs can be started right away when None is given.
3136 0 : fn activate(
3137 0 : self: &Arc<Self>,
3138 0 : broker_client: BrokerClientChannel,
3139 0 : background_jobs_can_start: Option<&completion::Barrier>,
3140 0 : ctx: &RequestContext,
3141 0 : ) {
3142 0 : span::debug_assert_current_span_has_tenant_id();
3143 0 :
3144 0 : let mut activating = false;
3145 0 : self.state.send_modify(|current_state| {
3146 : use pageserver_api::models::ActivatingFrom;
3147 0 : match &*current_state {
3148 : TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
3149 0 : panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {state:?}", state = current_state);
3150 : }
3151 0 : TenantState::Attaching => {
3152 0 : *current_state = TenantState::Activating(ActivatingFrom::Attaching);
3153 0 : }
3154 0 : }
3155 0 : debug!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), "Activating tenant");
3156 0 : activating = true;
3157 0 : // Continue outside the closure. We need to grab timelines.lock()
3158 0 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
3159 0 : });
3160 0 :
3161 0 : if activating {
3162 0 : let timelines_accessor = self.timelines.lock().unwrap();
3163 0 : let timelines_offloaded_accessor = self.timelines_offloaded.lock().unwrap();
3164 0 : let timelines_to_activate = timelines_accessor
3165 0 : .values()
3166 0 : .filter(|timeline| !(timeline.is_broken() || timeline.is_stopping()));
3167 0 :
3168 0 : // Before activation, populate each Timeline's GcInfo with information about its children
3169 0 : self.initialize_gc_info(&timelines_accessor, &timelines_offloaded_accessor, None);
3170 0 :
3171 0 : // Spawn gc and compaction loops. The loops will shut themselves
3172 0 : // down when they notice that the tenant is inactive.
3173 0 : tasks::start_background_loops(self, background_jobs_can_start);
3174 0 :
3175 0 : let mut activated_timelines = 0;
3176 :
3177 0 : for timeline in timelines_to_activate {
3178 0 : timeline.activate(
3179 0 : self.clone(),
3180 0 : broker_client.clone(),
3181 0 : background_jobs_can_start,
3182 0 : ctx,
3183 0 : );
3184 0 : activated_timelines += 1;
3185 0 : }
3186 :
3187 0 : self.state.send_modify(move |current_state| {
3188 0 : assert!(
3189 0 : matches!(current_state, TenantState::Activating(_)),
3190 0 : "set_stopping and set_broken wait for us to leave Activating state",
3191 : );
3192 0 : *current_state = TenantState::Active;
3193 0 :
3194 0 : let elapsed = self.constructed_at.elapsed();
3195 0 : let total_timelines = timelines_accessor.len();
3196 0 :
3197 0 : // log a lot of stuff, because some tenants sometimes suffer from user-visible
3198 0 : // times to activate. see https://github.com/neondatabase/neon/issues/4025
3199 0 : info!(
3200 0 : since_creation_millis = elapsed.as_millis(),
3201 0 : tenant_id = %self.tenant_shard_id.tenant_id,
3202 0 : shard_id = %self.tenant_shard_id.shard_slug(),
3203 0 : activated_timelines,
3204 0 : total_timelines,
3205 0 : post_state = <&'static str>::from(&*current_state),
3206 0 : "activation attempt finished"
3207 : );
3208 :
3209 0 : TENANT.activation.observe(elapsed.as_secs_f64());
3210 0 : });
3211 0 : }
3212 0 : }
3213 :
3214 : /// Shutdown the tenant and join all of the spawned tasks.
3215 : ///
3216 : /// The method caters for all use-cases:
3217 : /// - pageserver shutdown (freeze_and_flush == true)
3218 : /// - detach + ignore (freeze_and_flush == false)
3219 : ///
3220 : /// This will attempt to shutdown even if tenant is broken.
3221 : ///
3222 : /// `shutdown_progress` is a [`completion::Barrier`] for the shutdown initiated by this call.
3223 : /// If the tenant is already shutting down, we return a clone of the first shutdown call's
3224 : /// `Barrier` as an `Err`. This not-first caller can use the returned barrier to join with
3225 : /// the ongoing shutdown.
3226 6 : async fn shutdown(
3227 6 : &self,
3228 6 : shutdown_progress: completion::Barrier,
3229 6 : shutdown_mode: timeline::ShutdownMode,
3230 6 : ) -> Result<(), completion::Barrier> {
3231 6 : span::debug_assert_current_span_has_tenant_id();
3232 :
3233 : // Set tenant (and its timlines) to Stoppping state.
3234 : //
3235 : // Since we can only transition into Stopping state after activation is complete,
3236 : // run it in a JoinSet so all tenants have a chance to stop before we get SIGKILLed.
3237 : //
3238 : // Transitioning tenants to Stopping state has a couple of non-obvious side effects:
3239 : // 1. Lock out any new requests to the tenants.
3240 : // 2. Signal cancellation to WAL receivers (we wait on it below).
3241 : // 3. Signal cancellation for other tenant background loops.
3242 : // 4. ???
3243 : //
3244 : // The waiting for the cancellation is not done uniformly.
3245 : // We certainly wait for WAL receivers to shut down.
3246 : // That is necessary so that no new data comes in before the freeze_and_flush.
3247 : // But the tenant background loops are joined-on in our caller.
3248 : // It's mesed up.
3249 : // we just ignore the failure to stop
3250 :
3251 : // If we're still attaching, fire the cancellation token early to drop out: this
3252 : // will prevent us flushing, but ensures timely shutdown if some I/O during attach
3253 : // is very slow.
3254 6 : let shutdown_mode = if matches!(self.current_state(), TenantState::Attaching) {
3255 0 : self.cancel.cancel();
3256 0 :
3257 0 : // Having fired our cancellation token, do not try and flush timelines: their cancellation tokens
3258 0 : // are children of ours, so their flush loops will have shut down already
3259 0 : timeline::ShutdownMode::Hard
3260 : } else {
3261 6 : shutdown_mode
3262 : };
3263 :
3264 6 : match self.set_stopping(shutdown_progress, false, false).await {
3265 6 : Ok(()) => {}
3266 0 : Err(SetStoppingError::Broken) => {
3267 0 : // assume that this is acceptable
3268 0 : }
3269 0 : Err(SetStoppingError::AlreadyStopping(other)) => {
3270 0 : // give caller the option to wait for this this shutdown
3271 0 : info!("Tenant::shutdown: AlreadyStopping");
3272 0 : return Err(other);
3273 : }
3274 : };
3275 :
3276 6 : let mut js = tokio::task::JoinSet::new();
3277 6 : {
3278 6 : let timelines = self.timelines.lock().unwrap();
3279 6 : timelines.values().for_each(|timeline| {
3280 6 : let timeline = Arc::clone(timeline);
3281 6 : let timeline_id = timeline.timeline_id;
3282 6 : let span = tracing::info_span!("timeline_shutdown", %timeline_id, ?shutdown_mode);
3283 6 : js.spawn(async move { timeline.shutdown(shutdown_mode).instrument(span).await });
3284 6 : });
3285 6 : }
3286 6 : {
3287 6 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
3288 6 : timelines_offloaded.values().for_each(|timeline| {
3289 0 : timeline.defuse_for_tenant_drop();
3290 6 : });
3291 6 : }
3292 6 : // test_long_timeline_create_then_tenant_delete is leaning on this message
3293 6 : tracing::info!("Waiting for timelines...");
3294 12 : while let Some(res) = js.join_next().await {
3295 0 : match res {
3296 6 : Ok(()) => {}
3297 0 : Err(je) if je.is_cancelled() => unreachable!("no cancelling used"),
3298 0 : Err(je) if je.is_panic() => { /* logged already */ }
3299 0 : Err(je) => warn!("unexpected JoinError: {je:?}"),
3300 : }
3301 : }
3302 :
3303 6 : if let ShutdownMode::Reload = shutdown_mode {
3304 0 : tracing::info!("Flushing deletion queue");
3305 0 : if let Err(e) = self.deletion_queue_client.flush().await {
3306 0 : match e {
3307 0 : DeletionQueueError::ShuttingDown => {
3308 0 : // This is the only error we expect for now. In the future, if more error
3309 0 : // variants are added, we should handle them here.
3310 0 : }
3311 : }
3312 0 : }
3313 6 : }
3314 :
3315 : // We cancel the Tenant's cancellation token _after_ the timelines have all shut down. This permits
3316 : // them to continue to do work during their shutdown methods, e.g. flushing data.
3317 6 : tracing::debug!("Cancelling CancellationToken");
3318 6 : self.cancel.cancel();
3319 6 :
3320 6 : // shutdown all tenant and timeline tasks: gc, compaction, page service
3321 6 : // No new tasks will be started for this tenant because it's in `Stopping` state.
3322 6 : //
3323 6 : // this will additionally shutdown and await all timeline tasks.
3324 6 : tracing::debug!("Waiting for tasks...");
3325 6 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), None).await;
3326 :
3327 6 : if let Some(walredo_mgr) = self.walredo_mgr.as_ref() {
3328 6 : walredo_mgr.shutdown().await;
3329 0 : }
3330 :
3331 : // Wait for any in-flight operations to complete
3332 6 : self.gate.close().await;
3333 :
3334 6 : remove_tenant_metrics(&self.tenant_shard_id);
3335 6 :
3336 6 : Ok(())
3337 6 : }
3338 :
3339 : /// Change tenant status to Stopping, to mark that it is being shut down.
3340 : ///
3341 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
3342 : ///
3343 : /// This function is not cancel-safe!
3344 : ///
3345 : /// `allow_transition_from_loading` is needed for the special case of loading task deleting the tenant.
3346 : /// `allow_transition_from_attaching` is needed for the special case of attaching deleted tenant.
3347 6 : async fn set_stopping(
3348 6 : &self,
3349 6 : progress: completion::Barrier,
3350 6 : _allow_transition_from_loading: bool,
3351 6 : allow_transition_from_attaching: bool,
3352 6 : ) -> Result<(), SetStoppingError> {
3353 6 : let mut rx = self.state.subscribe();
3354 6 :
3355 6 : // cannot stop before we're done activating, so wait out until we're done activating
3356 6 : rx.wait_for(|state| match state {
3357 0 : TenantState::Attaching if allow_transition_from_attaching => true,
3358 : TenantState::Activating(_) | TenantState::Attaching => {
3359 0 : info!(
3360 0 : "waiting for {} to turn Active|Broken|Stopping",
3361 0 : <&'static str>::from(state)
3362 : );
3363 0 : false
3364 : }
3365 6 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3366 6 : })
3367 6 : .await
3368 6 : .expect("cannot drop self.state while on a &self method");
3369 6 :
3370 6 : // we now know we're done activating, let's see whether this task is the winner to transition into Stopping
3371 6 : let mut err = None;
3372 6 : let stopping = self.state.send_if_modified(|current_state| match current_state {
3373 : TenantState::Activating(_) => {
3374 0 : unreachable!("1we ensured above that we're done with activation, and, there is no re-activation")
3375 : }
3376 : TenantState::Attaching => {
3377 0 : if !allow_transition_from_attaching {
3378 0 : unreachable!("2we ensured above that we're done with activation, and, there is no re-activation")
3379 0 : };
3380 0 : *current_state = TenantState::Stopping { progress };
3381 0 : true
3382 : }
3383 : TenantState::Active => {
3384 : // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
3385 : // are created after the transition to Stopping. That's harmless, as the Timelines
3386 : // won't be accessible to anyone afterwards, because the Tenant is in Stopping state.
3387 6 : *current_state = TenantState::Stopping { progress };
3388 6 : // Continue stopping outside the closure. We need to grab timelines.lock()
3389 6 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
3390 6 : true
3391 : }
3392 0 : TenantState::Broken { reason, .. } => {
3393 0 : info!(
3394 0 : "Cannot set tenant to Stopping state, it is in Broken state due to: {reason}"
3395 : );
3396 0 : err = Some(SetStoppingError::Broken);
3397 0 : false
3398 : }
3399 0 : TenantState::Stopping { progress } => {
3400 0 : info!("Tenant is already in Stopping state");
3401 0 : err = Some(SetStoppingError::AlreadyStopping(progress.clone()));
3402 0 : false
3403 : }
3404 6 : });
3405 6 : match (stopping, err) {
3406 6 : (true, None) => {} // continue
3407 0 : (false, Some(err)) => return Err(err),
3408 0 : (true, Some(_)) => unreachable!(
3409 0 : "send_if_modified closure must error out if not transitioning to Stopping"
3410 0 : ),
3411 0 : (false, None) => unreachable!(
3412 0 : "send_if_modified closure must return true if transitioning to Stopping"
3413 0 : ),
3414 : }
3415 :
3416 6 : let timelines_accessor = self.timelines.lock().unwrap();
3417 6 : let not_broken_timelines = timelines_accessor
3418 6 : .values()
3419 6 : .filter(|timeline| !timeline.is_broken());
3420 12 : for timeline in not_broken_timelines {
3421 6 : timeline.set_state(TimelineState::Stopping);
3422 6 : }
3423 6 : Ok(())
3424 6 : }
3425 :
3426 : /// Method for tenant::mgr to transition us into Broken state in case of a late failure in
3427 : /// `remove_tenant_from_memory`
3428 : ///
3429 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
3430 : ///
3431 : /// In tests, we also use this to set tenants to Broken state on purpose.
3432 0 : pub(crate) async fn set_broken(&self, reason: String) {
3433 0 : let mut rx = self.state.subscribe();
3434 0 :
3435 0 : // The load & attach routines own the tenant state until it has reached `Active`.
3436 0 : // So, wait until it's done.
3437 0 : rx.wait_for(|state| match state {
3438 : TenantState::Activating(_) | TenantState::Attaching => {
3439 0 : info!(
3440 0 : "waiting for {} to turn Active|Broken|Stopping",
3441 0 : <&'static str>::from(state)
3442 : );
3443 0 : false
3444 : }
3445 0 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3446 0 : })
3447 0 : .await
3448 0 : .expect("cannot drop self.state while on a &self method");
3449 0 :
3450 0 : // we now know we're done activating, let's see whether this task is the winner to transition into Broken
3451 0 : self.set_broken_no_wait(reason)
3452 0 : }
3453 :
3454 0 : pub(crate) fn set_broken_no_wait(&self, reason: impl Display) {
3455 0 : let reason = reason.to_string();
3456 0 : self.state.send_modify(|current_state| {
3457 0 : match *current_state {
3458 : TenantState::Activating(_) | TenantState::Attaching => {
3459 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
3460 : }
3461 : TenantState::Active => {
3462 0 : if cfg!(feature = "testing") {
3463 0 : warn!("Changing Active tenant to Broken state, reason: {}", reason);
3464 0 : *current_state = TenantState::broken_from_reason(reason);
3465 : } else {
3466 0 : unreachable!("not allowed to call set_broken on Active tenants in non-testing builds")
3467 : }
3468 : }
3469 : TenantState::Broken { .. } => {
3470 0 : warn!("Tenant is already in Broken state");
3471 : }
3472 : // This is the only "expected" path, any other path is a bug.
3473 : TenantState::Stopping { .. } => {
3474 0 : warn!(
3475 0 : "Marking Stopping tenant as Broken state, reason: {}",
3476 : reason
3477 : );
3478 0 : *current_state = TenantState::broken_from_reason(reason);
3479 : }
3480 : }
3481 0 : });
3482 0 : }
3483 :
3484 0 : pub fn subscribe_for_state_updates(&self) -> watch::Receiver<TenantState> {
3485 0 : self.state.subscribe()
3486 0 : }
3487 :
3488 : /// The activate_now semaphore is initialized with zero units. As soon as
3489 : /// we add a unit, waiters will be able to acquire a unit and proceed.
3490 0 : pub(crate) fn activate_now(&self) {
3491 0 : self.activate_now_sem.add_permits(1);
3492 0 : }
3493 :
3494 0 : pub(crate) async fn wait_to_become_active(
3495 0 : &self,
3496 0 : timeout: Duration,
3497 0 : ) -> Result<(), GetActiveTenantError> {
3498 0 : let mut receiver = self.state.subscribe();
3499 : loop {
3500 0 : let current_state = receiver.borrow_and_update().clone();
3501 0 : match current_state {
3502 : TenantState::Attaching | TenantState::Activating(_) => {
3503 : // in these states, there's a chance that we can reach ::Active
3504 0 : self.activate_now();
3505 0 : match timeout_cancellable(timeout, &self.cancel, receiver.changed()).await {
3506 0 : Ok(r) => {
3507 0 : r.map_err(
3508 0 : |_e: tokio::sync::watch::error::RecvError|
3509 : // Tenant existed but was dropped: report it as non-existent
3510 0 : GetActiveTenantError::NotFound(GetTenantError::ShardNotFound(self.tenant_shard_id))
3511 0 : )?
3512 : }
3513 : Err(TimeoutCancellableError::Cancelled) => {
3514 0 : return Err(GetActiveTenantError::Cancelled);
3515 : }
3516 : Err(TimeoutCancellableError::Timeout) => {
3517 0 : return Err(GetActiveTenantError::WaitForActiveTimeout {
3518 0 : latest_state: Some(self.current_state()),
3519 0 : wait_time: timeout,
3520 0 : });
3521 : }
3522 : }
3523 : }
3524 : TenantState::Active { .. } => {
3525 0 : return Ok(());
3526 : }
3527 0 : TenantState::Broken { reason, .. } => {
3528 0 : // This is fatal, and reported distinctly from the general case of "will never be active" because
3529 0 : // it's logically a 500 to external API users (broken is always a bug).
3530 0 : return Err(GetActiveTenantError::Broken(reason));
3531 : }
3532 : TenantState::Stopping { .. } => {
3533 : // There's no chance the tenant can transition back into ::Active
3534 0 : return Err(GetActiveTenantError::WillNotBecomeActive(current_state));
3535 : }
3536 : }
3537 : }
3538 0 : }
3539 :
3540 0 : pub(crate) fn get_attach_mode(&self) -> AttachmentMode {
3541 0 : self.tenant_conf.load().location.attach_mode
3542 0 : }
3543 :
3544 : /// For API access: generate a LocationConfig equivalent to the one that would be used to
3545 : /// create a Tenant in the same state. Do not use this in hot paths: it's for relatively
3546 : /// rare external API calls, like a reconciliation at startup.
3547 0 : pub(crate) fn get_location_conf(&self) -> models::LocationConfig {
3548 0 : let conf = self.tenant_conf.load();
3549 :
3550 0 : let location_config_mode = match conf.location.attach_mode {
3551 0 : AttachmentMode::Single => models::LocationConfigMode::AttachedSingle,
3552 0 : AttachmentMode::Multi => models::LocationConfigMode::AttachedMulti,
3553 0 : AttachmentMode::Stale => models::LocationConfigMode::AttachedStale,
3554 : };
3555 :
3556 : // We have a pageserver TenantConf, we need the API-facing TenantConfig.
3557 0 : let tenant_config: models::TenantConfig = conf.tenant_conf.clone().into();
3558 0 :
3559 0 : models::LocationConfig {
3560 0 : mode: location_config_mode,
3561 0 : generation: self.generation.into(),
3562 0 : secondary_conf: None,
3563 0 : shard_number: self.shard_identity.number.0,
3564 0 : shard_count: self.shard_identity.count.literal(),
3565 0 : shard_stripe_size: self.shard_identity.stripe_size.0,
3566 0 : tenant_conf: tenant_config,
3567 0 : }
3568 0 : }
3569 :
3570 0 : pub(crate) fn get_tenant_shard_id(&self) -> &TenantShardId {
3571 0 : &self.tenant_shard_id
3572 0 : }
3573 :
3574 0 : pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
3575 0 : self.shard_identity.stripe_size
3576 0 : }
3577 :
3578 0 : pub(crate) fn get_generation(&self) -> Generation {
3579 0 : self.generation
3580 0 : }
3581 :
3582 : /// This function partially shuts down the tenant (it shuts down the Timelines) and is fallible,
3583 : /// and can leave the tenant in a bad state if it fails. The caller is responsible for
3584 : /// resetting this tenant to a valid state if we fail.
3585 0 : pub(crate) async fn split_prepare(
3586 0 : &self,
3587 0 : child_shards: &Vec<TenantShardId>,
3588 0 : ) -> anyhow::Result<()> {
3589 0 : let (timelines, offloaded) = {
3590 0 : let timelines = self.timelines.lock().unwrap();
3591 0 : let offloaded = self.timelines_offloaded.lock().unwrap();
3592 0 : (timelines.clone(), offloaded.clone())
3593 0 : };
3594 0 : let timelines_iter = timelines
3595 0 : .values()
3596 0 : .map(TimelineOrOffloadedArcRef::<'_>::from)
3597 0 : .chain(
3598 0 : offloaded
3599 0 : .values()
3600 0 : .map(TimelineOrOffloadedArcRef::<'_>::from),
3601 0 : );
3602 0 : for timeline in timelines_iter {
3603 : // We do not block timeline creation/deletion during splits inside the pageserver: it is up to higher levels
3604 : // to ensure that they do not start a split if currently in the process of doing these.
3605 :
3606 0 : let timeline_id = timeline.timeline_id();
3607 :
3608 0 : if let TimelineOrOffloadedArcRef::Timeline(timeline) = timeline {
3609 : // Upload an index from the parent: this is partly to provide freshness for the
3610 : // child tenants that will copy it, and partly for general ease-of-debugging: there will
3611 : // always be a parent shard index in the same generation as we wrote the child shard index.
3612 0 : tracing::info!(%timeline_id, "Uploading index");
3613 0 : timeline
3614 0 : .remote_client
3615 0 : .schedule_index_upload_for_file_changes()?;
3616 0 : timeline.remote_client.wait_completion().await?;
3617 0 : }
3618 :
3619 0 : let remote_client = match timeline {
3620 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.remote_client.clone(),
3621 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => {
3622 0 : let remote_client = self
3623 0 : .build_timeline_client(offloaded.timeline_id, self.remote_storage.clone());
3624 0 : Arc::new(remote_client)
3625 : }
3626 : };
3627 :
3628 : // Shut down the timeline's remote client: this means that the indices we write
3629 : // for child shards will not be invalidated by the parent shard deleting layers.
3630 0 : tracing::info!(%timeline_id, "Shutting down remote storage client");
3631 0 : remote_client.shutdown().await;
3632 :
3633 : // Download methods can still be used after shutdown, as they don't flow through the remote client's
3634 : // queue. In principal the RemoteTimelineClient could provide this without downloading it, but this
3635 : // operation is rare, so it's simpler to just download it (and robustly guarantees that the index
3636 : // we use here really is the remotely persistent one).
3637 0 : tracing::info!(%timeline_id, "Downloading index_part from parent");
3638 0 : let result = remote_client
3639 0 : .download_index_file(&self.cancel)
3640 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))
3641 0 : .await?;
3642 0 : let index_part = match result {
3643 : MaybeDeletedIndexPart::Deleted(_) => {
3644 0 : anyhow::bail!("Timeline deletion happened concurrently with split")
3645 : }
3646 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
3647 : };
3648 :
3649 0 : for child_shard in child_shards {
3650 0 : tracing::info!(%timeline_id, "Uploading index_part for child {}", child_shard.to_index());
3651 0 : upload_index_part(
3652 0 : &self.remote_storage,
3653 0 : child_shard,
3654 0 : &timeline_id,
3655 0 : self.generation,
3656 0 : &index_part,
3657 0 : &self.cancel,
3658 0 : )
3659 0 : .await?;
3660 : }
3661 : }
3662 :
3663 0 : let tenant_manifest = self.build_tenant_manifest();
3664 0 : for child_shard in child_shards {
3665 0 : tracing::info!(
3666 0 : "Uploading tenant manifest for child {}",
3667 0 : child_shard.to_index()
3668 : );
3669 0 : upload_tenant_manifest(
3670 0 : &self.remote_storage,
3671 0 : child_shard,
3672 0 : self.generation,
3673 0 : &tenant_manifest,
3674 0 : &self.cancel,
3675 0 : )
3676 0 : .await?;
3677 : }
3678 :
3679 0 : Ok(())
3680 0 : }
3681 :
3682 0 : pub(crate) fn get_sizes(&self) -> TopTenantShardItem {
3683 0 : let mut result = TopTenantShardItem {
3684 0 : id: self.tenant_shard_id,
3685 0 : resident_size: 0,
3686 0 : physical_size: 0,
3687 0 : max_logical_size: 0,
3688 0 : };
3689 :
3690 0 : for timeline in self.timelines.lock().unwrap().values() {
3691 0 : result.resident_size += timeline.metrics.resident_physical_size_gauge.get();
3692 0 :
3693 0 : result.physical_size += timeline
3694 0 : .remote_client
3695 0 : .metrics
3696 0 : .remote_physical_size_gauge
3697 0 : .get();
3698 0 : result.max_logical_size = std::cmp::max(
3699 0 : result.max_logical_size,
3700 0 : timeline.metrics.current_logical_size_gauge.get(),
3701 0 : );
3702 0 : }
3703 :
3704 0 : result
3705 0 : }
3706 : }
3707 :
3708 : /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
3709 : /// perform a topological sort, so that the parent of each timeline comes
3710 : /// before the children.
3711 : /// E extracts the ancestor from T
3712 : /// This allows for T to be different. It can be TimelineMetadata, can be Timeline itself, etc.
3713 220 : fn tree_sort_timelines<T, E>(
3714 220 : timelines: HashMap<TimelineId, T>,
3715 220 : extractor: E,
3716 220 : ) -> anyhow::Result<Vec<(TimelineId, T)>>
3717 220 : where
3718 220 : E: Fn(&T) -> Option<TimelineId>,
3719 220 : {
3720 220 : let mut result = Vec::with_capacity(timelines.len());
3721 220 :
3722 220 : let mut now = Vec::with_capacity(timelines.len());
3723 220 : // (ancestor, children)
3724 220 : let mut later: HashMap<TimelineId, Vec<(TimelineId, T)>> =
3725 220 : HashMap::with_capacity(timelines.len());
3726 :
3727 226 : for (timeline_id, value) in timelines {
3728 6 : if let Some(ancestor_id) = extractor(&value) {
3729 2 : let children = later.entry(ancestor_id).or_default();
3730 2 : children.push((timeline_id, value));
3731 4 : } else {
3732 4 : now.push((timeline_id, value));
3733 4 : }
3734 : }
3735 :
3736 226 : while let Some((timeline_id, metadata)) = now.pop() {
3737 6 : result.push((timeline_id, metadata));
3738 : // All children of this can be loaded now
3739 6 : if let Some(mut children) = later.remove(&timeline_id) {
3740 2 : now.append(&mut children);
3741 4 : }
3742 : }
3743 :
3744 : // All timelines should be visited now. Unless there were timelines with missing ancestors.
3745 220 : if !later.is_empty() {
3746 0 : for (missing_id, orphan_ids) in later {
3747 0 : for (orphan_id, _) in orphan_ids {
3748 0 : error!("could not load timeline {orphan_id} because its ancestor timeline {missing_id} could not be loaded");
3749 : }
3750 : }
3751 0 : bail!("could not load tenant because some timelines are missing ancestors");
3752 220 : }
3753 220 :
3754 220 : Ok(result)
3755 220 : }
3756 :
3757 : enum ActivateTimelineArgs {
3758 : Yes {
3759 : broker_client: storage_broker::BrokerClientChannel,
3760 : },
3761 : No,
3762 : }
3763 :
3764 : impl Tenant {
3765 0 : pub fn tenant_specific_overrides(&self) -> TenantConfOpt {
3766 0 : self.tenant_conf.load().tenant_conf.clone()
3767 0 : }
3768 :
3769 0 : pub fn effective_config(&self) -> TenantConf {
3770 0 : self.tenant_specific_overrides()
3771 0 : .merge(self.conf.default_tenant_conf.clone())
3772 0 : }
3773 :
3774 0 : pub fn get_checkpoint_distance(&self) -> u64 {
3775 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3776 0 : tenant_conf
3777 0 : .checkpoint_distance
3778 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
3779 0 : }
3780 :
3781 0 : pub fn get_checkpoint_timeout(&self) -> Duration {
3782 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3783 0 : tenant_conf
3784 0 : .checkpoint_timeout
3785 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
3786 0 : }
3787 :
3788 0 : pub fn get_compaction_target_size(&self) -> u64 {
3789 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3790 0 : tenant_conf
3791 0 : .compaction_target_size
3792 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
3793 0 : }
3794 :
3795 0 : pub fn get_compaction_period(&self) -> Duration {
3796 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3797 0 : tenant_conf
3798 0 : .compaction_period
3799 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
3800 0 : }
3801 :
3802 0 : pub fn get_compaction_threshold(&self) -> usize {
3803 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3804 0 : tenant_conf
3805 0 : .compaction_threshold
3806 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
3807 0 : }
3808 :
3809 0 : pub fn get_gc_horizon(&self) -> u64 {
3810 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3811 0 : tenant_conf
3812 0 : .gc_horizon
3813 0 : .unwrap_or(self.conf.default_tenant_conf.gc_horizon)
3814 0 : }
3815 :
3816 0 : pub fn get_gc_period(&self) -> Duration {
3817 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3818 0 : tenant_conf
3819 0 : .gc_period
3820 0 : .unwrap_or(self.conf.default_tenant_conf.gc_period)
3821 0 : }
3822 :
3823 0 : pub fn get_image_creation_threshold(&self) -> usize {
3824 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3825 0 : tenant_conf
3826 0 : .image_creation_threshold
3827 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
3828 0 : }
3829 :
3830 0 : pub fn get_pitr_interval(&self) -> Duration {
3831 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3832 0 : tenant_conf
3833 0 : .pitr_interval
3834 0 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
3835 0 : }
3836 :
3837 0 : pub fn get_min_resident_size_override(&self) -> Option<u64> {
3838 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3839 0 : tenant_conf
3840 0 : .min_resident_size_override
3841 0 : .or(self.conf.default_tenant_conf.min_resident_size_override)
3842 0 : }
3843 :
3844 0 : pub fn get_heatmap_period(&self) -> Option<Duration> {
3845 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3846 0 : let heatmap_period = tenant_conf
3847 0 : .heatmap_period
3848 0 : .unwrap_or(self.conf.default_tenant_conf.heatmap_period);
3849 0 : if heatmap_period.is_zero() {
3850 0 : None
3851 : } else {
3852 0 : Some(heatmap_period)
3853 : }
3854 0 : }
3855 :
3856 4 : pub fn get_lsn_lease_length(&self) -> Duration {
3857 4 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3858 4 : tenant_conf
3859 4 : .lsn_lease_length
3860 4 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
3861 4 : }
3862 :
3863 : /// Generate an up-to-date TenantManifest based on the state of this Tenant.
3864 2 : fn build_tenant_manifest(&self) -> TenantManifest {
3865 2 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
3866 2 :
3867 2 : let mut timeline_manifests = timelines_offloaded
3868 2 : .iter()
3869 2 : .map(|(_timeline_id, offloaded)| offloaded.manifest())
3870 2 : .collect::<Vec<_>>();
3871 2 : // Sort the manifests so that our output is deterministic
3872 2 : timeline_manifests.sort_by_key(|timeline_manifest| timeline_manifest.timeline_id);
3873 2 :
3874 2 : TenantManifest {
3875 2 : version: LATEST_TENANT_MANIFEST_VERSION,
3876 2 : offloaded_timelines: timeline_manifests,
3877 2 : }
3878 2 : }
3879 :
3880 0 : pub fn update_tenant_config<F: Fn(TenantConfOpt) -> anyhow::Result<TenantConfOpt>>(
3881 0 : &self,
3882 0 : update: F,
3883 0 : ) -> anyhow::Result<TenantConfOpt> {
3884 0 : // Use read-copy-update in order to avoid overwriting the location config
3885 0 : // state if this races with [`Tenant::set_new_location_config`]. Note that
3886 0 : // this race is not possible if both request types come from the storage
3887 0 : // controller (as they should!) because an exclusive op lock is required
3888 0 : // on the storage controller side.
3889 0 :
3890 0 : self.tenant_conf
3891 0 : .try_rcu(|attached_conf| -> Result<_, anyhow::Error> {
3892 0 : Ok(Arc::new(AttachedTenantConf {
3893 0 : tenant_conf: update(attached_conf.tenant_conf.clone())?,
3894 0 : location: attached_conf.location,
3895 0 : lsn_lease_deadline: attached_conf.lsn_lease_deadline,
3896 : }))
3897 0 : })?;
3898 :
3899 0 : let updated = self.tenant_conf.load();
3900 0 :
3901 0 : self.tenant_conf_updated(&updated.tenant_conf);
3902 0 : // Don't hold self.timelines.lock() during the notifies.
3903 0 : // There's no risk of deadlock right now, but there could be if we consolidate
3904 0 : // mutexes in struct Timeline in the future.
3905 0 : let timelines = self.list_timelines();
3906 0 : for timeline in timelines {
3907 0 : timeline.tenant_conf_updated(&updated);
3908 0 : }
3909 :
3910 0 : Ok(updated.tenant_conf.clone())
3911 0 : }
3912 :
3913 0 : pub(crate) fn set_new_location_config(&self, new_conf: AttachedTenantConf) {
3914 0 : let new_tenant_conf = new_conf.tenant_conf.clone();
3915 0 :
3916 0 : self.tenant_conf.store(Arc::new(new_conf.clone()));
3917 0 :
3918 0 : self.tenant_conf_updated(&new_tenant_conf);
3919 0 : // Don't hold self.timelines.lock() during the notifies.
3920 0 : // There's no risk of deadlock right now, but there could be if we consolidate
3921 0 : // mutexes in struct Timeline in the future.
3922 0 : let timelines = self.list_timelines();
3923 0 : for timeline in timelines {
3924 0 : timeline.tenant_conf_updated(&new_conf);
3925 0 : }
3926 0 : }
3927 :
3928 220 : fn get_pagestream_throttle_config(
3929 220 : psconf: &'static PageServerConf,
3930 220 : overrides: &TenantConfOpt,
3931 220 : ) -> throttle::Config {
3932 220 : overrides
3933 220 : .timeline_get_throttle
3934 220 : .clone()
3935 220 : .unwrap_or(psconf.default_tenant_conf.timeline_get_throttle.clone())
3936 220 : }
3937 :
3938 0 : pub(crate) fn tenant_conf_updated(&self, new_conf: &TenantConfOpt) {
3939 0 : let conf = Self::get_pagestream_throttle_config(self.conf, new_conf);
3940 0 : self.pagestream_throttle.reconfigure(conf)
3941 0 : }
3942 :
3943 : /// Helper function to create a new Timeline struct.
3944 : ///
3945 : /// The returned Timeline is in Loading state. The caller is responsible for
3946 : /// initializing any on-disk state, and for inserting the Timeline to the 'timelines'
3947 : /// map.
3948 : ///
3949 : /// `validate_ancestor == false` is used when a timeline is created for deletion
3950 : /// and we might not have the ancestor present anymore which is fine for to be
3951 : /// deleted timelines.
3952 : #[allow(clippy::too_many_arguments)]
3953 446 : fn create_timeline_struct(
3954 446 : &self,
3955 446 : new_timeline_id: TimelineId,
3956 446 : new_metadata: &TimelineMetadata,
3957 446 : ancestor: Option<Arc<Timeline>>,
3958 446 : resources: TimelineResources,
3959 446 : cause: CreateTimelineCause,
3960 446 : create_idempotency: CreateTimelineIdempotency,
3961 446 : ) -> anyhow::Result<Arc<Timeline>> {
3962 446 : let state = match cause {
3963 : CreateTimelineCause::Load => {
3964 446 : let ancestor_id = new_metadata.ancestor_timeline();
3965 446 : anyhow::ensure!(
3966 446 : ancestor_id == ancestor.as_ref().map(|t| t.timeline_id),
3967 0 : "Timeline's {new_timeline_id} ancestor {ancestor_id:?} was not found"
3968 : );
3969 446 : TimelineState::Loading
3970 : }
3971 0 : CreateTimelineCause::Delete => TimelineState::Stopping,
3972 : };
3973 :
3974 446 : let pg_version = new_metadata.pg_version();
3975 446 :
3976 446 : let timeline = Timeline::new(
3977 446 : self.conf,
3978 446 : Arc::clone(&self.tenant_conf),
3979 446 : new_metadata,
3980 446 : ancestor,
3981 446 : new_timeline_id,
3982 446 : self.tenant_shard_id,
3983 446 : self.generation,
3984 446 : self.shard_identity,
3985 446 : self.walredo_mgr.clone(),
3986 446 : resources,
3987 446 : pg_version,
3988 446 : state,
3989 446 : self.attach_wal_lag_cooldown.clone(),
3990 446 : create_idempotency,
3991 446 : self.cancel.child_token(),
3992 446 : );
3993 446 :
3994 446 : Ok(timeline)
3995 446 : }
3996 :
3997 : /// [`Tenant::shutdown`] must be called before dropping the returned [`Tenant`] object
3998 : /// to ensure proper cleanup of background tasks and metrics.
3999 : //
4000 : // Allow too_many_arguments because a constructor's argument list naturally grows with the
4001 : // number of attributes in the struct: breaking these out into a builder wouldn't be helpful.
4002 : #[allow(clippy::too_many_arguments)]
4003 220 : fn new(
4004 220 : state: TenantState,
4005 220 : conf: &'static PageServerConf,
4006 220 : attached_conf: AttachedTenantConf,
4007 220 : shard_identity: ShardIdentity,
4008 220 : walredo_mgr: Option<Arc<WalRedoManager>>,
4009 220 : tenant_shard_id: TenantShardId,
4010 220 : remote_storage: GenericRemoteStorage,
4011 220 : deletion_queue_client: DeletionQueueClient,
4012 220 : l0_flush_global_state: L0FlushGlobalState,
4013 220 : ) -> Tenant {
4014 220 : debug_assert!(
4015 220 : !attached_conf.location.generation.is_none() || conf.control_plane_api.is_none()
4016 : );
4017 :
4018 220 : let (state, mut rx) = watch::channel(state);
4019 220 :
4020 220 : tokio::spawn(async move {
4021 220 : // reflect tenant state in metrics:
4022 220 : // - global per tenant state: TENANT_STATE_METRIC
4023 220 : // - "set" of broken tenants: BROKEN_TENANTS_SET
4024 220 : //
4025 220 : // set of broken tenants should not have zero counts so that it remains accessible for
4026 220 : // alerting.
