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::LsnLease;
25 : use pageserver_api::models::TimelineArchivalState;
26 : use pageserver_api::models::TimelineState;
27 : use pageserver_api::models::TopTenantShardItem;
28 : use pageserver_api::models::WalRedoManagerStatus;
29 : use pageserver_api::shard::ShardIdentity;
30 : use pageserver_api::shard::ShardStripeSize;
31 : use pageserver_api::shard::TenantShardId;
32 : use remote_storage::DownloadError;
33 : use remote_storage::GenericRemoteStorage;
34 : use remote_storage::TimeoutOrCancel;
35 : use remote_timeline_client::manifest::{
36 : OffloadedTimelineManifest, TenantManifest, LATEST_TENANT_MANIFEST_VERSION,
37 : };
38 : use remote_timeline_client::UploadQueueNotReadyError;
39 : use std::collections::BTreeMap;
40 : use std::collections::VecDeque;
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::GcCompactJob;
48 : use timeline::compaction::ScheduledCompactionTask;
49 : use timeline::import_pgdata;
50 : use timeline::offload::offload_timeline;
51 : use timeline::CompactFlags;
52 : use timeline::CompactOptions;
53 : use timeline::CompactionError;
54 : use timeline::ShutdownMode;
55 : use tokio::io::BufReader;
56 : use tokio::sync::watch;
57 : use tokio::task::JoinSet;
58 : use tokio_util::sync::CancellationToken;
59 : use tracing::*;
60 : use upload_queue::NotInitialized;
61 : use utils::backoff;
62 : use utils::circuit_breaker::CircuitBreaker;
63 : use utils::completion;
64 : use utils::crashsafe::path_with_suffix_extension;
65 : use utils::failpoint_support;
66 : use utils::fs_ext;
67 : use utils::pausable_failpoint;
68 : use utils::sync::gate::Gate;
69 : use utils::sync::gate::GateGuard;
70 : use utils::timeout::timeout_cancellable;
71 : use utils::timeout::TimeoutCancellableError;
72 : use utils::try_rcu::ArcSwapExt;
73 : use utils::zstd::create_zst_tarball;
74 : use utils::zstd::extract_zst_tarball;
75 :
76 : use self::config::AttachedLocationConfig;
77 : use self::config::AttachmentMode;
78 : use self::config::LocationConf;
79 : use self::config::TenantConf;
80 : use self::metadata::TimelineMetadata;
81 : use self::mgr::GetActiveTenantError;
82 : use self::mgr::GetTenantError;
83 : use self::remote_timeline_client::upload::{upload_index_part, upload_tenant_manifest};
84 : use self::remote_timeline_client::{RemoteTimelineClient, WaitCompletionError};
85 : use self::timeline::uninit::TimelineCreateGuard;
86 : use self::timeline::uninit::TimelineExclusionError;
87 : use self::timeline::uninit::UninitializedTimeline;
88 : use self::timeline::EvictionTaskTenantState;
89 : use self::timeline::GcCutoffs;
90 : use self::timeline::TimelineDeleteProgress;
91 : use self::timeline::TimelineResources;
92 : use self::timeline::WaitLsnError;
93 : use crate::config::PageServerConf;
94 : use crate::context::{DownloadBehavior, RequestContext};
95 : use crate::deletion_queue::DeletionQueueClient;
96 : use crate::deletion_queue::DeletionQueueError;
97 : use crate::import_datadir;
98 : use crate::is_uninit_mark;
99 : use crate::l0_flush::L0FlushGlobalState;
100 : use crate::metrics::TENANT;
101 : use crate::metrics::{
102 : remove_tenant_metrics, BROKEN_TENANTS_SET, CIRCUIT_BREAKERS_BROKEN, CIRCUIT_BREAKERS_UNBROKEN,
103 : TENANT_STATE_METRIC, TENANT_SYNTHETIC_SIZE_METRIC,
104 : };
105 : use crate::task_mgr;
106 : use crate::task_mgr::TaskKind;
107 : use crate::tenant::config::LocationMode;
108 : use crate::tenant::config::TenantConfOpt;
109 : use crate::tenant::gc_result::GcResult;
110 : pub use crate::tenant::remote_timeline_client::index::IndexPart;
111 : use crate::tenant::remote_timeline_client::remote_initdb_archive_path;
112 : use crate::tenant::remote_timeline_client::MaybeDeletedIndexPart;
113 : use crate::tenant::remote_timeline_client::INITDB_PATH;
114 : use crate::tenant::storage_layer::DeltaLayer;
115 : use crate::tenant::storage_layer::ImageLayer;
116 : use crate::walingest::WalLagCooldown;
117 : use crate::walredo;
118 : use crate::InitializationOrder;
119 : use std::collections::hash_map::Entry;
120 : use std::collections::HashMap;
121 : use std::collections::HashSet;
122 : use std::fmt::Debug;
123 : use std::fmt::Display;
124 : use std::fs;
125 : use std::fs::File;
126 : use std::sync::atomic::{AtomicU64, Ordering};
127 : use std::sync::Arc;
128 : use std::sync::Mutex;
129 : use std::time::{Duration, Instant};
130 :
131 : use crate::span;
132 : use crate::tenant::timeline::delete::DeleteTimelineFlow;
133 : use crate::tenant::timeline::uninit::cleanup_timeline_directory;
134 : use crate::virtual_file::VirtualFile;
135 : use crate::walredo::PostgresRedoManager;
136 : use crate::TEMP_FILE_SUFFIX;
137 : use once_cell::sync::Lazy;
138 : pub use pageserver_api::models::TenantState;
139 : use tokio::sync::Semaphore;
140 :
141 0 : static INIT_DB_SEMAPHORE: Lazy<Semaphore> = Lazy::new(|| Semaphore::new(8));
142 : use utils::{
143 : crashsafe,
144 : generation::Generation,
145 : id::TimelineId,
146 : lsn::{Lsn, RecordLsn},
147 : };
148 :
149 : pub mod blob_io;
150 : pub mod block_io;
151 : pub mod vectored_blob_io;
152 :
153 : pub mod disk_btree;
154 : pub(crate) mod ephemeral_file;
155 : pub mod layer_map;
156 :
157 : pub mod metadata;
158 : pub mod remote_timeline_client;
159 : pub mod storage_layer;
160 :
161 : pub mod checks;
162 : pub mod config;
163 : pub mod mgr;
164 : pub mod secondary;
165 : pub mod tasks;
166 : pub mod upload_queue;
167 :
168 : pub(crate) mod timeline;
169 :
170 : pub mod size;
171 :
172 : mod gc_block;
173 : mod gc_result;
174 : pub(crate) mod throttle;
175 :
176 : pub(crate) use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
177 : pub(crate) use timeline::{LogicalSizeCalculationCause, PageReconstructError, Timeline};
178 :
179 : // re-export for use in walreceiver
180 : pub use crate::tenant::timeline::WalReceiverInfo;
181 :
182 : /// The "tenants" part of `tenants/<tenant>/timelines...`
183 : pub const TENANTS_SEGMENT_NAME: &str = "tenants";
184 :
185 : /// Parts of the `.neon/tenants/<tenant_id>/timelines/<timeline_id>` directory prefix.
186 : pub const TIMELINES_SEGMENT_NAME: &str = "timelines";
187 :
188 : /// References to shared objects that are passed into each tenant, such
189 : /// as the shared remote storage client and process initialization state.
190 : #[derive(Clone)]
191 : pub struct TenantSharedResources {
192 : pub broker_client: storage_broker::BrokerClientChannel,
193 : pub remote_storage: GenericRemoteStorage,
194 : pub deletion_queue_client: DeletionQueueClient,
195 : pub l0_flush_global_state: L0FlushGlobalState,
196 : }
197 :
198 : /// A [`Tenant`] is really an _attached_ tenant. The configuration
199 : /// for an attached tenant is a subset of the [`LocationConf`], represented
200 : /// in this struct.
201 : #[derive(Clone)]
202 : pub(super) struct AttachedTenantConf {
203 : tenant_conf: TenantConfOpt,
204 : location: AttachedLocationConfig,
205 : /// The deadline before which we are blocked from GC so that
206 : /// leases have a chance to be renewed.
207 : lsn_lease_deadline: Option<tokio::time::Instant>,
208 : }
209 :
210 : impl AttachedTenantConf {
211 196 : fn new(tenant_conf: TenantConfOpt, location: AttachedLocationConfig) -> Self {
212 : // Sets a deadline before which we cannot proceed to GC due to lsn lease.
213 : //
214 : // We do this as the leases mapping are not persisted to disk. By delaying GC by lease
215 : // length, we guarantee that all the leases we granted before will have a chance to renew
216 : // when we run GC for the first time after restart / transition from AttachedMulti to AttachedSingle.
217 196 : let lsn_lease_deadline = if location.attach_mode == AttachmentMode::Single {
218 196 : Some(
219 196 : tokio::time::Instant::now()
220 196 : + tenant_conf
221 196 : .lsn_lease_length
222 196 : .unwrap_or(LsnLease::DEFAULT_LENGTH),
223 196 : )
224 : } else {
225 : // We don't use `lsn_lease_deadline` to delay GC in AttachedMulti and AttachedStale
226 : // because we don't do GC in these modes.
227 0 : None
228 : };
229 :
230 196 : Self {
231 196 : tenant_conf,
232 196 : location,
233 196 : lsn_lease_deadline,
234 196 : }
235 196 : }
236 :
237 196 : fn try_from(location_conf: LocationConf) -> anyhow::Result<Self> {
238 196 : match &location_conf.mode {
239 196 : LocationMode::Attached(attach_conf) => {
240 196 : Ok(Self::new(location_conf.tenant_conf, *attach_conf))
241 : }
242 : LocationMode::Secondary(_) => {
243 0 : anyhow::bail!("Attempted to construct AttachedTenantConf from a LocationConf in secondary mode")
244 : }
245 : }
246 196 : }
247 :
248 762 : fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
249 762 : self.lsn_lease_deadline
250 762 : .map(|d| tokio::time::Instant::now() < d)
251 762 : .unwrap_or(false)
252 762 : }
253 : }
254 : struct TimelinePreload {
255 : timeline_id: TimelineId,
256 : client: RemoteTimelineClient,
257 : index_part: Result<MaybeDeletedIndexPart, DownloadError>,
258 : }
259 :
260 : pub(crate) struct TenantPreload {
261 : tenant_manifest: TenantManifest,
262 : /// Map from timeline ID to a possible timeline preload. It is None iff the timeline is offloaded according to the manifest.
263 : timelines: HashMap<TimelineId, Option<TimelinePreload>>,
264 : }
265 :
266 : /// When we spawn a tenant, there is a special mode for tenant creation that
267 : /// avoids trying to read anything from remote storage.
268 : pub(crate) enum SpawnMode {
269 : /// Activate as soon as possible
270 : Eager,
271 : /// Lazy activation in the background, with the option to skip the queue if the need comes up
272 : Lazy,
273 : }
274 :
275 : ///
276 : /// Tenant consists of multiple timelines. Keep them in a hash table.
277 : ///
278 : pub struct Tenant {
279 : // Global pageserver config parameters
280 : pub conf: &'static PageServerConf,
281 :
282 : /// The value creation timestamp, used to measure activation delay, see:
283 : /// <https://github.com/neondatabase/neon/issues/4025>
284 : constructed_at: Instant,
285 :
286 : state: watch::Sender<TenantState>,
287 :
288 : // Overridden tenant-specific config parameters.
289 : // We keep TenantConfOpt sturct here to preserve the information
290 : // about parameters that are not set.
291 : // This is necessary to allow global config updates.
292 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
293 :
294 : tenant_shard_id: TenantShardId,
295 :
296 : // The detailed sharding information, beyond the number/count in tenant_shard_id
297 : shard_identity: ShardIdentity,
298 :
299 : /// The remote storage generation, used to protect S3 objects from split-brain.
300 : /// Does not change over the lifetime of the [`Tenant`] object.
301 : ///
302 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
303 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
304 : generation: Generation,
305 :
306 : timelines: Mutex<HashMap<TimelineId, Arc<Timeline>>>,
307 :
308 : /// During timeline creation, we first insert the TimelineId to the
309 : /// creating map, then `timelines`, then remove it from the creating map.
310 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
311 : timelines_creating: std::sync::Mutex<HashSet<TimelineId>>,
312 :
313 : /// Possibly offloaded and archived timelines
314 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
315 : timelines_offloaded: Mutex<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
316 :
317 : /// Serialize writes of the tenant manifest to remote storage. If there are concurrent operations
318 : /// affecting the manifest, such as timeline deletion and timeline offload, they must wait for
319 : /// each other (this could be optimized to coalesce writes if necessary).
320 : ///
321 : /// The contents of the Mutex are the last manifest we successfully uploaded
322 : tenant_manifest_upload: tokio::sync::Mutex<Option<TenantManifest>>,
323 :
324 : // This mutex prevents creation of new timelines during GC.
325 : // Adding yet another mutex (in addition to `timelines`) is needed because holding
326 : // `timelines` mutex during all GC iteration
327 : // may block for a long time `get_timeline`, `get_timelines_state`,... and other operations
328 : // with timelines, which in turn may cause dropping replication connection, expiration of wait_for_lsn
329 : // timeout...
330 : gc_cs: tokio::sync::Mutex<()>,
331 : walredo_mgr: Option<Arc<WalRedoManager>>,
332 :
333 : // provides access to timeline data sitting in the remote storage
334 : pub(crate) remote_storage: GenericRemoteStorage,
335 :
336 : // Access to global deletion queue for when this tenant wants to schedule a deletion
337 : deletion_queue_client: DeletionQueueClient,
338 :
339 : /// Cached logical sizes updated updated on each [`Tenant::gather_size_inputs`].
340 : cached_logical_sizes: tokio::sync::Mutex<HashMap<(TimelineId, Lsn), u64>>,
341 : cached_synthetic_tenant_size: Arc<AtomicU64>,
342 :
343 : eviction_task_tenant_state: tokio::sync::Mutex<EvictionTaskTenantState>,
344 :
345 : /// Track repeated failures to compact, so that we can back off.
346 : /// Overhead of mutex is acceptable because compaction is done with a multi-second period.
347 : compaction_circuit_breaker: std::sync::Mutex<CircuitBreaker>,
348 :
349 : /// Scheduled compaction tasks. Currently, this can only be populated by triggering
350 : /// a manual gc-compaction from the manual compaction API.
351 : scheduled_compaction_tasks:
352 : std::sync::Mutex<HashMap<TimelineId, VecDeque<ScheduledCompactionTask>>>,
353 :
354 : /// If the tenant is in Activating state, notify this to encourage it
355 : /// to proceed to Active as soon as possible, rather than waiting for lazy
356 : /// background warmup.
357 : pub(crate) activate_now_sem: tokio::sync::Semaphore,
358 :
359 : /// Time it took for the tenant to activate. Zero if not active yet.
360 : attach_wal_lag_cooldown: Arc<std::sync::OnceLock<WalLagCooldown>>,
361 :
362 : // Cancellation token fires when we have entered shutdown(). This is a parent of
363 : // Timelines' cancellation token.
364 : pub(crate) cancel: CancellationToken,
365 :
366 : // Users of the Tenant such as the page service must take this Gate to avoid
367 : // trying to use a Tenant which is shutting down.
368 : pub(crate) gate: Gate,
369 :
370 : /// Throttle applied at the top of [`Timeline::get`].
371 : /// All [`Tenant::timelines`] of a given [`Tenant`] instance share the same [`throttle::Throttle`] instance.
372 : pub(crate) pagestream_throttle:
373 : Arc<throttle::Throttle<crate::metrics::tenant_throttling::Pagestream>>,
374 :
375 : /// An ongoing timeline detach concurrency limiter.
376 : ///
377 : /// As a tenant will likely be restarted as part of timeline detach ancestor it makes no sense
378 : /// to have two running at the same time. A different one can be started if an earlier one
379 : /// has failed for whatever reason.
380 : ongoing_timeline_detach: std::sync::Mutex<Option<(TimelineId, utils::completion::Barrier)>>,
381 :
382 : /// `index_part.json` based gc blocking reason tracking.
383 : ///
384 : /// New gc iterations must start a new iteration by acquiring `GcBlock::start` before
385 : /// proceeding.
386 : pub(crate) gc_block: gc_block::GcBlock,
387 :
388 : l0_flush_global_state: L0FlushGlobalState,
389 : }
390 : impl std::fmt::Debug for Tenant {
391 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
392 0 : write!(f, "{} ({})", self.tenant_shard_id, self.current_state())
393 0 : }
394 : }
395 :
396 : pub(crate) enum WalRedoManager {
397 : Prod(WalredoManagerId, PostgresRedoManager),
398 : #[cfg(test)]
399 : Test(harness::TestRedoManager),
400 : }
401 :
402 : #[derive(thiserror::Error, Debug)]
403 : #[error("pageserver is shutting down")]
404 : pub(crate) struct GlobalShutDown;
405 :
406 : impl WalRedoManager {
407 0 : pub(crate) fn new(mgr: PostgresRedoManager) -> Result<Arc<Self>, GlobalShutDown> {
408 0 : let id = WalredoManagerId::next();
409 0 : let arc = Arc::new(Self::Prod(id, mgr));
410 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
411 0 : match &mut *guard {
412 0 : Some(map) => {
413 0 : map.insert(id, Arc::downgrade(&arc));
414 0 : Ok(arc)
415 : }
416 0 : None => Err(GlobalShutDown),
417 : }
418 0 : }
419 : }
420 :
421 : impl Drop for WalRedoManager {
422 10 : fn drop(&mut self) {
423 10 : match self {
424 0 : Self::Prod(id, _) => {
425 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
426 0 : if let Some(map) = &mut *guard {
427 0 : map.remove(id).expect("new() registers, drop() unregisters");
428 0 : }
429 : }
430 : #[cfg(test)]
431 10 : Self::Test(_) => {
432 10 : // Not applicable to test redo manager
433 10 : }
434 : }
435 10 : }
436 : }
437 :
438 : /// Global registry of all walredo managers so that [`crate::shutdown_pageserver`] can shut down
439 : /// the walredo processes outside of the regular order.
440 : ///
441 : /// This is necessary to work around a systemd bug where it freezes if there are
442 : /// walredo processes left => <https://github.com/neondatabase/cloud/issues/11387>
443 : #[allow(clippy::type_complexity)]
444 : pub(crate) static WALREDO_MANAGERS: once_cell::sync::Lazy<
445 : Mutex<Option<HashMap<WalredoManagerId, Weak<WalRedoManager>>>>,
446 0 : > = once_cell::sync::Lazy::new(|| Mutex::new(Some(HashMap::new())));
447 : #[derive(PartialEq, Eq, Hash, Clone, Copy, Debug)]
448 : pub(crate) struct WalredoManagerId(u64);
449 : impl WalredoManagerId {
450 0 : pub fn next() -> Self {
451 : static NEXT: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(1);
452 0 : let id = NEXT.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
453 0 : if id == 0 {
454 0 : panic!("WalredoManagerId::new() returned 0, indicating wraparound, risking it's no longer unique");
455 0 : }
456 0 : Self(id)
457 0 : }
458 : }
459 :
460 : #[cfg(test)]
461 : impl From<harness::TestRedoManager> for WalRedoManager {
462 196 : fn from(mgr: harness::TestRedoManager) -> Self {
463 196 : Self::Test(mgr)
464 196 : }
465 : }
466 :
467 : impl WalRedoManager {
468 6 : pub(crate) async fn shutdown(&self) -> bool {
469 6 : match self {
470 0 : Self::Prod(_, mgr) => mgr.shutdown().await,
471 : #[cfg(test)]
472 : Self::Test(_) => {
473 : // Not applicable to test redo manager
474 6 : true
475 : }
476 : }
477 6 : }
478 :
479 0 : pub(crate) fn maybe_quiesce(&self, idle_timeout: Duration) {
480 0 : match self {
481 0 : Self::Prod(_, mgr) => mgr.maybe_quiesce(idle_timeout),
482 0 : #[cfg(test)]
483 0 : Self::Test(_) => {
484 0 : // Not applicable to test redo manager
485 0 : }
486 0 : }
487 0 : }
488 :
489 : /// # Cancel-Safety
490 : ///
491 : /// This method is cancellation-safe.
492 520 : pub async fn request_redo(
493 520 : &self,
494 520 : key: pageserver_api::key::Key,
495 520 : lsn: Lsn,
496 520 : base_img: Option<(Lsn, bytes::Bytes)>,
497 520 : records: Vec<(Lsn, pageserver_api::record::NeonWalRecord)>,
498 520 : pg_version: u32,
499 520 : ) -> Result<bytes::Bytes, walredo::Error> {
500 520 : match self {
501 0 : Self::Prod(_, mgr) => {
502 0 : mgr.request_redo(key, lsn, base_img, records, pg_version)
503 0 : .await
504 : }
505 : #[cfg(test)]
506 520 : Self::Test(mgr) => {
507 520 : mgr.request_redo(key, lsn, base_img, records, pg_version)
508 520 : .await
509 : }
510 : }
511 520 : }
512 :
513 0 : pub(crate) fn status(&self) -> Option<WalRedoManagerStatus> {
514 0 : match self {
515 0 : WalRedoManager::Prod(_, m) => Some(m.status()),
516 0 : #[cfg(test)]
517 0 : WalRedoManager::Test(_) => None,
518 0 : }
519 0 : }
520 : }
521 :
522 : /// A very lightweight memory representation of an offloaded timeline.
523 : ///
524 : /// We need to store the list of offloaded timelines so that we can perform operations on them,
525 : /// like unoffloading them, or (at a later date), decide to perform flattening.
526 : /// This type has a much smaller memory impact than [`Timeline`], and thus we can store many
527 : /// more offloaded timelines than we can manage ones that aren't.
528 : pub struct OffloadedTimeline {
529 : pub tenant_shard_id: TenantShardId,
530 : pub timeline_id: TimelineId,
531 : pub ancestor_timeline_id: Option<TimelineId>,
532 : /// Whether to retain the branch lsn at the ancestor or not
533 : pub ancestor_retain_lsn: Option<Lsn>,
534 :
535 : /// When the timeline was archived.
536 : ///
537 : /// Present for future flattening deliberations.
538 : pub archived_at: NaiveDateTime,
539 :
540 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
541 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
542 : pub delete_progress: TimelineDeleteProgress,
543 :
544 : /// Part of the `OffloadedTimeline` object's lifecycle: this needs to be set before we drop it
545 : pub deleted_from_ancestor: AtomicBool,
546 : }
547 :
548 : impl OffloadedTimeline {
549 : /// Obtains an offloaded timeline from a given timeline object.
550 : ///
551 : /// Returns `None` if the `archived_at` flag couldn't be obtained, i.e.
552 : /// the timeline is not in a stopped state.
553 : /// Panics if the timeline is not archived.
554 2 : fn from_timeline(timeline: &Timeline) -> Result<Self, UploadQueueNotReadyError> {
555 2 : let (ancestor_retain_lsn, ancestor_timeline_id) =
556 2 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
557 2 : let ancestor_lsn = timeline.get_ancestor_lsn();
558 2 : let ancestor_timeline_id = ancestor_timeline.timeline_id;
559 2 : let mut gc_info = ancestor_timeline.gc_info.write().unwrap();
560 2 : gc_info.insert_child(timeline.timeline_id, ancestor_lsn, MaybeOffloaded::Yes);
561 2 : (Some(ancestor_lsn), Some(ancestor_timeline_id))
562 : } else {
563 0 : (None, None)
564 : };
565 2 : let archived_at = timeline
566 2 : .remote_client
567 2 : .archived_at_stopped_queue()?
568 2 : .expect("must be called on an archived timeline");
569 2 : Ok(Self {
570 2 : tenant_shard_id: timeline.tenant_shard_id,
571 2 : timeline_id: timeline.timeline_id,
572 2 : ancestor_timeline_id,
573 2 : ancestor_retain_lsn,
574 2 : archived_at,
575 2 :
576 2 : delete_progress: timeline.delete_progress.clone(),
577 2 : deleted_from_ancestor: AtomicBool::new(false),
578 2 : })
579 2 : }
580 0 : fn from_manifest(tenant_shard_id: TenantShardId, manifest: &OffloadedTimelineManifest) -> Self {
581 0 : // We expect to reach this case in tenant loading, where the `retain_lsn` is populated in the parent's `gc_info`
582 0 : // by the `initialize_gc_info` function.
583 0 : let OffloadedTimelineManifest {
584 0 : timeline_id,
585 0 : ancestor_timeline_id,
586 0 : ancestor_retain_lsn,
587 0 : archived_at,
588 0 : } = *manifest;
589 0 : Self {
590 0 : tenant_shard_id,
591 0 : timeline_id,
592 0 : ancestor_timeline_id,
593 0 : ancestor_retain_lsn,
594 0 : archived_at,
595 0 : delete_progress: TimelineDeleteProgress::default(),
596 0 : deleted_from_ancestor: AtomicBool::new(false),
597 0 : }
598 0 : }
599 2 : fn manifest(&self) -> OffloadedTimelineManifest {
600 2 : let Self {
601 2 : timeline_id,
602 2 : ancestor_timeline_id,
603 2 : ancestor_retain_lsn,
604 2 : archived_at,
605 2 : ..
606 2 : } = self;
607 2 : OffloadedTimelineManifest {
608 2 : timeline_id: *timeline_id,
609 2 : ancestor_timeline_id: *ancestor_timeline_id,
610 2 : ancestor_retain_lsn: *ancestor_retain_lsn,
611 2 : archived_at: *archived_at,
612 2 : }
613 2 : }
614 : /// Delete this timeline's retain_lsn from its ancestor, if present in the given tenant
615 0 : fn delete_from_ancestor_with_timelines(
616 0 : &self,
617 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
618 0 : ) {
619 0 : if let (Some(_retain_lsn), Some(ancestor_timeline_id)) =
620 0 : (self.ancestor_retain_lsn, self.ancestor_timeline_id)
621 : {
622 0 : if let Some((_, ancestor_timeline)) = timelines
623 0 : .iter()
624 0 : .find(|(tid, _tl)| **tid == ancestor_timeline_id)
625 : {
626 0 : let removal_happened = ancestor_timeline
627 0 : .gc_info
628 0 : .write()
629 0 : .unwrap()
630 0 : .remove_child_offloaded(self.timeline_id);
631 0 : if !removal_happened {
632 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
633 0 : "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
634 0 : }
635 0 : }
636 0 : }
637 0 : self.deleted_from_ancestor.store(true, Ordering::Release);
638 0 : }
639 : /// Call [`Self::delete_from_ancestor_with_timelines`] instead if possible.
640 : ///
641 : /// As the entire tenant is being dropped, don't bother deregistering the `retain_lsn` from the ancestor.
642 2 : fn defuse_for_tenant_drop(&self) {
643 2 : self.deleted_from_ancestor.store(true, Ordering::Release);
644 2 : }
645 : }
646 :
647 : impl fmt::Debug for OffloadedTimeline {
648 0 : fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
649 0 : write!(f, "OffloadedTimeline<{}>", self.timeline_id)
650 0 : }
651 : }
652 :
653 : impl Drop for OffloadedTimeline {
654 2 : fn drop(&mut self) {
655 2 : if !self.deleted_from_ancestor.load(Ordering::Acquire) {
656 0 : tracing::warn!(
657 0 : "offloaded timeline {} was dropped without having cleaned it up at the ancestor",
658 : self.timeline_id
659 : );
660 2 : }
661 2 : }
662 : }
663 :
664 : #[derive(Copy, Clone, PartialEq, Eq, Hash, Debug)]
665 : pub enum MaybeOffloaded {
666 : Yes,
667 : No,
668 : }
669 :
670 : #[derive(Clone, Debug)]
671 : pub enum TimelineOrOffloaded {
672 : Timeline(Arc<Timeline>),
673 : Offloaded(Arc<OffloadedTimeline>),
674 : }
675 :
676 : impl TimelineOrOffloaded {
677 0 : pub fn arc_ref(&self) -> TimelineOrOffloadedArcRef<'_> {
678 0 : match self {
679 0 : TimelineOrOffloaded::Timeline(timeline) => {
680 0 : TimelineOrOffloadedArcRef::Timeline(timeline)
681 : }
682 0 : TimelineOrOffloaded::Offloaded(offloaded) => {
683 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded)
684 : }
685 : }
686 0 : }
687 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
688 0 : self.arc_ref().tenant_shard_id()
689 0 : }
690 0 : pub fn timeline_id(&self) -> TimelineId {
691 0 : self.arc_ref().timeline_id()
692 0 : }
693 2 : pub fn delete_progress(&self) -> &Arc<tokio::sync::Mutex<DeleteTimelineFlow>> {
694 2 : match self {
695 2 : TimelineOrOffloaded::Timeline(timeline) => &timeline.delete_progress,
696 0 : TimelineOrOffloaded::Offloaded(offloaded) => &offloaded.delete_progress,
697 : }
698 2 : }
699 0 : fn maybe_remote_client(&self) -> Option<Arc<RemoteTimelineClient>> {
700 0 : match self {
701 0 : TimelineOrOffloaded::Timeline(timeline) => Some(timeline.remote_client.clone()),
702 0 : TimelineOrOffloaded::Offloaded(_offloaded) => None,
703 : }
704 0 : }
705 : }
706 :
707 : pub enum TimelineOrOffloadedArcRef<'a> {
708 : Timeline(&'a Arc<Timeline>),
709 : Offloaded(&'a Arc<OffloadedTimeline>),
710 : }
711 :
712 : impl TimelineOrOffloadedArcRef<'_> {
713 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
714 0 : match self {
715 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.tenant_shard_id,
716 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.tenant_shard_id,
717 : }
718 0 : }
719 0 : pub fn timeline_id(&self) -> TimelineId {
720 0 : match self {
721 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.timeline_id,
722 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.timeline_id,
723 : }
724 0 : }
725 : }
726 :
727 : impl<'a> From<&'a Arc<Timeline>> for TimelineOrOffloadedArcRef<'a> {
728 0 : fn from(timeline: &'a Arc<Timeline>) -> Self {
729 0 : Self::Timeline(timeline)
730 0 : }
731 : }
732 :
733 : impl<'a> From<&'a Arc<OffloadedTimeline>> for TimelineOrOffloadedArcRef<'a> {
734 0 : fn from(timeline: &'a Arc<OffloadedTimeline>) -> Self {
735 0 : Self::Offloaded(timeline)
736 0 : }
737 : }
738 :
739 : #[derive(Debug, thiserror::Error, PartialEq, Eq)]
740 : pub enum GetTimelineError {
741 : #[error("Timeline is shutting down")]
742 : ShuttingDown,
743 : #[error("Timeline {tenant_id}/{timeline_id} is not active, state: {state:?}")]
744 : NotActive {
745 : tenant_id: TenantShardId,
746 : timeline_id: TimelineId,
747 : state: TimelineState,
748 : },
749 : #[error("Timeline {tenant_id}/{timeline_id} was not found")]
750 : NotFound {
751 : tenant_id: TenantShardId,
752 : timeline_id: TimelineId,
753 : },
754 : }
755 :
756 : #[derive(Debug, thiserror::Error)]
757 : pub enum LoadLocalTimelineError {
758 : #[error("FailedToLoad")]
759 : Load(#[source] anyhow::Error),
760 : #[error("FailedToResumeDeletion")]
761 : ResumeDeletion(#[source] anyhow::Error),
762 : }
763 :
764 : #[derive(thiserror::Error)]
765 : pub enum DeleteTimelineError {
766 : #[error("NotFound")]
767 : NotFound,
768 :
769 : #[error("HasChildren")]
770 : HasChildren(Vec<TimelineId>),
771 :
772 : #[error("Timeline deletion is already in progress")]
773 : AlreadyInProgress(Arc<tokio::sync::Mutex<DeleteTimelineFlow>>),
774 :
775 : #[error("Cancelled")]
776 : Cancelled,
777 :
778 : #[error(transparent)]
779 : Other(#[from] anyhow::Error),
780 : }
781 :
782 : impl Debug for DeleteTimelineError {
783 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
784 0 : match self {
785 0 : Self::NotFound => write!(f, "NotFound"),
786 0 : Self::HasChildren(c) => f.debug_tuple("HasChildren").field(c).finish(),
787 0 : Self::AlreadyInProgress(_) => f.debug_tuple("AlreadyInProgress").finish(),
788 0 : Self::Cancelled => f.debug_tuple("Cancelled").finish(),
789 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
790 : }
791 0 : }
792 : }
793 :
794 : #[derive(thiserror::Error)]
795 : pub enum TimelineArchivalError {
796 : #[error("NotFound")]
797 : NotFound,
798 :
799 : #[error("Timeout")]
800 : Timeout,
801 :
802 : #[error("Cancelled")]
803 : Cancelled,
804 :
805 : #[error("ancestor is archived: {}", .0)]
806 : HasArchivedParent(TimelineId),
807 :
808 : #[error("HasUnarchivedChildren")]
809 : HasUnarchivedChildren(Vec<TimelineId>),
810 :
811 : #[error("Timeline archival is already in progress")]
812 : AlreadyInProgress,
813 :
814 : #[error(transparent)]
815 : Other(anyhow::Error),
816 : }
817 :
818 : #[derive(thiserror::Error, Debug)]
819 : pub(crate) enum TenantManifestError {
820 : #[error("Remote storage error: {0}")]
821 : RemoteStorage(anyhow::Error),
822 :
823 : #[error("Cancelled")]
824 : Cancelled,
825 : }
826 :
827 : impl From<TenantManifestError> for TimelineArchivalError {
828 0 : fn from(e: TenantManifestError) -> Self {
829 0 : match e {
830 0 : TenantManifestError::RemoteStorage(e) => Self::Other(e),
831 0 : TenantManifestError::Cancelled => Self::Cancelled,
832 : }
833 0 : }
834 : }
835 :
836 : impl Debug for TimelineArchivalError {
837 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
838 0 : match self {
839 0 : Self::NotFound => write!(f, "NotFound"),
840 0 : Self::Timeout => write!(f, "Timeout"),
841 0 : Self::Cancelled => write!(f, "Cancelled"),
842 0 : Self::HasArchivedParent(p) => f.debug_tuple("HasArchivedParent").field(p).finish(),
843 0 : Self::HasUnarchivedChildren(c) => {
844 0 : f.debug_tuple("HasUnarchivedChildren").field(c).finish()
845 : }
846 0 : Self::AlreadyInProgress => f.debug_tuple("AlreadyInProgress").finish(),
847 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
848 : }
849 0 : }
850 : }
851 :
852 : pub enum SetStoppingError {
853 : AlreadyStopping(completion::Barrier),
854 : Broken,
855 : }
856 :
857 : impl Debug for SetStoppingError {
858 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
859 0 : match self {
860 0 : Self::AlreadyStopping(_) => f.debug_tuple("AlreadyStopping").finish(),
861 0 : Self::Broken => write!(f, "Broken"),
862 : }
863 0 : }
864 : }
865 :
866 : /// Arguments to [`Tenant::create_timeline`].
867 : ///
868 : /// Not usable as an idempotency key for timeline creation because if [`CreateTimelineParamsBranch::ancestor_start_lsn`]
869 : /// is `None`, the result of the timeline create call is not deterministic.
870 : ///
871 : /// See [`CreateTimelineIdempotency`] for an idempotency key.
872 : #[derive(Debug)]
873 : pub(crate) enum CreateTimelineParams {
874 : Bootstrap(CreateTimelineParamsBootstrap),
875 : Branch(CreateTimelineParamsBranch),
876 : ImportPgdata(CreateTimelineParamsImportPgdata),
877 : }
878 :
879 : #[derive(Debug)]
880 : pub(crate) struct CreateTimelineParamsBootstrap {
881 : pub(crate) new_timeline_id: TimelineId,
882 : pub(crate) existing_initdb_timeline_id: Option<TimelineId>,
883 : pub(crate) pg_version: u32,
884 : }
885 :
886 : /// NB: See comment on [`CreateTimelineIdempotency::Branch`] for why there's no `pg_version` here.
887 : #[derive(Debug)]
888 : pub(crate) struct CreateTimelineParamsBranch {
889 : pub(crate) new_timeline_id: TimelineId,
890 : pub(crate) ancestor_timeline_id: TimelineId,
891 : pub(crate) ancestor_start_lsn: Option<Lsn>,
892 : }
893 :
894 : #[derive(Debug)]
895 : pub(crate) struct CreateTimelineParamsImportPgdata {
896 : pub(crate) new_timeline_id: TimelineId,
897 : pub(crate) location: import_pgdata::index_part_format::Location,
898 : pub(crate) idempotency_key: import_pgdata::index_part_format::IdempotencyKey,
899 : }
900 :
901 : /// What is used to determine idempotency of a [`Tenant::create_timeline`] call in [`Tenant::start_creating_timeline`] in [`Tenant::start_creating_timeline`].
902 : ///
903 : /// Each [`Timeline`] object holds [`Self`] as an immutable property in [`Timeline::create_idempotency`].
904 : ///
905 : /// We lower timeline creation requests to [`Self`], and then use [`PartialEq::eq`] to compare [`Timeline::create_idempotency`] with the request.
906 : /// If they are equal, we return a reference to the existing timeline, otherwise it's an idempotency conflict.
907 : ///
908 : /// There is special treatment for [`Self::FailWithConflict`] to always return an idempotency conflict.
909 : /// It would be nice to have more advanced derive macros to make that special treatment declarative.
910 : ///
911 : /// Notes:
912 : /// - Unlike [`CreateTimelineParams`], ancestor LSN is fixed, so, branching will be at a deterministic LSN.
913 : /// - We make some trade-offs though, e.g., [`CreateTimelineParamsBootstrap::existing_initdb_timeline_id`]
914 : /// is not considered for idempotency. We can improve on this over time if we deem it necessary.
915 : ///
916 : #[derive(Debug, Clone, PartialEq, Eq)]
917 : pub(crate) enum CreateTimelineIdempotency {
918 : /// NB: special treatment, see comment in [`Self`].
919 : FailWithConflict,
920 : Bootstrap {
921 : pg_version: u32,
922 : },
923 : /// NB: branches always have the same `pg_version` as their ancestor.
924 : /// While [`pageserver_api::models::TimelineCreateRequestMode::Branch::pg_version`]
925 : /// exists as a field, and is set by cplane, it has always been ignored by pageserver when
926 : /// determining the child branch pg_version.
927 : Branch {
928 : ancestor_timeline_id: TimelineId,
929 : ancestor_start_lsn: Lsn,
930 : },
931 : ImportPgdata(CreatingTimelineIdempotencyImportPgdata),
932 : }
933 :
934 : #[derive(Debug, Clone, PartialEq, Eq)]
935 : pub(crate) struct CreatingTimelineIdempotencyImportPgdata {
936 : idempotency_key: import_pgdata::index_part_format::IdempotencyKey,
937 : }
938 :
939 : /// What is returned by [`Tenant::start_creating_timeline`].
940 : #[must_use]
941 : enum StartCreatingTimelineResult {
942 : CreateGuard(TimelineCreateGuard),
943 : Idempotent(Arc<Timeline>),
944 : }
945 :
946 : enum TimelineInitAndSyncResult {
947 : ReadyToActivate(Arc<Timeline>),
948 : NeedsSpawnImportPgdata(TimelineInitAndSyncNeedsSpawnImportPgdata),
949 : }
950 :
951 : impl TimelineInitAndSyncResult {
952 0 : fn ready_to_activate(self) -> Option<Arc<Timeline>> {
953 0 : match self {
954 0 : Self::ReadyToActivate(timeline) => Some(timeline),
955 0 : _ => None,
956 : }
957 0 : }
958 : }
959 :
960 : #[must_use]
961 : struct TimelineInitAndSyncNeedsSpawnImportPgdata {
962 : timeline: Arc<Timeline>,
963 : import_pgdata: import_pgdata::index_part_format::Root,
964 : guard: TimelineCreateGuard,
965 : }
966 :
967 : /// What is returned by [`Tenant::create_timeline`].
968 : enum CreateTimelineResult {
969 : Created(Arc<Timeline>),
970 : Idempotent(Arc<Timeline>),
971 : /// IMPORTANT: This [`Arc<Timeline>`] object is not in [`Tenant::timelines`] when
972 : /// we return this result, nor will this concrete object ever be added there.
973 : /// Cf method comment on [`Tenant::create_timeline_import_pgdata`].
974 : ImportSpawned(Arc<Timeline>),
975 : }
976 :
977 : impl CreateTimelineResult {
978 0 : fn discriminant(&self) -> &'static str {
979 0 : match self {
980 0 : Self::Created(_) => "Created",
981 0 : Self::Idempotent(_) => "Idempotent",
982 0 : Self::ImportSpawned(_) => "ImportSpawned",
983 : }
984 0 : }
985 0 : fn timeline(&self) -> &Arc<Timeline> {
986 0 : match self {
987 0 : Self::Created(t) | Self::Idempotent(t) | Self::ImportSpawned(t) => t,
988 0 : }
989 0 : }
990 : /// Unit test timelines aren't activated, test has to do it if it needs to.
991 : #[cfg(test)]
992 230 : fn into_timeline_for_test(self) -> Arc<Timeline> {
993 230 : match self {
994 230 : Self::Created(t) | Self::Idempotent(t) | Self::ImportSpawned(t) => t,
995 230 : }
996 230 : }
997 : }
998 :
999 : #[derive(thiserror::Error, Debug)]
1000 : pub enum CreateTimelineError {
1001 : #[error("creation of timeline with the given ID is in progress")]
1002 : AlreadyCreating,
1003 : #[error("timeline already exists with different parameters")]
1004 : Conflict,
1005 : #[error(transparent)]
1006 : AncestorLsn(anyhow::Error),
1007 : #[error("ancestor timeline is not active")]
1008 : AncestorNotActive,
1009 : #[error("ancestor timeline is archived")]
1010 : AncestorArchived,
1011 : #[error("tenant shutting down")]
1012 : ShuttingDown,
1013 : #[error(transparent)]
1014 : Other(#[from] anyhow::Error),
1015 : }
1016 :
1017 : #[derive(thiserror::Error, Debug)]
1018 : pub enum InitdbError {
1019 : #[error("Operation was cancelled")]
1020 : Cancelled,
1021 : #[error(transparent)]
1022 : Other(anyhow::Error),
1023 : #[error(transparent)]
1024 : Inner(postgres_initdb::Error),
1025 : }
1026 :
1027 : enum CreateTimelineCause {
1028 : Load,
1029 : Delete,
1030 : }
1031 :
1032 : enum LoadTimelineCause {
1033 : Attach,
1034 : Unoffload,
1035 : ImportPgdata {
1036 : create_guard: TimelineCreateGuard,
1037 : activate: ActivateTimelineArgs,
1038 : },
1039 : }
1040 :
1041 : #[derive(thiserror::Error, Debug)]
1042 : pub(crate) enum GcError {
1043 : // The tenant is shutting down
1044 : #[error("tenant shutting down")]
1045 : TenantCancelled,
1046 :
1047 : // The tenant is shutting down
1048 : #[error("timeline shutting down")]
1049 : TimelineCancelled,
1050 :
1051 : // The tenant is in a state inelegible to run GC
1052 : #[error("not active")]
1053 : NotActive,
1054 :
1055 : // A requested GC cutoff LSN was invalid, for example it tried to move backwards
1056 : #[error("not active")]
1057 : BadLsn { why: String },
1058 :
1059 : // A remote storage error while scheduling updates after compaction
1060 : #[error(transparent)]
1061 : Remote(anyhow::Error),
1062 :
1063 : // An error reading while calculating GC cutoffs
1064 : #[error(transparent)]
1065 : GcCutoffs(PageReconstructError),
1066 :
1067 : // If GC was invoked for a particular timeline, this error means it didn't exist
1068 : #[error("timeline not found")]
1069 : TimelineNotFound,
1070 : }
1071 :
1072 : impl From<PageReconstructError> for GcError {
1073 0 : fn from(value: PageReconstructError) -> Self {
1074 0 : match value {
1075 0 : PageReconstructError::Cancelled => Self::TimelineCancelled,
1076 0 : other => Self::GcCutoffs(other),
1077 : }
1078 0 : }
1079 : }
1080 :
1081 : impl From<NotInitialized> for GcError {
1082 0 : fn from(value: NotInitialized) -> Self {
1083 0 : match value {
1084 0 : NotInitialized::Uninitialized => GcError::Remote(value.into()),
1085 0 : NotInitialized::Stopped | NotInitialized::ShuttingDown => GcError::TimelineCancelled,
1086 : }
1087 0 : }
1088 : }
1089 :
1090 : impl From<timeline::layer_manager::Shutdown> for GcError {
1091 0 : fn from(_: timeline::layer_manager::Shutdown) -> Self {
1092 0 : GcError::TimelineCancelled
1093 0 : }
1094 : }
1095 :
1096 : #[derive(thiserror::Error, Debug)]
1097 : pub(crate) enum LoadConfigError {
1098 : #[error("TOML deserialization error: '{0}'")]
1099 : DeserializeToml(#[from] toml_edit::de::Error),
1100 :
1101 : #[error("Config not found at {0}")]
1102 : NotFound(Utf8PathBuf),
1103 : }
1104 :
1105 : impl Tenant {
1106 : /// Yet another helper for timeline initialization.
1107 : ///
1108 : /// - Initializes the Timeline struct and inserts it into the tenant's hash map
1109 : /// - Scans the local timeline directory for layer files and builds the layer map
1110 : /// - Downloads remote index file and adds remote files to the layer map
1111 : /// - Schedules remote upload tasks for any files that are present locally but missing from remote storage.
1112 : ///
1113 : /// If the operation fails, the timeline is left in the tenant's hash map in Broken state. On success,
1114 : /// it is marked as Active.
1115 : #[allow(clippy::too_many_arguments)]
1116 6 : async fn timeline_init_and_sync(
1117 6 : self: &Arc<Self>,
1118 6 : timeline_id: TimelineId,
1119 6 : resources: TimelineResources,
1120 6 : mut index_part: IndexPart,
1121 6 : metadata: TimelineMetadata,
1122 6 : ancestor: Option<Arc<Timeline>>,
1123 6 : cause: LoadTimelineCause,
1124 6 : ctx: &RequestContext,
1125 6 : ) -> anyhow::Result<TimelineInitAndSyncResult> {
1126 6 : let tenant_id = self.tenant_shard_id;
1127 6 :
1128 6 : let import_pgdata = index_part.import_pgdata.take();
1129 6 : let idempotency = match &import_pgdata {
1130 0 : Some(import_pgdata) => {
1131 0 : CreateTimelineIdempotency::ImportPgdata(CreatingTimelineIdempotencyImportPgdata {
1132 0 : idempotency_key: import_pgdata.idempotency_key().clone(),
1133 0 : })
1134 : }
1135 : None => {
1136 6 : if metadata.ancestor_timeline().is_none() {
1137 4 : CreateTimelineIdempotency::Bootstrap {
1138 4 : pg_version: metadata.pg_version(),
1139 4 : }
1140 : } else {
1141 2 : CreateTimelineIdempotency::Branch {
1142 2 : ancestor_timeline_id: metadata.ancestor_timeline().unwrap(),
1143 2 : ancestor_start_lsn: metadata.ancestor_lsn(),
1144 2 : }
1145 : }
1146 : }
1147 : };
1148 :
1149 6 : let timeline = self.create_timeline_struct(
1150 6 : timeline_id,
1151 6 : &metadata,
1152 6 : ancestor.clone(),
1153 6 : resources,
1154 6 : CreateTimelineCause::Load,
1155 6 : idempotency.clone(),
1156 6 : )?;
1157 6 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
1158 6 : anyhow::ensure!(
1159 6 : disk_consistent_lsn.is_valid(),
1160 0 : "Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn"
1161 : );
1162 6 : assert_eq!(
1163 6 : disk_consistent_lsn,
1164 6 : metadata.disk_consistent_lsn(),
1165 0 : "these are used interchangeably"
1166 : );
1167 :
1168 6 : timeline.remote_client.init_upload_queue(&index_part)?;
1169 :
1170 6 : timeline
1171 6 : .load_layer_map(disk_consistent_lsn, index_part)
1172 6 : .await
1173 6 : .with_context(|| {
1174 0 : format!("Failed to load layermap for timeline {tenant_id}/{timeline_id}")
1175 6 : })?;
1176 :
1177 0 : match import_pgdata {
1178 0 : Some(import_pgdata) if !import_pgdata.is_done() => {
1179 0 : match cause {
1180 0 : LoadTimelineCause::Attach | LoadTimelineCause::Unoffload => (),
1181 : LoadTimelineCause::ImportPgdata { .. } => {
1182 0 : unreachable!("ImportPgdata should not be reloading timeline import is done and persisted as such in s3")
1183 : }
1184 : }
1185 0 : let mut guard = self.timelines_creating.lock().unwrap();
1186 0 : if !guard.insert(timeline_id) {
1187 : // We should never try and load the same timeline twice during startup
1188 0 : unreachable!("Timeline {tenant_id}/{timeline_id} is already being created")
1189 0 : }
1190 0 : let timeline_create_guard = TimelineCreateGuard {
1191 0 : _tenant_gate_guard: self.gate.enter()?,
1192 0 : owning_tenant: self.clone(),
1193 0 : timeline_id,
1194 0 : idempotency,
1195 0 : // The users of this specific return value don't need the timline_path in there.
1196 0 : timeline_path: timeline
1197 0 : .conf
1198 0 : .timeline_path(&timeline.tenant_shard_id, &timeline.timeline_id),
1199 0 : };
1200 0 : Ok(TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
1201 0 : TimelineInitAndSyncNeedsSpawnImportPgdata {
1202 0 : timeline,
1203 0 : import_pgdata,
1204 0 : guard: timeline_create_guard,
1205 0 : },
1206 0 : ))
1207 : }
1208 : Some(_) | None => {
1209 : {
1210 6 : let mut timelines_accessor = self.timelines.lock().unwrap();
1211 6 : match timelines_accessor.entry(timeline_id) {
1212 : // We should never try and load the same timeline twice during startup
1213 : Entry::Occupied(_) => {
1214 0 : unreachable!(
1215 0 : "Timeline {tenant_id}/{timeline_id} already exists in the tenant map"
1216 0 : );
1217 : }
1218 6 : Entry::Vacant(v) => {
1219 6 : v.insert(Arc::clone(&timeline));
1220 6 : timeline.maybe_spawn_flush_loop();
1221 6 : }
1222 : }
1223 : }
1224 :
1225 : // Sanity check: a timeline should have some content.
1226 6 : anyhow::ensure!(
1227 6 : ancestor.is_some()
1228 4 : || timeline
1229 4 : .layers
1230 4 : .read()
1231 4 : .await
1232 4 : .layer_map()
1233 4 : .expect("currently loading, layer manager cannot be shutdown already")
1234 4 : .iter_historic_layers()
1235 4 : .next()
1236 4 : .is_some(),
1237 0 : "Timeline has no ancestor and no layer files"
1238 : );
1239 :
1240 6 : match cause {
1241 6 : LoadTimelineCause::Attach | LoadTimelineCause::Unoffload => (),
1242 : LoadTimelineCause::ImportPgdata {
1243 0 : create_guard,
1244 0 : activate,
1245 0 : } => {
1246 0 : // TODO: see the comment in the task code above how I'm not so certain
1247 0 : // it is safe to activate here because of concurrent shutdowns.
1248 0 : match activate {
1249 0 : ActivateTimelineArgs::Yes { broker_client } => {
1250 0 : info!("activating timeline after reload from pgdata import task");
1251 0 : timeline.activate(self.clone(), broker_client, None, ctx);
1252 : }
1253 0 : ActivateTimelineArgs::No => (),
1254 : }
1255 0 : drop(create_guard);
1256 : }
1257 : }
1258 :
1259 6 : Ok(TimelineInitAndSyncResult::ReadyToActivate(timeline))
1260 : }
1261 : }
1262 6 : }
1263 :
1264 : /// Attach a tenant that's available in cloud storage.
1265 : ///
1266 : /// This returns quickly, after just creating the in-memory object
1267 : /// Tenant struct and launching a background task to download
1268 : /// the remote index files. On return, the tenant is most likely still in
1269 : /// Attaching state, and it will become Active once the background task
1270 : /// finishes. You can use wait_until_active() to wait for the task to
1271 : /// complete.
1272 : ///
1273 : #[allow(clippy::too_many_arguments)]
1274 0 : pub(crate) fn spawn(
1275 0 : conf: &'static PageServerConf,
1276 0 : tenant_shard_id: TenantShardId,
1277 0 : resources: TenantSharedResources,
1278 0 : attached_conf: AttachedTenantConf,
1279 0 : shard_identity: ShardIdentity,
1280 0 : init_order: Option<InitializationOrder>,
1281 0 : mode: SpawnMode,
1282 0 : ctx: &RequestContext,
1283 0 : ) -> Result<Arc<Tenant>, GlobalShutDown> {
1284 0 : let wal_redo_manager =
1285 0 : WalRedoManager::new(PostgresRedoManager::new(conf, tenant_shard_id))?;
1286 :
1287 : let TenantSharedResources {
1288 0 : broker_client,
1289 0 : remote_storage,
1290 0 : deletion_queue_client,
1291 0 : l0_flush_global_state,
1292 0 : } = resources;
1293 0 :
1294 0 : let attach_mode = attached_conf.location.attach_mode;
1295 0 : let generation = attached_conf.location.generation;
1296 0 :
1297 0 : let tenant = Arc::new(Tenant::new(
1298 0 : TenantState::Attaching,
1299 0 : conf,
1300 0 : attached_conf,
1301 0 : shard_identity,
1302 0 : Some(wal_redo_manager),
1303 0 : tenant_shard_id,
1304 0 : remote_storage.clone(),
1305 0 : deletion_queue_client,
1306 0 : l0_flush_global_state,
1307 0 : ));
1308 0 :
1309 0 : // The attach task will carry a GateGuard, so that shutdown() reliably waits for it to drop out if
1310 0 : // we shut down while attaching.
1311 0 : let attach_gate_guard = tenant
1312 0 : .gate
1313 0 : .enter()
1314 0 : .expect("We just created the Tenant: nothing else can have shut it down yet");
1315 0 :
1316 0 : // Do all the hard work in the background
1317 0 : let tenant_clone = Arc::clone(&tenant);
1318 0 : let ctx = ctx.detached_child(TaskKind::Attach, DownloadBehavior::Warn);
1319 0 : task_mgr::spawn(
1320 0 : &tokio::runtime::Handle::current(),
1321 0 : TaskKind::Attach,
1322 0 : tenant_shard_id,
1323 0 : None,
1324 0 : "attach tenant",
1325 0 : async move {
1326 0 :
1327 0 : info!(
1328 : ?attach_mode,
1329 0 : "Attaching tenant"
1330 : );
1331 :
1332 0 : let _gate_guard = attach_gate_guard;
1333 0 :
1334 0 : // Is this tenant being spawned as part of process startup?
