Line data Source code
1 : //! Timeline repository implementation that keeps old data in layer files, and
2 : //! the recent changes in ephemeral files.
3 : //!
4 : //! See tenant/*_layer.rs files. The functions here are responsible for locating
5 : //! the correct layer for the get/put call, walking back the timeline branching
6 : //! history as needed.
7 : //!
8 : //! The files are stored in the .neon/tenants/<tenant_id>/timelines/<timeline_id>
9 : //! directory. See docs/pageserver-storage.md for how the files are managed.
10 : //! In addition to the layer files, there is a metadata file in the same
11 : //! directory that contains information about the timeline, in particular its
12 : //! parent timeline, and the last LSN that has been written to disk.
13 : //!
14 :
15 : use anyhow::{bail, Context};
16 : use arc_swap::ArcSwap;
17 : use camino::Utf8Path;
18 : use camino::Utf8PathBuf;
19 : use chrono::NaiveDateTime;
20 : use enumset::EnumSet;
21 : use futures::stream::FuturesUnordered;
22 : use futures::StreamExt;
23 : use pageserver_api::models;
24 : use pageserver_api::models::CompactInfoResponse;
25 : use pageserver_api::models::LsnLease;
26 : use pageserver_api::models::TimelineArchivalState;
27 : use pageserver_api::models::TimelineState;
28 : use pageserver_api::models::TopTenantShardItem;
29 : use pageserver_api::models::WalRedoManagerStatus;
30 : use pageserver_api::shard::ShardIdentity;
31 : use pageserver_api::shard::ShardStripeSize;
32 : use pageserver_api::shard::TenantShardId;
33 : use remote_storage::DownloadError;
34 : use remote_storage::GenericRemoteStorage;
35 : use remote_storage::TimeoutOrCancel;
36 : use remote_timeline_client::manifest::{
37 : OffloadedTimelineManifest, TenantManifest, LATEST_TENANT_MANIFEST_VERSION,
38 : };
39 : use remote_timeline_client::UploadQueueNotReadyError;
40 : use remote_timeline_client::FAILED_REMOTE_OP_RETRIES;
41 : use remote_timeline_client::FAILED_UPLOAD_WARN_THRESHOLD;
42 : use std::collections::BTreeMap;
43 : use std::fmt;
44 : use std::future::Future;
45 : use std::sync::atomic::AtomicBool;
46 : use std::sync::Weak;
47 : use std::time::SystemTime;
48 : use storage_broker::BrokerClientChannel;
49 : use timeline::compaction::GcCompactionQueue;
50 : use timeline::import_pgdata;
51 : use timeline::offload::offload_timeline;
52 : use timeline::offload::OffloadError;
53 : use timeline::CompactOptions;
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::CONCURRENT_INITDBS;
101 : use crate::metrics::INITDB_RUN_TIME;
102 : use crate::metrics::INITDB_SEMAPHORE_ACQUISITION_TIME;
103 : use crate::metrics::TENANT;
104 : use crate::metrics::{
105 : remove_tenant_metrics, BROKEN_TENANTS_SET, CIRCUIT_BREAKERS_BROKEN, CIRCUIT_BREAKERS_UNBROKEN,
106 : TENANT_STATE_METRIC, TENANT_SYNTHETIC_SIZE_METRIC,
107 : };
108 : use crate::task_mgr;
109 : use crate::task_mgr::TaskKind;
110 : use crate::tenant::config::LocationMode;
111 : use crate::tenant::config::TenantConfOpt;
112 : use crate::tenant::gc_result::GcResult;
113 : pub use crate::tenant::remote_timeline_client::index::IndexPart;
114 : use crate::tenant::remote_timeline_client::remote_initdb_archive_path;
115 : use crate::tenant::remote_timeline_client::MaybeDeletedIndexPart;
116 : use crate::tenant::remote_timeline_client::INITDB_PATH;
117 : use crate::tenant::storage_layer::DeltaLayer;
118 : use crate::tenant::storage_layer::ImageLayer;
119 : use crate::walingest::WalLagCooldown;
120 : use crate::walredo;
121 : use crate::InitializationOrder;
122 : use std::collections::hash_map::Entry;
123 : use std::collections::HashMap;
124 : use std::collections::HashSet;
125 : use std::fmt::Debug;
126 : use std::fmt::Display;
127 : use std::fs;
128 : use std::fs::File;
129 : use std::sync::atomic::{AtomicU64, Ordering};
130 : use std::sync::Arc;
131 : use std::sync::Mutex;
132 : use std::time::{Duration, Instant};
133 :
134 : use crate::span;
135 : use crate::tenant::timeline::delete::DeleteTimelineFlow;
136 : use crate::tenant::timeline::uninit::cleanup_timeline_directory;
137 : use crate::virtual_file::VirtualFile;
138 : use crate::walredo::PostgresRedoManager;
139 : use crate::TEMP_FILE_SUFFIX;
140 : use once_cell::sync::Lazy;
141 : pub use pageserver_api::models::TenantState;
142 : use tokio::sync::Semaphore;
143 :
144 0 : static INIT_DB_SEMAPHORE: Lazy<Semaphore> = Lazy::new(|| Semaphore::new(8));
145 : use utils::{
146 : crashsafe,
147 : generation::Generation,
148 : id::TimelineId,
149 : lsn::{Lsn, RecordLsn},
150 : };
151 :
152 : pub mod blob_io;
153 : pub mod block_io;
154 : pub mod vectored_blob_io;
155 :
156 : pub mod disk_btree;
157 : pub(crate) mod ephemeral_file;
158 : pub mod layer_map;
159 :
160 : pub mod metadata;
161 : pub mod remote_timeline_client;
162 : pub mod storage_layer;
163 :
164 : pub mod checks;
165 : pub mod config;
166 : pub mod mgr;
167 : pub mod secondary;
168 : pub mod tasks;
169 : pub mod upload_queue;
170 :
171 : pub(crate) mod timeline;
172 :
173 : pub mod size;
174 :
175 : mod gc_block;
176 : mod gc_result;
177 : pub(crate) mod throttle;
178 :
179 : pub(crate) use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
180 : pub(crate) use timeline::{LogicalSizeCalculationCause, PageReconstructError, Timeline};
181 :
182 : // re-export for use in walreceiver
183 : pub use crate::tenant::timeline::WalReceiverInfo;
184 :
185 : /// The "tenants" part of `tenants/<tenant>/timelines...`
186 : pub const TENANTS_SEGMENT_NAME: &str = "tenants";
187 :
188 : /// Parts of the `.neon/tenants/<tenant_id>/timelines/<timeline_id>` directory prefix.
189 : pub const TIMELINES_SEGMENT_NAME: &str = "timelines";
190 :
191 : /// References to shared objects that are passed into each tenant, such
192 : /// as the shared remote storage client and process initialization state.
193 : #[derive(Clone)]
194 : pub struct TenantSharedResources {
195 : pub broker_client: storage_broker::BrokerClientChannel,
196 : pub remote_storage: GenericRemoteStorage,
197 : pub deletion_queue_client: DeletionQueueClient,
198 : pub l0_flush_global_state: L0FlushGlobalState,
199 : }
200 :
201 : /// A [`Tenant`] is really an _attached_ tenant. The configuration
202 : /// for an attached tenant is a subset of the [`LocationConf`], represented
203 : /// in this struct.
204 : #[derive(Clone)]
205 : pub(super) struct AttachedTenantConf {
206 : tenant_conf: TenantConfOpt,
207 : location: AttachedLocationConfig,
208 : /// The deadline before which we are blocked from GC so that
209 : /// leases have a chance to be renewed.
210 : lsn_lease_deadline: Option<tokio::time::Instant>,
211 : }
212 :
213 : impl AttachedTenantConf {
214 440 : fn new(tenant_conf: TenantConfOpt, location: AttachedLocationConfig) -> Self {
215 : // Sets a deadline before which we cannot proceed to GC due to lsn lease.
216 : //
217 : // We do this as the leases mapping are not persisted to disk. By delaying GC by lease
218 : // length, we guarantee that all the leases we granted before will have a chance to renew
219 : // when we run GC for the first time after restart / transition from AttachedMulti to AttachedSingle.
220 440 : let lsn_lease_deadline = if location.attach_mode == AttachmentMode::Single {
221 440 : Some(
222 440 : tokio::time::Instant::now()
223 440 : + tenant_conf
224 440 : .lsn_lease_length
225 440 : .unwrap_or(LsnLease::DEFAULT_LENGTH),
226 440 : )
227 : } else {
228 : // We don't use `lsn_lease_deadline` to delay GC in AttachedMulti and AttachedStale
229 : // because we don't do GC in these modes.
230 0 : None
231 : };
232 :
233 440 : Self {
234 440 : tenant_conf,
235 440 : location,
236 440 : lsn_lease_deadline,
237 440 : }
238 440 : }
239 :
240 440 : fn try_from(location_conf: LocationConf) -> anyhow::Result<Self> {
241 440 : match &location_conf.mode {
242 440 : LocationMode::Attached(attach_conf) => {
243 440 : Ok(Self::new(location_conf.tenant_conf, *attach_conf))
244 : }
245 : LocationMode::Secondary(_) => {
246 0 : anyhow::bail!("Attempted to construct AttachedTenantConf from a LocationConf in secondary mode")
247 : }
248 : }
249 440 : }
250 :
251 1524 : fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
252 1524 : self.lsn_lease_deadline
253 1524 : .map(|d| tokio::time::Instant::now() < d)
254 1524 : .unwrap_or(false)
255 1524 : }
256 : }
257 : struct TimelinePreload {
258 : timeline_id: TimelineId,
259 : client: RemoteTimelineClient,
260 : index_part: Result<MaybeDeletedIndexPart, DownloadError>,
261 : }
262 :
263 : pub(crate) struct TenantPreload {
264 : tenant_manifest: TenantManifest,
265 : /// Map from timeline ID to a possible timeline preload. It is None iff the timeline is offloaded according to the manifest.
266 : timelines: HashMap<TimelineId, Option<TimelinePreload>>,
267 : }
268 :
269 : /// When we spawn a tenant, there is a special mode for tenant creation that
270 : /// avoids trying to read anything from remote storage.
271 : pub(crate) enum SpawnMode {
272 : /// Activate as soon as possible
273 : Eager,
274 : /// Lazy activation in the background, with the option to skip the queue if the need comes up
275 : Lazy,
276 : }
277 :
278 : ///
279 : /// Tenant consists of multiple timelines. Keep them in a hash table.
280 : ///
281 : pub struct Tenant {
282 : // Global pageserver config parameters
283 : pub conf: &'static PageServerConf,
284 :
285 : /// The value creation timestamp, used to measure activation delay, see:
286 : /// <https://github.com/neondatabase/neon/issues/4025>
287 : constructed_at: Instant,
288 :
289 : state: watch::Sender<TenantState>,
290 :
291 : // Overridden tenant-specific config parameters.
292 : // We keep TenantConfOpt sturct here to preserve the information
293 : // about parameters that are not set.
294 : // This is necessary to allow global config updates.
295 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
296 :
297 : tenant_shard_id: TenantShardId,
298 :
299 : // The detailed sharding information, beyond the number/count in tenant_shard_id
300 : shard_identity: ShardIdentity,
301 :
302 : /// The remote storage generation, used to protect S3 objects from split-brain.
303 : /// Does not change over the lifetime of the [`Tenant`] object.
304 : ///
305 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
306 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
307 : generation: Generation,
308 :
309 : timelines: Mutex<HashMap<TimelineId, Arc<Timeline>>>,
310 :
311 : /// During timeline creation, we first insert the TimelineId to the
312 : /// creating map, then `timelines`, then remove it from the creating map.
313 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
314 : timelines_creating: std::sync::Mutex<HashSet<TimelineId>>,
315 :
316 : /// Possibly offloaded and archived timelines
317 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
318 : timelines_offloaded: Mutex<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
319 :
320 : /// Serialize writes of the tenant manifest to remote storage. If there are concurrent operations
321 : /// affecting the manifest, such as timeline deletion and timeline offload, they must wait for
322 : /// each other (this could be optimized to coalesce writes if necessary).
323 : ///
324 : /// The contents of the Mutex are the last manifest we successfully uploaded
325 : tenant_manifest_upload: tokio::sync::Mutex<Option<TenantManifest>>,
326 :
327 : // This mutex prevents creation of new timelines during GC.
328 : // Adding yet another mutex (in addition to `timelines`) is needed because holding
329 : // `timelines` mutex during all GC iteration
330 : // may block for a long time `get_timeline`, `get_timelines_state`,... and other operations
331 : // with timelines, which in turn may cause dropping replication connection, expiration of wait_for_lsn
332 : // timeout...
333 : gc_cs: tokio::sync::Mutex<()>,
334 : walredo_mgr: Option<Arc<WalRedoManager>>,
335 :
336 : // provides access to timeline data sitting in the remote storage
337 : pub(crate) remote_storage: GenericRemoteStorage,
338 :
339 : // Access to global deletion queue for when this tenant wants to schedule a deletion
340 : deletion_queue_client: DeletionQueueClient,
341 :
342 : /// Cached logical sizes updated updated on each [`Tenant::gather_size_inputs`].
343 : cached_logical_sizes: tokio::sync::Mutex<HashMap<(TimelineId, Lsn), u64>>,
344 : cached_synthetic_tenant_size: Arc<AtomicU64>,
345 :
346 : eviction_task_tenant_state: tokio::sync::Mutex<EvictionTaskTenantState>,
347 :
348 : /// Track repeated failures to compact, so that we can back off.
349 : /// Overhead of mutex is acceptable because compaction is done with a multi-second period.
350 : compaction_circuit_breaker: std::sync::Mutex<CircuitBreaker>,
351 :
352 : /// Scheduled gc-compaction tasks.
353 : scheduled_compaction_tasks: std::sync::Mutex<HashMap<TimelineId, Arc<GcCompactionQueue>>>,
354 :
355 : /// If the tenant is in Activating state, notify this to encourage it
356 : /// to proceed to Active as soon as possible, rather than waiting for lazy
357 : /// background warmup.
358 : pub(crate) activate_now_sem: tokio::sync::Semaphore,
359 :
360 : /// Time it took for the tenant to activate. Zero if not active yet.
361 : attach_wal_lag_cooldown: Arc<std::sync::OnceLock<WalLagCooldown>>,
362 :
363 : // Cancellation token fires when we have entered shutdown(). This is a parent of
364 : // Timelines' cancellation token.
365 : pub(crate) cancel: CancellationToken,
366 :
367 : // Users of the Tenant such as the page service must take this Gate to avoid
368 : // trying to use a Tenant which is shutting down.
369 : pub(crate) gate: Gate,
370 :
371 : /// Throttle applied at the top of [`Timeline::get`].
372 : /// All [`Tenant::timelines`] of a given [`Tenant`] instance share the same [`throttle::Throttle`] instance.
373 : pub(crate) pagestream_throttle: Arc<throttle::Throttle>,
374 :
375 : pub(crate) pagestream_throttle_metrics: Arc<crate::metrics::tenant_throttling::Pagestream>,
376 :
377 : /// An ongoing timeline detach concurrency limiter.
378 : ///
379 : /// As a tenant will likely be restarted as part of timeline detach ancestor it makes no sense
380 : /// to have two running at the same time. A different one can be started if an earlier one
381 : /// has failed for whatever reason.
382 : ongoing_timeline_detach: std::sync::Mutex<Option<(TimelineId, utils::completion::Barrier)>>,
383 :
384 : /// `index_part.json` based gc blocking reason tracking.
385 : ///
386 : /// New gc iterations must start a new iteration by acquiring `GcBlock::start` before
387 : /// proceeding.
388 : pub(crate) gc_block: gc_block::GcBlock,
389 :
390 : l0_flush_global_state: L0FlushGlobalState,
391 : }
392 : impl std::fmt::Debug for Tenant {
393 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
394 0 : write!(f, "{} ({})", self.tenant_shard_id, self.current_state())
395 0 : }
396 : }
397 :
398 : pub(crate) enum WalRedoManager {
399 : Prod(WalredoManagerId, PostgresRedoManager),
400 : #[cfg(test)]
401 : Test(harness::TestRedoManager),
402 : }
403 :
404 : #[derive(thiserror::Error, Debug)]
405 : #[error("pageserver is shutting down")]
406 : pub(crate) struct GlobalShutDown;
407 :
408 : impl WalRedoManager {
409 0 : pub(crate) fn new(mgr: PostgresRedoManager) -> Result<Arc<Self>, GlobalShutDown> {
410 0 : let id = WalredoManagerId::next();
411 0 : let arc = Arc::new(Self::Prod(id, mgr));
412 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
413 0 : match &mut *guard {
414 0 : Some(map) => {
415 0 : map.insert(id, Arc::downgrade(&arc));
416 0 : Ok(arc)
417 : }
418 0 : None => Err(GlobalShutDown),
419 : }
420 0 : }
421 : }
422 :
423 : impl Drop for WalRedoManager {
424 20 : fn drop(&mut self) {
425 20 : match self {
426 0 : Self::Prod(id, _) => {
427 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
428 0 : if let Some(map) = &mut *guard {
429 0 : map.remove(id).expect("new() registers, drop() unregisters");
430 0 : }
431 : }
432 : #[cfg(test)]
433 20 : Self::Test(_) => {
434 20 : // Not applicable to test redo manager
435 20 : }
436 : }
437 20 : }
438 : }
439 :
440 : /// Global registry of all walredo managers so that [`crate::shutdown_pageserver`] can shut down
441 : /// the walredo processes outside of the regular order.
442 : ///
443 : /// This is necessary to work around a systemd bug where it freezes if there are
444 : /// walredo processes left => <https://github.com/neondatabase/cloud/issues/11387>
445 : #[allow(clippy::type_complexity)]
446 : pub(crate) static WALREDO_MANAGERS: once_cell::sync::Lazy<
447 : Mutex<Option<HashMap<WalredoManagerId, Weak<WalRedoManager>>>>,
448 0 : > = once_cell::sync::Lazy::new(|| Mutex::new(Some(HashMap::new())));
449 : #[derive(PartialEq, Eq, Hash, Clone, Copy, Debug)]
450 : pub(crate) struct WalredoManagerId(u64);
451 : impl WalredoManagerId {
452 0 : pub fn next() -> Self {
453 : static NEXT: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(1);
454 0 : let id = NEXT.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
455 0 : if id == 0 {
456 0 : panic!("WalredoManagerId::new() returned 0, indicating wraparound, risking it's no longer unique");
457 0 : }
458 0 : Self(id)
459 0 : }
460 : }
461 :
462 : #[cfg(test)]
463 : impl From<harness::TestRedoManager> for WalRedoManager {
464 440 : fn from(mgr: harness::TestRedoManager) -> Self {
465 440 : Self::Test(mgr)
466 440 : }
467 : }
468 :
469 : impl WalRedoManager {
470 12 : pub(crate) async fn shutdown(&self) -> bool {
471 12 : match self {
472 0 : Self::Prod(_, mgr) => mgr.shutdown().await,
473 : #[cfg(test)]
474 : Self::Test(_) => {
475 : // Not applicable to test redo manager
476 12 : true
477 : }
478 : }
479 12 : }
480 :
481 0 : pub(crate) fn maybe_quiesce(&self, idle_timeout: Duration) {
482 0 : match self {
483 0 : Self::Prod(_, mgr) => mgr.maybe_quiesce(idle_timeout),
484 0 : #[cfg(test)]
485 0 : Self::Test(_) => {
486 0 : // Not applicable to test redo manager
487 0 : }
488 0 : }
489 0 : }
490 :
491 : /// # Cancel-Safety
492 : ///
493 : /// This method is cancellation-safe.
494 1636 : pub async fn request_redo(
495 1636 : &self,
496 1636 : key: pageserver_api::key::Key,
497 1636 : lsn: Lsn,
498 1636 : base_img: Option<(Lsn, bytes::Bytes)>,
499 1636 : records: Vec<(Lsn, pageserver_api::record::NeonWalRecord)>,
500 1636 : pg_version: u32,
501 1636 : ) -> Result<bytes::Bytes, walredo::Error> {
502 1636 : match self {
503 0 : Self::Prod(_, mgr) => {
504 0 : mgr.request_redo(key, lsn, base_img, records, pg_version)
505 0 : .await
506 : }
507 : #[cfg(test)]
508 1636 : Self::Test(mgr) => {
509 1636 : mgr.request_redo(key, lsn, base_img, records, pg_version)
510 1636 : .await
511 : }
512 : }
513 1636 : }
514 :
515 0 : pub(crate) fn status(&self) -> Option<WalRedoManagerStatus> {
516 0 : match self {
517 0 : WalRedoManager::Prod(_, m) => Some(m.status()),
518 0 : #[cfg(test)]
519 0 : WalRedoManager::Test(_) => None,
520 0 : }
521 0 : }
522 : }
523 :
524 : /// A very lightweight memory representation of an offloaded timeline.
525 : ///
526 : /// We need to store the list of offloaded timelines so that we can perform operations on them,
527 : /// like unoffloading them, or (at a later date), decide to perform flattening.
528 : /// This type has a much smaller memory impact than [`Timeline`], and thus we can store many
529 : /// more offloaded timelines than we can manage ones that aren't.
530 : pub struct OffloadedTimeline {
531 : pub tenant_shard_id: TenantShardId,
532 : pub timeline_id: TimelineId,
533 : pub ancestor_timeline_id: Option<TimelineId>,
534 : /// Whether to retain the branch lsn at the ancestor or not
535 : pub ancestor_retain_lsn: Option<Lsn>,
536 :
537 : /// When the timeline was archived.
538 : ///
539 : /// Present for future flattening deliberations.
540 : pub archived_at: NaiveDateTime,
541 :
542 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
543 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
544 : pub delete_progress: TimelineDeleteProgress,
545 :
546 : /// Part of the `OffloadedTimeline` object's lifecycle: this needs to be set before we drop it
547 : pub deleted_from_ancestor: AtomicBool,
548 : }
549 :
550 : impl OffloadedTimeline {
551 : /// Obtains an offloaded timeline from a given timeline object.
552 : ///
553 : /// Returns `None` if the `archived_at` flag couldn't be obtained, i.e.
554 : /// the timeline is not in a stopped state.
555 : /// Panics if the timeline is not archived.
556 4 : fn from_timeline(timeline: &Timeline) -> Result<Self, UploadQueueNotReadyError> {
557 4 : let (ancestor_retain_lsn, ancestor_timeline_id) =
558 4 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
559 4 : let ancestor_lsn = timeline.get_ancestor_lsn();
560 4 : let ancestor_timeline_id = ancestor_timeline.timeline_id;
561 4 : let mut gc_info = ancestor_timeline.gc_info.write().unwrap();
562 4 : gc_info.insert_child(timeline.timeline_id, ancestor_lsn, MaybeOffloaded::Yes);
563 4 : (Some(ancestor_lsn), Some(ancestor_timeline_id))
564 : } else {
565 0 : (None, None)
566 : };
567 4 : let archived_at = timeline
568 4 : .remote_client
569 4 : .archived_at_stopped_queue()?
570 4 : .expect("must be called on an archived timeline");
571 4 : Ok(Self {
572 4 : tenant_shard_id: timeline.tenant_shard_id,
573 4 : timeline_id: timeline.timeline_id,
574 4 : ancestor_timeline_id,
575 4 : ancestor_retain_lsn,
576 4 : archived_at,
577 4 :
578 4 : delete_progress: timeline.delete_progress.clone(),
579 4 : deleted_from_ancestor: AtomicBool::new(false),
580 4 : })
581 4 : }
582 0 : fn from_manifest(tenant_shard_id: TenantShardId, manifest: &OffloadedTimelineManifest) -> Self {
583 0 : // We expect to reach this case in tenant loading, where the `retain_lsn` is populated in the parent's `gc_info`
584 0 : // by the `initialize_gc_info` function.
585 0 : let OffloadedTimelineManifest {
586 0 : timeline_id,
587 0 : ancestor_timeline_id,
588 0 : ancestor_retain_lsn,
589 0 : archived_at,
590 0 : } = *manifest;
591 0 : Self {
592 0 : tenant_shard_id,
593 0 : timeline_id,
594 0 : ancestor_timeline_id,
595 0 : ancestor_retain_lsn,
596 0 : archived_at,
597 0 : delete_progress: TimelineDeleteProgress::default(),
598 0 : deleted_from_ancestor: AtomicBool::new(false),
599 0 : }
600 0 : }
601 4 : fn manifest(&self) -> OffloadedTimelineManifest {
602 4 : let Self {
603 4 : timeline_id,
604 4 : ancestor_timeline_id,
605 4 : ancestor_retain_lsn,
606 4 : archived_at,
607 4 : ..
608 4 : } = self;
609 4 : OffloadedTimelineManifest {
610 4 : timeline_id: *timeline_id,
611 4 : ancestor_timeline_id: *ancestor_timeline_id,
612 4 : ancestor_retain_lsn: *ancestor_retain_lsn,
613 4 : archived_at: *archived_at,
614 4 : }
615 4 : }
616 : /// Delete this timeline's retain_lsn from its ancestor, if present in the given tenant
617 0 : fn delete_from_ancestor_with_timelines(
618 0 : &self,
619 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
620 0 : ) {
621 0 : if let (Some(_retain_lsn), Some(ancestor_timeline_id)) =
622 0 : (self.ancestor_retain_lsn, self.ancestor_timeline_id)
623 : {
624 0 : if let Some((_, ancestor_timeline)) = timelines
625 0 : .iter()
626 0 : .find(|(tid, _tl)| **tid == ancestor_timeline_id)
627 : {
628 0 : let removal_happened = ancestor_timeline
629 0 : .gc_info
630 0 : .write()
631 0 : .unwrap()
632 0 : .remove_child_offloaded(self.timeline_id);
633 0 : if !removal_happened {
634 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
635 0 : "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
636 0 : }
637 0 : }
638 0 : }
639 0 : self.deleted_from_ancestor.store(true, Ordering::Release);
640 0 : }
641 : /// Call [`Self::delete_from_ancestor_with_timelines`] instead if possible.
642 : ///
643 : /// As the entire tenant is being dropped, don't bother deregistering the `retain_lsn` from the ancestor.
644 4 : fn defuse_for_tenant_drop(&self) {
645 4 : self.deleted_from_ancestor.store(true, Ordering::Release);
646 4 : }
647 : }
648 :
649 : impl fmt::Debug for OffloadedTimeline {
650 0 : fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
651 0 : write!(f, "OffloadedTimeline<{}>", self.timeline_id)
652 0 : }
653 : }
654 :
655 : impl Drop for OffloadedTimeline {
656 4 : fn drop(&mut self) {
657 4 : if !self.deleted_from_ancestor.load(Ordering::Acquire) {
658 0 : tracing::warn!(
659 0 : "offloaded timeline {} was dropped without having cleaned it up at the ancestor",
660 : self.timeline_id
661 : );
662 4 : }
663 4 : }
664 : }
665 :
666 : #[derive(Copy, Clone, PartialEq, Eq, Hash, Debug)]
667 : pub enum MaybeOffloaded {
668 : Yes,
669 : No,
670 : }
671 :
672 : #[derive(Clone, Debug)]
673 : pub enum TimelineOrOffloaded {
674 : Timeline(Arc<Timeline>),
675 : Offloaded(Arc<OffloadedTimeline>),
676 : }
677 :
678 : impl TimelineOrOffloaded {
679 0 : pub fn arc_ref(&self) -> TimelineOrOffloadedArcRef<'_> {
680 0 : match self {
681 0 : TimelineOrOffloaded::Timeline(timeline) => {
682 0 : TimelineOrOffloadedArcRef::Timeline(timeline)
683 : }
684 0 : TimelineOrOffloaded::Offloaded(offloaded) => {
685 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded)
686 : }
687 : }
688 0 : }
689 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
690 0 : self.arc_ref().tenant_shard_id()
691 0 : }
692 0 : pub fn timeline_id(&self) -> TimelineId {
693 0 : self.arc_ref().timeline_id()
694 0 : }
695 4 : pub fn delete_progress(&self) -> &Arc<tokio::sync::Mutex<DeleteTimelineFlow>> {
696 4 : match self {
697 4 : TimelineOrOffloaded::Timeline(timeline) => &timeline.delete_progress,
698 0 : TimelineOrOffloaded::Offloaded(offloaded) => &offloaded.delete_progress,
699 : }
700 4 : }
701 0 : fn maybe_remote_client(&self) -> Option<Arc<RemoteTimelineClient>> {
702 0 : match self {
703 0 : TimelineOrOffloaded::Timeline(timeline) => Some(timeline.remote_client.clone()),
704 0 : TimelineOrOffloaded::Offloaded(_offloaded) => None,
705 : }
706 0 : }
707 : }
708 :
709 : pub enum TimelineOrOffloadedArcRef<'a> {
710 : Timeline(&'a Arc<Timeline>),
711 : Offloaded(&'a Arc<OffloadedTimeline>),
712 : }
713 :
714 : impl TimelineOrOffloadedArcRef<'_> {
715 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
716 0 : match self {
717 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.tenant_shard_id,
718 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.tenant_shard_id,
719 : }
720 0 : }
721 0 : pub fn timeline_id(&self) -> TimelineId {
722 0 : match self {
723 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.timeline_id,
724 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.timeline_id,
725 : }
726 0 : }
727 : }
728 :
729 : impl<'a> From<&'a Arc<Timeline>> for TimelineOrOffloadedArcRef<'a> {
730 0 : fn from(timeline: &'a Arc<Timeline>) -> Self {
731 0 : Self::Timeline(timeline)
732 0 : }
733 : }
734 :
735 : impl<'a> From<&'a Arc<OffloadedTimeline>> for TimelineOrOffloadedArcRef<'a> {
736 0 : fn from(timeline: &'a Arc<OffloadedTimeline>) -> Self {
737 0 : Self::Offloaded(timeline)
738 0 : }
739 : }
740 :
741 : #[derive(Debug, thiserror::Error, PartialEq, Eq)]
742 : pub enum GetTimelineError {
743 : #[error("Timeline is shutting down")]
744 : ShuttingDown,
745 : #[error("Timeline {tenant_id}/{timeline_id} is not active, state: {state:?}")]
746 : NotActive {
747 : tenant_id: TenantShardId,
748 : timeline_id: TimelineId,
749 : state: TimelineState,
750 : },
751 : #[error("Timeline {tenant_id}/{timeline_id} was not found")]
752 : NotFound {
753 : tenant_id: TenantShardId,
754 : timeline_id: TimelineId,
755 : },
756 : }
757 :
758 : #[derive(Debug, thiserror::Error)]
759 : pub enum LoadLocalTimelineError {
760 : #[error("FailedToLoad")]
761 : Load(#[source] anyhow::Error),
762 : #[error("FailedToResumeDeletion")]
763 : ResumeDeletion(#[source] anyhow::Error),
764 : }
765 :
766 : #[derive(thiserror::Error)]
767 : pub enum DeleteTimelineError {
768 : #[error("NotFound")]
769 : NotFound,
770 :
771 : #[error("HasChildren")]
772 : HasChildren(Vec<TimelineId>),
773 :
774 : #[error("Timeline deletion is already in progress")]
775 : AlreadyInProgress(Arc<tokio::sync::Mutex<DeleteTimelineFlow>>),
776 :
777 : #[error("Cancelled")]
778 : Cancelled,
779 :
780 : #[error(transparent)]
781 : Other(#[from] anyhow::Error),
782 : }
783 :
784 : impl Debug for DeleteTimelineError {
785 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
786 0 : match self {
787 0 : Self::NotFound => write!(f, "NotFound"),
788 0 : Self::HasChildren(c) => f.debug_tuple("HasChildren").field(c).finish(),
789 0 : Self::AlreadyInProgress(_) => f.debug_tuple("AlreadyInProgress").finish(),
790 0 : Self::Cancelled => f.debug_tuple("Cancelled").finish(),
791 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
792 : }
793 0 : }
794 : }
795 :
796 : #[derive(thiserror::Error)]
797 : pub enum TimelineArchivalError {
798 : #[error("NotFound")]
799 : NotFound,
800 :
801 : #[error("Timeout")]
802 : Timeout,
803 :
804 : #[error("Cancelled")]
805 : Cancelled,
806 :
807 : #[error("ancestor is archived: {}", .0)]
808 : HasArchivedParent(TimelineId),
809 :
810 : #[error("HasUnarchivedChildren")]
811 : HasUnarchivedChildren(Vec<TimelineId>),
812 :
813 : #[error("Timeline archival is already in progress")]
814 : AlreadyInProgress,
815 :
816 : #[error(transparent)]
817 : Other(anyhow::Error),
818 : }
819 :
820 : #[derive(thiserror::Error, Debug)]
821 : pub(crate) enum TenantManifestError {
822 : #[error("Remote storage error: {0}")]
823 : RemoteStorage(anyhow::Error),
824 :
825 : #[error("Cancelled")]
826 : Cancelled,
827 : }
828 :
829 : impl From<TenantManifestError> for TimelineArchivalError {
830 0 : fn from(e: TenantManifestError) -> Self {
831 0 : match e {
832 0 : TenantManifestError::RemoteStorage(e) => Self::Other(e),
833 0 : TenantManifestError::Cancelled => Self::Cancelled,
834 : }
835 0 : }
836 : }
837 :
838 : impl Debug for TimelineArchivalError {
839 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
840 0 : match self {
841 0 : Self::NotFound => write!(f, "NotFound"),
842 0 : Self::Timeout => write!(f, "Timeout"),
843 0 : Self::Cancelled => write!(f, "Cancelled"),
844 0 : Self::HasArchivedParent(p) => f.debug_tuple("HasArchivedParent").field(p).finish(),
845 0 : Self::HasUnarchivedChildren(c) => {
846 0 : f.debug_tuple("HasUnarchivedChildren").field(c).finish()
847 : }
848 0 : Self::AlreadyInProgress => f.debug_tuple("AlreadyInProgress").finish(),
849 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
850 : }
851 0 : }
852 : }
853 :
854 : pub enum SetStoppingError {
855 : AlreadyStopping(completion::Barrier),
856 : Broken,
857 : }
858 :
859 : impl Debug for SetStoppingError {
860 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
861 0 : match self {
862 0 : Self::AlreadyStopping(_) => f.debug_tuple("AlreadyStopping").finish(),
863 0 : Self::Broken => write!(f, "Broken"),
864 : }
865 0 : }
866 : }
867 :
868 : /// Arguments to [`Tenant::create_timeline`].
869 : ///
870 : /// Not usable as an idempotency key for timeline creation because if [`CreateTimelineParamsBranch::ancestor_start_lsn`]
871 : /// is `None`, the result of the timeline create call is not deterministic.
872 : ///
873 : /// See [`CreateTimelineIdempotency`] for an idempotency key.
874 : #[derive(Debug)]
875 : pub(crate) enum CreateTimelineParams {
876 : Bootstrap(CreateTimelineParamsBootstrap),
877 : Branch(CreateTimelineParamsBranch),
878 : ImportPgdata(CreateTimelineParamsImportPgdata),
879 : }
880 :
881 : #[derive(Debug)]
882 : pub(crate) struct CreateTimelineParamsBootstrap {
883 : pub(crate) new_timeline_id: TimelineId,
884 : pub(crate) existing_initdb_timeline_id: Option<TimelineId>,
885 : pub(crate) pg_version: u32,
886 : }
887 :
888 : /// NB: See comment on [`CreateTimelineIdempotency::Branch`] for why there's no `pg_version` here.
889 : #[derive(Debug)]
890 : pub(crate) struct CreateTimelineParamsBranch {
891 : pub(crate) new_timeline_id: TimelineId,
892 : pub(crate) ancestor_timeline_id: TimelineId,
893 : pub(crate) ancestor_start_lsn: Option<Lsn>,
894 : }
895 :
896 : #[derive(Debug)]
897 : pub(crate) struct CreateTimelineParamsImportPgdata {
898 : pub(crate) new_timeline_id: TimelineId,
899 : pub(crate) location: import_pgdata::index_part_format::Location,
900 : pub(crate) idempotency_key: import_pgdata::index_part_format::IdempotencyKey,
901 : }
902 :
903 : /// What is used to determine idempotency of a [`Tenant::create_timeline`] call in [`Tenant::start_creating_timeline`] in [`Tenant::start_creating_timeline`].
904 : ///
905 : /// Each [`Timeline`] object holds [`Self`] as an immutable property in [`Timeline::create_idempotency`].
906 : ///
907 : /// We lower timeline creation requests to [`Self`], and then use [`PartialEq::eq`] to compare [`Timeline::create_idempotency`] with the request.
908 : /// If they are equal, we return a reference to the existing timeline, otherwise it's an idempotency conflict.
909 : ///
910 : /// There is special treatment for [`Self::FailWithConflict`] to always return an idempotency conflict.
911 : /// It would be nice to have more advanced derive macros to make that special treatment declarative.
912 : ///
913 : /// Notes:
914 : /// - Unlike [`CreateTimelineParams`], ancestor LSN is fixed, so, branching will be at a deterministic LSN.
915 : /// - We make some trade-offs though, e.g., [`CreateTimelineParamsBootstrap::existing_initdb_timeline_id`]
916 : /// is not considered for idempotency. We can improve on this over time if we deem it necessary.
917 : ///
918 : #[derive(Debug, Clone, PartialEq, Eq)]
919 : pub(crate) enum CreateTimelineIdempotency {
920 : /// NB: special treatment, see comment in [`Self`].
921 : FailWithConflict,
922 : Bootstrap {
923 : pg_version: u32,
924 : },
925 : /// NB: branches always have the same `pg_version` as their ancestor.
926 : /// While [`pageserver_api::models::TimelineCreateRequestMode::Branch::pg_version`]
927 : /// exists as a field, and is set by cplane, it has always been ignored by pageserver when
928 : /// determining the child branch pg_version.
929 : Branch {
930 : ancestor_timeline_id: TimelineId,
931 : ancestor_start_lsn: Lsn,
932 : },
933 : ImportPgdata(CreatingTimelineIdempotencyImportPgdata),
934 : }
935 :
936 : #[derive(Debug, Clone, PartialEq, Eq)]
937 : pub(crate) struct CreatingTimelineIdempotencyImportPgdata {
938 : idempotency_key: import_pgdata::index_part_format::IdempotencyKey,
939 : }
940 :
941 : /// What is returned by [`Tenant::start_creating_timeline`].
942 : #[must_use]
943 : enum StartCreatingTimelineResult {
944 : CreateGuard(TimelineCreateGuard),
945 : Idempotent(Arc<Timeline>),
946 : }
947 :
948 : enum TimelineInitAndSyncResult {
949 : ReadyToActivate(Arc<Timeline>),
950 : NeedsSpawnImportPgdata(TimelineInitAndSyncNeedsSpawnImportPgdata),
951 : }
952 :
953 : impl TimelineInitAndSyncResult {
954 0 : fn ready_to_activate(self) -> Option<Arc<Timeline>> {
955 0 : match self {
956 0 : Self::ReadyToActivate(timeline) => Some(timeline),
957 0 : _ => None,
958 : }
959 0 : }
960 : }
961 :
962 : #[must_use]
963 : struct TimelineInitAndSyncNeedsSpawnImportPgdata {
964 : timeline: Arc<Timeline>,
965 : import_pgdata: import_pgdata::index_part_format::Root,
966 : guard: TimelineCreateGuard,
967 : }
968 :
969 : /// What is returned by [`Tenant::create_timeline`].
970 : enum CreateTimelineResult {
971 : Created(Arc<Timeline>),
972 : Idempotent(Arc<Timeline>),
973 : /// IMPORTANT: This [`Arc<Timeline>`] object is not in [`Tenant::timelines`] when
974 : /// we return this result, nor will this concrete object ever be added there.
975 : /// Cf method comment on [`Tenant::create_timeline_import_pgdata`].
976 : ImportSpawned(Arc<Timeline>),
977 : }
978 :
979 : impl CreateTimelineResult {
980 0 : fn discriminant(&self) -> &'static str {
981 0 : match self {
982 0 : Self::Created(_) => "Created",
983 0 : Self::Idempotent(_) => "Idempotent",
984 0 : Self::ImportSpawned(_) => "ImportSpawned",
985 : }
986 0 : }
987 0 : fn timeline(&self) -> &Arc<Timeline> {
988 0 : match self {
989 0 : Self::Created(t) | Self::Idempotent(t) | Self::ImportSpawned(t) => t,
990 0 : }
991 0 : }
992 : /// Unit test timelines aren't activated, test has to do it if it needs to.
993 : #[cfg(test)]
994 460 : fn into_timeline_for_test(self) -> Arc<Timeline> {
995 460 : match self {
996 460 : Self::Created(t) | Self::Idempotent(t) | Self::ImportSpawned(t) => t,
997 460 : }
998 460 : }
999 : }
1000 :
1001 : #[derive(thiserror::Error, Debug)]
1002 : pub enum CreateTimelineError {
1003 : #[error("creation of timeline with the given ID is in progress")]
1004 : AlreadyCreating,
1005 : #[error("timeline already exists with different parameters")]
1006 : Conflict,
1007 : #[error(transparent)]
1008 : AncestorLsn(anyhow::Error),
1009 : #[error("ancestor timeline is not active")]
1010 : AncestorNotActive,
1011 : #[error("ancestor timeline is archived")]
1012 : AncestorArchived,
1013 : #[error("tenant shutting down")]
1014 : ShuttingDown,
1015 : #[error(transparent)]
1016 : Other(#[from] anyhow::Error),
1017 : }
1018 :
1019 : #[derive(thiserror::Error, Debug)]
1020 : pub enum InitdbError {
1021 : #[error("Operation was cancelled")]
1022 : Cancelled,
1023 : #[error(transparent)]
1024 : Other(anyhow::Error),
1025 : #[error(transparent)]
1026 : Inner(postgres_initdb::Error),
1027 : }
1028 :
1029 : enum CreateTimelineCause {
1030 : Load,
1031 : Delete,
1032 : }
1033 :
1034 : enum LoadTimelineCause {
1035 : Attach,
1036 : Unoffload,
1037 : ImportPgdata {
1038 : create_guard: TimelineCreateGuard,
1039 : activate: ActivateTimelineArgs,
1040 : },
1041 : }
1042 :
1043 : #[derive(thiserror::Error, Debug)]
1044 : pub(crate) enum GcError {
1045 : // The tenant is shutting down
1046 : #[error("tenant shutting down")]
1047 : TenantCancelled,
1048 :
1049 : // The tenant is shutting down
1050 : #[error("timeline shutting down")]
1051 : TimelineCancelled,
1052 :
1053 : // The tenant is in a state inelegible to run GC
1054 : #[error("not active")]
1055 : NotActive,
1056 :
1057 : // A requested GC cutoff LSN was invalid, for example it tried to move backwards
1058 : #[error("not active")]
1059 : BadLsn { why: String },
1060 :
1061 : // A remote storage error while scheduling updates after compaction
1062 : #[error(transparent)]
1063 : Remote(anyhow::Error),
1064 :
1065 : // An error reading while calculating GC cutoffs
1066 : #[error(transparent)]
1067 : GcCutoffs(PageReconstructError),
1068 :
1069 : // If GC was invoked for a particular timeline, this error means it didn't exist
1070 : #[error("timeline not found")]
1071 : TimelineNotFound,
1072 : }
1073 :
1074 : impl From<PageReconstructError> for GcError {
1075 0 : fn from(value: PageReconstructError) -> Self {
1076 0 : match value {
1077 0 : PageReconstructError::Cancelled => Self::TimelineCancelled,
1078 0 : other => Self::GcCutoffs(other),
1079 : }
1080 0 : }
1081 : }
1082 :
1083 : impl From<NotInitialized> for GcError {
1084 0 : fn from(value: NotInitialized) -> Self {
1085 0 : match value {
1086 0 : NotInitialized::Uninitialized => GcError::Remote(value.into()),
1087 0 : NotInitialized::Stopped | NotInitialized::ShuttingDown => GcError::TimelineCancelled,
1088 : }
1089 0 : }
1090 : }
1091 :
1092 : impl From<timeline::layer_manager::Shutdown> for GcError {
1093 0 : fn from(_: timeline::layer_manager::Shutdown) -> Self {
1094 0 : GcError::TimelineCancelled
1095 0 : }
1096 : }
1097 :
1098 : #[derive(thiserror::Error, Debug)]
1099 : pub(crate) enum LoadConfigError {
1100 : #[error("TOML deserialization error: '{0}'")]
1101 : DeserializeToml(#[from] toml_edit::de::Error),
1102 :
1103 : #[error("Config not found at {0}")]
1104 : NotFound(Utf8PathBuf),
1105 : }
1106 :
1107 : impl Tenant {
1108 : /// Yet another helper for timeline initialization.
1109 : ///
1110 : /// - Initializes the Timeline struct and inserts it into the tenant's hash map
1111 : /// - Scans the local timeline directory for layer files and builds the layer map
1112 : /// - Downloads remote index file and adds remote files to the layer map
1113 : /// - Schedules remote upload tasks for any files that are present locally but missing from remote storage.
1114 : ///
1115 : /// If the operation fails, the timeline is left in the tenant's hash map in Broken state. On success,
1116 : /// it is marked as Active.
1117 : #[allow(clippy::too_many_arguments)]
1118 12 : async fn timeline_init_and_sync(
1119 12 : self: &Arc<Self>,
1120 12 : timeline_id: TimelineId,
1121 12 : resources: TimelineResources,
1122 12 : mut index_part: IndexPart,
1123 12 : metadata: TimelineMetadata,
1124 12 : ancestor: Option<Arc<Timeline>>,
1125 12 : cause: LoadTimelineCause,
1126 12 : ctx: &RequestContext,
1127 12 : ) -> anyhow::Result<TimelineInitAndSyncResult> {
1128 12 : let tenant_id = self.tenant_shard_id;
1129 12 :
1130 12 : let import_pgdata = index_part.import_pgdata.take();
1131 12 : let idempotency = match &import_pgdata {
1132 0 : Some(import_pgdata) => {
1133 0 : CreateTimelineIdempotency::ImportPgdata(CreatingTimelineIdempotencyImportPgdata {
1134 0 : idempotency_key: import_pgdata.idempotency_key().clone(),
1135 0 : })
1136 : }
1137 : None => {
1138 12 : if metadata.ancestor_timeline().is_none() {
1139 8 : CreateTimelineIdempotency::Bootstrap {
1140 8 : pg_version: metadata.pg_version(),
1141 8 : }
1142 : } else {
1143 4 : CreateTimelineIdempotency::Branch {
1144 4 : ancestor_timeline_id: metadata.ancestor_timeline().unwrap(),
1145 4 : ancestor_start_lsn: metadata.ancestor_lsn(),
1146 4 : }
1147 : }
1148 : }
1149 : };
1150 :
1151 12 : let timeline = self.create_timeline_struct(
1152 12 : timeline_id,
1153 12 : &metadata,
1154 12 : ancestor.clone(),
1155 12 : resources,
1156 12 : CreateTimelineCause::Load,
1157 12 : idempotency.clone(),
1158 12 : )?;
1159 12 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
1160 12 : anyhow::ensure!(
1161 12 : disk_consistent_lsn.is_valid(),
1162 0 : "Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn"
1163 : );
1164 12 : assert_eq!(
1165 12 : disk_consistent_lsn,
1166 12 : metadata.disk_consistent_lsn(),
1167 0 : "these are used interchangeably"
1168 : );
1169 :
1170 12 : timeline.remote_client.init_upload_queue(&index_part)?;
1171 :
1172 12 : timeline
1173 12 : .load_layer_map(disk_consistent_lsn, index_part)
1174 12 : .await
1175 12 : .with_context(|| {
1176 0 : format!("Failed to load layermap for timeline {tenant_id}/{timeline_id}")
1177 12 : })?;
1178 :
1179 0 : match import_pgdata {
1180 0 : Some(import_pgdata) if !import_pgdata.is_done() => {
1181 0 : match cause {
1182 0 : LoadTimelineCause::Attach | LoadTimelineCause::Unoffload => (),
1183 : LoadTimelineCause::ImportPgdata { .. } => {
1184 0 : unreachable!("ImportPgdata should not be reloading timeline import is done and persisted as such in s3")
1185 : }
1186 : }
1187 0 : let mut guard = self.timelines_creating.lock().unwrap();
1188 0 : if !guard.insert(timeline_id) {
1189 : // We should never try and load the same timeline twice during startup
1190 0 : unreachable!("Timeline {tenant_id}/{timeline_id} is already being created")
1191 0 : }
1192 0 : let timeline_create_guard = TimelineCreateGuard {
1193 0 : _tenant_gate_guard: self.gate.enter()?,
1194 0 : owning_tenant: self.clone(),
1195 0 : timeline_id,
1196 0 : idempotency,
1197 0 : // The users of this specific return value don't need the timline_path in there.
1198 0 : timeline_path: timeline
1199 0 : .conf
1200 0 : .timeline_path(&timeline.tenant_shard_id, &timeline.timeline_id),
1201 0 : };
1202 0 : Ok(TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
1203 0 : TimelineInitAndSyncNeedsSpawnImportPgdata {
1204 0 : timeline,
1205 0 : import_pgdata,
1206 0 : guard: timeline_create_guard,
1207 0 : },
1208 0 : ))
1209 : }
1210 : Some(_) | None => {
1211 : {
1212 12 : let mut timelines_accessor = self.timelines.lock().unwrap();
1213 12 : match timelines_accessor.entry(timeline_id) {
1214 : // We should never try and load the same timeline twice during startup
1215 : Entry::Occupied(_) => {
1216 0 : unreachable!(
1217 0 : "Timeline {tenant_id}/{timeline_id} already exists in the tenant map"
1218 0 : );
1219 : }
1220 12 : Entry::Vacant(v) => {
1221 12 : v.insert(Arc::clone(&timeline));
1222 12 : timeline.maybe_spawn_flush_loop();
1223 12 : }
1224 : }
1225 : }
1226 :
1227 : // Sanity check: a timeline should have some content.
1228 12 : anyhow::ensure!(
1229 12 : ancestor.is_some()
1230 8 : || timeline
1231 8 : .layers
1232 8 : .read()
1233 8 : .await
1234 8 : .layer_map()
1235 8 : .expect("currently loading, layer manager cannot be shutdown already")
1236 8 : .iter_historic_layers()
1237 8 : .next()
1238 8 : .is_some(),
1239 0 : "Timeline has no ancestor and no layer files"
1240 : );
1241 :
1242 12 : match cause {
1243 12 : LoadTimelineCause::Attach | LoadTimelineCause::Unoffload => (),
1244 : LoadTimelineCause::ImportPgdata {
1245 0 : create_guard,
1246 0 : activate,
1247 0 : } => {
1248 0 : // TODO: see the comment in the task code above how I'm not so certain
1249 0 : // it is safe to activate here because of concurrent shutdowns.
1250 0 : match activate {
1251 0 : ActivateTimelineArgs::Yes { broker_client } => {
1252 0 : info!("activating timeline after reload from pgdata import task");
1253 0 : timeline.activate(self.clone(), broker_client, None, ctx);
1254 : }
1255 0 : ActivateTimelineArgs::No => (),
1256 : }
1257 0 : drop(create_guard);
1258 : }
1259 : }
1260 :
1261 12 : Ok(TimelineInitAndSyncResult::ReadyToActivate(timeline))
1262 : }
1263 : }
1264 12 : }
1265 :
1266 : /// Attach a tenant that's available in cloud storage.
1267 : ///
1268 : /// This returns quickly, after just creating the in-memory object
1269 : /// Tenant struct and launching a background task to download
1270 : /// the remote index files. On return, the tenant is most likely still in
1271 : /// Attaching state, and it will become Active once the background task
1272 : /// finishes. You can use wait_until_active() to wait for the task to
1273 : /// complete.
1274 : ///
1275 : #[allow(clippy::too_many_arguments)]
1276 0 : pub(crate) fn spawn(
1277 0 : conf: &'static PageServerConf,
1278 0 : tenant_shard_id: TenantShardId,
1279 0 : resources: TenantSharedResources,
1280 0 : attached_conf: AttachedTenantConf,
1281 0 : shard_identity: ShardIdentity,
1282 0 : init_order: Option<InitializationOrder>,
1283 0 : mode: SpawnMode,
1284 0 : ctx: &RequestContext,
1285 0 : ) -> Result<Arc<Tenant>, GlobalShutDown> {
1286 0 : let wal_redo_manager =
1287 0 : WalRedoManager::new(PostgresRedoManager::new(conf, tenant_shard_id))?;
1288 :
1289 : let TenantSharedResources {
1290 0 : broker_client,
1291 0 : remote_storage,
1292 0 : deletion_queue_client,
1293 0 : l0_flush_global_state,
1294 0 : } = resources;
1295 0 :
1296 0 : let attach_mode = attached_conf.location.attach_mode;
1297 0 : let generation = attached_conf.location.generation;
1298 0 :
1299 0 : let tenant = Arc::new(Tenant::new(
1300 0 : TenantState::Attaching,
1301 0 : conf,
1302 0 : attached_conf,
1303 0 : shard_identity,
1304 0 : Some(wal_redo_manager),
1305 0 : tenant_shard_id,
1306 0 : remote_storage.clone(),
1307 0 : deletion_queue_client,
1308 0 : l0_flush_global_state,
1309 0 : ));
1310 0 :
1311 0 : // The attach task will carry a GateGuard, so that shutdown() reliably waits for it to drop out if
1312 0 : // we shut down while attaching.
1313 0 : let attach_gate_guard = tenant
1314 0 : .gate
1315 0 : .enter()
1316 0 : .expect("We just created the Tenant: nothing else can have shut it down yet");
1317 0 :
1318 0 : // Do all the hard work in the background
1319 0 : let tenant_clone = Arc::clone(&tenant);
1320 0 : let ctx = ctx.detached_child(TaskKind::Attach, DownloadBehavior::Warn);
1321 0 : task_mgr::spawn(
1322 0 : &tokio::runtime::Handle::current(),
1323 0 : TaskKind::Attach,
1324 0 : tenant_shard_id,
1325 0 : None,
1326 0 : "attach tenant",
1327 0 : async move {
1328 0 :
1329 0 : info!(
1330 : ?attach_mode,
1331 0 : "Attaching tenant"
1332 : );
1333 :
1334 0 : let _gate_guard = attach_gate_guard;
1335 0 :
1336 0 : // Is this tenant being spawned as part of process startup?
1337 0 : let starting_up = init_order.is_some();
1338 0 : scopeguard::defer! {
1339 0 : if starting_up {
1340 0 : TENANT.startup_complete.inc();
1341 0 : }
1342 0 : }
1343 :
1344 : // Ideally we should use Tenant::set_broken_no_wait, but it is not supposed to be used when tenant is in loading state.
1345 : enum BrokenVerbosity {
1346 : Error,
1347 : Info
1348 : }
1349 0 : let make_broken =
1350 0 : |t: &Tenant, err: anyhow::Error, verbosity: BrokenVerbosity| {
1351 0 : match verbosity {
1352 : BrokenVerbosity::Info => {
1353 0 : info!("attach cancelled, setting tenant state to Broken: {err}");
1354 : },
1355 : BrokenVerbosity::Error => {
1356 0 : error!("attach failed, setting tenant state to Broken: {err:?}");
1357 : }
1358 : }
1359 0 : t.state.send_modify(|state| {
1360 0 : // The Stopping case is for when we have passed control on to DeleteTenantFlow:
1361 0 : // if it errors, we will call make_broken when tenant is already in Stopping.
1362 0 : assert!(
1363 0 : matches!(*state, TenantState::Attaching | TenantState::Stopping { .. }),
1364 0 : "the attach task owns the tenant state until activation is complete"
1365 : );
1366 :
1367 0 : *state = TenantState::broken_from_reason(err.to_string());
1368 0 : });
1369 0 : };
1370 :
1371 : // TODO: should also be rejecting tenant conf changes that violate this check.