4027 220 :
4028 220 : let tid = tenant_shard_id.to_string();
4029 220 : let shard_id = tenant_shard_id.shard_slug().to_string();
4030 220 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
4031 :
4032 440 : fn inspect_state(state: &TenantState) -> ([&'static str; 1], bool) {
4033 440 : ([state.into()], matches!(state, TenantState::Broken { .. }))
4034 440 : }
4035 :
4036 220 : let mut tuple = inspect_state(&rx.borrow_and_update());
4037 220 :
4038 220 : let is_broken = tuple.1;
4039 220 : let mut counted_broken = if is_broken {
4040 : // add the id to the set right away, there should not be any updates on the channel
4041 : // after before tenant is removed, if ever
4042 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
4043 0 : true
4044 : } else {
4045 220 : false
4046 : };
4047 :
4048 : loop {
4049 440 : let labels = &tuple.0;
4050 440 : let current = TENANT_STATE_METRIC.with_label_values(labels);
4051 440 : current.inc();
4052 440 :
4053 440 : if rx.changed().await.is_err() {
4054 : // tenant has been dropped
4055 14 : current.dec();
4056 14 : drop(BROKEN_TENANTS_SET.remove_label_values(set_key));
4057 14 : break;
4058 220 : }
4059 220 :
4060 220 : current.dec();
4061 220 : tuple = inspect_state(&rx.borrow_and_update());
4062 220 :
4063 220 : let is_broken = tuple.1;
4064 220 : if is_broken && !counted_broken {
4065 0 : counted_broken = true;
4066 0 : // insert the tenant_id (back) into the set while avoiding needless counter
4067 0 : // access
4068 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
4069 220 : }
4070 : }
4071 220 : });
4072 220 :
4073 220 : Tenant {
4074 220 : tenant_shard_id,
4075 220 : shard_identity,
4076 220 : generation: attached_conf.location.generation,
4077 220 : conf,
4078 220 : // using now here is good enough approximation to catch tenants with really long
4079 220 : // activation times.
4080 220 : constructed_at: Instant::now(),
4081 220 : timelines: Mutex::new(HashMap::new()),
4082 220 : timelines_creating: Mutex::new(HashSet::new()),
4083 220 : timelines_offloaded: Mutex::new(HashMap::new()),
4084 220 : tenant_manifest_upload: Default::default(),
4085 220 : gc_cs: tokio::sync::Mutex::new(()),
4086 220 : walredo_mgr,
4087 220 : remote_storage,
4088 220 : deletion_queue_client,
4089 220 : state,
4090 220 : cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
4091 220 : cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
4092 220 : eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
4093 220 : compaction_circuit_breaker: std::sync::Mutex::new(CircuitBreaker::new(
4094 220 : format!("compaction-{tenant_shard_id}"),
4095 220 : 5,
4096 220 : // Compaction can be a very expensive operation, and might leak disk space. It also ought
4097 220 : // to be infallible, as long as remote storage is available. So if it repeatedly fails,
4098 220 : // use an extremely long backoff.
4099 220 : Some(Duration::from_secs(3600 * 24)),
4100 220 : )),
4101 220 : scheduled_compaction_tasks: Mutex::new(Default::default()),
4102 220 : activate_now_sem: tokio::sync::Semaphore::new(0),
4103 220 : attach_wal_lag_cooldown: Arc::new(std::sync::OnceLock::new()),
4104 220 : cancel: CancellationToken::default(),
4105 220 : gate: Gate::default(),
4106 220 : pagestream_throttle: Arc::new(throttle::Throttle::new(
4107 220 : Tenant::get_pagestream_throttle_config(conf, &attached_conf.tenant_conf),
4108 220 : )),
4109 220 : pagestream_throttle_metrics: Arc::new(
4110 220 : crate::metrics::tenant_throttling::Pagestream::new(&tenant_shard_id),
4111 220 : ),
4112 220 : tenant_conf: Arc::new(ArcSwap::from_pointee(attached_conf)),
4113 220 : ongoing_timeline_detach: std::sync::Mutex::default(),
4114 220 : gc_block: Default::default(),
4115 220 : l0_flush_global_state,
4116 220 : }
4117 220 : }
4118 :
4119 : /// Locate and load config
4120 0 : pub(super) fn load_tenant_config(
4121 0 : conf: &'static PageServerConf,
4122 0 : tenant_shard_id: &TenantShardId,
4123 0 : ) -> Result<LocationConf, LoadConfigError> {
4124 0 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4125 0 :
4126 0 : info!("loading tenant configuration from {config_path}");
4127 :
4128 : // load and parse file
4129 0 : let config = fs::read_to_string(&config_path).map_err(|e| {
4130 0 : match e.kind() {
4131 : std::io::ErrorKind::NotFound => {
4132 : // The config should almost always exist for a tenant directory:
4133 : // - When attaching a tenant, the config is the first thing we write
4134 : // - When detaching a tenant, we atomically move the directory to a tmp location
4135 : // before deleting contents.
4136 : //
4137 : // The very rare edge case that can result in a missing config is if we crash during attach
4138 : // between creating directory and writing config. Callers should handle that as if the
4139 : // directory didn't exist.
4140 :
4141 0 : LoadConfigError::NotFound(config_path)
4142 : }
4143 : _ => {
4144 : // No IO errors except NotFound are acceptable here: other kinds of error indicate local storage or permissions issues
4145 : // that we cannot cleanly recover
4146 0 : crate::virtual_file::on_fatal_io_error(&e, "Reading tenant config file")
4147 : }
4148 : }
4149 0 : })?;
4150 :
4151 0 : Ok(toml_edit::de::from_str::<LocationConf>(&config)?)
4152 0 : }
4153 :
4154 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4155 : pub(super) async fn persist_tenant_config(
4156 : conf: &'static PageServerConf,
4157 : tenant_shard_id: &TenantShardId,
4158 : location_conf: &LocationConf,
4159 : ) -> std::io::Result<()> {
4160 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4161 :
4162 : Self::persist_tenant_config_at(tenant_shard_id, &config_path, location_conf).await
4163 : }
4164 :
4165 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4166 : pub(super) async fn persist_tenant_config_at(
4167 : tenant_shard_id: &TenantShardId,
4168 : config_path: &Utf8Path,
4169 : location_conf: &LocationConf,
4170 : ) -> std::io::Result<()> {
4171 : debug!("persisting tenantconf to {config_path}");
4172 :
4173 : let mut conf_content = r#"# This file contains a specific per-tenant's config.
4174 : # It is read in case of pageserver restart.
4175 : "#
4176 : .to_string();
4177 :
4178 0 : fail::fail_point!("tenant-config-before-write", |_| {
4179 0 : Err(std::io::Error::new(
4180 0 : std::io::ErrorKind::Other,
4181 0 : "tenant-config-before-write",
4182 0 : ))
4183 0 : });
4184 :
4185 : // Convert the config to a toml file.
4186 : conf_content +=
4187 : &toml_edit::ser::to_string_pretty(&location_conf).expect("Config serialization failed");
4188 :
4189 : let temp_path = path_with_suffix_extension(config_path, TEMP_FILE_SUFFIX);
4190 :
4191 : let conf_content = conf_content.into_bytes();
4192 : VirtualFile::crashsafe_overwrite(config_path.to_owned(), temp_path, conf_content).await
4193 : }
4194 :
4195 : //
4196 : // How garbage collection works:
4197 : //
4198 : // +--bar------------->
4199 : // /
4200 : // +----+-----foo---------------->
4201 : // /
4202 : // ----main--+-------------------------->
4203 : // \
4204 : // +-----baz-------->
4205 : //
4206 : //
4207 : // 1. Grab 'gc_cs' mutex to prevent new timelines from being created while Timeline's
4208 : // `gc_infos` are being refreshed
4209 : // 2. Scan collected timelines, and on each timeline, make note of the
4210 : // all the points where other timelines have been branched off.
4211 : // We will refrain from removing page versions at those LSNs.
4212 : // 3. For each timeline, scan all layer files on the timeline.
4213 : // Remove all files for which a newer file exists and which
4214 : // don't cover any branch point LSNs.
4215 : //
4216 : // TODO:
4217 : // - if a relation has a non-incremental persistent layer on a child branch, then we
4218 : // don't need to keep that in the parent anymore. But currently
4219 : // we do.
4220 4 : async fn gc_iteration_internal(
4221 4 : &self,
4222 4 : target_timeline_id: Option<TimelineId>,
4223 4 : horizon: u64,
4224 4 : pitr: Duration,
4225 4 : cancel: &CancellationToken,
4226 4 : ctx: &RequestContext,
4227 4 : ) -> Result<GcResult, GcError> {
4228 4 : let mut totals: GcResult = Default::default();
4229 4 : let now = Instant::now();
4230 :
4231 4 : let gc_timelines = self
4232 4 : .refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4233 4 : .await?;
4234 :
4235 4 : failpoint_support::sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
4236 :
4237 : // If there is nothing to GC, we don't want any messages in the INFO log.
4238 4 : if !gc_timelines.is_empty() {
4239 4 : info!("{} timelines need GC", gc_timelines.len());
4240 : } else {
4241 0 : debug!("{} timelines need GC", gc_timelines.len());
4242 : }
4243 :
4244 : // Perform GC for each timeline.
4245 : //
4246 : // Note that we don't hold the `Tenant::gc_cs` lock here because we don't want to delay the
4247 : // branch creation task, which requires the GC lock. A GC iteration can run concurrently
4248 : // with branch creation.
4249 : //
4250 : // See comments in [`Tenant::branch_timeline`] for more information about why branch
4251 : // creation task can run concurrently with timeline's GC iteration.
4252 8 : for timeline in gc_timelines {
4253 4 : if cancel.is_cancelled() {
4254 : // We were requested to shut down. Stop and return with the progress we
4255 : // made.
4256 0 : break;
4257 4 : }
4258 4 : let result = match timeline.gc().await {
4259 : Err(GcError::TimelineCancelled) => {
4260 0 : if target_timeline_id.is_some() {
4261 : // If we were targetting this specific timeline, surface cancellation to caller
4262 0 : return Err(GcError::TimelineCancelled);
4263 : } else {
4264 : // A timeline may be shutting down independently of the tenant's lifecycle: we should
4265 : // skip past this and proceed to try GC on other timelines.
4266 0 : continue;
4267 : }
4268 : }
4269 4 : r => r?,
4270 : };
4271 4 : totals += result;
4272 : }
4273 :
4274 4 : totals.elapsed = now.elapsed();
4275 4 : Ok(totals)
4276 4 : }
4277 :
4278 : /// Refreshes the Timeline::gc_info for all timelines, returning the
4279 : /// vector of timelines which have [`Timeline::get_last_record_lsn`] past
4280 : /// [`Tenant::get_gc_horizon`].
4281 : ///
4282 : /// This is usually executed as part of periodic gc, but can now be triggered more often.
4283 0 : pub(crate) async fn refresh_gc_info(
4284 0 : &self,
4285 0 : cancel: &CancellationToken,
4286 0 : ctx: &RequestContext,
4287 0 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4288 0 : // since this method can now be called at different rates than the configured gc loop, it
4289 0 : // might be that these configuration values get applied faster than what it was previously,
4290 0 : // since these were only read from the gc task.
4291 0 : let horizon = self.get_gc_horizon();
4292 0 : let pitr = self.get_pitr_interval();
4293 0 :
4294 0 : // refresh all timelines
4295 0 : let target_timeline_id = None;
4296 0 :
4297 0 : self.refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4298 0 : .await
4299 0 : }
4300 :
4301 : /// Populate all Timelines' `GcInfo` with information about their children. We do not set the
4302 : /// PITR cutoffs here, because that requires I/O: this is done later, before GC, by [`Self::refresh_gc_info_internal`]
4303 : ///
4304 : /// Subsequently, parent-child relationships are updated incrementally inside [`Timeline::new`] and [`Timeline::drop`].
4305 0 : fn initialize_gc_info(
4306 0 : &self,
4307 0 : timelines: &std::sync::MutexGuard<HashMap<TimelineId, Arc<Timeline>>>,
4308 0 : timelines_offloaded: &std::sync::MutexGuard<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
4309 0 : restrict_to_timeline: Option<TimelineId>,
4310 0 : ) {
4311 0 : if restrict_to_timeline.is_none() {
4312 : // This function must be called before activation: after activation timeline create/delete operations
4313 : // might happen, and this function is not safe to run concurrently with those.
4314 0 : assert!(!self.is_active());
4315 0 : }
4316 :
4317 : // Scan all timelines. For each timeline, remember the timeline ID and
4318 : // the branch point where it was created.
4319 0 : let mut all_branchpoints: BTreeMap<TimelineId, Vec<(Lsn, TimelineId, MaybeOffloaded)>> =
4320 0 : BTreeMap::new();
4321 0 : timelines.iter().for_each(|(timeline_id, timeline_entry)| {
4322 0 : if let Some(ancestor_timeline_id) = &timeline_entry.get_ancestor_timeline_id() {
4323 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4324 0 : ancestor_children.push((
4325 0 : timeline_entry.get_ancestor_lsn(),
4326 0 : *timeline_id,
4327 0 : MaybeOffloaded::No,
4328 0 : ));
4329 0 : }
4330 0 : });
4331 0 : timelines_offloaded
4332 0 : .iter()
4333 0 : .for_each(|(timeline_id, timeline_entry)| {
4334 0 : let Some(ancestor_timeline_id) = &timeline_entry.ancestor_timeline_id else {
4335 0 : return;
4336 : };
4337 0 : let Some(retain_lsn) = timeline_entry.ancestor_retain_lsn else {
4338 0 : return;
4339 : };
4340 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4341 0 : ancestor_children.push((retain_lsn, *timeline_id, MaybeOffloaded::Yes));
4342 0 : });
4343 0 :
4344 0 : // The number of bytes we always keep, irrespective of PITR: this is a constant across timelines
4345 0 : let horizon = self.get_gc_horizon();
4346 :
4347 : // Populate each timeline's GcInfo with information about its child branches
4348 0 : let timelines_to_write = if let Some(timeline_id) = restrict_to_timeline {
4349 0 : itertools::Either::Left(timelines.get(&timeline_id).into_iter())
4350 : } else {
4351 0 : itertools::Either::Right(timelines.values())
4352 : };
4353 0 : for timeline in timelines_to_write {
4354 0 : let mut branchpoints: Vec<(Lsn, TimelineId, MaybeOffloaded)> = all_branchpoints
4355 0 : .remove(&timeline.timeline_id)
4356 0 : .unwrap_or_default();
4357 0 :
4358 0 : branchpoints.sort_by_key(|b| b.0);
4359 0 :
4360 0 : let mut target = timeline.gc_info.write().unwrap();
4361 0 :
4362 0 : target.retain_lsns = branchpoints;
4363 0 :
4364 0 : let space_cutoff = timeline
4365 0 : .get_last_record_lsn()
4366 0 : .checked_sub(horizon)
4367 0 : .unwrap_or(Lsn(0));
4368 0 :
4369 0 : target.cutoffs = GcCutoffs {
4370 0 : space: space_cutoff,
4371 0 : time: Lsn::INVALID,
4372 0 : };
4373 0 : }
4374 0 : }
4375 :
4376 4 : async fn refresh_gc_info_internal(
4377 4 : &self,
4378 4 : target_timeline_id: Option<TimelineId>,
4379 4 : horizon: u64,
4380 4 : pitr: Duration,
4381 4 : cancel: &CancellationToken,
4382 4 : ctx: &RequestContext,
4383 4 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4384 4 : // before taking the gc_cs lock, do the heavier weight finding of gc_cutoff points for
4385 4 : // currently visible timelines.
4386 4 : let timelines = self
4387 4 : .timelines
4388 4 : .lock()
4389 4 : .unwrap()
4390 4 : .values()
4391 4 : .filter(|tl| match target_timeline_id.as_ref() {
4392 4 : Some(target) => &tl.timeline_id == target,
4393 0 : None => true,
4394 4 : })
4395 4 : .cloned()
4396 4 : .collect::<Vec<_>>();
4397 4 :
4398 4 : if target_timeline_id.is_some() && timelines.is_empty() {
4399 : // We were to act on a particular timeline and it wasn't found
4400 0 : return Err(GcError::TimelineNotFound);
4401 4 : }
4402 4 :
4403 4 : let mut gc_cutoffs: HashMap<TimelineId, GcCutoffs> =
4404 4 : HashMap::with_capacity(timelines.len());
4405 4 :
4406 4 : // Ensures all timelines use the same start time when computing the time cutoff.
4407 4 : let now_ts_for_pitr_calc = SystemTime::now();
4408 4 : for timeline in timelines.iter() {
4409 4 : let cutoff = timeline
4410 4 : .get_last_record_lsn()
4411 4 : .checked_sub(horizon)
4412 4 : .unwrap_or(Lsn(0));
4413 :
4414 4 : let cutoffs = timeline
4415 4 : .find_gc_cutoffs(now_ts_for_pitr_calc, cutoff, pitr, cancel, ctx)
4416 4 : .await?;
4417 4 : let old = gc_cutoffs.insert(timeline.timeline_id, cutoffs);
4418 4 : assert!(old.is_none());
4419 : }
4420 :
4421 4 : if !self.is_active() || self.cancel.is_cancelled() {
4422 0 : return Err(GcError::TenantCancelled);
4423 4 : }
4424 :
4425 : // grab mutex to prevent new timelines from being created here; avoid doing long operations
4426 : // because that will stall branch creation.
4427 4 : let gc_cs = self.gc_cs.lock().await;
4428 :
4429 : // Ok, we now know all the branch points.
4430 : // Update the GC information for each timeline.
4431 4 : let mut gc_timelines = Vec::with_capacity(timelines.len());
4432 8 : for timeline in timelines {
4433 : // We filtered the timeline list above
4434 4 : if let Some(target_timeline_id) = target_timeline_id {
4435 4 : assert_eq!(target_timeline_id, timeline.timeline_id);
4436 0 : }
4437 :
4438 : {
4439 4 : let mut target = timeline.gc_info.write().unwrap();
4440 4 :
4441 4 : // Cull any expired leases
4442 4 : let now = SystemTime::now();
4443 6 : target.leases.retain(|_, lease| !lease.is_expired(&now));
4444 4 :
4445 4 : timeline
4446 4 : .metrics
4447 4 : .valid_lsn_lease_count_gauge
4448 4 : .set(target.leases.len() as u64);
4449 :
4450 : // Look up parent's PITR cutoff to update the child's knowledge of whether it is within parent's PITR
4451 4 : if let Some(ancestor_id) = timeline.get_ancestor_timeline_id() {
4452 0 : if let Some(ancestor_gc_cutoffs) = gc_cutoffs.get(&ancestor_id) {
4453 0 : target.within_ancestor_pitr =
4454 0 : timeline.get_ancestor_lsn() >= ancestor_gc_cutoffs.time;
4455 0 : }
4456 4 : }
4457 :
4458 : // Update metrics that depend on GC state
4459 4 : timeline
4460 4 : .metrics
4461 4 : .archival_size
4462 4 : .set(if target.within_ancestor_pitr {
4463 0 : timeline.metrics.current_logical_size_gauge.get()
4464 : } else {
4465 4 : 0
4466 : });
4467 4 : timeline.metrics.pitr_history_size.set(
4468 4 : timeline
4469 4 : .get_last_record_lsn()
4470 4 : .checked_sub(target.cutoffs.time)
4471 4 : .unwrap_or(Lsn(0))
4472 4 : .0,
4473 4 : );
4474 :
4475 : // Apply the cutoffs we found to the Timeline's GcInfo. Why might we _not_ have cutoffs for a timeline?
4476 : // - this timeline was created while we were finding cutoffs
4477 : // - lsn for timestamp search fails for this timeline repeatedly
4478 4 : if let Some(cutoffs) = gc_cutoffs.get(&timeline.timeline_id) {
4479 4 : let original_cutoffs = target.cutoffs.clone();
4480 4 : // GC cutoffs should never go back
4481 4 : target.cutoffs = GcCutoffs {
4482 4 : space: Lsn(cutoffs.space.0.max(original_cutoffs.space.0)),
4483 4 : time: Lsn(cutoffs.time.0.max(original_cutoffs.time.0)),
4484 4 : }
4485 0 : }
4486 : }
4487 :
4488 4 : gc_timelines.push(timeline);
4489 : }
4490 4 : drop(gc_cs);
4491 4 : Ok(gc_timelines)
4492 4 : }
4493 :
4494 : /// A substitute for `branch_timeline` for use in unit tests.
4495 : /// The returned timeline will have state value `Active` to make various `anyhow::ensure!()`
4496 : /// calls pass, but, we do not actually call `.activate()` under the hood. So, none of the
4497 : /// timeline background tasks are launched, except the flush loop.
4498 : #[cfg(test)]
4499 232 : async fn branch_timeline_test(
4500 232 : self: &Arc<Self>,
4501 232 : src_timeline: &Arc<Timeline>,
4502 232 : dst_id: TimelineId,
4503 232 : ancestor_lsn: Option<Lsn>,
4504 232 : ctx: &RequestContext,
4505 232 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
4506 232 : let tl = self
4507 232 : .branch_timeline_impl(src_timeline, dst_id, ancestor_lsn, ctx)
4508 232 : .await?
4509 228 : .into_timeline_for_test();
4510 228 : tl.set_state(TimelineState::Active);
4511 228 : Ok(tl)
4512 232 : }
4513 :
4514 : /// Helper for unit tests to branch a timeline with some pre-loaded states.
4515 : #[cfg(test)]
4516 : #[allow(clippy::too_many_arguments)]
4517 6 : pub async fn branch_timeline_test_with_layers(
4518 6 : self: &Arc<Self>,
4519 6 : src_timeline: &Arc<Timeline>,
4520 6 : dst_id: TimelineId,
4521 6 : ancestor_lsn: Option<Lsn>,
4522 6 : ctx: &RequestContext,
4523 6 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
4524 6 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
4525 6 : end_lsn: Lsn,
4526 6 : ) -> anyhow::Result<Arc<Timeline>> {
4527 : use checks::check_valid_layermap;
4528 : use itertools::Itertools;
4529 :
4530 6 : let tline = self
4531 6 : .branch_timeline_test(src_timeline, dst_id, ancestor_lsn, ctx)
4532 6 : .await?;
4533 6 : let ancestor_lsn = if let Some(ancestor_lsn) = ancestor_lsn {
4534 6 : ancestor_lsn
4535 : } else {
4536 0 : tline.get_last_record_lsn()
4537 : };
4538 6 : assert!(end_lsn >= ancestor_lsn);
4539 6 : tline.force_advance_lsn(end_lsn);
4540 12 : for deltas in delta_layer_desc {
4541 6 : tline
4542 6 : .force_create_delta_layer(deltas, Some(ancestor_lsn), ctx)
4543 6 : .await?;
4544 : }
4545 10 : for (lsn, images) in image_layer_desc {
4546 4 : tline
4547 4 : .force_create_image_layer(lsn, images, Some(ancestor_lsn), ctx)
4548 4 : .await?;
4549 : }
4550 6 : let layer_names = tline
4551 6 : .layers
4552 6 : .read()
4553 6 : .await
4554 6 : .layer_map()
4555 6 : .unwrap()
4556 6 : .iter_historic_layers()
4557 10 : .map(|layer| layer.layer_name())
4558 6 : .collect_vec();
4559 6 : if let Some(err) = check_valid_layermap(&layer_names) {
4560 0 : bail!("invalid layermap: {err}");
4561 6 : }
4562 6 : Ok(tline)
4563 6 : }
4564 :
4565 : /// Branch an existing timeline.
4566 0 : async fn branch_timeline(
4567 0 : self: &Arc<Self>,
4568 0 : src_timeline: &Arc<Timeline>,
4569 0 : dst_id: TimelineId,
4570 0 : start_lsn: Option<Lsn>,
4571 0 : ctx: &RequestContext,
4572 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4573 0 : self.branch_timeline_impl(src_timeline, dst_id, start_lsn, ctx)
4574 0 : .await
4575 0 : }
4576 :
4577 232 : async fn branch_timeline_impl(
4578 232 : self: &Arc<Self>,
4579 232 : src_timeline: &Arc<Timeline>,
4580 232 : dst_id: TimelineId,
4581 232 : start_lsn: Option<Lsn>,
4582 232 : _ctx: &RequestContext,
4583 232 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4584 232 : let src_id = src_timeline.timeline_id;
4585 :
4586 : // We will validate our ancestor LSN in this function. Acquire the GC lock so that
4587 : // this check cannot race with GC, and the ancestor LSN is guaranteed to remain
4588 : // valid while we are creating the branch.
4589 232 : let _gc_cs = self.gc_cs.lock().await;
4590 :
4591 : // If no start LSN is specified, we branch the new timeline from the source timeline's last record LSN
4592 232 : let start_lsn = start_lsn.unwrap_or_else(|| {
4593 2 : let lsn = src_timeline.get_last_record_lsn();
4594 2 : info!("branching timeline {dst_id} from timeline {src_id} at last record LSN: {lsn}");
4595 2 : lsn
4596 232 : });
4597 :
4598 : // we finally have determined the ancestor_start_lsn, so we can get claim exclusivity now
4599 232 : let timeline_create_guard = match self
4600 232 : .start_creating_timeline(
4601 232 : dst_id,
4602 232 : CreateTimelineIdempotency::Branch {
4603 232 : ancestor_timeline_id: src_timeline.timeline_id,
4604 232 : ancestor_start_lsn: start_lsn,
4605 232 : },
4606 232 : )
4607 232 : .await?
4608 : {
4609 232 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
4610 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
4611 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
4612 : }
4613 : };
4614 :
4615 : // Ensure that `start_lsn` is valid, i.e. the LSN is within the PITR
4616 : // horizon on the source timeline
4617 : //
4618 : // We check it against both the planned GC cutoff stored in 'gc_info',
4619 : // and the 'latest_gc_cutoff' of the last GC that was performed. The
4620 : // planned GC cutoff in 'gc_info' is normally larger than
4621 : // 'latest_gc_cutoff_lsn', but beware of corner cases like if you just
4622 : // changed the GC settings for the tenant to make the PITR window
4623 : // larger, but some of the data was already removed by an earlier GC
4624 : // iteration.
4625 :
4626 : // check against last actual 'latest_gc_cutoff' first
4627 232 : let latest_gc_cutoff_lsn = src_timeline.get_latest_gc_cutoff_lsn();
4628 232 : src_timeline
4629 232 : .check_lsn_is_in_scope(start_lsn, &latest_gc_cutoff_lsn)
4630 232 : .context(format!(
4631 232 : "invalid branch start lsn: less than latest GC cutoff {}",
4632 232 : *latest_gc_cutoff_lsn,
4633 232 : ))
4634 232 : .map_err(CreateTimelineError::AncestorLsn)?;
4635 :
4636 : // and then the planned GC cutoff
4637 : {
4638 228 : let gc_info = src_timeline.gc_info.read().unwrap();
4639 228 : let cutoff = gc_info.min_cutoff();
4640 228 : if start_lsn < cutoff {
4641 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
4642 0 : "invalid branch start lsn: less than planned GC cutoff {cutoff}"
4643 0 : )));
4644 228 : }
4645 228 : }
4646 228 :
4647 228 : //
4648 228 : // The branch point is valid, and we are still holding the 'gc_cs' lock
4649 228 : // so that GC cannot advance the GC cutoff until we are finished.
4650 228 : // Proceed with the branch creation.
4651 228 : //
4652 228 :
4653 228 : // Determine prev-LSN for the new timeline. We can only determine it if
4654 228 : // the timeline was branched at the current end of the source timeline.
4655 228 : let RecordLsn {
4656 228 : last: src_last,
4657 228 : prev: src_prev,
4658 228 : } = src_timeline.get_last_record_rlsn();
4659 228 : let dst_prev = if src_last == start_lsn {
4660 216 : Some(src_prev)
4661 : } else {
4662 12 : None
4663 : };
4664 :
4665 : // Create the metadata file, noting the ancestor of the new timeline.
4666 : // There is initially no data in it, but all the read-calls know to look
4667 : // into the ancestor.
4668 228 : let metadata = TimelineMetadata::new(
4669 228 : start_lsn,
4670 228 : dst_prev,
4671 228 : Some(src_id),
4672 228 : start_lsn,
4673 228 : *src_timeline.latest_gc_cutoff_lsn.read(), // FIXME: should we hold onto this guard longer?
4674 228 : src_timeline.initdb_lsn,
4675 228 : src_timeline.pg_version,
4676 228 : );
4677 :
4678 228 : let uninitialized_timeline = self
4679 228 : .prepare_new_timeline(
4680 228 : dst_id,
4681 228 : &metadata,
4682 228 : timeline_create_guard,
4683 228 : start_lsn + 1,
4684 228 : Some(Arc::clone(src_timeline)),
4685 228 : )
4686 228 : .await?;
4687 :
4688 228 : let new_timeline = uninitialized_timeline.finish_creation()?;
4689 :
4690 : // Root timeline gets its layers during creation and uploads them along with the metadata.
4691 : // A branch timeline though, when created, can get no writes for some time, hence won't get any layers created.
4692 : // We still need to upload its metadata eagerly: if other nodes `attach` the tenant and miss this timeline, their GC
4693 : // could get incorrect information and remove more layers, than needed.
4694 : // See also https://github.com/neondatabase/neon/issues/3865
4695 228 : new_timeline
4696 228 : .remote_client
4697 228 : .schedule_index_upload_for_full_metadata_update(&metadata)
4698 228 : .context("branch initial metadata upload")?;
4699 :
4700 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
4701 :
4702 228 : Ok(CreateTimelineResult::Created(new_timeline))
4703 232 : }
4704 :
4705 : /// For unit tests, make this visible so that other modules can directly create timelines
4706 : #[cfg(test)]
4707 : #[tracing::instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
4708 : pub(crate) async fn bootstrap_timeline_test(
4709 : self: &Arc<Self>,
4710 : timeline_id: TimelineId,
4711 : pg_version: u32,
4712 : load_existing_initdb: Option<TimelineId>,
4713 : ctx: &RequestContext,
4714 : ) -> anyhow::Result<Arc<Timeline>> {
4715 : self.bootstrap_timeline(timeline_id, pg_version, load_existing_initdb, ctx)
4716 : .await
4717 : .map_err(anyhow::Error::new)
4718 2 : .map(|r| r.into_timeline_for_test())
4719 : }
4720 :
4721 : /// Get exclusive access to the timeline ID for creation.
4722 : ///
4723 : /// Timeline-creating code paths must use this function before making changes
4724 : /// to in-memory or persistent state.
4725 : ///
4726 : /// The `state` parameter is a description of the timeline creation operation
4727 : /// we intend to perform.
4728 : /// If the timeline was already created in the meantime, we check whether this
4729 : /// request conflicts or is idempotent , based on `state`.
4730 446 : async fn start_creating_timeline(
4731 446 : self: &Arc<Self>,
4732 446 : new_timeline_id: TimelineId,
4733 446 : idempotency: CreateTimelineIdempotency,
4734 446 : ) -> Result<StartCreatingTimelineResult, CreateTimelineError> {
4735 446 : let allow_offloaded = false;
4736 446 : match self.create_timeline_create_guard(new_timeline_id, idempotency, allow_offloaded) {
4737 444 : Ok(create_guard) => {
4738 444 : pausable_failpoint!("timeline-creation-after-uninit");
4739 444 : Ok(StartCreatingTimelineResult::CreateGuard(create_guard))
4740 : }
4741 0 : Err(TimelineExclusionError::ShuttingDown) => Err(CreateTimelineError::ShuttingDown),
4742 : Err(TimelineExclusionError::AlreadyCreating) => {
4743 : // Creation is in progress, we cannot create it again, and we cannot
4744 : // check if this request matches the existing one, so caller must try
4745 : // again later.
4746 0 : Err(CreateTimelineError::AlreadyCreating)
4747 : }
4748 0 : Err(TimelineExclusionError::Other(e)) => Err(CreateTimelineError::Other(e)),
4749 : Err(TimelineExclusionError::AlreadyExists {
4750 0 : existing: TimelineOrOffloaded::Offloaded(_existing),
4751 0 : ..
4752 0 : }) => {
4753 0 : info!("timeline already exists but is offloaded");
4754 0 : Err(CreateTimelineError::Conflict)
4755 : }
4756 : Err(TimelineExclusionError::AlreadyExists {
4757 2 : existing: TimelineOrOffloaded::Timeline(existing),
4758 2 : arg,
4759 2 : }) => {
4760 2 : {
4761 2 : let existing = &existing.create_idempotency;
4762 2 : let _span = info_span!("idempotency_check", ?existing, ?arg).entered();
4763 2 : debug!("timeline already exists");
4764 :
4765 2 : match (existing, &arg) {
4766 : // FailWithConflict => no idempotency check
4767 : (CreateTimelineIdempotency::FailWithConflict, _)
4768 : | (_, CreateTimelineIdempotency::FailWithConflict) => {
4769 2 : warn!("timeline already exists, failing request");
4770 2 : return Err(CreateTimelineError::Conflict);
4771 : }
4772 : // Idempotent <=> CreateTimelineIdempotency is identical
4773 0 : (x, y) if x == y => {
4774 0 : info!("timeline already exists and idempotency matches, succeeding request");
4775 : // fallthrough
4776 : }
4777 : (_, _) => {
4778 0 : warn!("idempotency conflict, failing request");
4779 0 : return Err(CreateTimelineError::Conflict);
4780 : }
4781 : }
4782 : }
4783 :
4784 0 : Ok(StartCreatingTimelineResult::Idempotent(existing))
4785 : }
4786 : }
4787 446 : }
4788 :
4789 0 : async fn upload_initdb(
4790 0 : &self,
4791 0 : timelines_path: &Utf8PathBuf,
4792 0 : pgdata_path: &Utf8PathBuf,
4793 0 : timeline_id: &TimelineId,
4794 0 : ) -> anyhow::Result<()> {
4795 0 : let temp_path = timelines_path.join(format!(
4796 0 : "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
4797 0 : ));
4798 0 :
4799 0 : scopeguard::defer! {
4800 0 : if let Err(e) = fs::remove_file(&temp_path) {
4801 0 : error!("Failed to remove temporary initdb archive '{temp_path}': {e}");
4802 0 : }
4803 0 : }
4804 :
4805 0 : let (pgdata_zstd, tar_zst_size) = create_zst_tarball(pgdata_path, &temp_path).await?;
4806 : const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
4807 0 : if tar_zst_size > INITDB_TAR_ZST_WARN_LIMIT {
4808 0 : warn!(
4809 0 : "compressed {temp_path} size of {tar_zst_size} is above limit {INITDB_TAR_ZST_WARN_LIMIT}."
4810 : );
4811 0 : }
4812 :
4813 0 : pausable_failpoint!("before-initdb-upload");
4814 :
4815 0 : backoff::retry(
4816 0 : || async {
4817 0 : self::remote_timeline_client::upload_initdb_dir(
4818 0 : &self.remote_storage,
4819 0 : &self.tenant_shard_id.tenant_id,
4820 0 : timeline_id,
4821 0 : pgdata_zstd.try_clone().await?,
4822 0 : tar_zst_size,
4823 0 : &self.cancel,
4824 0 : )
4825 0 : .await
4826 0 : },
4827 0 : |_| false,
4828 0 : 3,
4829 0 : u32::MAX,
4830 0 : "persist_initdb_tar_zst",
4831 0 : &self.cancel,
4832 0 : )
4833 0 : .await
4834 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
4835 0 : .and_then(|x| x)
4836 0 : }
4837 :
4838 : /// - run initdb to init temporary instance and get bootstrap data
4839 : /// - after initialization completes, tar up the temp dir and upload it to S3.
4840 2 : async fn bootstrap_timeline(
4841 2 : self: &Arc<Self>,
4842 2 : timeline_id: TimelineId,
4843 2 : pg_version: u32,
4844 2 : load_existing_initdb: Option<TimelineId>,
4845 2 : ctx: &RequestContext,
4846 2 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4847 2 : let timeline_create_guard = match self
4848 2 : .start_creating_timeline(
4849 2 : timeline_id,
4850 2 : CreateTimelineIdempotency::Bootstrap { pg_version },
4851 2 : )
4852 2 : .await?
4853 : {
4854 2 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
4855 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
4856 0 : return Ok(CreateTimelineResult::Idempotent(timeline))
4857 : }
4858 : };
4859 :
4860 : // create a `tenant/{tenant_id}/timelines/basebackup-{timeline_id}.{TEMP_FILE_SUFFIX}/`
4861 : // temporary directory for basebackup files for the given timeline.
4862 :
4863 2 : let timelines_path = self.conf.timelines_path(&self.tenant_shard_id);
4864 2 : let pgdata_path = path_with_suffix_extension(
4865 2 : timelines_path.join(format!("basebackup-{timeline_id}")),
4866 2 : TEMP_FILE_SUFFIX,
4867 2 : );
4868 2 :
4869 2 : // Remove whatever was left from the previous runs: safe because TimelineCreateGuard guarantees
4870 2 : // we won't race with other creations or existent timelines with the same path.
4871 2 : if pgdata_path.exists() {
4872 0 : fs::remove_dir_all(&pgdata_path).with_context(|| {
4873 0 : format!("Failed to remove already existing initdb directory: {pgdata_path}")
4874 0 : })?;
4875 2 : }
4876 :
4877 : // this new directory is very temporary, set to remove it immediately after bootstrap, we don't need it
4878 2 : scopeguard::defer! {
4879 2 : if let Err(e) = fs::remove_dir_all(&pgdata_path) {
4880 2 : // this is unlikely, but we will remove the directory on pageserver restart or another bootstrap call
4881 2 : error!("Failed to remove temporary initdb directory '{pgdata_path}': {e}");
4882 2 : }
4883 2 : }
4884 2 : if let Some(existing_initdb_timeline_id) = load_existing_initdb {
4885 2 : if existing_initdb_timeline_id != timeline_id {
4886 0 : let source_path = &remote_initdb_archive_path(
4887 0 : &self.tenant_shard_id.tenant_id,
4888 0 : &existing_initdb_timeline_id,
4889 0 : );
4890 0 : let dest_path =
4891 0 : &remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &timeline_id);
4892 0 :
4893 0 : // if this fails, it will get retried by retried control plane requests
4894 0 : self.remote_storage
4895 0 : .copy_object(source_path, dest_path, &self.cancel)
4896 0 : .await
4897 0 : .context("copy initdb tar")?;
4898 2 : }
4899 2 : let (initdb_tar_zst_path, initdb_tar_zst) =
4900 2 : self::remote_timeline_client::download_initdb_tar_zst(
4901 2 : self.conf,
4902 2 : &self.remote_storage,
4903 2 : &self.tenant_shard_id,
4904 2 : &existing_initdb_timeline_id,
4905 2 : &self.cancel,
4906 2 : )
4907 2 : .await
4908 2 : .context("download initdb tar")?;
4909 :
4910 2 : scopeguard::defer! {
4911 2 : if let Err(e) = fs::remove_file(&initdb_tar_zst_path) {
4912 2 : error!("Failed to remove temporary initdb archive '{initdb_tar_zst_path}': {e}");
4913 2 : }
4914 2 : }
4915 2 :
4916 2 : let buf_read =
4917 2 : BufReader::with_capacity(remote_timeline_client::BUFFER_SIZE, initdb_tar_zst);
4918 2 : extract_zst_tarball(&pgdata_path, buf_read)
4919 2 : .await
4920 2 : .context("extract initdb tar")?;
4921 : } else {
4922 : // Init temporarily repo to get bootstrap data, this creates a directory in the `pgdata_path` path
4923 0 : run_initdb(self.conf, &pgdata_path, pg_version, &self.cancel)
4924 0 : .await
4925 0 : .context("run initdb")?;
4926 :
4927 : // Upload the created data dir to S3
4928 0 : if self.tenant_shard_id().is_shard_zero() {
4929 0 : self.upload_initdb(&timelines_path, &pgdata_path, &timeline_id)
4930 0 : .await?;
4931 0 : }
4932 : }
4933 2 : let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
4934 2 :
4935 2 : // Import the contents of the data directory at the initial checkpoint
4936 2 : // LSN, and any WAL after that.
