Line data Source code
1 : //! Timeline repository implementation that keeps old data in layer files, and
2 : //! the recent changes in ephemeral files.
3 : //!
4 : //! See tenant/*_layer.rs files. The functions here are responsible for locating
5 : //! the correct layer for the get/put call, walking back the timeline branching
6 : //! history as needed.
7 : //!
8 : //! The files are stored in the .neon/tenants/<tenant_id>/timelines/<timeline_id>
9 : //! directory. See docs/pageserver-storage.md for how the files are managed.
10 : //! In addition to the layer files, there is a metadata file in the same
11 : //! directory that contains information about the timeline, in particular its
12 : //! parent timeline, and the last LSN that has been written to disk.
13 : //!
14 :
15 : use anyhow::{bail, Context};
16 : use arc_swap::ArcSwap;
17 : use camino::Utf8Path;
18 : use camino::Utf8PathBuf;
19 : use chrono::NaiveDateTime;
20 : use enumset::EnumSet;
21 : use futures::stream::FuturesUnordered;
22 : use futures::StreamExt;
23 : use pageserver_api::models;
24 : use pageserver_api::models::LsnLease;
25 : use pageserver_api::models::TimelineArchivalState;
26 : use pageserver_api::models::TimelineState;
27 : use pageserver_api::models::TopTenantShardItem;
28 : use pageserver_api::models::WalRedoManagerStatus;
29 : use pageserver_api::shard::ShardIdentity;
30 : use pageserver_api::shard::ShardStripeSize;
31 : use pageserver_api::shard::TenantShardId;
32 : use remote_storage::DownloadError;
33 : use remote_storage::GenericRemoteStorage;
34 : use remote_storage::TimeoutOrCancel;
35 : use remote_timeline_client::manifest::{
36 : OffloadedTimelineManifest, TenantManifest, LATEST_TENANT_MANIFEST_VERSION,
37 : };
38 : use remote_timeline_client::UploadQueueNotReadyError;
39 : use std::collections::BTreeMap;
40 : use std::fmt;
41 : use std::future::Future;
42 : use std::sync::Weak;
43 : use std::time::SystemTime;
44 : use storage_broker::BrokerClientChannel;
45 : use timeline::offload::offload_timeline;
46 : use tokio::io::BufReader;
47 : use tokio::sync::watch;
48 : use tokio::task::JoinSet;
49 : use tokio_util::sync::CancellationToken;
50 : use tracing::*;
51 : use upload_queue::NotInitialized;
52 : use utils::backoff;
53 : use utils::circuit_breaker::CircuitBreaker;
54 : use utils::completion;
55 : use utils::crashsafe::path_with_suffix_extension;
56 : use utils::failpoint_support;
57 : use utils::fs_ext;
58 : use utils::pausable_failpoint;
59 : use utils::sync::gate::Gate;
60 : use utils::sync::gate::GateGuard;
61 : use utils::timeout::timeout_cancellable;
62 : use utils::timeout::TimeoutCancellableError;
63 : use utils::zstd::create_zst_tarball;
64 : use utils::zstd::extract_zst_tarball;
65 :
66 : use self::config::AttachedLocationConfig;
67 : use self::config::AttachmentMode;
68 : use self::config::LocationConf;
69 : use self::config::TenantConf;
70 : use self::metadata::TimelineMetadata;
71 : use self::mgr::GetActiveTenantError;
72 : use self::mgr::GetTenantError;
73 : use self::remote_timeline_client::upload::{upload_index_part, upload_tenant_manifest};
74 : use self::remote_timeline_client::{RemoteTimelineClient, WaitCompletionError};
75 : use self::timeline::uninit::TimelineCreateGuard;
76 : use self::timeline::uninit::TimelineExclusionError;
77 : use self::timeline::uninit::UninitializedTimeline;
78 : use self::timeline::EvictionTaskTenantState;
79 : use self::timeline::GcCutoffs;
80 : use self::timeline::TimelineDeleteProgress;
81 : use self::timeline::TimelineResources;
82 : use self::timeline::WaitLsnError;
83 : use crate::config::PageServerConf;
84 : use crate::context::{DownloadBehavior, RequestContext};
85 : use crate::deletion_queue::DeletionQueueClient;
86 : use crate::deletion_queue::DeletionQueueError;
87 : use crate::import_datadir;
88 : use crate::is_uninit_mark;
89 : use crate::l0_flush::L0FlushGlobalState;
90 : use crate::metrics::TENANT;
91 : use crate::metrics::{
92 : remove_tenant_metrics, BROKEN_TENANTS_SET, CIRCUIT_BREAKERS_BROKEN, CIRCUIT_BREAKERS_UNBROKEN,
93 : TENANT_STATE_METRIC, TENANT_SYNTHETIC_SIZE_METRIC,
94 : };
95 : use crate::repository::GcResult;
96 : use crate::task_mgr;
97 : use crate::task_mgr::TaskKind;
98 : use crate::tenant::config::LocationMode;
99 : use crate::tenant::config::TenantConfOpt;
100 : pub use crate::tenant::remote_timeline_client::index::IndexPart;
101 : use crate::tenant::remote_timeline_client::remote_initdb_archive_path;
102 : use crate::tenant::remote_timeline_client::MaybeDeletedIndexPart;
103 : use crate::tenant::remote_timeline_client::INITDB_PATH;
104 : use crate::tenant::storage_layer::DeltaLayer;
105 : use crate::tenant::storage_layer::ImageLayer;
106 : use crate::walingest::WalLagCooldown;
107 : use crate::walredo;
108 : use crate::InitializationOrder;
109 : use std::collections::hash_map::Entry;
110 : use std::collections::HashMap;
111 : use std::collections::HashSet;
112 : use std::fmt::Debug;
113 : use std::fmt::Display;
114 : use std::fs;
115 : use std::fs::File;
116 : use std::sync::atomic::{AtomicU64, Ordering};
117 : use std::sync::Arc;
118 : use std::sync::Mutex;
119 : use std::time::{Duration, Instant};
120 :
121 : use crate::span;
122 : use crate::tenant::timeline::delete::DeleteTimelineFlow;
123 : use crate::tenant::timeline::uninit::cleanup_timeline_directory;
124 : use crate::virtual_file::VirtualFile;
125 : use crate::walredo::PostgresRedoManager;
126 : use crate::TEMP_FILE_SUFFIX;
127 : use once_cell::sync::Lazy;
128 : pub use pageserver_api::models::TenantState;
129 : use tokio::sync::Semaphore;
130 :
131 0 : static INIT_DB_SEMAPHORE: Lazy<Semaphore> = Lazy::new(|| Semaphore::new(8));
132 : use utils::{
133 : crashsafe,
134 : generation::Generation,
135 : id::TimelineId,
136 : lsn::{Lsn, RecordLsn},
137 : };
138 :
139 : pub mod blob_io;
140 : pub mod block_io;
141 : pub mod vectored_blob_io;
142 :
143 : pub mod disk_btree;
144 : pub(crate) mod ephemeral_file;
145 : pub mod layer_map;
146 :
147 : pub mod metadata;
148 : pub mod remote_timeline_client;
149 : pub mod storage_layer;
150 :
151 : pub mod checks;
152 : pub mod config;
153 : pub mod mgr;
154 : pub mod secondary;
155 : pub mod tasks;
156 : pub mod upload_queue;
157 :
158 : pub(crate) mod timeline;
159 :
160 : pub mod size;
161 :
162 : mod gc_block;
163 : pub(crate) mod throttle;
164 :
165 : pub(crate) use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
166 : pub(crate) use timeline::{LogicalSizeCalculationCause, PageReconstructError, Timeline};
167 :
168 : // re-export for use in walreceiver
169 : pub use crate::tenant::timeline::WalReceiverInfo;
170 :
171 : /// The "tenants" part of `tenants/<tenant>/timelines...`
172 : pub const TENANTS_SEGMENT_NAME: &str = "tenants";
173 :
174 : /// Parts of the `.neon/tenants/<tenant_id>/timelines/<timeline_id>` directory prefix.
175 : pub const TIMELINES_SEGMENT_NAME: &str = "timelines";
176 :
177 : /// References to shared objects that are passed into each tenant, such
178 : /// as the shared remote storage client and process initialization state.
179 : #[derive(Clone)]
180 : pub struct TenantSharedResources {
181 : pub broker_client: storage_broker::BrokerClientChannel,
182 : pub remote_storage: GenericRemoteStorage,
183 : pub deletion_queue_client: DeletionQueueClient,
184 : pub l0_flush_global_state: L0FlushGlobalState,
185 : }
186 :
187 : /// A [`Tenant`] is really an _attached_ tenant. The configuration
188 : /// for an attached tenant is a subset of the [`LocationConf`], represented
189 : /// in this struct.
190 : pub(super) struct AttachedTenantConf {
191 : tenant_conf: TenantConfOpt,
192 : location: AttachedLocationConfig,
193 : /// The deadline before which we are blocked from GC so that
194 : /// leases have a chance to be renewed.
195 : lsn_lease_deadline: Option<tokio::time::Instant>,
196 : }
197 :
198 : impl AttachedTenantConf {
199 186 : fn new(tenant_conf: TenantConfOpt, location: AttachedLocationConfig) -> Self {
200 : // Sets a deadline before which we cannot proceed to GC due to lsn lease.
201 : //
202 : // We do this as the leases mapping are not persisted to disk. By delaying GC by lease
203 : // length, we guarantee that all the leases we granted before will have a chance to renew
204 : // when we run GC for the first time after restart / transition from AttachedMulti to AttachedSingle.
205 186 : let lsn_lease_deadline = if location.attach_mode == AttachmentMode::Single {
206 186 : Some(
207 186 : tokio::time::Instant::now()
208 186 : + tenant_conf
209 186 : .lsn_lease_length
210 186 : .unwrap_or(LsnLease::DEFAULT_LENGTH),
211 186 : )
212 : } else {
213 : // We don't use `lsn_lease_deadline` to delay GC in AttachedMulti and AttachedStale
214 : // because we don't do GC in these modes.
215 0 : None
216 : };
217 :
218 186 : Self {
219 186 : tenant_conf,
220 186 : location,
221 186 : lsn_lease_deadline,
222 186 : }
223 186 : }
224 :
225 186 : fn try_from(location_conf: LocationConf) -> anyhow::Result<Self> {
226 186 : match &location_conf.mode {
227 186 : LocationMode::Attached(attach_conf) => {
228 186 : Ok(Self::new(location_conf.tenant_conf, *attach_conf))
229 : }
230 : LocationMode::Secondary(_) => {
231 0 : anyhow::bail!("Attempted to construct AttachedTenantConf from a LocationConf in secondary mode")
232 : }
233 : }
234 186 : }
235 :
236 762 : fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
237 762 : self.lsn_lease_deadline
238 762 : .map(|d| tokio::time::Instant::now() < d)
239 762 : .unwrap_or(false)
240 762 : }
241 : }
242 : struct TimelinePreload {
243 : timeline_id: TimelineId,
244 : client: RemoteTimelineClient,
245 : index_part: Result<MaybeDeletedIndexPart, DownloadError>,
246 : }
247 :
248 : pub(crate) struct TenantPreload {
249 : tenant_manifest: TenantManifest,
250 : timelines: HashMap<TimelineId, TimelinePreload>,
251 : }
252 :
253 : /// When we spawn a tenant, there is a special mode for tenant creation that
254 : /// avoids trying to read anything from remote storage.
255 : pub(crate) enum SpawnMode {
256 : /// Activate as soon as possible
257 : Eager,
258 : /// Lazy activation in the background, with the option to skip the queue if the need comes up
259 : Lazy,
260 : }
261 :
262 : ///
263 : /// Tenant consists of multiple timelines. Keep them in a hash table.
264 : ///
265 : pub struct Tenant {
266 : // Global pageserver config parameters
267 : pub conf: &'static PageServerConf,
268 :
269 : /// The value creation timestamp, used to measure activation delay, see:
270 : /// <https://github.com/neondatabase/neon/issues/4025>
271 : constructed_at: Instant,
272 :
273 : state: watch::Sender<TenantState>,
274 :
275 : // Overridden tenant-specific config parameters.
276 : // We keep TenantConfOpt sturct here to preserve the information
277 : // about parameters that are not set.
278 : // This is necessary to allow global config updates.
279 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
280 :
281 : tenant_shard_id: TenantShardId,
282 :
283 : // The detailed sharding information, beyond the number/count in tenant_shard_id
284 : shard_identity: ShardIdentity,
285 :
286 : /// The remote storage generation, used to protect S3 objects from split-brain.
287 : /// Does not change over the lifetime of the [`Tenant`] object.
288 : ///
289 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
290 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
291 : generation: Generation,
292 :
293 : timelines: Mutex<HashMap<TimelineId, Arc<Timeline>>>,
294 :
295 : /// During timeline creation, we first insert the TimelineId to the
296 : /// creating map, then `timelines`, then remove it from the creating map.
297 : /// **Lock order**: if acquiring both, acquire`timelines` before `timelines_creating`
298 : timelines_creating: std::sync::Mutex<HashSet<TimelineId>>,
299 :
300 : /// Possibly offloaded and archived timelines
301 : /// **Lock order**: if acquiring both, acquire`timelines` before `timelines_offloaded`
302 : timelines_offloaded: Mutex<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
303 :
304 : // This mutex prevents creation of new timelines during GC.
305 : // Adding yet another mutex (in addition to `timelines`) is needed because holding
306 : // `timelines` mutex during all GC iteration
307 : // may block for a long time `get_timeline`, `get_timelines_state`,... and other operations
308 : // with timelines, which in turn may cause dropping replication connection, expiration of wait_for_lsn
309 : // timeout...
310 : gc_cs: tokio::sync::Mutex<()>,
311 : walredo_mgr: Option<Arc<WalRedoManager>>,
312 :
313 : // provides access to timeline data sitting in the remote storage
314 : pub(crate) remote_storage: GenericRemoteStorage,
315 :
316 : // Access to global deletion queue for when this tenant wants to schedule a deletion
317 : deletion_queue_client: DeletionQueueClient,
318 :
319 : /// Cached logical sizes updated updated on each [`Tenant::gather_size_inputs`].
320 : cached_logical_sizes: tokio::sync::Mutex<HashMap<(TimelineId, Lsn), u64>>,
321 : cached_synthetic_tenant_size: Arc<AtomicU64>,
322 :
323 : eviction_task_tenant_state: tokio::sync::Mutex<EvictionTaskTenantState>,
324 :
325 : /// Track repeated failures to compact, so that we can back off.
326 : /// Overhead of mutex is acceptable because compaction is done with a multi-second period.
327 : compaction_circuit_breaker: std::sync::Mutex<CircuitBreaker>,
328 :
329 : /// If the tenant is in Activating state, notify this to encourage it
330 : /// to proceed to Active as soon as possible, rather than waiting for lazy
331 : /// background warmup.
332 : pub(crate) activate_now_sem: tokio::sync::Semaphore,
333 :
334 : /// Time it took for the tenant to activate. Zero if not active yet.
335 : attach_wal_lag_cooldown: Arc<std::sync::OnceLock<WalLagCooldown>>,
336 :
337 : // Cancellation token fires when we have entered shutdown(). This is a parent of
338 : // Timelines' cancellation token.
339 : pub(crate) cancel: CancellationToken,
340 :
341 : // Users of the Tenant such as the page service must take this Gate to avoid
342 : // trying to use a Tenant which is shutting down.
343 : pub(crate) gate: Gate,
344 :
345 : /// Throttle applied at the top of [`Timeline::get`].
346 : /// All [`Tenant::timelines`] of a given [`Tenant`] instance share the same [`throttle::Throttle`] instance.
347 : pub(crate) timeline_get_throttle:
348 : Arc<throttle::Throttle<crate::metrics::tenant_throttling::TimelineGet>>,
349 :
350 : /// An ongoing timeline detach concurrency limiter.
351 : ///
352 : /// As a tenant will likely be restarted as part of timeline detach ancestor it makes no sense
353 : /// to have two running at the same time. A different one can be started if an earlier one
354 : /// has failed for whatever reason.
355 : ongoing_timeline_detach: std::sync::Mutex<Option<(TimelineId, utils::completion::Barrier)>>,
356 :
357 : /// `index_part.json` based gc blocking reason tracking.
358 : ///
359 : /// New gc iterations must start a new iteration by acquiring `GcBlock::start` before
360 : /// proceeding.
361 : pub(crate) gc_block: gc_block::GcBlock,
362 :
363 : l0_flush_global_state: L0FlushGlobalState,
364 : }
365 :
366 : impl std::fmt::Debug for Tenant {
367 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
368 0 : write!(f, "{} ({})", self.tenant_shard_id, self.current_state())
369 0 : }
370 : }
371 :
372 : pub(crate) enum WalRedoManager {
373 : Prod(WalredoManagerId, PostgresRedoManager),
374 : #[cfg(test)]
375 : Test(harness::TestRedoManager),
376 : }
377 :
378 0 : #[derive(thiserror::Error, Debug)]
379 : #[error("pageserver is shutting down")]
380 : pub(crate) struct GlobalShutDown;
381 :
382 : impl WalRedoManager {
383 0 : pub(crate) fn new(mgr: PostgresRedoManager) -> Result<Arc<Self>, GlobalShutDown> {
384 0 : let id = WalredoManagerId::next();
385 0 : let arc = Arc::new(Self::Prod(id, mgr));
386 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
387 0 : match &mut *guard {
388 0 : Some(map) => {
389 0 : map.insert(id, Arc::downgrade(&arc));
390 0 : Ok(arc)
391 : }
392 0 : None => Err(GlobalShutDown),
393 : }
394 0 : }
395 : }
396 :
397 : impl Drop for WalRedoManager {
398 10 : fn drop(&mut self) {
399 10 : match self {
400 0 : Self::Prod(id, _) => {
401 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
402 0 : if let Some(map) = &mut *guard {
403 0 : map.remove(id).expect("new() registers, drop() unregisters");
404 0 : }
405 : }
406 : #[cfg(test)]
407 10 : Self::Test(_) => {
408 10 : // Not applicable to test redo manager
409 10 : }
410 : }
411 10 : }
412 : }
413 :
414 : /// Global registry of all walredo managers so that [`crate::shutdown_pageserver`] can shut down
415 : /// the walredo processes outside of the regular order.
416 : ///
417 : /// This is necessary to work around a systemd bug where it freezes if there are
418 : /// walredo processes left => <https://github.com/neondatabase/cloud/issues/11387>
419 : #[allow(clippy::type_complexity)]
420 : pub(crate) static WALREDO_MANAGERS: once_cell::sync::Lazy<
421 : Mutex<Option<HashMap<WalredoManagerId, Weak<WalRedoManager>>>>,
422 0 : > = once_cell::sync::Lazy::new(|| Mutex::new(Some(HashMap::new())));
423 : #[derive(PartialEq, Eq, Hash, Clone, Copy, Debug)]
424 : pub(crate) struct WalredoManagerId(u64);
425 : impl WalredoManagerId {
426 0 : pub fn next() -> Self {
427 : static NEXT: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(1);
428 0 : let id = NEXT.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
429 0 : if id == 0 {
430 0 : panic!("WalredoManagerId::new() returned 0, indicating wraparound, risking it's no longer unique");
431 0 : }
432 0 : Self(id)
433 0 : }
434 : }
435 :
436 : #[cfg(test)]
437 : impl From<harness::TestRedoManager> for WalRedoManager {
438 186 : fn from(mgr: harness::TestRedoManager) -> Self {
439 186 : Self::Test(mgr)
440 186 : }
441 : }
442 :
443 : impl WalRedoManager {
444 6 : pub(crate) async fn shutdown(&self) -> bool {
445 6 : match self {
446 0 : Self::Prod(_, mgr) => mgr.shutdown().await,
447 : #[cfg(test)]
448 : Self::Test(_) => {
449 : // Not applicable to test redo manager
450 6 : true
451 : }
452 : }
453 6 : }
454 :
455 0 : pub(crate) fn maybe_quiesce(&self, idle_timeout: Duration) {
456 0 : match self {
457 0 : Self::Prod(_, mgr) => mgr.maybe_quiesce(idle_timeout),
458 0 : #[cfg(test)]
459 0 : Self::Test(_) => {
460 0 : // Not applicable to test redo manager
461 0 : }
462 0 : }
463 0 : }
464 :
465 : /// # Cancel-Safety
466 : ///
467 : /// This method is cancellation-safe.
468 342 : pub async fn request_redo(
469 342 : &self,
470 342 : key: crate::repository::Key,
471 342 : lsn: Lsn,
472 342 : base_img: Option<(Lsn, bytes::Bytes)>,
473 342 : records: Vec<(Lsn, crate::walrecord::NeonWalRecord)>,
474 342 : pg_version: u32,
475 342 : ) -> Result<bytes::Bytes, walredo::Error> {
476 342 : match self {
477 0 : Self::Prod(_, mgr) => {
478 0 : mgr.request_redo(key, lsn, base_img, records, pg_version)
479 0 : .await
480 : }
481 : #[cfg(test)]
482 342 : Self::Test(mgr) => {
483 342 : mgr.request_redo(key, lsn, base_img, records, pg_version)
484 0 : .await
485 : }
486 : }
487 342 : }
488 :
489 0 : pub(crate) fn status(&self) -> Option<WalRedoManagerStatus> {
490 0 : match self {
491 0 : WalRedoManager::Prod(_, m) => Some(m.status()),
492 0 : #[cfg(test)]
493 0 : WalRedoManager::Test(_) => None,
494 0 : }
495 0 : }
496 : }
497 :
498 : /// A very lightweight memory representation of an offloaded timeline.
499 : ///
500 : /// We need to store the list of offloaded timelines so that we can perform operations on them,
501 : /// like unoffloading them, or (at a later date), decide to perform flattening.
502 : /// This type has a much smaller memory impact than [`Timeline`], and thus we can store many
503 : /// more offloaded timelines than we can manage ones that aren't.
504 : pub struct OffloadedTimeline {
505 : pub tenant_shard_id: TenantShardId,
506 : pub timeline_id: TimelineId,
507 : pub ancestor_timeline_id: Option<TimelineId>,
508 : /// Whether to retain the branch lsn at the ancestor or not
509 : pub ancestor_retain_lsn: Option<Lsn>,
510 :
511 : /// When the timeline was archived.
512 : ///
513 : /// Present for future flattening deliberations.
514 : pub archived_at: NaiveDateTime,
515 :
516 : /// Lazily constructed remote client for the timeline
517 : ///
518 : /// If we offload a timeline, we keep around the remote client
519 : /// for the duration of the process. If we find it through the
520 : /// manifest, we don't construct it up until it's needed (deletion).
521 : pub remote_client: Option<Arc<RemoteTimelineClient>>,
522 :
523 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
524 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
525 : pub delete_progress: TimelineDeleteProgress,
526 : }
527 :
528 : impl OffloadedTimeline {
529 : /// Obtains an offloaded timeline from a given timeline object.
530 : ///
531 : /// Returns `None` if the `archived_at` flag couldn't be obtained, i.e.
532 : /// the timeline is not in a stopped state.
533 : /// Panics if the timeline is not archived.
534 0 : fn from_timeline(timeline: &Timeline) -> Result<Self, UploadQueueNotReadyError> {
535 0 : let ancestor_retain_lsn = timeline
536 0 : .get_ancestor_timeline_id()
537 0 : .map(|_timeline_id| timeline.get_ancestor_lsn());
538 0 : let archived_at = timeline
539 0 : .remote_client
540 0 : .archived_at_stopped_queue()?
541 0 : .expect("must be called on an archived timeline");
542 0 : Ok(Self {
543 0 : tenant_shard_id: timeline.tenant_shard_id,
544 0 : timeline_id: timeline.timeline_id,
545 0 : ancestor_timeline_id: timeline.get_ancestor_timeline_id(),
546 0 : ancestor_retain_lsn,
547 0 : archived_at,
548 0 :
549 0 : remote_client: Some(timeline.remote_client.clone()),
550 0 : delete_progress: timeline.delete_progress.clone(),
551 0 : })
552 0 : }
553 0 : fn from_manifest(tenant_shard_id: TenantShardId, manifest: &OffloadedTimelineManifest) -> Self {
554 0 : let OffloadedTimelineManifest {
555 0 : timeline_id,
556 0 : ancestor_timeline_id,
557 0 : ancestor_retain_lsn,
558 0 : archived_at,
559 0 : } = *manifest;
560 0 : Self {
561 0 : tenant_shard_id,
562 0 : timeline_id,
563 0 : ancestor_timeline_id,
564 0 : ancestor_retain_lsn,
565 0 : archived_at,
566 0 : remote_client: None,
567 0 : delete_progress: TimelineDeleteProgress::default(),
568 0 : }
569 0 : }
570 0 : fn manifest(&self) -> OffloadedTimelineManifest {
571 0 : let Self {
572 0 : timeline_id,
573 0 : ancestor_timeline_id,
574 0 : ancestor_retain_lsn,
575 0 : archived_at,
576 0 : ..
577 0 : } = self;
578 0 : OffloadedTimelineManifest {
579 0 : timeline_id: *timeline_id,
580 0 : ancestor_timeline_id: *ancestor_timeline_id,
581 0 : ancestor_retain_lsn: *ancestor_retain_lsn,
582 0 : archived_at: *archived_at,
583 0 : }
584 0 : }
585 : }
586 :
587 : #[derive(Copy, Clone, PartialEq, Eq, Hash, Debug)]
588 : pub enum MaybeOffloaded {
589 : Yes,
590 : No,
591 : }
592 :
593 : #[derive(Clone)]
594 : pub enum TimelineOrOffloaded {
595 : Timeline(Arc<Timeline>),
596 : Offloaded(Arc<OffloadedTimeline>),
597 : }
598 :
599 : impl TimelineOrOffloaded {
600 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
601 0 : match self {
602 0 : TimelineOrOffloaded::Timeline(timeline) => timeline.tenant_shard_id,
603 0 : TimelineOrOffloaded::Offloaded(offloaded) => offloaded.tenant_shard_id,
604 : }
605 0 : }
606 0 : pub fn timeline_id(&self) -> TimelineId {
607 0 : match self {
608 0 : TimelineOrOffloaded::Timeline(timeline) => timeline.timeline_id,
609 0 : TimelineOrOffloaded::Offloaded(offloaded) => offloaded.timeline_id,
610 : }
611 0 : }
612 0 : pub fn delete_progress(&self) -> &Arc<tokio::sync::Mutex<DeleteTimelineFlow>> {
613 0 : match self {
614 0 : TimelineOrOffloaded::Timeline(timeline) => &timeline.delete_progress,
615 0 : TimelineOrOffloaded::Offloaded(offloaded) => &offloaded.delete_progress,
616 : }
617 0 : }
618 0 : pub fn remote_client_maybe_construct(&self, tenant: &Tenant) -> Arc<RemoteTimelineClient> {
619 0 : match self {
620 0 : TimelineOrOffloaded::Timeline(timeline) => timeline.remote_client.clone(),
621 0 : TimelineOrOffloaded::Offloaded(offloaded) => match offloaded.remote_client.clone() {
622 0 : Some(remote_client) => remote_client,
623 : None => {
624 0 : let remote_client = tenant.build_timeline_client(
625 0 : offloaded.timeline_id,
626 0 : tenant.remote_storage.clone(),
627 0 : );
628 0 : Arc::new(remote_client)
629 : }
630 : },
631 : }
632 0 : }
633 : }
634 :
635 0 : #[derive(Debug, thiserror::Error, PartialEq, Eq)]
636 : pub enum GetTimelineError {
637 : #[error("Timeline is shutting down")]
638 : ShuttingDown,
639 : #[error("Timeline {tenant_id}/{timeline_id} is not active, state: {state:?}")]
640 : NotActive {
641 : tenant_id: TenantShardId,
642 : timeline_id: TimelineId,
643 : state: TimelineState,
644 : },
645 : #[error("Timeline {tenant_id}/{timeline_id} was not found")]
646 : NotFound {
647 : tenant_id: TenantShardId,
648 : timeline_id: TimelineId,
649 : },
650 : }
651 :
652 0 : #[derive(Debug, thiserror::Error)]
653 : pub enum LoadLocalTimelineError {
654 : #[error("FailedToLoad")]
655 : Load(#[source] anyhow::Error),
656 : #[error("FailedToResumeDeletion")]
657 : ResumeDeletion(#[source] anyhow::Error),
658 : }
659 :
660 0 : #[derive(thiserror::Error)]
661 : pub enum DeleteTimelineError {
662 : #[error("NotFound")]
663 : NotFound,
664 :
665 : #[error("HasChildren")]
666 : HasChildren(Vec<TimelineId>),
667 :
668 : #[error("Timeline deletion is already in progress")]
669 : AlreadyInProgress(Arc<tokio::sync::Mutex<DeleteTimelineFlow>>),
670 :
671 : #[error(transparent)]
672 : Other(#[from] anyhow::Error),
673 : }
674 :
675 : impl Debug for DeleteTimelineError {
676 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
677 0 : match self {
678 0 : Self::NotFound => write!(f, "NotFound"),
679 0 : Self::HasChildren(c) => f.debug_tuple("HasChildren").field(c).finish(),
680 0 : Self::AlreadyInProgress(_) => f.debug_tuple("AlreadyInProgress").finish(),
681 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
682 : }
683 0 : }
684 : }
685 :
686 0 : #[derive(thiserror::Error)]
687 : pub enum TimelineArchivalError {
688 : #[error("NotFound")]
689 : NotFound,
690 :
691 : #[error("Timeout")]
692 : Timeout,
693 :
694 : #[error("Cancelled")]
695 : Cancelled,
696 :
697 : #[error("ancestor is archived: {}", .0)]
698 : HasArchivedParent(TimelineId),
699 :
700 : #[error("HasUnarchivedChildren")]
701 : HasUnarchivedChildren(Vec<TimelineId>),
702 :
703 : #[error("Timeline archival is already in progress")]
704 : AlreadyInProgress,
705 :
706 : #[error(transparent)]
707 : Other(anyhow::Error),
708 : }
709 :
710 : impl Debug for TimelineArchivalError {
711 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
712 0 : match self {
713 0 : Self::NotFound => write!(f, "NotFound"),
714 0 : Self::Timeout => write!(f, "Timeout"),
715 0 : Self::Cancelled => write!(f, "Cancelled"),
716 0 : Self::HasArchivedParent(p) => f.debug_tuple("HasArchivedParent").field(p).finish(),
717 0 : Self::HasUnarchivedChildren(c) => {
718 0 : f.debug_tuple("HasUnarchivedChildren").field(c).finish()
719 : }
720 0 : Self::AlreadyInProgress => f.debug_tuple("AlreadyInProgress").finish(),
721 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
722 : }
723 0 : }
724 : }
725 :
726 : pub enum SetStoppingError {
727 : AlreadyStopping(completion::Barrier),
728 : Broken,
729 : }
730 :
731 : impl Debug for SetStoppingError {
732 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
733 0 : match self {
734 0 : Self::AlreadyStopping(_) => f.debug_tuple("AlreadyStopping").finish(),
735 0 : Self::Broken => write!(f, "Broken"),
736 : }
737 0 : }
738 : }
739 :
740 : /// Arguments to [`Tenant::create_timeline`].
741 : ///
742 : /// Not usable as an idempotency key for timeline creation because if [`CreateTimelineParamsBranch::ancestor_start_lsn`]
743 : /// is `None`, the result of the timeline create call is not deterministic.
744 : ///
745 : /// See [`CreateTimelineIdempotency`] for an idempotency key.
746 : #[derive(Debug)]
747 : pub(crate) enum CreateTimelineParams {
748 : Bootstrap(CreateTimelineParamsBootstrap),
749 : Branch(CreateTimelineParamsBranch),
750 : }
751 :
752 : #[derive(Debug)]
753 : pub(crate) struct CreateTimelineParamsBootstrap {
754 : pub(crate) new_timeline_id: TimelineId,
755 : pub(crate) existing_initdb_timeline_id: Option<TimelineId>,
756 : pub(crate) pg_version: u32,
757 : }
758 :
759 : /// NB: See comment on [`CreateTimelineIdempotency::Branch`] for why there's no `pg_version` here.
760 : #[derive(Debug)]
761 : pub(crate) struct CreateTimelineParamsBranch {
762 : pub(crate) new_timeline_id: TimelineId,
763 : pub(crate) ancestor_timeline_id: TimelineId,
764 : pub(crate) ancestor_start_lsn: Option<Lsn>,
765 : }
766 :
767 : /// What is used to determine idempotency of a [`Tenant::create_timeline`] call.
768 : ///
769 : /// Unlike [`CreateTimelineParams`], ancestor LSN is fixed, so, branching will be at a deterministic LSN.
770 : ///
771 : /// We make some trade-offs though, e.g., [`CreateTimelineParamsBootstrap::existing_initdb_timeline_id`]
772 : /// is not considered for idempotency.
773 : ///
774 : /// We can improve on this over time.
775 : pub(crate) enum CreateTimelineIdempotency {
776 : Bootstrap {
777 : pg_version: u32,
778 : },
779 : /// NB: branches always have the same `pg_version` as their ancestor.
780 : /// While [`pageserver_api::models::TimelineCreateRequestMode::Branch::pg_version`]
781 : /// exists as a field, and is set by cplane, it has always been ignored by pageserver when
782 : /// determining the child branch pg_version.
783 : Branch {
784 : ancestor_timeline_id: TimelineId,
785 : ancestor_start_lsn: Lsn,
786 : },
787 : }
788 :
789 : /// What is returned by [`Tenant::start_creating_timeline`].
790 : #[must_use]
791 : enum StartCreatingTimelineResult<'t> {
792 : CreateGuard(TimelineCreateGuard<'t>),
793 : Idempotent(Arc<Timeline>),
794 : }
795 :
796 : /// What is returned by [`Tenant::create_timeline`].
797 : enum CreateTimelineResult {
798 : Created(Arc<Timeline>),
799 : Idempotent(Arc<Timeline>),
800 : }
801 :
802 : impl CreateTimelineResult {
803 0 : fn timeline(&self) -> &Arc<Timeline> {
804 0 : match self {
805 0 : Self::Created(t) | Self::Idempotent(t) => t,
806 0 : }
807 0 : }
808 : /// Unit test timelines aren't activated, test has to do it if it needs to.
809 : #[cfg(test)]
810 228 : fn into_timeline_for_test(self) -> Arc<Timeline> {
811 228 : match self {
812 228 : Self::Created(t) | Self::Idempotent(t) => t,
813 228 : }
814 228 : }
815 : }
816 :
817 0 : #[derive(thiserror::Error, Debug)]
818 : pub enum CreateTimelineError {
819 : #[error("creation of timeline with the given ID is in progress")]
820 : AlreadyCreating,
821 : #[error("timeline already exists with different parameters")]
822 : Conflict,
823 : #[error(transparent)]
824 : AncestorLsn(anyhow::Error),
825 : #[error("ancestor timeline is not active")]
826 : AncestorNotActive,
827 : #[error("ancestor timeline is archived")]
828 : AncestorArchived,
829 : #[error("tenant shutting down")]
830 : ShuttingDown,
831 : #[error(transparent)]
832 : Other(#[from] anyhow::Error),
833 : }
834 :
835 : #[derive(thiserror::Error, Debug)]
836 : enum InitdbError {
837 : Other(anyhow::Error),
838 : Cancelled,
839 : Spawn(std::io::Result<()>),
840 : Failed(std::process::ExitStatus, Vec<u8>),
841 : }
842 :
843 : impl fmt::Display for InitdbError {
844 0 : fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
845 0 : match self {
846 0 : InitdbError::Cancelled => write!(f, "Operation was cancelled"),
847 0 : InitdbError::Spawn(e) => write!(f, "Spawn error: {:?}", e),
848 0 : InitdbError::Failed(status, stderr) => write!(
849 0 : f,
850 0 : "Command failed with status {:?}: {}",
851 0 : status,
852 0 : String::from_utf8_lossy(stderr)
853 0 : ),
854 0 : InitdbError::Other(e) => write!(f, "Error: {:?}", e),
855 : }
856 0 : }
857 : }
858 :
859 : impl From<std::io::Error> for InitdbError {
860 0 : fn from(error: std::io::Error) -> Self {
861 0 : InitdbError::Spawn(Err(error))
862 0 : }
863 : }
864 :
865 : enum CreateTimelineCause {
866 : Load,
867 : Delete,
868 : }
869 :
870 0 : #[derive(thiserror::Error, Debug)]
871 : pub(crate) enum GcError {
872 : // The tenant is shutting down
873 : #[error("tenant shutting down")]
874 : TenantCancelled,
875 :
876 : // The tenant is shutting down
877 : #[error("timeline shutting down")]
878 : TimelineCancelled,
879 :
880 : // The tenant is in a state inelegible to run GC
881 : #[error("not active")]
882 : NotActive,
883 :
884 : // A requested GC cutoff LSN was invalid, for example it tried to move backwards
885 : #[error("not active")]
886 : BadLsn { why: String },
887 :
888 : // A remote storage error while scheduling updates after compaction
889 : #[error(transparent)]
890 : Remote(anyhow::Error),
891 :
892 : // An error reading while calculating GC cutoffs
893 : #[error(transparent)]
894 : GcCutoffs(PageReconstructError),
895 :
896 : // If GC was invoked for a particular timeline, this error means it didn't exist
897 : #[error("timeline not found")]
898 : TimelineNotFound,
899 : }
900 :
901 : impl From<PageReconstructError> for GcError {
902 0 : fn from(value: PageReconstructError) -> Self {
903 0 : match value {
904 0 : PageReconstructError::Cancelled => Self::TimelineCancelled,
905 0 : other => Self::GcCutoffs(other),
906 : }
907 0 : }
908 : }
909 :
910 : impl From<NotInitialized> for GcError {
911 0 : fn from(value: NotInitialized) -> Self {
912 0 : match value {
913 0 : NotInitialized::Uninitialized => GcError::Remote(value.into()),
914 0 : NotInitialized::Stopped | NotInitialized::ShuttingDown => GcError::TimelineCancelled,
915 : }
916 0 : }
917 : }
918 :
919 : impl From<timeline::layer_manager::Shutdown> for GcError {
920 0 : fn from(_: timeline::layer_manager::Shutdown) -> Self {
921 0 : GcError::TimelineCancelled
922 0 : }
923 : }
924 :
925 0 : #[derive(thiserror::Error, Debug)]
926 : pub(crate) enum LoadConfigError {
927 : #[error("TOML deserialization error: '{0}'")]
928 : DeserializeToml(#[from] toml_edit::de::Error),
929 :
930 : #[error("Config not found at {0}")]
931 : NotFound(Utf8PathBuf),
932 : }
933 :
934 : impl Tenant {
935 : /// Yet another helper for timeline initialization.
936 : ///
937 : /// - Initializes the Timeline struct and inserts it into the tenant's hash map
938 : /// - Scans the local timeline directory for layer files and builds the layer map
939 : /// - Downloads remote index file and adds remote files to the layer map
940 : /// - Schedules remote upload tasks for any files that are present locally but missing from remote storage.
941 : ///
942 : /// If the operation fails, the timeline is left in the tenant's hash map in Broken state. On success,
943 : /// it is marked as Active.
944 : #[allow(clippy::too_many_arguments)]
945 6 : async fn timeline_init_and_sync(
946 6 : &self,
947 6 : timeline_id: TimelineId,
948 6 : resources: TimelineResources,
949 6 : index_part: IndexPart,
950 6 : metadata: TimelineMetadata,
951 6 : ancestor: Option<Arc<Timeline>>,
952 6 : _ctx: &RequestContext,
953 6 : ) -> anyhow::Result<()> {
954 6 : let tenant_id = self.tenant_shard_id;
955 :
956 6 : let timeline = self.create_timeline_struct(
957 6 : timeline_id,
958 6 : &metadata,
959 6 : ancestor.clone(),
960 6 : resources,
961 6 : CreateTimelineCause::Load,
962 6 : )?;
963 6 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
964 6 : anyhow::ensure!(
965 6 : disk_consistent_lsn.is_valid(),
966 0 : "Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn"
967 : );
968 6 : assert_eq!(
969 6 : disk_consistent_lsn,
970 6 : metadata.disk_consistent_lsn(),
971 0 : "these are used interchangeably"
972 : );
973 :
974 6 : timeline.remote_client.init_upload_queue(&index_part)?;
975 :
976 6 : timeline
977 6 : .load_layer_map(disk_consistent_lsn, index_part)
978 6 : .await
979 6 : .with_context(|| {
980 0 : format!("Failed to load layermap for timeline {tenant_id}/{timeline_id}")
981 6 : })?;
982 :
983 : {
984 : // avoiding holding it across awaits
985 6 : let mut timelines_accessor = self.timelines.lock().unwrap();
986 6 : match timelines_accessor.entry(timeline_id) {
987 : // We should never try and load the same timeline twice during startup
988 : Entry::Occupied(_) => {
989 0 : unreachable!(
990 0 : "Timeline {tenant_id}/{timeline_id} already exists in the tenant map"
991 0 : );
992 : }
993 6 : Entry::Vacant(v) => {
994 6 : v.insert(Arc::clone(&timeline));
995 6 : timeline.maybe_spawn_flush_loop();
996 6 : }
997 6 : }
998 6 : };
999 6 :
1000 6 : // Sanity check: a timeline should have some content.
1001 6 : anyhow::ensure!(
1002 6 : ancestor.is_some()
1003 4 : || timeline
1004 4 : .layers
1005 4 : .read()
1006 0 : .await
1007 4 : .layer_map()
1008 4 : .expect("currently loading, layer manager cannot be shutdown already")
1009 4 : .iter_historic_layers()
1010 4 : .next()
1011 4 : .is_some(),
1012 0 : "Timeline has no ancestor and no layer files"
1013 : );
1014 :
1015 6 : Ok(())
1016 6 : }
1017 :
1018 : /// Attach a tenant that's available in cloud storage.
1019 : ///
1020 : /// This returns quickly, after just creating the in-memory object
1021 : /// Tenant struct and launching a background task to download
1022 : /// the remote index files. On return, the tenant is most likely still in
1023 : /// Attaching state, and it will become Active once the background task
1024 : /// finishes. You can use wait_until_active() to wait for the task to
1025 : /// complete.
1026 : ///
1027 : #[allow(clippy::too_many_arguments)]
1028 0 : pub(crate) fn spawn(
1029 0 : conf: &'static PageServerConf,
1030 0 : tenant_shard_id: TenantShardId,
1031 0 : resources: TenantSharedResources,
1032 0 : attached_conf: AttachedTenantConf,
1033 0 : shard_identity: ShardIdentity,
1034 0 : init_order: Option<InitializationOrder>,
1035 0 : mode: SpawnMode,
1036 0 : ctx: &RequestContext,
1037 0 : ) -> Result<Arc<Tenant>, GlobalShutDown> {
1038 0 : let wal_redo_manager =
1039 0 : WalRedoManager::new(PostgresRedoManager::new(conf, tenant_shard_id))?;
1040 :
1041 : let TenantSharedResources {
1042 0 : broker_client,
1043 0 : remote_storage,
1044 0 : deletion_queue_client,
1045 0 : l0_flush_global_state,
1046 0 : } = resources;
1047 0 :
1048 0 : let attach_mode = attached_conf.location.attach_mode;
1049 0 : let generation = attached_conf.location.generation;
1050 0 :
1051 0 : let tenant = Arc::new(Tenant::new(
1052 0 : TenantState::Attaching,
1053 0 : conf,
1054 0 : attached_conf,
1055 0 : shard_identity,
1056 0 : Some(wal_redo_manager),
1057 0 : tenant_shard_id,
1058 0 : remote_storage.clone(),
1059 0 : deletion_queue_client,
1060 0 : l0_flush_global_state,
1061 0 : ));
1062 0 :
1063 0 : // The attach task will carry a GateGuard, so that shutdown() reliably waits for it to drop out if
1064 0 : // we shut down while attaching.
1065 0 : let attach_gate_guard = tenant
1066 0 : .gate
1067 0 : .enter()
1068 0 : .expect("We just created the Tenant: nothing else can have shut it down yet");
1069 0 :
1070 0 : // Do all the hard work in the background
1071 0 : let tenant_clone = Arc::clone(&tenant);
1072 0 : let ctx = ctx.detached_child(TaskKind::Attach, DownloadBehavior::Warn);
1073 0 : task_mgr::spawn(
1074 0 : &tokio::runtime::Handle::current(),
1075 0 : TaskKind::Attach,
1076 0 : tenant_shard_id,
1077 0 : None,
1078 0 : "attach tenant",
1079 0 : async move {
1080 0 :
1081 0 : info!(
1082 : ?attach_mode,
1083 0 : "Attaching tenant"
1084 : );
1085 :
1086 0 : let _gate_guard = attach_gate_guard;
1087 0 :
1088 0 : // Is this tenant being spawned as part of process startup?
1089 0 : let starting_up = init_order.is_some();
1090 0 : scopeguard::defer! {
1091 0 : if starting_up {
1092 0 : TENANT.startup_complete.inc();
1093 0 : }
1094 0 : }
1095 :
1096 : // Ideally we should use Tenant::set_broken_no_wait, but it is not supposed to be used when tenant is in loading state.
