Line data Source code
1 : //! Timeline repository implementation that keeps old data in layer files, and
2 : //! the recent changes in ephemeral files.
3 : //!
4 : //! See tenant/*_layer.rs files. The functions here are responsible for locating
5 : //! the correct layer for the get/put call, walking back the timeline branching
6 : //! history as needed.
7 : //!
8 : //! The files are stored in the .neon/tenants/<tenant_id>/timelines/<timeline_id>
9 : //! directory. See docs/pageserver-storage.md for how the files are managed.
10 : //! In addition to the layer files, there is a metadata file in the same
11 : //! directory that contains information about the timeline, in particular its
12 : //! parent timeline, and the last LSN that has been written to disk.
13 : //!
14 :
15 : use anyhow::{bail, Context};
16 : use arc_swap::ArcSwap;
17 : use camino::Utf8Path;
18 : use camino::Utf8PathBuf;
19 : use chrono::NaiveDateTime;
20 : use enumset::EnumSet;
21 : use futures::stream::FuturesUnordered;
22 : use futures::StreamExt;
23 : use pageserver_api::models;
24 : use pageserver_api::models::LsnLease;
25 : use pageserver_api::models::TimelineArchivalState;
26 : use pageserver_api::models::TimelineState;
27 : use pageserver_api::models::TopTenantShardItem;
28 : use pageserver_api::models::WalRedoManagerStatus;
29 : use pageserver_api::shard::ShardIdentity;
30 : use pageserver_api::shard::ShardStripeSize;
31 : use pageserver_api::shard::TenantShardId;
32 : use remote_storage::DownloadError;
33 : use remote_storage::GenericRemoteStorage;
34 : use remote_storage::TimeoutOrCancel;
35 : use remote_timeline_client::manifest::{
36 : OffloadedTimelineManifest, TenantManifest, LATEST_TENANT_MANIFEST_VERSION,
37 : };
38 : use remote_timeline_client::UploadQueueNotReadyError;
39 : use std::collections::BTreeMap;
40 : use std::collections::VecDeque;
41 : use std::fmt;
42 : use std::future::Future;
43 : use std::sync::atomic::AtomicBool;
44 : use std::sync::Weak;
45 : use std::time::SystemTime;
46 : use storage_broker::BrokerClientChannel;
47 : use timeline::compaction::ScheduledCompactionTask;
48 : use timeline::import_pgdata;
49 : use timeline::offload::offload_timeline;
50 : use timeline::CompactFlags;
51 : use timeline::CompactOptions;
52 : use timeline::CompactionError;
53 : use timeline::ShutdownMode;
54 : use tokio::io::BufReader;
55 : use tokio::sync::watch;
56 : use tokio::task::JoinSet;
57 : use tokio_util::sync::CancellationToken;
58 : use tracing::*;
59 : use upload_queue::NotInitialized;
60 : use utils::backoff;
61 : use utils::circuit_breaker::CircuitBreaker;
62 : use utils::completion;
63 : use utils::crashsafe::path_with_suffix_extension;
64 : use utils::failpoint_support;
65 : use utils::fs_ext;
66 : use utils::pausable_failpoint;
67 : use utils::sync::gate::Gate;
68 : use utils::sync::gate::GateGuard;
69 : use utils::timeout::timeout_cancellable;
70 : use utils::timeout::TimeoutCancellableError;
71 : use utils::zstd::create_zst_tarball;
72 : use utils::zstd::extract_zst_tarball;
73 :
74 : use self::config::AttachedLocationConfig;
75 : use self::config::AttachmentMode;
76 : use self::config::LocationConf;
77 : use self::config::TenantConf;
78 : use self::metadata::TimelineMetadata;
79 : use self::mgr::GetActiveTenantError;
80 : use self::mgr::GetTenantError;
81 : use self::remote_timeline_client::upload::{upload_index_part, upload_tenant_manifest};
82 : use self::remote_timeline_client::{RemoteTimelineClient, WaitCompletionError};
83 : use self::timeline::uninit::TimelineCreateGuard;
84 : use self::timeline::uninit::TimelineExclusionError;
85 : use self::timeline::uninit::UninitializedTimeline;
86 : use self::timeline::EvictionTaskTenantState;
87 : use self::timeline::GcCutoffs;
88 : use self::timeline::TimelineDeleteProgress;
89 : use self::timeline::TimelineResources;
90 : use self::timeline::WaitLsnError;
91 : use crate::config::PageServerConf;
92 : use crate::context::{DownloadBehavior, RequestContext};
93 : use crate::deletion_queue::DeletionQueueClient;
94 : use crate::deletion_queue::DeletionQueueError;
95 : use crate::import_datadir;
96 : use crate::is_uninit_mark;
97 : use crate::l0_flush::L0FlushGlobalState;
98 : use crate::metrics::TENANT;
99 : use crate::metrics::{
100 : remove_tenant_metrics, BROKEN_TENANTS_SET, CIRCUIT_BREAKERS_BROKEN, CIRCUIT_BREAKERS_UNBROKEN,
101 : TENANT_STATE_METRIC, TENANT_SYNTHETIC_SIZE_METRIC,
102 : };
103 : use crate::task_mgr;
104 : use crate::task_mgr::TaskKind;
105 : use crate::tenant::config::LocationMode;
106 : use crate::tenant::config::TenantConfOpt;
107 : use crate::tenant::gc_result::GcResult;
108 : pub use crate::tenant::remote_timeline_client::index::IndexPart;
109 : use crate::tenant::remote_timeline_client::remote_initdb_archive_path;
110 : use crate::tenant::remote_timeline_client::MaybeDeletedIndexPart;
111 : use crate::tenant::remote_timeline_client::INITDB_PATH;
112 : use crate::tenant::storage_layer::DeltaLayer;
113 : use crate::tenant::storage_layer::ImageLayer;
114 : use crate::walingest::WalLagCooldown;
115 : use crate::walredo;
116 : use crate::InitializationOrder;
117 : use std::collections::hash_map::Entry;
118 : use std::collections::HashMap;
119 : use std::collections::HashSet;
120 : use std::fmt::Debug;
121 : use std::fmt::Display;
122 : use std::fs;
123 : use std::fs::File;
124 : use std::sync::atomic::{AtomicU64, Ordering};
125 : use std::sync::Arc;
126 : use std::sync::Mutex;
127 : use std::time::{Duration, Instant};
128 :
129 : use crate::span;
130 : use crate::tenant::timeline::delete::DeleteTimelineFlow;
131 : use crate::tenant::timeline::uninit::cleanup_timeline_directory;
132 : use crate::virtual_file::VirtualFile;
133 : use crate::walredo::PostgresRedoManager;
134 : use crate::TEMP_FILE_SUFFIX;
135 : use once_cell::sync::Lazy;
136 : pub use pageserver_api::models::TenantState;
137 : use tokio::sync::Semaphore;
138 :
139 0 : static INIT_DB_SEMAPHORE: Lazy<Semaphore> = Lazy::new(|| Semaphore::new(8));
140 : use utils::{
141 : crashsafe,
142 : generation::Generation,
143 : id::TimelineId,
144 : lsn::{Lsn, RecordLsn},
145 : };
146 :
147 : pub mod blob_io;
148 : pub mod block_io;
149 : pub mod vectored_blob_io;
150 :
151 : pub mod disk_btree;
152 : pub(crate) mod ephemeral_file;
153 : pub mod layer_map;
154 :
155 : pub mod metadata;
156 : pub mod remote_timeline_client;
157 : pub mod storage_layer;
158 :
159 : pub mod checks;
160 : pub mod config;
161 : pub mod mgr;
162 : pub mod secondary;
163 : pub mod tasks;
164 : pub mod upload_queue;
165 :
166 : pub(crate) mod timeline;
167 :
168 : pub mod size;
169 :
170 : mod gc_block;
171 : mod gc_result;
172 : pub(crate) mod throttle;
173 :
174 : pub(crate) use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
175 : pub(crate) use timeline::{LogicalSizeCalculationCause, PageReconstructError, Timeline};
176 :
177 : // re-export for use in walreceiver
178 : pub use crate::tenant::timeline::WalReceiverInfo;
179 :
180 : /// The "tenants" part of `tenants/<tenant>/timelines...`
181 : pub const TENANTS_SEGMENT_NAME: &str = "tenants";
182 :
183 : /// Parts of the `.neon/tenants/<tenant_id>/timelines/<timeline_id>` directory prefix.
184 : pub const TIMELINES_SEGMENT_NAME: &str = "timelines";
185 :
186 : /// References to shared objects that are passed into each tenant, such
187 : /// as the shared remote storage client and process initialization state.
188 : #[derive(Clone)]
189 : pub struct TenantSharedResources {
190 : pub broker_client: storage_broker::BrokerClientChannel,
191 : pub remote_storage: GenericRemoteStorage,
192 : pub deletion_queue_client: DeletionQueueClient,
193 : pub l0_flush_global_state: L0FlushGlobalState,
194 : }
195 :
196 : /// A [`Tenant`] is really an _attached_ tenant. The configuration
197 : /// for an attached tenant is a subset of the [`LocationConf`], represented
198 : /// in this struct.
199 : #[derive(Clone)]
200 : pub(super) struct AttachedTenantConf {
201 : tenant_conf: TenantConfOpt,
202 : location: AttachedLocationConfig,
203 : /// The deadline before which we are blocked from GC so that
204 : /// leases have a chance to be renewed.
205 : lsn_lease_deadline: Option<tokio::time::Instant>,
206 : }
207 :
208 : impl AttachedTenantConf {
209 192 : fn new(tenant_conf: TenantConfOpt, location: AttachedLocationConfig) -> Self {
210 : // Sets a deadline before which we cannot proceed to GC due to lsn lease.
211 : //
212 : // We do this as the leases mapping are not persisted to disk. By delaying GC by lease
213 : // length, we guarantee that all the leases we granted before will have a chance to renew
214 : // when we run GC for the first time after restart / transition from AttachedMulti to AttachedSingle.
215 192 : let lsn_lease_deadline = if location.attach_mode == AttachmentMode::Single {
216 192 : Some(
217 192 : tokio::time::Instant::now()
218 192 : + tenant_conf
219 192 : .lsn_lease_length
220 192 : .unwrap_or(LsnLease::DEFAULT_LENGTH),
221 192 : )
222 : } else {
223 : // We don't use `lsn_lease_deadline` to delay GC in AttachedMulti and AttachedStale
224 : // because we don't do GC in these modes.
225 0 : None
226 : };
227 :
228 192 : Self {
229 192 : tenant_conf,
230 192 : location,
231 192 : lsn_lease_deadline,
232 192 : }
233 192 : }
234 :
235 192 : fn try_from(location_conf: LocationConf) -> anyhow::Result<Self> {
236 192 : match &location_conf.mode {
237 192 : LocationMode::Attached(attach_conf) => {
238 192 : Ok(Self::new(location_conf.tenant_conf, *attach_conf))
239 : }
240 : LocationMode::Secondary(_) => {
241 0 : anyhow::bail!("Attempted to construct AttachedTenantConf from a LocationConf in secondary mode")
242 : }
243 : }
244 192 : }
245 :
246 762 : fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
247 762 : self.lsn_lease_deadline
248 762 : .map(|d| tokio::time::Instant::now() < d)
249 762 : .unwrap_or(false)
250 762 : }
251 : }
252 : struct TimelinePreload {
253 : timeline_id: TimelineId,
254 : client: RemoteTimelineClient,
255 : index_part: Result<MaybeDeletedIndexPart, DownloadError>,
256 : }
257 :
258 : pub(crate) struct TenantPreload {
259 : tenant_manifest: TenantManifest,
260 : /// Map from timeline ID to a possible timeline preload. It is None iff the timeline is offloaded according to the manifest.
261 : timelines: HashMap<TimelineId, Option<TimelinePreload>>,
262 : }
263 :
264 : /// When we spawn a tenant, there is a special mode for tenant creation that
265 : /// avoids trying to read anything from remote storage.
266 : pub(crate) enum SpawnMode {
267 : /// Activate as soon as possible
268 : Eager,
269 : /// Lazy activation in the background, with the option to skip the queue if the need comes up
270 : Lazy,
271 : }
272 :
273 : ///
274 : /// Tenant consists of multiple timelines. Keep them in a hash table.
275 : ///
276 : pub struct Tenant {
277 : // Global pageserver config parameters
278 : pub conf: &'static PageServerConf,
279 :
280 : /// The value creation timestamp, used to measure activation delay, see:
281 : /// <https://github.com/neondatabase/neon/issues/4025>
282 : constructed_at: Instant,
283 :
284 : state: watch::Sender<TenantState>,
285 :
286 : // Overridden tenant-specific config parameters.
287 : // We keep TenantConfOpt sturct here to preserve the information
288 : // about parameters that are not set.
289 : // This is necessary to allow global config updates.
290 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
291 :
292 : tenant_shard_id: TenantShardId,
293 :
294 : // The detailed sharding information, beyond the number/count in tenant_shard_id
295 : shard_identity: ShardIdentity,
296 :
297 : /// The remote storage generation, used to protect S3 objects from split-brain.
298 : /// Does not change over the lifetime of the [`Tenant`] object.
299 : ///
300 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
301 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
302 : generation: Generation,
303 :
304 : timelines: Mutex<HashMap<TimelineId, Arc<Timeline>>>,
305 :
306 : /// During timeline creation, we first insert the TimelineId to the
307 : /// creating map, then `timelines`, then remove it from the creating map.
308 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
309 : timelines_creating: std::sync::Mutex<HashSet<TimelineId>>,
310 :
311 : /// Possibly offloaded and archived timelines
312 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
313 : timelines_offloaded: Mutex<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
314 :
315 : /// Serialize writes of the tenant manifest to remote storage. If there are concurrent operations
316 : /// affecting the manifest, such as timeline deletion and timeline offload, they must wait for
317 : /// each other (this could be optimized to coalesce writes if necessary).
318 : ///
319 : /// The contents of the Mutex are the last manifest we successfully uploaded
320 : tenant_manifest_upload: tokio::sync::Mutex<Option<TenantManifest>>,
321 :
322 : // This mutex prevents creation of new timelines during GC.
323 : // Adding yet another mutex (in addition to `timelines`) is needed because holding
324 : // `timelines` mutex during all GC iteration
325 : // may block for a long time `get_timeline`, `get_timelines_state`,... and other operations
326 : // with timelines, which in turn may cause dropping replication connection, expiration of wait_for_lsn
327 : // timeout...
328 : gc_cs: tokio::sync::Mutex<()>,
329 : walredo_mgr: Option<Arc<WalRedoManager>>,
330 :
331 : // provides access to timeline data sitting in the remote storage
332 : pub(crate) remote_storage: GenericRemoteStorage,
333 :
334 : // Access to global deletion queue for when this tenant wants to schedule a deletion
335 : deletion_queue_client: DeletionQueueClient,
336 :
337 : /// Cached logical sizes updated updated on each [`Tenant::gather_size_inputs`].
338 : cached_logical_sizes: tokio::sync::Mutex<HashMap<(TimelineId, Lsn), u64>>,
339 : cached_synthetic_tenant_size: Arc<AtomicU64>,
340 :
341 : eviction_task_tenant_state: tokio::sync::Mutex<EvictionTaskTenantState>,
342 :
343 : /// Track repeated failures to compact, so that we can back off.
344 : /// Overhead of mutex is acceptable because compaction is done with a multi-second period.
345 : compaction_circuit_breaker: std::sync::Mutex<CircuitBreaker>,
346 :
347 : /// Scheduled compaction tasks. Currently, this can only be populated by triggering
348 : /// a manual gc-compaction from the manual compaction API.
349 : scheduled_compaction_tasks:
350 : std::sync::Mutex<HashMap<TimelineId, VecDeque<ScheduledCompactionTask>>>,
351 :
352 : /// If the tenant is in Activating state, notify this to encourage it
353 : /// to proceed to Active as soon as possible, rather than waiting for lazy
354 : /// background warmup.
355 : pub(crate) activate_now_sem: tokio::sync::Semaphore,
356 :
357 : /// Time it took for the tenant to activate. Zero if not active yet.
358 : attach_wal_lag_cooldown: Arc<std::sync::OnceLock<WalLagCooldown>>,
359 :
360 : // Cancellation token fires when we have entered shutdown(). This is a parent of
361 : // Timelines' cancellation token.
362 : pub(crate) cancel: CancellationToken,
363 :
364 : // Users of the Tenant such as the page service must take this Gate to avoid
365 : // trying to use a Tenant which is shutting down.
366 : pub(crate) gate: Gate,
367 :
368 : /// Throttle applied at the top of [`Timeline::get`].
369 : /// All [`Tenant::timelines`] of a given [`Tenant`] instance share the same [`throttle::Throttle`] instance.
370 : pub(crate) pagestream_throttle:
371 : Arc<throttle::Throttle<crate::metrics::tenant_throttling::Pagestream>>,
372 :
373 : /// An ongoing timeline detach concurrency limiter.
374 : ///
375 : /// As a tenant will likely be restarted as part of timeline detach ancestor it makes no sense
376 : /// to have two running at the same time. A different one can be started if an earlier one
377 : /// has failed for whatever reason.
378 : ongoing_timeline_detach: std::sync::Mutex<Option<(TimelineId, utils::completion::Barrier)>>,
379 :
380 : /// `index_part.json` based gc blocking reason tracking.
381 : ///
382 : /// New gc iterations must start a new iteration by acquiring `GcBlock::start` before
383 : /// proceeding.
384 : pub(crate) gc_block: gc_block::GcBlock,
385 :
386 : l0_flush_global_state: L0FlushGlobalState,
387 : }
388 : impl std::fmt::Debug for Tenant {
389 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
390 0 : write!(f, "{} ({})", self.tenant_shard_id, self.current_state())
391 0 : }
392 : }
393 :
394 : pub(crate) enum WalRedoManager {
395 : Prod(WalredoManagerId, PostgresRedoManager),
396 : #[cfg(test)]
397 : Test(harness::TestRedoManager),
398 : }
399 :
400 : #[derive(thiserror::Error, Debug)]
401 : #[error("pageserver is shutting down")]
402 : pub(crate) struct GlobalShutDown;
403 :
404 : impl WalRedoManager {
405 0 : pub(crate) fn new(mgr: PostgresRedoManager) -> Result<Arc<Self>, GlobalShutDown> {
406 0 : let id = WalredoManagerId::next();
407 0 : let arc = Arc::new(Self::Prod(id, mgr));
408 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
409 0 : match &mut *guard {
410 0 : Some(map) => {
411 0 : map.insert(id, Arc::downgrade(&arc));
412 0 : Ok(arc)
413 : }
414 0 : None => Err(GlobalShutDown),
415 : }
416 0 : }
417 : }
418 :
419 : impl Drop for WalRedoManager {
420 10 : fn drop(&mut self) {
421 10 : match self {
422 0 : Self::Prod(id, _) => {
423 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
424 0 : if let Some(map) = &mut *guard {
425 0 : map.remove(id).expect("new() registers, drop() unregisters");
426 0 : }
427 : }
428 : #[cfg(test)]
429 10 : Self::Test(_) => {
430 10 : // Not applicable to test redo manager
431 10 : }
432 : }
433 10 : }
434 : }
435 :
436 : /// Global registry of all walredo managers so that [`crate::shutdown_pageserver`] can shut down
437 : /// the walredo processes outside of the regular order.
438 : ///
439 : /// This is necessary to work around a systemd bug where it freezes if there are
440 : /// walredo processes left => <https://github.com/neondatabase/cloud/issues/11387>
441 : #[allow(clippy::type_complexity)]
442 : pub(crate) static WALREDO_MANAGERS: once_cell::sync::Lazy<
443 : Mutex<Option<HashMap<WalredoManagerId, Weak<WalRedoManager>>>>,
444 0 : > = once_cell::sync::Lazy::new(|| Mutex::new(Some(HashMap::new())));
445 : #[derive(PartialEq, Eq, Hash, Clone, Copy, Debug)]
446 : pub(crate) struct WalredoManagerId(u64);
447 : impl WalredoManagerId {
448 0 : pub fn next() -> Self {
449 : static NEXT: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(1);
450 0 : let id = NEXT.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
451 0 : if id == 0 {
452 0 : panic!("WalredoManagerId::new() returned 0, indicating wraparound, risking it's no longer unique");
453 0 : }
454 0 : Self(id)
455 0 : }
456 : }
457 :
458 : #[cfg(test)]
459 : impl From<harness::TestRedoManager> for WalRedoManager {
460 192 : fn from(mgr: harness::TestRedoManager) -> Self {
461 192 : Self::Test(mgr)
462 192 : }
463 : }
464 :
465 : impl WalRedoManager {
466 6 : pub(crate) async fn shutdown(&self) -> bool {
467 6 : match self {
468 0 : Self::Prod(_, mgr) => mgr.shutdown().await,
469 : #[cfg(test)]
470 : Self::Test(_) => {
471 : // Not applicable to test redo manager
472 6 : true
473 : }
474 : }
475 6 : }
476 :
477 0 : pub(crate) fn maybe_quiesce(&self, idle_timeout: Duration) {
478 0 : match self {
479 0 : Self::Prod(_, mgr) => mgr.maybe_quiesce(idle_timeout),
480 0 : #[cfg(test)]
481 0 : Self::Test(_) => {
482 0 : // Not applicable to test redo manager
483 0 : }
484 0 : }
485 0 : }
486 :
487 : /// # Cancel-Safety
488 : ///
489 : /// This method is cancellation-safe.
490 410 : pub async fn request_redo(
491 410 : &self,
492 410 : key: pageserver_api::key::Key,
493 410 : lsn: Lsn,
494 410 : base_img: Option<(Lsn, bytes::Bytes)>,
495 410 : records: Vec<(Lsn, pageserver_api::record::NeonWalRecord)>,
496 410 : pg_version: u32,
497 410 : ) -> Result<bytes::Bytes, walredo::Error> {
498 410 : match self {
499 0 : Self::Prod(_, mgr) => {
500 0 : mgr.request_redo(key, lsn, base_img, records, pg_version)
501 0 : .await
502 : }
503 : #[cfg(test)]
504 410 : Self::Test(mgr) => {
505 410 : mgr.request_redo(key, lsn, base_img, records, pg_version)
506 410 : .await
507 : }
508 : }
509 410 : }
510 :
511 0 : pub(crate) fn status(&self) -> Option<WalRedoManagerStatus> {
512 0 : match self {
513 0 : WalRedoManager::Prod(_, m) => Some(m.status()),
514 0 : #[cfg(test)]
515 0 : WalRedoManager::Test(_) => None,
516 0 : }
517 0 : }
518 : }
519 :
520 : /// A very lightweight memory representation of an offloaded timeline.
521 : ///
522 : /// We need to store the list of offloaded timelines so that we can perform operations on them,
523 : /// like unoffloading them, or (at a later date), decide to perform flattening.
524 : /// This type has a much smaller memory impact than [`Timeline`], and thus we can store many
525 : /// more offloaded timelines than we can manage ones that aren't.
526 : pub struct OffloadedTimeline {
527 : pub tenant_shard_id: TenantShardId,
528 : pub timeline_id: TimelineId,
529 : pub ancestor_timeline_id: Option<TimelineId>,
530 : /// Whether to retain the branch lsn at the ancestor or not
531 : pub ancestor_retain_lsn: Option<Lsn>,
532 :
533 : /// When the timeline was archived.
534 : ///
535 : /// Present for future flattening deliberations.
536 : pub archived_at: NaiveDateTime,
537 :
538 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
539 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
540 : pub delete_progress: TimelineDeleteProgress,
541 :
542 : /// Part of the `OffloadedTimeline` object's lifecycle: this needs to be set before we drop it
543 : pub deleted_from_ancestor: AtomicBool,
544 : }
545 :
546 : impl OffloadedTimeline {
547 : /// Obtains an offloaded timeline from a given timeline object.
548 : ///
549 : /// Returns `None` if the `archived_at` flag couldn't be obtained, i.e.
550 : /// the timeline is not in a stopped state.
551 : /// Panics if the timeline is not archived.
552 2 : fn from_timeline(timeline: &Timeline) -> Result<Self, UploadQueueNotReadyError> {
553 2 : let (ancestor_retain_lsn, ancestor_timeline_id) =
554 2 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
555 2 : let ancestor_lsn = timeline.get_ancestor_lsn();
556 2 : let ancestor_timeline_id = ancestor_timeline.timeline_id;
557 2 : let mut gc_info = ancestor_timeline.gc_info.write().unwrap();
558 2 : gc_info.insert_child(timeline.timeline_id, ancestor_lsn, MaybeOffloaded::Yes);
559 2 : (Some(ancestor_lsn), Some(ancestor_timeline_id))
560 : } else {
561 0 : (None, None)
562 : };
563 2 : let archived_at = timeline
564 2 : .remote_client
565 2 : .archived_at_stopped_queue()?
566 2 : .expect("must be called on an archived timeline");
567 2 : Ok(Self {
568 2 : tenant_shard_id: timeline.tenant_shard_id,
569 2 : timeline_id: timeline.timeline_id,
570 2 : ancestor_timeline_id,
571 2 : ancestor_retain_lsn,
572 2 : archived_at,
573 2 :
574 2 : delete_progress: timeline.delete_progress.clone(),
575 2 : deleted_from_ancestor: AtomicBool::new(false),
576 2 : })
577 2 : }
578 0 : fn from_manifest(tenant_shard_id: TenantShardId, manifest: &OffloadedTimelineManifest) -> Self {
579 0 : // We expect to reach this case in tenant loading, where the `retain_lsn` is populated in the parent's `gc_info`
580 0 : // by the `initialize_gc_info` function.
581 0 : let OffloadedTimelineManifest {
582 0 : timeline_id,
583 0 : ancestor_timeline_id,
584 0 : ancestor_retain_lsn,
585 0 : archived_at,
586 0 : } = *manifest;
587 0 : Self {
588 0 : tenant_shard_id,
589 0 : timeline_id,
590 0 : ancestor_timeline_id,
591 0 : ancestor_retain_lsn,
592 0 : archived_at,
593 0 : delete_progress: TimelineDeleteProgress::default(),
594 0 : deleted_from_ancestor: AtomicBool::new(false),
595 0 : }
596 0 : }
597 2 : fn manifest(&self) -> OffloadedTimelineManifest {
598 2 : let Self {
599 2 : timeline_id,
600 2 : ancestor_timeline_id,
601 2 : ancestor_retain_lsn,
602 2 : archived_at,
603 2 : ..
604 2 : } = self;
605 2 : OffloadedTimelineManifest {
606 2 : timeline_id: *timeline_id,
607 2 : ancestor_timeline_id: *ancestor_timeline_id,
608 2 : ancestor_retain_lsn: *ancestor_retain_lsn,
609 2 : archived_at: *archived_at,
610 2 : }
611 2 : }
612 : /// Delete this timeline's retain_lsn from its ancestor, if present in the given tenant
613 0 : fn delete_from_ancestor_with_timelines(
614 0 : &self,
615 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
616 0 : ) {
617 0 : if let (Some(_retain_lsn), Some(ancestor_timeline_id)) =
618 0 : (self.ancestor_retain_lsn, self.ancestor_timeline_id)
619 : {
620 0 : if let Some((_, ancestor_timeline)) = timelines
621 0 : .iter()
622 0 : .find(|(tid, _tl)| **tid == ancestor_timeline_id)
623 : {
624 0 : let removal_happened = ancestor_timeline
625 0 : .gc_info
626 0 : .write()
627 0 : .unwrap()
628 0 : .remove_child_offloaded(self.timeline_id);
629 0 : if !removal_happened {
630 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
631 0 : "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
632 0 : }
633 0 : }
634 0 : }
635 0 : self.deleted_from_ancestor.store(true, Ordering::Release);
636 0 : }
637 : /// Call [`Self::delete_from_ancestor_with_timelines`] instead if possible.
638 : ///
639 : /// As the entire tenant is being dropped, don't bother deregistering the `retain_lsn` from the ancestor.
640 2 : fn defuse_for_tenant_drop(&self) {
641 2 : self.deleted_from_ancestor.store(true, Ordering::Release);
642 2 : }
643 : }
644 :
645 : impl fmt::Debug for OffloadedTimeline {
646 0 : fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
647 0 : write!(f, "OffloadedTimeline<{}>", self.timeline_id)
648 0 : }
649 : }
650 :
651 : impl Drop for OffloadedTimeline {
652 2 : fn drop(&mut self) {
653 2 : if !self.deleted_from_ancestor.load(Ordering::Acquire) {
654 0 : tracing::warn!(
655 0 : "offloaded timeline {} was dropped without having cleaned it up at the ancestor",
656 : self.timeline_id
657 : );
658 2 : }
659 2 : }
660 : }
661 :
662 : #[derive(Copy, Clone, PartialEq, Eq, Hash, Debug)]
663 : pub enum MaybeOffloaded {
664 : Yes,
665 : No,
666 : }
667 :
668 : #[derive(Clone, Debug)]
669 : pub enum TimelineOrOffloaded {
670 : Timeline(Arc<Timeline>),
671 : Offloaded(Arc<OffloadedTimeline>),
672 : }
673 :
674 : impl TimelineOrOffloaded {
675 0 : pub fn arc_ref(&self) -> TimelineOrOffloadedArcRef<'_> {
676 0 : match self {
677 0 : TimelineOrOffloaded::Timeline(timeline) => {
678 0 : TimelineOrOffloadedArcRef::Timeline(timeline)
679 : }
680 0 : TimelineOrOffloaded::Offloaded(offloaded) => {
681 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded)
682 : }
683 : }
684 0 : }
685 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
686 0 : self.arc_ref().tenant_shard_id()
687 0 : }
688 0 : pub fn timeline_id(&self) -> TimelineId {
689 0 : self.arc_ref().timeline_id()
690 0 : }
691 2 : pub fn delete_progress(&self) -> &Arc<tokio::sync::Mutex<DeleteTimelineFlow>> {
692 2 : match self {
693 2 : TimelineOrOffloaded::Timeline(timeline) => &timeline.delete_progress,
694 0 : TimelineOrOffloaded::Offloaded(offloaded) => &offloaded.delete_progress,
695 : }
696 2 : }
697 0 : fn maybe_remote_client(&self) -> Option<Arc<RemoteTimelineClient>> {
698 0 : match self {
699 0 : TimelineOrOffloaded::Timeline(timeline) => Some(timeline.remote_client.clone()),
700 0 : TimelineOrOffloaded::Offloaded(_offloaded) => None,
701 : }
702 0 : }
703 : }
704 :
705 : pub enum TimelineOrOffloadedArcRef<'a> {
706 : Timeline(&'a Arc<Timeline>),
707 : Offloaded(&'a Arc<OffloadedTimeline>),
708 : }
709 :
710 : impl TimelineOrOffloadedArcRef<'_> {
711 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
712 0 : match self {
713 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.tenant_shard_id,
714 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.tenant_shard_id,
715 : }
716 0 : }
717 0 : pub fn timeline_id(&self) -> TimelineId {
718 0 : match self {
719 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.timeline_id,
720 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.timeline_id,
721 : }
722 0 : }
723 : }
724 :
725 : impl<'a> From<&'a Arc<Timeline>> for TimelineOrOffloadedArcRef<'a> {
726 0 : fn from(timeline: &'a Arc<Timeline>) -> Self {
727 0 : Self::Timeline(timeline)
728 0 : }
729 : }
730 :
731 : impl<'a> From<&'a Arc<OffloadedTimeline>> for TimelineOrOffloadedArcRef<'a> {
732 0 : fn from(timeline: &'a Arc<OffloadedTimeline>) -> Self {
733 0 : Self::Offloaded(timeline)
734 0 : }
735 : }
736 :
737 : #[derive(Debug, thiserror::Error, PartialEq, Eq)]
738 : pub enum GetTimelineError {
739 : #[error("Timeline is shutting down")]
740 : ShuttingDown,
741 : #[error("Timeline {tenant_id}/{timeline_id} is not active, state: {state:?}")]
742 : NotActive {
743 : tenant_id: TenantShardId,
744 : timeline_id: TimelineId,
745 : state: TimelineState,
746 : },
747 : #[error("Timeline {tenant_id}/{timeline_id} was not found")]
748 : NotFound {
749 : tenant_id: TenantShardId,
750 : timeline_id: TimelineId,
751 : },
752 : }
753 :
754 : #[derive(Debug, thiserror::Error)]
755 : pub enum LoadLocalTimelineError {
756 : #[error("FailedToLoad")]
757 : Load(#[source] anyhow::Error),
758 : #[error("FailedToResumeDeletion")]
759 : ResumeDeletion(#[source] anyhow::Error),
760 : }
761 :
762 : #[derive(thiserror::Error)]
763 : pub enum DeleteTimelineError {
764 : #[error("NotFound")]
765 : NotFound,
766 :
767 : #[error("HasChildren")]
768 : HasChildren(Vec<TimelineId>),
769 :
770 : #[error("Timeline deletion is already in progress")]
771 : AlreadyInProgress(Arc<tokio::sync::Mutex<DeleteTimelineFlow>>),
772 :
773 : #[error("Cancelled")]
774 : Cancelled,
775 :
776 : #[error(transparent)]
777 : Other(#[from] anyhow::Error),
778 : }
779 :
780 : impl Debug for DeleteTimelineError {
781 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
782 0 : match self {
783 0 : Self::NotFound => write!(f, "NotFound"),
784 0 : Self::HasChildren(c) => f.debug_tuple("HasChildren").field(c).finish(),
785 0 : Self::AlreadyInProgress(_) => f.debug_tuple("AlreadyInProgress").finish(),
786 0 : Self::Cancelled => f.debug_tuple("Cancelled").finish(),
787 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
788 : }
789 0 : }
790 : }
791 :
792 : #[derive(thiserror::Error)]
793 : pub enum TimelineArchivalError {
794 : #[error("NotFound")]
795 : NotFound,
796 :
797 : #[error("Timeout")]
798 : Timeout,
799 :
800 : #[error("Cancelled")]
801 : Cancelled,
802 :
803 : #[error("ancestor is archived: {}", .0)]
804 : HasArchivedParent(TimelineId),
805 :
806 : #[error("HasUnarchivedChildren")]
807 : HasUnarchivedChildren(Vec<TimelineId>),
808 :
809 : #[error("Timeline archival is already in progress")]
810 : AlreadyInProgress,
811 :
812 : #[error(transparent)]
813 : Other(anyhow::Error),
814 : }
815 :
816 : #[derive(thiserror::Error, Debug)]
817 : pub(crate) enum TenantManifestError {
818 : #[error("Remote storage error: {0}")]
819 : RemoteStorage(anyhow::Error),
820 :
821 : #[error("Cancelled")]
822 : Cancelled,
823 : }
824 :
825 : impl From<TenantManifestError> for TimelineArchivalError {
826 0 : fn from(e: TenantManifestError) -> Self {
827 0 : match e {
828 0 : TenantManifestError::RemoteStorage(e) => Self::Other(e),
829 0 : TenantManifestError::Cancelled => Self::Cancelled,
830 : }
831 0 : }
832 : }
833 :
834 : impl Debug for TimelineArchivalError {
835 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
836 0 : match self {
837 0 : Self::NotFound => write!(f, "NotFound"),
838 0 : Self::Timeout => write!(f, "Timeout"),
839 0 : Self::Cancelled => write!(f, "Cancelled"),
840 0 : Self::HasArchivedParent(p) => f.debug_tuple("HasArchivedParent").field(p).finish(),
841 0 : Self::HasUnarchivedChildren(c) => {
842 0 : f.debug_tuple("HasUnarchivedChildren").field(c).finish()
843 : }
844 0 : Self::AlreadyInProgress => f.debug_tuple("AlreadyInProgress").finish(),
845 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
846 : }
847 0 : }
848 : }
849 :
850 : pub enum SetStoppingError {
851 : AlreadyStopping(completion::Barrier),
852 : Broken,
853 : }
854 :
855 : impl Debug for SetStoppingError {
856 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
857 0 : match self {
858 0 : Self::AlreadyStopping(_) => f.debug_tuple("AlreadyStopping").finish(),
859 0 : Self::Broken => write!(f, "Broken"),
860 : }
861 0 : }
862 : }
863 :
864 : /// Arguments to [`Tenant::create_timeline`].
865 : ///
866 : /// Not usable as an idempotency key for timeline creation because if [`CreateTimelineParamsBranch::ancestor_start_lsn`]
867 : /// is `None`, the result of the timeline create call is not deterministic.
868 : ///
869 : /// See [`CreateTimelineIdempotency`] for an idempotency key.
870 : #[derive(Debug)]
871 : pub(crate) enum CreateTimelineParams {
872 : Bootstrap(CreateTimelineParamsBootstrap),
873 : Branch(CreateTimelineParamsBranch),
874 : ImportPgdata(CreateTimelineParamsImportPgdata),
875 : }
876 :
877 : #[derive(Debug)]
878 : pub(crate) struct CreateTimelineParamsBootstrap {
879 : pub(crate) new_timeline_id: TimelineId,
880 : pub(crate) existing_initdb_timeline_id: Option<TimelineId>,
881 : pub(crate) pg_version: u32,
882 : }
883 :
884 : /// NB: See comment on [`CreateTimelineIdempotency::Branch`] for why there's no `pg_version` here.
885 : #[derive(Debug)]
886 : pub(crate) struct CreateTimelineParamsBranch {
887 : pub(crate) new_timeline_id: TimelineId,
888 : pub(crate) ancestor_timeline_id: TimelineId,
889 : pub(crate) ancestor_start_lsn: Option<Lsn>,
890 : }
891 :
892 : #[derive(Debug)]
893 : pub(crate) struct CreateTimelineParamsImportPgdata {
894 : pub(crate) new_timeline_id: TimelineId,
895 : pub(crate) location: import_pgdata::index_part_format::Location,
896 : pub(crate) idempotency_key: import_pgdata::index_part_format::IdempotencyKey,
897 : }
898 :
899 : /// What is used to determine idempotency of a [`Tenant::create_timeline`] call in [`Tenant::start_creating_timeline`] in [`Tenant::start_creating_timeline`].
900 : ///
901 : /// Each [`Timeline`] object holds [`Self`] as an immutable property in [`Timeline::create_idempotency`].
902 : ///
903 : /// We lower timeline creation requests to [`Self`], and then use [`PartialEq::eq`] to compare [`Timeline::create_idempotency`] with the request.
904 : /// If they are equal, we return a reference to the existing timeline, otherwise it's an idempotency conflict.
905 : ///
906 : /// There is special treatment for [`Self::FailWithConflict`] to always return an idempotency conflict.
907 : /// It would be nice to have more advanced derive macros to make that special treatment declarative.
908 : ///
909 : /// Notes:
910 : /// - Unlike [`CreateTimelineParams`], ancestor LSN is fixed, so, branching will be at a deterministic LSN.
911 : /// - We make some trade-offs though, e.g., [`CreateTimelineParamsBootstrap::existing_initdb_timeline_id`]
912 : /// is not considered for idempotency. We can improve on this over time if we deem it necessary.
913 : ///
914 : #[derive(Debug, Clone, PartialEq, Eq)]
915 : pub(crate) enum CreateTimelineIdempotency {
916 : /// NB: special treatment, see comment in [`Self`].
917 : FailWithConflict,
918 : Bootstrap {
919 : pg_version: u32,
920 : },
921 : /// NB: branches always have the same `pg_version` as their ancestor.
922 : /// While [`pageserver_api::models::TimelineCreateRequestMode::Branch::pg_version`]
923 : /// exists as a field, and is set by cplane, it has always been ignored by pageserver when
924 : /// determining the child branch pg_version.
925 : Branch {
926 : ancestor_timeline_id: TimelineId,
927 : ancestor_start_lsn: Lsn,
928 : },
929 : ImportPgdata(CreatingTimelineIdempotencyImportPgdata),
930 : }
931 :
932 : #[derive(Debug, Clone, PartialEq, Eq)]
933 : pub(crate) struct CreatingTimelineIdempotencyImportPgdata {
934 : idempotency_key: import_pgdata::index_part_format::IdempotencyKey,
935 : }
936 :
937 : /// What is returned by [`Tenant::start_creating_timeline`].
938 : #[must_use]
939 : enum StartCreatingTimelineResult {
940 : CreateGuard(TimelineCreateGuard),
941 : Idempotent(Arc<Timeline>),
942 : }
943 :
944 : enum TimelineInitAndSyncResult {
945 : ReadyToActivate(Arc<Timeline>),
946 : NeedsSpawnImportPgdata(TimelineInitAndSyncNeedsSpawnImportPgdata),
947 : }
948 :
949 : impl TimelineInitAndSyncResult {
950 0 : fn ready_to_activate(self) -> Option<Arc<Timeline>> {
951 0 : match self {
952 0 : Self::ReadyToActivate(timeline) => Some(timeline),
953 0 : _ => None,
954 : }
955 0 : }
956 : }
957 :
958 : #[must_use]
959 : struct TimelineInitAndSyncNeedsSpawnImportPgdata {
960 : timeline: Arc<Timeline>,
961 : import_pgdata: import_pgdata::index_part_format::Root,
962 : guard: TimelineCreateGuard,
963 : }
964 :
965 : /// What is returned by [`Tenant::create_timeline`].
966 : enum CreateTimelineResult {
967 : Created(Arc<Timeline>),
968 : Idempotent(Arc<Timeline>),
969 : /// IMPORTANT: This [`Arc<Timeline>`] object is not in [`Tenant::timelines`] when
970 : /// we return this result, nor will this concrete object ever be added there.
971 : /// Cf method comment on [`Tenant::create_timeline_import_pgdata`].
972 : ImportSpawned(Arc<Timeline>),
973 : }
974 :
975 : impl CreateTimelineResult {
976 0 : fn discriminant(&self) -> &'static str {
977 0 : match self {
978 0 : Self::Created(_) => "Created",
979 0 : Self::Idempotent(_) => "Idempotent",
980 0 : Self::ImportSpawned(_) => "ImportSpawned",
981 : }
982 0 : }
983 0 : fn timeline(&self) -> &Arc<Timeline> {
984 0 : match self {
985 0 : Self::Created(t) | Self::Idempotent(t) | Self::ImportSpawned(t) => t,
986 0 : }
987 0 : }
988 : /// Unit test timelines aren't activated, test has to do it if it needs to.
989 : #[cfg(test)]
990 230 : fn into_timeline_for_test(self) -> Arc<Timeline> {
991 230 : match self {
992 230 : Self::Created(t) | Self::Idempotent(t) | Self::ImportSpawned(t) => t,
993 230 : }
994 230 : }
995 : }
996 :
997 : #[derive(thiserror::Error, Debug)]
998 : pub enum CreateTimelineError {
999 : #[error("creation of timeline with the given ID is in progress")]
1000 : AlreadyCreating,
1001 : #[error("timeline already exists with different parameters")]
1002 : Conflict,
1003 : #[error(transparent)]
1004 : AncestorLsn(anyhow::Error),
1005 : #[error("ancestor timeline is not active")]
1006 : AncestorNotActive,
1007 : #[error("ancestor timeline is archived")]
1008 : AncestorArchived,
1009 : #[error("tenant shutting down")]
1010 : ShuttingDown,
1011 : #[error(transparent)]
1012 : Other(#[from] anyhow::Error),
1013 : }
1014 :
1015 : #[derive(thiserror::Error, Debug)]
1016 : pub enum InitdbError {
1017 : #[error("Operation was cancelled")]
1018 : Cancelled,
1019 : #[error(transparent)]
1020 : Other(anyhow::Error),
1021 : #[error(transparent)]
1022 : Inner(postgres_initdb::Error),
1023 : }
1024 :
1025 : enum CreateTimelineCause {
1026 : Load,
1027 : Delete,
1028 : }
1029 :
1030 : enum LoadTimelineCause {
1031 : Attach,
1032 : Unoffload,
1033 : ImportPgdata {
1034 : create_guard: TimelineCreateGuard,
1035 : activate: ActivateTimelineArgs,
1036 : },
1037 : }
1038 :
1039 : #[derive(thiserror::Error, Debug)]
1040 : pub(crate) enum GcError {
1041 : // The tenant is shutting down
1042 : #[error("tenant shutting down")]
1043 : TenantCancelled,
1044 :
1045 : // The tenant is shutting down
1046 : #[error("timeline shutting down")]
1047 : TimelineCancelled,
1048 :
1049 : // The tenant is in a state inelegible to run GC
1050 : #[error("not active")]
1051 : NotActive,
1052 :
1053 : // A requested GC cutoff LSN was invalid, for example it tried to move backwards
1054 : #[error("not active")]
1055 : BadLsn { why: String },
1056 :
1057 : // A remote storage error while scheduling updates after compaction
1058 : #[error(transparent)]
1059 : Remote(anyhow::Error),
1060 :
1061 : // An error reading while calculating GC cutoffs
1062 : #[error(transparent)]
1063 : GcCutoffs(PageReconstructError),
1064 :
1065 : // If GC was invoked for a particular timeline, this error means it didn't exist
1066 : #[error("timeline not found")]
1067 : TimelineNotFound,
1068 : }
1069 :
1070 : impl From<PageReconstructError> for GcError {
1071 0 : fn from(value: PageReconstructError) -> Self {
1072 0 : match value {
1073 0 : PageReconstructError::Cancelled => Self::TimelineCancelled,
1074 0 : other => Self::GcCutoffs(other),
1075 : }
1076 0 : }
1077 : }
1078 :
1079 : impl From<NotInitialized> for GcError {
1080 0 : fn from(value: NotInitialized) -> Self {
1081 0 : match value {
1082 0 : NotInitialized::Uninitialized => GcError::Remote(value.into()),
1083 0 : NotInitialized::Stopped | NotInitialized::ShuttingDown => GcError::TimelineCancelled,
1084 : }
1085 0 : }
1086 : }
1087 :
1088 : impl From<timeline::layer_manager::Shutdown> for GcError {
1089 0 : fn from(_: timeline::layer_manager::Shutdown) -> Self {
1090 0 : GcError::TimelineCancelled
1091 0 : }
1092 : }
1093 :
1094 : #[derive(thiserror::Error, Debug)]
1095 : pub(crate) enum LoadConfigError {
1096 : #[error("TOML deserialization error: '{0}'")]
1097 : DeserializeToml(#[from] toml_edit::de::Error),
1098 :
1099 : #[error("Config not found at {0}")]
1100 : NotFound(Utf8PathBuf),
1101 : }
1102 :
1103 : impl Tenant {
1104 : /// Yet another helper for timeline initialization.
1105 : ///
1106 : /// - Initializes the Timeline struct and inserts it into the tenant's hash map
1107 : /// - Scans the local timeline directory for layer files and builds the layer map
1108 : /// - Downloads remote index file and adds remote files to the layer map
1109 : /// - Schedules remote upload tasks for any files that are present locally but missing from remote storage.
1110 : ///
1111 : /// If the operation fails, the timeline is left in the tenant's hash map in Broken state. On success,
1112 : /// it is marked as Active.
1113 : #[allow(clippy::too_many_arguments)]
1114 6 : async fn timeline_init_and_sync(
1115 6 : self: &Arc<Self>,
1116 6 : timeline_id: TimelineId,
1117 6 : resources: TimelineResources,
1118 6 : mut index_part: IndexPart,
1119 6 : metadata: TimelineMetadata,
1120 6 : ancestor: Option<Arc<Timeline>>,
1121 6 : cause: LoadTimelineCause,
1122 6 : ctx: &RequestContext,
1123 6 : ) -> anyhow::Result<TimelineInitAndSyncResult> {
1124 6 : let tenant_id = self.tenant_shard_id;
1125 6 :
1126 6 : let import_pgdata = index_part.import_pgdata.take();
1127 6 : let idempotency = match &import_pgdata {
1128 0 : Some(import_pgdata) => {
1129 0 : CreateTimelineIdempotency::ImportPgdata(CreatingTimelineIdempotencyImportPgdata {
1130 0 : idempotency_key: import_pgdata.idempotency_key().clone(),
1131 0 : })
1132 : }
1133 : None => {
1134 6 : if metadata.ancestor_timeline().is_none() {
1135 4 : CreateTimelineIdempotency::Bootstrap {
1136 4 : pg_version: metadata.pg_version(),
1137 4 : }
1138 : } else {
1139 2 : CreateTimelineIdempotency::Branch {
1140 2 : ancestor_timeline_id: metadata.ancestor_timeline().unwrap(),
1141 2 : ancestor_start_lsn: metadata.ancestor_lsn(),
1142 2 : }
1143 : }
1144 : }
1145 : };
1146 :
1147 6 : let timeline = self.create_timeline_struct(
1148 6 : timeline_id,
1149 6 : &metadata,
1150 6 : ancestor.clone(),
1151 6 : resources,
1152 6 : CreateTimelineCause::Load,
1153 6 : idempotency.clone(),
1154 6 : )?;
1155 6 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
1156 6 : anyhow::ensure!(
1157 6 : disk_consistent_lsn.is_valid(),
1158 0 : "Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn"
1159 : );
1160 6 : assert_eq!(
1161 6 : disk_consistent_lsn,
1162 6 : metadata.disk_consistent_lsn(),
1163 0 : "these are used interchangeably"
1164 : );
1165 :
1166 6 : timeline.remote_client.init_upload_queue(&index_part)?;
1167 :
1168 6 : timeline
1169 6 : .load_layer_map(disk_consistent_lsn, index_part)
1170 6 : .await
1171 6 : .with_context(|| {
1172 0 : format!("Failed to load layermap for timeline {tenant_id}/{timeline_id}")
1173 6 : })?;
1174 :
1175 0 : match import_pgdata {
1176 0 : Some(import_pgdata) if !import_pgdata.is_done() => {
1177 0 : match cause {
1178 0 : LoadTimelineCause::Attach | LoadTimelineCause::Unoffload => (),
1179 : LoadTimelineCause::ImportPgdata { .. } => {
1180 0 : unreachable!("ImportPgdata should not be reloading timeline import is done and persisted as such in s3")
1181 : }
1182 : }
1183 0 : let mut guard = self.timelines_creating.lock().unwrap();
1184 0 : if !guard.insert(timeline_id) {
1185 : // We should never try and load the same timeline twice during startup
1186 0 : unreachable!("Timeline {tenant_id}/{timeline_id} is already being created")
1187 0 : }
1188 0 : let timeline_create_guard = TimelineCreateGuard {
1189 0 : _tenant_gate_guard: self.gate.enter()?,
1190 0 : owning_tenant: self.clone(),
1191 0 : timeline_id,
1192 0 : idempotency,
1193 0 : // The users of this specific return value don't need the timline_path in there.
1194 0 : timeline_path: timeline
1195 0 : .conf
1196 0 : .timeline_path(&timeline.tenant_shard_id, &timeline.timeline_id),
1197 0 : };
1198 0 : Ok(TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
1199 0 : TimelineInitAndSyncNeedsSpawnImportPgdata {
1200 0 : timeline,
1201 0 : import_pgdata,
1202 0 : guard: timeline_create_guard,
1203 0 : },
1204 0 : ))
1205 : }
1206 : Some(_) | None => {
1207 : {
1208 6 : let mut timelines_accessor = self.timelines.lock().unwrap();
1209 6 : match timelines_accessor.entry(timeline_id) {
1210 : // We should never try and load the same timeline twice during startup
1211 : Entry::Occupied(_) => {
1212 0 : unreachable!(
1213 0 : "Timeline {tenant_id}/{timeline_id} already exists in the tenant map"
1214 0 : );
1215 : }
1216 6 : Entry::Vacant(v) => {
1217 6 : v.insert(Arc::clone(&timeline));
1218 6 : timeline.maybe_spawn_flush_loop();
1219 6 : }
1220 : }
1221 : }
1222 :
1223 : // Sanity check: a timeline should have some content.