1372 0 : if let Err(e) = crate::tenant::storage_layer::inmemory_layer::IndexEntry::validate_checkpoint_distance(tenant_clone.get_checkpoint_distance()) {
1373 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1374 0 : return Ok(());
1375 0 : }
1376 0 :
1377 0 : let mut init_order = init_order;
1378 0 : // take the completion because initial tenant loading will complete when all of
1379 0 : // these tasks complete.
1380 0 : let _completion = init_order
1381 0 : .as_mut()
1382 0 : .and_then(|x| x.initial_tenant_load.take());
1383 0 : let remote_load_completion = init_order
1384 0 : .as_mut()
1385 0 : .and_then(|x| x.initial_tenant_load_remote.take());
1386 :
1387 : enum AttachType<'a> {
1388 : /// We are attaching this tenant lazily in the background.
1389 : Warmup {
1390 : _permit: tokio::sync::SemaphorePermit<'a>,
1391 : during_startup: bool
1392 : },
1393 : /// We are attaching this tenant as soon as we can, because for example an
1394 : /// endpoint tried to access it.
1395 : OnDemand,
1396 : /// During normal operations after startup, we are attaching a tenant, and
1397 : /// eager attach was requested.
1398 : Normal,
1399 : }
1400 :
1401 0 : let attach_type = if matches!(mode, SpawnMode::Lazy) {
1402 : // Before doing any I/O, wait for at least one of:
1403 : // - A client attempting to access to this tenant (on-demand loading)
1404 : // - A permit becoming available in the warmup semaphore (background warmup)
1405 :
1406 0 : tokio::select!(
1407 0 : permit = tenant_clone.activate_now_sem.acquire() => {
1408 0 : let _ = permit.expect("activate_now_sem is never closed");
1409 0 : tracing::info!("Activating tenant (on-demand)");
1410 0 : AttachType::OnDemand
1411 : },
1412 0 : permit = conf.concurrent_tenant_warmup.inner().acquire() => {
1413 0 : let _permit = permit.expect("concurrent_tenant_warmup semaphore is never closed");
1414 0 : tracing::info!("Activating tenant (warmup)");
1415 0 : AttachType::Warmup {
1416 0 : _permit,
1417 0 : during_startup: init_order.is_some()
1418 0 : }
1419 : }
1420 0 : _ = tenant_clone.cancel.cancelled() => {
1421 : // This is safe, but should be pretty rare: it is interesting if a tenant
1422 : // stayed in Activating for such a long time that shutdown found it in
1423 : // that state.
1424 0 : tracing::info!(state=%tenant_clone.current_state(), "Tenant shut down before activation");
1425 : // Make the tenant broken so that set_stopping will not hang waiting for it to leave
1426 : // the Attaching state. This is an over-reaction (nothing really broke, the tenant is
1427 : // just shutting down), but ensures progress.
1428 0 : make_broken(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"), BrokenVerbosity::Info);
1429 0 : return Ok(());
1430 : },
1431 : )
1432 : } else {
1433 : // SpawnMode::{Create,Eager} always cause jumping ahead of the
1434 : // concurrent_tenant_warmup queue
1435 0 : AttachType::Normal
1436 : };
1437 :
1438 0 : let preload = match &mode {
1439 : SpawnMode::Eager | SpawnMode::Lazy => {
1440 0 : let _preload_timer = TENANT.preload.start_timer();
1441 0 : let res = tenant_clone
1442 0 : .preload(&remote_storage, task_mgr::shutdown_token())
1443 0 : .await;
1444 0 : match res {
1445 0 : Ok(p) => Some(p),
1446 0 : Err(e) => {
1447 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1448 0 : return Ok(());
1449 : }
1450 : }
1451 : }
1452 :
1453 : };
1454 :
1455 : // Remote preload is complete.
1456 0 : drop(remote_load_completion);
1457 0 :
1458 0 :
1459 0 : // We will time the duration of the attach phase unless this is a creation (attach will do no work)
1460 0 : let attach_start = std::time::Instant::now();
1461 0 : let attached = {
1462 0 : let _attach_timer = Some(TENANT.attach.start_timer());
1463 0 : tenant_clone.attach(preload, &ctx).await
1464 : };
1465 0 : let attach_duration = attach_start.elapsed();
1466 0 : _ = tenant_clone.attach_wal_lag_cooldown.set(WalLagCooldown::new(attach_start, attach_duration));
1467 0 :
1468 0 : match attached {
1469 : Ok(()) => {
1470 0 : info!("attach finished, activating");
1471 0 : tenant_clone.activate(broker_client, None, &ctx);
1472 : }
1473 0 : Err(e) => {
1474 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1475 0 : }
1476 : }
1477 :
1478 : // If we are doing an opportunistic warmup attachment at startup, initialize
1479 : // logical size at the same time. This is better than starting a bunch of idle tenants
1480 : // with cold caches and then coming back later to initialize their logical sizes.
1481 : //
1482 : // It also prevents the warmup proccess competing with the concurrency limit on
1483 : // logical size calculations: if logical size calculation semaphore is saturated,
1484 : // then warmup will wait for that before proceeding to the next tenant.
1485 0 : if matches!(attach_type, AttachType::Warmup { during_startup: true, .. }) {
1486 0 : let mut futs: FuturesUnordered<_> = tenant_clone.timelines.lock().unwrap().values().cloned().map(|t| t.await_initial_logical_size()).collect();
1487 0 : tracing::info!("Waiting for initial logical sizes while warming up...");
1488 0 : while futs.next().await.is_some() {}
1489 0 : tracing::info!("Warm-up complete");
1490 0 : }
1491 :
1492 0 : Ok(())
1493 0 : }
1494 0 : .instrument(tracing::info_span!(parent: None, "attach", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), gen=?generation)),
1495 : );
1496 0 : Ok(tenant)
1497 0 : }
1498 :
1499 : #[instrument(skip_all)]
1500 : pub(crate) async fn preload(
1501 : self: &Arc<Self>,
1502 : remote_storage: &GenericRemoteStorage,
1503 : cancel: CancellationToken,
1504 : ) -> anyhow::Result<TenantPreload> {
1505 : span::debug_assert_current_span_has_tenant_id();
1506 : // Get list of remote timelines
1507 : // download index files for every tenant timeline
1508 : info!("listing remote timelines");
1509 : let (mut remote_timeline_ids, other_keys) = remote_timeline_client::list_remote_timelines(
1510 : remote_storage,
1511 : self.tenant_shard_id,
1512 : cancel.clone(),
1513 : )
1514 : .await?;
1515 : let (offloaded_add, tenant_manifest) =
1516 : match remote_timeline_client::download_tenant_manifest(
1517 : remote_storage,
1518 : &self.tenant_shard_id,
1519 : self.generation,
1520 : &cancel,
1521 : )
1522 : .await
1523 : {
1524 : Ok((tenant_manifest, _generation, _manifest_mtime)) => (
1525 : format!("{} offloaded", tenant_manifest.offloaded_timelines.len()),
1526 : tenant_manifest,
1527 : ),
1528 : Err(DownloadError::NotFound) => {
1529 : ("no manifest".to_string(), TenantManifest::empty())
1530 : }
1531 : Err(e) => Err(e)?,
1532 : };
1533 :
1534 : info!(
1535 : "found {} timelines, and {offloaded_add}",
1536 : remote_timeline_ids.len()
1537 : );
1538 :
1539 : for k in other_keys {
1540 : warn!("Unexpected non timeline key {k}");
1541 : }
1542 :
1543 : // Avoid downloading IndexPart of offloaded timelines.
1544 : let mut offloaded_with_prefix = HashSet::new();
1545 : for offloaded in tenant_manifest.offloaded_timelines.iter() {
1546 : if remote_timeline_ids.remove(&offloaded.timeline_id) {
1547 : offloaded_with_prefix.insert(offloaded.timeline_id);
1548 : } else {
1549 : // We'll take care later of timelines in the manifest without a prefix
1550 : }
1551 : }
1552 :
1553 : let timelines = self
1554 : .load_timelines_metadata(remote_timeline_ids, remote_storage, cancel)
1555 : .await?;
1556 :
1557 : Ok(TenantPreload {
1558 : tenant_manifest,
1559 : timelines: timelines
1560 : .into_iter()
1561 12 : .map(|(id, tl)| (id, Some(tl)))
1562 0 : .chain(offloaded_with_prefix.into_iter().map(|id| (id, None)))
1563 : .collect(),
1564 : })
1565 : }
1566 :
1567 : ///
1568 : /// Background task that downloads all data for a tenant and brings it to Active state.
1569 : ///
1570 : /// No background tasks are started as part of this routine.
1571 : ///
1572 440 : async fn attach(
1573 440 : self: &Arc<Tenant>,
1574 440 : preload: Option<TenantPreload>,
1575 440 : ctx: &RequestContext,
1576 440 : ) -> anyhow::Result<()> {
1577 440 : span::debug_assert_current_span_has_tenant_id();
1578 440 :
1579 440 : failpoint_support::sleep_millis_async!("before-attaching-tenant");
1580 :
1581 440 : let Some(preload) = preload else {
1582 0 : anyhow::bail!("local-only deployment is no longer supported, https://github.com/neondatabase/neon/issues/5624");
1583 : };
1584 :
1585 440 : let mut offloaded_timeline_ids = HashSet::new();
1586 440 : let mut offloaded_timelines_list = Vec::new();
1587 440 : for timeline_manifest in preload.tenant_manifest.offloaded_timelines.iter() {
1588 0 : let timeline_id = timeline_manifest.timeline_id;
1589 0 : let offloaded_timeline =
1590 0 : OffloadedTimeline::from_manifest(self.tenant_shard_id, timeline_manifest);
1591 0 : offloaded_timelines_list.push((timeline_id, Arc::new(offloaded_timeline)));
1592 0 : offloaded_timeline_ids.insert(timeline_id);
1593 0 : }
1594 : // Complete deletions for offloaded timeline id's from manifest.
1595 : // The manifest will be uploaded later in this function.
1596 440 : offloaded_timelines_list
1597 440 : .retain(|(offloaded_id, offloaded)| {
1598 0 : // Existence of a timeline is finally determined by the existence of an index-part.json in remote storage.
1599 0 : // If there is dangling references in another location, they need to be cleaned up.
1600 0 : let delete = !preload.timelines.contains_key(offloaded_id);
1601 0 : if delete {
1602 0 : tracing::info!("Removing offloaded timeline {offloaded_id} from manifest as no remote prefix was found");
1603 0 : offloaded.defuse_for_tenant_drop();
1604 0 : }
1605 0 : !delete
1606 440 : });
1607 440 :
1608 440 : let mut timelines_to_resume_deletions = vec![];
1609 440 :
1610 440 : let mut remote_index_and_client = HashMap::new();
1611 440 : let mut timeline_ancestors = HashMap::new();
1612 440 : let mut existent_timelines = HashSet::new();
1613 452 : for (timeline_id, preload) in preload.timelines {
1614 12 : let Some(preload) = preload else { continue };
1615 : // This is an invariant of the `preload` function's API
1616 12 : assert!(!offloaded_timeline_ids.contains(&timeline_id));
1617 12 : let index_part = match preload.index_part {
1618 12 : Ok(i) => {
1619 12 : debug!("remote index part exists for timeline {timeline_id}");
1620 : // We found index_part on the remote, this is the standard case.
1621 12 : existent_timelines.insert(timeline_id);
1622 12 : i
1623 : }
1624 : Err(DownloadError::NotFound) => {
1625 : // There is no index_part on the remote. We only get here
1626 : // if there is some prefix for the timeline in the remote storage.
1627 : // This can e.g. be the initdb.tar.zst archive, maybe a
1628 : // remnant from a prior incomplete creation or deletion attempt.
1629 : // Delete the local directory as the deciding criterion for a
1630 : // timeline's existence is presence of index_part.
1631 0 : info!(%timeline_id, "index_part not found on remote");
1632 0 : continue;
1633 : }
1634 0 : Err(DownloadError::Fatal(why)) => {
1635 0 : // If, while loading one remote timeline, we saw an indication that our generation
1636 0 : // number is likely invalid, then we should not load the whole tenant.
1637 0 : error!(%timeline_id, "Fatal error loading timeline: {why}");
1638 0 : anyhow::bail!(why.to_string());
1639 : }
1640 0 : Err(e) => {
1641 0 : // Some (possibly ephemeral) error happened during index_part download.
1642 0 : // Pretend the timeline exists to not delete the timeline directory,
1643 0 : // as it might be a temporary issue and we don't want to re-download
1644 0 : // everything after it resolves.
1645 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1646 :
1647 0 : existent_timelines.insert(timeline_id);
1648 0 : continue;
1649 : }
1650 : };
1651 12 : match index_part {
1652 12 : MaybeDeletedIndexPart::IndexPart(index_part) => {
1653 12 : timeline_ancestors.insert(timeline_id, index_part.metadata.clone());
1654 12 : remote_index_and_client.insert(timeline_id, (index_part, preload.client));
1655 12 : }
1656 0 : MaybeDeletedIndexPart::Deleted(index_part) => {
1657 0 : info!(
1658 0 : "timeline {} is deleted, picking to resume deletion",
1659 : timeline_id
1660 : );
1661 0 : timelines_to_resume_deletions.push((timeline_id, index_part, preload.client));
1662 : }
1663 : }
1664 : }
1665 :
1666 440 : let mut gc_blocks = HashMap::new();
1667 :
1668 : // For every timeline, download the metadata file, scan the local directory,
1669 : // and build a layer map that contains an entry for each remote and local
1670 : // layer file.
1671 440 : let sorted_timelines = tree_sort_timelines(timeline_ancestors, |m| m.ancestor_timeline())?;
1672 452 : for (timeline_id, remote_metadata) in sorted_timelines {
1673 12 : let (index_part, remote_client) = remote_index_and_client
1674 12 : .remove(&timeline_id)
1675 12 : .expect("just put it in above");
1676 :
1677 12 : if let Some(blocking) = index_part.gc_blocking.as_ref() {
1678 : // could just filter these away, but it helps while testing
1679 0 : anyhow::ensure!(
1680 0 : !blocking.reasons.is_empty(),
1681 0 : "index_part for {timeline_id} is malformed: it should not have gc blocking with zero reasons"
1682 : );
1683 0 : let prev = gc_blocks.insert(timeline_id, blocking.reasons);
1684 0 : assert!(prev.is_none());
1685 12 : }
1686 :
1687 : // TODO again handle early failure
1688 12 : let effect = self
1689 12 : .load_remote_timeline(
1690 12 : timeline_id,
1691 12 : index_part,
1692 12 : remote_metadata,
1693 12 : TimelineResources {
1694 12 : remote_client,
1695 12 : pagestream_throttle: self.pagestream_throttle.clone(),
1696 12 : pagestream_throttle_metrics: self.pagestream_throttle_metrics.clone(),
1697 12 : l0_flush_global_state: self.l0_flush_global_state.clone(),
1698 12 : },
1699 12 : LoadTimelineCause::Attach,
1700 12 : ctx,
1701 12 : )
1702 12 : .await
1703 12 : .with_context(|| {
1704 0 : format!(
1705 0 : "failed to load remote timeline {} for tenant {}",
1706 0 : timeline_id, self.tenant_shard_id
1707 0 : )
1708 12 : })?;
1709 :
1710 12 : match effect {
1711 12 : TimelineInitAndSyncResult::ReadyToActivate(_) => {
1712 12 : // activation happens later, on Tenant::activate
1713 12 : }
1714 : TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
1715 : TimelineInitAndSyncNeedsSpawnImportPgdata {
1716 0 : timeline,
1717 0 : import_pgdata,
1718 0 : guard,
1719 0 : },
1720 0 : ) => {
1721 0 : tokio::task::spawn(self.clone().create_timeline_import_pgdata_task(
1722 0 : timeline,
1723 0 : import_pgdata,
1724 0 : ActivateTimelineArgs::No,
1725 0 : guard,
1726 0 : ));
1727 0 : }
1728 : }
1729 : }
1730 :
1731 : // Walk through deleted timelines, resume deletion
1732 440 : for (timeline_id, index_part, remote_timeline_client) in timelines_to_resume_deletions {
1733 0 : remote_timeline_client
1734 0 : .init_upload_queue_stopped_to_continue_deletion(&index_part)
1735 0 : .context("init queue stopped")
1736 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1737 :
1738 0 : DeleteTimelineFlow::resume_deletion(
1739 0 : Arc::clone(self),
1740 0 : timeline_id,
1741 0 : &index_part.metadata,
1742 0 : remote_timeline_client,
1743 0 : )
1744 0 : .instrument(tracing::info_span!("timeline_delete", %timeline_id))
1745 0 : .await
1746 0 : .context("resume_deletion")
1747 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1748 : }
1749 440 : let needs_manifest_upload =
1750 440 : offloaded_timelines_list.len() != preload.tenant_manifest.offloaded_timelines.len();
1751 440 : {
1752 440 : let mut offloaded_timelines_accessor = self.timelines_offloaded.lock().unwrap();
1753 440 : offloaded_timelines_accessor.extend(offloaded_timelines_list.into_iter());
1754 440 : }
1755 440 : if needs_manifest_upload {
1756 0 : self.store_tenant_manifest().await?;
1757 440 : }
1758 :
1759 : // The local filesystem contents are a cache of what's in the remote IndexPart;
1760 : // IndexPart is the source of truth.
1761 440 : self.clean_up_timelines(&existent_timelines)?;
1762 :
1763 440 : self.gc_block.set_scanned(gc_blocks);
1764 440 :
1765 440 : fail::fail_point!("attach-before-activate", |_| {
1766 0 : anyhow::bail!("attach-before-activate");
1767 440 : });
1768 440 : failpoint_support::sleep_millis_async!("attach-before-activate-sleep", &self.cancel);
1769 :
1770 440 : info!("Done");
1771 :
1772 440 : Ok(())
1773 440 : }
1774 :
1775 : /// Check for any local timeline directories that are temporary, or do not correspond to a
1776 : /// timeline that still exists: this can happen if we crashed during a deletion/creation, or
1777 : /// if a timeline was deleted while the tenant was attached to a different pageserver.
1778 440 : fn clean_up_timelines(&self, existent_timelines: &HashSet<TimelineId>) -> anyhow::Result<()> {
1779 440 : let timelines_dir = self.conf.timelines_path(&self.tenant_shard_id);
1780 :
1781 440 : let entries = match timelines_dir.read_dir_utf8() {
1782 440 : Ok(d) => d,
1783 0 : Err(e) => {
1784 0 : if e.kind() == std::io::ErrorKind::NotFound {
1785 0 : return Ok(());
1786 : } else {
1787 0 : return Err(e).context("list timelines directory for tenant");
1788 : }
1789 : }
1790 : };
1791 :
1792 456 : for entry in entries {
1793 16 : let entry = entry.context("read timeline dir entry")?;
1794 16 : let entry_path = entry.path();
1795 :
1796 16 : let purge = if crate::is_temporary(entry_path)
1797 : // TODO: remove uninit mark code (https://github.com/neondatabase/neon/issues/5718)
1798 16 : || is_uninit_mark(entry_path)
1799 16 : || crate::is_delete_mark(entry_path)
1800 : {
1801 0 : true
1802 : } else {
1803 16 : match TimelineId::try_from(entry_path.file_name()) {
1804 16 : Ok(i) => {
1805 16 : // Purge if the timeline ID does not exist in remote storage: remote storage is the authority.
1806 16 : !existent_timelines.contains(&i)
1807 : }
1808 0 : Err(e) => {
1809 0 : tracing::warn!(
1810 0 : "Unparseable directory in timelines directory: {entry_path}, ignoring ({e})"
1811 : );
1812 : // Do not purge junk: if we don't recognize it, be cautious and leave it for a human.
1813 0 : false
1814 : }
1815 : }
1816 : };
1817 :
1818 16 : if purge {
1819 4 : tracing::info!("Purging stale timeline dentry {entry_path}");
1820 4 : if let Err(e) = match entry.file_type() {
1821 4 : Ok(t) => if t.is_dir() {
1822 4 : std::fs::remove_dir_all(entry_path)
1823 : } else {
1824 0 : std::fs::remove_file(entry_path)
1825 : }
1826 4 : .or_else(fs_ext::ignore_not_found),
1827 0 : Err(e) => Err(e),
1828 : } {
1829 0 : tracing::warn!("Failed to purge stale timeline dentry {entry_path}: {e}");
1830 4 : }
1831 12 : }
1832 : }
1833 :
1834 440 : Ok(())
1835 440 : }
1836 :
1837 : /// Get sum of all remote timelines sizes
1838 : ///
1839 : /// This function relies on the index_part instead of listing the remote storage
1840 0 : pub fn remote_size(&self) -> u64 {
1841 0 : let mut size = 0;
1842 :
1843 0 : for timeline in self.list_timelines() {
1844 0 : size += timeline.remote_client.get_remote_physical_size();
1845 0 : }
1846 :
1847 0 : size
1848 0 : }
1849 :
1850 : #[instrument(skip_all, fields(timeline_id=%timeline_id))]
1851 : async fn load_remote_timeline(
1852 : self: &Arc<Self>,
1853 : timeline_id: TimelineId,
1854 : index_part: IndexPart,
1855 : remote_metadata: TimelineMetadata,
1856 : resources: TimelineResources,
1857 : cause: LoadTimelineCause,
1858 : ctx: &RequestContext,
1859 : ) -> anyhow::Result<TimelineInitAndSyncResult> {
1860 : span::debug_assert_current_span_has_tenant_id();
1861 :
1862 : info!("downloading index file for timeline {}", timeline_id);
1863 : tokio::fs::create_dir_all(self.conf.timeline_path(&self.tenant_shard_id, &timeline_id))
1864 : .await
1865 : .context("Failed to create new timeline directory")?;
1866 :
1867 : let ancestor = if let Some(ancestor_id) = remote_metadata.ancestor_timeline() {
1868 : let timelines = self.timelines.lock().unwrap();
1869 : Some(Arc::clone(timelines.get(&ancestor_id).ok_or_else(
1870 0 : || {
1871 0 : anyhow::anyhow!(
1872 0 : "cannot find ancestor timeline {ancestor_id} for timeline {timeline_id}"
1873 0 : )
1874 0 : },
1875 : )?))
1876 : } else {
1877 : None
1878 : };
1879 :
1880 : self.timeline_init_and_sync(
1881 : timeline_id,
1882 : resources,
1883 : index_part,
1884 : remote_metadata,
1885 : ancestor,
1886 : cause,
1887 : ctx,
1888 : )
1889 : .await
1890 : }
1891 :
1892 440 : async fn load_timelines_metadata(
1893 440 : self: &Arc<Tenant>,
1894 440 : timeline_ids: HashSet<TimelineId>,
1895 440 : remote_storage: &GenericRemoteStorage,
1896 440 : cancel: CancellationToken,
1897 440 : ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
1898 440 : let mut part_downloads = JoinSet::new();
1899 452 : for timeline_id in timeline_ids {
1900 12 : let cancel_clone = cancel.clone();
1901 12 : part_downloads.spawn(
1902 12 : self.load_timeline_metadata(timeline_id, remote_storage.clone(), cancel_clone)
1903 12 : .instrument(info_span!("download_index_part", %timeline_id)),
1904 : );
1905 : }
1906 :
1907 440 : let mut timeline_preloads: HashMap<TimelineId, TimelinePreload> = HashMap::new();
1908 :
1909 : loop {
1910 452 : tokio::select!(
1911 452 : next = part_downloads.join_next() => {
1912 452 : match next {
1913 12 : Some(result) => {
1914 12 : let preload = result.context("join preload task")?;
1915 12 : timeline_preloads.insert(preload.timeline_id, preload);
1916 : },
1917 : None => {
1918 440 : break;
1919 : }
1920 : }
1921 : },
1922 452 : _ = cancel.cancelled() => {
1923 0 : anyhow::bail!("Cancelled while waiting for remote index download")
1924 : }
1925 : )
1926 : }
1927 :
1928 440 : Ok(timeline_preloads)
1929 440 : }
1930 :
1931 12 : fn build_timeline_client(
1932 12 : &self,
1933 12 : timeline_id: TimelineId,
1934 12 : remote_storage: GenericRemoteStorage,
1935 12 : ) -> RemoteTimelineClient {
1936 12 : RemoteTimelineClient::new(
1937 12 : remote_storage.clone(),
1938 12 : self.deletion_queue_client.clone(),
1939 12 : self.conf,
1940 12 : self.tenant_shard_id,
1941 12 : timeline_id,
1942 12 : self.generation,
1943 12 : &self.tenant_conf.load().location,
1944 12 : )
1945 12 : }
1946 :
1947 12 : fn load_timeline_metadata(
1948 12 : self: &Arc<Tenant>,
1949 12 : timeline_id: TimelineId,
1950 12 : remote_storage: GenericRemoteStorage,
1951 12 : cancel: CancellationToken,
1952 12 : ) -> impl Future<Output = TimelinePreload> {
1953 12 : let client = self.build_timeline_client(timeline_id, remote_storage);
1954 12 : async move {
1955 12 : debug_assert_current_span_has_tenant_and_timeline_id();
1956 12 : debug!("starting index part download");
1957 :
1958 12 : let index_part = client.download_index_file(&cancel).await;
1959 :
1960 12 : debug!("finished index part download");
1961 :
1962 12 : TimelinePreload {
1963 12 : client,
1964 12 : timeline_id,
1965 12 : index_part,
1966 12 : }
1967 12 : }
1968 12 : }
1969 :
1970 0 : fn check_to_be_archived_has_no_unarchived_children(
1971 0 : timeline_id: TimelineId,
1972 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
1973 0 : ) -> Result<(), TimelineArchivalError> {
1974 0 : let children: Vec<TimelineId> = timelines
1975 0 : .iter()
1976 0 : .filter_map(|(id, entry)| {
1977 0 : if entry.get_ancestor_timeline_id() != Some(timeline_id) {
1978 0 : return None;
1979 0 : }
1980 0 : if entry.is_archived() == Some(true) {
1981 0 : return None;
1982 0 : }
1983 0 : Some(*id)
1984 0 : })
1985 0 : .collect();
1986 0 :
1987 0 : if !children.is_empty() {
1988 0 : return Err(TimelineArchivalError::HasUnarchivedChildren(children));
1989 0 : }
1990 0 : Ok(())
1991 0 : }
1992 :
1993 0 : fn check_ancestor_of_to_be_unarchived_is_not_archived(
1994 0 : ancestor_timeline_id: TimelineId,
1995 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
1996 0 : offloaded_timelines: &std::sync::MutexGuard<
1997 0 : '_,
1998 0 : HashMap<TimelineId, Arc<OffloadedTimeline>>,
1999 0 : >,
2000 0 : ) -> Result<(), TimelineArchivalError> {
2001 0 : let has_archived_parent =
2002 0 : if let Some(ancestor_timeline) = timelines.get(&ancestor_timeline_id) {
2003 0 : ancestor_timeline.is_archived() == Some(true)
2004 0 : } else if offloaded_timelines.contains_key(&ancestor_timeline_id) {
2005 0 : true
2006 : } else {
2007 0 : error!("ancestor timeline {ancestor_timeline_id} not found");
2008 0 : if cfg!(debug_assertions) {
2009 0 : panic!("ancestor timeline {ancestor_timeline_id} not found");
2010 0 : }
2011 0 : return Err(TimelineArchivalError::NotFound);
2012 : };
2013 0 : if has_archived_parent {
2014 0 : return Err(TimelineArchivalError::HasArchivedParent(
2015 0 : ancestor_timeline_id,
2016 0 : ));
2017 0 : }
2018 0 : Ok(())
2019 0 : }
2020 :
2021 0 : fn check_to_be_unarchived_timeline_has_no_archived_parent(
2022 0 : timeline: &Arc<Timeline>,
2023 0 : ) -> Result<(), TimelineArchivalError> {
2024 0 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
2025 0 : if ancestor_timeline.is_archived() == Some(true) {
2026 0 : return Err(TimelineArchivalError::HasArchivedParent(
2027 0 : ancestor_timeline.timeline_id,
2028 0 : ));
2029 0 : }
2030 0 : }
2031 0 : Ok(())
2032 0 : }
2033 :
2034 : /// Loads the specified (offloaded) timeline from S3 and attaches it as a loaded timeline
2035 : ///
2036 : /// Counterpart to [`offload_timeline`].
2037 0 : async fn unoffload_timeline(
2038 0 : self: &Arc<Self>,
2039 0 : timeline_id: TimelineId,
2040 0 : broker_client: storage_broker::BrokerClientChannel,
2041 0 : ctx: RequestContext,
2042 0 : ) -> Result<Arc<Timeline>, TimelineArchivalError> {
2043 0 : info!("unoffloading timeline");
2044 :
2045 : // We activate the timeline below manually, so this must be called on an active tenant.
2046 : // We expect callers of this function to ensure this.
2047 0 : match self.current_state() {
2048 : TenantState::Activating { .. }
2049 : | TenantState::Attaching
2050 : | TenantState::Broken { .. } => {
2051 0 : panic!("Timeline expected to be active")
2052 : }
2053 0 : TenantState::Stopping { .. } => return Err(TimelineArchivalError::Cancelled),
2054 0 : TenantState::Active => {}
2055 0 : }
2056 0 : let cancel = self.cancel.clone();
2057 0 :
2058 0 : // Protect against concurrent attempts to use this TimelineId
2059 0 : // We don't care much about idempotency, as it's ensured a layer above.
2060 0 : let allow_offloaded = true;
2061 0 : let _create_guard = self
2062 0 : .create_timeline_create_guard(
2063 0 : timeline_id,
2064 0 : CreateTimelineIdempotency::FailWithConflict,
2065 0 : allow_offloaded,
2066 0 : )
2067 0 : .map_err(|err| match err {
2068 0 : TimelineExclusionError::AlreadyCreating => TimelineArchivalError::AlreadyInProgress,
2069 : TimelineExclusionError::AlreadyExists { .. } => {
2070 0 : TimelineArchivalError::Other(anyhow::anyhow!("Timeline already exists"))
2071 : }
2072 0 : TimelineExclusionError::Other(e) => TimelineArchivalError::Other(e),
2073 0 : TimelineExclusionError::ShuttingDown => TimelineArchivalError::Cancelled,
2074 0 : })?;
2075 :
2076 0 : let timeline_preload = self
2077 0 : .load_timeline_metadata(timeline_id, self.remote_storage.clone(), cancel.clone())
2078 0 : .await;
2079 :
2080 0 : let index_part = match timeline_preload.index_part {
2081 0 : Ok(index_part) => {
2082 0 : debug!("remote index part exists for timeline {timeline_id}");
2083 0 : index_part
2084 : }
2085 : Err(DownloadError::NotFound) => {
2086 0 : error!(%timeline_id, "index_part not found on remote");
2087 0 : return Err(TimelineArchivalError::NotFound);
2088 : }
2089 0 : Err(DownloadError::Cancelled) => return Err(TimelineArchivalError::Cancelled),
2090 0 : Err(e) => {
2091 0 : // Some (possibly ephemeral) error happened during index_part download.
2092 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
2093 0 : return Err(TimelineArchivalError::Other(
2094 0 : anyhow::Error::new(e).context("downloading index_part from remote storage"),
2095 0 : ));
2096 : }
2097 : };
2098 0 : let index_part = match index_part {
2099 0 : MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
2100 0 : MaybeDeletedIndexPart::Deleted(_index_part) => {
2101 0 : info!("timeline is deleted according to index_part.json");
2102 0 : return Err(TimelineArchivalError::NotFound);
2103 : }
2104 : };
2105 0 : let remote_metadata = index_part.metadata.clone();
2106 0 : let timeline_resources = self.build_timeline_resources(timeline_id);
2107 0 : self.load_remote_timeline(
2108 0 : timeline_id,
2109 0 : index_part,
2110 0 : remote_metadata,
2111 0 : timeline_resources,
2112 0 : LoadTimelineCause::Unoffload,
2113 0 : &ctx,
2114 0 : )
2115 0 : .await
2116 0 : .with_context(|| {
2117 0 : format!(
2118 0 : "failed to load remote timeline {} for tenant {}",
2119 0 : timeline_id, self.tenant_shard_id
2120 0 : )
2121 0 : })
2122 0 : .map_err(TimelineArchivalError::Other)?;
2123 :
2124 0 : let timeline = {
2125 0 : let timelines = self.timelines.lock().unwrap();
2126 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2127 0 : warn!("timeline not available directly after attach");
2128 : // This is not a panic because no locks are held between `load_remote_timeline`
2129 : // which puts the timeline into timelines, and our look into the timeline map.
2130 0 : return Err(TimelineArchivalError::Other(anyhow::anyhow!(
2131 0 : "timeline not available directly after attach"
2132 0 : )));
2133 : };
2134 0 : let mut offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2135 0 : match offloaded_timelines.remove(&timeline_id) {
2136 0 : Some(offloaded) => {
2137 0 : offloaded.delete_from_ancestor_with_timelines(&timelines);
2138 0 : }
2139 0 : None => warn!("timeline already removed from offloaded timelines"),
2140 : }
2141 :
2142 0 : self.initialize_gc_info(&timelines, &offloaded_timelines, Some(timeline_id));
2143 0 :
2144 0 : Arc::clone(timeline)
2145 0 : };
2146 0 :
2147 0 : // Upload new list of offloaded timelines to S3
2148 0 : self.store_tenant_manifest().await?;
2149 :
2150 : // Activate the timeline (if it makes sense)
2151 0 : if !(timeline.is_broken() || timeline.is_stopping()) {
2152 0 : let background_jobs_can_start = None;
2153 0 : timeline.activate(
2154 0 : self.clone(),
2155 0 : broker_client.clone(),
2156 0 : background_jobs_can_start,
2157 0 : &ctx,
2158 0 : );
2159 0 : }
2160 :
2161 0 : info!("timeline unoffloading complete");
2162 0 : Ok(timeline)
2163 0 : }
2164 :
2165 0 : pub(crate) async fn apply_timeline_archival_config(
2166 0 : self: &Arc<Self>,
2167 0 : timeline_id: TimelineId,
2168 0 : new_state: TimelineArchivalState,
2169 0 : broker_client: storage_broker::BrokerClientChannel,
2170 0 : ctx: RequestContext,
2171 0 : ) -> Result<(), TimelineArchivalError> {
2172 0 : info!("setting timeline archival config");
2173 : // First part: figure out what is needed to do, and do validation
2174 0 : let timeline_or_unarchive_offloaded = 'outer: {
2175 0 : let timelines = self.timelines.lock().unwrap();
2176 :
2177 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2178 0 : let offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2179 0 : let Some(offloaded) = offloaded_timelines.get(&timeline_id) else {
2180 0 : return Err(TimelineArchivalError::NotFound);
2181 : };
2182 0 : if new_state == TimelineArchivalState::Archived {
2183 : // It's offloaded already, so nothing to do
2184 0 : return Ok(());
2185 0 : }
2186 0 : if let Some(ancestor_timeline_id) = offloaded.ancestor_timeline_id {
2187 0 : Self::check_ancestor_of_to_be_unarchived_is_not_archived(
2188 0 : ancestor_timeline_id,
2189 0 : &timelines,
2190 0 : &offloaded_timelines,
2191 0 : )?;
2192 0 : }
2193 0 : break 'outer None;
2194 : };
2195 :
2196 : // Do some validation. We release the timelines lock below, so there is potential
2197 : // for race conditions: these checks are more present to prevent misunderstandings of
2198 : // the API's capabilities, instead of serving as the sole way to defend their invariants.
2199 0 : match new_state {
2200 : TimelineArchivalState::Unarchived => {
2201 0 : Self::check_to_be_unarchived_timeline_has_no_archived_parent(timeline)?
2202 : }
2203 : TimelineArchivalState::Archived => {
2204 0 : Self::check_to_be_archived_has_no_unarchived_children(timeline_id, &timelines)?
2205 : }
2206 : }
2207 0 : Some(Arc::clone(timeline))
2208 : };
2209 :
2210 : // Second part: unoffload timeline (if needed)
2211 0 : let timeline = if let Some(timeline) = timeline_or_unarchive_offloaded {
2212 0 : timeline
2213 : } else {
2214 : // Turn offloaded timeline into a non-offloaded one
2215 0 : self.unoffload_timeline(timeline_id, broker_client, ctx)
2216 0 : .await?
2217 : };
2218 :
2219 : // Third part: upload new timeline archival state and block until it is present in S3
2220 0 : let upload_needed = match timeline
2221 0 : .remote_client
2222 0 : .schedule_index_upload_for_timeline_archival_state(new_state)
2223 : {
2224 0 : Ok(upload_needed) => upload_needed,
2225 0 : Err(e) => {
2226 0 : if timeline.cancel.is_cancelled() {
2227 0 : return Err(TimelineArchivalError::Cancelled);
2228 : } else {
2229 0 : return Err(TimelineArchivalError::Other(e));
2230 : }
2231 : }
2232 : };
2233 :
2234 0 : if upload_needed {
2235 0 : info!("Uploading new state");
2236 : const MAX_WAIT: Duration = Duration::from_secs(10);
2237 0 : let Ok(v) =
2238 0 : tokio::time::timeout(MAX_WAIT, timeline.remote_client.wait_completion()).await
2239 : else {
2240 0 : tracing::warn!("reached timeout for waiting on upload queue");
2241 0 : return Err(TimelineArchivalError::Timeout);
2242 : };
2243 0 : v.map_err(|e| match e {
2244 0 : WaitCompletionError::NotInitialized(e) => {
2245 0 : TimelineArchivalError::Other(anyhow::anyhow!(e))
2246 : }
2247 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2248 0 : TimelineArchivalError::Cancelled
2249 : }
2250 0 : })?;
2251 0 : }
2252 0 : Ok(())
2253 0 : }
2254 :
2255 4 : pub fn get_offloaded_timeline(
2256 4 : &self,
2257 4 : timeline_id: TimelineId,
2258 4 : ) -> Result<Arc<OffloadedTimeline>, GetTimelineError> {
2259 4 : self.timelines_offloaded
2260 4 : .lock()
2261 4 : .unwrap()
2262 4 : .get(&timeline_id)
2263 4 : .map(Arc::clone)
2264 4 : .ok_or(GetTimelineError::NotFound {
2265 4 : tenant_id: self.tenant_shard_id,
2266 4 : timeline_id,
2267 4 : })
2268 4 : }
2269 :
2270 8 : pub(crate) fn tenant_shard_id(&self) -> TenantShardId {
2271 8 : self.tenant_shard_id
2272 8 : }
2273 :
2274 : /// Get Timeline handle for given Neon timeline ID.
2275 : /// This function is idempotent. It doesn't change internal state in any way.
2276 444 : pub fn get_timeline(
2277 444 : &self,
2278 444 : timeline_id: TimelineId,
2279 444 : active_only: bool,
2280 444 : ) -> Result<Arc<Timeline>, GetTimelineError> {
2281 444 : let timelines_accessor = self.timelines.lock().unwrap();
2282 444 : let timeline = timelines_accessor
2283 444 : .get(&timeline_id)
2284 444 : .ok_or(GetTimelineError::NotFound {
2285 444 : tenant_id: self.tenant_shard_id,
2286 444 : timeline_id,
2287 444 : })?;
2288 :
2289 440 : if active_only && !timeline.is_active() {
2290 0 : Err(GetTimelineError::NotActive {
2291 0 : tenant_id: self.tenant_shard_id,
2292 0 : timeline_id,
2293 0 : state: timeline.current_state(),
2294 0 : })
2295 : } else {
2296 440 : Ok(Arc::clone(timeline))
2297 : }
2298 444 : }
2299 :
2300 : /// Lists timelines the tenant contains.
2301 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2302 0 : pub fn list_timelines(&self) -> Vec<Arc<Timeline>> {
2303 0 : self.timelines
2304 0 : .lock()
2305 0 : .unwrap()
2306 0 : .values()
2307 0 : .map(Arc::clone)
2308 0 : .collect()
2309 0 : }
2310 :
2311 : /// Lists timelines the tenant manages, including offloaded ones.
2312 : ///
2313 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2314 0 : pub fn list_timelines_and_offloaded(
2315 0 : &self,
2316 0 : ) -> (Vec<Arc<Timeline>>, Vec<Arc<OffloadedTimeline>>) {
2317 0 : let timelines = self
2318 0 : .timelines
2319 0 : .lock()
2320 0 : .unwrap()
2321 0 : .values()
2322 0 : .map(Arc::clone)
2323 0 : .collect();
2324 0 : let offloaded = self
2325 0 : .timelines_offloaded
2326 0 : .lock()
2327 0 : .unwrap()
2328 0 : .values()
2329 0 : .map(Arc::clone)
2330 0 : .collect();
2331 0 : (timelines, offloaded)
2332 0 : }
2333 :
2334 0 : pub fn list_timeline_ids(&self) -> Vec<TimelineId> {
2335 0 : self.timelines.lock().unwrap().keys().cloned().collect()
2336 0 : }
2337 :
2338 : /// This is used by tests & import-from-basebackup.
2339 : ///
2340 : /// The returned [`UninitializedTimeline`] contains no data nor metadata and it is in
2341 : /// a state that will fail [`Tenant::load_remote_timeline`] because `disk_consistent_lsn=Lsn(0)`.
2342 : ///
2343 : /// The caller is responsible for getting the timeline into a state that will be accepted
2344 : /// by [`Tenant::load_remote_timeline`] / [`Tenant::attach`].
2345 : /// Then they may call [`UninitializedTimeline::finish_creation`] to add the timeline
2346 : /// to the [`Tenant::timelines`].
2347 : ///
2348 : /// Tests should use `Tenant::create_test_timeline` to set up the minimum required metadata keys.
2349 424 : pub(crate) async fn create_empty_timeline(
2350 424 : self: &Arc<Self>,
2351 424 : new_timeline_id: TimelineId,
2352 424 : initdb_lsn: Lsn,
2353 424 : pg_version: u32,
2354 424 : _ctx: &RequestContext,
2355 424 : ) -> anyhow::Result<UninitializedTimeline> {
2356 424 : anyhow::ensure!(
2357 424 : self.is_active(),
2358 0 : "Cannot create empty timelines on inactive tenant"
2359 : );
2360 :
2361 : // Protect against concurrent attempts to use this TimelineId
2362 424 : let create_guard = match self
2363 424 : .start_creating_timeline(new_timeline_id, CreateTimelineIdempotency::FailWithConflict)
2364 424 : .await?
2365 : {
2366 420 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2367 : StartCreatingTimelineResult::Idempotent(_) => {
2368 0 : unreachable!("FailWithConflict implies we get an error instead")
2369 : }
2370 : };
2371 :
2372 420 : let new_metadata = TimelineMetadata::new(
2373 420 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2374 420 : // make it valid, before calling finish_creation()
2375 420 : Lsn(0),
2376 420 : None,
2377 420 : None,
2378 420 : Lsn(0),
2379 420 : initdb_lsn,
2380 420 : initdb_lsn,
2381 420 : pg_version,
2382 420 : );
2383 420 : self.prepare_new_timeline(
2384 420 : new_timeline_id,
2385 420 : &new_metadata,
2386 420 : create_guard,
2387 420 : initdb_lsn,
2388 420 : None,
2389 420 : )
2390 420 : .await
2391 424 : }
2392 :
2393 : /// Helper for unit tests to create an empty timeline.
2394 : ///
2395 : /// The timeline is has state value `Active` but its background loops are not running.
2396 : // This makes the various functions which anyhow::ensure! for Active state work in tests.
2397 : // Our current tests don't need the background loops.
2398 : #[cfg(test)]
2399 404 : pub async fn create_test_timeline(
2400 404 : self: &Arc<Self>,
2401 404 : new_timeline_id: TimelineId,
2402 404 : initdb_lsn: Lsn,
2403 404 : pg_version: u32,
2404 404 : ctx: &RequestContext,
2405 404 : ) -> anyhow::Result<Arc<Timeline>> {
2406 404 : let uninit_tl = self
2407 404 : .create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2408 404 : .await?;
2409 404 : let tline = uninit_tl.raw_timeline().expect("we just created it");
2410 404 : assert_eq!(tline.get_last_record_lsn(), Lsn(0));
2411 :
2412 : // Setup minimum keys required for the timeline to be usable.
2413 404 : let mut modification = tline.begin_modification(initdb_lsn);
2414 404 : modification
2415 404 : .init_empty_test_timeline()
2416 404 : .context("init_empty_test_timeline")?;
2417 404 : modification
2418 404 : .commit(ctx)
2419 404 : .await
2420 404 : .context("commit init_empty_test_timeline modification")?;
2421 :
2422 : // Flush to disk so that uninit_tl's check for valid disk_consistent_lsn passes.
2423 404 : tline.maybe_spawn_flush_loop();
2424 404 : tline.freeze_and_flush().await.context("freeze_and_flush")?;
2425 :
2426 : // Make sure the freeze_and_flush reaches remote storage.
2427 404 : tline.remote_client.wait_completion().await.unwrap();
2428 :
2429 404 : let tl = uninit_tl.finish_creation()?;
2430 : // The non-test code would call tl.activate() here.
2431 404 : tl.set_state(TimelineState::Active);
2432 404 : Ok(tl)
2433 404 : }
2434 :
2435 : /// Helper for unit tests to create a timeline with some pre-loaded states.
2436 : #[cfg(test)]
2437 : #[allow(clippy::too_many_arguments)]
2438 76 : pub async fn create_test_timeline_with_layers(
2439 76 : self: &Arc<Self>,
2440 76 : new_timeline_id: TimelineId,
2441 76 : initdb_lsn: Lsn,
2442 76 : pg_version: u32,
2443 76 : ctx: &RequestContext,
2444 76 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
2445 76 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
2446 76 : end_lsn: Lsn,
2447 76 : ) -> anyhow::Result<Arc<Timeline>> {
2448 : use checks::check_valid_layermap;
2449 : use itertools::Itertools;
2450 :
2451 76 : let tline = self
2452 76 : .create_test_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2453 76 : .await?;
2454 76 : tline.force_advance_lsn(end_lsn);
2455 244 : for deltas in delta_layer_desc {
2456 168 : tline
2457 168 : .force_create_delta_layer(deltas, Some(initdb_lsn), ctx)
2458 168 : .await?;
2459 : }
2460 184 : for (lsn, images) in image_layer_desc {
2461 108 : tline
2462 108 : .force_create_image_layer(lsn, images, Some(initdb_lsn), ctx)
2463 108 : .await?;
2464 : }
2465 76 : let layer_names = tline
2466 76 : .layers
2467 76 : .read()
2468 76 : .await
2469 76 : .layer_map()
2470 76 : .unwrap()
2471 76 : .iter_historic_layers()
2472 352 : .map(|layer| layer.layer_name())
2473 76 : .collect_vec();
2474 76 : if let Some(err) = check_valid_layermap(&layer_names) {
2475 0 : bail!("invalid layermap: {err}");
2476 76 : }
2477 76 : Ok(tline)
2478 76 : }
2479 :
2480 : /// Create a new timeline.
2481 : ///
2482 : /// Returns the new timeline ID and reference to its Timeline object.
2483 : ///
2484 : /// If the caller specified the timeline ID to use (`new_timeline_id`), and timeline with
2485 : /// the same timeline ID already exists, returns CreateTimelineError::AlreadyExists.
2486 : #[allow(clippy::too_many_arguments)]
2487 0 : pub(crate) async fn create_timeline(
2488 0 : self: &Arc<Tenant>,
2489 0 : params: CreateTimelineParams,
2490 0 : broker_client: storage_broker::BrokerClientChannel,
2491 0 : ctx: &RequestContext,
2492 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
2493 0 : if !self.is_active() {
2494 0 : if matches!(self.current_state(), TenantState::Stopping { .. }) {
2495 0 : return Err(CreateTimelineError::ShuttingDown);
2496 : } else {
2497 0 : return Err(CreateTimelineError::Other(anyhow::anyhow!(
2498 0 : "Cannot create timelines on inactive tenant"
2499 0 : )));
2500 : }
2501 0 : }
2502 :
2503 0 : let _gate = self
2504 0 : .gate
2505 0 : .enter()
2506 0 : .map_err(|_| CreateTimelineError::ShuttingDown)?;
2507 :
2508 0 : let result: CreateTimelineResult = match params {
2509 : CreateTimelineParams::Bootstrap(CreateTimelineParamsBootstrap {
2510 0 : new_timeline_id,
2511 0 : existing_initdb_timeline_id,
2512 0 : pg_version,
2513 0 : }) => {
2514 0 : self.bootstrap_timeline(
2515 0 : new_timeline_id,
2516 0 : pg_version,
2517 0 : existing_initdb_timeline_id,
2518 0 : ctx,
2519 0 : )
2520 0 : .await?
2521 : }
2522 : CreateTimelineParams::Branch(CreateTimelineParamsBranch {
2523 0 : new_timeline_id,
2524 0 : ancestor_timeline_id,
2525 0 : mut ancestor_start_lsn,
2526 : }) => {
2527 0 : let ancestor_timeline = self
2528 0 : .get_timeline(ancestor_timeline_id, false)
2529 0 : .context("Cannot branch off the timeline that's not present in pageserver")?;
2530 :
2531 : // instead of waiting around, just deny the request because ancestor is not yet
2532 : // ready for other purposes either.
2533 0 : if !ancestor_timeline.is_active() {
2534 0 : return Err(CreateTimelineError::AncestorNotActive);
2535 0 : }
2536 0 :
2537 0 : if ancestor_timeline.is_archived() == Some(true) {
2538 0 : info!("tried to branch archived timeline");
2539 0 : return Err(CreateTimelineError::AncestorArchived);
2540 0 : }
2541 :
2542 0 : if let Some(lsn) = ancestor_start_lsn.as_mut() {
2543 0 : *lsn = lsn.align();
2544 0 :
2545 0 : let ancestor_ancestor_lsn = ancestor_timeline.get_ancestor_lsn();
2546 0 : if ancestor_ancestor_lsn > *lsn {
2547 : // can we safely just branch from the ancestor instead?
2548 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
2549 0 : "invalid start lsn {} for ancestor timeline {}: less than timeline ancestor lsn {}",
2550 0 : lsn,
2551 0 : ancestor_timeline_id,
2552 0 : ancestor_ancestor_lsn,
2553 0 : )));
2554 0 : }
2555 0 :
2556 0 : // Wait for the WAL to arrive and be processed on the parent branch up
2557 0 : // to the requested branch point. The repository code itself doesn't
2558 0 : // require it, but if we start to receive WAL on the new timeline,
2559 0 : // decoding the new WAL might need to look up previous pages, relation
2560 0 : // sizes etc. and that would get confused if the previous page versions
2561 0 : // are not in the repository yet.
2562 0 : ancestor_timeline
2563 0 : .wait_lsn(
2564 0 : *lsn,
2565 0 : timeline::WaitLsnWaiter::Tenant,
2566 0 : timeline::WaitLsnTimeout::Default,
2567 0 : ctx,
2568 0 : )
2569 0 : .await
2570 0 : .map_err(|e| match e {
2571 0 : e @ (WaitLsnError::Timeout(_) | WaitLsnError::BadState { .. }) => {
2572 0 : CreateTimelineError::AncestorLsn(anyhow::anyhow!(e))
2573 : }
2574 0 : WaitLsnError::Shutdown => CreateTimelineError::ShuttingDown,
2575 0 : })?;
2576 0 : }
2577 :
2578 0 : self.branch_timeline(&ancestor_timeline, new_timeline_id, ancestor_start_lsn, ctx)
2579 0 : .await?
2580 : }
2581 0 : CreateTimelineParams::ImportPgdata(params) => {
2582 0 : self.create_timeline_import_pgdata(
2583 0 : params,
2584 0 : ActivateTimelineArgs::Yes {
2585 0 : broker_client: broker_client.clone(),
2586 0 : },
2587 0 : ctx,
2588 0 : )
2589 0 : .await?
2590 : }
2591 : };
2592 :
2593 : // At this point we have dropped our guard on [`Self::timelines_creating`], and
2594 : // the timeline is visible in [`Self::timelines`], but it is _not_ durable yet. We must
2595 : // not send a success to the caller until it is. The same applies to idempotent retries.
2596 : //
2597 : // TODO: the timeline is already visible in [`Self::timelines`]; a caller could incorrectly
2598 : // assume that, because they can see the timeline via API, that the creation is done and
2599 : // that it is durable. Ideally, we would keep the timeline hidden (in [`Self::timelines_creating`])
2600 : // until it is durable, e.g., by extending the time we hold the creation guard. This also
2601 : // interacts with UninitializedTimeline and is generally a bit tricky.
2602 : //
2603 : // To re-emphasize: the only correct way to create a timeline is to repeat calling the
2604 : // creation API until it returns success. Only then is durability guaranteed.
2605 0 : info!(creation_result=%result.discriminant(), "waiting for timeline to be durable");
2606 0 : result
2607 0 : .timeline()
2608 0 : .remote_client
2609 0 : .wait_completion()
2610 0 : .await
2611 0 : .map_err(|e| match e {
2612 : WaitCompletionError::NotInitialized(
2613 0 : e, // If the queue is already stopped, it's a shutdown error.
2614 0 : ) if e.is_stopping() => CreateTimelineError::ShuttingDown,
2615 : WaitCompletionError::NotInitialized(_) => {
2616 : // This is a bug: we should never try to wait for uploads before initializing the timeline
2617 0 : debug_assert!(false);
2618 0 : CreateTimelineError::Other(anyhow::anyhow!("timeline not initialized"))
2619 : }
2620 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2621 0 : CreateTimelineError::ShuttingDown
2622 : }
2623 0 : })?;
2624 :
2625 : // The creating task is responsible for activating the timeline.
2626 : // We do this after `wait_completion()` so that we don't spin up tasks that start
2627 : // doing stuff before the IndexPart is durable in S3, which is done by the previous section.
2628 0 : let activated_timeline = match result {
2629 0 : CreateTimelineResult::Created(timeline) => {
2630 0 : timeline.activate(self.clone(), broker_client, None, ctx);
2631 0 : timeline
2632 : }
2633 0 : CreateTimelineResult::Idempotent(timeline) => {
2634 0 : info!(
2635 0 : "request was deemed idempotent, activation will be done by the creating task"
2636 : );
2637 0 : timeline
2638 : }
2639 0 : CreateTimelineResult::ImportSpawned(timeline) => {
2640 0 : info!("import task spawned, timeline will become visible and activated once the import is done");
2641 0 : timeline
2642 : }
2643 : };
2644 :
2645 0 : Ok(activated_timeline)
2646 0 : }
2647 :
2648 : /// The returned [`Arc<Timeline>`] is NOT in the [`Tenant::timelines`] map until the import
2649 : /// completes in the background. A DIFFERENT [`Arc<Timeline>`] will be inserted into the
2650 : /// [`Tenant::timelines`] map when the import completes.
2651 : /// We only return an [`Arc<Timeline>`] here so the API handler can create a [`pageserver_api::models::TimelineInfo`]
2652 : /// for the response.
2653 0 : async fn create_timeline_import_pgdata(
2654 0 : self: &Arc<Tenant>,
2655 0 : params: CreateTimelineParamsImportPgdata,
2656 0 : activate: ActivateTimelineArgs,
2657 0 : ctx: &RequestContext,
2658 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
2659 0 : let CreateTimelineParamsImportPgdata {
2660 0 : new_timeline_id,
2661 0 : location,
2662 0 : idempotency_key,
2663 0 : } = params;
2664 0 :
2665 0 : let started_at = chrono::Utc::now().naive_utc();
2666 :
2667 : //
2668 : // There's probably a simpler way to upload an index part, but, remote_timeline_client
2669 : // is the canonical way we do it.
2670 : // - create an empty timeline in-memory
2671 : // - use its remote_timeline_client to do the upload
2672 : // - dispose of the uninit timeline
2673 : // - keep the creation guard alive
2674 :
2675 0 : let timeline_create_guard = match self
2676 0 : .start_creating_timeline(
2677 0 : new_timeline_id,
2678 0 : CreateTimelineIdempotency::ImportPgdata(CreatingTimelineIdempotencyImportPgdata {
2679 0 : idempotency_key: idempotency_key.clone(),
2680 0 : }),
2681 0 : )
2682 0 : .await?