1097 : enum BrokenVerbosity {
1098 : Error,
1099 : Info
1100 : }
1101 0 : let make_broken =
1102 0 : |t: &Tenant, err: anyhow::Error, verbosity: BrokenVerbosity| {
1103 0 : match verbosity {
1104 : BrokenVerbosity::Info => {
1105 0 : info!("attach cancelled, setting tenant state to Broken: {err}");
1106 : },
1107 : BrokenVerbosity::Error => {
1108 0 : error!("attach failed, setting tenant state to Broken: {err:?}");
1109 : }
1110 : }
1111 0 : t.state.send_modify(|state| {
1112 0 : // The Stopping case is for when we have passed control on to DeleteTenantFlow:
1113 0 : // if it errors, we will call make_broken when tenant is already in Stopping.
1114 0 : assert!(
1115 0 : matches!(*state, TenantState::Attaching | TenantState::Stopping { .. }),
1116 0 : "the attach task owns the tenant state until activation is complete"
1117 : );
1118 :
1119 0 : *state = TenantState::broken_from_reason(err.to_string());
1120 0 : });
1121 0 : };
1122 :
1123 : // TODO: should also be rejecting tenant conf changes that violate this check.
1124 0 : if let Err(e) = crate::tenant::storage_layer::inmemory_layer::IndexEntry::validate_checkpoint_distance(tenant_clone.get_checkpoint_distance()) {
1125 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1126 0 : return Ok(());
1127 0 : }
1128 0 :
1129 0 : let mut init_order = init_order;
1130 0 : // take the completion because initial tenant loading will complete when all of
1131 0 : // these tasks complete.
1132 0 : let _completion = init_order
1133 0 : .as_mut()
1134 0 : .and_then(|x| x.initial_tenant_load.take());
1135 0 : let remote_load_completion = init_order
1136 0 : .as_mut()
1137 0 : .and_then(|x| x.initial_tenant_load_remote.take());
1138 :
1139 : enum AttachType<'a> {
1140 : /// We are attaching this tenant lazily in the background.
1141 : Warmup {
1142 : _permit: tokio::sync::SemaphorePermit<'a>,
1143 : during_startup: bool
1144 : },
1145 : /// We are attaching this tenant as soon as we can, because for example an
1146 : /// endpoint tried to access it.
1147 : OnDemand,
1148 : /// During normal operations after startup, we are attaching a tenant, and
1149 : /// eager attach was requested.
1150 : Normal,
1151 : }
1152 :
1153 0 : let attach_type = if matches!(mode, SpawnMode::Lazy) {
1154 : // Before doing any I/O, wait for at least one of:
1155 : // - A client attempting to access to this tenant (on-demand loading)
1156 : // - A permit becoming available in the warmup semaphore (background warmup)
1157 :
1158 0 : tokio::select!(
1159 0 : permit = tenant_clone.activate_now_sem.acquire() => {
1160 0 : let _ = permit.expect("activate_now_sem is never closed");
1161 0 : tracing::info!("Activating tenant (on-demand)");
1162 0 : AttachType::OnDemand
1163 : },
1164 0 : permit = conf.concurrent_tenant_warmup.inner().acquire() => {
1165 0 : let _permit = permit.expect("concurrent_tenant_warmup semaphore is never closed");
1166 0 : tracing::info!("Activating tenant (warmup)");
1167 0 : AttachType::Warmup {
1168 0 : _permit,
1169 0 : during_startup: init_order.is_some()
1170 0 : }
1171 : }
1172 0 : _ = tenant_clone.cancel.cancelled() => {
1173 : // This is safe, but should be pretty rare: it is interesting if a tenant
1174 : // stayed in Activating for such a long time that shutdown found it in
1175 : // that state.
1176 0 : tracing::info!(state=%tenant_clone.current_state(), "Tenant shut down before activation");
1177 : // Make the tenant broken so that set_stopping will not hang waiting for it to leave
1178 : // the Attaching state. This is an over-reaction (nothing really broke, the tenant is
1179 : // just shutting down), but ensures progress.
1180 0 : make_broken(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"), BrokenVerbosity::Info);
1181 0 : return Ok(());
1182 : },
1183 : )
1184 : } else {
1185 : // SpawnMode::{Create,Eager} always cause jumping ahead of the
1186 : // concurrent_tenant_warmup queue
1187 0 : AttachType::Normal
1188 : };
1189 :
1190 0 : let preload = match &mode {
1191 : SpawnMode::Eager | SpawnMode::Lazy => {
1192 0 : let _preload_timer = TENANT.preload.start_timer();
1193 0 : let res = tenant_clone
1194 0 : .preload(&remote_storage, task_mgr::shutdown_token())
1195 0 : .await;
1196 0 : match res {
1197 0 : Ok(p) => Some(p),
1198 0 : Err(e) => {
1199 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1200 0 : return Ok(());
1201 : }
1202 : }
1203 : }
1204 :
1205 : };
1206 :
1207 : // Remote preload is complete.
1208 0 : drop(remote_load_completion);
1209 0 :
1210 0 :
1211 0 : // We will time the duration of the attach phase unless this is a creation (attach will do no work)
1212 0 : let attach_start = std::time::Instant::now();
1213 0 : let attached = {
1214 0 : let _attach_timer = Some(TENANT.attach.start_timer());
1215 0 : tenant_clone.attach(preload, &ctx).await
1216 : };
1217 0 : let attach_duration = attach_start.elapsed();
1218 0 : _ = tenant_clone.attach_wal_lag_cooldown.set(WalLagCooldown::new(attach_start, attach_duration));
1219 0 :
1220 0 : match attached {
1221 : Ok(()) => {
1222 0 : info!("attach finished, activating");
1223 0 : tenant_clone.activate(broker_client, None, &ctx);
1224 : }
1225 0 : Err(e) => {
1226 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1227 0 : }
1228 : }
1229 :
1230 : // If we are doing an opportunistic warmup attachment at startup, initialize
1231 : // logical size at the same time. This is better than starting a bunch of idle tenants
1232 : // with cold caches and then coming back later to initialize their logical sizes.
1233 : //
1234 : // It also prevents the warmup proccess competing with the concurrency limit on
1235 : // logical size calculations: if logical size calculation semaphore is saturated,
1236 : // then warmup will wait for that before proceeding to the next tenant.
1237 0 : if matches!(attach_type, AttachType::Warmup { during_startup: true, .. }) {
1238 0 : let mut futs: FuturesUnordered<_> = tenant_clone.timelines.lock().unwrap().values().cloned().map(|t| t.await_initial_logical_size()).collect();
1239 0 : tracing::info!("Waiting for initial logical sizes while warming up...");
1240 0 : while futs.next().await.is_some() {}
1241 0 : tracing::info!("Warm-up complete");
1242 0 : }
1243 :
1244 0 : Ok(())
1245 0 : }
1246 0 : .instrument(tracing::info_span!(parent: None, "attach", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), gen=?generation)),
1247 : );
1248 0 : Ok(tenant)
1249 0 : }
1250 :
1251 186 : #[instrument(skip_all)]
1252 : pub(crate) async fn preload(
1253 : self: &Arc<Self>,
1254 : remote_storage: &GenericRemoteStorage,
1255 : cancel: CancellationToken,
1256 : ) -> anyhow::Result<TenantPreload> {
1257 : span::debug_assert_current_span_has_tenant_id();
1258 : // Get list of remote timelines
1259 : // download index files for every tenant timeline
1260 : info!("listing remote timelines");
1261 : let (remote_timeline_ids, other_keys) = remote_timeline_client::list_remote_timelines(
1262 : remote_storage,
1263 : self.tenant_shard_id,
1264 : cancel.clone(),
1265 : )
1266 : .await?;
1267 : let (offloaded_add, tenant_manifest) =
1268 : match remote_timeline_client::do_download_tenant_manifest(
1269 : remote_storage,
1270 : &self.tenant_shard_id,
1271 : &cancel,
1272 : )
1273 : .await
1274 : {
1275 : Ok((tenant_manifest, _generation)) => (
1276 : format!("{} offloaded", tenant_manifest.offloaded_timelines.len()),
1277 : tenant_manifest,
1278 : ),
1279 : Err(DownloadError::NotFound) => {
1280 : ("no manifest".to_string(), TenantManifest::empty())
1281 : }
1282 : Err(e) => Err(e)?,
1283 : };
1284 :
1285 : info!(
1286 : "found {} timelines, and {offloaded_add}",
1287 : remote_timeline_ids.len()
1288 : );
1289 :
1290 : for k in other_keys {
1291 : warn!("Unexpected non timeline key {k}");
1292 : }
1293 :
1294 : Ok(TenantPreload {
1295 : tenant_manifest,
1296 : timelines: self
1297 : .load_timelines_metadata(remote_timeline_ids, remote_storage, cancel)
1298 : .await?,
1299 : })
1300 : }
1301 :
1302 : ///
1303 : /// Background task that downloads all data for a tenant and brings it to Active state.
1304 : ///
1305 : /// No background tasks are started as part of this routine.
1306 : ///
1307 186 : async fn attach(
1308 186 : self: &Arc<Tenant>,
1309 186 : preload: Option<TenantPreload>,
1310 186 : ctx: &RequestContext,
1311 186 : ) -> anyhow::Result<()> {
1312 186 : span::debug_assert_current_span_has_tenant_id();
1313 186 :
1314 186 : failpoint_support::sleep_millis_async!("before-attaching-tenant");
1315 :
1316 186 : let Some(preload) = preload else {
1317 0 : anyhow::bail!("local-only deployment is no longer supported, https://github.com/neondatabase/neon/issues/5624");
1318 : };
1319 :
1320 186 : let mut offloaded_timeline_ids = HashSet::new();
1321 186 : let mut offloaded_timelines_list = Vec::new();
1322 186 : for timeline_manifest in preload.tenant_manifest.offloaded_timelines.iter() {
1323 0 : let timeline_id = timeline_manifest.timeline_id;
1324 0 : let offloaded_timeline =
1325 0 : OffloadedTimeline::from_manifest(self.tenant_shard_id, timeline_manifest);
1326 0 : offloaded_timelines_list.push((timeline_id, Arc::new(offloaded_timeline)));
1327 0 : offloaded_timeline_ids.insert(timeline_id);
1328 0 : }
1329 :
1330 186 : let mut timelines_to_resume_deletions = vec![];
1331 186 :
1332 186 : let mut remote_index_and_client = HashMap::new();
1333 186 : let mut timeline_ancestors = HashMap::new();
1334 186 : let mut existent_timelines = HashSet::new();
1335 192 : for (timeline_id, preload) in preload.timelines {
1336 6 : if offloaded_timeline_ids.remove(&timeline_id) {
1337 : // The timeline is offloaded, skip loading it.
1338 0 : continue;
1339 6 : }
1340 6 : let index_part = match preload.index_part {
1341 6 : Ok(i) => {
1342 6 : debug!("remote index part exists for timeline {timeline_id}");
1343 : // We found index_part on the remote, this is the standard case.
1344 6 : existent_timelines.insert(timeline_id);
1345 6 : i
1346 : }
1347 : Err(DownloadError::NotFound) => {
1348 : // There is no index_part on the remote. We only get here
1349 : // if there is some prefix for the timeline in the remote storage.
1350 : // This can e.g. be the initdb.tar.zst archive, maybe a
1351 : // remnant from a prior incomplete creation or deletion attempt.
1352 : // Delete the local directory as the deciding criterion for a
1353 : // timeline's existence is presence of index_part.
1354 0 : info!(%timeline_id, "index_part not found on remote");
1355 0 : continue;
1356 : }
1357 0 : Err(e) => {
1358 0 : // Some (possibly ephemeral) error happened during index_part download.
1359 0 : // Pretend the timeline exists to not delete the timeline directory,
1360 0 : // as it might be a temporary issue and we don't want to re-download
1361 0 : // everything after it resolves.
1362 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1363 :
1364 0 : existent_timelines.insert(timeline_id);
1365 0 : continue;
1366 : }
1367 : };
1368 6 : match index_part {
1369 6 : MaybeDeletedIndexPart::IndexPart(index_part) => {
1370 6 : timeline_ancestors.insert(timeline_id, index_part.metadata.clone());
1371 6 : remote_index_and_client.insert(timeline_id, (index_part, preload.client));
1372 6 : }
1373 0 : MaybeDeletedIndexPart::Deleted(index_part) => {
1374 0 : info!(
1375 0 : "timeline {} is deleted, picking to resume deletion",
1376 : timeline_id
1377 : );
1378 0 : timelines_to_resume_deletions.push((timeline_id, index_part, preload.client));
1379 : }
1380 : }
1381 : }
1382 :
1383 186 : let mut gc_blocks = HashMap::new();
1384 :
1385 : // For every timeline, download the metadata file, scan the local directory,
1386 : // and build a layer map that contains an entry for each remote and local
1387 : // layer file.
1388 186 : let sorted_timelines = tree_sort_timelines(timeline_ancestors, |m| m.ancestor_timeline())?;
1389 192 : for (timeline_id, remote_metadata) in sorted_timelines {
1390 6 : let (index_part, remote_client) = remote_index_and_client
1391 6 : .remove(&timeline_id)
1392 6 : .expect("just put it in above");
1393 :
1394 6 : if let Some(blocking) = index_part.gc_blocking.as_ref() {
1395 : // could just filter these away, but it helps while testing
1396 0 : anyhow::ensure!(
1397 0 : !blocking.reasons.is_empty(),
1398 0 : "index_part for {timeline_id} is malformed: it should not have gc blocking with zero reasons"
1399 : );
1400 0 : let prev = gc_blocks.insert(timeline_id, blocking.reasons);
1401 0 : assert!(prev.is_none());
1402 6 : }
1403 :
1404 : // TODO again handle early failure
1405 6 : self.load_remote_timeline(
1406 6 : timeline_id,
1407 6 : index_part,
1408 6 : remote_metadata,
1409 6 : TimelineResources {
1410 6 : remote_client,
1411 6 : timeline_get_throttle: self.timeline_get_throttle.clone(),
1412 6 : l0_flush_global_state: self.l0_flush_global_state.clone(),
1413 6 : },
1414 6 : ctx,
1415 6 : )
1416 12 : .await
1417 6 : .with_context(|| {
1418 0 : format!(
1419 0 : "failed to load remote timeline {} for tenant {}",
1420 0 : timeline_id, self.tenant_shard_id
1421 0 : )
1422 6 : })?;
1423 : }
1424 :
1425 : // Walk through deleted timelines, resume deletion
1426 186 : for (timeline_id, index_part, remote_timeline_client) in timelines_to_resume_deletions {
1427 0 : remote_timeline_client
1428 0 : .init_upload_queue_stopped_to_continue_deletion(&index_part)
1429 0 : .context("init queue stopped")
1430 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1431 :
1432 0 : DeleteTimelineFlow::resume_deletion(
1433 0 : Arc::clone(self),
1434 0 : timeline_id,
1435 0 : &index_part.metadata,
1436 0 : remote_timeline_client,
1437 0 : )
1438 0 : .instrument(tracing::info_span!("timeline_delete", %timeline_id))
1439 0 : .await
1440 0 : .context("resume_deletion")
1441 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1442 : }
1443 : // Complete deletions for offloaded timeline id's.
1444 186 : offloaded_timelines_list
1445 186 : .retain(|(offloaded_id, _offloaded)| {
1446 0 : // At this point, offloaded_timeline_ids has the list of all offloaded timelines
1447 0 : // without a prefix in S3, so they are inexistent.
1448 0 : // In the end, existence of a timeline is finally determined by the existence of an index-part.json in remote storage.
1449 0 : // If there is a dangling reference in another location, they need to be cleaned up.
1450 0 : let delete = offloaded_timeline_ids.contains(offloaded_id);
1451 0 : if delete {
1452 0 : tracing::info!("Removing offloaded timeline {offloaded_id} from manifest as no remote prefix was found");
1453 0 : }
1454 0 : !delete
1455 186 : });
1456 186 : if !offloaded_timelines_list.is_empty() {
1457 0 : tracing::info!(
1458 0 : "Tenant has {} offloaded timelines",
1459 0 : offloaded_timelines_list.len()
1460 : );
1461 186 : }
1462 186 : {
1463 186 : let mut offloaded_timelines_accessor = self.timelines_offloaded.lock().unwrap();
1464 186 : offloaded_timelines_accessor.extend(offloaded_timelines_list.into_iter());
1465 186 : }
1466 186 : if !offloaded_timeline_ids.is_empty() {
1467 0 : let manifest = self.tenant_manifest();
1468 0 : // TODO: generation support
1469 0 : let generation = remote_timeline_client::TENANT_MANIFEST_GENERATION;
1470 0 : upload_tenant_manifest(
1471 0 : &self.remote_storage,
1472 0 : &self.tenant_shard_id,
1473 0 : generation,
1474 0 : &manifest,
1475 0 : &self.cancel,
1476 0 : )
1477 0 : .await
1478 0 : .map_err(TimelineArchivalError::Other)?;
1479 186 : }
1480 :
1481 : // The local filesystem contents are a cache of what's in the remote IndexPart;
1482 : // IndexPart is the source of truth.
1483 186 : self.clean_up_timelines(&existent_timelines)?;
1484 :
1485 186 : self.gc_block.set_scanned(gc_blocks);
1486 186 :
1487 186 : fail::fail_point!("attach-before-activate", |_| {
1488 0 : anyhow::bail!("attach-before-activate");
1489 186 : });
1490 186 : failpoint_support::sleep_millis_async!("attach-before-activate-sleep", &self.cancel);
1491 :
1492 186 : info!("Done");
1493 :
1494 186 : Ok(())
1495 186 : }
1496 :
1497 : /// Check for any local timeline directories that are temporary, or do not correspond to a
1498 : /// timeline that still exists: this can happen if we crashed during a deletion/creation, or
1499 : /// if a timeline was deleted while the tenant was attached to a different pageserver.
1500 186 : fn clean_up_timelines(&self, existent_timelines: &HashSet<TimelineId>) -> anyhow::Result<()> {
1501 186 : let timelines_dir = self.conf.timelines_path(&self.tenant_shard_id);
1502 :
1503 186 : let entries = match timelines_dir.read_dir_utf8() {
1504 186 : Ok(d) => d,
1505 0 : Err(e) => {
1506 0 : if e.kind() == std::io::ErrorKind::NotFound {
1507 0 : return Ok(());
1508 : } else {
1509 0 : return Err(e).context("list timelines directory for tenant");
1510 : }
1511 : }
1512 : };
1513 :
1514 194 : for entry in entries {
1515 8 : let entry = entry.context("read timeline dir entry")?;
1516 8 : let entry_path = entry.path();
1517 :
1518 8 : let purge = if crate::is_temporary(entry_path)
1519 : // TODO: remove uninit mark code (https://github.com/neondatabase/neon/issues/5718)
1520 8 : || is_uninit_mark(entry_path)
1521 8 : || crate::is_delete_mark(entry_path)
1522 : {
1523 0 : true
1524 : } else {
1525 8 : match TimelineId::try_from(entry_path.file_name()) {
1526 8 : Ok(i) => {
1527 8 : // Purge if the timeline ID does not exist in remote storage: remote storage is the authority.
1528 8 : !existent_timelines.contains(&i)
1529 : }
1530 0 : Err(e) => {
1531 0 : tracing::warn!(
1532 0 : "Unparseable directory in timelines directory: {entry_path}, ignoring ({e})"
1533 : );
1534 : // Do not purge junk: if we don't recognize it, be cautious and leave it for a human.
1535 0 : false
1536 : }
1537 : }
1538 : };
1539 :
1540 8 : if purge {
1541 2 : tracing::info!("Purging stale timeline dentry {entry_path}");
1542 2 : if let Err(e) = match entry.file_type() {
1543 2 : Ok(t) => if t.is_dir() {
1544 2 : std::fs::remove_dir_all(entry_path)
1545 : } else {
1546 0 : std::fs::remove_file(entry_path)
1547 : }
1548 2 : .or_else(fs_ext::ignore_not_found),
1549 0 : Err(e) => Err(e),
1550 : } {
1551 0 : tracing::warn!("Failed to purge stale timeline dentry {entry_path}: {e}");
1552 2 : }
1553 6 : }
1554 : }
1555 :
1556 186 : Ok(())
1557 186 : }
1558 :
1559 : /// Get sum of all remote timelines sizes
1560 : ///
1561 : /// This function relies on the index_part instead of listing the remote storage
1562 0 : pub fn remote_size(&self) -> u64 {
1563 0 : let mut size = 0;
1564 :
1565 0 : for timeline in self.list_timelines() {
1566 0 : size += timeline.remote_client.get_remote_physical_size();
1567 0 : }
1568 :
1569 0 : size
1570 0 : }
1571 :
1572 6 : #[instrument(skip_all, fields(timeline_id=%timeline_id))]
1573 : async fn load_remote_timeline(
1574 : &self,
1575 : timeline_id: TimelineId,
1576 : index_part: IndexPart,
1577 : remote_metadata: TimelineMetadata,
1578 : resources: TimelineResources,
1579 : ctx: &RequestContext,
1580 : ) -> anyhow::Result<()> {
1581 : span::debug_assert_current_span_has_tenant_id();
1582 :
1583 : info!("downloading index file for timeline {}", timeline_id);
1584 : tokio::fs::create_dir_all(self.conf.timeline_path(&self.tenant_shard_id, &timeline_id))
1585 : .await
1586 : .context("Failed to create new timeline directory")?;
1587 :
1588 : let ancestor = if let Some(ancestor_id) = remote_metadata.ancestor_timeline() {
1589 : let timelines = self.timelines.lock().unwrap();
1590 : Some(Arc::clone(timelines.get(&ancestor_id).ok_or_else(
1591 0 : || {
1592 0 : anyhow::anyhow!(
1593 0 : "cannot find ancestor timeline {ancestor_id} for timeline {timeline_id}"
1594 0 : )
1595 0 : },
1596 : )?))
1597 : } else {
1598 : None
1599 : };
1600 :
1601 : self.timeline_init_and_sync(
1602 : timeline_id,
1603 : resources,
1604 : index_part,
1605 : remote_metadata,
1606 : ancestor,
1607 : ctx,
1608 : )
1609 : .await
1610 : }
1611 :
1612 186 : async fn load_timelines_metadata(
1613 186 : self: &Arc<Tenant>,
1614 186 : timeline_ids: HashSet<TimelineId>,
1615 186 : remote_storage: &GenericRemoteStorage,
1616 186 : cancel: CancellationToken,
1617 186 : ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
1618 186 : let mut part_downloads = JoinSet::new();
1619 192 : for timeline_id in timeline_ids {
1620 6 : let cancel_clone = cancel.clone();
1621 6 : part_downloads.spawn(
1622 6 : self.load_timeline_metadata(timeline_id, remote_storage.clone(), cancel_clone)
1623 6 : .instrument(info_span!("download_index_part", %timeline_id)),
1624 : );
1625 : }
1626 :
1627 186 : let mut timeline_preloads: HashMap<TimelineId, TimelinePreload> = HashMap::new();
1628 :
1629 : loop {
1630 192 : tokio::select!(
1631 192 : next = part_downloads.join_next() => {
1632 192 : match next {
1633 6 : Some(result) => {
1634 6 : let preload = result.context("join preload task")?;
1635 6 : timeline_preloads.insert(preload.timeline_id, preload);
1636 : },
1637 : None => {
1638 186 : break;
1639 : }
1640 : }
1641 : },
1642 192 : _ = cancel.cancelled() => {
1643 0 : anyhow::bail!("Cancelled while waiting for remote index download")
1644 : }
1645 : )
1646 : }
1647 :
1648 186 : Ok(timeline_preloads)
1649 186 : }
1650 :
1651 6 : fn build_timeline_client(
1652 6 : &self,
1653 6 : timeline_id: TimelineId,
1654 6 : remote_storage: GenericRemoteStorage,
1655 6 : ) -> RemoteTimelineClient {
1656 6 : RemoteTimelineClient::new(
1657 6 : remote_storage.clone(),
1658 6 : self.deletion_queue_client.clone(),
1659 6 : self.conf,
1660 6 : self.tenant_shard_id,
1661 6 : timeline_id,
1662 6 : self.generation,
1663 6 : )
1664 6 : }
1665 :
1666 6 : fn load_timeline_metadata(
1667 6 : self: &Arc<Tenant>,
1668 6 : timeline_id: TimelineId,
1669 6 : remote_storage: GenericRemoteStorage,
1670 6 : cancel: CancellationToken,
1671 6 : ) -> impl Future<Output = TimelinePreload> {
1672 6 : let client = self.build_timeline_client(timeline_id, remote_storage);
1673 6 : async move {
1674 6 : debug_assert_current_span_has_tenant_and_timeline_id();
1675 6 : debug!("starting index part download");
1676 :
1677 14 : let index_part = client.download_index_file(&cancel).await;
1678 :
1679 6 : debug!("finished index part download");
1680 :
1681 6 : TimelinePreload {
1682 6 : client,
1683 6 : timeline_id,
1684 6 : index_part,
1685 6 : }
1686 6 : }
1687 6 : }
1688 :
1689 0 : fn check_to_be_archived_has_no_unarchived_children(
1690 0 : timeline_id: TimelineId,
1691 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
1692 0 : ) -> Result<(), TimelineArchivalError> {
1693 0 : let children: Vec<TimelineId> = timelines
1694 0 : .iter()
1695 0 : .filter_map(|(id, entry)| {
1696 0 : if entry.get_ancestor_timeline_id() != Some(timeline_id) {
1697 0 : return None;
1698 0 : }
1699 0 : if entry.is_archived() == Some(true) {
1700 0 : return None;
1701 0 : }
1702 0 : Some(*id)
1703 0 : })
1704 0 : .collect();
1705 0 :
1706 0 : if !children.is_empty() {
1707 0 : return Err(TimelineArchivalError::HasUnarchivedChildren(children));
1708 0 : }
1709 0 : Ok(())
1710 0 : }
1711 :
1712 0 : fn check_ancestor_of_to_be_unarchived_is_not_archived(
1713 0 : ancestor_timeline_id: TimelineId,
1714 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
1715 0 : offloaded_timelines: &std::sync::MutexGuard<
1716 0 : '_,
1717 0 : HashMap<TimelineId, Arc<OffloadedTimeline>>,
1718 0 : >,
1719 0 : ) -> Result<(), TimelineArchivalError> {
1720 0 : let has_archived_parent =
1721 0 : if let Some(ancestor_timeline) = timelines.get(&ancestor_timeline_id) {
1722 0 : ancestor_timeline.is_archived() == Some(true)
1723 0 : } else if offloaded_timelines.contains_key(&ancestor_timeline_id) {
1724 0 : true
1725 : } else {
1726 0 : error!("ancestor timeline {ancestor_timeline_id} not found");
1727 0 : if cfg!(debug_assertions) {
1728 0 : panic!("ancestor timeline {ancestor_timeline_id} not found");
1729 0 : }
1730 0 : return Err(TimelineArchivalError::NotFound);
1731 : };
1732 0 : if has_archived_parent {
1733 0 : return Err(TimelineArchivalError::HasArchivedParent(
1734 0 : ancestor_timeline_id,
1735 0 : ));
1736 0 : }
1737 0 : Ok(())
1738 0 : }
1739 :
1740 0 : fn check_to_be_unarchived_timeline_has_no_archived_parent(
1741 0 : timeline: &Arc<Timeline>,
1742 0 : ) -> Result<(), TimelineArchivalError> {
1743 0 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
1744 0 : if ancestor_timeline.is_archived() == Some(true) {
1745 0 : return Err(TimelineArchivalError::HasArchivedParent(
1746 0 : ancestor_timeline.timeline_id,
1747 0 : ));
1748 0 : }
1749 0 : }
1750 0 : Ok(())
1751 0 : }
1752 :
1753 : /// Loads the specified (offloaded) timeline from S3 and attaches it as a loaded timeline
1754 0 : async fn unoffload_timeline(
1755 0 : self: &Arc<Self>,
1756 0 : timeline_id: TimelineId,
1757 0 : broker_client: storage_broker::BrokerClientChannel,
1758 0 : ctx: RequestContext,
1759 0 : ) -> Result<Arc<Timeline>, TimelineArchivalError> {
1760 0 : info!("unoffloading timeline");
1761 0 : let cancel = self.cancel.clone();
1762 0 : let timeline_preload = self
1763 0 : .load_timeline_metadata(timeline_id, self.remote_storage.clone(), cancel.clone())
1764 0 : .await;
1765 :
1766 0 : let index_part = match timeline_preload.index_part {
1767 0 : Ok(index_part) => {
1768 0 : debug!("remote index part exists for timeline {timeline_id}");
1769 0 : index_part
1770 : }
1771 : Err(DownloadError::NotFound) => {
1772 0 : error!(%timeline_id, "index_part not found on remote");
1773 0 : return Err(TimelineArchivalError::NotFound);
1774 : }
1775 0 : Err(DownloadError::Cancelled) => return Err(TimelineArchivalError::Cancelled),
1776 0 : Err(e) => {
1777 0 : // Some (possibly ephemeral) error happened during index_part download.
1778 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1779 0 : return Err(TimelineArchivalError::Other(
1780 0 : anyhow::Error::new(e).context("downloading index_part from remote storage"),
1781 0 : ));
1782 : }
1783 : };
1784 0 : let index_part = match index_part {
1785 0 : MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
1786 0 : MaybeDeletedIndexPart::Deleted(_index_part) => {
1787 0 : info!("timeline is deleted according to index_part.json");
1788 0 : return Err(TimelineArchivalError::NotFound);
1789 : }
1790 : };
1791 0 : let remote_metadata = index_part.metadata.clone();
1792 0 : let timeline_resources = self.build_timeline_resources(timeline_id);
1793 0 : self.load_remote_timeline(
1794 0 : timeline_id,
1795 0 : index_part,
1796 0 : remote_metadata,
1797 0 : timeline_resources,
1798 0 : &ctx,
1799 0 : )
1800 0 : .await
1801 0 : .with_context(|| {
1802 0 : format!(
1803 0 : "failed to load remote timeline {} for tenant {}",
1804 0 : timeline_id, self.tenant_shard_id
1805 0 : )
1806 0 : })
1807 0 : .map_err(TimelineArchivalError::Other)?;
1808 :
1809 0 : let timeline = {
1810 0 : let timelines = self.timelines.lock().unwrap();
1811 0 : let Some(timeline) = timelines.get(&timeline_id) else {
1812 0 : warn!("timeline not available directly after attach");
1813 : // This is not a panic because no locks are held between `load_remote_timeline`
1814 : // which puts the timeline into timelines, and our look into the timeline map.
1815 0 : return Err(TimelineArchivalError::Other(anyhow::anyhow!(
1816 0 : "timeline not available directly after attach"
1817 0 : )));
1818 : };
1819 0 : let mut offloaded_timelines = self.timelines_offloaded.lock().unwrap();
1820 0 : if offloaded_timelines.remove(&timeline_id).is_none() {
1821 0 : warn!("timeline already removed from offloaded timelines");
1822 0 : }
1823 0 : Arc::clone(timeline)
1824 0 : };
1825 0 :
1826 0 : // Upload new list of offloaded timelines to S3
1827 0 : let manifest = self.tenant_manifest();
1828 0 : // TODO: generation support
1829 0 : let generation = remote_timeline_client::TENANT_MANIFEST_GENERATION;
1830 0 : upload_tenant_manifest(
1831 0 : &self.remote_storage,
1832 0 : &self.tenant_shard_id,
1833 0 : generation,
1834 0 : &manifest,
1835 0 : &cancel,
1836 0 : )
1837 0 : .await
1838 0 : .map_err(TimelineArchivalError::Other)?;
1839 :
1840 : // Activate the timeline (if it makes sense)
1841 0 : if !(timeline.is_broken() || timeline.is_stopping()) {
1842 0 : let background_jobs_can_start = None;
1843 0 : timeline.activate(
1844 0 : self.clone(),
1845 0 : broker_client.clone(),
1846 0 : background_jobs_can_start,
1847 0 : &ctx,
1848 0 : );
1849 0 : }
1850 :
1851 0 : info!("timeline unoffloading complete");
1852 0 : Ok(timeline)
1853 0 : }
1854 :
1855 0 : pub(crate) async fn apply_timeline_archival_config(
1856 0 : self: &Arc<Self>,
1857 0 : timeline_id: TimelineId,
1858 0 : new_state: TimelineArchivalState,
1859 0 : broker_client: storage_broker::BrokerClientChannel,
1860 0 : ctx: RequestContext,
1861 0 : ) -> Result<(), TimelineArchivalError> {
1862 0 : info!("setting timeline archival config");
1863 : // First part: figure out what is needed to do, and do validation
1864 0 : let timeline_or_unarchive_offloaded = 'outer: {
1865 0 : let timelines = self.timelines.lock().unwrap();
1866 :
1867 0 : let Some(timeline) = timelines.get(&timeline_id) else {
1868 0 : let offloaded_timelines = self.timelines_offloaded.lock().unwrap();
1869 0 : let Some(offloaded) = offloaded_timelines.get(&timeline_id) else {
1870 0 : return Err(TimelineArchivalError::NotFound);
1871 : };
1872 0 : if new_state == TimelineArchivalState::Archived {
1873 : // It's offloaded already, so nothing to do
1874 0 : return Ok(());
1875 0 : }
1876 0 : if let Some(ancestor_timeline_id) = offloaded.ancestor_timeline_id {
1877 0 : Self::check_ancestor_of_to_be_unarchived_is_not_archived(
1878 0 : ancestor_timeline_id,
1879 0 : &timelines,
1880 0 : &offloaded_timelines,
1881 0 : )?;
1882 0 : }
1883 0 : break 'outer None;
1884 : };
1885 :
1886 : // Do some validation. We release the timelines lock below, so there is potential
1887 : // for race conditions: these checks are more present to prevent misunderstandings of
1888 : // the API's capabilities, instead of serving as the sole way to defend their invariants.
1889 0 : match new_state {
1890 : TimelineArchivalState::Unarchived => {
1891 0 : Self::check_to_be_unarchived_timeline_has_no_archived_parent(timeline)?
1892 : }
1893 : TimelineArchivalState::Archived => {
1894 0 : Self::check_to_be_archived_has_no_unarchived_children(timeline_id, &timelines)?
1895 : }
1896 : }
1897 0 : Some(Arc::clone(timeline))
1898 : };
1899 :
1900 : // Second part: unoffload timeline (if needed)
1901 0 : let timeline = if let Some(timeline) = timeline_or_unarchive_offloaded {
1902 0 : timeline
1903 : } else {
1904 : // Turn offloaded timeline into a non-offloaded one
1905 0 : self.unoffload_timeline(timeline_id, broker_client, ctx)
1906 0 : .await?
1907 : };
1908 :
1909 : // Third part: upload new timeline archival state and block until it is present in S3
1910 0 : let upload_needed = match timeline
1911 0 : .remote_client
1912 0 : .schedule_index_upload_for_timeline_archival_state(new_state)
1913 : {
1914 0 : Ok(upload_needed) => upload_needed,
1915 0 : Err(e) => {
1916 0 : if timeline.cancel.is_cancelled() {
1917 0 : return Err(TimelineArchivalError::Cancelled);
1918 : } else {
1919 0 : return Err(TimelineArchivalError::Other(e));
1920 : }
1921 : }
1922 : };
1923 :
1924 0 : if upload_needed {
1925 0 : info!("Uploading new state");
1926 : const MAX_WAIT: Duration = Duration::from_secs(10);
1927 0 : let Ok(v) =
1928 0 : tokio::time::timeout(MAX_WAIT, timeline.remote_client.wait_completion()).await
1929 : else {
1930 0 : tracing::warn!("reached timeout for waiting on upload queue");
1931 0 : return Err(TimelineArchivalError::Timeout);
1932 : };
1933 0 : v.map_err(|e| match e {
1934 0 : WaitCompletionError::NotInitialized(e) => {
1935 0 : TimelineArchivalError::Other(anyhow::anyhow!(e))
1936 : }
1937 : WaitCompletionError::UploadQueueShutDownOrStopped => {
1938 0 : TimelineArchivalError::Cancelled
1939 : }
1940 0 : })?;
1941 0 : }
1942 0 : Ok(())
1943 0 : }
1944 :
1945 0 : pub fn get_offloaded_timeline(
1946 0 : &self,
1947 0 : timeline_id: TimelineId,
1948 0 : ) -> Result<Arc<OffloadedTimeline>, GetTimelineError> {
1949 0 : self.timelines_offloaded
1950 0 : .lock()
1951 0 : .unwrap()
1952 0 : .get(&timeline_id)
1953 0 : .map(Arc::clone)
1954 0 : .ok_or(GetTimelineError::NotFound {
1955 0 : tenant_id: self.tenant_shard_id,
1956 0 : timeline_id,
1957 0 : })
1958 0 : }
1959 :
1960 4 : pub(crate) fn tenant_shard_id(&self) -> TenantShardId {
1961 4 : self.tenant_shard_id
1962 4 : }
1963 :
1964 : /// Get Timeline handle for given Neon timeline ID.
1965 : /// This function is idempotent. It doesn't change internal state in any way.
1966 222 : pub fn get_timeline(
1967 222 : &self,
1968 222 : timeline_id: TimelineId,
1969 222 : active_only: bool,
1970 222 : ) -> Result<Arc<Timeline>, GetTimelineError> {
1971 222 : let timelines_accessor = self.timelines.lock().unwrap();
1972 222 : let timeline = timelines_accessor
1973 222 : .get(&timeline_id)
1974 222 : .ok_or(GetTimelineError::NotFound {
1975 222 : tenant_id: self.tenant_shard_id,
1976 222 : timeline_id,
1977 222 : })?;
1978 :
1979 220 : if active_only && !timeline.is_active() {
1980 0 : Err(GetTimelineError::NotActive {
1981 0 : tenant_id: self.tenant_shard_id,
1982 0 : timeline_id,
1983 0 : state: timeline.current_state(),
1984 0 : })
1985 : } else {
1986 220 : Ok(Arc::clone(timeline))
1987 : }
1988 222 : }
1989 :
1990 : /// Lists timelines the tenant contains.
1991 : /// It's up to callers to omit certain timelines that are not considered ready for use.
1992 0 : pub fn list_timelines(&self) -> Vec<Arc<Timeline>> {
1993 0 : self.timelines
1994 0 : .lock()
1995 0 : .unwrap()
1996 0 : .values()
1997 0 : .map(Arc::clone)
1998 0 : .collect()
1999 0 : }
2000 :
2001 : /// Lists timelines the tenant manages, including offloaded ones.
2002 : ///
2003 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2004 0 : pub fn list_timelines_and_offloaded(
2005 0 : &self,
2006 0 : ) -> (Vec<Arc<Timeline>>, Vec<Arc<OffloadedTimeline>>) {
2007 0 : let timelines = self
2008 0 : .timelines
2009 0 : .lock()
2010 0 : .unwrap()
2011 0 : .values()
2012 0 : .map(Arc::clone)
2013 0 : .collect();
2014 0 : let offloaded = self
2015 0 : .timelines_offloaded
2016 0 : .lock()
2017 0 : .unwrap()
2018 0 : .values()
2019 0 : .map(Arc::clone)
2020 0 : .collect();
2021 0 : (timelines, offloaded)
2022 0 : }
2023 :
2024 0 : pub fn list_timeline_ids(&self) -> Vec<TimelineId> {
2025 0 : self.timelines.lock().unwrap().keys().cloned().collect()
2026 0 : }
2027 :
2028 : /// This is used to create the initial 'main' timeline during bootstrapping,
2029 : /// or when importing a new base backup. The caller is expected to load an
2030 : /// initial image of the datadir to the new timeline after this.
2031 : ///
2032 : /// Until that happens, the on-disk state is invalid (disk_consistent_lsn=Lsn(0))
2033 : /// and the timeline will fail to load at a restart.
2034 : ///
2035 : /// For tests, use `DatadirModification::init_empty_test_timeline` + `commit` to setup the
2036 : /// minimum amount of keys required to get a writable timeline.
2037 : /// (Without it, `put` might fail due to `repartition` failing.)
2038 178 : pub(crate) async fn create_empty_timeline(
2039 178 : &self,
2040 178 : new_timeline_id: TimelineId,
2041 178 : initdb_lsn: Lsn,
2042 178 : pg_version: u32,
2043 178 : _ctx: &RequestContext,
2044 178 : ) -> anyhow::Result<UninitializedTimeline> {
2045 178 : anyhow::ensure!(
2046 178 : self.is_active(),
2047 0 : "Cannot create empty timelines on inactive tenant"
2048 : );
2049 :
2050 : // Protect against concurrent attempts to use this TimelineId
2051 178 : let create_guard = self.create_timeline_create_guard(new_timeline_id)?;
2052 :
2053 176 : let new_metadata = TimelineMetadata::new(
2054 176 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2055 176 : // make it valid, before calling finish_creation()
2056 176 : Lsn(0),
2057 176 : None,
2058 176 : None,
2059 176 : Lsn(0),
2060 176 : initdb_lsn,
2061 176 : initdb_lsn,
2062 176 : pg_version,
2063 176 : );
2064 176 : self.prepare_new_timeline(
2065 176 : new_timeline_id,
2066 176 : &new_metadata,
2067 176 : create_guard,
2068 176 : initdb_lsn,
2069 176 : None,
2070 176 : )
2071 0 : .await
2072 178 : }
2073 :
2074 : /// Helper for unit tests to create an empty timeline.
2075 : ///
2076 : /// The timeline is has state value `Active` but its background loops are not running.
2077 : // This makes the various functions which anyhow::ensure! for Active state work in tests.
2078 : // Our current tests don't need the background loops.
2079 : #[cfg(test)]
2080 168 : pub async fn create_test_timeline(
2081 168 : &self,
2082 168 : new_timeline_id: TimelineId,
2083 168 : initdb_lsn: Lsn,
2084 168 : pg_version: u32,
2085 168 : ctx: &RequestContext,
2086 168 : ) -> anyhow::Result<Arc<Timeline>> {
2087 168 : let uninit_tl = self
2088 168 : .create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2089 0 : .await?;
2090 168 : let tline = uninit_tl.raw_timeline().expect("we just created it");
2091 168 : assert_eq!(tline.get_last_record_lsn(), Lsn(0));
2092 :
2093 : // Setup minimum keys required for the timeline to be usable.
2094 168 : let mut modification = tline.begin_modification(initdb_lsn);
2095 168 : modification
2096 168 : .init_empty_test_timeline()
2097 168 : .context("init_empty_test_timeline")?;
2098 168 : modification
2099 168 : .commit(ctx)
2100 162 : .await
2101 168 : .context("commit init_empty_test_timeline modification")?;
2102 :
2103 : // Flush to disk so that uninit_tl's check for valid disk_consistent_lsn passes.
2104 168 : tline.maybe_spawn_flush_loop();
2105 168 : tline.freeze_and_flush().await.context("freeze_and_flush")?;
2106 :
2107 : // Make sure the freeze_and_flush reaches remote storage.
2108 168 : tline.remote_client.wait_completion().await.unwrap();
2109 :
2110 168 : let tl = uninit_tl.finish_creation()?;
2111 : // The non-test code would call tl.activate() here.
2112 168 : tl.set_state(TimelineState::Active);
2113 168 : Ok(tl)
2114 168 : }
2115 :
2116 : /// Helper for unit tests to create a timeline with some pre-loaded states.
2117 : #[cfg(test)]
2118 : #[allow(clippy::too_many_arguments)]
2119 28 : pub async fn create_test_timeline_with_layers(
2120 28 : &self,
2121 28 : new_timeline_id: TimelineId,
2122 28 : initdb_lsn: Lsn,
2123 28 : pg_version: u32,
2124 28 : ctx: &RequestContext,
2125 28 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
2126 28 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
2127 28 : end_lsn: Lsn,
2128 28 : ) -> anyhow::Result<Arc<Timeline>> {
2129 : use checks::check_valid_layermap;
2130 : use itertools::Itertools;
2131 :
2132 28 : let tline = self
2133 28 : .create_test_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2134 56 : .await?;
2135 28 : tline.force_advance_lsn(end_lsn);
2136 82 : for deltas in delta_layer_desc {
2137 54 : tline
2138 54 : .force_create_delta_layer(deltas, Some(initdb_lsn), ctx)
2139 162 : .await?;
2140 : }
2141 74 : for (lsn, images) in image_layer_desc {
2142 46 : tline
2143 46 : .force_create_image_layer(lsn, images, Some(initdb_lsn), ctx)
2144 269 : .await?;
2145 : }
2146 28 : let layer_names = tline
2147 28 : .layers
2148 28 : .read()
2149 0 : .await
2150 28 : .layer_map()
2151 28 : .unwrap()
2152 28 : .iter_historic_layers()
2153 128 : .map(|layer| layer.layer_name())
2154 28 : .collect_vec();
2155 28 : if let Some(err) = check_valid_layermap(&layer_names) {
2156 0 : bail!("invalid layermap: {err}");
2157 28 : }
2158 28 : Ok(tline)
2159 28 : }
2160 :
2161 : /// Create a new timeline.
2162 : ///
2163 : /// Returns the new timeline ID and reference to its Timeline object.
2164 : ///
2165 : /// If the caller specified the timeline ID to use (`new_timeline_id`), and timeline with
2166 : /// the same timeline ID already exists, returns CreateTimelineError::AlreadyExists.