1224 6 : anyhow::ensure!(
1225 6 : ancestor.is_some()
1226 4 : || timeline
1227 4 : .layers
1228 4 : .read()
1229 4 : .await
1230 4 : .layer_map()
1231 4 : .expect("currently loading, layer manager cannot be shutdown already")
1232 4 : .iter_historic_layers()
1233 4 : .next()
1234 4 : .is_some(),
1235 0 : "Timeline has no ancestor and no layer files"
1236 : );
1237 :
1238 6 : match cause {
1239 6 : LoadTimelineCause::Attach | LoadTimelineCause::Unoffload => (),
1240 : LoadTimelineCause::ImportPgdata {
1241 0 : create_guard,
1242 0 : activate,
1243 0 : } => {
1244 0 : // TODO: see the comment in the task code above how I'm not so certain
1245 0 : // it is safe to activate here because of concurrent shutdowns.
1246 0 : match activate {
1247 0 : ActivateTimelineArgs::Yes { broker_client } => {
1248 0 : info!("activating timeline after reload from pgdata import task");
1249 0 : timeline.activate(self.clone(), broker_client, None, ctx);
1250 : }
1251 0 : ActivateTimelineArgs::No => (),
1252 : }
1253 0 : drop(create_guard);
1254 : }
1255 : }
1256 :
1257 6 : Ok(TimelineInitAndSyncResult::ReadyToActivate(timeline))
1258 : }
1259 : }
1260 6 : }
1261 :
1262 : /// Attach a tenant that's available in cloud storage.
1263 : ///
1264 : /// This returns quickly, after just creating the in-memory object
1265 : /// Tenant struct and launching a background task to download
1266 : /// the remote index files. On return, the tenant is most likely still in
1267 : /// Attaching state, and it will become Active once the background task
1268 : /// finishes. You can use wait_until_active() to wait for the task to
1269 : /// complete.
1270 : ///
1271 : #[allow(clippy::too_many_arguments)]
1272 0 : pub(crate) fn spawn(
1273 0 : conf: &'static PageServerConf,
1274 0 : tenant_shard_id: TenantShardId,
1275 0 : resources: TenantSharedResources,
1276 0 : attached_conf: AttachedTenantConf,
1277 0 : shard_identity: ShardIdentity,
1278 0 : init_order: Option<InitializationOrder>,
1279 0 : mode: SpawnMode,
1280 0 : ctx: &RequestContext,
1281 0 : ) -> Result<Arc<Tenant>, GlobalShutDown> {
1282 0 : let wal_redo_manager =
1283 0 : WalRedoManager::new(PostgresRedoManager::new(conf, tenant_shard_id))?;
1284 :
1285 : let TenantSharedResources {
1286 0 : broker_client,
1287 0 : remote_storage,
1288 0 : deletion_queue_client,
1289 0 : l0_flush_global_state,
1290 0 : } = resources;
1291 0 :
1292 0 : let attach_mode = attached_conf.location.attach_mode;
1293 0 : let generation = attached_conf.location.generation;
1294 0 :
1295 0 : let tenant = Arc::new(Tenant::new(
1296 0 : TenantState::Attaching,
1297 0 : conf,
1298 0 : attached_conf,
1299 0 : shard_identity,
1300 0 : Some(wal_redo_manager),
1301 0 : tenant_shard_id,
1302 0 : remote_storage.clone(),
1303 0 : deletion_queue_client,
1304 0 : l0_flush_global_state,
1305 0 : ));
1306 0 :
1307 0 : // The attach task will carry a GateGuard, so that shutdown() reliably waits for it to drop out if
1308 0 : // we shut down while attaching.
1309 0 : let attach_gate_guard = tenant
1310 0 : .gate
1311 0 : .enter()
1312 0 : .expect("We just created the Tenant: nothing else can have shut it down yet");
1313 0 :
1314 0 : // Do all the hard work in the background
1315 0 : let tenant_clone = Arc::clone(&tenant);
1316 0 : let ctx = ctx.detached_child(TaskKind::Attach, DownloadBehavior::Warn);
1317 0 : task_mgr::spawn(
1318 0 : &tokio::runtime::Handle::current(),
1319 0 : TaskKind::Attach,
1320 0 : tenant_shard_id,
1321 0 : None,
1322 0 : "attach tenant",
1323 0 : async move {
1324 0 :
1325 0 : info!(
1326 : ?attach_mode,
1327 0 : "Attaching tenant"
1328 : );
1329 :
1330 0 : let _gate_guard = attach_gate_guard;
1331 0 :
1332 0 : // Is this tenant being spawned as part of process startup?
1333 0 : let starting_up = init_order.is_some();
1334 0 : scopeguard::defer! {
1335 0 : if starting_up {
1336 0 : TENANT.startup_complete.inc();
1337 0 : }
1338 0 : }
1339 :
1340 : // Ideally we should use Tenant::set_broken_no_wait, but it is not supposed to be used when tenant is in loading state.
1341 : enum BrokenVerbosity {
1342 : Error,
1343 : Info
1344 : }
1345 0 : let make_broken =
1346 0 : |t: &Tenant, err: anyhow::Error, verbosity: BrokenVerbosity| {
1347 0 : match verbosity {
1348 : BrokenVerbosity::Info => {
1349 0 : info!("attach cancelled, setting tenant state to Broken: {err}");
1350 : },
1351 : BrokenVerbosity::Error => {
1352 0 : error!("attach failed, setting tenant state to Broken: {err:?}");
1353 : }
1354 : }
1355 0 : t.state.send_modify(|state| {
1356 0 : // The Stopping case is for when we have passed control on to DeleteTenantFlow:
1357 0 : // if it errors, we will call make_broken when tenant is already in Stopping.
1358 0 : assert!(
1359 0 : matches!(*state, TenantState::Attaching | TenantState::Stopping { .. }),
1360 0 : "the attach task owns the tenant state until activation is complete"
1361 : );
1362 :
1363 0 : *state = TenantState::broken_from_reason(err.to_string());
1364 0 : });
1365 0 : };
1366 :
1367 : // TODO: should also be rejecting tenant conf changes that violate this check.
1368 0 : if let Err(e) = crate::tenant::storage_layer::inmemory_layer::IndexEntry::validate_checkpoint_distance(tenant_clone.get_checkpoint_distance()) {
1369 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1370 0 : return Ok(());
1371 0 : }
1372 0 :
1373 0 : let mut init_order = init_order;
1374 0 : // take the completion because initial tenant loading will complete when all of
1375 0 : // these tasks complete.
1376 0 : let _completion = init_order
1377 0 : .as_mut()
1378 0 : .and_then(|x| x.initial_tenant_load.take());
1379 0 : let remote_load_completion = init_order
1380 0 : .as_mut()
1381 0 : .and_then(|x| x.initial_tenant_load_remote.take());
1382 :
1383 : enum AttachType<'a> {
1384 : /// We are attaching this tenant lazily in the background.
1385 : Warmup {
1386 : _permit: tokio::sync::SemaphorePermit<'a>,
1387 : during_startup: bool
1388 : },
1389 : /// We are attaching this tenant as soon as we can, because for example an
1390 : /// endpoint tried to access it.
1391 : OnDemand,
1392 : /// During normal operations after startup, we are attaching a tenant, and
1393 : /// eager attach was requested.
1394 : Normal,
1395 : }
1396 :
1397 0 : let attach_type = if matches!(mode, SpawnMode::Lazy) {
1398 : // Before doing any I/O, wait for at least one of:
1399 : // - A client attempting to access to this tenant (on-demand loading)
1400 : // - A permit becoming available in the warmup semaphore (background warmup)
1401 :
1402 0 : tokio::select!(
1403 0 : permit = tenant_clone.activate_now_sem.acquire() => {
1404 0 : let _ = permit.expect("activate_now_sem is never closed");
1405 0 : tracing::info!("Activating tenant (on-demand)");
1406 0 : AttachType::OnDemand
1407 : },
1408 0 : permit = conf.concurrent_tenant_warmup.inner().acquire() => {
1409 0 : let _permit = permit.expect("concurrent_tenant_warmup semaphore is never closed");
1410 0 : tracing::info!("Activating tenant (warmup)");
1411 0 : AttachType::Warmup {
1412 0 : _permit,
1413 0 : during_startup: init_order.is_some()
1414 0 : }
1415 : }
1416 0 : _ = tenant_clone.cancel.cancelled() => {
1417 : // This is safe, but should be pretty rare: it is interesting if a tenant
1418 : // stayed in Activating for such a long time that shutdown found it in
1419 : // that state.
1420 0 : tracing::info!(state=%tenant_clone.current_state(), "Tenant shut down before activation");
1421 : // Make the tenant broken so that set_stopping will not hang waiting for it to leave
1422 : // the Attaching state. This is an over-reaction (nothing really broke, the tenant is
1423 : // just shutting down), but ensures progress.
1424 0 : make_broken(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"), BrokenVerbosity::Info);
1425 0 : return Ok(());
1426 : },
1427 : )
1428 : } else {
1429 : // SpawnMode::{Create,Eager} always cause jumping ahead of the
1430 : // concurrent_tenant_warmup queue
1431 0 : AttachType::Normal
1432 : };
1433 :
1434 0 : let preload = match &mode {
1435 : SpawnMode::Eager | SpawnMode::Lazy => {
1436 0 : let _preload_timer = TENANT.preload.start_timer();
1437 0 : let res = tenant_clone
1438 0 : .preload(&remote_storage, task_mgr::shutdown_token())
1439 0 : .await;
1440 0 : match res {
1441 0 : Ok(p) => Some(p),
1442 0 : Err(e) => {
1443 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1444 0 : return Ok(());
1445 : }
1446 : }
1447 : }
1448 :
1449 : };
1450 :
1451 : // Remote preload is complete.
1452 0 : drop(remote_load_completion);
1453 0 :
1454 0 :
1455 0 : // We will time the duration of the attach phase unless this is a creation (attach will do no work)
1456 0 : let attach_start = std::time::Instant::now();
1457 0 : let attached = {
1458 0 : let _attach_timer = Some(TENANT.attach.start_timer());
1459 0 : tenant_clone.attach(preload, &ctx).await
1460 : };
1461 0 : let attach_duration = attach_start.elapsed();
1462 0 : _ = tenant_clone.attach_wal_lag_cooldown.set(WalLagCooldown::new(attach_start, attach_duration));
1463 0 :
1464 0 : match attached {
1465 : Ok(()) => {
1466 0 : info!("attach finished, activating");
1467 0 : tenant_clone.activate(broker_client, None, &ctx);
1468 : }
1469 0 : Err(e) => {
1470 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1471 0 : }
1472 : }
1473 :
1474 : // If we are doing an opportunistic warmup attachment at startup, initialize
1475 : // logical size at the same time. This is better than starting a bunch of idle tenants
1476 : // with cold caches and then coming back later to initialize their logical sizes.
1477 : //
1478 : // It also prevents the warmup proccess competing with the concurrency limit on
1479 : // logical size calculations: if logical size calculation semaphore is saturated,
1480 : // then warmup will wait for that before proceeding to the next tenant.
1481 0 : if matches!(attach_type, AttachType::Warmup { during_startup: true, .. }) {
1482 0 : let mut futs: FuturesUnordered<_> = tenant_clone.timelines.lock().unwrap().values().cloned().map(|t| t.await_initial_logical_size()).collect();
1483 0 : tracing::info!("Waiting for initial logical sizes while warming up...");
1484 0 : while futs.next().await.is_some() {}
1485 0 : tracing::info!("Warm-up complete");
1486 0 : }
1487 :
1488 0 : Ok(())
1489 0 : }
1490 0 : .instrument(tracing::info_span!(parent: None, "attach", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), gen=?generation)),
1491 : );
1492 0 : Ok(tenant)
1493 0 : }
1494 :
1495 192 : #[instrument(skip_all)]
1496 : pub(crate) async fn preload(
1497 : self: &Arc<Self>,
1498 : remote_storage: &GenericRemoteStorage,
1499 : cancel: CancellationToken,
1500 : ) -> anyhow::Result<TenantPreload> {
1501 : span::debug_assert_current_span_has_tenant_id();
1502 : // Get list of remote timelines
1503 : // download index files for every tenant timeline
1504 : info!("listing remote timelines");
1505 : let (mut remote_timeline_ids, other_keys) = remote_timeline_client::list_remote_timelines(
1506 : remote_storage,
1507 : self.tenant_shard_id,
1508 : cancel.clone(),
1509 : )
1510 : .await?;
1511 : let (offloaded_add, tenant_manifest) =
1512 : match remote_timeline_client::download_tenant_manifest(
1513 : remote_storage,
1514 : &self.tenant_shard_id,
1515 : self.generation,
1516 : &cancel,
1517 : )
1518 : .await
1519 : {
1520 : Ok((tenant_manifest, _generation, _manifest_mtime)) => (
1521 : format!("{} offloaded", tenant_manifest.offloaded_timelines.len()),
1522 : tenant_manifest,
1523 : ),
1524 : Err(DownloadError::NotFound) => {
1525 : ("no manifest".to_string(), TenantManifest::empty())
1526 : }
1527 : Err(e) => Err(e)?,
1528 : };
1529 :
1530 : info!(
1531 : "found {} timelines, and {offloaded_add}",
1532 : remote_timeline_ids.len()
1533 : );
1534 :
1535 : for k in other_keys {
1536 : warn!("Unexpected non timeline key {k}");
1537 : }
1538 :
1539 : // Avoid downloading IndexPart of offloaded timelines.
1540 : let mut offloaded_with_prefix = HashSet::new();
1541 : for offloaded in tenant_manifest.offloaded_timelines.iter() {
1542 : if remote_timeline_ids.remove(&offloaded.timeline_id) {
1543 : offloaded_with_prefix.insert(offloaded.timeline_id);
1544 : } else {
1545 : // We'll take care later of timelines in the manifest without a prefix
1546 : }
1547 : }
1548 :
1549 : let timelines = self
1550 : .load_timelines_metadata(remote_timeline_ids, remote_storage, cancel)
1551 : .await?;
1552 :
1553 : Ok(TenantPreload {
1554 : tenant_manifest,
1555 : timelines: timelines
1556 : .into_iter()
1557 6 : .map(|(id, tl)| (id, Some(tl)))
1558 0 : .chain(offloaded_with_prefix.into_iter().map(|id| (id, None)))
1559 : .collect(),
1560 : })
1561 : }
1562 :
1563 : ///
1564 : /// Background task that downloads all data for a tenant and brings it to Active state.
1565 : ///
1566 : /// No background tasks are started as part of this routine.
1567 : ///
1568 192 : async fn attach(
1569 192 : self: &Arc<Tenant>,
1570 192 : preload: Option<TenantPreload>,
1571 192 : ctx: &RequestContext,
1572 192 : ) -> anyhow::Result<()> {
1573 192 : span::debug_assert_current_span_has_tenant_id();
1574 192 :
1575 192 : failpoint_support::sleep_millis_async!("before-attaching-tenant");
1576 :
1577 192 : let Some(preload) = preload else {
1578 0 : anyhow::bail!("local-only deployment is no longer supported, https://github.com/neondatabase/neon/issues/5624");
1579 : };
1580 :
1581 192 : let mut offloaded_timeline_ids = HashSet::new();
1582 192 : let mut offloaded_timelines_list = Vec::new();
1583 192 : for timeline_manifest in preload.tenant_manifest.offloaded_timelines.iter() {
1584 0 : let timeline_id = timeline_manifest.timeline_id;
1585 0 : let offloaded_timeline =
1586 0 : OffloadedTimeline::from_manifest(self.tenant_shard_id, timeline_manifest);
1587 0 : offloaded_timelines_list.push((timeline_id, Arc::new(offloaded_timeline)));
1588 0 : offloaded_timeline_ids.insert(timeline_id);
1589 0 : }
1590 : // Complete deletions for offloaded timeline id's from manifest.
1591 : // The manifest will be uploaded later in this function.
1592 192 : offloaded_timelines_list
1593 192 : .retain(|(offloaded_id, offloaded)| {
1594 0 : // Existence of a timeline is finally determined by the existence of an index-part.json in remote storage.
1595 0 : // If there is dangling references in another location, they need to be cleaned up.
1596 0 : let delete = !preload.timelines.contains_key(offloaded_id);
1597 0 : if delete {
1598 0 : tracing::info!("Removing offloaded timeline {offloaded_id} from manifest as no remote prefix was found");
1599 0 : offloaded.defuse_for_tenant_drop();
1600 0 : }
1601 0 : !delete
1602 192 : });
1603 192 :
1604 192 : let mut timelines_to_resume_deletions = vec![];
1605 192 :
1606 192 : let mut remote_index_and_client = HashMap::new();
1607 192 : let mut timeline_ancestors = HashMap::new();
1608 192 : let mut existent_timelines = HashSet::new();
1609 198 : for (timeline_id, preload) in preload.timelines {
1610 6 : let Some(preload) = preload else { continue };
1611 : // This is an invariant of the `preload` function's API
1612 6 : assert!(!offloaded_timeline_ids.contains(&timeline_id));
1613 6 : let index_part = match preload.index_part {
1614 6 : Ok(i) => {
1615 6 : debug!("remote index part exists for timeline {timeline_id}");
1616 : // We found index_part on the remote, this is the standard case.
1617 6 : existent_timelines.insert(timeline_id);
1618 6 : i
1619 : }
1620 : Err(DownloadError::NotFound) => {
1621 : // There is no index_part on the remote. We only get here
1622 : // if there is some prefix for the timeline in the remote storage.
1623 : // This can e.g. be the initdb.tar.zst archive, maybe a
1624 : // remnant from a prior incomplete creation or deletion attempt.
1625 : // Delete the local directory as the deciding criterion for a
1626 : // timeline's existence is presence of index_part.
1627 0 : info!(%timeline_id, "index_part not found on remote");
1628 0 : continue;
1629 : }
1630 0 : Err(DownloadError::Fatal(why)) => {
1631 0 : // If, while loading one remote timeline, we saw an indication that our generation
1632 0 : // number is likely invalid, then we should not load the whole tenant.
1633 0 : error!(%timeline_id, "Fatal error loading timeline: {why}");
1634 0 : anyhow::bail!(why.to_string());
1635 : }
1636 0 : Err(e) => {
1637 0 : // Some (possibly ephemeral) error happened during index_part download.
1638 0 : // Pretend the timeline exists to not delete the timeline directory,
1639 0 : // as it might be a temporary issue and we don't want to re-download
1640 0 : // everything after it resolves.
1641 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1642 :
1643 0 : existent_timelines.insert(timeline_id);
1644 0 : continue;
1645 : }
1646 : };
1647 6 : match index_part {
1648 6 : MaybeDeletedIndexPart::IndexPart(index_part) => {
1649 6 : timeline_ancestors.insert(timeline_id, index_part.metadata.clone());
1650 6 : remote_index_and_client.insert(timeline_id, (index_part, preload.client));
1651 6 : }
1652 0 : MaybeDeletedIndexPart::Deleted(index_part) => {
1653 0 : info!(
1654 0 : "timeline {} is deleted, picking to resume deletion",
1655 : timeline_id
1656 : );
1657 0 : timelines_to_resume_deletions.push((timeline_id, index_part, preload.client));
1658 : }
1659 : }
1660 : }
1661 :
1662 192 : let mut gc_blocks = HashMap::new();
1663 :
1664 : // For every timeline, download the metadata file, scan the local directory,
1665 : // and build a layer map that contains an entry for each remote and local
1666 : // layer file.
1667 192 : let sorted_timelines = tree_sort_timelines(timeline_ancestors, |m| m.ancestor_timeline())?;
1668 198 : for (timeline_id, remote_metadata) in sorted_timelines {
1669 6 : let (index_part, remote_client) = remote_index_and_client
1670 6 : .remove(&timeline_id)
1671 6 : .expect("just put it in above");
1672 :
1673 6 : if let Some(blocking) = index_part.gc_blocking.as_ref() {
1674 : // could just filter these away, but it helps while testing
1675 0 : anyhow::ensure!(
1676 0 : !blocking.reasons.is_empty(),
1677 0 : "index_part for {timeline_id} is malformed: it should not have gc blocking with zero reasons"
1678 : );
1679 0 : let prev = gc_blocks.insert(timeline_id, blocking.reasons);
1680 0 : assert!(prev.is_none());
1681 6 : }
1682 :
1683 : // TODO again handle early failure
1684 6 : let effect = self
1685 6 : .load_remote_timeline(
1686 6 : timeline_id,
1687 6 : index_part,
1688 6 : remote_metadata,
1689 6 : TimelineResources {
1690 6 : remote_client,
1691 6 : pagestream_throttle: self.pagestream_throttle.clone(),
1692 6 : l0_flush_global_state: self.l0_flush_global_state.clone(),
1693 6 : },
1694 6 : LoadTimelineCause::Attach,
1695 6 : ctx,
1696 6 : )
1697 6 : .await
1698 6 : .with_context(|| {
1699 0 : format!(
1700 0 : "failed to load remote timeline {} for tenant {}",
1701 0 : timeline_id, self.tenant_shard_id
1702 0 : )
1703 6 : })?;
1704 :
1705 6 : match effect {
1706 6 : TimelineInitAndSyncResult::ReadyToActivate(_) => {
1707 6 : // activation happens later, on Tenant::activate
1708 6 : }
1709 : TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
1710 : TimelineInitAndSyncNeedsSpawnImportPgdata {
1711 0 : timeline,
1712 0 : import_pgdata,
1713 0 : guard,
1714 0 : },
1715 0 : ) => {
1716 0 : tokio::task::spawn(self.clone().create_timeline_import_pgdata_task(
1717 0 : timeline,
1718 0 : import_pgdata,
1719 0 : ActivateTimelineArgs::No,
1720 0 : guard,
1721 0 : ));
1722 0 : }
1723 : }
1724 : }
1725 :
1726 : // Walk through deleted timelines, resume deletion
1727 192 : for (timeline_id, index_part, remote_timeline_client) in timelines_to_resume_deletions {
1728 0 : remote_timeline_client
1729 0 : .init_upload_queue_stopped_to_continue_deletion(&index_part)
1730 0 : .context("init queue stopped")
1731 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1732 :
1733 0 : DeleteTimelineFlow::resume_deletion(
1734 0 : Arc::clone(self),
1735 0 : timeline_id,
1736 0 : &index_part.metadata,
1737 0 : remote_timeline_client,
1738 0 : )
1739 0 : .instrument(tracing::info_span!("timeline_delete", %timeline_id))
1740 0 : .await
1741 0 : .context("resume_deletion")
1742 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1743 : }
1744 192 : let needs_manifest_upload =
1745 192 : offloaded_timelines_list.len() != preload.tenant_manifest.offloaded_timelines.len();
1746 192 : {
1747 192 : let mut offloaded_timelines_accessor = self.timelines_offloaded.lock().unwrap();
1748 192 : offloaded_timelines_accessor.extend(offloaded_timelines_list.into_iter());
1749 192 : }
1750 192 : if needs_manifest_upload {
1751 0 : self.store_tenant_manifest().await?;
1752 192 : }
1753 :
1754 : // The local filesystem contents are a cache of what's in the remote IndexPart;
1755 : // IndexPart is the source of truth.
1756 192 : self.clean_up_timelines(&existent_timelines)?;
1757 :
1758 192 : self.gc_block.set_scanned(gc_blocks);
1759 192 :
1760 192 : fail::fail_point!("attach-before-activate", |_| {
1761 0 : anyhow::bail!("attach-before-activate");
1762 192 : });
1763 192 : failpoint_support::sleep_millis_async!("attach-before-activate-sleep", &self.cancel);
1764 :
1765 192 : info!("Done");
1766 :
1767 192 : Ok(())
1768 192 : }
1769 :
1770 : /// Check for any local timeline directories that are temporary, or do not correspond to a
1771 : /// timeline that still exists: this can happen if we crashed during a deletion/creation, or
1772 : /// if a timeline was deleted while the tenant was attached to a different pageserver.
1773 192 : fn clean_up_timelines(&self, existent_timelines: &HashSet<TimelineId>) -> anyhow::Result<()> {
1774 192 : let timelines_dir = self.conf.timelines_path(&self.tenant_shard_id);
1775 :
1776 192 : let entries = match timelines_dir.read_dir_utf8() {
1777 192 : Ok(d) => d,
1778 0 : Err(e) => {
1779 0 : if e.kind() == std::io::ErrorKind::NotFound {
1780 0 : return Ok(());
1781 : } else {
1782 0 : return Err(e).context("list timelines directory for tenant");
1783 : }
1784 : }
1785 : };
1786 :
1787 200 : for entry in entries {
1788 8 : let entry = entry.context("read timeline dir entry")?;
1789 8 : let entry_path = entry.path();
1790 :
1791 8 : let purge = if crate::is_temporary(entry_path)
1792 : // TODO: remove uninit mark code (https://github.com/neondatabase/neon/issues/5718)
1793 8 : || is_uninit_mark(entry_path)
1794 8 : || crate::is_delete_mark(entry_path)
1795 : {
1796 0 : true
1797 : } else {
1798 8 : match TimelineId::try_from(entry_path.file_name()) {
1799 8 : Ok(i) => {
1800 8 : // Purge if the timeline ID does not exist in remote storage: remote storage is the authority.
1801 8 : !existent_timelines.contains(&i)
1802 : }
1803 0 : Err(e) => {
1804 0 : tracing::warn!(
1805 0 : "Unparseable directory in timelines directory: {entry_path}, ignoring ({e})"
1806 : );
1807 : // Do not purge junk: if we don't recognize it, be cautious and leave it for a human.
1808 0 : false
1809 : }
1810 : }
1811 : };
1812 :
1813 8 : if purge {
1814 2 : tracing::info!("Purging stale timeline dentry {entry_path}");
1815 2 : if let Err(e) = match entry.file_type() {
1816 2 : Ok(t) => if t.is_dir() {
1817 2 : std::fs::remove_dir_all(entry_path)
1818 : } else {
1819 0 : std::fs::remove_file(entry_path)
1820 : }
1821 2 : .or_else(fs_ext::ignore_not_found),
1822 0 : Err(e) => Err(e),
1823 : } {
1824 0 : tracing::warn!("Failed to purge stale timeline dentry {entry_path}: {e}");
1825 2 : }
1826 6 : }
1827 : }
1828 :
1829 192 : Ok(())
1830 192 : }
1831 :
1832 : /// Get sum of all remote timelines sizes
1833 : ///
1834 : /// This function relies on the index_part instead of listing the remote storage
1835 0 : pub fn remote_size(&self) -> u64 {
1836 0 : let mut size = 0;
1837 :
1838 0 : for timeline in self.list_timelines() {
1839 0 : size += timeline.remote_client.get_remote_physical_size();
1840 0 : }
1841 :
1842 0 : size
1843 0 : }
1844 :
1845 6 : #[instrument(skip_all, fields(timeline_id=%timeline_id))]
1846 : async fn load_remote_timeline(
1847 : self: &Arc<Self>,
1848 : timeline_id: TimelineId,
1849 : index_part: IndexPart,
1850 : remote_metadata: TimelineMetadata,
1851 : resources: TimelineResources,
1852 : cause: LoadTimelineCause,
1853 : ctx: &RequestContext,
1854 : ) -> anyhow::Result<TimelineInitAndSyncResult> {
1855 : span::debug_assert_current_span_has_tenant_id();
1856 :
1857 : info!("downloading index file for timeline {}", timeline_id);
1858 : tokio::fs::create_dir_all(self.conf.timeline_path(&self.tenant_shard_id, &timeline_id))
1859 : .await
1860 : .context("Failed to create new timeline directory")?;
1861 :
1862 : let ancestor = if let Some(ancestor_id) = remote_metadata.ancestor_timeline() {
1863 : let timelines = self.timelines.lock().unwrap();
1864 : Some(Arc::clone(timelines.get(&ancestor_id).ok_or_else(
1865 0 : || {
1866 0 : anyhow::anyhow!(
1867 0 : "cannot find ancestor timeline {ancestor_id} for timeline {timeline_id}"
1868 0 : )
1869 0 : },
1870 : )?))
1871 : } else {
1872 : None
1873 : };
1874 :
1875 : self.timeline_init_and_sync(
1876 : timeline_id,
1877 : resources,
1878 : index_part,
1879 : remote_metadata,
1880 : ancestor,
1881 : cause,
1882 : ctx,
1883 : )
1884 : .await
1885 : }
1886 :
1887 192 : async fn load_timelines_metadata(
1888 192 : self: &Arc<Tenant>,
1889 192 : timeline_ids: HashSet<TimelineId>,
1890 192 : remote_storage: &GenericRemoteStorage,
1891 192 : cancel: CancellationToken,
1892 192 : ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
1893 192 : let mut part_downloads = JoinSet::new();
1894 198 : for timeline_id in timeline_ids {
1895 6 : let cancel_clone = cancel.clone();
1896 6 : part_downloads.spawn(
1897 6 : self.load_timeline_metadata(timeline_id, remote_storage.clone(), cancel_clone)
1898 6 : .instrument(info_span!("download_index_part", %timeline_id)),
1899 : );
1900 : }
1901 :
1902 192 : let mut timeline_preloads: HashMap<TimelineId, TimelinePreload> = HashMap::new();
1903 :
1904 : loop {
1905 198 : tokio::select!(
1906 198 : next = part_downloads.join_next() => {
1907 198 : match next {
1908 6 : Some(result) => {
1909 6 : let preload = result.context("join preload task")?;
1910 6 : timeline_preloads.insert(preload.timeline_id, preload);
1911 : },
1912 : None => {
1913 192 : break;
1914 : }
1915 : }
1916 : },
1917 198 : _ = cancel.cancelled() => {
1918 0 : anyhow::bail!("Cancelled while waiting for remote index download")
1919 : }
1920 : )
1921 : }
1922 :
1923 192 : Ok(timeline_preloads)
1924 192 : }
1925 :
1926 6 : fn build_timeline_client(
1927 6 : &self,
1928 6 : timeline_id: TimelineId,
1929 6 : remote_storage: GenericRemoteStorage,
1930 6 : ) -> RemoteTimelineClient {
1931 6 : RemoteTimelineClient::new(
1932 6 : remote_storage.clone(),
1933 6 : self.deletion_queue_client.clone(),
1934 6 : self.conf,
1935 6 : self.tenant_shard_id,
1936 6 : timeline_id,
1937 6 : self.generation,
1938 6 : &self.tenant_conf.load().location,
1939 6 : )
1940 6 : }
1941 :
1942 6 : fn load_timeline_metadata(
1943 6 : self: &Arc<Tenant>,
1944 6 : timeline_id: TimelineId,
1945 6 : remote_storage: GenericRemoteStorage,
1946 6 : cancel: CancellationToken,
1947 6 : ) -> impl Future<Output = TimelinePreload> {
1948 6 : let client = self.build_timeline_client(timeline_id, remote_storage);
1949 6 : async move {
1950 6 : debug_assert_current_span_has_tenant_and_timeline_id();
1951 6 : debug!("starting index part download");
1952 :
1953 6 : let index_part = client.download_index_file(&cancel).await;
1954 :
1955 6 : debug!("finished index part download");
1956 :
1957 6 : TimelinePreload {
1958 6 : client,
1959 6 : timeline_id,
1960 6 : index_part,
1961 6 : }
1962 6 : }
1963 6 : }
1964 :
1965 0 : fn check_to_be_archived_has_no_unarchived_children(
1966 0 : timeline_id: TimelineId,
1967 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
1968 0 : ) -> Result<(), TimelineArchivalError> {
1969 0 : let children: Vec<TimelineId> = timelines
1970 0 : .iter()
1971 0 : .filter_map(|(id, entry)| {
1972 0 : if entry.get_ancestor_timeline_id() != Some(timeline_id) {
1973 0 : return None;
1974 0 : }
1975 0 : if entry.is_archived() == Some(true) {
1976 0 : return None;
1977 0 : }
1978 0 : Some(*id)
1979 0 : })
1980 0 : .collect();
1981 0 :
1982 0 : if !children.is_empty() {
1983 0 : return Err(TimelineArchivalError::HasUnarchivedChildren(children));
1984 0 : }
1985 0 : Ok(())
1986 0 : }
1987 :
1988 0 : fn check_ancestor_of_to_be_unarchived_is_not_archived(
1989 0 : ancestor_timeline_id: TimelineId,
1990 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
1991 0 : offloaded_timelines: &std::sync::MutexGuard<
1992 0 : '_,
1993 0 : HashMap<TimelineId, Arc<OffloadedTimeline>>,
1994 0 : >,
1995 0 : ) -> Result<(), TimelineArchivalError> {
1996 0 : let has_archived_parent =
1997 0 : if let Some(ancestor_timeline) = timelines.get(&ancestor_timeline_id) {
1998 0 : ancestor_timeline.is_archived() == Some(true)
1999 0 : } else if offloaded_timelines.contains_key(&ancestor_timeline_id) {
2000 0 : true
2001 : } else {
2002 0 : error!("ancestor timeline {ancestor_timeline_id} not found");
2003 0 : if cfg!(debug_assertions) {
2004 0 : panic!("ancestor timeline {ancestor_timeline_id} not found");
2005 0 : }
2006 0 : return Err(TimelineArchivalError::NotFound);
2007 : };
2008 0 : if has_archived_parent {
2009 0 : return Err(TimelineArchivalError::HasArchivedParent(
2010 0 : ancestor_timeline_id,
2011 0 : ));
2012 0 : }
2013 0 : Ok(())
2014 0 : }
2015 :
2016 0 : fn check_to_be_unarchived_timeline_has_no_archived_parent(
2017 0 : timeline: &Arc<Timeline>,
2018 0 : ) -> Result<(), TimelineArchivalError> {
2019 0 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
2020 0 : if ancestor_timeline.is_archived() == Some(true) {
2021 0 : return Err(TimelineArchivalError::HasArchivedParent(
2022 0 : ancestor_timeline.timeline_id,
2023 0 : ));
2024 0 : }
2025 0 : }
2026 0 : Ok(())
2027 0 : }
2028 :
2029 : /// Loads the specified (offloaded) timeline from S3 and attaches it as a loaded timeline
2030 : ///
2031 : /// Counterpart to [`offload_timeline`].
2032 0 : async fn unoffload_timeline(
2033 0 : self: &Arc<Self>,
2034 0 : timeline_id: TimelineId,
2035 0 : broker_client: storage_broker::BrokerClientChannel,
2036 0 : ctx: RequestContext,
2037 0 : ) -> Result<Arc<Timeline>, TimelineArchivalError> {
2038 0 : info!("unoffloading timeline");
2039 :
2040 : // We activate the timeline below manually, so this must be called on an active timeline.
2041 : // We expect callers of this function to ensure this.
2042 0 : match self.current_state() {
2043 : TenantState::Activating { .. }
2044 : | TenantState::Attaching
2045 : | TenantState::Broken { .. } => {
2046 0 : panic!("Timeline expected to be active")
2047 : }
2048 0 : TenantState::Stopping { .. } => return Err(TimelineArchivalError::Cancelled),
2049 0 : TenantState::Active => {}
2050 0 : }
2051 0 : let cancel = self.cancel.clone();
2052 0 :
2053 0 : // Protect against concurrent attempts to use this TimelineId
2054 0 : // We don't care much about idempotency, as it's ensured a layer above.
2055 0 : let allow_offloaded = true;
2056 0 : let _create_guard = self
2057 0 : .create_timeline_create_guard(
2058 0 : timeline_id,
2059 0 : CreateTimelineIdempotency::FailWithConflict,
2060 0 : allow_offloaded,
2061 0 : )
2062 0 : .map_err(|err| match err {
2063 0 : TimelineExclusionError::AlreadyCreating => TimelineArchivalError::AlreadyInProgress,
2064 : TimelineExclusionError::AlreadyExists { .. } => {
2065 0 : TimelineArchivalError::Other(anyhow::anyhow!("Timeline already exists"))
2066 : }
2067 0 : TimelineExclusionError::Other(e) => TimelineArchivalError::Other(e),
2068 0 : TimelineExclusionError::ShuttingDown => TimelineArchivalError::Cancelled,
2069 0 : })?;
2070 :
2071 0 : let timeline_preload = self
2072 0 : .load_timeline_metadata(timeline_id, self.remote_storage.clone(), cancel.clone())
2073 0 : .await;
2074 :
2075 0 : let index_part = match timeline_preload.index_part {
2076 0 : Ok(index_part) => {
2077 0 : debug!("remote index part exists for timeline {timeline_id}");
2078 0 : index_part
2079 : }
2080 : Err(DownloadError::NotFound) => {
2081 0 : error!(%timeline_id, "index_part not found on remote");
2082 0 : return Err(TimelineArchivalError::NotFound);
2083 : }
2084 0 : Err(DownloadError::Cancelled) => return Err(TimelineArchivalError::Cancelled),
2085 0 : Err(e) => {
2086 0 : // Some (possibly ephemeral) error happened during index_part download.
2087 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
2088 0 : return Err(TimelineArchivalError::Other(
2089 0 : anyhow::Error::new(e).context("downloading index_part from remote storage"),
2090 0 : ));
2091 : }
2092 : };
2093 0 : let index_part = match index_part {
2094 0 : MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
2095 0 : MaybeDeletedIndexPart::Deleted(_index_part) => {
2096 0 : info!("timeline is deleted according to index_part.json");
2097 0 : return Err(TimelineArchivalError::NotFound);
2098 : }
2099 : };
2100 0 : let remote_metadata = index_part.metadata.clone();
2101 0 : let timeline_resources = self.build_timeline_resources(timeline_id);
2102 0 : self.load_remote_timeline(
2103 0 : timeline_id,
2104 0 : index_part,
2105 0 : remote_metadata,
2106 0 : timeline_resources,
2107 0 : LoadTimelineCause::Unoffload,
2108 0 : &ctx,
2109 0 : )
2110 0 : .await
2111 0 : .with_context(|| {
2112 0 : format!(
2113 0 : "failed to load remote timeline {} for tenant {}",
2114 0 : timeline_id, self.tenant_shard_id
2115 0 : )
2116 0 : })
2117 0 : .map_err(TimelineArchivalError::Other)?;
2118 :
2119 0 : let timeline = {
2120 0 : let timelines = self.timelines.lock().unwrap();
2121 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2122 0 : warn!("timeline not available directly after attach");
2123 : // This is not a panic because no locks are held between `load_remote_timeline`
2124 : // which puts the timeline into timelines, and our look into the timeline map.
2125 0 : return Err(TimelineArchivalError::Other(anyhow::anyhow!(
2126 0 : "timeline not available directly after attach"
2127 0 : )));
2128 : };
2129 0 : let mut offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2130 0 : match offloaded_timelines.remove(&timeline_id) {
2131 0 : Some(offloaded) => {
2132 0 : offloaded.delete_from_ancestor_with_timelines(&timelines);
2133 0 : }
2134 0 : None => warn!("timeline already removed from offloaded timelines"),
2135 : }
2136 :
2137 0 : self.initialize_gc_info(&timelines, &offloaded_timelines, Some(timeline_id));
2138 0 :
2139 0 : Arc::clone(timeline)
2140 0 : };
2141 0 :
2142 0 : // Upload new list of offloaded timelines to S3
2143 0 : self.store_tenant_manifest().await?;
2144 :
2145 : // Activate the timeline (if it makes sense)
2146 0 : if !(timeline.is_broken() || timeline.is_stopping()) {
2147 0 : let background_jobs_can_start = None;
2148 0 : timeline.activate(
2149 0 : self.clone(),
2150 0 : broker_client.clone(),
2151 0 : background_jobs_can_start,
2152 0 : &ctx,
2153 0 : );
2154 0 : }
2155 :
2156 0 : info!("timeline unoffloading complete");
2157 0 : Ok(timeline)
2158 0 : }
2159 :
2160 0 : pub(crate) async fn apply_timeline_archival_config(
2161 0 : self: &Arc<Self>,
2162 0 : timeline_id: TimelineId,
2163 0 : new_state: TimelineArchivalState,
2164 0 : broker_client: storage_broker::BrokerClientChannel,
2165 0 : ctx: RequestContext,
2166 0 : ) -> Result<(), TimelineArchivalError> {
2167 0 : info!("setting timeline archival config");
2168 : // First part: figure out what is needed to do, and do validation
2169 0 : let timeline_or_unarchive_offloaded = 'outer: {
2170 0 : let timelines = self.timelines.lock().unwrap();
2171 :
2172 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2173 0 : let offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2174 0 : let Some(offloaded) = offloaded_timelines.get(&timeline_id) else {
2175 0 : return Err(TimelineArchivalError::NotFound);
2176 : };
2177 0 : if new_state == TimelineArchivalState::Archived {
2178 : // It's offloaded already, so nothing to do
2179 0 : return Ok(());
2180 0 : }
2181 0 : if let Some(ancestor_timeline_id) = offloaded.ancestor_timeline_id {
2182 0 : Self::check_ancestor_of_to_be_unarchived_is_not_archived(
2183 0 : ancestor_timeline_id,
2184 0 : &timelines,
2185 0 : &offloaded_timelines,
2186 0 : )?;
2187 0 : }
2188 0 : break 'outer None;
2189 : };
2190 :
2191 : // Do some validation. We release the timelines lock below, so there is potential
2192 : // for race conditions: these checks are more present to prevent misunderstandings of
2193 : // the API's capabilities, instead of serving as the sole way to defend their invariants.
2194 0 : match new_state {
2195 : TimelineArchivalState::Unarchived => {
2196 0 : Self::check_to_be_unarchived_timeline_has_no_archived_parent(timeline)?
2197 : }
2198 : TimelineArchivalState::Archived => {
2199 0 : Self::check_to_be_archived_has_no_unarchived_children(timeline_id, &timelines)?
2200 : }
2201 : }
2202 0 : Some(Arc::clone(timeline))
2203 : };
2204 :
2205 : // Second part: unoffload timeline (if needed)
2206 0 : let timeline = if let Some(timeline) = timeline_or_unarchive_offloaded {
2207 0 : timeline
2208 : } else {
2209 : // Turn offloaded timeline into a non-offloaded one
2210 0 : self.unoffload_timeline(timeline_id, broker_client, ctx)
2211 0 : .await?
2212 : };
2213 :
2214 : // Third part: upload new timeline archival state and block until it is present in S3
2215 0 : let upload_needed = match timeline
2216 0 : .remote_client
2217 0 : .schedule_index_upload_for_timeline_archival_state(new_state)
2218 : {
2219 0 : Ok(upload_needed) => upload_needed,
2220 0 : Err(e) => {
2221 0 : if timeline.cancel.is_cancelled() {
2222 0 : return Err(TimelineArchivalError::Cancelled);
2223 : } else {
2224 0 : return Err(TimelineArchivalError::Other(e));
2225 : }
2226 : }
2227 : };
2228 :
2229 0 : if upload_needed {
2230 0 : info!("Uploading new state");
2231 : const MAX_WAIT: Duration = Duration::from_secs(10);
2232 0 : let Ok(v) =
2233 0 : tokio::time::timeout(MAX_WAIT, timeline.remote_client.wait_completion()).await
2234 : else {
2235 0 : tracing::warn!("reached timeout for waiting on upload queue");
2236 0 : return Err(TimelineArchivalError::Timeout);
2237 : };
2238 0 : v.map_err(|e| match e {
2239 0 : WaitCompletionError::NotInitialized(e) => {
2240 0 : TimelineArchivalError::Other(anyhow::anyhow!(e))
2241 : }
2242 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2243 0 : TimelineArchivalError::Cancelled
2244 : }
2245 0 : })?;
2246 0 : }
2247 0 : Ok(())
2248 0 : }
2249 :
2250 2 : pub fn get_offloaded_timeline(
2251 2 : &self,
2252 2 : timeline_id: TimelineId,
2253 2 : ) -> Result<Arc<OffloadedTimeline>, GetTimelineError> {
2254 2 : self.timelines_offloaded
2255 2 : .lock()
2256 2 : .unwrap()
2257 2 : .get(&timeline_id)
2258 2 : .map(Arc::clone)
2259 2 : .ok_or(GetTimelineError::NotFound {
2260 2 : tenant_id: self.tenant_shard_id,
2261 2 : timeline_id,
2262 2 : })
2263 2 : }
2264 :
2265 4 : pub(crate) fn tenant_shard_id(&self) -> TenantShardId {
2266 4 : self.tenant_shard_id
2267 4 : }
2268 :
2269 : /// Get Timeline handle for given Neon timeline ID.
2270 : /// This function is idempotent. It doesn't change internal state in any way.
2271 222 : pub fn get_timeline(
2272 222 : &self,
2273 222 : timeline_id: TimelineId,
2274 222 : active_only: bool,
2275 222 : ) -> Result<Arc<Timeline>, GetTimelineError> {
2276 222 : let timelines_accessor = self.timelines.lock().unwrap();
2277 222 : let timeline = timelines_accessor
2278 222 : .get(&timeline_id)
2279 222 : .ok_or(GetTimelineError::NotFound {
2280 222 : tenant_id: self.tenant_shard_id,
2281 222 : timeline_id,
2282 222 : })?;
2283 :
2284 220 : if active_only && !timeline.is_active() {
2285 0 : Err(GetTimelineError::NotActive {
2286 0 : tenant_id: self.tenant_shard_id,
2287 0 : timeline_id,
2288 0 : state: timeline.current_state(),
2289 0 : })
2290 : } else {
2291 220 : Ok(Arc::clone(timeline))
2292 : }
2293 222 : }
2294 :
2295 : /// Lists timelines the tenant contains.
2296 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2297 0 : pub fn list_timelines(&self) -> Vec<Arc<Timeline>> {
2298 0 : self.timelines
2299 0 : .lock()
2300 0 : .unwrap()
2301 0 : .values()
2302 0 : .map(Arc::clone)
2303 0 : .collect()
2304 0 : }
2305 :
2306 : /// Lists timelines the tenant manages, including offloaded ones.
2307 : ///
2308 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2309 0 : pub fn list_timelines_and_offloaded(
2310 0 : &self,
2311 0 : ) -> (Vec<Arc<Timeline>>, Vec<Arc<OffloadedTimeline>>) {
2312 0 : let timelines = self
2313 0 : .timelines
2314 0 : .lock()
2315 0 : .unwrap()
2316 0 : .values()
2317 0 : .map(Arc::clone)
2318 0 : .collect();
2319 0 : let offloaded = self
2320 0 : .timelines_offloaded
2321 0 : .lock()
2322 0 : .unwrap()
2323 0 : .values()
2324 0 : .map(Arc::clone)
2325 0 : .collect();
2326 0 : (timelines, offloaded)
2327 0 : }
2328 :
2329 0 : pub fn list_timeline_ids(&self) -> Vec<TimelineId> {
2330 0 : self.timelines.lock().unwrap().keys().cloned().collect()
2331 0 : }
2332 :
2333 : /// This is used by tests & import-from-basebackup.
2334 : ///
2335 : /// The returned [`UninitializedTimeline`] contains no data nor metadata and it is in
2336 : /// a state that will fail [`Tenant::load_remote_timeline`] because `disk_consistent_lsn=Lsn(0)`.
2337 : ///
2338 : /// The caller is responsible for getting the timeline into a state that will be accepted
2339 : /// by [`Tenant::load_remote_timeline`] / [`Tenant::attach`].
2340 : /// Then they may call [`UninitializedTimeline::finish_creation`] to add the timeline
2341 : /// to the [`Tenant::timelines`].
2342 : ///
2343 : /// Tests should use `Tenant::create_test_timeline` to set up the minimum required metadata keys.
2344 184 : pub(crate) async fn create_empty_timeline(
2345 184 : self: &Arc<Self>,
2346 184 : new_timeline_id: TimelineId,
2347 184 : initdb_lsn: Lsn,
2348 184 : pg_version: u32,
2349 184 : _ctx: &RequestContext,
2350 184 : ) -> anyhow::Result<UninitializedTimeline> {
2351 184 : anyhow::ensure!(
2352 184 : self.is_active(),
2353 0 : "Cannot create empty timelines on inactive tenant"
2354 : );
2355 :
2356 : // Protect against concurrent attempts to use this TimelineId
2357 184 : let create_guard = match self
2358 184 : .start_creating_timeline(new_timeline_id, CreateTimelineIdempotency::FailWithConflict)
2359 184 : .await?
2360 : {
2361 182 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2362 : StartCreatingTimelineResult::Idempotent(_) => {
2363 0 : unreachable!("FailWithConflict implies we get an error instead")
2364 : }
2365 : };
2366 :
2367 182 : let new_metadata = TimelineMetadata::new(
2368 182 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2369 182 : // make it valid, before calling finish_creation()
2370 182 : Lsn(0),
2371 182 : None,
2372 182 : None,
2373 182 : Lsn(0),
2374 182 : initdb_lsn,
2375 182 : initdb_lsn,
2376 182 : pg_version,
2377 182 : );
2378 182 : self.prepare_new_timeline(
2379 182 : new_timeline_id,
2380 182 : &new_metadata,
2381 182 : create_guard,
2382 182 : initdb_lsn,
2383 182 : None,
2384 182 : )
2385 182 : .await
2386 184 : }
2387 :
2388 : /// Helper for unit tests to create an empty timeline.
2389 : ///
2390 : /// The timeline is has state value `Active` but its background loops are not running.
2391 : // This makes the various functions which anyhow::ensure! for Active state work in tests.
2392 : // Our current tests don't need the background loops.
2393 : #[cfg(test)]
2394 174 : pub async fn create_test_timeline(
2395 174 : self: &Arc<Self>,
2396 174 : new_timeline_id: TimelineId,
2397 174 : initdb_lsn: Lsn,
2398 174 : pg_version: u32,
2399 174 : ctx: &RequestContext,
2400 174 : ) -> anyhow::Result<Arc<Timeline>> {
2401 174 : let uninit_tl = self
2402 174 : .create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2403 174 : .await?;
2404 174 : let tline = uninit_tl.raw_timeline().expect("we just created it");
2405 174 : assert_eq!(tline.get_last_record_lsn(), Lsn(0));
2406 :
2407 : // Setup minimum keys required for the timeline to be usable.
2408 174 : let mut modification = tline.begin_modification(initdb_lsn);
2409 174 : modification
2410 174 : .init_empty_test_timeline()
2411 174 : .context("init_empty_test_timeline")?;
2412 174 : modification
2413 174 : .commit(ctx)
2414 174 : .await
2415 174 : .context("commit init_empty_test_timeline modification")?;
2416 :
2417 : // Flush to disk so that uninit_tl's check for valid disk_consistent_lsn passes.
2418 174 : tline.maybe_spawn_flush_loop();
2419 174 : tline.freeze_and_flush().await.context("freeze_and_flush")?;
2420 :
2421 : // Make sure the freeze_and_flush reaches remote storage.
2422 174 : tline.remote_client.wait_completion().await.unwrap();
2423 :
2424 174 : let tl = uninit_tl.finish_creation()?;
2425 : // The non-test code would call tl.activate() here.
2426 174 : tl.set_state(TimelineState::Active);
2427 174 : Ok(tl)
2428 174 : }
2429 :
2430 : /// Helper for unit tests to create a timeline with some pre-loaded states.