4937 2 : // Initdb lsn will be equal to last_record_lsn which will be set after import.
4938 2 : // Because we know it upfront avoid having an option or dummy zero value by passing it to the metadata.
4939 2 : let new_metadata = TimelineMetadata::new(
4940 2 : Lsn(0),
4941 2 : None,
4942 2 : None,
4943 2 : Lsn(0),
4944 2 : pgdata_lsn,
4945 2 : pgdata_lsn,
4946 2 : pg_version,
4947 2 : );
4948 2 : let raw_timeline = self
4949 2 : .prepare_new_timeline(
4950 2 : timeline_id,
4951 2 : &new_metadata,
4952 2 : timeline_create_guard,
4953 2 : pgdata_lsn,
4954 2 : None,
4955 2 : )
4956 2 : .await?;
4957 :
4958 2 : let tenant_shard_id = raw_timeline.owning_tenant.tenant_shard_id;
4959 2 : let unfinished_timeline = raw_timeline.raw_timeline()?;
4960 :
4961 : // Flush the new layer files to disk, before we make the timeline as available to
4962 : // the outside world.
4963 : //
4964 : // Flush loop needs to be spawned in order to be able to flush.
4965 2 : unfinished_timeline.maybe_spawn_flush_loop();
4966 2 :
4967 2 : import_datadir::import_timeline_from_postgres_datadir(
4968 2 : unfinished_timeline,
4969 2 : &pgdata_path,
4970 2 : pgdata_lsn,
4971 2 : ctx,
4972 2 : )
4973 2 : .await
4974 2 : .with_context(|| {
4975 0 : format!("Failed to import pgdatadir for timeline {tenant_shard_id}/{timeline_id}")
4976 2 : })?;
4977 :
4978 2 : fail::fail_point!("before-checkpoint-new-timeline", |_| {
4979 0 : Err(CreateTimelineError::Other(anyhow::anyhow!(
4980 0 : "failpoint before-checkpoint-new-timeline"
4981 0 : )))
4982 2 : });
4983 :
4984 2 : unfinished_timeline
4985 2 : .freeze_and_flush()
4986 2 : .await
4987 2 : .with_context(|| {
4988 0 : format!(
4989 0 : "Failed to flush after pgdatadir import for timeline {tenant_shard_id}/{timeline_id}"
4990 0 : )
4991 2 : })?;
4992 :
4993 : // All done!
4994 2 : let timeline = raw_timeline.finish_creation()?;
4995 :
4996 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
4997 :
4998 2 : Ok(CreateTimelineResult::Created(timeline))
4999 2 : }
5000 :
5001 440 : fn build_timeline_remote_client(&self, timeline_id: TimelineId) -> RemoteTimelineClient {
5002 440 : RemoteTimelineClient::new(
5003 440 : self.remote_storage.clone(),
5004 440 : self.deletion_queue_client.clone(),
5005 440 : self.conf,
5006 440 : self.tenant_shard_id,
5007 440 : timeline_id,
5008 440 : self.generation,
5009 440 : &self.tenant_conf.load().location,
5010 440 : )
5011 440 : }
5012 :
5013 : /// Call this before constructing a timeline, to build its required structures
5014 440 : fn build_timeline_resources(&self, timeline_id: TimelineId) -> TimelineResources {
5015 440 : TimelineResources {
5016 440 : remote_client: self.build_timeline_remote_client(timeline_id),
5017 440 : pagestream_throttle: self.pagestream_throttle.clone(),
5018 440 : pagestream_throttle_metrics: self.pagestream_throttle_metrics.clone(),
5019 440 : l0_flush_global_state: self.l0_flush_global_state.clone(),
5020 440 : }
5021 440 : }
5022 :
5023 : /// Creates intermediate timeline structure and its files.
5024 : ///
5025 : /// An empty layer map is initialized, and new data and WAL can be imported starting
5026 : /// at 'disk_consistent_lsn'. After any initial data has been imported, call
5027 : /// `finish_creation` to insert the Timeline into the timelines map.
5028 440 : async fn prepare_new_timeline<'a>(
5029 440 : &'a self,
5030 440 : new_timeline_id: TimelineId,
5031 440 : new_metadata: &TimelineMetadata,
5032 440 : create_guard: TimelineCreateGuard,
5033 440 : start_lsn: Lsn,
5034 440 : ancestor: Option<Arc<Timeline>>,
5035 440 : ) -> anyhow::Result<UninitializedTimeline<'a>> {
5036 440 : let tenant_shard_id = self.tenant_shard_id;
5037 440 :
5038 440 : let resources = self.build_timeline_resources(new_timeline_id);
5039 440 : resources
5040 440 : .remote_client
5041 440 : .init_upload_queue_for_empty_remote(new_metadata)?;
5042 :
5043 440 : let timeline_struct = self
5044 440 : .create_timeline_struct(
5045 440 : new_timeline_id,
5046 440 : new_metadata,
5047 440 : ancestor,
5048 440 : resources,
5049 440 : CreateTimelineCause::Load,
5050 440 : create_guard.idempotency.clone(),
5051 440 : )
5052 440 : .context("Failed to create timeline data structure")?;
5053 :
5054 440 : timeline_struct.init_empty_layer_map(start_lsn);
5055 :
5056 440 : if let Err(e) = self
5057 440 : .create_timeline_files(&create_guard.timeline_path)
5058 440 : .await
5059 : {
5060 0 : error!("Failed to create initial files for timeline {tenant_shard_id}/{new_timeline_id}, cleaning up: {e:?}");
5061 0 : cleanup_timeline_directory(create_guard);
5062 0 : return Err(e);
5063 440 : }
5064 440 :
5065 440 : debug!(
5066 0 : "Successfully created initial files for timeline {tenant_shard_id}/{new_timeline_id}"
5067 : );
5068 :
5069 440 : Ok(UninitializedTimeline::new(
5070 440 : self,
5071 440 : new_timeline_id,
5072 440 : Some((timeline_struct, create_guard)),
5073 440 : ))
5074 440 : }
5075 :
5076 440 : async fn create_timeline_files(&self, timeline_path: &Utf8Path) -> anyhow::Result<()> {
5077 440 : crashsafe::create_dir(timeline_path).context("Failed to create timeline directory")?;
5078 :
5079 440 : fail::fail_point!("after-timeline-dir-creation", |_| {
5080 0 : anyhow::bail!("failpoint after-timeline-dir-creation");
5081 440 : });
5082 :
5083 440 : Ok(())
5084 440 : }
5085 :
5086 : /// Get a guard that provides exclusive access to the timeline directory, preventing
5087 : /// concurrent attempts to create the same timeline.
5088 : ///
5089 : /// The `allow_offloaded` parameter controls whether to tolerate the existence of
5090 : /// offloaded timelines or not.
5091 446 : fn create_timeline_create_guard(
5092 446 : self: &Arc<Self>,
5093 446 : timeline_id: TimelineId,
5094 446 : idempotency: CreateTimelineIdempotency,
5095 446 : allow_offloaded: bool,
5096 446 : ) -> Result<TimelineCreateGuard, TimelineExclusionError> {
5097 446 : let tenant_shard_id = self.tenant_shard_id;
5098 446 :
5099 446 : let timeline_path = self.conf.timeline_path(&tenant_shard_id, &timeline_id);
5100 :
5101 446 : let create_guard = TimelineCreateGuard::new(
5102 446 : self,
5103 446 : timeline_id,
5104 446 : timeline_path.clone(),
5105 446 : idempotency,
5106 446 : allow_offloaded,
5107 446 : )?;
5108 :
5109 : // At this stage, we have got exclusive access to in-memory state for this timeline ID
5110 : // for creation.
5111 : // A timeline directory should never exist on disk already:
5112 : // - a previous failed creation would have cleaned up after itself
5113 : // - a pageserver restart would clean up timeline directories that don't have valid remote state
5114 : //
5115 : // Therefore it is an unexpected internal error to encounter a timeline directory already existing here,
5116 : // this error may indicate a bug in cleanup on failed creations.
5117 444 : if timeline_path.exists() {
5118 0 : return Err(TimelineExclusionError::Other(anyhow::anyhow!(
5119 0 : "Timeline directory already exists! This is a bug."
5120 0 : )));
5121 444 : }
5122 444 :
5123 444 : Ok(create_guard)
5124 446 : }
5125 :
5126 : /// Gathers inputs from all of the timelines to produce a sizing model input.
5127 : ///
5128 : /// Future is cancellation safe. Only one calculation can be running at once per tenant.
5129 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5130 : pub async fn gather_size_inputs(
5131 : &self,
5132 : // `max_retention_period` overrides the cutoff that is used to calculate the size
5133 : // (only if it is shorter than the real cutoff).
5134 : max_retention_period: Option<u64>,
5135 : cause: LogicalSizeCalculationCause,
5136 : cancel: &CancellationToken,
5137 : ctx: &RequestContext,
5138 : ) -> Result<size::ModelInputs, size::CalculateSyntheticSizeError> {
5139 : let logical_sizes_at_once = self
5140 : .conf
5141 : .concurrent_tenant_size_logical_size_queries
5142 : .inner();
5143 :
5144 : // TODO: Having a single mutex block concurrent reads is not great for performance.
5145 : //
5146 : // But the only case where we need to run multiple of these at once is when we
5147 : // request a size for a tenant manually via API, while another background calculation
5148 : // is in progress (which is not a common case).
5149 : //
5150 : // See more for on the issue #2748 condenced out of the initial PR review.
5151 : let mut shared_cache = tokio::select! {
5152 : locked = self.cached_logical_sizes.lock() => locked,
5153 : _ = cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5154 : _ = self.cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5155 : };
5156 :
5157 : size::gather_inputs(
5158 : self,
5159 : logical_sizes_at_once,
5160 : max_retention_period,
5161 : &mut shared_cache,
5162 : cause,
5163 : cancel,
5164 : ctx,
5165 : )
5166 : .await
5167 : }
5168 :
5169 : /// Calculate synthetic tenant size and cache the result.
5170 : /// This is periodically called by background worker.
5171 : /// result is cached in tenant struct
5172 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5173 : pub async fn calculate_synthetic_size(
5174 : &self,
5175 : cause: LogicalSizeCalculationCause,
5176 : cancel: &CancellationToken,
5177 : ctx: &RequestContext,
5178 : ) -> Result<u64, size::CalculateSyntheticSizeError> {
5179 : let inputs = self.gather_size_inputs(None, cause, cancel, ctx).await?;
5180 :
5181 : let size = inputs.calculate();
5182 :
5183 : self.set_cached_synthetic_size(size);
5184 :
5185 : Ok(size)
5186 : }
5187 :
5188 : /// Cache given synthetic size and update the metric value
5189 0 : pub fn set_cached_synthetic_size(&self, size: u64) {
5190 0 : self.cached_synthetic_tenant_size
5191 0 : .store(size, Ordering::Relaxed);
5192 0 :
5193 0 : // Only shard zero should be calculating synthetic sizes
5194 0 : debug_assert!(self.shard_identity.is_shard_zero());
5195 :
5196 0 : TENANT_SYNTHETIC_SIZE_METRIC
5197 0 : .get_metric_with_label_values(&[&self.tenant_shard_id.tenant_id.to_string()])
5198 0 : .unwrap()
5199 0 : .set(size);
5200 0 : }
5201 :
5202 0 : pub fn cached_synthetic_size(&self) -> u64 {
5203 0 : self.cached_synthetic_tenant_size.load(Ordering::Relaxed)
5204 0 : }
5205 :
5206 : /// Flush any in-progress layers, schedule uploads, and wait for uploads to complete.
5207 : ///
5208 : /// This function can take a long time: callers should wrap it in a timeout if calling
5209 : /// from an external API handler.
5210 : ///
5211 : /// Cancel-safety: cancelling this function may leave I/O running, but such I/O is
5212 : /// still bounded by tenant/timeline shutdown.
5213 : #[tracing::instrument(skip_all)]
5214 : pub(crate) async fn flush_remote(&self) -> anyhow::Result<()> {
5215 : let timelines = self.timelines.lock().unwrap().clone();
5216 :
5217 0 : async fn flush_timeline(_gate: GateGuard, timeline: Arc<Timeline>) -> anyhow::Result<()> {
5218 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Flushing...");
5219 0 : timeline.freeze_and_flush().await?;
5220 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Waiting for uploads...");
5221 0 : timeline.remote_client.wait_completion().await?;
5222 :
5223 0 : Ok(())
5224 0 : }
5225 :
5226 : // We do not use a JoinSet for these tasks, because we don't want them to be
5227 : // aborted when this function's future is cancelled: they should stay alive
5228 : // holding their GateGuard until they complete, to ensure their I/Os complete
5229 : // before Timeline shutdown completes.
5230 : let mut results = FuturesUnordered::new();
5231 :
5232 : for (_timeline_id, timeline) in timelines {
5233 : // Run each timeline's flush in a task holding the timeline's gate: this
5234 : // means that if this function's future is cancelled, the Timeline shutdown
5235 : // will still wait for any I/O in here to complete.
5236 : let Ok(gate) = timeline.gate.enter() else {
5237 : continue;
5238 : };
5239 0 : let jh = tokio::task::spawn(async move { flush_timeline(gate, timeline).await });
5240 : results.push(jh);
5241 : }
5242 :
5243 : while let Some(r) = results.next().await {
5244 : if let Err(e) = r {
5245 : if !e.is_cancelled() && !e.is_panic() {
5246 : tracing::error!("unexpected join error: {e:?}");
5247 : }
5248 : }
5249 : }
5250 :
5251 : // The flushes we did above were just writes, but the Tenant might have had
5252 : // pending deletions as well from recent compaction/gc: we want to flush those
5253 : // as well. This requires flushing the global delete queue. This is cheap
5254 : // because it's typically a no-op.
5255 : match self.deletion_queue_client.flush_execute().await {
5256 : Ok(_) => {}
5257 : Err(DeletionQueueError::ShuttingDown) => {}
5258 : }
5259 :
5260 : Ok(())
5261 : }
5262 :
5263 0 : pub(crate) fn get_tenant_conf(&self) -> TenantConfOpt {
5264 0 : self.tenant_conf.load().tenant_conf.clone()
5265 0 : }
5266 :
5267 : /// How much local storage would this tenant like to have? It can cope with
5268 : /// less than this (via eviction and on-demand downloads), but this function enables
5269 : /// the Tenant to advertise how much storage it would prefer to have to provide fast I/O
5270 : /// by keeping important things on local disk.
5271 : ///
5272 : /// This is a heuristic, not a guarantee: tenants that are long-idle will actually use less
5273 : /// than they report here, due to layer eviction. Tenants with many active branches may
5274 : /// actually use more than they report here.
5275 0 : pub(crate) fn local_storage_wanted(&self) -> u64 {
5276 0 : let timelines = self.timelines.lock().unwrap();
5277 0 :
5278 0 : // Heuristic: we use the max() of the timelines' visible sizes, rather than the sum. This
5279 0 : // reflects the observation that on tenants with multiple large branches, typically only one
5280 0 : // of them is used actively enough to occupy space on disk.
5281 0 : timelines
5282 0 : .values()
5283 0 : .map(|t| t.metrics.visible_physical_size_gauge.get())
5284 0 : .max()
5285 0 : .unwrap_or(0)
5286 0 : }
5287 :
5288 : /// Serialize and write the latest TenantManifest to remote storage.
5289 2 : pub(crate) async fn store_tenant_manifest(&self) -> Result<(), TenantManifestError> {
5290 : // Only one manifest write may be done at at time, and the contents of the manifest
5291 : // must be loaded while holding this lock. This makes it safe to call this function
5292 : // from anywhere without worrying about colliding updates.
5293 2 : let mut guard = tokio::select! {
5294 2 : g = self.tenant_manifest_upload.lock() => {
5295 2 : g
5296 : },
5297 2 : _ = self.cancel.cancelled() => {
5298 0 : return Err(TenantManifestError::Cancelled);
5299 : }
5300 : };
5301 :
5302 2 : let manifest = self.build_tenant_manifest();
5303 2 : if Some(&manifest) == (*guard).as_ref() {
5304 : // Optimisation: skip uploads that don't change anything.
5305 0 : return Ok(());
5306 2 : }
5307 2 :
5308 2 : upload_tenant_manifest(
5309 2 : &self.remote_storage,
5310 2 : &self.tenant_shard_id,
5311 2 : self.generation,
5312 2 : &manifest,
5313 2 : &self.cancel,
5314 2 : )
5315 2 : .await
5316 2 : .map_err(|e| {
5317 0 : if self.cancel.is_cancelled() {
5318 0 : TenantManifestError::Cancelled
5319 : } else {
5320 0 : TenantManifestError::RemoteStorage(e)
5321 : }
5322 2 : })?;
5323 :
5324 : // Store the successfully uploaded manifest, so that future callers can avoid
5325 : // re-uploading the same thing.
5326 2 : *guard = Some(manifest);
5327 2 :
5328 2 : Ok(())
5329 2 : }
5330 : }
5331 :
5332 : /// Create the cluster temporarily in 'initdbpath' directory inside the repository
5333 : /// to get bootstrap data for timeline initialization.
5334 0 : async fn run_initdb(
5335 0 : conf: &'static PageServerConf,
5336 0 : initdb_target_dir: &Utf8Path,
5337 0 : pg_version: u32,
5338 0 : cancel: &CancellationToken,
5339 0 : ) -> Result<(), InitdbError> {
5340 0 : let initdb_bin_path = conf
5341 0 : .pg_bin_dir(pg_version)
5342 0 : .map_err(InitdbError::Other)?
5343 0 : .join("initdb");
5344 0 : let initdb_lib_dir = conf.pg_lib_dir(pg_version).map_err(InitdbError::Other)?;
5345 0 : info!(
5346 0 : "running {} in {}, libdir: {}",
5347 : initdb_bin_path, initdb_target_dir, initdb_lib_dir,
5348 : );
5349 :
5350 0 : let _permit = INIT_DB_SEMAPHORE.acquire().await;
5351 :
5352 0 : let res = postgres_initdb::do_run_initdb(postgres_initdb::RunInitdbArgs {
5353 0 : superuser: &conf.superuser,
5354 0 : locale: &conf.locale,
5355 0 : initdb_bin: &initdb_bin_path,
5356 0 : pg_version,
5357 0 : library_search_path: &initdb_lib_dir,
5358 0 : pgdata: initdb_target_dir,
5359 0 : })
5360 0 : .await
5361 0 : .map_err(InitdbError::Inner);
5362 0 :
5363 0 : // This isn't true cancellation support, see above. Still return an error to
5364 0 : // excercise the cancellation code path.
5365 0 : if cancel.is_cancelled() {
5366 0 : return Err(InitdbError::Cancelled);
5367 0 : }
5368 0 :
5369 0 : res
5370 0 : }
5371 :
5372 : /// Dump contents of a layer file to stdout.
5373 0 : pub async fn dump_layerfile_from_path(
5374 0 : path: &Utf8Path,
5375 0 : verbose: bool,
5376 0 : ctx: &RequestContext,
5377 0 : ) -> anyhow::Result<()> {
5378 : use std::os::unix::fs::FileExt;
5379 :
5380 : // All layer files start with a two-byte "magic" value, to identify the kind of
5381 : // file.
5382 0 : let file = File::open(path)?;
5383 0 : let mut header_buf = [0u8; 2];
5384 0 : file.read_exact_at(&mut header_buf, 0)?;
5385 :
5386 0 : match u16::from_be_bytes(header_buf) {
5387 : crate::IMAGE_FILE_MAGIC => {
5388 0 : ImageLayer::new_for_path(path, file)?
5389 0 : .dump(verbose, ctx)
5390 0 : .await?
5391 : }
5392 : crate::DELTA_FILE_MAGIC => {
5393 0 : DeltaLayer::new_for_path(path, file)?
5394 0 : .dump(verbose, ctx)
5395 0 : .await?
5396 : }
5397 0 : magic => bail!("unrecognized magic identifier: {:?}", magic),
5398 : }
5399 :
5400 0 : Ok(())
5401 0 : }
5402 :
5403 : #[cfg(test)]
5404 : pub(crate) mod harness {
5405 : use bytes::{Bytes, BytesMut};
5406 : use once_cell::sync::OnceCell;
5407 : use pageserver_api::models::ShardParameters;
5408 : use pageserver_api::shard::ShardIndex;
5409 : use utils::logging;
5410 :
5411 : use crate::deletion_queue::mock::MockDeletionQueue;
5412 : use crate::l0_flush::L0FlushConfig;
5413 : use crate::walredo::apply_neon;
5414 : use pageserver_api::key::Key;
5415 : use pageserver_api::record::NeonWalRecord;
5416 :
5417 : use super::*;
5418 : use hex_literal::hex;
5419 : use utils::id::TenantId;
5420 :
5421 : pub const TIMELINE_ID: TimelineId =
5422 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
5423 : pub const NEW_TIMELINE_ID: TimelineId =
5424 : TimelineId::from_array(hex!("AA223344556677881122334455667788"));
5425 :
5426 : /// Convenience function to create a page image with given string as the only content
5427 5028755 : pub fn test_img(s: &str) -> Bytes {
5428 5028755 : let mut buf = BytesMut::new();
5429 5028755 : buf.extend_from_slice(s.as_bytes());
5430 5028755 : buf.resize(64, 0);
5431 5028755 :
5432 5028755 : buf.freeze()
5433 5028755 : }
5434 :
5435 : impl From<TenantConf> for TenantConfOpt {
5436 220 : fn from(tenant_conf: TenantConf) -> Self {
5437 220 : Self {
5438 220 : checkpoint_distance: Some(tenant_conf.checkpoint_distance),
5439 220 : checkpoint_timeout: Some(tenant_conf.checkpoint_timeout),
5440 220 : compaction_target_size: Some(tenant_conf.compaction_target_size),
5441 220 : compaction_period: Some(tenant_conf.compaction_period),
5442 220 : compaction_threshold: Some(tenant_conf.compaction_threshold),
5443 220 : compaction_algorithm: Some(tenant_conf.compaction_algorithm),
5444 220 : gc_horizon: Some(tenant_conf.gc_horizon),
5445 220 : gc_period: Some(tenant_conf.gc_period),
5446 220 : image_creation_threshold: Some(tenant_conf.image_creation_threshold),
5447 220 : pitr_interval: Some(tenant_conf.pitr_interval),
5448 220 : walreceiver_connect_timeout: Some(tenant_conf.walreceiver_connect_timeout),
5449 220 : lagging_wal_timeout: Some(tenant_conf.lagging_wal_timeout),
5450 220 : max_lsn_wal_lag: Some(tenant_conf.max_lsn_wal_lag),
5451 220 : eviction_policy: Some(tenant_conf.eviction_policy),
5452 220 : min_resident_size_override: tenant_conf.min_resident_size_override,
5453 220 : evictions_low_residence_duration_metric_threshold: Some(
5454 220 : tenant_conf.evictions_low_residence_duration_metric_threshold,
5455 220 : ),
5456 220 : heatmap_period: Some(tenant_conf.heatmap_period),
5457 220 : lazy_slru_download: Some(tenant_conf.lazy_slru_download),
5458 220 : timeline_get_throttle: Some(tenant_conf.timeline_get_throttle),
5459 220 : image_layer_creation_check_threshold: Some(
5460 220 : tenant_conf.image_layer_creation_check_threshold,
5461 220 : ),
5462 220 : lsn_lease_length: Some(tenant_conf.lsn_lease_length),
5463 220 : lsn_lease_length_for_ts: Some(tenant_conf.lsn_lease_length_for_ts),
5464 220 : timeline_offloading: Some(tenant_conf.timeline_offloading),
5465 220 : wal_receiver_protocol_override: tenant_conf.wal_receiver_protocol_override,
5466 220 : }
5467 220 : }
5468 : }
5469 :
5470 : pub struct TenantHarness {
5471 : pub conf: &'static PageServerConf,
5472 : pub tenant_conf: TenantConf,
5473 : pub tenant_shard_id: TenantShardId,
5474 : pub generation: Generation,
5475 : pub shard: ShardIndex,
5476 : pub remote_storage: GenericRemoteStorage,
5477 : pub remote_fs_dir: Utf8PathBuf,
5478 : pub deletion_queue: MockDeletionQueue,
5479 : }
5480 :
5481 : static LOG_HANDLE: OnceCell<()> = OnceCell::new();
5482 :
5483 236 : pub(crate) fn setup_logging() {
5484 236 : LOG_HANDLE.get_or_init(|| {
5485 224 : logging::init(
5486 224 : logging::LogFormat::Test,
5487 224 : // enable it in case the tests exercise code paths that use
5488 224 : // debug_assert_current_span_has_tenant_and_timeline_id
5489 224 : logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
5490 224 : logging::Output::Stdout,
5491 224 : )
5492 224 : .expect("Failed to init test logging")
5493 236 : });
5494 236 : }
5495 :
5496 : impl TenantHarness {
5497 220 : pub async fn create_custom(
5498 220 : test_name: &'static str,
5499 220 : tenant_conf: TenantConf,
5500 220 : tenant_id: TenantId,
5501 220 : shard_identity: ShardIdentity,
5502 220 : generation: Generation,
5503 220 : ) -> anyhow::Result<Self> {
5504 220 : setup_logging();
5505 220 :
5506 220 : let repo_dir = PageServerConf::test_repo_dir(test_name);
5507 220 : let _ = fs::remove_dir_all(&repo_dir);
5508 220 : fs::create_dir_all(&repo_dir)?;
5509 :
5510 220 : let conf = PageServerConf::dummy_conf(repo_dir);
5511 220 : // Make a static copy of the config. This can never be free'd, but that's
5512 220 : // OK in a test.
5513 220 : let conf: &'static PageServerConf = Box::leak(Box::new(conf));
5514 220 :
5515 220 : let shard = shard_identity.shard_index();
5516 220 : let tenant_shard_id = TenantShardId {
5517 220 : tenant_id,
5518 220 : shard_number: shard.shard_number,
5519 220 : shard_count: shard.shard_count,
5520 220 : };
5521 220 : fs::create_dir_all(conf.tenant_path(&tenant_shard_id))?;
5522 220 : fs::create_dir_all(conf.timelines_path(&tenant_shard_id))?;
5523 :
5524 : use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
5525 220 : let remote_fs_dir = conf.workdir.join("localfs");
5526 220 : std::fs::create_dir_all(&remote_fs_dir).unwrap();
5527 220 : let config = RemoteStorageConfig {
5528 220 : storage: RemoteStorageKind::LocalFs {
5529 220 : local_path: remote_fs_dir.clone(),
5530 220 : },
5531 220 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
5532 220 : small_timeout: RemoteStorageConfig::DEFAULT_SMALL_TIMEOUT,
5533 220 : };
5534 220 : let remote_storage = GenericRemoteStorage::from_config(&config).await.unwrap();
5535 220 : let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()));
5536 220 :
5537 220 : Ok(Self {
5538 220 : conf,
5539 220 : tenant_conf,
5540 220 : tenant_shard_id,
5541 220 : generation,
5542 220 : shard,
5543 220 : remote_storage,
5544 220 : remote_fs_dir,
5545 220 : deletion_queue,
5546 220 : })
5547 220 : }
5548 :
5549 208 : pub async fn create(test_name: &'static str) -> anyhow::Result<Self> {
5550 208 : // Disable automatic GC and compaction to make the unit tests more deterministic.
5551 208 : // The tests perform them manually if needed.
5552 208 : let tenant_conf = TenantConf {
5553 208 : gc_period: Duration::ZERO,
5554 208 : compaction_period: Duration::ZERO,
5555 208 : ..TenantConf::default()
5556 208 : };
5557 208 : let tenant_id = TenantId::generate();
5558 208 : let shard = ShardIdentity::unsharded();
5559 208 : Self::create_custom(
5560 208 : test_name,
5561 208 : tenant_conf,
5562 208 : tenant_id,
5563 208 : shard,
5564 208 : Generation::new(0xdeadbeef),
5565 208 : )
5566 208 : .await
5567 208 : }
5568 :
5569 20 : pub fn span(&self) -> tracing::Span {
5570 20 : info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
5571 20 : }
5572 :
5573 220 : pub(crate) async fn load(&self) -> (Arc<Tenant>, RequestContext) {
5574 220 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
5575 220 : (
5576 220 : self.do_try_load(&ctx)
5577 220 : .await
5578 220 : .expect("failed to load test tenant"),
5579 220 : ctx,
5580 220 : )
5581 220 : }
5582 :
5583 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5584 : pub(crate) async fn do_try_load(
5585 : &self,
5586 : ctx: &RequestContext,
5587 : ) -> anyhow::Result<Arc<Tenant>> {
5588 : let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
5589 :
5590 : let tenant = Arc::new(Tenant::new(
5591 : TenantState::Attaching,
5592 : self.conf,
5593 : AttachedTenantConf::try_from(LocationConf::attached_single(
5594 : TenantConfOpt::from(self.tenant_conf.clone()),
5595 : self.generation,
5596 : &ShardParameters::default(),
5597 : ))
5598 : .unwrap(),
5599 : // This is a legacy/test code path: sharding isn't supported here.
5600 : ShardIdentity::unsharded(),
5601 : Some(walredo_mgr),
5602 : self.tenant_shard_id,
5603 : self.remote_storage.clone(),
5604 : self.deletion_queue.new_client(),
5605 : // TODO: ideally we should run all unit tests with both configs
5606 : L0FlushGlobalState::new(L0FlushConfig::default()),
5607 : ));
5608 :
5609 : let preload = tenant
5610 : .preload(&self.remote_storage, CancellationToken::new())
5611 : .await?;
5612 : tenant.attach(Some(preload), ctx).await?;
5613 :
5614 : tenant.state.send_replace(TenantState::Active);
5615 : for timeline in tenant.timelines.lock().unwrap().values() {
5616 : timeline.set_state(TimelineState::Active);
5617 : }
5618 : Ok(tenant)
5619 : }
5620 :
5621 2 : pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
5622 2 : self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
5623 2 : }
5624 : }
5625 :
5626 : // Mock WAL redo manager that doesn't do much
5627 : pub(crate) struct TestRedoManager;
5628 :
5629 : impl TestRedoManager {
5630 : /// # Cancel-Safety
5631 : ///
5632 : /// This method is cancellation-safe.
5633 818 : pub async fn request_redo(
5634 818 : &self,
5635 818 : key: Key,
5636 818 : lsn: Lsn,
5637 818 : base_img: Option<(Lsn, Bytes)>,
5638 818 : records: Vec<(Lsn, NeonWalRecord)>,
5639 818 : _pg_version: u32,
5640 818 : ) -> Result<Bytes, walredo::Error> {
5641 1196 : let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
5642 818 : if records_neon {
5643 : // For Neon wal records, we can decode without spawning postgres, so do so.
5644 818 : let mut page = match (base_img, records.first()) {
5645 752 : (Some((_lsn, img)), _) => {
5646 752 : let mut page = BytesMut::new();
5647 752 : page.extend_from_slice(&img);
5648 752 : page
5649 : }
5650 66 : (_, Some((_lsn, rec))) if rec.will_init() => BytesMut::new(),
5651 : _ => {
5652 0 : panic!("Neon WAL redo requires base image or will init record");
5653 : }
5654 : };
5655 :
5656 2014 : for (record_lsn, record) in records {
5657 1196 : apply_neon::apply_in_neon(&record, record_lsn, key, &mut page)?;
5658 : }
5659 818 : Ok(page.freeze())
5660 : } else {
5661 : // We never spawn a postgres walredo process in unit tests: just log what we might have done.
5662 0 : let s = format!(
5663 0 : "redo for {} to get to {}, with {} and {} records",
5664 0 : key,
5665 0 : lsn,
5666 0 : if base_img.is_some() {
5667 0 : "base image"
5668 : } else {
5669 0 : "no base image"
5670 : },
5671 0 : records.len()
5672 0 : );
5673 0 : println!("{s}");
5674 0 :
5675 0 : Ok(test_img(&s))
5676 : }
5677 818 : }
5678 : }
5679 : }
5680 :
5681 : #[cfg(test)]
5682 : mod tests {
5683 : use std::collections::{BTreeMap, BTreeSet};
5684 :
5685 : use super::*;
5686 : use crate::keyspace::KeySpaceAccum;
5687 : use crate::tenant::harness::*;
5688 : use crate::tenant::timeline::CompactFlags;
5689 : use crate::DEFAULT_PG_VERSION;
5690 : use bytes::{Bytes, BytesMut};
5691 : use hex_literal::hex;
5692 : use itertools::Itertools;
5693 : use pageserver_api::key::{Key, AUX_KEY_PREFIX, NON_INHERITED_RANGE, RELATION_SIZE_PREFIX};
5694 : use pageserver_api::keyspace::KeySpace;
5695 : use pageserver_api::models::{CompactionAlgorithm, CompactionAlgorithmSettings};
5696 : use pageserver_api::value::Value;
5697 : use pageserver_compaction::helpers::overlaps_with;
5698 : use rand::{thread_rng, Rng};
5699 : use storage_layer::PersistentLayerKey;
5700 : use tests::storage_layer::ValuesReconstructState;
5701 : use tests::timeline::{GetVectoredError, ShutdownMode};
5702 : use timeline::{CompactOptions, DeltaLayerTestDesc};
5703 : use utils::id::TenantId;
5704 :
5705 : #[cfg(feature = "testing")]
5706 : use models::CompactLsnRange;
5707 : #[cfg(feature = "testing")]
5708 : use pageserver_api::record::NeonWalRecord;
5709 : #[cfg(feature = "testing")]
5710 : use timeline::compaction::{KeyHistoryRetention, KeyLogAtLsn};
5711 : #[cfg(feature = "testing")]
5712 : use timeline::GcInfo;
5713 :
5714 : static TEST_KEY: Lazy<Key> =
5715 18 : Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));
5716 :
5717 : #[tokio::test]
5718 2 : async fn test_basic() -> anyhow::Result<()> {
5719 2 : let (tenant, ctx) = TenantHarness::create("test_basic").await?.load().await;
5720 2 : let tline = tenant
5721 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5722 2 : .await?;
5723 2 :
5724 2 : let mut writer = tline.writer().await;
5725 2 : writer
5726 2 : .put(
5727 2 : *TEST_KEY,
5728 2 : Lsn(0x10),
5729 2 : &Value::Image(test_img("foo at 0x10")),
5730 2 : &ctx,
5731 2 : )
5732 2 : .await?;
5733 2 : writer.finish_write(Lsn(0x10));
5734 2 : drop(writer);
5735 2 :
5736 2 : let mut writer = tline.writer().await;
5737 2 : writer
5738 2 : .put(
5739 2 : *TEST_KEY,
5740 2 : Lsn(0x20),
5741 2 : &Value::Image(test_img("foo at 0x20")),
5742 2 : &ctx,
5743 2 : )
5744 2 : .await?;
5745 2 : writer.finish_write(Lsn(0x20));
5746 2 : drop(writer);
5747 2 :
5748 2 : assert_eq!(
5749 2 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
5750 2 : test_img("foo at 0x10")
5751 2 : );
5752 2 : assert_eq!(
5753 2 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
5754 2 : test_img("foo at 0x10")
5755 2 : );
5756 2 : assert_eq!(
5757 2 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
5758 2 : test_img("foo at 0x20")
5759 2 : );
5760 2 :
5761 2 : Ok(())
5762 2 : }
5763 :
5764 : #[tokio::test]
5765 2 : async fn no_duplicate_timelines() -> anyhow::Result<()> {
5766 2 : let (tenant, ctx) = TenantHarness::create("no_duplicate_timelines")
5767 2 : .await?
