Line data Source code
1 : //! Timeline repository implementation that keeps old data in layer files, and
2 : //! the recent changes in ephemeral files.
3 : //!
4 : //! See tenant/*_layer.rs files. The functions here are responsible for locating
5 : //! the correct layer for the get/put call, walking back the timeline branching
6 : //! history as needed.
7 : //!
8 : //! The files are stored in the .neon/tenants/<tenant_id>/timelines/<timeline_id>
9 : //! directory. See docs/pageserver-storage.md for how the files are managed.
10 : //! In addition to the layer files, there is a metadata file in the same
11 : //! directory that contains information about the timeline, in particular its
12 : //! parent timeline, and the last LSN that has been written to disk.
13 : //!
14 :
15 : use anyhow::{bail, Context};
16 : use arc_swap::ArcSwap;
17 : use camino::Utf8Path;
18 : use camino::Utf8PathBuf;
19 : use chrono::NaiveDateTime;
20 : use enumset::EnumSet;
21 : use futures::stream::FuturesUnordered;
22 : use futures::StreamExt;
23 : use pageserver_api::models;
24 : use pageserver_api::models::LsnLease;
25 : use pageserver_api::models::TimelineArchivalState;
26 : use pageserver_api::models::TimelineState;
27 : use pageserver_api::models::TopTenantShardItem;
28 : use pageserver_api::models::WalRedoManagerStatus;
29 : use pageserver_api::shard::ShardIdentity;
30 : use pageserver_api::shard::ShardStripeSize;
31 : use pageserver_api::shard::TenantShardId;
32 : use remote_storage::DownloadError;
33 : use remote_storage::GenericRemoteStorage;
34 : use remote_storage::TimeoutOrCancel;
35 : use remote_timeline_client::manifest::{
36 : OffloadedTimelineManifest, TenantManifest, LATEST_TENANT_MANIFEST_VERSION,
37 : };
38 : use remote_timeline_client::UploadQueueNotReadyError;
39 : use std::collections::BTreeMap;
40 : use std::fmt;
41 : use std::future::Future;
42 : use std::sync::atomic::AtomicBool;
43 : use std::sync::Weak;
44 : use std::time::SystemTime;
45 : use storage_broker::BrokerClientChannel;
46 : use timeline::offload::offload_timeline;
47 : use tokio::io::BufReader;
48 : use tokio::sync::watch;
49 : use tokio::task::JoinSet;
50 : use tokio_util::sync::CancellationToken;
51 : use tracing::*;
52 : use upload_queue::NotInitialized;
53 : use utils::backoff;
54 : use utils::circuit_breaker::CircuitBreaker;
55 : use utils::completion;
56 : use utils::crashsafe::path_with_suffix_extension;
57 : use utils::failpoint_support;
58 : use utils::fs_ext;
59 : use utils::pausable_failpoint;
60 : use utils::sync::gate::Gate;
61 : use utils::sync::gate::GateGuard;
62 : use utils::timeout::timeout_cancellable;
63 : use utils::timeout::TimeoutCancellableError;
64 : use utils::zstd::create_zst_tarball;
65 : use utils::zstd::extract_zst_tarball;
66 :
67 : use self::config::AttachedLocationConfig;
68 : use self::config::AttachmentMode;
69 : use self::config::LocationConf;
70 : use self::config::TenantConf;
71 : use self::metadata::TimelineMetadata;
72 : use self::mgr::GetActiveTenantError;
73 : use self::mgr::GetTenantError;
74 : use self::remote_timeline_client::upload::{upload_index_part, upload_tenant_manifest};
75 : use self::remote_timeline_client::{RemoteTimelineClient, WaitCompletionError};
76 : use self::timeline::uninit::TimelineCreateGuard;
77 : use self::timeline::uninit::TimelineExclusionError;
78 : use self::timeline::uninit::UninitializedTimeline;
79 : use self::timeline::EvictionTaskTenantState;
80 : use self::timeline::GcCutoffs;
81 : use self::timeline::TimelineDeleteProgress;
82 : use self::timeline::TimelineResources;
83 : use self::timeline::WaitLsnError;
84 : use crate::config::PageServerConf;
85 : use crate::context::{DownloadBehavior, RequestContext};
86 : use crate::deletion_queue::DeletionQueueClient;
87 : use crate::deletion_queue::DeletionQueueError;
88 : use crate::import_datadir;
89 : use crate::is_uninit_mark;
90 : use crate::l0_flush::L0FlushGlobalState;
91 : use crate::metrics::TENANT;
92 : use crate::metrics::{
93 : remove_tenant_metrics, BROKEN_TENANTS_SET, CIRCUIT_BREAKERS_BROKEN, CIRCUIT_BREAKERS_UNBROKEN,
94 : TENANT_STATE_METRIC, TENANT_SYNTHETIC_SIZE_METRIC,
95 : };
96 : use crate::task_mgr;
97 : use crate::task_mgr::TaskKind;
98 : use crate::tenant::config::LocationMode;
99 : use crate::tenant::config::TenantConfOpt;
100 : use crate::tenant::gc_result::GcResult;
101 : pub use crate::tenant::remote_timeline_client::index::IndexPart;
102 : use crate::tenant::remote_timeline_client::remote_initdb_archive_path;
103 : use crate::tenant::remote_timeline_client::MaybeDeletedIndexPart;
104 : use crate::tenant::remote_timeline_client::INITDB_PATH;
105 : use crate::tenant::storage_layer::DeltaLayer;
106 : use crate::tenant::storage_layer::ImageLayer;
107 : use crate::walingest::WalLagCooldown;
108 : use crate::walredo;
109 : use crate::InitializationOrder;
110 : use std::collections::hash_map::Entry;
111 : use std::collections::HashMap;
112 : use std::collections::HashSet;
113 : use std::fmt::Debug;
114 : use std::fmt::Display;
115 : use std::fs;
116 : use std::fs::File;
117 : use std::sync::atomic::{AtomicU64, Ordering};
118 : use std::sync::Arc;
119 : use std::sync::Mutex;
120 : use std::time::{Duration, Instant};
121 :
122 : use crate::span;
123 : use crate::tenant::timeline::delete::DeleteTimelineFlow;
124 : use crate::tenant::timeline::uninit::cleanup_timeline_directory;
125 : use crate::virtual_file::VirtualFile;
126 : use crate::walredo::PostgresRedoManager;
127 : use crate::TEMP_FILE_SUFFIX;
128 : use once_cell::sync::Lazy;
129 : pub use pageserver_api::models::TenantState;
130 : use tokio::sync::Semaphore;
131 :
132 0 : static INIT_DB_SEMAPHORE: Lazy<Semaphore> = Lazy::new(|| Semaphore::new(8));
133 : use utils::{
134 : crashsafe,
135 : generation::Generation,
136 : id::TimelineId,
137 : lsn::{Lsn, RecordLsn},
138 : };
139 :
140 : pub mod blob_io;
141 : pub mod block_io;
142 : pub mod vectored_blob_io;
143 :
144 : pub mod disk_btree;
145 : pub(crate) mod ephemeral_file;
146 : pub mod layer_map;
147 :
148 : pub mod metadata;
149 : pub mod remote_timeline_client;
150 : pub mod storage_layer;
151 :
152 : pub mod checks;
153 : pub mod config;
154 : pub mod mgr;
155 : pub mod secondary;
156 : pub mod tasks;
157 : pub mod upload_queue;
158 :
159 : pub(crate) mod timeline;
160 :
161 : pub mod size;
162 :
163 : mod gc_block;
164 : mod gc_result;
165 : pub(crate) mod throttle;
166 :
167 : pub(crate) use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
168 : pub(crate) use timeline::{LogicalSizeCalculationCause, PageReconstructError, Timeline};
169 :
170 : // re-export for use in walreceiver
171 : pub use crate::tenant::timeline::WalReceiverInfo;
172 :
173 : /// The "tenants" part of `tenants/<tenant>/timelines...`
174 : pub const TENANTS_SEGMENT_NAME: &str = "tenants";
175 :
176 : /// Parts of the `.neon/tenants/<tenant_id>/timelines/<timeline_id>` directory prefix.
177 : pub const TIMELINES_SEGMENT_NAME: &str = "timelines";
178 :
179 : /// References to shared objects that are passed into each tenant, such
180 : /// as the shared remote storage client and process initialization state.
181 : #[derive(Clone)]
182 : pub struct TenantSharedResources {
183 : pub broker_client: storage_broker::BrokerClientChannel,
184 : pub remote_storage: GenericRemoteStorage,
185 : pub deletion_queue_client: DeletionQueueClient,
186 : pub l0_flush_global_state: L0FlushGlobalState,
187 : }
188 :
189 : /// A [`Tenant`] is really an _attached_ tenant. The configuration
190 : /// for an attached tenant is a subset of the [`LocationConf`], represented
191 : /// in this struct.
192 : pub(super) struct AttachedTenantConf {
193 : tenant_conf: TenantConfOpt,
194 : location: AttachedLocationConfig,
195 : /// The deadline before which we are blocked from GC so that
196 : /// leases have a chance to be renewed.
197 : lsn_lease_deadline: Option<tokio::time::Instant>,
198 : }
199 :
200 : impl AttachedTenantConf {
201 192 : fn new(tenant_conf: TenantConfOpt, location: AttachedLocationConfig) -> Self {
202 : // Sets a deadline before which we cannot proceed to GC due to lsn lease.
203 : //
204 : // We do this as the leases mapping are not persisted to disk. By delaying GC by lease
205 : // length, we guarantee that all the leases we granted before will have a chance to renew
206 : // when we run GC for the first time after restart / transition from AttachedMulti to AttachedSingle.
207 192 : let lsn_lease_deadline = if location.attach_mode == AttachmentMode::Single {
208 192 : Some(
209 192 : tokio::time::Instant::now()
210 192 : + tenant_conf
211 192 : .lsn_lease_length
212 192 : .unwrap_or(LsnLease::DEFAULT_LENGTH),
213 192 : )
214 : } else {
215 : // We don't use `lsn_lease_deadline` to delay GC in AttachedMulti and AttachedStale
216 : // because we don't do GC in these modes.
217 0 : None
218 : };
219 :
220 192 : Self {
221 192 : tenant_conf,
222 192 : location,
223 192 : lsn_lease_deadline,
224 192 : }
225 192 : }
226 :
227 192 : fn try_from(location_conf: LocationConf) -> anyhow::Result<Self> {
228 192 : match &location_conf.mode {
229 192 : LocationMode::Attached(attach_conf) => {
230 192 : Ok(Self::new(location_conf.tenant_conf, *attach_conf))
231 : }
232 : LocationMode::Secondary(_) => {
233 0 : anyhow::bail!("Attempted to construct AttachedTenantConf from a LocationConf in secondary mode")
234 : }
235 : }
236 192 : }
237 :
238 762 : fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
239 762 : self.lsn_lease_deadline
240 762 : .map(|d| tokio::time::Instant::now() < d)
241 762 : .unwrap_or(false)
242 762 : }
243 : }
244 : struct TimelinePreload {
245 : timeline_id: TimelineId,
246 : client: RemoteTimelineClient,
247 : index_part: Result<MaybeDeletedIndexPart, DownloadError>,
248 : }
249 :
250 : pub(crate) struct TenantPreload {
251 : tenant_manifest: TenantManifest,
252 : /// Map from timeline ID to a possible timeline preload. It is None iff the timeline is offloaded according to the manifest.
253 : timelines: HashMap<TimelineId, Option<TimelinePreload>>,
254 : }
255 :
256 : /// When we spawn a tenant, there is a special mode for tenant creation that
257 : /// avoids trying to read anything from remote storage.
258 : pub(crate) enum SpawnMode {
259 : /// Activate as soon as possible
260 : Eager,
261 : /// Lazy activation in the background, with the option to skip the queue if the need comes up
262 : Lazy,
263 : }
264 :
265 : ///
266 : /// Tenant consists of multiple timelines. Keep them in a hash table.
267 : ///
268 : pub struct Tenant {
269 : // Global pageserver config parameters
270 : pub conf: &'static PageServerConf,
271 :
272 : /// The value creation timestamp, used to measure activation delay, see:
273 : /// <https://github.com/neondatabase/neon/issues/4025>
274 : constructed_at: Instant,
275 :
276 : state: watch::Sender<TenantState>,
277 :
278 : // Overridden tenant-specific config parameters.
279 : // We keep TenantConfOpt sturct here to preserve the information
280 : // about parameters that are not set.
281 : // This is necessary to allow global config updates.
282 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
283 :
284 : tenant_shard_id: TenantShardId,
285 :
286 : // The detailed sharding information, beyond the number/count in tenant_shard_id
287 : shard_identity: ShardIdentity,
288 :
289 : /// The remote storage generation, used to protect S3 objects from split-brain.
290 : /// Does not change over the lifetime of the [`Tenant`] object.
291 : ///
292 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
293 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
294 : generation: Generation,
295 :
296 : timelines: Mutex<HashMap<TimelineId, Arc<Timeline>>>,
297 :
298 : /// During timeline creation, we first insert the TimelineId to the
299 : /// creating map, then `timelines`, then remove it from the creating map.
300 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
301 : timelines_creating: std::sync::Mutex<HashSet<TimelineId>>,
302 :
303 : /// Possibly offloaded and archived timelines
304 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
305 : timelines_offloaded: Mutex<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
306 :
307 : /// Serialize writes of the tenant manifest to remote storage. If there are concurrent operations
308 : /// affecting the manifest, such as timeline deletion and timeline offload, they must wait for
309 : /// each other (this could be optimized to coalesce writes if necessary).
310 : ///
311 : /// The contents of the Mutex are the last manifest we successfully uploaded
312 : tenant_manifest_upload: tokio::sync::Mutex<Option<TenantManifest>>,
313 :
314 : // This mutex prevents creation of new timelines during GC.
315 : // Adding yet another mutex (in addition to `timelines`) is needed because holding
316 : // `timelines` mutex during all GC iteration
317 : // may block for a long time `get_timeline`, `get_timelines_state`,... and other operations
318 : // with timelines, which in turn may cause dropping replication connection, expiration of wait_for_lsn
319 : // timeout...
320 : gc_cs: tokio::sync::Mutex<()>,
321 : walredo_mgr: Option<Arc<WalRedoManager>>,
322 :
323 : // provides access to timeline data sitting in the remote storage
324 : pub(crate) remote_storage: GenericRemoteStorage,
325 :
326 : // Access to global deletion queue for when this tenant wants to schedule a deletion
327 : deletion_queue_client: DeletionQueueClient,
328 :
329 : /// Cached logical sizes updated updated on each [`Tenant::gather_size_inputs`].
330 : cached_logical_sizes: tokio::sync::Mutex<HashMap<(TimelineId, Lsn), u64>>,
331 : cached_synthetic_tenant_size: Arc<AtomicU64>,
332 :
333 : eviction_task_tenant_state: tokio::sync::Mutex<EvictionTaskTenantState>,
334 :
335 : /// Track repeated failures to compact, so that we can back off.
336 : /// Overhead of mutex is acceptable because compaction is done with a multi-second period.
337 : compaction_circuit_breaker: std::sync::Mutex<CircuitBreaker>,
338 :
339 : /// If the tenant is in Activating state, notify this to encourage it
340 : /// to proceed to Active as soon as possible, rather than waiting for lazy
341 : /// background warmup.
342 : pub(crate) activate_now_sem: tokio::sync::Semaphore,
343 :
344 : /// Time it took for the tenant to activate. Zero if not active yet.
345 : attach_wal_lag_cooldown: Arc<std::sync::OnceLock<WalLagCooldown>>,
346 :
347 : // Cancellation token fires when we have entered shutdown(). This is a parent of
348 : // Timelines' cancellation token.
349 : pub(crate) cancel: CancellationToken,
350 :
351 : // Users of the Tenant such as the page service must take this Gate to avoid
352 : // trying to use a Tenant which is shutting down.
353 : pub(crate) gate: Gate,
354 :
355 : /// Throttle applied at the top of [`Timeline::get`].
356 : /// All [`Tenant::timelines`] of a given [`Tenant`] instance share the same [`throttle::Throttle`] instance.
357 : pub(crate) timeline_get_throttle:
358 : Arc<throttle::Throttle<crate::metrics::tenant_throttling::TimelineGet>>,
359 :
360 : /// An ongoing timeline detach concurrency limiter.
361 : ///
362 : /// As a tenant will likely be restarted as part of timeline detach ancestor it makes no sense
363 : /// to have two running at the same time. A different one can be started if an earlier one
364 : /// has failed for whatever reason.
365 : ongoing_timeline_detach: std::sync::Mutex<Option<(TimelineId, utils::completion::Barrier)>>,
366 :
367 : /// `index_part.json` based gc blocking reason tracking.
368 : ///
369 : /// New gc iterations must start a new iteration by acquiring `GcBlock::start` before
370 : /// proceeding.
371 : pub(crate) gc_block: gc_block::GcBlock,
372 :
373 : l0_flush_global_state: L0FlushGlobalState,
374 : }
375 :
376 : impl std::fmt::Debug for Tenant {
377 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
378 0 : write!(f, "{} ({})", self.tenant_shard_id, self.current_state())
379 0 : }
380 : }
381 :
382 : pub(crate) enum WalRedoManager {
383 : Prod(WalredoManagerId, PostgresRedoManager),
384 : #[cfg(test)]
385 : Test(harness::TestRedoManager),
386 : }
387 :
388 0 : #[derive(thiserror::Error, Debug)]
389 : #[error("pageserver is shutting down")]
390 : pub(crate) struct GlobalShutDown;
391 :
392 : impl WalRedoManager {
393 0 : pub(crate) fn new(mgr: PostgresRedoManager) -> Result<Arc<Self>, GlobalShutDown> {
394 0 : let id = WalredoManagerId::next();
395 0 : let arc = Arc::new(Self::Prod(id, mgr));
396 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
397 0 : match &mut *guard {
398 0 : Some(map) => {
399 0 : map.insert(id, Arc::downgrade(&arc));
400 0 : Ok(arc)
401 : }
402 0 : None => Err(GlobalShutDown),
403 : }
404 0 : }
405 : }
406 :
407 : impl Drop for WalRedoManager {
408 10 : fn drop(&mut self) {
409 10 : match self {
410 0 : Self::Prod(id, _) => {
411 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
412 0 : if let Some(map) = &mut *guard {
413 0 : map.remove(id).expect("new() registers, drop() unregisters");
414 0 : }
415 : }
416 : #[cfg(test)]
417 10 : Self::Test(_) => {
418 10 : // Not applicable to test redo manager
419 10 : }
420 : }
421 10 : }
422 : }
423 :
424 : /// Global registry of all walredo managers so that [`crate::shutdown_pageserver`] can shut down
425 : /// the walredo processes outside of the regular order.
426 : ///
427 : /// This is necessary to work around a systemd bug where it freezes if there are
428 : /// walredo processes left => <https://github.com/neondatabase/cloud/issues/11387>
429 : #[allow(clippy::type_complexity)]
430 : pub(crate) static WALREDO_MANAGERS: once_cell::sync::Lazy<
431 : Mutex<Option<HashMap<WalredoManagerId, Weak<WalRedoManager>>>>,
432 0 : > = once_cell::sync::Lazy::new(|| Mutex::new(Some(HashMap::new())));
433 : #[derive(PartialEq, Eq, Hash, Clone, Copy, Debug)]
434 : pub(crate) struct WalredoManagerId(u64);
435 : impl WalredoManagerId {
436 0 : pub fn next() -> Self {
437 : static NEXT: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(1);
438 0 : let id = NEXT.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
439 0 : if id == 0 {
440 0 : panic!("WalredoManagerId::new() returned 0, indicating wraparound, risking it's no longer unique");
441 0 : }
442 0 : Self(id)
443 0 : }
444 : }
445 :
446 : #[cfg(test)]
447 : impl From<harness::TestRedoManager> for WalRedoManager {
448 192 : fn from(mgr: harness::TestRedoManager) -> Self {
449 192 : Self::Test(mgr)
450 192 : }
451 : }
452 :
453 : impl WalRedoManager {
454 6 : pub(crate) async fn shutdown(&self) -> bool {
455 6 : match self {
456 0 : Self::Prod(_, mgr) => mgr.shutdown().await,
457 : #[cfg(test)]
458 : Self::Test(_) => {
459 : // Not applicable to test redo manager
460 6 : true
461 : }
462 : }
463 6 : }
464 :
465 0 : pub(crate) fn maybe_quiesce(&self, idle_timeout: Duration) {
466 0 : match self {
467 0 : Self::Prod(_, mgr) => mgr.maybe_quiesce(idle_timeout),
468 0 : #[cfg(test)]
469 0 : Self::Test(_) => {
470 0 : // Not applicable to test redo manager
471 0 : }
472 0 : }
473 0 : }
474 :
475 : /// # Cancel-Safety
476 : ///
477 : /// This method is cancellation-safe.
478 410 : pub async fn request_redo(
479 410 : &self,
480 410 : key: pageserver_api::key::Key,
481 410 : lsn: Lsn,
482 410 : base_img: Option<(Lsn, bytes::Bytes)>,
483 410 : records: Vec<(Lsn, pageserver_api::record::NeonWalRecord)>,
484 410 : pg_version: u32,
485 410 : ) -> Result<bytes::Bytes, walredo::Error> {
486 410 : match self {
487 0 : Self::Prod(_, mgr) => {
488 0 : mgr.request_redo(key, lsn, base_img, records, pg_version)
489 0 : .await
490 : }
491 : #[cfg(test)]
492 410 : Self::Test(mgr) => {
493 410 : mgr.request_redo(key, lsn, base_img, records, pg_version)
494 0 : .await
495 : }
496 : }
497 410 : }
498 :
499 0 : pub(crate) fn status(&self) -> Option<WalRedoManagerStatus> {
500 0 : match self {
501 0 : WalRedoManager::Prod(_, m) => Some(m.status()),
502 0 : #[cfg(test)]
503 0 : WalRedoManager::Test(_) => None,
504 0 : }
505 0 : }
506 : }
507 :
508 : /// A very lightweight memory representation of an offloaded timeline.
509 : ///
510 : /// We need to store the list of offloaded timelines so that we can perform operations on them,
511 : /// like unoffloading them, or (at a later date), decide to perform flattening.
512 : /// This type has a much smaller memory impact than [`Timeline`], and thus we can store many
513 : /// more offloaded timelines than we can manage ones that aren't.
514 : pub struct OffloadedTimeline {
515 : pub tenant_shard_id: TenantShardId,
516 : pub timeline_id: TimelineId,
517 : pub ancestor_timeline_id: Option<TimelineId>,
518 : /// Whether to retain the branch lsn at the ancestor or not
519 : pub ancestor_retain_lsn: Option<Lsn>,
520 :
521 : /// When the timeline was archived.
522 : ///
523 : /// Present for future flattening deliberations.
524 : pub archived_at: NaiveDateTime,
525 :
526 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
527 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
528 : pub delete_progress: TimelineDeleteProgress,
529 :
530 : /// Part of the `OffloadedTimeline` object's lifecycle: this needs to be set before we drop it
531 : pub deleted_from_ancestor: AtomicBool,
532 : }
533 :
534 : impl OffloadedTimeline {
535 : /// Obtains an offloaded timeline from a given timeline object.
536 : ///
537 : /// Returns `None` if the `archived_at` flag couldn't be obtained, i.e.
538 : /// the timeline is not in a stopped state.
539 : /// Panics if the timeline is not archived.
540 2 : fn from_timeline(timeline: &Timeline) -> Result<Self, UploadQueueNotReadyError> {
541 2 : let (ancestor_retain_lsn, ancestor_timeline_id) =
542 2 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
543 2 : let ancestor_lsn = timeline.get_ancestor_lsn();
544 2 : let ancestor_timeline_id = ancestor_timeline.timeline_id;
545 2 : let mut gc_info = ancestor_timeline.gc_info.write().unwrap();
546 2 : gc_info.insert_child(timeline.timeline_id, ancestor_lsn, MaybeOffloaded::Yes);
547 2 : (Some(ancestor_lsn), Some(ancestor_timeline_id))
548 : } else {
549 0 : (None, None)
550 : };
551 2 : let archived_at = timeline
552 2 : .remote_client
553 2 : .archived_at_stopped_queue()?
554 2 : .expect("must be called on an archived timeline");
555 2 : Ok(Self {
556 2 : tenant_shard_id: timeline.tenant_shard_id,
557 2 : timeline_id: timeline.timeline_id,
558 2 : ancestor_timeline_id,
559 2 : ancestor_retain_lsn,
560 2 : archived_at,
561 2 :
562 2 : delete_progress: timeline.delete_progress.clone(),
563 2 : deleted_from_ancestor: AtomicBool::new(false),
564 2 : })
565 2 : }
566 0 : fn from_manifest(tenant_shard_id: TenantShardId, manifest: &OffloadedTimelineManifest) -> Self {
567 0 : // We expect to reach this case in tenant loading, where the `retain_lsn` is populated in the parent's `gc_info`
568 0 : // by the `initialize_gc_info` function.
569 0 : let OffloadedTimelineManifest {
570 0 : timeline_id,
571 0 : ancestor_timeline_id,
572 0 : ancestor_retain_lsn,
573 0 : archived_at,
574 0 : } = *manifest;
575 0 : Self {
576 0 : tenant_shard_id,
577 0 : timeline_id,
578 0 : ancestor_timeline_id,
579 0 : ancestor_retain_lsn,
580 0 : archived_at,
581 0 : delete_progress: TimelineDeleteProgress::default(),
582 0 : deleted_from_ancestor: AtomicBool::new(false),
583 0 : }
584 0 : }
585 2 : fn manifest(&self) -> OffloadedTimelineManifest {
586 2 : let Self {
587 2 : timeline_id,
588 2 : ancestor_timeline_id,
589 2 : ancestor_retain_lsn,
590 2 : archived_at,
591 2 : ..
592 2 : } = self;
593 2 : OffloadedTimelineManifest {
594 2 : timeline_id: *timeline_id,
595 2 : ancestor_timeline_id: *ancestor_timeline_id,
596 2 : ancestor_retain_lsn: *ancestor_retain_lsn,
597 2 : archived_at: *archived_at,
598 2 : }
599 2 : }
600 : /// Delete this timeline's retain_lsn from its ancestor, if present in the given tenant
601 0 : fn delete_from_ancestor_with_timelines(
602 0 : &self,
603 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
604 0 : ) {
605 0 : if let (Some(_retain_lsn), Some(ancestor_timeline_id)) =
606 0 : (self.ancestor_retain_lsn, self.ancestor_timeline_id)
607 : {
608 0 : if let Some((_, ancestor_timeline)) = timelines
609 0 : .iter()
610 0 : .find(|(tid, _tl)| **tid == ancestor_timeline_id)
611 : {
612 0 : let removal_happened = ancestor_timeline
613 0 : .gc_info
614 0 : .write()
615 0 : .unwrap()
616 0 : .remove_child_offloaded(self.timeline_id);
617 0 : if !removal_happened {
618 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
619 0 : "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
620 0 : }
621 0 : }
622 0 : }
623 0 : self.deleted_from_ancestor.store(true, Ordering::Release);
624 0 : }
625 : /// Call [`Self::delete_from_ancestor_with_timelines`] instead if possible.
626 : ///
627 : /// As the entire tenant is being dropped, don't bother deregistering the `retain_lsn` from the ancestor.
628 2 : fn defuse_for_tenant_drop(&self) {
629 2 : self.deleted_from_ancestor.store(true, Ordering::Release);
630 2 : }
631 : }
632 :
633 : impl fmt::Debug for OffloadedTimeline {
634 0 : fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
635 0 : write!(f, "OffloadedTimeline<{}>", self.timeline_id)
636 0 : }
637 : }
638 :
639 : impl Drop for OffloadedTimeline {
640 2 : fn drop(&mut self) {
641 2 : if !self.deleted_from_ancestor.load(Ordering::Acquire) {
642 0 : tracing::warn!(
643 0 : "offloaded timeline {} was dropped without having cleaned it up at the ancestor",
644 : self.timeline_id
645 : );
646 2 : }
647 2 : }
648 : }
649 :
650 : #[derive(Copy, Clone, PartialEq, Eq, Hash, Debug)]
651 : pub enum MaybeOffloaded {
652 : Yes,
653 : No,
654 : }
655 :
656 : #[derive(Clone, Debug)]
657 : pub enum TimelineOrOffloaded {
658 : Timeline(Arc<Timeline>),
659 : Offloaded(Arc<OffloadedTimeline>),
660 : }
661 :
662 : impl TimelineOrOffloaded {
663 0 : pub fn arc_ref(&self) -> TimelineOrOffloadedArcRef<'_> {
664 0 : match self {
665 0 : TimelineOrOffloaded::Timeline(timeline) => {
666 0 : TimelineOrOffloadedArcRef::Timeline(timeline)
667 : }
668 0 : TimelineOrOffloaded::Offloaded(offloaded) => {
669 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded)
670 : }
671 : }
672 0 : }
673 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
674 0 : self.arc_ref().tenant_shard_id()
675 0 : }
676 0 : pub fn timeline_id(&self) -> TimelineId {
677 0 : self.arc_ref().timeline_id()
678 0 : }
679 2 : pub fn delete_progress(&self) -> &Arc<tokio::sync::Mutex<DeleteTimelineFlow>> {
680 2 : match self {
681 2 : TimelineOrOffloaded::Timeline(timeline) => &timeline.delete_progress,
682 0 : TimelineOrOffloaded::Offloaded(offloaded) => &offloaded.delete_progress,
683 : }
684 2 : }
685 0 : fn maybe_remote_client(&self) -> Option<Arc<RemoteTimelineClient>> {
686 0 : match self {
687 0 : TimelineOrOffloaded::Timeline(timeline) => Some(timeline.remote_client.clone()),
688 0 : TimelineOrOffloaded::Offloaded(_offloaded) => None,
689 : }
690 0 : }
691 : }
692 :
693 : pub enum TimelineOrOffloadedArcRef<'a> {
694 : Timeline(&'a Arc<Timeline>),
695 : Offloaded(&'a Arc<OffloadedTimeline>),
696 : }
697 :
698 : impl TimelineOrOffloadedArcRef<'_> {
699 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
700 0 : match self {
701 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.tenant_shard_id,
702 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.tenant_shard_id,
703 : }
704 0 : }
705 0 : pub fn timeline_id(&self) -> TimelineId {
706 0 : match self {
707 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.timeline_id,
708 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.timeline_id,
709 : }
710 0 : }
711 : }
712 :
713 : impl<'a> From<&'a Arc<Timeline>> for TimelineOrOffloadedArcRef<'a> {
714 0 : fn from(timeline: &'a Arc<Timeline>) -> Self {
715 0 : Self::Timeline(timeline)
716 0 : }
717 : }
718 :
719 : impl<'a> From<&'a Arc<OffloadedTimeline>> for TimelineOrOffloadedArcRef<'a> {
720 0 : fn from(timeline: &'a Arc<OffloadedTimeline>) -> Self {
721 0 : Self::Offloaded(timeline)
722 0 : }
723 : }
724 :
725 0 : #[derive(Debug, thiserror::Error, PartialEq, Eq)]
726 : pub enum GetTimelineError {
727 : #[error("Timeline is shutting down")]
728 : ShuttingDown,
729 : #[error("Timeline {tenant_id}/{timeline_id} is not active, state: {state:?}")]
730 : NotActive {
731 : tenant_id: TenantShardId,
732 : timeline_id: TimelineId,
733 : state: TimelineState,
734 : },
735 : #[error("Timeline {tenant_id}/{timeline_id} was not found")]
736 : NotFound {
737 : tenant_id: TenantShardId,
738 : timeline_id: TimelineId,
739 : },
740 : }
741 :
742 0 : #[derive(Debug, thiserror::Error)]
743 : pub enum LoadLocalTimelineError {
744 : #[error("FailedToLoad")]
745 : Load(#[source] anyhow::Error),
746 : #[error("FailedToResumeDeletion")]
747 : ResumeDeletion(#[source] anyhow::Error),
748 : }
749 :
750 0 : #[derive(thiserror::Error)]
751 : pub enum DeleteTimelineError {
752 : #[error("NotFound")]
753 : NotFound,
754 :
755 : #[error("HasChildren")]
756 : HasChildren(Vec<TimelineId>),
757 :
758 : #[error("Timeline deletion is already in progress")]
759 : AlreadyInProgress(Arc<tokio::sync::Mutex<DeleteTimelineFlow>>),
760 :
761 : #[error("Cancelled")]
762 : Cancelled,
763 :
764 : #[error(transparent)]
765 : Other(#[from] anyhow::Error),
766 : }
767 :
768 : impl Debug for DeleteTimelineError {
769 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
770 0 : match self {
771 0 : Self::NotFound => write!(f, "NotFound"),
772 0 : Self::HasChildren(c) => f.debug_tuple("HasChildren").field(c).finish(),
773 0 : Self::AlreadyInProgress(_) => f.debug_tuple("AlreadyInProgress").finish(),
774 0 : Self::Cancelled => f.debug_tuple("Cancelled").finish(),
775 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
776 : }
777 0 : }
778 : }
779 :
780 0 : #[derive(thiserror::Error)]
781 : pub enum TimelineArchivalError {
782 : #[error("NotFound")]
783 : NotFound,
784 :
785 : #[error("Timeout")]
786 : Timeout,
787 :
788 : #[error("Cancelled")]
789 : Cancelled,
790 :
791 : #[error("ancestor is archived: {}", .0)]
792 : HasArchivedParent(TimelineId),
793 :
794 : #[error("HasUnarchivedChildren")]
795 : HasUnarchivedChildren(Vec<TimelineId>),
796 :
797 : #[error("Timeline archival is already in progress")]
798 : AlreadyInProgress,
799 :
800 : #[error(transparent)]
801 : Other(anyhow::Error),
802 : }
803 :
804 0 : #[derive(thiserror::Error, Debug)]
805 : pub(crate) enum TenantManifestError {
806 : #[error("Remote storage error: {0}")]
807 : RemoteStorage(anyhow::Error),
808 :
809 : #[error("Cancelled")]
810 : Cancelled,
811 : }
812 :
813 : impl From<TenantManifestError> for TimelineArchivalError {
814 0 : fn from(e: TenantManifestError) -> Self {
815 0 : match e {
816 0 : TenantManifestError::RemoteStorage(e) => Self::Other(e),
817 0 : TenantManifestError::Cancelled => Self::Cancelled,
818 : }
819 0 : }
820 : }
821 :
822 : impl Debug for TimelineArchivalError {
823 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
824 0 : match self {
825 0 : Self::NotFound => write!(f, "NotFound"),
826 0 : Self::Timeout => write!(f, "Timeout"),
827 0 : Self::Cancelled => write!(f, "Cancelled"),
828 0 : Self::HasArchivedParent(p) => f.debug_tuple("HasArchivedParent").field(p).finish(),
829 0 : Self::HasUnarchivedChildren(c) => {
830 0 : f.debug_tuple("HasUnarchivedChildren").field(c).finish()
831 : }
832 0 : Self::AlreadyInProgress => f.debug_tuple("AlreadyInProgress").finish(),
833 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
834 : }
835 0 : }
836 : }
837 :
838 : pub enum SetStoppingError {
839 : AlreadyStopping(completion::Barrier),
840 : Broken,
841 : }
842 :
843 : impl Debug for SetStoppingError {
844 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
845 0 : match self {
846 0 : Self::AlreadyStopping(_) => f.debug_tuple("AlreadyStopping").finish(),
847 0 : Self::Broken => write!(f, "Broken"),
848 : }
849 0 : }
850 : }
851 :
852 : /// Arguments to [`Tenant::create_timeline`].
853 : ///
854 : /// Not usable as an idempotency key for timeline creation because if [`CreateTimelineParamsBranch::ancestor_start_lsn`]
855 : /// is `None`, the result of the timeline create call is not deterministic.
856 : ///
857 : /// See [`CreateTimelineIdempotency`] for an idempotency key.
858 : #[derive(Debug)]
859 : pub(crate) enum CreateTimelineParams {
860 : Bootstrap(CreateTimelineParamsBootstrap),
861 : Branch(CreateTimelineParamsBranch),
862 : }
863 :
864 : #[derive(Debug)]
865 : pub(crate) struct CreateTimelineParamsBootstrap {
866 : pub(crate) new_timeline_id: TimelineId,
867 : pub(crate) existing_initdb_timeline_id: Option<TimelineId>,
868 : pub(crate) pg_version: u32,
869 : }
870 :
871 : /// NB: See comment on [`CreateTimelineIdempotency::Branch`] for why there's no `pg_version` here.
872 : #[derive(Debug)]
873 : pub(crate) struct CreateTimelineParamsBranch {
874 : pub(crate) new_timeline_id: TimelineId,
875 : pub(crate) ancestor_timeline_id: TimelineId,
876 : pub(crate) ancestor_start_lsn: Option<Lsn>,
877 : }
878 :
879 : /// What is used to determine idempotency of a [`Tenant::create_timeline`] call in [`Tenant::start_creating_timeline`].
880 : ///
881 : /// Each [`Timeline`] object holds [`Self`] as an immutable property in [`Timeline::create_idempotency`].
882 : ///
883 : /// We lower timeline creation requests to [`Self`], and then use [`PartialEq::eq`] to compare [`Timeline::create_idempotency`] with the request.
884 : /// If they are equal, we return a reference to the existing timeline, otherwise it's an idempotency conflict.
885 : ///
886 : /// There is special treatment for [`Self::FailWithConflict`] to always return an idempotency conflict.
887 : /// It would be nice to have more advanced derive macros to make that special treatment declarative.
888 : ///
889 : /// Notes:
890 : /// - Unlike [`CreateTimelineParams`], ancestor LSN is fixed, so, branching will be at a deterministic LSN.
891 : /// - We make some trade-offs though, e.g., [`CreateTimelineParamsBootstrap::existing_initdb_timeline_id`]
892 : /// is not considered for idempotency. We can improve on this over time if we deem it necessary.
893 : ///
894 : #[derive(Debug, Clone, PartialEq, Eq)]
895 : pub(crate) enum CreateTimelineIdempotency {
896 : /// NB: special treatment, see comment in [`Self`].
897 : FailWithConflict,
898 : Bootstrap {
899 : pg_version: u32,
900 : },
901 : /// NB: branches always have the same `pg_version` as their ancestor.
902 : /// While [`pageserver_api::models::TimelineCreateRequestMode::Branch::pg_version`]
903 : /// exists as a field, and is set by cplane, it has always been ignored by pageserver when
904 : /// determining the child branch pg_version.
905 : Branch {
906 : ancestor_timeline_id: TimelineId,
907 : ancestor_start_lsn: Lsn,
908 : },
909 : }
910 :
911 : /// What is returned by [`Tenant::start_creating_timeline`].
912 : #[must_use]
913 : enum StartCreatingTimelineResult<'t> {
914 : CreateGuard(TimelineCreateGuard<'t>),
915 : Idempotent(Arc<Timeline>),
916 : }
917 :
918 : /// What is returned by [`Tenant::create_timeline`].
919 : enum CreateTimelineResult {
920 : Created(Arc<Timeline>),
921 : Idempotent(Arc<Timeline>),
922 : }
923 :
924 : impl CreateTimelineResult {
925 0 : fn discriminant(&self) -> &'static str {
926 0 : match self {
927 0 : Self::Created(_) => "Created",
928 0 : Self::Idempotent(_) => "Idempotent",
929 : }
930 0 : }
931 0 : fn timeline(&self) -> &Arc<Timeline> {
932 0 : match self {
933 0 : Self::Created(t) | Self::Idempotent(t) => t,
934 0 : }
935 0 : }
936 : /// Unit test timelines aren't activated, test has to do it if it needs to.
937 : #[cfg(test)]
938 230 : fn into_timeline_for_test(self) -> Arc<Timeline> {
939 230 : match self {
940 230 : Self::Created(t) | Self::Idempotent(t) => t,
941 230 : }
942 230 : }
943 : }
944 :
945 2 : #[derive(thiserror::Error, Debug)]
946 : pub enum CreateTimelineError {
947 : #[error("creation of timeline with the given ID is in progress")]
948 : AlreadyCreating,
949 : #[error("timeline already exists with different parameters")]
950 : Conflict,
951 : #[error(transparent)]
952 : AncestorLsn(anyhow::Error),
953 : #[error("ancestor timeline is not active")]
954 : AncestorNotActive,
955 : #[error("ancestor timeline is archived")]
956 : AncestorArchived,
957 : #[error("tenant shutting down")]
958 : ShuttingDown,
959 : #[error(transparent)]
960 : Other(#[from] anyhow::Error),
961 : }
962 :
963 : #[derive(thiserror::Error, Debug)]
964 : enum InitdbError {
965 : Other(anyhow::Error),
966 : Cancelled,
967 : Spawn(std::io::Result<()>),
968 : Failed(std::process::ExitStatus, Vec<u8>),
969 : }
970 :
971 : impl fmt::Display for InitdbError {
972 0 : fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
973 0 : match self {
974 0 : InitdbError::Cancelled => write!(f, "Operation was cancelled"),
975 0 : InitdbError::Spawn(e) => write!(f, "Spawn error: {:?}", e),
976 0 : InitdbError::Failed(status, stderr) => write!(
977 0 : f,
978 0 : "Command failed with status {:?}: {}",
979 0 : status,
980 0 : String::from_utf8_lossy(stderr)
981 0 : ),
982 0 : InitdbError::Other(e) => write!(f, "Error: {:?}", e),
983 : }
984 0 : }
985 : }
986 :
987 : impl From<std::io::Error> for InitdbError {
988 0 : fn from(error: std::io::Error) -> Self {
989 0 : InitdbError::Spawn(Err(error))
990 0 : }
991 : }
992 :
993 : enum CreateTimelineCause {
994 : Load,
995 : Delete,
996 : }
997 :
998 0 : #[derive(thiserror::Error, Debug)]
999 : pub(crate) enum GcError {
1000 : // The tenant is shutting down
1001 : #[error("tenant shutting down")]
1002 : TenantCancelled,
1003 :
1004 : // The tenant is shutting down
1005 : #[error("timeline shutting down")]
1006 : TimelineCancelled,
1007 :
1008 : // The tenant is in a state inelegible to run GC
1009 : #[error("not active")]
1010 : NotActive,
1011 :
1012 : // A requested GC cutoff LSN was invalid, for example it tried to move backwards
1013 : #[error("not active")]
1014 : BadLsn { why: String },
1015 :
1016 : // A remote storage error while scheduling updates after compaction
1017 : #[error(transparent)]
1018 : Remote(anyhow::Error),
1019 :
1020 : // An error reading while calculating GC cutoffs
1021 : #[error(transparent)]
1022 : GcCutoffs(PageReconstructError),
1023 :
1024 : // If GC was invoked for a particular timeline, this error means it didn't exist
1025 : #[error("timeline not found")]
1026 : TimelineNotFound,
1027 : }
1028 :
1029 : impl From<PageReconstructError> for GcError {
1030 0 : fn from(value: PageReconstructError) -> Self {
1031 0 : match value {
1032 0 : PageReconstructError::Cancelled => Self::TimelineCancelled,
1033 0 : other => Self::GcCutoffs(other),
1034 : }
1035 0 : }
1036 : }
1037 :
1038 : impl From<NotInitialized> for GcError {
1039 0 : fn from(value: NotInitialized) -> Self {
1040 0 : match value {
1041 0 : NotInitialized::Uninitialized => GcError::Remote(value.into()),
1042 0 : NotInitialized::Stopped | NotInitialized::ShuttingDown => GcError::TimelineCancelled,
1043 : }
1044 0 : }
1045 : }
1046 :
1047 : impl From<timeline::layer_manager::Shutdown> for GcError {
1048 0 : fn from(_: timeline::layer_manager::Shutdown) -> Self {
1049 0 : GcError::TimelineCancelled
1050 0 : }
1051 : }
1052 :
1053 0 : #[derive(thiserror::Error, Debug)]
1054 : pub(crate) enum LoadConfigError {
1055 : #[error("TOML deserialization error: '{0}'")]
1056 : DeserializeToml(#[from] toml_edit::de::Error),
1057 :
1058 : #[error("Config not found at {0}")]
1059 : NotFound(Utf8PathBuf),
1060 : }
1061 :
1062 : impl Tenant {
1063 : /// Yet another helper for timeline initialization.
1064 : ///
1065 : /// - Initializes the Timeline struct and inserts it into the tenant's hash map
1066 : /// - Scans the local timeline directory for layer files and builds the layer map
1067 : /// - Downloads remote index file and adds remote files to the layer map
1068 : /// - Schedules remote upload tasks for any files that are present locally but missing from remote storage.
1069 : ///
1070 : /// If the operation fails, the timeline is left in the tenant's hash map in Broken state. On success,
1071 : /// it is marked as Active.
1072 : #[allow(clippy::too_many_arguments)]
1073 6 : async fn timeline_init_and_sync(
1074 6 : &self,
1075 6 : timeline_id: TimelineId,
1076 6 : resources: TimelineResources,
1077 6 : index_part: IndexPart,
1078 6 : metadata: TimelineMetadata,
1079 6 : ancestor: Option<Arc<Timeline>>,
1080 6 : _ctx: &RequestContext,
1081 6 : ) -> anyhow::Result<()> {
1082 6 : let tenant_id = self.tenant_shard_id;
1083 :
1084 6 : let idempotency = if metadata.ancestor_timeline().is_none() {
1085 4 : CreateTimelineIdempotency::Bootstrap {
1086 4 : pg_version: metadata.pg_version(),
1087 4 : }
1088 : } else {
1089 2 : CreateTimelineIdempotency::Branch {
1090 2 : ancestor_timeline_id: metadata.ancestor_timeline().unwrap(),
1091 2 : ancestor_start_lsn: metadata.ancestor_lsn(),
1092 2 : }
1093 : };
1094 :
1095 6 : let timeline = self.create_timeline_struct(
1096 6 : timeline_id,
1097 6 : &metadata,
1098 6 : ancestor.clone(),
1099 6 : resources,
1100 6 : CreateTimelineCause::Load,
1101 6 : idempotency.clone(),
1102 6 : )?;
1103 6 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
1104 6 : anyhow::ensure!(
1105 6 : disk_consistent_lsn.is_valid(),
1106 0 : "Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn"
1107 : );
1108 6 : assert_eq!(
1109 6 : disk_consistent_lsn,
1110 6 : metadata.disk_consistent_lsn(),
1111 0 : "these are used interchangeably"
1112 : );
1113 :
1114 6 : timeline.remote_client.init_upload_queue(&index_part)?;
1115 :
1116 6 : timeline
1117 6 : .load_layer_map(disk_consistent_lsn, index_part)
1118 1 : .await
1119 6 : .with_context(|| {
1120 0 : format!("Failed to load layermap for timeline {tenant_id}/{timeline_id}")
1121 6 : })?;
1122 :
1123 : {
1124 : // avoiding holding it across awaits
1125 6 : let mut timelines_accessor = self.timelines.lock().unwrap();
1126 6 : match timelines_accessor.entry(timeline_id) {
1127 : // We should never try and load the same timeline twice during startup
1128 : Entry::Occupied(_) => {
1129 0 : unreachable!(
1130 0 : "Timeline {tenant_id}/{timeline_id} already exists in the tenant map"
1131 0 : );
1132 : }
1133 6 : Entry::Vacant(v) => {
1134 6 : v.insert(Arc::clone(&timeline));
1135 6 : timeline.maybe_spawn_flush_loop();
1136 6 : }
1137 6 : }
1138 6 : };
1139 6 :
1140 6 : // Sanity check: a timeline should have some content.
1141 6 : anyhow::ensure!(
1142 6 : ancestor.is_some()
1143 4 : || timeline
1144 4 : .layers
1145 4 : .read()
1146 0 : .await
1147 4 : .layer_map()
1148 4 : .expect("currently loading, layer manager cannot be shutdown already")
1149 4 : .iter_historic_layers()
1150 4 : .next()
1151 4 : .is_some(),
1152 0 : "Timeline has no ancestor and no layer files"
1153 : );
1154 :
1155 6 : Ok(())
1156 6 : }
1157 :
1158 : /// Attach a tenant that's available in cloud storage.
1159 : ///
1160 : /// This returns quickly, after just creating the in-memory object
1161 : /// Tenant struct and launching a background task to download
1162 : /// the remote index files. On return, the tenant is most likely still in
1163 : /// Attaching state, and it will become Active once the background task
1164 : /// finishes. You can use wait_until_active() to wait for the task to
1165 : /// complete.