2167 : #[allow(clippy::too_many_arguments)]
2168 0 : pub(crate) async fn create_timeline(
2169 0 : self: &Arc<Tenant>,
2170 0 : params: CreateTimelineParams,
2171 0 : broker_client: storage_broker::BrokerClientChannel,
2172 0 : ctx: &RequestContext,
2173 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
2174 0 : if !self.is_active() {
2175 0 : if matches!(self.current_state(), TenantState::Stopping { .. }) {
2176 0 : return Err(CreateTimelineError::ShuttingDown);
2177 : } else {
2178 0 : return Err(CreateTimelineError::Other(anyhow::anyhow!(
2179 0 : "Cannot create timelines on inactive tenant"
2180 0 : )));
2181 : }
2182 0 : }
2183 :
2184 0 : let _gate = self
2185 0 : .gate
2186 0 : .enter()
2187 0 : .map_err(|_| CreateTimelineError::ShuttingDown)?;
2188 :
2189 0 : let result: CreateTimelineResult = match params {
2190 : CreateTimelineParams::Bootstrap(CreateTimelineParamsBootstrap {
2191 0 : new_timeline_id,
2192 0 : existing_initdb_timeline_id,
2193 0 : pg_version,
2194 0 : }) => {
2195 0 : self.bootstrap_timeline(
2196 0 : new_timeline_id,
2197 0 : pg_version,
2198 0 : existing_initdb_timeline_id,
2199 0 : ctx,
2200 0 : )
2201 0 : .await?
2202 : }
2203 : CreateTimelineParams::Branch(CreateTimelineParamsBranch {
2204 0 : new_timeline_id,
2205 0 : ancestor_timeline_id,
2206 0 : mut ancestor_start_lsn,
2207 : }) => {
2208 0 : let ancestor_timeline = self
2209 0 : .get_timeline(ancestor_timeline_id, false)
2210 0 : .context("Cannot branch off the timeline that's not present in pageserver")?;
2211 :
2212 : // instead of waiting around, just deny the request because ancestor is not yet
2213 : // ready for other purposes either.
2214 0 : if !ancestor_timeline.is_active() {
2215 0 : return Err(CreateTimelineError::AncestorNotActive);
2216 0 : }
2217 0 :
2218 0 : if ancestor_timeline.is_archived() == Some(true) {
2219 0 : info!("tried to branch archived timeline");
2220 0 : return Err(CreateTimelineError::AncestorArchived);
2221 0 : }
2222 :
2223 0 : if let Some(lsn) = ancestor_start_lsn.as_mut() {
2224 0 : *lsn = lsn.align();
2225 0 :
2226 0 : let ancestor_ancestor_lsn = ancestor_timeline.get_ancestor_lsn();
2227 0 : if ancestor_ancestor_lsn > *lsn {
2228 : // can we safely just branch from the ancestor instead?
2229 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
2230 0 : "invalid start lsn {} for ancestor timeline {}: less than timeline ancestor lsn {}",
2231 0 : lsn,
2232 0 : ancestor_timeline_id,
2233 0 : ancestor_ancestor_lsn,
2234 0 : )));
2235 0 : }
2236 0 :
2237 0 : // Wait for the WAL to arrive and be processed on the parent branch up
2238 0 : // to the requested branch point. The repository code itself doesn't
2239 0 : // require it, but if we start to receive WAL on the new timeline,
2240 0 : // decoding the new WAL might need to look up previous pages, relation
2241 0 : // sizes etc. and that would get confused if the previous page versions
2242 0 : // are not in the repository yet.
2243 0 : ancestor_timeline
2244 0 : .wait_lsn(*lsn, timeline::WaitLsnWaiter::Tenant, ctx)
2245 0 : .await
2246 0 : .map_err(|e| match e {
2247 0 : e @ (WaitLsnError::Timeout(_) | WaitLsnError::BadState { .. }) => {
2248 0 : CreateTimelineError::AncestorLsn(anyhow::anyhow!(e))
2249 : }
2250 0 : WaitLsnError::Shutdown => CreateTimelineError::ShuttingDown,
2251 0 : })?;
2252 0 : }
2253 :
2254 0 : self.branch_timeline(&ancestor_timeline, new_timeline_id, ancestor_start_lsn, ctx)
2255 0 : .await?
2256 : }
2257 : };
2258 :
2259 : // At this point we have dropped our guard on [`Self::timelines_creating`], and
2260 : // the timeline is visible in [`Self::timelines`], but it is _not_ durable yet. We must
2261 : // not send a success to the caller until it is. The same applies to handling retries,
2262 : // that is done in [`Self::start_creating_timeline`].
2263 0 : result
2264 0 : .timeline()
2265 0 : .remote_client
2266 0 : .wait_completion()
2267 0 : .await
2268 0 : .context("wait for timeline initial uploads to complete")?;
2269 :
2270 : // The creating task is responsible for activating the timeline.
2271 : // We do this after `wait_completion()` so that we don't spin up tasks that start
2272 : // doing stuff before the IndexPart is durable in S3, which is done by the previous section.
2273 0 : let activated_timeline = match result {
2274 0 : CreateTimelineResult::Created(timeline) => {
2275 0 : timeline.activate(self.clone(), broker_client, None, ctx);
2276 0 : timeline
2277 : }
2278 0 : CreateTimelineResult::Idempotent(timeline) => {
2279 0 : info!(
2280 0 : "request was deemed idempotent, activation will be done by the creating task"
2281 : );
2282 0 : timeline
2283 : }
2284 : };
2285 :
2286 0 : Ok(activated_timeline)
2287 0 : }
2288 :
2289 0 : pub(crate) async fn delete_timeline(
2290 0 : self: Arc<Self>,
2291 0 : timeline_id: TimelineId,
2292 0 : ) -> Result<(), DeleteTimelineError> {
2293 0 : DeleteTimelineFlow::run(&self, timeline_id).await?;
2294 :
2295 0 : Ok(())
2296 0 : }
2297 :
2298 : /// perform one garbage collection iteration, removing old data files from disk.
2299 : /// this function is periodically called by gc task.
2300 : /// also it can be explicitly requested through page server api 'do_gc' command.
2301 : ///
2302 : /// `target_timeline_id` specifies the timeline to GC, or None for all.
2303 : ///
2304 : /// The `horizon` an `pitr` parameters determine how much WAL history needs to be retained.
2305 : /// Also known as the retention period, or the GC cutoff point. `horizon` specifies
2306 : /// the amount of history, as LSN difference from current latest LSN on each timeline.
2307 : /// `pitr` specifies the same as a time difference from the current time. The effective
2308 : /// GC cutoff point is determined conservatively by either `horizon` and `pitr`, whichever
2309 : /// requires more history to be retained.
2310 : //
2311 754 : pub(crate) async fn gc_iteration(
2312 754 : &self,
2313 754 : target_timeline_id: Option<TimelineId>,
2314 754 : horizon: u64,
2315 754 : pitr: Duration,
2316 754 : cancel: &CancellationToken,
2317 754 : ctx: &RequestContext,
2318 754 : ) -> Result<GcResult, GcError> {
2319 754 : // Don't start doing work during shutdown
2320 754 : if let TenantState::Stopping { .. } = self.current_state() {
2321 0 : return Ok(GcResult::default());
2322 754 : }
2323 754 :
2324 754 : // there is a global allowed_error for this
2325 754 : if !self.is_active() {
2326 0 : return Err(GcError::NotActive);
2327 754 : }
2328 754 :
2329 754 : {
2330 754 : let conf = self.tenant_conf.load();
2331 754 :
2332 754 : if !conf.location.may_delete_layers_hint() {
2333 0 : info!("Skipping GC in location state {:?}", conf.location);
2334 0 : return Ok(GcResult::default());
2335 754 : }
2336 754 :
2337 754 : if conf.is_gc_blocked_by_lsn_lease_deadline() {
2338 750 : info!("Skipping GC because lsn lease deadline is not reached");
2339 750 : return Ok(GcResult::default());
2340 4 : }
2341 : }
2342 :
2343 4 : let _guard = match self.gc_block.start().await {
2344 4 : Ok(guard) => guard,
2345 0 : Err(reasons) => {
2346 0 : info!("Skipping GC: {reasons}");
2347 0 : return Ok(GcResult::default());
2348 : }
2349 : };
2350 :
2351 4 : self.gc_iteration_internal(target_timeline_id, horizon, pitr, cancel, ctx)
2352 3 : .await
2353 754 : }
2354 :
2355 : /// Perform one compaction iteration.
2356 : /// This function is periodically called by compactor task.
2357 : /// Also it can be explicitly requested per timeline through page server
2358 : /// api's 'compact' command.
2359 : ///
2360 : /// Returns whether we have pending compaction task.
2361 0 : async fn compaction_iteration(
2362 0 : self: &Arc<Self>,
2363 0 : cancel: &CancellationToken,
2364 0 : ctx: &RequestContext,
2365 0 : ) -> Result<bool, timeline::CompactionError> {
2366 0 : // Don't start doing work during shutdown, or when broken, we do not need those in the logs
2367 0 : if !self.is_active() {
2368 0 : return Ok(false);
2369 0 : }
2370 0 :
2371 0 : {
2372 0 : let conf = self.tenant_conf.load();
2373 0 : if !conf.location.may_delete_layers_hint() || !conf.location.may_upload_layers_hint() {
2374 0 : info!("Skipping compaction in location state {:?}", conf.location);
2375 0 : return Ok(false);
2376 0 : }
2377 0 : }
2378 0 :
2379 0 : // Scan through the hashmap and collect a list of all the timelines,
2380 0 : // while holding the lock. Then drop the lock and actually perform the
2381 0 : // compactions. We don't want to block everything else while the
2382 0 : // compaction runs.
2383 0 : let timelines_to_compact_or_offload;
2384 0 : {
2385 0 : let timelines = self.timelines.lock().unwrap();
2386 0 : timelines_to_compact_or_offload = timelines
2387 0 : .iter()
2388 0 : .filter_map(|(timeline_id, timeline)| {
2389 0 : let (is_active, can_offload) = (timeline.is_active(), timeline.can_offload());
2390 0 : let has_no_unoffloaded_children = {
2391 0 : !timelines
2392 0 : .iter()
2393 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(*timeline_id))
2394 : };
2395 0 : let can_offload =
2396 0 : can_offload && has_no_unoffloaded_children && self.conf.timeline_offloading;
2397 0 : if (is_active, can_offload) == (false, false) {
2398 0 : None
2399 : } else {
2400 0 : Some((*timeline_id, timeline.clone(), (is_active, can_offload)))
2401 : }
2402 0 : })
2403 0 : .collect::<Vec<_>>();
2404 0 : drop(timelines);
2405 0 : }
2406 0 :
2407 0 : // Before doing any I/O work, check our circuit breaker
2408 0 : if self.compaction_circuit_breaker.lock().unwrap().is_broken() {
2409 0 : info!("Skipping compaction due to previous failures");
2410 0 : return Ok(false);
2411 0 : }
2412 0 :
2413 0 : let mut has_pending_task = false;
2414 :
2415 0 : for (timeline_id, timeline, (can_compact, can_offload)) in &timelines_to_compact_or_offload
2416 : {
2417 0 : let pending_task_left = if *can_compact {
2418 : Some(
2419 0 : timeline
2420 0 : .compact(cancel, EnumSet::empty(), ctx)
2421 0 : .instrument(info_span!("compact_timeline", %timeline_id))
2422 0 : .await
2423 0 : .inspect_err(|e| match e {
2424 0 : timeline::CompactionError::ShuttingDown => (),
2425 0 : timeline::CompactionError::Other(e) => {
2426 0 : self.compaction_circuit_breaker
2427 0 : .lock()
2428 0 : .unwrap()
2429 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, e);
2430 0 : }
2431 0 : })?,
2432 : )
2433 : } else {
2434 0 : None
2435 : };
2436 0 : has_pending_task |= pending_task_left.unwrap_or(false);
2437 0 : if pending_task_left == Some(false) && *can_offload {
2438 0 : offload_timeline(self, timeline)
2439 0 : .instrument(info_span!("offload_timeline", %timeline_id))
2440 0 : .await
2441 0 : .map_err(timeline::CompactionError::Other)?;
2442 0 : }
2443 : }
2444 :
2445 0 : self.compaction_circuit_breaker
2446 0 : .lock()
2447 0 : .unwrap()
2448 0 : .success(&CIRCUIT_BREAKERS_UNBROKEN);
2449 0 :
2450 0 : Ok(has_pending_task)
2451 0 : }
2452 :
2453 : // Call through to all timelines to freeze ephemeral layers if needed. Usually
2454 : // this happens during ingest: this background housekeeping is for freezing layers
2455 : // that are open but haven't been written to for some time.
2456 0 : async fn ingest_housekeeping(&self) {
2457 0 : // Scan through the hashmap and collect a list of all the timelines,
2458 0 : // while holding the lock. Then drop the lock and actually perform the
2459 0 : // compactions. We don't want to block everything else while the
2460 0 : // compaction runs.
2461 0 : let timelines = {
2462 0 : self.timelines
2463 0 : .lock()
2464 0 : .unwrap()
2465 0 : .values()
2466 0 : .filter_map(|timeline| {
2467 0 : if timeline.is_active() {
2468 0 : Some(timeline.clone())
2469 : } else {
2470 0 : None
2471 : }
2472 0 : })
2473 0 : .collect::<Vec<_>>()
2474 : };
2475 :
2476 0 : for timeline in &timelines {
2477 0 : timeline.maybe_freeze_ephemeral_layer().await;
2478 : }
2479 0 : }
2480 :
2481 0 : pub fn timeline_has_no_attached_children(&self, timeline_id: TimelineId) -> bool {
2482 0 : let timelines = self.timelines.lock().unwrap();
2483 0 : !timelines
2484 0 : .iter()
2485 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(timeline_id))
2486 0 : }
2487 :
2488 1696 : pub fn current_state(&self) -> TenantState {
2489 1696 : self.state.borrow().clone()
2490 1696 : }
2491 :
2492 936 : pub fn is_active(&self) -> bool {
2493 936 : self.current_state() == TenantState::Active
2494 936 : }
2495 :
2496 0 : pub fn generation(&self) -> Generation {
2497 0 : self.generation
2498 0 : }
2499 :
2500 0 : pub(crate) fn wal_redo_manager_status(&self) -> Option<WalRedoManagerStatus> {
2501 0 : self.walredo_mgr.as_ref().and_then(|mgr| mgr.status())
2502 0 : }
2503 :
2504 : /// Changes tenant status to active, unless shutdown was already requested.
2505 : ///
2506 : /// `background_jobs_can_start` is an optional barrier set to a value during pageserver startup
2507 : /// to delay background jobs. Background jobs can be started right away when None is given.
2508 0 : fn activate(
2509 0 : self: &Arc<Self>,
2510 0 : broker_client: BrokerClientChannel,
2511 0 : background_jobs_can_start: Option<&completion::Barrier>,
2512 0 : ctx: &RequestContext,
2513 0 : ) {
2514 0 : span::debug_assert_current_span_has_tenant_id();
2515 0 :
2516 0 : let mut activating = false;
2517 0 : self.state.send_modify(|current_state| {
2518 : use pageserver_api::models::ActivatingFrom;
2519 0 : match &*current_state {
2520 : TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
2521 0 : panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {state:?}", state = current_state);
2522 : }
2523 0 : TenantState::Attaching => {
2524 0 : *current_state = TenantState::Activating(ActivatingFrom::Attaching);
2525 0 : }
2526 0 : }
2527 0 : debug!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), "Activating tenant");
2528 0 : activating = true;
2529 0 : // Continue outside the closure. We need to grab timelines.lock()
2530 0 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
2531 0 : });
2532 0 :
2533 0 : if activating {
2534 0 : let timelines_accessor = self.timelines.lock().unwrap();
2535 0 : let timelines_offloaded_accessor = self.timelines_offloaded.lock().unwrap();
2536 0 : let timelines_to_activate = timelines_accessor
2537 0 : .values()
2538 0 : .filter(|timeline| !(timeline.is_broken() || timeline.is_stopping()));
2539 0 :
2540 0 : // Before activation, populate each Timeline's GcInfo with information about its children
2541 0 : self.initialize_gc_info(&timelines_accessor, &timelines_offloaded_accessor);
2542 0 :
2543 0 : // Spawn gc and compaction loops. The loops will shut themselves
2544 0 : // down when they notice that the tenant is inactive.
2545 0 : tasks::start_background_loops(self, background_jobs_can_start);
2546 0 :
2547 0 : let mut activated_timelines = 0;
2548 :
2549 0 : for timeline in timelines_to_activate {
2550 0 : timeline.activate(
2551 0 : self.clone(),
2552 0 : broker_client.clone(),
2553 0 : background_jobs_can_start,
2554 0 : ctx,
2555 0 : );
2556 0 : activated_timelines += 1;
2557 0 : }
2558 :
2559 0 : self.state.send_modify(move |current_state| {
2560 0 : assert!(
2561 0 : matches!(current_state, TenantState::Activating(_)),
2562 0 : "set_stopping and set_broken wait for us to leave Activating state",
2563 : );
2564 0 : *current_state = TenantState::Active;
2565 0 :
2566 0 : let elapsed = self.constructed_at.elapsed();
2567 0 : let total_timelines = timelines_accessor.len();
2568 0 :
2569 0 : // log a lot of stuff, because some tenants sometimes suffer from user-visible
2570 0 : // times to activate. see https://github.com/neondatabase/neon/issues/4025
2571 0 : info!(
2572 0 : since_creation_millis = elapsed.as_millis(),
2573 0 : tenant_id = %self.tenant_shard_id.tenant_id,
2574 0 : shard_id = %self.tenant_shard_id.shard_slug(),
2575 0 : activated_timelines,
2576 0 : total_timelines,
2577 0 : post_state = <&'static str>::from(&*current_state),
2578 0 : "activation attempt finished"
2579 : );
2580 :
2581 0 : TENANT.activation.observe(elapsed.as_secs_f64());
2582 0 : });
2583 0 : }
2584 0 : }
2585 :
2586 : /// Shutdown the tenant and join all of the spawned tasks.
2587 : ///
2588 : /// The method caters for all use-cases:
2589 : /// - pageserver shutdown (freeze_and_flush == true)
2590 : /// - detach + ignore (freeze_and_flush == false)
2591 : ///
2592 : /// This will attempt to shutdown even if tenant is broken.
2593 : ///
2594 : /// `shutdown_progress` is a [`completion::Barrier`] for the shutdown initiated by this call.
2595 : /// If the tenant is already shutting down, we return a clone of the first shutdown call's
2596 : /// `Barrier` as an `Err`. This not-first caller can use the returned barrier to join with
2597 : /// the ongoing shutdown.
2598 6 : async fn shutdown(
2599 6 : &self,
2600 6 : shutdown_progress: completion::Barrier,
2601 6 : shutdown_mode: timeline::ShutdownMode,
2602 6 : ) -> Result<(), completion::Barrier> {
2603 6 : span::debug_assert_current_span_has_tenant_id();
2604 :
2605 : // Set tenant (and its timlines) to Stoppping state.
2606 : //
2607 : // Since we can only transition into Stopping state after activation is complete,
2608 : // run it in a JoinSet so all tenants have a chance to stop before we get SIGKILLed.
2609 : //
2610 : // Transitioning tenants to Stopping state has a couple of non-obvious side effects:
2611 : // 1. Lock out any new requests to the tenants.
2612 : // 2. Signal cancellation to WAL receivers (we wait on it below).
2613 : // 3. Signal cancellation for other tenant background loops.
2614 : // 4. ???
2615 : //
2616 : // The waiting for the cancellation is not done uniformly.
2617 : // We certainly wait for WAL receivers to shut down.
2618 : // That is necessary so that no new data comes in before the freeze_and_flush.
2619 : // But the tenant background loops are joined-on in our caller.
2620 : // It's mesed up.
2621 : // we just ignore the failure to stop
2622 :
2623 : // If we're still attaching, fire the cancellation token early to drop out: this
2624 : // will prevent us flushing, but ensures timely shutdown if some I/O during attach
2625 : // is very slow.
2626 6 : let shutdown_mode = if matches!(self.current_state(), TenantState::Attaching) {
2627 0 : self.cancel.cancel();
2628 0 :
2629 0 : // Having fired our cancellation token, do not try and flush timelines: their cancellation tokens
2630 0 : // are children of ours, so their flush loops will have shut down already
2631 0 : timeline::ShutdownMode::Hard
2632 : } else {
2633 6 : shutdown_mode
2634 : };
2635 :
2636 6 : match self.set_stopping(shutdown_progress, false, false).await {
2637 6 : Ok(()) => {}
2638 0 : Err(SetStoppingError::Broken) => {
2639 0 : // assume that this is acceptable
2640 0 : }
2641 0 : Err(SetStoppingError::AlreadyStopping(other)) => {
2642 0 : // give caller the option to wait for this this shutdown
2643 0 : info!("Tenant::shutdown: AlreadyStopping");
2644 0 : return Err(other);
2645 : }
2646 : };
2647 :
2648 6 : let mut js = tokio::task::JoinSet::new();
2649 6 : {
2650 6 : let timelines = self.timelines.lock().unwrap();
2651 6 : timelines.values().for_each(|timeline| {
2652 6 : let timeline = Arc::clone(timeline);
2653 6 : let timeline_id = timeline.timeline_id;
2654 6 : let span = tracing::info_span!("timeline_shutdown", %timeline_id, ?shutdown_mode);
2655 11 : js.spawn(async move { timeline.shutdown(shutdown_mode).instrument(span).await });
2656 6 : })
2657 6 : };
2658 6 : // test_long_timeline_create_then_tenant_delete is leaning on this message
2659 6 : tracing::info!("Waiting for timelines...");
2660 12 : while let Some(res) = js.join_next().await {
2661 0 : match res {
2662 6 : Ok(()) => {}
2663 0 : Err(je) if je.is_cancelled() => unreachable!("no cancelling used"),
2664 0 : Err(je) if je.is_panic() => { /* logged already */ }
2665 0 : Err(je) => warn!("unexpected JoinError: {je:?}"),
2666 : }
2667 : }
2668 :
2669 : // We cancel the Tenant's cancellation token _after_ the timelines have all shut down. This permits
2670 : // them to continue to do work during their shutdown methods, e.g. flushing data.
2671 6 : tracing::debug!("Cancelling CancellationToken");
2672 6 : self.cancel.cancel();
2673 6 :
2674 6 : // shutdown all tenant and timeline tasks: gc, compaction, page service
2675 6 : // No new tasks will be started for this tenant because it's in `Stopping` state.
2676 6 : //
2677 6 : // this will additionally shutdown and await all timeline tasks.
2678 6 : tracing::debug!("Waiting for tasks...");
2679 6 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), None).await;
2680 :
2681 6 : if let Some(walredo_mgr) = self.walredo_mgr.as_ref() {
2682 6 : walredo_mgr.shutdown().await;
2683 0 : }
2684 :
2685 : // Wait for any in-flight operations to complete
2686 6 : self.gate.close().await;
2687 :
2688 6 : remove_tenant_metrics(&self.tenant_shard_id);
2689 6 :
2690 6 : Ok(())
2691 6 : }
2692 :
2693 : /// Change tenant status to Stopping, to mark that it is being shut down.
2694 : ///
2695 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
2696 : ///
2697 : /// This function is not cancel-safe!
2698 : ///
2699 : /// `allow_transition_from_loading` is needed for the special case of loading task deleting the tenant.
2700 : /// `allow_transition_from_attaching` is needed for the special case of attaching deleted tenant.
2701 6 : async fn set_stopping(
2702 6 : &self,
2703 6 : progress: completion::Barrier,
2704 6 : _allow_transition_from_loading: bool,
2705 6 : allow_transition_from_attaching: bool,
2706 6 : ) -> Result<(), SetStoppingError> {
2707 6 : let mut rx = self.state.subscribe();
2708 6 :
2709 6 : // cannot stop before we're done activating, so wait out until we're done activating
2710 6 : rx.wait_for(|state| match state {
2711 0 : TenantState::Attaching if allow_transition_from_attaching => true,
2712 : TenantState::Activating(_) | TenantState::Attaching => {
2713 0 : info!(
2714 0 : "waiting for {} to turn Active|Broken|Stopping",
2715 0 : <&'static str>::from(state)
2716 : );
2717 0 : false
2718 : }
2719 6 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
2720 6 : })
2721 0 : .await
2722 6 : .expect("cannot drop self.state while on a &self method");
2723 6 :
2724 6 : // we now know we're done activating, let's see whether this task is the winner to transition into Stopping
2725 6 : let mut err = None;
2726 6 : let stopping = self.state.send_if_modified(|current_state| match current_state {
2727 : TenantState::Activating(_) => {
2728 0 : unreachable!("1we ensured above that we're done with activation, and, there is no re-activation")
2729 : }
2730 : TenantState::Attaching => {
2731 0 : if !allow_transition_from_attaching {
2732 0 : unreachable!("2we ensured above that we're done with activation, and, there is no re-activation")
2733 0 : };
2734 0 : *current_state = TenantState::Stopping { progress };
2735 0 : true
2736 : }
2737 : TenantState::Active => {
2738 : // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
2739 : // are created after the transition to Stopping. That's harmless, as the Timelines
2740 : // won't be accessible to anyone afterwards, because the Tenant is in Stopping state.
2741 6 : *current_state = TenantState::Stopping { progress };
2742 6 : // Continue stopping outside the closure. We need to grab timelines.lock()
2743 6 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
2744 6 : true
2745 : }
2746 0 : TenantState::Broken { reason, .. } => {
2747 0 : info!(
2748 0 : "Cannot set tenant to Stopping state, it is in Broken state due to: {reason}"
2749 : );
2750 0 : err = Some(SetStoppingError::Broken);
2751 0 : false
2752 : }
2753 0 : TenantState::Stopping { progress } => {
2754 0 : info!("Tenant is already in Stopping state");
2755 0 : err = Some(SetStoppingError::AlreadyStopping(progress.clone()));
2756 0 : false
2757 : }
2758 6 : });
2759 6 : match (stopping, err) {
2760 6 : (true, None) => {} // continue
2761 0 : (false, Some(err)) => return Err(err),
2762 0 : (true, Some(_)) => unreachable!(
2763 0 : "send_if_modified closure must error out if not transitioning to Stopping"
2764 0 : ),
2765 0 : (false, None) => unreachable!(
2766 0 : "send_if_modified closure must return true if transitioning to Stopping"
2767 0 : ),
2768 : }
2769 :
2770 6 : let timelines_accessor = self.timelines.lock().unwrap();
2771 6 : let not_broken_timelines = timelines_accessor
2772 6 : .values()
2773 6 : .filter(|timeline| !timeline.is_broken());
2774 12 : for timeline in not_broken_timelines {
2775 6 : timeline.set_state(TimelineState::Stopping);
2776 6 : }
2777 6 : Ok(())
2778 6 : }
2779 :
2780 : /// Method for tenant::mgr to transition us into Broken state in case of a late failure in
2781 : /// `remove_tenant_from_memory`
2782 : ///
2783 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
2784 : ///
2785 : /// In tests, we also use this to set tenants to Broken state on purpose.
2786 0 : pub(crate) async fn set_broken(&self, reason: String) {
2787 0 : let mut rx = self.state.subscribe();
2788 0 :
2789 0 : // The load & attach routines own the tenant state until it has reached `Active`.
2790 0 : // So, wait until it's done.
2791 0 : rx.wait_for(|state| match state {
2792 : TenantState::Activating(_) | TenantState::Attaching => {
2793 0 : info!(
2794 0 : "waiting for {} to turn Active|Broken|Stopping",
2795 0 : <&'static str>::from(state)
2796 : );
2797 0 : false
2798 : }
2799 0 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
2800 0 : })
2801 0 : .await
2802 0 : .expect("cannot drop self.state while on a &self method");
2803 0 :
2804 0 : // we now know we're done activating, let's see whether this task is the winner to transition into Broken
2805 0 : self.set_broken_no_wait(reason)
2806 0 : }
2807 :
2808 0 : pub(crate) fn set_broken_no_wait(&self, reason: impl Display) {
2809 0 : let reason = reason.to_string();
2810 0 : self.state.send_modify(|current_state| {
2811 0 : match *current_state {
2812 : TenantState::Activating(_) | TenantState::Attaching => {
2813 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
2814 : }
2815 : TenantState::Active => {
2816 0 : if cfg!(feature = "testing") {
2817 0 : warn!("Changing Active tenant to Broken state, reason: {}", reason);
2818 0 : *current_state = TenantState::broken_from_reason(reason);
2819 : } else {
2820 0 : unreachable!("not allowed to call set_broken on Active tenants in non-testing builds")
2821 : }
2822 : }
2823 : TenantState::Broken { .. } => {
2824 0 : warn!("Tenant is already in Broken state");
2825 : }
2826 : // This is the only "expected" path, any other path is a bug.
2827 : TenantState::Stopping { .. } => {
2828 0 : warn!(
2829 0 : "Marking Stopping tenant as Broken state, reason: {}",
2830 : reason
2831 : );
2832 0 : *current_state = TenantState::broken_from_reason(reason);
2833 : }
2834 : }
2835 0 : });
2836 0 : }
2837 :
2838 0 : pub fn subscribe_for_state_updates(&self) -> watch::Receiver<TenantState> {
2839 0 : self.state.subscribe()
2840 0 : }
2841 :
2842 : /// The activate_now semaphore is initialized with zero units. As soon as
2843 : /// we add a unit, waiters will be able to acquire a unit and proceed.
2844 0 : pub(crate) fn activate_now(&self) {
2845 0 : self.activate_now_sem.add_permits(1);
2846 0 : }
2847 :
2848 0 : pub(crate) async fn wait_to_become_active(
2849 0 : &self,
2850 0 : timeout: Duration,
2851 0 : ) -> Result<(), GetActiveTenantError> {
2852 0 : let mut receiver = self.state.subscribe();
2853 : loop {
2854 0 : let current_state = receiver.borrow_and_update().clone();
2855 0 : match current_state {
2856 : TenantState::Attaching | TenantState::Activating(_) => {
2857 : // in these states, there's a chance that we can reach ::Active
2858 0 : self.activate_now();
2859 0 : match timeout_cancellable(timeout, &self.cancel, receiver.changed()).await {
2860 0 : Ok(r) => {
2861 0 : r.map_err(
2862 0 : |_e: tokio::sync::watch::error::RecvError|
2863 : // Tenant existed but was dropped: report it as non-existent
2864 0 : GetActiveTenantError::NotFound(GetTenantError::NotFound(self.tenant_shard_id.tenant_id))
2865 0 : )?
2866 : }
2867 : Err(TimeoutCancellableError::Cancelled) => {
2868 0 : return Err(GetActiveTenantError::Cancelled);
2869 : }
2870 : Err(TimeoutCancellableError::Timeout) => {
2871 0 : return Err(GetActiveTenantError::WaitForActiveTimeout {
2872 0 : latest_state: Some(self.current_state()),
2873 0 : wait_time: timeout,
2874 0 : });
2875 : }
2876 : }
2877 : }
2878 : TenantState::Active { .. } => {
2879 0 : return Ok(());
2880 : }
2881 0 : TenantState::Broken { reason, .. } => {
2882 0 : // This is fatal, and reported distinctly from the general case of "will never be active" because
2883 0 : // it's logically a 500 to external API users (broken is always a bug).
2884 0 : return Err(GetActiveTenantError::Broken(reason));
2885 : }
2886 : TenantState::Stopping { .. } => {
2887 : // There's no chance the tenant can transition back into ::Active
2888 0 : return Err(GetActiveTenantError::WillNotBecomeActive(current_state));
2889 : }
2890 : }
2891 : }
2892 0 : }
2893 :
2894 0 : pub(crate) fn get_attach_mode(&self) -> AttachmentMode {
2895 0 : self.tenant_conf.load().location.attach_mode
2896 0 : }
2897 :
2898 : /// For API access: generate a LocationConfig equivalent to the one that would be used to
2899 : /// create a Tenant in the same state. Do not use this in hot paths: it's for relatively
2900 : /// rare external API calls, like a reconciliation at startup.
2901 0 : pub(crate) fn get_location_conf(&self) -> models::LocationConfig {
2902 0 : let conf = self.tenant_conf.load();
2903 :
2904 0 : let location_config_mode = match conf.location.attach_mode {
2905 0 : AttachmentMode::Single => models::LocationConfigMode::AttachedSingle,
2906 0 : AttachmentMode::Multi => models::LocationConfigMode::AttachedMulti,
2907 0 : AttachmentMode::Stale => models::LocationConfigMode::AttachedStale,
2908 : };
2909 :
2910 : // We have a pageserver TenantConf, we need the API-facing TenantConfig.
2911 0 : let tenant_config: models::TenantConfig = conf.tenant_conf.clone().into();
2912 0 :
2913 0 : models::LocationConfig {
2914 0 : mode: location_config_mode,
2915 0 : generation: self.generation.into(),
2916 0 : secondary_conf: None,
2917 0 : shard_number: self.shard_identity.number.0,
2918 0 : shard_count: self.shard_identity.count.literal(),
2919 0 : shard_stripe_size: self.shard_identity.stripe_size.0,
2920 0 : tenant_conf: tenant_config,
2921 0 : }
2922 0 : }
2923 :
2924 0 : pub(crate) fn get_tenant_shard_id(&self) -> &TenantShardId {
2925 0 : &self.tenant_shard_id
2926 0 : }
2927 :
2928 0 : pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
2929 0 : self.shard_identity.stripe_size
2930 0 : }
2931 :
2932 0 : pub(crate) fn get_generation(&self) -> Generation {
2933 0 : self.generation
2934 0 : }
2935 :
2936 : /// This function partially shuts down the tenant (it shuts down the Timelines) and is fallible,
2937 : /// and can leave the tenant in a bad state if it fails. The caller is responsible for
2938 : /// resetting this tenant to a valid state if we fail.
2939 0 : pub(crate) async fn split_prepare(
2940 0 : &self,
2941 0 : child_shards: &Vec<TenantShardId>,
2942 0 : ) -> anyhow::Result<()> {
2943 0 : let timelines = self.timelines.lock().unwrap().clone();
2944 0 : for timeline in timelines.values() {
2945 : // We do not block timeline creation/deletion during splits inside the pageserver: it is up to higher levels
2946 : // to ensure that they do not start a split if currently in the process of doing these.
2947 :
2948 : // Upload an index from the parent: this is partly to provide freshness for the
2949 : // child tenants that will copy it, and partly for general ease-of-debugging: there will
2950 : // always be a parent shard index in the same generation as we wrote the child shard index.
2951 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Uploading index");
2952 0 : timeline
2953 0 : .remote_client
2954 0 : .schedule_index_upload_for_file_changes()?;
2955 0 : timeline.remote_client.wait_completion().await?;
2956 :
2957 : // Shut down the timeline's remote client: this means that the indices we write
2958 : // for child shards will not be invalidated by the parent shard deleting layers.
2959 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Shutting down remote storage client");
2960 0 : timeline.remote_client.shutdown().await;
2961 :
2962 : // Download methods can still be used after shutdown, as they don't flow through the remote client's
2963 : // queue. In principal the RemoteTimelineClient could provide this without downloading it, but this
2964 : // operation is rare, so it's simpler to just download it (and robustly guarantees that the index
2965 : // we use here really is the remotely persistent one).
2966 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Downloading index_part from parent");
2967 0 : let result = timeline.remote_client
2968 0 : .download_index_file(&self.cancel)
2969 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=%timeline.timeline_id))
2970 0 : .await?;
2971 0 : let index_part = match result {
2972 : MaybeDeletedIndexPart::Deleted(_) => {
2973 0 : anyhow::bail!("Timeline deletion happened concurrently with split")
2974 : }
2975 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
2976 : };
2977 :
2978 0 : for child_shard in child_shards {
2979 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Uploading index_part for child {}", child_shard.to_index());
2980 0 : upload_index_part(
2981 0 : &self.remote_storage,
2982 0 : child_shard,
2983 0 : &timeline.timeline_id,
2984 0 : self.generation,
2985 0 : &index_part,
2986 0 : &self.cancel,
2987 0 : )
2988 0 : .await?;
2989 : }
2990 : }
2991 :
2992 : // TODO: also copy index files of offloaded timelines
2993 :
2994 0 : let tenant_manifest = self.tenant_manifest();
2995 0 : // TODO: generation support
2996 0 : let generation = remote_timeline_client::TENANT_MANIFEST_GENERATION;
2997 0 : for child_shard in child_shards {
2998 0 : tracing::info!(
2999 0 : "Uploading tenant manifest for child {}",
3000 0 : child_shard.to_index()
3001 : );
3002 0 : upload_tenant_manifest(
3003 0 : &self.remote_storage,
3004 0 : child_shard,
3005 0 : generation,
3006 0 : &tenant_manifest,
3007 0 : &self.cancel,
3008 0 : )
3009 0 : .await?;
3010 : }
3011 :
3012 0 : Ok(())
3013 0 : }
3014 :
3015 0 : pub(crate) fn get_sizes(&self) -> TopTenantShardItem {
3016 0 : let mut result = TopTenantShardItem {
3017 0 : id: self.tenant_shard_id,
3018 0 : resident_size: 0,
3019 0 : physical_size: 0,
3020 0 : max_logical_size: 0,
3021 0 : };
3022 :
3023 0 : for timeline in self.timelines.lock().unwrap().values() {
3024 0 : result.resident_size += timeline.metrics.resident_physical_size_gauge.get();
3025 0 :
3026 0 : result.physical_size += timeline
3027 0 : .remote_client
3028 0 : .metrics
3029 0 : .remote_physical_size_gauge
3030 0 : .get();
3031 0 : result.max_logical_size = std::cmp::max(
3032 0 : result.max_logical_size,
3033 0 : timeline.metrics.current_logical_size_gauge.get(),
3034 0 : );
3035 0 : }
3036 :
3037 0 : result
3038 0 : }
3039 : }
3040 :
3041 : /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
3042 : /// perform a topological sort, so that the parent of each timeline comes
3043 : /// before the children.
3044 : /// E extracts the ancestor from T
3045 : /// This allows for T to be different. It can be TimelineMetadata, can be Timeline itself, etc.
3046 186 : fn tree_sort_timelines<T, E>(
3047 186 : timelines: HashMap<TimelineId, T>,
3048 186 : extractor: E,
3049 186 : ) -> anyhow::Result<Vec<(TimelineId, T)>>
3050 186 : where
3051 186 : E: Fn(&T) -> Option<TimelineId>,
3052 186 : {
3053 186 : let mut result = Vec::with_capacity(timelines.len());
3054 186 :
3055 186 : let mut now = Vec::with_capacity(timelines.len());
3056 186 : // (ancestor, children)
3057 186 : let mut later: HashMap<TimelineId, Vec<(TimelineId, T)>> =
3058 186 : HashMap::with_capacity(timelines.len());
3059 :
3060 192 : for (timeline_id, value) in timelines {
3061 6 : if let Some(ancestor_id) = extractor(&value) {
3062 2 : let children = later.entry(ancestor_id).or_default();
3063 2 : children.push((timeline_id, value));
3064 4 : } else {
3065 4 : now.push((timeline_id, value));
3066 4 : }
3067 : }
3068 :
3069 192 : while let Some((timeline_id, metadata)) = now.pop() {
3070 6 : result.push((timeline_id, metadata));
3071 : // All children of this can be loaded now
3072 6 : if let Some(mut children) = later.remove(&timeline_id) {
3073 2 : now.append(&mut children);
3074 4 : }
3075 : }
3076 :
3077 : // All timelines should be visited now. Unless there were timelines with missing ancestors.
3078 186 : if !later.is_empty() {
3079 0 : for (missing_id, orphan_ids) in later {
3080 0 : for (orphan_id, _) in orphan_ids {
3081 0 : error!("could not load timeline {orphan_id} because its ancestor timeline {missing_id} could not be loaded");
3082 : }
3083 : }
3084 0 : bail!("could not load tenant because some timelines are missing ancestors");
3085 186 : }
3086 186 :
3087 186 : Ok(result)
3088 186 : }
3089 :
3090 : impl Tenant {
3091 0 : pub fn tenant_specific_overrides(&self) -> TenantConfOpt {
3092 0 : self.tenant_conf.load().tenant_conf.clone()
3093 0 : }
3094 :
3095 0 : pub fn effective_config(&self) -> TenantConf {
3096 0 : self.tenant_specific_overrides()
3097 0 : .merge(self.conf.default_tenant_conf.clone())
3098 0 : }
3099 :
3100 0 : pub fn get_checkpoint_distance(&self) -> u64 {
3101 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3102 0 : tenant_conf
3103 0 : .checkpoint_distance
3104 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
3105 0 : }
3106 :
3107 0 : pub fn get_checkpoint_timeout(&self) -> Duration {
3108 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3109 0 : tenant_conf
3110 0 : .checkpoint_timeout
3111 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
3112 0 : }
3113 :
3114 0 : pub fn get_compaction_target_size(&self) -> u64 {
3115 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3116 0 : tenant_conf
3117 0 : .compaction_target_size
3118 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
3119 0 : }
3120 :
3121 0 : pub fn get_compaction_period(&self) -> Duration {
3122 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3123 0 : tenant_conf
3124 0 : .compaction_period
3125 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
3126 0 : }
3127 :
3128 0 : pub fn get_compaction_threshold(&self) -> usize {
3129 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3130 0 : tenant_conf
3131 0 : .compaction_threshold
3132 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
3133 0 : }
3134 :
3135 0 : pub fn get_gc_horizon(&self) -> u64 {
3136 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3137 0 : tenant_conf
3138 0 : .gc_horizon
3139 0 : .unwrap_or(self.conf.default_tenant_conf.gc_horizon)
3140 0 : }
3141 :
3142 0 : pub fn get_gc_period(&self) -> Duration {
3143 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3144 0 : tenant_conf
3145 0 : .gc_period
3146 0 : .unwrap_or(self.conf.default_tenant_conf.gc_period)
3147 0 : }
3148 :
3149 0 : pub fn get_image_creation_threshold(&self) -> usize {
3150 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3151 0 : tenant_conf
3152 0 : .image_creation_threshold
3153 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
3154 0 : }
3155 :
3156 0 : pub fn get_pitr_interval(&self) -> Duration {
3157 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3158 0 : tenant_conf
3159 0 : .pitr_interval
3160 0 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
3161 0 : }
3162 :
3163 0 : pub fn get_min_resident_size_override(&self) -> Option<u64> {
3164 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3165 0 : tenant_conf
3166 0 : .min_resident_size_override
3167 0 : .or(self.conf.default_tenant_conf.min_resident_size_override)
3168 0 : }
3169 :
3170 0 : pub fn get_heatmap_period(&self) -> Option<Duration> {
3171 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3172 0 : let heatmap_period = tenant_conf
3173 0 : .heatmap_period
3174 0 : .unwrap_or(self.conf.default_tenant_conf.heatmap_period);
3175 0 : if heatmap_period.is_zero() {
3176 0 : None
3177 : } else {
3178 0 : Some(heatmap_period)
3179 : }
3180 0 : }
3181 :
3182 4 : pub fn get_lsn_lease_length(&self) -> Duration {
3183 4 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3184 4 : tenant_conf
3185 4 : .lsn_lease_length
3186 4 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
3187 4 : }
3188 :
3189 0 : pub(crate) fn tenant_manifest(&self) -> TenantManifest {
3190 0 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
3191 0 :
3192 0 : let mut timeline_manifests = timelines_offloaded
3193 0 : .iter()
3194 0 : .map(|(_timeline_id, offloaded)| offloaded.manifest())
3195 0 : .collect::<Vec<_>>();
3196 0 : // Sort the manifests so that our output is deterministic
3197 0 : timeline_manifests.sort_by_key(|timeline_manifest| timeline_manifest.timeline_id);
3198 0 :
3199 0 : TenantManifest {
3200 0 : version: LATEST_TENANT_MANIFEST_VERSION,
3201 0 : offloaded_timelines: timeline_manifests,
3202 0 : }
3203 0 : }
3204 :
3205 0 : pub fn set_new_tenant_config(&self, new_tenant_conf: TenantConfOpt) {
3206 0 : // Use read-copy-update in order to avoid overwriting the location config
3207 0 : // state if this races with [`Tenant::set_new_location_config`]. Note that
3208 0 : // this race is not possible if both request types come from the storage
3209 0 : // controller (as they should!) because an exclusive op lock is required
3210 0 : // on the storage controller side.
3211 0 : self.tenant_conf.rcu(|inner| {
3212 0 : Arc::new(AttachedTenantConf {
3213 0 : tenant_conf: new_tenant_conf.clone(),
3214 0 : location: inner.location,
3215 0 : // Attached location is not changed, no need to update lsn lease deadline.
3216 0 : lsn_lease_deadline: inner.lsn_lease_deadline,
3217 0 : })
3218 0 : });
3219 0 :
3220 0 : self.tenant_conf_updated(&new_tenant_conf);
3221 0 : // Don't hold self.timelines.lock() during the notifies.
3222 0 : // There's no risk of deadlock right now, but there could be if we consolidate
3223 0 : // mutexes in struct Timeline in the future.