5768 2 : .load()
5769 2 : .await;
5770 2 : let _ = tenant
5771 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5772 2 : .await?;
5773 2 :
5774 2 : match tenant
5775 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5776 2 : .await
5777 2 : {
5778 2 : Ok(_) => panic!("duplicate timeline creation should fail"),
5779 2 : Err(e) => assert_eq!(
5780 2 : e.to_string(),
5781 2 : "timeline already exists with different parameters".to_string()
5782 2 : ),
5783 2 : }
5784 2 :
5785 2 : Ok(())
5786 2 : }
5787 :
5788 : /// Convenience function to create a page image with given string as the only content
5789 10 : pub fn test_value(s: &str) -> Value {
5790 10 : let mut buf = BytesMut::new();
5791 10 : buf.extend_from_slice(s.as_bytes());
5792 10 : Value::Image(buf.freeze())
5793 10 : }
5794 :
5795 : ///
5796 : /// Test branch creation
5797 : ///
5798 : #[tokio::test]
5799 2 : async fn test_branch() -> anyhow::Result<()> {
5800 2 : use std::str::from_utf8;
5801 2 :
5802 2 : let (tenant, ctx) = TenantHarness::create("test_branch").await?.load().await;
5803 2 : let tline = tenant
5804 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5805 2 : .await?;
5806 2 : let mut writer = tline.writer().await;
5807 2 :
5808 2 : #[allow(non_snake_case)]
5809 2 : let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
5810 2 : #[allow(non_snake_case)]
5811 2 : let TEST_KEY_B: Key = Key::from_hex("110000000033333333444444445500000002").unwrap();
5812 2 :
5813 2 : // Insert a value on the timeline
5814 2 : writer
5815 2 : .put(TEST_KEY_A, Lsn(0x20), &test_value("foo at 0x20"), &ctx)
5816 2 : .await?;
5817 2 : writer
5818 2 : .put(TEST_KEY_B, Lsn(0x20), &test_value("foobar at 0x20"), &ctx)
5819 2 : .await?;
5820 2 : writer.finish_write(Lsn(0x20));
5821 2 :
5822 2 : writer
5823 2 : .put(TEST_KEY_A, Lsn(0x30), &test_value("foo at 0x30"), &ctx)
5824 2 : .await?;
5825 2 : writer.finish_write(Lsn(0x30));
5826 2 : writer
5827 2 : .put(TEST_KEY_A, Lsn(0x40), &test_value("foo at 0x40"), &ctx)
5828 2 : .await?;
5829 2 : writer.finish_write(Lsn(0x40));
5830 2 :
5831 2 : //assert_current_logical_size(&tline, Lsn(0x40));
5832 2 :
5833 2 : // Branch the history, modify relation differently on the new timeline
5834 2 : tenant
5835 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x30)), &ctx)
5836 2 : .await?;
5837 2 : let newtline = tenant
5838 2 : .get_timeline(NEW_TIMELINE_ID, true)
5839 2 : .expect("Should have a local timeline");
5840 2 : let mut new_writer = newtline.writer().await;
5841 2 : new_writer
5842 2 : .put(TEST_KEY_A, Lsn(0x40), &test_value("bar at 0x40"), &ctx)
5843 2 : .await?;
5844 2 : new_writer.finish_write(Lsn(0x40));
5845 2 :
5846 2 : // Check page contents on both branches
5847 2 : assert_eq!(
5848 2 : from_utf8(&tline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
5849 2 : "foo at 0x40"
5850 2 : );
5851 2 : assert_eq!(
5852 2 : from_utf8(&newtline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
5853 2 : "bar at 0x40"
5854 2 : );
5855 2 : assert_eq!(
5856 2 : from_utf8(&newtline.get(TEST_KEY_B, Lsn(0x40), &ctx).await?)?,
5857 2 : "foobar at 0x20"
5858 2 : );
5859 2 :
5860 2 : //assert_current_logical_size(&tline, Lsn(0x40));
5861 2 :
5862 2 : Ok(())
5863 2 : }
5864 :
5865 20 : async fn make_some_layers(
5866 20 : tline: &Timeline,
5867 20 : start_lsn: Lsn,
5868 20 : ctx: &RequestContext,
5869 20 : ) -> anyhow::Result<()> {
5870 20 : let mut lsn = start_lsn;
5871 : {
5872 20 : let mut writer = tline.writer().await;
5873 : // Create a relation on the timeline
5874 20 : writer
5875 20 : .put(
5876 20 : *TEST_KEY,
5877 20 : lsn,
5878 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5879 20 : ctx,
5880 20 : )
5881 20 : .await?;
5882 20 : writer.finish_write(lsn);
5883 20 : lsn += 0x10;
5884 20 : writer
5885 20 : .put(
5886 20 : *TEST_KEY,
5887 20 : lsn,
5888 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5889 20 : ctx,
5890 20 : )
5891 20 : .await?;
5892 20 : writer.finish_write(lsn);
5893 20 : lsn += 0x10;
5894 20 : }
5895 20 : tline.freeze_and_flush().await?;
5896 : {
5897 20 : let mut writer = tline.writer().await;
5898 20 : writer
5899 20 : .put(
5900 20 : *TEST_KEY,
5901 20 : lsn,
5902 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5903 20 : ctx,
5904 20 : )
5905 20 : .await?;
5906 20 : writer.finish_write(lsn);
5907 20 : lsn += 0x10;
5908 20 : writer
5909 20 : .put(
5910 20 : *TEST_KEY,
5911 20 : lsn,
5912 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5913 20 : ctx,
5914 20 : )
5915 20 : .await?;
5916 20 : writer.finish_write(lsn);
5917 20 : }
5918 20 : tline.freeze_and_flush().await.map_err(|e| e.into())
5919 20 : }
5920 :
5921 : #[tokio::test(start_paused = true)]
5922 2 : async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
5923 2 : let (tenant, ctx) =
5924 2 : TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
5925 2 : .await?
5926 2 : .load()
5927 2 : .await;
5928 2 : // Advance to the lsn lease deadline so that GC is not blocked by
5929 2 : // initial transition into AttachedSingle.
5930 2 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
5931 2 : tokio::time::resume();
5932 2 : let tline = tenant
5933 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5934 2 : .await?;
5935 2 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5936 2 :
5937 2 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
5938 2 : // FIXME: this doesn't actually remove any layer currently, given how the flushing
5939 2 : // and compaction works. But it does set the 'cutoff' point so that the cross check
5940 2 : // below should fail.
5941 2 : tenant
5942 2 : .gc_iteration(
5943 2 : Some(TIMELINE_ID),
5944 2 : 0x10,
5945 2 : Duration::ZERO,
5946 2 : &CancellationToken::new(),
5947 2 : &ctx,
5948 2 : )
5949 2 : .await?;
5950 2 :
5951 2 : // try to branch at lsn 25, should fail because we already garbage collected the data
5952 2 : match tenant
5953 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
5954 2 : .await
5955 2 : {
5956 2 : Ok(_) => panic!("branching should have failed"),
5957 2 : Err(err) => {
5958 2 : let CreateTimelineError::AncestorLsn(err) = err else {
5959 2 : panic!("wrong error type")
5960 2 : };
5961 2 : assert!(err.to_string().contains("invalid branch start lsn"));
5962 2 : assert!(err
5963 2 : .source()
5964 2 : .unwrap()
5965 2 : .to_string()
5966 2 : .contains("we might've already garbage collected needed data"))
5967 2 : }
5968 2 : }
5969 2 :
5970 2 : Ok(())
5971 2 : }
5972 :
5973 : #[tokio::test]
5974 2 : async fn test_prohibit_branch_creation_on_pre_initdb_lsn() -> anyhow::Result<()> {
5975 2 : let (tenant, ctx) =
5976 2 : TenantHarness::create("test_prohibit_branch_creation_on_pre_initdb_lsn")
5977 2 : .await?
5978 2 : .load()
5979 2 : .await;
5980 2 :
5981 2 : let tline = tenant
5982 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x50), DEFAULT_PG_VERSION, &ctx)
5983 2 : .await?;
5984 2 : // try to branch at lsn 0x25, should fail because initdb lsn is 0x50
5985 2 : match tenant
5986 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
5987 2 : .await
5988 2 : {
5989 2 : Ok(_) => panic!("branching should have failed"),
5990 2 : Err(err) => {
5991 2 : let CreateTimelineError::AncestorLsn(err) = err else {
5992 2 : panic!("wrong error type");
5993 2 : };
5994 2 : assert!(&err.to_string().contains("invalid branch start lsn"));
5995 2 : assert!(&err
5996 2 : .source()
5997 2 : .unwrap()
5998 2 : .to_string()
5999 2 : .contains("is earlier than latest GC cutoff"));
6000 2 : }
6001 2 : }
6002 2 :
6003 2 : Ok(())
6004 2 : }
6005 :
6006 : /*
6007 : // FIXME: This currently fails to error out. Calling GC doesn't currently
6008 : // remove the old value, we'd need to work a little harder
6009 : #[tokio::test]
6010 : async fn test_prohibit_get_for_garbage_collected_data() -> anyhow::Result<()> {
6011 : let repo =
6012 : RepoHarness::create("test_prohibit_get_for_garbage_collected_data")?
6013 : .load();
6014 :
6015 : let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION)?;
6016 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6017 :
6018 : repo.gc_iteration(Some(TIMELINE_ID), 0x10, Duration::ZERO)?;
6019 : let latest_gc_cutoff_lsn = tline.get_latest_gc_cutoff_lsn();
6020 : assert!(*latest_gc_cutoff_lsn > Lsn(0x25));
6021 : match tline.get(*TEST_KEY, Lsn(0x25)) {
6022 : Ok(_) => panic!("request for page should have failed"),
6023 : Err(err) => assert!(err.to_string().contains("not found at")),
6024 : }
6025 : Ok(())
6026 : }
6027 : */
6028 :
6029 : #[tokio::test]
6030 2 : async fn test_get_branchpoints_from_an_inactive_timeline() -> anyhow::Result<()> {
6031 2 : let (tenant, ctx) =
6032 2 : TenantHarness::create("test_get_branchpoints_from_an_inactive_timeline")
6033 2 : .await?
6034 2 : .load()
6035 2 : .await;
6036 2 : let tline = tenant
6037 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6038 2 : .await?;
6039 2 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6040 2 :
6041 2 : tenant
6042 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6043 2 : .await?;
6044 2 : let newtline = tenant
6045 2 : .get_timeline(NEW_TIMELINE_ID, true)
6046 2 : .expect("Should have a local timeline");
6047 2 :
6048 2 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6049 2 :
6050 2 : tline.set_broken("test".to_owned());
6051 2 :
6052 2 : tenant
6053 2 : .gc_iteration(
6054 2 : Some(TIMELINE_ID),
6055 2 : 0x10,
6056 2 : Duration::ZERO,
6057 2 : &CancellationToken::new(),
6058 2 : &ctx,
6059 2 : )
6060 2 : .await?;
6061 2 :
6062 2 : // The branchpoints should contain all timelines, even ones marked
6063 2 : // as Broken.
6064 2 : {
6065 2 : let branchpoints = &tline.gc_info.read().unwrap().retain_lsns;
6066 2 : assert_eq!(branchpoints.len(), 1);
6067 2 : assert_eq!(
6068 2 : branchpoints[0],
6069 2 : (Lsn(0x40), NEW_TIMELINE_ID, MaybeOffloaded::No)
6070 2 : );
6071 2 : }
6072 2 :
6073 2 : // You can read the key from the child branch even though the parent is
6074 2 : // Broken, as long as you don't need to access data from the parent.
6075 2 : assert_eq!(
6076 2 : newtline.get(*TEST_KEY, Lsn(0x70), &ctx).await?,
6077 2 : test_img(&format!("foo at {}", Lsn(0x70)))
6078 2 : );
6079 2 :
6080 2 : // This needs to traverse to the parent, and fails.
6081 2 : let err = newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await.unwrap_err();
6082 2 : assert!(
6083 2 : err.to_string().starts_with(&format!(
6084 2 : "bad state on timeline {}: Broken",
6085 2 : tline.timeline_id
6086 2 : )),
6087 2 : "{err}"
6088 2 : );
6089 2 :
6090 2 : Ok(())
6091 2 : }
6092 :
6093 : #[tokio::test]
6094 2 : async fn test_retain_data_in_parent_which_is_needed_for_child() -> anyhow::Result<()> {
6095 2 : let (tenant, ctx) =
6096 2 : TenantHarness::create("test_retain_data_in_parent_which_is_needed_for_child")
6097 2 : .await?
6098 2 : .load()
6099 2 : .await;
6100 2 : let tline = tenant
6101 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6102 2 : .await?;
6103 2 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6104 2 :
6105 2 : tenant
6106 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6107 2 : .await?;
6108 2 : let newtline = tenant
6109 2 : .get_timeline(NEW_TIMELINE_ID, true)
6110 2 : .expect("Should have a local timeline");
6111 2 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
6112 2 : tenant
6113 2 : .gc_iteration(
6114 2 : Some(TIMELINE_ID),
6115 2 : 0x10,
6116 2 : Duration::ZERO,
6117 2 : &CancellationToken::new(),
6118 2 : &ctx,
6119 2 : )
6120 2 : .await?;
6121 2 : assert!(newtline.get(*TEST_KEY, Lsn(0x25), &ctx).await.is_ok());
6122 2 :
6123 2 : Ok(())
6124 2 : }
6125 : #[tokio::test]
6126 2 : async fn test_parent_keeps_data_forever_after_branching() -> anyhow::Result<()> {
6127 2 : let (tenant, ctx) = TenantHarness::create("test_parent_keeps_data_forever_after_branching")
6128 2 : .await?
6129 2 : .load()
6130 2 : .await;
6131 2 : let tline = tenant
6132 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6133 2 : .await?;
6134 2 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6135 2 :
6136 2 : tenant
6137 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6138 2 : .await?;
6139 2 : let newtline = tenant
6140 2 : .get_timeline(NEW_TIMELINE_ID, true)
6141 2 : .expect("Should have a local timeline");
6142 2 :
6143 2 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6144 2 :
6145 2 : // run gc on parent
6146 2 : tenant
6147 2 : .gc_iteration(
6148 2 : Some(TIMELINE_ID),
6149 2 : 0x10,
6150 2 : Duration::ZERO,
6151 2 : &CancellationToken::new(),
6152 2 : &ctx,
6153 2 : )
6154 2 : .await?;
6155 2 :
6156 2 : // Check that the data is still accessible on the branch.
6157 2 : assert_eq!(
6158 2 : newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
6159 2 : test_img(&format!("foo at {}", Lsn(0x40)))
6160 2 : );
6161 2 :
6162 2 : Ok(())
6163 2 : }
6164 :
6165 : #[tokio::test]
6166 2 : async fn timeline_load() -> anyhow::Result<()> {
6167 2 : const TEST_NAME: &str = "timeline_load";
6168 2 : let harness = TenantHarness::create(TEST_NAME).await?;
6169 2 : {
6170 2 : let (tenant, ctx) = harness.load().await;
6171 2 : let tline = tenant
6172 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x7000), DEFAULT_PG_VERSION, &ctx)
6173 2 : .await?;
6174 2 : make_some_layers(tline.as_ref(), Lsn(0x8000), &ctx).await?;
6175 2 : // so that all uploads finish & we can call harness.load() below again
6176 2 : tenant
6177 2 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
6178 2 : .instrument(harness.span())
6179 2 : .await
6180 2 : .ok()
6181 2 : .unwrap();
6182 2 : }
6183 2 :
6184 2 : let (tenant, _ctx) = harness.load().await;
6185 2 : tenant
6186 2 : .get_timeline(TIMELINE_ID, true)
6187 2 : .expect("cannot load timeline");
6188 2 :
6189 2 : Ok(())
6190 2 : }
6191 :
6192 : #[tokio::test]
6193 2 : async fn timeline_load_with_ancestor() -> anyhow::Result<()> {
6194 2 : const TEST_NAME: &str = "timeline_load_with_ancestor";
6195 2 : let harness = TenantHarness::create(TEST_NAME).await?;
6196 2 : // create two timelines
6197 2 : {
6198 2 : let (tenant, ctx) = harness.load().await;
6199 2 : let tline = tenant
6200 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6201 2 : .await?;
6202 2 :
6203 2 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6204 2 :
6205 2 : let child_tline = tenant
6206 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6207 2 : .await?;
6208 2 : child_tline.set_state(TimelineState::Active);
6209 2 :
6210 2 : let newtline = tenant
6211 2 : .get_timeline(NEW_TIMELINE_ID, true)
6212 2 : .expect("Should have a local timeline");
6213 2 :
6214 2 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6215 2 :
6216 2 : // so that all uploads finish & we can call harness.load() below again
6217 2 : tenant
6218 2 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
6219 2 : .instrument(harness.span())
6220 2 : .await
6221 2 : .ok()
6222 2 : .unwrap();
6223 2 : }
6224 2 :
6225 2 : // check that both of them are initially unloaded
6226 2 : let (tenant, _ctx) = harness.load().await;
6227 2 :
6228 2 : // check that both, child and ancestor are loaded
6229 2 : let _child_tline = tenant
6230 2 : .get_timeline(NEW_TIMELINE_ID, true)
6231 2 : .expect("cannot get child timeline loaded");
6232 2 :
6233 2 : let _ancestor_tline = tenant
6234 2 : .get_timeline(TIMELINE_ID, true)
6235 2 : .expect("cannot get ancestor timeline loaded");
6236 2 :
6237 2 : Ok(())
6238 2 : }
6239 :
6240 : #[tokio::test]
6241 2 : async fn delta_layer_dumping() -> anyhow::Result<()> {
6242 2 : use storage_layer::AsLayerDesc;
6243 2 : let (tenant, ctx) = TenantHarness::create("test_layer_dumping")
6244 2 : .await?
6245 2 : .load()
6246 2 : .await;
6247 2 : let tline = tenant
6248 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6249 2 : .await?;
6250 2 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6251 2 :
6252 2 : let layer_map = tline.layers.read().await;
6253 2 : let level0_deltas = layer_map
6254 2 : .layer_map()?
6255 2 : .level0_deltas()
6256 2 : .iter()
6257 4 : .map(|desc| layer_map.get_from_desc(desc))
6258 2 : .collect::<Vec<_>>();
6259 2 :
6260 2 : assert!(!level0_deltas.is_empty());
6261 2 :
6262 6 : for delta in level0_deltas {
6263 2 : // Ensure we are dumping a delta layer here
6264 4 : assert!(delta.layer_desc().is_delta);
6265 4 : delta.dump(true, &ctx).await.unwrap();
6266 2 : }
6267 2 :
6268 2 : Ok(())
6269 2 : }
6270 :
6271 : #[tokio::test]
6272 2 : async fn test_images() -> anyhow::Result<()> {
6273 2 : let (tenant, ctx) = TenantHarness::create("test_images").await?.load().await;
6274 2 : let tline = tenant
6275 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6276 2 : .await?;
6277 2 :
6278 2 : let mut writer = tline.writer().await;
6279 2 : writer
6280 2 : .put(
6281 2 : *TEST_KEY,
6282 2 : Lsn(0x10),
6283 2 : &Value::Image(test_img("foo at 0x10")),
6284 2 : &ctx,
6285 2 : )
6286 2 : .await?;
6287 2 : writer.finish_write(Lsn(0x10));
6288 2 : drop(writer);
6289 2 :
6290 2 : tline.freeze_and_flush().await?;
6291 2 : tline
6292 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6293 2 : .await?;
6294 2 :
6295 2 : let mut writer = tline.writer().await;
6296 2 : writer
6297 2 : .put(
6298 2 : *TEST_KEY,
6299 2 : Lsn(0x20),
6300 2 : &Value::Image(test_img("foo at 0x20")),
6301 2 : &ctx,
6302 2 : )
6303 2 : .await?;
6304 2 : writer.finish_write(Lsn(0x20));
6305 2 : drop(writer);
6306 2 :
6307 2 : tline.freeze_and_flush().await?;
6308 2 : tline
6309 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6310 2 : .await?;
6311 2 :
6312 2 : let mut writer = tline.writer().await;
6313 2 : writer
6314 2 : .put(
6315 2 : *TEST_KEY,
6316 2 : Lsn(0x30),
6317 2 : &Value::Image(test_img("foo at 0x30")),
6318 2 : &ctx,
6319 2 : )
6320 2 : .await?;
6321 2 : writer.finish_write(Lsn(0x30));
6322 2 : drop(writer);
6323 2 :
6324 2 : tline.freeze_and_flush().await?;
6325 2 : tline
6326 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6327 2 : .await?;
6328 2 :
6329 2 : let mut writer = tline.writer().await;
6330 2 : writer
6331 2 : .put(
6332 2 : *TEST_KEY,
6333 2 : Lsn(0x40),
6334 2 : &Value::Image(test_img("foo at 0x40")),
6335 2 : &ctx,
6336 2 : )
6337 2 : .await?;
6338 2 : writer.finish_write(Lsn(0x40));
6339 2 : drop(writer);
6340 2 :
6341 2 : tline.freeze_and_flush().await?;
6342 2 : tline
6343 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6344 2 : .await?;
6345 2 :
6346 2 : assert_eq!(
6347 2 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
6348 2 : test_img("foo at 0x10")
6349 2 : );
6350 2 : assert_eq!(
6351 2 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
6352 2 : test_img("foo at 0x10")
6353 2 : );
6354 2 : assert_eq!(
6355 2 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
6356 2 : test_img("foo at 0x20")
6357 2 : );
6358 2 : assert_eq!(
6359 2 : tline.get(*TEST_KEY, Lsn(0x30), &ctx).await?,
6360 2 : test_img("foo at 0x30")
6361 2 : );
6362 2 : assert_eq!(
6363 2 : tline.get(*TEST_KEY, Lsn(0x40), &ctx).await?,
6364 2 : test_img("foo at 0x40")
6365 2 : );
6366 2 :
6367 2 : Ok(())
6368 2 : }
6369 :
6370 4 : async fn bulk_insert_compact_gc(
6371 4 : tenant: &Tenant,
6372 4 : timeline: &Arc<Timeline>,
6373 4 : ctx: &RequestContext,
6374 4 : lsn: Lsn,
6375 4 : repeat: usize,
6376 4 : key_count: usize,
6377 4 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
6378 4 : let compact = true;
6379 4 : bulk_insert_maybe_compact_gc(tenant, timeline, ctx, lsn, repeat, key_count, compact).await
6380 4 : }
6381 :
6382 8 : async fn bulk_insert_maybe_compact_gc(
6383 8 : tenant: &Tenant,
6384 8 : timeline: &Arc<Timeline>,
6385 8 : ctx: &RequestContext,
6386 8 : mut lsn: Lsn,
6387 8 : repeat: usize,
6388 8 : key_count: usize,
6389 8 : compact: bool,
6390 8 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
6391 8 : let mut inserted: HashMap<Key, BTreeSet<Lsn>> = Default::default();
6392 8 :
6393 8 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6394 8 : let mut blknum = 0;
6395 8 :
6396 8 : // Enforce that key range is monotonously increasing
6397 8 : let mut keyspace = KeySpaceAccum::new();
6398 8 :
6399 8 : let cancel = CancellationToken::new();
6400 8 :
6401 8 : for _ in 0..repeat {
6402 400 : for _ in 0..key_count {
6403 4000000 : test_key.field6 = blknum;
6404 4000000 : let mut writer = timeline.writer().await;
6405 4000000 : writer
6406 4000000 : .put(
6407 4000000 : test_key,
6408 4000000 : lsn,
6409 4000000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6410 4000000 : ctx,
6411 4000000 : )
6412 4000000 : .await?;
6413 4000000 : inserted.entry(test_key).or_default().insert(lsn);
6414 4000000 : writer.finish_write(lsn);
6415 4000000 : drop(writer);
6416 4000000 :
6417 4000000 : keyspace.add_key(test_key);
6418 4000000 :
6419 4000000 : lsn = Lsn(lsn.0 + 0x10);
6420 4000000 : blknum += 1;
6421 : }
6422 :
6423 400 : timeline.freeze_and_flush().await?;
6424 400 : if compact {
6425 : // this requires timeline to be &Arc<Timeline>
6426 200 : timeline.compact(&cancel, EnumSet::empty(), ctx).await?;
6427 200 : }
6428 :
6429 : // this doesn't really need to use the timeline_id target, but it is closer to what it
6430 : // originally was.
6431 400 : let res = tenant
6432 400 : .gc_iteration(Some(timeline.timeline_id), 0, Duration::ZERO, &cancel, ctx)
6433 400 : .await?;
6434 :
6435 400 : assert_eq!(res.layers_removed, 0, "this never removes anything");
6436 : }
6437 :
6438 8 : Ok(inserted)
6439 8 : }
6440 :
6441 : //
6442 : // Insert 1000 key-value pairs with increasing keys, flush, compact, GC.
6443 : // Repeat 50 times.
6444 : //
6445 : #[tokio::test]
6446 2 : async fn test_bulk_insert() -> anyhow::Result<()> {
6447 2 : let harness = TenantHarness::create("test_bulk_insert").await?;
6448 2 : let (tenant, ctx) = harness.load().await;
6449 2 : let tline = tenant
6450 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6451 2 : .await?;
6452 2 :
6453 2 : let lsn = Lsn(0x10);
6454 2 : bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
6455 2 :
6456 2 : Ok(())
6457 2 : }
6458 :
6459 : // Test the vectored get real implementation against a simple sequential implementation.
6460 : //
6461 : // The test generates a keyspace by repeatedly flushing the in-memory layer and compacting.
6462 : // Projected to 2D the key space looks like below. Lsn grows upwards on the Y axis and keys
6463 : // grow to the right on the X axis.
6464 : // [Delta]
6465 : // [Delta]
6466 : // [Delta]
6467 : // [Delta]
6468 : // ------------ Image ---------------
6469 : //
6470 : // After layer generation we pick the ranges to query as follows:
6471 : // 1. The beginning of each delta layer
6472 : // 2. At the seam between two adjacent delta layers
6473 : //
6474 : // There's one major downside to this test: delta layers only contains images,
6475 : // so the search can stop at the first delta layer and doesn't traverse any deeper.
6476 : #[tokio::test]
6477 2 : async fn test_get_vectored() -> anyhow::Result<()> {
6478 2 : let harness = TenantHarness::create("test_get_vectored").await?;
6479 2 : let (tenant, ctx) = harness.load().await;
6480 2 : let tline = tenant
6481 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6482 2 : .await?;
6483 2 :
6484 2 : let lsn = Lsn(0x10);
6485 2 : let inserted = bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
6486 2 :
6487 2 : let guard = tline.layers.read().await;
6488 2 : let lm = guard.layer_map()?;
6489 2 :
6490 2 : lm.dump(true, &ctx).await?;
6491 2 :
6492 2 : let mut reads = Vec::new();
6493 2 : let mut prev = None;
6494 12 : lm.iter_historic_layers().for_each(|desc| {
6495 12 : if !desc.is_delta() {
6496 2 : prev = Some(desc.clone());
6497 2 : return;
6498 10 : }
6499 10 :
6500 10 : let start = desc.key_range.start;
6501 10 : let end = desc
6502 10 : .key_range
6503 10 : .start
6504 10 : .add(Timeline::MAX_GET_VECTORED_KEYS.try_into().unwrap());
6505 10 : reads.push(KeySpace {
6506 10 : ranges: vec![start..end],
6507 10 : });
6508 2 :
6509 10 : if let Some(prev) = &prev {
6510 10 : if !prev.is_delta() {
6511 10 : return;
6512 2 : }
6513 0 :
6514 0 : let first_range = Key {
6515 0 : field6: prev.key_range.end.field6 - 4,
6516 0 : ..prev.key_range.end
6517 0 : }..prev.key_range.end;
6518 0 :
6519 0 : let second_range = desc.key_range.start..Key {
6520 0 : field6: desc.key_range.start.field6 + 4,
6521 0 : ..desc.key_range.start
6522 0 : };
6523 0 :
6524 0 : reads.push(KeySpace {
6525 0 : ranges: vec![first_range, second_range],
6526 0 : });
6527 2 : };
6528 2 :
6529 2 : prev = Some(desc.clone());
6530 12 : });
6531 2 :
6532 2 : drop(guard);
6533 2 :
6534 2 : // Pick a big LSN such that we query over all the changes.
6535 2 : let reads_lsn = Lsn(u64::MAX - 1);
6536 2 :
6537 12 : for read in reads {
6538 10 : info!("Doing vectored read on {:?}", read);
6539 2 :
6540 10 : let vectored_res = tline
6541 10 : .get_vectored_impl(
6542 10 : read.clone(),
6543 10 : reads_lsn,
6544 10 : &mut ValuesReconstructState::new(),
6545 10 : &ctx,
6546 10 : )
6547 10 : .await;
6548 2 :
6549 10 : let mut expected_lsns: HashMap<Key, Lsn> = Default::default();
6550 10 : let mut expect_missing = false;
6551 10 : let mut key = read.start().unwrap();
6552 330 : while key != read.end().unwrap() {
6553 320 : if let Some(lsns) = inserted.get(&key) {
6554 320 : let expected_lsn = lsns.iter().rfind(|lsn| **lsn <= reads_lsn);
6555 320 : match expected_lsn {
6556 320 : Some(lsn) => {
6557 320 : expected_lsns.insert(key, *lsn);
6558 320 : }
6559 2 : None => {
6560 2 : expect_missing = true;
6561 0 : break;
6562 2 : }
6563 2 : }
6564 2 : } else {
6565 2 : expect_missing = true;
6566 0 : break;
6567 2 : }
6568 2 :
6569 320 : key = key.next();
6570 2 : }
6571 2 :
6572 10 : if expect_missing {
6573 2 : assert!(matches!(vectored_res, Err(GetVectoredError::MissingKey(_))));
6574 2 : } else {
6575 320 : for (key, image) in vectored_res? {
6576 320 : let expected_lsn = expected_lsns.get(&key).expect("determined above");
6577 320 : let expected_image = test_img(&format!("{} at {}", key.field6, expected_lsn));
6578 320 : assert_eq!(image?, expected_image);
6579 2 : }
6580 2 : }
6581 2 : }
6582 2 :
6583 2 : Ok(())
6584 2 : }
6585 :
6586 : #[tokio::test]
6587 2 : async fn test_get_vectored_aux_files() -> anyhow::Result<()> {
6588 2 : let harness = TenantHarness::create("test_get_vectored_aux_files").await?;
6589 2 :
6590 2 : let (tenant, ctx) = harness.load().await;
6591 2 : let tline = tenant
6592 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
6593 2 : .await?;
6594 2 : let tline = tline.raw_timeline().unwrap();
6595 2 :
6596 2 : let mut modification = tline.begin_modification(Lsn(0x1000));
6597 2 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
6598 2 : modification.set_lsn(Lsn(0x1008))?;
6599 2 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
6600 2 : modification.commit(&ctx).await?;
6601 2 :
6602 2 : let child_timeline_id = TimelineId::generate();
6603 2 : tenant
6604 2 : .branch_timeline_test(
6605 2 : tline,
6606 2 : child_timeline_id,
6607 2 : Some(tline.get_last_record_lsn()),
6608 2 : &ctx,
6609 2 : )
6610 2 : .await?;
6611 2 :
6612 2 : let child_timeline = tenant
6613 2 : .get_timeline(child_timeline_id, true)
6614 2 : .expect("Should have the branched timeline");
6615 2 :
6616 2 : let aux_keyspace = KeySpace {
6617 2 : ranges: vec![NON_INHERITED_RANGE],
6618 2 : };
6619 2 : let read_lsn = child_timeline.get_last_record_lsn();
6620 2 :
6621 2 : let vectored_res = child_timeline
6622 2 : .get_vectored_impl(
6623 2 : aux_keyspace.clone(),
6624 2 : read_lsn,
6625 2 : &mut ValuesReconstructState::new(),
6626 2 : &ctx,
6627 2 : )
6628 2 : .await;
6629 2 :
6630 2 : let images = vectored_res?;
6631 2 : assert!(images.is_empty());
6632 2 : Ok(())
6633 2 : }
6634 :
6635 : // Test that vectored get handles layer gaps correctly
6636 : // by advancing into the next ancestor timeline if required.
6637 : //
6638 : // The test generates timelines that look like the diagram below.
6639 : // We leave a gap in one of the L1 layers at `gap_at_key` (`/` in the diagram).
6640 : // The reconstruct data for that key lies in the ancestor timeline (`X` in the diagram).
6641 : //
6642 : // ```
6643 : //-------------------------------+
6644 : // ... |
6645 : // [ L1 ] |
6646 : // [ / L1 ] | Child Timeline
6647 : // ... |
6648 : // ------------------------------+
6649 : // [ X L1 ] | Parent Timeline
6650 : // ------------------------------+
6651 : // ```
6652 : #[tokio::test]
6653 2 : async fn test_get_vectored_key_gap() -> anyhow::Result<()> {
6654 2 : let tenant_conf = TenantConf {
6655 2 : // Make compaction deterministic
6656 2 : gc_period: Duration::ZERO,
6657 2 : compaction_period: Duration::ZERO,
6658 2 : // Encourage creation of L1 layers
6659 2 : checkpoint_distance: 16 * 1024,
6660 2 : compaction_target_size: 8 * 1024,
6661 2 : ..TenantConf::default()
6662 2 : };
6663 2 :
6664 2 : let harness = TenantHarness::create_custom(
6665 2 : "test_get_vectored_key_gap",
6666 2 : tenant_conf,
6667 2 : TenantId::generate(),
6668 2 : ShardIdentity::unsharded(),
6669 2 : Generation::new(0xdeadbeef),
6670 2 : )
6671 2 : .await?;
6672 2 : let (tenant, ctx) = harness.load().await;
6673 2 :
6674 2 : let mut current_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6675 2 : let gap_at_key = current_key.add(100);
6676 2 : let mut current_lsn = Lsn(0x10);
6677 2 :
6678 2 : const KEY_COUNT: usize = 10_000;
6679 2 :
6680 2 : let timeline_id = TimelineId::generate();
6681 2 : let current_timeline = tenant
6682 2 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
6683 2 : .await?;
6684 2 :
6685 2 : current_lsn += 0x100;
6686 2 :
6687 2 : let mut writer = current_timeline.writer().await;
6688 2 : writer
6689 2 : .put(
6690 2 : gap_at_key,
6691 2 : current_lsn,
6692 2 : &Value::Image(test_img(&format!("{} at {}", gap_at_key, current_lsn))),
6693 2 : &ctx,
6694 2 : )
6695 2 : .await?;
6696 2 : writer.finish_write(current_lsn);
6697 2 : drop(writer);
6698 2 :
6699 2 : let mut latest_lsns = HashMap::new();
6700 2 : latest_lsns.insert(gap_at_key, current_lsn);
6701 2 :
6702 2 : current_timeline.freeze_and_flush().await?;
6703 2 :
6704 2 : let child_timeline_id = TimelineId::generate();
6705 2 :
6706 2 : tenant
6707 2 : .branch_timeline_test(
6708 2 : ¤t_timeline,
6709 2 : child_timeline_id,
6710 2 : Some(current_lsn),
6711 2 : &ctx,
6712 2 : )
6713 2 : .await?;
6714 2 : let child_timeline = tenant
6715 2 : .get_timeline(child_timeline_id, true)
6716 2 : .expect("Should have the branched timeline");
6717 2 :
6718 20002 : for i in 0..KEY_COUNT {
6719 20000 : if current_key == gap_at_key {
6720 2 : current_key = current_key.next();
6721 2 : continue;
6722 19998 : }
6723 19998 :
6724 19998 : current_lsn += 0x10;
6725 2 :
6726 19998 : let mut writer = child_timeline.writer().await;
6727 19998 : writer
6728 19998 : .put(
6729 19998 : current_key,
6730 19998 : current_lsn,
6731 19998 : &Value::Image(test_img(&format!("{} at {}", current_key, current_lsn))),
6732 19998 : &ctx,
6733 19998 : )
6734 19998 : .await?;
6735 19998 : writer.finish_write(current_lsn);
6736 19998 : drop(writer);
6737 19998 :
6738 19998 : latest_lsns.insert(current_key, current_lsn);
6739 19998 : current_key = current_key.next();
6740 19998 :
6741 19998 : // Flush every now and then to encourage layer file creation.
6742 19998 : if i % 500 == 0 {
6743 40 : child_timeline.freeze_and_flush().await?;
6744 19958 : }
6745 2 : }
6746 2 :
6747 2 : child_timeline.freeze_and_flush().await?;
6748 2 : let mut flags = EnumSet::new();
6749 2 : flags.insert(CompactFlags::ForceRepartition);
6750 2 : child_timeline
6751 2 : .compact(&CancellationToken::new(), flags, &ctx)
6752 2 : .await?;
6753 2 :
6754 2 : let key_near_end = {
6755 2 : let mut tmp = current_key;
6756 2 : tmp.field6 -= 10;
6757 2 : tmp
6758 2 : };
6759 2 :
6760 2 : let key_near_gap = {
6761 2 : let mut tmp = gap_at_key;
6762 2 : tmp.field6 -= 10;
6763 2 : tmp
6764 2 : };
6765 2 :
6766 2 : let read = KeySpace {
6767 2 : ranges: vec![key_near_gap..gap_at_key.next(), key_near_end..current_key],
6768 2 : };
6769 2 : let results = child_timeline
6770 2 : .get_vectored_impl(
6771 2 : read.clone(),
6772 2 : current_lsn,
6773 2 : &mut ValuesReconstructState::new(),
6774 2 : &ctx,
6775 2 : )
6776 2 : .await?;
6777 2 :
6778 44 : for (key, img_res) in results {
6779 42 : let expected = test_img(&format!("{} at {}", key, latest_lsns[&key]));
6780 42 : assert_eq!(img_res?, expected);
6781 2 : }
6782 2 :
6783 2 : Ok(())
6784 2 : }
6785 :
6786 : // Test that vectored get descends into ancestor timelines correctly and
6787 : // does not return an image that's newer than requested.
6788 : //
6789 : // The diagram below ilustrates an interesting case. We have a parent timeline
6790 : // (top of the Lsn range) and a child timeline. The request key cannot be reconstructed
6791 : // from the child timeline, so the parent timeline must be visited. When advacing into
6792 : // the child timeline, the read path needs to remember what the requested Lsn was in
6793 : // order to avoid returning an image that's too new. The test below constructs such
6794 : // a timeline setup and does a few queries around the Lsn of each page image.