1166 : ///
1167 : #[allow(clippy::too_many_arguments)]
1168 0 : pub(crate) fn spawn(
1169 0 : conf: &'static PageServerConf,
1170 0 : tenant_shard_id: TenantShardId,
1171 0 : resources: TenantSharedResources,
1172 0 : attached_conf: AttachedTenantConf,
1173 0 : shard_identity: ShardIdentity,
1174 0 : init_order: Option<InitializationOrder>,
1175 0 : mode: SpawnMode,
1176 0 : ctx: &RequestContext,
1177 0 : ) -> Result<Arc<Tenant>, GlobalShutDown> {
1178 0 : let wal_redo_manager =
1179 0 : WalRedoManager::new(PostgresRedoManager::new(conf, tenant_shard_id))?;
1180 :
1181 : let TenantSharedResources {
1182 0 : broker_client,
1183 0 : remote_storage,
1184 0 : deletion_queue_client,
1185 0 : l0_flush_global_state,
1186 0 : } = resources;
1187 0 :
1188 0 : let attach_mode = attached_conf.location.attach_mode;
1189 0 : let generation = attached_conf.location.generation;
1190 0 :
1191 0 : let tenant = Arc::new(Tenant::new(
1192 0 : TenantState::Attaching,
1193 0 : conf,
1194 0 : attached_conf,
1195 0 : shard_identity,
1196 0 : Some(wal_redo_manager),
1197 0 : tenant_shard_id,
1198 0 : remote_storage.clone(),
1199 0 : deletion_queue_client,
1200 0 : l0_flush_global_state,
1201 0 : ));
1202 0 :
1203 0 : // The attach task will carry a GateGuard, so that shutdown() reliably waits for it to drop out if
1204 0 : // we shut down while attaching.
1205 0 : let attach_gate_guard = tenant
1206 0 : .gate
1207 0 : .enter()
1208 0 : .expect("We just created the Tenant: nothing else can have shut it down yet");
1209 0 :
1210 0 : // Do all the hard work in the background
1211 0 : let tenant_clone = Arc::clone(&tenant);
1212 0 : let ctx = ctx.detached_child(TaskKind::Attach, DownloadBehavior::Warn);
1213 0 : task_mgr::spawn(
1214 0 : &tokio::runtime::Handle::current(),
1215 0 : TaskKind::Attach,
1216 0 : tenant_shard_id,
1217 0 : None,
1218 0 : "attach tenant",
1219 0 : async move {
1220 0 :
1221 0 : info!(
1222 : ?attach_mode,
1223 0 : "Attaching tenant"
1224 : );
1225 :
1226 0 : let _gate_guard = attach_gate_guard;
1227 0 :
1228 0 : // Is this tenant being spawned as part of process startup?
1229 0 : let starting_up = init_order.is_some();
1230 0 : scopeguard::defer! {
1231 0 : if starting_up {
1232 0 : TENANT.startup_complete.inc();
1233 0 : }
1234 0 : }
1235 :
1236 : // Ideally we should use Tenant::set_broken_no_wait, but it is not supposed to be used when tenant is in loading state.
1237 : enum BrokenVerbosity {
1238 : Error,
1239 : Info
1240 : }
1241 0 : let make_broken =
1242 0 : |t: &Tenant, err: anyhow::Error, verbosity: BrokenVerbosity| {
1243 0 : match verbosity {
1244 : BrokenVerbosity::Info => {
1245 0 : info!("attach cancelled, setting tenant state to Broken: {err}");
1246 : },
1247 : BrokenVerbosity::Error => {
1248 0 : error!("attach failed, setting tenant state to Broken: {err:?}");
1249 : }
1250 : }
1251 0 : t.state.send_modify(|state| {
1252 0 : // The Stopping case is for when we have passed control on to DeleteTenantFlow:
1253 0 : // if it errors, we will call make_broken when tenant is already in Stopping.
1254 0 : assert!(
1255 0 : matches!(*state, TenantState::Attaching | TenantState::Stopping { .. }),
1256 0 : "the attach task owns the tenant state until activation is complete"
1257 : );
1258 :
1259 0 : *state = TenantState::broken_from_reason(err.to_string());
1260 0 : });
1261 0 : };
1262 :
1263 : // TODO: should also be rejecting tenant conf changes that violate this check.
1264 0 : if let Err(e) = crate::tenant::storage_layer::inmemory_layer::IndexEntry::validate_checkpoint_distance(tenant_clone.get_checkpoint_distance()) {
1265 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1266 0 : return Ok(());
1267 0 : }
1268 0 :
1269 0 : let mut init_order = init_order;
1270 0 : // take the completion because initial tenant loading will complete when all of
1271 0 : // these tasks complete.
1272 0 : let _completion = init_order
1273 0 : .as_mut()
1274 0 : .and_then(|x| x.initial_tenant_load.take());
1275 0 : let remote_load_completion = init_order
1276 0 : .as_mut()
1277 0 : .and_then(|x| x.initial_tenant_load_remote.take());
1278 :
1279 : enum AttachType<'a> {
1280 : /// We are attaching this tenant lazily in the background.
1281 : Warmup {
1282 : _permit: tokio::sync::SemaphorePermit<'a>,
1283 : during_startup: bool
1284 : },
1285 : /// We are attaching this tenant as soon as we can, because for example an
1286 : /// endpoint tried to access it.
1287 : OnDemand,
1288 : /// During normal operations after startup, we are attaching a tenant, and
1289 : /// eager attach was requested.
1290 : Normal,
1291 : }
1292 :
1293 0 : let attach_type = if matches!(mode, SpawnMode::Lazy) {
1294 : // Before doing any I/O, wait for at least one of:
1295 : // - A client attempting to access to this tenant (on-demand loading)
1296 : // - A permit becoming available in the warmup semaphore (background warmup)
1297 :
1298 0 : tokio::select!(
1299 0 : permit = tenant_clone.activate_now_sem.acquire() => {
1300 0 : let _ = permit.expect("activate_now_sem is never closed");
1301 0 : tracing::info!("Activating tenant (on-demand)");
1302 0 : AttachType::OnDemand
1303 : },
1304 0 : permit = conf.concurrent_tenant_warmup.inner().acquire() => {
1305 0 : let _permit = permit.expect("concurrent_tenant_warmup semaphore is never closed");
1306 0 : tracing::info!("Activating tenant (warmup)");
1307 0 : AttachType::Warmup {
1308 0 : _permit,
1309 0 : during_startup: init_order.is_some()
1310 0 : }
1311 : }
1312 0 : _ = tenant_clone.cancel.cancelled() => {
1313 : // This is safe, but should be pretty rare: it is interesting if a tenant
1314 : // stayed in Activating for such a long time that shutdown found it in
1315 : // that state.
1316 0 : tracing::info!(state=%tenant_clone.current_state(), "Tenant shut down before activation");
1317 : // Make the tenant broken so that set_stopping will not hang waiting for it to leave
1318 : // the Attaching state. This is an over-reaction (nothing really broke, the tenant is
1319 : // just shutting down), but ensures progress.
1320 0 : make_broken(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"), BrokenVerbosity::Info);
1321 0 : return Ok(());
1322 : },
1323 : )
1324 : } else {
1325 : // SpawnMode::{Create,Eager} always cause jumping ahead of the
1326 : // concurrent_tenant_warmup queue
1327 0 : AttachType::Normal
1328 : };
1329 :
1330 0 : let preload = match &mode {
1331 : SpawnMode::Eager | SpawnMode::Lazy => {
1332 0 : let _preload_timer = TENANT.preload.start_timer();
1333 0 : let res = tenant_clone
1334 0 : .preload(&remote_storage, task_mgr::shutdown_token())
1335 0 : .await;
1336 0 : match res {
1337 0 : Ok(p) => Some(p),
1338 0 : Err(e) => {
1339 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1340 0 : return Ok(());
1341 : }
1342 : }
1343 : }
1344 :
1345 : };
1346 :
1347 : // Remote preload is complete.
1348 0 : drop(remote_load_completion);
1349 0 :
1350 0 :
1351 0 : // We will time the duration of the attach phase unless this is a creation (attach will do no work)
1352 0 : let attach_start = std::time::Instant::now();
1353 0 : let attached = {
1354 0 : let _attach_timer = Some(TENANT.attach.start_timer());
1355 0 : tenant_clone.attach(preload, &ctx).await
1356 : };
1357 0 : let attach_duration = attach_start.elapsed();
1358 0 : _ = tenant_clone.attach_wal_lag_cooldown.set(WalLagCooldown::new(attach_start, attach_duration));
1359 0 :
1360 0 : match attached {
1361 : Ok(()) => {
1362 0 : info!("attach finished, activating");
1363 0 : tenant_clone.activate(broker_client, None, &ctx);
1364 : }
1365 0 : Err(e) => {
1366 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1367 0 : }
1368 : }
1369 :
1370 : // If we are doing an opportunistic warmup attachment at startup, initialize
1371 : // logical size at the same time. This is better than starting a bunch of idle tenants
1372 : // with cold caches and then coming back later to initialize their logical sizes.
1373 : //
1374 : // It also prevents the warmup proccess competing with the concurrency limit on
1375 : // logical size calculations: if logical size calculation semaphore is saturated,
1376 : // then warmup will wait for that before proceeding to the next tenant.
1377 0 : if matches!(attach_type, AttachType::Warmup { during_startup: true, .. }) {
1378 0 : let mut futs: FuturesUnordered<_> = tenant_clone.timelines.lock().unwrap().values().cloned().map(|t| t.await_initial_logical_size()).collect();
1379 0 : tracing::info!("Waiting for initial logical sizes while warming up...");
1380 0 : while futs.next().await.is_some() {}
1381 0 : tracing::info!("Warm-up complete");
1382 0 : }
1383 :
1384 0 : Ok(())
1385 0 : }
1386 0 : .instrument(tracing::info_span!(parent: None, "attach", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), gen=?generation)),
1387 : );
1388 0 : Ok(tenant)
1389 0 : }
1390 :
1391 192 : #[instrument(skip_all)]
1392 : pub(crate) async fn preload(
1393 : self: &Arc<Self>,
1394 : remote_storage: &GenericRemoteStorage,
1395 : cancel: CancellationToken,
1396 : ) -> anyhow::Result<TenantPreload> {
1397 : span::debug_assert_current_span_has_tenant_id();
1398 : // Get list of remote timelines
1399 : // download index files for every tenant timeline
1400 : info!("listing remote timelines");
1401 : let (mut remote_timeline_ids, other_keys) = remote_timeline_client::list_remote_timelines(
1402 : remote_storage,
1403 : self.tenant_shard_id,
1404 : cancel.clone(),
1405 : )
1406 : .await?;
1407 : let (offloaded_add, tenant_manifest) =
1408 : match remote_timeline_client::download_tenant_manifest(
1409 : remote_storage,
1410 : &self.tenant_shard_id,
1411 : self.generation,
1412 : &cancel,
1413 : )
1414 : .await
1415 : {
1416 : Ok((tenant_manifest, _generation, _manifest_mtime)) => (
1417 : format!("{} offloaded", tenant_manifest.offloaded_timelines.len()),
1418 : tenant_manifest,
1419 : ),
1420 : Err(DownloadError::NotFound) => {
1421 : ("no manifest".to_string(), TenantManifest::empty())
1422 : }
1423 : Err(e) => Err(e)?,
1424 : };
1425 :
1426 : info!(
1427 : "found {} timelines, and {offloaded_add}",
1428 : remote_timeline_ids.len()
1429 : );
1430 :
1431 : for k in other_keys {
1432 : warn!("Unexpected non timeline key {k}");
1433 : }
1434 :
1435 : // Avoid downloading IndexPart of offloaded timelines.
1436 : let mut offloaded_with_prefix = HashSet::new();
1437 : for offloaded in tenant_manifest.offloaded_timelines.iter() {
1438 : if remote_timeline_ids.remove(&offloaded.timeline_id) {
1439 : offloaded_with_prefix.insert(offloaded.timeline_id);
1440 : } else {
1441 : // We'll take care later of timelines in the manifest without a prefix
1442 : }
1443 : }
1444 :
1445 : let timelines = self
1446 : .load_timelines_metadata(remote_timeline_ids, remote_storage, cancel)
1447 : .await?;
1448 :
1449 : Ok(TenantPreload {
1450 : tenant_manifest,
1451 : timelines: timelines
1452 : .into_iter()
1453 6 : .map(|(id, tl)| (id, Some(tl)))
1454 0 : .chain(offloaded_with_prefix.into_iter().map(|id| (id, None)))
1455 : .collect(),
1456 : })
1457 : }
1458 :
1459 : ///
1460 : /// Background task that downloads all data for a tenant and brings it to Active state.
1461 : ///
1462 : /// No background tasks are started as part of this routine.
1463 : ///
1464 192 : async fn attach(
1465 192 : self: &Arc<Tenant>,
1466 192 : preload: Option<TenantPreload>,
1467 192 : ctx: &RequestContext,
1468 192 : ) -> anyhow::Result<()> {
1469 192 : span::debug_assert_current_span_has_tenant_id();
1470 192 :
1471 192 : failpoint_support::sleep_millis_async!("before-attaching-tenant");
1472 :
1473 192 : let Some(preload) = preload else {
1474 0 : anyhow::bail!("local-only deployment is no longer supported, https://github.com/neondatabase/neon/issues/5624");
1475 : };
1476 :
1477 192 : let mut offloaded_timeline_ids = HashSet::new();
1478 192 : let mut offloaded_timelines_list = Vec::new();
1479 192 : for timeline_manifest in preload.tenant_manifest.offloaded_timelines.iter() {
1480 0 : let timeline_id = timeline_manifest.timeline_id;
1481 0 : let offloaded_timeline =
1482 0 : OffloadedTimeline::from_manifest(self.tenant_shard_id, timeline_manifest);
1483 0 : offloaded_timelines_list.push((timeline_id, Arc::new(offloaded_timeline)));
1484 0 : offloaded_timeline_ids.insert(timeline_id);
1485 0 : }
1486 : // Complete deletions for offloaded timeline id's from manifest.
1487 : // The manifest will be uploaded later in this function.
1488 192 : offloaded_timelines_list
1489 192 : .retain(|(offloaded_id, offloaded)| {
1490 0 : // Existence of a timeline is finally determined by the existence of an index-part.json in remote storage.
1491 0 : // If there is dangling references in another location, they need to be cleaned up.
1492 0 : let delete = !preload.timelines.contains_key(offloaded_id);
1493 0 : if delete {
1494 0 : tracing::info!("Removing offloaded timeline {offloaded_id} from manifest as no remote prefix was found");
1495 0 : offloaded.defuse_for_tenant_drop();
1496 0 : }
1497 0 : !delete
1498 192 : });
1499 192 :
1500 192 : let mut timelines_to_resume_deletions = vec![];
1501 192 :
1502 192 : let mut remote_index_and_client = HashMap::new();
1503 192 : let mut timeline_ancestors = HashMap::new();
1504 192 : let mut existent_timelines = HashSet::new();
1505 198 : for (timeline_id, preload) in preload.timelines {
1506 6 : let Some(preload) = preload else { continue };
1507 : // This is an invariant of the `preload` function's API
1508 6 : assert!(!offloaded_timeline_ids.contains(&timeline_id));
1509 6 : let index_part = match preload.index_part {
1510 6 : Ok(i) => {
1511 6 : debug!("remote index part exists for timeline {timeline_id}");
1512 : // We found index_part on the remote, this is the standard case.
1513 6 : existent_timelines.insert(timeline_id);
1514 6 : i
1515 : }
1516 : Err(DownloadError::NotFound) => {
1517 : // There is no index_part on the remote. We only get here
1518 : // if there is some prefix for the timeline in the remote storage.
1519 : // This can e.g. be the initdb.tar.zst archive, maybe a
1520 : // remnant from a prior incomplete creation or deletion attempt.
1521 : // Delete the local directory as the deciding criterion for a
1522 : // timeline's existence is presence of index_part.
1523 0 : info!(%timeline_id, "index_part not found on remote");
1524 0 : continue;
1525 : }
1526 0 : Err(DownloadError::Fatal(why)) => {
1527 0 : // If, while loading one remote timeline, we saw an indication that our generation
1528 0 : // number is likely invalid, then we should not load the whole tenant.
1529 0 : error!(%timeline_id, "Fatal error loading timeline: {why}");
1530 0 : anyhow::bail!(why.to_string());
1531 : }
1532 0 : Err(e) => {
1533 0 : // Some (possibly ephemeral) error happened during index_part download.
1534 0 : // Pretend the timeline exists to not delete the timeline directory,
1535 0 : // as it might be a temporary issue and we don't want to re-download
1536 0 : // everything after it resolves.
1537 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1538 :
1539 0 : existent_timelines.insert(timeline_id);
1540 0 : continue;
1541 : }
1542 : };
1543 6 : match index_part {
1544 6 : MaybeDeletedIndexPart::IndexPart(index_part) => {
1545 6 : timeline_ancestors.insert(timeline_id, index_part.metadata.clone());
1546 6 : remote_index_and_client.insert(timeline_id, (index_part, preload.client));
1547 6 : }
1548 0 : MaybeDeletedIndexPart::Deleted(index_part) => {
1549 0 : info!(
1550 0 : "timeline {} is deleted, picking to resume deletion",
1551 : timeline_id
1552 : );
1553 0 : timelines_to_resume_deletions.push((timeline_id, index_part, preload.client));
1554 : }
1555 : }
1556 : }
1557 :
1558 192 : let mut gc_blocks = HashMap::new();
1559 :
1560 : // For every timeline, download the metadata file, scan the local directory,
1561 : // and build a layer map that contains an entry for each remote and local
1562 : // layer file.
1563 192 : let sorted_timelines = tree_sort_timelines(timeline_ancestors, |m| m.ancestor_timeline())?;
1564 198 : for (timeline_id, remote_metadata) in sorted_timelines {
1565 6 : let (index_part, remote_client) = remote_index_and_client
1566 6 : .remove(&timeline_id)
1567 6 : .expect("just put it in above");
1568 :
1569 6 : if let Some(blocking) = index_part.gc_blocking.as_ref() {
1570 : // could just filter these away, but it helps while testing
1571 0 : anyhow::ensure!(
1572 0 : !blocking.reasons.is_empty(),
1573 0 : "index_part for {timeline_id} is malformed: it should not have gc blocking with zero reasons"
1574 : );
1575 0 : let prev = gc_blocks.insert(timeline_id, blocking.reasons);
1576 0 : assert!(prev.is_none());
1577 6 : }
1578 :
1579 : // TODO again handle early failure
1580 6 : self.load_remote_timeline(
1581 6 : timeline_id,
1582 6 : index_part,
1583 6 : remote_metadata,
1584 6 : TimelineResources {
1585 6 : remote_client,
1586 6 : timeline_get_throttle: self.timeline_get_throttle.clone(),
1587 6 : l0_flush_global_state: self.l0_flush_global_state.clone(),
1588 6 : },
1589 6 : ctx,
1590 6 : )
1591 2 : .await
1592 6 : .with_context(|| {
1593 0 : format!(
1594 0 : "failed to load remote timeline {} for tenant {}",
1595 0 : timeline_id, self.tenant_shard_id
1596 0 : )
1597 6 : })?;
1598 : }
1599 :
1600 : // Walk through deleted timelines, resume deletion
1601 192 : for (timeline_id, index_part, remote_timeline_client) in timelines_to_resume_deletions {
1602 0 : remote_timeline_client
1603 0 : .init_upload_queue_stopped_to_continue_deletion(&index_part)
1604 0 : .context("init queue stopped")
1605 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1606 :
1607 0 : DeleteTimelineFlow::resume_deletion(
1608 0 : Arc::clone(self),
1609 0 : timeline_id,
1610 0 : &index_part.metadata,
1611 0 : remote_timeline_client,
1612 0 : )
1613 0 : .instrument(tracing::info_span!("timeline_delete", %timeline_id))
1614 0 : .await
1615 0 : .context("resume_deletion")
1616 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1617 : }
1618 192 : let needs_manifest_upload =
1619 192 : offloaded_timelines_list.len() != preload.tenant_manifest.offloaded_timelines.len();
1620 192 : {
1621 192 : let mut offloaded_timelines_accessor = self.timelines_offloaded.lock().unwrap();
1622 192 : offloaded_timelines_accessor.extend(offloaded_timelines_list.into_iter());
1623 192 : }
1624 192 : if needs_manifest_upload {
1625 0 : self.store_tenant_manifest().await?;
1626 192 : }
1627 :
1628 : // The local filesystem contents are a cache of what's in the remote IndexPart;
1629 : // IndexPart is the source of truth.
1630 192 : self.clean_up_timelines(&existent_timelines)?;
1631 :
1632 192 : self.gc_block.set_scanned(gc_blocks);
1633 192 :
1634 192 : fail::fail_point!("attach-before-activate", |_| {
1635 0 : anyhow::bail!("attach-before-activate");
1636 192 : });
1637 192 : failpoint_support::sleep_millis_async!("attach-before-activate-sleep", &self.cancel);
1638 :
1639 192 : info!("Done");
1640 :
1641 192 : Ok(())
1642 192 : }
1643 :
1644 : /// Check for any local timeline directories that are temporary, or do not correspond to a
1645 : /// timeline that still exists: this can happen if we crashed during a deletion/creation, or
1646 : /// if a timeline was deleted while the tenant was attached to a different pageserver.
1647 192 : fn clean_up_timelines(&self, existent_timelines: &HashSet<TimelineId>) -> anyhow::Result<()> {
1648 192 : let timelines_dir = self.conf.timelines_path(&self.tenant_shard_id);
1649 :
1650 192 : let entries = match timelines_dir.read_dir_utf8() {
1651 192 : Ok(d) => d,
1652 0 : Err(e) => {
1653 0 : if e.kind() == std::io::ErrorKind::NotFound {
1654 0 : return Ok(());
1655 : } else {
1656 0 : return Err(e).context("list timelines directory for tenant");
1657 : }
1658 : }
1659 : };
1660 :
1661 200 : for entry in entries {
1662 8 : let entry = entry.context("read timeline dir entry")?;
1663 8 : let entry_path = entry.path();
1664 :
1665 8 : let purge = if crate::is_temporary(entry_path)
1666 : // TODO: remove uninit mark code (https://github.com/neondatabase/neon/issues/5718)
1667 8 : || is_uninit_mark(entry_path)
1668 8 : || crate::is_delete_mark(entry_path)
1669 : {
1670 0 : true
1671 : } else {
1672 8 : match TimelineId::try_from(entry_path.file_name()) {
1673 8 : Ok(i) => {
1674 8 : // Purge if the timeline ID does not exist in remote storage: remote storage is the authority.
1675 8 : !existent_timelines.contains(&i)
1676 : }
1677 0 : Err(e) => {
1678 0 : tracing::warn!(
1679 0 : "Unparseable directory in timelines directory: {entry_path}, ignoring ({e})"
1680 : );
1681 : // Do not purge junk: if we don't recognize it, be cautious and leave it for a human.
1682 0 : false
1683 : }
1684 : }
1685 : };
1686 :
1687 8 : if purge {
1688 2 : tracing::info!("Purging stale timeline dentry {entry_path}");
1689 2 : if let Err(e) = match entry.file_type() {
1690 2 : Ok(t) => if t.is_dir() {
1691 2 : std::fs::remove_dir_all(entry_path)
1692 : } else {
1693 0 : std::fs::remove_file(entry_path)
1694 : }
1695 2 : .or_else(fs_ext::ignore_not_found),
1696 0 : Err(e) => Err(e),
1697 : } {
1698 0 : tracing::warn!("Failed to purge stale timeline dentry {entry_path}: {e}");
1699 2 : }
1700 6 : }
1701 : }
1702 :
1703 192 : Ok(())
1704 192 : }
1705 :
1706 : /// Get sum of all remote timelines sizes
1707 : ///
1708 : /// This function relies on the index_part instead of listing the remote storage
1709 0 : pub fn remote_size(&self) -> u64 {
1710 0 : let mut size = 0;
1711 :
1712 0 : for timeline in self.list_timelines() {
1713 0 : size += timeline.remote_client.get_remote_physical_size();
1714 0 : }
1715 :
1716 0 : size
1717 0 : }
1718 :
1719 6 : #[instrument(skip_all, fields(timeline_id=%timeline_id))]
1720 : async fn load_remote_timeline(
1721 : &self,
1722 : timeline_id: TimelineId,
1723 : index_part: IndexPart,
1724 : remote_metadata: TimelineMetadata,
1725 : resources: TimelineResources,
1726 : ctx: &RequestContext,
1727 : ) -> anyhow::Result<()> {
1728 : span::debug_assert_current_span_has_tenant_id();
1729 :
1730 : info!("downloading index file for timeline {}", timeline_id);
1731 : tokio::fs::create_dir_all(self.conf.timeline_path(&self.tenant_shard_id, &timeline_id))
1732 : .await
1733 : .context("Failed to create new timeline directory")?;
1734 :
1735 : let ancestor = if let Some(ancestor_id) = remote_metadata.ancestor_timeline() {
1736 : let timelines = self.timelines.lock().unwrap();
1737 : Some(Arc::clone(timelines.get(&ancestor_id).ok_or_else(
1738 0 : || {
1739 0 : anyhow::anyhow!(
1740 0 : "cannot find ancestor timeline {ancestor_id} for timeline {timeline_id}"
1741 0 : )
1742 0 : },
1743 : )?))
1744 : } else {
1745 : None
1746 : };
1747 :
1748 : self.timeline_init_and_sync(
1749 : timeline_id,
1750 : resources,
1751 : index_part,
1752 : remote_metadata,
1753 : ancestor,
1754 : ctx,
1755 : )
1756 : .await
1757 : }
1758 :
1759 192 : async fn load_timelines_metadata(
1760 192 : self: &Arc<Tenant>,
1761 192 : timeline_ids: HashSet<TimelineId>,
1762 192 : remote_storage: &GenericRemoteStorage,
1763 192 : cancel: CancellationToken,
1764 192 : ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
1765 192 : let mut part_downloads = JoinSet::new();
1766 198 : for timeline_id in timeline_ids {
1767 6 : let cancel_clone = cancel.clone();
1768 6 : part_downloads.spawn(
1769 6 : self.load_timeline_metadata(timeline_id, remote_storage.clone(), cancel_clone)
1770 6 : .instrument(info_span!("download_index_part", %timeline_id)),
1771 : );
1772 : }
1773 :
1774 192 : let mut timeline_preloads: HashMap<TimelineId, TimelinePreload> = HashMap::new();
1775 :
1776 : loop {
1777 198 : tokio::select!(
1778 198 : next = part_downloads.join_next() => {
1779 198 : match next {
1780 6 : Some(result) => {
1781 6 : let preload = result.context("join preload task")?;
1782 6 : timeline_preloads.insert(preload.timeline_id, preload);
1783 : },
1784 : None => {
1785 192 : break;
1786 : }
1787 : }
1788 : },
1789 198 : _ = cancel.cancelled() => {
1790 0 : anyhow::bail!("Cancelled while waiting for remote index download")
1791 : }
1792 : )
1793 : }
1794 :
1795 192 : Ok(timeline_preloads)
1796 192 : }
1797 :
1798 6 : fn build_timeline_client(
1799 6 : &self,
1800 6 : timeline_id: TimelineId,
1801 6 : remote_storage: GenericRemoteStorage,
1802 6 : ) -> RemoteTimelineClient {
1803 6 : RemoteTimelineClient::new(
1804 6 : remote_storage.clone(),
1805 6 : self.deletion_queue_client.clone(),
1806 6 : self.conf,
1807 6 : self.tenant_shard_id,
1808 6 : timeline_id,
1809 6 : self.generation,
1810 6 : )
1811 6 : }
1812 :
1813 6 : fn load_timeline_metadata(
1814 6 : self: &Arc<Tenant>,
1815 6 : timeline_id: TimelineId,
1816 6 : remote_storage: GenericRemoteStorage,
1817 6 : cancel: CancellationToken,
1818 6 : ) -> impl Future<Output = TimelinePreload> {
1819 6 : let client = self.build_timeline_client(timeline_id, remote_storage);
1820 6 : async move {
1821 6 : debug_assert_current_span_has_tenant_and_timeline_id();
1822 6 : debug!("starting index part download");
1823 :
1824 6 : let index_part = client.download_index_file(&cancel).await;
1825 :
1826 6 : debug!("finished index part download");
1827 :
1828 6 : TimelinePreload {
1829 6 : client,
1830 6 : timeline_id,
1831 6 : index_part,
1832 6 : }
1833 6 : }
1834 6 : }
1835 :
1836 0 : fn check_to_be_archived_has_no_unarchived_children(
1837 0 : timeline_id: TimelineId,
1838 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
1839 0 : ) -> Result<(), TimelineArchivalError> {
1840 0 : let children: Vec<TimelineId> = timelines
1841 0 : .iter()
1842 0 : .filter_map(|(id, entry)| {
1843 0 : if entry.get_ancestor_timeline_id() != Some(timeline_id) {
1844 0 : return None;
1845 0 : }
1846 0 : if entry.is_archived() == Some(true) {
1847 0 : return None;
1848 0 : }
1849 0 : Some(*id)
1850 0 : })
1851 0 : .collect();
1852 0 :
1853 0 : if !children.is_empty() {
1854 0 : return Err(TimelineArchivalError::HasUnarchivedChildren(children));
1855 0 : }
1856 0 : Ok(())
1857 0 : }
1858 :
1859 0 : fn check_ancestor_of_to_be_unarchived_is_not_archived(
1860 0 : ancestor_timeline_id: TimelineId,
1861 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
1862 0 : offloaded_timelines: &std::sync::MutexGuard<
1863 0 : '_,
1864 0 : HashMap<TimelineId, Arc<OffloadedTimeline>>,
1865 0 : >,
1866 0 : ) -> Result<(), TimelineArchivalError> {
1867 0 : let has_archived_parent =
1868 0 : if let Some(ancestor_timeline) = timelines.get(&ancestor_timeline_id) {
1869 0 : ancestor_timeline.is_archived() == Some(true)
1870 0 : } else if offloaded_timelines.contains_key(&ancestor_timeline_id) {
1871 0 : true
1872 : } else {
1873 0 : error!("ancestor timeline {ancestor_timeline_id} not found");
1874 0 : if cfg!(debug_assertions) {
1875 0 : panic!("ancestor timeline {ancestor_timeline_id} not found");
1876 0 : }
1877 0 : return Err(TimelineArchivalError::NotFound);
1878 : };
1879 0 : if has_archived_parent {
1880 0 : return Err(TimelineArchivalError::HasArchivedParent(
1881 0 : ancestor_timeline_id,
1882 0 : ));
1883 0 : }
1884 0 : Ok(())
1885 0 : }
1886 :
1887 0 : fn check_to_be_unarchived_timeline_has_no_archived_parent(
1888 0 : timeline: &Arc<Timeline>,
1889 0 : ) -> Result<(), TimelineArchivalError> {
1890 0 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
1891 0 : if ancestor_timeline.is_archived() == Some(true) {
1892 0 : return Err(TimelineArchivalError::HasArchivedParent(
1893 0 : ancestor_timeline.timeline_id,
1894 0 : ));
1895 0 : }
1896 0 : }
1897 0 : Ok(())
1898 0 : }
1899 :
1900 : /// Loads the specified (offloaded) timeline from S3 and attaches it as a loaded timeline
1901 : ///
1902 : /// Counterpart to [`offload_timeline`].
1903 0 : async fn unoffload_timeline(
1904 0 : self: &Arc<Self>,
1905 0 : timeline_id: TimelineId,
1906 0 : broker_client: storage_broker::BrokerClientChannel,
1907 0 : ctx: RequestContext,
1908 0 : ) -> Result<Arc<Timeline>, TimelineArchivalError> {
1909 0 : info!("unoffloading timeline");
1910 :
1911 : // We activate the timeline below manually, so this must be called on an active timeline.
1912 : // We expect callers of this function to ensure this.
1913 0 : match self.current_state() {
1914 : TenantState::Activating { .. }
1915 : | TenantState::Attaching
1916 : | TenantState::Broken { .. } => {
1917 0 : panic!("Timeline expected to be active")
1918 : }
1919 0 : TenantState::Stopping { .. } => return Err(TimelineArchivalError::Cancelled),
1920 0 : TenantState::Active => {}
1921 0 : }
1922 0 : let cancel = self.cancel.clone();
1923 0 :
1924 0 : // Protect against concurrent attempts to use this TimelineId
1925 0 : // We don't care much about idempotency, as it's ensured a layer above.
1926 0 : let allow_offloaded = true;
1927 0 : let _create_guard = self
1928 0 : .create_timeline_create_guard(
1929 0 : timeline_id,
1930 0 : CreateTimelineIdempotency::FailWithConflict,
1931 0 : allow_offloaded,
1932 0 : )
1933 0 : .map_err(|err| match err {
1934 0 : TimelineExclusionError::AlreadyCreating => TimelineArchivalError::AlreadyInProgress,
1935 : TimelineExclusionError::AlreadyExists { .. } => {
1936 0 : TimelineArchivalError::Other(anyhow::anyhow!("Timeline already exists"))
1937 : }
1938 0 : TimelineExclusionError::Other(e) => TimelineArchivalError::Other(e),
1939 0 : })?;
1940 :
1941 0 : let timeline_preload = self
1942 0 : .load_timeline_metadata(timeline_id, self.remote_storage.clone(), cancel.clone())
1943 0 : .await;
1944 :
1945 0 : let index_part = match timeline_preload.index_part {
1946 0 : Ok(index_part) => {
1947 0 : debug!("remote index part exists for timeline {timeline_id}");
1948 0 : index_part
1949 : }
1950 : Err(DownloadError::NotFound) => {
1951 0 : error!(%timeline_id, "index_part not found on remote");
1952 0 : return Err(TimelineArchivalError::NotFound);
1953 : }
1954 0 : Err(DownloadError::Cancelled) => return Err(TimelineArchivalError::Cancelled),
1955 0 : Err(e) => {
1956 0 : // Some (possibly ephemeral) error happened during index_part download.
1957 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1958 0 : return Err(TimelineArchivalError::Other(
1959 0 : anyhow::Error::new(e).context("downloading index_part from remote storage"),
1960 0 : ));
1961 : }
1962 : };
1963 0 : let index_part = match index_part {
1964 0 : MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
1965 0 : MaybeDeletedIndexPart::Deleted(_index_part) => {
1966 0 : info!("timeline is deleted according to index_part.json");
1967 0 : return Err(TimelineArchivalError::NotFound);
1968 : }
1969 : };
1970 0 : let remote_metadata = index_part.metadata.clone();
1971 0 : let timeline_resources = self.build_timeline_resources(timeline_id);
1972 0 : self.load_remote_timeline(
1973 0 : timeline_id,
1974 0 : index_part,
1975 0 : remote_metadata,
1976 0 : timeline_resources,
1977 0 : &ctx,
1978 0 : )
1979 0 : .await
1980 0 : .with_context(|| {
1981 0 : format!(
1982 0 : "failed to load remote timeline {} for tenant {}",
1983 0 : timeline_id, self.tenant_shard_id
1984 0 : )
1985 0 : })
1986 0 : .map_err(TimelineArchivalError::Other)?;
1987 :
1988 0 : let timeline = {
1989 0 : let timelines = self.timelines.lock().unwrap();
1990 0 : let Some(timeline) = timelines.get(&timeline_id) else {
1991 0 : warn!("timeline not available directly after attach");
1992 : // This is not a panic because no locks are held between `load_remote_timeline`
1993 : // which puts the timeline into timelines, and our look into the timeline map.
1994 0 : return Err(TimelineArchivalError::Other(anyhow::anyhow!(
1995 0 : "timeline not available directly after attach"
1996 0 : )));
1997 : };
1998 0 : let mut offloaded_timelines = self.timelines_offloaded.lock().unwrap();
1999 0 : match offloaded_timelines.remove(&timeline_id) {
2000 0 : Some(offloaded) => {
2001 0 : offloaded.delete_from_ancestor_with_timelines(&timelines);
2002 0 : }
2003 0 : None => warn!("timeline already removed from offloaded timelines"),
2004 : }
2005 :
2006 0 : self.initialize_gc_info(&timelines, &offloaded_timelines, Some(timeline_id));
2007 0 :
2008 0 : Arc::clone(timeline)
2009 0 : };
2010 0 :
2011 0 : // Upload new list of offloaded timelines to S3
2012 0 : self.store_tenant_manifest().await?;
2013 :
2014 : // Activate the timeline (if it makes sense)
2015 0 : if !(timeline.is_broken() || timeline.is_stopping()) {
2016 0 : let background_jobs_can_start = None;
2017 0 : timeline.activate(
2018 0 : self.clone(),
2019 0 : broker_client.clone(),
2020 0 : background_jobs_can_start,
2021 0 : &ctx,
2022 0 : );
2023 0 : }
2024 :
2025 0 : info!("timeline unoffloading complete");
2026 0 : Ok(timeline)
2027 0 : }
2028 :
2029 0 : pub(crate) async fn apply_timeline_archival_config(
2030 0 : self: &Arc<Self>,
2031 0 : timeline_id: TimelineId,
2032 0 : new_state: TimelineArchivalState,
2033 0 : broker_client: storage_broker::BrokerClientChannel,
2034 0 : ctx: RequestContext,
2035 0 : ) -> Result<(), TimelineArchivalError> {
2036 0 : info!("setting timeline archival config");
2037 : // First part: figure out what is needed to do, and do validation
2038 0 : let timeline_or_unarchive_offloaded = 'outer: {
2039 0 : let timelines = self.timelines.lock().unwrap();
2040 :
2041 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2042 0 : let offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2043 0 : let Some(offloaded) = offloaded_timelines.get(&timeline_id) else {
2044 0 : return Err(TimelineArchivalError::NotFound);
2045 : };
2046 0 : if new_state == TimelineArchivalState::Archived {
2047 : // It's offloaded already, so nothing to do
2048 0 : return Ok(());
2049 0 : }
2050 0 : if let Some(ancestor_timeline_id) = offloaded.ancestor_timeline_id {
2051 0 : Self::check_ancestor_of_to_be_unarchived_is_not_archived(
2052 0 : ancestor_timeline_id,
2053 0 : &timelines,
2054 0 : &offloaded_timelines,
2055 0 : )?;
2056 0 : }
2057 0 : break 'outer None;
2058 : };
2059 :
2060 : // Do some validation. We release the timelines lock below, so there is potential
2061 : // for race conditions: these checks are more present to prevent misunderstandings of
2062 : // the API's capabilities, instead of serving as the sole way to defend their invariants.
2063 0 : match new_state {
2064 : TimelineArchivalState::Unarchived => {
2065 0 : Self::check_to_be_unarchived_timeline_has_no_archived_parent(timeline)?
2066 : }
2067 : TimelineArchivalState::Archived => {
2068 0 : Self::check_to_be_archived_has_no_unarchived_children(timeline_id, &timelines)?
2069 : }
2070 : }
2071 0 : Some(Arc::clone(timeline))
2072 : };
2073 :
2074 : // Second part: unoffload timeline (if needed)
2075 0 : let timeline = if let Some(timeline) = timeline_or_unarchive_offloaded {
2076 0 : timeline
2077 : } else {
2078 : // Turn offloaded timeline into a non-offloaded one
2079 0 : self.unoffload_timeline(timeline_id, broker_client, ctx)
2080 0 : .await?
2081 : };
2082 :
2083 : // Third part: upload new timeline archival state and block until it is present in S3
2084 0 : let upload_needed = match timeline
2085 0 : .remote_client
2086 0 : .schedule_index_upload_for_timeline_archival_state(new_state)
2087 : {
2088 0 : Ok(upload_needed) => upload_needed,
2089 0 : Err(e) => {
2090 0 : if timeline.cancel.is_cancelled() {
2091 0 : return Err(TimelineArchivalError::Cancelled);
2092 : } else {
2093 0 : return Err(TimelineArchivalError::Other(e));
2094 : }
2095 : }
2096 : };
2097 :
2098 0 : if upload_needed {
2099 0 : info!("Uploading new state");
2100 : const MAX_WAIT: Duration = Duration::from_secs(10);
2101 0 : let Ok(v) =
2102 0 : tokio::time::timeout(MAX_WAIT, timeline.remote_client.wait_completion()).await
2103 : else {
2104 0 : tracing::warn!("reached timeout for waiting on upload queue");
2105 0 : return Err(TimelineArchivalError::Timeout);
2106 : };
2107 0 : v.map_err(|e| match e {
2108 0 : WaitCompletionError::NotInitialized(e) => {
2109 0 : TimelineArchivalError::Other(anyhow::anyhow!(e))
2110 : }
2111 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2112 0 : TimelineArchivalError::Cancelled
2113 : }
2114 0 : })?;
2115 0 : }
2116 0 : Ok(())
2117 0 : }
2118 :
2119 2 : pub fn get_offloaded_timeline(
2120 2 : &self,
2121 2 : timeline_id: TimelineId,
2122 2 : ) -> Result<Arc<OffloadedTimeline>, GetTimelineError> {
2123 2 : self.timelines_offloaded
2124 2 : .lock()
2125 2 : .unwrap()
2126 2 : .get(&timeline_id)
2127 2 : .map(Arc::clone)
2128 2 : .ok_or(GetTimelineError::NotFound {
2129 2 : tenant_id: self.tenant_shard_id,
2130 2 : timeline_id,
2131 2 : })
2132 2 : }
2133 :
2134 4 : pub(crate) fn tenant_shard_id(&self) -> TenantShardId {
2135 4 : self.tenant_shard_id
2136 4 : }
2137 :
2138 : /// Get Timeline handle for given Neon timeline ID.
2139 : /// This function is idempotent. It doesn't change internal state in any way.
2140 222 : pub fn get_timeline(
2141 222 : &self,
2142 222 : timeline_id: TimelineId,
2143 222 : active_only: bool,
2144 222 : ) -> Result<Arc<Timeline>, GetTimelineError> {
2145 222 : let timelines_accessor = self.timelines.lock().unwrap();
2146 222 : let timeline = timelines_accessor
2147 222 : .get(&timeline_id)
2148 222 : .ok_or(GetTimelineError::NotFound {
2149 222 : tenant_id: self.tenant_shard_id,
2150 222 : timeline_id,
2151 222 : })?;
2152 :
2153 220 : if active_only && !timeline.is_active() {
2154 0 : Err(GetTimelineError::NotActive {
2155 0 : tenant_id: self.tenant_shard_id,
2156 0 : timeline_id,
2157 0 : state: timeline.current_state(),
2158 0 : })
2159 : } else {
2160 220 : Ok(Arc::clone(timeline))
2161 : }
2162 222 : }
2163 :
2164 : /// Lists timelines the tenant contains.
2165 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2166 0 : pub fn list_timelines(&self) -> Vec<Arc<Timeline>> {
2167 0 : self.timelines
2168 0 : .lock()
2169 0 : .unwrap()
2170 0 : .values()
2171 0 : .map(Arc::clone)
2172 0 : .collect()
2173 0 : }
2174 :
2175 : /// Lists timelines the tenant manages, including offloaded ones.
2176 : ///
2177 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2178 0 : pub fn list_timelines_and_offloaded(
2179 0 : &self,
2180 0 : ) -> (Vec<Arc<Timeline>>, Vec<Arc<OffloadedTimeline>>) {
2181 0 : let timelines = self
2182 0 : .timelines
2183 0 : .lock()
2184 0 : .unwrap()
2185 0 : .values()
2186 0 : .map(Arc::clone)
2187 0 : .collect();
2188 0 : let offloaded = self
2189 0 : .timelines_offloaded
2190 0 : .lock()
2191 0 : .unwrap()
2192 0 : .values()
2193 0 : .map(Arc::clone)
2194 0 : .collect();
2195 0 : (timelines, offloaded)
2196 0 : }
2197 :
2198 0 : pub fn list_timeline_ids(&self) -> Vec<TimelineId> {
2199 0 : self.timelines.lock().unwrap().keys().cloned().collect()
2200 0 : }
2201 :
2202 : /// This is used by tests & import-from-basebackup.
2203 : ///
2204 : /// The returned [`UninitializedTimeline`] contains no data nor metadata and it is in
2205 : /// a state that will fail [`Tenant::load_remote_timeline`] because `disk_consistent_lsn=Lsn(0)`.
2206 : ///
2207 : /// The caller is responsible for getting the timeline into a state that will be accepted
2208 : /// by [`Tenant::load_remote_timeline`] / [`Tenant::attach`].
2209 : /// Then they may call [`UninitializedTimeline::finish_creation`] to add the timeline
2210 : /// to the [`Tenant::timelines`].
2211 : ///
2212 : /// Tests should use `Tenant::create_test_timeline` to set up the minimum required metadata keys.
2213 184 : pub(crate) async fn create_empty_timeline(
2214 184 : &self,
2215 184 : new_timeline_id: TimelineId,
2216 184 : initdb_lsn: Lsn,
2217 184 : pg_version: u32,
2218 184 : _ctx: &RequestContext,
2219 184 : ) -> anyhow::Result<UninitializedTimeline> {
2220 184 : anyhow::ensure!(
2221 184 : self.is_active(),
2222 0 : "Cannot create empty timelines on inactive tenant"
2223 : );
2224 :
2225 : // Protect against concurrent attempts to use this TimelineId
2226 184 : let create_guard = match self
2227 184 : .start_creating_timeline(new_timeline_id, CreateTimelineIdempotency::FailWithConflict)
2228 179 : .await?
2229 : {
2230 182 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2231 : StartCreatingTimelineResult::Idempotent(_) => {
2232 0 : unreachable!("FailWithConflict implies we get an error instead")
2233 : }
2234 : };
2235 :
2236 182 : let new_metadata = TimelineMetadata::new(
2237 182 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2238 182 : // make it valid, before calling finish_creation()
2239 182 : Lsn(0),
2240 182 : None,
2241 182 : None,
2242 182 : Lsn(0),
2243 182 : initdb_lsn,
2244 182 : initdb_lsn,
2245 182 : pg_version,
2246 182 : );
2247 182 : self.prepare_new_timeline(
2248 182 : new_timeline_id,
2249 182 : &new_metadata,
2250 182 : create_guard,
2251 182 : initdb_lsn,
2252 182 : None,
2253 182 : )
2254 0 : .await
2255 184 : }
2256 :
2257 : /// Helper for unit tests to create an empty timeline.
2258 : ///
2259 : /// The timeline is has state value `Active` but its background loops are not running.
2260 : // This makes the various functions which anyhow::ensure! for Active state work in tests.
2261 : // Our current tests don't need the background loops.
2262 : #[cfg(test)]
2263 174 : pub async fn create_test_timeline(
2264 174 : &self,
2265 174 : new_timeline_id: TimelineId,
2266 174 : initdb_lsn: Lsn,
2267 174 : pg_version: u32,
2268 174 : ctx: &RequestContext,
2269 174 : ) -> anyhow::Result<Arc<Timeline>> {
2270 174 : let uninit_tl = self
2271 174 : .create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2272 171 : .await?;
2273 174 : let tline = uninit_tl.raw_timeline().expect("we just created it");
2274 174 : assert_eq!(tline.get_last_record_lsn(), Lsn(0));
2275 :
2276 : // Setup minimum keys required for the timeline to be usable.
2277 174 : let mut modification = tline.begin_modification(initdb_lsn);
2278 174 : modification
2279 174 : .init_empty_test_timeline()
2280 174 : .context("init_empty_test_timeline")?;
2281 174 : modification
2282 174 : .commit(ctx)
2283 168 : .await
2284 174 : .context("commit init_empty_test_timeline modification")?;
2285 :
2286 : // Flush to disk so that uninit_tl's check for valid disk_consistent_lsn passes.
2287 174 : tline.maybe_spawn_flush_loop();
2288 174 : tline.freeze_and_flush().await.context("freeze_and_flush")?;
2289 :
2290 : // Make sure the freeze_and_flush reaches remote storage.
2291 174 : tline.remote_client.wait_completion().await.unwrap();
2292 :
2293 174 : let tl = uninit_tl.finish_creation()?;
2294 : // The non-test code would call tl.activate() here.
2295 174 : tl.set_state(TimelineState::Active);
2296 174 : Ok(tl)
2297 174 : }
2298 :
2299 : /// Helper for unit tests to create a timeline with some pre-loaded states.
2300 : #[cfg(test)]
2301 : #[allow(clippy::too_many_arguments)]
2302 32 : pub async fn create_test_timeline_with_layers(
2303 32 : &self,
2304 32 : new_timeline_id: TimelineId,
2305 32 : initdb_lsn: Lsn,
2306 32 : pg_version: u32,
2307 32 : ctx: &RequestContext,
2308 32 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
2309 32 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
2310 32 : end_lsn: Lsn,
2311 32 : ) -> anyhow::Result<Arc<Timeline>> {
2312 : use checks::check_valid_layermap;
2313 : use itertools::Itertools;
2314 :
2315 32 : let tline = self
2316 32 : .create_test_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2317 96 : .await?;
2318 32 : tline.force_advance_lsn(end_lsn);
2319 100 : for deltas in delta_layer_desc {
2320 68 : tline
2321 68 : .force_create_delta_layer(deltas, Some(initdb_lsn), ctx)
2322 204 : .await?;
2323 : }
2324 80 : for (lsn, images) in image_layer_desc {
2325 48 : tline
2326 48 : .force_create_image_layer(lsn, images, Some(initdb_lsn), ctx)
2327 294 : .await?;
2328 : }
2329 32 : let layer_names = tline
2330 32 : .layers
2331 32 : .read()
2332 0 : .await
2333 32 : .layer_map()
2334 32 : .unwrap()
2335 32 : .iter_historic_layers()
2336 148 : .map(|layer| layer.layer_name())
2337 32 : .collect_vec();
2338 32 : if let Some(err) = check_valid_layermap(&layer_names) {
2339 0 : bail!("invalid layermap: {err}");
2340 32 : }
2341 32 : Ok(tline)
2342 32 : }
2343 :
2344 : /// Create a new timeline.