2683 : {
2684 0 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2685 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
2686 0 : return Ok(CreateTimelineResult::Idempotent(timeline))
2687 : }
2688 : };
2689 :
2690 0 : let mut uninit_timeline = {
2691 0 : let this = &self;
2692 0 : let initdb_lsn = Lsn(0);
2693 0 : let _ctx = ctx;
2694 0 : async move {
2695 0 : let new_metadata = TimelineMetadata::new(
2696 0 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2697 0 : // make it valid, before calling finish_creation()
2698 0 : Lsn(0),
2699 0 : None,
2700 0 : None,
2701 0 : Lsn(0),
2702 0 : initdb_lsn,
2703 0 : initdb_lsn,
2704 0 : 15,
2705 0 : );
2706 0 : this.prepare_new_timeline(
2707 0 : new_timeline_id,
2708 0 : &new_metadata,
2709 0 : timeline_create_guard,
2710 0 : initdb_lsn,
2711 0 : None,
2712 0 : )
2713 0 : .await
2714 0 : }
2715 0 : }
2716 0 : .await?;
2717 :
2718 0 : let in_progress = import_pgdata::index_part_format::InProgress {
2719 0 : idempotency_key,
2720 0 : location,
2721 0 : started_at,
2722 0 : };
2723 0 : let index_part = import_pgdata::index_part_format::Root::V1(
2724 0 : import_pgdata::index_part_format::V1::InProgress(in_progress),
2725 0 : );
2726 0 : uninit_timeline
2727 0 : .raw_timeline()
2728 0 : .unwrap()
2729 0 : .remote_client
2730 0 : .schedule_index_upload_for_import_pgdata_state_update(Some(index_part.clone()))?;
2731 :
2732 : // wait_completion happens in caller
2733 :
2734 0 : let (timeline, timeline_create_guard) = uninit_timeline.finish_creation_myself();
2735 0 :
2736 0 : tokio::spawn(self.clone().create_timeline_import_pgdata_task(
2737 0 : timeline.clone(),
2738 0 : index_part,
2739 0 : activate,
2740 0 : timeline_create_guard,
2741 0 : ));
2742 0 :
2743 0 : // NB: the timeline doesn't exist in self.timelines at this point
2744 0 : Ok(CreateTimelineResult::ImportSpawned(timeline))
2745 0 : }
2746 :
2747 : #[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))]
2748 : async fn create_timeline_import_pgdata_task(
2749 : self: Arc<Tenant>,
2750 : timeline: Arc<Timeline>,
2751 : index_part: import_pgdata::index_part_format::Root,
2752 : activate: ActivateTimelineArgs,
2753 : timeline_create_guard: TimelineCreateGuard,
2754 : ) {
2755 : debug_assert_current_span_has_tenant_and_timeline_id();
2756 : info!("starting");
2757 : scopeguard::defer! {info!("exiting")};
2758 :
2759 : let res = self
2760 : .create_timeline_import_pgdata_task_impl(
2761 : timeline,
2762 : index_part,
2763 : activate,
2764 : timeline_create_guard,
2765 : )
2766 : .await;
2767 : if let Err(err) = &res {
2768 : error!(?err, "task failed");
2769 : // TODO sleep & retry, sensitive to tenant shutdown
2770 : // TODO: allow timeline deletion requests => should cancel the task
2771 : }
2772 : }
2773 :
2774 0 : async fn create_timeline_import_pgdata_task_impl(
2775 0 : self: Arc<Tenant>,
2776 0 : timeline: Arc<Timeline>,
2777 0 : index_part: import_pgdata::index_part_format::Root,
2778 0 : activate: ActivateTimelineArgs,
2779 0 : timeline_create_guard: TimelineCreateGuard,
2780 0 : ) -> Result<(), anyhow::Error> {
2781 0 : let ctx = RequestContext::new(TaskKind::ImportPgdata, DownloadBehavior::Warn);
2782 0 :
2783 0 : info!("importing pgdata");
2784 0 : import_pgdata::doit(&timeline, index_part, &ctx, self.cancel.clone())
2785 0 : .await
2786 0 : .context("import")?;
2787 0 : info!("import done");
2788 :
2789 : //
2790 : // Reload timeline from remote.
2791 : // This proves that the remote state is attachable, and it reuses the code.
2792 : //
2793 : // TODO: think about whether this is safe to do with concurrent Tenant::shutdown.
2794 : // timeline_create_guard hols the tenant gate open, so, shutdown cannot _complete_ until we exit.
2795 : // But our activate() call might launch new background tasks after Tenant::shutdown
2796 : // already went past shutting down the Tenant::timelines, which this timeline here is no part of.
2797 : // I think the same problem exists with the bootstrap & branch mgmt API tasks (tenant shutting
2798 : // down while bootstrapping/branching + activating), but, the race condition is much more likely
2799 : // to manifest because of the long runtime of this import task.
2800 :
2801 : // in theory this shouldn't even .await anything except for coop yield
2802 0 : info!("shutting down timeline");
2803 0 : timeline.shutdown(ShutdownMode::Hard).await;
2804 0 : info!("timeline shut down, reloading from remote");
2805 : // TODO: we can't do the following check because create_timeline_import_pgdata must return an Arc<Timeline>
2806 : // let Some(timeline) = Arc::into_inner(timeline) else {
2807 : // anyhow::bail!("implementation error: timeline that we shut down was still referenced from somewhere");
2808 : // };
2809 0 : let timeline_id = timeline.timeline_id;
2810 0 :
2811 0 : // load from object storage like Tenant::attach does
2812 0 : let resources = self.build_timeline_resources(timeline_id);
2813 0 : let index_part = resources
2814 0 : .remote_client
2815 0 : .download_index_file(&self.cancel)
2816 0 : .await?;
2817 0 : let index_part = match index_part {
2818 : MaybeDeletedIndexPart::Deleted(_) => {
2819 : // likely concurrent delete call, cplane should prevent this
2820 0 : anyhow::bail!("index part says deleted but we are not done creating yet, this should not happen but")
2821 : }
2822 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
2823 0 : };
2824 0 : let metadata = index_part.metadata.clone();
2825 0 : self
2826 0 : .load_remote_timeline(timeline_id, index_part, metadata, resources, LoadTimelineCause::ImportPgdata{
2827 0 : create_guard: timeline_create_guard, activate, }, &ctx)
2828 0 : .await?
2829 0 : .ready_to_activate()
2830 0 : .context("implementation error: reloaded timeline still needs import after import reported success")?;
2831 :
2832 0 : anyhow::Ok(())
2833 0 : }
2834 :
2835 0 : pub(crate) async fn delete_timeline(
2836 0 : self: Arc<Self>,
2837 0 : timeline_id: TimelineId,
2838 0 : ) -> Result<(), DeleteTimelineError> {
2839 0 : DeleteTimelineFlow::run(&self, timeline_id).await?;
2840 :
2841 0 : Ok(())
2842 0 : }
2843 :
2844 : /// perform one garbage collection iteration, removing old data files from disk.
2845 : /// this function is periodically called by gc task.
2846 : /// also it can be explicitly requested through page server api 'do_gc' command.
2847 : ///
2848 : /// `target_timeline_id` specifies the timeline to GC, or None for all.
2849 : ///
2850 : /// The `horizon` an `pitr` parameters determine how much WAL history needs to be retained.
2851 : /// Also known as the retention period, or the GC cutoff point. `horizon` specifies
2852 : /// the amount of history, as LSN difference from current latest LSN on each timeline.
2853 : /// `pitr` specifies the same as a time difference from the current time. The effective
2854 : /// GC cutoff point is determined conservatively by either `horizon` and `pitr`, whichever
2855 : /// requires more history to be retained.
2856 : //
2857 1508 : pub(crate) async fn gc_iteration(
2858 1508 : &self,
2859 1508 : target_timeline_id: Option<TimelineId>,
2860 1508 : horizon: u64,
2861 1508 : pitr: Duration,
2862 1508 : cancel: &CancellationToken,
2863 1508 : ctx: &RequestContext,
2864 1508 : ) -> Result<GcResult, GcError> {
2865 1508 : // Don't start doing work during shutdown
2866 1508 : if let TenantState::Stopping { .. } = self.current_state() {
2867 0 : return Ok(GcResult::default());
2868 1508 : }
2869 1508 :
2870 1508 : // there is a global allowed_error for this
2871 1508 : if !self.is_active() {
2872 0 : return Err(GcError::NotActive);
2873 1508 : }
2874 1508 :
2875 1508 : {
2876 1508 : let conf = self.tenant_conf.load();
2877 1508 :
2878 1508 : // If we may not delete layers, then simply skip GC. Even though a tenant
2879 1508 : // in AttachedMulti state could do GC and just enqueue the blocked deletions,
2880 1508 : // the only advantage to doing it is to perhaps shrink the LayerMap metadata
2881 1508 : // a bit sooner than we would achieve by waiting for AttachedSingle status.
2882 1508 : if !conf.location.may_delete_layers_hint() {
2883 0 : info!("Skipping GC in location state {:?}", conf.location);
2884 0 : return Ok(GcResult::default());
2885 1508 : }
2886 1508 :
2887 1508 : if conf.is_gc_blocked_by_lsn_lease_deadline() {
2888 1500 : info!("Skipping GC because lsn lease deadline is not reached");
2889 1500 : return Ok(GcResult::default());
2890 8 : }
2891 : }
2892 :
2893 8 : let _guard = match self.gc_block.start().await {
2894 8 : Ok(guard) => guard,
2895 0 : Err(reasons) => {
2896 0 : info!("Skipping GC: {reasons}");
2897 0 : return Ok(GcResult::default());
2898 : }
2899 : };
2900 :
2901 8 : self.gc_iteration_internal(target_timeline_id, horizon, pitr, cancel, ctx)
2902 8 : .await
2903 1508 : }
2904 :
2905 : /// Perform one compaction iteration.
2906 : /// This function is periodically called by compactor task.
2907 : /// Also it can be explicitly requested per timeline through page server
2908 : /// api's 'compact' command.
2909 : ///
2910 : /// Returns whether we have pending compaction task.
2911 0 : async fn compaction_iteration(
2912 0 : self: &Arc<Self>,
2913 0 : cancel: &CancellationToken,
2914 0 : ctx: &RequestContext,
2915 0 : ) -> Result<bool, timeline::CompactionError> {
2916 0 : // Don't start doing work during shutdown, or when broken, we do not need those in the logs
2917 0 : if !self.is_active() {
2918 0 : return Ok(false);
2919 0 : }
2920 0 :
2921 0 : {
2922 0 : let conf = self.tenant_conf.load();
2923 0 :
2924 0 : // Note that compaction usually requires deletions, but we don't respect
2925 0 : // may_delete_layers_hint here: that is because tenants in AttachedMulti
2926 0 : // should proceed with compaction even if they can't do deletion, to avoid
2927 0 : // accumulating dangerously deep stacks of L0 layers. Deletions will be
2928 0 : // enqueued inside RemoteTimelineClient, and executed layer if/when we transition
2929 0 : // to AttachedSingle state.
2930 0 : if !conf.location.may_upload_layers_hint() {
2931 0 : info!("Skipping compaction in location state {:?}", conf.location);
2932 0 : return Ok(false);
2933 0 : }
2934 0 : }
2935 0 :
2936 0 : // Scan through the hashmap and collect a list of all the timelines,
2937 0 : // while holding the lock. Then drop the lock and actually perform the
2938 0 : // compactions. We don't want to block everything else while the
2939 0 : // compaction runs.
2940 0 : let timelines_to_compact_or_offload;
2941 0 : {
2942 0 : let timelines = self.timelines.lock().unwrap();
2943 0 : timelines_to_compact_or_offload = timelines
2944 0 : .iter()
2945 0 : .filter_map(|(timeline_id, timeline)| {
2946 0 : let (is_active, (can_offload, _)) =
2947 0 : (timeline.is_active(), timeline.can_offload());
2948 0 : let has_no_unoffloaded_children = {
2949 0 : !timelines
2950 0 : .iter()
2951 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(*timeline_id))
2952 : };
2953 0 : let config_allows_offload = self.conf.timeline_offloading
2954 0 : || self
2955 0 : .tenant_conf
2956 0 : .load()
2957 0 : .tenant_conf
2958 0 : .timeline_offloading
2959 0 : .unwrap_or_default();
2960 0 : let can_offload =
2961 0 : can_offload && has_no_unoffloaded_children && config_allows_offload;
2962 0 : if (is_active, can_offload) == (false, false) {
2963 0 : None
2964 : } else {
2965 0 : Some((*timeline_id, timeline.clone(), (is_active, can_offload)))
2966 : }
2967 0 : })
2968 0 : .collect::<Vec<_>>();
2969 0 : drop(timelines);
2970 0 : }
2971 0 :
2972 0 : // Before doing any I/O work, check our circuit breaker
2973 0 : if self.compaction_circuit_breaker.lock().unwrap().is_broken() {
2974 0 : info!("Skipping compaction due to previous failures");
2975 0 : return Ok(false);
2976 0 : }
2977 0 :
2978 0 : let mut has_pending_task = false;
2979 :
2980 0 : for (timeline_id, timeline, (can_compact, can_offload)) in &timelines_to_compact_or_offload
2981 : {
2982 : // pending_task_left == None: cannot compact, maybe still pending tasks
2983 : // pending_task_left == Some(true): compaction task left
2984 : // pending_task_left == Some(false): no compaction task left
2985 0 : let pending_task_left = if *can_compact {
2986 0 : let has_pending_l0_compaction_task = timeline
2987 0 : .compact(cancel, EnumSet::empty(), ctx)
2988 0 : .instrument(info_span!("compact_timeline", %timeline_id))
2989 0 : .await
2990 0 : .inspect_err(|e| match e {
2991 0 : timeline::CompactionError::ShuttingDown => (),
2992 0 : timeline::CompactionError::Offload(_) => {
2993 0 : // Failures to offload timelines do not trip the circuit breaker, because
2994 0 : // they do not do lots of writes the way compaction itself does: it is cheap
2995 0 : // to retry, and it would be bad to stop all compaction because of an issue with offloading.
2996 0 : }
2997 0 : timeline::CompactionError::Other(e) => {
2998 0 : self.compaction_circuit_breaker
2999 0 : .lock()
3000 0 : .unwrap()
3001 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, e);
3002 0 : }
3003 0 : })?;
3004 0 : if has_pending_l0_compaction_task {
3005 0 : Some(true)
3006 : } else {
3007 0 : let queue = {
3008 0 : let guard = self.scheduled_compaction_tasks.lock().unwrap();
3009 0 : guard.get(timeline_id).cloned()
3010 : };
3011 0 : if let Some(queue) = queue {
3012 0 : let has_pending_tasks = queue
3013 0 : .iteration(cancel, ctx, &self.gc_block, timeline)
3014 0 : .await?;
3015 0 : Some(has_pending_tasks)
3016 : } else {
3017 0 : Some(false)
3018 : }
3019 : }
3020 : } else {
3021 0 : None
3022 : };
3023 0 : has_pending_task |= pending_task_left.unwrap_or(false);
3024 0 : if pending_task_left == Some(false) && *can_offload {
3025 0 : pausable_failpoint!("before-timeline-auto-offload");
3026 0 : match offload_timeline(self, timeline)
3027 0 : .instrument(info_span!("offload_timeline", %timeline_id))
3028 0 : .await
3029 : {
3030 : Err(OffloadError::NotArchived) => {
3031 : // Ignore this, we likely raced with unarchival
3032 0 : Ok(())
3033 : }
3034 0 : other => other,
3035 0 : }?;
3036 0 : }
3037 : }
3038 :
3039 0 : self.compaction_circuit_breaker
3040 0 : .lock()
3041 0 : .unwrap()
3042 0 : .success(&CIRCUIT_BREAKERS_UNBROKEN);
3043 0 :
3044 0 : Ok(has_pending_task)
3045 0 : }
3046 :
3047 : /// Cancel scheduled compaction tasks
3048 0 : pub(crate) fn cancel_scheduled_compaction(&self, timeline_id: TimelineId) {
3049 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3050 0 : if let Some(q) = guard.get_mut(&timeline_id) {
3051 0 : q.cancel_scheduled();
3052 0 : }
3053 0 : }
3054 :
3055 0 : pub(crate) fn get_scheduled_compaction_tasks(
3056 0 : &self,
3057 0 : timeline_id: TimelineId,
3058 0 : ) -> Vec<CompactInfoResponse> {
3059 0 : let res = {
3060 0 : let guard = self.scheduled_compaction_tasks.lock().unwrap();
3061 0 : guard.get(&timeline_id).map(|q| q.remaining_jobs())
3062 : };
3063 0 : let Some((running, remaining)) = res else {
3064 0 : return Vec::new();
3065 : };
3066 0 : let mut result = Vec::new();
3067 0 : if let Some((id, running)) = running {
3068 0 : result.extend(running.into_compact_info_resp(id, true));
3069 0 : }
3070 0 : for (id, job) in remaining {
3071 0 : result.extend(job.into_compact_info_resp(id, false));
3072 0 : }
3073 0 : result
3074 0 : }
3075 :
3076 : /// Schedule a compaction task for a timeline.
3077 0 : pub(crate) async fn schedule_compaction(
3078 0 : &self,
3079 0 : timeline_id: TimelineId,
3080 0 : options: CompactOptions,
3081 0 : ) -> anyhow::Result<tokio::sync::oneshot::Receiver<()>> {
3082 0 : let (tx, rx) = tokio::sync::oneshot::channel();
3083 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3084 0 : let q = guard
3085 0 : .entry(timeline_id)
3086 0 : .or_insert_with(|| Arc::new(GcCompactionQueue::new()));
3087 0 : q.schedule_manual_compaction(options, Some(tx));
3088 0 : Ok(rx)
3089 0 : }
3090 :
3091 : // Call through to all timelines to freeze ephemeral layers if needed. Usually
3092 : // this happens during ingest: this background housekeeping is for freezing layers
3093 : // that are open but haven't been written to for some time.
3094 0 : async fn ingest_housekeeping(&self) {
3095 0 : // Scan through the hashmap and collect a list of all the timelines,
3096 0 : // while holding the lock. Then drop the lock and actually perform the
3097 0 : // compactions. We don't want to block everything else while the
3098 0 : // compaction runs.
3099 0 : let timelines = {
3100 0 : self.timelines
3101 0 : .lock()
3102 0 : .unwrap()
3103 0 : .values()
3104 0 : .filter_map(|timeline| {
3105 0 : if timeline.is_active() {
3106 0 : Some(timeline.clone())
3107 : } else {
3108 0 : None
3109 : }
3110 0 : })
3111 0 : .collect::<Vec<_>>()
3112 : };
3113 :
3114 0 : for timeline in &timelines {
3115 0 : timeline.maybe_freeze_ephemeral_layer().await;
3116 : }
3117 0 : }
3118 :
3119 0 : pub fn timeline_has_no_attached_children(&self, timeline_id: TimelineId) -> bool {
3120 0 : let timelines = self.timelines.lock().unwrap();
3121 0 : !timelines
3122 0 : .iter()
3123 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(timeline_id))
3124 0 : }
3125 :
3126 3464 : pub fn current_state(&self) -> TenantState {
3127 3464 : self.state.borrow().clone()
3128 3464 : }
3129 :
3130 1940 : pub fn is_active(&self) -> bool {
3131 1940 : self.current_state() == TenantState::Active
3132 1940 : }
3133 :
3134 0 : pub fn generation(&self) -> Generation {
3135 0 : self.generation
3136 0 : }
3137 :
3138 0 : pub(crate) fn wal_redo_manager_status(&self) -> Option<WalRedoManagerStatus> {
3139 0 : self.walredo_mgr.as_ref().and_then(|mgr| mgr.status())
3140 0 : }
3141 :
3142 : /// Changes tenant status to active, unless shutdown was already requested.
3143 : ///
3144 : /// `background_jobs_can_start` is an optional barrier set to a value during pageserver startup
3145 : /// to delay background jobs. Background jobs can be started right away when None is given.
3146 0 : fn activate(
3147 0 : self: &Arc<Self>,
3148 0 : broker_client: BrokerClientChannel,
3149 0 : background_jobs_can_start: Option<&completion::Barrier>,
3150 0 : ctx: &RequestContext,
3151 0 : ) {
3152 0 : span::debug_assert_current_span_has_tenant_id();
3153 0 :
3154 0 : let mut activating = false;
3155 0 : self.state.send_modify(|current_state| {
3156 : use pageserver_api::models::ActivatingFrom;
3157 0 : match &*current_state {
3158 : TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
3159 0 : panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {state:?}", state = current_state);
3160 : }
3161 0 : TenantState::Attaching => {
3162 0 : *current_state = TenantState::Activating(ActivatingFrom::Attaching);
3163 0 : }
3164 0 : }
3165 0 : debug!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), "Activating tenant");
3166 0 : activating = true;
3167 0 : // Continue outside the closure. We need to grab timelines.lock()
3168 0 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
3169 0 : });
3170 0 :
3171 0 : if activating {
3172 0 : let timelines_accessor = self.timelines.lock().unwrap();
3173 0 : let timelines_offloaded_accessor = self.timelines_offloaded.lock().unwrap();
3174 0 : let timelines_to_activate = timelines_accessor
3175 0 : .values()
3176 0 : .filter(|timeline| !(timeline.is_broken() || timeline.is_stopping()));
3177 0 :
3178 0 : // Before activation, populate each Timeline's GcInfo with information about its children
3179 0 : self.initialize_gc_info(&timelines_accessor, &timelines_offloaded_accessor, None);
3180 0 :
3181 0 : // Spawn gc and compaction loops. The loops will shut themselves
3182 0 : // down when they notice that the tenant is inactive.
3183 0 : tasks::start_background_loops(self, background_jobs_can_start);
3184 0 :
3185 0 : let mut activated_timelines = 0;
3186 :
3187 0 : for timeline in timelines_to_activate {
3188 0 : timeline.activate(
3189 0 : self.clone(),
3190 0 : broker_client.clone(),
3191 0 : background_jobs_can_start,
3192 0 : ctx,
3193 0 : );
3194 0 : activated_timelines += 1;
3195 0 : }
3196 :
3197 0 : self.state.send_modify(move |current_state| {
3198 0 : assert!(
3199 0 : matches!(current_state, TenantState::Activating(_)),
3200 0 : "set_stopping and set_broken wait for us to leave Activating state",
3201 : );
3202 0 : *current_state = TenantState::Active;
3203 0 :
3204 0 : let elapsed = self.constructed_at.elapsed();
3205 0 : let total_timelines = timelines_accessor.len();
3206 0 :
3207 0 : // log a lot of stuff, because some tenants sometimes suffer from user-visible
3208 0 : // times to activate. see https://github.com/neondatabase/neon/issues/4025
3209 0 : info!(
3210 0 : since_creation_millis = elapsed.as_millis(),
3211 0 : tenant_id = %self.tenant_shard_id.tenant_id,
3212 0 : shard_id = %self.tenant_shard_id.shard_slug(),
3213 0 : activated_timelines,
3214 0 : total_timelines,
3215 0 : post_state = <&'static str>::from(&*current_state),
3216 0 : "activation attempt finished"
3217 : );
3218 :
3219 0 : TENANT.activation.observe(elapsed.as_secs_f64());
3220 0 : });
3221 0 : }
3222 0 : }
3223 :
3224 : /// Shutdown the tenant and join all of the spawned tasks.
3225 : ///
3226 : /// The method caters for all use-cases:
3227 : /// - pageserver shutdown (freeze_and_flush == true)
3228 : /// - detach + ignore (freeze_and_flush == false)
3229 : ///
3230 : /// This will attempt to shutdown even if tenant is broken.
3231 : ///
3232 : /// `shutdown_progress` is a [`completion::Barrier`] for the shutdown initiated by this call.
3233 : /// If the tenant is already shutting down, we return a clone of the first shutdown call's
3234 : /// `Barrier` as an `Err`. This not-first caller can use the returned barrier to join with
3235 : /// the ongoing shutdown.
3236 12 : async fn shutdown(
3237 12 : &self,
3238 12 : shutdown_progress: completion::Barrier,
3239 12 : shutdown_mode: timeline::ShutdownMode,
3240 12 : ) -> Result<(), completion::Barrier> {
3241 12 : span::debug_assert_current_span_has_tenant_id();
3242 :
3243 : // Set tenant (and its timlines) to Stoppping state.
3244 : //
3245 : // Since we can only transition into Stopping state after activation is complete,
3246 : // run it in a JoinSet so all tenants have a chance to stop before we get SIGKILLed.
3247 : //
3248 : // Transitioning tenants to Stopping state has a couple of non-obvious side effects:
3249 : // 1. Lock out any new requests to the tenants.
3250 : // 2. Signal cancellation to WAL receivers (we wait on it below).
3251 : // 3. Signal cancellation for other tenant background loops.
3252 : // 4. ???
3253 : //
3254 : // The waiting for the cancellation is not done uniformly.
3255 : // We certainly wait for WAL receivers to shut down.
3256 : // That is necessary so that no new data comes in before the freeze_and_flush.
3257 : // But the tenant background loops are joined-on in our caller.
3258 : // It's mesed up.
3259 : // we just ignore the failure to stop
3260 :
3261 : // If we're still attaching, fire the cancellation token early to drop out: this
3262 : // will prevent us flushing, but ensures timely shutdown if some I/O during attach
3263 : // is very slow.
3264 12 : let shutdown_mode = if matches!(self.current_state(), TenantState::Attaching) {
3265 0 : self.cancel.cancel();
3266 0 :
3267 0 : // Having fired our cancellation token, do not try and flush timelines: their cancellation tokens
3268 0 : // are children of ours, so their flush loops will have shut down already
3269 0 : timeline::ShutdownMode::Hard
3270 : } else {
3271 12 : shutdown_mode
3272 : };
3273 :
3274 12 : match self.set_stopping(shutdown_progress, false, false).await {
3275 12 : Ok(()) => {}
3276 0 : Err(SetStoppingError::Broken) => {
3277 0 : // assume that this is acceptable
3278 0 : }
3279 0 : Err(SetStoppingError::AlreadyStopping(other)) => {
3280 0 : // give caller the option to wait for this this shutdown
3281 0 : info!("Tenant::shutdown: AlreadyStopping");
3282 0 : return Err(other);
3283 : }
3284 : };
3285 :
3286 12 : let mut js = tokio::task::JoinSet::new();
3287 12 : {
3288 12 : let timelines = self.timelines.lock().unwrap();
3289 12 : timelines.values().for_each(|timeline| {
3290 12 : let timeline = Arc::clone(timeline);
3291 12 : let timeline_id = timeline.timeline_id;
3292 12 : let span = tracing::info_span!("timeline_shutdown", %timeline_id, ?shutdown_mode);
3293 12 : js.spawn(async move { timeline.shutdown(shutdown_mode).instrument(span).await });
3294 12 : });
3295 12 : }
3296 12 : {
3297 12 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
3298 12 : timelines_offloaded.values().for_each(|timeline| {
3299 0 : timeline.defuse_for_tenant_drop();
3300 12 : });
3301 12 : }
3302 12 : // test_long_timeline_create_then_tenant_delete is leaning on this message
3303 12 : tracing::info!("Waiting for timelines...");
3304 24 : while let Some(res) = js.join_next().await {
3305 0 : match res {
3306 12 : Ok(()) => {}
3307 0 : Err(je) if je.is_cancelled() => unreachable!("no cancelling used"),
3308 0 : Err(je) if je.is_panic() => { /* logged already */ }
3309 0 : Err(je) => warn!("unexpected JoinError: {je:?}"),
3310 : }
3311 : }
3312 :
3313 12 : if let ShutdownMode::Reload = shutdown_mode {
3314 0 : tracing::info!("Flushing deletion queue");
3315 0 : if let Err(e) = self.deletion_queue_client.flush().await {
3316 0 : match e {
3317 0 : DeletionQueueError::ShuttingDown => {
3318 0 : // This is the only error we expect for now. In the future, if more error
3319 0 : // variants are added, we should handle them here.
3320 0 : }
3321 : }
3322 0 : }
3323 12 : }
3324 :
3325 : // We cancel the Tenant's cancellation token _after_ the timelines have all shut down. This permits
3326 : // them to continue to do work during their shutdown methods, e.g. flushing data.
3327 12 : tracing::debug!("Cancelling CancellationToken");
3328 12 : self.cancel.cancel();
3329 12 :
3330 12 : // shutdown all tenant and timeline tasks: gc, compaction, page service
3331 12 : // No new tasks will be started for this tenant because it's in `Stopping` state.
3332 12 : //
3333 12 : // this will additionally shutdown and await all timeline tasks.
3334 12 : tracing::debug!("Waiting for tasks...");
3335 12 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), None).await;
3336 :
3337 12 : if let Some(walredo_mgr) = self.walredo_mgr.as_ref() {
3338 12 : walredo_mgr.shutdown().await;
3339 0 : }
3340 :
3341 : // Wait for any in-flight operations to complete
3342 12 : self.gate.close().await;
3343 :
3344 12 : remove_tenant_metrics(&self.tenant_shard_id);
3345 12 :
3346 12 : Ok(())
3347 12 : }
3348 :
3349 : /// Change tenant status to Stopping, to mark that it is being shut down.
3350 : ///
3351 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
3352 : ///
3353 : /// This function is not cancel-safe!
3354 : ///
3355 : /// `allow_transition_from_loading` is needed for the special case of loading task deleting the tenant.
3356 : /// `allow_transition_from_attaching` is needed for the special case of attaching deleted tenant.
3357 12 : async fn set_stopping(
3358 12 : &self,
3359 12 : progress: completion::Barrier,
3360 12 : _allow_transition_from_loading: bool,
3361 12 : allow_transition_from_attaching: bool,
3362 12 : ) -> Result<(), SetStoppingError> {
3363 12 : let mut rx = self.state.subscribe();
3364 12 :
3365 12 : // cannot stop before we're done activating, so wait out until we're done activating
3366 12 : rx.wait_for(|state| match state {
3367 0 : TenantState::Attaching if allow_transition_from_attaching => true,
3368 : TenantState::Activating(_) | TenantState::Attaching => {
3369 0 : info!(
3370 0 : "waiting for {} to turn Active|Broken|Stopping",
3371 0 : <&'static str>::from(state)
3372 : );
3373 0 : false
3374 : }
3375 12 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3376 12 : })
3377 12 : .await
3378 12 : .expect("cannot drop self.state while on a &self method");
3379 12 :
3380 12 : // we now know we're done activating, let's see whether this task is the winner to transition into Stopping
3381 12 : let mut err = None;
3382 12 : let stopping = self.state.send_if_modified(|current_state| match current_state {
3383 : TenantState::Activating(_) => {
3384 0 : unreachable!("1we ensured above that we're done with activation, and, there is no re-activation")
3385 : }
3386 : TenantState::Attaching => {
3387 0 : if !allow_transition_from_attaching {
3388 0 : unreachable!("2we ensured above that we're done with activation, and, there is no re-activation")
3389 0 : };
3390 0 : *current_state = TenantState::Stopping { progress };
3391 0 : true
3392 : }
3393 : TenantState::Active => {
3394 : // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
3395 : // are created after the transition to Stopping. That's harmless, as the Timelines
3396 : // won't be accessible to anyone afterwards, because the Tenant is in Stopping state.
3397 12 : *current_state = TenantState::Stopping { progress };
3398 12 : // Continue stopping outside the closure. We need to grab timelines.lock()
3399 12 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
3400 12 : true
3401 : }
3402 0 : TenantState::Broken { reason, .. } => {
3403 0 : info!(
3404 0 : "Cannot set tenant to Stopping state, it is in Broken state due to: {reason}"
3405 : );
3406 0 : err = Some(SetStoppingError::Broken);
3407 0 : false
3408 : }
3409 0 : TenantState::Stopping { progress } => {
3410 0 : info!("Tenant is already in Stopping state");
3411 0 : err = Some(SetStoppingError::AlreadyStopping(progress.clone()));
3412 0 : false
3413 : }
3414 12 : });
3415 12 : match (stopping, err) {
3416 12 : (true, None) => {} // continue
3417 0 : (false, Some(err)) => return Err(err),
3418 0 : (true, Some(_)) => unreachable!(
3419 0 : "send_if_modified closure must error out if not transitioning to Stopping"
3420 0 : ),
3421 0 : (false, None) => unreachable!(
3422 0 : "send_if_modified closure must return true if transitioning to Stopping"
3423 0 : ),
3424 : }
3425 :
3426 12 : let timelines_accessor = self.timelines.lock().unwrap();
3427 12 : let not_broken_timelines = timelines_accessor
3428 12 : .values()
3429 12 : .filter(|timeline| !timeline.is_broken());
3430 24 : for timeline in not_broken_timelines {
3431 12 : timeline.set_state(TimelineState::Stopping);
3432 12 : }
3433 12 : Ok(())
3434 12 : }
3435 :
3436 : /// Method for tenant::mgr to transition us into Broken state in case of a late failure in
3437 : /// `remove_tenant_from_memory`
3438 : ///
3439 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
3440 : ///
3441 : /// In tests, we also use this to set tenants to Broken state on purpose.
3442 0 : pub(crate) async fn set_broken(&self, reason: String) {
3443 0 : let mut rx = self.state.subscribe();
3444 0 :
3445 0 : // The load & attach routines own the tenant state until it has reached `Active`.
3446 0 : // So, wait until it's done.
3447 0 : rx.wait_for(|state| match state {
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 0 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3456 0 : })
3457 0 : .await
3458 0 : .expect("cannot drop self.state while on a &self method");
3459 0 :
3460 0 : // we now know we're done activating, let's see whether this task is the winner to transition into Broken
3461 0 : self.set_broken_no_wait(reason)
3462 0 : }
3463 :
3464 0 : pub(crate) fn set_broken_no_wait(&self, reason: impl Display) {
3465 0 : let reason = reason.to_string();
3466 0 : self.state.send_modify(|current_state| {
3467 0 : match *current_state {
3468 : TenantState::Activating(_) | TenantState::Attaching => {
3469 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
3470 : }
3471 : TenantState::Active => {
3472 0 : if cfg!(feature = "testing") {
3473 0 : warn!("Changing Active tenant to Broken state, reason: {}", reason);
3474 0 : *current_state = TenantState::broken_from_reason(reason);
3475 : } else {
3476 0 : unreachable!("not allowed to call set_broken on Active tenants in non-testing builds")
3477 : }
3478 : }
3479 : TenantState::Broken { .. } => {
3480 0 : warn!("Tenant is already in Broken state");
3481 : }
3482 : // This is the only "expected" path, any other path is a bug.
3483 : TenantState::Stopping { .. } => {
3484 0 : warn!(
3485 0 : "Marking Stopping tenant as Broken state, reason: {}",
3486 : reason
3487 : );
3488 0 : *current_state = TenantState::broken_from_reason(reason);
3489 : }
3490 : }
3491 0 : });
3492 0 : }
3493 :
3494 0 : pub fn subscribe_for_state_updates(&self) -> watch::Receiver<TenantState> {
3495 0 : self.state.subscribe()
3496 0 : }
3497 :
3498 : /// The activate_now semaphore is initialized with zero units. As soon as
3499 : /// we add a unit, waiters will be able to acquire a unit and proceed.
3500 0 : pub(crate) fn activate_now(&self) {
3501 0 : self.activate_now_sem.add_permits(1);
3502 0 : }
3503 :
3504 0 : pub(crate) async fn wait_to_become_active(
3505 0 : &self,
3506 0 : timeout: Duration,
3507 0 : ) -> Result<(), GetActiveTenantError> {
3508 0 : let mut receiver = self.state.subscribe();
3509 : loop {
3510 0 : let current_state = receiver.borrow_and_update().clone();
3511 0 : match current_state {
3512 : TenantState::Attaching | TenantState::Activating(_) => {
3513 : // in these states, there's a chance that we can reach ::Active
3514 0 : self.activate_now();
3515 0 : match timeout_cancellable(timeout, &self.cancel, receiver.changed()).await {
3516 0 : Ok(r) => {
3517 0 : r.map_err(
3518 0 : |_e: tokio::sync::watch::error::RecvError|
3519 : // Tenant existed but was dropped: report it as non-existent
3520 0 : GetActiveTenantError::NotFound(GetTenantError::ShardNotFound(self.tenant_shard_id))
3521 0 : )?
3522 : }
3523 : Err(TimeoutCancellableError::Cancelled) => {
3524 0 : return Err(GetActiveTenantError::Cancelled);
3525 : }
3526 : Err(TimeoutCancellableError::Timeout) => {
3527 0 : return Err(GetActiveTenantError::WaitForActiveTimeout {
3528 0 : latest_state: Some(self.current_state()),
3529 0 : wait_time: timeout,
3530 0 : });
3531 : }
3532 : }
3533 : }
3534 : TenantState::Active { .. } => {
3535 0 : return Ok(());
3536 : }
3537 0 : TenantState::Broken { reason, .. } => {
3538 0 : // This is fatal, and reported distinctly from the general case of "will never be active" because
3539 0 : // it's logically a 500 to external API users (broken is always a bug).
3540 0 : return Err(GetActiveTenantError::Broken(reason));
3541 : }
3542 : TenantState::Stopping { .. } => {
3543 : // There's no chance the tenant can transition back into ::Active
3544 0 : return Err(GetActiveTenantError::WillNotBecomeActive(current_state));
3545 : }
3546 : }
3547 : }
3548 0 : }
3549 :
3550 0 : pub(crate) fn get_attach_mode(&self) -> AttachmentMode {
3551 0 : self.tenant_conf.load().location.attach_mode
3552 0 : }
3553 :
3554 : /// For API access: generate a LocationConfig equivalent to the one that would be used to
3555 : /// create a Tenant in the same state. Do not use this in hot paths: it's for relatively
3556 : /// rare external API calls, like a reconciliation at startup.
3557 0 : pub(crate) fn get_location_conf(&self) -> models::LocationConfig {
3558 0 : let conf = self.tenant_conf.load();
3559 :
3560 0 : let location_config_mode = match conf.location.attach_mode {
3561 0 : AttachmentMode::Single => models::LocationConfigMode::AttachedSingle,
3562 0 : AttachmentMode::Multi => models::LocationConfigMode::AttachedMulti,
3563 0 : AttachmentMode::Stale => models::LocationConfigMode::AttachedStale,
3564 : };
3565 :
3566 : // We have a pageserver TenantConf, we need the API-facing TenantConfig.
3567 0 : let tenant_config: models::TenantConfig = conf.tenant_conf.clone().into();
3568 0 :
3569 0 : models::LocationConfig {
3570 0 : mode: location_config_mode,
3571 0 : generation: self.generation.into(),
3572 0 : secondary_conf: None,
3573 0 : shard_number: self.shard_identity.number.0,
3574 0 : shard_count: self.shard_identity.count.literal(),
3575 0 : shard_stripe_size: self.shard_identity.stripe_size.0,
3576 0 : tenant_conf: tenant_config,
3577 0 : }
3578 0 : }
3579 :
3580 0 : pub(crate) fn get_tenant_shard_id(&self) -> &TenantShardId {
3581 0 : &self.tenant_shard_id
3582 0 : }
3583 :
3584 0 : pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
3585 0 : self.shard_identity.stripe_size
3586 0 : }
3587 :
3588 0 : pub(crate) fn get_generation(&self) -> Generation {
3589 0 : self.generation
3590 0 : }
3591 :
3592 : /// This function partially shuts down the tenant (it shuts down the Timelines) and is fallible,
3593 : /// and can leave the tenant in a bad state if it fails. The caller is responsible for
3594 : /// resetting this tenant to a valid state if we fail.
3595 0 : pub(crate) async fn split_prepare(
3596 0 : &self,
3597 0 : child_shards: &Vec<TenantShardId>,
3598 0 : ) -> anyhow::Result<()> {
3599 0 : let (timelines, offloaded) = {
3600 0 : let timelines = self.timelines.lock().unwrap();
3601 0 : let offloaded = self.timelines_offloaded.lock().unwrap();
3602 0 : (timelines.clone(), offloaded.clone())
3603 0 : };
3604 0 : let timelines_iter = timelines
3605 0 : .values()
3606 0 : .map(TimelineOrOffloadedArcRef::<'_>::from)
3607 0 : .chain(
3608 0 : offloaded
3609 0 : .values()
3610 0 : .map(TimelineOrOffloadedArcRef::<'_>::from),
3611 0 : );
3612 0 : for timeline in timelines_iter {
3613 : // We do not block timeline creation/deletion during splits inside the pageserver: it is up to higher levels
3614 : // to ensure that they do not start a split if currently in the process of doing these.
3615 :
3616 0 : let timeline_id = timeline.timeline_id();
3617 :
3618 0 : if let TimelineOrOffloadedArcRef::Timeline(timeline) = timeline {
3619 : // Upload an index from the parent: this is partly to provide freshness for the
3620 : // child tenants that will copy it, and partly for general ease-of-debugging: there will
3621 : // always be a parent shard index in the same generation as we wrote the child shard index.
3622 0 : tracing::info!(%timeline_id, "Uploading index");
3623 0 : timeline
3624 0 : .remote_client
3625 0 : .schedule_index_upload_for_file_changes()?;
3626 0 : timeline.remote_client.wait_completion().await?;
3627 0 : }
3628 :
3629 0 : let remote_client = match timeline {
3630 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.remote_client.clone(),
3631 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => {
3632 0 : let remote_client = self
3633 0 : .build_timeline_client(offloaded.timeline_id, self.remote_storage.clone());
3634 0 : Arc::new(remote_client)
3635 : }
3636 : };
3637 :
3638 : // Shut down the timeline's remote client: this means that the indices we write
3639 : // for child shards will not be invalidated by the parent shard deleting layers.
3640 0 : tracing::info!(%timeline_id, "Shutting down remote storage client");
3641 0 : remote_client.shutdown().await;
3642 :
3643 : // Download methods can still be used after shutdown, as they don't flow through the remote client's
3644 : // queue. In principal the RemoteTimelineClient could provide this without downloading it, but this
3645 : // operation is rare, so it's simpler to just download it (and robustly guarantees that the index
3646 : // we use here really is the remotely persistent one).
3647 0 : tracing::info!(%timeline_id, "Downloading index_part from parent");
3648 0 : let result = remote_client
3649 0 : .download_index_file(&self.cancel)
3650 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))
3651 0 : .await?;
3652 0 : let index_part = match result {
3653 : MaybeDeletedIndexPart::Deleted(_) => {
3654 0 : anyhow::bail!("Timeline deletion happened concurrently with split")
3655 : }
3656 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
3657 : };
3658 :
3659 0 : for child_shard in child_shards {
3660 0 : tracing::info!(%timeline_id, "Uploading index_part for child {}", child_shard.to_index());
3661 0 : upload_index_part(
3662 0 : &self.remote_storage,
3663 0 : child_shard,
3664 0 : &timeline_id,
3665 0 : self.generation,
3666 0 : &index_part,
3667 0 : &self.cancel,
3668 0 : )
3669 0 : .await?;
3670 : }
3671 : }
3672 :
3673 0 : let tenant_manifest = self.build_tenant_manifest();
3674 0 : for child_shard in child_shards {
3675 0 : tracing::info!(
3676 0 : "Uploading tenant manifest for child {}",
3677 0 : child_shard.to_index()
3678 : );
3679 0 : upload_tenant_manifest(
3680 0 : &self.remote_storage,
3681 0 : child_shard,
3682 0 : self.generation,
3683 0 : &tenant_manifest,
3684 0 : &self.cancel,
3685 0 : )
3686 0 : .await?;
3687 : }
3688 :
3689 0 : Ok(())
3690 0 : }
3691 :
3692 0 : pub(crate) fn get_sizes(&self) -> TopTenantShardItem {
3693 0 : let mut result = TopTenantShardItem {
3694 0 : id: self.tenant_shard_id,
3695 0 : resident_size: 0,
3696 0 : physical_size: 0,
3697 0 : max_logical_size: 0,
3698 0 : };
3699 :
3700 0 : for timeline in self.timelines.lock().unwrap().values() {
3701 0 : result.resident_size += timeline.metrics.resident_physical_size_gauge.get();
3702 0 :
3703 0 : result.physical_size += timeline
3704 0 : .remote_client
3705 0 : .metrics
3706 0 : .remote_physical_size_gauge
3707 0 : .get();
3708 0 : result.max_logical_size = std::cmp::max(
3709 0 : result.max_logical_size,
3710 0 : timeline.metrics.current_logical_size_gauge.get(),
3711 0 : );
3712 0 : }
3713 :
3714 0 : result
3715 0 : }
3716 : }
3717 :
3718 : /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
3719 : /// perform a topological sort, so that the parent of each timeline comes
3720 : /// before the children.
3721 : /// E extracts the ancestor from T
3722 : /// This allows for T to be different. It can be TimelineMetadata, can be Timeline itself, etc.
3723 440 : fn tree_sort_timelines<T, E>(
3724 440 : timelines: HashMap<TimelineId, T>,
3725 440 : extractor: E,
3726 440 : ) -> anyhow::Result<Vec<(TimelineId, T)>>
3727 440 : where
3728 440 : E: Fn(&T) -> Option<TimelineId>,
3729 440 : {
3730 440 : let mut result = Vec::with_capacity(timelines.len());
3731 440 :
3732 440 : let mut now = Vec::with_capacity(timelines.len());
3733 440 : // (ancestor, children)
3734 440 : let mut later: HashMap<TimelineId, Vec<(TimelineId, T)>> =
3735 440 : HashMap::with_capacity(timelines.len());
3736 :
3737 452 : for (timeline_id, value) in timelines {
3738 12 : if let Some(ancestor_id) = extractor(&value) {
3739 4 : let children = later.entry(ancestor_id).or_default();
3740 4 : children.push((timeline_id, value));
3741 8 : } else {
3742 8 : now.push((timeline_id, value));
3743 8 : }
3744 : }
3745 :
3746 452 : while let Some((timeline_id, metadata)) = now.pop() {
3747 12 : result.push((timeline_id, metadata));
3748 : // All children of this can be loaded now
3749 12 : if let Some(mut children) = later.remove(&timeline_id) {
3750 4 : now.append(&mut children);
3751 8 : }
3752 : }
3753 :
3754 : // All timelines should be visited now. Unless there were timelines with missing ancestors.
3755 440 : if !later.is_empty() {
3756 0 : for (missing_id, orphan_ids) in later {
3757 0 : for (orphan_id, _) in orphan_ids {
3758 0 : error!("could not load timeline {orphan_id} because its ancestor timeline {missing_id} could not be loaded");
3759 : }
3760 : }
3761 0 : bail!("could not load tenant because some timelines are missing ancestors");
3762 440 : }
3763 440 :
3764 440 : Ok(result)
3765 440 : }
3766 :
3767 : enum ActivateTimelineArgs {
3768 : Yes {
3769 : broker_client: storage_broker::BrokerClientChannel,
3770 : },
3771 : No,
3772 : }
3773 :
3774 : impl Tenant {
3775 0 : pub fn tenant_specific_overrides(&self) -> TenantConfOpt {
3776 0 : self.tenant_conf.load().tenant_conf.clone()
3777 0 : }
3778 :
3779 0 : pub fn effective_config(&self) -> TenantConf {
3780 0 : self.tenant_specific_overrides()
3781 0 : .merge(self.conf.default_tenant_conf.clone())
3782 0 : }
3783 :
3784 0 : pub fn get_checkpoint_distance(&self) -> u64 {
3785 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3786 0 : tenant_conf
3787 0 : .checkpoint_distance
3788 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
3789 0 : }
3790 :
3791 0 : pub fn get_checkpoint_timeout(&self) -> Duration {
3792 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3793 0 : tenant_conf
3794 0 : .checkpoint_timeout
3795 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
3796 0 : }
3797 :
3798 0 : pub fn get_compaction_target_size(&self) -> u64 {
3799 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3800 0 : tenant_conf
3801 0 : .compaction_target_size
3802 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
3803 0 : }
3804 :
3805 0 : pub fn get_compaction_period(&self) -> Duration {
3806 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3807 0 : tenant_conf
3808 0 : .compaction_period
3809 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
3810 0 : }
3811 :
3812 0 : pub fn get_compaction_threshold(&self) -> usize {
3813 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3814 0 : tenant_conf
3815 0 : .compaction_threshold
3816 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
3817 0 : }
3818 :
3819 0 : pub fn get_compaction_upper_limit(&self) -> usize {
3820 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3821 0 : tenant_conf
3822 0 : .compaction_upper_limit
3823 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_upper_limit)
3824 0 : }
3825 :
3826 0 : pub fn get_gc_horizon(&self) -> u64 {
3827 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3828 0 : tenant_conf
3829 0 : .gc_horizon
3830 0 : .unwrap_or(self.conf.default_tenant_conf.gc_horizon)
3831 0 : }
3832 :
3833 0 : pub fn get_gc_period(&self) -> Duration {
3834 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3835 0 : tenant_conf
3836 0 : .gc_period
3837 0 : .unwrap_or(self.conf.default_tenant_conf.gc_period)
3838 0 : }
3839 :
3840 0 : pub fn get_image_creation_threshold(&self) -> usize {
3841 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3842 0 : tenant_conf
3843 0 : .image_creation_threshold
3844 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
3845 0 : }
3846 :
3847 0 : pub fn get_pitr_interval(&self) -> Duration {
3848 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3849 0 : tenant_conf
3850 0 : .pitr_interval
3851 0 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
3852 0 : }
3853 :
3854 0 : pub fn get_min_resident_size_override(&self) -> Option<u64> {
3855 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3856 0 : tenant_conf
3857 0 : .min_resident_size_override
3858 0 : .or(self.conf.default_tenant_conf.min_resident_size_override)
3859 0 : }
3860 :
3861 0 : pub fn get_heatmap_period(&self) -> Option<Duration> {
3862 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3863 0 : let heatmap_period = tenant_conf
3864 0 : .heatmap_period
3865 0 : .unwrap_or(self.conf.default_tenant_conf.heatmap_period);
3866 0 : if heatmap_period.is_zero() {
3867 0 : None
3868 : } else {
3869 0 : Some(heatmap_period)
3870 : }
3871 0 : }
3872 :
3873 8 : pub fn get_lsn_lease_length(&self) -> Duration {
3874 8 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3875 8 : tenant_conf
3876 8 : .lsn_lease_length
3877 8 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
3878 8 : }
3879 :
3880 : /// Generate an up-to-date TenantManifest based on the state of this Tenant.
3881 4 : fn build_tenant_manifest(&self) -> TenantManifest {
3882 4 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
3883 4 :
3884 4 : let mut timeline_manifests = timelines_offloaded
3885 4 : .iter()
3886 4 : .map(|(_timeline_id, offloaded)| offloaded.manifest())
3887 4 : .collect::<Vec<_>>();
3888 4 : // Sort the manifests so that our output is deterministic
3889 4 : timeline_manifests.sort_by_key(|timeline_manifest| timeline_manifest.timeline_id);
3890 4 :
3891 4 : TenantManifest {
3892 4 : version: LATEST_TENANT_MANIFEST_VERSION,
3893 4 : offloaded_timelines: timeline_manifests,
3894 4 : }
3895 4 : }
3896 :
3897 0 : pub fn update_tenant_config<F: Fn(TenantConfOpt) -> anyhow::Result<TenantConfOpt>>(
3898 0 : &self,
3899 0 : update: F,
3900 0 : ) -> anyhow::Result<TenantConfOpt> {
3901 0 : // Use read-copy-update in order to avoid overwriting the location config
3902 0 : // state if this races with [`Tenant::set_new_location_config`]. Note that
3903 0 : // this race is not possible if both request types come from the storage
3904 0 : // controller (as they should!) because an exclusive op lock is required
3905 0 : // on the storage controller side.
3906 0 :
3907 0 : self.tenant_conf
3908 0 : .try_rcu(|attached_conf| -> Result<_, anyhow::Error> {
3909 0 : Ok(Arc::new(AttachedTenantConf {
3910 0 : tenant_conf: update(attached_conf.tenant_conf.clone())?,
3911 0 : location: attached_conf.location,
3912 0 : lsn_lease_deadline: attached_conf.lsn_lease_deadline,
3913 : }))
3914 0 : })?;
3915 :
3916 0 : let updated = self.tenant_conf.load();
3917 0 :
3918 0 : self.tenant_conf_updated(&updated.tenant_conf);
3919 0 : // Don't hold self.timelines.lock() during the notifies.
3920 0 : // There's no risk of deadlock right now, but there could be if we consolidate
3921 0 : // mutexes in struct Timeline in the future.
3922 0 : let timelines = self.list_timelines();
3923 0 : for timeline in timelines {
3924 0 : timeline.tenant_conf_updated(&updated);
3925 0 : }
3926 :
3927 0 : Ok(updated.tenant_conf.clone())
3928 0 : }
3929 :
3930 0 : pub(crate) fn set_new_location_config(&self, new_conf: AttachedTenantConf) {
3931 0 : let new_tenant_conf = new_conf.tenant_conf.clone();
3932 0 :
3933 0 : self.tenant_conf.store(Arc::new(new_conf.clone()));
3934 0 :
3935 0 : self.tenant_conf_updated(&new_tenant_conf);
3936 0 : // Don't hold self.timelines.lock() during the notifies.
3937 0 : // There's no risk of deadlock right now, but there could be if we consolidate
3938 0 : // mutexes in struct Timeline in the future.
3939 0 : let timelines = self.list_timelines();
3940 0 : for timeline in timelines {
3941 0 : timeline.tenant_conf_updated(&new_conf);
3942 0 : }
3943 0 : }
3944 :
3945 440 : fn get_pagestream_throttle_config(
3946 440 : psconf: &'static PageServerConf,
3947 440 : overrides: &TenantConfOpt,
3948 440 : ) -> throttle::Config {
3949 440 : overrides
3950 440 : .timeline_get_throttle
3951 440 : .clone()
3952 440 : .unwrap_or(psconf.default_tenant_conf.timeline_get_throttle.clone())
3953 440 : }
3954 :
3955 0 : pub(crate) fn tenant_conf_updated(&self, new_conf: &TenantConfOpt) {
3956 0 : let conf = Self::get_pagestream_throttle_config(self.conf, new_conf);
3957 0 : self.pagestream_throttle.reconfigure(conf)
3958 0 : }
3959 :
3960 : /// Helper function to create a new Timeline struct.
3961 : ///
3962 : /// The returned Timeline is in Loading state. The caller is responsible for
3963 : /// initializing any on-disk state, and for inserting the Timeline to the 'timelines'
3964 : /// map.
3965 : ///
3966 : /// `validate_ancestor == false` is used when a timeline is created for deletion
3967 : /// and we might not have the ancestor present anymore which is fine for to be
3968 : /// deleted timelines.
3969 : #[allow(clippy::too_many_arguments)]
3970 892 : fn create_timeline_struct(
3971 892 : &self,
3972 892 : new_timeline_id: TimelineId,
3973 892 : new_metadata: &TimelineMetadata,
3974 892 : ancestor: Option<Arc<Timeline>>,
3975 892 : resources: TimelineResources,
3976 892 : cause: CreateTimelineCause,
3977 892 : create_idempotency: CreateTimelineIdempotency,
3978 892 : ) -> anyhow::Result<Arc<Timeline>> {
3979 892 : let state = match cause {
3980 : CreateTimelineCause::Load => {
3981 892 : let ancestor_id = new_metadata.ancestor_timeline();
3982 892 : anyhow::ensure!(
3983 892 : ancestor_id == ancestor.as_ref().map(|t| t.timeline_id),
3984 0 : "Timeline's {new_timeline_id} ancestor {ancestor_id:?} was not found"
3985 : );
3986 892 : TimelineState::Loading
3987 : }
3988 0 : CreateTimelineCause::Delete => TimelineState::Stopping,
3989 : };
3990 :
3991 892 : let pg_version = new_metadata.pg_version();
3992 892 :
3993 892 : let timeline = Timeline::new(
3994 892 : self.conf,
3995 892 : Arc::clone(&self.tenant_conf),
3996 892 : new_metadata,
3997 892 : ancestor,
3998 892 : new_timeline_id,
3999 892 : self.tenant_shard_id,
4000 892 : self.generation,
4001 892 : self.shard_identity,
4002 892 : self.walredo_mgr.clone(),
4003 892 : resources,
4004 892 : pg_version,
4005 892 : state,
4006 892 : self.attach_wal_lag_cooldown.clone(),
4007 892 : create_idempotency,
4008 892 : self.cancel.child_token(),
4009 892 : );
4010 892 :
4011 892 : Ok(timeline)
4012 892 : }
4013 :
4014 : /// [`Tenant::shutdown`] must be called before dropping the returned [`Tenant`] object
4015 : /// to ensure proper cleanup of background tasks and metrics.