2431 : #[cfg(test)]
2432 : #[allow(clippy::too_many_arguments)]
2433 32 : pub async fn create_test_timeline_with_layers(
2434 32 : self: &Arc<Self>,
2435 32 : new_timeline_id: TimelineId,
2436 32 : initdb_lsn: Lsn,
2437 32 : pg_version: u32,
2438 32 : ctx: &RequestContext,
2439 32 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
2440 32 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
2441 32 : end_lsn: Lsn,
2442 32 : ) -> anyhow::Result<Arc<Timeline>> {
2443 : use checks::check_valid_layermap;
2444 : use itertools::Itertools;
2445 :
2446 32 : let tline = self
2447 32 : .create_test_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2448 32 : .await?;
2449 32 : tline.force_advance_lsn(end_lsn);
2450 100 : for deltas in delta_layer_desc {
2451 68 : tline
2452 68 : .force_create_delta_layer(deltas, Some(initdb_lsn), ctx)
2453 68 : .await?;
2454 : }
2455 80 : for (lsn, images) in image_layer_desc {
2456 48 : tline
2457 48 : .force_create_image_layer(lsn, images, Some(initdb_lsn), ctx)
2458 48 : .await?;
2459 : }
2460 32 : let layer_names = tline
2461 32 : .layers
2462 32 : .read()
2463 32 : .await
2464 32 : .layer_map()
2465 32 : .unwrap()
2466 32 : .iter_historic_layers()
2467 148 : .map(|layer| layer.layer_name())
2468 32 : .collect_vec();
2469 32 : if let Some(err) = check_valid_layermap(&layer_names) {
2470 0 : bail!("invalid layermap: {err}");
2471 32 : }
2472 32 : Ok(tline)
2473 32 : }
2474 :
2475 : /// Create a new timeline.
2476 : ///
2477 : /// Returns the new timeline ID and reference to its Timeline object.
2478 : ///
2479 : /// If the caller specified the timeline ID to use (`new_timeline_id`), and timeline with
2480 : /// the same timeline ID already exists, returns CreateTimelineError::AlreadyExists.
2481 : #[allow(clippy::too_many_arguments)]
2482 0 : pub(crate) async fn create_timeline(
2483 0 : self: &Arc<Tenant>,
2484 0 : params: CreateTimelineParams,
2485 0 : broker_client: storage_broker::BrokerClientChannel,
2486 0 : ctx: &RequestContext,
2487 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
2488 0 : if !self.is_active() {
2489 0 : if matches!(self.current_state(), TenantState::Stopping { .. }) {
2490 0 : return Err(CreateTimelineError::ShuttingDown);
2491 : } else {
2492 0 : return Err(CreateTimelineError::Other(anyhow::anyhow!(
2493 0 : "Cannot create timelines on inactive tenant"
2494 0 : )));
2495 : }
2496 0 : }
2497 :
2498 0 : let _gate = self
2499 0 : .gate
2500 0 : .enter()
2501 0 : .map_err(|_| CreateTimelineError::ShuttingDown)?;
2502 :
2503 0 : let result: CreateTimelineResult = match params {
2504 : CreateTimelineParams::Bootstrap(CreateTimelineParamsBootstrap {
2505 0 : new_timeline_id,
2506 0 : existing_initdb_timeline_id,
2507 0 : pg_version,
2508 0 : }) => {
2509 0 : self.bootstrap_timeline(
2510 0 : new_timeline_id,
2511 0 : pg_version,
2512 0 : existing_initdb_timeline_id,
2513 0 : ctx,
2514 0 : )
2515 0 : .await?
2516 : }
2517 : CreateTimelineParams::Branch(CreateTimelineParamsBranch {
2518 0 : new_timeline_id,
2519 0 : ancestor_timeline_id,
2520 0 : mut ancestor_start_lsn,
2521 : }) => {
2522 0 : let ancestor_timeline = self
2523 0 : .get_timeline(ancestor_timeline_id, false)
2524 0 : .context("Cannot branch off the timeline that's not present in pageserver")?;
2525 :
2526 : // instead of waiting around, just deny the request because ancestor is not yet
2527 : // ready for other purposes either.
2528 0 : if !ancestor_timeline.is_active() {
2529 0 : return Err(CreateTimelineError::AncestorNotActive);
2530 0 : }
2531 0 :
2532 0 : if ancestor_timeline.is_archived() == Some(true) {
2533 0 : info!("tried to branch archived timeline");
2534 0 : return Err(CreateTimelineError::AncestorArchived);
2535 0 : }
2536 :
2537 0 : if let Some(lsn) = ancestor_start_lsn.as_mut() {
2538 0 : *lsn = lsn.align();
2539 0 :
2540 0 : let ancestor_ancestor_lsn = ancestor_timeline.get_ancestor_lsn();
2541 0 : if ancestor_ancestor_lsn > *lsn {
2542 : // can we safely just branch from the ancestor instead?
2543 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
2544 0 : "invalid start lsn {} for ancestor timeline {}: less than timeline ancestor lsn {}",
2545 0 : lsn,
2546 0 : ancestor_timeline_id,
2547 0 : ancestor_ancestor_lsn,
2548 0 : )));
2549 0 : }
2550 0 :
2551 0 : // Wait for the WAL to arrive and be processed on the parent branch up
2552 0 : // to the requested branch point. The repository code itself doesn't
2553 0 : // require it, but if we start to receive WAL on the new timeline,
2554 0 : // decoding the new WAL might need to look up previous pages, relation
2555 0 : // sizes etc. and that would get confused if the previous page versions
2556 0 : // are not in the repository yet.
2557 0 : ancestor_timeline
2558 0 : .wait_lsn(*lsn, timeline::WaitLsnWaiter::Tenant, ctx)
2559 0 : .await
2560 0 : .map_err(|e| match e {
2561 0 : e @ (WaitLsnError::Timeout(_) | WaitLsnError::BadState { .. }) => {
2562 0 : CreateTimelineError::AncestorLsn(anyhow::anyhow!(e))
2563 : }
2564 0 : WaitLsnError::Shutdown => CreateTimelineError::ShuttingDown,
2565 0 : })?;
2566 0 : }
2567 :
2568 0 : self.branch_timeline(&ancestor_timeline, new_timeline_id, ancestor_start_lsn, ctx)
2569 0 : .await?
2570 : }
2571 0 : CreateTimelineParams::ImportPgdata(params) => {
2572 0 : self.create_timeline_import_pgdata(
2573 0 : params,
2574 0 : ActivateTimelineArgs::Yes {
2575 0 : broker_client: broker_client.clone(),
2576 0 : },
2577 0 : ctx,
2578 0 : )
2579 0 : .await?
2580 : }
2581 : };
2582 :
2583 : // At this point we have dropped our guard on [`Self::timelines_creating`], and
2584 : // the timeline is visible in [`Self::timelines`], but it is _not_ durable yet. We must
2585 : // not send a success to the caller until it is. The same applies to idempotent retries.
2586 : //
2587 : // TODO: the timeline is already visible in [`Self::timelines`]; a caller could incorrectly
2588 : // assume that, because they can see the timeline via API, that the creation is done and
2589 : // that it is durable. Ideally, we would keep the timeline hidden (in [`Self::timelines_creating`])
2590 : // until it is durable, e.g., by extending the time we hold the creation guard. This also
2591 : // interacts with UninitializedTimeline and is generally a bit tricky.
2592 : //
2593 : // To re-emphasize: the only correct way to create a timeline is to repeat calling the
2594 : // creation API until it returns success. Only then is durability guaranteed.
2595 0 : info!(creation_result=%result.discriminant(), "waiting for timeline to be durable");
2596 0 : result
2597 0 : .timeline()
2598 0 : .remote_client
2599 0 : .wait_completion()
2600 0 : .await
2601 0 : .map_err(|e| match e {
2602 : WaitCompletionError::NotInitialized(
2603 0 : e, // If the queue is already stopped, it's a shutdown error.
2604 0 : ) if e.is_stopping() => CreateTimelineError::ShuttingDown,
2605 0 : e => CreateTimelineError::Other(e.into()),
2606 0 : })
2607 0 : .context("wait for timeline initial uploads to complete")?;
2608 :
2609 : // The creating task is responsible for activating the timeline.
2610 : // We do this after `wait_completion()` so that we don't spin up tasks that start
2611 : // doing stuff before the IndexPart is durable in S3, which is done by the previous section.
2612 0 : let activated_timeline = match result {
2613 0 : CreateTimelineResult::Created(timeline) => {
2614 0 : timeline.activate(self.clone(), broker_client, None, ctx);
2615 0 : timeline
2616 : }
2617 0 : CreateTimelineResult::Idempotent(timeline) => {
2618 0 : info!(
2619 0 : "request was deemed idempotent, activation will be done by the creating task"
2620 : );
2621 0 : timeline
2622 : }
2623 0 : CreateTimelineResult::ImportSpawned(timeline) => {
2624 0 : info!("import task spawned, timeline will become visible and activated once the import is done");
2625 0 : timeline
2626 : }
2627 : };
2628 :
2629 0 : Ok(activated_timeline)
2630 0 : }
2631 :
2632 : /// The returned [`Arc<Timeline>`] is NOT in the [`Tenant::timelines`] map until the import
2633 : /// completes in the background. A DIFFERENT [`Arc<Timeline>`] will be inserted into the
2634 : /// [`Tenant::timelines`] map when the import completes.
2635 : /// We only return an [`Arc<Timeline>`] here so the API handler can create a [`pageserver_api::models::TimelineInfo`]
2636 : /// for the response.
2637 0 : async fn create_timeline_import_pgdata(
2638 0 : self: &Arc<Tenant>,
2639 0 : params: CreateTimelineParamsImportPgdata,
2640 0 : activate: ActivateTimelineArgs,
2641 0 : ctx: &RequestContext,
2642 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
2643 0 : let CreateTimelineParamsImportPgdata {
2644 0 : new_timeline_id,
2645 0 : location,
2646 0 : idempotency_key,
2647 0 : } = params;
2648 0 :
2649 0 : let started_at = chrono::Utc::now().naive_utc();
2650 :
2651 : //
2652 : // There's probably a simpler way to upload an index part, but, remote_timeline_client
2653 : // is the canonical way we do it.
2654 : // - create an empty timeline in-memory
2655 : // - use its remote_timeline_client to do the upload
2656 : // - dispose of the uninit timeline
2657 : // - keep the creation guard alive
2658 :
2659 0 : let timeline_create_guard = match self
2660 0 : .start_creating_timeline(
2661 0 : new_timeline_id,
2662 0 : CreateTimelineIdempotency::ImportPgdata(CreatingTimelineIdempotencyImportPgdata {
2663 0 : idempotency_key: idempotency_key.clone(),
2664 0 : }),
2665 0 : )
2666 0 : .await?
2667 : {
2668 0 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2669 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
2670 0 : return Ok(CreateTimelineResult::Idempotent(timeline))
2671 : }
2672 : };
2673 :
2674 0 : let mut uninit_timeline = {
2675 0 : let this = &self;
2676 0 : let initdb_lsn = Lsn(0);
2677 0 : let _ctx = ctx;
2678 0 : async move {
2679 0 : let new_metadata = TimelineMetadata::new(
2680 0 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2681 0 : // make it valid, before calling finish_creation()
2682 0 : Lsn(0),
2683 0 : None,
2684 0 : None,
2685 0 : Lsn(0),
2686 0 : initdb_lsn,
2687 0 : initdb_lsn,
2688 0 : 15,
2689 0 : );
2690 0 : this.prepare_new_timeline(
2691 0 : new_timeline_id,
2692 0 : &new_metadata,
2693 0 : timeline_create_guard,
2694 0 : initdb_lsn,
2695 0 : None,
2696 0 : )
2697 0 : .await
2698 0 : }
2699 0 : }
2700 0 : .await?;
2701 :
2702 0 : let in_progress = import_pgdata::index_part_format::InProgress {
2703 0 : idempotency_key,
2704 0 : location,
2705 0 : started_at,
2706 0 : };
2707 0 : let index_part = import_pgdata::index_part_format::Root::V1(
2708 0 : import_pgdata::index_part_format::V1::InProgress(in_progress),
2709 0 : );
2710 0 : uninit_timeline
2711 0 : .raw_timeline()
2712 0 : .unwrap()
2713 0 : .remote_client
2714 0 : .schedule_index_upload_for_import_pgdata_state_update(Some(index_part.clone()))?;
2715 :
2716 : // wait_completion happens in caller
2717 :
2718 0 : let (timeline, timeline_create_guard) = uninit_timeline.finish_creation_myself();
2719 0 :
2720 0 : tokio::spawn(self.clone().create_timeline_import_pgdata_task(
2721 0 : timeline.clone(),
2722 0 : index_part,
2723 0 : activate,
2724 0 : timeline_create_guard,
2725 0 : ));
2726 0 :
2727 0 : // NB: the timeline doesn't exist in self.timelines at this point
2728 0 : Ok(CreateTimelineResult::ImportSpawned(timeline))
2729 0 : }
2730 :
2731 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), timeline_id=%timeline.timeline_id))]
2732 : async fn create_timeline_import_pgdata_task(
2733 : self: Arc<Tenant>,
2734 : timeline: Arc<Timeline>,
2735 : index_part: import_pgdata::index_part_format::Root,
2736 : activate: ActivateTimelineArgs,
2737 : timeline_create_guard: TimelineCreateGuard,
2738 : ) {
2739 : debug_assert_current_span_has_tenant_and_timeline_id();
2740 : info!("starting");
2741 : scopeguard::defer! {info!("exiting")};
2742 :
2743 : let res = self
2744 : .create_timeline_import_pgdata_task_impl(
2745 : timeline,
2746 : index_part,
2747 : activate,
2748 : timeline_create_guard,
2749 : )
2750 : .await;
2751 : if let Err(err) = &res {
2752 : error!(?err, "task failed");
2753 : // TODO sleep & retry, sensitive to tenant shutdown
2754 : // TODO: allow timeline deletion requests => should cancel the task
2755 : }
2756 : }
2757 :
2758 0 : async fn create_timeline_import_pgdata_task_impl(
2759 0 : self: Arc<Tenant>,
2760 0 : timeline: Arc<Timeline>,
2761 0 : index_part: import_pgdata::index_part_format::Root,
2762 0 : activate: ActivateTimelineArgs,
2763 0 : timeline_create_guard: TimelineCreateGuard,
2764 0 : ) -> Result<(), anyhow::Error> {
2765 0 : let ctx = RequestContext::new(TaskKind::ImportPgdata, DownloadBehavior::Warn);
2766 0 :
2767 0 : info!("importing pgdata");
2768 0 : import_pgdata::doit(&timeline, index_part, &ctx, self.cancel.clone())
2769 0 : .await
2770 0 : .context("import")?;
2771 0 : info!("import done");
2772 :
2773 : //
2774 : // Reload timeline from remote.
2775 : // This proves that the remote state is attachable, and it reuses the code.
2776 : //
2777 : // TODO: think about whether this is safe to do with concurrent Tenant::shutdown.
2778 : // timeline_create_guard hols the tenant gate open, so, shutdown cannot _complete_ until we exit.
2779 : // But our activate() call might launch new background tasks after Tenant::shutdown
2780 : // already went past shutting down the Tenant::timelines, which this timeline here is no part of.
2781 : // I think the same problem exists with the bootstrap & branch mgmt API tasks (tenant shutting
2782 : // down while bootstrapping/branching + activating), but, the race condition is much more likely
2783 : // to manifest because of the long runtime of this import task.
2784 :
2785 : // in theory this shouldn't even .await anything except for coop yield
2786 0 : info!("shutting down timeline");
2787 0 : timeline.shutdown(ShutdownMode::Hard).await;
2788 0 : info!("timeline shut down, reloading from remote");
2789 : // TODO: we can't do the following check because create_timeline_import_pgdata must return an Arc<Timeline>
2790 : // let Some(timeline) = Arc::into_inner(timeline) else {
2791 : // anyhow::bail!("implementation error: timeline that we shut down was still referenced from somewhere");
2792 : // };
2793 0 : let timeline_id = timeline.timeline_id;
2794 0 :
2795 0 : // load from object storage like Tenant::attach does
2796 0 : let resources = self.build_timeline_resources(timeline_id);
2797 0 : let index_part = resources
2798 0 : .remote_client
2799 0 : .download_index_file(&self.cancel)
2800 0 : .await?;
2801 0 : let index_part = match index_part {
2802 : MaybeDeletedIndexPart::Deleted(_) => {
2803 : // likely concurrent delete call, cplane should prevent this
2804 0 : anyhow::bail!("index part says deleted but we are not done creating yet, this should not happen but")
2805 : }
2806 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
2807 0 : };
2808 0 : let metadata = index_part.metadata.clone();
2809 0 : self
2810 0 : .load_remote_timeline(timeline_id, index_part, metadata, resources, LoadTimelineCause::ImportPgdata{
2811 0 : create_guard: timeline_create_guard, activate, }, &ctx)
2812 0 : .await?
2813 0 : .ready_to_activate()
2814 0 : .context("implementation error: reloaded timeline still needs import after import reported success")?;
2815 :
2816 0 : anyhow::Ok(())
2817 0 : }
2818 :
2819 0 : pub(crate) async fn delete_timeline(
2820 0 : self: Arc<Self>,
2821 0 : timeline_id: TimelineId,
2822 0 : ) -> Result<(), DeleteTimelineError> {
2823 0 : DeleteTimelineFlow::run(&self, timeline_id).await?;
2824 :
2825 0 : Ok(())
2826 0 : }
2827 :
2828 : /// perform one garbage collection iteration, removing old data files from disk.
2829 : /// this function is periodically called by gc task.
2830 : /// also it can be explicitly requested through page server api 'do_gc' command.
2831 : ///
2832 : /// `target_timeline_id` specifies the timeline to GC, or None for all.
2833 : ///
2834 : /// The `horizon` an `pitr` parameters determine how much WAL history needs to be retained.
2835 : /// Also known as the retention period, or the GC cutoff point. `horizon` specifies
2836 : /// the amount of history, as LSN difference from current latest LSN on each timeline.
2837 : /// `pitr` specifies the same as a time difference from the current time. The effective
2838 : /// GC cutoff point is determined conservatively by either `horizon` and `pitr`, whichever
2839 : /// requires more history to be retained.
2840 : //
2841 754 : pub(crate) async fn gc_iteration(
2842 754 : &self,
2843 754 : target_timeline_id: Option<TimelineId>,
2844 754 : horizon: u64,
2845 754 : pitr: Duration,
2846 754 : cancel: &CancellationToken,
2847 754 : ctx: &RequestContext,
2848 754 : ) -> Result<GcResult, GcError> {
2849 754 : // Don't start doing work during shutdown
2850 754 : if let TenantState::Stopping { .. } = self.current_state() {
2851 0 : return Ok(GcResult::default());
2852 754 : }
2853 754 :
2854 754 : // there is a global allowed_error for this
2855 754 : if !self.is_active() {
2856 0 : return Err(GcError::NotActive);
2857 754 : }
2858 754 :
2859 754 : {
2860 754 : let conf = self.tenant_conf.load();
2861 754 :
2862 754 : // If we may not delete layers, then simply skip GC. Even though a tenant
2863 754 : // in AttachedMulti state could do GC and just enqueue the blocked deletions,
2864 754 : // the only advantage to doing it is to perhaps shrink the LayerMap metadata
2865 754 : // a bit sooner than we would achieve by waiting for AttachedSingle status.
2866 754 : if !conf.location.may_delete_layers_hint() {
2867 0 : info!("Skipping GC in location state {:?}", conf.location);
2868 0 : return Ok(GcResult::default());
2869 754 : }
2870 754 :
2871 754 : if conf.is_gc_blocked_by_lsn_lease_deadline() {
2872 750 : info!("Skipping GC because lsn lease deadline is not reached");
2873 750 : return Ok(GcResult::default());
2874 4 : }
2875 : }
2876 :
2877 4 : let _guard = match self.gc_block.start().await {
2878 4 : Ok(guard) => guard,
2879 0 : Err(reasons) => {
2880 0 : info!("Skipping GC: {reasons}");
2881 0 : return Ok(GcResult::default());
2882 : }
2883 : };
2884 :
2885 4 : self.gc_iteration_internal(target_timeline_id, horizon, pitr, cancel, ctx)
2886 4 : .await
2887 754 : }
2888 :
2889 : /// Perform one compaction iteration.
2890 : /// This function is periodically called by compactor task.
2891 : /// Also it can be explicitly requested per timeline through page server
2892 : /// api's 'compact' command.
2893 : ///
2894 : /// Returns whether we have pending compaction task.
2895 0 : async fn compaction_iteration(
2896 0 : self: &Arc<Self>,
2897 0 : cancel: &CancellationToken,
2898 0 : ctx: &RequestContext,
2899 0 : ) -> Result<bool, timeline::CompactionError> {
2900 0 : // Don't start doing work during shutdown, or when broken, we do not need those in the logs
2901 0 : if !self.is_active() {
2902 0 : return Ok(false);
2903 0 : }
2904 0 :
2905 0 : {
2906 0 : let conf = self.tenant_conf.load();
2907 0 :
2908 0 : // Note that compaction usually requires deletions, but we don't respect
2909 0 : // may_delete_layers_hint here: that is because tenants in AttachedMulti
2910 0 : // should proceed with compaction even if they can't do deletion, to avoid
2911 0 : // accumulating dangerously deep stacks of L0 layers. Deletions will be
2912 0 : // enqueued inside RemoteTimelineClient, and executed layer if/when we transition
2913 0 : // to AttachedSingle state.
2914 0 : if !conf.location.may_upload_layers_hint() {
2915 0 : info!("Skipping compaction in location state {:?}", conf.location);
2916 0 : return Ok(false);
2917 0 : }
2918 0 : }
2919 0 :
2920 0 : // Scan through the hashmap and collect a list of all the timelines,
2921 0 : // while holding the lock. Then drop the lock and actually perform the
2922 0 : // compactions. We don't want to block everything else while the
2923 0 : // compaction runs.
2924 0 : let timelines_to_compact_or_offload;
2925 0 : {
2926 0 : let timelines = self.timelines.lock().unwrap();
2927 0 : timelines_to_compact_or_offload = timelines
2928 0 : .iter()
2929 0 : .filter_map(|(timeline_id, timeline)| {
2930 0 : let (is_active, (can_offload, _)) =
2931 0 : (timeline.is_active(), timeline.can_offload());
2932 0 : let has_no_unoffloaded_children = {
2933 0 : !timelines
2934 0 : .iter()
2935 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(*timeline_id))
2936 : };
2937 0 : let config_allows_offload = self.conf.timeline_offloading
2938 0 : || self
2939 0 : .tenant_conf
2940 0 : .load()
2941 0 : .tenant_conf
2942 0 : .timeline_offloading
2943 0 : .unwrap_or_default();
2944 0 : let can_offload =
2945 0 : can_offload && has_no_unoffloaded_children && config_allows_offload;
2946 0 : if (is_active, can_offload) == (false, false) {
2947 0 : None
2948 : } else {
2949 0 : Some((*timeline_id, timeline.clone(), (is_active, can_offload)))
2950 : }
2951 0 : })
2952 0 : .collect::<Vec<_>>();
2953 0 : drop(timelines);
2954 0 : }
2955 0 :
2956 0 : // Before doing any I/O work, check our circuit breaker
2957 0 : if self.compaction_circuit_breaker.lock().unwrap().is_broken() {
2958 0 : info!("Skipping compaction due to previous failures");
2959 0 : return Ok(false);
2960 0 : }
2961 0 :
2962 0 : let mut has_pending_task = false;
2963 :
2964 0 : for (timeline_id, timeline, (can_compact, can_offload)) in &timelines_to_compact_or_offload
2965 : {
2966 : // pending_task_left == None: cannot compact, maybe still pending tasks
2967 : // pending_task_left == Some(true): compaction task left
2968 : // pending_task_left == Some(false): no compaction task left
2969 0 : let pending_task_left = if *can_compact {
2970 0 : let has_pending_l0_compaction_task = timeline
2971 0 : .compact(cancel, EnumSet::empty(), ctx)
2972 0 : .instrument(info_span!("compact_timeline", %timeline_id))
2973 0 : .await
2974 0 : .inspect_err(|e| match e {
2975 0 : timeline::CompactionError::ShuttingDown => (),
2976 0 : timeline::CompactionError::Offload(_) => {
2977 0 : // Failures to offload timelines do not trip the circuit breaker, because
2978 0 : // they do not do lots of writes the way compaction itself does: it is cheap
2979 0 : // to retry, and it would be bad to stop all compaction because of an issue with offloading.
2980 0 : }
2981 0 : timeline::CompactionError::Other(e) => {
2982 0 : self.compaction_circuit_breaker
2983 0 : .lock()
2984 0 : .unwrap()
2985 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, e);
2986 0 : }
2987 0 : })?;
2988 0 : if has_pending_l0_compaction_task {
2989 0 : Some(true)
2990 : } else {
2991 : let mut has_pending_scheduled_compaction_task;
2992 0 : let next_scheduled_compaction_task = {
2993 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
2994 0 : if let Some(tline_pending_tasks) = guard.get_mut(timeline_id) {
2995 0 : if !tline_pending_tasks.is_empty() {
2996 0 : info!(
2997 0 : "{} tasks left in the compaction schedule queue",
2998 0 : tline_pending_tasks.len()
2999 : );
3000 0 : }
3001 0 : let next_task = tline_pending_tasks.pop_front();
3002 0 : has_pending_scheduled_compaction_task = !tline_pending_tasks.is_empty();
3003 0 : next_task
3004 : } else {
3005 0 : has_pending_scheduled_compaction_task = false;
3006 0 : None
3007 : }
3008 : };
3009 0 : if let Some(mut next_scheduled_compaction_task) = next_scheduled_compaction_task
3010 : {
3011 0 : if !next_scheduled_compaction_task
3012 0 : .options
3013 0 : .flags
3014 0 : .contains(CompactFlags::EnhancedGcBottomMostCompaction)
3015 : {
3016 0 : warn!("ignoring scheduled compaction task: scheduled task must be gc compaction: {:?}", next_scheduled_compaction_task.options);
3017 0 : } else if next_scheduled_compaction_task.options.sub_compaction {
3018 0 : info!("running scheduled enhanced gc bottom-most compaction with sub-compaction, splitting compaction jobs");
3019 0 : let jobs = timeline
3020 0 : .gc_compaction_split_jobs(next_scheduled_compaction_task.options)
3021 0 : .await
3022 0 : .map_err(CompactionError::Other)?;
3023 0 : if jobs.is_empty() {
3024 0 : info!("no jobs to run, skipping scheduled compaction task");
3025 : } else {
3026 0 : has_pending_scheduled_compaction_task = true;
3027 0 : let jobs_len = jobs.len();
3028 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3029 0 : let tline_pending_tasks = guard.entry(*timeline_id).or_default();
3030 0 : for (idx, job) in jobs.into_iter().enumerate() {
3031 0 : tline_pending_tasks.push_back(if idx == jobs_len - 1 {
3032 0 : ScheduledCompactionTask {
3033 0 : options: job,
3034 0 : // The last job in the queue sends the signal and releases the gc guard
3035 0 : result_tx: next_scheduled_compaction_task
3036 0 : .result_tx
3037 0 : .take(),
3038 0 : gc_block: next_scheduled_compaction_task
3039 0 : .gc_block
3040 0 : .take(),
3041 0 : }
3042 : } else {
3043 0 : ScheduledCompactionTask {
3044 0 : options: job,
3045 0 : result_tx: None,
3046 0 : gc_block: None,
3047 0 : }
3048 : });
3049 : }
3050 0 : info!("scheduled enhanced gc bottom-most compaction with sub-compaction, split into {} jobs", jobs_len);
3051 : }
3052 : } else {
3053 0 : let _ = timeline
3054 0 : .compact_with_options(
3055 0 : cancel,
3056 0 : next_scheduled_compaction_task.options,
3057 0 : ctx,
3058 0 : )
3059 0 : .instrument(info_span!("scheduled_compact_timeline", %timeline_id))
3060 0 : .await?;
3061 0 : if let Some(tx) = next_scheduled_compaction_task.result_tx.take() {
3062 0 : // TODO: we can send compaction statistics in the future
3063 0 : tx.send(()).ok();
3064 0 : }
3065 : }
3066 0 : }
3067 0 : Some(has_pending_scheduled_compaction_task)
3068 : }
3069 : } else {
3070 0 : None
3071 : };
3072 0 : has_pending_task |= pending_task_left.unwrap_or(false);
3073 0 : if pending_task_left == Some(false) && *can_offload {
3074 0 : offload_timeline(self, timeline)
3075 0 : .instrument(info_span!("offload_timeline", %timeline_id))
3076 0 : .await?;
3077 0 : }
3078 : }
3079 :
3080 0 : self.compaction_circuit_breaker
3081 0 : .lock()
3082 0 : .unwrap()
3083 0 : .success(&CIRCUIT_BREAKERS_UNBROKEN);
3084 0 :
3085 0 : Ok(has_pending_task)
3086 0 : }
3087 :
3088 : /// Cancel scheduled compaction tasks
3089 0 : pub(crate) fn cancel_scheduled_compaction(
3090 0 : &self,
3091 0 : timeline_id: TimelineId,
3092 0 : ) -> Vec<ScheduledCompactionTask> {
3093 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3094 0 : if let Some(tline_pending_tasks) = guard.get_mut(&timeline_id) {
3095 0 : let current_tline_pending_tasks = std::mem::take(tline_pending_tasks);
3096 0 : current_tline_pending_tasks.into_iter().collect()
3097 : } else {
3098 0 : Vec::new()
3099 : }
3100 0 : }
3101 :
3102 : /// Schedule a compaction task for a timeline.
3103 0 : pub(crate) async fn schedule_compaction(
3104 0 : &self,
3105 0 : timeline_id: TimelineId,
3106 0 : options: CompactOptions,
3107 0 : ) -> anyhow::Result<tokio::sync::oneshot::Receiver<()>> {
3108 0 : let gc_guard = match self.gc_block.start().await {
3109 0 : Ok(guard) => guard,
3110 0 : Err(e) => {
3111 0 : bail!("cannot run gc-compaction because gc is blocked: {}", e);
3112 : }
3113 : };
3114 0 : let (tx, rx) = tokio::sync::oneshot::channel();
3115 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3116 0 : let tline_pending_tasks = guard.entry(timeline_id).or_default();
3117 0 : tline_pending_tasks.push_back(ScheduledCompactionTask {
3118 0 : options,
3119 0 : result_tx: Some(tx),
3120 0 : gc_block: Some(gc_guard),
3121 0 : });
3122 0 : Ok(rx)
3123 0 : }
3124 :
3125 : // Call through to all timelines to freeze ephemeral layers if needed. Usually
3126 : // this happens during ingest: this background housekeeping is for freezing layers
3127 : // that are open but haven't been written to for some time.
3128 0 : async fn ingest_housekeeping(&self) {
3129 0 : // Scan through the hashmap and collect a list of all the timelines,
3130 0 : // while holding the lock. Then drop the lock and actually perform the
3131 0 : // compactions. We don't want to block everything else while the
3132 0 : // compaction runs.
3133 0 : let timelines = {
3134 0 : self.timelines
3135 0 : .lock()
3136 0 : .unwrap()
3137 0 : .values()
3138 0 : .filter_map(|timeline| {
3139 0 : if timeline.is_active() {
3140 0 : Some(timeline.clone())
3141 : } else {
3142 0 : None
3143 : }
3144 0 : })
3145 0 : .collect::<Vec<_>>()
3146 : };
3147 :
3148 0 : for timeline in &timelines {
3149 0 : timeline.maybe_freeze_ephemeral_layer().await;
3150 : }
3151 0 : }
3152 :
3153 0 : pub fn timeline_has_no_attached_children(&self, timeline_id: TimelineId) -> bool {
3154 0 : let timelines = self.timelines.lock().unwrap();
3155 0 : !timelines
3156 0 : .iter()
3157 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(timeline_id))
3158 0 : }
3159 :
3160 1702 : pub fn current_state(&self) -> TenantState {
3161 1702 : self.state.borrow().clone()
3162 1702 : }
3163 :
3164 942 : pub fn is_active(&self) -> bool {
3165 942 : self.current_state() == TenantState::Active
3166 942 : }
3167 :
3168 0 : pub fn generation(&self) -> Generation {
3169 0 : self.generation
3170 0 : }
3171 :
3172 0 : pub(crate) fn wal_redo_manager_status(&self) -> Option<WalRedoManagerStatus> {
3173 0 : self.walredo_mgr.as_ref().and_then(|mgr| mgr.status())
3174 0 : }
3175 :
3176 : /// Changes tenant status to active, unless shutdown was already requested.
3177 : ///
3178 : /// `background_jobs_can_start` is an optional barrier set to a value during pageserver startup
3179 : /// to delay background jobs. Background jobs can be started right away when None is given.
3180 0 : fn activate(
3181 0 : self: &Arc<Self>,
3182 0 : broker_client: BrokerClientChannel,
3183 0 : background_jobs_can_start: Option<&completion::Barrier>,
3184 0 : ctx: &RequestContext,
3185 0 : ) {
3186 0 : span::debug_assert_current_span_has_tenant_id();
3187 0 :
3188 0 : let mut activating = false;
3189 0 : self.state.send_modify(|current_state| {
3190 : use pageserver_api::models::ActivatingFrom;
3191 0 : match &*current_state {
3192 : TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
3193 0 : panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {state:?}", state = current_state);
3194 : }
3195 0 : TenantState::Attaching => {
3196 0 : *current_state = TenantState::Activating(ActivatingFrom::Attaching);
3197 0 : }
3198 0 : }
3199 0 : debug!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), "Activating tenant");
3200 0 : activating = true;
3201 0 : // Continue outside the closure. We need to grab timelines.lock()
3202 0 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
3203 0 : });
3204 0 :
3205 0 : if activating {
3206 0 : let timelines_accessor = self.timelines.lock().unwrap();
3207 0 : let timelines_offloaded_accessor = self.timelines_offloaded.lock().unwrap();
3208 0 : let timelines_to_activate = timelines_accessor
3209 0 : .values()
3210 0 : .filter(|timeline| !(timeline.is_broken() || timeline.is_stopping()));
3211 0 :
3212 0 : // Before activation, populate each Timeline's GcInfo with information about its children
3213 0 : self.initialize_gc_info(&timelines_accessor, &timelines_offloaded_accessor, None);
3214 0 :
3215 0 : // Spawn gc and compaction loops. The loops will shut themselves
3216 0 : // down when they notice that the tenant is inactive.
3217 0 : tasks::start_background_loops(self, background_jobs_can_start);
3218 0 :
3219 0 : let mut activated_timelines = 0;
3220 :
3221 0 : for timeline in timelines_to_activate {
3222 0 : timeline.activate(
3223 0 : self.clone(),
3224 0 : broker_client.clone(),
3225 0 : background_jobs_can_start,
3226 0 : ctx,
3227 0 : );
3228 0 : activated_timelines += 1;
3229 0 : }
3230 :
3231 0 : self.state.send_modify(move |current_state| {
3232 0 : assert!(
3233 0 : matches!(current_state, TenantState::Activating(_)),
3234 0 : "set_stopping and set_broken wait for us to leave Activating state",
3235 : );
3236 0 : *current_state = TenantState::Active;
3237 0 :
3238 0 : let elapsed = self.constructed_at.elapsed();
3239 0 : let total_timelines = timelines_accessor.len();
3240 0 :
3241 0 : // log a lot of stuff, because some tenants sometimes suffer from user-visible
3242 0 : // times to activate. see https://github.com/neondatabase/neon/issues/4025
3243 0 : info!(
3244 0 : since_creation_millis = elapsed.as_millis(),
3245 0 : tenant_id = %self.tenant_shard_id.tenant_id,
3246 0 : shard_id = %self.tenant_shard_id.shard_slug(),
3247 0 : activated_timelines,
3248 0 : total_timelines,
3249 0 : post_state = <&'static str>::from(&*current_state),
3250 0 : "activation attempt finished"
3251 : );
3252 :
3253 0 : TENANT.activation.observe(elapsed.as_secs_f64());
3254 0 : });
3255 0 : }
3256 0 : }
3257 :
3258 : /// Shutdown the tenant and join all of the spawned tasks.
3259 : ///
3260 : /// The method caters for all use-cases:
3261 : /// - pageserver shutdown (freeze_and_flush == true)
3262 : /// - detach + ignore (freeze_and_flush == false)
3263 : ///
3264 : /// This will attempt to shutdown even if tenant is broken.
3265 : ///
3266 : /// `shutdown_progress` is a [`completion::Barrier`] for the shutdown initiated by this call.
3267 : /// If the tenant is already shutting down, we return a clone of the first shutdown call's
3268 : /// `Barrier` as an `Err`. This not-first caller can use the returned barrier to join with
3269 : /// the ongoing shutdown.
3270 6 : async fn shutdown(
3271 6 : &self,
3272 6 : shutdown_progress: completion::Barrier,
3273 6 : shutdown_mode: timeline::ShutdownMode,
3274 6 : ) -> Result<(), completion::Barrier> {
3275 6 : span::debug_assert_current_span_has_tenant_id();
3276 :
3277 : // Set tenant (and its timlines) to Stoppping state.
3278 : //
3279 : // Since we can only transition into Stopping state after activation is complete,
3280 : // run it in a JoinSet so all tenants have a chance to stop before we get SIGKILLed.
3281 : //
3282 : // Transitioning tenants to Stopping state has a couple of non-obvious side effects:
3283 : // 1. Lock out any new requests to the tenants.
3284 : // 2. Signal cancellation to WAL receivers (we wait on it below).
3285 : // 3. Signal cancellation for other tenant background loops.
3286 : // 4. ???
3287 : //
3288 : // The waiting for the cancellation is not done uniformly.
3289 : // We certainly wait for WAL receivers to shut down.
3290 : // That is necessary so that no new data comes in before the freeze_and_flush.
3291 : // But the tenant background loops are joined-on in our caller.
3292 : // It's mesed up.
3293 : // we just ignore the failure to stop
3294 :
3295 : // If we're still attaching, fire the cancellation token early to drop out: this
3296 : // will prevent us flushing, but ensures timely shutdown if some I/O during attach
3297 : // is very slow.
3298 6 : let shutdown_mode = if matches!(self.current_state(), TenantState::Attaching) {
3299 0 : self.cancel.cancel();
3300 0 :
3301 0 : // Having fired our cancellation token, do not try and flush timelines: their cancellation tokens
3302 0 : // are children of ours, so their flush loops will have shut down already
3303 0 : timeline::ShutdownMode::Hard
3304 : } else {
3305 6 : shutdown_mode
3306 : };
3307 :
3308 6 : match self.set_stopping(shutdown_progress, false, false).await {
3309 6 : Ok(()) => {}
3310 0 : Err(SetStoppingError::Broken) => {
3311 0 : // assume that this is acceptable
3312 0 : }
3313 0 : Err(SetStoppingError::AlreadyStopping(other)) => {
3314 0 : // give caller the option to wait for this this shutdown
3315 0 : info!("Tenant::shutdown: AlreadyStopping");
3316 0 : return Err(other);
3317 : }
3318 : };
3319 :
3320 6 : let mut js = tokio::task::JoinSet::new();
3321 6 : {
3322 6 : let timelines = self.timelines.lock().unwrap();
3323 6 : timelines.values().for_each(|timeline| {
3324 6 : let timeline = Arc::clone(timeline);
3325 6 : let timeline_id = timeline.timeline_id;
3326 6 : let span = tracing::info_span!("timeline_shutdown", %timeline_id, ?shutdown_mode);
3327 6 : js.spawn(async move { timeline.shutdown(shutdown_mode).instrument(span).await });
3328 6 : });
3329 6 : }
3330 6 : {
3331 6 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
3332 6 : timelines_offloaded.values().for_each(|timeline| {
3333 0 : timeline.defuse_for_tenant_drop();
3334 6 : });
3335 6 : }
3336 6 : // test_long_timeline_create_then_tenant_delete is leaning on this message
3337 6 : tracing::info!("Waiting for timelines...");
3338 12 : while let Some(res) = js.join_next().await {
3339 0 : match res {
3340 6 : Ok(()) => {}
3341 0 : Err(je) if je.is_cancelled() => unreachable!("no cancelling used"),
3342 0 : Err(je) if je.is_panic() => { /* logged already */ }
3343 0 : Err(je) => warn!("unexpected JoinError: {je:?}"),
3344 : }
3345 : }
3346 :
3347 6 : if let ShutdownMode::Reload = shutdown_mode {
3348 0 : tracing::info!("Flushing deletion queue");
3349 0 : if let Err(e) = self.deletion_queue_client.flush().await {
3350 0 : match e {
3351 0 : DeletionQueueError::ShuttingDown => {
3352 0 : // This is the only error we expect for now. In the future, if more error
3353 0 : // variants are added, we should handle them here.
3354 0 : }
3355 : }
3356 0 : }
3357 6 : }
3358 :
3359 : // We cancel the Tenant's cancellation token _after_ the timelines have all shut down. This permits
3360 : // them to continue to do work during their shutdown methods, e.g. flushing data.
3361 6 : tracing::debug!("Cancelling CancellationToken");
3362 6 : self.cancel.cancel();
3363 6 :
3364 6 : // shutdown all tenant and timeline tasks: gc, compaction, page service
3365 6 : // No new tasks will be started for this tenant because it's in `Stopping` state.
3366 6 : //
3367 6 : // this will additionally shutdown and await all timeline tasks.
3368 6 : tracing::debug!("Waiting for tasks...");
3369 6 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), None).await;
3370 :
3371 6 : if let Some(walredo_mgr) = self.walredo_mgr.as_ref() {
3372 6 : walredo_mgr.shutdown().await;
3373 0 : }
3374 :
3375 : // Wait for any in-flight operations to complete
3376 6 : self.gate.close().await;
3377 :
3378 6 : remove_tenant_metrics(&self.tenant_shard_id);
3379 6 :
3380 6 : Ok(())
3381 6 : }
3382 :
3383 : /// Change tenant status to Stopping, to mark that it is being shut down.
3384 : ///
3385 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
3386 : ///
3387 : /// This function is not cancel-safe!
3388 : ///
3389 : /// `allow_transition_from_loading` is needed for the special case of loading task deleting the tenant.
3390 : /// `allow_transition_from_attaching` is needed for the special case of attaching deleted tenant.
3391 6 : async fn set_stopping(
3392 6 : &self,
3393 6 : progress: completion::Barrier,
3394 6 : _allow_transition_from_loading: bool,
3395 6 : allow_transition_from_attaching: bool,
3396 6 : ) -> Result<(), SetStoppingError> {
3397 6 : let mut rx = self.state.subscribe();
3398 6 :
3399 6 : // cannot stop before we're done activating, so wait out until we're done activating
3400 6 : rx.wait_for(|state| match state {
3401 0 : TenantState::Attaching if allow_transition_from_attaching => true,
3402 : TenantState::Activating(_) | TenantState::Attaching => {
3403 0 : info!(
3404 0 : "waiting for {} to turn Active|Broken|Stopping",
3405 0 : <&'static str>::from(state)
3406 : );
3407 0 : false
3408 : }
3409 6 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3410 6 : })
3411 6 : .await
3412 6 : .expect("cannot drop self.state while on a &self method");
3413 6 :
3414 6 : // we now know we're done activating, let's see whether this task is the winner to transition into Stopping
3415 6 : let mut err = None;
3416 6 : let stopping = self.state.send_if_modified(|current_state| match current_state {
3417 : TenantState::Activating(_) => {
3418 0 : unreachable!("1we ensured above that we're done with activation, and, there is no re-activation")
3419 : }
3420 : TenantState::Attaching => {
3421 0 : if !allow_transition_from_attaching {
3422 0 : unreachable!("2we ensured above that we're done with activation, and, there is no re-activation")
3423 0 : };
3424 0 : *current_state = TenantState::Stopping { progress };
3425 0 : true
3426 : }
3427 : TenantState::Active => {
3428 : // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
3429 : // are created after the transition to Stopping. That's harmless, as the Timelines
3430 : // won't be accessible to anyone afterwards, because the Tenant is in Stopping state.
3431 6 : *current_state = TenantState::Stopping { progress };
3432 6 : // Continue stopping outside the closure. We need to grab timelines.lock()
3433 6 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
3434 6 : true
3435 : }
3436 0 : TenantState::Broken { reason, .. } => {
3437 0 : info!(
3438 0 : "Cannot set tenant to Stopping state, it is in Broken state due to: {reason}"
3439 : );
3440 0 : err = Some(SetStoppingError::Broken);
3441 0 : false
3442 : }
3443 0 : TenantState::Stopping { progress } => {
3444 0 : info!("Tenant is already in Stopping state");
3445 0 : err = Some(SetStoppingError::AlreadyStopping(progress.clone()));
3446 0 : false
3447 : }
3448 6 : });
3449 6 : match (stopping, err) {
3450 6 : (true, None) => {} // continue
3451 0 : (false, Some(err)) => return Err(err),
3452 0 : (true, Some(_)) => unreachable!(
3453 0 : "send_if_modified closure must error out if not transitioning to Stopping"
3454 0 : ),
3455 0 : (false, None) => unreachable!(
3456 0 : "send_if_modified closure must return true if transitioning to Stopping"
3457 0 : ),
3458 : }
3459 :
3460 6 : let timelines_accessor = self.timelines.lock().unwrap();
3461 6 : let not_broken_timelines = timelines_accessor
3462 6 : .values()
3463 6 : .filter(|timeline| !timeline.is_broken());
3464 12 : for timeline in not_broken_timelines {
3465 6 : timeline.set_state(TimelineState::Stopping);
3466 6 : }
3467 6 : Ok(())
3468 6 : }
3469 :
3470 : /// Method for tenant::mgr to transition us into Broken state in case of a late failure in
3471 : /// `remove_tenant_from_memory`
3472 : ///
3473 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
3474 : ///
3475 : /// In tests, we also use this to set tenants to Broken state on purpose.
3476 0 : pub(crate) async fn set_broken(&self, reason: String) {
3477 0 : let mut rx = self.state.subscribe();
3478 0 :
3479 0 : // The load & attach routines own the tenant state until it has reached `Active`.
3480 0 : // So, wait until it's done.
3481 0 : rx.wait_for(|state| match state {
3482 : TenantState::Activating(_) | TenantState::Attaching => {
3483 0 : info!(
3484 0 : "waiting for {} to turn Active|Broken|Stopping",
3485 0 : <&'static str>::from(state)
3486 : );
3487 0 : false
3488 : }
3489 0 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3490 0 : })
3491 0 : .await
3492 0 : .expect("cannot drop self.state while on a &self method");
3493 0 :
3494 0 : // we now know we're done activating, let's see whether this task is the winner to transition into Broken
3495 0 : self.set_broken_no_wait(reason)
3496 0 : }
3497 :
3498 0 : pub(crate) fn set_broken_no_wait(&self, reason: impl Display) {
3499 0 : let reason = reason.to_string();
3500 0 : self.state.send_modify(|current_state| {
3501 0 : match *current_state {
3502 : TenantState::Activating(_) | TenantState::Attaching => {
3503 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
3504 : }
3505 : TenantState::Active => {
3506 0 : if cfg!(feature = "testing") {
3507 0 : warn!("Changing Active tenant to Broken state, reason: {}", reason);
3508 0 : *current_state = TenantState::broken_from_reason(reason);
3509 : } else {
3510 0 : unreachable!("not allowed to call set_broken on Active tenants in non-testing builds")
3511 : }
3512 : }
3513 : TenantState::Broken { .. } => {
3514 0 : warn!("Tenant is already in Broken state");
3515 : }
3516 : // This is the only "expected" path, any other path is a bug.
3517 : TenantState::Stopping { .. } => {
3518 0 : warn!(
3519 0 : "Marking Stopping tenant as Broken state, reason: {}",
3520 : reason
3521 : );
3522 0 : *current_state = TenantState::broken_from_reason(reason);
3523 : }
3524 : }
3525 0 : });
3526 0 : }
3527 :
3528 0 : pub fn subscribe_for_state_updates(&self) -> watch::Receiver<TenantState> {
3529 0 : self.state.subscribe()
3530 0 : }
3531 :
3532 : /// The activate_now semaphore is initialized with zero units. As soon as
3533 : /// we add a unit, waiters will be able to acquire a unit and proceed.
3534 0 : pub(crate) fn activate_now(&self) {
3535 0 : self.activate_now_sem.add_permits(1);
3536 0 : }
3537 :
3538 0 : pub(crate) async fn wait_to_become_active(
3539 0 : &self,
3540 0 : timeout: Duration,
3541 0 : ) -> Result<(), GetActiveTenantError> {
3542 0 : let mut receiver = self.state.subscribe();
3543 : loop {
3544 0 : let current_state = receiver.borrow_and_update().clone();
3545 0 : match current_state {
3546 : TenantState::Attaching | TenantState::Activating(_) => {
3547 : // in these states, there's a chance that we can reach ::Active
3548 0 : self.activate_now();
3549 0 : match timeout_cancellable(timeout, &self.cancel, receiver.changed()).await {
3550 0 : Ok(r) => {
3551 0 : r.map_err(
3552 0 : |_e: tokio::sync::watch::error::RecvError|
3553 : // Tenant existed but was dropped: report it as non-existent
3554 0 : GetActiveTenantError::NotFound(GetTenantError::ShardNotFound(self.tenant_shard_id))
3555 0 : )?
3556 : }
3557 : Err(TimeoutCancellableError::Cancelled) => {
3558 0 : return Err(GetActiveTenantError::Cancelled);
3559 : }
3560 : Err(TimeoutCancellableError::Timeout) => {
3561 0 : return Err(GetActiveTenantError::WaitForActiveTimeout {
3562 0 : latest_state: Some(self.current_state()),
3563 0 : wait_time: timeout,
3564 0 : });
3565 : }
3566 : }
3567 : }
3568 : TenantState::Active { .. } => {
3569 0 : return Ok(());
3570 : }
3571 0 : TenantState::Broken { reason, .. } => {
3572 0 : // This is fatal, and reported distinctly from the general case of "will never be active" because
3573 0 : // it's logically a 500 to external API users (broken is always a bug).
3574 0 : return Err(GetActiveTenantError::Broken(reason));
3575 : }
3576 : TenantState::Stopping { .. } => {
3577 : // There's no chance the tenant can transition back into ::Active
3578 0 : return Err(GetActiveTenantError::WillNotBecomeActive(current_state));
3579 : }
3580 : }
3581 : }
3582 0 : }
3583 :
3584 0 : pub(crate) fn get_attach_mode(&self) -> AttachmentMode {
3585 0 : self.tenant_conf.load().location.attach_mode
3586 0 : }
3587 :
3588 : /// For API access: generate a LocationConfig equivalent to the one that would be used to
3589 : /// create a Tenant in the same state. Do not use this in hot paths: it's for relatively
3590 : /// rare external API calls, like a reconciliation at startup.
3591 0 : pub(crate) fn get_location_conf(&self) -> models::LocationConfig {
3592 0 : let conf = self.tenant_conf.load();
3593 :
3594 0 : let location_config_mode = match conf.location.attach_mode {
3595 0 : AttachmentMode::Single => models::LocationConfigMode::AttachedSingle,
3596 0 : AttachmentMode::Multi => models::LocationConfigMode::AttachedMulti,
3597 0 : AttachmentMode::Stale => models::LocationConfigMode::AttachedStale,
3598 : };
3599 :
3600 : // We have a pageserver TenantConf, we need the API-facing TenantConfig.
3601 0 : let tenant_config: models::TenantConfig = conf.tenant_conf.clone().into();
3602 0 :
3603 0 : models::LocationConfig {
3604 0 : mode: location_config_mode,
3605 0 : generation: self.generation.into(),
3606 0 : secondary_conf: None,
3607 0 : shard_number: self.shard_identity.number.0,
3608 0 : shard_count: self.shard_identity.count.literal(),
3609 0 : shard_stripe_size: self.shard_identity.stripe_size.0,
3610 0 : tenant_conf: tenant_config,
3611 0 : }
3612 0 : }
3613 :
3614 0 : pub(crate) fn get_tenant_shard_id(&self) -> &TenantShardId {
3615 0 : &self.tenant_shard_id
3616 0 : }
3617 :
3618 0 : pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
3619 0 : self.shard_identity.stripe_size
3620 0 : }
3621 :
3622 0 : pub(crate) fn get_generation(&self) -> Generation {
3623 0 : self.generation
3624 0 : }
3625 :
3626 : /// This function partially shuts down the tenant (it shuts down the Timelines) and is fallible,
3627 : /// and can leave the tenant in a bad state if it fails. The caller is responsible for
3628 : /// resetting this tenant to a valid state if we fail.
