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::task_mgr;
96 : use crate::task_mgr::TaskKind;
97 : use crate::tenant::config::LocationMode;
98 : use crate::tenant::config::TenantConfOpt;
99 : use crate::tenant::gc_result::GcResult;
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 : mod gc_result;
164 : pub(crate) mod throttle;
165 :
166 : pub(crate) use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
167 : pub(crate) use timeline::{LogicalSizeCalculationCause, PageReconstructError, Timeline};
168 :
169 : // re-export for use in walreceiver
170 : pub use crate::tenant::timeline::WalReceiverInfo;
171 :
172 : /// The "tenants" part of `tenants/<tenant>/timelines...`
173 : pub const TENANTS_SEGMENT_NAME: &str = "tenants";
174 :
175 : /// Parts of the `.neon/tenants/<tenant_id>/timelines/<timeline_id>` directory prefix.
176 : pub const TIMELINES_SEGMENT_NAME: &str = "timelines";
177 :
178 : /// References to shared objects that are passed into each tenant, such
179 : /// as the shared remote storage client and process initialization state.
180 : #[derive(Clone)]
181 : pub struct TenantSharedResources {
182 : pub broker_client: storage_broker::BrokerClientChannel,
183 : pub remote_storage: GenericRemoteStorage,
184 : pub deletion_queue_client: DeletionQueueClient,
185 : pub l0_flush_global_state: L0FlushGlobalState,
186 : }
187 :
188 : /// A [`Tenant`] is really an _attached_ tenant. The configuration
189 : /// for an attached tenant is a subset of the [`LocationConf`], represented
190 : /// in this struct.
191 : pub(super) struct AttachedTenantConf {
192 : tenant_conf: TenantConfOpt,
193 : location: AttachedLocationConfig,
194 : /// The deadline before which we are blocked from GC so that
195 : /// leases have a chance to be renewed.
196 : lsn_lease_deadline: Option<tokio::time::Instant>,
197 : }
198 :
199 : impl AttachedTenantConf {
200 190 : fn new(tenant_conf: TenantConfOpt, location: AttachedLocationConfig) -> Self {
201 : // Sets a deadline before which we cannot proceed to GC due to lsn lease.
202 : //
203 : // We do this as the leases mapping are not persisted to disk. By delaying GC by lease
204 : // length, we guarantee that all the leases we granted before will have a chance to renew
205 : // when we run GC for the first time after restart / transition from AttachedMulti to AttachedSingle.
206 190 : let lsn_lease_deadline = if location.attach_mode == AttachmentMode::Single {
207 190 : Some(
208 190 : tokio::time::Instant::now()
209 190 : + tenant_conf
210 190 : .lsn_lease_length
211 190 : .unwrap_or(LsnLease::DEFAULT_LENGTH),
212 190 : )
213 : } else {
214 : // We don't use `lsn_lease_deadline` to delay GC in AttachedMulti and AttachedStale
215 : // because we don't do GC in these modes.
216 0 : None
217 : };
218 :
219 190 : Self {
220 190 : tenant_conf,
221 190 : location,
222 190 : lsn_lease_deadline,
223 190 : }
224 190 : }
225 :
226 190 : fn try_from(location_conf: LocationConf) -> anyhow::Result<Self> {
227 190 : match &location_conf.mode {
228 190 : LocationMode::Attached(attach_conf) => {
229 190 : Ok(Self::new(location_conf.tenant_conf, *attach_conf))
230 : }
231 : LocationMode::Secondary(_) => {
232 0 : anyhow::bail!("Attempted to construct AttachedTenantConf from a LocationConf in secondary mode")
233 : }
234 : }
235 190 : }
236 :
237 762 : fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
238 762 : self.lsn_lease_deadline
239 762 : .map(|d| tokio::time::Instant::now() < d)
240 762 : .unwrap_or(false)
241 762 : }
242 : }
243 : struct TimelinePreload {
244 : timeline_id: TimelineId,
245 : client: RemoteTimelineClient,
246 : index_part: Result<MaybeDeletedIndexPart, DownloadError>,
247 : }
248 :
249 : pub(crate) struct TenantPreload {
250 : tenant_manifest: TenantManifest,
251 : timelines: HashMap<TimelineId, TimelinePreload>,
252 : }
253 :
254 : /// When we spawn a tenant, there is a special mode for tenant creation that
255 : /// avoids trying to read anything from remote storage.
256 : pub(crate) enum SpawnMode {
257 : /// Activate as soon as possible
258 : Eager,
259 : /// Lazy activation in the background, with the option to skip the queue if the need comes up
260 : Lazy,
261 : }
262 :
263 : ///
264 : /// Tenant consists of multiple timelines. Keep them in a hash table.
265 : ///
266 : pub struct Tenant {
267 : // Global pageserver config parameters
268 : pub conf: &'static PageServerConf,
269 :
270 : /// The value creation timestamp, used to measure activation delay, see:
271 : /// <https://github.com/neondatabase/neon/issues/4025>
272 : constructed_at: Instant,
273 :
274 : state: watch::Sender<TenantState>,
275 :
276 : // Overridden tenant-specific config parameters.
277 : // We keep TenantConfOpt sturct here to preserve the information
278 : // about parameters that are not set.
279 : // This is necessary to allow global config updates.
280 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
281 :
282 : tenant_shard_id: TenantShardId,
283 :
284 : // The detailed sharding information, beyond the number/count in tenant_shard_id
285 : shard_identity: ShardIdentity,
286 :
287 : /// The remote storage generation, used to protect S3 objects from split-brain.
288 : /// Does not change over the lifetime of the [`Tenant`] object.
289 : ///
290 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
291 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
292 : generation: Generation,
293 :
294 : timelines: Mutex<HashMap<TimelineId, Arc<Timeline>>>,
295 :
296 : /// During timeline creation, we first insert the TimelineId to the
297 : /// creating map, then `timelines`, then remove it from the creating map.
298 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
299 : timelines_creating: std::sync::Mutex<HashSet<TimelineId>>,
300 :
301 : /// Possibly offloaded and archived timelines
302 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
303 : timelines_offloaded: Mutex<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
304 :
305 : /// Serialize writes of the tenant manifest to remote storage. If there are concurrent operations
306 : /// affecting the manifest, such as timeline deletion and timeline offload, they must wait for
307 : /// each other (this could be optimized to coalesce writes if necessary).
308 : ///
309 : /// The contents of the Mutex are the last manifest we successfully uploaded
310 : tenant_manifest_upload: tokio::sync::Mutex<Option<TenantManifest>>,
311 :
312 : // This mutex prevents creation of new timelines during GC.
313 : // Adding yet another mutex (in addition to `timelines`) is needed because holding
314 : // `timelines` mutex during all GC iteration
315 : // may block for a long time `get_timeline`, `get_timelines_state`,... and other operations
316 : // with timelines, which in turn may cause dropping replication connection, expiration of wait_for_lsn
317 : // timeout...
318 : gc_cs: tokio::sync::Mutex<()>,
319 : walredo_mgr: Option<Arc<WalRedoManager>>,
320 :
321 : // provides access to timeline data sitting in the remote storage
322 : pub(crate) remote_storage: GenericRemoteStorage,
323 :
324 : // Access to global deletion queue for when this tenant wants to schedule a deletion
325 : deletion_queue_client: DeletionQueueClient,
326 :
327 : /// Cached logical sizes updated updated on each [`Tenant::gather_size_inputs`].
328 : cached_logical_sizes: tokio::sync::Mutex<HashMap<(TimelineId, Lsn), u64>>,
329 : cached_synthetic_tenant_size: Arc<AtomicU64>,
330 :
331 : eviction_task_tenant_state: tokio::sync::Mutex<EvictionTaskTenantState>,
332 :
333 : /// Track repeated failures to compact, so that we can back off.
334 : /// Overhead of mutex is acceptable because compaction is done with a multi-second period.
335 : compaction_circuit_breaker: std::sync::Mutex<CircuitBreaker>,
336 :
337 : /// If the tenant is in Activating state, notify this to encourage it
338 : /// to proceed to Active as soon as possible, rather than waiting for lazy
339 : /// background warmup.
340 : pub(crate) activate_now_sem: tokio::sync::Semaphore,
341 :
342 : /// Time it took for the tenant to activate. Zero if not active yet.
343 : attach_wal_lag_cooldown: Arc<std::sync::OnceLock<WalLagCooldown>>,
344 :
345 : // Cancellation token fires when we have entered shutdown(). This is a parent of
346 : // Timelines' cancellation token.
347 : pub(crate) cancel: CancellationToken,
348 :
349 : // Users of the Tenant such as the page service must take this Gate to avoid
350 : // trying to use a Tenant which is shutting down.
351 : pub(crate) gate: Gate,
352 :
353 : /// Throttle applied at the top of [`Timeline::get`].
354 : /// All [`Tenant::timelines`] of a given [`Tenant`] instance share the same [`throttle::Throttle`] instance.
355 : pub(crate) timeline_get_throttle:
356 : Arc<throttle::Throttle<crate::metrics::tenant_throttling::TimelineGet>>,
357 :
358 : /// An ongoing timeline detach concurrency limiter.
359 : ///
360 : /// As a tenant will likely be restarted as part of timeline detach ancestor it makes no sense
361 : /// to have two running at the same time. A different one can be started if an earlier one
362 : /// has failed for whatever reason.
363 : ongoing_timeline_detach: std::sync::Mutex<Option<(TimelineId, utils::completion::Barrier)>>,
364 :
365 : /// `index_part.json` based gc blocking reason tracking.
366 : ///
367 : /// New gc iterations must start a new iteration by acquiring `GcBlock::start` before
368 : /// proceeding.
369 : pub(crate) gc_block: gc_block::GcBlock,
370 :
371 : l0_flush_global_state: L0FlushGlobalState,
372 : }
373 :
374 : impl std::fmt::Debug for Tenant {
375 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
376 0 : write!(f, "{} ({})", self.tenant_shard_id, self.current_state())
377 0 : }
378 : }
379 :
380 : pub(crate) enum WalRedoManager {
381 : Prod(WalredoManagerId, PostgresRedoManager),
382 : #[cfg(test)]
383 : Test(harness::TestRedoManager),
384 : }
385 :
386 0 : #[derive(thiserror::Error, Debug)]
387 : #[error("pageserver is shutting down")]
388 : pub(crate) struct GlobalShutDown;
389 :
390 : impl WalRedoManager {
391 0 : pub(crate) fn new(mgr: PostgresRedoManager) -> Result<Arc<Self>, GlobalShutDown> {
392 0 : let id = WalredoManagerId::next();
393 0 : let arc = Arc::new(Self::Prod(id, mgr));
394 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
395 0 : match &mut *guard {
396 0 : Some(map) => {
397 0 : map.insert(id, Arc::downgrade(&arc));
398 0 : Ok(arc)
399 : }
400 0 : None => Err(GlobalShutDown),
401 : }
402 0 : }
403 : }
404 :
405 : impl Drop for WalRedoManager {
406 10 : fn drop(&mut self) {
407 10 : match self {
408 0 : Self::Prod(id, _) => {
409 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
410 0 : if let Some(map) = &mut *guard {
411 0 : map.remove(id).expect("new() registers, drop() unregisters");
412 0 : }
413 : }
414 : #[cfg(test)]
415 10 : Self::Test(_) => {
416 10 : // Not applicable to test redo manager
417 10 : }
418 : }
419 10 : }
420 : }
421 :
422 : /// Global registry of all walredo managers so that [`crate::shutdown_pageserver`] can shut down
423 : /// the walredo processes outside of the regular order.
424 : ///
425 : /// This is necessary to work around a systemd bug where it freezes if there are
426 : /// walredo processes left => <https://github.com/neondatabase/cloud/issues/11387>
427 : #[allow(clippy::type_complexity)]
428 : pub(crate) static WALREDO_MANAGERS: once_cell::sync::Lazy<
429 : Mutex<Option<HashMap<WalredoManagerId, Weak<WalRedoManager>>>>,
430 0 : > = once_cell::sync::Lazy::new(|| Mutex::new(Some(HashMap::new())));
431 : #[derive(PartialEq, Eq, Hash, Clone, Copy, Debug)]
432 : pub(crate) struct WalredoManagerId(u64);
433 : impl WalredoManagerId {
434 0 : pub fn next() -> Self {
435 : static NEXT: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(1);
436 0 : let id = NEXT.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
437 0 : if id == 0 {
438 0 : panic!("WalredoManagerId::new() returned 0, indicating wraparound, risking it's no longer unique");
439 0 : }
440 0 : Self(id)
441 0 : }
442 : }
443 :
444 : #[cfg(test)]
445 : impl From<harness::TestRedoManager> for WalRedoManager {
446 190 : fn from(mgr: harness::TestRedoManager) -> Self {
447 190 : Self::Test(mgr)
448 190 : }
449 : }
450 :
451 : impl WalRedoManager {
452 6 : pub(crate) async fn shutdown(&self) -> bool {
453 6 : match self {
454 0 : Self::Prod(_, mgr) => mgr.shutdown().await,
455 : #[cfg(test)]
456 : Self::Test(_) => {
457 : // Not applicable to test redo manager
458 6 : true
459 : }
460 : }
461 6 : }
462 :
463 0 : pub(crate) fn maybe_quiesce(&self, idle_timeout: Duration) {
464 0 : match self {
465 0 : Self::Prod(_, mgr) => mgr.maybe_quiesce(idle_timeout),
466 0 : #[cfg(test)]
467 0 : Self::Test(_) => {
468 0 : // Not applicable to test redo manager
469 0 : }
470 0 : }
471 0 : }
472 :
473 : /// # Cancel-Safety
474 : ///
475 : /// This method is cancellation-safe.
476 410 : pub async fn request_redo(
477 410 : &self,
478 410 : key: pageserver_api::key::Key,
479 410 : lsn: Lsn,
480 410 : base_img: Option<(Lsn, bytes::Bytes)>,
481 410 : records: Vec<(Lsn, pageserver_api::record::NeonWalRecord)>,
482 410 : pg_version: u32,
483 410 : ) -> Result<bytes::Bytes, walredo::Error> {
484 410 : match self {
485 0 : Self::Prod(_, mgr) => {
486 0 : mgr.request_redo(key, lsn, base_img, records, pg_version)
487 0 : .await
488 : }
489 : #[cfg(test)]
490 410 : Self::Test(mgr) => {
491 410 : mgr.request_redo(key, lsn, base_img, records, pg_version)
492 0 : .await
493 : }
494 : }
495 410 : }
496 :
497 0 : pub(crate) fn status(&self) -> Option<WalRedoManagerStatus> {
498 0 : match self {
499 0 : WalRedoManager::Prod(_, m) => Some(m.status()),
500 0 : #[cfg(test)]
501 0 : WalRedoManager::Test(_) => None,
502 0 : }
503 0 : }
504 : }
505 :
506 : /// A very lightweight memory representation of an offloaded timeline.
507 : ///
508 : /// We need to store the list of offloaded timelines so that we can perform operations on them,
509 : /// like unoffloading them, or (at a later date), decide to perform flattening.
510 : /// This type has a much smaller memory impact than [`Timeline`], and thus we can store many
511 : /// more offloaded timelines than we can manage ones that aren't.
512 : pub struct OffloadedTimeline {
513 : pub tenant_shard_id: TenantShardId,
514 : pub timeline_id: TimelineId,
515 : pub ancestor_timeline_id: Option<TimelineId>,
516 : /// Whether to retain the branch lsn at the ancestor or not
517 : pub ancestor_retain_lsn: Option<Lsn>,
518 :
519 : /// When the timeline was archived.
520 : ///
521 : /// Present for future flattening deliberations.
522 : pub archived_at: NaiveDateTime,
523 :
524 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
525 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
526 : pub delete_progress: TimelineDeleteProgress,
527 : }
528 :
529 : impl OffloadedTimeline {
530 : /// Obtains an offloaded timeline from a given timeline object.
531 : ///
532 : /// Returns `None` if the `archived_at` flag couldn't be obtained, i.e.
533 : /// the timeline is not in a stopped state.
534 : /// Panics if the timeline is not archived.
535 0 : fn from_timeline(timeline: &Timeline) -> Result<Self, UploadQueueNotReadyError> {
536 0 : let ancestor_retain_lsn = timeline
537 0 : .get_ancestor_timeline_id()
538 0 : .map(|_timeline_id| timeline.get_ancestor_lsn());
539 0 : let archived_at = timeline
540 0 : .remote_client
541 0 : .archived_at_stopped_queue()?
542 0 : .expect("must be called on an archived timeline");
543 0 : Ok(Self {
544 0 : tenant_shard_id: timeline.tenant_shard_id,
545 0 : timeline_id: timeline.timeline_id,
546 0 : ancestor_timeline_id: timeline.get_ancestor_timeline_id(),
547 0 : ancestor_retain_lsn,
548 0 : archived_at,
549 0 :
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 : delete_progress: TimelineDeleteProgress::default(),
567 0 : }
568 0 : }
569 0 : fn manifest(&self) -> OffloadedTimelineManifest {
570 0 : let Self {
571 0 : timeline_id,
572 0 : ancestor_timeline_id,
573 0 : ancestor_retain_lsn,
574 0 : archived_at,
575 0 : ..
576 0 : } = self;
577 0 : OffloadedTimelineManifest {
578 0 : timeline_id: *timeline_id,
579 0 : ancestor_timeline_id: *ancestor_timeline_id,
580 0 : ancestor_retain_lsn: *ancestor_retain_lsn,
581 0 : archived_at: *archived_at,
582 0 : }
583 0 : }
584 : }
585 :
586 : impl fmt::Debug for OffloadedTimeline {
587 0 : fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
588 0 : write!(f, "OffloadedTimeline<{}>", self.timeline_id)
589 0 : }
590 : }
591 :
592 : #[derive(Copy, Clone, PartialEq, Eq, Hash, Debug)]
593 : pub enum MaybeOffloaded {
594 : Yes,
595 : No,
596 : }
597 :
598 : #[derive(Clone, Debug)]
599 : pub enum TimelineOrOffloaded {
600 : Timeline(Arc<Timeline>),
601 : Offloaded(Arc<OffloadedTimeline>),
602 : }
603 :
604 : impl TimelineOrOffloaded {
605 0 : pub fn arc_ref(&self) -> TimelineOrOffloadedArcRef<'_> {
606 0 : match self {
607 0 : TimelineOrOffloaded::Timeline(timeline) => {
608 0 : TimelineOrOffloadedArcRef::Timeline(timeline)
609 : }
610 0 : TimelineOrOffloaded::Offloaded(offloaded) => {
611 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded)
612 : }
613 : }
614 0 : }
615 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
616 0 : self.arc_ref().tenant_shard_id()
617 0 : }
618 0 : pub fn timeline_id(&self) -> TimelineId {
619 0 : self.arc_ref().timeline_id()
620 0 : }
621 0 : pub fn delete_progress(&self) -> &Arc<tokio::sync::Mutex<DeleteTimelineFlow>> {
622 0 : match self {
623 0 : TimelineOrOffloaded::Timeline(timeline) => &timeline.delete_progress,
624 0 : TimelineOrOffloaded::Offloaded(offloaded) => &offloaded.delete_progress,
625 : }
626 0 : }
627 0 : fn maybe_remote_client(&self) -> Option<Arc<RemoteTimelineClient>> {
628 0 : match self {
629 0 : TimelineOrOffloaded::Timeline(timeline) => Some(timeline.remote_client.clone()),
630 0 : TimelineOrOffloaded::Offloaded(_offloaded) => None,
631 : }
632 0 : }
633 : }
634 :
635 : pub enum TimelineOrOffloadedArcRef<'a> {
636 : Timeline(&'a Arc<Timeline>),
637 : Offloaded(&'a Arc<OffloadedTimeline>),
638 : }
639 :
640 : impl TimelineOrOffloadedArcRef<'_> {
641 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
642 0 : match self {
643 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.tenant_shard_id,
644 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.tenant_shard_id,
645 : }
646 0 : }
647 0 : pub fn timeline_id(&self) -> TimelineId {
648 0 : match self {
649 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.timeline_id,
650 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.timeline_id,
651 : }
652 0 : }
653 : }
654 :
655 : impl<'a> From<&'a Arc<Timeline>> for TimelineOrOffloadedArcRef<'a> {
656 0 : fn from(timeline: &'a Arc<Timeline>) -> Self {
657 0 : Self::Timeline(timeline)
658 0 : }
659 : }
660 :
661 : impl<'a> From<&'a Arc<OffloadedTimeline>> for TimelineOrOffloadedArcRef<'a> {
662 0 : fn from(timeline: &'a Arc<OffloadedTimeline>) -> Self {
663 0 : Self::Offloaded(timeline)
664 0 : }
665 : }
666 :
667 0 : #[derive(Debug, thiserror::Error, PartialEq, Eq)]
668 : pub enum GetTimelineError {
669 : #[error("Timeline is shutting down")]
670 : ShuttingDown,
671 : #[error("Timeline {tenant_id}/{timeline_id} is not active, state: {state:?}")]
672 : NotActive {
673 : tenant_id: TenantShardId,
674 : timeline_id: TimelineId,
675 : state: TimelineState,
676 : },
677 : #[error("Timeline {tenant_id}/{timeline_id} was not found")]
678 : NotFound {
679 : tenant_id: TenantShardId,
680 : timeline_id: TimelineId,
681 : },
682 : }
683 :
684 0 : #[derive(Debug, thiserror::Error)]
685 : pub enum LoadLocalTimelineError {
686 : #[error("FailedToLoad")]
687 : Load(#[source] anyhow::Error),
688 : #[error("FailedToResumeDeletion")]
689 : ResumeDeletion(#[source] anyhow::Error),
690 : }
691 :
692 0 : #[derive(thiserror::Error)]
693 : pub enum DeleteTimelineError {
694 : #[error("NotFound")]
695 : NotFound,
696 :
697 : #[error("HasChildren")]
698 : HasChildren(Vec<TimelineId>),
699 :
700 : #[error("Timeline deletion is already in progress")]
701 : AlreadyInProgress(Arc<tokio::sync::Mutex<DeleteTimelineFlow>>),
702 :
703 : #[error(transparent)]
704 : Other(#[from] anyhow::Error),
705 : }
706 :
707 : impl Debug for DeleteTimelineError {
708 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
709 0 : match self {
710 0 : Self::NotFound => write!(f, "NotFound"),
711 0 : Self::HasChildren(c) => f.debug_tuple("HasChildren").field(c).finish(),
712 0 : Self::AlreadyInProgress(_) => f.debug_tuple("AlreadyInProgress").finish(),
713 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
714 : }
715 0 : }
716 : }
717 :
718 0 : #[derive(thiserror::Error)]
719 : pub enum TimelineArchivalError {
720 : #[error("NotFound")]
721 : NotFound,
722 :
723 : #[error("Timeout")]
724 : Timeout,
725 :
726 : #[error("Cancelled")]
727 : Cancelled,
728 :
729 : #[error("ancestor is archived: {}", .0)]
730 : HasArchivedParent(TimelineId),
731 :
732 : #[error("HasUnarchivedChildren")]
733 : HasUnarchivedChildren(Vec<TimelineId>),
734 :
735 : #[error("Timeline archival is already in progress")]
736 : AlreadyInProgress,
737 :
738 : #[error(transparent)]
739 : Other(anyhow::Error),
740 : }
741 :
742 0 : #[derive(thiserror::Error, Debug)]
743 : pub(crate) enum TenantManifestError {
744 : #[error("Remote storage error: {0}")]
745 : RemoteStorage(anyhow::Error),
746 :
747 : #[error("Cancelled")]
748 : Cancelled,
749 : }
750 :
751 : impl From<TenantManifestError> for TimelineArchivalError {
752 0 : fn from(e: TenantManifestError) -> Self {
753 0 : match e {
754 0 : TenantManifestError::RemoteStorage(e) => Self::Other(e),
755 0 : TenantManifestError::Cancelled => Self::Cancelled,
756 : }
757 0 : }
758 : }
759 :
760 : impl Debug for TimelineArchivalError {
761 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
762 0 : match self {
763 0 : Self::NotFound => write!(f, "NotFound"),
764 0 : Self::Timeout => write!(f, "Timeout"),
765 0 : Self::Cancelled => write!(f, "Cancelled"),
766 0 : Self::HasArchivedParent(p) => f.debug_tuple("HasArchivedParent").field(p).finish(),
767 0 : Self::HasUnarchivedChildren(c) => {
768 0 : f.debug_tuple("HasUnarchivedChildren").field(c).finish()
769 : }
770 0 : Self::AlreadyInProgress => f.debug_tuple("AlreadyInProgress").finish(),
771 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
772 : }
773 0 : }
774 : }
775 :
776 : pub enum SetStoppingError {
777 : AlreadyStopping(completion::Barrier),
778 : Broken,
779 : }
780 :
781 : impl Debug for SetStoppingError {
782 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
783 0 : match self {
784 0 : Self::AlreadyStopping(_) => f.debug_tuple("AlreadyStopping").finish(),
785 0 : Self::Broken => write!(f, "Broken"),
786 : }
787 0 : }
788 : }
789 :
790 : /// Arguments to [`Tenant::create_timeline`].
791 : ///
792 : /// Not usable as an idempotency key for timeline creation because if [`CreateTimelineParamsBranch::ancestor_start_lsn`]
793 : /// is `None`, the result of the timeline create call is not deterministic.
794 : ///
795 : /// See [`CreateTimelineIdempotency`] for an idempotency key.
796 : #[derive(Debug)]
797 : pub(crate) enum CreateTimelineParams {
798 : Bootstrap(CreateTimelineParamsBootstrap),
799 : Branch(CreateTimelineParamsBranch),
800 : }
801 :
802 : #[derive(Debug)]
803 : pub(crate) struct CreateTimelineParamsBootstrap {
804 : pub(crate) new_timeline_id: TimelineId,
805 : pub(crate) existing_initdb_timeline_id: Option<TimelineId>,
806 : pub(crate) pg_version: u32,
807 : }
808 :
809 : /// NB: See comment on [`CreateTimelineIdempotency::Branch`] for why there's no `pg_version` here.
810 : #[derive(Debug)]
811 : pub(crate) struct CreateTimelineParamsBranch {
812 : pub(crate) new_timeline_id: TimelineId,
813 : pub(crate) ancestor_timeline_id: TimelineId,
814 : pub(crate) ancestor_start_lsn: Option<Lsn>,
815 : }
816 :
817 : /// What is used to determine idempotency of a [`Tenant::create_timeline`] call in [`Tenant::start_creating_timeline`].
818 : ///
819 : /// Each [`Timeline`] object holds [`Self`] as an immutable property in [`Timeline::create_idempotency`].
820 : ///
821 : /// We lower timeline creation requests to [`Self`], and then use [`PartialEq::eq`] to compare [`Timeline::create_idempotency`] with the request.
822 : /// If they are equal, we return a reference to the existing timeline, otherwise it's an idempotency conflict.
823 : ///
824 : /// There is special treatment for [`Self::FailWithConflict`] to always return an idempotency conflict.
825 : /// It would be nice to have more advanced derive macros to make that special treatment declarative.
826 : ///
827 : /// Notes:
828 : /// - Unlike [`CreateTimelineParams`], ancestor LSN is fixed, so, branching will be at a deterministic LSN.
829 : /// - We make some trade-offs though, e.g., [`CreateTimelineParamsBootstrap::existing_initdb_timeline_id`]
830 : /// is not considered for idempotency. We can improve on this over time if we deem it necessary.
831 : ///
832 : #[derive(Debug, Clone, PartialEq, Eq)]
833 : pub(crate) enum CreateTimelineIdempotency {
834 : /// NB: special treatment, see comment in [`Self`].
835 : FailWithConflict,
836 : Bootstrap {
837 : pg_version: u32,
838 : },
839 : /// NB: branches always have the same `pg_version` as their ancestor.
840 : /// While [`pageserver_api::models::TimelineCreateRequestMode::Branch::pg_version`]
841 : /// exists as a field, and is set by cplane, it has always been ignored by pageserver when
842 : /// determining the child branch pg_version.
843 : Branch {
844 : ancestor_timeline_id: TimelineId,
845 : ancestor_start_lsn: Lsn,
846 : },
847 : }
848 :
849 : /// What is returned by [`Tenant::start_creating_timeline`].
850 : #[must_use]
851 : enum StartCreatingTimelineResult<'t> {
852 : CreateGuard(TimelineCreateGuard<'t>),
853 : Idempotent(Arc<Timeline>),
854 : }
855 :
856 : /// What is returned by [`Tenant::create_timeline`].
857 : enum CreateTimelineResult {
858 : Created(Arc<Timeline>),
859 : Idempotent(Arc<Timeline>),
860 : }
861 :
862 : impl CreateTimelineResult {
863 0 : fn discriminant(&self) -> &'static str {
864 0 : match self {
865 0 : Self::Created(_) => "Created",
866 0 : Self::Idempotent(_) => "Idempotent",
867 : }
868 0 : }
869 0 : fn timeline(&self) -> &Arc<Timeline> {
870 0 : match self {
871 0 : Self::Created(t) | Self::Idempotent(t) => t,
872 0 : }
873 0 : }
874 : /// Unit test timelines aren't activated, test has to do it if it needs to.
875 : #[cfg(test)]
876 228 : fn into_timeline_for_test(self) -> Arc<Timeline> {
877 228 : match self {
878 228 : Self::Created(t) | Self::Idempotent(t) => t,
879 228 : }
880 228 : }
881 : }
882 :
883 2 : #[derive(thiserror::Error, Debug)]
884 : pub enum CreateTimelineError {
885 : #[error("creation of timeline with the given ID is in progress")]
886 : AlreadyCreating,
887 : #[error("timeline already exists with different parameters")]
888 : Conflict,
889 : #[error(transparent)]
890 : AncestorLsn(anyhow::Error),
891 : #[error("ancestor timeline is not active")]
892 : AncestorNotActive,
893 : #[error("ancestor timeline is archived")]
894 : AncestorArchived,
895 : #[error("tenant shutting down")]
896 : ShuttingDown,
897 : #[error(transparent)]
898 : Other(#[from] anyhow::Error),
899 : }
900 :
901 : #[derive(thiserror::Error, Debug)]
902 : enum InitdbError {
903 : Other(anyhow::Error),
904 : Cancelled,
905 : Spawn(std::io::Result<()>),
906 : Failed(std::process::ExitStatus, Vec<u8>),
907 : }
908 :
909 : impl fmt::Display for InitdbError {
910 0 : fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
911 0 : match self {
912 0 : InitdbError::Cancelled => write!(f, "Operation was cancelled"),
913 0 : InitdbError::Spawn(e) => write!(f, "Spawn error: {:?}", e),
914 0 : InitdbError::Failed(status, stderr) => write!(
915 0 : f,
916 0 : "Command failed with status {:?}: {}",
917 0 : status,
918 0 : String::from_utf8_lossy(stderr)
919 0 : ),
920 0 : InitdbError::Other(e) => write!(f, "Error: {:?}", e),
921 : }
922 0 : }
923 : }
924 :
925 : impl From<std::io::Error> for InitdbError {
926 0 : fn from(error: std::io::Error) -> Self {
927 0 : InitdbError::Spawn(Err(error))
928 0 : }
929 : }
930 :
931 : enum CreateTimelineCause {
932 : Load,
933 : Delete,
934 : }
935 :
936 0 : #[derive(thiserror::Error, Debug)]
937 : pub(crate) enum GcError {
938 : // The tenant is shutting down
939 : #[error("tenant shutting down")]
940 : TenantCancelled,
941 :
942 : // The tenant is shutting down
943 : #[error("timeline shutting down")]
944 : TimelineCancelled,
945 :
946 : // The tenant is in a state inelegible to run GC
947 : #[error("not active")]
948 : NotActive,
949 :
950 : // A requested GC cutoff LSN was invalid, for example it tried to move backwards
951 : #[error("not active")]
952 : BadLsn { why: String },
953 :
954 : // A remote storage error while scheduling updates after compaction
955 : #[error(transparent)]
956 : Remote(anyhow::Error),
957 :
958 : // An error reading while calculating GC cutoffs
959 : #[error(transparent)]
960 : GcCutoffs(PageReconstructError),
961 :
962 : // If GC was invoked for a particular timeline, this error means it didn't exist
963 : #[error("timeline not found")]
964 : TimelineNotFound,
965 : }
966 :
967 : impl From<PageReconstructError> for GcError {
968 0 : fn from(value: PageReconstructError) -> Self {
969 0 : match value {
970 0 : PageReconstructError::Cancelled => Self::TimelineCancelled,
971 0 : other => Self::GcCutoffs(other),
972 : }
973 0 : }
974 : }
975 :
976 : impl From<NotInitialized> for GcError {
977 0 : fn from(value: NotInitialized) -> Self {
978 0 : match value {
979 0 : NotInitialized::Uninitialized => GcError::Remote(value.into()),
980 0 : NotInitialized::Stopped | NotInitialized::ShuttingDown => GcError::TimelineCancelled,
981 : }
982 0 : }
983 : }
984 :
985 : impl From<timeline::layer_manager::Shutdown> for GcError {
986 0 : fn from(_: timeline::layer_manager::Shutdown) -> Self {
987 0 : GcError::TimelineCancelled
988 0 : }
989 : }
990 :
991 0 : #[derive(thiserror::Error, Debug)]
992 : pub(crate) enum LoadConfigError {
993 : #[error("TOML deserialization error: '{0}'")]
994 : DeserializeToml(#[from] toml_edit::de::Error),
995 :
996 : #[error("Config not found at {0}")]
997 : NotFound(Utf8PathBuf),
998 : }
999 :
1000 : impl Tenant {
1001 : /// Yet another helper for timeline initialization.
1002 : ///
1003 : /// - Initializes the Timeline struct and inserts it into the tenant's hash map
1004 : /// - Scans the local timeline directory for layer files and builds the layer map
1005 : /// - Downloads remote index file and adds remote files to the layer map
1006 : /// - Schedules remote upload tasks for any files that are present locally but missing from remote storage.
1007 : ///
1008 : /// If the operation fails, the timeline is left in the tenant's hash map in Broken state. On success,
1009 : /// it is marked as Active.
1010 : #[allow(clippy::too_many_arguments)]
1011 6 : async fn timeline_init_and_sync(
1012 6 : &self,
1013 6 : timeline_id: TimelineId,
1014 6 : resources: TimelineResources,
1015 6 : index_part: IndexPart,
1016 6 : metadata: TimelineMetadata,
1017 6 : ancestor: Option<Arc<Timeline>>,
1018 6 : _ctx: &RequestContext,
1019 6 : ) -> anyhow::Result<()> {
1020 6 : let tenant_id = self.tenant_shard_id;
1021 :
1022 6 : let idempotency = if metadata.ancestor_timeline().is_none() {
1023 4 : CreateTimelineIdempotency::Bootstrap {
1024 4 : pg_version: metadata.pg_version(),
1025 4 : }
1026 : } else {
1027 2 : CreateTimelineIdempotency::Branch {
1028 2 : ancestor_timeline_id: metadata.ancestor_timeline().unwrap(),
1029 2 : ancestor_start_lsn: metadata.ancestor_lsn(),
1030 2 : }
1031 : };
1032 :
1033 6 : let timeline = self.create_timeline_struct(
1034 6 : timeline_id,
1035 6 : &metadata,
1036 6 : ancestor.clone(),
1037 6 : resources,
1038 6 : CreateTimelineCause::Load,
1039 6 : idempotency.clone(),
1040 6 : )?;
1041 6 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
1042 6 : anyhow::ensure!(
1043 6 : disk_consistent_lsn.is_valid(),
1044 0 : "Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn"
1045 : );
1046 6 : assert_eq!(
1047 6 : disk_consistent_lsn,
1048 6 : metadata.disk_consistent_lsn(),
1049 0 : "these are used interchangeably"
1050 : );
1051 :
1052 6 : timeline.remote_client.init_upload_queue(&index_part)?;
1053 :
1054 6 : timeline
1055 6 : .load_layer_map(disk_consistent_lsn, index_part)
1056 0 : .await
1057 6 : .with_context(|| {
1058 0 : format!("Failed to load layermap for timeline {tenant_id}/{timeline_id}")
1059 6 : })?;
1060 :
1061 : {
1062 : // avoiding holding it across awaits
1063 6 : let mut timelines_accessor = self.timelines.lock().unwrap();
1064 6 : match timelines_accessor.entry(timeline_id) {
1065 : // We should never try and load the same timeline twice during startup
1066 : Entry::Occupied(_) => {
1067 0 : unreachable!(
1068 0 : "Timeline {tenant_id}/{timeline_id} already exists in the tenant map"
1069 0 : );
1070 : }
1071 6 : Entry::Vacant(v) => {
1072 6 : v.insert(Arc::clone(&timeline));
1073 6 : timeline.maybe_spawn_flush_loop();
1074 6 : }
1075 6 : }
1076 6 : };
1077 6 :
1078 6 : // Sanity check: a timeline should have some content.
1079 6 : anyhow::ensure!(
1080 6 : ancestor.is_some()
1081 4 : || timeline
1082 4 : .layers
1083 4 : .read()
1084 0 : .await
1085 4 : .layer_map()
1086 4 : .expect("currently loading, layer manager cannot be shutdown already")
1087 4 : .iter_historic_layers()
1088 4 : .next()
1089 4 : .is_some(),
1090 0 : "Timeline has no ancestor and no layer files"
1091 : );
1092 :
1093 6 : Ok(())
1094 6 : }
1095 :
1096 : /// Attach a tenant that's available in cloud storage.
1097 : ///
1098 : /// This returns quickly, after just creating the in-memory object
1099 : /// Tenant struct and launching a background task to download
1100 : /// the remote index files. On return, the tenant is most likely still in
1101 : /// Attaching state, and it will become Active once the background task
1102 : /// finishes. You can use wait_until_active() to wait for the task to
1103 : /// complete.
1104 : ///
1105 : #[allow(clippy::too_many_arguments)]
1106 0 : pub(crate) fn spawn(
1107 0 : conf: &'static PageServerConf,
1108 0 : tenant_shard_id: TenantShardId,
1109 0 : resources: TenantSharedResources,
1110 0 : attached_conf: AttachedTenantConf,
1111 0 : shard_identity: ShardIdentity,
1112 0 : init_order: Option<InitializationOrder>,
1113 0 : mode: SpawnMode,
1114 0 : ctx: &RequestContext,
1115 0 : ) -> Result<Arc<Tenant>, GlobalShutDown> {
1116 0 : let wal_redo_manager =
1117 0 : WalRedoManager::new(PostgresRedoManager::new(conf, tenant_shard_id))?;
1118 :
1119 : let TenantSharedResources {
1120 0 : broker_client,
1121 0 : remote_storage,
1122 0 : deletion_queue_client,
1123 0 : l0_flush_global_state,
1124 0 : } = resources;
1125 0 :
1126 0 : let attach_mode = attached_conf.location.attach_mode;
1127 0 : let generation = attached_conf.location.generation;
1128 0 :
1129 0 : let tenant = Arc::new(Tenant::new(
1130 0 : TenantState::Attaching,
1131 0 : conf,
1132 0 : attached_conf,
1133 0 : shard_identity,
1134 0 : Some(wal_redo_manager),
1135 0 : tenant_shard_id,
1136 0 : remote_storage.clone(),
1137 0 : deletion_queue_client,
1138 0 : l0_flush_global_state,
1139 0 : ));
1140 0 :
1141 0 : // The attach task will carry a GateGuard, so that shutdown() reliably waits for it to drop out if
1142 0 : // we shut down while attaching.
1143 0 : let attach_gate_guard = tenant
1144 0 : .gate
1145 0 : .enter()
1146 0 : .expect("We just created the Tenant: nothing else can have shut it down yet");
1147 0 :
1148 0 : // Do all the hard work in the background
1149 0 : let tenant_clone = Arc::clone(&tenant);
1150 0 : let ctx = ctx.detached_child(TaskKind::Attach, DownloadBehavior::Warn);
1151 0 : task_mgr::spawn(
1152 0 : &tokio::runtime::Handle::current(),
1153 0 : TaskKind::Attach,
1154 0 : tenant_shard_id,
1155 0 : None,
1156 0 : "attach tenant",
1157 0 : async move {
1158 0 :
1159 0 : info!(
1160 : ?attach_mode,
1161 0 : "Attaching tenant"
1162 : );
1163 :
1164 0 : let _gate_guard = attach_gate_guard;
1165 0 :
1166 0 : // Is this tenant being spawned as part of process startup?
1167 0 : let starting_up = init_order.is_some();
1168 0 : scopeguard::defer! {
1169 0 : if starting_up {
1170 0 : TENANT.startup_complete.inc();
1171 0 : }
1172 0 : }
1173 :
1174 : // Ideally we should use Tenant::set_broken_no_wait, but it is not supposed to be used when tenant is in loading state.
1175 : enum BrokenVerbosity {
1176 : Error,
1177 : Info
1178 : }
1179 0 : let make_broken =
1180 0 : |t: &Tenant, err: anyhow::Error, verbosity: BrokenVerbosity| {
1181 0 : match verbosity {
1182 : BrokenVerbosity::Info => {
1183 0 : info!("attach cancelled, setting tenant state to Broken: {err}");
1184 : },
1185 : BrokenVerbosity::Error => {
1186 0 : error!("attach failed, setting tenant state to Broken: {err:?}");
1187 : }
1188 : }
1189 0 : t.state.send_modify(|state| {
1190 0 : // The Stopping case is for when we have passed control on to DeleteTenantFlow:
1191 0 : // if it errors, we will call make_broken when tenant is already in Stopping.
1192 0 : assert!(
1193 0 : matches!(*state, TenantState::Attaching | TenantState::Stopping { .. }),
1194 0 : "the attach task owns the tenant state until activation is complete"
1195 : );
1196 :
1197 0 : *state = TenantState::broken_from_reason(err.to_string());
1198 0 : });
1199 0 : };
1200 :
1201 : // TODO: should also be rejecting tenant conf changes that violate this check.
1202 0 : if let Err(e) = crate::tenant::storage_layer::inmemory_layer::IndexEntry::validate_checkpoint_distance(tenant_clone.get_checkpoint_distance()) {
1203 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1204 0 : return Ok(());
1205 0 : }
1206 0 :
1207 0 : let mut init_order = init_order;
1208 0 : // take the completion because initial tenant loading will complete when all of
1209 0 : // these tasks complete.
1210 0 : let _completion = init_order
1211 0 : .as_mut()
1212 0 : .and_then(|x| x.initial_tenant_load.take());
1213 0 : let remote_load_completion = init_order
1214 0 : .as_mut()
1215 0 : .and_then(|x| x.initial_tenant_load_remote.take());
1216 :
1217 : enum AttachType<'a> {
1218 : /// We are attaching this tenant lazily in the background.
1219 : Warmup {
1220 : _permit: tokio::sync::SemaphorePermit<'a>,
1221 : during_startup: bool
1222 : },
1223 : /// We are attaching this tenant as soon as we can, because for example an
1224 : /// endpoint tried to access it.
1225 : OnDemand,
1226 : /// During normal operations after startup, we are attaching a tenant, and
1227 : /// eager attach was requested.
1228 : Normal,
1229 : }
1230 :
1231 0 : let attach_type = if matches!(mode, SpawnMode::Lazy) {
1232 : // Before doing any I/O, wait for at least one of:
1233 : // - A client attempting to access to this tenant (on-demand loading)
1234 : // - A permit becoming available in the warmup semaphore (background warmup)
1235 :
1236 0 : tokio::select!(
1237 0 : permit = tenant_clone.activate_now_sem.acquire() => {
1238 0 : let _ = permit.expect("activate_now_sem is never closed");
1239 0 : tracing::info!("Activating tenant (on-demand)");
1240 0 : AttachType::OnDemand
1241 : },
1242 0 : permit = conf.concurrent_tenant_warmup.inner().acquire() => {
1243 0 : let _permit = permit.expect("concurrent_tenant_warmup semaphore is never closed");
1244 0 : tracing::info!("Activating tenant (warmup)");
1245 0 : AttachType::Warmup {
1246 0 : _permit,
1247 0 : during_startup: init_order.is_some()
1248 0 : }
1249 : }
1250 0 : _ = tenant_clone.cancel.cancelled() => {
1251 : // This is safe, but should be pretty rare: it is interesting if a tenant
1252 : // stayed in Activating for such a long time that shutdown found it in
1253 : // that state.
1254 0 : tracing::info!(state=%tenant_clone.current_state(), "Tenant shut down before activation");
1255 : // Make the tenant broken so that set_stopping will not hang waiting for it to leave
1256 : // the Attaching state. This is an over-reaction (nothing really broke, the tenant is
1257 : // just shutting down), but ensures progress.
1258 0 : make_broken(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"), BrokenVerbosity::Info);
1259 0 : return Ok(());
1260 : },
1261 : )
1262 : } else {
1263 : // SpawnMode::{Create,Eager} always cause jumping ahead of the
1264 : // concurrent_tenant_warmup queue
1265 0 : AttachType::Normal
1266 : };
1267 :
1268 0 : let preload = match &mode {
1269 : SpawnMode::Eager | SpawnMode::Lazy => {
1270 0 : let _preload_timer = TENANT.preload.start_timer();
1271 0 : let res = tenant_clone
1272 0 : .preload(&remote_storage, task_mgr::shutdown_token())
1273 0 : .await;
1274 0 : match res {
1275 0 : Ok(p) => Some(p),
1276 0 : Err(e) => {
1277 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1278 0 : return Ok(());
1279 : }
1280 : }
1281 : }
1282 :
1283 : };
1284 :
1285 : // Remote preload is complete.
1286 0 : drop(remote_load_completion);
1287 0 :
1288 0 :
1289 0 : // We will time the duration of the attach phase unless this is a creation (attach will do no work)
1290 0 : let attach_start = std::time::Instant::now();
1291 0 : let attached = {
1292 0 : let _attach_timer = Some(TENANT.attach.start_timer());
1293 0 : tenant_clone.attach(preload, &ctx).await
1294 : };
1295 0 : let attach_duration = attach_start.elapsed();
1296 0 : _ = tenant_clone.attach_wal_lag_cooldown.set(WalLagCooldown::new(attach_start, attach_duration));
1297 0 :
1298 0 : match attached {
1299 : Ok(()) => {
1300 0 : info!("attach finished, activating");
1301 0 : tenant_clone.activate(broker_client, None, &ctx);
1302 : }
1303 0 : Err(e) => {
1304 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1305 0 : }
1306 : }
1307 :
1308 : // If we are doing an opportunistic warmup attachment at startup, initialize
1309 : // logical size at the same time. This is better than starting a bunch of idle tenants
1310 : // with cold caches and then coming back later to initialize their logical sizes.
1311 : //
1312 : // It also prevents the warmup proccess competing with the concurrency limit on
1313 : // logical size calculations: if logical size calculation semaphore is saturated,
1314 : // then warmup will wait for that before proceeding to the next tenant.