6795 : // ```
6796 : // LSN
6797 : // ^
6798 : // |
6799 : // |
6800 : // 500 | --------------------------------------> branch point
6801 : // 400 | X
6802 : // 300 | X
6803 : // 200 | --------------------------------------> requested lsn
6804 : // 100 | X
6805 : // |---------------------------------------> Key
6806 : // |
6807 : // ------> requested key
6808 : //
6809 : // Legend:
6810 : // * X - page images
6811 : // ```
6812 : #[tokio::test]
6813 2 : async fn test_get_vectored_ancestor_descent() -> anyhow::Result<()> {
6814 2 : let harness = TenantHarness::create("test_get_vectored_on_lsn_axis").await?;
6815 2 : let (tenant, ctx) = harness.load().await;
6816 2 :
6817 2 : let start_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6818 2 : let end_key = start_key.add(1000);
6819 2 : let child_gap_at_key = start_key.add(500);
6820 2 : let mut parent_gap_lsns: BTreeMap<Lsn, String> = BTreeMap::new();
6821 2 :
6822 2 : let mut current_lsn = Lsn(0x10);
6823 2 :
6824 2 : let timeline_id = TimelineId::generate();
6825 2 : let parent_timeline = tenant
6826 2 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
6827 2 : .await?;
6828 2 :
6829 2 : current_lsn += 0x100;
6830 2 :
6831 8 : for _ in 0..3 {
6832 6 : let mut key = start_key;
6833 6006 : while key < end_key {
6834 6000 : current_lsn += 0x10;
6835 6000 :
6836 6000 : let image_value = format!("{} at {}", child_gap_at_key, current_lsn);
6837 2 :
6838 6000 : let mut writer = parent_timeline.writer().await;
6839 6000 : writer
6840 6000 : .put(
6841 6000 : key,
6842 6000 : current_lsn,
6843 6000 : &Value::Image(test_img(&image_value)),
6844 6000 : &ctx,
6845 6000 : )
6846 6000 : .await?;
6847 6000 : writer.finish_write(current_lsn);
6848 6000 :
6849 6000 : if key == child_gap_at_key {
6850 6 : parent_gap_lsns.insert(current_lsn, image_value);
6851 5994 : }
6852 2 :
6853 6000 : key = key.next();
6854 2 : }
6855 2 :
6856 6 : parent_timeline.freeze_and_flush().await?;
6857 2 : }
6858 2 :
6859 2 : let child_timeline_id = TimelineId::generate();
6860 2 :
6861 2 : let child_timeline = tenant
6862 2 : .branch_timeline_test(&parent_timeline, child_timeline_id, Some(current_lsn), &ctx)
6863 2 : .await?;
6864 2 :
6865 2 : let mut key = start_key;
6866 2002 : while key < end_key {
6867 2000 : if key == child_gap_at_key {
6868 2 : key = key.next();
6869 2 : continue;
6870 1998 : }
6871 1998 :
6872 1998 : current_lsn += 0x10;
6873 2 :
6874 1998 : let mut writer = child_timeline.writer().await;
6875 1998 : writer
6876 1998 : .put(
6877 1998 : key,
6878 1998 : current_lsn,
6879 1998 : &Value::Image(test_img(&format!("{} at {}", key, current_lsn))),
6880 1998 : &ctx,
6881 1998 : )
6882 1998 : .await?;
6883 1998 : writer.finish_write(current_lsn);
6884 1998 :
6885 1998 : key = key.next();
6886 2 : }
6887 2 :
6888 2 : child_timeline.freeze_and_flush().await?;
6889 2 :
6890 2 : let lsn_offsets: [i64; 5] = [-10, -1, 0, 1, 10];
6891 2 : let mut query_lsns = Vec::new();
6892 6 : for image_lsn in parent_gap_lsns.keys().rev() {
6893 36 : for offset in lsn_offsets {
6894 30 : query_lsns.push(Lsn(image_lsn
6895 30 : .0
6896 30 : .checked_add_signed(offset)
6897 30 : .expect("Shouldn't overflow")));
6898 30 : }
6899 2 : }
6900 2 :
6901 32 : for query_lsn in query_lsns {
6902 30 : let results = child_timeline
6903 30 : .get_vectored_impl(
6904 30 : KeySpace {
6905 30 : ranges: vec![child_gap_at_key..child_gap_at_key.next()],
6906 30 : },
6907 30 : query_lsn,
6908 30 : &mut ValuesReconstructState::new(),
6909 30 : &ctx,
6910 30 : )
6911 30 : .await;
6912 2 :
6913 30 : let expected_item = parent_gap_lsns
6914 30 : .iter()
6915 30 : .rev()
6916 68 : .find(|(lsn, _)| **lsn <= query_lsn);
6917 30 :
6918 30 : info!(
6919 2 : "Doing vectored read at LSN {}. Expecting image to be: {:?}",
6920 2 : query_lsn, expected_item
6921 2 : );
6922 2 :
6923 30 : match expected_item {
6924 26 : Some((_, img_value)) => {
6925 26 : let key_results = results.expect("No vectored get error expected");
6926 26 : let key_result = &key_results[&child_gap_at_key];
6927 26 : let returned_img = key_result
6928 26 : .as_ref()
6929 26 : .expect("No page reconstruct error expected");
6930 26 :
6931 26 : info!(
6932 2 : "Vectored read at LSN {} returned image {}",
6933 0 : query_lsn,
6934 0 : std::str::from_utf8(returned_img)?
6935 2 : );
6936 26 : assert_eq!(*returned_img, test_img(img_value));
6937 2 : }
6938 2 : None => {
6939 4 : assert!(matches!(results, Err(GetVectoredError::MissingKey(_))));
6940 2 : }
6941 2 : }
6942 2 : }
6943 2 :
6944 2 : Ok(())
6945 2 : }
6946 :
6947 : #[tokio::test]
6948 2 : async fn test_random_updates() -> anyhow::Result<()> {
6949 2 : let names_algorithms = [
6950 2 : ("test_random_updates_legacy", CompactionAlgorithm::Legacy),
6951 2 : ("test_random_updates_tiered", CompactionAlgorithm::Tiered),
6952 2 : ];
6953 6 : for (name, algorithm) in names_algorithms {
6954 4 : test_random_updates_algorithm(name, algorithm).await?;
6955 2 : }
6956 2 : Ok(())
6957 2 : }
6958 :
6959 4 : async fn test_random_updates_algorithm(
6960 4 : name: &'static str,
6961 4 : compaction_algorithm: CompactionAlgorithm,
6962 4 : ) -> anyhow::Result<()> {
6963 4 : let mut harness = TenantHarness::create(name).await?;
6964 4 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
6965 4 : kind: compaction_algorithm,
6966 4 : };
6967 4 : let (tenant, ctx) = harness.load().await;
6968 4 : let tline = tenant
6969 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6970 4 : .await?;
6971 :
6972 : const NUM_KEYS: usize = 1000;
6973 4 : let cancel = CancellationToken::new();
6974 4 :
6975 4 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6976 4 : let mut test_key_end = test_key;
6977 4 : test_key_end.field6 = NUM_KEYS as u32;
6978 4 : tline.add_extra_test_dense_keyspace(KeySpace::single(test_key..test_key_end));
6979 4 :
6980 4 : let mut keyspace = KeySpaceAccum::new();
6981 4 :
6982 4 : // Track when each page was last modified. Used to assert that
6983 4 : // a read sees the latest page version.
6984 4 : let mut updated = [Lsn(0); NUM_KEYS];
6985 4 :
6986 4 : let mut lsn = Lsn(0x10);
6987 : #[allow(clippy::needless_range_loop)]
6988 4004 : for blknum in 0..NUM_KEYS {
6989 4000 : lsn = Lsn(lsn.0 + 0x10);
6990 4000 : test_key.field6 = blknum as u32;
6991 4000 : let mut writer = tline.writer().await;
6992 4000 : writer
6993 4000 : .put(
6994 4000 : test_key,
6995 4000 : lsn,
6996 4000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6997 4000 : &ctx,
6998 4000 : )
6999 4000 : .await?;
7000 4000 : writer.finish_write(lsn);
7001 4000 : updated[blknum] = lsn;
7002 4000 : drop(writer);
7003 4000 :
7004 4000 : keyspace.add_key(test_key);
7005 : }
7006 :
7007 204 : for _ in 0..50 {
7008 200200 : for _ in 0..NUM_KEYS {
7009 200000 : lsn = Lsn(lsn.0 + 0x10);
7010 200000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7011 200000 : test_key.field6 = blknum as u32;
7012 200000 : let mut writer = tline.writer().await;
7013 200000 : writer
7014 200000 : .put(
7015 200000 : test_key,
7016 200000 : lsn,
7017 200000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7018 200000 : &ctx,
7019 200000 : )
7020 200000 : .await?;
7021 200000 : writer.finish_write(lsn);
7022 200000 : drop(writer);
7023 200000 : updated[blknum] = lsn;
7024 : }
7025 :
7026 : // Read all the blocks
7027 200000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7028 200000 : test_key.field6 = blknum as u32;
7029 200000 : assert_eq!(
7030 200000 : tline.get(test_key, lsn, &ctx).await?,
7031 200000 : test_img(&format!("{} at {}", blknum, last_lsn))
7032 : );
7033 : }
7034 :
7035 : // Perform a cycle of flush, and GC
7036 200 : tline.freeze_and_flush().await?;
7037 200 : tenant
7038 200 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7039 200 : .await?;
7040 : }
7041 :
7042 4 : Ok(())
7043 4 : }
7044 :
7045 : #[tokio::test]
7046 2 : async fn test_traverse_branches() -> anyhow::Result<()> {
7047 2 : let (tenant, ctx) = TenantHarness::create("test_traverse_branches")
7048 2 : .await?
7049 2 : .load()
7050 2 : .await;
7051 2 : let mut tline = tenant
7052 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7053 2 : .await?;
7054 2 :
7055 2 : const NUM_KEYS: usize = 1000;
7056 2 :
7057 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7058 2 :
7059 2 : let mut keyspace = KeySpaceAccum::new();
7060 2 :
7061 2 : let cancel = CancellationToken::new();
7062 2 :
7063 2 : // Track when each page was last modified. Used to assert that
7064 2 : // a read sees the latest page version.
7065 2 : let mut updated = [Lsn(0); NUM_KEYS];
7066 2 :
7067 2 : let mut lsn = Lsn(0x10);
7068 2 : #[allow(clippy::needless_range_loop)]
7069 2002 : for blknum in 0..NUM_KEYS {
7070 2000 : lsn = Lsn(lsn.0 + 0x10);
7071 2000 : test_key.field6 = blknum as u32;
7072 2000 : let mut writer = tline.writer().await;
7073 2000 : writer
7074 2000 : .put(
7075 2000 : test_key,
7076 2000 : lsn,
7077 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7078 2000 : &ctx,
7079 2000 : )
7080 2000 : .await?;
7081 2000 : writer.finish_write(lsn);
7082 2000 : updated[blknum] = lsn;
7083 2000 : drop(writer);
7084 2000 :
7085 2000 : keyspace.add_key(test_key);
7086 2 : }
7087 2 :
7088 102 : for _ in 0..50 {
7089 100 : let new_tline_id = TimelineId::generate();
7090 100 : tenant
7091 100 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7092 100 : .await?;
7093 100 : tline = tenant
7094 100 : .get_timeline(new_tline_id, true)
7095 100 : .expect("Should have the branched timeline");
7096 2 :
7097 100100 : for _ in 0..NUM_KEYS {
7098 100000 : lsn = Lsn(lsn.0 + 0x10);
7099 100000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7100 100000 : test_key.field6 = blknum as u32;
7101 100000 : let mut writer = tline.writer().await;
7102 100000 : writer
7103 100000 : .put(
7104 100000 : test_key,
7105 100000 : lsn,
7106 100000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7107 100000 : &ctx,
7108 100000 : )
7109 100000 : .await?;
7110 100000 : println!("updating {} at {}", blknum, lsn);
7111 100000 : writer.finish_write(lsn);
7112 100000 : drop(writer);
7113 100000 : updated[blknum] = lsn;
7114 2 : }
7115 2 :
7116 2 : // Read all the blocks
7117 100000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7118 100000 : test_key.field6 = blknum as u32;
7119 100000 : assert_eq!(
7120 100000 : tline.get(test_key, lsn, &ctx).await?,
7121 100000 : test_img(&format!("{} at {}", blknum, last_lsn))
7122 2 : );
7123 2 : }
7124 2 :
7125 2 : // Perform a cycle of flush, compact, and GC
7126 100 : tline.freeze_and_flush().await?;
7127 100 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
7128 100 : tenant
7129 100 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7130 100 : .await?;
7131 2 : }
7132 2 :
7133 2 : Ok(())
7134 2 : }
7135 :
7136 : #[tokio::test]
7137 2 : async fn test_traverse_ancestors() -> anyhow::Result<()> {
7138 2 : let (tenant, ctx) = TenantHarness::create("test_traverse_ancestors")
7139 2 : .await?
7140 2 : .load()
7141 2 : .await;
7142 2 : let mut tline = tenant
7143 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7144 2 : .await?;
7145 2 :
7146 2 : const NUM_KEYS: usize = 100;
7147 2 : const NUM_TLINES: usize = 50;
7148 2 :
7149 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7150 2 : // Track page mutation lsns across different timelines.
7151 2 : let mut updated = [[Lsn(0); NUM_KEYS]; NUM_TLINES];
7152 2 :
7153 2 : let mut lsn = Lsn(0x10);
7154 2 :
7155 2 : #[allow(clippy::needless_range_loop)]
7156 102 : for idx in 0..NUM_TLINES {
7157 100 : let new_tline_id = TimelineId::generate();
7158 100 : tenant
7159 100 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7160 100 : .await?;
7161 100 : tline = tenant
7162 100 : .get_timeline(new_tline_id, true)
7163 100 : .expect("Should have the branched timeline");
7164 2 :
7165 10100 : for _ in 0..NUM_KEYS {
7166 10000 : lsn = Lsn(lsn.0 + 0x10);
7167 10000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7168 10000 : test_key.field6 = blknum as u32;
7169 10000 : let mut writer = tline.writer().await;
7170 10000 : writer
7171 10000 : .put(
7172 10000 : test_key,
7173 10000 : lsn,
7174 10000 : &Value::Image(test_img(&format!("{} {} at {}", idx, blknum, lsn))),
7175 10000 : &ctx,
7176 10000 : )
7177 10000 : .await?;
7178 10000 : println!("updating [{}][{}] at {}", idx, blknum, lsn);
7179 10000 : writer.finish_write(lsn);
7180 10000 : drop(writer);
7181 10000 : updated[idx][blknum] = lsn;
7182 2 : }
7183 2 : }
7184 2 :
7185 2 : // Read pages from leaf timeline across all ancestors.
7186 100 : for (idx, lsns) in updated.iter().enumerate() {
7187 10000 : for (blknum, lsn) in lsns.iter().enumerate() {
7188 2 : // Skip empty mutations.
7189 10000 : if lsn.0 == 0 {
7190 3683 : continue;
7191 6317 : }
7192 6317 : println!("checking [{idx}][{blknum}] at {lsn}");
7193 6317 : test_key.field6 = blknum as u32;
7194 6317 : assert_eq!(
7195 6317 : tline.get(test_key, *lsn, &ctx).await?,
7196 6317 : test_img(&format!("{idx} {blknum} at {lsn}"))
7197 2 : );
7198 2 : }
7199 2 : }
7200 2 : Ok(())
7201 2 : }
7202 :
7203 : #[tokio::test]
7204 2 : async fn test_write_at_initdb_lsn_takes_optimization_code_path() -> anyhow::Result<()> {
7205 2 : let (tenant, ctx) = TenantHarness::create("test_empty_test_timeline_is_usable")
7206 2 : .await?
7207 2 : .load()
7208 2 : .await;
7209 2 :
7210 2 : let initdb_lsn = Lsn(0x20);
7211 2 : let utline = tenant
7212 2 : .create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)
7213 2 : .await?;
7214 2 : let tline = utline.raw_timeline().unwrap();
7215 2 :
7216 2 : // Spawn flush loop now so that we can set the `expect_initdb_optimization`
7217 2 : tline.maybe_spawn_flush_loop();
7218 2 :
7219 2 : // Make sure the timeline has the minimum set of required keys for operation.
7220 2 : // The only operation you can always do on an empty timeline is to `put` new data.
7221 2 : // Except if you `put` at `initdb_lsn`.
7222 2 : // In that case, there's an optimization to directly create image layers instead of delta layers.
7223 2 : // It uses `repartition()`, which assumes some keys to be present.
7224 2 : // Let's make sure the test timeline can handle that case.
7225 2 : {
7226 2 : let mut state = tline.flush_loop_state.lock().unwrap();
7227 2 : assert_eq!(
7228 2 : timeline::FlushLoopState::Running {
7229 2 : expect_initdb_optimization: false,
7230 2 : initdb_optimization_count: 0,
7231 2 : },
7232 2 : *state
7233 2 : );
7234 2 : *state = timeline::FlushLoopState::Running {
7235 2 : expect_initdb_optimization: true,
7236 2 : initdb_optimization_count: 0,
7237 2 : };
7238 2 : }
7239 2 :
7240 2 : // Make writes at the initdb_lsn. When we flush it below, it should be handled by the optimization.
7241 2 : // As explained above, the optimization requires some keys to be present.
7242 2 : // As per `create_empty_timeline` documentation, use init_empty to set them.
7243 2 : // This is what `create_test_timeline` does, by the way.
7244 2 : let mut modification = tline.begin_modification(initdb_lsn);
7245 2 : modification
7246 2 : .init_empty_test_timeline()
7247 2 : .context("init_empty_test_timeline")?;
7248 2 : modification
7249 2 : .commit(&ctx)
7250 2 : .await
7251 2 : .context("commit init_empty_test_timeline modification")?;
7252 2 :
7253 2 : // Do the flush. The flush code will check the expectations that we set above.
7254 2 : tline.freeze_and_flush().await?;
7255 2 :
7256 2 : // assert freeze_and_flush exercised the initdb optimization
7257 2 : {
7258 2 : let state = tline.flush_loop_state.lock().unwrap();
7259 2 : let timeline::FlushLoopState::Running {
7260 2 : expect_initdb_optimization,
7261 2 : initdb_optimization_count,
7262 2 : } = *state
7263 2 : else {
7264 2 : panic!("unexpected state: {:?}", *state);
7265 2 : };
7266 2 : assert!(expect_initdb_optimization);
7267 2 : assert!(initdb_optimization_count > 0);
7268 2 : }
7269 2 : Ok(())
7270 2 : }
7271 :
7272 : #[tokio::test]
7273 2 : async fn test_create_guard_crash() -> anyhow::Result<()> {
7274 2 : let name = "test_create_guard_crash";
7275 2 : let harness = TenantHarness::create(name).await?;
7276 2 : {
7277 2 : let (tenant, ctx) = harness.load().await;
7278 2 : let tline = tenant
7279 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
7280 2 : .await?;
7281 2 : // Leave the timeline ID in [`Tenant::timelines_creating`] to exclude attempting to create it again
7282 2 : let raw_tline = tline.raw_timeline().unwrap();
7283 2 : raw_tline
7284 2 : .shutdown(super::timeline::ShutdownMode::Hard)
7285 2 : .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))
7286 2 : .await;
7287 2 : std::mem::forget(tline);
7288 2 : }
7289 2 :
7290 2 : let (tenant, _) = harness.load().await;
7291 2 : match tenant.get_timeline(TIMELINE_ID, false) {
7292 2 : Ok(_) => panic!("timeline should've been removed during load"),
7293 2 : Err(e) => {
7294 2 : assert_eq!(
7295 2 : e,
7296 2 : GetTimelineError::NotFound {
7297 2 : tenant_id: tenant.tenant_shard_id,
7298 2 : timeline_id: TIMELINE_ID,
7299 2 : }
7300 2 : )
7301 2 : }
7302 2 : }
7303 2 :
7304 2 : assert!(!harness
7305 2 : .conf
7306 2 : .timeline_path(&tenant.tenant_shard_id, &TIMELINE_ID)
7307 2 : .exists());
7308 2 :
7309 2 : Ok(())
7310 2 : }
7311 :
7312 : #[tokio::test]
7313 2 : async fn test_read_at_max_lsn() -> anyhow::Result<()> {
7314 2 : let names_algorithms = [
7315 2 : ("test_read_at_max_lsn_legacy", CompactionAlgorithm::Legacy),
7316 2 : ("test_read_at_max_lsn_tiered", CompactionAlgorithm::Tiered),
7317 2 : ];
7318 6 : for (name, algorithm) in names_algorithms {
7319 4 : test_read_at_max_lsn_algorithm(name, algorithm).await?;
7320 2 : }
7321 2 : Ok(())
7322 2 : }
7323 :
7324 4 : async fn test_read_at_max_lsn_algorithm(
7325 4 : name: &'static str,
7326 4 : compaction_algorithm: CompactionAlgorithm,
7327 4 : ) -> anyhow::Result<()> {
7328 4 : let mut harness = TenantHarness::create(name).await?;
7329 4 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
7330 4 : kind: compaction_algorithm,
7331 4 : };
7332 4 : let (tenant, ctx) = harness.load().await;
7333 4 : let tline = tenant
7334 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
7335 4 : .await?;
7336 :
7337 4 : let lsn = Lsn(0x10);
7338 4 : let compact = false;
7339 4 : bulk_insert_maybe_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000, compact).await?;
7340 :
7341 4 : let test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7342 4 : let read_lsn = Lsn(u64::MAX - 1);
7343 :
7344 4 : let result = tline.get(test_key, read_lsn, &ctx).await;
7345 4 : assert!(result.is_ok(), "result is not Ok: {}", result.unwrap_err());
7346 :
7347 4 : Ok(())
7348 4 : }
7349 :
7350 : #[tokio::test]
7351 2 : async fn test_metadata_scan() -> anyhow::Result<()> {
7352 2 : let harness = TenantHarness::create("test_metadata_scan").await?;
7353 2 : let (tenant, ctx) = harness.load().await;
7354 2 : let tline = tenant
7355 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7356 2 : .await?;
7357 2 :
7358 2 : const NUM_KEYS: usize = 1000;
7359 2 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
7360 2 :
7361 2 : let cancel = CancellationToken::new();
7362 2 :
7363 2 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7364 2 : base_key.field1 = AUX_KEY_PREFIX;
7365 2 : let mut test_key = base_key;
7366 2 :
7367 2 : // Track when each page was last modified. Used to assert that
7368 2 : // a read sees the latest page version.
7369 2 : let mut updated = [Lsn(0); NUM_KEYS];
7370 2 :
7371 2 : let mut lsn = Lsn(0x10);
7372 2 : #[allow(clippy::needless_range_loop)]
7373 2002 : for blknum in 0..NUM_KEYS {
7374 2000 : lsn = Lsn(lsn.0 + 0x10);
7375 2000 : test_key.field6 = (blknum * STEP) as u32;
7376 2000 : let mut writer = tline.writer().await;
7377 2000 : writer
7378 2000 : .put(
7379 2000 : test_key,
7380 2000 : lsn,
7381 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7382 2000 : &ctx,
7383 2000 : )
7384 2000 : .await?;
7385 2000 : writer.finish_write(lsn);
7386 2000 : updated[blknum] = lsn;
7387 2000 : drop(writer);
7388 2 : }
7389 2 :
7390 2 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
7391 2 :
7392 24 : for iter in 0..=10 {
7393 2 : // Read all the blocks
7394 22000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7395 22000 : test_key.field6 = (blknum * STEP) as u32;
7396 22000 : assert_eq!(
7397 22000 : tline.get(test_key, lsn, &ctx).await?,
7398 22000 : test_img(&format!("{} at {}", blknum, last_lsn))
7399 2 : );
7400 2 : }
7401 2 :
7402 22 : let mut cnt = 0;
7403 22000 : for (key, value) in tline
7404 22 : .get_vectored_impl(
7405 22 : keyspace.clone(),
7406 22 : lsn,
7407 22 : &mut ValuesReconstructState::default(),
7408 22 : &ctx,
7409 22 : )
7410 22 : .await?
7411 2 : {
7412 22000 : let blknum = key.field6 as usize;
7413 22000 : let value = value?;
7414 22000 : assert!(blknum % STEP == 0);
7415 22000 : let blknum = blknum / STEP;
7416 22000 : assert_eq!(
7417 22000 : value,
7418 22000 : test_img(&format!("{} at {}", blknum, updated[blknum]))
7419 22000 : );
7420 22000 : cnt += 1;
7421 2 : }
7422 2 :
7423 22 : assert_eq!(cnt, NUM_KEYS);
7424 2 :
7425 22022 : for _ in 0..NUM_KEYS {
7426 22000 : lsn = Lsn(lsn.0 + 0x10);
7427 22000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7428 22000 : test_key.field6 = (blknum * STEP) as u32;
7429 22000 : let mut writer = tline.writer().await;
7430 22000 : writer
7431 22000 : .put(
7432 22000 : test_key,
7433 22000 : lsn,
7434 22000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7435 22000 : &ctx,
7436 22000 : )
7437 22000 : .await?;
7438 22000 : writer.finish_write(lsn);
7439 22000 : drop(writer);
7440 22000 : updated[blknum] = lsn;
7441 2 : }
7442 2 :
7443 2 : // Perform two cycles of flush, compact, and GC
7444 66 : for round in 0..2 {
7445 44 : tline.freeze_and_flush().await?;
7446 44 : tline
7447 44 : .compact(
7448 44 : &cancel,
7449 44 : if iter % 5 == 0 && round == 0 {
7450 6 : let mut flags = EnumSet::new();
7451 6 : flags.insert(CompactFlags::ForceImageLayerCreation);
7452 6 : flags.insert(CompactFlags::ForceRepartition);
7453 6 : flags
7454 2 : } else {
7455 38 : EnumSet::empty()
7456 2 : },
7457 44 : &ctx,
7458 44 : )
7459 44 : .await?;
7460 44 : tenant
7461 44 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7462 44 : .await?;
7463 2 : }
7464 2 : }
7465 2 :
7466 2 : Ok(())
7467 2 : }
7468 :
7469 : #[tokio::test]
7470 2 : async fn test_metadata_compaction_trigger() -> anyhow::Result<()> {
7471 2 : let harness = TenantHarness::create("test_metadata_compaction_trigger").await?;
7472 2 : let (tenant, ctx) = harness.load().await;
7473 2 : let tline = tenant
7474 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7475 2 : .await?;
7476 2 :
7477 2 : let cancel = CancellationToken::new();
7478 2 :
7479 2 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7480 2 : base_key.field1 = AUX_KEY_PREFIX;
7481 2 : let test_key = base_key;
7482 2 : let mut lsn = Lsn(0x10);
7483 2 :
7484 42 : for _ in 0..20 {
7485 40 : lsn = Lsn(lsn.0 + 0x10);
7486 40 : let mut writer = tline.writer().await;
7487 40 : writer
7488 40 : .put(
7489 40 : test_key,
7490 40 : lsn,
7491 40 : &Value::Image(test_img(&format!("{} at {}", 0, lsn))),
7492 40 : &ctx,
7493 40 : )
7494 40 : .await?;
7495 40 : writer.finish_write(lsn);
7496 40 : drop(writer);
7497 40 : tline.freeze_and_flush().await?; // force create a delta layer
7498 2 : }
7499 2 :
7500 2 : let before_num_l0_delta_files =
7501 2 : tline.layers.read().await.layer_map()?.level0_deltas().len();
7502 2 :
7503 2 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
7504 2 :
7505 2 : let after_num_l0_delta_files = tline.layers.read().await.layer_map()?.level0_deltas().len();
7506 2 :
7507 2 : 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}");
7508 2 :
7509 2 : assert_eq!(
7510 2 : tline.get(test_key, lsn, &ctx).await?,
7511 2 : test_img(&format!("{} at {}", 0, lsn))
7512 2 : );
7513 2 :
7514 2 : Ok(())
7515 2 : }
7516 :
7517 : #[tokio::test]
7518 2 : async fn test_aux_file_e2e() {
7519 2 : let harness = TenantHarness::create("test_aux_file_e2e").await.unwrap();
7520 2 :
7521 2 : let (tenant, ctx) = harness.load().await;
7522 2 :
7523 2 : let mut lsn = Lsn(0x08);
7524 2 :
7525 2 : let tline: Arc<Timeline> = tenant
7526 2 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
7527 2 : .await
7528 2 : .unwrap();
7529 2 :
7530 2 : {
7531 2 : lsn += 8;
7532 2 : let mut modification = tline.begin_modification(lsn);
7533 2 : modification
7534 2 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
7535 2 : .await
7536 2 : .unwrap();
7537 2 : modification.commit(&ctx).await.unwrap();
7538 2 : }
7539 2 :
7540 2 : // we can read everything from the storage
7541 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
7542 2 : assert_eq!(
7543 2 : files.get("pg_logical/mappings/test1"),
7544 2 : Some(&bytes::Bytes::from_static(b"first"))
7545 2 : );
7546 2 :
7547 2 : {
7548 2 : lsn += 8;
7549 2 : let mut modification = tline.begin_modification(lsn);
7550 2 : modification
7551 2 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
7552 2 : .await
7553 2 : .unwrap();
7554 2 : modification.commit(&ctx).await.unwrap();
7555 2 : }
7556 2 :
7557 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
7558 2 : assert_eq!(
7559 2 : files.get("pg_logical/mappings/test2"),
7560 2 : Some(&bytes::Bytes::from_static(b"second"))
7561 2 : );
7562 2 :
7563 2 : let child = tenant
7564 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(lsn), &ctx)
7565 2 : .await
7566 2 : .unwrap();
7567 2 :
7568 2 : let files = child.list_aux_files(lsn, &ctx).await.unwrap();
7569 2 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
7570 2 : assert_eq!(files.get("pg_logical/mappings/test2"), None);
7571 2 : }
7572 :
7573 : #[tokio::test]
7574 2 : async fn test_metadata_image_creation() -> anyhow::Result<()> {
7575 2 : let harness = TenantHarness::create("test_metadata_image_creation").await?;
7576 2 : let (tenant, ctx) = harness.load().await;
7577 2 : let tline = tenant
7578 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7579 2 : .await?;
7580 2 :
7581 2 : const NUM_KEYS: usize = 1000;
7582 2 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
7583 2 :
7584 2 : let cancel = CancellationToken::new();
7585 2 :
7586 2 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7587 2 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
7588 2 : let mut test_key = base_key;
7589 2 : let mut lsn = Lsn(0x10);
7590 2 :
7591 8 : async fn scan_with_statistics(
7592 8 : tline: &Timeline,
7593 8 : keyspace: &KeySpace,
7594 8 : lsn: Lsn,
7595 8 : ctx: &RequestContext,
7596 8 : ) -> anyhow::Result<(BTreeMap<Key, Result<Bytes, PageReconstructError>>, usize)> {
7597 8 : let mut reconstruct_state = ValuesReconstructState::default();
7598 8 : let res = tline
7599 8 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
7600 8 : .await?;
7601 8 : Ok((res, reconstruct_state.get_delta_layers_visited() as usize))
7602 8 : }
7603 2 :
7604 2 : #[allow(clippy::needless_range_loop)]
7605 2002 : for blknum in 0..NUM_KEYS {
7606 2000 : lsn = Lsn(lsn.0 + 0x10);
7607 2000 : test_key.field6 = (blknum * STEP) as u32;
7608 2000 : let mut writer = tline.writer().await;
7609 2000 : writer
7610 2000 : .put(
7611 2000 : test_key,
7612 2000 : lsn,
7613 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7614 2000 : &ctx,
7615 2000 : )
7616 2000 : .await?;
7617 2000 : writer.finish_write(lsn);
7618 2000 : drop(writer);
7619 2 : }
7620 2 :
7621 2 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
7622 2 :
7623 22 : for iter in 1..=10 {
7624 20020 : for _ in 0..NUM_KEYS {
7625 20000 : lsn = Lsn(lsn.0 + 0x10);
7626 20000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7627 20000 : test_key.field6 = (blknum * STEP) as u32;
7628 20000 : let mut writer = tline.writer().await;
7629 20000 : writer
7630 20000 : .put(
7631 20000 : test_key,
7632 20000 : lsn,
7633 20000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7634 20000 : &ctx,
7635 20000 : )
7636 20000 : .await?;
7637 20000 : writer.finish_write(lsn);
7638 20000 : drop(writer);
7639 2 : }
7640 2 :
7641 20 : tline.freeze_and_flush().await?;
7642 2 :
7643 20 : if iter % 5 == 0 {
7644 4 : let (_, before_delta_file_accessed) =
7645 4 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
7646 4 : tline
7647 4 : .compact(
7648 4 : &cancel,
7649 4 : {
7650 4 : let mut flags = EnumSet::new();
7651 4 : flags.insert(CompactFlags::ForceImageLayerCreation);
7652 4 : flags.insert(CompactFlags::ForceRepartition);
7653 4 : flags
7654 4 : },
7655 4 : &ctx,
7656 4 : )
7657 4 : .await?;
7658 4 : let (_, after_delta_file_accessed) =
7659 4 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
7660 4 : 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}");
7661 2 : // 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.
