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 : #[derive(Clone)]
193 : pub(super) struct AttachedTenantConf {
194 : tenant_conf: TenantConfOpt,
195 : location: AttachedLocationConfig,
196 : /// The deadline before which we are blocked from GC so that
197 : /// leases have a chance to be renewed.
198 : lsn_lease_deadline: Option<tokio::time::Instant>,
199 : }
200 :
201 : impl AttachedTenantConf {
202 192 : fn new(tenant_conf: TenantConfOpt, location: AttachedLocationConfig) -> Self {
203 : // Sets a deadline before which we cannot proceed to GC due to lsn lease.
204 : //
205 : // We do this as the leases mapping are not persisted to disk. By delaying GC by lease
206 : // length, we guarantee that all the leases we granted before will have a chance to renew
207 : // when we run GC for the first time after restart / transition from AttachedMulti to AttachedSingle.
208 192 : let lsn_lease_deadline = if location.attach_mode == AttachmentMode::Single {
209 192 : Some(
210 192 : tokio::time::Instant::now()
211 192 : + tenant_conf
212 192 : .lsn_lease_length
213 192 : .unwrap_or(LsnLease::DEFAULT_LENGTH),
214 192 : )
215 : } else {
216 : // We don't use `lsn_lease_deadline` to delay GC in AttachedMulti and AttachedStale
217 : // because we don't do GC in these modes.
218 0 : None
219 : };
220 :
221 192 : Self {
222 192 : tenant_conf,
223 192 : location,
224 192 : lsn_lease_deadline,
225 192 : }
226 192 : }
227 :
228 192 : fn try_from(location_conf: LocationConf) -> anyhow::Result<Self> {
229 192 : match &location_conf.mode {
230 192 : LocationMode::Attached(attach_conf) => {
231 192 : Ok(Self::new(location_conf.tenant_conf, *attach_conf))
232 : }
233 : LocationMode::Secondary(_) => {
234 0 : anyhow::bail!("Attempted to construct AttachedTenantConf from a LocationConf in secondary mode")
235 : }
236 : }
237 192 : }
238 :
239 762 : fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
240 762 : self.lsn_lease_deadline
241 762 : .map(|d| tokio::time::Instant::now() < d)
242 762 : .unwrap_or(false)
243 762 : }
244 : }
245 : struct TimelinePreload {
246 : timeline_id: TimelineId,
247 : client: RemoteTimelineClient,
248 : index_part: Result<MaybeDeletedIndexPart, DownloadError>,
249 : }
250 :
251 : pub(crate) struct TenantPreload {
252 : tenant_manifest: TenantManifest,
253 : /// Map from timeline ID to a possible timeline preload. It is None iff the timeline is offloaded according to the manifest.
254 : timelines: HashMap<TimelineId, Option<TimelinePreload>>,
255 : }
256 :
257 : /// When we spawn a tenant, there is a special mode for tenant creation that
258 : /// avoids trying to read anything from remote storage.
259 : pub(crate) enum SpawnMode {
260 : /// Activate as soon as possible
261 : Eager,
262 : /// Lazy activation in the background, with the option to skip the queue if the need comes up
263 : Lazy,
264 : }
265 :
266 : ///
267 : /// Tenant consists of multiple timelines. Keep them in a hash table.
268 : ///
269 : pub struct Tenant {
270 : // Global pageserver config parameters
271 : pub conf: &'static PageServerConf,
272 :
273 : /// The value creation timestamp, used to measure activation delay, see:
274 : /// <https://github.com/neondatabase/neon/issues/4025>
275 : constructed_at: Instant,
276 :
277 : state: watch::Sender<TenantState>,
278 :
279 : // Overridden tenant-specific config parameters.
280 : // We keep TenantConfOpt sturct here to preserve the information
281 : // about parameters that are not set.
282 : // This is necessary to allow global config updates.
283 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
284 :
285 : tenant_shard_id: TenantShardId,
286 :
287 : // The detailed sharding information, beyond the number/count in tenant_shard_id
288 : shard_identity: ShardIdentity,
289 :
290 : /// The remote storage generation, used to protect S3 objects from split-brain.
291 : /// Does not change over the lifetime of the [`Tenant`] object.
292 : ///
293 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
294 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
295 : generation: Generation,
296 :
297 : timelines: Mutex<HashMap<TimelineId, Arc<Timeline>>>,
298 :
299 : /// During timeline creation, we first insert the TimelineId to the
300 : /// creating map, then `timelines`, then remove it from the creating map.
301 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
302 : timelines_creating: std::sync::Mutex<HashSet<TimelineId>>,
303 :
304 : /// Possibly offloaded and archived timelines
305 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
306 : timelines_offloaded: Mutex<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
307 :
308 : /// Serialize writes of the tenant manifest to remote storage. If there are concurrent operations
309 : /// affecting the manifest, such as timeline deletion and timeline offload, they must wait for
310 : /// each other (this could be optimized to coalesce writes if necessary).
311 : ///
312 : /// The contents of the Mutex are the last manifest we successfully uploaded
313 : tenant_manifest_upload: tokio::sync::Mutex<Option<TenantManifest>>,
314 :
315 : // This mutex prevents creation of new timelines during GC.
316 : // Adding yet another mutex (in addition to `timelines`) is needed because holding
317 : // `timelines` mutex during all GC iteration
318 : // may block for a long time `get_timeline`, `get_timelines_state`,... and other operations
319 : // with timelines, which in turn may cause dropping replication connection, expiration of wait_for_lsn
320 : // timeout...
321 : gc_cs: tokio::sync::Mutex<()>,
322 : walredo_mgr: Option<Arc<WalRedoManager>>,
323 :
324 : // provides access to timeline data sitting in the remote storage
325 : pub(crate) remote_storage: GenericRemoteStorage,
326 :
327 : // Access to global deletion queue for when this tenant wants to schedule a deletion
328 : deletion_queue_client: DeletionQueueClient,
329 :
330 : /// Cached logical sizes updated updated on each [`Tenant::gather_size_inputs`].
331 : cached_logical_sizes: tokio::sync::Mutex<HashMap<(TimelineId, Lsn), u64>>,
332 : cached_synthetic_tenant_size: Arc<AtomicU64>,
333 :
334 : eviction_task_tenant_state: tokio::sync::Mutex<EvictionTaskTenantState>,
335 :
336 : /// Track repeated failures to compact, so that we can back off.
337 : /// Overhead of mutex is acceptable because compaction is done with a multi-second period.
338 : compaction_circuit_breaker: std::sync::Mutex<CircuitBreaker>,
339 :
340 : /// If the tenant is in Activating state, notify this to encourage it
341 : /// to proceed to Active as soon as possible, rather than waiting for lazy
342 : /// background warmup.
343 : pub(crate) activate_now_sem: tokio::sync::Semaphore,
344 :
345 : /// Time it took for the tenant to activate. Zero if not active yet.
346 : attach_wal_lag_cooldown: Arc<std::sync::OnceLock<WalLagCooldown>>,
347 :
348 : // Cancellation token fires when we have entered shutdown(). This is a parent of
349 : // Timelines' cancellation token.
350 : pub(crate) cancel: CancellationToken,
351 :
352 : // Users of the Tenant such as the page service must take this Gate to avoid
353 : // trying to use a Tenant which is shutting down.
354 : pub(crate) gate: Gate,
355 :
356 : /// Throttle applied at the top of [`Timeline::get`].
357 : /// All [`Tenant::timelines`] of a given [`Tenant`] instance share the same [`throttle::Throttle`] instance.
358 : pub(crate) timeline_get_throttle:
359 : Arc<throttle::Throttle<crate::metrics::tenant_throttling::TimelineGet>>,
360 :
361 : /// An ongoing timeline detach concurrency limiter.
362 : ///
363 : /// As a tenant will likely be restarted as part of timeline detach ancestor it makes no sense
364 : /// to have two running at the same time. A different one can be started if an earlier one
365 : /// has failed for whatever reason.
366 : ongoing_timeline_detach: std::sync::Mutex<Option<(TimelineId, utils::completion::Barrier)>>,
367 :
368 : /// `index_part.json` based gc blocking reason tracking.
369 : ///
370 : /// New gc iterations must start a new iteration by acquiring `GcBlock::start` before
371 : /// proceeding.
372 : pub(crate) gc_block: gc_block::GcBlock,
373 :
374 : l0_flush_global_state: L0FlushGlobalState,
375 : }
376 :
377 : impl std::fmt::Debug for Tenant {
378 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
379 0 : write!(f, "{} ({})", self.tenant_shard_id, self.current_state())
380 0 : }
381 : }
382 :
383 : pub(crate) enum WalRedoManager {
384 : Prod(WalredoManagerId, PostgresRedoManager),
385 : #[cfg(test)]
386 : Test(harness::TestRedoManager),
387 : }
388 :
389 0 : #[derive(thiserror::Error, Debug)]
390 : #[error("pageserver is shutting down")]
391 : pub(crate) struct GlobalShutDown;
392 :
393 : impl WalRedoManager {
394 0 : pub(crate) fn new(mgr: PostgresRedoManager) -> Result<Arc<Self>, GlobalShutDown> {
395 0 : let id = WalredoManagerId::next();
396 0 : let arc = Arc::new(Self::Prod(id, mgr));
397 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
398 0 : match &mut *guard {
399 0 : Some(map) => {
400 0 : map.insert(id, Arc::downgrade(&arc));
401 0 : Ok(arc)
402 : }
403 0 : None => Err(GlobalShutDown),
404 : }
405 0 : }
406 : }
407 :
408 : impl Drop for WalRedoManager {
409 10 : fn drop(&mut self) {
410 10 : match self {
411 0 : Self::Prod(id, _) => {
412 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
413 0 : if let Some(map) = &mut *guard {
414 0 : map.remove(id).expect("new() registers, drop() unregisters");
415 0 : }
416 : }
417 : #[cfg(test)]
418 10 : Self::Test(_) => {
419 10 : // Not applicable to test redo manager
420 10 : }
421 : }
422 10 : }
423 : }
424 :
425 : /// Global registry of all walredo managers so that [`crate::shutdown_pageserver`] can shut down
426 : /// the walredo processes outside of the regular order.
427 : ///
428 : /// This is necessary to work around a systemd bug where it freezes if there are
429 : /// walredo processes left => <https://github.com/neondatabase/cloud/issues/11387>
430 : #[allow(clippy::type_complexity)]
431 : pub(crate) static WALREDO_MANAGERS: once_cell::sync::Lazy<
432 : Mutex<Option<HashMap<WalredoManagerId, Weak<WalRedoManager>>>>,
433 0 : > = once_cell::sync::Lazy::new(|| Mutex::new(Some(HashMap::new())));
434 : #[derive(PartialEq, Eq, Hash, Clone, Copy, Debug)]
435 : pub(crate) struct WalredoManagerId(u64);
436 : impl WalredoManagerId {
437 0 : pub fn next() -> Self {
438 : static NEXT: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(1);
439 0 : let id = NEXT.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
440 0 : if id == 0 {
441 0 : panic!("WalredoManagerId::new() returned 0, indicating wraparound, risking it's no longer unique");
442 0 : }
443 0 : Self(id)
444 0 : }
445 : }
446 :
447 : #[cfg(test)]
448 : impl From<harness::TestRedoManager> for WalRedoManager {
449 192 : fn from(mgr: harness::TestRedoManager) -> Self {
450 192 : Self::Test(mgr)
451 192 : }
452 : }
453 :
454 : impl WalRedoManager {
455 6 : pub(crate) async fn shutdown(&self) -> bool {
456 6 : match self {
457 0 : Self::Prod(_, mgr) => mgr.shutdown().await,
458 : #[cfg(test)]
459 : Self::Test(_) => {
460 : // Not applicable to test redo manager
461 6 : true
462 : }
463 : }
464 6 : }
465 :
466 0 : pub(crate) fn maybe_quiesce(&self, idle_timeout: Duration) {
467 0 : match self {
468 0 : Self::Prod(_, mgr) => mgr.maybe_quiesce(idle_timeout),
469 0 : #[cfg(test)]
470 0 : Self::Test(_) => {
471 0 : // Not applicable to test redo manager
472 0 : }
473 0 : }
474 0 : }
475 :
476 : /// # Cancel-Safety
477 : ///
478 : /// This method is cancellation-safe.
479 410 : pub async fn request_redo(
480 410 : &self,
481 410 : key: pageserver_api::key::Key,
482 410 : lsn: Lsn,
483 410 : base_img: Option<(Lsn, bytes::Bytes)>,
484 410 : records: Vec<(Lsn, pageserver_api::record::NeonWalRecord)>,
485 410 : pg_version: u32,
486 410 : ) -> Result<bytes::Bytes, walredo::Error> {
487 410 : match self {
488 0 : Self::Prod(_, mgr) => {
489 0 : mgr.request_redo(key, lsn, base_img, records, pg_version)
490 0 : .await
491 : }
492 : #[cfg(test)]
493 410 : Self::Test(mgr) => {
494 410 : mgr.request_redo(key, lsn, base_img, records, pg_version)
495 0 : .await
496 : }
497 : }
498 410 : }
499 :
500 0 : pub(crate) fn status(&self) -> Option<WalRedoManagerStatus> {
501 0 : match self {
502 0 : WalRedoManager::Prod(_, m) => Some(m.status()),
503 0 : #[cfg(test)]
504 0 : WalRedoManager::Test(_) => None,
505 0 : }
506 0 : }
507 : }
508 :
509 : /// A very lightweight memory representation of an offloaded timeline.
510 : ///
511 : /// We need to store the list of offloaded timelines so that we can perform operations on them,
512 : /// like unoffloading them, or (at a later date), decide to perform flattening.
513 : /// This type has a much smaller memory impact than [`Timeline`], and thus we can store many
514 : /// more offloaded timelines than we can manage ones that aren't.
515 : pub struct OffloadedTimeline {
516 : pub tenant_shard_id: TenantShardId,
517 : pub timeline_id: TimelineId,
518 : pub ancestor_timeline_id: Option<TimelineId>,
519 : /// Whether to retain the branch lsn at the ancestor or not
520 : pub ancestor_retain_lsn: Option<Lsn>,
521 :
522 : /// When the timeline was archived.
523 : ///
524 : /// Present for future flattening deliberations.
525 : pub archived_at: NaiveDateTime,
526 :
527 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
528 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
529 : pub delete_progress: TimelineDeleteProgress,
530 :
531 : /// Part of the `OffloadedTimeline` object's lifecycle: this needs to be set before we drop it
532 : pub deleted_from_ancestor: AtomicBool,
533 : }
534 :
535 : impl OffloadedTimeline {
536 : /// Obtains an offloaded timeline from a given timeline object.
537 : ///
538 : /// Returns `None` if the `archived_at` flag couldn't be obtained, i.e.
539 : /// the timeline is not in a stopped state.
540 : /// Panics if the timeline is not archived.
541 2 : fn from_timeline(timeline: &Timeline) -> Result<Self, UploadQueueNotReadyError> {
542 2 : let (ancestor_retain_lsn, ancestor_timeline_id) =
543 2 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
544 2 : let ancestor_lsn = timeline.get_ancestor_lsn();
545 2 : let ancestor_timeline_id = ancestor_timeline.timeline_id;
546 2 : let mut gc_info = ancestor_timeline.gc_info.write().unwrap();
547 2 : gc_info.insert_child(timeline.timeline_id, ancestor_lsn, MaybeOffloaded::Yes);
548 2 : (Some(ancestor_lsn), Some(ancestor_timeline_id))
549 : } else {
550 0 : (None, None)
551 : };
552 2 : let archived_at = timeline
553 2 : .remote_client
554 2 : .archived_at_stopped_queue()?
555 2 : .expect("must be called on an archived timeline");
556 2 : Ok(Self {
557 2 : tenant_shard_id: timeline.tenant_shard_id,
558 2 : timeline_id: timeline.timeline_id,
559 2 : ancestor_timeline_id,
560 2 : ancestor_retain_lsn,
561 2 : archived_at,
562 2 :
563 2 : delete_progress: timeline.delete_progress.clone(),
564 2 : deleted_from_ancestor: AtomicBool::new(false),
565 2 : })
566 2 : }
567 0 : fn from_manifest(tenant_shard_id: TenantShardId, manifest: &OffloadedTimelineManifest) -> Self {
568 0 : // We expect to reach this case in tenant loading, where the `retain_lsn` is populated in the parent's `gc_info`
569 0 : // by the `initialize_gc_info` function.
570 0 : let OffloadedTimelineManifest {
571 0 : timeline_id,
572 0 : ancestor_timeline_id,
573 0 : ancestor_retain_lsn,
574 0 : archived_at,
575 0 : } = *manifest;
576 0 : Self {
577 0 : tenant_shard_id,
578 0 : timeline_id,
579 0 : ancestor_timeline_id,
580 0 : ancestor_retain_lsn,
581 0 : archived_at,
582 0 : delete_progress: TimelineDeleteProgress::default(),
583 0 : deleted_from_ancestor: AtomicBool::new(false),
584 0 : }
585 0 : }
586 2 : fn manifest(&self) -> OffloadedTimelineManifest {
587 2 : let Self {
588 2 : timeline_id,
589 2 : ancestor_timeline_id,
590 2 : ancestor_retain_lsn,
591 2 : archived_at,
592 2 : ..
593 2 : } = self;
594 2 : OffloadedTimelineManifest {
595 2 : timeline_id: *timeline_id,
596 2 : ancestor_timeline_id: *ancestor_timeline_id,
597 2 : ancestor_retain_lsn: *ancestor_retain_lsn,
598 2 : archived_at: *archived_at,
599 2 : }
600 2 : }
601 : /// Delete this timeline's retain_lsn from its ancestor, if present in the given tenant
602 0 : fn delete_from_ancestor_with_timelines(
603 0 : &self,
604 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
605 0 : ) {
606 0 : if let (Some(_retain_lsn), Some(ancestor_timeline_id)) =
607 0 : (self.ancestor_retain_lsn, self.ancestor_timeline_id)
608 : {
609 0 : if let Some((_, ancestor_timeline)) = timelines
610 0 : .iter()
611 0 : .find(|(tid, _tl)| **tid == ancestor_timeline_id)
612 : {
613 0 : let removal_happened = ancestor_timeline
614 0 : .gc_info
615 0 : .write()
616 0 : .unwrap()
617 0 : .remove_child_offloaded(self.timeline_id);
618 0 : if !removal_happened {
619 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
620 0 : "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
621 0 : }
622 0 : }
623 0 : }
624 0 : self.deleted_from_ancestor.store(true, Ordering::Release);
625 0 : }
626 : /// Call [`Self::delete_from_ancestor_with_timelines`] instead if possible.
627 : ///
628 : /// As the entire tenant is being dropped, don't bother deregistering the `retain_lsn` from the ancestor.