1335 0 : let starting_up = init_order.is_some();
1336 0 : scopeguard::defer! {
1337 0 : if starting_up {
1338 0 : TENANT.startup_complete.inc();
1339 0 : }
1340 0 : }
1341 :
1342 : // Ideally we should use Tenant::set_broken_no_wait, but it is not supposed to be used when tenant is in loading state.
1343 : enum BrokenVerbosity {
1344 : Error,
1345 : Info
1346 : }
1347 0 : let make_broken =
1348 0 : |t: &Tenant, err: anyhow::Error, verbosity: BrokenVerbosity| {
1349 0 : match verbosity {
1350 : BrokenVerbosity::Info => {
1351 0 : info!("attach cancelled, setting tenant state to Broken: {err}");
1352 : },
1353 : BrokenVerbosity::Error => {
1354 0 : error!("attach failed, setting tenant state to Broken: {err:?}");
1355 : }
1356 : }
1357 0 : t.state.send_modify(|state| {
1358 0 : // The Stopping case is for when we have passed control on to DeleteTenantFlow:
1359 0 : // if it errors, we will call make_broken when tenant is already in Stopping.
1360 0 : assert!(
1361 0 : matches!(*state, TenantState::Attaching | TenantState::Stopping { .. }),
1362 0 : "the attach task owns the tenant state until activation is complete"
1363 : );
1364 :
1365 0 : *state = TenantState::broken_from_reason(err.to_string());
1366 0 : });
1367 0 : };
1368 :
1369 : // TODO: should also be rejecting tenant conf changes that violate this check.
1370 0 : if let Err(e) = crate::tenant::storage_layer::inmemory_layer::IndexEntry::validate_checkpoint_distance(tenant_clone.get_checkpoint_distance()) {
1371 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1372 0 : return Ok(());
1373 0 : }
1374 0 :
1375 0 : let mut init_order = init_order;
1376 0 : // take the completion because initial tenant loading will complete when all of
1377 0 : // these tasks complete.
1378 0 : let _completion = init_order
1379 0 : .as_mut()
1380 0 : .and_then(|x| x.initial_tenant_load.take());
1381 0 : let remote_load_completion = init_order
1382 0 : .as_mut()
1383 0 : .and_then(|x| x.initial_tenant_load_remote.take());
1384 :
1385 : enum AttachType<'a> {
1386 : /// We are attaching this tenant lazily in the background.
1387 : Warmup {
1388 : _permit: tokio::sync::SemaphorePermit<'a>,
1389 : during_startup: bool
1390 : },
1391 : /// We are attaching this tenant as soon as we can, because for example an
1392 : /// endpoint tried to access it.
1393 : OnDemand,
1394 : /// During normal operations after startup, we are attaching a tenant, and
1395 : /// eager attach was requested.
1396 : Normal,
1397 : }
1398 :
1399 0 : let attach_type = if matches!(mode, SpawnMode::Lazy) {
1400 : // Before doing any I/O, wait for at least one of:
1401 : // - A client attempting to access to this tenant (on-demand loading)
1402 : // - A permit becoming available in the warmup semaphore (background warmup)
1403 :
1404 0 : tokio::select!(
1405 0 : permit = tenant_clone.activate_now_sem.acquire() => {
1406 0 : let _ = permit.expect("activate_now_sem is never closed");
1407 0 : tracing::info!("Activating tenant (on-demand)");
1408 0 : AttachType::OnDemand
1409 : },
1410 0 : permit = conf.concurrent_tenant_warmup.inner().acquire() => {
1411 0 : let _permit = permit.expect("concurrent_tenant_warmup semaphore is never closed");
1412 0 : tracing::info!("Activating tenant (warmup)");
1413 0 : AttachType::Warmup {
1414 0 : _permit,
1415 0 : during_startup: init_order.is_some()
1416 0 : }
1417 : }
1418 0 : _ = tenant_clone.cancel.cancelled() => {
1419 : // This is safe, but should be pretty rare: it is interesting if a tenant
1420 : // stayed in Activating for such a long time that shutdown found it in
1421 : // that state.
1422 0 : tracing::info!(state=%tenant_clone.current_state(), "Tenant shut down before activation");
1423 : // Make the tenant broken so that set_stopping will not hang waiting for it to leave
1424 : // the Attaching state. This is an over-reaction (nothing really broke, the tenant is
1425 : // just shutting down), but ensures progress.
1426 0 : make_broken(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"), BrokenVerbosity::Info);
1427 0 : return Ok(());
1428 : },
1429 : )
1430 : } else {
1431 : // SpawnMode::{Create,Eager} always cause jumping ahead of the
1432 : // concurrent_tenant_warmup queue
1433 0 : AttachType::Normal
1434 : };
1435 :
1436 0 : let preload = match &mode {
1437 : SpawnMode::Eager | SpawnMode::Lazy => {
1438 0 : let _preload_timer = TENANT.preload.start_timer();
1439 0 : let res = tenant_clone
1440 0 : .preload(&remote_storage, task_mgr::shutdown_token())
1441 0 : .await;
1442 0 : match res {
1443 0 : Ok(p) => Some(p),
1444 0 : Err(e) => {
1445 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1446 0 : return Ok(());
1447 : }
1448 : }
1449 : }
1450 :
1451 : };
1452 :
1453 : // Remote preload is complete.
1454 0 : drop(remote_load_completion);
1455 0 :
1456 0 :
1457 0 : // We will time the duration of the attach phase unless this is a creation (attach will do no work)
1458 0 : let attach_start = std::time::Instant::now();
1459 0 : let attached = {
1460 0 : let _attach_timer = Some(TENANT.attach.start_timer());
1461 0 : tenant_clone.attach(preload, &ctx).await
1462 : };
1463 0 : let attach_duration = attach_start.elapsed();
1464 0 : _ = tenant_clone.attach_wal_lag_cooldown.set(WalLagCooldown::new(attach_start, attach_duration));
1465 0 :
1466 0 : match attached {
1467 : Ok(()) => {
1468 0 : info!("attach finished, activating");
1469 0 : tenant_clone.activate(broker_client, None, &ctx);
1470 : }
1471 0 : Err(e) => {
1472 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1473 0 : }
1474 : }
1475 :
1476 : // If we are doing an opportunistic warmup attachment at startup, initialize
1477 : // logical size at the same time. This is better than starting a bunch of idle tenants
1478 : // with cold caches and then coming back later to initialize their logical sizes.
1479 : //
1480 : // It also prevents the warmup proccess competing with the concurrency limit on
1481 : // logical size calculations: if logical size calculation semaphore is saturated,
1482 : // then warmup will wait for that before proceeding to the next tenant.
1483 0 : if matches!(attach_type, AttachType::Warmup { during_startup: true, .. }) {
1484 0 : let mut futs: FuturesUnordered<_> = tenant_clone.timelines.lock().unwrap().values().cloned().map(|t| t.await_initial_logical_size()).collect();
1485 0 : tracing::info!("Waiting for initial logical sizes while warming up...");
1486 0 : while futs.next().await.is_some() {}
1487 0 : tracing::info!("Warm-up complete");
1488 0 : }
1489 :
1490 0 : Ok(())
1491 0 : }
1492 0 : .instrument(tracing::info_span!(parent: None, "attach", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), gen=?generation)),
1493 : );
1494 0 : Ok(tenant)
1495 0 : }
1496 :
1497 196 : #[instrument(skip_all)]
1498 : pub(crate) async fn preload(
1499 : self: &Arc<Self>,
1500 : remote_storage: &GenericRemoteStorage,
1501 : cancel: CancellationToken,
1502 : ) -> anyhow::Result<TenantPreload> {
1503 : span::debug_assert_current_span_has_tenant_id();
1504 : // Get list of remote timelines
1505 : // download index files for every tenant timeline
1506 : info!("listing remote timelines");
1507 : let (mut remote_timeline_ids, other_keys) = remote_timeline_client::list_remote_timelines(
1508 : remote_storage,
1509 : self.tenant_shard_id,
1510 : cancel.clone(),
1511 : )
1512 : .await?;
1513 : let (offloaded_add, tenant_manifest) =
1514 : match remote_timeline_client::download_tenant_manifest(
1515 : remote_storage,
1516 : &self.tenant_shard_id,
1517 : self.generation,
1518 : &cancel,
1519 : )
1520 : .await
1521 : {
1522 : Ok((tenant_manifest, _generation, _manifest_mtime)) => (
1523 : format!("{} offloaded", tenant_manifest.offloaded_timelines.len()),
1524 : tenant_manifest,
1525 : ),
1526 : Err(DownloadError::NotFound) => {
1527 : ("no manifest".to_string(), TenantManifest::empty())
1528 : }
1529 : Err(e) => Err(e)?,
1530 : };
1531 :
1532 : info!(
1533 : "found {} timelines, and {offloaded_add}",
1534 : remote_timeline_ids.len()
1535 : );
1536 :
1537 : for k in other_keys {
1538 : warn!("Unexpected non timeline key {k}");
1539 : }
1540 :
1541 : // Avoid downloading IndexPart of offloaded timelines.
1542 : let mut offloaded_with_prefix = HashSet::new();
1543 : for offloaded in tenant_manifest.offloaded_timelines.iter() {
1544 : if remote_timeline_ids.remove(&offloaded.timeline_id) {
1545 : offloaded_with_prefix.insert(offloaded.timeline_id);
1546 : } else {
1547 : // We'll take care later of timelines in the manifest without a prefix
1548 : }
1549 : }
1550 :
1551 : let timelines = self
1552 : .load_timelines_metadata(remote_timeline_ids, remote_storage, cancel)
1553 : .await?;
1554 :
1555 : Ok(TenantPreload {
1556 : tenant_manifest,
1557 : timelines: timelines
1558 : .into_iter()
1559 6 : .map(|(id, tl)| (id, Some(tl)))
1560 0 : .chain(offloaded_with_prefix.into_iter().map(|id| (id, None)))
1561 : .collect(),
1562 : })
1563 : }
1564 :
1565 : ///
1566 : /// Background task that downloads all data for a tenant and brings it to Active state.
1567 : ///
1568 : /// No background tasks are started as part of this routine.
1569 : ///
1570 196 : async fn attach(
1571 196 : self: &Arc<Tenant>,
1572 196 : preload: Option<TenantPreload>,
1573 196 : ctx: &RequestContext,
1574 196 : ) -> anyhow::Result<()> {
1575 196 : span::debug_assert_current_span_has_tenant_id();
1576 196 :
1577 196 : failpoint_support::sleep_millis_async!("before-attaching-tenant");
1578 :
1579 196 : let Some(preload) = preload else {
1580 0 : anyhow::bail!("local-only deployment is no longer supported, https://github.com/neondatabase/neon/issues/5624");
1581 : };
1582 :
1583 196 : let mut offloaded_timeline_ids = HashSet::new();
1584 196 : let mut offloaded_timelines_list = Vec::new();
1585 196 : for timeline_manifest in preload.tenant_manifest.offloaded_timelines.iter() {
1586 0 : let timeline_id = timeline_manifest.timeline_id;
1587 0 : let offloaded_timeline =
1588 0 : OffloadedTimeline::from_manifest(self.tenant_shard_id, timeline_manifest);
1589 0 : offloaded_timelines_list.push((timeline_id, Arc::new(offloaded_timeline)));
1590 0 : offloaded_timeline_ids.insert(timeline_id);
1591 0 : }
1592 : // Complete deletions for offloaded timeline id's from manifest.
1593 : // The manifest will be uploaded later in this function.
1594 196 : offloaded_timelines_list
1595 196 : .retain(|(offloaded_id, offloaded)| {
1596 0 : // Existence of a timeline is finally determined by the existence of an index-part.json in remote storage.
1597 0 : // If there is dangling references in another location, they need to be cleaned up.
1598 0 : let delete = !preload.timelines.contains_key(offloaded_id);
1599 0 : if delete {
1600 0 : tracing::info!("Removing offloaded timeline {offloaded_id} from manifest as no remote prefix was found");
1601 0 : offloaded.defuse_for_tenant_drop();
1602 0 : }
1603 0 : !delete
1604 196 : });
1605 196 :
1606 196 : let mut timelines_to_resume_deletions = vec![];
1607 196 :
1608 196 : let mut remote_index_and_client = HashMap::new();
1609 196 : let mut timeline_ancestors = HashMap::new();
1610 196 : let mut existent_timelines = HashSet::new();
1611 202 : for (timeline_id, preload) in preload.timelines {
1612 6 : let Some(preload) = preload else { continue };
1613 : // This is an invariant of the `preload` function's API
1614 6 : assert!(!offloaded_timeline_ids.contains(&timeline_id));
1615 6 : let index_part = match preload.index_part {
1616 6 : Ok(i) => {
1617 6 : debug!("remote index part exists for timeline {timeline_id}");
1618 : // We found index_part on the remote, this is the standard case.
1619 6 : existent_timelines.insert(timeline_id);
1620 6 : i
1621 : }
1622 : Err(DownloadError::NotFound) => {
1623 : // There is no index_part on the remote. We only get here
1624 : // if there is some prefix for the timeline in the remote storage.
1625 : // This can e.g. be the initdb.tar.zst archive, maybe a
1626 : // remnant from a prior incomplete creation or deletion attempt.
1627 : // Delete the local directory as the deciding criterion for a
1628 : // timeline's existence is presence of index_part.
1629 0 : info!(%timeline_id, "index_part not found on remote");
1630 0 : continue;
1631 : }
1632 0 : Err(DownloadError::Fatal(why)) => {
1633 0 : // If, while loading one remote timeline, we saw an indication that our generation
1634 0 : // number is likely invalid, then we should not load the whole tenant.
1635 0 : error!(%timeline_id, "Fatal error loading timeline: {why}");
1636 0 : anyhow::bail!(why.to_string());
1637 : }
1638 0 : Err(e) => {
1639 0 : // Some (possibly ephemeral) error happened during index_part download.
1640 0 : // Pretend the timeline exists to not delete the timeline directory,
1641 0 : // as it might be a temporary issue and we don't want to re-download
1642 0 : // everything after it resolves.
1643 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1644 :
1645 0 : existent_timelines.insert(timeline_id);
1646 0 : continue;
1647 : }
1648 : };
1649 6 : match index_part {
1650 6 : MaybeDeletedIndexPart::IndexPart(index_part) => {
1651 6 : timeline_ancestors.insert(timeline_id, index_part.metadata.clone());
1652 6 : remote_index_and_client.insert(timeline_id, (index_part, preload.client));
1653 6 : }
1654 0 : MaybeDeletedIndexPart::Deleted(index_part) => {
1655 0 : info!(
1656 0 : "timeline {} is deleted, picking to resume deletion",
1657 : timeline_id
1658 : );
1659 0 : timelines_to_resume_deletions.push((timeline_id, index_part, preload.client));
1660 : }
1661 : }
1662 : }
1663 :
1664 196 : let mut gc_blocks = HashMap::new();
1665 :
1666 : // For every timeline, download the metadata file, scan the local directory,
1667 : // and build a layer map that contains an entry for each remote and local
1668 : // layer file.
1669 196 : let sorted_timelines = tree_sort_timelines(timeline_ancestors, |m| m.ancestor_timeline())?;
1670 202 : for (timeline_id, remote_metadata) in sorted_timelines {
1671 6 : let (index_part, remote_client) = remote_index_and_client
1672 6 : .remove(&timeline_id)
1673 6 : .expect("just put it in above");
1674 :
1675 6 : if let Some(blocking) = index_part.gc_blocking.as_ref() {
1676 : // could just filter these away, but it helps while testing
1677 0 : anyhow::ensure!(
1678 0 : !blocking.reasons.is_empty(),
1679 0 : "index_part for {timeline_id} is malformed: it should not have gc blocking with zero reasons"
1680 : );
1681 0 : let prev = gc_blocks.insert(timeline_id, blocking.reasons);
1682 0 : assert!(prev.is_none());
1683 6 : }
1684 :
1685 : // TODO again handle early failure
1686 6 : let effect = self
1687 6 : .load_remote_timeline(
1688 6 : timeline_id,
1689 6 : index_part,
1690 6 : remote_metadata,
1691 6 : TimelineResources {
1692 6 : remote_client,
1693 6 : pagestream_throttle: self.pagestream_throttle.clone(),
1694 6 : l0_flush_global_state: self.l0_flush_global_state.clone(),
1695 6 : },
1696 6 : LoadTimelineCause::Attach,
1697 6 : ctx,
1698 6 : )
1699 6 : .await
1700 6 : .with_context(|| {
1701 0 : format!(
1702 0 : "failed to load remote timeline {} for tenant {}",
1703 0 : timeline_id, self.tenant_shard_id
1704 0 : )
1705 6 : })?;
1706 :
1707 6 : match effect {
1708 6 : TimelineInitAndSyncResult::ReadyToActivate(_) => {
1709 6 : // activation happens later, on Tenant::activate
1710 6 : }
1711 : TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
1712 : TimelineInitAndSyncNeedsSpawnImportPgdata {
1713 0 : timeline,
1714 0 : import_pgdata,
1715 0 : guard,
1716 0 : },
1717 0 : ) => {
1718 0 : tokio::task::spawn(self.clone().create_timeline_import_pgdata_task(
1719 0 : timeline,
1720 0 : import_pgdata,
1721 0 : ActivateTimelineArgs::No,
1722 0 : guard,
1723 0 : ));
1724 0 : }
1725 : }
1726 : }
1727 :
1728 : // Walk through deleted timelines, resume deletion
1729 196 : for (timeline_id, index_part, remote_timeline_client) in timelines_to_resume_deletions {
1730 0 : remote_timeline_client
1731 0 : .init_upload_queue_stopped_to_continue_deletion(&index_part)
1732 0 : .context("init queue stopped")
1733 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1734 :
1735 0 : DeleteTimelineFlow::resume_deletion(
1736 0 : Arc::clone(self),
1737 0 : timeline_id,
1738 0 : &index_part.metadata,
1739 0 : remote_timeline_client,
1740 0 : )
1741 0 : .instrument(tracing::info_span!("timeline_delete", %timeline_id))
1742 0 : .await
1743 0 : .context("resume_deletion")
1744 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1745 : }
1746 196 : let needs_manifest_upload =
1747 196 : offloaded_timelines_list.len() != preload.tenant_manifest.offloaded_timelines.len();
1748 196 : {
1749 196 : let mut offloaded_timelines_accessor = self.timelines_offloaded.lock().unwrap();
1750 196 : offloaded_timelines_accessor.extend(offloaded_timelines_list.into_iter());
1751 196 : }
1752 196 : if needs_manifest_upload {
1753 0 : self.store_tenant_manifest().await?;
1754 196 : }
1755 :
1756 : // The local filesystem contents are a cache of what's in the remote IndexPart;
1757 : // IndexPart is the source of truth.
1758 196 : self.clean_up_timelines(&existent_timelines)?;
1759 :
1760 196 : self.gc_block.set_scanned(gc_blocks);
1761 196 :
1762 196 : fail::fail_point!("attach-before-activate", |_| {
1763 0 : anyhow::bail!("attach-before-activate");
1764 196 : });
1765 196 : failpoint_support::sleep_millis_async!("attach-before-activate-sleep", &self.cancel);
1766 :
1767 196 : info!("Done");
1768 :
1769 196 : Ok(())
1770 196 : }
1771 :
1772 : /// Check for any local timeline directories that are temporary, or do not correspond to a
1773 : /// timeline that still exists: this can happen if we crashed during a deletion/creation, or
1774 : /// if a timeline was deleted while the tenant was attached to a different pageserver.
1775 196 : fn clean_up_timelines(&self, existent_timelines: &HashSet<TimelineId>) -> anyhow::Result<()> {
1776 196 : let timelines_dir = self.conf.timelines_path(&self.tenant_shard_id);
1777 :
1778 196 : let entries = match timelines_dir.read_dir_utf8() {
1779 196 : Ok(d) => d,
1780 0 : Err(e) => {
1781 0 : if e.kind() == std::io::ErrorKind::NotFound {
1782 0 : return Ok(());
1783 : } else {
1784 0 : return Err(e).context("list timelines directory for tenant");
1785 : }
1786 : }
1787 : };
1788 :
1789 204 : for entry in entries {
1790 8 : let entry = entry.context("read timeline dir entry")?;
1791 8 : let entry_path = entry.path();
1792 :
1793 8 : let purge = if crate::is_temporary(entry_path)
1794 : // TODO: remove uninit mark code (https://github.com/neondatabase/neon/issues/5718)
1795 8 : || is_uninit_mark(entry_path)
1796 8 : || crate::is_delete_mark(entry_path)
1797 : {
1798 0 : true
1799 : } else {
1800 8 : match TimelineId::try_from(entry_path.file_name()) {
1801 8 : Ok(i) => {
1802 8 : // Purge if the timeline ID does not exist in remote storage: remote storage is the authority.
1803 8 : !existent_timelines.contains(&i)
1804 : }
1805 0 : Err(e) => {
1806 0 : tracing::warn!(
1807 0 : "Unparseable directory in timelines directory: {entry_path}, ignoring ({e})"
1808 : );
1809 : // Do not purge junk: if we don't recognize it, be cautious and leave it for a human.
1810 0 : false
1811 : }
1812 : }
1813 : };
1814 :
1815 8 : if purge {
1816 2 : tracing::info!("Purging stale timeline dentry {entry_path}");
1817 2 : if let Err(e) = match entry.file_type() {
1818 2 : Ok(t) => if t.is_dir() {
1819 2 : std::fs::remove_dir_all(entry_path)
1820 : } else {
1821 0 : std::fs::remove_file(entry_path)
1822 : }
1823 2 : .or_else(fs_ext::ignore_not_found),
1824 0 : Err(e) => Err(e),
1825 : } {
1826 0 : tracing::warn!("Failed to purge stale timeline dentry {entry_path}: {e}");
1827 2 : }
1828 6 : }
1829 : }
1830 :
1831 196 : Ok(())
1832 196 : }
1833 :
1834 : /// Get sum of all remote timelines sizes
1835 : ///
1836 : /// This function relies on the index_part instead of listing the remote storage
1837 0 : pub fn remote_size(&self) -> u64 {
1838 0 : let mut size = 0;
1839 :
1840 0 : for timeline in self.list_timelines() {
1841 0 : size += timeline.remote_client.get_remote_physical_size();
1842 0 : }
1843 :
1844 0 : size
1845 0 : }
1846 :
1847 6 : #[instrument(skip_all, fields(timeline_id=%timeline_id))]
1848 : async fn load_remote_timeline(
1849 : self: &Arc<Self>,
1850 : timeline_id: TimelineId,
1851 : index_part: IndexPart,
1852 : remote_metadata: TimelineMetadata,
1853 : resources: TimelineResources,
1854 : cause: LoadTimelineCause,
1855 : ctx: &RequestContext,
1856 : ) -> anyhow::Result<TimelineInitAndSyncResult> {
1857 : span::debug_assert_current_span_has_tenant_id();
1858 :
1859 : info!("downloading index file for timeline {}", timeline_id);
1860 : tokio::fs::create_dir_all(self.conf.timeline_path(&self.tenant_shard_id, &timeline_id))
1861 : .await
1862 : .context("Failed to create new timeline directory")?;
1863 :
1864 : let ancestor = if let Some(ancestor_id) = remote_metadata.ancestor_timeline() {
1865 : let timelines = self.timelines.lock().unwrap();
1866 : Some(Arc::clone(timelines.get(&ancestor_id).ok_or_else(
1867 0 : || {
1868 0 : anyhow::anyhow!(
1869 0 : "cannot find ancestor timeline {ancestor_id} for timeline {timeline_id}"
1870 0 : )
1871 0 : },
1872 : )?))
1873 : } else {
1874 : None
1875 : };
1876 :
1877 : self.timeline_init_and_sync(
1878 : timeline_id,
1879 : resources,
1880 : index_part,
1881 : remote_metadata,
1882 : ancestor,
1883 : cause,
1884 : ctx,
1885 : )
1886 : .await
1887 : }
1888 :
1889 196 : async fn load_timelines_metadata(
1890 196 : self: &Arc<Tenant>,
1891 196 : timeline_ids: HashSet<TimelineId>,
1892 196 : remote_storage: &GenericRemoteStorage,
1893 196 : cancel: CancellationToken,
1894 196 : ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
1895 196 : let mut part_downloads = JoinSet::new();
1896 202 : for timeline_id in timeline_ids {
1897 6 : let cancel_clone = cancel.clone();
1898 6 : part_downloads.spawn(
1899 6 : self.load_timeline_metadata(timeline_id, remote_storage.clone(), cancel_clone)
1900 6 : .instrument(info_span!("download_index_part", %timeline_id)),
1901 : );
1902 : }
1903 :
1904 196 : let mut timeline_preloads: HashMap<TimelineId, TimelinePreload> = HashMap::new();
1905 :
1906 : loop {
1907 202 : tokio::select!(
1908 202 : next = part_downloads.join_next() => {
1909 202 : match next {
1910 6 : Some(result) => {
1911 6 : let preload = result.context("join preload task")?;
1912 6 : timeline_preloads.insert(preload.timeline_id, preload);
1913 : },
1914 : None => {
1915 196 : break;
1916 : }
1917 : }
1918 : },
1919 202 : _ = cancel.cancelled() => {
1920 0 : anyhow::bail!("Cancelled while waiting for remote index download")
1921 : }
1922 : )
1923 : }
1924 :
1925 196 : Ok(timeline_preloads)
1926 196 : }
1927 :
1928 6 : fn build_timeline_client(
1929 6 : &self,
1930 6 : timeline_id: TimelineId,
1931 6 : remote_storage: GenericRemoteStorage,
1932 6 : ) -> RemoteTimelineClient {
1933 6 : RemoteTimelineClient::new(
1934 6 : remote_storage.clone(),
1935 6 : self.deletion_queue_client.clone(),
1936 6 : self.conf,
1937 6 : self.tenant_shard_id,
1938 6 : timeline_id,
1939 6 : self.generation,
1940 6 : &self.tenant_conf.load().location,
1941 6 : )
1942 6 : }
1943 :
1944 6 : fn load_timeline_metadata(
1945 6 : self: &Arc<Tenant>,
1946 6 : timeline_id: TimelineId,
1947 6 : remote_storage: GenericRemoteStorage,
1948 6 : cancel: CancellationToken,
1949 6 : ) -> impl Future<Output = TimelinePreload> {
1950 6 : let client = self.build_timeline_client(timeline_id, remote_storage);
1951 6 : async move {
1952 6 : debug_assert_current_span_has_tenant_and_timeline_id();
1953 6 : debug!("starting index part download");
1954 :
1955 6 : let index_part = client.download_index_file(&cancel).await;
1956 :
1957 6 : debug!("finished index part download");
1958 :
1959 6 : TimelinePreload {
1960 6 : client,
1961 6 : timeline_id,
1962 6 : index_part,
1963 6 : }
1964 6 : }
1965 6 : }
1966 :
1967 0 : fn check_to_be_archived_has_no_unarchived_children(
1968 0 : timeline_id: TimelineId,
1969 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
1970 0 : ) -> Result<(), TimelineArchivalError> {
1971 0 : let children: Vec<TimelineId> = timelines
1972 0 : .iter()
1973 0 : .filter_map(|(id, entry)| {
1974 0 : if entry.get_ancestor_timeline_id() != Some(timeline_id) {
1975 0 : return None;
1976 0 : }
1977 0 : if entry.is_archived() == Some(true) {
1978 0 : return None;
1979 0 : }
1980 0 : Some(*id)
1981 0 : })
1982 0 : .collect();
1983 0 :
1984 0 : if !children.is_empty() {
1985 0 : return Err(TimelineArchivalError::HasUnarchivedChildren(children));
1986 0 : }
1987 0 : Ok(())
1988 0 : }
1989 :
1990 0 : fn check_ancestor_of_to_be_unarchived_is_not_archived(
1991 0 : ancestor_timeline_id: TimelineId,
1992 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
1993 0 : offloaded_timelines: &std::sync::MutexGuard<
1994 0 : '_,
1995 0 : HashMap<TimelineId, Arc<OffloadedTimeline>>,
1996 0 : >,
1997 0 : ) -> Result<(), TimelineArchivalError> {
1998 0 : let has_archived_parent =
1999 0 : if let Some(ancestor_timeline) = timelines.get(&ancestor_timeline_id) {
2000 0 : ancestor_timeline.is_archived() == Some(true)
2001 0 : } else if offloaded_timelines.contains_key(&ancestor_timeline_id) {
2002 0 : true
2003 : } else {
2004 0 : error!("ancestor timeline {ancestor_timeline_id} not found");
2005 0 : if cfg!(debug_assertions) {
2006 0 : panic!("ancestor timeline {ancestor_timeline_id} not found");
2007 0 : }
2008 0 : return Err(TimelineArchivalError::NotFound);
2009 : };
2010 0 : if has_archived_parent {
2011 0 : return Err(TimelineArchivalError::HasArchivedParent(
2012 0 : ancestor_timeline_id,
2013 0 : ));
2014 0 : }
2015 0 : Ok(())
2016 0 : }
2017 :
2018 0 : fn check_to_be_unarchived_timeline_has_no_archived_parent(
2019 0 : timeline: &Arc<Timeline>,
2020 0 : ) -> Result<(), TimelineArchivalError> {
2021 0 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
2022 0 : if ancestor_timeline.is_archived() == Some(true) {
2023 0 : return Err(TimelineArchivalError::HasArchivedParent(
2024 0 : ancestor_timeline.timeline_id,
2025 0 : ));
2026 0 : }
2027 0 : }
2028 0 : Ok(())
2029 0 : }
2030 :
2031 : /// Loads the specified (offloaded) timeline from S3 and attaches it as a loaded timeline
2032 : ///
2033 : /// Counterpart to [`offload_timeline`].
2034 0 : async fn unoffload_timeline(
2035 0 : self: &Arc<Self>,
2036 0 : timeline_id: TimelineId,
2037 0 : broker_client: storage_broker::BrokerClientChannel,
2038 0 : ctx: RequestContext,
2039 0 : ) -> Result<Arc<Timeline>, TimelineArchivalError> {
2040 0 : info!("unoffloading timeline");
2041 :
2042 : // We activate the timeline below manually, so this must be called on an active timeline.
2043 : // We expect callers of this function to ensure this.
2044 0 : match self.current_state() {
2045 : TenantState::Activating { .. }
2046 : | TenantState::Attaching
2047 : | TenantState::Broken { .. } => {
2048 0 : panic!("Timeline expected to be active")
2049 : }
2050 0 : TenantState::Stopping { .. } => return Err(TimelineArchivalError::Cancelled),
2051 0 : TenantState::Active => {}
2052 0 : }
2053 0 : let cancel = self.cancel.clone();
2054 0 :
2055 0 : // Protect against concurrent attempts to use this TimelineId
2056 0 : // We don't care much about idempotency, as it's ensured a layer above.
2057 0 : let allow_offloaded = true;
2058 0 : let _create_guard = self
2059 0 : .create_timeline_create_guard(
2060 0 : timeline_id,
2061 0 : CreateTimelineIdempotency::FailWithConflict,
2062 0 : allow_offloaded,
2063 0 : )
2064 0 : .map_err(|err| match err {
2065 0 : TimelineExclusionError::AlreadyCreating => TimelineArchivalError::AlreadyInProgress,
2066 : TimelineExclusionError::AlreadyExists { .. } => {
2067 0 : TimelineArchivalError::Other(anyhow::anyhow!("Timeline already exists"))
2068 : }
2069 0 : TimelineExclusionError::Other(e) => TimelineArchivalError::Other(e),
2070 0 : TimelineExclusionError::ShuttingDown => TimelineArchivalError::Cancelled,
2071 0 : })?;
2072 :
2073 0 : let timeline_preload = self
2074 0 : .load_timeline_metadata(timeline_id, self.remote_storage.clone(), cancel.clone())
2075 0 : .await;
2076 :
2077 0 : let index_part = match timeline_preload.index_part {
2078 0 : Ok(index_part) => {
2079 0 : debug!("remote index part exists for timeline {timeline_id}");
2080 0 : index_part
2081 : }
2082 : Err(DownloadError::NotFound) => {
2083 0 : error!(%timeline_id, "index_part not found on remote");
2084 0 : return Err(TimelineArchivalError::NotFound);
2085 : }
2086 0 : Err(DownloadError::Cancelled) => return Err(TimelineArchivalError::Cancelled),
2087 0 : Err(e) => {
2088 0 : // Some (possibly ephemeral) error happened during index_part download.
2089 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
2090 0 : return Err(TimelineArchivalError::Other(
2091 0 : anyhow::Error::new(e).context("downloading index_part from remote storage"),
2092 0 : ));
2093 : }
2094 : };
2095 0 : let index_part = match index_part {
2096 0 : MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
2097 0 : MaybeDeletedIndexPart::Deleted(_index_part) => {
2098 0 : info!("timeline is deleted according to index_part.json");
2099 0 : return Err(TimelineArchivalError::NotFound);
2100 : }
2101 : };
2102 0 : let remote_metadata = index_part.metadata.clone();
2103 0 : let timeline_resources = self.build_timeline_resources(timeline_id);
2104 0 : self.load_remote_timeline(
2105 0 : timeline_id,
2106 0 : index_part,
2107 0 : remote_metadata,
2108 0 : timeline_resources,
2109 0 : LoadTimelineCause::Unoffload,
2110 0 : &ctx,
2111 0 : )
2112 0 : .await
2113 0 : .with_context(|| {
2114 0 : format!(
2115 0 : "failed to load remote timeline {} for tenant {}",
2116 0 : timeline_id, self.tenant_shard_id
2117 0 : )
2118 0 : })
2119 0 : .map_err(TimelineArchivalError::Other)?;
2120 :
2121 0 : let timeline = {
2122 0 : let timelines = self.timelines.lock().unwrap();
2123 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2124 0 : warn!("timeline not available directly after attach");
2125 : // This is not a panic because no locks are held between `load_remote_timeline`
2126 : // which puts the timeline into timelines, and our look into the timeline map.
2127 0 : return Err(TimelineArchivalError::Other(anyhow::anyhow!(
2128 0 : "timeline not available directly after attach"
2129 0 : )));
2130 : };
2131 0 : let mut offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2132 0 : match offloaded_timelines.remove(&timeline_id) {
2133 0 : Some(offloaded) => {
2134 0 : offloaded.delete_from_ancestor_with_timelines(&timelines);
2135 0 : }
2136 0 : None => warn!("timeline already removed from offloaded timelines"),
2137 : }
2138 :
2139 0 : self.initialize_gc_info(&timelines, &offloaded_timelines, Some(timeline_id));
2140 0 :
2141 0 : Arc::clone(timeline)
2142 0 : };
2143 0 :
2144 0 : // Upload new list of offloaded timelines to S3
2145 0 : self.store_tenant_manifest().await?;
2146 :
2147 : // Activate the timeline (if it makes sense)
2148 0 : if !(timeline.is_broken() || timeline.is_stopping()) {
2149 0 : let background_jobs_can_start = None;
2150 0 : timeline.activate(
2151 0 : self.clone(),
2152 0 : broker_client.clone(),
2153 0 : background_jobs_can_start,
2154 0 : &ctx,
2155 0 : );
2156 0 : }
2157 :
2158 0 : info!("timeline unoffloading complete");
2159 0 : Ok(timeline)
2160 0 : }
2161 :
2162 0 : pub(crate) async fn apply_timeline_archival_config(
2163 0 : self: &Arc<Self>,
2164 0 : timeline_id: TimelineId,
2165 0 : new_state: TimelineArchivalState,
2166 0 : broker_client: storage_broker::BrokerClientChannel,
2167 0 : ctx: RequestContext,
2168 0 : ) -> Result<(), TimelineArchivalError> {
2169 0 : info!("setting timeline archival config");
2170 : // First part: figure out what is needed to do, and do validation
2171 0 : let timeline_or_unarchive_offloaded = 'outer: {
2172 0 : let timelines = self.timelines.lock().unwrap();
2173 :
2174 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2175 0 : let offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2176 0 : let Some(offloaded) = offloaded_timelines.get(&timeline_id) else {
2177 0 : return Err(TimelineArchivalError::NotFound);
2178 : };
2179 0 : if new_state == TimelineArchivalState::Archived {
2180 : // It's offloaded already, so nothing to do
2181 0 : return Ok(());
2182 0 : }
2183 0 : if let Some(ancestor_timeline_id) = offloaded.ancestor_timeline_id {
2184 0 : Self::check_ancestor_of_to_be_unarchived_is_not_archived(
2185 0 : ancestor_timeline_id,
2186 0 : &timelines,
2187 0 : &offloaded_timelines,
2188 0 : )?;
2189 0 : }
2190 0 : break 'outer None;
2191 : };
2192 :
2193 : // Do some validation. We release the timelines lock below, so there is potential
2194 : // for race conditions: these checks are more present to prevent misunderstandings of
2195 : // the API's capabilities, instead of serving as the sole way to defend their invariants.
2196 0 : match new_state {
2197 : TimelineArchivalState::Unarchived => {
2198 0 : Self::check_to_be_unarchived_timeline_has_no_archived_parent(timeline)?
2199 : }
2200 : TimelineArchivalState::Archived => {
2201 0 : Self::check_to_be_archived_has_no_unarchived_children(timeline_id, &timelines)?
2202 : }
2203 : }
2204 0 : Some(Arc::clone(timeline))
2205 : };
2206 :
2207 : // Second part: unoffload timeline (if needed)
2208 0 : let timeline = if let Some(timeline) = timeline_or_unarchive_offloaded {
2209 0 : timeline
2210 : } else {
2211 : // Turn offloaded timeline into a non-offloaded one
2212 0 : self.unoffload_timeline(timeline_id, broker_client, ctx)
2213 0 : .await?
2214 : };
2215 :
2216 : // Third part: upload new timeline archival state and block until it is present in S3
2217 0 : let upload_needed = match timeline
2218 0 : .remote_client
2219 0 : .schedule_index_upload_for_timeline_archival_state(new_state)
2220 : {
2221 0 : Ok(upload_needed) => upload_needed,
2222 0 : Err(e) => {
2223 0 : if timeline.cancel.is_cancelled() {
2224 0 : return Err(TimelineArchivalError::Cancelled);
2225 : } else {
2226 0 : return Err(TimelineArchivalError::Other(e));
2227 : }
2228 : }
2229 : };
2230 :
2231 0 : if upload_needed {
2232 0 : info!("Uploading new state");
2233 : const MAX_WAIT: Duration = Duration::from_secs(10);
2234 0 : let Ok(v) =
2235 0 : tokio::time::timeout(MAX_WAIT, timeline.remote_client.wait_completion()).await
2236 : else {
2237 0 : tracing::warn!("reached timeout for waiting on upload queue");
2238 0 : return Err(TimelineArchivalError::Timeout);
2239 : };
2240 0 : v.map_err(|e| match e {
2241 0 : WaitCompletionError::NotInitialized(e) => {
2242 0 : TimelineArchivalError::Other(anyhow::anyhow!(e))
2243 : }
2244 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2245 0 : TimelineArchivalError::Cancelled
2246 : }
2247 0 : })?;
2248 0 : }
2249 0 : Ok(())
2250 0 : }
2251 :
2252 2 : pub fn get_offloaded_timeline(
2253 2 : &self,
2254 2 : timeline_id: TimelineId,
2255 2 : ) -> Result<Arc<OffloadedTimeline>, GetTimelineError> {
2256 2 : self.timelines_offloaded
2257 2 : .lock()
2258 2 : .unwrap()
2259 2 : .get(&timeline_id)
2260 2 : .map(Arc::clone)
2261 2 : .ok_or(GetTimelineError::NotFound {
2262 2 : tenant_id: self.tenant_shard_id,
2263 2 : timeline_id,
2264 2 : })
2265 2 : }
2266 :
2267 4 : pub(crate) fn tenant_shard_id(&self) -> TenantShardId {
2268 4 : self.tenant_shard_id
2269 4 : }
2270 :
2271 : /// Get Timeline handle for given Neon timeline ID.
2272 : /// This function is idempotent. It doesn't change internal state in any way.
2273 222 : pub fn get_timeline(
2274 222 : &self,
2275 222 : timeline_id: TimelineId,
2276 222 : active_only: bool,
2277 222 : ) -> Result<Arc<Timeline>, GetTimelineError> {
2278 222 : let timelines_accessor = self.timelines.lock().unwrap();
2279 222 : let timeline = timelines_accessor
2280 222 : .get(&timeline_id)
2281 222 : .ok_or(GetTimelineError::NotFound {
2282 222 : tenant_id: self.tenant_shard_id,
2283 222 : timeline_id,
2284 222 : })?;
2285 :
2286 220 : if active_only && !timeline.is_active() {
2287 0 : Err(GetTimelineError::NotActive {
2288 0 : tenant_id: self.tenant_shard_id,
2289 0 : timeline_id,
2290 0 : state: timeline.current_state(),
2291 0 : })
2292 : } else {
2293 220 : Ok(Arc::clone(timeline))
2294 : }
2295 222 : }
2296 :
2297 : /// Lists timelines the tenant contains.
2298 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2299 0 : pub fn list_timelines(&self) -> Vec<Arc<Timeline>> {
2300 0 : self.timelines
2301 0 : .lock()
2302 0 : .unwrap()
2303 0 : .values()
2304 0 : .map(Arc::clone)
2305 0 : .collect()
2306 0 : }
2307 :
2308 : /// Lists timelines the tenant manages, including offloaded ones.
2309 : ///
2310 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2311 0 : pub fn list_timelines_and_offloaded(
2312 0 : &self,
2313 0 : ) -> (Vec<Arc<Timeline>>, Vec<Arc<OffloadedTimeline>>) {
2314 0 : let timelines = self
2315 0 : .timelines
2316 0 : .lock()
2317 0 : .unwrap()
2318 0 : .values()
2319 0 : .map(Arc::clone)
2320 0 : .collect();
2321 0 : let offloaded = self
2322 0 : .timelines_offloaded
2323 0 : .lock()
2324 0 : .unwrap()
2325 0 : .values()
2326 0 : .map(Arc::clone)
2327 0 : .collect();
2328 0 : (timelines, offloaded)
2329 0 : }
2330 :
2331 0 : pub fn list_timeline_ids(&self) -> Vec<TimelineId> {
2332 0 : self.timelines.lock().unwrap().keys().cloned().collect()
2333 0 : }
2334 :
2335 : /// This is used by tests & import-from-basebackup.
2336 : ///
2337 : /// The returned [`UninitializedTimeline`] contains no data nor metadata and it is in
2338 : /// a state that will fail [`Tenant::load_remote_timeline`] because `disk_consistent_lsn=Lsn(0)`.
2339 : ///
2340 : /// The caller is responsible for getting the timeline into a state that will be accepted
2341 : /// by [`Tenant::load_remote_timeline`] / [`Tenant::attach`].
2342 : /// Then they may call [`UninitializedTimeline::finish_creation`] to add the timeline
2343 : /// to the [`Tenant::timelines`].
2344 : ///
2345 : /// Tests should use `Tenant::create_test_timeline` to set up the minimum required metadata keys.
2346 188 : pub(crate) async fn create_empty_timeline(
2347 188 : self: &Arc<Self>,
2348 188 : new_timeline_id: TimelineId,
2349 188 : initdb_lsn: Lsn,
2350 188 : pg_version: u32,
2351 188 : _ctx: &RequestContext,
2352 188 : ) -> anyhow::Result<UninitializedTimeline> {
2353 188 : anyhow::ensure!(
2354 188 : self.is_active(),
2355 0 : "Cannot create empty timelines on inactive tenant"
2356 : );
2357 :
2358 : // Protect against concurrent attempts to use this TimelineId
2359 188 : let create_guard = match self
2360 188 : .start_creating_timeline(new_timeline_id, CreateTimelineIdempotency::FailWithConflict)
2361 188 : .await?
2362 : {
2363 186 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2364 : StartCreatingTimelineResult::Idempotent(_) => {
2365 0 : unreachable!("FailWithConflict implies we get an error instead")
2366 : }
2367 : };
2368 :
2369 186 : let new_metadata = TimelineMetadata::new(
2370 186 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2371 186 : // make it valid, before calling finish_creation()
2372 186 : Lsn(0),
2373 186 : None,
2374 186 : None,
2375 186 : Lsn(0),
2376 186 : initdb_lsn,
2377 186 : initdb_lsn,
2378 186 : pg_version,
2379 186 : );
2380 186 : self.prepare_new_timeline(
2381 186 : new_timeline_id,
2382 186 : &new_metadata,
2383 186 : create_guard,
2384 186 : initdb_lsn,
2385 186 : None,
2386 186 : )
2387 186 : .await
2388 188 : }
2389 :
2390 : /// Helper for unit tests to create an empty timeline.
2391 : ///
2392 : /// The timeline is has state value `Active` but its background loops are not running.
2393 : // This makes the various functions which anyhow::ensure! for Active state work in tests.
2394 : // Our current tests don't need the background loops.
2395 : #[cfg(test)]
2396 178 : pub async fn create_test_timeline(
2397 178 : self: &Arc<Self>,
2398 178 : new_timeline_id: TimelineId,
2399 178 : initdb_lsn: Lsn,
2400 178 : pg_version: u32,
2401 178 : ctx: &RequestContext,
2402 178 : ) -> anyhow::Result<Arc<Timeline>> {
2403 178 : let uninit_tl = self
2404 178 : .create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2405 178 : .await?;
2406 178 : let tline = uninit_tl.raw_timeline().expect("we just created it");
2407 178 : assert_eq!(tline.get_last_record_lsn(), Lsn(0));
2408 :
2409 : // Setup minimum keys required for the timeline to be usable.
2410 178 : let mut modification = tline.begin_modification(initdb_lsn);
2411 178 : modification
2412 178 : .init_empty_test_timeline()
2413 178 : .context("init_empty_test_timeline")?;
2414 178 : modification
2415 178 : .commit(ctx)
2416 178 : .await
2417 178 : .context("commit init_empty_test_timeline modification")?;
2418 :
2419 : // Flush to disk so that uninit_tl's check for valid disk_consistent_lsn passes.
2420 178 : tline.maybe_spawn_flush_loop();
2421 178 : tline.freeze_and_flush().await.context("freeze_and_flush")?;
2422 :
2423 : // Make sure the freeze_and_flush reaches remote storage.
2424 178 : tline.remote_client.wait_completion().await.unwrap();
2425 :
2426 178 : let tl = uninit_tl.finish_creation()?;
2427 : // The non-test code would call tl.activate() here.
2428 178 : tl.set_state(TimelineState::Active);
2429 178 : Ok(tl)
2430 178 : }
2431 :
2432 : /// Helper for unit tests to create a timeline with some pre-loaded states.
2433 : #[cfg(test)]
2434 : #[allow(clippy::too_many_arguments)]
2435 36 : pub async fn create_test_timeline_with_layers(
2436 36 : self: &Arc<Self>,
2437 36 : new_timeline_id: TimelineId,
2438 36 : initdb_lsn: Lsn,
2439 36 : pg_version: u32,
2440 36 : ctx: &RequestContext,
2441 36 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
2442 36 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
2443 36 : end_lsn: Lsn,
2444 36 : ) -> anyhow::Result<Arc<Timeline>> {
2445 : use checks::check_valid_layermap;
2446 : use itertools::Itertools;
2447 :
2448 36 : let tline = self
2449 36 : .create_test_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2450 36 : .await?;
2451 36 : tline.force_advance_lsn(end_lsn);
2452 120 : for deltas in delta_layer_desc {
2453 84 : tline
2454 84 : .force_create_delta_layer(deltas, Some(initdb_lsn), ctx)
2455 84 : .await?;
2456 : }
2457 88 : for (lsn, images) in image_layer_desc {
2458 52 : tline
2459 52 : .force_create_image_layer(lsn, images, Some(initdb_lsn), ctx)
2460 52 : .await?;
2461 : }
2462 36 : let layer_names = tline
2463 36 : .layers
2464 36 : .read()
2465 36 : .await
2466 36 : .layer_map()
2467 36 : .unwrap()
2468 36 : .iter_historic_layers()
2469 172 : .map(|layer| layer.layer_name())
2470 36 : .collect_vec();
2471 36 : if let Some(err) = check_valid_layermap(&layer_names) {
2472 0 : bail!("invalid layermap: {err}");
2473 36 : }
2474 36 : Ok(tline)
2475 36 : }
2476 :
2477 : /// Create a new timeline.
2478 : ///
2479 : /// Returns the new timeline ID and reference to its Timeline object.
2480 : ///
2481 : /// If the caller specified the timeline ID to use (`new_timeline_id`), and timeline with
2482 : /// the same timeline ID already exists, returns CreateTimelineError::AlreadyExists.
2483 : #[allow(clippy::too_many_arguments)]
2484 0 : pub(crate) async fn create_timeline(
2485 0 : self: &Arc<Tenant>,
2486 0 : params: CreateTimelineParams,
2487 0 : broker_client: storage_broker::BrokerClientChannel,
2488 0 : ctx: &RequestContext,
2489 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
2490 0 : if !self.is_active() {
2491 0 : if matches!(self.current_state(), TenantState::Stopping { .. }) {
2492 0 : return Err(CreateTimelineError::ShuttingDown);
2493 : } else {
2494 0 : return Err(CreateTimelineError::Other(anyhow::anyhow!(
2495 0 : "Cannot create timelines on inactive tenant"
2496 0 : )));
2497 : }
2498 0 : }
2499 :
2500 0 : let _gate = self
2501 0 : .gate
2502 0 : .enter()
2503 0 : .map_err(|_| CreateTimelineError::ShuttingDown)?;
2504 :
2505 0 : let result: CreateTimelineResult = match params {
2506 : CreateTimelineParams::Bootstrap(CreateTimelineParamsBootstrap {
2507 0 : new_timeline_id,
2508 0 : existing_initdb_timeline_id,
2509 0 : pg_version,
2510 0 : }) => {
2511 0 : self.bootstrap_timeline(
2512 0 : new_timeline_id,
2513 0 : pg_version,
2514 0 : existing_initdb_timeline_id,
2515 0 : ctx,
2516 0 : )
2517 0 : .await?
2518 : }
2519 : CreateTimelineParams::Branch(CreateTimelineParamsBranch {
2520 0 : new_timeline_id,
2521 0 : ancestor_timeline_id,
2522 0 : mut ancestor_start_lsn,
2523 : }) => {
2524 0 : let ancestor_timeline = self
2525 0 : .get_timeline(ancestor_timeline_id, false)
2526 0 : .context("Cannot branch off the timeline that's not present in pageserver")?;
2527 :
2528 : // instead of waiting around, just deny the request because ancestor is not yet
2529 : // ready for other purposes either.
2530 0 : if !ancestor_timeline.is_active() {
2531 0 : return Err(CreateTimelineError::AncestorNotActive);
2532 0 : }
2533 0 :
2534 0 : if ancestor_timeline.is_archived() == Some(true) {
2535 0 : info!("tried to branch archived timeline");
2536 0 : return Err(CreateTimelineError::AncestorArchived);
2537 0 : }
2538 :
2539 0 : if let Some(lsn) = ancestor_start_lsn.as_mut() {
2540 0 : *lsn = lsn.align();
2541 0 :
2542 0 : let ancestor_ancestor_lsn = ancestor_timeline.get_ancestor_lsn();
2543 0 : if ancestor_ancestor_lsn > *lsn {
2544 : // can we safely just branch from the ancestor instead?
2545 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
2546 0 : "invalid start lsn {} for ancestor timeline {}: less than timeline ancestor lsn {}",
2547 0 : lsn,
2548 0 : ancestor_timeline_id,
2549 0 : ancestor_ancestor_lsn,
2550 0 : )));
2551 0 : }
2552 0 :
2553 0 : // Wait for the WAL to arrive and be processed on the parent branch up
2554 0 : // to the requested branch point. The repository code itself doesn't
2555 0 : // require it, but if we start to receive WAL on the new timeline,
2556 0 : // decoding the new WAL might need to look up previous pages, relation
2557 0 : // sizes etc. and that would get confused if the previous page versions
2558 0 : // are not in the repository yet.
2559 0 : ancestor_timeline
2560 0 : .wait_lsn(*lsn, timeline::WaitLsnWaiter::Tenant, ctx)
2561 0 : .await
2562 0 : .map_err(|e| match e {
2563 0 : e @ (WaitLsnError::Timeout(_) | WaitLsnError::BadState { .. }) => {
2564 0 : CreateTimelineError::AncestorLsn(anyhow::anyhow!(e))
2565 : }
2566 0 : WaitLsnError::Shutdown => CreateTimelineError::ShuttingDown,
2567 0 : })?;
2568 0 : }
2569 :
2570 0 : self.branch_timeline(&ancestor_timeline, new_timeline_id, ancestor_start_lsn, ctx)
2571 0 : .await?
2572 : }
2573 0 : CreateTimelineParams::ImportPgdata(params) => {
2574 0 : self.create_timeline_import_pgdata(
2575 0 : params,
2576 0 : ActivateTimelineArgs::Yes {
2577 0 : broker_client: broker_client.clone(),
2578 0 : },
2579 0 : ctx,
2580 0 : )
2581 0 : .await?
2582 : }
2583 : };
2584 :
2585 : // At this point we have dropped our guard on [`Self::timelines_creating`], and
2586 : // the timeline is visible in [`Self::timelines`], but it is _not_ durable yet. We must
2587 : // not send a success to the caller until it is. The same applies to idempotent retries.
2588 : //
2589 : // TODO: the timeline is already visible in [`Self::timelines`]; a caller could incorrectly
2590 : // assume that, because they can see the timeline via API, that the creation is done and
2591 : // that it is durable. Ideally, we would keep the timeline hidden (in [`Self::timelines_creating`])
2592 : // until it is durable, e.g., by extending the time we hold the creation guard. This also
2593 : // interacts with UninitializedTimeline and is generally a bit tricky.
2594 : //
2595 : // To re-emphasize: the only correct way to create a timeline is to repeat calling the
2596 : // creation API until it returns success. Only then is durability guaranteed.
2597 0 : info!(creation_result=%result.discriminant(), "waiting for timeline to be durable");
2598 0 : result
2599 0 : .timeline()
2600 0 : .remote_client
2601 0 : .wait_completion()
2602 0 : .await
2603 0 : .map_err(|e| match e {
2604 : WaitCompletionError::NotInitialized(
2605 0 : e, // If the queue is already stopped, it's a shutdown error.