1315 0 : if matches!(attach_type, AttachType::Warmup { during_startup: true, .. }) {
1316 0 : let mut futs: FuturesUnordered<_> = tenant_clone.timelines.lock().unwrap().values().cloned().map(|t| t.await_initial_logical_size()).collect();
1317 0 : tracing::info!("Waiting for initial logical sizes while warming up...");
1318 0 : while futs.next().await.is_some() {}
1319 0 : tracing::info!("Warm-up complete");
1320 0 : }
1321 :
1322 0 : Ok(())
1323 0 : }
1324 0 : .instrument(tracing::info_span!(parent: None, "attach", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), gen=?generation)),
1325 : );
1326 0 : Ok(tenant)
1327 0 : }
1328 :
1329 190 : #[instrument(skip_all)]
1330 : pub(crate) async fn preload(
1331 : self: &Arc<Self>,
1332 : remote_storage: &GenericRemoteStorage,
1333 : cancel: CancellationToken,
1334 : ) -> anyhow::Result<TenantPreload> {
1335 : span::debug_assert_current_span_has_tenant_id();
1336 : // Get list of remote timelines
1337 : // download index files for every tenant timeline
1338 : info!("listing remote timelines");
1339 : let (remote_timeline_ids, other_keys) = remote_timeline_client::list_remote_timelines(
1340 : remote_storage,
1341 : self.tenant_shard_id,
1342 : cancel.clone(),
1343 : )
1344 : .await?;
1345 : let (offloaded_add, tenant_manifest) =
1346 : match remote_timeline_client::download_tenant_manifest(
1347 : remote_storage,
1348 : &self.tenant_shard_id,
1349 : self.generation,
1350 : &cancel,
1351 : )
1352 : .await
1353 : {
1354 : Ok((tenant_manifest, _generation, _manifest_mtime)) => (
1355 : format!("{} offloaded", tenant_manifest.offloaded_timelines.len()),
1356 : tenant_manifest,
1357 : ),
1358 : Err(DownloadError::NotFound) => {
1359 : ("no manifest".to_string(), TenantManifest::empty())
1360 : }
1361 : Err(e) => Err(e)?,
1362 : };
1363 :
1364 : info!(
1365 : "found {} timelines, and {offloaded_add}",
1366 : remote_timeline_ids.len()
1367 : );
1368 :
1369 : for k in other_keys {
1370 : warn!("Unexpected non timeline key {k}");
1371 : }
1372 :
1373 : Ok(TenantPreload {
1374 : tenant_manifest,
1375 : timelines: self
1376 : .load_timelines_metadata(remote_timeline_ids, remote_storage, cancel)
1377 : .await?,
1378 : })
1379 : }
1380 :
1381 : ///
1382 : /// Background task that downloads all data for a tenant and brings it to Active state.
1383 : ///
1384 : /// No background tasks are started as part of this routine.
1385 : ///
1386 190 : async fn attach(
1387 190 : self: &Arc<Tenant>,
1388 190 : preload: Option<TenantPreload>,
1389 190 : ctx: &RequestContext,
1390 190 : ) -> anyhow::Result<()> {
1391 190 : span::debug_assert_current_span_has_tenant_id();
1392 190 :
1393 190 : failpoint_support::sleep_millis_async!("before-attaching-tenant");
1394 :
1395 190 : let Some(preload) = preload else {
1396 0 : anyhow::bail!("local-only deployment is no longer supported, https://github.com/neondatabase/neon/issues/5624");
1397 : };
1398 :
1399 190 : let mut offloaded_timeline_ids = HashSet::new();
1400 190 : let mut offloaded_timelines_list = Vec::new();
1401 190 : for timeline_manifest in preload.tenant_manifest.offloaded_timelines.iter() {
1402 0 : let timeline_id = timeline_manifest.timeline_id;
1403 0 : let offloaded_timeline =
1404 0 : OffloadedTimeline::from_manifest(self.tenant_shard_id, timeline_manifest);
1405 0 : offloaded_timelines_list.push((timeline_id, Arc::new(offloaded_timeline)));
1406 0 : offloaded_timeline_ids.insert(timeline_id);
1407 0 : }
1408 :
1409 190 : let mut timelines_to_resume_deletions = vec![];
1410 190 :
1411 190 : let mut remote_index_and_client = HashMap::new();
1412 190 : let mut timeline_ancestors = HashMap::new();
1413 190 : let mut existent_timelines = HashSet::new();
1414 196 : for (timeline_id, preload) in preload.timelines {
1415 6 : if offloaded_timeline_ids.remove(&timeline_id) {
1416 : // The timeline is offloaded, skip loading it.
1417 0 : continue;
1418 6 : }
1419 6 : let index_part = match preload.index_part {
1420 6 : Ok(i) => {
1421 6 : debug!("remote index part exists for timeline {timeline_id}");
1422 : // We found index_part on the remote, this is the standard case.
1423 6 : existent_timelines.insert(timeline_id);
1424 6 : i
1425 : }
1426 : Err(DownloadError::NotFound) => {
1427 : // There is no index_part on the remote. We only get here
1428 : // if there is some prefix for the timeline in the remote storage.
1429 : // This can e.g. be the initdb.tar.zst archive, maybe a
1430 : // remnant from a prior incomplete creation or deletion attempt.
1431 : // Delete the local directory as the deciding criterion for a
1432 : // timeline's existence is presence of index_part.
1433 0 : info!(%timeline_id, "index_part not found on remote");
1434 0 : continue;
1435 : }
1436 0 : Err(DownloadError::Fatal(why)) => {
1437 0 : // If, while loading one remote timeline, we saw an indication that our generation
1438 0 : // number is likely invalid, then we should not load the whole tenant.
1439 0 : error!(%timeline_id, "Fatal error loading timeline: {why}");
1440 0 : anyhow::bail!(why.to_string());
1441 : }
1442 0 : Err(e) => {
1443 0 : // Some (possibly ephemeral) error happened during index_part download.
1444 0 : // Pretend the timeline exists to not delete the timeline directory,
1445 0 : // as it might be a temporary issue and we don't want to re-download
1446 0 : // everything after it resolves.
1447 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1448 :
1449 0 : existent_timelines.insert(timeline_id);
1450 0 : continue;
1451 : }
1452 : };
1453 6 : match index_part {
1454 6 : MaybeDeletedIndexPart::IndexPart(index_part) => {
1455 6 : timeline_ancestors.insert(timeline_id, index_part.metadata.clone());
1456 6 : remote_index_and_client.insert(timeline_id, (index_part, preload.client));
1457 6 : }
1458 0 : MaybeDeletedIndexPart::Deleted(index_part) => {
1459 0 : info!(
1460 0 : "timeline {} is deleted, picking to resume deletion",
1461 : timeline_id
1462 : );
1463 0 : timelines_to_resume_deletions.push((timeline_id, index_part, preload.client));
1464 : }
1465 : }
1466 : }
1467 :
1468 190 : let mut gc_blocks = HashMap::new();
1469 :
1470 : // For every timeline, download the metadata file, scan the local directory,
1471 : // and build a layer map that contains an entry for each remote and local
1472 : // layer file.
1473 190 : let sorted_timelines = tree_sort_timelines(timeline_ancestors, |m| m.ancestor_timeline())?;
1474 196 : for (timeline_id, remote_metadata) in sorted_timelines {
1475 6 : let (index_part, remote_client) = remote_index_and_client
1476 6 : .remove(&timeline_id)
1477 6 : .expect("just put it in above");
1478 :
1479 6 : if let Some(blocking) = index_part.gc_blocking.as_ref() {
1480 : // could just filter these away, but it helps while testing
1481 0 : anyhow::ensure!(
1482 0 : !blocking.reasons.is_empty(),
1483 0 : "index_part for {timeline_id} is malformed: it should not have gc blocking with zero reasons"
1484 : );
1485 0 : let prev = gc_blocks.insert(timeline_id, blocking.reasons);
1486 0 : assert!(prev.is_none());
1487 6 : }
1488 :
1489 : // TODO again handle early failure
1490 6 : self.load_remote_timeline(
1491 6 : timeline_id,
1492 6 : index_part,
1493 6 : remote_metadata,
1494 6 : TimelineResources {
1495 6 : remote_client,
1496 6 : timeline_get_throttle: self.timeline_get_throttle.clone(),
1497 6 : l0_flush_global_state: self.l0_flush_global_state.clone(),
1498 6 : },
1499 6 : ctx,
1500 6 : )
1501 3 : .await
1502 6 : .with_context(|| {
1503 0 : format!(
1504 0 : "failed to load remote timeline {} for tenant {}",
1505 0 : timeline_id, self.tenant_shard_id
1506 0 : )
1507 6 : })?;
1508 : }
1509 :
1510 : // Walk through deleted timelines, resume deletion
1511 190 : for (timeline_id, index_part, remote_timeline_client) in timelines_to_resume_deletions {
1512 0 : remote_timeline_client
1513 0 : .init_upload_queue_stopped_to_continue_deletion(&index_part)
1514 0 : .context("init queue stopped")
1515 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1516 :
1517 0 : DeleteTimelineFlow::resume_deletion(
1518 0 : Arc::clone(self),
1519 0 : timeline_id,
1520 0 : &index_part.metadata,
1521 0 : remote_timeline_client,
1522 0 : )
1523 0 : .instrument(tracing::info_span!("timeline_delete", %timeline_id))
1524 0 : .await
1525 0 : .context("resume_deletion")
1526 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1527 : }
1528 : // Complete deletions for offloaded timeline id's.
1529 190 : offloaded_timelines_list
1530 190 : .retain(|(offloaded_id, _offloaded)| {
1531 0 : // At this point, offloaded_timeline_ids has the list of all offloaded timelines
1532 0 : // without a prefix in S3, so they are inexistent.
1533 0 : // In the end, existence of a timeline is finally determined by the existence of an index-part.json in remote storage.
1534 0 : // If there is a dangling reference in another location, they need to be cleaned up.
1535 0 : let delete = offloaded_timeline_ids.contains(offloaded_id);
1536 0 : if delete {
1537 0 : tracing::info!("Removing offloaded timeline {offloaded_id} from manifest as no remote prefix was found");
1538 0 : }
1539 0 : !delete
1540 190 : });
1541 190 : if !offloaded_timelines_list.is_empty() {
1542 0 : tracing::info!(
1543 0 : "Tenant has {} offloaded timelines",
1544 0 : offloaded_timelines_list.len()
1545 : );
1546 190 : }
1547 190 : {
1548 190 : let mut offloaded_timelines_accessor = self.timelines_offloaded.lock().unwrap();
1549 190 : offloaded_timelines_accessor.extend(offloaded_timelines_list.into_iter());
1550 190 : }
1551 190 : if !offloaded_timeline_ids.is_empty() {
1552 0 : self.store_tenant_manifest().await?;
1553 190 : }
1554 :
1555 : // The local filesystem contents are a cache of what's in the remote IndexPart;
1556 : // IndexPart is the source of truth.
1557 190 : self.clean_up_timelines(&existent_timelines)?;
1558 :
1559 190 : self.gc_block.set_scanned(gc_blocks);
1560 190 :
1561 190 : fail::fail_point!("attach-before-activate", |_| {
1562 0 : anyhow::bail!("attach-before-activate");
1563 190 : });
1564 190 : failpoint_support::sleep_millis_async!("attach-before-activate-sleep", &self.cancel);
1565 :
1566 190 : info!("Done");
1567 :
1568 190 : Ok(())
1569 190 : }
1570 :
1571 : /// Check for any local timeline directories that are temporary, or do not correspond to a
1572 : /// timeline that still exists: this can happen if we crashed during a deletion/creation, or
1573 : /// if a timeline was deleted while the tenant was attached to a different pageserver.
1574 190 : fn clean_up_timelines(&self, existent_timelines: &HashSet<TimelineId>) -> anyhow::Result<()> {
1575 190 : let timelines_dir = self.conf.timelines_path(&self.tenant_shard_id);
1576 :
1577 190 : let entries = match timelines_dir.read_dir_utf8() {
1578 190 : Ok(d) => d,
1579 0 : Err(e) => {
1580 0 : if e.kind() == std::io::ErrorKind::NotFound {
1581 0 : return Ok(());
1582 : } else {
1583 0 : return Err(e).context("list timelines directory for tenant");
1584 : }
1585 : }
1586 : };
1587 :
1588 198 : for entry in entries {
1589 8 : let entry = entry.context("read timeline dir entry")?;
1590 8 : let entry_path = entry.path();
1591 :
1592 8 : let purge = if crate::is_temporary(entry_path)
1593 : // TODO: remove uninit mark code (https://github.com/neondatabase/neon/issues/5718)
1594 8 : || is_uninit_mark(entry_path)
1595 8 : || crate::is_delete_mark(entry_path)
1596 : {
1597 0 : true
1598 : } else {
1599 8 : match TimelineId::try_from(entry_path.file_name()) {
1600 8 : Ok(i) => {
1601 8 : // Purge if the timeline ID does not exist in remote storage: remote storage is the authority.
1602 8 : !existent_timelines.contains(&i)
1603 : }
1604 0 : Err(e) => {
1605 0 : tracing::warn!(
1606 0 : "Unparseable directory in timelines directory: {entry_path}, ignoring ({e})"
1607 : );
1608 : // Do not purge junk: if we don't recognize it, be cautious and leave it for a human.
1609 0 : false
1610 : }
1611 : }
1612 : };
1613 :
1614 8 : if purge {
1615 2 : tracing::info!("Purging stale timeline dentry {entry_path}");
1616 2 : if let Err(e) = match entry.file_type() {
1617 2 : Ok(t) => if t.is_dir() {
1618 2 : std::fs::remove_dir_all(entry_path)
1619 : } else {
1620 0 : std::fs::remove_file(entry_path)
1621 : }
1622 2 : .or_else(fs_ext::ignore_not_found),
1623 0 : Err(e) => Err(e),
1624 : } {
1625 0 : tracing::warn!("Failed to purge stale timeline dentry {entry_path}: {e}");
1626 2 : }
1627 6 : }
1628 : }
1629 :
1630 190 : Ok(())
1631 190 : }
1632 :
1633 : /// Get sum of all remote timelines sizes
1634 : ///
1635 : /// This function relies on the index_part instead of listing the remote storage
1636 0 : pub fn remote_size(&self) -> u64 {
1637 0 : let mut size = 0;
1638 :
1639 0 : for timeline in self.list_timelines() {
1640 0 : size += timeline.remote_client.get_remote_physical_size();
1641 0 : }
1642 :
1643 0 : size
1644 0 : }
1645 :
1646 6 : #[instrument(skip_all, fields(timeline_id=%timeline_id))]
1647 : async fn load_remote_timeline(
1648 : &self,
1649 : timeline_id: TimelineId,
1650 : index_part: IndexPart,
1651 : remote_metadata: TimelineMetadata,
1652 : resources: TimelineResources,
1653 : ctx: &RequestContext,
1654 : ) -> anyhow::Result<()> {
1655 : span::debug_assert_current_span_has_tenant_id();
1656 :
1657 : info!("downloading index file for timeline {}", timeline_id);
1658 : tokio::fs::create_dir_all(self.conf.timeline_path(&self.tenant_shard_id, &timeline_id))
1659 : .await
1660 : .context("Failed to create new timeline directory")?;
1661 :
1662 : let ancestor = if let Some(ancestor_id) = remote_metadata.ancestor_timeline() {
1663 : let timelines = self.timelines.lock().unwrap();
1664 : Some(Arc::clone(timelines.get(&ancestor_id).ok_or_else(
1665 0 : || {
1666 0 : anyhow::anyhow!(
1667 0 : "cannot find ancestor timeline {ancestor_id} for timeline {timeline_id}"
1668 0 : )
1669 0 : },
1670 : )?))
1671 : } else {
1672 : None
1673 : };
1674 :
1675 : self.timeline_init_and_sync(
1676 : timeline_id,
1677 : resources,
1678 : index_part,
1679 : remote_metadata,
1680 : ancestor,
1681 : ctx,
1682 : )
1683 : .await
1684 : }
1685 :
1686 190 : async fn load_timelines_metadata(
1687 190 : self: &Arc<Tenant>,
1688 190 : timeline_ids: HashSet<TimelineId>,
1689 190 : remote_storage: &GenericRemoteStorage,
1690 190 : cancel: CancellationToken,
1691 190 : ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
1692 190 : let mut part_downloads = JoinSet::new();
1693 196 : for timeline_id in timeline_ids {
1694 6 : let cancel_clone = cancel.clone();
1695 6 : part_downloads.spawn(
1696 6 : self.load_timeline_metadata(timeline_id, remote_storage.clone(), cancel_clone)
1697 6 : .instrument(info_span!("download_index_part", %timeline_id)),
1698 : );
1699 : }
1700 :
1701 190 : let mut timeline_preloads: HashMap<TimelineId, TimelinePreload> = HashMap::new();
1702 :
1703 : loop {
1704 196 : tokio::select!(
1705 196 : next = part_downloads.join_next() => {
1706 196 : match next {
1707 6 : Some(result) => {
1708 6 : let preload = result.context("join preload task")?;
1709 6 : timeline_preloads.insert(preload.timeline_id, preload);
1710 : },
1711 : None => {
1712 190 : break;
1713 : }
1714 : }
1715 : },
1716 196 : _ = cancel.cancelled() => {
1717 0 : anyhow::bail!("Cancelled while waiting for remote index download")
1718 : }
1719 : )
1720 : }
1721 :
1722 190 : Ok(timeline_preloads)
1723 190 : }
1724 :
1725 6 : fn build_timeline_client(
1726 6 : &self,
1727 6 : timeline_id: TimelineId,
1728 6 : remote_storage: GenericRemoteStorage,
1729 6 : ) -> RemoteTimelineClient {
1730 6 : RemoteTimelineClient::new(
1731 6 : remote_storage.clone(),
1732 6 : self.deletion_queue_client.clone(),
1733 6 : self.conf,
1734 6 : self.tenant_shard_id,
1735 6 : timeline_id,
1736 6 : self.generation,
1737 6 : )
1738 6 : }
1739 :
1740 6 : fn load_timeline_metadata(
1741 6 : self: &Arc<Tenant>,
1742 6 : timeline_id: TimelineId,
1743 6 : remote_storage: GenericRemoteStorage,
1744 6 : cancel: CancellationToken,
1745 6 : ) -> impl Future<Output = TimelinePreload> {
1746 6 : let client = self.build_timeline_client(timeline_id, remote_storage);
1747 6 : async move {
1748 6 : debug_assert_current_span_has_tenant_and_timeline_id();
1749 6 : debug!("starting index part download");
1750 :
1751 15 : let index_part = client.download_index_file(&cancel).await;
1752 :
1753 6 : debug!("finished index part download");
1754 :
1755 6 : TimelinePreload {
1756 6 : client,
1757 6 : timeline_id,
1758 6 : index_part,
1759 6 : }
1760 6 : }
1761 6 : }
1762 :
1763 0 : fn check_to_be_archived_has_no_unarchived_children(
1764 0 : timeline_id: TimelineId,
1765 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
1766 0 : ) -> Result<(), TimelineArchivalError> {
1767 0 : let children: Vec<TimelineId> = timelines
1768 0 : .iter()
1769 0 : .filter_map(|(id, entry)| {
1770 0 : if entry.get_ancestor_timeline_id() != Some(timeline_id) {
1771 0 : return None;
1772 0 : }
1773 0 : if entry.is_archived() == Some(true) {
1774 0 : return None;
1775 0 : }
1776 0 : Some(*id)
1777 0 : })
1778 0 : .collect();
1779 0 :
1780 0 : if !children.is_empty() {
1781 0 : return Err(TimelineArchivalError::HasUnarchivedChildren(children));
1782 0 : }
1783 0 : Ok(())
1784 0 : }
1785 :
1786 0 : fn check_ancestor_of_to_be_unarchived_is_not_archived(
1787 0 : ancestor_timeline_id: TimelineId,
1788 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
1789 0 : offloaded_timelines: &std::sync::MutexGuard<
1790 0 : '_,
1791 0 : HashMap<TimelineId, Arc<OffloadedTimeline>>,
1792 0 : >,
1793 0 : ) -> Result<(), TimelineArchivalError> {
1794 0 : let has_archived_parent =
1795 0 : if let Some(ancestor_timeline) = timelines.get(&ancestor_timeline_id) {
1796 0 : ancestor_timeline.is_archived() == Some(true)
1797 0 : } else if offloaded_timelines.contains_key(&ancestor_timeline_id) {
1798 0 : true
1799 : } else {
1800 0 : error!("ancestor timeline {ancestor_timeline_id} not found");
1801 0 : if cfg!(debug_assertions) {
1802 0 : panic!("ancestor timeline {ancestor_timeline_id} not found");
1803 0 : }
1804 0 : return Err(TimelineArchivalError::NotFound);
1805 : };
1806 0 : if has_archived_parent {
1807 0 : return Err(TimelineArchivalError::HasArchivedParent(
1808 0 : ancestor_timeline_id,
1809 0 : ));
1810 0 : }
1811 0 : Ok(())
1812 0 : }
1813 :
1814 0 : fn check_to_be_unarchived_timeline_has_no_archived_parent(
1815 0 : timeline: &Arc<Timeline>,
1816 0 : ) -> Result<(), TimelineArchivalError> {
1817 0 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
1818 0 : if ancestor_timeline.is_archived() == Some(true) {
1819 0 : return Err(TimelineArchivalError::HasArchivedParent(
1820 0 : ancestor_timeline.timeline_id,
1821 0 : ));
1822 0 : }
1823 0 : }
1824 0 : Ok(())
1825 0 : }
1826 :
1827 : /// Loads the specified (offloaded) timeline from S3 and attaches it as a loaded timeline
1828 : ///
1829 : /// Counterpart to [`offload_timeline`].
1830 0 : async fn unoffload_timeline(
1831 0 : self: &Arc<Self>,
1832 0 : timeline_id: TimelineId,
1833 0 : broker_client: storage_broker::BrokerClientChannel,
1834 0 : ctx: RequestContext,
1835 0 : ) -> Result<Arc<Timeline>, TimelineArchivalError> {
1836 0 : info!("unoffloading timeline");
1837 :
1838 : // We activate the timeline below manually, so this must be called on an active timeline.
1839 : // We expect callers of this function to ensure this.
1840 0 : match self.current_state() {
1841 : TenantState::Activating { .. }
1842 : | TenantState::Attaching
1843 : | TenantState::Broken { .. } => {
1844 0 : panic!("Timeline expected to be active")
1845 : }
1846 0 : TenantState::Stopping { .. } => return Err(TimelineArchivalError::Cancelled),
1847 0 : TenantState::Active => {}
1848 0 : }
1849 0 : let cancel = self.cancel.clone();
1850 0 :
1851 0 : // Protect against concurrent attempts to use this TimelineId
1852 0 : // We don't care much about idempotency, as it's ensured a layer above.
1853 0 : let allow_offloaded = true;
1854 0 : let _create_guard = self
1855 0 : .create_timeline_create_guard(
1856 0 : timeline_id,
1857 0 : CreateTimelineIdempotency::FailWithConflict,
1858 0 : allow_offloaded,
1859 0 : )
1860 0 : .map_err(|err| match err {
1861 0 : TimelineExclusionError::AlreadyCreating => TimelineArchivalError::AlreadyInProgress,
1862 : TimelineExclusionError::AlreadyExists { .. } => {
1863 0 : TimelineArchivalError::Other(anyhow::anyhow!("Timeline already exists"))
1864 : }
1865 0 : TimelineExclusionError::Other(e) => TimelineArchivalError::Other(e),
1866 0 : })?;
1867 :
1868 0 : let timeline_preload = self
1869 0 : .load_timeline_metadata(timeline_id, self.remote_storage.clone(), cancel.clone())
1870 0 : .await;
1871 :
1872 0 : let index_part = match timeline_preload.index_part {
1873 0 : Ok(index_part) => {
1874 0 : debug!("remote index part exists for timeline {timeline_id}");
1875 0 : index_part
1876 : }
1877 : Err(DownloadError::NotFound) => {
1878 0 : error!(%timeline_id, "index_part not found on remote");
1879 0 : return Err(TimelineArchivalError::NotFound);
1880 : }
1881 0 : Err(DownloadError::Cancelled) => return Err(TimelineArchivalError::Cancelled),
1882 0 : Err(e) => {
1883 0 : // Some (possibly ephemeral) error happened during index_part download.
1884 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1885 0 : return Err(TimelineArchivalError::Other(
1886 0 : anyhow::Error::new(e).context("downloading index_part from remote storage"),
1887 0 : ));
1888 : }
1889 : };
1890 0 : let index_part = match index_part {
1891 0 : MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
1892 0 : MaybeDeletedIndexPart::Deleted(_index_part) => {
1893 0 : info!("timeline is deleted according to index_part.json");
1894 0 : return Err(TimelineArchivalError::NotFound);
1895 : }
1896 : };
1897 0 : let remote_metadata = index_part.metadata.clone();
1898 0 : let timeline_resources = self.build_timeline_resources(timeline_id);
1899 0 : self.load_remote_timeline(
1900 0 : timeline_id,
1901 0 : index_part,
1902 0 : remote_metadata,
1903 0 : timeline_resources,
1904 0 : &ctx,
1905 0 : )
1906 0 : .await
1907 0 : .with_context(|| {
1908 0 : format!(
1909 0 : "failed to load remote timeline {} for tenant {}",
1910 0 : timeline_id, self.tenant_shard_id
1911 0 : )
1912 0 : })
1913 0 : .map_err(TimelineArchivalError::Other)?;
1914 :
1915 0 : let timeline = {
1916 0 : let timelines = self.timelines.lock().unwrap();
1917 0 : let Some(timeline) = timelines.get(&timeline_id) else {
1918 0 : warn!("timeline not available directly after attach");
1919 : // This is not a panic because no locks are held between `load_remote_timeline`
1920 : // which puts the timeline into timelines, and our look into the timeline map.
1921 0 : return Err(TimelineArchivalError::Other(anyhow::anyhow!(
1922 0 : "timeline not available directly after attach"
1923 0 : )));
1924 : };
1925 0 : let mut offloaded_timelines = self.timelines_offloaded.lock().unwrap();
1926 0 : if offloaded_timelines.remove(&timeline_id).is_none() {
1927 0 : warn!("timeline already removed from offloaded timelines");
1928 0 : }
1929 0 : Arc::clone(timeline)
1930 0 : };
1931 0 :
1932 0 : // Upload new list of offloaded timelines to S3
1933 0 : self.store_tenant_manifest().await?;
1934 :
1935 : // Activate the timeline (if it makes sense)
1936 0 : if !(timeline.is_broken() || timeline.is_stopping()) {
1937 0 : let background_jobs_can_start = None;
1938 0 : timeline.activate(
1939 0 : self.clone(),
1940 0 : broker_client.clone(),
1941 0 : background_jobs_can_start,
1942 0 : &ctx,
1943 0 : );
1944 0 : }
1945 :
1946 0 : info!("timeline unoffloading complete");
1947 0 : Ok(timeline)
1948 0 : }
1949 :
1950 0 : pub(crate) async fn apply_timeline_archival_config(
1951 0 : self: &Arc<Self>,
1952 0 : timeline_id: TimelineId,
1953 0 : new_state: TimelineArchivalState,
1954 0 : broker_client: storage_broker::BrokerClientChannel,
1955 0 : ctx: RequestContext,
1956 0 : ) -> Result<(), TimelineArchivalError> {
1957 0 : info!("setting timeline archival config");
1958 : // First part: figure out what is needed to do, and do validation
1959 0 : let timeline_or_unarchive_offloaded = 'outer: {
1960 0 : let timelines = self.timelines.lock().unwrap();
1961 :
1962 0 : let Some(timeline) = timelines.get(&timeline_id) else {
1963 0 : let offloaded_timelines = self.timelines_offloaded.lock().unwrap();
1964 0 : let Some(offloaded) = offloaded_timelines.get(&timeline_id) else {
1965 0 : return Err(TimelineArchivalError::NotFound);
1966 : };
1967 0 : if new_state == TimelineArchivalState::Archived {
1968 : // It's offloaded already, so nothing to do
1969 0 : return Ok(());
1970 0 : }
1971 0 : if let Some(ancestor_timeline_id) = offloaded.ancestor_timeline_id {
1972 0 : Self::check_ancestor_of_to_be_unarchived_is_not_archived(
1973 0 : ancestor_timeline_id,
1974 0 : &timelines,
1975 0 : &offloaded_timelines,
1976 0 : )?;
1977 0 : }
1978 0 : break 'outer None;
1979 : };
1980 :
1981 : // Do some validation. We release the timelines lock below, so there is potential
1982 : // for race conditions: these checks are more present to prevent misunderstandings of
1983 : // the API's capabilities, instead of serving as the sole way to defend their invariants.
1984 0 : match new_state {
1985 : TimelineArchivalState::Unarchived => {
1986 0 : Self::check_to_be_unarchived_timeline_has_no_archived_parent(timeline)?
1987 : }
1988 : TimelineArchivalState::Archived => {
1989 0 : Self::check_to_be_archived_has_no_unarchived_children(timeline_id, &timelines)?
1990 : }
1991 : }
1992 0 : Some(Arc::clone(timeline))
1993 : };
1994 :
1995 : // Second part: unoffload timeline (if needed)
1996 0 : let timeline = if let Some(timeline) = timeline_or_unarchive_offloaded {
1997 0 : timeline
1998 : } else {
1999 : // Turn offloaded timeline into a non-offloaded one
2000 0 : self.unoffload_timeline(timeline_id, broker_client, ctx)
2001 0 : .await?
2002 : };
2003 :
2004 : // Third part: upload new timeline archival state and block until it is present in S3
2005 0 : let upload_needed = match timeline
2006 0 : .remote_client
2007 0 : .schedule_index_upload_for_timeline_archival_state(new_state)
2008 : {
2009 0 : Ok(upload_needed) => upload_needed,
2010 0 : Err(e) => {
2011 0 : if timeline.cancel.is_cancelled() {
2012 0 : return Err(TimelineArchivalError::Cancelled);
2013 : } else {
2014 0 : return Err(TimelineArchivalError::Other(e));
2015 : }
2016 : }
2017 : };
2018 :
2019 0 : if upload_needed {
2020 0 : info!("Uploading new state");
2021 : const MAX_WAIT: Duration = Duration::from_secs(10);
2022 0 : let Ok(v) =
2023 0 : tokio::time::timeout(MAX_WAIT, timeline.remote_client.wait_completion()).await
2024 : else {
2025 0 : tracing::warn!("reached timeout for waiting on upload queue");
2026 0 : return Err(TimelineArchivalError::Timeout);
2027 : };
2028 0 : v.map_err(|e| match e {
2029 0 : WaitCompletionError::NotInitialized(e) => {
2030 0 : TimelineArchivalError::Other(anyhow::anyhow!(e))
2031 : }
2032 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2033 0 : TimelineArchivalError::Cancelled
2034 : }
2035 0 : })?;
2036 0 : }
2037 0 : Ok(())
2038 0 : }
2039 :
2040 0 : pub fn get_offloaded_timeline(
2041 0 : &self,
2042 0 : timeline_id: TimelineId,
2043 0 : ) -> Result<Arc<OffloadedTimeline>, GetTimelineError> {
2044 0 : self.timelines_offloaded
2045 0 : .lock()
2046 0 : .unwrap()
2047 0 : .get(&timeline_id)
2048 0 : .map(Arc::clone)
2049 0 : .ok_or(GetTimelineError::NotFound {
2050 0 : tenant_id: self.tenant_shard_id,
2051 0 : timeline_id,
2052 0 : })
2053 0 : }
2054 :
2055 4 : pub(crate) fn tenant_shard_id(&self) -> TenantShardId {
2056 4 : self.tenant_shard_id
2057 4 : }
2058 :
2059 : /// Get Timeline handle for given Neon timeline ID.
2060 : /// This function is idempotent. It doesn't change internal state in any way.
2061 222 : pub fn get_timeline(
2062 222 : &self,
2063 222 : timeline_id: TimelineId,
2064 222 : active_only: bool,
2065 222 : ) -> Result<Arc<Timeline>, GetTimelineError> {
2066 222 : let timelines_accessor = self.timelines.lock().unwrap();
2067 222 : let timeline = timelines_accessor
2068 222 : .get(&timeline_id)
2069 222 : .ok_or(GetTimelineError::NotFound {
2070 222 : tenant_id: self.tenant_shard_id,
2071 222 : timeline_id,
2072 222 : })?;
2073 :
2074 220 : if active_only && !timeline.is_active() {
2075 0 : Err(GetTimelineError::NotActive {
2076 0 : tenant_id: self.tenant_shard_id,
2077 0 : timeline_id,
2078 0 : state: timeline.current_state(),
2079 0 : })
2080 : } else {
2081 220 : Ok(Arc::clone(timeline))
2082 : }
2083 222 : }
2084 :
2085 : /// Lists timelines the tenant contains.
2086 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2087 0 : pub fn list_timelines(&self) -> Vec<Arc<Timeline>> {
2088 0 : self.timelines
2089 0 : .lock()
2090 0 : .unwrap()
2091 0 : .values()
2092 0 : .map(Arc::clone)
2093 0 : .collect()
2094 0 : }
2095 :
2096 : /// Lists timelines the tenant manages, including offloaded ones.
2097 : ///
2098 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2099 0 : pub fn list_timelines_and_offloaded(
2100 0 : &self,
2101 0 : ) -> (Vec<Arc<Timeline>>, Vec<Arc<OffloadedTimeline>>) {
2102 0 : let timelines = self
2103 0 : .timelines
2104 0 : .lock()
2105 0 : .unwrap()
2106 0 : .values()
2107 0 : .map(Arc::clone)
2108 0 : .collect();
2109 0 : let offloaded = self
2110 0 : .timelines_offloaded
2111 0 : .lock()
2112 0 : .unwrap()
2113 0 : .values()
2114 0 : .map(Arc::clone)
2115 0 : .collect();
2116 0 : (timelines, offloaded)
2117 0 : }
2118 :
2119 0 : pub fn list_timeline_ids(&self) -> Vec<TimelineId> {
2120 0 : self.timelines.lock().unwrap().keys().cloned().collect()
2121 0 : }
2122 :
2123 : /// This is used by tests & import-from-basebackup.
2124 : ///
2125 : /// The returned [`UninitializedTimeline`] contains no data nor metadata and it is in
2126 : /// a state that will fail [`Tenant::load_remote_timeline`] because `disk_consistent_lsn=Lsn(0)`.
2127 : ///
2128 : /// The caller is responsible for getting the timeline into a state that will be accepted
2129 : /// by [`Tenant::load_remote_timeline`] / [`Tenant::attach`].
2130 : /// Then they may call [`UninitializedTimeline::finish_creation`] to add the timeline
2131 : /// to the [`Tenant::timelines`].
2132 : ///
2133 : /// Tests should use `Tenant::create_test_timeline` to set up the minimum required metadata keys.
2134 182 : pub(crate) async fn create_empty_timeline(
2135 182 : &self,
2136 182 : new_timeline_id: TimelineId,
2137 182 : initdb_lsn: Lsn,
2138 182 : pg_version: u32,
2139 182 : _ctx: &RequestContext,
2140 182 : ) -> anyhow::Result<UninitializedTimeline> {
2141 182 : anyhow::ensure!(
2142 182 : self.is_active(),
2143 0 : "Cannot create empty timelines on inactive tenant"
2144 : );
2145 :
2146 : // Protect against concurrent attempts to use this TimelineId
2147 182 : let create_guard = match self
2148 182 : .start_creating_timeline(new_timeline_id, CreateTimelineIdempotency::FailWithConflict)
2149 177 : .await?
2150 : {
2151 180 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2152 : StartCreatingTimelineResult::Idempotent(_) => {
2153 0 : unreachable!("FailWithConflict implies we get an error instead")
2154 : }
2155 : };
2156 :
2157 180 : let new_metadata = TimelineMetadata::new(
2158 180 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2159 180 : // make it valid, before calling finish_creation()
2160 180 : Lsn(0),
2161 180 : None,
2162 180 : None,
2163 180 : Lsn(0),
2164 180 : initdb_lsn,
2165 180 : initdb_lsn,
2166 180 : pg_version,
2167 180 : );
2168 180 : self.prepare_new_timeline(
2169 180 : new_timeline_id,
2170 180 : &new_metadata,
2171 180 : create_guard,
2172 180 : initdb_lsn,
2173 180 : None,
2174 180 : )
2175 0 : .await
2176 182 : }
2177 :
2178 : /// Helper for unit tests to create an empty timeline.
2179 : ///
2180 : /// The timeline is has state value `Active` but its background loops are not running.
2181 : // This makes the various functions which anyhow::ensure! for Active state work in tests.
2182 : // Our current tests don't need the background loops.
2183 : #[cfg(test)]
2184 172 : pub async fn create_test_timeline(
2185 172 : &self,
2186 172 : new_timeline_id: TimelineId,
2187 172 : initdb_lsn: Lsn,
2188 172 : pg_version: u32,
2189 172 : ctx: &RequestContext,
2190 172 : ) -> anyhow::Result<Arc<Timeline>> {
2191 172 : let uninit_tl = self
2192 172 : .create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2193 169 : .await?;
2194 172 : let tline = uninit_tl.raw_timeline().expect("we just created it");
2195 172 : assert_eq!(tline.get_last_record_lsn(), Lsn(0));
2196 :
2197 : // Setup minimum keys required for the timeline to be usable.
2198 172 : let mut modification = tline.begin_modification(initdb_lsn);
2199 172 : modification
2200 172 : .init_empty_test_timeline()
2201 172 : .context("init_empty_test_timeline")?;
2202 172 : modification
2203 172 : .commit(ctx)
2204 166 : .await
2205 172 : .context("commit init_empty_test_timeline modification")?;
2206 :
2207 : // Flush to disk so that uninit_tl's check for valid disk_consistent_lsn passes.
2208 172 : tline.maybe_spawn_flush_loop();
2209 172 : tline.freeze_and_flush().await.context("freeze_and_flush")?;
2210 :
2211 : // Make sure the freeze_and_flush reaches remote storage.
2212 172 : tline.remote_client.wait_completion().await.unwrap();
2213 :
2214 172 : let tl = uninit_tl.finish_creation()?;
2215 : // The non-test code would call tl.activate() here.
2216 172 : tl.set_state(TimelineState::Active);
2217 172 : Ok(tl)
2218 172 : }
2219 :
2220 : /// Helper for unit tests to create a timeline with some pre-loaded states.
2221 : #[cfg(test)]
2222 : #[allow(clippy::too_many_arguments)]
2223 32 : pub async fn create_test_timeline_with_layers(
2224 32 : &self,
2225 32 : new_timeline_id: TimelineId,
2226 32 : initdb_lsn: Lsn,
2227 32 : pg_version: u32,
2228 32 : ctx: &RequestContext,
2229 32 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
2230 32 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
2231 32 : end_lsn: Lsn,
2232 32 : ) -> anyhow::Result<Arc<Timeline>> {
2233 : use checks::check_valid_layermap;
2234 : use itertools::Itertools;
2235 :
2236 32 : let tline = self
2237 32 : .create_test_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2238 96 : .await?;
2239 32 : tline.force_advance_lsn(end_lsn);
2240 100 : for deltas in delta_layer_desc {
2241 68 : tline
2242 68 : .force_create_delta_layer(deltas, Some(initdb_lsn), ctx)
2243 204 : .await?;
2244 : }
2245 80 : for (lsn, images) in image_layer_desc {
2246 48 : tline
2247 48 : .force_create_image_layer(lsn, images, Some(initdb_lsn), ctx)
2248 294 : .await?;
2249 : }
2250 32 : let layer_names = tline
2251 32 : .layers
2252 32 : .read()
2253 0 : .await
2254 32 : .layer_map()
2255 32 : .unwrap()
2256 32 : .iter_historic_layers()
2257 148 : .map(|layer| layer.layer_name())
2258 32 : .collect_vec();
2259 32 : if let Some(err) = check_valid_layermap(&layer_names) {
2260 0 : bail!("invalid layermap: {err}");
2261 32 : }
2262 32 : Ok(tline)
2263 32 : }
2264 :
2265 : /// Create a new timeline.
2266 : ///
2267 : /// Returns the new timeline ID and reference to its Timeline object.
2268 : ///
2269 : /// If the caller specified the timeline ID to use (`new_timeline_id`), and timeline with
2270 : /// the same timeline ID already exists, returns CreateTimelineError::AlreadyExists.
2271 : #[allow(clippy::too_many_arguments)]
2272 0 : pub(crate) async fn create_timeline(
2273 0 : self: &Arc<Tenant>,
2274 0 : params: CreateTimelineParams,
2275 0 : broker_client: storage_broker::BrokerClientChannel,
2276 0 : ctx: &RequestContext,
2277 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
2278 0 : if !self.is_active() {
2279 0 : if matches!(self.current_state(), TenantState::Stopping { .. }) {
2280 0 : return Err(CreateTimelineError::ShuttingDown);
2281 : } else {
2282 0 : return Err(CreateTimelineError::Other(anyhow::anyhow!(
2283 0 : "Cannot create timelines on inactive tenant"
2284 0 : )));
2285 : }
2286 0 : }
2287 :
2288 0 : let _gate = self
2289 0 : .gate
2290 0 : .enter()
2291 0 : .map_err(|_| CreateTimelineError::ShuttingDown)?;
2292 :
2293 0 : let result: CreateTimelineResult = match params {
2294 : CreateTimelineParams::Bootstrap(CreateTimelineParamsBootstrap {
2295 0 : new_timeline_id,
2296 0 : existing_initdb_timeline_id,
2297 0 : pg_version,
2298 0 : }) => {
2299 0 : self.bootstrap_timeline(
2300 0 : new_timeline_id,
2301 0 : pg_version,
2302 0 : existing_initdb_timeline_id,
2303 0 : ctx,
2304 0 : )
2305 0 : .await?
2306 : }
2307 : CreateTimelineParams::Branch(CreateTimelineParamsBranch {
2308 0 : new_timeline_id,
2309 0 : ancestor_timeline_id,
2310 0 : mut ancestor_start_lsn,
2311 : }) => {
2312 0 : let ancestor_timeline = self
2313 0 : .get_timeline(ancestor_timeline_id, false)
2314 0 : .context("Cannot branch off the timeline that's not present in pageserver")?;
2315 :
2316 : // instead of waiting around, just deny the request because ancestor is not yet
2317 : // ready for other purposes either.
2318 0 : if !ancestor_timeline.is_active() {
2319 0 : return Err(CreateTimelineError::AncestorNotActive);
2320 0 : }
2321 0 :
2322 0 : if ancestor_timeline.is_archived() == Some(true) {
2323 0 : info!("tried to branch archived timeline");
2324 0 : return Err(CreateTimelineError::AncestorArchived);
2325 0 : }
2326 :
2327 0 : if let Some(lsn) = ancestor_start_lsn.as_mut() {
2328 0 : *lsn = lsn.align();
2329 0 :
2330 0 : let ancestor_ancestor_lsn = ancestor_timeline.get_ancestor_lsn();
2331 0 : if ancestor_ancestor_lsn > *lsn {
2332 : // can we safely just branch from the ancestor instead?
2333 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
2334 0 : "invalid start lsn {} for ancestor timeline {}: less than timeline ancestor lsn {}",
2335 0 : lsn,
2336 0 : ancestor_timeline_id,
2337 0 : ancestor_ancestor_lsn,
2338 0 : )));
2339 0 : }
2340 0 :
2341 0 : // Wait for the WAL to arrive and be processed on the parent branch up
2342 0 : // to the requested branch point. The repository code itself doesn't
2343 0 : // require it, but if we start to receive WAL on the new timeline,
2344 0 : // decoding the new WAL might need to look up previous pages, relation
2345 0 : // sizes etc. and that would get confused if the previous page versions
2346 0 : // are not in the repository yet.
2347 0 : ancestor_timeline
2348 0 : .wait_lsn(*lsn, timeline::WaitLsnWaiter::Tenant, ctx)
2349 0 : .await
2350 0 : .map_err(|e| match e {
2351 0 : e @ (WaitLsnError::Timeout(_) | WaitLsnError::BadState { .. }) => {
2352 0 : CreateTimelineError::AncestorLsn(anyhow::anyhow!(e))
2353 : }
2354 0 : WaitLsnError::Shutdown => CreateTimelineError::ShuttingDown,
2355 0 : })?;
2356 0 : }
2357 :
2358 0 : self.branch_timeline(&ancestor_timeline, new_timeline_id, ancestor_start_lsn, ctx)
2359 0 : .await?
2360 : }
2361 : };
2362 :
2363 : // At this point we have dropped our guard on [`Self::timelines_creating`], and
2364 : // the timeline is visible in [`Self::timelines`], but it is _not_ durable yet. We must
2365 : // not send a success to the caller until it is. The same applies to idempotent retries.
2366 : //
2367 : // TODO: the timeline is already visible in [`Self::timelines`]; a caller could incorrectly
2368 : // assume that, because they can see the timeline via API, that the creation is done and
2369 : // that it is durable. Ideally, we would keep the timeline hidden (in [`Self::timelines_creating`])
2370 : // until it is durable, e.g., by extending the time we hold the creation guard. This also
2371 : // interacts with UninitializedTimeline and is generally a bit tricky.
2372 : //
2373 : // To re-emphasize: the only correct way to create a timeline is to repeat calling the
2374 : // creation API until it returns success. Only then is durability guaranteed.
2375 0 : info!(creation_result=%result.discriminant(), "waiting for timeline to be durable");
2376 0 : result
2377 0 : .timeline()
2378 0 : .remote_client
2379 0 : .wait_completion()
2380 0 : .await
2381 0 : .context("wait for timeline initial uploads to complete")?;
2382 :
2383 : // The creating task is responsible for activating the timeline.
2384 : // We do this after `wait_completion()` so that we don't spin up tasks that start
2385 : // doing stuff before the IndexPart is durable in S3, which is done by the previous section.
2386 0 : let activated_timeline = match result {
2387 0 : CreateTimelineResult::Created(timeline) => {
2388 0 : timeline.activate(self.clone(), broker_client, None, ctx);
2389 0 : timeline
2390 : }
2391 0 : CreateTimelineResult::Idempotent(timeline) => {
2392 0 : info!(
2393 0 : "request was deemed idempotent, activation will be done by the creating task"
2394 : );
2395 0 : timeline
2396 : }
2397 : };
2398 :
2399 0 : Ok(activated_timeline)
2400 0 : }
2401 :
2402 0 : pub(crate) async fn delete_timeline(
2403 0 : self: Arc<Self>,
2404 0 : timeline_id: TimelineId,
2405 0 : ) -> Result<(), DeleteTimelineError> {
2406 0 : DeleteTimelineFlow::run(&self, timeline_id).await?;
2407 :
2408 0 : Ok(())
2409 0 : }
2410 :
2411 : /// perform one garbage collection iteration, removing old data files from disk.
2412 : /// this function is periodically called by gc task.
2413 : /// also it can be explicitly requested through page server api 'do_gc' command.
2414 : ///
2415 : /// `target_timeline_id` specifies the timeline to GC, or None for all.
2416 : ///
2417 : /// The `horizon` an `pitr` parameters determine how much WAL history needs to be retained.