3224 0 : let timelines = self.list_timelines();
3225 0 : for timeline in timelines {
3226 0 : timeline.tenant_conf_updated(&new_tenant_conf);
3227 0 : }
3228 0 : }
3229 :
3230 0 : pub(crate) fn set_new_location_config(&self, new_conf: AttachedTenantConf) {
3231 0 : let new_tenant_conf = new_conf.tenant_conf.clone();
3232 0 :
3233 0 : self.tenant_conf.store(Arc::new(new_conf));
3234 0 :
3235 0 : self.tenant_conf_updated(&new_tenant_conf);
3236 0 : // Don't hold self.timelines.lock() during the notifies.
3237 0 : // There's no risk of deadlock right now, but there could be if we consolidate
3238 0 : // mutexes in struct Timeline in the future.
3239 0 : let timelines = self.list_timelines();
3240 0 : for timeline in timelines {
3241 0 : timeline.tenant_conf_updated(&new_tenant_conf);
3242 0 : }
3243 0 : }
3244 :
3245 186 : fn get_timeline_get_throttle_config(
3246 186 : psconf: &'static PageServerConf,
3247 186 : overrides: &TenantConfOpt,
3248 186 : ) -> throttle::Config {
3249 186 : overrides
3250 186 : .timeline_get_throttle
3251 186 : .clone()
3252 186 : .unwrap_or(psconf.default_tenant_conf.timeline_get_throttle.clone())
3253 186 : }
3254 :
3255 0 : pub(crate) fn tenant_conf_updated(&self, new_conf: &TenantConfOpt) {
3256 0 : let conf = Self::get_timeline_get_throttle_config(self.conf, new_conf);
3257 0 : self.timeline_get_throttle.reconfigure(conf)
3258 0 : }
3259 :
3260 : /// Helper function to create a new Timeline struct.
3261 : ///
3262 : /// The returned Timeline is in Loading state. The caller is responsible for
3263 : /// initializing any on-disk state, and for inserting the Timeline to the 'timelines'
3264 : /// map.
3265 : ///
3266 : /// `validate_ancestor == false` is used when a timeline is created for deletion
3267 : /// and we might not have the ancestor present anymore which is fine for to be
3268 : /// deleted timelines.
3269 410 : fn create_timeline_struct(
3270 410 : &self,
3271 410 : new_timeline_id: TimelineId,
3272 410 : new_metadata: &TimelineMetadata,
3273 410 : ancestor: Option<Arc<Timeline>>,
3274 410 : resources: TimelineResources,
3275 410 : cause: CreateTimelineCause,
3276 410 : ) -> anyhow::Result<Arc<Timeline>> {
3277 410 : let state = match cause {
3278 : CreateTimelineCause::Load => {
3279 410 : let ancestor_id = new_metadata.ancestor_timeline();
3280 410 : anyhow::ensure!(
3281 410 : ancestor_id == ancestor.as_ref().map(|t| t.timeline_id),
3282 0 : "Timeline's {new_timeline_id} ancestor {ancestor_id:?} was not found"
3283 : );
3284 410 : TimelineState::Loading
3285 : }
3286 0 : CreateTimelineCause::Delete => TimelineState::Stopping,
3287 : };
3288 :
3289 410 : let pg_version = new_metadata.pg_version();
3290 410 :
3291 410 : let timeline = Timeline::new(
3292 410 : self.conf,
3293 410 : Arc::clone(&self.tenant_conf),
3294 410 : new_metadata,
3295 410 : ancestor,
3296 410 : new_timeline_id,
3297 410 : self.tenant_shard_id,
3298 410 : self.generation,
3299 410 : self.shard_identity,
3300 410 : self.walredo_mgr.clone(),
3301 410 : resources,
3302 410 : pg_version,
3303 410 : state,
3304 410 : self.attach_wal_lag_cooldown.clone(),
3305 410 : self.cancel.child_token(),
3306 410 : );
3307 410 :
3308 410 : Ok(timeline)
3309 410 : }
3310 :
3311 : // Allow too_many_arguments because a constructor's argument list naturally grows with the
3312 : // number of attributes in the struct: breaking these out into a builder wouldn't be helpful.
3313 : #[allow(clippy::too_many_arguments)]
3314 186 : fn new(
3315 186 : state: TenantState,
3316 186 : conf: &'static PageServerConf,
3317 186 : attached_conf: AttachedTenantConf,
3318 186 : shard_identity: ShardIdentity,
3319 186 : walredo_mgr: Option<Arc<WalRedoManager>>,
3320 186 : tenant_shard_id: TenantShardId,
3321 186 : remote_storage: GenericRemoteStorage,
3322 186 : deletion_queue_client: DeletionQueueClient,
3323 186 : l0_flush_global_state: L0FlushGlobalState,
3324 186 : ) -> Tenant {
3325 186 : debug_assert!(
3326 186 : !attached_conf.location.generation.is_none() || conf.control_plane_api.is_none()
3327 : );
3328 :
3329 186 : let (state, mut rx) = watch::channel(state);
3330 186 :
3331 186 : tokio::spawn(async move {
3332 186 : // reflect tenant state in metrics:
3333 186 : // - global per tenant state: TENANT_STATE_METRIC
3334 186 : // - "set" of broken tenants: BROKEN_TENANTS_SET
3335 186 : //
3336 186 : // set of broken tenants should not have zero counts so that it remains accessible for
3337 186 : // alerting.
3338 186 :
3339 186 : let tid = tenant_shard_id.to_string();
3340 186 : let shard_id = tenant_shard_id.shard_slug().to_string();
3341 186 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
3342 :
3343 369 : fn inspect_state(state: &TenantState) -> ([&'static str; 1], bool) {
3344 369 : ([state.into()], matches!(state, TenantState::Broken { .. }))
3345 369 : }
3346 :
3347 186 : let mut tuple = inspect_state(&rx.borrow_and_update());
3348 186 :
3349 186 : let is_broken = tuple.1;
3350 186 : let mut counted_broken = if is_broken {
3351 : // add the id to the set right away, there should not be any updates on the channel
3352 : // after before tenant is removed, if ever
3353 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
3354 0 : true
3355 : } else {
3356 186 : false
3357 : };
3358 :
3359 : loop {
3360 369 : let labels = &tuple.0;
3361 369 : let current = TENANT_STATE_METRIC.with_label_values(labels);
3362 369 : current.inc();
3363 369 :
3364 369 : if rx.changed().await.is_err() {
3365 : // tenant has been dropped
3366 16 : current.dec();
3367 16 : drop(BROKEN_TENANTS_SET.remove_label_values(set_key));
3368 16 : break;
3369 183 : }
3370 183 :
3371 183 : current.dec();
3372 183 : tuple = inspect_state(&rx.borrow_and_update());
3373 183 :
3374 183 : let is_broken = tuple.1;
3375 183 : if is_broken && !counted_broken {
3376 0 : counted_broken = true;
3377 0 : // insert the tenant_id (back) into the set while avoiding needless counter
3378 0 : // access
3379 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
3380 183 : }
3381 : }
3382 186 : });
3383 186 :
3384 186 : Tenant {
3385 186 : tenant_shard_id,
3386 186 : shard_identity,
3387 186 : generation: attached_conf.location.generation,
3388 186 : conf,
3389 186 : // using now here is good enough approximation to catch tenants with really long
3390 186 : // activation times.
3391 186 : constructed_at: Instant::now(),
3392 186 : timelines: Mutex::new(HashMap::new()),
3393 186 : timelines_creating: Mutex::new(HashSet::new()),
3394 186 : timelines_offloaded: Mutex::new(HashMap::new()),
3395 186 : gc_cs: tokio::sync::Mutex::new(()),
3396 186 : walredo_mgr,
3397 186 : remote_storage,
3398 186 : deletion_queue_client,
3399 186 : state,
3400 186 : cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
3401 186 : cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
3402 186 : eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
3403 186 : compaction_circuit_breaker: std::sync::Mutex::new(CircuitBreaker::new(
3404 186 : format!("compaction-{tenant_shard_id}"),
3405 186 : 5,
3406 186 : // Compaction can be a very expensive operation, and might leak disk space. It also ought
3407 186 : // to be infallible, as long as remote storage is available. So if it repeatedly fails,
3408 186 : // use an extremely long backoff.
3409 186 : Some(Duration::from_secs(3600 * 24)),
3410 186 : )),
3411 186 : activate_now_sem: tokio::sync::Semaphore::new(0),
3412 186 : attach_wal_lag_cooldown: Arc::new(std::sync::OnceLock::new()),
3413 186 : cancel: CancellationToken::default(),
3414 186 : gate: Gate::default(),
3415 186 : timeline_get_throttle: Arc::new(throttle::Throttle::new(
3416 186 : Tenant::get_timeline_get_throttle_config(conf, &attached_conf.tenant_conf),
3417 186 : crate::metrics::tenant_throttling::TimelineGet::new(&tenant_shard_id),
3418 186 : )),
3419 186 : tenant_conf: Arc::new(ArcSwap::from_pointee(attached_conf)),
3420 186 : ongoing_timeline_detach: std::sync::Mutex::default(),
3421 186 : gc_block: Default::default(),
3422 186 : l0_flush_global_state,
3423 186 : }
3424 186 : }
3425 :
3426 : /// Locate and load config
3427 0 : pub(super) fn load_tenant_config(
3428 0 : conf: &'static PageServerConf,
3429 0 : tenant_shard_id: &TenantShardId,
3430 0 : ) -> Result<LocationConf, LoadConfigError> {
3431 0 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
3432 0 :
3433 0 : info!("loading tenant configuration from {config_path}");
3434 :
3435 : // load and parse file
3436 0 : let config = fs::read_to_string(&config_path).map_err(|e| {
3437 0 : match e.kind() {
3438 : std::io::ErrorKind::NotFound => {
3439 : // The config should almost always exist for a tenant directory:
3440 : // - When attaching a tenant, the config is the first thing we write
3441 : // - When detaching a tenant, we atomically move the directory to a tmp location
3442 : // before deleting contents.
3443 : //
3444 : // The very rare edge case that can result in a missing config is if we crash during attach
3445 : // between creating directory and writing config. Callers should handle that as if the
3446 : // directory didn't exist.
3447 :
3448 0 : LoadConfigError::NotFound(config_path)
3449 : }
3450 : _ => {
3451 : // No IO errors except NotFound are acceptable here: other kinds of error indicate local storage or permissions issues
3452 : // that we cannot cleanly recover
3453 0 : crate::virtual_file::on_fatal_io_error(&e, "Reading tenant config file")
3454 : }
3455 : }
3456 0 : })?;
3457 :
3458 0 : Ok(toml_edit::de::from_str::<LocationConf>(&config)?)
3459 0 : }
3460 :
3461 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
3462 : pub(super) async fn persist_tenant_config(
3463 : conf: &'static PageServerConf,
3464 : tenant_shard_id: &TenantShardId,
3465 : location_conf: &LocationConf,
3466 : ) -> std::io::Result<()> {
3467 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
3468 :
3469 : Self::persist_tenant_config_at(tenant_shard_id, &config_path, location_conf).await
3470 : }
3471 :
3472 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
3473 : pub(super) async fn persist_tenant_config_at(
3474 : tenant_shard_id: &TenantShardId,
3475 : config_path: &Utf8Path,
3476 : location_conf: &LocationConf,
3477 : ) -> std::io::Result<()> {
3478 : debug!("persisting tenantconf to {config_path}");
3479 :
3480 : let mut conf_content = r#"# This file contains a specific per-tenant's config.
3481 : # It is read in case of pageserver restart.
3482 : "#
3483 : .to_string();
3484 :
3485 0 : fail::fail_point!("tenant-config-before-write", |_| {
3486 0 : Err(std::io::Error::new(
3487 0 : std::io::ErrorKind::Other,
3488 0 : "tenant-config-before-write",
3489 0 : ))
3490 0 : });
3491 :
3492 : // Convert the config to a toml file.
3493 : conf_content +=
3494 : &toml_edit::ser::to_string_pretty(&location_conf).expect("Config serialization failed");
3495 :
3496 : let temp_path = path_with_suffix_extension(config_path, TEMP_FILE_SUFFIX);
3497 :
3498 : let conf_content = conf_content.into_bytes();
3499 : VirtualFile::crashsafe_overwrite(config_path.to_owned(), temp_path, conf_content).await
3500 : }
3501 :
3502 : //
3503 : // How garbage collection works:
3504 : //
3505 : // +--bar------------->
3506 : // /
3507 : // +----+-----foo---------------->
3508 : // /
3509 : // ----main--+-------------------------->
3510 : // \
3511 : // +-----baz-------->
3512 : //
3513 : //
3514 : // 1. Grab 'gc_cs' mutex to prevent new timelines from being created while Timeline's
3515 : // `gc_infos` are being refreshed
3516 : // 2. Scan collected timelines, and on each timeline, make note of the
3517 : // all the points where other timelines have been branched off.
3518 : // We will refrain from removing page versions at those LSNs.
3519 : // 3. For each timeline, scan all layer files on the timeline.
3520 : // Remove all files for which a newer file exists and which
3521 : // don't cover any branch point LSNs.
3522 : //
3523 : // TODO:
3524 : // - if a relation has a non-incremental persistent layer on a child branch, then we
3525 : // don't need to keep that in the parent anymore. But currently
3526 : // we do.
3527 4 : async fn gc_iteration_internal(
3528 4 : &self,
3529 4 : target_timeline_id: Option<TimelineId>,
3530 4 : horizon: u64,
3531 4 : pitr: Duration,
3532 4 : cancel: &CancellationToken,
3533 4 : ctx: &RequestContext,
3534 4 : ) -> Result<GcResult, GcError> {
3535 4 : let mut totals: GcResult = Default::default();
3536 4 : let now = Instant::now();
3537 :
3538 4 : let gc_timelines = self
3539 4 : .refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
3540 3 : .await?;
3541 :
3542 4 : failpoint_support::sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
3543 :
3544 : // If there is nothing to GC, we don't want any messages in the INFO log.
3545 4 : if !gc_timelines.is_empty() {
3546 4 : info!("{} timelines need GC", gc_timelines.len());
3547 : } else {
3548 0 : debug!("{} timelines need GC", gc_timelines.len());
3549 : }
3550 :
3551 : // Perform GC for each timeline.
3552 : //
3553 : // Note that we don't hold the `Tenant::gc_cs` lock here because we don't want to delay the
3554 : // branch creation task, which requires the GC lock. A GC iteration can run concurrently
3555 : // with branch creation.
3556 : //
3557 : // See comments in [`Tenant::branch_timeline`] for more information about why branch
3558 : // creation task can run concurrently with timeline's GC iteration.
3559 8 : for timeline in gc_timelines {
3560 4 : if cancel.is_cancelled() {
3561 : // We were requested to shut down. Stop and return with the progress we
3562 : // made.
3563 0 : break;
3564 4 : }
3565 4 : let result = match timeline.gc().await {
3566 : Err(GcError::TimelineCancelled) => {
3567 0 : if target_timeline_id.is_some() {
3568 : // If we were targetting this specific timeline, surface cancellation to caller
3569 0 : return Err(GcError::TimelineCancelled);
3570 : } else {
3571 : // A timeline may be shutting down independently of the tenant's lifecycle: we should
3572 : // skip past this and proceed to try GC on other timelines.
3573 0 : continue;
3574 : }
3575 : }
3576 4 : r => r?,
3577 : };
3578 4 : totals += result;
3579 : }
3580 :
3581 4 : totals.elapsed = now.elapsed();
3582 4 : Ok(totals)
3583 4 : }
3584 :
3585 : /// Refreshes the Timeline::gc_info for all timelines, returning the
3586 : /// vector of timelines which have [`Timeline::get_last_record_lsn`] past
3587 : /// [`Tenant::get_gc_horizon`].
3588 : ///
3589 : /// This is usually executed as part of periodic gc, but can now be triggered more often.
3590 0 : pub(crate) async fn refresh_gc_info(
3591 0 : &self,
3592 0 : cancel: &CancellationToken,
3593 0 : ctx: &RequestContext,
3594 0 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
3595 0 : // since this method can now be called at different rates than the configured gc loop, it
3596 0 : // might be that these configuration values get applied faster than what it was previously,
3597 0 : // since these were only read from the gc task.
3598 0 : let horizon = self.get_gc_horizon();
3599 0 : let pitr = self.get_pitr_interval();
3600 0 :
3601 0 : // refresh all timelines
3602 0 : let target_timeline_id = None;
3603 0 :
3604 0 : self.refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
3605 0 : .await
3606 0 : }
3607 :
3608 : /// Populate all Timelines' `GcInfo` with information about their children. We do not set the
3609 : /// PITR cutoffs here, because that requires I/O: this is done later, before GC, by [`Self::refresh_gc_info_internal`]
3610 : ///
3611 : /// Subsequently, parent-child relationships are updated incrementally inside [`Timeline::new`] and [`Timeline::drop`].
3612 0 : fn initialize_gc_info(
3613 0 : &self,
3614 0 : timelines: &std::sync::MutexGuard<HashMap<TimelineId, Arc<Timeline>>>,
3615 0 : timelines_offloaded: &std::sync::MutexGuard<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
3616 0 : ) {
3617 0 : // This function must be called before activation: after activation timeline create/delete operations
3618 0 : // might happen, and this function is not safe to run concurrently with those.
3619 0 : assert!(!self.is_active());
3620 :
3621 : // Scan all timelines. For each timeline, remember the timeline ID and
3622 : // the branch point where it was created.
3623 0 : let mut all_branchpoints: BTreeMap<TimelineId, Vec<(Lsn, TimelineId, MaybeOffloaded)>> =
3624 0 : BTreeMap::new();
3625 0 : timelines.iter().for_each(|(timeline_id, timeline_entry)| {
3626 0 : if let Some(ancestor_timeline_id) = &timeline_entry.get_ancestor_timeline_id() {
3627 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
3628 0 : ancestor_children.push((
3629 0 : timeline_entry.get_ancestor_lsn(),
3630 0 : *timeline_id,
3631 0 : MaybeOffloaded::No,
3632 0 : ));
3633 0 : }
3634 0 : });
3635 0 : timelines_offloaded
3636 0 : .iter()
3637 0 : .for_each(|(timeline_id, timeline_entry)| {
3638 0 : let Some(ancestor_timeline_id) = &timeline_entry.ancestor_timeline_id else {
3639 0 : return;
3640 : };
3641 0 : let Some(retain_lsn) = timeline_entry.ancestor_retain_lsn else {
3642 0 : return;
3643 : };
3644 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
3645 0 : ancestor_children.push((retain_lsn, *timeline_id, MaybeOffloaded::Yes));
3646 0 : });
3647 0 :
3648 0 : // The number of bytes we always keep, irrespective of PITR: this is a constant across timelines
3649 0 : let horizon = self.get_gc_horizon();
3650 :
3651 : // Populate each timeline's GcInfo with information about its child branches
3652 0 : for timeline in timelines.values() {
3653 0 : let mut branchpoints: Vec<(Lsn, TimelineId, MaybeOffloaded)> = all_branchpoints
3654 0 : .remove(&timeline.timeline_id)
3655 0 : .unwrap_or_default();
3656 0 :
3657 0 : branchpoints.sort_by_key(|b| b.0);
3658 0 :
3659 0 : let mut target = timeline.gc_info.write().unwrap();
3660 0 :
3661 0 : target.retain_lsns = branchpoints;
3662 0 :
3663 0 : let space_cutoff = timeline
3664 0 : .get_last_record_lsn()
3665 0 : .checked_sub(horizon)
3666 0 : .unwrap_or(Lsn(0));
3667 0 :
3668 0 : target.cutoffs = GcCutoffs {
3669 0 : space: space_cutoff,
3670 0 : time: Lsn::INVALID,
3671 0 : };
3672 0 : }
3673 0 : }
3674 :
3675 4 : async fn refresh_gc_info_internal(
3676 4 : &self,
3677 4 : target_timeline_id: Option<TimelineId>,
3678 4 : horizon: u64,
3679 4 : pitr: Duration,
3680 4 : cancel: &CancellationToken,
3681 4 : ctx: &RequestContext,
3682 4 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
3683 4 : // before taking the gc_cs lock, do the heavier weight finding of gc_cutoff points for
3684 4 : // currently visible timelines.
3685 4 : let timelines = self
3686 4 : .timelines
3687 4 : .lock()
3688 4 : .unwrap()
3689 4 : .values()
3690 4 : .filter(|tl| match target_timeline_id.as_ref() {
3691 4 : Some(target) => &tl.timeline_id == target,
3692 0 : None => true,
3693 4 : })
3694 4 : .cloned()
3695 4 : .collect::<Vec<_>>();
3696 4 :
3697 4 : if target_timeline_id.is_some() && timelines.is_empty() {
3698 : // We were to act on a particular timeline and it wasn't found
3699 0 : return Err(GcError::TimelineNotFound);
3700 4 : }
3701 4 :
3702 4 : let mut gc_cutoffs: HashMap<TimelineId, GcCutoffs> =
3703 4 : HashMap::with_capacity(timelines.len());
3704 :
3705 4 : for timeline in timelines.iter() {
3706 4 : let cutoff = timeline
3707 4 : .get_last_record_lsn()
3708 4 : .checked_sub(horizon)
3709 4 : .unwrap_or(Lsn(0));
3710 :
3711 4 : let cutoffs = timeline.find_gc_cutoffs(cutoff, pitr, cancel, ctx).await?;
3712 4 : let old = gc_cutoffs.insert(timeline.timeline_id, cutoffs);
3713 4 : assert!(old.is_none());
3714 : }
3715 :
3716 4 : if !self.is_active() || self.cancel.is_cancelled() {
3717 0 : return Err(GcError::TenantCancelled);
3718 4 : }
3719 :
3720 : // grab mutex to prevent new timelines from being created here; avoid doing long operations
3721 : // because that will stall branch creation.
3722 4 : let gc_cs = self.gc_cs.lock().await;
3723 :
3724 : // Ok, we now know all the branch points.
3725 : // Update the GC information for each timeline.
3726 4 : let mut gc_timelines = Vec::with_capacity(timelines.len());
3727 8 : for timeline in timelines {
3728 : // We filtered the timeline list above
3729 4 : if let Some(target_timeline_id) = target_timeline_id {
3730 4 : assert_eq!(target_timeline_id, timeline.timeline_id);
3731 0 : }
3732 :
3733 : {
3734 4 : let mut target = timeline.gc_info.write().unwrap();
3735 4 :
3736 4 : // Cull any expired leases
3737 4 : let now = SystemTime::now();
3738 6 : target.leases.retain(|_, lease| !lease.is_expired(&now));
3739 4 :
3740 4 : timeline
3741 4 : .metrics
3742 4 : .valid_lsn_lease_count_gauge
3743 4 : .set(target.leases.len() as u64);
3744 :
3745 : // Look up parent's PITR cutoff to update the child's knowledge of whether it is within parent's PITR
3746 4 : if let Some(ancestor_id) = timeline.get_ancestor_timeline_id() {
3747 0 : if let Some(ancestor_gc_cutoffs) = gc_cutoffs.get(&ancestor_id) {
3748 0 : target.within_ancestor_pitr =
3749 0 : timeline.get_ancestor_lsn() >= ancestor_gc_cutoffs.time;
3750 0 : }
3751 4 : }
3752 :
3753 : // Update metrics that depend on GC state
3754 4 : timeline
3755 4 : .metrics
3756 4 : .archival_size
3757 4 : .set(if target.within_ancestor_pitr {
3758 0 : timeline.metrics.current_logical_size_gauge.get()
3759 : } else {
3760 4 : 0
3761 : });
3762 4 : timeline.metrics.pitr_history_size.set(
3763 4 : timeline
3764 4 : .get_last_record_lsn()
3765 4 : .checked_sub(target.cutoffs.time)
3766 4 : .unwrap_or(Lsn(0))
3767 4 : .0,
3768 4 : );
3769 :
3770 : // Apply the cutoffs we found to the Timeline's GcInfo. Why might we _not_ have cutoffs for a timeline?
3771 : // - this timeline was created while we were finding cutoffs
3772 : // - lsn for timestamp search fails for this timeline repeatedly
3773 4 : if let Some(cutoffs) = gc_cutoffs.get(&timeline.timeline_id) {
3774 4 : target.cutoffs = cutoffs.clone();
3775 4 : }
3776 : }
3777 :
3778 4 : gc_timelines.push(timeline);
3779 : }
3780 4 : drop(gc_cs);
3781 4 : Ok(gc_timelines)
3782 4 : }
3783 :
3784 : /// A substitute for `branch_timeline` for use in unit tests.
3785 : /// The returned timeline will have state value `Active` to make various `anyhow::ensure!()`
3786 : /// calls pass, but, we do not actually call `.activate()` under the hood. So, none of the
3787 : /// timeline background tasks are launched, except the flush loop.
3788 : #[cfg(test)]
3789 230 : async fn branch_timeline_test(
3790 230 : self: &Arc<Self>,
3791 230 : src_timeline: &Arc<Timeline>,
3792 230 : dst_id: TimelineId,
3793 230 : ancestor_lsn: Option<Lsn>,
3794 230 : ctx: &RequestContext,
3795 230 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
3796 230 : let tl = self
3797 230 : .branch_timeline_impl(src_timeline, dst_id, ancestor_lsn, ctx)
3798 179 : .await?
3799 226 : .into_timeline_for_test();
3800 226 : tl.set_state(TimelineState::Active);
3801 226 : Ok(tl)
3802 230 : }
3803 :
3804 : /// Helper for unit tests to branch a timeline with some pre-loaded states.
3805 : #[cfg(test)]
3806 : #[allow(clippy::too_many_arguments)]
3807 6 : pub async fn branch_timeline_test_with_layers(
3808 6 : self: &Arc<Self>,
3809 6 : src_timeline: &Arc<Timeline>,
3810 6 : dst_id: TimelineId,
3811 6 : ancestor_lsn: Option<Lsn>,
3812 6 : ctx: &RequestContext,
3813 6 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
3814 6 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
3815 6 : end_lsn: Lsn,
3816 6 : ) -> anyhow::Result<Arc<Timeline>> {
3817 : use checks::check_valid_layermap;
3818 : use itertools::Itertools;
3819 :
3820 6 : let tline = self
3821 6 : .branch_timeline_test(src_timeline, dst_id, ancestor_lsn, ctx)
3822 6 : .await?;
3823 6 : let ancestor_lsn = if let Some(ancestor_lsn) = ancestor_lsn {
3824 6 : ancestor_lsn
3825 : } else {
3826 0 : tline.get_last_record_lsn()
3827 : };
3828 6 : assert!(end_lsn >= ancestor_lsn);
3829 6 : tline.force_advance_lsn(end_lsn);
3830 12 : for deltas in delta_layer_desc {
3831 6 : tline
3832 6 : .force_create_delta_layer(deltas, Some(ancestor_lsn), ctx)
3833 18 : .await?;
3834 : }
3835 10 : for (lsn, images) in image_layer_desc {
3836 4 : tline
3837 4 : .force_create_image_layer(lsn, images, Some(ancestor_lsn), ctx)
3838 14 : .await?;
3839 : }
3840 6 : let layer_names = tline
3841 6 : .layers
3842 6 : .read()
3843 0 : .await
3844 6 : .layer_map()
3845 6 : .unwrap()
3846 6 : .iter_historic_layers()
3847 10 : .map(|layer| layer.layer_name())
3848 6 : .collect_vec();
3849 6 : if let Some(err) = check_valid_layermap(&layer_names) {
3850 0 : bail!("invalid layermap: {err}");
3851 6 : }
3852 6 : Ok(tline)
3853 6 : }
3854 :
3855 : /// Branch an existing timeline.
3856 0 : async fn branch_timeline(
3857 0 : self: &Arc<Self>,
3858 0 : src_timeline: &Arc<Timeline>,
3859 0 : dst_id: TimelineId,
3860 0 : start_lsn: Option<Lsn>,
3861 0 : ctx: &RequestContext,
3862 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
3863 0 : self.branch_timeline_impl(src_timeline, dst_id, start_lsn, ctx)
3864 0 : .await
3865 0 : }
3866 :
3867 230 : async fn branch_timeline_impl(
3868 230 : self: &Arc<Self>,
3869 230 : src_timeline: &Arc<Timeline>,
3870 230 : dst_id: TimelineId,
3871 230 : start_lsn: Option<Lsn>,
3872 230 : _ctx: &RequestContext,
3873 230 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
3874 230 : let src_id = src_timeline.timeline_id;
3875 :
3876 : // We will validate our ancestor LSN in this function. Acquire the GC lock so that
3877 : // this check cannot race with GC, and the ancestor LSN is guaranteed to remain
3878 : // valid while we are creating the branch.
3879 230 : let _gc_cs = self.gc_cs.lock().await;
3880 :
3881 : // If no start LSN is specified, we branch the new timeline from the source timeline's last record LSN
3882 230 : let start_lsn = start_lsn.unwrap_or_else(|| {
3883 2 : let lsn = src_timeline.get_last_record_lsn();
3884 2 : info!("branching timeline {dst_id} from timeline {src_id} at last record LSN: {lsn}");
3885 2 : lsn
3886 230 : });
3887 :
3888 : // we finally have determined the ancestor_start_lsn, so we can get claim exclusivity now
3889 230 : let timeline_create_guard = match self
3890 230 : .start_creating_timeline(
3891 230 : dst_id,
3892 230 : CreateTimelineIdempotency::Branch {
3893 230 : ancestor_timeline_id: src_timeline.timeline_id,
3894 230 : ancestor_start_lsn: start_lsn,
3895 230 : },
3896 230 : )
3897 179 : .await?
3898 : {
3899 230 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
3900 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
3901 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
3902 : }
3903 : };
3904 :
3905 : // Ensure that `start_lsn` is valid, i.e. the LSN is within the PITR
3906 : // horizon on the source timeline
3907 : //
3908 : // We check it against both the planned GC cutoff stored in 'gc_info',
3909 : // and the 'latest_gc_cutoff' of the last GC that was performed. The
3910 : // planned GC cutoff in 'gc_info' is normally larger than
3911 : // 'latest_gc_cutoff_lsn', but beware of corner cases like if you just
3912 : // changed the GC settings for the tenant to make the PITR window
3913 : // larger, but some of the data was already removed by an earlier GC
3914 : // iteration.
3915 :
3916 : // check against last actual 'latest_gc_cutoff' first
3917 230 : let latest_gc_cutoff_lsn = src_timeline.get_latest_gc_cutoff_lsn();
3918 230 : src_timeline
3919 230 : .check_lsn_is_in_scope(start_lsn, &latest_gc_cutoff_lsn)
3920 230 : .context(format!(
3921 230 : "invalid branch start lsn: less than latest GC cutoff {}",
3922 230 : *latest_gc_cutoff_lsn,
3923 230 : ))
3924 230 : .map_err(CreateTimelineError::AncestorLsn)?;
3925 :
3926 : // and then the planned GC cutoff
3927 : {
3928 226 : let gc_info = src_timeline.gc_info.read().unwrap();
3929 226 : let cutoff = gc_info.min_cutoff();
3930 226 : if start_lsn < cutoff {
3931 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
3932 0 : "invalid branch start lsn: less than planned GC cutoff {cutoff}"
3933 0 : )));
3934 226 : }
3935 226 : }
3936 226 :
3937 226 : //
3938 226 : // The branch point is valid, and we are still holding the 'gc_cs' lock
3939 226 : // so that GC cannot advance the GC cutoff until we are finished.
3940 226 : // Proceed with the branch creation.
3941 226 : //
3942 226 :
3943 226 : // Determine prev-LSN for the new timeline. We can only determine it if
3944 226 : // the timeline was branched at the current end of the source timeline.
3945 226 : let RecordLsn {
3946 226 : last: src_last,
3947 226 : prev: src_prev,
3948 226 : } = src_timeline.get_last_record_rlsn();
3949 226 : let dst_prev = if src_last == start_lsn {
3950 216 : Some(src_prev)
3951 : } else {
3952 10 : None
3953 : };
3954 :
3955 : // Create the metadata file, noting the ancestor of the new timeline.
3956 : // There is initially no data in it, but all the read-calls know to look
3957 : // into the ancestor.
3958 226 : let metadata = TimelineMetadata::new(
3959 226 : start_lsn,
3960 226 : dst_prev,
3961 226 : Some(src_id),
3962 226 : start_lsn,
3963 226 : *src_timeline.latest_gc_cutoff_lsn.read(), // FIXME: should we hold onto this guard longer?
3964 226 : src_timeline.initdb_lsn,
3965 226 : src_timeline.pg_version,
3966 226 : );
3967 :
3968 226 : let uninitialized_timeline = self
3969 226 : .prepare_new_timeline(
3970 226 : dst_id,
3971 226 : &metadata,
3972 226 : timeline_create_guard,
3973 226 : start_lsn + 1,
3974 226 : Some(Arc::clone(src_timeline)),
3975 226 : )
3976 0 : .await?;
3977 :
3978 226 : let new_timeline = uninitialized_timeline.finish_creation()?;
3979 :
3980 : // Root timeline gets its layers during creation and uploads them along with the metadata.
3981 : // A branch timeline though, when created, can get no writes for some time, hence won't get any layers created.
3982 : // We still need to upload its metadata eagerly: if other nodes `attach` the tenant and miss this timeline, their GC
3983 : // could get incorrect information and remove more layers, than needed.
3984 : // See also https://github.com/neondatabase/neon/issues/3865
3985 226 : new_timeline
3986 226 : .remote_client
3987 226 : .schedule_index_upload_for_full_metadata_update(&metadata)
3988 226 : .context("branch initial metadata upload")?;
3989 :
3990 226 : Ok(CreateTimelineResult::Created(new_timeline))
3991 230 : }
3992 :
3993 : /// For unit tests, make this visible so that other modules can directly create timelines
3994 : #[cfg(test)]
3995 2 : #[tracing::instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
3996 : pub(crate) async fn bootstrap_timeline_test(
3997 : self: &Arc<Self>,
3998 : timeline_id: TimelineId,
3999 : pg_version: u32,
4000 : load_existing_initdb: Option<TimelineId>,
4001 : ctx: &RequestContext,
4002 : ) -> anyhow::Result<Arc<Timeline>> {
4003 : self.bootstrap_timeline(timeline_id, pg_version, load_existing_initdb, ctx)
4004 : .await
4005 : .map_err(anyhow::Error::new)
4006 2 : .map(|r| r.into_timeline_for_test())
4007 : }
4008 :
4009 : /// Get exclusive access to the timeline ID for creation.
4010 : ///
4011 : /// Timeline-creating code paths must use this function before making changes
4012 : /// to in-memory or persistent state.
4013 : ///
4014 : /// The `state` parameter is a description of the timeline creation operation
4015 : /// we intend to perform.
4016 : /// If the timeline was already created in the meantime, we check whether this
4017 : /// request conflicts or is idempotent , based on `state`.
4018 232 : async fn start_creating_timeline(
4019 232 : &self,
4020 232 : new_timeline_id: TimelineId,
4021 232 : idempotency: CreateTimelineIdempotency,
4022 232 : ) -> Result<StartCreatingTimelineResult<'_>, CreateTimelineError> {
4023 232 : match self.create_timeline_create_guard(new_timeline_id) {
4024 232 : Ok(create_guard) => {
4025 232 : pausable_failpoint!("timeline-creation-after-uninit");
4026 232 : Ok(StartCreatingTimelineResult::CreateGuard(create_guard))
4027 : }
4028 : Err(TimelineExclusionError::AlreadyCreating) => {
4029 : // Creation is in progress, we cannot create it again, and we cannot
4030 : // check if this request matches the existing one, so caller must try
4031 : // again later.
4032 0 : Err(CreateTimelineError::AlreadyCreating)
4033 : }
4034 0 : Err(TimelineExclusionError::Other(e)) => Err(CreateTimelineError::Other(e)),
4035 0 : Err(TimelineExclusionError::AlreadyExists(existing)) => {
4036 0 : debug!("timeline already exists");
4037 :
4038 : // Idempotency: creating the same timeline twice is not an error, unless
4039 : // the second creation has different parameters.
4040 : //
4041 : // TODO: this is a crutch; we should store the CreateTimelineState as an
4042 : // immutable attribute in the index part, and compare them using derive(`Eq`).
4043 0 : match idempotency {
4044 0 : CreateTimelineIdempotency::Bootstrap { pg_version } => {
4045 0 : if existing.pg_version != pg_version {
4046 0 : info!("timeline already exists with different pg_version");
4047 0 : return Err(CreateTimelineError::Conflict);
4048 0 : }
4049 0 : if existing.get_ancestor_timeline_id().is_some() {
4050 0 : info!("timeline already exists with an ancestor");
4051 0 : return Err(CreateTimelineError::Conflict);
4052 0 : }
4053 0 : if existing.get_ancestor_lsn() != Lsn::INVALID {
4054 0 : info!("timeline already exists with an ancestor LSN");
4055 0 : return Err(CreateTimelineError::Conflict);
4056 0 : }
4057 : }
4058 : CreateTimelineIdempotency::Branch {
4059 0 : ancestor_timeline_id,
4060 0 : ancestor_start_lsn,
4061 0 : } => {
4062 0 : if existing.get_ancestor_timeline_id() != Some(ancestor_timeline_id) {
4063 0 : info!("timeline already exists with different ancestor");
4064 0 : return Err(CreateTimelineError::Conflict);
4065 0 : }
4066 0 : if existing.get_ancestor_lsn() != ancestor_start_lsn {
4067 0 : info!("timeline already exists with different ancestor LSN");
4068 0 : return Err(CreateTimelineError::Conflict);
4069 0 : }
4070 : }
4071 : }
4072 :
4073 : // Wait for uploads to complete, so that when we return Ok, the timeline
4074 : // is known to be durable on remote storage. Just like we do at the end of
4075 : // this function, after we have created the timeline ourselves.
4076 : //
4077 : // We only really care that the initial version of `index_part.json` has
4078 : // been uploaded. That's enough to remember that the timeline
4079 : // exists. However, there is no function to wait specifically for that so
4080 : // we just wait for all in-progress uploads to finish.
4081 0 : existing
4082 0 : .remote_client
4083 0 : .wait_completion()
4084 0 : .await
4085 0 : .context("wait for timeline uploads to complete")?;
4086 :
4087 : // TODO: shouldn't we also wait for timeline to become active?
4088 : // Code before this(https://github.com/neondatabase/neon/pull/9366) refactoring
4089 : // didn't do it.
4090 :
4091 0 : Ok(StartCreatingTimelineResult::Idempotent(existing))
4092 : }
4093 : }
4094 232 : }
4095 :
4096 0 : async fn upload_initdb(
4097 0 : &self,
4098 0 : timelines_path: &Utf8PathBuf,
4099 0 : pgdata_path: &Utf8PathBuf,
4100 0 : timeline_id: &TimelineId,
4101 0 : ) -> anyhow::Result<()> {
4102 0 : let temp_path = timelines_path.join(format!(
4103 0 : "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
4104 0 : ));
4105 0 :
4106 0 : scopeguard::defer! {
4107 0 : if let Err(e) = fs::remove_file(&temp_path) {
4108 0 : error!("Failed to remove temporary initdb archive '{temp_path}': {e}");
4109 0 : }
4110 0 : }
4111 :
4112 0 : let (pgdata_zstd, tar_zst_size) = create_zst_tarball(pgdata_path, &temp_path).await?;
4113 : const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
4114 0 : if tar_zst_size > INITDB_TAR_ZST_WARN_LIMIT {
4115 0 : warn!(
4116 0 : "compressed {temp_path} size of {tar_zst_size} is above limit {INITDB_TAR_ZST_WARN_LIMIT}."
4117 : );
4118 0 : }
4119 :
4120 0 : pausable_failpoint!("before-initdb-upload");
4121 :
4122 0 : backoff::retry(
4123 0 : || async {
4124 0 : self::remote_timeline_client::upload_initdb_dir(
4125 0 : &self.remote_storage,
4126 0 : &self.tenant_shard_id.tenant_id,
4127 0 : timeline_id,
4128 0 : pgdata_zstd.try_clone().await?,
4129 0 : tar_zst_size,
4130 0 : &self.cancel,
4131 : )
4132 0 : .await
4133 0 : },
4134 0 : |_| false,
4135 0 : 3,
4136 0 : u32::MAX,
4137 0 : "persist_initdb_tar_zst",
4138 0 : &self.cancel,
4139 0 : )
4140 0 : .await
4141 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
4142 0 : .and_then(|x| x)
4143 0 : }
4144 :
4145 : /// - run initdb to init temporary instance and get bootstrap data
4146 : /// - after initialization completes, tar up the temp dir and upload it to S3.
4147 2 : async fn bootstrap_timeline(
4148 2 : self: &Arc<Self>,
4149 2 : timeline_id: TimelineId,
4150 2 : pg_version: u32,
4151 2 : load_existing_initdb: Option<TimelineId>,
4152 2 : ctx: &RequestContext,
4153 2 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4154 2 : let timeline_create_guard = match self
4155 2 : .start_creating_timeline(
4156 2 : timeline_id,
4157 2 : CreateTimelineIdempotency::Bootstrap { pg_version },
4158 2 : )
4159 2 : .await?
4160 : {
4161 2 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
4162 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
4163 0 : return Ok(CreateTimelineResult::Idempotent(timeline))
4164 : }
4165 : };
4166 :
4167 : // create a `tenant/{tenant_id}/timelines/basebackup-{timeline_id}.{TEMP_FILE_SUFFIX}/`
4168 : // temporary directory for basebackup files for the given timeline.
4169 :
4170 2 : let timelines_path = self.conf.timelines_path(&self.tenant_shard_id);
4171 2 : let pgdata_path = path_with_suffix_extension(
4172 2 : timelines_path.join(format!("basebackup-{timeline_id}")),
4173 2 : TEMP_FILE_SUFFIX,
4174 2 : );
4175 2 :
4176 2 : // Remove whatever was left from the previous runs: safe because TimelineCreateGuard guarantees
4177 2 : // we won't race with other creations or existent timelines with the same path.
4178 2 : if pgdata_path.exists() {
4179 0 : fs::remove_dir_all(&pgdata_path).with_context(|| {
4180 0 : format!("Failed to remove already existing initdb directory: {pgdata_path}")
4181 0 : })?;
4182 2 : }
4183 :
4184 : // this new directory is very temporary, set to remove it immediately after bootstrap, we don't need it
4185 2 : scopeguard::defer! {
4186 2 : if let Err(e) = fs::remove_dir_all(&pgdata_path) {
4187 2 : // this is unlikely, but we will remove the directory on pageserver restart or another bootstrap call
4188 2 : error!("Failed to remove temporary initdb directory '{pgdata_path}': {e}");
4189 2 : }
4190 2 : }
4191 2 : if let Some(existing_initdb_timeline_id) = load_existing_initdb {
4192 2 : if existing_initdb_timeline_id != timeline_id {
4193 0 : let source_path = &remote_initdb_archive_path(
4194 0 : &self.tenant_shard_id.tenant_id,
4195 0 : &existing_initdb_timeline_id,
4196 0 : );
4197 0 : let dest_path =
4198 0 : &remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &timeline_id);
4199 0 :
4200 0 : // if this fails, it will get retried by retried control plane requests
4201 0 : self.remote_storage
4202 0 : .copy_object(source_path, dest_path, &self.cancel)
4203 0 : .await
4204 0 : .context("copy initdb tar")?;
4205 2 : }
4206 2 : let (initdb_tar_zst_path, initdb_tar_zst) =
4207 2 : self::remote_timeline_client::download_initdb_tar_zst(
4208 2 : self.conf,
4209 2 : &self.remote_storage,
4210 2 : &self.tenant_shard_id,
4211 2 : &existing_initdb_timeline_id,
4212 2 : &self.cancel,
4213 2 : )
4214 421 : .await
4215 2 : .context("download initdb tar")?;
4216 :
4217 2 : scopeguard::defer! {
4218 2 : if let Err(e) = fs::remove_file(&initdb_tar_zst_path) {
4219 2 : error!("Failed to remove temporary initdb archive '{initdb_tar_zst_path}': {e}");
4220 2 : }
4221 2 : }
4222 2 :
4223 2 : let buf_read =
4224 2 : BufReader::with_capacity(remote_timeline_client::BUFFER_SIZE, initdb_tar_zst);
4225 2 : extract_zst_tarball(&pgdata_path, buf_read)
4226 7742 : .await
4227 2 : .context("extract initdb tar")?;
4228 : } else {
4229 : // Init temporarily repo to get bootstrap data, this creates a directory in the `pgdata_path` path
4230 0 : run_initdb(self.conf, &pgdata_path, pg_version, &self.cancel)
4231 0 : .await
4232 0 : .context("run initdb")?;
4233 :
4234 : // Upload the created data dir to S3
4235 0 : if self.tenant_shard_id().is_shard_zero() {
4236 0 : self.upload_initdb(&timelines_path, &pgdata_path, &timeline_id)
4237 0 : .await?;
4238 0 : }
4239 : }
4240 2 : let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
4241 2 :
4242 2 : // Import the contents of the data directory at the initial checkpoint
4243 2 : // LSN, and any WAL after that.
4244 2 : // Initdb lsn will be equal to last_record_lsn which will be set after import.
4245 2 : // Because we know it upfront avoid having an option or dummy zero value by passing it to the metadata.