2606 0 : ) if e.is_stopping() => CreateTimelineError::ShuttingDown,
2607 : WaitCompletionError::NotInitialized(_) => {
2608 : // This is a bug: we should never try to wait for uploads before initializing the timeline
2609 0 : debug_assert!(false);
2610 0 : CreateTimelineError::Other(anyhow::anyhow!("timeline not initialized"))
2611 : }
2612 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2613 0 : CreateTimelineError::ShuttingDown
2614 : }
2615 0 : })?;
2616 :
2617 : // The creating task is responsible for activating the timeline.
2618 : // We do this after `wait_completion()` so that we don't spin up tasks that start
2619 : // doing stuff before the IndexPart is durable in S3, which is done by the previous section.
2620 0 : let activated_timeline = match result {
2621 0 : CreateTimelineResult::Created(timeline) => {
2622 0 : timeline.activate(self.clone(), broker_client, None, ctx);
2623 0 : timeline
2624 : }
2625 0 : CreateTimelineResult::Idempotent(timeline) => {
2626 0 : info!(
2627 0 : "request was deemed idempotent, activation will be done by the creating task"
2628 : );
2629 0 : timeline
2630 : }
2631 0 : CreateTimelineResult::ImportSpawned(timeline) => {
2632 0 : info!("import task spawned, timeline will become visible and activated once the import is done");
2633 0 : timeline
2634 : }
2635 : };
2636 :
2637 0 : Ok(activated_timeline)
2638 0 : }
2639 :
2640 : /// The returned [`Arc<Timeline>`] is NOT in the [`Tenant::timelines`] map until the import
2641 : /// completes in the background. A DIFFERENT [`Arc<Timeline>`] will be inserted into the
2642 : /// [`Tenant::timelines`] map when the import completes.
2643 : /// We only return an [`Arc<Timeline>`] here so the API handler can create a [`pageserver_api::models::TimelineInfo`]
2644 : /// for the response.
2645 0 : async fn create_timeline_import_pgdata(
2646 0 : self: &Arc<Tenant>,
2647 0 : params: CreateTimelineParamsImportPgdata,
2648 0 : activate: ActivateTimelineArgs,
2649 0 : ctx: &RequestContext,
2650 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
2651 0 : let CreateTimelineParamsImportPgdata {
2652 0 : new_timeline_id,
2653 0 : location,
2654 0 : idempotency_key,
2655 0 : } = params;
2656 0 :
2657 0 : let started_at = chrono::Utc::now().naive_utc();
2658 :
2659 : //
2660 : // There's probably a simpler way to upload an index part, but, remote_timeline_client
2661 : // is the canonical way we do it.
2662 : // - create an empty timeline in-memory
2663 : // - use its remote_timeline_client to do the upload
2664 : // - dispose of the uninit timeline
2665 : // - keep the creation guard alive
2666 :
2667 0 : let timeline_create_guard = match self
2668 0 : .start_creating_timeline(
2669 0 : new_timeline_id,
2670 0 : CreateTimelineIdempotency::ImportPgdata(CreatingTimelineIdempotencyImportPgdata {
2671 0 : idempotency_key: idempotency_key.clone(),
2672 0 : }),
2673 0 : )
2674 0 : .await?
2675 : {
2676 0 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2677 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
2678 0 : return Ok(CreateTimelineResult::Idempotent(timeline))
2679 : }
2680 : };
2681 :
2682 0 : let mut uninit_timeline = {
2683 0 : let this = &self;
2684 0 : let initdb_lsn = Lsn(0);
2685 0 : let _ctx = ctx;
2686 0 : async move {
2687 0 : let new_metadata = TimelineMetadata::new(
2688 0 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2689 0 : // make it valid, before calling finish_creation()
2690 0 : Lsn(0),
2691 0 : None,
2692 0 : None,
2693 0 : Lsn(0),
2694 0 : initdb_lsn,
2695 0 : initdb_lsn,
2696 0 : 15,
2697 0 : );
2698 0 : this.prepare_new_timeline(
2699 0 : new_timeline_id,
2700 0 : &new_metadata,
2701 0 : timeline_create_guard,
2702 0 : initdb_lsn,
2703 0 : None,
2704 0 : )
2705 0 : .await
2706 0 : }
2707 0 : }
2708 0 : .await?;
2709 :
2710 0 : let in_progress = import_pgdata::index_part_format::InProgress {
2711 0 : idempotency_key,
2712 0 : location,
2713 0 : started_at,
2714 0 : };
2715 0 : let index_part = import_pgdata::index_part_format::Root::V1(
2716 0 : import_pgdata::index_part_format::V1::InProgress(in_progress),
2717 0 : );
2718 0 : uninit_timeline
2719 0 : .raw_timeline()
2720 0 : .unwrap()
2721 0 : .remote_client
2722 0 : .schedule_index_upload_for_import_pgdata_state_update(Some(index_part.clone()))?;
2723 :
2724 : // wait_completion happens in caller
2725 :
2726 0 : let (timeline, timeline_create_guard) = uninit_timeline.finish_creation_myself();
2727 0 :
2728 0 : tokio::spawn(self.clone().create_timeline_import_pgdata_task(
2729 0 : timeline.clone(),
2730 0 : index_part,
2731 0 : activate,
2732 0 : timeline_create_guard,
2733 0 : ));
2734 0 :
2735 0 : // NB: the timeline doesn't exist in self.timelines at this point
2736 0 : Ok(CreateTimelineResult::ImportSpawned(timeline))
2737 0 : }
2738 :
2739 0 : #[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))]
2740 : async fn create_timeline_import_pgdata_task(
2741 : self: Arc<Tenant>,
2742 : timeline: Arc<Timeline>,
2743 : index_part: import_pgdata::index_part_format::Root,
2744 : activate: ActivateTimelineArgs,
2745 : timeline_create_guard: TimelineCreateGuard,
2746 : ) {
2747 : debug_assert_current_span_has_tenant_and_timeline_id();
2748 : info!("starting");
2749 : scopeguard::defer! {info!("exiting")};
2750 :
2751 : let res = self
2752 : .create_timeline_import_pgdata_task_impl(
2753 : timeline,
2754 : index_part,
2755 : activate,
2756 : timeline_create_guard,
2757 : )
2758 : .await;
2759 : if let Err(err) = &res {
2760 : error!(?err, "task failed");
2761 : // TODO sleep & retry, sensitive to tenant shutdown
2762 : // TODO: allow timeline deletion requests => should cancel the task
2763 : }
2764 : }
2765 :
2766 0 : async fn create_timeline_import_pgdata_task_impl(
2767 0 : self: Arc<Tenant>,
2768 0 : timeline: Arc<Timeline>,
2769 0 : index_part: import_pgdata::index_part_format::Root,
2770 0 : activate: ActivateTimelineArgs,
2771 0 : timeline_create_guard: TimelineCreateGuard,
2772 0 : ) -> Result<(), anyhow::Error> {
2773 0 : let ctx = RequestContext::new(TaskKind::ImportPgdata, DownloadBehavior::Warn);
2774 0 :
2775 0 : info!("importing pgdata");
2776 0 : import_pgdata::doit(&timeline, index_part, &ctx, self.cancel.clone())
2777 0 : .await
2778 0 : .context("import")?;
2779 0 : info!("import done");
2780 :
2781 : //
2782 : // Reload timeline from remote.
2783 : // This proves that the remote state is attachable, and it reuses the code.
2784 : //
2785 : // TODO: think about whether this is safe to do with concurrent Tenant::shutdown.
2786 : // timeline_create_guard hols the tenant gate open, so, shutdown cannot _complete_ until we exit.
2787 : // But our activate() call might launch new background tasks after Tenant::shutdown
2788 : // already went past shutting down the Tenant::timelines, which this timeline here is no part of.
2789 : // I think the same problem exists with the bootstrap & branch mgmt API tasks (tenant shutting
2790 : // down while bootstrapping/branching + activating), but, the race condition is much more likely
2791 : // to manifest because of the long runtime of this import task.
2792 :
2793 : // in theory this shouldn't even .await anything except for coop yield
2794 0 : info!("shutting down timeline");
2795 0 : timeline.shutdown(ShutdownMode::Hard).await;
2796 0 : info!("timeline shut down, reloading from remote");
2797 : // TODO: we can't do the following check because create_timeline_import_pgdata must return an Arc<Timeline>
2798 : // let Some(timeline) = Arc::into_inner(timeline) else {
2799 : // anyhow::bail!("implementation error: timeline that we shut down was still referenced from somewhere");
2800 : // };
2801 0 : let timeline_id = timeline.timeline_id;
2802 0 :
2803 0 : // load from object storage like Tenant::attach does
2804 0 : let resources = self.build_timeline_resources(timeline_id);
2805 0 : let index_part = resources
2806 0 : .remote_client
2807 0 : .download_index_file(&self.cancel)
2808 0 : .await?;
2809 0 : let index_part = match index_part {
2810 : MaybeDeletedIndexPart::Deleted(_) => {
2811 : // likely concurrent delete call, cplane should prevent this
2812 0 : anyhow::bail!("index part says deleted but we are not done creating yet, this should not happen but")
2813 : }
2814 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
2815 0 : };
2816 0 : let metadata = index_part.metadata.clone();
2817 0 : self
2818 0 : .load_remote_timeline(timeline_id, index_part, metadata, resources, LoadTimelineCause::ImportPgdata{
2819 0 : create_guard: timeline_create_guard, activate, }, &ctx)
2820 0 : .await?
2821 0 : .ready_to_activate()
2822 0 : .context("implementation error: reloaded timeline still needs import after import reported success")?;
2823 :
2824 0 : anyhow::Ok(())
2825 0 : }
2826 :
2827 0 : pub(crate) async fn delete_timeline(
2828 0 : self: Arc<Self>,
2829 0 : timeline_id: TimelineId,
2830 0 : ) -> Result<(), DeleteTimelineError> {
2831 0 : DeleteTimelineFlow::run(&self, timeline_id).await?;
2832 :
2833 0 : Ok(())
2834 0 : }
2835 :
2836 : /// perform one garbage collection iteration, removing old data files from disk.
2837 : /// this function is periodically called by gc task.
2838 : /// also it can be explicitly requested through page server api 'do_gc' command.
2839 : ///
2840 : /// `target_timeline_id` specifies the timeline to GC, or None for all.
2841 : ///
2842 : /// The `horizon` an `pitr` parameters determine how much WAL history needs to be retained.
2843 : /// Also known as the retention period, or the GC cutoff point. `horizon` specifies
2844 : /// the amount of history, as LSN difference from current latest LSN on each timeline.
2845 : /// `pitr` specifies the same as a time difference from the current time. The effective
2846 : /// GC cutoff point is determined conservatively by either `horizon` and `pitr`, whichever
2847 : /// requires more history to be retained.
2848 : //
2849 754 : pub(crate) async fn gc_iteration(
2850 754 : &self,
2851 754 : target_timeline_id: Option<TimelineId>,
2852 754 : horizon: u64,
2853 754 : pitr: Duration,
2854 754 : cancel: &CancellationToken,
2855 754 : ctx: &RequestContext,
2856 754 : ) -> Result<GcResult, GcError> {
2857 754 : // Don't start doing work during shutdown
2858 754 : if let TenantState::Stopping { .. } = self.current_state() {
2859 0 : return Ok(GcResult::default());
2860 754 : }
2861 754 :
2862 754 : // there is a global allowed_error for this
2863 754 : if !self.is_active() {
2864 0 : return Err(GcError::NotActive);
2865 754 : }
2866 754 :
2867 754 : {
2868 754 : let conf = self.tenant_conf.load();
2869 754 :
2870 754 : // If we may not delete layers, then simply skip GC. Even though a tenant
2871 754 : // in AttachedMulti state could do GC and just enqueue the blocked deletions,
2872 754 : // the only advantage to doing it is to perhaps shrink the LayerMap metadata
2873 754 : // a bit sooner than we would achieve by waiting for AttachedSingle status.
2874 754 : if !conf.location.may_delete_layers_hint() {
2875 0 : info!("Skipping GC in location state {:?}", conf.location);
2876 0 : return Ok(GcResult::default());
2877 754 : }
2878 754 :
2879 754 : if conf.is_gc_blocked_by_lsn_lease_deadline() {
2880 750 : info!("Skipping GC because lsn lease deadline is not reached");
2881 750 : return Ok(GcResult::default());
2882 4 : }
2883 : }
2884 :
2885 4 : let _guard = match self.gc_block.start().await {
2886 4 : Ok(guard) => guard,
2887 0 : Err(reasons) => {
2888 0 : info!("Skipping GC: {reasons}");
2889 0 : return Ok(GcResult::default());
2890 : }
2891 : };
2892 :
2893 4 : self.gc_iteration_internal(target_timeline_id, horizon, pitr, cancel, ctx)
2894 4 : .await
2895 754 : }
2896 :
2897 : /// Perform one compaction iteration.
2898 : /// This function is periodically called by compactor task.
2899 : /// Also it can be explicitly requested per timeline through page server
2900 : /// api's 'compact' command.
2901 : ///
2902 : /// Returns whether we have pending compaction task.
2903 0 : async fn compaction_iteration(
2904 0 : self: &Arc<Self>,
2905 0 : cancel: &CancellationToken,
2906 0 : ctx: &RequestContext,
2907 0 : ) -> Result<bool, timeline::CompactionError> {
2908 0 : // Don't start doing work during shutdown, or when broken, we do not need those in the logs
2909 0 : if !self.is_active() {
2910 0 : return Ok(false);
2911 0 : }
2912 0 :
2913 0 : {
2914 0 : let conf = self.tenant_conf.load();
2915 0 :
2916 0 : // Note that compaction usually requires deletions, but we don't respect
2917 0 : // may_delete_layers_hint here: that is because tenants in AttachedMulti
2918 0 : // should proceed with compaction even if they can't do deletion, to avoid
2919 0 : // accumulating dangerously deep stacks of L0 layers. Deletions will be
2920 0 : // enqueued inside RemoteTimelineClient, and executed layer if/when we transition
2921 0 : // to AttachedSingle state.
2922 0 : if !conf.location.may_upload_layers_hint() {
2923 0 : info!("Skipping compaction in location state {:?}", conf.location);
2924 0 : return Ok(false);
2925 0 : }
2926 0 : }
2927 0 :
2928 0 : // Scan through the hashmap and collect a list of all the timelines,
2929 0 : // while holding the lock. Then drop the lock and actually perform the
2930 0 : // compactions. We don't want to block everything else while the
2931 0 : // compaction runs.
2932 0 : let timelines_to_compact_or_offload;
2933 0 : {
2934 0 : let timelines = self.timelines.lock().unwrap();
2935 0 : timelines_to_compact_or_offload = timelines
2936 0 : .iter()
2937 0 : .filter_map(|(timeline_id, timeline)| {
2938 0 : let (is_active, (can_offload, _)) =
2939 0 : (timeline.is_active(), timeline.can_offload());
2940 0 : let has_no_unoffloaded_children = {
2941 0 : !timelines
2942 0 : .iter()
2943 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(*timeline_id))
2944 : };
2945 0 : let config_allows_offload = self.conf.timeline_offloading
2946 0 : || self
2947 0 : .tenant_conf
2948 0 : .load()
2949 0 : .tenant_conf
2950 0 : .timeline_offloading
2951 0 : .unwrap_or_default();
2952 0 : let can_offload =
2953 0 : can_offload && has_no_unoffloaded_children && config_allows_offload;
2954 0 : if (is_active, can_offload) == (false, false) {
2955 0 : None
2956 : } else {
2957 0 : Some((*timeline_id, timeline.clone(), (is_active, can_offload)))
2958 : }
2959 0 : })
2960 0 : .collect::<Vec<_>>();
2961 0 : drop(timelines);
2962 0 : }
2963 0 :
2964 0 : // Before doing any I/O work, check our circuit breaker
2965 0 : if self.compaction_circuit_breaker.lock().unwrap().is_broken() {
2966 0 : info!("Skipping compaction due to previous failures");
2967 0 : return Ok(false);
2968 0 : }
2969 0 :
2970 0 : let mut has_pending_task = false;
2971 :
2972 0 : for (timeline_id, timeline, (can_compact, can_offload)) in &timelines_to_compact_or_offload
2973 : {
2974 : // pending_task_left == None: cannot compact, maybe still pending tasks
2975 : // pending_task_left == Some(true): compaction task left
2976 : // pending_task_left == Some(false): no compaction task left
2977 0 : let pending_task_left = if *can_compact {
2978 0 : let has_pending_l0_compaction_task = timeline
2979 0 : .compact(cancel, EnumSet::empty(), ctx)
2980 0 : .instrument(info_span!("compact_timeline", %timeline_id))
2981 0 : .await
2982 0 : .inspect_err(|e| match e {
2983 0 : timeline::CompactionError::ShuttingDown => (),
2984 0 : timeline::CompactionError::Offload(_) => {
2985 0 : // Failures to offload timelines do not trip the circuit breaker, because
2986 0 : // they do not do lots of writes the way compaction itself does: it is cheap
2987 0 : // to retry, and it would be bad to stop all compaction because of an issue with offloading.
2988 0 : }
2989 0 : timeline::CompactionError::Other(e) => {
2990 0 : self.compaction_circuit_breaker
2991 0 : .lock()
2992 0 : .unwrap()
2993 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, e);
2994 0 : }
2995 0 : })?;
2996 0 : if has_pending_l0_compaction_task {
2997 0 : Some(true)
2998 : } else {
2999 : let mut has_pending_scheduled_compaction_task;
3000 0 : let next_scheduled_compaction_task = {
3001 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3002 0 : if let Some(tline_pending_tasks) = guard.get_mut(timeline_id) {
3003 0 : if !tline_pending_tasks.is_empty() {
3004 0 : info!(
3005 0 : "{} tasks left in the compaction schedule queue",
3006 0 : tline_pending_tasks.len()
3007 : );
3008 0 : }
3009 0 : let next_task = tline_pending_tasks.pop_front();
3010 0 : has_pending_scheduled_compaction_task = !tline_pending_tasks.is_empty();
3011 0 : next_task
3012 : } else {
3013 0 : has_pending_scheduled_compaction_task = false;
3014 0 : None
3015 : }
3016 : };
3017 0 : if let Some(mut next_scheduled_compaction_task) = next_scheduled_compaction_task
3018 : {
3019 0 : if !next_scheduled_compaction_task
3020 0 : .options
3021 0 : .flags
3022 0 : .contains(CompactFlags::EnhancedGcBottomMostCompaction)
3023 : {
3024 0 : warn!("ignoring scheduled compaction task: scheduled task must be gc compaction: {:?}", next_scheduled_compaction_task.options);
3025 0 : } else if next_scheduled_compaction_task.options.sub_compaction {
3026 0 : info!("running scheduled enhanced gc bottom-most compaction with sub-compaction, splitting compaction jobs");
3027 0 : let jobs: Vec<GcCompactJob> = timeline
3028 0 : .gc_compaction_split_jobs(
3029 0 : GcCompactJob::from_compact_options(
3030 0 : next_scheduled_compaction_task.options.clone(),
3031 0 : ),
3032 0 : next_scheduled_compaction_task
3033 0 : .options
3034 0 : .sub_compaction_max_job_size_mb,
3035 0 : )
3036 0 : .await
3037 0 : .map_err(CompactionError::Other)?;
3038 0 : if jobs.is_empty() {
3039 0 : info!("no jobs to run, skipping scheduled compaction task");
3040 : } else {
3041 0 : has_pending_scheduled_compaction_task = true;
3042 0 : let jobs_len = jobs.len();
3043 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3044 0 : let tline_pending_tasks = guard.entry(*timeline_id).or_default();
3045 0 : for (idx, job) in jobs.into_iter().enumerate() {
3046 : // Unfortunately we need to convert the `GcCompactJob` back to `CompactionOptions`
3047 : // until we do further refactors to allow directly call `compact_with_gc`.
3048 0 : let mut flags: EnumSet<CompactFlags> = EnumSet::default();
3049 0 : flags |= CompactFlags::EnhancedGcBottomMostCompaction;
3050 0 : if job.dry_run {
3051 0 : flags |= CompactFlags::DryRun;
3052 0 : }
3053 0 : let options = CompactOptions {
3054 0 : flags,
3055 0 : sub_compaction: false,
3056 0 : compact_key_range: Some(job.compact_key_range.into()),
3057 0 : compact_lsn_range: Some(job.compact_lsn_range.into()),
3058 0 : sub_compaction_max_job_size_mb: None,
3059 0 : };
3060 0 : tline_pending_tasks.push_back(if idx == jobs_len - 1 {
3061 0 : ScheduledCompactionTask {
3062 0 : options,
3063 0 : // The last job in the queue sends the signal and releases the gc guard
3064 0 : result_tx: next_scheduled_compaction_task
3065 0 : .result_tx
3066 0 : .take(),
3067 0 : gc_block: next_scheduled_compaction_task
3068 0 : .gc_block
3069 0 : .take(),
3070 0 : }
3071 : } else {
3072 0 : ScheduledCompactionTask {
3073 0 : options,
3074 0 : result_tx: None,
3075 0 : gc_block: None,
3076 0 : }
3077 : });
3078 : }
3079 0 : info!("scheduled enhanced gc bottom-most compaction with sub-compaction, split into {} jobs", jobs_len);
3080 : }
3081 : } else {
3082 0 : let _ = timeline
3083 0 : .compact_with_options(
3084 0 : cancel,
3085 0 : next_scheduled_compaction_task.options,
3086 0 : ctx,
3087 0 : )
3088 0 : .instrument(info_span!("scheduled_compact_timeline", %timeline_id))
3089 0 : .await?;
3090 0 : if let Some(tx) = next_scheduled_compaction_task.result_tx.take() {
3091 0 : // TODO: we can send compaction statistics in the future
3092 0 : tx.send(()).ok();
3093 0 : }
3094 : }
3095 0 : }
3096 0 : Some(has_pending_scheduled_compaction_task)
3097 : }
3098 : } else {
3099 0 : None
3100 : };
3101 0 : has_pending_task |= pending_task_left.unwrap_or(false);
3102 0 : if pending_task_left == Some(false) && *can_offload {
3103 0 : offload_timeline(self, timeline)
3104 0 : .instrument(info_span!("offload_timeline", %timeline_id))
3105 0 : .await?;
3106 0 : }
3107 : }
3108 :
3109 0 : self.compaction_circuit_breaker
3110 0 : .lock()
3111 0 : .unwrap()
3112 0 : .success(&CIRCUIT_BREAKERS_UNBROKEN);
3113 0 :
3114 0 : Ok(has_pending_task)
3115 0 : }
3116 :
3117 : /// Cancel scheduled compaction tasks
3118 0 : pub(crate) fn cancel_scheduled_compaction(
3119 0 : &self,
3120 0 : timeline_id: TimelineId,
3121 0 : ) -> Vec<ScheduledCompactionTask> {
3122 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3123 0 : if let Some(tline_pending_tasks) = guard.get_mut(&timeline_id) {
3124 0 : let current_tline_pending_tasks = std::mem::take(tline_pending_tasks);
3125 0 : current_tline_pending_tasks.into_iter().collect()
3126 : } else {
3127 0 : Vec::new()
3128 : }
3129 0 : }
3130 :
3131 0 : pub(crate) fn get_scheduled_compaction_tasks(
3132 0 : &self,
3133 0 : timeline_id: TimelineId,
3134 0 : ) -> Vec<CompactOptions> {
3135 : use itertools::Itertools;
3136 0 : let guard = self.scheduled_compaction_tasks.lock().unwrap();
3137 0 : guard
3138 0 : .get(&timeline_id)
3139 0 : .map(|tline_pending_tasks| {
3140 0 : tline_pending_tasks
3141 0 : .iter()
3142 0 : .map(|x| x.options.clone())
3143 0 : .collect_vec()
3144 0 : })
3145 0 : .unwrap_or_default()
3146 0 : }
3147 :
3148 : /// Schedule a compaction task for a timeline.
3149 0 : pub(crate) async fn schedule_compaction(
3150 0 : &self,
3151 0 : timeline_id: TimelineId,
3152 0 : options: CompactOptions,
3153 0 : ) -> anyhow::Result<tokio::sync::oneshot::Receiver<()>> {
3154 0 : let gc_guard = match self.gc_block.start().await {
3155 0 : Ok(guard) => guard,
3156 0 : Err(e) => {
3157 0 : bail!("cannot run gc-compaction because gc is blocked: {}", e);
3158 : }
3159 : };
3160 0 : let (tx, rx) = tokio::sync::oneshot::channel();
3161 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3162 0 : let tline_pending_tasks = guard.entry(timeline_id).or_default();
3163 0 : tline_pending_tasks.push_back(ScheduledCompactionTask {
3164 0 : options,
3165 0 : result_tx: Some(tx),
3166 0 : gc_block: Some(gc_guard),
3167 0 : });
3168 0 : Ok(rx)
3169 0 : }
3170 :
3171 : // Call through to all timelines to freeze ephemeral layers if needed. Usually
3172 : // this happens during ingest: this background housekeeping is for freezing layers
3173 : // that are open but haven't been written to for some time.
3174 0 : async fn ingest_housekeeping(&self) {
3175 0 : // Scan through the hashmap and collect a list of all the timelines,
3176 0 : // while holding the lock. Then drop the lock and actually perform the
3177 0 : // compactions. We don't want to block everything else while the
3178 0 : // compaction runs.
3179 0 : let timelines = {
3180 0 : self.timelines
3181 0 : .lock()
3182 0 : .unwrap()
3183 0 : .values()
3184 0 : .filter_map(|timeline| {
3185 0 : if timeline.is_active() {
3186 0 : Some(timeline.clone())
3187 : } else {
3188 0 : None
3189 : }
3190 0 : })
3191 0 : .collect::<Vec<_>>()
3192 : };
3193 :
3194 0 : for timeline in &timelines {
3195 0 : timeline.maybe_freeze_ephemeral_layer().await;
3196 : }
3197 0 : }
3198 :
3199 0 : pub fn timeline_has_no_attached_children(&self, timeline_id: TimelineId) -> bool {
3200 0 : let timelines = self.timelines.lock().unwrap();
3201 0 : !timelines
3202 0 : .iter()
3203 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(timeline_id))
3204 0 : }
3205 :
3206 1708 : pub fn current_state(&self) -> TenantState {
3207 1708 : self.state.borrow().clone()
3208 1708 : }
3209 :
3210 946 : pub fn is_active(&self) -> bool {
3211 946 : self.current_state() == TenantState::Active
3212 946 : }
3213 :
3214 0 : pub fn generation(&self) -> Generation {
3215 0 : self.generation
3216 0 : }
3217 :
3218 0 : pub(crate) fn wal_redo_manager_status(&self) -> Option<WalRedoManagerStatus> {
3219 0 : self.walredo_mgr.as_ref().and_then(|mgr| mgr.status())
3220 0 : }
3221 :
3222 : /// Changes tenant status to active, unless shutdown was already requested.
3223 : ///
3224 : /// `background_jobs_can_start` is an optional barrier set to a value during pageserver startup
3225 : /// to delay background jobs. Background jobs can be started right away when None is given.
3226 0 : fn activate(
3227 0 : self: &Arc<Self>,
3228 0 : broker_client: BrokerClientChannel,
3229 0 : background_jobs_can_start: Option<&completion::Barrier>,
3230 0 : ctx: &RequestContext,
3231 0 : ) {
3232 0 : span::debug_assert_current_span_has_tenant_id();
3233 0 :
3234 0 : let mut activating = false;
3235 0 : self.state.send_modify(|current_state| {
3236 : use pageserver_api::models::ActivatingFrom;
3237 0 : match &*current_state {
3238 : TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
3239 0 : panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {state:?}", state = current_state);
3240 : }
3241 0 : TenantState::Attaching => {
3242 0 : *current_state = TenantState::Activating(ActivatingFrom::Attaching);
3243 0 : }
3244 0 : }
3245 0 : debug!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), "Activating tenant");
3246 0 : activating = true;
3247 0 : // Continue outside the closure. We need to grab timelines.lock()
3248 0 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
3249 0 : });
3250 0 :
3251 0 : if activating {
3252 0 : let timelines_accessor = self.timelines.lock().unwrap();
3253 0 : let timelines_offloaded_accessor = self.timelines_offloaded.lock().unwrap();
3254 0 : let timelines_to_activate = timelines_accessor
3255 0 : .values()
3256 0 : .filter(|timeline| !(timeline.is_broken() || timeline.is_stopping()));
3257 0 :
3258 0 : // Before activation, populate each Timeline's GcInfo with information about its children
3259 0 : self.initialize_gc_info(&timelines_accessor, &timelines_offloaded_accessor, None);
3260 0 :
3261 0 : // Spawn gc and compaction loops. The loops will shut themselves
3262 0 : // down when they notice that the tenant is inactive.
3263 0 : tasks::start_background_loops(self, background_jobs_can_start);
3264 0 :
3265 0 : let mut activated_timelines = 0;
3266 :
3267 0 : for timeline in timelines_to_activate {
3268 0 : timeline.activate(
3269 0 : self.clone(),
3270 0 : broker_client.clone(),
3271 0 : background_jobs_can_start,
3272 0 : ctx,
3273 0 : );
3274 0 : activated_timelines += 1;
3275 0 : }
3276 :
3277 0 : self.state.send_modify(move |current_state| {
3278 0 : assert!(
3279 0 : matches!(current_state, TenantState::Activating(_)),
3280 0 : "set_stopping and set_broken wait for us to leave Activating state",
3281 : );
3282 0 : *current_state = TenantState::Active;
3283 0 :
3284 0 : let elapsed = self.constructed_at.elapsed();
3285 0 : let total_timelines = timelines_accessor.len();
3286 0 :
3287 0 : // log a lot of stuff, because some tenants sometimes suffer from user-visible
3288 0 : // times to activate. see https://github.com/neondatabase/neon/issues/4025
3289 0 : info!(
3290 0 : since_creation_millis = elapsed.as_millis(),
3291 0 : tenant_id = %self.tenant_shard_id.tenant_id,
3292 0 : shard_id = %self.tenant_shard_id.shard_slug(),
3293 0 : activated_timelines,
3294 0 : total_timelines,
3295 0 : post_state = <&'static str>::from(&*current_state),
3296 0 : "activation attempt finished"
3297 : );
3298 :
3299 0 : TENANT.activation.observe(elapsed.as_secs_f64());
3300 0 : });
3301 0 : }
3302 0 : }
3303 :
3304 : /// Shutdown the tenant and join all of the spawned tasks.
3305 : ///
3306 : /// The method caters for all use-cases:
3307 : /// - pageserver shutdown (freeze_and_flush == true)
3308 : /// - detach + ignore (freeze_and_flush == false)
3309 : ///
3310 : /// This will attempt to shutdown even if tenant is broken.
3311 : ///
3312 : /// `shutdown_progress` is a [`completion::Barrier`] for the shutdown initiated by this call.
3313 : /// If the tenant is already shutting down, we return a clone of the first shutdown call's
3314 : /// `Barrier` as an `Err`. This not-first caller can use the returned barrier to join with
3315 : /// the ongoing shutdown.
3316 6 : async fn shutdown(
3317 6 : &self,
3318 6 : shutdown_progress: completion::Barrier,
3319 6 : shutdown_mode: timeline::ShutdownMode,
3320 6 : ) -> Result<(), completion::Barrier> {
3321 6 : span::debug_assert_current_span_has_tenant_id();
3322 :
3323 : // Set tenant (and its timlines) to Stoppping state.
3324 : //
3325 : // Since we can only transition into Stopping state after activation is complete,
3326 : // run it in a JoinSet so all tenants have a chance to stop before we get SIGKILLed.
3327 : //
3328 : // Transitioning tenants to Stopping state has a couple of non-obvious side effects:
3329 : // 1. Lock out any new requests to the tenants.
3330 : // 2. Signal cancellation to WAL receivers (we wait on it below).
3331 : // 3. Signal cancellation for other tenant background loops.
3332 : // 4. ???
3333 : //
3334 : // The waiting for the cancellation is not done uniformly.
3335 : // We certainly wait for WAL receivers to shut down.
3336 : // That is necessary so that no new data comes in before the freeze_and_flush.
3337 : // But the tenant background loops are joined-on in our caller.
3338 : // It's mesed up.
3339 : // we just ignore the failure to stop
3340 :
3341 : // If we're still attaching, fire the cancellation token early to drop out: this
3342 : // will prevent us flushing, but ensures timely shutdown if some I/O during attach
3343 : // is very slow.
3344 6 : let shutdown_mode = if matches!(self.current_state(), TenantState::Attaching) {
3345 0 : self.cancel.cancel();
3346 0 :
3347 0 : // Having fired our cancellation token, do not try and flush timelines: their cancellation tokens
3348 0 : // are children of ours, so their flush loops will have shut down already
3349 0 : timeline::ShutdownMode::Hard
3350 : } else {
3351 6 : shutdown_mode
3352 : };
3353 :
3354 6 : match self.set_stopping(shutdown_progress, false, false).await {
3355 6 : Ok(()) => {}
3356 0 : Err(SetStoppingError::Broken) => {
3357 0 : // assume that this is acceptable
3358 0 : }
3359 0 : Err(SetStoppingError::AlreadyStopping(other)) => {
3360 0 : // give caller the option to wait for this this shutdown
3361 0 : info!("Tenant::shutdown: AlreadyStopping");
3362 0 : return Err(other);
3363 : }
3364 : };
3365 :
3366 6 : let mut js = tokio::task::JoinSet::new();
3367 6 : {
3368 6 : let timelines = self.timelines.lock().unwrap();
3369 6 : timelines.values().for_each(|timeline| {
3370 6 : let timeline = Arc::clone(timeline);
3371 6 : let timeline_id = timeline.timeline_id;
3372 6 : let span = tracing::info_span!("timeline_shutdown", %timeline_id, ?shutdown_mode);
3373 6 : js.spawn(async move { timeline.shutdown(shutdown_mode).instrument(span).await });
3374 6 : });
3375 6 : }
3376 6 : {
3377 6 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
3378 6 : timelines_offloaded.values().for_each(|timeline| {
3379 0 : timeline.defuse_for_tenant_drop();
3380 6 : });
3381 6 : }
3382 6 : // test_long_timeline_create_then_tenant_delete is leaning on this message
3383 6 : tracing::info!("Waiting for timelines...");
3384 12 : while let Some(res) = js.join_next().await {
3385 0 : match res {
3386 6 : Ok(()) => {}
3387 0 : Err(je) if je.is_cancelled() => unreachable!("no cancelling used"),
3388 0 : Err(je) if je.is_panic() => { /* logged already */ }
3389 0 : Err(je) => warn!("unexpected JoinError: {je:?}"),
3390 : }
3391 : }
3392 :
3393 6 : if let ShutdownMode::Reload = shutdown_mode {
3394 0 : tracing::info!("Flushing deletion queue");
3395 0 : if let Err(e) = self.deletion_queue_client.flush().await {
3396 0 : match e {
3397 0 : DeletionQueueError::ShuttingDown => {
3398 0 : // This is the only error we expect for now. In the future, if more error
3399 0 : // variants are added, we should handle them here.
3400 0 : }
3401 : }
3402 0 : }
3403 6 : }
3404 :
3405 : // We cancel the Tenant's cancellation token _after_ the timelines have all shut down. This permits
3406 : // them to continue to do work during their shutdown methods, e.g. flushing data.
3407 6 : tracing::debug!("Cancelling CancellationToken");
3408 6 : self.cancel.cancel();
3409 6 :
3410 6 : // shutdown all tenant and timeline tasks: gc, compaction, page service
3411 6 : // No new tasks will be started for this tenant because it's in `Stopping` state.
3412 6 : //
3413 6 : // this will additionally shutdown and await all timeline tasks.
3414 6 : tracing::debug!("Waiting for tasks...");
3415 6 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), None).await;
3416 :
3417 6 : if let Some(walredo_mgr) = self.walredo_mgr.as_ref() {
3418 6 : walredo_mgr.shutdown().await;
3419 0 : }
3420 :
3421 : // Wait for any in-flight operations to complete
3422 6 : self.gate.close().await;
3423 :
3424 6 : remove_tenant_metrics(&self.tenant_shard_id);
3425 6 :
3426 6 : Ok(())
3427 6 : }
3428 :
3429 : /// Change tenant status to Stopping, to mark that it is being shut down.
3430 : ///
3431 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
3432 : ///
3433 : /// This function is not cancel-safe!
3434 : ///
3435 : /// `allow_transition_from_loading` is needed for the special case of loading task deleting the tenant.
3436 : /// `allow_transition_from_attaching` is needed for the special case of attaching deleted tenant.
3437 6 : async fn set_stopping(
3438 6 : &self,
3439 6 : progress: completion::Barrier,
3440 6 : _allow_transition_from_loading: bool,
3441 6 : allow_transition_from_attaching: bool,
3442 6 : ) -> Result<(), SetStoppingError> {
3443 6 : let mut rx = self.state.subscribe();
3444 6 :
3445 6 : // cannot stop before we're done activating, so wait out until we're done activating
3446 6 : rx.wait_for(|state| match state {
3447 0 : TenantState::Attaching if allow_transition_from_attaching => true,
3448 : TenantState::Activating(_) | TenantState::Attaching => {
3449 0 : info!(
3450 0 : "waiting for {} to turn Active|Broken|Stopping",
3451 0 : <&'static str>::from(state)
3452 : );
3453 0 : false
3454 : }
3455 6 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3456 6 : })
3457 6 : .await
3458 6 : .expect("cannot drop self.state while on a &self method");
3459 6 :
3460 6 : // we now know we're done activating, let's see whether this task is the winner to transition into Stopping
3461 6 : let mut err = None;
3462 6 : let stopping = self.state.send_if_modified(|current_state| match current_state {
3463 : TenantState::Activating(_) => {
3464 0 : unreachable!("1we ensured above that we're done with activation, and, there is no re-activation")
3465 : }
3466 : TenantState::Attaching => {
3467 0 : if !allow_transition_from_attaching {
3468 0 : unreachable!("2we ensured above that we're done with activation, and, there is no re-activation")
3469 0 : };
3470 0 : *current_state = TenantState::Stopping { progress };
3471 0 : true
3472 : }
3473 : TenantState::Active => {
3474 : // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
3475 : // are created after the transition to Stopping. That's harmless, as the Timelines
3476 : // won't be accessible to anyone afterwards, because the Tenant is in Stopping state.
3477 6 : *current_state = TenantState::Stopping { progress };
3478 6 : // Continue stopping outside the closure. We need to grab timelines.lock()
3479 6 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
3480 6 : true
3481 : }
3482 0 : TenantState::Broken { reason, .. } => {
3483 0 : info!(
3484 0 : "Cannot set tenant to Stopping state, it is in Broken state due to: {reason}"
3485 : );
3486 0 : err = Some(SetStoppingError::Broken);
3487 0 : false
3488 : }
3489 0 : TenantState::Stopping { progress } => {
3490 0 : info!("Tenant is already in Stopping state");
3491 0 : err = Some(SetStoppingError::AlreadyStopping(progress.clone()));
3492 0 : false
3493 : }
3494 6 : });
3495 6 : match (stopping, err) {
3496 6 : (true, None) => {} // continue
3497 0 : (false, Some(err)) => return Err(err),
3498 0 : (true, Some(_)) => unreachable!(
3499 0 : "send_if_modified closure must error out if not transitioning to Stopping"
3500 0 : ),
3501 0 : (false, None) => unreachable!(
3502 0 : "send_if_modified closure must return true if transitioning to Stopping"
3503 0 : ),
3504 : }
3505 :
3506 6 : let timelines_accessor = self.timelines.lock().unwrap();
3507 6 : let not_broken_timelines = timelines_accessor
3508 6 : .values()
3509 6 : .filter(|timeline| !timeline.is_broken());
3510 12 : for timeline in not_broken_timelines {
3511 6 : timeline.set_state(TimelineState::Stopping);
3512 6 : }
3513 6 : Ok(())
3514 6 : }
3515 :
3516 : /// Method for tenant::mgr to transition us into Broken state in case of a late failure in
3517 : /// `remove_tenant_from_memory`
3518 : ///
3519 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
3520 : ///
3521 : /// In tests, we also use this to set tenants to Broken state on purpose.
3522 0 : pub(crate) async fn set_broken(&self, reason: String) {
3523 0 : let mut rx = self.state.subscribe();
3524 0 :
3525 0 : // The load & attach routines own the tenant state until it has reached `Active`.
3526 0 : // So, wait until it's done.
3527 0 : rx.wait_for(|state| match state {
3528 : TenantState::Activating(_) | TenantState::Attaching => {
3529 0 : info!(
3530 0 : "waiting for {} to turn Active|Broken|Stopping",
3531 0 : <&'static str>::from(state)
3532 : );
3533 0 : false
3534 : }
3535 0 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3536 0 : })
3537 0 : .await
3538 0 : .expect("cannot drop self.state while on a &self method");
3539 0 :
3540 0 : // we now know we're done activating, let's see whether this task is the winner to transition into Broken
3541 0 : self.set_broken_no_wait(reason)
3542 0 : }
3543 :
3544 0 : pub(crate) fn set_broken_no_wait(&self, reason: impl Display) {
3545 0 : let reason = reason.to_string();
3546 0 : self.state.send_modify(|current_state| {
3547 0 : match *current_state {
3548 : TenantState::Activating(_) | TenantState::Attaching => {
3549 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
3550 : }
3551 : TenantState::Active => {
3552 0 : if cfg!(feature = "testing") {
3553 0 : warn!("Changing Active tenant to Broken state, reason: {}", reason);
3554 0 : *current_state = TenantState::broken_from_reason(reason);
3555 : } else {
3556 0 : unreachable!("not allowed to call set_broken on Active tenants in non-testing builds")
3557 : }
3558 : }
3559 : TenantState::Broken { .. } => {
3560 0 : warn!("Tenant is already in Broken state");
3561 : }
3562 : // This is the only "expected" path, any other path is a bug.
3563 : TenantState::Stopping { .. } => {
3564 0 : warn!(
3565 0 : "Marking Stopping tenant as Broken state, reason: {}",
3566 : reason
3567 : );
3568 0 : *current_state = TenantState::broken_from_reason(reason);
3569 : }
3570 : }
3571 0 : });
3572 0 : }
3573 :
3574 0 : pub fn subscribe_for_state_updates(&self) -> watch::Receiver<TenantState> {
3575 0 : self.state.subscribe()
3576 0 : }
3577 :
3578 : /// The activate_now semaphore is initialized with zero units. As soon as
3579 : /// we add a unit, waiters will be able to acquire a unit and proceed.
3580 0 : pub(crate) fn activate_now(&self) {
3581 0 : self.activate_now_sem.add_permits(1);
3582 0 : }
3583 :
3584 0 : pub(crate) async fn wait_to_become_active(
3585 0 : &self,
3586 0 : timeout: Duration,
3587 0 : ) -> Result<(), GetActiveTenantError> {
3588 0 : let mut receiver = self.state.subscribe();
3589 : loop {
3590 0 : let current_state = receiver.borrow_and_update().clone();
3591 0 : match current_state {
3592 : TenantState::Attaching | TenantState::Activating(_) => {
3593 : // in these states, there's a chance that we can reach ::Active
3594 0 : self.activate_now();
3595 0 : match timeout_cancellable(timeout, &self.cancel, receiver.changed()).await {
3596 0 : Ok(r) => {
3597 0 : r.map_err(
3598 0 : |_e: tokio::sync::watch::error::RecvError|
3599 : // Tenant existed but was dropped: report it as non-existent
3600 0 : GetActiveTenantError::NotFound(GetTenantError::ShardNotFound(self.tenant_shard_id))
3601 0 : )?
3602 : }
3603 : Err(TimeoutCancellableError::Cancelled) => {
3604 0 : return Err(GetActiveTenantError::Cancelled);
3605 : }
3606 : Err(TimeoutCancellableError::Timeout) => {
3607 0 : return Err(GetActiveTenantError::WaitForActiveTimeout {
3608 0 : latest_state: Some(self.current_state()),
3609 0 : wait_time: timeout,
3610 0 : });
3611 : }
3612 : }
3613 : }
3614 : TenantState::Active { .. } => {
3615 0 : return Ok(());
3616 : }
3617 0 : TenantState::Broken { reason, .. } => {
3618 0 : // This is fatal, and reported distinctly from the general case of "will never be active" because
3619 0 : // it's logically a 500 to external API users (broken is always a bug).
3620 0 : return Err(GetActiveTenantError::Broken(reason));
3621 : }
3622 : TenantState::Stopping { .. } => {
3623 : // There's no chance the tenant can transition back into ::Active
3624 0 : return Err(GetActiveTenantError::WillNotBecomeActive(current_state));
3625 : }
3626 : }
3627 : }
3628 0 : }
3629 :
3630 0 : pub(crate) fn get_attach_mode(&self) -> AttachmentMode {
3631 0 : self.tenant_conf.load().location.attach_mode
3632 0 : }
3633 :
3634 : /// For API access: generate a LocationConfig equivalent to the one that would be used to
3635 : /// create a Tenant in the same state. Do not use this in hot paths: it's for relatively
3636 : /// rare external API calls, like a reconciliation at startup.
3637 0 : pub(crate) fn get_location_conf(&self) -> models::LocationConfig {
3638 0 : let conf = self.tenant_conf.load();
3639 :
3640 0 : let location_config_mode = match conf.location.attach_mode {
3641 0 : AttachmentMode::Single => models::LocationConfigMode::AttachedSingle,
3642 0 : AttachmentMode::Multi => models::LocationConfigMode::AttachedMulti,
3643 0 : AttachmentMode::Stale => models::LocationConfigMode::AttachedStale,
3644 : };
3645 :
3646 : // We have a pageserver TenantConf, we need the API-facing TenantConfig.
3647 0 : let tenant_config: models::TenantConfig = conf.tenant_conf.clone().into();
3648 0 :
3649 0 : models::LocationConfig {
3650 0 : mode: location_config_mode,
3651 0 : generation: self.generation.into(),
3652 0 : secondary_conf: None,
3653 0 : shard_number: self.shard_identity.number.0,
3654 0 : shard_count: self.shard_identity.count.literal(),
3655 0 : shard_stripe_size: self.shard_identity.stripe_size.0,
3656 0 : tenant_conf: tenant_config,
3657 0 : }
3658 0 : }
3659 :
3660 0 : pub(crate) fn get_tenant_shard_id(&self) -> &TenantShardId {
3661 0 : &self.tenant_shard_id
3662 0 : }
3663 :
3664 0 : pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
3665 0 : self.shard_identity.stripe_size
3666 0 : }
3667 :
3668 0 : pub(crate) fn get_generation(&self) -> Generation {
3669 0 : self.generation
3670 0 : }
3671 :
3672 : /// This function partially shuts down the tenant (it shuts down the Timelines) and is fallible,
3673 : /// and can leave the tenant in a bad state if it fails. The caller is responsible for
3674 : /// resetting this tenant to a valid state if we fail.
3675 0 : pub(crate) async fn split_prepare(
3676 0 : &self,
3677 0 : child_shards: &Vec<TenantShardId>,
3678 0 : ) -> anyhow::Result<()> {
3679 0 : let (timelines, offloaded) = {
3680 0 : let timelines = self.timelines.lock().unwrap();
3681 0 : let offloaded = self.timelines_offloaded.lock().unwrap();
3682 0 : (timelines.clone(), offloaded.clone())
3683 0 : };
3684 0 : let timelines_iter = timelines
3685 0 : .values()
3686 0 : .map(TimelineOrOffloadedArcRef::<'_>::from)
3687 0 : .chain(
3688 0 : offloaded
3689 0 : .values()
3690 0 : .map(TimelineOrOffloadedArcRef::<'_>::from),
3691 0 : );
3692 0 : for timeline in timelines_iter {
3693 : // We do not block timeline creation/deletion during splits inside the pageserver: it is up to higher levels
3694 : // to ensure that they do not start a split if currently in the process of doing these.
3695 :
3696 0 : let timeline_id = timeline.timeline_id();
3697 :
3698 0 : if let TimelineOrOffloadedArcRef::Timeline(timeline) = timeline {
3699 : // Upload an index from the parent: this is partly to provide freshness for the
3700 : // child tenants that will copy it, and partly for general ease-of-debugging: there will
3701 : // always be a parent shard index in the same generation as we wrote the child shard index.
3702 0 : tracing::info!(%timeline_id, "Uploading index");
3703 0 : timeline
3704 0 : .remote_client
3705 0 : .schedule_index_upload_for_file_changes()?;
3706 0 : timeline.remote_client.wait_completion().await?;
3707 0 : }
3708 :
3709 0 : let remote_client = match timeline {
3710 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.remote_client.clone(),
3711 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => {
3712 0 : let remote_client = self
3713 0 : .build_timeline_client(offloaded.timeline_id, self.remote_storage.clone());
3714 0 : Arc::new(remote_client)
3715 : }
3716 : };
3717 :
3718 : // Shut down the timeline's remote client: this means that the indices we write
3719 : // for child shards will not be invalidated by the parent shard deleting layers.
3720 0 : tracing::info!(%timeline_id, "Shutting down remote storage client");
3721 0 : remote_client.shutdown().await;
3722 :
3723 : // Download methods can still be used after shutdown, as they don't flow through the remote client's
3724 : // queue. In principal the RemoteTimelineClient could provide this without downloading it, but this
3725 : // operation is rare, so it's simpler to just download it (and robustly guarantees that the index
3726 : // we use here really is the remotely persistent one).
3727 0 : tracing::info!(%timeline_id, "Downloading index_part from parent");
3728 0 : let result = remote_client
3729 0 : .download_index_file(&self.cancel)
3730 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))
3731 0 : .await?;
3732 0 : let index_part = match result {
3733 : MaybeDeletedIndexPart::Deleted(_) => {
3734 0 : anyhow::bail!("Timeline deletion happened concurrently with split")
3735 : }
3736 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
3737 : };
3738 :
3739 0 : for child_shard in child_shards {
3740 0 : tracing::info!(%timeline_id, "Uploading index_part for child {}", child_shard.to_index());
3741 0 : upload_index_part(
3742 0 : &self.remote_storage,
3743 0 : child_shard,
3744 0 : &timeline_id,
3745 0 : self.generation,
3746 0 : &index_part,
3747 0 : &self.cancel,
3748 0 : )
3749 0 : .await?;
3750 : }
3751 : }
3752 :
3753 0 : let tenant_manifest = self.build_tenant_manifest();
3754 0 : for child_shard in child_shards {
3755 0 : tracing::info!(
3756 0 : "Uploading tenant manifest for child {}",
3757 0 : child_shard.to_index()
3758 : );
3759 0 : upload_tenant_manifest(
3760 0 : &self.remote_storage,
3761 0 : child_shard,
3762 0 : self.generation,
3763 0 : &tenant_manifest,
3764 0 : &self.cancel,
3765 0 : )
3766 0 : .await?;
3767 : }
3768 :
3769 0 : Ok(())
3770 0 : }
3771 :
3772 0 : pub(crate) fn get_sizes(&self) -> TopTenantShardItem {
3773 0 : let mut result = TopTenantShardItem {
3774 0 : id: self.tenant_shard_id,
3775 0 : resident_size: 0,
3776 0 : physical_size: 0,
3777 0 : max_logical_size: 0,
3778 0 : };
3779 :
3780 0 : for timeline in self.timelines.lock().unwrap().values() {
3781 0 : result.resident_size += timeline.metrics.resident_physical_size_gauge.get();
3782 0 :
3783 0 : result.physical_size += timeline
3784 0 : .remote_client
3785 0 : .metrics
3786 0 : .remote_physical_size_gauge
3787 0 : .get();
3788 0 : result.max_logical_size = std::cmp::max(
3789 0 : result.max_logical_size,
3790 0 : timeline.metrics.current_logical_size_gauge.get(),
3791 0 : );
3792 0 : }
3793 :
3794 0 : result
3795 0 : }
3796 : }
3797 :
3798 : /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
3799 : /// perform a topological sort, so that the parent of each timeline comes
3800 : /// before the children.
3801 : /// E extracts the ancestor from T
3802 : /// This allows for T to be different. It can be TimelineMetadata, can be Timeline itself, etc.
3803 196 : fn tree_sort_timelines<T, E>(
3804 196 : timelines: HashMap<TimelineId, T>,
3805 196 : extractor: E,
3806 196 : ) -> anyhow::Result<Vec<(TimelineId, T)>>
3807 196 : where
3808 196 : E: Fn(&T) -> Option<TimelineId>,
3809 196 : {
3810 196 : let mut result = Vec::with_capacity(timelines.len());
3811 196 :
3812 196 : let mut now = Vec::with_capacity(timelines.len());
3813 196 : // (ancestor, children)
3814 196 : let mut later: HashMap<TimelineId, Vec<(TimelineId, T)>> =
3815 196 : HashMap::with_capacity(timelines.len());
3816 :
3817 202 : for (timeline_id, value) in timelines {
3818 6 : if let Some(ancestor_id) = extractor(&value) {
3819 2 : let children = later.entry(ancestor_id).or_default();
3820 2 : children.push((timeline_id, value));
3821 4 : } else {
3822 4 : now.push((timeline_id, value));
3823 4 : }
3824 : }
3825 :
3826 202 : while let Some((timeline_id, metadata)) = now.pop() {
3827 6 : result.push((timeline_id, metadata));
3828 : // All children of this can be loaded now
3829 6 : if let Some(mut children) = later.remove(&timeline_id) {
3830 2 : now.append(&mut children);
3831 4 : }
3832 : }
3833 :
3834 : // All timelines should be visited now. Unless there were timelines with missing ancestors.