2418 : /// Also known as the retention period, or the GC cutoff point. `horizon` specifies
2419 : /// the amount of history, as LSN difference from current latest LSN on each timeline.
2420 : /// `pitr` specifies the same as a time difference from the current time. The effective
2421 : /// GC cutoff point is determined conservatively by either `horizon` and `pitr`, whichever
2422 : /// requires more history to be retained.
2423 : //
2424 754 : pub(crate) async fn gc_iteration(
2425 754 : &self,
2426 754 : target_timeline_id: Option<TimelineId>,
2427 754 : horizon: u64,
2428 754 : pitr: Duration,
2429 754 : cancel: &CancellationToken,
2430 754 : ctx: &RequestContext,
2431 754 : ) -> Result<GcResult, GcError> {
2432 754 : // Don't start doing work during shutdown
2433 754 : if let TenantState::Stopping { .. } = self.current_state() {
2434 0 : return Ok(GcResult::default());
2435 754 : }
2436 754 :
2437 754 : // there is a global allowed_error for this
2438 754 : if !self.is_active() {
2439 0 : return Err(GcError::NotActive);
2440 754 : }
2441 754 :
2442 754 : {
2443 754 : let conf = self.tenant_conf.load();
2444 754 :
2445 754 : if !conf.location.may_delete_layers_hint() {
2446 0 : info!("Skipping GC in location state {:?}", conf.location);
2447 0 : return Ok(GcResult::default());
2448 754 : }
2449 754 :
2450 754 : if conf.is_gc_blocked_by_lsn_lease_deadline() {
2451 750 : info!("Skipping GC because lsn lease deadline is not reached");
2452 750 : return Ok(GcResult::default());
2453 4 : }
2454 : }
2455 :
2456 4 : let _guard = match self.gc_block.start().await {
2457 4 : Ok(guard) => guard,
2458 0 : Err(reasons) => {
2459 0 : info!("Skipping GC: {reasons}");
2460 0 : return Ok(GcResult::default());
2461 : }
2462 : };
2463 :
2464 4 : self.gc_iteration_internal(target_timeline_id, horizon, pitr, cancel, ctx)
2465 4 : .await
2466 754 : }
2467 :
2468 : /// Perform one compaction iteration.
2469 : /// This function is periodically called by compactor task.
2470 : /// Also it can be explicitly requested per timeline through page server
2471 : /// api's 'compact' command.
2472 : ///
2473 : /// Returns whether we have pending compaction task.
2474 0 : async fn compaction_iteration(
2475 0 : self: &Arc<Self>,
2476 0 : cancel: &CancellationToken,
2477 0 : ctx: &RequestContext,
2478 0 : ) -> Result<bool, timeline::CompactionError> {
2479 0 : // Don't start doing work during shutdown, or when broken, we do not need those in the logs
2480 0 : if !self.is_active() {
2481 0 : return Ok(false);
2482 0 : }
2483 0 :
2484 0 : {
2485 0 : let conf = self.tenant_conf.load();
2486 0 : if !conf.location.may_delete_layers_hint() || !conf.location.may_upload_layers_hint() {
2487 0 : info!("Skipping compaction in location state {:?}", conf.location);
2488 0 : return Ok(false);
2489 0 : }
2490 0 : }
2491 0 :
2492 0 : // Scan through the hashmap and collect a list of all the timelines,
2493 0 : // while holding the lock. Then drop the lock and actually perform the
2494 0 : // compactions. We don't want to block everything else while the
2495 0 : // compaction runs.
2496 0 : let timelines_to_compact_or_offload;
2497 0 : {
2498 0 : let timelines = self.timelines.lock().unwrap();
2499 0 : timelines_to_compact_or_offload = timelines
2500 0 : .iter()
2501 0 : .filter_map(|(timeline_id, timeline)| {
2502 0 : let (is_active, (can_offload, _)) =
2503 0 : (timeline.is_active(), timeline.can_offload());
2504 0 : let has_no_unoffloaded_children = {
2505 0 : !timelines
2506 0 : .iter()
2507 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(*timeline_id))
2508 : };
2509 0 : let config_allows_offload = self.conf.timeline_offloading
2510 0 : || self
2511 0 : .tenant_conf
2512 0 : .load()
2513 0 : .tenant_conf
2514 0 : .timeline_offloading
2515 0 : .unwrap_or_default();
2516 0 : let can_offload =
2517 0 : can_offload && has_no_unoffloaded_children && config_allows_offload;
2518 0 : if (is_active, can_offload) == (false, false) {
2519 0 : None
2520 : } else {
2521 0 : Some((*timeline_id, timeline.clone(), (is_active, can_offload)))
2522 : }
2523 0 : })
2524 0 : .collect::<Vec<_>>();
2525 0 : drop(timelines);
2526 0 : }
2527 0 :
2528 0 : // Before doing any I/O work, check our circuit breaker
2529 0 : if self.compaction_circuit_breaker.lock().unwrap().is_broken() {
2530 0 : info!("Skipping compaction due to previous failures");
2531 0 : return Ok(false);
2532 0 : }
2533 0 :
2534 0 : let mut has_pending_task = false;
2535 :
2536 0 : for (timeline_id, timeline, (can_compact, can_offload)) in &timelines_to_compact_or_offload
2537 : {
2538 0 : let pending_task_left = if *can_compact {
2539 : Some(
2540 0 : timeline
2541 0 : .compact(cancel, EnumSet::empty(), ctx)
2542 0 : .instrument(info_span!("compact_timeline", %timeline_id))
2543 0 : .await
2544 0 : .inspect_err(|e| match e {
2545 0 : timeline::CompactionError::ShuttingDown => (),
2546 0 : timeline::CompactionError::Offload(_) => {
2547 0 : // Failures to offload timelines do not trip the circuit breaker, because
2548 0 : // they do not do lots of writes the way compaction itself does: it is cheap
2549 0 : // to retry, and it would be bad to stop all compaction because of an issue with offloading.
2550 0 : }
2551 0 : timeline::CompactionError::Other(e) => {
2552 0 : self.compaction_circuit_breaker
2553 0 : .lock()
2554 0 : .unwrap()
2555 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, e);
2556 0 : }
2557 0 : })?,
2558 : )
2559 : } else {
2560 0 : None
2561 : };
2562 0 : has_pending_task |= pending_task_left.unwrap_or(false);
2563 0 : if pending_task_left == Some(false) && *can_offload {
2564 0 : offload_timeline(self, timeline)
2565 0 : .instrument(info_span!("offload_timeline", %timeline_id))
2566 0 : .await?;
2567 0 : }
2568 : }
2569 :
2570 0 : self.compaction_circuit_breaker
2571 0 : .lock()
2572 0 : .unwrap()
2573 0 : .success(&CIRCUIT_BREAKERS_UNBROKEN);
2574 0 :
2575 0 : Ok(has_pending_task)
2576 0 : }
2577 :
2578 : // Call through to all timelines to freeze ephemeral layers if needed. Usually
2579 : // this happens during ingest: this background housekeeping is for freezing layers
2580 : // that are open but haven't been written to for some time.
2581 0 : async fn ingest_housekeeping(&self) {
2582 0 : // Scan through the hashmap and collect a list of all the timelines,
2583 0 : // while holding the lock. Then drop the lock and actually perform the
2584 0 : // compactions. We don't want to block everything else while the
2585 0 : // compaction runs.
2586 0 : let timelines = {
2587 0 : self.timelines
2588 0 : .lock()
2589 0 : .unwrap()
2590 0 : .values()
2591 0 : .filter_map(|timeline| {
2592 0 : if timeline.is_active() {
2593 0 : Some(timeline.clone())
2594 : } else {
2595 0 : None
2596 : }
2597 0 : })
2598 0 : .collect::<Vec<_>>()
2599 : };
2600 :
2601 0 : for timeline in &timelines {
2602 0 : timeline.maybe_freeze_ephemeral_layer().await;
2603 : }
2604 0 : }
2605 :
2606 0 : pub fn timeline_has_no_attached_children(&self, timeline_id: TimelineId) -> bool {
2607 0 : let timelines = self.timelines.lock().unwrap();
2608 0 : !timelines
2609 0 : .iter()
2610 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(timeline_id))
2611 0 : }
2612 :
2613 1700 : pub fn current_state(&self) -> TenantState {
2614 1700 : self.state.borrow().clone()
2615 1700 : }
2616 :
2617 940 : pub fn is_active(&self) -> bool {
2618 940 : self.current_state() == TenantState::Active
2619 940 : }
2620 :
2621 0 : pub fn generation(&self) -> Generation {
2622 0 : self.generation
2623 0 : }
2624 :
2625 0 : pub(crate) fn wal_redo_manager_status(&self) -> Option<WalRedoManagerStatus> {
2626 0 : self.walredo_mgr.as_ref().and_then(|mgr| mgr.status())
2627 0 : }
2628 :
2629 : /// Changes tenant status to active, unless shutdown was already requested.
2630 : ///
2631 : /// `background_jobs_can_start` is an optional barrier set to a value during pageserver startup
2632 : /// to delay background jobs. Background jobs can be started right away when None is given.
2633 0 : fn activate(
2634 0 : self: &Arc<Self>,
2635 0 : broker_client: BrokerClientChannel,
2636 0 : background_jobs_can_start: Option<&completion::Barrier>,
2637 0 : ctx: &RequestContext,
2638 0 : ) {
2639 0 : span::debug_assert_current_span_has_tenant_id();
2640 0 :
2641 0 : let mut activating = false;
2642 0 : self.state.send_modify(|current_state| {
2643 : use pageserver_api::models::ActivatingFrom;
2644 0 : match &*current_state {
2645 : TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
2646 0 : panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {state:?}", state = current_state);
2647 : }
2648 0 : TenantState::Attaching => {
2649 0 : *current_state = TenantState::Activating(ActivatingFrom::Attaching);
2650 0 : }
2651 0 : }
2652 0 : debug!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), "Activating tenant");
2653 0 : activating = true;
2654 0 : // Continue outside the closure. We need to grab timelines.lock()
2655 0 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
2656 0 : });
2657 0 :
2658 0 : if activating {
2659 0 : let timelines_accessor = self.timelines.lock().unwrap();
2660 0 : let timelines_offloaded_accessor = self.timelines_offloaded.lock().unwrap();
2661 0 : let timelines_to_activate = timelines_accessor
2662 0 : .values()
2663 0 : .filter(|timeline| !(timeline.is_broken() || timeline.is_stopping()));
2664 0 :
2665 0 : // Before activation, populate each Timeline's GcInfo with information about its children
2666 0 : self.initialize_gc_info(&timelines_accessor, &timelines_offloaded_accessor);
2667 0 :
2668 0 : // Spawn gc and compaction loops. The loops will shut themselves
2669 0 : // down when they notice that the tenant is inactive.
2670 0 : tasks::start_background_loops(self, background_jobs_can_start);
2671 0 :
2672 0 : let mut activated_timelines = 0;
2673 :
2674 0 : for timeline in timelines_to_activate {
2675 0 : timeline.activate(
2676 0 : self.clone(),
2677 0 : broker_client.clone(),
2678 0 : background_jobs_can_start,
2679 0 : ctx,
2680 0 : );
2681 0 : activated_timelines += 1;
2682 0 : }
2683 :
2684 0 : self.state.send_modify(move |current_state| {
2685 0 : assert!(
2686 0 : matches!(current_state, TenantState::Activating(_)),
2687 0 : "set_stopping and set_broken wait for us to leave Activating state",
2688 : );
2689 0 : *current_state = TenantState::Active;
2690 0 :
2691 0 : let elapsed = self.constructed_at.elapsed();
2692 0 : let total_timelines = timelines_accessor.len();
2693 0 :
2694 0 : // log a lot of stuff, because some tenants sometimes suffer from user-visible
2695 0 : // times to activate. see https://github.com/neondatabase/neon/issues/4025
2696 0 : info!(
2697 0 : since_creation_millis = elapsed.as_millis(),
2698 0 : tenant_id = %self.tenant_shard_id.tenant_id,
2699 0 : shard_id = %self.tenant_shard_id.shard_slug(),
2700 0 : activated_timelines,
2701 0 : total_timelines,
2702 0 : post_state = <&'static str>::from(&*current_state),
2703 0 : "activation attempt finished"
2704 : );
2705 :
2706 0 : TENANT.activation.observe(elapsed.as_secs_f64());
2707 0 : });
2708 0 : }
2709 0 : }
2710 :
2711 : /// Shutdown the tenant and join all of the spawned tasks.
2712 : ///
2713 : /// The method caters for all use-cases:
2714 : /// - pageserver shutdown (freeze_and_flush == true)
2715 : /// - detach + ignore (freeze_and_flush == false)
2716 : ///
2717 : /// This will attempt to shutdown even if tenant is broken.
2718 : ///
2719 : /// `shutdown_progress` is a [`completion::Barrier`] for the shutdown initiated by this call.
2720 : /// If the tenant is already shutting down, we return a clone of the first shutdown call's
2721 : /// `Barrier` as an `Err`. This not-first caller can use the returned barrier to join with
2722 : /// the ongoing shutdown.
2723 6 : async fn shutdown(
2724 6 : &self,
2725 6 : shutdown_progress: completion::Barrier,
2726 6 : shutdown_mode: timeline::ShutdownMode,
2727 6 : ) -> Result<(), completion::Barrier> {
2728 6 : span::debug_assert_current_span_has_tenant_id();
2729 :
2730 : // Set tenant (and its timlines) to Stoppping state.
2731 : //
2732 : // Since we can only transition into Stopping state after activation is complete,
2733 : // run it in a JoinSet so all tenants have a chance to stop before we get SIGKILLed.
2734 : //
2735 : // Transitioning tenants to Stopping state has a couple of non-obvious side effects:
2736 : // 1. Lock out any new requests to the tenants.
2737 : // 2. Signal cancellation to WAL receivers (we wait on it below).
2738 : // 3. Signal cancellation for other tenant background loops.
2739 : // 4. ???
2740 : //
2741 : // The waiting for the cancellation is not done uniformly.
2742 : // We certainly wait for WAL receivers to shut down.
2743 : // That is necessary so that no new data comes in before the freeze_and_flush.
2744 : // But the tenant background loops are joined-on in our caller.
2745 : // It's mesed up.
2746 : // we just ignore the failure to stop
2747 :
2748 : // If we're still attaching, fire the cancellation token early to drop out: this
2749 : // will prevent us flushing, but ensures timely shutdown if some I/O during attach
2750 : // is very slow.
2751 6 : let shutdown_mode = if matches!(self.current_state(), TenantState::Attaching) {
2752 0 : self.cancel.cancel();
2753 0 :
2754 0 : // Having fired our cancellation token, do not try and flush timelines: their cancellation tokens
2755 0 : // are children of ours, so their flush loops will have shut down already
2756 0 : timeline::ShutdownMode::Hard
2757 : } else {
2758 6 : shutdown_mode
2759 : };
2760 :
2761 6 : match self.set_stopping(shutdown_progress, false, false).await {
2762 6 : Ok(()) => {}
2763 0 : Err(SetStoppingError::Broken) => {
2764 0 : // assume that this is acceptable
2765 0 : }
2766 0 : Err(SetStoppingError::AlreadyStopping(other)) => {
2767 0 : // give caller the option to wait for this this shutdown
2768 0 : info!("Tenant::shutdown: AlreadyStopping");
2769 0 : return Err(other);
2770 : }
2771 : };
2772 :
2773 6 : let mut js = tokio::task::JoinSet::new();
2774 6 : {
2775 6 : let timelines = self.timelines.lock().unwrap();
2776 6 : timelines.values().for_each(|timeline| {
2777 6 : let timeline = Arc::clone(timeline);
2778 6 : let timeline_id = timeline.timeline_id;
2779 6 : let span = tracing::info_span!("timeline_shutdown", %timeline_id, ?shutdown_mode);
2780 13 : js.spawn(async move { timeline.shutdown(shutdown_mode).instrument(span).await });
2781 6 : })
2782 6 : };
2783 6 : // test_long_timeline_create_then_tenant_delete is leaning on this message
2784 6 : tracing::info!("Waiting for timelines...");
2785 12 : while let Some(res) = js.join_next().await {
2786 0 : match res {
2787 6 : Ok(()) => {}
2788 0 : Err(je) if je.is_cancelled() => unreachable!("no cancelling used"),
2789 0 : Err(je) if je.is_panic() => { /* logged already */ }
2790 0 : Err(je) => warn!("unexpected JoinError: {je:?}"),
2791 : }
2792 : }
2793 :
2794 : // We cancel the Tenant's cancellation token _after_ the timelines have all shut down. This permits
2795 : // them to continue to do work during their shutdown methods, e.g. flushing data.
2796 6 : tracing::debug!("Cancelling CancellationToken");
2797 6 : self.cancel.cancel();
2798 6 :
2799 6 : // shutdown all tenant and timeline tasks: gc, compaction, page service
2800 6 : // No new tasks will be started for this tenant because it's in `Stopping` state.
2801 6 : //
2802 6 : // this will additionally shutdown and await all timeline tasks.
2803 6 : tracing::debug!("Waiting for tasks...");
2804 6 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), None).await;
2805 :
2806 6 : if let Some(walredo_mgr) = self.walredo_mgr.as_ref() {
2807 6 : walredo_mgr.shutdown().await;
2808 0 : }
2809 :
2810 : // Wait for any in-flight operations to complete
2811 6 : self.gate.close().await;
2812 :
2813 6 : remove_tenant_metrics(&self.tenant_shard_id);
2814 6 :
2815 6 : Ok(())
2816 6 : }
2817 :
2818 : /// Change tenant status to Stopping, to mark that it is being shut down.
2819 : ///
2820 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
2821 : ///
2822 : /// This function is not cancel-safe!
2823 : ///
2824 : /// `allow_transition_from_loading` is needed for the special case of loading task deleting the tenant.
2825 : /// `allow_transition_from_attaching` is needed for the special case of attaching deleted tenant.
2826 6 : async fn set_stopping(
2827 6 : &self,
2828 6 : progress: completion::Barrier,
2829 6 : _allow_transition_from_loading: bool,
2830 6 : allow_transition_from_attaching: bool,
2831 6 : ) -> Result<(), SetStoppingError> {
2832 6 : let mut rx = self.state.subscribe();
2833 6 :
2834 6 : // cannot stop before we're done activating, so wait out until we're done activating
2835 6 : rx.wait_for(|state| match state {
2836 0 : TenantState::Attaching if allow_transition_from_attaching => true,
2837 : TenantState::Activating(_) | TenantState::Attaching => {
2838 0 : info!(
2839 0 : "waiting for {} to turn Active|Broken|Stopping",
2840 0 : <&'static str>::from(state)
2841 : );
2842 0 : false
2843 : }
2844 6 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
2845 6 : })
2846 0 : .await
2847 6 : .expect("cannot drop self.state while on a &self method");
2848 6 :
2849 6 : // we now know we're done activating, let's see whether this task is the winner to transition into Stopping
2850 6 : let mut err = None;
2851 6 : let stopping = self.state.send_if_modified(|current_state| match current_state {
2852 : TenantState::Activating(_) => {
2853 0 : unreachable!("1we ensured above that we're done with activation, and, there is no re-activation")
2854 : }
2855 : TenantState::Attaching => {
2856 0 : if !allow_transition_from_attaching {
2857 0 : unreachable!("2we ensured above that we're done with activation, and, there is no re-activation")
2858 0 : };
2859 0 : *current_state = TenantState::Stopping { progress };
2860 0 : true
2861 : }
2862 : TenantState::Active => {
2863 : // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
2864 : // are created after the transition to Stopping. That's harmless, as the Timelines
2865 : // won't be accessible to anyone afterwards, because the Tenant is in Stopping state.
2866 6 : *current_state = TenantState::Stopping { progress };
2867 6 : // Continue stopping outside the closure. We need to grab timelines.lock()
2868 6 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
2869 6 : true
2870 : }
2871 0 : TenantState::Broken { reason, .. } => {
2872 0 : info!(
2873 0 : "Cannot set tenant to Stopping state, it is in Broken state due to: {reason}"
2874 : );
2875 0 : err = Some(SetStoppingError::Broken);
2876 0 : false
2877 : }
2878 0 : TenantState::Stopping { progress } => {
2879 0 : info!("Tenant is already in Stopping state");
2880 0 : err = Some(SetStoppingError::AlreadyStopping(progress.clone()));
2881 0 : false
2882 : }
2883 6 : });
2884 6 : match (stopping, err) {
2885 6 : (true, None) => {} // continue
2886 0 : (false, Some(err)) => return Err(err),
2887 0 : (true, Some(_)) => unreachable!(
2888 0 : "send_if_modified closure must error out if not transitioning to Stopping"
2889 0 : ),
2890 0 : (false, None) => unreachable!(
2891 0 : "send_if_modified closure must return true if transitioning to Stopping"
2892 0 : ),
2893 : }
2894 :
2895 6 : let timelines_accessor = self.timelines.lock().unwrap();
2896 6 : let not_broken_timelines = timelines_accessor
2897 6 : .values()
2898 6 : .filter(|timeline| !timeline.is_broken());
2899 12 : for timeline in not_broken_timelines {
2900 6 : timeline.set_state(TimelineState::Stopping);
2901 6 : }
2902 6 : Ok(())
2903 6 : }
2904 :
2905 : /// Method for tenant::mgr to transition us into Broken state in case of a late failure in
2906 : /// `remove_tenant_from_memory`
2907 : ///
2908 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
2909 : ///
2910 : /// In tests, we also use this to set tenants to Broken state on purpose.
2911 0 : pub(crate) async fn set_broken(&self, reason: String) {
2912 0 : let mut rx = self.state.subscribe();
2913 0 :
2914 0 : // The load & attach routines own the tenant state until it has reached `Active`.
2915 0 : // So, wait until it's done.
2916 0 : rx.wait_for(|state| match state {
2917 : TenantState::Activating(_) | TenantState::Attaching => {
2918 0 : info!(
2919 0 : "waiting for {} to turn Active|Broken|Stopping",
2920 0 : <&'static str>::from(state)
2921 : );
2922 0 : false
2923 : }
2924 0 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
2925 0 : })
2926 0 : .await
2927 0 : .expect("cannot drop self.state while on a &self method");
2928 0 :
2929 0 : // we now know we're done activating, let's see whether this task is the winner to transition into Broken
2930 0 : self.set_broken_no_wait(reason)
2931 0 : }
2932 :
2933 0 : pub(crate) fn set_broken_no_wait(&self, reason: impl Display) {
2934 0 : let reason = reason.to_string();
2935 0 : self.state.send_modify(|current_state| {
2936 0 : match *current_state {
2937 : TenantState::Activating(_) | TenantState::Attaching => {
2938 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
2939 : }
2940 : TenantState::Active => {
2941 0 : if cfg!(feature = "testing") {
2942 0 : warn!("Changing Active tenant to Broken state, reason: {}", reason);
2943 0 : *current_state = TenantState::broken_from_reason(reason);
2944 : } else {
2945 0 : unreachable!("not allowed to call set_broken on Active tenants in non-testing builds")
2946 : }
2947 : }
2948 : TenantState::Broken { .. } => {
2949 0 : warn!("Tenant is already in Broken state");
2950 : }
2951 : // This is the only "expected" path, any other path is a bug.
2952 : TenantState::Stopping { .. } => {
2953 0 : warn!(
2954 0 : "Marking Stopping tenant as Broken state, reason: {}",
2955 : reason
2956 : );
2957 0 : *current_state = TenantState::broken_from_reason(reason);
2958 : }
2959 : }
2960 0 : });
2961 0 : }
2962 :
2963 0 : pub fn subscribe_for_state_updates(&self) -> watch::Receiver<TenantState> {
2964 0 : self.state.subscribe()
2965 0 : }
2966 :
2967 : /// The activate_now semaphore is initialized with zero units. As soon as
2968 : /// we add a unit, waiters will be able to acquire a unit and proceed.
2969 0 : pub(crate) fn activate_now(&self) {
2970 0 : self.activate_now_sem.add_permits(1);
2971 0 : }
2972 :
2973 0 : pub(crate) async fn wait_to_become_active(
2974 0 : &self,
2975 0 : timeout: Duration,
2976 0 : ) -> Result<(), GetActiveTenantError> {
2977 0 : let mut receiver = self.state.subscribe();
2978 : loop {
2979 0 : let current_state = receiver.borrow_and_update().clone();
2980 0 : match current_state {
2981 : TenantState::Attaching | TenantState::Activating(_) => {
2982 : // in these states, there's a chance that we can reach ::Active
2983 0 : self.activate_now();
2984 0 : match timeout_cancellable(timeout, &self.cancel, receiver.changed()).await {
2985 0 : Ok(r) => {
2986 0 : r.map_err(
2987 0 : |_e: tokio::sync::watch::error::RecvError|
2988 : // Tenant existed but was dropped: report it as non-existent
2989 0 : GetActiveTenantError::NotFound(GetTenantError::NotFound(self.tenant_shard_id.tenant_id))
2990 0 : )?
2991 : }
2992 : Err(TimeoutCancellableError::Cancelled) => {
2993 0 : return Err(GetActiveTenantError::Cancelled);
2994 : }
2995 : Err(TimeoutCancellableError::Timeout) => {
2996 0 : return Err(GetActiveTenantError::WaitForActiveTimeout {
2997 0 : latest_state: Some(self.current_state()),
2998 0 : wait_time: timeout,
2999 0 : });
3000 : }
3001 : }
3002 : }
3003 : TenantState::Active { .. } => {
3004 0 : return Ok(());
3005 : }
3006 0 : TenantState::Broken { reason, .. } => {
3007 0 : // This is fatal, and reported distinctly from the general case of "will never be active" because
3008 0 : // it's logically a 500 to external API users (broken is always a bug).
3009 0 : return Err(GetActiveTenantError::Broken(reason));
3010 : }
3011 : TenantState::Stopping { .. } => {
3012 : // There's no chance the tenant can transition back into ::Active
3013 0 : return Err(GetActiveTenantError::WillNotBecomeActive(current_state));
3014 : }
3015 : }
3016 : }
3017 0 : }
3018 :
3019 0 : pub(crate) fn get_attach_mode(&self) -> AttachmentMode {
3020 0 : self.tenant_conf.load().location.attach_mode
3021 0 : }
3022 :
3023 : /// For API access: generate a LocationConfig equivalent to the one that would be used to
3024 : /// create a Tenant in the same state. Do not use this in hot paths: it's for relatively
3025 : /// rare external API calls, like a reconciliation at startup.
3026 0 : pub(crate) fn get_location_conf(&self) -> models::LocationConfig {
3027 0 : let conf = self.tenant_conf.load();
3028 :
3029 0 : let location_config_mode = match conf.location.attach_mode {
3030 0 : AttachmentMode::Single => models::LocationConfigMode::AttachedSingle,
3031 0 : AttachmentMode::Multi => models::LocationConfigMode::AttachedMulti,
3032 0 : AttachmentMode::Stale => models::LocationConfigMode::AttachedStale,
3033 : };
3034 :
3035 : // We have a pageserver TenantConf, we need the API-facing TenantConfig.
3036 0 : let tenant_config: models::TenantConfig = conf.tenant_conf.clone().into();
3037 0 :
3038 0 : models::LocationConfig {
3039 0 : mode: location_config_mode,
3040 0 : generation: self.generation.into(),
3041 0 : secondary_conf: None,
3042 0 : shard_number: self.shard_identity.number.0,
3043 0 : shard_count: self.shard_identity.count.literal(),
3044 0 : shard_stripe_size: self.shard_identity.stripe_size.0,
3045 0 : tenant_conf: tenant_config,
3046 0 : }
3047 0 : }
3048 :
3049 0 : pub(crate) fn get_tenant_shard_id(&self) -> &TenantShardId {
3050 0 : &self.tenant_shard_id
3051 0 : }
3052 :
3053 0 : pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
3054 0 : self.shard_identity.stripe_size
3055 0 : }
3056 :
3057 0 : pub(crate) fn get_generation(&self) -> Generation {
3058 0 : self.generation
3059 0 : }
3060 :
3061 : /// This function partially shuts down the tenant (it shuts down the Timelines) and is fallible,
3062 : /// and can leave the tenant in a bad state if it fails. The caller is responsible for
3063 : /// resetting this tenant to a valid state if we fail.
3064 0 : pub(crate) async fn split_prepare(
3065 0 : &self,
3066 0 : child_shards: &Vec<TenantShardId>,
3067 0 : ) -> anyhow::Result<()> {
3068 0 : let (timelines, offloaded) = {
3069 0 : let timelines = self.timelines.lock().unwrap();
3070 0 : let offloaded = self.timelines_offloaded.lock().unwrap();
3071 0 : (timelines.clone(), offloaded.clone())
3072 0 : };
3073 0 : let timelines_iter = timelines
3074 0 : .values()
3075 0 : .map(TimelineOrOffloadedArcRef::<'_>::from)
3076 0 : .chain(
3077 0 : offloaded
3078 0 : .values()
3079 0 : .map(TimelineOrOffloadedArcRef::<'_>::from),
3080 0 : );
3081 0 : for timeline in timelines_iter {
3082 : // We do not block timeline creation/deletion during splits inside the pageserver: it is up to higher levels
3083 : // to ensure that they do not start a split if currently in the process of doing these.
3084 :
3085 0 : let timeline_id = timeline.timeline_id();
3086 :
3087 0 : if let TimelineOrOffloadedArcRef::Timeline(timeline) = timeline {
3088 : // Upload an index from the parent: this is partly to provide freshness for the
3089 : // child tenants that will copy it, and partly for general ease-of-debugging: there will
3090 : // always be a parent shard index in the same generation as we wrote the child shard index.
3091 0 : tracing::info!(%timeline_id, "Uploading index");
3092 0 : timeline
3093 0 : .remote_client
3094 0 : .schedule_index_upload_for_file_changes()?;
3095 0 : timeline.remote_client.wait_completion().await?;
3096 0 : }
3097 :
3098 0 : let remote_client = match timeline {
3099 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.remote_client.clone(),
3100 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => {
3101 0 : let remote_client = self
3102 0 : .build_timeline_client(offloaded.timeline_id, self.remote_storage.clone());
3103 0 : Arc::new(remote_client)
3104 : }
3105 : };
3106 :
3107 : // Shut down the timeline's remote client: this means that the indices we write
3108 : // for child shards will not be invalidated by the parent shard deleting layers.
3109 0 : tracing::info!(%timeline_id, "Shutting down remote storage client");
3110 0 : remote_client.shutdown().await;
3111 :
3112 : // Download methods can still be used after shutdown, as they don't flow through the remote client's
3113 : // queue. In principal the RemoteTimelineClient could provide this without downloading it, but this
3114 : // operation is rare, so it's simpler to just download it (and robustly guarantees that the index
3115 : // we use here really is the remotely persistent one).
3116 0 : tracing::info!(%timeline_id, "Downloading index_part from parent");
3117 0 : let result = remote_client
3118 0 : .download_index_file(&self.cancel)
3119 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))
3120 0 : .await?;
3121 0 : let index_part = match result {
3122 : MaybeDeletedIndexPart::Deleted(_) => {
3123 0 : anyhow::bail!("Timeline deletion happened concurrently with split")
3124 : }
3125 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
3126 : };
3127 :
3128 0 : for child_shard in child_shards {
3129 0 : tracing::info!(%timeline_id, "Uploading index_part for child {}", child_shard.to_index());
3130 0 : upload_index_part(
3131 0 : &self.remote_storage,
3132 0 : child_shard,
3133 0 : &timeline_id,
3134 0 : self.generation,
3135 0 : &index_part,
3136 0 : &self.cancel,
3137 0 : )
3138 0 : .await?;
3139 : }
3140 : }
3141 :
3142 0 : let tenant_manifest = self.build_tenant_manifest();
3143 0 : for child_shard in child_shards {
3144 0 : tracing::info!(
3145 0 : "Uploading tenant manifest for child {}",
3146 0 : child_shard.to_index()
3147 : );
3148 0 : upload_tenant_manifest(
3149 0 : &self.remote_storage,
3150 0 : child_shard,
3151 0 : self.generation,
3152 0 : &tenant_manifest,
3153 0 : &self.cancel,
3154 0 : )
3155 0 : .await?;
3156 : }
3157 :
3158 0 : Ok(())
3159 0 : }
3160 :
3161 0 : pub(crate) fn get_sizes(&self) -> TopTenantShardItem {
3162 0 : let mut result = TopTenantShardItem {
3163 0 : id: self.tenant_shard_id,
3164 0 : resident_size: 0,
3165 0 : physical_size: 0,
3166 0 : max_logical_size: 0,
3167 0 : };
3168 :
3169 0 : for timeline in self.timelines.lock().unwrap().values() {
3170 0 : result.resident_size += timeline.metrics.resident_physical_size_gauge.get();
3171 0 :
3172 0 : result.physical_size += timeline
3173 0 : .remote_client
3174 0 : .metrics
3175 0 : .remote_physical_size_gauge
3176 0 : .get();
3177 0 : result.max_logical_size = std::cmp::max(
3178 0 : result.max_logical_size,
3179 0 : timeline.metrics.current_logical_size_gauge.get(),
3180 0 : );
3181 0 : }
3182 :
3183 0 : result
3184 0 : }
3185 : }
3186 :
3187 : /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
3188 : /// perform a topological sort, so that the parent of each timeline comes
3189 : /// before the children.
3190 : /// E extracts the ancestor from T
3191 : /// This allows for T to be different. It can be TimelineMetadata, can be Timeline itself, etc.
3192 190 : fn tree_sort_timelines<T, E>(
3193 190 : timelines: HashMap<TimelineId, T>,
3194 190 : extractor: E,
3195 190 : ) -> anyhow::Result<Vec<(TimelineId, T)>>
3196 190 : where
3197 190 : E: Fn(&T) -> Option<TimelineId>,
3198 190 : {
3199 190 : let mut result = Vec::with_capacity(timelines.len());
3200 190 :
3201 190 : let mut now = Vec::with_capacity(timelines.len());
3202 190 : // (ancestor, children)
3203 190 : let mut later: HashMap<TimelineId, Vec<(TimelineId, T)>> =
3204 190 : HashMap::with_capacity(timelines.len());
3205 :
3206 196 : for (timeline_id, value) in timelines {
3207 6 : if let Some(ancestor_id) = extractor(&value) {
3208 2 : let children = later.entry(ancestor_id).or_default();
3209 2 : children.push((timeline_id, value));
3210 4 : } else {
3211 4 : now.push((timeline_id, value));
3212 4 : }
3213 : }
3214 :
3215 196 : while let Some((timeline_id, metadata)) = now.pop() {
3216 6 : result.push((timeline_id, metadata));
3217 : // All children of this can be loaded now
3218 6 : if let Some(mut children) = later.remove(&timeline_id) {
3219 2 : now.append(&mut children);
3220 4 : }
3221 : }
3222 :
3223 : // All timelines should be visited now. Unless there were timelines with missing ancestors.
3224 190 : if !later.is_empty() {
3225 0 : for (missing_id, orphan_ids) in later {
3226 0 : for (orphan_id, _) in orphan_ids {
3227 0 : error!("could not load timeline {orphan_id} because its ancestor timeline {missing_id} could not be loaded");
3228 : }
3229 : }
3230 0 : bail!("could not load tenant because some timelines are missing ancestors");
3231 190 : }
3232 190 :
3233 190 : Ok(result)
3234 190 : }
3235 :
3236 : impl Tenant {
3237 0 : pub fn tenant_specific_overrides(&self) -> TenantConfOpt {
3238 0 : self.tenant_conf.load().tenant_conf.clone()
3239 0 : }
3240 :
3241 0 : pub fn effective_config(&self) -> TenantConf {
3242 0 : self.tenant_specific_overrides()
3243 0 : .merge(self.conf.default_tenant_conf.clone())
3244 0 : }
3245 :
3246 0 : pub fn get_checkpoint_distance(&self) -> u64 {
3247 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3248 0 : tenant_conf
3249 0 : .checkpoint_distance
3250 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
3251 0 : }
3252 :
3253 0 : pub fn get_checkpoint_timeout(&self) -> Duration {
3254 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3255 0 : tenant_conf
3256 0 : .checkpoint_timeout
3257 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
3258 0 : }
3259 :
3260 0 : pub fn get_compaction_target_size(&self) -> u64 {
3261 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3262 0 : tenant_conf
3263 0 : .compaction_target_size
3264 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
3265 0 : }
3266 :
3267 0 : pub fn get_compaction_period(&self) -> Duration {
3268 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3269 0 : tenant_conf
3270 0 : .compaction_period
3271 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
3272 0 : }
3273 :
3274 0 : pub fn get_compaction_threshold(&self) -> usize {
3275 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3276 0 : tenant_conf
3277 0 : .compaction_threshold
3278 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
3279 0 : }
3280 :
3281 0 : pub fn get_gc_horizon(&self) -> u64 {
3282 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3283 0 : tenant_conf
3284 0 : .gc_horizon
3285 0 : .unwrap_or(self.conf.default_tenant_conf.gc_horizon)
3286 0 : }
3287 :
3288 0 : pub fn get_gc_period(&self) -> Duration {
3289 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3290 0 : tenant_conf
3291 0 : .gc_period
3292 0 : .unwrap_or(self.conf.default_tenant_conf.gc_period)
3293 0 : }
3294 :
3295 0 : pub fn get_image_creation_threshold(&self) -> usize {
3296 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3297 0 : tenant_conf
3298 0 : .image_creation_threshold
3299 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
3300 0 : }
3301 :
3302 0 : pub fn get_pitr_interval(&self) -> Duration {
3303 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3304 0 : tenant_conf
3305 0 : .pitr_interval
3306 0 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
3307 0 : }
3308 :
3309 0 : pub fn get_min_resident_size_override(&self) -> Option<u64> {
3310 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3311 0 : tenant_conf
3312 0 : .min_resident_size_override
3313 0 : .or(self.conf.default_tenant_conf.min_resident_size_override)
3314 0 : }
3315 :
3316 0 : pub fn get_heatmap_period(&self) -> Option<Duration> {
3317 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3318 0 : let heatmap_period = tenant_conf
3319 0 : .heatmap_period
3320 0 : .unwrap_or(self.conf.default_tenant_conf.heatmap_period);
3321 0 : if heatmap_period.is_zero() {
3322 0 : None
3323 : } else {
3324 0 : Some(heatmap_period)
3325 : }
3326 0 : }
3327 :
3328 4 : pub fn get_lsn_lease_length(&self) -> Duration {
3329 4 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3330 4 : tenant_conf
3331 4 : .lsn_lease_length
3332 4 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
3333 4 : }
3334 :
3335 : /// Generate an up-to-date TenantManifest based on the state of this Tenant.
3336 0 : fn build_tenant_manifest(&self) -> TenantManifest {
3337 0 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
3338 0 :
3339 0 : let mut timeline_manifests = timelines_offloaded
3340 0 : .iter()
3341 0 : .map(|(_timeline_id, offloaded)| offloaded.manifest())
3342 0 : .collect::<Vec<_>>();
3343 0 : // Sort the manifests so that our output is deterministic
3344 0 : timeline_manifests.sort_by_key(|timeline_manifest| timeline_manifest.timeline_id);
3345 0 :
3346 0 : TenantManifest {
3347 0 : version: LATEST_TENANT_MANIFEST_VERSION,
3348 0 : offloaded_timelines: timeline_manifests,
3349 0 : }
3350 0 : }
3351 :
3352 0 : pub fn set_new_tenant_config(&self, new_tenant_conf: TenantConfOpt) {
3353 0 : // Use read-copy-update in order to avoid overwriting the location config
3354 0 : // state if this races with [`Tenant::set_new_location_config`]. Note that
3355 0 : // this race is not possible if both request types come from the storage
3356 0 : // controller (as they should!) because an exclusive op lock is required
3357 0 : // on the storage controller side.
3358 0 : self.tenant_conf.rcu(|inner| {
3359 0 : Arc::new(AttachedTenantConf {
3360 0 : tenant_conf: new_tenant_conf.clone(),
3361 0 : location: inner.location,
3362 0 : // Attached location is not changed, no need to update lsn lease deadline.
3363 0 : lsn_lease_deadline: inner.lsn_lease_deadline,
3364 0 : })
3365 0 : });
3366 0 :
3367 0 : self.tenant_conf_updated(&new_tenant_conf);
3368 0 : // Don't hold self.timelines.lock() during the notifies.
3369 0 : // There's no risk of deadlock right now, but there could be if we consolidate
3370 0 : // mutexes in struct Timeline in the future.
3371 0 : let timelines = self.list_timelines();
3372 0 : for timeline in timelines {
3373 0 : timeline.tenant_conf_updated(&new_tenant_conf);
3374 0 : }
3375 0 : }
3376 :
3377 0 : pub(crate) fn set_new_location_config(&self, new_conf: AttachedTenantConf) {
3378 0 : let new_tenant_conf = new_conf.tenant_conf.clone();
3379 0 :
3380 0 : self.tenant_conf.store(Arc::new(new_conf));
3381 0 :
3382 0 : self.tenant_conf_updated(&new_tenant_conf);
3383 0 : // Don't hold self.timelines.lock() during the notifies.
3384 0 : // There's no risk of deadlock right now, but there could be if we consolidate
3385 0 : // mutexes in struct Timeline in the future.
3386 0 : let timelines = self.list_timelines();
3387 0 : for timeline in timelines {
3388 0 : timeline.tenant_conf_updated(&new_tenant_conf);
3389 0 : }
3390 0 : }
3391 :
3392 190 : fn get_timeline_get_throttle_config(
3393 190 : psconf: &'static PageServerConf,
3394 190 : overrides: &TenantConfOpt,
3395 190 : ) -> throttle::Config {
3396 190 : overrides
3397 190 : .timeline_get_throttle
3398 190 : .clone()
3399 190 : .unwrap_or(psconf.default_tenant_conf.timeline_get_throttle.clone())
3400 190 : }
3401 :
3402 0 : pub(crate) fn tenant_conf_updated(&self, new_conf: &TenantConfOpt) {
3403 0 : let conf = Self::get_timeline_get_throttle_config(self.conf, new_conf);
3404 0 : self.timeline_get_throttle.reconfigure(conf)
3405 0 : }
3406 :
3407 : /// Helper function to create a new Timeline struct.
3408 : ///
3409 : /// The returned Timeline is in Loading state. The caller is responsible for
3410 : /// initializing any on-disk state, and for inserting the Timeline to the 'timelines'
3411 : /// map.
3412 : ///
3413 : /// `validate_ancestor == false` is used when a timeline is created for deletion
3414 : /// and we might not have the ancestor present anymore which is fine for to be
3415 : /// deleted timelines.
3416 414 : fn create_timeline_struct(
3417 414 : &self,
3418 414 : new_timeline_id: TimelineId,
3419 414 : new_metadata: &TimelineMetadata,
3420 414 : ancestor: Option<Arc<Timeline>>,
3421 414 : resources: TimelineResources,
3422 414 : cause: CreateTimelineCause,
3423 414 : create_idempotency: CreateTimelineIdempotency,
3424 414 : ) -> anyhow::Result<Arc<Timeline>> {
3425 414 : let state = match cause {
3426 : CreateTimelineCause::Load => {
3427 414 : let ancestor_id = new_metadata.ancestor_timeline();
3428 414 : anyhow::ensure!(
3429 414 : ancestor_id == ancestor.as_ref().map(|t| t.timeline_id),
3430 0 : "Timeline's {new_timeline_id} ancestor {ancestor_id:?} was not found"
3431 : );
3432 414 : TimelineState::Loading
3433 : }
3434 0 : CreateTimelineCause::Delete => TimelineState::Stopping,
3435 : };
3436 :
3437 414 : let pg_version = new_metadata.pg_version();
3438 414 :
3439 414 : let timeline = Timeline::new(
3440 414 : self.conf,
3441 414 : Arc::clone(&self.tenant_conf),
3442 414 : new_metadata,
3443 414 : ancestor,
3444 414 : new_timeline_id,
3445 414 : self.tenant_shard_id,
3446 414 : self.generation,
3447 414 : self.shard_identity,
3448 414 : self.walredo_mgr.clone(),
3449 414 : resources,
3450 414 : pg_version,
3451 414 : state,
3452 414 : self.attach_wal_lag_cooldown.clone(),
3453 414 : create_idempotency,
3454 414 : self.cancel.child_token(),
3455 414 : );
3456 414 :
3457 414 : Ok(timeline)
3458 414 : }
3459 :
3460 : // Allow too_many_arguments because a constructor's argument list naturally grows with the
3461 : // number of attributes in the struct: breaking these out into a builder wouldn't be helpful.
3462 : #[allow(clippy::too_many_arguments)]
3463 190 : fn new(
3464 190 : state: TenantState,
3465 190 : conf: &'static PageServerConf,
3466 190 : attached_conf: AttachedTenantConf,
3467 190 : shard_identity: ShardIdentity,
3468 190 : walredo_mgr: Option<Arc<WalRedoManager>>,
3469 190 : tenant_shard_id: TenantShardId,
3470 190 : remote_storage: GenericRemoteStorage,
3471 190 : deletion_queue_client: DeletionQueueClient,
3472 190 : l0_flush_global_state: L0FlushGlobalState,
3473 190 : ) -> Tenant {
3474 190 : debug_assert!(
3475 190 : !attached_conf.location.generation.is_none() || conf.control_plane_api.is_none()
3476 : );
3477 :
3478 190 : let (state, mut rx) = watch::channel(state);
3479 190 :
3480 190 : tokio::spawn(async move {
3481 190 : // reflect tenant state in metrics:
3482 190 : // - global per tenant state: TENANT_STATE_METRIC
3483 190 : // - "set" of broken tenants: BROKEN_TENANTS_SET
3484 190 : //
3485 190 : // set of broken tenants should not have zero counts so that it remains accessible for
3486 190 : // alerting.
3487 190 :
3488 190 : let tid = tenant_shard_id.to_string();
3489 190 : let shard_id = tenant_shard_id.shard_slug().to_string();
3490 190 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
3491 :
3492 380 : fn inspect_state(state: &TenantState) -> ([&'static str; 1], bool) {
3493 380 : ([state.into()], matches!(state, TenantState::Broken { .. }))
3494 380 : }
3495 :
3496 190 : let mut tuple = inspect_state(&rx.borrow_and_update());
3497 190 :
3498 190 : let is_broken = tuple.1;
3499 190 : let mut counted_broken = if is_broken {
3500 : // add the id to the set right away, there should not be any updates on the channel
3501 : // after before tenant is removed, if ever
3502 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
3503 0 : true
3504 : } else {
3505 190 : false
3506 : };
3507 :
3508 : loop {
3509 380 : let labels = &tuple.0;
3510 380 : let current = TENANT_STATE_METRIC.with_label_values(labels);
3511 380 : current.inc();
3512 380 :
3513 380 : if rx.changed().await.is_err() {
3514 : // tenant has been dropped
3515 16 : current.dec();
3516 16 : drop(BROKEN_TENANTS_SET.remove_label_values(set_key));
3517 16 : break;
3518 190 : }
3519 190 :
3520 190 : current.dec();
3521 190 : tuple = inspect_state(&rx.borrow_and_update());
3522 190 :
3523 190 : let is_broken = tuple.1;
3524 190 : if is_broken && !counted_broken {
3525 0 : counted_broken = true;
3526 0 : // insert the tenant_id (back) into the set while avoiding needless counter
3527 0 : // access
3528 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
3529 190 : }
3530 : }
3531 190 : });
3532 190 :
3533 190 : Tenant {
3534 190 : tenant_shard_id,
3535 190 : shard_identity,
3536 190 : generation: attached_conf.location.generation,
3537 190 : conf,
3538 190 : // using now here is good enough approximation to catch tenants with really long
3539 190 : // activation times.