7662 4 : assert!(
7663 4 : after_delta_file_accessed <= 2,
7664 2 : "after_delta_file_accessed={after_delta_file_accessed}"
7665 2 : );
7666 16 : }
7667 2 : }
7668 2 :
7669 2 : Ok(())
7670 2 : }
7671 :
7672 : #[tokio::test]
7673 2 : async fn test_vectored_missing_data_key_reads() -> anyhow::Result<()> {
7674 2 : let harness = TenantHarness::create("test_vectored_missing_data_key_reads").await?;
7675 2 : let (tenant, ctx) = harness.load().await;
7676 2 :
7677 2 : let base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7678 2 : let base_key_child = Key::from_hex("000000000033333333444444445500000001").unwrap();
7679 2 : let base_key_nonexist = Key::from_hex("000000000033333333444444445500000002").unwrap();
7680 2 :
7681 2 : let tline = tenant
7682 2 : .create_test_timeline_with_layers(
7683 2 : TIMELINE_ID,
7684 2 : Lsn(0x10),
7685 2 : DEFAULT_PG_VERSION,
7686 2 : &ctx,
7687 2 : Vec::new(), // delta layers
7688 2 : vec![(Lsn(0x20), vec![(base_key, test_img("data key 1"))])], // image layers
7689 2 : 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
7690 2 : )
7691 2 : .await?;
7692 2 : tline.add_extra_test_dense_keyspace(KeySpace::single(base_key..(base_key_nonexist.next())));
7693 2 :
7694 2 : let child = tenant
7695 2 : .branch_timeline_test_with_layers(
7696 2 : &tline,
7697 2 : NEW_TIMELINE_ID,
7698 2 : Some(Lsn(0x20)),
7699 2 : &ctx,
7700 2 : Vec::new(), // delta layers
7701 2 : vec![(Lsn(0x30), vec![(base_key_child, test_img("data key 2"))])], // image layers
7702 2 : Lsn(0x30),
7703 2 : )
7704 2 : .await
7705 2 : .unwrap();
7706 2 :
7707 2 : let lsn = Lsn(0x30);
7708 2 :
7709 2 : // test vectored get on parent timeline
7710 2 : assert_eq!(
7711 2 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
7712 2 : Some(test_img("data key 1"))
7713 2 : );
7714 2 : assert!(get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx)
7715 2 : .await
7716 2 : .unwrap_err()
7717 2 : .is_missing_key_error());
7718 2 : assert!(
7719 2 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx)
7720 2 : .await
7721 2 : .unwrap_err()
7722 2 : .is_missing_key_error()
7723 2 : );
7724 2 :
7725 2 : // test vectored get on child timeline
7726 2 : assert_eq!(
7727 2 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
7728 2 : Some(test_img("data key 1"))
7729 2 : );
7730 2 : assert_eq!(
7731 2 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
7732 2 : Some(test_img("data key 2"))
7733 2 : );
7734 2 : assert!(
7735 2 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx)
7736 2 : .await
7737 2 : .unwrap_err()
7738 2 : .is_missing_key_error()
7739 2 : );
7740 2 :
7741 2 : Ok(())
7742 2 : }
7743 :
7744 : #[tokio::test]
7745 2 : async fn test_vectored_missing_metadata_key_reads() -> anyhow::Result<()> {
7746 2 : let harness = TenantHarness::create("test_vectored_missing_metadata_key_reads").await?;
7747 2 : let (tenant, ctx) = harness.load().await;
7748 2 :
7749 2 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7750 2 : let base_key_child = Key::from_hex("620000000033333333444444445500000001").unwrap();
7751 2 : let base_key_nonexist = Key::from_hex("620000000033333333444444445500000002").unwrap();
7752 2 : let base_key_overwrite = Key::from_hex("620000000033333333444444445500000003").unwrap();
7753 2 :
7754 2 : let base_inherited_key = Key::from_hex("610000000033333333444444445500000000").unwrap();
7755 2 : let base_inherited_key_child =
7756 2 : Key::from_hex("610000000033333333444444445500000001").unwrap();
7757 2 : let base_inherited_key_nonexist =
7758 2 : Key::from_hex("610000000033333333444444445500000002").unwrap();
7759 2 : let base_inherited_key_overwrite =
7760 2 : Key::from_hex("610000000033333333444444445500000003").unwrap();
7761 2 :
7762 2 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
7763 2 : assert_eq!(base_inherited_key.field1, RELATION_SIZE_PREFIX);
7764 2 :
7765 2 : let tline = tenant
7766 2 : .create_test_timeline_with_layers(
7767 2 : TIMELINE_ID,
7768 2 : Lsn(0x10),
7769 2 : DEFAULT_PG_VERSION,
7770 2 : &ctx,
7771 2 : Vec::new(), // delta layers
7772 2 : vec![(
7773 2 : Lsn(0x20),
7774 2 : vec![
7775 2 : (base_inherited_key, test_img("metadata inherited key 1")),
7776 2 : (
7777 2 : base_inherited_key_overwrite,
7778 2 : test_img("metadata key overwrite 1a"),
7779 2 : ),
7780 2 : (base_key, test_img("metadata key 1")),
7781 2 : (base_key_overwrite, test_img("metadata key overwrite 1b")),
7782 2 : ],
7783 2 : )], // image layers
7784 2 : 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
7785 2 : )
7786 2 : .await?;
7787 2 :
7788 2 : let child = tenant
7789 2 : .branch_timeline_test_with_layers(
7790 2 : &tline,
7791 2 : NEW_TIMELINE_ID,
7792 2 : Some(Lsn(0x20)),
7793 2 : &ctx,
7794 2 : Vec::new(), // delta layers
7795 2 : vec![(
7796 2 : Lsn(0x30),
7797 2 : vec![
7798 2 : (
7799 2 : base_inherited_key_child,
7800 2 : test_img("metadata inherited key 2"),
7801 2 : ),
7802 2 : (
7803 2 : base_inherited_key_overwrite,
7804 2 : test_img("metadata key overwrite 2a"),
7805 2 : ),
7806 2 : (base_key_child, test_img("metadata key 2")),
7807 2 : (base_key_overwrite, test_img("metadata key overwrite 2b")),
7808 2 : ],
7809 2 : )], // image layers
7810 2 : Lsn(0x30),
7811 2 : )
7812 2 : .await
7813 2 : .unwrap();
7814 2 :
7815 2 : let lsn = Lsn(0x30);
7816 2 :
7817 2 : // test vectored get on parent timeline
7818 2 : assert_eq!(
7819 2 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
7820 2 : Some(test_img("metadata key 1"))
7821 2 : );
7822 2 : assert_eq!(
7823 2 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx).await?,
7824 2 : None
7825 2 : );
7826 2 : assert_eq!(
7827 2 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx).await?,
7828 2 : None
7829 2 : );
7830 2 : assert_eq!(
7831 2 : get_vectored_impl_wrapper(&tline, base_key_overwrite, lsn, &ctx).await?,
7832 2 : Some(test_img("metadata key overwrite 1b"))
7833 2 : );
7834 2 : assert_eq!(
7835 2 : get_vectored_impl_wrapper(&tline, base_inherited_key, lsn, &ctx).await?,
7836 2 : Some(test_img("metadata inherited key 1"))
7837 2 : );
7838 2 : assert_eq!(
7839 2 : get_vectored_impl_wrapper(&tline, base_inherited_key_child, lsn, &ctx).await?,
7840 2 : None
7841 2 : );
7842 2 : assert_eq!(
7843 2 : get_vectored_impl_wrapper(&tline, base_inherited_key_nonexist, lsn, &ctx).await?,
7844 2 : None
7845 2 : );
7846 2 : assert_eq!(
7847 2 : get_vectored_impl_wrapper(&tline, base_inherited_key_overwrite, lsn, &ctx).await?,
7848 2 : Some(test_img("metadata key overwrite 1a"))
7849 2 : );
7850 2 :
7851 2 : // test vectored get on child timeline
7852 2 : assert_eq!(
7853 2 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
7854 2 : None
7855 2 : );
7856 2 : assert_eq!(
7857 2 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
7858 2 : Some(test_img("metadata key 2"))
7859 2 : );
7860 2 : assert_eq!(
7861 2 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx).await?,
7862 2 : None
7863 2 : );
7864 2 : assert_eq!(
7865 2 : get_vectored_impl_wrapper(&child, base_inherited_key, lsn, &ctx).await?,
7866 2 : Some(test_img("metadata inherited key 1"))
7867 2 : );
7868 2 : assert_eq!(
7869 2 : get_vectored_impl_wrapper(&child, base_inherited_key_child, lsn, &ctx).await?,
7870 2 : Some(test_img("metadata inherited key 2"))
7871 2 : );
7872 2 : assert_eq!(
7873 2 : get_vectored_impl_wrapper(&child, base_inherited_key_nonexist, lsn, &ctx).await?,
7874 2 : None
7875 2 : );
7876 2 : assert_eq!(
7877 2 : get_vectored_impl_wrapper(&child, base_key_overwrite, lsn, &ctx).await?,
7878 2 : Some(test_img("metadata key overwrite 2b"))
7879 2 : );
7880 2 : assert_eq!(
7881 2 : get_vectored_impl_wrapper(&child, base_inherited_key_overwrite, lsn, &ctx).await?,
7882 2 : Some(test_img("metadata key overwrite 2a"))
7883 2 : );
7884 2 :
7885 2 : // test vectored scan on parent timeline
7886 2 : let mut reconstruct_state = ValuesReconstructState::new();
7887 2 : let res = tline
7888 2 : .get_vectored_impl(
7889 2 : KeySpace::single(Key::metadata_key_range()),
7890 2 : lsn,
7891 2 : &mut reconstruct_state,
7892 2 : &ctx,
7893 2 : )
7894 2 : .await?;
7895 2 :
7896 2 : assert_eq!(
7897 2 : res.into_iter()
7898 8 : .map(|(k, v)| (k, v.unwrap()))
7899 2 : .collect::<Vec<_>>(),
7900 2 : vec![
7901 2 : (base_inherited_key, test_img("metadata inherited key 1")),
7902 2 : (
7903 2 : base_inherited_key_overwrite,
7904 2 : test_img("metadata key overwrite 1a")
7905 2 : ),
7906 2 : (base_key, test_img("metadata key 1")),
7907 2 : (base_key_overwrite, test_img("metadata key overwrite 1b")),
7908 2 : ]
7909 2 : );
7910 2 :
7911 2 : // test vectored scan on child timeline
7912 2 : let mut reconstruct_state = ValuesReconstructState::new();
7913 2 : let res = child
7914 2 : .get_vectored_impl(
7915 2 : KeySpace::single(Key::metadata_key_range()),
7916 2 : lsn,
7917 2 : &mut reconstruct_state,
7918 2 : &ctx,
7919 2 : )
7920 2 : .await?;
7921 2 :
7922 2 : assert_eq!(
7923 2 : res.into_iter()
7924 10 : .map(|(k, v)| (k, v.unwrap()))
7925 2 : .collect::<Vec<_>>(),
7926 2 : vec![
7927 2 : (base_inherited_key, test_img("metadata inherited key 1")),
7928 2 : (
7929 2 : base_inherited_key_child,
7930 2 : test_img("metadata inherited key 2")
7931 2 : ),
7932 2 : (
7933 2 : base_inherited_key_overwrite,
7934 2 : test_img("metadata key overwrite 2a")
7935 2 : ),
7936 2 : (base_key_child, test_img("metadata key 2")),
7937 2 : (base_key_overwrite, test_img("metadata key overwrite 2b")),
7938 2 : ]
7939 2 : );
7940 2 :
7941 2 : Ok(())
7942 2 : }
7943 :
7944 56 : async fn get_vectored_impl_wrapper(
7945 56 : tline: &Arc<Timeline>,
7946 56 : key: Key,
7947 56 : lsn: Lsn,
7948 56 : ctx: &RequestContext,
7949 56 : ) -> Result<Option<Bytes>, GetVectoredError> {
7950 56 : let mut reconstruct_state = ValuesReconstructState::new();
7951 56 : let mut res = tline
7952 56 : .get_vectored_impl(
7953 56 : KeySpace::single(key..key.next()),
7954 56 : lsn,
7955 56 : &mut reconstruct_state,
7956 56 : ctx,
7957 56 : )
7958 56 : .await?;
7959 50 : Ok(res.pop_last().map(|(k, v)| {
7960 32 : assert_eq!(k, key);
7961 32 : v.unwrap()
7962 50 : }))
7963 56 : }
7964 :
7965 : #[tokio::test]
7966 2 : async fn test_metadata_tombstone_reads() -> anyhow::Result<()> {
7967 2 : let harness = TenantHarness::create("test_metadata_tombstone_reads").await?;
7968 2 : let (tenant, ctx) = harness.load().await;
7969 2 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
7970 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
7971 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
7972 2 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
7973 2 :
7974 2 : // We emulate the situation that the compaction algorithm creates an image layer that removes the tombstones
7975 2 : // Lsn 0x30 key0, key3, no key1+key2
7976 2 : // Lsn 0x20 key1+key2 tomestones
7977 2 : // Lsn 0x10 key1 in image, key2 in delta
7978 2 : let tline = tenant
7979 2 : .create_test_timeline_with_layers(
7980 2 : TIMELINE_ID,
7981 2 : Lsn(0x10),
7982 2 : DEFAULT_PG_VERSION,
7983 2 : &ctx,
7984 2 : // delta layers
7985 2 : vec![
7986 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7987 2 : Lsn(0x10)..Lsn(0x20),
7988 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
7989 2 : ),
7990 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7991 2 : Lsn(0x20)..Lsn(0x30),
7992 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
7993 2 : ),
7994 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7995 2 : Lsn(0x20)..Lsn(0x30),
7996 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
7997 2 : ),
7998 2 : ],
7999 2 : // image layers
8000 2 : vec![
8001 2 : (Lsn(0x10), vec![(key1, test_img("metadata key 1"))]),
8002 2 : (
8003 2 : Lsn(0x30),
8004 2 : vec![
8005 2 : (key0, test_img("metadata key 0")),
8006 2 : (key3, test_img("metadata key 3")),
8007 2 : ],
8008 2 : ),
8009 2 : ],
8010 2 : Lsn(0x30),
8011 2 : )
8012 2 : .await?;
8013 2 :
8014 2 : let lsn = Lsn(0x30);
8015 2 : let old_lsn = Lsn(0x20);
8016 2 :
8017 2 : assert_eq!(
8018 2 : get_vectored_impl_wrapper(&tline, key0, lsn, &ctx).await?,
8019 2 : Some(test_img("metadata key 0"))
8020 2 : );
8021 2 : assert_eq!(
8022 2 : get_vectored_impl_wrapper(&tline, key1, lsn, &ctx).await?,
8023 2 : None,
8024 2 : );
8025 2 : assert_eq!(
8026 2 : get_vectored_impl_wrapper(&tline, key2, lsn, &ctx).await?,
8027 2 : None,
8028 2 : );
8029 2 : assert_eq!(
8030 2 : get_vectored_impl_wrapper(&tline, key1, old_lsn, &ctx).await?,
8031 2 : Some(Bytes::new()),
8032 2 : );
8033 2 : assert_eq!(
8034 2 : get_vectored_impl_wrapper(&tline, key2, old_lsn, &ctx).await?,
8035 2 : Some(Bytes::new()),
8036 2 : );
8037 2 : assert_eq!(
8038 2 : get_vectored_impl_wrapper(&tline, key3, lsn, &ctx).await?,
8039 2 : Some(test_img("metadata key 3"))
8040 2 : );
8041 2 :
8042 2 : Ok(())
8043 2 : }
8044 :
8045 : #[tokio::test]
8046 2 : async fn test_metadata_tombstone_image_creation() {
8047 2 : let harness = TenantHarness::create("test_metadata_tombstone_image_creation")
8048 2 : .await
8049 2 : .unwrap();
8050 2 : let (tenant, ctx) = harness.load().await;
8051 2 :
8052 2 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
8053 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8054 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8055 2 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
8056 2 :
8057 2 : let tline = tenant
8058 2 : .create_test_timeline_with_layers(
8059 2 : TIMELINE_ID,
8060 2 : Lsn(0x10),
8061 2 : DEFAULT_PG_VERSION,
8062 2 : &ctx,
8063 2 : // delta layers
8064 2 : vec![
8065 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8066 2 : Lsn(0x10)..Lsn(0x20),
8067 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8068 2 : ),
8069 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8070 2 : Lsn(0x20)..Lsn(0x30),
8071 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8072 2 : ),
8073 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8074 2 : Lsn(0x20)..Lsn(0x30),
8075 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8076 2 : ),
8077 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8078 2 : Lsn(0x30)..Lsn(0x40),
8079 2 : vec![
8080 2 : (key0, Lsn(0x30), Value::Image(test_img("metadata key 0"))),
8081 2 : (key3, Lsn(0x30), Value::Image(test_img("metadata key 3"))),
8082 2 : ],
8083 2 : ),
8084 2 : ],
8085 2 : // image layers
8086 2 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
8087 2 : Lsn(0x40),
8088 2 : )
8089 2 : .await
8090 2 : .unwrap();
8091 2 :
8092 2 : let cancel = CancellationToken::new();
8093 2 :
8094 2 : tline
8095 2 : .compact(
8096 2 : &cancel,
8097 2 : {
8098 2 : let mut flags = EnumSet::new();
8099 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
8100 2 : flags.insert(CompactFlags::ForceRepartition);
8101 2 : flags
8102 2 : },
8103 2 : &ctx,
8104 2 : )
8105 2 : .await
8106 2 : .unwrap();
8107 2 :
8108 2 : // Image layers are created at last_record_lsn
8109 2 : let images = tline
8110 2 : .inspect_image_layers(Lsn(0x40), &ctx)
8111 2 : .await
8112 2 : .unwrap()
8113 2 : .into_iter()
8114 18 : .filter(|(k, _)| k.is_metadata_key())
8115 2 : .collect::<Vec<_>>();
8116 2 : assert_eq!(images.len(), 2); // the image layer should only contain two existing keys, tombstones should be removed.
8117 2 : }
8118 :
8119 : #[tokio::test]
8120 2 : async fn test_metadata_tombstone_empty_image_creation() {
8121 2 : let harness = TenantHarness::create("test_metadata_tombstone_empty_image_creation")
8122 2 : .await
8123 2 : .unwrap();
8124 2 : let (tenant, ctx) = harness.load().await;
8125 2 :
8126 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8127 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8128 2 :
8129 2 : let tline = tenant
8130 2 : .create_test_timeline_with_layers(
8131 2 : TIMELINE_ID,
8132 2 : Lsn(0x10),
8133 2 : DEFAULT_PG_VERSION,
8134 2 : &ctx,
8135 2 : // delta layers
8136 2 : vec![
8137 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8138 2 : Lsn(0x10)..Lsn(0x20),
8139 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8140 2 : ),
8141 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8142 2 : Lsn(0x20)..Lsn(0x30),
8143 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8144 2 : ),
8145 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8146 2 : Lsn(0x20)..Lsn(0x30),
8147 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8148 2 : ),
8149 2 : ],
8150 2 : // image layers
8151 2 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
8152 2 : Lsn(0x30),
8153 2 : )
8154 2 : .await
8155 2 : .unwrap();
8156 2 :
8157 2 : let cancel = CancellationToken::new();
8158 2 :
8159 2 : tline
8160 2 : .compact(
8161 2 : &cancel,
8162 2 : {
8163 2 : let mut flags = EnumSet::new();
8164 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
8165 2 : flags.insert(CompactFlags::ForceRepartition);
8166 2 : flags
8167 2 : },
8168 2 : &ctx,
8169 2 : )
8170 2 : .await
8171 2 : .unwrap();
8172 2 :
8173 2 : // Image layers are created at last_record_lsn
8174 2 : let images = tline
8175 2 : .inspect_image_layers(Lsn(0x30), &ctx)
8176 2 : .await
8177 2 : .unwrap()
8178 2 : .into_iter()
8179 14 : .filter(|(k, _)| k.is_metadata_key())
8180 2 : .collect::<Vec<_>>();
8181 2 : assert_eq!(images.len(), 0); // the image layer should not contain tombstones, or it is not created
8182 2 : }
8183 :
8184 : #[tokio::test]
8185 2 : async fn test_simple_bottom_most_compaction_images() -> anyhow::Result<()> {
8186 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_images").await?;
8187 2 : let (tenant, ctx) = harness.load().await;
8188 2 :
8189 102 : fn get_key(id: u32) -> Key {
8190 102 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8191 102 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8192 102 : key.field6 = id;
8193 102 : key
8194 102 : }
8195 2 :
8196 2 : // We create
8197 2 : // - one bottom-most image layer,
8198 2 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
8199 2 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
8200 2 : // - a delta layer D3 above the horizon.
8201 2 : //
8202 2 : // | D3 |
8203 2 : // | D1 |
8204 2 : // -| |-- gc horizon -----------------
8205 2 : // | | | D2 |
8206 2 : // --------- img layer ------------------
8207 2 : //
8208 2 : // What we should expact from this compaction is:
8209 2 : // | D3 |
8210 2 : // | Part of D1 |
8211 2 : // --------- img layer with D1+D2 at GC horizon------------------
8212 2 :
8213 2 : // img layer at 0x10
8214 2 : let img_layer = (0..10)
8215 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8216 2 : .collect_vec();
8217 2 :
8218 2 : let delta1 = vec![
8219 2 : (
8220 2 : get_key(1),
8221 2 : Lsn(0x20),
8222 2 : Value::Image(Bytes::from("value 1@0x20")),
8223 2 : ),
8224 2 : (
8225 2 : get_key(2),
8226 2 : Lsn(0x30),
8227 2 : Value::Image(Bytes::from("value 2@0x30")),
8228 2 : ),
8229 2 : (
8230 2 : get_key(3),
8231 2 : Lsn(0x40),
8232 2 : Value::Image(Bytes::from("value 3@0x40")),
8233 2 : ),
8234 2 : ];
8235 2 : let delta2 = vec![
8236 2 : (
8237 2 : get_key(5),
8238 2 : Lsn(0x20),
8239 2 : Value::Image(Bytes::from("value 5@0x20")),
8240 2 : ),
8241 2 : (
8242 2 : get_key(6),
8243 2 : Lsn(0x20),
8244 2 : Value::Image(Bytes::from("value 6@0x20")),
8245 2 : ),
8246 2 : ];
8247 2 : let delta3 = vec![
8248 2 : (
8249 2 : get_key(8),
8250 2 : Lsn(0x48),
8251 2 : Value::Image(Bytes::from("value 8@0x48")),
8252 2 : ),
8253 2 : (
8254 2 : get_key(9),
8255 2 : Lsn(0x48),
8256 2 : Value::Image(Bytes::from("value 9@0x48")),
8257 2 : ),
8258 2 : ];
8259 2 :
8260 2 : let tline = tenant
8261 2 : .create_test_timeline_with_layers(
8262 2 : TIMELINE_ID,
8263 2 : Lsn(0x10),
8264 2 : DEFAULT_PG_VERSION,
8265 2 : &ctx,
8266 2 : vec![
8267 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
8268 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
8269 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
8270 2 : ], // delta layers
8271 2 : vec![(Lsn(0x10), img_layer)], // image layers
8272 2 : Lsn(0x50),
8273 2 : )
8274 2 : .await?;
8275 2 : {
8276 2 : tline
8277 2 : .latest_gc_cutoff_lsn
8278 2 : .lock_for_write()
8279 2 : .store_and_unlock(Lsn(0x30))
8280 2 : .wait()
8281 2 : .await;
8282 2 : // Update GC info
8283 2 : let mut guard = tline.gc_info.write().unwrap();
8284 2 : guard.cutoffs.time = Lsn(0x30);
8285 2 : guard.cutoffs.space = Lsn(0x30);
8286 2 : }
8287 2 :
8288 2 : let expected_result = [
8289 2 : Bytes::from_static(b"value 0@0x10"),
8290 2 : Bytes::from_static(b"value 1@0x20"),
8291 2 : Bytes::from_static(b"value 2@0x30"),
8292 2 : Bytes::from_static(b"value 3@0x40"),
8293 2 : Bytes::from_static(b"value 4@0x10"),
8294 2 : Bytes::from_static(b"value 5@0x20"),
8295 2 : Bytes::from_static(b"value 6@0x20"),
8296 2 : Bytes::from_static(b"value 7@0x10"),
8297 2 : Bytes::from_static(b"value 8@0x48"),
8298 2 : Bytes::from_static(b"value 9@0x48"),
8299 2 : ];
8300 2 :
8301 20 : for (idx, expected) in expected_result.iter().enumerate() {
8302 20 : assert_eq!(
8303 20 : tline
8304 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8305 20 : .await
8306 20 : .unwrap(),
8307 2 : expected
8308 2 : );
8309 2 : }
8310 2 :
8311 2 : let cancel = CancellationToken::new();
8312 2 : tline
8313 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8314 2 : .await
8315 2 : .unwrap();
8316 2 :
8317 20 : for (idx, expected) in expected_result.iter().enumerate() {
8318 20 : assert_eq!(
8319 20 : tline
8320 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8321 20 : .await
8322 20 : .unwrap(),
8323 2 : expected
8324 2 : );
8325 2 : }
8326 2 :
8327 2 : // Check if the image layer at the GC horizon contains exactly what we want
8328 2 : let image_at_gc_horizon = tline
8329 2 : .inspect_image_layers(Lsn(0x30), &ctx)
8330 2 : .await
8331 2 : .unwrap()
8332 2 : .into_iter()
8333 34 : .filter(|(k, _)| k.is_metadata_key())
8334 2 : .collect::<Vec<_>>();
8335 2 :
8336 2 : assert_eq!(image_at_gc_horizon.len(), 10);
8337 2 : let expected_result = [
8338 2 : Bytes::from_static(b"value 0@0x10"),
8339 2 : Bytes::from_static(b"value 1@0x20"),
8340 2 : Bytes::from_static(b"value 2@0x30"),
8341 2 : Bytes::from_static(b"value 3@0x10"),
8342 2 : Bytes::from_static(b"value 4@0x10"),
8343 2 : Bytes::from_static(b"value 5@0x20"),
8344 2 : Bytes::from_static(b"value 6@0x20"),
8345 2 : Bytes::from_static(b"value 7@0x10"),
8346 2 : Bytes::from_static(b"value 8@0x10"),
8347 2 : Bytes::from_static(b"value 9@0x10"),
8348 2 : ];
8349 22 : for idx in 0..10 {
8350 20 : assert_eq!(
8351 20 : image_at_gc_horizon[idx],
8352 20 : (get_key(idx as u32), expected_result[idx].clone())
8353 20 : );
8354 2 : }
8355 2 :
8356 2 : // Check if old layers are removed / new layers have the expected LSN
8357 2 : let all_layers = inspect_and_sort(&tline, None).await;
8358 2 : assert_eq!(
8359 2 : all_layers,
8360 2 : vec![
8361 2 : // Image layer at GC horizon
8362 2 : PersistentLayerKey {
8363 2 : key_range: Key::MIN..Key::MAX,
8364 2 : lsn_range: Lsn(0x30)..Lsn(0x31),
8365 2 : is_delta: false
8366 2 : },
8367 2 : // The delta layer below the horizon
8368 2 : PersistentLayerKey {
8369 2 : key_range: get_key(3)..get_key(4),
8370 2 : lsn_range: Lsn(0x30)..Lsn(0x48),
8371 2 : is_delta: true
8372 2 : },
8373 2 : // The delta3 layer that should not be picked for the compaction
8374 2 : PersistentLayerKey {
8375 2 : key_range: get_key(8)..get_key(10),
8376 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
8377 2 : is_delta: true
8378 2 : }
8379 2 : ]
8380 2 : );
8381 2 :
8382 2 : // increase GC horizon and compact again
8383 2 : {
8384 2 : tline
8385 2 : .latest_gc_cutoff_lsn
8386 2 : .lock_for_write()
8387 2 : .store_and_unlock(Lsn(0x40))
8388 2 : .wait()
8389 2 : .await;
8390 2 : // Update GC info
8391 2 : let mut guard = tline.gc_info.write().unwrap();
8392 2 : guard.cutoffs.time = Lsn(0x40);
8393 2 : guard.cutoffs.space = Lsn(0x40);
8394 2 : }
8395 2 : tline
8396 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8397 2 : .await
8398 2 : .unwrap();
8399 2 :
8400 2 : Ok(())
8401 2 : }
8402 :
8403 : #[cfg(feature = "testing")]
8404 : #[tokio::test]
8405 2 : async fn test_neon_test_record() -> anyhow::Result<()> {
8406 2 : let harness = TenantHarness::create("test_neon_test_record").await?;
8407 2 : let (tenant, ctx) = harness.load().await;
8408 2 :
8409 24 : fn get_key(id: u32) -> Key {
8410 24 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8411 24 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8412 24 : key.field6 = id;
8413 24 : key
8414 24 : }
8415 2 :
8416 2 : let delta1 = vec![
8417 2 : (
8418 2 : get_key(1),
8419 2 : Lsn(0x20),
8420 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
8421 2 : ),
8422 2 : (
8423 2 : get_key(1),
8424 2 : Lsn(0x30),
8425 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
8426 2 : ),
8427 2 : (get_key(2), Lsn(0x10), Value::Image("0x10".into())),
8428 2 : (
8429 2 : get_key(2),
8430 2 : Lsn(0x20),
8431 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
8432 2 : ),
8433 2 : (
8434 2 : get_key(2),
8435 2 : Lsn(0x30),
8436 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
8437 2 : ),
8438 2 : (get_key(3), Lsn(0x10), Value::Image("0x10".into())),
8439 2 : (
8440 2 : get_key(3),
8441 2 : Lsn(0x20),
8442 2 : Value::WalRecord(NeonWalRecord::wal_clear("c")),
8443 2 : ),
8444 2 : (get_key(4), Lsn(0x10), Value::Image("0x10".into())),
8445 2 : (
8446 2 : get_key(4),
8447 2 : Lsn(0x20),
8448 2 : Value::WalRecord(NeonWalRecord::wal_init("i")),
8449 2 : ),
8450 2 : ];
8451 2 : let image1 = vec![(get_key(1), "0x10".into())];
8452 2 :
8453 2 : let tline = tenant
8454 2 : .create_test_timeline_with_layers(
8455 2 : TIMELINE_ID,
8456 2 : Lsn(0x10),
8457 2 : DEFAULT_PG_VERSION,
8458 2 : &ctx,
8459 2 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
8460 2 : Lsn(0x10)..Lsn(0x40),
8461 2 : delta1,
8462 2 : )], // delta layers
8463 2 : vec![(Lsn(0x10), image1)], // image layers
8464 2 : Lsn(0x50),
8465 2 : )
8466 2 : .await?;
8467 2 :
8468 2 : assert_eq!(
8469 2 : tline.get(get_key(1), Lsn(0x50), &ctx).await?,
8470 2 : Bytes::from_static(b"0x10,0x20,0x30")
8471 2 : );
8472 2 : assert_eq!(
8473 2 : tline.get(get_key(2), Lsn(0x50), &ctx).await?,
8474 2 : Bytes::from_static(b"0x10,0x20,0x30")
8475 2 : );
8476 2 :
8477 2 : // Need to remove the limit of "Neon WAL redo requires base image".
8478 2 :
8479 2 : // assert_eq!(tline.get(get_key(3), Lsn(0x50), &ctx).await?, Bytes::new());
8480 2 : // assert_eq!(tline.get(get_key(4), Lsn(0x50), &ctx).await?, Bytes::new());
8481 2 :
8482 2 : Ok(())
8483 2 : }
8484 :
8485 : #[tokio::test(start_paused = true)]
8486 2 : async fn test_lsn_lease() -> anyhow::Result<()> {
8487 2 : let (tenant, ctx) = TenantHarness::create("test_lsn_lease")
8488 2 : .await
8489 2 : .unwrap()
8490 2 : .load()
8491 2 : .await;
8492 2 : // Advance to the lsn lease deadline so that GC is not blocked by
8493 2 : // initial transition into AttachedSingle.
8494 2 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
8495 2 : tokio::time::resume();
8496 2 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
8497 2 :
8498 2 : let end_lsn = Lsn(0x100);
8499 2 : let image_layers = (0x20..=0x90)
8500 2 : .step_by(0x10)
8501 16 : .map(|n| {
8502 16 : (
8503 16 : Lsn(n),
8504 16 : vec![(key, test_img(&format!("data key at {:x}", n)))],
8505 16 : )
8506 16 : })
8507 2 : .collect();
8508 2 :
8509 2 : let timeline = tenant
8510 2 : .create_test_timeline_with_layers(
8511 2 : TIMELINE_ID,
8512 2 : Lsn(0x10),
8513 2 : DEFAULT_PG_VERSION,
8514 2 : &ctx,
8515 2 : Vec::new(),
8516 2 : image_layers,
8517 2 : end_lsn,
8518 2 : )
8519 2 : .await?;
8520 2 :
8521 2 : let leased_lsns = [0x30, 0x50, 0x70];
8522 2 : let mut leases = Vec::new();
8523 6 : leased_lsns.iter().for_each(|n| {
8524 6 : leases.push(
8525 6 : timeline
8526 6 : .init_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)
8527 6 : .expect("lease request should succeed"),
8528 6 : );
8529 6 : });
8530 2 :
8531 2 : let updated_lease_0 = timeline
8532 2 : .renew_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)
8533 2 : .expect("lease renewal should succeed");
8534 2 : assert_eq!(
8535 2 : updated_lease_0.valid_until, leases[0].valid_until,
8536 2 : " Renewing with shorter lease should not change the lease."
8537 2 : );
8538 2 :
8539 2 : let updated_lease_1 = timeline
8540 2 : .renew_lsn_lease(
8541 2 : Lsn(leased_lsns[1]),
8542 2 : timeline.get_lsn_lease_length() * 2,
8543 2 : &ctx,
8544 2 : )
8545 2 : .expect("lease renewal should succeed");
8546 2 : assert!(
8547 2 : updated_lease_1.valid_until > leases[1].valid_until,
8548 2 : "Renewing with a long lease should renew lease with later expiration time."
8549 2 : );
8550 2 :
8551 2 : // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
8552 2 : info!(
8553 2 : "latest_gc_cutoff_lsn: {}",
8554 0 : *timeline.get_latest_gc_cutoff_lsn()
8555 2 : );
8556 2 : timeline.force_set_disk_consistent_lsn(end_lsn);
8557 2 :
8558 2 : let res = tenant
8559 2 : .gc_iteration(
8560 2 : Some(TIMELINE_ID),
8561 2 : 0,
8562 2 : Duration::ZERO,
8563 2 : &CancellationToken::new(),
8564 2 : &ctx,
8565 2 : )
8566 2 : .await
8567 2 : .unwrap();
8568 2 :
8569 2 : // Keeping everything <= Lsn(0x80) b/c leases:
8570 2 : // 0/10: initdb layer
8571 2 : // (0/20..=0/70).step_by(0x10): image layers added when creating the timeline.
8572 2 : assert_eq!(res.layers_needed_by_leases, 7);
8573 2 : // Keeping 0/90 b/c it is the latest layer.
8574 2 : assert_eq!(res.layers_not_updated, 1);
8575 2 : // Removed 0/80.
8576 2 : assert_eq!(res.layers_removed, 1);
8577 2 :
8578 2 : // Make lease on a already GC-ed LSN.
8579 2 : // 0/80 does not have a valid lease + is below latest_gc_cutoff
8580 2 : assert!(Lsn(0x80) < *timeline.get_latest_gc_cutoff_lsn());
8581 2 : timeline
8582 2 : .init_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx)
8583 2 : .expect_err("lease request on GC-ed LSN should fail");
8584 2 :
8585 2 : // Should still be able to renew a currently valid lease
8586 2 : // Assumption: original lease to is still valid for 0/50.
8587 2 : // (use `Timeline::init_lsn_lease` for testing so it always does validation)
8588 2 : timeline
8589 2 : .init_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)
8590 2 : .expect("lease renewal with validation should succeed");
8591 2 :
8592 2 : Ok(())
8593 2 : }
8594 :
8595 : #[cfg(feature = "testing")]
8596 : #[tokio::test]
8597 2 : async fn test_simple_bottom_most_compaction_deltas_1() -> anyhow::Result<()> {
8598 2 : test_simple_bottom_most_compaction_deltas_helper(
8599 2 : "test_simple_bottom_most_compaction_deltas_1",
8600 2 : false,
8601 2 : )
8602 2 : .await
8603 2 : }
8604 :
8605 : #[cfg(feature = "testing")]
8606 : #[tokio::test]
8607 2 : async fn test_simple_bottom_most_compaction_deltas_2() -> anyhow::Result<()> {
8608 2 : test_simple_bottom_most_compaction_deltas_helper(
8609 2 : "test_simple_bottom_most_compaction_deltas_2",
8610 2 : true,
8611 2 : )
8612 2 : .await
8613 2 : }
8614 :
8615 : #[cfg(feature = "testing")]
8616 4 : async fn test_simple_bottom_most_compaction_deltas_helper(
8617 4 : test_name: &'static str,
8618 4 : use_delta_bottom_layer: bool,
8619 4 : ) -> anyhow::Result<()> {
8620 4 : let harness = TenantHarness::create(test_name).await?;
8621 4 : let (tenant, ctx) = harness.load().await;
8622 :
8623 276 : fn get_key(id: u32) -> Key {
8624 276 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8625 276 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8626 276 : key.field6 = id;
8627 276 : key
8628 276 : }
8629 :
8630 : // We create
8631 : // - one bottom-most image layer,
8632 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
8633 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
8634 : // - a delta layer D3 above the horizon.
8635 : //
8636 : // | D3 |
8637 : // | D1 |
8638 : // -| |-- gc horizon -----------------
8639 : // | | | D2 |
8640 : // --------- img layer ------------------
8641 : //
8642 : // What we should expact from this compaction is:
8643 : // | D3 |
8644 : // | Part of D1 |
8645 : // --------- img layer with D1+D2 at GC horizon------------------
8646 :
8647 : // img layer at 0x10
8648 4 : let img_layer = (0..10)
8649 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8650 4 : .collect_vec();
8651 4 : // or, delta layer at 0x10 if `use_delta_bottom_layer` is true
8652 4 : let delta4 = (0..10)
8653 40 : .map(|id| {
8654 40 : (
8655 40 : get_key(id),
8656 40 : Lsn(0x08),
8657 40 : Value::WalRecord(NeonWalRecord::wal_init(format!("value {id}@0x10"))),
8658 40 : )
8659 40 : })
8660 4 : .collect_vec();
8661 4 :
8662 4 : let delta1 = vec![
8663 4 : (
8664 4 : get_key(1),
8665 4 : Lsn(0x20),
8666 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8667 4 : ),
8668 4 : (
8669 4 : get_key(2),
8670 4 : Lsn(0x30),
8671 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8672 4 : ),
8673 4 : (
8674 4 : get_key(3),
8675 4 : Lsn(0x28),
8676 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8677 4 : ),
8678 4 : (
8679 4 : get_key(3),
8680 4 : Lsn(0x30),
8681 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8682 4 : ),
8683 4 : (
8684 4 : get_key(3),
8685 4 : Lsn(0x40),
8686 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
8687 4 : ),
8688 4 : ];
8689 4 : let delta2 = vec![
8690 4 : (
8691 4 : get_key(5),
8692 4 : Lsn(0x20),
8693 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8694 4 : ),
8695 4 : (
8696 4 : get_key(6),
8697 4 : Lsn(0x20),
8698 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8699 4 : ),
8700 4 : ];
8701 4 : let delta3 = vec![
8702 4 : (
8703 4 : get_key(8),
8704 4 : Lsn(0x48),
8705 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8706 4 : ),
8707 4 : (
8708 4 : get_key(9),
8709 4 : Lsn(0x48),
8710 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8711 4 : ),
8712 4 : ];
8713 :
8714 4 : let tline = if use_delta_bottom_layer {
8715 2 : tenant
8716 2 : .create_test_timeline_with_layers(
8717 2 : TIMELINE_ID,
8718 2 : Lsn(0x08),
8719 2 : DEFAULT_PG_VERSION,
8720 2 : &ctx,
8721 2 : vec![
8722 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8723 2 : Lsn(0x08)..Lsn(0x10),
8724 2 : delta4,
8725 2 : ),
8726 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8727 2 : Lsn(0x20)..Lsn(0x48),
8728 2 : delta1,
8729 2 : ),
8730 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8731 2 : Lsn(0x20)..Lsn(0x48),
8732 2 : delta2,
8733 2 : ),
8734 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8735 2 : Lsn(0x48)..Lsn(0x50),
8736 2 : delta3,
8737 2 : ),
8738 2 : ], // delta layers
8739 2 : vec![], // image layers
8740 2 : Lsn(0x50),
8741 2 : )
8742 2 : .await?
8743 : } else {
8744 2 : tenant
8745 2 : .create_test_timeline_with_layers(
8746 2 : TIMELINE_ID,
8747 2 : Lsn(0x10),
8748 2 : DEFAULT_PG_VERSION,
8749 2 : &ctx,
8750 2 : vec![
8751 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8752 2 : Lsn(0x10)..Lsn(0x48),
8753 2 : delta1,
8754 2 : ),
8755 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8756 2 : Lsn(0x10)..Lsn(0x48),
8757 2 : delta2,
8758 2 : ),
8759 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8760 2 : Lsn(0x48)..Lsn(0x50),
8761 2 : delta3,
8762 2 : ),
8763 2 : ], // delta layers
8764 2 : vec![(Lsn(0x10), img_layer)], // image layers
8765 2 : Lsn(0x50),
8766 2 : )
8767 2 : .await?
8768 : };
8769 : {
8770 4 : tline
8771 4 : .latest_gc_cutoff_lsn
8772 4 : .lock_for_write()
8773 4 : .store_and_unlock(Lsn(0x30))
8774 4 : .wait()
8775 4 : .await;
8776 : // Update GC info
8777 4 : let mut guard = tline.gc_info.write().unwrap();
8778 4 : *guard = GcInfo {
8779 4 : retain_lsns: vec![],
8780 4 : cutoffs: GcCutoffs {
8781 4 : time: Lsn(0x30),
8782 4 : space: Lsn(0x30),
8783 4 : },
8784 4 : leases: Default::default(),
8785 4 : within_ancestor_pitr: false,
8786 4 : };
8787 4 : }
8788 4 :
8789 4 : let expected_result = [
8790 4 : Bytes::from_static(b"value 0@0x10"),
8791 4 : Bytes::from_static(b"value 1@0x10@0x20"),
8792 4 : Bytes::from_static(b"value 2@0x10@0x30"),
8793 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
8794 4 : Bytes::from_static(b"value 4@0x10"),
8795 4 : Bytes::from_static(b"value 5@0x10@0x20"),
8796 4 : Bytes::from_static(b"value 6@0x10@0x20"),
8797 4 : Bytes::from_static(b"value 7@0x10"),
8798 4 : Bytes::from_static(b"value 8@0x10@0x48"),
8799 4 : Bytes::from_static(b"value 9@0x10@0x48"),
8800 4 : ];
8801 4 :
8802 4 : let expected_result_at_gc_horizon = [
8803 4 : Bytes::from_static(b"value 0@0x10"),
8804 4 : Bytes::from_static(b"value 1@0x10@0x20"),
8805 4 : Bytes::from_static(b"value 2@0x10@0x30"),
8806 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
8807 4 : Bytes::from_static(b"value 4@0x10"),
8808 4 : Bytes::from_static(b"value 5@0x10@0x20"),
8809 4 : Bytes::from_static(b"value 6@0x10@0x20"),
8810 4 : Bytes::from_static(b"value 7@0x10"),
8811 4 : Bytes::from_static(b"value 8@0x10"),
8812 4 : Bytes::from_static(b"value 9@0x10"),
8813 4 : ];
8814 :
8815 44 : for idx in 0..10 {
8816 40 : assert_eq!(
8817 40 : tline
8818 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8819 40 : .await
8820 40 : .unwrap(),
8821 40 : &expected_result[idx]
8822 : );
8823 40 : assert_eq!(
8824 40 : tline
8825 40 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
8826 40 : .await
8827 40 : .unwrap(),
8828 40 : &expected_result_at_gc_horizon[idx]
8829 : );
8830 : }
8831 :
8832 4 : let cancel = CancellationToken::new();
8833 4 : tline
8834 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8835 4 : .await
8836 4 : .unwrap();
8837 :
8838 44 : for idx in 0..10 {
8839 40 : assert_eq!(
8840 40 : tline
8841 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8842 40 : .await
8843 40 : .unwrap(),
8844 40 : &expected_result[idx]
8845 : );
8846 40 : assert_eq!(
8847 40 : tline
8848 40 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
8849 40 : .await
8850 40 : .unwrap(),
8851 40 : &expected_result_at_gc_horizon[idx]
8852 : );
8853 : }
8854 :
8855 : // increase GC horizon and compact again
8856 : {
8857 4 : tline
8858 4 : .latest_gc_cutoff_lsn
8859 4 : .lock_for_write()
8860 4 : .store_and_unlock(Lsn(0x40))
8861 4 : .wait()
8862 4 : .await;
8863 : // Update GC info
8864 4 : let mut guard = tline.gc_info.write().unwrap();
8865 4 : guard.cutoffs.time = Lsn(0x40);
8866 4 : guard.cutoffs.space = Lsn(0x40);
8867 4 : }
8868 4 : tline
8869 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8870 4 : .await
8871 4 : .unwrap();
8872 4 :
8873 4 : Ok(())
8874 4 : }
8875 :
8876 : #[cfg(feature = "testing")]
8877 : #[tokio::test]
8878 2 : async fn test_generate_key_retention() -> anyhow::Result<()> {
8879 2 : let harness = TenantHarness::create("test_generate_key_retention").await?;
8880 2 : let (tenant, ctx) = harness.load().await;
8881 2 : let tline = tenant
8882 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
8883 2 : .await?;
8884 2 : tline.force_advance_lsn(Lsn(0x70));
8885 2 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
8886 2 : let history = vec![
8887 2 : (
8888 2 : key,
8889 2 : Lsn(0x10),
8890 2 : Value::WalRecord(NeonWalRecord::wal_init("0x10")),
8891 2 : ),
8892 2 : (
8893 2 : key,
8894 2 : Lsn(0x20),
8895 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8896 2 : ),
8897 2 : (
8898 2 : key,
8899 2 : Lsn(0x30),
8900 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8901 2 : ),
8902 2 : (
8903 2 : key,
8904 2 : Lsn(0x40),
8905 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8906 2 : ),
8907 2 : (
8908 2 : key,
8909 2 : Lsn(0x50),
8910 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
8911 2 : ),
8912 2 : (
8913 2 : key,
8914 2 : Lsn(0x60),
8915 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8916 2 : ),
8917 2 : (
8918 2 : key,
8919 2 : Lsn(0x70),
8920 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8921 2 : ),
8922 2 : (
8923 2 : key,
8924 2 : Lsn(0x80),
8925 2 : Value::Image(Bytes::copy_from_slice(
8926 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8927 2 : )),
8928 2 : ),
8929 2 : (
8930 2 : key,
8931 2 : Lsn(0x90),
8932 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8933 2 : ),
8934 2 : ];
8935 2 : let res = tline
8936 2 : .generate_key_retention(
8937 2 : key,
8938 2 : &history,
8939 2 : Lsn(0x60),
8940 2 : &[Lsn(0x20), Lsn(0x40), Lsn(0x50)],
8941 2 : 3,
8942 2 : None,
8943 2 : )
8944 2 : .await
8945 2 : .unwrap();
8946 2 : let expected_res = KeyHistoryRetention {
8947 2 : below_horizon: vec![
8948 2 : (
8949 2 : Lsn(0x20),
8950 2 : KeyLogAtLsn(vec![(
8951 2 : Lsn(0x20),
8952 2 : Value::Image(Bytes::from_static(b"0x10;0x20")),
8953 2 : )]),
8954 2 : ),
8955 2 : (
8956 2 : Lsn(0x40),
8957 2 : KeyLogAtLsn(vec![
8958 2 : (
8959 2 : Lsn(0x30),
8960 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8961 2 : ),
8962 2 : (
8963 2 : Lsn(0x40),
8964 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8965 2 : ),
8966 2 : ]),
8967 2 : ),
8968 2 : (
8969 2 : Lsn(0x50),
8970 2 : KeyLogAtLsn(vec![(
8971 2 : Lsn(0x50),
8972 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40;0x50")),
8973 2 : )]),
8974 2 : ),
8975 2 : (
8976 2 : Lsn(0x60),
8977 2 : KeyLogAtLsn(vec![(
8978 2 : Lsn(0x60),
8979 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8980 2 : )]),
8981 2 : ),
8982 2 : ],
8983 2 : above_horizon: KeyLogAtLsn(vec![
8984 2 : (
8985 2 : Lsn(0x70),
8986 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8987 2 : ),
8988 2 : (
8989 2 : Lsn(0x80),
8990 2 : Value::Image(Bytes::copy_from_slice(
8991 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8992 2 : )),
8993 2 : ),
8994 2 : (
8995 2 : Lsn(0x90),
8996 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8997 2 : ),
8998 2 : ]),
8999 2 : };
9000 2 : assert_eq!(res, expected_res);
9001 2 :
9002 2 : // We expect GC-compaction to run with the original GC. This would create a situation that
9003 2 : // the original GC algorithm removes some delta layers b/c there are full image coverage,
9004 2 : // therefore causing some keys to have an incomplete history below the lowest retain LSN.