3835 196 : if !later.is_empty() {
3836 0 : for (missing_id, orphan_ids) in later {
3837 0 : for (orphan_id, _) in orphan_ids {
3838 0 : error!("could not load timeline {orphan_id} because its ancestor timeline {missing_id} could not be loaded");
3839 : }
3840 : }
3841 0 : bail!("could not load tenant because some timelines are missing ancestors");
3842 196 : }
3843 196 :
3844 196 : Ok(result)
3845 196 : }
3846 :
3847 : enum ActivateTimelineArgs {
3848 : Yes {
3849 : broker_client: storage_broker::BrokerClientChannel,
3850 : },
3851 : No,
3852 : }
3853 :
3854 : impl Tenant {
3855 0 : pub fn tenant_specific_overrides(&self) -> TenantConfOpt {
3856 0 : self.tenant_conf.load().tenant_conf.clone()
3857 0 : }
3858 :
3859 0 : pub fn effective_config(&self) -> TenantConf {
3860 0 : self.tenant_specific_overrides()
3861 0 : .merge(self.conf.default_tenant_conf.clone())
3862 0 : }
3863 :
3864 0 : pub fn get_checkpoint_distance(&self) -> u64 {
3865 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3866 0 : tenant_conf
3867 0 : .checkpoint_distance
3868 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
3869 0 : }
3870 :
3871 0 : pub fn get_checkpoint_timeout(&self) -> Duration {
3872 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3873 0 : tenant_conf
3874 0 : .checkpoint_timeout
3875 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
3876 0 : }
3877 :
3878 0 : pub fn get_compaction_target_size(&self) -> u64 {
3879 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3880 0 : tenant_conf
3881 0 : .compaction_target_size
3882 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
3883 0 : }
3884 :
3885 0 : pub fn get_compaction_period(&self) -> Duration {
3886 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3887 0 : tenant_conf
3888 0 : .compaction_period
3889 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
3890 0 : }
3891 :
3892 0 : pub fn get_compaction_threshold(&self) -> usize {
3893 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3894 0 : tenant_conf
3895 0 : .compaction_threshold
3896 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
3897 0 : }
3898 :
3899 0 : pub fn get_gc_horizon(&self) -> u64 {
3900 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3901 0 : tenant_conf
3902 0 : .gc_horizon
3903 0 : .unwrap_or(self.conf.default_tenant_conf.gc_horizon)
3904 0 : }
3905 :
3906 0 : pub fn get_gc_period(&self) -> Duration {
3907 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3908 0 : tenant_conf
3909 0 : .gc_period
3910 0 : .unwrap_or(self.conf.default_tenant_conf.gc_period)
3911 0 : }
3912 :
3913 0 : pub fn get_image_creation_threshold(&self) -> usize {
3914 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3915 0 : tenant_conf
3916 0 : .image_creation_threshold
3917 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
3918 0 : }
3919 :
3920 0 : pub fn get_pitr_interval(&self) -> Duration {
3921 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3922 0 : tenant_conf
3923 0 : .pitr_interval
3924 0 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
3925 0 : }
3926 :
3927 0 : pub fn get_min_resident_size_override(&self) -> Option<u64> {
3928 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3929 0 : tenant_conf
3930 0 : .min_resident_size_override
3931 0 : .or(self.conf.default_tenant_conf.min_resident_size_override)
3932 0 : }
3933 :
3934 0 : pub fn get_heatmap_period(&self) -> Option<Duration> {
3935 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3936 0 : let heatmap_period = tenant_conf
3937 0 : .heatmap_period
3938 0 : .unwrap_or(self.conf.default_tenant_conf.heatmap_period);
3939 0 : if heatmap_period.is_zero() {
3940 0 : None
3941 : } else {
3942 0 : Some(heatmap_period)
3943 : }
3944 0 : }
3945 :
3946 4 : pub fn get_lsn_lease_length(&self) -> Duration {
3947 4 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3948 4 : tenant_conf
3949 4 : .lsn_lease_length
3950 4 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
3951 4 : }
3952 :
3953 : /// Generate an up-to-date TenantManifest based on the state of this Tenant.
3954 2 : fn build_tenant_manifest(&self) -> TenantManifest {
3955 2 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
3956 2 :
3957 2 : let mut timeline_manifests = timelines_offloaded
3958 2 : .iter()
3959 2 : .map(|(_timeline_id, offloaded)| offloaded.manifest())
3960 2 : .collect::<Vec<_>>();
3961 2 : // Sort the manifests so that our output is deterministic
3962 2 : timeline_manifests.sort_by_key(|timeline_manifest| timeline_manifest.timeline_id);
3963 2 :
3964 2 : TenantManifest {
3965 2 : version: LATEST_TENANT_MANIFEST_VERSION,
3966 2 : offloaded_timelines: timeline_manifests,
3967 2 : }
3968 2 : }
3969 :
3970 0 : pub fn update_tenant_config<F: Fn(TenantConfOpt) -> anyhow::Result<TenantConfOpt>>(
3971 0 : &self,
3972 0 : update: F,
3973 0 : ) -> anyhow::Result<TenantConfOpt> {
3974 0 : // Use read-copy-update in order to avoid overwriting the location config
3975 0 : // state if this races with [`Tenant::set_new_location_config`]. Note that
3976 0 : // this race is not possible if both request types come from the storage
3977 0 : // controller (as they should!) because an exclusive op lock is required
3978 0 : // on the storage controller side.
3979 0 :
3980 0 : self.tenant_conf
3981 0 : .try_rcu(|attached_conf| -> Result<_, anyhow::Error> {
3982 0 : Ok(Arc::new(AttachedTenantConf {
3983 0 : tenant_conf: update(attached_conf.tenant_conf.clone())?,
3984 0 : location: attached_conf.location,
3985 0 : lsn_lease_deadline: attached_conf.lsn_lease_deadline,
3986 : }))
3987 0 : })?;
3988 :
3989 0 : let updated = self.tenant_conf.load();
3990 0 :
3991 0 : self.tenant_conf_updated(&updated.tenant_conf);
3992 0 : // Don't hold self.timelines.lock() during the notifies.
3993 0 : // There's no risk of deadlock right now, but there could be if we consolidate
3994 0 : // mutexes in struct Timeline in the future.
3995 0 : let timelines = self.list_timelines();
3996 0 : for timeline in timelines {
3997 0 : timeline.tenant_conf_updated(&updated);
3998 0 : }
3999 :
4000 0 : Ok(updated.tenant_conf.clone())
4001 0 : }
4002 :
4003 0 : pub(crate) fn set_new_location_config(&self, new_conf: AttachedTenantConf) {
4004 0 : let new_tenant_conf = new_conf.tenant_conf.clone();
4005 0 :
4006 0 : self.tenant_conf.store(Arc::new(new_conf.clone()));
4007 0 :
4008 0 : self.tenant_conf_updated(&new_tenant_conf);
4009 0 : // Don't hold self.timelines.lock() during the notifies.
4010 0 : // There's no risk of deadlock right now, but there could be if we consolidate
4011 0 : // mutexes in struct Timeline in the future.
4012 0 : let timelines = self.list_timelines();
4013 0 : for timeline in timelines {
4014 0 : timeline.tenant_conf_updated(&new_conf);
4015 0 : }
4016 0 : }
4017 :
4018 196 : fn get_pagestream_throttle_config(
4019 196 : psconf: &'static PageServerConf,
4020 196 : overrides: &TenantConfOpt,
4021 196 : ) -> throttle::Config {
4022 196 : overrides
4023 196 : .timeline_get_throttle
4024 196 : .clone()
4025 196 : .unwrap_or(psconf.default_tenant_conf.timeline_get_throttle.clone())
4026 196 : }
4027 :
4028 0 : pub(crate) fn tenant_conf_updated(&self, new_conf: &TenantConfOpt) {
4029 0 : let conf = Self::get_pagestream_throttle_config(self.conf, new_conf);
4030 0 : self.pagestream_throttle.reconfigure(conf)
4031 0 : }
4032 :
4033 : /// Helper function to create a new Timeline struct.
4034 : ///
4035 : /// The returned Timeline is in Loading state. The caller is responsible for
4036 : /// initializing any on-disk state, and for inserting the Timeline to the 'timelines'
4037 : /// map.
4038 : ///
4039 : /// `validate_ancestor == false` is used when a timeline is created for deletion
4040 : /// and we might not have the ancestor present anymore which is fine for to be
4041 : /// deleted timelines.
4042 : #[allow(clippy::too_many_arguments)]
4043 422 : fn create_timeline_struct(
4044 422 : &self,
4045 422 : new_timeline_id: TimelineId,
4046 422 : new_metadata: &TimelineMetadata,
4047 422 : ancestor: Option<Arc<Timeline>>,
4048 422 : resources: TimelineResources,
4049 422 : cause: CreateTimelineCause,
4050 422 : create_idempotency: CreateTimelineIdempotency,
4051 422 : ) -> anyhow::Result<Arc<Timeline>> {
4052 422 : let state = match cause {
4053 : CreateTimelineCause::Load => {
4054 422 : let ancestor_id = new_metadata.ancestor_timeline();
4055 422 : anyhow::ensure!(
4056 422 : ancestor_id == ancestor.as_ref().map(|t| t.timeline_id),
4057 0 : "Timeline's {new_timeline_id} ancestor {ancestor_id:?} was not found"
4058 : );
4059 422 : TimelineState::Loading
4060 : }
4061 0 : CreateTimelineCause::Delete => TimelineState::Stopping,
4062 : };
4063 :
4064 422 : let pg_version = new_metadata.pg_version();
4065 422 :
4066 422 : let timeline = Timeline::new(
4067 422 : self.conf,
4068 422 : Arc::clone(&self.tenant_conf),
4069 422 : new_metadata,
4070 422 : ancestor,
4071 422 : new_timeline_id,
4072 422 : self.tenant_shard_id,
4073 422 : self.generation,
4074 422 : self.shard_identity,
4075 422 : self.walredo_mgr.clone(),
4076 422 : resources,
4077 422 : pg_version,
4078 422 : state,
4079 422 : self.attach_wal_lag_cooldown.clone(),
4080 422 : create_idempotency,
4081 422 : self.cancel.child_token(),
4082 422 : );
4083 422 :
4084 422 : Ok(timeline)
4085 422 : }
4086 :
4087 : // Allow too_many_arguments because a constructor's argument list naturally grows with the
4088 : // number of attributes in the struct: breaking these out into a builder wouldn't be helpful.
4089 : #[allow(clippy::too_many_arguments)]
4090 196 : fn new(
4091 196 : state: TenantState,
4092 196 : conf: &'static PageServerConf,
4093 196 : attached_conf: AttachedTenantConf,
4094 196 : shard_identity: ShardIdentity,
4095 196 : walredo_mgr: Option<Arc<WalRedoManager>>,
4096 196 : tenant_shard_id: TenantShardId,
4097 196 : remote_storage: GenericRemoteStorage,
4098 196 : deletion_queue_client: DeletionQueueClient,
4099 196 : l0_flush_global_state: L0FlushGlobalState,
4100 196 : ) -> Tenant {
4101 196 : debug_assert!(
4102 196 : !attached_conf.location.generation.is_none() || conf.control_plane_api.is_none()
4103 : );
4104 :
4105 196 : let (state, mut rx) = watch::channel(state);
4106 196 :
4107 196 : tokio::spawn(async move {
4108 196 : // reflect tenant state in metrics:
4109 196 : // - global per tenant state: TENANT_STATE_METRIC
4110 196 : // - "set" of broken tenants: BROKEN_TENANTS_SET
4111 196 : //
4112 196 : // set of broken tenants should not have zero counts so that it remains accessible for
4113 196 : // alerting.
4114 196 :
4115 196 : let tid = tenant_shard_id.to_string();
4116 196 : let shard_id = tenant_shard_id.shard_slug().to_string();
4117 196 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
4118 :
4119 392 : fn inspect_state(state: &TenantState) -> ([&'static str; 1], bool) {
4120 392 : ([state.into()], matches!(state, TenantState::Broken { .. }))
4121 392 : }
4122 :
4123 196 : let mut tuple = inspect_state(&rx.borrow_and_update());
4124 196 :
4125 196 : let is_broken = tuple.1;
4126 196 : let mut counted_broken = if is_broken {
4127 : // add the id to the set right away, there should not be any updates on the channel
4128 : // after before tenant is removed, if ever
4129 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
4130 0 : true
4131 : } else {
4132 196 : false
4133 : };
4134 :
4135 : loop {
4136 392 : let labels = &tuple.0;
4137 392 : let current = TENANT_STATE_METRIC.with_label_values(labels);
4138 392 : current.inc();
4139 392 :
4140 392 : if rx.changed().await.is_err() {
4141 : // tenant has been dropped
4142 14 : current.dec();
4143 14 : drop(BROKEN_TENANTS_SET.remove_label_values(set_key));
4144 14 : break;
4145 196 : }
4146 196 :
4147 196 : current.dec();
4148 196 : tuple = inspect_state(&rx.borrow_and_update());
4149 196 :
4150 196 : let is_broken = tuple.1;
4151 196 : if is_broken && !counted_broken {
4152 0 : counted_broken = true;
4153 0 : // insert the tenant_id (back) into the set while avoiding needless counter
4154 0 : // access
4155 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
4156 196 : }
4157 : }
4158 196 : });
4159 196 :
4160 196 : Tenant {
4161 196 : tenant_shard_id,
4162 196 : shard_identity,
4163 196 : generation: attached_conf.location.generation,
4164 196 : conf,
4165 196 : // using now here is good enough approximation to catch tenants with really long
4166 196 : // activation times.
4167 196 : constructed_at: Instant::now(),
4168 196 : timelines: Mutex::new(HashMap::new()),
4169 196 : timelines_creating: Mutex::new(HashSet::new()),
4170 196 : timelines_offloaded: Mutex::new(HashMap::new()),
4171 196 : tenant_manifest_upload: Default::default(),
4172 196 : gc_cs: tokio::sync::Mutex::new(()),
4173 196 : walredo_mgr,
4174 196 : remote_storage,
4175 196 : deletion_queue_client,
4176 196 : state,
4177 196 : cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
4178 196 : cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
4179 196 : eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
4180 196 : compaction_circuit_breaker: std::sync::Mutex::new(CircuitBreaker::new(
4181 196 : format!("compaction-{tenant_shard_id}"),
4182 196 : 5,
4183 196 : // Compaction can be a very expensive operation, and might leak disk space. It also ought
4184 196 : // to be infallible, as long as remote storage is available. So if it repeatedly fails,
4185 196 : // use an extremely long backoff.
4186 196 : Some(Duration::from_secs(3600 * 24)),
4187 196 : )),
4188 196 : scheduled_compaction_tasks: Mutex::new(Default::default()),
4189 196 : activate_now_sem: tokio::sync::Semaphore::new(0),
4190 196 : attach_wal_lag_cooldown: Arc::new(std::sync::OnceLock::new()),
4191 196 : cancel: CancellationToken::default(),
4192 196 : gate: Gate::default(),
4193 196 : pagestream_throttle: Arc::new(throttle::Throttle::new(
4194 196 : Tenant::get_pagestream_throttle_config(conf, &attached_conf.tenant_conf),
4195 196 : crate::metrics::tenant_throttling::Metrics::new(&tenant_shard_id),
4196 196 : )),
4197 196 : tenant_conf: Arc::new(ArcSwap::from_pointee(attached_conf)),
4198 196 : ongoing_timeline_detach: std::sync::Mutex::default(),
4199 196 : gc_block: Default::default(),
4200 196 : l0_flush_global_state,
4201 196 : }
4202 196 : }
4203 :
4204 : /// Locate and load config
4205 0 : pub(super) fn load_tenant_config(
4206 0 : conf: &'static PageServerConf,
4207 0 : tenant_shard_id: &TenantShardId,
4208 0 : ) -> Result<LocationConf, LoadConfigError> {
4209 0 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4210 0 :
4211 0 : info!("loading tenant configuration from {config_path}");
4212 :
4213 : // load and parse file
4214 0 : let config = fs::read_to_string(&config_path).map_err(|e| {
4215 0 : match e.kind() {
4216 : std::io::ErrorKind::NotFound => {
4217 : // The config should almost always exist for a tenant directory:
4218 : // - When attaching a tenant, the config is the first thing we write
4219 : // - When detaching a tenant, we atomically move the directory to a tmp location
4220 : // before deleting contents.
4221 : //
4222 : // The very rare edge case that can result in a missing config is if we crash during attach
4223 : // between creating directory and writing config. Callers should handle that as if the
4224 : // directory didn't exist.
4225 :
4226 0 : LoadConfigError::NotFound(config_path)
4227 : }
4228 : _ => {
4229 : // No IO errors except NotFound are acceptable here: other kinds of error indicate local storage or permissions issues
4230 : // that we cannot cleanly recover
4231 0 : crate::virtual_file::on_fatal_io_error(&e, "Reading tenant config file")
4232 : }
4233 : }
4234 0 : })?;
4235 :
4236 0 : Ok(toml_edit::de::from_str::<LocationConf>(&config)?)
4237 0 : }
4238 :
4239 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4240 : pub(super) async fn persist_tenant_config(
4241 : conf: &'static PageServerConf,
4242 : tenant_shard_id: &TenantShardId,
4243 : location_conf: &LocationConf,
4244 : ) -> std::io::Result<()> {
4245 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4246 :
4247 : Self::persist_tenant_config_at(tenant_shard_id, &config_path, location_conf).await
4248 : }
4249 :
4250 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4251 : pub(super) async fn persist_tenant_config_at(
4252 : tenant_shard_id: &TenantShardId,
4253 : config_path: &Utf8Path,
4254 : location_conf: &LocationConf,
4255 : ) -> std::io::Result<()> {
4256 : debug!("persisting tenantconf to {config_path}");
4257 :
4258 : let mut conf_content = r#"# This file contains a specific per-tenant's config.
4259 : # It is read in case of pageserver restart.
4260 : "#
4261 : .to_string();
4262 :
4263 0 : fail::fail_point!("tenant-config-before-write", |_| {
4264 0 : Err(std::io::Error::new(
4265 0 : std::io::ErrorKind::Other,
4266 0 : "tenant-config-before-write",
4267 0 : ))
4268 0 : });
4269 :
4270 : // Convert the config to a toml file.
4271 : conf_content +=
4272 : &toml_edit::ser::to_string_pretty(&location_conf).expect("Config serialization failed");
4273 :
4274 : let temp_path = path_with_suffix_extension(config_path, TEMP_FILE_SUFFIX);
4275 :
4276 : let conf_content = conf_content.into_bytes();
4277 : VirtualFile::crashsafe_overwrite(config_path.to_owned(), temp_path, conf_content).await
4278 : }
4279 :
4280 : //
4281 : // How garbage collection works:
4282 : //
4283 : // +--bar------------->
4284 : // /
4285 : // +----+-----foo---------------->
4286 : // /
4287 : // ----main--+-------------------------->
4288 : // \
4289 : // +-----baz-------->
4290 : //
4291 : //
4292 : // 1. Grab 'gc_cs' mutex to prevent new timelines from being created while Timeline's
4293 : // `gc_infos` are being refreshed
4294 : // 2. Scan collected timelines, and on each timeline, make note of the
4295 : // all the points where other timelines have been branched off.
4296 : // We will refrain from removing page versions at those LSNs.
4297 : // 3. For each timeline, scan all layer files on the timeline.
4298 : // Remove all files for which a newer file exists and which
4299 : // don't cover any branch point LSNs.
4300 : //
4301 : // TODO:
4302 : // - if a relation has a non-incremental persistent layer on a child branch, then we
4303 : // don't need to keep that in the parent anymore. But currently
4304 : // we do.
4305 4 : async fn gc_iteration_internal(
4306 4 : &self,
4307 4 : target_timeline_id: Option<TimelineId>,
4308 4 : horizon: u64,
4309 4 : pitr: Duration,
4310 4 : cancel: &CancellationToken,
4311 4 : ctx: &RequestContext,
4312 4 : ) -> Result<GcResult, GcError> {
4313 4 : let mut totals: GcResult = Default::default();
4314 4 : let now = Instant::now();
4315 :
4316 4 : let gc_timelines = self
4317 4 : .refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4318 4 : .await?;
4319 :
4320 4 : failpoint_support::sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
4321 :
4322 : // If there is nothing to GC, we don't want any messages in the INFO log.
4323 4 : if !gc_timelines.is_empty() {
4324 4 : info!("{} timelines need GC", gc_timelines.len());
4325 : } else {
4326 0 : debug!("{} timelines need GC", gc_timelines.len());
4327 : }
4328 :
4329 : // Perform GC for each timeline.
4330 : //
4331 : // Note that we don't hold the `Tenant::gc_cs` lock here because we don't want to delay the
4332 : // branch creation task, which requires the GC lock. A GC iteration can run concurrently
4333 : // with branch creation.
4334 : //
4335 : // See comments in [`Tenant::branch_timeline`] for more information about why branch
4336 : // creation task can run concurrently with timeline's GC iteration.
4337 8 : for timeline in gc_timelines {
4338 4 : if cancel.is_cancelled() {
4339 : // We were requested to shut down. Stop and return with the progress we
4340 : // made.
4341 0 : break;
4342 4 : }
4343 4 : let result = match timeline.gc().await {
4344 : Err(GcError::TimelineCancelled) => {
4345 0 : if target_timeline_id.is_some() {
4346 : // If we were targetting this specific timeline, surface cancellation to caller
4347 0 : return Err(GcError::TimelineCancelled);
4348 : } else {
4349 : // A timeline may be shutting down independently of the tenant's lifecycle: we should
4350 : // skip past this and proceed to try GC on other timelines.
4351 0 : continue;
4352 : }
4353 : }
4354 4 : r => r?,
4355 : };
4356 4 : totals += result;
4357 : }
4358 :
4359 4 : totals.elapsed = now.elapsed();
4360 4 : Ok(totals)
4361 4 : }
4362 :
4363 : /// Refreshes the Timeline::gc_info for all timelines, returning the
4364 : /// vector of timelines which have [`Timeline::get_last_record_lsn`] past
4365 : /// [`Tenant::get_gc_horizon`].
4366 : ///
4367 : /// This is usually executed as part of periodic gc, but can now be triggered more often.
4368 0 : pub(crate) async fn refresh_gc_info(
4369 0 : &self,
4370 0 : cancel: &CancellationToken,
4371 0 : ctx: &RequestContext,
4372 0 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4373 0 : // since this method can now be called at different rates than the configured gc loop, it
4374 0 : // might be that these configuration values get applied faster than what it was previously,
4375 0 : // since these were only read from the gc task.
4376 0 : let horizon = self.get_gc_horizon();
4377 0 : let pitr = self.get_pitr_interval();
4378 0 :
4379 0 : // refresh all timelines
4380 0 : let target_timeline_id = None;
4381 0 :
4382 0 : self.refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4383 0 : .await
4384 0 : }
4385 :
4386 : /// Populate all Timelines' `GcInfo` with information about their children. We do not set the
4387 : /// PITR cutoffs here, because that requires I/O: this is done later, before GC, by [`Self::refresh_gc_info_internal`]
4388 : ///
4389 : /// Subsequently, parent-child relationships are updated incrementally inside [`Timeline::new`] and [`Timeline::drop`].
4390 0 : fn initialize_gc_info(
4391 0 : &self,
4392 0 : timelines: &std::sync::MutexGuard<HashMap<TimelineId, Arc<Timeline>>>,
4393 0 : timelines_offloaded: &std::sync::MutexGuard<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
4394 0 : restrict_to_timeline: Option<TimelineId>,
4395 0 : ) {
4396 0 : if restrict_to_timeline.is_none() {
4397 : // This function must be called before activation: after activation timeline create/delete operations
4398 : // might happen, and this function is not safe to run concurrently with those.
4399 0 : assert!(!self.is_active());
4400 0 : }
4401 :
4402 : // Scan all timelines. For each timeline, remember the timeline ID and
4403 : // the branch point where it was created.
4404 0 : let mut all_branchpoints: BTreeMap<TimelineId, Vec<(Lsn, TimelineId, MaybeOffloaded)>> =
4405 0 : BTreeMap::new();
4406 0 : timelines.iter().for_each(|(timeline_id, timeline_entry)| {
4407 0 : if let Some(ancestor_timeline_id) = &timeline_entry.get_ancestor_timeline_id() {
4408 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4409 0 : ancestor_children.push((
4410 0 : timeline_entry.get_ancestor_lsn(),
4411 0 : *timeline_id,
4412 0 : MaybeOffloaded::No,
4413 0 : ));
4414 0 : }
4415 0 : });
4416 0 : timelines_offloaded
4417 0 : .iter()
4418 0 : .for_each(|(timeline_id, timeline_entry)| {
4419 0 : let Some(ancestor_timeline_id) = &timeline_entry.ancestor_timeline_id else {
4420 0 : return;
4421 : };
4422 0 : let Some(retain_lsn) = timeline_entry.ancestor_retain_lsn else {
4423 0 : return;
4424 : };
4425 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4426 0 : ancestor_children.push((retain_lsn, *timeline_id, MaybeOffloaded::Yes));
4427 0 : });
4428 0 :
4429 0 : // The number of bytes we always keep, irrespective of PITR: this is a constant across timelines
4430 0 : let horizon = self.get_gc_horizon();
4431 :
4432 : // Populate each timeline's GcInfo with information about its child branches
4433 0 : let timelines_to_write = if let Some(timeline_id) = restrict_to_timeline {
4434 0 : itertools::Either::Left(timelines.get(&timeline_id).into_iter())
4435 : } else {
4436 0 : itertools::Either::Right(timelines.values())
4437 : };
4438 0 : for timeline in timelines_to_write {
4439 0 : let mut branchpoints: Vec<(Lsn, TimelineId, MaybeOffloaded)> = all_branchpoints
4440 0 : .remove(&timeline.timeline_id)
4441 0 : .unwrap_or_default();
4442 0 :
4443 0 : branchpoints.sort_by_key(|b| b.0);
4444 0 :
4445 0 : let mut target = timeline.gc_info.write().unwrap();
4446 0 :
4447 0 : target.retain_lsns = branchpoints;
4448 0 :
4449 0 : let space_cutoff = timeline
4450 0 : .get_last_record_lsn()
4451 0 : .checked_sub(horizon)
4452 0 : .unwrap_or(Lsn(0));
4453 0 :
4454 0 : target.cutoffs = GcCutoffs {
4455 0 : space: space_cutoff,
4456 0 : time: Lsn::INVALID,
4457 0 : };
4458 0 : }
4459 0 : }
4460 :
4461 4 : async fn refresh_gc_info_internal(
4462 4 : &self,
4463 4 : target_timeline_id: Option<TimelineId>,
4464 4 : horizon: u64,
4465 4 : pitr: Duration,
4466 4 : cancel: &CancellationToken,
4467 4 : ctx: &RequestContext,
4468 4 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4469 4 : // before taking the gc_cs lock, do the heavier weight finding of gc_cutoff points for
4470 4 : // currently visible timelines.
4471 4 : let timelines = self
4472 4 : .timelines
4473 4 : .lock()
4474 4 : .unwrap()
4475 4 : .values()
4476 4 : .filter(|tl| match target_timeline_id.as_ref() {
4477 4 : Some(target) => &tl.timeline_id == target,
4478 0 : None => true,
4479 4 : })
4480 4 : .cloned()
4481 4 : .collect::<Vec<_>>();
4482 4 :
4483 4 : if target_timeline_id.is_some() && timelines.is_empty() {
4484 : // We were to act on a particular timeline and it wasn't found
4485 0 : return Err(GcError::TimelineNotFound);
4486 4 : }
4487 4 :
4488 4 : let mut gc_cutoffs: HashMap<TimelineId, GcCutoffs> =
4489 4 : HashMap::with_capacity(timelines.len());
4490 4 :
4491 4 : // Ensures all timelines use the same start time when computing the time cutoff.
4492 4 : let now_ts_for_pitr_calc = SystemTime::now();
4493 4 : for timeline in timelines.iter() {
4494 4 : let cutoff = timeline
4495 4 : .get_last_record_lsn()
4496 4 : .checked_sub(horizon)
4497 4 : .unwrap_or(Lsn(0));
4498 :
4499 4 : let cutoffs = timeline
4500 4 : .find_gc_cutoffs(now_ts_for_pitr_calc, cutoff, pitr, cancel, ctx)
4501 4 : .await?;
4502 4 : let old = gc_cutoffs.insert(timeline.timeline_id, cutoffs);
4503 4 : assert!(old.is_none());
4504 : }
4505 :
4506 4 : if !self.is_active() || self.cancel.is_cancelled() {
4507 0 : return Err(GcError::TenantCancelled);
4508 4 : }
4509 :
4510 : // grab mutex to prevent new timelines from being created here; avoid doing long operations
4511 : // because that will stall branch creation.
4512 4 : let gc_cs = self.gc_cs.lock().await;
4513 :
4514 : // Ok, we now know all the branch points.
4515 : // Update the GC information for each timeline.
4516 4 : let mut gc_timelines = Vec::with_capacity(timelines.len());
4517 8 : for timeline in timelines {
4518 : // We filtered the timeline list above
4519 4 : if let Some(target_timeline_id) = target_timeline_id {
4520 4 : assert_eq!(target_timeline_id, timeline.timeline_id);
4521 0 : }
4522 :
4523 : {
4524 4 : let mut target = timeline.gc_info.write().unwrap();
4525 4 :
4526 4 : // Cull any expired leases
4527 4 : let now = SystemTime::now();
4528 6 : target.leases.retain(|_, lease| !lease.is_expired(&now));
4529 4 :
4530 4 : timeline
4531 4 : .metrics
4532 4 : .valid_lsn_lease_count_gauge
4533 4 : .set(target.leases.len() as u64);
4534 :
4535 : // Look up parent's PITR cutoff to update the child's knowledge of whether it is within parent's PITR
4536 4 : if let Some(ancestor_id) = timeline.get_ancestor_timeline_id() {
4537 0 : if let Some(ancestor_gc_cutoffs) = gc_cutoffs.get(&ancestor_id) {
4538 0 : target.within_ancestor_pitr =
4539 0 : timeline.get_ancestor_lsn() >= ancestor_gc_cutoffs.time;
4540 0 : }
4541 4 : }
4542 :
4543 : // Update metrics that depend on GC state
4544 4 : timeline
4545 4 : .metrics
4546 4 : .archival_size
4547 4 : .set(if target.within_ancestor_pitr {
4548 0 : timeline.metrics.current_logical_size_gauge.get()
4549 : } else {
4550 4 : 0
4551 : });
4552 4 : timeline.metrics.pitr_history_size.set(
4553 4 : timeline
4554 4 : .get_last_record_lsn()
4555 4 : .checked_sub(target.cutoffs.time)
4556 4 : .unwrap_or(Lsn(0))
4557 4 : .0,
4558 4 : );
4559 :
4560 : // Apply the cutoffs we found to the Timeline's GcInfo. Why might we _not_ have cutoffs for a timeline?
4561 : // - this timeline was created while we were finding cutoffs
4562 : // - lsn for timestamp search fails for this timeline repeatedly
4563 4 : if let Some(cutoffs) = gc_cutoffs.get(&timeline.timeline_id) {
4564 4 : let original_cutoffs = target.cutoffs.clone();
4565 4 : // GC cutoffs should never go back
4566 4 : target.cutoffs = GcCutoffs {
4567 4 : space: Lsn(cutoffs.space.0.max(original_cutoffs.space.0)),
4568 4 : time: Lsn(cutoffs.time.0.max(original_cutoffs.time.0)),
4569 4 : }
4570 0 : }
4571 : }
4572 :
4573 4 : gc_timelines.push(timeline);
4574 : }
4575 4 : drop(gc_cs);
4576 4 : Ok(gc_timelines)
4577 4 : }
4578 :
4579 : /// A substitute for `branch_timeline` for use in unit tests.
4580 : /// The returned timeline will have state value `Active` to make various `anyhow::ensure!()`
4581 : /// calls pass, but, we do not actually call `.activate()` under the hood. So, none of the
4582 : /// timeline background tasks are launched, except the flush loop.
4583 : #[cfg(test)]
4584 232 : async fn branch_timeline_test(
4585 232 : self: &Arc<Self>,
4586 232 : src_timeline: &Arc<Timeline>,
4587 232 : dst_id: TimelineId,
4588 232 : ancestor_lsn: Option<Lsn>,
4589 232 : ctx: &RequestContext,
4590 232 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
4591 232 : let tl = self
4592 232 : .branch_timeline_impl(src_timeline, dst_id, ancestor_lsn, ctx)
4593 232 : .await?
4594 228 : .into_timeline_for_test();
4595 228 : tl.set_state(TimelineState::Active);
4596 228 : Ok(tl)
4597 232 : }
4598 :
4599 : /// Helper for unit tests to branch a timeline with some pre-loaded states.
4600 : #[cfg(test)]
4601 : #[allow(clippy::too_many_arguments)]
4602 6 : pub async fn branch_timeline_test_with_layers(
4603 6 : self: &Arc<Self>,
4604 6 : src_timeline: &Arc<Timeline>,
4605 6 : dst_id: TimelineId,
4606 6 : ancestor_lsn: Option<Lsn>,
4607 6 : ctx: &RequestContext,
4608 6 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
4609 6 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
4610 6 : end_lsn: Lsn,
4611 6 : ) -> anyhow::Result<Arc<Timeline>> {
4612 : use checks::check_valid_layermap;
4613 : use itertools::Itertools;
4614 :
4615 6 : let tline = self
4616 6 : .branch_timeline_test(src_timeline, dst_id, ancestor_lsn, ctx)
4617 6 : .await?;
4618 6 : let ancestor_lsn = if let Some(ancestor_lsn) = ancestor_lsn {
4619 6 : ancestor_lsn
4620 : } else {
4621 0 : tline.get_last_record_lsn()
4622 : };
4623 6 : assert!(end_lsn >= ancestor_lsn);
4624 6 : tline.force_advance_lsn(end_lsn);
4625 12 : for deltas in delta_layer_desc {
4626 6 : tline
4627 6 : .force_create_delta_layer(deltas, Some(ancestor_lsn), ctx)
4628 6 : .await?;
4629 : }
4630 10 : for (lsn, images) in image_layer_desc {
4631 4 : tline
4632 4 : .force_create_image_layer(lsn, images, Some(ancestor_lsn), ctx)
4633 4 : .await?;
4634 : }
4635 6 : let layer_names = tline
4636 6 : .layers
4637 6 : .read()
4638 6 : .await
4639 6 : .layer_map()
4640 6 : .unwrap()
4641 6 : .iter_historic_layers()
4642 10 : .map(|layer| layer.layer_name())
4643 6 : .collect_vec();
4644 6 : if let Some(err) = check_valid_layermap(&layer_names) {
4645 0 : bail!("invalid layermap: {err}");
4646 6 : }
4647 6 : Ok(tline)
4648 6 : }
4649 :
4650 : /// Branch an existing timeline.
4651 0 : async fn branch_timeline(
4652 0 : self: &Arc<Self>,
4653 0 : src_timeline: &Arc<Timeline>,
4654 0 : dst_id: TimelineId,
4655 0 : start_lsn: Option<Lsn>,
4656 0 : ctx: &RequestContext,
4657 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4658 0 : self.branch_timeline_impl(src_timeline, dst_id, start_lsn, ctx)
4659 0 : .await
4660 0 : }
4661 :
4662 232 : async fn branch_timeline_impl(
4663 232 : self: &Arc<Self>,
4664 232 : src_timeline: &Arc<Timeline>,
4665 232 : dst_id: TimelineId,
4666 232 : start_lsn: Option<Lsn>,
4667 232 : _ctx: &RequestContext,
4668 232 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4669 232 : let src_id = src_timeline.timeline_id;
4670 :
4671 : // We will validate our ancestor LSN in this function. Acquire the GC lock so that
4672 : // this check cannot race with GC, and the ancestor LSN is guaranteed to remain
4673 : // valid while we are creating the branch.
4674 232 : let _gc_cs = self.gc_cs.lock().await;
4675 :
4676 : // If no start LSN is specified, we branch the new timeline from the source timeline's last record LSN
4677 232 : let start_lsn = start_lsn.unwrap_or_else(|| {
4678 2 : let lsn = src_timeline.get_last_record_lsn();
4679 2 : info!("branching timeline {dst_id} from timeline {src_id} at last record LSN: {lsn}");
4680 2 : lsn
4681 232 : });
4682 :
4683 : // we finally have determined the ancestor_start_lsn, so we can get claim exclusivity now
4684 232 : let timeline_create_guard = match self
4685 232 : .start_creating_timeline(
4686 232 : dst_id,
4687 232 : CreateTimelineIdempotency::Branch {
4688 232 : ancestor_timeline_id: src_timeline.timeline_id,
4689 232 : ancestor_start_lsn: start_lsn,
4690 232 : },
4691 232 : )
4692 232 : .await?
4693 : {
4694 232 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
4695 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
4696 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
4697 : }
4698 : };
4699 :
4700 : // Ensure that `start_lsn` is valid, i.e. the LSN is within the PITR
4701 : // horizon on the source timeline
4702 : //
4703 : // We check it against both the planned GC cutoff stored in 'gc_info',
4704 : // and the 'latest_gc_cutoff' of the last GC that was performed. The
4705 : // planned GC cutoff in 'gc_info' is normally larger than
4706 : // 'latest_gc_cutoff_lsn', but beware of corner cases like if you just
4707 : // changed the GC settings for the tenant to make the PITR window
4708 : // larger, but some of the data was already removed by an earlier GC
4709 : // iteration.
4710 :
4711 : // check against last actual 'latest_gc_cutoff' first
4712 232 : let latest_gc_cutoff_lsn = src_timeline.get_latest_gc_cutoff_lsn();
4713 232 : src_timeline
4714 232 : .check_lsn_is_in_scope(start_lsn, &latest_gc_cutoff_lsn)
4715 232 : .context(format!(
4716 232 : "invalid branch start lsn: less than latest GC cutoff {}",
4717 232 : *latest_gc_cutoff_lsn,
4718 232 : ))
4719 232 : .map_err(CreateTimelineError::AncestorLsn)?;
4720 :
4721 : // and then the planned GC cutoff
4722 : {
4723 228 : let gc_info = src_timeline.gc_info.read().unwrap();
4724 228 : let cutoff = gc_info.min_cutoff();
4725 228 : if start_lsn < cutoff {
4726 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
4727 0 : "invalid branch start lsn: less than planned GC cutoff {cutoff}"
4728 0 : )));
4729 228 : }
4730 228 : }
4731 228 :
4732 228 : //
4733 228 : // The branch point is valid, and we are still holding the 'gc_cs' lock
4734 228 : // so that GC cannot advance the GC cutoff until we are finished.
4735 228 : // Proceed with the branch creation.
4736 228 : //
4737 228 :
4738 228 : // Determine prev-LSN for the new timeline. We can only determine it if
4739 228 : // the timeline was branched at the current end of the source timeline.
4740 228 : let RecordLsn {
4741 228 : last: src_last,
4742 228 : prev: src_prev,
4743 228 : } = src_timeline.get_last_record_rlsn();
4744 228 : let dst_prev = if src_last == start_lsn {
4745 216 : Some(src_prev)
4746 : } else {
4747 12 : None
4748 : };
4749 :
4750 : // Create the metadata file, noting the ancestor of the new timeline.
4751 : // There is initially no data in it, but all the read-calls know to look
4752 : // into the ancestor.
4753 228 : let metadata = TimelineMetadata::new(
4754 228 : start_lsn,
4755 228 : dst_prev,
4756 228 : Some(src_id),
4757 228 : start_lsn,
4758 228 : *src_timeline.latest_gc_cutoff_lsn.read(), // FIXME: should we hold onto this guard longer?
4759 228 : src_timeline.initdb_lsn,
4760 228 : src_timeline.pg_version,
4761 228 : );
4762 :
4763 228 : let uninitialized_timeline = self
4764 228 : .prepare_new_timeline(
4765 228 : dst_id,
4766 228 : &metadata,
4767 228 : timeline_create_guard,
4768 228 : start_lsn + 1,
4769 228 : Some(Arc::clone(src_timeline)),
4770 228 : )
4771 228 : .await?;
4772 :
4773 228 : let new_timeline = uninitialized_timeline.finish_creation()?;
4774 :
4775 : // Root timeline gets its layers during creation and uploads them along with the metadata.
4776 : // A branch timeline though, when created, can get no writes for some time, hence won't get any layers created.
4777 : // We still need to upload its metadata eagerly: if other nodes `attach` the tenant and miss this timeline, their GC
4778 : // could get incorrect information and remove more layers, than needed.
4779 : // See also https://github.com/neondatabase/neon/issues/3865
4780 228 : new_timeline
4781 228 : .remote_client
4782 228 : .schedule_index_upload_for_full_metadata_update(&metadata)
4783 228 : .context("branch initial metadata upload")?;
4784 :
4785 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
4786 :
4787 228 : Ok(CreateTimelineResult::Created(new_timeline))
4788 232 : }
4789 :
4790 : /// For unit tests, make this visible so that other modules can directly create timelines
4791 : #[cfg(test)]
4792 2 : #[tracing::instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
4793 : pub(crate) async fn bootstrap_timeline_test(
4794 : self: &Arc<Self>,
4795 : timeline_id: TimelineId,
4796 : pg_version: u32,
4797 : load_existing_initdb: Option<TimelineId>,
4798 : ctx: &RequestContext,
4799 : ) -> anyhow::Result<Arc<Timeline>> {
4800 : self.bootstrap_timeline(timeline_id, pg_version, load_existing_initdb, ctx)
4801 : .await
4802 : .map_err(anyhow::Error::new)
4803 2 : .map(|r| r.into_timeline_for_test())
4804 : }
4805 :
4806 : /// Get exclusive access to the timeline ID for creation.
4807 : ///
4808 : /// Timeline-creating code paths must use this function before making changes
4809 : /// to in-memory or persistent state.
4810 : ///
4811 : /// The `state` parameter is a description of the timeline creation operation
4812 : /// we intend to perform.
4813 : /// If the timeline was already created in the meantime, we check whether this
4814 : /// request conflicts or is idempotent , based on `state`.
4815 422 : async fn start_creating_timeline(
4816 422 : self: &Arc<Self>,
4817 422 : new_timeline_id: TimelineId,
4818 422 : idempotency: CreateTimelineIdempotency,
4819 422 : ) -> Result<StartCreatingTimelineResult, CreateTimelineError> {
4820 422 : let allow_offloaded = false;
4821 422 : match self.create_timeline_create_guard(new_timeline_id, idempotency, allow_offloaded) {
4822 420 : Ok(create_guard) => {
4823 420 : pausable_failpoint!("timeline-creation-after-uninit");
4824 420 : Ok(StartCreatingTimelineResult::CreateGuard(create_guard))
4825 : }
4826 0 : Err(TimelineExclusionError::ShuttingDown) => Err(CreateTimelineError::ShuttingDown),
4827 : Err(TimelineExclusionError::AlreadyCreating) => {
4828 : // Creation is in progress, we cannot create it again, and we cannot
4829 : // check if this request matches the existing one, so caller must try
4830 : // again later.
4831 0 : Err(CreateTimelineError::AlreadyCreating)
4832 : }
4833 0 : Err(TimelineExclusionError::Other(e)) => Err(CreateTimelineError::Other(e)),
4834 : Err(TimelineExclusionError::AlreadyExists {
4835 0 : existing: TimelineOrOffloaded::Offloaded(_existing),
4836 0 : ..
4837 0 : }) => {
4838 0 : info!("timeline already exists but is offloaded");
4839 0 : Err(CreateTimelineError::Conflict)
4840 : }
4841 : Err(TimelineExclusionError::AlreadyExists {
4842 2 : existing: TimelineOrOffloaded::Timeline(existing),
4843 2 : arg,
4844 2 : }) => {
4845 2 : {
4846 2 : let existing = &existing.create_idempotency;
4847 2 : let _span = info_span!("idempotency_check", ?existing, ?arg).entered();
4848 2 : debug!("timeline already exists");
4849 :
4850 2 : match (existing, &arg) {
4851 : // FailWithConflict => no idempotency check
4852 : (CreateTimelineIdempotency::FailWithConflict, _)
4853 : | (_, CreateTimelineIdempotency::FailWithConflict) => {
4854 2 : warn!("timeline already exists, failing request");
4855 2 : return Err(CreateTimelineError::Conflict);
4856 : }
4857 : // Idempotent <=> CreateTimelineIdempotency is identical
4858 0 : (x, y) if x == y => {
4859 0 : info!("timeline already exists and idempotency matches, succeeding request");
4860 : // fallthrough
4861 : }
4862 : (_, _) => {
4863 0 : warn!("idempotency conflict, failing request");
4864 0 : return Err(CreateTimelineError::Conflict);
4865 : }
4866 : }
4867 : }
4868 :
4869 0 : Ok(StartCreatingTimelineResult::Idempotent(existing))
4870 : }
4871 : }
4872 422 : }
4873 :
4874 0 : async fn upload_initdb(
4875 0 : &self,
4876 0 : timelines_path: &Utf8PathBuf,
4877 0 : pgdata_path: &Utf8PathBuf,
4878 0 : timeline_id: &TimelineId,
4879 0 : ) -> anyhow::Result<()> {
4880 0 : let temp_path = timelines_path.join(format!(
4881 0 : "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
4882 0 : ));
4883 0 :
4884 0 : scopeguard::defer! {
4885 0 : if let Err(e) = fs::remove_file(&temp_path) {
4886 0 : error!("Failed to remove temporary initdb archive '{temp_path}': {e}");
4887 0 : }
4888 0 : }
4889 :
4890 0 : let (pgdata_zstd, tar_zst_size) = create_zst_tarball(pgdata_path, &temp_path).await?;
4891 : const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
4892 0 : if tar_zst_size > INITDB_TAR_ZST_WARN_LIMIT {
4893 0 : warn!(
4894 0 : "compressed {temp_path} size of {tar_zst_size} is above limit {INITDB_TAR_ZST_WARN_LIMIT}."
4895 : );
4896 0 : }
4897 :
4898 0 : pausable_failpoint!("before-initdb-upload");
4899 :
4900 0 : backoff::retry(
4901 0 : || async {
4902 0 : self::remote_timeline_client::upload_initdb_dir(
4903 0 : &self.remote_storage,
4904 0 : &self.tenant_shard_id.tenant_id,
4905 0 : timeline_id,
4906 0 : pgdata_zstd.try_clone().await?,
4907 0 : tar_zst_size,
4908 0 : &self.cancel,
4909 0 : )
4910 0 : .await
4911 0 : },
4912 0 : |_| false,
4913 0 : 3,
4914 0 : u32::MAX,
4915 0 : "persist_initdb_tar_zst",
4916 0 : &self.cancel,
4917 0 : )
4918 0 : .await
4919 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
4920 0 : .and_then(|x| x)
4921 0 : }
4922 :
4923 : /// - run initdb to init temporary instance and get bootstrap data
4924 : /// - after initialization completes, tar up the temp dir and upload it to S3.
4925 2 : async fn bootstrap_timeline(
4926 2 : self: &Arc<Self>,
4927 2 : timeline_id: TimelineId,
4928 2 : pg_version: u32,
4929 2 : load_existing_initdb: Option<TimelineId>,
4930 2 : ctx: &RequestContext,
4931 2 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4932 2 : let timeline_create_guard = match self
4933 2 : .start_creating_timeline(
4934 2 : timeline_id,
4935 2 : CreateTimelineIdempotency::Bootstrap { pg_version },
4936 2 : )
4937 2 : .await?
4938 : {
4939 2 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
4940 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
4941 0 : return Ok(CreateTimelineResult::Idempotent(timeline))
4942 : }
4943 : };
4944 :
4945 : // create a `tenant/{tenant_id}/timelines/basebackup-{timeline_id}.{TEMP_FILE_SUFFIX}/`
4946 : // temporary directory for basebackup files for the given timeline.
4947 :
4948 2 : let timelines_path = self.conf.timelines_path(&self.tenant_shard_id);
4949 2 : let pgdata_path = path_with_suffix_extension(
4950 2 : timelines_path.join(format!("basebackup-{timeline_id}")),
4951 2 : TEMP_FILE_SUFFIX,
4952 2 : );
4953 2 :
4954 2 : // Remove whatever was left from the previous runs: safe because TimelineCreateGuard guarantees
4955 2 : // we won't race with other creations or existent timelines with the same path.
4956 2 : if pgdata_path.exists() {
4957 0 : fs::remove_dir_all(&pgdata_path).with_context(|| {
4958 0 : format!("Failed to remove already existing initdb directory: {pgdata_path}")
4959 0 : })?;
4960 2 : }
4961 :
4962 : // this new directory is very temporary, set to remove it immediately after bootstrap, we don't need it
4963 2 : scopeguard::defer! {
4964 2 : if let Err(e) = fs::remove_dir_all(&pgdata_path) {
4965 2 : // this is unlikely, but we will remove the directory on pageserver restart or another bootstrap call
4966 2 : error!("Failed to remove temporary initdb directory '{pgdata_path}': {e}");
4967 2 : }
4968 2 : }
4969 2 : if let Some(existing_initdb_timeline_id) = load_existing_initdb {
4970 2 : if existing_initdb_timeline_id != timeline_id {
4971 0 : let source_path = &remote_initdb_archive_path(
4972 0 : &self.tenant_shard_id.tenant_id,
4973 0 : &existing_initdb_timeline_id,
4974 0 : );
4975 0 : let dest_path =
4976 0 : &remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &timeline_id);
4977 0 :
4978 0 : // if this fails, it will get retried by retried control plane requests
4979 0 : self.remote_storage
4980 0 : .copy_object(source_path, dest_path, &self.cancel)
4981 0 : .await
4982 0 : .context("copy initdb tar")?;
4983 2 : }
4984 2 : let (initdb_tar_zst_path, initdb_tar_zst) =
4985 2 : self::remote_timeline_client::download_initdb_tar_zst(
4986 2 : self.conf,
4987 2 : &self.remote_storage,
4988 2 : &self.tenant_shard_id,
4989 2 : &existing_initdb_timeline_id,
4990 2 : &self.cancel,
4991 2 : )
4992 2 : .await
4993 2 : .context("download initdb tar")?;
4994 :
4995 2 : scopeguard::defer! {
4996 2 : if let Err(e) = fs::remove_file(&initdb_tar_zst_path) {
4997 2 : error!("Failed to remove temporary initdb archive '{initdb_tar_zst_path}': {e}");
4998 2 : }
4999 2 : }
5000 2 :
5001 2 : let buf_read =
5002 2 : BufReader::with_capacity(remote_timeline_client::BUFFER_SIZE, initdb_tar_zst);
5003 2 : extract_zst_tarball(&pgdata_path, buf_read)
5004 2 : .await
5005 2 : .context("extract initdb tar")?;
5006 : } else {
5007 : // Init temporarily repo to get bootstrap data, this creates a directory in the `pgdata_path` path
5008 0 : run_initdb(self.conf, &pgdata_path, pg_version, &self.cancel)
5009 0 : .await
5010 0 : .context("run initdb")?;
5011 :
5012 : // Upload the created data dir to S3
5013 0 : if self.tenant_shard_id().is_shard_zero() {
5014 0 : self.upload_initdb(&timelines_path, &pgdata_path, &timeline_id)
5015 0 : .await?;
5016 0 : }
5017 : }
5018 2 : let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
5019 2 :
5020 2 : // Import the contents of the data directory at the initial checkpoint
5021 2 : // LSN, and any WAL after that.
5022 2 : // Initdb lsn will be equal to last_record_lsn which will be set after import.
5023 2 : // Because we know it upfront avoid having an option or dummy zero value by passing it to the metadata.
5024 2 : let new_metadata = TimelineMetadata::new(
5025 2 : Lsn(0),
5026 2 : None,
5027 2 : None,
5028 2 : Lsn(0),
5029 2 : pgdata_lsn,
5030 2 : pgdata_lsn,
5031 2 : pg_version,
5032 2 : );
5033 2 : let raw_timeline = self
5034 2 : .prepare_new_timeline(
5035 2 : timeline_id,
5036 2 : &new_metadata,
5037 2 : timeline_create_guard,
5038 2 : pgdata_lsn,
5039 2 : None,
5040 2 : )
5041 2 : .await?;
5042 :
5043 2 : let tenant_shard_id = raw_timeline.owning_tenant.tenant_shard_id;
5044 2 : let unfinished_timeline = raw_timeline.raw_timeline()?;
5045 :
5046 : // Flush the new layer files to disk, before we make the timeline as available to
5047 : // the outside world.
5048 : //
5049 : // Flush loop needs to be spawned in order to be able to flush.
5050 2 : unfinished_timeline.maybe_spawn_flush_loop();
5051 2 :
5052 2 : import_datadir::import_timeline_from_postgres_datadir(
5053 2 : unfinished_timeline,
5054 2 : &pgdata_path,
5055 2 : pgdata_lsn,
5056 2 : ctx,
5057 2 : )
5058 2 : .await
5059 2 : .with_context(|| {
5060 0 : format!("Failed to import pgdatadir for timeline {tenant_shard_id}/{timeline_id}")
5061 2 : })?;
5062 :
5063 2 : fail::fail_point!("before-checkpoint-new-timeline", |_| {
5064 0 : Err(CreateTimelineError::Other(anyhow::anyhow!(
5065 0 : "failpoint before-checkpoint-new-timeline"
5066 0 : )))
5067 2 : });
5068 :
5069 2 : unfinished_timeline
5070 2 : .freeze_and_flush()
5071 2 : .await
5072 2 : .with_context(|| {
5073 0 : format!(
5074 0 : "Failed to flush after pgdatadir import for timeline {tenant_shard_id}/{timeline_id}"
5075 0 : )
5076 2 : })?;
5077 :
5078 : // All done!
5079 2 : let timeline = raw_timeline.finish_creation()?;
5080 :
5081 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
5082 :
5083 2 : Ok(CreateTimelineResult::Created(timeline))
5084 2 : }
5085 :
5086 416 : fn build_timeline_remote_client(&self, timeline_id: TimelineId) -> RemoteTimelineClient {
5087 416 : RemoteTimelineClient::new(
5088 416 : self.remote_storage.clone(),
5089 416 : self.deletion_queue_client.clone(),
5090 416 : self.conf,
5091 416 : self.tenant_shard_id,
5092 416 : timeline_id,
5093 416 : self.generation,
5094 416 : &self.tenant_conf.load().location,
5095 416 : )
5096 416 : }
5097 :
5098 : /// Call this before constructing a timeline, to build its required structures
5099 416 : fn build_timeline_resources(&self, timeline_id: TimelineId) -> TimelineResources {
5100 416 : TimelineResources {
5101 416 : remote_client: self.build_timeline_remote_client(timeline_id),
5102 416 : pagestream_throttle: self.pagestream_throttle.clone(),
5103 416 : l0_flush_global_state: self.l0_flush_global_state.clone(),
5104 416 : }
5105 416 : }
5106 :
5107 : /// Creates intermediate timeline structure and its files.
5108 : ///
5109 : /// An empty layer map is initialized, and new data and WAL can be imported starting
5110 : /// at 'disk_consistent_lsn'. After any initial data has been imported, call
5111 : /// `finish_creation` to insert the Timeline into the timelines map.