4246 2 : let new_metadata = TimelineMetadata::new(
4247 2 : Lsn(0),
4248 2 : None,
4249 2 : None,
4250 2 : Lsn(0),
4251 2 : pgdata_lsn,
4252 2 : pgdata_lsn,
4253 2 : pg_version,
4254 2 : );
4255 2 : let raw_timeline = self
4256 2 : .prepare_new_timeline(
4257 2 : timeline_id,
4258 2 : &new_metadata,
4259 2 : timeline_create_guard,
4260 2 : pgdata_lsn,
4261 2 : None,
4262 2 : )
4263 0 : .await?;
4264 :
4265 2 : let tenant_shard_id = raw_timeline.owning_tenant.tenant_shard_id;
4266 2 : let unfinished_timeline = raw_timeline.raw_timeline()?;
4267 :
4268 : // Flush the new layer files to disk, before we make the timeline as available to
4269 : // the outside world.
4270 : //
4271 : // Flush loop needs to be spawned in order to be able to flush.
4272 2 : unfinished_timeline.maybe_spawn_flush_loop();
4273 2 :
4274 2 : import_datadir::import_timeline_from_postgres_datadir(
4275 2 : unfinished_timeline,
4276 2 : &pgdata_path,
4277 2 : pgdata_lsn,
4278 2 : ctx,
4279 2 : )
4280 8495 : .await
4281 2 : .with_context(|| {
4282 0 : format!("Failed to import pgdatadir for timeline {tenant_shard_id}/{timeline_id}")
4283 2 : })?;
4284 :
4285 2 : fail::fail_point!("before-checkpoint-new-timeline", |_| {
4286 0 : Err(CreateTimelineError::Other(anyhow::anyhow!(
4287 0 : "failpoint before-checkpoint-new-timeline"
4288 0 : )))
4289 2 : });
4290 :
4291 2 : unfinished_timeline
4292 2 : .freeze_and_flush()
4293 2 : .await
4294 2 : .with_context(|| {
4295 0 : format!(
4296 0 : "Failed to flush after pgdatadir import for timeline {tenant_shard_id}/{timeline_id}"
4297 0 : )
4298 2 : })?;
4299 :
4300 : // All done!
4301 2 : let timeline = raw_timeline.finish_creation()?;
4302 :
4303 2 : Ok(CreateTimelineResult::Created(timeline))
4304 2 : }
4305 :
4306 404 : fn build_timeline_remote_client(&self, timeline_id: TimelineId) -> RemoteTimelineClient {
4307 404 : RemoteTimelineClient::new(
4308 404 : self.remote_storage.clone(),
4309 404 : self.deletion_queue_client.clone(),
4310 404 : self.conf,
4311 404 : self.tenant_shard_id,
4312 404 : timeline_id,
4313 404 : self.generation,
4314 404 : )
4315 404 : }
4316 :
4317 : /// Call this before constructing a timeline, to build its required structures
4318 404 : fn build_timeline_resources(&self, timeline_id: TimelineId) -> TimelineResources {
4319 404 : TimelineResources {
4320 404 : remote_client: self.build_timeline_remote_client(timeline_id),
4321 404 : timeline_get_throttle: self.timeline_get_throttle.clone(),
4322 404 : l0_flush_global_state: self.l0_flush_global_state.clone(),
4323 404 : }
4324 404 : }
4325 :
4326 : /// Creates intermediate timeline structure and its files.
4327 : ///
4328 : /// An empty layer map is initialized, and new data and WAL can be imported starting
4329 : /// at 'disk_consistent_lsn'. After any initial data has been imported, call
4330 : /// `finish_creation` to insert the Timeline into the timelines map.
4331 404 : async fn prepare_new_timeline<'a>(
4332 404 : &'a self,
4333 404 : new_timeline_id: TimelineId,
4334 404 : new_metadata: &TimelineMetadata,
4335 404 : create_guard: TimelineCreateGuard<'a>,
4336 404 : start_lsn: Lsn,
4337 404 : ancestor: Option<Arc<Timeline>>,
4338 404 : ) -> anyhow::Result<UninitializedTimeline<'a>> {
4339 404 : let tenant_shard_id = self.tenant_shard_id;
4340 404 :
4341 404 : let resources = self.build_timeline_resources(new_timeline_id);
4342 404 : resources
4343 404 : .remote_client
4344 404 : .init_upload_queue_for_empty_remote(new_metadata)?;
4345 :
4346 404 : let timeline_struct = self
4347 404 : .create_timeline_struct(
4348 404 : new_timeline_id,
4349 404 : new_metadata,
4350 404 : ancestor,
4351 404 : resources,
4352 404 : CreateTimelineCause::Load,
4353 404 : )
4354 404 : .context("Failed to create timeline data structure")?;
4355 :
4356 404 : timeline_struct.init_empty_layer_map(start_lsn);
4357 :
4358 404 : if let Err(e) = self
4359 404 : .create_timeline_files(&create_guard.timeline_path)
4360 0 : .await
4361 : {
4362 0 : error!("Failed to create initial files for timeline {tenant_shard_id}/{new_timeline_id}, cleaning up: {e:?}");
4363 0 : cleanup_timeline_directory(create_guard);
4364 0 : return Err(e);
4365 404 : }
4366 404 :
4367 404 : debug!(
4368 0 : "Successfully created initial files for timeline {tenant_shard_id}/{new_timeline_id}"
4369 : );
4370 :
4371 404 : Ok(UninitializedTimeline::new(
4372 404 : self,
4373 404 : new_timeline_id,
4374 404 : Some((timeline_struct, create_guard)),
4375 404 : ))
4376 404 : }
4377 :
4378 404 : async fn create_timeline_files(&self, timeline_path: &Utf8Path) -> anyhow::Result<()> {
4379 404 : crashsafe::create_dir(timeline_path).context("Failed to create timeline directory")?;
4380 :
4381 404 : fail::fail_point!("after-timeline-dir-creation", |_| {
4382 0 : anyhow::bail!("failpoint after-timeline-dir-creation");
4383 404 : });
4384 :
4385 404 : Ok(())
4386 404 : }
4387 :
4388 : /// Get a guard that provides exclusive access to the timeline directory, preventing
4389 : /// concurrent attempts to create the same timeline.
4390 410 : fn create_timeline_create_guard(
4391 410 : &self,
4392 410 : timeline_id: TimelineId,
4393 410 : ) -> Result<TimelineCreateGuard, TimelineExclusionError> {
4394 410 : let tenant_shard_id = self.tenant_shard_id;
4395 410 :
4396 410 : let timeline_path = self.conf.timeline_path(&tenant_shard_id, &timeline_id);
4397 :
4398 410 : let create_guard = TimelineCreateGuard::new(self, timeline_id, timeline_path.clone())?;
4399 :
4400 : // At this stage, we have got exclusive access to in-memory state for this timeline ID
4401 : // for creation.
4402 : // A timeline directory should never exist on disk already:
4403 : // - a previous failed creation would have cleaned up after itself
4404 : // - a pageserver restart would clean up timeline directories that don't have valid remote state
4405 : //
4406 : // Therefore it is an unexpected internal error to encounter a timeline directory already existing here,
4407 : // this error may indicate a bug in cleanup on failed creations.
4408 408 : if timeline_path.exists() {
4409 0 : return Err(TimelineExclusionError::Other(anyhow::anyhow!(
4410 0 : "Timeline directory already exists! This is a bug."
4411 0 : )));
4412 408 : }
4413 408 :
4414 408 : Ok(create_guard)
4415 410 : }
4416 :
4417 : /// Gathers inputs from all of the timelines to produce a sizing model input.
4418 : ///
4419 : /// Future is cancellation safe. Only one calculation can be running at once per tenant.
4420 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
4421 : pub async fn gather_size_inputs(
4422 : &self,
4423 : // `max_retention_period` overrides the cutoff that is used to calculate the size
4424 : // (only if it is shorter than the real cutoff).
4425 : max_retention_period: Option<u64>,
4426 : cause: LogicalSizeCalculationCause,
4427 : cancel: &CancellationToken,
4428 : ctx: &RequestContext,
4429 : ) -> Result<size::ModelInputs, size::CalculateSyntheticSizeError> {
4430 : let logical_sizes_at_once = self
4431 : .conf
4432 : .concurrent_tenant_size_logical_size_queries
4433 : .inner();
4434 :
4435 : // TODO: Having a single mutex block concurrent reads is not great for performance.
4436 : //
4437 : // But the only case where we need to run multiple of these at once is when we
4438 : // request a size for a tenant manually via API, while another background calculation
4439 : // is in progress (which is not a common case).
4440 : //
4441 : // See more for on the issue #2748 condenced out of the initial PR review.
4442 : let mut shared_cache = tokio::select! {
4443 : locked = self.cached_logical_sizes.lock() => locked,
4444 : _ = cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
4445 : _ = self.cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
4446 : };
4447 :
4448 : size::gather_inputs(
4449 : self,
4450 : logical_sizes_at_once,
4451 : max_retention_period,
4452 : &mut shared_cache,
4453 : cause,
4454 : cancel,
4455 : ctx,
4456 : )
4457 : .await
4458 : }
4459 :
4460 : /// Calculate synthetic tenant size and cache the result.
4461 : /// This is periodically called by background worker.
4462 : /// result is cached in tenant struct
4463 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
4464 : pub async fn calculate_synthetic_size(
4465 : &self,
4466 : cause: LogicalSizeCalculationCause,
4467 : cancel: &CancellationToken,
4468 : ctx: &RequestContext,
4469 : ) -> Result<u64, size::CalculateSyntheticSizeError> {
4470 : let inputs = self.gather_size_inputs(None, cause, cancel, ctx).await?;
4471 :
4472 : let size = inputs.calculate();
4473 :
4474 : self.set_cached_synthetic_size(size);
4475 :
4476 : Ok(size)
4477 : }
4478 :
4479 : /// Cache given synthetic size and update the metric value
4480 0 : pub fn set_cached_synthetic_size(&self, size: u64) {
4481 0 : self.cached_synthetic_tenant_size
4482 0 : .store(size, Ordering::Relaxed);
4483 0 :
4484 0 : // Only shard zero should be calculating synthetic sizes
4485 0 : debug_assert!(self.shard_identity.is_shard_zero());
4486 :
4487 0 : TENANT_SYNTHETIC_SIZE_METRIC
4488 0 : .get_metric_with_label_values(&[&self.tenant_shard_id.tenant_id.to_string()])
4489 0 : .unwrap()
4490 0 : .set(size);
4491 0 : }
4492 :
4493 0 : pub fn cached_synthetic_size(&self) -> u64 {
4494 0 : self.cached_synthetic_tenant_size.load(Ordering::Relaxed)
4495 0 : }
4496 :
4497 : /// Flush any in-progress layers, schedule uploads, and wait for uploads to complete.
4498 : ///
4499 : /// This function can take a long time: callers should wrap it in a timeout if calling
4500 : /// from an external API handler.
4501 : ///
4502 : /// Cancel-safety: cancelling this function may leave I/O running, but such I/O is
4503 : /// still bounded by tenant/timeline shutdown.
4504 0 : #[tracing::instrument(skip_all)]
4505 : pub(crate) async fn flush_remote(&self) -> anyhow::Result<()> {
4506 : let timelines = self.timelines.lock().unwrap().clone();
4507 :
4508 0 : async fn flush_timeline(_gate: GateGuard, timeline: Arc<Timeline>) -> anyhow::Result<()> {
4509 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Flushing...");
4510 0 : timeline.freeze_and_flush().await?;
4511 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Waiting for uploads...");
4512 0 : timeline.remote_client.wait_completion().await?;
4513 :
4514 0 : Ok(())
4515 0 : }
4516 :
4517 : // We do not use a JoinSet for these tasks, because we don't want them to be
4518 : // aborted when this function's future is cancelled: they should stay alive
4519 : // holding their GateGuard until they complete, to ensure their I/Os complete
4520 : // before Timeline shutdown completes.
4521 : let mut results = FuturesUnordered::new();
4522 :
4523 : for (_timeline_id, timeline) in timelines {
4524 : // Run each timeline's flush in a task holding the timeline's gate: this
4525 : // means that if this function's future is cancelled, the Timeline shutdown
4526 : // will still wait for any I/O in here to complete.
4527 : let Ok(gate) = timeline.gate.enter() else {
4528 : continue;
4529 : };
4530 0 : let jh = tokio::task::spawn(async move { flush_timeline(gate, timeline).await });
4531 : results.push(jh);
4532 : }
4533 :
4534 : while let Some(r) = results.next().await {
4535 : if let Err(e) = r {
4536 : if !e.is_cancelled() && !e.is_panic() {
4537 : tracing::error!("unexpected join error: {e:?}");
4538 : }
4539 : }
4540 : }
4541 :
4542 : // The flushes we did above were just writes, but the Tenant might have had
4543 : // pending deletions as well from recent compaction/gc: we want to flush those
4544 : // as well. This requires flushing the global delete queue. This is cheap
4545 : // because it's typically a no-op.
4546 : match self.deletion_queue_client.flush_execute().await {
4547 : Ok(_) => {}
4548 : Err(DeletionQueueError::ShuttingDown) => {}
4549 : }
4550 :
4551 : Ok(())
4552 : }
4553 :
4554 0 : pub(crate) fn get_tenant_conf(&self) -> TenantConfOpt {
4555 0 : self.tenant_conf.load().tenant_conf.clone()
4556 0 : }
4557 :
4558 : /// How much local storage would this tenant like to have? It can cope with
4559 : /// less than this (via eviction and on-demand downloads), but this function enables
4560 : /// the Tenant to advertise how much storage it would prefer to have to provide fast I/O
4561 : /// by keeping important things on local disk.
4562 : ///
4563 : /// This is a heuristic, not a guarantee: tenants that are long-idle will actually use less
4564 : /// than they report here, due to layer eviction. Tenants with many active branches may
4565 : /// actually use more than they report here.
4566 0 : pub(crate) fn local_storage_wanted(&self) -> u64 {
4567 0 : let timelines = self.timelines.lock().unwrap();
4568 0 :
4569 0 : // Heuristic: we use the max() of the timelines' visible sizes, rather than the sum. This
4570 0 : // reflects the observation that on tenants with multiple large branches, typically only one
4571 0 : // of them is used actively enough to occupy space on disk.
4572 0 : timelines
4573 0 : .values()
4574 0 : .map(|t| t.metrics.visible_physical_size_gauge.get())
4575 0 : .max()
4576 0 : .unwrap_or(0)
4577 0 : }
4578 : }
4579 :
4580 : /// Create the cluster temporarily in 'initdbpath' directory inside the repository
4581 : /// to get bootstrap data for timeline initialization.
4582 0 : async fn run_initdb(
4583 0 : conf: &'static PageServerConf,
4584 0 : initdb_target_dir: &Utf8Path,
4585 0 : pg_version: u32,
4586 0 : cancel: &CancellationToken,
4587 0 : ) -> Result<(), InitdbError> {
4588 0 : let initdb_bin_path = conf
4589 0 : .pg_bin_dir(pg_version)
4590 0 : .map_err(InitdbError::Other)?
4591 0 : .join("initdb");
4592 0 : let initdb_lib_dir = conf.pg_lib_dir(pg_version).map_err(InitdbError::Other)?;
4593 0 : info!(
4594 0 : "running {} in {}, libdir: {}",
4595 : initdb_bin_path, initdb_target_dir, initdb_lib_dir,
4596 : );
4597 :
4598 0 : let _permit = INIT_DB_SEMAPHORE.acquire().await;
4599 :
4600 0 : let initdb_command = tokio::process::Command::new(&initdb_bin_path)
4601 0 : .args(["--pgdata", initdb_target_dir.as_ref()])
4602 0 : .args(["--username", &conf.superuser])
4603 0 : .args(["--encoding", "utf8"])
4604 0 : .arg("--no-instructions")
4605 0 : .arg("--no-sync")
4606 0 : .env_clear()
4607 0 : .env("LD_LIBRARY_PATH", &initdb_lib_dir)
4608 0 : .env("DYLD_LIBRARY_PATH", &initdb_lib_dir)
4609 0 : .stdin(std::process::Stdio::null())
4610 0 : // stdout invocation produces the same output every time, we don't need it
4611 0 : .stdout(std::process::Stdio::null())
4612 0 : // we would be interested in the stderr output, if there was any
4613 0 : .stderr(std::process::Stdio::piped())
4614 0 : .spawn()?;
4615 :
4616 : // Ideally we'd select here with the cancellation token, but the problem is that
4617 : // we can't safely terminate initdb: it launches processes of its own, and killing
4618 : // initdb doesn't kill them. After we return from this function, we want the target
4619 : // directory to be able to be cleaned up.
4620 : // See https://github.com/neondatabase/neon/issues/6385
4621 0 : let initdb_output = initdb_command.wait_with_output().await?;
4622 0 : if !initdb_output.status.success() {
4623 0 : return Err(InitdbError::Failed(
4624 0 : initdb_output.status,
4625 0 : initdb_output.stderr,
4626 0 : ));
4627 0 : }
4628 0 :
4629 0 : // This isn't true cancellation support, see above. Still return an error to
4630 0 : // excercise the cancellation code path.
4631 0 : if cancel.is_cancelled() {
4632 0 : return Err(InitdbError::Cancelled);
4633 0 : }
4634 0 :
4635 0 : Ok(())
4636 0 : }
4637 :
4638 : /// Dump contents of a layer file to stdout.
4639 0 : pub async fn dump_layerfile_from_path(
4640 0 : path: &Utf8Path,
4641 0 : verbose: bool,
4642 0 : ctx: &RequestContext,
4643 0 : ) -> anyhow::Result<()> {
4644 : use std::os::unix::fs::FileExt;
4645 :
4646 : // All layer files start with a two-byte "magic" value, to identify the kind of
4647 : // file.
4648 0 : let file = File::open(path)?;
4649 0 : let mut header_buf = [0u8; 2];
4650 0 : file.read_exact_at(&mut header_buf, 0)?;
4651 :
4652 0 : match u16::from_be_bytes(header_buf) {
4653 : crate::IMAGE_FILE_MAGIC => {
4654 0 : ImageLayer::new_for_path(path, file)?
4655 0 : .dump(verbose, ctx)
4656 0 : .await?
4657 : }
4658 : crate::DELTA_FILE_MAGIC => {
4659 0 : DeltaLayer::new_for_path(path, file)?
4660 0 : .dump(verbose, ctx)
4661 0 : .await?
4662 : }
4663 0 : magic => bail!("unrecognized magic identifier: {:?}", magic),
4664 : }
4665 :
4666 0 : Ok(())
4667 0 : }
4668 :
4669 : #[cfg(test)]
4670 : pub(crate) mod harness {
4671 : use bytes::{Bytes, BytesMut};
4672 : use once_cell::sync::OnceCell;
4673 : use pageserver_api::models::ShardParameters;
4674 : use pageserver_api::shard::ShardIndex;
4675 : use utils::logging;
4676 :
4677 : use crate::deletion_queue::mock::MockDeletionQueue;
4678 : use crate::l0_flush::L0FlushConfig;
4679 : use crate::walredo::apply_neon;
4680 : use crate::{repository::Key, walrecord::NeonWalRecord};
4681 :
4682 : use super::*;
4683 : use hex_literal::hex;
4684 : use utils::id::TenantId;
4685 :
4686 : pub const TIMELINE_ID: TimelineId =
4687 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
4688 : pub const NEW_TIMELINE_ID: TimelineId =
4689 : TimelineId::from_array(hex!("AA223344556677881122334455667788"));
4690 :
4691 : /// Convenience function to create a page image with given string as the only content
4692 5028717 : pub fn test_img(s: &str) -> Bytes {
4693 5028717 : let mut buf = BytesMut::new();
4694 5028717 : buf.extend_from_slice(s.as_bytes());
4695 5028717 : buf.resize(64, 0);
4696 5028717 :
4697 5028717 : buf.freeze()
4698 5028717 : }
4699 :
4700 : impl From<TenantConf> for TenantConfOpt {
4701 186 : fn from(tenant_conf: TenantConf) -> Self {
4702 186 : Self {
4703 186 : checkpoint_distance: Some(tenant_conf.checkpoint_distance),
4704 186 : checkpoint_timeout: Some(tenant_conf.checkpoint_timeout),
4705 186 : compaction_target_size: Some(tenant_conf.compaction_target_size),
4706 186 : compaction_period: Some(tenant_conf.compaction_period),
4707 186 : compaction_threshold: Some(tenant_conf.compaction_threshold),
4708 186 : compaction_algorithm: Some(tenant_conf.compaction_algorithm),
4709 186 : gc_horizon: Some(tenant_conf.gc_horizon),
4710 186 : gc_period: Some(tenant_conf.gc_period),
4711 186 : image_creation_threshold: Some(tenant_conf.image_creation_threshold),
4712 186 : pitr_interval: Some(tenant_conf.pitr_interval),
4713 186 : walreceiver_connect_timeout: Some(tenant_conf.walreceiver_connect_timeout),
4714 186 : lagging_wal_timeout: Some(tenant_conf.lagging_wal_timeout),
4715 186 : max_lsn_wal_lag: Some(tenant_conf.max_lsn_wal_lag),
4716 186 : eviction_policy: Some(tenant_conf.eviction_policy),
4717 186 : min_resident_size_override: tenant_conf.min_resident_size_override,
4718 186 : evictions_low_residence_duration_metric_threshold: Some(
4719 186 : tenant_conf.evictions_low_residence_duration_metric_threshold,
4720 186 : ),
4721 186 : heatmap_period: Some(tenant_conf.heatmap_period),
4722 186 : lazy_slru_download: Some(tenant_conf.lazy_slru_download),
4723 186 : timeline_get_throttle: Some(tenant_conf.timeline_get_throttle),
4724 186 : image_layer_creation_check_threshold: Some(
4725 186 : tenant_conf.image_layer_creation_check_threshold,
4726 186 : ),
4727 186 : switch_aux_file_policy: Some(tenant_conf.switch_aux_file_policy),
4728 186 : lsn_lease_length: Some(tenant_conf.lsn_lease_length),
4729 186 : lsn_lease_length_for_ts: Some(tenant_conf.lsn_lease_length_for_ts),
4730 186 : }
4731 186 : }
4732 : }
4733 :
4734 : pub struct TenantHarness {
4735 : pub conf: &'static PageServerConf,
4736 : pub tenant_conf: TenantConf,
4737 : pub tenant_shard_id: TenantShardId,
4738 : pub generation: Generation,
4739 : pub shard: ShardIndex,
4740 : pub remote_storage: GenericRemoteStorage,
4741 : pub remote_fs_dir: Utf8PathBuf,
4742 : pub deletion_queue: MockDeletionQueue,
4743 : }
4744 :
4745 : static LOG_HANDLE: OnceCell<()> = OnceCell::new();
4746 :
4747 202 : pub(crate) fn setup_logging() {
4748 202 : LOG_HANDLE.get_or_init(|| {
4749 190 : logging::init(
4750 190 : logging::LogFormat::Test,
4751 190 : // enable it in case the tests exercise code paths that use
4752 190 : // debug_assert_current_span_has_tenant_and_timeline_id
4753 190 : logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
4754 190 : logging::Output::Stdout,
4755 190 : )
4756 190 : .expect("Failed to init test logging")
4757 202 : });
4758 202 : }
4759 :
4760 : impl TenantHarness {
4761 186 : pub async fn create_custom(
4762 186 : test_name: &'static str,
4763 186 : tenant_conf: TenantConf,
4764 186 : tenant_id: TenantId,
4765 186 : shard_identity: ShardIdentity,
4766 186 : generation: Generation,
4767 186 : ) -> anyhow::Result<Self> {
4768 186 : setup_logging();
4769 186 :
4770 186 : let repo_dir = PageServerConf::test_repo_dir(test_name);
4771 186 : let _ = fs::remove_dir_all(&repo_dir);
4772 186 : fs::create_dir_all(&repo_dir)?;
4773 :
4774 186 : let conf = PageServerConf::dummy_conf(repo_dir);
4775 186 : // Make a static copy of the config. This can never be free'd, but that's
4776 186 : // OK in a test.
4777 186 : let conf: &'static PageServerConf = Box::leak(Box::new(conf));
4778 186 :
4779 186 : let shard = shard_identity.shard_index();
4780 186 : let tenant_shard_id = TenantShardId {
4781 186 : tenant_id,
4782 186 : shard_number: shard.shard_number,
4783 186 : shard_count: shard.shard_count,
4784 186 : };
4785 186 : fs::create_dir_all(conf.tenant_path(&tenant_shard_id))?;
4786 186 : fs::create_dir_all(conf.timelines_path(&tenant_shard_id))?;
4787 :
4788 : use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
4789 186 : let remote_fs_dir = conf.workdir.join("localfs");
4790 186 : std::fs::create_dir_all(&remote_fs_dir).unwrap();
4791 186 : let config = RemoteStorageConfig {
4792 186 : storage: RemoteStorageKind::LocalFs {
4793 186 : local_path: remote_fs_dir.clone(),
4794 186 : },
4795 186 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
4796 186 : };
4797 186 : let remote_storage = GenericRemoteStorage::from_config(&config).await.unwrap();
4798 186 : let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()));
4799 186 :
4800 186 : Ok(Self {
4801 186 : conf,
4802 186 : tenant_conf,
4803 186 : tenant_shard_id,
4804 186 : generation,
4805 186 : shard,
4806 186 : remote_storage,
4807 186 : remote_fs_dir,
4808 186 : deletion_queue,
4809 186 : })
4810 186 : }
4811 :
4812 174 : pub async fn create(test_name: &'static str) -> anyhow::Result<Self> {
4813 174 : // Disable automatic GC and compaction to make the unit tests more deterministic.
4814 174 : // The tests perform them manually if needed.
4815 174 : let tenant_conf = TenantConf {
4816 174 : gc_period: Duration::ZERO,
4817 174 : compaction_period: Duration::ZERO,
4818 174 : ..TenantConf::default()
4819 174 : };
4820 174 : let tenant_id = TenantId::generate();
4821 174 : let shard = ShardIdentity::unsharded();
4822 174 : Self::create_custom(
4823 174 : test_name,
4824 174 : tenant_conf,
4825 174 : tenant_id,
4826 174 : shard,
4827 174 : Generation::new(0xdeadbeef),
4828 174 : )
4829 0 : .await
4830 174 : }
4831 :
4832 20 : pub fn span(&self) -> tracing::Span {
4833 20 : info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
4834 20 : }
4835 :
4836 186 : pub(crate) async fn load(&self) -> (Arc<Tenant>, RequestContext) {
4837 186 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
4838 186 : (
4839 186 : self.do_try_load(&ctx)
4840 912 : .await
4841 186 : .expect("failed to load test tenant"),
4842 186 : ctx,
4843 186 : )
4844 186 : }
4845 :
4846 186 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
4847 : pub(crate) async fn do_try_load(
4848 : &self,
4849 : ctx: &RequestContext,
4850 : ) -> anyhow::Result<Arc<Tenant>> {
4851 : let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
4852 :
4853 : let tenant = Arc::new(Tenant::new(
4854 : TenantState::Attaching,
4855 : self.conf,
4856 : AttachedTenantConf::try_from(LocationConf::attached_single(
4857 : TenantConfOpt::from(self.tenant_conf.clone()),
4858 : self.generation,
4859 : &ShardParameters::default(),
4860 : ))
4861 : .unwrap(),
4862 : // This is a legacy/test code path: sharding isn't supported here.
4863 : ShardIdentity::unsharded(),
4864 : Some(walredo_mgr),
4865 : self.tenant_shard_id,
4866 : self.remote_storage.clone(),
4867 : self.deletion_queue.new_client(),
4868 : // TODO: ideally we should run all unit tests with both configs
4869 : L0FlushGlobalState::new(L0FlushConfig::default()),
4870 : ));
4871 :
4872 : let preload = tenant
4873 : .preload(&self.remote_storage, CancellationToken::new())
4874 : .await?;
4875 : tenant.attach(Some(preload), ctx).await?;
4876 :
4877 : tenant.state.send_replace(TenantState::Active);
4878 : for timeline in tenant.timelines.lock().unwrap().values() {
4879 : timeline.set_state(TimelineState::Active);
4880 : }
4881 : Ok(tenant)
4882 : }
4883 :
4884 2 : pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
4885 2 : self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
4886 2 : }
4887 : }
4888 :
4889 : // Mock WAL redo manager that doesn't do much
4890 : pub(crate) struct TestRedoManager;
4891 :
4892 : impl TestRedoManager {
4893 : /// # Cancel-Safety
4894 : ///
4895 : /// This method is cancellation-safe.
4896 342 : pub async fn request_redo(
4897 342 : &self,
4898 342 : key: Key,
4899 342 : lsn: Lsn,
4900 342 : base_img: Option<(Lsn, Bytes)>,
4901 342 : records: Vec<(Lsn, NeonWalRecord)>,
4902 342 : _pg_version: u32,
4903 342 : ) -> Result<Bytes, walredo::Error> {
4904 458 : let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
4905 342 : if records_neon {
4906 : // For Neon wal records, we can decode without spawning postgres, so do so.
4907 342 : let mut page = match (base_img, records.first()) {
4908 338 : (Some((_lsn, img)), _) => {
4909 338 : let mut page = BytesMut::new();
4910 338 : page.extend_from_slice(&img);
4911 338 : page
4912 : }
4913 4 : (_, Some((_lsn, rec))) if rec.will_init() => BytesMut::new(),
4914 : _ => {
4915 0 : panic!("Neon WAL redo requires base image or will init record");
4916 : }
4917 : };
4918 :
4919 800 : for (record_lsn, record) in records {
4920 458 : apply_neon::apply_in_neon(&record, record_lsn, key, &mut page)?;
4921 : }
4922 342 : Ok(page.freeze())
4923 : } else {
4924 : // We never spawn a postgres walredo process in unit tests: just log what we might have done.
4925 0 : let s = format!(
4926 0 : "redo for {} to get to {}, with {} and {} records",
4927 0 : key,
4928 0 : lsn,
4929 0 : if base_img.is_some() {
4930 0 : "base image"
4931 : } else {
4932 0 : "no base image"
4933 : },
4934 0 : records.len()
4935 0 : );
4936 0 : println!("{s}");
4937 0 :
4938 0 : Ok(test_img(&s))
4939 : }
4940 342 : }
4941 : }
4942 : }
4943 :
4944 : #[cfg(test)]
4945 : mod tests {
4946 : use std::collections::{BTreeMap, BTreeSet};
4947 :
4948 : use super::*;
4949 : use crate::keyspace::KeySpaceAccum;
4950 : use crate::repository::{Key, Value};
4951 : use crate::tenant::harness::*;
4952 : use crate::tenant::timeline::CompactFlags;
4953 : use crate::walrecord::NeonWalRecord;
4954 : use crate::DEFAULT_PG_VERSION;
4955 : use bytes::{Bytes, BytesMut};
4956 : use hex_literal::hex;
4957 : use itertools::Itertools;
4958 : use pageserver_api::key::{AUX_KEY_PREFIX, NON_INHERITED_RANGE};
4959 : use pageserver_api::keyspace::KeySpace;
4960 : use pageserver_api::models::{CompactionAlgorithm, CompactionAlgorithmSettings};
4961 : use rand::{thread_rng, Rng};
4962 : use storage_layer::PersistentLayerKey;
4963 : use tests::storage_layer::ValuesReconstructState;
4964 : use tests::timeline::{GetVectoredError, ShutdownMode};
4965 : use timeline::compaction::{KeyHistoryRetention, KeyLogAtLsn};
4966 : use timeline::{DeltaLayerTestDesc, GcInfo};
4967 : use utils::id::TenantId;
4968 :
4969 : static TEST_KEY: Lazy<Key> =
4970 18 : Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));
4971 :
4972 : #[tokio::test]
4973 2 : async fn test_basic() -> anyhow::Result<()> {
4974 10 : let (tenant, ctx) = TenantHarness::create("test_basic").await?.load().await;
4975 2 : let tline = tenant
4976 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
4977 4 : .await?;
4978 2 :
4979 2 : let mut writer = tline.writer().await;
4980 2 : writer
4981 2 : .put(
4982 2 : *TEST_KEY,
4983 2 : Lsn(0x10),
4984 2 : &Value::Image(test_img("foo at 0x10")),
4985 2 : &ctx,
4986 2 : )
4987 2 : .await?;
4988 2 : writer.finish_write(Lsn(0x10));
4989 2 : drop(writer);
4990 2 :
4991 2 : let mut writer = tline.writer().await;
4992 2 : writer
4993 2 : .put(
4994 2 : *TEST_KEY,
4995 2 : Lsn(0x20),
4996 2 : &Value::Image(test_img("foo at 0x20")),
4997 2 : &ctx,
4998 2 : )
4999 2 : .await?;
5000 2 : writer.finish_write(Lsn(0x20));
5001 2 : drop(writer);
5002 2 :
5003 2 : assert_eq!(
5004 2 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
5005 2 : test_img("foo at 0x10")
5006 2 : );
5007 2 : assert_eq!(
5008 2 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
5009 2 : test_img("foo at 0x10")
5010 2 : );
5011 2 : assert_eq!(
5012 2 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
5013 2 : test_img("foo at 0x20")
5014 2 : );
5015 2 :
5016 2 : Ok(())
5017 2 : }
5018 :
5019 : #[tokio::test]
5020 2 : async fn no_duplicate_timelines() -> anyhow::Result<()> {
5021 2 : let (tenant, ctx) = TenantHarness::create("no_duplicate_timelines")
5022 2 : .await?
5023 2 : .load()
5024 10 : .await;
5025 2 : let _ = tenant
5026 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5027 4 : .await?;
5028 2 :
5029 2 : match tenant
5030 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5031 2 : .await
5032 2 : {
5033 2 : Ok(_) => panic!("duplicate timeline creation should fail"),
5034 2 : Err(e) => assert_eq!(e.to_string(), "Already exists".to_string()),
5035 2 : }
5036 2 :
5037 2 : Ok(())
5038 2 : }
5039 :
5040 : /// Convenience function to create a page image with given string as the only content
5041 10 : pub fn test_value(s: &str) -> Value {
5042 10 : let mut buf = BytesMut::new();
5043 10 : buf.extend_from_slice(s.as_bytes());
5044 10 : Value::Image(buf.freeze())
5045 10 : }
5046 :
5047 : ///
5048 : /// Test branch creation
5049 : ///
5050 : #[tokio::test]
5051 2 : async fn test_branch() -> anyhow::Result<()> {
5052 2 : use std::str::from_utf8;
5053 2 :
5054 7 : let (tenant, ctx) = TenantHarness::create("test_branch").await?.load().await;
5055 2 : let tline = tenant
5056 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5057 4 : .await?;
5058 2 : let mut writer = tline.writer().await;
5059 2 :
5060 2 : #[allow(non_snake_case)]
5061 2 : let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
5062 2 : #[allow(non_snake_case)]
5063 2 : let TEST_KEY_B: Key = Key::from_hex("110000000033333333444444445500000002").unwrap();
5064 2 :
5065 2 : // Insert a value on the timeline
5066 2 : writer
5067 2 : .put(TEST_KEY_A, Lsn(0x20), &test_value("foo at 0x20"), &ctx)
5068 2 : .await?;
5069 2 : writer
5070 2 : .put(TEST_KEY_B, Lsn(0x20), &test_value("foobar at 0x20"), &ctx)
5071 2 : .await?;
5072 2 : writer.finish_write(Lsn(0x20));
5073 2 :
5074 2 : writer
5075 2 : .put(TEST_KEY_A, Lsn(0x30), &test_value("foo at 0x30"), &ctx)
5076 2 : .await?;
5077 2 : writer.finish_write(Lsn(0x30));
5078 2 : writer
5079 2 : .put(TEST_KEY_A, Lsn(0x40), &test_value("foo at 0x40"), &ctx)
5080 2 : .await?;
5081 2 : writer.finish_write(Lsn(0x40));
5082 2 :
5083 2 : //assert_current_logical_size(&tline, Lsn(0x40));
5084 2 :
5085 2 : // Branch the history, modify relation differently on the new timeline
5086 2 : tenant
5087 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x30)), &ctx)
5088 2 : .await?;
5089 2 : let newtline = tenant
5090 2 : .get_timeline(NEW_TIMELINE_ID, true)
5091 2 : .expect("Should have a local timeline");
5092 2 : let mut new_writer = newtline.writer().await;
5093 2 : new_writer
5094 2 : .put(TEST_KEY_A, Lsn(0x40), &test_value("bar at 0x40"), &ctx)
5095 2 : .await?;
5096 2 : new_writer.finish_write(Lsn(0x40));
5097 2 :
5098 2 : // Check page contents on both branches
5099 2 : assert_eq!(
5100 2 : from_utf8(&tline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
5101 2 : "foo at 0x40"
5102 2 : );
5103 2 : assert_eq!(
5104 2 : from_utf8(&newtline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
5105 2 : "bar at 0x40"
5106 2 : );
5107 2 : assert_eq!(
5108 2 : from_utf8(&newtline.get(TEST_KEY_B, Lsn(0x40), &ctx).await?)?,
5109 2 : "foobar at 0x20"
5110 2 : );
5111 2 :
5112 2 : //assert_current_logical_size(&tline, Lsn(0x40));
5113 2 :
5114 2 : Ok(())
5115 2 : }
5116 :
5117 20 : async fn make_some_layers(
5118 20 : tline: &Timeline,
5119 20 : start_lsn: Lsn,
5120 20 : ctx: &RequestContext,
5121 20 : ) -> anyhow::Result<()> {
5122 20 : let mut lsn = start_lsn;
5123 : {
5124 20 : let mut writer = tline.writer().await;
5125 : // Create a relation on the timeline
5126 20 : writer
5127 20 : .put(
5128 20 : *TEST_KEY,
5129 20 : lsn,
5130 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5131 20 : ctx,
5132 20 : )
5133 10 : .await?;
5134 20 : writer.finish_write(lsn);
5135 20 : lsn += 0x10;
5136 20 : writer
5137 20 : .put(
5138 20 : *TEST_KEY,
5139 20 : lsn,
5140 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5141 20 : ctx,
5142 20 : )
5143 0 : .await?;
5144 20 : writer.finish_write(lsn);
5145 20 : lsn += 0x10;
5146 20 : }
5147 20 : tline.freeze_and_flush().await?;
5148 : {
5149 20 : let mut writer = tline.writer().await;
5150 20 : writer
5151 20 : .put(
5152 20 : *TEST_KEY,
5153 20 : lsn,
5154 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5155 20 : ctx,
5156 20 : )
5157 10 : .await?;
5158 20 : writer.finish_write(lsn);
5159 20 : lsn += 0x10;
5160 20 : writer
5161 20 : .put(
5162 20 : *TEST_KEY,
5163 20 : lsn,
5164 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5165 20 : ctx,
5166 20 : )
5167 0 : .await?;
5168 20 : writer.finish_write(lsn);
5169 20 : }
5170 20 : tline.freeze_and_flush().await.map_err(|e| e.into())
5171 20 : }
5172 :
5173 : #[tokio::test(start_paused = true)]
5174 2 : async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
5175 2 : let (tenant, ctx) =
5176 2 : TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
5177 2 : .await?
5178 2 : .load()
5179 9 : .await;
5180 2 : // Advance to the lsn lease deadline so that GC is not blocked by
5181 2 : // initial transition into AttachedSingle.
5182 2 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
5183 2 : tokio::time::resume();
5184 2 : let tline = tenant
5185 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5186 4 : .await?;
5187 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5188 2 :
5189 2 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
5190 2 : // FIXME: this doesn't actually remove any layer currently, given how the flushing
5191 2 : // and compaction works. But it does set the 'cutoff' point so that the cross check
5192 2 : // below should fail.
5193 2 : tenant
5194 2 : .gc_iteration(
5195 2 : Some(TIMELINE_ID),
5196 2 : 0x10,
5197 2 : Duration::ZERO,
5198 2 : &CancellationToken::new(),
5199 2 : &ctx,
5200 2 : )
5201 2 : .await?;
5202 2 :
5203 2 : // try to branch at lsn 25, should fail because we already garbage collected the data
5204 2 : match tenant
5205 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
5206 2 : .await
5207 2 : {
5208 2 : Ok(_) => panic!("branching should have failed"),
5209 2 : Err(err) => {
5210 2 : let CreateTimelineError::AncestorLsn(err) = err else {
5211 2 : panic!("wrong error type")
5212 2 : };
5213 2 : assert!(err.to_string().contains("invalid branch start lsn"));
5214 2 : assert!(err
5215 2 : .source()
5216 2 : .unwrap()
5217 2 : .to_string()
5218 2 : .contains("we might've already garbage collected needed data"))
5219 2 : }
5220 2 : }
5221 2 :
5222 2 : Ok(())
5223 2 : }
5224 :
5225 : #[tokio::test]
5226 2 : async fn test_prohibit_branch_creation_on_pre_initdb_lsn() -> anyhow::Result<()> {
5227 2 : let (tenant, ctx) =
5228 2 : TenantHarness::create("test_prohibit_branch_creation_on_pre_initdb_lsn")
5229 2 : .await?
5230 2 : .load()
5231 10 : .await;
5232 2 :
5233 2 : let tline = tenant
5234 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x50), DEFAULT_PG_VERSION, &ctx)
5235 4 : .await?;
5236 2 : // try to branch at lsn 0x25, should fail because initdb lsn is 0x50
5237 2 : match tenant
5238 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
5239 2 : .await
5240 2 : {
5241 2 : Ok(_) => panic!("branching should have failed"),
5242 2 : Err(err) => {
5243 2 : let CreateTimelineError::AncestorLsn(err) = err else {
5244 2 : panic!("wrong error type");
5245 2 : };
5246 2 : assert!(&err.to_string().contains("invalid branch start lsn"));
5247 2 : assert!(&err
5248 2 : .source()
5249 2 : .unwrap()
5250 2 : .to_string()
5251 2 : .contains("is earlier than latest GC cutoff"));
5252 2 : }
5253 2 : }
5254 2 :
5255 2 : Ok(())
5256 2 : }
5257 :
5258 : /*
5259 : // FIXME: This currently fails to error out. Calling GC doesn't currently
5260 : // remove the old value, we'd need to work a little harder
5261 : #[tokio::test]
5262 : async fn test_prohibit_get_for_garbage_collected_data() -> anyhow::Result<()> {
5263 : let repo =
5264 : RepoHarness::create("test_prohibit_get_for_garbage_collected_data")?
5265 : .load();
5266 :
5267 : let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION)?;
5268 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5269 :
5270 : repo.gc_iteration(Some(TIMELINE_ID), 0x10, Duration::ZERO)?;
5271 : let latest_gc_cutoff_lsn = tline.get_latest_gc_cutoff_lsn();
5272 : assert!(*latest_gc_cutoff_lsn > Lsn(0x25));
5273 : match tline.get(*TEST_KEY, Lsn(0x25)) {
5274 : Ok(_) => panic!("request for page should have failed"),
5275 : Err(err) => assert!(err.to_string().contains("not found at")),
5276 : }
5277 : Ok(())
5278 : }
5279 : */
5280 :
5281 : #[tokio::test]
5282 2 : async fn test_get_branchpoints_from_an_inactive_timeline() -> anyhow::Result<()> {
5283 2 : let (tenant, ctx) =
5284 2 : TenantHarness::create("test_get_branchpoints_from_an_inactive_timeline")
5285 2 : .await?
5286 2 : .load()
5287 10 : .await;
5288 2 : let tline = tenant
5289 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5290 4 : .await?;
5291 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5292 2 :
5293 2 : tenant
5294 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
5295 2 : .await?;
5296 2 : let newtline = tenant
5297 2 : .get_timeline(NEW_TIMELINE_ID, true)
5298 2 : .expect("Should have a local timeline");
5299 2 :
5300 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
5301 2 :
5302 2 : tline.set_broken("test".to_owned());
5303 2 :
5304 2 : tenant
5305 2 : .gc_iteration(
5306 2 : Some(TIMELINE_ID),
5307 2 : 0x10,
5308 2 : Duration::ZERO,
5309 2 : &CancellationToken::new(),
5310 2 : &ctx,
5311 2 : )
5312 2 : .await?;
5313 2 :
5314 2 : // The branchpoints should contain all timelines, even ones marked
5315 2 : // as Broken.
5316 2 : {
5317 2 : let branchpoints = &tline.gc_info.read().unwrap().retain_lsns;
5318 2 : assert_eq!(branchpoints.len(), 1);
5319 2 : assert_eq!(
5320 2 : branchpoints[0],
5321 2 : (Lsn(0x40), NEW_TIMELINE_ID, MaybeOffloaded::No)
5322 2 : );
5323 2 : }
5324 2 :
5325 2 : // You can read the key from the child branch even though the parent is
5326 2 : // Broken, as long as you don't need to access data from the parent.
5327 2 : assert_eq!(
5328 4 : newtline.get(*TEST_KEY, Lsn(0x70), &ctx).await?,
5329 2 : test_img(&format!("foo at {}", Lsn(0x70)))
5330 2 : );
5331 2 :
5332 2 : // This needs to traverse to the parent, and fails.
5333 2 : let err = newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await.unwrap_err();
5334 2 : assert!(
5335 2 : err.to_string().starts_with(&format!(
5336 2 : "bad state on timeline {}: Broken",
5337 2 : tline.timeline_id
5338 2 : )),
5339 2 : "{err}"
5340 2 : );
5341 2 :
5342 2 : Ok(())
5343 2 : }
5344 :
5345 : #[tokio::test]
5346 2 : async fn test_retain_data_in_parent_which_is_needed_for_child() -> anyhow::Result<()> {
5347 2 : let (tenant, ctx) =
5348 2 : TenantHarness::create("test_retain_data_in_parent_which_is_needed_for_child")
5349 2 : .await?