3629 0 : pub(crate) async fn split_prepare(
3630 0 : &self,
3631 0 : child_shards: &Vec<TenantShardId>,
3632 0 : ) -> anyhow::Result<()> {
3633 0 : let (timelines, offloaded) = {
3634 0 : let timelines = self.timelines.lock().unwrap();
3635 0 : let offloaded = self.timelines_offloaded.lock().unwrap();
3636 0 : (timelines.clone(), offloaded.clone())
3637 0 : };
3638 0 : let timelines_iter = timelines
3639 0 : .values()
3640 0 : .map(TimelineOrOffloadedArcRef::<'_>::from)
3641 0 : .chain(
3642 0 : offloaded
3643 0 : .values()
3644 0 : .map(TimelineOrOffloadedArcRef::<'_>::from),
3645 0 : );
3646 0 : for timeline in timelines_iter {
3647 : // We do not block timeline creation/deletion during splits inside the pageserver: it is up to higher levels
3648 : // to ensure that they do not start a split if currently in the process of doing these.
3649 :
3650 0 : let timeline_id = timeline.timeline_id();
3651 :
3652 0 : if let TimelineOrOffloadedArcRef::Timeline(timeline) = timeline {
3653 : // Upload an index from the parent: this is partly to provide freshness for the
3654 : // child tenants that will copy it, and partly for general ease-of-debugging: there will
3655 : // always be a parent shard index in the same generation as we wrote the child shard index.
3656 0 : tracing::info!(%timeline_id, "Uploading index");
3657 0 : timeline
3658 0 : .remote_client
3659 0 : .schedule_index_upload_for_file_changes()?;
3660 0 : timeline.remote_client.wait_completion().await?;
3661 0 : }
3662 :
3663 0 : let remote_client = match timeline {
3664 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.remote_client.clone(),
3665 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => {
3666 0 : let remote_client = self
3667 0 : .build_timeline_client(offloaded.timeline_id, self.remote_storage.clone());
3668 0 : Arc::new(remote_client)
3669 : }
3670 : };
3671 :
3672 : // Shut down the timeline's remote client: this means that the indices we write
3673 : // for child shards will not be invalidated by the parent shard deleting layers.
3674 0 : tracing::info!(%timeline_id, "Shutting down remote storage client");
3675 0 : remote_client.shutdown().await;
3676 :
3677 : // Download methods can still be used after shutdown, as they don't flow through the remote client's
3678 : // queue. In principal the RemoteTimelineClient could provide this without downloading it, but this
3679 : // operation is rare, so it's simpler to just download it (and robustly guarantees that the index
3680 : // we use here really is the remotely persistent one).
3681 0 : tracing::info!(%timeline_id, "Downloading index_part from parent");
3682 0 : let result = remote_client
3683 0 : .download_index_file(&self.cancel)
3684 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))
3685 0 : .await?;
3686 0 : let index_part = match result {
3687 : MaybeDeletedIndexPart::Deleted(_) => {
3688 0 : anyhow::bail!("Timeline deletion happened concurrently with split")
3689 : }
3690 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
3691 : };
3692 :
3693 0 : for child_shard in child_shards {
3694 0 : tracing::info!(%timeline_id, "Uploading index_part for child {}", child_shard.to_index());
3695 0 : upload_index_part(
3696 0 : &self.remote_storage,
3697 0 : child_shard,
3698 0 : &timeline_id,
3699 0 : self.generation,
3700 0 : &index_part,
3701 0 : &self.cancel,
3702 0 : )
3703 0 : .await?;
3704 : }
3705 : }
3706 :
3707 0 : let tenant_manifest = self.build_tenant_manifest();
3708 0 : for child_shard in child_shards {
3709 0 : tracing::info!(
3710 0 : "Uploading tenant manifest for child {}",
3711 0 : child_shard.to_index()
3712 : );
3713 0 : upload_tenant_manifest(
3714 0 : &self.remote_storage,
3715 0 : child_shard,
3716 0 : self.generation,
3717 0 : &tenant_manifest,
3718 0 : &self.cancel,
3719 0 : )
3720 0 : .await?;
3721 : }
3722 :
3723 0 : Ok(())
3724 0 : }
3725 :
3726 0 : pub(crate) fn get_sizes(&self) -> TopTenantShardItem {
3727 0 : let mut result = TopTenantShardItem {
3728 0 : id: self.tenant_shard_id,
3729 0 : resident_size: 0,
3730 0 : physical_size: 0,
3731 0 : max_logical_size: 0,
3732 0 : };
3733 :
3734 0 : for timeline in self.timelines.lock().unwrap().values() {
3735 0 : result.resident_size += timeline.metrics.resident_physical_size_gauge.get();
3736 0 :
3737 0 : result.physical_size += timeline
3738 0 : .remote_client
3739 0 : .metrics
3740 0 : .remote_physical_size_gauge
3741 0 : .get();
3742 0 : result.max_logical_size = std::cmp::max(
3743 0 : result.max_logical_size,
3744 0 : timeline.metrics.current_logical_size_gauge.get(),
3745 0 : );
3746 0 : }
3747 :
3748 0 : result
3749 0 : }
3750 : }
3751 :
3752 : /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
3753 : /// perform a topological sort, so that the parent of each timeline comes
3754 : /// before the children.
3755 : /// E extracts the ancestor from T
3756 : /// This allows for T to be different. It can be TimelineMetadata, can be Timeline itself, etc.
3757 192 : fn tree_sort_timelines<T, E>(
3758 192 : timelines: HashMap<TimelineId, T>,
3759 192 : extractor: E,
3760 192 : ) -> anyhow::Result<Vec<(TimelineId, T)>>
3761 192 : where
3762 192 : E: Fn(&T) -> Option<TimelineId>,
3763 192 : {
3764 192 : let mut result = Vec::with_capacity(timelines.len());
3765 192 :
3766 192 : let mut now = Vec::with_capacity(timelines.len());
3767 192 : // (ancestor, children)
3768 192 : let mut later: HashMap<TimelineId, Vec<(TimelineId, T)>> =
3769 192 : HashMap::with_capacity(timelines.len());
3770 :
3771 198 : for (timeline_id, value) in timelines {
3772 6 : if let Some(ancestor_id) = extractor(&value) {
3773 2 : let children = later.entry(ancestor_id).or_default();
3774 2 : children.push((timeline_id, value));
3775 4 : } else {
3776 4 : now.push((timeline_id, value));
3777 4 : }
3778 : }
3779 :
3780 198 : while let Some((timeline_id, metadata)) = now.pop() {
3781 6 : result.push((timeline_id, metadata));
3782 : // All children of this can be loaded now
3783 6 : if let Some(mut children) = later.remove(&timeline_id) {
3784 2 : now.append(&mut children);
3785 4 : }
3786 : }
3787 :
3788 : // All timelines should be visited now. Unless there were timelines with missing ancestors.
3789 192 : if !later.is_empty() {
3790 0 : for (missing_id, orphan_ids) in later {
3791 0 : for (orphan_id, _) in orphan_ids {
3792 0 : error!("could not load timeline {orphan_id} because its ancestor timeline {missing_id} could not be loaded");
3793 : }
3794 : }
3795 0 : bail!("could not load tenant because some timelines are missing ancestors");
3796 192 : }
3797 192 :
3798 192 : Ok(result)
3799 192 : }
3800 :
3801 : enum ActivateTimelineArgs {
3802 : Yes {
3803 : broker_client: storage_broker::BrokerClientChannel,
3804 : },
3805 : No,
3806 : }
3807 :
3808 : impl Tenant {
3809 0 : pub fn tenant_specific_overrides(&self) -> TenantConfOpt {
3810 0 : self.tenant_conf.load().tenant_conf.clone()
3811 0 : }
3812 :
3813 0 : pub fn effective_config(&self) -> TenantConf {
3814 0 : self.tenant_specific_overrides()
3815 0 : .merge(self.conf.default_tenant_conf.clone())
3816 0 : }
3817 :
3818 0 : pub fn get_checkpoint_distance(&self) -> u64 {
3819 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3820 0 : tenant_conf
3821 0 : .checkpoint_distance
3822 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
3823 0 : }
3824 :
3825 0 : pub fn get_checkpoint_timeout(&self) -> Duration {
3826 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3827 0 : tenant_conf
3828 0 : .checkpoint_timeout
3829 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
3830 0 : }
3831 :
3832 0 : pub fn get_compaction_target_size(&self) -> u64 {
3833 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3834 0 : tenant_conf
3835 0 : .compaction_target_size
3836 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
3837 0 : }
3838 :
3839 0 : pub fn get_compaction_period(&self) -> Duration {
3840 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3841 0 : tenant_conf
3842 0 : .compaction_period
3843 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
3844 0 : }
3845 :
3846 0 : pub fn get_compaction_threshold(&self) -> usize {
3847 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3848 0 : tenant_conf
3849 0 : .compaction_threshold
3850 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
3851 0 : }
3852 :
3853 0 : pub fn get_gc_horizon(&self) -> u64 {
3854 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3855 0 : tenant_conf
3856 0 : .gc_horizon
3857 0 : .unwrap_or(self.conf.default_tenant_conf.gc_horizon)
3858 0 : }
3859 :
3860 0 : pub fn get_gc_period(&self) -> Duration {
3861 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3862 0 : tenant_conf
3863 0 : .gc_period
3864 0 : .unwrap_or(self.conf.default_tenant_conf.gc_period)
3865 0 : }
3866 :
3867 0 : pub fn get_image_creation_threshold(&self) -> usize {
3868 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3869 0 : tenant_conf
3870 0 : .image_creation_threshold
3871 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
3872 0 : }
3873 :
3874 0 : pub fn get_pitr_interval(&self) -> Duration {
3875 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3876 0 : tenant_conf
3877 0 : .pitr_interval
3878 0 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
3879 0 : }
3880 :
3881 0 : pub fn get_min_resident_size_override(&self) -> Option<u64> {
3882 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3883 0 : tenant_conf
3884 0 : .min_resident_size_override
3885 0 : .or(self.conf.default_tenant_conf.min_resident_size_override)
3886 0 : }
3887 :
3888 0 : pub fn get_heatmap_period(&self) -> Option<Duration> {
3889 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3890 0 : let heatmap_period = tenant_conf
3891 0 : .heatmap_period
3892 0 : .unwrap_or(self.conf.default_tenant_conf.heatmap_period);
3893 0 : if heatmap_period.is_zero() {
3894 0 : None
3895 : } else {
3896 0 : Some(heatmap_period)
3897 : }
3898 0 : }
3899 :
3900 4 : pub fn get_lsn_lease_length(&self) -> Duration {
3901 4 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3902 4 : tenant_conf
3903 4 : .lsn_lease_length
3904 4 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
3905 4 : }
3906 :
3907 : /// Generate an up-to-date TenantManifest based on the state of this Tenant.
3908 2 : fn build_tenant_manifest(&self) -> TenantManifest {
3909 2 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
3910 2 :
3911 2 : let mut timeline_manifests = timelines_offloaded
3912 2 : .iter()
3913 2 : .map(|(_timeline_id, offloaded)| offloaded.manifest())
3914 2 : .collect::<Vec<_>>();
3915 2 : // Sort the manifests so that our output is deterministic
3916 2 : timeline_manifests.sort_by_key(|timeline_manifest| timeline_manifest.timeline_id);
3917 2 :
3918 2 : TenantManifest {
3919 2 : version: LATEST_TENANT_MANIFEST_VERSION,
3920 2 : offloaded_timelines: timeline_manifests,
3921 2 : }
3922 2 : }
3923 :
3924 0 : pub fn set_new_tenant_config(&self, new_tenant_conf: TenantConfOpt) {
3925 0 : // Use read-copy-update in order to avoid overwriting the location config
3926 0 : // state if this races with [`Tenant::set_new_location_config`]. Note that
3927 0 : // this race is not possible if both request types come from the storage
3928 0 : // controller (as they should!) because an exclusive op lock is required
3929 0 : // on the storage controller side.
3930 0 :
3931 0 : self.tenant_conf.rcu(|inner| {
3932 0 : Arc::new(AttachedTenantConf {
3933 0 : tenant_conf: new_tenant_conf.clone(),
3934 0 : location: inner.location,
3935 0 : // Attached location is not changed, no need to update lsn lease deadline.
3936 0 : lsn_lease_deadline: inner.lsn_lease_deadline,
3937 0 : })
3938 0 : });
3939 0 :
3940 0 : let updated = self.tenant_conf.load().clone();
3941 0 :
3942 0 : self.tenant_conf_updated(&new_tenant_conf);
3943 0 : // Don't hold self.timelines.lock() during the notifies.
3944 0 : // There's no risk of deadlock right now, but there could be if we consolidate
3945 0 : // mutexes in struct Timeline in the future.
3946 0 : let timelines = self.list_timelines();
3947 0 : for timeline in timelines {
3948 0 : timeline.tenant_conf_updated(&updated);
3949 0 : }
3950 0 : }
3951 :
3952 0 : pub(crate) fn set_new_location_config(&self, new_conf: AttachedTenantConf) {
3953 0 : let new_tenant_conf = new_conf.tenant_conf.clone();
3954 0 :
3955 0 : self.tenant_conf.store(Arc::new(new_conf.clone()));
3956 0 :
3957 0 : self.tenant_conf_updated(&new_tenant_conf);
3958 0 : // Don't hold self.timelines.lock() during the notifies.
3959 0 : // There's no risk of deadlock right now, but there could be if we consolidate
3960 0 : // mutexes in struct Timeline in the future.
3961 0 : let timelines = self.list_timelines();
3962 0 : for timeline in timelines {
3963 0 : timeline.tenant_conf_updated(&new_conf);
3964 0 : }
3965 0 : }
3966 :
3967 192 : fn get_pagestream_throttle_config(
3968 192 : psconf: &'static PageServerConf,
3969 192 : overrides: &TenantConfOpt,
3970 192 : ) -> throttle::Config {
3971 192 : overrides
3972 192 : .timeline_get_throttle
3973 192 : .clone()
3974 192 : .unwrap_or(psconf.default_tenant_conf.timeline_get_throttle.clone())
3975 192 : }
3976 :
3977 0 : pub(crate) fn tenant_conf_updated(&self, new_conf: &TenantConfOpt) {
3978 0 : let conf = Self::get_pagestream_throttle_config(self.conf, new_conf);
3979 0 : self.pagestream_throttle.reconfigure(conf)
3980 0 : }
3981 :
3982 : /// Helper function to create a new Timeline struct.
3983 : ///
3984 : /// The returned Timeline is in Loading state. The caller is responsible for
3985 : /// initializing any on-disk state, and for inserting the Timeline to the 'timelines'
3986 : /// map.
3987 : ///
3988 : /// `validate_ancestor == false` is used when a timeline is created for deletion
3989 : /// and we might not have the ancestor present anymore which is fine for to be
3990 : /// deleted timelines.
3991 : #[allow(clippy::too_many_arguments)]
3992 418 : fn create_timeline_struct(
3993 418 : &self,
3994 418 : new_timeline_id: TimelineId,
3995 418 : new_metadata: &TimelineMetadata,
3996 418 : ancestor: Option<Arc<Timeline>>,
3997 418 : resources: TimelineResources,
3998 418 : cause: CreateTimelineCause,
3999 418 : create_idempotency: CreateTimelineIdempotency,
4000 418 : ) -> anyhow::Result<Arc<Timeline>> {
4001 418 : let state = match cause {
4002 : CreateTimelineCause::Load => {
4003 418 : let ancestor_id = new_metadata.ancestor_timeline();
4004 418 : anyhow::ensure!(
4005 418 : ancestor_id == ancestor.as_ref().map(|t| t.timeline_id),
4006 0 : "Timeline's {new_timeline_id} ancestor {ancestor_id:?} was not found"
4007 : );
4008 418 : TimelineState::Loading
4009 : }
4010 0 : CreateTimelineCause::Delete => TimelineState::Stopping,
4011 : };
4012 :
4013 418 : let pg_version = new_metadata.pg_version();
4014 418 :
4015 418 : let timeline = Timeline::new(
4016 418 : self.conf,
4017 418 : Arc::clone(&self.tenant_conf),
4018 418 : new_metadata,
4019 418 : ancestor,
4020 418 : new_timeline_id,
4021 418 : self.tenant_shard_id,
4022 418 : self.generation,
4023 418 : self.shard_identity,
4024 418 : self.walredo_mgr.clone(),
4025 418 : resources,
4026 418 : pg_version,
4027 418 : state,
4028 418 : self.attach_wal_lag_cooldown.clone(),
4029 418 : create_idempotency,
4030 418 : self.cancel.child_token(),
4031 418 : );
4032 418 :
4033 418 : Ok(timeline)
4034 418 : }
4035 :
4036 : // Allow too_many_arguments because a constructor's argument list naturally grows with the
4037 : // number of attributes in the struct: breaking these out into a builder wouldn't be helpful.
4038 : #[allow(clippy::too_many_arguments)]
4039 192 : fn new(
4040 192 : state: TenantState,
4041 192 : conf: &'static PageServerConf,
4042 192 : attached_conf: AttachedTenantConf,
4043 192 : shard_identity: ShardIdentity,
4044 192 : walredo_mgr: Option<Arc<WalRedoManager>>,
4045 192 : tenant_shard_id: TenantShardId,
4046 192 : remote_storage: GenericRemoteStorage,
4047 192 : deletion_queue_client: DeletionQueueClient,
4048 192 : l0_flush_global_state: L0FlushGlobalState,
4049 192 : ) -> Tenant {
4050 192 : debug_assert!(
4051 192 : !attached_conf.location.generation.is_none() || conf.control_plane_api.is_none()
4052 : );
4053 :
4054 192 : let (state, mut rx) = watch::channel(state);
4055 192 :
4056 192 : tokio::spawn(async move {
4057 192 : // reflect tenant state in metrics:
4058 192 : // - global per tenant state: TENANT_STATE_METRIC
4059 192 : // - "set" of broken tenants: BROKEN_TENANTS_SET
4060 192 : //
4061 192 : // set of broken tenants should not have zero counts so that it remains accessible for
4062 192 : // alerting.
4063 192 :
4064 192 : let tid = tenant_shard_id.to_string();
4065 192 : let shard_id = tenant_shard_id.shard_slug().to_string();
4066 192 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
4067 :
4068 384 : fn inspect_state(state: &TenantState) -> ([&'static str; 1], bool) {
4069 384 : ([state.into()], matches!(state, TenantState::Broken { .. }))
4070 384 : }
4071 :
4072 192 : let mut tuple = inspect_state(&rx.borrow_and_update());
4073 192 :
4074 192 : let is_broken = tuple.1;
4075 192 : let mut counted_broken = if is_broken {
4076 : // add the id to the set right away, there should not be any updates on the channel
4077 : // after before tenant is removed, if ever
4078 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
4079 0 : true
4080 : } else {
4081 192 : false
4082 : };
4083 :
4084 : loop {
4085 384 : let labels = &tuple.0;
4086 384 : let current = TENANT_STATE_METRIC.with_label_values(labels);
4087 384 : current.inc();
4088 384 :
4089 384 : if rx.changed().await.is_err() {
4090 : // tenant has been dropped
4091 14 : current.dec();
4092 14 : drop(BROKEN_TENANTS_SET.remove_label_values(set_key));
4093 14 : break;
4094 192 : }
4095 192 :
4096 192 : current.dec();
4097 192 : tuple = inspect_state(&rx.borrow_and_update());
4098 192 :
4099 192 : let is_broken = tuple.1;
4100 192 : if is_broken && !counted_broken {
4101 0 : counted_broken = true;
4102 0 : // insert the tenant_id (back) into the set while avoiding needless counter
4103 0 : // access
4104 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
4105 192 : }
4106 : }
4107 192 : });
4108 192 :
4109 192 : Tenant {
4110 192 : tenant_shard_id,
4111 192 : shard_identity,
4112 192 : generation: attached_conf.location.generation,
4113 192 : conf,
4114 192 : // using now here is good enough approximation to catch tenants with really long
4115 192 : // activation times.
4116 192 : constructed_at: Instant::now(),
4117 192 : timelines: Mutex::new(HashMap::new()),
4118 192 : timelines_creating: Mutex::new(HashSet::new()),
4119 192 : timelines_offloaded: Mutex::new(HashMap::new()),
4120 192 : tenant_manifest_upload: Default::default(),
4121 192 : gc_cs: tokio::sync::Mutex::new(()),
4122 192 : walredo_mgr,
4123 192 : remote_storage,
4124 192 : deletion_queue_client,
4125 192 : state,
4126 192 : cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
4127 192 : cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
4128 192 : eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
4129 192 : compaction_circuit_breaker: std::sync::Mutex::new(CircuitBreaker::new(
4130 192 : format!("compaction-{tenant_shard_id}"),
4131 192 : 5,
4132 192 : // Compaction can be a very expensive operation, and might leak disk space. It also ought
4133 192 : // to be infallible, as long as remote storage is available. So if it repeatedly fails,
4134 192 : // use an extremely long backoff.
4135 192 : Some(Duration::from_secs(3600 * 24)),
4136 192 : )),
4137 192 : scheduled_compaction_tasks: Mutex::new(Default::default()),
4138 192 : activate_now_sem: tokio::sync::Semaphore::new(0),
4139 192 : attach_wal_lag_cooldown: Arc::new(std::sync::OnceLock::new()),
4140 192 : cancel: CancellationToken::default(),
4141 192 : gate: Gate::default(),
4142 192 : pagestream_throttle: Arc::new(throttle::Throttle::new(
4143 192 : Tenant::get_pagestream_throttle_config(conf, &attached_conf.tenant_conf),
4144 192 : crate::metrics::tenant_throttling::Metrics::new(&tenant_shard_id),
4145 192 : )),
4146 192 : tenant_conf: Arc::new(ArcSwap::from_pointee(attached_conf)),
4147 192 : ongoing_timeline_detach: std::sync::Mutex::default(),
4148 192 : gc_block: Default::default(),
4149 192 : l0_flush_global_state,
4150 192 : }
4151 192 : }
4152 :
4153 : /// Locate and load config
4154 0 : pub(super) fn load_tenant_config(
4155 0 : conf: &'static PageServerConf,
4156 0 : tenant_shard_id: &TenantShardId,
4157 0 : ) -> Result<LocationConf, LoadConfigError> {
4158 0 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4159 0 :
4160 0 : info!("loading tenant configuration from {config_path}");
4161 :
4162 : // load and parse file
4163 0 : let config = fs::read_to_string(&config_path).map_err(|e| {
4164 0 : match e.kind() {
4165 : std::io::ErrorKind::NotFound => {
4166 : // The config should almost always exist for a tenant directory:
4167 : // - When attaching a tenant, the config is the first thing we write
4168 : // - When detaching a tenant, we atomically move the directory to a tmp location
4169 : // before deleting contents.
4170 : //
4171 : // The very rare edge case that can result in a missing config is if we crash during attach
4172 : // between creating directory and writing config. Callers should handle that as if the
4173 : // directory didn't exist.
4174 :
4175 0 : LoadConfigError::NotFound(config_path)
4176 : }
4177 : _ => {
4178 : // No IO errors except NotFound are acceptable here: other kinds of error indicate local storage or permissions issues
4179 : // that we cannot cleanly recover
4180 0 : crate::virtual_file::on_fatal_io_error(&e, "Reading tenant config file")
4181 : }
4182 : }
4183 0 : })?;
4184 :
4185 0 : Ok(toml_edit::de::from_str::<LocationConf>(&config)?)
4186 0 : }
4187 :
4188 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4189 : pub(super) async fn persist_tenant_config(
4190 : conf: &'static PageServerConf,
4191 : tenant_shard_id: &TenantShardId,
4192 : location_conf: &LocationConf,
4193 : ) -> std::io::Result<()> {
4194 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4195 :
4196 : Self::persist_tenant_config_at(tenant_shard_id, &config_path, location_conf).await
4197 : }
4198 :
4199 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4200 : pub(super) async fn persist_tenant_config_at(
4201 : tenant_shard_id: &TenantShardId,
4202 : config_path: &Utf8Path,
4203 : location_conf: &LocationConf,
4204 : ) -> std::io::Result<()> {
4205 : debug!("persisting tenantconf to {config_path}");
4206 :
4207 : let mut conf_content = r#"# This file contains a specific per-tenant's config.
4208 : # It is read in case of pageserver restart.
4209 : "#
4210 : .to_string();
4211 :
4212 0 : fail::fail_point!("tenant-config-before-write", |_| {
4213 0 : Err(std::io::Error::new(
4214 0 : std::io::ErrorKind::Other,
4215 0 : "tenant-config-before-write",
4216 0 : ))
4217 0 : });
4218 :
4219 : // Convert the config to a toml file.
4220 : conf_content +=
4221 : &toml_edit::ser::to_string_pretty(&location_conf).expect("Config serialization failed");
4222 :
4223 : let temp_path = path_with_suffix_extension(config_path, TEMP_FILE_SUFFIX);
4224 :
4225 : let conf_content = conf_content.into_bytes();
4226 : VirtualFile::crashsafe_overwrite(config_path.to_owned(), temp_path, conf_content).await
4227 : }
4228 :
4229 : //
4230 : // How garbage collection works:
4231 : //
4232 : // +--bar------------->
4233 : // /
4234 : // +----+-----foo---------------->
4235 : // /
4236 : // ----main--+-------------------------->
4237 : // \
4238 : // +-----baz-------->
4239 : //
4240 : //
4241 : // 1. Grab 'gc_cs' mutex to prevent new timelines from being created while Timeline's
4242 : // `gc_infos` are being refreshed
4243 : // 2. Scan collected timelines, and on each timeline, make note of the
4244 : // all the points where other timelines have been branched off.
4245 : // We will refrain from removing page versions at those LSNs.
4246 : // 3. For each timeline, scan all layer files on the timeline.
4247 : // Remove all files for which a newer file exists and which
4248 : // don't cover any branch point LSNs.
4249 : //
4250 : // TODO:
4251 : // - if a relation has a non-incremental persistent layer on a child branch, then we
4252 : // don't need to keep that in the parent anymore. But currently
4253 : // we do.
4254 4 : async fn gc_iteration_internal(
4255 4 : &self,
4256 4 : target_timeline_id: Option<TimelineId>,
4257 4 : horizon: u64,
4258 4 : pitr: Duration,
4259 4 : cancel: &CancellationToken,
4260 4 : ctx: &RequestContext,
4261 4 : ) -> Result<GcResult, GcError> {
4262 4 : let mut totals: GcResult = Default::default();
4263 4 : let now = Instant::now();
4264 :
4265 4 : let gc_timelines = self
4266 4 : .refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4267 4 : .await?;
4268 :
4269 4 : failpoint_support::sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
4270 :
4271 : // If there is nothing to GC, we don't want any messages in the INFO log.
4272 4 : if !gc_timelines.is_empty() {
4273 4 : info!("{} timelines need GC", gc_timelines.len());
4274 : } else {
4275 0 : debug!("{} timelines need GC", gc_timelines.len());
4276 : }
4277 :
4278 : // Perform GC for each timeline.
4279 : //
4280 : // Note that we don't hold the `Tenant::gc_cs` lock here because we don't want to delay the
4281 : // branch creation task, which requires the GC lock. A GC iteration can run concurrently
4282 : // with branch creation.
4283 : //
4284 : // See comments in [`Tenant::branch_timeline`] for more information about why branch
4285 : // creation task can run concurrently with timeline's GC iteration.
4286 8 : for timeline in gc_timelines {
4287 4 : if cancel.is_cancelled() {
4288 : // We were requested to shut down. Stop and return with the progress we
4289 : // made.
4290 0 : break;
4291 4 : }
4292 4 : let result = match timeline.gc().await {
4293 : Err(GcError::TimelineCancelled) => {
4294 0 : if target_timeline_id.is_some() {
4295 : // If we were targetting this specific timeline, surface cancellation to caller
4296 0 : return Err(GcError::TimelineCancelled);
4297 : } else {
4298 : // A timeline may be shutting down independently of the tenant's lifecycle: we should
4299 : // skip past this and proceed to try GC on other timelines.
4300 0 : continue;
4301 : }
4302 : }
4303 4 : r => r?,
4304 : };
4305 4 : totals += result;
4306 : }
4307 :
4308 4 : totals.elapsed = now.elapsed();
4309 4 : Ok(totals)
4310 4 : }
4311 :
4312 : /// Refreshes the Timeline::gc_info for all timelines, returning the
4313 : /// vector of timelines which have [`Timeline::get_last_record_lsn`] past
4314 : /// [`Tenant::get_gc_horizon`].
4315 : ///
4316 : /// This is usually executed as part of periodic gc, but can now be triggered more often.
4317 0 : pub(crate) async fn refresh_gc_info(
4318 0 : &self,
4319 0 : cancel: &CancellationToken,
4320 0 : ctx: &RequestContext,
4321 0 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4322 0 : // since this method can now be called at different rates than the configured gc loop, it
4323 0 : // might be that these configuration values get applied faster than what it was previously,
4324 0 : // since these were only read from the gc task.
4325 0 : let horizon = self.get_gc_horizon();
4326 0 : let pitr = self.get_pitr_interval();
4327 0 :
4328 0 : // refresh all timelines
4329 0 : let target_timeline_id = None;
4330 0 :
4331 0 : self.refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4332 0 : .await
4333 0 : }
4334 :
4335 : /// Populate all Timelines' `GcInfo` with information about their children. We do not set the
4336 : /// PITR cutoffs here, because that requires I/O: this is done later, before GC, by [`Self::refresh_gc_info_internal`]
4337 : ///
4338 : /// Subsequently, parent-child relationships are updated incrementally inside [`Timeline::new`] and [`Timeline::drop`].
4339 0 : fn initialize_gc_info(
4340 0 : &self,
4341 0 : timelines: &std::sync::MutexGuard<HashMap<TimelineId, Arc<Timeline>>>,
4342 0 : timelines_offloaded: &std::sync::MutexGuard<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
4343 0 : restrict_to_timeline: Option<TimelineId>,
4344 0 : ) {
4345 0 : if restrict_to_timeline.is_none() {
4346 : // This function must be called before activation: after activation timeline create/delete operations
4347 : // might happen, and this function is not safe to run concurrently with those.
4348 0 : assert!(!self.is_active());
4349 0 : }
4350 :
4351 : // Scan all timelines. For each timeline, remember the timeline ID and
4352 : // the branch point where it was created.
4353 0 : let mut all_branchpoints: BTreeMap<TimelineId, Vec<(Lsn, TimelineId, MaybeOffloaded)>> =
4354 0 : BTreeMap::new();
4355 0 : timelines.iter().for_each(|(timeline_id, timeline_entry)| {
4356 0 : if let Some(ancestor_timeline_id) = &timeline_entry.get_ancestor_timeline_id() {
4357 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4358 0 : ancestor_children.push((
4359 0 : timeline_entry.get_ancestor_lsn(),
4360 0 : *timeline_id,
4361 0 : MaybeOffloaded::No,
4362 0 : ));
4363 0 : }
4364 0 : });
4365 0 : timelines_offloaded
4366 0 : .iter()
4367 0 : .for_each(|(timeline_id, timeline_entry)| {
4368 0 : let Some(ancestor_timeline_id) = &timeline_entry.ancestor_timeline_id else {
4369 0 : return;
4370 : };
4371 0 : let Some(retain_lsn) = timeline_entry.ancestor_retain_lsn else {
4372 0 : return;
4373 : };
4374 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4375 0 : ancestor_children.push((retain_lsn, *timeline_id, MaybeOffloaded::Yes));
4376 0 : });
4377 0 :
4378 0 : // The number of bytes we always keep, irrespective of PITR: this is a constant across timelines
4379 0 : let horizon = self.get_gc_horizon();
4380 :
4381 : // Populate each timeline's GcInfo with information about its child branches
4382 0 : let timelines_to_write = if let Some(timeline_id) = restrict_to_timeline {
4383 0 : itertools::Either::Left(timelines.get(&timeline_id).into_iter())
4384 : } else {
4385 0 : itertools::Either::Right(timelines.values())
4386 : };
4387 0 : for timeline in timelines_to_write {
4388 0 : let mut branchpoints: Vec<(Lsn, TimelineId, MaybeOffloaded)> = all_branchpoints
4389 0 : .remove(&timeline.timeline_id)
4390 0 : .unwrap_or_default();
4391 0 :
4392 0 : branchpoints.sort_by_key(|b| b.0);
4393 0 :
4394 0 : let mut target = timeline.gc_info.write().unwrap();
4395 0 :
4396 0 : target.retain_lsns = branchpoints;
4397 0 :
4398 0 : let space_cutoff = timeline
4399 0 : .get_last_record_lsn()
4400 0 : .checked_sub(horizon)
4401 0 : .unwrap_or(Lsn(0));
4402 0 :
4403 0 : target.cutoffs = GcCutoffs {
4404 0 : space: space_cutoff,
4405 0 : time: Lsn::INVALID,
4406 0 : };
4407 0 : }
4408 0 : }
4409 :
4410 4 : async fn refresh_gc_info_internal(
4411 4 : &self,
4412 4 : target_timeline_id: Option<TimelineId>,
4413 4 : horizon: u64,
4414 4 : pitr: Duration,
4415 4 : cancel: &CancellationToken,
4416 4 : ctx: &RequestContext,
4417 4 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4418 4 : // before taking the gc_cs lock, do the heavier weight finding of gc_cutoff points for
4419 4 : // currently visible timelines.
4420 4 : let timelines = self
4421 4 : .timelines
4422 4 : .lock()
4423 4 : .unwrap()
4424 4 : .values()
4425 4 : .filter(|tl| match target_timeline_id.as_ref() {
4426 4 : Some(target) => &tl.timeline_id == target,
4427 0 : None => true,
4428 4 : })
4429 4 : .cloned()
4430 4 : .collect::<Vec<_>>();
4431 4 :
4432 4 : if target_timeline_id.is_some() && timelines.is_empty() {
4433 : // We were to act on a particular timeline and it wasn't found
4434 0 : return Err(GcError::TimelineNotFound);
4435 4 : }
4436 4 :
4437 4 : let mut gc_cutoffs: HashMap<TimelineId, GcCutoffs> =
4438 4 : HashMap::with_capacity(timelines.len());
4439 :
4440 4 : for timeline in timelines.iter() {
4441 4 : let cutoff = timeline
4442 4 : .get_last_record_lsn()
4443 4 : .checked_sub(horizon)
4444 4 : .unwrap_or(Lsn(0));
4445 :
4446 4 : let cutoffs = timeline.find_gc_cutoffs(cutoff, pitr, cancel, ctx).await?;
4447 4 : let old = gc_cutoffs.insert(timeline.timeline_id, cutoffs);
4448 4 : assert!(old.is_none());
4449 : }
4450 :
4451 4 : if !self.is_active() || self.cancel.is_cancelled() {
4452 0 : return Err(GcError::TenantCancelled);
4453 4 : }
4454 :
4455 : // grab mutex to prevent new timelines from being created here; avoid doing long operations
4456 : // because that will stall branch creation.
4457 4 : let gc_cs = self.gc_cs.lock().await;
4458 :
4459 : // Ok, we now know all the branch points.
4460 : // Update the GC information for each timeline.
4461 4 : let mut gc_timelines = Vec::with_capacity(timelines.len());
4462 8 : for timeline in timelines {
4463 : // We filtered the timeline list above
4464 4 : if let Some(target_timeline_id) = target_timeline_id {
4465 4 : assert_eq!(target_timeline_id, timeline.timeline_id);
4466 0 : }
4467 :
4468 : {
4469 4 : let mut target = timeline.gc_info.write().unwrap();
4470 4 :
4471 4 : // Cull any expired leases
4472 4 : let now = SystemTime::now();
4473 6 : target.leases.retain(|_, lease| !lease.is_expired(&now));
4474 4 :
4475 4 : timeline
4476 4 : .metrics
4477 4 : .valid_lsn_lease_count_gauge
4478 4 : .set(target.leases.len() as u64);
4479 :
4480 : // Look up parent's PITR cutoff to update the child's knowledge of whether it is within parent's PITR
4481 4 : if let Some(ancestor_id) = timeline.get_ancestor_timeline_id() {
4482 0 : if let Some(ancestor_gc_cutoffs) = gc_cutoffs.get(&ancestor_id) {
4483 0 : target.within_ancestor_pitr =
4484 0 : timeline.get_ancestor_lsn() >= ancestor_gc_cutoffs.time;
4485 0 : }
4486 4 : }
4487 :
4488 : // Update metrics that depend on GC state
4489 4 : timeline
4490 4 : .metrics
4491 4 : .archival_size
4492 4 : .set(if target.within_ancestor_pitr {
4493 0 : timeline.metrics.current_logical_size_gauge.get()
4494 : } else {
4495 4 : 0
4496 : });
4497 4 : timeline.metrics.pitr_history_size.set(
4498 4 : timeline
4499 4 : .get_last_record_lsn()
4500 4 : .checked_sub(target.cutoffs.time)
4501 4 : .unwrap_or(Lsn(0))
4502 4 : .0,
4503 4 : );
4504 :
4505 : // Apply the cutoffs we found to the Timeline's GcInfo. Why might we _not_ have cutoffs for a timeline?
4506 : // - this timeline was created while we were finding cutoffs
4507 : // - lsn for timestamp search fails for this timeline repeatedly
4508 4 : if let Some(cutoffs) = gc_cutoffs.get(&timeline.timeline_id) {
4509 4 : let original_cutoffs = target.cutoffs.clone();
4510 4 : // GC cutoffs should never go back
4511 4 : target.cutoffs = GcCutoffs {
4512 4 : space: Lsn(cutoffs.space.0.max(original_cutoffs.space.0)),
4513 4 : time: Lsn(cutoffs.time.0.max(original_cutoffs.time.0)),
4514 4 : }
4515 0 : }
4516 : }
4517 :
4518 4 : gc_timelines.push(timeline);
4519 : }
4520 4 : drop(gc_cs);
4521 4 : Ok(gc_timelines)
4522 4 : }
4523 :
4524 : /// A substitute for `branch_timeline` for use in unit tests.
4525 : /// The returned timeline will have state value `Active` to make various `anyhow::ensure!()`
4526 : /// calls pass, but, we do not actually call `.activate()` under the hood. So, none of the
4527 : /// timeline background tasks are launched, except the flush loop.
4528 : #[cfg(test)]
4529 232 : async fn branch_timeline_test(
4530 232 : self: &Arc<Self>,
4531 232 : src_timeline: &Arc<Timeline>,
4532 232 : dst_id: TimelineId,
4533 232 : ancestor_lsn: Option<Lsn>,
4534 232 : ctx: &RequestContext,
4535 232 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
4536 232 : let tl = self
4537 232 : .branch_timeline_impl(src_timeline, dst_id, ancestor_lsn, ctx)
4538 232 : .await?
4539 228 : .into_timeline_for_test();
4540 228 : tl.set_state(TimelineState::Active);
4541 228 : Ok(tl)
4542 232 : }
4543 :
4544 : /// Helper for unit tests to branch a timeline with some pre-loaded states.
4545 : #[cfg(test)]
4546 : #[allow(clippy::too_many_arguments)]
4547 6 : pub async fn branch_timeline_test_with_layers(
4548 6 : self: &Arc<Self>,
4549 6 : src_timeline: &Arc<Timeline>,
4550 6 : dst_id: TimelineId,
4551 6 : ancestor_lsn: Option<Lsn>,
4552 6 : ctx: &RequestContext,
4553 6 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
4554 6 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
4555 6 : end_lsn: Lsn,
4556 6 : ) -> anyhow::Result<Arc<Timeline>> {
4557 : use checks::check_valid_layermap;
4558 : use itertools::Itertools;
4559 :
4560 6 : let tline = self
4561 6 : .branch_timeline_test(src_timeline, dst_id, ancestor_lsn, ctx)
4562 6 : .await?;
4563 6 : let ancestor_lsn = if let Some(ancestor_lsn) = ancestor_lsn {
4564 6 : ancestor_lsn
4565 : } else {
4566 0 : tline.get_last_record_lsn()
4567 : };
4568 6 : assert!(end_lsn >= ancestor_lsn);
4569 6 : tline.force_advance_lsn(end_lsn);
4570 12 : for deltas in delta_layer_desc {
4571 6 : tline
4572 6 : .force_create_delta_layer(deltas, Some(ancestor_lsn), ctx)
4573 6 : .await?;
4574 : }
4575 10 : for (lsn, images) in image_layer_desc {
4576 4 : tline
4577 4 : .force_create_image_layer(lsn, images, Some(ancestor_lsn), ctx)
4578 4 : .await?;
4579 : }
4580 6 : let layer_names = tline
4581 6 : .layers
4582 6 : .read()
4583 6 : .await
4584 6 : .layer_map()
4585 6 : .unwrap()
4586 6 : .iter_historic_layers()
4587 10 : .map(|layer| layer.layer_name())
4588 6 : .collect_vec();
4589 6 : if let Some(err) = check_valid_layermap(&layer_names) {
4590 0 : bail!("invalid layermap: {err}");
4591 6 : }
4592 6 : Ok(tline)
4593 6 : }
4594 :
4595 : /// Branch an existing timeline.
4596 0 : async fn branch_timeline(
4597 0 : self: &Arc<Self>,
4598 0 : src_timeline: &Arc<Timeline>,
4599 0 : dst_id: TimelineId,
4600 0 : start_lsn: Option<Lsn>,
4601 0 : ctx: &RequestContext,
4602 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4603 0 : self.branch_timeline_impl(src_timeline, dst_id, start_lsn, ctx)
4604 0 : .await
4605 0 : }
4606 :
4607 232 : async fn branch_timeline_impl(
4608 232 : self: &Arc<Self>,
4609 232 : src_timeline: &Arc<Timeline>,
4610 232 : dst_id: TimelineId,
4611 232 : start_lsn: Option<Lsn>,
4612 232 : _ctx: &RequestContext,
4613 232 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4614 232 : let src_id = src_timeline.timeline_id;
4615 :
4616 : // We will validate our ancestor LSN in this function. Acquire the GC lock so that
4617 : // this check cannot race with GC, and the ancestor LSN is guaranteed to remain
4618 : // valid while we are creating the branch.
4619 232 : let _gc_cs = self.gc_cs.lock().await;
4620 :
4621 : // If no start LSN is specified, we branch the new timeline from the source timeline's last record LSN
4622 232 : let start_lsn = start_lsn.unwrap_or_else(|| {
4623 2 : let lsn = src_timeline.get_last_record_lsn();
4624 2 : info!("branching timeline {dst_id} from timeline {src_id} at last record LSN: {lsn}");
4625 2 : lsn
4626 232 : });
4627 :
4628 : // we finally have determined the ancestor_start_lsn, so we can get claim exclusivity now
4629 232 : let timeline_create_guard = match self
4630 232 : .start_creating_timeline(
4631 232 : dst_id,
4632 232 : CreateTimelineIdempotency::Branch {
4633 232 : ancestor_timeline_id: src_timeline.timeline_id,
4634 232 : ancestor_start_lsn: start_lsn,
4635 232 : },
4636 232 : )
4637 232 : .await?
4638 : {
4639 232 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
4640 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
4641 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
4642 : }
4643 : };
4644 :
4645 : // Ensure that `start_lsn` is valid, i.e. the LSN is within the PITR
4646 : // horizon on the source timeline
4647 : //
4648 : // We check it against both the planned GC cutoff stored in 'gc_info',
4649 : // and the 'latest_gc_cutoff' of the last GC that was performed. The
4650 : // planned GC cutoff in 'gc_info' is normally larger than
4651 : // 'latest_gc_cutoff_lsn', but beware of corner cases like if you just
4652 : // changed the GC settings for the tenant to make the PITR window
4653 : // larger, but some of the data was already removed by an earlier GC
4654 : // iteration.
4655 :
4656 : // check against last actual 'latest_gc_cutoff' first
4657 232 : let latest_gc_cutoff_lsn = src_timeline.get_latest_gc_cutoff_lsn();
4658 232 : src_timeline
4659 232 : .check_lsn_is_in_scope(start_lsn, &latest_gc_cutoff_lsn)
4660 232 : .context(format!(
4661 232 : "invalid branch start lsn: less than latest GC cutoff {}",
4662 232 : *latest_gc_cutoff_lsn,
4663 232 : ))
4664 232 : .map_err(CreateTimelineError::AncestorLsn)?;
4665 :
4666 : // and then the planned GC cutoff
4667 : {
4668 228 : let gc_info = src_timeline.gc_info.read().unwrap();
4669 228 : let cutoff = gc_info.min_cutoff();
4670 228 : if start_lsn < cutoff {
4671 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
4672 0 : "invalid branch start lsn: less than planned GC cutoff {cutoff}"
4673 0 : )));
4674 228 : }
4675 228 : }
4676 228 :
4677 228 : //
4678 228 : // The branch point is valid, and we are still holding the 'gc_cs' lock
4679 228 : // so that GC cannot advance the GC cutoff until we are finished.
4680 228 : // Proceed with the branch creation.
4681 228 : //
4682 228 :
4683 228 : // Determine prev-LSN for the new timeline. We can only determine it if
4684 228 : // the timeline was branched at the current end of the source timeline.
4685 228 : let RecordLsn {
4686 228 : last: src_last,
4687 228 : prev: src_prev,
4688 228 : } = src_timeline.get_last_record_rlsn();
4689 228 : let dst_prev = if src_last == start_lsn {
4690 216 : Some(src_prev)
4691 : } else {
4692 12 : None
4693 : };
4694 :
4695 : // Create the metadata file, noting the ancestor of the new timeline.
4696 : // There is initially no data in it, but all the read-calls know to look
4697 : // into the ancestor.
4698 228 : let metadata = TimelineMetadata::new(
4699 228 : start_lsn,
4700 228 : dst_prev,
4701 228 : Some(src_id),
4702 228 : start_lsn,
4703 228 : *src_timeline.latest_gc_cutoff_lsn.read(), // FIXME: should we hold onto this guard longer?
4704 228 : src_timeline.initdb_lsn,
4705 228 : src_timeline.pg_version,
4706 228 : );
4707 :
4708 228 : let uninitialized_timeline = self
4709 228 : .prepare_new_timeline(
4710 228 : dst_id,
4711 228 : &metadata,
4712 228 : timeline_create_guard,
4713 228 : start_lsn + 1,
4714 228 : Some(Arc::clone(src_timeline)),
4715 228 : )
4716 228 : .await?;
4717 :
4718 228 : let new_timeline = uninitialized_timeline.finish_creation()?;
4719 :
4720 : // Root timeline gets its layers during creation and uploads them along with the metadata.
4721 : // A branch timeline though, when created, can get no writes for some time, hence won't get any layers created.
4722 : // We still need to upload its metadata eagerly: if other nodes `attach` the tenant and miss this timeline, their GC
4723 : // could get incorrect information and remove more layers, than needed.
4724 : // See also https://github.com/neondatabase/neon/issues/3865
4725 228 : new_timeline
4726 228 : .remote_client
4727 228 : .schedule_index_upload_for_full_metadata_update(&metadata)
4728 228 : .context("branch initial metadata upload")?;
4729 :
4730 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
4731 :
4732 228 : Ok(CreateTimelineResult::Created(new_timeline))
4733 232 : }
4734 :
4735 : /// For unit tests, make this visible so that other modules can directly create timelines
4736 : #[cfg(test)]
4737 2 : #[tracing::instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
4738 : pub(crate) async fn bootstrap_timeline_test(
4739 : self: &Arc<Self>,
4740 : timeline_id: TimelineId,
4741 : pg_version: u32,
4742 : load_existing_initdb: Option<TimelineId>,
4743 : ctx: &RequestContext,
4744 : ) -> anyhow::Result<Arc<Timeline>> {
4745 : self.bootstrap_timeline(timeline_id, pg_version, load_existing_initdb, ctx)
4746 : .await
4747 : .map_err(anyhow::Error::new)
4748 2 : .map(|r| r.into_timeline_for_test())
4749 : }
4750 :
4751 : /// Get exclusive access to the timeline ID for creation.
4752 : ///
4753 : /// Timeline-creating code paths must use this function before making changes
4754 : /// to in-memory or persistent state.
4755 : ///
4756 : /// The `state` parameter is a description of the timeline creation operation
4757 : /// we intend to perform.
4758 : /// If the timeline was already created in the meantime, we check whether this
4759 : /// request conflicts or is idempotent , based on `state`.
4760 418 : async fn start_creating_timeline(
4761 418 : self: &Arc<Self>,
4762 418 : new_timeline_id: TimelineId,
4763 418 : idempotency: CreateTimelineIdempotency,
4764 418 : ) -> Result<StartCreatingTimelineResult, CreateTimelineError> {
4765 418 : let allow_offloaded = false;
4766 418 : match self.create_timeline_create_guard(new_timeline_id, idempotency, allow_offloaded) {
4767 416 : Ok(create_guard) => {
4768 416 : pausable_failpoint!("timeline-creation-after-uninit");
4769 416 : Ok(StartCreatingTimelineResult::CreateGuard(create_guard))
4770 : }
4771 0 : Err(TimelineExclusionError::ShuttingDown) => Err(CreateTimelineError::ShuttingDown),
4772 : Err(TimelineExclusionError::AlreadyCreating) => {
4773 : // Creation is in progress, we cannot create it again, and we cannot
4774 : // check if this request matches the existing one, so caller must try
4775 : // again later.
4776 0 : Err(CreateTimelineError::AlreadyCreating)
4777 : }
4778 0 : Err(TimelineExclusionError::Other(e)) => Err(CreateTimelineError::Other(e)),
4779 : Err(TimelineExclusionError::AlreadyExists {
4780 0 : existing: TimelineOrOffloaded::Offloaded(_existing),
4781 0 : ..
4782 0 : }) => {
4783 0 : info!("timeline already exists but is offloaded");
4784 0 : Err(CreateTimelineError::Conflict)
4785 : }
4786 : Err(TimelineExclusionError::AlreadyExists {
4787 2 : existing: TimelineOrOffloaded::Timeline(existing),
4788 2 : arg,
4789 2 : }) => {
4790 2 : {
4791 2 : let existing = &existing.create_idempotency;
4792 2 : let _span = info_span!("idempotency_check", ?existing, ?arg).entered();
4793 2 : debug!("timeline already exists");
4794 :
4795 2 : match (existing, &arg) {
4796 : // FailWithConflict => no idempotency check
4797 : (CreateTimelineIdempotency::FailWithConflict, _)
4798 : | (_, CreateTimelineIdempotency::FailWithConflict) => {
4799 2 : warn!("timeline already exists, failing request");
4800 2 : return Err(CreateTimelineError::Conflict);
4801 : }
4802 : // Idempotent <=> CreateTimelineIdempotency is identical
4803 0 : (x, y) if x == y => {
4804 0 : info!("timeline already exists and idempotency matches, succeeding request");
4805 : // fallthrough
4806 : }
4807 : (_, _) => {
4808 0 : warn!("idempotency conflict, failing request");
4809 0 : return Err(CreateTimelineError::Conflict);
4810 : }
4811 : }
4812 : }
4813 :
4814 0 : Ok(StartCreatingTimelineResult::Idempotent(existing))
4815 : }
4816 : }
4817 418 : }
4818 :
4819 0 : async fn upload_initdb(
4820 0 : &self,
4821 0 : timelines_path: &Utf8PathBuf,
4822 0 : pgdata_path: &Utf8PathBuf,
4823 0 : timeline_id: &TimelineId,
4824 0 : ) -> anyhow::Result<()> {
4825 0 : let temp_path = timelines_path.join(format!(
4826 0 : "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
4827 0 : ));
4828 0 :
4829 0 : scopeguard::defer! {
4830 0 : if let Err(e) = fs::remove_file(&temp_path) {
4831 0 : error!("Failed to remove temporary initdb archive '{temp_path}': {e}");
4832 0 : }
4833 0 : }
4834 :
4835 0 : let (pgdata_zstd, tar_zst_size) = create_zst_tarball(pgdata_path, &temp_path).await?;
4836 : const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
4837 0 : if tar_zst_size > INITDB_TAR_ZST_WARN_LIMIT {
4838 0 : warn!(
4839 0 : "compressed {temp_path} size of {tar_zst_size} is above limit {INITDB_TAR_ZST_WARN_LIMIT}."