2345 : ///
2346 : /// Returns the new timeline ID and reference to its Timeline object.
2347 : ///
2348 : /// If the caller specified the timeline ID to use (`new_timeline_id`), and timeline with
2349 : /// the same timeline ID already exists, returns CreateTimelineError::AlreadyExists.
2350 : #[allow(clippy::too_many_arguments)]
2351 0 : pub(crate) async fn create_timeline(
2352 0 : self: &Arc<Tenant>,
2353 0 : params: CreateTimelineParams,
2354 0 : broker_client: storage_broker::BrokerClientChannel,
2355 0 : ctx: &RequestContext,
2356 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
2357 0 : if !self.is_active() {
2358 0 : if matches!(self.current_state(), TenantState::Stopping { .. }) {
2359 0 : return Err(CreateTimelineError::ShuttingDown);
2360 : } else {
2361 0 : return Err(CreateTimelineError::Other(anyhow::anyhow!(
2362 0 : "Cannot create timelines on inactive tenant"
2363 0 : )));
2364 : }
2365 0 : }
2366 :
2367 0 : let _gate = self
2368 0 : .gate
2369 0 : .enter()
2370 0 : .map_err(|_| CreateTimelineError::ShuttingDown)?;
2371 :
2372 0 : let result: CreateTimelineResult = match params {
2373 : CreateTimelineParams::Bootstrap(CreateTimelineParamsBootstrap {
2374 0 : new_timeline_id,
2375 0 : existing_initdb_timeline_id,
2376 0 : pg_version,
2377 0 : }) => {
2378 0 : self.bootstrap_timeline(
2379 0 : new_timeline_id,
2380 0 : pg_version,
2381 0 : existing_initdb_timeline_id,
2382 0 : ctx,
2383 0 : )
2384 0 : .await?
2385 : }
2386 : CreateTimelineParams::Branch(CreateTimelineParamsBranch {
2387 0 : new_timeline_id,
2388 0 : ancestor_timeline_id,
2389 0 : mut ancestor_start_lsn,
2390 : }) => {
2391 0 : let ancestor_timeline = self
2392 0 : .get_timeline(ancestor_timeline_id, false)
2393 0 : .context("Cannot branch off the timeline that's not present in pageserver")?;
2394 :
2395 : // instead of waiting around, just deny the request because ancestor is not yet
2396 : // ready for other purposes either.
2397 0 : if !ancestor_timeline.is_active() {
2398 0 : return Err(CreateTimelineError::AncestorNotActive);
2399 0 : }
2400 0 :
2401 0 : if ancestor_timeline.is_archived() == Some(true) {
2402 0 : info!("tried to branch archived timeline");
2403 0 : return Err(CreateTimelineError::AncestorArchived);
2404 0 : }
2405 :
2406 0 : if let Some(lsn) = ancestor_start_lsn.as_mut() {
2407 0 : *lsn = lsn.align();
2408 0 :
2409 0 : let ancestor_ancestor_lsn = ancestor_timeline.get_ancestor_lsn();
2410 0 : if ancestor_ancestor_lsn > *lsn {
2411 : // can we safely just branch from the ancestor instead?
2412 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
2413 0 : "invalid start lsn {} for ancestor timeline {}: less than timeline ancestor lsn {}",
2414 0 : lsn,
2415 0 : ancestor_timeline_id,
2416 0 : ancestor_ancestor_lsn,
2417 0 : )));
2418 0 : }
2419 0 :
2420 0 : // Wait for the WAL to arrive and be processed on the parent branch up
2421 0 : // to the requested branch point. The repository code itself doesn't
2422 0 : // require it, but if we start to receive WAL on the new timeline,
2423 0 : // decoding the new WAL might need to look up previous pages, relation
2424 0 : // sizes etc. and that would get confused if the previous page versions
2425 0 : // are not in the repository yet.
2426 0 : ancestor_timeline
2427 0 : .wait_lsn(*lsn, timeline::WaitLsnWaiter::Tenant, ctx)
2428 0 : .await
2429 0 : .map_err(|e| match e {
2430 0 : e @ (WaitLsnError::Timeout(_) | WaitLsnError::BadState { .. }) => {
2431 0 : CreateTimelineError::AncestorLsn(anyhow::anyhow!(e))
2432 : }
2433 0 : WaitLsnError::Shutdown => CreateTimelineError::ShuttingDown,
2434 0 : })?;
2435 0 : }
2436 :
2437 0 : self.branch_timeline(&ancestor_timeline, new_timeline_id, ancestor_start_lsn, ctx)
2438 0 : .await?
2439 : }
2440 : };
2441 :
2442 : // At this point we have dropped our guard on [`Self::timelines_creating`], and
2443 : // the timeline is visible in [`Self::timelines`], but it is _not_ durable yet. We must
2444 : // not send a success to the caller until it is. The same applies to idempotent retries.
2445 : //
2446 : // TODO: the timeline is already visible in [`Self::timelines`]; a caller could incorrectly
2447 : // assume that, because they can see the timeline via API, that the creation is done and
2448 : // that it is durable. Ideally, we would keep the timeline hidden (in [`Self::timelines_creating`])
2449 : // until it is durable, e.g., by extending the time we hold the creation guard. This also
2450 : // interacts with UninitializedTimeline and is generally a bit tricky.
2451 : //
2452 : // To re-emphasize: the only correct way to create a timeline is to repeat calling the
2453 : // creation API until it returns success. Only then is durability guaranteed.
2454 0 : info!(creation_result=%result.discriminant(), "waiting for timeline to be durable");
2455 0 : result
2456 0 : .timeline()
2457 0 : .remote_client
2458 0 : .wait_completion()
2459 0 : .await
2460 0 : .map_err(|e| match e {
2461 : WaitCompletionError::NotInitialized(
2462 0 : e, // If the queue is already stopped, it's a shutdown error.
2463 0 : ) if e.is_stopping() => CreateTimelineError::ShuttingDown,
2464 0 : e => CreateTimelineError::Other(e.into()),
2465 0 : })
2466 0 : .context("wait for timeline initial uploads to complete")?;
2467 :
2468 : // The creating task is responsible for activating the timeline.
2469 : // We do this after `wait_completion()` so that we don't spin up tasks that start
2470 : // doing stuff before the IndexPart is durable in S3, which is done by the previous section.
2471 0 : let activated_timeline = match result {
2472 0 : CreateTimelineResult::Created(timeline) => {
2473 0 : timeline.activate(self.clone(), broker_client, None, ctx);
2474 0 : timeline
2475 : }
2476 0 : CreateTimelineResult::Idempotent(timeline) => {
2477 0 : info!(
2478 0 : "request was deemed idempotent, activation will be done by the creating task"
2479 : );
2480 0 : timeline
2481 : }
2482 : };
2483 :
2484 0 : Ok(activated_timeline)
2485 0 : }
2486 :
2487 0 : pub(crate) async fn delete_timeline(
2488 0 : self: Arc<Self>,
2489 0 : timeline_id: TimelineId,
2490 0 : ) -> Result<(), DeleteTimelineError> {
2491 0 : DeleteTimelineFlow::run(&self, timeline_id).await?;
2492 :
2493 0 : Ok(())
2494 0 : }
2495 :
2496 : /// perform one garbage collection iteration, removing old data files from disk.
2497 : /// this function is periodically called by gc task.
2498 : /// also it can be explicitly requested through page server api 'do_gc' command.
2499 : ///
2500 : /// `target_timeline_id` specifies the timeline to GC, or None for all.
2501 : ///
2502 : /// The `horizon` an `pitr` parameters determine how much WAL history needs to be retained.
2503 : /// Also known as the retention period, or the GC cutoff point. `horizon` specifies
2504 : /// the amount of history, as LSN difference from current latest LSN on each timeline.
2505 : /// `pitr` specifies the same as a time difference from the current time. The effective
2506 : /// GC cutoff point is determined conservatively by either `horizon` and `pitr`, whichever
2507 : /// requires more history to be retained.
2508 : //
2509 754 : pub(crate) async fn gc_iteration(
2510 754 : &self,
2511 754 : target_timeline_id: Option<TimelineId>,
2512 754 : horizon: u64,
2513 754 : pitr: Duration,
2514 754 : cancel: &CancellationToken,
2515 754 : ctx: &RequestContext,
2516 754 : ) -> Result<GcResult, GcError> {
2517 754 : // Don't start doing work during shutdown
2518 754 : if let TenantState::Stopping { .. } = self.current_state() {
2519 0 : return Ok(GcResult::default());
2520 754 : }
2521 754 :
2522 754 : // there is a global allowed_error for this
2523 754 : if !self.is_active() {
2524 0 : return Err(GcError::NotActive);
2525 754 : }
2526 754 :
2527 754 : {
2528 754 : let conf = self.tenant_conf.load();
2529 754 :
2530 754 : if !conf.location.may_delete_layers_hint() {
2531 0 : info!("Skipping GC in location state {:?}", conf.location);
2532 0 : return Ok(GcResult::default());
2533 754 : }
2534 754 :
2535 754 : if conf.is_gc_blocked_by_lsn_lease_deadline() {
2536 750 : info!("Skipping GC because lsn lease deadline is not reached");
2537 750 : return Ok(GcResult::default());
2538 4 : }
2539 : }
2540 :
2541 4 : let _guard = match self.gc_block.start().await {
2542 4 : Ok(guard) => guard,
2543 0 : Err(reasons) => {
2544 0 : info!("Skipping GC: {reasons}");
2545 0 : return Ok(GcResult::default());
2546 : }
2547 : };
2548 :
2549 4 : self.gc_iteration_internal(target_timeline_id, horizon, pitr, cancel, ctx)
2550 4 : .await
2551 754 : }
2552 :
2553 : /// Perform one compaction iteration.
2554 : /// This function is periodically called by compactor task.
2555 : /// Also it can be explicitly requested per timeline through page server
2556 : /// api's 'compact' command.
2557 : ///
2558 : /// Returns whether we have pending compaction task.
2559 0 : async fn compaction_iteration(
2560 0 : self: &Arc<Self>,
2561 0 : cancel: &CancellationToken,
2562 0 : ctx: &RequestContext,
2563 0 : ) -> Result<bool, timeline::CompactionError> {
2564 0 : // Don't start doing work during shutdown, or when broken, we do not need those in the logs
2565 0 : if !self.is_active() {
2566 0 : return Ok(false);
2567 0 : }
2568 0 :
2569 0 : {
2570 0 : let conf = self.tenant_conf.load();
2571 0 : if !conf.location.may_delete_layers_hint() || !conf.location.may_upload_layers_hint() {
2572 0 : info!("Skipping compaction in location state {:?}", conf.location);
2573 0 : return Ok(false);
2574 0 : }
2575 0 : }
2576 0 :
2577 0 : // Scan through the hashmap and collect a list of all the timelines,
2578 0 : // while holding the lock. Then drop the lock and actually perform the
2579 0 : // compactions. We don't want to block everything else while the
2580 0 : // compaction runs.
2581 0 : let timelines_to_compact_or_offload;
2582 0 : {
2583 0 : let timelines = self.timelines.lock().unwrap();
2584 0 : timelines_to_compact_or_offload = timelines
2585 0 : .iter()
2586 0 : .filter_map(|(timeline_id, timeline)| {
2587 0 : let (is_active, (can_offload, _)) =
2588 0 : (timeline.is_active(), timeline.can_offload());
2589 0 : let has_no_unoffloaded_children = {
2590 0 : !timelines
2591 0 : .iter()
2592 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(*timeline_id))
2593 : };
2594 0 : let config_allows_offload = self.conf.timeline_offloading
2595 0 : || self
2596 0 : .tenant_conf
2597 0 : .load()
2598 0 : .tenant_conf
2599 0 : .timeline_offloading
2600 0 : .unwrap_or_default();
2601 0 : let can_offload =
2602 0 : can_offload && has_no_unoffloaded_children && config_allows_offload;
2603 0 : if (is_active, can_offload) == (false, false) {
2604 0 : None
2605 : } else {
2606 0 : Some((*timeline_id, timeline.clone(), (is_active, can_offload)))
2607 : }
2608 0 : })
2609 0 : .collect::<Vec<_>>();
2610 0 : drop(timelines);
2611 0 : }
2612 0 :
2613 0 : // Before doing any I/O work, check our circuit breaker
2614 0 : if self.compaction_circuit_breaker.lock().unwrap().is_broken() {
2615 0 : info!("Skipping compaction due to previous failures");
2616 0 : return Ok(false);
2617 0 : }
2618 0 :
2619 0 : let mut has_pending_task = false;
2620 :
2621 0 : for (timeline_id, timeline, (can_compact, can_offload)) in &timelines_to_compact_or_offload
2622 : {
2623 0 : let pending_task_left = if *can_compact {
2624 : Some(
2625 0 : timeline
2626 0 : .compact(cancel, EnumSet::empty(), ctx)
2627 0 : .instrument(info_span!("compact_timeline", %timeline_id))
2628 0 : .await
2629 0 : .inspect_err(|e| match e {
2630 0 : timeline::CompactionError::ShuttingDown => (),
2631 0 : timeline::CompactionError::Offload(_) => {
2632 0 : // Failures to offload timelines do not trip the circuit breaker, because
2633 0 : // they do not do lots of writes the way compaction itself does: it is cheap
2634 0 : // to retry, and it would be bad to stop all compaction because of an issue with offloading.
2635 0 : }
2636 0 : timeline::CompactionError::Other(e) => {
2637 0 : self.compaction_circuit_breaker
2638 0 : .lock()
2639 0 : .unwrap()
2640 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, e);
2641 0 : }
2642 0 : })?,
2643 : )
2644 : } else {
2645 0 : None
2646 : };
2647 0 : has_pending_task |= pending_task_left.unwrap_or(false);
2648 0 : if pending_task_left == Some(false) && *can_offload {
2649 0 : offload_timeline(self, timeline)
2650 0 : .instrument(info_span!("offload_timeline", %timeline_id))
2651 0 : .await?;
2652 0 : }
2653 : }
2654 :
2655 0 : self.compaction_circuit_breaker
2656 0 : .lock()
2657 0 : .unwrap()
2658 0 : .success(&CIRCUIT_BREAKERS_UNBROKEN);
2659 0 :
2660 0 : Ok(has_pending_task)
2661 0 : }
2662 :
2663 : // Call through to all timelines to freeze ephemeral layers if needed. Usually
2664 : // this happens during ingest: this background housekeeping is for freezing layers
2665 : // that are open but haven't been written to for some time.
2666 0 : async fn ingest_housekeeping(&self) {
2667 0 : // Scan through the hashmap and collect a list of all the timelines,
2668 0 : // while holding the lock. Then drop the lock and actually perform the
2669 0 : // compactions. We don't want to block everything else while the
2670 0 : // compaction runs.
2671 0 : let timelines = {
2672 0 : self.timelines
2673 0 : .lock()
2674 0 : .unwrap()
2675 0 : .values()
2676 0 : .filter_map(|timeline| {
2677 0 : if timeline.is_active() {
2678 0 : Some(timeline.clone())
2679 : } else {
2680 0 : None
2681 : }
2682 0 : })
2683 0 : .collect::<Vec<_>>()
2684 : };
2685 :
2686 0 : for timeline in &timelines {
2687 0 : timeline.maybe_freeze_ephemeral_layer().await;
2688 : }
2689 0 : }
2690 :
2691 0 : pub fn timeline_has_no_attached_children(&self, timeline_id: TimelineId) -> bool {
2692 0 : let timelines = self.timelines.lock().unwrap();
2693 0 : !timelines
2694 0 : .iter()
2695 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(timeline_id))
2696 0 : }
2697 :
2698 1702 : pub fn current_state(&self) -> TenantState {
2699 1702 : self.state.borrow().clone()
2700 1702 : }
2701 :
2702 942 : pub fn is_active(&self) -> bool {
2703 942 : self.current_state() == TenantState::Active
2704 942 : }
2705 :
2706 0 : pub fn generation(&self) -> Generation {
2707 0 : self.generation
2708 0 : }
2709 :
2710 0 : pub(crate) fn wal_redo_manager_status(&self) -> Option<WalRedoManagerStatus> {
2711 0 : self.walredo_mgr.as_ref().and_then(|mgr| mgr.status())
2712 0 : }
2713 :
2714 : /// Changes tenant status to active, unless shutdown was already requested.
2715 : ///
2716 : /// `background_jobs_can_start` is an optional barrier set to a value during pageserver startup
2717 : /// to delay background jobs. Background jobs can be started right away when None is given.
2718 0 : fn activate(
2719 0 : self: &Arc<Self>,
2720 0 : broker_client: BrokerClientChannel,
2721 0 : background_jobs_can_start: Option<&completion::Barrier>,
2722 0 : ctx: &RequestContext,
2723 0 : ) {
2724 0 : span::debug_assert_current_span_has_tenant_id();
2725 0 :
2726 0 : let mut activating = false;
2727 0 : self.state.send_modify(|current_state| {
2728 : use pageserver_api::models::ActivatingFrom;
2729 0 : match &*current_state {
2730 : TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
2731 0 : panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {state:?}", state = current_state);
2732 : }
2733 0 : TenantState::Attaching => {
2734 0 : *current_state = TenantState::Activating(ActivatingFrom::Attaching);
2735 0 : }
2736 0 : }
2737 0 : debug!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), "Activating tenant");
2738 0 : activating = true;
2739 0 : // Continue outside the closure. We need to grab timelines.lock()
2740 0 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
2741 0 : });
2742 0 :
2743 0 : if activating {
2744 0 : let timelines_accessor = self.timelines.lock().unwrap();
2745 0 : let timelines_offloaded_accessor = self.timelines_offloaded.lock().unwrap();
2746 0 : let timelines_to_activate = timelines_accessor
2747 0 : .values()
2748 0 : .filter(|timeline| !(timeline.is_broken() || timeline.is_stopping()));
2749 0 :
2750 0 : // Before activation, populate each Timeline's GcInfo with information about its children
2751 0 : self.initialize_gc_info(&timelines_accessor, &timelines_offloaded_accessor, None);
2752 0 :
2753 0 : // Spawn gc and compaction loops. The loops will shut themselves
2754 0 : // down when they notice that the tenant is inactive.
2755 0 : tasks::start_background_loops(self, background_jobs_can_start);
2756 0 :
2757 0 : let mut activated_timelines = 0;
2758 :
2759 0 : for timeline in timelines_to_activate {
2760 0 : timeline.activate(
2761 0 : self.clone(),
2762 0 : broker_client.clone(),
2763 0 : background_jobs_can_start,
2764 0 : ctx,
2765 0 : );
2766 0 : activated_timelines += 1;
2767 0 : }
2768 :
2769 0 : self.state.send_modify(move |current_state| {
2770 0 : assert!(
2771 0 : matches!(current_state, TenantState::Activating(_)),
2772 0 : "set_stopping and set_broken wait for us to leave Activating state",
2773 : );
2774 0 : *current_state = TenantState::Active;
2775 0 :
2776 0 : let elapsed = self.constructed_at.elapsed();
2777 0 : let total_timelines = timelines_accessor.len();
2778 0 :
2779 0 : // log a lot of stuff, because some tenants sometimes suffer from user-visible
2780 0 : // times to activate. see https://github.com/neondatabase/neon/issues/4025
2781 0 : info!(
2782 0 : since_creation_millis = elapsed.as_millis(),
2783 0 : tenant_id = %self.tenant_shard_id.tenant_id,
2784 0 : shard_id = %self.tenant_shard_id.shard_slug(),
2785 0 : activated_timelines,
2786 0 : total_timelines,
2787 0 : post_state = <&'static str>::from(&*current_state),
2788 0 : "activation attempt finished"
2789 : );
2790 :
2791 0 : TENANT.activation.observe(elapsed.as_secs_f64());
2792 0 : });
2793 0 : }
2794 0 : }
2795 :
2796 : /// Shutdown the tenant and join all of the spawned tasks.
2797 : ///
2798 : /// The method caters for all use-cases:
2799 : /// - pageserver shutdown (freeze_and_flush == true)
2800 : /// - detach + ignore (freeze_and_flush == false)
2801 : ///
2802 : /// This will attempt to shutdown even if tenant is broken.
2803 : ///
2804 : /// `shutdown_progress` is a [`completion::Barrier`] for the shutdown initiated by this call.
2805 : /// If the tenant is already shutting down, we return a clone of the first shutdown call's
2806 : /// `Barrier` as an `Err`. This not-first caller can use the returned barrier to join with
2807 : /// the ongoing shutdown.
2808 6 : async fn shutdown(
2809 6 : &self,
2810 6 : shutdown_progress: completion::Barrier,
2811 6 : shutdown_mode: timeline::ShutdownMode,
2812 6 : ) -> Result<(), completion::Barrier> {
2813 6 : span::debug_assert_current_span_has_tenant_id();
2814 :
2815 : // Set tenant (and its timlines) to Stoppping state.
2816 : //
2817 : // Since we can only transition into Stopping state after activation is complete,
2818 : // run it in a JoinSet so all tenants have a chance to stop before we get SIGKILLed.
2819 : //
2820 : // Transitioning tenants to Stopping state has a couple of non-obvious side effects:
2821 : // 1. Lock out any new requests to the tenants.
2822 : // 2. Signal cancellation to WAL receivers (we wait on it below).
2823 : // 3. Signal cancellation for other tenant background loops.
2824 : // 4. ???
2825 : //
2826 : // The waiting for the cancellation is not done uniformly.
2827 : // We certainly wait for WAL receivers to shut down.
2828 : // That is necessary so that no new data comes in before the freeze_and_flush.
2829 : // But the tenant background loops are joined-on in our caller.
2830 : // It's mesed up.
2831 : // we just ignore the failure to stop
2832 :
2833 : // If we're still attaching, fire the cancellation token early to drop out: this
2834 : // will prevent us flushing, but ensures timely shutdown if some I/O during attach
2835 : // is very slow.
2836 6 : let shutdown_mode = if matches!(self.current_state(), TenantState::Attaching) {
2837 0 : self.cancel.cancel();
2838 0 :
2839 0 : // Having fired our cancellation token, do not try and flush timelines: their cancellation tokens
2840 0 : // are children of ours, so their flush loops will have shut down already
2841 0 : timeline::ShutdownMode::Hard
2842 : } else {
2843 6 : shutdown_mode
2844 : };
2845 :
2846 6 : match self.set_stopping(shutdown_progress, false, false).await {
2847 6 : Ok(()) => {}
2848 0 : Err(SetStoppingError::Broken) => {
2849 0 : // assume that this is acceptable
2850 0 : }
2851 0 : Err(SetStoppingError::AlreadyStopping(other)) => {
2852 0 : // give caller the option to wait for this this shutdown
2853 0 : info!("Tenant::shutdown: AlreadyStopping");
2854 0 : return Err(other);
2855 : }
2856 : };
2857 :
2858 6 : let mut js = tokio::task::JoinSet::new();
2859 6 : {
2860 6 : let timelines = self.timelines.lock().unwrap();
2861 6 : timelines.values().for_each(|timeline| {
2862 6 : let timeline = Arc::clone(timeline);
2863 6 : let timeline_id = timeline.timeline_id;
2864 6 : let span = tracing::info_span!("timeline_shutdown", %timeline_id, ?shutdown_mode);
2865 9 : js.spawn(async move { timeline.shutdown(shutdown_mode).instrument(span).await });
2866 6 : });
2867 6 : }
2868 6 : {
2869 6 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
2870 6 : timelines_offloaded.values().for_each(|timeline| {
2871 0 : timeline.defuse_for_tenant_drop();
2872 6 : });
2873 6 : }
2874 6 : // test_long_timeline_create_then_tenant_delete is leaning on this message
2875 6 : tracing::info!("Waiting for timelines...");
2876 12 : while let Some(res) = js.join_next().await {
2877 0 : match res {
2878 6 : Ok(()) => {}
2879 0 : Err(je) if je.is_cancelled() => unreachable!("no cancelling used"),
2880 0 : Err(je) if je.is_panic() => { /* logged already */ }
2881 0 : Err(je) => warn!("unexpected JoinError: {je:?}"),
2882 : }
2883 : }
2884 :
2885 : // We cancel the Tenant's cancellation token _after_ the timelines have all shut down. This permits
2886 : // them to continue to do work during their shutdown methods, e.g. flushing data.
2887 6 : tracing::debug!("Cancelling CancellationToken");
2888 6 : self.cancel.cancel();
2889 6 :
2890 6 : // shutdown all tenant and timeline tasks: gc, compaction, page service
2891 6 : // No new tasks will be started for this tenant because it's in `Stopping` state.
2892 6 : //
2893 6 : // this will additionally shutdown and await all timeline tasks.
2894 6 : tracing::debug!("Waiting for tasks...");
2895 6 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), None).await;
2896 :
2897 6 : if let Some(walredo_mgr) = self.walredo_mgr.as_ref() {
2898 6 : walredo_mgr.shutdown().await;
2899 0 : }
2900 :
2901 : // Wait for any in-flight operations to complete
2902 6 : self.gate.close().await;
2903 :
2904 6 : remove_tenant_metrics(&self.tenant_shard_id);
2905 6 :
2906 6 : Ok(())
2907 6 : }
2908 :
2909 : /// Change tenant status to Stopping, to mark that it is being shut down.
2910 : ///
2911 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
2912 : ///
2913 : /// This function is not cancel-safe!
2914 : ///
2915 : /// `allow_transition_from_loading` is needed for the special case of loading task deleting the tenant.
2916 : /// `allow_transition_from_attaching` is needed for the special case of attaching deleted tenant.
2917 6 : async fn set_stopping(
2918 6 : &self,
2919 6 : progress: completion::Barrier,
2920 6 : _allow_transition_from_loading: bool,
2921 6 : allow_transition_from_attaching: bool,
2922 6 : ) -> Result<(), SetStoppingError> {
2923 6 : let mut rx = self.state.subscribe();
2924 6 :
2925 6 : // cannot stop before we're done activating, so wait out until we're done activating
2926 6 : rx.wait_for(|state| match state {
2927 0 : TenantState::Attaching if allow_transition_from_attaching => true,
2928 : TenantState::Activating(_) | TenantState::Attaching => {
2929 0 : info!(
2930 0 : "waiting for {} to turn Active|Broken|Stopping",
2931 0 : <&'static str>::from(state)
2932 : );
2933 0 : false
2934 : }
2935 6 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
2936 6 : })
2937 0 : .await
2938 6 : .expect("cannot drop self.state while on a &self method");
2939 6 :
2940 6 : // we now know we're done activating, let's see whether this task is the winner to transition into Stopping
2941 6 : let mut err = None;
2942 6 : let stopping = self.state.send_if_modified(|current_state| match current_state {
2943 : TenantState::Activating(_) => {
2944 0 : unreachable!("1we ensured above that we're done with activation, and, there is no re-activation")
2945 : }
2946 : TenantState::Attaching => {
2947 0 : if !allow_transition_from_attaching {
2948 0 : unreachable!("2we ensured above that we're done with activation, and, there is no re-activation")
2949 0 : };
2950 0 : *current_state = TenantState::Stopping { progress };
2951 0 : true
2952 : }
2953 : TenantState::Active => {
2954 : // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
2955 : // are created after the transition to Stopping. That's harmless, as the Timelines
2956 : // won't be accessible to anyone afterwards, because the Tenant is in Stopping state.
2957 6 : *current_state = TenantState::Stopping { progress };
2958 6 : // Continue stopping outside the closure. We need to grab timelines.lock()
2959 6 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
2960 6 : true
2961 : }
2962 0 : TenantState::Broken { reason, .. } => {
2963 0 : info!(
2964 0 : "Cannot set tenant to Stopping state, it is in Broken state due to: {reason}"
2965 : );
2966 0 : err = Some(SetStoppingError::Broken);
2967 0 : false
2968 : }
2969 0 : TenantState::Stopping { progress } => {
2970 0 : info!("Tenant is already in Stopping state");
2971 0 : err = Some(SetStoppingError::AlreadyStopping(progress.clone()));
2972 0 : false
2973 : }
2974 6 : });
2975 6 : match (stopping, err) {
2976 6 : (true, None) => {} // continue
2977 0 : (false, Some(err)) => return Err(err),
2978 0 : (true, Some(_)) => unreachable!(
2979 0 : "send_if_modified closure must error out if not transitioning to Stopping"
2980 0 : ),
2981 0 : (false, None) => unreachable!(
2982 0 : "send_if_modified closure must return true if transitioning to Stopping"
2983 0 : ),
2984 : }
2985 :
2986 6 : let timelines_accessor = self.timelines.lock().unwrap();
2987 6 : let not_broken_timelines = timelines_accessor
2988 6 : .values()
2989 6 : .filter(|timeline| !timeline.is_broken());
2990 12 : for timeline in not_broken_timelines {
2991 6 : timeline.set_state(TimelineState::Stopping);
2992 6 : }
2993 6 : Ok(())
2994 6 : }
2995 :
2996 : /// Method for tenant::mgr to transition us into Broken state in case of a late failure in
2997 : /// `remove_tenant_from_memory`
2998 : ///
2999 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
3000 : ///
3001 : /// In tests, we also use this to set tenants to Broken state on purpose.
3002 0 : pub(crate) async fn set_broken(&self, reason: String) {
3003 0 : let mut rx = self.state.subscribe();
3004 0 :
3005 0 : // The load & attach routines own the tenant state until it has reached `Active`.
3006 0 : // So, wait until it's done.
3007 0 : rx.wait_for(|state| match state {
3008 : TenantState::Activating(_) | TenantState::Attaching => {
3009 0 : info!(
3010 0 : "waiting for {} to turn Active|Broken|Stopping",
3011 0 : <&'static str>::from(state)
3012 : );
3013 0 : false
3014 : }
3015 0 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3016 0 : })
3017 0 : .await
3018 0 : .expect("cannot drop self.state while on a &self method");
3019 0 :
3020 0 : // we now know we're done activating, let's see whether this task is the winner to transition into Broken
3021 0 : self.set_broken_no_wait(reason)
3022 0 : }
3023 :
3024 0 : pub(crate) fn set_broken_no_wait(&self, reason: impl Display) {
3025 0 : let reason = reason.to_string();
3026 0 : self.state.send_modify(|current_state| {
3027 0 : match *current_state {
3028 : TenantState::Activating(_) | TenantState::Attaching => {
3029 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
3030 : }
3031 : TenantState::Active => {
3032 0 : if cfg!(feature = "testing") {
3033 0 : warn!("Changing Active tenant to Broken state, reason: {}", reason);
3034 0 : *current_state = TenantState::broken_from_reason(reason);
3035 : } else {
3036 0 : unreachable!("not allowed to call set_broken on Active tenants in non-testing builds")
3037 : }
3038 : }
3039 : TenantState::Broken { .. } => {
3040 0 : warn!("Tenant is already in Broken state");
3041 : }
3042 : // This is the only "expected" path, any other path is a bug.
3043 : TenantState::Stopping { .. } => {
3044 0 : warn!(
3045 0 : "Marking Stopping tenant as Broken state, reason: {}",
3046 : reason
3047 : );
3048 0 : *current_state = TenantState::broken_from_reason(reason);
3049 : }
3050 : }
3051 0 : });
3052 0 : }
3053 :
3054 0 : pub fn subscribe_for_state_updates(&self) -> watch::Receiver<TenantState> {
3055 0 : self.state.subscribe()
3056 0 : }
3057 :
3058 : /// The activate_now semaphore is initialized with zero units. As soon as
3059 : /// we add a unit, waiters will be able to acquire a unit and proceed.
3060 0 : pub(crate) fn activate_now(&self) {
3061 0 : self.activate_now_sem.add_permits(1);
3062 0 : }
3063 :
3064 0 : pub(crate) async fn wait_to_become_active(
3065 0 : &self,
3066 0 : timeout: Duration,
3067 0 : ) -> Result<(), GetActiveTenantError> {
3068 0 : let mut receiver = self.state.subscribe();
3069 : loop {
3070 0 : let current_state = receiver.borrow_and_update().clone();
3071 0 : match current_state {
3072 : TenantState::Attaching | TenantState::Activating(_) => {
3073 : // in these states, there's a chance that we can reach ::Active
3074 0 : self.activate_now();
3075 0 : match timeout_cancellable(timeout, &self.cancel, receiver.changed()).await {
3076 0 : Ok(r) => {
3077 0 : r.map_err(
3078 0 : |_e: tokio::sync::watch::error::RecvError|
3079 : // Tenant existed but was dropped: report it as non-existent
3080 0 : GetActiveTenantError::NotFound(GetTenantError::NotFound(self.tenant_shard_id.tenant_id))
3081 0 : )?
3082 : }
3083 : Err(TimeoutCancellableError::Cancelled) => {
3084 0 : return Err(GetActiveTenantError::Cancelled);
3085 : }
3086 : Err(TimeoutCancellableError::Timeout) => {
3087 0 : return Err(GetActiveTenantError::WaitForActiveTimeout {
3088 0 : latest_state: Some(self.current_state()),
3089 0 : wait_time: timeout,
3090 0 : });
3091 : }
3092 : }
3093 : }
3094 : TenantState::Active { .. } => {
3095 0 : return Ok(());
3096 : }
3097 0 : TenantState::Broken { reason, .. } => {
3098 0 : // This is fatal, and reported distinctly from the general case of "will never be active" because
3099 0 : // it's logically a 500 to external API users (broken is always a bug).
3100 0 : return Err(GetActiveTenantError::Broken(reason));
3101 : }
3102 : TenantState::Stopping { .. } => {
3103 : // There's no chance the tenant can transition back into ::Active
3104 0 : return Err(GetActiveTenantError::WillNotBecomeActive(current_state));
3105 : }
3106 : }
3107 : }
3108 0 : }
3109 :
3110 0 : pub(crate) fn get_attach_mode(&self) -> AttachmentMode {
3111 0 : self.tenant_conf.load().location.attach_mode
3112 0 : }
3113 :
3114 : /// For API access: generate a LocationConfig equivalent to the one that would be used to
3115 : /// create a Tenant in the same state. Do not use this in hot paths: it's for relatively
3116 : /// rare external API calls, like a reconciliation at startup.
3117 0 : pub(crate) fn get_location_conf(&self) -> models::LocationConfig {
3118 0 : let conf = self.tenant_conf.load();
3119 :
3120 0 : let location_config_mode = match conf.location.attach_mode {
3121 0 : AttachmentMode::Single => models::LocationConfigMode::AttachedSingle,
3122 0 : AttachmentMode::Multi => models::LocationConfigMode::AttachedMulti,
3123 0 : AttachmentMode::Stale => models::LocationConfigMode::AttachedStale,
3124 : };
3125 :
3126 : // We have a pageserver TenantConf, we need the API-facing TenantConfig.
3127 0 : let tenant_config: models::TenantConfig = conf.tenant_conf.clone().into();
3128 0 :
3129 0 : models::LocationConfig {
3130 0 : mode: location_config_mode,
3131 0 : generation: self.generation.into(),
3132 0 : secondary_conf: None,
3133 0 : shard_number: self.shard_identity.number.0,
3134 0 : shard_count: self.shard_identity.count.literal(),
3135 0 : shard_stripe_size: self.shard_identity.stripe_size.0,
3136 0 : tenant_conf: tenant_config,
3137 0 : }
3138 0 : }
3139 :
3140 0 : pub(crate) fn get_tenant_shard_id(&self) -> &TenantShardId {
3141 0 : &self.tenant_shard_id
3142 0 : }
3143 :
3144 0 : pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
3145 0 : self.shard_identity.stripe_size
3146 0 : }
3147 :
3148 0 : pub(crate) fn get_generation(&self) -> Generation {
3149 0 : self.generation
3150 0 : }
3151 :
3152 : /// This function partially shuts down the tenant (it shuts down the Timelines) and is fallible,
3153 : /// and can leave the tenant in a bad state if it fails. The caller is responsible for
3154 : /// resetting this tenant to a valid state if we fail.
3155 0 : pub(crate) async fn split_prepare(
3156 0 : &self,
3157 0 : child_shards: &Vec<TenantShardId>,
3158 0 : ) -> anyhow::Result<()> {
3159 0 : let (timelines, offloaded) = {
3160 0 : let timelines = self.timelines.lock().unwrap();
3161 0 : let offloaded = self.timelines_offloaded.lock().unwrap();
3162 0 : (timelines.clone(), offloaded.clone())
3163 0 : };
3164 0 : let timelines_iter = timelines
3165 0 : .values()
3166 0 : .map(TimelineOrOffloadedArcRef::<'_>::from)
3167 0 : .chain(
3168 0 : offloaded
3169 0 : .values()
3170 0 : .map(TimelineOrOffloadedArcRef::<'_>::from),
3171 0 : );
3172 0 : for timeline in timelines_iter {
3173 : // We do not block timeline creation/deletion during splits inside the pageserver: it is up to higher levels
3174 : // to ensure that they do not start a split if currently in the process of doing these.
3175 :
3176 0 : let timeline_id = timeline.timeline_id();
3177 :
3178 0 : if let TimelineOrOffloadedArcRef::Timeline(timeline) = timeline {
3179 : // Upload an index from the parent: this is partly to provide freshness for the
3180 : // child tenants that will copy it, and partly for general ease-of-debugging: there will
3181 : // always be a parent shard index in the same generation as we wrote the child shard index.
3182 0 : tracing::info!(%timeline_id, "Uploading index");
3183 0 : timeline
3184 0 : .remote_client
3185 0 : .schedule_index_upload_for_file_changes()?;
3186 0 : timeline.remote_client.wait_completion().await?;
3187 0 : }
3188 :
3189 0 : let remote_client = match timeline {
3190 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.remote_client.clone(),
3191 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => {
3192 0 : let remote_client = self
3193 0 : .build_timeline_client(offloaded.timeline_id, self.remote_storage.clone());
3194 0 : Arc::new(remote_client)
3195 : }
3196 : };
3197 :
3198 : // Shut down the timeline's remote client: this means that the indices we write
3199 : // for child shards will not be invalidated by the parent shard deleting layers.
3200 0 : tracing::info!(%timeline_id, "Shutting down remote storage client");
3201 0 : remote_client.shutdown().await;
3202 :
3203 : // Download methods can still be used after shutdown, as they don't flow through the remote client's
3204 : // queue. In principal the RemoteTimelineClient could provide this without downloading it, but this
3205 : // operation is rare, so it's simpler to just download it (and robustly guarantees that the index
3206 : // we use here really is the remotely persistent one).
3207 0 : tracing::info!(%timeline_id, "Downloading index_part from parent");
3208 0 : let result = remote_client
3209 0 : .download_index_file(&self.cancel)
3210 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))
3211 0 : .await?;
3212 0 : let index_part = match result {
3213 : MaybeDeletedIndexPart::Deleted(_) => {
3214 0 : anyhow::bail!("Timeline deletion happened concurrently with split")
3215 : }
3216 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
3217 : };
3218 :
3219 0 : for child_shard in child_shards {
3220 0 : tracing::info!(%timeline_id, "Uploading index_part for child {}", child_shard.to_index());
3221 0 : upload_index_part(
3222 0 : &self.remote_storage,
3223 0 : child_shard,
3224 0 : &timeline_id,
3225 0 : self.generation,
3226 0 : &index_part,
3227 0 : &self.cancel,
3228 0 : )
3229 0 : .await?;
3230 : }
3231 : }
3232 :
3233 0 : let tenant_manifest = self.build_tenant_manifest();
3234 0 : for child_shard in child_shards {
3235 0 : tracing::info!(
3236 0 : "Uploading tenant manifest for child {}",
3237 0 : child_shard.to_index()
3238 : );
3239 0 : upload_tenant_manifest(
3240 0 : &self.remote_storage,
3241 0 : child_shard,
3242 0 : self.generation,
3243 0 : &tenant_manifest,
3244 0 : &self.cancel,
3245 0 : )
3246 0 : .await?;
3247 : }
3248 :
3249 0 : Ok(())
3250 0 : }
3251 :
3252 0 : pub(crate) fn get_sizes(&self) -> TopTenantShardItem {
3253 0 : let mut result = TopTenantShardItem {
3254 0 : id: self.tenant_shard_id,
3255 0 : resident_size: 0,
3256 0 : physical_size: 0,
3257 0 : max_logical_size: 0,
3258 0 : };
3259 :
3260 0 : for timeline in self.timelines.lock().unwrap().values() {
3261 0 : result.resident_size += timeline.metrics.resident_physical_size_gauge.get();
3262 0 :
3263 0 : result.physical_size += timeline
3264 0 : .remote_client
3265 0 : .metrics
3266 0 : .remote_physical_size_gauge
3267 0 : .get();
3268 0 : result.max_logical_size = std::cmp::max(
3269 0 : result.max_logical_size,
3270 0 : timeline.metrics.current_logical_size_gauge.get(),
3271 0 : );
3272 0 : }
3273 :
3274 0 : result
3275 0 : }
3276 : }
3277 :
3278 : /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
3279 : /// perform a topological sort, so that the parent of each timeline comes
3280 : /// before the children.
3281 : /// E extracts the ancestor from T
3282 : /// This allows for T to be different. It can be TimelineMetadata, can be Timeline itself, etc.
3283 192 : fn tree_sort_timelines<T, E>(
3284 192 : timelines: HashMap<TimelineId, T>,
3285 192 : extractor: E,
3286 192 : ) -> anyhow::Result<Vec<(TimelineId, T)>>
3287 192 : where
3288 192 : E: Fn(&T) -> Option<TimelineId>,
3289 192 : {
3290 192 : let mut result = Vec::with_capacity(timelines.len());
3291 192 :
3292 192 : let mut now = Vec::with_capacity(timelines.len());
3293 192 : // (ancestor, children)
3294 192 : let mut later: HashMap<TimelineId, Vec<(TimelineId, T)>> =
3295 192 : HashMap::with_capacity(timelines.len());
3296 :
3297 198 : for (timeline_id, value) in timelines {
3298 6 : if let Some(ancestor_id) = extractor(&value) {
3299 2 : let children = later.entry(ancestor_id).or_default();
3300 2 : children.push((timeline_id, value));
3301 4 : } else {
3302 4 : now.push((timeline_id, value));
3303 4 : }
3304 : }
3305 :
3306 198 : while let Some((timeline_id, metadata)) = now.pop() {
3307 6 : result.push((timeline_id, metadata));
3308 : // All children of this can be loaded now
3309 6 : if let Some(mut children) = later.remove(&timeline_id) {
3310 2 : now.append(&mut children);
3311 4 : }
3312 : }
3313 :
3314 : // All timelines should be visited now. Unless there were timelines with missing ancestors.
3315 192 : if !later.is_empty() {
3316 0 : for (missing_id, orphan_ids) in later {
3317 0 : for (orphan_id, _) in orphan_ids {
3318 0 : error!("could not load timeline {orphan_id} because its ancestor timeline {missing_id} could not be loaded");
3319 : }
3320 : }
3321 0 : bail!("could not load tenant because some timelines are missing ancestors");
3322 192 : }
3323 192 :
3324 192 : Ok(result)
3325 192 : }
3326 :
3327 : impl Tenant {
3328 0 : pub fn tenant_specific_overrides(&self) -> TenantConfOpt {
3329 0 : self.tenant_conf.load().tenant_conf.clone()
3330 0 : }
3331 :
3332 0 : pub fn effective_config(&self) -> TenantConf {
3333 0 : self.tenant_specific_overrides()
3334 0 : .merge(self.conf.default_tenant_conf.clone())
3335 0 : }
3336 :
3337 0 : pub fn get_checkpoint_distance(&self) -> u64 {
3338 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3339 0 : tenant_conf
3340 0 : .checkpoint_distance
3341 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
3342 0 : }
3343 :
3344 0 : pub fn get_checkpoint_timeout(&self) -> Duration {
3345 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3346 0 : tenant_conf
3347 0 : .checkpoint_timeout
3348 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
3349 0 : }
3350 :
3351 0 : pub fn get_compaction_target_size(&self) -> u64 {
3352 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3353 0 : tenant_conf
3354 0 : .compaction_target_size
3355 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
3356 0 : }
3357 :
3358 0 : pub fn get_compaction_period(&self) -> Duration {
3359 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3360 0 : tenant_conf
3361 0 : .compaction_period
3362 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
3363 0 : }
3364 :
3365 0 : pub fn get_compaction_threshold(&self) -> usize {
3366 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3367 0 : tenant_conf
3368 0 : .compaction_threshold
3369 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
3370 0 : }
3371 :
3372 0 : pub fn get_gc_horizon(&self) -> u64 {
3373 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3374 0 : tenant_conf
3375 0 : .gc_horizon
3376 0 : .unwrap_or(self.conf.default_tenant_conf.gc_horizon)
3377 0 : }
3378 :
3379 0 : pub fn get_gc_period(&self) -> Duration {
3380 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3381 0 : tenant_conf
3382 0 : .gc_period
3383 0 : .unwrap_or(self.conf.default_tenant_conf.gc_period)
3384 0 : }
3385 :
3386 0 : pub fn get_image_creation_threshold(&self) -> usize {
3387 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3388 0 : tenant_conf
3389 0 : .image_creation_threshold
3390 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
3391 0 : }
3392 :
3393 0 : pub fn get_pitr_interval(&self) -> Duration {
3394 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3395 0 : tenant_conf
3396 0 : .pitr_interval
3397 0 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
3398 0 : }
3399 :
3400 0 : pub fn get_min_resident_size_override(&self) -> Option<u64> {
3401 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3402 0 : tenant_conf
3403 0 : .min_resident_size_override
3404 0 : .or(self.conf.default_tenant_conf.min_resident_size_override)
3405 0 : }
3406 :
3407 0 : pub fn get_heatmap_period(&self) -> Option<Duration> {
3408 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3409 0 : let heatmap_period = tenant_conf
3410 0 : .heatmap_period
3411 0 : .unwrap_or(self.conf.default_tenant_conf.heatmap_period);
3412 0 : if heatmap_period.is_zero() {
3413 0 : None
3414 : } else {
3415 0 : Some(heatmap_period)
3416 : }
3417 0 : }
3418 :
3419 4 : pub fn get_lsn_lease_length(&self) -> Duration {
3420 4 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3421 4 : tenant_conf
3422 4 : .lsn_lease_length
3423 4 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
3424 4 : }
3425 :
3426 : /// Generate an up-to-date TenantManifest based on the state of this Tenant.
3427 2 : fn build_tenant_manifest(&self) -> TenantManifest {
3428 2 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
3429 2 :
3430 2 : let mut timeline_manifests = timelines_offloaded
3431 2 : .iter()
3432 2 : .map(|(_timeline_id, offloaded)| offloaded.manifest())
3433 2 : .collect::<Vec<_>>();
3434 2 : // Sort the manifests so that our output is deterministic
3435 2 : timeline_manifests.sort_by_key(|timeline_manifest| timeline_manifest.timeline_id);
3436 2 :
3437 2 : TenantManifest {
3438 2 : version: LATEST_TENANT_MANIFEST_VERSION,
3439 2 : offloaded_timelines: timeline_manifests,
3440 2 : }
3441 2 : }
3442 :
3443 0 : pub fn set_new_tenant_config(&self, new_tenant_conf: TenantConfOpt) {
3444 0 : // Use read-copy-update in order to avoid overwriting the location config
3445 0 : // state if this races with [`Tenant::set_new_location_config`]. Note that
3446 0 : // this race is not possible if both request types come from the storage
3447 0 : // controller (as they should!) because an exclusive op lock is required
3448 0 : // on the storage controller side.
3449 0 : self.tenant_conf.rcu(|inner| {
3450 0 : Arc::new(AttachedTenantConf {
3451 0 : tenant_conf: new_tenant_conf.clone(),
3452 0 : location: inner.location,
3453 0 : // Attached location is not changed, no need to update lsn lease deadline.
3454 0 : lsn_lease_deadline: inner.lsn_lease_deadline,
3455 0 : })
3456 0 : });
3457 0 :
3458 0 : self.tenant_conf_updated(&new_tenant_conf);
3459 0 : // Don't hold self.timelines.lock() during the notifies.
3460 0 : // There's no risk of deadlock right now, but there could be if we consolidate
3461 0 : // mutexes in struct Timeline in the future.