9005 2 : // For example, we have
9006 2 : // ```plain
9007 2 : // init delta @ 0x10, image @ 0x20, delta @ 0x30 (gc_horizon), image @ 0x40.
9008 2 : // ```
9009 2 : // Now the GC horizon moves up, and we have
9010 2 : // ```plain
9011 2 : // init delta @ 0x10, image @ 0x20, delta @ 0x30, image @ 0x40 (gc_horizon)
9012 2 : // ```
9013 2 : // The original GC algorithm kicks in, and removes delta @ 0x10, image @ 0x20.
9014 2 : // We will end up with
9015 2 : // ```plain
9016 2 : // delta @ 0x30, image @ 0x40 (gc_horizon)
9017 2 : // ```
9018 2 : // Now we run the GC-compaction, and this key does not have a full history.
9019 2 : // We should be able to handle this partial history and drop everything before the
9020 2 : // gc_horizon image.
9021 2 :
9022 2 : let history = vec![
9023 2 : (
9024 2 : key,
9025 2 : Lsn(0x20),
9026 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9027 2 : ),
9028 2 : (
9029 2 : key,
9030 2 : Lsn(0x30),
9031 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9032 2 : ),
9033 2 : (
9034 2 : key,
9035 2 : Lsn(0x40),
9036 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
9037 2 : ),
9038 2 : (
9039 2 : key,
9040 2 : Lsn(0x50),
9041 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
9042 2 : ),
9043 2 : (
9044 2 : key,
9045 2 : Lsn(0x60),
9046 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9047 2 : ),
9048 2 : (
9049 2 : key,
9050 2 : Lsn(0x70),
9051 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9052 2 : ),
9053 2 : (
9054 2 : key,
9055 2 : Lsn(0x80),
9056 2 : Value::Image(Bytes::copy_from_slice(
9057 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9058 2 : )),
9059 2 : ),
9060 2 : (
9061 2 : key,
9062 2 : Lsn(0x90),
9063 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9064 2 : ),
9065 2 : ];
9066 2 : let res = tline
9067 2 : .generate_key_retention(key, &history, Lsn(0x60), &[Lsn(0x40), Lsn(0x50)], 3, None)
9068 2 : .await
9069 2 : .unwrap();
9070 2 : let expected_res = KeyHistoryRetention {
9071 2 : below_horizon: vec![
9072 2 : (
9073 2 : Lsn(0x40),
9074 2 : KeyLogAtLsn(vec![(
9075 2 : Lsn(0x40),
9076 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
9077 2 : )]),
9078 2 : ),
9079 2 : (
9080 2 : Lsn(0x50),
9081 2 : KeyLogAtLsn(vec![(
9082 2 : Lsn(0x50),
9083 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
9084 2 : )]),
9085 2 : ),
9086 2 : (
9087 2 : Lsn(0x60),
9088 2 : KeyLogAtLsn(vec![(
9089 2 : Lsn(0x60),
9090 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9091 2 : )]),
9092 2 : ),
9093 2 : ],
9094 2 : above_horizon: KeyLogAtLsn(vec![
9095 2 : (
9096 2 : Lsn(0x70),
9097 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9098 2 : ),
9099 2 : (
9100 2 : Lsn(0x80),
9101 2 : Value::Image(Bytes::copy_from_slice(
9102 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9103 2 : )),
9104 2 : ),
9105 2 : (
9106 2 : Lsn(0x90),
9107 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9108 2 : ),
9109 2 : ]),
9110 2 : };
9111 2 : assert_eq!(res, expected_res);
9112 2 :
9113 2 : // In case of branch compaction, the branch itself does not have the full history, and we need to provide
9114 2 : // the ancestor image in the test case.
9115 2 :
9116 2 : let history = vec![
9117 2 : (
9118 2 : key,
9119 2 : Lsn(0x20),
9120 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9121 2 : ),
9122 2 : (
9123 2 : key,
9124 2 : Lsn(0x30),
9125 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9126 2 : ),
9127 2 : (
9128 2 : key,
9129 2 : Lsn(0x40),
9130 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9131 2 : ),
9132 2 : (
9133 2 : key,
9134 2 : Lsn(0x70),
9135 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9136 2 : ),
9137 2 : ];
9138 2 : let res = tline
9139 2 : .generate_key_retention(
9140 2 : key,
9141 2 : &history,
9142 2 : Lsn(0x60),
9143 2 : &[],
9144 2 : 3,
9145 2 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
9146 2 : )
9147 2 : .await
9148 2 : .unwrap();
9149 2 : let expected_res = KeyHistoryRetention {
9150 2 : below_horizon: vec![(
9151 2 : Lsn(0x60),
9152 2 : KeyLogAtLsn(vec![(
9153 2 : Lsn(0x60),
9154 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")), // use the ancestor image to reconstruct the page
9155 2 : )]),
9156 2 : )],
9157 2 : above_horizon: KeyLogAtLsn(vec![(
9158 2 : Lsn(0x70),
9159 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9160 2 : )]),
9161 2 : };
9162 2 : assert_eq!(res, expected_res);
9163 2 :
9164 2 : let history = vec![
9165 2 : (
9166 2 : key,
9167 2 : Lsn(0x20),
9168 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9169 2 : ),
9170 2 : (
9171 2 : key,
9172 2 : Lsn(0x40),
9173 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9174 2 : ),
9175 2 : (
9176 2 : key,
9177 2 : Lsn(0x60),
9178 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9179 2 : ),
9180 2 : (
9181 2 : key,
9182 2 : Lsn(0x70),
9183 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9184 2 : ),
9185 2 : ];
9186 2 : let res = tline
9187 2 : .generate_key_retention(
9188 2 : key,
9189 2 : &history,
9190 2 : Lsn(0x60),
9191 2 : &[Lsn(0x30)],
9192 2 : 3,
9193 2 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
9194 2 : )
9195 2 : .await
9196 2 : .unwrap();
9197 2 : let expected_res = KeyHistoryRetention {
9198 2 : below_horizon: vec![
9199 2 : (
9200 2 : Lsn(0x30),
9201 2 : KeyLogAtLsn(vec![(
9202 2 : Lsn(0x20),
9203 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9204 2 : )]),
9205 2 : ),
9206 2 : (
9207 2 : Lsn(0x60),
9208 2 : KeyLogAtLsn(vec![(
9209 2 : Lsn(0x60),
9210 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x40;0x60")),
9211 2 : )]),
9212 2 : ),
9213 2 : ],
9214 2 : above_horizon: KeyLogAtLsn(vec![(
9215 2 : Lsn(0x70),
9216 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9217 2 : )]),
9218 2 : };
9219 2 : assert_eq!(res, expected_res);
9220 2 :
9221 2 : Ok(())
9222 2 : }
9223 :
9224 : #[cfg(feature = "testing")]
9225 : #[tokio::test]
9226 2 : async fn test_simple_bottom_most_compaction_with_retain_lsns() -> anyhow::Result<()> {
9227 2 : let harness =
9228 2 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns").await?;
9229 2 : let (tenant, ctx) = harness.load().await;
9230 2 :
9231 518 : fn get_key(id: u32) -> Key {
9232 518 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9233 518 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9234 518 : key.field6 = id;
9235 518 : key
9236 518 : }
9237 2 :
9238 2 : let img_layer = (0..10)
9239 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9240 2 : .collect_vec();
9241 2 :
9242 2 : let delta1 = vec![
9243 2 : (
9244 2 : get_key(1),
9245 2 : Lsn(0x20),
9246 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9247 2 : ),
9248 2 : (
9249 2 : get_key(2),
9250 2 : Lsn(0x30),
9251 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9252 2 : ),
9253 2 : (
9254 2 : get_key(3),
9255 2 : Lsn(0x28),
9256 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9257 2 : ),
9258 2 : (
9259 2 : get_key(3),
9260 2 : Lsn(0x30),
9261 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9262 2 : ),
9263 2 : (
9264 2 : get_key(3),
9265 2 : Lsn(0x40),
9266 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9267 2 : ),
9268 2 : ];
9269 2 : let delta2 = vec![
9270 2 : (
9271 2 : get_key(5),
9272 2 : Lsn(0x20),
9273 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9274 2 : ),
9275 2 : (
9276 2 : get_key(6),
9277 2 : Lsn(0x20),
9278 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9279 2 : ),
9280 2 : ];
9281 2 : let delta3 = vec![
9282 2 : (
9283 2 : get_key(8),
9284 2 : Lsn(0x48),
9285 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9286 2 : ),
9287 2 : (
9288 2 : get_key(9),
9289 2 : Lsn(0x48),
9290 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9291 2 : ),
9292 2 : ];
9293 2 :
9294 2 : let tline = tenant
9295 2 : .create_test_timeline_with_layers(
9296 2 : TIMELINE_ID,
9297 2 : Lsn(0x10),
9298 2 : DEFAULT_PG_VERSION,
9299 2 : &ctx,
9300 2 : vec![
9301 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
9302 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
9303 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
9304 2 : ], // delta layers
9305 2 : vec![(Lsn(0x10), img_layer)], // image layers
9306 2 : Lsn(0x50),
9307 2 : )
9308 2 : .await?;
9309 2 : {
9310 2 : tline
9311 2 : .latest_gc_cutoff_lsn
9312 2 : .lock_for_write()
9313 2 : .store_and_unlock(Lsn(0x30))
9314 2 : .wait()
9315 2 : .await;
9316 2 : // Update GC info
9317 2 : let mut guard = tline.gc_info.write().unwrap();
9318 2 : *guard = GcInfo {
9319 2 : retain_lsns: vec![
9320 2 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
9321 2 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
9322 2 : ],
9323 2 : cutoffs: GcCutoffs {
9324 2 : time: Lsn(0x30),
9325 2 : space: Lsn(0x30),
9326 2 : },
9327 2 : leases: Default::default(),
9328 2 : within_ancestor_pitr: false,
9329 2 : };
9330 2 : }
9331 2 :
9332 2 : let expected_result = [
9333 2 : Bytes::from_static(b"value 0@0x10"),
9334 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9335 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9336 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9337 2 : Bytes::from_static(b"value 4@0x10"),
9338 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9339 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9340 2 : Bytes::from_static(b"value 7@0x10"),
9341 2 : Bytes::from_static(b"value 8@0x10@0x48"),
9342 2 : Bytes::from_static(b"value 9@0x10@0x48"),
9343 2 : ];
9344 2 :
9345 2 : let expected_result_at_gc_horizon = [
9346 2 : Bytes::from_static(b"value 0@0x10"),
9347 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9348 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9349 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
9350 2 : Bytes::from_static(b"value 4@0x10"),
9351 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9352 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9353 2 : Bytes::from_static(b"value 7@0x10"),
9354 2 : Bytes::from_static(b"value 8@0x10"),
9355 2 : Bytes::from_static(b"value 9@0x10"),
9356 2 : ];
9357 2 :
9358 2 : let expected_result_at_lsn_20 = [
9359 2 : Bytes::from_static(b"value 0@0x10"),
9360 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9361 2 : Bytes::from_static(b"value 2@0x10"),
9362 2 : Bytes::from_static(b"value 3@0x10"),
9363 2 : Bytes::from_static(b"value 4@0x10"),
9364 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9365 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9366 2 : Bytes::from_static(b"value 7@0x10"),
9367 2 : Bytes::from_static(b"value 8@0x10"),
9368 2 : Bytes::from_static(b"value 9@0x10"),
9369 2 : ];
9370 2 :
9371 2 : let expected_result_at_lsn_10 = [
9372 2 : Bytes::from_static(b"value 0@0x10"),
9373 2 : Bytes::from_static(b"value 1@0x10"),
9374 2 : Bytes::from_static(b"value 2@0x10"),
9375 2 : Bytes::from_static(b"value 3@0x10"),
9376 2 : Bytes::from_static(b"value 4@0x10"),
9377 2 : Bytes::from_static(b"value 5@0x10"),
9378 2 : Bytes::from_static(b"value 6@0x10"),
9379 2 : Bytes::from_static(b"value 7@0x10"),
9380 2 : Bytes::from_static(b"value 8@0x10"),
9381 2 : Bytes::from_static(b"value 9@0x10"),
9382 2 : ];
9383 2 :
9384 12 : let verify_result = || async {
9385 12 : let gc_horizon = {
9386 12 : let gc_info = tline.gc_info.read().unwrap();
9387 12 : gc_info.cutoffs.time
9388 2 : };
9389 132 : for idx in 0..10 {
9390 120 : assert_eq!(
9391 120 : tline
9392 120 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9393 120 : .await
9394 120 : .unwrap(),
9395 120 : &expected_result[idx]
9396 2 : );
9397 120 : assert_eq!(
9398 120 : tline
9399 120 : .get(get_key(idx as u32), gc_horizon, &ctx)
9400 120 : .await
9401 120 : .unwrap(),
9402 120 : &expected_result_at_gc_horizon[idx]
9403 2 : );
9404 120 : assert_eq!(
9405 120 : tline
9406 120 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
9407 120 : .await
9408 120 : .unwrap(),
9409 120 : &expected_result_at_lsn_20[idx]
9410 2 : );
9411 120 : assert_eq!(
9412 120 : tline
9413 120 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
9414 120 : .await
9415 120 : .unwrap(),
9416 120 : &expected_result_at_lsn_10[idx]
9417 2 : );
9418 2 : }
9419 24 : };
9420 2 :
9421 2 : verify_result().await;
9422 2 :
9423 2 : let cancel = CancellationToken::new();
9424 2 : let mut dryrun_flags = EnumSet::new();
9425 2 : dryrun_flags.insert(CompactFlags::DryRun);
9426 2 :
9427 2 : tline
9428 2 : .compact_with_gc(
9429 2 : &cancel,
9430 2 : CompactOptions {
9431 2 : flags: dryrun_flags,
9432 2 : ..Default::default()
9433 2 : },
9434 2 : &ctx,
9435 2 : )
9436 2 : .await
9437 2 : .unwrap();
9438 2 : // 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
9439 2 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
9440 2 : verify_result().await;
9441 2 :
9442 2 : tline
9443 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9444 2 : .await
9445 2 : .unwrap();
9446 2 : verify_result().await;
9447 2 :
9448 2 : // compact again
9449 2 : tline
9450 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9451 2 : .await
9452 2 : .unwrap();
9453 2 : verify_result().await;
9454 2 :
9455 2 : // increase GC horizon and compact again
9456 2 : {
9457 2 : tline
9458 2 : .latest_gc_cutoff_lsn
9459 2 : .lock_for_write()
9460 2 : .store_and_unlock(Lsn(0x38))
9461 2 : .wait()
9462 2 : .await;
9463 2 : // Update GC info
9464 2 : let mut guard = tline.gc_info.write().unwrap();
9465 2 : guard.cutoffs.time = Lsn(0x38);
9466 2 : guard.cutoffs.space = Lsn(0x38);
9467 2 : }
9468 2 : tline
9469 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9470 2 : .await
9471 2 : .unwrap();
9472 2 : verify_result().await; // no wals between 0x30 and 0x38, so we should obtain the same result
9473 2 :
9474 2 : // not increasing the GC horizon and compact again
9475 2 : tline
9476 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9477 2 : .await
9478 2 : .unwrap();
9479 2 : verify_result().await;
9480 2 :
9481 2 : Ok(())
9482 2 : }
9483 :
9484 : #[cfg(feature = "testing")]
9485 : #[tokio::test]
9486 2 : async fn test_simple_bottom_most_compaction_with_retain_lsns_single_key() -> anyhow::Result<()>
9487 2 : {
9488 2 : let harness =
9489 2 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns_single_key")
9490 2 : .await?;
9491 2 : let (tenant, ctx) = harness.load().await;
9492 2 :
9493 352 : fn get_key(id: u32) -> Key {
9494 352 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9495 352 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9496 352 : key.field6 = id;
9497 352 : key
9498 352 : }
9499 2 :
9500 2 : let img_layer = (0..10)
9501 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9502 2 : .collect_vec();
9503 2 :
9504 2 : let delta1 = vec![
9505 2 : (
9506 2 : get_key(1),
9507 2 : Lsn(0x20),
9508 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9509 2 : ),
9510 2 : (
9511 2 : get_key(1),
9512 2 : Lsn(0x28),
9513 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9514 2 : ),
9515 2 : ];
9516 2 : let delta2 = vec![
9517 2 : (
9518 2 : get_key(1),
9519 2 : Lsn(0x30),
9520 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9521 2 : ),
9522 2 : (
9523 2 : get_key(1),
9524 2 : Lsn(0x38),
9525 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
9526 2 : ),
9527 2 : ];
9528 2 : let delta3 = vec![
9529 2 : (
9530 2 : get_key(8),
9531 2 : Lsn(0x48),
9532 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9533 2 : ),
9534 2 : (
9535 2 : get_key(9),
9536 2 : Lsn(0x48),
9537 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9538 2 : ),
9539 2 : ];
9540 2 :
9541 2 : let tline = tenant
9542 2 : .create_test_timeline_with_layers(
9543 2 : TIMELINE_ID,
9544 2 : Lsn(0x10),
9545 2 : DEFAULT_PG_VERSION,
9546 2 : &ctx,
9547 2 : vec![
9548 2 : // delta1 and delta 2 only contain a single key but multiple updates
9549 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x30), delta1),
9550 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
9551 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x50), delta3),
9552 2 : ], // delta layers
9553 2 : vec![(Lsn(0x10), img_layer)], // image layers
9554 2 : Lsn(0x50),
9555 2 : )
9556 2 : .await?;
9557 2 : {
9558 2 : tline
9559 2 : .latest_gc_cutoff_lsn
9560 2 : .lock_for_write()
9561 2 : .store_and_unlock(Lsn(0x30))
9562 2 : .wait()
9563 2 : .await;
9564 2 : // Update GC info
9565 2 : let mut guard = tline.gc_info.write().unwrap();
9566 2 : *guard = GcInfo {
9567 2 : retain_lsns: vec![
9568 2 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
9569 2 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
9570 2 : ],
9571 2 : cutoffs: GcCutoffs {
9572 2 : time: Lsn(0x30),
9573 2 : space: Lsn(0x30),
9574 2 : },
9575 2 : leases: Default::default(),
9576 2 : within_ancestor_pitr: false,
9577 2 : };
9578 2 : }
9579 2 :
9580 2 : let expected_result = [
9581 2 : Bytes::from_static(b"value 0@0x10"),
9582 2 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
9583 2 : Bytes::from_static(b"value 2@0x10"),
9584 2 : Bytes::from_static(b"value 3@0x10"),
9585 2 : Bytes::from_static(b"value 4@0x10"),
9586 2 : Bytes::from_static(b"value 5@0x10"),
9587 2 : Bytes::from_static(b"value 6@0x10"),
9588 2 : Bytes::from_static(b"value 7@0x10"),
9589 2 : Bytes::from_static(b"value 8@0x10@0x48"),
9590 2 : Bytes::from_static(b"value 9@0x10@0x48"),
9591 2 : ];
9592 2 :
9593 2 : let expected_result_at_gc_horizon = [
9594 2 : Bytes::from_static(b"value 0@0x10"),
9595 2 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
9596 2 : Bytes::from_static(b"value 2@0x10"),
9597 2 : Bytes::from_static(b"value 3@0x10"),
9598 2 : Bytes::from_static(b"value 4@0x10"),
9599 2 : Bytes::from_static(b"value 5@0x10"),
9600 2 : Bytes::from_static(b"value 6@0x10"),
9601 2 : Bytes::from_static(b"value 7@0x10"),
9602 2 : Bytes::from_static(b"value 8@0x10"),
9603 2 : Bytes::from_static(b"value 9@0x10"),
9604 2 : ];
9605 2 :
9606 2 : let expected_result_at_lsn_20 = [
9607 2 : Bytes::from_static(b"value 0@0x10"),
9608 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9609 2 : Bytes::from_static(b"value 2@0x10"),
9610 2 : Bytes::from_static(b"value 3@0x10"),
9611 2 : Bytes::from_static(b"value 4@0x10"),
9612 2 : Bytes::from_static(b"value 5@0x10"),
9613 2 : Bytes::from_static(b"value 6@0x10"),
9614 2 : Bytes::from_static(b"value 7@0x10"),
9615 2 : Bytes::from_static(b"value 8@0x10"),
9616 2 : Bytes::from_static(b"value 9@0x10"),
9617 2 : ];
9618 2 :
9619 2 : let expected_result_at_lsn_10 = [
9620 2 : Bytes::from_static(b"value 0@0x10"),
9621 2 : Bytes::from_static(b"value 1@0x10"),
9622 2 : Bytes::from_static(b"value 2@0x10"),
9623 2 : Bytes::from_static(b"value 3@0x10"),
9624 2 : Bytes::from_static(b"value 4@0x10"),
9625 2 : Bytes::from_static(b"value 5@0x10"),
9626 2 : Bytes::from_static(b"value 6@0x10"),
9627 2 : Bytes::from_static(b"value 7@0x10"),
9628 2 : Bytes::from_static(b"value 8@0x10"),
9629 2 : Bytes::from_static(b"value 9@0x10"),
9630 2 : ];
9631 2 :
9632 8 : let verify_result = || async {
9633 8 : let gc_horizon = {
9634 8 : let gc_info = tline.gc_info.read().unwrap();
9635 8 : gc_info.cutoffs.time
9636 2 : };
9637 88 : for idx in 0..10 {
9638 80 : assert_eq!(
9639 80 : tline
9640 80 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9641 80 : .await
9642 80 : .unwrap(),
9643 80 : &expected_result[idx]
9644 2 : );
9645 80 : assert_eq!(
9646 80 : tline
9647 80 : .get(get_key(idx as u32), gc_horizon, &ctx)
9648 80 : .await
9649 80 : .unwrap(),
9650 80 : &expected_result_at_gc_horizon[idx]
9651 2 : );
9652 80 : assert_eq!(
9653 80 : tline
9654 80 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
9655 80 : .await
9656 80 : .unwrap(),
9657 80 : &expected_result_at_lsn_20[idx]
9658 2 : );
9659 80 : assert_eq!(
9660 80 : tline
9661 80 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
9662 80 : .await
9663 80 : .unwrap(),
9664 80 : &expected_result_at_lsn_10[idx]
9665 2 : );
9666 2 : }
9667 16 : };
9668 2 :
9669 2 : verify_result().await;
9670 2 :
9671 2 : let cancel = CancellationToken::new();
9672 2 : let mut dryrun_flags = EnumSet::new();
9673 2 : dryrun_flags.insert(CompactFlags::DryRun);
9674 2 :
9675 2 : tline
9676 2 : .compact_with_gc(
9677 2 : &cancel,
9678 2 : CompactOptions {
9679 2 : flags: dryrun_flags,
9680 2 : ..Default::default()
9681 2 : },
9682 2 : &ctx,
9683 2 : )
9684 2 : .await
9685 2 : .unwrap();
9686 2 : // 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
9687 2 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
9688 2 : verify_result().await;
9689 2 :
9690 2 : tline
9691 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9692 2 : .await
9693 2 : .unwrap();
9694 2 : verify_result().await;
9695 2 :
9696 2 : // compact again
9697 2 : tline
9698 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9699 2 : .await
9700 2 : .unwrap();
9701 2 : verify_result().await;
9702 2 :
9703 2 : Ok(())
9704 2 : }
9705 :
9706 : #[cfg(feature = "testing")]
9707 : #[tokio::test]
9708 2 : async fn test_simple_bottom_most_compaction_on_branch() -> anyhow::Result<()> {
9709 2 : use models::CompactLsnRange;
9710 2 :
9711 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_on_branch").await?;
9712 2 : let (tenant, ctx) = harness.load().await;
9713 2 :
9714 166 : fn get_key(id: u32) -> Key {
9715 166 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
9716 166 : key.field6 = id;
9717 166 : key
9718 166 : }
9719 2 :
9720 2 : let img_layer = (0..10)
9721 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9722 2 : .collect_vec();
9723 2 :
9724 2 : let delta1 = vec![
9725 2 : (
9726 2 : get_key(1),
9727 2 : Lsn(0x20),
9728 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9729 2 : ),
9730 2 : (
9731 2 : get_key(2),
9732 2 : Lsn(0x30),
9733 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9734 2 : ),
9735 2 : (
9736 2 : get_key(3),
9737 2 : Lsn(0x28),
9738 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9739 2 : ),
9740 2 : (
9741 2 : get_key(3),
9742 2 : Lsn(0x30),
9743 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9744 2 : ),
9745 2 : (
9746 2 : get_key(3),
9747 2 : Lsn(0x40),
9748 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9749 2 : ),
9750 2 : ];
9751 2 : let delta2 = vec![
9752 2 : (
9753 2 : get_key(5),
9754 2 : Lsn(0x20),
9755 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9756 2 : ),
9757 2 : (
9758 2 : get_key(6),
9759 2 : Lsn(0x20),
9760 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9761 2 : ),
9762 2 : ];
9763 2 : let delta3 = vec![
9764 2 : (
9765 2 : get_key(8),
9766 2 : Lsn(0x48),
9767 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9768 2 : ),
9769 2 : (
9770 2 : get_key(9),
9771 2 : Lsn(0x48),
9772 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9773 2 : ),
9774 2 : ];
9775 2 :
9776 2 : let parent_tline = tenant
9777 2 : .create_test_timeline_with_layers(
9778 2 : TIMELINE_ID,
9779 2 : Lsn(0x10),
9780 2 : DEFAULT_PG_VERSION,
9781 2 : &ctx,
9782 2 : vec![], // delta layers
9783 2 : vec![(Lsn(0x18), img_layer)], // image layers
9784 2 : Lsn(0x18),
9785 2 : )
9786 2 : .await?;
9787 2 :
9788 2 : parent_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
9789 2 :
9790 2 : let branch_tline = tenant
9791 2 : .branch_timeline_test_with_layers(
9792 2 : &parent_tline,
9793 2 : NEW_TIMELINE_ID,
9794 2 : Some(Lsn(0x18)),
9795 2 : &ctx,
9796 2 : vec![
9797 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
9798 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
9799 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
9800 2 : ], // delta layers
9801 2 : vec![], // image layers
9802 2 : Lsn(0x50),
9803 2 : )
9804 2 : .await?;
9805 2 :
9806 2 : branch_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
9807 2 :
9808 2 : {
9809 2 : parent_tline
9810 2 : .latest_gc_cutoff_lsn
9811 2 : .lock_for_write()
9812 2 : .store_and_unlock(Lsn(0x10))
9813 2 : .wait()
9814 2 : .await;
9815 2 : // Update GC info
9816 2 : let mut guard = parent_tline.gc_info.write().unwrap();
9817 2 : *guard = GcInfo {
9818 2 : retain_lsns: vec![(Lsn(0x18), branch_tline.timeline_id, MaybeOffloaded::No)],
9819 2 : cutoffs: GcCutoffs {
9820 2 : time: Lsn(0x10),
9821 2 : space: Lsn(0x10),
9822 2 : },
9823 2 : leases: Default::default(),
9824 2 : within_ancestor_pitr: false,
9825 2 : };
9826 2 : }
9827 2 :
9828 2 : {
9829 2 : branch_tline
9830 2 : .latest_gc_cutoff_lsn
9831 2 : .lock_for_write()
9832 2 : .store_and_unlock(Lsn(0x50))
9833 2 : .wait()
9834 2 : .await;
9835 2 : // Update GC info
9836 2 : let mut guard = branch_tline.gc_info.write().unwrap();
9837 2 : *guard = GcInfo {
9838 2 : retain_lsns: vec![(Lsn(0x40), branch_tline.timeline_id, MaybeOffloaded::No)],
9839 2 : cutoffs: GcCutoffs {
9840 2 : time: Lsn(0x50),
9841 2 : space: Lsn(0x50),
9842 2 : },
9843 2 : leases: Default::default(),
9844 2 : within_ancestor_pitr: false,
9845 2 : };
9846 2 : }
9847 2 :
9848 2 : let expected_result_at_gc_horizon = [
9849 2 : Bytes::from_static(b"value 0@0x10"),
9850 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9851 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9852 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9853 2 : Bytes::from_static(b"value 4@0x10"),
9854 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9855 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9856 2 : Bytes::from_static(b"value 7@0x10"),
9857 2 : Bytes::from_static(b"value 8@0x10@0x48"),
9858 2 : Bytes::from_static(b"value 9@0x10@0x48"),
9859 2 : ];
9860 2 :
9861 2 : let expected_result_at_lsn_40 = [
9862 2 : Bytes::from_static(b"value 0@0x10"),
9863 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9864 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9865 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9866 2 : Bytes::from_static(b"value 4@0x10"),
9867 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9868 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9869 2 : Bytes::from_static(b"value 7@0x10"),
9870 2 : Bytes::from_static(b"value 8@0x10"),
9871 2 : Bytes::from_static(b"value 9@0x10"),
9872 2 : ];
9873 2 :
9874 6 : let verify_result = || async {
9875 66 : for idx in 0..10 {
9876 60 : assert_eq!(
9877 60 : branch_tline
9878 60 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9879 60 : .await
9880 60 : .unwrap(),
9881 60 : &expected_result_at_gc_horizon[idx]
9882 2 : );
9883 60 : assert_eq!(
9884 60 : branch_tline
9885 60 : .get(get_key(idx as u32), Lsn(0x40), &ctx)
9886 60 : .await
9887 60 : .unwrap(),
9888 60 : &expected_result_at_lsn_40[idx]
9889 2 : );
9890 2 : }
9891 12 : };
9892 2 :
9893 2 : verify_result().await;
9894 2 :
9895 2 : let cancel = CancellationToken::new();
9896 2 : branch_tline
9897 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9898 2 : .await
9899 2 : .unwrap();
9900 2 :
9901 2 : verify_result().await;
9902 2 :
9903 2 : // Piggyback a compaction with above_lsn. Ensure it works correctly when the specified LSN intersects with the layer files.
9904 2 : // Now we already have a single large delta layer, so the compaction min_layer_lsn should be the same as ancestor LSN (0x18).
9905 2 : branch_tline
9906 2 : .compact_with_gc(
9907 2 : &cancel,
9908 2 : CompactOptions {
9909 2 : compact_lsn_range: Some(CompactLsnRange::above(Lsn(0x40))),
9910 2 : ..Default::default()
9911 2 : },
9912 2 : &ctx,
9913 2 : )
9914 2 : .await
9915 2 : .unwrap();
9916 2 :
9917 2 : verify_result().await;
9918 2 :
9919 2 : Ok(())
9920 2 : }
9921 :
9922 : // Regression test for https://github.com/neondatabase/neon/issues/9012
9923 : // Create an image arrangement where we have to read at different LSN ranges
9924 : // from a delta layer. This is achieved by overlapping an image layer on top of
9925 : // a delta layer. Like so:
9926 : //
9927 : // A B
9928 : // +----------------+ -> delta_layer
9929 : // | | ^ lsn
9930 : // | =========|-> nested_image_layer |
9931 : // | C | |
9932 : // +----------------+ |
9933 : // ======== -> baseline_image_layer +-------> key
9934 : //
9935 : //
9936 : // When querying the key range [A, B) we need to read at different LSN ranges
9937 : // for [A, C) and [C, B). This test checks that the described edge case is handled correctly.