4016 : //
4017 : // Allow too_many_arguments because a constructor's argument list naturally grows with the
4018 : // number of attributes in the struct: breaking these out into a builder wouldn't be helpful.
4019 : #[allow(clippy::too_many_arguments)]
4020 440 : fn new(
4021 440 : state: TenantState,
4022 440 : conf: &'static PageServerConf,
4023 440 : attached_conf: AttachedTenantConf,
4024 440 : shard_identity: ShardIdentity,
4025 440 : walredo_mgr: Option<Arc<WalRedoManager>>,
4026 440 : tenant_shard_id: TenantShardId,
4027 440 : remote_storage: GenericRemoteStorage,
4028 440 : deletion_queue_client: DeletionQueueClient,
4029 440 : l0_flush_global_state: L0FlushGlobalState,
4030 440 : ) -> Tenant {
4031 440 : debug_assert!(
4032 440 : !attached_conf.location.generation.is_none() || conf.control_plane_api.is_none()
4033 : );
4034 :
4035 440 : let (state, mut rx) = watch::channel(state);
4036 440 :
4037 440 : tokio::spawn(async move {
4038 439 : // reflect tenant state in metrics:
4039 439 : // - global per tenant state: TENANT_STATE_METRIC
4040 439 : // - "set" of broken tenants: BROKEN_TENANTS_SET
4041 439 : //
4042 439 : // set of broken tenants should not have zero counts so that it remains accessible for
4043 439 : // alerting.
4044 439 :
4045 439 : let tid = tenant_shard_id.to_string();
4046 439 : let shard_id = tenant_shard_id.shard_slug().to_string();
4047 439 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
4048 :
4049 879 : fn inspect_state(state: &TenantState) -> ([&'static str; 1], bool) {
4050 879 : ([state.into()], matches!(state, TenantState::Broken { .. }))
4051 879 : }
4052 :
4053 439 : let mut tuple = inspect_state(&rx.borrow_and_update());
4054 439 :
4055 439 : let is_broken = tuple.1;
4056 439 : let mut counted_broken = if is_broken {
4057 : // add the id to the set right away, there should not be any updates on the channel
4058 : // after before tenant is removed, if ever
4059 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
4060 0 : true
4061 : } else {
4062 439 : false
4063 : };
4064 :
4065 : loop {
4066 879 : let labels = &tuple.0;
4067 879 : let current = TENANT_STATE_METRIC.with_label_values(labels);
4068 879 : current.inc();
4069 879 :
4070 879 : if rx.changed().await.is_err() {
4071 : // tenant has been dropped
4072 28 : current.dec();
4073 28 : drop(BROKEN_TENANTS_SET.remove_label_values(set_key));
4074 28 : break;
4075 440 : }
4076 440 :
4077 440 : current.dec();
4078 440 : tuple = inspect_state(&rx.borrow_and_update());
4079 440 :
4080 440 : let is_broken = tuple.1;
4081 440 : if is_broken && !counted_broken {
4082 0 : counted_broken = true;
4083 0 : // insert the tenant_id (back) into the set while avoiding needless counter
4084 0 : // access
4085 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
4086 440 : }
4087 : }
4088 440 : });
4089 440 :
4090 440 : Tenant {
4091 440 : tenant_shard_id,
4092 440 : shard_identity,
4093 440 : generation: attached_conf.location.generation,
4094 440 : conf,
4095 440 : // using now here is good enough approximation to catch tenants with really long
4096 440 : // activation times.
4097 440 : constructed_at: Instant::now(),
4098 440 : timelines: Mutex::new(HashMap::new()),
4099 440 : timelines_creating: Mutex::new(HashSet::new()),
4100 440 : timelines_offloaded: Mutex::new(HashMap::new()),
4101 440 : tenant_manifest_upload: Default::default(),
4102 440 : gc_cs: tokio::sync::Mutex::new(()),
4103 440 : walredo_mgr,
4104 440 : remote_storage,
4105 440 : deletion_queue_client,
4106 440 : state,
4107 440 : cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
4108 440 : cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
4109 440 : eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
4110 440 : compaction_circuit_breaker: std::sync::Mutex::new(CircuitBreaker::new(
4111 440 : format!("compaction-{tenant_shard_id}"),
4112 440 : 5,
4113 440 : // Compaction can be a very expensive operation, and might leak disk space. It also ought
4114 440 : // to be infallible, as long as remote storage is available. So if it repeatedly fails,
4115 440 : // use an extremely long backoff.
4116 440 : Some(Duration::from_secs(3600 * 24)),
4117 440 : )),
4118 440 : scheduled_compaction_tasks: Mutex::new(Default::default()),
4119 440 : activate_now_sem: tokio::sync::Semaphore::new(0),
4120 440 : attach_wal_lag_cooldown: Arc::new(std::sync::OnceLock::new()),
4121 440 : cancel: CancellationToken::default(),
4122 440 : gate: Gate::default(),
4123 440 : pagestream_throttle: Arc::new(throttle::Throttle::new(
4124 440 : Tenant::get_pagestream_throttle_config(conf, &attached_conf.tenant_conf),
4125 440 : )),
4126 440 : pagestream_throttle_metrics: Arc::new(
4127 440 : crate::metrics::tenant_throttling::Pagestream::new(&tenant_shard_id),
4128 440 : ),
4129 440 : tenant_conf: Arc::new(ArcSwap::from_pointee(attached_conf)),
4130 440 : ongoing_timeline_detach: std::sync::Mutex::default(),
4131 440 : gc_block: Default::default(),
4132 440 : l0_flush_global_state,
4133 440 : }
4134 440 : }
4135 :
4136 : /// Locate and load config
4137 0 : pub(super) fn load_tenant_config(
4138 0 : conf: &'static PageServerConf,
4139 0 : tenant_shard_id: &TenantShardId,
4140 0 : ) -> Result<LocationConf, LoadConfigError> {
4141 0 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4142 0 :
4143 0 : info!("loading tenant configuration from {config_path}");
4144 :
4145 : // load and parse file
4146 0 : let config = fs::read_to_string(&config_path).map_err(|e| {
4147 0 : match e.kind() {
4148 : std::io::ErrorKind::NotFound => {
4149 : // The config should almost always exist for a tenant directory:
4150 : // - When attaching a tenant, the config is the first thing we write
4151 : // - When detaching a tenant, we atomically move the directory to a tmp location
4152 : // before deleting contents.
4153 : //
4154 : // The very rare edge case that can result in a missing config is if we crash during attach
4155 : // between creating directory and writing config. Callers should handle that as if the
4156 : // directory didn't exist.
4157 :
4158 0 : LoadConfigError::NotFound(config_path)
4159 : }
4160 : _ => {
4161 : // No IO errors except NotFound are acceptable here: other kinds of error indicate local storage or permissions issues
4162 : // that we cannot cleanly recover
4163 0 : crate::virtual_file::on_fatal_io_error(&e, "Reading tenant config file")
4164 : }
4165 : }
4166 0 : })?;
4167 :
4168 0 : Ok(toml_edit::de::from_str::<LocationConf>(&config)?)
4169 0 : }
4170 :
4171 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4172 : pub(super) async fn persist_tenant_config(
4173 : conf: &'static PageServerConf,
4174 : tenant_shard_id: &TenantShardId,
4175 : location_conf: &LocationConf,
4176 : ) -> std::io::Result<()> {
4177 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4178 :
4179 : Self::persist_tenant_config_at(tenant_shard_id, &config_path, location_conf).await
4180 : }
4181 :
4182 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4183 : pub(super) async fn persist_tenant_config_at(
4184 : tenant_shard_id: &TenantShardId,
4185 : config_path: &Utf8Path,
4186 : location_conf: &LocationConf,
4187 : ) -> std::io::Result<()> {
4188 : debug!("persisting tenantconf to {config_path}");
4189 :
4190 : let mut conf_content = r#"# This file contains a specific per-tenant's config.
4191 : # It is read in case of pageserver restart.
4192 : "#
4193 : .to_string();
4194 :
4195 0 : fail::fail_point!("tenant-config-before-write", |_| {
4196 0 : Err(std::io::Error::new(
4197 0 : std::io::ErrorKind::Other,
4198 0 : "tenant-config-before-write",
4199 0 : ))
4200 0 : });
4201 :
4202 : // Convert the config to a toml file.
4203 : conf_content +=
4204 : &toml_edit::ser::to_string_pretty(&location_conf).expect("Config serialization failed");
4205 :
4206 : let temp_path = path_with_suffix_extension(config_path, TEMP_FILE_SUFFIX);
4207 :
4208 : let conf_content = conf_content.into_bytes();
4209 : VirtualFile::crashsafe_overwrite(config_path.to_owned(), temp_path, conf_content).await
4210 : }
4211 :
4212 : //
4213 : // How garbage collection works:
4214 : //
4215 : // +--bar------------->
4216 : // /
4217 : // +----+-----foo---------------->
4218 : // /
4219 : // ----main--+-------------------------->
4220 : // \
4221 : // +-----baz-------->
4222 : //
4223 : //
4224 : // 1. Grab 'gc_cs' mutex to prevent new timelines from being created while Timeline's
4225 : // `gc_infos` are being refreshed
4226 : // 2. Scan collected timelines, and on each timeline, make note of the
4227 : // all the points where other timelines have been branched off.
4228 : // We will refrain from removing page versions at those LSNs.
4229 : // 3. For each timeline, scan all layer files on the timeline.
4230 : // Remove all files for which a newer file exists and which
4231 : // don't cover any branch point LSNs.
4232 : //
4233 : // TODO:
4234 : // - if a relation has a non-incremental persistent layer on a child branch, then we
4235 : // don't need to keep that in the parent anymore. But currently
4236 : // we do.
4237 8 : async fn gc_iteration_internal(
4238 8 : &self,
4239 8 : target_timeline_id: Option<TimelineId>,
4240 8 : horizon: u64,
4241 8 : pitr: Duration,
4242 8 : cancel: &CancellationToken,
4243 8 : ctx: &RequestContext,
4244 8 : ) -> Result<GcResult, GcError> {
4245 8 : let mut totals: GcResult = Default::default();
4246 8 : let now = Instant::now();
4247 :
4248 8 : let gc_timelines = self
4249 8 : .refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4250 8 : .await?;
4251 :
4252 8 : failpoint_support::sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
4253 :
4254 : // If there is nothing to GC, we don't want any messages in the INFO log.
4255 8 : if !gc_timelines.is_empty() {
4256 8 : info!("{} timelines need GC", gc_timelines.len());
4257 : } else {
4258 0 : debug!("{} timelines need GC", gc_timelines.len());
4259 : }
4260 :
4261 : // Perform GC for each timeline.
4262 : //
4263 : // Note that we don't hold the `Tenant::gc_cs` lock here because we don't want to delay the
4264 : // branch creation task, which requires the GC lock. A GC iteration can run concurrently
4265 : // with branch creation.
4266 : //
4267 : // See comments in [`Tenant::branch_timeline`] for more information about why branch
4268 : // creation task can run concurrently with timeline's GC iteration.
4269 16 : for timeline in gc_timelines {
4270 8 : if cancel.is_cancelled() {
4271 : // We were requested to shut down. Stop and return with the progress we
4272 : // made.
4273 0 : break;
4274 8 : }
4275 8 : let result = match timeline.gc().await {
4276 : Err(GcError::TimelineCancelled) => {
4277 0 : if target_timeline_id.is_some() {
4278 : // If we were targetting this specific timeline, surface cancellation to caller
4279 0 : return Err(GcError::TimelineCancelled);
4280 : } else {
4281 : // A timeline may be shutting down independently of the tenant's lifecycle: we should
4282 : // skip past this and proceed to try GC on other timelines.
4283 0 : continue;
4284 : }
4285 : }
4286 8 : r => r?,
4287 : };
4288 8 : totals += result;
4289 : }
4290 :
4291 8 : totals.elapsed = now.elapsed();
4292 8 : Ok(totals)
4293 8 : }
4294 :
4295 : /// Refreshes the Timeline::gc_info for all timelines, returning the
4296 : /// vector of timelines which have [`Timeline::get_last_record_lsn`] past
4297 : /// [`Tenant::get_gc_horizon`].
4298 : ///
4299 : /// This is usually executed as part of periodic gc, but can now be triggered more often.
4300 0 : pub(crate) async fn refresh_gc_info(
4301 0 : &self,
4302 0 : cancel: &CancellationToken,
4303 0 : ctx: &RequestContext,
4304 0 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4305 0 : // since this method can now be called at different rates than the configured gc loop, it
4306 0 : // might be that these configuration values get applied faster than what it was previously,
4307 0 : // since these were only read from the gc task.
4308 0 : let horizon = self.get_gc_horizon();
4309 0 : let pitr = self.get_pitr_interval();
4310 0 :
4311 0 : // refresh all timelines
4312 0 : let target_timeline_id = None;
4313 0 :
4314 0 : self.refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4315 0 : .await
4316 0 : }
4317 :
4318 : /// Populate all Timelines' `GcInfo` with information about their children. We do not set the
4319 : /// PITR cutoffs here, because that requires I/O: this is done later, before GC, by [`Self::refresh_gc_info_internal`]
4320 : ///
4321 : /// Subsequently, parent-child relationships are updated incrementally inside [`Timeline::new`] and [`Timeline::drop`].
4322 0 : fn initialize_gc_info(
4323 0 : &self,
4324 0 : timelines: &std::sync::MutexGuard<HashMap<TimelineId, Arc<Timeline>>>,
4325 0 : timelines_offloaded: &std::sync::MutexGuard<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
4326 0 : restrict_to_timeline: Option<TimelineId>,
4327 0 : ) {
4328 0 : if restrict_to_timeline.is_none() {
4329 : // This function must be called before activation: after activation timeline create/delete operations
4330 : // might happen, and this function is not safe to run concurrently with those.
4331 0 : assert!(!self.is_active());
4332 0 : }
4333 :
4334 : // Scan all timelines. For each timeline, remember the timeline ID and
4335 : // the branch point where it was created.
4336 0 : let mut all_branchpoints: BTreeMap<TimelineId, Vec<(Lsn, TimelineId, MaybeOffloaded)>> =
4337 0 : BTreeMap::new();
4338 0 : timelines.iter().for_each(|(timeline_id, timeline_entry)| {
4339 0 : if let Some(ancestor_timeline_id) = &timeline_entry.get_ancestor_timeline_id() {
4340 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4341 0 : ancestor_children.push((
4342 0 : timeline_entry.get_ancestor_lsn(),
4343 0 : *timeline_id,
4344 0 : MaybeOffloaded::No,
4345 0 : ));
4346 0 : }
4347 0 : });
4348 0 : timelines_offloaded
4349 0 : .iter()
4350 0 : .for_each(|(timeline_id, timeline_entry)| {
4351 0 : let Some(ancestor_timeline_id) = &timeline_entry.ancestor_timeline_id else {
4352 0 : return;
4353 : };
4354 0 : let Some(retain_lsn) = timeline_entry.ancestor_retain_lsn else {
4355 0 : return;
4356 : };
4357 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4358 0 : ancestor_children.push((retain_lsn, *timeline_id, MaybeOffloaded::Yes));
4359 0 : });
4360 0 :
4361 0 : // The number of bytes we always keep, irrespective of PITR: this is a constant across timelines
4362 0 : let horizon = self.get_gc_horizon();
4363 :
4364 : // Populate each timeline's GcInfo with information about its child branches
4365 0 : let timelines_to_write = if let Some(timeline_id) = restrict_to_timeline {
4366 0 : itertools::Either::Left(timelines.get(&timeline_id).into_iter())
4367 : } else {
4368 0 : itertools::Either::Right(timelines.values())
4369 : };
4370 0 : for timeline in timelines_to_write {
4371 0 : let mut branchpoints: Vec<(Lsn, TimelineId, MaybeOffloaded)> = all_branchpoints
4372 0 : .remove(&timeline.timeline_id)
4373 0 : .unwrap_or_default();
4374 0 :
4375 0 : branchpoints.sort_by_key(|b| b.0);
4376 0 :
4377 0 : let mut target = timeline.gc_info.write().unwrap();
4378 0 :
4379 0 : target.retain_lsns = branchpoints;
4380 0 :
4381 0 : let space_cutoff = timeline
4382 0 : .get_last_record_lsn()
4383 0 : .checked_sub(horizon)
4384 0 : .unwrap_or(Lsn(0));
4385 0 :
4386 0 : target.cutoffs = GcCutoffs {
4387 0 : space: space_cutoff,
4388 0 : time: Lsn::INVALID,
4389 0 : };
4390 0 : }
4391 0 : }
4392 :
4393 8 : async fn refresh_gc_info_internal(
4394 8 : &self,
4395 8 : target_timeline_id: Option<TimelineId>,
4396 8 : horizon: u64,
4397 8 : pitr: Duration,
4398 8 : cancel: &CancellationToken,
4399 8 : ctx: &RequestContext,
4400 8 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4401 8 : // before taking the gc_cs lock, do the heavier weight finding of gc_cutoff points for
4402 8 : // currently visible timelines.
4403 8 : let timelines = self
4404 8 : .timelines
4405 8 : .lock()
4406 8 : .unwrap()
4407 8 : .values()
4408 8 : .filter(|tl| match target_timeline_id.as_ref() {
4409 8 : Some(target) => &tl.timeline_id == target,
4410 0 : None => true,
4411 8 : })
4412 8 : .cloned()
4413 8 : .collect::<Vec<_>>();
4414 8 :
4415 8 : if target_timeline_id.is_some() && timelines.is_empty() {
4416 : // We were to act on a particular timeline and it wasn't found
4417 0 : return Err(GcError::TimelineNotFound);
4418 8 : }
4419 8 :
4420 8 : let mut gc_cutoffs: HashMap<TimelineId, GcCutoffs> =
4421 8 : HashMap::with_capacity(timelines.len());
4422 8 :
4423 8 : // Ensures all timelines use the same start time when computing the time cutoff.
4424 8 : let now_ts_for_pitr_calc = SystemTime::now();
4425 8 : for timeline in timelines.iter() {
4426 8 : let cutoff = timeline
4427 8 : .get_last_record_lsn()
4428 8 : .checked_sub(horizon)
4429 8 : .unwrap_or(Lsn(0));
4430 :
4431 8 : let cutoffs = timeline
4432 8 : .find_gc_cutoffs(now_ts_for_pitr_calc, cutoff, pitr, cancel, ctx)
4433 8 : .await?;
4434 8 : let old = gc_cutoffs.insert(timeline.timeline_id, cutoffs);
4435 8 : assert!(old.is_none());
4436 : }
4437 :
4438 8 : if !self.is_active() || self.cancel.is_cancelled() {
4439 0 : return Err(GcError::TenantCancelled);
4440 8 : }
4441 :
4442 : // grab mutex to prevent new timelines from being created here; avoid doing long operations
4443 : // because that will stall branch creation.
4444 8 : let gc_cs = self.gc_cs.lock().await;
4445 :
4446 : // Ok, we now know all the branch points.
4447 : // Update the GC information for each timeline.
4448 8 : let mut gc_timelines = Vec::with_capacity(timelines.len());
4449 16 : for timeline in timelines {
4450 : // We filtered the timeline list above
4451 8 : if let Some(target_timeline_id) = target_timeline_id {
4452 8 : assert_eq!(target_timeline_id, timeline.timeline_id);
4453 0 : }
4454 :
4455 : {
4456 8 : let mut target = timeline.gc_info.write().unwrap();
4457 8 :
4458 8 : // Cull any expired leases
4459 8 : let now = SystemTime::now();
4460 12 : target.leases.retain(|_, lease| !lease.is_expired(&now));
4461 8 :
4462 8 : timeline
4463 8 : .metrics
4464 8 : .valid_lsn_lease_count_gauge
4465 8 : .set(target.leases.len() as u64);
4466 :
4467 : // Look up parent's PITR cutoff to update the child's knowledge of whether it is within parent's PITR
4468 8 : if let Some(ancestor_id) = timeline.get_ancestor_timeline_id() {
4469 0 : if let Some(ancestor_gc_cutoffs) = gc_cutoffs.get(&ancestor_id) {
4470 0 : target.within_ancestor_pitr =
4471 0 : timeline.get_ancestor_lsn() >= ancestor_gc_cutoffs.time;
4472 0 : }
4473 8 : }
4474 :
4475 : // Update metrics that depend on GC state
4476 8 : timeline
4477 8 : .metrics
4478 8 : .archival_size
4479 8 : .set(if target.within_ancestor_pitr {
4480 0 : timeline.metrics.current_logical_size_gauge.get()
4481 : } else {
4482 8 : 0
4483 : });
4484 8 : timeline.metrics.pitr_history_size.set(
4485 8 : timeline
4486 8 : .get_last_record_lsn()
4487 8 : .checked_sub(target.cutoffs.time)
4488 8 : .unwrap_or(Lsn(0))
4489 8 : .0,
4490 8 : );
4491 :
4492 : // Apply the cutoffs we found to the Timeline's GcInfo. Why might we _not_ have cutoffs for a timeline?
4493 : // - this timeline was created while we were finding cutoffs
4494 : // - lsn for timestamp search fails for this timeline repeatedly
4495 8 : if let Some(cutoffs) = gc_cutoffs.get(&timeline.timeline_id) {
4496 8 : let original_cutoffs = target.cutoffs.clone();
4497 8 : // GC cutoffs should never go back
4498 8 : target.cutoffs = GcCutoffs {
4499 8 : space: Lsn(cutoffs.space.0.max(original_cutoffs.space.0)),
4500 8 : time: Lsn(cutoffs.time.0.max(original_cutoffs.time.0)),
4501 8 : }
4502 0 : }
4503 : }
4504 :
4505 8 : gc_timelines.push(timeline);
4506 : }
4507 8 : drop(gc_cs);
4508 8 : Ok(gc_timelines)
4509 8 : }
4510 :
4511 : /// A substitute for `branch_timeline` for use in unit tests.
4512 : /// The returned timeline will have state value `Active` to make various `anyhow::ensure!()`
4513 : /// calls pass, but, we do not actually call `.activate()` under the hood. So, none of the
4514 : /// timeline background tasks are launched, except the flush loop.
4515 : #[cfg(test)]
4516 464 : async fn branch_timeline_test(
4517 464 : self: &Arc<Self>,
4518 464 : src_timeline: &Arc<Timeline>,
4519 464 : dst_id: TimelineId,
4520 464 : ancestor_lsn: Option<Lsn>,
4521 464 : ctx: &RequestContext,
4522 464 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
4523 464 : let tl = self
4524 464 : .branch_timeline_impl(src_timeline, dst_id, ancestor_lsn, ctx)
4525 464 : .await?
4526 456 : .into_timeline_for_test();
4527 456 : tl.set_state(TimelineState::Active);
4528 456 : Ok(tl)
4529 464 : }
4530 :
4531 : /// Helper for unit tests to branch a timeline with some pre-loaded states.
4532 : #[cfg(test)]
4533 : #[allow(clippy::too_many_arguments)]
4534 12 : pub async fn branch_timeline_test_with_layers(
4535 12 : self: &Arc<Self>,
4536 12 : src_timeline: &Arc<Timeline>,
4537 12 : dst_id: TimelineId,
4538 12 : ancestor_lsn: Option<Lsn>,
4539 12 : ctx: &RequestContext,
4540 12 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
4541 12 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
4542 12 : end_lsn: Lsn,
4543 12 : ) -> anyhow::Result<Arc<Timeline>> {
4544 : use checks::check_valid_layermap;
4545 : use itertools::Itertools;
4546 :
4547 12 : let tline = self
4548 12 : .branch_timeline_test(src_timeline, dst_id, ancestor_lsn, ctx)
4549 12 : .await?;
4550 12 : let ancestor_lsn = if let Some(ancestor_lsn) = ancestor_lsn {
4551 12 : ancestor_lsn
4552 : } else {
4553 0 : tline.get_last_record_lsn()
4554 : };
4555 12 : assert!(end_lsn >= ancestor_lsn);
4556 12 : tline.force_advance_lsn(end_lsn);
4557 24 : for deltas in delta_layer_desc {
4558 12 : tline
4559 12 : .force_create_delta_layer(deltas, Some(ancestor_lsn), ctx)
4560 12 : .await?;
4561 : }
4562 20 : for (lsn, images) in image_layer_desc {
4563 8 : tline
4564 8 : .force_create_image_layer(lsn, images, Some(ancestor_lsn), ctx)
4565 8 : .await?;
4566 : }
4567 12 : let layer_names = tline
4568 12 : .layers
4569 12 : .read()
4570 12 : .await
4571 12 : .layer_map()
4572 12 : .unwrap()
4573 12 : .iter_historic_layers()
4574 20 : .map(|layer| layer.layer_name())
4575 12 : .collect_vec();
4576 12 : if let Some(err) = check_valid_layermap(&layer_names) {
4577 0 : bail!("invalid layermap: {err}");
4578 12 : }
4579 12 : Ok(tline)
4580 12 : }
4581 :
4582 : /// Branch an existing timeline.
4583 0 : async fn branch_timeline(
4584 0 : self: &Arc<Self>,
4585 0 : src_timeline: &Arc<Timeline>,
4586 0 : dst_id: TimelineId,
4587 0 : start_lsn: Option<Lsn>,
4588 0 : ctx: &RequestContext,
4589 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4590 0 : self.branch_timeline_impl(src_timeline, dst_id, start_lsn, ctx)
4591 0 : .await
4592 0 : }
4593 :
4594 464 : async fn branch_timeline_impl(
4595 464 : self: &Arc<Self>,
4596 464 : src_timeline: &Arc<Timeline>,
4597 464 : dst_id: TimelineId,
4598 464 : start_lsn: Option<Lsn>,
4599 464 : _ctx: &RequestContext,
4600 464 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4601 464 : let src_id = src_timeline.timeline_id;
4602 :
4603 : // We will validate our ancestor LSN in this function. Acquire the GC lock so that
4604 : // this check cannot race with GC, and the ancestor LSN is guaranteed to remain
4605 : // valid while we are creating the branch.
4606 464 : let _gc_cs = self.gc_cs.lock().await;
4607 :
4608 : // If no start LSN is specified, we branch the new timeline from the source timeline's last record LSN
4609 464 : let start_lsn = start_lsn.unwrap_or_else(|| {
4610 4 : let lsn = src_timeline.get_last_record_lsn();
4611 4 : info!("branching timeline {dst_id} from timeline {src_id} at last record LSN: {lsn}");
4612 4 : lsn
4613 464 : });
4614 :
4615 : // we finally have determined the ancestor_start_lsn, so we can get claim exclusivity now
4616 464 : let timeline_create_guard = match self
4617 464 : .start_creating_timeline(
4618 464 : dst_id,
4619 464 : CreateTimelineIdempotency::Branch {
4620 464 : ancestor_timeline_id: src_timeline.timeline_id,
4621 464 : ancestor_start_lsn: start_lsn,
4622 464 : },
4623 464 : )
4624 464 : .await?
4625 : {
4626 464 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
4627 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
4628 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
4629 : }
4630 : };
4631 :
4632 : // Ensure that `start_lsn` is valid, i.e. the LSN is within the PITR
4633 : // horizon on the source timeline
4634 : //
4635 : // We check it against both the planned GC cutoff stored in 'gc_info',
4636 : // and the 'latest_gc_cutoff' of the last GC that was performed. The
4637 : // planned GC cutoff in 'gc_info' is normally larger than
4638 : // 'latest_gc_cutoff_lsn', but beware of corner cases like if you just
4639 : // changed the GC settings for the tenant to make the PITR window
4640 : // larger, but some of the data was already removed by an earlier GC
4641 : // iteration.
4642 :
4643 : // check against last actual 'latest_gc_cutoff' first
4644 464 : let latest_gc_cutoff_lsn = src_timeline.get_latest_gc_cutoff_lsn();
4645 464 : src_timeline
4646 464 : .check_lsn_is_in_scope(start_lsn, &latest_gc_cutoff_lsn)
4647 464 : .context(format!(
4648 464 : "invalid branch start lsn: less than latest GC cutoff {}",
4649 464 : *latest_gc_cutoff_lsn,
4650 464 : ))
4651 464 : .map_err(CreateTimelineError::AncestorLsn)?;
4652 :
4653 : // and then the planned GC cutoff
4654 : {
4655 456 : let gc_info = src_timeline.gc_info.read().unwrap();
4656 456 : let cutoff = gc_info.min_cutoff();
4657 456 : if start_lsn < cutoff {
4658 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
4659 0 : "invalid branch start lsn: less than planned GC cutoff {cutoff}"
4660 0 : )));
4661 456 : }
4662 456 : }
4663 456 :
4664 456 : //
4665 456 : // The branch point is valid, and we are still holding the 'gc_cs' lock
4666 456 : // so that GC cannot advance the GC cutoff until we are finished.
4667 456 : // Proceed with the branch creation.
4668 456 : //
4669 456 :
4670 456 : // Determine prev-LSN for the new timeline. We can only determine it if
4671 456 : // the timeline was branched at the current end of the source timeline.
4672 456 : let RecordLsn {
4673 456 : last: src_last,
4674 456 : prev: src_prev,
4675 456 : } = src_timeline.get_last_record_rlsn();
4676 456 : let dst_prev = if src_last == start_lsn {
4677 432 : Some(src_prev)
4678 : } else {
4679 24 : None
4680 : };
4681 :
4682 : // Create the metadata file, noting the ancestor of the new timeline.
4683 : // There is initially no data in it, but all the read-calls know to look
4684 : // into the ancestor.
4685 456 : let metadata = TimelineMetadata::new(
4686 456 : start_lsn,
4687 456 : dst_prev,
4688 456 : Some(src_id),
4689 456 : start_lsn,
4690 456 : *src_timeline.latest_gc_cutoff_lsn.read(), // FIXME: should we hold onto this guard longer?
4691 456 : src_timeline.initdb_lsn,
4692 456 : src_timeline.pg_version,
4693 456 : );
4694 :
4695 456 : let uninitialized_timeline = self
4696 456 : .prepare_new_timeline(
4697 456 : dst_id,
4698 456 : &metadata,
4699 456 : timeline_create_guard,
4700 456 : start_lsn + 1,
4701 456 : Some(Arc::clone(src_timeline)),
4702 456 : )
4703 456 : .await?;
4704 :
4705 456 : let new_timeline = uninitialized_timeline.finish_creation()?;
4706 :
4707 : // Root timeline gets its layers during creation and uploads them along with the metadata.
4708 : // A branch timeline though, when created, can get no writes for some time, hence won't get any layers created.
4709 : // We still need to upload its metadata eagerly: if other nodes `attach` the tenant and miss this timeline, their GC
4710 : // could get incorrect information and remove more layers, than needed.
4711 : // See also https://github.com/neondatabase/neon/issues/3865
4712 456 : new_timeline
4713 456 : .remote_client
4714 456 : .schedule_index_upload_for_full_metadata_update(&metadata)
4715 456 : .context("branch initial metadata upload")?;
4716 :
4717 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
4718 :
4719 456 : Ok(CreateTimelineResult::Created(new_timeline))
4720 464 : }
4721 :
4722 : /// For unit tests, make this visible so that other modules can directly create timelines
4723 : #[cfg(test)]
4724 : #[tracing::instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
4725 : pub(crate) async fn bootstrap_timeline_test(
4726 : self: &Arc<Self>,
4727 : timeline_id: TimelineId,
4728 : pg_version: u32,
4729 : load_existing_initdb: Option<TimelineId>,
4730 : ctx: &RequestContext,
4731 : ) -> anyhow::Result<Arc<Timeline>> {
4732 : self.bootstrap_timeline(timeline_id, pg_version, load_existing_initdb, ctx)
4733 : .await
4734 : .map_err(anyhow::Error::new)
4735 4 : .map(|r| r.into_timeline_for_test())
4736 : }
4737 :
4738 : /// Get exclusive access to the timeline ID for creation.
4739 : ///
4740 : /// Timeline-creating code paths must use this function before making changes
4741 : /// to in-memory or persistent state.
4742 : ///
4743 : /// The `state` parameter is a description of the timeline creation operation
4744 : /// we intend to perform.
4745 : /// If the timeline was already created in the meantime, we check whether this
4746 : /// request conflicts or is idempotent , based on `state`.
4747 892 : async fn start_creating_timeline(
4748 892 : self: &Arc<Self>,
4749 892 : new_timeline_id: TimelineId,
4750 892 : idempotency: CreateTimelineIdempotency,
4751 892 : ) -> Result<StartCreatingTimelineResult, CreateTimelineError> {
4752 892 : let allow_offloaded = false;
4753 892 : match self.create_timeline_create_guard(new_timeline_id, idempotency, allow_offloaded) {
4754 888 : Ok(create_guard) => {
4755 888 : pausable_failpoint!("timeline-creation-after-uninit");
4756 888 : Ok(StartCreatingTimelineResult::CreateGuard(create_guard))
4757 : }
4758 0 : Err(TimelineExclusionError::ShuttingDown) => Err(CreateTimelineError::ShuttingDown),
4759 : Err(TimelineExclusionError::AlreadyCreating) => {
4760 : // Creation is in progress, we cannot create it again, and we cannot
4761 : // check if this request matches the existing one, so caller must try
4762 : // again later.
4763 0 : Err(CreateTimelineError::AlreadyCreating)
4764 : }
4765 0 : Err(TimelineExclusionError::Other(e)) => Err(CreateTimelineError::Other(e)),
4766 : Err(TimelineExclusionError::AlreadyExists {
4767 0 : existing: TimelineOrOffloaded::Offloaded(_existing),
4768 0 : ..
4769 0 : }) => {
4770 0 : info!("timeline already exists but is offloaded");
4771 0 : Err(CreateTimelineError::Conflict)
4772 : }
4773 : Err(TimelineExclusionError::AlreadyExists {
4774 4 : existing: TimelineOrOffloaded::Timeline(existing),
4775 4 : arg,
4776 4 : }) => {
4777 4 : {
4778 4 : let existing = &existing.create_idempotency;
4779 4 : let _span = info_span!("idempotency_check", ?existing, ?arg).entered();
4780 4 : debug!("timeline already exists");
4781 :
4782 4 : match (existing, &arg) {
4783 : // FailWithConflict => no idempotency check
4784 : (CreateTimelineIdempotency::FailWithConflict, _)
4785 : | (_, CreateTimelineIdempotency::FailWithConflict) => {
4786 4 : warn!("timeline already exists, failing request");
4787 4 : return Err(CreateTimelineError::Conflict);
4788 : }
4789 : // Idempotent <=> CreateTimelineIdempotency is identical
4790 0 : (x, y) if x == y => {
4791 0 : info!("timeline already exists and idempotency matches, succeeding request");
4792 : // fallthrough
4793 : }
4794 : (_, _) => {
4795 0 : warn!("idempotency conflict, failing request");
4796 0 : return Err(CreateTimelineError::Conflict);
4797 : }
4798 : }
4799 : }
4800 :
4801 0 : Ok(StartCreatingTimelineResult::Idempotent(existing))
4802 : }
4803 : }
4804 892 : }
4805 :
4806 0 : async fn upload_initdb(
4807 0 : &self,
4808 0 : timelines_path: &Utf8PathBuf,
4809 0 : pgdata_path: &Utf8PathBuf,
4810 0 : timeline_id: &TimelineId,
4811 0 : ) -> anyhow::Result<()> {
4812 0 : let temp_path = timelines_path.join(format!(
4813 0 : "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
4814 0 : ));
4815 0 :
4816 0 : scopeguard::defer! {
4817 0 : if let Err(e) = fs::remove_file(&temp_path) {
4818 0 : error!("Failed to remove temporary initdb archive '{temp_path}': {e}");
4819 0 : }
4820 0 : }
4821 :
4822 0 : let (pgdata_zstd, tar_zst_size) = create_zst_tarball(pgdata_path, &temp_path).await?;
4823 : const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
4824 0 : if tar_zst_size > INITDB_TAR_ZST_WARN_LIMIT {
4825 0 : warn!(
4826 0 : "compressed {temp_path} size of {tar_zst_size} is above limit {INITDB_TAR_ZST_WARN_LIMIT}."
4827 : );
4828 0 : }
4829 :
4830 0 : pausable_failpoint!("before-initdb-upload");
4831 :
4832 0 : backoff::retry(
4833 0 : || async {
4834 0 : self::remote_timeline_client::upload_initdb_dir(
4835 0 : &self.remote_storage,
4836 0 : &self.tenant_shard_id.tenant_id,
4837 0 : timeline_id,
4838 0 : pgdata_zstd.try_clone().await?,
4839 0 : tar_zst_size,
4840 0 : &self.cancel,
4841 0 : )
4842 0 : .await
4843 0 : },
4844 0 : |_| false,
4845 0 : 3,
4846 0 : u32::MAX,
4847 0 : "persist_initdb_tar_zst",
4848 0 : &self.cancel,
4849 0 : )
4850 0 : .await
4851 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
4852 0 : .and_then(|x| x)
4853 0 : }
4854 :
4855 : /// - run initdb to init temporary instance and get bootstrap data
4856 : /// - after initialization completes, tar up the temp dir and upload it to S3.
4857 4 : async fn bootstrap_timeline(
4858 4 : self: &Arc<Self>,
4859 4 : timeline_id: TimelineId,
4860 4 : pg_version: u32,
4861 4 : load_existing_initdb: Option<TimelineId>,
4862 4 : ctx: &RequestContext,
4863 4 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4864 4 : let timeline_create_guard = match self
4865 4 : .start_creating_timeline(
4866 4 : timeline_id,
4867 4 : CreateTimelineIdempotency::Bootstrap { pg_version },
4868 4 : )
4869 4 : .await?
4870 : {
4871 4 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
4872 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
4873 0 : return Ok(CreateTimelineResult::Idempotent(timeline))
4874 : }
4875 : };
4876 :
4877 : // create a `tenant/{tenant_id}/timelines/basebackup-{timeline_id}.{TEMP_FILE_SUFFIX}/`
4878 : // temporary directory for basebackup files for the given timeline.
4879 :
4880 4 : let timelines_path = self.conf.timelines_path(&self.tenant_shard_id);
4881 4 : let pgdata_path = path_with_suffix_extension(
4882 4 : timelines_path.join(format!("basebackup-{timeline_id}")),
4883 4 : TEMP_FILE_SUFFIX,
4884 4 : );
4885 4 :
4886 4 : // Remove whatever was left from the previous runs: safe because TimelineCreateGuard guarantees
4887 4 : // we won't race with other creations or existent timelines with the same path.
4888 4 : if pgdata_path.exists() {
4889 0 : fs::remove_dir_all(&pgdata_path).with_context(|| {
4890 0 : format!("Failed to remove already existing initdb directory: {pgdata_path}")
4891 0 : })?;
4892 4 : }
4893 :
4894 : // this new directory is very temporary, set to remove it immediately after bootstrap, we don't need it
4895 4 : scopeguard::defer! {
4896 4 : if let Err(e) = fs::remove_dir_all(&pgdata_path) {
4897 4 : // this is unlikely, but we will remove the directory on pageserver restart or another bootstrap call
4898 4 : error!("Failed to remove temporary initdb directory '{pgdata_path}': {e}");
4899 4 : }
4900 4 : }
4901 4 : if let Some(existing_initdb_timeline_id) = load_existing_initdb {
4902 4 : if existing_initdb_timeline_id != timeline_id {
4903 0 : let source_path = &remote_initdb_archive_path(
4904 0 : &self.tenant_shard_id.tenant_id,
4905 0 : &existing_initdb_timeline_id,
4906 0 : );
4907 0 : let dest_path =
4908 0 : &remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &timeline_id);
4909 0 :
4910 0 : // if this fails, it will get retried by retried control plane requests
4911 0 : self.remote_storage
4912 0 : .copy_object(source_path, dest_path, &self.cancel)
4913 0 : .await
4914 0 : .context("copy initdb tar")?;
4915 4 : }
4916 4 : let (initdb_tar_zst_path, initdb_tar_zst) =
4917 4 : self::remote_timeline_client::download_initdb_tar_zst(
4918 4 : self.conf,
4919 4 : &self.remote_storage,
4920 4 : &self.tenant_shard_id,
4921 4 : &existing_initdb_timeline_id,
4922 4 : &self.cancel,
4923 4 : )
4924 4 : .await
4925 4 : .context("download initdb tar")?;
4926 :
4927 4 : scopeguard::defer! {
4928 4 : if let Err(e) = fs::remove_file(&initdb_tar_zst_path) {
4929 4 : error!("Failed to remove temporary initdb archive '{initdb_tar_zst_path}': {e}");
4930 4 : }
4931 4 : }
4932 4 :
4933 4 : let buf_read =
4934 4 : BufReader::with_capacity(remote_timeline_client::BUFFER_SIZE, initdb_tar_zst);
4935 4 : extract_zst_tarball(&pgdata_path, buf_read)
4936 4 : .await
4937 4 : .context("extract initdb tar")?;
4938 : } else {
4939 : // Init temporarily repo to get bootstrap data, this creates a directory in the `pgdata_path` path
4940 0 : run_initdb(self.conf, &pgdata_path, pg_version, &self.cancel)
4941 0 : .await
4942 0 : .context("run initdb")?;
4943 :
4944 : // Upload the created data dir to S3
4945 0 : if self.tenant_shard_id().is_shard_zero() {
4946 0 : self.upload_initdb(&timelines_path, &pgdata_path, &timeline_id)
4947 0 : .await?;
4948 0 : }
4949 : }
4950 4 : let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
4951 4 :
4952 4 : // Import the contents of the data directory at the initial checkpoint
4953 4 : // LSN, and any WAL after that.
4954 4 : // Initdb lsn will be equal to last_record_lsn which will be set after import.
4955 4 : // Because we know it upfront avoid having an option or dummy zero value by passing it to the metadata.
4956 4 : let new_metadata = TimelineMetadata::new(
4957 4 : Lsn(0),
4958 4 : None,
4959 4 : None,
4960 4 : Lsn(0),
4961 4 : pgdata_lsn,
4962 4 : pgdata_lsn,
4963 4 : pg_version,
4964 4 : );
4965 4 : let raw_timeline = self
4966 4 : .prepare_new_timeline(
4967 4 : timeline_id,
4968 4 : &new_metadata,
4969 4 : timeline_create_guard,
4970 4 : pgdata_lsn,
4971 4 : None,
4972 4 : )
4973 4 : .await?;
4974 :
4975 4 : let tenant_shard_id = raw_timeline.owning_tenant.tenant_shard_id;
4976 4 : let unfinished_timeline = raw_timeline.raw_timeline()?;
4977 :
4978 : // Flush the new layer files to disk, before we make the timeline as available to
4979 : // the outside world.
4980 : //
4981 : // Flush loop needs to be spawned in order to be able to flush.
4982 4 : unfinished_timeline.maybe_spawn_flush_loop();
4983 4 :
4984 4 : import_datadir::import_timeline_from_postgres_datadir(
4985 4 : unfinished_timeline,
4986 4 : &pgdata_path,
4987 4 : pgdata_lsn,
4988 4 : ctx,
4989 4 : )
4990 4 : .await
4991 4 : .with_context(|| {
4992 0 : format!("Failed to import pgdatadir for timeline {tenant_shard_id}/{timeline_id}")
4993 4 : })?;
4994 :
4995 4 : fail::fail_point!("before-checkpoint-new-timeline", |_| {
4996 0 : Err(CreateTimelineError::Other(anyhow::anyhow!(
4997 0 : "failpoint before-checkpoint-new-timeline"
4998 0 : )))
4999 4 : });
5000 :
5001 4 : unfinished_timeline
5002 4 : .freeze_and_flush()
5003 4 : .await
5004 4 : .with_context(|| {
5005 0 : format!(
5006 0 : "Failed to flush after pgdatadir import for timeline {tenant_shard_id}/{timeline_id}"
5007 0 : )
5008 4 : })?;
5009 :
5010 : // All done!
5011 4 : let timeline = raw_timeline.finish_creation()?;
5012 :
5013 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
5014 :
5015 4 : Ok(CreateTimelineResult::Created(timeline))
5016 4 : }
5017 :
5018 880 : fn build_timeline_remote_client(&self, timeline_id: TimelineId) -> RemoteTimelineClient {
5019 880 : RemoteTimelineClient::new(
5020 880 : self.remote_storage.clone(),
5021 880 : self.deletion_queue_client.clone(),
5022 880 : self.conf,
5023 880 : self.tenant_shard_id,
5024 880 : timeline_id,
5025 880 : self.generation,
5026 880 : &self.tenant_conf.load().location,
5027 880 : )
5028 880 : }
5029 :
5030 : /// Call this before constructing a timeline, to build its required structures
5031 880 : fn build_timeline_resources(&self, timeline_id: TimelineId) -> TimelineResources {
5032 880 : TimelineResources {
5033 880 : remote_client: self.build_timeline_remote_client(timeline_id),
5034 880 : pagestream_throttle: self.pagestream_throttle.clone(),
5035 880 : pagestream_throttle_metrics: self.pagestream_throttle_metrics.clone(),
5036 880 : l0_flush_global_state: self.l0_flush_global_state.clone(),
5037 880 : }
5038 880 : }
5039 :
5040 : /// Creates intermediate timeline structure and its files.
5041 : ///
5042 : /// An empty layer map is initialized, and new data and WAL can be imported starting
5043 : /// at 'disk_consistent_lsn'. After any initial data has been imported, call
5044 : /// `finish_creation` to insert the Timeline into the timelines map.
5045 880 : async fn prepare_new_timeline<'a>(
5046 880 : &'a self,
5047 880 : new_timeline_id: TimelineId,
5048 880 : new_metadata: &TimelineMetadata,
5049 880 : create_guard: TimelineCreateGuard,
5050 880 : start_lsn: Lsn,
5051 880 : ancestor: Option<Arc<Timeline>>,
5052 880 : ) -> anyhow::Result<UninitializedTimeline<'a>> {
5053 880 : let tenant_shard_id = self.tenant_shard_id;
5054 880 :
5055 880 : let resources = self.build_timeline_resources(new_timeline_id);
5056 880 : resources
5057 880 : .remote_client
5058 880 : .init_upload_queue_for_empty_remote(new_metadata)?;
5059 :
5060 880 : let timeline_struct = self
5061 880 : .create_timeline_struct(
5062 880 : new_timeline_id,
5063 880 : new_metadata,
5064 880 : ancestor,
5065 880 : resources,
5066 880 : CreateTimelineCause::Load,
5067 880 : create_guard.idempotency.clone(),
5068 880 : )
5069 880 : .context("Failed to create timeline data structure")?;
5070 :
5071 880 : timeline_struct.init_empty_layer_map(start_lsn);
5072 :
5073 880 : if let Err(e) = self
5074 880 : .create_timeline_files(&create_guard.timeline_path)
5075 880 : .await
5076 : {
5077 0 : error!("Failed to create initial files for timeline {tenant_shard_id}/{new_timeline_id}, cleaning up: {e:?}");
5078 0 : cleanup_timeline_directory(create_guard);
5079 0 : return Err(e);
5080 880 : }
5081 880 :
5082 880 : debug!(
5083 0 : "Successfully created initial files for timeline {tenant_shard_id}/{new_timeline_id}"
5084 : );
5085 :
5086 880 : Ok(UninitializedTimeline::new(
5087 880 : self,
5088 880 : new_timeline_id,
5089 880 : Some((timeline_struct, create_guard)),
5090 880 : ))
5091 880 : }
5092 :
5093 880 : async fn create_timeline_files(&self, timeline_path: &Utf8Path) -> anyhow::Result<()> {
5094 880 : crashsafe::create_dir(timeline_path).context("Failed to create timeline directory")?;
5095 :
5096 880 : fail::fail_point!("after-timeline-dir-creation", |_| {
5097 0 : anyhow::bail!("failpoint after-timeline-dir-creation");
5098 880 : });
5099 :
5100 880 : Ok(())
5101 880 : }
5102 :
5103 : /// Get a guard that provides exclusive access to the timeline directory, preventing
5104 : /// concurrent attempts to create the same timeline.
5105 : ///
5106 : /// The `allow_offloaded` parameter controls whether to tolerate the existence of
5107 : /// offloaded timelines or not.
5108 892 : fn create_timeline_create_guard(
5109 892 : self: &Arc<Self>,
5110 892 : timeline_id: TimelineId,
5111 892 : idempotency: CreateTimelineIdempotency,
5112 892 : allow_offloaded: bool,
5113 892 : ) -> Result<TimelineCreateGuard, TimelineExclusionError> {
5114 892 : let tenant_shard_id = self.tenant_shard_id;
5115 892 :
5116 892 : let timeline_path = self.conf.timeline_path(&tenant_shard_id, &timeline_id);
5117 :
5118 892 : let create_guard = TimelineCreateGuard::new(
5119 892 : self,
5120 892 : timeline_id,
5121 892 : timeline_path.clone(),
5122 892 : idempotency,
5123 892 : allow_offloaded,
5124 892 : )?;
5125 :
5126 : // At this stage, we have got exclusive access to in-memory state for this timeline ID
5127 : // for creation.
5128 : // A timeline directory should never exist on disk already:
5129 : // - a previous failed creation would have cleaned up after itself
5130 : // - a pageserver restart would clean up timeline directories that don't have valid remote state
5131 : //
5132 : // Therefore it is an unexpected internal error to encounter a timeline directory already existing here,
5133 : // this error may indicate a bug in cleanup on failed creations.
5134 888 : if timeline_path.exists() {
5135 0 : return Err(TimelineExclusionError::Other(anyhow::anyhow!(
5136 0 : "Timeline directory already exists! This is a bug."
5137 0 : )));
5138 888 : }
5139 888 :
5140 888 : Ok(create_guard)
5141 892 : }
5142 :
5143 : /// Gathers inputs from all of the timelines to produce a sizing model input.
5144 : ///
5145 : /// Future is cancellation safe. Only one calculation can be running at once per tenant.
5146 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5147 : pub async fn gather_size_inputs(
5148 : &self,
5149 : // `max_retention_period` overrides the cutoff that is used to calculate the size
5150 : // (only if it is shorter than the real cutoff).
5151 : max_retention_period: Option<u64>,
5152 : cause: LogicalSizeCalculationCause,
5153 : cancel: &CancellationToken,
5154 : ctx: &RequestContext,
5155 : ) -> Result<size::ModelInputs, size::CalculateSyntheticSizeError> {
5156 : let logical_sizes_at_once = self
5157 : .conf
5158 : .concurrent_tenant_size_logical_size_queries
5159 : .inner();
5160 :
5161 : // TODO: Having a single mutex block concurrent reads is not great for performance.
5162 : //
5163 : // But the only case where we need to run multiple of these at once is when we
5164 : // request a size for a tenant manually via API, while another background calculation
5165 : // is in progress (which is not a common case).
5166 : //
5167 : // See more for on the issue #2748 condenced out of the initial PR review.
5168 : let mut shared_cache = tokio::select! {
5169 : locked = self.cached_logical_sizes.lock() => locked,
5170 : _ = cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5171 : _ = self.cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5172 : };
5173 :
5174 : size::gather_inputs(
5175 : self,
5176 : logical_sizes_at_once,
5177 : max_retention_period,
5178 : &mut shared_cache,
5179 : cause,
5180 : cancel,
5181 : ctx,
5182 : )
5183 : .await
5184 : }
5185 :
5186 : /// Calculate synthetic tenant size and cache the result.
5187 : /// This is periodically called by background worker.
5188 : /// result is cached in tenant struct
5189 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5190 : pub async fn calculate_synthetic_size(
5191 : &self,
5192 : cause: LogicalSizeCalculationCause,
5193 : cancel: &CancellationToken,
5194 : ctx: &RequestContext,
5195 : ) -> Result<u64, size::CalculateSyntheticSizeError> {
5196 : let inputs = self.gather_size_inputs(None, cause, cancel, ctx).await?;
5197 :
5198 : let size = inputs.calculate();
5199 :
5200 : self.set_cached_synthetic_size(size);
5201 :
5202 : Ok(size)
5203 : }
5204 :
5205 : /// Cache given synthetic size and update the metric value
5206 0 : pub fn set_cached_synthetic_size(&self, size: u64) {
5207 0 : self.cached_synthetic_tenant_size
5208 0 : .store(size, Ordering::Relaxed);
5209 0 :
5210 0 : // Only shard zero should be calculating synthetic sizes
5211 0 : debug_assert!(self.shard_identity.is_shard_zero());
5212 :
5213 0 : TENANT_SYNTHETIC_SIZE_METRIC
5214 0 : .get_metric_with_label_values(&[&self.tenant_shard_id.tenant_id.to_string()])
5215 0 : .unwrap()
5216 0 : .set(size);
5217 0 : }
5218 :
5219 0 : pub fn cached_synthetic_size(&self) -> u64 {
5220 0 : self.cached_synthetic_tenant_size.load(Ordering::Relaxed)
5221 0 : }
5222 :
5223 : /// Flush any in-progress layers, schedule uploads, and wait for uploads to complete.
5224 : ///
5225 : /// This function can take a long time: callers should wrap it in a timeout if calling
5226 : /// from an external API handler.
5227 : ///
5228 : /// Cancel-safety: cancelling this function may leave I/O running, but such I/O is
5229 : /// still bounded by tenant/timeline shutdown.
5230 : #[tracing::instrument(skip_all)]
5231 : pub(crate) async fn flush_remote(&self) -> anyhow::Result<()> {
5232 : let timelines = self.timelines.lock().unwrap().clone();
5233 :
5234 0 : async fn flush_timeline(_gate: GateGuard, timeline: Arc<Timeline>) -> anyhow::Result<()> {
5235 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Flushing...");
5236 0 : timeline.freeze_and_flush().await?;
5237 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Waiting for uploads...");
5238 0 : timeline.remote_client.wait_completion().await?;
5239 :
5240 0 : Ok(())
5241 0 : }
5242 :
5243 : // We do not use a JoinSet for these tasks, because we don't want them to be
5244 : // aborted when this function's future is cancelled: they should stay alive
5245 : // holding their GateGuard until they complete, to ensure their I/Os complete
5246 : // before Timeline shutdown completes.
5247 : let mut results = FuturesUnordered::new();
5248 :
5249 : for (_timeline_id, timeline) in timelines {
5250 : // Run each timeline's flush in a task holding the timeline's gate: this
5251 : // means that if this function's future is cancelled, the Timeline shutdown
5252 : // will still wait for any I/O in here to complete.
5253 : let Ok(gate) = timeline.gate.enter() else {
5254 : continue;
5255 : };
5256 0 : let jh = tokio::task::spawn(async move { flush_timeline(gate, timeline).await });
5257 : results.push(jh);
5258 : }
5259 :
5260 : while let Some(r) = results.next().await {
5261 : if let Err(e) = r {
5262 : if !e.is_cancelled() && !e.is_panic() {
5263 : tracing::error!("unexpected join error: {e:?}");
5264 : }
5265 : }
5266 : }
5267 :
5268 : // The flushes we did above were just writes, but the Tenant might have had
5269 : // pending deletions as well from recent compaction/gc: we want to flush those
5270 : // as well. This requires flushing the global delete queue. This is cheap
5271 : // because it's typically a no-op.