629 2 : fn defuse_for_tenant_drop(&self) {
630 2 : self.deleted_from_ancestor.store(true, Ordering::Release);
631 2 : }
632 : }
633 :
634 : impl fmt::Debug for OffloadedTimeline {
635 0 : fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
636 0 : write!(f, "OffloadedTimeline<{}>", self.timeline_id)
637 0 : }
638 : }
639 :
640 : impl Drop for OffloadedTimeline {
641 2 : fn drop(&mut self) {
642 2 : if !self.deleted_from_ancestor.load(Ordering::Acquire) {
643 0 : tracing::warn!(
644 0 : "offloaded timeline {} was dropped without having cleaned it up at the ancestor",
645 : self.timeline_id
646 : );
647 2 : }
648 2 : }
649 : }
650 :
651 : #[derive(Copy, Clone, PartialEq, Eq, Hash, Debug)]
652 : pub enum MaybeOffloaded {
653 : Yes,
654 : No,
655 : }
656 :
657 : #[derive(Clone, Debug)]
658 : pub enum TimelineOrOffloaded {
659 : Timeline(Arc<Timeline>),
660 : Offloaded(Arc<OffloadedTimeline>),
661 : }
662 :
663 : impl TimelineOrOffloaded {
664 0 : pub fn arc_ref(&self) -> TimelineOrOffloadedArcRef<'_> {
665 0 : match self {
666 0 : TimelineOrOffloaded::Timeline(timeline) => {
667 0 : TimelineOrOffloadedArcRef::Timeline(timeline)
668 : }
669 0 : TimelineOrOffloaded::Offloaded(offloaded) => {
670 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded)
671 : }
672 : }
673 0 : }
674 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
675 0 : self.arc_ref().tenant_shard_id()
676 0 : }
677 0 : pub fn timeline_id(&self) -> TimelineId {
678 0 : self.arc_ref().timeline_id()
679 0 : }
680 2 : pub fn delete_progress(&self) -> &Arc<tokio::sync::Mutex<DeleteTimelineFlow>> {
681 2 : match self {
682 2 : TimelineOrOffloaded::Timeline(timeline) => &timeline.delete_progress,
683 0 : TimelineOrOffloaded::Offloaded(offloaded) => &offloaded.delete_progress,
684 : }
685 2 : }
686 0 : fn maybe_remote_client(&self) -> Option<Arc<RemoteTimelineClient>> {
687 0 : match self {
688 0 : TimelineOrOffloaded::Timeline(timeline) => Some(timeline.remote_client.clone()),
689 0 : TimelineOrOffloaded::Offloaded(_offloaded) => None,
690 : }
691 0 : }
692 : }
693 :
694 : pub enum TimelineOrOffloadedArcRef<'a> {
695 : Timeline(&'a Arc<Timeline>),
696 : Offloaded(&'a Arc<OffloadedTimeline>),
697 : }
698 :
699 : impl TimelineOrOffloadedArcRef<'_> {
700 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
701 0 : match self {
702 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.tenant_shard_id,
703 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.tenant_shard_id,
704 : }
705 0 : }
706 0 : pub fn timeline_id(&self) -> TimelineId {
707 0 : match self {
708 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.timeline_id,
709 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.timeline_id,
710 : }
711 0 : }
712 : }
713 :
714 : impl<'a> From<&'a Arc<Timeline>> for TimelineOrOffloadedArcRef<'a> {
715 0 : fn from(timeline: &'a Arc<Timeline>) -> Self {
716 0 : Self::Timeline(timeline)
717 0 : }
718 : }
719 :
720 : impl<'a> From<&'a Arc<OffloadedTimeline>> for TimelineOrOffloadedArcRef<'a> {
721 0 : fn from(timeline: &'a Arc<OffloadedTimeline>) -> Self {
722 0 : Self::Offloaded(timeline)
723 0 : }
724 : }
725 :
726 0 : #[derive(Debug, thiserror::Error, PartialEq, Eq)]
727 : pub enum GetTimelineError {
728 : #[error("Timeline is shutting down")]
729 : ShuttingDown,
730 : #[error("Timeline {tenant_id}/{timeline_id} is not active, state: {state:?}")]
731 : NotActive {
732 : tenant_id: TenantShardId,
733 : timeline_id: TimelineId,
734 : state: TimelineState,
735 : },
736 : #[error("Timeline {tenant_id}/{timeline_id} was not found")]
737 : NotFound {
738 : tenant_id: TenantShardId,
739 : timeline_id: TimelineId,
740 : },
741 : }
742 :
743 0 : #[derive(Debug, thiserror::Error)]
744 : pub enum LoadLocalTimelineError {
745 : #[error("FailedToLoad")]
746 : Load(#[source] anyhow::Error),
747 : #[error("FailedToResumeDeletion")]
748 : ResumeDeletion(#[source] anyhow::Error),
749 : }
750 :
751 0 : #[derive(thiserror::Error)]
752 : pub enum DeleteTimelineError {
753 : #[error("NotFound")]
754 : NotFound,
755 :
756 : #[error("HasChildren")]
757 : HasChildren(Vec<TimelineId>),
758 :
759 : #[error("Timeline deletion is already in progress")]
760 : AlreadyInProgress(Arc<tokio::sync::Mutex<DeleteTimelineFlow>>),
761 :
762 : #[error("Cancelled")]
763 : Cancelled,
764 :
765 : #[error(transparent)]
766 : Other(#[from] anyhow::Error),
767 : }
768 :
769 : impl Debug for DeleteTimelineError {
770 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
771 0 : match self {
772 0 : Self::NotFound => write!(f, "NotFound"),
773 0 : Self::HasChildren(c) => f.debug_tuple("HasChildren").field(c).finish(),
774 0 : Self::AlreadyInProgress(_) => f.debug_tuple("AlreadyInProgress").finish(),
775 0 : Self::Cancelled => f.debug_tuple("Cancelled").finish(),
776 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
777 : }
778 0 : }
779 : }
780 :
781 0 : #[derive(thiserror::Error)]
782 : pub enum TimelineArchivalError {
783 : #[error("NotFound")]
784 : NotFound,
785 :
786 : #[error("Timeout")]
787 : Timeout,
788 :
789 : #[error("Cancelled")]
790 : Cancelled,
791 :
792 : #[error("ancestor is archived: {}", .0)]
793 : HasArchivedParent(TimelineId),
794 :
795 : #[error("HasUnarchivedChildren")]
796 : HasUnarchivedChildren(Vec<TimelineId>),
797 :
798 : #[error("Timeline archival is already in progress")]
799 : AlreadyInProgress,
800 :
801 : #[error(transparent)]
802 : Other(anyhow::Error),
803 : }
804 :
805 0 : #[derive(thiserror::Error, Debug)]
806 : pub(crate) enum TenantManifestError {
807 : #[error("Remote storage error: {0}")]
808 : RemoteStorage(anyhow::Error),
809 :
810 : #[error("Cancelled")]
811 : Cancelled,
812 : }
813 :
814 : impl From<TenantManifestError> for TimelineArchivalError {
815 0 : fn from(e: TenantManifestError) -> Self {
816 0 : match e {
817 0 : TenantManifestError::RemoteStorage(e) => Self::Other(e),
818 0 : TenantManifestError::Cancelled => Self::Cancelled,
819 : }
820 0 : }
821 : }
822 :
823 : impl Debug for TimelineArchivalError {
824 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
825 0 : match self {
826 0 : Self::NotFound => write!(f, "NotFound"),
827 0 : Self::Timeout => write!(f, "Timeout"),
828 0 : Self::Cancelled => write!(f, "Cancelled"),
829 0 : Self::HasArchivedParent(p) => f.debug_tuple("HasArchivedParent").field(p).finish(),
830 0 : Self::HasUnarchivedChildren(c) => {
831 0 : f.debug_tuple("HasUnarchivedChildren").field(c).finish()
832 : }
833 0 : Self::AlreadyInProgress => f.debug_tuple("AlreadyInProgress").finish(),
834 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
835 : }
836 0 : }
837 : }
838 :
839 : pub enum SetStoppingError {
840 : AlreadyStopping(completion::Barrier),
841 : Broken,
842 : }
843 :
844 : impl Debug for SetStoppingError {
845 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
846 0 : match self {
847 0 : Self::AlreadyStopping(_) => f.debug_tuple("AlreadyStopping").finish(),
848 0 : Self::Broken => write!(f, "Broken"),
849 : }
850 0 : }
851 : }
852 :
853 : /// Arguments to [`Tenant::create_timeline`].
854 : ///
855 : /// Not usable as an idempotency key for timeline creation because if [`CreateTimelineParamsBranch::ancestor_start_lsn`]
856 : /// is `None`, the result of the timeline create call is not deterministic.
857 : ///
858 : /// See [`CreateTimelineIdempotency`] for an idempotency key.
859 : #[derive(Debug)]
860 : pub(crate) enum CreateTimelineParams {
861 : Bootstrap(CreateTimelineParamsBootstrap),
862 : Branch(CreateTimelineParamsBranch),
863 : }
864 :
865 : #[derive(Debug)]
866 : pub(crate) struct CreateTimelineParamsBootstrap {
867 : pub(crate) new_timeline_id: TimelineId,
868 : pub(crate) existing_initdb_timeline_id: Option<TimelineId>,
869 : pub(crate) pg_version: u32,
870 : }
871 :
872 : /// NB: See comment on [`CreateTimelineIdempotency::Branch`] for why there's no `pg_version` here.
873 : #[derive(Debug)]
874 : pub(crate) struct CreateTimelineParamsBranch {
875 : pub(crate) new_timeline_id: TimelineId,
876 : pub(crate) ancestor_timeline_id: TimelineId,
877 : pub(crate) ancestor_start_lsn: Option<Lsn>,
878 : }
879 :
880 : /// What is used to determine idempotency of a [`Tenant::create_timeline`] call in [`Tenant::start_creating_timeline`].
881 : ///
882 : /// Each [`Timeline`] object holds [`Self`] as an immutable property in [`Timeline::create_idempotency`].
883 : ///
884 : /// We lower timeline creation requests to [`Self`], and then use [`PartialEq::eq`] to compare [`Timeline::create_idempotency`] with the request.
885 : /// If they are equal, we return a reference to the existing timeline, otherwise it's an idempotency conflict.
886 : ///
887 : /// There is special treatment for [`Self::FailWithConflict`] to always return an idempotency conflict.
888 : /// It would be nice to have more advanced derive macros to make that special treatment declarative.
889 : ///
890 : /// Notes:
891 : /// - Unlike [`CreateTimelineParams`], ancestor LSN is fixed, so, branching will be at a deterministic LSN.
892 : /// - We make some trade-offs though, e.g., [`CreateTimelineParamsBootstrap::existing_initdb_timeline_id`]
893 : /// is not considered for idempotency. We can improve on this over time if we deem it necessary.
894 : ///
895 : #[derive(Debug, Clone, PartialEq, Eq)]
896 : pub(crate) enum CreateTimelineIdempotency {
897 : /// NB: special treatment, see comment in [`Self`].
898 : FailWithConflict,
899 : Bootstrap {
900 : pg_version: u32,
901 : },
902 : /// NB: branches always have the same `pg_version` as their ancestor.
903 : /// While [`pageserver_api::models::TimelineCreateRequestMode::Branch::pg_version`]
904 : /// exists as a field, and is set by cplane, it has always been ignored by pageserver when
905 : /// determining the child branch pg_version.
906 : Branch {
907 : ancestor_timeline_id: TimelineId,
908 : ancestor_start_lsn: Lsn,
909 : },
910 : }
911 :
912 : /// What is returned by [`Tenant::start_creating_timeline`].
913 : #[must_use]
914 : enum StartCreatingTimelineResult<'t> {
915 : CreateGuard(TimelineCreateGuard<'t>),
916 : Idempotent(Arc<Timeline>),
917 : }
918 :
919 : /// What is returned by [`Tenant::create_timeline`].
920 : enum CreateTimelineResult {
921 : Created(Arc<Timeline>),
922 : Idempotent(Arc<Timeline>),
923 : }
924 :
925 : impl CreateTimelineResult {
926 0 : fn discriminant(&self) -> &'static str {
927 0 : match self {
928 0 : Self::Created(_) => "Created",
929 0 : Self::Idempotent(_) => "Idempotent",
930 : }
931 0 : }
932 0 : fn timeline(&self) -> &Arc<Timeline> {
933 0 : match self {
934 0 : Self::Created(t) | Self::Idempotent(t) => t,
935 0 : }
936 0 : }
937 : /// Unit test timelines aren't activated, test has to do it if it needs to.
938 : #[cfg(test)]
939 230 : fn into_timeline_for_test(self) -> Arc<Timeline> {
940 230 : match self {
941 230 : Self::Created(t) | Self::Idempotent(t) => t,
942 230 : }
943 230 : }
944 : }
945 :
946 2 : #[derive(thiserror::Error, Debug)]
947 : pub enum CreateTimelineError {
948 : #[error("creation of timeline with the given ID is in progress")]
949 : AlreadyCreating,
950 : #[error("timeline already exists with different parameters")]
951 : Conflict,
952 : #[error(transparent)]
953 : AncestorLsn(anyhow::Error),
954 : #[error("ancestor timeline is not active")]
955 : AncestorNotActive,
956 : #[error("ancestor timeline is archived")]
957 : AncestorArchived,
958 : #[error("tenant shutting down")]
959 : ShuttingDown,
960 : #[error(transparent)]
961 : Other(#[from] anyhow::Error),
962 : }
963 :
964 : #[derive(thiserror::Error, Debug)]
965 : enum InitdbError {
966 : Other(anyhow::Error),
967 : Cancelled,
968 : Spawn(std::io::Result<()>),
969 : Failed(std::process::ExitStatus, Vec<u8>),
970 : }
971 :
972 : impl fmt::Display for InitdbError {
973 0 : fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
974 0 : match self {
975 0 : InitdbError::Cancelled => write!(f, "Operation was cancelled"),
976 0 : InitdbError::Spawn(e) => write!(f, "Spawn error: {:?}", e),
977 0 : InitdbError::Failed(status, stderr) => write!(
978 0 : f,
979 0 : "Command failed with status {:?}: {}",
980 0 : status,
981 0 : String::from_utf8_lossy(stderr)
982 0 : ),
983 0 : InitdbError::Other(e) => write!(f, "Error: {:?}", e),
984 : }
985 0 : }
986 : }
987 :
988 : impl From<std::io::Error> for InitdbError {
989 0 : fn from(error: std::io::Error) -> Self {
990 0 : InitdbError::Spawn(Err(error))
991 0 : }
992 : }
993 :
994 : enum CreateTimelineCause {
995 : Load,
996 : Delete,
997 : }
998 :
999 0 : #[derive(thiserror::Error, Debug)]
1000 : pub(crate) enum GcError {
1001 : // The tenant is shutting down
1002 : #[error("tenant shutting down")]
1003 : TenantCancelled,
1004 :
1005 : // The tenant is shutting down
1006 : #[error("timeline shutting down")]
1007 : TimelineCancelled,
1008 :
1009 : // The tenant is in a state inelegible to run GC
1010 : #[error("not active")]
1011 : NotActive,
1012 :
1013 : // A requested GC cutoff LSN was invalid, for example it tried to move backwards
1014 : #[error("not active")]
1015 : BadLsn { why: String },
1016 :
1017 : // A remote storage error while scheduling updates after compaction
1018 : #[error(transparent)]
1019 : Remote(anyhow::Error),
1020 :
1021 : // An error reading while calculating GC cutoffs
1022 : #[error(transparent)]
1023 : GcCutoffs(PageReconstructError),
1024 :
1025 : // If GC was invoked for a particular timeline, this error means it didn't exist
1026 : #[error("timeline not found")]
1027 : TimelineNotFound,
1028 : }
1029 :
1030 : impl From<PageReconstructError> for GcError {
1031 0 : fn from(value: PageReconstructError) -> Self {
1032 0 : match value {
1033 0 : PageReconstructError::Cancelled => Self::TimelineCancelled,
1034 0 : other => Self::GcCutoffs(other),
1035 : }
1036 0 : }
1037 : }
1038 :
1039 : impl From<NotInitialized> for GcError {
1040 0 : fn from(value: NotInitialized) -> Self {
1041 0 : match value {
1042 0 : NotInitialized::Uninitialized => GcError::Remote(value.into()),
1043 0 : NotInitialized::Stopped | NotInitialized::ShuttingDown => GcError::TimelineCancelled,
1044 : }
1045 0 : }
1046 : }
1047 :
1048 : impl From<timeline::layer_manager::Shutdown> for GcError {
1049 0 : fn from(_: timeline::layer_manager::Shutdown) -> Self {
1050 0 : GcError::TimelineCancelled
1051 0 : }
1052 : }
1053 :
1054 0 : #[derive(thiserror::Error, Debug)]
1055 : pub(crate) enum LoadConfigError {
1056 : #[error("TOML deserialization error: '{0}'")]
1057 : DeserializeToml(#[from] toml_edit::de::Error),
1058 :
1059 : #[error("Config not found at {0}")]
1060 : NotFound(Utf8PathBuf),
1061 : }
1062 :
1063 : impl Tenant {
1064 : /// Yet another helper for timeline initialization.
1065 : ///
1066 : /// - Initializes the Timeline struct and inserts it into the tenant's hash map
1067 : /// - Scans the local timeline directory for layer files and builds the layer map
1068 : /// - Downloads remote index file and adds remote files to the layer map
1069 : /// - Schedules remote upload tasks for any files that are present locally but missing from remote storage.
1070 : ///
1071 : /// If the operation fails, the timeline is left in the tenant's hash map in Broken state. On success,
1072 : /// it is marked as Active.
1073 : #[allow(clippy::too_many_arguments)]
1074 6 : async fn timeline_init_and_sync(
1075 6 : &self,
1076 6 : timeline_id: TimelineId,
1077 6 : resources: TimelineResources,
1078 6 : index_part: IndexPart,
1079 6 : metadata: TimelineMetadata,
1080 6 : ancestor: Option<Arc<Timeline>>,
1081 6 : _ctx: &RequestContext,
1082 6 : ) -> anyhow::Result<()> {
1083 6 : let tenant_id = self.tenant_shard_id;
1084 :
1085 6 : let idempotency = if metadata.ancestor_timeline().is_none() {
1086 4 : CreateTimelineIdempotency::Bootstrap {
1087 4 : pg_version: metadata.pg_version(),
1088 4 : }
1089 : } else {
1090 2 : CreateTimelineIdempotency::Branch {
1091 2 : ancestor_timeline_id: metadata.ancestor_timeline().unwrap(),
1092 2 : ancestor_start_lsn: metadata.ancestor_lsn(),
1093 2 : }
1094 : };
1095 :
1096 6 : let timeline = self.create_timeline_struct(
1097 6 : timeline_id,
1098 6 : &metadata,
1099 6 : ancestor.clone(),
1100 6 : resources,
1101 6 : CreateTimelineCause::Load,
1102 6 : idempotency.clone(),
1103 6 : )?;
1104 6 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
1105 6 : anyhow::ensure!(
1106 6 : disk_consistent_lsn.is_valid(),
1107 0 : "Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn"
1108 : );
1109 6 : assert_eq!(
1110 6 : disk_consistent_lsn,
1111 6 : metadata.disk_consistent_lsn(),
1112 0 : "these are used interchangeably"
1113 : );
1114 :
1115 6 : timeline.remote_client.init_upload_queue(&index_part)?;
1116 :
1117 6 : timeline
1118 6 : .load_layer_map(disk_consistent_lsn, index_part)
1119 1 : .await
1120 6 : .with_context(|| {
1121 0 : format!("Failed to load layermap for timeline {tenant_id}/{timeline_id}")
1122 6 : })?;
1123 :
1124 : {
1125 : // avoiding holding it across awaits
1126 6 : let mut timelines_accessor = self.timelines.lock().unwrap();
1127 6 : match timelines_accessor.entry(timeline_id) {
1128 : // We should never try and load the same timeline twice during startup
1129 : Entry::Occupied(_) => {
1130 0 : unreachable!(
1131 0 : "Timeline {tenant_id}/{timeline_id} already exists in the tenant map"
1132 0 : );
1133 : }
1134 6 : Entry::Vacant(v) => {
1135 6 : v.insert(Arc::clone(&timeline));
1136 6 : timeline.maybe_spawn_flush_loop();
1137 6 : }
1138 6 : }
1139 6 : };
1140 6 :
1141 6 : // Sanity check: a timeline should have some content.
1142 6 : anyhow::ensure!(
1143 6 : ancestor.is_some()
1144 4 : || timeline
1145 4 : .layers
1146 4 : .read()
1147 0 : .await
1148 4 : .layer_map()
1149 4 : .expect("currently loading, layer manager cannot be shutdown already")
1150 4 : .iter_historic_layers()
1151 4 : .next()
1152 4 : .is_some(),
1153 0 : "Timeline has no ancestor and no layer files"
1154 : );
1155 :
1156 6 : Ok(())
1157 6 : }
1158 :
1159 : /// Attach a tenant that's available in cloud storage.
1160 : ///
1161 : /// This returns quickly, after just creating the in-memory object
1162 : /// Tenant struct and launching a background task to download
1163 : /// the remote index files. On return, the tenant is most likely still in
1164 : /// Attaching state, and it will become Active once the background task
1165 : /// finishes. You can use wait_until_active() to wait for the task to
1166 : /// complete.