5112 416 : async fn prepare_new_timeline<'a>(
5113 416 : &'a self,
5114 416 : new_timeline_id: TimelineId,
5115 416 : new_metadata: &TimelineMetadata,
5116 416 : create_guard: TimelineCreateGuard,
5117 416 : start_lsn: Lsn,
5118 416 : ancestor: Option<Arc<Timeline>>,
5119 416 : ) -> anyhow::Result<UninitializedTimeline<'a>> {
5120 416 : let tenant_shard_id = self.tenant_shard_id;
5121 416 :
5122 416 : let resources = self.build_timeline_resources(new_timeline_id);
5123 416 : resources
5124 416 : .remote_client
5125 416 : .init_upload_queue_for_empty_remote(new_metadata)?;
5126 :
5127 416 : let timeline_struct = self
5128 416 : .create_timeline_struct(
5129 416 : new_timeline_id,
5130 416 : new_metadata,
5131 416 : ancestor,
5132 416 : resources,
5133 416 : CreateTimelineCause::Load,
5134 416 : create_guard.idempotency.clone(),
5135 416 : )
5136 416 : .context("Failed to create timeline data structure")?;
5137 :
5138 416 : timeline_struct.init_empty_layer_map(start_lsn);
5139 :
5140 416 : if let Err(e) = self
5141 416 : .create_timeline_files(&create_guard.timeline_path)
5142 416 : .await
5143 : {
5144 0 : error!("Failed to create initial files for timeline {tenant_shard_id}/{new_timeline_id}, cleaning up: {e:?}");
5145 0 : cleanup_timeline_directory(create_guard);
5146 0 : return Err(e);
5147 416 : }
5148 416 :
5149 416 : debug!(
5150 0 : "Successfully created initial files for timeline {tenant_shard_id}/{new_timeline_id}"
5151 : );
5152 :
5153 416 : Ok(UninitializedTimeline::new(
5154 416 : self,
5155 416 : new_timeline_id,
5156 416 : Some((timeline_struct, create_guard)),
5157 416 : ))
5158 416 : }
5159 :
5160 416 : async fn create_timeline_files(&self, timeline_path: &Utf8Path) -> anyhow::Result<()> {
5161 416 : crashsafe::create_dir(timeline_path).context("Failed to create timeline directory")?;
5162 :
5163 416 : fail::fail_point!("after-timeline-dir-creation", |_| {
5164 0 : anyhow::bail!("failpoint after-timeline-dir-creation");
5165 416 : });
5166 :
5167 416 : Ok(())
5168 416 : }
5169 :
5170 : /// Get a guard that provides exclusive access to the timeline directory, preventing
5171 : /// concurrent attempts to create the same timeline.
5172 : ///
5173 : /// The `allow_offloaded` parameter controls whether to tolerate the existence of
5174 : /// offloaded timelines or not.
5175 422 : fn create_timeline_create_guard(
5176 422 : self: &Arc<Self>,
5177 422 : timeline_id: TimelineId,
5178 422 : idempotency: CreateTimelineIdempotency,
5179 422 : allow_offloaded: bool,
5180 422 : ) -> Result<TimelineCreateGuard, TimelineExclusionError> {
5181 422 : let tenant_shard_id = self.tenant_shard_id;
5182 422 :
5183 422 : let timeline_path = self.conf.timeline_path(&tenant_shard_id, &timeline_id);
5184 :
5185 422 : let create_guard = TimelineCreateGuard::new(
5186 422 : self,
5187 422 : timeline_id,
5188 422 : timeline_path.clone(),
5189 422 : idempotency,
5190 422 : allow_offloaded,
5191 422 : )?;
5192 :
5193 : // At this stage, we have got exclusive access to in-memory state for this timeline ID
5194 : // for creation.
5195 : // A timeline directory should never exist on disk already:
5196 : // - a previous failed creation would have cleaned up after itself
5197 : // - a pageserver restart would clean up timeline directories that don't have valid remote state
5198 : //
5199 : // Therefore it is an unexpected internal error to encounter a timeline directory already existing here,
5200 : // this error may indicate a bug in cleanup on failed creations.
5201 420 : if timeline_path.exists() {
5202 0 : return Err(TimelineExclusionError::Other(anyhow::anyhow!(
5203 0 : "Timeline directory already exists! This is a bug."
5204 0 : )));
5205 420 : }
5206 420 :
5207 420 : Ok(create_guard)
5208 422 : }
5209 :
5210 : /// Gathers inputs from all of the timelines to produce a sizing model input.
5211 : ///
5212 : /// Future is cancellation safe. Only one calculation can be running at once per tenant.
5213 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5214 : pub async fn gather_size_inputs(
5215 : &self,
5216 : // `max_retention_period` overrides the cutoff that is used to calculate the size
5217 : // (only if it is shorter than the real cutoff).
5218 : max_retention_period: Option<u64>,
5219 : cause: LogicalSizeCalculationCause,
5220 : cancel: &CancellationToken,
5221 : ctx: &RequestContext,
5222 : ) -> Result<size::ModelInputs, size::CalculateSyntheticSizeError> {
5223 : let logical_sizes_at_once = self
5224 : .conf
5225 : .concurrent_tenant_size_logical_size_queries
5226 : .inner();
5227 :
5228 : // TODO: Having a single mutex block concurrent reads is not great for performance.
5229 : //
5230 : // But the only case where we need to run multiple of these at once is when we
5231 : // request a size for a tenant manually via API, while another background calculation
5232 : // is in progress (which is not a common case).
5233 : //
5234 : // See more for on the issue #2748 condenced out of the initial PR review.
5235 : let mut shared_cache = tokio::select! {
5236 : locked = self.cached_logical_sizes.lock() => locked,
5237 : _ = cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5238 : _ = self.cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5239 : };
5240 :
5241 : size::gather_inputs(
5242 : self,
5243 : logical_sizes_at_once,
5244 : max_retention_period,
5245 : &mut shared_cache,
5246 : cause,
5247 : cancel,
5248 : ctx,
5249 : )
5250 : .await
5251 : }
5252 :
5253 : /// Calculate synthetic tenant size and cache the result.
5254 : /// This is periodically called by background worker.
5255 : /// result is cached in tenant struct
5256 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5257 : pub async fn calculate_synthetic_size(
5258 : &self,
5259 : cause: LogicalSizeCalculationCause,
5260 : cancel: &CancellationToken,
5261 : ctx: &RequestContext,
5262 : ) -> Result<u64, size::CalculateSyntheticSizeError> {
5263 : let inputs = self.gather_size_inputs(None, cause, cancel, ctx).await?;
5264 :
5265 : let size = inputs.calculate();
5266 :
5267 : self.set_cached_synthetic_size(size);
5268 :
5269 : Ok(size)
5270 : }
5271 :
5272 : /// Cache given synthetic size and update the metric value
5273 0 : pub fn set_cached_synthetic_size(&self, size: u64) {
5274 0 : self.cached_synthetic_tenant_size
5275 0 : .store(size, Ordering::Relaxed);
5276 0 :
5277 0 : // Only shard zero should be calculating synthetic sizes
5278 0 : debug_assert!(self.shard_identity.is_shard_zero());
5279 :
5280 0 : TENANT_SYNTHETIC_SIZE_METRIC
5281 0 : .get_metric_with_label_values(&[&self.tenant_shard_id.tenant_id.to_string()])
5282 0 : .unwrap()
5283 0 : .set(size);
5284 0 : }
5285 :
5286 0 : pub fn cached_synthetic_size(&self) -> u64 {
5287 0 : self.cached_synthetic_tenant_size.load(Ordering::Relaxed)
5288 0 : }
5289 :
5290 : /// Flush any in-progress layers, schedule uploads, and wait for uploads to complete.
5291 : ///
5292 : /// This function can take a long time: callers should wrap it in a timeout if calling
5293 : /// from an external API handler.
5294 : ///
5295 : /// Cancel-safety: cancelling this function may leave I/O running, but such I/O is
5296 : /// still bounded by tenant/timeline shutdown.
5297 0 : #[tracing::instrument(skip_all)]
5298 : pub(crate) async fn flush_remote(&self) -> anyhow::Result<()> {
5299 : let timelines = self.timelines.lock().unwrap().clone();
5300 :
5301 0 : async fn flush_timeline(_gate: GateGuard, timeline: Arc<Timeline>) -> anyhow::Result<()> {
5302 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Flushing...");
5303 0 : timeline.freeze_and_flush().await?;
5304 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Waiting for uploads...");
5305 0 : timeline.remote_client.wait_completion().await?;
5306 :
5307 0 : Ok(())
5308 0 : }
5309 :
5310 : // We do not use a JoinSet for these tasks, because we don't want them to be
5311 : // aborted when this function's future is cancelled: they should stay alive
5312 : // holding their GateGuard until they complete, to ensure their I/Os complete
5313 : // before Timeline shutdown completes.
5314 : let mut results = FuturesUnordered::new();
5315 :
5316 : for (_timeline_id, timeline) in timelines {
5317 : // Run each timeline's flush in a task holding the timeline's gate: this
5318 : // means that if this function's future is cancelled, the Timeline shutdown
5319 : // will still wait for any I/O in here to complete.
5320 : let Ok(gate) = timeline.gate.enter() else {
5321 : continue;
5322 : };
5323 0 : let jh = tokio::task::spawn(async move { flush_timeline(gate, timeline).await });
5324 : results.push(jh);
5325 : }
5326 :
5327 : while let Some(r) = results.next().await {
5328 : if let Err(e) = r {
5329 : if !e.is_cancelled() && !e.is_panic() {
5330 : tracing::error!("unexpected join error: {e:?}");
5331 : }
5332 : }
5333 : }
5334 :
5335 : // The flushes we did above were just writes, but the Tenant might have had
5336 : // pending deletions as well from recent compaction/gc: we want to flush those
5337 : // as well. This requires flushing the global delete queue. This is cheap
5338 : // because it's typically a no-op.
5339 : match self.deletion_queue_client.flush_execute().await {
5340 : Ok(_) => {}
5341 : Err(DeletionQueueError::ShuttingDown) => {}
5342 : }
5343 :
5344 : Ok(())
5345 : }
5346 :
5347 0 : pub(crate) fn get_tenant_conf(&self) -> TenantConfOpt {
5348 0 : self.tenant_conf.load().tenant_conf.clone()
5349 0 : }
5350 :
5351 : /// How much local storage would this tenant like to have? It can cope with
5352 : /// less than this (via eviction and on-demand downloads), but this function enables
5353 : /// the Tenant to advertise how much storage it would prefer to have to provide fast I/O
5354 : /// by keeping important things on local disk.
5355 : ///
5356 : /// This is a heuristic, not a guarantee: tenants that are long-idle will actually use less
5357 : /// than they report here, due to layer eviction. Tenants with many active branches may
5358 : /// actually use more than they report here.
5359 0 : pub(crate) fn local_storage_wanted(&self) -> u64 {
5360 0 : let timelines = self.timelines.lock().unwrap();
5361 0 :
5362 0 : // Heuristic: we use the max() of the timelines' visible sizes, rather than the sum. This
5363 0 : // reflects the observation that on tenants with multiple large branches, typically only one
5364 0 : // of them is used actively enough to occupy space on disk.
5365 0 : timelines
5366 0 : .values()
5367 0 : .map(|t| t.metrics.visible_physical_size_gauge.get())
5368 0 : .max()
5369 0 : .unwrap_or(0)
5370 0 : }
5371 :
5372 : /// Serialize and write the latest TenantManifest to remote storage.
5373 2 : pub(crate) async fn store_tenant_manifest(&self) -> Result<(), TenantManifestError> {
5374 : // Only one manifest write may be done at at time, and the contents of the manifest
5375 : // must be loaded while holding this lock. This makes it safe to call this function
5376 : // from anywhere without worrying about colliding updates.
5377 2 : let mut guard = tokio::select! {
5378 2 : g = self.tenant_manifest_upload.lock() => {
5379 2 : g
5380 : },
5381 2 : _ = self.cancel.cancelled() => {
5382 0 : return Err(TenantManifestError::Cancelled);
5383 : }
5384 : };
5385 :
5386 2 : let manifest = self.build_tenant_manifest();
5387 2 : if Some(&manifest) == (*guard).as_ref() {
5388 : // Optimisation: skip uploads that don't change anything.
5389 0 : return Ok(());
5390 2 : }
5391 2 :
5392 2 : upload_tenant_manifest(
5393 2 : &self.remote_storage,
5394 2 : &self.tenant_shard_id,
5395 2 : self.generation,
5396 2 : &manifest,
5397 2 : &self.cancel,
5398 2 : )
5399 2 : .await
5400 2 : .map_err(|e| {
5401 0 : if self.cancel.is_cancelled() {
5402 0 : TenantManifestError::Cancelled
5403 : } else {
5404 0 : TenantManifestError::RemoteStorage(e)
5405 : }
5406 2 : })?;
5407 :
5408 : // Store the successfully uploaded manifest, so that future callers can avoid
5409 : // re-uploading the same thing.
5410 2 : *guard = Some(manifest);
5411 2 :
5412 2 : Ok(())
5413 2 : }
5414 : }
5415 :
5416 : /// Create the cluster temporarily in 'initdbpath' directory inside the repository
5417 : /// to get bootstrap data for timeline initialization.
5418 0 : async fn run_initdb(
5419 0 : conf: &'static PageServerConf,
5420 0 : initdb_target_dir: &Utf8Path,
5421 0 : pg_version: u32,
5422 0 : cancel: &CancellationToken,
5423 0 : ) -> Result<(), InitdbError> {
5424 0 : let initdb_bin_path = conf
5425 0 : .pg_bin_dir(pg_version)
5426 0 : .map_err(InitdbError::Other)?
5427 0 : .join("initdb");
5428 0 : let initdb_lib_dir = conf.pg_lib_dir(pg_version).map_err(InitdbError::Other)?;
5429 0 : info!(
5430 0 : "running {} in {}, libdir: {}",
5431 : initdb_bin_path, initdb_target_dir, initdb_lib_dir,
5432 : );
5433 :
5434 0 : let _permit = INIT_DB_SEMAPHORE.acquire().await;
5435 :
5436 0 : let res = postgres_initdb::do_run_initdb(postgres_initdb::RunInitdbArgs {
5437 0 : superuser: &conf.superuser,
5438 0 : locale: &conf.locale,
5439 0 : initdb_bin: &initdb_bin_path,
5440 0 : pg_version,
5441 0 : library_search_path: &initdb_lib_dir,
5442 0 : pgdata: initdb_target_dir,
5443 0 : })
5444 0 : .await
5445 0 : .map_err(InitdbError::Inner);
5446 0 :
5447 0 : // This isn't true cancellation support, see above. Still return an error to
5448 0 : // excercise the cancellation code path.
5449 0 : if cancel.is_cancelled() {
5450 0 : return Err(InitdbError::Cancelled);
5451 0 : }
5452 0 :
5453 0 : res
5454 0 : }
5455 :
5456 : /// Dump contents of a layer file to stdout.
5457 0 : pub async fn dump_layerfile_from_path(
5458 0 : path: &Utf8Path,
5459 0 : verbose: bool,
5460 0 : ctx: &RequestContext,
5461 0 : ) -> anyhow::Result<()> {
5462 : use std::os::unix::fs::FileExt;
5463 :
5464 : // All layer files start with a two-byte "magic" value, to identify the kind of
5465 : // file.
5466 0 : let file = File::open(path)?;
5467 0 : let mut header_buf = [0u8; 2];
5468 0 : file.read_exact_at(&mut header_buf, 0)?;
5469 :
5470 0 : match u16::from_be_bytes(header_buf) {
5471 : crate::IMAGE_FILE_MAGIC => {
5472 0 : ImageLayer::new_for_path(path, file)?
5473 0 : .dump(verbose, ctx)
5474 0 : .await?
5475 : }
5476 : crate::DELTA_FILE_MAGIC => {
5477 0 : DeltaLayer::new_for_path(path, file)?
5478 0 : .dump(verbose, ctx)
5479 0 : .await?
5480 : }
5481 0 : magic => bail!("unrecognized magic identifier: {:?}", magic),
5482 : }
5483 :
5484 0 : Ok(())
5485 0 : }
5486 :
5487 : #[cfg(test)]
5488 : pub(crate) mod harness {
5489 : use bytes::{Bytes, BytesMut};
5490 : use once_cell::sync::OnceCell;
5491 : use pageserver_api::models::ShardParameters;
5492 : use pageserver_api::shard::ShardIndex;
5493 : use utils::logging;
5494 :
5495 : use crate::deletion_queue::mock::MockDeletionQueue;
5496 : use crate::l0_flush::L0FlushConfig;
5497 : use crate::walredo::apply_neon;
5498 : use pageserver_api::key::Key;
5499 : use pageserver_api::record::NeonWalRecord;
5500 :
5501 : use super::*;
5502 : use hex_literal::hex;
5503 : use utils::id::TenantId;
5504 :
5505 : pub const TIMELINE_ID: TimelineId =
5506 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
5507 : pub const NEW_TIMELINE_ID: TimelineId =
5508 : TimelineId::from_array(hex!("AA223344556677881122334455667788"));
5509 :
5510 : /// Convenience function to create a page image with given string as the only content
5511 5028831 : pub fn test_img(s: &str) -> Bytes {
5512 5028831 : let mut buf = BytesMut::new();
5513 5028831 : buf.extend_from_slice(s.as_bytes());
5514 5028831 : buf.resize(64, 0);
5515 5028831 :
5516 5028831 : buf.freeze()
5517 5028831 : }
5518 :
5519 : impl From<TenantConf> for TenantConfOpt {
5520 196 : fn from(tenant_conf: TenantConf) -> Self {
5521 196 : Self {
5522 196 : checkpoint_distance: Some(tenant_conf.checkpoint_distance),
5523 196 : checkpoint_timeout: Some(tenant_conf.checkpoint_timeout),
5524 196 : compaction_target_size: Some(tenant_conf.compaction_target_size),
5525 196 : compaction_period: Some(tenant_conf.compaction_period),
5526 196 : compaction_threshold: Some(tenant_conf.compaction_threshold),
5527 196 : compaction_algorithm: Some(tenant_conf.compaction_algorithm),
5528 196 : gc_horizon: Some(tenant_conf.gc_horizon),
5529 196 : gc_period: Some(tenant_conf.gc_period),
5530 196 : image_creation_threshold: Some(tenant_conf.image_creation_threshold),
5531 196 : pitr_interval: Some(tenant_conf.pitr_interval),
5532 196 : walreceiver_connect_timeout: Some(tenant_conf.walreceiver_connect_timeout),
5533 196 : lagging_wal_timeout: Some(tenant_conf.lagging_wal_timeout),
5534 196 : max_lsn_wal_lag: Some(tenant_conf.max_lsn_wal_lag),
5535 196 : eviction_policy: Some(tenant_conf.eviction_policy),
5536 196 : min_resident_size_override: tenant_conf.min_resident_size_override,
5537 196 : evictions_low_residence_duration_metric_threshold: Some(
5538 196 : tenant_conf.evictions_low_residence_duration_metric_threshold,
5539 196 : ),
5540 196 : heatmap_period: Some(tenant_conf.heatmap_period),
5541 196 : lazy_slru_download: Some(tenant_conf.lazy_slru_download),
5542 196 : timeline_get_throttle: Some(tenant_conf.timeline_get_throttle),
5543 196 : image_layer_creation_check_threshold: Some(
5544 196 : tenant_conf.image_layer_creation_check_threshold,
5545 196 : ),
5546 196 : lsn_lease_length: Some(tenant_conf.lsn_lease_length),
5547 196 : lsn_lease_length_for_ts: Some(tenant_conf.lsn_lease_length_for_ts),
5548 196 : timeline_offloading: Some(tenant_conf.timeline_offloading),
5549 196 : wal_receiver_protocol_override: tenant_conf.wal_receiver_protocol_override,
5550 196 : }
5551 196 : }
5552 : }
5553 :
5554 : pub struct TenantHarness {
5555 : pub conf: &'static PageServerConf,
5556 : pub tenant_conf: TenantConf,
5557 : pub tenant_shard_id: TenantShardId,
5558 : pub generation: Generation,
5559 : pub shard: ShardIndex,
5560 : pub remote_storage: GenericRemoteStorage,
5561 : pub remote_fs_dir: Utf8PathBuf,
5562 : pub deletion_queue: MockDeletionQueue,
5563 : }
5564 :
5565 : static LOG_HANDLE: OnceCell<()> = OnceCell::new();
5566 :
5567 212 : pub(crate) fn setup_logging() {
5568 212 : LOG_HANDLE.get_or_init(|| {
5569 200 : logging::init(
5570 200 : logging::LogFormat::Test,
5571 200 : // enable it in case the tests exercise code paths that use
5572 200 : // debug_assert_current_span_has_tenant_and_timeline_id
5573 200 : logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
5574 200 : logging::Output::Stdout,
5575 200 : )
5576 200 : .expect("Failed to init test logging")
5577 212 : });
5578 212 : }
5579 :
5580 : impl TenantHarness {
5581 196 : pub async fn create_custom(
5582 196 : test_name: &'static str,
5583 196 : tenant_conf: TenantConf,
5584 196 : tenant_id: TenantId,
5585 196 : shard_identity: ShardIdentity,
5586 196 : generation: Generation,
5587 196 : ) -> anyhow::Result<Self> {
5588 196 : setup_logging();
5589 196 :
5590 196 : let repo_dir = PageServerConf::test_repo_dir(test_name);
5591 196 : let _ = fs::remove_dir_all(&repo_dir);
5592 196 : fs::create_dir_all(&repo_dir)?;
5593 :
5594 196 : let conf = PageServerConf::dummy_conf(repo_dir);
5595 196 : // Make a static copy of the config. This can never be free'd, but that's
5596 196 : // OK in a test.
5597 196 : let conf: &'static PageServerConf = Box::leak(Box::new(conf));
5598 196 :
5599 196 : let shard = shard_identity.shard_index();
5600 196 : let tenant_shard_id = TenantShardId {
5601 196 : tenant_id,
5602 196 : shard_number: shard.shard_number,
5603 196 : shard_count: shard.shard_count,
5604 196 : };
5605 196 : fs::create_dir_all(conf.tenant_path(&tenant_shard_id))?;
5606 196 : fs::create_dir_all(conf.timelines_path(&tenant_shard_id))?;
5607 :
5608 : use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
5609 196 : let remote_fs_dir = conf.workdir.join("localfs");
5610 196 : std::fs::create_dir_all(&remote_fs_dir).unwrap();
5611 196 : let config = RemoteStorageConfig {
5612 196 : storage: RemoteStorageKind::LocalFs {
5613 196 : local_path: remote_fs_dir.clone(),
5614 196 : },
5615 196 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
5616 196 : small_timeout: RemoteStorageConfig::DEFAULT_SMALL_TIMEOUT,
5617 196 : };
5618 196 : let remote_storage = GenericRemoteStorage::from_config(&config).await.unwrap();
5619 196 : let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()));
5620 196 :
5621 196 : Ok(Self {
5622 196 : conf,
5623 196 : tenant_conf,
5624 196 : tenant_shard_id,
5625 196 : generation,
5626 196 : shard,
5627 196 : remote_storage,
5628 196 : remote_fs_dir,
5629 196 : deletion_queue,
5630 196 : })
5631 196 : }
5632 :
5633 184 : pub async fn create(test_name: &'static str) -> anyhow::Result<Self> {
5634 184 : // Disable automatic GC and compaction to make the unit tests more deterministic.
5635 184 : // The tests perform them manually if needed.
5636 184 : let tenant_conf = TenantConf {
5637 184 : gc_period: Duration::ZERO,
5638 184 : compaction_period: Duration::ZERO,
5639 184 : ..TenantConf::default()
5640 184 : };
5641 184 : let tenant_id = TenantId::generate();
5642 184 : let shard = ShardIdentity::unsharded();
5643 184 : Self::create_custom(
5644 184 : test_name,
5645 184 : tenant_conf,
5646 184 : tenant_id,
5647 184 : shard,
5648 184 : Generation::new(0xdeadbeef),
5649 184 : )
5650 184 : .await
5651 184 : }
5652 :
5653 20 : pub fn span(&self) -> tracing::Span {
5654 20 : info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
5655 20 : }
5656 :
5657 196 : pub(crate) async fn load(&self) -> (Arc<Tenant>, RequestContext) {
5658 196 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
5659 196 : (
5660 196 : self.do_try_load(&ctx)
5661 196 : .await
5662 196 : .expect("failed to load test tenant"),
5663 196 : ctx,
5664 196 : )
5665 196 : }
5666 :
5667 196 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5668 : pub(crate) async fn do_try_load(
5669 : &self,
5670 : ctx: &RequestContext,
5671 : ) -> anyhow::Result<Arc<Tenant>> {
5672 : let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
5673 :
5674 : let tenant = Arc::new(Tenant::new(
5675 : TenantState::Attaching,
5676 : self.conf,
5677 : AttachedTenantConf::try_from(LocationConf::attached_single(
5678 : TenantConfOpt::from(self.tenant_conf.clone()),
5679 : self.generation,
5680 : &ShardParameters::default(),
5681 : ))
5682 : .unwrap(),
5683 : // This is a legacy/test code path: sharding isn't supported here.
5684 : ShardIdentity::unsharded(),
5685 : Some(walredo_mgr),
5686 : self.tenant_shard_id,
5687 : self.remote_storage.clone(),
5688 : self.deletion_queue.new_client(),
5689 : // TODO: ideally we should run all unit tests with both configs
5690 : L0FlushGlobalState::new(L0FlushConfig::default()),
5691 : ));
5692 :
5693 : let preload = tenant
5694 : .preload(&self.remote_storage, CancellationToken::new())
5695 : .await?;
5696 : tenant.attach(Some(preload), ctx).await?;
5697 :
5698 : tenant.state.send_replace(TenantState::Active);
5699 : for timeline in tenant.timelines.lock().unwrap().values() {
5700 : timeline.set_state(TimelineState::Active);
5701 : }
5702 : Ok(tenant)
5703 : }
5704 :
5705 2 : pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
5706 2 : self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
5707 2 : }
5708 : }
5709 :
5710 : // Mock WAL redo manager that doesn't do much
5711 : pub(crate) struct TestRedoManager;
5712 :
5713 : impl TestRedoManager {
5714 : /// # Cancel-Safety
5715 : ///
5716 : /// This method is cancellation-safe.
5717 520 : pub async fn request_redo(
5718 520 : &self,
5719 520 : key: Key,
5720 520 : lsn: Lsn,
5721 520 : base_img: Option<(Lsn, Bytes)>,
5722 520 : records: Vec<(Lsn, NeonWalRecord)>,
5723 520 : _pg_version: u32,
5724 520 : ) -> Result<Bytes, walredo::Error> {
5725 770 : let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
5726 520 : if records_neon {
5727 : // For Neon wal records, we can decode without spawning postgres, so do so.
5728 520 : let mut page = match (base_img, records.first()) {
5729 454 : (Some((_lsn, img)), _) => {
5730 454 : let mut page = BytesMut::new();
5731 454 : page.extend_from_slice(&img);
5732 454 : page
5733 : }
5734 66 : (_, Some((_lsn, rec))) if rec.will_init() => BytesMut::new(),
5735 : _ => {
5736 0 : panic!("Neon WAL redo requires base image or will init record");
5737 : }
5738 : };
5739 :
5740 1290 : for (record_lsn, record) in records {
5741 770 : apply_neon::apply_in_neon(&record, record_lsn, key, &mut page)?;
5742 : }
5743 520 : Ok(page.freeze())
5744 : } else {
5745 : // We never spawn a postgres walredo process in unit tests: just log what we might have done.
5746 0 : let s = format!(
5747 0 : "redo for {} to get to {}, with {} and {} records",
5748 0 : key,
5749 0 : lsn,
5750 0 : if base_img.is_some() {
5751 0 : "base image"
5752 : } else {
5753 0 : "no base image"
5754 : },
5755 0 : records.len()
5756 0 : );
5757 0 : println!("{s}");
5758 0 :
5759 0 : Ok(test_img(&s))
5760 : }
5761 520 : }
5762 : }
5763 : }
5764 :
5765 : #[cfg(test)]
5766 : mod tests {
5767 : use std::collections::{BTreeMap, BTreeSet};
5768 :
5769 : use super::*;
5770 : use crate::keyspace::KeySpaceAccum;
5771 : use crate::tenant::harness::*;
5772 : use crate::tenant::timeline::CompactFlags;
5773 : use crate::DEFAULT_PG_VERSION;
5774 : use bytes::{Bytes, BytesMut};
5775 : use hex_literal::hex;
5776 : use itertools::Itertools;
5777 : use pageserver_api::key::{Key, AUX_KEY_PREFIX, NON_INHERITED_RANGE};
5778 : use pageserver_api::keyspace::KeySpace;
5779 : use pageserver_api::models::{CompactionAlgorithm, CompactionAlgorithmSettings};
5780 : use pageserver_api::value::Value;
5781 : use pageserver_compaction::helpers::overlaps_with;
5782 : use rand::{thread_rng, Rng};
5783 : use storage_layer::PersistentLayerKey;
5784 : use tests::storage_layer::ValuesReconstructState;
5785 : use tests::timeline::{GetVectoredError, ShutdownMode};
5786 : use timeline::{CompactOptions, DeltaLayerTestDesc};
5787 : use utils::id::TenantId;
5788 :
5789 : #[cfg(feature = "testing")]
5790 : use models::CompactLsnRange;
5791 : #[cfg(feature = "testing")]
5792 : use pageserver_api::record::NeonWalRecord;
5793 : #[cfg(feature = "testing")]
5794 : use timeline::compaction::{KeyHistoryRetention, KeyLogAtLsn};
5795 : #[cfg(feature = "testing")]
5796 : use timeline::GcInfo;
5797 :
5798 : static TEST_KEY: Lazy<Key> =
5799 18 : Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));
5800 :
5801 : #[tokio::test]
5802 2 : async fn test_basic() -> anyhow::Result<()> {
5803 2 : let (tenant, ctx) = TenantHarness::create("test_basic").await?.load().await;
5804 2 : let tline = tenant
5805 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5806 2 : .await?;
5807 2 :
5808 2 : let mut writer = tline.writer().await;
5809 2 : writer
5810 2 : .put(
5811 2 : *TEST_KEY,
5812 2 : Lsn(0x10),
5813 2 : &Value::Image(test_img("foo at 0x10")),
5814 2 : &ctx,
5815 2 : )
5816 2 : .await?;
5817 2 : writer.finish_write(Lsn(0x10));
5818 2 : drop(writer);
5819 2 :
5820 2 : let mut writer = tline.writer().await;
5821 2 : writer
5822 2 : .put(
5823 2 : *TEST_KEY,
5824 2 : Lsn(0x20),
5825 2 : &Value::Image(test_img("foo at 0x20")),
5826 2 : &ctx,
5827 2 : )
5828 2 : .await?;
5829 2 : writer.finish_write(Lsn(0x20));
5830 2 : drop(writer);
5831 2 :
5832 2 : assert_eq!(
5833 2 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
5834 2 : test_img("foo at 0x10")
5835 2 : );
5836 2 : assert_eq!(
5837 2 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
5838 2 : test_img("foo at 0x10")
5839 2 : );
5840 2 : assert_eq!(
5841 2 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
5842 2 : test_img("foo at 0x20")
5843 2 : );
5844 2 :
5845 2 : Ok(())
5846 2 : }
5847 :
5848 : #[tokio::test]
5849 2 : async fn no_duplicate_timelines() -> anyhow::Result<()> {
5850 2 : let (tenant, ctx) = TenantHarness::create("no_duplicate_timelines")
5851 2 : .await?
5852 2 : .load()
5853 2 : .await;
5854 2 : let _ = tenant
5855 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5856 2 : .await?;
5857 2 :
5858 2 : match tenant
5859 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5860 2 : .await
5861 2 : {
5862 2 : Ok(_) => panic!("duplicate timeline creation should fail"),
5863 2 : Err(e) => assert_eq!(
5864 2 : e.to_string(),
5865 2 : "timeline already exists with different parameters".to_string()
5866 2 : ),
5867 2 : }
5868 2 :
5869 2 : Ok(())
5870 2 : }
5871 :
5872 : /// Convenience function to create a page image with given string as the only content
5873 10 : pub fn test_value(s: &str) -> Value {
5874 10 : let mut buf = BytesMut::new();
5875 10 : buf.extend_from_slice(s.as_bytes());
5876 10 : Value::Image(buf.freeze())
5877 10 : }
5878 :
5879 : ///
5880 : /// Test branch creation
5881 : ///
5882 : #[tokio::test]
5883 2 : async fn test_branch() -> anyhow::Result<()> {
5884 2 : use std::str::from_utf8;
5885 2 :
5886 2 : let (tenant, ctx) = TenantHarness::create("test_branch").await?.load().await;
5887 2 : let tline = tenant
5888 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5889 2 : .await?;
5890 2 : let mut writer = tline.writer().await;
5891 2 :
5892 2 : #[allow(non_snake_case)]
5893 2 : let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
5894 2 : #[allow(non_snake_case)]
5895 2 : let TEST_KEY_B: Key = Key::from_hex("110000000033333333444444445500000002").unwrap();
5896 2 :
5897 2 : // Insert a value on the timeline
5898 2 : writer
5899 2 : .put(TEST_KEY_A, Lsn(0x20), &test_value("foo at 0x20"), &ctx)
5900 2 : .await?;
5901 2 : writer
5902 2 : .put(TEST_KEY_B, Lsn(0x20), &test_value("foobar at 0x20"), &ctx)
5903 2 : .await?;
5904 2 : writer.finish_write(Lsn(0x20));
5905 2 :
5906 2 : writer
5907 2 : .put(TEST_KEY_A, Lsn(0x30), &test_value("foo at 0x30"), &ctx)
5908 2 : .await?;
5909 2 : writer.finish_write(Lsn(0x30));
5910 2 : writer
5911 2 : .put(TEST_KEY_A, Lsn(0x40), &test_value("foo at 0x40"), &ctx)
5912 2 : .await?;
5913 2 : writer.finish_write(Lsn(0x40));
5914 2 :
5915 2 : //assert_current_logical_size(&tline, Lsn(0x40));
5916 2 :
5917 2 : // Branch the history, modify relation differently on the new timeline
5918 2 : tenant
5919 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x30)), &ctx)
5920 2 : .await?;
5921 2 : let newtline = tenant
5922 2 : .get_timeline(NEW_TIMELINE_ID, true)
5923 2 : .expect("Should have a local timeline");
5924 2 : let mut new_writer = newtline.writer().await;
5925 2 : new_writer
5926 2 : .put(TEST_KEY_A, Lsn(0x40), &test_value("bar at 0x40"), &ctx)
5927 2 : .await?;
5928 2 : new_writer.finish_write(Lsn(0x40));
5929 2 :
5930 2 : // Check page contents on both branches
5931 2 : assert_eq!(
5932 2 : from_utf8(&tline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
5933 2 : "foo at 0x40"
5934 2 : );
5935 2 : assert_eq!(
5936 2 : from_utf8(&newtline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
5937 2 : "bar at 0x40"
5938 2 : );
5939 2 : assert_eq!(
5940 2 : from_utf8(&newtline.get(TEST_KEY_B, Lsn(0x40), &ctx).await?)?,
5941 2 : "foobar at 0x20"
5942 2 : );
5943 2 :
5944 2 : //assert_current_logical_size(&tline, Lsn(0x40));
5945 2 :
5946 2 : Ok(())
5947 2 : }
5948 :
5949 20 : async fn make_some_layers(
5950 20 : tline: &Timeline,
5951 20 : start_lsn: Lsn,
5952 20 : ctx: &RequestContext,
5953 20 : ) -> anyhow::Result<()> {
5954 20 : let mut lsn = start_lsn;
5955 : {
5956 20 : let mut writer = tline.writer().await;
5957 : // Create a relation on the timeline
5958 20 : writer
5959 20 : .put(
5960 20 : *TEST_KEY,
5961 20 : lsn,
5962 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5963 20 : ctx,
5964 20 : )
5965 20 : .await?;
5966 20 : writer.finish_write(lsn);
5967 20 : lsn += 0x10;
5968 20 : writer
5969 20 : .put(
5970 20 : *TEST_KEY,
5971 20 : lsn,
5972 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5973 20 : ctx,
5974 20 : )
5975 20 : .await?;
5976 20 : writer.finish_write(lsn);
5977 20 : lsn += 0x10;
5978 20 : }
5979 20 : tline.freeze_and_flush().await?;
5980 : {
5981 20 : let mut writer = tline.writer().await;
5982 20 : writer
5983 20 : .put(
5984 20 : *TEST_KEY,
5985 20 : lsn,
5986 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5987 20 : ctx,
5988 20 : )
5989 20 : .await?;
5990 20 : writer.finish_write(lsn);
5991 20 : lsn += 0x10;
5992 20 : writer
5993 20 : .put(
5994 20 : *TEST_KEY,
5995 20 : lsn,
5996 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5997 20 : ctx,
5998 20 : )
5999 20 : .await?;
6000 20 : writer.finish_write(lsn);
6001 20 : }
6002 20 : tline.freeze_and_flush().await.map_err(|e| e.into())
6003 20 : }
6004 :
6005 : #[tokio::test(start_paused = true)]
6006 2 : async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
6007 2 : let (tenant, ctx) =
6008 2 : TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
6009 2 : .await?
6010 2 : .load()
6011 2 : .await;
6012 2 : // Advance to the lsn lease deadline so that GC is not blocked by
6013 2 : // initial transition into AttachedSingle.
6014 2 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
6015 2 : tokio::time::resume();
6016 2 : let tline = tenant
6017 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6018 2 : .await?;
6019 2 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6020 2 :
6021 2 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
6022 2 : // FIXME: this doesn't actually remove any layer currently, given how the flushing
6023 2 : // and compaction works. But it does set the 'cutoff' point so that the cross check
6024 2 : // below should fail.
6025 2 : tenant
6026 2 : .gc_iteration(
6027 2 : Some(TIMELINE_ID),
6028 2 : 0x10,
6029 2 : Duration::ZERO,
6030 2 : &CancellationToken::new(),
6031 2 : &ctx,
6032 2 : )
6033 2 : .await?;
6034 2 :
6035 2 : // try to branch at lsn 25, should fail because we already garbage collected the data
6036 2 : match tenant
6037 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
6038 2 : .await
6039 2 : {
6040 2 : Ok(_) => panic!("branching should have failed"),
6041 2 : Err(err) => {
6042 2 : let CreateTimelineError::AncestorLsn(err) = err else {
6043 2 : panic!("wrong error type")
6044 2 : };
6045 2 : assert!(err.to_string().contains("invalid branch start lsn"));
6046 2 : assert!(err
6047 2 : .source()
6048 2 : .unwrap()
6049 2 : .to_string()
6050 2 : .contains("we might've already garbage collected needed data"))
6051 2 : }
6052 2 : }
6053 2 :
6054 2 : Ok(())
6055 2 : }
6056 :
6057 : #[tokio::test]
6058 2 : async fn test_prohibit_branch_creation_on_pre_initdb_lsn() -> anyhow::Result<()> {
6059 2 : let (tenant, ctx) =
6060 2 : TenantHarness::create("test_prohibit_branch_creation_on_pre_initdb_lsn")
6061 2 : .await?
6062 2 : .load()
6063 2 : .await;
6064 2 :
6065 2 : let tline = tenant
6066 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x50), DEFAULT_PG_VERSION, &ctx)
6067 2 : .await?;
6068 2 : // try to branch at lsn 0x25, should fail because initdb lsn is 0x50
6069 2 : match tenant
6070 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
6071 2 : .await
6072 2 : {
6073 2 : Ok(_) => panic!("branching should have failed"),
6074 2 : Err(err) => {
6075 2 : let CreateTimelineError::AncestorLsn(err) = err else {
6076 2 : panic!("wrong error type");
6077 2 : };
6078 2 : assert!(&err.to_string().contains("invalid branch start lsn"));
6079 2 : assert!(&err
6080 2 : .source()
6081 2 : .unwrap()
6082 2 : .to_string()
6083 2 : .contains("is earlier than latest GC cutoff"));
6084 2 : }
6085 2 : }
6086 2 :
6087 2 : Ok(())
6088 2 : }
6089 :
6090 : /*
6091 : // FIXME: This currently fails to error out. Calling GC doesn't currently
6092 : // remove the old value, we'd need to work a little harder
6093 : #[tokio::test]
6094 : async fn test_prohibit_get_for_garbage_collected_data() -> anyhow::Result<()> {
6095 : let repo =
6096 : RepoHarness::create("test_prohibit_get_for_garbage_collected_data")?
6097 : .load();
6098 :
6099 : let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION)?;
6100 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6101 :
6102 : repo.gc_iteration(Some(TIMELINE_ID), 0x10, Duration::ZERO)?;
6103 : let latest_gc_cutoff_lsn = tline.get_latest_gc_cutoff_lsn();
6104 : assert!(*latest_gc_cutoff_lsn > Lsn(0x25));
6105 : match tline.get(*TEST_KEY, Lsn(0x25)) {
6106 : Ok(_) => panic!("request for page should have failed"),
6107 : Err(err) => assert!(err.to_string().contains("not found at")),
6108 : }
6109 : Ok(())
6110 : }
6111 : */
6112 :
6113 : #[tokio::test]
6114 2 : async fn test_get_branchpoints_from_an_inactive_timeline() -> anyhow::Result<()> {
6115 2 : let (tenant, ctx) =
6116 2 : TenantHarness::create("test_get_branchpoints_from_an_inactive_timeline")
6117 2 : .await?
6118 2 : .load()
6119 2 : .await;
6120 2 : let tline = tenant
6121 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6122 2 : .await?;
6123 2 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6124 2 :
6125 2 : tenant
6126 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6127 2 : .await?;
6128 2 : let newtline = tenant
6129 2 : .get_timeline(NEW_TIMELINE_ID, true)
6130 2 : .expect("Should have a local timeline");
6131 2 :
6132 2 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6133 2 :
6134 2 : tline.set_broken("test".to_owned());
6135 2 :
6136 2 : tenant
6137 2 : .gc_iteration(
6138 2 : Some(TIMELINE_ID),
6139 2 : 0x10,
6140 2 : Duration::ZERO,
6141 2 : &CancellationToken::new(),
6142 2 : &ctx,
6143 2 : )
6144 2 : .await?;
6145 2 :
6146 2 : // The branchpoints should contain all timelines, even ones marked
6147 2 : // as Broken.
6148 2 : {
6149 2 : let branchpoints = &tline.gc_info.read().unwrap().retain_lsns;
6150 2 : assert_eq!(branchpoints.len(), 1);
6151 2 : assert_eq!(
6152 2 : branchpoints[0],
6153 2 : (Lsn(0x40), NEW_TIMELINE_ID, MaybeOffloaded::No)
6154 2 : );
6155 2 : }
6156 2 :
6157 2 : // You can read the key from the child branch even though the parent is
6158 2 : // Broken, as long as you don't need to access data from the parent.
6159 2 : assert_eq!(
6160 2 : newtline.get(*TEST_KEY, Lsn(0x70), &ctx).await?,
6161 2 : test_img(&format!("foo at {}", Lsn(0x70)))
6162 2 : );
6163 2 :
6164 2 : // This needs to traverse to the parent, and fails.
6165 2 : let err = newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await.unwrap_err();
6166 2 : assert!(
6167 2 : err.to_string().starts_with(&format!(
6168 2 : "bad state on timeline {}: Broken",
6169 2 : tline.timeline_id
6170 2 : )),
6171 2 : "{err}"
6172 2 : );
6173 2 :
6174 2 : Ok(())
6175 2 : }
6176 :
6177 : #[tokio::test]
6178 2 : async fn test_retain_data_in_parent_which_is_needed_for_child() -> anyhow::Result<()> {
6179 2 : let (tenant, ctx) =
6180 2 : TenantHarness::create("test_retain_data_in_parent_which_is_needed_for_child")
6181 2 : .await?
6182 2 : .load()
6183 2 : .await;
6184 2 : let tline = tenant
6185 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6186 2 : .await?;
6187 2 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6188 2 :
6189 2 : tenant
6190 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6191 2 : .await?;
6192 2 : let newtline = tenant
6193 2 : .get_timeline(NEW_TIMELINE_ID, true)
6194 2 : .expect("Should have a local timeline");
6195 2 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
6196 2 : tenant
6197 2 : .gc_iteration(
6198 2 : Some(TIMELINE_ID),
6199 2 : 0x10,
6200 2 : Duration::ZERO,
6201 2 : &CancellationToken::new(),
6202 2 : &ctx,
6203 2 : )
6204 2 : .await?;
6205 2 : assert!(newtline.get(*TEST_KEY, Lsn(0x25), &ctx).await.is_ok());
6206 2 :
6207 2 : Ok(())
6208 2 : }
6209 : #[tokio::test]
6210 2 : async fn test_parent_keeps_data_forever_after_branching() -> anyhow::Result<()> {
6211 2 : let (tenant, ctx) = TenantHarness::create("test_parent_keeps_data_forever_after_branching")
6212 2 : .await?
6213 2 : .load()
6214 2 : .await;
6215 2 : let tline = tenant
6216 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6217 2 : .await?;
6218 2 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6219 2 :
6220 2 : tenant
6221 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6222 2 : .await?;
6223 2 : let newtline = tenant
6224 2 : .get_timeline(NEW_TIMELINE_ID, true)
6225 2 : .expect("Should have a local timeline");
6226 2 :
6227 2 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6228 2 :
6229 2 : // run gc on parent
6230 2 : tenant
6231 2 : .gc_iteration(
6232 2 : Some(TIMELINE_ID),
6233 2 : 0x10,
6234 2 : Duration::ZERO,
6235 2 : &CancellationToken::new(),
6236 2 : &ctx,
6237 2 : )
6238 2 : .await?;
6239 2 :
6240 2 : // Check that the data is still accessible on the branch.
6241 2 : assert_eq!(
6242 2 : newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
6243 2 : test_img(&format!("foo at {}", Lsn(0x40)))
6244 2 : );
6245 2 :
6246 2 : Ok(())
6247 2 : }
6248 :
6249 : #[tokio::test]
6250 2 : async fn timeline_load() -> anyhow::Result<()> {
6251 2 : const TEST_NAME: &str = "timeline_load";
6252 2 : let harness = TenantHarness::create(TEST_NAME).await?;
6253 2 : {
6254 2 : let (tenant, ctx) = harness.load().await;
6255 2 : let tline = tenant
6256 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x7000), DEFAULT_PG_VERSION, &ctx)
6257 2 : .await?;
6258 2 : make_some_layers(tline.as_ref(), Lsn(0x8000), &ctx).await?;
6259 2 : // so that all uploads finish & we can call harness.load() below again
6260 2 : tenant
6261 2 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
6262 2 : .instrument(harness.span())
6263 2 : .await
6264 2 : .ok()
6265 2 : .unwrap();
6266 2 : }
6267 2 :
6268 2 : let (tenant, _ctx) = harness.load().await;
6269 2 : tenant
6270 2 : .get_timeline(TIMELINE_ID, true)
6271 2 : .expect("cannot load timeline");
6272 2 :
6273 2 : Ok(())
6274 2 : }
6275 :
6276 : #[tokio::test]
6277 2 : async fn timeline_load_with_ancestor() -> anyhow::Result<()> {
6278 2 : const TEST_NAME: &str = "timeline_load_with_ancestor";
6279 2 : let harness = TenantHarness::create(TEST_NAME).await?;
6280 2 : // create two timelines
6281 2 : {
6282 2 : let (tenant, ctx) = harness.load().await;
6283 2 : let tline = tenant
6284 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6285 2 : .await?;
6286 2 :
6287 2 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6288 2 :
6289 2 : let child_tline = tenant
6290 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6291 2 : .await?;
6292 2 : child_tline.set_state(TimelineState::Active);
6293 2 :
6294 2 : let newtline = tenant
6295 2 : .get_timeline(NEW_TIMELINE_ID, true)
6296 2 : .expect("Should have a local timeline");
6297 2 :
6298 2 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6299 2 :
6300 2 : // so that all uploads finish & we can call harness.load() below again
6301 2 : tenant
6302 2 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
6303 2 : .instrument(harness.span())
6304 2 : .await
6305 2 : .ok()
6306 2 : .unwrap();
6307 2 : }
6308 2 :
6309 2 : // check that both of them are initially unloaded
6310 2 : let (tenant, _ctx) = harness.load().await;
6311 2 :
6312 2 : // check that both, child and ancestor are loaded
6313 2 : let _child_tline = tenant
6314 2 : .get_timeline(NEW_TIMELINE_ID, true)
6315 2 : .expect("cannot get child timeline loaded");
6316 2 :
6317 2 : let _ancestor_tline = tenant
6318 2 : .get_timeline(TIMELINE_ID, true)
6319 2 : .expect("cannot get ancestor timeline loaded");
6320 2 :
6321 2 : Ok(())
6322 2 : }
6323 :
6324 : #[tokio::test]
6325 2 : async fn delta_layer_dumping() -> anyhow::Result<()> {
6326 2 : use storage_layer::AsLayerDesc;
6327 2 : let (tenant, ctx) = TenantHarness::create("test_layer_dumping")
6328 2 : .await?
6329 2 : .load()
6330 2 : .await;
6331 2 : let tline = tenant
6332 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6333 2 : .await?;
6334 2 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6335 2 :
6336 2 : let layer_map = tline.layers.read().await;
6337 2 : let level0_deltas = layer_map
6338 2 : .layer_map()?
6339 2 : .level0_deltas()
6340 2 : .iter()
6341 4 : .map(|desc| layer_map.get_from_desc(desc))
6342 2 : .collect::<Vec<_>>();
6343 2 :
6344 2 : assert!(!level0_deltas.is_empty());
6345 2 :
6346 6 : for delta in level0_deltas {
6347 2 : // Ensure we are dumping a delta layer here
6348 4 : assert!(delta.layer_desc().is_delta);
6349 4 : delta.dump(true, &ctx).await.unwrap();
6350 2 : }
6351 2 :
6352 2 : Ok(())
6353 2 : }
6354 :
6355 : #[tokio::test]
6356 2 : async fn test_images() -> anyhow::Result<()> {
6357 2 : let (tenant, ctx) = TenantHarness::create("test_images").await?.load().await;
6358 2 : let tline = tenant
6359 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6360 2 : .await?;
6361 2 :
6362 2 : let mut writer = tline.writer().await;
6363 2 : writer
6364 2 : .put(
6365 2 : *TEST_KEY,
6366 2 : Lsn(0x10),
6367 2 : &Value::Image(test_img("foo at 0x10")),
6368 2 : &ctx,
6369 2 : )
6370 2 : .await?;
6371 2 : writer.finish_write(Lsn(0x10));
6372 2 : drop(writer);
6373 2 :
6374 2 : tline.freeze_and_flush().await?;
6375 2 : tline
6376 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6377 2 : .await?;
6378 2 :
6379 2 : let mut writer = tline.writer().await;
6380 2 : writer
6381 2 : .put(
6382 2 : *TEST_KEY,
6383 2 : Lsn(0x20),
6384 2 : &Value::Image(test_img("foo at 0x20")),
6385 2 : &ctx,
6386 2 : )
6387 2 : .await?;
6388 2 : writer.finish_write(Lsn(0x20));
6389 2 : drop(writer);
6390 2 :
6391 2 : tline.freeze_and_flush().await?;
6392 2 : tline
6393 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6394 2 : .await?;
6395 2 :
6396 2 : let mut writer = tline.writer().await;
6397 2 : writer
6398 2 : .put(
6399 2 : *TEST_KEY,
6400 2 : Lsn(0x30),
6401 2 : &Value::Image(test_img("foo at 0x30")),
6402 2 : &ctx,
6403 2 : )
6404 2 : .await?;
6405 2 : writer.finish_write(Lsn(0x30));
6406 2 : drop(writer);
6407 2 :
6408 2 : tline.freeze_and_flush().await?;
6409 2 : tline
6410 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6411 2 : .await?;
6412 2 :
6413 2 : let mut writer = tline.writer().await;
6414 2 : writer
6415 2 : .put(
6416 2 : *TEST_KEY,
6417 2 : Lsn(0x40),
6418 2 : &Value::Image(test_img("foo at 0x40")),
6419 2 : &ctx,
6420 2 : )
6421 2 : .await?;
6422 2 : writer.finish_write(Lsn(0x40));
6423 2 : drop(writer);
6424 2 :
6425 2 : tline.freeze_and_flush().await?;
6426 2 : tline
6427 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6428 2 : .await?;
6429 2 :
6430 2 : assert_eq!(
6431 2 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
6432 2 : test_img("foo at 0x10")
6433 2 : );
6434 2 : assert_eq!(
6435 2 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
6436 2 : test_img("foo at 0x10")
6437 2 : );
6438 2 : assert_eq!(
6439 2 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
6440 2 : test_img("foo at 0x20")
6441 2 : );
6442 2 : assert_eq!(
6443 2 : tline.get(*TEST_KEY, Lsn(0x30), &ctx).await?,
6444 2 : test_img("foo at 0x30")
6445 2 : );
6446 2 : assert_eq!(
6447 2 : tline.get(*TEST_KEY, Lsn(0x40), &ctx).await?,
6448 2 : test_img("foo at 0x40")
6449 2 : );
6450 2 :
6451 2 : Ok(())
6452 2 : }
6453 :
6454 4 : async fn bulk_insert_compact_gc(
6455 4 : tenant: &Tenant,
6456 4 : timeline: &Arc<Timeline>,
6457 4 : ctx: &RequestContext,
6458 4 : lsn: Lsn,
6459 4 : repeat: usize,
6460 4 : key_count: usize,
6461 4 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
6462 4 : let compact = true;
6463 4 : bulk_insert_maybe_compact_gc(tenant, timeline, ctx, lsn, repeat, key_count, compact).await
6464 4 : }
6465 :
6466 8 : async fn bulk_insert_maybe_compact_gc(
6467 8 : tenant: &Tenant,
6468 8 : timeline: &Arc<Timeline>,
6469 8 : ctx: &RequestContext,
6470 8 : mut lsn: Lsn,
6471 8 : repeat: usize,
6472 8 : key_count: usize,
6473 8 : compact: bool,
6474 8 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
6475 8 : let mut inserted: HashMap<Key, BTreeSet<Lsn>> = Default::default();
6476 8 :
6477 8 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6478 8 : let mut blknum = 0;
6479 8 :
6480 8 : // Enforce that key range is monotonously increasing
6481 8 : let mut keyspace = KeySpaceAccum::new();
6482 8 :
6483 8 : let cancel = CancellationToken::new();
6484 8 :
6485 8 : for _ in 0..repeat {
6486 400 : for _ in 0..key_count {
6487 4000000 : test_key.field6 = blknum;
6488 4000000 : let mut writer = timeline.writer().await;
6489 4000000 : writer
6490 4000000 : .put(
6491 4000000 : test_key,
6492 4000000 : lsn,
6493 4000000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6494 4000000 : ctx,
6495 4000000 : )
6496 4000000 : .await?;
6497 4000000 : inserted.entry(test_key).or_default().insert(lsn);
6498 4000000 : writer.finish_write(lsn);
6499 4000000 : drop(writer);
6500 4000000 :
6501 4000000 : keyspace.add_key(test_key);
6502 4000000 :
6503 4000000 : lsn = Lsn(lsn.0 + 0x10);
6504 4000000 : blknum += 1;
6505 : }
6506 :
6507 400 : timeline.freeze_and_flush().await?;
6508 400 : if compact {
6509 : // this requires timeline to be &Arc<Timeline>
6510 200 : timeline.compact(&cancel, EnumSet::empty(), ctx).await?;
6511 200 : }
6512 :
6513 : // this doesn't really need to use the timeline_id target, but it is closer to what it
6514 : // originally was.