5272 : match self.deletion_queue_client.flush_execute().await {
5273 : Ok(_) => {}
5274 : Err(DeletionQueueError::ShuttingDown) => {}
5275 : }
5276 :
5277 : Ok(())
5278 : }
5279 :
5280 0 : pub(crate) fn get_tenant_conf(&self) -> TenantConfOpt {
5281 0 : self.tenant_conf.load().tenant_conf.clone()
5282 0 : }
5283 :
5284 : /// How much local storage would this tenant like to have? It can cope with
5285 : /// less than this (via eviction and on-demand downloads), but this function enables
5286 : /// the Tenant to advertise how much storage it would prefer to have to provide fast I/O
5287 : /// by keeping important things on local disk.
5288 : ///
5289 : /// This is a heuristic, not a guarantee: tenants that are long-idle will actually use less
5290 : /// than they report here, due to layer eviction. Tenants with many active branches may
5291 : /// actually use more than they report here.
5292 0 : pub(crate) fn local_storage_wanted(&self) -> u64 {
5293 0 : let timelines = self.timelines.lock().unwrap();
5294 0 :
5295 0 : // Heuristic: we use the max() of the timelines' visible sizes, rather than the sum. This
5296 0 : // reflects the observation that on tenants with multiple large branches, typically only one
5297 0 : // of them is used actively enough to occupy space on disk.
5298 0 : timelines
5299 0 : .values()
5300 0 : .map(|t| t.metrics.visible_physical_size_gauge.get())
5301 0 : .max()
5302 0 : .unwrap_or(0)
5303 0 : }
5304 :
5305 : /// Serialize and write the latest TenantManifest to remote storage.
5306 4 : pub(crate) async fn store_tenant_manifest(&self) -> Result<(), TenantManifestError> {
5307 : // Only one manifest write may be done at at time, and the contents of the manifest
5308 : // must be loaded while holding this lock. This makes it safe to call this function
5309 : // from anywhere without worrying about colliding updates.
5310 4 : let mut guard = tokio::select! {
5311 4 : g = self.tenant_manifest_upload.lock() => {
5312 4 : g
5313 : },
5314 4 : _ = self.cancel.cancelled() => {
5315 0 : return Err(TenantManifestError::Cancelled);
5316 : }
5317 : };
5318 :
5319 4 : let manifest = self.build_tenant_manifest();
5320 4 : if Some(&manifest) == (*guard).as_ref() {
5321 : // Optimisation: skip uploads that don't change anything.
5322 0 : return Ok(());
5323 4 : }
5324 4 :
5325 4 : // Remote storage does no retries internally, so wrap it
5326 4 : match backoff::retry(
5327 4 : || async {
5328 4 : upload_tenant_manifest(
5329 4 : &self.remote_storage,
5330 4 : &self.tenant_shard_id,
5331 4 : self.generation,
5332 4 : &manifest,
5333 4 : &self.cancel,
5334 4 : )
5335 4 : .await
5336 8 : },
5337 4 : |_e| self.cancel.is_cancelled(),
5338 4 : FAILED_UPLOAD_WARN_THRESHOLD,
5339 4 : FAILED_REMOTE_OP_RETRIES,
5340 4 : "uploading tenant manifest",
5341 4 : &self.cancel,
5342 4 : )
5343 4 : .await
5344 : {
5345 0 : None => Err(TenantManifestError::Cancelled),
5346 0 : Some(Err(_)) if self.cancel.is_cancelled() => Err(TenantManifestError::Cancelled),
5347 0 : Some(Err(e)) => Err(TenantManifestError::RemoteStorage(e)),
5348 : Some(Ok(_)) => {
5349 : // Store the successfully uploaded manifest, so that future callers can avoid
5350 : // re-uploading the same thing.
5351 4 : *guard = Some(manifest);
5352 4 :
5353 4 : Ok(())
5354 : }
5355 : }
5356 4 : }
5357 : }
5358 :
5359 : /// Create the cluster temporarily in 'initdbpath' directory inside the repository
5360 : /// to get bootstrap data for timeline initialization.
5361 0 : async fn run_initdb(
5362 0 : conf: &'static PageServerConf,
5363 0 : initdb_target_dir: &Utf8Path,
5364 0 : pg_version: u32,
5365 0 : cancel: &CancellationToken,
5366 0 : ) -> Result<(), InitdbError> {
5367 0 : let initdb_bin_path = conf
5368 0 : .pg_bin_dir(pg_version)
5369 0 : .map_err(InitdbError::Other)?
5370 0 : .join("initdb");
5371 0 : let initdb_lib_dir = conf.pg_lib_dir(pg_version).map_err(InitdbError::Other)?;
5372 0 : info!(
5373 0 : "running {} in {}, libdir: {}",
5374 : initdb_bin_path, initdb_target_dir, initdb_lib_dir,
5375 : );
5376 :
5377 0 : let _permit = {
5378 0 : let _timer = INITDB_SEMAPHORE_ACQUISITION_TIME.start_timer();
5379 0 : INIT_DB_SEMAPHORE.acquire().await
5380 : };
5381 :
5382 0 : CONCURRENT_INITDBS.inc();
5383 0 : scopeguard::defer! {
5384 0 : CONCURRENT_INITDBS.dec();
5385 0 : }
5386 0 :
5387 0 : let _timer = INITDB_RUN_TIME.start_timer();
5388 0 : let res = postgres_initdb::do_run_initdb(postgres_initdb::RunInitdbArgs {
5389 0 : superuser: &conf.superuser,
5390 0 : locale: &conf.locale,
5391 0 : initdb_bin: &initdb_bin_path,
5392 0 : pg_version,
5393 0 : library_search_path: &initdb_lib_dir,
5394 0 : pgdata: initdb_target_dir,
5395 0 : })
5396 0 : .await
5397 0 : .map_err(InitdbError::Inner);
5398 0 :
5399 0 : // This isn't true cancellation support, see above. Still return an error to
5400 0 : // excercise the cancellation code path.
5401 0 : if cancel.is_cancelled() {
5402 0 : return Err(InitdbError::Cancelled);
5403 0 : }
5404 0 :
5405 0 : res
5406 0 : }
5407 :
5408 : /// Dump contents of a layer file to stdout.
5409 0 : pub async fn dump_layerfile_from_path(
5410 0 : path: &Utf8Path,
5411 0 : verbose: bool,
5412 0 : ctx: &RequestContext,
5413 0 : ) -> anyhow::Result<()> {
5414 : use std::os::unix::fs::FileExt;
5415 :
5416 : // All layer files start with a two-byte "magic" value, to identify the kind of
5417 : // file.
5418 0 : let file = File::open(path)?;
5419 0 : let mut header_buf = [0u8; 2];
5420 0 : file.read_exact_at(&mut header_buf, 0)?;
5421 :
5422 0 : match u16::from_be_bytes(header_buf) {
5423 : crate::IMAGE_FILE_MAGIC => {
5424 0 : ImageLayer::new_for_path(path, file)?
5425 0 : .dump(verbose, ctx)
5426 0 : .await?
5427 : }
5428 : crate::DELTA_FILE_MAGIC => {
5429 0 : DeltaLayer::new_for_path(path, file)?
5430 0 : .dump(verbose, ctx)
5431 0 : .await?
5432 : }
5433 0 : magic => bail!("unrecognized magic identifier: {:?}", magic),
5434 : }
5435 :
5436 0 : Ok(())
5437 0 : }
5438 :
5439 : #[cfg(test)]
5440 : pub(crate) mod harness {
5441 : use bytes::{Bytes, BytesMut};
5442 : use once_cell::sync::OnceCell;
5443 : use pageserver_api::models::ShardParameters;
5444 : use pageserver_api::shard::ShardIndex;
5445 : use utils::logging;
5446 :
5447 : use crate::deletion_queue::mock::MockDeletionQueue;
5448 : use crate::l0_flush::L0FlushConfig;
5449 : use crate::walredo::apply_neon;
5450 : use pageserver_api::key::Key;
5451 : use pageserver_api::record::NeonWalRecord;
5452 :
5453 : use super::*;
5454 : use hex_literal::hex;
5455 : use utils::id::TenantId;
5456 :
5457 : pub const TIMELINE_ID: TimelineId =
5458 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
5459 : pub const NEW_TIMELINE_ID: TimelineId =
5460 : TimelineId::from_array(hex!("AA223344556677881122334455667788"));
5461 :
5462 : /// Convenience function to create a page image with given string as the only content
5463 10057556 : pub fn test_img(s: &str) -> Bytes {
5464 10057556 : let mut buf = BytesMut::new();
5465 10057556 : buf.extend_from_slice(s.as_bytes());
5466 10057556 : buf.resize(64, 0);
5467 10057556 :
5468 10057556 : buf.freeze()
5469 10057556 : }
5470 :
5471 : impl From<TenantConf> for TenantConfOpt {
5472 440 : fn from(tenant_conf: TenantConf) -> Self {
5473 440 : Self {
5474 440 : checkpoint_distance: Some(tenant_conf.checkpoint_distance),
5475 440 : checkpoint_timeout: Some(tenant_conf.checkpoint_timeout),
5476 440 : compaction_target_size: Some(tenant_conf.compaction_target_size),
5477 440 : compaction_period: Some(tenant_conf.compaction_period),
5478 440 : compaction_threshold: Some(tenant_conf.compaction_threshold),
5479 440 : compaction_upper_limit: Some(tenant_conf.compaction_upper_limit),
5480 440 : compaction_algorithm: Some(tenant_conf.compaction_algorithm),
5481 440 : l0_flush_delay_threshold: tenant_conf.l0_flush_delay_threshold,
5482 440 : l0_flush_stall_threshold: tenant_conf.l0_flush_stall_threshold,
5483 440 : l0_flush_wait_upload: Some(tenant_conf.l0_flush_wait_upload),
5484 440 : gc_horizon: Some(tenant_conf.gc_horizon),
5485 440 : gc_period: Some(tenant_conf.gc_period),
5486 440 : image_creation_threshold: Some(tenant_conf.image_creation_threshold),
5487 440 : pitr_interval: Some(tenant_conf.pitr_interval),
5488 440 : walreceiver_connect_timeout: Some(tenant_conf.walreceiver_connect_timeout),
5489 440 : lagging_wal_timeout: Some(tenant_conf.lagging_wal_timeout),
5490 440 : max_lsn_wal_lag: Some(tenant_conf.max_lsn_wal_lag),
5491 440 : eviction_policy: Some(tenant_conf.eviction_policy),
5492 440 : min_resident_size_override: tenant_conf.min_resident_size_override,
5493 440 : evictions_low_residence_duration_metric_threshold: Some(
5494 440 : tenant_conf.evictions_low_residence_duration_metric_threshold,
5495 440 : ),
5496 440 : heatmap_period: Some(tenant_conf.heatmap_period),
5497 440 : lazy_slru_download: Some(tenant_conf.lazy_slru_download),
5498 440 : timeline_get_throttle: Some(tenant_conf.timeline_get_throttle),
5499 440 : image_layer_creation_check_threshold: Some(
5500 440 : tenant_conf.image_layer_creation_check_threshold,
5501 440 : ),
5502 440 : lsn_lease_length: Some(tenant_conf.lsn_lease_length),
5503 440 : lsn_lease_length_for_ts: Some(tenant_conf.lsn_lease_length_for_ts),
5504 440 : timeline_offloading: Some(tenant_conf.timeline_offloading),
5505 440 : wal_receiver_protocol_override: tenant_conf.wal_receiver_protocol_override,
5506 440 : rel_size_v2_enabled: tenant_conf.rel_size_v2_enabled,
5507 440 : gc_compaction_enabled: Some(tenant_conf.gc_compaction_enabled),
5508 440 : gc_compaction_initial_threshold_kb: Some(
5509 440 : tenant_conf.gc_compaction_initial_threshold_kb,
5510 440 : ),
5511 440 : gc_compaction_ratio_percent: Some(tenant_conf.gc_compaction_ratio_percent),
5512 440 : }
5513 440 : }
5514 : }
5515 :
5516 : pub struct TenantHarness {
5517 : pub conf: &'static PageServerConf,
5518 : pub tenant_conf: TenantConf,
5519 : pub tenant_shard_id: TenantShardId,
5520 : pub generation: Generation,
5521 : pub shard: ShardIndex,
5522 : pub remote_storage: GenericRemoteStorage,
5523 : pub remote_fs_dir: Utf8PathBuf,
5524 : pub deletion_queue: MockDeletionQueue,
5525 : }
5526 :
5527 : static LOG_HANDLE: OnceCell<()> = OnceCell::new();
5528 :
5529 488 : pub(crate) fn setup_logging() {
5530 488 : LOG_HANDLE.get_or_init(|| {
5531 464 : logging::init(
5532 464 : logging::LogFormat::Test,
5533 464 : // enable it in case the tests exercise code paths that use
5534 464 : // debug_assert_current_span_has_tenant_and_timeline_id
5535 464 : logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
5536 464 : logging::Output::Stdout,
5537 464 : )
5538 464 : .expect("Failed to init test logging")
5539 488 : });
5540 488 : }
5541 :
5542 : impl TenantHarness {
5543 440 : pub async fn create_custom(
5544 440 : test_name: &'static str,
5545 440 : tenant_conf: TenantConf,
5546 440 : tenant_id: TenantId,
5547 440 : shard_identity: ShardIdentity,
5548 440 : generation: Generation,
5549 440 : ) -> anyhow::Result<Self> {
5550 440 : setup_logging();
5551 440 :
5552 440 : let repo_dir = PageServerConf::test_repo_dir(test_name);
5553 440 : let _ = fs::remove_dir_all(&repo_dir);
5554 440 : fs::create_dir_all(&repo_dir)?;
5555 :
5556 440 : let conf = PageServerConf::dummy_conf(repo_dir);
5557 440 : // Make a static copy of the config. This can never be free'd, but that's
5558 440 : // OK in a test.
5559 440 : let conf: &'static PageServerConf = Box::leak(Box::new(conf));
5560 440 :
5561 440 : let shard = shard_identity.shard_index();
5562 440 : let tenant_shard_id = TenantShardId {
5563 440 : tenant_id,
5564 440 : shard_number: shard.shard_number,
5565 440 : shard_count: shard.shard_count,
5566 440 : };
5567 440 : fs::create_dir_all(conf.tenant_path(&tenant_shard_id))?;
5568 440 : fs::create_dir_all(conf.timelines_path(&tenant_shard_id))?;
5569 :
5570 : use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
5571 440 : let remote_fs_dir = conf.workdir.join("localfs");
5572 440 : std::fs::create_dir_all(&remote_fs_dir).unwrap();
5573 440 : let config = RemoteStorageConfig {
5574 440 : storage: RemoteStorageKind::LocalFs {
5575 440 : local_path: remote_fs_dir.clone(),
5576 440 : },
5577 440 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
5578 440 : small_timeout: RemoteStorageConfig::DEFAULT_SMALL_TIMEOUT,
5579 440 : };
5580 440 : let remote_storage = GenericRemoteStorage::from_config(&config).await.unwrap();
5581 440 : let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()));
5582 440 :
5583 440 : Ok(Self {
5584 440 : conf,
5585 440 : tenant_conf,
5586 440 : tenant_shard_id,
5587 440 : generation,
5588 440 : shard,
5589 440 : remote_storage,
5590 440 : remote_fs_dir,
5591 440 : deletion_queue,
5592 440 : })
5593 440 : }
5594 :
5595 416 : pub async fn create(test_name: &'static str) -> anyhow::Result<Self> {
5596 416 : // Disable automatic GC and compaction to make the unit tests more deterministic.
5597 416 : // The tests perform them manually if needed.
5598 416 : let tenant_conf = TenantConf {
5599 416 : gc_period: Duration::ZERO,
5600 416 : compaction_period: Duration::ZERO,
5601 416 : ..TenantConf::default()
5602 416 : };
5603 416 : let tenant_id = TenantId::generate();
5604 416 : let shard = ShardIdentity::unsharded();
5605 416 : Self::create_custom(
5606 416 : test_name,
5607 416 : tenant_conf,
5608 416 : tenant_id,
5609 416 : shard,
5610 416 : Generation::new(0xdeadbeef),
5611 416 : )
5612 416 : .await
5613 416 : }
5614 :
5615 40 : pub fn span(&self) -> tracing::Span {
5616 40 : info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
5617 40 : }
5618 :
5619 440 : pub(crate) async fn load(&self) -> (Arc<Tenant>, RequestContext) {
5620 440 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
5621 440 : (
5622 440 : self.do_try_load(&ctx)
5623 440 : .await
5624 440 : .expect("failed to load test tenant"),
5625 440 : ctx,
5626 440 : )
5627 440 : }
5628 :
5629 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5630 : pub(crate) async fn do_try_load(
5631 : &self,
5632 : ctx: &RequestContext,
5633 : ) -> anyhow::Result<Arc<Tenant>> {
5634 : let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
5635 :
5636 : let tenant = Arc::new(Tenant::new(
5637 : TenantState::Attaching,
5638 : self.conf,
5639 : AttachedTenantConf::try_from(LocationConf::attached_single(
5640 : TenantConfOpt::from(self.tenant_conf.clone()),
5641 : self.generation,
5642 : &ShardParameters::default(),
5643 : ))
5644 : .unwrap(),
5645 : // This is a legacy/test code path: sharding isn't supported here.
5646 : ShardIdentity::unsharded(),
5647 : Some(walredo_mgr),
5648 : self.tenant_shard_id,
5649 : self.remote_storage.clone(),
5650 : self.deletion_queue.new_client(),
5651 : // TODO: ideally we should run all unit tests with both configs
5652 : L0FlushGlobalState::new(L0FlushConfig::default()),
5653 : ));
5654 :
5655 : let preload = tenant
5656 : .preload(&self.remote_storage, CancellationToken::new())
5657 : .await?;
5658 : tenant.attach(Some(preload), ctx).await?;
5659 :
5660 : tenant.state.send_replace(TenantState::Active);
5661 : for timeline in tenant.timelines.lock().unwrap().values() {
5662 : timeline.set_state(TimelineState::Active);
5663 : }
5664 : Ok(tenant)
5665 : }
5666 :
5667 4 : pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
5668 4 : self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
5669 4 : }
5670 : }
5671 :
5672 : // Mock WAL redo manager that doesn't do much
5673 : pub(crate) struct TestRedoManager;
5674 :
5675 : impl TestRedoManager {
5676 : /// # Cancel-Safety
5677 : ///
5678 : /// This method is cancellation-safe.
5679 1636 : pub async fn request_redo(
5680 1636 : &self,
5681 1636 : key: Key,
5682 1636 : lsn: Lsn,
5683 1636 : base_img: Option<(Lsn, Bytes)>,
5684 1636 : records: Vec<(Lsn, NeonWalRecord)>,
5685 1636 : _pg_version: u32,
5686 1636 : ) -> Result<Bytes, walredo::Error> {
5687 2392 : let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
5688 1636 : if records_neon {
5689 : // For Neon wal records, we can decode without spawning postgres, so do so.
5690 1636 : let mut page = match (base_img, records.first()) {
5691 1504 : (Some((_lsn, img)), _) => {
5692 1504 : let mut page = BytesMut::new();
5693 1504 : page.extend_from_slice(&img);
5694 1504 : page
5695 : }
5696 132 : (_, Some((_lsn, rec))) if rec.will_init() => BytesMut::new(),
5697 : _ => {
5698 0 : panic!("Neon WAL redo requires base image or will init record");
5699 : }
5700 : };
5701 :
5702 4028 : for (record_lsn, record) in records {
5703 2392 : apply_neon::apply_in_neon(&record, record_lsn, key, &mut page)?;
5704 : }
5705 1636 : Ok(page.freeze())
5706 : } else {
5707 : // We never spawn a postgres walredo process in unit tests: just log what we might have done.
5708 0 : let s = format!(
5709 0 : "redo for {} to get to {}, with {} and {} records",
5710 0 : key,
5711 0 : lsn,
5712 0 : if base_img.is_some() {
5713 0 : "base image"
5714 : } else {
5715 0 : "no base image"
5716 : },
5717 0 : records.len()
5718 0 : );
5719 0 : println!("{s}");
5720 0 :
5721 0 : Ok(test_img(&s))
5722 : }
5723 1636 : }
5724 : }
5725 : }
5726 :
5727 : #[cfg(test)]
5728 : mod tests {
5729 : use std::collections::{BTreeMap, BTreeSet};
5730 :
5731 : use super::*;
5732 : use crate::keyspace::KeySpaceAccum;
5733 : use crate::tenant::harness::*;
5734 : use crate::tenant::timeline::CompactFlags;
5735 : use crate::DEFAULT_PG_VERSION;
5736 : use bytes::{Bytes, BytesMut};
5737 : use hex_literal::hex;
5738 : use itertools::Itertools;
5739 : use pageserver_api::key::{Key, AUX_KEY_PREFIX, NON_INHERITED_RANGE, RELATION_SIZE_PREFIX};
5740 : use pageserver_api::keyspace::KeySpace;
5741 : use pageserver_api::models::{CompactionAlgorithm, CompactionAlgorithmSettings};
5742 : use pageserver_api::value::Value;
5743 : use pageserver_compaction::helpers::overlaps_with;
5744 : use rand::{thread_rng, Rng};
5745 : use storage_layer::{IoConcurrency, PersistentLayerKey};
5746 : use tests::storage_layer::ValuesReconstructState;
5747 : use tests::timeline::{GetVectoredError, ShutdownMode};
5748 : use timeline::{CompactOptions, DeltaLayerTestDesc};
5749 : use utils::id::TenantId;
5750 :
5751 : #[cfg(feature = "testing")]
5752 : use models::CompactLsnRange;
5753 : #[cfg(feature = "testing")]
5754 : use pageserver_api::record::NeonWalRecord;
5755 : #[cfg(feature = "testing")]
5756 : use timeline::compaction::{KeyHistoryRetention, KeyLogAtLsn};
5757 : #[cfg(feature = "testing")]
5758 : use timeline::GcInfo;
5759 :
5760 : static TEST_KEY: Lazy<Key> =
5761 36 : Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));
5762 :
5763 : #[tokio::test]
5764 4 : async fn test_basic() -> anyhow::Result<()> {
5765 4 : let (tenant, ctx) = TenantHarness::create("test_basic").await?.load().await;
5766 4 : let tline = tenant
5767 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5768 4 : .await?;
5769 4 :
5770 4 : let mut writer = tline.writer().await;
5771 4 : writer
5772 4 : .put(
5773 4 : *TEST_KEY,
5774 4 : Lsn(0x10),
5775 4 : &Value::Image(test_img("foo at 0x10")),
5776 4 : &ctx,
5777 4 : )
5778 4 : .await?;
5779 4 : writer.finish_write(Lsn(0x10));
5780 4 : drop(writer);
5781 4 :
5782 4 : let mut writer = tline.writer().await;
5783 4 : writer
5784 4 : .put(
5785 4 : *TEST_KEY,
5786 4 : Lsn(0x20),
5787 4 : &Value::Image(test_img("foo at 0x20")),
5788 4 : &ctx,
5789 4 : )
5790 4 : .await?;
5791 4 : writer.finish_write(Lsn(0x20));
5792 4 : drop(writer);
5793 4 :
5794 4 : assert_eq!(
5795 4 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
5796 4 : test_img("foo at 0x10")
5797 4 : );
5798 4 : assert_eq!(
5799 4 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
5800 4 : test_img("foo at 0x10")
5801 4 : );
5802 4 : assert_eq!(
5803 4 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
5804 4 : test_img("foo at 0x20")
5805 4 : );
5806 4 :
5807 4 : Ok(())
5808 4 : }
5809 :
5810 : #[tokio::test]
5811 4 : async fn no_duplicate_timelines() -> anyhow::Result<()> {
5812 4 : let (tenant, ctx) = TenantHarness::create("no_duplicate_timelines")
5813 4 : .await?
5814 4 : .load()
5815 4 : .await;
5816 4 : let _ = tenant
5817 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5818 4 : .await?;
5819 4 :
5820 4 : match tenant
5821 4 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5822 4 : .await
5823 4 : {
5824 4 : Ok(_) => panic!("duplicate timeline creation should fail"),
5825 4 : Err(e) => assert_eq!(
5826 4 : e.to_string(),
5827 4 : "timeline already exists with different parameters".to_string()
5828 4 : ),
5829 4 : }
5830 4 :
5831 4 : Ok(())
5832 4 : }
5833 :
5834 : /// Convenience function to create a page image with given string as the only content
5835 20 : pub fn test_value(s: &str) -> Value {
5836 20 : let mut buf = BytesMut::new();
5837 20 : buf.extend_from_slice(s.as_bytes());
5838 20 : Value::Image(buf.freeze())
5839 20 : }
5840 :
5841 : ///
5842 : /// Test branch creation
5843 : ///
5844 : #[tokio::test]
5845 4 : async fn test_branch() -> anyhow::Result<()> {
5846 4 : use std::str::from_utf8;
5847 4 :
5848 4 : let (tenant, ctx) = TenantHarness::create("test_branch").await?.load().await;
5849 4 : let tline = tenant
5850 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5851 4 : .await?;
5852 4 : let mut writer = tline.writer().await;
5853 4 :
5854 4 : #[allow(non_snake_case)]
5855 4 : let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
5856 4 : #[allow(non_snake_case)]
5857 4 : let TEST_KEY_B: Key = Key::from_hex("110000000033333333444444445500000002").unwrap();
5858 4 :
5859 4 : // Insert a value on the timeline
5860 4 : writer
5861 4 : .put(TEST_KEY_A, Lsn(0x20), &test_value("foo at 0x20"), &ctx)
5862 4 : .await?;
5863 4 : writer
5864 4 : .put(TEST_KEY_B, Lsn(0x20), &test_value("foobar at 0x20"), &ctx)
5865 4 : .await?;
5866 4 : writer.finish_write(Lsn(0x20));
5867 4 :
5868 4 : writer
5869 4 : .put(TEST_KEY_A, Lsn(0x30), &test_value("foo at 0x30"), &ctx)
5870 4 : .await?;
5871 4 : writer.finish_write(Lsn(0x30));
5872 4 : writer
5873 4 : .put(TEST_KEY_A, Lsn(0x40), &test_value("foo at 0x40"), &ctx)
5874 4 : .await?;
5875 4 : writer.finish_write(Lsn(0x40));
5876 4 :
5877 4 : //assert_current_logical_size(&tline, Lsn(0x40));
5878 4 :
5879 4 : // Branch the history, modify relation differently on the new timeline
5880 4 : tenant
5881 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x30)), &ctx)
5882 4 : .await?;
5883 4 : let newtline = tenant
5884 4 : .get_timeline(NEW_TIMELINE_ID, true)
5885 4 : .expect("Should have a local timeline");
5886 4 : let mut new_writer = newtline.writer().await;
5887 4 : new_writer
5888 4 : .put(TEST_KEY_A, Lsn(0x40), &test_value("bar at 0x40"), &ctx)
5889 4 : .await?;
5890 4 : new_writer.finish_write(Lsn(0x40));
5891 4 :
5892 4 : // Check page contents on both branches
5893 4 : assert_eq!(
5894 4 : from_utf8(&tline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
5895 4 : "foo at 0x40"
5896 4 : );
5897 4 : assert_eq!(
5898 4 : from_utf8(&newtline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
5899 4 : "bar at 0x40"
5900 4 : );
5901 4 : assert_eq!(
5902 4 : from_utf8(&newtline.get(TEST_KEY_B, Lsn(0x40), &ctx).await?)?,
5903 4 : "foobar at 0x20"
5904 4 : );
5905 4 :
5906 4 : //assert_current_logical_size(&tline, Lsn(0x40));
5907 4 :
5908 4 : Ok(())
5909 4 : }
5910 :
5911 40 : async fn make_some_layers(
5912 40 : tline: &Timeline,
5913 40 : start_lsn: Lsn,
5914 40 : ctx: &RequestContext,
5915 40 : ) -> anyhow::Result<()> {
5916 40 : let mut lsn = start_lsn;
5917 : {
5918 40 : let mut writer = tline.writer().await;
5919 : // Create a relation on the timeline
5920 40 : writer
5921 40 : .put(
5922 40 : *TEST_KEY,
5923 40 : lsn,
5924 40 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5925 40 : ctx,
5926 40 : )
5927 40 : .await?;
5928 40 : writer.finish_write(lsn);
5929 40 : lsn += 0x10;
5930 40 : writer
5931 40 : .put(
5932 40 : *TEST_KEY,
5933 40 : lsn,
5934 40 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5935 40 : ctx,
5936 40 : )
5937 40 : .await?;
5938 40 : writer.finish_write(lsn);
5939 40 : lsn += 0x10;
5940 40 : }
5941 40 : tline.freeze_and_flush().await?;
5942 : {
5943 40 : let mut writer = tline.writer().await;
5944 40 : writer
5945 40 : .put(
5946 40 : *TEST_KEY,
5947 40 : lsn,
5948 40 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5949 40 : ctx,
5950 40 : )
5951 40 : .await?;
5952 40 : writer.finish_write(lsn);
5953 40 : lsn += 0x10;
5954 40 : writer
5955 40 : .put(
5956 40 : *TEST_KEY,
5957 40 : lsn,
5958 40 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5959 40 : ctx,
5960 40 : )
5961 40 : .await?;
5962 40 : writer.finish_write(lsn);
5963 40 : }
5964 40 : tline.freeze_and_flush().await.map_err(|e| e.into())
5965 40 : }
5966 :
5967 : #[tokio::test(start_paused = true)]
5968 4 : async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
5969 4 : let (tenant, ctx) =
5970 4 : TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
5971 4 : .await?
5972 4 : .load()
5973 4 : .await;
5974 4 : // Advance to the lsn lease deadline so that GC is not blocked by
5975 4 : // initial transition into AttachedSingle.
5976 4 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
5977 4 : tokio::time::resume();
5978 4 : let tline = tenant
5979 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5980 4 : .await?;
5981 4 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5982 4 :
5983 4 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
5984 4 : // FIXME: this doesn't actually remove any layer currently, given how the flushing
5985 4 : // and compaction works. But it does set the 'cutoff' point so that the cross check
5986 4 : // below should fail.
5987 4 : tenant
5988 4 : .gc_iteration(
5989 4 : Some(TIMELINE_ID),
5990 4 : 0x10,
5991 4 : Duration::ZERO,
5992 4 : &CancellationToken::new(),
5993 4 : &ctx,
5994 4 : )
5995 4 : .await?;
5996 4 :
5997 4 : // try to branch at lsn 25, should fail because we already garbage collected the data
5998 4 : match tenant
5999 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
6000 4 : .await
6001 4 : {
6002 4 : Ok(_) => panic!("branching should have failed"),
6003 4 : Err(err) => {
6004 4 : let CreateTimelineError::AncestorLsn(err) = err else {
6005 4 : panic!("wrong error type")
6006 4 : };
6007 4 : assert!(err.to_string().contains("invalid branch start lsn"));
6008 4 : assert!(err
6009 4 : .source()
6010 4 : .unwrap()
6011 4 : .to_string()
6012 4 : .contains("we might've already garbage collected needed data"))
6013 4 : }
6014 4 : }
6015 4 :
6016 4 : Ok(())
6017 4 : }
6018 :
6019 : #[tokio::test]
6020 4 : async fn test_prohibit_branch_creation_on_pre_initdb_lsn() -> anyhow::Result<()> {
6021 4 : let (tenant, ctx) =
6022 4 : TenantHarness::create("test_prohibit_branch_creation_on_pre_initdb_lsn")
6023 4 : .await?
6024 4 : .load()
6025 4 : .await;
6026 4 :
6027 4 : let tline = tenant
6028 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x50), DEFAULT_PG_VERSION, &ctx)
6029 4 : .await?;
6030 4 : // try to branch at lsn 0x25, should fail because initdb lsn is 0x50
6031 4 : match tenant
6032 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
6033 4 : .await
6034 4 : {
6035 4 : Ok(_) => panic!("branching should have failed"),
6036 4 : Err(err) => {
6037 4 : let CreateTimelineError::AncestorLsn(err) = err else {
6038 4 : panic!("wrong error type");
6039 4 : };
6040 4 : assert!(&err.to_string().contains("invalid branch start lsn"));
6041 4 : assert!(&err
6042 4 : .source()
6043 4 : .unwrap()
6044 4 : .to_string()
6045 4 : .contains("is earlier than latest GC cutoff"));
6046 4 : }
6047 4 : }
6048 4 :
6049 4 : Ok(())
6050 4 : }
6051 :
6052 : /*
6053 : // FIXME: This currently fails to error out. Calling GC doesn't currently
6054 : // remove the old value, we'd need to work a little harder
6055 : #[tokio::test]
6056 : async fn test_prohibit_get_for_garbage_collected_data() -> anyhow::Result<()> {
6057 : let repo =
6058 : RepoHarness::create("test_prohibit_get_for_garbage_collected_data")?
6059 : .load();
6060 :
6061 : let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION)?;
6062 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6063 :
6064 : repo.gc_iteration(Some(TIMELINE_ID), 0x10, Duration::ZERO)?;
6065 : let latest_gc_cutoff_lsn = tline.get_latest_gc_cutoff_lsn();
6066 : assert!(*latest_gc_cutoff_lsn > Lsn(0x25));
6067 : match tline.get(*TEST_KEY, Lsn(0x25)) {
6068 : Ok(_) => panic!("request for page should have failed"),
6069 : Err(err) => assert!(err.to_string().contains("not found at")),
6070 : }
6071 : Ok(())
6072 : }
6073 : */
6074 :
6075 : #[tokio::test]
6076 4 : async fn test_get_branchpoints_from_an_inactive_timeline() -> anyhow::Result<()> {
6077 4 : let (tenant, ctx) =
6078 4 : TenantHarness::create("test_get_branchpoints_from_an_inactive_timeline")
6079 4 : .await?
6080 4 : .load()
6081 4 : .await;
6082 4 : let tline = tenant
6083 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6084 4 : .await?;
6085 4 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6086 4 :
6087 4 : tenant
6088 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6089 4 : .await?;
6090 4 : let newtline = tenant
6091 4 : .get_timeline(NEW_TIMELINE_ID, true)
6092 4 : .expect("Should have a local timeline");
6093 4 :
6094 4 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6095 4 :
6096 4 : tline.set_broken("test".to_owned());
6097 4 :
6098 4 : tenant
6099 4 : .gc_iteration(
6100 4 : Some(TIMELINE_ID),
6101 4 : 0x10,
6102 4 : Duration::ZERO,
6103 4 : &CancellationToken::new(),
6104 4 : &ctx,
6105 4 : )
6106 4 : .await?;
6107 4 :
6108 4 : // The branchpoints should contain all timelines, even ones marked
6109 4 : // as Broken.
6110 4 : {
6111 4 : let branchpoints = &tline.gc_info.read().unwrap().retain_lsns;
6112 4 : assert_eq!(branchpoints.len(), 1);
6113 4 : assert_eq!(
6114 4 : branchpoints[0],
6115 4 : (Lsn(0x40), NEW_TIMELINE_ID, MaybeOffloaded::No)
6116 4 : );
6117 4 : }
6118 4 :
6119 4 : // You can read the key from the child branch even though the parent is
6120 4 : // Broken, as long as you don't need to access data from the parent.
6121 4 : assert_eq!(
6122 4 : newtline.get(*TEST_KEY, Lsn(0x70), &ctx).await?,
6123 4 : test_img(&format!("foo at {}", Lsn(0x70)))
6124 4 : );
6125 4 :
6126 4 : // This needs to traverse to the parent, and fails.
6127 4 : let err = newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await.unwrap_err();
6128 4 : assert!(
6129 4 : err.to_string().starts_with(&format!(
6130 4 : "bad state on timeline {}: Broken",
6131 4 : tline.timeline_id
6132 4 : )),
6133 4 : "{err}"
6134 4 : );
6135 4 :
6136 4 : Ok(())
6137 4 : }
6138 :
6139 : #[tokio::test]
6140 4 : async fn test_retain_data_in_parent_which_is_needed_for_child() -> anyhow::Result<()> {
6141 4 : let (tenant, ctx) =
6142 4 : TenantHarness::create("test_retain_data_in_parent_which_is_needed_for_child")
6143 4 : .await?
6144 4 : .load()
6145 4 : .await;
6146 4 : let tline = tenant
6147 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6148 4 : .await?;
6149 4 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6150 4 :
6151 4 : tenant
6152 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6153 4 : .await?;
6154 4 : let newtline = tenant
6155 4 : .get_timeline(NEW_TIMELINE_ID, true)
6156 4 : .expect("Should have a local timeline");
6157 4 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
6158 4 : tenant
6159 4 : .gc_iteration(
6160 4 : Some(TIMELINE_ID),
6161 4 : 0x10,
6162 4 : Duration::ZERO,
6163 4 : &CancellationToken::new(),
6164 4 : &ctx,
6165 4 : )
6166 4 : .await?;
6167 4 : assert!(newtline.get(*TEST_KEY, Lsn(0x25), &ctx).await.is_ok());
6168 4 :
6169 4 : Ok(())
6170 4 : }
6171 : #[tokio::test]
6172 4 : async fn test_parent_keeps_data_forever_after_branching() -> anyhow::Result<()> {
6173 4 : let (tenant, ctx) = TenantHarness::create("test_parent_keeps_data_forever_after_branching")
6174 4 : .await?
6175 4 : .load()
6176 4 : .await;
6177 4 : let tline = tenant
6178 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6179 4 : .await?;
6180 4 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6181 4 :
6182 4 : tenant
6183 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6184 4 : .await?;
6185 4 : let newtline = tenant
6186 4 : .get_timeline(NEW_TIMELINE_ID, true)
6187 4 : .expect("Should have a local timeline");
6188 4 :
6189 4 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6190 4 :
6191 4 : // run gc on parent
6192 4 : tenant
6193 4 : .gc_iteration(
6194 4 : Some(TIMELINE_ID),
6195 4 : 0x10,
6196 4 : Duration::ZERO,
6197 4 : &CancellationToken::new(),
6198 4 : &ctx,
6199 4 : )
6200 4 : .await?;
6201 4 :
6202 4 : // Check that the data is still accessible on the branch.
6203 4 : assert_eq!(
6204 4 : newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
6205 4 : test_img(&format!("foo at {}", Lsn(0x40)))
6206 4 : );
6207 4 :
6208 4 : Ok(())
6209 4 : }
6210 :
6211 : #[tokio::test]
6212 4 : async fn timeline_load() -> anyhow::Result<()> {
6213 4 : const TEST_NAME: &str = "timeline_load";
6214 4 : let harness = TenantHarness::create(TEST_NAME).await?;
6215 4 : {
6216 4 : let (tenant, ctx) = harness.load().await;
6217 4 : let tline = tenant
6218 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x7000), DEFAULT_PG_VERSION, &ctx)
6219 4 : .await?;
6220 4 : make_some_layers(tline.as_ref(), Lsn(0x8000), &ctx).await?;
6221 4 : // so that all uploads finish & we can call harness.load() below again
6222 4 : tenant
6223 4 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
6224 4 : .instrument(harness.span())
6225 4 : .await
6226 4 : .ok()
6227 4 : .unwrap();
6228 4 : }
6229 4 :
6230 4 : let (tenant, _ctx) = harness.load().await;
6231 4 : tenant
6232 4 : .get_timeline(TIMELINE_ID, true)
6233 4 : .expect("cannot load timeline");
6234 4 :
6235 4 : Ok(())
6236 4 : }
6237 :
6238 : #[tokio::test]
6239 4 : async fn timeline_load_with_ancestor() -> anyhow::Result<()> {
6240 4 : const TEST_NAME: &str = "timeline_load_with_ancestor";
6241 4 : let harness = TenantHarness::create(TEST_NAME).await?;
6242 4 : // create two timelines
6243 4 : {
6244 4 : let (tenant, ctx) = harness.load().await;
6245 4 : let tline = tenant
6246 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6247 4 : .await?;
6248 4 :
6249 4 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6250 4 :
6251 4 : let child_tline = tenant
6252 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6253 4 : .await?;
6254 4 : child_tline.set_state(TimelineState::Active);
6255 4 :
6256 4 : let newtline = tenant
6257 4 : .get_timeline(NEW_TIMELINE_ID, true)
6258 4 : .expect("Should have a local timeline");
6259 4 :
6260 4 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6261 4 :
6262 4 : // so that all uploads finish & we can call harness.load() below again
6263 4 : tenant
6264 4 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
6265 4 : .instrument(harness.span())
6266 4 : .await
6267 4 : .ok()
6268 4 : .unwrap();
6269 4 : }
6270 4 :
6271 4 : // check that both of them are initially unloaded
6272 4 : let (tenant, _ctx) = harness.load().await;
6273 4 :
6274 4 : // check that both, child and ancestor are loaded
6275 4 : let _child_tline = tenant
6276 4 : .get_timeline(NEW_TIMELINE_ID, true)
6277 4 : .expect("cannot get child timeline loaded");
6278 4 :
6279 4 : let _ancestor_tline = tenant
6280 4 : .get_timeline(TIMELINE_ID, true)
6281 4 : .expect("cannot get ancestor timeline loaded");
6282 4 :
6283 4 : Ok(())
6284 4 : }
6285 :
6286 : #[tokio::test]
6287 4 : async fn delta_layer_dumping() -> anyhow::Result<()> {
6288 4 : use storage_layer::AsLayerDesc;
6289 4 : let (tenant, ctx) = TenantHarness::create("test_layer_dumping")
6290 4 : .await?
6291 4 : .load()
6292 4 : .await;
6293 4 : let tline = tenant
6294 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6295 4 : .await?;
6296 4 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6297 4 :
6298 4 : let layer_map = tline.layers.read().await;
6299 4 : let level0_deltas = layer_map
6300 4 : .layer_map()?
6301 4 : .level0_deltas()
6302 4 : .iter()
6303 8 : .map(|desc| layer_map.get_from_desc(desc))
6304 4 : .collect::<Vec<_>>();
6305 4 :
6306 4 : assert!(!level0_deltas.is_empty());
6307 4 :
6308 12 : for delta in level0_deltas {
6309 4 : // Ensure we are dumping a delta layer here
6310 8 : assert!(delta.layer_desc().is_delta);
6311 8 : delta.dump(true, &ctx).await.unwrap();
6312 4 : }
6313 4 :
6314 4 : Ok(())
6315 4 : }
6316 :
6317 : #[tokio::test]
6318 4 : async fn test_images() -> anyhow::Result<()> {
6319 4 : let (tenant, ctx) = TenantHarness::create("test_images").await?.load().await;
6320 4 : let tline = tenant
6321 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6322 4 : .await?;
6323 4 :
6324 4 : let mut writer = tline.writer().await;
6325 4 : writer
6326 4 : .put(
6327 4 : *TEST_KEY,
6328 4 : Lsn(0x10),
6329 4 : &Value::Image(test_img("foo at 0x10")),
6330 4 : &ctx,
6331 4 : )
6332 4 : .await?;
6333 4 : writer.finish_write(Lsn(0x10));
6334 4 : drop(writer);
6335 4 :
6336 4 : tline.freeze_and_flush().await?;
6337 4 : tline
6338 4 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6339 4 : .await?;
6340 4 :
6341 4 : let mut writer = tline.writer().await;
6342 4 : writer
6343 4 : .put(
6344 4 : *TEST_KEY,
6345 4 : Lsn(0x20),
6346 4 : &Value::Image(test_img("foo at 0x20")),
6347 4 : &ctx,
6348 4 : )
6349 4 : .await?;
6350 4 : writer.finish_write(Lsn(0x20));
6351 4 : drop(writer);
6352 4 :
6353 4 : tline.freeze_and_flush().await?;
6354 4 : tline
6355 4 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6356 4 : .await?;
6357 4 :
6358 4 : let mut writer = tline.writer().await;
6359 4 : writer
6360 4 : .put(
6361 4 : *TEST_KEY,
6362 4 : Lsn(0x30),
6363 4 : &Value::Image(test_img("foo at 0x30")),
6364 4 : &ctx,
6365 4 : )
6366 4 : .await?;
6367 4 : writer.finish_write(Lsn(0x30));
6368 4 : drop(writer);
6369 4 :
6370 4 : tline.freeze_and_flush().await?;
6371 4 : tline
6372 4 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6373 4 : .await?;
6374 4 :
6375 4 : let mut writer = tline.writer().await;
6376 4 : writer
6377 4 : .put(
6378 4 : *TEST_KEY,
6379 4 : Lsn(0x40),
6380 4 : &Value::Image(test_img("foo at 0x40")),
6381 4 : &ctx,
6382 4 : )
6383 4 : .await?;
6384 4 : writer.finish_write(Lsn(0x40));
6385 4 : drop(writer);
6386 4 :
6387 4 : tline.freeze_and_flush().await?;
6388 4 : tline
6389 4 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6390 4 : .await?;
6391 4 :
6392 4 : assert_eq!(
6393 4 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
6394 4 : test_img("foo at 0x10")
6395 4 : );
6396 4 : assert_eq!(
6397 4 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
6398 4 : test_img("foo at 0x10")
6399 4 : );
6400 4 : assert_eq!(
6401 4 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
6402 4 : test_img("foo at 0x20")
6403 4 : );
6404 4 : assert_eq!(
6405 4 : tline.get(*TEST_KEY, Lsn(0x30), &ctx).await?,
6406 4 : test_img("foo at 0x30")
6407 4 : );
6408 4 : assert_eq!(
6409 4 : tline.get(*TEST_KEY, Lsn(0x40), &ctx).await?,
6410 4 : test_img("foo at 0x40")
6411 4 : );
6412 4 :
6413 4 : Ok(())
6414 4 : }
6415 :
6416 8 : async fn bulk_insert_compact_gc(
6417 8 : tenant: &Tenant,
6418 8 : timeline: &Arc<Timeline>,
6419 8 : ctx: &RequestContext,
6420 8 : lsn: Lsn,
6421 8 : repeat: usize,
6422 8 : key_count: usize,
6423 8 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
6424 8 : let compact = true;
6425 8 : bulk_insert_maybe_compact_gc(tenant, timeline, ctx, lsn, repeat, key_count, compact).await
6426 8 : }
6427 :
6428 16 : async fn bulk_insert_maybe_compact_gc(
6429 16 : tenant: &Tenant,
6430 16 : timeline: &Arc<Timeline>,
6431 16 : ctx: &RequestContext,
6432 16 : mut lsn: Lsn,
6433 16 : repeat: usize,
6434 16 : key_count: usize,
6435 16 : compact: bool,
6436 16 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
6437 16 : let mut inserted: HashMap<Key, BTreeSet<Lsn>> = Default::default();
6438 16 :
6439 16 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6440 16 : let mut blknum = 0;
6441 16 :
6442 16 : // Enforce that key range is monotonously increasing
6443 16 : let mut keyspace = KeySpaceAccum::new();
6444 16 :
6445 16 : let cancel = CancellationToken::new();
6446 16 :
6447 16 : for _ in 0..repeat {
6448 800 : for _ in 0..key_count {
6449 8000000 : test_key.field6 = blknum;
6450 8000000 : let mut writer = timeline.writer().await;
6451 8000000 : writer
6452 8000000 : .put(
6453 8000000 : test_key,
6454 8000000 : lsn,
6455 8000000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6456 8000000 : ctx,
6457 8000000 : )
6458 8000000 : .await?;
6459 8000000 : inserted.entry(test_key).or_default().insert(lsn);
6460 8000000 : writer.finish_write(lsn);
6461 8000000 : drop(writer);
6462 8000000 :
6463 8000000 : keyspace.add_key(test_key);
6464 8000000 :
6465 8000000 : lsn = Lsn(lsn.0 + 0x10);
6466 8000000 : blknum += 1;
6467 : }
6468 :
6469 800 : timeline.freeze_and_flush().await?;
6470 800 : if compact {
6471 : // this requires timeline to be &Arc<Timeline>
6472 400 : timeline.compact(&cancel, EnumSet::empty(), ctx).await?;
6473 400 : }
6474 :
6475 : // this doesn't really need to use the timeline_id target, but it is closer to what it
6476 : // originally was.
6477 800 : let res = tenant
6478 800 : .gc_iteration(Some(timeline.timeline_id), 0, Duration::ZERO, &cancel, ctx)
6479 800 : .await?;
6480 :
6481 800 : assert_eq!(res.layers_removed, 0, "this never removes anything");
6482 : }
6483 :
6484 16 : Ok(inserted)
6485 16 : }
6486 :
6487 : //
6488 : // Insert 1000 key-value pairs with increasing keys, flush, compact, GC.
6489 : // Repeat 50 times.
6490 : //
6491 : #[tokio::test]
6492 4 : async fn test_bulk_insert() -> anyhow::Result<()> {
6493 4 : let harness = TenantHarness::create("test_bulk_insert").await?;
6494 4 : let (tenant, ctx) = harness.load().await;
6495 4 : let tline = tenant
6496 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6497 4 : .await?;
6498 4 :
6499 4 : let lsn = Lsn(0x10);
6500 4 : bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
6501 4 :
6502 4 : Ok(())
6503 4 : }
6504 :
6505 : // Test the vectored get real implementation against a simple sequential implementation.
6506 : //
6507 : // The test generates a keyspace by repeatedly flushing the in-memory layer and compacting.
6508 : // Projected to 2D the key space looks like below. Lsn grows upwards on the Y axis and keys
6509 : // grow to the right on the X axis.
6510 : // [Delta]
6511 : // [Delta]
6512 : // [Delta]
6513 : // [Delta]
6514 : // ------------ Image ---------------
6515 : //
6516 : // After layer generation we pick the ranges to query as follows:
6517 : // 1. The beginning of each delta layer
6518 : // 2. At the seam between two adjacent delta layers
6519 : //
6520 : // There's one major downside to this test: delta layers only contains images,
6521 : // so the search can stop at the first delta layer and doesn't traverse any deeper.
6522 : #[tokio::test]
6523 4 : async fn test_get_vectored() -> anyhow::Result<()> {
6524 4 : let harness = TenantHarness::create("test_get_vectored").await?;
6525 4 : let (tenant, ctx) = harness.load().await;
6526 4 : let io_concurrency = IoConcurrency::spawn_for_test();
6527 4 : let tline = tenant
6528 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6529 4 : .await?;
6530 4 :
6531 4 : let lsn = Lsn(0x10);
6532 4 : let inserted = bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
6533 4 :
6534 4 : let guard = tline.layers.read().await;
6535 4 : let lm = guard.layer_map()?;
6536 4 :
6537 4 : lm.dump(true, &ctx).await?;
6538 4 :
6539 4 : let mut reads = Vec::new();
6540 4 : let mut prev = None;
6541 24 : lm.iter_historic_layers().for_each(|desc| {
6542 24 : if !desc.is_delta() {
6543 4 : prev = Some(desc.clone());
6544 4 : return;
6545 20 : }
6546 20 :
6547 20 : let start = desc.key_range.start;
6548 20 : let end = desc
6549 20 : .key_range
6550 20 : .start
6551 20 : .add(Timeline::MAX_GET_VECTORED_KEYS.try_into().unwrap());
6552 20 : reads.push(KeySpace {
6553 20 : ranges: vec![start..end],
6554 20 : });
6555 4 :
6556 20 : if let Some(prev) = &prev {
6557 20 : if !prev.is_delta() {
6558 20 : return;
6559 4 : }
6560 0 :
6561 0 : let first_range = Key {
6562 0 : field6: prev.key_range.end.field6 - 4,
6563 0 : ..prev.key_range.end
6564 0 : }..prev.key_range.end;
6565 0 :
6566 0 : let second_range = desc.key_range.start..Key {
6567 0 : field6: desc.key_range.start.field6 + 4,
6568 0 : ..desc.key_range.start
6569 0 : };
6570 0 :
6571 0 : reads.push(KeySpace {
6572 0 : ranges: vec![first_range, second_range],
6573 0 : });
6574 4 : };
6575 4 :
6576 4 : prev = Some(desc.clone());
6577 24 : });
6578 4 :
6579 4 : drop(guard);
6580 4 :
6581 4 : // Pick a big LSN such that we query over all the changes.
6582 4 : let reads_lsn = Lsn(u64::MAX - 1);
6583 4 :
6584 24 : for read in reads {
6585 20 : info!("Doing vectored read on {:?}", read);
6586 4 :
6587 20 : let vectored_res = tline
6588 20 : .get_vectored_impl(
6589 20 : read.clone(),
6590 20 : reads_lsn,
6591 20 : &mut ValuesReconstructState::new(io_concurrency.clone()),
6592 20 : &ctx,
6593 20 : )
6594 20 : .await;
6595 4 :
6596 20 : let mut expected_lsns: HashMap<Key, Lsn> = Default::default();
6597 20 : let mut expect_missing = false;
6598 20 : let mut key = read.start().unwrap();
6599 660 : while key != read.end().unwrap() {
6600 640 : if let Some(lsns) = inserted.get(&key) {
6601 640 : let expected_lsn = lsns.iter().rfind(|lsn| **lsn <= reads_lsn);
6602 640 : match expected_lsn {
6603 640 : Some(lsn) => {
6604 640 : expected_lsns.insert(key, *lsn);
6605 640 : }
6606 4 : None => {
6607 4 : expect_missing = true;
6608 0 : break;
6609 4 : }
6610 4 : }
6611 4 : } else {
6612 4 : expect_missing = true;
6613 0 : break;
6614 4 : }
6615 4 :
6616 640 : key = key.next();
6617 4 : }
6618 4 :
6619 20 : if expect_missing {
6620 4 : assert!(matches!(vectored_res, Err(GetVectoredError::MissingKey(_))));
6621 4 : } else {
6622 640 : for (key, image) in vectored_res? {
6623 640 : let expected_lsn = expected_lsns.get(&key).expect("determined above");
6624 640 : let expected_image = test_img(&format!("{} at {}", key.field6, expected_lsn));
6625 640 : assert_eq!(image?, expected_image);
6626 4 : }
6627 4 : }
6628 4 : }
6629 4 :
6630 4 : Ok(())
6631 4 : }
6632 :
6633 : #[tokio::test]
6634 4 : async fn test_get_vectored_aux_files() -> anyhow::Result<()> {
6635 4 : let harness = TenantHarness::create("test_get_vectored_aux_files").await?;
6636 4 :
6637 4 : let (tenant, ctx) = harness.load().await;
6638 4 : let io_concurrency = IoConcurrency::spawn_for_test();
6639 4 : let tline = tenant
6640 4 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
6641 4 : .await?;
6642 4 : let tline = tline.raw_timeline().unwrap();
6643 4 :
6644 4 : let mut modification = tline.begin_modification(Lsn(0x1000));
6645 4 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
6646 4 : modification.set_lsn(Lsn(0x1008))?;
6647 4 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
6648 4 : modification.commit(&ctx).await?;
6649 4 :
6650 4 : let child_timeline_id = TimelineId::generate();
6651 4 : tenant
6652 4 : .branch_timeline_test(
6653 4 : tline,
6654 4 : child_timeline_id,
6655 4 : Some(tline.get_last_record_lsn()),
6656 4 : &ctx,
6657 4 : )
6658 4 : .await?;
6659 4 :
6660 4 : let child_timeline = tenant
6661 4 : .get_timeline(child_timeline_id, true)
6662 4 : .expect("Should have the branched timeline");
6663 4 :
6664 4 : let aux_keyspace = KeySpace {
6665 4 : ranges: vec![NON_INHERITED_RANGE],
6666 4 : };
6667 4 : let read_lsn = child_timeline.get_last_record_lsn();
6668 4 :
6669 4 : let vectored_res = child_timeline
6670 4 : .get_vectored_impl(
6671 4 : aux_keyspace.clone(),
6672 4 : read_lsn,
6673 4 : &mut ValuesReconstructState::new(io_concurrency.clone()),
6674 4 : &ctx,
6675 4 : )
6676 4 : .await;
6677 4 :
6678 4 : let images = vectored_res?;
6679 4 : assert!(images.is_empty());
6680 4 : Ok(())
6681 4 : }
6682 :
6683 : // Test that vectored get handles layer gaps correctly
6684 : // by advancing into the next ancestor timeline if required.