1167 : ///
1168 : #[allow(clippy::too_many_arguments)]
1169 0 : pub(crate) fn spawn(
1170 0 : conf: &'static PageServerConf,
1171 0 : tenant_shard_id: TenantShardId,
1172 0 : resources: TenantSharedResources,
1173 0 : attached_conf: AttachedTenantConf,
1174 0 : shard_identity: ShardIdentity,
1175 0 : init_order: Option<InitializationOrder>,
1176 0 : mode: SpawnMode,
1177 0 : ctx: &RequestContext,
1178 0 : ) -> Result<Arc<Tenant>, GlobalShutDown> {
1179 0 : let wal_redo_manager =
1180 0 : WalRedoManager::new(PostgresRedoManager::new(conf, tenant_shard_id))?;
1181 :
1182 : let TenantSharedResources {
1183 0 : broker_client,
1184 0 : remote_storage,
1185 0 : deletion_queue_client,
1186 0 : l0_flush_global_state,
1187 0 : } = resources;
1188 0 :
1189 0 : let attach_mode = attached_conf.location.attach_mode;
1190 0 : let generation = attached_conf.location.generation;
1191 0 :
1192 0 : let tenant = Arc::new(Tenant::new(
1193 0 : TenantState::Attaching,
1194 0 : conf,
1195 0 : attached_conf,
1196 0 : shard_identity,
1197 0 : Some(wal_redo_manager),
1198 0 : tenant_shard_id,
1199 0 : remote_storage.clone(),
1200 0 : deletion_queue_client,
1201 0 : l0_flush_global_state,
1202 0 : ));
1203 0 :
1204 0 : // The attach task will carry a GateGuard, so that shutdown() reliably waits for it to drop out if
1205 0 : // we shut down while attaching.
1206 0 : let attach_gate_guard = tenant
1207 0 : .gate
1208 0 : .enter()
1209 0 : .expect("We just created the Tenant: nothing else can have shut it down yet");
1210 0 :
1211 0 : // Do all the hard work in the background
1212 0 : let tenant_clone = Arc::clone(&tenant);
1213 0 : let ctx = ctx.detached_child(TaskKind::Attach, DownloadBehavior::Warn);
1214 0 : task_mgr::spawn(
1215 0 : &tokio::runtime::Handle::current(),
1216 0 : TaskKind::Attach,
1217 0 : tenant_shard_id,
1218 0 : None,
1219 0 : "attach tenant",
1220 0 : async move {
1221 0 :
1222 0 : info!(
1223 : ?attach_mode,
1224 0 : "Attaching tenant"
1225 : );
1226 :
1227 0 : let _gate_guard = attach_gate_guard;
1228 0 :
1229 0 : // Is this tenant being spawned as part of process startup?
1230 0 : let starting_up = init_order.is_some();
1231 0 : scopeguard::defer! {
1232 0 : if starting_up {
1233 0 : TENANT.startup_complete.inc();
1234 0 : }
1235 0 : }
1236 :
1237 : // Ideally we should use Tenant::set_broken_no_wait, but it is not supposed to be used when tenant is in loading state.
1238 : enum BrokenVerbosity {
1239 : Error,
1240 : Info
1241 : }
1242 0 : let make_broken =
1243 0 : |t: &Tenant, err: anyhow::Error, verbosity: BrokenVerbosity| {
1244 0 : match verbosity {
1245 : BrokenVerbosity::Info => {
1246 0 : info!("attach cancelled, setting tenant state to Broken: {err}");
1247 : },
1248 : BrokenVerbosity::Error => {
1249 0 : error!("attach failed, setting tenant state to Broken: {err:?}");
1250 : }
1251 : }
1252 0 : t.state.send_modify(|state| {
1253 0 : // The Stopping case is for when we have passed control on to DeleteTenantFlow:
1254 0 : // if it errors, we will call make_broken when tenant is already in Stopping.
1255 0 : assert!(
1256 0 : matches!(*state, TenantState::Attaching | TenantState::Stopping { .. }),
1257 0 : "the attach task owns the tenant state until activation is complete"
1258 : );
1259 :
1260 0 : *state = TenantState::broken_from_reason(err.to_string());
1261 0 : });
1262 0 : };
1263 :
1264 : // TODO: should also be rejecting tenant conf changes that violate this check.
1265 0 : if let Err(e) = crate::tenant::storage_layer::inmemory_layer::IndexEntry::validate_checkpoint_distance(tenant_clone.get_checkpoint_distance()) {
1266 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1267 0 : return Ok(());
1268 0 : }
1269 0 :
1270 0 : let mut init_order = init_order;
1271 0 : // take the completion because initial tenant loading will complete when all of
1272 0 : // these tasks complete.
1273 0 : let _completion = init_order
1274 0 : .as_mut()
1275 0 : .and_then(|x| x.initial_tenant_load.take());
1276 0 : let remote_load_completion = init_order
1277 0 : .as_mut()
1278 0 : .and_then(|x| x.initial_tenant_load_remote.take());
1279 :
1280 : enum AttachType<'a> {
1281 : /// We are attaching this tenant lazily in the background.
1282 : Warmup {
1283 : _permit: tokio::sync::SemaphorePermit<'a>,
1284 : during_startup: bool
1285 : },
1286 : /// We are attaching this tenant as soon as we can, because for example an
1287 : /// endpoint tried to access it.
1288 : OnDemand,
1289 : /// During normal operations after startup, we are attaching a tenant, and
1290 : /// eager attach was requested.
1291 : Normal,
1292 : }
1293 :
1294 0 : let attach_type = if matches!(mode, SpawnMode::Lazy) {
1295 : // Before doing any I/O, wait for at least one of:
1296 : // - A client attempting to access to this tenant (on-demand loading)
1297 : // - A permit becoming available in the warmup semaphore (background warmup)
1298 :
1299 0 : tokio::select!(
1300 0 : permit = tenant_clone.activate_now_sem.acquire() => {
1301 0 : let _ = permit.expect("activate_now_sem is never closed");
1302 0 : tracing::info!("Activating tenant (on-demand)");
1303 0 : AttachType::OnDemand
1304 : },
1305 0 : permit = conf.concurrent_tenant_warmup.inner().acquire() => {
1306 0 : let _permit = permit.expect("concurrent_tenant_warmup semaphore is never closed");
1307 0 : tracing::info!("Activating tenant (warmup)");
1308 0 : AttachType::Warmup {
1309 0 : _permit,
1310 0 : during_startup: init_order.is_some()
1311 0 : }
1312 : }
1313 0 : _ = tenant_clone.cancel.cancelled() => {
1314 : // This is safe, but should be pretty rare: it is interesting if a tenant
1315 : // stayed in Activating for such a long time that shutdown found it in
1316 : // that state.
1317 0 : tracing::info!(state=%tenant_clone.current_state(), "Tenant shut down before activation");
1318 : // Make the tenant broken so that set_stopping will not hang waiting for it to leave
1319 : // the Attaching state. This is an over-reaction (nothing really broke, the tenant is
1320 : // just shutting down), but ensures progress.
1321 0 : make_broken(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"), BrokenVerbosity::Info);
1322 0 : return Ok(());
1323 : },
1324 : )
1325 : } else {
1326 : // SpawnMode::{Create,Eager} always cause jumping ahead of the
1327 : // concurrent_tenant_warmup queue
1328 0 : AttachType::Normal
1329 : };
1330 :
1331 0 : let preload = match &mode {
1332 : SpawnMode::Eager | SpawnMode::Lazy => {
1333 0 : let _preload_timer = TENANT.preload.start_timer();
1334 0 : let res = tenant_clone
1335 0 : .preload(&remote_storage, task_mgr::shutdown_token())
1336 0 : .await;
1337 0 : match res {
1338 0 : Ok(p) => Some(p),
1339 0 : Err(e) => {
1340 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1341 0 : return Ok(());
1342 : }
1343 : }
1344 : }
1345 :
1346 : };
1347 :
1348 : // Remote preload is complete.
1349 0 : drop(remote_load_completion);
1350 0 :
1351 0 :
1352 0 : // We will time the duration of the attach phase unless this is a creation (attach will do no work)
1353 0 : let attach_start = std::time::Instant::now();
1354 0 : let attached = {
1355 0 : let _attach_timer = Some(TENANT.attach.start_timer());
1356 0 : tenant_clone.attach(preload, &ctx).await
1357 : };
1358 0 : let attach_duration = attach_start.elapsed();
1359 0 : _ = tenant_clone.attach_wal_lag_cooldown.set(WalLagCooldown::new(attach_start, attach_duration));
1360 0 :
1361 0 : match attached {
1362 : Ok(()) => {
1363 0 : info!("attach finished, activating");
1364 0 : tenant_clone.activate(broker_client, None, &ctx);
1365 : }
1366 0 : Err(e) => {
1367 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1368 0 : }
1369 : }
1370 :
1371 : // If we are doing an opportunistic warmup attachment at startup, initialize
1372 : // logical size at the same time. This is better than starting a bunch of idle tenants
1373 : // with cold caches and then coming back later to initialize their logical sizes.
1374 : //
1375 : // It also prevents the warmup proccess competing with the concurrency limit on
1376 : // logical size calculations: if logical size calculation semaphore is saturated,
1377 : // then warmup will wait for that before proceeding to the next tenant.
1378 0 : if matches!(attach_type, AttachType::Warmup { during_startup: true, .. }) {
1379 0 : let mut futs: FuturesUnordered<_> = tenant_clone.timelines.lock().unwrap().values().cloned().map(|t| t.await_initial_logical_size()).collect();
1380 0 : tracing::info!("Waiting for initial logical sizes while warming up...");
1381 0 : while futs.next().await.is_some() {}
1382 0 : tracing::info!("Warm-up complete");
1383 0 : }
1384 :
1385 0 : Ok(())
1386 0 : }
1387 0 : .instrument(tracing::info_span!(parent: None, "attach", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), gen=?generation)),
1388 : );
1389 0 : Ok(tenant)
1390 0 : }
1391 :
1392 192 : #[instrument(skip_all)]
1393 : pub(crate) async fn preload(
1394 : self: &Arc<Self>,
1395 : remote_storage: &GenericRemoteStorage,
1396 : cancel: CancellationToken,
1397 : ) -> anyhow::Result<TenantPreload> {
1398 : span::debug_assert_current_span_has_tenant_id();
1399 : // Get list of remote timelines
1400 : // download index files for every tenant timeline
1401 : info!("listing remote timelines");
1402 : let (mut remote_timeline_ids, other_keys) = remote_timeline_client::list_remote_timelines(
1403 : remote_storage,
1404 : self.tenant_shard_id,
1405 : cancel.clone(),
1406 : )
1407 : .await?;
1408 : let (offloaded_add, tenant_manifest) =
1409 : match remote_timeline_client::download_tenant_manifest(
1410 : remote_storage,
1411 : &self.tenant_shard_id,
1412 : self.generation,
1413 : &cancel,
1414 : )
1415 : .await
1416 : {
1417 : Ok((tenant_manifest, _generation, _manifest_mtime)) => (
1418 : format!("{} offloaded", tenant_manifest.offloaded_timelines.len()),
1419 : tenant_manifest,
1420 : ),
1421 : Err(DownloadError::NotFound) => {
1422 : ("no manifest".to_string(), TenantManifest::empty())
1423 : }
1424 : Err(e) => Err(e)?,
1425 : };
1426 :
1427 : info!(
1428 : "found {} timelines, and {offloaded_add}",
1429 : remote_timeline_ids.len()
1430 : );
1431 :
1432 : for k in other_keys {
1433 : warn!("Unexpected non timeline key {k}");
1434 : }
1435 :
1436 : // Avoid downloading IndexPart of offloaded timelines.
1437 : let mut offloaded_with_prefix = HashSet::new();
1438 : for offloaded in tenant_manifest.offloaded_timelines.iter() {
1439 : if remote_timeline_ids.remove(&offloaded.timeline_id) {
1440 : offloaded_with_prefix.insert(offloaded.timeline_id);
1441 : } else {
1442 : // We'll take care later of timelines in the manifest without a prefix
1443 : }
1444 : }
1445 :
1446 : let timelines = self
1447 : .load_timelines_metadata(remote_timeline_ids, remote_storage, cancel)
1448 : .await?;
1449 :
1450 : Ok(TenantPreload {
1451 : tenant_manifest,
1452 : timelines: timelines
1453 : .into_iter()
1454 6 : .map(|(id, tl)| (id, Some(tl)))
1455 0 : .chain(offloaded_with_prefix.into_iter().map(|id| (id, None)))
1456 : .collect(),
1457 : })
1458 : }
1459 :
1460 : ///
1461 : /// Background task that downloads all data for a tenant and brings it to Active state.
1462 : ///
1463 : /// No background tasks are started as part of this routine.
1464 : ///
1465 192 : async fn attach(
1466 192 : self: &Arc<Tenant>,
1467 192 : preload: Option<TenantPreload>,
1468 192 : ctx: &RequestContext,
1469 192 : ) -> anyhow::Result<()> {
1470 192 : span::debug_assert_current_span_has_tenant_id();
1471 192 :
1472 192 : failpoint_support::sleep_millis_async!("before-attaching-tenant");
1473 :
1474 192 : let Some(preload) = preload else {
1475 0 : anyhow::bail!("local-only deployment is no longer supported, https://github.com/neondatabase/neon/issues/5624");
1476 : };
1477 :
1478 192 : let mut offloaded_timeline_ids = HashSet::new();
1479 192 : let mut offloaded_timelines_list = Vec::new();
1480 192 : for timeline_manifest in preload.tenant_manifest.offloaded_timelines.iter() {
1481 0 : let timeline_id = timeline_manifest.timeline_id;
1482 0 : let offloaded_timeline =
1483 0 : OffloadedTimeline::from_manifest(self.tenant_shard_id, timeline_manifest);
1484 0 : offloaded_timelines_list.push((timeline_id, Arc::new(offloaded_timeline)));
1485 0 : offloaded_timeline_ids.insert(timeline_id);
1486 0 : }
1487 : // Complete deletions for offloaded timeline id's from manifest.
1488 : // The manifest will be uploaded later in this function.
1489 192 : offloaded_timelines_list
1490 192 : .retain(|(offloaded_id, offloaded)| {
1491 0 : // Existence of a timeline is finally determined by the existence of an index-part.json in remote storage.
1492 0 : // If there is dangling references in another location, they need to be cleaned up.
1493 0 : let delete = !preload.timelines.contains_key(offloaded_id);
1494 0 : if delete {
1495 0 : tracing::info!("Removing offloaded timeline {offloaded_id} from manifest as no remote prefix was found");
1496 0 : offloaded.defuse_for_tenant_drop();
1497 0 : }
1498 0 : !delete
1499 192 : });
1500 192 :
1501 192 : let mut timelines_to_resume_deletions = vec![];
1502 192 :
1503 192 : let mut remote_index_and_client = HashMap::new();
1504 192 : let mut timeline_ancestors = HashMap::new();
1505 192 : let mut existent_timelines = HashSet::new();
1506 198 : for (timeline_id, preload) in preload.timelines {
1507 6 : let Some(preload) = preload else { continue };
1508 : // This is an invariant of the `preload` function's API
1509 6 : assert!(!offloaded_timeline_ids.contains(&timeline_id));
1510 6 : let index_part = match preload.index_part {
1511 6 : Ok(i) => {
1512 6 : debug!("remote index part exists for timeline {timeline_id}");
1513 : // We found index_part on the remote, this is the standard case.
1514 6 : existent_timelines.insert(timeline_id);
1515 6 : i
1516 : }
1517 : Err(DownloadError::NotFound) => {
1518 : // There is no index_part on the remote. We only get here
1519 : // if there is some prefix for the timeline in the remote storage.
1520 : // This can e.g. be the initdb.tar.zst archive, maybe a
1521 : // remnant from a prior incomplete creation or deletion attempt.
1522 : // Delete the local directory as the deciding criterion for a
1523 : // timeline's existence is presence of index_part.
1524 0 : info!(%timeline_id, "index_part not found on remote");
1525 0 : continue;
1526 : }
1527 0 : Err(DownloadError::Fatal(why)) => {
1528 0 : // If, while loading one remote timeline, we saw an indication that our generation
1529 0 : // number is likely invalid, then we should not load the whole tenant.
1530 0 : error!(%timeline_id, "Fatal error loading timeline: {why}");
1531 0 : anyhow::bail!(why.to_string());
1532 : }
1533 0 : Err(e) => {
1534 0 : // Some (possibly ephemeral) error happened during index_part download.
1535 0 : // Pretend the timeline exists to not delete the timeline directory,
1536 0 : // as it might be a temporary issue and we don't want to re-download
1537 0 : // everything after it resolves.
1538 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1539 :
1540 0 : existent_timelines.insert(timeline_id);
1541 0 : continue;
1542 : }
1543 : };
1544 6 : match index_part {
1545 6 : MaybeDeletedIndexPart::IndexPart(index_part) => {
1546 6 : timeline_ancestors.insert(timeline_id, index_part.metadata.clone());
1547 6 : remote_index_and_client.insert(timeline_id, (index_part, preload.client));
1548 6 : }
1549 0 : MaybeDeletedIndexPart::Deleted(index_part) => {
1550 0 : info!(
1551 0 : "timeline {} is deleted, picking to resume deletion",
1552 : timeline_id
1553 : );
1554 0 : timelines_to_resume_deletions.push((timeline_id, index_part, preload.client));
1555 : }
1556 : }
1557 : }
1558 :
1559 192 : let mut gc_blocks = HashMap::new();
1560 :
1561 : // For every timeline, download the metadata file, scan the local directory,
1562 : // and build a layer map that contains an entry for each remote and local
1563 : // layer file.
1564 192 : let sorted_timelines = tree_sort_timelines(timeline_ancestors, |m| m.ancestor_timeline())?;
1565 198 : for (timeline_id, remote_metadata) in sorted_timelines {
1566 6 : let (index_part, remote_client) = remote_index_and_client
1567 6 : .remove(&timeline_id)
1568 6 : .expect("just put it in above");
1569 :
1570 6 : if let Some(blocking) = index_part.gc_blocking.as_ref() {
1571 : // could just filter these away, but it helps while testing
1572 0 : anyhow::ensure!(
1573 0 : !blocking.reasons.is_empty(),
1574 0 : "index_part for {timeline_id} is malformed: it should not have gc blocking with zero reasons"
1575 : );
1576 0 : let prev = gc_blocks.insert(timeline_id, blocking.reasons);
1577 0 : assert!(prev.is_none());
1578 6 : }
1579 :
1580 : // TODO again handle early failure
1581 6 : self.load_remote_timeline(
1582 6 : timeline_id,
1583 6 : index_part,
1584 6 : remote_metadata,
1585 6 : TimelineResources {
1586 6 : remote_client,
1587 6 : timeline_get_throttle: self.timeline_get_throttle.clone(),
1588 6 : l0_flush_global_state: self.l0_flush_global_state.clone(),
1589 6 : },
1590 6 : ctx,
1591 6 : )
1592 5 : .await
1593 6 : .with_context(|| {
1594 0 : format!(
1595 0 : "failed to load remote timeline {} for tenant {}",
1596 0 : timeline_id, self.tenant_shard_id
1597 0 : )
1598 6 : })?;
1599 : }
1600 :
1601 : // Walk through deleted timelines, resume deletion
1602 192 : for (timeline_id, index_part, remote_timeline_client) in timelines_to_resume_deletions {
1603 0 : remote_timeline_client
1604 0 : .init_upload_queue_stopped_to_continue_deletion(&index_part)
1605 0 : .context("init queue stopped")
1606 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1607 :
1608 0 : DeleteTimelineFlow::resume_deletion(
1609 0 : Arc::clone(self),
1610 0 : timeline_id,
1611 0 : &index_part.metadata,
1612 0 : remote_timeline_client,
1613 0 : )
1614 0 : .instrument(tracing::info_span!("timeline_delete", %timeline_id))
1615 0 : .await
1616 0 : .context("resume_deletion")
1617 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1618 : }
1619 192 : let needs_manifest_upload =
1620 192 : offloaded_timelines_list.len() != preload.tenant_manifest.offloaded_timelines.len();
1621 192 : {
1622 192 : let mut offloaded_timelines_accessor = self.timelines_offloaded.lock().unwrap();
1623 192 : offloaded_timelines_accessor.extend(offloaded_timelines_list.into_iter());
1624 192 : }
1625 192 : if needs_manifest_upload {
1626 0 : self.store_tenant_manifest().await?;
1627 192 : }
1628 :
1629 : // The local filesystem contents are a cache of what's in the remote IndexPart;
1630 : // IndexPart is the source of truth.