3540 190 : constructed_at: Instant::now(),
3541 190 : timelines: Mutex::new(HashMap::new()),
3542 190 : timelines_creating: Mutex::new(HashSet::new()),
3543 190 : timelines_offloaded: Mutex::new(HashMap::new()),
3544 190 : tenant_manifest_upload: Default::default(),
3545 190 : gc_cs: tokio::sync::Mutex::new(()),
3546 190 : walredo_mgr,
3547 190 : remote_storage,
3548 190 : deletion_queue_client,
3549 190 : state,
3550 190 : cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
3551 190 : cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
3552 190 : eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
3553 190 : compaction_circuit_breaker: std::sync::Mutex::new(CircuitBreaker::new(
3554 190 : format!("compaction-{tenant_shard_id}"),
3555 190 : 5,
3556 190 : // Compaction can be a very expensive operation, and might leak disk space. It also ought
3557 190 : // to be infallible, as long as remote storage is available. So if it repeatedly fails,
3558 190 : // use an extremely long backoff.
3559 190 : Some(Duration::from_secs(3600 * 24)),
3560 190 : )),
3561 190 : activate_now_sem: tokio::sync::Semaphore::new(0),
3562 190 : attach_wal_lag_cooldown: Arc::new(std::sync::OnceLock::new()),
3563 190 : cancel: CancellationToken::default(),
3564 190 : gate: Gate::default(),
3565 190 : timeline_get_throttle: Arc::new(throttle::Throttle::new(
3566 190 : Tenant::get_timeline_get_throttle_config(conf, &attached_conf.tenant_conf),
3567 190 : crate::metrics::tenant_throttling::TimelineGet::new(&tenant_shard_id),
3568 190 : )),
3569 190 : tenant_conf: Arc::new(ArcSwap::from_pointee(attached_conf)),
3570 190 : ongoing_timeline_detach: std::sync::Mutex::default(),
3571 190 : gc_block: Default::default(),
3572 190 : l0_flush_global_state,
3573 190 : }
3574 190 : }
3575 :
3576 : /// Locate and load config
3577 0 : pub(super) fn load_tenant_config(
3578 0 : conf: &'static PageServerConf,
3579 0 : tenant_shard_id: &TenantShardId,
3580 0 : ) -> Result<LocationConf, LoadConfigError> {
3581 0 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
3582 0 :
3583 0 : info!("loading tenant configuration from {config_path}");
3584 :
3585 : // load and parse file
3586 0 : let config = fs::read_to_string(&config_path).map_err(|e| {
3587 0 : match e.kind() {
3588 : std::io::ErrorKind::NotFound => {
3589 : // The config should almost always exist for a tenant directory:
3590 : // - When attaching a tenant, the config is the first thing we write
3591 : // - When detaching a tenant, we atomically move the directory to a tmp location
3592 : // before deleting contents.
3593 : //
3594 : // The very rare edge case that can result in a missing config is if we crash during attach
3595 : // between creating directory and writing config. Callers should handle that as if the
3596 : // directory didn't exist.
3597 :
3598 0 : LoadConfigError::NotFound(config_path)
3599 : }
3600 : _ => {
3601 : // No IO errors except NotFound are acceptable here: other kinds of error indicate local storage or permissions issues
3602 : // that we cannot cleanly recover
3603 0 : crate::virtual_file::on_fatal_io_error(&e, "Reading tenant config file")
3604 : }
3605 : }
3606 0 : })?;
3607 :
3608 0 : Ok(toml_edit::de::from_str::<LocationConf>(&config)?)
3609 0 : }
3610 :
3611 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
3612 : pub(super) async fn persist_tenant_config(
3613 : conf: &'static PageServerConf,
3614 : tenant_shard_id: &TenantShardId,
3615 : location_conf: &LocationConf,
3616 : ) -> std::io::Result<()> {
3617 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
3618 :
3619 : Self::persist_tenant_config_at(tenant_shard_id, &config_path, location_conf).await
3620 : }
3621 :
3622 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
3623 : pub(super) async fn persist_tenant_config_at(
3624 : tenant_shard_id: &TenantShardId,
3625 : config_path: &Utf8Path,
3626 : location_conf: &LocationConf,
3627 : ) -> std::io::Result<()> {
3628 : debug!("persisting tenantconf to {config_path}");
3629 :
3630 : let mut conf_content = r#"# This file contains a specific per-tenant's config.
3631 : # It is read in case of pageserver restart.
3632 : "#
3633 : .to_string();
3634 :
3635 0 : fail::fail_point!("tenant-config-before-write", |_| {
3636 0 : Err(std::io::Error::new(
3637 0 : std::io::ErrorKind::Other,
3638 0 : "tenant-config-before-write",
3639 0 : ))
3640 0 : });
3641 :
3642 : // Convert the config to a toml file.
3643 : conf_content +=
3644 : &toml_edit::ser::to_string_pretty(&location_conf).expect("Config serialization failed");
3645 :
3646 : let temp_path = path_with_suffix_extension(config_path, TEMP_FILE_SUFFIX);
3647 :
3648 : let conf_content = conf_content.into_bytes();
3649 : VirtualFile::crashsafe_overwrite(config_path.to_owned(), temp_path, conf_content).await
3650 : }
3651 :
3652 : //
3653 : // How garbage collection works:
3654 : //
3655 : // +--bar------------->
3656 : // /
3657 : // +----+-----foo---------------->
3658 : // /
3659 : // ----main--+-------------------------->
3660 : // \
3661 : // +-----baz-------->
3662 : //
3663 : //
3664 : // 1. Grab 'gc_cs' mutex to prevent new timelines from being created while Timeline's
3665 : // `gc_infos` are being refreshed
3666 : // 2. Scan collected timelines, and on each timeline, make note of the
3667 : // all the points where other timelines have been branched off.
3668 : // We will refrain from removing page versions at those LSNs.
3669 : // 3. For each timeline, scan all layer files on the timeline.
3670 : // Remove all files for which a newer file exists and which
3671 : // don't cover any branch point LSNs.
3672 : //
3673 : // TODO:
3674 : // - if a relation has a non-incremental persistent layer on a child branch, then we
3675 : // don't need to keep that in the parent anymore. But currently
3676 : // we do.
3677 4 : async fn gc_iteration_internal(
3678 4 : &self,
3679 4 : target_timeline_id: Option<TimelineId>,
3680 4 : horizon: u64,
3681 4 : pitr: Duration,
3682 4 : cancel: &CancellationToken,
3683 4 : ctx: &RequestContext,
3684 4 : ) -> Result<GcResult, GcError> {
3685 4 : let mut totals: GcResult = Default::default();
3686 4 : let now = Instant::now();
3687 :
3688 4 : let gc_timelines = self
3689 4 : .refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
3690 4 : .await?;
3691 :
3692 4 : failpoint_support::sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
3693 :
3694 : // If there is nothing to GC, we don't want any messages in the INFO log.
3695 4 : if !gc_timelines.is_empty() {
3696 4 : info!("{} timelines need GC", gc_timelines.len());
3697 : } else {
3698 0 : debug!("{} timelines need GC", gc_timelines.len());
3699 : }
3700 :
3701 : // Perform GC for each timeline.
3702 : //
3703 : // Note that we don't hold the `Tenant::gc_cs` lock here because we don't want to delay the
3704 : // branch creation task, which requires the GC lock. A GC iteration can run concurrently
3705 : // with branch creation.
3706 : //
3707 : // See comments in [`Tenant::branch_timeline`] for more information about why branch
3708 : // creation task can run concurrently with timeline's GC iteration.
3709 8 : for timeline in gc_timelines {
3710 4 : if cancel.is_cancelled() {
3711 : // We were requested to shut down. Stop and return with the progress we
3712 : // made.
3713 0 : break;
3714 4 : }
3715 4 : let result = match timeline.gc().await {
3716 : Err(GcError::TimelineCancelled) => {
3717 0 : if target_timeline_id.is_some() {
3718 : // If we were targetting this specific timeline, surface cancellation to caller
3719 0 : return Err(GcError::TimelineCancelled);
3720 : } else {
3721 : // A timeline may be shutting down independently of the tenant's lifecycle: we should
3722 : // skip past this and proceed to try GC on other timelines.
3723 0 : continue;
3724 : }
3725 : }
3726 4 : r => r?,
3727 : };
3728 4 : totals += result;
3729 : }
3730 :
3731 4 : totals.elapsed = now.elapsed();
3732 4 : Ok(totals)
3733 4 : }
3734 :
3735 : /// Refreshes the Timeline::gc_info for all timelines, returning the
3736 : /// vector of timelines which have [`Timeline::get_last_record_lsn`] past
3737 : /// [`Tenant::get_gc_horizon`].
3738 : ///
3739 : /// This is usually executed as part of periodic gc, but can now be triggered more often.
3740 0 : pub(crate) async fn refresh_gc_info(
3741 0 : &self,
3742 0 : cancel: &CancellationToken,
3743 0 : ctx: &RequestContext,
3744 0 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
3745 0 : // since this method can now be called at different rates than the configured gc loop, it
3746 0 : // might be that these configuration values get applied faster than what it was previously,
3747 0 : // since these were only read from the gc task.
3748 0 : let horizon = self.get_gc_horizon();
3749 0 : let pitr = self.get_pitr_interval();
3750 0 :
3751 0 : // refresh all timelines
3752 0 : let target_timeline_id = None;
3753 0 :
3754 0 : self.refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
3755 0 : .await
3756 0 : }
3757 :
3758 : /// Populate all Timelines' `GcInfo` with information about their children. We do not set the
3759 : /// PITR cutoffs here, because that requires I/O: this is done later, before GC, by [`Self::refresh_gc_info_internal`]
3760 : ///
3761 : /// Subsequently, parent-child relationships are updated incrementally inside [`Timeline::new`] and [`Timeline::drop`].
3762 0 : fn initialize_gc_info(
3763 0 : &self,
3764 0 : timelines: &std::sync::MutexGuard<HashMap<TimelineId, Arc<Timeline>>>,
3765 0 : timelines_offloaded: &std::sync::MutexGuard<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
3766 0 : ) {
3767 0 : // This function must be called before activation: after activation timeline create/delete operations
3768 0 : // might happen, and this function is not safe to run concurrently with those.
3769 0 : assert!(!self.is_active());
3770 :
3771 : // Scan all timelines. For each timeline, remember the timeline ID and
3772 : // the branch point where it was created.
3773 0 : let mut all_branchpoints: BTreeMap<TimelineId, Vec<(Lsn, TimelineId, MaybeOffloaded)>> =
3774 0 : BTreeMap::new();
3775 0 : timelines.iter().for_each(|(timeline_id, timeline_entry)| {
3776 0 : if let Some(ancestor_timeline_id) = &timeline_entry.get_ancestor_timeline_id() {
3777 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
3778 0 : ancestor_children.push((
3779 0 : timeline_entry.get_ancestor_lsn(),
3780 0 : *timeline_id,
3781 0 : MaybeOffloaded::No,
3782 0 : ));
3783 0 : }
3784 0 : });
3785 0 : timelines_offloaded
3786 0 : .iter()
3787 0 : .for_each(|(timeline_id, timeline_entry)| {
3788 0 : let Some(ancestor_timeline_id) = &timeline_entry.ancestor_timeline_id else {
3789 0 : return;
3790 : };
3791 0 : let Some(retain_lsn) = timeline_entry.ancestor_retain_lsn else {
3792 0 : return;
3793 : };
3794 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
3795 0 : ancestor_children.push((retain_lsn, *timeline_id, MaybeOffloaded::Yes));
3796 0 : });
3797 0 :
3798 0 : // The number of bytes we always keep, irrespective of PITR: this is a constant across timelines
3799 0 : let horizon = self.get_gc_horizon();
3800 :
3801 : // Populate each timeline's GcInfo with information about its child branches
3802 0 : for timeline in timelines.values() {
3803 0 : let mut branchpoints: Vec<(Lsn, TimelineId, MaybeOffloaded)> = all_branchpoints
3804 0 : .remove(&timeline.timeline_id)
3805 0 : .unwrap_or_default();
3806 0 :
3807 0 : branchpoints.sort_by_key(|b| b.0);
3808 0 :
3809 0 : let mut target = timeline.gc_info.write().unwrap();
3810 0 :
3811 0 : target.retain_lsns = branchpoints;
3812 0 :
3813 0 : let space_cutoff = timeline
3814 0 : .get_last_record_lsn()
3815 0 : .checked_sub(horizon)
3816 0 : .unwrap_or(Lsn(0));
3817 0 :
3818 0 : target.cutoffs = GcCutoffs {
3819 0 : space: space_cutoff,
3820 0 : time: Lsn::INVALID,
3821 0 : };
3822 0 : }
3823 0 : }
3824 :
3825 4 : async fn refresh_gc_info_internal(
3826 4 : &self,
3827 4 : target_timeline_id: Option<TimelineId>,
3828 4 : horizon: u64,
3829 4 : pitr: Duration,
3830 4 : cancel: &CancellationToken,
3831 4 : ctx: &RequestContext,
3832 4 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
3833 4 : // before taking the gc_cs lock, do the heavier weight finding of gc_cutoff points for
3834 4 : // currently visible timelines.
3835 4 : let timelines = self
3836 4 : .timelines
3837 4 : .lock()
3838 4 : .unwrap()
3839 4 : .values()
3840 4 : .filter(|tl| match target_timeline_id.as_ref() {
3841 4 : Some(target) => &tl.timeline_id == target,
3842 0 : None => true,
3843 4 : })
3844 4 : .cloned()
3845 4 : .collect::<Vec<_>>();
3846 4 :
3847 4 : if target_timeline_id.is_some() && timelines.is_empty() {
3848 : // We were to act on a particular timeline and it wasn't found
3849 0 : return Err(GcError::TimelineNotFound);
3850 4 : }
3851 4 :
3852 4 : let mut gc_cutoffs: HashMap<TimelineId, GcCutoffs> =
3853 4 : HashMap::with_capacity(timelines.len());
3854 :
3855 4 : for timeline in timelines.iter() {
3856 4 : let cutoff = timeline
3857 4 : .get_last_record_lsn()
3858 4 : .checked_sub(horizon)
3859 4 : .unwrap_or(Lsn(0));
3860 :
3861 4 : let cutoffs = timeline.find_gc_cutoffs(cutoff, pitr, cancel, ctx).await?;
3862 4 : let old = gc_cutoffs.insert(timeline.timeline_id, cutoffs);
3863 4 : assert!(old.is_none());
3864 : }
3865 :
3866 4 : if !self.is_active() || self.cancel.is_cancelled() {
3867 0 : return Err(GcError::TenantCancelled);
3868 4 : }
3869 :
3870 : // grab mutex to prevent new timelines from being created here; avoid doing long operations
3871 : // because that will stall branch creation.
3872 4 : let gc_cs = self.gc_cs.lock().await;
3873 :
3874 : // Ok, we now know all the branch points.
3875 : // Update the GC information for each timeline.
3876 4 : let mut gc_timelines = Vec::with_capacity(timelines.len());
3877 8 : for timeline in timelines {
3878 : // We filtered the timeline list above
3879 4 : if let Some(target_timeline_id) = target_timeline_id {
3880 4 : assert_eq!(target_timeline_id, timeline.timeline_id);
3881 0 : }
3882 :
3883 : {
3884 4 : let mut target = timeline.gc_info.write().unwrap();
3885 4 :
3886 4 : // Cull any expired leases
3887 4 : let now = SystemTime::now();
3888 6 : target.leases.retain(|_, lease| !lease.is_expired(&now));
3889 4 :
3890 4 : timeline
3891 4 : .metrics
3892 4 : .valid_lsn_lease_count_gauge
3893 4 : .set(target.leases.len() as u64);
3894 :
3895 : // Look up parent's PITR cutoff to update the child's knowledge of whether it is within parent's PITR
3896 4 : if let Some(ancestor_id) = timeline.get_ancestor_timeline_id() {
3897 0 : if let Some(ancestor_gc_cutoffs) = gc_cutoffs.get(&ancestor_id) {
3898 0 : target.within_ancestor_pitr =
3899 0 : timeline.get_ancestor_lsn() >= ancestor_gc_cutoffs.time;
3900 0 : }
3901 4 : }
3902 :
3903 : // Update metrics that depend on GC state
3904 4 : timeline
3905 4 : .metrics
3906 4 : .archival_size
3907 4 : .set(if target.within_ancestor_pitr {
3908 0 : timeline.metrics.current_logical_size_gauge.get()
3909 : } else {
3910 4 : 0
3911 : });
3912 4 : timeline.metrics.pitr_history_size.set(
3913 4 : timeline
3914 4 : .get_last_record_lsn()
3915 4 : .checked_sub(target.cutoffs.time)
3916 4 : .unwrap_or(Lsn(0))
3917 4 : .0,
3918 4 : );
3919 :
3920 : // Apply the cutoffs we found to the Timeline's GcInfo. Why might we _not_ have cutoffs for a timeline?
3921 : // - this timeline was created while we were finding cutoffs
3922 : // - lsn for timestamp search fails for this timeline repeatedly
3923 4 : if let Some(cutoffs) = gc_cutoffs.get(&timeline.timeline_id) {
3924 4 : target.cutoffs = cutoffs.clone();
3925 4 : }
3926 : }
3927 :
3928 4 : gc_timelines.push(timeline);
3929 : }
3930 4 : drop(gc_cs);
3931 4 : Ok(gc_timelines)
3932 4 : }
3933 :
3934 : /// A substitute for `branch_timeline` for use in unit tests.
3935 : /// The returned timeline will have state value `Active` to make various `anyhow::ensure!()`
3936 : /// calls pass, but, we do not actually call `.activate()` under the hood. So, none of the
3937 : /// timeline background tasks are launched, except the flush loop.
3938 : #[cfg(test)]
3939 230 : async fn branch_timeline_test(
3940 230 : self: &Arc<Self>,
3941 230 : src_timeline: &Arc<Timeline>,
3942 230 : dst_id: TimelineId,
3943 230 : ancestor_lsn: Option<Lsn>,
3944 230 : ctx: &RequestContext,
3945 230 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
3946 230 : let tl = self
3947 230 : .branch_timeline_impl(src_timeline, dst_id, ancestor_lsn, ctx)
3948 193 : .await?
3949 226 : .into_timeline_for_test();
3950 226 : tl.set_state(TimelineState::Active);
3951 226 : Ok(tl)
3952 230 : }
3953 :
3954 : /// Helper for unit tests to branch a timeline with some pre-loaded states.
3955 : #[cfg(test)]
3956 : #[allow(clippy::too_many_arguments)]
3957 6 : pub async fn branch_timeline_test_with_layers(
3958 6 : self: &Arc<Self>,
3959 6 : src_timeline: &Arc<Timeline>,
3960 6 : dst_id: TimelineId,
3961 6 : ancestor_lsn: Option<Lsn>,
3962 6 : ctx: &RequestContext,
3963 6 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
3964 6 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
3965 6 : end_lsn: Lsn,
3966 6 : ) -> anyhow::Result<Arc<Timeline>> {
3967 : use checks::check_valid_layermap;
3968 : use itertools::Itertools;
3969 :
3970 6 : let tline = self
3971 6 : .branch_timeline_test(src_timeline, dst_id, ancestor_lsn, ctx)
3972 6 : .await?;
3973 6 : let ancestor_lsn = if let Some(ancestor_lsn) = ancestor_lsn {
3974 6 : ancestor_lsn
3975 : } else {
3976 0 : tline.get_last_record_lsn()
3977 : };
3978 6 : assert!(end_lsn >= ancestor_lsn);
3979 6 : tline.force_advance_lsn(end_lsn);
3980 12 : for deltas in delta_layer_desc {
3981 6 : tline
3982 6 : .force_create_delta_layer(deltas, Some(ancestor_lsn), ctx)
3983 18 : .await?;
3984 : }
3985 10 : for (lsn, images) in image_layer_desc {
3986 4 : tline
3987 4 : .force_create_image_layer(lsn, images, Some(ancestor_lsn), ctx)
3988 14 : .await?;
3989 : }
3990 6 : let layer_names = tline
3991 6 : .layers
3992 6 : .read()
3993 0 : .await
3994 6 : .layer_map()
3995 6 : .unwrap()
3996 6 : .iter_historic_layers()
3997 10 : .map(|layer| layer.layer_name())
3998 6 : .collect_vec();
3999 6 : if let Some(err) = check_valid_layermap(&layer_names) {
4000 0 : bail!("invalid layermap: {err}");
4001 6 : }
4002 6 : Ok(tline)
4003 6 : }
4004 :
4005 : /// Branch an existing timeline.
4006 0 : async fn branch_timeline(
4007 0 : self: &Arc<Self>,
4008 0 : src_timeline: &Arc<Timeline>,
4009 0 : dst_id: TimelineId,
4010 0 : start_lsn: Option<Lsn>,
4011 0 : ctx: &RequestContext,
4012 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4013 0 : self.branch_timeline_impl(src_timeline, dst_id, start_lsn, ctx)
4014 0 : .await
4015 0 : }
4016 :
4017 230 : async fn branch_timeline_impl(
4018 230 : self: &Arc<Self>,
4019 230 : src_timeline: &Arc<Timeline>,
4020 230 : dst_id: TimelineId,
4021 230 : start_lsn: Option<Lsn>,
4022 230 : _ctx: &RequestContext,
4023 230 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4024 230 : let src_id = src_timeline.timeline_id;
4025 :
4026 : // We will validate our ancestor LSN in this function. Acquire the GC lock so that
4027 : // this check cannot race with GC, and the ancestor LSN is guaranteed to remain
4028 : // valid while we are creating the branch.
4029 230 : let _gc_cs = self.gc_cs.lock().await;
4030 :
4031 : // If no start LSN is specified, we branch the new timeline from the source timeline's last record LSN
4032 230 : let start_lsn = start_lsn.unwrap_or_else(|| {
4033 2 : let lsn = src_timeline.get_last_record_lsn();
4034 2 : info!("branching timeline {dst_id} from timeline {src_id} at last record LSN: {lsn}");
4035 2 : lsn
4036 230 : });
4037 :
4038 : // we finally have determined the ancestor_start_lsn, so we can get claim exclusivity now
4039 230 : let timeline_create_guard = match self
4040 230 : .start_creating_timeline(
4041 230 : dst_id,
4042 230 : CreateTimelineIdempotency::Branch {
4043 230 : ancestor_timeline_id: src_timeline.timeline_id,
4044 230 : ancestor_start_lsn: start_lsn,
4045 230 : },
4046 230 : )
4047 192 : .await?
4048 : {
4049 230 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
4050 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
4051 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
4052 : }
4053 : };
4054 :
4055 : // Ensure that `start_lsn` is valid, i.e. the LSN is within the PITR
4056 : // horizon on the source timeline
4057 : //
4058 : // We check it against both the planned GC cutoff stored in 'gc_info',
4059 : // and the 'latest_gc_cutoff' of the last GC that was performed. The
4060 : // planned GC cutoff in 'gc_info' is normally larger than
4061 : // 'latest_gc_cutoff_lsn', but beware of corner cases like if you just
4062 : // changed the GC settings for the tenant to make the PITR window
4063 : // larger, but some of the data was already removed by an earlier GC
4064 : // iteration.
4065 :
4066 : // check against last actual 'latest_gc_cutoff' first
4067 230 : let latest_gc_cutoff_lsn = src_timeline.get_latest_gc_cutoff_lsn();
4068 230 : src_timeline
4069 230 : .check_lsn_is_in_scope(start_lsn, &latest_gc_cutoff_lsn)
4070 230 : .context(format!(
4071 230 : "invalid branch start lsn: less than latest GC cutoff {}",
4072 230 : *latest_gc_cutoff_lsn,
4073 230 : ))
4074 230 : .map_err(CreateTimelineError::AncestorLsn)?;
4075 :
4076 : // and then the planned GC cutoff
4077 : {
4078 226 : let gc_info = src_timeline.gc_info.read().unwrap();
4079 226 : let cutoff = gc_info.min_cutoff();
4080 226 : if start_lsn < cutoff {
4081 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
4082 0 : "invalid branch start lsn: less than planned GC cutoff {cutoff}"
4083 0 : )));
4084 226 : }
4085 226 : }
4086 226 :
4087 226 : //
4088 226 : // The branch point is valid, and we are still holding the 'gc_cs' lock
4089 226 : // so that GC cannot advance the GC cutoff until we are finished.
4090 226 : // Proceed with the branch creation.
4091 226 : //
4092 226 :
4093 226 : // Determine prev-LSN for the new timeline. We can only determine it if
4094 226 : // the timeline was branched at the current end of the source timeline.
4095 226 : let RecordLsn {
4096 226 : last: src_last,
4097 226 : prev: src_prev,
4098 226 : } = src_timeline.get_last_record_rlsn();
4099 226 : let dst_prev = if src_last == start_lsn {
4100 216 : Some(src_prev)
4101 : } else {
4102 10 : None
4103 : };
4104 :
4105 : // Create the metadata file, noting the ancestor of the new timeline.
4106 : // There is initially no data in it, but all the read-calls know to look
4107 : // into the ancestor.
4108 226 : let metadata = TimelineMetadata::new(
4109 226 : start_lsn,
4110 226 : dst_prev,
4111 226 : Some(src_id),
4112 226 : start_lsn,
4113 226 : *src_timeline.latest_gc_cutoff_lsn.read(), // FIXME: should we hold onto this guard longer?
4114 226 : src_timeline.initdb_lsn,
4115 226 : src_timeline.pg_version,
4116 226 : );
4117 :
4118 226 : let uninitialized_timeline = self
4119 226 : .prepare_new_timeline(
4120 226 : dst_id,
4121 226 : &metadata,
4122 226 : timeline_create_guard,
4123 226 : start_lsn + 1,
4124 226 : Some(Arc::clone(src_timeline)),
4125 226 : )
4126 0 : .await?;
4127 :
4128 226 : let new_timeline = uninitialized_timeline.finish_creation()?;
4129 :
4130 : // Root timeline gets its layers during creation and uploads them along with the metadata.
4131 : // A branch timeline though, when created, can get no writes for some time, hence won't get any layers created.
4132 : // We still need to upload its metadata eagerly: if other nodes `attach` the tenant and miss this timeline, their GC
4133 : // could get incorrect information and remove more layers, than needed.
4134 : // See also https://github.com/neondatabase/neon/issues/3865
4135 226 : new_timeline
4136 226 : .remote_client
4137 226 : .schedule_index_upload_for_full_metadata_update(&metadata)
4138 226 : .context("branch initial metadata upload")?;
4139 :
4140 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
4141 :
4142 226 : Ok(CreateTimelineResult::Created(new_timeline))
4143 230 : }
4144 :
4145 : /// For unit tests, make this visible so that other modules can directly create timelines
4146 : #[cfg(test)]
4147 2 : #[tracing::instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
4148 : pub(crate) async fn bootstrap_timeline_test(
4149 : self: &Arc<Self>,
4150 : timeline_id: TimelineId,
4151 : pg_version: u32,
4152 : load_existing_initdb: Option<TimelineId>,
4153 : ctx: &RequestContext,
4154 : ) -> anyhow::Result<Arc<Timeline>> {
4155 : self.bootstrap_timeline(timeline_id, pg_version, load_existing_initdb, ctx)
4156 : .await
4157 : .map_err(anyhow::Error::new)
4158 2 : .map(|r| r.into_timeline_for_test())
4159 : }
4160 :
4161 : /// Get exclusive access to the timeline ID for creation.
4162 : ///
4163 : /// Timeline-creating code paths must use this function before making changes
4164 : /// to in-memory or persistent state.
4165 : ///
4166 : /// The `state` parameter is a description of the timeline creation operation
4167 : /// we intend to perform.
4168 : /// If the timeline was already created in the meantime, we check whether this
4169 : /// request conflicts or is idempotent , based on `state`.
4170 414 : async fn start_creating_timeline(
4171 414 : &self,
4172 414 : new_timeline_id: TimelineId,
4173 414 : idempotency: CreateTimelineIdempotency,
4174 414 : ) -> Result<StartCreatingTimelineResult<'_>, CreateTimelineError> {
4175 414 : let allow_offloaded = false;
4176 414 : match self.create_timeline_create_guard(new_timeline_id, idempotency, allow_offloaded) {
4177 412 : Ok(create_guard) => {
4178 412 : pausable_failpoint!("timeline-creation-after-uninit");
4179 412 : Ok(StartCreatingTimelineResult::CreateGuard(create_guard))
4180 : }
4181 : Err(TimelineExclusionError::AlreadyCreating) => {
4182 : // Creation is in progress, we cannot create it again, and we cannot
4183 : // check if this request matches the existing one, so caller must try
4184 : // again later.
4185 0 : Err(CreateTimelineError::AlreadyCreating)
4186 : }
4187 0 : Err(TimelineExclusionError::Other(e)) => Err(CreateTimelineError::Other(e)),
4188 : Err(TimelineExclusionError::AlreadyExists {
4189 0 : existing: TimelineOrOffloaded::Offloaded(_existing),
4190 0 : ..
4191 0 : }) => {
4192 0 : info!("timeline already exists but is offloaded");
4193 0 : Err(CreateTimelineError::Conflict)
4194 : }
4195 : Err(TimelineExclusionError::AlreadyExists {
4196 2 : existing: TimelineOrOffloaded::Timeline(existing),
4197 2 : arg,
4198 2 : }) => {
4199 2 : {
4200 2 : let existing = &existing.create_idempotency;
4201 2 : let _span = info_span!("idempotency_check", ?existing, ?arg).entered();
4202 2 : debug!("timeline already exists");
4203 :
4204 2 : match (existing, &arg) {
4205 : // FailWithConflict => no idempotency check
4206 : (CreateTimelineIdempotency::FailWithConflict, _)
4207 : | (_, CreateTimelineIdempotency::FailWithConflict) => {
4208 2 : warn!("timeline already exists, failing request");
4209 2 : return Err(CreateTimelineError::Conflict);
4210 : }
4211 : // Idempotent <=> CreateTimelineIdempotency is identical
4212 0 : (x, y) if x == y => {
4213 0 : info!("timeline already exists and idempotency matches, succeeding request");
4214 : // fallthrough
4215 : }
4216 : (_, _) => {
4217 0 : warn!("idempotency conflict, failing request");
4218 0 : return Err(CreateTimelineError::Conflict);
4219 : }
4220 : }
4221 : }
4222 :
4223 0 : Ok(StartCreatingTimelineResult::Idempotent(existing))
4224 : }
4225 : }
4226 414 : }
4227 :
4228 0 : async fn upload_initdb(
4229 0 : &self,
4230 0 : timelines_path: &Utf8PathBuf,
4231 0 : pgdata_path: &Utf8PathBuf,
4232 0 : timeline_id: &TimelineId,
4233 0 : ) -> anyhow::Result<()> {
4234 0 : let temp_path = timelines_path.join(format!(
4235 0 : "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
4236 0 : ));
4237 0 :
4238 0 : scopeguard::defer! {
4239 0 : if let Err(e) = fs::remove_file(&temp_path) {
4240 0 : error!("Failed to remove temporary initdb archive '{temp_path}': {e}");
4241 0 : }
4242 0 : }
4243 :
4244 0 : let (pgdata_zstd, tar_zst_size) = create_zst_tarball(pgdata_path, &temp_path).await?;
4245 : const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
4246 0 : if tar_zst_size > INITDB_TAR_ZST_WARN_LIMIT {
4247 0 : warn!(
4248 0 : "compressed {temp_path} size of {tar_zst_size} is above limit {INITDB_TAR_ZST_WARN_LIMIT}."
4249 : );
4250 0 : }
4251 :
4252 0 : pausable_failpoint!("before-initdb-upload");
4253 :
4254 0 : backoff::retry(
4255 0 : || async {
4256 0 : self::remote_timeline_client::upload_initdb_dir(
4257 0 : &self.remote_storage,
4258 0 : &self.tenant_shard_id.tenant_id,
4259 0 : timeline_id,
4260 0 : pgdata_zstd.try_clone().await?,
4261 0 : tar_zst_size,
4262 0 : &self.cancel,
4263 : )
4264 0 : .await
4265 0 : },
4266 0 : |_| false,
4267 0 : 3,
4268 0 : u32::MAX,
4269 0 : "persist_initdb_tar_zst",
4270 0 : &self.cancel,
4271 0 : )
4272 0 : .await
4273 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
4274 0 : .and_then(|x| x)
4275 0 : }
4276 :
4277 : /// - run initdb to init temporary instance and get bootstrap data
4278 : /// - after initialization completes, tar up the temp dir and upload it to S3.
4279 2 : async fn bootstrap_timeline(
4280 2 : self: &Arc<Self>,
4281 2 : timeline_id: TimelineId,
4282 2 : pg_version: u32,
4283 2 : load_existing_initdb: Option<TimelineId>,
4284 2 : ctx: &RequestContext,
4285 2 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4286 2 : let timeline_create_guard = match self
4287 2 : .start_creating_timeline(
4288 2 : timeline_id,
4289 2 : CreateTimelineIdempotency::Bootstrap { pg_version },
4290 2 : )
4291 2 : .await?
4292 : {
4293 2 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
4294 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
4295 0 : return Ok(CreateTimelineResult::Idempotent(timeline))
4296 : }
4297 : };
4298 :
4299 : // create a `tenant/{tenant_id}/timelines/basebackup-{timeline_id}.{TEMP_FILE_SUFFIX}/`
4300 : // temporary directory for basebackup files for the given timeline.
4301 :
4302 2 : let timelines_path = self.conf.timelines_path(&self.tenant_shard_id);
4303 2 : let pgdata_path = path_with_suffix_extension(
4304 2 : timelines_path.join(format!("basebackup-{timeline_id}")),
4305 2 : TEMP_FILE_SUFFIX,
4306 2 : );
4307 2 :
4308 2 : // Remove whatever was left from the previous runs: safe because TimelineCreateGuard guarantees
4309 2 : // we won't race with other creations or existent timelines with the same path.
4310 2 : if pgdata_path.exists() {
4311 0 : fs::remove_dir_all(&pgdata_path).with_context(|| {
4312 0 : format!("Failed to remove already existing initdb directory: {pgdata_path}")
4313 0 : })?;
4314 2 : }
4315 :
4316 : // this new directory is very temporary, set to remove it immediately after bootstrap, we don't need it
4317 2 : scopeguard::defer! {
4318 2 : if let Err(e) = fs::remove_dir_all(&pgdata_path) {
4319 2 : // this is unlikely, but we will remove the directory on pageserver restart or another bootstrap call
4320 2 : error!("Failed to remove temporary initdb directory '{pgdata_path}': {e}");
4321 2 : }
4322 2 : }
4323 2 : if let Some(existing_initdb_timeline_id) = load_existing_initdb {
4324 2 : if existing_initdb_timeline_id != timeline_id {
4325 0 : let source_path = &remote_initdb_archive_path(
4326 0 : &self.tenant_shard_id.tenant_id,
4327 0 : &existing_initdb_timeline_id,
4328 0 : );
4329 0 : let dest_path =
4330 0 : &remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &timeline_id);
4331 0 :
4332 0 : // if this fails, it will get retried by retried control plane requests
4333 0 : self.remote_storage
4334 0 : .copy_object(source_path, dest_path, &self.cancel)
4335 0 : .await
4336 0 : .context("copy initdb tar")?;
4337 2 : }
4338 2 : let (initdb_tar_zst_path, initdb_tar_zst) =
4339 2 : self::remote_timeline_client::download_initdb_tar_zst(
4340 2 : self.conf,
4341 2 : &self.remote_storage,
4342 2 : &self.tenant_shard_id,
4343 2 : &existing_initdb_timeline_id,
4344 2 : &self.cancel,
4345 2 : )
4346 725 : .await
4347 2 : .context("download initdb tar")?;
4348 :
4349 2 : scopeguard::defer! {
4350 2 : if let Err(e) = fs::remove_file(&initdb_tar_zst_path) {
4351 2 : error!("Failed to remove temporary initdb archive '{initdb_tar_zst_path}': {e}");
4352 2 : }
4353 2 : }
4354 2 :
4355 2 : let buf_read =
4356 2 : BufReader::with_capacity(remote_timeline_client::BUFFER_SIZE, initdb_tar_zst);
4357 2 : extract_zst_tarball(&pgdata_path, buf_read)
4358 10556 : .await
4359 2 : .context("extract initdb tar")?;
4360 : } else {
4361 : // Init temporarily repo to get bootstrap data, this creates a directory in the `pgdata_path` path
4362 0 : run_initdb(self.conf, &pgdata_path, pg_version, &self.cancel)
4363 0 : .await
4364 0 : .context("run initdb")?;
4365 :
4366 : // Upload the created data dir to S3
4367 0 : if self.tenant_shard_id().is_shard_zero() {
4368 0 : self.upload_initdb(&timelines_path, &pgdata_path, &timeline_id)
4369 0 : .await?;
4370 0 : }
4371 : }
4372 2 : let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
4373 2 :
4374 2 : // Import the contents of the data directory at the initial checkpoint
4375 2 : // LSN, and any WAL after that.
4376 2 : // Initdb lsn will be equal to last_record_lsn which will be set after import.
4377 2 : // Because we know it upfront avoid having an option or dummy zero value by passing it to the metadata.
4378 2 : let new_metadata = TimelineMetadata::new(
4379 2 : Lsn(0),
4380 2 : None,
4381 2 : None,
4382 2 : Lsn(0),
4383 2 : pgdata_lsn,
4384 2 : pgdata_lsn,
4385 2 : pg_version,
4386 2 : );
4387 2 : let raw_timeline = self
4388 2 : .prepare_new_timeline(
4389 2 : timeline_id,
4390 2 : &new_metadata,
4391 2 : timeline_create_guard,
4392 2 : pgdata_lsn,
4393 2 : None,
4394 2 : )
4395 0 : .await?;
4396 :
4397 2 : let tenant_shard_id = raw_timeline.owning_tenant.tenant_shard_id;
4398 2 : let unfinished_timeline = raw_timeline.raw_timeline()?;
4399 :
4400 : // Flush the new layer files to disk, before we make the timeline as available to
4401 : // the outside world.
4402 : //
4403 : // Flush loop needs to be spawned in order to be able to flush.
4404 2 : unfinished_timeline.maybe_spawn_flush_loop();
4405 2 :
4406 2 : import_datadir::import_timeline_from_postgres_datadir(
4407 2 : unfinished_timeline,
4408 2 : &pgdata_path,
4409 2 : pgdata_lsn,
4410 2 : ctx,
4411 2 : )
4412 9522 : .await
4413 2 : .with_context(|| {
4414 0 : format!("Failed to import pgdatadir for timeline {tenant_shard_id}/{timeline_id}")
4415 2 : })?;
4416 :
4417 2 : fail::fail_point!("before-checkpoint-new-timeline", |_| {
4418 0 : Err(CreateTimelineError::Other(anyhow::anyhow!(
4419 0 : "failpoint before-checkpoint-new-timeline"
4420 0 : )))
4421 2 : });
4422 :
4423 2 : unfinished_timeline
4424 2 : .freeze_and_flush()
4425 2 : .await
4426 2 : .with_context(|| {
4427 0 : format!(
4428 0 : "Failed to flush after pgdatadir import for timeline {tenant_shard_id}/{timeline_id}"
4429 0 : )
4430 2 : })?;
4431 :
4432 : // All done!
4433 2 : let timeline = raw_timeline.finish_creation()?;
4434 :
4435 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
4436 :
4437 2 : Ok(CreateTimelineResult::Created(timeline))
4438 2 : }
4439 :
4440 408 : fn build_timeline_remote_client(&self, timeline_id: TimelineId) -> RemoteTimelineClient {
4441 408 : RemoteTimelineClient::new(
4442 408 : self.remote_storage.clone(),
4443 408 : self.deletion_queue_client.clone(),
4444 408 : self.conf,
4445 408 : self.tenant_shard_id,
4446 408 : timeline_id,
4447 408 : self.generation,
4448 408 : )
4449 408 : }
4450 :
4451 : /// Call this before constructing a timeline, to build its required structures
4452 408 : fn build_timeline_resources(&self, timeline_id: TimelineId) -> TimelineResources {
4453 408 : TimelineResources {
4454 408 : remote_client: self.build_timeline_remote_client(timeline_id),
4455 408 : timeline_get_throttle: self.timeline_get_throttle.clone(),
4456 408 : l0_flush_global_state: self.l0_flush_global_state.clone(),
4457 408 : }
4458 408 : }
4459 :
4460 : /// Creates intermediate timeline structure and its files.
4461 : ///
4462 : /// An empty layer map is initialized, and new data and WAL can be imported starting
4463 : /// at 'disk_consistent_lsn'. After any initial data has been imported, call
4464 : /// `finish_creation` to insert the Timeline into the timelines map.
4465 408 : async fn prepare_new_timeline<'a>(
4466 408 : &'a self,
4467 408 : new_timeline_id: TimelineId,
4468 408 : new_metadata: &TimelineMetadata,
4469 408 : create_guard: TimelineCreateGuard<'a>,
4470 408 : start_lsn: Lsn,
4471 408 : ancestor: Option<Arc<Timeline>>,
4472 408 : ) -> anyhow::Result<UninitializedTimeline<'a>> {
4473 408 : let tenant_shard_id = self.tenant_shard_id;
4474 408 :
4475 408 : let resources = self.build_timeline_resources(new_timeline_id);
4476 408 : resources
4477 408 : .remote_client
4478 408 : .init_upload_queue_for_empty_remote(new_metadata)?;
4479 :
4480 408 : let timeline_struct = self
4481 408 : .create_timeline_struct(
4482 408 : new_timeline_id,
4483 408 : new_metadata,
4484 408 : ancestor,
4485 408 : resources,
4486 408 : CreateTimelineCause::Load,
4487 408 : create_guard.idempotency.clone(),
4488 408 : )
4489 408 : .context("Failed to create timeline data structure")?;
4490 :
4491 408 : timeline_struct.init_empty_layer_map(start_lsn);
4492 :
4493 408 : if let Err(e) = self
4494 408 : .create_timeline_files(&create_guard.timeline_path)
4495 0 : .await
4496 : {
4497 0 : error!("Failed to create initial files for timeline {tenant_shard_id}/{new_timeline_id}, cleaning up: {e:?}");
4498 0 : cleanup_timeline_directory(create_guard);
4499 0 : return Err(e);
4500 408 : }
4501 408 :
4502 408 : debug!(
4503 0 : "Successfully created initial files for timeline {tenant_shard_id}/{new_timeline_id}"
4504 : );
4505 :
4506 408 : Ok(UninitializedTimeline::new(
4507 408 : self,
4508 408 : new_timeline_id,
4509 408 : Some((timeline_struct, create_guard)),
4510 408 : ))
4511 408 : }
4512 :
4513 408 : async fn create_timeline_files(&self, timeline_path: &Utf8Path) -> anyhow::Result<()> {
4514 408 : crashsafe::create_dir(timeline_path).context("Failed to create timeline directory")?;
4515 :
4516 408 : fail::fail_point!("after-timeline-dir-creation", |_| {
4517 0 : anyhow::bail!("failpoint after-timeline-dir-creation");
4518 408 : });
4519 :
4520 408 : Ok(())
4521 408 : }
4522 :
4523 : /// Get a guard that provides exclusive access to the timeline directory, preventing
4524 : /// concurrent attempts to create the same timeline.
4525 : ///
4526 : /// The `allow_offloaded` parameter controls whether to tolerate the existence of
4527 : /// offloaded timelines or not.
4528 414 : fn create_timeline_create_guard(
4529 414 : &self,
4530 414 : timeline_id: TimelineId,
4531 414 : idempotency: CreateTimelineIdempotency,
4532 414 : allow_offloaded: bool,
4533 414 : ) -> Result<TimelineCreateGuard, TimelineExclusionError> {
4534 414 : let tenant_shard_id = self.tenant_shard_id;
4535 414 :
4536 414 : let timeline_path = self.conf.timeline_path(&tenant_shard_id, &timeline_id);
4537 :
4538 414 : let create_guard = TimelineCreateGuard::new(
4539 414 : self,
4540 414 : timeline_id,
4541 414 : timeline_path.clone(),
4542 414 : idempotency,
4543 414 : allow_offloaded,
4544 414 : )?;
4545 :
4546 : // At this stage, we have got exclusive access to in-memory state for this timeline ID
4547 : // for creation.
4548 : // A timeline directory should never exist on disk already:
4549 : // - a previous failed creation would have cleaned up after itself
4550 : // - a pageserver restart would clean up timeline directories that don't have valid remote state
4551 : //
4552 : // Therefore it is an unexpected internal error to encounter a timeline directory already existing here,
4553 : // this error may indicate a bug in cleanup on failed creations.
4554 412 : if timeline_path.exists() {
4555 0 : return Err(TimelineExclusionError::Other(anyhow::anyhow!(
4556 0 : "Timeline directory already exists! This is a bug."
4557 0 : )));
4558 412 : }
4559 412 :
4560 412 : Ok(create_guard)
4561 414 : }
4562 :
4563 : /// Gathers inputs from all of the timelines to produce a sizing model input.
4564 : ///
4565 : /// Future is cancellation safe. Only one calculation can be running at once per tenant.
4566 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
4567 : pub async fn gather_size_inputs(
4568 : &self,
4569 : // `max_retention_period` overrides the cutoff that is used to calculate the size
4570 : // (only if it is shorter than the real cutoff).
4571 : max_retention_period: Option<u64>,
4572 : cause: LogicalSizeCalculationCause,
4573 : cancel: &CancellationToken,
4574 : ctx: &RequestContext,
4575 : ) -> Result<size::ModelInputs, size::CalculateSyntheticSizeError> {
4576 : let logical_sizes_at_once = self
4577 : .conf
4578 : .concurrent_tenant_size_logical_size_queries
4579 : .inner();
4580 :
4581 : // TODO: Having a single mutex block concurrent reads is not great for performance.
4582 : //
4583 : // But the only case where we need to run multiple of these at once is when we
4584 : // request a size for a tenant manually via API, while another background calculation
4585 : // is in progress (which is not a common case).
4586 : //
4587 : // See more for on the issue #2748 condenced out of the initial PR review.
4588 : let mut shared_cache = tokio::select! {
4589 : locked = self.cached_logical_sizes.lock() => locked,
4590 : _ = cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
4591 : _ = self.cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
4592 : };
4593 :
4594 : size::gather_inputs(
4595 : self,
4596 : logical_sizes_at_once,
4597 : max_retention_period,
4598 : &mut shared_cache,
4599 : cause,
4600 : cancel,
4601 : ctx,
4602 : )
4603 : .await
4604 : }
4605 :
4606 : /// Calculate synthetic tenant size and cache the result.