6515 400 : let res = tenant
6516 400 : .gc_iteration(Some(timeline.timeline_id), 0, Duration::ZERO, &cancel, ctx)
6517 400 : .await?;
6518 :
6519 400 : assert_eq!(res.layers_removed, 0, "this never removes anything");
6520 : }
6521 :
6522 8 : Ok(inserted)
6523 8 : }
6524 :
6525 : //
6526 : // Insert 1000 key-value pairs with increasing keys, flush, compact, GC.
6527 : // Repeat 50 times.
6528 : //
6529 : #[tokio::test]
6530 2 : async fn test_bulk_insert() -> anyhow::Result<()> {
6531 2 : let harness = TenantHarness::create("test_bulk_insert").await?;
6532 2 : let (tenant, ctx) = harness.load().await;
6533 2 : let tline = tenant
6534 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6535 2 : .await?;
6536 2 :
6537 2 : let lsn = Lsn(0x10);
6538 2 : bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
6539 2 :
6540 2 : Ok(())
6541 2 : }
6542 :
6543 : // Test the vectored get real implementation against a simple sequential implementation.
6544 : //
6545 : // The test generates a keyspace by repeatedly flushing the in-memory layer and compacting.
6546 : // Projected to 2D the key space looks like below. Lsn grows upwards on the Y axis and keys
6547 : // grow to the right on the X axis.
6548 : // [Delta]
6549 : // [Delta]
6550 : // [Delta]
6551 : // [Delta]
6552 : // ------------ Image ---------------
6553 : //
6554 : // After layer generation we pick the ranges to query as follows:
6555 : // 1. The beginning of each delta layer
6556 : // 2. At the seam between two adjacent delta layers
6557 : //
6558 : // There's one major downside to this test: delta layers only contains images,
6559 : // so the search can stop at the first delta layer and doesn't traverse any deeper.
6560 : #[tokio::test]
6561 2 : async fn test_get_vectored() -> anyhow::Result<()> {
6562 2 : let harness = TenantHarness::create("test_get_vectored").await?;
6563 2 : let (tenant, ctx) = harness.load().await;
6564 2 : let tline = tenant
6565 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6566 2 : .await?;
6567 2 :
6568 2 : let lsn = Lsn(0x10);
6569 2 : let inserted = bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
6570 2 :
6571 2 : let guard = tline.layers.read().await;
6572 2 : let lm = guard.layer_map()?;
6573 2 :
6574 2 : lm.dump(true, &ctx).await?;
6575 2 :
6576 2 : let mut reads = Vec::new();
6577 2 : let mut prev = None;
6578 12 : lm.iter_historic_layers().for_each(|desc| {
6579 12 : if !desc.is_delta() {
6580 2 : prev = Some(desc.clone());
6581 2 : return;
6582 10 : }
6583 10 :
6584 10 : let start = desc.key_range.start;
6585 10 : let end = desc
6586 10 : .key_range
6587 10 : .start
6588 10 : .add(Timeline::MAX_GET_VECTORED_KEYS.try_into().unwrap());
6589 10 : reads.push(KeySpace {
6590 10 : ranges: vec![start..end],
6591 10 : });
6592 2 :
6593 10 : if let Some(prev) = &prev {
6594 10 : if !prev.is_delta() {
6595 10 : return;
6596 2 : }
6597 0 :
6598 0 : let first_range = Key {
6599 0 : field6: prev.key_range.end.field6 - 4,
6600 0 : ..prev.key_range.end
6601 0 : }..prev.key_range.end;
6602 0 :
6603 0 : let second_range = desc.key_range.start..Key {
6604 0 : field6: desc.key_range.start.field6 + 4,
6605 0 : ..desc.key_range.start
6606 0 : };
6607 0 :
6608 0 : reads.push(KeySpace {
6609 0 : ranges: vec![first_range, second_range],
6610 0 : });
6611 2 : };
6612 2 :
6613 2 : prev = Some(desc.clone());
6614 12 : });
6615 2 :
6616 2 : drop(guard);
6617 2 :
6618 2 : // Pick a big LSN such that we query over all the changes.
6619 2 : let reads_lsn = Lsn(u64::MAX - 1);
6620 2 :
6621 12 : for read in reads {
6622 10 : info!("Doing vectored read on {:?}", read);
6623 2 :
6624 10 : let vectored_res = tline
6625 10 : .get_vectored_impl(
6626 10 : read.clone(),
6627 10 : reads_lsn,
6628 10 : &mut ValuesReconstructState::new(),
6629 10 : &ctx,
6630 10 : )
6631 10 : .await;
6632 2 :
6633 10 : let mut expected_lsns: HashMap<Key, Lsn> = Default::default();
6634 10 : let mut expect_missing = false;
6635 10 : let mut key = read.start().unwrap();
6636 330 : while key != read.end().unwrap() {
6637 320 : if let Some(lsns) = inserted.get(&key) {
6638 320 : let expected_lsn = lsns.iter().rfind(|lsn| **lsn <= reads_lsn);
6639 320 : match expected_lsn {
6640 320 : Some(lsn) => {
6641 320 : expected_lsns.insert(key, *lsn);
6642 320 : }
6643 2 : None => {
6644 2 : expect_missing = true;
6645 0 : break;
6646 2 : }
6647 2 : }
6648 2 : } else {
6649 2 : expect_missing = true;
6650 0 : break;
6651 2 : }
6652 2 :
6653 320 : key = key.next();
6654 2 : }
6655 2 :
6656 10 : if expect_missing {
6657 2 : assert!(matches!(vectored_res, Err(GetVectoredError::MissingKey(_))));
6658 2 : } else {
6659 320 : for (key, image) in vectored_res? {
6660 320 : let expected_lsn = expected_lsns.get(&key).expect("determined above");
6661 320 : let expected_image = test_img(&format!("{} at {}", key.field6, expected_lsn));
6662 320 : assert_eq!(image?, expected_image);
6663 2 : }
6664 2 : }
6665 2 : }
6666 2 :
6667 2 : Ok(())
6668 2 : }
6669 :
6670 : #[tokio::test]
6671 2 : async fn test_get_vectored_aux_files() -> anyhow::Result<()> {
6672 2 : let harness = TenantHarness::create("test_get_vectored_aux_files").await?;
6673 2 :
6674 2 : let (tenant, ctx) = harness.load().await;
6675 2 : let tline = tenant
6676 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
6677 2 : .await?;
6678 2 : let tline = tline.raw_timeline().unwrap();
6679 2 :
6680 2 : let mut modification = tline.begin_modification(Lsn(0x1000));
6681 2 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
6682 2 : modification.set_lsn(Lsn(0x1008))?;
6683 2 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
6684 2 : modification.commit(&ctx).await?;
6685 2 :
6686 2 : let child_timeline_id = TimelineId::generate();
6687 2 : tenant
6688 2 : .branch_timeline_test(
6689 2 : tline,
6690 2 : child_timeline_id,
6691 2 : Some(tline.get_last_record_lsn()),
6692 2 : &ctx,
6693 2 : )
6694 2 : .await?;
6695 2 :
6696 2 : let child_timeline = tenant
6697 2 : .get_timeline(child_timeline_id, true)
6698 2 : .expect("Should have the branched timeline");
6699 2 :
6700 2 : let aux_keyspace = KeySpace {
6701 2 : ranges: vec![NON_INHERITED_RANGE],
6702 2 : };
6703 2 : let read_lsn = child_timeline.get_last_record_lsn();
6704 2 :
6705 2 : let vectored_res = child_timeline
6706 2 : .get_vectored_impl(
6707 2 : aux_keyspace.clone(),
6708 2 : read_lsn,
6709 2 : &mut ValuesReconstructState::new(),
6710 2 : &ctx,
6711 2 : )
6712 2 : .await;
6713 2 :
6714 2 : let images = vectored_res?;
6715 2 : assert!(images.is_empty());
6716 2 : Ok(())
6717 2 : }
6718 :
6719 : // Test that vectored get handles layer gaps correctly
6720 : // by advancing into the next ancestor timeline if required.
6721 : //
6722 : // The test generates timelines that look like the diagram below.
6723 : // We leave a gap in one of the L1 layers at `gap_at_key` (`/` in the diagram).
6724 : // The reconstruct data for that key lies in the ancestor timeline (`X` in the diagram).
6725 : //
6726 : // ```
6727 : //-------------------------------+
6728 : // ... |
6729 : // [ L1 ] |
6730 : // [ / L1 ] | Child Timeline
6731 : // ... |
6732 : // ------------------------------+
6733 : // [ X L1 ] | Parent Timeline
6734 : // ------------------------------+
6735 : // ```
6736 : #[tokio::test]
6737 2 : async fn test_get_vectored_key_gap() -> anyhow::Result<()> {
6738 2 : let tenant_conf = TenantConf {
6739 2 : // Make compaction deterministic
6740 2 : gc_period: Duration::ZERO,
6741 2 : compaction_period: Duration::ZERO,
6742 2 : // Encourage creation of L1 layers
6743 2 : checkpoint_distance: 16 * 1024,
6744 2 : compaction_target_size: 8 * 1024,
6745 2 : ..TenantConf::default()
6746 2 : };
6747 2 :
6748 2 : let harness = TenantHarness::create_custom(
6749 2 : "test_get_vectored_key_gap",
6750 2 : tenant_conf,
6751 2 : TenantId::generate(),
6752 2 : ShardIdentity::unsharded(),
6753 2 : Generation::new(0xdeadbeef),
6754 2 : )
6755 2 : .await?;
6756 2 : let (tenant, ctx) = harness.load().await;
6757 2 :
6758 2 : let mut current_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6759 2 : let gap_at_key = current_key.add(100);
6760 2 : let mut current_lsn = Lsn(0x10);
6761 2 :
6762 2 : const KEY_COUNT: usize = 10_000;
6763 2 :
6764 2 : let timeline_id = TimelineId::generate();
6765 2 : let current_timeline = tenant
6766 2 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
6767 2 : .await?;
6768 2 :
6769 2 : current_lsn += 0x100;
6770 2 :
6771 2 : let mut writer = current_timeline.writer().await;
6772 2 : writer
6773 2 : .put(
6774 2 : gap_at_key,
6775 2 : current_lsn,
6776 2 : &Value::Image(test_img(&format!("{} at {}", gap_at_key, current_lsn))),
6777 2 : &ctx,
6778 2 : )
6779 2 : .await?;
6780 2 : writer.finish_write(current_lsn);
6781 2 : drop(writer);
6782 2 :
6783 2 : let mut latest_lsns = HashMap::new();
6784 2 : latest_lsns.insert(gap_at_key, current_lsn);
6785 2 :
6786 2 : current_timeline.freeze_and_flush().await?;
6787 2 :
6788 2 : let child_timeline_id = TimelineId::generate();
6789 2 :
6790 2 : tenant
6791 2 : .branch_timeline_test(
6792 2 : ¤t_timeline,
6793 2 : child_timeline_id,
6794 2 : Some(current_lsn),
6795 2 : &ctx,
6796 2 : )
6797 2 : .await?;
6798 2 : let child_timeline = tenant
6799 2 : .get_timeline(child_timeline_id, true)
6800 2 : .expect("Should have the branched timeline");
6801 2 :
6802 20002 : for i in 0..KEY_COUNT {
6803 20000 : if current_key == gap_at_key {
6804 2 : current_key = current_key.next();
6805 2 : continue;
6806 19998 : }
6807 19998 :
6808 19998 : current_lsn += 0x10;
6809 2 :
6810 19998 : let mut writer = child_timeline.writer().await;
6811 19998 : writer
6812 19998 : .put(
6813 19998 : current_key,
6814 19998 : current_lsn,
6815 19998 : &Value::Image(test_img(&format!("{} at {}", current_key, current_lsn))),
6816 19998 : &ctx,
6817 19998 : )
6818 19998 : .await?;
6819 19998 : writer.finish_write(current_lsn);
6820 19998 : drop(writer);
6821 19998 :
6822 19998 : latest_lsns.insert(current_key, current_lsn);
6823 19998 : current_key = current_key.next();
6824 19998 :
6825 19998 : // Flush every now and then to encourage layer file creation.
6826 19998 : if i % 500 == 0 {
6827 40 : child_timeline.freeze_and_flush().await?;
6828 19958 : }
6829 2 : }
6830 2 :
6831 2 : child_timeline.freeze_and_flush().await?;
6832 2 : let mut flags = EnumSet::new();
6833 2 : flags.insert(CompactFlags::ForceRepartition);
6834 2 : child_timeline
6835 2 : .compact(&CancellationToken::new(), flags, &ctx)
6836 2 : .await?;
6837 2 :
6838 2 : let key_near_end = {
6839 2 : let mut tmp = current_key;
6840 2 : tmp.field6 -= 10;
6841 2 : tmp
6842 2 : };
6843 2 :
6844 2 : let key_near_gap = {
6845 2 : let mut tmp = gap_at_key;
6846 2 : tmp.field6 -= 10;
6847 2 : tmp
6848 2 : };
6849 2 :
6850 2 : let read = KeySpace {
6851 2 : ranges: vec![key_near_gap..gap_at_key.next(), key_near_end..current_key],
6852 2 : };
6853 2 : let results = child_timeline
6854 2 : .get_vectored_impl(
6855 2 : read.clone(),
6856 2 : current_lsn,
6857 2 : &mut ValuesReconstructState::new(),
6858 2 : &ctx,
6859 2 : )
6860 2 : .await?;
6861 2 :
6862 44 : for (key, img_res) in results {
6863 42 : let expected = test_img(&format!("{} at {}", key, latest_lsns[&key]));
6864 42 : assert_eq!(img_res?, expected);
6865 2 : }
6866 2 :
6867 2 : Ok(())
6868 2 : }
6869 :
6870 : // Test that vectored get descends into ancestor timelines correctly and
6871 : // does not return an image that's newer than requested.
6872 : //
6873 : // The diagram below ilustrates an interesting case. We have a parent timeline
6874 : // (top of the Lsn range) and a child timeline. The request key cannot be reconstructed
6875 : // from the child timeline, so the parent timeline must be visited. When advacing into
6876 : // the child timeline, the read path needs to remember what the requested Lsn was in
6877 : // order to avoid returning an image that's too new. The test below constructs such
6878 : // a timeline setup and does a few queries around the Lsn of each page image.
6879 : // ```
6880 : // LSN
6881 : // ^
6882 : // |
6883 : // |
6884 : // 500 | --------------------------------------> branch point
6885 : // 400 | X
6886 : // 300 | X
6887 : // 200 | --------------------------------------> requested lsn
6888 : // 100 | X
6889 : // |---------------------------------------> Key
6890 : // |
6891 : // ------> requested key
6892 : //
6893 : // Legend:
6894 : // * X - page images
6895 : // ```
6896 : #[tokio::test]
6897 2 : async fn test_get_vectored_ancestor_descent() -> anyhow::Result<()> {
6898 2 : let harness = TenantHarness::create("test_get_vectored_on_lsn_axis").await?;
6899 2 : let (tenant, ctx) = harness.load().await;
6900 2 :
6901 2 : let start_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6902 2 : let end_key = start_key.add(1000);
6903 2 : let child_gap_at_key = start_key.add(500);
6904 2 : let mut parent_gap_lsns: BTreeMap<Lsn, String> = BTreeMap::new();
6905 2 :
6906 2 : let mut current_lsn = Lsn(0x10);
6907 2 :
6908 2 : let timeline_id = TimelineId::generate();
6909 2 : let parent_timeline = tenant
6910 2 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
6911 2 : .await?;
6912 2 :
6913 2 : current_lsn += 0x100;
6914 2 :
6915 8 : for _ in 0..3 {
6916 6 : let mut key = start_key;
6917 6006 : while key < end_key {
6918 6000 : current_lsn += 0x10;
6919 6000 :
6920 6000 : let image_value = format!("{} at {}", child_gap_at_key, current_lsn);
6921 2 :
6922 6000 : let mut writer = parent_timeline.writer().await;
6923 6000 : writer
6924 6000 : .put(
6925 6000 : key,
6926 6000 : current_lsn,
6927 6000 : &Value::Image(test_img(&image_value)),
6928 6000 : &ctx,
6929 6000 : )
6930 6000 : .await?;
6931 6000 : writer.finish_write(current_lsn);
6932 6000 :
6933 6000 : if key == child_gap_at_key {
6934 6 : parent_gap_lsns.insert(current_lsn, image_value);
6935 5994 : }
6936 2 :
6937 6000 : key = key.next();
6938 2 : }
6939 2 :
6940 6 : parent_timeline.freeze_and_flush().await?;
6941 2 : }
6942 2 :
6943 2 : let child_timeline_id = TimelineId::generate();
6944 2 :
6945 2 : let child_timeline = tenant
6946 2 : .branch_timeline_test(&parent_timeline, child_timeline_id, Some(current_lsn), &ctx)
6947 2 : .await?;
6948 2 :
6949 2 : let mut key = start_key;
6950 2002 : while key < end_key {
6951 2000 : if key == child_gap_at_key {
6952 2 : key = key.next();
6953 2 : continue;
6954 1998 : }
6955 1998 :
6956 1998 : current_lsn += 0x10;
6957 2 :
6958 1998 : let mut writer = child_timeline.writer().await;
6959 1998 : writer
6960 1998 : .put(
6961 1998 : key,
6962 1998 : current_lsn,
6963 1998 : &Value::Image(test_img(&format!("{} at {}", key, current_lsn))),
6964 1998 : &ctx,
6965 1998 : )
6966 1998 : .await?;
6967 1998 : writer.finish_write(current_lsn);
6968 1998 :
6969 1998 : key = key.next();
6970 2 : }
6971 2 :
6972 2 : child_timeline.freeze_and_flush().await?;
6973 2 :
6974 2 : let lsn_offsets: [i64; 5] = [-10, -1, 0, 1, 10];
6975 2 : let mut query_lsns = Vec::new();
6976 6 : for image_lsn in parent_gap_lsns.keys().rev() {
6977 36 : for offset in lsn_offsets {
6978 30 : query_lsns.push(Lsn(image_lsn
6979 30 : .0
6980 30 : .checked_add_signed(offset)
6981 30 : .expect("Shouldn't overflow")));
6982 30 : }
6983 2 : }
6984 2 :
6985 32 : for query_lsn in query_lsns {
6986 30 : let results = child_timeline
6987 30 : .get_vectored_impl(
6988 30 : KeySpace {
6989 30 : ranges: vec![child_gap_at_key..child_gap_at_key.next()],
6990 30 : },
6991 30 : query_lsn,
6992 30 : &mut ValuesReconstructState::new(),
6993 30 : &ctx,
6994 30 : )
6995 30 : .await;
6996 2 :
6997 30 : let expected_item = parent_gap_lsns
6998 30 : .iter()
6999 30 : .rev()
7000 68 : .find(|(lsn, _)| **lsn <= query_lsn);
7001 30 :
7002 30 : info!(
7003 2 : "Doing vectored read at LSN {}. Expecting image to be: {:?}",
7004 2 : query_lsn, expected_item
7005 2 : );
7006 2 :
7007 30 : match expected_item {
7008 26 : Some((_, img_value)) => {
7009 26 : let key_results = results.expect("No vectored get error expected");
7010 26 : let key_result = &key_results[&child_gap_at_key];
7011 26 : let returned_img = key_result
7012 26 : .as_ref()
7013 26 : .expect("No page reconstruct error expected");
7014 26 :
7015 26 : info!(
7016 2 : "Vectored read at LSN {} returned image {}",
7017 0 : query_lsn,
7018 0 : std::str::from_utf8(returned_img)?
7019 2 : );
7020 26 : assert_eq!(*returned_img, test_img(img_value));
7021 2 : }
7022 2 : None => {
7023 4 : assert!(matches!(results, Err(GetVectoredError::MissingKey(_))));
7024 2 : }
7025 2 : }
7026 2 : }
7027 2 :
7028 2 : Ok(())
7029 2 : }
7030 :
7031 : #[tokio::test]
7032 2 : async fn test_random_updates() -> anyhow::Result<()> {
7033 2 : let names_algorithms = [
7034 2 : ("test_random_updates_legacy", CompactionAlgorithm::Legacy),
7035 2 : ("test_random_updates_tiered", CompactionAlgorithm::Tiered),
7036 2 : ];
7037 6 : for (name, algorithm) in names_algorithms {
7038 4 : test_random_updates_algorithm(name, algorithm).await?;
7039 2 : }
7040 2 : Ok(())
7041 2 : }
7042 :
7043 4 : async fn test_random_updates_algorithm(
7044 4 : name: &'static str,
7045 4 : compaction_algorithm: CompactionAlgorithm,
7046 4 : ) -> anyhow::Result<()> {
7047 4 : let mut harness = TenantHarness::create(name).await?;
7048 4 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
7049 4 : kind: compaction_algorithm,
7050 4 : };
7051 4 : let (tenant, ctx) = harness.load().await;
7052 4 : let tline = tenant
7053 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7054 4 : .await?;
7055 :
7056 : const NUM_KEYS: usize = 1000;
7057 4 : let cancel = CancellationToken::new();
7058 4 :
7059 4 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7060 4 : let mut test_key_end = test_key;
7061 4 : test_key_end.field6 = NUM_KEYS as u32;
7062 4 : tline.add_extra_test_dense_keyspace(KeySpace::single(test_key..test_key_end));
7063 4 :
7064 4 : let mut keyspace = KeySpaceAccum::new();
7065 4 :
7066 4 : // Track when each page was last modified. Used to assert that
7067 4 : // a read sees the latest page version.
7068 4 : let mut updated = [Lsn(0); NUM_KEYS];
7069 4 :
7070 4 : let mut lsn = Lsn(0x10);
7071 : #[allow(clippy::needless_range_loop)]
7072 4004 : for blknum in 0..NUM_KEYS {
7073 4000 : lsn = Lsn(lsn.0 + 0x10);
7074 4000 : test_key.field6 = blknum as u32;
7075 4000 : let mut writer = tline.writer().await;
7076 4000 : writer
7077 4000 : .put(
7078 4000 : test_key,
7079 4000 : lsn,
7080 4000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7081 4000 : &ctx,
7082 4000 : )
7083 4000 : .await?;
7084 4000 : writer.finish_write(lsn);
7085 4000 : updated[blknum] = lsn;
7086 4000 : drop(writer);
7087 4000 :
7088 4000 : keyspace.add_key(test_key);
7089 : }
7090 :
7091 204 : for _ in 0..50 {
7092 200200 : for _ in 0..NUM_KEYS {
7093 200000 : lsn = Lsn(lsn.0 + 0x10);
7094 200000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7095 200000 : test_key.field6 = blknum as u32;
7096 200000 : let mut writer = tline.writer().await;
7097 200000 : writer
7098 200000 : .put(
7099 200000 : test_key,
7100 200000 : lsn,
7101 200000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7102 200000 : &ctx,
7103 200000 : )
7104 200000 : .await?;
7105 200000 : writer.finish_write(lsn);
7106 200000 : drop(writer);
7107 200000 : updated[blknum] = lsn;
7108 : }
7109 :
7110 : // Read all the blocks
7111 200000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7112 200000 : test_key.field6 = blknum as u32;
7113 200000 : assert_eq!(
7114 200000 : tline.get(test_key, lsn, &ctx).await?,
7115 200000 : test_img(&format!("{} at {}", blknum, last_lsn))
7116 : );
7117 : }
7118 :
7119 : // Perform a cycle of flush, and GC
7120 200 : tline.freeze_and_flush().await?;
7121 200 : tenant
7122 200 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7123 200 : .await?;
7124 : }
7125 :
7126 4 : Ok(())
7127 4 : }
7128 :
7129 : #[tokio::test]
7130 2 : async fn test_traverse_branches() -> anyhow::Result<()> {
7131 2 : let (tenant, ctx) = TenantHarness::create("test_traverse_branches")
7132 2 : .await?
7133 2 : .load()
7134 2 : .await;
7135 2 : let mut tline = tenant
7136 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7137 2 : .await?;
7138 2 :
7139 2 : const NUM_KEYS: usize = 1000;
7140 2 :
7141 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7142 2 :
7143 2 : let mut keyspace = KeySpaceAccum::new();
7144 2 :
7145 2 : let cancel = CancellationToken::new();
7146 2 :
7147 2 : // Track when each page was last modified. Used to assert that
7148 2 : // a read sees the latest page version.
7149 2 : let mut updated = [Lsn(0); NUM_KEYS];
7150 2 :
7151 2 : let mut lsn = Lsn(0x10);
7152 2 : #[allow(clippy::needless_range_loop)]
7153 2002 : for blknum in 0..NUM_KEYS {
7154 2000 : lsn = Lsn(lsn.0 + 0x10);
7155 2000 : test_key.field6 = blknum as u32;
7156 2000 : let mut writer = tline.writer().await;
7157 2000 : writer
7158 2000 : .put(
7159 2000 : test_key,
7160 2000 : lsn,
7161 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7162 2000 : &ctx,
7163 2000 : )
7164 2000 : .await?;
7165 2000 : writer.finish_write(lsn);
7166 2000 : updated[blknum] = lsn;
7167 2000 : drop(writer);
7168 2000 :
7169 2000 : keyspace.add_key(test_key);
7170 2 : }
7171 2 :
7172 102 : for _ in 0..50 {
7173 100 : let new_tline_id = TimelineId::generate();
7174 100 : tenant
7175 100 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7176 100 : .await?;
7177 100 : tline = tenant
7178 100 : .get_timeline(new_tline_id, true)
7179 100 : .expect("Should have the branched timeline");
7180 2 :
7181 100100 : for _ in 0..NUM_KEYS {
7182 100000 : lsn = Lsn(lsn.0 + 0x10);
7183 100000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7184 100000 : test_key.field6 = blknum as u32;
7185 100000 : let mut writer = tline.writer().await;
7186 100000 : writer
7187 100000 : .put(
7188 100000 : test_key,
7189 100000 : lsn,
7190 100000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7191 100000 : &ctx,
7192 100000 : )
7193 100000 : .await?;
7194 100000 : println!("updating {} at {}", blknum, lsn);
7195 100000 : writer.finish_write(lsn);
7196 100000 : drop(writer);
7197 100000 : updated[blknum] = lsn;
7198 2 : }
7199 2 :
7200 2 : // Read all the blocks
7201 100000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7202 100000 : test_key.field6 = blknum as u32;
7203 100000 : assert_eq!(
7204 100000 : tline.get(test_key, lsn, &ctx).await?,
7205 100000 : test_img(&format!("{} at {}", blknum, last_lsn))
7206 2 : );
7207 2 : }
7208 2 :
7209 2 : // Perform a cycle of flush, compact, and GC
7210 100 : tline.freeze_and_flush().await?;
7211 100 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
7212 100 : tenant
7213 100 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7214 100 : .await?;
7215 2 : }
7216 2 :
7217 2 : Ok(())
7218 2 : }
7219 :
7220 : #[tokio::test]
7221 2 : async fn test_traverse_ancestors() -> anyhow::Result<()> {
7222 2 : let (tenant, ctx) = TenantHarness::create("test_traverse_ancestors")
7223 2 : .await?
7224 2 : .load()
7225 2 : .await;
7226 2 : let mut tline = tenant
7227 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7228 2 : .await?;
7229 2 :
7230 2 : const NUM_KEYS: usize = 100;
7231 2 : const NUM_TLINES: usize = 50;
7232 2 :
7233 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7234 2 : // Track page mutation lsns across different timelines.
7235 2 : let mut updated = [[Lsn(0); NUM_KEYS]; NUM_TLINES];
7236 2 :
7237 2 : let mut lsn = Lsn(0x10);
7238 2 :
7239 2 : #[allow(clippy::needless_range_loop)]
7240 102 : for idx in 0..NUM_TLINES {
7241 100 : let new_tline_id = TimelineId::generate();
7242 100 : tenant
7243 100 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7244 100 : .await?;
7245 100 : tline = tenant
7246 100 : .get_timeline(new_tline_id, true)
7247 100 : .expect("Should have the branched timeline");
7248 2 :
7249 10100 : for _ in 0..NUM_KEYS {
7250 10000 : lsn = Lsn(lsn.0 + 0x10);
7251 10000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7252 10000 : test_key.field6 = blknum as u32;
7253 10000 : let mut writer = tline.writer().await;
7254 10000 : writer
7255 10000 : .put(
7256 10000 : test_key,
7257 10000 : lsn,
7258 10000 : &Value::Image(test_img(&format!("{} {} at {}", idx, blknum, lsn))),
7259 10000 : &ctx,
7260 10000 : )
7261 10000 : .await?;
7262 10000 : println!("updating [{}][{}] at {}", idx, blknum, lsn);
7263 10000 : writer.finish_write(lsn);
7264 10000 : drop(writer);
7265 10000 : updated[idx][blknum] = lsn;
7266 2 : }
7267 2 : }
7268 2 :
7269 2 : // Read pages from leaf timeline across all ancestors.
7270 100 : for (idx, lsns) in updated.iter().enumerate() {
7271 10000 : for (blknum, lsn) in lsns.iter().enumerate() {
7272 2 : // Skip empty mutations.
7273 10000 : if lsn.0 == 0 {
7274 3563 : continue;
7275 6437 : }
7276 6437 : println!("checking [{idx}][{blknum}] at {lsn}");
7277 6437 : test_key.field6 = blknum as u32;
7278 6437 : assert_eq!(
7279 6437 : tline.get(test_key, *lsn, &ctx).await?,
7280 6437 : test_img(&format!("{idx} {blknum} at {lsn}"))
7281 2 : );
7282 2 : }
7283 2 : }
7284 2 : Ok(())
7285 2 : }
7286 :
7287 : #[tokio::test]
7288 2 : async fn test_write_at_initdb_lsn_takes_optimization_code_path() -> anyhow::Result<()> {
7289 2 : let (tenant, ctx) = TenantHarness::create("test_empty_test_timeline_is_usable")
7290 2 : .await?
7291 2 : .load()
7292 2 : .await;
7293 2 :
7294 2 : let initdb_lsn = Lsn(0x20);
7295 2 : let utline = tenant
7296 2 : .create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)
7297 2 : .await?;
7298 2 : let tline = utline.raw_timeline().unwrap();
7299 2 :
7300 2 : // Spawn flush loop now so that we can set the `expect_initdb_optimization`
7301 2 : tline.maybe_spawn_flush_loop();
7302 2 :
7303 2 : // Make sure the timeline has the minimum set of required keys for operation.
7304 2 : // The only operation you can always do on an empty timeline is to `put` new data.
7305 2 : // Except if you `put` at `initdb_lsn`.
7306 2 : // In that case, there's an optimization to directly create image layers instead of delta layers.
7307 2 : // It uses `repartition()`, which assumes some keys to be present.
7308 2 : // Let's make sure the test timeline can handle that case.
7309 2 : {
7310 2 : let mut state = tline.flush_loop_state.lock().unwrap();
7311 2 : assert_eq!(
7312 2 : timeline::FlushLoopState::Running {
7313 2 : expect_initdb_optimization: false,
7314 2 : initdb_optimization_count: 0,
7315 2 : },
7316 2 : *state
7317 2 : );
7318 2 : *state = timeline::FlushLoopState::Running {
7319 2 : expect_initdb_optimization: true,
7320 2 : initdb_optimization_count: 0,
7321 2 : };
7322 2 : }
7323 2 :
7324 2 : // Make writes at the initdb_lsn. When we flush it below, it should be handled by the optimization.
7325 2 : // As explained above, the optimization requires some keys to be present.
7326 2 : // As per `create_empty_timeline` documentation, use init_empty to set them.
7327 2 : // This is what `create_test_timeline` does, by the way.
7328 2 : let mut modification = tline.begin_modification(initdb_lsn);
7329 2 : modification
7330 2 : .init_empty_test_timeline()
7331 2 : .context("init_empty_test_timeline")?;
7332 2 : modification
7333 2 : .commit(&ctx)
7334 2 : .await
7335 2 : .context("commit init_empty_test_timeline modification")?;
7336 2 :
7337 2 : // Do the flush. The flush code will check the expectations that we set above.
7338 2 : tline.freeze_and_flush().await?;
7339 2 :
7340 2 : // assert freeze_and_flush exercised the initdb optimization
7341 2 : {
7342 2 : let state = tline.flush_loop_state.lock().unwrap();
7343 2 : let timeline::FlushLoopState::Running {
7344 2 : expect_initdb_optimization,
7345 2 : initdb_optimization_count,
7346 2 : } = *state
7347 2 : else {
7348 2 : panic!("unexpected state: {:?}", *state);
7349 2 : };
7350 2 : assert!(expect_initdb_optimization);
7351 2 : assert!(initdb_optimization_count > 0);
7352 2 : }
7353 2 : Ok(())
7354 2 : }
7355 :
7356 : #[tokio::test]
7357 2 : async fn test_create_guard_crash() -> anyhow::Result<()> {
7358 2 : let name = "test_create_guard_crash";
7359 2 : let harness = TenantHarness::create(name).await?;
7360 2 : {
7361 2 : let (tenant, ctx) = harness.load().await;
7362 2 : let tline = tenant
7363 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
7364 2 : .await?;
7365 2 : // Leave the timeline ID in [`Tenant::timelines_creating`] to exclude attempting to create it again
7366 2 : let raw_tline = tline.raw_timeline().unwrap();
7367 2 : raw_tline
7368 2 : .shutdown(super::timeline::ShutdownMode::Hard)
7369 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))
7370 2 : .await;
7371 2 : std::mem::forget(tline);
7372 2 : }
7373 2 :
7374 2 : let (tenant, _) = harness.load().await;
7375 2 : match tenant.get_timeline(TIMELINE_ID, false) {
7376 2 : Ok(_) => panic!("timeline should've been removed during load"),
7377 2 : Err(e) => {
7378 2 : assert_eq!(
7379 2 : e,
7380 2 : GetTimelineError::NotFound {
7381 2 : tenant_id: tenant.tenant_shard_id,
7382 2 : timeline_id: TIMELINE_ID,
7383 2 : }
7384 2 : )
7385 2 : }
7386 2 : }
7387 2 :
7388 2 : assert!(!harness
7389 2 : .conf
7390 2 : .timeline_path(&tenant.tenant_shard_id, &TIMELINE_ID)
7391 2 : .exists());
7392 2 :
7393 2 : Ok(())
7394 2 : }
7395 :
7396 : #[tokio::test]
7397 2 : async fn test_read_at_max_lsn() -> anyhow::Result<()> {
7398 2 : let names_algorithms = [
7399 2 : ("test_read_at_max_lsn_legacy", CompactionAlgorithm::Legacy),
7400 2 : ("test_read_at_max_lsn_tiered", CompactionAlgorithm::Tiered),
7401 2 : ];
7402 6 : for (name, algorithm) in names_algorithms {
7403 4 : test_read_at_max_lsn_algorithm(name, algorithm).await?;
7404 2 : }
7405 2 : Ok(())
7406 2 : }
7407 :
7408 4 : async fn test_read_at_max_lsn_algorithm(
7409 4 : name: &'static str,
7410 4 : compaction_algorithm: CompactionAlgorithm,
7411 4 : ) -> anyhow::Result<()> {
7412 4 : let mut harness = TenantHarness::create(name).await?;
7413 4 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
7414 4 : kind: compaction_algorithm,
7415 4 : };
7416 4 : let (tenant, ctx) = harness.load().await;
7417 4 : let tline = tenant
7418 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
7419 4 : .await?;
7420 :
7421 4 : let lsn = Lsn(0x10);
7422 4 : let compact = false;
7423 4 : bulk_insert_maybe_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000, compact).await?;
7424 :
7425 4 : let test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7426 4 : let read_lsn = Lsn(u64::MAX - 1);
7427 :
7428 4 : let result = tline.get(test_key, read_lsn, &ctx).await;
7429 4 : assert!(result.is_ok(), "result is not Ok: {}", result.unwrap_err());
7430 :
7431 4 : Ok(())
7432 4 : }
7433 :
7434 : #[tokio::test]
7435 2 : async fn test_metadata_scan() -> anyhow::Result<()> {
7436 2 : let harness = TenantHarness::create("test_metadata_scan").await?;
7437 2 : let (tenant, ctx) = harness.load().await;
7438 2 : let tline = tenant
7439 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7440 2 : .await?;
7441 2 :
7442 2 : const NUM_KEYS: usize = 1000;
7443 2 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
7444 2 :
7445 2 : let cancel = CancellationToken::new();
7446 2 :
7447 2 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7448 2 : base_key.field1 = AUX_KEY_PREFIX;
7449 2 : let mut test_key = base_key;
7450 2 :
7451 2 : // Track when each page was last modified. Used to assert that
7452 2 : // a read sees the latest page version.
7453 2 : let mut updated = [Lsn(0); NUM_KEYS];
7454 2 :
7455 2 : let mut lsn = Lsn(0x10);
7456 2 : #[allow(clippy::needless_range_loop)]
7457 2002 : for blknum in 0..NUM_KEYS {
7458 2000 : lsn = Lsn(lsn.0 + 0x10);
7459 2000 : test_key.field6 = (blknum * STEP) as u32;
7460 2000 : let mut writer = tline.writer().await;
7461 2000 : writer
7462 2000 : .put(
7463 2000 : test_key,
7464 2000 : lsn,
7465 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7466 2000 : &ctx,
7467 2000 : )
7468 2000 : .await?;
7469 2000 : writer.finish_write(lsn);
7470 2000 : updated[blknum] = lsn;
7471 2000 : drop(writer);
7472 2 : }
7473 2 :
7474 2 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
7475 2 :
7476 24 : for iter in 0..=10 {
7477 2 : // Read all the blocks
7478 22000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7479 22000 : test_key.field6 = (blknum * STEP) as u32;
7480 22000 : assert_eq!(
7481 22000 : tline.get(test_key, lsn, &ctx).await?,
7482 22000 : test_img(&format!("{} at {}", blknum, last_lsn))
7483 2 : );
7484 2 : }
7485 2 :
7486 22 : let mut cnt = 0;
7487 22000 : for (key, value) in tline
7488 22 : .get_vectored_impl(
7489 22 : keyspace.clone(),
7490 22 : lsn,
7491 22 : &mut ValuesReconstructState::default(),
7492 22 : &ctx,
7493 22 : )
7494 22 : .await?
7495 2 : {
7496 22000 : let blknum = key.field6 as usize;
7497 22000 : let value = value?;
7498 22000 : assert!(blknum % STEP == 0);
7499 22000 : let blknum = blknum / STEP;
7500 22000 : assert_eq!(
7501 22000 : value,
7502 22000 : test_img(&format!("{} at {}", blknum, updated[blknum]))
7503 22000 : );
7504 22000 : cnt += 1;
7505 2 : }
7506 2 :
7507 22 : assert_eq!(cnt, NUM_KEYS);
7508 2 :
7509 22022 : for _ in 0..NUM_KEYS {
7510 22000 : lsn = Lsn(lsn.0 + 0x10);
7511 22000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7512 22000 : test_key.field6 = (blknum * STEP) as u32;
7513 22000 : let mut writer = tline.writer().await;
7514 22000 : writer
7515 22000 : .put(
7516 22000 : test_key,
7517 22000 : lsn,
7518 22000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7519 22000 : &ctx,
7520 22000 : )
7521 22000 : .await?;
7522 22000 : writer.finish_write(lsn);
7523 22000 : drop(writer);
7524 22000 : updated[blknum] = lsn;
7525 2 : }
7526 2 :
7527 2 : // Perform two cycles of flush, compact, and GC
7528 66 : for round in 0..2 {
7529 44 : tline.freeze_and_flush().await?;
7530 44 : tline
7531 44 : .compact(
7532 44 : &cancel,
7533 44 : if iter % 5 == 0 && round == 0 {
7534 6 : let mut flags = EnumSet::new();
7535 6 : flags.insert(CompactFlags::ForceImageLayerCreation);
7536 6 : flags.insert(CompactFlags::ForceRepartition);
7537 6 : flags
7538 2 : } else {
7539 38 : EnumSet::empty()
7540 2 : },
7541 44 : &ctx,
7542 44 : )
7543 44 : .await?;
7544 44 : tenant
7545 44 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7546 44 : .await?;
7547 2 : }
7548 2 : }
7549 2 :
7550 2 : Ok(())
7551 2 : }
7552 :
7553 : #[tokio::test]
7554 2 : async fn test_metadata_compaction_trigger() -> anyhow::Result<()> {
7555 2 : let harness = TenantHarness::create("test_metadata_compaction_trigger").await?;
7556 2 : let (tenant, ctx) = harness.load().await;
7557 2 : let tline = tenant
7558 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7559 2 : .await?;
7560 2 :
7561 2 : let cancel = CancellationToken::new();
7562 2 :
7563 2 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7564 2 : base_key.field1 = AUX_KEY_PREFIX;
7565 2 : let test_key = base_key;
7566 2 : let mut lsn = Lsn(0x10);
7567 2 :
7568 42 : for _ in 0..20 {
7569 40 : lsn = Lsn(lsn.0 + 0x10);
7570 40 : let mut writer = tline.writer().await;
7571 40 : writer
7572 40 : .put(
7573 40 : test_key,
7574 40 : lsn,
7575 40 : &Value::Image(test_img(&format!("{} at {}", 0, lsn))),
7576 40 : &ctx,
7577 40 : )
7578 40 : .await?;
7579 40 : writer.finish_write(lsn);
7580 40 : drop(writer);
7581 40 : tline.freeze_and_flush().await?; // force create a delta layer
7582 2 : }
7583 2 :
7584 2 : let before_num_l0_delta_files =
7585 2 : tline.layers.read().await.layer_map()?.level0_deltas().len();
7586 2 :
7587 2 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
7588 2 :
7589 2 : let after_num_l0_delta_files = tline.layers.read().await.layer_map()?.level0_deltas().len();
7590 2 :
7591 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}");
7592 2 :
7593 2 : assert_eq!(
7594 2 : tline.get(test_key, lsn, &ctx).await?,
7595 2 : test_img(&format!("{} at {}", 0, lsn))
7596 2 : );
7597 2 :
7598 2 : Ok(())
7599 2 : }
7600 :
7601 : #[tokio::test]
7602 2 : async fn test_aux_file_e2e() {
7603 2 : let harness = TenantHarness::create("test_aux_file_e2e").await.unwrap();
7604 2 :
7605 2 : let (tenant, ctx) = harness.load().await;
7606 2 :
7607 2 : let mut lsn = Lsn(0x08);
7608 2 :
7609 2 : let tline: Arc<Timeline> = tenant
7610 2 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
7611 2 : .await
7612 2 : .unwrap();
7613 2 :
7614 2 : {
7615 2 : lsn += 8;
7616 2 : let mut modification = tline.begin_modification(lsn);
7617 2 : modification
7618 2 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
7619 2 : .await
7620 2 : .unwrap();
7621 2 : modification.commit(&ctx).await.unwrap();
7622 2 : }
7623 2 :
7624 2 : // we can read everything from the storage
7625 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
7626 2 : assert_eq!(
7627 2 : files.get("pg_logical/mappings/test1"),
7628 2 : Some(&bytes::Bytes::from_static(b"first"))
7629 2 : );
7630 2 :
7631 2 : {
7632 2 : lsn += 8;
7633 2 : let mut modification = tline.begin_modification(lsn);
7634 2 : modification
7635 2 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
7636 2 : .await
7637 2 : .unwrap();
7638 2 : modification.commit(&ctx).await.unwrap();
7639 2 : }
7640 2 :
7641 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
7642 2 : assert_eq!(
7643 2 : files.get("pg_logical/mappings/test2"),
7644 2 : Some(&bytes::Bytes::from_static(b"second"))
7645 2 : );
7646 2 :
7647 2 : let child = tenant
7648 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(lsn), &ctx)
7649 2 : .await
7650 2 : .unwrap();
7651 2 :
7652 2 : let files = child.list_aux_files(lsn, &ctx).await.unwrap();
7653 2 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
7654 2 : assert_eq!(files.get("pg_logical/mappings/test2"), None);
7655 2 : }
7656 :
7657 : #[tokio::test]
7658 2 : async fn test_metadata_image_creation() -> anyhow::Result<()> {
7659 2 : let harness = TenantHarness::create("test_metadata_image_creation").await?;
7660 2 : let (tenant, ctx) = harness.load().await;
7661 2 : let tline = tenant
7662 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7663 2 : .await?;
7664 2 :
7665 2 : const NUM_KEYS: usize = 1000;
7666 2 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
7667 2 :
7668 2 : let cancel = CancellationToken::new();
7669 2 :
7670 2 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7671 2 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
7672 2 : let mut test_key = base_key;
7673 2 : let mut lsn = Lsn(0x10);
7674 2 :
7675 8 : async fn scan_with_statistics(
7676 8 : tline: &Timeline,
7677 8 : keyspace: &KeySpace,
7678 8 : lsn: Lsn,
7679 8 : ctx: &RequestContext,
7680 8 : ) -> anyhow::Result<(BTreeMap<Key, Result<Bytes, PageReconstructError>>, usize)> {
7681 8 : let mut reconstruct_state = ValuesReconstructState::default();
7682 8 : let res = tline
7683 8 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
7684 8 : .await?;
7685 8 : Ok((res, reconstruct_state.get_delta_layers_visited() as usize))
7686 8 : }
7687 2 :
7688 2 : #[allow(clippy::needless_range_loop)]
7689 2002 : for blknum in 0..NUM_KEYS {
7690 2000 : lsn = Lsn(lsn.0 + 0x10);
7691 2000 : test_key.field6 = (blknum * STEP) as u32;
7692 2000 : let mut writer = tline.writer().await;
7693 2000 : writer
7694 2000 : .put(
7695 2000 : test_key,
7696 2000 : lsn,
7697 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7698 2000 : &ctx,
7699 2000 : )
7700 2000 : .await?;
7701 2000 : writer.finish_write(lsn);
7702 2000 : drop(writer);
7703 2 : }
7704 2 :
7705 2 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
7706 2 :
7707 22 : for iter in 1..=10 {
7708 20020 : for _ in 0..NUM_KEYS {
7709 20000 : lsn = Lsn(lsn.0 + 0x10);
7710 20000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7711 20000 : test_key.field6 = (blknum * STEP) as u32;
7712 20000 : let mut writer = tline.writer().await;
7713 20000 : writer
7714 20000 : .put(
7715 20000 : test_key,
7716 20000 : lsn,
7717 20000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7718 20000 : &ctx,
7719 20000 : )
7720 20000 : .await?;
7721 20000 : writer.finish_write(lsn);
7722 20000 : drop(writer);
7723 2 : }
7724 2 :
7725 20 : tline.freeze_and_flush().await?;
7726 2 :
7727 20 : if iter % 5 == 0 {
7728 4 : let (_, before_delta_file_accessed) =
7729 4 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
7730 4 : tline
7731 4 : .compact(
7732 4 : &cancel,
7733 4 : {
7734 4 : let mut flags = EnumSet::new();
7735 4 : flags.insert(CompactFlags::ForceImageLayerCreation);
7736 4 : flags.insert(CompactFlags::ForceRepartition);
7737 4 : flags
7738 4 : },
7739 4 : &ctx,
7740 4 : )
7741 4 : .await?;
7742 4 : let (_, after_delta_file_accessed) =
7743 4 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
7744 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}");
7745 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.
7746 4 : assert!(
7747 4 : after_delta_file_accessed <= 2,
7748 2 : "after_delta_file_accessed={after_delta_file_accessed}"
7749 2 : );
7750 16 : }
7751 2 : }
7752 2 :
7753 2 : Ok(())
7754 2 : }
7755 :
7756 : #[tokio::test]
7757 2 : async fn test_vectored_missing_data_key_reads() -> anyhow::Result<()> {
7758 2 : let harness = TenantHarness::create("test_vectored_missing_data_key_reads").await?;
7759 2 : let (tenant, ctx) = harness.load().await;
7760 2 :
7761 2 : let base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7762 2 : let base_key_child = Key::from_hex("000000000033333333444444445500000001").unwrap();
7763 2 : let base_key_nonexist = Key::from_hex("000000000033333333444444445500000002").unwrap();
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![(Lsn(0x20), vec![(base_key, test_img("data key 1"))])], // image layers
7773 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
7774 2 : )
7775 2 : .await?;
7776 2 : tline.add_extra_test_dense_keyspace(KeySpace::single(base_key..(base_key_nonexist.next())));
7777 2 :
7778 2 : let child = tenant
7779 2 : .branch_timeline_test_with_layers(
7780 2 : &tline,
7781 2 : NEW_TIMELINE_ID,
7782 2 : Some(Lsn(0x20)),
7783 2 : &ctx,
7784 2 : Vec::new(), // delta layers
7785 2 : vec![(Lsn(0x30), vec![(base_key_child, test_img("data key 2"))])], // image layers
7786 2 : Lsn(0x30),
7787 2 : )
7788 2 : .await
7789 2 : .unwrap();
7790 2 :
7791 2 : let lsn = Lsn(0x30);
7792 2 :
7793 2 : // test vectored get on parent timeline
7794 2 : assert_eq!(
7795 2 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
7796 2 : Some(test_img("data key 1"))
7797 2 : );
7798 2 : assert!(get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx)
7799 2 : .await
7800 2 : .unwrap_err()
7801 2 : .is_missing_key_error());
7802 2 : assert!(
7803 2 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx)
7804 2 : .await
7805 2 : .unwrap_err()
7806 2 : .is_missing_key_error()
7807 2 : );
7808 2 :
7809 2 : // test vectored get on child timeline
7810 2 : assert_eq!(
7811 2 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
7812 2 : Some(test_img("data key 1"))
7813 2 : );
7814 2 : assert_eq!(
7815 2 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
7816 2 : Some(test_img("data key 2"))
7817 2 : );
7818 2 : assert!(
7819 2 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx)
7820 2 : .await
7821 2 : .unwrap_err()
7822 2 : .is_missing_key_error()
7823 2 : );
7824 2 :
7825 2 : Ok(())
7826 2 : }
7827 :
7828 : #[tokio::test]
7829 2 : async fn test_vectored_missing_metadata_key_reads() -> anyhow::Result<()> {
7830 2 : let harness = TenantHarness::create("test_vectored_missing_metadata_key_reads").await?;
7831 2 : let (tenant, ctx) = harness.load().await;
7832 2 :
7833 2 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7834 2 : let base_key_child = Key::from_hex("620000000033333333444444445500000001").unwrap();
7835 2 : let base_key_nonexist = Key::from_hex("620000000033333333444444445500000002").unwrap();
7836 2 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
7837 2 :
7838 2 : let tline = tenant
7839 2 : .create_test_timeline_with_layers(
7840 2 : TIMELINE_ID,
7841 2 : Lsn(0x10),
7842 2 : DEFAULT_PG_VERSION,
7843 2 : &ctx,
7844 2 : Vec::new(), // delta layers
7845 2 : vec![(Lsn(0x20), vec![(base_key, test_img("metadata key 1"))])], // image layers
7846 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
7847 2 : )
7848 2 : .await?;
7849 2 :
7850 2 : let child = tenant
7851 2 : .branch_timeline_test_with_layers(
7852 2 : &tline,
7853 2 : NEW_TIMELINE_ID,
7854 2 : Some(Lsn(0x20)),
7855 2 : &ctx,
7856 2 : Vec::new(), // delta layers
7857 2 : vec![(
7858 2 : Lsn(0x30),
7859 2 : vec![(base_key_child, test_img("metadata key 2"))],
7860 2 : )], // image layers
7861 2 : Lsn(0x30),
7862 2 : )
7863 2 : .await
7864 2 : .unwrap();
7865 2 :
7866 2 : let lsn = Lsn(0x30);
7867 2 :
7868 2 : // test vectored get on parent timeline
7869 2 : assert_eq!(
7870 2 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
7871 2 : Some(test_img("metadata key 1"))
7872 2 : );
7873 2 : assert_eq!(
7874 2 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx).await?,
7875 2 : None
7876 2 : );
7877 2 : assert_eq!(
7878 2 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx).await?,
7879 2 : None
7880 2 : );
7881 2 :
7882 2 : // test vectored get on child timeline
7883 2 : assert_eq!(
7884 2 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
7885 2 : None
7886 2 : );
7887 2 : assert_eq!(
7888 2 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
7889 2 : Some(test_img("metadata key 2"))
7890 2 : );
7891 2 : assert_eq!(
7892 2 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx).await?,
7893 2 : None
7894 2 : );
7895 2 :
7896 2 : Ok(())
7897 2 : }
7898 :
7899 36 : async fn get_vectored_impl_wrapper(
7900 36 : tline: &Arc<Timeline>,
7901 36 : key: Key,
7902 36 : lsn: Lsn,
7903 36 : ctx: &RequestContext,
7904 36 : ) -> Result<Option<Bytes>, GetVectoredError> {
7905 36 : let mut reconstruct_state = ValuesReconstructState::new();
7906 36 : let mut res = tline
7907 36 : .get_vectored_impl(
7908 36 : KeySpace::single(key..key.next()),
7909 36 : lsn,
7910 36 : &mut reconstruct_state,
7911 36 : ctx,
7912 36 : )
7913 36 : .await?;
7914 30 : Ok(res.pop_last().map(|(k, v)| {
7915 18 : assert_eq!(k, key);
7916 18 : v.unwrap()
7917 30 : }))
7918 36 : }
7919 :
7920 : #[tokio::test]
7921 2 : async fn test_metadata_tombstone_reads() -> anyhow::Result<()> {
7922 2 : let harness = TenantHarness::create("test_metadata_tombstone_reads").await?;
7923 2 : let (tenant, ctx) = harness.load().await;
7924 2 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
7925 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
7926 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
7927 2 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
7928 2 :
7929 2 : // We emulate the situation that the compaction algorithm creates an image layer that removes the tombstones
7930 2 : // Lsn 0x30 key0, key3, no key1+key2
7931 2 : // Lsn 0x20 key1+key2 tomestones
7932 2 : // Lsn 0x10 key1 in image, key2 in delta
7933 2 : let tline = tenant
7934 2 : .create_test_timeline_with_layers(
7935 2 : TIMELINE_ID,
7936 2 : Lsn(0x10),
7937 2 : DEFAULT_PG_VERSION,
7938 2 : &ctx,
7939 2 : // delta layers
7940 2 : vec![
7941 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7942 2 : Lsn(0x10)..Lsn(0x20),
7943 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
7944 2 : ),
7945 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7946 2 : Lsn(0x20)..Lsn(0x30),
7947 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
7948 2 : ),
7949 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7950 2 : Lsn(0x20)..Lsn(0x30),
7951 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
7952 2 : ),
7953 2 : ],
7954 2 : // image layers
7955 2 : vec![
7956 2 : (Lsn(0x10), vec![(key1, test_img("metadata key 1"))]),
7957 2 : (
7958 2 : Lsn(0x30),
7959 2 : vec![
7960 2 : (key0, test_img("metadata key 0")),
7961 2 : (key3, test_img("metadata key 3")),
7962 2 : ],
7963 2 : ),
7964 2 : ],
7965 2 : Lsn(0x30),
7966 2 : )
7967 2 : .await?;
7968 2 :
7969 2 : let lsn = Lsn(0x30);
7970 2 : let old_lsn = Lsn(0x20);
7971 2 :
7972 2 : assert_eq!(
7973 2 : get_vectored_impl_wrapper(&tline, key0, lsn, &ctx).await?,
7974 2 : Some(test_img("metadata key 0"))
7975 2 : );
7976 2 : assert_eq!(
7977 2 : get_vectored_impl_wrapper(&tline, key1, lsn, &ctx).await?,
7978 2 : None,
7979 2 : );
7980 2 : assert_eq!(
7981 2 : get_vectored_impl_wrapper(&tline, key2, lsn, &ctx).await?,
7982 2 : None,
7983 2 : );
7984 2 : assert_eq!(
7985 2 : get_vectored_impl_wrapper(&tline, key1, old_lsn, &ctx).await?,
7986 2 : Some(Bytes::new()),
7987 2 : );
7988 2 : assert_eq!(
7989 2 : get_vectored_impl_wrapper(&tline, key2, old_lsn, &ctx).await?,
7990 2 : Some(Bytes::new()),
7991 2 : );
7992 2 : assert_eq!(
7993 2 : get_vectored_impl_wrapper(&tline, key3, lsn, &ctx).await?,
7994 2 : Some(test_img("metadata key 3"))
7995 2 : );
7996 2 :
7997 2 : Ok(())
7998 2 : }
7999 :
8000 : #[tokio::test]
8001 2 : async fn test_metadata_tombstone_image_creation() {
8002 2 : let harness = TenantHarness::create("test_metadata_tombstone_image_creation")
8003 2 : .await
8004 2 : .unwrap();
8005 2 : let (tenant, ctx) = harness.load().await;
8006 2 :
8007 2 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
8008 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8009 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8010 2 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
8011 2 :
8012 2 : let tline = tenant
8013 2 : .create_test_timeline_with_layers(
8014 2 : TIMELINE_ID,
8015 2 : Lsn(0x10),
8016 2 : DEFAULT_PG_VERSION,
8017 2 : &ctx,
8018 2 : // delta layers
8019 2 : vec![
8020 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8021 2 : Lsn(0x10)..Lsn(0x20),
8022 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8023 2 : ),
8024 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8025 2 : Lsn(0x20)..Lsn(0x30),
8026 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8027 2 : ),
8028 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8029 2 : Lsn(0x20)..Lsn(0x30),
8030 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8031 2 : ),
8032 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8033 2 : Lsn(0x30)..Lsn(0x40),
8034 2 : vec![
8035 2 : (key0, Lsn(0x30), Value::Image(test_img("metadata key 0"))),
8036 2 : (key3, Lsn(0x30), Value::Image(test_img("metadata key 3"))),
8037 2 : ],
8038 2 : ),
8039 2 : ],
8040 2 : // image layers
8041 2 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
8042 2 : Lsn(0x40),
8043 2 : )
8044 2 : .await
8045 2 : .unwrap();
8046 2 :
8047 2 : let cancel = CancellationToken::new();
8048 2 :
8049 2 : tline
8050 2 : .compact(
8051 2 : &cancel,
8052 2 : {
8053 2 : let mut flags = EnumSet::new();
8054 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
8055 2 : flags.insert(CompactFlags::ForceRepartition);
8056 2 : flags
8057 2 : },
8058 2 : &ctx,
8059 2 : )
8060 2 : .await
8061 2 : .unwrap();
8062 2 :
8063 2 : // Image layers are created at last_record_lsn
8064 2 : let images = tline
8065 2 : .inspect_image_layers(Lsn(0x40), &ctx)
8066 2 : .await
8067 2 : .unwrap()
8068 2 : .into_iter()
8069 18 : .filter(|(k, _)| k.is_metadata_key())
8070 2 : .collect::<Vec<_>>();
8071 2 : assert_eq!(images.len(), 2); // the image layer should only contain two existing keys, tombstones should be removed.