3462 0 : let timelines = self.list_timelines();
3463 0 : for timeline in timelines {
3464 0 : timeline.tenant_conf_updated(&new_tenant_conf);
3465 0 : }
3466 0 : }
3467 :
3468 0 : pub(crate) fn set_new_location_config(&self, new_conf: AttachedTenantConf) {
3469 0 : let new_tenant_conf = new_conf.tenant_conf.clone();
3470 0 :
3471 0 : self.tenant_conf.store(Arc::new(new_conf));
3472 0 :
3473 0 : self.tenant_conf_updated(&new_tenant_conf);
3474 0 : // Don't hold self.timelines.lock() during the notifies.
3475 0 : // There's no risk of deadlock right now, but there could be if we consolidate
3476 0 : // mutexes in struct Timeline in the future.
3477 0 : let timelines = self.list_timelines();
3478 0 : for timeline in timelines {
3479 0 : timeline.tenant_conf_updated(&new_tenant_conf);
3480 0 : }
3481 0 : }
3482 :
3483 192 : fn get_timeline_get_throttle_config(
3484 192 : psconf: &'static PageServerConf,
3485 192 : overrides: &TenantConfOpt,
3486 192 : ) -> throttle::Config {
3487 192 : overrides
3488 192 : .timeline_get_throttle
3489 192 : .clone()
3490 192 : .unwrap_or(psconf.default_tenant_conf.timeline_get_throttle.clone())
3491 192 : }
3492 :
3493 0 : pub(crate) fn tenant_conf_updated(&self, new_conf: &TenantConfOpt) {
3494 0 : let conf = Self::get_timeline_get_throttle_config(self.conf, new_conf);
3495 0 : self.timeline_get_throttle.reconfigure(conf)
3496 0 : }
3497 :
3498 : /// Helper function to create a new Timeline struct.
3499 : ///
3500 : /// The returned Timeline is in Loading state. The caller is responsible for
3501 : /// initializing any on-disk state, and for inserting the Timeline to the 'timelines'
3502 : /// map.
3503 : ///
3504 : /// `validate_ancestor == false` is used when a timeline is created for deletion
3505 : /// and we might not have the ancestor present anymore which is fine for to be
3506 : /// deleted timelines.
3507 418 : fn create_timeline_struct(
3508 418 : &self,
3509 418 : new_timeline_id: TimelineId,
3510 418 : new_metadata: &TimelineMetadata,
3511 418 : ancestor: Option<Arc<Timeline>>,
3512 418 : resources: TimelineResources,
3513 418 : cause: CreateTimelineCause,
3514 418 : create_idempotency: CreateTimelineIdempotency,
3515 418 : ) -> anyhow::Result<Arc<Timeline>> {
3516 418 : let state = match cause {
3517 : CreateTimelineCause::Load => {
3518 418 : let ancestor_id = new_metadata.ancestor_timeline();
3519 418 : anyhow::ensure!(
3520 418 : ancestor_id == ancestor.as_ref().map(|t| t.timeline_id),
3521 0 : "Timeline's {new_timeline_id} ancestor {ancestor_id:?} was not found"
3522 : );
3523 418 : TimelineState::Loading
3524 : }
3525 0 : CreateTimelineCause::Delete => TimelineState::Stopping,
3526 : };
3527 :
3528 418 : let pg_version = new_metadata.pg_version();
3529 418 :
3530 418 : let timeline = Timeline::new(
3531 418 : self.conf,
3532 418 : Arc::clone(&self.tenant_conf),
3533 418 : new_metadata,
3534 418 : ancestor,
3535 418 : new_timeline_id,
3536 418 : self.tenant_shard_id,
3537 418 : self.generation,
3538 418 : self.shard_identity,
3539 418 : self.walredo_mgr.clone(),
3540 418 : resources,
3541 418 : pg_version,
3542 418 : state,
3543 418 : self.attach_wal_lag_cooldown.clone(),
3544 418 : create_idempotency,
3545 418 : self.cancel.child_token(),
3546 418 : );
3547 418 :
3548 418 : Ok(timeline)
3549 418 : }
3550 :
3551 : // Allow too_many_arguments because a constructor's argument list naturally grows with the
3552 : // number of attributes in the struct: breaking these out into a builder wouldn't be helpful.
3553 : #[allow(clippy::too_many_arguments)]
3554 192 : fn new(
3555 192 : state: TenantState,
3556 192 : conf: &'static PageServerConf,
3557 192 : attached_conf: AttachedTenantConf,
3558 192 : shard_identity: ShardIdentity,
3559 192 : walredo_mgr: Option<Arc<WalRedoManager>>,
3560 192 : tenant_shard_id: TenantShardId,
3561 192 : remote_storage: GenericRemoteStorage,
3562 192 : deletion_queue_client: DeletionQueueClient,
3563 192 : l0_flush_global_state: L0FlushGlobalState,
3564 192 : ) -> Tenant {
3565 192 : debug_assert!(
3566 192 : !attached_conf.location.generation.is_none() || conf.control_plane_api.is_none()
3567 : );
3568 :
3569 192 : let (state, mut rx) = watch::channel(state);
3570 192 :
3571 192 : tokio::spawn(async move {
3572 191 : // reflect tenant state in metrics:
3573 191 : // - global per tenant state: TENANT_STATE_METRIC
3574 191 : // - "set" of broken tenants: BROKEN_TENANTS_SET
3575 191 : //
3576 191 : // set of broken tenants should not have zero counts so that it remains accessible for
3577 191 : // alerting.
3578 191 :
3579 191 : let tid = tenant_shard_id.to_string();
3580 191 : let shard_id = tenant_shard_id.shard_slug().to_string();
3581 191 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
3582 :
3583 383 : fn inspect_state(state: &TenantState) -> ([&'static str; 1], bool) {
3584 383 : ([state.into()], matches!(state, TenantState::Broken { .. }))
3585 383 : }
3586 :
3587 191 : let mut tuple = inspect_state(&rx.borrow_and_update());
3588 191 :
3589 191 : let is_broken = tuple.1;
3590 191 : let mut counted_broken = if is_broken {
3591 : // add the id to the set right away, there should not be any updates on the channel
3592 : // after before tenant is removed, if ever
3593 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
3594 0 : true
3595 : } else {
3596 191 : false
3597 : };
3598 :
3599 : loop {
3600 383 : let labels = &tuple.0;
3601 383 : let current = TENANT_STATE_METRIC.with_label_values(labels);
3602 383 : current.inc();
3603 383 :
3604 383 : if rx.changed().await.is_err() {
3605 : // tenant has been dropped
3606 15 : current.dec();
3607 15 : drop(BROKEN_TENANTS_SET.remove_label_values(set_key));
3608 15 : break;
3609 192 : }
3610 192 :
3611 192 : current.dec();
3612 192 : tuple = inspect_state(&rx.borrow_and_update());
3613 192 :
3614 192 : let is_broken = tuple.1;
3615 192 : if is_broken && !counted_broken {
3616 0 : counted_broken = true;
3617 0 : // insert the tenant_id (back) into the set while avoiding needless counter
3618 0 : // access
3619 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
3620 192 : }
3621 : }
3622 192 : });
3623 192 :
3624 192 : Tenant {
3625 192 : tenant_shard_id,
3626 192 : shard_identity,
3627 192 : generation: attached_conf.location.generation,
3628 192 : conf,
3629 192 : // using now here is good enough approximation to catch tenants with really long
3630 192 : // activation times.
3631 192 : constructed_at: Instant::now(),
3632 192 : timelines: Mutex::new(HashMap::new()),
3633 192 : timelines_creating: Mutex::new(HashSet::new()),
3634 192 : timelines_offloaded: Mutex::new(HashMap::new()),
3635 192 : tenant_manifest_upload: Default::default(),
3636 192 : gc_cs: tokio::sync::Mutex::new(()),
3637 192 : walredo_mgr,
3638 192 : remote_storage,
3639 192 : deletion_queue_client,
3640 192 : state,
3641 192 : cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
3642 192 : cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
3643 192 : eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
3644 192 : compaction_circuit_breaker: std::sync::Mutex::new(CircuitBreaker::new(
3645 192 : format!("compaction-{tenant_shard_id}"),
3646 192 : 5,
3647 192 : // Compaction can be a very expensive operation, and might leak disk space. It also ought
3648 192 : // to be infallible, as long as remote storage is available. So if it repeatedly fails,
3649 192 : // use an extremely long backoff.
3650 192 : Some(Duration::from_secs(3600 * 24)),
3651 192 : )),
3652 192 : activate_now_sem: tokio::sync::Semaphore::new(0),
3653 192 : attach_wal_lag_cooldown: Arc::new(std::sync::OnceLock::new()),
3654 192 : cancel: CancellationToken::default(),
3655 192 : gate: Gate::default(),
3656 192 : timeline_get_throttle: Arc::new(throttle::Throttle::new(
3657 192 : Tenant::get_timeline_get_throttle_config(conf, &attached_conf.tenant_conf),
3658 192 : crate::metrics::tenant_throttling::TimelineGet::new(&tenant_shard_id),
3659 192 : )),
3660 192 : tenant_conf: Arc::new(ArcSwap::from_pointee(attached_conf)),
3661 192 : ongoing_timeline_detach: std::sync::Mutex::default(),
3662 192 : gc_block: Default::default(),
3663 192 : l0_flush_global_state,
3664 192 : }
3665 192 : }
3666 :
3667 : /// Locate and load config
3668 0 : pub(super) fn load_tenant_config(
3669 0 : conf: &'static PageServerConf,
3670 0 : tenant_shard_id: &TenantShardId,
3671 0 : ) -> Result<LocationConf, LoadConfigError> {
3672 0 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
3673 0 :
3674 0 : info!("loading tenant configuration from {config_path}");
3675 :
3676 : // load and parse file
3677 0 : let config = fs::read_to_string(&config_path).map_err(|e| {
3678 0 : match e.kind() {
3679 : std::io::ErrorKind::NotFound => {
3680 : // The config should almost always exist for a tenant directory:
3681 : // - When attaching a tenant, the config is the first thing we write
3682 : // - When detaching a tenant, we atomically move the directory to a tmp location
3683 : // before deleting contents.
3684 : //
3685 : // The very rare edge case that can result in a missing config is if we crash during attach
3686 : // between creating directory and writing config. Callers should handle that as if the
3687 : // directory didn't exist.
3688 :
3689 0 : LoadConfigError::NotFound(config_path)
3690 : }
3691 : _ => {
3692 : // No IO errors except NotFound are acceptable here: other kinds of error indicate local storage or permissions issues
3693 : // that we cannot cleanly recover
3694 0 : crate::virtual_file::on_fatal_io_error(&e, "Reading tenant config file")
3695 : }
3696 : }
3697 0 : })?;
3698 :
3699 0 : Ok(toml_edit::de::from_str::<LocationConf>(&config)?)
3700 0 : }
3701 :
3702 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
3703 : pub(super) async fn persist_tenant_config(
3704 : conf: &'static PageServerConf,
3705 : tenant_shard_id: &TenantShardId,
3706 : location_conf: &LocationConf,
3707 : ) -> std::io::Result<()> {
3708 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
3709 :
3710 : Self::persist_tenant_config_at(tenant_shard_id, &config_path, location_conf).await
3711 : }
3712 :
3713 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
3714 : pub(super) async fn persist_tenant_config_at(
3715 : tenant_shard_id: &TenantShardId,
3716 : config_path: &Utf8Path,
3717 : location_conf: &LocationConf,
3718 : ) -> std::io::Result<()> {
3719 : debug!("persisting tenantconf to {config_path}");
3720 :
3721 : let mut conf_content = r#"# This file contains a specific per-tenant's config.
3722 : # It is read in case of pageserver restart.
3723 : "#
3724 : .to_string();
3725 :
3726 0 : fail::fail_point!("tenant-config-before-write", |_| {
3727 0 : Err(std::io::Error::new(
3728 0 : std::io::ErrorKind::Other,
3729 0 : "tenant-config-before-write",
3730 0 : ))
3731 0 : });
3732 :
3733 : // Convert the config to a toml file.
3734 : conf_content +=
3735 : &toml_edit::ser::to_string_pretty(&location_conf).expect("Config serialization failed");
3736 :
3737 : let temp_path = path_with_suffix_extension(config_path, TEMP_FILE_SUFFIX);
3738 :
3739 : let conf_content = conf_content.into_bytes();
3740 : VirtualFile::crashsafe_overwrite(config_path.to_owned(), temp_path, conf_content).await
3741 : }
3742 :
3743 : //
3744 : // How garbage collection works:
3745 : //
3746 : // +--bar------------->
3747 : // /
3748 : // +----+-----foo---------------->
3749 : // /
3750 : // ----main--+-------------------------->
3751 : // \
3752 : // +-----baz-------->
3753 : //
3754 : //
3755 : // 1. Grab 'gc_cs' mutex to prevent new timelines from being created while Timeline's
3756 : // `gc_infos` are being refreshed
3757 : // 2. Scan collected timelines, and on each timeline, make note of the
3758 : // all the points where other timelines have been branched off.
3759 : // We will refrain from removing page versions at those LSNs.
3760 : // 3. For each timeline, scan all layer files on the timeline.
3761 : // Remove all files for which a newer file exists and which
3762 : // don't cover any branch point LSNs.
3763 : //
3764 : // TODO:
3765 : // - if a relation has a non-incremental persistent layer on a child branch, then we
3766 : // don't need to keep that in the parent anymore. But currently
3767 : // we do.
3768 4 : async fn gc_iteration_internal(
3769 4 : &self,
3770 4 : target_timeline_id: Option<TimelineId>,
3771 4 : horizon: u64,
3772 4 : pitr: Duration,
3773 4 : cancel: &CancellationToken,
3774 4 : ctx: &RequestContext,
3775 4 : ) -> Result<GcResult, GcError> {
3776 4 : let mut totals: GcResult = Default::default();
3777 4 : let now = Instant::now();
3778 :
3779 4 : let gc_timelines = self
3780 4 : .refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
3781 4 : .await?;
3782 :
3783 4 : failpoint_support::sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
3784 :
3785 : // If there is nothing to GC, we don't want any messages in the INFO log.
3786 4 : if !gc_timelines.is_empty() {
3787 4 : info!("{} timelines need GC", gc_timelines.len());
3788 : } else {
3789 0 : debug!("{} timelines need GC", gc_timelines.len());
3790 : }
3791 :
3792 : // Perform GC for each timeline.
3793 : //
3794 : // Note that we don't hold the `Tenant::gc_cs` lock here because we don't want to delay the
3795 : // branch creation task, which requires the GC lock. A GC iteration can run concurrently
3796 : // with branch creation.
3797 : //
3798 : // See comments in [`Tenant::branch_timeline`] for more information about why branch
3799 : // creation task can run concurrently with timeline's GC iteration.
3800 8 : for timeline in gc_timelines {
3801 4 : if cancel.is_cancelled() {
3802 : // We were requested to shut down. Stop and return with the progress we
3803 : // made.
3804 0 : break;
3805 4 : }
3806 4 : let result = match timeline.gc().await {
3807 : Err(GcError::TimelineCancelled) => {
3808 0 : if target_timeline_id.is_some() {
3809 : // If we were targetting this specific timeline, surface cancellation to caller
3810 0 : return Err(GcError::TimelineCancelled);
3811 : } else {
3812 : // A timeline may be shutting down independently of the tenant's lifecycle: we should
3813 : // skip past this and proceed to try GC on other timelines.
3814 0 : continue;
3815 : }
3816 : }
3817 4 : r => r?,
3818 : };
3819 4 : totals += result;
3820 : }
3821 :
3822 4 : totals.elapsed = now.elapsed();
3823 4 : Ok(totals)
3824 4 : }
3825 :
3826 : /// Refreshes the Timeline::gc_info for all timelines, returning the
3827 : /// vector of timelines which have [`Timeline::get_last_record_lsn`] past
3828 : /// [`Tenant::get_gc_horizon`].
3829 : ///
3830 : /// This is usually executed as part of periodic gc, but can now be triggered more often.
3831 0 : pub(crate) async fn refresh_gc_info(
3832 0 : &self,
3833 0 : cancel: &CancellationToken,
3834 0 : ctx: &RequestContext,
3835 0 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
3836 0 : // since this method can now be called at different rates than the configured gc loop, it
3837 0 : // might be that these configuration values get applied faster than what it was previously,
3838 0 : // since these were only read from the gc task.
3839 0 : let horizon = self.get_gc_horizon();
3840 0 : let pitr = self.get_pitr_interval();
3841 0 :
3842 0 : // refresh all timelines
3843 0 : let target_timeline_id = None;
3844 0 :
3845 0 : self.refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
3846 0 : .await
3847 0 : }
3848 :
3849 : /// Populate all Timelines' `GcInfo` with information about their children. We do not set the
3850 : /// PITR cutoffs here, because that requires I/O: this is done later, before GC, by [`Self::refresh_gc_info_internal`]
3851 : ///
3852 : /// Subsequently, parent-child relationships are updated incrementally inside [`Timeline::new`] and [`Timeline::drop`].
3853 0 : fn initialize_gc_info(
3854 0 : &self,
3855 0 : timelines: &std::sync::MutexGuard<HashMap<TimelineId, Arc<Timeline>>>,
3856 0 : timelines_offloaded: &std::sync::MutexGuard<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
3857 0 : restrict_to_timeline: Option<TimelineId>,
3858 0 : ) {
3859 0 : if restrict_to_timeline.is_none() {
3860 : // This function must be called before activation: after activation timeline create/delete operations
3861 : // might happen, and this function is not safe to run concurrently with those.
3862 0 : assert!(!self.is_active());
3863 0 : }
3864 :
3865 : // Scan all timelines. For each timeline, remember the timeline ID and
3866 : // the branch point where it was created.
3867 0 : let mut all_branchpoints: BTreeMap<TimelineId, Vec<(Lsn, TimelineId, MaybeOffloaded)>> =
3868 0 : BTreeMap::new();
3869 0 : timelines.iter().for_each(|(timeline_id, timeline_entry)| {
3870 0 : if let Some(ancestor_timeline_id) = &timeline_entry.get_ancestor_timeline_id() {
3871 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
3872 0 : ancestor_children.push((
3873 0 : timeline_entry.get_ancestor_lsn(),
3874 0 : *timeline_id,
3875 0 : MaybeOffloaded::No,
3876 0 : ));
3877 0 : }
3878 0 : });
3879 0 : timelines_offloaded
3880 0 : .iter()
3881 0 : .for_each(|(timeline_id, timeline_entry)| {
3882 0 : let Some(ancestor_timeline_id) = &timeline_entry.ancestor_timeline_id else {
3883 0 : return;
3884 : };
3885 0 : let Some(retain_lsn) = timeline_entry.ancestor_retain_lsn else {
3886 0 : return;
3887 : };
3888 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
3889 0 : ancestor_children.push((retain_lsn, *timeline_id, MaybeOffloaded::Yes));
3890 0 : });
3891 0 :
3892 0 : // The number of bytes we always keep, irrespective of PITR: this is a constant across timelines
3893 0 : let horizon = self.get_gc_horizon();
3894 :
3895 : // Populate each timeline's GcInfo with information about its child branches
3896 0 : let timelines_to_write = if let Some(timeline_id) = restrict_to_timeline {
3897 0 : itertools::Either::Left(timelines.get(&timeline_id).into_iter())
3898 : } else {
3899 0 : itertools::Either::Right(timelines.values())
3900 : };
3901 0 : for timeline in timelines_to_write {
3902 0 : let mut branchpoints: Vec<(Lsn, TimelineId, MaybeOffloaded)> = all_branchpoints
3903 0 : .remove(&timeline.timeline_id)
3904 0 : .unwrap_or_default();
3905 0 :
3906 0 : branchpoints.sort_by_key(|b| b.0);
3907 0 :
3908 0 : let mut target = timeline.gc_info.write().unwrap();
3909 0 :
3910 0 : target.retain_lsns = branchpoints;
3911 0 :
3912 0 : let space_cutoff = timeline
3913 0 : .get_last_record_lsn()
3914 0 : .checked_sub(horizon)
3915 0 : .unwrap_or(Lsn(0));
3916 0 :
3917 0 : target.cutoffs = GcCutoffs {
3918 0 : space: space_cutoff,
3919 0 : time: Lsn::INVALID,
3920 0 : };
3921 0 : }
3922 0 : }
3923 :
3924 4 : async fn refresh_gc_info_internal(
3925 4 : &self,
3926 4 : target_timeline_id: Option<TimelineId>,
3927 4 : horizon: u64,
3928 4 : pitr: Duration,
3929 4 : cancel: &CancellationToken,
3930 4 : ctx: &RequestContext,
3931 4 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
3932 4 : // before taking the gc_cs lock, do the heavier weight finding of gc_cutoff points for
3933 4 : // currently visible timelines.
3934 4 : let timelines = self
3935 4 : .timelines
3936 4 : .lock()
3937 4 : .unwrap()
3938 4 : .values()
3939 4 : .filter(|tl| match target_timeline_id.as_ref() {
3940 4 : Some(target) => &tl.timeline_id == target,
3941 0 : None => true,
3942 4 : })
3943 4 : .cloned()
3944 4 : .collect::<Vec<_>>();
3945 4 :
3946 4 : if target_timeline_id.is_some() && timelines.is_empty() {
3947 : // We were to act on a particular timeline and it wasn't found
3948 0 : return Err(GcError::TimelineNotFound);
3949 4 : }
3950 4 :
3951 4 : let mut gc_cutoffs: HashMap<TimelineId, GcCutoffs> =
3952 4 : HashMap::with_capacity(timelines.len());
3953 :
3954 4 : for timeline in timelines.iter() {
3955 4 : let cutoff = timeline
3956 4 : .get_last_record_lsn()
3957 4 : .checked_sub(horizon)
3958 4 : .unwrap_or(Lsn(0));
3959 :
3960 4 : let cutoffs = timeline.find_gc_cutoffs(cutoff, pitr, cancel, ctx).await?;
3961 4 : let old = gc_cutoffs.insert(timeline.timeline_id, cutoffs);
3962 4 : assert!(old.is_none());
3963 : }
3964 :
3965 4 : if !self.is_active() || self.cancel.is_cancelled() {
3966 0 : return Err(GcError::TenantCancelled);
3967 4 : }
3968 :
3969 : // grab mutex to prevent new timelines from being created here; avoid doing long operations
3970 : // because that will stall branch creation.
3971 4 : let gc_cs = self.gc_cs.lock().await;
3972 :
3973 : // Ok, we now know all the branch points.
3974 : // Update the GC information for each timeline.
3975 4 : let mut gc_timelines = Vec::with_capacity(timelines.len());
3976 8 : for timeline in timelines {
3977 : // We filtered the timeline list above
3978 4 : if let Some(target_timeline_id) = target_timeline_id {
3979 4 : assert_eq!(target_timeline_id, timeline.timeline_id);
3980 0 : }
3981 :
3982 : {
3983 4 : let mut target = timeline.gc_info.write().unwrap();
3984 4 :
3985 4 : // Cull any expired leases
3986 4 : let now = SystemTime::now();
3987 6 : target.leases.retain(|_, lease| !lease.is_expired(&now));
3988 4 :
3989 4 : timeline
3990 4 : .metrics
3991 4 : .valid_lsn_lease_count_gauge
3992 4 : .set(target.leases.len() as u64);
3993 :
3994 : // Look up parent's PITR cutoff to update the child's knowledge of whether it is within parent's PITR
3995 4 : if let Some(ancestor_id) = timeline.get_ancestor_timeline_id() {
3996 0 : if let Some(ancestor_gc_cutoffs) = gc_cutoffs.get(&ancestor_id) {
3997 0 : target.within_ancestor_pitr =
3998 0 : timeline.get_ancestor_lsn() >= ancestor_gc_cutoffs.time;
3999 0 : }
4000 4 : }
4001 :
4002 : // Update metrics that depend on GC state
4003 4 : timeline
4004 4 : .metrics
4005 4 : .archival_size
4006 4 : .set(if target.within_ancestor_pitr {
4007 0 : timeline.metrics.current_logical_size_gauge.get()
4008 : } else {
4009 4 : 0
4010 : });
4011 4 : timeline.metrics.pitr_history_size.set(
4012 4 : timeline
4013 4 : .get_last_record_lsn()
4014 4 : .checked_sub(target.cutoffs.time)
4015 4 : .unwrap_or(Lsn(0))
4016 4 : .0,
4017 4 : );
4018 :
4019 : // Apply the cutoffs we found to the Timeline's GcInfo. Why might we _not_ have cutoffs for a timeline?
4020 : // - this timeline was created while we were finding cutoffs
4021 : // - lsn for timestamp search fails for this timeline repeatedly
4022 4 : if let Some(cutoffs) = gc_cutoffs.get(&timeline.timeline_id) {
4023 4 : target.cutoffs = cutoffs.clone();
4024 4 : }
4025 : }
4026 :
4027 4 : gc_timelines.push(timeline);
4028 : }
4029 4 : drop(gc_cs);
4030 4 : Ok(gc_timelines)
4031 4 : }
4032 :
4033 : /// A substitute for `branch_timeline` for use in unit tests.
4034 : /// The returned timeline will have state value `Active` to make various `anyhow::ensure!()`
4035 : /// calls pass, but, we do not actually call `.activate()` under the hood. So, none of the
4036 : /// timeline background tasks are launched, except the flush loop.
4037 : #[cfg(test)]
4038 232 : async fn branch_timeline_test(
4039 232 : self: &Arc<Self>,
4040 232 : src_timeline: &Arc<Timeline>,
4041 232 : dst_id: TimelineId,
4042 232 : ancestor_lsn: Option<Lsn>,
4043 232 : ctx: &RequestContext,
4044 232 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
4045 232 : let tl = self
4046 232 : .branch_timeline_impl(src_timeline, dst_id, ancestor_lsn, ctx)
4047 197 : .await?
4048 228 : .into_timeline_for_test();
4049 228 : tl.set_state(TimelineState::Active);
4050 228 : Ok(tl)
4051 232 : }
4052 :
4053 : /// Helper for unit tests to branch a timeline with some pre-loaded states.
4054 : #[cfg(test)]
4055 : #[allow(clippy::too_many_arguments)]
4056 6 : pub async fn branch_timeline_test_with_layers(
4057 6 : self: &Arc<Self>,
4058 6 : src_timeline: &Arc<Timeline>,
4059 6 : dst_id: TimelineId,
4060 6 : ancestor_lsn: Option<Lsn>,
4061 6 : ctx: &RequestContext,
4062 6 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
4063 6 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
4064 6 : end_lsn: Lsn,
4065 6 : ) -> anyhow::Result<Arc<Timeline>> {
4066 : use checks::check_valid_layermap;
4067 : use itertools::Itertools;
4068 :
4069 6 : let tline = self
4070 6 : .branch_timeline_test(src_timeline, dst_id, ancestor_lsn, ctx)
4071 5 : .await?;
4072 6 : let ancestor_lsn = if let Some(ancestor_lsn) = ancestor_lsn {
4073 6 : ancestor_lsn
4074 : } else {
4075 0 : tline.get_last_record_lsn()
4076 : };
4077 6 : assert!(end_lsn >= ancestor_lsn);
4078 6 : tline.force_advance_lsn(end_lsn);
4079 12 : for deltas in delta_layer_desc {
4080 6 : tline
4081 6 : .force_create_delta_layer(deltas, Some(ancestor_lsn), ctx)
4082 18 : .await?;
4083 : }
4084 10 : for (lsn, images) in image_layer_desc {
4085 4 : tline
4086 4 : .force_create_image_layer(lsn, images, Some(ancestor_lsn), ctx)
4087 14 : .await?;
4088 : }
4089 6 : let layer_names = tline
4090 6 : .layers
4091 6 : .read()
4092 0 : .await
4093 6 : .layer_map()
4094 6 : .unwrap()
4095 6 : .iter_historic_layers()
4096 10 : .map(|layer| layer.layer_name())
4097 6 : .collect_vec();
4098 6 : if let Some(err) = check_valid_layermap(&layer_names) {
4099 0 : bail!("invalid layermap: {err}");
4100 6 : }
4101 6 : Ok(tline)
4102 6 : }
4103 :
4104 : /// Branch an existing timeline.
4105 0 : async fn branch_timeline(
4106 0 : self: &Arc<Self>,
4107 0 : src_timeline: &Arc<Timeline>,
4108 0 : dst_id: TimelineId,
4109 0 : start_lsn: Option<Lsn>,
4110 0 : ctx: &RequestContext,
4111 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4112 0 : self.branch_timeline_impl(src_timeline, dst_id, start_lsn, ctx)
4113 0 : .await
4114 0 : }
4115 :
4116 232 : async fn branch_timeline_impl(
4117 232 : self: &Arc<Self>,
4118 232 : src_timeline: &Arc<Timeline>,
4119 232 : dst_id: TimelineId,
4120 232 : start_lsn: Option<Lsn>,
4121 232 : _ctx: &RequestContext,
4122 232 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4123 232 : let src_id = src_timeline.timeline_id;
4124 :
4125 : // We will validate our ancestor LSN in this function. Acquire the GC lock so that
4126 : // this check cannot race with GC, and the ancestor LSN is guaranteed to remain
4127 : // valid while we are creating the branch.
4128 232 : let _gc_cs = self.gc_cs.lock().await;
4129 :
4130 : // If no start LSN is specified, we branch the new timeline from the source timeline's last record LSN
4131 232 : let start_lsn = start_lsn.unwrap_or_else(|| {
4132 2 : let lsn = src_timeline.get_last_record_lsn();
4133 2 : info!("branching timeline {dst_id} from timeline {src_id} at last record LSN: {lsn}");
4134 2 : lsn
4135 232 : });
4136 :
4137 : // we finally have determined the ancestor_start_lsn, so we can get claim exclusivity now
4138 232 : let timeline_create_guard = match self
4139 232 : .start_creating_timeline(
4140 232 : dst_id,
4141 232 : CreateTimelineIdempotency::Branch {
4142 232 : ancestor_timeline_id: src_timeline.timeline_id,
4143 232 : ancestor_start_lsn: start_lsn,
4144 232 : },
4145 232 : )
4146 197 : .await?
4147 : {
4148 232 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
4149 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
4150 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
4151 : }
4152 : };
4153 :
4154 : // Ensure that `start_lsn` is valid, i.e. the LSN is within the PITR
4155 : // horizon on the source timeline
4156 : //
4157 : // We check it against both the planned GC cutoff stored in 'gc_info',
4158 : // and the 'latest_gc_cutoff' of the last GC that was performed. The
4159 : // planned GC cutoff in 'gc_info' is normally larger than
4160 : // 'latest_gc_cutoff_lsn', but beware of corner cases like if you just
4161 : // changed the GC settings for the tenant to make the PITR window
4162 : // larger, but some of the data was already removed by an earlier GC
4163 : // iteration.
4164 :
4165 : // check against last actual 'latest_gc_cutoff' first
4166 232 : let latest_gc_cutoff_lsn = src_timeline.get_latest_gc_cutoff_lsn();
4167 232 : src_timeline
4168 232 : .check_lsn_is_in_scope(start_lsn, &latest_gc_cutoff_lsn)
4169 232 : .context(format!(
4170 232 : "invalid branch start lsn: less than latest GC cutoff {}",
4171 232 : *latest_gc_cutoff_lsn,
4172 232 : ))
4173 232 : .map_err(CreateTimelineError::AncestorLsn)?;
4174 :
4175 : // and then the planned GC cutoff
4176 : {
4177 228 : let gc_info = src_timeline.gc_info.read().unwrap();
4178 228 : let cutoff = gc_info.min_cutoff();
4179 228 : if start_lsn < cutoff {
4180 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
4181 0 : "invalid branch start lsn: less than planned GC cutoff {cutoff}"
4182 0 : )));
4183 228 : }
4184 228 : }
4185 228 :
4186 228 : //
4187 228 : // The branch point is valid, and we are still holding the 'gc_cs' lock
4188 228 : // so that GC cannot advance the GC cutoff until we are finished.
4189 228 : // Proceed with the branch creation.
4190 228 : //
4191 228 :
4192 228 : // Determine prev-LSN for the new timeline. We can only determine it if
4193 228 : // the timeline was branched at the current end of the source timeline.
4194 228 : let RecordLsn {
4195 228 : last: src_last,
4196 228 : prev: src_prev,
4197 228 : } = src_timeline.get_last_record_rlsn();
4198 228 : let dst_prev = if src_last == start_lsn {
4199 216 : Some(src_prev)
4200 : } else {
4201 12 : None
4202 : };
4203 :
4204 : // Create the metadata file, noting the ancestor of the new timeline.
4205 : // There is initially no data in it, but all the read-calls know to look
4206 : // into the ancestor.
4207 228 : let metadata = TimelineMetadata::new(
4208 228 : start_lsn,
4209 228 : dst_prev,
4210 228 : Some(src_id),
4211 228 : start_lsn,
4212 228 : *src_timeline.latest_gc_cutoff_lsn.read(), // FIXME: should we hold onto this guard longer?
4213 228 : src_timeline.initdb_lsn,
4214 228 : src_timeline.pg_version,
4215 228 : );
4216 :
4217 228 : let uninitialized_timeline = self
4218 228 : .prepare_new_timeline(
4219 228 : dst_id,
4220 228 : &metadata,
4221 228 : timeline_create_guard,
4222 228 : start_lsn + 1,
4223 228 : Some(Arc::clone(src_timeline)),
4224 228 : )
4225 0 : .await?;
4226 :
4227 228 : let new_timeline = uninitialized_timeline.finish_creation()?;
4228 :
4229 : // Root timeline gets its layers during creation and uploads them along with the metadata.
4230 : // A branch timeline though, when created, can get no writes for some time, hence won't get any layers created.
4231 : // We still need to upload its metadata eagerly: if other nodes `attach` the tenant and miss this timeline, their GC
4232 : // could get incorrect information and remove more layers, than needed.
4233 : // See also https://github.com/neondatabase/neon/issues/3865
4234 228 : new_timeline
4235 228 : .remote_client
4236 228 : .schedule_index_upload_for_full_metadata_update(&metadata)
4237 228 : .context("branch initial metadata upload")?;
4238 :
4239 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
4240 :
4241 228 : Ok(CreateTimelineResult::Created(new_timeline))
4242 232 : }
4243 :
4244 : /// For unit tests, make this visible so that other modules can directly create timelines
4245 : #[cfg(test)]
4246 2 : #[tracing::instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
4247 : pub(crate) async fn bootstrap_timeline_test(
4248 : self: &Arc<Self>,
4249 : timeline_id: TimelineId,
4250 : pg_version: u32,
4251 : load_existing_initdb: Option<TimelineId>,
4252 : ctx: &RequestContext,
4253 : ) -> anyhow::Result<Arc<Timeline>> {
4254 : self.bootstrap_timeline(timeline_id, pg_version, load_existing_initdb, ctx)
4255 : .await
4256 : .map_err(anyhow::Error::new)
4257 2 : .map(|r| r.into_timeline_for_test())
4258 : }
4259 :
4260 : /// Get exclusive access to the timeline ID for creation.
4261 : ///
4262 : /// Timeline-creating code paths must use this function before making changes
4263 : /// to in-memory or persistent state.
4264 : ///
4265 : /// The `state` parameter is a description of the timeline creation operation
4266 : /// we intend to perform.
4267 : /// If the timeline was already created in the meantime, we check whether this
4268 : /// request conflicts or is idempotent , based on `state`.
4269 418 : async fn start_creating_timeline(
4270 418 : &self,
4271 418 : new_timeline_id: TimelineId,
4272 418 : idempotency: CreateTimelineIdempotency,
4273 418 : ) -> Result<StartCreatingTimelineResult<'_>, CreateTimelineError> {
4274 418 : let allow_offloaded = false;
4275 418 : match self.create_timeline_create_guard(new_timeline_id, idempotency, allow_offloaded) {
4276 416 : Ok(create_guard) => {
4277 416 : pausable_failpoint!("timeline-creation-after-uninit");
4278 416 : Ok(StartCreatingTimelineResult::CreateGuard(create_guard))
4279 : }
4280 : Err(TimelineExclusionError::AlreadyCreating) => {
4281 : // Creation is in progress, we cannot create it again, and we cannot
4282 : // check if this request matches the existing one, so caller must try
4283 : // again later.
4284 0 : Err(CreateTimelineError::AlreadyCreating)
4285 : }
4286 0 : Err(TimelineExclusionError::Other(e)) => Err(CreateTimelineError::Other(e)),
4287 : Err(TimelineExclusionError::AlreadyExists {
4288 0 : existing: TimelineOrOffloaded::Offloaded(_existing),
4289 0 : ..
4290 0 : }) => {
4291 0 : info!("timeline already exists but is offloaded");
4292 0 : Err(CreateTimelineError::Conflict)
4293 : }
4294 : Err(TimelineExclusionError::AlreadyExists {
4295 2 : existing: TimelineOrOffloaded::Timeline(existing),
4296 2 : arg,
4297 2 : }) => {
4298 2 : {
4299 2 : let existing = &existing.create_idempotency;
4300 2 : let _span = info_span!("idempotency_check", ?existing, ?arg).entered();
4301 2 : debug!("timeline already exists");
4302 :
4303 2 : match (existing, &arg) {
4304 : // FailWithConflict => no idempotency check
4305 : (CreateTimelineIdempotency::FailWithConflict, _)
4306 : | (_, CreateTimelineIdempotency::FailWithConflict) => {
4307 2 : warn!("timeline already exists, failing request");
4308 2 : return Err(CreateTimelineError::Conflict);
4309 : }
4310 : // Idempotent <=> CreateTimelineIdempotency is identical
4311 0 : (x, y) if x == y => {
4312 0 : info!("timeline already exists and idempotency matches, succeeding request");
4313 : // fallthrough
4314 : }
4315 : (_, _) => {
4316 0 : warn!("idempotency conflict, failing request");
4317 0 : return Err(CreateTimelineError::Conflict);
4318 : }
4319 : }
4320 : }
4321 :
4322 0 : Ok(StartCreatingTimelineResult::Idempotent(existing))
4323 : }
4324 : }
4325 418 : }
4326 :
4327 0 : async fn upload_initdb(
4328 0 : &self,
4329 0 : timelines_path: &Utf8PathBuf,
4330 0 : pgdata_path: &Utf8PathBuf,
4331 0 : timeline_id: &TimelineId,
4332 0 : ) -> anyhow::Result<()> {
4333 0 : let temp_path = timelines_path.join(format!(
4334 0 : "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
4335 0 : ));
4336 0 :
4337 0 : scopeguard::defer! {
4338 0 : if let Err(e) = fs::remove_file(&temp_path) {
4339 0 : error!("Failed to remove temporary initdb archive '{temp_path}': {e}");
4340 0 : }
4341 0 : }
4342 :
4343 0 : let (pgdata_zstd, tar_zst_size) = create_zst_tarball(pgdata_path, &temp_path).await?;
4344 : const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
4345 0 : if tar_zst_size > INITDB_TAR_ZST_WARN_LIMIT {
4346 0 : warn!(
4347 0 : "compressed {temp_path} size of {tar_zst_size} is above limit {INITDB_TAR_ZST_WARN_LIMIT}."
4348 : );
4349 0 : }
4350 :
4351 0 : pausable_failpoint!("before-initdb-upload");
4352 :
4353 0 : backoff::retry(
4354 0 : || async {
4355 0 : self::remote_timeline_client::upload_initdb_dir(
4356 0 : &self.remote_storage,
4357 0 : &self.tenant_shard_id.tenant_id,
4358 0 : timeline_id,
4359 0 : pgdata_zstd.try_clone().await?,
4360 0 : tar_zst_size,
4361 0 : &self.cancel,
4362 : )
4363 0 : .await
4364 0 : },
4365 0 : |_| false,
4366 0 : 3,
4367 0 : u32::MAX,
4368 0 : "persist_initdb_tar_zst",
4369 0 : &self.cancel,
4370 0 : )
4371 0 : .await
4372 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
4373 0 : .and_then(|x| x)
4374 0 : }
4375 :
4376 : /// - run initdb to init temporary instance and get bootstrap data
4377 : /// - after initialization completes, tar up the temp dir and upload it to S3.
4378 2 : async fn bootstrap_timeline(
4379 2 : self: &Arc<Self>,
4380 2 : timeline_id: TimelineId,
4381 2 : pg_version: u32,
4382 2 : load_existing_initdb: Option<TimelineId>,
4383 2 : ctx: &RequestContext,
4384 2 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4385 2 : let timeline_create_guard = match self
4386 2 : .start_creating_timeline(
4387 2 : timeline_id,
4388 2 : CreateTimelineIdempotency::Bootstrap { pg_version },
4389 2 : )
4390 1 : .await?
4391 : {
4392 2 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
4393 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
4394 0 : return Ok(CreateTimelineResult::Idempotent(timeline))
4395 : }
4396 : };
4397 :
4398 : // create a `tenant/{tenant_id}/timelines/basebackup-{timeline_id}.{TEMP_FILE_SUFFIX}/`
4399 : // temporary directory for basebackup files for the given timeline.
4400 :
4401 2 : let timelines_path = self.conf.timelines_path(&self.tenant_shard_id);
4402 2 : let pgdata_path = path_with_suffix_extension(
4403 2 : timelines_path.join(format!("basebackup-{timeline_id}")),
4404 2 : TEMP_FILE_SUFFIX,
4405 2 : );
4406 2 :
4407 2 : // Remove whatever was left from the previous runs: safe because TimelineCreateGuard guarantees
4408 2 : // we won't race with other creations or existent timelines with the same path.
4409 2 : if pgdata_path.exists() {
4410 0 : fs::remove_dir_all(&pgdata_path).with_context(|| {
4411 0 : format!("Failed to remove already existing initdb directory: {pgdata_path}")
4412 0 : })?;
4413 2 : }
4414 :
4415 : // this new directory is very temporary, set to remove it immediately after bootstrap, we don't need it
4416 2 : scopeguard::defer! {
4417 2 : if let Err(e) = fs::remove_dir_all(&pgdata_path) {
4418 2 : // this is unlikely, but we will remove the directory on pageserver restart or another bootstrap call
4419 2 : error!("Failed to remove temporary initdb directory '{pgdata_path}': {e}");
4420 2 : }
4421 2 : }
4422 2 : if let Some(existing_initdb_timeline_id) = load_existing_initdb {
4423 2 : if existing_initdb_timeline_id != timeline_id {
4424 0 : let source_path = &remote_initdb_archive_path(
4425 0 : &self.tenant_shard_id.tenant_id,
4426 0 : &existing_initdb_timeline_id,
4427 0 : );
4428 0 : let dest_path =
4429 0 : &remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &timeline_id);
4430 0 :
4431 0 : // if this fails, it will get retried by retried control plane requests
4432 0 : self.remote_storage
4433 0 : .copy_object(source_path, dest_path, &self.cancel)
4434 0 : .await
4435 0 : .context("copy initdb tar")?;
4436 2 : }
4437 2 : let (initdb_tar_zst_path, initdb_tar_zst) =
4438 2 : self::remote_timeline_client::download_initdb_tar_zst(
4439 2 : self.conf,
4440 2 : &self.remote_storage,
4441 2 : &self.tenant_shard_id,
4442 2 : &existing_initdb_timeline_id,
4443 2 : &self.cancel,
4444 2 : )
4445 571 : .await
4446 2 : .context("download initdb tar")?;
4447 :
4448 2 : scopeguard::defer! {
4449 2 : if let Err(e) = fs::remove_file(&initdb_tar_zst_path) {
4450 2 : error!("Failed to remove temporary initdb archive '{initdb_tar_zst_path}': {e}");
4451 2 : }
4452 2 : }
4453 2 :
4454 2 : let buf_read =
4455 2 : BufReader::with_capacity(remote_timeline_client::BUFFER_SIZE, initdb_tar_zst);
4456 2 : extract_zst_tarball(&pgdata_path, buf_read)
4457 11175 : .await
4458 2 : .context("extract initdb tar")?;
4459 : } else {
4460 : // Init temporarily repo to get bootstrap data, this creates a directory in the `pgdata_path` path
4461 0 : run_initdb(self.conf, &pgdata_path, pg_version, &self.cancel)
4462 0 : .await
4463 0 : .context("run initdb")?;
4464 :
4465 : // Upload the created data dir to S3
4466 0 : if self.tenant_shard_id().is_shard_zero() {
4467 0 : self.upload_initdb(&timelines_path, &pgdata_path, &timeline_id)
4468 0 : .await?;
4469 0 : }
4470 : }
4471 2 : let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
4472 2 :
4473 2 : // Import the contents of the data directory at the initial checkpoint
4474 2 : // LSN, and any WAL after that.
4475 2 : // Initdb lsn will be equal to last_record_lsn which will be set after import.
4476 2 : // Because we know it upfront avoid having an option or dummy zero value by passing it to the metadata.
4477 2 : let new_metadata = TimelineMetadata::new(
4478 2 : Lsn(0),
4479 2 : None,
4480 2 : None,
4481 2 : Lsn(0),
4482 2 : pgdata_lsn,
4483 2 : pgdata_lsn,
4484 2 : pg_version,
4485 2 : );
4486 2 : let raw_timeline = self
4487 2 : .prepare_new_timeline(
4488 2 : timeline_id,
4489 2 : &new_metadata,
4490 2 : timeline_create_guard,
4491 2 : pgdata_lsn,
4492 2 : None,
4493 2 : )
4494 0 : .await?;
4495 :
4496 2 : let tenant_shard_id = raw_timeline.owning_tenant.tenant_shard_id;
4497 2 : let unfinished_timeline = raw_timeline.raw_timeline()?;
4498 :
4499 : // Flush the new layer files to disk, before we make the timeline as available to
4500 : // the outside world.
4501 : //
4502 : // Flush loop needs to be spawned in order to be able to flush.
4503 2 : unfinished_timeline.maybe_spawn_flush_loop();
4504 2 :
4505 2 : import_datadir::import_timeline_from_postgres_datadir(
4506 2 : unfinished_timeline,
4507 2 : &pgdata_path,
4508 2 : pgdata_lsn,
4509 2 : ctx,
4510 2 : )
4511 8835 : .await
4512 2 : .with_context(|| {
4513 0 : format!("Failed to import pgdatadir for timeline {tenant_shard_id}/{timeline_id}")
4514 2 : })?;
4515 :
4516 2 : fail::fail_point!("before-checkpoint-new-timeline", |_| {
4517 0 : Err(CreateTimelineError::Other(anyhow::anyhow!(
4518 0 : "failpoint before-checkpoint-new-timeline"
4519 0 : )))
4520 2 : });
4521 :
4522 2 : unfinished_timeline
4523 2 : .freeze_and_flush()
4524 2 : .await
4525 2 : .with_context(|| {
4526 0 : format!(
4527 0 : "Failed to flush after pgdatadir import for timeline {tenant_shard_id}/{timeline_id}"
4528 0 : )
4529 2 : })?;
4530 :
4531 : // All done!
4532 2 : let timeline = raw_timeline.finish_creation()?;
4533 :
4534 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
4535 :
4536 2 : Ok(CreateTimelineResult::Created(timeline))
4537 2 : }
4538 :
4539 412 : fn build_timeline_remote_client(&self, timeline_id: TimelineId) -> RemoteTimelineClient {
4540 412 : RemoteTimelineClient::new(
4541 412 : self.remote_storage.clone(),
4542 412 : self.deletion_queue_client.clone(),
4543 412 : self.conf,
4544 412 : self.tenant_shard_id,
4545 412 : timeline_id,
4546 412 : self.generation,
4547 412 : )
4548 412 : }
4549 :
4550 : /// Call this before constructing a timeline, to build its required structures
4551 412 : fn build_timeline_resources(&self, timeline_id: TimelineId) -> TimelineResources {
4552 412 : TimelineResources {
4553 412 : remote_client: self.build_timeline_remote_client(timeline_id),
4554 412 : timeline_get_throttle: self.timeline_get_throttle.clone(),
4555 412 : l0_flush_global_state: self.l0_flush_global_state.clone(),
4556 412 : }
4557 412 : }
4558 :
4559 : /// Creates intermediate timeline structure and its files.
4560 : ///
4561 : /// An empty layer map is initialized, and new data and WAL can be imported starting
4562 : /// at 'disk_consistent_lsn'. After any initial data has been imported, call
4563 : /// `finish_creation` to insert the Timeline into the timelines map.