9938 : #[cfg(feature = "testing")]
9939 : #[tokio::test]
9940 2 : async fn test_vectored_read_with_nested_image_layer() -> anyhow::Result<()> {
9941 2 : let harness = TenantHarness::create("test_vectored_read_with_nested_image_layer").await?;
9942 2 : let (tenant, ctx) = harness.load().await;
9943 2 :
9944 2 : let will_init_keys = [2, 6];
9945 44 : fn get_key(id: u32) -> Key {
9946 44 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
9947 44 : key.field6 = id;
9948 44 : key
9949 44 : }
9950 2 :
9951 2 : let mut expected_key_values = HashMap::new();
9952 2 :
9953 2 : let baseline_image_layer_lsn = Lsn(0x10);
9954 2 : let mut baseline_img_layer = Vec::new();
9955 12 : for i in 0..5 {
9956 10 : let key = get_key(i);
9957 10 : let value = format!("value {i}@{baseline_image_layer_lsn}");
9958 10 :
9959 10 : let removed = expected_key_values.insert(key, value.clone());
9960 10 : assert!(removed.is_none());
9961 2 :
9962 10 : baseline_img_layer.push((key, Bytes::from(value)));
9963 2 : }
9964 2 :
9965 2 : let nested_image_layer_lsn = Lsn(0x50);
9966 2 : let mut nested_img_layer = Vec::new();
9967 12 : for i in 5..10 {
9968 10 : let key = get_key(i);
9969 10 : let value = format!("value {i}@{nested_image_layer_lsn}");
9970 10 :
9971 10 : let removed = expected_key_values.insert(key, value.clone());
9972 10 : assert!(removed.is_none());
9973 2 :
9974 10 : nested_img_layer.push((key, Bytes::from(value)));
9975 2 : }
9976 2 :
9977 2 : let mut delta_layer_spec = Vec::default();
9978 2 : let delta_layer_start_lsn = Lsn(0x20);
9979 2 : let mut delta_layer_end_lsn = delta_layer_start_lsn;
9980 2 :
9981 22 : for i in 0..10 {
9982 20 : let key = get_key(i);
9983 20 : let key_in_nested = nested_img_layer
9984 20 : .iter()
9985 80 : .any(|(key_with_img, _)| *key_with_img == key);
9986 20 : let lsn = {
9987 20 : if key_in_nested {
9988 10 : Lsn(nested_image_layer_lsn.0 + 0x10)
9989 2 : } else {
9990 10 : delta_layer_start_lsn
9991 2 : }
9992 2 : };
9993 2 :
9994 20 : let will_init = will_init_keys.contains(&i);
9995 20 : if will_init {
9996 4 : delta_layer_spec.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init(""))));
9997 4 :
9998 4 : expected_key_values.insert(key, "".to_string());
9999 16 : } else {
10000 16 : let delta = format!("@{lsn}");
10001 16 : delta_layer_spec.push((
10002 16 : key,
10003 16 : lsn,
10004 16 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
10005 16 : ));
10006 16 :
10007 16 : expected_key_values
10008 16 : .get_mut(&key)
10009 16 : .expect("An image exists for each key")
10010 16 : .push_str(delta.as_str());
10011 16 : }
10012 20 : delta_layer_end_lsn = std::cmp::max(delta_layer_start_lsn, lsn);
10013 2 : }
10014 2 :
10015 2 : delta_layer_end_lsn = Lsn(delta_layer_end_lsn.0 + 1);
10016 2 :
10017 2 : assert!(
10018 2 : nested_image_layer_lsn > delta_layer_start_lsn
10019 2 : && nested_image_layer_lsn < delta_layer_end_lsn
10020 2 : );
10021 2 :
10022 2 : let tline = tenant
10023 2 : .create_test_timeline_with_layers(
10024 2 : TIMELINE_ID,
10025 2 : baseline_image_layer_lsn,
10026 2 : DEFAULT_PG_VERSION,
10027 2 : &ctx,
10028 2 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
10029 2 : delta_layer_start_lsn..delta_layer_end_lsn,
10030 2 : delta_layer_spec,
10031 2 : )], // delta layers
10032 2 : vec![
10033 2 : (baseline_image_layer_lsn, baseline_img_layer),
10034 2 : (nested_image_layer_lsn, nested_img_layer),
10035 2 : ], // image layers
10036 2 : delta_layer_end_lsn,
10037 2 : )
10038 2 : .await?;
10039 2 :
10040 2 : let keyspace = KeySpace::single(get_key(0)..get_key(10));
10041 2 : let results = tline
10042 2 : .get_vectored(keyspace, delta_layer_end_lsn, &ctx)
10043 2 : .await
10044 2 : .expect("No vectored errors");
10045 22 : for (key, res) in results {
10046 20 : let value = res.expect("No key errors");
10047 20 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
10048 20 : assert_eq!(value, Bytes::from(expected_value));
10049 2 : }
10050 2 :
10051 2 : Ok(())
10052 2 : }
10053 :
10054 214 : fn sort_layer_key(k1: &PersistentLayerKey, k2: &PersistentLayerKey) -> std::cmp::Ordering {
10055 214 : (
10056 214 : k1.is_delta,
10057 214 : k1.key_range.start,
10058 214 : k1.key_range.end,
10059 214 : k1.lsn_range.start,
10060 214 : k1.lsn_range.end,
10061 214 : )
10062 214 : .cmp(&(
10063 214 : k2.is_delta,
10064 214 : k2.key_range.start,
10065 214 : k2.key_range.end,
10066 214 : k2.lsn_range.start,
10067 214 : k2.lsn_range.end,
10068 214 : ))
10069 214 : }
10070 :
10071 24 : async fn inspect_and_sort(
10072 24 : tline: &Arc<Timeline>,
10073 24 : filter: Option<std::ops::Range<Key>>,
10074 24 : ) -> Vec<PersistentLayerKey> {
10075 24 : let mut all_layers = tline.inspect_historic_layers().await.unwrap();
10076 24 : if let Some(filter) = filter {
10077 108 : all_layers.retain(|layer| overlaps_with(&layer.key_range, &filter));
10078 22 : }
10079 24 : all_layers.sort_by(sort_layer_key);
10080 24 : all_layers
10081 24 : }
10082 :
10083 : #[cfg(feature = "testing")]
10084 22 : fn check_layer_map_key_eq(
10085 22 : mut left: Vec<PersistentLayerKey>,
10086 22 : mut right: Vec<PersistentLayerKey>,
10087 22 : ) {
10088 22 : left.sort_by(sort_layer_key);
10089 22 : right.sort_by(sort_layer_key);
10090 22 : if left != right {
10091 0 : eprintln!("---LEFT---");
10092 0 : for left in left.iter() {
10093 0 : eprintln!("{}", left);
10094 0 : }
10095 0 : eprintln!("---RIGHT---");
10096 0 : for right in right.iter() {
10097 0 : eprintln!("{}", right);
10098 0 : }
10099 0 : assert_eq!(left, right);
10100 22 : }
10101 22 : }
10102 :
10103 : #[cfg(feature = "testing")]
10104 : #[tokio::test]
10105 2 : async fn test_simple_partial_bottom_most_compaction() -> anyhow::Result<()> {
10106 2 : let harness = TenantHarness::create("test_simple_partial_bottom_most_compaction").await?;
10107 2 : let (tenant, ctx) = harness.load().await;
10108 2 :
10109 182 : fn get_key(id: u32) -> Key {
10110 182 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
10111 182 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
10112 182 : key.field6 = id;
10113 182 : key
10114 182 : }
10115 2 :
10116 2 : // img layer at 0x10
10117 2 : let img_layer = (0..10)
10118 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10119 2 : .collect_vec();
10120 2 :
10121 2 : let delta1 = vec![
10122 2 : (
10123 2 : get_key(1),
10124 2 : Lsn(0x20),
10125 2 : Value::Image(Bytes::from("value 1@0x20")),
10126 2 : ),
10127 2 : (
10128 2 : get_key(2),
10129 2 : Lsn(0x30),
10130 2 : Value::Image(Bytes::from("value 2@0x30")),
10131 2 : ),
10132 2 : (
10133 2 : get_key(3),
10134 2 : Lsn(0x40),
10135 2 : Value::Image(Bytes::from("value 3@0x40")),
10136 2 : ),
10137 2 : ];
10138 2 : let delta2 = vec![
10139 2 : (
10140 2 : get_key(5),
10141 2 : Lsn(0x20),
10142 2 : Value::Image(Bytes::from("value 5@0x20")),
10143 2 : ),
10144 2 : (
10145 2 : get_key(6),
10146 2 : Lsn(0x20),
10147 2 : Value::Image(Bytes::from("value 6@0x20")),
10148 2 : ),
10149 2 : ];
10150 2 : let delta3 = vec![
10151 2 : (
10152 2 : get_key(8),
10153 2 : Lsn(0x48),
10154 2 : Value::Image(Bytes::from("value 8@0x48")),
10155 2 : ),
10156 2 : (
10157 2 : get_key(9),
10158 2 : Lsn(0x48),
10159 2 : Value::Image(Bytes::from("value 9@0x48")),
10160 2 : ),
10161 2 : ];
10162 2 :
10163 2 : let tline = tenant
10164 2 : .create_test_timeline_with_layers(
10165 2 : TIMELINE_ID,
10166 2 : Lsn(0x10),
10167 2 : DEFAULT_PG_VERSION,
10168 2 : &ctx,
10169 2 : vec![
10170 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
10171 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
10172 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
10173 2 : ], // delta layers
10174 2 : vec![(Lsn(0x10), img_layer)], // image layers
10175 2 : Lsn(0x50),
10176 2 : )
10177 2 : .await?;
10178 2 :
10179 2 : {
10180 2 : tline
10181 2 : .latest_gc_cutoff_lsn
10182 2 : .lock_for_write()
10183 2 : .store_and_unlock(Lsn(0x30))
10184 2 : .wait()
10185 2 : .await;
10186 2 : // Update GC info
10187 2 : let mut guard = tline.gc_info.write().unwrap();
10188 2 : *guard = GcInfo {
10189 2 : retain_lsns: vec![(Lsn(0x20), tline.timeline_id, MaybeOffloaded::No)],
10190 2 : cutoffs: GcCutoffs {
10191 2 : time: Lsn(0x30),
10192 2 : space: Lsn(0x30),
10193 2 : },
10194 2 : leases: Default::default(),
10195 2 : within_ancestor_pitr: false,
10196 2 : };
10197 2 : }
10198 2 :
10199 2 : let cancel = CancellationToken::new();
10200 2 :
10201 2 : // Do a partial compaction on key range 0..2
10202 2 : tline
10203 2 : .compact_with_gc(
10204 2 : &cancel,
10205 2 : CompactOptions {
10206 2 : flags: EnumSet::new(),
10207 2 : compact_key_range: Some((get_key(0)..get_key(2)).into()),
10208 2 : ..Default::default()
10209 2 : },
10210 2 : &ctx,
10211 2 : )
10212 2 : .await
10213 2 : .unwrap();
10214 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10215 2 : check_layer_map_key_eq(
10216 2 : all_layers,
10217 2 : vec![
10218 2 : // newly-generated image layer for the partial compaction range 0-2
10219 2 : PersistentLayerKey {
10220 2 : key_range: get_key(0)..get_key(2),
10221 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10222 2 : is_delta: false,
10223 2 : },
10224 2 : PersistentLayerKey {
10225 2 : key_range: get_key(0)..get_key(10),
10226 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
10227 2 : is_delta: false,
10228 2 : },
10229 2 : // delta1 is split and the second part is rewritten
10230 2 : PersistentLayerKey {
10231 2 : key_range: get_key(2)..get_key(4),
10232 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
10233 2 : is_delta: true,
10234 2 : },
10235 2 : PersistentLayerKey {
10236 2 : key_range: get_key(5)..get_key(7),
10237 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
10238 2 : is_delta: true,
10239 2 : },
10240 2 : PersistentLayerKey {
10241 2 : key_range: get_key(8)..get_key(10),
10242 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
10243 2 : is_delta: true,
10244 2 : },
10245 2 : ],
10246 2 : );
10247 2 :
10248 2 : // Do a partial compaction on key range 2..4
10249 2 : tline
10250 2 : .compact_with_gc(
10251 2 : &cancel,
10252 2 : CompactOptions {
10253 2 : flags: EnumSet::new(),
10254 2 : compact_key_range: Some((get_key(2)..get_key(4)).into()),
10255 2 : ..Default::default()
10256 2 : },
10257 2 : &ctx,
10258 2 : )
10259 2 : .await
10260 2 : .unwrap();
10261 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10262 2 : check_layer_map_key_eq(
10263 2 : all_layers,
10264 2 : vec![
10265 2 : PersistentLayerKey {
10266 2 : key_range: get_key(0)..get_key(2),
10267 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10268 2 : is_delta: false,
10269 2 : },
10270 2 : PersistentLayerKey {
10271 2 : key_range: get_key(0)..get_key(10),
10272 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
10273 2 : is_delta: false,
10274 2 : },
10275 2 : // image layer generated for the compaction range 2-4
10276 2 : PersistentLayerKey {
10277 2 : key_range: get_key(2)..get_key(4),
10278 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10279 2 : is_delta: false,
10280 2 : },
10281 2 : // we have key2/key3 above the retain_lsn, so we still need this delta layer
10282 2 : PersistentLayerKey {
10283 2 : key_range: get_key(2)..get_key(4),
10284 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
10285 2 : is_delta: true,
10286 2 : },
10287 2 : PersistentLayerKey {
10288 2 : key_range: get_key(5)..get_key(7),
10289 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
10290 2 : is_delta: true,
10291 2 : },
10292 2 : PersistentLayerKey {
10293 2 : key_range: get_key(8)..get_key(10),
10294 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
10295 2 : is_delta: true,
10296 2 : },
10297 2 : ],
10298 2 : );
10299 2 :
10300 2 : // Do a partial compaction on key range 4..9
10301 2 : tline
10302 2 : .compact_with_gc(
10303 2 : &cancel,
10304 2 : CompactOptions {
10305 2 : flags: EnumSet::new(),
10306 2 : compact_key_range: Some((get_key(4)..get_key(9)).into()),
10307 2 : ..Default::default()
10308 2 : },
10309 2 : &ctx,
10310 2 : )
10311 2 : .await
10312 2 : .unwrap();
10313 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10314 2 : check_layer_map_key_eq(
10315 2 : all_layers,
10316 2 : vec![
10317 2 : PersistentLayerKey {
10318 2 : key_range: get_key(0)..get_key(2),
10319 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10320 2 : is_delta: false,
10321 2 : },
10322 2 : PersistentLayerKey {
10323 2 : key_range: get_key(0)..get_key(10),
10324 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
10325 2 : is_delta: false,
10326 2 : },
10327 2 : PersistentLayerKey {
10328 2 : key_range: get_key(2)..get_key(4),
10329 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10330 2 : is_delta: false,
10331 2 : },
10332 2 : PersistentLayerKey {
10333 2 : key_range: get_key(2)..get_key(4),
10334 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
10335 2 : is_delta: true,
10336 2 : },
10337 2 : // image layer generated for this compaction range
10338 2 : PersistentLayerKey {
10339 2 : key_range: get_key(4)..get_key(9),
10340 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10341 2 : is_delta: false,
10342 2 : },
10343 2 : PersistentLayerKey {
10344 2 : key_range: get_key(8)..get_key(10),
10345 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
10346 2 : is_delta: true,
10347 2 : },
10348 2 : ],
10349 2 : );
10350 2 :
10351 2 : // Do a partial compaction on key range 9..10
10352 2 : tline
10353 2 : .compact_with_gc(
10354 2 : &cancel,
10355 2 : CompactOptions {
10356 2 : flags: EnumSet::new(),
10357 2 : compact_key_range: Some((get_key(9)..get_key(10)).into()),
10358 2 : ..Default::default()
10359 2 : },
10360 2 : &ctx,
10361 2 : )
10362 2 : .await
10363 2 : .unwrap();
10364 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10365 2 : check_layer_map_key_eq(
10366 2 : all_layers,
10367 2 : vec![
10368 2 : PersistentLayerKey {
10369 2 : key_range: get_key(0)..get_key(2),
10370 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10371 2 : is_delta: false,
10372 2 : },
10373 2 : PersistentLayerKey {
10374 2 : key_range: get_key(0)..get_key(10),
10375 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
10376 2 : is_delta: false,
10377 2 : },
10378 2 : PersistentLayerKey {
10379 2 : key_range: get_key(2)..get_key(4),
10380 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10381 2 : is_delta: false,
10382 2 : },
10383 2 : PersistentLayerKey {
10384 2 : key_range: get_key(2)..get_key(4),
10385 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
10386 2 : is_delta: true,
10387 2 : },
10388 2 : PersistentLayerKey {
10389 2 : key_range: get_key(4)..get_key(9),
10390 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10391 2 : is_delta: false,
10392 2 : },
10393 2 : // image layer generated for the compaction range
10394 2 : PersistentLayerKey {
10395 2 : key_range: get_key(9)..get_key(10),
10396 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10397 2 : is_delta: false,
10398 2 : },
10399 2 : PersistentLayerKey {
10400 2 : key_range: get_key(8)..get_key(10),
10401 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
10402 2 : is_delta: true,
10403 2 : },
10404 2 : ],
10405 2 : );
10406 2 :
10407 2 : // Do a partial compaction on key range 0..10, all image layers below LSN 20 can be replaced with new ones.
10408 2 : tline
10409 2 : .compact_with_gc(
10410 2 : &cancel,
10411 2 : CompactOptions {
10412 2 : flags: EnumSet::new(),
10413 2 : compact_key_range: Some((get_key(0)..get_key(10)).into()),
10414 2 : ..Default::default()
10415 2 : },
10416 2 : &ctx,
10417 2 : )
10418 2 : .await
10419 2 : .unwrap();
10420 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10421 2 : check_layer_map_key_eq(
10422 2 : all_layers,
10423 2 : vec![
10424 2 : // aha, we removed all unnecessary image/delta layers and got a very clean layer map!
10425 2 : PersistentLayerKey {
10426 2 : key_range: get_key(0)..get_key(10),
10427 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10428 2 : is_delta: false,
10429 2 : },
10430 2 : PersistentLayerKey {
10431 2 : key_range: get_key(2)..get_key(4),
10432 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
10433 2 : is_delta: true,
10434 2 : },
10435 2 : PersistentLayerKey {
10436 2 : key_range: get_key(8)..get_key(10),
10437 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
10438 2 : is_delta: true,
10439 2 : },
10440 2 : ],
10441 2 : );
10442 2 : Ok(())
10443 2 : }
10444 :
10445 : #[cfg(feature = "testing")]
10446 : #[tokio::test]
10447 2 : async fn test_timeline_offload_retain_lsn() -> anyhow::Result<()> {
10448 2 : let harness = TenantHarness::create("test_timeline_offload_retain_lsn")
10449 2 : .await
10450 2 : .unwrap();
10451 2 : let (tenant, ctx) = harness.load().await;
10452 2 : let tline_parent = tenant
10453 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
10454 2 : .await
10455 2 : .unwrap();
10456 2 : let tline_child = tenant
10457 2 : .branch_timeline_test(&tline_parent, NEW_TIMELINE_ID, Some(Lsn(0x20)), &ctx)
10458 2 : .await
10459 2 : .unwrap();
10460 2 : {
10461 2 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
10462 2 : assert_eq!(
10463 2 : gc_info_parent.retain_lsns,
10464 2 : vec![(Lsn(0x20), tline_child.timeline_id, MaybeOffloaded::No)]
10465 2 : );
10466 2 : }
10467 2 : // We have to directly call the remote_client instead of using the archive function to avoid constructing broker client...
10468 2 : tline_child
10469 2 : .remote_client
10470 2 : .schedule_index_upload_for_timeline_archival_state(TimelineArchivalState::Archived)
10471 2 : .unwrap();
10472 2 : tline_child.remote_client.wait_completion().await.unwrap();
10473 2 : offload_timeline(&tenant, &tline_child)
10474 2 : .instrument(tracing::info_span!(parent: None, "offload_test", tenant_id=%"test", shard_id=%"test", timeline_id=%"test"))
10475 2 : .await.unwrap();
10476 2 : let child_timeline_id = tline_child.timeline_id;
10477 2 : Arc::try_unwrap(tline_child).unwrap();
10478 2 :
10479 2 : {
10480 2 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
10481 2 : assert_eq!(
10482 2 : gc_info_parent.retain_lsns,
10483 2 : vec![(Lsn(0x20), child_timeline_id, MaybeOffloaded::Yes)]
10484 2 : );
10485 2 : }
10486 2 :
10487 2 : tenant
10488 2 : .get_offloaded_timeline(child_timeline_id)
10489 2 : .unwrap()
10490 2 : .defuse_for_tenant_drop();
10491 2 :
10492 2 : Ok(())
10493 2 : }
10494 :
10495 : #[cfg(feature = "testing")]
10496 : #[tokio::test]
10497 2 : async fn test_simple_bottom_most_compaction_above_lsn() -> anyhow::Result<()> {
10498 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_above_lsn").await?;
10499 2 : let (tenant, ctx) = harness.load().await;
10500 2 :
10501 296 : fn get_key(id: u32) -> Key {
10502 296 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
10503 296 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
10504 296 : key.field6 = id;
10505 296 : key
10506 296 : }
10507 2 :
10508 2 : let img_layer = (0..10)
10509 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10510 2 : .collect_vec();
10511 2 :
10512 2 : let delta1 = vec![(
10513 2 : get_key(1),
10514 2 : Lsn(0x20),
10515 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10516 2 : )];
10517 2 : let delta4 = vec![(
10518 2 : get_key(1),
10519 2 : Lsn(0x28),
10520 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
10521 2 : )];
10522 2 : let delta2 = vec![
10523 2 : (
10524 2 : get_key(1),
10525 2 : Lsn(0x30),
10526 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10527 2 : ),
10528 2 : (
10529 2 : get_key(1),
10530 2 : Lsn(0x38),
10531 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
10532 2 : ),
10533 2 : ];
10534 2 : let delta3 = vec![
10535 2 : (
10536 2 : get_key(8),
10537 2 : Lsn(0x48),
10538 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10539 2 : ),
10540 2 : (
10541 2 : get_key(9),
10542 2 : Lsn(0x48),
10543 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10544 2 : ),
10545 2 : ];
10546 2 :
10547 2 : let tline = tenant
10548 2 : .create_test_timeline_with_layers(
10549 2 : TIMELINE_ID,
10550 2 : Lsn(0x10),
10551 2 : DEFAULT_PG_VERSION,
10552 2 : &ctx,
10553 2 : vec![
10554 2 : // delta1/2/4 only contain a single key but multiple updates
10555 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x28), delta1),
10556 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
10557 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x28)..Lsn(0x30), delta4),
10558 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta3),
10559 2 : ], // delta layers
10560 2 : vec![(Lsn(0x10), img_layer)], // image layers
10561 2 : Lsn(0x50),
10562 2 : )
10563 2 : .await?;
10564 2 : {
10565 2 : tline
10566 2 : .latest_gc_cutoff_lsn
10567 2 : .lock_for_write()
10568 2 : .store_and_unlock(Lsn(0x30))
10569 2 : .wait()
10570 2 : .await;
10571 2 : // Update GC info
10572 2 : let mut guard = tline.gc_info.write().unwrap();
10573 2 : *guard = GcInfo {
10574 2 : retain_lsns: vec![
10575 2 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
10576 2 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
10577 2 : ],
10578 2 : cutoffs: GcCutoffs {
10579 2 : time: Lsn(0x30),
10580 2 : space: Lsn(0x30),
10581 2 : },
10582 2 : leases: Default::default(),
10583 2 : within_ancestor_pitr: false,
10584 2 : };
10585 2 : }
10586 2 :
10587 2 : let expected_result = [
10588 2 : Bytes::from_static(b"value 0@0x10"),
10589 2 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
10590 2 : Bytes::from_static(b"value 2@0x10"),
10591 2 : Bytes::from_static(b"value 3@0x10"),
10592 2 : Bytes::from_static(b"value 4@0x10"),
10593 2 : Bytes::from_static(b"value 5@0x10"),
10594 2 : Bytes::from_static(b"value 6@0x10"),
10595 2 : Bytes::from_static(b"value 7@0x10"),
10596 2 : Bytes::from_static(b"value 8@0x10@0x48"),
10597 2 : Bytes::from_static(b"value 9@0x10@0x48"),
10598 2 : ];
10599 2 :
10600 2 : let expected_result_at_gc_horizon = [
10601 2 : Bytes::from_static(b"value 0@0x10"),
10602 2 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
10603 2 : Bytes::from_static(b"value 2@0x10"),
10604 2 : Bytes::from_static(b"value 3@0x10"),
10605 2 : Bytes::from_static(b"value 4@0x10"),
10606 2 : Bytes::from_static(b"value 5@0x10"),
10607 2 : Bytes::from_static(b"value 6@0x10"),
10608 2 : Bytes::from_static(b"value 7@0x10"),
10609 2 : Bytes::from_static(b"value 8@0x10"),
10610 2 : Bytes::from_static(b"value 9@0x10"),
10611 2 : ];
10612 2 :
10613 2 : let expected_result_at_lsn_20 = [
10614 2 : Bytes::from_static(b"value 0@0x10"),
10615 2 : Bytes::from_static(b"value 1@0x10@0x20"),
10616 2 : Bytes::from_static(b"value 2@0x10"),
10617 2 : Bytes::from_static(b"value 3@0x10"),
10618 2 : Bytes::from_static(b"value 4@0x10"),
10619 2 : Bytes::from_static(b"value 5@0x10"),
10620 2 : Bytes::from_static(b"value 6@0x10"),
10621 2 : Bytes::from_static(b"value 7@0x10"),
10622 2 : Bytes::from_static(b"value 8@0x10"),
10623 2 : Bytes::from_static(b"value 9@0x10"),
10624 2 : ];
10625 2 :
10626 2 : let expected_result_at_lsn_10 = [
10627 2 : Bytes::from_static(b"value 0@0x10"),
10628 2 : Bytes::from_static(b"value 1@0x10"),
10629 2 : Bytes::from_static(b"value 2@0x10"),
10630 2 : Bytes::from_static(b"value 3@0x10"),
10631 2 : Bytes::from_static(b"value 4@0x10"),
10632 2 : Bytes::from_static(b"value 5@0x10"),
10633 2 : Bytes::from_static(b"value 6@0x10"),
10634 2 : Bytes::from_static(b"value 7@0x10"),
10635 2 : Bytes::from_static(b"value 8@0x10"),
10636 2 : Bytes::from_static(b"value 9@0x10"),
10637 2 : ];
10638 2 :
10639 6 : let verify_result = || async {
10640 6 : let gc_horizon = {
10641 6 : let gc_info = tline.gc_info.read().unwrap();
10642 6 : gc_info.cutoffs.time
10643 2 : };
10644 66 : for idx in 0..10 {
10645 60 : assert_eq!(
10646 60 : tline
10647 60 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10648 60 : .await
10649 60 : .unwrap(),
10650 60 : &expected_result[idx]
10651 2 : );
10652 60 : assert_eq!(
10653 60 : tline
10654 60 : .get(get_key(idx as u32), gc_horizon, &ctx)
10655 60 : .await
10656 60 : .unwrap(),
10657 60 : &expected_result_at_gc_horizon[idx]
10658 2 : );
10659 60 : assert_eq!(
10660 60 : tline
10661 60 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
10662 60 : .await
10663 60 : .unwrap(),
10664 60 : &expected_result_at_lsn_20[idx]
10665 2 : );
10666 60 : assert_eq!(
10667 60 : tline
10668 60 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
10669 60 : .await
10670 60 : .unwrap(),
10671 60 : &expected_result_at_lsn_10[idx]
10672 2 : );
10673 2 : }
10674 12 : };
10675 2 :
10676 2 : verify_result().await;
10677 2 :
10678 2 : let cancel = CancellationToken::new();
10679 2 : tline
10680 2 : .compact_with_gc(
10681 2 : &cancel,
10682 2 : CompactOptions {
10683 2 : compact_lsn_range: Some(CompactLsnRange::above(Lsn(0x28))),
10684 2 : ..Default::default()
10685 2 : },
10686 2 : &ctx,
10687 2 : )
10688 2 : .await
10689 2 : .unwrap();
10690 2 : verify_result().await;
10691 2 :
10692 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10693 2 : check_layer_map_key_eq(
10694 2 : all_layers,
10695 2 : vec![
10696 2 : // The original image layer, not compacted
10697 2 : PersistentLayerKey {
10698 2 : key_range: get_key(0)..get_key(10),
10699 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
10700 2 : is_delta: false,
10701 2 : },
10702 2 : // Delta layer below the specified above_lsn not compacted
10703 2 : PersistentLayerKey {
10704 2 : key_range: get_key(1)..get_key(2),
10705 2 : lsn_range: Lsn(0x20)..Lsn(0x28),
10706 2 : is_delta: true,
10707 2 : },
10708 2 : // Delta layer compacted above the LSN
10709 2 : PersistentLayerKey {
10710 2 : key_range: get_key(1)..get_key(10),
10711 2 : lsn_range: Lsn(0x28)..Lsn(0x50),
10712 2 : is_delta: true,
10713 2 : },
10714 2 : ],
10715 2 : );
10716 2 :
10717 2 : // compact again
10718 2 : tline
10719 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10720 2 : .await
10721 2 : .unwrap();
10722 2 : verify_result().await;
10723 2 :
10724 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10725 2 : check_layer_map_key_eq(
10726 2 : all_layers,
10727 2 : vec![
10728 2 : // The compacted image layer (full key range)
10729 2 : PersistentLayerKey {
10730 2 : key_range: Key::MIN..Key::MAX,
10731 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
10732 2 : is_delta: false,
10733 2 : },
10734 2 : // All other data in the delta layer
10735 2 : PersistentLayerKey {
10736 2 : key_range: get_key(1)..get_key(10),
10737 2 : lsn_range: Lsn(0x10)..Lsn(0x50),
10738 2 : is_delta: true,
10739 2 : },
10740 2 : ],
10741 2 : );
10742 2 :
10743 2 : Ok(())
10744 2 : }
10745 :
10746 : #[cfg(feature = "testing")]
10747 : #[tokio::test]
10748 2 : async fn test_simple_bottom_most_compaction_rectangle() -> anyhow::Result<()> {
10749 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_rectangle").await?;
10750 2 : let (tenant, ctx) = harness.load().await;
10751 2 :
10752 508 : fn get_key(id: u32) -> Key {
10753 508 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
10754 508 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
10755 508 : key.field6 = id;
10756 508 : key
10757 508 : }
10758 2 :
10759 2 : let img_layer = (0..10)
10760 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10761 2 : .collect_vec();
10762 2 :
10763 2 : let delta1 = vec![(
10764 2 : get_key(1),
10765 2 : Lsn(0x20),
10766 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10767 2 : )];
10768 2 : let delta4 = vec![(
10769 2 : get_key(1),
10770 2 : Lsn(0x28),
10771 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
10772 2 : )];
10773 2 : let delta2 = vec![
10774 2 : (
10775 2 : get_key(1),
10776 2 : Lsn(0x30),
10777 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10778 2 : ),
10779 2 : (
10780 2 : get_key(1),
10781 2 : Lsn(0x38),
10782 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
10783 2 : ),
10784 2 : ];
10785 2 : let delta3 = vec![
10786 2 : (
10787 2 : get_key(8),
10788 2 : Lsn(0x48),
10789 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10790 2 : ),
10791 2 : (
10792 2 : get_key(9),
10793 2 : Lsn(0x48),
10794 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10795 2 : ),
10796 2 : ];
10797 2 :
10798 2 : let tline = tenant
10799 2 : .create_test_timeline_with_layers(
10800 2 : TIMELINE_ID,
10801 2 : Lsn(0x10),
10802 2 : DEFAULT_PG_VERSION,
10803 2 : &ctx,
10804 2 : vec![
10805 2 : // delta1/2/4 only contain a single key but multiple updates
10806 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x28), delta1),
10807 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
10808 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x28)..Lsn(0x30), delta4),
10809 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta3),
10810 2 : ], // delta layers
10811 2 : vec![(Lsn(0x10), img_layer)], // image layers
10812 2 : Lsn(0x50),
10813 2 : )
10814 2 : .await?;
10815 2 : {
10816 2 : tline
10817 2 : .latest_gc_cutoff_lsn
10818 2 : .lock_for_write()
10819 2 : .store_and_unlock(Lsn(0x30))
10820 2 : .wait()
10821 2 : .await;
10822 2 : // Update GC info
10823 2 : let mut guard = tline.gc_info.write().unwrap();
10824 2 : *guard = GcInfo {
10825 2 : retain_lsns: vec![
10826 2 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
10827 2 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
10828 2 : ],
10829 2 : cutoffs: GcCutoffs {
10830 2 : time: Lsn(0x30),
10831 2 : space: Lsn(0x30),
10832 2 : },
10833 2 : leases: Default::default(),
10834 2 : within_ancestor_pitr: false,
10835 2 : };
10836 2 : }
10837 2 :
10838 2 : let expected_result = [
10839 2 : Bytes::from_static(b"value 0@0x10"),
10840 2 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
10841 2 : Bytes::from_static(b"value 2@0x10"),
10842 2 : Bytes::from_static(b"value 3@0x10"),
10843 2 : Bytes::from_static(b"value 4@0x10"),
10844 2 : Bytes::from_static(b"value 5@0x10"),
10845 2 : Bytes::from_static(b"value 6@0x10"),
10846 2 : Bytes::from_static(b"value 7@0x10"),
10847 2 : Bytes::from_static(b"value 8@0x10@0x48"),
10848 2 : Bytes::from_static(b"value 9@0x10@0x48"),
10849 2 : ];
10850 2 :
10851 2 : let expected_result_at_gc_horizon = [
10852 2 : Bytes::from_static(b"value 0@0x10"),
10853 2 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
10854 2 : Bytes::from_static(b"value 2@0x10"),
10855 2 : Bytes::from_static(b"value 3@0x10"),
10856 2 : Bytes::from_static(b"value 4@0x10"),
10857 2 : Bytes::from_static(b"value 5@0x10"),
10858 2 : Bytes::from_static(b"value 6@0x10"),
10859 2 : Bytes::from_static(b"value 7@0x10"),
10860 2 : Bytes::from_static(b"value 8@0x10"),
10861 2 : Bytes::from_static(b"value 9@0x10"),
10862 2 : ];
10863 2 :
10864 2 : let expected_result_at_lsn_20 = [
10865 2 : Bytes::from_static(b"value 0@0x10"),
10866 2 : Bytes::from_static(b"value 1@0x10@0x20"),
10867 2 : Bytes::from_static(b"value 2@0x10"),
10868 2 : Bytes::from_static(b"value 3@0x10"),
10869 2 : Bytes::from_static(b"value 4@0x10"),
10870 2 : Bytes::from_static(b"value 5@0x10"),
10871 2 : Bytes::from_static(b"value 6@0x10"),
10872 2 : Bytes::from_static(b"value 7@0x10"),
10873 2 : Bytes::from_static(b"value 8@0x10"),
10874 2 : Bytes::from_static(b"value 9@0x10"),
10875 2 : ];
10876 2 :
10877 2 : let expected_result_at_lsn_10 = [
10878 2 : Bytes::from_static(b"value 0@0x10"),
10879 2 : Bytes::from_static(b"value 1@0x10"),
10880 2 : Bytes::from_static(b"value 2@0x10"),
10881 2 : Bytes::from_static(b"value 3@0x10"),
10882 2 : Bytes::from_static(b"value 4@0x10"),
10883 2 : Bytes::from_static(b"value 5@0x10"),
10884 2 : Bytes::from_static(b"value 6@0x10"),
10885 2 : Bytes::from_static(b"value 7@0x10"),
10886 2 : Bytes::from_static(b"value 8@0x10"),
10887 2 : Bytes::from_static(b"value 9@0x10"),
10888 2 : ];
10889 2 :
10890 10 : let verify_result = || async {
10891 10 : let gc_horizon = {
10892 10 : let gc_info = tline.gc_info.read().unwrap();
10893 10 : gc_info.cutoffs.time
10894 2 : };
10895 110 : for idx in 0..10 {
10896 100 : assert_eq!(
10897 100 : tline
10898 100 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10899 100 : .await
10900 100 : .unwrap(),
10901 100 : &expected_result[idx]
10902 2 : );
10903 100 : assert_eq!(
10904 100 : tline
10905 100 : .get(get_key(idx as u32), gc_horizon, &ctx)
10906 100 : .await
10907 100 : .unwrap(),
10908 100 : &expected_result_at_gc_horizon[idx]
10909 2 : );
10910 100 : assert_eq!(
10911 100 : tline
10912 100 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
10913 100 : .await
10914 100 : .unwrap(),
10915 100 : &expected_result_at_lsn_20[idx]
10916 2 : );
10917 100 : assert_eq!(
10918 100 : tline
10919 100 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
10920 100 : .await
10921 100 : .unwrap(),
10922 100 : &expected_result_at_lsn_10[idx]
10923 2 : );
10924 2 : }
10925 20 : };
10926 2 :
10927 2 : verify_result().await;
10928 2 :
10929 2 : let cancel = CancellationToken::new();
10930 2 :
10931 2 : tline
10932 2 : .compact_with_gc(
10933 2 : &cancel,
10934 2 : CompactOptions {
10935 2 : compact_key_range: Some((get_key(0)..get_key(2)).into()),
10936 2 : compact_lsn_range: Some((Lsn(0x20)..Lsn(0x28)).into()),
10937 2 : ..Default::default()
10938 2 : },
10939 2 : &ctx,
10940 2 : )
10941 2 : .await
10942 2 : .unwrap();
10943 2 : verify_result().await;
10944 2 :
10945 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10946 2 : check_layer_map_key_eq(
10947 2 : all_layers,
10948 2 : vec![
10949 2 : // The original image layer, not compacted
10950 2 : PersistentLayerKey {
10951 2 : key_range: get_key(0)..get_key(10),
10952 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
10953 2 : is_delta: false,
10954 2 : },
10955 2 : // According the selection logic, we select all layers with start key <= 0x28, so we would merge the layer 0x20-0x28 and
10956 2 : // the layer 0x28-0x30 into one.
10957 2 : PersistentLayerKey {
10958 2 : key_range: get_key(1)..get_key(2),
10959 2 : lsn_range: Lsn(0x20)..Lsn(0x30),
10960 2 : is_delta: true,
10961 2 : },
10962 2 : // Above the upper bound and untouched
10963 2 : PersistentLayerKey {
10964 2 : key_range: get_key(1)..get_key(2),
10965 2 : lsn_range: Lsn(0x30)..Lsn(0x50),
10966 2 : is_delta: true,
10967 2 : },
10968 2 : // This layer is untouched
10969 2 : PersistentLayerKey {
10970 2 : key_range: get_key(8)..get_key(10),
10971 2 : lsn_range: Lsn(0x30)..Lsn(0x50),
10972 2 : is_delta: true,
10973 2 : },
10974 2 : ],
10975 2 : );
10976 2 :
10977 2 : tline
10978 2 : .compact_with_gc(
10979 2 : &cancel,
10980 2 : CompactOptions {
10981 2 : compact_key_range: Some((get_key(3)..get_key(8)).into()),
10982 2 : compact_lsn_range: Some((Lsn(0x28)..Lsn(0x40)).into()),
10983 2 : ..Default::default()
10984 2 : },
10985 2 : &ctx,
10986 2 : )
10987 2 : .await
10988 2 : .unwrap();
10989 2 : verify_result().await;
10990 2 :
10991 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10992 2 : check_layer_map_key_eq(
10993 2 : all_layers,
10994 2 : vec![
10995 2 : // The original image layer, not compacted
10996 2 : PersistentLayerKey {
10997 2 : key_range: get_key(0)..get_key(10),
10998 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
10999 2 : is_delta: false,
11000 2 : },
11001 2 : // Not in the compaction key range, uncompacted
11002 2 : PersistentLayerKey {
11003 2 : key_range: get_key(1)..get_key(2),
11004 2 : lsn_range: Lsn(0x20)..Lsn(0x30),
11005 2 : is_delta: true,
11006 2 : },
11007 2 : // Not in the compaction key range, uncompacted but need rewrite because the delta layer overlaps with the range
11008 2 : PersistentLayerKey {
11009 2 : key_range: get_key(1)..get_key(2),
11010 2 : lsn_range: Lsn(0x30)..Lsn(0x50),
11011 2 : is_delta: true,
11012 2 : },
11013 2 : // Note that when we specify the LSN upper bound to be 0x40, the compaction algorithm will not try to cut the layer
11014 2 : // horizontally in half. Instead, it will include all LSNs that overlap with 0x40. So the real max_lsn of the compaction
11015 2 : // becomes 0x50.
11016 2 : PersistentLayerKey {
11017 2 : key_range: get_key(8)..get_key(10),
11018 2 : lsn_range: Lsn(0x30)..Lsn(0x50),
11019 2 : is_delta: true,
11020 2 : },
11021 2 : ],
11022 2 : );
11023 2 :
11024 2 : // compact again
11025 2 : tline
11026 2 : .compact_with_gc(
11027 2 : &cancel,
11028 2 : CompactOptions {
11029 2 : compact_key_range: Some((get_key(0)..get_key(5)).into()),
11030 2 : compact_lsn_range: Some((Lsn(0x20)..Lsn(0x50)).into()),
11031 2 : ..Default::default()
11032 2 : },
11033 2 : &ctx,
11034 2 : )
11035 2 : .await
11036 2 : .unwrap();
11037 2 : verify_result().await;
11038 2 :
11039 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11040 2 : check_layer_map_key_eq(
11041 2 : all_layers,
11042 2 : vec![
11043 2 : // The original image layer, not compacted
11044 2 : PersistentLayerKey {
11045 2 : key_range: get_key(0)..get_key(10),
11046 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
11047 2 : is_delta: false,
11048 2 : },
11049 2 : // The range gets compacted
11050 2 : PersistentLayerKey {
11051 2 : key_range: get_key(1)..get_key(2),
11052 2 : lsn_range: Lsn(0x20)..Lsn(0x50),
11053 2 : is_delta: true,
11054 2 : },
11055 2 : // Not touched during this iteration of compaction
11056 2 : PersistentLayerKey {
11057 2 : key_range: get_key(8)..get_key(10),
11058 2 : lsn_range: Lsn(0x30)..Lsn(0x50),
11059 2 : is_delta: true,
11060 2 : },
11061 2 : ],
11062 2 : );
11063 2 :
11064 2 : // final full compaction
11065 2 : tline
11066 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
11067 2 : .await
11068 2 : .unwrap();
11069 2 : verify_result().await;
11070 2 :
11071 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11072 2 : check_layer_map_key_eq(
11073 2 : all_layers,
11074 2 : vec![
11075 2 : // The compacted image layer (full key range)
11076 2 : PersistentLayerKey {
11077 2 : key_range: Key::MIN..Key::MAX,
11078 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
11079 2 : is_delta: false,
11080 2 : },
11081 2 : // All other data in the delta layer
11082 2 : PersistentLayerKey {
11083 2 : key_range: get_key(1)..get_key(10),
11084 2 : lsn_range: Lsn(0x10)..Lsn(0x50),
11085 2 : is_delta: true,
11086 2 : },
11087 2 : ],
11088 2 : );
11089 2 :
11090 2 : Ok(())
11091 2 : }
11092 : }
|