4840 : );
4841 0 : }
4842 :
4843 0 : pausable_failpoint!("before-initdb-upload");
4844 :
4845 0 : backoff::retry(
4846 0 : || async {
4847 0 : self::remote_timeline_client::upload_initdb_dir(
4848 0 : &self.remote_storage,
4849 0 : &self.tenant_shard_id.tenant_id,
4850 0 : timeline_id,
4851 0 : pgdata_zstd.try_clone().await?,
4852 0 : tar_zst_size,
4853 0 : &self.cancel,
4854 0 : )
4855 0 : .await
4856 0 : },
4857 0 : |_| false,
4858 0 : 3,
4859 0 : u32::MAX,
4860 0 : "persist_initdb_tar_zst",
4861 0 : &self.cancel,
4862 0 : )
4863 0 : .await
4864 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
4865 0 : .and_then(|x| x)
4866 0 : }
4867 :
4868 : /// - run initdb to init temporary instance and get bootstrap data
4869 : /// - after initialization completes, tar up the temp dir and upload it to S3.
4870 2 : async fn bootstrap_timeline(
4871 2 : self: &Arc<Self>,
4872 2 : timeline_id: TimelineId,
4873 2 : pg_version: u32,
4874 2 : load_existing_initdb: Option<TimelineId>,
4875 2 : ctx: &RequestContext,
4876 2 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4877 2 : let timeline_create_guard = match self
4878 2 : .start_creating_timeline(
4879 2 : timeline_id,
4880 2 : CreateTimelineIdempotency::Bootstrap { pg_version },
4881 2 : )
4882 2 : .await?
4883 : {
4884 2 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
4885 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
4886 0 : return Ok(CreateTimelineResult::Idempotent(timeline))
4887 : }
4888 : };
4889 :
4890 : // create a `tenant/{tenant_id}/timelines/basebackup-{timeline_id}.{TEMP_FILE_SUFFIX}/`
4891 : // temporary directory for basebackup files for the given timeline.
4892 :
4893 2 : let timelines_path = self.conf.timelines_path(&self.tenant_shard_id);
4894 2 : let pgdata_path = path_with_suffix_extension(
4895 2 : timelines_path.join(format!("basebackup-{timeline_id}")),
4896 2 : TEMP_FILE_SUFFIX,
4897 2 : );
4898 2 :
4899 2 : // Remove whatever was left from the previous runs: safe because TimelineCreateGuard guarantees
4900 2 : // we won't race with other creations or existent timelines with the same path.
4901 2 : if pgdata_path.exists() {
4902 0 : fs::remove_dir_all(&pgdata_path).with_context(|| {
4903 0 : format!("Failed to remove already existing initdb directory: {pgdata_path}")
4904 0 : })?;
4905 2 : }
4906 :
4907 : // this new directory is very temporary, set to remove it immediately after bootstrap, we don't need it
4908 2 : scopeguard::defer! {
4909 2 : if let Err(e) = fs::remove_dir_all(&pgdata_path) {
4910 2 : // this is unlikely, but we will remove the directory on pageserver restart or another bootstrap call
4911 2 : error!("Failed to remove temporary initdb directory '{pgdata_path}': {e}");
4912 2 : }
4913 2 : }
4914 2 : if let Some(existing_initdb_timeline_id) = load_existing_initdb {
4915 2 : if existing_initdb_timeline_id != timeline_id {
4916 0 : let source_path = &remote_initdb_archive_path(
4917 0 : &self.tenant_shard_id.tenant_id,
4918 0 : &existing_initdb_timeline_id,
4919 0 : );
4920 0 : let dest_path =
4921 0 : &remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &timeline_id);
4922 0 :
4923 0 : // if this fails, it will get retried by retried control plane requests
4924 0 : self.remote_storage
4925 0 : .copy_object(source_path, dest_path, &self.cancel)
4926 0 : .await
4927 0 : .context("copy initdb tar")?;
4928 2 : }
4929 2 : let (initdb_tar_zst_path, initdb_tar_zst) =
4930 2 : self::remote_timeline_client::download_initdb_tar_zst(
4931 2 : self.conf,
4932 2 : &self.remote_storage,
4933 2 : &self.tenant_shard_id,
4934 2 : &existing_initdb_timeline_id,
4935 2 : &self.cancel,
4936 2 : )
4937 2 : .await
4938 2 : .context("download initdb tar")?;
4939 :
4940 2 : scopeguard::defer! {
4941 2 : if let Err(e) = fs::remove_file(&initdb_tar_zst_path) {
4942 2 : error!("Failed to remove temporary initdb archive '{initdb_tar_zst_path}': {e}");
4943 2 : }
4944 2 : }
4945 2 :
4946 2 : let buf_read =
4947 2 : BufReader::with_capacity(remote_timeline_client::BUFFER_SIZE, initdb_tar_zst);
4948 2 : extract_zst_tarball(&pgdata_path, buf_read)
4949 2 : .await
4950 2 : .context("extract initdb tar")?;
4951 : } else {
4952 : // Init temporarily repo to get bootstrap data, this creates a directory in the `pgdata_path` path
4953 0 : run_initdb(self.conf, &pgdata_path, pg_version, &self.cancel)
4954 0 : .await
4955 0 : .context("run initdb")?;
4956 :
4957 : // Upload the created data dir to S3
4958 0 : if self.tenant_shard_id().is_shard_zero() {
4959 0 : self.upload_initdb(&timelines_path, &pgdata_path, &timeline_id)
4960 0 : .await?;
4961 0 : }
4962 : }
4963 2 : let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
4964 2 :
4965 2 : // Import the contents of the data directory at the initial checkpoint
4966 2 : // LSN, and any WAL after that.
4967 2 : // Initdb lsn will be equal to last_record_lsn which will be set after import.
4968 2 : // Because we know it upfront avoid having an option or dummy zero value by passing it to the metadata.
4969 2 : let new_metadata = TimelineMetadata::new(
4970 2 : Lsn(0),
4971 2 : None,
4972 2 : None,
4973 2 : Lsn(0),
4974 2 : pgdata_lsn,
4975 2 : pgdata_lsn,
4976 2 : pg_version,
4977 2 : );
4978 2 : let raw_timeline = self
4979 2 : .prepare_new_timeline(
4980 2 : timeline_id,
4981 2 : &new_metadata,
4982 2 : timeline_create_guard,
4983 2 : pgdata_lsn,
4984 2 : None,
4985 2 : )
4986 2 : .await?;
4987 :
4988 2 : let tenant_shard_id = raw_timeline.owning_tenant.tenant_shard_id;
4989 2 : let unfinished_timeline = raw_timeline.raw_timeline()?;
4990 :
4991 : // Flush the new layer files to disk, before we make the timeline as available to
4992 : // the outside world.
4993 : //
4994 : // Flush loop needs to be spawned in order to be able to flush.
4995 2 : unfinished_timeline.maybe_spawn_flush_loop();
4996 2 :
4997 2 : import_datadir::import_timeline_from_postgres_datadir(
4998 2 : unfinished_timeline,
4999 2 : &pgdata_path,
5000 2 : pgdata_lsn,
5001 2 : ctx,
5002 2 : )
5003 2 : .await
5004 2 : .with_context(|| {
5005 0 : format!("Failed to import pgdatadir for timeline {tenant_shard_id}/{timeline_id}")
5006 2 : })?;
5007 :
5008 2 : fail::fail_point!("before-checkpoint-new-timeline", |_| {
5009 0 : Err(CreateTimelineError::Other(anyhow::anyhow!(
5010 0 : "failpoint before-checkpoint-new-timeline"
5011 0 : )))
5012 2 : });
5013 :
5014 2 : unfinished_timeline
5015 2 : .freeze_and_flush()
5016 2 : .await
5017 2 : .with_context(|| {
5018 0 : format!(
5019 0 : "Failed to flush after pgdatadir import for timeline {tenant_shard_id}/{timeline_id}"
5020 0 : )
5021 2 : })?;
5022 :
5023 : // All done!
5024 2 : let timeline = raw_timeline.finish_creation()?;
5025 :
5026 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
5027 :
5028 2 : Ok(CreateTimelineResult::Created(timeline))
5029 2 : }
5030 :
5031 412 : fn build_timeline_remote_client(&self, timeline_id: TimelineId) -> RemoteTimelineClient {
5032 412 : RemoteTimelineClient::new(
5033 412 : self.remote_storage.clone(),
5034 412 : self.deletion_queue_client.clone(),
5035 412 : self.conf,
5036 412 : self.tenant_shard_id,
5037 412 : timeline_id,
5038 412 : self.generation,
5039 412 : &self.tenant_conf.load().location,
5040 412 : )
5041 412 : }
5042 :
5043 : /// Call this before constructing a timeline, to build its required structures
5044 412 : fn build_timeline_resources(&self, timeline_id: TimelineId) -> TimelineResources {
5045 412 : TimelineResources {
5046 412 : remote_client: self.build_timeline_remote_client(timeline_id),
5047 412 : pagestream_throttle: self.pagestream_throttle.clone(),
5048 412 : l0_flush_global_state: self.l0_flush_global_state.clone(),
5049 412 : }
5050 412 : }
5051 :
5052 : /// Creates intermediate timeline structure and its files.
5053 : ///
5054 : /// An empty layer map is initialized, and new data and WAL can be imported starting
5055 : /// at 'disk_consistent_lsn'. After any initial data has been imported, call
5056 : /// `finish_creation` to insert the Timeline into the timelines map.
5057 412 : async fn prepare_new_timeline<'a>(
5058 412 : &'a self,
5059 412 : new_timeline_id: TimelineId,
5060 412 : new_metadata: &TimelineMetadata,
5061 412 : create_guard: TimelineCreateGuard,
5062 412 : start_lsn: Lsn,
5063 412 : ancestor: Option<Arc<Timeline>>,
5064 412 : ) -> anyhow::Result<UninitializedTimeline<'a>> {
5065 412 : let tenant_shard_id = self.tenant_shard_id;
5066 412 :
5067 412 : let resources = self.build_timeline_resources(new_timeline_id);
5068 412 : resources
5069 412 : .remote_client
5070 412 : .init_upload_queue_for_empty_remote(new_metadata)?;
5071 :
5072 412 : let timeline_struct = self
5073 412 : .create_timeline_struct(
5074 412 : new_timeline_id,
5075 412 : new_metadata,
5076 412 : ancestor,
5077 412 : resources,
5078 412 : CreateTimelineCause::Load,
5079 412 : create_guard.idempotency.clone(),
5080 412 : )
5081 412 : .context("Failed to create timeline data structure")?;
5082 :
5083 412 : timeline_struct.init_empty_layer_map(start_lsn);
5084 :
5085 412 : if let Err(e) = self
5086 412 : .create_timeline_files(&create_guard.timeline_path)
5087 412 : .await
5088 : {
5089 0 : error!("Failed to create initial files for timeline {tenant_shard_id}/{new_timeline_id}, cleaning up: {e:?}");
5090 0 : cleanup_timeline_directory(create_guard);
5091 0 : return Err(e);
5092 412 : }
5093 412 :
5094 412 : debug!(
5095 0 : "Successfully created initial files for timeline {tenant_shard_id}/{new_timeline_id}"
5096 : );
5097 :
5098 412 : Ok(UninitializedTimeline::new(
5099 412 : self,
5100 412 : new_timeline_id,
5101 412 : Some((timeline_struct, create_guard)),
5102 412 : ))
5103 412 : }
5104 :
5105 412 : async fn create_timeline_files(&self, timeline_path: &Utf8Path) -> anyhow::Result<()> {
5106 412 : crashsafe::create_dir(timeline_path).context("Failed to create timeline directory")?;
5107 :
5108 412 : fail::fail_point!("after-timeline-dir-creation", |_| {
5109 0 : anyhow::bail!("failpoint after-timeline-dir-creation");
5110 412 : });
5111 :
5112 412 : Ok(())
5113 412 : }
5114 :
5115 : /// Get a guard that provides exclusive access to the timeline directory, preventing
5116 : /// concurrent attempts to create the same timeline.
5117 : ///
5118 : /// The `allow_offloaded` parameter controls whether to tolerate the existence of
5119 : /// offloaded timelines or not.
5120 418 : fn create_timeline_create_guard(
5121 418 : self: &Arc<Self>,
5122 418 : timeline_id: TimelineId,
5123 418 : idempotency: CreateTimelineIdempotency,
5124 418 : allow_offloaded: bool,
5125 418 : ) -> Result<TimelineCreateGuard, TimelineExclusionError> {
5126 418 : let tenant_shard_id = self.tenant_shard_id;
5127 418 :
5128 418 : let timeline_path = self.conf.timeline_path(&tenant_shard_id, &timeline_id);
5129 :
5130 418 : let create_guard = TimelineCreateGuard::new(
5131 418 : self,
5132 418 : timeline_id,
5133 418 : timeline_path.clone(),
5134 418 : idempotency,
5135 418 : allow_offloaded,
5136 418 : )?;
5137 :
5138 : // At this stage, we have got exclusive access to in-memory state for this timeline ID
5139 : // for creation.
5140 : // A timeline directory should never exist on disk already:
5141 : // - a previous failed creation would have cleaned up after itself
5142 : // - a pageserver restart would clean up timeline directories that don't have valid remote state
5143 : //
5144 : // Therefore it is an unexpected internal error to encounter a timeline directory already existing here,
5145 : // this error may indicate a bug in cleanup on failed creations.
5146 416 : if timeline_path.exists() {
5147 0 : return Err(TimelineExclusionError::Other(anyhow::anyhow!(
5148 0 : "Timeline directory already exists! This is a bug."
5149 0 : )));
5150 416 : }
5151 416 :
5152 416 : Ok(create_guard)
5153 418 : }
5154 :
5155 : /// Gathers inputs from all of the timelines to produce a sizing model input.
5156 : ///
5157 : /// Future is cancellation safe. Only one calculation can be running at once per tenant.
5158 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5159 : pub async fn gather_size_inputs(
5160 : &self,
5161 : // `max_retention_period` overrides the cutoff that is used to calculate the size
5162 : // (only if it is shorter than the real cutoff).
5163 : max_retention_period: Option<u64>,
5164 : cause: LogicalSizeCalculationCause,
5165 : cancel: &CancellationToken,
5166 : ctx: &RequestContext,
5167 : ) -> Result<size::ModelInputs, size::CalculateSyntheticSizeError> {
5168 : let logical_sizes_at_once = self
5169 : .conf
5170 : .concurrent_tenant_size_logical_size_queries
5171 : .inner();
5172 :
5173 : // TODO: Having a single mutex block concurrent reads is not great for performance.
5174 : //
5175 : // But the only case where we need to run multiple of these at once is when we
5176 : // request a size for a tenant manually via API, while another background calculation
5177 : // is in progress (which is not a common case).
5178 : //
5179 : // See more for on the issue #2748 condenced out of the initial PR review.
5180 : let mut shared_cache = tokio::select! {
5181 : locked = self.cached_logical_sizes.lock() => locked,
5182 : _ = cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5183 : _ = self.cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5184 : };
5185 :
5186 : size::gather_inputs(
5187 : self,
5188 : logical_sizes_at_once,
5189 : max_retention_period,
5190 : &mut shared_cache,
5191 : cause,
5192 : cancel,
5193 : ctx,
5194 : )
5195 : .await
5196 : }
5197 :
5198 : /// Calculate synthetic tenant size and cache the result.
5199 : /// This is periodically called by background worker.
5200 : /// result is cached in tenant struct
5201 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5202 : pub async fn calculate_synthetic_size(
5203 : &self,
5204 : cause: LogicalSizeCalculationCause,
5205 : cancel: &CancellationToken,
5206 : ctx: &RequestContext,
5207 : ) -> Result<u64, size::CalculateSyntheticSizeError> {
5208 : let inputs = self.gather_size_inputs(None, cause, cancel, ctx).await?;
5209 :
5210 : let size = inputs.calculate();
5211 :
5212 : self.set_cached_synthetic_size(size);
5213 :
5214 : Ok(size)
5215 : }
5216 :
5217 : /// Cache given synthetic size and update the metric value
5218 0 : pub fn set_cached_synthetic_size(&self, size: u64) {
5219 0 : self.cached_synthetic_tenant_size
5220 0 : .store(size, Ordering::Relaxed);
5221 0 :
5222 0 : // Only shard zero should be calculating synthetic sizes
5223 0 : debug_assert!(self.shard_identity.is_shard_zero());
5224 :
5225 0 : TENANT_SYNTHETIC_SIZE_METRIC
5226 0 : .get_metric_with_label_values(&[&self.tenant_shard_id.tenant_id.to_string()])
5227 0 : .unwrap()
5228 0 : .set(size);
5229 0 : }
5230 :
5231 0 : pub fn cached_synthetic_size(&self) -> u64 {
5232 0 : self.cached_synthetic_tenant_size.load(Ordering::Relaxed)
5233 0 : }
5234 :
5235 : /// Flush any in-progress layers, schedule uploads, and wait for uploads to complete.
5236 : ///
5237 : /// This function can take a long time: callers should wrap it in a timeout if calling
5238 : /// from an external API handler.
5239 : ///
5240 : /// Cancel-safety: cancelling this function may leave I/O running, but such I/O is
5241 : /// still bounded by tenant/timeline shutdown.
5242 0 : #[tracing::instrument(skip_all)]
5243 : pub(crate) async fn flush_remote(&self) -> anyhow::Result<()> {
5244 : let timelines = self.timelines.lock().unwrap().clone();
5245 :
5246 0 : async fn flush_timeline(_gate: GateGuard, timeline: Arc<Timeline>) -> anyhow::Result<()> {
5247 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Flushing...");
5248 0 : timeline.freeze_and_flush().await?;
5249 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Waiting for uploads...");
5250 0 : timeline.remote_client.wait_completion().await?;
5251 :
5252 0 : Ok(())
5253 0 : }
5254 :
5255 : // We do not use a JoinSet for these tasks, because we don't want them to be
5256 : // aborted when this function's future is cancelled: they should stay alive
5257 : // holding their GateGuard until they complete, to ensure their I/Os complete
5258 : // before Timeline shutdown completes.
5259 : let mut results = FuturesUnordered::new();
5260 :
5261 : for (_timeline_id, timeline) in timelines {
5262 : // Run each timeline's flush in a task holding the timeline's gate: this
5263 : // means that if this function's future is cancelled, the Timeline shutdown
5264 : // will still wait for any I/O in here to complete.
5265 : let Ok(gate) = timeline.gate.enter() else {
5266 : continue;
5267 : };
5268 0 : let jh = tokio::task::spawn(async move { flush_timeline(gate, timeline).await });
5269 : results.push(jh);
5270 : }
5271 :
5272 : while let Some(r) = results.next().await {
5273 : if let Err(e) = r {
5274 : if !e.is_cancelled() && !e.is_panic() {
5275 : tracing::error!("unexpected join error: {e:?}");
5276 : }
5277 : }
5278 : }
5279 :
5280 : // The flushes we did above were just writes, but the Tenant might have had
5281 : // pending deletions as well from recent compaction/gc: we want to flush those
5282 : // as well. This requires flushing the global delete queue. This is cheap
5283 : // because it's typically a no-op.
5284 : match self.deletion_queue_client.flush_execute().await {
5285 : Ok(_) => {}
5286 : Err(DeletionQueueError::ShuttingDown) => {}
5287 : }
5288 :
5289 : Ok(())
5290 : }
5291 :
5292 0 : pub(crate) fn get_tenant_conf(&self) -> TenantConfOpt {
5293 0 : self.tenant_conf.load().tenant_conf.clone()
5294 0 : }
5295 :
5296 : /// How much local storage would this tenant like to have? It can cope with
5297 : /// less than this (via eviction and on-demand downloads), but this function enables
5298 : /// the Tenant to advertise how much storage it would prefer to have to provide fast I/O
5299 : /// by keeping important things on local disk.
5300 : ///
5301 : /// This is a heuristic, not a guarantee: tenants that are long-idle will actually use less
5302 : /// than they report here, due to layer eviction. Tenants with many active branches may
5303 : /// actually use more than they report here.
5304 0 : pub(crate) fn local_storage_wanted(&self) -> u64 {
5305 0 : let timelines = self.timelines.lock().unwrap();
5306 0 :
5307 0 : // Heuristic: we use the max() of the timelines' visible sizes, rather than the sum. This
5308 0 : // reflects the observation that on tenants with multiple large branches, typically only one
5309 0 : // of them is used actively enough to occupy space on disk.
5310 0 : timelines
5311 0 : .values()
5312 0 : .map(|t| t.metrics.visible_physical_size_gauge.get())
5313 0 : .max()
5314 0 : .unwrap_or(0)
5315 0 : }
5316 :
5317 : /// Serialize and write the latest TenantManifest to remote storage.
5318 2 : pub(crate) async fn store_tenant_manifest(&self) -> Result<(), TenantManifestError> {
5319 : // Only one manifest write may be done at at time, and the contents of the manifest
5320 : // must be loaded while holding this lock. This makes it safe to call this function
5321 : // from anywhere without worrying about colliding updates.
5322 2 : let mut guard = tokio::select! {
5323 2 : g = self.tenant_manifest_upload.lock() => {
5324 2 : g
5325 : },
5326 2 : _ = self.cancel.cancelled() => {
5327 0 : return Err(TenantManifestError::Cancelled);
5328 : }
5329 : };
5330 :
5331 2 : let manifest = self.build_tenant_manifest();
5332 2 : if Some(&manifest) == (*guard).as_ref() {
5333 : // Optimisation: skip uploads that don't change anything.
5334 0 : return Ok(());
5335 2 : }
5336 2 :
5337 2 : upload_tenant_manifest(
5338 2 : &self.remote_storage,
5339 2 : &self.tenant_shard_id,
5340 2 : self.generation,
5341 2 : &manifest,
5342 2 : &self.cancel,
5343 2 : )
5344 2 : .await
5345 2 : .map_err(|e| {
5346 0 : if self.cancel.is_cancelled() {
5347 0 : TenantManifestError::Cancelled
5348 : } else {
5349 0 : TenantManifestError::RemoteStorage(e)
5350 : }
5351 2 : })?;
5352 :
5353 : // Store the successfully uploaded manifest, so that future callers can avoid
5354 : // re-uploading the same thing.
5355 2 : *guard = Some(manifest);
5356 2 :
5357 2 : Ok(())
5358 2 : }
5359 : }
5360 :
5361 : /// Create the cluster temporarily in 'initdbpath' directory inside the repository
5362 : /// to get bootstrap data for timeline initialization.
5363 0 : async fn run_initdb(
5364 0 : conf: &'static PageServerConf,
5365 0 : initdb_target_dir: &Utf8Path,
5366 0 : pg_version: u32,
5367 0 : cancel: &CancellationToken,
5368 0 : ) -> Result<(), InitdbError> {
5369 0 : let initdb_bin_path = conf
5370 0 : .pg_bin_dir(pg_version)
5371 0 : .map_err(InitdbError::Other)?
5372 0 : .join("initdb");
5373 0 : let initdb_lib_dir = conf.pg_lib_dir(pg_version).map_err(InitdbError::Other)?;
5374 0 : info!(
5375 0 : "running {} in {}, libdir: {}",
5376 : initdb_bin_path, initdb_target_dir, initdb_lib_dir,
5377 : );
5378 :
5379 0 : let _permit = INIT_DB_SEMAPHORE.acquire().await;
5380 :
5381 0 : let res = postgres_initdb::do_run_initdb(postgres_initdb::RunInitdbArgs {
5382 0 : superuser: &conf.superuser,
5383 0 : locale: &conf.locale,
5384 0 : initdb_bin: &initdb_bin_path,
5385 0 : pg_version,
5386 0 : library_search_path: &initdb_lib_dir,
5387 0 : pgdata: initdb_target_dir,
5388 0 : })
5389 0 : .await
5390 0 : .map_err(InitdbError::Inner);
5391 0 :
5392 0 : // This isn't true cancellation support, see above. Still return an error to
5393 0 : // excercise the cancellation code path.
5394 0 : if cancel.is_cancelled() {
5395 0 : return Err(InitdbError::Cancelled);
5396 0 : }
5397 0 :
5398 0 : res
5399 0 : }
5400 :
5401 : /// Dump contents of a layer file to stdout.
5402 0 : pub async fn dump_layerfile_from_path(
5403 0 : path: &Utf8Path,
5404 0 : verbose: bool,
5405 0 : ctx: &RequestContext,
5406 0 : ) -> anyhow::Result<()> {
5407 : use std::os::unix::fs::FileExt;
5408 :
5409 : // All layer files start with a two-byte "magic" value, to identify the kind of
5410 : // file.
5411 0 : let file = File::open(path)?;
5412 0 : let mut header_buf = [0u8; 2];
5413 0 : file.read_exact_at(&mut header_buf, 0)?;
5414 :
5415 0 : match u16::from_be_bytes(header_buf) {
5416 : crate::IMAGE_FILE_MAGIC => {
5417 0 : ImageLayer::new_for_path(path, file)?
5418 0 : .dump(verbose, ctx)
5419 0 : .await?
5420 : }
5421 : crate::DELTA_FILE_MAGIC => {
5422 0 : DeltaLayer::new_for_path(path, file)?
5423 0 : .dump(verbose, ctx)
5424 0 : .await?
5425 : }
5426 0 : magic => bail!("unrecognized magic identifier: {:?}", magic),
5427 : }
5428 :
5429 0 : Ok(())
5430 0 : }
5431 :
5432 : #[cfg(test)]
5433 : pub(crate) mod harness {
5434 : use bytes::{Bytes, BytesMut};
5435 : use once_cell::sync::OnceCell;
5436 : use pageserver_api::models::ShardParameters;
5437 : use pageserver_api::shard::ShardIndex;
5438 : use utils::logging;
5439 :
5440 : use crate::deletion_queue::mock::MockDeletionQueue;
5441 : use crate::l0_flush::L0FlushConfig;
5442 : use crate::walredo::apply_neon;
5443 : use pageserver_api::key::Key;
5444 : use pageserver_api::record::NeonWalRecord;
5445 :
5446 : use super::*;
5447 : use hex_literal::hex;
5448 : use utils::id::TenantId;
5449 :
5450 : pub const TIMELINE_ID: TimelineId =
5451 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
5452 : pub const NEW_TIMELINE_ID: TimelineId =
5453 : TimelineId::from_array(hex!("AA223344556677881122334455667788"));
5454 :
5455 : /// Convenience function to create a page image with given string as the only content
5456 5028714 : pub fn test_img(s: &str) -> Bytes {
5457 5028714 : let mut buf = BytesMut::new();
5458 5028714 : buf.extend_from_slice(s.as_bytes());
5459 5028714 : buf.resize(64, 0);
5460 5028714 :
5461 5028714 : buf.freeze()
5462 5028714 : }
5463 :
5464 : impl From<TenantConf> for TenantConfOpt {
5465 192 : fn from(tenant_conf: TenantConf) -> Self {
5466 192 : Self {
5467 192 : checkpoint_distance: Some(tenant_conf.checkpoint_distance),
5468 192 : checkpoint_timeout: Some(tenant_conf.checkpoint_timeout),
5469 192 : compaction_target_size: Some(tenant_conf.compaction_target_size),
5470 192 : compaction_period: Some(tenant_conf.compaction_period),
5471 192 : compaction_threshold: Some(tenant_conf.compaction_threshold),
5472 192 : compaction_algorithm: Some(tenant_conf.compaction_algorithm),
5473 192 : gc_horizon: Some(tenant_conf.gc_horizon),
5474 192 : gc_period: Some(tenant_conf.gc_period),
5475 192 : image_creation_threshold: Some(tenant_conf.image_creation_threshold),
5476 192 : pitr_interval: Some(tenant_conf.pitr_interval),
5477 192 : walreceiver_connect_timeout: Some(tenant_conf.walreceiver_connect_timeout),
5478 192 : lagging_wal_timeout: Some(tenant_conf.lagging_wal_timeout),
5479 192 : max_lsn_wal_lag: Some(tenant_conf.max_lsn_wal_lag),
5480 192 : eviction_policy: Some(tenant_conf.eviction_policy),
5481 192 : min_resident_size_override: tenant_conf.min_resident_size_override,
5482 192 : evictions_low_residence_duration_metric_threshold: Some(
5483 192 : tenant_conf.evictions_low_residence_duration_metric_threshold,
5484 192 : ),
5485 192 : heatmap_period: Some(tenant_conf.heatmap_period),
5486 192 : lazy_slru_download: Some(tenant_conf.lazy_slru_download),
5487 192 : timeline_get_throttle: Some(tenant_conf.timeline_get_throttle),
5488 192 : image_layer_creation_check_threshold: Some(
5489 192 : tenant_conf.image_layer_creation_check_threshold,
5490 192 : ),
5491 192 : lsn_lease_length: Some(tenant_conf.lsn_lease_length),
5492 192 : lsn_lease_length_for_ts: Some(tenant_conf.lsn_lease_length_for_ts),
5493 192 : timeline_offloading: Some(tenant_conf.timeline_offloading),
5494 192 : wal_receiver_protocol_override: tenant_conf.wal_receiver_protocol_override,
5495 192 : }
5496 192 : }
5497 : }
5498 :
5499 : pub struct TenantHarness {
5500 : pub conf: &'static PageServerConf,
5501 : pub tenant_conf: TenantConf,
5502 : pub tenant_shard_id: TenantShardId,
5503 : pub generation: Generation,
5504 : pub shard: ShardIndex,
5505 : pub remote_storage: GenericRemoteStorage,
5506 : pub remote_fs_dir: Utf8PathBuf,
5507 : pub deletion_queue: MockDeletionQueue,
5508 : }
5509 :
5510 : static LOG_HANDLE: OnceCell<()> = OnceCell::new();
5511 :
5512 208 : pub(crate) fn setup_logging() {
5513 208 : LOG_HANDLE.get_or_init(|| {
5514 196 : logging::init(
5515 196 : logging::LogFormat::Test,
5516 196 : // enable it in case the tests exercise code paths that use
5517 196 : // debug_assert_current_span_has_tenant_and_timeline_id
5518 196 : logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
5519 196 : logging::Output::Stdout,
5520 196 : )
5521 196 : .expect("Failed to init test logging")
5522 208 : });
5523 208 : }
5524 :
5525 : impl TenantHarness {
5526 192 : pub async fn create_custom(
5527 192 : test_name: &'static str,
5528 192 : tenant_conf: TenantConf,
5529 192 : tenant_id: TenantId,
5530 192 : shard_identity: ShardIdentity,
5531 192 : generation: Generation,
5532 192 : ) -> anyhow::Result<Self> {
5533 192 : setup_logging();
5534 192 :
5535 192 : let repo_dir = PageServerConf::test_repo_dir(test_name);
5536 192 : let _ = fs::remove_dir_all(&repo_dir);
5537 192 : fs::create_dir_all(&repo_dir)?;
5538 :
5539 192 : let conf = PageServerConf::dummy_conf(repo_dir);
5540 192 : // Make a static copy of the config. This can never be free'd, but that's
5541 192 : // OK in a test.
5542 192 : let conf: &'static PageServerConf = Box::leak(Box::new(conf));
5543 192 :
5544 192 : let shard = shard_identity.shard_index();
5545 192 : let tenant_shard_id = TenantShardId {
5546 192 : tenant_id,
5547 192 : shard_number: shard.shard_number,
5548 192 : shard_count: shard.shard_count,
5549 192 : };
5550 192 : fs::create_dir_all(conf.tenant_path(&tenant_shard_id))?;
5551 192 : fs::create_dir_all(conf.timelines_path(&tenant_shard_id))?;
5552 :
5553 : use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
5554 192 : let remote_fs_dir = conf.workdir.join("localfs");
5555 192 : std::fs::create_dir_all(&remote_fs_dir).unwrap();
5556 192 : let config = RemoteStorageConfig {
5557 192 : storage: RemoteStorageKind::LocalFs {
5558 192 : local_path: remote_fs_dir.clone(),
5559 192 : },
5560 192 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
5561 192 : small_timeout: RemoteStorageConfig::DEFAULT_SMALL_TIMEOUT,
5562 192 : };
5563 192 : let remote_storage = GenericRemoteStorage::from_config(&config).await.unwrap();
5564 192 : let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()));
5565 192 :
5566 192 : Ok(Self {
5567 192 : conf,
5568 192 : tenant_conf,
5569 192 : tenant_shard_id,
5570 192 : generation,
5571 192 : shard,
5572 192 : remote_storage,
5573 192 : remote_fs_dir,
5574 192 : deletion_queue,
5575 192 : })
5576 192 : }
5577 :
5578 180 : pub async fn create(test_name: &'static str) -> anyhow::Result<Self> {
5579 180 : // Disable automatic GC and compaction to make the unit tests more deterministic.
5580 180 : // The tests perform them manually if needed.
5581 180 : let tenant_conf = TenantConf {
5582 180 : gc_period: Duration::ZERO,
5583 180 : compaction_period: Duration::ZERO,
5584 180 : ..TenantConf::default()
5585 180 : };
5586 180 : let tenant_id = TenantId::generate();
5587 180 : let shard = ShardIdentity::unsharded();
5588 180 : Self::create_custom(
5589 180 : test_name,
5590 180 : tenant_conf,
5591 180 : tenant_id,
5592 180 : shard,
5593 180 : Generation::new(0xdeadbeef),
5594 180 : )
5595 180 : .await
5596 180 : }
5597 :
5598 20 : pub fn span(&self) -> tracing::Span {
5599 20 : info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
5600 20 : }
5601 :
5602 192 : pub(crate) async fn load(&self) -> (Arc<Tenant>, RequestContext) {
5603 192 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
5604 192 : (
5605 192 : self.do_try_load(&ctx)
5606 192 : .await
5607 192 : .expect("failed to load test tenant"),
5608 192 : ctx,
5609 192 : )
5610 192 : }
5611 :
5612 192 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5613 : pub(crate) async fn do_try_load(
5614 : &self,
5615 : ctx: &RequestContext,
5616 : ) -> anyhow::Result<Arc<Tenant>> {
5617 : let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
5618 :
5619 : let tenant = Arc::new(Tenant::new(
5620 : TenantState::Attaching,
5621 : self.conf,
5622 : AttachedTenantConf::try_from(LocationConf::attached_single(
5623 : TenantConfOpt::from(self.tenant_conf.clone()),
5624 : self.generation,
5625 : &ShardParameters::default(),
5626 : ))
5627 : .unwrap(),
5628 : // This is a legacy/test code path: sharding isn't supported here.
5629 : ShardIdentity::unsharded(),
5630 : Some(walredo_mgr),
5631 : self.tenant_shard_id,
5632 : self.remote_storage.clone(),
5633 : self.deletion_queue.new_client(),
5634 : // TODO: ideally we should run all unit tests with both configs
5635 : L0FlushGlobalState::new(L0FlushConfig::default()),
5636 : ));
5637 :
5638 : let preload = tenant
5639 : .preload(&self.remote_storage, CancellationToken::new())
5640 : .await?;
5641 : tenant.attach(Some(preload), ctx).await?;
5642 :
5643 : tenant.state.send_replace(TenantState::Active);
5644 : for timeline in tenant.timelines.lock().unwrap().values() {
5645 : timeline.set_state(TimelineState::Active);
5646 : }
5647 : Ok(tenant)
5648 : }
5649 :
5650 2 : pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
5651 2 : self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
5652 2 : }
5653 : }
5654 :
5655 : // Mock WAL redo manager that doesn't do much
5656 : pub(crate) struct TestRedoManager;
5657 :
5658 : impl TestRedoManager {
5659 : /// # Cancel-Safety
5660 : ///
5661 : /// This method is cancellation-safe.
5662 410 : pub async fn request_redo(
5663 410 : &self,
5664 410 : key: Key,
5665 410 : lsn: Lsn,
5666 410 : base_img: Option<(Lsn, Bytes)>,
5667 410 : records: Vec<(Lsn, NeonWalRecord)>,
5668 410 : _pg_version: u32,
5669 410 : ) -> Result<Bytes, walredo::Error> {
5670 570 : let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
5671 410 : if records_neon {
5672 : // For Neon wal records, we can decode without spawning postgres, so do so.
5673 410 : let mut page = match (base_img, records.first()) {
5674 344 : (Some((_lsn, img)), _) => {
5675 344 : let mut page = BytesMut::new();
5676 344 : page.extend_from_slice(&img);
5677 344 : page
5678 : }
5679 66 : (_, Some((_lsn, rec))) if rec.will_init() => BytesMut::new(),
5680 : _ => {
5681 0 : panic!("Neon WAL redo requires base image or will init record");
5682 : }
5683 : };
5684 :
5685 980 : for (record_lsn, record) in records {
5686 570 : apply_neon::apply_in_neon(&record, record_lsn, key, &mut page)?;
5687 : }
5688 410 : Ok(page.freeze())
5689 : } else {
5690 : // We never spawn a postgres walredo process in unit tests: just log what we might have done.
5691 0 : let s = format!(
5692 0 : "redo for {} to get to {}, with {} and {} records",
5693 0 : key,
5694 0 : lsn,
5695 0 : if base_img.is_some() {
5696 0 : "base image"
5697 : } else {
5698 0 : "no base image"
5699 : },
5700 0 : records.len()
5701 0 : );
5702 0 : println!("{s}");
5703 0 :
5704 0 : Ok(test_img(&s))
5705 : }
5706 410 : }
5707 : }
5708 : }
5709 :
5710 : #[cfg(test)]
5711 : mod tests {
5712 : use std::collections::{BTreeMap, BTreeSet};
5713 :
5714 : use super::*;
5715 : use crate::keyspace::KeySpaceAccum;
5716 : use crate::tenant::harness::*;
5717 : use crate::tenant::timeline::CompactFlags;
5718 : use crate::DEFAULT_PG_VERSION;
5719 : use bytes::{Bytes, BytesMut};
5720 : use hex_literal::hex;
5721 : use itertools::Itertools;
5722 : use pageserver_api::key::{Key, AUX_KEY_PREFIX, NON_INHERITED_RANGE};
5723 : use pageserver_api::keyspace::KeySpace;
5724 : use pageserver_api::models::{CompactionAlgorithm, CompactionAlgorithmSettings};
5725 : use pageserver_api::value::Value;
5726 : use pageserver_compaction::helpers::overlaps_with;
5727 : use rand::{thread_rng, Rng};
5728 : use storage_layer::PersistentLayerKey;
5729 : use tests::storage_layer::ValuesReconstructState;
5730 : use tests::timeline::{GetVectoredError, ShutdownMode};
5731 : use timeline::{CompactOptions, DeltaLayerTestDesc};
5732 : use utils::id::TenantId;
5733 :
5734 : #[cfg(feature = "testing")]
5735 : use pageserver_api::record::NeonWalRecord;
5736 : #[cfg(feature = "testing")]
5737 : use timeline::compaction::{KeyHistoryRetention, KeyLogAtLsn};
5738 : #[cfg(feature = "testing")]
5739 : use timeline::GcInfo;
5740 :
5741 : static TEST_KEY: Lazy<Key> =
5742 18 : Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));
5743 :
5744 : #[tokio::test]
5745 2 : async fn test_basic() -> anyhow::Result<()> {
5746 2 : let (tenant, ctx) = TenantHarness::create("test_basic").await?.load().await;
5747 2 : let tline = tenant
5748 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5749 2 : .await?;
5750 2 :
5751 2 : let mut writer = tline.writer().await;
5752 2 : writer
5753 2 : .put(
5754 2 : *TEST_KEY,
5755 2 : Lsn(0x10),
5756 2 : &Value::Image(test_img("foo at 0x10")),
5757 2 : &ctx,
5758 2 : )
5759 2 : .await?;
5760 2 : writer.finish_write(Lsn(0x10));
5761 2 : drop(writer);
5762 2 :
5763 2 : let mut writer = tline.writer().await;
5764 2 : writer
5765 2 : .put(
5766 2 : *TEST_KEY,
5767 2 : Lsn(0x20),
5768 2 : &Value::Image(test_img("foo at 0x20")),
5769 2 : &ctx,
5770 2 : )
5771 2 : .await?;
5772 2 : writer.finish_write(Lsn(0x20));
5773 2 : drop(writer);
5774 2 :
5775 2 : assert_eq!(
5776 2 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
5777 2 : test_img("foo at 0x10")
5778 2 : );
5779 2 : assert_eq!(
5780 2 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
5781 2 : test_img("foo at 0x10")
5782 2 : );
5783 2 : assert_eq!(
5784 2 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
5785 2 : test_img("foo at 0x20")
5786 2 : );
5787 2 :
5788 2 : Ok(())
5789 2 : }
5790 :
5791 : #[tokio::test]
5792 2 : async fn no_duplicate_timelines() -> anyhow::Result<()> {
5793 2 : let (tenant, ctx) = TenantHarness::create("no_duplicate_timelines")
5794 2 : .await?
5795 2 : .load()
5796 2 : .await;
5797 2 : let _ = tenant
5798 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5799 2 : .await?;
5800 2 :
5801 2 : match tenant
5802 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5803 2 : .await
5804 2 : {
5805 2 : Ok(_) => panic!("duplicate timeline creation should fail"),
5806 2 : Err(e) => assert_eq!(
5807 2 : e.to_string(),
5808 2 : "timeline already exists with different parameters".to_string()
5809 2 : ),
5810 2 : }
5811 2 :
5812 2 : Ok(())
5813 2 : }
5814 :
5815 : /// Convenience function to create a page image with given string as the only content
5816 10 : pub fn test_value(s: &str) -> Value {
5817 10 : let mut buf = BytesMut::new();
5818 10 : buf.extend_from_slice(s.as_bytes());
5819 10 : Value::Image(buf.freeze())
5820 10 : }
5821 :
5822 : ///
5823 : /// Test branch creation
5824 : ///
5825 : #[tokio::test]
5826 2 : async fn test_branch() -> anyhow::Result<()> {
5827 2 : use std::str::from_utf8;
5828 2 :
5829 2 : let (tenant, ctx) = TenantHarness::create("test_branch").await?.load().await;
5830 2 : let tline = tenant
5831 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5832 2 : .await?;
5833 2 : let mut writer = tline.writer().await;
5834 2 :
5835 2 : #[allow(non_snake_case)]
5836 2 : let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
5837 2 : #[allow(non_snake_case)]
5838 2 : let TEST_KEY_B: Key = Key::from_hex("110000000033333333444444445500000002").unwrap();
5839 2 :
5840 2 : // Insert a value on the timeline
5841 2 : writer
5842 2 : .put(TEST_KEY_A, Lsn(0x20), &test_value("foo at 0x20"), &ctx)
5843 2 : .await?;
5844 2 : writer
5845 2 : .put(TEST_KEY_B, Lsn(0x20), &test_value("foobar at 0x20"), &ctx)
5846 2 : .await?;
5847 2 : writer.finish_write(Lsn(0x20));
5848 2 :
5849 2 : writer
5850 2 : .put(TEST_KEY_A, Lsn(0x30), &test_value("foo at 0x30"), &ctx)
5851 2 : .await?;
5852 2 : writer.finish_write(Lsn(0x30));
5853 2 : writer
5854 2 : .put(TEST_KEY_A, Lsn(0x40), &test_value("foo at 0x40"), &ctx)
5855 2 : .await?;
5856 2 : writer.finish_write(Lsn(0x40));
5857 2 :
5858 2 : //assert_current_logical_size(&tline, Lsn(0x40));
5859 2 :
5860 2 : // Branch the history, modify relation differently on the new timeline
5861 2 : tenant
5862 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x30)), &ctx)
5863 2 : .await?;
5864 2 : let newtline = tenant
5865 2 : .get_timeline(NEW_TIMELINE_ID, true)
5866 2 : .expect("Should have a local timeline");
5867 2 : let mut new_writer = newtline.writer().await;
5868 2 : new_writer
5869 2 : .put(TEST_KEY_A, Lsn(0x40), &test_value("bar at 0x40"), &ctx)
5870 2 : .await?;
5871 2 : new_writer.finish_write(Lsn(0x40));
5872 2 :
5873 2 : // Check page contents on both branches
5874 2 : assert_eq!(
5875 2 : from_utf8(&tline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
5876 2 : "foo at 0x40"
5877 2 : );
5878 2 : assert_eq!(
5879 2 : from_utf8(&newtline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
5880 2 : "bar at 0x40"
5881 2 : );
5882 2 : assert_eq!(
5883 2 : from_utf8(&newtline.get(TEST_KEY_B, Lsn(0x40), &ctx).await?)?,
5884 2 : "foobar at 0x20"
5885 2 : );
5886 2 :
5887 2 : //assert_current_logical_size(&tline, Lsn(0x40));
5888 2 :
5889 2 : Ok(())
5890 2 : }
5891 :
5892 20 : async fn make_some_layers(
5893 20 : tline: &Timeline,
5894 20 : start_lsn: Lsn,
5895 20 : ctx: &RequestContext,
5896 20 : ) -> anyhow::Result<()> {
5897 20 : let mut lsn = start_lsn;
5898 : {
5899 20 : let mut writer = tline.writer().await;
5900 : // Create a relation on the timeline
5901 20 : writer
5902 20 : .put(
5903 20 : *TEST_KEY,
5904 20 : lsn,
5905 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5906 20 : ctx,
5907 20 : )
5908 20 : .await?;
5909 20 : writer.finish_write(lsn);
5910 20 : lsn += 0x10;
5911 20 : writer
5912 20 : .put(
5913 20 : *TEST_KEY,
5914 20 : lsn,
5915 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5916 20 : ctx,
5917 20 : )
5918 20 : .await?;
5919 20 : writer.finish_write(lsn);
5920 20 : lsn += 0x10;
5921 20 : }
5922 20 : tline.freeze_and_flush().await?;
5923 : {
5924 20 : let mut writer = tline.writer().await;
5925 20 : writer
5926 20 : .put(
5927 20 : *TEST_KEY,
5928 20 : lsn,
5929 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5930 20 : ctx,
5931 20 : )
5932 20 : .await?;
5933 20 : writer.finish_write(lsn);
5934 20 : lsn += 0x10;
5935 20 : writer
5936 20 : .put(
5937 20 : *TEST_KEY,
5938 20 : lsn,
5939 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5940 20 : ctx,
5941 20 : )
5942 20 : .await?;
5943 20 : writer.finish_write(lsn);
5944 20 : }
5945 20 : tline.freeze_and_flush().await.map_err(|e| e.into())
5946 20 : }
5947 :
5948 : #[tokio::test(start_paused = true)]
5949 2 : async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
5950 2 : let (tenant, ctx) =
5951 2 : TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
5952 2 : .await?
5953 2 : .load()
5954 2 : .await;
5955 2 : // Advance to the lsn lease deadline so that GC is not blocked by
5956 2 : // initial transition into AttachedSingle.
5957 2 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
5958 2 : tokio::time::resume();
5959 2 : let tline = tenant
5960 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5961 2 : .await?;
5962 2 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5963 2 :
5964 2 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
5965 2 : // FIXME: this doesn't actually remove any layer currently, given how the flushing
5966 2 : // and compaction works. But it does set the 'cutoff' point so that the cross check
5967 2 : // below should fail.
5968 2 : tenant
5969 2 : .gc_iteration(
5970 2 : Some(TIMELINE_ID),
5971 2 : 0x10,
5972 2 : Duration::ZERO,
5973 2 : &CancellationToken::new(),
5974 2 : &ctx,
5975 2 : )
5976 2 : .await?;
5977 2 :
5978 2 : // try to branch at lsn 25, should fail because we already garbage collected the data
5979 2 : match tenant
5980 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
5981 2 : .await
5982 2 : {
5983 2 : Ok(_) => panic!("branching should have failed"),
5984 2 : Err(err) => {
5985 2 : let CreateTimelineError::AncestorLsn(err) = err else {
5986 2 : panic!("wrong error type")
5987 2 : };
5988 2 : assert!(err.to_string().contains("invalid branch start lsn"));
5989 2 : assert!(err
5990 2 : .source()
5991 2 : .unwrap()
5992 2 : .to_string()
5993 2 : .contains("we might've already garbage collected needed data"))
5994 2 : }
5995 2 : }
5996 2 :
5997 2 : Ok(())
5998 2 : }
5999 :
6000 : #[tokio::test]
6001 2 : async fn test_prohibit_branch_creation_on_pre_initdb_lsn() -> anyhow::Result<()> {
6002 2 : let (tenant, ctx) =
6003 2 : TenantHarness::create("test_prohibit_branch_creation_on_pre_initdb_lsn")
6004 2 : .await?
6005 2 : .load()
6006 2 : .await;
6007 2 :
6008 2 : let tline = tenant
6009 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x50), DEFAULT_PG_VERSION, &ctx)
6010 2 : .await?;
6011 2 : // try to branch at lsn 0x25, should fail because initdb lsn is 0x50
6012 2 : match tenant
6013 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
6014 2 : .await
6015 2 : {
6016 2 : Ok(_) => panic!("branching should have failed"),
6017 2 : Err(err) => {
6018 2 : let CreateTimelineError::AncestorLsn(err) = err else {
6019 2 : panic!("wrong error type");
6020 2 : };
6021 2 : assert!(&err.to_string().contains("invalid branch start lsn"));
6022 2 : assert!(&err
6023 2 : .source()
6024 2 : .unwrap()
6025 2 : .to_string()
6026 2 : .contains("is earlier than latest GC cutoff"));
6027 2 : }
6028 2 : }
6029 2 :
6030 2 : Ok(())
6031 2 : }
6032 :
6033 : /*
6034 : // FIXME: This currently fails to error out. Calling GC doesn't currently
6035 : // remove the old value, we'd need to work a little harder
6036 : #[tokio::test]
6037 : async fn test_prohibit_get_for_garbage_collected_data() -> anyhow::Result<()> {
6038 : let repo =
6039 : RepoHarness::create("test_prohibit_get_for_garbage_collected_data")?
6040 : .load();
6041 :
6042 : let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION)?;
6043 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6044 :
6045 : repo.gc_iteration(Some(TIMELINE_ID), 0x10, Duration::ZERO)?;
6046 : let latest_gc_cutoff_lsn = tline.get_latest_gc_cutoff_lsn();
6047 : assert!(*latest_gc_cutoff_lsn > Lsn(0x25));
6048 : match tline.get(*TEST_KEY, Lsn(0x25)) {
6049 : Ok(_) => panic!("request for page should have failed"),
6050 : Err(err) => assert!(err.to_string().contains("not found at")),
6051 : }
6052 : Ok(())
6053 : }
6054 : */
6055 :
6056 : #[tokio::test]
6057 2 : async fn test_get_branchpoints_from_an_inactive_timeline() -> anyhow::Result<()> {
6058 2 : let (tenant, ctx) =
6059 2 : TenantHarness::create("test_get_branchpoints_from_an_inactive_timeline")
6060 2 : .await?
6061 2 : .load()
6062 2 : .await;
6063 2 : let tline = tenant
6064 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6065 2 : .await?;
6066 2 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6067 2 :
6068 2 : tenant
6069 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6070 2 : .await?;
6071 2 : let newtline = tenant
6072 2 : .get_timeline(NEW_TIMELINE_ID, true)
6073 2 : .expect("Should have a local timeline");
6074 2 :
6075 2 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6076 2 :
6077 2 : tline.set_broken("test".to_owned());
6078 2 :
6079 2 : tenant
6080 2 : .gc_iteration(
6081 2 : Some(TIMELINE_ID),
6082 2 : 0x10,
6083 2 : Duration::ZERO,
6084 2 : &CancellationToken::new(),
6085 2 : &ctx,
6086 2 : )
6087 2 : .await?;
6088 2 :
6089 2 : // The branchpoints should contain all timelines, even ones marked
6090 2 : // as Broken.