4607 : /// This is periodically called by background worker.
4608 : /// result is cached in tenant struct
4609 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
4610 : pub async fn calculate_synthetic_size(
4611 : &self,
4612 : cause: LogicalSizeCalculationCause,
4613 : cancel: &CancellationToken,
4614 : ctx: &RequestContext,
4615 : ) -> Result<u64, size::CalculateSyntheticSizeError> {
4616 : let inputs = self.gather_size_inputs(None, cause, cancel, ctx).await?;
4617 :
4618 : let size = inputs.calculate();
4619 :
4620 : self.set_cached_synthetic_size(size);
4621 :
4622 : Ok(size)
4623 : }
4624 :
4625 : /// Cache given synthetic size and update the metric value
4626 0 : pub fn set_cached_synthetic_size(&self, size: u64) {
4627 0 : self.cached_synthetic_tenant_size
4628 0 : .store(size, Ordering::Relaxed);
4629 0 :
4630 0 : // Only shard zero should be calculating synthetic sizes
4631 0 : debug_assert!(self.shard_identity.is_shard_zero());
4632 :
4633 0 : TENANT_SYNTHETIC_SIZE_METRIC
4634 0 : .get_metric_with_label_values(&[&self.tenant_shard_id.tenant_id.to_string()])
4635 0 : .unwrap()
4636 0 : .set(size);
4637 0 : }
4638 :
4639 0 : pub fn cached_synthetic_size(&self) -> u64 {
4640 0 : self.cached_synthetic_tenant_size.load(Ordering::Relaxed)
4641 0 : }
4642 :
4643 : /// Flush any in-progress layers, schedule uploads, and wait for uploads to complete.
4644 : ///
4645 : /// This function can take a long time: callers should wrap it in a timeout if calling
4646 : /// from an external API handler.
4647 : ///
4648 : /// Cancel-safety: cancelling this function may leave I/O running, but such I/O is
4649 : /// still bounded by tenant/timeline shutdown.
4650 0 : #[tracing::instrument(skip_all)]
4651 : pub(crate) async fn flush_remote(&self) -> anyhow::Result<()> {
4652 : let timelines = self.timelines.lock().unwrap().clone();
4653 :
4654 0 : async fn flush_timeline(_gate: GateGuard, timeline: Arc<Timeline>) -> anyhow::Result<()> {
4655 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Flushing...");
4656 0 : timeline.freeze_and_flush().await?;
4657 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Waiting for uploads...");
4658 0 : timeline.remote_client.wait_completion().await?;
4659 :
4660 0 : Ok(())
4661 0 : }
4662 :
4663 : // We do not use a JoinSet for these tasks, because we don't want them to be
4664 : // aborted when this function's future is cancelled: they should stay alive
4665 : // holding their GateGuard until they complete, to ensure their I/Os complete
4666 : // before Timeline shutdown completes.
4667 : let mut results = FuturesUnordered::new();
4668 :
4669 : for (_timeline_id, timeline) in timelines {
4670 : // Run each timeline's flush in a task holding the timeline's gate: this
4671 : // means that if this function's future is cancelled, the Timeline shutdown
4672 : // will still wait for any I/O in here to complete.
4673 : let Ok(gate) = timeline.gate.enter() else {
4674 : continue;
4675 : };
4676 0 : let jh = tokio::task::spawn(async move { flush_timeline(gate, timeline).await });
4677 : results.push(jh);
4678 : }
4679 :
4680 : while let Some(r) = results.next().await {
4681 : if let Err(e) = r {
4682 : if !e.is_cancelled() && !e.is_panic() {
4683 : tracing::error!("unexpected join error: {e:?}");
4684 : }
4685 : }
4686 : }
4687 :
4688 : // The flushes we did above were just writes, but the Tenant might have had
4689 : // pending deletions as well from recent compaction/gc: we want to flush those
4690 : // as well. This requires flushing the global delete queue. This is cheap
4691 : // because it's typically a no-op.
4692 : match self.deletion_queue_client.flush_execute().await {
4693 : Ok(_) => {}
4694 : Err(DeletionQueueError::ShuttingDown) => {}
4695 : }
4696 :
4697 : Ok(())
4698 : }
4699 :
4700 0 : pub(crate) fn get_tenant_conf(&self) -> TenantConfOpt {
4701 0 : self.tenant_conf.load().tenant_conf.clone()
4702 0 : }
4703 :
4704 : /// How much local storage would this tenant like to have? It can cope with
4705 : /// less than this (via eviction and on-demand downloads), but this function enables
4706 : /// the Tenant to advertise how much storage it would prefer to have to provide fast I/O
4707 : /// by keeping important things on local disk.
4708 : ///
4709 : /// This is a heuristic, not a guarantee: tenants that are long-idle will actually use less
4710 : /// than they report here, due to layer eviction. Tenants with many active branches may
4711 : /// actually use more than they report here.
4712 0 : pub(crate) fn local_storage_wanted(&self) -> u64 {
4713 0 : let timelines = self.timelines.lock().unwrap();
4714 0 :
4715 0 : // Heuristic: we use the max() of the timelines' visible sizes, rather than the sum. This
4716 0 : // reflects the observation that on tenants with multiple large branches, typically only one
4717 0 : // of them is used actively enough to occupy space on disk.
4718 0 : timelines
4719 0 : .values()
4720 0 : .map(|t| t.metrics.visible_physical_size_gauge.get())
4721 0 : .max()
4722 0 : .unwrap_or(0)
4723 0 : }
4724 :
4725 : /// Serialize and write the latest TenantManifest to remote storage.
4726 0 : pub(crate) async fn store_tenant_manifest(&self) -> Result<(), TenantManifestError> {
4727 : // Only one manifest write may be done at at time, and the contents of the manifest
4728 : // must be loaded while holding this lock. This makes it safe to call this function
4729 : // from anywhere without worrying about colliding updates.
4730 0 : let mut guard = tokio::select! {
4731 0 : g = self.tenant_manifest_upload.lock() => {
4732 0 : g
4733 : },
4734 0 : _ = self.cancel.cancelled() => {
4735 0 : return Err(TenantManifestError::Cancelled);
4736 : }
4737 : };
4738 :
4739 0 : let manifest = self.build_tenant_manifest();
4740 0 : if Some(&manifest) == (*guard).as_ref() {
4741 : // Optimisation: skip uploads that don't change anything.
4742 0 : return Ok(());
4743 0 : }
4744 0 :
4745 0 : upload_tenant_manifest(
4746 0 : &self.remote_storage,
4747 0 : &self.tenant_shard_id,
4748 0 : self.generation,
4749 0 : &manifest,
4750 0 : &self.cancel,
4751 0 : )
4752 0 : .await
4753 0 : .map_err(|e| {
4754 0 : if self.cancel.is_cancelled() {
4755 0 : TenantManifestError::Cancelled
4756 : } else {
4757 0 : TenantManifestError::RemoteStorage(e)
4758 : }
4759 0 : })?;
4760 :
4761 : // Store the successfully uploaded manifest, so that future callers can avoid
4762 : // re-uploading the same thing.
4763 0 : *guard = Some(manifest);
4764 0 :
4765 0 : Ok(())
4766 0 : }
4767 : }
4768 :
4769 : /// Create the cluster temporarily in 'initdbpath' directory inside the repository
4770 : /// to get bootstrap data for timeline initialization.
4771 0 : async fn run_initdb(
4772 0 : conf: &'static PageServerConf,
4773 0 : initdb_target_dir: &Utf8Path,
4774 0 : pg_version: u32,
4775 0 : cancel: &CancellationToken,
4776 0 : ) -> Result<(), InitdbError> {
4777 0 : let initdb_bin_path = conf
4778 0 : .pg_bin_dir(pg_version)
4779 0 : .map_err(InitdbError::Other)?
4780 0 : .join("initdb");
4781 0 : let initdb_lib_dir = conf.pg_lib_dir(pg_version).map_err(InitdbError::Other)?;
4782 0 : info!(
4783 0 : "running {} in {}, libdir: {}",
4784 : initdb_bin_path, initdb_target_dir, initdb_lib_dir,
4785 : );
4786 :
4787 0 : let _permit = INIT_DB_SEMAPHORE.acquire().await;
4788 :
4789 0 : let mut initdb_command = tokio::process::Command::new(&initdb_bin_path);
4790 0 : initdb_command
4791 0 : .args(["--pgdata", initdb_target_dir.as_ref()])
4792 0 : .args(["--username", &conf.superuser])
4793 0 : .args(["--encoding", "utf8"])
4794 0 : .args(["--locale", &conf.locale])
4795 0 : .arg("--no-instructions")
4796 0 : .arg("--no-sync")
4797 0 : .env_clear()
4798 0 : .env("LD_LIBRARY_PATH", &initdb_lib_dir)
4799 0 : .env("DYLD_LIBRARY_PATH", &initdb_lib_dir)
4800 0 : .stdin(std::process::Stdio::null())
4801 0 : // stdout invocation produces the same output every time, we don't need it
4802 0 : .stdout(std::process::Stdio::null())
4803 0 : // we would be interested in the stderr output, if there was any
4804 0 : .stderr(std::process::Stdio::piped());
4805 0 :
4806 0 : // Before version 14, only the libc provide was available.
4807 0 : if pg_version > 14 {
4808 : // Version 17 brought with it a builtin locale provider which only provides
4809 : // C and C.UTF-8. While being safer for collation purposes since it is
4810 : // guaranteed to be consistent throughout a major release, it is also more
4811 : // performant.
4812 0 : let locale_provider = if pg_version >= 17 { "builtin" } else { "libc" };
4813 :
4814 0 : initdb_command.args(["--locale-provider", locale_provider]);
4815 0 : }
4816 :
4817 0 : let initdb_proc = initdb_command.spawn()?;
4818 :
4819 : // Ideally we'd select here with the cancellation token, but the problem is that
4820 : // we can't safely terminate initdb: it launches processes of its own, and killing
4821 : // initdb doesn't kill them. After we return from this function, we want the target
4822 : // directory to be able to be cleaned up.
4823 : // See https://github.com/neondatabase/neon/issues/6385
4824 0 : let initdb_output = initdb_proc.wait_with_output().await?;
4825 0 : if !initdb_output.status.success() {
4826 0 : return Err(InitdbError::Failed(
4827 0 : initdb_output.status,
4828 0 : initdb_output.stderr,
4829 0 : ));
4830 0 : }
4831 0 :
4832 0 : // This isn't true cancellation support, see above. Still return an error to
4833 0 : // excercise the cancellation code path.
4834 0 : if cancel.is_cancelled() {
4835 0 : return Err(InitdbError::Cancelled);
4836 0 : }
4837 0 :
4838 0 : Ok(())
4839 0 : }
4840 :
4841 : /// Dump contents of a layer file to stdout.
4842 0 : pub async fn dump_layerfile_from_path(
4843 0 : path: &Utf8Path,
4844 0 : verbose: bool,
4845 0 : ctx: &RequestContext,
4846 0 : ) -> anyhow::Result<()> {
4847 : use std::os::unix::fs::FileExt;
4848 :
4849 : // All layer files start with a two-byte "magic" value, to identify the kind of
4850 : // file.
4851 0 : let file = File::open(path)?;
4852 0 : let mut header_buf = [0u8; 2];
4853 0 : file.read_exact_at(&mut header_buf, 0)?;
4854 :
4855 0 : match u16::from_be_bytes(header_buf) {
4856 : crate::IMAGE_FILE_MAGIC => {
4857 0 : ImageLayer::new_for_path(path, file)?
4858 0 : .dump(verbose, ctx)
4859 0 : .await?
4860 : }
4861 : crate::DELTA_FILE_MAGIC => {
4862 0 : DeltaLayer::new_for_path(path, file)?
4863 0 : .dump(verbose, ctx)
4864 0 : .await?
4865 : }
4866 0 : magic => bail!("unrecognized magic identifier: {:?}", magic),
4867 : }
4868 :
4869 0 : Ok(())
4870 0 : }
4871 :
4872 : #[cfg(test)]
4873 : pub(crate) mod harness {
4874 : use bytes::{Bytes, BytesMut};
4875 : use once_cell::sync::OnceCell;
4876 : use pageserver_api::models::ShardParameters;
4877 : use pageserver_api::shard::ShardIndex;
4878 : use utils::logging;
4879 :
4880 : use crate::deletion_queue::mock::MockDeletionQueue;
4881 : use crate::l0_flush::L0FlushConfig;
4882 : use crate::walredo::apply_neon;
4883 : use pageserver_api::key::Key;
4884 : use pageserver_api::record::NeonWalRecord;
4885 :
4886 : use super::*;
4887 : use hex_literal::hex;
4888 : use utils::id::TenantId;
4889 :
4890 : pub const TIMELINE_ID: TimelineId =
4891 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
4892 : pub const NEW_TIMELINE_ID: TimelineId =
4893 : TimelineId::from_array(hex!("AA223344556677881122334455667788"));
4894 :
4895 : /// Convenience function to create a page image with given string as the only content
4896 5028729 : pub fn test_img(s: &str) -> Bytes {
4897 5028729 : let mut buf = BytesMut::new();
4898 5028729 : buf.extend_from_slice(s.as_bytes());
4899 5028729 : buf.resize(64, 0);
4900 5028729 :
4901 5028729 : buf.freeze()
4902 5028729 : }
4903 :
4904 : impl From<TenantConf> for TenantConfOpt {
4905 190 : fn from(tenant_conf: TenantConf) -> Self {
4906 190 : Self {
4907 190 : checkpoint_distance: Some(tenant_conf.checkpoint_distance),
4908 190 : checkpoint_timeout: Some(tenant_conf.checkpoint_timeout),
4909 190 : compaction_target_size: Some(tenant_conf.compaction_target_size),
4910 190 : compaction_period: Some(tenant_conf.compaction_period),
4911 190 : compaction_threshold: Some(tenant_conf.compaction_threshold),
4912 190 : compaction_algorithm: Some(tenant_conf.compaction_algorithm),
4913 190 : gc_horizon: Some(tenant_conf.gc_horizon),
4914 190 : gc_period: Some(tenant_conf.gc_period),
4915 190 : image_creation_threshold: Some(tenant_conf.image_creation_threshold),
4916 190 : pitr_interval: Some(tenant_conf.pitr_interval),
4917 190 : walreceiver_connect_timeout: Some(tenant_conf.walreceiver_connect_timeout),
4918 190 : lagging_wal_timeout: Some(tenant_conf.lagging_wal_timeout),
4919 190 : max_lsn_wal_lag: Some(tenant_conf.max_lsn_wal_lag),
4920 190 : eviction_policy: Some(tenant_conf.eviction_policy),
4921 190 : min_resident_size_override: tenant_conf.min_resident_size_override,
4922 190 : evictions_low_residence_duration_metric_threshold: Some(
4923 190 : tenant_conf.evictions_low_residence_duration_metric_threshold,
4924 190 : ),
4925 190 : heatmap_period: Some(tenant_conf.heatmap_period),
4926 190 : lazy_slru_download: Some(tenant_conf.lazy_slru_download),
4927 190 : timeline_get_throttle: Some(tenant_conf.timeline_get_throttle),
4928 190 : image_layer_creation_check_threshold: Some(
4929 190 : tenant_conf.image_layer_creation_check_threshold,
4930 190 : ),
4931 190 : lsn_lease_length: Some(tenant_conf.lsn_lease_length),
4932 190 : lsn_lease_length_for_ts: Some(tenant_conf.lsn_lease_length_for_ts),
4933 190 : timeline_offloading: Some(tenant_conf.timeline_offloading),
4934 190 : }
4935 190 : }
4936 : }
4937 :
4938 : pub struct TenantHarness {
4939 : pub conf: &'static PageServerConf,
4940 : pub tenant_conf: TenantConf,
4941 : pub tenant_shard_id: TenantShardId,
4942 : pub generation: Generation,
4943 : pub shard: ShardIndex,
4944 : pub remote_storage: GenericRemoteStorage,
4945 : pub remote_fs_dir: Utf8PathBuf,
4946 : pub deletion_queue: MockDeletionQueue,
4947 : }
4948 :
4949 : static LOG_HANDLE: OnceCell<()> = OnceCell::new();
4950 :
4951 206 : pub(crate) fn setup_logging() {
4952 206 : LOG_HANDLE.get_or_init(|| {
4953 194 : logging::init(
4954 194 : logging::LogFormat::Test,
4955 194 : // enable it in case the tests exercise code paths that use
4956 194 : // debug_assert_current_span_has_tenant_and_timeline_id
4957 194 : logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
4958 194 : logging::Output::Stdout,
4959 194 : )
4960 194 : .expect("Failed to init test logging")
4961 206 : });
4962 206 : }
4963 :
4964 : impl TenantHarness {
4965 190 : pub async fn create_custom(
4966 190 : test_name: &'static str,
4967 190 : tenant_conf: TenantConf,
4968 190 : tenant_id: TenantId,
4969 190 : shard_identity: ShardIdentity,
4970 190 : generation: Generation,
4971 190 : ) -> anyhow::Result<Self> {
4972 190 : setup_logging();
4973 190 :
4974 190 : let repo_dir = PageServerConf::test_repo_dir(test_name);
4975 190 : let _ = fs::remove_dir_all(&repo_dir);
4976 190 : fs::create_dir_all(&repo_dir)?;
4977 :
4978 190 : let conf = PageServerConf::dummy_conf(repo_dir);
4979 190 : // Make a static copy of the config. This can never be free'd, but that's
4980 190 : // OK in a test.
4981 190 : let conf: &'static PageServerConf = Box::leak(Box::new(conf));
4982 190 :
4983 190 : let shard = shard_identity.shard_index();
4984 190 : let tenant_shard_id = TenantShardId {
4985 190 : tenant_id,
4986 190 : shard_number: shard.shard_number,
4987 190 : shard_count: shard.shard_count,
4988 190 : };
4989 190 : fs::create_dir_all(conf.tenant_path(&tenant_shard_id))?;
4990 190 : fs::create_dir_all(conf.timelines_path(&tenant_shard_id))?;
4991 :
4992 : use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
4993 190 : let remote_fs_dir = conf.workdir.join("localfs");
4994 190 : std::fs::create_dir_all(&remote_fs_dir).unwrap();
4995 190 : let config = RemoteStorageConfig {
4996 190 : storage: RemoteStorageKind::LocalFs {
4997 190 : local_path: remote_fs_dir.clone(),
4998 190 : },
4999 190 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
5000 190 : };
5001 190 : let remote_storage = GenericRemoteStorage::from_config(&config).await.unwrap();
5002 190 : let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()));
5003 190 :
5004 190 : Ok(Self {
5005 190 : conf,
5006 190 : tenant_conf,
5007 190 : tenant_shard_id,
5008 190 : generation,
5009 190 : shard,
5010 190 : remote_storage,
5011 190 : remote_fs_dir,
5012 190 : deletion_queue,
5013 190 : })
5014 190 : }
5015 :
5016 178 : pub async fn create(test_name: &'static str) -> anyhow::Result<Self> {
5017 178 : // Disable automatic GC and compaction to make the unit tests more deterministic.
5018 178 : // The tests perform them manually if needed.
5019 178 : let tenant_conf = TenantConf {
5020 178 : gc_period: Duration::ZERO,
5021 178 : compaction_period: Duration::ZERO,
5022 178 : ..TenantConf::default()
5023 178 : };
5024 178 : let tenant_id = TenantId::generate();
5025 178 : let shard = ShardIdentity::unsharded();
5026 178 : Self::create_custom(
5027 178 : test_name,
5028 178 : tenant_conf,
5029 178 : tenant_id,
5030 178 : shard,
5031 178 : Generation::new(0xdeadbeef),
5032 178 : )
5033 0 : .await
5034 178 : }
5035 :
5036 20 : pub fn span(&self) -> tracing::Span {
5037 20 : info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
5038 20 : }
5039 :
5040 190 : pub(crate) async fn load(&self) -> (Arc<Tenant>, RequestContext) {
5041 190 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
5042 190 : (
5043 190 : self.do_try_load(&ctx)
5044 1903 : .await
5045 190 : .expect("failed to load test tenant"),
5046 190 : ctx,
5047 190 : )
5048 190 : }
5049 :
5050 190 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5051 : pub(crate) async fn do_try_load(
5052 : &self,
5053 : ctx: &RequestContext,
5054 : ) -> anyhow::Result<Arc<Tenant>> {
5055 : let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
5056 :
5057 : let tenant = Arc::new(Tenant::new(
5058 : TenantState::Attaching,
5059 : self.conf,
5060 : AttachedTenantConf::try_from(LocationConf::attached_single(
5061 : TenantConfOpt::from(self.tenant_conf.clone()),
5062 : self.generation,
5063 : &ShardParameters::default(),
5064 : ))
5065 : .unwrap(),
5066 : // This is a legacy/test code path: sharding isn't supported here.
5067 : ShardIdentity::unsharded(),
5068 : Some(walredo_mgr),
5069 : self.tenant_shard_id,
5070 : self.remote_storage.clone(),
5071 : self.deletion_queue.new_client(),
5072 : // TODO: ideally we should run all unit tests with both configs
5073 : L0FlushGlobalState::new(L0FlushConfig::default()),
5074 : ));
5075 :
5076 : let preload = tenant
5077 : .preload(&self.remote_storage, CancellationToken::new())
5078 : .await?;
5079 : tenant.attach(Some(preload), ctx).await?;
5080 :
5081 : tenant.state.send_replace(TenantState::Active);
5082 : for timeline in tenant.timelines.lock().unwrap().values() {
5083 : timeline.set_state(TimelineState::Active);
5084 : }
5085 : Ok(tenant)
5086 : }
5087 :
5088 2 : pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
5089 2 : self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
5090 2 : }
5091 : }
5092 :
5093 : // Mock WAL redo manager that doesn't do much
5094 : pub(crate) struct TestRedoManager;
5095 :
5096 : impl TestRedoManager {
5097 : /// # Cancel-Safety
5098 : ///
5099 : /// This method is cancellation-safe.
5100 410 : pub async fn request_redo(
5101 410 : &self,
5102 410 : key: Key,
5103 410 : lsn: Lsn,
5104 410 : base_img: Option<(Lsn, Bytes)>,
5105 410 : records: Vec<(Lsn, NeonWalRecord)>,
5106 410 : _pg_version: u32,
5107 410 : ) -> Result<Bytes, walredo::Error> {
5108 570 : let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
5109 410 : if records_neon {
5110 : // For Neon wal records, we can decode without spawning postgres, so do so.
5111 410 : let mut page = match (base_img, records.first()) {
5112 344 : (Some((_lsn, img)), _) => {
5113 344 : let mut page = BytesMut::new();
5114 344 : page.extend_from_slice(&img);
5115 344 : page
5116 : }
5117 66 : (_, Some((_lsn, rec))) if rec.will_init() => BytesMut::new(),
5118 : _ => {
5119 0 : panic!("Neon WAL redo requires base image or will init record");
5120 : }
5121 : };
5122 :
5123 980 : for (record_lsn, record) in records {
5124 570 : apply_neon::apply_in_neon(&record, record_lsn, key, &mut page)?;
5125 : }
5126 410 : Ok(page.freeze())
5127 : } else {
5128 : // We never spawn a postgres walredo process in unit tests: just log what we might have done.
5129 0 : let s = format!(
5130 0 : "redo for {} to get to {}, with {} and {} records",
5131 0 : key,
5132 0 : lsn,
5133 0 : if base_img.is_some() {
5134 0 : "base image"
5135 : } else {
5136 0 : "no base image"
5137 : },
5138 0 : records.len()
5139 0 : );
5140 0 : println!("{s}");
5141 0 :
5142 0 : Ok(test_img(&s))
5143 : }
5144 410 : }
5145 : }
5146 : }
5147 :
5148 : #[cfg(test)]
5149 : mod tests {
5150 : use std::collections::{BTreeMap, BTreeSet};
5151 :
5152 : use super::*;
5153 : use crate::keyspace::KeySpaceAccum;
5154 : use crate::tenant::harness::*;
5155 : use crate::tenant::timeline::CompactFlags;
5156 : use crate::DEFAULT_PG_VERSION;
5157 : use bytes::{Bytes, BytesMut};
5158 : use hex_literal::hex;
5159 : use itertools::Itertools;
5160 : use pageserver_api::key::{Key, AUX_KEY_PREFIX, NON_INHERITED_RANGE};
5161 : use pageserver_api::keyspace::KeySpace;
5162 : use pageserver_api::models::{CompactionAlgorithm, CompactionAlgorithmSettings};
5163 : use pageserver_api::value::Value;
5164 : use pageserver_compaction::helpers::overlaps_with;
5165 : use rand::{thread_rng, Rng};
5166 : use storage_layer::PersistentLayerKey;
5167 : use tests::storage_layer::ValuesReconstructState;
5168 : use tests::timeline::{GetVectoredError, ShutdownMode};
5169 : use timeline::DeltaLayerTestDesc;
5170 : use utils::id::TenantId;
5171 :
5172 : #[cfg(feature = "testing")]
5173 : use pageserver_api::record::NeonWalRecord;
5174 : #[cfg(feature = "testing")]
5175 : use timeline::compaction::{KeyHistoryRetention, KeyLogAtLsn};
5176 : #[cfg(feature = "testing")]
5177 : use timeline::GcInfo;
5178 :
5179 : static TEST_KEY: Lazy<Key> =
5180 18 : Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));
5181 :
5182 : #[tokio::test]
5183 2 : async fn test_basic() -> anyhow::Result<()> {
5184 20 : let (tenant, ctx) = TenantHarness::create("test_basic").await?.load().await;
5185 2 : let tline = tenant
5186 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5187 6 : .await?;
5188 2 :
5189 2 : let mut writer = tline.writer().await;
5190 2 : writer
5191 2 : .put(
5192 2 : *TEST_KEY,
5193 2 : Lsn(0x10),
5194 2 : &Value::Image(test_img("foo at 0x10")),
5195 2 : &ctx,
5196 2 : )
5197 2 : .await?;
5198 2 : writer.finish_write(Lsn(0x10));
5199 2 : drop(writer);
5200 2 :
5201 2 : let mut writer = tline.writer().await;
5202 2 : writer
5203 2 : .put(
5204 2 : *TEST_KEY,
5205 2 : Lsn(0x20),
5206 2 : &Value::Image(test_img("foo at 0x20")),
5207 2 : &ctx,
5208 2 : )
5209 2 : .await?;
5210 2 : writer.finish_write(Lsn(0x20));
5211 2 : drop(writer);
5212 2 :
5213 2 : assert_eq!(
5214 2 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
5215 2 : test_img("foo at 0x10")
5216 2 : );
5217 2 : assert_eq!(
5218 2 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
5219 2 : test_img("foo at 0x10")
5220 2 : );
5221 2 : assert_eq!(
5222 2 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
5223 2 : test_img("foo at 0x20")
5224 2 : );
5225 2 :
5226 2 : Ok(())
5227 2 : }
5228 :
5229 : #[tokio::test]
5230 2 : async fn no_duplicate_timelines() -> anyhow::Result<()> {
5231 2 : let (tenant, ctx) = TenantHarness::create("no_duplicate_timelines")
5232 2 : .await?
5233 2 : .load()
5234 20 : .await;
5235 2 : let _ = tenant
5236 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5237 5 : .await?;
5238 2 :
5239 2 : match tenant
5240 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5241 2 : .await
5242 2 : {
5243 2 : Ok(_) => panic!("duplicate timeline creation should fail"),
5244 2 : Err(e) => assert_eq!(
5245 2 : e.to_string(),
5246 2 : "timeline already exists with different parameters".to_string()
5247 2 : ),
5248 2 : }
5249 2 :
5250 2 : Ok(())
5251 2 : }
5252 :
5253 : /// Convenience function to create a page image with given string as the only content
5254 10 : pub fn test_value(s: &str) -> Value {
5255 10 : let mut buf = BytesMut::new();
5256 10 : buf.extend_from_slice(s.as_bytes());
5257 10 : Value::Image(buf.freeze())
5258 10 : }
5259 :
5260 : ///
5261 : /// Test branch creation
5262 : ///
5263 : #[tokio::test]
5264 2 : async fn test_branch() -> anyhow::Result<()> {
5265 2 : use std::str::from_utf8;
5266 2 :
5267 20 : let (tenant, ctx) = TenantHarness::create("test_branch").await?.load().await;
5268 2 : let tline = tenant
5269 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5270 6 : .await?;
5271 2 : let mut writer = tline.writer().await;
5272 2 :
5273 2 : #[allow(non_snake_case)]
5274 2 : let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
5275 2 : #[allow(non_snake_case)]
5276 2 : let TEST_KEY_B: Key = Key::from_hex("110000000033333333444444445500000002").unwrap();
5277 2 :
5278 2 : // Insert a value on the timeline
5279 2 : writer
5280 2 : .put(TEST_KEY_A, Lsn(0x20), &test_value("foo at 0x20"), &ctx)
5281 2 : .await?;
5282 2 : writer
5283 2 : .put(TEST_KEY_B, Lsn(0x20), &test_value("foobar at 0x20"), &ctx)
5284 2 : .await?;
5285 2 : writer.finish_write(Lsn(0x20));
5286 2 :
5287 2 : writer
5288 2 : .put(TEST_KEY_A, Lsn(0x30), &test_value("foo at 0x30"), &ctx)
5289 2 : .await?;
5290 2 : writer.finish_write(Lsn(0x30));
5291 2 : writer
5292 2 : .put(TEST_KEY_A, Lsn(0x40), &test_value("foo at 0x40"), &ctx)
5293 2 : .await?;
5294 2 : writer.finish_write(Lsn(0x40));
5295 2 :
5296 2 : //assert_current_logical_size(&tline, Lsn(0x40));
5297 2 :
5298 2 : // Branch the history, modify relation differently on the new timeline
5299 2 : tenant
5300 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x30)), &ctx)
5301 2 : .await?;
5302 2 : let newtline = tenant
5303 2 : .get_timeline(NEW_TIMELINE_ID, true)
5304 2 : .expect("Should have a local timeline");
5305 2 : let mut new_writer = newtline.writer().await;
5306 2 : new_writer
5307 2 : .put(TEST_KEY_A, Lsn(0x40), &test_value("bar at 0x40"), &ctx)
5308 2 : .await?;
5309 2 : new_writer.finish_write(Lsn(0x40));
5310 2 :
5311 2 : // Check page contents on both branches
5312 2 : assert_eq!(
5313 2 : from_utf8(&tline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
5314 2 : "foo at 0x40"
5315 2 : );
5316 2 : assert_eq!(
5317 2 : from_utf8(&newtline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
5318 2 : "bar at 0x40"
5319 2 : );
5320 2 : assert_eq!(
5321 2 : from_utf8(&newtline.get(TEST_KEY_B, Lsn(0x40), &ctx).await?)?,
5322 2 : "foobar at 0x20"
5323 2 : );
5324 2 :
5325 2 : //assert_current_logical_size(&tline, Lsn(0x40));
5326 2 :
5327 2 : Ok(())
5328 2 : }
5329 :
5330 20 : async fn make_some_layers(
5331 20 : tline: &Timeline,
5332 20 : start_lsn: Lsn,
5333 20 : ctx: &RequestContext,
5334 20 : ) -> anyhow::Result<()> {
5335 20 : let mut lsn = start_lsn;
5336 : {
5337 20 : let mut writer = tline.writer().await;
5338 : // Create a relation on the timeline
5339 20 : writer
5340 20 : .put(
5341 20 : *TEST_KEY,
5342 20 : lsn,
5343 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5344 20 : ctx,
5345 20 : )
5346 10 : .await?;
5347 20 : writer.finish_write(lsn);
5348 20 : lsn += 0x10;
5349 20 : writer
5350 20 : .put(
5351 20 : *TEST_KEY,
5352 20 : lsn,
5353 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5354 20 : ctx,
5355 20 : )
5356 0 : .await?;
5357 20 : writer.finish_write(lsn);
5358 20 : lsn += 0x10;
5359 20 : }
5360 20 : tline.freeze_and_flush().await?;
5361 : {
5362 20 : let mut writer = tline.writer().await;
5363 20 : writer
5364 20 : .put(
5365 20 : *TEST_KEY,
5366 20 : lsn,
5367 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5368 20 : ctx,
5369 20 : )
5370 10 : .await?;
5371 20 : writer.finish_write(lsn);
5372 20 : lsn += 0x10;
5373 20 : writer
5374 20 : .put(
5375 20 : *TEST_KEY,
5376 20 : lsn,
5377 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5378 20 : ctx,
5379 20 : )
5380 0 : .await?;
5381 20 : writer.finish_write(lsn);
5382 20 : }
5383 21 : tline.freeze_and_flush().await.map_err(|e| e.into())
5384 20 : }
5385 :
5386 : #[tokio::test(start_paused = true)]
5387 2 : async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
5388 2 : let (tenant, ctx) =
5389 2 : TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
5390 2 : .await?
5391 2 : .load()
5392 20 : .await;
5393 2 : // Advance to the lsn lease deadline so that GC is not blocked by
5394 2 : // initial transition into AttachedSingle.
5395 2 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
5396 2 : tokio::time::resume();
5397 2 : let tline = tenant
5398 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5399 6 : .await?;
5400 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5401 2 :
5402 2 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
5403 2 : // FIXME: this doesn't actually remove any layer currently, given how the flushing
5404 2 : // and compaction works. But it does set the 'cutoff' point so that the cross check
5405 2 : // below should fail.
5406 2 : tenant
5407 2 : .gc_iteration(
5408 2 : Some(TIMELINE_ID),
5409 2 : 0x10,
5410 2 : Duration::ZERO,
5411 2 : &CancellationToken::new(),
5412 2 : &ctx,
5413 2 : )
5414 2 : .await?;
5415 2 :
5416 2 : // try to branch at lsn 25, should fail because we already garbage collected the data
5417 2 : match tenant
5418 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
5419 2 : .await
5420 2 : {
5421 2 : Ok(_) => panic!("branching should have failed"),
5422 2 : Err(err) => {
5423 2 : let CreateTimelineError::AncestorLsn(err) = err else {
5424 2 : panic!("wrong error type")
5425 2 : };
5426 2 : assert!(err.to_string().contains("invalid branch start lsn"));
5427 2 : assert!(err
5428 2 : .source()
5429 2 : .unwrap()
5430 2 : .to_string()
5431 2 : .contains("we might've already garbage collected needed data"))
5432 2 : }
5433 2 : }
5434 2 :
5435 2 : Ok(())
5436 2 : }
5437 :
5438 : #[tokio::test]
5439 2 : async fn test_prohibit_branch_creation_on_pre_initdb_lsn() -> anyhow::Result<()> {
5440 2 : let (tenant, ctx) =
5441 2 : TenantHarness::create("test_prohibit_branch_creation_on_pre_initdb_lsn")
5442 2 : .await?
5443 2 : .load()
5444 20 : .await;
5445 2 :
5446 2 : let tline = tenant
5447 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x50), DEFAULT_PG_VERSION, &ctx)
5448 6 : .await?;
5449 2 : // try to branch at lsn 0x25, should fail because initdb lsn is 0x50
5450 2 : match tenant
5451 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
5452 2 : .await
5453 2 : {
5454 2 : Ok(_) => panic!("branching should have failed"),
5455 2 : Err(err) => {
5456 2 : let CreateTimelineError::AncestorLsn(err) = err else {
5457 2 : panic!("wrong error type");
5458 2 : };
5459 2 : assert!(&err.to_string().contains("invalid branch start lsn"));
5460 2 : assert!(&err
5461 2 : .source()
5462 2 : .unwrap()
5463 2 : .to_string()
5464 2 : .contains("is earlier than latest GC cutoff"));
5465 2 : }
5466 2 : }
5467 2 :
5468 2 : Ok(())
5469 2 : }
5470 :
5471 : /*
5472 : // FIXME: This currently fails to error out. Calling GC doesn't currently
5473 : // remove the old value, we'd need to work a little harder
5474 : #[tokio::test]
5475 : async fn test_prohibit_get_for_garbage_collected_data() -> anyhow::Result<()> {
5476 : let repo =
5477 : RepoHarness::create("test_prohibit_get_for_garbage_collected_data")?
5478 : .load();
5479 :
5480 : let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION)?;
5481 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5482 :
5483 : repo.gc_iteration(Some(TIMELINE_ID), 0x10, Duration::ZERO)?;
5484 : let latest_gc_cutoff_lsn = tline.get_latest_gc_cutoff_lsn();
5485 : assert!(*latest_gc_cutoff_lsn > Lsn(0x25));
5486 : match tline.get(*TEST_KEY, Lsn(0x25)) {
5487 : Ok(_) => panic!("request for page should have failed"),
5488 : Err(err) => assert!(err.to_string().contains("not found at")),
5489 : }
5490 : Ok(())
5491 : }
5492 : */
5493 :
5494 : #[tokio::test]
5495 2 : async fn test_get_branchpoints_from_an_inactive_timeline() -> anyhow::Result<()> {
5496 2 : let (tenant, ctx) =
5497 2 : TenantHarness::create("test_get_branchpoints_from_an_inactive_timeline")
5498 2 : .await?
5499 2 : .load()
5500 20 : .await;
5501 2 : let tline = tenant
5502 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5503 6 : .await?;
5504 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5505 2 :
5506 2 : tenant
5507 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
5508 2 : .await?;
5509 2 : let newtline = tenant
5510 2 : .get_timeline(NEW_TIMELINE_ID, true)
5511 2 : .expect("Should have a local timeline");
5512 2 :
5513 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
5514 2 :
5515 2 : tline.set_broken("test".to_owned());
5516 2 :
5517 2 : tenant
5518 2 : .gc_iteration(
5519 2 : Some(TIMELINE_ID),
5520 2 : 0x10,
5521 2 : Duration::ZERO,
5522 2 : &CancellationToken::new(),
5523 2 : &ctx,
5524 2 : )
5525 2 : .await?;
5526 2 :
5527 2 : // The branchpoints should contain all timelines, even ones marked
5528 2 : // as Broken.
5529 2 : {
5530 2 : let branchpoints = &tline.gc_info.read().unwrap().retain_lsns;
5531 2 : assert_eq!(branchpoints.len(), 1);
5532 2 : assert_eq!(
5533 2 : branchpoints[0],
5534 2 : (Lsn(0x40), NEW_TIMELINE_ID, MaybeOffloaded::No)
5535 2 : );
5536 2 : }
5537 2 :
5538 2 : // You can read the key from the child branch even though the parent is
5539 2 : // Broken, as long as you don't need to access data from the parent.
5540 2 : assert_eq!(
5541 4 : newtline.get(*TEST_KEY, Lsn(0x70), &ctx).await?,
5542 2 : test_img(&format!("foo at {}", Lsn(0x70)))
5543 2 : );
5544 2 :
5545 2 : // This needs to traverse to the parent, and fails.
5546 2 : let err = newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await.unwrap_err();
5547 2 : assert!(
5548 2 : err.to_string().starts_with(&format!(
5549 2 : "bad state on timeline {}: Broken",
5550 2 : tline.timeline_id
5551 2 : )),
5552 2 : "{err}"
5553 2 : );
5554 2 :
5555 2 : Ok(())
5556 2 : }
5557 :
5558 : #[tokio::test]
5559 2 : async fn test_retain_data_in_parent_which_is_needed_for_child() -> anyhow::Result<()> {
5560 2 : let (tenant, ctx) =
5561 2 : TenantHarness::create("test_retain_data_in_parent_which_is_needed_for_child")
5562 2 : .await?
5563 2 : .load()
5564 18 : .await;
5565 2 : let tline = tenant
5566 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5567 5 : .await?;
5568 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5569 2 :
5570 2 : tenant
5571 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
5572 2 : .await?;
5573 2 : let newtline = tenant
5574 2 : .get_timeline(NEW_TIMELINE_ID, true)
5575 2 : .expect("Should have a local timeline");
5576 2 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
5577 2 : tenant
5578 2 : .gc_iteration(
5579 2 : Some(TIMELINE_ID),
5580 2 : 0x10,
5581 2 : Duration::ZERO,
5582 2 : &CancellationToken::new(),
5583 2 : &ctx,
5584 2 : )
5585 2 : .await?;
5586 4 : assert!(newtline.get(*TEST_KEY, Lsn(0x25), &ctx).await.is_ok());
5587 2 :
5588 2 : Ok(())
5589 2 : }
5590 : #[tokio::test]
5591 2 : async fn test_parent_keeps_data_forever_after_branching() -> anyhow::Result<()> {
5592 2 : let (tenant, ctx) = TenantHarness::create("test_parent_keeps_data_forever_after_branching")
5593 2 : .await?
5594 2 : .load()
5595 19 : .await;
5596 2 : let tline = tenant
5597 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5598 6 : .await?;
5599 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5600 2 :
5601 2 : tenant
5602 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
5603 2 : .await?;
5604 2 : let newtline = tenant
5605 2 : .get_timeline(NEW_TIMELINE_ID, true)
5606 2 : .expect("Should have a local timeline");
5607 2 :
5608 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
5609 2 :
5610 2 : // run gc on parent
5611 2 : tenant
5612 2 : .gc_iteration(
5613 2 : Some(TIMELINE_ID),
5614 2 : 0x10,
5615 2 : Duration::ZERO,
5616 2 : &CancellationToken::new(),
5617 2 : &ctx,
5618 2 : )
5619 2 : .await?;
5620 2 :
5621 2 : // Check that the data is still accessible on the branch.
5622 2 : assert_eq!(
5623 7 : newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
5624 2 : test_img(&format!("foo at {}", Lsn(0x40)))
5625 2 : );
5626 2 :
5627 2 : Ok(())
5628 2 : }
5629 :
5630 : #[tokio::test]
5631 2 : async fn timeline_load() -> anyhow::Result<()> {
5632 2 : const TEST_NAME: &str = "timeline_load";
5633 2 : let harness = TenantHarness::create(TEST_NAME).await?;
5634 2 : {
5635 20 : let (tenant, ctx) = harness.load().await;
5636 2 : let tline = tenant
5637 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x7000), DEFAULT_PG_VERSION, &ctx)
5638 6 : .await?;
5639 7 : make_some_layers(tline.as_ref(), Lsn(0x8000), &ctx).await?;
5640 2 : // so that all uploads finish & we can call harness.load() below again
5641 2 : tenant
5642 2 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
5643 2 : .instrument(harness.span())
5644 2 : .await
5645 2 : .ok()
5646 2 : .unwrap();
5647 2 : }
5648 2 :
5649 12 : let (tenant, _ctx) = harness.load().await;
5650 2 : tenant
5651 2 : .get_timeline(TIMELINE_ID, true)
5652 2 : .expect("cannot load timeline");
5653 2 :
5654 2 : Ok(())
5655 2 : }
5656 :
5657 : #[tokio::test]
5658 2 : async fn timeline_load_with_ancestor() -> anyhow::Result<()> {
5659 2 : const TEST_NAME: &str = "timeline_load_with_ancestor";
5660 2 : let harness = TenantHarness::create(TEST_NAME).await?;
5661 2 : // create two timelines
5662 2 : {
5663 20 : let (tenant, ctx) = harness.load().await;
5664 2 : let tline = tenant
5665 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5666 6 : .await?;
5667 2 :
5668 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5669 2 :
5670 2 : let child_tline = tenant
5671 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
5672 2 : .await?;
5673 2 : child_tline.set_state(TimelineState::Active);
5674 2 :
5675 2 : let newtline = tenant
5676 2 : .get_timeline(NEW_TIMELINE_ID, true)
5677 2 : .expect("Should have a local timeline");
5678 2 :
5679 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
5680 2 :
5681 2 : // so that all uploads finish & we can call harness.load() below again
5682 2 : tenant
5683 2 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
5684 2 : .instrument(harness.span())
5685 2 : .await
5686 2 : .ok()
5687 2 : .unwrap();
5688 2 : }
5689 2 :
5690 2 : // check that both of them are initially unloaded
5691 34 : let (tenant, _ctx) = harness.load().await;
5692 2 :
5693 2 : // check that both, child and ancestor are loaded
5694 2 : let _child_tline = tenant
5695 2 : .get_timeline(NEW_TIMELINE_ID, true)
5696 2 : .expect("cannot get child timeline loaded");
5697 2 :
5698 2 : let _ancestor_tline = tenant
5699 2 : .get_timeline(TIMELINE_ID, true)
5700 2 : .expect("cannot get ancestor timeline loaded");
5701 2 :
5702 2 : Ok(())
5703 2 : }
5704 :
5705 : #[tokio::test]
5706 2 : async fn delta_layer_dumping() -> anyhow::Result<()> {
5707 2 : use storage_layer::AsLayerDesc;
5708 2 : let (tenant, ctx) = TenantHarness::create("test_layer_dumping")
5709 2 : .await?
5710 2 : .load()
5711 20 : .await;
5712 2 : let tline = tenant
5713 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5714 6 : .await?;
5715 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5716 2 :
5717 2 : let layer_map = tline.layers.read().await;
5718 2 : let level0_deltas = layer_map
5719 2 : .layer_map()?