1631 192 : self.clean_up_timelines(&existent_timelines)?;
1632 :
1633 192 : self.gc_block.set_scanned(gc_blocks);
1634 192 :
1635 192 : fail::fail_point!("attach-before-activate", |_| {
1636 0 : anyhow::bail!("attach-before-activate");
1637 192 : });
1638 192 : failpoint_support::sleep_millis_async!("attach-before-activate-sleep", &self.cancel);
1639 :
1640 192 : info!("Done");
1641 :
1642 192 : Ok(())
1643 192 : }
1644 :
1645 : /// Check for any local timeline directories that are temporary, or do not correspond to a
1646 : /// timeline that still exists: this can happen if we crashed during a deletion/creation, or
1647 : /// if a timeline was deleted while the tenant was attached to a different pageserver.
1648 192 : fn clean_up_timelines(&self, existent_timelines: &HashSet<TimelineId>) -> anyhow::Result<()> {
1649 192 : let timelines_dir = self.conf.timelines_path(&self.tenant_shard_id);
1650 :
1651 192 : let entries = match timelines_dir.read_dir_utf8() {
1652 192 : Ok(d) => d,
1653 0 : Err(e) => {
1654 0 : if e.kind() == std::io::ErrorKind::NotFound {
1655 0 : return Ok(());
1656 : } else {
1657 0 : return Err(e).context("list timelines directory for tenant");
1658 : }
1659 : }
1660 : };
1661 :
1662 200 : for entry in entries {
1663 8 : let entry = entry.context("read timeline dir entry")?;
1664 8 : let entry_path = entry.path();
1665 :
1666 8 : let purge = if crate::is_temporary(entry_path)
1667 : // TODO: remove uninit mark code (https://github.com/neondatabase/neon/issues/5718)
1668 8 : || is_uninit_mark(entry_path)
1669 8 : || crate::is_delete_mark(entry_path)
1670 : {
1671 0 : true
1672 : } else {
1673 8 : match TimelineId::try_from(entry_path.file_name()) {
1674 8 : Ok(i) => {
1675 8 : // Purge if the timeline ID does not exist in remote storage: remote storage is the authority.
1676 8 : !existent_timelines.contains(&i)
1677 : }
1678 0 : Err(e) => {
1679 0 : tracing::warn!(
1680 0 : "Unparseable directory in timelines directory: {entry_path}, ignoring ({e})"
1681 : );
1682 : // Do not purge junk: if we don't recognize it, be cautious and leave it for a human.
1683 0 : false
1684 : }
1685 : }
1686 : };
1687 :
1688 8 : if purge {
1689 2 : tracing::info!("Purging stale timeline dentry {entry_path}");
1690 2 : if let Err(e) = match entry.file_type() {
1691 2 : Ok(t) => if t.is_dir() {
1692 2 : std::fs::remove_dir_all(entry_path)
1693 : } else {
1694 0 : std::fs::remove_file(entry_path)
1695 : }
1696 2 : .or_else(fs_ext::ignore_not_found),
1697 0 : Err(e) => Err(e),
1698 : } {
1699 0 : tracing::warn!("Failed to purge stale timeline dentry {entry_path}: {e}");
1700 2 : }
1701 6 : }
1702 : }
1703 :
1704 192 : Ok(())
1705 192 : }
1706 :
1707 : /// Get sum of all remote timelines sizes
1708 : ///
1709 : /// This function relies on the index_part instead of listing the remote storage
1710 0 : pub fn remote_size(&self) -> u64 {
1711 0 : let mut size = 0;
1712 :
1713 0 : for timeline in self.list_timelines() {
1714 0 : size += timeline.remote_client.get_remote_physical_size();
1715 0 : }
1716 :
1717 0 : size
1718 0 : }
1719 :
1720 6 : #[instrument(skip_all, fields(timeline_id=%timeline_id))]
1721 : async fn load_remote_timeline(
1722 : &self,
1723 : timeline_id: TimelineId,
1724 : index_part: IndexPart,
1725 : remote_metadata: TimelineMetadata,
1726 : resources: TimelineResources,
1727 : ctx: &RequestContext,
1728 : ) -> anyhow::Result<()> {
1729 : span::debug_assert_current_span_has_tenant_id();
1730 :
1731 : info!("downloading index file for timeline {}", timeline_id);
1732 : tokio::fs::create_dir_all(self.conf.timeline_path(&self.tenant_shard_id, &timeline_id))
1733 : .await
1734 : .context("Failed to create new timeline directory")?;
1735 :
1736 : let ancestor = if let Some(ancestor_id) = remote_metadata.ancestor_timeline() {
1737 : let timelines = self.timelines.lock().unwrap();
1738 : Some(Arc::clone(timelines.get(&ancestor_id).ok_or_else(
1739 0 : || {
1740 0 : anyhow::anyhow!(
1741 0 : "cannot find ancestor timeline {ancestor_id} for timeline {timeline_id}"
1742 0 : )
1743 0 : },
1744 : )?))
1745 : } else {
1746 : None
1747 : };
1748 :
1749 : self.timeline_init_and_sync(
1750 : timeline_id,
1751 : resources,
1752 : index_part,
1753 : remote_metadata,
1754 : ancestor,
1755 : ctx,
1756 : )
1757 : .await
1758 : }
1759 :
1760 192 : async fn load_timelines_metadata(
1761 192 : self: &Arc<Tenant>,
1762 192 : timeline_ids: HashSet<TimelineId>,
1763 192 : remote_storage: &GenericRemoteStorage,
1764 192 : cancel: CancellationToken,
1765 192 : ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
1766 192 : let mut part_downloads = JoinSet::new();
1767 198 : for timeline_id in timeline_ids {
1768 6 : let cancel_clone = cancel.clone();
1769 6 : part_downloads.spawn(
1770 6 : self.load_timeline_metadata(timeline_id, remote_storage.clone(), cancel_clone)
1771 6 : .instrument(info_span!("download_index_part", %timeline_id)),
1772 : );
1773 : }
1774 :
1775 192 : let mut timeline_preloads: HashMap<TimelineId, TimelinePreload> = HashMap::new();
1776 :
1777 : loop {
1778 198 : tokio::select!(
1779 198 : next = part_downloads.join_next() => {
1780 198 : match next {
1781 6 : Some(result) => {
1782 6 : let preload = result.context("join preload task")?;
1783 6 : timeline_preloads.insert(preload.timeline_id, preload);
1784 : },
1785 : None => {
1786 192 : break;
1787 : }
1788 : }
1789 : },
1790 198 : _ = cancel.cancelled() => {
1791 0 : anyhow::bail!("Cancelled while waiting for remote index download")
1792 : }
1793 : )
1794 : }
1795 :
1796 192 : Ok(timeline_preloads)
1797 192 : }
1798 :
1799 6 : fn build_timeline_client(
1800 6 : &self,
1801 6 : timeline_id: TimelineId,
1802 6 : remote_storage: GenericRemoteStorage,
1803 6 : ) -> RemoteTimelineClient {
1804 6 : RemoteTimelineClient::new(
1805 6 : remote_storage.clone(),
1806 6 : self.deletion_queue_client.clone(),
1807 6 : self.conf,
1808 6 : self.tenant_shard_id,
1809 6 : timeline_id,
1810 6 : self.generation,
1811 6 : &self.tenant_conf.load().location,
1812 6 : )
1813 6 : }
1814 :
1815 6 : fn load_timeline_metadata(
1816 6 : self: &Arc<Tenant>,
1817 6 : timeline_id: TimelineId,
1818 6 : remote_storage: GenericRemoteStorage,
1819 6 : cancel: CancellationToken,
1820 6 : ) -> impl Future<Output = TimelinePreload> {
1821 6 : let client = self.build_timeline_client(timeline_id, remote_storage);
1822 6 : async move {
1823 6 : debug_assert_current_span_has_tenant_and_timeline_id();
1824 6 : debug!("starting index part download");
1825 :
1826 15 : let index_part = client.download_index_file(&cancel).await;
1827 :
1828 6 : debug!("finished index part download");
1829 :
1830 6 : TimelinePreload {
1831 6 : client,
1832 6 : timeline_id,
1833 6 : index_part,
1834 6 : }
1835 6 : }
1836 6 : }
1837 :
1838 0 : fn check_to_be_archived_has_no_unarchived_children(
1839 0 : timeline_id: TimelineId,
1840 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
1841 0 : ) -> Result<(), TimelineArchivalError> {
1842 0 : let children: Vec<TimelineId> = timelines
1843 0 : .iter()
1844 0 : .filter_map(|(id, entry)| {
1845 0 : if entry.get_ancestor_timeline_id() != Some(timeline_id) {
1846 0 : return None;
1847 0 : }
1848 0 : if entry.is_archived() == Some(true) {
1849 0 : return None;
1850 0 : }
1851 0 : Some(*id)
1852 0 : })
1853 0 : .collect();
1854 0 :
1855 0 : if !children.is_empty() {
1856 0 : return Err(TimelineArchivalError::HasUnarchivedChildren(children));
1857 0 : }
1858 0 : Ok(())
1859 0 : }
1860 :
1861 0 : fn check_ancestor_of_to_be_unarchived_is_not_archived(
1862 0 : ancestor_timeline_id: TimelineId,
1863 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
1864 0 : offloaded_timelines: &std::sync::MutexGuard<
1865 0 : '_,
1866 0 : HashMap<TimelineId, Arc<OffloadedTimeline>>,
1867 0 : >,
1868 0 : ) -> Result<(), TimelineArchivalError> {
1869 0 : let has_archived_parent =
1870 0 : if let Some(ancestor_timeline) = timelines.get(&ancestor_timeline_id) {
1871 0 : ancestor_timeline.is_archived() == Some(true)
1872 0 : } else if offloaded_timelines.contains_key(&ancestor_timeline_id) {
1873 0 : true
1874 : } else {
1875 0 : error!("ancestor timeline {ancestor_timeline_id} not found");
1876 0 : if cfg!(debug_assertions) {
1877 0 : panic!("ancestor timeline {ancestor_timeline_id} not found");
1878 0 : }
1879 0 : return Err(TimelineArchivalError::NotFound);
1880 : };
1881 0 : if has_archived_parent {
1882 0 : return Err(TimelineArchivalError::HasArchivedParent(
1883 0 : ancestor_timeline_id,
1884 0 : ));
1885 0 : }
1886 0 : Ok(())
1887 0 : }
1888 :
1889 0 : fn check_to_be_unarchived_timeline_has_no_archived_parent(
1890 0 : timeline: &Arc<Timeline>,
1891 0 : ) -> Result<(), TimelineArchivalError> {
1892 0 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
1893 0 : if ancestor_timeline.is_archived() == Some(true) {
1894 0 : return Err(TimelineArchivalError::HasArchivedParent(
1895 0 : ancestor_timeline.timeline_id,
1896 0 : ));
1897 0 : }
1898 0 : }
1899 0 : Ok(())
1900 0 : }
1901 :
1902 : /// Loads the specified (offloaded) timeline from S3 and attaches it as a loaded timeline
1903 : ///
1904 : /// Counterpart to [`offload_timeline`].
1905 0 : async fn unoffload_timeline(
1906 0 : self: &Arc<Self>,
1907 0 : timeline_id: TimelineId,
1908 0 : broker_client: storage_broker::BrokerClientChannel,
1909 0 : ctx: RequestContext,
1910 0 : ) -> Result<Arc<Timeline>, TimelineArchivalError> {
1911 0 : info!("unoffloading timeline");
1912 :
1913 : // We activate the timeline below manually, so this must be called on an active timeline.
1914 : // We expect callers of this function to ensure this.
1915 0 : match self.current_state() {
1916 : TenantState::Activating { .. }
1917 : | TenantState::Attaching
1918 : | TenantState::Broken { .. } => {
1919 0 : panic!("Timeline expected to be active")
1920 : }
1921 0 : TenantState::Stopping { .. } => return Err(TimelineArchivalError::Cancelled),
1922 0 : TenantState::Active => {}
1923 0 : }
1924 0 : let cancel = self.cancel.clone();
1925 0 :
1926 0 : // Protect against concurrent attempts to use this TimelineId
1927 0 : // We don't care much about idempotency, as it's ensured a layer above.
1928 0 : let allow_offloaded = true;
1929 0 : let _create_guard = self
1930 0 : .create_timeline_create_guard(
1931 0 : timeline_id,
1932 0 : CreateTimelineIdempotency::FailWithConflict,
1933 0 : allow_offloaded,
1934 0 : )
1935 0 : .map_err(|err| match err {
1936 0 : TimelineExclusionError::AlreadyCreating => TimelineArchivalError::AlreadyInProgress,
1937 : TimelineExclusionError::AlreadyExists { .. } => {
1938 0 : TimelineArchivalError::Other(anyhow::anyhow!("Timeline already exists"))
1939 : }
1940 0 : TimelineExclusionError::Other(e) => TimelineArchivalError::Other(e),
1941 0 : })?;
1942 :
1943 0 : let timeline_preload = self
1944 0 : .load_timeline_metadata(timeline_id, self.remote_storage.clone(), cancel.clone())
1945 0 : .await;
1946 :
1947 0 : let index_part = match timeline_preload.index_part {
1948 0 : Ok(index_part) => {
1949 0 : debug!("remote index part exists for timeline {timeline_id}");
1950 0 : index_part
1951 : }
1952 : Err(DownloadError::NotFound) => {
1953 0 : error!(%timeline_id, "index_part not found on remote");
1954 0 : return Err(TimelineArchivalError::NotFound);
1955 : }
1956 0 : Err(DownloadError::Cancelled) => return Err(TimelineArchivalError::Cancelled),
1957 0 : Err(e) => {
1958 0 : // Some (possibly ephemeral) error happened during index_part download.
1959 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1960 0 : return Err(TimelineArchivalError::Other(
1961 0 : anyhow::Error::new(e).context("downloading index_part from remote storage"),
1962 0 : ));
1963 : }
1964 : };
1965 0 : let index_part = match index_part {
1966 0 : MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
1967 0 : MaybeDeletedIndexPart::Deleted(_index_part) => {
1968 0 : info!("timeline is deleted according to index_part.json");
1969 0 : return Err(TimelineArchivalError::NotFound);
1970 : }
1971 : };
1972 0 : let remote_metadata = index_part.metadata.clone();
1973 0 : let timeline_resources = self.build_timeline_resources(timeline_id);
1974 0 : self.load_remote_timeline(
1975 0 : timeline_id,
1976 0 : index_part,
1977 0 : remote_metadata,
1978 0 : timeline_resources,
1979 0 : &ctx,
1980 0 : )
1981 0 : .await
1982 0 : .with_context(|| {
1983 0 : format!(
1984 0 : "failed to load remote timeline {} for tenant {}",
1985 0 : timeline_id, self.tenant_shard_id
1986 0 : )
1987 0 : })
1988 0 : .map_err(TimelineArchivalError::Other)?;
1989 :
1990 0 : let timeline = {
1991 0 : let timelines = self.timelines.lock().unwrap();
1992 0 : let Some(timeline) = timelines.get(&timeline_id) else {
1993 0 : warn!("timeline not available directly after attach");
1994 : // This is not a panic because no locks are held between `load_remote_timeline`
1995 : // which puts the timeline into timelines, and our look into the timeline map.
1996 0 : return Err(TimelineArchivalError::Other(anyhow::anyhow!(
1997 0 : "timeline not available directly after attach"
1998 0 : )));
1999 : };
2000 0 : let mut offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2001 0 : match offloaded_timelines.remove(&timeline_id) {
2002 0 : Some(offloaded) => {
2003 0 : offloaded.delete_from_ancestor_with_timelines(&timelines);
2004 0 : }
2005 0 : None => warn!("timeline already removed from offloaded timelines"),
2006 : }
2007 :
2008 0 : self.initialize_gc_info(&timelines, &offloaded_timelines, Some(timeline_id));
2009 0 :
2010 0 : Arc::clone(timeline)
2011 0 : };
2012 0 :
2013 0 : // Upload new list of offloaded timelines to S3
2014 0 : self.store_tenant_manifest().await?;
2015 :
2016 : // Activate the timeline (if it makes sense)
2017 0 : if !(timeline.is_broken() || timeline.is_stopping()) {
2018 0 : let background_jobs_can_start = None;
2019 0 : timeline.activate(
2020 0 : self.clone(),
2021 0 : broker_client.clone(),
2022 0 : background_jobs_can_start,
2023 0 : &ctx,
2024 0 : );
2025 0 : }
2026 :
2027 0 : info!("timeline unoffloading complete");
2028 0 : Ok(timeline)
2029 0 : }
2030 :
2031 0 : pub(crate) async fn apply_timeline_archival_config(
2032 0 : self: &Arc<Self>,
2033 0 : timeline_id: TimelineId,
2034 0 : new_state: TimelineArchivalState,
2035 0 : broker_client: storage_broker::BrokerClientChannel,
2036 0 : ctx: RequestContext,
2037 0 : ) -> Result<(), TimelineArchivalError> {
2038 0 : info!("setting timeline archival config");
2039 : // First part: figure out what is needed to do, and do validation
2040 0 : let timeline_or_unarchive_offloaded = 'outer: {
2041 0 : let timelines = self.timelines.lock().unwrap();
2042 :
2043 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2044 0 : let offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2045 0 : let Some(offloaded) = offloaded_timelines.get(&timeline_id) else {
2046 0 : return Err(TimelineArchivalError::NotFound);
2047 : };
2048 0 : if new_state == TimelineArchivalState::Archived {
2049 : // It's offloaded already, so nothing to do
2050 0 : return Ok(());
2051 0 : }
2052 0 : if let Some(ancestor_timeline_id) = offloaded.ancestor_timeline_id {
2053 0 : Self::check_ancestor_of_to_be_unarchived_is_not_archived(
2054 0 : ancestor_timeline_id,
2055 0 : &timelines,
2056 0 : &offloaded_timelines,
2057 0 : )?;
2058 0 : }
2059 0 : break 'outer None;
2060 : };
2061 :
2062 : // Do some validation. We release the timelines lock below, so there is potential
2063 : // for race conditions: these checks are more present to prevent misunderstandings of
2064 : // the API's capabilities, instead of serving as the sole way to defend their invariants.
2065 0 : match new_state {
2066 : TimelineArchivalState::Unarchived => {
2067 0 : Self::check_to_be_unarchived_timeline_has_no_archived_parent(timeline)?
2068 : }
2069 : TimelineArchivalState::Archived => {
2070 0 : Self::check_to_be_archived_has_no_unarchived_children(timeline_id, &timelines)?
2071 : }
2072 : }
2073 0 : Some(Arc::clone(timeline))
2074 : };
2075 :
2076 : // Second part: unoffload timeline (if needed)
2077 0 : let timeline = if let Some(timeline) = timeline_or_unarchive_offloaded {
2078 0 : timeline
2079 : } else {
2080 : // Turn offloaded timeline into a non-offloaded one
2081 0 : self.unoffload_timeline(timeline_id, broker_client, ctx)
2082 0 : .await?
2083 : };
2084 :
2085 : // Third part: upload new timeline archival state and block until it is present in S3
2086 0 : let upload_needed = match timeline
2087 0 : .remote_client
2088 0 : .schedule_index_upload_for_timeline_archival_state(new_state)
2089 : {
2090 0 : Ok(upload_needed) => upload_needed,
2091 0 : Err(e) => {
2092 0 : if timeline.cancel.is_cancelled() {
2093 0 : return Err(TimelineArchivalError::Cancelled);
2094 : } else {
2095 0 : return Err(TimelineArchivalError::Other(e));
2096 : }
2097 : }
2098 : };
2099 :
2100 0 : if upload_needed {
2101 0 : info!("Uploading new state");
2102 : const MAX_WAIT: Duration = Duration::from_secs(10);
2103 0 : let Ok(v) =
2104 0 : tokio::time::timeout(MAX_WAIT, timeline.remote_client.wait_completion()).await
2105 : else {
2106 0 : tracing::warn!("reached timeout for waiting on upload queue");
2107 0 : return Err(TimelineArchivalError::Timeout);
2108 : };
2109 0 : v.map_err(|e| match e {
2110 0 : WaitCompletionError::NotInitialized(e) => {
2111 0 : TimelineArchivalError::Other(anyhow::anyhow!(e))
2112 : }
2113 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2114 0 : TimelineArchivalError::Cancelled
2115 : }
2116 0 : })?;
2117 0 : }
2118 0 : Ok(())
2119 0 : }
2120 :
2121 2 : pub fn get_offloaded_timeline(
2122 2 : &self,
2123 2 : timeline_id: TimelineId,
2124 2 : ) -> Result<Arc<OffloadedTimeline>, GetTimelineError> {
2125 2 : self.timelines_offloaded
2126 2 : .lock()
2127 2 : .unwrap()
2128 2 : .get(&timeline_id)
2129 2 : .map(Arc::clone)
2130 2 : .ok_or(GetTimelineError::NotFound {
2131 2 : tenant_id: self.tenant_shard_id,
2132 2 : timeline_id,
2133 2 : })
2134 2 : }
2135 :
2136 4 : pub(crate) fn tenant_shard_id(&self) -> TenantShardId {
2137 4 : self.tenant_shard_id
2138 4 : }
2139 :
2140 : /// Get Timeline handle for given Neon timeline ID.