4564 412 : async fn prepare_new_timeline<'a>(
4565 412 : &'a self,
4566 412 : new_timeline_id: TimelineId,
4567 412 : new_metadata: &TimelineMetadata,
4568 412 : create_guard: TimelineCreateGuard<'a>,
4569 412 : start_lsn: Lsn,
4570 412 : ancestor: Option<Arc<Timeline>>,
4571 412 : ) -> anyhow::Result<UninitializedTimeline<'a>> {
4572 412 : let tenant_shard_id = self.tenant_shard_id;
4573 412 :
4574 412 : let resources = self.build_timeline_resources(new_timeline_id);
4575 412 : resources
4576 412 : .remote_client
4577 412 : .init_upload_queue_for_empty_remote(new_metadata)?;
4578 :
4579 412 : let timeline_struct = self
4580 412 : .create_timeline_struct(
4581 412 : new_timeline_id,
4582 412 : new_metadata,
4583 412 : ancestor,
4584 412 : resources,
4585 412 : CreateTimelineCause::Load,
4586 412 : create_guard.idempotency.clone(),
4587 412 : )
4588 412 : .context("Failed to create timeline data structure")?;
4589 :
4590 412 : timeline_struct.init_empty_layer_map(start_lsn);
4591 :
4592 412 : if let Err(e) = self
4593 412 : .create_timeline_files(&create_guard.timeline_path)
4594 0 : .await
4595 : {
4596 0 : error!("Failed to create initial files for timeline {tenant_shard_id}/{new_timeline_id}, cleaning up: {e:?}");
4597 0 : cleanup_timeline_directory(create_guard);
4598 0 : return Err(e);
4599 412 : }
4600 412 :
4601 412 : debug!(
4602 0 : "Successfully created initial files for timeline {tenant_shard_id}/{new_timeline_id}"
4603 : );
4604 :
4605 412 : Ok(UninitializedTimeline::new(
4606 412 : self,
4607 412 : new_timeline_id,
4608 412 : Some((timeline_struct, create_guard)),
4609 412 : ))
4610 412 : }
4611 :
4612 412 : async fn create_timeline_files(&self, timeline_path: &Utf8Path) -> anyhow::Result<()> {
4613 412 : crashsafe::create_dir(timeline_path).context("Failed to create timeline directory")?;
4614 :
4615 412 : fail::fail_point!("after-timeline-dir-creation", |_| {
4616 0 : anyhow::bail!("failpoint after-timeline-dir-creation");
4617 412 : });
4618 :
4619 412 : Ok(())
4620 412 : }
4621 :
4622 : /// Get a guard that provides exclusive access to the timeline directory, preventing
4623 : /// concurrent attempts to create the same timeline.
4624 : ///
4625 : /// The `allow_offloaded` parameter controls whether to tolerate the existence of
4626 : /// offloaded timelines or not.
4627 418 : fn create_timeline_create_guard(
4628 418 : &self,
4629 418 : timeline_id: TimelineId,
4630 418 : idempotency: CreateTimelineIdempotency,
4631 418 : allow_offloaded: bool,
4632 418 : ) -> Result<TimelineCreateGuard, TimelineExclusionError> {
4633 418 : let tenant_shard_id = self.tenant_shard_id;
4634 418 :
4635 418 : let timeline_path = self.conf.timeline_path(&tenant_shard_id, &timeline_id);
4636 :
4637 418 : let create_guard = TimelineCreateGuard::new(
4638 418 : self,
4639 418 : timeline_id,
4640 418 : timeline_path.clone(),
4641 418 : idempotency,
4642 418 : allow_offloaded,
4643 418 : )?;
4644 :
4645 : // At this stage, we have got exclusive access to in-memory state for this timeline ID
4646 : // for creation.
4647 : // A timeline directory should never exist on disk already:
4648 : // - a previous failed creation would have cleaned up after itself
4649 : // - a pageserver restart would clean up timeline directories that don't have valid remote state
4650 : //
4651 : // Therefore it is an unexpected internal error to encounter a timeline directory already existing here,
4652 : // this error may indicate a bug in cleanup on failed creations.
4653 416 : if timeline_path.exists() {
4654 0 : return Err(TimelineExclusionError::Other(anyhow::anyhow!(
4655 0 : "Timeline directory already exists! This is a bug."
4656 0 : )));
4657 416 : }
4658 416 :
4659 416 : Ok(create_guard)
4660 418 : }
4661 :
4662 : /// Gathers inputs from all of the timelines to produce a sizing model input.
4663 : ///
4664 : /// Future is cancellation safe. Only one calculation can be running at once per tenant.
4665 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
4666 : pub async fn gather_size_inputs(
4667 : &self,
4668 : // `max_retention_period` overrides the cutoff that is used to calculate the size
4669 : // (only if it is shorter than the real cutoff).
4670 : max_retention_period: Option<u64>,
4671 : cause: LogicalSizeCalculationCause,
4672 : cancel: &CancellationToken,
4673 : ctx: &RequestContext,
4674 : ) -> Result<size::ModelInputs, size::CalculateSyntheticSizeError> {
4675 : let logical_sizes_at_once = self
4676 : .conf
4677 : .concurrent_tenant_size_logical_size_queries
4678 : .inner();
4679 :
4680 : // TODO: Having a single mutex block concurrent reads is not great for performance.
4681 : //
4682 : // But the only case where we need to run multiple of these at once is when we
4683 : // request a size for a tenant manually via API, while another background calculation
4684 : // is in progress (which is not a common case).
4685 : //
4686 : // See more for on the issue #2748 condenced out of the initial PR review.
4687 : let mut shared_cache = tokio::select! {
4688 : locked = self.cached_logical_sizes.lock() => locked,
4689 : _ = cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
4690 : _ = self.cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
4691 : };
4692 :
4693 : size::gather_inputs(
4694 : self,
4695 : logical_sizes_at_once,
4696 : max_retention_period,
4697 : &mut shared_cache,
4698 : cause,
4699 : cancel,
4700 : ctx,
4701 : )
4702 : .await
4703 : }
4704 :
4705 : /// Calculate synthetic tenant size and cache the result.
4706 : /// This is periodically called by background worker.
4707 : /// result is cached in tenant struct
4708 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
4709 : pub async fn calculate_synthetic_size(
4710 : &self,
4711 : cause: LogicalSizeCalculationCause,
4712 : cancel: &CancellationToken,
4713 : ctx: &RequestContext,
4714 : ) -> Result<u64, size::CalculateSyntheticSizeError> {
4715 : let inputs = self.gather_size_inputs(None, cause, cancel, ctx).await?;
4716 :
4717 : let size = inputs.calculate();
4718 :
4719 : self.set_cached_synthetic_size(size);
4720 :
4721 : Ok(size)
4722 : }
4723 :
4724 : /// Cache given synthetic size and update the metric value
4725 0 : pub fn set_cached_synthetic_size(&self, size: u64) {
4726 0 : self.cached_synthetic_tenant_size
4727 0 : .store(size, Ordering::Relaxed);
4728 0 :
4729 0 : // Only shard zero should be calculating synthetic sizes
4730 0 : debug_assert!(self.shard_identity.is_shard_zero());
4731 :
4732 0 : TENANT_SYNTHETIC_SIZE_METRIC
4733 0 : .get_metric_with_label_values(&[&self.tenant_shard_id.tenant_id.to_string()])
4734 0 : .unwrap()
4735 0 : .set(size);
4736 0 : }
4737 :
4738 0 : pub fn cached_synthetic_size(&self) -> u64 {
4739 0 : self.cached_synthetic_tenant_size.load(Ordering::Relaxed)
4740 0 : }
4741 :
4742 : /// Flush any in-progress layers, schedule uploads, and wait for uploads to complete.
4743 : ///
4744 : /// This function can take a long time: callers should wrap it in a timeout if calling
4745 : /// from an external API handler.
4746 : ///
4747 : /// Cancel-safety: cancelling this function may leave I/O running, but such I/O is
4748 : /// still bounded by tenant/timeline shutdown.
4749 0 : #[tracing::instrument(skip_all)]
4750 : pub(crate) async fn flush_remote(&self) -> anyhow::Result<()> {
4751 : let timelines = self.timelines.lock().unwrap().clone();
4752 :
4753 0 : async fn flush_timeline(_gate: GateGuard, timeline: Arc<Timeline>) -> anyhow::Result<()> {
4754 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Flushing...");
4755 0 : timeline.freeze_and_flush().await?;
4756 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Waiting for uploads...");
4757 0 : timeline.remote_client.wait_completion().await?;
4758 :
4759 0 : Ok(())
4760 0 : }
4761 :
4762 : // We do not use a JoinSet for these tasks, because we don't want them to be
4763 : // aborted when this function's future is cancelled: they should stay alive
4764 : // holding their GateGuard until they complete, to ensure their I/Os complete
4765 : // before Timeline shutdown completes.
4766 : let mut results = FuturesUnordered::new();
4767 :
4768 : for (_timeline_id, timeline) in timelines {
4769 : // Run each timeline's flush in a task holding the timeline's gate: this
4770 : // means that if this function's future is cancelled, the Timeline shutdown
4771 : // will still wait for any I/O in here to complete.
4772 : let Ok(gate) = timeline.gate.enter() else {
4773 : continue;
4774 : };
4775 0 : let jh = tokio::task::spawn(async move { flush_timeline(gate, timeline).await });
4776 : results.push(jh);
4777 : }
4778 :
4779 : while let Some(r) = results.next().await {
4780 : if let Err(e) = r {
4781 : if !e.is_cancelled() && !e.is_panic() {
4782 : tracing::error!("unexpected join error: {e:?}");
4783 : }
4784 : }
4785 : }
4786 :
4787 : // The flushes we did above were just writes, but the Tenant might have had
4788 : // pending deletions as well from recent compaction/gc: we want to flush those
4789 : // as well. This requires flushing the global delete queue. This is cheap
4790 : // because it's typically a no-op.
4791 : match self.deletion_queue_client.flush_execute().await {
4792 : Ok(_) => {}
4793 : Err(DeletionQueueError::ShuttingDown) => {}
4794 : }
4795 :
4796 : Ok(())
4797 : }
4798 :
4799 0 : pub(crate) fn get_tenant_conf(&self) -> TenantConfOpt {
4800 0 : self.tenant_conf.load().tenant_conf.clone()
4801 0 : }
4802 :
4803 : /// How much local storage would this tenant like to have? It can cope with
4804 : /// less than this (via eviction and on-demand downloads), but this function enables
4805 : /// the Tenant to advertise how much storage it would prefer to have to provide fast I/O
4806 : /// by keeping important things on local disk.
4807 : ///
4808 : /// This is a heuristic, not a guarantee: tenants that are long-idle will actually use less
4809 : /// than they report here, due to layer eviction. Tenants with many active branches may
4810 : /// actually use more than they report here.
4811 0 : pub(crate) fn local_storage_wanted(&self) -> u64 {
4812 0 : let timelines = self.timelines.lock().unwrap();
4813 0 :
4814 0 : // Heuristic: we use the max() of the timelines' visible sizes, rather than the sum. This
4815 0 : // reflects the observation that on tenants with multiple large branches, typically only one
4816 0 : // of them is used actively enough to occupy space on disk.
4817 0 : timelines
4818 0 : .values()
4819 0 : .map(|t| t.metrics.visible_physical_size_gauge.get())
4820 0 : .max()
4821 0 : .unwrap_or(0)
4822 0 : }
4823 :
4824 : /// Serialize and write the latest TenantManifest to remote storage.
4825 2 : pub(crate) async fn store_tenant_manifest(&self) -> Result<(), TenantManifestError> {
4826 : // Only one manifest write may be done at at time, and the contents of the manifest
4827 : // must be loaded while holding this lock. This makes it safe to call this function
4828 : // from anywhere without worrying about colliding updates.
4829 2 : let mut guard = tokio::select! {
4830 2 : g = self.tenant_manifest_upload.lock() => {
4831 2 : g
4832 : },
4833 2 : _ = self.cancel.cancelled() => {
4834 0 : return Err(TenantManifestError::Cancelled);
4835 : }
4836 : };
4837 :
4838 2 : let manifest = self.build_tenant_manifest();
4839 2 : if Some(&manifest) == (*guard).as_ref() {
4840 : // Optimisation: skip uploads that don't change anything.
4841 0 : return Ok(());
4842 2 : }
4843 2 :
4844 2 : upload_tenant_manifest(
4845 2 : &self.remote_storage,
4846 2 : &self.tenant_shard_id,
4847 2 : self.generation,
4848 2 : &manifest,
4849 2 : &self.cancel,
4850 2 : )
4851 8 : .await
4852 2 : .map_err(|e| {
4853 0 : if self.cancel.is_cancelled() {
4854 0 : TenantManifestError::Cancelled
4855 : } else {
4856 0 : TenantManifestError::RemoteStorage(e)
4857 : }
4858 2 : })?;
4859 :
4860 : // Store the successfully uploaded manifest, so that future callers can avoid
4861 : // re-uploading the same thing.
4862 2 : *guard = Some(manifest);
4863 2 :
4864 2 : Ok(())
4865 2 : }
4866 : }
4867 :
4868 : /// Create the cluster temporarily in 'initdbpath' directory inside the repository
4869 : /// to get bootstrap data for timeline initialization.
4870 0 : async fn run_initdb(
4871 0 : conf: &'static PageServerConf,
4872 0 : initdb_target_dir: &Utf8Path,
4873 0 : pg_version: u32,
4874 0 : cancel: &CancellationToken,
4875 0 : ) -> Result<(), InitdbError> {
4876 0 : let initdb_bin_path = conf
4877 0 : .pg_bin_dir(pg_version)
4878 0 : .map_err(InitdbError::Other)?
4879 0 : .join("initdb");
4880 0 : let initdb_lib_dir = conf.pg_lib_dir(pg_version).map_err(InitdbError::Other)?;
4881 0 : info!(
4882 0 : "running {} in {}, libdir: {}",
4883 : initdb_bin_path, initdb_target_dir, initdb_lib_dir,
4884 : );
4885 :
4886 0 : let _permit = INIT_DB_SEMAPHORE.acquire().await;
4887 :
4888 0 : let mut initdb_command = tokio::process::Command::new(&initdb_bin_path);
4889 0 : initdb_command
4890 0 : .args(["--pgdata", initdb_target_dir.as_ref()])
4891 0 : .args(["--username", &conf.superuser])
4892 0 : .args(["--encoding", "utf8"])
4893 0 : .args(["--locale", &conf.locale])
4894 0 : .arg("--no-instructions")
4895 0 : .arg("--no-sync")
4896 0 : .env_clear()
4897 0 : .env("LD_LIBRARY_PATH", &initdb_lib_dir)
4898 0 : .env("DYLD_LIBRARY_PATH", &initdb_lib_dir)
4899 0 : .stdin(std::process::Stdio::null())
4900 0 : // stdout invocation produces the same output every time, we don't need it
4901 0 : .stdout(std::process::Stdio::null())
4902 0 : // we would be interested in the stderr output, if there was any
4903 0 : .stderr(std::process::Stdio::piped());
4904 0 :
4905 0 : // Before version 14, only the libc provide was available.
4906 0 : if pg_version > 14 {
4907 : // Version 17 brought with it a builtin locale provider which only provides
4908 : // C and C.UTF-8. While being safer for collation purposes since it is
4909 : // guaranteed to be consistent throughout a major release, it is also more
4910 : // performant.
4911 0 : let locale_provider = if pg_version >= 17 { "builtin" } else { "libc" };
4912 :
4913 0 : initdb_command.args(["--locale-provider", locale_provider]);
4914 0 : }
4915 :
4916 0 : let initdb_proc = initdb_command.spawn()?;
4917 :
4918 : // Ideally we'd select here with the cancellation token, but the problem is that
4919 : // we can't safely terminate initdb: it launches processes of its own, and killing
4920 : // initdb doesn't kill them. After we return from this function, we want the target
4921 : // directory to be able to be cleaned up.
4922 : // See https://github.com/neondatabase/neon/issues/6385
4923 0 : let initdb_output = initdb_proc.wait_with_output().await?;
4924 0 : if !initdb_output.status.success() {
4925 0 : return Err(InitdbError::Failed(
4926 0 : initdb_output.status,
4927 0 : initdb_output.stderr,
4928 0 : ));
4929 0 : }
4930 0 :
4931 0 : // This isn't true cancellation support, see above. Still return an error to
4932 0 : // excercise the cancellation code path.
4933 0 : if cancel.is_cancelled() {
4934 0 : return Err(InitdbError::Cancelled);
4935 0 : }
4936 0 :
4937 0 : Ok(())
4938 0 : }
4939 :
4940 : /// Dump contents of a layer file to stdout.
4941 0 : pub async fn dump_layerfile_from_path(
4942 0 : path: &Utf8Path,
4943 0 : verbose: bool,
4944 0 : ctx: &RequestContext,
4945 0 : ) -> anyhow::Result<()> {
4946 : use std::os::unix::fs::FileExt;
4947 :
4948 : // All layer files start with a two-byte "magic" value, to identify the kind of
4949 : // file.
4950 0 : let file = File::open(path)?;
4951 0 : let mut header_buf = [0u8; 2];
4952 0 : file.read_exact_at(&mut header_buf, 0)?;
4953 :
4954 0 : match u16::from_be_bytes(header_buf) {
4955 : crate::IMAGE_FILE_MAGIC => {
4956 0 : ImageLayer::new_for_path(path, file)?
4957 0 : .dump(verbose, ctx)
4958 0 : .await?
4959 : }
4960 : crate::DELTA_FILE_MAGIC => {
4961 0 : DeltaLayer::new_for_path(path, file)?
4962 0 : .dump(verbose, ctx)
4963 0 : .await?
4964 : }
4965 0 : magic => bail!("unrecognized magic identifier: {:?}", magic),
4966 : }
4967 :
4968 0 : Ok(())
4969 0 : }
4970 :
4971 : #[cfg(test)]
4972 : pub(crate) mod harness {
4973 : use bytes::{Bytes, BytesMut};
4974 : use once_cell::sync::OnceCell;
4975 : use pageserver_api::models::ShardParameters;
4976 : use pageserver_api::shard::ShardIndex;
4977 : use utils::logging;
4978 :
4979 : use crate::deletion_queue::mock::MockDeletionQueue;
4980 : use crate::l0_flush::L0FlushConfig;
4981 : use crate::walredo::apply_neon;
4982 : use pageserver_api::key::Key;
4983 : use pageserver_api::record::NeonWalRecord;
4984 :
4985 : use super::*;
4986 : use hex_literal::hex;
4987 : use utils::id::TenantId;
4988 :
4989 : pub const TIMELINE_ID: TimelineId =
4990 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
4991 : pub const NEW_TIMELINE_ID: TimelineId =
4992 : TimelineId::from_array(hex!("AA223344556677881122334455667788"));
4993 :
4994 : /// Convenience function to create a page image with given string as the only content
4995 5028713 : pub fn test_img(s: &str) -> Bytes {
4996 5028713 : let mut buf = BytesMut::new();
4997 5028713 : buf.extend_from_slice(s.as_bytes());
4998 5028713 : buf.resize(64, 0);
4999 5028713 :
5000 5028713 : buf.freeze()
5001 5028713 : }
5002 :
5003 : impl From<TenantConf> for TenantConfOpt {
5004 192 : fn from(tenant_conf: TenantConf) -> Self {
5005 192 : Self {
5006 192 : checkpoint_distance: Some(tenant_conf.checkpoint_distance),
5007 192 : checkpoint_timeout: Some(tenant_conf.checkpoint_timeout),
5008 192 : compaction_target_size: Some(tenant_conf.compaction_target_size),
5009 192 : compaction_period: Some(tenant_conf.compaction_period),
5010 192 : compaction_threshold: Some(tenant_conf.compaction_threshold),
5011 192 : compaction_algorithm: Some(tenant_conf.compaction_algorithm),
5012 192 : gc_horizon: Some(tenant_conf.gc_horizon),
5013 192 : gc_period: Some(tenant_conf.gc_period),
5014 192 : image_creation_threshold: Some(tenant_conf.image_creation_threshold),
5015 192 : pitr_interval: Some(tenant_conf.pitr_interval),
5016 192 : walreceiver_connect_timeout: Some(tenant_conf.walreceiver_connect_timeout),
5017 192 : lagging_wal_timeout: Some(tenant_conf.lagging_wal_timeout),
5018 192 : max_lsn_wal_lag: Some(tenant_conf.max_lsn_wal_lag),
5019 192 : eviction_policy: Some(tenant_conf.eviction_policy),
5020 192 : min_resident_size_override: tenant_conf.min_resident_size_override,
5021 192 : evictions_low_residence_duration_metric_threshold: Some(
5022 192 : tenant_conf.evictions_low_residence_duration_metric_threshold,
5023 192 : ),
5024 192 : heatmap_period: Some(tenant_conf.heatmap_period),
5025 192 : lazy_slru_download: Some(tenant_conf.lazy_slru_download),
5026 192 : timeline_get_throttle: Some(tenant_conf.timeline_get_throttle),
5027 192 : image_layer_creation_check_threshold: Some(
5028 192 : tenant_conf.image_layer_creation_check_threshold,
5029 192 : ),
5030 192 : lsn_lease_length: Some(tenant_conf.lsn_lease_length),
5031 192 : lsn_lease_length_for_ts: Some(tenant_conf.lsn_lease_length_for_ts),
5032 192 : timeline_offloading: Some(tenant_conf.timeline_offloading),
5033 192 : }
5034 192 : }
5035 : }
5036 :
5037 : pub struct TenantHarness {
5038 : pub conf: &'static PageServerConf,
5039 : pub tenant_conf: TenantConf,
5040 : pub tenant_shard_id: TenantShardId,
5041 : pub generation: Generation,
5042 : pub shard: ShardIndex,
5043 : pub remote_storage: GenericRemoteStorage,
5044 : pub remote_fs_dir: Utf8PathBuf,
5045 : pub deletion_queue: MockDeletionQueue,
5046 : }
5047 :
5048 : static LOG_HANDLE: OnceCell<()> = OnceCell::new();
5049 :
5050 208 : pub(crate) fn setup_logging() {
5051 208 : LOG_HANDLE.get_or_init(|| {
5052 196 : logging::init(
5053 196 : logging::LogFormat::Test,
5054 196 : // enable it in case the tests exercise code paths that use
5055 196 : // debug_assert_current_span_has_tenant_and_timeline_id
5056 196 : logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
5057 196 : logging::Output::Stdout,
5058 196 : )
5059 196 : .expect("Failed to init test logging")
5060 208 : });
5061 208 : }
5062 :
5063 : impl TenantHarness {
5064 192 : pub async fn create_custom(
5065 192 : test_name: &'static str,
5066 192 : tenant_conf: TenantConf,
5067 192 : tenant_id: TenantId,
5068 192 : shard_identity: ShardIdentity,
5069 192 : generation: Generation,
5070 192 : ) -> anyhow::Result<Self> {
5071 192 : setup_logging();
5072 192 :
5073 192 : let repo_dir = PageServerConf::test_repo_dir(test_name);
5074 192 : let _ = fs::remove_dir_all(&repo_dir);
5075 192 : fs::create_dir_all(&repo_dir)?;
5076 :
5077 192 : let conf = PageServerConf::dummy_conf(repo_dir);
5078 192 : // Make a static copy of the config. This can never be free'd, but that's
5079 192 : // OK in a test.
5080 192 : let conf: &'static PageServerConf = Box::leak(Box::new(conf));
5081 192 :
5082 192 : let shard = shard_identity.shard_index();
5083 192 : let tenant_shard_id = TenantShardId {
5084 192 : tenant_id,
5085 192 : shard_number: shard.shard_number,
5086 192 : shard_count: shard.shard_count,
5087 192 : };
5088 192 : fs::create_dir_all(conf.tenant_path(&tenant_shard_id))?;
5089 192 : fs::create_dir_all(conf.timelines_path(&tenant_shard_id))?;
5090 :
5091 : use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
5092 192 : let remote_fs_dir = conf.workdir.join("localfs");
5093 192 : std::fs::create_dir_all(&remote_fs_dir).unwrap();
5094 192 : let config = RemoteStorageConfig {
5095 192 : storage: RemoteStorageKind::LocalFs {
5096 192 : local_path: remote_fs_dir.clone(),
5097 192 : },
5098 192 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
5099 192 : };
5100 192 : let remote_storage = GenericRemoteStorage::from_config(&config).await.unwrap();
5101 192 : let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()));
5102 192 :
5103 192 : Ok(Self {
5104 192 : conf,
5105 192 : tenant_conf,
5106 192 : tenant_shard_id,
5107 192 : generation,
5108 192 : shard,
5109 192 : remote_storage,
5110 192 : remote_fs_dir,
5111 192 : deletion_queue,
5112 192 : })
5113 192 : }
5114 :
5115 180 : pub async fn create(test_name: &'static str) -> anyhow::Result<Self> {
5116 180 : // Disable automatic GC and compaction to make the unit tests more deterministic.
5117 180 : // The tests perform them manually if needed.
5118 180 : let tenant_conf = TenantConf {
5119 180 : gc_period: Duration::ZERO,
5120 180 : compaction_period: Duration::ZERO,
5121 180 : ..TenantConf::default()
5122 180 : };
5123 180 : let tenant_id = TenantId::generate();
5124 180 : let shard = ShardIdentity::unsharded();
5125 180 : Self::create_custom(
5126 180 : test_name,
5127 180 : tenant_conf,
5128 180 : tenant_id,
5129 180 : shard,
5130 180 : Generation::new(0xdeadbeef),
5131 180 : )
5132 0 : .await
5133 180 : }
5134 :
5135 20 : pub fn span(&self) -> tracing::Span {
5136 20 : info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
5137 20 : }
5138 :
5139 192 : pub(crate) async fn load(&self) -> (Arc<Tenant>, RequestContext) {
5140 192 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
5141 192 : (
5142 192 : self.do_try_load(&ctx)
5143 1900 : .await
5144 192 : .expect("failed to load test tenant"),
5145 192 : ctx,
5146 192 : )
5147 192 : }
5148 :
5149 192 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5150 : pub(crate) async fn do_try_load(
5151 : &self,
5152 : ctx: &RequestContext,
5153 : ) -> anyhow::Result<Arc<Tenant>> {
5154 : let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
5155 :
5156 : let tenant = Arc::new(Tenant::new(
5157 : TenantState::Attaching,
5158 : self.conf,
5159 : AttachedTenantConf::try_from(LocationConf::attached_single(
5160 : TenantConfOpt::from(self.tenant_conf.clone()),
5161 : self.generation,
5162 : &ShardParameters::default(),
5163 : ))
5164 : .unwrap(),
5165 : // This is a legacy/test code path: sharding isn't supported here.
5166 : ShardIdentity::unsharded(),
5167 : Some(walredo_mgr),
5168 : self.tenant_shard_id,
5169 : self.remote_storage.clone(),
5170 : self.deletion_queue.new_client(),
5171 : // TODO: ideally we should run all unit tests with both configs
5172 : L0FlushGlobalState::new(L0FlushConfig::default()),
5173 : ));
5174 :
5175 : let preload = tenant
5176 : .preload(&self.remote_storage, CancellationToken::new())
5177 : .await?;
5178 : tenant.attach(Some(preload), ctx).await?;
5179 :
5180 : tenant.state.send_replace(TenantState::Active);
5181 : for timeline in tenant.timelines.lock().unwrap().values() {
5182 : timeline.set_state(TimelineState::Active);
5183 : }
5184 : Ok(tenant)
5185 : }
5186 :
5187 2 : pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
5188 2 : self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
5189 2 : }
5190 : }
5191 :
5192 : // Mock WAL redo manager that doesn't do much
5193 : pub(crate) struct TestRedoManager;
5194 :
5195 : impl TestRedoManager {
5196 : /// # Cancel-Safety
5197 : ///
5198 : /// This method is cancellation-safe.
5199 410 : pub async fn request_redo(
5200 410 : &self,
5201 410 : key: Key,
5202 410 : lsn: Lsn,
5203 410 : base_img: Option<(Lsn, Bytes)>,
5204 410 : records: Vec<(Lsn, NeonWalRecord)>,
5205 410 : _pg_version: u32,
5206 410 : ) -> Result<Bytes, walredo::Error> {
5207 570 : let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
5208 410 : if records_neon {
5209 : // For Neon wal records, we can decode without spawning postgres, so do so.
5210 410 : let mut page = match (base_img, records.first()) {
5211 344 : (Some((_lsn, img)), _) => {
5212 344 : let mut page = BytesMut::new();
5213 344 : page.extend_from_slice(&img);
5214 344 : page
5215 : }
5216 66 : (_, Some((_lsn, rec))) if rec.will_init() => BytesMut::new(),
5217 : _ => {
5218 0 : panic!("Neon WAL redo requires base image or will init record");
5219 : }
5220 : };
5221 :
5222 980 : for (record_lsn, record) in records {
5223 570 : apply_neon::apply_in_neon(&record, record_lsn, key, &mut page)?;
5224 : }
5225 410 : Ok(page.freeze())
5226 : } else {
5227 : // We never spawn a postgres walredo process in unit tests: just log what we might have done.
5228 0 : let s = format!(
5229 0 : "redo for {} to get to {}, with {} and {} records",
5230 0 : key,
5231 0 : lsn,
5232 0 : if base_img.is_some() {
5233 0 : "base image"
5234 : } else {
5235 0 : "no base image"
5236 : },
5237 0 : records.len()
5238 0 : );
5239 0 : println!("{s}");
5240 0 :
5241 0 : Ok(test_img(&s))
5242 : }
5243 410 : }
5244 : }
5245 : }
5246 :
5247 : #[cfg(test)]
5248 : mod tests {
5249 : use std::collections::{BTreeMap, BTreeSet};
5250 :
5251 : use super::*;
5252 : use crate::keyspace::KeySpaceAccum;
5253 : use crate::tenant::harness::*;
5254 : use crate::tenant::timeline::CompactFlags;
5255 : use crate::DEFAULT_PG_VERSION;
5256 : use bytes::{Bytes, BytesMut};
5257 : use hex_literal::hex;
5258 : use itertools::Itertools;
5259 : use pageserver_api::key::{Key, AUX_KEY_PREFIX, NON_INHERITED_RANGE};
5260 : use pageserver_api::keyspace::KeySpace;
5261 : use pageserver_api::models::{CompactionAlgorithm, CompactionAlgorithmSettings};
5262 : use pageserver_api::value::Value;
5263 : use pageserver_compaction::helpers::overlaps_with;
5264 : use rand::{thread_rng, Rng};
5265 : use storage_layer::PersistentLayerKey;
5266 : use tests::storage_layer::ValuesReconstructState;
5267 : use tests::timeline::{GetVectoredError, ShutdownMode};
5268 : use timeline::{CompactOptions, DeltaLayerTestDesc};
5269 : use utils::id::TenantId;
5270 :
5271 : #[cfg(feature = "testing")]
5272 : use pageserver_api::record::NeonWalRecord;
5273 : #[cfg(feature = "testing")]
5274 : use timeline::compaction::{KeyHistoryRetention, KeyLogAtLsn};
5275 : #[cfg(feature = "testing")]
5276 : use timeline::GcInfo;
5277 :
5278 : static TEST_KEY: Lazy<Key> =
5279 18 : Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));
5280 :
5281 : #[tokio::test]
5282 2 : async fn test_basic() -> anyhow::Result<()> {
5283 20 : let (tenant, ctx) = TenantHarness::create("test_basic").await?.load().await;
5284 2 : let tline = tenant
5285 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5286 6 : .await?;
5287 2 :
5288 2 : let mut writer = tline.writer().await;
5289 2 : writer
5290 2 : .put(
5291 2 : *TEST_KEY,
5292 2 : Lsn(0x10),
5293 2 : &Value::Image(test_img("foo at 0x10")),
5294 2 : &ctx,
5295 2 : )
5296 2 : .await?;
5297 2 : writer.finish_write(Lsn(0x10));
5298 2 : drop(writer);
5299 2 :
5300 2 : let mut writer = tline.writer().await;
5301 2 : writer
5302 2 : .put(
5303 2 : *TEST_KEY,
5304 2 : Lsn(0x20),
5305 2 : &Value::Image(test_img("foo at 0x20")),
5306 2 : &ctx,
5307 2 : )
5308 2 : .await?;
5309 2 : writer.finish_write(Lsn(0x20));
5310 2 : drop(writer);
5311 2 :
5312 2 : assert_eq!(
5313 2 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
5314 2 : test_img("foo at 0x10")
5315 2 : );
5316 2 : assert_eq!(
5317 2 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
5318 2 : test_img("foo at 0x10")
5319 2 : );
5320 2 : assert_eq!(
5321 2 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
5322 2 : test_img("foo at 0x20")
5323 2 : );
5324 2 :
5325 2 : Ok(())
5326 2 : }
5327 :
5328 : #[tokio::test]
5329 2 : async fn no_duplicate_timelines() -> anyhow::Result<()> {
5330 2 : let (tenant, ctx) = TenantHarness::create("no_duplicate_timelines")
5331 2 : .await?
5332 2 : .load()
5333 20 : .await;
5334 2 : let _ = tenant
5335 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5336 6 : .await?;
5337 2 :
5338 2 : match tenant
5339 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5340 2 : .await
5341 2 : {
5342 2 : Ok(_) => panic!("duplicate timeline creation should fail"),
5343 2 : Err(e) => assert_eq!(
5344 2 : e.to_string(),
5345 2 : "timeline already exists with different parameters".to_string()
5346 2 : ),
5347 2 : }
5348 2 :
5349 2 : Ok(())
5350 2 : }
5351 :
5352 : /// Convenience function to create a page image with given string as the only content
5353 10 : pub fn test_value(s: &str) -> Value {
5354 10 : let mut buf = BytesMut::new();
5355 10 : buf.extend_from_slice(s.as_bytes());
5356 10 : Value::Image(buf.freeze())
5357 10 : }
5358 :
5359 : ///
5360 : /// Test branch creation
5361 : ///
5362 : #[tokio::test]
5363 2 : async fn test_branch() -> anyhow::Result<()> {
5364 2 : use std::str::from_utf8;
5365 2 :
5366 19 : let (tenant, ctx) = TenantHarness::create("test_branch").await?.load().await;
5367 2 : let tline = tenant
5368 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5369 5 : .await?;
5370 2 : let mut writer = tline.writer().await;
5371 2 :
5372 2 : #[allow(non_snake_case)]
5373 2 : let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
5374 2 : #[allow(non_snake_case)]
5375 2 : let TEST_KEY_B: Key = Key::from_hex("110000000033333333444444445500000002").unwrap();
5376 2 :
5377 2 : // Insert a value on the timeline
5378 2 : writer
5379 2 : .put(TEST_KEY_A, Lsn(0x20), &test_value("foo at 0x20"), &ctx)
5380 2 : .await?;
5381 2 : writer
5382 2 : .put(TEST_KEY_B, Lsn(0x20), &test_value("foobar at 0x20"), &ctx)
5383 2 : .await?;
5384 2 : writer.finish_write(Lsn(0x20));
5385 2 :
5386 2 : writer
5387 2 : .put(TEST_KEY_A, Lsn(0x30), &test_value("foo at 0x30"), &ctx)
5388 2 : .await?;
5389 2 : writer.finish_write(Lsn(0x30));
5390 2 : writer
5391 2 : .put(TEST_KEY_A, Lsn(0x40), &test_value("foo at 0x40"), &ctx)
5392 2 : .await?;
5393 2 : writer.finish_write(Lsn(0x40));
5394 2 :
5395 2 : //assert_current_logical_size(&tline, Lsn(0x40));
5396 2 :
5397 2 : // Branch the history, modify relation differently on the new timeline
5398 2 : tenant
5399 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x30)), &ctx)
5400 2 : .await?;
5401 2 : let newtline = tenant
5402 2 : .get_timeline(NEW_TIMELINE_ID, true)
5403 2 : .expect("Should have a local timeline");
5404 2 : let mut new_writer = newtline.writer().await;
5405 2 : new_writer
5406 2 : .put(TEST_KEY_A, Lsn(0x40), &test_value("bar at 0x40"), &ctx)
5407 2 : .await?;
5408 2 : new_writer.finish_write(Lsn(0x40));
5409 2 :
5410 2 : // Check page contents on both branches
5411 2 : assert_eq!(
5412 2 : from_utf8(&tline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
5413 2 : "foo at 0x40"
5414 2 : );
5415 2 : assert_eq!(
5416 2 : from_utf8(&newtline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
5417 2 : "bar at 0x40"
5418 2 : );
5419 2 : assert_eq!(
5420 2 : from_utf8(&newtline.get(TEST_KEY_B, Lsn(0x40), &ctx).await?)?,
5421 2 : "foobar at 0x20"
5422 2 : );
5423 2 :
5424 2 : //assert_current_logical_size(&tline, Lsn(0x40));
5425 2 :
5426 2 : Ok(())
5427 2 : }
5428 :
5429 20 : async fn make_some_layers(
5430 20 : tline: &Timeline,
5431 20 : start_lsn: Lsn,
5432 20 : ctx: &RequestContext,
5433 20 : ) -> anyhow::Result<()> {
5434 20 : let mut lsn = start_lsn;
5435 : {
5436 20 : let mut writer = tline.writer().await;
5437 : // Create a relation on the timeline
5438 20 : writer
5439 20 : .put(
5440 20 : *TEST_KEY,
5441 20 : lsn,
5442 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5443 20 : ctx,
5444 20 : )
5445 10 : .await?;
5446 20 : writer.finish_write(lsn);
5447 20 : lsn += 0x10;
5448 20 : writer
5449 20 : .put(
5450 20 : *TEST_KEY,
5451 20 : lsn,
5452 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5453 20 : ctx,
5454 20 : )
5455 0 : .await?;
5456 20 : writer.finish_write(lsn);
5457 20 : lsn += 0x10;
5458 20 : }
5459 21 : tline.freeze_and_flush().await?;
5460 : {
5461 20 : let mut writer = tline.writer().await;
5462 20 : writer
5463 20 : .put(
5464 20 : *TEST_KEY,
5465 20 : lsn,
5466 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5467 20 : ctx,
5468 20 : )
5469 10 : .await?;
5470 20 : writer.finish_write(lsn);
5471 20 : lsn += 0x10;
5472 20 : writer
5473 20 : .put(
5474 20 : *TEST_KEY,
5475 20 : lsn,
5476 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5477 20 : ctx,
5478 20 : )
5479 0 : .await?;
5480 20 : writer.finish_write(lsn);
5481 20 : }
5482 20 : tline.freeze_and_flush().await.map_err(|e| e.into())
5483 20 : }
5484 :
5485 : #[tokio::test(start_paused = true)]
5486 2 : async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
5487 2 : let (tenant, ctx) =
5488 2 : TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
5489 2 : .await?
5490 2 : .load()
5491 20 : .await;
5492 2 : // Advance to the lsn lease deadline so that GC is not blocked by
5493 2 : // initial transition into AttachedSingle.
5494 2 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
5495 2 : tokio::time::resume();
5496 2 : let tline = tenant
5497 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5498 6 : .await?;
5499 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5500 2 :
5501 2 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
5502 2 : // FIXME: this doesn't actually remove any layer currently, given how the flushing
5503 2 : // and compaction works. But it does set the 'cutoff' point so that the cross check
5504 2 : // below should fail.
5505 2 : tenant
5506 2 : .gc_iteration(
5507 2 : Some(TIMELINE_ID),
5508 2 : 0x10,
5509 2 : Duration::ZERO,
5510 2 : &CancellationToken::new(),
5511 2 : &ctx,
5512 2 : )
5513 2 : .await?;
5514 2 :
5515 2 : // try to branch at lsn 25, should fail because we already garbage collected the data
5516 2 : match tenant
5517 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
5518 2 : .await
5519 2 : {
5520 2 : Ok(_) => panic!("branching should have failed"),
5521 2 : Err(err) => {
5522 2 : let CreateTimelineError::AncestorLsn(err) = err else {
5523 2 : panic!("wrong error type")
5524 2 : };
5525 2 : assert!(err.to_string().contains("invalid branch start lsn"));
5526 2 : assert!(err
5527 2 : .source()
5528 2 : .unwrap()
5529 2 : .to_string()
5530 2 : .contains("we might've already garbage collected needed data"))
5531 2 : }
5532 2 : }
5533 2 :
5534 2 : Ok(())
5535 2 : }
5536 :
5537 : #[tokio::test]
5538 2 : async fn test_prohibit_branch_creation_on_pre_initdb_lsn() -> anyhow::Result<()> {
5539 2 : let (tenant, ctx) =
5540 2 : TenantHarness::create("test_prohibit_branch_creation_on_pre_initdb_lsn")
5541 2 : .await?
5542 2 : .load()
5543 20 : .await;
5544 2 :
5545 2 : let tline = tenant
5546 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x50), DEFAULT_PG_VERSION, &ctx)
5547 6 : .await?;
5548 2 : // try to branch at lsn 0x25, should fail because initdb lsn is 0x50
5549 2 : match tenant
5550 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
5551 2 : .await
5552 2 : {
5553 2 : Ok(_) => panic!("branching should have failed"),
5554 2 : Err(err) => {
5555 2 : let CreateTimelineError::AncestorLsn(err) = err else {
5556 2 : panic!("wrong error type");
5557 2 : };
5558 2 : assert!(&err.to_string().contains("invalid branch start lsn"));
5559 2 : assert!(&err
5560 2 : .source()
5561 2 : .unwrap()
5562 2 : .to_string()
5563 2 : .contains("is earlier than latest GC cutoff"));
5564 2 : }
5565 2 : }
5566 2 :
5567 2 : Ok(())
5568 2 : }
5569 :
5570 : /*
5571 : // FIXME: This currently fails to error out. Calling GC doesn't currently
5572 : // remove the old value, we'd need to work a little harder
5573 : #[tokio::test]
5574 : async fn test_prohibit_get_for_garbage_collected_data() -> anyhow::Result<()> {
5575 : let repo =
5576 : RepoHarness::create("test_prohibit_get_for_garbage_collected_data")?
5577 : .load();
5578 :
5579 : let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION)?;
5580 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5581 :
5582 : repo.gc_iteration(Some(TIMELINE_ID), 0x10, Duration::ZERO)?;
5583 : let latest_gc_cutoff_lsn = tline.get_latest_gc_cutoff_lsn();
5584 : assert!(*latest_gc_cutoff_lsn > Lsn(0x25));
5585 : match tline.get(*TEST_KEY, Lsn(0x25)) {
5586 : Ok(_) => panic!("request for page should have failed"),
5587 : Err(err) => assert!(err.to_string().contains("not found at")),
5588 : }
5589 : Ok(())
5590 : }
5591 : */
5592 :
5593 : #[tokio::test]
5594 2 : async fn test_get_branchpoints_from_an_inactive_timeline() -> anyhow::Result<()> {
5595 2 : let (tenant, ctx) =
5596 2 : TenantHarness::create("test_get_branchpoints_from_an_inactive_timeline")
5597 2 : .await?
5598 2 : .load()
5599 20 : .await;
5600 2 : let tline = tenant
5601 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5602 6 : .await?;
5603 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5604 2 :
5605 2 : tenant
5606 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
5607 2 : .await?;
5608 2 : let newtline = tenant
5609 2 : .get_timeline(NEW_TIMELINE_ID, true)
5610 2 : .expect("Should have a local timeline");
5611 2 :
5612 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
5613 2 :
5614 2 : tline.set_broken("test".to_owned());
5615 2 :
5616 2 : tenant
5617 2 : .gc_iteration(
5618 2 : Some(TIMELINE_ID),
5619 2 : 0x10,
5620 2 : Duration::ZERO,
5621 2 : &CancellationToken::new(),
5622 2 : &ctx,
5623 2 : )
5624 2 : .await?;
5625 2 :
5626 2 : // The branchpoints should contain all timelines, even ones marked
5627 2 : // as Broken.
5628 2 : {
5629 2 : let branchpoints = &tline.gc_info.read().unwrap().retain_lsns;
5630 2 : assert_eq!(branchpoints.len(), 1);
5631 2 : assert_eq!(
5632 2 : branchpoints[0],
5633 2 : (Lsn(0x40), NEW_TIMELINE_ID, MaybeOffloaded::No)
5634 2 : );
5635 2 : }
5636 2 :
5637 2 : // You can read the key from the child branch even though the parent is
5638 2 : // Broken, as long as you don't need to access data from the parent.
5639 2 : assert_eq!(
5640 4 : newtline.get(*TEST_KEY, Lsn(0x70), &ctx).await?,
5641 2 : test_img(&format!("foo at {}", Lsn(0x70)))
5642 2 : );
5643 2 :
5644 2 : // This needs to traverse to the parent, and fails.
5645 2 : let err = newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await.unwrap_err();
5646 2 : assert!(
5647 2 : err.to_string().starts_with(&format!(
5648 2 : "bad state on timeline {}: Broken",
5649 2 : tline.timeline_id
5650 2 : )),
5651 2 : "{err}"
5652 2 : );
5653 2 :
5654 2 : Ok(())
5655 2 : }
5656 :
5657 : #[tokio::test]
5658 2 : async fn test_retain_data_in_parent_which_is_needed_for_child() -> anyhow::Result<()> {
5659 2 : let (tenant, ctx) =
5660 2 : TenantHarness::create("test_retain_data_in_parent_which_is_needed_for_child")
5661 2 : .await?
5662 2 : .load()
5663 20 : .await;
5664 2 : let tline = tenant
5665 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5666 6 : .await?;
5667 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5668 2 :
5669 2 : tenant
5670 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
5671 2 : .await?;
5672 2 : let newtline = tenant
5673 2 : .get_timeline(NEW_TIMELINE_ID, true)
5674 2 : .expect("Should have a local timeline");
5675 2 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
5676 2 : tenant
5677 2 : .gc_iteration(
5678 2 : Some(TIMELINE_ID),
5679 2 : 0x10,
5680 2 : Duration::ZERO,
5681 2 : &CancellationToken::new(),
5682 2 : &ctx,
5683 2 : )
5684 2 : .await?;
5685 4 : assert!(newtline.get(*TEST_KEY, Lsn(0x25), &ctx).await.is_ok());
5686 2 :
5687 2 : Ok(())
5688 2 : }
5689 : #[tokio::test]
5690 2 : async fn test_parent_keeps_data_forever_after_branching() -> anyhow::Result<()> {
5691 2 : let (tenant, ctx) = TenantHarness::create("test_parent_keeps_data_forever_after_branching")
5692 2 : .await?
5693 2 : .load()
5694 20 : .await;
5695 2 : let tline = tenant
5696 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5697 6 : .await?;
5698 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5699 2 :
5700 2 : tenant
5701 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
5702 2 : .await?;
5703 2 : let newtline = tenant
5704 2 : .get_timeline(NEW_TIMELINE_ID, true)
5705 2 : .expect("Should have a local timeline");
5706 2 :
5707 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
5708 2 :
5709 2 : // run gc on parent
5710 2 : tenant
5711 2 : .gc_iteration(
5712 2 : Some(TIMELINE_ID),
5713 2 : 0x10,
5714 2 : Duration::ZERO,
5715 2 : &CancellationToken::new(),
5716 2 : &ctx,
5717 2 : )
5718 2 : .await?;
5719 2 :
5720 2 : // Check that the data is still accessible on the branch.
5721 2 : assert_eq!(
5722 7 : newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
5723 2 : test_img(&format!("foo at {}", Lsn(0x40)))
5724 2 : );
5725 2 :
5726 2 : Ok(())
5727 2 : }
5728 :
5729 : #[tokio::test]
5730 2 : async fn timeline_load() -> anyhow::Result<()> {
5731 2 : const TEST_NAME: &str = "timeline_load";
5732 2 : let harness = TenantHarness::create(TEST_NAME).await?;
5733 2 : {
5734 20 : let (tenant, ctx) = harness.load().await;
5735 2 : let tline = tenant
5736 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x7000), DEFAULT_PG_VERSION, &ctx)
5737 6 : .await?;
5738 7 : make_some_layers(tline.as_ref(), Lsn(0x8000), &ctx).await?;
5739 2 : // so that all uploads finish & we can call harness.load() below again
5740 2 : tenant
5741 2 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
5742 2 : .instrument(harness.span())
5743 2 : .await
5744 2 : .ok()
5745 2 : .unwrap();
5746 2 : }
5747 2 :
5748 24 : let (tenant, _ctx) = harness.load().await;
5749 2 : tenant
5750 2 : .get_timeline(TIMELINE_ID, true)
5751 2 : .expect("cannot load timeline");
5752 2 :
5753 2 : Ok(())
5754 2 : }
5755 :
5756 : #[tokio::test]
5757 2 : async fn timeline_load_with_ancestor() -> anyhow::Result<()> {
5758 2 : const TEST_NAME: &str = "timeline_load_with_ancestor";
5759 2 : let harness = TenantHarness::create(TEST_NAME).await?;
5760 2 : // create two timelines
5761 2 : {
5762 20 : let (tenant, ctx) = harness.load().await;
5763 2 : let tline = tenant
5764 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5765 6 : .await?;
5766 2 :
5767 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5768 2 :
5769 2 : let child_tline = tenant
5770 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
5771 2 : .await?;
5772 2 : child_tline.set_state(TimelineState::Active);
5773 2 :
5774 2 : let newtline = tenant
5775 2 : .get_timeline(NEW_TIMELINE_ID, true)
5776 2 : .expect("Should have a local timeline");
5777 2 :
5778 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
5779 2 :
5780 2 : // so that all uploads finish & we can call harness.load() below again
5781 2 : tenant
5782 2 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
5783 2 : .instrument(harness.span())
5784 2 : .await
5785 2 : .ok()
5786 2 : .unwrap();
5787 2 : }
5788 2 :
5789 2 : // check that both of them are initially unloaded
5790 30 : let (tenant, _ctx) = harness.load().await;
5791 2 :
5792 2 : // check that both, child and ancestor are loaded
5793 2 : let _child_tline = tenant
5794 2 : .get_timeline(NEW_TIMELINE_ID, true)
5795 2 : .expect("cannot get child timeline loaded");
5796 2 :
5797 2 : let _ancestor_tline = tenant
5798 2 : .get_timeline(TIMELINE_ID, true)
5799 2 : .expect("cannot get ancestor timeline loaded");
5800 2 :
5801 2 : Ok(())
5802 2 : }
5803 :
5804 : #[tokio::test]
5805 2 : async fn delta_layer_dumping() -> anyhow::Result<()> {
5806 2 : use storage_layer::AsLayerDesc;
5807 2 : let (tenant, ctx) = TenantHarness::create("test_layer_dumping")
5808 2 : .await?