8072 2 : }
8073 :
8074 : #[tokio::test]
8075 2 : async fn test_metadata_tombstone_empty_image_creation() {
8076 2 : let harness = TenantHarness::create("test_metadata_tombstone_empty_image_creation")
8077 2 : .await
8078 2 : .unwrap();
8079 2 : let (tenant, ctx) = harness.load().await;
8080 2 :
8081 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8082 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8083 2 :
8084 2 : let tline = tenant
8085 2 : .create_test_timeline_with_layers(
8086 2 : TIMELINE_ID,
8087 2 : Lsn(0x10),
8088 2 : DEFAULT_PG_VERSION,
8089 2 : &ctx,
8090 2 : // delta layers
8091 2 : vec![
8092 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8093 2 : Lsn(0x10)..Lsn(0x20),
8094 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8095 2 : ),
8096 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8097 2 : Lsn(0x20)..Lsn(0x30),
8098 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8099 2 : ),
8100 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8101 2 : Lsn(0x20)..Lsn(0x30),
8102 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8103 2 : ),
8104 2 : ],
8105 2 : // image layers
8106 2 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
8107 2 : Lsn(0x30),
8108 2 : )
8109 2 : .await
8110 2 : .unwrap();
8111 2 :
8112 2 : let cancel = CancellationToken::new();
8113 2 :
8114 2 : tline
8115 2 : .compact(
8116 2 : &cancel,
8117 2 : {
8118 2 : let mut flags = EnumSet::new();
8119 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
8120 2 : flags.insert(CompactFlags::ForceRepartition);
8121 2 : flags
8122 2 : },
8123 2 : &ctx,
8124 2 : )
8125 2 : .await
8126 2 : .unwrap();
8127 2 :
8128 2 : // Image layers are created at last_record_lsn
8129 2 : let images = tline
8130 2 : .inspect_image_layers(Lsn(0x30), &ctx)
8131 2 : .await
8132 2 : .unwrap()
8133 2 : .into_iter()
8134 14 : .filter(|(k, _)| k.is_metadata_key())
8135 2 : .collect::<Vec<_>>();
8136 2 : assert_eq!(images.len(), 0); // the image layer should not contain tombstones, or it is not created
8137 2 : }
8138 :
8139 : #[tokio::test]
8140 2 : async fn test_simple_bottom_most_compaction_images() -> anyhow::Result<()> {
8141 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_images").await?;
8142 2 : let (tenant, ctx) = harness.load().await;
8143 2 :
8144 102 : fn get_key(id: u32) -> Key {
8145 102 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8146 102 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8147 102 : key.field6 = id;
8148 102 : key
8149 102 : }
8150 2 :
8151 2 : // We create
8152 2 : // - one bottom-most image layer,
8153 2 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
8154 2 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
8155 2 : // - a delta layer D3 above the horizon.
8156 2 : //
8157 2 : // | D3 |
8158 2 : // | D1 |
8159 2 : // -| |-- gc horizon -----------------
8160 2 : // | | | D2 |
8161 2 : // --------- img layer ------------------
8162 2 : //
8163 2 : // What we should expact from this compaction is:
8164 2 : // | D3 |
8165 2 : // | Part of D1 |
8166 2 : // --------- img layer with D1+D2 at GC horizon------------------
8167 2 :
8168 2 : // img layer at 0x10
8169 2 : let img_layer = (0..10)
8170 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8171 2 : .collect_vec();
8172 2 :
8173 2 : let delta1 = vec![
8174 2 : (
8175 2 : get_key(1),
8176 2 : Lsn(0x20),
8177 2 : Value::Image(Bytes::from("value 1@0x20")),
8178 2 : ),
8179 2 : (
8180 2 : get_key(2),
8181 2 : Lsn(0x30),
8182 2 : Value::Image(Bytes::from("value 2@0x30")),
8183 2 : ),
8184 2 : (
8185 2 : get_key(3),
8186 2 : Lsn(0x40),
8187 2 : Value::Image(Bytes::from("value 3@0x40")),
8188 2 : ),
8189 2 : ];
8190 2 : let delta2 = vec![
8191 2 : (
8192 2 : get_key(5),
8193 2 : Lsn(0x20),
8194 2 : Value::Image(Bytes::from("value 5@0x20")),
8195 2 : ),
8196 2 : (
8197 2 : get_key(6),
8198 2 : Lsn(0x20),
8199 2 : Value::Image(Bytes::from("value 6@0x20")),
8200 2 : ),
8201 2 : ];
8202 2 : let delta3 = vec![
8203 2 : (
8204 2 : get_key(8),
8205 2 : Lsn(0x48),
8206 2 : Value::Image(Bytes::from("value 8@0x48")),
8207 2 : ),
8208 2 : (
8209 2 : get_key(9),
8210 2 : Lsn(0x48),
8211 2 : Value::Image(Bytes::from("value 9@0x48")),
8212 2 : ),
8213 2 : ];
8214 2 :
8215 2 : let tline = tenant
8216 2 : .create_test_timeline_with_layers(
8217 2 : TIMELINE_ID,
8218 2 : Lsn(0x10),
8219 2 : DEFAULT_PG_VERSION,
8220 2 : &ctx,
8221 2 : vec![
8222 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
8223 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
8224 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
8225 2 : ], // delta layers
8226 2 : vec![(Lsn(0x10), img_layer)], // image layers
8227 2 : Lsn(0x50),
8228 2 : )
8229 2 : .await?;
8230 2 : {
8231 2 : tline
8232 2 : .latest_gc_cutoff_lsn
8233 2 : .lock_for_write()
8234 2 : .store_and_unlock(Lsn(0x30))
8235 2 : .wait()
8236 2 : .await;
8237 2 : // Update GC info
8238 2 : let mut guard = tline.gc_info.write().unwrap();
8239 2 : guard.cutoffs.time = Lsn(0x30);
8240 2 : guard.cutoffs.space = Lsn(0x30);
8241 2 : }
8242 2 :
8243 2 : let expected_result = [
8244 2 : Bytes::from_static(b"value 0@0x10"),
8245 2 : Bytes::from_static(b"value 1@0x20"),
8246 2 : Bytes::from_static(b"value 2@0x30"),
8247 2 : Bytes::from_static(b"value 3@0x40"),
8248 2 : Bytes::from_static(b"value 4@0x10"),
8249 2 : Bytes::from_static(b"value 5@0x20"),
8250 2 : Bytes::from_static(b"value 6@0x20"),
8251 2 : Bytes::from_static(b"value 7@0x10"),
8252 2 : Bytes::from_static(b"value 8@0x48"),
8253 2 : Bytes::from_static(b"value 9@0x48"),
8254 2 : ];
8255 2 :
8256 20 : for (idx, expected) in expected_result.iter().enumerate() {
8257 20 : assert_eq!(
8258 20 : tline
8259 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8260 20 : .await
8261 20 : .unwrap(),
8262 2 : expected
8263 2 : );
8264 2 : }
8265 2 :
8266 2 : let cancel = CancellationToken::new();
8267 2 : tline
8268 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8269 2 : .await
8270 2 : .unwrap();
8271 2 :
8272 20 : for (idx, expected) in expected_result.iter().enumerate() {
8273 20 : assert_eq!(
8274 20 : tline
8275 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8276 20 : .await
8277 20 : .unwrap(),
8278 2 : expected
8279 2 : );
8280 2 : }
8281 2 :
8282 2 : // Check if the image layer at the GC horizon contains exactly what we want
8283 2 : let image_at_gc_horizon = tline
8284 2 : .inspect_image_layers(Lsn(0x30), &ctx)
8285 2 : .await
8286 2 : .unwrap()
8287 2 : .into_iter()
8288 34 : .filter(|(k, _)| k.is_metadata_key())
8289 2 : .collect::<Vec<_>>();
8290 2 :
8291 2 : assert_eq!(image_at_gc_horizon.len(), 10);
8292 2 : let expected_result = [
8293 2 : Bytes::from_static(b"value 0@0x10"),
8294 2 : Bytes::from_static(b"value 1@0x20"),
8295 2 : Bytes::from_static(b"value 2@0x30"),
8296 2 : Bytes::from_static(b"value 3@0x10"),
8297 2 : Bytes::from_static(b"value 4@0x10"),
8298 2 : Bytes::from_static(b"value 5@0x20"),
8299 2 : Bytes::from_static(b"value 6@0x20"),
8300 2 : Bytes::from_static(b"value 7@0x10"),
8301 2 : Bytes::from_static(b"value 8@0x10"),
8302 2 : Bytes::from_static(b"value 9@0x10"),
8303 2 : ];
8304 22 : for idx in 0..10 {
8305 20 : assert_eq!(
8306 20 : image_at_gc_horizon[idx],
8307 20 : (get_key(idx as u32), expected_result[idx].clone())
8308 20 : );
8309 2 : }
8310 2 :
8311 2 : // Check if old layers are removed / new layers have the expected LSN
8312 2 : let all_layers = inspect_and_sort(&tline, None).await;
8313 2 : assert_eq!(
8314 2 : all_layers,
8315 2 : vec![
8316 2 : // Image layer at GC horizon
8317 2 : PersistentLayerKey {
8318 2 : key_range: Key::MIN..Key::MAX,
8319 2 : lsn_range: Lsn(0x30)..Lsn(0x31),
8320 2 : is_delta: false
8321 2 : },
8322 2 : // The delta layer below the horizon
8323 2 : PersistentLayerKey {
8324 2 : key_range: get_key(3)..get_key(4),
8325 2 : lsn_range: Lsn(0x30)..Lsn(0x48),
8326 2 : is_delta: true
8327 2 : },
8328 2 : // The delta3 layer that should not be picked for the compaction
8329 2 : PersistentLayerKey {
8330 2 : key_range: get_key(8)..get_key(10),
8331 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
8332 2 : is_delta: true
8333 2 : }
8334 2 : ]
8335 2 : );
8336 2 :
8337 2 : // increase GC horizon and compact again
8338 2 : {
8339 2 : tline
8340 2 : .latest_gc_cutoff_lsn
8341 2 : .lock_for_write()
8342 2 : .store_and_unlock(Lsn(0x40))
8343 2 : .wait()
8344 2 : .await;
8345 2 : // Update GC info
8346 2 : let mut guard = tline.gc_info.write().unwrap();
8347 2 : guard.cutoffs.time = Lsn(0x40);
8348 2 : guard.cutoffs.space = Lsn(0x40);
8349 2 : }
8350 2 : tline
8351 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8352 2 : .await
8353 2 : .unwrap();
8354 2 :
8355 2 : Ok(())
8356 2 : }
8357 :
8358 : #[cfg(feature = "testing")]
8359 : #[tokio::test]
8360 2 : async fn test_neon_test_record() -> anyhow::Result<()> {
8361 2 : let harness = TenantHarness::create("test_neon_test_record").await?;
8362 2 : let (tenant, ctx) = harness.load().await;
8363 2 :
8364 24 : fn get_key(id: u32) -> Key {
8365 24 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8366 24 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8367 24 : key.field6 = id;
8368 24 : key
8369 24 : }
8370 2 :
8371 2 : let delta1 = vec![
8372 2 : (
8373 2 : get_key(1),
8374 2 : Lsn(0x20),
8375 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
8376 2 : ),
8377 2 : (
8378 2 : get_key(1),
8379 2 : Lsn(0x30),
8380 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
8381 2 : ),
8382 2 : (get_key(2), Lsn(0x10), Value::Image("0x10".into())),
8383 2 : (
8384 2 : get_key(2),
8385 2 : Lsn(0x20),
8386 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
8387 2 : ),
8388 2 : (
8389 2 : get_key(2),
8390 2 : Lsn(0x30),
8391 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
8392 2 : ),
8393 2 : (get_key(3), Lsn(0x10), Value::Image("0x10".into())),
8394 2 : (
8395 2 : get_key(3),
8396 2 : Lsn(0x20),
8397 2 : Value::WalRecord(NeonWalRecord::wal_clear("c")),
8398 2 : ),
8399 2 : (get_key(4), Lsn(0x10), Value::Image("0x10".into())),
8400 2 : (
8401 2 : get_key(4),
8402 2 : Lsn(0x20),
8403 2 : Value::WalRecord(NeonWalRecord::wal_init("i")),
8404 2 : ),
8405 2 : ];
8406 2 : let image1 = vec![(get_key(1), "0x10".into())];
8407 2 :
8408 2 : let tline = tenant
8409 2 : .create_test_timeline_with_layers(
8410 2 : TIMELINE_ID,
8411 2 : Lsn(0x10),
8412 2 : DEFAULT_PG_VERSION,
8413 2 : &ctx,
8414 2 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
8415 2 : Lsn(0x10)..Lsn(0x40),
8416 2 : delta1,
8417 2 : )], // delta layers
8418 2 : vec![(Lsn(0x10), image1)], // image layers
8419 2 : Lsn(0x50),
8420 2 : )
8421 2 : .await?;
8422 2 :
8423 2 : assert_eq!(
8424 2 : tline.get(get_key(1), Lsn(0x50), &ctx).await?,
8425 2 : Bytes::from_static(b"0x10,0x20,0x30")
8426 2 : );
8427 2 : assert_eq!(
8428 2 : tline.get(get_key(2), Lsn(0x50), &ctx).await?,
8429 2 : Bytes::from_static(b"0x10,0x20,0x30")
8430 2 : );
8431 2 :
8432 2 : // Need to remove the limit of "Neon WAL redo requires base image".
8433 2 :
8434 2 : // assert_eq!(tline.get(get_key(3), Lsn(0x50), &ctx).await?, Bytes::new());
8435 2 : // assert_eq!(tline.get(get_key(4), Lsn(0x50), &ctx).await?, Bytes::new());
8436 2 :
8437 2 : Ok(())
8438 2 : }
8439 :
8440 : #[tokio::test(start_paused = true)]
8441 2 : async fn test_lsn_lease() -> anyhow::Result<()> {
8442 2 : let (tenant, ctx) = TenantHarness::create("test_lsn_lease")
8443 2 : .await
8444 2 : .unwrap()
8445 2 : .load()
8446 2 : .await;
8447 2 : // Advance to the lsn lease deadline so that GC is not blocked by
8448 2 : // initial transition into AttachedSingle.
8449 2 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
8450 2 : tokio::time::resume();
8451 2 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
8452 2 :
8453 2 : let end_lsn = Lsn(0x100);
8454 2 : let image_layers = (0x20..=0x90)
8455 2 : .step_by(0x10)
8456 16 : .map(|n| {
8457 16 : (
8458 16 : Lsn(n),
8459 16 : vec![(key, test_img(&format!("data key at {:x}", n)))],
8460 16 : )
8461 16 : })
8462 2 : .collect();
8463 2 :
8464 2 : let timeline = tenant
8465 2 : .create_test_timeline_with_layers(
8466 2 : TIMELINE_ID,
8467 2 : Lsn(0x10),
8468 2 : DEFAULT_PG_VERSION,
8469 2 : &ctx,
8470 2 : Vec::new(),
8471 2 : image_layers,
8472 2 : end_lsn,
8473 2 : )
8474 2 : .await?;
8475 2 :
8476 2 : let leased_lsns = [0x30, 0x50, 0x70];
8477 2 : let mut leases = Vec::new();
8478 6 : leased_lsns.iter().for_each(|n| {
8479 6 : leases.push(
8480 6 : timeline
8481 6 : .init_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)
8482 6 : .expect("lease request should succeed"),
8483 6 : );
8484 6 : });
8485 2 :
8486 2 : let updated_lease_0 = timeline
8487 2 : .renew_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)
8488 2 : .expect("lease renewal should succeed");
8489 2 : assert_eq!(
8490 2 : updated_lease_0.valid_until, leases[0].valid_until,
8491 2 : " Renewing with shorter lease should not change the lease."
8492 2 : );
8493 2 :
8494 2 : let updated_lease_1 = timeline
8495 2 : .renew_lsn_lease(
8496 2 : Lsn(leased_lsns[1]),
8497 2 : timeline.get_lsn_lease_length() * 2,
8498 2 : &ctx,
8499 2 : )
8500 2 : .expect("lease renewal should succeed");
8501 2 : assert!(
8502 2 : updated_lease_1.valid_until > leases[1].valid_until,
8503 2 : "Renewing with a long lease should renew lease with later expiration time."
8504 2 : );
8505 2 :
8506 2 : // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
8507 2 : info!(
8508 2 : "latest_gc_cutoff_lsn: {}",
8509 0 : *timeline.get_latest_gc_cutoff_lsn()
8510 2 : );
8511 2 : timeline.force_set_disk_consistent_lsn(end_lsn);
8512 2 :
8513 2 : let res = tenant
8514 2 : .gc_iteration(
8515 2 : Some(TIMELINE_ID),
8516 2 : 0,
8517 2 : Duration::ZERO,
8518 2 : &CancellationToken::new(),
8519 2 : &ctx,
8520 2 : )
8521 2 : .await
8522 2 : .unwrap();
8523 2 :
8524 2 : // Keeping everything <= Lsn(0x80) b/c leases:
8525 2 : // 0/10: initdb layer
8526 2 : // (0/20..=0/70).step_by(0x10): image layers added when creating the timeline.
8527 2 : assert_eq!(res.layers_needed_by_leases, 7);
8528 2 : // Keeping 0/90 b/c it is the latest layer.
8529 2 : assert_eq!(res.layers_not_updated, 1);
8530 2 : // Removed 0/80.
8531 2 : assert_eq!(res.layers_removed, 1);
8532 2 :
8533 2 : // Make lease on a already GC-ed LSN.
8534 2 : // 0/80 does not have a valid lease + is below latest_gc_cutoff
8535 2 : assert!(Lsn(0x80) < *timeline.get_latest_gc_cutoff_lsn());
8536 2 : timeline
8537 2 : .init_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx)
8538 2 : .expect_err("lease request on GC-ed LSN should fail");
8539 2 :
8540 2 : // Should still be able to renew a currently valid lease
8541 2 : // Assumption: original lease to is still valid for 0/50.
8542 2 : // (use `Timeline::init_lsn_lease` for testing so it always does validation)
8543 2 : timeline
8544 2 : .init_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)
8545 2 : .expect("lease renewal with validation should succeed");
8546 2 :
8547 2 : Ok(())
8548 2 : }
8549 :
8550 : #[cfg(feature = "testing")]
8551 : #[tokio::test]
8552 2 : async fn test_simple_bottom_most_compaction_deltas_1() -> anyhow::Result<()> {
8553 2 : test_simple_bottom_most_compaction_deltas_helper(
8554 2 : "test_simple_bottom_most_compaction_deltas_1",
8555 2 : false,
8556 2 : )
8557 2 : .await
8558 2 : }
8559 :
8560 : #[cfg(feature = "testing")]
8561 : #[tokio::test]
8562 2 : async fn test_simple_bottom_most_compaction_deltas_2() -> anyhow::Result<()> {
8563 2 : test_simple_bottom_most_compaction_deltas_helper(
8564 2 : "test_simple_bottom_most_compaction_deltas_2",
8565 2 : true,
8566 2 : )
8567 2 : .await
8568 2 : }
8569 :
8570 : #[cfg(feature = "testing")]
8571 4 : async fn test_simple_bottom_most_compaction_deltas_helper(
8572 4 : test_name: &'static str,
8573 4 : use_delta_bottom_layer: bool,
8574 4 : ) -> anyhow::Result<()> {
8575 4 : let harness = TenantHarness::create(test_name).await?;
8576 4 : let (tenant, ctx) = harness.load().await;
8577 :
8578 276 : fn get_key(id: u32) -> Key {
8579 276 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8580 276 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8581 276 : key.field6 = id;
8582 276 : key
8583 276 : }
8584 :
8585 : // We create
8586 : // - one bottom-most image layer,
8587 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
8588 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
8589 : // - a delta layer D3 above the horizon.
8590 : //
8591 : // | D3 |
8592 : // | D1 |
8593 : // -| |-- gc horizon -----------------
8594 : // | | | D2 |
8595 : // --------- img layer ------------------
8596 : //
8597 : // What we should expact from this compaction is:
8598 : // | D3 |
8599 : // | Part of D1 |
8600 : // --------- img layer with D1+D2 at GC horizon------------------
8601 :
8602 : // img layer at 0x10
8603 4 : let img_layer = (0..10)
8604 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8605 4 : .collect_vec();
8606 4 : // or, delta layer at 0x10 if `use_delta_bottom_layer` is true
8607 4 : let delta4 = (0..10)
8608 40 : .map(|id| {
8609 40 : (
8610 40 : get_key(id),
8611 40 : Lsn(0x08),
8612 40 : Value::WalRecord(NeonWalRecord::wal_init(format!("value {id}@0x10"))),
8613 40 : )
8614 40 : })
8615 4 : .collect_vec();
8616 4 :
8617 4 : let delta1 = vec![
8618 4 : (
8619 4 : get_key(1),
8620 4 : Lsn(0x20),
8621 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8622 4 : ),
8623 4 : (
8624 4 : get_key(2),
8625 4 : Lsn(0x30),
8626 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8627 4 : ),
8628 4 : (
8629 4 : get_key(3),
8630 4 : Lsn(0x28),
8631 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8632 4 : ),
8633 4 : (
8634 4 : get_key(3),
8635 4 : Lsn(0x30),
8636 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8637 4 : ),
8638 4 : (
8639 4 : get_key(3),
8640 4 : Lsn(0x40),
8641 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
8642 4 : ),
8643 4 : ];
8644 4 : let delta2 = vec![
8645 4 : (
8646 4 : get_key(5),
8647 4 : Lsn(0x20),
8648 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8649 4 : ),
8650 4 : (
8651 4 : get_key(6),
8652 4 : Lsn(0x20),
8653 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8654 4 : ),
8655 4 : ];
8656 4 : let delta3 = vec![
8657 4 : (
8658 4 : get_key(8),
8659 4 : Lsn(0x48),
8660 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8661 4 : ),
8662 4 : (
8663 4 : get_key(9),
8664 4 : Lsn(0x48),
8665 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8666 4 : ),
8667 4 : ];
8668 :
8669 4 : let tline = if use_delta_bottom_layer {
8670 2 : tenant
8671 2 : .create_test_timeline_with_layers(
8672 2 : TIMELINE_ID,
8673 2 : Lsn(0x08),
8674 2 : DEFAULT_PG_VERSION,
8675 2 : &ctx,
8676 2 : vec![
8677 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8678 2 : Lsn(0x08)..Lsn(0x10),
8679 2 : delta4,
8680 2 : ),
8681 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8682 2 : Lsn(0x20)..Lsn(0x48),
8683 2 : delta1,
8684 2 : ),
8685 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8686 2 : Lsn(0x20)..Lsn(0x48),
8687 2 : delta2,
8688 2 : ),
8689 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8690 2 : Lsn(0x48)..Lsn(0x50),
8691 2 : delta3,
8692 2 : ),
8693 2 : ], // delta layers
8694 2 : vec![], // image layers
8695 2 : Lsn(0x50),
8696 2 : )
8697 2 : .await?
8698 : } else {
8699 2 : tenant
8700 2 : .create_test_timeline_with_layers(
8701 2 : TIMELINE_ID,
8702 2 : Lsn(0x10),
8703 2 : DEFAULT_PG_VERSION,
8704 2 : &ctx,
8705 2 : vec![
8706 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8707 2 : Lsn(0x10)..Lsn(0x48),
8708 2 : delta1,
8709 2 : ),
8710 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8711 2 : Lsn(0x10)..Lsn(0x48),
8712 2 : delta2,
8713 2 : ),
8714 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8715 2 : Lsn(0x48)..Lsn(0x50),
8716 2 : delta3,
8717 2 : ),
8718 2 : ], // delta layers
8719 2 : vec![(Lsn(0x10), img_layer)], // image layers
8720 2 : Lsn(0x50),
8721 2 : )
8722 2 : .await?
8723 : };
8724 : {
8725 4 : tline
8726 4 : .latest_gc_cutoff_lsn
8727 4 : .lock_for_write()
8728 4 : .store_and_unlock(Lsn(0x30))
8729 4 : .wait()
8730 4 : .await;
8731 : // Update GC info
8732 4 : let mut guard = tline.gc_info.write().unwrap();
8733 4 : *guard = GcInfo {
8734 4 : retain_lsns: vec![],
8735 4 : cutoffs: GcCutoffs {
8736 4 : time: Lsn(0x30),
8737 4 : space: Lsn(0x30),
8738 4 : },
8739 4 : leases: Default::default(),
8740 4 : within_ancestor_pitr: false,
8741 4 : };
8742 4 : }
8743 4 :
8744 4 : let expected_result = [
8745 4 : Bytes::from_static(b"value 0@0x10"),
8746 4 : Bytes::from_static(b"value 1@0x10@0x20"),
8747 4 : Bytes::from_static(b"value 2@0x10@0x30"),
8748 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
8749 4 : Bytes::from_static(b"value 4@0x10"),
8750 4 : Bytes::from_static(b"value 5@0x10@0x20"),
8751 4 : Bytes::from_static(b"value 6@0x10@0x20"),
8752 4 : Bytes::from_static(b"value 7@0x10"),
8753 4 : Bytes::from_static(b"value 8@0x10@0x48"),
8754 4 : Bytes::from_static(b"value 9@0x10@0x48"),
8755 4 : ];
8756 4 :
8757 4 : let expected_result_at_gc_horizon = [
8758 4 : Bytes::from_static(b"value 0@0x10"),
8759 4 : Bytes::from_static(b"value 1@0x10@0x20"),
8760 4 : Bytes::from_static(b"value 2@0x10@0x30"),
8761 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
8762 4 : Bytes::from_static(b"value 4@0x10"),
8763 4 : Bytes::from_static(b"value 5@0x10@0x20"),
8764 4 : Bytes::from_static(b"value 6@0x10@0x20"),
8765 4 : Bytes::from_static(b"value 7@0x10"),
8766 4 : Bytes::from_static(b"value 8@0x10"),
8767 4 : Bytes::from_static(b"value 9@0x10"),
8768 4 : ];
8769 :
8770 44 : for idx in 0..10 {
8771 40 : assert_eq!(
8772 40 : tline
8773 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8774 40 : .await
8775 40 : .unwrap(),
8776 40 : &expected_result[idx]
8777 : );
8778 40 : assert_eq!(
8779 40 : tline
8780 40 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
8781 40 : .await
8782 40 : .unwrap(),
8783 40 : &expected_result_at_gc_horizon[idx]
8784 : );
8785 : }
8786 :
8787 4 : let cancel = CancellationToken::new();
8788 4 : tline
8789 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8790 4 : .await
8791 4 : .unwrap();
8792 :
8793 44 : for idx in 0..10 {
8794 40 : assert_eq!(
8795 40 : tline
8796 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8797 40 : .await
8798 40 : .unwrap(),
8799 40 : &expected_result[idx]
8800 : );
8801 40 : assert_eq!(
8802 40 : tline
8803 40 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
8804 40 : .await
8805 40 : .unwrap(),
8806 40 : &expected_result_at_gc_horizon[idx]
8807 : );
8808 : }
8809 :
8810 : // increase GC horizon and compact again
8811 : {
8812 4 : tline
8813 4 : .latest_gc_cutoff_lsn
8814 4 : .lock_for_write()
8815 4 : .store_and_unlock(Lsn(0x40))
8816 4 : .wait()
8817 4 : .await;
8818 : // Update GC info
8819 4 : let mut guard = tline.gc_info.write().unwrap();
8820 4 : guard.cutoffs.time = Lsn(0x40);
8821 4 : guard.cutoffs.space = Lsn(0x40);
8822 4 : }
8823 4 : tline
8824 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8825 4 : .await
8826 4 : .unwrap();
8827 4 :
8828 4 : Ok(())
8829 4 : }
8830 :
8831 : #[cfg(feature = "testing")]
8832 : #[tokio::test]
8833 2 : async fn test_generate_key_retention() -> anyhow::Result<()> {
8834 2 : let harness = TenantHarness::create("test_generate_key_retention").await?;
8835 2 : let (tenant, ctx) = harness.load().await;
8836 2 : let tline = tenant
8837 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
8838 2 : .await?;
8839 2 : tline.force_advance_lsn(Lsn(0x70));
8840 2 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
8841 2 : let history = vec![
8842 2 : (
8843 2 : key,
8844 2 : Lsn(0x10),
8845 2 : Value::WalRecord(NeonWalRecord::wal_init("0x10")),
8846 2 : ),
8847 2 : (
8848 2 : key,
8849 2 : Lsn(0x20),
8850 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8851 2 : ),
8852 2 : (
8853 2 : key,
8854 2 : Lsn(0x30),
8855 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8856 2 : ),
8857 2 : (
8858 2 : key,
8859 2 : Lsn(0x40),
8860 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8861 2 : ),
8862 2 : (
8863 2 : key,
8864 2 : Lsn(0x50),
8865 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
8866 2 : ),
8867 2 : (
8868 2 : key,
8869 2 : Lsn(0x60),
8870 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8871 2 : ),
8872 2 : (
8873 2 : key,
8874 2 : Lsn(0x70),
8875 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8876 2 : ),
8877 2 : (
8878 2 : key,
8879 2 : Lsn(0x80),
8880 2 : Value::Image(Bytes::copy_from_slice(
8881 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8882 2 : )),
8883 2 : ),
8884 2 : (
8885 2 : key,
8886 2 : Lsn(0x90),
8887 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8888 2 : ),
8889 2 : ];
8890 2 : let res = tline
8891 2 : .generate_key_retention(
8892 2 : key,
8893 2 : &history,
8894 2 : Lsn(0x60),
8895 2 : &[Lsn(0x20), Lsn(0x40), Lsn(0x50)],
8896 2 : 3,
8897 2 : None,
8898 2 : )
8899 2 : .await
8900 2 : .unwrap();
8901 2 : let expected_res = KeyHistoryRetention {
8902 2 : below_horizon: vec![
8903 2 : (
8904 2 : Lsn(0x20),
8905 2 : KeyLogAtLsn(vec![(
8906 2 : Lsn(0x20),
8907 2 : Value::Image(Bytes::from_static(b"0x10;0x20")),
8908 2 : )]),
8909 2 : ),
8910 2 : (
8911 2 : Lsn(0x40),
8912 2 : KeyLogAtLsn(vec![
8913 2 : (
8914 2 : Lsn(0x30),
8915 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8916 2 : ),
8917 2 : (
8918 2 : Lsn(0x40),
8919 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8920 2 : ),
8921 2 : ]),
8922 2 : ),
8923 2 : (
8924 2 : Lsn(0x50),
8925 2 : KeyLogAtLsn(vec![(
8926 2 : Lsn(0x50),
8927 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40;0x50")),
8928 2 : )]),
8929 2 : ),
8930 2 : (
8931 2 : Lsn(0x60),
8932 2 : KeyLogAtLsn(vec![(
8933 2 : Lsn(0x60),
8934 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8935 2 : )]),
8936 2 : ),
8937 2 : ],
8938 2 : above_horizon: KeyLogAtLsn(vec![
8939 2 : (
8940 2 : Lsn(0x70),
8941 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8942 2 : ),
8943 2 : (
8944 2 : Lsn(0x80),
8945 2 : Value::Image(Bytes::copy_from_slice(
8946 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8947 2 : )),
8948 2 : ),
8949 2 : (
8950 2 : Lsn(0x90),
8951 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8952 2 : ),
8953 2 : ]),
8954 2 : };
8955 2 : assert_eq!(res, expected_res);
8956 2 :
8957 2 : // We expect GC-compaction to run with the original GC. This would create a situation that
8958 2 : // the original GC algorithm removes some delta layers b/c there are full image coverage,
8959 2 : // therefore causing some keys to have an incomplete history below the lowest retain LSN.
8960 2 : // For example, we have
8961 2 : // ```plain
8962 2 : // init delta @ 0x10, image @ 0x20, delta @ 0x30 (gc_horizon), image @ 0x40.
8963 2 : // ```
8964 2 : // Now the GC horizon moves up, and we have
8965 2 : // ```plain
8966 2 : // init delta @ 0x10, image @ 0x20, delta @ 0x30, image @ 0x40 (gc_horizon)
8967 2 : // ```
8968 2 : // The original GC algorithm kicks in, and removes delta @ 0x10, image @ 0x20.
8969 2 : // We will end up with
8970 2 : // ```plain
8971 2 : // delta @ 0x30, image @ 0x40 (gc_horizon)
8972 2 : // ```
8973 2 : // Now we run the GC-compaction, and this key does not have a full history.
8974 2 : // We should be able to handle this partial history and drop everything before the
8975 2 : // gc_horizon image.
8976 2 :
8977 2 : let history = vec![
8978 2 : (
8979 2 : key,
8980 2 : Lsn(0x20),
8981 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8982 2 : ),
8983 2 : (
8984 2 : key,
8985 2 : Lsn(0x30),
8986 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8987 2 : ),
8988 2 : (
8989 2 : key,
8990 2 : Lsn(0x40),
8991 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
8992 2 : ),
8993 2 : (
8994 2 : key,
8995 2 : Lsn(0x50),
8996 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
8997 2 : ),
8998 2 : (
8999 2 : key,
9000 2 : Lsn(0x60),
9001 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9002 2 : ),
9003 2 : (
9004 2 : key,
9005 2 : Lsn(0x70),
9006 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9007 2 : ),
9008 2 : (
9009 2 : key,
9010 2 : Lsn(0x80),
9011 2 : Value::Image(Bytes::copy_from_slice(
9012 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9013 2 : )),
9014 2 : ),
9015 2 : (
9016 2 : key,
9017 2 : Lsn(0x90),
9018 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9019 2 : ),
9020 2 : ];
9021 2 : let res = tline
9022 2 : .generate_key_retention(key, &history, Lsn(0x60), &[Lsn(0x40), Lsn(0x50)], 3, None)
9023 2 : .await
9024 2 : .unwrap();
9025 2 : let expected_res = KeyHistoryRetention {
9026 2 : below_horizon: vec![
9027 2 : (
9028 2 : Lsn(0x40),
9029 2 : KeyLogAtLsn(vec![(
9030 2 : Lsn(0x40),
9031 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
9032 2 : )]),
9033 2 : ),
9034 2 : (
9035 2 : Lsn(0x50),
9036 2 : KeyLogAtLsn(vec![(
9037 2 : Lsn(0x50),
9038 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
9039 2 : )]),
9040 2 : ),
9041 2 : (
9042 2 : Lsn(0x60),
9043 2 : KeyLogAtLsn(vec![(
9044 2 : Lsn(0x60),
9045 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9046 2 : )]),
9047 2 : ),
9048 2 : ],
9049 2 : above_horizon: KeyLogAtLsn(vec![
9050 2 : (
9051 2 : Lsn(0x70),
9052 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9053 2 : ),
9054 2 : (
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 : Lsn(0x90),
9062 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9063 2 : ),
9064 2 : ]),
9065 2 : };
9066 2 : assert_eq!(res, expected_res);
9067 2 :
9068 2 : // In case of branch compaction, the branch itself does not have the full history, and we need to provide
9069 2 : // the ancestor image in the test case.
9070 2 :
9071 2 : let history = vec![
9072 2 : (
9073 2 : key,
9074 2 : Lsn(0x20),
9075 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9076 2 : ),
9077 2 : (
9078 2 : key,
9079 2 : Lsn(0x30),
9080 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9081 2 : ),
9082 2 : (
9083 2 : key,
9084 2 : Lsn(0x40),
9085 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9086 2 : ),
9087 2 : (
9088 2 : key,
9089 2 : Lsn(0x70),
9090 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9091 2 : ),
9092 2 : ];
9093 2 : let res = tline
9094 2 : .generate_key_retention(
9095 2 : key,
9096 2 : &history,
9097 2 : Lsn(0x60),
9098 2 : &[],
9099 2 : 3,
9100 2 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
9101 2 : )
9102 2 : .await
9103 2 : .unwrap();
9104 2 : let expected_res = KeyHistoryRetention {
9105 2 : below_horizon: vec![(
9106 2 : Lsn(0x60),
9107 2 : KeyLogAtLsn(vec![(
9108 2 : Lsn(0x60),
9109 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")), // use the ancestor image to reconstruct the page
9110 2 : )]),
9111 2 : )],
9112 2 : above_horizon: KeyLogAtLsn(vec![(
9113 2 : Lsn(0x70),
9114 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9115 2 : )]),
9116 2 : };
9117 2 : assert_eq!(res, expected_res);
9118 2 :
9119 2 : let history = vec![
9120 2 : (
9121 2 : key,
9122 2 : Lsn(0x20),
9123 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9124 2 : ),
9125 2 : (
9126 2 : key,
9127 2 : Lsn(0x40),
9128 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9129 2 : ),
9130 2 : (
9131 2 : key,
9132 2 : Lsn(0x60),
9133 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9134 2 : ),
9135 2 : (
9136 2 : key,
9137 2 : Lsn(0x70),
9138 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9139 2 : ),
9140 2 : ];
9141 2 : let res = tline
9142 2 : .generate_key_retention(
9143 2 : key,
9144 2 : &history,
9145 2 : Lsn(0x60),
9146 2 : &[Lsn(0x30)],
9147 2 : 3,
9148 2 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
9149 2 : )
9150 2 : .await
9151 2 : .unwrap();
9152 2 : let expected_res = KeyHistoryRetention {
9153 2 : below_horizon: vec![
9154 2 : (
9155 2 : Lsn(0x30),
9156 2 : KeyLogAtLsn(vec![(
9157 2 : Lsn(0x20),
9158 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9159 2 : )]),
9160 2 : ),
9161 2 : (
9162 2 : Lsn(0x60),
9163 2 : KeyLogAtLsn(vec![(
9164 2 : Lsn(0x60),
9165 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x40;0x60")),
9166 2 : )]),
9167 2 : ),
9168 2 : ],
9169 2 : above_horizon: KeyLogAtLsn(vec![(
9170 2 : Lsn(0x70),
9171 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9172 2 : )]),
9173 2 : };
9174 2 : assert_eq!(res, expected_res);
9175 2 :
9176 2 : Ok(())
9177 2 : }
9178 :
9179 : #[cfg(feature = "testing")]
9180 : #[tokio::test]
9181 2 : async fn test_simple_bottom_most_compaction_with_retain_lsns() -> anyhow::Result<()> {
9182 2 : let harness =
9183 2 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns").await?;
9184 2 : let (tenant, ctx) = harness.load().await;
9185 2 :
9186 518 : fn get_key(id: u32) -> Key {
9187 518 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9188 518 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9189 518 : key.field6 = id;
9190 518 : key
9191 518 : }
9192 2 :
9193 2 : let img_layer = (0..10)
9194 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9195 2 : .collect_vec();
9196 2 :
9197 2 : let delta1 = vec![
9198 2 : (
9199 2 : get_key(1),
9200 2 : Lsn(0x20),
9201 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9202 2 : ),
9203 2 : (
9204 2 : get_key(2),
9205 2 : Lsn(0x30),
9206 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9207 2 : ),
9208 2 : (
9209 2 : get_key(3),
9210 2 : Lsn(0x28),
9211 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9212 2 : ),
9213 2 : (
9214 2 : get_key(3),
9215 2 : Lsn(0x30),
9216 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9217 2 : ),
9218 2 : (
9219 2 : get_key(3),
9220 2 : Lsn(0x40),
9221 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9222 2 : ),
9223 2 : ];
9224 2 : let delta2 = vec![
9225 2 : (
9226 2 : get_key(5),
9227 2 : Lsn(0x20),
9228 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9229 2 : ),
9230 2 : (
9231 2 : get_key(6),
9232 2 : Lsn(0x20),
9233 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9234 2 : ),
9235 2 : ];
9236 2 : let delta3 = vec![
9237 2 : (
9238 2 : get_key(8),
9239 2 : Lsn(0x48),
9240 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9241 2 : ),
9242 2 : (
9243 2 : get_key(9),
9244 2 : Lsn(0x48),
9245 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9246 2 : ),
9247 2 : ];
9248 2 :
9249 2 : let tline = tenant
9250 2 : .create_test_timeline_with_layers(
9251 2 : TIMELINE_ID,
9252 2 : Lsn(0x10),
9253 2 : DEFAULT_PG_VERSION,
9254 2 : &ctx,
9255 2 : vec![
9256 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
9257 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
9258 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
9259 2 : ], // delta layers
9260 2 : vec![(Lsn(0x10), img_layer)], // image layers
9261 2 : Lsn(0x50),
9262 2 : )
9263 2 : .await?;
9264 2 : {
9265 2 : tline
9266 2 : .latest_gc_cutoff_lsn
9267 2 : .lock_for_write()
9268 2 : .store_and_unlock(Lsn(0x30))
9269 2 : .wait()
9270 2 : .await;
9271 2 : // Update GC info
9272 2 : let mut guard = tline.gc_info.write().unwrap();
9273 2 : *guard = GcInfo {
9274 2 : retain_lsns: vec![
9275 2 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
9276 2 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
9277 2 : ],
9278 2 : cutoffs: GcCutoffs {
9279 2 : time: Lsn(0x30),
9280 2 : space: Lsn(0x30),
9281 2 : },
9282 2 : leases: Default::default(),
9283 2 : within_ancestor_pitr: false,
9284 2 : };
9285 2 : }
9286 2 :
9287 2 : let expected_result = [
9288 2 : Bytes::from_static(b"value 0@0x10"),
9289 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9290 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9291 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9292 2 : Bytes::from_static(b"value 4@0x10"),
9293 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9294 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9295 2 : Bytes::from_static(b"value 7@0x10"),
9296 2 : Bytes::from_static(b"value 8@0x10@0x48"),
9297 2 : Bytes::from_static(b"value 9@0x10@0x48"),
9298 2 : ];
9299 2 :
9300 2 : let expected_result_at_gc_horizon = [
9301 2 : Bytes::from_static(b"value 0@0x10"),
9302 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9303 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9304 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
9305 2 : Bytes::from_static(b"value 4@0x10"),
9306 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9307 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9308 2 : Bytes::from_static(b"value 7@0x10"),
9309 2 : Bytes::from_static(b"value 8@0x10"),
9310 2 : Bytes::from_static(b"value 9@0x10"),
9311 2 : ];
9312 2 :
9313 2 : let expected_result_at_lsn_20 = [
9314 2 : Bytes::from_static(b"value 0@0x10"),
9315 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9316 2 : Bytes::from_static(b"value 2@0x10"),
9317 2 : Bytes::from_static(b"value 3@0x10"),
9318 2 : Bytes::from_static(b"value 4@0x10"),
9319 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9320 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9321 2 : Bytes::from_static(b"value 7@0x10"),
9322 2 : Bytes::from_static(b"value 8@0x10"),
9323 2 : Bytes::from_static(b"value 9@0x10"),
9324 2 : ];
9325 2 :
9326 2 : let expected_result_at_lsn_10 = [
9327 2 : Bytes::from_static(b"value 0@0x10"),
9328 2 : Bytes::from_static(b"value 1@0x10"),
9329 2 : Bytes::from_static(b"value 2@0x10"),
9330 2 : Bytes::from_static(b"value 3@0x10"),
9331 2 : Bytes::from_static(b"value 4@0x10"),
9332 2 : Bytes::from_static(b"value 5@0x10"),
9333 2 : Bytes::from_static(b"value 6@0x10"),
9334 2 : Bytes::from_static(b"value 7@0x10"),
9335 2 : Bytes::from_static(b"value 8@0x10"),
9336 2 : Bytes::from_static(b"value 9@0x10"),
9337 2 : ];
9338 2 :
9339 12 : let verify_result = || async {
9340 12 : let gc_horizon = {
9341 12 : let gc_info = tline.gc_info.read().unwrap();
9342 12 : gc_info.cutoffs.time
9343 2 : };
9344 132 : for idx in 0..10 {
9345 120 : assert_eq!(
9346 120 : tline
9347 120 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9348 120 : .await
9349 120 : .unwrap(),
9350 120 : &expected_result[idx]
9351 2 : );
9352 120 : assert_eq!(
9353 120 : tline
9354 120 : .get(get_key(idx as u32), gc_horizon, &ctx)
9355 120 : .await
9356 120 : .unwrap(),
9357 120 : &expected_result_at_gc_horizon[idx]
9358 2 : );
9359 120 : assert_eq!(
9360 120 : tline
9361 120 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
9362 120 : .await
9363 120 : .unwrap(),
9364 120 : &expected_result_at_lsn_20[idx]
9365 2 : );
9366 120 : assert_eq!(
9367 120 : tline
9368 120 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
9369 120 : .await
9370 120 : .unwrap(),
9371 120 : &expected_result_at_lsn_10[idx]
9372 2 : );
9373 2 : }
9374 24 : };
9375 2 :
9376 2 : verify_result().await;
9377 2 :
9378 2 : let cancel = CancellationToken::new();
9379 2 : let mut dryrun_flags = EnumSet::new();
9380 2 : dryrun_flags.insert(CompactFlags::DryRun);
9381 2 :
9382 2 : tline
9383 2 : .compact_with_gc(
9384 2 : &cancel,
9385 2 : CompactOptions {
9386 2 : flags: dryrun_flags,
9387 2 : ..Default::default()
9388 2 : },
9389 2 : &ctx,
9390 2 : )
9391 2 : .await
9392 2 : .unwrap();
9393 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
9394 2 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
9395 2 : verify_result().await;
9396 2 :
9397 2 : tline
9398 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9399 2 : .await
9400 2 : .unwrap();
9401 2 : verify_result().await;
9402 2 :
9403 2 : // compact again
9404 2 : tline
9405 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9406 2 : .await
9407 2 : .unwrap();
9408 2 : verify_result().await;
9409 2 :
9410 2 : // increase GC horizon and compact again
9411 2 : {
9412 2 : tline
9413 2 : .latest_gc_cutoff_lsn
9414 2 : .lock_for_write()
9415 2 : .store_and_unlock(Lsn(0x38))
9416 2 : .wait()
9417 2 : .await;
9418 2 : // Update GC info
9419 2 : let mut guard = tline.gc_info.write().unwrap();
9420 2 : guard.cutoffs.time = Lsn(0x38);
9421 2 : guard.cutoffs.space = Lsn(0x38);
9422 2 : }
9423 2 : tline
9424 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9425 2 : .await
9426 2 : .unwrap();
9427 2 : verify_result().await; // no wals between 0x30 and 0x38, so we should obtain the same result
9428 2 :
9429 2 : // not increasing the GC horizon and compact again
9430 2 : tline
9431 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9432 2 : .await
9433 2 : .unwrap();
9434 2 : verify_result().await;
9435 2 :
9436 2 : Ok(())
9437 2 : }
9438 :
9439 : #[cfg(feature = "testing")]
9440 : #[tokio::test]
9441 2 : async fn test_simple_bottom_most_compaction_with_retain_lsns_single_key() -> anyhow::Result<()>
9442 2 : {
9443 2 : let harness =
9444 2 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns_single_key")
9445 2 : .await?;
9446 2 : let (tenant, ctx) = harness.load().await;
9447 2 :
9448 352 : fn get_key(id: u32) -> Key {
9449 352 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9450 352 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9451 352 : key.field6 = id;
9452 352 : key
9453 352 : }
9454 2 :
9455 2 : let img_layer = (0..10)
9456 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9457 2 : .collect_vec();
9458 2 :
9459 2 : let delta1 = vec![
9460 2 : (
9461 2 : get_key(1),
9462 2 : Lsn(0x20),
9463 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9464 2 : ),
9465 2 : (
9466 2 : get_key(1),
9467 2 : Lsn(0x28),
9468 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9469 2 : ),
9470 2 : ];
9471 2 : let delta2 = vec![
9472 2 : (
9473 2 : get_key(1),
9474 2 : Lsn(0x30),
9475 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9476 2 : ),
9477 2 : (
9478 2 : get_key(1),
9479 2 : Lsn(0x38),
9480 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
9481 2 : ),
9482 2 : ];
9483 2 : let delta3 = vec![
9484 2 : (
9485 2 : get_key(8),
9486 2 : Lsn(0x48),
9487 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9488 2 : ),
9489 2 : (
9490 2 : get_key(9),
9491 2 : Lsn(0x48),
9492 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9493 2 : ),
9494 2 : ];
9495 2 :
9496 2 : let tline = tenant
9497 2 : .create_test_timeline_with_layers(
9498 2 : TIMELINE_ID,
9499 2 : Lsn(0x10),
9500 2 : DEFAULT_PG_VERSION,
9501 2 : &ctx,
9502 2 : vec![
9503 2 : // delta1 and delta 2 only contain a single key but multiple updates
9504 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x30), delta1),
9505 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
9506 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x50), delta3),
9507 2 : ], // delta layers
9508 2 : vec![(Lsn(0x10), img_layer)], // image layers
9509 2 : Lsn(0x50),
9510 2 : )
9511 2 : .await?;
9512 2 : {
9513 2 : tline
9514 2 : .latest_gc_cutoff_lsn
9515 2 : .lock_for_write()
9516 2 : .store_and_unlock(Lsn(0x30))
9517 2 : .wait()
9518 2 : .await;
9519 2 : // Update GC info
9520 2 : let mut guard = tline.gc_info.write().unwrap();
9521 2 : *guard = GcInfo {
9522 2 : retain_lsns: vec![
9523 2 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
9524 2 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
9525 2 : ],
9526 2 : cutoffs: GcCutoffs {
9527 2 : time: Lsn(0x30),
9528 2 : space: Lsn(0x30),
9529 2 : },
9530 2 : leases: Default::default(),
9531 2 : within_ancestor_pitr: false,
9532 2 : };
9533 2 : }
9534 2 :
9535 2 : let expected_result = [
9536 2 : Bytes::from_static(b"value 0@0x10"),
9537 2 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
9538 2 : Bytes::from_static(b"value 2@0x10"),
9539 2 : Bytes::from_static(b"value 3@0x10"),
9540 2 : Bytes::from_static(b"value 4@0x10"),
9541 2 : Bytes::from_static(b"value 5@0x10"),
9542 2 : Bytes::from_static(b"value 6@0x10"),
9543 2 : Bytes::from_static(b"value 7@0x10"),
9544 2 : Bytes::from_static(b"value 8@0x10@0x48"),
9545 2 : Bytes::from_static(b"value 9@0x10@0x48"),
9546 2 : ];
9547 2 :
9548 2 : let expected_result_at_gc_horizon = [
9549 2 : Bytes::from_static(b"value 0@0x10"),
9550 2 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
9551 2 : Bytes::from_static(b"value 2@0x10"),
9552 2 : Bytes::from_static(b"value 3@0x10"),
9553 2 : Bytes::from_static(b"value 4@0x10"),
9554 2 : Bytes::from_static(b"value 5@0x10"),
9555 2 : Bytes::from_static(b"value 6@0x10"),
9556 2 : Bytes::from_static(b"value 7@0x10"),
9557 2 : Bytes::from_static(b"value 8@0x10"),
9558 2 : Bytes::from_static(b"value 9@0x10"),
9559 2 : ];
9560 2 :
9561 2 : let expected_result_at_lsn_20 = [
9562 2 : Bytes::from_static(b"value 0@0x10"),
9563 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9564 2 : Bytes::from_static(b"value 2@0x10"),
9565 2 : Bytes::from_static(b"value 3@0x10"),
9566 2 : Bytes::from_static(b"value 4@0x10"),
9567 2 : Bytes::from_static(b"value 5@0x10"),
9568 2 : Bytes::from_static(b"value 6@0x10"),
9569 2 : Bytes::from_static(b"value 7@0x10"),
9570 2 : Bytes::from_static(b"value 8@0x10"),
9571 2 : Bytes::from_static(b"value 9@0x10"),
9572 2 : ];
9573 2 :
9574 2 : let expected_result_at_lsn_10 = [
9575 2 : Bytes::from_static(b"value 0@0x10"),
9576 2 : Bytes::from_static(b"value 1@0x10"),
9577 2 : Bytes::from_static(b"value 2@0x10"),
9578 2 : Bytes::from_static(b"value 3@0x10"),
9579 2 : Bytes::from_static(b"value 4@0x10"),
9580 2 : Bytes::from_static(b"value 5@0x10"),
9581 2 : Bytes::from_static(b"value 6@0x10"),
9582 2 : Bytes::from_static(b"value 7@0x10"),
9583 2 : Bytes::from_static(b"value 8@0x10"),
9584 2 : Bytes::from_static(b"value 9@0x10"),
9585 2 : ];
9586 2 :
9587 8 : let verify_result = || async {
9588 8 : let gc_horizon = {
9589 8 : let gc_info = tline.gc_info.read().unwrap();
9590 8 : gc_info.cutoffs.time
9591 2 : };
9592 88 : for idx in 0..10 {
9593 80 : assert_eq!(
9594 80 : tline
9595 80 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9596 80 : .await
9597 80 : .unwrap(),
9598 80 : &expected_result[idx]
9599 2 : );
9600 80 : assert_eq!(
9601 80 : tline
9602 80 : .get(get_key(idx as u32), gc_horizon, &ctx)
9603 80 : .await
9604 80 : .unwrap(),
9605 80 : &expected_result_at_gc_horizon[idx]
9606 2 : );
9607 80 : assert_eq!(
9608 80 : tline
9609 80 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
9610 80 : .await
9611 80 : .unwrap(),
9612 80 : &expected_result_at_lsn_20[idx]
9613 2 : );
9614 80 : assert_eq!(
9615 80 : tline
9616 80 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
9617 80 : .await
9618 80 : .unwrap(),
9619 80 : &expected_result_at_lsn_10[idx]
9620 2 : );
9621 2 : }
9622 16 : };
9623 2 :
9624 2 : verify_result().await;
9625 2 :
9626 2 : let cancel = CancellationToken::new();
9627 2 : let mut dryrun_flags = EnumSet::new();
9628 2 : dryrun_flags.insert(CompactFlags::DryRun);
9629 2 :
9630 2 : tline
9631 2 : .compact_with_gc(
9632 2 : &cancel,
9633 2 : CompactOptions {
9634 2 : flags: dryrun_flags,
9635 2 : ..Default::default()
9636 2 : },
9637 2 : &ctx,
9638 2 : )
9639 2 : .await
9640 2 : .unwrap();
9641 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
9642 2 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
9643 2 : verify_result().await;
9644 2 :
9645 2 : tline
9646 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9647 2 : .await
9648 2 : .unwrap();
9649 2 : verify_result().await;
9650 2 :
9651 2 : // compact again
9652 2 : tline
9653 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9654 2 : .await
9655 2 : .unwrap();
9656 2 : verify_result().await;
9657 2 :
9658 2 : Ok(())
9659 2 : }
9660 :
9661 : #[cfg(feature = "testing")]
9662 : #[tokio::test]
9663 2 : async fn test_simple_bottom_most_compaction_on_branch() -> anyhow::Result<()> {
9664 2 : use models::CompactLsnRange;
9665 2 :
9666 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_on_branch").await?;
9667 2 : let (tenant, ctx) = harness.load().await;
9668 2 :
9669 166 : fn get_key(id: u32) -> Key {
9670 166 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
9671 166 : key.field6 = id;
9672 166 : key
9673 166 : }
9674 2 :
9675 2 : let img_layer = (0..10)
9676 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9677 2 : .collect_vec();
9678 2 :
9679 2 : let delta1 = vec![
9680 2 : (
9681 2 : get_key(1),
9682 2 : Lsn(0x20),
9683 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9684 2 : ),
9685 2 : (
9686 2 : get_key(2),
9687 2 : Lsn(0x30),
9688 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9689 2 : ),
9690 2 : (
9691 2 : get_key(3),
9692 2 : Lsn(0x28),
9693 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9694 2 : ),
9695 2 : (
9696 2 : get_key(3),
9697 2 : Lsn(0x30),
9698 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9699 2 : ),
9700 2 : (
9701 2 : get_key(3),
9702 2 : Lsn(0x40),
9703 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9704 2 : ),
9705 2 : ];
9706 2 : let delta2 = vec![
9707 2 : (
9708 2 : get_key(5),
9709 2 : Lsn(0x20),
9710 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9711 2 : ),
9712 2 : (
9713 2 : get_key(6),
9714 2 : Lsn(0x20),
9715 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9716 2 : ),
9717 2 : ];
9718 2 : let delta3 = vec![
9719 2 : (
9720 2 : get_key(8),
9721 2 : Lsn(0x48),
9722 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9723 2 : ),
9724 2 : (
9725 2 : get_key(9),
9726 2 : Lsn(0x48),
9727 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9728 2 : ),
9729 2 : ];
9730 2 :
9731 2 : let parent_tline = tenant
9732 2 : .create_test_timeline_with_layers(
9733 2 : TIMELINE_ID,
9734 2 : Lsn(0x10),
9735 2 : DEFAULT_PG_VERSION,
9736 2 : &ctx,
9737 2 : vec![], // delta layers
9738 2 : vec![(Lsn(0x18), img_layer)], // image layers
9739 2 : Lsn(0x18),
9740 2 : )
9741 2 : .await?;
9742 2 :
9743 2 : parent_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
9744 2 :
9745 2 : let branch_tline = tenant
9746 2 : .branch_timeline_test_with_layers(
9747 2 : &parent_tline,
9748 2 : NEW_TIMELINE_ID,
9749 2 : Some(Lsn(0x18)),
9750 2 : &ctx,
9751 2 : vec![
9752 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
9753 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
9754 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
9755 2 : ], // delta layers
9756 2 : vec![], // image layers
9757 2 : Lsn(0x50),
9758 2 : )
9759 2 : .await?;
9760 2 :
9761 2 : branch_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
9762 2 :
9763 2 : {
9764 2 : parent_tline
9765 2 : .latest_gc_cutoff_lsn
9766 2 : .lock_for_write()
9767 2 : .store_and_unlock(Lsn(0x10))
9768 2 : .wait()
9769 2 : .await;
9770 2 : // Update GC info
9771 2 : let mut guard = parent_tline.gc_info.write().unwrap();
9772 2 : *guard = GcInfo {
9773 2 : retain_lsns: vec![(Lsn(0x18), branch_tline.timeline_id, MaybeOffloaded::No)],
9774 2 : cutoffs: GcCutoffs {
9775 2 : time: Lsn(0x10),
9776 2 : space: Lsn(0x10),
9777 2 : },
9778 2 : leases: Default::default(),
9779 2 : within_ancestor_pitr: false,
9780 2 : };
9781 2 : }
9782 2 :
9783 2 : {
9784 2 : branch_tline
9785 2 : .latest_gc_cutoff_lsn
9786 2 : .lock_for_write()
9787 2 : .store_and_unlock(Lsn(0x50))
9788 2 : .wait()
9789 2 : .await;
9790 2 : // Update GC info
9791 2 : let mut guard = branch_tline.gc_info.write().unwrap();
9792 2 : *guard = GcInfo {
9793 2 : retain_lsns: vec![(Lsn(0x40), branch_tline.timeline_id, MaybeOffloaded::No)],
9794 2 : cutoffs: GcCutoffs {
9795 2 : time: Lsn(0x50),
9796 2 : space: Lsn(0x50),
9797 2 : },
9798 2 : leases: Default::default(),
9799 2 : within_ancestor_pitr: false,
9800 2 : };
9801 2 : }
9802 2 :
9803 2 : let expected_result_at_gc_horizon = [
9804 2 : Bytes::from_static(b"value 0@0x10"),
9805 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9806 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9807 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9808 2 : Bytes::from_static(b"value 4@0x10"),
9809 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9810 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9811 2 : Bytes::from_static(b"value 7@0x10"),
9812 2 : Bytes::from_static(b"value 8@0x10@0x48"),
9813 2 : Bytes::from_static(b"value 9@0x10@0x48"),
9814 2 : ];
9815 2 :
9816 2 : let expected_result_at_lsn_40 = [
9817 2 : Bytes::from_static(b"value 0@0x10"),
9818 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9819 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9820 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9821 2 : Bytes::from_static(b"value 4@0x10"),
9822 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9823 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9824 2 : Bytes::from_static(b"value 7@0x10"),
9825 2 : Bytes::from_static(b"value 8@0x10"),
9826 2 : Bytes::from_static(b"value 9@0x10"),
9827 2 : ];
9828 2 :
9829 6 : let verify_result = || async {
9830 66 : for idx in 0..10 {
9831 60 : assert_eq!(
9832 60 : branch_tline
9833 60 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9834 60 : .await
9835 60 : .unwrap(),
9836 60 : &expected_result_at_gc_horizon[idx]
9837 2 : );
9838 60 : assert_eq!(
9839 60 : branch_tline
9840 60 : .get(get_key(idx as u32), Lsn(0x40), &ctx)
9841 60 : .await
9842 60 : .unwrap(),
9843 60 : &expected_result_at_lsn_40[idx]
9844 2 : );
9845 2 : }
9846 12 : };
9847 2 :
9848 2 : verify_result().await;
9849 2 :
9850 2 : let cancel = CancellationToken::new();
9851 2 : branch_tline
9852 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9853 2 : .await
9854 2 : .unwrap();
9855 2 :
9856 2 : verify_result().await;
9857 2 :
9858 2 : // Piggyback a compaction with above_lsn. Ensure it works correctly when the specified LSN intersects with the layer files.