5350 2 : .load()
5351 10 : .await;
5352 2 : let tline = tenant
5353 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5354 4 : .await?;
5355 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5356 2 :
5357 2 : tenant
5358 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
5359 2 : .await?;
5360 2 : let newtline = tenant
5361 2 : .get_timeline(NEW_TIMELINE_ID, true)
5362 2 : .expect("Should have a local timeline");
5363 2 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
5364 2 : tenant
5365 2 : .gc_iteration(
5366 2 : Some(TIMELINE_ID),
5367 2 : 0x10,
5368 2 : Duration::ZERO,
5369 2 : &CancellationToken::new(),
5370 2 : &ctx,
5371 2 : )
5372 2 : .await?;
5373 4 : assert!(newtline.get(*TEST_KEY, Lsn(0x25), &ctx).await.is_ok());
5374 2 :
5375 2 : Ok(())
5376 2 : }
5377 : #[tokio::test]
5378 2 : async fn test_parent_keeps_data_forever_after_branching() -> anyhow::Result<()> {
5379 2 : let (tenant, ctx) = TenantHarness::create("test_parent_keeps_data_forever_after_branching")
5380 2 : .await?
5381 2 : .load()
5382 10 : .await;
5383 2 : let tline = tenant
5384 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5385 4 : .await?;
5386 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5387 2 :
5388 2 : tenant
5389 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
5390 2 : .await?;
5391 2 : let newtline = tenant
5392 2 : .get_timeline(NEW_TIMELINE_ID, true)
5393 2 : .expect("Should have a local timeline");
5394 2 :
5395 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
5396 2 :
5397 2 : // run gc on parent
5398 2 : tenant
5399 2 : .gc_iteration(
5400 2 : Some(TIMELINE_ID),
5401 2 : 0x10,
5402 2 : Duration::ZERO,
5403 2 : &CancellationToken::new(),
5404 2 : &ctx,
5405 2 : )
5406 2 : .await?;
5407 2 :
5408 2 : // Check that the data is still accessible on the branch.
5409 2 : assert_eq!(
5410 7 : newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
5411 2 : test_img(&format!("foo at {}", Lsn(0x40)))
5412 2 : );
5413 2 :
5414 2 : Ok(())
5415 2 : }
5416 :
5417 : #[tokio::test]
5418 2 : async fn timeline_load() -> anyhow::Result<()> {
5419 2 : const TEST_NAME: &str = "timeline_load";
5420 2 : let harness = TenantHarness::create(TEST_NAME).await?;
5421 2 : {
5422 10 : let (tenant, ctx) = harness.load().await;
5423 2 : let tline = tenant
5424 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x7000), DEFAULT_PG_VERSION, &ctx)
5425 4 : .await?;
5426 6 : make_some_layers(tline.as_ref(), Lsn(0x8000), &ctx).await?;
5427 2 : // so that all uploads finish & we can call harness.load() below again
5428 2 : tenant
5429 2 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
5430 2 : .instrument(harness.span())
5431 2 : .await
5432 2 : .ok()
5433 2 : .unwrap();
5434 2 : }
5435 2 :
5436 10 : let (tenant, _ctx) = harness.load().await;
5437 2 : tenant
5438 2 : .get_timeline(TIMELINE_ID, true)
5439 2 : .expect("cannot load timeline");
5440 2 :
5441 2 : Ok(())
5442 2 : }
5443 :
5444 : #[tokio::test]
5445 2 : async fn timeline_load_with_ancestor() -> anyhow::Result<()> {
5446 2 : const TEST_NAME: &str = "timeline_load_with_ancestor";
5447 2 : let harness = TenantHarness::create(TEST_NAME).await?;
5448 2 : // create two timelines
5449 2 : {
5450 10 : let (tenant, ctx) = harness.load().await;
5451 2 : let tline = tenant
5452 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5453 4 : .await?;
5454 2 :
5455 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5456 2 :
5457 2 : let child_tline = tenant
5458 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
5459 2 : .await?;
5460 2 : child_tline.set_state(TimelineState::Active);
5461 2 :
5462 2 : let newtline = tenant
5463 2 : .get_timeline(NEW_TIMELINE_ID, true)
5464 2 : .expect("Should have a local timeline");
5465 2 :
5466 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
5467 2 :
5468 2 : // so that all uploads finish & we can call harness.load() below again
5469 2 : tenant
5470 2 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
5471 2 : .instrument(harness.span())
5472 3 : .await
5473 2 : .ok()
5474 2 : .unwrap();
5475 2 : }
5476 2 :
5477 2 : // check that both of them are initially unloaded
5478 13 : let (tenant, _ctx) = harness.load().await;
5479 2 :
5480 2 : // check that both, child and ancestor are loaded
5481 2 : let _child_tline = tenant
5482 2 : .get_timeline(NEW_TIMELINE_ID, true)
5483 2 : .expect("cannot get child timeline loaded");
5484 2 :
5485 2 : let _ancestor_tline = tenant
5486 2 : .get_timeline(TIMELINE_ID, true)
5487 2 : .expect("cannot get ancestor timeline loaded");
5488 2 :
5489 2 : Ok(())
5490 2 : }
5491 :
5492 : #[tokio::test]
5493 2 : async fn delta_layer_dumping() -> anyhow::Result<()> {
5494 2 : use storage_layer::AsLayerDesc;
5495 2 : let (tenant, ctx) = TenantHarness::create("test_layer_dumping")
5496 2 : .await?
5497 2 : .load()
5498 10 : .await;
5499 2 : let tline = tenant
5500 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5501 4 : .await?;
5502 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5503 2 :
5504 2 : let layer_map = tline.layers.read().await;
5505 2 : let level0_deltas = layer_map
5506 2 : .layer_map()?
5507 2 : .level0_deltas()
5508 2 : .iter()
5509 4 : .map(|desc| layer_map.get_from_desc(desc))
5510 2 : .collect::<Vec<_>>();
5511 2 :
5512 2 : assert!(!level0_deltas.is_empty());
5513 2 :
5514 6 : for delta in level0_deltas {
5515 2 : // Ensure we are dumping a delta layer here
5516 4 : assert!(delta.layer_desc().is_delta);
5517 8 : delta.dump(true, &ctx).await.unwrap();
5518 2 : }
5519 2 :
5520 2 : Ok(())
5521 2 : }
5522 :
5523 : #[tokio::test]
5524 2 : async fn test_images() -> anyhow::Result<()> {
5525 9 : let (tenant, ctx) = TenantHarness::create("test_images").await?.load().await;
5526 2 : let tline = tenant
5527 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5528 4 : .await?;
5529 2 :
5530 2 : let mut writer = tline.writer().await;
5531 2 : writer
5532 2 : .put(
5533 2 : *TEST_KEY,
5534 2 : Lsn(0x10),
5535 2 : &Value::Image(test_img("foo at 0x10")),
5536 2 : &ctx,
5537 2 : )
5538 2 : .await?;
5539 2 : writer.finish_write(Lsn(0x10));
5540 2 : drop(writer);
5541 2 :
5542 2 : tline.freeze_and_flush().await?;
5543 2 : tline
5544 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
5545 2 : .await?;
5546 2 :
5547 2 : let mut writer = tline.writer().await;
5548 2 : writer
5549 2 : .put(
5550 2 : *TEST_KEY,
5551 2 : Lsn(0x20),
5552 2 : &Value::Image(test_img("foo at 0x20")),
5553 2 : &ctx,
5554 2 : )
5555 2 : .await?;
5556 2 : writer.finish_write(Lsn(0x20));
5557 2 : drop(writer);
5558 2 :
5559 2 : tline.freeze_and_flush().await?;
5560 2 : tline
5561 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
5562 2 : .await?;
5563 2 :
5564 2 : let mut writer = tline.writer().await;
5565 2 : writer
5566 2 : .put(
5567 2 : *TEST_KEY,
5568 2 : Lsn(0x30),
5569 2 : &Value::Image(test_img("foo at 0x30")),
5570 2 : &ctx,
5571 2 : )
5572 2 : .await?;
5573 2 : writer.finish_write(Lsn(0x30));
5574 2 : drop(writer);
5575 2 :
5576 2 : tline.freeze_and_flush().await?;
5577 2 : tline
5578 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
5579 2 : .await?;
5580 2 :
5581 2 : let mut writer = tline.writer().await;
5582 2 : writer
5583 2 : .put(
5584 2 : *TEST_KEY,
5585 2 : Lsn(0x40),
5586 2 : &Value::Image(test_img("foo at 0x40")),
5587 2 : &ctx,
5588 2 : )
5589 2 : .await?;
5590 2 : writer.finish_write(Lsn(0x40));
5591 2 : drop(writer);
5592 2 :
5593 2 : tline.freeze_and_flush().await?;
5594 2 : tline
5595 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
5596 2 : .await?;
5597 2 :
5598 2 : assert_eq!(
5599 4 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
5600 2 : test_img("foo at 0x10")
5601 2 : );
5602 2 : assert_eq!(
5603 4 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
5604 2 : test_img("foo at 0x10")
5605 2 : );
5606 2 : assert_eq!(
5607 2 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
5608 2 : test_img("foo at 0x20")
5609 2 : );
5610 2 : assert_eq!(
5611 4 : tline.get(*TEST_KEY, Lsn(0x30), &ctx).await?,
5612 2 : test_img("foo at 0x30")
5613 2 : );
5614 2 : assert_eq!(
5615 4 : tline.get(*TEST_KEY, Lsn(0x40), &ctx).await?,
5616 2 : test_img("foo at 0x40")
5617 2 : );
5618 2 :
5619 2 : Ok(())
5620 2 : }
5621 :
5622 4 : async fn bulk_insert_compact_gc(
5623 4 : tenant: &Tenant,
5624 4 : timeline: &Arc<Timeline>,
5625 4 : ctx: &RequestContext,
5626 4 : lsn: Lsn,
5627 4 : repeat: usize,
5628 4 : key_count: usize,
5629 4 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
5630 4 : let compact = true;
5631 40718 : bulk_insert_maybe_compact_gc(tenant, timeline, ctx, lsn, repeat, key_count, compact).await
5632 4 : }
5633 :
5634 8 : async fn bulk_insert_maybe_compact_gc(
5635 8 : tenant: &Tenant,
5636 8 : timeline: &Arc<Timeline>,
5637 8 : ctx: &RequestContext,
5638 8 : mut lsn: Lsn,
5639 8 : repeat: usize,
5640 8 : key_count: usize,
5641 8 : compact: bool,
5642 8 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
5643 8 : let mut inserted: HashMap<Key, BTreeSet<Lsn>> = Default::default();
5644 8 :
5645 8 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5646 8 : let mut blknum = 0;
5647 8 :
5648 8 : // Enforce that key range is monotonously increasing
5649 8 : let mut keyspace = KeySpaceAccum::new();
5650 8 :
5651 8 : let cancel = CancellationToken::new();
5652 8 :
5653 8 : for _ in 0..repeat {
5654 400 : for _ in 0..key_count {
5655 4000000 : test_key.field6 = blknum;
5656 4000000 : let mut writer = timeline.writer().await;
5657 4000000 : writer
5658 4000000 : .put(
5659 4000000 : test_key,
5660 4000000 : lsn,
5661 4000000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5662 4000000 : ctx,
5663 4000000 : )
5664 3426 : .await?;
5665 4000000 : inserted.entry(test_key).or_default().insert(lsn);
5666 4000000 : writer.finish_write(lsn);
5667 4000000 : drop(writer);
5668 4000000 :
5669 4000000 : keyspace.add_key(test_key);
5670 4000000 :
5671 4000000 : lsn = Lsn(lsn.0 + 0x10);
5672 4000000 : blknum += 1;
5673 : }
5674 :
5675 400 : timeline.freeze_and_flush().await?;
5676 400 : if compact {
5677 : // this requires timeline to be &Arc<Timeline>
5678 8618 : timeline.compact(&cancel, EnumSet::empty(), ctx).await?;
5679 200 : }
5680 :
5681 : // this doesn't really need to use the timeline_id target, but it is closer to what it
5682 : // originally was.
5683 400 : let res = tenant
5684 400 : .gc_iteration(Some(timeline.timeline_id), 0, Duration::ZERO, &cancel, ctx)
5685 0 : .await?;
5686 :
5687 400 : assert_eq!(res.layers_removed, 0, "this never removes anything");
5688 : }
5689 :
5690 8 : Ok(inserted)
5691 8 : }
5692 :
5693 : //
5694 : // Insert 1000 key-value pairs with increasing keys, flush, compact, GC.
5695 : // Repeat 50 times.
5696 : //
5697 : #[tokio::test]
5698 2 : async fn test_bulk_insert() -> anyhow::Result<()> {
5699 2 : let harness = TenantHarness::create("test_bulk_insert").await?;
5700 9 : let (tenant, ctx) = harness.load().await;
5701 2 : let tline = tenant
5702 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5703 4 : .await?;
5704 2 :
5705 2 : let lsn = Lsn(0x10);
5706 20359 : bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
5707 2 :
5708 2 : Ok(())
5709 2 : }
5710 :
5711 : // Test the vectored get real implementation against a simple sequential implementation.
5712 : //
5713 : // The test generates a keyspace by repeatedly flushing the in-memory layer and compacting.
5714 : // Projected to 2D the key space looks like below. Lsn grows upwards on the Y axis and keys
5715 : // grow to the right on the X axis.
5716 : // [Delta]
5717 : // [Delta]
5718 : // [Delta]
5719 : // [Delta]
5720 : // ------------ Image ---------------
5721 : //
5722 : // After layer generation we pick the ranges to query as follows:
5723 : // 1. The beginning of each delta layer
5724 : // 2. At the seam between two adjacent delta layers
5725 : //
5726 : // There's one major downside to this test: delta layers only contains images,
5727 : // so the search can stop at the first delta layer and doesn't traverse any deeper.
5728 : #[tokio::test]
5729 2 : async fn test_get_vectored() -> anyhow::Result<()> {
5730 2 : let harness = TenantHarness::create("test_get_vectored").await?;
5731 10 : let (tenant, ctx) = harness.load().await;
5732 2 : let tline = tenant
5733 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5734 4 : .await?;
5735 2 :
5736 2 : let lsn = Lsn(0x10);
5737 20359 : let inserted = bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
5738 2 :
5739 2 : let guard = tline.layers.read().await;
5740 2 : let lm = guard.layer_map()?;
5741 2 :
5742 2 : lm.dump(true, &ctx).await?;
5743 2 :
5744 2 : let mut reads = Vec::new();
5745 2 : let mut prev = None;
5746 12 : lm.iter_historic_layers().for_each(|desc| {
5747 12 : if !desc.is_delta() {
5748 2 : prev = Some(desc.clone());
5749 2 : return;
5750 10 : }
5751 10 :
5752 10 : let start = desc.key_range.start;
5753 10 : let end = desc
5754 10 : .key_range
5755 10 : .start
5756 10 : .add(Timeline::MAX_GET_VECTORED_KEYS.try_into().unwrap());
5757 10 : reads.push(KeySpace {
5758 10 : ranges: vec![start..end],
5759 10 : });
5760 2 :
5761 10 : if let Some(prev) = &prev {
5762 10 : if !prev.is_delta() {
5763 10 : return;
5764 2 : }
5765 0 :
5766 0 : let first_range = Key {
5767 0 : field6: prev.key_range.end.field6 - 4,
5768 0 : ..prev.key_range.end
5769 0 : }..prev.key_range.end;
5770 0 :
5771 0 : let second_range = desc.key_range.start..Key {
5772 0 : field6: desc.key_range.start.field6 + 4,
5773 0 : ..desc.key_range.start
5774 0 : };
5775 0 :
5776 0 : reads.push(KeySpace {
5777 0 : ranges: vec![first_range, second_range],
5778 0 : });
5779 2 : };
5780 2 :
5781 2 : prev = Some(desc.clone());
5782 12 : });
5783 2 :
5784 2 : drop(guard);
5785 2 :
5786 2 : // Pick a big LSN such that we query over all the changes.
5787 2 : let reads_lsn = Lsn(u64::MAX - 1);
5788 2 :
5789 12 : for read in reads {
5790 10 : info!("Doing vectored read on {:?}", read);
5791 2 :
5792 10 : let vectored_res = tline
5793 10 : .get_vectored_impl(
5794 10 : read.clone(),
5795 10 : reads_lsn,
5796 10 : &mut ValuesReconstructState::new(),
5797 10 : &ctx,
5798 10 : )
5799 25 : .await;
5800 2 :
5801 10 : let mut expected_lsns: HashMap<Key, Lsn> = Default::default();
5802 10 : let mut expect_missing = false;
5803 10 : let mut key = read.start().unwrap();
5804 330 : while key != read.end().unwrap() {
5805 320 : if let Some(lsns) = inserted.get(&key) {
5806 320 : let expected_lsn = lsns.iter().rfind(|lsn| **lsn <= reads_lsn);
5807 320 : match expected_lsn {
5808 320 : Some(lsn) => {
5809 320 : expected_lsns.insert(key, *lsn);
5810 320 : }
5811 2 : None => {
5812 2 : expect_missing = true;
5813 0 : break;
5814 2 : }
5815 2 : }
5816 2 : } else {
5817 2 : expect_missing = true;
5818 0 : break;
5819 2 : }
5820 2 :
5821 320 : key = key.next();
5822 2 : }
5823 2 :
5824 10 : if expect_missing {
5825 2 : assert!(matches!(vectored_res, Err(GetVectoredError::MissingKey(_))));
5826 2 : } else {
5827 320 : for (key, image) in vectored_res? {
5828 320 : let expected_lsn = expected_lsns.get(&key).expect("determined above");
5829 320 : let expected_image = test_img(&format!("{} at {}", key.field6, expected_lsn));
5830 320 : assert_eq!(image?, expected_image);
5831 2 : }
5832 2 : }
5833 2 : }
5834 2 :
5835 2 : Ok(())
5836 2 : }
5837 :
5838 : #[tokio::test]
5839 2 : async fn test_get_vectored_aux_files() -> anyhow::Result<()> {
5840 2 : let harness = TenantHarness::create("test_get_vectored_aux_files").await?;
5841 2 :
5842 10 : let (tenant, ctx) = harness.load().await;
5843 2 : let tline = tenant
5844 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
5845 2 : .await?;
5846 2 : let tline = tline.raw_timeline().unwrap();
5847 2 :
5848 2 : let mut modification = tline.begin_modification(Lsn(0x1000));
5849 2 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
5850 2 : modification.set_lsn(Lsn(0x1008))?;
5851 2 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
5852 2 : modification.commit(&ctx).await?;
5853 2 :
5854 2 : let child_timeline_id = TimelineId::generate();
5855 2 : tenant
5856 2 : .branch_timeline_test(
5857 2 : tline,
5858 2 : child_timeline_id,
5859 2 : Some(tline.get_last_record_lsn()),
5860 2 : &ctx,
5861 2 : )
5862 2 : .await?;
5863 2 :
5864 2 : let child_timeline = tenant
5865 2 : .get_timeline(child_timeline_id, true)
5866 2 : .expect("Should have the branched timeline");
5867 2 :
5868 2 : let aux_keyspace = KeySpace {
5869 2 : ranges: vec![NON_INHERITED_RANGE],
5870 2 : };
5871 2 : let read_lsn = child_timeline.get_last_record_lsn();
5872 2 :
5873 2 : let vectored_res = child_timeline
5874 2 : .get_vectored_impl(
5875 2 : aux_keyspace.clone(),
5876 2 : read_lsn,
5877 2 : &mut ValuesReconstructState::new(),
5878 2 : &ctx,
5879 2 : )
5880 2 : .await;
5881 2 :
5882 2 : let images = vectored_res?;
5883 2 : assert!(images.is_empty());
5884 2 : Ok(())
5885 2 : }
5886 :
5887 : // Test that vectored get handles layer gaps correctly
5888 : // by advancing into the next ancestor timeline if required.
5889 : //
5890 : // The test generates timelines that look like the diagram below.
5891 : // We leave a gap in one of the L1 layers at `gap_at_key` (`/` in the diagram).
5892 : // The reconstruct data for that key lies in the ancestor timeline (`X` in the diagram).
5893 : //
5894 : // ```
5895 : //-------------------------------+
5896 : // ... |
5897 : // [ L1 ] |
5898 : // [ / L1 ] | Child Timeline
5899 : // ... |
5900 : // ------------------------------+
5901 : // [ X L1 ] | Parent Timeline
5902 : // ------------------------------+
5903 : // ```
5904 : #[tokio::test]
5905 2 : async fn test_get_vectored_key_gap() -> anyhow::Result<()> {
5906 2 : let tenant_conf = TenantConf {
5907 2 : // Make compaction deterministic
5908 2 : gc_period: Duration::ZERO,
5909 2 : compaction_period: Duration::ZERO,
5910 2 : // Encourage creation of L1 layers
5911 2 : checkpoint_distance: 16 * 1024,
5912 2 : compaction_target_size: 8 * 1024,
5913 2 : ..TenantConf::default()
5914 2 : };
5915 2 :
5916 2 : let harness = TenantHarness::create_custom(
5917 2 : "test_get_vectored_key_gap",
5918 2 : tenant_conf,
5919 2 : TenantId::generate(),
5920 2 : ShardIdentity::unsharded(),
5921 2 : Generation::new(0xdeadbeef),
5922 2 : )
5923 2 : .await?;
5924 10 : let (tenant, ctx) = harness.load().await;
5925 2 :
5926 2 : let mut current_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5927 2 : let gap_at_key = current_key.add(100);
5928 2 : let mut current_lsn = Lsn(0x10);
5929 2 :
5930 2 : const KEY_COUNT: usize = 10_000;
5931 2 :
5932 2 : let timeline_id = TimelineId::generate();
5933 2 : let current_timeline = tenant
5934 2 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
5935 4 : .await?;
5936 2 :
5937 2 : current_lsn += 0x100;
5938 2 :
5939 2 : let mut writer = current_timeline.writer().await;
5940 2 : writer
5941 2 : .put(
5942 2 : gap_at_key,
5943 2 : current_lsn,
5944 2 : &Value::Image(test_img(&format!("{} at {}", gap_at_key, current_lsn))),
5945 2 : &ctx,
5946 2 : )
5947 2 : .await?;
5948 2 : writer.finish_write(current_lsn);
5949 2 : drop(writer);
5950 2 :
5951 2 : let mut latest_lsns = HashMap::new();
5952 2 : latest_lsns.insert(gap_at_key, current_lsn);
5953 2 :
5954 2 : current_timeline.freeze_and_flush().await?;
5955 2 :
5956 2 : let child_timeline_id = TimelineId::generate();
5957 2 :
5958 2 : tenant
5959 2 : .branch_timeline_test(
5960 2 : ¤t_timeline,
5961 2 : child_timeline_id,
5962 2 : Some(current_lsn),
5963 2 : &ctx,
5964 2 : )
5965 2 : .await?;
5966 2 : let child_timeline = tenant
5967 2 : .get_timeline(child_timeline_id, true)
5968 2 : .expect("Should have the branched timeline");
5969 2 :
5970 20002 : for i in 0..KEY_COUNT {
5971 20000 : if current_key == gap_at_key {
5972 2 : current_key = current_key.next();
5973 2 : continue;
5974 19998 : }
5975 19998 :
5976 19998 : current_lsn += 0x10;
5977 2 :
5978 19998 : let mut writer = child_timeline.writer().await;
5979 19998 : writer
5980 19998 : .put(
5981 19998 : current_key,
5982 19998 : current_lsn,
5983 19998 : &Value::Image(test_img(&format!("{} at {}", current_key, current_lsn))),
5984 19998 : &ctx,
5985 19998 : )
5986 66 : .await?;
5987 19998 : writer.finish_write(current_lsn);
5988 19998 : drop(writer);
5989 19998 :
5990 19998 : latest_lsns.insert(current_key, current_lsn);
5991 19998 : current_key = current_key.next();
5992 19998 :
5993 19998 : // Flush every now and then to encourage layer file creation.
5994 19998 : if i % 500 == 0 {
5995 45 : child_timeline.freeze_and_flush().await?;
5996 19958 : }
5997 2 : }
5998 2 :
5999 2 : child_timeline.freeze_and_flush().await?;
6000 2 : let mut flags = EnumSet::new();
6001 2 : flags.insert(CompactFlags::ForceRepartition);
6002 2 : child_timeline
6003 2 : .compact(&CancellationToken::new(), flags, &ctx)
6004 1757 : .await?;
6005 2 :
6006 2 : let key_near_end = {
6007 2 : let mut tmp = current_key;
6008 2 : tmp.field6 -= 10;
6009 2 : tmp
6010 2 : };
6011 2 :
6012 2 : let key_near_gap = {
6013 2 : let mut tmp = gap_at_key;
6014 2 : tmp.field6 -= 10;
6015 2 : tmp
6016 2 : };
6017 2 :
6018 2 : let read = KeySpace {
6019 2 : ranges: vec![key_near_gap..gap_at_key.next(), key_near_end..current_key],
6020 2 : };
6021 2 : let results = child_timeline
6022 2 : .get_vectored_impl(
6023 2 : read.clone(),
6024 2 : current_lsn,
6025 2 : &mut ValuesReconstructState::new(),
6026 2 : &ctx,
6027 2 : )
6028 16 : .await?;
6029 2 :
6030 44 : for (key, img_res) in results {
6031 42 : let expected = test_img(&format!("{} at {}", key, latest_lsns[&key]));
6032 42 : assert_eq!(img_res?, expected);
6033 2 : }
6034 2 :
6035 2 : Ok(())
6036 2 : }
6037 :
6038 : // Test that vectored get descends into ancestor timelines correctly and
6039 : // does not return an image that's newer than requested.
6040 : //
6041 : // The diagram below ilustrates an interesting case. We have a parent timeline
6042 : // (top of the Lsn range) and a child timeline. The request key cannot be reconstructed
6043 : // from the child timeline, so the parent timeline must be visited. When advacing into
6044 : // the child timeline, the read path needs to remember what the requested Lsn was in
6045 : // order to avoid returning an image that's too new. The test below constructs such
6046 : // a timeline setup and does a few queries around the Lsn of each page image.
6047 : // ```
6048 : // LSN
6049 : // ^
6050 : // |
6051 : // |
6052 : // 500 | --------------------------------------> branch point
6053 : // 400 | X
6054 : // 300 | X
6055 : // 200 | --------------------------------------> requested lsn
6056 : // 100 | X
6057 : // |---------------------------------------> Key
6058 : // |
6059 : // ------> requested key
6060 : //
6061 : // Legend:
6062 : // * X - page images
6063 : // ```
6064 : #[tokio::test]
6065 2 : async fn test_get_vectored_ancestor_descent() -> anyhow::Result<()> {
6066 2 : let harness = TenantHarness::create("test_get_vectored_on_lsn_axis").await?;
6067 10 : let (tenant, ctx) = harness.load().await;
6068 2 :
6069 2 : let start_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6070 2 : let end_key = start_key.add(1000);
6071 2 : let child_gap_at_key = start_key.add(500);
6072 2 : let mut parent_gap_lsns: BTreeMap<Lsn, String> = BTreeMap::new();
6073 2 :
6074 2 : let mut current_lsn = Lsn(0x10);
6075 2 :
6076 2 : let timeline_id = TimelineId::generate();
6077 2 : let parent_timeline = tenant
6078 2 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
6079 4 : .await?;
6080 2 :
6081 2 : current_lsn += 0x100;
6082 2 :
6083 8 : for _ in 0..3 {
6084 6 : let mut key = start_key;
6085 6006 : while key < end_key {
6086 6000 : current_lsn += 0x10;
6087 6000 :
6088 6000 : let image_value = format!("{} at {}", child_gap_at_key, current_lsn);
6089 2 :
6090 6000 : let mut writer = parent_timeline.writer().await;
6091 6000 : writer
6092 6000 : .put(
6093 6000 : key,
6094 6000 : current_lsn,
6095 6000 : &Value::Image(test_img(&image_value)),
6096 6000 : &ctx,
6097 6000 : )
6098 6 : .await?;
6099 6000 : writer.finish_write(current_lsn);
6100 6000 :
6101 6000 : if key == child_gap_at_key {
6102 6 : parent_gap_lsns.insert(current_lsn, image_value);
6103 5994 : }
6104 2 :
6105 6000 : key = key.next();
6106 2 : }
6107 2 :
6108 6 : parent_timeline.freeze_and_flush().await?;
6109 2 : }
6110 2 :
6111 2 : let child_timeline_id = TimelineId::generate();
6112 2 :
6113 2 : let child_timeline = tenant
6114 2 : .branch_timeline_test(&parent_timeline, child_timeline_id, Some(current_lsn), &ctx)
6115 2 : .await?;
6116 2 :
6117 2 : let mut key = start_key;
6118 2002 : while key < end_key {
6119 2000 : if key == child_gap_at_key {
6120 2 : key = key.next();
6121 2 : continue;
6122 1998 : }
6123 1998 :
6124 1998 : current_lsn += 0x10;
6125 2 :
6126 1998 : let mut writer = child_timeline.writer().await;
6127 1998 : writer
6128 1998 : .put(
6129 1998 : key,
6130 1998 : current_lsn,
6131 1998 : &Value::Image(test_img(&format!("{} at {}", key, current_lsn))),
6132 1998 : &ctx,
6133 1998 : )
6134 17 : .await?;
6135 1998 : writer.finish_write(current_lsn);
6136 1998 :
6137 1998 : key = key.next();
6138 2 : }
6139 2 :
6140 2 : child_timeline.freeze_and_flush().await?;
6141 2 :
6142 2 : let lsn_offsets: [i64; 5] = [-10, -1, 0, 1, 10];
6143 2 : let mut query_lsns = Vec::new();
6144 6 : for image_lsn in parent_gap_lsns.keys().rev() {
6145 36 : for offset in lsn_offsets {
6146 30 : query_lsns.push(Lsn(image_lsn
6147 30 : .0
6148 30 : .checked_add_signed(offset)
6149 30 : .expect("Shouldn't overflow")));
6150 30 : }
6151 2 : }
6152 2 :
6153 32 : for query_lsn in query_lsns {
6154 30 : let results = child_timeline
6155 30 : .get_vectored_impl(
6156 30 : KeySpace {
6157 30 : ranges: vec![child_gap_at_key..child_gap_at_key.next()],
6158 30 : },
6159 30 : query_lsn,
6160 30 : &mut ValuesReconstructState::new(),
6161 30 : &ctx,
6162 30 : )
6163 29 : .await;
6164 2 :
6165 30 : let expected_item = parent_gap_lsns
6166 30 : .iter()
6167 30 : .rev()
6168 68 : .find(|(lsn, _)| **lsn <= query_lsn);
6169 30 :
6170 30 : info!(
6171 2 : "Doing vectored read at LSN {}. Expecting image to be: {:?}",
6172 2 : query_lsn, expected_item
6173 2 : );
6174 2 :
6175 30 : match expected_item {
6176 26 : Some((_, img_value)) => {
6177 26 : let key_results = results.expect("No vectored get error expected");
6178 26 : let key_result = &key_results[&child_gap_at_key];
6179 26 : let returned_img = key_result
6180 26 : .as_ref()
6181 26 : .expect("No page reconstruct error expected");
6182 26 :
6183 26 : info!(
6184 2 : "Vectored read at LSN {} returned image {}",
6185 0 : query_lsn,
6186 0 : std::str::from_utf8(returned_img)?
6187 2 : );
6188 26 : assert_eq!(*returned_img, test_img(img_value));
6189 2 : }
6190 2 : None => {
6191 4 : assert!(matches!(results, Err(GetVectoredError::MissingKey(_))));
6192 2 : }
6193 2 : }
6194 2 : }
6195 2 :
6196 2 : Ok(())
6197 2 : }
6198 :
6199 : #[tokio::test]
6200 2 : async fn test_random_updates() -> anyhow::Result<()> {
6201 2 : let names_algorithms = [
6202 2 : ("test_random_updates_legacy", CompactionAlgorithm::Legacy),
6203 2 : ("test_random_updates_tiered", CompactionAlgorithm::Tiered),
6204 2 : ];
6205 6 : for (name, algorithm) in names_algorithms {
6206 96146 : test_random_updates_algorithm(name, algorithm).await?;
6207 2 : }
6208 2 : Ok(())
6209 2 : }
6210 :
6211 4 : async fn test_random_updates_algorithm(
6212 4 : name: &'static str,
6213 4 : compaction_algorithm: CompactionAlgorithm,
6214 4 : ) -> anyhow::Result<()> {
6215 4 : let mut harness = TenantHarness::create(name).await?;
6216 4 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
6217 4 : kind: compaction_algorithm,
6218 4 : };
6219 19 : let (tenant, ctx) = harness.load().await;
6220 4 : let tline = tenant
6221 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6222 7 : .await?;
6223 :
6224 : const NUM_KEYS: usize = 1000;
6225 4 : let cancel = CancellationToken::new();
6226 4 :
6227 4 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6228 4 : let mut test_key_end = test_key;
6229 4 : test_key_end.field6 = NUM_KEYS as u32;
6230 4 : tline.add_extra_test_dense_keyspace(KeySpace::single(test_key..test_key_end));
6231 4 :
6232 4 : let mut keyspace = KeySpaceAccum::new();
6233 4 :
6234 4 : // Track when each page was last modified. Used to assert that
6235 4 : // a read sees the latest page version.
6236 4 : let mut updated = [Lsn(0); NUM_KEYS];
6237 4 :
6238 4 : let mut lsn = Lsn(0x10);
6239 : #[allow(clippy::needless_range_loop)]
6240 4004 : for blknum in 0..NUM_KEYS {
6241 4000 : lsn = Lsn(lsn.0 + 0x10);
6242 4000 : test_key.field6 = blknum as u32;
6243 4000 : let mut writer = tline.writer().await;
6244 4000 : writer
6245 4000 : .put(
6246 4000 : test_key,
6247 4000 : lsn,
6248 4000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6249 4000 : &ctx,
6250 4000 : )
6251 4 : .await?;
6252 4000 : writer.finish_write(lsn);
6253 4000 : updated[blknum] = lsn;
6254 4000 : drop(writer);
6255 4000 :
6256 4000 : keyspace.add_key(test_key);
6257 : }
6258 :
6259 204 : for _ in 0..50 {
6260 200200 : for _ in 0..NUM_KEYS {
6261 200000 : lsn = Lsn(lsn.0 + 0x10);
6262 200000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6263 200000 : test_key.field6 = blknum as u32;
6264 200000 : let mut writer = tline.writer().await;
6265 200000 : writer
6266 200000 : .put(
6267 200000 : test_key,
6268 200000 : lsn,
6269 200000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6270 200000 : &ctx,
6271 200000 : )
6272 198 : .await?;
6273 200000 : writer.finish_write(lsn);
6274 200000 : drop(writer);
6275 200000 : updated[blknum] = lsn;
6276 : }
6277 :
6278 : // Read all the blocks
6279 200000 : for (blknum, last_lsn) in updated.iter().enumerate() {
6280 200000 : test_key.field6 = blknum as u32;
6281 200000 : assert_eq!(
6282 200000 : tline.get(test_key, lsn, &ctx).await?,
6283 200000 : test_img(&format!("{} at {}", blknum, last_lsn))
6284 : );
6285 : }
6286 :
6287 : // Perform a cycle of flush, and GC
6288 205 : tline.freeze_and_flush().await?;
6289 200 : tenant
6290 200 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
6291 0 : .await?;
6292 : }
6293 :
6294 4 : Ok(())
6295 4 : }
6296 :
6297 : #[tokio::test]
6298 2 : async fn test_traverse_branches() -> anyhow::Result<()> {
6299 2 : let (tenant, ctx) = TenantHarness::create("test_traverse_branches")
6300 2 : .await?
6301 2 : .load()
6302 5 : .await;
6303 2 : let mut tline = tenant
6304 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6305 4 : .await?;
6306 2 :
6307 2 : const NUM_KEYS: usize = 1000;
6308 2 :
6309 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6310 2 :
6311 2 : let mut keyspace = KeySpaceAccum::new();
6312 2 :
6313 2 : let cancel = CancellationToken::new();
6314 2 :
6315 2 : // Track when each page was last modified. Used to assert that
6316 2 : // a read sees the latest page version.
6317 2 : let mut updated = [Lsn(0); NUM_KEYS];
6318 2 :
6319 2 : let mut lsn = Lsn(0x10);
6320 2 : #[allow(clippy::needless_range_loop)]
6321 2002 : for blknum in 0..NUM_KEYS {
6322 2000 : lsn = Lsn(lsn.0 + 0x10);
6323 2000 : test_key.field6 = blknum as u32;
6324 2000 : let mut writer = tline.writer().await;
6325 2000 : writer
6326 2000 : .put(
6327 2000 : test_key,
6328 2000 : lsn,
6329 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6330 2000 : &ctx,
6331 2000 : )
6332 2 : .await?;
6333 2000 : writer.finish_write(lsn);
6334 2000 : updated[blknum] = lsn;
6335 2000 : drop(writer);
6336 2000 :
6337 2000 : keyspace.add_key(test_key);
6338 2 : }
6339 2 :
6340 102 : for _ in 0..50 {
6341 100 : let new_tline_id = TimelineId::generate();
6342 100 : tenant
6343 100 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
6344 93 : .await?;
6345 100 : tline = tenant
6346 100 : .get_timeline(new_tline_id, true)
6347 100 : .expect("Should have the branched timeline");
6348 2 :
6349 100100 : for _ in 0..NUM_KEYS {
6350 100000 : lsn = Lsn(lsn.0 + 0x10);
6351 100000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6352 100000 : test_key.field6 = blknum as u32;
6353 100000 : let mut writer = tline.writer().await;
6354 100000 : writer
6355 100000 : .put(
6356 100000 : test_key,
6357 100000 : lsn,
6358 100000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6359 100000 : &ctx,
6360 100000 : )
6361 820 : .await?;
6362 100000 : println!("updating {} at {}", blknum, lsn);
6363 100000 : writer.finish_write(lsn);
6364 100000 : drop(writer);
6365 100000 : updated[blknum] = lsn;
6366 2 : }
6367 2 :
6368 2 : // Read all the blocks
6369 100000 : for (blknum, last_lsn) in updated.iter().enumerate() {
6370 100000 : test_key.field6 = blknum as u32;
6371 100000 : assert_eq!(
6372 100000 : tline.get(test_key, lsn, &ctx).await?,
6373 100000 : test_img(&format!("{} at {}", blknum, last_lsn))
6374 2 : );
6375 2 : }
6376 2 :
6377 2 : // Perform a cycle of flush, compact, and GC
6378 102 : tline.freeze_and_flush().await?;
6379 15014 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
6380 100 : tenant
6381 100 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
6382 2 : .await?;
6383 2 : }
6384 2 :
6385 2 : Ok(())
6386 2 : }
6387 :
6388 : #[tokio::test]
6389 2 : async fn test_traverse_ancestors() -> anyhow::Result<()> {
6390 2 : let (tenant, ctx) = TenantHarness::create("test_traverse_ancestors")
6391 2 : .await?
6392 2 : .load()
6393 10 : .await;
6394 2 : let mut tline = tenant
6395 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6396 4 : .await?;
6397 2 :
6398 2 : const NUM_KEYS: usize = 100;
6399 2 : const NUM_TLINES: usize = 50;
6400 2 :
6401 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6402 2 : // Track page mutation lsns across different timelines.
6403 2 : let mut updated = [[Lsn(0); NUM_KEYS]; NUM_TLINES];
6404 2 :
6405 2 : let mut lsn = Lsn(0x10);
6406 2 :
6407 2 : #[allow(clippy::needless_range_loop)]
6408 102 : for idx in 0..NUM_TLINES {
6409 100 : let new_tline_id = TimelineId::generate();
6410 100 : tenant
6411 100 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
6412 59 : .await?;
6413 100 : tline = tenant
6414 100 : .get_timeline(new_tline_id, true)
6415 100 : .expect("Should have the branched timeline");
6416 2 :
6417 10100 : for _ in 0..NUM_KEYS {
6418 10000 : lsn = Lsn(lsn.0 + 0x10);
6419 10000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6420 10000 : test_key.field6 = blknum as u32;
6421 10000 : let mut writer = tline.writer().await;
6422 10000 : writer
6423 10000 : .put(
6424 10000 : test_key,
6425 10000 : lsn,
6426 10000 : &Value::Image(test_img(&format!("{} {} at {}", idx, blknum, lsn))),
6427 10000 : &ctx,
6428 10000 : )
6429 110 : .await?;
6430 10000 : println!("updating [{}][{}] at {}", idx, blknum, lsn);
6431 10000 : writer.finish_write(lsn);
6432 10000 : drop(writer);
6433 10000 : updated[idx][blknum] = lsn;
6434 2 : }
6435 2 : }
6436 2 :
6437 2 : // Read pages from leaf timeline across all ancestors.
6438 100 : for (idx, lsns) in updated.iter().enumerate() {
6439 10000 : for (blknum, lsn) in lsns.iter().enumerate() {
6440 2 : // Skip empty mutations.
6441 10000 : if lsn.0 == 0 {
6442 3677 : continue;
6443 6323 : }
6444 6323 : println!("checking [{idx}][{blknum}] at {lsn}");
6445 6323 : test_key.field6 = blknum as u32;
6446 6323 : assert_eq!(
6447 6323 : tline.get(test_key, *lsn, &ctx).await?,
6448 6323 : test_img(&format!("{idx} {blknum} at {lsn}"))
6449 2 : );
6450 2 : }
6451 2 : }
6452 2 : Ok(())
6453 2 : }
6454 :
6455 : #[tokio::test]
6456 2 : async fn test_write_at_initdb_lsn_takes_optimization_code_path() -> anyhow::Result<()> {
6457 2 : let (tenant, ctx) = TenantHarness::create("test_empty_test_timeline_is_usable")
6458 2 : .await?
6459 2 : .load()
6460 10 : .await;
6461 2 :
6462 2 : let initdb_lsn = Lsn(0x20);
6463 2 : let utline = tenant
6464 2 : .create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)
6465 2 : .await?;
6466 2 : let tline = utline.raw_timeline().unwrap();
6467 2 :
6468 2 : // Spawn flush loop now so that we can set the `expect_initdb_optimization`
6469 2 : tline.maybe_spawn_flush_loop();
6470 2 :
6471 2 : // Make sure the timeline has the minimum set of required keys for operation.
6472 2 : // The only operation you can always do on an empty timeline is to `put` new data.
6473 2 : // Except if you `put` at `initdb_lsn`.
6474 2 : // In that case, there's an optimization to directly create image layers instead of delta layers.
6475 2 : // It uses `repartition()`, which assumes some keys to be present.
6476 2 : // Let's make sure the test timeline can handle that case.
6477 2 : {
6478 2 : let mut state = tline.flush_loop_state.lock().unwrap();
6479 2 : assert_eq!(
6480 2 : timeline::FlushLoopState::Running {
6481 2 : expect_initdb_optimization: false,
6482 2 : initdb_optimization_count: 0,
6483 2 : },
6484 2 : *state
6485 2 : );
6486 2 : *state = timeline::FlushLoopState::Running {
6487 2 : expect_initdb_optimization: true,
6488 2 : initdb_optimization_count: 0,
6489 2 : };
6490 2 : }
6491 2 :
6492 2 : // Make writes at the initdb_lsn. When we flush it below, it should be handled by the optimization.
6493 2 : // As explained above, the optimization requires some keys to be present.
6494 2 : // As per `create_empty_timeline` documentation, use init_empty to set them.
6495 2 : // This is what `create_test_timeline` does, by the way.
6496 2 : let mut modification = tline.begin_modification(initdb_lsn);
6497 2 : modification
6498 2 : .init_empty_test_timeline()
6499 2 : .context("init_empty_test_timeline")?;
6500 2 : modification
6501 2 : .commit(&ctx)
6502 2 : .await
6503 2 : .context("commit init_empty_test_timeline modification")?;
6504 2 :
6505 2 : // Do the flush. The flush code will check the expectations that we set above.