2141 : /// This function is idempotent. It doesn't change internal state in any way.
2142 222 : pub fn get_timeline(
2143 222 : &self,
2144 222 : timeline_id: TimelineId,
2145 222 : active_only: bool,
2146 222 : ) -> Result<Arc<Timeline>, GetTimelineError> {
2147 222 : let timelines_accessor = self.timelines.lock().unwrap();
2148 222 : let timeline = timelines_accessor
2149 222 : .get(&timeline_id)
2150 222 : .ok_or(GetTimelineError::NotFound {
2151 222 : tenant_id: self.tenant_shard_id,
2152 222 : timeline_id,
2153 222 : })?;
2154 :
2155 220 : if active_only && !timeline.is_active() {
2156 0 : Err(GetTimelineError::NotActive {
2157 0 : tenant_id: self.tenant_shard_id,
2158 0 : timeline_id,
2159 0 : state: timeline.current_state(),
2160 0 : })
2161 : } else {
2162 220 : Ok(Arc::clone(timeline))
2163 : }
2164 222 : }
2165 :
2166 : /// Lists timelines the tenant contains.
2167 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2168 0 : pub fn list_timelines(&self) -> Vec<Arc<Timeline>> {
2169 0 : self.timelines
2170 0 : .lock()
2171 0 : .unwrap()
2172 0 : .values()
2173 0 : .map(Arc::clone)
2174 0 : .collect()
2175 0 : }
2176 :
2177 : /// Lists timelines the tenant manages, including offloaded ones.
2178 : ///
2179 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2180 0 : pub fn list_timelines_and_offloaded(
2181 0 : &self,
2182 0 : ) -> (Vec<Arc<Timeline>>, Vec<Arc<OffloadedTimeline>>) {
2183 0 : let timelines = self
2184 0 : .timelines
2185 0 : .lock()
2186 0 : .unwrap()
2187 0 : .values()
2188 0 : .map(Arc::clone)
2189 0 : .collect();
2190 0 : let offloaded = self
2191 0 : .timelines_offloaded
2192 0 : .lock()
2193 0 : .unwrap()
2194 0 : .values()
2195 0 : .map(Arc::clone)
2196 0 : .collect();
2197 0 : (timelines, offloaded)
2198 0 : }
2199 :
2200 0 : pub fn list_timeline_ids(&self) -> Vec<TimelineId> {
2201 0 : self.timelines.lock().unwrap().keys().cloned().collect()
2202 0 : }
2203 :
2204 : /// This is used by tests & import-from-basebackup.
2205 : ///
2206 : /// The returned [`UninitializedTimeline`] contains no data nor metadata and it is in
2207 : /// a state that will fail [`Tenant::load_remote_timeline`] because `disk_consistent_lsn=Lsn(0)`.
2208 : ///
2209 : /// The caller is responsible for getting the timeline into a state that will be accepted
2210 : /// by [`Tenant::load_remote_timeline`] / [`Tenant::attach`].
2211 : /// Then they may call [`UninitializedTimeline::finish_creation`] to add the timeline
2212 : /// to the [`Tenant::timelines`].
2213 : ///
2214 : /// Tests should use `Tenant::create_test_timeline` to set up the minimum required metadata keys.
2215 184 : pub(crate) async fn create_empty_timeline(
2216 184 : &self,
2217 184 : new_timeline_id: TimelineId,
2218 184 : initdb_lsn: Lsn,
2219 184 : pg_version: u32,
2220 184 : _ctx: &RequestContext,
2221 184 : ) -> anyhow::Result<UninitializedTimeline> {
2222 184 : anyhow::ensure!(
2223 184 : self.is_active(),
2224 0 : "Cannot create empty timelines on inactive tenant"
2225 : );
2226 :
2227 : // Protect against concurrent attempts to use this TimelineId
2228 184 : let create_guard = match self
2229 184 : .start_creating_timeline(new_timeline_id, CreateTimelineIdempotency::FailWithConflict)
2230 179 : .await?
2231 : {
2232 182 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2233 : StartCreatingTimelineResult::Idempotent(_) => {
2234 0 : unreachable!("FailWithConflict implies we get an error instead")
2235 : }
2236 : };
2237 :
2238 182 : let new_metadata = TimelineMetadata::new(
2239 182 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2240 182 : // make it valid, before calling finish_creation()
2241 182 : Lsn(0),
2242 182 : None,
2243 182 : None,
2244 182 : Lsn(0),
2245 182 : initdb_lsn,
2246 182 : initdb_lsn,
2247 182 : pg_version,
2248 182 : );
2249 182 : self.prepare_new_timeline(
2250 182 : new_timeline_id,
2251 182 : &new_metadata,
2252 182 : create_guard,
2253 182 : initdb_lsn,
2254 182 : None,
2255 182 : )
2256 0 : .await
2257 184 : }
2258 :
2259 : /// Helper for unit tests to create an empty timeline.
2260 : ///
2261 : /// The timeline is has state value `Active` but its background loops are not running.
2262 : // This makes the various functions which anyhow::ensure! for Active state work in tests.
2263 : // Our current tests don't need the background loops.
2264 : #[cfg(test)]
2265 174 : pub async fn create_test_timeline(
2266 174 : &self,
2267 174 : new_timeline_id: TimelineId,
2268 174 : initdb_lsn: Lsn,
2269 174 : pg_version: u32,
2270 174 : ctx: &RequestContext,
2271 174 : ) -> anyhow::Result<Arc<Timeline>> {
2272 174 : let uninit_tl = self
2273 174 : .create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2274 171 : .await?;
2275 174 : let tline = uninit_tl.raw_timeline().expect("we just created it");
2276 174 : assert_eq!(tline.get_last_record_lsn(), Lsn(0));
2277 :
2278 : // Setup minimum keys required for the timeline to be usable.
2279 174 : let mut modification = tline.begin_modification(initdb_lsn);
2280 174 : modification
2281 174 : .init_empty_test_timeline()
2282 174 : .context("init_empty_test_timeline")?;
2283 174 : modification
2284 174 : .commit(ctx)
2285 168 : .await
2286 174 : .context("commit init_empty_test_timeline modification")?;
2287 :
2288 : // Flush to disk so that uninit_tl's check for valid disk_consistent_lsn passes.
2289 174 : tline.maybe_spawn_flush_loop();
2290 174 : tline.freeze_and_flush().await.context("freeze_and_flush")?;
2291 :
2292 : // Make sure the freeze_and_flush reaches remote storage.
2293 174 : tline.remote_client.wait_completion().await.unwrap();
2294 :
2295 174 : let tl = uninit_tl.finish_creation()?;
2296 : // The non-test code would call tl.activate() here.
2297 174 : tl.set_state(TimelineState::Active);
2298 174 : Ok(tl)
2299 174 : }
2300 :
2301 : /// Helper for unit tests to create a timeline with some pre-loaded states.
2302 : #[cfg(test)]
2303 : #[allow(clippy::too_many_arguments)]
2304 32 : pub async fn create_test_timeline_with_layers(
2305 32 : &self,
2306 32 : new_timeline_id: TimelineId,
2307 32 : initdb_lsn: Lsn,
2308 32 : pg_version: u32,
2309 32 : ctx: &RequestContext,
2310 32 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
2311 32 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
2312 32 : end_lsn: Lsn,
2313 32 : ) -> anyhow::Result<Arc<Timeline>> {
2314 : use checks::check_valid_layermap;
2315 : use itertools::Itertools;
2316 :
2317 32 : let tline = self
2318 32 : .create_test_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2319 95 : .await?;
2320 32 : tline.force_advance_lsn(end_lsn);
2321 100 : for deltas in delta_layer_desc {
2322 68 : tline
2323 68 : .force_create_delta_layer(deltas, Some(initdb_lsn), ctx)
2324 204 : .await?;
2325 : }
2326 80 : for (lsn, images) in image_layer_desc {
2327 48 : tline
2328 48 : .force_create_image_layer(lsn, images, Some(initdb_lsn), ctx)
2329 294 : .await?;
2330 : }
2331 32 : let layer_names = tline
2332 32 : .layers
2333 32 : .read()
2334 0 : .await
2335 32 : .layer_map()
2336 32 : .unwrap()
2337 32 : .iter_historic_layers()
2338 148 : .map(|layer| layer.layer_name())
2339 32 : .collect_vec();
2340 32 : if let Some(err) = check_valid_layermap(&layer_names) {
2341 0 : bail!("invalid layermap: {err}");
2342 32 : }
2343 32 : Ok(tline)
2344 32 : }
2345 :
2346 : /// Create a new timeline.
2347 : ///
2348 : /// Returns the new timeline ID and reference to its Timeline object.
2349 : ///
2350 : /// If the caller specified the timeline ID to use (`new_timeline_id`), and timeline with
2351 : /// the same timeline ID already exists, returns CreateTimelineError::AlreadyExists.
2352 : #[allow(clippy::too_many_arguments)]
2353 0 : pub(crate) async fn create_timeline(
2354 0 : self: &Arc<Tenant>,
2355 0 : params: CreateTimelineParams,
2356 0 : broker_client: storage_broker::BrokerClientChannel,
2357 0 : ctx: &RequestContext,
2358 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
2359 0 : if !self.is_active() {
2360 0 : if matches!(self.current_state(), TenantState::Stopping { .. }) {
2361 0 : return Err(CreateTimelineError::ShuttingDown);
2362 : } else {
2363 0 : return Err(CreateTimelineError::Other(anyhow::anyhow!(
2364 0 : "Cannot create timelines on inactive tenant"
2365 0 : )));
2366 : }
2367 0 : }
2368 :
2369 0 : let _gate = self
2370 0 : .gate
2371 0 : .enter()
2372 0 : .map_err(|_| CreateTimelineError::ShuttingDown)?;
2373 :
2374 0 : let result: CreateTimelineResult = match params {
2375 : CreateTimelineParams::Bootstrap(CreateTimelineParamsBootstrap {
2376 0 : new_timeline_id,
2377 0 : existing_initdb_timeline_id,
2378 0 : pg_version,
2379 0 : }) => {
2380 0 : self.bootstrap_timeline(
2381 0 : new_timeline_id,
2382 0 : pg_version,
2383 0 : existing_initdb_timeline_id,
2384 0 : ctx,
2385 0 : )
2386 0 : .await?
2387 : }
2388 : CreateTimelineParams::Branch(CreateTimelineParamsBranch {
2389 0 : new_timeline_id,
2390 0 : ancestor_timeline_id,
2391 0 : mut ancestor_start_lsn,
2392 : }) => {
2393 0 : let ancestor_timeline = self
2394 0 : .get_timeline(ancestor_timeline_id, false)
2395 0 : .context("Cannot branch off the timeline that's not present in pageserver")?;
2396 :
2397 : // instead of waiting around, just deny the request because ancestor is not yet
2398 : // ready for other purposes either.
2399 0 : if !ancestor_timeline.is_active() {
2400 0 : return Err(CreateTimelineError::AncestorNotActive);
2401 0 : }
2402 0 :
2403 0 : if ancestor_timeline.is_archived() == Some(true) {
2404 0 : info!("tried to branch archived timeline");
2405 0 : return Err(CreateTimelineError::AncestorArchived);
2406 0 : }
2407 :
2408 0 : if let Some(lsn) = ancestor_start_lsn.as_mut() {
2409 0 : *lsn = lsn.align();
2410 0 :
2411 0 : let ancestor_ancestor_lsn = ancestor_timeline.get_ancestor_lsn();
2412 0 : if ancestor_ancestor_lsn > *lsn {
2413 : // can we safely just branch from the ancestor instead?
2414 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
2415 0 : "invalid start lsn {} for ancestor timeline {}: less than timeline ancestor lsn {}",
2416 0 : lsn,
2417 0 : ancestor_timeline_id,
2418 0 : ancestor_ancestor_lsn,
2419 0 : )));
2420 0 : }
2421 0 :
2422 0 : // Wait for the WAL to arrive and be processed on the parent branch up
2423 0 : // to the requested branch point. The repository code itself doesn't
2424 0 : // require it, but if we start to receive WAL on the new timeline,
2425 0 : // decoding the new WAL might need to look up previous pages, relation
2426 0 : // sizes etc. and that would get confused if the previous page versions
2427 0 : // are not in the repository yet.
2428 0 : ancestor_timeline
2429 0 : .wait_lsn(*lsn, timeline::WaitLsnWaiter::Tenant, ctx)
2430 0 : .await
2431 0 : .map_err(|e| match e {
2432 0 : e @ (WaitLsnError::Timeout(_) | WaitLsnError::BadState { .. }) => {
2433 0 : CreateTimelineError::AncestorLsn(anyhow::anyhow!(e))
2434 : }
2435 0 : WaitLsnError::Shutdown => CreateTimelineError::ShuttingDown,
2436 0 : })?;
2437 0 : }
2438 :
2439 0 : self.branch_timeline(&ancestor_timeline, new_timeline_id, ancestor_start_lsn, ctx)
2440 0 : .await?
2441 : }
2442 : };
2443 :
2444 : // At this point we have dropped our guard on [`Self::timelines_creating`], and
2445 : // the timeline is visible in [`Self::timelines`], but it is _not_ durable yet. We must
2446 : // not send a success to the caller until it is. The same applies to idempotent retries.
2447 : //
2448 : // TODO: the timeline is already visible in [`Self::timelines`]; a caller could incorrectly
2449 : // assume that, because they can see the timeline via API, that the creation is done and
2450 : // that it is durable. Ideally, we would keep the timeline hidden (in [`Self::timelines_creating`])
2451 : // until it is durable, e.g., by extending the time we hold the creation guard. This also
2452 : // interacts with UninitializedTimeline and is generally a bit tricky.
2453 : //
2454 : // To re-emphasize: the only correct way to create a timeline is to repeat calling the
2455 : // creation API until it returns success. Only then is durability guaranteed.
2456 0 : info!(creation_result=%result.discriminant(), "waiting for timeline to be durable");
2457 0 : result
2458 0 : .timeline()
2459 0 : .remote_client
2460 0 : .wait_completion()
2461 0 : .await
2462 0 : .map_err(|e| match e {
2463 : WaitCompletionError::NotInitialized(
2464 0 : e, // If the queue is already stopped, it's a shutdown error.
2465 0 : ) if e.is_stopping() => CreateTimelineError::ShuttingDown,
2466 0 : e => CreateTimelineError::Other(e.into()),
2467 0 : })
2468 0 : .context("wait for timeline initial uploads to complete")?;
2469 :
2470 : // The creating task is responsible for activating the timeline.
2471 : // We do this after `wait_completion()` so that we don't spin up tasks that start
2472 : // doing stuff before the IndexPart is durable in S3, which is done by the previous section.
2473 0 : let activated_timeline = match result {
2474 0 : CreateTimelineResult::Created(timeline) => {
2475 0 : timeline.activate(self.clone(), broker_client, None, ctx);
2476 0 : timeline
2477 : }
2478 0 : CreateTimelineResult::Idempotent(timeline) => {
2479 0 : info!(
2480 0 : "request was deemed idempotent, activation will be done by the creating task"
2481 : );
2482 0 : timeline
2483 : }
2484 : };
2485 :
2486 0 : Ok(activated_timeline)
2487 0 : }
2488 :
2489 0 : pub(crate) async fn delete_timeline(
2490 0 : self: Arc<Self>,
2491 0 : timeline_id: TimelineId,
2492 0 : ) -> Result<(), DeleteTimelineError> {
2493 0 : DeleteTimelineFlow::run(&self, timeline_id).await?;
2494 :
2495 0 : Ok(())
2496 0 : }
2497 :
2498 : /// perform one garbage collection iteration, removing old data files from disk.
2499 : /// this function is periodically called by gc task.
2500 : /// also it can be explicitly requested through page server api 'do_gc' command.
2501 : ///
2502 : /// `target_timeline_id` specifies the timeline to GC, or None for all.
2503 : ///
2504 : /// The `horizon` an `pitr` parameters determine how much WAL history needs to be retained.
2505 : /// Also known as the retention period, or the GC cutoff point. `horizon` specifies
2506 : /// the amount of history, as LSN difference from current latest LSN on each timeline.
2507 : /// `pitr` specifies the same as a time difference from the current time. The effective
2508 : /// GC cutoff point is determined conservatively by either `horizon` and `pitr`, whichever
2509 : /// requires more history to be retained.
2510 : //
2511 754 : pub(crate) async fn gc_iteration(
2512 754 : &self,
2513 754 : target_timeline_id: Option<TimelineId>,
2514 754 : horizon: u64,
2515 754 : pitr: Duration,
2516 754 : cancel: &CancellationToken,
2517 754 : ctx: &RequestContext,
2518 754 : ) -> Result<GcResult, GcError> {
2519 754 : // Don't start doing work during shutdown
2520 754 : if let TenantState::Stopping { .. } = self.current_state() {
2521 0 : return Ok(GcResult::default());
2522 754 : }
2523 754 :
2524 754 : // there is a global allowed_error for this
2525 754 : if !self.is_active() {
2526 0 : return Err(GcError::NotActive);
2527 754 : }
2528 754 :
2529 754 : {
2530 754 : let conf = self.tenant_conf.load();
2531 754 :
2532 754 : // If we may not delete layers, then simply skip GC. Even though a tenant
2533 754 : // in AttachedMulti state could do GC and just enqueue the blocked deletions,
2534 754 : // the only advantage to doing it is to perhaps shrink the LayerMap metadata
2535 754 : // a bit sooner than we would achieve by waiting for AttachedSingle status.
2536 754 : if !conf.location.may_delete_layers_hint() {
2537 0 : info!("Skipping GC in location state {:?}", conf.location);
2538 0 : return Ok(GcResult::default());
2539 754 : }
2540 754 :
2541 754 : if conf.is_gc_blocked_by_lsn_lease_deadline() {
2542 750 : info!("Skipping GC because lsn lease deadline is not reached");
2543 750 : return Ok(GcResult::default());
2544 4 : }
2545 : }
2546 :
2547 4 : let _guard = match self.gc_block.start().await {
2548 4 : Ok(guard) => guard,
2549 0 : Err(reasons) => {
2550 0 : info!("Skipping GC: {reasons}");
2551 0 : return Ok(GcResult::default());
2552 : }
2553 : };
2554 :
2555 4 : self.gc_iteration_internal(target_timeline_id, horizon, pitr, cancel, ctx)
2556 4 : .await
2557 754 : }
2558 :
2559 : /// Perform one compaction iteration.
2560 : /// This function is periodically called by compactor task.
2561 : /// Also it can be explicitly requested per timeline through page server
2562 : /// api's 'compact' command.
2563 : ///
2564 : /// Returns whether we have pending compaction task.
2565 0 : async fn compaction_iteration(
2566 0 : self: &Arc<Self>,
2567 0 : cancel: &CancellationToken,
2568 0 : ctx: &RequestContext,
2569 0 : ) -> Result<bool, timeline::CompactionError> {
2570 0 : // Don't start doing work during shutdown, or when broken, we do not need those in the logs
2571 0 : if !self.is_active() {
2572 0 : return Ok(false);
2573 0 : }
2574 0 :
2575 0 : {
2576 0 : let conf = self.tenant_conf.load();
2577 0 :
2578 0 : // Note that compaction usually requires deletions, but we don't respect
2579 0 : // may_delete_layers_hint here: that is because tenants in AttachedMulti
2580 0 : // should proceed with compaction even if they can't do deletion, to avoid
2581 0 : // accumulating dangerously deep stacks of L0 layers. Deletions will be
2582 0 : // enqueued inside RemoteTimelineClient, and executed layer if/when we transition
2583 0 : // to AttachedSingle state.