6091 2 : {
6092 2 : let branchpoints = &tline.gc_info.read().unwrap().retain_lsns;
6093 2 : assert_eq!(branchpoints.len(), 1);
6094 2 : assert_eq!(
6095 2 : branchpoints[0],
6096 2 : (Lsn(0x40), NEW_TIMELINE_ID, MaybeOffloaded::No)
6097 2 : );
6098 2 : }
6099 2 :
6100 2 : // You can read the key from the child branch even though the parent is
6101 2 : // Broken, as long as you don't need to access data from the parent.
6102 2 : assert_eq!(
6103 2 : newtline.get(*TEST_KEY, Lsn(0x70), &ctx).await?,
6104 2 : test_img(&format!("foo at {}", Lsn(0x70)))
6105 2 : );
6106 2 :
6107 2 : // This needs to traverse to the parent, and fails.
6108 2 : let err = newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await.unwrap_err();
6109 2 : assert!(
6110 2 : err.to_string().starts_with(&format!(
6111 2 : "bad state on timeline {}: Broken",
6112 2 : tline.timeline_id
6113 2 : )),
6114 2 : "{err}"
6115 2 : );
6116 2 :
6117 2 : Ok(())
6118 2 : }
6119 :
6120 : #[tokio::test]
6121 2 : async fn test_retain_data_in_parent_which_is_needed_for_child() -> anyhow::Result<()> {
6122 2 : let (tenant, ctx) =
6123 2 : TenantHarness::create("test_retain_data_in_parent_which_is_needed_for_child")
6124 2 : .await?
6125 2 : .load()
6126 2 : .await;
6127 2 : let tline = tenant
6128 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6129 2 : .await?;
6130 2 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6131 2 :
6132 2 : tenant
6133 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6134 2 : .await?;
6135 2 : let newtline = tenant
6136 2 : .get_timeline(NEW_TIMELINE_ID, true)
6137 2 : .expect("Should have a local timeline");
6138 2 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
6139 2 : tenant
6140 2 : .gc_iteration(
6141 2 : Some(TIMELINE_ID),
6142 2 : 0x10,
6143 2 : Duration::ZERO,
6144 2 : &CancellationToken::new(),
6145 2 : &ctx,
6146 2 : )
6147 2 : .await?;
6148 2 : assert!(newtline.get(*TEST_KEY, Lsn(0x25), &ctx).await.is_ok());
6149 2 :
6150 2 : Ok(())
6151 2 : }
6152 : #[tokio::test]
6153 2 : async fn test_parent_keeps_data_forever_after_branching() -> anyhow::Result<()> {
6154 2 : let (tenant, ctx) = TenantHarness::create("test_parent_keeps_data_forever_after_branching")
6155 2 : .await?
6156 2 : .load()
6157 2 : .await;
6158 2 : let tline = tenant
6159 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6160 2 : .await?;
6161 2 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6162 2 :
6163 2 : tenant
6164 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6165 2 : .await?;
6166 2 : let newtline = tenant
6167 2 : .get_timeline(NEW_TIMELINE_ID, true)
6168 2 : .expect("Should have a local timeline");
6169 2 :
6170 2 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6171 2 :
6172 2 : // run gc on parent
6173 2 : tenant
6174 2 : .gc_iteration(
6175 2 : Some(TIMELINE_ID),
6176 2 : 0x10,
6177 2 : Duration::ZERO,
6178 2 : &CancellationToken::new(),
6179 2 : &ctx,
6180 2 : )
6181 2 : .await?;
6182 2 :
6183 2 : // Check that the data is still accessible on the branch.
6184 2 : assert_eq!(
6185 2 : newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
6186 2 : test_img(&format!("foo at {}", Lsn(0x40)))
6187 2 : );
6188 2 :
6189 2 : Ok(())
6190 2 : }
6191 :
6192 : #[tokio::test]
6193 2 : async fn timeline_load() -> anyhow::Result<()> {
6194 2 : const TEST_NAME: &str = "timeline_load";
6195 2 : let harness = TenantHarness::create(TEST_NAME).await?;
6196 2 : {
6197 2 : let (tenant, ctx) = harness.load().await;
6198 2 : let tline = tenant
6199 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x7000), DEFAULT_PG_VERSION, &ctx)
6200 2 : .await?;
6201 2 : make_some_layers(tline.as_ref(), Lsn(0x8000), &ctx).await?;
6202 2 : // so that all uploads finish & we can call harness.load() below again
6203 2 : tenant
6204 2 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
6205 2 : .instrument(harness.span())
6206 2 : .await
6207 2 : .ok()
6208 2 : .unwrap();
6209 2 : }
6210 2 :
6211 2 : let (tenant, _ctx) = harness.load().await;
6212 2 : tenant
6213 2 : .get_timeline(TIMELINE_ID, true)
6214 2 : .expect("cannot load timeline");
6215 2 :
6216 2 : Ok(())
6217 2 : }
6218 :
6219 : #[tokio::test]
6220 2 : async fn timeline_load_with_ancestor() -> anyhow::Result<()> {
6221 2 : const TEST_NAME: &str = "timeline_load_with_ancestor";
6222 2 : let harness = TenantHarness::create(TEST_NAME).await?;
6223 2 : // create two timelines
6224 2 : {
6225 2 : let (tenant, ctx) = harness.load().await;
6226 2 : let tline = tenant
6227 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6228 2 : .await?;
6229 2 :
6230 2 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6231 2 :
6232 2 : let child_tline = tenant
6233 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6234 2 : .await?;
6235 2 : child_tline.set_state(TimelineState::Active);
6236 2 :
6237 2 : let newtline = tenant
6238 2 : .get_timeline(NEW_TIMELINE_ID, true)
6239 2 : .expect("Should have a local timeline");
6240 2 :
6241 2 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6242 2 :
6243 2 : // so that all uploads finish & we can call harness.load() below again
6244 2 : tenant
6245 2 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
6246 2 : .instrument(harness.span())
6247 2 : .await
6248 2 : .ok()
6249 2 : .unwrap();
6250 2 : }
6251 2 :
6252 2 : // check that both of them are initially unloaded
6253 2 : let (tenant, _ctx) = harness.load().await;
6254 2 :
6255 2 : // check that both, child and ancestor are loaded
6256 2 : let _child_tline = tenant
6257 2 : .get_timeline(NEW_TIMELINE_ID, true)
6258 2 : .expect("cannot get child timeline loaded");
6259 2 :
6260 2 : let _ancestor_tline = tenant
6261 2 : .get_timeline(TIMELINE_ID, true)
6262 2 : .expect("cannot get ancestor timeline loaded");
6263 2 :
6264 2 : Ok(())
6265 2 : }
6266 :
6267 : #[tokio::test]
6268 2 : async fn delta_layer_dumping() -> anyhow::Result<()> {
6269 2 : use storage_layer::AsLayerDesc;
6270 2 : let (tenant, ctx) = TenantHarness::create("test_layer_dumping")
6271 2 : .await?
6272 2 : .load()
6273 2 : .await;
6274 2 : let tline = tenant
6275 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6276 2 : .await?;
6277 2 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6278 2 :
6279 2 : let layer_map = tline.layers.read().await;
6280 2 : let level0_deltas = layer_map
6281 2 : .layer_map()?
6282 2 : .level0_deltas()
6283 2 : .iter()
6284 4 : .map(|desc| layer_map.get_from_desc(desc))
6285 2 : .collect::<Vec<_>>();
6286 2 :
6287 2 : assert!(!level0_deltas.is_empty());
6288 2 :
6289 6 : for delta in level0_deltas {
6290 2 : // Ensure we are dumping a delta layer here
6291 4 : assert!(delta.layer_desc().is_delta);
6292 4 : delta.dump(true, &ctx).await.unwrap();
6293 2 : }
6294 2 :
6295 2 : Ok(())
6296 2 : }
6297 :
6298 : #[tokio::test]
6299 2 : async fn test_images() -> anyhow::Result<()> {
6300 2 : let (tenant, ctx) = TenantHarness::create("test_images").await?.load().await;
6301 2 : let tline = tenant
6302 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6303 2 : .await?;
6304 2 :
6305 2 : let mut writer = tline.writer().await;
6306 2 : writer
6307 2 : .put(
6308 2 : *TEST_KEY,
6309 2 : Lsn(0x10),
6310 2 : &Value::Image(test_img("foo at 0x10")),
6311 2 : &ctx,
6312 2 : )
6313 2 : .await?;
6314 2 : writer.finish_write(Lsn(0x10));
6315 2 : drop(writer);
6316 2 :
6317 2 : tline.freeze_and_flush().await?;
6318 2 : tline
6319 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6320 2 : .await?;
6321 2 :
6322 2 : let mut writer = tline.writer().await;
6323 2 : writer
6324 2 : .put(
6325 2 : *TEST_KEY,
6326 2 : Lsn(0x20),
6327 2 : &Value::Image(test_img("foo at 0x20")),
6328 2 : &ctx,
6329 2 : )
6330 2 : .await?;
6331 2 : writer.finish_write(Lsn(0x20));
6332 2 : drop(writer);
6333 2 :
6334 2 : tline.freeze_and_flush().await?;
6335 2 : tline
6336 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6337 2 : .await?;
6338 2 :
6339 2 : let mut writer = tline.writer().await;
6340 2 : writer
6341 2 : .put(
6342 2 : *TEST_KEY,
6343 2 : Lsn(0x30),
6344 2 : &Value::Image(test_img("foo at 0x30")),
6345 2 : &ctx,
6346 2 : )
6347 2 : .await?;
6348 2 : writer.finish_write(Lsn(0x30));
6349 2 : drop(writer);
6350 2 :
6351 2 : tline.freeze_and_flush().await?;
6352 2 : tline
6353 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6354 2 : .await?;
6355 2 :
6356 2 : let mut writer = tline.writer().await;
6357 2 : writer
6358 2 : .put(
6359 2 : *TEST_KEY,
6360 2 : Lsn(0x40),
6361 2 : &Value::Image(test_img("foo at 0x40")),
6362 2 : &ctx,
6363 2 : )
6364 2 : .await?;
6365 2 : writer.finish_write(Lsn(0x40));
6366 2 : drop(writer);
6367 2 :
6368 2 : tline.freeze_and_flush().await?;
6369 2 : tline
6370 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6371 2 : .await?;
6372 2 :
6373 2 : assert_eq!(
6374 2 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
6375 2 : test_img("foo at 0x10")
6376 2 : );
6377 2 : assert_eq!(
6378 2 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
6379 2 : test_img("foo at 0x10")
6380 2 : );
6381 2 : assert_eq!(
6382 2 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
6383 2 : test_img("foo at 0x20")
6384 2 : );
6385 2 : assert_eq!(
6386 2 : tline.get(*TEST_KEY, Lsn(0x30), &ctx).await?,
6387 2 : test_img("foo at 0x30")
6388 2 : );
6389 2 : assert_eq!(
6390 2 : tline.get(*TEST_KEY, Lsn(0x40), &ctx).await?,
6391 2 : test_img("foo at 0x40")
6392 2 : );
6393 2 :
6394 2 : Ok(())
6395 2 : }
6396 :
6397 4 : async fn bulk_insert_compact_gc(
6398 4 : tenant: &Tenant,
6399 4 : timeline: &Arc<Timeline>,
6400 4 : ctx: &RequestContext,
6401 4 : lsn: Lsn,
6402 4 : repeat: usize,
6403 4 : key_count: usize,
6404 4 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
6405 4 : let compact = true;
6406 4 : bulk_insert_maybe_compact_gc(tenant, timeline, ctx, lsn, repeat, key_count, compact).await
6407 4 : }
6408 :
6409 8 : async fn bulk_insert_maybe_compact_gc(
6410 8 : tenant: &Tenant,
6411 8 : timeline: &Arc<Timeline>,
6412 8 : ctx: &RequestContext,
6413 8 : mut lsn: Lsn,
6414 8 : repeat: usize,
6415 8 : key_count: usize,
6416 8 : compact: bool,
6417 8 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
6418 8 : let mut inserted: HashMap<Key, BTreeSet<Lsn>> = Default::default();
6419 8 :
6420 8 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6421 8 : let mut blknum = 0;
6422 8 :
6423 8 : // Enforce that key range is monotonously increasing
6424 8 : let mut keyspace = KeySpaceAccum::new();
6425 8 :
6426 8 : let cancel = CancellationToken::new();
6427 8 :
6428 8 : for _ in 0..repeat {
6429 400 : for _ in 0..key_count {
6430 4000000 : test_key.field6 = blknum;
6431 4000000 : let mut writer = timeline.writer().await;
6432 4000000 : writer
6433 4000000 : .put(
6434 4000000 : test_key,
6435 4000000 : lsn,
6436 4000000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6437 4000000 : ctx,
6438 4000000 : )
6439 4000000 : .await?;
6440 4000000 : inserted.entry(test_key).or_default().insert(lsn);
6441 4000000 : writer.finish_write(lsn);
6442 4000000 : drop(writer);
6443 4000000 :
6444 4000000 : keyspace.add_key(test_key);
6445 4000000 :
6446 4000000 : lsn = Lsn(lsn.0 + 0x10);
6447 4000000 : blknum += 1;
6448 : }
6449 :
6450 400 : timeline.freeze_and_flush().await?;
6451 400 : if compact {
6452 : // this requires timeline to be &Arc<Timeline>
6453 200 : timeline.compact(&cancel, EnumSet::empty(), ctx).await?;
6454 200 : }
6455 :
6456 : // this doesn't really need to use the timeline_id target, but it is closer to what it
6457 : // originally was.
6458 400 : let res = tenant
6459 400 : .gc_iteration(Some(timeline.timeline_id), 0, Duration::ZERO, &cancel, ctx)
6460 400 : .await?;
6461 :
6462 400 : assert_eq!(res.layers_removed, 0, "this never removes anything");
6463 : }
6464 :
6465 8 : Ok(inserted)
6466 8 : }
6467 :
6468 : //
6469 : // Insert 1000 key-value pairs with increasing keys, flush, compact, GC.
6470 : // Repeat 50 times.
6471 : //
6472 : #[tokio::test]
6473 2 : async fn test_bulk_insert() -> anyhow::Result<()> {
6474 2 : let harness = TenantHarness::create("test_bulk_insert").await?;
6475 2 : let (tenant, ctx) = harness.load().await;
6476 2 : let tline = tenant
6477 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6478 2 : .await?;
6479 2 :
6480 2 : let lsn = Lsn(0x10);
6481 2 : bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
6482 2 :
6483 2 : Ok(())
6484 2 : }
6485 :
6486 : // Test the vectored get real implementation against a simple sequential implementation.
6487 : //
6488 : // The test generates a keyspace by repeatedly flushing the in-memory layer and compacting.
6489 : // Projected to 2D the key space looks like below. Lsn grows upwards on the Y axis and keys
6490 : // grow to the right on the X axis.
6491 : // [Delta]
6492 : // [Delta]
6493 : // [Delta]
6494 : // [Delta]
6495 : // ------------ Image ---------------
6496 : //
6497 : // After layer generation we pick the ranges to query as follows:
6498 : // 1. The beginning of each delta layer
6499 : // 2. At the seam between two adjacent delta layers
6500 : //
6501 : // There's one major downside to this test: delta layers only contains images,
6502 : // so the search can stop at the first delta layer and doesn't traverse any deeper.
6503 : #[tokio::test]
6504 2 : async fn test_get_vectored() -> anyhow::Result<()> {
6505 2 : let harness = TenantHarness::create("test_get_vectored").await?;
6506 2 : let (tenant, ctx) = harness.load().await;
6507 2 : let tline = tenant
6508 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6509 2 : .await?;
6510 2 :
6511 2 : let lsn = Lsn(0x10);
6512 2 : let inserted = bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
6513 2 :
6514 2 : let guard = tline.layers.read().await;
6515 2 : let lm = guard.layer_map()?;
6516 2 :
6517 2 : lm.dump(true, &ctx).await?;
6518 2 :
6519 2 : let mut reads = Vec::new();
6520 2 : let mut prev = None;
6521 12 : lm.iter_historic_layers().for_each(|desc| {
6522 12 : if !desc.is_delta() {
6523 2 : prev = Some(desc.clone());
6524 2 : return;
6525 10 : }
6526 10 :
6527 10 : let start = desc.key_range.start;
6528 10 : let end = desc
6529 10 : .key_range
6530 10 : .start
6531 10 : .add(Timeline::MAX_GET_VECTORED_KEYS.try_into().unwrap());
6532 10 : reads.push(KeySpace {
6533 10 : ranges: vec![start..end],
6534 10 : });
6535 2 :
6536 10 : if let Some(prev) = &prev {
6537 10 : if !prev.is_delta() {
6538 10 : return;
6539 2 : }
6540 0 :
6541 0 : let first_range = Key {
6542 0 : field6: prev.key_range.end.field6 - 4,
6543 0 : ..prev.key_range.end
6544 0 : }..prev.key_range.end;
6545 0 :
6546 0 : let second_range = desc.key_range.start..Key {
6547 0 : field6: desc.key_range.start.field6 + 4,
6548 0 : ..desc.key_range.start
6549 0 : };
6550 0 :
6551 0 : reads.push(KeySpace {
6552 0 : ranges: vec![first_range, second_range],
6553 0 : });
6554 2 : };
6555 2 :
6556 2 : prev = Some(desc.clone());
6557 12 : });
6558 2 :
6559 2 : drop(guard);
6560 2 :
6561 2 : // Pick a big LSN such that we query over all the changes.
6562 2 : let reads_lsn = Lsn(u64::MAX - 1);
6563 2 :
6564 12 : for read in reads {
6565 10 : info!("Doing vectored read on {:?}", read);
6566 2 :
6567 10 : let vectored_res = tline
6568 10 : .get_vectored_impl(
6569 10 : read.clone(),
6570 10 : reads_lsn,
6571 10 : &mut ValuesReconstructState::new(),
6572 10 : &ctx,
6573 10 : )
6574 10 : .await;
6575 2 :
6576 10 : let mut expected_lsns: HashMap<Key, Lsn> = Default::default();
6577 10 : let mut expect_missing = false;
6578 10 : let mut key = read.start().unwrap();
6579 330 : while key != read.end().unwrap() {
6580 320 : if let Some(lsns) = inserted.get(&key) {
6581 320 : let expected_lsn = lsns.iter().rfind(|lsn| **lsn <= reads_lsn);
6582 320 : match expected_lsn {
6583 320 : Some(lsn) => {
6584 320 : expected_lsns.insert(key, *lsn);
6585 320 : }
6586 2 : None => {
6587 2 : expect_missing = true;
6588 0 : break;
6589 2 : }
6590 2 : }
6591 2 : } else {
6592 2 : expect_missing = true;
6593 0 : break;
6594 2 : }
6595 2 :
6596 320 : key = key.next();
6597 2 : }
6598 2 :
6599 10 : if expect_missing {
6600 2 : assert!(matches!(vectored_res, Err(GetVectoredError::MissingKey(_))));
6601 2 : } else {
6602 320 : for (key, image) in vectored_res? {
6603 320 : let expected_lsn = expected_lsns.get(&key).expect("determined above");
6604 320 : let expected_image = test_img(&format!("{} at {}", key.field6, expected_lsn));
6605 320 : assert_eq!(image?, expected_image);
6606 2 : }
6607 2 : }
6608 2 : }
6609 2 :
6610 2 : Ok(())
6611 2 : }
6612 :
6613 : #[tokio::test]
6614 2 : async fn test_get_vectored_aux_files() -> anyhow::Result<()> {
6615 2 : let harness = TenantHarness::create("test_get_vectored_aux_files").await?;
6616 2 :
6617 2 : let (tenant, ctx) = harness.load().await;
6618 2 : let tline = tenant
6619 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
6620 2 : .await?;
6621 2 : let tline = tline.raw_timeline().unwrap();
6622 2 :
6623 2 : let mut modification = tline.begin_modification(Lsn(0x1000));
6624 2 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
6625 2 : modification.set_lsn(Lsn(0x1008))?;
6626 2 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
6627 2 : modification.commit(&ctx).await?;
6628 2 :
6629 2 : let child_timeline_id = TimelineId::generate();
6630 2 : tenant
6631 2 : .branch_timeline_test(
6632 2 : tline,
6633 2 : child_timeline_id,
6634 2 : Some(tline.get_last_record_lsn()),
6635 2 : &ctx,
6636 2 : )
6637 2 : .await?;
6638 2 :
6639 2 : let child_timeline = tenant
6640 2 : .get_timeline(child_timeline_id, true)
6641 2 : .expect("Should have the branched timeline");
6642 2 :
6643 2 : let aux_keyspace = KeySpace {
6644 2 : ranges: vec![NON_INHERITED_RANGE],
6645 2 : };
6646 2 : let read_lsn = child_timeline.get_last_record_lsn();
6647 2 :
6648 2 : let vectored_res = child_timeline
6649 2 : .get_vectored_impl(
6650 2 : aux_keyspace.clone(),
6651 2 : read_lsn,
6652 2 : &mut ValuesReconstructState::new(),
6653 2 : &ctx,
6654 2 : )
6655 2 : .await;
6656 2 :
6657 2 : let images = vectored_res?;
6658 2 : assert!(images.is_empty());
6659 2 : Ok(())
6660 2 : }
6661 :
6662 : // Test that vectored get handles layer gaps correctly
6663 : // by advancing into the next ancestor timeline if required.
6664 : //
6665 : // The test generates timelines that look like the diagram below.
6666 : // We leave a gap in one of the L1 layers at `gap_at_key` (`/` in the diagram).
6667 : // The reconstruct data for that key lies in the ancestor timeline (`X` in the diagram).
6668 : //
6669 : // ```
6670 : //-------------------------------+
6671 : // ... |
6672 : // [ L1 ] |
6673 : // [ / L1 ] | Child Timeline
6674 : // ... |
6675 : // ------------------------------+
6676 : // [ X L1 ] | Parent Timeline
6677 : // ------------------------------+
6678 : // ```
6679 : #[tokio::test]
6680 2 : async fn test_get_vectored_key_gap() -> anyhow::Result<()> {
6681 2 : let tenant_conf = TenantConf {
6682 2 : // Make compaction deterministic
6683 2 : gc_period: Duration::ZERO,
6684 2 : compaction_period: Duration::ZERO,
6685 2 : // Encourage creation of L1 layers
6686 2 : checkpoint_distance: 16 * 1024,
6687 2 : compaction_target_size: 8 * 1024,
6688 2 : ..TenantConf::default()
6689 2 : };
6690 2 :
6691 2 : let harness = TenantHarness::create_custom(
6692 2 : "test_get_vectored_key_gap",
6693 2 : tenant_conf,
6694 2 : TenantId::generate(),
6695 2 : ShardIdentity::unsharded(),
6696 2 : Generation::new(0xdeadbeef),
6697 2 : )
6698 2 : .await?;
6699 2 : let (tenant, ctx) = harness.load().await;
6700 2 :
6701 2 : let mut current_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6702 2 : let gap_at_key = current_key.add(100);
6703 2 : let mut current_lsn = Lsn(0x10);
6704 2 :
6705 2 : const KEY_COUNT: usize = 10_000;
6706 2 :
6707 2 : let timeline_id = TimelineId::generate();
6708 2 : let current_timeline = tenant
6709 2 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
6710 2 : .await?;
6711 2 :
6712 2 : current_lsn += 0x100;
6713 2 :
6714 2 : let mut writer = current_timeline.writer().await;
6715 2 : writer
6716 2 : .put(
6717 2 : gap_at_key,
6718 2 : current_lsn,
6719 2 : &Value::Image(test_img(&format!("{} at {}", gap_at_key, current_lsn))),
6720 2 : &ctx,
6721 2 : )
6722 2 : .await?;
6723 2 : writer.finish_write(current_lsn);
6724 2 : drop(writer);
6725 2 :
6726 2 : let mut latest_lsns = HashMap::new();
6727 2 : latest_lsns.insert(gap_at_key, current_lsn);
6728 2 :
6729 2 : current_timeline.freeze_and_flush().await?;
6730 2 :
6731 2 : let child_timeline_id = TimelineId::generate();
6732 2 :
6733 2 : tenant
6734 2 : .branch_timeline_test(
6735 2 : ¤t_timeline,
6736 2 : child_timeline_id,
6737 2 : Some(current_lsn),
6738 2 : &ctx,
6739 2 : )
6740 2 : .await?;
6741 2 : let child_timeline = tenant
6742 2 : .get_timeline(child_timeline_id, true)
6743 2 : .expect("Should have the branched timeline");
6744 2 :
6745 20002 : for i in 0..KEY_COUNT {
6746 20000 : if current_key == gap_at_key {
6747 2 : current_key = current_key.next();
6748 2 : continue;
6749 19998 : }
6750 19998 :
6751 19998 : current_lsn += 0x10;
6752 2 :
6753 19998 : let mut writer = child_timeline.writer().await;
6754 19998 : writer
6755 19998 : .put(
6756 19998 : current_key,
6757 19998 : current_lsn,
6758 19998 : &Value::Image(test_img(&format!("{} at {}", current_key, current_lsn))),
6759 19998 : &ctx,
6760 19998 : )
6761 19998 : .await?;
6762 19998 : writer.finish_write(current_lsn);
6763 19998 : drop(writer);
6764 19998 :
6765 19998 : latest_lsns.insert(current_key, current_lsn);
6766 19998 : current_key = current_key.next();
6767 19998 :
6768 19998 : // Flush every now and then to encourage layer file creation.
6769 19998 : if i % 500 == 0 {
6770 40 : child_timeline.freeze_and_flush().await?;
6771 19958 : }
6772 2 : }
6773 2 :
6774 2 : child_timeline.freeze_and_flush().await?;
6775 2 : let mut flags = EnumSet::new();
6776 2 : flags.insert(CompactFlags::ForceRepartition);
6777 2 : child_timeline
6778 2 : .compact(&CancellationToken::new(), flags, &ctx)
6779 2 : .await?;
6780 2 :
6781 2 : let key_near_end = {
6782 2 : let mut tmp = current_key;
6783 2 : tmp.field6 -= 10;
6784 2 : tmp
6785 2 : };
6786 2 :
6787 2 : let key_near_gap = {
6788 2 : let mut tmp = gap_at_key;
6789 2 : tmp.field6 -= 10;
6790 2 : tmp
6791 2 : };
6792 2 :
6793 2 : let read = KeySpace {
6794 2 : ranges: vec![key_near_gap..gap_at_key.next(), key_near_end..current_key],
6795 2 : };
6796 2 : let results = child_timeline
6797 2 : .get_vectored_impl(
6798 2 : read.clone(),
6799 2 : current_lsn,
6800 2 : &mut ValuesReconstructState::new(),
6801 2 : &ctx,
6802 2 : )
6803 2 : .await?;
6804 2 :
6805 44 : for (key, img_res) in results {
6806 42 : let expected = test_img(&format!("{} at {}", key, latest_lsns[&key]));
6807 42 : assert_eq!(img_res?, expected);
6808 2 : }
6809 2 :
6810 2 : Ok(())
6811 2 : }
6812 :
6813 : // Test that vectored get descends into ancestor timelines correctly and
6814 : // does not return an image that's newer than requested.
6815 : //
6816 : // The diagram below ilustrates an interesting case. We have a parent timeline
6817 : // (top of the Lsn range) and a child timeline. The request key cannot be reconstructed
6818 : // from the child timeline, so the parent timeline must be visited. When advacing into
6819 : // the child timeline, the read path needs to remember what the requested Lsn was in
6820 : // order to avoid returning an image that's too new. The test below constructs such
6821 : // a timeline setup and does a few queries around the Lsn of each page image.
6822 : // ```
6823 : // LSN
6824 : // ^
6825 : // |
6826 : // |
6827 : // 500 | --------------------------------------> branch point
6828 : // 400 | X
6829 : // 300 | X
6830 : // 200 | --------------------------------------> requested lsn
6831 : // 100 | X
6832 : // |---------------------------------------> Key
6833 : // |
6834 : // ------> requested key
6835 : //
6836 : // Legend:
6837 : // * X - page images
6838 : // ```
6839 : #[tokio::test]
6840 2 : async fn test_get_vectored_ancestor_descent() -> anyhow::Result<()> {
6841 2 : let harness = TenantHarness::create("test_get_vectored_on_lsn_axis").await?;
6842 2 : let (tenant, ctx) = harness.load().await;
6843 2 :
6844 2 : let start_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6845 2 : let end_key = start_key.add(1000);
6846 2 : let child_gap_at_key = start_key.add(500);
6847 2 : let mut parent_gap_lsns: BTreeMap<Lsn, String> = BTreeMap::new();
6848 2 :
6849 2 : let mut current_lsn = Lsn(0x10);
6850 2 :
6851 2 : let timeline_id = TimelineId::generate();
6852 2 : let parent_timeline = tenant
6853 2 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
6854 2 : .await?;
6855 2 :
6856 2 : current_lsn += 0x100;
6857 2 :
6858 8 : for _ in 0..3 {
6859 6 : let mut key = start_key;
6860 6006 : while key < end_key {
6861 6000 : current_lsn += 0x10;
6862 6000 :
6863 6000 : let image_value = format!("{} at {}", child_gap_at_key, current_lsn);
6864 2 :
6865 6000 : let mut writer = parent_timeline.writer().await;
6866 6000 : writer
6867 6000 : .put(
6868 6000 : key,
6869 6000 : current_lsn,
6870 6000 : &Value::Image(test_img(&image_value)),
6871 6000 : &ctx,
6872 6000 : )
6873 6000 : .await?;
6874 6000 : writer.finish_write(current_lsn);
6875 6000 :
6876 6000 : if key == child_gap_at_key {
6877 6 : parent_gap_lsns.insert(current_lsn, image_value);
6878 5994 : }
6879 2 :
6880 6000 : key = key.next();
6881 2 : }
6882 2 :
6883 6 : parent_timeline.freeze_and_flush().await?;
6884 2 : }
6885 2 :
6886 2 : let child_timeline_id = TimelineId::generate();
6887 2 :
6888 2 : let child_timeline = tenant
6889 2 : .branch_timeline_test(&parent_timeline, child_timeline_id, Some(current_lsn), &ctx)
6890 2 : .await?;
6891 2 :
6892 2 : let mut key = start_key;
6893 2002 : while key < end_key {
6894 2000 : if key == child_gap_at_key {
6895 2 : key = key.next();
6896 2 : continue;
6897 1998 : }
6898 1998 :
6899 1998 : current_lsn += 0x10;
6900 2 :
6901 1998 : let mut writer = child_timeline.writer().await;
6902 1998 : writer
6903 1998 : .put(
6904 1998 : key,
6905 1998 : current_lsn,
6906 1998 : &Value::Image(test_img(&format!("{} at {}", key, current_lsn))),
6907 1998 : &ctx,
6908 1998 : )
6909 1998 : .await?;
6910 1998 : writer.finish_write(current_lsn);
6911 1998 :
6912 1998 : key = key.next();
6913 2 : }
6914 2 :
6915 2 : child_timeline.freeze_and_flush().await?;
6916 2 :
6917 2 : let lsn_offsets: [i64; 5] = [-10, -1, 0, 1, 10];
6918 2 : let mut query_lsns = Vec::new();
6919 6 : for image_lsn in parent_gap_lsns.keys().rev() {
6920 36 : for offset in lsn_offsets {
6921 30 : query_lsns.push(Lsn(image_lsn
6922 30 : .0
6923 30 : .checked_add_signed(offset)
6924 30 : .expect("Shouldn't overflow")));
6925 30 : }
6926 2 : }
6927 2 :
6928 32 : for query_lsn in query_lsns {
6929 30 : let results = child_timeline
6930 30 : .get_vectored_impl(
6931 30 : KeySpace {
6932 30 : ranges: vec![child_gap_at_key..child_gap_at_key.next()],
6933 30 : },
6934 30 : query_lsn,
6935 30 : &mut ValuesReconstructState::new(),
6936 30 : &ctx,
6937 30 : )
6938 30 : .await;
6939 2 :
6940 30 : let expected_item = parent_gap_lsns
6941 30 : .iter()
6942 30 : .rev()
6943 68 : .find(|(lsn, _)| **lsn <= query_lsn);
6944 30 :
6945 30 : info!(
6946 2 : "Doing vectored read at LSN {}. Expecting image to be: {:?}",
6947 2 : query_lsn, expected_item
6948 2 : );
6949 2 :
6950 30 : match expected_item {
6951 26 : Some((_, img_value)) => {
6952 26 : let key_results = results.expect("No vectored get error expected");
6953 26 : let key_result = &key_results[&child_gap_at_key];
6954 26 : let returned_img = key_result
6955 26 : .as_ref()
6956 26 : .expect("No page reconstruct error expected");
6957 26 :
6958 26 : info!(
6959 2 : "Vectored read at LSN {} returned image {}",
6960 0 : query_lsn,
6961 0 : std::str::from_utf8(returned_img)?
6962 2 : );
6963 26 : assert_eq!(*returned_img, test_img(img_value));
6964 2 : }
6965 2 : None => {
6966 4 : assert!(matches!(results, Err(GetVectoredError::MissingKey(_))));
6967 2 : }
6968 2 : }
6969 2 : }
6970 2 :
6971 2 : Ok(())
6972 2 : }
6973 :
6974 : #[tokio::test]
6975 2 : async fn test_random_updates() -> anyhow::Result<()> {
6976 2 : let names_algorithms = [
6977 2 : ("test_random_updates_legacy", CompactionAlgorithm::Legacy),
6978 2 : ("test_random_updates_tiered", CompactionAlgorithm::Tiered),
6979 2 : ];
6980 6 : for (name, algorithm) in names_algorithms {
6981 4 : test_random_updates_algorithm(name, algorithm).await?;
6982 2 : }
6983 2 : Ok(())
6984 2 : }
6985 :
6986 4 : async fn test_random_updates_algorithm(
6987 4 : name: &'static str,
6988 4 : compaction_algorithm: CompactionAlgorithm,
6989 4 : ) -> anyhow::Result<()> {
6990 4 : let mut harness = TenantHarness::create(name).await?;
6991 4 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
6992 4 : kind: compaction_algorithm,
6993 4 : };
6994 4 : let (tenant, ctx) = harness.load().await;
6995 4 : let tline = tenant
6996 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6997 4 : .await?;
6998 :
6999 : const NUM_KEYS: usize = 1000;
7000 4 : let cancel = CancellationToken::new();
7001 4 :
7002 4 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7003 4 : let mut test_key_end = test_key;
7004 4 : test_key_end.field6 = NUM_KEYS as u32;
7005 4 : tline.add_extra_test_dense_keyspace(KeySpace::single(test_key..test_key_end));
7006 4 :
7007 4 : let mut keyspace = KeySpaceAccum::new();
7008 4 :
7009 4 : // Track when each page was last modified. Used to assert that
7010 4 : // a read sees the latest page version.
7011 4 : let mut updated = [Lsn(0); NUM_KEYS];
7012 4 :
7013 4 : let mut lsn = Lsn(0x10);
7014 : #[allow(clippy::needless_range_loop)]
7015 4004 : for blknum in 0..NUM_KEYS {
7016 4000 : lsn = Lsn(lsn.0 + 0x10);
7017 4000 : test_key.field6 = blknum as u32;
7018 4000 : let mut writer = tline.writer().await;
7019 4000 : writer
7020 4000 : .put(
7021 4000 : test_key,
7022 4000 : lsn,
7023 4000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7024 4000 : &ctx,
7025 4000 : )
7026 4000 : .await?;
7027 4000 : writer.finish_write(lsn);
7028 4000 : updated[blknum] = lsn;
7029 4000 : drop(writer);
7030 4000 :
7031 4000 : keyspace.add_key(test_key);
7032 : }
7033 :
7034 204 : for _ in 0..50 {
7035 200200 : for _ in 0..NUM_KEYS {
7036 200000 : lsn = Lsn(lsn.0 + 0x10);
7037 200000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7038 200000 : test_key.field6 = blknum as u32;
7039 200000 : let mut writer = tline.writer().await;
7040 200000 : writer
7041 200000 : .put(
7042 200000 : test_key,
7043 200000 : lsn,
7044 200000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7045 200000 : &ctx,
7046 200000 : )
7047 200000 : .await?;
7048 200000 : writer.finish_write(lsn);
7049 200000 : drop(writer);
7050 200000 : updated[blknum] = lsn;
7051 : }
7052 :
7053 : // Read all the blocks
7054 200000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7055 200000 : test_key.field6 = blknum as u32;
7056 200000 : assert_eq!(
7057 200000 : tline.get(test_key, lsn, &ctx).await?,
7058 200000 : test_img(&format!("{} at {}", blknum, last_lsn))
7059 : );
7060 : }
7061 :
7062 : // Perform a cycle of flush, and GC
7063 200 : tline.freeze_and_flush().await?;
7064 200 : tenant
7065 200 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7066 200 : .await?;
7067 : }
7068 :
7069 4 : Ok(())
7070 4 : }
7071 :
7072 : #[tokio::test]
7073 2 : async fn test_traverse_branches() -> anyhow::Result<()> {
7074 2 : let (tenant, ctx) = TenantHarness::create("test_traverse_branches")
7075 2 : .await?
7076 2 : .load()
7077 2 : .await;
7078 2 : let mut tline = tenant
7079 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7080 2 : .await?;
7081 2 :
7082 2 : const NUM_KEYS: usize = 1000;
7083 2 :
7084 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7085 2 :
7086 2 : let mut keyspace = KeySpaceAccum::new();
7087 2 :
7088 2 : let cancel = CancellationToken::new();
7089 2 :
7090 2 : // Track when each page was last modified. Used to assert that
7091 2 : // a read sees the latest page version.
7092 2 : let mut updated = [Lsn(0); NUM_KEYS];
7093 2 :
7094 2 : let mut lsn = Lsn(0x10);
7095 2 : #[allow(clippy::needless_range_loop)]
7096 2002 : for blknum in 0..NUM_KEYS {
7097 2000 : lsn = Lsn(lsn.0 + 0x10);
7098 2000 : test_key.field6 = blknum as u32;
7099 2000 : let mut writer = tline.writer().await;
7100 2000 : writer
7101 2000 : .put(
7102 2000 : test_key,
7103 2000 : lsn,
7104 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7105 2000 : &ctx,
7106 2000 : )
7107 2000 : .await?;
7108 2000 : writer.finish_write(lsn);
7109 2000 : updated[blknum] = lsn;
7110 2000 : drop(writer);
7111 2000 :
7112 2000 : keyspace.add_key(test_key);
7113 2 : }
7114 2 :
7115 102 : for _ in 0..50 {
7116 100 : let new_tline_id = TimelineId::generate();
7117 100 : tenant
7118 100 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7119 100 : .await?;
7120 100 : tline = tenant
7121 100 : .get_timeline(new_tline_id, true)
7122 100 : .expect("Should have the branched timeline");
7123 2 :
7124 100100 : for _ in 0..NUM_KEYS {
7125 100000 : lsn = Lsn(lsn.0 + 0x10);
7126 100000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7127 100000 : test_key.field6 = blknum as u32;
7128 100000 : let mut writer = tline.writer().await;
7129 100000 : writer
7130 100000 : .put(
7131 100000 : test_key,
7132 100000 : lsn,
7133 100000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7134 100000 : &ctx,
7135 100000 : )
7136 100000 : .await?;
7137 100000 : println!("updating {} at {}", blknum, lsn);
7138 100000 : writer.finish_write(lsn);
7139 100000 : drop(writer);
7140 100000 : updated[blknum] = lsn;
7141 2 : }
7142 2 :
7143 2 : // Read all the blocks
7144 100000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7145 100000 : test_key.field6 = blknum as u32;
7146 100000 : assert_eq!(
7147 100000 : tline.get(test_key, lsn, &ctx).await?,
7148 100000 : test_img(&format!("{} at {}", blknum, last_lsn))
7149 2 : );
7150 2 : }
7151 2 :
7152 2 : // Perform a cycle of flush, compact, and GC
7153 100 : tline.freeze_and_flush().await?;
7154 100 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
7155 100 : tenant
7156 100 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7157 100 : .await?;
7158 2 : }
7159 2 :
7160 2 : Ok(())
7161 2 : }
7162 :
7163 : #[tokio::test]
7164 2 : async fn test_traverse_ancestors() -> anyhow::Result<()> {
7165 2 : let (tenant, ctx) = TenantHarness::create("test_traverse_ancestors")
7166 2 : .await?
7167 2 : .load()
7168 2 : .await;
7169 2 : let mut tline = tenant
7170 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7171 2 : .await?;
7172 2 :
7173 2 : const NUM_KEYS: usize = 100;
7174 2 : const NUM_TLINES: usize = 50;
7175 2 :
7176 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7177 2 : // Track page mutation lsns across different timelines.
7178 2 : let mut updated = [[Lsn(0); NUM_KEYS]; NUM_TLINES];
7179 2 :
7180 2 : let mut lsn = Lsn(0x10);
7181 2 :
7182 2 : #[allow(clippy::needless_range_loop)]
7183 102 : for idx in 0..NUM_TLINES {
7184 100 : let new_tline_id = TimelineId::generate();
7185 100 : tenant
7186 100 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7187 100 : .await?;
7188 100 : tline = tenant
7189 100 : .get_timeline(new_tline_id, true)
7190 100 : .expect("Should have the branched timeline");
7191 2 :
7192 10100 : for _ in 0..NUM_KEYS {
7193 10000 : lsn = Lsn(lsn.0 + 0x10);
7194 10000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7195 10000 : test_key.field6 = blknum as u32;
7196 10000 : let mut writer = tline.writer().await;
7197 10000 : writer
7198 10000 : .put(
7199 10000 : test_key,
7200 10000 : lsn,
7201 10000 : &Value::Image(test_img(&format!("{} {} at {}", idx, blknum, lsn))),
7202 10000 : &ctx,
7203 10000 : )
7204 10000 : .await?;
7205 10000 : println!("updating [{}][{}] at {}", idx, blknum, lsn);
7206 10000 : writer.finish_write(lsn);
7207 10000 : drop(writer);
7208 10000 : updated[idx][blknum] = lsn;
7209 2 : }
7210 2 : }
7211 2 :
7212 2 : // Read pages from leaf timeline across all ancestors.
7213 100 : for (idx, lsns) in updated.iter().enumerate() {
7214 10000 : for (blknum, lsn) in lsns.iter().enumerate() {
7215 2 : // Skip empty mutations.
7216 10000 : if lsn.0 == 0 {
7217 3680 : continue;
7218 6320 : }
7219 6320 : println!("checking [{idx}][{blknum}] at {lsn}");
7220 6320 : test_key.field6 = blknum as u32;
7221 6320 : assert_eq!(
7222 6320 : tline.get(test_key, *lsn, &ctx).await?,
7223 6320 : test_img(&format!("{idx} {blknum} at {lsn}"))
7224 2 : );
7225 2 : }
7226 2 : }
7227 2 : Ok(())
7228 2 : }
7229 :
7230 : #[tokio::test]
7231 2 : async fn test_write_at_initdb_lsn_takes_optimization_code_path() -> anyhow::Result<()> {
7232 2 : let (tenant, ctx) = TenantHarness::create("test_empty_test_timeline_is_usable")
7233 2 : .await?
7234 2 : .load()
7235 2 : .await;
7236 2 :
7237 2 : let initdb_lsn = Lsn(0x20);
7238 2 : let utline = tenant
7239 2 : .create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)
7240 2 : .await?;
7241 2 : let tline = utline.raw_timeline().unwrap();
7242 2 :
7243 2 : // Spawn flush loop now so that we can set the `expect_initdb_optimization`
7244 2 : tline.maybe_spawn_flush_loop();
7245 2 :
7246 2 : // Make sure the timeline has the minimum set of required keys for operation.
7247 2 : // The only operation you can always do on an empty timeline is to `put` new data.
7248 2 : // Except if you `put` at `initdb_lsn`.
7249 2 : // In that case, there's an optimization to directly create image layers instead of delta layers.
7250 2 : // It uses `repartition()`, which assumes some keys to be present.
7251 2 : // Let's make sure the test timeline can handle that case.
7252 2 : {
7253 2 : let mut state = tline.flush_loop_state.lock().unwrap();
7254 2 : assert_eq!(
7255 2 : timeline::FlushLoopState::Running {
7256 2 : expect_initdb_optimization: false,
7257 2 : initdb_optimization_count: 0,
7258 2 : },
7259 2 : *state
7260 2 : );
7261 2 : *state = timeline::FlushLoopState::Running {
7262 2 : expect_initdb_optimization: true,
7263 2 : initdb_optimization_count: 0,
7264 2 : };
7265 2 : }
7266 2 :
7267 2 : // Make writes at the initdb_lsn. When we flush it below, it should be handled by the optimization.
7268 2 : // As explained above, the optimization requires some keys to be present.
7269 2 : // As per `create_empty_timeline` documentation, use init_empty to set them.
7270 2 : // This is what `create_test_timeline` does, by the way.
7271 2 : let mut modification = tline.begin_modification(initdb_lsn);
7272 2 : modification
7273 2 : .init_empty_test_timeline()
7274 2 : .context("init_empty_test_timeline")?;
7275 2 : modification
7276 2 : .commit(&ctx)
7277 2 : .await
7278 2 : .context("commit init_empty_test_timeline modification")?;
7279 2 :
7280 2 : // Do the flush. The flush code will check the expectations that we set above.
7281 2 : tline.freeze_and_flush().await?;
7282 2 :
7283 2 : // assert freeze_and_flush exercised the initdb optimization
7284 2 : {
7285 2 : let state = tline.flush_loop_state.lock().unwrap();
7286 2 : let timeline::FlushLoopState::Running {
7287 2 : expect_initdb_optimization,
7288 2 : initdb_optimization_count,
7289 2 : } = *state
7290 2 : else {
7291 2 : panic!("unexpected state: {:?}", *state);
7292 2 : };
7293 2 : assert!(expect_initdb_optimization);
7294 2 : assert!(initdb_optimization_count > 0);
7295 2 : }
7296 2 : Ok(())
7297 2 : }
7298 :
7299 : #[tokio::test]
7300 2 : async fn test_create_guard_crash() -> anyhow::Result<()> {
7301 2 : let name = "test_create_guard_crash";
7302 2 : let harness = TenantHarness::create(name).await?;
7303 2 : {
7304 2 : let (tenant, ctx) = harness.load().await;
7305 2 : let tline = tenant
7306 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
7307 2 : .await?;
7308 2 : // Leave the timeline ID in [`Tenant::timelines_creating`] to exclude attempting to create it again
7309 2 : let raw_tline = tline.raw_timeline().unwrap();
7310 2 : raw_tline
7311 2 : .shutdown(super::timeline::ShutdownMode::Hard)
7312 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))
7313 2 : .await;
7314 2 : std::mem::forget(tline);
7315 2 : }
7316 2 :
7317 2 : let (tenant, _) = harness.load().await;
7318 2 : match tenant.get_timeline(TIMELINE_ID, false) {
7319 2 : Ok(_) => panic!("timeline should've been removed during load"),
7320 2 : Err(e) => {
7321 2 : assert_eq!(
7322 2 : e,
7323 2 : GetTimelineError::NotFound {
7324 2 : tenant_id: tenant.tenant_shard_id,
7325 2 : timeline_id: TIMELINE_ID,
7326 2 : }
7327 2 : )
7328 2 : }
7329 2 : }
7330 2 :
7331 2 : assert!(!harness
7332 2 : .conf
7333 2 : .timeline_path(&tenant.tenant_shard_id, &TIMELINE_ID)
7334 2 : .exists());
7335 2 :
7336 2 : Ok(())
7337 2 : }
7338 :
7339 : #[tokio::test]
7340 2 : async fn test_read_at_max_lsn() -> anyhow::Result<()> {
7341 2 : let names_algorithms = [
7342 2 : ("test_read_at_max_lsn_legacy", CompactionAlgorithm::Legacy),
7343 2 : ("test_read_at_max_lsn_tiered", CompactionAlgorithm::Tiered),
7344 2 : ];
7345 6 : for (name, algorithm) in names_algorithms {
7346 4 : test_read_at_max_lsn_algorithm(name, algorithm).await?;
7347 2 : }
7348 2 : Ok(())
7349 2 : }
7350 :
7351 4 : async fn test_read_at_max_lsn_algorithm(
7352 4 : name: &'static str,
7353 4 : compaction_algorithm: CompactionAlgorithm,
7354 4 : ) -> anyhow::Result<()> {
7355 4 : let mut harness = TenantHarness::create(name).await?;
7356 4 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
7357 4 : kind: compaction_algorithm,
7358 4 : };
7359 4 : let (tenant, ctx) = harness.load().await;
7360 4 : let tline = tenant
7361 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
7362 4 : .await?;
7363 :
7364 4 : let lsn = Lsn(0x10);
7365 4 : let compact = false;
7366 4 : bulk_insert_maybe_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000, compact).await?;
7367 :
7368 4 : let test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7369 4 : let read_lsn = Lsn(u64::MAX - 1);
7370 :
7371 4 : let result = tline.get(test_key, read_lsn, &ctx).await;
7372 4 : assert!(result.is_ok(), "result is not Ok: {}", result.unwrap_err());
7373 :
7374 4 : Ok(())
7375 4 : }
7376 :
7377 : #[tokio::test]
7378 2 : async fn test_metadata_scan() -> anyhow::Result<()> {
7379 2 : let harness = TenantHarness::create("test_metadata_scan").await?;
7380 2 : let (tenant, ctx) = harness.load().await;
7381 2 : let tline = tenant
7382 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7383 2 : .await?;
7384 2 :
7385 2 : const NUM_KEYS: usize = 1000;
7386 2 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
7387 2 :
7388 2 : let cancel = CancellationToken::new();
7389 2 :
7390 2 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7391 2 : base_key.field1 = AUX_KEY_PREFIX;
7392 2 : let mut test_key = base_key;
7393 2 :
7394 2 : // Track when each page was last modified. Used to assert that
7395 2 : // a read sees the latest page version.
7396 2 : let mut updated = [Lsn(0); NUM_KEYS];
7397 2 :
7398 2 : let mut lsn = Lsn(0x10);
7399 2 : #[allow(clippy::needless_range_loop)]
7400 2002 : for blknum in 0..NUM_KEYS {
7401 2000 : lsn = Lsn(lsn.0 + 0x10);
7402 2000 : test_key.field6 = (blknum * STEP) as u32;
7403 2000 : let mut writer = tline.writer().await;
7404 2000 : writer
7405 2000 : .put(
7406 2000 : test_key,
7407 2000 : lsn,
7408 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7409 2000 : &ctx,
7410 2000 : )
7411 2000 : .await?;
7412 2000 : writer.finish_write(lsn);
7413 2000 : updated[blknum] = lsn;
7414 2000 : drop(writer);
7415 2 : }
7416 2 :
7417 2 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
7418 2 :
7419 24 : for iter in 0..=10 {
7420 2 : // Read all the blocks
7421 22000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7422 22000 : test_key.field6 = (blknum * STEP) as u32;
7423 22000 : assert_eq!(
7424 22000 : tline.get(test_key, lsn, &ctx).await?,
7425 22000 : test_img(&format!("{} at {}", blknum, last_lsn))
7426 2 : );
7427 2 : }
7428 2 :
7429 22 : let mut cnt = 0;
7430 22000 : for (key, value) in tline
7431 22 : .get_vectored_impl(
7432 22 : keyspace.clone(),
7433 22 : lsn,
7434 22 : &mut ValuesReconstructState::default(),
7435 22 : &ctx,
7436 22 : )
7437 22 : .await?