5809 2 : .load()
5810 20 : .await;
5811 2 : let tline = tenant
5812 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5813 6 : .await?;
5814 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5815 2 :
5816 2 : let layer_map = tline.layers.read().await;
5817 2 : let level0_deltas = layer_map
5818 2 : .layer_map()?
5819 2 : .level0_deltas()
5820 2 : .iter()
5821 4 : .map(|desc| layer_map.get_from_desc(desc))
5822 2 : .collect::<Vec<_>>();
5823 2 :
5824 2 : assert!(!level0_deltas.is_empty());
5825 2 :
5826 6 : for delta in level0_deltas {
5827 2 : // Ensure we are dumping a delta layer here
5828 4 : assert!(delta.layer_desc().is_delta);
5829 8 : delta.dump(true, &ctx).await.unwrap();
5830 2 : }
5831 2 :
5832 2 : Ok(())
5833 2 : }
5834 :
5835 : #[tokio::test]
5836 2 : async fn test_images() -> anyhow::Result<()> {
5837 20 : let (tenant, ctx) = TenantHarness::create("test_images").await?.load().await;
5838 2 : let tline = tenant
5839 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5840 6 : .await?;
5841 2 :
5842 2 : let mut writer = tline.writer().await;
5843 2 : writer
5844 2 : .put(
5845 2 : *TEST_KEY,
5846 2 : Lsn(0x10),
5847 2 : &Value::Image(test_img("foo at 0x10")),
5848 2 : &ctx,
5849 2 : )
5850 2 : .await?;
5851 2 : writer.finish_write(Lsn(0x10));
5852 2 : drop(writer);
5853 2 :
5854 2 : tline.freeze_and_flush().await?;
5855 2 : tline
5856 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
5857 2 : .await?;
5858 2 :
5859 2 : let mut writer = tline.writer().await;
5860 2 : writer
5861 2 : .put(
5862 2 : *TEST_KEY,
5863 2 : Lsn(0x20),
5864 2 : &Value::Image(test_img("foo at 0x20")),
5865 2 : &ctx,
5866 2 : )
5867 2 : .await?;
5868 2 : writer.finish_write(Lsn(0x20));
5869 2 : drop(writer);
5870 2 :
5871 2 : tline.freeze_and_flush().await?;
5872 2 : tline
5873 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
5874 2 : .await?;
5875 2 :
5876 2 : let mut writer = tline.writer().await;
5877 2 : writer
5878 2 : .put(
5879 2 : *TEST_KEY,
5880 2 : Lsn(0x30),
5881 2 : &Value::Image(test_img("foo at 0x30")),
5882 2 : &ctx,
5883 2 : )
5884 2 : .await?;
5885 2 : writer.finish_write(Lsn(0x30));
5886 2 : drop(writer);
5887 2 :
5888 2 : tline.freeze_and_flush().await?;
5889 2 : tline
5890 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
5891 2 : .await?;
5892 2 :
5893 2 : let mut writer = tline.writer().await;
5894 2 : writer
5895 2 : .put(
5896 2 : *TEST_KEY,
5897 2 : Lsn(0x40),
5898 2 : &Value::Image(test_img("foo at 0x40")),
5899 2 : &ctx,
5900 2 : )
5901 2 : .await?;
5902 2 : writer.finish_write(Lsn(0x40));
5903 2 : drop(writer);
5904 2 :
5905 2 : tline.freeze_and_flush().await?;
5906 2 : tline
5907 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
5908 2 : .await?;
5909 2 :
5910 2 : assert_eq!(
5911 4 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
5912 2 : test_img("foo at 0x10")
5913 2 : );
5914 2 : assert_eq!(
5915 4 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
5916 2 : test_img("foo at 0x10")
5917 2 : );
5918 2 : assert_eq!(
5919 2 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
5920 2 : test_img("foo at 0x20")
5921 2 : );
5922 2 : assert_eq!(
5923 4 : tline.get(*TEST_KEY, Lsn(0x30), &ctx).await?,
5924 2 : test_img("foo at 0x30")
5925 2 : );
5926 2 : assert_eq!(
5927 4 : tline.get(*TEST_KEY, Lsn(0x40), &ctx).await?,
5928 2 : test_img("foo at 0x40")
5929 2 : );
5930 2 :
5931 2 : Ok(())
5932 2 : }
5933 :
5934 4 : async fn bulk_insert_compact_gc(
5935 4 : tenant: &Tenant,
5936 4 : timeline: &Arc<Timeline>,
5937 4 : ctx: &RequestContext,
5938 4 : lsn: Lsn,
5939 4 : repeat: usize,
5940 4 : key_count: usize,
5941 4 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
5942 4 : let compact = true;
5943 40718 : bulk_insert_maybe_compact_gc(tenant, timeline, ctx, lsn, repeat, key_count, compact).await
5944 4 : }
5945 :
5946 8 : async fn bulk_insert_maybe_compact_gc(
5947 8 : tenant: &Tenant,
5948 8 : timeline: &Arc<Timeline>,
5949 8 : ctx: &RequestContext,
5950 8 : mut lsn: Lsn,
5951 8 : repeat: usize,
5952 8 : key_count: usize,
5953 8 : compact: bool,
5954 8 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
5955 8 : let mut inserted: HashMap<Key, BTreeSet<Lsn>> = Default::default();
5956 8 :
5957 8 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5958 8 : let mut blknum = 0;
5959 8 :
5960 8 : // Enforce that key range is monotonously increasing
5961 8 : let mut keyspace = KeySpaceAccum::new();
5962 8 :
5963 8 : let cancel = CancellationToken::new();
5964 8 :
5965 8 : for _ in 0..repeat {
5966 400 : for _ in 0..key_count {
5967 4000000 : test_key.field6 = blknum;
5968 4000000 : let mut writer = timeline.writer().await;
5969 4000000 : writer
5970 4000000 : .put(
5971 4000000 : test_key,
5972 4000000 : lsn,
5973 4000000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5974 4000000 : ctx,
5975 4000000 : )
5976 3426 : .await?;
5977 4000000 : inserted.entry(test_key).or_default().insert(lsn);
5978 4000000 : writer.finish_write(lsn);
5979 4000000 : drop(writer);
5980 4000000 :
5981 4000000 : keyspace.add_key(test_key);
5982 4000000 :
5983 4000000 : lsn = Lsn(lsn.0 + 0x10);
5984 4000000 : blknum += 1;
5985 : }
5986 :
5987 400 : timeline.freeze_and_flush().await?;
5988 400 : if compact {
5989 : // this requires timeline to be &Arc<Timeline>
5990 8618 : timeline.compact(&cancel, EnumSet::empty(), ctx).await?;
5991 200 : }
5992 :
5993 : // this doesn't really need to use the timeline_id target, but it is closer to what it
5994 : // originally was.
5995 400 : let res = tenant
5996 400 : .gc_iteration(Some(timeline.timeline_id), 0, Duration::ZERO, &cancel, ctx)
5997 0 : .await?;
5998 :
5999 400 : assert_eq!(res.layers_removed, 0, "this never removes anything");
6000 : }
6001 :
6002 8 : Ok(inserted)
6003 8 : }
6004 :
6005 : //
6006 : // Insert 1000 key-value pairs with increasing keys, flush, compact, GC.
6007 : // Repeat 50 times.
6008 : //
6009 : #[tokio::test]
6010 2 : async fn test_bulk_insert() -> anyhow::Result<()> {
6011 2 : let harness = TenantHarness::create("test_bulk_insert").await?;
6012 20 : let (tenant, ctx) = harness.load().await;
6013 2 : let tline = tenant
6014 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6015 6 : .await?;
6016 2 :
6017 2 : let lsn = Lsn(0x10);
6018 20359 : bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
6019 2 :
6020 2 : Ok(())
6021 2 : }
6022 :
6023 : // Test the vectored get real implementation against a simple sequential implementation.
6024 : //
6025 : // The test generates a keyspace by repeatedly flushing the in-memory layer and compacting.
6026 : // Projected to 2D the key space looks like below. Lsn grows upwards on the Y axis and keys
6027 : // grow to the right on the X axis.
6028 : // [Delta]
6029 : // [Delta]
6030 : // [Delta]
6031 : // [Delta]
6032 : // ------------ Image ---------------
6033 : //
6034 : // After layer generation we pick the ranges to query as follows:
6035 : // 1. The beginning of each delta layer
6036 : // 2. At the seam between two adjacent delta layers
6037 : //
6038 : // There's one major downside to this test: delta layers only contains images,
6039 : // so the search can stop at the first delta layer and doesn't traverse any deeper.
6040 : #[tokio::test]
6041 2 : async fn test_get_vectored() -> anyhow::Result<()> {
6042 2 : let harness = TenantHarness::create("test_get_vectored").await?;
6043 20 : let (tenant, ctx) = harness.load().await;
6044 2 : let tline = tenant
6045 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6046 6 : .await?;
6047 2 :
6048 2 : let lsn = Lsn(0x10);
6049 20359 : let inserted = bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
6050 2 :
6051 2 : let guard = tline.layers.read().await;
6052 2 : let lm = guard.layer_map()?;
6053 2 :
6054 2 : lm.dump(true, &ctx).await?;
6055 2 :
6056 2 : let mut reads = Vec::new();
6057 2 : let mut prev = None;
6058 12 : lm.iter_historic_layers().for_each(|desc| {
6059 12 : if !desc.is_delta() {
6060 2 : prev = Some(desc.clone());
6061 2 : return;
6062 10 : }
6063 10 :
6064 10 : let start = desc.key_range.start;
6065 10 : let end = desc
6066 10 : .key_range
6067 10 : .start
6068 10 : .add(Timeline::MAX_GET_VECTORED_KEYS.try_into().unwrap());
6069 10 : reads.push(KeySpace {
6070 10 : ranges: vec![start..end],
6071 10 : });
6072 2 :
6073 10 : if let Some(prev) = &prev {
6074 10 : if !prev.is_delta() {
6075 10 : return;
6076 2 : }
6077 0 :
6078 0 : let first_range = Key {
6079 0 : field6: prev.key_range.end.field6 - 4,
6080 0 : ..prev.key_range.end
6081 0 : }..prev.key_range.end;
6082 0 :
6083 0 : let second_range = desc.key_range.start..Key {
6084 0 : field6: desc.key_range.start.field6 + 4,
6085 0 : ..desc.key_range.start
6086 0 : };
6087 0 :
6088 0 : reads.push(KeySpace {
6089 0 : ranges: vec![first_range, second_range],
6090 0 : });
6091 2 : };
6092 2 :
6093 2 : prev = Some(desc.clone());
6094 12 : });
6095 2 :
6096 2 : drop(guard);
6097 2 :
6098 2 : // Pick a big LSN such that we query over all the changes.
6099 2 : let reads_lsn = Lsn(u64::MAX - 1);
6100 2 :
6101 12 : for read in reads {
6102 10 : info!("Doing vectored read on {:?}", read);
6103 2 :
6104 10 : let vectored_res = tline
6105 10 : .get_vectored_impl(
6106 10 : read.clone(),
6107 10 : reads_lsn,
6108 10 : &mut ValuesReconstructState::new(),
6109 10 : &ctx,
6110 10 : )
6111 25 : .await;
6112 2 :
6113 10 : let mut expected_lsns: HashMap<Key, Lsn> = Default::default();
6114 10 : let mut expect_missing = false;
6115 10 : let mut key = read.start().unwrap();
6116 330 : while key != read.end().unwrap() {
6117 320 : if let Some(lsns) = inserted.get(&key) {
6118 320 : let expected_lsn = lsns.iter().rfind(|lsn| **lsn <= reads_lsn);
6119 320 : match expected_lsn {
6120 320 : Some(lsn) => {
6121 320 : expected_lsns.insert(key, *lsn);
6122 320 : }
6123 2 : None => {
6124 2 : expect_missing = true;
6125 0 : break;
6126 2 : }
6127 2 : }
6128 2 : } else {
6129 2 : expect_missing = true;
6130 0 : break;
6131 2 : }
6132 2 :
6133 320 : key = key.next();
6134 2 : }
6135 2 :
6136 10 : if expect_missing {
6137 2 : assert!(matches!(vectored_res, Err(GetVectoredError::MissingKey(_))));
6138 2 : } else {
6139 320 : for (key, image) in vectored_res? {
6140 320 : let expected_lsn = expected_lsns.get(&key).expect("determined above");
6141 320 : let expected_image = test_img(&format!("{} at {}", key.field6, expected_lsn));
6142 320 : assert_eq!(image?, expected_image);
6143 2 : }
6144 2 : }
6145 2 : }
6146 2 :
6147 2 : Ok(())
6148 2 : }
6149 :
6150 : #[tokio::test]
6151 2 : async fn test_get_vectored_aux_files() -> anyhow::Result<()> {
6152 2 : let harness = TenantHarness::create("test_get_vectored_aux_files").await?;
6153 2 :
6154 20 : let (tenant, ctx) = harness.load().await;
6155 2 : let tline = tenant
6156 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
6157 2 : .await?;
6158 2 : let tline = tline.raw_timeline().unwrap();
6159 2 :
6160 2 : let mut modification = tline.begin_modification(Lsn(0x1000));
6161 2 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
6162 2 : modification.set_lsn(Lsn(0x1008))?;
6163 2 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
6164 2 : modification.commit(&ctx).await?;
6165 2 :
6166 2 : let child_timeline_id = TimelineId::generate();
6167 2 : tenant
6168 2 : .branch_timeline_test(
6169 2 : tline,
6170 2 : child_timeline_id,
6171 2 : Some(tline.get_last_record_lsn()),
6172 2 : &ctx,
6173 2 : )
6174 2 : .await?;
6175 2 :
6176 2 : let child_timeline = tenant
6177 2 : .get_timeline(child_timeline_id, true)
6178 2 : .expect("Should have the branched timeline");
6179 2 :
6180 2 : let aux_keyspace = KeySpace {
6181 2 : ranges: vec![NON_INHERITED_RANGE],
6182 2 : };
6183 2 : let read_lsn = child_timeline.get_last_record_lsn();
6184 2 :
6185 2 : let vectored_res = child_timeline
6186 2 : .get_vectored_impl(
6187 2 : aux_keyspace.clone(),
6188 2 : read_lsn,
6189 2 : &mut ValuesReconstructState::new(),
6190 2 : &ctx,
6191 2 : )
6192 2 : .await;
6193 2 :
6194 2 : let images = vectored_res?;
6195 2 : assert!(images.is_empty());
6196 2 : Ok(())
6197 2 : }
6198 :
6199 : // Test that vectored get handles layer gaps correctly
6200 : // by advancing into the next ancestor timeline if required.
6201 : //
6202 : // The test generates timelines that look like the diagram below.
6203 : // We leave a gap in one of the L1 layers at `gap_at_key` (`/` in the diagram).
6204 : // The reconstruct data for that key lies in the ancestor timeline (`X` in the diagram).
6205 : //
6206 : // ```
6207 : //-------------------------------+
6208 : // ... |
6209 : // [ L1 ] |
6210 : // [ / L1 ] | Child Timeline
6211 : // ... |
6212 : // ------------------------------+
6213 : // [ X L1 ] | Parent Timeline
6214 : // ------------------------------+
6215 : // ```
6216 : #[tokio::test]
6217 2 : async fn test_get_vectored_key_gap() -> anyhow::Result<()> {
6218 2 : let tenant_conf = TenantConf {
6219 2 : // Make compaction deterministic
6220 2 : gc_period: Duration::ZERO,
6221 2 : compaction_period: Duration::ZERO,
6222 2 : // Encourage creation of L1 layers
6223 2 : checkpoint_distance: 16 * 1024,
6224 2 : compaction_target_size: 8 * 1024,
6225 2 : ..TenantConf::default()
6226 2 : };
6227 2 :
6228 2 : let harness = TenantHarness::create_custom(
6229 2 : "test_get_vectored_key_gap",
6230 2 : tenant_conf,
6231 2 : TenantId::generate(),
6232 2 : ShardIdentity::unsharded(),
6233 2 : Generation::new(0xdeadbeef),
6234 2 : )
6235 2 : .await?;
6236 20 : let (tenant, ctx) = harness.load().await;
6237 2 :
6238 2 : let mut current_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6239 2 : let gap_at_key = current_key.add(100);
6240 2 : let mut current_lsn = Lsn(0x10);
6241 2 :
6242 2 : const KEY_COUNT: usize = 10_000;
6243 2 :
6244 2 : let timeline_id = TimelineId::generate();
6245 2 : let current_timeline = tenant
6246 2 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
6247 6 : .await?;
6248 2 :
6249 2 : current_lsn += 0x100;
6250 2 :
6251 2 : let mut writer = current_timeline.writer().await;
6252 2 : writer
6253 2 : .put(
6254 2 : gap_at_key,
6255 2 : current_lsn,
6256 2 : &Value::Image(test_img(&format!("{} at {}", gap_at_key, current_lsn))),
6257 2 : &ctx,
6258 2 : )
6259 2 : .await?;
6260 2 : writer.finish_write(current_lsn);
6261 2 : drop(writer);
6262 2 :
6263 2 : let mut latest_lsns = HashMap::new();
6264 2 : latest_lsns.insert(gap_at_key, current_lsn);
6265 2 :
6266 2 : current_timeline.freeze_and_flush().await?;
6267 2 :
6268 2 : let child_timeline_id = TimelineId::generate();
6269 2 :
6270 2 : tenant
6271 2 : .branch_timeline_test(
6272 2 : ¤t_timeline,
6273 2 : child_timeline_id,
6274 2 : Some(current_lsn),
6275 2 : &ctx,
6276 2 : )
6277 2 : .await?;
6278 2 : let child_timeline = tenant
6279 2 : .get_timeline(child_timeline_id, true)
6280 2 : .expect("Should have the branched timeline");
6281 2 :
6282 20002 : for i in 0..KEY_COUNT {
6283 20000 : if current_key == gap_at_key {
6284 2 : current_key = current_key.next();
6285 2 : continue;
6286 19998 : }
6287 19998 :
6288 19998 : current_lsn += 0x10;
6289 2 :
6290 19998 : let mut writer = child_timeline.writer().await;
6291 19998 : writer
6292 19998 : .put(
6293 19998 : current_key,
6294 19998 : current_lsn,
6295 19998 : &Value::Image(test_img(&format!("{} at {}", current_key, current_lsn))),
6296 19998 : &ctx,
6297 19998 : )
6298 69 : .await?;
6299 19998 : writer.finish_write(current_lsn);
6300 19998 : drop(writer);
6301 19998 :
6302 19998 : latest_lsns.insert(current_key, current_lsn);
6303 19998 : current_key = current_key.next();
6304 19998 :
6305 19998 : // Flush every now and then to encourage layer file creation.
6306 19998 : if i % 500 == 0 {
6307 45 : child_timeline.freeze_and_flush().await?;
6308 19958 : }
6309 2 : }
6310 2 :
6311 2 : child_timeline.freeze_and_flush().await?;
6312 2 : let mut flags = EnumSet::new();
6313 2 : flags.insert(CompactFlags::ForceRepartition);
6314 2 : child_timeline
6315 2 : .compact(&CancellationToken::new(), flags, &ctx)
6316 1757 : .await?;
6317 2 :
6318 2 : let key_near_end = {
6319 2 : let mut tmp = current_key;
6320 2 : tmp.field6 -= 10;
6321 2 : tmp
6322 2 : };
6323 2 :
6324 2 : let key_near_gap = {
6325 2 : let mut tmp = gap_at_key;
6326 2 : tmp.field6 -= 10;
6327 2 : tmp
6328 2 : };
6329 2 :
6330 2 : let read = KeySpace {
6331 2 : ranges: vec![key_near_gap..gap_at_key.next(), key_near_end..current_key],
6332 2 : };
6333 2 : let results = child_timeline
6334 2 : .get_vectored_impl(
6335 2 : read.clone(),
6336 2 : current_lsn,
6337 2 : &mut ValuesReconstructState::new(),
6338 2 : &ctx,
6339 2 : )
6340 16 : .await?;
6341 2 :
6342 44 : for (key, img_res) in results {
6343 42 : let expected = test_img(&format!("{} at {}", key, latest_lsns[&key]));
6344 42 : assert_eq!(img_res?, expected);
6345 2 : }
6346 2 :
6347 2 : Ok(())
6348 2 : }
6349 :
6350 : // Test that vectored get descends into ancestor timelines correctly and
6351 : // does not return an image that's newer than requested.
6352 : //
6353 : // The diagram below ilustrates an interesting case. We have a parent timeline
6354 : // (top of the Lsn range) and a child timeline. The request key cannot be reconstructed
6355 : // from the child timeline, so the parent timeline must be visited. When advacing into
6356 : // the child timeline, the read path needs to remember what the requested Lsn was in
6357 : // order to avoid returning an image that's too new. The test below constructs such
6358 : // a timeline setup and does a few queries around the Lsn of each page image.
6359 : // ```
6360 : // LSN
6361 : // ^
6362 : // |
6363 : // |
6364 : // 500 | --------------------------------------> branch point
6365 : // 400 | X
6366 : // 300 | X
6367 : // 200 | --------------------------------------> requested lsn
6368 : // 100 | X
6369 : // |---------------------------------------> Key
6370 : // |
6371 : // ------> requested key
6372 : //
6373 : // Legend:
6374 : // * X - page images
6375 : // ```
6376 : #[tokio::test]
6377 2 : async fn test_get_vectored_ancestor_descent() -> anyhow::Result<()> {
6378 2 : let harness = TenantHarness::create("test_get_vectored_on_lsn_axis").await?;
6379 20 : let (tenant, ctx) = harness.load().await;
6380 2 :
6381 2 : let start_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6382 2 : let end_key = start_key.add(1000);
6383 2 : let child_gap_at_key = start_key.add(500);
6384 2 : let mut parent_gap_lsns: BTreeMap<Lsn, String> = BTreeMap::new();
6385 2 :
6386 2 : let mut current_lsn = Lsn(0x10);
6387 2 :
6388 2 : let timeline_id = TimelineId::generate();
6389 2 : let parent_timeline = tenant
6390 2 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
6391 6 : .await?;
6392 2 :
6393 2 : current_lsn += 0x100;
6394 2 :
6395 8 : for _ in 0..3 {
6396 6 : let mut key = start_key;
6397 6006 : while key < end_key {
6398 6000 : current_lsn += 0x10;
6399 6000 :
6400 6000 : let image_value = format!("{} at {}", child_gap_at_key, current_lsn);
6401 2 :
6402 6000 : let mut writer = parent_timeline.writer().await;
6403 6000 : writer
6404 6000 : .put(
6405 6000 : key,
6406 6000 : current_lsn,
6407 6000 : &Value::Image(test_img(&image_value)),
6408 6000 : &ctx,
6409 6000 : )
6410 6 : .await?;
6411 6000 : writer.finish_write(current_lsn);
6412 6000 :
6413 6000 : if key == child_gap_at_key {
6414 6 : parent_gap_lsns.insert(current_lsn, image_value);
6415 5994 : }
6416 2 :
6417 6000 : key = key.next();
6418 2 : }
6419 2 :
6420 6 : parent_timeline.freeze_and_flush().await?;
6421 2 : }
6422 2 :
6423 2 : let child_timeline_id = TimelineId::generate();
6424 2 :
6425 2 : let child_timeline = tenant
6426 2 : .branch_timeline_test(&parent_timeline, child_timeline_id, Some(current_lsn), &ctx)
6427 2 : .await?;
6428 2 :
6429 2 : let mut key = start_key;
6430 2002 : while key < end_key {
6431 2000 : if key == child_gap_at_key {
6432 2 : key = key.next();
6433 2 : continue;
6434 1998 : }
6435 1998 :
6436 1998 : current_lsn += 0x10;
6437 2 :
6438 1998 : let mut writer = child_timeline.writer().await;
6439 1998 : writer
6440 1998 : .put(
6441 1998 : key,
6442 1998 : current_lsn,
6443 1998 : &Value::Image(test_img(&format!("{} at {}", key, current_lsn))),
6444 1998 : &ctx,
6445 1998 : )
6446 17 : .await?;
6447 1998 : writer.finish_write(current_lsn);
6448 1998 :
6449 1998 : key = key.next();
6450 2 : }
6451 2 :
6452 2 : child_timeline.freeze_and_flush().await?;
6453 2 :
6454 2 : let lsn_offsets: [i64; 5] = [-10, -1, 0, 1, 10];
6455 2 : let mut query_lsns = Vec::new();
6456 6 : for image_lsn in parent_gap_lsns.keys().rev() {
6457 36 : for offset in lsn_offsets {
6458 30 : query_lsns.push(Lsn(image_lsn
6459 30 : .0
6460 30 : .checked_add_signed(offset)
6461 30 : .expect("Shouldn't overflow")));
6462 30 : }
6463 2 : }
6464 2 :
6465 32 : for query_lsn in query_lsns {
6466 30 : let results = child_timeline
6467 30 : .get_vectored_impl(
6468 30 : KeySpace {
6469 30 : ranges: vec![child_gap_at_key..child_gap_at_key.next()],
6470 30 : },
6471 30 : query_lsn,
6472 30 : &mut ValuesReconstructState::new(),
6473 30 : &ctx,
6474 30 : )
6475 29 : .await;
6476 2 :
6477 30 : let expected_item = parent_gap_lsns
6478 30 : .iter()
6479 30 : .rev()
6480 68 : .find(|(lsn, _)| **lsn <= query_lsn);
6481 30 :
6482 30 : info!(
6483 2 : "Doing vectored read at LSN {}. Expecting image to be: {:?}",
6484 2 : query_lsn, expected_item
6485 2 : );
6486 2 :
6487 30 : match expected_item {
6488 26 : Some((_, img_value)) => {
6489 26 : let key_results = results.expect("No vectored get error expected");
6490 26 : let key_result = &key_results[&child_gap_at_key];
6491 26 : let returned_img = key_result
6492 26 : .as_ref()
6493 26 : .expect("No page reconstruct error expected");
6494 26 :
6495 26 : info!(
6496 2 : "Vectored read at LSN {} returned image {}",
6497 0 : query_lsn,
6498 0 : std::str::from_utf8(returned_img)?
6499 2 : );
6500 26 : assert_eq!(*returned_img, test_img(img_value));
6501 2 : }
6502 2 : None => {
6503 4 : assert!(matches!(results, Err(GetVectoredError::MissingKey(_))));
6504 2 : }
6505 2 : }
6506 2 : }
6507 2 :
6508 2 : Ok(())
6509 2 : }
6510 :
6511 : #[tokio::test]
6512 2 : async fn test_random_updates() -> anyhow::Result<()> {
6513 2 : let names_algorithms = [
6514 2 : ("test_random_updates_legacy", CompactionAlgorithm::Legacy),
6515 2 : ("test_random_updates_tiered", CompactionAlgorithm::Tiered),
6516 2 : ];
6517 6 : for (name, algorithm) in names_algorithms {
6518 96208 : test_random_updates_algorithm(name, algorithm).await?;
6519 2 : }
6520 2 : Ok(())
6521 2 : }
6522 :
6523 4 : async fn test_random_updates_algorithm(
6524 4 : name: &'static str,
6525 4 : compaction_algorithm: CompactionAlgorithm,
6526 4 : ) -> anyhow::Result<()> {
6527 4 : let mut harness = TenantHarness::create(name).await?;
6528 4 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
6529 4 : kind: compaction_algorithm,
6530 4 : };
6531 40 : let (tenant, ctx) = harness.load().await;
6532 4 : let tline = tenant
6533 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6534 11 : .await?;
6535 :
6536 : const NUM_KEYS: usize = 1000;
6537 4 : let cancel = CancellationToken::new();
6538 4 :
6539 4 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6540 4 : let mut test_key_end = test_key;
6541 4 : test_key_end.field6 = NUM_KEYS as u32;
6542 4 : tline.add_extra_test_dense_keyspace(KeySpace::single(test_key..test_key_end));
6543 4 :
6544 4 : let mut keyspace = KeySpaceAccum::new();
6545 4 :
6546 4 : // Track when each page was last modified. Used to assert that
6547 4 : // a read sees the latest page version.
6548 4 : let mut updated = [Lsn(0); NUM_KEYS];
6549 4 :
6550 4 : let mut lsn = Lsn(0x10);
6551 : #[allow(clippy::needless_range_loop)]
6552 4004 : for blknum in 0..NUM_KEYS {
6553 4000 : lsn = Lsn(lsn.0 + 0x10);
6554 4000 : test_key.field6 = blknum as u32;
6555 4000 : let mut writer = tline.writer().await;
6556 4000 : writer
6557 4000 : .put(
6558 4000 : test_key,
6559 4000 : lsn,
6560 4000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6561 4000 : &ctx,
6562 4000 : )
6563 4 : .await?;
6564 4000 : writer.finish_write(lsn);
6565 4000 : updated[blknum] = lsn;
6566 4000 : drop(writer);
6567 4000 :
6568 4000 : keyspace.add_key(test_key);
6569 : }
6570 :
6571 204 : for _ in 0..50 {
6572 200200 : for _ in 0..NUM_KEYS {
6573 200000 : lsn = Lsn(lsn.0 + 0x10);
6574 200000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6575 200000 : test_key.field6 = blknum as u32;
6576 200000 : let mut writer = tline.writer().await;
6577 200000 : writer
6578 200000 : .put(
6579 200000 : test_key,
6580 200000 : lsn,
6581 200000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6582 200000 : &ctx,
6583 200000 : )
6584 198 : .await?;
6585 200000 : writer.finish_write(lsn);
6586 200000 : drop(writer);
6587 200000 : updated[blknum] = lsn;
6588 : }
6589 :
6590 : // Read all the blocks
6591 200000 : for (blknum, last_lsn) in updated.iter().enumerate() {
6592 200000 : test_key.field6 = blknum as u32;
6593 200000 : assert_eq!(
6594 200000 : tline.get(test_key, lsn, &ctx).await?,
6595 200000 : test_img(&format!("{} at {}", blknum, last_lsn))
6596 : );
6597 : }
6598 :
6599 : // Perform a cycle of flush, and GC
6600 203 : tline.freeze_and_flush().await?;
6601 200 : tenant
6602 200 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
6603 0 : .await?;
6604 : }
6605 :
6606 4 : Ok(())
6607 4 : }
6608 :
6609 : #[tokio::test]
6610 2 : async fn test_traverse_branches() -> anyhow::Result<()> {
6611 2 : let (tenant, ctx) = TenantHarness::create("test_traverse_branches")
6612 2 : .await?
6613 2 : .load()
6614 20 : .await;
6615 2 : let mut tline = tenant
6616 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6617 6 : .await?;
6618 2 :
6619 2 : const NUM_KEYS: usize = 1000;
6620 2 :
6621 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6622 2 :
6623 2 : let mut keyspace = KeySpaceAccum::new();
6624 2 :
6625 2 : let cancel = CancellationToken::new();
6626 2 :
6627 2 : // Track when each page was last modified. Used to assert that
6628 2 : // a read sees the latest page version.
6629 2 : let mut updated = [Lsn(0); NUM_KEYS];
6630 2 :
6631 2 : let mut lsn = Lsn(0x10);
6632 2 : #[allow(clippy::needless_range_loop)]
6633 2002 : for blknum in 0..NUM_KEYS {
6634 2000 : lsn = Lsn(lsn.0 + 0x10);
6635 2000 : test_key.field6 = blknum as u32;
6636 2000 : let mut writer = tline.writer().await;
6637 2000 : writer
6638 2000 : .put(
6639 2000 : test_key,
6640 2000 : lsn,
6641 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6642 2000 : &ctx,
6643 2000 : )
6644 2 : .await?;
6645 2000 : writer.finish_write(lsn);
6646 2000 : updated[blknum] = lsn;
6647 2000 : drop(writer);
6648 2000 :
6649 2000 : keyspace.add_key(test_key);
6650 2 : }
6651 2 :
6652 102 : for _ in 0..50 {
6653 100 : let new_tline_id = TimelineId::generate();
6654 100 : tenant
6655 100 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
6656 87 : .await?;
6657 100 : tline = tenant
6658 100 : .get_timeline(new_tline_id, true)
6659 100 : .expect("Should have the branched timeline");
6660 2 :
6661 100100 : for _ in 0..NUM_KEYS {
6662 100000 : lsn = Lsn(lsn.0 + 0x10);
6663 100000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6664 100000 : test_key.field6 = blknum as u32;
6665 100000 : let mut writer = tline.writer().await;
6666 100000 : writer
6667 100000 : .put(
6668 100000 : test_key,
6669 100000 : lsn,
6670 100000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6671 100000 : &ctx,
6672 100000 : )
6673 820 : .await?;
6674 100000 : println!("updating {} at {}", blknum, lsn);
6675 100000 : writer.finish_write(lsn);
6676 100000 : drop(writer);
6677 100000 : updated[blknum] = lsn;
6678 2 : }
6679 2 :
6680 2 : // Read all the blocks
6681 100000 : for (blknum, last_lsn) in updated.iter().enumerate() {
6682 100000 : test_key.field6 = blknum as u32;
6683 100000 : assert_eq!(
6684 100000 : tline.get(test_key, lsn, &ctx).await?,
6685 100000 : test_img(&format!("{} at {}", blknum, last_lsn))
6686 2 : );
6687 2 : }
6688 2 :
6689 2 : // Perform a cycle of flush, compact, and GC
6690 100 : tline.freeze_and_flush().await?;
6691 14885 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
6692 100 : tenant
6693 100 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
6694 2 : .await?;
6695 2 : }
6696 2 :
6697 2 : Ok(())
6698 2 : }
6699 :
6700 : #[tokio::test]
6701 2 : async fn test_traverse_ancestors() -> anyhow::Result<()> {
6702 2 : let (tenant, ctx) = TenantHarness::create("test_traverse_ancestors")
6703 2 : .await?
6704 2 : .load()
6705 20 : .await;
6706 2 : let mut tline = tenant
6707 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6708 6 : .await?;
6709 2 :
6710 2 : const NUM_KEYS: usize = 100;
6711 2 : const NUM_TLINES: usize = 50;
6712 2 :
6713 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6714 2 : // Track page mutation lsns across different timelines.
6715 2 : let mut updated = [[Lsn(0); NUM_KEYS]; NUM_TLINES];
6716 2 :
6717 2 : let mut lsn = Lsn(0x10);
6718 2 :
6719 2 : #[allow(clippy::needless_range_loop)]
6720 102 : for idx in 0..NUM_TLINES {
6721 100 : let new_tline_id = TimelineId::generate();
6722 100 : tenant
6723 100 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
6724 80 : .await?;
6725 100 : tline = tenant
6726 100 : .get_timeline(new_tline_id, true)
6727 100 : .expect("Should have the branched timeline");
6728 2 :
6729 10100 : for _ in 0..NUM_KEYS {
6730 10000 : lsn = Lsn(lsn.0 + 0x10);
6731 10000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6732 10000 : test_key.field6 = blknum as u32;
6733 10000 : let mut writer = tline.writer().await;
6734 10000 : writer
6735 10000 : .put(
6736 10000 : test_key,
6737 10000 : lsn,
6738 10000 : &Value::Image(test_img(&format!("{} {} at {}", idx, blknum, lsn))),
6739 10000 : &ctx,
6740 10000 : )
6741 108 : .await?;
6742 10000 : println!("updating [{}][{}] at {}", idx, blknum, lsn);
6743 10000 : writer.finish_write(lsn);
6744 10000 : drop(writer);
6745 10000 : updated[idx][blknum] = lsn;
6746 2 : }
6747 2 : }
6748 2 :
6749 2 : // Read pages from leaf timeline across all ancestors.
6750 100 : for (idx, lsns) in updated.iter().enumerate() {
6751 10000 : for (blknum, lsn) in lsns.iter().enumerate() {
6752 2 : // Skip empty mutations.
6753 10000 : if lsn.0 == 0 {
6754 3681 : continue;
6755 6319 : }
6756 6319 : println!("checking [{idx}][{blknum}] at {lsn}");
6757 6319 : test_key.field6 = blknum as u32;
6758 6319 : assert_eq!(
6759 6319 : tline.get(test_key, *lsn, &ctx).await?,
6760 6319 : test_img(&format!("{idx} {blknum} at {lsn}"))
6761 2 : );
6762 2 : }
6763 2 : }
6764 2 : Ok(())
6765 2 : }
6766 :
6767 : #[tokio::test]
6768 2 : async fn test_write_at_initdb_lsn_takes_optimization_code_path() -> anyhow::Result<()> {
6769 2 : let (tenant, ctx) = TenantHarness::create("test_empty_test_timeline_is_usable")
6770 2 : .await?
6771 2 : .load()
6772 20 : .await;
6773 2 :
6774 2 : let initdb_lsn = Lsn(0x20);
6775 2 : let utline = tenant
6776 2 : .create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)
6777 2 : .await?;
6778 2 : let tline = utline.raw_timeline().unwrap();
6779 2 :
6780 2 : // Spawn flush loop now so that we can set the `expect_initdb_optimization`
6781 2 : tline.maybe_spawn_flush_loop();
6782 2 :
6783 2 : // Make sure the timeline has the minimum set of required keys for operation.
6784 2 : // The only operation you can always do on an empty timeline is to `put` new data.
6785 2 : // Except if you `put` at `initdb_lsn`.
6786 2 : // In that case, there's an optimization to directly create image layers instead of delta layers.
6787 2 : // It uses `repartition()`, which assumes some keys to be present.
6788 2 : // Let's make sure the test timeline can handle that case.
6789 2 : {
6790 2 : let mut state = tline.flush_loop_state.lock().unwrap();
6791 2 : assert_eq!(
6792 2 : timeline::FlushLoopState::Running {
6793 2 : expect_initdb_optimization: false,
6794 2 : initdb_optimization_count: 0,
6795 2 : },
6796 2 : *state
6797 2 : );
6798 2 : *state = timeline::FlushLoopState::Running {
6799 2 : expect_initdb_optimization: true,
6800 2 : initdb_optimization_count: 0,
6801 2 : };
6802 2 : }
6803 2 :
6804 2 : // Make writes at the initdb_lsn. When we flush it below, it should be handled by the optimization.
6805 2 : // As explained above, the optimization requires some keys to be present.
6806 2 : // As per `create_empty_timeline` documentation, use init_empty to set them.
6807 2 : // This is what `create_test_timeline` does, by the way.
6808 2 : let mut modification = tline.begin_modification(initdb_lsn);
6809 2 : modification
6810 2 : .init_empty_test_timeline()
6811 2 : .context("init_empty_test_timeline")?;
6812 2 : modification
6813 2 : .commit(&ctx)
6814 2 : .await
6815 2 : .context("commit init_empty_test_timeline modification")?;
6816 2 :
6817 2 : // Do the flush. The flush code will check the expectations that we set above.
6818 2 : tline.freeze_and_flush().await?;
6819 2 :
6820 2 : // assert freeze_and_flush exercised the initdb optimization
6821 2 : {
6822 2 : let state = tline.flush_loop_state.lock().unwrap();
6823 2 : let timeline::FlushLoopState::Running {
6824 2 : expect_initdb_optimization,
6825 2 : initdb_optimization_count,
6826 2 : } = *state
6827 2 : else {
6828 2 : panic!("unexpected state: {:?}", *state);
6829 2 : };
6830 2 : assert!(expect_initdb_optimization);
6831 2 : assert!(initdb_optimization_count > 0);
6832 2 : }
6833 2 : Ok(())
6834 2 : }
6835 :
6836 : #[tokio::test]
6837 2 : async fn test_create_guard_crash() -> anyhow::Result<()> {
6838 2 : let name = "test_create_guard_crash";
6839 2 : let harness = TenantHarness::create(name).await?;
6840 2 : {
6841 20 : let (tenant, ctx) = harness.load().await;
6842 2 : let tline = tenant
6843 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
6844 2 : .await?;
6845 2 : // Leave the timeline ID in [`Tenant::timelines_creating`] to exclude attempting to create it again
6846 2 : let raw_tline = tline.raw_timeline().unwrap();
6847 2 : raw_tline
6848 2 : .shutdown(super::timeline::ShutdownMode::Hard)
6849 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))
6850 2 : .await;
6851 2 : std::mem::forget(tline);
6852 2 : }
6853 2 :
6854 10 : let (tenant, _) = harness.load().await;
6855 2 : match tenant.get_timeline(TIMELINE_ID, false) {
6856 2 : Ok(_) => panic!("timeline should've been removed during load"),
6857 2 : Err(e) => {
6858 2 : assert_eq!(
6859 2 : e,
6860 2 : GetTimelineError::NotFound {
6861 2 : tenant_id: tenant.tenant_shard_id,
6862 2 : timeline_id: TIMELINE_ID,
6863 2 : }
6864 2 : )
6865 2 : }
6866 2 : }
6867 2 :
6868 2 : assert!(!harness
6869 2 : .conf
6870 2 : .timeline_path(&tenant.tenant_shard_id, &TIMELINE_ID)
6871 2 : .exists());
6872 2 :
6873 2 : Ok(())
6874 2 : }
6875 :
6876 : #[tokio::test]
6877 2 : async fn test_read_at_max_lsn() -> anyhow::Result<()> {
6878 2 : let names_algorithms = [
6879 2 : ("test_read_at_max_lsn_legacy", CompactionAlgorithm::Legacy),
6880 2 : ("test_read_at_max_lsn_tiered", CompactionAlgorithm::Tiered),
6881 2 : ];
6882 6 : for (name, algorithm) in names_algorithms {
6883 32561 : test_read_at_max_lsn_algorithm(name, algorithm).await?;
6884 2 : }
6885 2 : Ok(())
6886 2 : }
6887 :
6888 4 : async fn test_read_at_max_lsn_algorithm(
6889 4 : name: &'static str,
6890 4 : compaction_algorithm: CompactionAlgorithm,
6891 4 : ) -> anyhow::Result<()> {
6892 4 : let mut harness = TenantHarness::create(name).await?;
6893 4 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
6894 4 : kind: compaction_algorithm,
6895 4 : };
6896 40 : let (tenant, ctx) = harness.load().await;
6897 4 : let tline = tenant
6898 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6899 11 : .await?;
6900 :
6901 4 : let lsn = Lsn(0x10);
6902 4 : let compact = false;
6903 32100 : bulk_insert_maybe_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000, compact).await?;
6904 :
6905 4 : let test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6906 4 : let read_lsn = Lsn(u64::MAX - 1);
6907 :
6908 410 : let result = tline.get(test_key, read_lsn, &ctx).await;
6909 4 : assert!(result.is_ok(), "result is not Ok: {}", result.unwrap_err());
6910 :
6911 4 : Ok(())
6912 4 : }
6913 :
6914 : #[tokio::test]
6915 2 : async fn test_metadata_scan() -> anyhow::Result<()> {
6916 2 : let harness = TenantHarness::create("test_metadata_scan").await?;
6917 14 : let (tenant, ctx) = harness.load().await;
6918 2 : let tline = tenant
6919 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6920 5 : .await?;
6921 2 :
6922 2 : const NUM_KEYS: usize = 1000;
6923 2 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
6924 2 :
6925 2 : let cancel = CancellationToken::new();
6926 2 :
6927 2 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
6928 2 : base_key.field1 = AUX_KEY_PREFIX;
6929 2 : let mut test_key = base_key;
6930 2 :
6931 2 : // Track when each page was last modified. Used to assert that
6932 2 : // a read sees the latest page version.
6933 2 : let mut updated = [Lsn(0); NUM_KEYS];
6934 2 :
6935 2 : let mut lsn = Lsn(0x10);
6936 2 : #[allow(clippy::needless_range_loop)]
6937 2002 : for blknum in 0..NUM_KEYS {
6938 2000 : lsn = Lsn(lsn.0 + 0x10);
6939 2000 : test_key.field6 = (blknum * STEP) as u32;
6940 2000 : let mut writer = tline.writer().await;
6941 2000 : writer
6942 2000 : .put(
6943 2000 : test_key,
6944 2000 : lsn,
6945 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6946 2000 : &ctx,
6947 2000 : )
6948 2 : .await?;
6949 2000 : writer.finish_write(lsn);
6950 2000 : updated[blknum] = lsn;
6951 2000 : drop(writer);
6952 2 : }
6953 2 :
6954 2 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
6955 2 :
6956 24 : for iter in 0..=10 {
6957 2 : // Read all the blocks
6958 22000 : for (blknum, last_lsn) in updated.iter().enumerate() {
6959 22000 : test_key.field6 = (blknum * STEP) as u32;
6960 22000 : assert_eq!(
6961 22000 : tline.get(test_key, lsn, &ctx).await?,
6962 22000 : test_img(&format!("{} at {}", blknum, last_lsn))
6963 2 : );
6964 2 : }
6965 2 :
6966 22 : let mut cnt = 0;
6967 22000 : for (key, value) in tline
6968 22 : .get_vectored_impl(
6969 22 : keyspace.clone(),
6970 22 : lsn,
6971 22 : &mut ValuesReconstructState::default(),
6972 22 : &ctx,
6973 22 : )
6974 714 : .await?