9859 2 : // Now we already have a single large delta layer, so the compaction min_layer_lsn should be the same as ancestor LSN (0x18).
9860 2 : branch_tline
9861 2 : .compact_with_gc(
9862 2 : &cancel,
9863 2 : CompactOptions {
9864 2 : compact_lsn_range: Some(CompactLsnRange::above(Lsn(0x40))),
9865 2 : ..Default::default()
9866 2 : },
9867 2 : &ctx,
9868 2 : )
9869 2 : .await
9870 2 : .unwrap();
9871 2 :
9872 2 : verify_result().await;
9873 2 :
9874 2 : Ok(())
9875 2 : }
9876 :
9877 : // Regression test for https://github.com/neondatabase/neon/issues/9012
9878 : // Create an image arrangement where we have to read at different LSN ranges
9879 : // from a delta layer. This is achieved by overlapping an image layer on top of
9880 : // a delta layer. Like so:
9881 : //
9882 : // A B
9883 : // +----------------+ -> delta_layer
9884 : // | | ^ lsn
9885 : // | =========|-> nested_image_layer |
9886 : // | C | |
9887 : // +----------------+ |
9888 : // ======== -> baseline_image_layer +-------> key
9889 : //
9890 : //
9891 : // When querying the key range [A, B) we need to read at different LSN ranges
9892 : // for [A, C) and [C, B). This test checks that the described edge case is handled correctly.
9893 : #[cfg(feature = "testing")]
9894 : #[tokio::test]
9895 2 : async fn test_vectored_read_with_nested_image_layer() -> anyhow::Result<()> {
9896 2 : let harness = TenantHarness::create("test_vectored_read_with_nested_image_layer").await?;
9897 2 : let (tenant, ctx) = harness.load().await;
9898 2 :
9899 2 : let will_init_keys = [2, 6];
9900 44 : fn get_key(id: u32) -> Key {
9901 44 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
9902 44 : key.field6 = id;
9903 44 : key
9904 44 : }
9905 2 :
9906 2 : let mut expected_key_values = HashMap::new();
9907 2 :
9908 2 : let baseline_image_layer_lsn = Lsn(0x10);
9909 2 : let mut baseline_img_layer = Vec::new();
9910 12 : for i in 0..5 {
9911 10 : let key = get_key(i);
9912 10 : let value = format!("value {i}@{baseline_image_layer_lsn}");
9913 10 :
9914 10 : let removed = expected_key_values.insert(key, value.clone());
9915 10 : assert!(removed.is_none());
9916 2 :
9917 10 : baseline_img_layer.push((key, Bytes::from(value)));
9918 2 : }
9919 2 :
9920 2 : let nested_image_layer_lsn = Lsn(0x50);
9921 2 : let mut nested_img_layer = Vec::new();
9922 12 : for i in 5..10 {
9923 10 : let key = get_key(i);
9924 10 : let value = format!("value {i}@{nested_image_layer_lsn}");
9925 10 :
9926 10 : let removed = expected_key_values.insert(key, value.clone());
9927 10 : assert!(removed.is_none());
9928 2 :
9929 10 : nested_img_layer.push((key, Bytes::from(value)));
9930 2 : }
9931 2 :
9932 2 : let mut delta_layer_spec = Vec::default();
9933 2 : let delta_layer_start_lsn = Lsn(0x20);
9934 2 : let mut delta_layer_end_lsn = delta_layer_start_lsn;
9935 2 :
9936 22 : for i in 0..10 {
9937 20 : let key = get_key(i);
9938 20 : let key_in_nested = nested_img_layer
9939 20 : .iter()
9940 80 : .any(|(key_with_img, _)| *key_with_img == key);
9941 20 : let lsn = {
9942 20 : if key_in_nested {
9943 10 : Lsn(nested_image_layer_lsn.0 + 0x10)
9944 2 : } else {
9945 10 : delta_layer_start_lsn
9946 2 : }
9947 2 : };
9948 2 :
9949 20 : let will_init = will_init_keys.contains(&i);
9950 20 : if will_init {
9951 4 : delta_layer_spec.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init(""))));
9952 4 :
9953 4 : expected_key_values.insert(key, "".to_string());
9954 16 : } else {
9955 16 : let delta = format!("@{lsn}");
9956 16 : delta_layer_spec.push((
9957 16 : key,
9958 16 : lsn,
9959 16 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
9960 16 : ));
9961 16 :
9962 16 : expected_key_values
9963 16 : .get_mut(&key)
9964 16 : .expect("An image exists for each key")
9965 16 : .push_str(delta.as_str());
9966 16 : }
9967 20 : delta_layer_end_lsn = std::cmp::max(delta_layer_start_lsn, lsn);
9968 2 : }
9969 2 :
9970 2 : delta_layer_end_lsn = Lsn(delta_layer_end_lsn.0 + 1);
9971 2 :
9972 2 : assert!(
9973 2 : nested_image_layer_lsn > delta_layer_start_lsn
9974 2 : && nested_image_layer_lsn < delta_layer_end_lsn
9975 2 : );
9976 2 :
9977 2 : let tline = tenant
9978 2 : .create_test_timeline_with_layers(
9979 2 : TIMELINE_ID,
9980 2 : baseline_image_layer_lsn,
9981 2 : DEFAULT_PG_VERSION,
9982 2 : &ctx,
9983 2 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
9984 2 : delta_layer_start_lsn..delta_layer_end_lsn,
9985 2 : delta_layer_spec,
9986 2 : )], // delta layers
9987 2 : vec![
9988 2 : (baseline_image_layer_lsn, baseline_img_layer),
9989 2 : (nested_image_layer_lsn, nested_img_layer),
9990 2 : ], // image layers
9991 2 : delta_layer_end_lsn,
9992 2 : )
9993 2 : .await?;
9994 2 :
9995 2 : let keyspace = KeySpace::single(get_key(0)..get_key(10));
9996 2 : let results = tline
9997 2 : .get_vectored(keyspace, delta_layer_end_lsn, &ctx)
9998 2 : .await
9999 2 : .expect("No vectored errors");
10000 22 : for (key, res) in results {
10001 20 : let value = res.expect("No key errors");
10002 20 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
10003 20 : assert_eq!(value, Bytes::from(expected_value));
10004 2 : }
10005 2 :
10006 2 : Ok(())
10007 2 : }
10008 :
10009 214 : fn sort_layer_key(k1: &PersistentLayerKey, k2: &PersistentLayerKey) -> std::cmp::Ordering {
10010 214 : (
10011 214 : k1.is_delta,
10012 214 : k1.key_range.start,
10013 214 : k1.key_range.end,
10014 214 : k1.lsn_range.start,
10015 214 : k1.lsn_range.end,
10016 214 : )
10017 214 : .cmp(&(
10018 214 : k2.is_delta,
10019 214 : k2.key_range.start,
10020 214 : k2.key_range.end,
10021 214 : k2.lsn_range.start,
10022 214 : k2.lsn_range.end,
10023 214 : ))
10024 214 : }
10025 :
10026 24 : async fn inspect_and_sort(
10027 24 : tline: &Arc<Timeline>,
10028 24 : filter: Option<std::ops::Range<Key>>,
10029 24 : ) -> Vec<PersistentLayerKey> {
10030 24 : let mut all_layers = tline.inspect_historic_layers().await.unwrap();
10031 24 : if let Some(filter) = filter {
10032 108 : all_layers.retain(|layer| overlaps_with(&layer.key_range, &filter));
10033 22 : }
10034 24 : all_layers.sort_by(sort_layer_key);
10035 24 : all_layers
10036 24 : }
10037 :
10038 : #[cfg(feature = "testing")]
10039 22 : fn check_layer_map_key_eq(
10040 22 : mut left: Vec<PersistentLayerKey>,
10041 22 : mut right: Vec<PersistentLayerKey>,
10042 22 : ) {
10043 22 : left.sort_by(sort_layer_key);
10044 22 : right.sort_by(sort_layer_key);
10045 22 : if left != right {
10046 0 : eprintln!("---LEFT---");
10047 0 : for left in left.iter() {
10048 0 : eprintln!("{}", left);
10049 0 : }
10050 0 : eprintln!("---RIGHT---");
10051 0 : for right in right.iter() {
10052 0 : eprintln!("{}", right);
10053 0 : }
10054 0 : assert_eq!(left, right);
10055 22 : }
10056 22 : }
10057 :
10058 : #[cfg(feature = "testing")]
10059 : #[tokio::test]
10060 2 : async fn test_simple_partial_bottom_most_compaction() -> anyhow::Result<()> {
10061 2 : let harness = TenantHarness::create("test_simple_partial_bottom_most_compaction").await?;
10062 2 : let (tenant, ctx) = harness.load().await;
10063 2 :
10064 182 : fn get_key(id: u32) -> Key {
10065 182 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
10066 182 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
10067 182 : key.field6 = id;
10068 182 : key
10069 182 : }
10070 2 :
10071 2 : // img layer at 0x10
10072 2 : let img_layer = (0..10)
10073 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10074 2 : .collect_vec();
10075 2 :
10076 2 : let delta1 = vec![
10077 2 : (
10078 2 : get_key(1),
10079 2 : Lsn(0x20),
10080 2 : Value::Image(Bytes::from("value 1@0x20")),
10081 2 : ),
10082 2 : (
10083 2 : get_key(2),
10084 2 : Lsn(0x30),
10085 2 : Value::Image(Bytes::from("value 2@0x30")),
10086 2 : ),
10087 2 : (
10088 2 : get_key(3),
10089 2 : Lsn(0x40),
10090 2 : Value::Image(Bytes::from("value 3@0x40")),
10091 2 : ),
10092 2 : ];
10093 2 : let delta2 = vec![
10094 2 : (
10095 2 : get_key(5),
10096 2 : Lsn(0x20),
10097 2 : Value::Image(Bytes::from("value 5@0x20")),
10098 2 : ),
10099 2 : (
10100 2 : get_key(6),
10101 2 : Lsn(0x20),
10102 2 : Value::Image(Bytes::from("value 6@0x20")),
10103 2 : ),
10104 2 : ];
10105 2 : let delta3 = vec![
10106 2 : (
10107 2 : get_key(8),
10108 2 : Lsn(0x48),
10109 2 : Value::Image(Bytes::from("value 8@0x48")),
10110 2 : ),
10111 2 : (
10112 2 : get_key(9),
10113 2 : Lsn(0x48),
10114 2 : Value::Image(Bytes::from("value 9@0x48")),
10115 2 : ),
10116 2 : ];
10117 2 :
10118 2 : let tline = tenant
10119 2 : .create_test_timeline_with_layers(
10120 2 : TIMELINE_ID,
10121 2 : Lsn(0x10),
10122 2 : DEFAULT_PG_VERSION,
10123 2 : &ctx,
10124 2 : vec![
10125 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
10126 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
10127 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
10128 2 : ], // delta layers
10129 2 : vec![(Lsn(0x10), img_layer)], // image layers
10130 2 : Lsn(0x50),
10131 2 : )
10132 2 : .await?;
10133 2 :
10134 2 : {
10135 2 : tline
10136 2 : .latest_gc_cutoff_lsn
10137 2 : .lock_for_write()
10138 2 : .store_and_unlock(Lsn(0x30))
10139 2 : .wait()
10140 2 : .await;
10141 2 : // Update GC info
10142 2 : let mut guard = tline.gc_info.write().unwrap();
10143 2 : *guard = GcInfo {
10144 2 : retain_lsns: vec![(Lsn(0x20), tline.timeline_id, MaybeOffloaded::No)],
10145 2 : cutoffs: GcCutoffs {
10146 2 : time: Lsn(0x30),
10147 2 : space: Lsn(0x30),
10148 2 : },
10149 2 : leases: Default::default(),
10150 2 : within_ancestor_pitr: false,
10151 2 : };
10152 2 : }
10153 2 :
10154 2 : let cancel = CancellationToken::new();
10155 2 :
10156 2 : // Do a partial compaction on key range 0..2
10157 2 : tline
10158 2 : .compact_with_gc(
10159 2 : &cancel,
10160 2 : CompactOptions {
10161 2 : flags: EnumSet::new(),
10162 2 : compact_key_range: Some((get_key(0)..get_key(2)).into()),
10163 2 : ..Default::default()
10164 2 : },
10165 2 : &ctx,
10166 2 : )
10167 2 : .await
10168 2 : .unwrap();
10169 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10170 2 : check_layer_map_key_eq(
10171 2 : all_layers,
10172 2 : vec![
10173 2 : // newly-generated image layer for the partial compaction range 0-2
10174 2 : PersistentLayerKey {
10175 2 : key_range: get_key(0)..get_key(2),
10176 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10177 2 : is_delta: false,
10178 2 : },
10179 2 : PersistentLayerKey {
10180 2 : key_range: get_key(0)..get_key(10),
10181 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
10182 2 : is_delta: false,
10183 2 : },
10184 2 : // delta1 is split and the second part is rewritten
10185 2 : PersistentLayerKey {
10186 2 : key_range: get_key(2)..get_key(4),
10187 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
10188 2 : is_delta: true,
10189 2 : },
10190 2 : PersistentLayerKey {
10191 2 : key_range: get_key(5)..get_key(7),
10192 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
10193 2 : is_delta: true,
10194 2 : },
10195 2 : PersistentLayerKey {
10196 2 : key_range: get_key(8)..get_key(10),
10197 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
10198 2 : is_delta: true,
10199 2 : },
10200 2 : ],
10201 2 : );
10202 2 :
10203 2 : // Do a partial compaction on key range 2..4
10204 2 : tline
10205 2 : .compact_with_gc(
10206 2 : &cancel,
10207 2 : CompactOptions {
10208 2 : flags: EnumSet::new(),
10209 2 : compact_key_range: Some((get_key(2)..get_key(4)).into()),
10210 2 : ..Default::default()
10211 2 : },
10212 2 : &ctx,
10213 2 : )
10214 2 : .await
10215 2 : .unwrap();
10216 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10217 2 : check_layer_map_key_eq(
10218 2 : all_layers,
10219 2 : vec![
10220 2 : PersistentLayerKey {
10221 2 : key_range: get_key(0)..get_key(2),
10222 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10223 2 : is_delta: false,
10224 2 : },
10225 2 : PersistentLayerKey {
10226 2 : key_range: get_key(0)..get_key(10),
10227 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
10228 2 : is_delta: false,
10229 2 : },
10230 2 : // image layer generated for the compaction range 2-4
10231 2 : PersistentLayerKey {
10232 2 : key_range: get_key(2)..get_key(4),
10233 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10234 2 : is_delta: false,
10235 2 : },
10236 2 : // we have key2/key3 above the retain_lsn, so we still need this delta layer
10237 2 : PersistentLayerKey {
10238 2 : key_range: get_key(2)..get_key(4),
10239 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
10240 2 : is_delta: true,
10241 2 : },
10242 2 : PersistentLayerKey {
10243 2 : key_range: get_key(5)..get_key(7),
10244 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
10245 2 : is_delta: true,
10246 2 : },
10247 2 : PersistentLayerKey {
10248 2 : key_range: get_key(8)..get_key(10),
10249 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
10250 2 : is_delta: true,
10251 2 : },
10252 2 : ],
10253 2 : );
10254 2 :
10255 2 : // Do a partial compaction on key range 4..9
10256 2 : tline
10257 2 : .compact_with_gc(
10258 2 : &cancel,
10259 2 : CompactOptions {
10260 2 : flags: EnumSet::new(),
10261 2 : compact_key_range: Some((get_key(4)..get_key(9)).into()),
10262 2 : ..Default::default()
10263 2 : },
10264 2 : &ctx,
10265 2 : )
10266 2 : .await
10267 2 : .unwrap();
10268 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10269 2 : check_layer_map_key_eq(
10270 2 : all_layers,
10271 2 : vec![
10272 2 : PersistentLayerKey {
10273 2 : key_range: get_key(0)..get_key(2),
10274 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10275 2 : is_delta: false,
10276 2 : },
10277 2 : PersistentLayerKey {
10278 2 : key_range: get_key(0)..get_key(10),
10279 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
10280 2 : is_delta: false,
10281 2 : },
10282 2 : PersistentLayerKey {
10283 2 : key_range: get_key(2)..get_key(4),
10284 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10285 2 : is_delta: false,
10286 2 : },
10287 2 : PersistentLayerKey {
10288 2 : key_range: get_key(2)..get_key(4),
10289 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
10290 2 : is_delta: true,
10291 2 : },
10292 2 : // image layer generated for this compaction range
10293 2 : PersistentLayerKey {
10294 2 : key_range: get_key(4)..get_key(9),
10295 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10296 2 : is_delta: false,
10297 2 : },
10298 2 : PersistentLayerKey {
10299 2 : key_range: get_key(8)..get_key(10),
10300 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
10301 2 : is_delta: true,
10302 2 : },
10303 2 : ],
10304 2 : );
10305 2 :
10306 2 : // Do a partial compaction on key range 9..10
10307 2 : tline
10308 2 : .compact_with_gc(
10309 2 : &cancel,
10310 2 : CompactOptions {
10311 2 : flags: EnumSet::new(),
10312 2 : compact_key_range: Some((get_key(9)..get_key(10)).into()),
10313 2 : ..Default::default()
10314 2 : },
10315 2 : &ctx,
10316 2 : )
10317 2 : .await
10318 2 : .unwrap();
10319 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10320 2 : check_layer_map_key_eq(
10321 2 : all_layers,
10322 2 : vec![
10323 2 : PersistentLayerKey {
10324 2 : key_range: get_key(0)..get_key(2),
10325 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10326 2 : is_delta: false,
10327 2 : },
10328 2 : PersistentLayerKey {
10329 2 : key_range: get_key(0)..get_key(10),
10330 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
10331 2 : is_delta: false,
10332 2 : },
10333 2 : PersistentLayerKey {
10334 2 : key_range: get_key(2)..get_key(4),
10335 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10336 2 : is_delta: false,
10337 2 : },
10338 2 : PersistentLayerKey {
10339 2 : key_range: get_key(2)..get_key(4),
10340 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
10341 2 : is_delta: true,
10342 2 : },
10343 2 : PersistentLayerKey {
10344 2 : key_range: get_key(4)..get_key(9),
10345 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10346 2 : is_delta: false,
10347 2 : },
10348 2 : // image layer generated for the compaction range
10349 2 : PersistentLayerKey {
10350 2 : key_range: get_key(9)..get_key(10),
10351 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10352 2 : is_delta: false,
10353 2 : },
10354 2 : PersistentLayerKey {
10355 2 : key_range: get_key(8)..get_key(10),
10356 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
10357 2 : is_delta: true,
10358 2 : },
10359 2 : ],
10360 2 : );
10361 2 :
10362 2 : // Do a partial compaction on key range 0..10, all image layers below LSN 20 can be replaced with new ones.
10363 2 : tline
10364 2 : .compact_with_gc(
10365 2 : &cancel,
10366 2 : CompactOptions {
10367 2 : flags: EnumSet::new(),
10368 2 : compact_key_range: Some((get_key(0)..get_key(10)).into()),
10369 2 : ..Default::default()
10370 2 : },
10371 2 : &ctx,
10372 2 : )
10373 2 : .await
10374 2 : .unwrap();
10375 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10376 2 : check_layer_map_key_eq(
10377 2 : all_layers,
10378 2 : vec![
10379 2 : // aha, we removed all unnecessary image/delta layers and got a very clean layer map!
10380 2 : PersistentLayerKey {
10381 2 : key_range: get_key(0)..get_key(10),
10382 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10383 2 : is_delta: false,
10384 2 : },
10385 2 : PersistentLayerKey {
10386 2 : key_range: get_key(2)..get_key(4),
10387 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
10388 2 : is_delta: true,
10389 2 : },
10390 2 : PersistentLayerKey {
10391 2 : key_range: get_key(8)..get_key(10),
10392 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
10393 2 : is_delta: true,
10394 2 : },
10395 2 : ],
10396 2 : );
10397 2 : Ok(())
10398 2 : }
10399 :
10400 : #[cfg(feature = "testing")]
10401 : #[tokio::test]
10402 2 : async fn test_timeline_offload_retain_lsn() -> anyhow::Result<()> {
10403 2 : let harness = TenantHarness::create("test_timeline_offload_retain_lsn")
10404 2 : .await
10405 2 : .unwrap();
10406 2 : let (tenant, ctx) = harness.load().await;
10407 2 : let tline_parent = tenant
10408 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
10409 2 : .await
10410 2 : .unwrap();
10411 2 : let tline_child = tenant
10412 2 : .branch_timeline_test(&tline_parent, NEW_TIMELINE_ID, Some(Lsn(0x20)), &ctx)
10413 2 : .await
10414 2 : .unwrap();
10415 2 : {
10416 2 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
10417 2 : assert_eq!(
10418 2 : gc_info_parent.retain_lsns,
10419 2 : vec![(Lsn(0x20), tline_child.timeline_id, MaybeOffloaded::No)]
10420 2 : );
10421 2 : }
10422 2 : // We have to directly call the remote_client instead of using the archive function to avoid constructing broker client...
10423 2 : tline_child
10424 2 : .remote_client
10425 2 : .schedule_index_upload_for_timeline_archival_state(TimelineArchivalState::Archived)
10426 2 : .unwrap();
10427 2 : tline_child.remote_client.wait_completion().await.unwrap();
10428 2 : offload_timeline(&tenant, &tline_child)
10429 2 : .instrument(tracing::info_span!(parent: None, "offload_test", tenant_id=%"test", shard_id=%"test", timeline_id=%"test"))
10430 2 : .await.unwrap();
10431 2 : let child_timeline_id = tline_child.timeline_id;
10432 2 : Arc::try_unwrap(tline_child).unwrap();
10433 2 :
10434 2 : {
10435 2 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
10436 2 : assert_eq!(
10437 2 : gc_info_parent.retain_lsns,
10438 2 : vec![(Lsn(0x20), child_timeline_id, MaybeOffloaded::Yes)]
10439 2 : );
10440 2 : }
10441 2 :
10442 2 : tenant
10443 2 : .get_offloaded_timeline(child_timeline_id)
10444 2 : .unwrap()
10445 2 : .defuse_for_tenant_drop();
10446 2 :
10447 2 : Ok(())
10448 2 : }
10449 :
10450 : #[cfg(feature = "testing")]
10451 : #[tokio::test]
10452 2 : async fn test_simple_bottom_most_compaction_above_lsn() -> anyhow::Result<()> {
10453 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_above_lsn").await?;
10454 2 : let (tenant, ctx) = harness.load().await;
10455 2 :
10456 296 : fn get_key(id: u32) -> Key {
10457 296 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
10458 296 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
10459 296 : key.field6 = id;
10460 296 : key
10461 296 : }
10462 2 :
10463 2 : let img_layer = (0..10)
10464 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10465 2 : .collect_vec();
10466 2 :
10467 2 : let delta1 = vec![(
10468 2 : get_key(1),
10469 2 : Lsn(0x20),
10470 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10471 2 : )];
10472 2 : let delta4 = vec![(
10473 2 : get_key(1),
10474 2 : Lsn(0x28),
10475 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
10476 2 : )];
10477 2 : let delta2 = vec![
10478 2 : (
10479 2 : get_key(1),
10480 2 : Lsn(0x30),
10481 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10482 2 : ),
10483 2 : (
10484 2 : get_key(1),
10485 2 : Lsn(0x38),
10486 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
10487 2 : ),
10488 2 : ];
10489 2 : let delta3 = vec![
10490 2 : (
10491 2 : get_key(8),
10492 2 : Lsn(0x48),
10493 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10494 2 : ),
10495 2 : (
10496 2 : get_key(9),
10497 2 : Lsn(0x48),
10498 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10499 2 : ),
10500 2 : ];
10501 2 :
10502 2 : let tline = tenant
10503 2 : .create_test_timeline_with_layers(
10504 2 : TIMELINE_ID,
10505 2 : Lsn(0x10),
10506 2 : DEFAULT_PG_VERSION,
10507 2 : &ctx,
10508 2 : vec![
10509 2 : // delta1/2/4 only contain a single key but multiple updates
10510 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x28), delta1),
10511 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
10512 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x28)..Lsn(0x30), delta4),
10513 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta3),
10514 2 : ], // delta layers
10515 2 : vec![(Lsn(0x10), img_layer)], // image layers
10516 2 : Lsn(0x50),
10517 2 : )
10518 2 : .await?;
10519 2 : {
10520 2 : tline
10521 2 : .latest_gc_cutoff_lsn
10522 2 : .lock_for_write()
10523 2 : .store_and_unlock(Lsn(0x30))
10524 2 : .wait()
10525 2 : .await;
10526 2 : // Update GC info
10527 2 : let mut guard = tline.gc_info.write().unwrap();
10528 2 : *guard = GcInfo {
10529 2 : retain_lsns: vec![
10530 2 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
10531 2 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
10532 2 : ],
10533 2 : cutoffs: GcCutoffs {
10534 2 : time: Lsn(0x30),
10535 2 : space: Lsn(0x30),
10536 2 : },
10537 2 : leases: Default::default(),
10538 2 : within_ancestor_pitr: false,
10539 2 : };
10540 2 : }
10541 2 :
10542 2 : let expected_result = [
10543 2 : Bytes::from_static(b"value 0@0x10"),
10544 2 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
10545 2 : Bytes::from_static(b"value 2@0x10"),
10546 2 : Bytes::from_static(b"value 3@0x10"),
10547 2 : Bytes::from_static(b"value 4@0x10"),
10548 2 : Bytes::from_static(b"value 5@0x10"),
10549 2 : Bytes::from_static(b"value 6@0x10"),
10550 2 : Bytes::from_static(b"value 7@0x10"),
10551 2 : Bytes::from_static(b"value 8@0x10@0x48"),
10552 2 : Bytes::from_static(b"value 9@0x10@0x48"),
10553 2 : ];
10554 2 :
10555 2 : let expected_result_at_gc_horizon = [
10556 2 : Bytes::from_static(b"value 0@0x10"),
10557 2 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
10558 2 : Bytes::from_static(b"value 2@0x10"),
10559 2 : Bytes::from_static(b"value 3@0x10"),
10560 2 : Bytes::from_static(b"value 4@0x10"),
10561 2 : Bytes::from_static(b"value 5@0x10"),
10562 2 : Bytes::from_static(b"value 6@0x10"),
10563 2 : Bytes::from_static(b"value 7@0x10"),
10564 2 : Bytes::from_static(b"value 8@0x10"),
10565 2 : Bytes::from_static(b"value 9@0x10"),
10566 2 : ];
10567 2 :
10568 2 : let expected_result_at_lsn_20 = [
10569 2 : Bytes::from_static(b"value 0@0x10"),
10570 2 : Bytes::from_static(b"value 1@0x10@0x20"),
10571 2 : Bytes::from_static(b"value 2@0x10"),
10572 2 : Bytes::from_static(b"value 3@0x10"),
10573 2 : Bytes::from_static(b"value 4@0x10"),
10574 2 : Bytes::from_static(b"value 5@0x10"),
10575 2 : Bytes::from_static(b"value 6@0x10"),
10576 2 : Bytes::from_static(b"value 7@0x10"),
10577 2 : Bytes::from_static(b"value 8@0x10"),
10578 2 : Bytes::from_static(b"value 9@0x10"),
10579 2 : ];
10580 2 :
10581 2 : let expected_result_at_lsn_10 = [
10582 2 : Bytes::from_static(b"value 0@0x10"),
10583 2 : Bytes::from_static(b"value 1@0x10"),
10584 2 : Bytes::from_static(b"value 2@0x10"),
10585 2 : Bytes::from_static(b"value 3@0x10"),
10586 2 : Bytes::from_static(b"value 4@0x10"),
10587 2 : Bytes::from_static(b"value 5@0x10"),
10588 2 : Bytes::from_static(b"value 6@0x10"),
10589 2 : Bytes::from_static(b"value 7@0x10"),
10590 2 : Bytes::from_static(b"value 8@0x10"),
10591 2 : Bytes::from_static(b"value 9@0x10"),
10592 2 : ];
10593 2 :
10594 6 : let verify_result = || async {
10595 6 : let gc_horizon = {
10596 6 : let gc_info = tline.gc_info.read().unwrap();
10597 6 : gc_info.cutoffs.time
10598 2 : };
10599 66 : for idx in 0..10 {
10600 60 : assert_eq!(
10601 60 : tline
10602 60 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10603 60 : .await
10604 60 : .unwrap(),
10605 60 : &expected_result[idx]
10606 2 : );
10607 60 : assert_eq!(
10608 60 : tline
10609 60 : .get(get_key(idx as u32), gc_horizon, &ctx)
10610 60 : .await
10611 60 : .unwrap(),
10612 60 : &expected_result_at_gc_horizon[idx]
10613 2 : );
10614 60 : assert_eq!(
10615 60 : tline
10616 60 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
10617 60 : .await
10618 60 : .unwrap(),
10619 60 : &expected_result_at_lsn_20[idx]
10620 2 : );
10621 60 : assert_eq!(
10622 60 : tline
10623 60 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
10624 60 : .await
10625 60 : .unwrap(),
10626 60 : &expected_result_at_lsn_10[idx]
10627 2 : );
10628 2 : }
10629 12 : };
10630 2 :
10631 2 : verify_result().await;
10632 2 :
10633 2 : let cancel = CancellationToken::new();
10634 2 : tline
10635 2 : .compact_with_gc(
10636 2 : &cancel,
10637 2 : CompactOptions {
10638 2 : compact_lsn_range: Some(CompactLsnRange::above(Lsn(0x28))),
10639 2 : ..Default::default()
10640 2 : },
10641 2 : &ctx,
10642 2 : )
10643 2 : .await
10644 2 : .unwrap();
10645 2 : verify_result().await;
10646 2 :
10647 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10648 2 : check_layer_map_key_eq(
10649 2 : all_layers,
10650 2 : vec![
10651 2 : // The original image layer, not compacted
10652 2 : PersistentLayerKey {
10653 2 : key_range: get_key(0)..get_key(10),
10654 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
10655 2 : is_delta: false,
10656 2 : },
10657 2 : // Delta layer below the specified above_lsn not compacted
10658 2 : PersistentLayerKey {
10659 2 : key_range: get_key(1)..get_key(2),
10660 2 : lsn_range: Lsn(0x20)..Lsn(0x28),
10661 2 : is_delta: true,
10662 2 : },
10663 2 : // Delta layer compacted above the LSN
10664 2 : PersistentLayerKey {
10665 2 : key_range: get_key(1)..get_key(10),
10666 2 : lsn_range: Lsn(0x28)..Lsn(0x50),
10667 2 : is_delta: true,
10668 2 : },
10669 2 : ],
10670 2 : );
10671 2 :
10672 2 : // compact again
10673 2 : tline
10674 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10675 2 : .await
10676 2 : .unwrap();
10677 2 : verify_result().await;
10678 2 :
10679 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10680 2 : check_layer_map_key_eq(
10681 2 : all_layers,
10682 2 : vec![
10683 2 : // The compacted image layer (full key range)
10684 2 : PersistentLayerKey {
10685 2 : key_range: Key::MIN..Key::MAX,
10686 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
10687 2 : is_delta: false,
10688 2 : },
10689 2 : // All other data in the delta layer
10690 2 : PersistentLayerKey {
10691 2 : key_range: get_key(1)..get_key(10),
10692 2 : lsn_range: Lsn(0x10)..Lsn(0x50),
10693 2 : is_delta: true,
10694 2 : },
10695 2 : ],
10696 2 : );
10697 2 :
10698 2 : Ok(())
10699 2 : }
10700 :
10701 : #[cfg(feature = "testing")]
10702 : #[tokio::test]
10703 2 : async fn test_simple_bottom_most_compaction_rectangle() -> anyhow::Result<()> {
10704 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_rectangle").await?;
10705 2 : let (tenant, ctx) = harness.load().await;
10706 2 :
10707 508 : fn get_key(id: u32) -> Key {
10708 508 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
10709 508 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
10710 508 : key.field6 = id;
10711 508 : key
10712 508 : }
10713 2 :
10714 2 : let img_layer = (0..10)
10715 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10716 2 : .collect_vec();
10717 2 :
10718 2 : let delta1 = vec![(
10719 2 : get_key(1),
10720 2 : Lsn(0x20),
10721 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10722 2 : )];
10723 2 : let delta4 = vec![(
10724 2 : get_key(1),
10725 2 : Lsn(0x28),
10726 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
10727 2 : )];
10728 2 : let delta2 = vec![
10729 2 : (
10730 2 : get_key(1),
10731 2 : Lsn(0x30),
10732 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10733 2 : ),
10734 2 : (
10735 2 : get_key(1),
10736 2 : Lsn(0x38),
10737 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
10738 2 : ),
10739 2 : ];
10740 2 : let delta3 = vec![
10741 2 : (
10742 2 : get_key(8),
10743 2 : Lsn(0x48),
10744 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10745 2 : ),
10746 2 : (
10747 2 : get_key(9),
10748 2 : Lsn(0x48),
10749 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10750 2 : ),
10751 2 : ];
10752 2 :
10753 2 : let tline = tenant
10754 2 : .create_test_timeline_with_layers(
10755 2 : TIMELINE_ID,
10756 2 : Lsn(0x10),
10757 2 : DEFAULT_PG_VERSION,
10758 2 : &ctx,
10759 2 : vec![
10760 2 : // delta1/2/4 only contain a single key but multiple updates
10761 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x28), delta1),
10762 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
10763 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x28)..Lsn(0x30), delta4),
10764 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta3),
10765 2 : ], // delta layers
10766 2 : vec![(Lsn(0x10), img_layer)], // image layers
10767 2 : Lsn(0x50),
10768 2 : )
10769 2 : .await?;
10770 2 : {
10771 2 : tline
10772 2 : .latest_gc_cutoff_lsn
10773 2 : .lock_for_write()
10774 2 : .store_and_unlock(Lsn(0x30))
10775 2 : .wait()
10776 2 : .await;
10777 2 : // Update GC info
10778 2 : let mut guard = tline.gc_info.write().unwrap();
10779 2 : *guard = GcInfo {
10780 2 : retain_lsns: vec![
10781 2 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
10782 2 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
10783 2 : ],
10784 2 : cutoffs: GcCutoffs {
10785 2 : time: Lsn(0x30),
10786 2 : space: Lsn(0x30),
10787 2 : },
10788 2 : leases: Default::default(),
10789 2 : within_ancestor_pitr: false,
10790 2 : };
10791 2 : }
10792 2 :
10793 2 : let expected_result = [
10794 2 : Bytes::from_static(b"value 0@0x10"),
10795 2 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
10796 2 : Bytes::from_static(b"value 2@0x10"),
10797 2 : Bytes::from_static(b"value 3@0x10"),
10798 2 : Bytes::from_static(b"value 4@0x10"),
10799 2 : Bytes::from_static(b"value 5@0x10"),
10800 2 : Bytes::from_static(b"value 6@0x10"),
10801 2 : Bytes::from_static(b"value 7@0x10"),
10802 2 : Bytes::from_static(b"value 8@0x10@0x48"),
10803 2 : Bytes::from_static(b"value 9@0x10@0x48"),
10804 2 : ];
10805 2 :
10806 2 : let expected_result_at_gc_horizon = [
10807 2 : Bytes::from_static(b"value 0@0x10"),
10808 2 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
10809 2 : Bytes::from_static(b"value 2@0x10"),
10810 2 : Bytes::from_static(b"value 3@0x10"),
10811 2 : Bytes::from_static(b"value 4@0x10"),
10812 2 : Bytes::from_static(b"value 5@0x10"),
10813 2 : Bytes::from_static(b"value 6@0x10"),
10814 2 : Bytes::from_static(b"value 7@0x10"),
10815 2 : Bytes::from_static(b"value 8@0x10"),
10816 2 : Bytes::from_static(b"value 9@0x10"),
10817 2 : ];
10818 2 :
10819 2 : let expected_result_at_lsn_20 = [
10820 2 : Bytes::from_static(b"value 0@0x10"),
10821 2 : Bytes::from_static(b"value 1@0x10@0x20"),
10822 2 : Bytes::from_static(b"value 2@0x10"),
10823 2 : Bytes::from_static(b"value 3@0x10"),
10824 2 : Bytes::from_static(b"value 4@0x10"),
10825 2 : Bytes::from_static(b"value 5@0x10"),
10826 2 : Bytes::from_static(b"value 6@0x10"),
10827 2 : Bytes::from_static(b"value 7@0x10"),
10828 2 : Bytes::from_static(b"value 8@0x10"),
10829 2 : Bytes::from_static(b"value 9@0x10"),
10830 2 : ];
10831 2 :
10832 2 : let expected_result_at_lsn_10 = [
10833 2 : Bytes::from_static(b"value 0@0x10"),
10834 2 : Bytes::from_static(b"value 1@0x10"),
10835 2 : Bytes::from_static(b"value 2@0x10"),
10836 2 : Bytes::from_static(b"value 3@0x10"),
10837 2 : Bytes::from_static(b"value 4@0x10"),
10838 2 : Bytes::from_static(b"value 5@0x10"),
10839 2 : Bytes::from_static(b"value 6@0x10"),
10840 2 : Bytes::from_static(b"value 7@0x10"),
10841 2 : Bytes::from_static(b"value 8@0x10"),
10842 2 : Bytes::from_static(b"value 9@0x10"),
10843 2 : ];
10844 2 :
10845 10 : let verify_result = || async {
10846 10 : let gc_horizon = {
10847 10 : let gc_info = tline.gc_info.read().unwrap();
10848 10 : gc_info.cutoffs.time
10849 2 : };
10850 110 : for idx in 0..10 {
10851 100 : assert_eq!(
10852 100 : tline
10853 100 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10854 100 : .await
10855 100 : .unwrap(),
10856 100 : &expected_result[idx]
10857 2 : );
10858 100 : assert_eq!(
10859 100 : tline
10860 100 : .get(get_key(idx as u32), gc_horizon, &ctx)
10861 100 : .await
10862 100 : .unwrap(),
10863 100 : &expected_result_at_gc_horizon[idx]
10864 2 : );
10865 100 : assert_eq!(
10866 100 : tline
10867 100 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
10868 100 : .await
10869 100 : .unwrap(),
10870 100 : &expected_result_at_lsn_20[idx]
10871 2 : );
10872 100 : assert_eq!(
10873 100 : tline
10874 100 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
10875 100 : .await
10876 100 : .unwrap(),
10877 100 : &expected_result_at_lsn_10[idx]
10878 2 : );
10879 2 : }
10880 20 : };
10881 2 :
10882 2 : verify_result().await;
10883 2 :
10884 2 : let cancel = CancellationToken::new();
10885 2 :
10886 2 : tline
10887 2 : .compact_with_gc(
10888 2 : &cancel,
10889 2 : CompactOptions {
10890 2 : compact_key_range: Some((get_key(0)..get_key(2)).into()),
10891 2 : compact_lsn_range: Some((Lsn(0x20)..Lsn(0x28)).into()),
10892 2 : ..Default::default()
10893 2 : },
10894 2 : &ctx,
10895 2 : )
10896 2 : .await
10897 2 : .unwrap();
10898 2 : verify_result().await;
10899 2 :
10900 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10901 2 : check_layer_map_key_eq(
10902 2 : all_layers,
10903 2 : vec![
10904 2 : // The original image layer, not compacted
10905 2 : PersistentLayerKey {
10906 2 : key_range: get_key(0)..get_key(10),
10907 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
10908 2 : is_delta: false,
10909 2 : },
10910 2 : // According the selection logic, we select all layers with start key <= 0x28, so we would merge the layer 0x20-0x28 and
10911 2 : // the layer 0x28-0x30 into one.
10912 2 : PersistentLayerKey {
10913 2 : key_range: get_key(1)..get_key(2),
10914 2 : lsn_range: Lsn(0x20)..Lsn(0x30),
10915 2 : is_delta: true,
10916 2 : },
10917 2 : // Above the upper bound and untouched
10918 2 : PersistentLayerKey {
10919 2 : key_range: get_key(1)..get_key(2),
10920 2 : lsn_range: Lsn(0x30)..Lsn(0x50),
10921 2 : is_delta: true,
10922 2 : },
10923 2 : // This layer is untouched
10924 2 : PersistentLayerKey {
10925 2 : key_range: get_key(8)..get_key(10),
10926 2 : lsn_range: Lsn(0x30)..Lsn(0x50),
10927 2 : is_delta: true,
10928 2 : },
10929 2 : ],
10930 2 : );
10931 2 :
10932 2 : tline
10933 2 : .compact_with_gc(
10934 2 : &cancel,
10935 2 : CompactOptions {
10936 2 : compact_key_range: Some((get_key(3)..get_key(8)).into()),
10937 2 : compact_lsn_range: Some((Lsn(0x28)..Lsn(0x40)).into()),
10938 2 : ..Default::default()
10939 2 : },
10940 2 : &ctx,
10941 2 : )
10942 2 : .await
10943 2 : .unwrap();
10944 2 : verify_result().await;
10945 2 :
10946 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10947 2 : check_layer_map_key_eq(
10948 2 : all_layers,
10949 2 : vec![
10950 2 : // The original image layer, not compacted
10951 2 : PersistentLayerKey {
10952 2 : key_range: get_key(0)..get_key(10),
10953 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
10954 2 : is_delta: false,
10955 2 : },
10956 2 : // Not in the compaction key range, uncompacted
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 : // Not in the compaction key range, uncompacted but need rewrite because the delta layer overlaps with the range
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 : // Note that when we specify the LSN upper bound to be 0x40, the compaction algorithm will not try to cut the layer
10969 2 : // horizontally in half. Instead, it will include all LSNs that overlap with 0x40. So the real max_lsn of the compaction
10970 2 : // becomes 0x50.
10971 2 : PersistentLayerKey {
10972 2 : key_range: get_key(8)..get_key(10),
10973 2 : lsn_range: Lsn(0x30)..Lsn(0x50),
10974 2 : is_delta: true,
10975 2 : },
10976 2 : ],
10977 2 : );
10978 2 :
10979 2 : // compact again
10980 2 : tline
10981 2 : .compact_with_gc(
10982 2 : &cancel,
10983 2 : CompactOptions {
10984 2 : compact_key_range: Some((get_key(0)..get_key(5)).into()),
10985 2 : compact_lsn_range: Some((Lsn(0x20)..Lsn(0x50)).into()),
10986 2 : ..Default::default()
10987 2 : },
10988 2 : &ctx,
10989 2 : )
10990 2 : .await
10991 2 : .unwrap();
10992 2 : verify_result().await;
10993 2 :
10994 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10995 2 : check_layer_map_key_eq(
10996 2 : all_layers,
10997 2 : vec![
10998 2 : // The original image layer, not compacted
10999 2 : PersistentLayerKey {
11000 2 : key_range: get_key(0)..get_key(10),
11001 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
11002 2 : is_delta: false,
11003 2 : },
11004 2 : // The range gets compacted
11005 2 : PersistentLayerKey {
11006 2 : key_range: get_key(1)..get_key(2),
11007 2 : lsn_range: Lsn(0x20)..Lsn(0x50),
11008 2 : is_delta: true,
11009 2 : },
11010 2 : // Not touched during this iteration of compaction
11011 2 : PersistentLayerKey {
11012 2 : key_range: get_key(8)..get_key(10),
11013 2 : lsn_range: Lsn(0x30)..Lsn(0x50),
11014 2 : is_delta: true,
11015 2 : },
11016 2 : ],
11017 2 : );
11018 2 :
11019 2 : // final full compaction
11020 2 : tline
11021 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
11022 2 : .await
11023 2 : .unwrap();
11024 2 : verify_result().await;
11025 2 :
11026 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11027 2 : check_layer_map_key_eq(
11028 2 : all_layers,
11029 2 : vec![
11030 2 : // The compacted image layer (full key range)
11031 2 : PersistentLayerKey {
11032 2 : key_range: Key::MIN..Key::MAX,
11033 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
11034 2 : is_delta: false,
11035 2 : },
11036 2 : // All other data in the delta layer
11037 2 : PersistentLayerKey {
11038 2 : key_range: get_key(1)..get_key(10),
11039 2 : lsn_range: Lsn(0x10)..Lsn(0x50),
11040 2 : is_delta: true,
11041 2 : },
11042 2 : ],
11043 2 : );
11044 2 :
11045 2 : Ok(())
11046 2 : }
11047 : }
|