7438 2 : {
7439 22000 : let blknum = key.field6 as usize;
7440 22000 : let value = value?;
7441 22000 : assert!(blknum % STEP == 0);
7442 22000 : let blknum = blknum / STEP;
7443 22000 : assert_eq!(
7444 22000 : value,
7445 22000 : test_img(&format!("{} at {}", blknum, updated[blknum]))
7446 22000 : );
7447 22000 : cnt += 1;
7448 2 : }
7449 2 :
7450 22 : assert_eq!(cnt, NUM_KEYS);
7451 2 :
7452 22022 : for _ in 0..NUM_KEYS {
7453 22000 : lsn = Lsn(lsn.0 + 0x10);
7454 22000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7455 22000 : test_key.field6 = (blknum * STEP) as u32;
7456 22000 : let mut writer = tline.writer().await;
7457 22000 : writer
7458 22000 : .put(
7459 22000 : test_key,
7460 22000 : lsn,
7461 22000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7462 22000 : &ctx,
7463 22000 : )
7464 22000 : .await?;
7465 22000 : writer.finish_write(lsn);
7466 22000 : drop(writer);
7467 22000 : updated[blknum] = lsn;
7468 2 : }
7469 2 :
7470 2 : // Perform two cycles of flush, compact, and GC
7471 66 : for round in 0..2 {
7472 44 : tline.freeze_and_flush().await?;
7473 44 : tline
7474 44 : .compact(
7475 44 : &cancel,
7476 44 : if iter % 5 == 0 && round == 0 {
7477 6 : let mut flags = EnumSet::new();
7478 6 : flags.insert(CompactFlags::ForceImageLayerCreation);
7479 6 : flags.insert(CompactFlags::ForceRepartition);
7480 6 : flags
7481 2 : } else {
7482 38 : EnumSet::empty()
7483 2 : },
7484 44 : &ctx,
7485 44 : )
7486 44 : .await?;
7487 44 : tenant
7488 44 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7489 44 : .await?;
7490 2 : }
7491 2 : }
7492 2 :
7493 2 : Ok(())
7494 2 : }
7495 :
7496 : #[tokio::test]
7497 2 : async fn test_metadata_compaction_trigger() -> anyhow::Result<()> {
7498 2 : let harness = TenantHarness::create("test_metadata_compaction_trigger").await?;
7499 2 : let (tenant, ctx) = harness.load().await;
7500 2 : let tline = tenant
7501 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7502 2 : .await?;
7503 2 :
7504 2 : let cancel = CancellationToken::new();
7505 2 :
7506 2 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7507 2 : base_key.field1 = AUX_KEY_PREFIX;
7508 2 : let test_key = base_key;
7509 2 : let mut lsn = Lsn(0x10);
7510 2 :
7511 42 : for _ in 0..20 {
7512 40 : lsn = Lsn(lsn.0 + 0x10);
7513 40 : let mut writer = tline.writer().await;
7514 40 : writer
7515 40 : .put(
7516 40 : test_key,
7517 40 : lsn,
7518 40 : &Value::Image(test_img(&format!("{} at {}", 0, lsn))),
7519 40 : &ctx,
7520 40 : )
7521 40 : .await?;
7522 40 : writer.finish_write(lsn);
7523 40 : drop(writer);
7524 40 : tline.freeze_and_flush().await?; // force create a delta layer
7525 2 : }
7526 2 :
7527 2 : let before_num_l0_delta_files =
7528 2 : tline.layers.read().await.layer_map()?.level0_deltas().len();
7529 2 :
7530 2 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
7531 2 :
7532 2 : let after_num_l0_delta_files = tline.layers.read().await.layer_map()?.level0_deltas().len();
7533 2 :
7534 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}");
7535 2 :
7536 2 : assert_eq!(
7537 2 : tline.get(test_key, lsn, &ctx).await?,
7538 2 : test_img(&format!("{} at {}", 0, lsn))
7539 2 : );
7540 2 :
7541 2 : Ok(())
7542 2 : }
7543 :
7544 : #[tokio::test]
7545 2 : async fn test_aux_file_e2e() {
7546 2 : let harness = TenantHarness::create("test_aux_file_e2e").await.unwrap();
7547 2 :
7548 2 : let (tenant, ctx) = harness.load().await;
7549 2 :
7550 2 : let mut lsn = Lsn(0x08);
7551 2 :
7552 2 : let tline: Arc<Timeline> = tenant
7553 2 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
7554 2 : .await
7555 2 : .unwrap();
7556 2 :
7557 2 : {
7558 2 : lsn += 8;
7559 2 : let mut modification = tline.begin_modification(lsn);
7560 2 : modification
7561 2 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
7562 2 : .await
7563 2 : .unwrap();
7564 2 : modification.commit(&ctx).await.unwrap();
7565 2 : }
7566 2 :
7567 2 : // we can read everything from the storage
7568 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
7569 2 : assert_eq!(
7570 2 : files.get("pg_logical/mappings/test1"),
7571 2 : Some(&bytes::Bytes::from_static(b"first"))
7572 2 : );
7573 2 :
7574 2 : {
7575 2 : lsn += 8;
7576 2 : let mut modification = tline.begin_modification(lsn);
7577 2 : modification
7578 2 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
7579 2 : .await
7580 2 : .unwrap();
7581 2 : modification.commit(&ctx).await.unwrap();
7582 2 : }
7583 2 :
7584 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
7585 2 : assert_eq!(
7586 2 : files.get("pg_logical/mappings/test2"),
7587 2 : Some(&bytes::Bytes::from_static(b"second"))
7588 2 : );
7589 2 :
7590 2 : let child = tenant
7591 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(lsn), &ctx)
7592 2 : .await
7593 2 : .unwrap();
7594 2 :
7595 2 : let files = child.list_aux_files(lsn, &ctx).await.unwrap();
7596 2 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
7597 2 : assert_eq!(files.get("pg_logical/mappings/test2"), None);
7598 2 : }
7599 :
7600 : #[tokio::test]
7601 2 : async fn test_metadata_image_creation() -> anyhow::Result<()> {
7602 2 : let harness = TenantHarness::create("test_metadata_image_creation").await?;
7603 2 : let (tenant, ctx) = harness.load().await;
7604 2 : let tline = tenant
7605 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7606 2 : .await?;
7607 2 :
7608 2 : const NUM_KEYS: usize = 1000;
7609 2 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
7610 2 :
7611 2 : let cancel = CancellationToken::new();
7612 2 :
7613 2 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7614 2 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
7615 2 : let mut test_key = base_key;
7616 2 : let mut lsn = Lsn(0x10);
7617 2 :
7618 8 : async fn scan_with_statistics(
7619 8 : tline: &Timeline,
7620 8 : keyspace: &KeySpace,
7621 8 : lsn: Lsn,
7622 8 : ctx: &RequestContext,
7623 8 : ) -> anyhow::Result<(BTreeMap<Key, Result<Bytes, PageReconstructError>>, usize)> {
7624 8 : let mut reconstruct_state = ValuesReconstructState::default();
7625 8 : let res = tline
7626 8 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
7627 8 : .await?;
7628 8 : Ok((res, reconstruct_state.get_delta_layers_visited() as usize))
7629 8 : }
7630 2 :
7631 2 : #[allow(clippy::needless_range_loop)]
7632 2002 : for blknum in 0..NUM_KEYS {
7633 2000 : lsn = Lsn(lsn.0 + 0x10);
7634 2000 : test_key.field6 = (blknum * STEP) as u32;
7635 2000 : let mut writer = tline.writer().await;
7636 2000 : writer
7637 2000 : .put(
7638 2000 : test_key,
7639 2000 : lsn,
7640 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7641 2000 : &ctx,
7642 2000 : )
7643 2000 : .await?;
7644 2000 : writer.finish_write(lsn);
7645 2000 : drop(writer);
7646 2 : }
7647 2 :
7648 2 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
7649 2 :
7650 22 : for iter in 1..=10 {
7651 20020 : for _ in 0..NUM_KEYS {
7652 20000 : lsn = Lsn(lsn.0 + 0x10);
7653 20000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7654 20000 : test_key.field6 = (blknum * STEP) as u32;
7655 20000 : let mut writer = tline.writer().await;
7656 20000 : writer
7657 20000 : .put(
7658 20000 : test_key,
7659 20000 : lsn,
7660 20000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7661 20000 : &ctx,
7662 20000 : )
7663 20000 : .await?;
7664 20000 : writer.finish_write(lsn);
7665 20000 : drop(writer);
7666 2 : }
7667 2 :
7668 20 : tline.freeze_and_flush().await?;
7669 2 :
7670 20 : if iter % 5 == 0 {
7671 4 : let (_, before_delta_file_accessed) =
7672 4 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
7673 4 : tline
7674 4 : .compact(
7675 4 : &cancel,
7676 4 : {
7677 4 : let mut flags = EnumSet::new();
7678 4 : flags.insert(CompactFlags::ForceImageLayerCreation);
7679 4 : flags.insert(CompactFlags::ForceRepartition);
7680 4 : flags
7681 4 : },
7682 4 : &ctx,
7683 4 : )
7684 4 : .await?;
7685 4 : let (_, after_delta_file_accessed) =
7686 4 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
7687 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}");
7688 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.
7689 4 : assert!(
7690 4 : after_delta_file_accessed <= 2,
7691 2 : "after_delta_file_accessed={after_delta_file_accessed}"
7692 2 : );
7693 16 : }
7694 2 : }
7695 2 :
7696 2 : Ok(())
7697 2 : }
7698 :
7699 : #[tokio::test]
7700 2 : async fn test_vectored_missing_data_key_reads() -> anyhow::Result<()> {
7701 2 : let harness = TenantHarness::create("test_vectored_missing_data_key_reads").await?;
7702 2 : let (tenant, ctx) = harness.load().await;
7703 2 :
7704 2 : let base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7705 2 : let base_key_child = Key::from_hex("000000000033333333444444445500000001").unwrap();
7706 2 : let base_key_nonexist = Key::from_hex("000000000033333333444444445500000002").unwrap();
7707 2 :
7708 2 : let tline = tenant
7709 2 : .create_test_timeline_with_layers(
7710 2 : TIMELINE_ID,
7711 2 : Lsn(0x10),
7712 2 : DEFAULT_PG_VERSION,
7713 2 : &ctx,
7714 2 : Vec::new(), // delta layers
7715 2 : vec![(Lsn(0x20), vec![(base_key, test_img("data key 1"))])], // image layers
7716 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
7717 2 : )
7718 2 : .await?;
7719 2 : tline.add_extra_test_dense_keyspace(KeySpace::single(base_key..(base_key_nonexist.next())));
7720 2 :
7721 2 : let child = tenant
7722 2 : .branch_timeline_test_with_layers(
7723 2 : &tline,
7724 2 : NEW_TIMELINE_ID,
7725 2 : Some(Lsn(0x20)),
7726 2 : &ctx,
7727 2 : Vec::new(), // delta layers
7728 2 : vec![(Lsn(0x30), vec![(base_key_child, test_img("data key 2"))])], // image layers
7729 2 : Lsn(0x30),
7730 2 : )
7731 2 : .await
7732 2 : .unwrap();
7733 2 :
7734 2 : let lsn = Lsn(0x30);
7735 2 :
7736 2 : // test vectored get on parent timeline
7737 2 : assert_eq!(
7738 2 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
7739 2 : Some(test_img("data key 1"))
7740 2 : );
7741 2 : assert!(get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx)
7742 2 : .await
7743 2 : .unwrap_err()
7744 2 : .is_missing_key_error());
7745 2 : assert!(
7746 2 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx)
7747 2 : .await
7748 2 : .unwrap_err()
7749 2 : .is_missing_key_error()
7750 2 : );
7751 2 :
7752 2 : // test vectored get on child timeline
7753 2 : assert_eq!(
7754 2 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
7755 2 : Some(test_img("data key 1"))
7756 2 : );
7757 2 : assert_eq!(
7758 2 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
7759 2 : Some(test_img("data key 2"))
7760 2 : );
7761 2 : assert!(
7762 2 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx)
7763 2 : .await
7764 2 : .unwrap_err()
7765 2 : .is_missing_key_error()
7766 2 : );
7767 2 :
7768 2 : Ok(())
7769 2 : }
7770 :
7771 : #[tokio::test]
7772 2 : async fn test_vectored_missing_metadata_key_reads() -> anyhow::Result<()> {
7773 2 : let harness = TenantHarness::create("test_vectored_missing_metadata_key_reads").await?;
7774 2 : let (tenant, ctx) = harness.load().await;
7775 2 :
7776 2 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7777 2 : let base_key_child = Key::from_hex("620000000033333333444444445500000001").unwrap();
7778 2 : let base_key_nonexist = Key::from_hex("620000000033333333444444445500000002").unwrap();
7779 2 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
7780 2 :
7781 2 : let tline = tenant
7782 2 : .create_test_timeline_with_layers(
7783 2 : TIMELINE_ID,
7784 2 : Lsn(0x10),
7785 2 : DEFAULT_PG_VERSION,
7786 2 : &ctx,
7787 2 : Vec::new(), // delta layers
7788 2 : vec![(Lsn(0x20), vec![(base_key, test_img("metadata key 1"))])], // image layers
7789 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
7790 2 : )
7791 2 : .await?;
7792 2 :
7793 2 : let child = tenant
7794 2 : .branch_timeline_test_with_layers(
7795 2 : &tline,
7796 2 : NEW_TIMELINE_ID,
7797 2 : Some(Lsn(0x20)),
7798 2 : &ctx,
7799 2 : Vec::new(), // delta layers
7800 2 : vec![(
7801 2 : Lsn(0x30),
7802 2 : vec![(base_key_child, test_img("metadata key 2"))],
7803 2 : )], // image layers
7804 2 : Lsn(0x30),
7805 2 : )
7806 2 : .await
7807 2 : .unwrap();
7808 2 :
7809 2 : let lsn = Lsn(0x30);
7810 2 :
7811 2 : // test vectored get on parent timeline
7812 2 : assert_eq!(
7813 2 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
7814 2 : Some(test_img("metadata key 1"))
7815 2 : );
7816 2 : assert_eq!(
7817 2 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx).await?,
7818 2 : None
7819 2 : );
7820 2 : assert_eq!(
7821 2 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx).await?,
7822 2 : None
7823 2 : );
7824 2 :
7825 2 : // test vectored get on child timeline
7826 2 : assert_eq!(
7827 2 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
7828 2 : None
7829 2 : );
7830 2 : assert_eq!(
7831 2 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
7832 2 : Some(test_img("metadata key 2"))
7833 2 : );
7834 2 : assert_eq!(
7835 2 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx).await?,
7836 2 : None
7837 2 : );
7838 2 :
7839 2 : Ok(())
7840 2 : }
7841 :
7842 36 : async fn get_vectored_impl_wrapper(
7843 36 : tline: &Arc<Timeline>,
7844 36 : key: Key,
7845 36 : lsn: Lsn,
7846 36 : ctx: &RequestContext,
7847 36 : ) -> Result<Option<Bytes>, GetVectoredError> {
7848 36 : let mut reconstruct_state = ValuesReconstructState::new();
7849 36 : let mut res = tline
7850 36 : .get_vectored_impl(
7851 36 : KeySpace::single(key..key.next()),
7852 36 : lsn,
7853 36 : &mut reconstruct_state,
7854 36 : ctx,
7855 36 : )
7856 36 : .await?;
7857 30 : Ok(res.pop_last().map(|(k, v)| {
7858 18 : assert_eq!(k, key);
7859 18 : v.unwrap()
7860 30 : }))
7861 36 : }
7862 :
7863 : #[tokio::test]
7864 2 : async fn test_metadata_tombstone_reads() -> anyhow::Result<()> {
7865 2 : let harness = TenantHarness::create("test_metadata_tombstone_reads").await?;
7866 2 : let (tenant, ctx) = harness.load().await;
7867 2 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
7868 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
7869 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
7870 2 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
7871 2 :
7872 2 : // We emulate the situation that the compaction algorithm creates an image layer that removes the tombstones
7873 2 : // Lsn 0x30 key0, key3, no key1+key2
7874 2 : // Lsn 0x20 key1+key2 tomestones
7875 2 : // Lsn 0x10 key1 in image, key2 in delta
7876 2 : let tline = tenant
7877 2 : .create_test_timeline_with_layers(
7878 2 : TIMELINE_ID,
7879 2 : Lsn(0x10),
7880 2 : DEFAULT_PG_VERSION,
7881 2 : &ctx,
7882 2 : // delta layers
7883 2 : vec![
7884 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7885 2 : Lsn(0x10)..Lsn(0x20),
7886 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
7887 2 : ),
7888 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7889 2 : Lsn(0x20)..Lsn(0x30),
7890 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
7891 2 : ),
7892 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7893 2 : Lsn(0x20)..Lsn(0x30),
7894 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
7895 2 : ),
7896 2 : ],
7897 2 : // image layers
7898 2 : vec![
7899 2 : (Lsn(0x10), vec![(key1, test_img("metadata key 1"))]),
7900 2 : (
7901 2 : Lsn(0x30),
7902 2 : vec![
7903 2 : (key0, test_img("metadata key 0")),
7904 2 : (key3, test_img("metadata key 3")),
7905 2 : ],
7906 2 : ),
7907 2 : ],
7908 2 : Lsn(0x30),
7909 2 : )
7910 2 : .await?;
7911 2 :
7912 2 : let lsn = Lsn(0x30);
7913 2 : let old_lsn = Lsn(0x20);
7914 2 :
7915 2 : assert_eq!(
7916 2 : get_vectored_impl_wrapper(&tline, key0, lsn, &ctx).await?,
7917 2 : Some(test_img("metadata key 0"))
7918 2 : );
7919 2 : assert_eq!(
7920 2 : get_vectored_impl_wrapper(&tline, key1, lsn, &ctx).await?,
7921 2 : None,
7922 2 : );
7923 2 : assert_eq!(
7924 2 : get_vectored_impl_wrapper(&tline, key2, lsn, &ctx).await?,
7925 2 : None,
7926 2 : );
7927 2 : assert_eq!(
7928 2 : get_vectored_impl_wrapper(&tline, key1, old_lsn, &ctx).await?,
7929 2 : Some(Bytes::new()),
7930 2 : );
7931 2 : assert_eq!(
7932 2 : get_vectored_impl_wrapper(&tline, key2, old_lsn, &ctx).await?,
7933 2 : Some(Bytes::new()),
7934 2 : );
7935 2 : assert_eq!(
7936 2 : get_vectored_impl_wrapper(&tline, key3, lsn, &ctx).await?,
7937 2 : Some(test_img("metadata key 3"))
7938 2 : );
7939 2 :
7940 2 : Ok(())
7941 2 : }
7942 :
7943 : #[tokio::test]
7944 2 : async fn test_metadata_tombstone_image_creation() {
7945 2 : let harness = TenantHarness::create("test_metadata_tombstone_image_creation")
7946 2 : .await
7947 2 : .unwrap();
7948 2 : let (tenant, ctx) = harness.load().await;
7949 2 :
7950 2 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
7951 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
7952 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
7953 2 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
7954 2 :
7955 2 : let tline = tenant
7956 2 : .create_test_timeline_with_layers(
7957 2 : TIMELINE_ID,
7958 2 : Lsn(0x10),
7959 2 : DEFAULT_PG_VERSION,
7960 2 : &ctx,
7961 2 : // delta layers
7962 2 : vec![
7963 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7964 2 : Lsn(0x10)..Lsn(0x20),
7965 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
7966 2 : ),
7967 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7968 2 : Lsn(0x20)..Lsn(0x30),
7969 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
7970 2 : ),
7971 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7972 2 : Lsn(0x20)..Lsn(0x30),
7973 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
7974 2 : ),
7975 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7976 2 : Lsn(0x30)..Lsn(0x40),
7977 2 : vec![
7978 2 : (key0, Lsn(0x30), Value::Image(test_img("metadata key 0"))),
7979 2 : (key3, Lsn(0x30), Value::Image(test_img("metadata key 3"))),
7980 2 : ],
7981 2 : ),
7982 2 : ],
7983 2 : // image layers
7984 2 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
7985 2 : Lsn(0x40),
7986 2 : )
7987 2 : .await
7988 2 : .unwrap();
7989 2 :
7990 2 : let cancel = CancellationToken::new();
7991 2 :
7992 2 : tline
7993 2 : .compact(
7994 2 : &cancel,
7995 2 : {
7996 2 : let mut flags = EnumSet::new();
7997 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
7998 2 : flags.insert(CompactFlags::ForceRepartition);
7999 2 : flags
8000 2 : },
8001 2 : &ctx,
8002 2 : )
8003 2 : .await
8004 2 : .unwrap();
8005 2 :
8006 2 : // Image layers are created at last_record_lsn
8007 2 : let images = tline
8008 2 : .inspect_image_layers(Lsn(0x40), &ctx)
8009 2 : .await
8010 2 : .unwrap()
8011 2 : .into_iter()
8012 18 : .filter(|(k, _)| k.is_metadata_key())
8013 2 : .collect::<Vec<_>>();
8014 2 : assert_eq!(images.len(), 2); // the image layer should only contain two existing keys, tombstones should be removed.
8015 2 : }
8016 :
8017 : #[tokio::test]
8018 2 : async fn test_metadata_tombstone_empty_image_creation() {
8019 2 : let harness = TenantHarness::create("test_metadata_tombstone_empty_image_creation")
8020 2 : .await
8021 2 : .unwrap();
8022 2 : let (tenant, ctx) = harness.load().await;
8023 2 :
8024 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8025 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8026 2 :
8027 2 : let tline = tenant
8028 2 : .create_test_timeline_with_layers(
8029 2 : TIMELINE_ID,
8030 2 : Lsn(0x10),
8031 2 : DEFAULT_PG_VERSION,
8032 2 : &ctx,
8033 2 : // delta layers
8034 2 : vec![
8035 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8036 2 : Lsn(0x10)..Lsn(0x20),
8037 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8038 2 : ),
8039 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8040 2 : Lsn(0x20)..Lsn(0x30),
8041 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8042 2 : ),
8043 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8044 2 : Lsn(0x20)..Lsn(0x30),
8045 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8046 2 : ),
8047 2 : ],
8048 2 : // image layers
8049 2 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
8050 2 : Lsn(0x30),
8051 2 : )
8052 2 : .await
8053 2 : .unwrap();
8054 2 :
8055 2 : let cancel = CancellationToken::new();
8056 2 :
8057 2 : tline
8058 2 : .compact(
8059 2 : &cancel,
8060 2 : {
8061 2 : let mut flags = EnumSet::new();
8062 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
8063 2 : flags.insert(CompactFlags::ForceRepartition);
8064 2 : flags
8065 2 : },
8066 2 : &ctx,
8067 2 : )
8068 2 : .await
8069 2 : .unwrap();
8070 2 :
8071 2 : // Image layers are created at last_record_lsn
8072 2 : let images = tline
8073 2 : .inspect_image_layers(Lsn(0x30), &ctx)
8074 2 : .await
8075 2 : .unwrap()
8076 2 : .into_iter()
8077 14 : .filter(|(k, _)| k.is_metadata_key())
8078 2 : .collect::<Vec<_>>();
8079 2 : assert_eq!(images.len(), 0); // the image layer should not contain tombstones, or it is not created
8080 2 : }
8081 :
8082 : #[tokio::test]
8083 2 : async fn test_simple_bottom_most_compaction_images() -> anyhow::Result<()> {
8084 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_images").await?;
8085 2 : let (tenant, ctx) = harness.load().await;
8086 2 :
8087 102 : fn get_key(id: u32) -> Key {
8088 102 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8089 102 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8090 102 : key.field6 = id;
8091 102 : key
8092 102 : }
8093 2 :
8094 2 : // We create
8095 2 : // - one bottom-most image layer,
8096 2 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
8097 2 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
8098 2 : // - a delta layer D3 above the horizon.
8099 2 : //
8100 2 : // | D3 |
8101 2 : // | D1 |
8102 2 : // -| |-- gc horizon -----------------
8103 2 : // | | | D2 |
8104 2 : // --------- img layer ------------------
8105 2 : //
8106 2 : // What we should expact from this compaction is:
8107 2 : // | D3 |
8108 2 : // | Part of D1 |
8109 2 : // --------- img layer with D1+D2 at GC horizon------------------
8110 2 :
8111 2 : // img layer at 0x10
8112 2 : let img_layer = (0..10)
8113 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8114 2 : .collect_vec();
8115 2 :
8116 2 : let delta1 = vec![
8117 2 : (
8118 2 : get_key(1),
8119 2 : Lsn(0x20),
8120 2 : Value::Image(Bytes::from("value 1@0x20")),
8121 2 : ),
8122 2 : (
8123 2 : get_key(2),
8124 2 : Lsn(0x30),
8125 2 : Value::Image(Bytes::from("value 2@0x30")),
8126 2 : ),
8127 2 : (
8128 2 : get_key(3),
8129 2 : Lsn(0x40),
8130 2 : Value::Image(Bytes::from("value 3@0x40")),
8131 2 : ),
8132 2 : ];
8133 2 : let delta2 = vec![
8134 2 : (
8135 2 : get_key(5),
8136 2 : Lsn(0x20),
8137 2 : Value::Image(Bytes::from("value 5@0x20")),
8138 2 : ),
8139 2 : (
8140 2 : get_key(6),
8141 2 : Lsn(0x20),
8142 2 : Value::Image(Bytes::from("value 6@0x20")),
8143 2 : ),
8144 2 : ];
8145 2 : let delta3 = vec![
8146 2 : (
8147 2 : get_key(8),
8148 2 : Lsn(0x48),
8149 2 : Value::Image(Bytes::from("value 8@0x48")),
8150 2 : ),
8151 2 : (
8152 2 : get_key(9),
8153 2 : Lsn(0x48),
8154 2 : Value::Image(Bytes::from("value 9@0x48")),
8155 2 : ),
8156 2 : ];
8157 2 :
8158 2 : let tline = tenant
8159 2 : .create_test_timeline_with_layers(
8160 2 : TIMELINE_ID,
8161 2 : Lsn(0x10),
8162 2 : DEFAULT_PG_VERSION,
8163 2 : &ctx,
8164 2 : vec![
8165 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
8166 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
8167 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
8168 2 : ], // delta layers
8169 2 : vec![(Lsn(0x10), img_layer)], // image layers
8170 2 : Lsn(0x50),
8171 2 : )
8172 2 : .await?;
8173 2 : {
8174 2 : tline
8175 2 : .latest_gc_cutoff_lsn
8176 2 : .lock_for_write()
8177 2 : .store_and_unlock(Lsn(0x30))
8178 2 : .wait()
8179 2 : .await;
8180 2 : // Update GC info
8181 2 : let mut guard = tline.gc_info.write().unwrap();
8182 2 : guard.cutoffs.time = Lsn(0x30);
8183 2 : guard.cutoffs.space = Lsn(0x30);
8184 2 : }
8185 2 :
8186 2 : let expected_result = [
8187 2 : Bytes::from_static(b"value 0@0x10"),
8188 2 : Bytes::from_static(b"value 1@0x20"),
8189 2 : Bytes::from_static(b"value 2@0x30"),
8190 2 : Bytes::from_static(b"value 3@0x40"),
8191 2 : Bytes::from_static(b"value 4@0x10"),
8192 2 : Bytes::from_static(b"value 5@0x20"),
8193 2 : Bytes::from_static(b"value 6@0x20"),
8194 2 : Bytes::from_static(b"value 7@0x10"),
8195 2 : Bytes::from_static(b"value 8@0x48"),
8196 2 : Bytes::from_static(b"value 9@0x48"),
8197 2 : ];
8198 2 :
8199 20 : for (idx, expected) in expected_result.iter().enumerate() {
8200 20 : assert_eq!(
8201 20 : tline
8202 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8203 20 : .await
8204 20 : .unwrap(),
8205 2 : expected
8206 2 : );
8207 2 : }
8208 2 :
8209 2 : let cancel = CancellationToken::new();
8210 2 : tline
8211 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8212 2 : .await
8213 2 : .unwrap();
8214 2 :
8215 20 : for (idx, expected) in expected_result.iter().enumerate() {
8216 20 : assert_eq!(
8217 20 : tline
8218 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8219 20 : .await
8220 20 : .unwrap(),
8221 2 : expected
8222 2 : );
8223 2 : }
8224 2 :
8225 2 : // Check if the image layer at the GC horizon contains exactly what we want
8226 2 : let image_at_gc_horizon = tline
8227 2 : .inspect_image_layers(Lsn(0x30), &ctx)
8228 2 : .await
8229 2 : .unwrap()
8230 2 : .into_iter()
8231 34 : .filter(|(k, _)| k.is_metadata_key())
8232 2 : .collect::<Vec<_>>();
8233 2 :
8234 2 : assert_eq!(image_at_gc_horizon.len(), 10);
8235 2 : let expected_result = [
8236 2 : Bytes::from_static(b"value 0@0x10"),
8237 2 : Bytes::from_static(b"value 1@0x20"),
8238 2 : Bytes::from_static(b"value 2@0x30"),
8239 2 : Bytes::from_static(b"value 3@0x10"),
8240 2 : Bytes::from_static(b"value 4@0x10"),
8241 2 : Bytes::from_static(b"value 5@0x20"),
8242 2 : Bytes::from_static(b"value 6@0x20"),
8243 2 : Bytes::from_static(b"value 7@0x10"),
8244 2 : Bytes::from_static(b"value 8@0x10"),
8245 2 : Bytes::from_static(b"value 9@0x10"),
8246 2 : ];
8247 22 : for idx in 0..10 {
8248 20 : assert_eq!(
8249 20 : image_at_gc_horizon[idx],
8250 20 : (get_key(idx as u32), expected_result[idx].clone())
8251 20 : );
8252 2 : }
8253 2 :
8254 2 : // Check if old layers are removed / new layers have the expected LSN
8255 2 : let all_layers = inspect_and_sort(&tline, None).await;
8256 2 : assert_eq!(
8257 2 : all_layers,
8258 2 : vec![
8259 2 : // Image layer at GC horizon
8260 2 : PersistentLayerKey {
8261 2 : key_range: Key::MIN..Key::MAX,
8262 2 : lsn_range: Lsn(0x30)..Lsn(0x31),
8263 2 : is_delta: false
8264 2 : },
8265 2 : // The delta layer below the horizon
8266 2 : PersistentLayerKey {
8267 2 : key_range: get_key(3)..get_key(4),
8268 2 : lsn_range: Lsn(0x30)..Lsn(0x48),
8269 2 : is_delta: true
8270 2 : },
8271 2 : // The delta3 layer that should not be picked for the compaction
8272 2 : PersistentLayerKey {
8273 2 : key_range: get_key(8)..get_key(10),
8274 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
8275 2 : is_delta: true
8276 2 : }
8277 2 : ]
8278 2 : );
8279 2 :
8280 2 : // increase GC horizon and compact again
8281 2 : {
8282 2 : tline
8283 2 : .latest_gc_cutoff_lsn
8284 2 : .lock_for_write()
8285 2 : .store_and_unlock(Lsn(0x40))
8286 2 : .wait()
8287 2 : .await;
8288 2 : // Update GC info
8289 2 : let mut guard = tline.gc_info.write().unwrap();
8290 2 : guard.cutoffs.time = Lsn(0x40);
8291 2 : guard.cutoffs.space = Lsn(0x40);
8292 2 : }
8293 2 : tline
8294 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8295 2 : .await
8296 2 : .unwrap();
8297 2 :
8298 2 : Ok(())
8299 2 : }
8300 :
8301 : #[cfg(feature = "testing")]
8302 : #[tokio::test]
8303 2 : async fn test_neon_test_record() -> anyhow::Result<()> {
8304 2 : let harness = TenantHarness::create("test_neon_test_record").await?;
8305 2 : let (tenant, ctx) = harness.load().await;
8306 2 :
8307 24 : fn get_key(id: u32) -> Key {
8308 24 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8309 24 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8310 24 : key.field6 = id;
8311 24 : key
8312 24 : }
8313 2 :
8314 2 : let delta1 = vec![
8315 2 : (
8316 2 : get_key(1),
8317 2 : Lsn(0x20),
8318 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
8319 2 : ),
8320 2 : (
8321 2 : get_key(1),
8322 2 : Lsn(0x30),
8323 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
8324 2 : ),
8325 2 : (get_key(2), Lsn(0x10), Value::Image("0x10".into())),
8326 2 : (
8327 2 : get_key(2),
8328 2 : Lsn(0x20),
8329 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
8330 2 : ),
8331 2 : (
8332 2 : get_key(2),
8333 2 : Lsn(0x30),
8334 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
8335 2 : ),
8336 2 : (get_key(3), Lsn(0x10), Value::Image("0x10".into())),
8337 2 : (
8338 2 : get_key(3),
8339 2 : Lsn(0x20),
8340 2 : Value::WalRecord(NeonWalRecord::wal_clear("c")),
8341 2 : ),
8342 2 : (get_key(4), Lsn(0x10), Value::Image("0x10".into())),
8343 2 : (
8344 2 : get_key(4),
8345 2 : Lsn(0x20),
8346 2 : Value::WalRecord(NeonWalRecord::wal_init("i")),
8347 2 : ),
8348 2 : ];
8349 2 : let image1 = vec![(get_key(1), "0x10".into())];
8350 2 :
8351 2 : let tline = tenant
8352 2 : .create_test_timeline_with_layers(
8353 2 : TIMELINE_ID,
8354 2 : Lsn(0x10),
8355 2 : DEFAULT_PG_VERSION,
8356 2 : &ctx,
8357 2 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
8358 2 : Lsn(0x10)..Lsn(0x40),
8359 2 : delta1,
8360 2 : )], // delta layers
8361 2 : vec![(Lsn(0x10), image1)], // image layers
8362 2 : Lsn(0x50),
8363 2 : )
8364 2 : .await?;
8365 2 :
8366 2 : assert_eq!(
8367 2 : tline.get(get_key(1), Lsn(0x50), &ctx).await?,
8368 2 : Bytes::from_static(b"0x10,0x20,0x30")
8369 2 : );
8370 2 : assert_eq!(
8371 2 : tline.get(get_key(2), Lsn(0x50), &ctx).await?,
8372 2 : Bytes::from_static(b"0x10,0x20,0x30")
8373 2 : );
8374 2 :
8375 2 : // Need to remove the limit of "Neon WAL redo requires base image".
8376 2 :
8377 2 : // assert_eq!(tline.get(get_key(3), Lsn(0x50), &ctx).await?, Bytes::new());
8378 2 : // assert_eq!(tline.get(get_key(4), Lsn(0x50), &ctx).await?, Bytes::new());
8379 2 :
8380 2 : Ok(())
8381 2 : }
8382 :
8383 : #[tokio::test(start_paused = true)]
8384 2 : async fn test_lsn_lease() -> anyhow::Result<()> {
8385 2 : let (tenant, ctx) = TenantHarness::create("test_lsn_lease")
8386 2 : .await
8387 2 : .unwrap()
8388 2 : .load()
8389 2 : .await;
8390 2 : // Advance to the lsn lease deadline so that GC is not blocked by
8391 2 : // initial transition into AttachedSingle.
8392 2 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
8393 2 : tokio::time::resume();
8394 2 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
8395 2 :
8396 2 : let end_lsn = Lsn(0x100);
8397 2 : let image_layers = (0x20..=0x90)
8398 2 : .step_by(0x10)
8399 16 : .map(|n| {
8400 16 : (
8401 16 : Lsn(n),
8402 16 : vec![(key, test_img(&format!("data key at {:x}", n)))],
8403 16 : )
8404 16 : })
8405 2 : .collect();
8406 2 :
8407 2 : let timeline = tenant
8408 2 : .create_test_timeline_with_layers(
8409 2 : TIMELINE_ID,
8410 2 : Lsn(0x10),
8411 2 : DEFAULT_PG_VERSION,
8412 2 : &ctx,
8413 2 : Vec::new(),
8414 2 : image_layers,
8415 2 : end_lsn,
8416 2 : )
8417 2 : .await?;
8418 2 :
8419 2 : let leased_lsns = [0x30, 0x50, 0x70];
8420 2 : let mut leases = Vec::new();
8421 6 : leased_lsns.iter().for_each(|n| {
8422 6 : leases.push(
8423 6 : timeline
8424 6 : .init_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)
8425 6 : .expect("lease request should succeed"),
8426 6 : );
8427 6 : });
8428 2 :
8429 2 : let updated_lease_0 = timeline
8430 2 : .renew_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)
8431 2 : .expect("lease renewal should succeed");
8432 2 : assert_eq!(
8433 2 : updated_lease_0.valid_until, leases[0].valid_until,
8434 2 : " Renewing with shorter lease should not change the lease."
8435 2 : );
8436 2 :
8437 2 : let updated_lease_1 = timeline
8438 2 : .renew_lsn_lease(
8439 2 : Lsn(leased_lsns[1]),
8440 2 : timeline.get_lsn_lease_length() * 2,
8441 2 : &ctx,
8442 2 : )
8443 2 : .expect("lease renewal should succeed");
8444 2 : assert!(
8445 2 : updated_lease_1.valid_until > leases[1].valid_until,
8446 2 : "Renewing with a long lease should renew lease with later expiration time."
8447 2 : );
8448 2 :
8449 2 : // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
8450 2 : info!(
8451 2 : "latest_gc_cutoff_lsn: {}",
8452 0 : *timeline.get_latest_gc_cutoff_lsn()
8453 2 : );
8454 2 : timeline.force_set_disk_consistent_lsn(end_lsn);
8455 2 :
8456 2 : let res = tenant
8457 2 : .gc_iteration(
8458 2 : Some(TIMELINE_ID),
8459 2 : 0,
8460 2 : Duration::ZERO,
8461 2 : &CancellationToken::new(),
8462 2 : &ctx,
8463 2 : )
8464 2 : .await
8465 2 : .unwrap();
8466 2 :
8467 2 : // Keeping everything <= Lsn(0x80) b/c leases:
8468 2 : // 0/10: initdb layer
8469 2 : // (0/20..=0/70).step_by(0x10): image layers added when creating the timeline.
8470 2 : assert_eq!(res.layers_needed_by_leases, 7);
8471 2 : // Keeping 0/90 b/c it is the latest layer.
8472 2 : assert_eq!(res.layers_not_updated, 1);
8473 2 : // Removed 0/80.
8474 2 : assert_eq!(res.layers_removed, 1);
8475 2 :
8476 2 : // Make lease on a already GC-ed LSN.
8477 2 : // 0/80 does not have a valid lease + is below latest_gc_cutoff
8478 2 : assert!(Lsn(0x80) < *timeline.get_latest_gc_cutoff_lsn());
8479 2 : timeline
8480 2 : .init_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx)
8481 2 : .expect_err("lease request on GC-ed LSN should fail");
8482 2 :
8483 2 : // Should still be able to renew a currently valid lease
8484 2 : // Assumption: original lease to is still valid for 0/50.
8485 2 : // (use `Timeline::init_lsn_lease` for testing so it always does validation)
8486 2 : timeline
8487 2 : .init_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)
8488 2 : .expect("lease renewal with validation should succeed");
8489 2 :
8490 2 : Ok(())
8491 2 : }
8492 :
8493 : #[cfg(feature = "testing")]
8494 : #[tokio::test]
8495 2 : async fn test_simple_bottom_most_compaction_deltas_1() -> anyhow::Result<()> {
8496 2 : test_simple_bottom_most_compaction_deltas_helper(
8497 2 : "test_simple_bottom_most_compaction_deltas_1",
8498 2 : false,
8499 2 : )
8500 2 : .await
8501 2 : }
8502 :
8503 : #[cfg(feature = "testing")]
8504 : #[tokio::test]
8505 2 : async fn test_simple_bottom_most_compaction_deltas_2() -> anyhow::Result<()> {
8506 2 : test_simple_bottom_most_compaction_deltas_helper(
8507 2 : "test_simple_bottom_most_compaction_deltas_2",
8508 2 : true,
8509 2 : )
8510 2 : .await
8511 2 : }
8512 :
8513 : #[cfg(feature = "testing")]
8514 4 : async fn test_simple_bottom_most_compaction_deltas_helper(
8515 4 : test_name: &'static str,
8516 4 : use_delta_bottom_layer: bool,
8517 4 : ) -> anyhow::Result<()> {
8518 4 : let harness = TenantHarness::create(test_name).await?;
8519 4 : let (tenant, ctx) = harness.load().await;
8520 :
8521 276 : fn get_key(id: u32) -> Key {
8522 276 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8523 276 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8524 276 : key.field6 = id;
8525 276 : key
8526 276 : }
8527 :
8528 : // We create
8529 : // - one bottom-most image layer,
8530 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
8531 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
8532 : // - a delta layer D3 above the horizon.
8533 : //
8534 : // | D3 |
8535 : // | D1 |
8536 : // -| |-- gc horizon -----------------
8537 : // | | | D2 |
8538 : // --------- img layer ------------------
8539 : //
8540 : // What we should expact from this compaction is:
8541 : // | D3 |
8542 : // | Part of D1 |
8543 : // --------- img layer with D1+D2 at GC horizon------------------
8544 :
8545 : // img layer at 0x10
8546 4 : let img_layer = (0..10)
8547 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8548 4 : .collect_vec();
8549 4 : // or, delta layer at 0x10 if `use_delta_bottom_layer` is true
8550 4 : let delta4 = (0..10)
8551 40 : .map(|id| {
8552 40 : (
8553 40 : get_key(id),
8554 40 : Lsn(0x08),
8555 40 : Value::WalRecord(NeonWalRecord::wal_init(format!("value {id}@0x10"))),
8556 40 : )
8557 40 : })
8558 4 : .collect_vec();
8559 4 :
8560 4 : let delta1 = vec![
8561 4 : (
8562 4 : get_key(1),
8563 4 : Lsn(0x20),
8564 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8565 4 : ),
8566 4 : (
8567 4 : get_key(2),
8568 4 : Lsn(0x30),
8569 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8570 4 : ),
8571 4 : (
8572 4 : get_key(3),
8573 4 : Lsn(0x28),
8574 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8575 4 : ),
8576 4 : (
8577 4 : get_key(3),
8578 4 : Lsn(0x30),
8579 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8580 4 : ),
8581 4 : (
8582 4 : get_key(3),
8583 4 : Lsn(0x40),
8584 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
8585 4 : ),
8586 4 : ];
8587 4 : let delta2 = vec![
8588 4 : (
8589 4 : get_key(5),
8590 4 : Lsn(0x20),
8591 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8592 4 : ),
8593 4 : (
8594 4 : get_key(6),
8595 4 : Lsn(0x20),
8596 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8597 4 : ),
8598 4 : ];
8599 4 : let delta3 = vec![
8600 4 : (
8601 4 : get_key(8),
8602 4 : Lsn(0x48),
8603 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8604 4 : ),
8605 4 : (
8606 4 : get_key(9),
8607 4 : Lsn(0x48),
8608 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8609 4 : ),
8610 4 : ];
8611 :
8612 4 : let tline = if use_delta_bottom_layer {
8613 2 : tenant
8614 2 : .create_test_timeline_with_layers(
8615 2 : TIMELINE_ID,
8616 2 : Lsn(0x08),
8617 2 : DEFAULT_PG_VERSION,
8618 2 : &ctx,
8619 2 : vec![
8620 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8621 2 : Lsn(0x08)..Lsn(0x10),
8622 2 : delta4,
8623 2 : ),
8624 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8625 2 : Lsn(0x20)..Lsn(0x48),
8626 2 : delta1,
8627 2 : ),
8628 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8629 2 : Lsn(0x20)..Lsn(0x48),
8630 2 : delta2,
8631 2 : ),
8632 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8633 2 : Lsn(0x48)..Lsn(0x50),
8634 2 : delta3,
8635 2 : ),
8636 2 : ], // delta layers
8637 2 : vec![], // image layers
8638 2 : Lsn(0x50),
8639 2 : )
8640 2 : .await?
8641 : } else {
8642 2 : tenant
8643 2 : .create_test_timeline_with_layers(
8644 2 : TIMELINE_ID,
8645 2 : Lsn(0x10),
8646 2 : DEFAULT_PG_VERSION,
8647 2 : &ctx,
8648 2 : vec![
8649 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8650 2 : Lsn(0x10)..Lsn(0x48),
8651 2 : delta1,
8652 2 : ),
8653 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8654 2 : Lsn(0x10)..Lsn(0x48),
8655 2 : delta2,
8656 2 : ),
8657 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8658 2 : Lsn(0x48)..Lsn(0x50),
8659 2 : delta3,
8660 2 : ),
8661 2 : ], // delta layers
8662 2 : vec![(Lsn(0x10), img_layer)], // image layers
8663 2 : Lsn(0x50),
8664 2 : )
8665 2 : .await?
8666 : };
8667 : {
8668 4 : tline
8669 4 : .latest_gc_cutoff_lsn
8670 4 : .lock_for_write()
8671 4 : .store_and_unlock(Lsn(0x30))
8672 4 : .wait()
8673 4 : .await;
8674 : // Update GC info
8675 4 : let mut guard = tline.gc_info.write().unwrap();
8676 4 : *guard = GcInfo {
8677 4 : retain_lsns: vec![],
8678 4 : cutoffs: GcCutoffs {
8679 4 : time: Lsn(0x30),
8680 4 : space: Lsn(0x30),
8681 4 : },
8682 4 : leases: Default::default(),
8683 4 : within_ancestor_pitr: false,
8684 4 : };
8685 4 : }
8686 4 :
8687 4 : let expected_result = [
8688 4 : Bytes::from_static(b"value 0@0x10"),
8689 4 : Bytes::from_static(b"value 1@0x10@0x20"),
8690 4 : Bytes::from_static(b"value 2@0x10@0x30"),
8691 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
8692 4 : Bytes::from_static(b"value 4@0x10"),
8693 4 : Bytes::from_static(b"value 5@0x10@0x20"),
8694 4 : Bytes::from_static(b"value 6@0x10@0x20"),
8695 4 : Bytes::from_static(b"value 7@0x10"),
8696 4 : Bytes::from_static(b"value 8@0x10@0x48"),
8697 4 : Bytes::from_static(b"value 9@0x10@0x48"),
8698 4 : ];
8699 4 :
8700 4 : let expected_result_at_gc_horizon = [
8701 4 : Bytes::from_static(b"value 0@0x10"),
8702 4 : Bytes::from_static(b"value 1@0x10@0x20"),
8703 4 : Bytes::from_static(b"value 2@0x10@0x30"),
8704 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
8705 4 : Bytes::from_static(b"value 4@0x10"),
8706 4 : Bytes::from_static(b"value 5@0x10@0x20"),
8707 4 : Bytes::from_static(b"value 6@0x10@0x20"),
8708 4 : Bytes::from_static(b"value 7@0x10"),
8709 4 : Bytes::from_static(b"value 8@0x10"),
8710 4 : Bytes::from_static(b"value 9@0x10"),
8711 4 : ];
8712 :
8713 44 : for idx in 0..10 {
8714 40 : assert_eq!(
8715 40 : tline
8716 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8717 40 : .await
8718 40 : .unwrap(),
8719 40 : &expected_result[idx]
8720 : );
8721 40 : assert_eq!(
8722 40 : tline
8723 40 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
8724 40 : .await
8725 40 : .unwrap(),
8726 40 : &expected_result_at_gc_horizon[idx]
8727 : );
8728 : }
8729 :
8730 4 : let cancel = CancellationToken::new();
8731 4 : tline
8732 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8733 4 : .await
8734 4 : .unwrap();
8735 :
8736 44 : for idx in 0..10 {
8737 40 : assert_eq!(
8738 40 : tline
8739 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8740 40 : .await
8741 40 : .unwrap(),
8742 40 : &expected_result[idx]
8743 : );
8744 40 : assert_eq!(
8745 40 : tline
8746 40 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
8747 40 : .await
8748 40 : .unwrap(),
8749 40 : &expected_result_at_gc_horizon[idx]
8750 : );
8751 : }
8752 :
8753 : // increase GC horizon and compact again
8754 : {
8755 4 : tline
8756 4 : .latest_gc_cutoff_lsn
8757 4 : .lock_for_write()
8758 4 : .store_and_unlock(Lsn(0x40))
8759 4 : .wait()
8760 4 : .await;
8761 : // Update GC info
8762 4 : let mut guard = tline.gc_info.write().unwrap();
8763 4 : guard.cutoffs.time = Lsn(0x40);
8764 4 : guard.cutoffs.space = Lsn(0x40);
8765 4 : }
8766 4 : tline
8767 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8768 4 : .await
8769 4 : .unwrap();
8770 4 :
8771 4 : Ok(())
8772 4 : }
8773 :
8774 : #[cfg(feature = "testing")]
8775 : #[tokio::test]
8776 2 : async fn test_generate_key_retention() -> anyhow::Result<()> {
8777 2 : let harness = TenantHarness::create("test_generate_key_retention").await?;
8778 2 : let (tenant, ctx) = harness.load().await;
8779 2 : let tline = tenant
8780 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
8781 2 : .await?;
8782 2 : tline.force_advance_lsn(Lsn(0x70));
8783 2 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
8784 2 : let history = vec![
8785 2 : (
8786 2 : key,
8787 2 : Lsn(0x10),
8788 2 : Value::WalRecord(NeonWalRecord::wal_init("0x10")),
8789 2 : ),
8790 2 : (
8791 2 : key,
8792 2 : Lsn(0x20),
8793 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8794 2 : ),
8795 2 : (
8796 2 : key,
8797 2 : Lsn(0x30),
8798 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8799 2 : ),
8800 2 : (
8801 2 : key,
8802 2 : Lsn(0x40),
8803 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8804 2 : ),
8805 2 : (
8806 2 : key,
8807 2 : Lsn(0x50),
8808 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
8809 2 : ),
8810 2 : (
8811 2 : key,
8812 2 : Lsn(0x60),
8813 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8814 2 : ),
8815 2 : (
8816 2 : key,
8817 2 : Lsn(0x70),
8818 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8819 2 : ),
8820 2 : (
8821 2 : key,
8822 2 : Lsn(0x80),
8823 2 : Value::Image(Bytes::copy_from_slice(
8824 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8825 2 : )),
8826 2 : ),
8827 2 : (
8828 2 : key,
8829 2 : Lsn(0x90),
8830 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8831 2 : ),
8832 2 : ];
8833 2 : let res = tline
8834 2 : .generate_key_retention(
8835 2 : key,
8836 2 : &history,
8837 2 : Lsn(0x60),
8838 2 : &[Lsn(0x20), Lsn(0x40), Lsn(0x50)],
8839 2 : 3,
8840 2 : None,
8841 2 : )
8842 2 : .await
8843 2 : .unwrap();
8844 2 : let expected_res = KeyHistoryRetention {
8845 2 : below_horizon: vec![
8846 2 : (
8847 2 : Lsn(0x20),
8848 2 : KeyLogAtLsn(vec![(
8849 2 : Lsn(0x20),
8850 2 : Value::Image(Bytes::from_static(b"0x10;0x20")),
8851 2 : )]),
8852 2 : ),
8853 2 : (
8854 2 : Lsn(0x40),
8855 2 : KeyLogAtLsn(vec![
8856 2 : (
8857 2 : Lsn(0x30),
8858 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8859 2 : ),
8860 2 : (
8861 2 : Lsn(0x40),
8862 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8863 2 : ),
8864 2 : ]),
8865 2 : ),
8866 2 : (
8867 2 : Lsn(0x50),
8868 2 : KeyLogAtLsn(vec![(
8869 2 : Lsn(0x50),
8870 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40;0x50")),
8871 2 : )]),
8872 2 : ),
8873 2 : (
8874 2 : Lsn(0x60),
8875 2 : KeyLogAtLsn(vec![(
8876 2 : Lsn(0x60),
8877 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8878 2 : )]),
8879 2 : ),
8880 2 : ],
8881 2 : above_horizon: KeyLogAtLsn(vec![
8882 2 : (
8883 2 : Lsn(0x70),
8884 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8885 2 : ),
8886 2 : (
8887 2 : Lsn(0x80),
8888 2 : Value::Image(Bytes::copy_from_slice(
8889 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8890 2 : )),
8891 2 : ),
8892 2 : (
8893 2 : Lsn(0x90),
8894 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8895 2 : ),
8896 2 : ]),
8897 2 : };
8898 2 : assert_eq!(res, expected_res);
8899 2 :
8900 2 : // We expect GC-compaction to run with the original GC. This would create a situation that
8901 2 : // the original GC algorithm removes some delta layers b/c there are full image coverage,
8902 2 : // therefore causing some keys to have an incomplete history below the lowest retain LSN.