6975 2 : {
6976 22000 : let blknum = key.field6 as usize;
6977 22000 : let value = value?;
6978 22000 : assert!(blknum % STEP == 0);
6979 22000 : let blknum = blknum / STEP;
6980 22000 : assert_eq!(
6981 22000 : value,
6982 22000 : test_img(&format!("{} at {}", blknum, updated[blknum]))
6983 22000 : );
6984 22000 : cnt += 1;
6985 2 : }
6986 2 :
6987 22 : assert_eq!(cnt, NUM_KEYS);
6988 2 :
6989 22022 : for _ in 0..NUM_KEYS {
6990 22000 : lsn = Lsn(lsn.0 + 0x10);
6991 22000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6992 22000 : test_key.field6 = (blknum * STEP) as u32;
6993 22000 : let mut writer = tline.writer().await;
6994 22000 : writer
6995 22000 : .put(
6996 22000 : test_key,
6997 22000 : lsn,
6998 22000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6999 22000 : &ctx,
7000 22000 : )
7001 110 : .await?;
7002 22000 : writer.finish_write(lsn);
7003 22000 : drop(writer);
7004 22000 : updated[blknum] = lsn;
7005 2 : }
7006 2 :
7007 2 : // Perform two cycles of flush, compact, and GC
7008 66 : for round in 0..2 {
7009 44 : tline.freeze_and_flush().await?;
7010 44 : tline
7011 44 : .compact(
7012 44 : &cancel,
7013 44 : if iter % 5 == 0 && round == 0 {
7014 6 : let mut flags = EnumSet::new();
7015 6 : flags.insert(CompactFlags::ForceImageLayerCreation);
7016 6 : flags.insert(CompactFlags::ForceRepartition);
7017 6 : flags
7018 2 : } else {
7019 38 : EnumSet::empty()
7020 2 : },
7021 44 : &ctx,
7022 2 : )
7023 6689 : .await?;
7024 44 : tenant
7025 44 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7026 2 : .await?;
7027 2 : }
7028 2 : }
7029 2 :
7030 2 : Ok(())
7031 2 : }
7032 :
7033 : #[tokio::test]
7034 2 : async fn test_metadata_compaction_trigger() -> anyhow::Result<()> {
7035 2 : let harness = TenantHarness::create("test_metadata_compaction_trigger").await?;
7036 20 : let (tenant, ctx) = harness.load().await;
7037 2 : let tline = tenant
7038 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7039 6 : .await?;
7040 2 :
7041 2 : let cancel = CancellationToken::new();
7042 2 :
7043 2 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7044 2 : base_key.field1 = AUX_KEY_PREFIX;
7045 2 : let test_key = base_key;
7046 2 : let mut lsn = Lsn(0x10);
7047 2 :
7048 42 : for _ in 0..20 {
7049 40 : lsn = Lsn(lsn.0 + 0x10);
7050 40 : let mut writer = tline.writer().await;
7051 40 : writer
7052 40 : .put(
7053 40 : test_key,
7054 40 : lsn,
7055 40 : &Value::Image(test_img(&format!("{} at {}", 0, lsn))),
7056 40 : &ctx,
7057 40 : )
7058 20 : .await?;
7059 40 : writer.finish_write(lsn);
7060 40 : drop(writer);
7061 43 : tline.freeze_and_flush().await?; // force create a delta layer
7062 2 : }
7063 2 :
7064 2 : let before_num_l0_delta_files =
7065 2 : tline.layers.read().await.layer_map()?.level0_deltas().len();
7066 2 :
7067 110 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
7068 2 :
7069 2 : let after_num_l0_delta_files = tline.layers.read().await.layer_map()?.level0_deltas().len();
7070 2 :
7071 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}");
7072 2 :
7073 2 : assert_eq!(
7074 4 : tline.get(test_key, lsn, &ctx).await?,
7075 2 : test_img(&format!("{} at {}", 0, lsn))
7076 2 : );
7077 2 :
7078 2 : Ok(())
7079 2 : }
7080 :
7081 : #[tokio::test]
7082 2 : async fn test_aux_file_e2e() {
7083 2 : let harness = TenantHarness::create("test_aux_file_e2e").await.unwrap();
7084 2 :
7085 20 : let (tenant, ctx) = harness.load().await;
7086 2 :
7087 2 : let mut lsn = Lsn(0x08);
7088 2 :
7089 2 : let tline: Arc<Timeline> = tenant
7090 2 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
7091 6 : .await
7092 2 : .unwrap();
7093 2 :
7094 2 : {
7095 2 : lsn += 8;
7096 2 : let mut modification = tline.begin_modification(lsn);
7097 2 : modification
7098 2 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
7099 2 : .await
7100 2 : .unwrap();
7101 2 : modification.commit(&ctx).await.unwrap();
7102 2 : }
7103 2 :
7104 2 : // we can read everything from the storage
7105 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
7106 2 : assert_eq!(
7107 2 : files.get("pg_logical/mappings/test1"),
7108 2 : Some(&bytes::Bytes::from_static(b"first"))
7109 2 : );
7110 2 :
7111 2 : {
7112 2 : lsn += 8;
7113 2 : let mut modification = tline.begin_modification(lsn);
7114 2 : modification
7115 2 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
7116 2 : .await
7117 2 : .unwrap();
7118 2 : modification.commit(&ctx).await.unwrap();
7119 2 : }
7120 2 :
7121 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
7122 2 : assert_eq!(
7123 2 : files.get("pg_logical/mappings/test2"),
7124 2 : Some(&bytes::Bytes::from_static(b"second"))
7125 2 : );
7126 2 :
7127 2 : let child = tenant
7128 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(lsn), &ctx)
7129 2 : .await
7130 2 : .unwrap();
7131 2 :
7132 2 : let files = child.list_aux_files(lsn, &ctx).await.unwrap();
7133 2 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
7134 2 : assert_eq!(files.get("pg_logical/mappings/test2"), None);
7135 2 : }
7136 :
7137 : #[tokio::test]
7138 2 : async fn test_metadata_image_creation() -> anyhow::Result<()> {
7139 2 : let harness = TenantHarness::create("test_metadata_image_creation").await?;
7140 20 : let (tenant, ctx) = harness.load().await;
7141 2 : let tline = tenant
7142 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7143 6 : .await?;
7144 2 :
7145 2 : const NUM_KEYS: usize = 1000;
7146 2 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
7147 2 :
7148 2 : let cancel = CancellationToken::new();
7149 2 :
7150 2 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7151 2 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
7152 2 : let mut test_key = base_key;
7153 2 : let mut lsn = Lsn(0x10);
7154 2 :
7155 8 : async fn scan_with_statistics(
7156 8 : tline: &Timeline,
7157 8 : keyspace: &KeySpace,
7158 8 : lsn: Lsn,
7159 8 : ctx: &RequestContext,
7160 8 : ) -> anyhow::Result<(BTreeMap<Key, Result<Bytes, PageReconstructError>>, usize)> {
7161 8 : let mut reconstruct_state = ValuesReconstructState::default();
7162 8 : let res = tline
7163 8 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
7164 231 : .await?;
7165 8 : Ok((res, reconstruct_state.get_delta_layers_visited() as usize))
7166 8 : }
7167 2 :
7168 2 : #[allow(clippy::needless_range_loop)]
7169 2002 : for blknum in 0..NUM_KEYS {
7170 2000 : lsn = Lsn(lsn.0 + 0x10);
7171 2000 : test_key.field6 = (blknum * STEP) as u32;
7172 2000 : let mut writer = tline.writer().await;
7173 2000 : writer
7174 2000 : .put(
7175 2000 : test_key,
7176 2000 : lsn,
7177 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7178 2000 : &ctx,
7179 2000 : )
7180 2 : .await?;
7181 2000 : writer.finish_write(lsn);
7182 2000 : drop(writer);
7183 2 : }
7184 2 :
7185 2 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
7186 2 :
7187 22 : for iter in 1..=10 {
7188 20020 : for _ in 0..NUM_KEYS {
7189 20000 : lsn = Lsn(lsn.0 + 0x10);
7190 20000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7191 20000 : test_key.field6 = (blknum * STEP) as u32;
7192 20000 : let mut writer = tline.writer().await;
7193 20000 : writer
7194 20000 : .put(
7195 20000 : test_key,
7196 20000 : lsn,
7197 20000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7198 20000 : &ctx,
7199 20000 : )
7200 35 : .await?;
7201 20000 : writer.finish_write(lsn);
7202 20000 : drop(writer);
7203 2 : }
7204 2 :
7205 20 : tline.freeze_and_flush().await?;
7206 2 :
7207 20 : if iter % 5 == 0 {
7208 4 : let (_, before_delta_file_accessed) =
7209 223 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
7210 4 : tline
7211 4 : .compact(
7212 4 : &cancel,
7213 4 : {
7214 4 : let mut flags = EnumSet::new();
7215 4 : flags.insert(CompactFlags::ForceImageLayerCreation);
7216 4 : flags.insert(CompactFlags::ForceRepartition);
7217 4 : flags
7218 4 : },
7219 4 : &ctx,
7220 4 : )
7221 4811 : .await?;
7222 4 : let (_, after_delta_file_accessed) =
7223 8 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
7224 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}");
7225 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.
7226 4 : assert!(
7227 4 : after_delta_file_accessed <= 2,
7228 2 : "after_delta_file_accessed={after_delta_file_accessed}"
7229 2 : );
7230 16 : }
7231 2 : }
7232 2 :
7233 2 : Ok(())
7234 2 : }
7235 :
7236 : #[tokio::test]
7237 2 : async fn test_vectored_missing_data_key_reads() -> anyhow::Result<()> {
7238 2 : let harness = TenantHarness::create("test_vectored_missing_data_key_reads").await?;
7239 20 : let (tenant, ctx) = harness.load().await;
7240 2 :
7241 2 : let base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7242 2 : let base_key_child = Key::from_hex("000000000033333333444444445500000001").unwrap();
7243 2 : let base_key_nonexist = Key::from_hex("000000000033333333444444445500000002").unwrap();
7244 2 :
7245 2 : let tline = tenant
7246 2 : .create_test_timeline_with_layers(
7247 2 : TIMELINE_ID,
7248 2 : Lsn(0x10),
7249 2 : DEFAULT_PG_VERSION,
7250 2 : &ctx,
7251 2 : Vec::new(), // delta layers
7252 2 : vec![(Lsn(0x20), vec![(base_key, test_img("data key 1"))])], // image layers
7253 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
7254 2 : )
7255 13 : .await?;
7256 2 : tline.add_extra_test_dense_keyspace(KeySpace::single(base_key..(base_key_nonexist.next())));
7257 2 :
7258 2 : let child = tenant
7259 2 : .branch_timeline_test_with_layers(
7260 2 : &tline,
7261 2 : NEW_TIMELINE_ID,
7262 2 : Some(Lsn(0x20)),
7263 2 : &ctx,
7264 2 : Vec::new(), // delta layers
7265 2 : vec![(Lsn(0x30), vec![(base_key_child, test_img("data key 2"))])], // image layers
7266 2 : Lsn(0x30),
7267 2 : )
7268 9 : .await
7269 2 : .unwrap();
7270 2 :
7271 2 : let lsn = Lsn(0x30);
7272 2 :
7273 2 : // test vectored get on parent timeline
7274 2 : assert_eq!(
7275 4 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
7276 2 : Some(test_img("data key 1"))
7277 2 : );
7278 2 : assert!(get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx)
7279 3 : .await
7280 2 : .unwrap_err()
7281 2 : .is_missing_key_error());
7282 2 : assert!(
7283 2 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx)
7284 2 : .await
7285 2 : .unwrap_err()
7286 2 : .is_missing_key_error()
7287 2 : );
7288 2 :
7289 2 : // test vectored get on child timeline
7290 2 : assert_eq!(
7291 2 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
7292 2 : Some(test_img("data key 1"))
7293 2 : );
7294 2 : assert_eq!(
7295 4 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
7296 2 : Some(test_img("data key 2"))
7297 2 : );
7298 2 : assert!(
7299 2 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx)
7300 2 : .await
7301 2 : .unwrap_err()
7302 2 : .is_missing_key_error()
7303 2 : );
7304 2 :
7305 2 : Ok(())
7306 2 : }
7307 :
7308 : #[tokio::test]
7309 2 : async fn test_vectored_missing_metadata_key_reads() -> anyhow::Result<()> {
7310 2 : let harness = TenantHarness::create("test_vectored_missing_metadata_key_reads").await?;
7311 20 : let (tenant, ctx) = harness.load().await;
7312 2 :
7313 2 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7314 2 : let base_key_child = Key::from_hex("620000000033333333444444445500000001").unwrap();
7315 2 : let base_key_nonexist = Key::from_hex("620000000033333333444444445500000002").unwrap();
7316 2 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
7317 2 :
7318 2 : let tline = tenant
7319 2 : .create_test_timeline_with_layers(
7320 2 : TIMELINE_ID,
7321 2 : Lsn(0x10),
7322 2 : DEFAULT_PG_VERSION,
7323 2 : &ctx,
7324 2 : Vec::new(), // delta layers
7325 2 : vec![(Lsn(0x20), vec![(base_key, test_img("metadata key 1"))])], // image layers
7326 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
7327 2 : )
7328 13 : .await?;
7329 2 :
7330 2 : let child = tenant
7331 2 : .branch_timeline_test_with_layers(
7332 2 : &tline,
7333 2 : NEW_TIMELINE_ID,
7334 2 : Some(Lsn(0x20)),
7335 2 : &ctx,
7336 2 : Vec::new(), // delta layers
7337 2 : vec![(
7338 2 : Lsn(0x30),
7339 2 : vec![(base_key_child, test_img("metadata key 2"))],
7340 2 : )], // image layers
7341 2 : Lsn(0x30),
7342 2 : )
7343 9 : .await
7344 2 : .unwrap();
7345 2 :
7346 2 : let lsn = Lsn(0x30);
7347 2 :
7348 2 : // test vectored get on parent timeline
7349 2 : assert_eq!(
7350 4 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
7351 2 : Some(test_img("metadata key 1"))
7352 2 : );
7353 2 : assert_eq!(
7354 2 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx).await?,
7355 2 : None
7356 2 : );
7357 2 : assert_eq!(
7358 2 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx).await?,
7359 2 : None
7360 2 : );
7361 2 :
7362 2 : // test vectored get on child timeline
7363 2 : assert_eq!(
7364 2 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
7365 2 : None
7366 2 : );
7367 2 : assert_eq!(
7368 4 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
7369 2 : Some(test_img("metadata key 2"))
7370 2 : );
7371 2 : assert_eq!(
7372 2 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx).await?,
7373 2 : None
7374 2 : );
7375 2 :
7376 2 : Ok(())
7377 2 : }
7378 :
7379 36 : async fn get_vectored_impl_wrapper(
7380 36 : tline: &Arc<Timeline>,
7381 36 : key: Key,
7382 36 : lsn: Lsn,
7383 36 : ctx: &RequestContext,
7384 36 : ) -> Result<Option<Bytes>, GetVectoredError> {
7385 36 : let mut reconstruct_state = ValuesReconstructState::new();
7386 36 : let mut res = tline
7387 36 : .get_vectored_impl(
7388 36 : KeySpace::single(key..key.next()),
7389 36 : lsn,
7390 36 : &mut reconstruct_state,
7391 36 : ctx,
7392 36 : )
7393 40 : .await?;
7394 30 : Ok(res.pop_last().map(|(k, v)| {
7395 18 : assert_eq!(k, key);
7396 18 : v.unwrap()
7397 30 : }))
7398 36 : }
7399 :
7400 : #[tokio::test]
7401 2 : async fn test_metadata_tombstone_reads() -> anyhow::Result<()> {
7402 2 : let harness = TenantHarness::create("test_metadata_tombstone_reads").await?;
7403 20 : let (tenant, ctx) = harness.load().await;
7404 2 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
7405 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
7406 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
7407 2 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
7408 2 :
7409 2 : // We emulate the situation that the compaction algorithm creates an image layer that removes the tombstones
7410 2 : // Lsn 0x30 key0, key3, no key1+key2
7411 2 : // Lsn 0x20 key1+key2 tomestones
7412 2 : // Lsn 0x10 key1 in image, key2 in delta
7413 2 : let tline = tenant
7414 2 : .create_test_timeline_with_layers(
7415 2 : TIMELINE_ID,
7416 2 : Lsn(0x10),
7417 2 : DEFAULT_PG_VERSION,
7418 2 : &ctx,
7419 2 : // delta layers
7420 2 : vec![
7421 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7422 2 : Lsn(0x10)..Lsn(0x20),
7423 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
7424 2 : ),
7425 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7426 2 : Lsn(0x20)..Lsn(0x30),
7427 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
7428 2 : ),
7429 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7430 2 : Lsn(0x20)..Lsn(0x30),
7431 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
7432 2 : ),
7433 2 : ],
7434 2 : // image layers
7435 2 : vec![
7436 2 : (Lsn(0x10), vec![(key1, test_img("metadata key 1"))]),
7437 2 : (
7438 2 : Lsn(0x30),
7439 2 : vec![
7440 2 : (key0, test_img("metadata key 0")),
7441 2 : (key3, test_img("metadata key 3")),
7442 2 : ],
7443 2 : ),
7444 2 : ],
7445 2 : Lsn(0x30),
7446 2 : )
7447 40 : .await?;
7448 2 :
7449 2 : let lsn = Lsn(0x30);
7450 2 : let old_lsn = Lsn(0x20);
7451 2 :
7452 2 : assert_eq!(
7453 4 : get_vectored_impl_wrapper(&tline, key0, lsn, &ctx).await?,
7454 2 : Some(test_img("metadata key 0"))
7455 2 : );
7456 2 : assert_eq!(
7457 2 : get_vectored_impl_wrapper(&tline, key1, lsn, &ctx).await?,
7458 2 : None,
7459 2 : );
7460 2 : assert_eq!(
7461 2 : get_vectored_impl_wrapper(&tline, key2, lsn, &ctx).await?,
7462 2 : None,
7463 2 : );
7464 2 : assert_eq!(
7465 8 : get_vectored_impl_wrapper(&tline, key1, old_lsn, &ctx).await?,
7466 2 : Some(Bytes::new()),
7467 2 : );
7468 2 : assert_eq!(
7469 7 : get_vectored_impl_wrapper(&tline, key2, old_lsn, &ctx).await?,
7470 2 : Some(Bytes::new()),
7471 2 : );
7472 2 : assert_eq!(
7473 2 : get_vectored_impl_wrapper(&tline, key3, lsn, &ctx).await?,
7474 2 : Some(test_img("metadata key 3"))
7475 2 : );
7476 2 :
7477 2 : Ok(())
7478 2 : }
7479 :
7480 : #[tokio::test]
7481 2 : async fn test_metadata_tombstone_image_creation() {
7482 2 : let harness = TenantHarness::create("test_metadata_tombstone_image_creation")
7483 2 : .await
7484 2 : .unwrap();
7485 20 : let (tenant, ctx) = harness.load().await;
7486 2 :
7487 2 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
7488 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
7489 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
7490 2 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
7491 2 :
7492 2 : let tline = tenant
7493 2 : .create_test_timeline_with_layers(
7494 2 : TIMELINE_ID,
7495 2 : Lsn(0x10),
7496 2 : DEFAULT_PG_VERSION,
7497 2 : &ctx,
7498 2 : // delta layers
7499 2 : vec![
7500 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7501 2 : Lsn(0x10)..Lsn(0x20),
7502 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
7503 2 : ),
7504 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7505 2 : Lsn(0x20)..Lsn(0x30),
7506 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
7507 2 : ),
7508 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7509 2 : Lsn(0x20)..Lsn(0x30),
7510 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
7511 2 : ),
7512 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7513 2 : Lsn(0x30)..Lsn(0x40),
7514 2 : vec![
7515 2 : (key0, Lsn(0x30), Value::Image(test_img("metadata key 0"))),
7516 2 : (key3, Lsn(0x30), Value::Image(test_img("metadata key 3"))),
7517 2 : ],
7518 2 : ),
7519 2 : ],
7520 2 : // image layers
7521 2 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
7522 2 : Lsn(0x40),
7523 2 : )
7524 37 : .await
7525 2 : .unwrap();
7526 2 :
7527 2 : let cancel = CancellationToken::new();
7528 2 :
7529 2 : tline
7530 2 : .compact(
7531 2 : &cancel,
7532 2 : {
7533 2 : let mut flags = EnumSet::new();
7534 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
7535 2 : flags.insert(CompactFlags::ForceRepartition);
7536 2 : flags
7537 2 : },
7538 2 : &ctx,
7539 2 : )
7540 62 : .await
7541 2 : .unwrap();
7542 2 :
7543 2 : // Image layers are created at last_record_lsn
7544 2 : let images = tline
7545 2 : .inspect_image_layers(Lsn(0x40), &ctx)
7546 8 : .await
7547 2 : .unwrap()
7548 2 : .into_iter()
7549 18 : .filter(|(k, _)| k.is_metadata_key())
7550 2 : .collect::<Vec<_>>();
7551 2 : assert_eq!(images.len(), 2); // the image layer should only contain two existing keys, tombstones should be removed.
7552 2 : }
7553 :
7554 : #[tokio::test]
7555 2 : async fn test_metadata_tombstone_empty_image_creation() {
7556 2 : let harness = TenantHarness::create("test_metadata_tombstone_empty_image_creation")
7557 2 : .await
7558 2 : .unwrap();
7559 20 : let (tenant, ctx) = harness.load().await;
7560 2 :
7561 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
7562 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
7563 2 :
7564 2 : let tline = tenant
7565 2 : .create_test_timeline_with_layers(
7566 2 : TIMELINE_ID,
7567 2 : Lsn(0x10),
7568 2 : DEFAULT_PG_VERSION,
7569 2 : &ctx,
7570 2 : // delta layers
7571 2 : vec![
7572 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7573 2 : Lsn(0x10)..Lsn(0x20),
7574 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
7575 2 : ),
7576 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7577 2 : Lsn(0x20)..Lsn(0x30),
7578 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
7579 2 : ),
7580 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7581 2 : Lsn(0x20)..Lsn(0x30),
7582 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
7583 2 : ),
7584 2 : ],
7585 2 : // image layers
7586 2 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
7587 2 : Lsn(0x30),
7588 2 : )
7589 31 : .await
7590 2 : .unwrap();
7591 2 :
7592 2 : let cancel = CancellationToken::new();
7593 2 :
7594 2 : tline
7595 2 : .compact(
7596 2 : &cancel,
7597 2 : {
7598 2 : let mut flags = EnumSet::new();
7599 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
7600 2 : flags.insert(CompactFlags::ForceRepartition);
7601 2 : flags
7602 2 : },
7603 2 : &ctx,
7604 2 : )
7605 50 : .await
7606 2 : .unwrap();
7607 2 :
7608 2 : // Image layers are created at last_record_lsn
7609 2 : let images = tline
7610 2 : .inspect_image_layers(Lsn(0x30), &ctx)
7611 4 : .await
7612 2 : .unwrap()
7613 2 : .into_iter()
7614 14 : .filter(|(k, _)| k.is_metadata_key())
7615 2 : .collect::<Vec<_>>();
7616 2 : assert_eq!(images.len(), 0); // the image layer should not contain tombstones, or it is not created
7617 2 : }
7618 :
7619 : #[tokio::test]
7620 2 : async fn test_simple_bottom_most_compaction_images() -> anyhow::Result<()> {
7621 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_images").await?;
7622 20 : let (tenant, ctx) = harness.load().await;
7623 2 :
7624 102 : fn get_key(id: u32) -> Key {
7625 102 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7626 102 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7627 102 : key.field6 = id;
7628 102 : key
7629 102 : }
7630 2 :
7631 2 : // We create
7632 2 : // - one bottom-most image layer,
7633 2 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
7634 2 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
7635 2 : // - a delta layer D3 above the horizon.
7636 2 : //
7637 2 : // | D3 |
7638 2 : // | D1 |
7639 2 : // -| |-- gc horizon -----------------
7640 2 : // | | | D2 |
7641 2 : // --------- img layer ------------------
7642 2 : //
7643 2 : // What we should expact from this compaction is:
7644 2 : // | D3 |
7645 2 : // | Part of D1 |
7646 2 : // --------- img layer with D1+D2 at GC horizon------------------
7647 2 :
7648 2 : // img layer at 0x10
7649 2 : let img_layer = (0..10)
7650 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
7651 2 : .collect_vec();
7652 2 :
7653 2 : let delta1 = vec![
7654 2 : (
7655 2 : get_key(1),
7656 2 : Lsn(0x20),
7657 2 : Value::Image(Bytes::from("value 1@0x20")),
7658 2 : ),
7659 2 : (
7660 2 : get_key(2),
7661 2 : Lsn(0x30),
7662 2 : Value::Image(Bytes::from("value 2@0x30")),
7663 2 : ),
7664 2 : (
7665 2 : get_key(3),
7666 2 : Lsn(0x40),
7667 2 : Value::Image(Bytes::from("value 3@0x40")),
7668 2 : ),
7669 2 : ];
7670 2 : let delta2 = vec![
7671 2 : (
7672 2 : get_key(5),
7673 2 : Lsn(0x20),
7674 2 : Value::Image(Bytes::from("value 5@0x20")),
7675 2 : ),
7676 2 : (
7677 2 : get_key(6),
7678 2 : Lsn(0x20),
7679 2 : Value::Image(Bytes::from("value 6@0x20")),
7680 2 : ),
7681 2 : ];
7682 2 : let delta3 = vec![
7683 2 : (
7684 2 : get_key(8),
7685 2 : Lsn(0x48),
7686 2 : Value::Image(Bytes::from("value 8@0x48")),
7687 2 : ),
7688 2 : (
7689 2 : get_key(9),
7690 2 : Lsn(0x48),
7691 2 : Value::Image(Bytes::from("value 9@0x48")),
7692 2 : ),
7693 2 : ];
7694 2 :
7695 2 : let tline = tenant
7696 2 : .create_test_timeline_with_layers(
7697 2 : TIMELINE_ID,
7698 2 : Lsn(0x10),
7699 2 : DEFAULT_PG_VERSION,
7700 2 : &ctx,
7701 2 : vec![
7702 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
7703 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
7704 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
7705 2 : ], // delta layers
7706 2 : vec![(Lsn(0x10), img_layer)], // image layers
7707 2 : Lsn(0x50),
7708 2 : )
7709 49 : .await?;
7710 2 : {
7711 2 : // Update GC info
7712 2 : let mut guard = tline.gc_info.write().unwrap();
7713 2 : guard.cutoffs.time = Lsn(0x30);
7714 2 : guard.cutoffs.space = Lsn(0x30);
7715 2 : }
7716 2 :
7717 2 : let expected_result = [
7718 2 : Bytes::from_static(b"value 0@0x10"),
7719 2 : Bytes::from_static(b"value 1@0x20"),
7720 2 : Bytes::from_static(b"value 2@0x30"),
7721 2 : Bytes::from_static(b"value 3@0x40"),
7722 2 : Bytes::from_static(b"value 4@0x10"),
7723 2 : Bytes::from_static(b"value 5@0x20"),
7724 2 : Bytes::from_static(b"value 6@0x20"),
7725 2 : Bytes::from_static(b"value 7@0x10"),
7726 2 : Bytes::from_static(b"value 8@0x48"),
7727 2 : Bytes::from_static(b"value 9@0x48"),
7728 2 : ];
7729 2 :
7730 20 : for (idx, expected) in expected_result.iter().enumerate() {
7731 20 : assert_eq!(
7732 20 : tline
7733 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7734 30 : .await
7735 20 : .unwrap(),
7736 2 : expected
7737 2 : );
7738 2 : }
7739 2 :
7740 2 : let cancel = CancellationToken::new();
7741 2 : tline
7742 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
7743 56 : .await
7744 2 : .unwrap();
7745 2 :
7746 20 : for (idx, expected) in expected_result.iter().enumerate() {
7747 20 : assert_eq!(
7748 20 : tline
7749 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7750 20 : .await
7751 20 : .unwrap(),
7752 2 : expected
7753 2 : );
7754 2 : }
7755 2 :
7756 2 : // Check if the image layer at the GC horizon contains exactly what we want
7757 2 : let image_at_gc_horizon = tline
7758 2 : .inspect_image_layers(Lsn(0x30), &ctx)
7759 2 : .await
7760 2 : .unwrap()
7761 2 : .into_iter()
7762 34 : .filter(|(k, _)| k.is_metadata_key())
7763 2 : .collect::<Vec<_>>();
7764 2 :
7765 2 : assert_eq!(image_at_gc_horizon.len(), 10);
7766 2 : let expected_result = [
7767 2 : Bytes::from_static(b"value 0@0x10"),
7768 2 : Bytes::from_static(b"value 1@0x20"),
7769 2 : Bytes::from_static(b"value 2@0x30"),
7770 2 : Bytes::from_static(b"value 3@0x10"),
7771 2 : Bytes::from_static(b"value 4@0x10"),
7772 2 : Bytes::from_static(b"value 5@0x20"),
7773 2 : Bytes::from_static(b"value 6@0x20"),
7774 2 : Bytes::from_static(b"value 7@0x10"),
7775 2 : Bytes::from_static(b"value 8@0x10"),
7776 2 : Bytes::from_static(b"value 9@0x10"),
7777 2 : ];
7778 22 : for idx in 0..10 {
7779 20 : assert_eq!(
7780 20 : image_at_gc_horizon[idx],
7781 20 : (get_key(idx as u32), expected_result[idx].clone())
7782 20 : );
7783 2 : }
7784 2 :
7785 2 : // Check if old layers are removed / new layers have the expected LSN
7786 2 : let all_layers = inspect_and_sort(&tline, None).await;
7787 2 : assert_eq!(
7788 2 : all_layers,
7789 2 : vec![
7790 2 : // Image layer at GC horizon
7791 2 : PersistentLayerKey {
7792 2 : key_range: Key::MIN..Key::MAX,
7793 2 : lsn_range: Lsn(0x30)..Lsn(0x31),
7794 2 : is_delta: false
7795 2 : },
7796 2 : // The delta layer below the horizon
7797 2 : PersistentLayerKey {
7798 2 : key_range: get_key(3)..get_key(4),
7799 2 : lsn_range: Lsn(0x30)..Lsn(0x48),
7800 2 : is_delta: true
7801 2 : },
7802 2 : // The delta3 layer that should not be picked for the compaction
7803 2 : PersistentLayerKey {
7804 2 : key_range: get_key(8)..get_key(10),
7805 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
7806 2 : is_delta: true
7807 2 : }
7808 2 : ]
7809 2 : );
7810 2 :
7811 2 : // increase GC horizon and compact again
7812 2 : {
7813 2 : // Update GC info
7814 2 : let mut guard = tline.gc_info.write().unwrap();
7815 2 : guard.cutoffs.time = Lsn(0x40);
7816 2 : guard.cutoffs.space = Lsn(0x40);
7817 2 : }
7818 2 : tline
7819 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
7820 43 : .await
7821 2 : .unwrap();
7822 2 :
7823 2 : Ok(())
7824 2 : }
7825 :
7826 : #[cfg(feature = "testing")]
7827 : #[tokio::test]
7828 2 : async fn test_neon_test_record() -> anyhow::Result<()> {
7829 2 : let harness = TenantHarness::create("test_neon_test_record").await?;
7830 20 : let (tenant, ctx) = harness.load().await;
7831 2 :
7832 24 : fn get_key(id: u32) -> Key {
7833 24 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7834 24 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7835 24 : key.field6 = id;
7836 24 : key
7837 24 : }
7838 2 :
7839 2 : let delta1 = vec![
7840 2 : (
7841 2 : get_key(1),
7842 2 : Lsn(0x20),
7843 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
7844 2 : ),
7845 2 : (
7846 2 : get_key(1),
7847 2 : Lsn(0x30),
7848 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
7849 2 : ),
7850 2 : (get_key(2), Lsn(0x10), Value::Image("0x10".into())),
7851 2 : (
7852 2 : get_key(2),
7853 2 : Lsn(0x20),
7854 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
7855 2 : ),
7856 2 : (
7857 2 : get_key(2),
7858 2 : Lsn(0x30),
7859 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
7860 2 : ),
7861 2 : (get_key(3), Lsn(0x10), Value::Image("0x10".into())),
7862 2 : (
7863 2 : get_key(3),
7864 2 : Lsn(0x20),
7865 2 : Value::WalRecord(NeonWalRecord::wal_clear("c")),
7866 2 : ),
7867 2 : (get_key(4), Lsn(0x10), Value::Image("0x10".into())),
7868 2 : (
7869 2 : get_key(4),
7870 2 : Lsn(0x20),
7871 2 : Value::WalRecord(NeonWalRecord::wal_init("i")),
7872 2 : ),
7873 2 : ];
7874 2 : let image1 = vec![(get_key(1), "0x10".into())];
7875 2 :
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 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
7883 2 : Lsn(0x10)..Lsn(0x40),
7884 2 : delta1,
7885 2 : )], // delta layers
7886 2 : vec![(Lsn(0x10), image1)], // image layers
7887 2 : Lsn(0x50),
7888 2 : )
7889 19 : .await?;
7890 2 :
7891 2 : assert_eq!(
7892 8 : tline.get(get_key(1), Lsn(0x50), &ctx).await?,
7893 2 : Bytes::from_static(b"0x10,0x20,0x30")
7894 2 : );
7895 2 : assert_eq!(
7896 2 : tline.get(get_key(2), Lsn(0x50), &ctx).await?,
7897 2 : Bytes::from_static(b"0x10,0x20,0x30")
7898 2 : );
7899 2 :
7900 2 : // Need to remove the limit of "Neon WAL redo requires base image".
7901 2 :
7902 2 : // assert_eq!(tline.get(get_key(3), Lsn(0x50), &ctx).await?, Bytes::new());
7903 2 : // assert_eq!(tline.get(get_key(4), Lsn(0x50), &ctx).await?, Bytes::new());
7904 2 :
7905 2 : Ok(())
7906 2 : }
7907 :
7908 : #[tokio::test(start_paused = true)]
7909 2 : async fn test_lsn_lease() -> anyhow::Result<()> {
7910 2 : let (tenant, ctx) = TenantHarness::create("test_lsn_lease")
7911 2 : .await
7912 2 : .unwrap()
7913 2 : .load()
7914 11 : .await;
7915 2 : // Advance to the lsn lease deadline so that GC is not blocked by
7916 2 : // initial transition into AttachedSingle.
7917 2 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
7918 2 : tokio::time::resume();
7919 2 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7920 2 :
7921 2 : let end_lsn = Lsn(0x100);
7922 2 : let image_layers = (0x20..=0x90)
7923 2 : .step_by(0x10)
7924 16 : .map(|n| {
7925 16 : (
7926 16 : Lsn(n),
7927 16 : vec![(key, test_img(&format!("data key at {:x}", n)))],
7928 16 : )
7929 16 : })
7930 2 : .collect();
7931 2 :
7932 2 : let timeline = tenant
7933 2 : .create_test_timeline_with_layers(
7934 2 : TIMELINE_ID,
7935 2 : Lsn(0x10),
7936 2 : DEFAULT_PG_VERSION,
7937 2 : &ctx,
7938 2 : Vec::new(),
7939 2 : image_layers,
7940 2 : end_lsn,
7941 2 : )
7942 62 : .await?;
7943 2 :
7944 2 : let leased_lsns = [0x30, 0x50, 0x70];
7945 2 : let mut leases = Vec::new();
7946 6 : leased_lsns.iter().for_each(|n| {
7947 6 : leases.push(
7948 6 : timeline
7949 6 : .init_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)
7950 6 : .expect("lease request should succeed"),
7951 6 : );
7952 6 : });
7953 2 :
7954 2 : let updated_lease_0 = timeline
7955 2 : .renew_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)
7956 2 : .expect("lease renewal should succeed");
7957 2 : assert_eq!(
7958 2 : updated_lease_0.valid_until, leases[0].valid_until,
7959 2 : " Renewing with shorter lease should not change the lease."
7960 2 : );
7961 2 :
7962 2 : let updated_lease_1 = timeline
7963 2 : .renew_lsn_lease(
7964 2 : Lsn(leased_lsns[1]),
7965 2 : timeline.get_lsn_lease_length() * 2,
7966 2 : &ctx,
7967 2 : )
7968 2 : .expect("lease renewal should succeed");
7969 2 : assert!(
7970 2 : updated_lease_1.valid_until > leases[1].valid_until,
7971 2 : "Renewing with a long lease should renew lease with later expiration time."
7972 2 : );
7973 2 :
7974 2 : // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
7975 2 : info!(
7976 2 : "latest_gc_cutoff_lsn: {}",
7977 0 : *timeline.get_latest_gc_cutoff_lsn()
7978 2 : );
7979 2 : timeline.force_set_disk_consistent_lsn(end_lsn);
7980 2 :
7981 2 : let res = tenant
7982 2 : .gc_iteration(
7983 2 : Some(TIMELINE_ID),
7984 2 : 0,
7985 2 : Duration::ZERO,
7986 2 : &CancellationToken::new(),
7987 2 : &ctx,
7988 2 : )
7989 2 : .await
7990 2 : .unwrap();
7991 2 :
7992 2 : // Keeping everything <= Lsn(0x80) b/c leases:
7993 2 : // 0/10: initdb layer
7994 2 : // (0/20..=0/70).step_by(0x10): image layers added when creating the timeline.
7995 2 : assert_eq!(res.layers_needed_by_leases, 7);
7996 2 : // Keeping 0/90 b/c it is the latest layer.
7997 2 : assert_eq!(res.layers_not_updated, 1);
7998 2 : // Removed 0/80.
7999 2 : assert_eq!(res.layers_removed, 1);
8000 2 :
8001 2 : // Make lease on a already GC-ed LSN.
8002 2 : // 0/80 does not have a valid lease + is below latest_gc_cutoff
8003 2 : assert!(Lsn(0x80) < *timeline.get_latest_gc_cutoff_lsn());
8004 2 : timeline
8005 2 : .init_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx)
8006 2 : .expect_err("lease request on GC-ed LSN should fail");
8007 2 :
8008 2 : // Should still be able to renew a currently valid lease
8009 2 : // Assumption: original lease to is still valid for 0/50.
8010 2 : // (use `Timeline::init_lsn_lease` for testing so it always does validation)
8011 2 : timeline
8012 2 : .init_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)
8013 2 : .expect("lease renewal with validation should succeed");
8014 2 :
8015 2 : Ok(())
8016 2 : }
8017 :
8018 : #[cfg(feature = "testing")]
8019 : #[tokio::test]
8020 2 : async fn test_simple_bottom_most_compaction_deltas_1() -> anyhow::Result<()> {
8021 2 : test_simple_bottom_most_compaction_deltas_helper(
8022 2 : "test_simple_bottom_most_compaction_deltas_1",
8023 2 : false,
8024 2 : )
8025 242 : .await
8026 2 : }
8027 :
8028 : #[cfg(feature = "testing")]
8029 : #[tokio::test]
8030 2 : async fn test_simple_bottom_most_compaction_deltas_2() -> anyhow::Result<()> {
8031 2 : test_simple_bottom_most_compaction_deltas_helper(
8032 2 : "test_simple_bottom_most_compaction_deltas_2",
8033 2 : true,
8034 2 : )
8035 221 : .await
8036 2 : }
8037 :
8038 : #[cfg(feature = "testing")]
8039 4 : async fn test_simple_bottom_most_compaction_deltas_helper(
8040 4 : test_name: &'static str,
8041 4 : use_delta_bottom_layer: bool,
8042 4 : ) -> anyhow::Result<()> {
8043 4 : let harness = TenantHarness::create(test_name).await?;
8044 40 : let (tenant, ctx) = harness.load().await;
8045 :
8046 276 : fn get_key(id: u32) -> Key {
8047 276 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8048 276 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8049 276 : key.field6 = id;
8050 276 : key
8051 276 : }
8052 :
8053 : // We create
8054 : // - one bottom-most image layer,
8055 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
8056 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
8057 : // - a delta layer D3 above the horizon.
8058 : //
8059 : // | D3 |
8060 : // | D1 |
8061 : // -| |-- gc horizon -----------------
8062 : // | | | D2 |
8063 : // --------- img layer ------------------
8064 : //
8065 : // What we should expact from this compaction is:
8066 : // | D3 |
8067 : // | Part of D1 |
8068 : // --------- img layer with D1+D2 at GC horizon------------------
8069 :
8070 : // img layer at 0x10
8071 4 : let img_layer = (0..10)
8072 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8073 4 : .collect_vec();
8074 4 : // or, delta layer at 0x10 if `use_delta_bottom_layer` is true
8075 4 : let delta4 = (0..10)
8076 40 : .map(|id| {
8077 40 : (
8078 40 : get_key(id),
8079 40 : Lsn(0x08),
8080 40 : Value::WalRecord(NeonWalRecord::wal_init(format!("value {id}@0x10"))),
8081 40 : )
8082 40 : })
8083 4 : .collect_vec();
8084 4 :
8085 4 : let delta1 = vec![
8086 4 : (
8087 4 : get_key(1),
8088 4 : Lsn(0x20),
8089 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8090 4 : ),
8091 4 : (
8092 4 : get_key(2),
8093 4 : Lsn(0x30),
8094 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8095 4 : ),
8096 4 : (
8097 4 : get_key(3),
8098 4 : Lsn(0x28),
8099 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8100 4 : ),
8101 4 : (
8102 4 : get_key(3),
8103 4 : Lsn(0x30),
8104 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8105 4 : ),
8106 4 : (
8107 4 : get_key(3),
8108 4 : Lsn(0x40),
8109 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
8110 4 : ),
8111 4 : ];
8112 4 : let delta2 = vec![
8113 4 : (
8114 4 : get_key(5),
8115 4 : Lsn(0x20),
8116 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8117 4 : ),
8118 4 : (
8119 4 : get_key(6),
8120 4 : Lsn(0x20),
8121 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8122 4 : ),
8123 4 : ];
8124 4 : let delta3 = vec![
8125 4 : (
8126 4 : get_key(8),
8127 4 : Lsn(0x48),
8128 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8129 4 : ),
8130 4 : (
8131 4 : get_key(9),
8132 4 : Lsn(0x48),
8133 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8134 4 : ),
8135 4 : ];
8136 :
8137 4 : let tline = if use_delta_bottom_layer {
8138 2 : tenant
8139 2 : .create_test_timeline_with_layers(
8140 2 : TIMELINE_ID,
8141 2 : Lsn(0x08),
8142 2 : DEFAULT_PG_VERSION,
8143 2 : &ctx,
8144 2 : vec![
8145 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8146 2 : Lsn(0x08)..Lsn(0x10),
8147 2 : delta4,
8148 2 : ),
8149 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8150 2 : Lsn(0x20)..Lsn(0x48),
8151 2 : delta1,
8152 2 : ),
8153 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8154 2 : Lsn(0x20)..Lsn(0x48),
8155 2 : delta2,
8156 2 : ),
8157 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8158 2 : Lsn(0x48)..Lsn(0x50),
8159 2 : delta3,
8160 2 : ),
8161 2 : ], // delta layers
8162 2 : vec![], // image layers
8163 2 : Lsn(0x50),
8164 2 : )
8165 30 : .await?
8166 : } else {
8167 2 : tenant
8168 2 : .create_test_timeline_with_layers(
8169 2 : TIMELINE_ID,
8170 2 : Lsn(0x10),
8171 2 : DEFAULT_PG_VERSION,
8172 2 : &ctx,
8173 2 : vec![
8174 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8175 2 : Lsn(0x10)..Lsn(0x48),
8176 2 : delta1,
8177 2 : ),
8178 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8179 2 : Lsn(0x10)..Lsn(0x48),
8180 2 : delta2,
8181 2 : ),
8182 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8183 2 : Lsn(0x48)..Lsn(0x50),
8184 2 : delta3,
8185 2 : ),
8186 2 : ], // delta layers
8187 2 : vec![(Lsn(0x10), img_layer)], // image layers
8188 2 : Lsn(0x50),
8189 2 : )
8190 49 : .await?
8191 : };
8192 4 : {
8193 4 : // Update GC info
8194 4 : let mut guard = tline.gc_info.write().unwrap();
8195 4 : *guard = GcInfo {
8196 4 : retain_lsns: vec![],
8197 4 : cutoffs: GcCutoffs {
8198 4 : time: Lsn(0x30),
8199 4 : space: Lsn(0x30),
8200 4 : },
8201 4 : leases: Default::default(),
8202 4 : within_ancestor_pitr: false,
8203 4 : };
8204 4 : }
8205 4 :
8206 4 : let expected_result = [
8207 4 : Bytes::from_static(b"value 0@0x10"),
8208 4 : Bytes::from_static(b"value 1@0x10@0x20"),
8209 4 : Bytes::from_static(b"value 2@0x10@0x30"),
8210 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
8211 4 : Bytes::from_static(b"value 4@0x10"),
8212 4 : Bytes::from_static(b"value 5@0x10@0x20"),
8213 4 : Bytes::from_static(b"value 6@0x10@0x20"),
8214 4 : Bytes::from_static(b"value 7@0x10"),
8215 4 : Bytes::from_static(b"value 8@0x10@0x48"),
8216 4 : Bytes::from_static(b"value 9@0x10@0x48"),
8217 4 : ];
8218 4 :
8219 4 : let expected_result_at_gc_horizon = [
8220 4 : Bytes::from_static(b"value 0@0x10"),
8221 4 : Bytes::from_static(b"value 1@0x10@0x20"),
8222 4 : Bytes::from_static(b"value 2@0x10@0x30"),
8223 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
8224 4 : Bytes::from_static(b"value 4@0x10"),
8225 4 : Bytes::from_static(b"value 5@0x10@0x20"),
8226 4 : Bytes::from_static(b"value 6@0x10@0x20"),
8227 4 : Bytes::from_static(b"value 7@0x10"),
8228 4 : Bytes::from_static(b"value 8@0x10"),
8229 4 : Bytes::from_static(b"value 9@0x10"),
8230 4 : ];
8231 :
8232 44 : for idx in 0..10 {
8233 40 : assert_eq!(
8234 40 : tline
8235 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8236 60 : .await
8237 40 : .unwrap(),
8238 40 : &expected_result[idx]
8239 : );
8240 40 : assert_eq!(
8241 40 : tline
8242 40 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
8243 31 : .await
8244 40 : .unwrap(),
8245 40 : &expected_result_at_gc_horizon[idx]
8246 : );
8247 : }
8248 :
8249 4 : let cancel = CancellationToken::new();
8250 4 : tline
8251 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8252 107 : .await
8253 4 : .unwrap();
8254 :
8255 44 : for idx in 0..10 {
8256 40 : assert_eq!(
8257 40 : tline
8258 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8259 39 : .await
8260 40 : .unwrap(),
8261 40 : &expected_result[idx]
8262 : );
8263 40 : assert_eq!(
8264 40 : tline
8265 40 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
8266 21 : .await
8267 40 : .unwrap(),
8268 40 : &expected_result_at_gc_horizon[idx]
8269 : );
8270 : }
8271 :
8272 : // increase GC horizon and compact again
8273 4 : {
8274 4 : // Update GC info
8275 4 : let mut guard = tline.gc_info.write().unwrap();
8276 4 : guard.cutoffs.time = Lsn(0x40);
8277 4 : guard.cutoffs.space = Lsn(0x40);
8278 4 : }
8279 4 : tline
8280 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8281 86 : .await
8282 4 : .unwrap();
8283 4 :
8284 4 : Ok(())
8285 4 : }
8286 :
8287 : #[cfg(feature = "testing")]
8288 : #[tokio::test]
8289 2 : async fn test_generate_key_retention() -> anyhow::Result<()> {
8290 2 : let harness = TenantHarness::create("test_generate_key_retention").await?;
8291 20 : let (tenant, ctx) = harness.load().await;
8292 2 : let tline = tenant
8293 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
8294 6 : .await?;
8295 2 : tline.force_advance_lsn(Lsn(0x70));
8296 2 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
8297 2 : let history = vec![
8298 2 : (
8299 2 : key,
8300 2 : Lsn(0x10),
8301 2 : Value::WalRecord(NeonWalRecord::wal_init("0x10")),
8302 2 : ),
8303 2 : (
8304 2 : key,
8305 2 : Lsn(0x20),
8306 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8307 2 : ),
8308 2 : (
8309 2 : key,
8310 2 : Lsn(0x30),
8311 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8312 2 : ),
8313 2 : (
8314 2 : key,
8315 2 : Lsn(0x40),
8316 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8317 2 : ),
8318 2 : (
8319 2 : key,
8320 2 : Lsn(0x50),
8321 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
8322 2 : ),
8323 2 : (
8324 2 : key,
8325 2 : Lsn(0x60),
8326 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8327 2 : ),
8328 2 : (
8329 2 : key,
8330 2 : Lsn(0x70),
8331 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8332 2 : ),
8333 2 : (
8334 2 : key,
8335 2 : Lsn(0x80),
8336 2 : Value::Image(Bytes::copy_from_slice(
8337 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8338 2 : )),
8339 2 : ),
8340 2 : (
8341 2 : key,
8342 2 : Lsn(0x90),
8343 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8344 2 : ),
8345 2 : ];
8346 2 : let res = tline
8347 2 : .generate_key_retention(
8348 2 : key,
8349 2 : &history,
8350 2 : Lsn(0x60),
8351 2 : &[Lsn(0x20), Lsn(0x40), Lsn(0x50)],
8352 2 : 3,
8353 2 : None,
8354 2 : )
8355 2 : .await
8356 2 : .unwrap();
8357 2 : let expected_res = KeyHistoryRetention {
8358 2 : below_horizon: vec![
8359 2 : (
8360 2 : Lsn(0x20),
8361 2 : KeyLogAtLsn(vec![(
8362 2 : Lsn(0x20),
8363 2 : Value::Image(Bytes::from_static(b"0x10;0x20")),
8364 2 : )]),
8365 2 : ),
8366 2 : (
8367 2 : Lsn(0x40),
8368 2 : KeyLogAtLsn(vec![
8369 2 : (
8370 2 : Lsn(0x30),
8371 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8372 2 : ),
8373 2 : (
8374 2 : Lsn(0x40),
8375 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8376 2 : ),
8377 2 : ]),
8378 2 : ),
8379 2 : (
8380 2 : Lsn(0x50),
8381 2 : KeyLogAtLsn(vec![(
8382 2 : Lsn(0x50),
8383 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40;0x50")),
8384 2 : )]),
8385 2 : ),
8386 2 : (
8387 2 : Lsn(0x60),
8388 2 : KeyLogAtLsn(vec![(
8389 2 : Lsn(0x60),
8390 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8391 2 : )]),
8392 2 : ),
8393 2 : ],
8394 2 : above_horizon: KeyLogAtLsn(vec![
8395 2 : (
8396 2 : Lsn(0x70),
8397 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8398 2 : ),
8399 2 : (
8400 2 : Lsn(0x80),
8401 2 : Value::Image(Bytes::copy_from_slice(
8402 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8403 2 : )),
8404 2 : ),
8405 2 : (
8406 2 : Lsn(0x90),
8407 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8408 2 : ),
8409 2 : ]),
8410 2 : };
8411 2 : assert_eq!(res, expected_res);
8412 2 :
8413 2 : // We expect GC-compaction to run with the original GC. This would create a situation that
8414 2 : // the original GC algorithm removes some delta layers b/c there are full image coverage,
8415 2 : // therefore causing some keys to have an incomplete history below the lowest retain LSN.
8416 2 : // For example, we have
8417 2 : // ```plain
8418 2 : // init delta @ 0x10, image @ 0x20, delta @ 0x30 (gc_horizon), image @ 0x40.
8419 2 : // ```
8420 2 : // Now the GC horizon moves up, and we have
8421 2 : // ```plain
8422 2 : // init delta @ 0x10, image @ 0x20, delta @ 0x30, image @ 0x40 (gc_horizon)
8423 2 : // ```
8424 2 : // The original GC algorithm kicks in, and removes delta @ 0x10, image @ 0x20.