5720 2 : .level0_deltas()
5721 2 : .iter()
5722 4 : .map(|desc| layer_map.get_from_desc(desc))
5723 2 : .collect::<Vec<_>>();
5724 2 :
5725 2 : assert!(!level0_deltas.is_empty());
5726 2 :
5727 6 : for delta in level0_deltas {
5728 2 : // Ensure we are dumping a delta layer here
5729 4 : assert!(delta.layer_desc().is_delta);
5730 8 : delta.dump(true, &ctx).await.unwrap();
5731 2 : }
5732 2 :
5733 2 : Ok(())
5734 2 : }
5735 :
5736 : #[tokio::test]
5737 2 : async fn test_images() -> anyhow::Result<()> {
5738 20 : let (tenant, ctx) = TenantHarness::create("test_images").await?.load().await;
5739 2 : let tline = tenant
5740 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5741 6 : .await?;
5742 2 :
5743 2 : let mut writer = tline.writer().await;
5744 2 : writer
5745 2 : .put(
5746 2 : *TEST_KEY,
5747 2 : Lsn(0x10),
5748 2 : &Value::Image(test_img("foo at 0x10")),
5749 2 : &ctx,
5750 2 : )
5751 2 : .await?;
5752 2 : writer.finish_write(Lsn(0x10));
5753 2 : drop(writer);
5754 2 :
5755 2 : tline.freeze_and_flush().await?;
5756 2 : tline
5757 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
5758 2 : .await?;
5759 2 :
5760 2 : let mut writer = tline.writer().await;
5761 2 : writer
5762 2 : .put(
5763 2 : *TEST_KEY,
5764 2 : Lsn(0x20),
5765 2 : &Value::Image(test_img("foo at 0x20")),
5766 2 : &ctx,
5767 2 : )
5768 2 : .await?;
5769 2 : writer.finish_write(Lsn(0x20));
5770 2 : drop(writer);
5771 2 :
5772 2 : tline.freeze_and_flush().await?;
5773 2 : tline
5774 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
5775 2 : .await?;
5776 2 :
5777 2 : let mut writer = tline.writer().await;
5778 2 : writer
5779 2 : .put(
5780 2 : *TEST_KEY,
5781 2 : Lsn(0x30),
5782 2 : &Value::Image(test_img("foo at 0x30")),
5783 2 : &ctx,
5784 2 : )
5785 2 : .await?;
5786 2 : writer.finish_write(Lsn(0x30));
5787 2 : drop(writer);
5788 2 :
5789 2 : tline.freeze_and_flush().await?;
5790 2 : tline
5791 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
5792 2 : .await?;
5793 2 :
5794 2 : let mut writer = tline.writer().await;
5795 2 : writer
5796 2 : .put(
5797 2 : *TEST_KEY,
5798 2 : Lsn(0x40),
5799 2 : &Value::Image(test_img("foo at 0x40")),
5800 2 : &ctx,
5801 2 : )
5802 2 : .await?;
5803 2 : writer.finish_write(Lsn(0x40));
5804 2 : drop(writer);
5805 2 :
5806 2 : tline.freeze_and_flush().await?;
5807 2 : tline
5808 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
5809 2 : .await?;
5810 2 :
5811 2 : assert_eq!(
5812 4 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
5813 2 : test_img("foo at 0x10")
5814 2 : );
5815 2 : assert_eq!(
5816 4 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
5817 2 : test_img("foo at 0x10")
5818 2 : );
5819 2 : assert_eq!(
5820 2 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
5821 2 : test_img("foo at 0x20")
5822 2 : );
5823 2 : assert_eq!(
5824 4 : tline.get(*TEST_KEY, Lsn(0x30), &ctx).await?,
5825 2 : test_img("foo at 0x30")
5826 2 : );
5827 2 : assert_eq!(
5828 4 : tline.get(*TEST_KEY, Lsn(0x40), &ctx).await?,
5829 2 : test_img("foo at 0x40")
5830 2 : );
5831 2 :
5832 2 : Ok(())
5833 2 : }
5834 :
5835 4 : async fn bulk_insert_compact_gc(
5836 4 : tenant: &Tenant,
5837 4 : timeline: &Arc<Timeline>,
5838 4 : ctx: &RequestContext,
5839 4 : lsn: Lsn,
5840 4 : repeat: usize,
5841 4 : key_count: usize,
5842 4 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
5843 4 : let compact = true;
5844 40718 : bulk_insert_maybe_compact_gc(tenant, timeline, ctx, lsn, repeat, key_count, compact).await
5845 4 : }
5846 :
5847 8 : async fn bulk_insert_maybe_compact_gc(
5848 8 : tenant: &Tenant,
5849 8 : timeline: &Arc<Timeline>,
5850 8 : ctx: &RequestContext,
5851 8 : mut lsn: Lsn,
5852 8 : repeat: usize,
5853 8 : key_count: usize,
5854 8 : compact: bool,
5855 8 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
5856 8 : let mut inserted: HashMap<Key, BTreeSet<Lsn>> = Default::default();
5857 8 :
5858 8 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5859 8 : let mut blknum = 0;
5860 8 :
5861 8 : // Enforce that key range is monotonously increasing
5862 8 : let mut keyspace = KeySpaceAccum::new();
5863 8 :
5864 8 : let cancel = CancellationToken::new();
5865 8 :
5866 8 : for _ in 0..repeat {
5867 400 : for _ in 0..key_count {
5868 4000000 : test_key.field6 = blknum;
5869 4000000 : let mut writer = timeline.writer().await;
5870 4000000 : writer
5871 4000000 : .put(
5872 4000000 : test_key,
5873 4000000 : lsn,
5874 4000000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5875 4000000 : ctx,
5876 4000000 : )
5877 3426 : .await?;
5878 4000000 : inserted.entry(test_key).or_default().insert(lsn);
5879 4000000 : writer.finish_write(lsn);
5880 4000000 : drop(writer);
5881 4000000 :
5882 4000000 : keyspace.add_key(test_key);
5883 4000000 :
5884 4000000 : lsn = Lsn(lsn.0 + 0x10);
5885 4000000 : blknum += 1;
5886 : }
5887 :
5888 400 : timeline.freeze_and_flush().await?;
5889 400 : if compact {
5890 : // this requires timeline to be &Arc<Timeline>
5891 8618 : timeline.compact(&cancel, EnumSet::empty(), ctx).await?;
5892 200 : }
5893 :
5894 : // this doesn't really need to use the timeline_id target, but it is closer to what it
5895 : // originally was.
5896 400 : let res = tenant
5897 400 : .gc_iteration(Some(timeline.timeline_id), 0, Duration::ZERO, &cancel, ctx)
5898 0 : .await?;
5899 :
5900 400 : assert_eq!(res.layers_removed, 0, "this never removes anything");
5901 : }
5902 :
5903 8 : Ok(inserted)
5904 8 : }
5905 :
5906 : //
5907 : // Insert 1000 key-value pairs with increasing keys, flush, compact, GC.
5908 : // Repeat 50 times.
5909 : //
5910 : #[tokio::test]
5911 2 : async fn test_bulk_insert() -> anyhow::Result<()> {
5912 2 : let harness = TenantHarness::create("test_bulk_insert").await?;
5913 20 : let (tenant, ctx) = harness.load().await;
5914 2 : let tline = tenant
5915 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5916 5 : .await?;
5917 2 :
5918 2 : let lsn = Lsn(0x10);
5919 20359 : bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
5920 2 :
5921 2 : Ok(())
5922 2 : }
5923 :
5924 : // Test the vectored get real implementation against a simple sequential implementation.
5925 : //
5926 : // The test generates a keyspace by repeatedly flushing the in-memory layer and compacting.
5927 : // Projected to 2D the key space looks like below. Lsn grows upwards on the Y axis and keys
5928 : // grow to the right on the X axis.
5929 : // [Delta]
5930 : // [Delta]
5931 : // [Delta]
5932 : // [Delta]
5933 : // ------------ Image ---------------
5934 : //
5935 : // After layer generation we pick the ranges to query as follows:
5936 : // 1. The beginning of each delta layer
5937 : // 2. At the seam between two adjacent delta layers
5938 : //
5939 : // There's one major downside to this test: delta layers only contains images,
5940 : // so the search can stop at the first delta layer and doesn't traverse any deeper.
5941 : #[tokio::test]
5942 2 : async fn test_get_vectored() -> anyhow::Result<()> {
5943 2 : let harness = TenantHarness::create("test_get_vectored").await?;
5944 20 : let (tenant, ctx) = harness.load().await;
5945 2 : let tline = tenant
5946 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5947 6 : .await?;
5948 2 :
5949 2 : let lsn = Lsn(0x10);
5950 20359 : let inserted = bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
5951 2 :
5952 2 : let guard = tline.layers.read().await;
5953 2 : let lm = guard.layer_map()?;
5954 2 :
5955 2 : lm.dump(true, &ctx).await?;
5956 2 :
5957 2 : let mut reads = Vec::new();
5958 2 : let mut prev = None;
5959 12 : lm.iter_historic_layers().for_each(|desc| {
5960 12 : if !desc.is_delta() {
5961 2 : prev = Some(desc.clone());
5962 2 : return;
5963 10 : }
5964 10 :
5965 10 : let start = desc.key_range.start;
5966 10 : let end = desc
5967 10 : .key_range
5968 10 : .start
5969 10 : .add(Timeline::MAX_GET_VECTORED_KEYS.try_into().unwrap());
5970 10 : reads.push(KeySpace {
5971 10 : ranges: vec![start..end],
5972 10 : });
5973 2 :
5974 10 : if let Some(prev) = &prev {
5975 10 : if !prev.is_delta() {
5976 10 : return;
5977 2 : }
5978 0 :
5979 0 : let first_range = Key {
5980 0 : field6: prev.key_range.end.field6 - 4,
5981 0 : ..prev.key_range.end
5982 0 : }..prev.key_range.end;
5983 0 :
5984 0 : let second_range = desc.key_range.start..Key {
5985 0 : field6: desc.key_range.start.field6 + 4,
5986 0 : ..desc.key_range.start
5987 0 : };
5988 0 :
5989 0 : reads.push(KeySpace {
5990 0 : ranges: vec![first_range, second_range],
5991 0 : });
5992 2 : };
5993 2 :
5994 2 : prev = Some(desc.clone());
5995 12 : });
5996 2 :
5997 2 : drop(guard);
5998 2 :
5999 2 : // Pick a big LSN such that we query over all the changes.
6000 2 : let reads_lsn = Lsn(u64::MAX - 1);
6001 2 :
6002 12 : for read in reads {
6003 10 : info!("Doing vectored read on {:?}", read);
6004 2 :
6005 10 : let vectored_res = tline
6006 10 : .get_vectored_impl(
6007 10 : read.clone(),
6008 10 : reads_lsn,
6009 10 : &mut ValuesReconstructState::new(),
6010 10 : &ctx,
6011 10 : )
6012 25 : .await;
6013 2 :
6014 10 : let mut expected_lsns: HashMap<Key, Lsn> = Default::default();
6015 10 : let mut expect_missing = false;
6016 10 : let mut key = read.start().unwrap();
6017 330 : while key != read.end().unwrap() {
6018 320 : if let Some(lsns) = inserted.get(&key) {
6019 320 : let expected_lsn = lsns.iter().rfind(|lsn| **lsn <= reads_lsn);
6020 320 : match expected_lsn {
6021 320 : Some(lsn) => {
6022 320 : expected_lsns.insert(key, *lsn);
6023 320 : }
6024 2 : None => {
6025 2 : expect_missing = true;
6026 0 : break;
6027 2 : }
6028 2 : }
6029 2 : } else {
6030 2 : expect_missing = true;
6031 0 : break;
6032 2 : }
6033 2 :
6034 320 : key = key.next();
6035 2 : }
6036 2 :
6037 10 : if expect_missing {
6038 2 : assert!(matches!(vectored_res, Err(GetVectoredError::MissingKey(_))));
6039 2 : } else {
6040 320 : for (key, image) in vectored_res? {
6041 320 : let expected_lsn = expected_lsns.get(&key).expect("determined above");
6042 320 : let expected_image = test_img(&format!("{} at {}", key.field6, expected_lsn));
6043 320 : assert_eq!(image?, expected_image);
6044 2 : }
6045 2 : }
6046 2 : }
6047 2 :
6048 2 : Ok(())
6049 2 : }
6050 :
6051 : #[tokio::test]
6052 2 : async fn test_get_vectored_aux_files() -> anyhow::Result<()> {
6053 2 : let harness = TenantHarness::create("test_get_vectored_aux_files").await?;
6054 2 :
6055 20 : let (tenant, ctx) = harness.load().await;
6056 2 : let tline = tenant
6057 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
6058 2 : .await?;
6059 2 : let tline = tline.raw_timeline().unwrap();
6060 2 :
6061 2 : let mut modification = tline.begin_modification(Lsn(0x1000));
6062 2 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
6063 2 : modification.set_lsn(Lsn(0x1008))?;
6064 2 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
6065 2 : modification.commit(&ctx).await?;
6066 2 :
6067 2 : let child_timeline_id = TimelineId::generate();
6068 2 : tenant
6069 2 : .branch_timeline_test(
6070 2 : tline,
6071 2 : child_timeline_id,
6072 2 : Some(tline.get_last_record_lsn()),
6073 2 : &ctx,
6074 2 : )
6075 2 : .await?;
6076 2 :
6077 2 : let child_timeline = tenant
6078 2 : .get_timeline(child_timeline_id, true)
6079 2 : .expect("Should have the branched timeline");
6080 2 :
6081 2 : let aux_keyspace = KeySpace {
6082 2 : ranges: vec![NON_INHERITED_RANGE],
6083 2 : };
6084 2 : let read_lsn = child_timeline.get_last_record_lsn();
6085 2 :
6086 2 : let vectored_res = child_timeline
6087 2 : .get_vectored_impl(
6088 2 : aux_keyspace.clone(),
6089 2 : read_lsn,
6090 2 : &mut ValuesReconstructState::new(),
6091 2 : &ctx,
6092 2 : )
6093 2 : .await;
6094 2 :
6095 2 : let images = vectored_res?;
6096 2 : assert!(images.is_empty());
6097 2 : Ok(())
6098 2 : }
6099 :
6100 : // Test that vectored get handles layer gaps correctly
6101 : // by advancing into the next ancestor timeline if required.
6102 : //
6103 : // The test generates timelines that look like the diagram below.
6104 : // We leave a gap in one of the L1 layers at `gap_at_key` (`/` in the diagram).
6105 : // The reconstruct data for that key lies in the ancestor timeline (`X` in the diagram).
6106 : //
6107 : // ```
6108 : //-------------------------------+
6109 : // ... |
6110 : // [ L1 ] |
6111 : // [ / L1 ] | Child Timeline
6112 : // ... |
6113 : // ------------------------------+
6114 : // [ X L1 ] | Parent Timeline
6115 : // ------------------------------+
6116 : // ```
6117 : #[tokio::test]
6118 2 : async fn test_get_vectored_key_gap() -> anyhow::Result<()> {
6119 2 : let tenant_conf = TenantConf {
6120 2 : // Make compaction deterministic
6121 2 : gc_period: Duration::ZERO,
6122 2 : compaction_period: Duration::ZERO,
6123 2 : // Encourage creation of L1 layers
6124 2 : checkpoint_distance: 16 * 1024,
6125 2 : compaction_target_size: 8 * 1024,
6126 2 : ..TenantConf::default()
6127 2 : };
6128 2 :
6129 2 : let harness = TenantHarness::create_custom(
6130 2 : "test_get_vectored_key_gap",
6131 2 : tenant_conf,
6132 2 : TenantId::generate(),
6133 2 : ShardIdentity::unsharded(),
6134 2 : Generation::new(0xdeadbeef),
6135 2 : )
6136 2 : .await?;
6137 20 : let (tenant, ctx) = harness.load().await;
6138 2 :
6139 2 : let mut current_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6140 2 : let gap_at_key = current_key.add(100);
6141 2 : let mut current_lsn = Lsn(0x10);
6142 2 :
6143 2 : const KEY_COUNT: usize = 10_000;
6144 2 :
6145 2 : let timeline_id = TimelineId::generate();
6146 2 : let current_timeline = tenant
6147 2 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
6148 6 : .await?;
6149 2 :
6150 2 : current_lsn += 0x100;
6151 2 :
6152 2 : let mut writer = current_timeline.writer().await;
6153 2 : writer
6154 2 : .put(
6155 2 : gap_at_key,
6156 2 : current_lsn,
6157 2 : &Value::Image(test_img(&format!("{} at {}", gap_at_key, current_lsn))),
6158 2 : &ctx,
6159 2 : )
6160 2 : .await?;
6161 2 : writer.finish_write(current_lsn);
6162 2 : drop(writer);
6163 2 :
6164 2 : let mut latest_lsns = HashMap::new();
6165 2 : latest_lsns.insert(gap_at_key, current_lsn);
6166 2 :
6167 2 : current_timeline.freeze_and_flush().await?;
6168 2 :
6169 2 : let child_timeline_id = TimelineId::generate();
6170 2 :
6171 2 : tenant
6172 2 : .branch_timeline_test(
6173 2 : ¤t_timeline,
6174 2 : child_timeline_id,
6175 2 : Some(current_lsn),
6176 2 : &ctx,
6177 2 : )
6178 2 : .await?;
6179 2 : let child_timeline = tenant
6180 2 : .get_timeline(child_timeline_id, true)
6181 2 : .expect("Should have the branched timeline");
6182 2 :
6183 20002 : for i in 0..KEY_COUNT {
6184 20000 : if current_key == gap_at_key {
6185 2 : current_key = current_key.next();
6186 2 : continue;
6187 19998 : }
6188 19998 :
6189 19998 : current_lsn += 0x10;
6190 2 :
6191 19998 : let mut writer = child_timeline.writer().await;
6192 19998 : writer
6193 19998 : .put(
6194 19998 : current_key,
6195 19998 : current_lsn,
6196 19998 : &Value::Image(test_img(&format!("{} at {}", current_key, current_lsn))),
6197 19998 : &ctx,
6198 19998 : )
6199 71 : .await?;
6200 19998 : writer.finish_write(current_lsn);
6201 19998 : drop(writer);
6202 19998 :
6203 19998 : latest_lsns.insert(current_key, current_lsn);
6204 19998 : current_key = current_key.next();
6205 19998 :
6206 19998 : // Flush every now and then to encourage layer file creation.
6207 19998 : if i % 500 == 0 {
6208 43 : child_timeline.freeze_and_flush().await?;
6209 19958 : }
6210 2 : }
6211 2 :
6212 2 : child_timeline.freeze_and_flush().await?;
6213 2 : let mut flags = EnumSet::new();
6214 2 : flags.insert(CompactFlags::ForceRepartition);
6215 2 : child_timeline
6216 2 : .compact(&CancellationToken::new(), flags, &ctx)
6217 1757 : .await?;
6218 2 :
6219 2 : let key_near_end = {
6220 2 : let mut tmp = current_key;
6221 2 : tmp.field6 -= 10;
6222 2 : tmp
6223 2 : };
6224 2 :
6225 2 : let key_near_gap = {
6226 2 : let mut tmp = gap_at_key;
6227 2 : tmp.field6 -= 10;
6228 2 : tmp
6229 2 : };
6230 2 :
6231 2 : let read = KeySpace {
6232 2 : ranges: vec![key_near_gap..gap_at_key.next(), key_near_end..current_key],
6233 2 : };
6234 2 : let results = child_timeline
6235 2 : .get_vectored_impl(
6236 2 : read.clone(),
6237 2 : current_lsn,
6238 2 : &mut ValuesReconstructState::new(),
6239 2 : &ctx,
6240 2 : )
6241 16 : .await?;
6242 2 :
6243 44 : for (key, img_res) in results {
6244 42 : let expected = test_img(&format!("{} at {}", key, latest_lsns[&key]));
6245 42 : assert_eq!(img_res?, expected);
6246 2 : }
6247 2 :
6248 2 : Ok(())
6249 2 : }
6250 :
6251 : // Test that vectored get descends into ancestor timelines correctly and
6252 : // does not return an image that's newer than requested.
6253 : //
6254 : // The diagram below ilustrates an interesting case. We have a parent timeline
6255 : // (top of the Lsn range) and a child timeline. The request key cannot be reconstructed
6256 : // from the child timeline, so the parent timeline must be visited. When advacing into
6257 : // the child timeline, the read path needs to remember what the requested Lsn was in
6258 : // order to avoid returning an image that's too new. The test below constructs such
6259 : // a timeline setup and does a few queries around the Lsn of each page image.
6260 : // ```
6261 : // LSN
6262 : // ^
6263 : // |
6264 : // |
6265 : // 500 | --------------------------------------> branch point
6266 : // 400 | X
6267 : // 300 | X
6268 : // 200 | --------------------------------------> requested lsn
6269 : // 100 | X
6270 : // |---------------------------------------> Key
6271 : // |
6272 : // ------> requested key
6273 : //
6274 : // Legend:
6275 : // * X - page images
6276 : // ```
6277 : #[tokio::test]
6278 2 : async fn test_get_vectored_ancestor_descent() -> anyhow::Result<()> {
6279 2 : let harness = TenantHarness::create("test_get_vectored_on_lsn_axis").await?;
6280 20 : let (tenant, ctx) = harness.load().await;
6281 2 :
6282 2 : let start_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6283 2 : let end_key = start_key.add(1000);
6284 2 : let child_gap_at_key = start_key.add(500);
6285 2 : let mut parent_gap_lsns: BTreeMap<Lsn, String> = BTreeMap::new();
6286 2 :
6287 2 : let mut current_lsn = Lsn(0x10);
6288 2 :
6289 2 : let timeline_id = TimelineId::generate();
6290 2 : let parent_timeline = tenant
6291 2 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
6292 6 : .await?;
6293 2 :
6294 2 : current_lsn += 0x100;
6295 2 :
6296 8 : for _ in 0..3 {
6297 6 : let mut key = start_key;
6298 6006 : while key < end_key {
6299 6000 : current_lsn += 0x10;
6300 6000 :
6301 6000 : let image_value = format!("{} at {}", child_gap_at_key, current_lsn);
6302 2 :
6303 6000 : let mut writer = parent_timeline.writer().await;
6304 6000 : writer
6305 6000 : .put(
6306 6000 : key,
6307 6000 : current_lsn,
6308 6000 : &Value::Image(test_img(&image_value)),
6309 6000 : &ctx,
6310 6000 : )
6311 6 : .await?;
6312 6000 : writer.finish_write(current_lsn);
6313 6000 :
6314 6000 : if key == child_gap_at_key {
6315 6 : parent_gap_lsns.insert(current_lsn, image_value);
6316 5994 : }
6317 2 :
6318 6000 : key = key.next();
6319 2 : }
6320 2 :
6321 7 : parent_timeline.freeze_and_flush().await?;
6322 2 : }
6323 2 :
6324 2 : let child_timeline_id = TimelineId::generate();
6325 2 :
6326 2 : let child_timeline = tenant
6327 2 : .branch_timeline_test(&parent_timeline, child_timeline_id, Some(current_lsn), &ctx)
6328 2 : .await?;
6329 2 :
6330 2 : let mut key = start_key;
6331 2002 : while key < end_key {
6332 2000 : if key == child_gap_at_key {
6333 2 : key = key.next();
6334 2 : continue;
6335 1998 : }
6336 1998 :
6337 1998 : current_lsn += 0x10;
6338 2 :
6339 1998 : let mut writer = child_timeline.writer().await;
6340 1998 : writer
6341 1998 : .put(
6342 1998 : key,
6343 1998 : current_lsn,
6344 1998 : &Value::Image(test_img(&format!("{} at {}", key, current_lsn))),
6345 1998 : &ctx,
6346 1998 : )
6347 17 : .await?;
6348 1998 : writer.finish_write(current_lsn);
6349 1998 :
6350 1998 : key = key.next();
6351 2 : }
6352 2 :
6353 2 : child_timeline.freeze_and_flush().await?;
6354 2 :
6355 2 : let lsn_offsets: [i64; 5] = [-10, -1, 0, 1, 10];
6356 2 : let mut query_lsns = Vec::new();
6357 6 : for image_lsn in parent_gap_lsns.keys().rev() {
6358 36 : for offset in lsn_offsets {
6359 30 : query_lsns.push(Lsn(image_lsn
6360 30 : .0
6361 30 : .checked_add_signed(offset)
6362 30 : .expect("Shouldn't overflow")));
6363 30 : }
6364 2 : }
6365 2 :
6366 32 : for query_lsn in query_lsns {
6367 30 : let results = child_timeline
6368 30 : .get_vectored_impl(
6369 30 : KeySpace {
6370 30 : ranges: vec![child_gap_at_key..child_gap_at_key.next()],
6371 30 : },
6372 30 : query_lsn,
6373 30 : &mut ValuesReconstructState::new(),
6374 30 : &ctx,
6375 30 : )
6376 29 : .await;
6377 2 :
6378 30 : let expected_item = parent_gap_lsns
6379 30 : .iter()
6380 30 : .rev()
6381 68 : .find(|(lsn, _)| **lsn <= query_lsn);
6382 30 :
6383 30 : info!(
6384 2 : "Doing vectored read at LSN {}. Expecting image to be: {:?}",
6385 2 : query_lsn, expected_item
6386 2 : );
6387 2 :
6388 30 : match expected_item {
6389 26 : Some((_, img_value)) => {
6390 26 : let key_results = results.expect("No vectored get error expected");
6391 26 : let key_result = &key_results[&child_gap_at_key];
6392 26 : let returned_img = key_result
6393 26 : .as_ref()
6394 26 : .expect("No page reconstruct error expected");
6395 26 :
6396 26 : info!(
6397 2 : "Vectored read at LSN {} returned image {}",
6398 0 : query_lsn,
6399 0 : std::str::from_utf8(returned_img)?
6400 2 : );
6401 26 : assert_eq!(*returned_img, test_img(img_value));
6402 2 : }
6403 2 : None => {
6404 4 : assert!(matches!(results, Err(GetVectoredError::MissingKey(_))));
6405 2 : }
6406 2 : }
6407 2 : }
6408 2 :
6409 2 : Ok(())
6410 2 : }
6411 :
6412 : #[tokio::test]
6413 2 : async fn test_random_updates() -> anyhow::Result<()> {
6414 2 : let names_algorithms = [
6415 2 : ("test_random_updates_legacy", CompactionAlgorithm::Legacy),
6416 2 : ("test_random_updates_tiered", CompactionAlgorithm::Tiered),
6417 2 : ];
6418 6 : for (name, algorithm) in names_algorithms {
6419 96229 : test_random_updates_algorithm(name, algorithm).await?;
6420 2 : }
6421 2 : Ok(())
6422 2 : }
6423 :
6424 4 : async fn test_random_updates_algorithm(
6425 4 : name: &'static str,
6426 4 : compaction_algorithm: CompactionAlgorithm,
6427 4 : ) -> anyhow::Result<()> {
6428 4 : let mut harness = TenantHarness::create(name).await?;
6429 4 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
6430 4 : kind: compaction_algorithm,
6431 4 : };
6432 40 : let (tenant, ctx) = harness.load().await;
6433 4 : let tline = tenant
6434 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6435 11 : .await?;
6436 :
6437 : const NUM_KEYS: usize = 1000;
6438 4 : let cancel = CancellationToken::new();
6439 4 :
6440 4 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6441 4 : let mut test_key_end = test_key;
6442 4 : test_key_end.field6 = NUM_KEYS as u32;
6443 4 : tline.add_extra_test_dense_keyspace(KeySpace::single(test_key..test_key_end));
6444 4 :
6445 4 : let mut keyspace = KeySpaceAccum::new();
6446 4 :
6447 4 : // Track when each page was last modified. Used to assert that
6448 4 : // a read sees the latest page version.
6449 4 : let mut updated = [Lsn(0); NUM_KEYS];
6450 4 :
6451 4 : let mut lsn = Lsn(0x10);
6452 : #[allow(clippy::needless_range_loop)]
6453 4004 : for blknum in 0..NUM_KEYS {
6454 4000 : lsn = Lsn(lsn.0 + 0x10);
6455 4000 : test_key.field6 = blknum as u32;
6456 4000 : let mut writer = tline.writer().await;
6457 4000 : writer
6458 4000 : .put(
6459 4000 : test_key,
6460 4000 : lsn,
6461 4000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6462 4000 : &ctx,
6463 4000 : )
6464 4 : .await?;
6465 4000 : writer.finish_write(lsn);
6466 4000 : updated[blknum] = lsn;
6467 4000 : drop(writer);
6468 4000 :
6469 4000 : keyspace.add_key(test_key);
6470 : }
6471 :
6472 204 : for _ in 0..50 {
6473 200200 : for _ in 0..NUM_KEYS {
6474 200000 : lsn = Lsn(lsn.0 + 0x10);
6475 200000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6476 200000 : test_key.field6 = blknum as u32;
6477 200000 : let mut writer = tline.writer().await;
6478 200000 : writer
6479 200000 : .put(
6480 200000 : test_key,
6481 200000 : lsn,
6482 200000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6483 200000 : &ctx,
6484 200000 : )
6485 198 : .await?;
6486 200000 : writer.finish_write(lsn);
6487 200000 : drop(writer);
6488 200000 : updated[blknum] = lsn;
6489 : }
6490 :
6491 : // Read all the blocks
6492 200000 : for (blknum, last_lsn) in updated.iter().enumerate() {
6493 200000 : test_key.field6 = blknum as u32;
6494 200000 : assert_eq!(
6495 200000 : tline.get(test_key, lsn, &ctx).await?,
6496 200000 : test_img(&format!("{} at {}", blknum, last_lsn))
6497 : );
6498 : }
6499 :
6500 : // Perform a cycle of flush, and GC
6501 202 : tline.freeze_and_flush().await?;
6502 200 : tenant
6503 200 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
6504 0 : .await?;
6505 : }
6506 :
6507 4 : Ok(())
6508 4 : }
6509 :
6510 : #[tokio::test]
6511 2 : async fn test_traverse_branches() -> anyhow::Result<()> {
6512 2 : let (tenant, ctx) = TenantHarness::create("test_traverse_branches")
6513 2 : .await?
6514 2 : .load()
6515 20 : .await;
6516 2 : let mut tline = tenant
6517 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6518 6 : .await?;
6519 2 :
6520 2 : const NUM_KEYS: usize = 1000;
6521 2 :
6522 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6523 2 :
6524 2 : let mut keyspace = KeySpaceAccum::new();
6525 2 :
6526 2 : let cancel = CancellationToken::new();
6527 2 :
6528 2 : // Track when each page was last modified. Used to assert that
6529 2 : // a read sees the latest page version.
6530 2 : let mut updated = [Lsn(0); NUM_KEYS];
6531 2 :
6532 2 : let mut lsn = Lsn(0x10);
6533 2 : #[allow(clippy::needless_range_loop)]
6534 2002 : for blknum in 0..NUM_KEYS {
6535 2000 : lsn = Lsn(lsn.0 + 0x10);
6536 2000 : test_key.field6 = blknum as u32;
6537 2000 : let mut writer = tline.writer().await;
6538 2000 : writer
6539 2000 : .put(
6540 2000 : test_key,
6541 2000 : lsn,
6542 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6543 2000 : &ctx,
6544 2000 : )
6545 2 : .await?;
6546 2000 : writer.finish_write(lsn);
6547 2000 : updated[blknum] = lsn;
6548 2000 : drop(writer);
6549 2000 :
6550 2000 : keyspace.add_key(test_key);
6551 2 : }
6552 2 :
6553 102 : for _ in 0..50 {
6554 100 : let new_tline_id = TimelineId::generate();
6555 100 : tenant
6556 100 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
6557 98 : .await?;
6558 100 : tline = tenant
6559 100 : .get_timeline(new_tline_id, true)
6560 100 : .expect("Should have the branched timeline");
6561 2 :
6562 100100 : for _ in 0..NUM_KEYS {
6563 100000 : lsn = Lsn(lsn.0 + 0x10);
6564 100000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6565 100000 : test_key.field6 = blknum as u32;
6566 100000 : let mut writer = tline.writer().await;
6567 100000 : writer
6568 100000 : .put(
6569 100000 : test_key,
6570 100000 : lsn,
6571 100000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6572 100000 : &ctx,
6573 100000 : )
6574 835 : .await?;
6575 100000 : println!("updating {} at {}", blknum, lsn);
6576 100000 : writer.finish_write(lsn);
6577 100000 : drop(writer);
6578 100000 : updated[blknum] = lsn;
6579 2 : }
6580 2 :
6581 2 : // Read all the blocks
6582 100000 : for (blknum, last_lsn) in updated.iter().enumerate() {
6583 100000 : test_key.field6 = blknum as u32;
6584 100000 : assert_eq!(
6585 100000 : tline.get(test_key, lsn, &ctx).await?,
6586 100000 : test_img(&format!("{} at {}", blknum, last_lsn))
6587 2 : );
6588 2 : }
6589 2 :
6590 2 : // Perform a cycle of flush, compact, and GC
6591 100 : tline.freeze_and_flush().await?;
6592 14992 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
6593 100 : tenant
6594 100 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
6595 2 : .await?;
6596 2 : }
6597 2 :
6598 2 : Ok(())
6599 2 : }
6600 :
6601 : #[tokio::test]
6602 2 : async fn test_traverse_ancestors() -> anyhow::Result<()> {
6603 2 : let (tenant, ctx) = TenantHarness::create("test_traverse_ancestors")
6604 2 : .await?
6605 2 : .load()
6606 20 : .await;
6607 2 : let mut tline = tenant
6608 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6609 6 : .await?;
6610 2 :
6611 2 : const NUM_KEYS: usize = 100;
6612 2 : const NUM_TLINES: usize = 50;
6613 2 :
6614 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6615 2 : // Track page mutation lsns across different timelines.
6616 2 : let mut updated = [[Lsn(0); NUM_KEYS]; NUM_TLINES];
6617 2 :
6618 2 : let mut lsn = Lsn(0x10);
6619 2 :
6620 2 : #[allow(clippy::needless_range_loop)]
6621 102 : for idx in 0..NUM_TLINES {
6622 100 : let new_tline_id = TimelineId::generate();
6623 100 : tenant
6624 100 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
6625 67 : .await?;
6626 100 : tline = tenant
6627 100 : .get_timeline(new_tline_id, true)
6628 100 : .expect("Should have the branched timeline");
6629 2 :
6630 10100 : for _ in 0..NUM_KEYS {
6631 10000 : lsn = Lsn(lsn.0 + 0x10);
6632 10000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6633 10000 : test_key.field6 = blknum as u32;
6634 10000 : let mut writer = tline.writer().await;
6635 10000 : writer
6636 10000 : .put(
6637 10000 : test_key,
6638 10000 : lsn,
6639 10000 : &Value::Image(test_img(&format!("{} {} at {}", idx, blknum, lsn))),
6640 10000 : &ctx,
6641 10000 : )
6642 108 : .await?;
6643 10000 : println!("updating [{}][{}] at {}", idx, blknum, lsn);
6644 10000 : writer.finish_write(lsn);
6645 10000 : drop(writer);
6646 10000 : updated[idx][blknum] = lsn;
6647 2 : }
6648 2 : }
6649 2 :
6650 2 : // Read pages from leaf timeline across all ancestors.
6651 100 : for (idx, lsns) in updated.iter().enumerate() {
6652 10000 : for (blknum, lsn) in lsns.iter().enumerate() {
6653 2 : // Skip empty mutations.
6654 10000 : if lsn.0 == 0 {
6655 3665 : continue;
6656 6335 : }
6657 6335 : println!("checking [{idx}][{blknum}] at {lsn}");
6658 6335 : test_key.field6 = blknum as u32;
6659 6335 : assert_eq!(
6660 6335 : tline.get(test_key, *lsn, &ctx).await?,
6661 6335 : test_img(&format!("{idx} {blknum} at {lsn}"))
6662 2 : );
6663 2 : }
6664 2 : }
6665 2 : Ok(())
6666 2 : }
6667 :
6668 : #[tokio::test]
6669 2 : async fn test_write_at_initdb_lsn_takes_optimization_code_path() -> anyhow::Result<()> {
6670 2 : let (tenant, ctx) = TenantHarness::create("test_empty_test_timeline_is_usable")
6671 2 : .await?
6672 2 : .load()
6673 20 : .await;
6674 2 :
6675 2 : let initdb_lsn = Lsn(0x20);
6676 2 : let utline = tenant
6677 2 : .create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)
6678 2 : .await?;
6679 2 : let tline = utline.raw_timeline().unwrap();
6680 2 :
6681 2 : // Spawn flush loop now so that we can set the `expect_initdb_optimization`
6682 2 : tline.maybe_spawn_flush_loop();
6683 2 :
6684 2 : // Make sure the timeline has the minimum set of required keys for operation.
6685 2 : // The only operation you can always do on an empty timeline is to `put` new data.
6686 2 : // Except if you `put` at `initdb_lsn`.
6687 2 : // In that case, there's an optimization to directly create image layers instead of delta layers.
6688 2 : // It uses `repartition()`, which assumes some keys to be present.
6689 2 : // Let's make sure the test timeline can handle that case.
6690 2 : {
6691 2 : let mut state = tline.flush_loop_state.lock().unwrap();
6692 2 : assert_eq!(
6693 2 : timeline::FlushLoopState::Running {
6694 2 : expect_initdb_optimization: false,
6695 2 : initdb_optimization_count: 0,
6696 2 : },
6697 2 : *state
6698 2 : );
6699 2 : *state = timeline::FlushLoopState::Running {
6700 2 : expect_initdb_optimization: true,
6701 2 : initdb_optimization_count: 0,
6702 2 : };
6703 2 : }
6704 2 :
6705 2 : // Make writes at the initdb_lsn. When we flush it below, it should be handled by the optimization.
6706 2 : // As explained above, the optimization requires some keys to be present.
6707 2 : // As per `create_empty_timeline` documentation, use init_empty to set them.
6708 2 : // This is what `create_test_timeline` does, by the way.
6709 2 : let mut modification = tline.begin_modification(initdb_lsn);
6710 2 : modification
6711 2 : .init_empty_test_timeline()
6712 2 : .context("init_empty_test_timeline")?;
6713 2 : modification
6714 2 : .commit(&ctx)
6715 2 : .await
6716 2 : .context("commit init_empty_test_timeline modification")?;
6717 2 :
6718 2 : // Do the flush. The flush code will check the expectations that we set above.
6719 2 : tline.freeze_and_flush().await?;
6720 2 :
6721 2 : // assert freeze_and_flush exercised the initdb optimization
6722 2 : {
6723 2 : let state = tline.flush_loop_state.lock().unwrap();
6724 2 : let timeline::FlushLoopState::Running {
6725 2 : expect_initdb_optimization,
6726 2 : initdb_optimization_count,
6727 2 : } = *state
6728 2 : else {
6729 2 : panic!("unexpected state: {:?}", *state);
6730 2 : };
6731 2 : assert!(expect_initdb_optimization);
6732 2 : assert!(initdb_optimization_count > 0);
6733 2 : }
6734 2 : Ok(())
6735 2 : }
6736 :
6737 : #[tokio::test]
6738 2 : async fn test_create_guard_crash() -> anyhow::Result<()> {
6739 2 : let name = "test_create_guard_crash";
6740 2 : let harness = TenantHarness::create(name).await?;
6741 2 : {
6742 20 : let (tenant, ctx) = harness.load().await;
6743 2 : let tline = tenant
6744 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
6745 2 : .await?;
6746 2 : // Leave the timeline ID in [`Tenant::timelines_creating`] to exclude attempting to create it again
6747 2 : let raw_tline = tline.raw_timeline().unwrap();
6748 2 : raw_tline
6749 2 : .shutdown(super::timeline::ShutdownMode::Hard)
6750 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))
6751 2 : .await;
6752 2 : std::mem::forget(tline);
6753 2 : }
6754 2 :
6755 20 : let (tenant, _) = harness.load().await;
6756 2 : match tenant.get_timeline(TIMELINE_ID, false) {
6757 2 : Ok(_) => panic!("timeline should've been removed during load"),
6758 2 : Err(e) => {
6759 2 : assert_eq!(
6760 2 : e,
6761 2 : GetTimelineError::NotFound {
6762 2 : tenant_id: tenant.tenant_shard_id,
6763 2 : timeline_id: TIMELINE_ID,
6764 2 : }
6765 2 : )
6766 2 : }
6767 2 : }
6768 2 :
6769 2 : assert!(!harness
6770 2 : .conf
6771 2 : .timeline_path(&tenant.tenant_shard_id, &TIMELINE_ID)
6772 2 : .exists());
6773 2 :
6774 2 : Ok(())
6775 2 : }
6776 :
6777 : #[tokio::test]
6778 2 : async fn test_read_at_max_lsn() -> anyhow::Result<()> {
6779 2 : let names_algorithms = [
6780 2 : ("test_read_at_max_lsn_legacy", CompactionAlgorithm::Legacy),
6781 2 : ("test_read_at_max_lsn_tiered", CompactionAlgorithm::Tiered),
6782 2 : ];
6783 6 : for (name, algorithm) in names_algorithms {
6784 32561 : test_read_at_max_lsn_algorithm(name, algorithm).await?;
6785 2 : }
6786 2 : Ok(())
6787 2 : }
6788 :
6789 4 : async fn test_read_at_max_lsn_algorithm(
6790 4 : name: &'static str,
6791 4 : compaction_algorithm: CompactionAlgorithm,
6792 4 : ) -> anyhow::Result<()> {
6793 4 : let mut harness = TenantHarness::create(name).await?;
6794 4 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
6795 4 : kind: compaction_algorithm,
6796 4 : };
6797 40 : let (tenant, ctx) = harness.load().await;
6798 4 : let tline = tenant
6799 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6800 11 : .await?;
6801 :
6802 4 : let lsn = Lsn(0x10);
6803 4 : let compact = false;
6804 32100 : bulk_insert_maybe_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000, compact).await?;
6805 :
6806 4 : let test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6807 4 : let read_lsn = Lsn(u64::MAX - 1);
6808 :
6809 410 : let result = tline.get(test_key, read_lsn, &ctx).await;
6810 4 : assert!(result.is_ok(), "result is not Ok: {}", result.unwrap_err());
6811 :
6812 4 : Ok(())
6813 4 : }
6814 :
6815 : #[tokio::test]
6816 2 : async fn test_metadata_scan() -> anyhow::Result<()> {
6817 2 : let harness = TenantHarness::create("test_metadata_scan").await?;
6818 20 : let (tenant, ctx) = harness.load().await;
6819 2 : let tline = tenant
6820 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6821 6 : .await?;
6822 2 :
6823 2 : const NUM_KEYS: usize = 1000;
6824 2 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
6825 2 :
6826 2 : let cancel = CancellationToken::new();
6827 2 :
6828 2 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
6829 2 : base_key.field1 = AUX_KEY_PREFIX;
6830 2 : let mut test_key = base_key;
6831 2 :
6832 2 : // Track when each page was last modified. Used to assert that
6833 2 : // a read sees the latest page version.
6834 2 : let mut updated = [Lsn(0); NUM_KEYS];
6835 2 :
6836 2 : let mut lsn = Lsn(0x10);
6837 2 : #[allow(clippy::needless_range_loop)]
6838 2002 : for blknum in 0..NUM_KEYS {
6839 2000 : lsn = Lsn(lsn.0 + 0x10);
6840 2000 : test_key.field6 = (blknum * STEP) as u32;
6841 2000 : let mut writer = tline.writer().await;
6842 2000 : writer
6843 2000 : .put(
6844 2000 : test_key,
6845 2000 : lsn,
6846 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6847 2000 : &ctx,
6848 2000 : )
6849 2 : .await?;
6850 2000 : writer.finish_write(lsn);
6851 2000 : updated[blknum] = lsn;
6852 2000 : drop(writer);
6853 2 : }
6854 2 :
6855 2 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
6856 2 :
6857 24 : for iter in 0..=10 {
6858 2 : // Read all the blocks
6859 22000 : for (blknum, last_lsn) in updated.iter().enumerate() {
6860 22000 : test_key.field6 = (blknum * STEP) as u32;
6861 22000 : assert_eq!(
6862 22000 : tline.get(test_key, lsn, &ctx).await?,
6863 22000 : test_img(&format!("{} at {}", blknum, last_lsn))
6864 2 : );
6865 2 : }
6866 2 :
6867 22 : let mut cnt = 0;
6868 22000 : for (key, value) in tline
6869 22 : .get_vectored_impl(
6870 22 : keyspace.clone(),
6871 22 : lsn,
6872 22 : &mut ValuesReconstructState::default(),
6873 22 : &ctx,
6874 22 : )
6875 741 : .await?