2584 0 : if !conf.location.may_upload_layers_hint() {
2585 0 : info!("Skipping compaction in location state {:?}", conf.location);
2586 0 : return Ok(false);
2587 0 : }
2588 0 : }
2589 0 :
2590 0 : // Scan through the hashmap and collect a list of all the timelines,
2591 0 : // while holding the lock. Then drop the lock and actually perform the
2592 0 : // compactions. We don't want to block everything else while the
2593 0 : // compaction runs.
2594 0 : let timelines_to_compact_or_offload;
2595 0 : {
2596 0 : let timelines = self.timelines.lock().unwrap();
2597 0 : timelines_to_compact_or_offload = timelines
2598 0 : .iter()
2599 0 : .filter_map(|(timeline_id, timeline)| {
2600 0 : let (is_active, (can_offload, _)) =
2601 0 : (timeline.is_active(), timeline.can_offload());
2602 0 : let has_no_unoffloaded_children = {
2603 0 : !timelines
2604 0 : .iter()
2605 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(*timeline_id))
2606 : };
2607 0 : let config_allows_offload = self.conf.timeline_offloading
2608 0 : || self
2609 0 : .tenant_conf
2610 0 : .load()
2611 0 : .tenant_conf
2612 0 : .timeline_offloading
2613 0 : .unwrap_or_default();
2614 0 : let can_offload =
2615 0 : can_offload && has_no_unoffloaded_children && config_allows_offload;
2616 0 : if (is_active, can_offload) == (false, false) {
2617 0 : None
2618 : } else {
2619 0 : Some((*timeline_id, timeline.clone(), (is_active, can_offload)))
2620 : }
2621 0 : })
2622 0 : .collect::<Vec<_>>();
2623 0 : drop(timelines);
2624 0 : }
2625 0 :
2626 0 : // Before doing any I/O work, check our circuit breaker
2627 0 : if self.compaction_circuit_breaker.lock().unwrap().is_broken() {
2628 0 : info!("Skipping compaction due to previous failures");
2629 0 : return Ok(false);
2630 0 : }
2631 0 :
2632 0 : let mut has_pending_task = false;
2633 :
2634 0 : for (timeline_id, timeline, (can_compact, can_offload)) in &timelines_to_compact_or_offload
2635 : {
2636 0 : let pending_task_left = if *can_compact {
2637 : Some(
2638 0 : timeline
2639 0 : .compact(cancel, EnumSet::empty(), ctx)
2640 0 : .instrument(info_span!("compact_timeline", %timeline_id))
2641 0 : .await
2642 0 : .inspect_err(|e| match e {
2643 0 : timeline::CompactionError::ShuttingDown => (),
2644 0 : timeline::CompactionError::Offload(_) => {
2645 0 : // Failures to offload timelines do not trip the circuit breaker, because
2646 0 : // they do not do lots of writes the way compaction itself does: it is cheap
2647 0 : // to retry, and it would be bad to stop all compaction because of an issue with offloading.
2648 0 : }
2649 0 : timeline::CompactionError::Other(e) => {
2650 0 : self.compaction_circuit_breaker
2651 0 : .lock()
2652 0 : .unwrap()
2653 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, e);
2654 0 : }
2655 0 : })?,
2656 : )
2657 : } else {
2658 0 : None
2659 : };
2660 0 : has_pending_task |= pending_task_left.unwrap_or(false);
2661 0 : if pending_task_left == Some(false) && *can_offload {
2662 0 : offload_timeline(self, timeline)
2663 0 : .instrument(info_span!("offload_timeline", %timeline_id))
2664 0 : .await?;
2665 0 : }
2666 : }
2667 :
2668 0 : self.compaction_circuit_breaker
2669 0 : .lock()
2670 0 : .unwrap()
2671 0 : .success(&CIRCUIT_BREAKERS_UNBROKEN);
2672 0 :
2673 0 : Ok(has_pending_task)
2674 0 : }
2675 :
2676 : // Call through to all timelines to freeze ephemeral layers if needed. Usually
2677 : // this happens during ingest: this background housekeeping is for freezing layers
2678 : // that are open but haven't been written to for some time.
2679 0 : async fn ingest_housekeeping(&self) {
2680 0 : // Scan through the hashmap and collect a list of all the timelines,
2681 0 : // while holding the lock. Then drop the lock and actually perform the
2682 0 : // compactions. We don't want to block everything else while the
2683 0 : // compaction runs.
2684 0 : let timelines = {
2685 0 : self.timelines
2686 0 : .lock()
2687 0 : .unwrap()
2688 0 : .values()
2689 0 : .filter_map(|timeline| {
2690 0 : if timeline.is_active() {
2691 0 : Some(timeline.clone())
2692 : } else {
2693 0 : None
2694 : }
2695 0 : })
2696 0 : .collect::<Vec<_>>()
2697 : };
2698 :
2699 0 : for timeline in &timelines {
2700 0 : timeline.maybe_freeze_ephemeral_layer().await;
2701 : }
2702 0 : }
2703 :
2704 0 : pub fn timeline_has_no_attached_children(&self, timeline_id: TimelineId) -> bool {
2705 0 : let timelines = self.timelines.lock().unwrap();
2706 0 : !timelines
2707 0 : .iter()
2708 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(timeline_id))
2709 0 : }
2710 :
2711 1702 : pub fn current_state(&self) -> TenantState {
2712 1702 : self.state.borrow().clone()
2713 1702 : }
2714 :
2715 942 : pub fn is_active(&self) -> bool {
2716 942 : self.current_state() == TenantState::Active
2717 942 : }
2718 :
2719 0 : pub fn generation(&self) -> Generation {
2720 0 : self.generation
2721 0 : }
2722 :
2723 0 : pub(crate) fn wal_redo_manager_status(&self) -> Option<WalRedoManagerStatus> {
2724 0 : self.walredo_mgr.as_ref().and_then(|mgr| mgr.status())
2725 0 : }
2726 :
2727 : /// Changes tenant status to active, unless shutdown was already requested.
2728 : ///
2729 : /// `background_jobs_can_start` is an optional barrier set to a value during pageserver startup
2730 : /// to delay background jobs. Background jobs can be started right away when None is given.
2731 0 : fn activate(
2732 0 : self: &Arc<Self>,
2733 0 : broker_client: BrokerClientChannel,
2734 0 : background_jobs_can_start: Option<&completion::Barrier>,
2735 0 : ctx: &RequestContext,
2736 0 : ) {
2737 0 : span::debug_assert_current_span_has_tenant_id();
2738 0 :
2739 0 : let mut activating = false;
2740 0 : self.state.send_modify(|current_state| {
2741 : use pageserver_api::models::ActivatingFrom;
2742 0 : match &*current_state {
2743 : TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
2744 0 : panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {state:?}", state = current_state);
2745 : }
2746 0 : TenantState::Attaching => {
2747 0 : *current_state = TenantState::Activating(ActivatingFrom::Attaching);
2748 0 : }
2749 0 : }
2750 0 : debug!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), "Activating tenant");
2751 0 : activating = true;
2752 0 : // Continue outside the closure. We need to grab timelines.lock()
2753 0 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
2754 0 : });
2755 0 :
2756 0 : if activating {
2757 0 : let timelines_accessor = self.timelines.lock().unwrap();
2758 0 : let timelines_offloaded_accessor = self.timelines_offloaded.lock().unwrap();
2759 0 : let timelines_to_activate = timelines_accessor
2760 0 : .values()
2761 0 : .filter(|timeline| !(timeline.is_broken() || timeline.is_stopping()));
2762 0 :
2763 0 : // Before activation, populate each Timeline's GcInfo with information about its children
2764 0 : self.initialize_gc_info(&timelines_accessor, &timelines_offloaded_accessor, None);
2765 0 :
2766 0 : // Spawn gc and compaction loops. The loops will shut themselves
2767 0 : // down when they notice that the tenant is inactive.
2768 0 : tasks::start_background_loops(self, background_jobs_can_start);
2769 0 :
2770 0 : let mut activated_timelines = 0;
2771 :
2772 0 : for timeline in timelines_to_activate {
2773 0 : timeline.activate(
2774 0 : self.clone(),
2775 0 : broker_client.clone(),
2776 0 : background_jobs_can_start,
2777 0 : ctx,
2778 0 : );
2779 0 : activated_timelines += 1;
2780 0 : }
2781 :
2782 0 : self.state.send_modify(move |current_state| {
2783 0 : assert!(
2784 0 : matches!(current_state, TenantState::Activating(_)),
2785 0 : "set_stopping and set_broken wait for us to leave Activating state",
2786 : );
2787 0 : *current_state = TenantState::Active;
2788 0 :
2789 0 : let elapsed = self.constructed_at.elapsed();
2790 0 : let total_timelines = timelines_accessor.len();
2791 0 :
2792 0 : // log a lot of stuff, because some tenants sometimes suffer from user-visible
2793 0 : // times to activate. see https://github.com/neondatabase/neon/issues/4025
2794 0 : info!(
2795 0 : since_creation_millis = elapsed.as_millis(),
2796 0 : tenant_id = %self.tenant_shard_id.tenant_id,
2797 0 : shard_id = %self.tenant_shard_id.shard_slug(),
2798 0 : activated_timelines,
2799 0 : total_timelines,
2800 0 : post_state = <&'static str>::from(&*current_state),
2801 0 : "activation attempt finished"
2802 : );
2803 :
2804 0 : TENANT.activation.observe(elapsed.as_secs_f64());
2805 0 : });
2806 0 : }
2807 0 : }
2808 :
2809 : /// Shutdown the tenant and join all of the spawned tasks.
2810 : ///
2811 : /// The method caters for all use-cases:
2812 : /// - pageserver shutdown (freeze_and_flush == true)
2813 : /// - detach + ignore (freeze_and_flush == false)
2814 : ///
2815 : /// This will attempt to shutdown even if tenant is broken.
2816 : ///
2817 : /// `shutdown_progress` is a [`completion::Barrier`] for the shutdown initiated by this call.
2818 : /// If the tenant is already shutting down, we return a clone of the first shutdown call's
2819 : /// `Barrier` as an `Err`. This not-first caller can use the returned barrier to join with
2820 : /// the ongoing shutdown.
2821 6 : async fn shutdown(
2822 6 : &self,
2823 6 : shutdown_progress: completion::Barrier,
2824 6 : shutdown_mode: timeline::ShutdownMode,
2825 6 : ) -> Result<(), completion::Barrier> {
2826 6 : span::debug_assert_current_span_has_tenant_id();
2827 :
2828 : // Set tenant (and its timlines) to Stoppping state.
2829 : //
2830 : // Since we can only transition into Stopping state after activation is complete,
2831 : // run it in a JoinSet so all tenants have a chance to stop before we get SIGKILLed.
2832 : //
2833 : // Transitioning tenants to Stopping state has a couple of non-obvious side effects:
2834 : // 1. Lock out any new requests to the tenants.
2835 : // 2. Signal cancellation to WAL receivers (we wait on it below).
2836 : // 3. Signal cancellation for other tenant background loops.
2837 : // 4. ???
2838 : //
2839 : // The waiting for the cancellation is not done uniformly.
2840 : // We certainly wait for WAL receivers to shut down.
2841 : // That is necessary so that no new data comes in before the freeze_and_flush.
2842 : // But the tenant background loops are joined-on in our caller.
2843 : // It's mesed up.
2844 : // we just ignore the failure to stop
2845 :
2846 : // If we're still attaching, fire the cancellation token early to drop out: this
2847 : // will prevent us flushing, but ensures timely shutdown if some I/O during attach
2848 : // is very slow.
2849 6 : let shutdown_mode = if matches!(self.current_state(), TenantState::Attaching) {
2850 0 : self.cancel.cancel();
2851 0 :
2852 0 : // Having fired our cancellation token, do not try and flush timelines: their cancellation tokens
2853 0 : // are children of ours, so their flush loops will have shut down already
2854 0 : timeline::ShutdownMode::Hard
2855 : } else {
2856 6 : shutdown_mode
2857 : };
2858 :
2859 6 : match self.set_stopping(shutdown_progress, false, false).await {
2860 6 : Ok(()) => {}
2861 0 : Err(SetStoppingError::Broken) => {
2862 0 : // assume that this is acceptable
2863 0 : }
2864 0 : Err(SetStoppingError::AlreadyStopping(other)) => {
2865 0 : // give caller the option to wait for this this shutdown
2866 0 : info!("Tenant::shutdown: AlreadyStopping");
2867 0 : return Err(other);
2868 : }
2869 : };
2870 :
2871 6 : let mut js = tokio::task::JoinSet::new();
2872 6 : {
2873 6 : let timelines = self.timelines.lock().unwrap();
2874 6 : timelines.values().for_each(|timeline| {
2875 6 : let timeline = Arc::clone(timeline);
2876 6 : let timeline_id = timeline.timeline_id;
2877 6 : let span = tracing::info_span!("timeline_shutdown", %timeline_id, ?shutdown_mode);
2878 12 : js.spawn(async move { timeline.shutdown(shutdown_mode).instrument(span).await });
2879 6 : });
2880 6 : }
2881 6 : {
2882 6 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
2883 6 : timelines_offloaded.values().for_each(|timeline| {
2884 0 : timeline.defuse_for_tenant_drop();
2885 6 : });
2886 6 : }
2887 6 : // test_long_timeline_create_then_tenant_delete is leaning on this message
2888 6 : tracing::info!("Waiting for timelines...");
2889 12 : while let Some(res) = js.join_next().await {
2890 0 : match res {
2891 6 : Ok(()) => {}
2892 0 : Err(je) if je.is_cancelled() => unreachable!("no cancelling used"),
2893 0 : Err(je) if je.is_panic() => { /* logged already */ }
2894 0 : Err(je) => warn!("unexpected JoinError: {je:?}"),
2895 : }
2896 : }
2897 :
2898 : // We cancel the Tenant's cancellation token _after_ the timelines have all shut down. This permits
2899 : // them to continue to do work during their shutdown methods, e.g. flushing data.
2900 6 : tracing::debug!("Cancelling CancellationToken");
2901 6 : self.cancel.cancel();
2902 6 :
2903 6 : // shutdown all tenant and timeline tasks: gc, compaction, page service
2904 6 : // No new tasks will be started for this tenant because it's in `Stopping` state.
2905 6 : //
2906 6 : // this will additionally shutdown and await all timeline tasks.
2907 6 : tracing::debug!("Waiting for tasks...");
2908 6 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), None).await;
2909 :
2910 6 : if let Some(walredo_mgr) = self.walredo_mgr.as_ref() {
2911 6 : walredo_mgr.shutdown().await;
2912 0 : }
2913 :
2914 : // Wait for any in-flight operations to complete
2915 6 : self.gate.close().await;
2916 :
2917 6 : remove_tenant_metrics(&self.tenant_shard_id);
2918 6 :
2919 6 : Ok(())
2920 6 : }
2921 :
2922 : /// Change tenant status to Stopping, to mark that it is being shut down.
2923 : ///
2924 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
2925 : ///
2926 : /// This function is not cancel-safe!
2927 : ///
2928 : /// `allow_transition_from_loading` is needed for the special case of loading task deleting the tenant.
2929 : /// `allow_transition_from_attaching` is needed for the special case of attaching deleted tenant.
2930 6 : async fn set_stopping(
2931 6 : &self,
2932 6 : progress: completion::Barrier,
2933 6 : _allow_transition_from_loading: bool,
2934 6 : allow_transition_from_attaching: bool,
2935 6 : ) -> Result<(), SetStoppingError> {
2936 6 : let mut rx = self.state.subscribe();
2937 6 :
2938 6 : // cannot stop before we're done activating, so wait out until we're done activating
2939 6 : rx.wait_for(|state| match state {
2940 0 : TenantState::Attaching if allow_transition_from_attaching => true,
2941 : TenantState::Activating(_) | TenantState::Attaching => {
2942 0 : info!(
2943 0 : "waiting for {} to turn Active|Broken|Stopping",
2944 0 : <&'static str>::from(state)
2945 : );
2946 0 : false
2947 : }
2948 6 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
2949 6 : })
2950 0 : .await
2951 6 : .expect("cannot drop self.state while on a &self method");
2952 6 :
2953 6 : // we now know we're done activating, let's see whether this task is the winner to transition into Stopping
2954 6 : let mut err = None;
2955 6 : let stopping = self.state.send_if_modified(|current_state| match current_state {
2956 : TenantState::Activating(_) => {
2957 0 : unreachable!("1we ensured above that we're done with activation, and, there is no re-activation")
2958 : }
2959 : TenantState::Attaching => {
2960 0 : if !allow_transition_from_attaching {
2961 0 : unreachable!("2we ensured above that we're done with activation, and, there is no re-activation")
2962 0 : };
2963 0 : *current_state = TenantState::Stopping { progress };
2964 0 : true
2965 : }
2966 : TenantState::Active => {
2967 : // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
2968 : // are created after the transition to Stopping. That's harmless, as the Timelines
2969 : // won't be accessible to anyone afterwards, because the Tenant is in Stopping state.
2970 6 : *current_state = TenantState::Stopping { progress };
2971 6 : // Continue stopping outside the closure. We need to grab timelines.lock()
2972 6 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
2973 6 : true
2974 : }
2975 0 : TenantState::Broken { reason, .. } => {
2976 0 : info!(
2977 0 : "Cannot set tenant to Stopping state, it is in Broken state due to: {reason}"
2978 : );
2979 0 : err = Some(SetStoppingError::Broken);
2980 0 : false
2981 : }
2982 0 : TenantState::Stopping { progress } => {
2983 0 : info!("Tenant is already in Stopping state");
2984 0 : err = Some(SetStoppingError::AlreadyStopping(progress.clone()));
2985 0 : false
2986 : }
2987 6 : });
2988 6 : match (stopping, err) {
2989 6 : (true, None) => {} // continue
2990 0 : (false, Some(err)) => return Err(err),
2991 0 : (true, Some(_)) => unreachable!(
2992 0 : "send_if_modified closure must error out if not transitioning to Stopping"
2993 0 : ),
2994 0 : (false, None) => unreachable!(
2995 0 : "send_if_modified closure must return true if transitioning to Stopping"
2996 0 : ),
2997 : }
2998 :
2999 6 : let timelines_accessor = self.timelines.lock().unwrap();
3000 6 : let not_broken_timelines = timelines_accessor
3001 6 : .values()
3002 6 : .filter(|timeline| !timeline.is_broken());
3003 12 : for timeline in not_broken_timelines {
3004 6 : timeline.set_state(TimelineState::Stopping);
3005 6 : }
3006 6 : Ok(())
3007 6 : }
3008 :
3009 : /// Method for tenant::mgr to transition us into Broken state in case of a late failure in
3010 : /// `remove_tenant_from_memory`
3011 : ///
3012 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
3013 : ///
3014 : /// In tests, we also use this to set tenants to Broken state on purpose.
3015 0 : pub(crate) async fn set_broken(&self, reason: String) {
3016 0 : let mut rx = self.state.subscribe();
3017 0 :
3018 0 : // The load & attach routines own the tenant state until it has reached `Active`.
3019 0 : // So, wait until it's done.