8903 2 : // For example, we have
8904 2 : // ```plain
8905 2 : // init delta @ 0x10, image @ 0x20, delta @ 0x30 (gc_horizon), image @ 0x40.
8906 2 : // ```
8907 2 : // Now the GC horizon moves up, and we have
8908 2 : // ```plain
8909 2 : // init delta @ 0x10, image @ 0x20, delta @ 0x30, image @ 0x40 (gc_horizon)
8910 2 : // ```
8911 2 : // The original GC algorithm kicks in, and removes delta @ 0x10, image @ 0x20.
8912 2 : // We will end up with
8913 2 : // ```plain
8914 2 : // delta @ 0x30, image @ 0x40 (gc_horizon)
8915 2 : // ```
8916 2 : // Now we run the GC-compaction, and this key does not have a full history.
8917 2 : // We should be able to handle this partial history and drop everything before the
8918 2 : // gc_horizon image.
8919 2 :
8920 2 : let history = vec![
8921 2 : (
8922 2 : key,
8923 2 : Lsn(0x20),
8924 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8925 2 : ),
8926 2 : (
8927 2 : key,
8928 2 : Lsn(0x30),
8929 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8930 2 : ),
8931 2 : (
8932 2 : key,
8933 2 : Lsn(0x40),
8934 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
8935 2 : ),
8936 2 : (
8937 2 : key,
8938 2 : Lsn(0x50),
8939 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
8940 2 : ),
8941 2 : (
8942 2 : key,
8943 2 : Lsn(0x60),
8944 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8945 2 : ),
8946 2 : (
8947 2 : key,
8948 2 : Lsn(0x70),
8949 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8950 2 : ),
8951 2 : (
8952 2 : key,
8953 2 : Lsn(0x80),
8954 2 : Value::Image(Bytes::copy_from_slice(
8955 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8956 2 : )),
8957 2 : ),
8958 2 : (
8959 2 : key,
8960 2 : Lsn(0x90),
8961 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8962 2 : ),
8963 2 : ];
8964 2 : let res = tline
8965 2 : .generate_key_retention(key, &history, Lsn(0x60), &[Lsn(0x40), Lsn(0x50)], 3, None)
8966 2 : .await
8967 2 : .unwrap();
8968 2 : let expected_res = KeyHistoryRetention {
8969 2 : below_horizon: vec![
8970 2 : (
8971 2 : Lsn(0x40),
8972 2 : KeyLogAtLsn(vec![(
8973 2 : Lsn(0x40),
8974 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
8975 2 : )]),
8976 2 : ),
8977 2 : (
8978 2 : Lsn(0x50),
8979 2 : KeyLogAtLsn(vec![(
8980 2 : Lsn(0x50),
8981 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
8982 2 : )]),
8983 2 : ),
8984 2 : (
8985 2 : Lsn(0x60),
8986 2 : KeyLogAtLsn(vec![(
8987 2 : Lsn(0x60),
8988 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8989 2 : )]),
8990 2 : ),
8991 2 : ],
8992 2 : above_horizon: KeyLogAtLsn(vec![
8993 2 : (
8994 2 : Lsn(0x70),
8995 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8996 2 : ),
8997 2 : (
8998 2 : Lsn(0x80),
8999 2 : Value::Image(Bytes::copy_from_slice(
9000 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9001 2 : )),
9002 2 : ),
9003 2 : (
9004 2 : Lsn(0x90),
9005 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9006 2 : ),
9007 2 : ]),
9008 2 : };
9009 2 : assert_eq!(res, expected_res);
9010 2 :
9011 2 : // In case of branch compaction, the branch itself does not have the full history, and we need to provide
9012 2 : // the ancestor image in the test case.
9013 2 :
9014 2 : let history = vec![
9015 2 : (
9016 2 : key,
9017 2 : Lsn(0x20),
9018 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9019 2 : ),
9020 2 : (
9021 2 : key,
9022 2 : Lsn(0x30),
9023 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9024 2 : ),
9025 2 : (
9026 2 : key,
9027 2 : Lsn(0x40),
9028 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9029 2 : ),
9030 2 : (
9031 2 : key,
9032 2 : Lsn(0x70),
9033 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9034 2 : ),
9035 2 : ];
9036 2 : let res = tline
9037 2 : .generate_key_retention(
9038 2 : key,
9039 2 : &history,
9040 2 : Lsn(0x60),
9041 2 : &[],
9042 2 : 3,
9043 2 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
9044 2 : )
9045 2 : .await
9046 2 : .unwrap();
9047 2 : let expected_res = KeyHistoryRetention {
9048 2 : below_horizon: vec![(
9049 2 : Lsn(0x60),
9050 2 : KeyLogAtLsn(vec![(
9051 2 : Lsn(0x60),
9052 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")), // use the ancestor image to reconstruct the page
9053 2 : )]),
9054 2 : )],
9055 2 : above_horizon: KeyLogAtLsn(vec![(
9056 2 : Lsn(0x70),
9057 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9058 2 : )]),
9059 2 : };
9060 2 : assert_eq!(res, expected_res);
9061 2 :
9062 2 : let history = vec![
9063 2 : (
9064 2 : key,
9065 2 : Lsn(0x20),
9066 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9067 2 : ),
9068 2 : (
9069 2 : key,
9070 2 : Lsn(0x40),
9071 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9072 2 : ),
9073 2 : (
9074 2 : key,
9075 2 : Lsn(0x60),
9076 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9077 2 : ),
9078 2 : (
9079 2 : key,
9080 2 : Lsn(0x70),
9081 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9082 2 : ),
9083 2 : ];
9084 2 : let res = tline
9085 2 : .generate_key_retention(
9086 2 : key,
9087 2 : &history,
9088 2 : Lsn(0x60),
9089 2 : &[Lsn(0x30)],
9090 2 : 3,
9091 2 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
9092 2 : )
9093 2 : .await
9094 2 : .unwrap();
9095 2 : let expected_res = KeyHistoryRetention {
9096 2 : below_horizon: vec![
9097 2 : (
9098 2 : Lsn(0x30),
9099 2 : KeyLogAtLsn(vec![(
9100 2 : Lsn(0x20),
9101 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9102 2 : )]),
9103 2 : ),
9104 2 : (
9105 2 : Lsn(0x60),
9106 2 : KeyLogAtLsn(vec![(
9107 2 : Lsn(0x60),
9108 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x40;0x60")),
9109 2 : )]),
9110 2 : ),
9111 2 : ],
9112 2 : above_horizon: KeyLogAtLsn(vec![(
9113 2 : Lsn(0x70),
9114 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9115 2 : )]),
9116 2 : };
9117 2 : assert_eq!(res, expected_res);
9118 2 :
9119 2 : Ok(())
9120 2 : }
9121 :
9122 : #[cfg(feature = "testing")]
9123 : #[tokio::test]
9124 2 : async fn test_simple_bottom_most_compaction_with_retain_lsns() -> anyhow::Result<()> {
9125 2 : let harness =
9126 2 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns").await?;
9127 2 : let (tenant, ctx) = harness.load().await;
9128 2 :
9129 518 : fn get_key(id: u32) -> Key {
9130 518 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9131 518 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9132 518 : key.field6 = id;
9133 518 : key
9134 518 : }
9135 2 :
9136 2 : let img_layer = (0..10)
9137 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9138 2 : .collect_vec();
9139 2 :
9140 2 : let delta1 = vec![
9141 2 : (
9142 2 : get_key(1),
9143 2 : Lsn(0x20),
9144 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9145 2 : ),
9146 2 : (
9147 2 : get_key(2),
9148 2 : Lsn(0x30),
9149 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9150 2 : ),
9151 2 : (
9152 2 : get_key(3),
9153 2 : Lsn(0x28),
9154 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9155 2 : ),
9156 2 : (
9157 2 : get_key(3),
9158 2 : Lsn(0x30),
9159 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9160 2 : ),
9161 2 : (
9162 2 : get_key(3),
9163 2 : Lsn(0x40),
9164 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9165 2 : ),
9166 2 : ];
9167 2 : let delta2 = vec![
9168 2 : (
9169 2 : get_key(5),
9170 2 : Lsn(0x20),
9171 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9172 2 : ),
9173 2 : (
9174 2 : get_key(6),
9175 2 : Lsn(0x20),
9176 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9177 2 : ),
9178 2 : ];
9179 2 : let delta3 = vec![
9180 2 : (
9181 2 : get_key(8),
9182 2 : Lsn(0x48),
9183 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9184 2 : ),
9185 2 : (
9186 2 : get_key(9),
9187 2 : Lsn(0x48),
9188 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9189 2 : ),
9190 2 : ];
9191 2 :
9192 2 : let tline = tenant
9193 2 : .create_test_timeline_with_layers(
9194 2 : TIMELINE_ID,
9195 2 : Lsn(0x10),
9196 2 : DEFAULT_PG_VERSION,
9197 2 : &ctx,
9198 2 : vec![
9199 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
9200 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
9201 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
9202 2 : ], // delta layers
9203 2 : vec![(Lsn(0x10), img_layer)], // image layers
9204 2 : Lsn(0x50),
9205 2 : )
9206 2 : .await?;
9207 2 : {
9208 2 : tline
9209 2 : .latest_gc_cutoff_lsn
9210 2 : .lock_for_write()
9211 2 : .store_and_unlock(Lsn(0x30))
9212 2 : .wait()
9213 2 : .await;
9214 2 : // Update GC info
9215 2 : let mut guard = tline.gc_info.write().unwrap();
9216 2 : *guard = GcInfo {
9217 2 : retain_lsns: vec![
9218 2 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
9219 2 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
9220 2 : ],
9221 2 : cutoffs: GcCutoffs {
9222 2 : time: Lsn(0x30),
9223 2 : space: Lsn(0x30),
9224 2 : },
9225 2 : leases: Default::default(),
9226 2 : within_ancestor_pitr: false,
9227 2 : };
9228 2 : }
9229 2 :
9230 2 : let expected_result = [
9231 2 : Bytes::from_static(b"value 0@0x10"),
9232 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9233 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9234 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9235 2 : Bytes::from_static(b"value 4@0x10"),
9236 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9237 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9238 2 : Bytes::from_static(b"value 7@0x10"),
9239 2 : Bytes::from_static(b"value 8@0x10@0x48"),
9240 2 : Bytes::from_static(b"value 9@0x10@0x48"),
9241 2 : ];
9242 2 :
9243 2 : let expected_result_at_gc_horizon = [
9244 2 : Bytes::from_static(b"value 0@0x10"),
9245 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9246 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9247 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
9248 2 : Bytes::from_static(b"value 4@0x10"),
9249 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9250 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9251 2 : Bytes::from_static(b"value 7@0x10"),
9252 2 : Bytes::from_static(b"value 8@0x10"),
9253 2 : Bytes::from_static(b"value 9@0x10"),
9254 2 : ];
9255 2 :
9256 2 : let expected_result_at_lsn_20 = [
9257 2 : Bytes::from_static(b"value 0@0x10"),
9258 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9259 2 : Bytes::from_static(b"value 2@0x10"),
9260 2 : Bytes::from_static(b"value 3@0x10"),
9261 2 : Bytes::from_static(b"value 4@0x10"),
9262 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9263 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9264 2 : Bytes::from_static(b"value 7@0x10"),
9265 2 : Bytes::from_static(b"value 8@0x10"),
9266 2 : Bytes::from_static(b"value 9@0x10"),
9267 2 : ];
9268 2 :
9269 2 : let expected_result_at_lsn_10 = [
9270 2 : Bytes::from_static(b"value 0@0x10"),
9271 2 : Bytes::from_static(b"value 1@0x10"),
9272 2 : Bytes::from_static(b"value 2@0x10"),
9273 2 : Bytes::from_static(b"value 3@0x10"),
9274 2 : Bytes::from_static(b"value 4@0x10"),
9275 2 : Bytes::from_static(b"value 5@0x10"),
9276 2 : Bytes::from_static(b"value 6@0x10"),
9277 2 : Bytes::from_static(b"value 7@0x10"),
9278 2 : Bytes::from_static(b"value 8@0x10"),
9279 2 : Bytes::from_static(b"value 9@0x10"),
9280 2 : ];
9281 2 :
9282 12 : let verify_result = || async {
9283 12 : let gc_horizon = {
9284 12 : let gc_info = tline.gc_info.read().unwrap();
9285 12 : gc_info.cutoffs.time
9286 2 : };
9287 132 : for idx in 0..10 {
9288 120 : assert_eq!(
9289 120 : tline
9290 120 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9291 120 : .await
9292 120 : .unwrap(),
9293 120 : &expected_result[idx]
9294 2 : );
9295 120 : assert_eq!(
9296 120 : tline
9297 120 : .get(get_key(idx as u32), gc_horizon, &ctx)
9298 120 : .await
9299 120 : .unwrap(),
9300 120 : &expected_result_at_gc_horizon[idx]
9301 2 : );
9302 120 : assert_eq!(
9303 120 : tline
9304 120 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
9305 120 : .await
9306 120 : .unwrap(),
9307 120 : &expected_result_at_lsn_20[idx]
9308 2 : );
9309 120 : assert_eq!(
9310 120 : tline
9311 120 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
9312 120 : .await
9313 120 : .unwrap(),
9314 120 : &expected_result_at_lsn_10[idx]
9315 2 : );
9316 2 : }
9317 24 : };
9318 2 :
9319 2 : verify_result().await;
9320 2 :
9321 2 : let cancel = CancellationToken::new();
9322 2 : let mut dryrun_flags = EnumSet::new();
9323 2 : dryrun_flags.insert(CompactFlags::DryRun);
9324 2 :
9325 2 : tline
9326 2 : .compact_with_gc(
9327 2 : &cancel,
9328 2 : CompactOptions {
9329 2 : flags: dryrun_flags,
9330 2 : compact_range: None,
9331 2 : ..Default::default()
9332 2 : },
9333 2 : &ctx,
9334 2 : )
9335 2 : .await
9336 2 : .unwrap();
9337 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
9338 2 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
9339 2 : verify_result().await;
9340 2 :
9341 2 : tline
9342 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9343 2 : .await
9344 2 : .unwrap();
9345 2 : verify_result().await;
9346 2 :
9347 2 : // compact again
9348 2 : tline
9349 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9350 2 : .await
9351 2 : .unwrap();
9352 2 : verify_result().await;
9353 2 :
9354 2 : // increase GC horizon and compact again
9355 2 : {
9356 2 : tline
9357 2 : .latest_gc_cutoff_lsn
9358 2 : .lock_for_write()
9359 2 : .store_and_unlock(Lsn(0x38))
9360 2 : .wait()
9361 2 : .await;
9362 2 : // Update GC info
9363 2 : let mut guard = tline.gc_info.write().unwrap();
9364 2 : guard.cutoffs.time = Lsn(0x38);
9365 2 : guard.cutoffs.space = Lsn(0x38);
9366 2 : }
9367 2 : tline
9368 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9369 2 : .await
9370 2 : .unwrap();
9371 2 : verify_result().await; // no wals between 0x30 and 0x38, so we should obtain the same result
9372 2 :
9373 2 : // not increasing the GC horizon and compact again
9374 2 : tline
9375 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9376 2 : .await
9377 2 : .unwrap();
9378 2 : verify_result().await;
9379 2 :
9380 2 : Ok(())
9381 2 : }
9382 :
9383 : #[cfg(feature = "testing")]
9384 : #[tokio::test]
9385 2 : async fn test_simple_bottom_most_compaction_with_retain_lsns_single_key() -> anyhow::Result<()>
9386 2 : {
9387 2 : let harness =
9388 2 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns_single_key")
9389 2 : .await?;
9390 2 : let (tenant, ctx) = harness.load().await;
9391 2 :
9392 352 : fn get_key(id: u32) -> Key {
9393 352 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9394 352 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9395 352 : key.field6 = id;
9396 352 : key
9397 352 : }
9398 2 :
9399 2 : let img_layer = (0..10)
9400 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9401 2 : .collect_vec();
9402 2 :
9403 2 : let delta1 = vec![
9404 2 : (
9405 2 : get_key(1),
9406 2 : Lsn(0x20),
9407 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9408 2 : ),
9409 2 : (
9410 2 : get_key(1),
9411 2 : Lsn(0x28),
9412 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9413 2 : ),
9414 2 : ];
9415 2 : let delta2 = vec![
9416 2 : (
9417 2 : get_key(1),
9418 2 : Lsn(0x30),
9419 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9420 2 : ),
9421 2 : (
9422 2 : get_key(1),
9423 2 : Lsn(0x38),
9424 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
9425 2 : ),
9426 2 : ];
9427 2 : let delta3 = vec![
9428 2 : (
9429 2 : get_key(8),
9430 2 : Lsn(0x48),
9431 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9432 2 : ),
9433 2 : (
9434 2 : get_key(9),
9435 2 : Lsn(0x48),
9436 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9437 2 : ),
9438 2 : ];
9439 2 :
9440 2 : let tline = tenant
9441 2 : .create_test_timeline_with_layers(
9442 2 : TIMELINE_ID,
9443 2 : Lsn(0x10),
9444 2 : DEFAULT_PG_VERSION,
9445 2 : &ctx,
9446 2 : vec![
9447 2 : // delta1 and delta 2 only contain a single key but multiple updates
9448 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x30), delta1),
9449 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
9450 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x50), delta3),
9451 2 : ], // delta layers
9452 2 : vec![(Lsn(0x10), img_layer)], // image layers
9453 2 : Lsn(0x50),
9454 2 : )
9455 2 : .await?;
9456 2 : {
9457 2 : tline
9458 2 : .latest_gc_cutoff_lsn
9459 2 : .lock_for_write()
9460 2 : .store_and_unlock(Lsn(0x30))
9461 2 : .wait()
9462 2 : .await;
9463 2 : // Update GC info
9464 2 : let mut guard = tline.gc_info.write().unwrap();
9465 2 : *guard = GcInfo {
9466 2 : retain_lsns: vec![
9467 2 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
9468 2 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
9469 2 : ],
9470 2 : cutoffs: GcCutoffs {
9471 2 : time: Lsn(0x30),
9472 2 : space: Lsn(0x30),
9473 2 : },
9474 2 : leases: Default::default(),
9475 2 : within_ancestor_pitr: false,
9476 2 : };
9477 2 : }
9478 2 :
9479 2 : let expected_result = [
9480 2 : Bytes::from_static(b"value 0@0x10"),
9481 2 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
9482 2 : Bytes::from_static(b"value 2@0x10"),
9483 2 : Bytes::from_static(b"value 3@0x10"),
9484 2 : Bytes::from_static(b"value 4@0x10"),
9485 2 : Bytes::from_static(b"value 5@0x10"),
9486 2 : Bytes::from_static(b"value 6@0x10"),
9487 2 : Bytes::from_static(b"value 7@0x10"),
9488 2 : Bytes::from_static(b"value 8@0x10@0x48"),
9489 2 : Bytes::from_static(b"value 9@0x10@0x48"),
9490 2 : ];
9491 2 :
9492 2 : let expected_result_at_gc_horizon = [
9493 2 : Bytes::from_static(b"value 0@0x10"),
9494 2 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
9495 2 : Bytes::from_static(b"value 2@0x10"),
9496 2 : Bytes::from_static(b"value 3@0x10"),
9497 2 : Bytes::from_static(b"value 4@0x10"),
9498 2 : Bytes::from_static(b"value 5@0x10"),
9499 2 : Bytes::from_static(b"value 6@0x10"),
9500 2 : Bytes::from_static(b"value 7@0x10"),
9501 2 : Bytes::from_static(b"value 8@0x10"),
9502 2 : Bytes::from_static(b"value 9@0x10"),
9503 2 : ];
9504 2 :
9505 2 : let expected_result_at_lsn_20 = [
9506 2 : Bytes::from_static(b"value 0@0x10"),
9507 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9508 2 : Bytes::from_static(b"value 2@0x10"),
9509 2 : Bytes::from_static(b"value 3@0x10"),
9510 2 : Bytes::from_static(b"value 4@0x10"),
9511 2 : Bytes::from_static(b"value 5@0x10"),
9512 2 : Bytes::from_static(b"value 6@0x10"),
9513 2 : Bytes::from_static(b"value 7@0x10"),
9514 2 : Bytes::from_static(b"value 8@0x10"),
9515 2 : Bytes::from_static(b"value 9@0x10"),
9516 2 : ];
9517 2 :
9518 2 : let expected_result_at_lsn_10 = [
9519 2 : Bytes::from_static(b"value 0@0x10"),
9520 2 : Bytes::from_static(b"value 1@0x10"),
9521 2 : Bytes::from_static(b"value 2@0x10"),
9522 2 : Bytes::from_static(b"value 3@0x10"),
9523 2 : Bytes::from_static(b"value 4@0x10"),
9524 2 : Bytes::from_static(b"value 5@0x10"),
9525 2 : Bytes::from_static(b"value 6@0x10"),
9526 2 : Bytes::from_static(b"value 7@0x10"),
9527 2 : Bytes::from_static(b"value 8@0x10"),
9528 2 : Bytes::from_static(b"value 9@0x10"),
9529 2 : ];
9530 2 :
9531 8 : let verify_result = || async {
9532 8 : let gc_horizon = {
9533 8 : let gc_info = tline.gc_info.read().unwrap();
9534 8 : gc_info.cutoffs.time
9535 2 : };
9536 88 : for idx in 0..10 {
9537 80 : assert_eq!(
9538 80 : tline
9539 80 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9540 80 : .await
9541 80 : .unwrap(),
9542 80 : &expected_result[idx]
9543 2 : );
9544 80 : assert_eq!(
9545 80 : tline
9546 80 : .get(get_key(idx as u32), gc_horizon, &ctx)
9547 80 : .await
9548 80 : .unwrap(),
9549 80 : &expected_result_at_gc_horizon[idx]
9550 2 : );
9551 80 : assert_eq!(
9552 80 : tline
9553 80 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
9554 80 : .await
9555 80 : .unwrap(),
9556 80 : &expected_result_at_lsn_20[idx]
9557 2 : );
9558 80 : assert_eq!(
9559 80 : tline
9560 80 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
9561 80 : .await
9562 80 : .unwrap(),
9563 80 : &expected_result_at_lsn_10[idx]
9564 2 : );
9565 2 : }
9566 16 : };
9567 2 :
9568 2 : verify_result().await;
9569 2 :
9570 2 : let cancel = CancellationToken::new();
9571 2 : let mut dryrun_flags = EnumSet::new();
9572 2 : dryrun_flags.insert(CompactFlags::DryRun);
9573 2 :
9574 2 : tline
9575 2 : .compact_with_gc(
9576 2 : &cancel,
9577 2 : CompactOptions {
9578 2 : flags: dryrun_flags,
9579 2 : compact_range: None,
9580 2 : ..Default::default()
9581 2 : },
9582 2 : &ctx,
9583 2 : )
9584 2 : .await
9585 2 : .unwrap();
9586 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
9587 2 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
9588 2 : verify_result().await;
9589 2 :
9590 2 : tline
9591 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9592 2 : .await
9593 2 : .unwrap();
9594 2 : verify_result().await;
9595 2 :
9596 2 : // compact again
9597 2 : tline
9598 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9599 2 : .await
9600 2 : .unwrap();
9601 2 : verify_result().await;
9602 2 :
9603 2 : Ok(())
9604 2 : }
9605 :
9606 : #[cfg(feature = "testing")]
9607 : #[tokio::test]
9608 2 : async fn test_simple_bottom_most_compaction_on_branch() -> anyhow::Result<()> {
9609 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_on_branch").await?;
9610 2 : let (tenant, ctx) = harness.load().await;
9611 2 :
9612 126 : fn get_key(id: u32) -> Key {
9613 126 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
9614 126 : key.field6 = id;
9615 126 : key
9616 126 : }
9617 2 :
9618 2 : let img_layer = (0..10)
9619 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9620 2 : .collect_vec();
9621 2 :
9622 2 : let delta1 = vec![
9623 2 : (
9624 2 : get_key(1),
9625 2 : Lsn(0x20),
9626 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9627 2 : ),
9628 2 : (
9629 2 : get_key(2),
9630 2 : Lsn(0x30),
9631 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9632 2 : ),
9633 2 : (
9634 2 : get_key(3),
9635 2 : Lsn(0x28),
9636 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9637 2 : ),
9638 2 : (
9639 2 : get_key(3),
9640 2 : Lsn(0x30),
9641 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9642 2 : ),
9643 2 : (
9644 2 : get_key(3),
9645 2 : Lsn(0x40),
9646 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9647 2 : ),
9648 2 : ];
9649 2 : let delta2 = vec![
9650 2 : (
9651 2 : get_key(5),
9652 2 : Lsn(0x20),
9653 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9654 2 : ),
9655 2 : (
9656 2 : get_key(6),
9657 2 : Lsn(0x20),
9658 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9659 2 : ),
9660 2 : ];
9661 2 : let delta3 = vec![
9662 2 : (
9663 2 : get_key(8),
9664 2 : Lsn(0x48),
9665 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9666 2 : ),
9667 2 : (
9668 2 : get_key(9),
9669 2 : Lsn(0x48),
9670 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9671 2 : ),
9672 2 : ];
9673 2 :
9674 2 : let parent_tline = tenant
9675 2 : .create_test_timeline_with_layers(
9676 2 : TIMELINE_ID,
9677 2 : Lsn(0x10),
9678 2 : DEFAULT_PG_VERSION,
9679 2 : &ctx,
9680 2 : vec![], // delta layers
9681 2 : vec![(Lsn(0x18), img_layer)], // image layers
9682 2 : Lsn(0x18),
9683 2 : )
9684 2 : .await?;
9685 2 :
9686 2 : parent_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
9687 2 :
9688 2 : let branch_tline = tenant
9689 2 : .branch_timeline_test_with_layers(
9690 2 : &parent_tline,
9691 2 : NEW_TIMELINE_ID,
9692 2 : Some(Lsn(0x18)),
9693 2 : &ctx,
9694 2 : vec![
9695 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
9696 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
9697 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
9698 2 : ], // delta layers
9699 2 : vec![], // image layers
9700 2 : Lsn(0x50),
9701 2 : )
9702 2 : .await?;
9703 2 :
9704 2 : branch_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
9705 2 :
9706 2 : {
9707 2 : parent_tline
9708 2 : .latest_gc_cutoff_lsn
9709 2 : .lock_for_write()
9710 2 : .store_and_unlock(Lsn(0x10))
9711 2 : .wait()
9712 2 : .await;
9713 2 : // Update GC info
9714 2 : let mut guard = parent_tline.gc_info.write().unwrap();
9715 2 : *guard = GcInfo {
9716 2 : retain_lsns: vec![(Lsn(0x18), branch_tline.timeline_id, MaybeOffloaded::No)],
9717 2 : cutoffs: GcCutoffs {
9718 2 : time: Lsn(0x10),
9719 2 : space: Lsn(0x10),
9720 2 : },
9721 2 : leases: Default::default(),
9722 2 : within_ancestor_pitr: false,
9723 2 : };
9724 2 : }
9725 2 :
9726 2 : {
9727 2 : branch_tline
9728 2 : .latest_gc_cutoff_lsn
9729 2 : .lock_for_write()
9730 2 : .store_and_unlock(Lsn(0x50))
9731 2 : .wait()
9732 2 : .await;
9733 2 : // Update GC info
9734 2 : let mut guard = branch_tline.gc_info.write().unwrap();
9735 2 : *guard = GcInfo {
9736 2 : retain_lsns: vec![(Lsn(0x40), branch_tline.timeline_id, MaybeOffloaded::No)],
9737 2 : cutoffs: GcCutoffs {
9738 2 : time: Lsn(0x50),
9739 2 : space: Lsn(0x50),
9740 2 : },
9741 2 : leases: Default::default(),
9742 2 : within_ancestor_pitr: false,
9743 2 : };
9744 2 : }
9745 2 :
9746 2 : let expected_result_at_gc_horizon = [
9747 2 : Bytes::from_static(b"value 0@0x10"),
9748 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9749 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9750 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9751 2 : Bytes::from_static(b"value 4@0x10"),
9752 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9753 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9754 2 : Bytes::from_static(b"value 7@0x10"),
9755 2 : Bytes::from_static(b"value 8@0x10@0x48"),
9756 2 : Bytes::from_static(b"value 9@0x10@0x48"),
9757 2 : ];
9758 2 :
9759 2 : let expected_result_at_lsn_40 = [
9760 2 : Bytes::from_static(b"value 0@0x10"),
9761 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9762 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9763 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9764 2 : Bytes::from_static(b"value 4@0x10"),
9765 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9766 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9767 2 : Bytes::from_static(b"value 7@0x10"),
9768 2 : Bytes::from_static(b"value 8@0x10"),
9769 2 : Bytes::from_static(b"value 9@0x10"),
9770 2 : ];
9771 2 :
9772 4 : let verify_result = || async {
9773 44 : for idx in 0..10 {
9774 40 : assert_eq!(
9775 40 : branch_tline
9776 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9777 40 : .await
9778 40 : .unwrap(),
9779 40 : &expected_result_at_gc_horizon[idx]
9780 2 : );
9781 40 : assert_eq!(
9782 40 : branch_tline
9783 40 : .get(get_key(idx as u32), Lsn(0x40), &ctx)
9784 40 : .await
9785 40 : .unwrap(),
9786 40 : &expected_result_at_lsn_40[idx]
9787 2 : );
9788 2 : }
9789 8 : };
9790 2 :
9791 2 : verify_result().await;
9792 2 :
9793 2 : let cancel = CancellationToken::new();
9794 2 : branch_tline
9795 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9796 2 : .await
9797 2 : .unwrap();
9798 2 :
9799 2 : verify_result().await;
9800 2 :
9801 2 : Ok(())
9802 2 : }
9803 :
9804 : // Regression test for https://github.com/neondatabase/neon/issues/9012
9805 : // Create an image arrangement where we have to read at different LSN ranges
9806 : // from a delta layer. This is achieved by overlapping an image layer on top of
9807 : // a delta layer. Like so:
9808 : //
9809 : // A B
9810 : // +----------------+ -> delta_layer
9811 : // | | ^ lsn
9812 : // | =========|-> nested_image_layer |
9813 : // | C | |
9814 : // +----------------+ |
9815 : // ======== -> baseline_image_layer +-------> key
9816 : //
9817 : //
9818 : // When querying the key range [A, B) we need to read at different LSN ranges
9819 : // for [A, C) and [C, B). This test checks that the described edge case is handled correctly.
9820 : #[cfg(feature = "testing")]
9821 : #[tokio::test]
9822 2 : async fn test_vectored_read_with_nested_image_layer() -> anyhow::Result<()> {
9823 2 : let harness = TenantHarness::create("test_vectored_read_with_nested_image_layer").await?;
9824 2 : let (tenant, ctx) = harness.load().await;
9825 2 :
9826 2 : let will_init_keys = [2, 6];
9827 44 : fn get_key(id: u32) -> Key {
9828 44 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
9829 44 : key.field6 = id;
9830 44 : key
9831 44 : }
9832 2 :
9833 2 : let mut expected_key_values = HashMap::new();
9834 2 :
9835 2 : let baseline_image_layer_lsn = Lsn(0x10);
9836 2 : let mut baseline_img_layer = Vec::new();
9837 12 : for i in 0..5 {
9838 10 : let key = get_key(i);
9839 10 : let value = format!("value {i}@{baseline_image_layer_lsn}");
9840 10 :
9841 10 : let removed = expected_key_values.insert(key, value.clone());
9842 10 : assert!(removed.is_none());
9843 2 :
9844 10 : baseline_img_layer.push((key, Bytes::from(value)));
9845 2 : }
9846 2 :
9847 2 : let nested_image_layer_lsn = Lsn(0x50);
9848 2 : let mut nested_img_layer = Vec::new();
9849 12 : for i in 5..10 {
9850 10 : let key = get_key(i);
9851 10 : let value = format!("value {i}@{nested_image_layer_lsn}");
9852 10 :
9853 10 : let removed = expected_key_values.insert(key, value.clone());
9854 10 : assert!(removed.is_none());
9855 2 :
9856 10 : nested_img_layer.push((key, Bytes::from(value)));
9857 2 : }
9858 2 :
9859 2 : let mut delta_layer_spec = Vec::default();
9860 2 : let delta_layer_start_lsn = Lsn(0x20);
9861 2 : let mut delta_layer_end_lsn = delta_layer_start_lsn;
9862 2 :
9863 22 : for i in 0..10 {
9864 20 : let key = get_key(i);
9865 20 : let key_in_nested = nested_img_layer
9866 20 : .iter()
9867 80 : .any(|(key_with_img, _)| *key_with_img == key);
9868 20 : let lsn = {
9869 20 : if key_in_nested {
9870 10 : Lsn(nested_image_layer_lsn.0 + 0x10)
9871 2 : } else {
9872 10 : delta_layer_start_lsn
9873 2 : }
9874 2 : };
9875 2 :
9876 20 : let will_init = will_init_keys.contains(&i);
9877 20 : if will_init {
9878 4 : delta_layer_spec.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init(""))));
9879 4 :
9880 4 : expected_key_values.insert(key, "".to_string());
9881 16 : } else {
9882 16 : let delta = format!("@{lsn}");
9883 16 : delta_layer_spec.push((
9884 16 : key,
9885 16 : lsn,
9886 16 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
9887 16 : ));
9888 16 :
9889 16 : expected_key_values
9890 16 : .get_mut(&key)
9891 16 : .expect("An image exists for each key")
9892 16 : .push_str(delta.as_str());
9893 16 : }
9894 20 : delta_layer_end_lsn = std::cmp::max(delta_layer_start_lsn, lsn);
9895 2 : }
9896 2 :
9897 2 : delta_layer_end_lsn = Lsn(delta_layer_end_lsn.0 + 1);
9898 2 :
9899 2 : assert!(
9900 2 : nested_image_layer_lsn > delta_layer_start_lsn
9901 2 : && nested_image_layer_lsn < delta_layer_end_lsn
9902 2 : );
9903 2 :
9904 2 : let tline = tenant
9905 2 : .create_test_timeline_with_layers(
9906 2 : TIMELINE_ID,
9907 2 : baseline_image_layer_lsn,
9908 2 : DEFAULT_PG_VERSION,
9909 2 : &ctx,
9910 2 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
9911 2 : delta_layer_start_lsn..delta_layer_end_lsn,
9912 2 : delta_layer_spec,
9913 2 : )], // delta layers
9914 2 : vec![
9915 2 : (baseline_image_layer_lsn, baseline_img_layer),
9916 2 : (nested_image_layer_lsn, nested_img_layer),
9917 2 : ], // image layers
9918 2 : delta_layer_end_lsn,
9919 2 : )
9920 2 : .await?;
9921 2 :
9922 2 : let keyspace = KeySpace::single(get_key(0)..get_key(10));
9923 2 : let results = tline
9924 2 : .get_vectored(keyspace, delta_layer_end_lsn, &ctx)
9925 2 : .await
9926 2 : .expect("No vectored errors");
9927 22 : for (key, res) in results {
9928 20 : let value = res.expect("No key errors");
9929 20 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
9930 20 : assert_eq!(value, Bytes::from(expected_value));
9931 2 : }
9932 2 :
9933 2 : Ok(())
9934 2 : }
9935 :
9936 142 : fn sort_layer_key(k1: &PersistentLayerKey, k2: &PersistentLayerKey) -> std::cmp::Ordering {
9937 142 : (
9938 142 : k1.is_delta,
9939 142 : k1.key_range.start,
9940 142 : k1.key_range.end,
9941 142 : k1.lsn_range.start,
9942 142 : k1.lsn_range.end,
9943 142 : )
9944 142 : .cmp(&(
9945 142 : k2.is_delta,
9946 142 : k2.key_range.start,
9947 142 : k2.key_range.end,
9948 142 : k2.lsn_range.start,
9949 142 : k2.lsn_range.end,
9950 142 : ))
9951 142 : }
9952 :
9953 12 : async fn inspect_and_sort(
9954 12 : tline: &Arc<Timeline>,
9955 12 : filter: Option<std::ops::Range<Key>>,
9956 12 : ) -> Vec<PersistentLayerKey> {
9957 12 : let mut all_layers = tline.inspect_historic_layers().await.unwrap();
9958 12 : if let Some(filter) = filter {
9959 64 : all_layers.retain(|layer| overlaps_with(&layer.key_range, &filter));
9960 10 : }
9961 12 : all_layers.sort_by(sort_layer_key);
9962 12 : all_layers
9963 12 : }
9964 :
9965 : #[cfg(feature = "testing")]
9966 10 : fn check_layer_map_key_eq(
9967 10 : mut left: Vec<PersistentLayerKey>,
9968 10 : mut right: Vec<PersistentLayerKey>,
9969 10 : ) {
9970 10 : left.sort_by(sort_layer_key);
9971 10 : right.sort_by(sort_layer_key);
9972 10 : if left != right {
9973 0 : eprintln!("---LEFT---");
9974 0 : for left in left.iter() {
9975 0 : eprintln!("{}", left);
9976 0 : }
9977 0 : eprintln!("---RIGHT---");
9978 0 : for right in right.iter() {
9979 0 : eprintln!("{}", right);
9980 0 : }
9981 0 : assert_eq!(left, right);
9982 10 : }
9983 10 : }
9984 :
9985 : #[cfg(feature = "testing")]
9986 : #[tokio::test]
9987 2 : async fn test_simple_partial_bottom_most_compaction() -> anyhow::Result<()> {
9988 2 : let harness = TenantHarness::create("test_simple_partial_bottom_most_compaction").await?;
9989 2 : let (tenant, ctx) = harness.load().await;
9990 2 :
9991 182 : fn get_key(id: u32) -> Key {
9992 182 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9993 182 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9994 182 : key.field6 = id;
9995 182 : key
9996 182 : }
9997 2 :
9998 2 : // img layer at 0x10
9999 2 : let img_layer = (0..10)
10000 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10001 2 : .collect_vec();
10002 2 :
10003 2 : let delta1 = vec![
10004 2 : (
10005 2 : get_key(1),
10006 2 : Lsn(0x20),
10007 2 : Value::Image(Bytes::from("value 1@0x20")),
10008 2 : ),
10009 2 : (
10010 2 : get_key(2),
10011 2 : Lsn(0x30),
10012 2 : Value::Image(Bytes::from("value 2@0x30")),
10013 2 : ),
10014 2 : (
10015 2 : get_key(3),
10016 2 : Lsn(0x40),
10017 2 : Value::Image(Bytes::from("value 3@0x40")),
10018 2 : ),
10019 2 : ];
10020 2 : let delta2 = vec![
10021 2 : (
10022 2 : get_key(5),
10023 2 : Lsn(0x20),
10024 2 : Value::Image(Bytes::from("value 5@0x20")),
10025 2 : ),
10026 2 : (
10027 2 : get_key(6),
10028 2 : Lsn(0x20),
10029 2 : Value::Image(Bytes::from("value 6@0x20")),
10030 2 : ),
10031 2 : ];
10032 2 : let delta3 = vec![
10033 2 : (
10034 2 : get_key(8),
10035 2 : Lsn(0x48),
10036 2 : Value::Image(Bytes::from("value 8@0x48")),
10037 2 : ),
10038 2 : (
10039 2 : get_key(9),
10040 2 : Lsn(0x48),
10041 2 : Value::Image(Bytes::from("value 9@0x48")),
10042 2 : ),
10043 2 : ];
10044 2 :
10045 2 : let tline = tenant
10046 2 : .create_test_timeline_with_layers(
10047 2 : TIMELINE_ID,
10048 2 : Lsn(0x10),
10049 2 : DEFAULT_PG_VERSION,
10050 2 : &ctx,
10051 2 : vec![
10052 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
10053 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
10054 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
10055 2 : ], // delta layers
10056 2 : vec![(Lsn(0x10), img_layer)], // image layers
10057 2 : Lsn(0x50),
10058 2 : )
10059 2 : .await?;
10060 2 :
10061 2 : {
10062 2 : tline
10063 2 : .latest_gc_cutoff_lsn
10064 2 : .lock_for_write()
10065 2 : .store_and_unlock(Lsn(0x30))
10066 2 : .wait()
10067 2 : .await;
10068 2 : // Update GC info
10069 2 : let mut guard = tline.gc_info.write().unwrap();
10070 2 : *guard = GcInfo {
10071 2 : retain_lsns: vec![(Lsn(0x20), tline.timeline_id, MaybeOffloaded::No)],
10072 2 : cutoffs: GcCutoffs {
10073 2 : time: Lsn(0x30),
10074 2 : space: Lsn(0x30),
10075 2 : },
10076 2 : leases: Default::default(),
10077 2 : within_ancestor_pitr: false,
10078 2 : };
10079 2 : }
10080 2 :
10081 2 : let cancel = CancellationToken::new();
10082 2 :
10083 2 : // Do a partial compaction on key range 0..2
10084 2 : tline
10085 2 : .compact_with_gc(
10086 2 : &cancel,
10087 2 : CompactOptions {
10088 2 : flags: EnumSet::new(),
10089 2 : compact_range: Some((get_key(0)..get_key(2)).into()),
10090 2 : ..Default::default()
10091 2 : },
10092 2 : &ctx,
10093 2 : )
10094 2 : .await
10095 2 : .unwrap();
10096 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10097 2 : check_layer_map_key_eq(
10098 2 : all_layers,
10099 2 : vec![
10100 2 : // newly-generated image layer for the partial compaction range 0-2
10101 2 : PersistentLayerKey {
10102 2 : key_range: get_key(0)..get_key(2),
10103 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10104 2 : is_delta: false,
10105 2 : },
10106 2 : PersistentLayerKey {
10107 2 : key_range: get_key(0)..get_key(10),
10108 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
10109 2 : is_delta: false,
10110 2 : },
10111 2 : // delta1 is split and the second part is rewritten
10112 2 : PersistentLayerKey {
10113 2 : key_range: get_key(2)..get_key(4),
10114 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
10115 2 : is_delta: true,
10116 2 : },
10117 2 : PersistentLayerKey {
10118 2 : key_range: get_key(5)..get_key(7),
10119 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
10120 2 : is_delta: true,
10121 2 : },
10122 2 : PersistentLayerKey {
10123 2 : key_range: get_key(8)..get_key(10),
10124 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
10125 2 : is_delta: true,
10126 2 : },
10127 2 : ],
10128 2 : );
10129 2 :
10130 2 : // Do a partial compaction on key range 2..4
10131 2 : tline
10132 2 : .compact_with_gc(
10133 2 : &cancel,
10134 2 : CompactOptions {
10135 2 : flags: EnumSet::new(),
10136 2 : compact_range: Some((get_key(2)..get_key(4)).into()),
10137 2 : ..Default::default()
10138 2 : },
10139 2 : &ctx,
10140 2 : )
10141 2 : .await
10142 2 : .unwrap();
10143 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10144 2 : check_layer_map_key_eq(
10145 2 : all_layers,
10146 2 : vec![
10147 2 : PersistentLayerKey {
10148 2 : key_range: get_key(0)..get_key(2),
10149 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10150 2 : is_delta: false,
10151 2 : },
10152 2 : PersistentLayerKey {
10153 2 : key_range: get_key(0)..get_key(10),
10154 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
10155 2 : is_delta: false,
10156 2 : },
10157 2 : // image layer generated for the compaction range 2-4
10158 2 : PersistentLayerKey {
10159 2 : key_range: get_key(2)..get_key(4),
10160 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10161 2 : is_delta: false,
10162 2 : },
10163 2 : // we have key2/key3 above the retain_lsn, so we still need this delta layer
10164 2 : PersistentLayerKey {
10165 2 : key_range: get_key(2)..get_key(4),
10166 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
10167 2 : is_delta: true,
10168 2 : },
10169 2 : PersistentLayerKey {
10170 2 : key_range: get_key(5)..get_key(7),
10171 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
10172 2 : is_delta: true,
10173 2 : },
10174 2 : PersistentLayerKey {
10175 2 : key_range: get_key(8)..get_key(10),
10176 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
10177 2 : is_delta: true,
10178 2 : },
10179 2 : ],
10180 2 : );
10181 2 :
10182 2 : // Do a partial compaction on key range 4..9
10183 2 : tline
10184 2 : .compact_with_gc(
10185 2 : &cancel,
10186 2 : CompactOptions {
10187 2 : flags: EnumSet::new(),
10188 2 : compact_range: Some((get_key(4)..get_key(9)).into()),
10189 2 : ..Default::default()
10190 2 : },
10191 2 : &ctx,
10192 2 : )
10193 2 : .await
10194 2 : .unwrap();
10195 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10196 2 : check_layer_map_key_eq(
10197 2 : all_layers,
10198 2 : vec![
10199 2 : PersistentLayerKey {
10200 2 : key_range: get_key(0)..get_key(2),
10201 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10202 2 : is_delta: false,
10203 2 : },
10204 2 : PersistentLayerKey {
10205 2 : key_range: get_key(0)..get_key(10),
10206 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
10207 2 : is_delta: false,
10208 2 : },
10209 2 : PersistentLayerKey {
10210 2 : key_range: get_key(2)..get_key(4),
10211 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10212 2 : is_delta: false,
10213 2 : },
10214 2 : PersistentLayerKey {
10215 2 : key_range: get_key(2)..get_key(4),
10216 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
10217 2 : is_delta: true,
10218 2 : },
10219 2 : // image layer generated for this compaction range
10220 2 : PersistentLayerKey {
10221 2 : key_range: get_key(4)..get_key(9),
10222 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10223 2 : is_delta: false,
10224 2 : },
10225 2 : PersistentLayerKey {
10226 2 : key_range: get_key(8)..get_key(10),
10227 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
10228 2 : is_delta: true,
10229 2 : },
10230 2 : ],
10231 2 : );
10232 2 :
10233 2 : // Do a partial compaction on key range 9..10
10234 2 : tline
10235 2 : .compact_with_gc(
10236 2 : &cancel,
10237 2 : CompactOptions {
10238 2 : flags: EnumSet::new(),
10239 2 : compact_range: Some((get_key(9)..get_key(10)).into()),
10240 2 : ..Default::default()
10241 2 : },
10242 2 : &ctx,
10243 2 : )
10244 2 : .await
10245 2 : .unwrap();
10246 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10247 2 : check_layer_map_key_eq(
10248 2 : all_layers,
10249 2 : vec![
10250 2 : PersistentLayerKey {
10251 2 : key_range: get_key(0)..get_key(2),
10252 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10253 2 : is_delta: false,
10254 2 : },
10255 2 : PersistentLayerKey {
10256 2 : key_range: get_key(0)..get_key(10),
10257 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
10258 2 : is_delta: false,
10259 2 : },
10260 2 : PersistentLayerKey {
10261 2 : key_range: get_key(2)..get_key(4),
10262 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10263 2 : is_delta: false,
10264 2 : },
10265 2 : PersistentLayerKey {
10266 2 : key_range: get_key(2)..get_key(4),
10267 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
10268 2 : is_delta: true,
10269 2 : },
10270 2 : PersistentLayerKey {
10271 2 : key_range: get_key(4)..get_key(9),
10272 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10273 2 : is_delta: false,
10274 2 : },
10275 2 : // image layer generated for the compaction range
10276 2 : PersistentLayerKey {
10277 2 : key_range: get_key(9)..get_key(10),
10278 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10279 2 : is_delta: false,
10280 2 : },
10281 2 : PersistentLayerKey {
10282 2 : key_range: get_key(8)..get_key(10),
10283 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
10284 2 : is_delta: true,
10285 2 : },
10286 2 : ],
10287 2 : );
10288 2 :
10289 2 : // Do a partial compaction on key range 0..10, all image layers below LSN 20 can be replaced with new ones.
10290 2 : tline
10291 2 : .compact_with_gc(
10292 2 : &cancel,
10293 2 : CompactOptions {
10294 2 : flags: EnumSet::new(),
10295 2 : compact_range: Some((get_key(0)..get_key(10)).into()),
10296 2 : ..Default::default()
10297 2 : },
10298 2 : &ctx,
10299 2 : )
10300 2 : .await
10301 2 : .unwrap();
10302 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10303 2 : check_layer_map_key_eq(
10304 2 : all_layers,
10305 2 : vec![
10306 2 : // aha, we removed all unnecessary image/delta layers and got a very clean layer map!
10307 2 : PersistentLayerKey {
10308 2 : key_range: get_key(0)..get_key(10),
10309 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10310 2 : is_delta: false,
10311 2 : },
10312 2 : PersistentLayerKey {
10313 2 : key_range: get_key(2)..get_key(4),
10314 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
10315 2 : is_delta: true,
10316 2 : },
10317 2 : PersistentLayerKey {
10318 2 : key_range: get_key(8)..get_key(10),
10319 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
10320 2 : is_delta: true,
10321 2 : },
10322 2 : ],
10323 2 : );
10324 2 :
10325 2 : Ok(())
10326 2 : }
10327 :
10328 : #[cfg(feature = "testing")]
10329 : #[tokio::test]
10330 2 : async fn test_timeline_offload_retain_lsn() -> anyhow::Result<()> {
10331 2 : let harness = TenantHarness::create("test_timeline_offload_retain_lsn")
10332 2 : .await
10333 2 : .unwrap();
10334 2 : let (tenant, ctx) = harness.load().await;
10335 2 : let tline_parent = tenant
10336 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
10337 2 : .await
10338 2 : .unwrap();
10339 2 : let tline_child = tenant
10340 2 : .branch_timeline_test(&tline_parent, NEW_TIMELINE_ID, Some(Lsn(0x20)), &ctx)
10341 2 : .await
10342 2 : .unwrap();
10343 2 : {
10344 2 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
10345 2 : assert_eq!(
10346 2 : gc_info_parent.retain_lsns,
10347 2 : vec![(Lsn(0x20), tline_child.timeline_id, MaybeOffloaded::No)]
10348 2 : );
10349 2 : }
10350 2 : // We have to directly call the remote_client instead of using the archive function to avoid constructing broker client...
10351 2 : tline_child
10352 2 : .remote_client
10353 2 : .schedule_index_upload_for_timeline_archival_state(TimelineArchivalState::Archived)
10354 2 : .unwrap();
10355 2 : tline_child.remote_client.wait_completion().await.unwrap();
10356 2 : offload_timeline(&tenant, &tline_child)
10357 2 : .instrument(tracing::info_span!(parent: None, "offload_test", tenant_id=%"test", shard_id=%"test", timeline_id=%"test"))
10358 2 : .await.unwrap();
10359 2 : let child_timeline_id = tline_child.timeline_id;
10360 2 : Arc::try_unwrap(tline_child).unwrap();
10361 2 :
10362 2 : {
10363 2 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
10364 2 : assert_eq!(
10365 2 : gc_info_parent.retain_lsns,
10366 2 : vec![(Lsn(0x20), child_timeline_id, MaybeOffloaded::Yes)]
10367 2 : );
10368 2 : }
10369 2 :
10370 2 : tenant
10371 2 : .get_offloaded_timeline(child_timeline_id)
10372 2 : .unwrap()
10373 2 : .defuse_for_tenant_drop();
10374 2 :
10375 2 : Ok(())
10376 2 : }
10377 : }
|