6685 : //
6686 : // The test generates timelines that look like the diagram below.
6687 : // We leave a gap in one of the L1 layers at `gap_at_key` (`/` in the diagram).
6688 : // The reconstruct data for that key lies in the ancestor timeline (`X` in the diagram).
6689 : //
6690 : // ```
6691 : //-------------------------------+
6692 : // ... |
6693 : // [ L1 ] |
6694 : // [ / L1 ] | Child Timeline
6695 : // ... |
6696 : // ------------------------------+
6697 : // [ X L1 ] | Parent Timeline
6698 : // ------------------------------+
6699 : // ```
6700 : #[tokio::test]
6701 4 : async fn test_get_vectored_key_gap() -> anyhow::Result<()> {
6702 4 : let tenant_conf = TenantConf {
6703 4 : // Make compaction deterministic
6704 4 : gc_period: Duration::ZERO,
6705 4 : compaction_period: Duration::ZERO,
6706 4 : // Encourage creation of L1 layers
6707 4 : checkpoint_distance: 16 * 1024,
6708 4 : compaction_target_size: 8 * 1024,
6709 4 : ..TenantConf::default()
6710 4 : };
6711 4 :
6712 4 : let harness = TenantHarness::create_custom(
6713 4 : "test_get_vectored_key_gap",
6714 4 : tenant_conf,
6715 4 : TenantId::generate(),
6716 4 : ShardIdentity::unsharded(),
6717 4 : Generation::new(0xdeadbeef),
6718 4 : )
6719 4 : .await?;
6720 4 : let (tenant, ctx) = harness.load().await;
6721 4 : let io_concurrency = IoConcurrency::spawn_for_test();
6722 4 :
6723 4 : let mut current_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6724 4 : let gap_at_key = current_key.add(100);
6725 4 : let mut current_lsn = Lsn(0x10);
6726 4 :
6727 4 : const KEY_COUNT: usize = 10_000;
6728 4 :
6729 4 : let timeline_id = TimelineId::generate();
6730 4 : let current_timeline = tenant
6731 4 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
6732 4 : .await?;
6733 4 :
6734 4 : current_lsn += 0x100;
6735 4 :
6736 4 : let mut writer = current_timeline.writer().await;
6737 4 : writer
6738 4 : .put(
6739 4 : gap_at_key,
6740 4 : current_lsn,
6741 4 : &Value::Image(test_img(&format!("{} at {}", gap_at_key, current_lsn))),
6742 4 : &ctx,
6743 4 : )
6744 4 : .await?;
6745 4 : writer.finish_write(current_lsn);
6746 4 : drop(writer);
6747 4 :
6748 4 : let mut latest_lsns = HashMap::new();
6749 4 : latest_lsns.insert(gap_at_key, current_lsn);
6750 4 :
6751 4 : current_timeline.freeze_and_flush().await?;
6752 4 :
6753 4 : let child_timeline_id = TimelineId::generate();
6754 4 :
6755 4 : tenant
6756 4 : .branch_timeline_test(
6757 4 : ¤t_timeline,
6758 4 : child_timeline_id,
6759 4 : Some(current_lsn),
6760 4 : &ctx,
6761 4 : )
6762 4 : .await?;
6763 4 : let child_timeline = tenant
6764 4 : .get_timeline(child_timeline_id, true)
6765 4 : .expect("Should have the branched timeline");
6766 4 :
6767 40004 : for i in 0..KEY_COUNT {
6768 40000 : if current_key == gap_at_key {
6769 4 : current_key = current_key.next();
6770 4 : continue;
6771 39996 : }
6772 39996 :
6773 39996 : current_lsn += 0x10;
6774 4 :
6775 39996 : let mut writer = child_timeline.writer().await;
6776 39996 : writer
6777 39996 : .put(
6778 39996 : current_key,
6779 39996 : current_lsn,
6780 39996 : &Value::Image(test_img(&format!("{} at {}", current_key, current_lsn))),
6781 39996 : &ctx,
6782 39996 : )
6783 39996 : .await?;
6784 39996 : writer.finish_write(current_lsn);
6785 39996 : drop(writer);
6786 39996 :
6787 39996 : latest_lsns.insert(current_key, current_lsn);
6788 39996 : current_key = current_key.next();
6789 39996 :
6790 39996 : // Flush every now and then to encourage layer file creation.
6791 39996 : if i % 500 == 0 {
6792 80 : child_timeline.freeze_and_flush().await?;
6793 39916 : }
6794 4 : }
6795 4 :
6796 4 : child_timeline.freeze_and_flush().await?;
6797 4 : let mut flags = EnumSet::new();
6798 4 : flags.insert(CompactFlags::ForceRepartition);
6799 4 : child_timeline
6800 4 : .compact(&CancellationToken::new(), flags, &ctx)
6801 4 : .await?;
6802 4 :
6803 4 : let key_near_end = {
6804 4 : let mut tmp = current_key;
6805 4 : tmp.field6 -= 10;
6806 4 : tmp
6807 4 : };
6808 4 :
6809 4 : let key_near_gap = {
6810 4 : let mut tmp = gap_at_key;
6811 4 : tmp.field6 -= 10;
6812 4 : tmp
6813 4 : };
6814 4 :
6815 4 : let read = KeySpace {
6816 4 : ranges: vec![key_near_gap..gap_at_key.next(), key_near_end..current_key],
6817 4 : };
6818 4 : let results = child_timeline
6819 4 : .get_vectored_impl(
6820 4 : read.clone(),
6821 4 : current_lsn,
6822 4 : &mut ValuesReconstructState::new(io_concurrency.clone()),
6823 4 : &ctx,
6824 4 : )
6825 4 : .await?;
6826 4 :
6827 88 : for (key, img_res) in results {
6828 84 : let expected = test_img(&format!("{} at {}", key, latest_lsns[&key]));
6829 84 : assert_eq!(img_res?, expected);
6830 4 : }
6831 4 :
6832 4 : Ok(())
6833 4 : }
6834 :
6835 : // Test that vectored get descends into ancestor timelines correctly and
6836 : // does not return an image that's newer than requested.
6837 : //
6838 : // The diagram below ilustrates an interesting case. We have a parent timeline
6839 : // (top of the Lsn range) and a child timeline. The request key cannot be reconstructed
6840 : // from the child timeline, so the parent timeline must be visited. When advacing into
6841 : // the child timeline, the read path needs to remember what the requested Lsn was in
6842 : // order to avoid returning an image that's too new. The test below constructs such
6843 : // a timeline setup and does a few queries around the Lsn of each page image.
6844 : // ```
6845 : // LSN
6846 : // ^
6847 : // |
6848 : // |
6849 : // 500 | --------------------------------------> branch point
6850 : // 400 | X
6851 : // 300 | X
6852 : // 200 | --------------------------------------> requested lsn
6853 : // 100 | X
6854 : // |---------------------------------------> Key
6855 : // |
6856 : // ------> requested key
6857 : //
6858 : // Legend:
6859 : // * X - page images
6860 : // ```
6861 : #[tokio::test]
6862 4 : async fn test_get_vectored_ancestor_descent() -> anyhow::Result<()> {
6863 4 : let harness = TenantHarness::create("test_get_vectored_on_lsn_axis").await?;
6864 4 : let (tenant, ctx) = harness.load().await;
6865 4 : let io_concurrency = IoConcurrency::spawn_for_test();
6866 4 :
6867 4 : let start_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6868 4 : let end_key = start_key.add(1000);
6869 4 : let child_gap_at_key = start_key.add(500);
6870 4 : let mut parent_gap_lsns: BTreeMap<Lsn, String> = BTreeMap::new();
6871 4 :
6872 4 : let mut current_lsn = Lsn(0x10);
6873 4 :
6874 4 : let timeline_id = TimelineId::generate();
6875 4 : let parent_timeline = tenant
6876 4 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
6877 4 : .await?;
6878 4 :
6879 4 : current_lsn += 0x100;
6880 4 :
6881 16 : for _ in 0..3 {
6882 12 : let mut key = start_key;
6883 12012 : while key < end_key {
6884 12000 : current_lsn += 0x10;
6885 12000 :
6886 12000 : let image_value = format!("{} at {}", child_gap_at_key, current_lsn);
6887 4 :
6888 12000 : let mut writer = parent_timeline.writer().await;
6889 12000 : writer
6890 12000 : .put(
6891 12000 : key,
6892 12000 : current_lsn,
6893 12000 : &Value::Image(test_img(&image_value)),
6894 12000 : &ctx,
6895 12000 : )
6896 12000 : .await?;
6897 12000 : writer.finish_write(current_lsn);
6898 12000 :
6899 12000 : if key == child_gap_at_key {
6900 12 : parent_gap_lsns.insert(current_lsn, image_value);
6901 11988 : }
6902 4 :
6903 12000 : key = key.next();
6904 4 : }
6905 4 :
6906 12 : parent_timeline.freeze_and_flush().await?;
6907 4 : }
6908 4 :
6909 4 : let child_timeline_id = TimelineId::generate();
6910 4 :
6911 4 : let child_timeline = tenant
6912 4 : .branch_timeline_test(&parent_timeline, child_timeline_id, Some(current_lsn), &ctx)
6913 4 : .await?;
6914 4 :
6915 4 : let mut key = start_key;
6916 4004 : while key < end_key {
6917 4000 : if key == child_gap_at_key {
6918 4 : key = key.next();
6919 4 : continue;
6920 3996 : }
6921 3996 :
6922 3996 : current_lsn += 0x10;
6923 4 :
6924 3996 : let mut writer = child_timeline.writer().await;
6925 3996 : writer
6926 3996 : .put(
6927 3996 : key,
6928 3996 : current_lsn,
6929 3996 : &Value::Image(test_img(&format!("{} at {}", key, current_lsn))),
6930 3996 : &ctx,
6931 3996 : )
6932 3996 : .await?;
6933 3996 : writer.finish_write(current_lsn);
6934 3996 :
6935 3996 : key = key.next();
6936 4 : }
6937 4 :
6938 4 : child_timeline.freeze_and_flush().await?;
6939 4 :
6940 4 : let lsn_offsets: [i64; 5] = [-10, -1, 0, 1, 10];
6941 4 : let mut query_lsns = Vec::new();
6942 12 : for image_lsn in parent_gap_lsns.keys().rev() {
6943 72 : for offset in lsn_offsets {
6944 60 : query_lsns.push(Lsn(image_lsn
6945 60 : .0
6946 60 : .checked_add_signed(offset)
6947 60 : .expect("Shouldn't overflow")));
6948 60 : }
6949 4 : }
6950 4 :
6951 64 : for query_lsn in query_lsns {
6952 60 : let results = child_timeline
6953 60 : .get_vectored_impl(
6954 60 : KeySpace {
6955 60 : ranges: vec![child_gap_at_key..child_gap_at_key.next()],
6956 60 : },
6957 60 : query_lsn,
6958 60 : &mut ValuesReconstructState::new(io_concurrency.clone()),
6959 60 : &ctx,
6960 60 : )
6961 60 : .await;
6962 4 :
6963 60 : let expected_item = parent_gap_lsns
6964 60 : .iter()
6965 60 : .rev()
6966 136 : .find(|(lsn, _)| **lsn <= query_lsn);
6967 60 :
6968 60 : info!(
6969 4 : "Doing vectored read at LSN {}. Expecting image to be: {:?}",
6970 4 : query_lsn, expected_item
6971 4 : );
6972 4 :
6973 60 : match expected_item {
6974 52 : Some((_, img_value)) => {
6975 52 : let key_results = results.expect("No vectored get error expected");
6976 52 : let key_result = &key_results[&child_gap_at_key];
6977 52 : let returned_img = key_result
6978 52 : .as_ref()
6979 52 : .expect("No page reconstruct error expected");
6980 52 :
6981 52 : info!(
6982 4 : "Vectored read at LSN {} returned image {}",
6983 0 : query_lsn,
6984 0 : std::str::from_utf8(returned_img)?
6985 4 : );
6986 52 : assert_eq!(*returned_img, test_img(img_value));
6987 4 : }
6988 4 : None => {
6989 8 : assert!(matches!(results, Err(GetVectoredError::MissingKey(_))));
6990 4 : }
6991 4 : }
6992 4 : }
6993 4 :
6994 4 : Ok(())
6995 4 : }
6996 :
6997 : #[tokio::test]
6998 4 : async fn test_random_updates() -> anyhow::Result<()> {
6999 4 : let names_algorithms = [
7000 4 : ("test_random_updates_legacy", CompactionAlgorithm::Legacy),
7001 4 : ("test_random_updates_tiered", CompactionAlgorithm::Tiered),
7002 4 : ];
7003 12 : for (name, algorithm) in names_algorithms {
7004 8 : test_random_updates_algorithm(name, algorithm).await?;
7005 4 : }
7006 4 : Ok(())
7007 4 : }
7008 :
7009 8 : async fn test_random_updates_algorithm(
7010 8 : name: &'static str,
7011 8 : compaction_algorithm: CompactionAlgorithm,
7012 8 : ) -> anyhow::Result<()> {
7013 8 : let mut harness = TenantHarness::create(name).await?;
7014 8 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
7015 8 : kind: compaction_algorithm,
7016 8 : };
7017 8 : let (tenant, ctx) = harness.load().await;
7018 8 : let tline = tenant
7019 8 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7020 8 : .await?;
7021 :
7022 : const NUM_KEYS: usize = 1000;
7023 8 : let cancel = CancellationToken::new();
7024 8 :
7025 8 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7026 8 : let mut test_key_end = test_key;
7027 8 : test_key_end.field6 = NUM_KEYS as u32;
7028 8 : tline.add_extra_test_dense_keyspace(KeySpace::single(test_key..test_key_end));
7029 8 :
7030 8 : let mut keyspace = KeySpaceAccum::new();
7031 8 :
7032 8 : // Track when each page was last modified. Used to assert that
7033 8 : // a read sees the latest page version.
7034 8 : let mut updated = [Lsn(0); NUM_KEYS];
7035 8 :
7036 8 : let mut lsn = Lsn(0x10);
7037 : #[allow(clippy::needless_range_loop)]
7038 8008 : for blknum in 0..NUM_KEYS {
7039 8000 : lsn = Lsn(lsn.0 + 0x10);
7040 8000 : test_key.field6 = blknum as u32;
7041 8000 : let mut writer = tline.writer().await;
7042 8000 : writer
7043 8000 : .put(
7044 8000 : test_key,
7045 8000 : lsn,
7046 8000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7047 8000 : &ctx,
7048 8000 : )
7049 8000 : .await?;
7050 8000 : writer.finish_write(lsn);
7051 8000 : updated[blknum] = lsn;
7052 8000 : drop(writer);
7053 8000 :
7054 8000 : keyspace.add_key(test_key);
7055 : }
7056 :
7057 408 : for _ in 0..50 {
7058 400400 : for _ in 0..NUM_KEYS {
7059 400000 : lsn = Lsn(lsn.0 + 0x10);
7060 400000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7061 400000 : test_key.field6 = blknum as u32;
7062 400000 : let mut writer = tline.writer().await;
7063 400000 : writer
7064 400000 : .put(
7065 400000 : test_key,
7066 400000 : lsn,
7067 400000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7068 400000 : &ctx,
7069 400000 : )
7070 400000 : .await?;
7071 400000 : writer.finish_write(lsn);
7072 400000 : drop(writer);
7073 400000 : updated[blknum] = lsn;
7074 : }
7075 :
7076 : // Read all the blocks
7077 400000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7078 400000 : test_key.field6 = blknum as u32;
7079 400000 : assert_eq!(
7080 400000 : tline.get(test_key, lsn, &ctx).await?,
7081 400000 : test_img(&format!("{} at {}", blknum, last_lsn))
7082 : );
7083 : }
7084 :
7085 : // Perform a cycle of flush, and GC
7086 400 : tline.freeze_and_flush().await?;
7087 400 : tenant
7088 400 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7089 400 : .await?;
7090 : }
7091 :
7092 8 : Ok(())
7093 8 : }
7094 :
7095 : #[tokio::test]
7096 4 : async fn test_traverse_branches() -> anyhow::Result<()> {
7097 4 : let (tenant, ctx) = TenantHarness::create("test_traverse_branches")
7098 4 : .await?
7099 4 : .load()
7100 4 : .await;
7101 4 : let mut tline = tenant
7102 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7103 4 : .await?;
7104 4 :
7105 4 : const NUM_KEYS: usize = 1000;
7106 4 :
7107 4 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7108 4 :
7109 4 : let mut keyspace = KeySpaceAccum::new();
7110 4 :
7111 4 : let cancel = CancellationToken::new();
7112 4 :
7113 4 : // Track when each page was last modified. Used to assert that
7114 4 : // a read sees the latest page version.
7115 4 : let mut updated = [Lsn(0); NUM_KEYS];
7116 4 :
7117 4 : let mut lsn = Lsn(0x10);
7118 4 : #[allow(clippy::needless_range_loop)]
7119 4004 : for blknum in 0..NUM_KEYS {
7120 4000 : lsn = Lsn(lsn.0 + 0x10);
7121 4000 : test_key.field6 = blknum as u32;
7122 4000 : let mut writer = tline.writer().await;
7123 4000 : writer
7124 4000 : .put(
7125 4000 : test_key,
7126 4000 : lsn,
7127 4000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7128 4000 : &ctx,
7129 4000 : )
7130 4000 : .await?;
7131 4000 : writer.finish_write(lsn);
7132 4000 : updated[blknum] = lsn;
7133 4000 : drop(writer);
7134 4000 :
7135 4000 : keyspace.add_key(test_key);
7136 4 : }
7137 4 :
7138 204 : for _ in 0..50 {
7139 200 : let new_tline_id = TimelineId::generate();
7140 200 : tenant
7141 200 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7142 200 : .await?;
7143 200 : tline = tenant
7144 200 : .get_timeline(new_tline_id, true)
7145 200 : .expect("Should have the branched timeline");
7146 4 :
7147 200200 : for _ in 0..NUM_KEYS {
7148 200000 : lsn = Lsn(lsn.0 + 0x10);
7149 200000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7150 200000 : test_key.field6 = blknum as u32;
7151 200000 : let mut writer = tline.writer().await;
7152 200000 : writer
7153 200000 : .put(
7154 200000 : test_key,
7155 200000 : lsn,
7156 200000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7157 200000 : &ctx,
7158 200000 : )
7159 200000 : .await?;
7160 200000 : println!("updating {} at {}", blknum, lsn);
7161 200000 : writer.finish_write(lsn);
7162 200000 : drop(writer);
7163 200000 : updated[blknum] = lsn;
7164 4 : }
7165 4 :
7166 4 : // Read all the blocks
7167 200000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7168 200000 : test_key.field6 = blknum as u32;
7169 200000 : assert_eq!(
7170 200000 : tline.get(test_key, lsn, &ctx).await?,
7171 200000 : test_img(&format!("{} at {}", blknum, last_lsn))
7172 4 : );
7173 4 : }
7174 4 :
7175 4 : // Perform a cycle of flush, compact, and GC
7176 200 : tline.freeze_and_flush().await?;
7177 200 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
7178 200 : tenant
7179 200 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7180 200 : .await?;
7181 4 : }
7182 4 :
7183 4 : Ok(())
7184 4 : }
7185 :
7186 : #[tokio::test]
7187 4 : async fn test_traverse_ancestors() -> anyhow::Result<()> {
7188 4 : let (tenant, ctx) = TenantHarness::create("test_traverse_ancestors")
7189 4 : .await?
7190 4 : .load()
7191 4 : .await;
7192 4 : let mut tline = tenant
7193 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7194 4 : .await?;
7195 4 :
7196 4 : const NUM_KEYS: usize = 100;
7197 4 : const NUM_TLINES: usize = 50;
7198 4 :
7199 4 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7200 4 : // Track page mutation lsns across different timelines.
7201 4 : let mut updated = [[Lsn(0); NUM_KEYS]; NUM_TLINES];
7202 4 :
7203 4 : let mut lsn = Lsn(0x10);
7204 4 :
7205 4 : #[allow(clippy::needless_range_loop)]
7206 204 : for idx in 0..NUM_TLINES {
7207 200 : let new_tline_id = TimelineId::generate();
7208 200 : tenant
7209 200 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7210 200 : .await?;
7211 200 : tline = tenant
7212 200 : .get_timeline(new_tline_id, true)
7213 200 : .expect("Should have the branched timeline");
7214 4 :
7215 20200 : for _ in 0..NUM_KEYS {
7216 20000 : lsn = Lsn(lsn.0 + 0x10);
7217 20000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7218 20000 : test_key.field6 = blknum as u32;
7219 20000 : let mut writer = tline.writer().await;
7220 20000 : writer
7221 20000 : .put(
7222 20000 : test_key,
7223 20000 : lsn,
7224 20000 : &Value::Image(test_img(&format!("{} {} at {}", idx, blknum, lsn))),
7225 20000 : &ctx,
7226 20000 : )
7227 20000 : .await?;
7228 20000 : println!("updating [{}][{}] at {}", idx, blknum, lsn);
7229 20000 : writer.finish_write(lsn);
7230 20000 : drop(writer);
7231 20000 : updated[idx][blknum] = lsn;
7232 4 : }
7233 4 : }
7234 4 :
7235 4 : // Read pages from leaf timeline across all ancestors.
7236 200 : for (idx, lsns) in updated.iter().enumerate() {
7237 20000 : for (blknum, lsn) in lsns.iter().enumerate() {
7238 4 : // Skip empty mutations.
7239 20000 : if lsn.0 == 0 {
7240 7324 : continue;
7241 12676 : }
7242 12676 : println!("checking [{idx}][{blknum}] at {lsn}");
7243 12676 : test_key.field6 = blknum as u32;
7244 12676 : assert_eq!(
7245 12676 : tline.get(test_key, *lsn, &ctx).await?,
7246 12676 : test_img(&format!("{idx} {blknum} at {lsn}"))
7247 4 : );
7248 4 : }
7249 4 : }
7250 4 : Ok(())
7251 4 : }
7252 :
7253 : #[tokio::test]
7254 4 : async fn test_write_at_initdb_lsn_takes_optimization_code_path() -> anyhow::Result<()> {
7255 4 : let (tenant, ctx) = TenantHarness::create("test_empty_test_timeline_is_usable")
7256 4 : .await?
7257 4 : .load()
7258 4 : .await;
7259 4 :
7260 4 : let initdb_lsn = Lsn(0x20);
7261 4 : let utline = tenant
7262 4 : .create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)
7263 4 : .await?;
7264 4 : let tline = utline.raw_timeline().unwrap();
7265 4 :
7266 4 : // Spawn flush loop now so that we can set the `expect_initdb_optimization`
7267 4 : tline.maybe_spawn_flush_loop();
7268 4 :
7269 4 : // Make sure the timeline has the minimum set of required keys for operation.
7270 4 : // The only operation you can always do on an empty timeline is to `put` new data.
7271 4 : // Except if you `put` at `initdb_lsn`.
7272 4 : // In that case, there's an optimization to directly create image layers instead of delta layers.
7273 4 : // It uses `repartition()`, which assumes some keys to be present.
7274 4 : // Let's make sure the test timeline can handle that case.
7275 4 : {
7276 4 : let mut state = tline.flush_loop_state.lock().unwrap();
7277 4 : assert_eq!(
7278 4 : timeline::FlushLoopState::Running {
7279 4 : expect_initdb_optimization: false,
7280 4 : initdb_optimization_count: 0,
7281 4 : },
7282 4 : *state
7283 4 : );
7284 4 : *state = timeline::FlushLoopState::Running {
7285 4 : expect_initdb_optimization: true,
7286 4 : initdb_optimization_count: 0,
7287 4 : };
7288 4 : }
7289 4 :
7290 4 : // Make writes at the initdb_lsn. When we flush it below, it should be handled by the optimization.
7291 4 : // As explained above, the optimization requires some keys to be present.
7292 4 : // As per `create_empty_timeline` documentation, use init_empty to set them.
7293 4 : // This is what `create_test_timeline` does, by the way.
7294 4 : let mut modification = tline.begin_modification(initdb_lsn);
7295 4 : modification
7296 4 : .init_empty_test_timeline()
7297 4 : .context("init_empty_test_timeline")?;
7298 4 : modification
7299 4 : .commit(&ctx)
7300 4 : .await
7301 4 : .context("commit init_empty_test_timeline modification")?;
7302 4 :
7303 4 : // Do the flush. The flush code will check the expectations that we set above.
7304 4 : tline.freeze_and_flush().await?;
7305 4 :
7306 4 : // assert freeze_and_flush exercised the initdb optimization
7307 4 : {
7308 4 : let state = tline.flush_loop_state.lock().unwrap();
7309 4 : let timeline::FlushLoopState::Running {
7310 4 : expect_initdb_optimization,
7311 4 : initdb_optimization_count,
7312 4 : } = *state
7313 4 : else {
7314 4 : panic!("unexpected state: {:?}", *state);
7315 4 : };
7316 4 : assert!(expect_initdb_optimization);
7317 4 : assert!(initdb_optimization_count > 0);
7318 4 : }
7319 4 : Ok(())
7320 4 : }
7321 :
7322 : #[tokio::test]
7323 4 : async fn test_create_guard_crash() -> anyhow::Result<()> {
7324 4 : let name = "test_create_guard_crash";
7325 4 : let harness = TenantHarness::create(name).await?;
7326 4 : {
7327 4 : let (tenant, ctx) = harness.load().await;
7328 4 : let tline = tenant
7329 4 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
7330 4 : .await?;
7331 4 : // Leave the timeline ID in [`Tenant::timelines_creating`] to exclude attempting to create it again
7332 4 : let raw_tline = tline.raw_timeline().unwrap();
7333 4 : raw_tline
7334 4 : .shutdown(super::timeline::ShutdownMode::Hard)
7335 4 : .instrument(info_span!("test_shutdown", tenant_id=%raw_tline.tenant_shard_id, shard_id=%raw_tline.tenant_shard_id.shard_slug(), timeline_id=%TIMELINE_ID))
7336 4 : .await;
7337 4 : std::mem::forget(tline);
7338 4 : }
7339 4 :
7340 4 : let (tenant, _) = harness.load().await;
7341 4 : match tenant.get_timeline(TIMELINE_ID, false) {
7342 4 : Ok(_) => panic!("timeline should've been removed during load"),
7343 4 : Err(e) => {
7344 4 : assert_eq!(
7345 4 : e,
7346 4 : GetTimelineError::NotFound {
7347 4 : tenant_id: tenant.tenant_shard_id,
7348 4 : timeline_id: TIMELINE_ID,
7349 4 : }
7350 4 : )
7351 4 : }
7352 4 : }
7353 4 :
7354 4 : assert!(!harness
7355 4 : .conf
7356 4 : .timeline_path(&tenant.tenant_shard_id, &TIMELINE_ID)
7357 4 : .exists());
7358 4 :
7359 4 : Ok(())
7360 4 : }
7361 :
7362 : #[tokio::test]
7363 4 : async fn test_read_at_max_lsn() -> anyhow::Result<()> {
7364 4 : let names_algorithms = [
7365 4 : ("test_read_at_max_lsn_legacy", CompactionAlgorithm::Legacy),
7366 4 : ("test_read_at_max_lsn_tiered", CompactionAlgorithm::Tiered),
7367 4 : ];
7368 12 : for (name, algorithm) in names_algorithms {
7369 8 : test_read_at_max_lsn_algorithm(name, algorithm).await?;
7370 4 : }
7371 4 : Ok(())
7372 4 : }
7373 :
7374 8 : async fn test_read_at_max_lsn_algorithm(
7375 8 : name: &'static str,
7376 8 : compaction_algorithm: CompactionAlgorithm,
7377 8 : ) -> anyhow::Result<()> {
7378 8 : let mut harness = TenantHarness::create(name).await?;
7379 8 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
7380 8 : kind: compaction_algorithm,
7381 8 : };
7382 8 : let (tenant, ctx) = harness.load().await;
7383 8 : let tline = tenant
7384 8 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
7385 8 : .await?;
7386 :
7387 8 : let lsn = Lsn(0x10);
7388 8 : let compact = false;
7389 8 : bulk_insert_maybe_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000, compact).await?;
7390 :
7391 8 : let test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7392 8 : let read_lsn = Lsn(u64::MAX - 1);
7393 :
7394 8 : let result = tline.get(test_key, read_lsn, &ctx).await;
7395 8 : assert!(result.is_ok(), "result is not Ok: {}", result.unwrap_err());
7396 :
7397 8 : Ok(())
7398 8 : }
7399 :
7400 : #[tokio::test]
7401 4 : async fn test_metadata_scan() -> anyhow::Result<()> {
7402 4 : let harness = TenantHarness::create("test_metadata_scan").await?;
7403 4 : let (tenant, ctx) = harness.load().await;
7404 4 : let io_concurrency = IoConcurrency::spawn_for_test();
7405 4 : let tline = tenant
7406 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7407 4 : .await?;
7408 4 :
7409 4 : const NUM_KEYS: usize = 1000;
7410 4 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
7411 4 :
7412 4 : let cancel = CancellationToken::new();
7413 4 :
7414 4 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7415 4 : base_key.field1 = AUX_KEY_PREFIX;
7416 4 : let mut test_key = base_key;
7417 4 :
7418 4 : // Track when each page was last modified. Used to assert that
7419 4 : // a read sees the latest page version.
7420 4 : let mut updated = [Lsn(0); NUM_KEYS];
7421 4 :
7422 4 : let mut lsn = Lsn(0x10);
7423 4 : #[allow(clippy::needless_range_loop)]
7424 4004 : for blknum in 0..NUM_KEYS {
7425 4000 : lsn = Lsn(lsn.0 + 0x10);
7426 4000 : test_key.field6 = (blknum * STEP) as u32;
7427 4000 : let mut writer = tline.writer().await;
7428 4000 : writer
7429 4000 : .put(
7430 4000 : test_key,
7431 4000 : lsn,
7432 4000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7433 4000 : &ctx,
7434 4000 : )
7435 4000 : .await?;
7436 4000 : writer.finish_write(lsn);
7437 4000 : updated[blknum] = lsn;
7438 4000 : drop(writer);
7439 4 : }
7440 4 :
7441 4 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
7442 4 :
7443 48 : for iter in 0..=10 {
7444 4 : // Read all the blocks
7445 44000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7446 44000 : test_key.field6 = (blknum * STEP) as u32;
7447 44000 : assert_eq!(
7448 44000 : tline.get(test_key, lsn, &ctx).await?,
7449 44000 : test_img(&format!("{} at {}", blknum, last_lsn))
7450 4 : );
7451 4 : }
7452 4 :
7453 44 : let mut cnt = 0;
7454 44000 : for (key, value) in tline
7455 44 : .get_vectored_impl(
7456 44 : keyspace.clone(),
7457 44 : lsn,
7458 44 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7459 44 : &ctx,
7460 44 : )
7461 44 : .await?
7462 4 : {
7463 44000 : let blknum = key.field6 as usize;
7464 44000 : let value = value?;
7465 44000 : assert!(blknum % STEP == 0);
7466 44000 : let blknum = blknum / STEP;
7467 44000 : assert_eq!(
7468 44000 : value,
7469 44000 : test_img(&format!("{} at {}", blknum, updated[blknum]))
7470 44000 : );
7471 44000 : cnt += 1;
7472 4 : }
7473 4 :
7474 44 : assert_eq!(cnt, NUM_KEYS);
7475 4 :
7476 44044 : for _ in 0..NUM_KEYS {
7477 44000 : lsn = Lsn(lsn.0 + 0x10);
7478 44000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7479 44000 : test_key.field6 = (blknum * STEP) as u32;
7480 44000 : let mut writer = tline.writer().await;
7481 44000 : writer
7482 44000 : .put(
7483 44000 : test_key,
7484 44000 : lsn,
7485 44000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7486 44000 : &ctx,
7487 44000 : )
7488 44000 : .await?;
7489 44000 : writer.finish_write(lsn);
7490 44000 : drop(writer);
7491 44000 : updated[blknum] = lsn;
7492 4 : }
7493 4 :
7494 4 : // Perform two cycles of flush, compact, and GC
7495 132 : for round in 0..2 {
7496 88 : tline.freeze_and_flush().await?;
7497 88 : tline
7498 88 : .compact(
7499 88 : &cancel,
7500 88 : if iter % 5 == 0 && round == 0 {
7501 12 : let mut flags = EnumSet::new();
7502 12 : flags.insert(CompactFlags::ForceImageLayerCreation);
7503 12 : flags.insert(CompactFlags::ForceRepartition);
7504 12 : flags
7505 4 : } else {
7506 76 : EnumSet::empty()
7507 4 : },
7508 88 : &ctx,
7509 88 : )
7510 88 : .await?;
7511 88 : tenant
7512 88 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7513 88 : .await?;
7514 4 : }
7515 4 : }
7516 4 :
7517 4 : Ok(())
7518 4 : }
7519 :
7520 : #[tokio::test]
7521 4 : async fn test_metadata_compaction_trigger() -> anyhow::Result<()> {
7522 4 : let harness = TenantHarness::create("test_metadata_compaction_trigger").await?;
7523 4 : let (tenant, ctx) = harness.load().await;
7524 4 : let tline = tenant
7525 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7526 4 : .await?;
7527 4 :
7528 4 : let cancel = CancellationToken::new();
7529 4 :
7530 4 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7531 4 : base_key.field1 = AUX_KEY_PREFIX;
7532 4 : let test_key = base_key;
7533 4 : let mut lsn = Lsn(0x10);
7534 4 :
7535 84 : for _ in 0..20 {
7536 80 : lsn = Lsn(lsn.0 + 0x10);
7537 80 : let mut writer = tline.writer().await;
7538 80 : writer
7539 80 : .put(
7540 80 : test_key,
7541 80 : lsn,
7542 80 : &Value::Image(test_img(&format!("{} at {}", 0, lsn))),
7543 80 : &ctx,
7544 80 : )
7545 80 : .await?;
7546 80 : writer.finish_write(lsn);
7547 80 : drop(writer);
7548 80 : tline.freeze_and_flush().await?; // force create a delta layer
7549 4 : }
7550 4 :
7551 4 : let before_num_l0_delta_files =
7552 4 : tline.layers.read().await.layer_map()?.level0_deltas().len();
7553 4 :
7554 4 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
7555 4 :
7556 4 : let after_num_l0_delta_files = tline.layers.read().await.layer_map()?.level0_deltas().len();
7557 4 :
7558 4 : assert!(after_num_l0_delta_files < before_num_l0_delta_files, "after_num_l0_delta_files={after_num_l0_delta_files}, before_num_l0_delta_files={before_num_l0_delta_files}");
7559 4 :
7560 4 : assert_eq!(
7561 4 : tline.get(test_key, lsn, &ctx).await?,
7562 4 : test_img(&format!("{} at {}", 0, lsn))
7563 4 : );
7564 4 :
7565 4 : Ok(())
7566 4 : }
7567 :
7568 : #[tokio::test]
7569 4 : async fn test_aux_file_e2e() {
7570 4 : let harness = TenantHarness::create("test_aux_file_e2e").await.unwrap();
7571 4 :
7572 4 : let (tenant, ctx) = harness.load().await;
7573 4 : let io_concurrency = IoConcurrency::spawn_for_test();
7574 4 :
7575 4 : let mut lsn = Lsn(0x08);
7576 4 :
7577 4 : let tline: Arc<Timeline> = tenant
7578 4 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
7579 4 : .await
7580 4 : .unwrap();
7581 4 :
7582 4 : {
7583 4 : lsn += 8;
7584 4 : let mut modification = tline.begin_modification(lsn);
7585 4 : modification
7586 4 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
7587 4 : .await
7588 4 : .unwrap();
7589 4 : modification.commit(&ctx).await.unwrap();
7590 4 : }
7591 4 :
7592 4 : // we can read everything from the storage
7593 4 : let files = tline
7594 4 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
7595 4 : .await
7596 4 : .unwrap();
7597 4 : assert_eq!(
7598 4 : files.get("pg_logical/mappings/test1"),
7599 4 : Some(&bytes::Bytes::from_static(b"first"))
7600 4 : );
7601 4 :
7602 4 : {
7603 4 : lsn += 8;
7604 4 : let mut modification = tline.begin_modification(lsn);
7605 4 : modification
7606 4 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
7607 4 : .await
7608 4 : .unwrap();
7609 4 : modification.commit(&ctx).await.unwrap();
7610 4 : }
7611 4 :
7612 4 : let files = tline
7613 4 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
7614 4 : .await
7615 4 : .unwrap();
7616 4 : assert_eq!(
7617 4 : files.get("pg_logical/mappings/test2"),
7618 4 : Some(&bytes::Bytes::from_static(b"second"))
7619 4 : );
7620 4 :
7621 4 : let child = tenant
7622 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(lsn), &ctx)
7623 4 : .await
7624 4 : .unwrap();
7625 4 :
7626 4 : let files = child
7627 4 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
7628 4 : .await
7629 4 : .unwrap();
7630 4 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
7631 4 : assert_eq!(files.get("pg_logical/mappings/test2"), None);
7632 4 : }
7633 :
7634 : #[tokio::test]
7635 4 : async fn test_metadata_image_creation() -> anyhow::Result<()> {
7636 4 : let harness = TenantHarness::create("test_metadata_image_creation").await?;
7637 4 : let (tenant, ctx) = harness.load().await;
7638 4 : let io_concurrency = IoConcurrency::spawn_for_test();
7639 4 : let tline = tenant
7640 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7641 4 : .await?;
7642 4 :
7643 4 : const NUM_KEYS: usize = 1000;
7644 4 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
7645 4 :
7646 4 : let cancel = CancellationToken::new();
7647 4 :
7648 4 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7649 4 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
7650 4 : let mut test_key = base_key;
7651 4 : let mut lsn = Lsn(0x10);
7652 4 :
7653 16 : async fn scan_with_statistics(
7654 16 : tline: &Timeline,
7655 16 : keyspace: &KeySpace,
7656 16 : lsn: Lsn,
7657 16 : ctx: &RequestContext,
7658 16 : io_concurrency: IoConcurrency,
7659 16 : ) -> anyhow::Result<(BTreeMap<Key, Result<Bytes, PageReconstructError>>, usize)> {
7660 16 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
7661 16 : let res = tline
7662 16 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
7663 16 : .await?;
7664 16 : Ok((res, reconstruct_state.get_delta_layers_visited() as usize))
7665 16 : }
7666 4 :
7667 4 : #[allow(clippy::needless_range_loop)]
7668 4004 : for blknum in 0..NUM_KEYS {
7669 4000 : lsn = Lsn(lsn.0 + 0x10);
7670 4000 : test_key.field6 = (blknum * STEP) as u32;
7671 4000 : let mut writer = tline.writer().await;
7672 4000 : writer
7673 4000 : .put(
7674 4000 : test_key,
7675 4000 : lsn,
7676 4000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7677 4000 : &ctx,
7678 4000 : )
7679 4000 : .await?;
7680 4000 : writer.finish_write(lsn);
7681 4000 : drop(writer);
7682 4 : }
7683 4 :
7684 4 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
7685 4 :
7686 44 : for iter in 1..=10 {
7687 40040 : for _ in 0..NUM_KEYS {
7688 40000 : lsn = Lsn(lsn.0 + 0x10);
7689 40000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7690 40000 : test_key.field6 = (blknum * STEP) as u32;
7691 40000 : let mut writer = tline.writer().await;
7692 40000 : writer
7693 40000 : .put(
7694 40000 : test_key,
7695 40000 : lsn,
7696 40000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7697 40000 : &ctx,
7698 40000 : )
7699 40000 : .await?;
7700 40000 : writer.finish_write(lsn);
7701 40000 : drop(writer);
7702 4 : }
7703 4 :
7704 40 : tline.freeze_and_flush().await?;
7705 4 :
7706 40 : if iter % 5 == 0 {
7707 8 : let (_, before_delta_file_accessed) =
7708 8 : scan_with_statistics(&tline, &keyspace, lsn, &ctx, io_concurrency.clone())
7709 8 : .await?;
7710 8 : tline
7711 8 : .compact(
7712 8 : &cancel,
7713 8 : {
7714 8 : let mut flags = EnumSet::new();
7715 8 : flags.insert(CompactFlags::ForceImageLayerCreation);
7716 8 : flags.insert(CompactFlags::ForceRepartition);
7717 8 : flags
7718 8 : },
7719 8 : &ctx,
7720 8 : )
7721 8 : .await?;
7722 8 : let (_, after_delta_file_accessed) =
7723 8 : scan_with_statistics(&tline, &keyspace, lsn, &ctx, io_concurrency.clone())
7724 8 : .await?;
7725 8 : assert!(after_delta_file_accessed < before_delta_file_accessed, "after_delta_file_accessed={after_delta_file_accessed}, before_delta_file_accessed={before_delta_file_accessed}");
7726 4 : // Given that we already produced an image layer, there should be no delta layer needed for the scan, but still setting a low threshold there for unforeseen circumstances.
7727 8 : assert!(
7728 8 : after_delta_file_accessed <= 2,
7729 4 : "after_delta_file_accessed={after_delta_file_accessed}"
7730 4 : );
7731 32 : }
7732 4 : }
7733 4 :
7734 4 : Ok(())
7735 4 : }
7736 :
7737 : #[tokio::test]
7738 4 : async fn test_vectored_missing_data_key_reads() -> anyhow::Result<()> {
7739 4 : let harness = TenantHarness::create("test_vectored_missing_data_key_reads").await?;
7740 4 : let (tenant, ctx) = harness.load().await;
7741 4 :
7742 4 : let base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7743 4 : let base_key_child = Key::from_hex("000000000033333333444444445500000001").unwrap();
7744 4 : let base_key_nonexist = Key::from_hex("000000000033333333444444445500000002").unwrap();
7745 4 :
7746 4 : let tline = tenant
7747 4 : .create_test_timeline_with_layers(
7748 4 : TIMELINE_ID,
7749 4 : Lsn(0x10),
7750 4 : DEFAULT_PG_VERSION,
7751 4 : &ctx,
7752 4 : Vec::new(), // delta layers
7753 4 : vec![(Lsn(0x20), vec![(base_key, test_img("data key 1"))])], // image layers
7754 4 : Lsn(0x20), // it's fine to not advance LSN to 0x30 while using 0x30 to get below because `get_vectored_impl` does not wait for LSN
7755 4 : )
7756 4 : .await?;
7757 4 : tline.add_extra_test_dense_keyspace(KeySpace::single(base_key..(base_key_nonexist.next())));
7758 4 :
7759 4 : let child = tenant
7760 4 : .branch_timeline_test_with_layers(
7761 4 : &tline,
7762 4 : NEW_TIMELINE_ID,
7763 4 : Some(Lsn(0x20)),
7764 4 : &ctx,
7765 4 : Vec::new(), // delta layers
7766 4 : vec![(Lsn(0x30), vec![(base_key_child, test_img("data key 2"))])], // image layers
7767 4 : Lsn(0x30),
7768 4 : )
7769 4 : .await
7770 4 : .unwrap();
7771 4 :
7772 4 : let lsn = Lsn(0x30);
7773 4 :
7774 4 : // test vectored get on parent timeline
7775 4 : assert_eq!(
7776 4 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
7777 4 : Some(test_img("data key 1"))
7778 4 : );
7779 4 : assert!(get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx)
7780 4 : .await
7781 4 : .unwrap_err()
7782 4 : .is_missing_key_error());
7783 4 : assert!(
7784 4 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx)
7785 4 : .await
7786 4 : .unwrap_err()
7787 4 : .is_missing_key_error()
7788 4 : );
7789 4 :
7790 4 : // test vectored get on child timeline
7791 4 : assert_eq!(
7792 4 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
7793 4 : Some(test_img("data key 1"))
7794 4 : );
7795 4 : assert_eq!(
7796 4 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
7797 4 : Some(test_img("data key 2"))
7798 4 : );
7799 4 : assert!(
7800 4 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx)
7801 4 : .await
7802 4 : .unwrap_err()
7803 4 : .is_missing_key_error()
7804 4 : );
7805 4 :
7806 4 : Ok(())
7807 4 : }
7808 :
7809 : #[tokio::test]
7810 4 : async fn test_vectored_missing_metadata_key_reads() -> anyhow::Result<()> {
7811 4 : let harness = TenantHarness::create("test_vectored_missing_metadata_key_reads").await?;
7812 4 : let (tenant, ctx) = harness.load().await;
7813 4 : let io_concurrency = IoConcurrency::spawn_for_test();
7814 4 :
7815 4 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7816 4 : let base_key_child = Key::from_hex("620000000033333333444444445500000001").unwrap();
7817 4 : let base_key_nonexist = Key::from_hex("620000000033333333444444445500000002").unwrap();
7818 4 : let base_key_overwrite = Key::from_hex("620000000033333333444444445500000003").unwrap();
7819 4 :
7820 4 : let base_inherited_key = Key::from_hex("610000000033333333444444445500000000").unwrap();
7821 4 : let base_inherited_key_child =
7822 4 : Key::from_hex("610000000033333333444444445500000001").unwrap();
7823 4 : let base_inherited_key_nonexist =
7824 4 : Key::from_hex("610000000033333333444444445500000002").unwrap();
7825 4 : let base_inherited_key_overwrite =
7826 4 : Key::from_hex("610000000033333333444444445500000003").unwrap();
7827 4 :
7828 4 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
7829 4 : assert_eq!(base_inherited_key.field1, RELATION_SIZE_PREFIX);
7830 4 :
7831 4 : let tline = tenant
7832 4 : .create_test_timeline_with_layers(
7833 4 : TIMELINE_ID,
7834 4 : Lsn(0x10),
7835 4 : DEFAULT_PG_VERSION,
7836 4 : &ctx,
7837 4 : Vec::new(), // delta layers
7838 4 : vec![(
7839 4 : Lsn(0x20),
7840 4 : vec![
7841 4 : (base_inherited_key, test_img("metadata inherited key 1")),
7842 4 : (
7843 4 : base_inherited_key_overwrite,
7844 4 : test_img("metadata key overwrite 1a"),
7845 4 : ),
7846 4 : (base_key, test_img("metadata key 1")),
7847 4 : (base_key_overwrite, test_img("metadata key overwrite 1b")),
7848 4 : ],
7849 4 : )], // image layers
7850 4 : Lsn(0x20), // it's fine to not advance LSN to 0x30 while using 0x30 to get below because `get_vectored_impl` does not wait for LSN
7851 4 : )
7852 4 : .await?;
7853 4 :
7854 4 : let child = tenant
7855 4 : .branch_timeline_test_with_layers(
7856 4 : &tline,
7857 4 : NEW_TIMELINE_ID,
7858 4 : Some(Lsn(0x20)),
7859 4 : &ctx,
7860 4 : Vec::new(), // delta layers
7861 4 : vec![(
7862 4 : Lsn(0x30),
7863 4 : vec![
7864 4 : (
7865 4 : base_inherited_key_child,
7866 4 : test_img("metadata inherited key 2"),
7867 4 : ),
7868 4 : (
7869 4 : base_inherited_key_overwrite,
7870 4 : test_img("metadata key overwrite 2a"),
7871 4 : ),
7872 4 : (base_key_child, test_img("metadata key 2")),
7873 4 : (base_key_overwrite, test_img("metadata key overwrite 2b")),
7874 4 : ],
7875 4 : )], // image layers
7876 4 : Lsn(0x30),
7877 4 : )
7878 4 : .await
7879 4 : .unwrap();
7880 4 :
7881 4 : let lsn = Lsn(0x30);
7882 4 :
7883 4 : // test vectored get on parent timeline
7884 4 : assert_eq!(
7885 4 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
7886 4 : Some(test_img("metadata key 1"))
7887 4 : );
7888 4 : assert_eq!(
7889 4 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx).await?,
7890 4 : None
7891 4 : );
7892 4 : assert_eq!(
7893 4 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx).await?,
7894 4 : None
7895 4 : );
7896 4 : assert_eq!(
7897 4 : get_vectored_impl_wrapper(&tline, base_key_overwrite, lsn, &ctx).await?,
7898 4 : Some(test_img("metadata key overwrite 1b"))
7899 4 : );
7900 4 : assert_eq!(
7901 4 : get_vectored_impl_wrapper(&tline, base_inherited_key, lsn, &ctx).await?,
7902 4 : Some(test_img("metadata inherited key 1"))
7903 4 : );
7904 4 : assert_eq!(
7905 4 : get_vectored_impl_wrapper(&tline, base_inherited_key_child, lsn, &ctx).await?,
7906 4 : None
7907 4 : );
7908 4 : assert_eq!(
7909 4 : get_vectored_impl_wrapper(&tline, base_inherited_key_nonexist, lsn, &ctx).await?,
7910 4 : None
7911 4 : );
7912 4 : assert_eq!(
7913 4 : get_vectored_impl_wrapper(&tline, base_inherited_key_overwrite, lsn, &ctx).await?,
7914 4 : Some(test_img("metadata key overwrite 1a"))
7915 4 : );
7916 4 :
7917 4 : // test vectored get on child timeline
7918 4 : assert_eq!(
7919 4 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
7920 4 : None
7921 4 : );
7922 4 : assert_eq!(
7923 4 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
7924 4 : Some(test_img("metadata key 2"))
7925 4 : );
7926 4 : assert_eq!(
7927 4 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx).await?,
7928 4 : None
7929 4 : );
7930 4 : assert_eq!(
7931 4 : get_vectored_impl_wrapper(&child, base_inherited_key, lsn, &ctx).await?,
7932 4 : Some(test_img("metadata inherited key 1"))
7933 4 : );
7934 4 : assert_eq!(
7935 4 : get_vectored_impl_wrapper(&child, base_inherited_key_child, lsn, &ctx).await?,
7936 4 : Some(test_img("metadata inherited key 2"))
7937 4 : );
7938 4 : assert_eq!(
7939 4 : get_vectored_impl_wrapper(&child, base_inherited_key_nonexist, lsn, &ctx).await?,
7940 4 : None
7941 4 : );
7942 4 : assert_eq!(
7943 4 : get_vectored_impl_wrapper(&child, base_key_overwrite, lsn, &ctx).await?,
7944 4 : Some(test_img("metadata key overwrite 2b"))
7945 4 : );
7946 4 : assert_eq!(
7947 4 : get_vectored_impl_wrapper(&child, base_inherited_key_overwrite, lsn, &ctx).await?,
7948 4 : Some(test_img("metadata key overwrite 2a"))
7949 4 : );
7950 4 :
7951 4 : // test vectored scan on parent timeline
7952 4 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency.clone());
7953 4 : let res = tline
7954 4 : .get_vectored_impl(
7955 4 : KeySpace::single(Key::metadata_key_range()),
7956 4 : lsn,
7957 4 : &mut reconstruct_state,
7958 4 : &ctx,
7959 4 : )
7960 4 : .await?;
7961 4 :
7962 4 : assert_eq!(
7963 4 : res.into_iter()
7964 16 : .map(|(k, v)| (k, v.unwrap()))
7965 4 : .collect::<Vec<_>>(),
7966 4 : vec![
7967 4 : (base_inherited_key, test_img("metadata inherited key 1")),
7968 4 : (
7969 4 : base_inherited_key_overwrite,
7970 4 : test_img("metadata key overwrite 1a")
7971 4 : ),
7972 4 : (base_key, test_img("metadata key 1")),
7973 4 : (base_key_overwrite, test_img("metadata key overwrite 1b")),
7974 4 : ]
7975 4 : );
7976 4 :
7977 4 : // test vectored scan on child timeline
7978 4 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency.clone());
7979 4 : let res = child
7980 4 : .get_vectored_impl(
7981 4 : KeySpace::single(Key::metadata_key_range()),
7982 4 : lsn,
7983 4 : &mut reconstruct_state,
7984 4 : &ctx,
7985 4 : )
7986 4 : .await?;
7987 4 :
7988 4 : assert_eq!(
7989 4 : res.into_iter()
7990 20 : .map(|(k, v)| (k, v.unwrap()))
7991 4 : .collect::<Vec<_>>(),
7992 4 : vec![
7993 4 : (base_inherited_key, test_img("metadata inherited key 1")),
7994 4 : (
7995 4 : base_inherited_key_child,
7996 4 : test_img("metadata inherited key 2")
7997 4 : ),
7998 4 : (
7999 4 : base_inherited_key_overwrite,
8000 4 : test_img("metadata key overwrite 2a")
8001 4 : ),
8002 4 : (base_key_child, test_img("metadata key 2")),
8003 4 : (base_key_overwrite, test_img("metadata key overwrite 2b")),
8004 4 : ]
8005 4 : );
8006 4 :
8007 4 : Ok(())
8008 4 : }
8009 :
8010 112 : async fn get_vectored_impl_wrapper(
8011 112 : tline: &Arc<Timeline>,
8012 112 : key: Key,
8013 112 : lsn: Lsn,
8014 112 : ctx: &RequestContext,
8015 112 : ) -> Result<Option<Bytes>, GetVectoredError> {
8016 112 : let io_concurrency =
8017 112 : IoConcurrency::spawn_from_conf(tline.conf, tline.gate.enter().unwrap());
8018 112 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
8019 112 : let mut res = tline
8020 112 : .get_vectored_impl(
8021 112 : KeySpace::single(key..key.next()),
8022 112 : lsn,
8023 112 : &mut reconstruct_state,
8024 112 : ctx,
8025 112 : )
8026 112 : .await?;
8027 100 : Ok(res.pop_last().map(|(k, v)| {
8028 64 : assert_eq!(k, key);
8029 64 : v.unwrap()
8030 100 : }))
8031 112 : }
8032 :
8033 : #[tokio::test]
8034 4 : async fn test_metadata_tombstone_reads() -> anyhow::Result<()> {
8035 4 : let harness = TenantHarness::create("test_metadata_tombstone_reads").await?;
8036 4 : let (tenant, ctx) = harness.load().await;
8037 4 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
8038 4 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8039 4 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8040 4 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
8041 4 :
8042 4 : // We emulate the situation that the compaction algorithm creates an image layer that removes the tombstones
8043 4 : // Lsn 0x30 key0, key3, no key1+key2
8044 4 : // Lsn 0x20 key1+key2 tomestones
8045 4 : // Lsn 0x10 key1 in image, key2 in delta
8046 4 : let tline = tenant
8047 4 : .create_test_timeline_with_layers(
8048 4 : TIMELINE_ID,
8049 4 : Lsn(0x10),
8050 4 : DEFAULT_PG_VERSION,
8051 4 : &ctx,
8052 4 : // delta layers
8053 4 : vec![
8054 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8055 4 : Lsn(0x10)..Lsn(0x20),
8056 4 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8057 4 : ),
8058 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8059 4 : Lsn(0x20)..Lsn(0x30),
8060 4 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8061 4 : ),
8062 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8063 4 : Lsn(0x20)..Lsn(0x30),
8064 4 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8065 4 : ),
8066 4 : ],
8067 4 : // image layers
8068 4 : vec![
8069 4 : (Lsn(0x10), vec![(key1, test_img("metadata key 1"))]),
8070 4 : (
8071 4 : Lsn(0x30),
8072 4 : vec![
8073 4 : (key0, test_img("metadata key 0")),
8074 4 : (key3, test_img("metadata key 3")),
8075 4 : ],
8076 4 : ),
8077 4 : ],
8078 4 : Lsn(0x30),
8079 4 : )
8080 4 : .await?;
8081 4 :
8082 4 : let lsn = Lsn(0x30);
8083 4 : let old_lsn = Lsn(0x20);
8084 4 :
8085 4 : assert_eq!(
8086 4 : get_vectored_impl_wrapper(&tline, key0, lsn, &ctx).await?,
8087 4 : Some(test_img("metadata key 0"))
8088 4 : );
8089 4 : assert_eq!(
8090 4 : get_vectored_impl_wrapper(&tline, key1, lsn, &ctx).await?,
8091 4 : None,
8092 4 : );
8093 4 : assert_eq!(
8094 4 : get_vectored_impl_wrapper(&tline, key2, lsn, &ctx).await?,
8095 4 : None,
8096 4 : );
8097 4 : assert_eq!(
8098 4 : get_vectored_impl_wrapper(&tline, key1, old_lsn, &ctx).await?,
8099 4 : Some(Bytes::new()),
8100 4 : );
8101 4 : assert_eq!(
8102 4 : get_vectored_impl_wrapper(&tline, key2, old_lsn, &ctx).await?,
8103 4 : Some(Bytes::new()),
8104 4 : );
8105 4 : assert_eq!(
8106 4 : get_vectored_impl_wrapper(&tline, key3, lsn, &ctx).await?,
8107 4 : Some(test_img("metadata key 3"))
8108 4 : );
8109 4 :
8110 4 : Ok(())
8111 4 : }
8112 :
8113 : #[tokio::test]
8114 4 : async fn test_metadata_tombstone_image_creation() {
8115 4 : let harness = TenantHarness::create("test_metadata_tombstone_image_creation")
8116 4 : .await
8117 4 : .unwrap();
8118 4 : let (tenant, ctx) = harness.load().await;
8119 4 : let io_concurrency = IoConcurrency::spawn_for_test();
8120 4 :
8121 4 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
8122 4 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8123 4 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8124 4 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
8125 4 :
8126 4 : let tline = tenant
8127 4 : .create_test_timeline_with_layers(
8128 4 : TIMELINE_ID,
8129 4 : Lsn(0x10),
8130 4 : DEFAULT_PG_VERSION,
8131 4 : &ctx,
8132 4 : // delta layers
8133 4 : vec![
8134 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8135 4 : Lsn(0x10)..Lsn(0x20),
8136 4 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8137 4 : ),
8138 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8139 4 : Lsn(0x20)..Lsn(0x30),
8140 4 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8141 4 : ),
8142 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8143 4 : Lsn(0x20)..Lsn(0x30),
8144 4 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8145 4 : ),
8146 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8147 4 : Lsn(0x30)..Lsn(0x40),
8148 4 : vec![
8149 4 : (key0, Lsn(0x30), Value::Image(test_img("metadata key 0"))),
8150 4 : (key3, Lsn(0x30), Value::Image(test_img("metadata key 3"))),
8151 4 : ],
8152 4 : ),
8153 4 : ],
8154 4 : // image layers
8155 4 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
8156 4 : Lsn(0x40),
8157 4 : )
8158 4 : .await
8159 4 : .unwrap();
8160 4 :
8161 4 : let cancel = CancellationToken::new();
8162 4 :
8163 4 : tline
8164 4 : .compact(
8165 4 : &cancel,
8166 4 : {
8167 4 : let mut flags = EnumSet::new();
8168 4 : flags.insert(CompactFlags::ForceImageLayerCreation);
8169 4 : flags.insert(CompactFlags::ForceRepartition);
8170 4 : flags
8171 4 : },
8172 4 : &ctx,
8173 4 : )
8174 4 : .await
8175 4 : .unwrap();
8176 4 :
8177 4 : // Image layers are created at last_record_lsn
8178 4 : let images = tline
8179 4 : .inspect_image_layers(Lsn(0x40), &ctx, io_concurrency.clone())
8180 4 : .await
8181 4 : .unwrap()
8182 4 : .into_iter()
8183 36 : .filter(|(k, _)| k.is_metadata_key())
8184 4 : .collect::<Vec<_>>();
8185 4 : assert_eq!(images.len(), 2); // the image layer should only contain two existing keys, tombstones should be removed.