3020 0 : rx.wait_for(|state| match state {
3021 : TenantState::Activating(_) | TenantState::Attaching => {
3022 0 : info!(
3023 0 : "waiting for {} to turn Active|Broken|Stopping",
3024 0 : <&'static str>::from(state)
3025 : );
3026 0 : false
3027 : }
3028 0 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3029 0 : })
3030 0 : .await
3031 0 : .expect("cannot drop self.state while on a &self method");
3032 0 :
3033 0 : // we now know we're done activating, let's see whether this task is the winner to transition into Broken
3034 0 : self.set_broken_no_wait(reason)
3035 0 : }
3036 :
3037 0 : pub(crate) fn set_broken_no_wait(&self, reason: impl Display) {
3038 0 : let reason = reason.to_string();
3039 0 : self.state.send_modify(|current_state| {
3040 0 : match *current_state {
3041 : TenantState::Activating(_) | TenantState::Attaching => {
3042 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
3043 : }
3044 : TenantState::Active => {
3045 0 : if cfg!(feature = "testing") {
3046 0 : warn!("Changing Active tenant to Broken state, reason: {}", reason);
3047 0 : *current_state = TenantState::broken_from_reason(reason);
3048 : } else {
3049 0 : unreachable!("not allowed to call set_broken on Active tenants in non-testing builds")
3050 : }
3051 : }
3052 : TenantState::Broken { .. } => {
3053 0 : warn!("Tenant is already in Broken state");
3054 : }
3055 : // This is the only "expected" path, any other path is a bug.
3056 : TenantState::Stopping { .. } => {
3057 0 : warn!(
3058 0 : "Marking Stopping tenant as Broken state, reason: {}",
3059 : reason
3060 : );
3061 0 : *current_state = TenantState::broken_from_reason(reason);
3062 : }
3063 : }
3064 0 : });
3065 0 : }
3066 :
3067 0 : pub fn subscribe_for_state_updates(&self) -> watch::Receiver<TenantState> {
3068 0 : self.state.subscribe()
3069 0 : }
3070 :
3071 : /// The activate_now semaphore is initialized with zero units. As soon as
3072 : /// we add a unit, waiters will be able to acquire a unit and proceed.
3073 0 : pub(crate) fn activate_now(&self) {
3074 0 : self.activate_now_sem.add_permits(1);
3075 0 : }
3076 :
3077 0 : pub(crate) async fn wait_to_become_active(
3078 0 : &self,
3079 0 : timeout: Duration,
3080 0 : ) -> Result<(), GetActiveTenantError> {
3081 0 : let mut receiver = self.state.subscribe();
3082 : loop {
3083 0 : let current_state = receiver.borrow_and_update().clone();
3084 0 : match current_state {
3085 : TenantState::Attaching | TenantState::Activating(_) => {
3086 : // in these states, there's a chance that we can reach ::Active
3087 0 : self.activate_now();
3088 0 : match timeout_cancellable(timeout, &self.cancel, receiver.changed()).await {
3089 0 : Ok(r) => {
3090 0 : r.map_err(
3091 0 : |_e: tokio::sync::watch::error::RecvError|
3092 : // Tenant existed but was dropped: report it as non-existent
3093 0 : GetActiveTenantError::NotFound(GetTenantError::NotFound(self.tenant_shard_id.tenant_id))
3094 0 : )?
3095 : }
3096 : Err(TimeoutCancellableError::Cancelled) => {
3097 0 : return Err(GetActiveTenantError::Cancelled);
3098 : }
3099 : Err(TimeoutCancellableError::Timeout) => {
3100 0 : return Err(GetActiveTenantError::WaitForActiveTimeout {
3101 0 : latest_state: Some(self.current_state()),
3102 0 : wait_time: timeout,
3103 0 : });
3104 : }
3105 : }
3106 : }
3107 : TenantState::Active { .. } => {
3108 0 : return Ok(());
3109 : }
3110 0 : TenantState::Broken { reason, .. } => {
3111 0 : // This is fatal, and reported distinctly from the general case of "will never be active" because
3112 0 : // it's logically a 500 to external API users (broken is always a bug).
3113 0 : return Err(GetActiveTenantError::Broken(reason));
3114 : }
3115 : TenantState::Stopping { .. } => {
3116 : // There's no chance the tenant can transition back into ::Active
3117 0 : return Err(GetActiveTenantError::WillNotBecomeActive(current_state));
3118 : }
3119 : }
3120 : }
3121 0 : }
3122 :
3123 0 : pub(crate) fn get_attach_mode(&self) -> AttachmentMode {
3124 0 : self.tenant_conf.load().location.attach_mode
3125 0 : }
3126 :
3127 : /// For API access: generate a LocationConfig equivalent to the one that would be used to
3128 : /// create a Tenant in the same state. Do not use this in hot paths: it's for relatively
3129 : /// rare external API calls, like a reconciliation at startup.
3130 0 : pub(crate) fn get_location_conf(&self) -> models::LocationConfig {
3131 0 : let conf = self.tenant_conf.load();
3132 :
3133 0 : let location_config_mode = match conf.location.attach_mode {
3134 0 : AttachmentMode::Single => models::LocationConfigMode::AttachedSingle,
3135 0 : AttachmentMode::Multi => models::LocationConfigMode::AttachedMulti,
3136 0 : AttachmentMode::Stale => models::LocationConfigMode::AttachedStale,
3137 : };
3138 :
3139 : // We have a pageserver TenantConf, we need the API-facing TenantConfig.
3140 0 : let tenant_config: models::TenantConfig = conf.tenant_conf.clone().into();
3141 0 :
3142 0 : models::LocationConfig {
3143 0 : mode: location_config_mode,
3144 0 : generation: self.generation.into(),
3145 0 : secondary_conf: None,
3146 0 : shard_number: self.shard_identity.number.0,
3147 0 : shard_count: self.shard_identity.count.literal(),
3148 0 : shard_stripe_size: self.shard_identity.stripe_size.0,
3149 0 : tenant_conf: tenant_config,
3150 0 : }
3151 0 : }
3152 :
3153 0 : pub(crate) fn get_tenant_shard_id(&self) -> &TenantShardId {
3154 0 : &self.tenant_shard_id
3155 0 : }
3156 :
3157 0 : pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
3158 0 : self.shard_identity.stripe_size
3159 0 : }
3160 :
3161 0 : pub(crate) fn get_generation(&self) -> Generation {
3162 0 : self.generation
3163 0 : }
3164 :
3165 : /// This function partially shuts down the tenant (it shuts down the Timelines) and is fallible,
3166 : /// and can leave the tenant in a bad state if it fails. The caller is responsible for
3167 : /// resetting this tenant to a valid state if we fail.
3168 0 : pub(crate) async fn split_prepare(
3169 0 : &self,
3170 0 : child_shards: &Vec<TenantShardId>,
3171 0 : ) -> anyhow::Result<()> {
3172 0 : let (timelines, offloaded) = {
3173 0 : let timelines = self.timelines.lock().unwrap();
3174 0 : let offloaded = self.timelines_offloaded.lock().unwrap();
3175 0 : (timelines.clone(), offloaded.clone())
3176 0 : };
3177 0 : let timelines_iter = timelines
3178 0 : .values()
3179 0 : .map(TimelineOrOffloadedArcRef::<'_>::from)
3180 0 : .chain(
3181 0 : offloaded
3182 0 : .values()
3183 0 : .map(TimelineOrOffloadedArcRef::<'_>::from),
3184 0 : );
3185 0 : for timeline in timelines_iter {
3186 : // We do not block timeline creation/deletion during splits inside the pageserver: it is up to higher levels
3187 : // to ensure that they do not start a split if currently in the process of doing these.
3188 :
3189 0 : let timeline_id = timeline.timeline_id();
3190 :
3191 0 : if let TimelineOrOffloadedArcRef::Timeline(timeline) = timeline {
3192 : // Upload an index from the parent: this is partly to provide freshness for the
3193 : // child tenants that will copy it, and partly for general ease-of-debugging: there will
3194 : // always be a parent shard index in the same generation as we wrote the child shard index.
3195 0 : tracing::info!(%timeline_id, "Uploading index");
3196 0 : timeline
3197 0 : .remote_client
3198 0 : .schedule_index_upload_for_file_changes()?;
3199 0 : timeline.remote_client.wait_completion().await?;
3200 0 : }
3201 :
3202 0 : let remote_client = match timeline {
3203 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.remote_client.clone(),
3204 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => {
3205 0 : let remote_client = self
3206 0 : .build_timeline_client(offloaded.timeline_id, self.remote_storage.clone());
3207 0 : Arc::new(remote_client)
3208 : }
3209 : };
3210 :
3211 : // Shut down the timeline's remote client: this means that the indices we write
3212 : // for child shards will not be invalidated by the parent shard deleting layers.
3213 0 : tracing::info!(%timeline_id, "Shutting down remote storage client");
3214 0 : remote_client.shutdown().await;
3215 :
3216 : // Download methods can still be used after shutdown, as they don't flow through the remote client's
3217 : // queue. In principal the RemoteTimelineClient could provide this without downloading it, but this
3218 : // operation is rare, so it's simpler to just download it (and robustly guarantees that the index
3219 : // we use here really is the remotely persistent one).
3220 0 : tracing::info!(%timeline_id, "Downloading index_part from parent");
3221 0 : let result = remote_client
3222 0 : .download_index_file(&self.cancel)
3223 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))
3224 0 : .await?;
3225 0 : let index_part = match result {
3226 : MaybeDeletedIndexPart::Deleted(_) => {
3227 0 : anyhow::bail!("Timeline deletion happened concurrently with split")
3228 : }
3229 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
3230 : };
3231 :
3232 0 : for child_shard in child_shards {
3233 0 : tracing::info!(%timeline_id, "Uploading index_part for child {}", child_shard.to_index());
3234 0 : upload_index_part(
3235 0 : &self.remote_storage,
3236 0 : child_shard,
3237 0 : &timeline_id,
3238 0 : self.generation,
3239 0 : &index_part,
3240 0 : &self.cancel,
3241 0 : )
3242 0 : .await?;
3243 : }
3244 : }
3245 :
3246 0 : let tenant_manifest = self.build_tenant_manifest();
3247 0 : for child_shard in child_shards {
3248 0 : tracing::info!(
3249 0 : "Uploading tenant manifest for child {}",
3250 0 : child_shard.to_index()
3251 : );
3252 0 : upload_tenant_manifest(
3253 0 : &self.remote_storage,
3254 0 : child_shard,
3255 0 : self.generation,
3256 0 : &tenant_manifest,
3257 0 : &self.cancel,
3258 0 : )
3259 0 : .await?;
3260 : }
3261 :
3262 0 : Ok(())
3263 0 : }
3264 :
3265 0 : pub(crate) fn get_sizes(&self) -> TopTenantShardItem {
3266 0 : let mut result = TopTenantShardItem {
3267 0 : id: self.tenant_shard_id,
3268 0 : resident_size: 0,
3269 0 : physical_size: 0,
3270 0 : max_logical_size: 0,
3271 0 : };
3272 :
3273 0 : for timeline in self.timelines.lock().unwrap().values() {
3274 0 : result.resident_size += timeline.metrics.resident_physical_size_gauge.get();
3275 0 :
3276 0 : result.physical_size += timeline
3277 0 : .remote_client
3278 0 : .metrics
3279 0 : .remote_physical_size_gauge
3280 0 : .get();
3281 0 : result.max_logical_size = std::cmp::max(
3282 0 : result.max_logical_size,
3283 0 : timeline.metrics.current_logical_size_gauge.get(),
3284 0 : );
3285 0 : }
3286 :
3287 0 : result
3288 0 : }
3289 : }
3290 :
3291 : /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
3292 : /// perform a topological sort, so that the parent of each timeline comes
3293 : /// before the children.
3294 : /// E extracts the ancestor from T
3295 : /// This allows for T to be different. It can be TimelineMetadata, can be Timeline itself, etc.
3296 192 : fn tree_sort_timelines<T, E>(
3297 192 : timelines: HashMap<TimelineId, T>,
3298 192 : extractor: E,
3299 192 : ) -> anyhow::Result<Vec<(TimelineId, T)>>
3300 192 : where
3301 192 : E: Fn(&T) -> Option<TimelineId>,
3302 192 : {
3303 192 : let mut result = Vec::with_capacity(timelines.len());
3304 192 :
3305 192 : let mut now = Vec::with_capacity(timelines.len());
3306 192 : // (ancestor, children)
3307 192 : let mut later: HashMap<TimelineId, Vec<(TimelineId, T)>> =
3308 192 : HashMap::with_capacity(timelines.len());
3309 :
3310 198 : for (timeline_id, value) in timelines {
3311 6 : if let Some(ancestor_id) = extractor(&value) {
3312 2 : let children = later.entry(ancestor_id).or_default();
3313 2 : children.push((timeline_id, value));
3314 4 : } else {
3315 4 : now.push((timeline_id, value));
3316 4 : }
3317 : }
3318 :
3319 198 : while let Some((timeline_id, metadata)) = now.pop() {
3320 6 : result.push((timeline_id, metadata));
3321 : // All children of this can be loaded now
3322 6 : if let Some(mut children) = later.remove(&timeline_id) {
3323 2 : now.append(&mut children);
3324 4 : }
3325 : }
3326 :
3327 : // All timelines should be visited now. Unless there were timelines with missing ancestors.
3328 192 : if !later.is_empty() {
3329 0 : for (missing_id, orphan_ids) in later {
3330 0 : for (orphan_id, _) in orphan_ids {
3331 0 : error!("could not load timeline {orphan_id} because its ancestor timeline {missing_id} could not be loaded");
3332 : }
3333 : }
3334 0 : bail!("could not load tenant because some timelines are missing ancestors");
3335 192 : }
3336 192 :
3337 192 : Ok(result)
3338 192 : }
3339 :
3340 : impl Tenant {
3341 0 : pub fn tenant_specific_overrides(&self) -> TenantConfOpt {
3342 0 : self.tenant_conf.load().tenant_conf.clone()
3343 0 : }
3344 :
3345 0 : pub fn effective_config(&self) -> TenantConf {
3346 0 : self.tenant_specific_overrides()
3347 0 : .merge(self.conf.default_tenant_conf.clone())
3348 0 : }
3349 :
3350 0 : pub fn get_checkpoint_distance(&self) -> u64 {
3351 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3352 0 : tenant_conf
3353 0 : .checkpoint_distance
3354 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
3355 0 : }
3356 :
3357 0 : pub fn get_checkpoint_timeout(&self) -> Duration {
3358 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3359 0 : tenant_conf
3360 0 : .checkpoint_timeout
3361 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
3362 0 : }
3363 :
3364 0 : pub fn get_compaction_target_size(&self) -> u64 {
3365 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3366 0 : tenant_conf
3367 0 : .compaction_target_size
3368 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
3369 0 : }
3370 :
3371 0 : pub fn get_compaction_period(&self) -> Duration {
3372 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3373 0 : tenant_conf
3374 0 : .compaction_period
3375 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
3376 0 : }
3377 :
3378 0 : pub fn get_compaction_threshold(&self) -> usize {
3379 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3380 0 : tenant_conf
3381 0 : .compaction_threshold
3382 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
3383 0 : }
3384 :
3385 0 : pub fn get_gc_horizon(&self) -> u64 {
3386 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3387 0 : tenant_conf
3388 0 : .gc_horizon
3389 0 : .unwrap_or(self.conf.default_tenant_conf.gc_horizon)
3390 0 : }
3391 :
3392 0 : pub fn get_gc_period(&self) -> Duration {
3393 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3394 0 : tenant_conf
3395 0 : .gc_period
3396 0 : .unwrap_or(self.conf.default_tenant_conf.gc_period)
3397 0 : }
3398 :
3399 0 : pub fn get_image_creation_threshold(&self) -> usize {
3400 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3401 0 : tenant_conf
3402 0 : .image_creation_threshold
3403 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
3404 0 : }
3405 :
3406 0 : pub fn get_pitr_interval(&self) -> Duration {
3407 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3408 0 : tenant_conf
3409 0 : .pitr_interval
3410 0 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
3411 0 : }
3412 :
3413 0 : pub fn get_min_resident_size_override(&self) -> Option<u64> {
3414 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3415 0 : tenant_conf
3416 0 : .min_resident_size_override
3417 0 : .or(self.conf.default_tenant_conf.min_resident_size_override)
3418 0 : }
3419 :
3420 0 : pub fn get_heatmap_period(&self) -> Option<Duration> {
3421 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3422 0 : let heatmap_period = tenant_conf
3423 0 : .heatmap_period
3424 0 : .unwrap_or(self.conf.default_tenant_conf.heatmap_period);
3425 0 : if heatmap_period.is_zero() {
3426 0 : None
3427 : } else {
3428 0 : Some(heatmap_period)
3429 : }
3430 0 : }
3431 :
3432 4 : pub fn get_lsn_lease_length(&self) -> Duration {
3433 4 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3434 4 : tenant_conf
3435 4 : .lsn_lease_length
3436 4 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
3437 4 : }
3438 :
3439 : /// Generate an up-to-date TenantManifest based on the state of this Tenant.
3440 2 : fn build_tenant_manifest(&self) -> TenantManifest {
3441 2 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
3442 2 :
3443 2 : let mut timeline_manifests = timelines_offloaded
3444 2 : .iter()
3445 2 : .map(|(_timeline_id, offloaded)| offloaded.manifest())
3446 2 : .collect::<Vec<_>>();
3447 2 : // Sort the manifests so that our output is deterministic
3448 2 : timeline_manifests.sort_by_key(|timeline_manifest| timeline_manifest.timeline_id);
3449 2 :
3450 2 : TenantManifest {
3451 2 : version: LATEST_TENANT_MANIFEST_VERSION,
3452 2 : offloaded_timelines: timeline_manifests,
3453 2 : }
3454 2 : }
3455 :
3456 0 : pub fn set_new_tenant_config(&self, new_tenant_conf: TenantConfOpt) {
3457 0 : // Use read-copy-update in order to avoid overwriting the location config
3458 0 : // state if this races with [`Tenant::set_new_location_config`]. Note that
3459 0 : // this race is not possible if both request types come from the storage
3460 0 : // controller (as they should!) because an exclusive op lock is required
3461 0 : // on the storage controller side.
3462 0 :
3463 0 : self.tenant_conf.rcu(|inner| {
3464 0 : Arc::new(AttachedTenantConf {
3465 0 : tenant_conf: new_tenant_conf.clone(),
3466 0 : location: inner.location,
3467 0 : // Attached location is not changed, no need to update lsn lease deadline.
3468 0 : lsn_lease_deadline: inner.lsn_lease_deadline,
3469 0 : })
3470 0 : });
3471 0 :
3472 0 : let updated = self.tenant_conf.load().clone();
3473 0 :
3474 0 : self.tenant_conf_updated(&new_tenant_conf);
3475 0 : // Don't hold self.timelines.lock() during the notifies.
3476 0 : // There's no risk of deadlock right now, but there could be if we consolidate
3477 0 : // mutexes in struct Timeline in the future.
3478 0 : let timelines = self.list_timelines();
3479 0 : for timeline in timelines {
3480 0 : timeline.tenant_conf_updated(&updated);
3481 0 : }
3482 0 : }
3483 :
3484 0 : pub(crate) fn set_new_location_config(&self, new_conf: AttachedTenantConf) {
3485 0 : let new_tenant_conf = new_conf.tenant_conf.clone();
3486 0 :
3487 0 : self.tenant_conf.store(Arc::new(new_conf.clone()));
3488 0 :
3489 0 : self.tenant_conf_updated(&new_tenant_conf);
3490 0 : // Don't hold self.timelines.lock() during the notifies.
3491 0 : // There's no risk of deadlock right now, but there could be if we consolidate
3492 0 : // mutexes in struct Timeline in the future.
3493 0 : let timelines = self.list_timelines();
3494 0 : for timeline in timelines {
3495 0 : timeline.tenant_conf_updated(&new_conf);
3496 0 : }
3497 0 : }
3498 :
3499 192 : fn get_timeline_get_throttle_config(
3500 192 : psconf: &'static PageServerConf,
3501 192 : overrides: &TenantConfOpt,
3502 192 : ) -> throttle::Config {
3503 192 : overrides
3504 192 : .timeline_get_throttle
3505 192 : .clone()
3506 192 : .unwrap_or(psconf.default_tenant_conf.timeline_get_throttle.clone())
3507 192 : }
3508 :
3509 0 : pub(crate) fn tenant_conf_updated(&self, new_conf: &TenantConfOpt) {
3510 0 : let conf = Self::get_timeline_get_throttle_config(self.conf, new_conf);
3511 0 : self.timeline_get_throttle.reconfigure(conf)
3512 0 : }
3513 :
3514 : /// Helper function to create a new Timeline struct.
3515 : ///
3516 : /// The returned Timeline is in Loading state. The caller is responsible for
3517 : /// initializing any on-disk state, and for inserting the Timeline to the 'timelines'
3518 : /// map.