6876 2 : {
6877 22000 : let blknum = key.field6 as usize;
6878 22000 : let value = value?;
6879 22000 : assert!(blknum % STEP == 0);
6880 22000 : let blknum = blknum / STEP;
6881 22000 : assert_eq!(
6882 22000 : value,
6883 22000 : test_img(&format!("{} at {}", blknum, updated[blknum]))
6884 22000 : );
6885 22000 : cnt += 1;
6886 2 : }
6887 2 :
6888 22 : assert_eq!(cnt, NUM_KEYS);
6889 2 :
6890 22022 : for _ in 0..NUM_KEYS {
6891 22000 : lsn = Lsn(lsn.0 + 0x10);
6892 22000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6893 22000 : test_key.field6 = (blknum * STEP) as u32;
6894 22000 : let mut writer = tline.writer().await;
6895 22000 : writer
6896 22000 : .put(
6897 22000 : test_key,
6898 22000 : lsn,
6899 22000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6900 22000 : &ctx,
6901 22000 : )
6902 111 : .await?;
6903 22000 : writer.finish_write(lsn);
6904 22000 : drop(writer);
6905 22000 : updated[blknum] = lsn;
6906 2 : }
6907 2 :
6908 2 : // Perform two cycles of flush, compact, and GC
6909 66 : for round in 0..2 {
6910 44 : tline.freeze_and_flush().await?;
6911 44 : tline
6912 44 : .compact(
6913 44 : &cancel,
6914 44 : if iter % 5 == 0 && round == 0 {
6915 6 : let mut flags = EnumSet::new();
6916 6 : flags.insert(CompactFlags::ForceImageLayerCreation);
6917 6 : flags.insert(CompactFlags::ForceRepartition);
6918 6 : flags
6919 2 : } else {
6920 38 : EnumSet::empty()
6921 2 : },
6922 44 : &ctx,
6923 2 : )
6924 6702 : .await?;
6925 44 : tenant
6926 44 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
6927 2 : .await?;
6928 2 : }
6929 2 : }
6930 2 :
6931 2 : Ok(())
6932 2 : }
6933 :
6934 : #[tokio::test]
6935 2 : async fn test_metadata_compaction_trigger() -> anyhow::Result<()> {
6936 2 : let harness = TenantHarness::create("test_metadata_compaction_trigger").await?;
6937 20 : let (tenant, ctx) = harness.load().await;
6938 2 : let tline = tenant
6939 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6940 6 : .await?;
6941 2 :
6942 2 : let cancel = CancellationToken::new();
6943 2 :
6944 2 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
6945 2 : base_key.field1 = AUX_KEY_PREFIX;
6946 2 : let test_key = base_key;
6947 2 : let mut lsn = Lsn(0x10);
6948 2 :
6949 42 : for _ in 0..20 {
6950 40 : lsn = Lsn(lsn.0 + 0x10);
6951 40 : let mut writer = tline.writer().await;
6952 40 : writer
6953 40 : .put(
6954 40 : test_key,
6955 40 : lsn,
6956 40 : &Value::Image(test_img(&format!("{} at {}", 0, lsn))),
6957 40 : &ctx,
6958 40 : )
6959 20 : .await?;
6960 40 : writer.finish_write(lsn);
6961 40 : drop(writer);
6962 40 : tline.freeze_and_flush().await?; // force create a delta layer
6963 2 : }
6964 2 :
6965 2 : let before_num_l0_delta_files =
6966 2 : tline.layers.read().await.layer_map()?.level0_deltas().len();
6967 2 :
6968 110 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
6969 2 :
6970 2 : let after_num_l0_delta_files = tline.layers.read().await.layer_map()?.level0_deltas().len();
6971 2 :
6972 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}");
6973 2 :
6974 2 : assert_eq!(
6975 4 : tline.get(test_key, lsn, &ctx).await?,
6976 2 : test_img(&format!("{} at {}", 0, lsn))
6977 2 : );
6978 2 :
6979 2 : Ok(())
6980 2 : }
6981 :
6982 : #[tokio::test]
6983 2 : async fn test_aux_file_e2e() {
6984 2 : let harness = TenantHarness::create("test_aux_file_e2e").await.unwrap();
6985 2 :
6986 20 : let (tenant, ctx) = harness.load().await;
6987 2 :
6988 2 : let mut lsn = Lsn(0x08);
6989 2 :
6990 2 : let tline: Arc<Timeline> = tenant
6991 2 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
6992 6 : .await
6993 2 : .unwrap();
6994 2 :
6995 2 : {
6996 2 : lsn += 8;
6997 2 : let mut modification = tline.begin_modification(lsn);
6998 2 : modification
6999 2 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
7000 2 : .await
7001 2 : .unwrap();
7002 2 : modification.commit(&ctx).await.unwrap();
7003 2 : }
7004 2 :
7005 2 : // we can read everything from the storage
7006 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
7007 2 : assert_eq!(
7008 2 : files.get("pg_logical/mappings/test1"),
7009 2 : Some(&bytes::Bytes::from_static(b"first"))
7010 2 : );
7011 2 :
7012 2 : {
7013 2 : lsn += 8;
7014 2 : let mut modification = tline.begin_modification(lsn);
7015 2 : modification
7016 2 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
7017 2 : .await
7018 2 : .unwrap();
7019 2 : modification.commit(&ctx).await.unwrap();
7020 2 : }
7021 2 :
7022 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
7023 2 : assert_eq!(
7024 2 : files.get("pg_logical/mappings/test2"),
7025 2 : Some(&bytes::Bytes::from_static(b"second"))
7026 2 : );
7027 2 :
7028 2 : let child = tenant
7029 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(lsn), &ctx)
7030 2 : .await
7031 2 : .unwrap();
7032 2 :
7033 2 : let files = child.list_aux_files(lsn, &ctx).await.unwrap();
7034 2 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
7035 2 : assert_eq!(files.get("pg_logical/mappings/test2"), None);
7036 2 : }
7037 :
7038 : #[tokio::test]
7039 2 : async fn test_metadata_image_creation() -> anyhow::Result<()> {
7040 2 : let harness = TenantHarness::create("test_metadata_image_creation").await?;
7041 20 : let (tenant, ctx) = harness.load().await;
7042 2 : let tline = tenant
7043 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7044 6 : .await?;
7045 2 :
7046 2 : const NUM_KEYS: usize = 1000;
7047 2 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
7048 2 :
7049 2 : let cancel = CancellationToken::new();
7050 2 :
7051 2 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7052 2 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
7053 2 : let mut test_key = base_key;
7054 2 : let mut lsn = Lsn(0x10);
7055 2 :
7056 8 : async fn scan_with_statistics(
7057 8 : tline: &Timeline,
7058 8 : keyspace: &KeySpace,
7059 8 : lsn: Lsn,
7060 8 : ctx: &RequestContext,
7061 8 : ) -> anyhow::Result<(BTreeMap<Key, Result<Bytes, PageReconstructError>>, usize)> {
7062 8 : let mut reconstruct_state = ValuesReconstructState::default();
7063 8 : let res = tline
7064 8 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
7065 241 : .await?;
7066 8 : Ok((res, reconstruct_state.get_delta_layers_visited() as usize))
7067 8 : }
7068 2 :
7069 2 : #[allow(clippy::needless_range_loop)]
7070 2002 : for blknum in 0..NUM_KEYS {
7071 2000 : lsn = Lsn(lsn.0 + 0x10);
7072 2000 : test_key.field6 = (blknum * STEP) as u32;
7073 2000 : let mut writer = tline.writer().await;
7074 2000 : writer
7075 2000 : .put(
7076 2000 : test_key,
7077 2000 : lsn,
7078 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7079 2000 : &ctx,
7080 2000 : )
7081 2 : .await?;
7082 2000 : writer.finish_write(lsn);
7083 2000 : drop(writer);
7084 2 : }
7085 2 :
7086 2 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
7087 2 :
7088 22 : for iter in 1..=10 {
7089 20020 : for _ in 0..NUM_KEYS {
7090 20000 : lsn = Lsn(lsn.0 + 0x10);
7091 20000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7092 20000 : test_key.field6 = (blknum * STEP) as u32;
7093 20000 : let mut writer = tline.writer().await;
7094 20000 : writer
7095 20000 : .put(
7096 20000 : test_key,
7097 20000 : lsn,
7098 20000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7099 20000 : &ctx,
7100 20000 : )
7101 35 : .await?;
7102 20000 : writer.finish_write(lsn);
7103 20000 : drop(writer);
7104 2 : }
7105 2 :
7106 20 : tline.freeze_and_flush().await?;
7107 2 :
7108 20 : if iter % 5 == 0 {
7109 4 : let (_, before_delta_file_accessed) =
7110 233 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
7111 4 : tline
7112 4 : .compact(
7113 4 : &cancel,
7114 4 : {
7115 4 : let mut flags = EnumSet::new();
7116 4 : flags.insert(CompactFlags::ForceImageLayerCreation);
7117 4 : flags.insert(CompactFlags::ForceRepartition);
7118 4 : flags
7119 4 : },
7120 4 : &ctx,
7121 4 : )
7122 4829 : .await?;
7123 4 : let (_, after_delta_file_accessed) =
7124 8 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
7125 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}");
7126 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.
7127 4 : assert!(
7128 4 : after_delta_file_accessed <= 2,
7129 2 : "after_delta_file_accessed={after_delta_file_accessed}"
7130 2 : );
7131 16 : }
7132 2 : }
7133 2 :
7134 2 : Ok(())
7135 2 : }
7136 :
7137 : #[tokio::test]
7138 2 : async fn test_vectored_missing_data_key_reads() -> anyhow::Result<()> {
7139 2 : let harness = TenantHarness::create("test_vectored_missing_data_key_reads").await?;
7140 20 : let (tenant, ctx) = harness.load().await;
7141 2 :
7142 2 : let base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7143 2 : let base_key_child = Key::from_hex("000000000033333333444444445500000001").unwrap();
7144 2 : let base_key_nonexist = Key::from_hex("000000000033333333444444445500000002").unwrap();
7145 2 :
7146 2 : let tline = tenant
7147 2 : .create_test_timeline_with_layers(
7148 2 : TIMELINE_ID,
7149 2 : Lsn(0x10),
7150 2 : DEFAULT_PG_VERSION,
7151 2 : &ctx,
7152 2 : Vec::new(), // delta layers
7153 2 : vec![(Lsn(0x20), vec![(base_key, test_img("data key 1"))])], // image layers
7154 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
7155 2 : )
7156 13 : .await?;
7157 2 : tline.add_extra_test_dense_keyspace(KeySpace::single(base_key..(base_key_nonexist.next())));
7158 2 :
7159 2 : let child = tenant
7160 2 : .branch_timeline_test_with_layers(
7161 2 : &tline,
7162 2 : NEW_TIMELINE_ID,
7163 2 : Some(Lsn(0x20)),
7164 2 : &ctx,
7165 2 : Vec::new(), // delta layers
7166 2 : vec![(Lsn(0x30), vec![(base_key_child, test_img("data key 2"))])], // image layers
7167 2 : Lsn(0x30),
7168 2 : )
7169 9 : .await
7170 2 : .unwrap();
7171 2 :
7172 2 : let lsn = Lsn(0x30);
7173 2 :
7174 2 : // test vectored get on parent timeline
7175 2 : assert_eq!(
7176 4 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
7177 2 : Some(test_img("data key 1"))
7178 2 : );
7179 2 : assert!(get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx)
7180 3 : .await
7181 2 : .unwrap_err()
7182 2 : .is_missing_key_error());
7183 2 : assert!(
7184 2 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx)
7185 2 : .await
7186 2 : .unwrap_err()
7187 2 : .is_missing_key_error()
7188 2 : );
7189 2 :
7190 2 : // test vectored get on child timeline
7191 2 : assert_eq!(
7192 2 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
7193 2 : Some(test_img("data key 1"))
7194 2 : );
7195 2 : assert_eq!(
7196 4 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
7197 2 : Some(test_img("data key 2"))
7198 2 : );
7199 2 : assert!(
7200 2 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx)
7201 2 : .await
7202 2 : .unwrap_err()
7203 2 : .is_missing_key_error()
7204 2 : );
7205 2 :
7206 2 : Ok(())
7207 2 : }
7208 :
7209 : #[tokio::test]
7210 2 : async fn test_vectored_missing_metadata_key_reads() -> anyhow::Result<()> {
7211 2 : let harness = TenantHarness::create("test_vectored_missing_metadata_key_reads").await?;
7212 20 : let (tenant, ctx) = harness.load().await;
7213 2 :
7214 2 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7215 2 : let base_key_child = Key::from_hex("620000000033333333444444445500000001").unwrap();
7216 2 : let base_key_nonexist = Key::from_hex("620000000033333333444444445500000002").unwrap();
7217 2 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
7218 2 :
7219 2 : let tline = tenant
7220 2 : .create_test_timeline_with_layers(
7221 2 : TIMELINE_ID,
7222 2 : Lsn(0x10),
7223 2 : DEFAULT_PG_VERSION,
7224 2 : &ctx,
7225 2 : Vec::new(), // delta layers
7226 2 : vec![(Lsn(0x20), vec![(base_key, test_img("metadata key 1"))])], // image layers
7227 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
7228 2 : )
7229 13 : .await?;
7230 2 :
7231 2 : let child = tenant
7232 2 : .branch_timeline_test_with_layers(
7233 2 : &tline,
7234 2 : NEW_TIMELINE_ID,
7235 2 : Some(Lsn(0x20)),
7236 2 : &ctx,
7237 2 : Vec::new(), // delta layers
7238 2 : vec![(
7239 2 : Lsn(0x30),
7240 2 : vec![(base_key_child, test_img("metadata key 2"))],
7241 2 : )], // image layers
7242 2 : Lsn(0x30),
7243 2 : )
7244 9 : .await
7245 2 : .unwrap();
7246 2 :
7247 2 : let lsn = Lsn(0x30);
7248 2 :
7249 2 : // test vectored get on parent timeline
7250 2 : assert_eq!(
7251 4 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
7252 2 : Some(test_img("metadata key 1"))
7253 2 : );
7254 2 : assert_eq!(
7255 2 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx).await?,
7256 2 : None
7257 2 : );
7258 2 : assert_eq!(
7259 2 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx).await?,
7260 2 : None
7261 2 : );
7262 2 :
7263 2 : // test vectored get on child timeline
7264 2 : assert_eq!(
7265 2 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
7266 2 : None
7267 2 : );
7268 2 : assert_eq!(
7269 4 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
7270 2 : Some(test_img("metadata key 2"))
7271 2 : );
7272 2 : assert_eq!(
7273 2 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx).await?,
7274 2 : None
7275 2 : );
7276 2 :
7277 2 : Ok(())
7278 2 : }
7279 :
7280 36 : async fn get_vectored_impl_wrapper(
7281 36 : tline: &Arc<Timeline>,
7282 36 : key: Key,
7283 36 : lsn: Lsn,
7284 36 : ctx: &RequestContext,
7285 36 : ) -> Result<Option<Bytes>, GetVectoredError> {
7286 36 : let mut reconstruct_state = ValuesReconstructState::new();
7287 36 : let mut res = tline
7288 36 : .get_vectored_impl(
7289 36 : KeySpace::single(key..key.next()),
7290 36 : lsn,
7291 36 : &mut reconstruct_state,
7292 36 : ctx,
7293 36 : )
7294 40 : .await?;
7295 30 : Ok(res.pop_last().map(|(k, v)| {
7296 18 : assert_eq!(k, key);
7297 18 : v.unwrap()
7298 30 : }))
7299 36 : }
7300 :
7301 : #[tokio::test]
7302 2 : async fn test_metadata_tombstone_reads() -> anyhow::Result<()> {
7303 2 : let harness = TenantHarness::create("test_metadata_tombstone_reads").await?;
7304 20 : let (tenant, ctx) = harness.load().await;
7305 2 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
7306 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
7307 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
7308 2 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
7309 2 :
7310 2 : // We emulate the situation that the compaction algorithm creates an image layer that removes the tombstones
7311 2 : // Lsn 0x30 key0, key3, no key1+key2
7312 2 : // Lsn 0x20 key1+key2 tomestones
7313 2 : // Lsn 0x10 key1 in image, key2 in delta
7314 2 : let tline = tenant
7315 2 : .create_test_timeline_with_layers(
7316 2 : TIMELINE_ID,
7317 2 : Lsn(0x10),
7318 2 : DEFAULT_PG_VERSION,
7319 2 : &ctx,
7320 2 : // delta layers
7321 2 : vec![
7322 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7323 2 : Lsn(0x10)..Lsn(0x20),
7324 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
7325 2 : ),
7326 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7327 2 : Lsn(0x20)..Lsn(0x30),
7328 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
7329 2 : ),
7330 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7331 2 : Lsn(0x20)..Lsn(0x30),
7332 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
7333 2 : ),
7334 2 : ],
7335 2 : // image layers
7336 2 : vec![
7337 2 : (Lsn(0x10), vec![(key1, test_img("metadata key 1"))]),
7338 2 : (
7339 2 : Lsn(0x30),
7340 2 : vec![
7341 2 : (key0, test_img("metadata key 0")),
7342 2 : (key3, test_img("metadata key 3")),
7343 2 : ],
7344 2 : ),
7345 2 : ],
7346 2 : Lsn(0x30),
7347 2 : )
7348 40 : .await?;
7349 2 :
7350 2 : let lsn = Lsn(0x30);
7351 2 : let old_lsn = Lsn(0x20);
7352 2 :
7353 2 : assert_eq!(
7354 4 : get_vectored_impl_wrapper(&tline, key0, lsn, &ctx).await?,
7355 2 : Some(test_img("metadata key 0"))
7356 2 : );
7357 2 : assert_eq!(
7358 2 : get_vectored_impl_wrapper(&tline, key1, lsn, &ctx).await?,
7359 2 : None,
7360 2 : );
7361 2 : assert_eq!(
7362 2 : get_vectored_impl_wrapper(&tline, key2, lsn, &ctx).await?,
7363 2 : None,
7364 2 : );
7365 2 : assert_eq!(
7366 8 : get_vectored_impl_wrapper(&tline, key1, old_lsn, &ctx).await?,
7367 2 : Some(Bytes::new()),
7368 2 : );
7369 2 : assert_eq!(
7370 7 : get_vectored_impl_wrapper(&tline, key2, old_lsn, &ctx).await?,
7371 2 : Some(Bytes::new()),
7372 2 : );
7373 2 : assert_eq!(
7374 2 : get_vectored_impl_wrapper(&tline, key3, lsn, &ctx).await?,
7375 2 : Some(test_img("metadata key 3"))
7376 2 : );
7377 2 :
7378 2 : Ok(())
7379 2 : }
7380 :
7381 : #[tokio::test]
7382 2 : async fn test_metadata_tombstone_image_creation() {
7383 2 : let harness = TenantHarness::create("test_metadata_tombstone_image_creation")
7384 2 : .await
7385 2 : .unwrap();
7386 20 : let (tenant, ctx) = harness.load().await;
7387 2 :
7388 2 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
7389 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
7390 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
7391 2 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
7392 2 :
7393 2 : let tline = tenant
7394 2 : .create_test_timeline_with_layers(
7395 2 : TIMELINE_ID,
7396 2 : Lsn(0x10),
7397 2 : DEFAULT_PG_VERSION,
7398 2 : &ctx,
7399 2 : // delta layers
7400 2 : vec![
7401 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7402 2 : Lsn(0x10)..Lsn(0x20),
7403 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
7404 2 : ),
7405 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7406 2 : Lsn(0x20)..Lsn(0x30),
7407 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
7408 2 : ),
7409 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7410 2 : Lsn(0x20)..Lsn(0x30),
7411 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
7412 2 : ),
7413 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7414 2 : Lsn(0x30)..Lsn(0x40),
7415 2 : vec![
7416 2 : (key0, Lsn(0x30), Value::Image(test_img("metadata key 0"))),
7417 2 : (key3, Lsn(0x30), Value::Image(test_img("metadata key 3"))),
7418 2 : ],
7419 2 : ),
7420 2 : ],
7421 2 : // image layers
7422 2 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
7423 2 : Lsn(0x40),
7424 2 : )
7425 37 : .await
7426 2 : .unwrap();
7427 2 :
7428 2 : let cancel = CancellationToken::new();
7429 2 :
7430 2 : tline
7431 2 : .compact(
7432 2 : &cancel,
7433 2 : {
7434 2 : let mut flags = EnumSet::new();
7435 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
7436 2 : flags.insert(CompactFlags::ForceRepartition);
7437 2 : flags
7438 2 : },
7439 2 : &ctx,
7440 2 : )
7441 62 : .await
7442 2 : .unwrap();
7443 2 :
7444 2 : // Image layers are created at last_record_lsn
7445 2 : let images = tline
7446 2 : .inspect_image_layers(Lsn(0x40), &ctx)
7447 8 : .await
7448 2 : .unwrap()
7449 2 : .into_iter()
7450 18 : .filter(|(k, _)| k.is_metadata_key())
7451 2 : .collect::<Vec<_>>();
7452 2 : assert_eq!(images.len(), 2); // the image layer should only contain two existing keys, tombstones should be removed.
7453 2 : }
7454 :
7455 : #[tokio::test]
7456 2 : async fn test_metadata_tombstone_empty_image_creation() {
7457 2 : let harness = TenantHarness::create("test_metadata_tombstone_empty_image_creation")
7458 2 : .await
7459 2 : .unwrap();
7460 20 : let (tenant, ctx) = harness.load().await;
7461 2 :
7462 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
7463 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
7464 2 :
7465 2 : let tline = tenant
7466 2 : .create_test_timeline_with_layers(
7467 2 : TIMELINE_ID,
7468 2 : Lsn(0x10),
7469 2 : DEFAULT_PG_VERSION,
7470 2 : &ctx,
7471 2 : // delta layers
7472 2 : vec![
7473 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7474 2 : Lsn(0x10)..Lsn(0x20),
7475 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
7476 2 : ),
7477 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7478 2 : Lsn(0x20)..Lsn(0x30),
7479 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
7480 2 : ),
7481 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7482 2 : Lsn(0x20)..Lsn(0x30),
7483 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
7484 2 : ),
7485 2 : ],
7486 2 : // image layers
7487 2 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
7488 2 : Lsn(0x30),
7489 2 : )
7490 31 : .await
7491 2 : .unwrap();
7492 2 :
7493 2 : let cancel = CancellationToken::new();
7494 2 :
7495 2 : tline
7496 2 : .compact(
7497 2 : &cancel,
7498 2 : {
7499 2 : let mut flags = EnumSet::new();
7500 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
7501 2 : flags.insert(CompactFlags::ForceRepartition);
7502 2 : flags
7503 2 : },
7504 2 : &ctx,
7505 2 : )
7506 50 : .await
7507 2 : .unwrap();
7508 2 :
7509 2 : // Image layers are created at last_record_lsn
7510 2 : let images = tline
7511 2 : .inspect_image_layers(Lsn(0x30), &ctx)
7512 4 : .await
7513 2 : .unwrap()
7514 2 : .into_iter()
7515 14 : .filter(|(k, _)| k.is_metadata_key())
7516 2 : .collect::<Vec<_>>();
7517 2 : assert_eq!(images.len(), 0); // the image layer should not contain tombstones, or it is not created
7518 2 : }
7519 :
7520 : #[tokio::test]
7521 2 : async fn test_simple_bottom_most_compaction_images() -> anyhow::Result<()> {
7522 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_images").await?;
7523 20 : let (tenant, ctx) = harness.load().await;
7524 2 :
7525 102 : fn get_key(id: u32) -> Key {
7526 102 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7527 102 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7528 102 : key.field6 = id;
7529 102 : key
7530 102 : }
7531 2 :
7532 2 : // We create
7533 2 : // - one bottom-most image layer,
7534 2 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
7535 2 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
7536 2 : // - a delta layer D3 above the horizon.
7537 2 : //
7538 2 : // | D3 |
7539 2 : // | D1 |
7540 2 : // -| |-- gc horizon -----------------
7541 2 : // | | | D2 |
7542 2 : // --------- img layer ------------------
7543 2 : //
7544 2 : // What we should expact from this compaction is:
7545 2 : // | D3 |
7546 2 : // | Part of D1 |
7547 2 : // --------- img layer with D1+D2 at GC horizon------------------
7548 2 :
7549 2 : // img layer at 0x10
7550 2 : let img_layer = (0..10)
7551 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
7552 2 : .collect_vec();
7553 2 :
7554 2 : let delta1 = vec![
7555 2 : (
7556 2 : get_key(1),
7557 2 : Lsn(0x20),
7558 2 : Value::Image(Bytes::from("value 1@0x20")),
7559 2 : ),
7560 2 : (
7561 2 : get_key(2),
7562 2 : Lsn(0x30),
7563 2 : Value::Image(Bytes::from("value 2@0x30")),
7564 2 : ),
7565 2 : (
7566 2 : get_key(3),
7567 2 : Lsn(0x40),
7568 2 : Value::Image(Bytes::from("value 3@0x40")),
7569 2 : ),
7570 2 : ];
7571 2 : let delta2 = vec![
7572 2 : (
7573 2 : get_key(5),
7574 2 : Lsn(0x20),
7575 2 : Value::Image(Bytes::from("value 5@0x20")),
7576 2 : ),
7577 2 : (
7578 2 : get_key(6),
7579 2 : Lsn(0x20),
7580 2 : Value::Image(Bytes::from("value 6@0x20")),
7581 2 : ),
7582 2 : ];
7583 2 : let delta3 = vec![
7584 2 : (
7585 2 : get_key(8),
7586 2 : Lsn(0x48),
7587 2 : Value::Image(Bytes::from("value 8@0x48")),
7588 2 : ),
7589 2 : (
7590 2 : get_key(9),
7591 2 : Lsn(0x48),
7592 2 : Value::Image(Bytes::from("value 9@0x48")),
7593 2 : ),
7594 2 : ];
7595 2 :
7596 2 : let tline = tenant
7597 2 : .create_test_timeline_with_layers(
7598 2 : TIMELINE_ID,
7599 2 : Lsn(0x10),
7600 2 : DEFAULT_PG_VERSION,
7601 2 : &ctx,
7602 2 : vec![
7603 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
7604 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
7605 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
7606 2 : ], // delta layers
7607 2 : vec![(Lsn(0x10), img_layer)], // image layers
7608 2 : Lsn(0x50),
7609 2 : )
7610 49 : .await?;
7611 2 : {
7612 2 : // Update GC info
7613 2 : let mut guard = tline.gc_info.write().unwrap();
7614 2 : guard.cutoffs.time = Lsn(0x30);
7615 2 : guard.cutoffs.space = Lsn(0x30);
7616 2 : }
7617 2 :
7618 2 : let expected_result = [
7619 2 : Bytes::from_static(b"value 0@0x10"),
7620 2 : Bytes::from_static(b"value 1@0x20"),
7621 2 : Bytes::from_static(b"value 2@0x30"),
7622 2 : Bytes::from_static(b"value 3@0x40"),
7623 2 : Bytes::from_static(b"value 4@0x10"),
7624 2 : Bytes::from_static(b"value 5@0x20"),
7625 2 : Bytes::from_static(b"value 6@0x20"),
7626 2 : Bytes::from_static(b"value 7@0x10"),
7627 2 : Bytes::from_static(b"value 8@0x48"),
7628 2 : Bytes::from_static(b"value 9@0x48"),
7629 2 : ];
7630 2 :
7631 20 : for (idx, expected) in expected_result.iter().enumerate() {
7632 20 : assert_eq!(
7633 20 : tline
7634 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7635 30 : .await
7636 20 : .unwrap(),
7637 2 : expected
7638 2 : );
7639 2 : }
7640 2 :
7641 2 : let cancel = CancellationToken::new();
7642 2 : tline
7643 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7644 56 : .await
7645 2 : .unwrap();
7646 2 :
7647 20 : for (idx, expected) in expected_result.iter().enumerate() {
7648 20 : assert_eq!(
7649 20 : tline
7650 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7651 20 : .await
7652 20 : .unwrap(),
7653 2 : expected
7654 2 : );
7655 2 : }
7656 2 :
7657 2 : // Check if the image layer at the GC horizon contains exactly what we want
7658 2 : let image_at_gc_horizon = tline
7659 2 : .inspect_image_layers(Lsn(0x30), &ctx)
7660 2 : .await
7661 2 : .unwrap()
7662 2 : .into_iter()
7663 34 : .filter(|(k, _)| k.is_metadata_key())
7664 2 : .collect::<Vec<_>>();
7665 2 :
7666 2 : assert_eq!(image_at_gc_horizon.len(), 10);
7667 2 : let expected_result = [
7668 2 : Bytes::from_static(b"value 0@0x10"),
7669 2 : Bytes::from_static(b"value 1@0x20"),
7670 2 : Bytes::from_static(b"value 2@0x30"),
7671 2 : Bytes::from_static(b"value 3@0x10"),
7672 2 : Bytes::from_static(b"value 4@0x10"),
7673 2 : Bytes::from_static(b"value 5@0x20"),
7674 2 : Bytes::from_static(b"value 6@0x20"),
7675 2 : Bytes::from_static(b"value 7@0x10"),
7676 2 : Bytes::from_static(b"value 8@0x10"),
7677 2 : Bytes::from_static(b"value 9@0x10"),
7678 2 : ];
7679 22 : for idx in 0..10 {
7680 20 : assert_eq!(
7681 20 : image_at_gc_horizon[idx],
7682 20 : (get_key(idx as u32), expected_result[idx].clone())
7683 20 : );
7684 2 : }
7685 2 :
7686 2 : // Check if old layers are removed / new layers have the expected LSN
7687 2 : let all_layers = inspect_and_sort(&tline, None).await;
7688 2 : assert_eq!(
7689 2 : all_layers,
7690 2 : vec![
7691 2 : // Image layer at GC horizon
7692 2 : PersistentLayerKey {
7693 2 : key_range: Key::MIN..Key::MAX,
7694 2 : lsn_range: Lsn(0x30)..Lsn(0x31),
7695 2 : is_delta: false
7696 2 : },
7697 2 : // The delta layer below the horizon
7698 2 : PersistentLayerKey {
7699 2 : key_range: get_key(3)..get_key(4),
7700 2 : lsn_range: Lsn(0x30)..Lsn(0x48),
7701 2 : is_delta: true
7702 2 : },
7703 2 : // The delta3 layer that should not be picked for the compaction
7704 2 : PersistentLayerKey {
7705 2 : key_range: get_key(8)..get_key(10),
7706 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
7707 2 : is_delta: true
7708 2 : }
7709 2 : ]
7710 2 : );
7711 2 :
7712 2 : // increase GC horizon and compact again
7713 2 : {
7714 2 : // Update GC info
7715 2 : let mut guard = tline.gc_info.write().unwrap();
7716 2 : guard.cutoffs.time = Lsn(0x40);
7717 2 : guard.cutoffs.space = Lsn(0x40);
7718 2 : }
7719 2 : tline
7720 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7721 45 : .await
7722 2 : .unwrap();
7723 2 :
7724 2 : Ok(())
7725 2 : }
7726 :
7727 : #[cfg(feature = "testing")]
7728 : #[tokio::test]
7729 2 : async fn test_neon_test_record() -> anyhow::Result<()> {
7730 2 : let harness = TenantHarness::create("test_neon_test_record").await?;
7731 20 : let (tenant, ctx) = harness.load().await;
7732 2 :
7733 24 : fn get_key(id: u32) -> Key {
7734 24 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7735 24 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7736 24 : key.field6 = id;
7737 24 : key
7738 24 : }
7739 2 :
7740 2 : let delta1 = vec![
7741 2 : (
7742 2 : get_key(1),
7743 2 : Lsn(0x20),
7744 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
7745 2 : ),
7746 2 : (
7747 2 : get_key(1),
7748 2 : Lsn(0x30),
7749 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
7750 2 : ),
7751 2 : (get_key(2), Lsn(0x10), Value::Image("0x10".into())),
7752 2 : (
7753 2 : get_key(2),
7754 2 : Lsn(0x20),
7755 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
7756 2 : ),
7757 2 : (
7758 2 : get_key(2),
7759 2 : Lsn(0x30),
7760 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
7761 2 : ),
7762 2 : (get_key(3), Lsn(0x10), Value::Image("0x10".into())),
7763 2 : (
7764 2 : get_key(3),
7765 2 : Lsn(0x20),
7766 2 : Value::WalRecord(NeonWalRecord::wal_clear("c")),
7767 2 : ),
7768 2 : (get_key(4), Lsn(0x10), Value::Image("0x10".into())),
7769 2 : (
7770 2 : get_key(4),
7771 2 : Lsn(0x20),
7772 2 : Value::WalRecord(NeonWalRecord::wal_init("i")),
7773 2 : ),
7774 2 : ];
7775 2 : let image1 = vec![(get_key(1), "0x10".into())];
7776 2 :
7777 2 : let tline = tenant
7778 2 : .create_test_timeline_with_layers(
7779 2 : TIMELINE_ID,
7780 2 : Lsn(0x10),
7781 2 : DEFAULT_PG_VERSION,
7782 2 : &ctx,
7783 2 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
7784 2 : Lsn(0x10)..Lsn(0x40),
7785 2 : delta1,
7786 2 : )], // delta layers
7787 2 : vec![(Lsn(0x10), image1)], // image layers
7788 2 : Lsn(0x50),
7789 2 : )
7790 19 : .await?;
7791 2 :
7792 2 : assert_eq!(
7793 8 : tline.get(get_key(1), Lsn(0x50), &ctx).await?,
7794 2 : Bytes::from_static(b"0x10,0x20,0x30")
7795 2 : );
7796 2 : assert_eq!(
7797 2 : tline.get(get_key(2), Lsn(0x50), &ctx).await?,
7798 2 : Bytes::from_static(b"0x10,0x20,0x30")
7799 2 : );
7800 2 :
7801 2 : // Need to remove the limit of "Neon WAL redo requires base image".
7802 2 :
7803 2 : // assert_eq!(tline.get(get_key(3), Lsn(0x50), &ctx).await?, Bytes::new());
7804 2 : // assert_eq!(tline.get(get_key(4), Lsn(0x50), &ctx).await?, Bytes::new());
7805 2 :
7806 2 : Ok(())
7807 2 : }
7808 :
7809 : #[tokio::test(start_paused = true)]
7810 2 : async fn test_lsn_lease() -> anyhow::Result<()> {
7811 2 : let (tenant, ctx) = TenantHarness::create("test_lsn_lease")
7812 2 : .await
7813 2 : .unwrap()
7814 2 : .load()
7815 20 : .await;
7816 2 : // Advance to the lsn lease deadline so that GC is not blocked by
7817 2 : // initial transition into AttachedSingle.
7818 2 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
7819 2 : tokio::time::resume();
7820 2 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7821 2 :
7822 2 : let end_lsn = Lsn(0x100);
7823 2 : let image_layers = (0x20..=0x90)
7824 2 : .step_by(0x10)
7825 16 : .map(|n| {
7826 16 : (
7827 16 : Lsn(n),
7828 16 : vec![(key, test_img(&format!("data key at {:x}", n)))],
7829 16 : )
7830 16 : })
7831 2 : .collect();
7832 2 :
7833 2 : let timeline = tenant
7834 2 : .create_test_timeline_with_layers(
7835 2 : TIMELINE_ID,
7836 2 : Lsn(0x10),
7837 2 : DEFAULT_PG_VERSION,
7838 2 : &ctx,
7839 2 : Vec::new(),
7840 2 : image_layers,
7841 2 : end_lsn,
7842 2 : )
7843 62 : .await?;
7844 2 :
7845 2 : let leased_lsns = [0x30, 0x50, 0x70];
7846 2 : let mut leases = Vec::new();
7847 6 : leased_lsns.iter().for_each(|n| {
7848 6 : leases.push(
7849 6 : timeline
7850 6 : .init_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)
7851 6 : .expect("lease request should succeed"),
7852 6 : );
7853 6 : });
7854 2 :
7855 2 : let updated_lease_0 = timeline
7856 2 : .renew_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)
7857 2 : .expect("lease renewal should succeed");
7858 2 : assert_eq!(
7859 2 : updated_lease_0.valid_until, leases[0].valid_until,
7860 2 : " Renewing with shorter lease should not change the lease."
7861 2 : );
7862 2 :
7863 2 : let updated_lease_1 = timeline
7864 2 : .renew_lsn_lease(
7865 2 : Lsn(leased_lsns[1]),
7866 2 : timeline.get_lsn_lease_length() * 2,
7867 2 : &ctx,
7868 2 : )
7869 2 : .expect("lease renewal should succeed");
7870 2 : assert!(
7871 2 : updated_lease_1.valid_until > leases[1].valid_until,
7872 2 : "Renewing with a long lease should renew lease with later expiration time."
7873 2 : );
7874 2 :
7875 2 : // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
7876 2 : info!(
7877 2 : "latest_gc_cutoff_lsn: {}",
7878 0 : *timeline.get_latest_gc_cutoff_lsn()
7879 2 : );
7880 2 : timeline.force_set_disk_consistent_lsn(end_lsn);
7881 2 :
7882 2 : let res = tenant
7883 2 : .gc_iteration(
7884 2 : Some(TIMELINE_ID),
7885 2 : 0,
7886 2 : Duration::ZERO,
7887 2 : &CancellationToken::new(),
7888 2 : &ctx,
7889 2 : )
7890 2 : .await
7891 2 : .unwrap();
7892 2 :
7893 2 : // Keeping everything <= Lsn(0x80) b/c leases:
7894 2 : // 0/10: initdb layer
7895 2 : // (0/20..=0/70).step_by(0x10): image layers added when creating the timeline.
7896 2 : assert_eq!(res.layers_needed_by_leases, 7);
7897 2 : // Keeping 0/90 b/c it is the latest layer.
7898 2 : assert_eq!(res.layers_not_updated, 1);
7899 2 : // Removed 0/80.
7900 2 : assert_eq!(res.layers_removed, 1);
7901 2 :
7902 2 : // Make lease on a already GC-ed LSN.
7903 2 : // 0/80 does not have a valid lease + is below latest_gc_cutoff
7904 2 : assert!(Lsn(0x80) < *timeline.get_latest_gc_cutoff_lsn());
7905 2 : timeline
7906 2 : .init_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx)
7907 2 : .expect_err("lease request on GC-ed LSN should fail");
7908 2 :
7909 2 : // Should still be able to renew a currently valid lease
7910 2 : // Assumption: original lease to is still valid for 0/50.
7911 2 : // (use `Timeline::init_lsn_lease` for testing so it always does validation)
7912 2 : timeline
7913 2 : .init_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)
7914 2 : .expect("lease renewal with validation should succeed");
7915 2 :
7916 2 : Ok(())
7917 2 : }
7918 :
7919 : #[cfg(feature = "testing")]
7920 : #[tokio::test]
7921 2 : async fn test_simple_bottom_most_compaction_deltas_1() -> anyhow::Result<()> {
7922 2 : test_simple_bottom_most_compaction_deltas_helper(
7923 2 : "test_simple_bottom_most_compaction_deltas_1",
7924 2 : false,
7925 2 : )
7926 245 : .await
7927 2 : }
7928 :
7929 : #[cfg(feature = "testing")]
7930 : #[tokio::test]
7931 2 : async fn test_simple_bottom_most_compaction_deltas_2() -> anyhow::Result<()> {
7932 2 : test_simple_bottom_most_compaction_deltas_helper(
7933 2 : "test_simple_bottom_most_compaction_deltas_2",
7934 2 : true,
7935 2 : )
7936 226 : .await
7937 2 : }
7938 :
7939 : #[cfg(feature = "testing")]
7940 4 : async fn test_simple_bottom_most_compaction_deltas_helper(
7941 4 : test_name: &'static str,
7942 4 : use_delta_bottom_layer: bool,
7943 4 : ) -> anyhow::Result<()> {
7944 4 : let harness = TenantHarness::create(test_name).await?;
7945 40 : let (tenant, ctx) = harness.load().await;
7946 :
7947 276 : fn get_key(id: u32) -> Key {
7948 276 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7949 276 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7950 276 : key.field6 = id;
7951 276 : key
7952 276 : }
7953 :
7954 : // We create
7955 : // - one bottom-most image layer,
7956 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
7957 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
7958 : // - a delta layer D3 above the horizon.
7959 : //
7960 : // | D3 |
7961 : // | D1 |
7962 : // -| |-- gc horizon -----------------
7963 : // | | | D2 |
7964 : // --------- img layer ------------------
7965 : //
7966 : // What we should expact from this compaction is:
7967 : // | D3 |
7968 : // | Part of D1 |
7969 : // --------- img layer with D1+D2 at GC horizon------------------
7970 :
7971 : // img layer at 0x10
7972 4 : let img_layer = (0..10)
7973 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
7974 4 : .collect_vec();
7975 4 : // or, delta layer at 0x10 if `use_delta_bottom_layer` is true
7976 4 : let delta4 = (0..10)
7977 40 : .map(|id| {
7978 40 : (
7979 40 : get_key(id),
7980 40 : Lsn(0x08),
7981 40 : Value::WalRecord(NeonWalRecord::wal_init(format!("value {id}@0x10"))),
7982 40 : )
7983 40 : })
7984 4 : .collect_vec();
7985 4 :
7986 4 : let delta1 = vec![
7987 4 : (
7988 4 : get_key(1),
7989 4 : Lsn(0x20),
7990 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7991 4 : ),
7992 4 : (
7993 4 : get_key(2),
7994 4 : Lsn(0x30),
7995 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7996 4 : ),
7997 4 : (
7998 4 : get_key(3),
7999 4 : Lsn(0x28),
8000 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8001 4 : ),
8002 4 : (
8003 4 : get_key(3),
8004 4 : Lsn(0x30),
8005 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8006 4 : ),
8007 4 : (
8008 4 : get_key(3),
8009 4 : Lsn(0x40),
8010 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
8011 4 : ),
8012 4 : ];
8013 4 : let delta2 = vec![
8014 4 : (
8015 4 : get_key(5),
8016 4 : Lsn(0x20),
8017 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8018 4 : ),
8019 4 : (
8020 4 : get_key(6),
8021 4 : Lsn(0x20),
8022 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8023 4 : ),
8024 4 : ];
8025 4 : let delta3 = vec![
8026 4 : (
8027 4 : get_key(8),
8028 4 : Lsn(0x48),
8029 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8030 4 : ),
8031 4 : (
8032 4 : get_key(9),
8033 4 : Lsn(0x48),
8034 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8035 4 : ),
8036 4 : ];
8037 :
8038 4 : let tline = if use_delta_bottom_layer {
8039 2 : tenant
8040 2 : .create_test_timeline_with_layers(
8041 2 : TIMELINE_ID,
8042 2 : Lsn(0x08),
8043 2 : DEFAULT_PG_VERSION,
8044 2 : &ctx,
8045 2 : vec![
8046 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8047 2 : Lsn(0x08)..Lsn(0x10),
8048 2 : delta4,
8049 2 : ),
8050 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8051 2 : Lsn(0x20)..Lsn(0x48),
8052 2 : delta1,
8053 2 : ),
8054 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8055 2 : Lsn(0x20)..Lsn(0x48),
8056 2 : delta2,
8057 2 : ),
8058 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8059 2 : Lsn(0x48)..Lsn(0x50),
8060 2 : delta3,
8061 2 : ),
8062 2 : ], // delta layers
8063 2 : vec![], // image layers
8064 2 : Lsn(0x50),
8065 2 : )
8066 30 : .await?
8067 : } else {
8068 2 : tenant
8069 2 : .create_test_timeline_with_layers(
8070 2 : TIMELINE_ID,
8071 2 : Lsn(0x10),
8072 2 : DEFAULT_PG_VERSION,
8073 2 : &ctx,
8074 2 : vec![
8075 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8076 2 : Lsn(0x10)..Lsn(0x48),
8077 2 : delta1,
8078 2 : ),
8079 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8080 2 : Lsn(0x10)..Lsn(0x48),
8081 2 : delta2,
8082 2 : ),
8083 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8084 2 : Lsn(0x48)..Lsn(0x50),
8085 2 : delta3,
8086 2 : ),
8087 2 : ], // delta layers
8088 2 : vec![(Lsn(0x10), img_layer)], // image layers
8089 2 : Lsn(0x50),
8090 2 : )
8091 49 : .await?
8092 : };
8093 4 : {
8094 4 : // Update GC info
8095 4 : let mut guard = tline.gc_info.write().unwrap();
8096 4 : *guard = GcInfo {
8097 4 : retain_lsns: vec![],
8098 4 : cutoffs: GcCutoffs {
8099 4 : time: Lsn(0x30),
8100 4 : space: Lsn(0x30),
8101 4 : },
8102 4 : leases: Default::default(),
8103 4 : within_ancestor_pitr: false,
8104 4 : };
8105 4 : }
8106 4 :
8107 4 : let expected_result = [
8108 4 : Bytes::from_static(b"value 0@0x10"),
8109 4 : Bytes::from_static(b"value 1@0x10@0x20"),
8110 4 : Bytes::from_static(b"value 2@0x10@0x30"),
8111 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
8112 4 : Bytes::from_static(b"value 4@0x10"),
8113 4 : Bytes::from_static(b"value 5@0x10@0x20"),
8114 4 : Bytes::from_static(b"value 6@0x10@0x20"),
8115 4 : Bytes::from_static(b"value 7@0x10"),
8116 4 : Bytes::from_static(b"value 8@0x10@0x48"),
8117 4 : Bytes::from_static(b"value 9@0x10@0x48"),
8118 4 : ];
8119 4 :
8120 4 : let expected_result_at_gc_horizon = [
8121 4 : Bytes::from_static(b"value 0@0x10"),
8122 4 : Bytes::from_static(b"value 1@0x10@0x20"),
8123 4 : Bytes::from_static(b"value 2@0x10@0x30"),
8124 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
8125 4 : Bytes::from_static(b"value 4@0x10"),
8126 4 : Bytes::from_static(b"value 5@0x10@0x20"),
8127 4 : Bytes::from_static(b"value 6@0x10@0x20"),
8128 4 : Bytes::from_static(b"value 7@0x10"),
8129 4 : Bytes::from_static(b"value 8@0x10"),
8130 4 : Bytes::from_static(b"value 9@0x10"),
8131 4 : ];
8132 :
8133 44 : for idx in 0..10 {
8134 40 : assert_eq!(
8135 40 : tline
8136 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8137 60 : .await
8138 40 : .unwrap(),
8139 40 : &expected_result[idx]
8140 : );
8141 40 : assert_eq!(
8142 40 : tline
8143 40 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
8144 31 : .await
8145 40 : .unwrap(),
8146 40 : &expected_result_at_gc_horizon[idx]
8147 : );
8148 : }
8149 :
8150 4 : let cancel = CancellationToken::new();
8151 4 : tline
8152 4 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8153 112 : .await
8154 4 : .unwrap();
8155 :
8156 44 : for idx in 0..10 {
8157 40 : assert_eq!(
8158 40 : tline
8159 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8160 40 : .await
8161 40 : .unwrap(),
8162 40 : &expected_result[idx]
8163 : );
8164 40 : assert_eq!(
8165 40 : tline
8166 40 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
8167 20 : .await
8168 40 : .unwrap(),
8169 40 : &expected_result_at_gc_horizon[idx]
8170 : );
8171 : }
8172 :
8173 : // increase GC horizon and compact again
8174 4 : {
8175 4 : // Update GC info
8176 4 : let mut guard = tline.gc_info.write().unwrap();
8177 4 : guard.cutoffs.time = Lsn(0x40);
8178 4 : guard.cutoffs.space = Lsn(0x40);
8179 4 : }
8180 4 : tline
8181 4 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8182 89 : .await
8183 4 : .unwrap();
8184 4 :
8185 4 : Ok(())
8186 4 : }
8187 :
8188 : #[cfg(feature = "testing")]
8189 : #[tokio::test]
8190 2 : async fn test_generate_key_retention() -> anyhow::Result<()> {
8191 2 : let harness = TenantHarness::create("test_generate_key_retention").await?;
8192 20 : let (tenant, ctx) = harness.load().await;
8193 2 : let tline = tenant
8194 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
8195 6 : .await?;
8196 2 : tline.force_advance_lsn(Lsn(0x70));
8197 2 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
8198 2 : let history = vec![
8199 2 : (
8200 2 : key,
8201 2 : Lsn(0x10),
8202 2 : Value::WalRecord(NeonWalRecord::wal_init("0x10")),
8203 2 : ),
8204 2 : (
8205 2 : key,
8206 2 : Lsn(0x20),
8207 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8208 2 : ),
8209 2 : (
8210 2 : key,
8211 2 : Lsn(0x30),
8212 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8213 2 : ),
8214 2 : (
8215 2 : key,
8216 2 : Lsn(0x40),
8217 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8218 2 : ),
8219 2 : (
8220 2 : key,
8221 2 : Lsn(0x50),
8222 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
8223 2 : ),
8224 2 : (
8225 2 : key,
8226 2 : Lsn(0x60),
8227 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8228 2 : ),
8229 2 : (
8230 2 : key,
8231 2 : Lsn(0x70),
8232 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8233 2 : ),
8234 2 : (
8235 2 : key,
8236 2 : Lsn(0x80),
8237 2 : Value::Image(Bytes::copy_from_slice(
8238 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8239 2 : )),
8240 2 : ),
8241 2 : (
8242 2 : key,
8243 2 : Lsn(0x90),
8244 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8245 2 : ),
8246 2 : ];
8247 2 : let res = tline
8248 2 : .generate_key_retention(
8249 2 : key,
8250 2 : &history,
8251 2 : Lsn(0x60),
8252 2 : &[Lsn(0x20), Lsn(0x40), Lsn(0x50)],
8253 2 : 3,
8254 2 : None,
8255 2 : )
8256 2 : .await
8257 2 : .unwrap();
8258 2 : let expected_res = KeyHistoryRetention {
8259 2 : below_horizon: vec![
8260 2 : (
8261 2 : Lsn(0x20),
8262 2 : KeyLogAtLsn(vec![(
8263 2 : Lsn(0x20),
8264 2 : Value::Image(Bytes::from_static(b"0x10;0x20")),
8265 2 : )]),
8266 2 : ),
8267 2 : (
8268 2 : Lsn(0x40),
8269 2 : KeyLogAtLsn(vec![
8270 2 : (
8271 2 : Lsn(0x30),
8272 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8273 2 : ),
8274 2 : (
8275 2 : Lsn(0x40),
8276 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8277 2 : ),
8278 2 : ]),
8279 2 : ),
8280 2 : (
8281 2 : Lsn(0x50),
8282 2 : KeyLogAtLsn(vec![(
8283 2 : Lsn(0x50),
8284 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40;0x50")),
8285 2 : )]),
8286 2 : ),
8287 2 : (
8288 2 : Lsn(0x60),
8289 2 : KeyLogAtLsn(vec![(
8290 2 : Lsn(0x60),
8291 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8292 2 : )]),
8293 2 : ),
8294 2 : ],
8295 2 : above_horizon: KeyLogAtLsn(vec![
8296 2 : (
8297 2 : Lsn(0x70),
8298 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8299 2 : ),
8300 2 : (
8301 2 : Lsn(0x80),
8302 2 : Value::Image(Bytes::copy_from_slice(
8303 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8304 2 : )),
8305 2 : ),
8306 2 : (
8307 2 : Lsn(0x90),
8308 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8309 2 : ),
8310 2 : ]),
8311 2 : };
8312 2 : assert_eq!(res, expected_res);
8313 2 :
8314 2 : // We expect GC-compaction to run with the original GC. This would create a situation that
8315 2 : // the original GC algorithm removes some delta layers b/c there are full image coverage,
8316 2 : // therefore causing some keys to have an incomplete history below the lowest retain LSN.