8186 4 : }
8187 :
8188 : #[tokio::test]
8189 4 : async fn test_metadata_tombstone_empty_image_creation() {
8190 4 : let harness = TenantHarness::create("test_metadata_tombstone_empty_image_creation")
8191 4 : .await
8192 4 : .unwrap();
8193 4 : let (tenant, ctx) = harness.load().await;
8194 4 : let io_concurrency = IoConcurrency::spawn_for_test();
8195 4 :
8196 4 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8197 4 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8198 4 :
8199 4 : let tline = tenant
8200 4 : .create_test_timeline_with_layers(
8201 4 : TIMELINE_ID,
8202 4 : Lsn(0x10),
8203 4 : DEFAULT_PG_VERSION,
8204 4 : &ctx,
8205 4 : // delta layers
8206 4 : vec![
8207 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8208 4 : Lsn(0x10)..Lsn(0x20),
8209 4 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8210 4 : ),
8211 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8212 4 : Lsn(0x20)..Lsn(0x30),
8213 4 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8214 4 : ),
8215 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8216 4 : Lsn(0x20)..Lsn(0x30),
8217 4 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8218 4 : ),
8219 4 : ],
8220 4 : // image layers
8221 4 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
8222 4 : Lsn(0x30),
8223 4 : )
8224 4 : .await
8225 4 : .unwrap();
8226 4 :
8227 4 : let cancel = CancellationToken::new();
8228 4 :
8229 4 : tline
8230 4 : .compact(
8231 4 : &cancel,
8232 4 : {
8233 4 : let mut flags = EnumSet::new();
8234 4 : flags.insert(CompactFlags::ForceImageLayerCreation);
8235 4 : flags.insert(CompactFlags::ForceRepartition);
8236 4 : flags
8237 4 : },
8238 4 : &ctx,
8239 4 : )
8240 4 : .await
8241 4 : .unwrap();
8242 4 :
8243 4 : // Image layers are created at last_record_lsn
8244 4 : let images = tline
8245 4 : .inspect_image_layers(Lsn(0x30), &ctx, io_concurrency.clone())
8246 4 : .await
8247 4 : .unwrap()
8248 4 : .into_iter()
8249 28 : .filter(|(k, _)| k.is_metadata_key())
8250 4 : .collect::<Vec<_>>();
8251 4 : assert_eq!(images.len(), 0); // the image layer should not contain tombstones, or it is not created
8252 4 : }
8253 :
8254 : #[tokio::test]
8255 4 : async fn test_simple_bottom_most_compaction_images() -> anyhow::Result<()> {
8256 4 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_images").await?;
8257 4 : let (tenant, ctx) = harness.load().await;
8258 4 : let io_concurrency = IoConcurrency::spawn_for_test();
8259 4 :
8260 204 : fn get_key(id: u32) -> Key {
8261 204 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8262 204 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8263 204 : key.field6 = id;
8264 204 : key
8265 204 : }
8266 4 :
8267 4 : // We create
8268 4 : // - one bottom-most image layer,
8269 4 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
8270 4 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
8271 4 : // - a delta layer D3 above the horizon.
8272 4 : //
8273 4 : // | D3 |
8274 4 : // | D1 |
8275 4 : // -| |-- gc horizon -----------------
8276 4 : // | | | D2 |
8277 4 : // --------- img layer ------------------
8278 4 : //
8279 4 : // What we should expact from this compaction is:
8280 4 : // | D3 |
8281 4 : // | Part of D1 |
8282 4 : // --------- img layer with D1+D2 at GC horizon------------------
8283 4 :
8284 4 : // img layer at 0x10
8285 4 : let img_layer = (0..10)
8286 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8287 4 : .collect_vec();
8288 4 :
8289 4 : let delta1 = vec![
8290 4 : (
8291 4 : get_key(1),
8292 4 : Lsn(0x20),
8293 4 : Value::Image(Bytes::from("value 1@0x20")),
8294 4 : ),
8295 4 : (
8296 4 : get_key(2),
8297 4 : Lsn(0x30),
8298 4 : Value::Image(Bytes::from("value 2@0x30")),
8299 4 : ),
8300 4 : (
8301 4 : get_key(3),
8302 4 : Lsn(0x40),
8303 4 : Value::Image(Bytes::from("value 3@0x40")),
8304 4 : ),
8305 4 : ];
8306 4 : let delta2 = vec![
8307 4 : (
8308 4 : get_key(5),
8309 4 : Lsn(0x20),
8310 4 : Value::Image(Bytes::from("value 5@0x20")),
8311 4 : ),
8312 4 : (
8313 4 : get_key(6),
8314 4 : Lsn(0x20),
8315 4 : Value::Image(Bytes::from("value 6@0x20")),
8316 4 : ),
8317 4 : ];
8318 4 : let delta3 = vec![
8319 4 : (
8320 4 : get_key(8),
8321 4 : Lsn(0x48),
8322 4 : Value::Image(Bytes::from("value 8@0x48")),
8323 4 : ),
8324 4 : (
8325 4 : get_key(9),
8326 4 : Lsn(0x48),
8327 4 : Value::Image(Bytes::from("value 9@0x48")),
8328 4 : ),
8329 4 : ];
8330 4 :
8331 4 : let tline = tenant
8332 4 : .create_test_timeline_with_layers(
8333 4 : TIMELINE_ID,
8334 4 : Lsn(0x10),
8335 4 : DEFAULT_PG_VERSION,
8336 4 : &ctx,
8337 4 : vec![
8338 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
8339 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
8340 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
8341 4 : ], // delta layers
8342 4 : vec![(Lsn(0x10), img_layer)], // image layers
8343 4 : Lsn(0x50),
8344 4 : )
8345 4 : .await?;
8346 4 : {
8347 4 : tline
8348 4 : .latest_gc_cutoff_lsn
8349 4 : .lock_for_write()
8350 4 : .store_and_unlock(Lsn(0x30))
8351 4 : .wait()
8352 4 : .await;
8353 4 : // Update GC info
8354 4 : let mut guard = tline.gc_info.write().unwrap();
8355 4 : guard.cutoffs.time = Lsn(0x30);
8356 4 : guard.cutoffs.space = Lsn(0x30);
8357 4 : }
8358 4 :
8359 4 : let expected_result = [
8360 4 : Bytes::from_static(b"value 0@0x10"),
8361 4 : Bytes::from_static(b"value 1@0x20"),
8362 4 : Bytes::from_static(b"value 2@0x30"),
8363 4 : Bytes::from_static(b"value 3@0x40"),
8364 4 : Bytes::from_static(b"value 4@0x10"),
8365 4 : Bytes::from_static(b"value 5@0x20"),
8366 4 : Bytes::from_static(b"value 6@0x20"),
8367 4 : Bytes::from_static(b"value 7@0x10"),
8368 4 : Bytes::from_static(b"value 8@0x48"),
8369 4 : Bytes::from_static(b"value 9@0x48"),
8370 4 : ];
8371 4 :
8372 40 : for (idx, expected) in expected_result.iter().enumerate() {
8373 40 : assert_eq!(
8374 40 : tline
8375 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8376 40 : .await
8377 40 : .unwrap(),
8378 4 : expected
8379 4 : );
8380 4 : }
8381 4 :
8382 4 : let cancel = CancellationToken::new();
8383 4 : tline
8384 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8385 4 : .await
8386 4 : .unwrap();
8387 4 :
8388 40 : for (idx, expected) in expected_result.iter().enumerate() {
8389 40 : assert_eq!(
8390 40 : tline
8391 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8392 40 : .await
8393 40 : .unwrap(),
8394 4 : expected
8395 4 : );
8396 4 : }
8397 4 :
8398 4 : // Check if the image layer at the GC horizon contains exactly what we want
8399 4 : let image_at_gc_horizon = tline
8400 4 : .inspect_image_layers(Lsn(0x30), &ctx, io_concurrency.clone())
8401 4 : .await
8402 4 : .unwrap()
8403 4 : .into_iter()
8404 68 : .filter(|(k, _)| k.is_metadata_key())
8405 4 : .collect::<Vec<_>>();
8406 4 :
8407 4 : assert_eq!(image_at_gc_horizon.len(), 10);
8408 4 : let expected_result = [
8409 4 : Bytes::from_static(b"value 0@0x10"),
8410 4 : Bytes::from_static(b"value 1@0x20"),
8411 4 : Bytes::from_static(b"value 2@0x30"),
8412 4 : Bytes::from_static(b"value 3@0x10"),
8413 4 : Bytes::from_static(b"value 4@0x10"),
8414 4 : Bytes::from_static(b"value 5@0x20"),
8415 4 : Bytes::from_static(b"value 6@0x20"),
8416 4 : Bytes::from_static(b"value 7@0x10"),
8417 4 : Bytes::from_static(b"value 8@0x10"),
8418 4 : Bytes::from_static(b"value 9@0x10"),
8419 4 : ];
8420 44 : for idx in 0..10 {
8421 40 : assert_eq!(
8422 40 : image_at_gc_horizon[idx],
8423 40 : (get_key(idx as u32), expected_result[idx].clone())
8424 40 : );
8425 4 : }
8426 4 :
8427 4 : // Check if old layers are removed / new layers have the expected LSN
8428 4 : let all_layers = inspect_and_sort(&tline, None).await;
8429 4 : assert_eq!(
8430 4 : all_layers,
8431 4 : vec![
8432 4 : // Image layer at GC horizon
8433 4 : PersistentLayerKey {
8434 4 : key_range: Key::MIN..Key::MAX,
8435 4 : lsn_range: Lsn(0x30)..Lsn(0x31),
8436 4 : is_delta: false
8437 4 : },
8438 4 : // The delta layer below the horizon
8439 4 : PersistentLayerKey {
8440 4 : key_range: get_key(3)..get_key(4),
8441 4 : lsn_range: Lsn(0x30)..Lsn(0x48),
8442 4 : is_delta: true
8443 4 : },
8444 4 : // The delta3 layer that should not be picked for the compaction
8445 4 : PersistentLayerKey {
8446 4 : key_range: get_key(8)..get_key(10),
8447 4 : lsn_range: Lsn(0x48)..Lsn(0x50),
8448 4 : is_delta: true
8449 4 : }
8450 4 : ]
8451 4 : );
8452 4 :
8453 4 : // increase GC horizon and compact again
8454 4 : {
8455 4 : tline
8456 4 : .latest_gc_cutoff_lsn
8457 4 : .lock_for_write()
8458 4 : .store_and_unlock(Lsn(0x40))
8459 4 : .wait()
8460 4 : .await;
8461 4 : // Update GC info
8462 4 : let mut guard = tline.gc_info.write().unwrap();
8463 4 : guard.cutoffs.time = Lsn(0x40);
8464 4 : guard.cutoffs.space = Lsn(0x40);
8465 4 : }
8466 4 : tline
8467 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8468 4 : .await
8469 4 : .unwrap();
8470 4 :
8471 4 : Ok(())
8472 4 : }
8473 :
8474 : #[cfg(feature = "testing")]
8475 : #[tokio::test]
8476 4 : async fn test_neon_test_record() -> anyhow::Result<()> {
8477 4 : let harness = TenantHarness::create("test_neon_test_record").await?;
8478 4 : let (tenant, ctx) = harness.load().await;
8479 4 :
8480 48 : fn get_key(id: u32) -> Key {
8481 48 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8482 48 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8483 48 : key.field6 = id;
8484 48 : key
8485 48 : }
8486 4 :
8487 4 : let delta1 = vec![
8488 4 : (
8489 4 : get_key(1),
8490 4 : Lsn(0x20),
8491 4 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
8492 4 : ),
8493 4 : (
8494 4 : get_key(1),
8495 4 : Lsn(0x30),
8496 4 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
8497 4 : ),
8498 4 : (get_key(2), Lsn(0x10), Value::Image("0x10".into())),
8499 4 : (
8500 4 : get_key(2),
8501 4 : Lsn(0x20),
8502 4 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
8503 4 : ),
8504 4 : (
8505 4 : get_key(2),
8506 4 : Lsn(0x30),
8507 4 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
8508 4 : ),
8509 4 : (get_key(3), Lsn(0x10), Value::Image("0x10".into())),
8510 4 : (
8511 4 : get_key(3),
8512 4 : Lsn(0x20),
8513 4 : Value::WalRecord(NeonWalRecord::wal_clear("c")),
8514 4 : ),
8515 4 : (get_key(4), Lsn(0x10), Value::Image("0x10".into())),
8516 4 : (
8517 4 : get_key(4),
8518 4 : Lsn(0x20),
8519 4 : Value::WalRecord(NeonWalRecord::wal_init("i")),
8520 4 : ),
8521 4 : ];
8522 4 : let image1 = vec![(get_key(1), "0x10".into())];
8523 4 :
8524 4 : let tline = tenant
8525 4 : .create_test_timeline_with_layers(
8526 4 : TIMELINE_ID,
8527 4 : Lsn(0x10),
8528 4 : DEFAULT_PG_VERSION,
8529 4 : &ctx,
8530 4 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
8531 4 : Lsn(0x10)..Lsn(0x40),
8532 4 : delta1,
8533 4 : )], // delta layers
8534 4 : vec![(Lsn(0x10), image1)], // image layers
8535 4 : Lsn(0x50),
8536 4 : )
8537 4 : .await?;
8538 4 :
8539 4 : assert_eq!(
8540 4 : tline.get(get_key(1), Lsn(0x50), &ctx).await?,
8541 4 : Bytes::from_static(b"0x10,0x20,0x30")
8542 4 : );
8543 4 : assert_eq!(
8544 4 : tline.get(get_key(2), Lsn(0x50), &ctx).await?,
8545 4 : Bytes::from_static(b"0x10,0x20,0x30")
8546 4 : );
8547 4 :
8548 4 : // Need to remove the limit of "Neon WAL redo requires base image".
8549 4 :
8550 4 : // assert_eq!(tline.get(get_key(3), Lsn(0x50), &ctx).await?, Bytes::new());
8551 4 : // assert_eq!(tline.get(get_key(4), Lsn(0x50), &ctx).await?, Bytes::new());
8552 4 :
8553 4 : Ok(())
8554 4 : }
8555 :
8556 : #[tokio::test(start_paused = true)]
8557 4 : async fn test_lsn_lease() -> anyhow::Result<()> {
8558 4 : let (tenant, ctx) = TenantHarness::create("test_lsn_lease")
8559 4 : .await
8560 4 : .unwrap()
8561 4 : .load()
8562 4 : .await;
8563 4 : // Advance to the lsn lease deadline so that GC is not blocked by
8564 4 : // initial transition into AttachedSingle.
8565 4 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
8566 4 : tokio::time::resume();
8567 4 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
8568 4 :
8569 4 : let end_lsn = Lsn(0x100);
8570 4 : let image_layers = (0x20..=0x90)
8571 4 : .step_by(0x10)
8572 32 : .map(|n| {
8573 32 : (
8574 32 : Lsn(n),
8575 32 : vec![(key, test_img(&format!("data key at {:x}", n)))],
8576 32 : )
8577 32 : })
8578 4 : .collect();
8579 4 :
8580 4 : let timeline = tenant
8581 4 : .create_test_timeline_with_layers(
8582 4 : TIMELINE_ID,
8583 4 : Lsn(0x10),
8584 4 : DEFAULT_PG_VERSION,
8585 4 : &ctx,
8586 4 : Vec::new(),
8587 4 : image_layers,
8588 4 : end_lsn,
8589 4 : )
8590 4 : .await?;
8591 4 :
8592 4 : let leased_lsns = [0x30, 0x50, 0x70];
8593 4 : let mut leases = Vec::new();
8594 12 : leased_lsns.iter().for_each(|n| {
8595 12 : leases.push(
8596 12 : timeline
8597 12 : .init_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)
8598 12 : .expect("lease request should succeed"),
8599 12 : );
8600 12 : });
8601 4 :
8602 4 : let updated_lease_0 = timeline
8603 4 : .renew_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)
8604 4 : .expect("lease renewal should succeed");
8605 4 : assert_eq!(
8606 4 : updated_lease_0.valid_until, leases[0].valid_until,
8607 4 : " Renewing with shorter lease should not change the lease."
8608 4 : );
8609 4 :
8610 4 : let updated_lease_1 = timeline
8611 4 : .renew_lsn_lease(
8612 4 : Lsn(leased_lsns[1]),
8613 4 : timeline.get_lsn_lease_length() * 2,
8614 4 : &ctx,
8615 4 : )
8616 4 : .expect("lease renewal should succeed");
8617 4 : assert!(
8618 4 : updated_lease_1.valid_until > leases[1].valid_until,
8619 4 : "Renewing with a long lease should renew lease with later expiration time."
8620 4 : );
8621 4 :
8622 4 : // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
8623 4 : info!(
8624 4 : "latest_gc_cutoff_lsn: {}",
8625 0 : *timeline.get_latest_gc_cutoff_lsn()
8626 4 : );
8627 4 : timeline.force_set_disk_consistent_lsn(end_lsn);
8628 4 :
8629 4 : let res = tenant
8630 4 : .gc_iteration(
8631 4 : Some(TIMELINE_ID),
8632 4 : 0,
8633 4 : Duration::ZERO,
8634 4 : &CancellationToken::new(),
8635 4 : &ctx,
8636 4 : )
8637 4 : .await
8638 4 : .unwrap();
8639 4 :
8640 4 : // Keeping everything <= Lsn(0x80) b/c leases:
8641 4 : // 0/10: initdb layer
8642 4 : // (0/20..=0/70).step_by(0x10): image layers added when creating the timeline.
8643 4 : assert_eq!(res.layers_needed_by_leases, 7);
8644 4 : // Keeping 0/90 b/c it is the latest layer.
8645 4 : assert_eq!(res.layers_not_updated, 1);
8646 4 : // Removed 0/80.
8647 4 : assert_eq!(res.layers_removed, 1);
8648 4 :
8649 4 : // Make lease on a already GC-ed LSN.
8650 4 : // 0/80 does not have a valid lease + is below latest_gc_cutoff
8651 4 : assert!(Lsn(0x80) < *timeline.get_latest_gc_cutoff_lsn());
8652 4 : timeline
8653 4 : .init_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx)
8654 4 : .expect_err("lease request on GC-ed LSN should fail");
8655 4 :
8656 4 : // Should still be able to renew a currently valid lease
8657 4 : // Assumption: original lease to is still valid for 0/50.
8658 4 : // (use `Timeline::init_lsn_lease` for testing so it always does validation)
8659 4 : timeline
8660 4 : .init_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)
8661 4 : .expect("lease renewal with validation should succeed");
8662 4 :
8663 4 : Ok(())
8664 4 : }
8665 :
8666 : #[cfg(feature = "testing")]
8667 : #[tokio::test]
8668 4 : async fn test_simple_bottom_most_compaction_deltas_1() -> anyhow::Result<()> {
8669 4 : test_simple_bottom_most_compaction_deltas_helper(
8670 4 : "test_simple_bottom_most_compaction_deltas_1",
8671 4 : false,
8672 4 : )
8673 4 : .await
8674 4 : }
8675 :
8676 : #[cfg(feature = "testing")]
8677 : #[tokio::test]
8678 4 : async fn test_simple_bottom_most_compaction_deltas_2() -> anyhow::Result<()> {
8679 4 : test_simple_bottom_most_compaction_deltas_helper(
8680 4 : "test_simple_bottom_most_compaction_deltas_2",
8681 4 : true,
8682 4 : )
8683 4 : .await
8684 4 : }
8685 :
8686 : #[cfg(feature = "testing")]
8687 8 : async fn test_simple_bottom_most_compaction_deltas_helper(
8688 8 : test_name: &'static str,
8689 8 : use_delta_bottom_layer: bool,
8690 8 : ) -> anyhow::Result<()> {
8691 8 : let harness = TenantHarness::create(test_name).await?;
8692 8 : let (tenant, ctx) = harness.load().await;
8693 :
8694 552 : fn get_key(id: u32) -> Key {
8695 552 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8696 552 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8697 552 : key.field6 = id;
8698 552 : key
8699 552 : }
8700 :
8701 : // We create
8702 : // - one bottom-most image layer,
8703 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
8704 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
8705 : // - a delta layer D3 above the horizon.
8706 : //
8707 : // | D3 |
8708 : // | D1 |
8709 : // -| |-- gc horizon -----------------
8710 : // | | | D2 |
8711 : // --------- img layer ------------------
8712 : //
8713 : // What we should expact from this compaction is:
8714 : // | D3 |
8715 : // | Part of D1 |
8716 : // --------- img layer with D1+D2 at GC horizon------------------
8717 :
8718 : // img layer at 0x10
8719 8 : let img_layer = (0..10)
8720 80 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8721 8 : .collect_vec();
8722 8 : // or, delta layer at 0x10 if `use_delta_bottom_layer` is true
8723 8 : let delta4 = (0..10)
8724 80 : .map(|id| {
8725 80 : (
8726 80 : get_key(id),
8727 80 : Lsn(0x08),
8728 80 : Value::WalRecord(NeonWalRecord::wal_init(format!("value {id}@0x10"))),
8729 80 : )
8730 80 : })
8731 8 : .collect_vec();
8732 8 :
8733 8 : let delta1 = vec![
8734 8 : (
8735 8 : get_key(1),
8736 8 : Lsn(0x20),
8737 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8738 8 : ),
8739 8 : (
8740 8 : get_key(2),
8741 8 : Lsn(0x30),
8742 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8743 8 : ),
8744 8 : (
8745 8 : get_key(3),
8746 8 : Lsn(0x28),
8747 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8748 8 : ),
8749 8 : (
8750 8 : get_key(3),
8751 8 : Lsn(0x30),
8752 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8753 8 : ),
8754 8 : (
8755 8 : get_key(3),
8756 8 : Lsn(0x40),
8757 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
8758 8 : ),
8759 8 : ];
8760 8 : let delta2 = vec![
8761 8 : (
8762 8 : get_key(5),
8763 8 : Lsn(0x20),
8764 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8765 8 : ),
8766 8 : (
8767 8 : get_key(6),
8768 8 : Lsn(0x20),
8769 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8770 8 : ),
8771 8 : ];
8772 8 : let delta3 = vec![
8773 8 : (
8774 8 : get_key(8),
8775 8 : Lsn(0x48),
8776 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8777 8 : ),
8778 8 : (
8779 8 : get_key(9),
8780 8 : Lsn(0x48),
8781 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8782 8 : ),
8783 8 : ];
8784 :
8785 8 : let tline = if use_delta_bottom_layer {
8786 4 : tenant
8787 4 : .create_test_timeline_with_layers(
8788 4 : TIMELINE_ID,
8789 4 : Lsn(0x08),
8790 4 : DEFAULT_PG_VERSION,
8791 4 : &ctx,
8792 4 : vec![
8793 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8794 4 : Lsn(0x08)..Lsn(0x10),
8795 4 : delta4,
8796 4 : ),
8797 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8798 4 : Lsn(0x20)..Lsn(0x48),
8799 4 : delta1,
8800 4 : ),
8801 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8802 4 : Lsn(0x20)..Lsn(0x48),
8803 4 : delta2,
8804 4 : ),
8805 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8806 4 : Lsn(0x48)..Lsn(0x50),
8807 4 : delta3,
8808 4 : ),
8809 4 : ], // delta layers
8810 4 : vec![], // image layers
8811 4 : Lsn(0x50),
8812 4 : )
8813 4 : .await?
8814 : } else {
8815 4 : tenant
8816 4 : .create_test_timeline_with_layers(
8817 4 : TIMELINE_ID,
8818 4 : Lsn(0x10),
8819 4 : DEFAULT_PG_VERSION,
8820 4 : &ctx,
8821 4 : vec![
8822 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8823 4 : Lsn(0x10)..Lsn(0x48),
8824 4 : delta1,
8825 4 : ),
8826 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8827 4 : Lsn(0x10)..Lsn(0x48),
8828 4 : delta2,
8829 4 : ),
8830 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8831 4 : Lsn(0x48)..Lsn(0x50),
8832 4 : delta3,
8833 4 : ),
8834 4 : ], // delta layers
8835 4 : vec![(Lsn(0x10), img_layer)], // image layers
8836 4 : Lsn(0x50),
8837 4 : )
8838 4 : .await?
8839 : };
8840 : {
8841 8 : tline
8842 8 : .latest_gc_cutoff_lsn
8843 8 : .lock_for_write()
8844 8 : .store_and_unlock(Lsn(0x30))
8845 8 : .wait()
8846 8 : .await;
8847 : // Update GC info
8848 8 : let mut guard = tline.gc_info.write().unwrap();
8849 8 : *guard = GcInfo {
8850 8 : retain_lsns: vec![],
8851 8 : cutoffs: GcCutoffs {
8852 8 : time: Lsn(0x30),
8853 8 : space: Lsn(0x30),
8854 8 : },
8855 8 : leases: Default::default(),
8856 8 : within_ancestor_pitr: false,
8857 8 : };
8858 8 : }
8859 8 :
8860 8 : let expected_result = [
8861 8 : Bytes::from_static(b"value 0@0x10"),
8862 8 : Bytes::from_static(b"value 1@0x10@0x20"),
8863 8 : Bytes::from_static(b"value 2@0x10@0x30"),
8864 8 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
8865 8 : Bytes::from_static(b"value 4@0x10"),
8866 8 : Bytes::from_static(b"value 5@0x10@0x20"),
8867 8 : Bytes::from_static(b"value 6@0x10@0x20"),
8868 8 : Bytes::from_static(b"value 7@0x10"),
8869 8 : Bytes::from_static(b"value 8@0x10@0x48"),
8870 8 : Bytes::from_static(b"value 9@0x10@0x48"),
8871 8 : ];
8872 8 :
8873 8 : let expected_result_at_gc_horizon = [
8874 8 : Bytes::from_static(b"value 0@0x10"),
8875 8 : Bytes::from_static(b"value 1@0x10@0x20"),
8876 8 : Bytes::from_static(b"value 2@0x10@0x30"),
8877 8 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
8878 8 : Bytes::from_static(b"value 4@0x10"),
8879 8 : Bytes::from_static(b"value 5@0x10@0x20"),
8880 8 : Bytes::from_static(b"value 6@0x10@0x20"),
8881 8 : Bytes::from_static(b"value 7@0x10"),
8882 8 : Bytes::from_static(b"value 8@0x10"),
8883 8 : Bytes::from_static(b"value 9@0x10"),
8884 8 : ];
8885 :
8886 88 : for idx in 0..10 {
8887 80 : assert_eq!(
8888 80 : tline
8889 80 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8890 80 : .await
8891 80 : .unwrap(),
8892 80 : &expected_result[idx]
8893 : );
8894 80 : assert_eq!(
8895 80 : tline
8896 80 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
8897 80 : .await
8898 80 : .unwrap(),
8899 80 : &expected_result_at_gc_horizon[idx]
8900 : );
8901 : }
8902 :
8903 8 : let cancel = CancellationToken::new();
8904 8 : tline
8905 8 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8906 8 : .await
8907 8 : .unwrap();
8908 :
8909 88 : for idx in 0..10 {
8910 80 : assert_eq!(
8911 80 : tline
8912 80 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8913 80 : .await
8914 80 : .unwrap(),
8915 80 : &expected_result[idx]
8916 : );
8917 80 : assert_eq!(
8918 80 : tline
8919 80 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
8920 80 : .await
8921 80 : .unwrap(),
8922 80 : &expected_result_at_gc_horizon[idx]
8923 : );
8924 : }
8925 :
8926 : // increase GC horizon and compact again
8927 : {
8928 8 : tline
8929 8 : .latest_gc_cutoff_lsn
8930 8 : .lock_for_write()
8931 8 : .store_and_unlock(Lsn(0x40))
8932 8 : .wait()
8933 8 : .await;
8934 : // Update GC info
8935 8 : let mut guard = tline.gc_info.write().unwrap();
8936 8 : guard.cutoffs.time = Lsn(0x40);
8937 8 : guard.cutoffs.space = Lsn(0x40);
8938 8 : }
8939 8 : tline
8940 8 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8941 8 : .await
8942 8 : .unwrap();
8943 8 :
8944 8 : Ok(())
8945 8 : }
8946 :
8947 : #[cfg(feature = "testing")]
8948 : #[tokio::test]
8949 4 : async fn test_generate_key_retention() -> anyhow::Result<()> {
8950 4 : let harness = TenantHarness::create("test_generate_key_retention").await?;
8951 4 : let (tenant, ctx) = harness.load().await;
8952 4 : let tline = tenant
8953 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
8954 4 : .await?;
8955 4 : tline.force_advance_lsn(Lsn(0x70));
8956 4 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
8957 4 : let history = vec![
8958 4 : (
8959 4 : key,
8960 4 : Lsn(0x10),
8961 4 : Value::WalRecord(NeonWalRecord::wal_init("0x10")),
8962 4 : ),
8963 4 : (
8964 4 : key,
8965 4 : Lsn(0x20),
8966 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8967 4 : ),
8968 4 : (
8969 4 : key,
8970 4 : Lsn(0x30),
8971 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8972 4 : ),
8973 4 : (
8974 4 : key,
8975 4 : Lsn(0x40),
8976 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8977 4 : ),
8978 4 : (
8979 4 : key,
8980 4 : Lsn(0x50),
8981 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
8982 4 : ),
8983 4 : (
8984 4 : key,
8985 4 : Lsn(0x60),
8986 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8987 4 : ),
8988 4 : (
8989 4 : key,
8990 4 : Lsn(0x70),
8991 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8992 4 : ),
8993 4 : (
8994 4 : key,
8995 4 : Lsn(0x80),
8996 4 : Value::Image(Bytes::copy_from_slice(
8997 4 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8998 4 : )),
8999 4 : ),
9000 4 : (
9001 4 : key,
9002 4 : Lsn(0x90),
9003 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9004 4 : ),
9005 4 : ];
9006 4 : let res = tline
9007 4 : .generate_key_retention(
9008 4 : key,
9009 4 : &history,
9010 4 : Lsn(0x60),
9011 4 : &[Lsn(0x20), Lsn(0x40), Lsn(0x50)],
9012 4 : 3,
9013 4 : None,
9014 4 : )
9015 4 : .await
9016 4 : .unwrap();
9017 4 : let expected_res = KeyHistoryRetention {
9018 4 : below_horizon: vec![
9019 4 : (
9020 4 : Lsn(0x20),
9021 4 : KeyLogAtLsn(vec![(
9022 4 : Lsn(0x20),
9023 4 : Value::Image(Bytes::from_static(b"0x10;0x20")),
9024 4 : )]),
9025 4 : ),
9026 4 : (
9027 4 : Lsn(0x40),
9028 4 : KeyLogAtLsn(vec![
9029 4 : (
9030 4 : Lsn(0x30),
9031 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9032 4 : ),
9033 4 : (
9034 4 : Lsn(0x40),
9035 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9036 4 : ),
9037 4 : ]),
9038 4 : ),
9039 4 : (
9040 4 : Lsn(0x50),
9041 4 : KeyLogAtLsn(vec![(
9042 4 : Lsn(0x50),
9043 4 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40;0x50")),
9044 4 : )]),
9045 4 : ),
9046 4 : (
9047 4 : Lsn(0x60),
9048 4 : KeyLogAtLsn(vec![(
9049 4 : Lsn(0x60),
9050 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9051 4 : )]),
9052 4 : ),
9053 4 : ],
9054 4 : above_horizon: KeyLogAtLsn(vec![
9055 4 : (
9056 4 : Lsn(0x70),
9057 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9058 4 : ),
9059 4 : (
9060 4 : Lsn(0x80),
9061 4 : Value::Image(Bytes::copy_from_slice(
9062 4 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9063 4 : )),
9064 4 : ),
9065 4 : (
9066 4 : Lsn(0x90),
9067 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9068 4 : ),
9069 4 : ]),
9070 4 : };
9071 4 : assert_eq!(res, expected_res);
9072 4 :
9073 4 : // We expect GC-compaction to run with the original GC. This would create a situation that
9074 4 : // the original GC algorithm removes some delta layers b/c there are full image coverage,
9075 4 : // therefore causing some keys to have an incomplete history below the lowest retain LSN.
9076 4 : // For example, we have
9077 4 : // ```plain
9078 4 : // init delta @ 0x10, image @ 0x20, delta @ 0x30 (gc_horizon), image @ 0x40.
9079 4 : // ```
9080 4 : // Now the GC horizon moves up, and we have
9081 4 : // ```plain
9082 4 : // init delta @ 0x10, image @ 0x20, delta @ 0x30, image @ 0x40 (gc_horizon)
9083 4 : // ```
9084 4 : // The original GC algorithm kicks in, and removes delta @ 0x10, image @ 0x20.
9085 4 : // We will end up with
9086 4 : // ```plain
9087 4 : // delta @ 0x30, image @ 0x40 (gc_horizon)
9088 4 : // ```
9089 4 : // Now we run the GC-compaction, and this key does not have a full history.
9090 4 : // We should be able to handle this partial history and drop everything before the
9091 4 : // gc_horizon image.
9092 4 :
9093 4 : let history = vec![
9094 4 : (
9095 4 : key,
9096 4 : Lsn(0x20),
9097 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9098 4 : ),
9099 4 : (
9100 4 : key,
9101 4 : Lsn(0x30),
9102 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9103 4 : ),
9104 4 : (
9105 4 : key,
9106 4 : Lsn(0x40),
9107 4 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
9108 4 : ),
9109 4 : (
9110 4 : key,
9111 4 : Lsn(0x50),
9112 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
9113 4 : ),
9114 4 : (
9115 4 : key,
9116 4 : Lsn(0x60),
9117 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9118 4 : ),
9119 4 : (
9120 4 : key,
9121 4 : Lsn(0x70),
9122 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9123 4 : ),
9124 4 : (
9125 4 : key,
9126 4 : Lsn(0x80),
9127 4 : Value::Image(Bytes::copy_from_slice(
9128 4 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9129 4 : )),
9130 4 : ),
9131 4 : (
9132 4 : key,
9133 4 : Lsn(0x90),
9134 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9135 4 : ),
9136 4 : ];
9137 4 : let res = tline
9138 4 : .generate_key_retention(key, &history, Lsn(0x60), &[Lsn(0x40), Lsn(0x50)], 3, None)
9139 4 : .await
9140 4 : .unwrap();
9141 4 : let expected_res = KeyHistoryRetention {
9142 4 : below_horizon: vec![
9143 4 : (
9144 4 : Lsn(0x40),
9145 4 : KeyLogAtLsn(vec![(
9146 4 : Lsn(0x40),
9147 4 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
9148 4 : )]),
9149 4 : ),
9150 4 : (
9151 4 : Lsn(0x50),
9152 4 : KeyLogAtLsn(vec![(
9153 4 : Lsn(0x50),
9154 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
9155 4 : )]),
9156 4 : ),
9157 4 : (
9158 4 : Lsn(0x60),
9159 4 : KeyLogAtLsn(vec![(
9160 4 : Lsn(0x60),
9161 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9162 4 : )]),
9163 4 : ),
9164 4 : ],
9165 4 : above_horizon: KeyLogAtLsn(vec![
9166 4 : (
9167 4 : Lsn(0x70),
9168 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9169 4 : ),
9170 4 : (
9171 4 : Lsn(0x80),
9172 4 : Value::Image(Bytes::copy_from_slice(
9173 4 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9174 4 : )),
9175 4 : ),
9176 4 : (
9177 4 : Lsn(0x90),
9178 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9179 4 : ),
9180 4 : ]),
9181 4 : };
9182 4 : assert_eq!(res, expected_res);
9183 4 :
9184 4 : // In case of branch compaction, the branch itself does not have the full history, and we need to provide
9185 4 : // the ancestor image in the test case.
9186 4 :
9187 4 : let history = vec![
9188 4 : (
9189 4 : key,
9190 4 : Lsn(0x20),
9191 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9192 4 : ),
9193 4 : (
9194 4 : key,
9195 4 : Lsn(0x30),
9196 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9197 4 : ),
9198 4 : (
9199 4 : key,
9200 4 : Lsn(0x40),
9201 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9202 4 : ),
9203 4 : (
9204 4 : key,
9205 4 : Lsn(0x70),
9206 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9207 4 : ),
9208 4 : ];
9209 4 : let res = tline
9210 4 : .generate_key_retention(
9211 4 : key,
9212 4 : &history,
9213 4 : Lsn(0x60),
9214 4 : &[],
9215 4 : 3,
9216 4 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
9217 4 : )
9218 4 : .await
9219 4 : .unwrap();
9220 4 : let expected_res = KeyHistoryRetention {
9221 4 : below_horizon: vec![(
9222 4 : Lsn(0x60),
9223 4 : KeyLogAtLsn(vec![(
9224 4 : Lsn(0x60),
9225 4 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")), // use the ancestor image to reconstruct the page
9226 4 : )]),
9227 4 : )],
9228 4 : above_horizon: KeyLogAtLsn(vec![(
9229 4 : Lsn(0x70),
9230 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9231 4 : )]),
9232 4 : };
9233 4 : assert_eq!(res, expected_res);
9234 4 :
9235 4 : let history = vec![
9236 4 : (
9237 4 : key,
9238 4 : Lsn(0x20),
9239 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9240 4 : ),
9241 4 : (
9242 4 : key,
9243 4 : Lsn(0x40),
9244 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9245 4 : ),
9246 4 : (
9247 4 : key,
9248 4 : Lsn(0x60),
9249 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9250 4 : ),
9251 4 : (
9252 4 : key,
9253 4 : Lsn(0x70),
9254 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9255 4 : ),
9256 4 : ];
9257 4 : let res = tline
9258 4 : .generate_key_retention(
9259 4 : key,
9260 4 : &history,
9261 4 : Lsn(0x60),
9262 4 : &[Lsn(0x30)],
9263 4 : 3,
9264 4 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
9265 4 : )
9266 4 : .await
9267 4 : .unwrap();
9268 4 : let expected_res = KeyHistoryRetention {
9269 4 : below_horizon: vec![
9270 4 : (
9271 4 : Lsn(0x30),
9272 4 : KeyLogAtLsn(vec![(
9273 4 : Lsn(0x20),
9274 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9275 4 : )]),
9276 4 : ),
9277 4 : (
9278 4 : Lsn(0x60),
9279 4 : KeyLogAtLsn(vec![(
9280 4 : Lsn(0x60),
9281 4 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x40;0x60")),
9282 4 : )]),
9283 4 : ),
9284 4 : ],
9285 4 : above_horizon: KeyLogAtLsn(vec![(
9286 4 : Lsn(0x70),
9287 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9288 4 : )]),
9289 4 : };
9290 4 : assert_eq!(res, expected_res);
9291 4 :
9292 4 : Ok(())
9293 4 : }
9294 :
9295 : #[cfg(feature = "testing")]
9296 : #[tokio::test]
9297 4 : async fn test_simple_bottom_most_compaction_with_retain_lsns() -> anyhow::Result<()> {
9298 4 : let harness =
9299 4 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns").await?;
9300 4 : let (tenant, ctx) = harness.load().await;
9301 4 :
9302 1036 : fn get_key(id: u32) -> Key {
9303 1036 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9304 1036 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9305 1036 : key.field6 = id;
9306 1036 : key
9307 1036 : }
9308 4 :
9309 4 : let img_layer = (0..10)
9310 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9311 4 : .collect_vec();
9312 4 :
9313 4 : let delta1 = vec![
9314 4 : (
9315 4 : get_key(1),
9316 4 : Lsn(0x20),
9317 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9318 4 : ),
9319 4 : (
9320 4 : get_key(2),
9321 4 : Lsn(0x30),
9322 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9323 4 : ),
9324 4 : (
9325 4 : get_key(3),
9326 4 : Lsn(0x28),
9327 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9328 4 : ),
9329 4 : (
9330 4 : get_key(3),
9331 4 : Lsn(0x30),
9332 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9333 4 : ),
9334 4 : (
9335 4 : get_key(3),
9336 4 : Lsn(0x40),
9337 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9338 4 : ),
9339 4 : ];
9340 4 : let delta2 = vec![
9341 4 : (
9342 4 : get_key(5),
9343 4 : Lsn(0x20),
9344 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9345 4 : ),
9346 4 : (
9347 4 : get_key(6),
9348 4 : Lsn(0x20),
9349 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9350 4 : ),
9351 4 : ];
9352 4 : let delta3 = vec![
9353 4 : (
9354 4 : get_key(8),
9355 4 : Lsn(0x48),
9356 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9357 4 : ),
9358 4 : (
9359 4 : get_key(9),
9360 4 : Lsn(0x48),
9361 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9362 4 : ),
9363 4 : ];
9364 4 :
9365 4 : let tline = tenant
9366 4 : .create_test_timeline_with_layers(
9367 4 : TIMELINE_ID,
9368 4 : Lsn(0x10),
9369 4 : DEFAULT_PG_VERSION,
9370 4 : &ctx,
9371 4 : vec![
9372 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
9373 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
9374 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
9375 4 : ], // delta layers
9376 4 : vec![(Lsn(0x10), img_layer)], // image layers
9377 4 : Lsn(0x50),
9378 4 : )
9379 4 : .await?;
9380 4 : {
9381 4 : tline
9382 4 : .latest_gc_cutoff_lsn
9383 4 : .lock_for_write()
9384 4 : .store_and_unlock(Lsn(0x30))
9385 4 : .wait()
9386 4 : .await;
9387 4 : // Update GC info
9388 4 : let mut guard = tline.gc_info.write().unwrap();
9389 4 : *guard = GcInfo {
9390 4 : retain_lsns: vec![
9391 4 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
9392 4 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
9393 4 : ],
9394 4 : cutoffs: GcCutoffs {
9395 4 : time: Lsn(0x30),
9396 4 : space: Lsn(0x30),
9397 4 : },
9398 4 : leases: Default::default(),
9399 4 : within_ancestor_pitr: false,
9400 4 : };
9401 4 : }
9402 4 :
9403 4 : let expected_result = [
9404 4 : Bytes::from_static(b"value 0@0x10"),
9405 4 : Bytes::from_static(b"value 1@0x10@0x20"),
9406 4 : Bytes::from_static(b"value 2@0x10@0x30"),
9407 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9408 4 : Bytes::from_static(b"value 4@0x10"),
9409 4 : Bytes::from_static(b"value 5@0x10@0x20"),
9410 4 : Bytes::from_static(b"value 6@0x10@0x20"),
9411 4 : Bytes::from_static(b"value 7@0x10"),
9412 4 : Bytes::from_static(b"value 8@0x10@0x48"),
9413 4 : Bytes::from_static(b"value 9@0x10@0x48"),
9414 4 : ];
9415 4 :
9416 4 : let expected_result_at_gc_horizon = [
9417 4 : Bytes::from_static(b"value 0@0x10"),
9418 4 : Bytes::from_static(b"value 1@0x10@0x20"),
9419 4 : Bytes::from_static(b"value 2@0x10@0x30"),
9420 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
9421 4 : Bytes::from_static(b"value 4@0x10"),
9422 4 : Bytes::from_static(b"value 5@0x10@0x20"),
9423 4 : Bytes::from_static(b"value 6@0x10@0x20"),
9424 4 : Bytes::from_static(b"value 7@0x10"),
9425 4 : Bytes::from_static(b"value 8@0x10"),
9426 4 : Bytes::from_static(b"value 9@0x10"),
9427 4 : ];
9428 4 :
9429 4 : let expected_result_at_lsn_20 = [
9430 4 : Bytes::from_static(b"value 0@0x10"),
9431 4 : Bytes::from_static(b"value 1@0x10@0x20"),
9432 4 : Bytes::from_static(b"value 2@0x10"),
9433 4 : Bytes::from_static(b"value 3@0x10"),
9434 4 : Bytes::from_static(b"value 4@0x10"),
9435 4 : Bytes::from_static(b"value 5@0x10@0x20"),
9436 4 : Bytes::from_static(b"value 6@0x10@0x20"),
9437 4 : Bytes::from_static(b"value 7@0x10"),
9438 4 : Bytes::from_static(b"value 8@0x10"),
9439 4 : Bytes::from_static(b"value 9@0x10"),
9440 4 : ];
9441 4 :
9442 4 : let expected_result_at_lsn_10 = [
9443 4 : Bytes::from_static(b"value 0@0x10"),
9444 4 : Bytes::from_static(b"value 1@0x10"),
9445 4 : Bytes::from_static(b"value 2@0x10"),
9446 4 : Bytes::from_static(b"value 3@0x10"),
9447 4 : Bytes::from_static(b"value 4@0x10"),
9448 4 : Bytes::from_static(b"value 5@0x10"),
9449 4 : Bytes::from_static(b"value 6@0x10"),
9450 4 : Bytes::from_static(b"value 7@0x10"),
9451 4 : Bytes::from_static(b"value 8@0x10"),
9452 4 : Bytes::from_static(b"value 9@0x10"),
9453 4 : ];
9454 4 :
9455 24 : let verify_result = || async {
9456 24 : let gc_horizon = {
9457 24 : let gc_info = tline.gc_info.read().unwrap();
9458 24 : gc_info.cutoffs.time
9459 4 : };
9460 264 : for idx in 0..10 {
9461 240 : assert_eq!(
9462 240 : tline
9463 240 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9464 240 : .await
9465 240 : .unwrap(),
9466 240 : &expected_result[idx]
9467 4 : );
9468 240 : assert_eq!(
9469 240 : tline
9470 240 : .get(get_key(idx as u32), gc_horizon, &ctx)
9471 240 : .await
9472 240 : .unwrap(),
9473 240 : &expected_result_at_gc_horizon[idx]
9474 4 : );
9475 240 : assert_eq!(
9476 240 : tline
9477 240 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
9478 240 : .await
9479 240 : .unwrap(),
9480 240 : &expected_result_at_lsn_20[idx]
9481 4 : );
9482 240 : assert_eq!(
9483 240 : tline
9484 240 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
9485 240 : .await
9486 240 : .unwrap(),
9487 240 : &expected_result_at_lsn_10[idx]
9488 4 : );
9489 4 : }
9490 48 : };
9491 4 :
9492 4 : verify_result().await;
9493 4 :
9494 4 : let cancel = CancellationToken::new();
9495 4 : let mut dryrun_flags = EnumSet::new();
9496 4 : dryrun_flags.insert(CompactFlags::DryRun);
9497 4 :
9498 4 : tline
9499 4 : .compact_with_gc(
9500 4 : &cancel,
9501 4 : CompactOptions {
9502 4 : flags: dryrun_flags,
9503 4 : ..Default::default()
9504 4 : },
9505 4 : &ctx,
9506 4 : )
9507 4 : .await
9508 4 : .unwrap();
9509 4 : // We expect layer map to be the same b/c the dry run flag, but we don't know whether there will be other background jobs
9510 4 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
9511 4 : verify_result().await;
9512 4 :
9513 4 : tline
9514 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9515 4 : .await
9516 4 : .unwrap();
9517 4 : verify_result().await;
9518 4 :
9519 4 : // compact again
9520 4 : tline
9521 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9522 4 : .await
9523 4 : .unwrap();
9524 4 : verify_result().await;
9525 4 :
9526 4 : // increase GC horizon and compact again
9527 4 : {
9528 4 : tline
9529 4 : .latest_gc_cutoff_lsn
9530 4 : .lock_for_write()
9531 4 : .store_and_unlock(Lsn(0x38))
9532 4 : .wait()
9533 4 : .await;
9534 4 : // Update GC info
9535 4 : let mut guard = tline.gc_info.write().unwrap();
9536 4 : guard.cutoffs.time = Lsn(0x38);
9537 4 : guard.cutoffs.space = Lsn(0x38);
9538 4 : }
9539 4 : tline
9540 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9541 4 : .await
9542 4 : .unwrap();
9543 4 : verify_result().await; // no wals between 0x30 and 0x38, so we should obtain the same result
9544 4 :
9545 4 : // not increasing the GC horizon and compact again
9546 4 : tline
9547 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9548 4 : .await
9549 4 : .unwrap();
9550 4 : verify_result().await;
9551 4 :
9552 4 : Ok(())
9553 4 : }
9554 :
9555 : #[cfg(feature = "testing")]
9556 : #[tokio::test]
9557 4 : async fn test_simple_bottom_most_compaction_with_retain_lsns_single_key() -> anyhow::Result<()>
9558 4 : {
9559 4 : let harness =
9560 4 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns_single_key")
9561 4 : .await?;
9562 4 : let (tenant, ctx) = harness.load().await;
9563 4 :
9564 704 : fn get_key(id: u32) -> Key {
9565 704 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9566 704 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9567 704 : key.field6 = id;
9568 704 : key
9569 704 : }
9570 4 :
9571 4 : let img_layer = (0..10)
9572 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9573 4 : .collect_vec();
9574 4 :
9575 4 : let delta1 = vec![
9576 4 : (
9577 4 : get_key(1),
9578 4 : Lsn(0x20),
9579 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9580 4 : ),
9581 4 : (
9582 4 : get_key(1),
9583 4 : Lsn(0x28),
9584 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9585 4 : ),
9586 4 : ];
9587 4 : let delta2 = vec![
9588 4 : (
9589 4 : get_key(1),
9590 4 : Lsn(0x30),
9591 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9592 4 : ),
9593 4 : (
9594 4 : get_key(1),
9595 4 : Lsn(0x38),
9596 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
9597 4 : ),
9598 4 : ];
9599 4 : let delta3 = vec![
9600 4 : (
9601 4 : get_key(8),
9602 4 : Lsn(0x48),
9603 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9604 4 : ),
9605 4 : (
9606 4 : get_key(9),
9607 4 : Lsn(0x48),
9608 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9609 4 : ),
9610 4 : ];
9611 4 :
9612 4 : let tline = tenant
9613 4 : .create_test_timeline_with_layers(
9614 4 : TIMELINE_ID,
9615 4 : Lsn(0x10),
9616 4 : DEFAULT_PG_VERSION,
9617 4 : &ctx,
9618 4 : vec![
9619 4 : // delta1 and delta 2 only contain a single key but multiple updates
9620 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x30), delta1),
9621 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
9622 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x50), delta3),
9623 4 : ], // delta layers
9624 4 : vec![(Lsn(0x10), img_layer)], // image layers
9625 4 : Lsn(0x50),
9626 4 : )
9627 4 : .await?;
9628 4 : {
9629 4 : tline
9630 4 : .latest_gc_cutoff_lsn
9631 4 : .lock_for_write()
9632 4 : .store_and_unlock(Lsn(0x30))
9633 4 : .wait()
9634 4 : .await;
9635 4 : // Update GC info
9636 4 : let mut guard = tline.gc_info.write().unwrap();
9637 4 : *guard = GcInfo {
9638 4 : retain_lsns: vec![
9639 4 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
9640 4 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
9641 4 : ],
9642 4 : cutoffs: GcCutoffs {
9643 4 : time: Lsn(0x30),
9644 4 : space: Lsn(0x30),
9645 4 : },
9646 4 : leases: Default::default(),
9647 4 : within_ancestor_pitr: false,
9648 4 : };
9649 4 : }
9650 4 :
9651 4 : let expected_result = [
9652 4 : Bytes::from_static(b"value 0@0x10"),
9653 4 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
9654 4 : Bytes::from_static(b"value 2@0x10"),
9655 4 : Bytes::from_static(b"value 3@0x10"),
9656 4 : Bytes::from_static(b"value 4@0x10"),
9657 4 : Bytes::from_static(b"value 5@0x10"),
9658 4 : Bytes::from_static(b"value 6@0x10"),
9659 4 : Bytes::from_static(b"value 7@0x10"),
9660 4 : Bytes::from_static(b"value 8@0x10@0x48"),
9661 4 : Bytes::from_static(b"value 9@0x10@0x48"),
9662 4 : ];
9663 4 :
9664 4 : let expected_result_at_gc_horizon = [
9665 4 : Bytes::from_static(b"value 0@0x10"),
9666 4 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
9667 4 : Bytes::from_static(b"value 2@0x10"),
9668 4 : Bytes::from_static(b"value 3@0x10"),
9669 4 : Bytes::from_static(b"value 4@0x10"),
9670 4 : Bytes::from_static(b"value 5@0x10"),
9671 4 : Bytes::from_static(b"value 6@0x10"),
9672 4 : Bytes::from_static(b"value 7@0x10"),
9673 4 : Bytes::from_static(b"value 8@0x10"),
9674 4 : Bytes::from_static(b"value 9@0x10"),
9675 4 : ];
9676 4 :
9677 4 : let expected_result_at_lsn_20 = [
9678 4 : Bytes::from_static(b"value 0@0x10"),
9679 4 : Bytes::from_static(b"value 1@0x10@0x20"),
9680 4 : Bytes::from_static(b"value 2@0x10"),
9681 4 : Bytes::from_static(b"value 3@0x10"),
9682 4 : Bytes::from_static(b"value 4@0x10"),
9683 4 : Bytes::from_static(b"value 5@0x10"),
9684 4 : Bytes::from_static(b"value 6@0x10"),
9685 4 : Bytes::from_static(b"value 7@0x10"),
9686 4 : Bytes::from_static(b"value 8@0x10"),
9687 4 : Bytes::from_static(b"value 9@0x10"),
9688 4 : ];
9689 4 :
9690 4 : let expected_result_at_lsn_10 = [
9691 4 : Bytes::from_static(b"value 0@0x10"),
9692 4 : Bytes::from_static(b"value 1@0x10"),
9693 4 : Bytes::from_static(b"value 2@0x10"),
9694 4 : Bytes::from_static(b"value 3@0x10"),
9695 4 : Bytes::from_static(b"value 4@0x10"),
9696 4 : Bytes::from_static(b"value 5@0x10"),
9697 4 : Bytes::from_static(b"value 6@0x10"),
9698 4 : Bytes::from_static(b"value 7@0x10"),
9699 4 : Bytes::from_static(b"value 8@0x10"),
9700 4 : Bytes::from_static(b"value 9@0x10"),
9701 4 : ];
9702 4 :
9703 16 : let verify_result = || async {
9704 16 : let gc_horizon = {
9705 16 : let gc_info = tline.gc_info.read().unwrap();
9706 16 : gc_info.cutoffs.time
9707 4 : };
9708 176 : for idx in 0..10 {
9709 160 : assert_eq!(
9710 160 : tline
9711 160 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9712 160 : .await
9713 160 : .unwrap(),
9714 160 : &expected_result[idx]
9715 4 : );
9716 160 : assert_eq!(
9717 160 : tline
9718 160 : .get(get_key(idx as u32), gc_horizon, &ctx)
9719 160 : .await
9720 160 : .unwrap(),
9721 160 : &expected_result_at_gc_horizon[idx]
9722 4 : );
9723 160 : assert_eq!(
9724 160 : tline
9725 160 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
9726 160 : .await
9727 160 : .unwrap(),
9728 160 : &expected_result_at_lsn_20[idx]
9729 4 : );
9730 160 : assert_eq!(
9731 160 : tline
9732 160 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
9733 160 : .await
9734 160 : .unwrap(),
9735 160 : &expected_result_at_lsn_10[idx]
9736 4 : );
9737 4 : }
9738 32 : };
9739 4 :
9740 4 : verify_result().await;
9741 4 :
9742 4 : let cancel = CancellationToken::new();
9743 4 : let mut dryrun_flags = EnumSet::new();
9744 4 : dryrun_flags.insert(CompactFlags::DryRun);
9745 4 :
9746 4 : tline
9747 4 : .compact_with_gc(
9748 4 : &cancel,
9749 4 : CompactOptions {
9750 4 : flags: dryrun_flags,
9751 4 : ..Default::default()
9752 4 : },
9753 4 : &ctx,
9754 4 : )
9755 4 : .await
9756 4 : .unwrap();
9757 4 : // We expect layer map to be the same b/c the dry run flag, but we don't know whether there will be other background jobs
9758 4 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
9759 4 : verify_result().await;
9760 4 :
9761 4 : tline
9762 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9763 4 : .await
9764 4 : .unwrap();
9765 4 : verify_result().await;
9766 4 :
9767 4 : // compact again
9768 4 : tline
9769 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9770 4 : .await
9771 4 : .unwrap();
9772 4 : verify_result().await;
9773 4 :
9774 4 : Ok(())
9775 4 : }
9776 :
9777 : #[cfg(feature = "testing")]
9778 : #[tokio::test]
9779 4 : async fn test_simple_bottom_most_compaction_on_branch() -> anyhow::Result<()> {
9780 4 : use models::CompactLsnRange;
9781 4 :
9782 4 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_on_branch").await?;
9783 4 : let (tenant, ctx) = harness.load().await;
9784 4 :
9785 332 : fn get_key(id: u32) -> Key {
9786 332 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
9787 332 : key.field6 = id;
9788 332 : key
9789 332 : }
9790 4 :
9791 4 : let img_layer = (0..10)
9792 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9793 4 : .collect_vec();
9794 4 :
9795 4 : let delta1 = vec![
9796 4 : (
9797 4 : get_key(1),
9798 4 : Lsn(0x20),
9799 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9800 4 : ),
9801 4 : (
9802 4 : get_key(2),
9803 4 : Lsn(0x30),
9804 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9805 4 : ),
9806 4 : (
9807 4 : get_key(3),
9808 4 : Lsn(0x28),
9809 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9810 4 : ),
9811 4 : (
9812 4 : get_key(3),
9813 4 : Lsn(0x30),
9814 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9815 4 : ),
9816 4 : (
9817 4 : get_key(3),
9818 4 : Lsn(0x40),
9819 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9820 4 : ),
9821 4 : ];
9822 4 : let delta2 = vec![
9823 4 : (
9824 4 : get_key(5),
9825 4 : Lsn(0x20),
9826 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9827 4 : ),
9828 4 : (
9829 4 : get_key(6),
9830 4 : Lsn(0x20),
9831 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9832 4 : ),
9833 4 : ];
9834 4 : let delta3 = vec![
9835 4 : (
9836 4 : get_key(8),
9837 4 : Lsn(0x48),
9838 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9839 4 : ),
9840 4 : (
9841 4 : get_key(9),
9842 4 : Lsn(0x48),
9843 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9844 4 : ),
9845 4 : ];
9846 4 :
9847 4 : let parent_tline = tenant
9848 4 : .create_test_timeline_with_layers(
9849 4 : TIMELINE_ID,
9850 4 : Lsn(0x10),
9851 4 : DEFAULT_PG_VERSION,
9852 4 : &ctx,
9853 4 : vec![], // delta layers
9854 4 : vec![(Lsn(0x18), img_layer)], // image layers
9855 4 : Lsn(0x18),
9856 4 : )
9857 4 : .await?;
9858 4 :
9859 4 : parent_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
9860 4 :
9861 4 : let branch_tline = tenant
9862 4 : .branch_timeline_test_with_layers(
9863 4 : &parent_tline,
9864 4 : NEW_TIMELINE_ID,
9865 4 : Some(Lsn(0x18)),
9866 4 : &ctx,
9867 4 : vec![
9868 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
9869 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
9870 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
9871 4 : ], // delta layers
9872 4 : vec![], // image layers
9873 4 : Lsn(0x50),
9874 4 : )
9875 4 : .await?;
9876 4 :
9877 4 : branch_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
9878 4 :
9879 4 : {
9880 4 : parent_tline
9881 4 : .latest_gc_cutoff_lsn
9882 4 : .lock_for_write()
9883 4 : .store_and_unlock(Lsn(0x10))
9884 4 : .wait()
9885 4 : .await;
9886 4 : // Update GC info
9887 4 : let mut guard = parent_tline.gc_info.write().unwrap();
9888 4 : *guard = GcInfo {
9889 4 : retain_lsns: vec![(Lsn(0x18), branch_tline.timeline_id, MaybeOffloaded::No)],
9890 4 : cutoffs: GcCutoffs {
9891 4 : time: Lsn(0x10),
9892 4 : space: Lsn(0x10),
9893 4 : },
9894 4 : leases: Default::default(),
9895 4 : within_ancestor_pitr: false,
9896 4 : };
9897 4 : }
9898 4 :
9899 4 : {
9900 4 : branch_tline
9901 4 : .latest_gc_cutoff_lsn
9902 4 : .lock_for_write()
9903 4 : .store_and_unlock(Lsn(0x50))
9904 4 : .wait()
9905 4 : .await;
9906 4 : // Update GC info
9907 4 : let mut guard = branch_tline.gc_info.write().unwrap();
9908 4 : *guard = GcInfo {
9909 4 : retain_lsns: vec![(Lsn(0x40), branch_tline.timeline_id, MaybeOffloaded::No)],
9910 4 : cutoffs: GcCutoffs {
9911 4 : time: Lsn(0x50),
9912 4 : space: Lsn(0x50),
9913 4 : },
9914 4 : leases: Default::default(),
9915 4 : within_ancestor_pitr: false,
9916 4 : };
9917 4 : }
9918 4 :
9919 4 : let expected_result_at_gc_horizon = [
9920 4 : Bytes::from_static(b"value 0@0x10"),
9921 4 : Bytes::from_static(b"value 1@0x10@0x20"),
9922 4 : Bytes::from_static(b"value 2@0x10@0x30"),
9923 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9924 4 : Bytes::from_static(b"value 4@0x10"),
9925 4 : Bytes::from_static(b"value 5@0x10@0x20"),
9926 4 : Bytes::from_static(b"value 6@0x10@0x20"),
9927 4 : Bytes::from_static(b"value 7@0x10"),
9928 4 : Bytes::from_static(b"value 8@0x10@0x48"),
9929 4 : Bytes::from_static(b"value 9@0x10@0x48"),
9930 4 : ];
9931 4 :
9932 4 : let expected_result_at_lsn_40 = [
9933 4 : Bytes::from_static(b"value 0@0x10"),
9934 4 : Bytes::from_static(b"value 1@0x10@0x20"),
9935 4 : Bytes::from_static(b"value 2@0x10@0x30"),
9936 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9937 4 : Bytes::from_static(b"value 4@0x10"),
9938 4 : Bytes::from_static(b"value 5@0x10@0x20"),
9939 4 : Bytes::from_static(b"value 6@0x10@0x20"),
9940 4 : Bytes::from_static(b"value 7@0x10"),
9941 4 : Bytes::from_static(b"value 8@0x10"),
9942 4 : Bytes::from_static(b"value 9@0x10"),
9943 4 : ];
9944 4 :
9945 12 : let verify_result = || async {
9946 132 : for idx in 0..10 {
9947 120 : assert_eq!(
9948 120 : branch_tline
9949 120 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9950 120 : .await
9951 120 : .unwrap(),
9952 120 : &expected_result_at_gc_horizon[idx]
9953 4 : );
9954 120 : assert_eq!(
9955 120 : branch_tline
9956 120 : .get(get_key(idx as u32), Lsn(0x40), &ctx)
9957 120 : .await
9958 120 : .unwrap(),
9959 120 : &expected_result_at_lsn_40[idx]
9960 4 : );
9961 4 : }
9962 24 : };
9963 4 :
9964 4 : verify_result().await;
9965 4 :
9966 4 : let cancel = CancellationToken::new();
9967 4 : branch_tline
9968 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9969 4 : .await
9970 4 : .unwrap();
9971 4 :
9972 4 : verify_result().await;
9973 4 :
9974 4 : // Piggyback a compaction with above_lsn. Ensure it works correctly when the specified LSN intersects with the layer files.