6506 2 : tline.freeze_and_flush().await?;
6507 2 :
6508 2 : // assert freeze_and_flush exercised the initdb optimization
6509 2 : {
6510 2 : let state = tline.flush_loop_state.lock().unwrap();
6511 2 : let timeline::FlushLoopState::Running {
6512 2 : expect_initdb_optimization,
6513 2 : initdb_optimization_count,
6514 2 : } = *state
6515 2 : else {
6516 2 : panic!("unexpected state: {:?}", *state);
6517 2 : };
6518 2 : assert!(expect_initdb_optimization);
6519 2 : assert!(initdb_optimization_count > 0);
6520 2 : }
6521 2 : Ok(())
6522 2 : }
6523 :
6524 : #[tokio::test]
6525 2 : async fn test_create_guard_crash() -> anyhow::Result<()> {
6526 2 : let name = "test_create_guard_crash";
6527 2 : let harness = TenantHarness::create(name).await?;
6528 2 : {
6529 10 : let (tenant, ctx) = harness.load().await;
6530 2 : let tline = tenant
6531 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
6532 2 : .await?;
6533 2 : // Leave the timeline ID in [`Tenant::timelines_creating`] to exclude attempting to create it again
6534 2 : let raw_tline = tline.raw_timeline().unwrap();
6535 2 : raw_tline
6536 2 : .shutdown(super::timeline::ShutdownMode::Hard)
6537 2 : .instrument(info_span!("test_shutdown", tenant_id=%raw_tline.tenant_shard_id, shard_id=%raw_tline.tenant_shard_id.shard_slug(), timeline_id=%TIMELINE_ID))
6538 2 : .await;
6539 2 : std::mem::forget(tline);
6540 2 : }
6541 2 :
6542 10 : let (tenant, _) = harness.load().await;
6543 2 : match tenant.get_timeline(TIMELINE_ID, false) {
6544 2 : Ok(_) => panic!("timeline should've been removed during load"),
6545 2 : Err(e) => {
6546 2 : assert_eq!(
6547 2 : e,
6548 2 : GetTimelineError::NotFound {
6549 2 : tenant_id: tenant.tenant_shard_id,
6550 2 : timeline_id: TIMELINE_ID,
6551 2 : }
6552 2 : )
6553 2 : }
6554 2 : }
6555 2 :
6556 2 : assert!(!harness
6557 2 : .conf
6558 2 : .timeline_path(&tenant.tenant_shard_id, &TIMELINE_ID)
6559 2 : .exists());
6560 2 :
6561 2 : Ok(())
6562 2 : }
6563 :
6564 : #[tokio::test]
6565 2 : async fn test_read_at_max_lsn() -> anyhow::Result<()> {
6566 2 : let names_algorithms = [
6567 2 : ("test_read_at_max_lsn_legacy", CompactionAlgorithm::Legacy),
6568 2 : ("test_read_at_max_lsn_tiered", CompactionAlgorithm::Tiered),
6569 2 : ];
6570 6 : for (name, algorithm) in names_algorithms {
6571 32532 : test_read_at_max_lsn_algorithm(name, algorithm).await?;
6572 2 : }
6573 2 : Ok(())
6574 2 : }
6575 :
6576 4 : async fn test_read_at_max_lsn_algorithm(
6577 4 : name: &'static str,
6578 4 : compaction_algorithm: CompactionAlgorithm,
6579 4 : ) -> anyhow::Result<()> {
6580 4 : let mut harness = TenantHarness::create(name).await?;
6581 4 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
6582 4 : kind: compaction_algorithm,
6583 4 : };
6584 15 : let (tenant, ctx) = harness.load().await;
6585 4 : let tline = tenant
6586 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6587 7 : .await?;
6588 :
6589 4 : let lsn = Lsn(0x10);
6590 4 : let compact = false;
6591 32100 : bulk_insert_maybe_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000, compact).await?;
6592 :
6593 4 : let test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6594 4 : let read_lsn = Lsn(u64::MAX - 1);
6595 :
6596 410 : let result = tline.get(test_key, read_lsn, &ctx).await;
6597 4 : assert!(result.is_ok(), "result is not Ok: {}", result.unwrap_err());
6598 :
6599 4 : Ok(())
6600 4 : }
6601 :
6602 : #[tokio::test]
6603 2 : async fn test_metadata_scan() -> anyhow::Result<()> {
6604 2 : let harness = TenantHarness::create("test_metadata_scan").await?;
6605 10 : let (tenant, ctx) = harness.load().await;
6606 2 : let tline = tenant
6607 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6608 4 : .await?;
6609 2 :
6610 2 : const NUM_KEYS: usize = 1000;
6611 2 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
6612 2 :
6613 2 : let cancel = CancellationToken::new();
6614 2 :
6615 2 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
6616 2 : base_key.field1 = AUX_KEY_PREFIX;
6617 2 : let mut test_key = base_key;
6618 2 :
6619 2 : // Track when each page was last modified. Used to assert that
6620 2 : // a read sees the latest page version.
6621 2 : let mut updated = [Lsn(0); NUM_KEYS];
6622 2 :
6623 2 : let mut lsn = Lsn(0x10);
6624 2 : #[allow(clippy::needless_range_loop)]
6625 2002 : for blknum in 0..NUM_KEYS {
6626 2000 : lsn = Lsn(lsn.0 + 0x10);
6627 2000 : test_key.field6 = (blknum * STEP) as u32;
6628 2000 : let mut writer = tline.writer().await;
6629 2000 : writer
6630 2000 : .put(
6631 2000 : test_key,
6632 2000 : lsn,
6633 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6634 2000 : &ctx,
6635 2000 : )
6636 2 : .await?;
6637 2000 : writer.finish_write(lsn);
6638 2000 : updated[blknum] = lsn;
6639 2000 : drop(writer);
6640 2 : }
6641 2 :
6642 2 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
6643 2 :
6644 24 : for iter in 0..=10 {
6645 2 : // Read all the blocks
6646 22000 : for (blknum, last_lsn) in updated.iter().enumerate() {
6647 22000 : test_key.field6 = (blknum * STEP) as u32;
6648 22000 : assert_eq!(
6649 22000 : tline.get(test_key, lsn, &ctx).await?,
6650 22000 : test_img(&format!("{} at {}", blknum, last_lsn))
6651 2 : );
6652 2 : }
6653 2 :
6654 22 : let mut cnt = 0;
6655 22000 : for (key, value) in tline
6656 22 : .get_vectored_impl(
6657 22 : keyspace.clone(),
6658 22 : lsn,
6659 22 : &mut ValuesReconstructState::default(),
6660 22 : &ctx,
6661 22 : )
6662 1510 : .await?
6663 2 : {
6664 22000 : let blknum = key.field6 as usize;
6665 22000 : let value = value?;
6666 22000 : assert!(blknum % STEP == 0);
6667 22000 : let blknum = blknum / STEP;
6668 22000 : assert_eq!(
6669 22000 : value,
6670 22000 : test_img(&format!("{} at {}", blknum, updated[blknum]))
6671 22000 : );
6672 22000 : cnt += 1;
6673 2 : }
6674 2 :
6675 22 : assert_eq!(cnt, NUM_KEYS);
6676 2 :
6677 22022 : for _ in 0..NUM_KEYS {
6678 22000 : lsn = Lsn(lsn.0 + 0x10);
6679 22000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6680 22000 : test_key.field6 = (blknum * STEP) as u32;
6681 22000 : let mut writer = tline.writer().await;
6682 22000 : writer
6683 22000 : .put(
6684 22000 : test_key,
6685 22000 : lsn,
6686 22000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6687 22000 : &ctx,
6688 22000 : )
6689 96 : .await?;
6690 22000 : writer.finish_write(lsn);
6691 22000 : drop(writer);
6692 22000 : updated[blknum] = lsn;
6693 2 : }
6694 2 :
6695 2 : // Perform two cycles of flush, compact, and GC
6696 66 : for round in 0..2 {
6697 44 : tline.freeze_and_flush().await?;
6698 44 : tline
6699 44 : .compact(
6700 44 : &cancel,
6701 44 : if iter % 5 == 0 && round == 0 {
6702 6 : let mut flags = EnumSet::new();
6703 6 : flags.insert(CompactFlags::ForceImageLayerCreation);
6704 6 : flags.insert(CompactFlags::ForceRepartition);
6705 6 : flags
6706 2 : } else {
6707 38 : EnumSet::empty()
6708 2 : },
6709 44 : &ctx,
6710 2 : )
6711 7004 : .await?;
6712 44 : tenant
6713 44 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
6714 2 : .await?;
6715 2 : }
6716 2 : }
6717 2 :
6718 2 : Ok(())
6719 2 : }
6720 :
6721 : #[tokio::test]
6722 2 : async fn test_metadata_compaction_trigger() -> anyhow::Result<()> {
6723 2 : let harness = TenantHarness::create("test_metadata_compaction_trigger").await?;
6724 10 : let (tenant, ctx) = harness.load().await;
6725 2 : let tline = tenant
6726 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6727 4 : .await?;
6728 2 :
6729 2 : let cancel = CancellationToken::new();
6730 2 :
6731 2 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
6732 2 : base_key.field1 = AUX_KEY_PREFIX;
6733 2 : let test_key = base_key;
6734 2 : let mut lsn = Lsn(0x10);
6735 2 :
6736 42 : for _ in 0..20 {
6737 40 : lsn = Lsn(lsn.0 + 0x10);
6738 40 : let mut writer = tline.writer().await;
6739 40 : writer
6740 40 : .put(
6741 40 : test_key,
6742 40 : lsn,
6743 40 : &Value::Image(test_img(&format!("{} at {}", 0, lsn))),
6744 40 : &ctx,
6745 40 : )
6746 20 : .await?;
6747 40 : writer.finish_write(lsn);
6748 40 : drop(writer);
6749 40 : tline.freeze_and_flush().await?; // force create a delta layer
6750 2 : }
6751 2 :
6752 2 : let before_num_l0_delta_files =
6753 2 : tline.layers.read().await.layer_map()?.level0_deltas().len();
6754 2 :
6755 110 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
6756 2 :
6757 2 : let after_num_l0_delta_files = tline.layers.read().await.layer_map()?.level0_deltas().len();
6758 2 :
6759 2 : assert!(after_num_l0_delta_files < before_num_l0_delta_files, "after_num_l0_delta_files={after_num_l0_delta_files}, before_num_l0_delta_files={before_num_l0_delta_files}");
6760 2 :
6761 2 : assert_eq!(
6762 4 : tline.get(test_key, lsn, &ctx).await?,
6763 2 : test_img(&format!("{} at {}", 0, lsn))
6764 2 : );
6765 2 :
6766 2 : Ok(())
6767 2 : }
6768 :
6769 : #[tokio::test]
6770 2 : async fn test_aux_file_e2e() {
6771 2 : let harness = TenantHarness::create("test_aux_file_e2e").await.unwrap();
6772 2 :
6773 10 : let (tenant, ctx) = harness.load().await;
6774 2 :
6775 2 : let mut lsn = Lsn(0x08);
6776 2 :
6777 2 : let tline: Arc<Timeline> = tenant
6778 2 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
6779 4 : .await
6780 2 : .unwrap();
6781 2 :
6782 2 : {
6783 2 : lsn += 8;
6784 2 : let mut modification = tline.begin_modification(lsn);
6785 2 : modification
6786 2 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
6787 2 : .await
6788 2 : .unwrap();
6789 2 : modification.commit(&ctx).await.unwrap();
6790 2 : }
6791 2 :
6792 2 : // we can read everything from the storage
6793 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6794 2 : assert_eq!(
6795 2 : files.get("pg_logical/mappings/test1"),
6796 2 : Some(&bytes::Bytes::from_static(b"first"))
6797 2 : );
6798 2 :
6799 2 : {
6800 2 : lsn += 8;
6801 2 : let mut modification = tline.begin_modification(lsn);
6802 2 : modification
6803 2 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
6804 2 : .await
6805 2 : .unwrap();
6806 2 : modification.commit(&ctx).await.unwrap();
6807 2 : }
6808 2 :
6809 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6810 2 : assert_eq!(
6811 2 : files.get("pg_logical/mappings/test2"),
6812 2 : Some(&bytes::Bytes::from_static(b"second"))
6813 2 : );
6814 2 :
6815 2 : let child = tenant
6816 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(lsn), &ctx)
6817 2 : .await
6818 2 : .unwrap();
6819 2 :
6820 2 : let files = child.list_aux_files(lsn, &ctx).await.unwrap();
6821 2 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
6822 2 : assert_eq!(files.get("pg_logical/mappings/test2"), None);
6823 2 : }
6824 :
6825 : #[tokio::test]
6826 2 : async fn test_metadata_image_creation() -> anyhow::Result<()> {
6827 2 : let harness = TenantHarness::create("test_metadata_image_creation").await?;
6828 10 : let (tenant, ctx) = harness.load().await;
6829 2 : let tline = tenant
6830 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6831 4 : .await?;
6832 2 :
6833 2 : const NUM_KEYS: usize = 1000;
6834 2 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
6835 2 :
6836 2 : let cancel = CancellationToken::new();
6837 2 :
6838 2 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
6839 2 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
6840 2 : let mut test_key = base_key;
6841 2 : let mut lsn = Lsn(0x10);
6842 2 :
6843 8 : async fn scan_with_statistics(
6844 8 : tline: &Timeline,
6845 8 : keyspace: &KeySpace,
6846 8 : lsn: Lsn,
6847 8 : ctx: &RequestContext,
6848 8 : ) -> anyhow::Result<(BTreeMap<Key, Result<Bytes, PageReconstructError>>, usize)> {
6849 8 : let mut reconstruct_state = ValuesReconstructState::default();
6850 8 : let res = tline
6851 8 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
6852 703 : .await?;
6853 8 : Ok((res, reconstruct_state.get_delta_layers_visited() as usize))
6854 8 : }
6855 2 :
6856 2 : #[allow(clippy::needless_range_loop)]
6857 2002 : for blknum in 0..NUM_KEYS {
6858 2000 : lsn = Lsn(lsn.0 + 0x10);
6859 2000 : test_key.field6 = (blknum * STEP) as u32;
6860 2000 : let mut writer = tline.writer().await;
6861 2000 : writer
6862 2000 : .put(
6863 2000 : test_key,
6864 2000 : lsn,
6865 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6866 2000 : &ctx,
6867 2000 : )
6868 2 : .await?;
6869 2000 : writer.finish_write(lsn);
6870 2000 : drop(writer);
6871 2 : }
6872 2 :
6873 2 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
6874 2 :
6875 22 : for iter in 1..=10 {
6876 20020 : for _ in 0..NUM_KEYS {
6877 20000 : lsn = Lsn(lsn.0 + 0x10);
6878 20000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6879 20000 : test_key.field6 = (blknum * STEP) as u32;
6880 20000 : let mut writer = tline.writer().await;
6881 20000 : writer
6882 20000 : .put(
6883 20000 : test_key,
6884 20000 : lsn,
6885 20000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6886 20000 : &ctx,
6887 20000 : )
6888 35 : .await?;
6889 20000 : writer.finish_write(lsn);
6890 20000 : drop(writer);
6891 2 : }
6892 2 :
6893 20 : tline.freeze_and_flush().await?;
6894 2 :
6895 20 : if iter % 5 == 0 {
6896 4 : let (_, before_delta_file_accessed) =
6897 695 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
6898 4 : tline
6899 4 : .compact(
6900 4 : &cancel,
6901 4 : {
6902 4 : let mut flags = EnumSet::new();
6903 4 : flags.insert(CompactFlags::ForceImageLayerCreation);
6904 4 : flags.insert(CompactFlags::ForceRepartition);
6905 4 : flags
6906 4 : },
6907 4 : &ctx,
6908 4 : )
6909 5109 : .await?;
6910 4 : let (_, after_delta_file_accessed) =
6911 8 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
6912 4 : assert!(after_delta_file_accessed < before_delta_file_accessed, "after_delta_file_accessed={after_delta_file_accessed}, before_delta_file_accessed={before_delta_file_accessed}");
6913 2 : // Given that we already produced an image layer, there should be no delta layer needed for the scan, but still setting a low threshold there for unforeseen circumstances.
6914 4 : assert!(
6915 4 : after_delta_file_accessed <= 2,
6916 2 : "after_delta_file_accessed={after_delta_file_accessed}"
6917 2 : );
6918 16 : }
6919 2 : }
6920 2 :
6921 2 : Ok(())
6922 2 : }
6923 :
6924 : #[tokio::test]
6925 2 : async fn test_vectored_missing_data_key_reads() -> anyhow::Result<()> {
6926 2 : let harness = TenantHarness::create("test_vectored_missing_data_key_reads").await?;
6927 10 : let (tenant, ctx) = harness.load().await;
6928 2 :
6929 2 : let base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
6930 2 : let base_key_child = Key::from_hex("000000000033333333444444445500000001").unwrap();
6931 2 : let base_key_nonexist = Key::from_hex("000000000033333333444444445500000002").unwrap();
6932 2 :
6933 2 : let tline = tenant
6934 2 : .create_test_timeline_with_layers(
6935 2 : TIMELINE_ID,
6936 2 : Lsn(0x10),
6937 2 : DEFAULT_PG_VERSION,
6938 2 : &ctx,
6939 2 : Vec::new(), // delta layers
6940 2 : vec![(Lsn(0x20), vec![(base_key, test_img("data key 1"))])], // image layers
6941 2 : Lsn(0x20), // it's fine to not advance LSN to 0x30 while using 0x30 to get below because `get_vectored_impl` does not wait for LSN
6942 2 : )
6943 11 : .await?;
6944 2 : tline.add_extra_test_dense_keyspace(KeySpace::single(base_key..(base_key_nonexist.next())));
6945 2 :
6946 2 : let child = tenant
6947 2 : .branch_timeline_test_with_layers(
6948 2 : &tline,
6949 2 : NEW_TIMELINE_ID,
6950 2 : Some(Lsn(0x20)),
6951 2 : &ctx,
6952 2 : Vec::new(), // delta layers
6953 2 : vec![(Lsn(0x30), vec![(base_key_child, test_img("data key 2"))])], // image layers
6954 2 : Lsn(0x30),
6955 2 : )
6956 9 : .await
6957 2 : .unwrap();
6958 2 :
6959 2 : let lsn = Lsn(0x30);
6960 2 :
6961 2 : // test vectored get on parent timeline
6962 2 : assert_eq!(
6963 4 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
6964 2 : Some(test_img("data key 1"))
6965 2 : );
6966 2 : assert!(get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx)
6967 3 : .await
6968 2 : .unwrap_err()
6969 2 : .is_missing_key_error());
6970 2 : assert!(
6971 2 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx)
6972 2 : .await
6973 2 : .unwrap_err()
6974 2 : .is_missing_key_error()
6975 2 : );
6976 2 :
6977 2 : // test vectored get on child timeline
6978 2 : assert_eq!(
6979 2 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
6980 2 : Some(test_img("data key 1"))
6981 2 : );
6982 2 : assert_eq!(
6983 4 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
6984 2 : Some(test_img("data key 2"))
6985 2 : );
6986 2 : assert!(
6987 2 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx)
6988 2 : .await
6989 2 : .unwrap_err()
6990 2 : .is_missing_key_error()
6991 2 : );
6992 2 :
6993 2 : Ok(())
6994 2 : }
6995 :
6996 : #[tokio::test]
6997 2 : async fn test_vectored_missing_metadata_key_reads() -> anyhow::Result<()> {
6998 2 : let harness = TenantHarness::create("test_vectored_missing_metadata_key_reads").await?;
6999 10 : let (tenant, ctx) = harness.load().await;
7000 2 :
7001 2 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7002 2 : let base_key_child = Key::from_hex("620000000033333333444444445500000001").unwrap();
7003 2 : let base_key_nonexist = Key::from_hex("620000000033333333444444445500000002").unwrap();
7004 2 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
7005 2 :
7006 2 : let tline = tenant
7007 2 : .create_test_timeline_with_layers(
7008 2 : TIMELINE_ID,
7009 2 : Lsn(0x10),
7010 2 : DEFAULT_PG_VERSION,
7011 2 : &ctx,
7012 2 : Vec::new(), // delta layers
7013 2 : vec![(Lsn(0x20), vec![(base_key, test_img("metadata key 1"))])], // image layers
7014 2 : Lsn(0x20), // it's fine to not advance LSN to 0x30 while using 0x30 to get below because `get_vectored_impl` does not wait for LSN
7015 2 : )
7016 11 : .await?;
7017 2 :
7018 2 : let child = tenant
7019 2 : .branch_timeline_test_with_layers(
7020 2 : &tline,
7021 2 : NEW_TIMELINE_ID,
7022 2 : Some(Lsn(0x20)),
7023 2 : &ctx,
7024 2 : Vec::new(), // delta layers
7025 2 : vec![(
7026 2 : Lsn(0x30),
7027 2 : vec![(base_key_child, test_img("metadata key 2"))],
7028 2 : )], // image layers
7029 2 : Lsn(0x30),
7030 2 : )
7031 9 : .await
7032 2 : .unwrap();
7033 2 :
7034 2 : let lsn = Lsn(0x30);
7035 2 :
7036 2 : // test vectored get on parent timeline
7037 2 : assert_eq!(
7038 4 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
7039 2 : Some(test_img("metadata key 1"))
7040 2 : );
7041 2 : assert_eq!(
7042 2 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx).await?,
7043 2 : None
7044 2 : );
7045 2 : assert_eq!(
7046 2 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx).await?,
7047 2 : None
7048 2 : );
7049 2 :
7050 2 : // test vectored get on child timeline
7051 2 : assert_eq!(
7052 2 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
7053 2 : None
7054 2 : );
7055 2 : assert_eq!(
7056 4 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
7057 2 : Some(test_img("metadata key 2"))
7058 2 : );
7059 2 : assert_eq!(
7060 2 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx).await?,
7061 2 : None
7062 2 : );
7063 2 :
7064 2 : Ok(())
7065 2 : }
7066 :
7067 36 : async fn get_vectored_impl_wrapper(
7068 36 : tline: &Arc<Timeline>,
7069 36 : key: Key,
7070 36 : lsn: Lsn,
7071 36 : ctx: &RequestContext,
7072 36 : ) -> Result<Option<Bytes>, GetVectoredError> {
7073 36 : let mut reconstruct_state = ValuesReconstructState::new();
7074 36 : let mut res = tline
7075 36 : .get_vectored_impl(
7076 36 : KeySpace::single(key..key.next()),
7077 36 : lsn,
7078 36 : &mut reconstruct_state,
7079 36 : ctx,
7080 36 : )
7081 33 : .await?;
7082 30 : Ok(res.pop_last().map(|(k, v)| {
7083 18 : assert_eq!(k, key);
7084 18 : v.unwrap()
7085 30 : }))
7086 36 : }
7087 :
7088 : #[tokio::test]
7089 2 : async fn test_metadata_tombstone_reads() -> anyhow::Result<()> {
7090 2 : let harness = TenantHarness::create("test_metadata_tombstone_reads").await?;
7091 10 : let (tenant, ctx) = harness.load().await;
7092 2 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
7093 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
7094 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
7095 2 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
7096 2 :
7097 2 : // We emulate the situation that the compaction algorithm creates an image layer that removes the tombstones
7098 2 : // Lsn 0x30 key0, key3, no key1+key2
7099 2 : // Lsn 0x20 key1+key2 tomestones
7100 2 : // Lsn 0x10 key1 in image, key2 in delta
7101 2 : let tline = tenant
7102 2 : .create_test_timeline_with_layers(
7103 2 : TIMELINE_ID,
7104 2 : Lsn(0x10),
7105 2 : DEFAULT_PG_VERSION,
7106 2 : &ctx,
7107 2 : // delta layers
7108 2 : vec![
7109 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7110 2 : Lsn(0x10)..Lsn(0x20),
7111 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
7112 2 : ),
7113 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7114 2 : Lsn(0x20)..Lsn(0x30),
7115 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
7116 2 : ),
7117 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7118 2 : Lsn(0x20)..Lsn(0x30),
7119 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
7120 2 : ),
7121 2 : ],
7122 2 : // image layers
7123 2 : vec![
7124 2 : (Lsn(0x10), vec![(key1, test_img("metadata key 1"))]),
7125 2 : (
7126 2 : Lsn(0x30),
7127 2 : vec![
7128 2 : (key0, test_img("metadata key 0")),
7129 2 : (key3, test_img("metadata key 3")),
7130 2 : ],
7131 2 : ),
7132 2 : ],
7133 2 : Lsn(0x30),
7134 2 : )
7135 38 : .await?;
7136 2 :
7137 2 : let lsn = Lsn(0x30);
7138 2 : let old_lsn = Lsn(0x20);
7139 2 :
7140 2 : assert_eq!(
7141 4 : get_vectored_impl_wrapper(&tline, key0, lsn, &ctx).await?,
7142 2 : Some(test_img("metadata key 0"))
7143 2 : );
7144 2 : assert_eq!(
7145 2 : get_vectored_impl_wrapper(&tline, key1, lsn, &ctx).await?,
7146 2 : None,
7147 2 : );
7148 2 : assert_eq!(
7149 2 : get_vectored_impl_wrapper(&tline, key2, lsn, &ctx).await?,
7150 2 : None,
7151 2 : );
7152 2 : assert_eq!(
7153 4 : get_vectored_impl_wrapper(&tline, key1, old_lsn, &ctx).await?,
7154 2 : Some(Bytes::new()),
7155 2 : );
7156 2 : assert_eq!(
7157 4 : get_vectored_impl_wrapper(&tline, key2, old_lsn, &ctx).await?,
7158 2 : Some(Bytes::new()),
7159 2 : );
7160 2 : assert_eq!(
7161 2 : get_vectored_impl_wrapper(&tline, key3, lsn, &ctx).await?,
7162 2 : Some(test_img("metadata key 3"))
7163 2 : );
7164 2 :
7165 2 : Ok(())
7166 2 : }
7167 :
7168 : #[tokio::test]
7169 2 : async fn test_metadata_tombstone_image_creation() {
7170 2 : let harness = TenantHarness::create("test_metadata_tombstone_image_creation")
7171 2 : .await
7172 2 : .unwrap();
7173 10 : let (tenant, ctx) = harness.load().await;
7174 2 :
7175 2 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
7176 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
7177 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
7178 2 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
7179 2 :
7180 2 : let tline = tenant
7181 2 : .create_test_timeline_with_layers(
7182 2 : TIMELINE_ID,
7183 2 : Lsn(0x10),
7184 2 : DEFAULT_PG_VERSION,
7185 2 : &ctx,
7186 2 : // delta layers
7187 2 : vec![
7188 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7189 2 : Lsn(0x10)..Lsn(0x20),
7190 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
7191 2 : ),
7192 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7193 2 : Lsn(0x20)..Lsn(0x30),
7194 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
7195 2 : ),
7196 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7197 2 : Lsn(0x20)..Lsn(0x30),
7198 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
7199 2 : ),
7200 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7201 2 : Lsn(0x30)..Lsn(0x40),
7202 2 : vec![
7203 2 : (key0, Lsn(0x30), Value::Image(test_img("metadata key 0"))),
7204 2 : (key3, Lsn(0x30), Value::Image(test_img("metadata key 3"))),
7205 2 : ],
7206 2 : ),
7207 2 : ],
7208 2 : // image layers
7209 2 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
7210 2 : Lsn(0x40),
7211 2 : )
7212 35 : .await
7213 2 : .unwrap();
7214 2 :
7215 2 : let cancel = CancellationToken::new();
7216 2 :
7217 2 : tline
7218 2 : .compact(
7219 2 : &cancel,
7220 2 : {
7221 2 : let mut flags = EnumSet::new();
7222 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
7223 2 : flags.insert(CompactFlags::ForceRepartition);
7224 2 : flags
7225 2 : },
7226 2 : &ctx,
7227 2 : )
7228 55 : .await
7229 2 : .unwrap();
7230 2 :
7231 2 : // Image layers are created at last_record_lsn
7232 2 : let images = tline
7233 2 : .inspect_image_layers(Lsn(0x40), &ctx)
7234 8 : .await
7235 2 : .unwrap()
7236 2 : .into_iter()
7237 18 : .filter(|(k, _)| k.is_metadata_key())
7238 2 : .collect::<Vec<_>>();
7239 2 : assert_eq!(images.len(), 2); // the image layer should only contain two existing keys, tombstones should be removed.
7240 2 : }
7241 :
7242 : #[tokio::test]
7243 2 : async fn test_metadata_tombstone_empty_image_creation() {
7244 2 : let harness = TenantHarness::create("test_metadata_tombstone_empty_image_creation")
7245 2 : .await
7246 2 : .unwrap();
7247 10 : let (tenant, ctx) = harness.load().await;
7248 2 :
7249 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
7250 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
7251 2 :
7252 2 : let tline = tenant
7253 2 : .create_test_timeline_with_layers(
7254 2 : TIMELINE_ID,
7255 2 : Lsn(0x10),
7256 2 : DEFAULT_PG_VERSION,
7257 2 : &ctx,
7258 2 : // delta layers
7259 2 : vec![
7260 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7261 2 : Lsn(0x10)..Lsn(0x20),
7262 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
7263 2 : ),
7264 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7265 2 : Lsn(0x20)..Lsn(0x30),
7266 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
7267 2 : ),
7268 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7269 2 : Lsn(0x20)..Lsn(0x30),
7270 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
7271 2 : ),
7272 2 : ],
7273 2 : // image layers
7274 2 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
7275 2 : Lsn(0x30),
7276 2 : )
7277 29 : .await
7278 2 : .unwrap();
7279 2 :
7280 2 : let cancel = CancellationToken::new();
7281 2 :
7282 2 : tline
7283 2 : .compact(
7284 2 : &cancel,
7285 2 : {
7286 2 : let mut flags = EnumSet::new();
7287 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
7288 2 : flags.insert(CompactFlags::ForceRepartition);
7289 2 : flags
7290 2 : },
7291 2 : &ctx,
7292 2 : )
7293 43 : .await
7294 2 : .unwrap();
7295 2 :
7296 2 : // Image layers are created at last_record_lsn
7297 2 : let images = tline
7298 2 : .inspect_image_layers(Lsn(0x30), &ctx)
7299 4 : .await
7300 2 : .unwrap()
7301 2 : .into_iter()
7302 14 : .filter(|(k, _)| k.is_metadata_key())
7303 2 : .collect::<Vec<_>>();
7304 2 : assert_eq!(images.len(), 0); // the image layer should not contain tombstones, or it is not created
7305 2 : }
7306 :
7307 : #[tokio::test]
7308 2 : async fn test_simple_bottom_most_compaction_images() -> anyhow::Result<()> {
7309 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_images").await?;
7310 10 : let (tenant, ctx) = harness.load().await;
7311 2 :
7312 102 : fn get_key(id: u32) -> Key {
7313 102 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7314 102 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7315 102 : key.field6 = id;
7316 102 : key
7317 102 : }
7318 2 :
7319 2 : // We create
7320 2 : // - one bottom-most image layer,
7321 2 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
7322 2 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
7323 2 : // - a delta layer D3 above the horizon.
7324 2 : //
7325 2 : // | D3 |
7326 2 : // | D1 |
7327 2 : // -| |-- gc horizon -----------------
7328 2 : // | | | D2 |
7329 2 : // --------- img layer ------------------
7330 2 : //
7331 2 : // What we should expact from this compaction is:
7332 2 : // | D3 |
7333 2 : // | Part of D1 |
7334 2 : // --------- img layer with D1+D2 at GC horizon------------------
7335 2 :
7336 2 : // img layer at 0x10
7337 2 : let img_layer = (0..10)
7338 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
7339 2 : .collect_vec();
7340 2 :
7341 2 : let delta1 = vec![
7342 2 : (
7343 2 : get_key(1),
7344 2 : Lsn(0x20),
7345 2 : Value::Image(Bytes::from("value 1@0x20")),
7346 2 : ),
7347 2 : (
7348 2 : get_key(2),
7349 2 : Lsn(0x30),
7350 2 : Value::Image(Bytes::from("value 2@0x30")),
7351 2 : ),
7352 2 : (
7353 2 : get_key(3),
7354 2 : Lsn(0x40),
7355 2 : Value::Image(Bytes::from("value 3@0x40")),
7356 2 : ),
7357 2 : ];
7358 2 : let delta2 = vec![
7359 2 : (
7360 2 : get_key(5),
7361 2 : Lsn(0x20),
7362 2 : Value::Image(Bytes::from("value 5@0x20")),
7363 2 : ),
7364 2 : (
7365 2 : get_key(6),
7366 2 : Lsn(0x20),
7367 2 : Value::Image(Bytes::from("value 6@0x20")),
7368 2 : ),
7369 2 : ];
7370 2 : let delta3 = vec![
7371 2 : (
7372 2 : get_key(8),
7373 2 : Lsn(0x48),
7374 2 : Value::Image(Bytes::from("value 8@0x48")),
7375 2 : ),
7376 2 : (
7377 2 : get_key(9),
7378 2 : Lsn(0x48),
7379 2 : Value::Image(Bytes::from("value 9@0x48")),
7380 2 : ),
7381 2 : ];
7382 2 :
7383 2 : let tline = tenant
7384 2 : .create_test_timeline_with_layers(
7385 2 : TIMELINE_ID,
7386 2 : Lsn(0x10),
7387 2 : DEFAULT_PG_VERSION,
7388 2 : &ctx,
7389 2 : vec![
7390 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
7391 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
7392 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
7393 2 : ], // delta layers
7394 2 : vec![(Lsn(0x10), img_layer)], // image layers
7395 2 : Lsn(0x50),
7396 2 : )
7397 47 : .await?;
7398 2 : {
7399 2 : // Update GC info
7400 2 : let mut guard = tline.gc_info.write().unwrap();
7401 2 : guard.cutoffs.time = Lsn(0x30);
7402 2 : guard.cutoffs.space = Lsn(0x30);
7403 2 : }
7404 2 :
7405 2 : let expected_result = [
7406 2 : Bytes::from_static(b"value 0@0x10"),
7407 2 : Bytes::from_static(b"value 1@0x20"),
7408 2 : Bytes::from_static(b"value 2@0x30"),
7409 2 : Bytes::from_static(b"value 3@0x40"),
7410 2 : Bytes::from_static(b"value 4@0x10"),
7411 2 : Bytes::from_static(b"value 5@0x20"),
7412 2 : Bytes::from_static(b"value 6@0x20"),
7413 2 : Bytes::from_static(b"value 7@0x10"),
7414 2 : Bytes::from_static(b"value 8@0x48"),
7415 2 : Bytes::from_static(b"value 9@0x48"),
7416 2 : ];
7417 2 :
7418 20 : for (idx, expected) in expected_result.iter().enumerate() {
7419 20 : assert_eq!(
7420 20 : tline
7421 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7422 23 : .await
7423 20 : .unwrap(),
7424 2 : expected
7425 2 : );
7426 2 : }
7427 2 :
7428 2 : let cancel = CancellationToken::new();
7429 2 : tline
7430 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7431 53 : .await
7432 2 : .unwrap();
7433 2 :
7434 20 : for (idx, expected) in expected_result.iter().enumerate() {
7435 20 : assert_eq!(
7436 20 : tline
7437 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7438 17 : .await
7439 20 : .unwrap(),
7440 2 : expected
7441 2 : );
7442 2 : }
7443 2 :
7444 2 : // Check if the image layer at the GC horizon contains exactly what we want
7445 2 : let image_at_gc_horizon = tline
7446 2 : .inspect_image_layers(Lsn(0x30), &ctx)
7447 2 : .await
7448 2 : .unwrap()
7449 2 : .into_iter()
7450 34 : .filter(|(k, _)| k.is_metadata_key())
7451 2 : .collect::<Vec<_>>();
7452 2 :
7453 2 : assert_eq!(image_at_gc_horizon.len(), 10);
7454 2 : let expected_result = [
7455 2 : Bytes::from_static(b"value 0@0x10"),
7456 2 : Bytes::from_static(b"value 1@0x20"),
7457 2 : Bytes::from_static(b"value 2@0x30"),
7458 2 : Bytes::from_static(b"value 3@0x10"),
7459 2 : Bytes::from_static(b"value 4@0x10"),
7460 2 : Bytes::from_static(b"value 5@0x20"),
7461 2 : Bytes::from_static(b"value 6@0x20"),
7462 2 : Bytes::from_static(b"value 7@0x10"),
7463 2 : Bytes::from_static(b"value 8@0x10"),
7464 2 : Bytes::from_static(b"value 9@0x10"),
7465 2 : ];
7466 22 : for idx in 0..10 {
7467 20 : assert_eq!(
7468 20 : image_at_gc_horizon[idx],
7469 20 : (get_key(idx as u32), expected_result[idx].clone())
7470 20 : );
7471 2 : }
7472 2 :
7473 2 : // Check if old layers are removed / new layers have the expected LSN
7474 2 : let mut all_layers = tline.inspect_historic_layers().await.unwrap();
7475 4 : all_layers.sort_by(|k1, k2| {
7476 4 : (
7477 4 : k1.is_delta,
7478 4 : k1.key_range.start,
7479 4 : k1.key_range.end,
7480 4 : k1.lsn_range.start,
7481 4 : k1.lsn_range.end,
7482 4 : )
7483 4 : .cmp(&(
7484 4 : k2.is_delta,
7485 4 : k2.key_range.start,
7486 4 : k2.key_range.end,
7487 4 : k2.lsn_range.start,
7488 4 : k2.lsn_range.end,
7489 4 : ))
7490 4 : });
7491 2 : assert_eq!(
7492 2 : all_layers,
7493 2 : vec![
7494 2 : // Image layer at GC horizon
7495 2 : PersistentLayerKey {
7496 2 : key_range: Key::MIN..Key::MAX,
7497 2 : lsn_range: Lsn(0x30)..Lsn(0x31),
7498 2 : is_delta: false
7499 2 : },
7500 2 : // The delta layer below the horizon
7501 2 : PersistentLayerKey {
7502 2 : key_range: get_key(3)..get_key(4),
7503 2 : lsn_range: Lsn(0x30)..Lsn(0x48),
7504 2 : is_delta: true
7505 2 : },
7506 2 : // The delta3 layer that should not be picked for the compaction
7507 2 : PersistentLayerKey {
7508 2 : key_range: get_key(8)..get_key(10),
7509 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
7510 2 : is_delta: true
7511 2 : }
7512 2 : ]
7513 2 : );
7514 2 :
7515 2 : // increase GC horizon and compact again
7516 2 : {
7517 2 : // Update GC info
7518 2 : let mut guard = tline.gc_info.write().unwrap();
7519 2 : guard.cutoffs.time = Lsn(0x40);
7520 2 : guard.cutoffs.space = Lsn(0x40);
7521 2 : }
7522 2 : tline
7523 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7524 43 : .await
7525 2 : .unwrap();
7526 2 :
7527 2 : Ok(())
7528 2 : }
7529 :
7530 : #[tokio::test]
7531 2 : async fn test_neon_test_record() -> anyhow::Result<()> {
7532 2 : let harness = TenantHarness::create("test_neon_test_record").await?;
7533 10 : let (tenant, ctx) = harness.load().await;
7534 2 :
7535 24 : fn get_key(id: u32) -> Key {
7536 24 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7537 24 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7538 24 : key.field6 = id;
7539 24 : key
7540 24 : }
7541 2 :
7542 2 : let delta1 = vec![
7543 2 : (
7544 2 : get_key(1),
7545 2 : Lsn(0x20),
7546 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
7547 2 : ),
7548 2 : (
7549 2 : get_key(1),
7550 2 : Lsn(0x30),
7551 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
7552 2 : ),
7553 2 : (get_key(2), Lsn(0x10), Value::Image("0x10".into())),
7554 2 : (
7555 2 : get_key(2),
7556 2 : Lsn(0x20),
7557 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
7558 2 : ),
7559 2 : (
7560 2 : get_key(2),
7561 2 : Lsn(0x30),
7562 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
7563 2 : ),
7564 2 : (get_key(3), Lsn(0x10), Value::Image("0x10".into())),
7565 2 : (
7566 2 : get_key(3),
7567 2 : Lsn(0x20),
7568 2 : Value::WalRecord(NeonWalRecord::wal_clear()),
7569 2 : ),
7570 2 : (get_key(4), Lsn(0x10), Value::Image("0x10".into())),
7571 2 : (
7572 2 : get_key(4),
7573 2 : Lsn(0x20),
7574 2 : Value::WalRecord(NeonWalRecord::wal_init()),
7575 2 : ),
7576 2 : ];
7577 2 : let image1 = vec![(get_key(1), "0x10".into())];
7578 2 :
7579 2 : let tline = tenant
7580 2 : .create_test_timeline_with_layers(
7581 2 : TIMELINE_ID,
7582 2 : Lsn(0x10),
7583 2 : DEFAULT_PG_VERSION,
7584 2 : &ctx,
7585 2 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
7586 2 : Lsn(0x10)..Lsn(0x40),
7587 2 : delta1,
7588 2 : )], // delta layers
7589 2 : vec![(Lsn(0x10), image1)], // image layers
7590 2 : Lsn(0x50),
7591 2 : )
7592 17 : .await?;
7593 2 :
7594 2 : assert_eq!(
7595 8 : tline.get(get_key(1), Lsn(0x50), &ctx).await?,
7596 2 : Bytes::from_static(b"0x10,0x20,0x30")
7597 2 : );
7598 2 : assert_eq!(
7599 2 : tline.get(get_key(2), Lsn(0x50), &ctx).await?,
7600 2 : Bytes::from_static(b"0x10,0x20,0x30")
7601 2 : );
7602 2 :
7603 2 : // Need to remove the limit of "Neon WAL redo requires base image".
7604 2 :
7605 2 : // assert_eq!(tline.get(get_key(3), Lsn(0x50), &ctx).await?, Bytes::new());
7606 2 : // assert_eq!(tline.get(get_key(4), Lsn(0x50), &ctx).await?, Bytes::new());
7607 2 :
7608 2 : Ok(())
7609 2 : }
7610 :
7611 : #[tokio::test(start_paused = true)]
7612 2 : async fn test_lsn_lease() -> anyhow::Result<()> {
7613 2 : let (tenant, ctx) = TenantHarness::create("test_lsn_lease")
7614 2 : .await
7615 2 : .unwrap()
7616 2 : .load()
7617 10 : .await;
7618 2 : // Advance to the lsn lease deadline so that GC is not blocked by
7619 2 : // initial transition into AttachedSingle.
7620 2 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
7621 2 : tokio::time::resume();
7622 2 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7623 2 :
7624 2 : let end_lsn = Lsn(0x100);
7625 2 : let image_layers = (0x20..=0x90)
7626 2 : .step_by(0x10)
7627 16 : .map(|n| {
7628 16 : (
7629 16 : Lsn(n),
7630 16 : vec![(key, test_img(&format!("data key at {:x}", n)))],
7631 16 : )
7632 16 : })
7633 2 : .collect();
7634 2 :
7635 2 : let timeline = tenant
7636 2 : .create_test_timeline_with_layers(
7637 2 : TIMELINE_ID,
7638 2 : Lsn(0x10),
7639 2 : DEFAULT_PG_VERSION,
7640 2 : &ctx,
7641 2 : Vec::new(),
7642 2 : image_layers,
7643 2 : end_lsn,
7644 2 : )
7645 60 : .await?;
7646 2 :
7647 2 : let leased_lsns = [0x30, 0x50, 0x70];
7648 2 : let mut leases = Vec::new();
7649 6 : leased_lsns.iter().for_each(|n| {
7650 6 : leases.push(
7651 6 : timeline
7652 6 : .init_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)
7653 6 : .expect("lease request should succeed"),
7654 6 : );
7655 6 : });
7656 2 :
7657 2 : let updated_lease_0 = timeline
7658 2 : .renew_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)
7659 2 : .expect("lease renewal should succeed");
7660 2 : assert_eq!(
7661 2 : updated_lease_0.valid_until, leases[0].valid_until,
7662 2 : " Renewing with shorter lease should not change the lease."
7663 2 : );
7664 2 :
7665 2 : let updated_lease_1 = timeline
7666 2 : .renew_lsn_lease(
7667 2 : Lsn(leased_lsns[1]),
7668 2 : timeline.get_lsn_lease_length() * 2,
7669 2 : &ctx,
7670 2 : )
7671 2 : .expect("lease renewal should succeed");
7672 2 : assert!(
7673 2 : updated_lease_1.valid_until > leases[1].valid_until,
7674 2 : "Renewing with a long lease should renew lease with later expiration time."
7675 2 : );
7676 2 :
7677 2 : // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
7678 2 : info!(
7679 2 : "latest_gc_cutoff_lsn: {}",
7680 0 : *timeline.get_latest_gc_cutoff_lsn()
7681 2 : );
7682 2 : timeline.force_set_disk_consistent_lsn(end_lsn);
7683 2 :
7684 2 : let res = tenant
7685 2 : .gc_iteration(
7686 2 : Some(TIMELINE_ID),
7687 2 : 0,
7688 2 : Duration::ZERO,
7689 2 : &CancellationToken::new(),
7690 2 : &ctx,
7691 2 : )
7692 2 : .await
7693 2 : .unwrap();
7694 2 :
7695 2 : // Keeping everything <= Lsn(0x80) b/c leases:
7696 2 : // 0/10: initdb layer
7697 2 : // (0/20..=0/70).step_by(0x10): image layers added when creating the timeline.
7698 2 : assert_eq!(res.layers_needed_by_leases, 7);
7699 2 : // Keeping 0/90 b/c it is the latest layer.
7700 2 : assert_eq!(res.layers_not_updated, 1);
7701 2 : // Removed 0/80.
7702 2 : assert_eq!(res.layers_removed, 1);
7703 2 :
7704 2 : // Make lease on a already GC-ed LSN.
7705 2 : // 0/80 does not have a valid lease + is below latest_gc_cutoff
7706 2 : assert!(Lsn(0x80) < *timeline.get_latest_gc_cutoff_lsn());
7707 2 : timeline
7708 2 : .init_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx)
7709 2 : .expect_err("lease request on GC-ed LSN should fail");
7710 2 :
7711 2 : // Should still be able to renew a currently valid lease
7712 2 : // Assumption: original lease to is still valid for 0/50.