8317 2 : // For example, we have
8318 2 : // ```plain
8319 2 : // init delta @ 0x10, image @ 0x20, delta @ 0x30 (gc_horizon), image @ 0x40.
8320 2 : // ```
8321 2 : // Now the GC horizon moves up, and we have
8322 2 : // ```plain
8323 2 : // init delta @ 0x10, image @ 0x20, delta @ 0x30, image @ 0x40 (gc_horizon)
8324 2 : // ```
8325 2 : // The original GC algorithm kicks in, and removes delta @ 0x10, image @ 0x20.
8326 2 : // We will end up with
8327 2 : // ```plain
8328 2 : // delta @ 0x30, image @ 0x40 (gc_horizon)
8329 2 : // ```
8330 2 : // Now we run the GC-compaction, and this key does not have a full history.
8331 2 : // We should be able to handle this partial history and drop everything before the
8332 2 : // gc_horizon image.
8333 2 :
8334 2 : let history = vec![
8335 2 : (
8336 2 : key,
8337 2 : Lsn(0x20),
8338 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8339 2 : ),
8340 2 : (
8341 2 : key,
8342 2 : Lsn(0x30),
8343 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8344 2 : ),
8345 2 : (
8346 2 : key,
8347 2 : Lsn(0x40),
8348 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
8349 2 : ),
8350 2 : (
8351 2 : key,
8352 2 : Lsn(0x50),
8353 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
8354 2 : ),
8355 2 : (
8356 2 : key,
8357 2 : Lsn(0x60),
8358 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8359 2 : ),
8360 2 : (
8361 2 : key,
8362 2 : Lsn(0x70),
8363 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8364 2 : ),
8365 2 : (
8366 2 : key,
8367 2 : Lsn(0x80),
8368 2 : Value::Image(Bytes::copy_from_slice(
8369 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8370 2 : )),
8371 2 : ),
8372 2 : (
8373 2 : key,
8374 2 : Lsn(0x90),
8375 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8376 2 : ),
8377 2 : ];
8378 2 : let res = tline
8379 2 : .generate_key_retention(key, &history, Lsn(0x60), &[Lsn(0x40), Lsn(0x50)], 3, None)
8380 2 : .await
8381 2 : .unwrap();
8382 2 : let expected_res = KeyHistoryRetention {
8383 2 : below_horizon: vec![
8384 2 : (
8385 2 : Lsn(0x40),
8386 2 : KeyLogAtLsn(vec![(
8387 2 : Lsn(0x40),
8388 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
8389 2 : )]),
8390 2 : ),
8391 2 : (
8392 2 : Lsn(0x50),
8393 2 : KeyLogAtLsn(vec![(
8394 2 : Lsn(0x50),
8395 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
8396 2 : )]),
8397 2 : ),
8398 2 : (
8399 2 : Lsn(0x60),
8400 2 : KeyLogAtLsn(vec![(
8401 2 : Lsn(0x60),
8402 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8403 2 : )]),
8404 2 : ),
8405 2 : ],
8406 2 : above_horizon: KeyLogAtLsn(vec![
8407 2 : (
8408 2 : Lsn(0x70),
8409 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8410 2 : ),
8411 2 : (
8412 2 : Lsn(0x80),
8413 2 : Value::Image(Bytes::copy_from_slice(
8414 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8415 2 : )),
8416 2 : ),
8417 2 : (
8418 2 : Lsn(0x90),
8419 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8420 2 : ),
8421 2 : ]),
8422 2 : };
8423 2 : assert_eq!(res, expected_res);
8424 2 :
8425 2 : // In case of branch compaction, the branch itself does not have the full history, and we need to provide
8426 2 : // the ancestor image in the test case.
8427 2 :
8428 2 : let history = vec![
8429 2 : (
8430 2 : key,
8431 2 : Lsn(0x20),
8432 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8433 2 : ),
8434 2 : (
8435 2 : key,
8436 2 : Lsn(0x30),
8437 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8438 2 : ),
8439 2 : (
8440 2 : key,
8441 2 : Lsn(0x40),
8442 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8443 2 : ),
8444 2 : (
8445 2 : key,
8446 2 : Lsn(0x70),
8447 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8448 2 : ),
8449 2 : ];
8450 2 : let res = tline
8451 2 : .generate_key_retention(
8452 2 : key,
8453 2 : &history,
8454 2 : Lsn(0x60),
8455 2 : &[],
8456 2 : 3,
8457 2 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
8458 2 : )
8459 2 : .await
8460 2 : .unwrap();
8461 2 : let expected_res = KeyHistoryRetention {
8462 2 : below_horizon: vec![(
8463 2 : Lsn(0x60),
8464 2 : KeyLogAtLsn(vec![(
8465 2 : Lsn(0x60),
8466 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")), // use the ancestor image to reconstruct the page
8467 2 : )]),
8468 2 : )],
8469 2 : above_horizon: KeyLogAtLsn(vec![(
8470 2 : Lsn(0x70),
8471 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8472 2 : )]),
8473 2 : };
8474 2 : assert_eq!(res, expected_res);
8475 2 :
8476 2 : let history = vec![
8477 2 : (
8478 2 : key,
8479 2 : Lsn(0x20),
8480 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8481 2 : ),
8482 2 : (
8483 2 : key,
8484 2 : Lsn(0x40),
8485 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8486 2 : ),
8487 2 : (
8488 2 : key,
8489 2 : Lsn(0x60),
8490 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8491 2 : ),
8492 2 : (
8493 2 : key,
8494 2 : Lsn(0x70),
8495 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8496 2 : ),
8497 2 : ];
8498 2 : let res = tline
8499 2 : .generate_key_retention(
8500 2 : key,
8501 2 : &history,
8502 2 : Lsn(0x60),
8503 2 : &[Lsn(0x30)],
8504 2 : 3,
8505 2 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
8506 2 : )
8507 2 : .await
8508 2 : .unwrap();
8509 2 : let expected_res = KeyHistoryRetention {
8510 2 : below_horizon: vec![
8511 2 : (
8512 2 : Lsn(0x30),
8513 2 : KeyLogAtLsn(vec![(
8514 2 : Lsn(0x20),
8515 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8516 2 : )]),
8517 2 : ),
8518 2 : (
8519 2 : Lsn(0x60),
8520 2 : KeyLogAtLsn(vec![(
8521 2 : Lsn(0x60),
8522 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x40;0x60")),
8523 2 : )]),
8524 2 : ),
8525 2 : ],
8526 2 : above_horizon: KeyLogAtLsn(vec![(
8527 2 : Lsn(0x70),
8528 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8529 2 : )]),
8530 2 : };
8531 2 : assert_eq!(res, expected_res);
8532 2 :
8533 2 : Ok(())
8534 2 : }
8535 :
8536 : #[cfg(feature = "testing")]
8537 : #[tokio::test]
8538 2 : async fn test_simple_bottom_most_compaction_with_retain_lsns() -> anyhow::Result<()> {
8539 2 : let harness =
8540 2 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns").await?;
8541 20 : let (tenant, ctx) = harness.load().await;
8542 2 :
8543 518 : fn get_key(id: u32) -> Key {
8544 518 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8545 518 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8546 518 : key.field6 = id;
8547 518 : key
8548 518 : }
8549 2 :
8550 2 : let img_layer = (0..10)
8551 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8552 2 : .collect_vec();
8553 2 :
8554 2 : let delta1 = vec![
8555 2 : (
8556 2 : get_key(1),
8557 2 : Lsn(0x20),
8558 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8559 2 : ),
8560 2 : (
8561 2 : get_key(2),
8562 2 : Lsn(0x30),
8563 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8564 2 : ),
8565 2 : (
8566 2 : get_key(3),
8567 2 : Lsn(0x28),
8568 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8569 2 : ),
8570 2 : (
8571 2 : get_key(3),
8572 2 : Lsn(0x30),
8573 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8574 2 : ),
8575 2 : (
8576 2 : get_key(3),
8577 2 : Lsn(0x40),
8578 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
8579 2 : ),
8580 2 : ];
8581 2 : let delta2 = vec![
8582 2 : (
8583 2 : get_key(5),
8584 2 : Lsn(0x20),
8585 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8586 2 : ),
8587 2 : (
8588 2 : get_key(6),
8589 2 : Lsn(0x20),
8590 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8591 2 : ),
8592 2 : ];
8593 2 : let delta3 = vec![
8594 2 : (
8595 2 : get_key(8),
8596 2 : Lsn(0x48),
8597 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8598 2 : ),
8599 2 : (
8600 2 : get_key(9),
8601 2 : Lsn(0x48),
8602 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8603 2 : ),
8604 2 : ];
8605 2 :
8606 2 : let tline = tenant
8607 2 : .create_test_timeline_with_layers(
8608 2 : TIMELINE_ID,
8609 2 : Lsn(0x10),
8610 2 : DEFAULT_PG_VERSION,
8611 2 : &ctx,
8612 2 : vec![
8613 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
8614 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
8615 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
8616 2 : ], // delta layers
8617 2 : vec![(Lsn(0x10), img_layer)], // image layers
8618 2 : Lsn(0x50),
8619 2 : )
8620 49 : .await?;
8621 2 : {
8622 2 : // Update GC info
8623 2 : let mut guard = tline.gc_info.write().unwrap();
8624 2 : *guard = GcInfo {
8625 2 : retain_lsns: vec![
8626 2 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
8627 2 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
8628 2 : ],
8629 2 : cutoffs: GcCutoffs {
8630 2 : time: Lsn(0x30),
8631 2 : space: Lsn(0x30),
8632 2 : },
8633 2 : leases: Default::default(),
8634 2 : within_ancestor_pitr: false,
8635 2 : };
8636 2 : }
8637 2 :
8638 2 : let expected_result = [
8639 2 : Bytes::from_static(b"value 0@0x10"),
8640 2 : Bytes::from_static(b"value 1@0x10@0x20"),
8641 2 : Bytes::from_static(b"value 2@0x10@0x30"),
8642 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
8643 2 : Bytes::from_static(b"value 4@0x10"),
8644 2 : Bytes::from_static(b"value 5@0x10@0x20"),
8645 2 : Bytes::from_static(b"value 6@0x10@0x20"),
8646 2 : Bytes::from_static(b"value 7@0x10"),
8647 2 : Bytes::from_static(b"value 8@0x10@0x48"),
8648 2 : Bytes::from_static(b"value 9@0x10@0x48"),
8649 2 : ];
8650 2 :
8651 2 : let expected_result_at_gc_horizon = [
8652 2 : Bytes::from_static(b"value 0@0x10"),
8653 2 : Bytes::from_static(b"value 1@0x10@0x20"),
8654 2 : Bytes::from_static(b"value 2@0x10@0x30"),
8655 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
8656 2 : Bytes::from_static(b"value 4@0x10"),
8657 2 : Bytes::from_static(b"value 5@0x10@0x20"),
8658 2 : Bytes::from_static(b"value 6@0x10@0x20"),
8659 2 : Bytes::from_static(b"value 7@0x10"),
8660 2 : Bytes::from_static(b"value 8@0x10"),
8661 2 : Bytes::from_static(b"value 9@0x10"),
8662 2 : ];
8663 2 :
8664 2 : let expected_result_at_lsn_20 = [
8665 2 : Bytes::from_static(b"value 0@0x10"),
8666 2 : Bytes::from_static(b"value 1@0x10@0x20"),
8667 2 : Bytes::from_static(b"value 2@0x10"),
8668 2 : Bytes::from_static(b"value 3@0x10"),
8669 2 : Bytes::from_static(b"value 4@0x10"),
8670 2 : Bytes::from_static(b"value 5@0x10@0x20"),
8671 2 : Bytes::from_static(b"value 6@0x10@0x20"),
8672 2 : Bytes::from_static(b"value 7@0x10"),
8673 2 : Bytes::from_static(b"value 8@0x10"),
8674 2 : Bytes::from_static(b"value 9@0x10"),
8675 2 : ];
8676 2 :
8677 2 : let expected_result_at_lsn_10 = [
8678 2 : Bytes::from_static(b"value 0@0x10"),
8679 2 : Bytes::from_static(b"value 1@0x10"),
8680 2 : Bytes::from_static(b"value 2@0x10"),
8681 2 : Bytes::from_static(b"value 3@0x10"),
8682 2 : Bytes::from_static(b"value 4@0x10"),
8683 2 : Bytes::from_static(b"value 5@0x10"),
8684 2 : Bytes::from_static(b"value 6@0x10"),
8685 2 : Bytes::from_static(b"value 7@0x10"),
8686 2 : Bytes::from_static(b"value 8@0x10"),
8687 2 : Bytes::from_static(b"value 9@0x10"),
8688 2 : ];
8689 2 :
8690 12 : let verify_result = || async {
8691 12 : let gc_horizon = {
8692 12 : let gc_info = tline.gc_info.read().unwrap();
8693 12 : gc_info.cutoffs.time
8694 2 : };
8695 132 : for idx in 0..10 {
8696 120 : assert_eq!(
8697 120 : tline
8698 120 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8699 121 : .await
8700 120 : .unwrap(),
8701 120 : &expected_result[idx]
8702 2 : );
8703 120 : assert_eq!(
8704 120 : tline
8705 120 : .get(get_key(idx as u32), gc_horizon, &ctx)
8706 93 : .await
8707 120 : .unwrap(),
8708 120 : &expected_result_at_gc_horizon[idx]
8709 2 : );
8710 120 : assert_eq!(
8711 120 : tline
8712 120 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
8713 86 : .await
8714 120 : .unwrap(),
8715 120 : &expected_result_at_lsn_20[idx]
8716 2 : );
8717 120 : assert_eq!(
8718 120 : tline
8719 120 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
8720 61 : .await
8721 120 : .unwrap(),
8722 120 : &expected_result_at_lsn_10[idx]
8723 2 : );
8724 2 : }
8725 24 : };
8726 2 :
8727 69 : verify_result().await;
8728 2 :
8729 2 : let cancel = CancellationToken::new();
8730 2 : let mut dryrun_flags = EnumSet::new();
8731 2 : dryrun_flags.insert(CompactFlags::DryRun);
8732 2 :
8733 2 : tline
8734 2 : .compact_with_gc(&cancel, dryrun_flags, &ctx)
8735 52 : .await
8736 2 : .unwrap();
8737 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
8738 2 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
8739 57 : verify_result().await;
8740 2 :
8741 2 : tline
8742 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8743 53 : .await
8744 2 : .unwrap();
8745 64 : verify_result().await;
8746 2 :
8747 2 : // compact again
8748 2 : tline
8749 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8750 42 : .await
8751 2 : .unwrap();
8752 57 : verify_result().await;
8753 2 :
8754 2 : // increase GC horizon and compact again
8755 2 : {
8756 2 : // Update GC info
8757 2 : let mut guard = tline.gc_info.write().unwrap();
8758 2 : guard.cutoffs.time = Lsn(0x38);
8759 2 : guard.cutoffs.space = Lsn(0x38);
8760 2 : }
8761 2 : tline
8762 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8763 40 : .await
8764 2 : .unwrap();
8765 57 : verify_result().await; // no wals between 0x30 and 0x38, so we should obtain the same result
8766 2 :
8767 2 : // not increasing the GC horizon and compact again
8768 2 : tline
8769 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8770 42 : .await
8771 2 : .unwrap();
8772 57 : verify_result().await;
8773 2 :
8774 2 : Ok(())
8775 2 : }
8776 :
8777 : #[cfg(feature = "testing")]
8778 : #[tokio::test]
8779 2 : async fn test_simple_bottom_most_compaction_with_retain_lsns_single_key() -> anyhow::Result<()>
8780 2 : {
8781 2 : let harness =
8782 2 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns_single_key")
8783 2 : .await?;
8784 20 : let (tenant, ctx) = harness.load().await;
8785 2 :
8786 352 : fn get_key(id: u32) -> Key {
8787 352 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8788 352 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8789 352 : key.field6 = id;
8790 352 : key
8791 352 : }
8792 2 :
8793 2 : let img_layer = (0..10)
8794 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8795 2 : .collect_vec();
8796 2 :
8797 2 : let delta1 = vec![
8798 2 : (
8799 2 : get_key(1),
8800 2 : Lsn(0x20),
8801 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8802 2 : ),
8803 2 : (
8804 2 : get_key(1),
8805 2 : Lsn(0x28),
8806 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8807 2 : ),
8808 2 : ];
8809 2 : let delta2 = vec![
8810 2 : (
8811 2 : get_key(1),
8812 2 : Lsn(0x30),
8813 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8814 2 : ),
8815 2 : (
8816 2 : get_key(1),
8817 2 : Lsn(0x38),
8818 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
8819 2 : ),
8820 2 : ];
8821 2 : let delta3 = vec![
8822 2 : (
8823 2 : get_key(8),
8824 2 : Lsn(0x48),
8825 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8826 2 : ),
8827 2 : (
8828 2 : get_key(9),
8829 2 : Lsn(0x48),
8830 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8831 2 : ),
8832 2 : ];
8833 2 :
8834 2 : let tline = tenant
8835 2 : .create_test_timeline_with_layers(
8836 2 : TIMELINE_ID,
8837 2 : Lsn(0x10),
8838 2 : DEFAULT_PG_VERSION,
8839 2 : &ctx,
8840 2 : vec![
8841 2 : // delta1 and delta 2 only contain a single key but multiple updates
8842 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x30), delta1),
8843 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
8844 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x50), delta3),
8845 2 : ], // delta layers
8846 2 : vec![(Lsn(0x10), img_layer)], // image layers
8847 2 : Lsn(0x50),
8848 2 : )
8849 49 : .await?;
8850 2 : {
8851 2 : // Update GC info
8852 2 : let mut guard = tline.gc_info.write().unwrap();
8853 2 : *guard = GcInfo {
8854 2 : retain_lsns: vec![
8855 2 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
8856 2 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
8857 2 : ],
8858 2 : cutoffs: GcCutoffs {
8859 2 : time: Lsn(0x30),
8860 2 : space: Lsn(0x30),
8861 2 : },
8862 2 : leases: Default::default(),
8863 2 : within_ancestor_pitr: false,
8864 2 : };
8865 2 : }
8866 2 :
8867 2 : let expected_result = [
8868 2 : Bytes::from_static(b"value 0@0x10"),
8869 2 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
8870 2 : Bytes::from_static(b"value 2@0x10"),
8871 2 : Bytes::from_static(b"value 3@0x10"),
8872 2 : Bytes::from_static(b"value 4@0x10"),
8873 2 : Bytes::from_static(b"value 5@0x10"),
8874 2 : Bytes::from_static(b"value 6@0x10"),
8875 2 : Bytes::from_static(b"value 7@0x10"),
8876 2 : Bytes::from_static(b"value 8@0x10@0x48"),
8877 2 : Bytes::from_static(b"value 9@0x10@0x48"),
8878 2 : ];
8879 2 :
8880 2 : let expected_result_at_gc_horizon = [
8881 2 : Bytes::from_static(b"value 0@0x10"),
8882 2 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
8883 2 : Bytes::from_static(b"value 2@0x10"),
8884 2 : Bytes::from_static(b"value 3@0x10"),
8885 2 : Bytes::from_static(b"value 4@0x10"),
8886 2 : Bytes::from_static(b"value 5@0x10"),
8887 2 : Bytes::from_static(b"value 6@0x10"),
8888 2 : Bytes::from_static(b"value 7@0x10"),
8889 2 : Bytes::from_static(b"value 8@0x10"),
8890 2 : Bytes::from_static(b"value 9@0x10"),
8891 2 : ];
8892 2 :
8893 2 : let expected_result_at_lsn_20 = [
8894 2 : Bytes::from_static(b"value 0@0x10"),
8895 2 : Bytes::from_static(b"value 1@0x10@0x20"),
8896 2 : Bytes::from_static(b"value 2@0x10"),
8897 2 : Bytes::from_static(b"value 3@0x10"),
8898 2 : Bytes::from_static(b"value 4@0x10"),
8899 2 : Bytes::from_static(b"value 5@0x10"),
8900 2 : Bytes::from_static(b"value 6@0x10"),
8901 2 : Bytes::from_static(b"value 7@0x10"),
8902 2 : Bytes::from_static(b"value 8@0x10"),
8903 2 : Bytes::from_static(b"value 9@0x10"),
8904 2 : ];
8905 2 :
8906 2 : let expected_result_at_lsn_10 = [
8907 2 : Bytes::from_static(b"value 0@0x10"),
8908 2 : Bytes::from_static(b"value 1@0x10"),
8909 2 : Bytes::from_static(b"value 2@0x10"),
8910 2 : Bytes::from_static(b"value 3@0x10"),
8911 2 : Bytes::from_static(b"value 4@0x10"),
8912 2 : Bytes::from_static(b"value 5@0x10"),
8913 2 : Bytes::from_static(b"value 6@0x10"),
8914 2 : Bytes::from_static(b"value 7@0x10"),
8915 2 : Bytes::from_static(b"value 8@0x10"),
8916 2 : Bytes::from_static(b"value 9@0x10"),
8917 2 : ];
8918 2 :
8919 8 : let verify_result = || async {
8920 8 : let gc_horizon = {
8921 8 : let gc_info = tline.gc_info.read().unwrap();
8922 8 : gc_info.cutoffs.time
8923 2 : };
8924 88 : for idx in 0..10 {
8925 80 : assert_eq!(
8926 80 : tline
8927 80 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8928 76 : .await
8929 80 : .unwrap(),
8930 80 : &expected_result[idx]
8931 2 : );
8932 80 : assert_eq!(
8933 80 : tline
8934 80 : .get(get_key(idx as u32), gc_horizon, &ctx)
8935 46 : .await
8936 80 : .unwrap(),
8937 80 : &expected_result_at_gc_horizon[idx]
8938 2 : );
8939 80 : assert_eq!(
8940 80 : tline
8941 80 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
8942 48 : .await
8943 80 : .unwrap(),
8944 80 : &expected_result_at_lsn_20[idx]
8945 2 : );
8946 80 : assert_eq!(
8947 80 : tline
8948 80 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
8949 41 : .await
8950 80 : .unwrap(),
8951 80 : &expected_result_at_lsn_10[idx]
8952 2 : );
8953 2 : }
8954 16 : };
8955 2 :
8956 61 : verify_result().await;
8957 2 :
8958 2 : let cancel = CancellationToken::new();
8959 2 : let mut dryrun_flags = EnumSet::new();
8960 2 : dryrun_flags.insert(CompactFlags::DryRun);
8961 2 :
8962 2 : tline
8963 2 : .compact_with_gc(&cancel, dryrun_flags, &ctx)
8964 49 : .await
8965 2 : .unwrap();
8966 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
8967 2 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
8968 49 : verify_result().await;
8969 2 :
8970 2 : tline
8971 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8972 55 : .await
8973 2 : .unwrap();
8974 54 : verify_result().await;
8975 2 :
8976 2 : // compact again
8977 2 : tline
8978 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8979 41 : .await
8980 2 : .unwrap();
8981 47 : verify_result().await;
8982 2 :
8983 2 : Ok(())
8984 2 : }
8985 :
8986 : #[cfg(feature = "testing")]
8987 : #[tokio::test]
8988 2 : async fn test_simple_bottom_most_compaction_on_branch() -> anyhow::Result<()> {
8989 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_on_branch").await?;
8990 20 : let (tenant, ctx) = harness.load().await;
8991 2 :
8992 126 : fn get_key(id: u32) -> Key {
8993 126 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
8994 126 : key.field6 = id;
8995 126 : key
8996 126 : }
8997 2 :
8998 2 : let img_layer = (0..10)
8999 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9000 2 : .collect_vec();
9001 2 :
9002 2 : let delta1 = vec![
9003 2 : (
9004 2 : get_key(1),
9005 2 : Lsn(0x20),
9006 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9007 2 : ),
9008 2 : (
9009 2 : get_key(2),
9010 2 : Lsn(0x30),
9011 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9012 2 : ),
9013 2 : (
9014 2 : get_key(3),
9015 2 : Lsn(0x28),
9016 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9017 2 : ),
9018 2 : (
9019 2 : get_key(3),
9020 2 : Lsn(0x30),
9021 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9022 2 : ),
9023 2 : (
9024 2 : get_key(3),
9025 2 : Lsn(0x40),
9026 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9027 2 : ),
9028 2 : ];
9029 2 : let delta2 = vec![
9030 2 : (
9031 2 : get_key(5),
9032 2 : Lsn(0x20),
9033 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9034 2 : ),
9035 2 : (
9036 2 : get_key(6),
9037 2 : Lsn(0x20),
9038 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9039 2 : ),
9040 2 : ];
9041 2 : let delta3 = vec![
9042 2 : (
9043 2 : get_key(8),
9044 2 : Lsn(0x48),
9045 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9046 2 : ),
9047 2 : (
9048 2 : get_key(9),
9049 2 : Lsn(0x48),
9050 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9051 2 : ),
9052 2 : ];
9053 2 :
9054 2 : let parent_tline = tenant
9055 2 : .create_test_timeline_with_layers(
9056 2 : TIMELINE_ID,
9057 2 : Lsn(0x10),
9058 2 : DEFAULT_PG_VERSION,
9059 2 : &ctx,
9060 2 : vec![], // delta layers
9061 2 : vec![(Lsn(0x18), img_layer)], // image layers
9062 2 : Lsn(0x18),
9063 2 : )
9064 31 : .await?;
9065 2 :
9066 2 : parent_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
9067 2 :
9068 2 : let branch_tline = tenant
9069 2 : .branch_timeline_test_with_layers(
9070 2 : &parent_tline,
9071 2 : NEW_TIMELINE_ID,
9072 2 : Some(Lsn(0x18)),
9073 2 : &ctx,
9074 2 : vec![
9075 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
9076 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
9077 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
9078 2 : ], // delta layers
9079 2 : vec![], // image layers
9080 2 : Lsn(0x50),
9081 2 : )
9082 20 : .await?;
9083 2 :
9084 2 : branch_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
9085 2 :
9086 2 : {
9087 2 : // Update GC info
9088 2 : let mut guard = parent_tline.gc_info.write().unwrap();
9089 2 : *guard = GcInfo {
9090 2 : retain_lsns: vec![(Lsn(0x18), branch_tline.timeline_id, MaybeOffloaded::No)],
9091 2 : cutoffs: GcCutoffs {
9092 2 : time: Lsn(0x10),
9093 2 : space: Lsn(0x10),
9094 2 : },
9095 2 : leases: Default::default(),
9096 2 : within_ancestor_pitr: false,
9097 2 : };
9098 2 : }
9099 2 :
9100 2 : {
9101 2 : // Update GC info
9102 2 : let mut guard = branch_tline.gc_info.write().unwrap();
9103 2 : *guard = GcInfo {
9104 2 : retain_lsns: vec![(Lsn(0x40), branch_tline.timeline_id, MaybeOffloaded::No)],
9105 2 : cutoffs: GcCutoffs {
9106 2 : time: Lsn(0x50),
9107 2 : space: Lsn(0x50),
9108 2 : },
9109 2 : leases: Default::default(),
9110 2 : within_ancestor_pitr: false,
9111 2 : };
9112 2 : }
9113 2 :
9114 2 : let expected_result_at_gc_horizon = [
9115 2 : Bytes::from_static(b"value 0@0x10"),
9116 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9117 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9118 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9119 2 : Bytes::from_static(b"value 4@0x10"),
9120 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9121 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9122 2 : Bytes::from_static(b"value 7@0x10"),
9123 2 : Bytes::from_static(b"value 8@0x10@0x48"),
9124 2 : Bytes::from_static(b"value 9@0x10@0x48"),
9125 2 : ];
9126 2 :
9127 2 : let expected_result_at_lsn_40 = [
9128 2 : Bytes::from_static(b"value 0@0x10"),
9129 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9130 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9131 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9132 2 : Bytes::from_static(b"value 4@0x10"),
9133 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9134 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9135 2 : Bytes::from_static(b"value 7@0x10"),
9136 2 : Bytes::from_static(b"value 8@0x10"),
9137 2 : Bytes::from_static(b"value 9@0x10"),
9138 2 : ];
9139 2 :
9140 4 : let verify_result = || async {
9141 44 : for idx in 0..10 {
9142 40 : assert_eq!(
9143 40 : branch_tline
9144 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9145 50 : .await
9146 40 : .unwrap(),
9147 40 : &expected_result_at_gc_horizon[idx]
9148 2 : );
9149 40 : assert_eq!(
9150 40 : branch_tline
9151 40 : .get(get_key(idx as u32), Lsn(0x40), &ctx)
9152 31 : .await
9153 40 : .unwrap(),
9154 40 : &expected_result_at_lsn_40[idx]
9155 2 : );
9156 2 : }
9157 8 : };
9158 2 :
9159 45 : verify_result().await;
9160 2 :
9161 2 : let cancel = CancellationToken::new();
9162 2 : branch_tline
9163 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
9164 19 : .await
9165 2 : .unwrap();
9166 2 :
9167 36 : verify_result().await;
9168 2 :
9169 2 : Ok(())
9170 2 : }
9171 :
9172 : // Regression test for https://github.com/neondatabase/neon/issues/9012
9173 : // Create an image arrangement where we have to read at different LSN ranges
9174 : // from a delta layer. This is achieved by overlapping an image layer on top of
9175 : // a delta layer. Like so:
9176 : //
9177 : // A B
9178 : // +----------------+ -> delta_layer
9179 : // | | ^ lsn
9180 : // | =========|-> nested_image_layer |
9181 : // | C | |
9182 : // +----------------+ |
9183 : // ======== -> baseline_image_layer +-------> key
9184 : //
9185 : //
9186 : // When querying the key range [A, B) we need to read at different LSN ranges
9187 : // for [A, C) and [C, B). This test checks that the described edge case is handled correctly.
9188 : #[cfg(feature = "testing")]
9189 : #[tokio::test]
9190 2 : async fn test_vectored_read_with_nested_image_layer() -> anyhow::Result<()> {
9191 2 : let harness = TenantHarness::create("test_vectored_read_with_nested_image_layer").await?;
9192 20 : let (tenant, ctx) = harness.load().await;
9193 2 :
9194 2 : let will_init_keys = [2, 6];
9195 44 : fn get_key(id: u32) -> Key {
9196 44 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
9197 44 : key.field6 = id;
9198 44 : key
9199 44 : }
9200 2 :
9201 2 : let mut expected_key_values = HashMap::new();
9202 2 :
9203 2 : let baseline_image_layer_lsn = Lsn(0x10);
9204 2 : let mut baseline_img_layer = Vec::new();
9205 12 : for i in 0..5 {
9206 10 : let key = get_key(i);
9207 10 : let value = format!("value {i}@{baseline_image_layer_lsn}");
9208 10 :
9209 10 : let removed = expected_key_values.insert(key, value.clone());
9210 10 : assert!(removed.is_none());
9211 2 :
9212 10 : baseline_img_layer.push((key, Bytes::from(value)));
9213 2 : }
9214 2 :
9215 2 : let nested_image_layer_lsn = Lsn(0x50);
9216 2 : let mut nested_img_layer = Vec::new();
9217 12 : for i in 5..10 {
9218 10 : let key = get_key(i);
9219 10 : let value = format!("value {i}@{nested_image_layer_lsn}");
9220 10 :
9221 10 : let removed = expected_key_values.insert(key, value.clone());
9222 10 : assert!(removed.is_none());
9223 2 :
9224 10 : nested_img_layer.push((key, Bytes::from(value)));
9225 2 : }
9226 2 :
9227 2 : let mut delta_layer_spec = Vec::default();
9228 2 : let delta_layer_start_lsn = Lsn(0x20);
9229 2 : let mut delta_layer_end_lsn = delta_layer_start_lsn;
9230 2 :
9231 22 : for i in 0..10 {
9232 20 : let key = get_key(i);
9233 20 : let key_in_nested = nested_img_layer
9234 20 : .iter()
9235 80 : .any(|(key_with_img, _)| *key_with_img == key);
9236 20 : let lsn = {
9237 20 : if key_in_nested {
9238 10 : Lsn(nested_image_layer_lsn.0 + 0x10)
9239 2 : } else {
9240 10 : delta_layer_start_lsn
9241 2 : }
9242 2 : };
9243 2 :
9244 20 : let will_init = will_init_keys.contains(&i);
9245 20 : if will_init {
9246 4 : delta_layer_spec.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init(""))));
9247 4 :
9248 4 : expected_key_values.insert(key, "".to_string());
9249 16 : } else {
9250 16 : let delta = format!("@{lsn}");
9251 16 : delta_layer_spec.push((
9252 16 : key,
9253 16 : lsn,
9254 16 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
9255 16 : ));
9256 16 :
9257 16 : expected_key_values
9258 16 : .get_mut(&key)
9259 16 : .expect("An image exists for each key")
9260 16 : .push_str(delta.as_str());
9261 16 : }
9262 20 : delta_layer_end_lsn = std::cmp::max(delta_layer_start_lsn, lsn);
9263 2 : }
9264 2 :
9265 2 : delta_layer_end_lsn = Lsn(delta_layer_end_lsn.0 + 1);
9266 2 :
9267 2 : assert!(
9268 2 : nested_image_layer_lsn > delta_layer_start_lsn
9269 2 : && nested_image_layer_lsn < delta_layer_end_lsn
9270 2 : );
9271 2 :
9272 2 : let tline = tenant
9273 2 : .create_test_timeline_with_layers(
9274 2 : TIMELINE_ID,
9275 2 : baseline_image_layer_lsn,
9276 2 : DEFAULT_PG_VERSION,
9277 2 : &ctx,
9278 2 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
9279 2 : delta_layer_start_lsn..delta_layer_end_lsn,
9280 2 : delta_layer_spec,
9281 2 : )], // delta layers
9282 2 : vec![
9283 2 : (baseline_image_layer_lsn, baseline_img_layer),
9284 2 : (nested_image_layer_lsn, nested_img_layer),
9285 2 : ], // image layers
9286 2 : delta_layer_end_lsn,
9287 2 : )
9288 42 : .await?;
9289 2 :
9290 2 : let keyspace = KeySpace::single(get_key(0)..get_key(10));
9291 2 : let results = tline
9292 2 : .get_vectored(keyspace, delta_layer_end_lsn, &ctx)
9293 13 : .await
9294 2 : .expect("No vectored errors");
9295 22 : for (key, res) in results {
9296 20 : let value = res.expect("No key errors");
9297 20 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
9298 20 : assert_eq!(value, Bytes::from(expected_value));
9299 2 : }
9300 2 :
9301 2 : Ok(())
9302 2 : }
9303 :
9304 142 : fn sort_layer_key(k1: &PersistentLayerKey, k2: &PersistentLayerKey) -> std::cmp::Ordering {
9305 142 : (
9306 142 : k1.is_delta,
9307 142 : k1.key_range.start,
9308 142 : k1.key_range.end,
9309 142 : k1.lsn_range.start,
9310 142 : k1.lsn_range.end,
9311 142 : )
9312 142 : .cmp(&(
9313 142 : k2.is_delta,
9314 142 : k2.key_range.start,
9315 142 : k2.key_range.end,
9316 142 : k2.lsn_range.start,
9317 142 : k2.lsn_range.end,
9318 142 : ))
9319 142 : }
9320 :
9321 12 : async fn inspect_and_sort(
9322 12 : tline: &Arc<Timeline>,
9323 12 : filter: Option<std::ops::Range<Key>>,
9324 12 : ) -> Vec<PersistentLayerKey> {
9325 12 : let mut all_layers = tline.inspect_historic_layers().await.unwrap();
9326 12 : if let Some(filter) = filter {
9327 64 : all_layers.retain(|layer| overlaps_with(&layer.key_range, &filter));
9328 10 : }
9329 12 : all_layers.sort_by(sort_layer_key);
9330 12 : all_layers
9331 12 : }
9332 :
9333 : #[cfg(feature = "testing")]
9334 10 : fn check_layer_map_key_eq(
9335 10 : mut left: Vec<PersistentLayerKey>,
9336 10 : mut right: Vec<PersistentLayerKey>,
9337 10 : ) {
9338 10 : left.sort_by(sort_layer_key);
9339 10 : right.sort_by(sort_layer_key);
9340 10 : if left != right {
9341 0 : eprintln!("---LEFT---");
9342 0 : for left in left.iter() {
9343 0 : eprintln!("{}", left);
9344 0 : }
9345 0 : eprintln!("---RIGHT---");
9346 0 : for right in right.iter() {
9347 0 : eprintln!("{}", right);
9348 0 : }
9349 0 : assert_eq!(left, right);
9350 10 : }
9351 10 : }
9352 :
9353 : #[cfg(feature = "testing")]
9354 : #[tokio::test]
9355 2 : async fn test_simple_partial_bottom_most_compaction() -> anyhow::Result<()> {
9356 2 : let harness = TenantHarness::create("test_simple_partial_bottom_most_compaction").await?;
9357 20 : let (tenant, ctx) = harness.load().await;
9358 2 :
9359 182 : fn get_key(id: u32) -> Key {
9360 182 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9361 182 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9362 182 : key.field6 = id;
9363 182 : key
9364 182 : }
9365 2 :
9366 2 : // img layer at 0x10
9367 2 : let img_layer = (0..10)
9368 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9369 2 : .collect_vec();
9370 2 :
9371 2 : let delta1 = vec![
9372 2 : (
9373 2 : get_key(1),
9374 2 : Lsn(0x20),
9375 2 : Value::Image(Bytes::from("value 1@0x20")),
9376 2 : ),
9377 2 : (
9378 2 : get_key(2),
9379 2 : Lsn(0x30),
9380 2 : Value::Image(Bytes::from("value 2@0x30")),
9381 2 : ),
9382 2 : (
9383 2 : get_key(3),
9384 2 : Lsn(0x40),
9385 2 : Value::Image(Bytes::from("value 3@0x40")),
9386 2 : ),
9387 2 : ];
9388 2 : let delta2 = vec![
9389 2 : (
9390 2 : get_key(5),
9391 2 : Lsn(0x20),
9392 2 : Value::Image(Bytes::from("value 5@0x20")),
9393 2 : ),
9394 2 : (
9395 2 : get_key(6),
9396 2 : Lsn(0x20),
9397 2 : Value::Image(Bytes::from("value 6@0x20")),
9398 2 : ),
9399 2 : ];
9400 2 : let delta3 = vec![
9401 2 : (
9402 2 : get_key(8),
9403 2 : Lsn(0x48),
9404 2 : Value::Image(Bytes::from("value 8@0x48")),
9405 2 : ),
9406 2 : (
9407 2 : get_key(9),
9408 2 : Lsn(0x48),
9409 2 : Value::Image(Bytes::from("value 9@0x48")),
9410 2 : ),
9411 2 : ];
9412 2 :
9413 2 : let tline = tenant
9414 2 : .create_test_timeline_with_layers(
9415 2 : TIMELINE_ID,
9416 2 : Lsn(0x10),
9417 2 : DEFAULT_PG_VERSION,
9418 2 : &ctx,
9419 2 : vec![
9420 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
9421 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
9422 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
9423 2 : ], // delta layers
9424 2 : vec![(Lsn(0x10), img_layer)], // image layers
9425 2 : Lsn(0x50),
9426 2 : )
9427 49 : .await?;
9428 2 :
9429 2 : {
9430 2 : // Update GC info
9431 2 : let mut guard = tline.gc_info.write().unwrap();
9432 2 : *guard = GcInfo {
9433 2 : retain_lsns: vec![(Lsn(0x20), tline.timeline_id, MaybeOffloaded::No)],
9434 2 : cutoffs: GcCutoffs {
9435 2 : time: Lsn(0x30),
9436 2 : space: Lsn(0x30),
9437 2 : },
9438 2 : leases: Default::default(),
9439 2 : within_ancestor_pitr: false,
9440 2 : };
9441 2 : }
9442 2 :
9443 2 : let cancel = CancellationToken::new();
9444 2 :
9445 2 : // Do a partial compaction on key range 0..2
9446 2 : tline
9447 2 : .partial_compact_with_gc(get_key(0)..get_key(2), &cancel, EnumSet::new(), &ctx)
9448 27 : .await
9449 2 : .unwrap();
9450 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
9451 2 : check_layer_map_key_eq(
9452 2 : all_layers,
9453 2 : vec![
9454 2 : // newly-generated image layer for the partial compaction range 0-2
9455 2 : PersistentLayerKey {
9456 2 : key_range: get_key(0)..get_key(2),
9457 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9458 2 : is_delta: false,
9459 2 : },
9460 2 : PersistentLayerKey {
9461 2 : key_range: get_key(0)..get_key(10),
9462 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
9463 2 : is_delta: false,
9464 2 : },
9465 2 : // delta1 is split and the second part is rewritten
9466 2 : PersistentLayerKey {
9467 2 : key_range: get_key(2)..get_key(4),
9468 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9469 2 : is_delta: true,
9470 2 : },
9471 2 : PersistentLayerKey {
9472 2 : key_range: get_key(5)..get_key(7),
9473 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9474 2 : is_delta: true,
9475 2 : },
9476 2 : PersistentLayerKey {
9477 2 : key_range: get_key(8)..get_key(10),
9478 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
9479 2 : is_delta: true,
9480 2 : },
9481 2 : ],
9482 2 : );
9483 2 :
9484 2 : // Do a partial compaction on key range 2..4
9485 2 : tline
9486 2 : .partial_compact_with_gc(get_key(2)..get_key(4), &cancel, EnumSet::new(), &ctx)
9487 17 : .await
9488 2 : .unwrap();
9489 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
9490 2 : check_layer_map_key_eq(
9491 2 : all_layers,
9492 2 : vec![
9493 2 : PersistentLayerKey {
9494 2 : key_range: get_key(0)..get_key(2),
9495 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9496 2 : is_delta: false,
9497 2 : },
9498 2 : PersistentLayerKey {
9499 2 : key_range: get_key(0)..get_key(10),
9500 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
9501 2 : is_delta: false,
9502 2 : },
9503 2 : // image layer generated for the compaction range 2-4
9504 2 : PersistentLayerKey {
9505 2 : key_range: get_key(2)..get_key(4),
9506 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9507 2 : is_delta: false,
9508 2 : },
9509 2 : // we have key2/key3 above the retain_lsn, so we still need this delta layer
9510 2 : PersistentLayerKey {
9511 2 : key_range: get_key(2)..get_key(4),
9512 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9513 2 : is_delta: true,
9514 2 : },
9515 2 : PersistentLayerKey {
9516 2 : key_range: get_key(5)..get_key(7),
9517 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9518 2 : is_delta: true,
9519 2 : },
9520 2 : PersistentLayerKey {
9521 2 : key_range: get_key(8)..get_key(10),
9522 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
9523 2 : is_delta: true,
9524 2 : },
9525 2 : ],
9526 2 : );
9527 2 :
9528 2 : // Do a partial compaction on key range 4..9
9529 2 : tline
9530 2 : .partial_compact_with_gc(get_key(4)..get_key(9), &cancel, EnumSet::new(), &ctx)
9531 22 : .await
9532 2 : .unwrap();
9533 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
9534 2 : check_layer_map_key_eq(
9535 2 : all_layers,
9536 2 : vec![
9537 2 : PersistentLayerKey {
9538 2 : key_range: get_key(0)..get_key(2),
9539 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9540 2 : is_delta: false,
9541 2 : },
9542 2 : PersistentLayerKey {
9543 2 : key_range: get_key(0)..get_key(10),
9544 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
9545 2 : is_delta: false,
9546 2 : },
9547 2 : PersistentLayerKey {
9548 2 : key_range: get_key(2)..get_key(4),
9549 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9550 2 : is_delta: false,
9551 2 : },
9552 2 : PersistentLayerKey {
9553 2 : key_range: get_key(2)..get_key(4),
9554 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9555 2 : is_delta: true,
9556 2 : },
9557 2 : // image layer generated for this compaction range
9558 2 : PersistentLayerKey {
9559 2 : key_range: get_key(4)..get_key(9),
9560 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9561 2 : is_delta: false,
9562 2 : },
9563 2 : PersistentLayerKey {
9564 2 : key_range: get_key(8)..get_key(10),
9565 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
9566 2 : is_delta: true,
9567 2 : },
9568 2 : ],
9569 2 : );
9570 2 :
9571 2 : // Do a partial compaction on key range 9..10
9572 2 : tline
9573 2 : .partial_compact_with_gc(get_key(9)..get_key(10), &cancel, EnumSet::new(), &ctx)
9574 9 : .await
9575 2 : .unwrap();
9576 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
9577 2 : check_layer_map_key_eq(
9578 2 : all_layers,
9579 2 : vec![
9580 2 : PersistentLayerKey {
9581 2 : key_range: get_key(0)..get_key(2),
9582 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9583 2 : is_delta: false,
9584 2 : },
9585 2 : PersistentLayerKey {
9586 2 : key_range: get_key(0)..get_key(10),
9587 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
9588 2 : is_delta: false,
9589 2 : },
9590 2 : PersistentLayerKey {
9591 2 : key_range: get_key(2)..get_key(4),
9592 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9593 2 : is_delta: false,
9594 2 : },
9595 2 : PersistentLayerKey {
9596 2 : key_range: get_key(2)..get_key(4),
9597 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9598 2 : is_delta: true,
9599 2 : },
9600 2 : PersistentLayerKey {
9601 2 : key_range: get_key(4)..get_key(9),
9602 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9603 2 : is_delta: false,
9604 2 : },
9605 2 : // image layer generated for the compaction range
9606 2 : PersistentLayerKey {
9607 2 : key_range: get_key(9)..get_key(10),
9608 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9609 2 : is_delta: false,
9610 2 : },
9611 2 : PersistentLayerKey {
9612 2 : key_range: get_key(8)..get_key(10),
9613 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
9614 2 : is_delta: true,
9615 2 : },
9616 2 : ],
9617 2 : );
9618 2 :
9619 2 : // Do a partial compaction on key range 0..10, all image layers below LSN 20 can be replaced with new ones.
9620 2 : tline
9621 2 : .partial_compact_with_gc(get_key(0)..get_key(10), &cancel, EnumSet::new(), &ctx)
9622 50 : .await
9623 2 : .unwrap();
9624 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
9625 2 : check_layer_map_key_eq(
9626 2 : all_layers,
9627 2 : vec![
9628 2 : // aha, we removed all unnecessary image/delta layers and got a very clean layer map!
9629 2 : PersistentLayerKey {
9630 2 : key_range: get_key(0)..get_key(10),
9631 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9632 2 : is_delta: false,
9633 2 : },
9634 2 : PersistentLayerKey {
9635 2 : key_range: get_key(2)..get_key(4),
9636 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9637 2 : is_delta: true,
9638 2 : },
9639 2 : PersistentLayerKey {
9640 2 : key_range: get_key(8)..get_key(10),
9641 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
9642 2 : is_delta: true,
9643 2 : },
9644 2 : ],
9645 2 : );
9646 2 :
9647 2 : Ok(())
9648 2 : }
9649 : }
|