9975 4 : // Now we already have a single large delta layer, so the compaction min_layer_lsn should be the same as ancestor LSN (0x18).
9976 4 : branch_tline
9977 4 : .compact_with_gc(
9978 4 : &cancel,
9979 4 : CompactOptions {
9980 4 : compact_lsn_range: Some(CompactLsnRange::above(Lsn(0x40))),
9981 4 : ..Default::default()
9982 4 : },
9983 4 : &ctx,
9984 4 : )
9985 4 : .await
9986 4 : .unwrap();
9987 4 :
9988 4 : verify_result().await;
9989 4 :
9990 4 : Ok(())
9991 4 : }
9992 :
9993 : // Regression test for https://github.com/neondatabase/neon/issues/9012
9994 : // Create an image arrangement where we have to read at different LSN ranges
9995 : // from a delta layer. This is achieved by overlapping an image layer on top of
9996 : // a delta layer. Like so:
9997 : //
9998 : // A B
9999 : // +----------------+ -> delta_layer
10000 : // | | ^ lsn
10001 : // | =========|-> nested_image_layer |
10002 : // | C | |
10003 : // +----------------+ |
10004 : // ======== -> baseline_image_layer +-------> key
10005 : //
10006 : //
10007 : // When querying the key range [A, B) we need to read at different LSN ranges
10008 : // for [A, C) and [C, B). This test checks that the described edge case is handled correctly.
10009 : #[cfg(feature = "testing")]
10010 : #[tokio::test]
10011 4 : async fn test_vectored_read_with_nested_image_layer() -> anyhow::Result<()> {
10012 4 : let harness = TenantHarness::create("test_vectored_read_with_nested_image_layer").await?;
10013 4 : let (tenant, ctx) = harness.load().await;
10014 4 :
10015 4 : let will_init_keys = [2, 6];
10016 88 : fn get_key(id: u32) -> Key {
10017 88 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
10018 88 : key.field6 = id;
10019 88 : key
10020 88 : }
10021 4 :
10022 4 : let mut expected_key_values = HashMap::new();
10023 4 :
10024 4 : let baseline_image_layer_lsn = Lsn(0x10);
10025 4 : let mut baseline_img_layer = Vec::new();
10026 24 : for i in 0..5 {
10027 20 : let key = get_key(i);
10028 20 : let value = format!("value {i}@{baseline_image_layer_lsn}");
10029 20 :
10030 20 : let removed = expected_key_values.insert(key, value.clone());
10031 20 : assert!(removed.is_none());
10032 4 :
10033 20 : baseline_img_layer.push((key, Bytes::from(value)));
10034 4 : }
10035 4 :
10036 4 : let nested_image_layer_lsn = Lsn(0x50);
10037 4 : let mut nested_img_layer = Vec::new();
10038 24 : for i in 5..10 {
10039 20 : let key = get_key(i);
10040 20 : let value = format!("value {i}@{nested_image_layer_lsn}");
10041 20 :
10042 20 : let removed = expected_key_values.insert(key, value.clone());
10043 20 : assert!(removed.is_none());
10044 4 :
10045 20 : nested_img_layer.push((key, Bytes::from(value)));
10046 4 : }
10047 4 :
10048 4 : let mut delta_layer_spec = Vec::default();
10049 4 : let delta_layer_start_lsn = Lsn(0x20);
10050 4 : let mut delta_layer_end_lsn = delta_layer_start_lsn;
10051 4 :
10052 44 : for i in 0..10 {
10053 40 : let key = get_key(i);
10054 40 : let key_in_nested = nested_img_layer
10055 40 : .iter()
10056 160 : .any(|(key_with_img, _)| *key_with_img == key);
10057 40 : let lsn = {
10058 40 : if key_in_nested {
10059 20 : Lsn(nested_image_layer_lsn.0 + 0x10)
10060 4 : } else {
10061 20 : delta_layer_start_lsn
10062 4 : }
10063 4 : };
10064 4 :
10065 40 : let will_init = will_init_keys.contains(&i);
10066 40 : if will_init {
10067 8 : delta_layer_spec.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init(""))));
10068 8 :
10069 8 : expected_key_values.insert(key, "".to_string());
10070 32 : } else {
10071 32 : let delta = format!("@{lsn}");
10072 32 : delta_layer_spec.push((
10073 32 : key,
10074 32 : lsn,
10075 32 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
10076 32 : ));
10077 32 :
10078 32 : expected_key_values
10079 32 : .get_mut(&key)
10080 32 : .expect("An image exists for each key")
10081 32 : .push_str(delta.as_str());
10082 32 : }
10083 40 : delta_layer_end_lsn = std::cmp::max(delta_layer_start_lsn, lsn);
10084 4 : }
10085 4 :
10086 4 : delta_layer_end_lsn = Lsn(delta_layer_end_lsn.0 + 1);
10087 4 :
10088 4 : assert!(
10089 4 : nested_image_layer_lsn > delta_layer_start_lsn
10090 4 : && nested_image_layer_lsn < delta_layer_end_lsn
10091 4 : );
10092 4 :
10093 4 : let tline = tenant
10094 4 : .create_test_timeline_with_layers(
10095 4 : TIMELINE_ID,
10096 4 : baseline_image_layer_lsn,
10097 4 : DEFAULT_PG_VERSION,
10098 4 : &ctx,
10099 4 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
10100 4 : delta_layer_start_lsn..delta_layer_end_lsn,
10101 4 : delta_layer_spec,
10102 4 : )], // delta layers
10103 4 : vec![
10104 4 : (baseline_image_layer_lsn, baseline_img_layer),
10105 4 : (nested_image_layer_lsn, nested_img_layer),
10106 4 : ], // image layers
10107 4 : delta_layer_end_lsn,
10108 4 : )
10109 4 : .await?;
10110 4 :
10111 4 : let keyspace = KeySpace::single(get_key(0)..get_key(10));
10112 4 : let results = tline
10113 4 : .get_vectored(
10114 4 : keyspace,
10115 4 : delta_layer_end_lsn,
10116 4 : IoConcurrency::sequential(),
10117 4 : &ctx,
10118 4 : )
10119 4 : .await
10120 4 : .expect("No vectored errors");
10121 44 : for (key, res) in results {
10122 40 : let value = res.expect("No key errors");
10123 40 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
10124 40 : assert_eq!(value, Bytes::from(expected_value));
10125 4 : }
10126 4 :
10127 4 : Ok(())
10128 4 : }
10129 :
10130 428 : fn sort_layer_key(k1: &PersistentLayerKey, k2: &PersistentLayerKey) -> std::cmp::Ordering {
10131 428 : (
10132 428 : k1.is_delta,
10133 428 : k1.key_range.start,
10134 428 : k1.key_range.end,
10135 428 : k1.lsn_range.start,
10136 428 : k1.lsn_range.end,
10137 428 : )
10138 428 : .cmp(&(
10139 428 : k2.is_delta,
10140 428 : k2.key_range.start,
10141 428 : k2.key_range.end,
10142 428 : k2.lsn_range.start,
10143 428 : k2.lsn_range.end,
10144 428 : ))
10145 428 : }
10146 :
10147 48 : async fn inspect_and_sort(
10148 48 : tline: &Arc<Timeline>,
10149 48 : filter: Option<std::ops::Range<Key>>,
10150 48 : ) -> Vec<PersistentLayerKey> {
10151 48 : let mut all_layers = tline.inspect_historic_layers().await.unwrap();
10152 48 : if let Some(filter) = filter {
10153 216 : all_layers.retain(|layer| overlaps_with(&layer.key_range, &filter));
10154 44 : }
10155 48 : all_layers.sort_by(sort_layer_key);
10156 48 : all_layers
10157 48 : }
10158 :
10159 : #[cfg(feature = "testing")]
10160 44 : fn check_layer_map_key_eq(
10161 44 : mut left: Vec<PersistentLayerKey>,
10162 44 : mut right: Vec<PersistentLayerKey>,
10163 44 : ) {
10164 44 : left.sort_by(sort_layer_key);
10165 44 : right.sort_by(sort_layer_key);
10166 44 : if left != right {
10167 0 : eprintln!("---LEFT---");
10168 0 : for left in left.iter() {
10169 0 : eprintln!("{}", left);
10170 0 : }
10171 0 : eprintln!("---RIGHT---");
10172 0 : for right in right.iter() {
10173 0 : eprintln!("{}", right);
10174 0 : }
10175 0 : assert_eq!(left, right);
10176 44 : }
10177 44 : }
10178 :
10179 : #[cfg(feature = "testing")]
10180 : #[tokio::test]
10181 4 : async fn test_simple_partial_bottom_most_compaction() -> anyhow::Result<()> {
10182 4 : let harness = TenantHarness::create("test_simple_partial_bottom_most_compaction").await?;
10183 4 : let (tenant, ctx) = harness.load().await;
10184 4 :
10185 364 : fn get_key(id: u32) -> Key {
10186 364 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
10187 364 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
10188 364 : key.field6 = id;
10189 364 : key
10190 364 : }
10191 4 :
10192 4 : // img layer at 0x10
10193 4 : let img_layer = (0..10)
10194 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10195 4 : .collect_vec();
10196 4 :
10197 4 : let delta1 = vec![
10198 4 : (
10199 4 : get_key(1),
10200 4 : Lsn(0x20),
10201 4 : Value::Image(Bytes::from("value 1@0x20")),
10202 4 : ),
10203 4 : (
10204 4 : get_key(2),
10205 4 : Lsn(0x30),
10206 4 : Value::Image(Bytes::from("value 2@0x30")),
10207 4 : ),
10208 4 : (
10209 4 : get_key(3),
10210 4 : Lsn(0x40),
10211 4 : Value::Image(Bytes::from("value 3@0x40")),
10212 4 : ),
10213 4 : ];
10214 4 : let delta2 = vec![
10215 4 : (
10216 4 : get_key(5),
10217 4 : Lsn(0x20),
10218 4 : Value::Image(Bytes::from("value 5@0x20")),
10219 4 : ),
10220 4 : (
10221 4 : get_key(6),
10222 4 : Lsn(0x20),
10223 4 : Value::Image(Bytes::from("value 6@0x20")),
10224 4 : ),
10225 4 : ];
10226 4 : let delta3 = vec![
10227 4 : (
10228 4 : get_key(8),
10229 4 : Lsn(0x48),
10230 4 : Value::Image(Bytes::from("value 8@0x48")),
10231 4 : ),
10232 4 : (
10233 4 : get_key(9),
10234 4 : Lsn(0x48),
10235 4 : Value::Image(Bytes::from("value 9@0x48")),
10236 4 : ),
10237 4 : ];
10238 4 :
10239 4 : let tline = tenant
10240 4 : .create_test_timeline_with_layers(
10241 4 : TIMELINE_ID,
10242 4 : Lsn(0x10),
10243 4 : DEFAULT_PG_VERSION,
10244 4 : &ctx,
10245 4 : vec![
10246 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
10247 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
10248 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
10249 4 : ], // delta layers
10250 4 : vec![(Lsn(0x10), img_layer)], // image layers
10251 4 : Lsn(0x50),
10252 4 : )
10253 4 : .await?;
10254 4 :
10255 4 : {
10256 4 : tline
10257 4 : .latest_gc_cutoff_lsn
10258 4 : .lock_for_write()
10259 4 : .store_and_unlock(Lsn(0x30))
10260 4 : .wait()
10261 4 : .await;
10262 4 : // Update GC info
10263 4 : let mut guard = tline.gc_info.write().unwrap();
10264 4 : *guard = GcInfo {
10265 4 : retain_lsns: vec![(Lsn(0x20), tline.timeline_id, MaybeOffloaded::No)],
10266 4 : cutoffs: GcCutoffs {
10267 4 : time: Lsn(0x30),
10268 4 : space: Lsn(0x30),
10269 4 : },
10270 4 : leases: Default::default(),
10271 4 : within_ancestor_pitr: false,
10272 4 : };
10273 4 : }
10274 4 :
10275 4 : let cancel = CancellationToken::new();
10276 4 :
10277 4 : // Do a partial compaction on key range 0..2
10278 4 : tline
10279 4 : .compact_with_gc(
10280 4 : &cancel,
10281 4 : CompactOptions {
10282 4 : flags: EnumSet::new(),
10283 4 : compact_key_range: Some((get_key(0)..get_key(2)).into()),
10284 4 : ..Default::default()
10285 4 : },
10286 4 : &ctx,
10287 4 : )
10288 4 : .await
10289 4 : .unwrap();
10290 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10291 4 : check_layer_map_key_eq(
10292 4 : all_layers,
10293 4 : vec![
10294 4 : // newly-generated image layer for the partial compaction range 0-2
10295 4 : PersistentLayerKey {
10296 4 : key_range: get_key(0)..get_key(2),
10297 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10298 4 : is_delta: false,
10299 4 : },
10300 4 : PersistentLayerKey {
10301 4 : key_range: get_key(0)..get_key(10),
10302 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
10303 4 : is_delta: false,
10304 4 : },
10305 4 : // delta1 is split and the second part is rewritten
10306 4 : PersistentLayerKey {
10307 4 : key_range: get_key(2)..get_key(4),
10308 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10309 4 : is_delta: true,
10310 4 : },
10311 4 : PersistentLayerKey {
10312 4 : key_range: get_key(5)..get_key(7),
10313 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10314 4 : is_delta: true,
10315 4 : },
10316 4 : PersistentLayerKey {
10317 4 : key_range: get_key(8)..get_key(10),
10318 4 : lsn_range: Lsn(0x48)..Lsn(0x50),
10319 4 : is_delta: true,
10320 4 : },
10321 4 : ],
10322 4 : );
10323 4 :
10324 4 : // Do a partial compaction on key range 2..4
10325 4 : tline
10326 4 : .compact_with_gc(
10327 4 : &cancel,
10328 4 : CompactOptions {
10329 4 : flags: EnumSet::new(),
10330 4 : compact_key_range: Some((get_key(2)..get_key(4)).into()),
10331 4 : ..Default::default()
10332 4 : },
10333 4 : &ctx,
10334 4 : )
10335 4 : .await
10336 4 : .unwrap();
10337 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10338 4 : check_layer_map_key_eq(
10339 4 : all_layers,
10340 4 : vec![
10341 4 : PersistentLayerKey {
10342 4 : key_range: get_key(0)..get_key(2),
10343 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10344 4 : is_delta: false,
10345 4 : },
10346 4 : PersistentLayerKey {
10347 4 : key_range: get_key(0)..get_key(10),
10348 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
10349 4 : is_delta: false,
10350 4 : },
10351 4 : // image layer generated for the compaction range 2-4
10352 4 : PersistentLayerKey {
10353 4 : key_range: get_key(2)..get_key(4),
10354 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10355 4 : is_delta: false,
10356 4 : },
10357 4 : // we have key2/key3 above the retain_lsn, so we still need this delta layer
10358 4 : PersistentLayerKey {
10359 4 : key_range: get_key(2)..get_key(4),
10360 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10361 4 : is_delta: true,
10362 4 : },
10363 4 : PersistentLayerKey {
10364 4 : key_range: get_key(5)..get_key(7),
10365 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10366 4 : is_delta: true,
10367 4 : },
10368 4 : PersistentLayerKey {
10369 4 : key_range: get_key(8)..get_key(10),
10370 4 : lsn_range: Lsn(0x48)..Lsn(0x50),
10371 4 : is_delta: true,
10372 4 : },
10373 4 : ],
10374 4 : );
10375 4 :
10376 4 : // Do a partial compaction on key range 4..9
10377 4 : tline
10378 4 : .compact_with_gc(
10379 4 : &cancel,
10380 4 : CompactOptions {
10381 4 : flags: EnumSet::new(),
10382 4 : compact_key_range: Some((get_key(4)..get_key(9)).into()),
10383 4 : ..Default::default()
10384 4 : },
10385 4 : &ctx,
10386 4 : )
10387 4 : .await
10388 4 : .unwrap();
10389 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10390 4 : check_layer_map_key_eq(
10391 4 : all_layers,
10392 4 : vec![
10393 4 : PersistentLayerKey {
10394 4 : key_range: get_key(0)..get_key(2),
10395 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10396 4 : is_delta: false,
10397 4 : },
10398 4 : PersistentLayerKey {
10399 4 : key_range: get_key(0)..get_key(10),
10400 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
10401 4 : is_delta: false,
10402 4 : },
10403 4 : PersistentLayerKey {
10404 4 : key_range: get_key(2)..get_key(4),
10405 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10406 4 : is_delta: false,
10407 4 : },
10408 4 : PersistentLayerKey {
10409 4 : key_range: get_key(2)..get_key(4),
10410 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10411 4 : is_delta: true,
10412 4 : },
10413 4 : // image layer generated for this compaction range
10414 4 : PersistentLayerKey {
10415 4 : key_range: get_key(4)..get_key(9),
10416 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10417 4 : is_delta: false,
10418 4 : },
10419 4 : PersistentLayerKey {
10420 4 : key_range: get_key(8)..get_key(10),
10421 4 : lsn_range: Lsn(0x48)..Lsn(0x50),
10422 4 : is_delta: true,
10423 4 : },
10424 4 : ],
10425 4 : );
10426 4 :
10427 4 : // Do a partial compaction on key range 9..10
10428 4 : tline
10429 4 : .compact_with_gc(
10430 4 : &cancel,
10431 4 : CompactOptions {
10432 4 : flags: EnumSet::new(),
10433 4 : compact_key_range: Some((get_key(9)..get_key(10)).into()),
10434 4 : ..Default::default()
10435 4 : },
10436 4 : &ctx,
10437 4 : )
10438 4 : .await
10439 4 : .unwrap();
10440 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10441 4 : check_layer_map_key_eq(
10442 4 : all_layers,
10443 4 : vec![
10444 4 : PersistentLayerKey {
10445 4 : key_range: get_key(0)..get_key(2),
10446 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10447 4 : is_delta: false,
10448 4 : },
10449 4 : PersistentLayerKey {
10450 4 : key_range: get_key(0)..get_key(10),
10451 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
10452 4 : is_delta: false,
10453 4 : },
10454 4 : PersistentLayerKey {
10455 4 : key_range: get_key(2)..get_key(4),
10456 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10457 4 : is_delta: false,
10458 4 : },
10459 4 : PersistentLayerKey {
10460 4 : key_range: get_key(2)..get_key(4),
10461 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10462 4 : is_delta: true,
10463 4 : },
10464 4 : PersistentLayerKey {
10465 4 : key_range: get_key(4)..get_key(9),
10466 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10467 4 : is_delta: false,
10468 4 : },
10469 4 : // image layer generated for the compaction range
10470 4 : PersistentLayerKey {
10471 4 : key_range: get_key(9)..get_key(10),
10472 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10473 4 : is_delta: false,
10474 4 : },
10475 4 : PersistentLayerKey {
10476 4 : key_range: get_key(8)..get_key(10),
10477 4 : lsn_range: Lsn(0x48)..Lsn(0x50),
10478 4 : is_delta: true,
10479 4 : },
10480 4 : ],
10481 4 : );
10482 4 :
10483 4 : // Do a partial compaction on key range 0..10, all image layers below LSN 20 can be replaced with new ones.
10484 4 : tline
10485 4 : .compact_with_gc(
10486 4 : &cancel,
10487 4 : CompactOptions {
10488 4 : flags: EnumSet::new(),
10489 4 : compact_key_range: Some((get_key(0)..get_key(10)).into()),
10490 4 : ..Default::default()
10491 4 : },
10492 4 : &ctx,
10493 4 : )
10494 4 : .await
10495 4 : .unwrap();
10496 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10497 4 : check_layer_map_key_eq(
10498 4 : all_layers,
10499 4 : vec![
10500 4 : // aha, we removed all unnecessary image/delta layers and got a very clean layer map!
10501 4 : PersistentLayerKey {
10502 4 : key_range: get_key(0)..get_key(10),
10503 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10504 4 : is_delta: false,
10505 4 : },
10506 4 : PersistentLayerKey {
10507 4 : key_range: get_key(2)..get_key(4),
10508 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10509 4 : is_delta: true,
10510 4 : },
10511 4 : PersistentLayerKey {
10512 4 : key_range: get_key(8)..get_key(10),
10513 4 : lsn_range: Lsn(0x48)..Lsn(0x50),
10514 4 : is_delta: true,
10515 4 : },
10516 4 : ],
10517 4 : );
10518 4 : Ok(())
10519 4 : }
10520 :
10521 : #[cfg(feature = "testing")]
10522 : #[tokio::test]
10523 4 : async fn test_timeline_offload_retain_lsn() -> anyhow::Result<()> {
10524 4 : let harness = TenantHarness::create("test_timeline_offload_retain_lsn")
10525 4 : .await
10526 4 : .unwrap();
10527 4 : let (tenant, ctx) = harness.load().await;
10528 4 : let tline_parent = tenant
10529 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
10530 4 : .await
10531 4 : .unwrap();
10532 4 : let tline_child = tenant
10533 4 : .branch_timeline_test(&tline_parent, NEW_TIMELINE_ID, Some(Lsn(0x20)), &ctx)
10534 4 : .await
10535 4 : .unwrap();
10536 4 : {
10537 4 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
10538 4 : assert_eq!(
10539 4 : gc_info_parent.retain_lsns,
10540 4 : vec![(Lsn(0x20), tline_child.timeline_id, MaybeOffloaded::No)]
10541 4 : );
10542 4 : }
10543 4 : // We have to directly call the remote_client instead of using the archive function to avoid constructing broker client...
10544 4 : tline_child
10545 4 : .remote_client
10546 4 : .schedule_index_upload_for_timeline_archival_state(TimelineArchivalState::Archived)
10547 4 : .unwrap();
10548 4 : tline_child.remote_client.wait_completion().await.unwrap();
10549 4 : offload_timeline(&tenant, &tline_child)
10550 4 : .instrument(tracing::info_span!(parent: None, "offload_test", tenant_id=%"test", shard_id=%"test", timeline_id=%"test"))
10551 4 : .await.unwrap();
10552 4 : let child_timeline_id = tline_child.timeline_id;
10553 4 : Arc::try_unwrap(tline_child).unwrap();
10554 4 :
10555 4 : {
10556 4 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
10557 4 : assert_eq!(
10558 4 : gc_info_parent.retain_lsns,
10559 4 : vec![(Lsn(0x20), child_timeline_id, MaybeOffloaded::Yes)]
10560 4 : );
10561 4 : }
10562 4 :
10563 4 : tenant
10564 4 : .get_offloaded_timeline(child_timeline_id)
10565 4 : .unwrap()
10566 4 : .defuse_for_tenant_drop();
10567 4 :
10568 4 : Ok(())
10569 4 : }
10570 :
10571 : #[cfg(feature = "testing")]
10572 : #[tokio::test]
10573 4 : async fn test_simple_bottom_most_compaction_above_lsn() -> anyhow::Result<()> {
10574 4 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_above_lsn").await?;
10575 4 : let (tenant, ctx) = harness.load().await;
10576 4 :
10577 592 : fn get_key(id: u32) -> Key {
10578 592 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
10579 592 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
10580 592 : key.field6 = id;
10581 592 : key
10582 592 : }
10583 4 :
10584 4 : let img_layer = (0..10)
10585 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10586 4 : .collect_vec();
10587 4 :
10588 4 : let delta1 = vec![(
10589 4 : get_key(1),
10590 4 : Lsn(0x20),
10591 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10592 4 : )];
10593 4 : let delta4 = vec![(
10594 4 : get_key(1),
10595 4 : Lsn(0x28),
10596 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
10597 4 : )];
10598 4 : let delta2 = vec![
10599 4 : (
10600 4 : get_key(1),
10601 4 : Lsn(0x30),
10602 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10603 4 : ),
10604 4 : (
10605 4 : get_key(1),
10606 4 : Lsn(0x38),
10607 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
10608 4 : ),
10609 4 : ];
10610 4 : let delta3 = vec![
10611 4 : (
10612 4 : get_key(8),
10613 4 : Lsn(0x48),
10614 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10615 4 : ),
10616 4 : (
10617 4 : get_key(9),
10618 4 : Lsn(0x48),
10619 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10620 4 : ),
10621 4 : ];
10622 4 :
10623 4 : let tline = tenant
10624 4 : .create_test_timeline_with_layers(
10625 4 : TIMELINE_ID,
10626 4 : Lsn(0x10),
10627 4 : DEFAULT_PG_VERSION,
10628 4 : &ctx,
10629 4 : vec![
10630 4 : // delta1/2/4 only contain a single key but multiple updates
10631 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x28), delta1),
10632 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
10633 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x28)..Lsn(0x30), delta4),
10634 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta3),
10635 4 : ], // delta layers
10636 4 : vec![(Lsn(0x10), img_layer)], // image layers
10637 4 : Lsn(0x50),
10638 4 : )
10639 4 : .await?;
10640 4 : {
10641 4 : tline
10642 4 : .latest_gc_cutoff_lsn
10643 4 : .lock_for_write()
10644 4 : .store_and_unlock(Lsn(0x30))
10645 4 : .wait()
10646 4 : .await;
10647 4 : // Update GC info
10648 4 : let mut guard = tline.gc_info.write().unwrap();
10649 4 : *guard = GcInfo {
10650 4 : retain_lsns: vec![
10651 4 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
10652 4 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
10653 4 : ],
10654 4 : cutoffs: GcCutoffs {
10655 4 : time: Lsn(0x30),
10656 4 : space: Lsn(0x30),
10657 4 : },
10658 4 : leases: Default::default(),
10659 4 : within_ancestor_pitr: false,
10660 4 : };
10661 4 : }
10662 4 :
10663 4 : let expected_result = [
10664 4 : Bytes::from_static(b"value 0@0x10"),
10665 4 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
10666 4 : Bytes::from_static(b"value 2@0x10"),
10667 4 : Bytes::from_static(b"value 3@0x10"),
10668 4 : Bytes::from_static(b"value 4@0x10"),
10669 4 : Bytes::from_static(b"value 5@0x10"),
10670 4 : Bytes::from_static(b"value 6@0x10"),
10671 4 : Bytes::from_static(b"value 7@0x10"),
10672 4 : Bytes::from_static(b"value 8@0x10@0x48"),
10673 4 : Bytes::from_static(b"value 9@0x10@0x48"),
10674 4 : ];
10675 4 :
10676 4 : let expected_result_at_gc_horizon = [
10677 4 : Bytes::from_static(b"value 0@0x10"),
10678 4 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
10679 4 : Bytes::from_static(b"value 2@0x10"),
10680 4 : Bytes::from_static(b"value 3@0x10"),
10681 4 : Bytes::from_static(b"value 4@0x10"),
10682 4 : Bytes::from_static(b"value 5@0x10"),
10683 4 : Bytes::from_static(b"value 6@0x10"),
10684 4 : Bytes::from_static(b"value 7@0x10"),
10685 4 : Bytes::from_static(b"value 8@0x10"),
10686 4 : Bytes::from_static(b"value 9@0x10"),
10687 4 : ];
10688 4 :
10689 4 : let expected_result_at_lsn_20 = [
10690 4 : Bytes::from_static(b"value 0@0x10"),
10691 4 : Bytes::from_static(b"value 1@0x10@0x20"),
10692 4 : Bytes::from_static(b"value 2@0x10"),
10693 4 : Bytes::from_static(b"value 3@0x10"),
10694 4 : Bytes::from_static(b"value 4@0x10"),
10695 4 : Bytes::from_static(b"value 5@0x10"),
10696 4 : Bytes::from_static(b"value 6@0x10"),
10697 4 : Bytes::from_static(b"value 7@0x10"),
10698 4 : Bytes::from_static(b"value 8@0x10"),
10699 4 : Bytes::from_static(b"value 9@0x10"),
10700 4 : ];
10701 4 :
10702 4 : let expected_result_at_lsn_10 = [
10703 4 : Bytes::from_static(b"value 0@0x10"),
10704 4 : Bytes::from_static(b"value 1@0x10"),
10705 4 : Bytes::from_static(b"value 2@0x10"),
10706 4 : Bytes::from_static(b"value 3@0x10"),
10707 4 : Bytes::from_static(b"value 4@0x10"),
10708 4 : Bytes::from_static(b"value 5@0x10"),
10709 4 : Bytes::from_static(b"value 6@0x10"),
10710 4 : Bytes::from_static(b"value 7@0x10"),
10711 4 : Bytes::from_static(b"value 8@0x10"),
10712 4 : Bytes::from_static(b"value 9@0x10"),
10713 4 : ];
10714 4 :
10715 12 : let verify_result = || async {
10716 12 : let gc_horizon = {
10717 12 : let gc_info = tline.gc_info.read().unwrap();
10718 12 : gc_info.cutoffs.time
10719 4 : };
10720 132 : for idx in 0..10 {
10721 120 : assert_eq!(
10722 120 : tline
10723 120 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10724 120 : .await
10725 120 : .unwrap(),
10726 120 : &expected_result[idx]
10727 4 : );
10728 120 : assert_eq!(
10729 120 : tline
10730 120 : .get(get_key(idx as u32), gc_horizon, &ctx)
10731 120 : .await
10732 120 : .unwrap(),
10733 120 : &expected_result_at_gc_horizon[idx]
10734 4 : );
10735 120 : assert_eq!(
10736 120 : tline
10737 120 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
10738 120 : .await
10739 120 : .unwrap(),
10740 120 : &expected_result_at_lsn_20[idx]
10741 4 : );
10742 120 : assert_eq!(
10743 120 : tline
10744 120 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
10745 120 : .await
10746 120 : .unwrap(),
10747 120 : &expected_result_at_lsn_10[idx]
10748 4 : );
10749 4 : }
10750 24 : };
10751 4 :
10752 4 : verify_result().await;
10753 4 :
10754 4 : let cancel = CancellationToken::new();
10755 4 : tline
10756 4 : .compact_with_gc(
10757 4 : &cancel,
10758 4 : CompactOptions {
10759 4 : compact_lsn_range: Some(CompactLsnRange::above(Lsn(0x28))),
10760 4 : ..Default::default()
10761 4 : },
10762 4 : &ctx,
10763 4 : )
10764 4 : .await
10765 4 : .unwrap();
10766 4 : verify_result().await;
10767 4 :
10768 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10769 4 : check_layer_map_key_eq(
10770 4 : all_layers,
10771 4 : vec![
10772 4 : // The original image layer, not compacted
10773 4 : PersistentLayerKey {
10774 4 : key_range: get_key(0)..get_key(10),
10775 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
10776 4 : is_delta: false,
10777 4 : },
10778 4 : // Delta layer below the specified above_lsn not compacted
10779 4 : PersistentLayerKey {
10780 4 : key_range: get_key(1)..get_key(2),
10781 4 : lsn_range: Lsn(0x20)..Lsn(0x28),
10782 4 : is_delta: true,
10783 4 : },
10784 4 : // Delta layer compacted above the LSN
10785 4 : PersistentLayerKey {
10786 4 : key_range: get_key(1)..get_key(10),
10787 4 : lsn_range: Lsn(0x28)..Lsn(0x50),
10788 4 : is_delta: true,
10789 4 : },
10790 4 : ],
10791 4 : );
10792 4 :
10793 4 : // compact again
10794 4 : tline
10795 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10796 4 : .await
10797 4 : .unwrap();
10798 4 : verify_result().await;
10799 4 :
10800 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10801 4 : check_layer_map_key_eq(
10802 4 : all_layers,
10803 4 : vec![
10804 4 : // The compacted image layer (full key range)
10805 4 : PersistentLayerKey {
10806 4 : key_range: Key::MIN..Key::MAX,
10807 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
10808 4 : is_delta: false,
10809 4 : },
10810 4 : // All other data in the delta layer
10811 4 : PersistentLayerKey {
10812 4 : key_range: get_key(1)..get_key(10),
10813 4 : lsn_range: Lsn(0x10)..Lsn(0x50),
10814 4 : is_delta: true,
10815 4 : },
10816 4 : ],
10817 4 : );
10818 4 :
10819 4 : Ok(())
10820 4 : }
10821 :
10822 : #[cfg(feature = "testing")]
10823 : #[tokio::test]
10824 4 : async fn test_simple_bottom_most_compaction_rectangle() -> anyhow::Result<()> {
10825 4 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_rectangle").await?;
10826 4 : let (tenant, ctx) = harness.load().await;
10827 4 :
10828 1016 : fn get_key(id: u32) -> Key {
10829 1016 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
10830 1016 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
10831 1016 : key.field6 = id;
10832 1016 : key
10833 1016 : }
10834 4 :
10835 4 : let img_layer = (0..10)
10836 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10837 4 : .collect_vec();
10838 4 :
10839 4 : let delta1 = vec![(
10840 4 : get_key(1),
10841 4 : Lsn(0x20),
10842 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10843 4 : )];
10844 4 : let delta4 = vec![(
10845 4 : get_key(1),
10846 4 : Lsn(0x28),
10847 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
10848 4 : )];
10849 4 : let delta2 = vec![
10850 4 : (
10851 4 : get_key(1),
10852 4 : Lsn(0x30),
10853 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10854 4 : ),
10855 4 : (
10856 4 : get_key(1),
10857 4 : Lsn(0x38),
10858 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
10859 4 : ),
10860 4 : ];
10861 4 : let delta3 = vec![
10862 4 : (
10863 4 : get_key(8),
10864 4 : Lsn(0x48),
10865 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10866 4 : ),
10867 4 : (
10868 4 : get_key(9),
10869 4 : Lsn(0x48),
10870 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10871 4 : ),
10872 4 : ];
10873 4 :
10874 4 : let tline = tenant
10875 4 : .create_test_timeline_with_layers(
10876 4 : TIMELINE_ID,
10877 4 : Lsn(0x10),
10878 4 : DEFAULT_PG_VERSION,
10879 4 : &ctx,
10880 4 : vec![
10881 4 : // delta1/2/4 only contain a single key but multiple updates
10882 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x28), delta1),
10883 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
10884 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x28)..Lsn(0x30), delta4),
10885 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta3),
10886 4 : ], // delta layers
10887 4 : vec![(Lsn(0x10), img_layer)], // image layers
10888 4 : Lsn(0x50),
10889 4 : )
10890 4 : .await?;
10891 4 : {
10892 4 : tline
10893 4 : .latest_gc_cutoff_lsn
10894 4 : .lock_for_write()
10895 4 : .store_and_unlock(Lsn(0x30))
10896 4 : .wait()
10897 4 : .await;
10898 4 : // Update GC info
10899 4 : let mut guard = tline.gc_info.write().unwrap();
10900 4 : *guard = GcInfo {
10901 4 : retain_lsns: vec![
10902 4 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
10903 4 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
10904 4 : ],
10905 4 : cutoffs: GcCutoffs {
10906 4 : time: Lsn(0x30),
10907 4 : space: Lsn(0x30),
10908 4 : },
10909 4 : leases: Default::default(),
10910 4 : within_ancestor_pitr: false,
10911 4 : };
10912 4 : }
10913 4 :
10914 4 : let expected_result = [
10915 4 : Bytes::from_static(b"value 0@0x10"),
10916 4 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
10917 4 : Bytes::from_static(b"value 2@0x10"),
10918 4 : Bytes::from_static(b"value 3@0x10"),
10919 4 : Bytes::from_static(b"value 4@0x10"),
10920 4 : Bytes::from_static(b"value 5@0x10"),
10921 4 : Bytes::from_static(b"value 6@0x10"),
10922 4 : Bytes::from_static(b"value 7@0x10"),
10923 4 : Bytes::from_static(b"value 8@0x10@0x48"),
10924 4 : Bytes::from_static(b"value 9@0x10@0x48"),
10925 4 : ];
10926 4 :
10927 4 : let expected_result_at_gc_horizon = [
10928 4 : Bytes::from_static(b"value 0@0x10"),
10929 4 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
10930 4 : Bytes::from_static(b"value 2@0x10"),
10931 4 : Bytes::from_static(b"value 3@0x10"),
10932 4 : Bytes::from_static(b"value 4@0x10"),
10933 4 : Bytes::from_static(b"value 5@0x10"),
10934 4 : Bytes::from_static(b"value 6@0x10"),
10935 4 : Bytes::from_static(b"value 7@0x10"),
10936 4 : Bytes::from_static(b"value 8@0x10"),
10937 4 : Bytes::from_static(b"value 9@0x10"),
10938 4 : ];
10939 4 :
10940 4 : let expected_result_at_lsn_20 = [
10941 4 : Bytes::from_static(b"value 0@0x10"),
10942 4 : Bytes::from_static(b"value 1@0x10@0x20"),
10943 4 : Bytes::from_static(b"value 2@0x10"),
10944 4 : Bytes::from_static(b"value 3@0x10"),
10945 4 : Bytes::from_static(b"value 4@0x10"),
10946 4 : Bytes::from_static(b"value 5@0x10"),
10947 4 : Bytes::from_static(b"value 6@0x10"),
10948 4 : Bytes::from_static(b"value 7@0x10"),
10949 4 : Bytes::from_static(b"value 8@0x10"),
10950 4 : Bytes::from_static(b"value 9@0x10"),
10951 4 : ];
10952 4 :
10953 4 : let expected_result_at_lsn_10 = [
10954 4 : Bytes::from_static(b"value 0@0x10"),
10955 4 : Bytes::from_static(b"value 1@0x10"),
10956 4 : Bytes::from_static(b"value 2@0x10"),
10957 4 : Bytes::from_static(b"value 3@0x10"),
10958 4 : Bytes::from_static(b"value 4@0x10"),
10959 4 : Bytes::from_static(b"value 5@0x10"),
10960 4 : Bytes::from_static(b"value 6@0x10"),
10961 4 : Bytes::from_static(b"value 7@0x10"),
10962 4 : Bytes::from_static(b"value 8@0x10"),
10963 4 : Bytes::from_static(b"value 9@0x10"),
10964 4 : ];
10965 4 :
10966 20 : let verify_result = || async {
10967 20 : let gc_horizon = {
10968 20 : let gc_info = tline.gc_info.read().unwrap();
10969 20 : gc_info.cutoffs.time
10970 4 : };
10971 220 : for idx in 0..10 {
10972 200 : assert_eq!(
10973 200 : tline
10974 200 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10975 200 : .await
10976 200 : .unwrap(),
10977 200 : &expected_result[idx]
10978 4 : );
10979 200 : assert_eq!(
10980 200 : tline
10981 200 : .get(get_key(idx as u32), gc_horizon, &ctx)
10982 200 : .await
10983 200 : .unwrap(),
10984 200 : &expected_result_at_gc_horizon[idx]
10985 4 : );
10986 200 : assert_eq!(
10987 200 : tline
10988 200 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
10989 200 : .await
10990 200 : .unwrap(),
10991 200 : &expected_result_at_lsn_20[idx]
10992 4 : );
10993 200 : assert_eq!(
10994 200 : tline
10995 200 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
10996 200 : .await
10997 200 : .unwrap(),
10998 200 : &expected_result_at_lsn_10[idx]
10999 4 : );
11000 4 : }
11001 40 : };
11002 4 :
11003 4 : verify_result().await;
11004 4 :
11005 4 : let cancel = CancellationToken::new();
11006 4 :
11007 4 : tline
11008 4 : .compact_with_gc(
11009 4 : &cancel,
11010 4 : CompactOptions {
11011 4 : compact_key_range: Some((get_key(0)..get_key(2)).into()),
11012 4 : compact_lsn_range: Some((Lsn(0x20)..Lsn(0x28)).into()),
11013 4 : ..Default::default()
11014 4 : },
11015 4 : &ctx,
11016 4 : )
11017 4 : .await
11018 4 : .unwrap();
11019 4 : verify_result().await;
11020 4 :
11021 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11022 4 : check_layer_map_key_eq(
11023 4 : all_layers,
11024 4 : vec![
11025 4 : // The original image layer, not compacted
11026 4 : PersistentLayerKey {
11027 4 : key_range: get_key(0)..get_key(10),
11028 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
11029 4 : is_delta: false,
11030 4 : },
11031 4 : // According the selection logic, we select all layers with start key <= 0x28, so we would merge the layer 0x20-0x28 and
11032 4 : // the layer 0x28-0x30 into one.
11033 4 : PersistentLayerKey {
11034 4 : key_range: get_key(1)..get_key(2),
11035 4 : lsn_range: Lsn(0x20)..Lsn(0x30),
11036 4 : is_delta: true,
11037 4 : },
11038 4 : // Above the upper bound and untouched
11039 4 : PersistentLayerKey {
11040 4 : key_range: get_key(1)..get_key(2),
11041 4 : lsn_range: Lsn(0x30)..Lsn(0x50),
11042 4 : is_delta: true,
11043 4 : },
11044 4 : // This layer is untouched
11045 4 : PersistentLayerKey {
11046 4 : key_range: get_key(8)..get_key(10),
11047 4 : lsn_range: Lsn(0x30)..Lsn(0x50),
11048 4 : is_delta: true,
11049 4 : },
11050 4 : ],
11051 4 : );
11052 4 :
11053 4 : tline
11054 4 : .compact_with_gc(
11055 4 : &cancel,
11056 4 : CompactOptions {
11057 4 : compact_key_range: Some((get_key(3)..get_key(8)).into()),
11058 4 : compact_lsn_range: Some((Lsn(0x28)..Lsn(0x40)).into()),
11059 4 : ..Default::default()
11060 4 : },
11061 4 : &ctx,
11062 4 : )
11063 4 : .await
11064 4 : .unwrap();
11065 4 : verify_result().await;
11066 4 :
11067 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11068 4 : check_layer_map_key_eq(
11069 4 : all_layers,
11070 4 : vec![
11071 4 : // The original image layer, not compacted
11072 4 : PersistentLayerKey {
11073 4 : key_range: get_key(0)..get_key(10),
11074 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
11075 4 : is_delta: false,
11076 4 : },
11077 4 : // Not in the compaction key range, uncompacted
11078 4 : PersistentLayerKey {
11079 4 : key_range: get_key(1)..get_key(2),
11080 4 : lsn_range: Lsn(0x20)..Lsn(0x30),
11081 4 : is_delta: true,
11082 4 : },
11083 4 : // Not in the compaction key range, uncompacted but need rewrite because the delta layer overlaps with the range
11084 4 : PersistentLayerKey {
11085 4 : key_range: get_key(1)..get_key(2),
11086 4 : lsn_range: Lsn(0x30)..Lsn(0x50),
11087 4 : is_delta: true,
11088 4 : },
11089 4 : // Note that when we specify the LSN upper bound to be 0x40, the compaction algorithm will not try to cut the layer
11090 4 : // horizontally in half. Instead, it will include all LSNs that overlap with 0x40. So the real max_lsn of the compaction
11091 4 : // becomes 0x50.
11092 4 : PersistentLayerKey {
11093 4 : key_range: get_key(8)..get_key(10),
11094 4 : lsn_range: Lsn(0x30)..Lsn(0x50),
11095 4 : is_delta: true,
11096 4 : },
11097 4 : ],
11098 4 : );
11099 4 :
11100 4 : // compact again
11101 4 : tline
11102 4 : .compact_with_gc(
11103 4 : &cancel,
11104 4 : CompactOptions {
11105 4 : compact_key_range: Some((get_key(0)..get_key(5)).into()),
11106 4 : compact_lsn_range: Some((Lsn(0x20)..Lsn(0x50)).into()),
11107 4 : ..Default::default()
11108 4 : },
11109 4 : &ctx,
11110 4 : )
11111 4 : .await
11112 4 : .unwrap();
11113 4 : verify_result().await;
11114 4 :
11115 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11116 4 : check_layer_map_key_eq(
11117 4 : all_layers,
11118 4 : vec![
11119 4 : // The original image layer, not compacted
11120 4 : PersistentLayerKey {
11121 4 : key_range: get_key(0)..get_key(10),
11122 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
11123 4 : is_delta: false,
11124 4 : },
11125 4 : // The range gets compacted
11126 4 : PersistentLayerKey {
11127 4 : key_range: get_key(1)..get_key(2),
11128 4 : lsn_range: Lsn(0x20)..Lsn(0x50),
11129 4 : is_delta: true,
11130 4 : },
11131 4 : // Not touched during this iteration of compaction
11132 4 : PersistentLayerKey {
11133 4 : key_range: get_key(8)..get_key(10),
11134 4 : lsn_range: Lsn(0x30)..Lsn(0x50),
11135 4 : is_delta: true,
11136 4 : },
11137 4 : ],
11138 4 : );
11139 4 :
11140 4 : // final full compaction
11141 4 : tline
11142 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
11143 4 : .await
11144 4 : .unwrap();
11145 4 : verify_result().await;
11146 4 :
11147 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11148 4 : check_layer_map_key_eq(
11149 4 : all_layers,
11150 4 : vec![
11151 4 : // The compacted image layer (full key range)
11152 4 : PersistentLayerKey {
11153 4 : key_range: Key::MIN..Key::MAX,
11154 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
11155 4 : is_delta: false,
11156 4 : },
11157 4 : // All other data in the delta layer
11158 4 : PersistentLayerKey {
11159 4 : key_range: get_key(1)..get_key(10),
11160 4 : lsn_range: Lsn(0x10)..Lsn(0x50),
11161 4 : is_delta: true,
11162 4 : },
11163 4 : ],
11164 4 : );
11165 4 :
11166 4 : Ok(())
11167 4 : }
11168 : }
|