7713 2 : // (use `Timeline::init_lsn_lease` for testing so it always does validation)
7714 2 : timeline
7715 2 : .init_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)
7716 2 : .expect("lease renewal with validation should succeed");
7717 2 :
7718 2 : Ok(())
7719 2 : }
7720 :
7721 : #[tokio::test]
7722 2 : async fn test_simple_bottom_most_compaction_deltas() -> anyhow::Result<()> {
7723 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_deltas").await?;
7724 10 : let (tenant, ctx) = harness.load().await;
7725 2 :
7726 118 : fn get_key(id: u32) -> Key {
7727 118 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7728 118 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7729 118 : key.field6 = id;
7730 118 : key
7731 118 : }
7732 2 :
7733 2 : // We create
7734 2 : // - one bottom-most image layer,
7735 2 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
7736 2 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
7737 2 : // - a delta layer D3 above the horizon.
7738 2 : //
7739 2 : // | D3 |
7740 2 : // | D1 |
7741 2 : // -| |-- gc horizon -----------------
7742 2 : // | | | D2 |
7743 2 : // --------- img layer ------------------
7744 2 : //
7745 2 : // What we should expact from this compaction is:
7746 2 : // | D3 |
7747 2 : // | Part of D1 |
7748 2 : // --------- img layer with D1+D2 at GC horizon------------------
7749 2 :
7750 2 : // img layer at 0x10
7751 2 : let img_layer = (0..10)
7752 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
7753 2 : .collect_vec();
7754 2 :
7755 2 : let delta1 = vec![
7756 2 : (
7757 2 : get_key(1),
7758 2 : Lsn(0x20),
7759 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7760 2 : ),
7761 2 : (
7762 2 : get_key(2),
7763 2 : Lsn(0x30),
7764 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7765 2 : ),
7766 2 : (
7767 2 : get_key(3),
7768 2 : Lsn(0x28),
7769 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
7770 2 : ),
7771 2 : (
7772 2 : get_key(3),
7773 2 : Lsn(0x30),
7774 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7775 2 : ),
7776 2 : (
7777 2 : get_key(3),
7778 2 : Lsn(0x40),
7779 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
7780 2 : ),
7781 2 : ];
7782 2 : let delta2 = vec![
7783 2 : (
7784 2 : get_key(5),
7785 2 : Lsn(0x20),
7786 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7787 2 : ),
7788 2 : (
7789 2 : get_key(6),
7790 2 : Lsn(0x20),
7791 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7792 2 : ),
7793 2 : ];
7794 2 : let delta3 = vec![
7795 2 : (
7796 2 : get_key(8),
7797 2 : Lsn(0x48),
7798 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
7799 2 : ),
7800 2 : (
7801 2 : get_key(9),
7802 2 : Lsn(0x48),
7803 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
7804 2 : ),
7805 2 : ];
7806 2 :
7807 2 : let tline = tenant
7808 2 : .create_test_timeline_with_layers(
7809 2 : TIMELINE_ID,
7810 2 : Lsn(0x10),
7811 2 : DEFAULT_PG_VERSION,
7812 2 : &ctx,
7813 2 : vec![
7814 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
7815 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
7816 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
7817 2 : ], // delta layers
7818 2 : vec![(Lsn(0x10), img_layer)], // image layers
7819 2 : Lsn(0x50),
7820 2 : )
7821 47 : .await?;
7822 2 : {
7823 2 : // Update GC info
7824 2 : let mut guard = tline.gc_info.write().unwrap();
7825 2 : *guard = GcInfo {
7826 2 : retain_lsns: vec![],
7827 2 : cutoffs: GcCutoffs {
7828 2 : time: Lsn(0x30),
7829 2 : space: Lsn(0x30),
7830 2 : },
7831 2 : leases: Default::default(),
7832 2 : within_ancestor_pitr: false,
7833 2 : };
7834 2 : }
7835 2 :
7836 2 : let expected_result = [
7837 2 : Bytes::from_static(b"value 0@0x10"),
7838 2 : Bytes::from_static(b"value 1@0x10@0x20"),
7839 2 : Bytes::from_static(b"value 2@0x10@0x30"),
7840 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
7841 2 : Bytes::from_static(b"value 4@0x10"),
7842 2 : Bytes::from_static(b"value 5@0x10@0x20"),
7843 2 : Bytes::from_static(b"value 6@0x10@0x20"),
7844 2 : Bytes::from_static(b"value 7@0x10"),
7845 2 : Bytes::from_static(b"value 8@0x10@0x48"),
7846 2 : Bytes::from_static(b"value 9@0x10@0x48"),
7847 2 : ];
7848 2 :
7849 2 : let expected_result_at_gc_horizon = [
7850 2 : Bytes::from_static(b"value 0@0x10"),
7851 2 : Bytes::from_static(b"value 1@0x10@0x20"),
7852 2 : Bytes::from_static(b"value 2@0x10@0x30"),
7853 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
7854 2 : Bytes::from_static(b"value 4@0x10"),
7855 2 : Bytes::from_static(b"value 5@0x10@0x20"),
7856 2 : Bytes::from_static(b"value 6@0x10@0x20"),
7857 2 : Bytes::from_static(b"value 7@0x10"),
7858 2 : Bytes::from_static(b"value 8@0x10"),
7859 2 : Bytes::from_static(b"value 9@0x10"),
7860 2 : ];
7861 2 :
7862 22 : for idx in 0..10 {
7863 20 : assert_eq!(
7864 20 : tline
7865 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7866 31 : .await
7867 20 : .unwrap(),
7868 20 : &expected_result[idx]
7869 2 : );
7870 20 : assert_eq!(
7871 20 : tline
7872 20 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
7873 15 : .await
7874 20 : .unwrap(),
7875 20 : &expected_result_at_gc_horizon[idx]
7876 2 : );
7877 2 : }
7878 2 :
7879 2 : let cancel = CancellationToken::new();
7880 2 : tline
7881 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7882 60 : .await
7883 2 : .unwrap();
7884 2 :
7885 22 : for idx in 0..10 {
7886 20 : assert_eq!(
7887 20 : tline
7888 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7889 20 : .await
7890 20 : .unwrap(),
7891 20 : &expected_result[idx]
7892 2 : );
7893 20 : assert_eq!(
7894 20 : tline
7895 20 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
7896 10 : .await
7897 20 : .unwrap(),
7898 20 : &expected_result_at_gc_horizon[idx]
7899 2 : );
7900 2 : }
7901 2 :
7902 2 : // increase GC horizon and compact again
7903 2 : {
7904 2 : // Update GC info
7905 2 : let mut guard = tline.gc_info.write().unwrap();
7906 2 : guard.cutoffs.time = Lsn(0x40);
7907 2 : guard.cutoffs.space = Lsn(0x40);
7908 2 : }
7909 2 : tline
7910 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7911 43 : .await
7912 2 : .unwrap();
7913 2 :
7914 2 : Ok(())
7915 2 : }
7916 :
7917 : #[tokio::test]
7918 2 : async fn test_generate_key_retention() -> anyhow::Result<()> {
7919 2 : let harness = TenantHarness::create("test_generate_key_retention").await?;
7920 10 : let (tenant, ctx) = harness.load().await;
7921 2 : let tline = tenant
7922 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7923 4 : .await?;
7924 2 : tline.force_advance_lsn(Lsn(0x70));
7925 2 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7926 2 : let history = vec![
7927 2 : (
7928 2 : key,
7929 2 : Lsn(0x10),
7930 2 : Value::Image(Bytes::copy_from_slice(b"0x10")),
7931 2 : ),
7932 2 : (
7933 2 : key,
7934 2 : Lsn(0x20),
7935 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7936 2 : ),
7937 2 : (
7938 2 : key,
7939 2 : Lsn(0x30),
7940 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
7941 2 : ),
7942 2 : (
7943 2 : key,
7944 2 : Lsn(0x40),
7945 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
7946 2 : ),
7947 2 : (
7948 2 : key,
7949 2 : Lsn(0x50),
7950 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
7951 2 : ),
7952 2 : (
7953 2 : key,
7954 2 : Lsn(0x60),
7955 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7956 2 : ),
7957 2 : (
7958 2 : key,
7959 2 : Lsn(0x70),
7960 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7961 2 : ),
7962 2 : (
7963 2 : key,
7964 2 : Lsn(0x80),
7965 2 : Value::Image(Bytes::copy_from_slice(
7966 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
7967 2 : )),
7968 2 : ),
7969 2 : (
7970 2 : key,
7971 2 : Lsn(0x90),
7972 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
7973 2 : ),
7974 2 : ];
7975 2 : let res = tline
7976 2 : .generate_key_retention(
7977 2 : key,
7978 2 : &history,
7979 2 : Lsn(0x60),
7980 2 : &[Lsn(0x20), Lsn(0x40), Lsn(0x50)],
7981 2 : 3,
7982 2 : None,
7983 2 : )
7984 2 : .await
7985 2 : .unwrap();
7986 2 : let expected_res = KeyHistoryRetention {
7987 2 : below_horizon: vec![
7988 2 : (
7989 2 : Lsn(0x20),
7990 2 : KeyLogAtLsn(vec![(
7991 2 : Lsn(0x20),
7992 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20")),
7993 2 : )]),
7994 2 : ),
7995 2 : (
7996 2 : Lsn(0x40),
7997 2 : KeyLogAtLsn(vec![
7998 2 : (
7999 2 : Lsn(0x30),
8000 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8001 2 : ),
8002 2 : (
8003 2 : Lsn(0x40),
8004 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8005 2 : ),
8006 2 : ]),
8007 2 : ),
8008 2 : (
8009 2 : Lsn(0x50),
8010 2 : KeyLogAtLsn(vec![(
8011 2 : Lsn(0x50),
8012 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40;0x50")),
8013 2 : )]),
8014 2 : ),
8015 2 : (
8016 2 : Lsn(0x60),
8017 2 : KeyLogAtLsn(vec![(
8018 2 : Lsn(0x60),
8019 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8020 2 : )]),
8021 2 : ),
8022 2 : ],
8023 2 : above_horizon: KeyLogAtLsn(vec![
8024 2 : (
8025 2 : Lsn(0x70),
8026 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8027 2 : ),
8028 2 : (
8029 2 : Lsn(0x80),
8030 2 : Value::Image(Bytes::copy_from_slice(
8031 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8032 2 : )),
8033 2 : ),
8034 2 : (
8035 2 : Lsn(0x90),
8036 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8037 2 : ),
8038 2 : ]),
8039 2 : };
8040 2 : assert_eq!(res, expected_res);
8041 2 :
8042 2 : // We expect GC-compaction to run with the original GC. This would create a situation that
8043 2 : // the original GC algorithm removes some delta layers b/c there are full image coverage,
8044 2 : // therefore causing some keys to have an incomplete history below the lowest retain LSN.
8045 2 : // For example, we have
8046 2 : // ```plain
8047 2 : // init delta @ 0x10, image @ 0x20, delta @ 0x30 (gc_horizon), image @ 0x40.
8048 2 : // ```
8049 2 : // Now the GC horizon moves up, and we have
8050 2 : // ```plain
8051 2 : // init delta @ 0x10, image @ 0x20, delta @ 0x30, image @ 0x40 (gc_horizon)
8052 2 : // ```
8053 2 : // The original GC algorithm kicks in, and removes delta @ 0x10, image @ 0x20.
8054 2 : // We will end up with
8055 2 : // ```plain
8056 2 : // delta @ 0x30, image @ 0x40 (gc_horizon)
8057 2 : // ```
8058 2 : // Now we run the GC-compaction, and this key does not have a full history.
8059 2 : // We should be able to handle this partial history and drop everything before the
8060 2 : // gc_horizon image.
8061 2 :
8062 2 : let history = vec![
8063 2 : (
8064 2 : key,
8065 2 : Lsn(0x20),
8066 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8067 2 : ),
8068 2 : (
8069 2 : key,
8070 2 : Lsn(0x30),
8071 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8072 2 : ),
8073 2 : (
8074 2 : key,
8075 2 : Lsn(0x40),
8076 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
8077 2 : ),
8078 2 : (
8079 2 : key,
8080 2 : Lsn(0x50),
8081 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
8082 2 : ),
8083 2 : (
8084 2 : key,
8085 2 : Lsn(0x60),
8086 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8087 2 : ),
8088 2 : (
8089 2 : key,
8090 2 : Lsn(0x70),
8091 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8092 2 : ),
8093 2 : (
8094 2 : key,
8095 2 : Lsn(0x80),
8096 2 : Value::Image(Bytes::copy_from_slice(
8097 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8098 2 : )),
8099 2 : ),
8100 2 : (
8101 2 : key,
8102 2 : Lsn(0x90),
8103 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8104 2 : ),
8105 2 : ];
8106 2 : let res = tline
8107 2 : .generate_key_retention(key, &history, Lsn(0x60), &[Lsn(0x40), Lsn(0x50)], 3, None)
8108 2 : .await
8109 2 : .unwrap();
8110 2 : let expected_res = KeyHistoryRetention {
8111 2 : below_horizon: vec![
8112 2 : (
8113 2 : Lsn(0x40),
8114 2 : KeyLogAtLsn(vec![(
8115 2 : Lsn(0x40),
8116 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
8117 2 : )]),
8118 2 : ),
8119 2 : (
8120 2 : Lsn(0x50),
8121 2 : KeyLogAtLsn(vec![(
8122 2 : Lsn(0x50),
8123 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
8124 2 : )]),
8125 2 : ),
8126 2 : (
8127 2 : Lsn(0x60),
8128 2 : KeyLogAtLsn(vec![(
8129 2 : Lsn(0x60),
8130 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8131 2 : )]),
8132 2 : ),
8133 2 : ],
8134 2 : above_horizon: KeyLogAtLsn(vec![
8135 2 : (
8136 2 : Lsn(0x70),
8137 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8138 2 : ),
8139 2 : (
8140 2 : Lsn(0x80),
8141 2 : Value::Image(Bytes::copy_from_slice(
8142 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8143 2 : )),
8144 2 : ),
8145 2 : (
8146 2 : Lsn(0x90),
8147 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8148 2 : ),
8149 2 : ]),
8150 2 : };
8151 2 : assert_eq!(res, expected_res);
8152 2 :
8153 2 : // In case of branch compaction, the branch itself does not have the full history, and we need to provide
8154 2 : // the ancestor image in the test case.
8155 2 :
8156 2 : let history = vec![
8157 2 : (
8158 2 : key,
8159 2 : Lsn(0x20),
8160 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8161 2 : ),
8162 2 : (
8163 2 : key,
8164 2 : Lsn(0x30),
8165 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8166 2 : ),
8167 2 : (
8168 2 : key,
8169 2 : Lsn(0x40),
8170 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8171 2 : ),
8172 2 : (
8173 2 : key,
8174 2 : Lsn(0x70),
8175 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8176 2 : ),
8177 2 : ];
8178 2 : let res = tline
8179 2 : .generate_key_retention(
8180 2 : key,
8181 2 : &history,
8182 2 : Lsn(0x60),
8183 2 : &[],
8184 2 : 3,
8185 2 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
8186 2 : )
8187 2 : .await
8188 2 : .unwrap();
8189 2 : let expected_res = KeyHistoryRetention {
8190 2 : below_horizon: vec![(
8191 2 : Lsn(0x60),
8192 2 : KeyLogAtLsn(vec![(
8193 2 : Lsn(0x60),
8194 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")), // use the ancestor image to reconstruct the page
8195 2 : )]),
8196 2 : )],
8197 2 : above_horizon: KeyLogAtLsn(vec![(
8198 2 : Lsn(0x70),
8199 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8200 2 : )]),
8201 2 : };
8202 2 : assert_eq!(res, expected_res);
8203 2 :
8204 2 : let history = vec![
8205 2 : (
8206 2 : key,
8207 2 : Lsn(0x20),
8208 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8209 2 : ),
8210 2 : (
8211 2 : key,
8212 2 : Lsn(0x40),
8213 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8214 2 : ),
8215 2 : (
8216 2 : key,
8217 2 : Lsn(0x60),
8218 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8219 2 : ),
8220 2 : (
8221 2 : key,
8222 2 : Lsn(0x70),
8223 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8224 2 : ),
8225 2 : ];
8226 2 : let res = tline
8227 2 : .generate_key_retention(
8228 2 : key,
8229 2 : &history,
8230 2 : Lsn(0x60),
8231 2 : &[Lsn(0x30)],
8232 2 : 3,
8233 2 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
8234 2 : )
8235 2 : .await
8236 2 : .unwrap();
8237 2 : let expected_res = KeyHistoryRetention {
8238 2 : below_horizon: vec![
8239 2 : (
8240 2 : Lsn(0x30),
8241 2 : KeyLogAtLsn(vec![(
8242 2 : Lsn(0x20),
8243 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8244 2 : )]),
8245 2 : ),
8246 2 : (
8247 2 : Lsn(0x60),
8248 2 : KeyLogAtLsn(vec![(
8249 2 : Lsn(0x60),
8250 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x40;0x60")),
8251 2 : )]),
8252 2 : ),
8253 2 : ],
8254 2 : above_horizon: KeyLogAtLsn(vec![(
8255 2 : Lsn(0x70),
8256 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8257 2 : )]),
8258 2 : };
8259 2 : assert_eq!(res, expected_res);
8260 2 :
8261 2 : Ok(())
8262 2 : }
8263 :
8264 : #[tokio::test]
8265 2 : async fn test_simple_bottom_most_compaction_with_retain_lsns() -> anyhow::Result<()> {
8266 2 : let harness =
8267 2 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns").await?;
8268 10 : let (tenant, ctx) = harness.load().await;
8269 2 :
8270 518 : fn get_key(id: u32) -> Key {
8271 518 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8272 518 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8273 518 : key.field6 = id;
8274 518 : key
8275 518 : }
8276 2 :
8277 2 : let img_layer = (0..10)
8278 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8279 2 : .collect_vec();
8280 2 :
8281 2 : let delta1 = vec![
8282 2 : (
8283 2 : get_key(1),
8284 2 : Lsn(0x20),
8285 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8286 2 : ),
8287 2 : (
8288 2 : get_key(2),
8289 2 : Lsn(0x30),
8290 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8291 2 : ),
8292 2 : (
8293 2 : get_key(3),
8294 2 : Lsn(0x28),
8295 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8296 2 : ),
8297 2 : (
8298 2 : get_key(3),
8299 2 : Lsn(0x30),
8300 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8301 2 : ),
8302 2 : (
8303 2 : get_key(3),
8304 2 : Lsn(0x40),
8305 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
8306 2 : ),
8307 2 : ];
8308 2 : let delta2 = vec![
8309 2 : (
8310 2 : get_key(5),
8311 2 : Lsn(0x20),
8312 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8313 2 : ),
8314 2 : (
8315 2 : get_key(6),
8316 2 : Lsn(0x20),
8317 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8318 2 : ),
8319 2 : ];
8320 2 : let delta3 = vec![
8321 2 : (
8322 2 : get_key(8),
8323 2 : Lsn(0x48),
8324 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8325 2 : ),
8326 2 : (
8327 2 : get_key(9),
8328 2 : Lsn(0x48),
8329 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8330 2 : ),
8331 2 : ];
8332 2 :
8333 2 : let tline = tenant
8334 2 : .create_test_timeline_with_layers(
8335 2 : TIMELINE_ID,
8336 2 : Lsn(0x10),
8337 2 : DEFAULT_PG_VERSION,
8338 2 : &ctx,
8339 2 : vec![
8340 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
8341 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
8342 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
8343 2 : ], // delta layers
8344 2 : vec![(Lsn(0x10), img_layer)], // image layers
8345 2 : Lsn(0x50),
8346 2 : )
8347 47 : .await?;
8348 2 : {
8349 2 : // Update GC info
8350 2 : let mut guard = tline.gc_info.write().unwrap();
8351 2 : *guard = GcInfo {
8352 2 : retain_lsns: vec![
8353 2 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
8354 2 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
8355 2 : ],
8356 2 : cutoffs: GcCutoffs {
8357 2 : time: Lsn(0x30),
8358 2 : space: Lsn(0x30),
8359 2 : },
8360 2 : leases: Default::default(),
8361 2 : within_ancestor_pitr: false,
8362 2 : };
8363 2 : }
8364 2 :
8365 2 : let expected_result = [
8366 2 : Bytes::from_static(b"value 0@0x10"),
8367 2 : Bytes::from_static(b"value 1@0x10@0x20"),
8368 2 : Bytes::from_static(b"value 2@0x10@0x30"),
8369 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
8370 2 : Bytes::from_static(b"value 4@0x10"),
8371 2 : Bytes::from_static(b"value 5@0x10@0x20"),
8372 2 : Bytes::from_static(b"value 6@0x10@0x20"),
8373 2 : Bytes::from_static(b"value 7@0x10"),
8374 2 : Bytes::from_static(b"value 8@0x10@0x48"),
8375 2 : Bytes::from_static(b"value 9@0x10@0x48"),
8376 2 : ];
8377 2 :
8378 2 : let expected_result_at_gc_horizon = [
8379 2 : Bytes::from_static(b"value 0@0x10"),
8380 2 : Bytes::from_static(b"value 1@0x10@0x20"),
8381 2 : Bytes::from_static(b"value 2@0x10@0x30"),
8382 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
8383 2 : Bytes::from_static(b"value 4@0x10"),
8384 2 : Bytes::from_static(b"value 5@0x10@0x20"),
8385 2 : Bytes::from_static(b"value 6@0x10@0x20"),
8386 2 : Bytes::from_static(b"value 7@0x10"),
8387 2 : Bytes::from_static(b"value 8@0x10"),
8388 2 : Bytes::from_static(b"value 9@0x10"),
8389 2 : ];
8390 2 :
8391 2 : let expected_result_at_lsn_20 = [
8392 2 : Bytes::from_static(b"value 0@0x10"),
8393 2 : Bytes::from_static(b"value 1@0x10@0x20"),
8394 2 : Bytes::from_static(b"value 2@0x10"),
8395 2 : Bytes::from_static(b"value 3@0x10"),
8396 2 : Bytes::from_static(b"value 4@0x10"),
8397 2 : Bytes::from_static(b"value 5@0x10@0x20"),
8398 2 : Bytes::from_static(b"value 6@0x10@0x20"),
8399 2 : Bytes::from_static(b"value 7@0x10"),
8400 2 : Bytes::from_static(b"value 8@0x10"),
8401 2 : Bytes::from_static(b"value 9@0x10"),
8402 2 : ];
8403 2 :
8404 2 : let expected_result_at_lsn_10 = [
8405 2 : Bytes::from_static(b"value 0@0x10"),
8406 2 : Bytes::from_static(b"value 1@0x10"),
8407 2 : Bytes::from_static(b"value 2@0x10"),
8408 2 : Bytes::from_static(b"value 3@0x10"),
8409 2 : Bytes::from_static(b"value 4@0x10"),
8410 2 : Bytes::from_static(b"value 5@0x10"),
8411 2 : Bytes::from_static(b"value 6@0x10"),
8412 2 : Bytes::from_static(b"value 7@0x10"),
8413 2 : Bytes::from_static(b"value 8@0x10"),
8414 2 : Bytes::from_static(b"value 9@0x10"),
8415 2 : ];
8416 2 :
8417 12 : let verify_result = || async {
8418 12 : let gc_horizon = {
8419 12 : let gc_info = tline.gc_info.read().unwrap();
8420 12 : gc_info.cutoffs.time
8421 2 : };
8422 132 : for idx in 0..10 {
8423 120 : assert_eq!(
8424 120 : tline
8425 120 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8426 122 : .await
8427 120 : .unwrap(),
8428 120 : &expected_result[idx]
8429 2 : );
8430 120 : assert_eq!(
8431 120 : tline
8432 120 : .get(get_key(idx as u32), gc_horizon, &ctx)
8433 93 : .await
8434 120 : .unwrap(),
8435 120 : &expected_result_at_gc_horizon[idx]
8436 2 : );
8437 120 : assert_eq!(
8438 120 : tline
8439 120 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
8440 85 : .await
8441 120 : .unwrap(),
8442 120 : &expected_result_at_lsn_20[idx]
8443 2 : );
8444 120 : assert_eq!(
8445 120 : tline
8446 120 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
8447 61 : .await
8448 120 : .unwrap(),
8449 120 : &expected_result_at_lsn_10[idx]
8450 2 : );
8451 2 : }
8452 24 : };
8453 2 :
8454 69 : verify_result().await;
8455 2 :
8456 2 : let cancel = CancellationToken::new();
8457 2 : let mut dryrun_flags = EnumSet::new();
8458 2 : dryrun_flags.insert(CompactFlags::DryRun);
8459 2 :
8460 2 : tline
8461 2 : .compact_with_gc(&cancel, dryrun_flags, &ctx)
8462 49 : .await
8463 2 : .unwrap();
8464 2 : // We expect layer map to be the same b/c the dry run flag, but we don't know whether there will be other background jobs
8465 2 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
8466 57 : verify_result().await;
8467 2 :
8468 2 : tline
8469 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8470 50 : .await
8471 2 : .unwrap();
8472 64 : verify_result().await;
8473 2 :
8474 2 : // compact again
8475 2 : tline
8476 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8477 40 : .await
8478 2 : .unwrap();
8479 57 : verify_result().await;
8480 2 :
8481 2 : // increase GC horizon and compact again
8482 2 : {
8483 2 : // Update GC info
8484 2 : let mut guard = tline.gc_info.write().unwrap();
8485 2 : guard.cutoffs.time = Lsn(0x38);
8486 2 : guard.cutoffs.space = Lsn(0x38);
8487 2 : }
8488 2 : tline
8489 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8490 40 : .await
8491 2 : .unwrap();
8492 57 : verify_result().await; // no wals between 0x30 and 0x38, so we should obtain the same result
8493 2 :
8494 2 : // not increasing the GC horizon and compact again
8495 2 : tline
8496 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8497 39 : .await
8498 2 : .unwrap();
8499 57 : verify_result().await;
8500 2 :
8501 2 : Ok(())
8502 2 : }
8503 :
8504 : #[tokio::test]
8505 2 : async fn test_simple_bottom_most_compaction_with_retain_lsns_single_key() -> anyhow::Result<()>
8506 2 : {
8507 2 : let harness =
8508 2 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns_single_key")
8509 2 : .await?;
8510 10 : let (tenant, ctx) = harness.load().await;
8511 2 :
8512 352 : fn get_key(id: u32) -> Key {
8513 352 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8514 352 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8515 352 : key.field6 = id;
8516 352 : key
8517 352 : }
8518 2 :
8519 2 : let img_layer = (0..10)
8520 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8521 2 : .collect_vec();
8522 2 :
8523 2 : let delta1 = vec![
8524 2 : (
8525 2 : get_key(1),
8526 2 : Lsn(0x20),
8527 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8528 2 : ),
8529 2 : (
8530 2 : get_key(1),
8531 2 : Lsn(0x28),
8532 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8533 2 : ),
8534 2 : ];
8535 2 : let delta2 = vec![
8536 2 : (
8537 2 : get_key(1),
8538 2 : Lsn(0x30),
8539 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8540 2 : ),
8541 2 : (
8542 2 : get_key(1),
8543 2 : Lsn(0x38),
8544 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
8545 2 : ),
8546 2 : ];
8547 2 : let delta3 = vec![
8548 2 : (
8549 2 : get_key(8),
8550 2 : Lsn(0x48),
8551 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8552 2 : ),
8553 2 : (
8554 2 : get_key(9),
8555 2 : Lsn(0x48),
8556 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8557 2 : ),
8558 2 : ];
8559 2 :
8560 2 : let tline = tenant
8561 2 : .create_test_timeline_with_layers(
8562 2 : TIMELINE_ID,
8563 2 : Lsn(0x10),
8564 2 : DEFAULT_PG_VERSION,
8565 2 : &ctx,
8566 2 : vec![
8567 2 : // delta1 and delta 2 only contain a single key but multiple updates
8568 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x30), delta1),
8569 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
8570 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x50), delta3),
8571 2 : ], // delta layers
8572 2 : vec![(Lsn(0x10), img_layer)], // image layers
8573 2 : Lsn(0x50),
8574 2 : )
8575 47 : .await?;
8576 2 : {
8577 2 : // Update GC info
8578 2 : let mut guard = tline.gc_info.write().unwrap();
8579 2 : *guard = GcInfo {
8580 2 : retain_lsns: vec![
8581 2 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
8582 2 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
8583 2 : ],
8584 2 : cutoffs: GcCutoffs {
8585 2 : time: Lsn(0x30),
8586 2 : space: Lsn(0x30),
8587 2 : },
8588 2 : leases: Default::default(),
8589 2 : within_ancestor_pitr: false,
8590 2 : };
8591 2 : }
8592 2 :
8593 2 : let expected_result = [
8594 2 : Bytes::from_static(b"value 0@0x10"),
8595 2 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
8596 2 : Bytes::from_static(b"value 2@0x10"),
8597 2 : Bytes::from_static(b"value 3@0x10"),
8598 2 : Bytes::from_static(b"value 4@0x10"),
8599 2 : Bytes::from_static(b"value 5@0x10"),
8600 2 : Bytes::from_static(b"value 6@0x10"),
8601 2 : Bytes::from_static(b"value 7@0x10"),
8602 2 : Bytes::from_static(b"value 8@0x10@0x48"),
8603 2 : Bytes::from_static(b"value 9@0x10@0x48"),
8604 2 : ];
8605 2 :
8606 2 : let expected_result_at_gc_horizon = [
8607 2 : Bytes::from_static(b"value 0@0x10"),
8608 2 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
8609 2 : Bytes::from_static(b"value 2@0x10"),
8610 2 : Bytes::from_static(b"value 3@0x10"),
8611 2 : Bytes::from_static(b"value 4@0x10"),
8612 2 : Bytes::from_static(b"value 5@0x10"),
8613 2 : Bytes::from_static(b"value 6@0x10"),
8614 2 : Bytes::from_static(b"value 7@0x10"),
8615 2 : Bytes::from_static(b"value 8@0x10"),
8616 2 : Bytes::from_static(b"value 9@0x10"),
8617 2 : ];
8618 2 :
8619 2 : let expected_result_at_lsn_20 = [
8620 2 : Bytes::from_static(b"value 0@0x10"),
8621 2 : Bytes::from_static(b"value 1@0x10@0x20"),
8622 2 : Bytes::from_static(b"value 2@0x10"),
8623 2 : Bytes::from_static(b"value 3@0x10"),
8624 2 : Bytes::from_static(b"value 4@0x10"),
8625 2 : Bytes::from_static(b"value 5@0x10"),
8626 2 : Bytes::from_static(b"value 6@0x10"),
8627 2 : Bytes::from_static(b"value 7@0x10"),
8628 2 : Bytes::from_static(b"value 8@0x10"),
8629 2 : Bytes::from_static(b"value 9@0x10"),
8630 2 : ];
8631 2 :
8632 2 : let expected_result_at_lsn_10 = [
8633 2 : Bytes::from_static(b"value 0@0x10"),
8634 2 : Bytes::from_static(b"value 1@0x10"),
8635 2 : Bytes::from_static(b"value 2@0x10"),
8636 2 : Bytes::from_static(b"value 3@0x10"),
8637 2 : Bytes::from_static(b"value 4@0x10"),
8638 2 : Bytes::from_static(b"value 5@0x10"),
8639 2 : Bytes::from_static(b"value 6@0x10"),
8640 2 : Bytes::from_static(b"value 7@0x10"),
8641 2 : Bytes::from_static(b"value 8@0x10"),
8642 2 : Bytes::from_static(b"value 9@0x10"),
8643 2 : ];
8644 2 :
8645 8 : let verify_result = || async {
8646 8 : let gc_horizon = {
8647 8 : let gc_info = tline.gc_info.read().unwrap();
8648 8 : gc_info.cutoffs.time
8649 2 : };
8650 88 : for idx in 0..10 {
8651 80 : assert_eq!(
8652 80 : tline
8653 80 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8654 75 : .await
8655 80 : .unwrap(),
8656 80 : &expected_result[idx]
8657 2 : );
8658 80 : assert_eq!(
8659 80 : tline
8660 80 : .get(get_key(idx as u32), gc_horizon, &ctx)
8661 47 : .await
8662 80 : .unwrap(),
8663 80 : &expected_result_at_gc_horizon[idx]
8664 2 : );
8665 80 : assert_eq!(
8666 80 : tline
8667 80 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
8668 46 : .await
8669 80 : .unwrap(),
8670 80 : &expected_result_at_lsn_20[idx]
8671 2 : );
8672 80 : assert_eq!(
8673 80 : tline
8674 80 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
8675 42 : .await
8676 80 : .unwrap(),
8677 80 : &expected_result_at_lsn_10[idx]
8678 2 : );
8679 2 : }
8680 16 : };
8681 2 :
8682 61 : verify_result().await;
8683 2 :
8684 2 : let cancel = CancellationToken::new();
8685 2 : let mut dryrun_flags = EnumSet::new();
8686 2 : dryrun_flags.insert(CompactFlags::DryRun);
8687 2 :
8688 2 : tline
8689 2 : .compact_with_gc(&cancel, dryrun_flags, &ctx)
8690 50 : .await
8691 2 : .unwrap();
8692 2 : // We expect layer map to be the same b/c the dry run flag, but we don't know whether there will be other background jobs
8693 2 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
8694 49 : verify_result().await;
8695 2 :
8696 2 : tline
8697 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8698 51 : .await
8699 2 : .unwrap();
8700 53 : verify_result().await;
8701 2 :
8702 2 : // compact again
8703 2 : tline
8704 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8705 41 : .await
8706 2 : .unwrap();
8707 47 : verify_result().await;
8708 2 :
8709 2 : Ok(())
8710 2 : }
8711 :
8712 : #[tokio::test]
8713 2 : async fn test_simple_bottom_most_compaction_on_branch() -> anyhow::Result<()> {
8714 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_on_branch").await?;
8715 10 : let (tenant, ctx) = harness.load().await;
8716 2 :
8717 126 : fn get_key(id: u32) -> Key {
8718 126 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
8719 126 : key.field6 = id;
8720 126 : key
8721 126 : }
8722 2 :
8723 2 : let img_layer = (0..10)
8724 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8725 2 : .collect_vec();
8726 2 :
8727 2 : let delta1 = vec![
8728 2 : (
8729 2 : get_key(1),
8730 2 : Lsn(0x20),
8731 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8732 2 : ),
8733 2 : (
8734 2 : get_key(2),
8735 2 : Lsn(0x30),
8736 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8737 2 : ),
8738 2 : (
8739 2 : get_key(3),
8740 2 : Lsn(0x28),
8741 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8742 2 : ),
8743 2 : (
8744 2 : get_key(3),
8745 2 : Lsn(0x30),
8746 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8747 2 : ),
8748 2 : (
8749 2 : get_key(3),
8750 2 : Lsn(0x40),
8751 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
8752 2 : ),
8753 2 : ];
8754 2 : let delta2 = vec![
8755 2 : (
8756 2 : get_key(5),
8757 2 : Lsn(0x20),
8758 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8759 2 : ),
8760 2 : (
8761 2 : get_key(6),
8762 2 : Lsn(0x20),
8763 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8764 2 : ),
8765 2 : ];
8766 2 : let delta3 = vec![
8767 2 : (
8768 2 : get_key(8),
8769 2 : Lsn(0x48),
8770 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8771 2 : ),
8772 2 : (
8773 2 : get_key(9),
8774 2 : Lsn(0x48),
8775 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8776 2 : ),
8777 2 : ];
8778 2 :
8779 2 : let parent_tline = tenant
8780 2 : .create_test_timeline_with_layers(
8781 2 : TIMELINE_ID,
8782 2 : Lsn(0x10),
8783 2 : DEFAULT_PG_VERSION,
8784 2 : &ctx,
8785 2 : vec![], // delta layers
8786 2 : vec![(Lsn(0x18), img_layer)], // image layers
8787 2 : Lsn(0x18),
8788 2 : )
8789 29 : .await?;
8790 2 :
8791 2 : parent_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
8792 2 :
8793 2 : let branch_tline = tenant
8794 2 : .branch_timeline_test_with_layers(
8795 2 : &parent_tline,
8796 2 : NEW_TIMELINE_ID,
8797 2 : Some(Lsn(0x18)),
8798 2 : &ctx,
8799 2 : vec![
8800 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
8801 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
8802 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
8803 2 : ], // delta layers
8804 2 : vec![], // image layers
8805 2 : Lsn(0x50),
8806 2 : )
8807 20 : .await?;
8808 2 :
8809 2 : branch_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
8810 2 :
8811 2 : {
8812 2 : // Update GC info
8813 2 : let mut guard = parent_tline.gc_info.write().unwrap();
8814 2 : *guard = GcInfo {
8815 2 : retain_lsns: vec![(Lsn(0x18), branch_tline.timeline_id, MaybeOffloaded::No)],
8816 2 : cutoffs: GcCutoffs {
8817 2 : time: Lsn(0x10),
8818 2 : space: Lsn(0x10),
8819 2 : },
8820 2 : leases: Default::default(),
8821 2 : within_ancestor_pitr: false,
8822 2 : };
8823 2 : }
8824 2 :
8825 2 : {
8826 2 : // Update GC info
8827 2 : let mut guard = branch_tline.gc_info.write().unwrap();
8828 2 : *guard = GcInfo {
8829 2 : retain_lsns: vec![(Lsn(0x40), branch_tline.timeline_id, MaybeOffloaded::No)],
8830 2 : cutoffs: GcCutoffs {
8831 2 : time: Lsn(0x50),
8832 2 : space: Lsn(0x50),
8833 2 : },
8834 2 : leases: Default::default(),
8835 2 : within_ancestor_pitr: false,
8836 2 : };
8837 2 : }
8838 2 :
8839 2 : let expected_result_at_gc_horizon = [
8840 2 : Bytes::from_static(b"value 0@0x10"),
8841 2 : Bytes::from_static(b"value 1@0x10@0x20"),
8842 2 : Bytes::from_static(b"value 2@0x10@0x30"),
8843 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
8844 2 : Bytes::from_static(b"value 4@0x10"),
8845 2 : Bytes::from_static(b"value 5@0x10@0x20"),
8846 2 : Bytes::from_static(b"value 6@0x10@0x20"),
8847 2 : Bytes::from_static(b"value 7@0x10"),
8848 2 : Bytes::from_static(b"value 8@0x10@0x48"),
8849 2 : Bytes::from_static(b"value 9@0x10@0x48"),
8850 2 : ];
8851 2 :
8852 2 : let expected_result_at_lsn_40 = [
8853 2 : Bytes::from_static(b"value 0@0x10"),
8854 2 : Bytes::from_static(b"value 1@0x10@0x20"),
8855 2 : Bytes::from_static(b"value 2@0x10@0x30"),
8856 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
8857 2 : Bytes::from_static(b"value 4@0x10"),
8858 2 : Bytes::from_static(b"value 5@0x10@0x20"),
8859 2 : Bytes::from_static(b"value 6@0x10@0x20"),
8860 2 : Bytes::from_static(b"value 7@0x10"),
8861 2 : Bytes::from_static(b"value 8@0x10"),
8862 2 : Bytes::from_static(b"value 9@0x10"),
8863 2 : ];
8864 2 :
8865 4 : let verify_result = || async {
8866 44 : for idx in 0..10 {
8867 40 : assert_eq!(
8868 40 : branch_tline
8869 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8870 52 : .await
8871 40 : .unwrap(),
8872 40 : &expected_result_at_gc_horizon[idx]
8873 2 : );
8874 40 : assert_eq!(
8875 40 : branch_tline
8876 40 : .get(get_key(idx as u32), Lsn(0x40), &ctx)
8877 30 : .await
8878 40 : .unwrap(),
8879 40 : &expected_result_at_lsn_40[idx]
8880 2 : );
8881 2 : }
8882 8 : };
8883 2 :
8884 45 : verify_result().await;
8885 2 :
8886 2 : let cancel = CancellationToken::new();
8887 2 : branch_tline
8888 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8889 19 : .await
8890 2 : .unwrap();
8891 2 :
8892 37 : verify_result().await;
8893 2 :
8894 2 : Ok(())
8895 2 : }
8896 :
8897 : // Regression test for https://github.com/neondatabase/neon/issues/9012
8898 : // Create an image arrangement where we have to read at different LSN ranges
8899 : // from a delta layer. This is achieved by overlapping an image layer on top of
8900 : // a delta layer. Like so:
8901 : //
8902 : // A B
8903 : // +----------------+ -> delta_layer
8904 : // | | ^ lsn
8905 : // | =========|-> nested_image_layer |
8906 : // | C | |
8907 : // +----------------+ |
8908 : // ======== -> baseline_image_layer +-------> key
8909 : //
8910 : //
8911 : // When querying the key range [A, B) we need to read at different LSN ranges
8912 : // for [A, C) and [C, B). This test checks that the described edge case is handled correctly.
8913 : #[tokio::test]
8914 2 : async fn test_vectored_read_with_nested_image_layer() -> anyhow::Result<()> {
8915 2 : let harness = TenantHarness::create("test_vectored_read_with_nested_image_layer").await?;
8916 10 : let (tenant, ctx) = harness.load().await;
8917 2 :
8918 2 : let will_init_keys = [2, 6];
8919 44 : fn get_key(id: u32) -> Key {
8920 44 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
8921 44 : key.field6 = id;
8922 44 : key
8923 44 : }
8924 2 :
8925 2 : let mut expected_key_values = HashMap::new();
8926 2 :
8927 2 : let baseline_image_layer_lsn = Lsn(0x10);
8928 2 : let mut baseline_img_layer = Vec::new();
8929 12 : for i in 0..5 {
8930 10 : let key = get_key(i);
8931 10 : let value = format!("value {i}@{baseline_image_layer_lsn}");
8932 10 :
8933 10 : let removed = expected_key_values.insert(key, value.clone());
8934 10 : assert!(removed.is_none());
8935 2 :
8936 10 : baseline_img_layer.push((key, Bytes::from(value)));
8937 2 : }
8938 2 :
8939 2 : let nested_image_layer_lsn = Lsn(0x50);
8940 2 : let mut nested_img_layer = Vec::new();
8941 12 : for i in 5..10 {
8942 10 : let key = get_key(i);
8943 10 : let value = format!("value {i}@{nested_image_layer_lsn}");
8944 10 :
8945 10 : let removed = expected_key_values.insert(key, value.clone());
8946 10 : assert!(removed.is_none());
8947 2 :
8948 10 : nested_img_layer.push((key, Bytes::from(value)));
8949 2 : }
8950 2 :
8951 2 : let mut delta_layer_spec = Vec::default();
8952 2 : let delta_layer_start_lsn = Lsn(0x20);
8953 2 : let mut delta_layer_end_lsn = delta_layer_start_lsn;
8954 2 :
8955 22 : for i in 0..10 {
8956 20 : let key = get_key(i);
8957 20 : let key_in_nested = nested_img_layer
8958 20 : .iter()
8959 80 : .any(|(key_with_img, _)| *key_with_img == key);
8960 20 : let lsn = {
8961 20 : if key_in_nested {
8962 10 : Lsn(nested_image_layer_lsn.0 + 0x10)
8963 2 : } else {
8964 10 : delta_layer_start_lsn
8965 2 : }
8966 2 : };
8967 2 :
8968 20 : let will_init = will_init_keys.contains(&i);
8969 20 : if will_init {
8970 4 : delta_layer_spec.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init())));
8971 4 :
8972 4 : expected_key_values.insert(key, "".to_string());
8973 16 : } else {
8974 16 : let delta = format!("@{lsn}");
8975 16 : delta_layer_spec.push((
8976 16 : key,
8977 16 : lsn,
8978 16 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
8979 16 : ));
8980 16 :
8981 16 : expected_key_values
8982 16 : .get_mut(&key)
8983 16 : .expect("An image exists for each key")
8984 16 : .push_str(delta.as_str());
8985 16 : }
8986 20 : delta_layer_end_lsn = std::cmp::max(delta_layer_start_lsn, lsn);
8987 2 : }
8988 2 :
8989 2 : delta_layer_end_lsn = Lsn(delta_layer_end_lsn.0 + 1);
8990 2 :
8991 2 : assert!(
8992 2 : nested_image_layer_lsn > delta_layer_start_lsn
8993 2 : && nested_image_layer_lsn < delta_layer_end_lsn
8994 2 : );
8995 2 :
8996 2 : let tline = tenant
8997 2 : .create_test_timeline_with_layers(
8998 2 : TIMELINE_ID,
8999 2 : baseline_image_layer_lsn,
9000 2 : DEFAULT_PG_VERSION,
9001 2 : &ctx,
9002 2 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
9003 2 : delta_layer_start_lsn..delta_layer_end_lsn,
9004 2 : delta_layer_spec,
9005 2 : )], // delta layers
9006 2 : vec![
9007 2 : (baseline_image_layer_lsn, baseline_img_layer),
9008 2 : (nested_image_layer_lsn, nested_img_layer),
9009 2 : ], // image layers
9010 2 : delta_layer_end_lsn,
9011 2 : )
9012 40 : .await?;
9013 2 :
9014 2 : let keyspace = KeySpace::single(get_key(0)..get_key(10));
9015 2 : let results = tline
9016 2 : .get_vectored(keyspace, delta_layer_end_lsn, &ctx)
9017 13 : .await
9018 2 : .expect("No vectored errors");
9019 22 : for (key, res) in results {
9020 20 : let value = res.expect("No key errors");
9021 20 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
9022 20 : assert_eq!(value, Bytes::from(expected_value));
9023 2 : }
9024 2 :
9025 2 : Ok(())
9026 2 : }
9027 : }
|