8425 2 : // We will end up with
8426 2 : // ```plain
8427 2 : // delta @ 0x30, image @ 0x40 (gc_horizon)
8428 2 : // ```
8429 2 : // Now we run the GC-compaction, and this key does not have a full history.
8430 2 : // We should be able to handle this partial history and drop everything before the
8431 2 : // gc_horizon image.
8432 2 :
8433 2 : let history = vec![
8434 2 : (
8435 2 : key,
8436 2 : Lsn(0x20),
8437 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8438 2 : ),
8439 2 : (
8440 2 : key,
8441 2 : Lsn(0x30),
8442 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8443 2 : ),
8444 2 : (
8445 2 : key,
8446 2 : Lsn(0x40),
8447 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
8448 2 : ),
8449 2 : (
8450 2 : key,
8451 2 : Lsn(0x50),
8452 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
8453 2 : ),
8454 2 : (
8455 2 : key,
8456 2 : Lsn(0x60),
8457 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8458 2 : ),
8459 2 : (
8460 2 : key,
8461 2 : Lsn(0x70),
8462 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8463 2 : ),
8464 2 : (
8465 2 : key,
8466 2 : Lsn(0x80),
8467 2 : Value::Image(Bytes::copy_from_slice(
8468 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8469 2 : )),
8470 2 : ),
8471 2 : (
8472 2 : key,
8473 2 : Lsn(0x90),
8474 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8475 2 : ),
8476 2 : ];
8477 2 : let res = tline
8478 2 : .generate_key_retention(key, &history, Lsn(0x60), &[Lsn(0x40), Lsn(0x50)], 3, None)
8479 2 : .await
8480 2 : .unwrap();
8481 2 : let expected_res = KeyHistoryRetention {
8482 2 : below_horizon: vec![
8483 2 : (
8484 2 : Lsn(0x40),
8485 2 : KeyLogAtLsn(vec![(
8486 2 : Lsn(0x40),
8487 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
8488 2 : )]),
8489 2 : ),
8490 2 : (
8491 2 : Lsn(0x50),
8492 2 : KeyLogAtLsn(vec![(
8493 2 : Lsn(0x50),
8494 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
8495 2 : )]),
8496 2 : ),
8497 2 : (
8498 2 : Lsn(0x60),
8499 2 : KeyLogAtLsn(vec![(
8500 2 : Lsn(0x60),
8501 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8502 2 : )]),
8503 2 : ),
8504 2 : ],
8505 2 : above_horizon: KeyLogAtLsn(vec![
8506 2 : (
8507 2 : Lsn(0x70),
8508 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8509 2 : ),
8510 2 : (
8511 2 : Lsn(0x80),
8512 2 : Value::Image(Bytes::copy_from_slice(
8513 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8514 2 : )),
8515 2 : ),
8516 2 : (
8517 2 : Lsn(0x90),
8518 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8519 2 : ),
8520 2 : ]),
8521 2 : };
8522 2 : assert_eq!(res, expected_res);
8523 2 :
8524 2 : // In case of branch compaction, the branch itself does not have the full history, and we need to provide
8525 2 : // the ancestor image in the test case.
8526 2 :
8527 2 : let history = vec![
8528 2 : (
8529 2 : key,
8530 2 : Lsn(0x20),
8531 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8532 2 : ),
8533 2 : (
8534 2 : key,
8535 2 : Lsn(0x30),
8536 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8537 2 : ),
8538 2 : (
8539 2 : key,
8540 2 : Lsn(0x40),
8541 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8542 2 : ),
8543 2 : (
8544 2 : key,
8545 2 : Lsn(0x70),
8546 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8547 2 : ),
8548 2 : ];
8549 2 : let res = tline
8550 2 : .generate_key_retention(
8551 2 : key,
8552 2 : &history,
8553 2 : Lsn(0x60),
8554 2 : &[],
8555 2 : 3,
8556 2 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
8557 2 : )
8558 2 : .await
8559 2 : .unwrap();
8560 2 : let expected_res = KeyHistoryRetention {
8561 2 : below_horizon: vec![(
8562 2 : Lsn(0x60),
8563 2 : KeyLogAtLsn(vec![(
8564 2 : Lsn(0x60),
8565 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")), // use the ancestor image to reconstruct the page
8566 2 : )]),
8567 2 : )],
8568 2 : above_horizon: KeyLogAtLsn(vec![(
8569 2 : Lsn(0x70),
8570 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8571 2 : )]),
8572 2 : };
8573 2 : assert_eq!(res, expected_res);
8574 2 :
8575 2 : let history = vec![
8576 2 : (
8577 2 : key,
8578 2 : Lsn(0x20),
8579 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8580 2 : ),
8581 2 : (
8582 2 : key,
8583 2 : Lsn(0x40),
8584 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8585 2 : ),
8586 2 : (
8587 2 : key,
8588 2 : Lsn(0x60),
8589 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8590 2 : ),
8591 2 : (
8592 2 : key,
8593 2 : Lsn(0x70),
8594 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8595 2 : ),
8596 2 : ];
8597 2 : let res = tline
8598 2 : .generate_key_retention(
8599 2 : key,
8600 2 : &history,
8601 2 : Lsn(0x60),
8602 2 : &[Lsn(0x30)],
8603 2 : 3,
8604 2 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
8605 2 : )
8606 2 : .await
8607 2 : .unwrap();
8608 2 : let expected_res = KeyHistoryRetention {
8609 2 : below_horizon: vec![
8610 2 : (
8611 2 : Lsn(0x30),
8612 2 : KeyLogAtLsn(vec![(
8613 2 : Lsn(0x20),
8614 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8615 2 : )]),
8616 2 : ),
8617 2 : (
8618 2 : Lsn(0x60),
8619 2 : KeyLogAtLsn(vec![(
8620 2 : Lsn(0x60),
8621 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x40;0x60")),
8622 2 : )]),
8623 2 : ),
8624 2 : ],
8625 2 : above_horizon: KeyLogAtLsn(vec![(
8626 2 : Lsn(0x70),
8627 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8628 2 : )]),
8629 2 : };
8630 2 : assert_eq!(res, expected_res);
8631 2 :
8632 2 : Ok(())
8633 2 : }
8634 :
8635 : #[cfg(feature = "testing")]
8636 : #[tokio::test]
8637 2 : async fn test_simple_bottom_most_compaction_with_retain_lsns() -> anyhow::Result<()> {
8638 2 : let harness =
8639 2 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns").await?;
8640 20 : let (tenant, ctx) = harness.load().await;
8641 2 :
8642 518 : fn get_key(id: u32) -> Key {
8643 518 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8644 518 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8645 518 : key.field6 = id;
8646 518 : key
8647 518 : }
8648 2 :
8649 2 : let img_layer = (0..10)
8650 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8651 2 : .collect_vec();
8652 2 :
8653 2 : let delta1 = vec![
8654 2 : (
8655 2 : get_key(1),
8656 2 : Lsn(0x20),
8657 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8658 2 : ),
8659 2 : (
8660 2 : get_key(2),
8661 2 : Lsn(0x30),
8662 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8663 2 : ),
8664 2 : (
8665 2 : get_key(3),
8666 2 : Lsn(0x28),
8667 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8668 2 : ),
8669 2 : (
8670 2 : get_key(3),
8671 2 : Lsn(0x30),
8672 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8673 2 : ),
8674 2 : (
8675 2 : get_key(3),
8676 2 : Lsn(0x40),
8677 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
8678 2 : ),
8679 2 : ];
8680 2 : let delta2 = vec![
8681 2 : (
8682 2 : get_key(5),
8683 2 : Lsn(0x20),
8684 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8685 2 : ),
8686 2 : (
8687 2 : get_key(6),
8688 2 : Lsn(0x20),
8689 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8690 2 : ),
8691 2 : ];
8692 2 : let delta3 = vec![
8693 2 : (
8694 2 : get_key(8),
8695 2 : Lsn(0x48),
8696 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8697 2 : ),
8698 2 : (
8699 2 : get_key(9),
8700 2 : Lsn(0x48),
8701 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8702 2 : ),
8703 2 : ];
8704 2 :
8705 2 : let tline = tenant
8706 2 : .create_test_timeline_with_layers(
8707 2 : TIMELINE_ID,
8708 2 : Lsn(0x10),
8709 2 : DEFAULT_PG_VERSION,
8710 2 : &ctx,
8711 2 : vec![
8712 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
8713 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
8714 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
8715 2 : ], // delta layers
8716 2 : vec![(Lsn(0x10), img_layer)], // image layers
8717 2 : Lsn(0x50),
8718 2 : )
8719 49 : .await?;
8720 2 : {
8721 2 : // Update GC info
8722 2 : let mut guard = tline.gc_info.write().unwrap();
8723 2 : *guard = GcInfo {
8724 2 : retain_lsns: vec![
8725 2 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
8726 2 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
8727 2 : ],
8728 2 : cutoffs: GcCutoffs {
8729 2 : time: Lsn(0x30),
8730 2 : space: Lsn(0x30),
8731 2 : },
8732 2 : leases: Default::default(),
8733 2 : within_ancestor_pitr: false,
8734 2 : };
8735 2 : }
8736 2 :
8737 2 : let expected_result = [
8738 2 : Bytes::from_static(b"value 0@0x10"),
8739 2 : Bytes::from_static(b"value 1@0x10@0x20"),
8740 2 : Bytes::from_static(b"value 2@0x10@0x30"),
8741 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
8742 2 : Bytes::from_static(b"value 4@0x10"),
8743 2 : Bytes::from_static(b"value 5@0x10@0x20"),
8744 2 : Bytes::from_static(b"value 6@0x10@0x20"),
8745 2 : Bytes::from_static(b"value 7@0x10"),
8746 2 : Bytes::from_static(b"value 8@0x10@0x48"),
8747 2 : Bytes::from_static(b"value 9@0x10@0x48"),
8748 2 : ];
8749 2 :
8750 2 : let expected_result_at_gc_horizon = [
8751 2 : Bytes::from_static(b"value 0@0x10"),
8752 2 : Bytes::from_static(b"value 1@0x10@0x20"),
8753 2 : Bytes::from_static(b"value 2@0x10@0x30"),
8754 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
8755 2 : Bytes::from_static(b"value 4@0x10"),
8756 2 : Bytes::from_static(b"value 5@0x10@0x20"),
8757 2 : Bytes::from_static(b"value 6@0x10@0x20"),
8758 2 : Bytes::from_static(b"value 7@0x10"),
8759 2 : Bytes::from_static(b"value 8@0x10"),
8760 2 : Bytes::from_static(b"value 9@0x10"),
8761 2 : ];
8762 2 :
8763 2 : let expected_result_at_lsn_20 = [
8764 2 : Bytes::from_static(b"value 0@0x10"),
8765 2 : Bytes::from_static(b"value 1@0x10@0x20"),
8766 2 : Bytes::from_static(b"value 2@0x10"),
8767 2 : Bytes::from_static(b"value 3@0x10"),
8768 2 : Bytes::from_static(b"value 4@0x10"),
8769 2 : Bytes::from_static(b"value 5@0x10@0x20"),
8770 2 : Bytes::from_static(b"value 6@0x10@0x20"),
8771 2 : Bytes::from_static(b"value 7@0x10"),
8772 2 : Bytes::from_static(b"value 8@0x10"),
8773 2 : Bytes::from_static(b"value 9@0x10"),
8774 2 : ];
8775 2 :
8776 2 : let expected_result_at_lsn_10 = [
8777 2 : Bytes::from_static(b"value 0@0x10"),
8778 2 : Bytes::from_static(b"value 1@0x10"),
8779 2 : Bytes::from_static(b"value 2@0x10"),
8780 2 : Bytes::from_static(b"value 3@0x10"),
8781 2 : Bytes::from_static(b"value 4@0x10"),
8782 2 : Bytes::from_static(b"value 5@0x10"),
8783 2 : Bytes::from_static(b"value 6@0x10"),
8784 2 : Bytes::from_static(b"value 7@0x10"),
8785 2 : Bytes::from_static(b"value 8@0x10"),
8786 2 : Bytes::from_static(b"value 9@0x10"),
8787 2 : ];
8788 2 :
8789 12 : let verify_result = || async {
8790 12 : let gc_horizon = {
8791 12 : let gc_info = tline.gc_info.read().unwrap();
8792 12 : gc_info.cutoffs.time
8793 2 : };
8794 132 : for idx in 0..10 {
8795 120 : assert_eq!(
8796 120 : tline
8797 120 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8798 121 : .await
8799 120 : .unwrap(),
8800 120 : &expected_result[idx]
8801 2 : );
8802 120 : assert_eq!(
8803 120 : tline
8804 120 : .get(get_key(idx as u32), gc_horizon, &ctx)
8805 93 : .await
8806 120 : .unwrap(),
8807 120 : &expected_result_at_gc_horizon[idx]
8808 2 : );
8809 120 : assert_eq!(
8810 120 : tline
8811 120 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
8812 86 : .await
8813 120 : .unwrap(),
8814 120 : &expected_result_at_lsn_20[idx]
8815 2 : );
8816 120 : assert_eq!(
8817 120 : tline
8818 120 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
8819 61 : .await
8820 120 : .unwrap(),
8821 120 : &expected_result_at_lsn_10[idx]
8822 2 : );
8823 2 : }
8824 24 : };
8825 2 :
8826 69 : verify_result().await;
8827 2 :
8828 2 : let cancel = CancellationToken::new();
8829 2 : let mut dryrun_flags = EnumSet::new();
8830 2 : dryrun_flags.insert(CompactFlags::DryRun);
8831 2 :
8832 2 : tline
8833 2 : .compact_with_gc(
8834 2 : &cancel,
8835 2 : CompactOptions {
8836 2 : flags: dryrun_flags,
8837 2 : compact_range: None,
8838 2 : },
8839 2 : &ctx,
8840 2 : )
8841 52 : .await
8842 2 : .unwrap();
8843 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
8844 2 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
8845 57 : verify_result().await;
8846 2 :
8847 2 : tline
8848 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8849 57 : .await
8850 2 : .unwrap();
8851 64 : verify_result().await;
8852 2 :
8853 2 : // compact again
8854 2 : tline
8855 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8856 42 : .await
8857 2 : .unwrap();
8858 57 : verify_result().await;
8859 2 :
8860 2 : // increase GC horizon and compact again
8861 2 : {
8862 2 : // Update GC info
8863 2 : let mut guard = tline.gc_info.write().unwrap();
8864 2 : guard.cutoffs.time = Lsn(0x38);
8865 2 : guard.cutoffs.space = Lsn(0x38);
8866 2 : }
8867 2 : tline
8868 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8869 42 : .await
8870 2 : .unwrap();
8871 57 : verify_result().await; // no wals between 0x30 and 0x38, so we should obtain the same result
8872 2 :
8873 2 : // not increasing the GC horizon and compact again
8874 2 : tline
8875 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8876 41 : .await
8877 2 : .unwrap();
8878 57 : verify_result().await;
8879 2 :
8880 2 : Ok(())
8881 2 : }
8882 :
8883 : #[cfg(feature = "testing")]
8884 : #[tokio::test]
8885 2 : async fn test_simple_bottom_most_compaction_with_retain_lsns_single_key() -> anyhow::Result<()>
8886 2 : {
8887 2 : let harness =
8888 2 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns_single_key")
8889 2 : .await?;
8890 20 : let (tenant, ctx) = harness.load().await;
8891 2 :
8892 352 : fn get_key(id: u32) -> Key {
8893 352 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8894 352 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8895 352 : key.field6 = id;
8896 352 : key
8897 352 : }
8898 2 :
8899 2 : let img_layer = (0..10)
8900 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8901 2 : .collect_vec();
8902 2 :
8903 2 : let delta1 = vec![
8904 2 : (
8905 2 : get_key(1),
8906 2 : Lsn(0x20),
8907 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8908 2 : ),
8909 2 : (
8910 2 : get_key(1),
8911 2 : Lsn(0x28),
8912 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8913 2 : ),
8914 2 : ];
8915 2 : let delta2 = vec![
8916 2 : (
8917 2 : get_key(1),
8918 2 : Lsn(0x30),
8919 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8920 2 : ),
8921 2 : (
8922 2 : get_key(1),
8923 2 : Lsn(0x38),
8924 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
8925 2 : ),
8926 2 : ];
8927 2 : let delta3 = vec![
8928 2 : (
8929 2 : get_key(8),
8930 2 : Lsn(0x48),
8931 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8932 2 : ),
8933 2 : (
8934 2 : get_key(9),
8935 2 : Lsn(0x48),
8936 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8937 2 : ),
8938 2 : ];
8939 2 :
8940 2 : let tline = tenant
8941 2 : .create_test_timeline_with_layers(
8942 2 : TIMELINE_ID,
8943 2 : Lsn(0x10),
8944 2 : DEFAULT_PG_VERSION,
8945 2 : &ctx,
8946 2 : vec![
8947 2 : // delta1 and delta 2 only contain a single key but multiple updates
8948 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x30), delta1),
8949 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
8950 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x50), delta3),
8951 2 : ], // delta layers
8952 2 : vec![(Lsn(0x10), img_layer)], // image layers
8953 2 : Lsn(0x50),
8954 2 : )
8955 49 : .await?;
8956 2 : {
8957 2 : // Update GC info
8958 2 : let mut guard = tline.gc_info.write().unwrap();
8959 2 : *guard = GcInfo {
8960 2 : retain_lsns: vec![
8961 2 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
8962 2 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
8963 2 : ],
8964 2 : cutoffs: GcCutoffs {
8965 2 : time: Lsn(0x30),
8966 2 : space: Lsn(0x30),
8967 2 : },
8968 2 : leases: Default::default(),
8969 2 : within_ancestor_pitr: false,
8970 2 : };
8971 2 : }
8972 2 :
8973 2 : let expected_result = [
8974 2 : Bytes::from_static(b"value 0@0x10"),
8975 2 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
8976 2 : Bytes::from_static(b"value 2@0x10"),
8977 2 : Bytes::from_static(b"value 3@0x10"),
8978 2 : Bytes::from_static(b"value 4@0x10"),
8979 2 : Bytes::from_static(b"value 5@0x10"),
8980 2 : Bytes::from_static(b"value 6@0x10"),
8981 2 : Bytes::from_static(b"value 7@0x10"),
8982 2 : Bytes::from_static(b"value 8@0x10@0x48"),
8983 2 : Bytes::from_static(b"value 9@0x10@0x48"),
8984 2 : ];
8985 2 :
8986 2 : let expected_result_at_gc_horizon = [
8987 2 : Bytes::from_static(b"value 0@0x10"),
8988 2 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
8989 2 : Bytes::from_static(b"value 2@0x10"),
8990 2 : Bytes::from_static(b"value 3@0x10"),
8991 2 : Bytes::from_static(b"value 4@0x10"),
8992 2 : Bytes::from_static(b"value 5@0x10"),
8993 2 : Bytes::from_static(b"value 6@0x10"),
8994 2 : Bytes::from_static(b"value 7@0x10"),
8995 2 : Bytes::from_static(b"value 8@0x10"),
8996 2 : Bytes::from_static(b"value 9@0x10"),
8997 2 : ];
8998 2 :
8999 2 : let expected_result_at_lsn_20 = [
9000 2 : Bytes::from_static(b"value 0@0x10"),
9001 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9002 2 : Bytes::from_static(b"value 2@0x10"),
9003 2 : Bytes::from_static(b"value 3@0x10"),
9004 2 : Bytes::from_static(b"value 4@0x10"),
9005 2 : Bytes::from_static(b"value 5@0x10"),
9006 2 : Bytes::from_static(b"value 6@0x10"),
9007 2 : Bytes::from_static(b"value 7@0x10"),
9008 2 : Bytes::from_static(b"value 8@0x10"),
9009 2 : Bytes::from_static(b"value 9@0x10"),
9010 2 : ];
9011 2 :
9012 2 : let expected_result_at_lsn_10 = [
9013 2 : Bytes::from_static(b"value 0@0x10"),
9014 2 : Bytes::from_static(b"value 1@0x10"),
9015 2 : Bytes::from_static(b"value 2@0x10"),
9016 2 : Bytes::from_static(b"value 3@0x10"),
9017 2 : Bytes::from_static(b"value 4@0x10"),
9018 2 : Bytes::from_static(b"value 5@0x10"),
9019 2 : Bytes::from_static(b"value 6@0x10"),
9020 2 : Bytes::from_static(b"value 7@0x10"),
9021 2 : Bytes::from_static(b"value 8@0x10"),
9022 2 : Bytes::from_static(b"value 9@0x10"),
9023 2 : ];
9024 2 :
9025 8 : let verify_result = || async {
9026 8 : let gc_horizon = {
9027 8 : let gc_info = tline.gc_info.read().unwrap();
9028 8 : gc_info.cutoffs.time
9029 2 : };
9030 88 : for idx in 0..10 {
9031 80 : assert_eq!(
9032 80 : tline
9033 80 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9034 76 : .await
9035 80 : .unwrap(),
9036 80 : &expected_result[idx]
9037 2 : );
9038 80 : assert_eq!(
9039 80 : tline
9040 80 : .get(get_key(idx as u32), gc_horizon, &ctx)
9041 46 : .await
9042 80 : .unwrap(),
9043 80 : &expected_result_at_gc_horizon[idx]
9044 2 : );
9045 80 : assert_eq!(
9046 80 : tline
9047 80 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
9048 48 : .await
9049 80 : .unwrap(),
9050 80 : &expected_result_at_lsn_20[idx]
9051 2 : );
9052 80 : assert_eq!(
9053 80 : tline
9054 80 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
9055 41 : .await
9056 80 : .unwrap(),
9057 80 : &expected_result_at_lsn_10[idx]
9058 2 : );
9059 2 : }
9060 16 : };
9061 2 :
9062 61 : verify_result().await;
9063 2 :
9064 2 : let cancel = CancellationToken::new();
9065 2 : let mut dryrun_flags = EnumSet::new();
9066 2 : dryrun_flags.insert(CompactFlags::DryRun);
9067 2 :
9068 2 : tline
9069 2 : .compact_with_gc(
9070 2 : &cancel,
9071 2 : CompactOptions {
9072 2 : flags: dryrun_flags,
9073 2 : compact_range: None,
9074 2 : },
9075 2 : &ctx,
9076 2 : )
9077 49 : .await
9078 2 : .unwrap();
9079 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
9080 2 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
9081 49 : verify_result().await;
9082 2 :
9083 2 : tline
9084 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9085 55 : .await
9086 2 : .unwrap();
9087 54 : verify_result().await;
9088 2 :
9089 2 : // compact again
9090 2 : tline
9091 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9092 42 : .await
9093 2 : .unwrap();
9094 47 : verify_result().await;
9095 2 :
9096 2 : Ok(())
9097 2 : }
9098 :
9099 : #[cfg(feature = "testing")]
9100 : #[tokio::test]
9101 2 : async fn test_simple_bottom_most_compaction_on_branch() -> anyhow::Result<()> {
9102 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_on_branch").await?;
9103 20 : let (tenant, ctx) = harness.load().await;
9104 2 :
9105 126 : fn get_key(id: u32) -> Key {
9106 126 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
9107 126 : key.field6 = id;
9108 126 : key
9109 126 : }
9110 2 :
9111 2 : let img_layer = (0..10)
9112 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9113 2 : .collect_vec();
9114 2 :
9115 2 : let delta1 = vec![
9116 2 : (
9117 2 : get_key(1),
9118 2 : Lsn(0x20),
9119 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9120 2 : ),
9121 2 : (
9122 2 : get_key(2),
9123 2 : Lsn(0x30),
9124 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9125 2 : ),
9126 2 : (
9127 2 : get_key(3),
9128 2 : Lsn(0x28),
9129 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9130 2 : ),
9131 2 : (
9132 2 : get_key(3),
9133 2 : Lsn(0x30),
9134 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9135 2 : ),
9136 2 : (
9137 2 : get_key(3),
9138 2 : Lsn(0x40),
9139 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9140 2 : ),
9141 2 : ];
9142 2 : let delta2 = vec![
9143 2 : (
9144 2 : get_key(5),
9145 2 : Lsn(0x20),
9146 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9147 2 : ),
9148 2 : (
9149 2 : get_key(6),
9150 2 : Lsn(0x20),
9151 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9152 2 : ),
9153 2 : ];
9154 2 : let delta3 = vec![
9155 2 : (
9156 2 : get_key(8),
9157 2 : Lsn(0x48),
9158 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9159 2 : ),
9160 2 : (
9161 2 : get_key(9),
9162 2 : Lsn(0x48),
9163 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9164 2 : ),
9165 2 : ];
9166 2 :
9167 2 : let parent_tline = tenant
9168 2 : .create_test_timeline_with_layers(
9169 2 : TIMELINE_ID,
9170 2 : Lsn(0x10),
9171 2 : DEFAULT_PG_VERSION,
9172 2 : &ctx,
9173 2 : vec![], // delta layers
9174 2 : vec![(Lsn(0x18), img_layer)], // image layers
9175 2 : Lsn(0x18),
9176 2 : )
9177 31 : .await?;
9178 2 :
9179 2 : parent_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
9180 2 :
9181 2 : let branch_tline = tenant
9182 2 : .branch_timeline_test_with_layers(
9183 2 : &parent_tline,
9184 2 : NEW_TIMELINE_ID,
9185 2 : Some(Lsn(0x18)),
9186 2 : &ctx,
9187 2 : vec![
9188 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
9189 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
9190 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
9191 2 : ], // delta layers
9192 2 : vec![], // image layers
9193 2 : Lsn(0x50),
9194 2 : )
9195 19 : .await?;
9196 2 :
9197 2 : branch_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
9198 2 :
9199 2 : {
9200 2 : // Update GC info
9201 2 : let mut guard = parent_tline.gc_info.write().unwrap();
9202 2 : *guard = GcInfo {
9203 2 : retain_lsns: vec![(Lsn(0x18), branch_tline.timeline_id, MaybeOffloaded::No)],
9204 2 : cutoffs: GcCutoffs {
9205 2 : time: Lsn(0x10),
9206 2 : space: Lsn(0x10),
9207 2 : },
9208 2 : leases: Default::default(),
9209 2 : within_ancestor_pitr: false,
9210 2 : };
9211 2 : }
9212 2 :
9213 2 : {
9214 2 : // Update GC info
9215 2 : let mut guard = branch_tline.gc_info.write().unwrap();
9216 2 : *guard = GcInfo {
9217 2 : retain_lsns: vec![(Lsn(0x40), branch_tline.timeline_id, MaybeOffloaded::No)],
9218 2 : cutoffs: GcCutoffs {
9219 2 : time: Lsn(0x50),
9220 2 : space: Lsn(0x50),
9221 2 : },
9222 2 : leases: Default::default(),
9223 2 : within_ancestor_pitr: false,
9224 2 : };
9225 2 : }
9226 2 :
9227 2 : let expected_result_at_gc_horizon = [
9228 2 : Bytes::from_static(b"value 0@0x10"),
9229 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9230 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9231 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9232 2 : Bytes::from_static(b"value 4@0x10"),
9233 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9234 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9235 2 : Bytes::from_static(b"value 7@0x10"),
9236 2 : Bytes::from_static(b"value 8@0x10@0x48"),
9237 2 : Bytes::from_static(b"value 9@0x10@0x48"),
9238 2 : ];
9239 2 :
9240 2 : let expected_result_at_lsn_40 = [
9241 2 : Bytes::from_static(b"value 0@0x10"),
9242 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9243 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9244 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9245 2 : Bytes::from_static(b"value 4@0x10"),
9246 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9247 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9248 2 : Bytes::from_static(b"value 7@0x10"),
9249 2 : Bytes::from_static(b"value 8@0x10"),
9250 2 : Bytes::from_static(b"value 9@0x10"),
9251 2 : ];
9252 2 :
9253 4 : let verify_result = || async {
9254 44 : for idx in 0..10 {
9255 40 : assert_eq!(
9256 40 : branch_tline
9257 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9258 52 : .await
9259 40 : .unwrap(),
9260 40 : &expected_result_at_gc_horizon[idx]
9261 2 : );
9262 40 : assert_eq!(
9263 40 : branch_tline
9264 40 : .get(get_key(idx as u32), Lsn(0x40), &ctx)
9265 30 : .await
9266 40 : .unwrap(),
9267 40 : &expected_result_at_lsn_40[idx]
9268 2 : );
9269 2 : }
9270 8 : };
9271 2 :
9272 45 : verify_result().await;
9273 2 :
9274 2 : let cancel = CancellationToken::new();
9275 2 : branch_tline
9276 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9277 22 : .await
9278 2 : .unwrap();
9279 2 :
9280 37 : verify_result().await;
9281 2 :
9282 2 : Ok(())
9283 2 : }
9284 :
9285 : // Regression test for https://github.com/neondatabase/neon/issues/9012
9286 : // Create an image arrangement where we have to read at different LSN ranges
9287 : // from a delta layer. This is achieved by overlapping an image layer on top of
9288 : // a delta layer. Like so:
9289 : //
9290 : // A B
9291 : // +----------------+ -> delta_layer
9292 : // | | ^ lsn
9293 : // | =========|-> nested_image_layer |
9294 : // | C | |
9295 : // +----------------+ |
9296 : // ======== -> baseline_image_layer +-------> key
9297 : //
9298 : //
9299 : // When querying the key range [A, B) we need to read at different LSN ranges
9300 : // for [A, C) and [C, B). This test checks that the described edge case is handled correctly.
9301 : #[cfg(feature = "testing")]
9302 : #[tokio::test]
9303 2 : async fn test_vectored_read_with_nested_image_layer() -> anyhow::Result<()> {
9304 2 : let harness = TenantHarness::create("test_vectored_read_with_nested_image_layer").await?;
9305 20 : let (tenant, ctx) = harness.load().await;
9306 2 :
9307 2 : let will_init_keys = [2, 6];
9308 44 : fn get_key(id: u32) -> Key {
9309 44 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
9310 44 : key.field6 = id;
9311 44 : key
9312 44 : }
9313 2 :
9314 2 : let mut expected_key_values = HashMap::new();
9315 2 :
9316 2 : let baseline_image_layer_lsn = Lsn(0x10);
9317 2 : let mut baseline_img_layer = Vec::new();
9318 12 : for i in 0..5 {
9319 10 : let key = get_key(i);
9320 10 : let value = format!("value {i}@{baseline_image_layer_lsn}");
9321 10 :
9322 10 : let removed = expected_key_values.insert(key, value.clone());
9323 10 : assert!(removed.is_none());
9324 2 :
9325 10 : baseline_img_layer.push((key, Bytes::from(value)));
9326 2 : }
9327 2 :
9328 2 : let nested_image_layer_lsn = Lsn(0x50);
9329 2 : let mut nested_img_layer = Vec::new();
9330 12 : for i in 5..10 {
9331 10 : let key = get_key(i);
9332 10 : let value = format!("value {i}@{nested_image_layer_lsn}");
9333 10 :
9334 10 : let removed = expected_key_values.insert(key, value.clone());
9335 10 : assert!(removed.is_none());
9336 2 :
9337 10 : nested_img_layer.push((key, Bytes::from(value)));
9338 2 : }
9339 2 :
9340 2 : let mut delta_layer_spec = Vec::default();
9341 2 : let delta_layer_start_lsn = Lsn(0x20);
9342 2 : let mut delta_layer_end_lsn = delta_layer_start_lsn;
9343 2 :
9344 22 : for i in 0..10 {
9345 20 : let key = get_key(i);
9346 20 : let key_in_nested = nested_img_layer
9347 20 : .iter()
9348 80 : .any(|(key_with_img, _)| *key_with_img == key);
9349 20 : let lsn = {
9350 20 : if key_in_nested {
9351 10 : Lsn(nested_image_layer_lsn.0 + 0x10)
9352 2 : } else {
9353 10 : delta_layer_start_lsn
9354 2 : }
9355 2 : };
9356 2 :
9357 20 : let will_init = will_init_keys.contains(&i);
9358 20 : if will_init {
9359 4 : delta_layer_spec.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init(""))));
9360 4 :
9361 4 : expected_key_values.insert(key, "".to_string());
9362 16 : } else {
9363 16 : let delta = format!("@{lsn}");
9364 16 : delta_layer_spec.push((
9365 16 : key,
9366 16 : lsn,
9367 16 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
9368 16 : ));
9369 16 :
9370 16 : expected_key_values
9371 16 : .get_mut(&key)
9372 16 : .expect("An image exists for each key")
9373 16 : .push_str(delta.as_str());
9374 16 : }
9375 20 : delta_layer_end_lsn = std::cmp::max(delta_layer_start_lsn, lsn);
9376 2 : }
9377 2 :
9378 2 : delta_layer_end_lsn = Lsn(delta_layer_end_lsn.0 + 1);
9379 2 :
9380 2 : assert!(
9381 2 : nested_image_layer_lsn > delta_layer_start_lsn
9382 2 : && nested_image_layer_lsn < delta_layer_end_lsn
9383 2 : );
9384 2 :
9385 2 : let tline = tenant
9386 2 : .create_test_timeline_with_layers(
9387 2 : TIMELINE_ID,
9388 2 : baseline_image_layer_lsn,
9389 2 : DEFAULT_PG_VERSION,
9390 2 : &ctx,
9391 2 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
9392 2 : delta_layer_start_lsn..delta_layer_end_lsn,
9393 2 : delta_layer_spec,
9394 2 : )], // delta layers
9395 2 : vec![
9396 2 : (baseline_image_layer_lsn, baseline_img_layer),
9397 2 : (nested_image_layer_lsn, nested_img_layer),
9398 2 : ], // image layers
9399 2 : delta_layer_end_lsn,
9400 2 : )
9401 42 : .await?;
9402 2 :
9403 2 : let keyspace = KeySpace::single(get_key(0)..get_key(10));
9404 2 : let results = tline
9405 2 : .get_vectored(keyspace, delta_layer_end_lsn, &ctx)
9406 13 : .await
9407 2 : .expect("No vectored errors");
9408 22 : for (key, res) in results {
9409 20 : let value = res.expect("No key errors");
9410 20 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
9411 20 : assert_eq!(value, Bytes::from(expected_value));
9412 2 : }
9413 2 :
9414 2 : Ok(())
9415 2 : }
9416 :
9417 142 : fn sort_layer_key(k1: &PersistentLayerKey, k2: &PersistentLayerKey) -> std::cmp::Ordering {
9418 142 : (
9419 142 : k1.is_delta,
9420 142 : k1.key_range.start,
9421 142 : k1.key_range.end,
9422 142 : k1.lsn_range.start,
9423 142 : k1.lsn_range.end,
9424 142 : )
9425 142 : .cmp(&(
9426 142 : k2.is_delta,
9427 142 : k2.key_range.start,
9428 142 : k2.key_range.end,
9429 142 : k2.lsn_range.start,
9430 142 : k2.lsn_range.end,
9431 142 : ))
9432 142 : }
9433 :
9434 12 : async fn inspect_and_sort(
9435 12 : tline: &Arc<Timeline>,
9436 12 : filter: Option<std::ops::Range<Key>>,
9437 12 : ) -> Vec<PersistentLayerKey> {
9438 12 : let mut all_layers = tline.inspect_historic_layers().await.unwrap();
9439 12 : if let Some(filter) = filter {
9440 64 : all_layers.retain(|layer| overlaps_with(&layer.key_range, &filter));
9441 10 : }
9442 12 : all_layers.sort_by(sort_layer_key);
9443 12 : all_layers
9444 12 : }
9445 :
9446 : #[cfg(feature = "testing")]
9447 10 : fn check_layer_map_key_eq(
9448 10 : mut left: Vec<PersistentLayerKey>,
9449 10 : mut right: Vec<PersistentLayerKey>,
9450 10 : ) {
9451 10 : left.sort_by(sort_layer_key);
9452 10 : right.sort_by(sort_layer_key);
9453 10 : if left != right {
9454 0 : eprintln!("---LEFT---");
9455 0 : for left in left.iter() {
9456 0 : eprintln!("{}", left);
9457 0 : }
9458 0 : eprintln!("---RIGHT---");
9459 0 : for right in right.iter() {
9460 0 : eprintln!("{}", right);
9461 0 : }
9462 0 : assert_eq!(left, right);
9463 10 : }
9464 10 : }
9465 :
9466 : #[cfg(feature = "testing")]
9467 : #[tokio::test]
9468 2 : async fn test_simple_partial_bottom_most_compaction() -> anyhow::Result<()> {
9469 2 : let harness = TenantHarness::create("test_simple_partial_bottom_most_compaction").await?;
9470 20 : let (tenant, ctx) = harness.load().await;
9471 2 :
9472 182 : fn get_key(id: u32) -> Key {
9473 182 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9474 182 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9475 182 : key.field6 = id;
9476 182 : key
9477 182 : }
9478 2 :
9479 2 : // img layer at 0x10
9480 2 : let img_layer = (0..10)
9481 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9482 2 : .collect_vec();
9483 2 :
9484 2 : let delta1 = vec![
9485 2 : (
9486 2 : get_key(1),
9487 2 : Lsn(0x20),
9488 2 : Value::Image(Bytes::from("value 1@0x20")),
9489 2 : ),
9490 2 : (
9491 2 : get_key(2),
9492 2 : Lsn(0x30),
9493 2 : Value::Image(Bytes::from("value 2@0x30")),
9494 2 : ),
9495 2 : (
9496 2 : get_key(3),
9497 2 : Lsn(0x40),
9498 2 : Value::Image(Bytes::from("value 3@0x40")),
9499 2 : ),
9500 2 : ];
9501 2 : let delta2 = vec![
9502 2 : (
9503 2 : get_key(5),
9504 2 : Lsn(0x20),
9505 2 : Value::Image(Bytes::from("value 5@0x20")),
9506 2 : ),
9507 2 : (
9508 2 : get_key(6),
9509 2 : Lsn(0x20),
9510 2 : Value::Image(Bytes::from("value 6@0x20")),
9511 2 : ),
9512 2 : ];
9513 2 : let delta3 = vec![
9514 2 : (
9515 2 : get_key(8),
9516 2 : Lsn(0x48),
9517 2 : Value::Image(Bytes::from("value 8@0x48")),
9518 2 : ),
9519 2 : (
9520 2 : get_key(9),
9521 2 : Lsn(0x48),
9522 2 : Value::Image(Bytes::from("value 9@0x48")),
9523 2 : ),
9524 2 : ];
9525 2 :
9526 2 : let tline = tenant
9527 2 : .create_test_timeline_with_layers(
9528 2 : TIMELINE_ID,
9529 2 : Lsn(0x10),
9530 2 : DEFAULT_PG_VERSION,
9531 2 : &ctx,
9532 2 : vec![
9533 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
9534 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
9535 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
9536 2 : ], // delta layers
9537 2 : vec![(Lsn(0x10), img_layer)], // image layers
9538 2 : Lsn(0x50),
9539 2 : )
9540 49 : .await?;
9541 2 :
9542 2 : {
9543 2 : // Update GC info
9544 2 : let mut guard = tline.gc_info.write().unwrap();
9545 2 : *guard = GcInfo {
9546 2 : retain_lsns: vec![(Lsn(0x20), tline.timeline_id, MaybeOffloaded::No)],
9547 2 : cutoffs: GcCutoffs {
9548 2 : time: Lsn(0x30),
9549 2 : space: Lsn(0x30),
9550 2 : },
9551 2 : leases: Default::default(),
9552 2 : within_ancestor_pitr: false,
9553 2 : };
9554 2 : }
9555 2 :
9556 2 : let cancel = CancellationToken::new();
9557 2 :
9558 2 : // Do a partial compaction on key range 0..2
9559 2 : tline
9560 2 : .partial_compact_with_gc(get_key(0)..get_key(2), &cancel, EnumSet::new(), &ctx)
9561 27 : .await
9562 2 : .unwrap();
9563 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
9564 2 : check_layer_map_key_eq(
9565 2 : all_layers,
9566 2 : vec![
9567 2 : // newly-generated image layer for the partial compaction range 0-2
9568 2 : PersistentLayerKey {
9569 2 : key_range: get_key(0)..get_key(2),
9570 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9571 2 : is_delta: false,
9572 2 : },
9573 2 : PersistentLayerKey {
9574 2 : key_range: get_key(0)..get_key(10),
9575 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
9576 2 : is_delta: false,
9577 2 : },
9578 2 : // delta1 is split and the second part is rewritten
9579 2 : PersistentLayerKey {
9580 2 : key_range: get_key(2)..get_key(4),
9581 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9582 2 : is_delta: true,
9583 2 : },
9584 2 : PersistentLayerKey {
9585 2 : key_range: get_key(5)..get_key(7),
9586 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9587 2 : is_delta: true,
9588 2 : },
9589 2 : PersistentLayerKey {
9590 2 : key_range: get_key(8)..get_key(10),
9591 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
9592 2 : is_delta: true,
9593 2 : },
9594 2 : ],
9595 2 : );
9596 2 :
9597 2 : // Do a partial compaction on key range 2..4
9598 2 : tline
9599 2 : .partial_compact_with_gc(get_key(2)..get_key(4), &cancel, EnumSet::new(), &ctx)
9600 19 : .await
9601 2 : .unwrap();
9602 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
9603 2 : check_layer_map_key_eq(
9604 2 : all_layers,
9605 2 : vec![
9606 2 : PersistentLayerKey {
9607 2 : key_range: get_key(0)..get_key(2),
9608 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9609 2 : is_delta: false,
9610 2 : },
9611 2 : PersistentLayerKey {
9612 2 : key_range: get_key(0)..get_key(10),
9613 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
9614 2 : is_delta: false,
9615 2 : },
9616 2 : // image layer generated for the compaction range 2-4
9617 2 : PersistentLayerKey {
9618 2 : key_range: get_key(2)..get_key(4),
9619 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9620 2 : is_delta: false,
9621 2 : },
9622 2 : // we have key2/key3 above the retain_lsn, so we still need this delta layer
9623 2 : PersistentLayerKey {
9624 2 : key_range: get_key(2)..get_key(4),
9625 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9626 2 : is_delta: true,
9627 2 : },
9628 2 : PersistentLayerKey {
9629 2 : key_range: get_key(5)..get_key(7),
9630 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9631 2 : is_delta: true,
9632 2 : },
9633 2 : PersistentLayerKey {
9634 2 : key_range: get_key(8)..get_key(10),
9635 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
9636 2 : is_delta: true,
9637 2 : },
9638 2 : ],
9639 2 : );
9640 2 :
9641 2 : // Do a partial compaction on key range 4..9
9642 2 : tline
9643 2 : .partial_compact_with_gc(get_key(4)..get_key(9), &cancel, EnumSet::new(), &ctx)
9644 23 : .await
9645 2 : .unwrap();
9646 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
9647 2 : check_layer_map_key_eq(
9648 2 : all_layers,
9649 2 : vec![
9650 2 : PersistentLayerKey {
9651 2 : key_range: get_key(0)..get_key(2),
9652 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9653 2 : is_delta: false,
9654 2 : },
9655 2 : PersistentLayerKey {
9656 2 : key_range: get_key(0)..get_key(10),
9657 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
9658 2 : is_delta: false,
9659 2 : },
9660 2 : PersistentLayerKey {
9661 2 : key_range: get_key(2)..get_key(4),
9662 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9663 2 : is_delta: false,
9664 2 : },
9665 2 : PersistentLayerKey {
9666 2 : key_range: get_key(2)..get_key(4),
9667 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9668 2 : is_delta: true,
9669 2 : },
9670 2 : // image layer generated for this compaction range
9671 2 : PersistentLayerKey {
9672 2 : key_range: get_key(4)..get_key(9),
9673 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9674 2 : is_delta: false,
9675 2 : },
9676 2 : PersistentLayerKey {
9677 2 : key_range: get_key(8)..get_key(10),
9678 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
9679 2 : is_delta: true,
9680 2 : },
9681 2 : ],
9682 2 : );
9683 2 :
9684 2 : // Do a partial compaction on key range 9..10
9685 2 : tline
9686 2 : .partial_compact_with_gc(get_key(9)..get_key(10), &cancel, EnumSet::new(), &ctx)
9687 9 : .await
9688 2 : .unwrap();
9689 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
9690 2 : check_layer_map_key_eq(
9691 2 : all_layers,
9692 2 : vec![
9693 2 : PersistentLayerKey {
9694 2 : key_range: get_key(0)..get_key(2),
9695 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9696 2 : is_delta: false,
9697 2 : },
9698 2 : PersistentLayerKey {
9699 2 : key_range: get_key(0)..get_key(10),
9700 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
9701 2 : is_delta: false,
9702 2 : },
9703 2 : PersistentLayerKey {
9704 2 : key_range: get_key(2)..get_key(4),
9705 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9706 2 : is_delta: false,
9707 2 : },
9708 2 : PersistentLayerKey {
9709 2 : key_range: get_key(2)..get_key(4),
9710 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9711 2 : is_delta: true,
9712 2 : },
9713 2 : PersistentLayerKey {
9714 2 : key_range: get_key(4)..get_key(9),
9715 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9716 2 : is_delta: false,
9717 2 : },
9718 2 : // image layer generated for the compaction range
9719 2 : PersistentLayerKey {
9720 2 : key_range: get_key(9)..get_key(10),
9721 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9722 2 : is_delta: false,
9723 2 : },
9724 2 : PersistentLayerKey {
9725 2 : key_range: get_key(8)..get_key(10),
9726 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
9727 2 : is_delta: true,
9728 2 : },
9729 2 : ],
9730 2 : );
9731 2 :
9732 2 : // Do a partial compaction on key range 0..10, all image layers below LSN 20 can be replaced with new ones.
9733 2 : tline
9734 2 : .partial_compact_with_gc(get_key(0)..get_key(10), &cancel, EnumSet::new(), &ctx)
9735 55 : .await
9736 2 : .unwrap();
9737 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
9738 2 : check_layer_map_key_eq(
9739 2 : all_layers,
9740 2 : vec![
9741 2 : // aha, we removed all unnecessary image/delta layers and got a very clean layer map!
9742 2 : PersistentLayerKey {
9743 2 : key_range: get_key(0)..get_key(10),
9744 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9745 2 : is_delta: false,
9746 2 : },
9747 2 : PersistentLayerKey {
9748 2 : key_range: get_key(2)..get_key(4),
9749 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9750 2 : is_delta: true,
9751 2 : },
9752 2 : PersistentLayerKey {
9753 2 : key_range: get_key(8)..get_key(10),
9754 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
9755 2 : is_delta: true,
9756 2 : },
9757 2 : ],
9758 2 : );
9759 2 :
9760 2 : Ok(())
9761 2 : }
9762 :
9763 : #[cfg(feature = "testing")]
9764 : #[tokio::test]
9765 2 : async fn test_timeline_offload_retain_lsn() -> anyhow::Result<()> {
9766 2 : let harness = TenantHarness::create("test_timeline_offload_retain_lsn")
9767 2 : .await
9768 2 : .unwrap();
9769 20 : let (tenant, ctx) = harness.load().await;
9770 2 : let tline_parent = tenant
9771 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
9772 6 : .await
9773 2 : .unwrap();
9774 2 : let tline_child = tenant
9775 2 : .branch_timeline_test(&tline_parent, NEW_TIMELINE_ID, Some(Lsn(0x20)), &ctx)
9776 2 : .await
9777 2 : .unwrap();
9778 2 : {
9779 2 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
9780 2 : assert_eq!(
9781 2 : gc_info_parent.retain_lsns,
9782 2 : vec![(Lsn(0x20), tline_child.timeline_id, MaybeOffloaded::No)]
9783 2 : );
9784 2 : }
9785 2 : // We have to directly call the remote_client instead of using the archive function to avoid constructing broker client...
9786 2 : tline_child
9787 2 : .remote_client
9788 2 : .schedule_index_upload_for_timeline_archival_state(TimelineArchivalState::Archived)
9789 2 : .unwrap();
9790 2 : tline_child.remote_client.wait_completion().await.unwrap();
9791 2 : offload_timeline(&tenant, &tline_child)
9792 2 : .instrument(tracing::info_span!(parent: None, "offload_test", tenant_id=%"test", shard_id=%"test", timeline_id=%"test"))
9793 16 : .await.unwrap();
9794 2 : let child_timeline_id = tline_child.timeline_id;
9795 2 : Arc::try_unwrap(tline_child).unwrap();
9796 2 :
9797 2 : {
9798 2 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
9799 2 : assert_eq!(
9800 2 : gc_info_parent.retain_lsns,
9801 2 : vec![(Lsn(0x20), child_timeline_id, MaybeOffloaded::Yes)]
9802 2 : );
9803 2 : }
9804 2 :
9805 2 : tenant
9806 2 : .get_offloaded_timeline(child_timeline_id)
9807 2 : .unwrap()
9808 2 : .defuse_for_tenant_drop();
9809 2 :
9810 2 : Ok(())
9811 2 : }
9812 : }
|