3519 : ///
3520 : /// `validate_ancestor == false` is used when a timeline is created for deletion
3521 : /// and we might not have the ancestor present anymore which is fine for to be
3522 : /// deleted timelines.
3523 418 : fn create_timeline_struct(
3524 418 : &self,
3525 418 : new_timeline_id: TimelineId,
3526 418 : new_metadata: &TimelineMetadata,
3527 418 : ancestor: Option<Arc<Timeline>>,
3528 418 : resources: TimelineResources,
3529 418 : cause: CreateTimelineCause,
3530 418 : create_idempotency: CreateTimelineIdempotency,
3531 418 : ) -> anyhow::Result<Arc<Timeline>> {
3532 418 : let state = match cause {
3533 : CreateTimelineCause::Load => {
3534 418 : let ancestor_id = new_metadata.ancestor_timeline();
3535 418 : anyhow::ensure!(
3536 418 : ancestor_id == ancestor.as_ref().map(|t| t.timeline_id),
3537 0 : "Timeline's {new_timeline_id} ancestor {ancestor_id:?} was not found"
3538 : );
3539 418 : TimelineState::Loading
3540 : }
3541 0 : CreateTimelineCause::Delete => TimelineState::Stopping,
3542 : };
3543 :
3544 418 : let pg_version = new_metadata.pg_version();
3545 418 :
3546 418 : let timeline = Timeline::new(
3547 418 : self.conf,
3548 418 : Arc::clone(&self.tenant_conf),
3549 418 : new_metadata,
3550 418 : ancestor,
3551 418 : new_timeline_id,
3552 418 : self.tenant_shard_id,
3553 418 : self.generation,
3554 418 : self.shard_identity,
3555 418 : self.walredo_mgr.clone(),
3556 418 : resources,
3557 418 : pg_version,
3558 418 : state,
3559 418 : self.attach_wal_lag_cooldown.clone(),
3560 418 : create_idempotency,
3561 418 : self.cancel.child_token(),
3562 418 : );
3563 418 :
3564 418 : Ok(timeline)
3565 418 : }
3566 :
3567 : // Allow too_many_arguments because a constructor's argument list naturally grows with the
3568 : // number of attributes in the struct: breaking these out into a builder wouldn't be helpful.
3569 : #[allow(clippy::too_many_arguments)]
3570 192 : fn new(
3571 192 : state: TenantState,
3572 192 : conf: &'static PageServerConf,
3573 192 : attached_conf: AttachedTenantConf,
3574 192 : shard_identity: ShardIdentity,
3575 192 : walredo_mgr: Option<Arc<WalRedoManager>>,
3576 192 : tenant_shard_id: TenantShardId,
3577 192 : remote_storage: GenericRemoteStorage,
3578 192 : deletion_queue_client: DeletionQueueClient,
3579 192 : l0_flush_global_state: L0FlushGlobalState,
3580 192 : ) -> Tenant {
3581 192 : debug_assert!(
3582 192 : !attached_conf.location.generation.is_none() || conf.control_plane_api.is_none()
3583 : );
3584 :
3585 192 : let (state, mut rx) = watch::channel(state);
3586 192 :
3587 192 : tokio::spawn(async move {
3588 192 : // reflect tenant state in metrics:
3589 192 : // - global per tenant state: TENANT_STATE_METRIC
3590 192 : // - "set" of broken tenants: BROKEN_TENANTS_SET
3591 192 : //
3592 192 : // set of broken tenants should not have zero counts so that it remains accessible for
3593 192 : // alerting.
3594 192 :
3595 192 : let tid = tenant_shard_id.to_string();
3596 192 : let shard_id = tenant_shard_id.shard_slug().to_string();
3597 192 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
3598 :
3599 384 : fn inspect_state(state: &TenantState) -> ([&'static str; 1], bool) {
3600 384 : ([state.into()], matches!(state, TenantState::Broken { .. }))
3601 384 : }
3602 :
3603 192 : let mut tuple = inspect_state(&rx.borrow_and_update());
3604 192 :
3605 192 : let is_broken = tuple.1;
3606 192 : let mut counted_broken = if is_broken {
3607 : // add the id to the set right away, there should not be any updates on the channel
3608 : // after before tenant is removed, if ever
3609 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
3610 0 : true
3611 : } else {
3612 192 : false
3613 : };
3614 :
3615 : loop {
3616 384 : let labels = &tuple.0;
3617 384 : let current = TENANT_STATE_METRIC.with_label_values(labels);
3618 384 : current.inc();
3619 384 :
3620 384 : if rx.changed().await.is_err() {
3621 : // tenant has been dropped
3622 16 : current.dec();
3623 16 : drop(BROKEN_TENANTS_SET.remove_label_values(set_key));
3624 16 : break;
3625 192 : }
3626 192 :
3627 192 : current.dec();
3628 192 : tuple = inspect_state(&rx.borrow_and_update());
3629 192 :
3630 192 : let is_broken = tuple.1;
3631 192 : if is_broken && !counted_broken {
3632 0 : counted_broken = true;
3633 0 : // insert the tenant_id (back) into the set while avoiding needless counter
3634 0 : // access
3635 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
3636 192 : }
3637 : }
3638 192 : });
3639 192 :
3640 192 : Tenant {
3641 192 : tenant_shard_id,
3642 192 : shard_identity,
3643 192 : generation: attached_conf.location.generation,
3644 192 : conf,
3645 192 : // using now here is good enough approximation to catch tenants with really long
3646 192 : // activation times.
3647 192 : constructed_at: Instant::now(),
3648 192 : timelines: Mutex::new(HashMap::new()),
3649 192 : timelines_creating: Mutex::new(HashSet::new()),
3650 192 : timelines_offloaded: Mutex::new(HashMap::new()),
3651 192 : tenant_manifest_upload: Default::default(),
3652 192 : gc_cs: tokio::sync::Mutex::new(()),
3653 192 : walredo_mgr,
3654 192 : remote_storage,
3655 192 : deletion_queue_client,
3656 192 : state,
3657 192 : cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
3658 192 : cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
3659 192 : eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
3660 192 : compaction_circuit_breaker: std::sync::Mutex::new(CircuitBreaker::new(
3661 192 : format!("compaction-{tenant_shard_id}"),
3662 192 : 5,
3663 192 : // Compaction can be a very expensive operation, and might leak disk space. It also ought
3664 192 : // to be infallible, as long as remote storage is available. So if it repeatedly fails,
3665 192 : // use an extremely long backoff.
3666 192 : Some(Duration::from_secs(3600 * 24)),
3667 192 : )),
3668 192 : activate_now_sem: tokio::sync::Semaphore::new(0),
3669 192 : attach_wal_lag_cooldown: Arc::new(std::sync::OnceLock::new()),
3670 192 : cancel: CancellationToken::default(),
3671 192 : gate: Gate::default(),
3672 192 : timeline_get_throttle: Arc::new(throttle::Throttle::new(
3673 192 : Tenant::get_timeline_get_throttle_config(conf, &attached_conf.tenant_conf),
3674 192 : crate::metrics::tenant_throttling::TimelineGet::new(&tenant_shard_id),
3675 192 : )),
3676 192 : tenant_conf: Arc::new(ArcSwap::from_pointee(attached_conf)),
3677 192 : ongoing_timeline_detach: std::sync::Mutex::default(),
3678 192 : gc_block: Default::default(),
3679 192 : l0_flush_global_state,
3680 192 : }
3681 192 : }
3682 :
3683 : /// Locate and load config
3684 0 : pub(super) fn load_tenant_config(
3685 0 : conf: &'static PageServerConf,
3686 0 : tenant_shard_id: &TenantShardId,
3687 0 : ) -> Result<LocationConf, LoadConfigError> {
3688 0 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
3689 0 :
3690 0 : info!("loading tenant configuration from {config_path}");
3691 :
3692 : // load and parse file
3693 0 : let config = fs::read_to_string(&config_path).map_err(|e| {
3694 0 : match e.kind() {
3695 : std::io::ErrorKind::NotFound => {
3696 : // The config should almost always exist for a tenant directory:
3697 : // - When attaching a tenant, the config is the first thing we write
3698 : // - When detaching a tenant, we atomically move the directory to a tmp location
3699 : // before deleting contents.
3700 : //
3701 : // The very rare edge case that can result in a missing config is if we crash during attach
3702 : // between creating directory and writing config. Callers should handle that as if the
3703 : // directory didn't exist.
3704 :
3705 0 : LoadConfigError::NotFound(config_path)
3706 : }
3707 : _ => {
3708 : // No IO errors except NotFound are acceptable here: other kinds of error indicate local storage or permissions issues
3709 : // that we cannot cleanly recover
3710 0 : crate::virtual_file::on_fatal_io_error(&e, "Reading tenant config file")
3711 : }
3712 : }
3713 0 : })?;
3714 :
3715 0 : Ok(toml_edit::de::from_str::<LocationConf>(&config)?)
3716 0 : }
3717 :
3718 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
3719 : pub(super) async fn persist_tenant_config(
3720 : conf: &'static PageServerConf,
3721 : tenant_shard_id: &TenantShardId,
3722 : location_conf: &LocationConf,
3723 : ) -> std::io::Result<()> {
3724 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
3725 :
3726 : Self::persist_tenant_config_at(tenant_shard_id, &config_path, location_conf).await
3727 : }
3728 :
3729 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
3730 : pub(super) async fn persist_tenant_config_at(
3731 : tenant_shard_id: &TenantShardId,
3732 : config_path: &Utf8Path,
3733 : location_conf: &LocationConf,
3734 : ) -> std::io::Result<()> {
3735 : debug!("persisting tenantconf to {config_path}");
3736 :
3737 : let mut conf_content = r#"# This file contains a specific per-tenant's config.
3738 : # It is read in case of pageserver restart.
3739 : "#
3740 : .to_string();
3741 :
3742 0 : fail::fail_point!("tenant-config-before-write", |_| {
3743 0 : Err(std::io::Error::new(
3744 0 : std::io::ErrorKind::Other,
3745 0 : "tenant-config-before-write",
3746 0 : ))
3747 0 : });
3748 :
3749 : // Convert the config to a toml file.
3750 : conf_content +=
3751 : &toml_edit::ser::to_string_pretty(&location_conf).expect("Config serialization failed");
3752 :
3753 : let temp_path = path_with_suffix_extension(config_path, TEMP_FILE_SUFFIX);
3754 :
3755 : let conf_content = conf_content.into_bytes();
3756 : VirtualFile::crashsafe_overwrite(config_path.to_owned(), temp_path, conf_content).await
3757 : }
3758 :
3759 : //
3760 : // How garbage collection works:
3761 : //
3762 : // +--bar------------->
3763 : // /
3764 : // +----+-----foo---------------->
3765 : // /
3766 : // ----main--+-------------------------->
3767 : // \
3768 : // +-----baz-------->
3769 : //
3770 : //
3771 : // 1. Grab 'gc_cs' mutex to prevent new timelines from being created while Timeline's
3772 : // `gc_infos` are being refreshed
3773 : // 2. Scan collected timelines, and on each timeline, make note of the
3774 : // all the points where other timelines have been branched off.
3775 : // We will refrain from removing page versions at those LSNs.
3776 : // 3. For each timeline, scan all layer files on the timeline.
3777 : // Remove all files for which a newer file exists and which
3778 : // don't cover any branch point LSNs.
3779 : //
3780 : // TODO:
3781 : // - if a relation has a non-incremental persistent layer on a child branch, then we
3782 : // don't need to keep that in the parent anymore. But currently
3783 : // we do.
3784 4 : async fn gc_iteration_internal(
3785 4 : &self,
3786 4 : target_timeline_id: Option<TimelineId>,
3787 4 : horizon: u64,
3788 4 : pitr: Duration,
3789 4 : cancel: &CancellationToken,
3790 4 : ctx: &RequestContext,
3791 4 : ) -> Result<GcResult, GcError> {
3792 4 : let mut totals: GcResult = Default::default();
3793 4 : let now = Instant::now();
3794 :
3795 4 : let gc_timelines = self
3796 4 : .refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
3797 4 : .await?;
3798 :
3799 4 : failpoint_support::sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
3800 :
3801 : // If there is nothing to GC, we don't want any messages in the INFO log.
3802 4 : if !gc_timelines.is_empty() {
3803 4 : info!("{} timelines need GC", gc_timelines.len());
3804 : } else {
3805 0 : debug!("{} timelines need GC", gc_timelines.len());
3806 : }
3807 :
3808 : // Perform GC for each timeline.
3809 : //
3810 : // Note that we don't hold the `Tenant::gc_cs` lock here because we don't want to delay the
3811 : // branch creation task, which requires the GC lock. A GC iteration can run concurrently
3812 : // with branch creation.
3813 : //
3814 : // See comments in [`Tenant::branch_timeline`] for more information about why branch
3815 : // creation task can run concurrently with timeline's GC iteration.
3816 8 : for timeline in gc_timelines {
3817 4 : if cancel.is_cancelled() {
3818 : // We were requested to shut down. Stop and return with the progress we
3819 : // made.
3820 0 : break;
3821 4 : }
3822 4 : let result = match timeline.gc().await {
3823 : Err(GcError::TimelineCancelled) => {
3824 0 : if target_timeline_id.is_some() {
3825 : // If we were targetting this specific timeline, surface cancellation to caller
3826 0 : return Err(GcError::TimelineCancelled);
3827 : } else {
3828 : // A timeline may be shutting down independently of the tenant's lifecycle: we should
3829 : // skip past this and proceed to try GC on other timelines.
3830 0 : continue;
3831 : }
3832 : }
3833 4 : r => r?,
3834 : };
3835 4 : totals += result;
3836 : }
3837 :
3838 4 : totals.elapsed = now.elapsed();
3839 4 : Ok(totals)
3840 4 : }
3841 :
3842 : /// Refreshes the Timeline::gc_info for all timelines, returning the
3843 : /// vector of timelines which have [`Timeline::get_last_record_lsn`] past
3844 : /// [`Tenant::get_gc_horizon`].
3845 : ///
3846 : /// This is usually executed as part of periodic gc, but can now be triggered more often.
3847 0 : pub(crate) async fn refresh_gc_info(
3848 0 : &self,
3849 0 : cancel: &CancellationToken,
3850 0 : ctx: &RequestContext,
3851 0 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
3852 0 : // since this method can now be called at different rates than the configured gc loop, it
3853 0 : // might be that these configuration values get applied faster than what it was previously,
3854 0 : // since these were only read from the gc task.
3855 0 : let horizon = self.get_gc_horizon();
3856 0 : let pitr = self.get_pitr_interval();
3857 0 :
3858 0 : // refresh all timelines
3859 0 : let target_timeline_id = None;
3860 0 :
3861 0 : self.refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
3862 0 : .await
3863 0 : }
3864 :
3865 : /// Populate all Timelines' `GcInfo` with information about their children. We do not set the
3866 : /// PITR cutoffs here, because that requires I/O: this is done later, before GC, by [`Self::refresh_gc_info_internal`]
3867 : ///
3868 : /// Subsequently, parent-child relationships are updated incrementally inside [`Timeline::new`] and [`Timeline::drop`].
3869 0 : fn initialize_gc_info(
3870 0 : &self,
3871 0 : timelines: &std::sync::MutexGuard<HashMap<TimelineId, Arc<Timeline>>>,
3872 0 : timelines_offloaded: &std::sync::MutexGuard<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
3873 0 : restrict_to_timeline: Option<TimelineId>,
3874 0 : ) {
3875 0 : if restrict_to_timeline.is_none() {
3876 : // This function must be called before activation: after activation timeline create/delete operations
3877 : // might happen, and this function is not safe to run concurrently with those.
3878 0 : assert!(!self.is_active());
3879 0 : }
3880 :
3881 : // Scan all timelines. For each timeline, remember the timeline ID and
3882 : // the branch point where it was created.
3883 0 : let mut all_branchpoints: BTreeMap<TimelineId, Vec<(Lsn, TimelineId, MaybeOffloaded)>> =
3884 0 : BTreeMap::new();
3885 0 : timelines.iter().for_each(|(timeline_id, timeline_entry)| {
3886 0 : if let Some(ancestor_timeline_id) = &timeline_entry.get_ancestor_timeline_id() {
3887 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
3888 0 : ancestor_children.push((
3889 0 : timeline_entry.get_ancestor_lsn(),
3890 0 : *timeline_id,
3891 0 : MaybeOffloaded::No,
3892 0 : ));
3893 0 : }
3894 0 : });
3895 0 : timelines_offloaded
3896 0 : .iter()
3897 0 : .for_each(|(timeline_id, timeline_entry)| {
3898 0 : let Some(ancestor_timeline_id) = &timeline_entry.ancestor_timeline_id else {
3899 0 : return;
3900 : };
3901 0 : let Some(retain_lsn) = timeline_entry.ancestor_retain_lsn else {
3902 0 : return;
3903 : };
3904 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
3905 0 : ancestor_children.push((retain_lsn, *timeline_id, MaybeOffloaded::Yes));
3906 0 : });
3907 0 :
3908 0 : // The number of bytes we always keep, irrespective of PITR: this is a constant across timelines
3909 0 : let horizon = self.get_gc_horizon();
3910 :
3911 : // Populate each timeline's GcInfo with information about its child branches
3912 0 : let timelines_to_write = if let Some(timeline_id) = restrict_to_timeline {
3913 0 : itertools::Either::Left(timelines.get(&timeline_id).into_iter())
3914 : } else {
3915 0 : itertools::Either::Right(timelines.values())
3916 : };
3917 0 : for timeline in timelines_to_write {
3918 0 : let mut branchpoints: Vec<(Lsn, TimelineId, MaybeOffloaded)> = all_branchpoints
3919 0 : .remove(&timeline.timeline_id)
3920 0 : .unwrap_or_default();
3921 0 :
3922 0 : branchpoints.sort_by_key(|b| b.0);
3923 0 :
3924 0 : let mut target = timeline.gc_info.write().unwrap();
3925 0 :
3926 0 : target.retain_lsns = branchpoints;
3927 0 :
3928 0 : let space_cutoff = timeline
3929 0 : .get_last_record_lsn()
3930 0 : .checked_sub(horizon)
3931 0 : .unwrap_or(Lsn(0));
3932 0 :
3933 0 : target.cutoffs = GcCutoffs {
3934 0 : space: space_cutoff,
3935 0 : time: Lsn::INVALID,
3936 0 : };
3937 0 : }
3938 0 : }
3939 :
3940 4 : async fn refresh_gc_info_internal(
3941 4 : &self,
3942 4 : target_timeline_id: Option<TimelineId>,
3943 4 : horizon: u64,
3944 4 : pitr: Duration,
3945 4 : cancel: &CancellationToken,
3946 4 : ctx: &RequestContext,
3947 4 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
3948 4 : // before taking the gc_cs lock, do the heavier weight finding of gc_cutoff points for
3949 4 : // currently visible timelines.
3950 4 : let timelines = self
3951 4 : .timelines
3952 4 : .lock()
3953 4 : .unwrap()
3954 4 : .values()
3955 4 : .filter(|tl| match target_timeline_id.as_ref() {
3956 4 : Some(target) => &tl.timeline_id == target,
3957 0 : None => true,
3958 4 : })
3959 4 : .cloned()
3960 4 : .collect::<Vec<_>>();
3961 4 :
3962 4 : if target_timeline_id.is_some() && timelines.is_empty() {
3963 : // We were to act on a particular timeline and it wasn't found
3964 0 : return Err(GcError::TimelineNotFound);
3965 4 : }
3966 4 :
3967 4 : let mut gc_cutoffs: HashMap<TimelineId, GcCutoffs> =
3968 4 : HashMap::with_capacity(timelines.len());
3969 :
3970 4 : for timeline in timelines.iter() {
3971 4 : let cutoff = timeline
3972 4 : .get_last_record_lsn()
3973 4 : .checked_sub(horizon)
3974 4 : .unwrap_or(Lsn(0));
3975 :
3976 4 : let cutoffs = timeline.find_gc_cutoffs(cutoff, pitr, cancel, ctx).await?;
3977 4 : let old = gc_cutoffs.insert(timeline.timeline_id, cutoffs);
3978 4 : assert!(old.is_none());
3979 : }
3980 :
3981 4 : if !self.is_active() || self.cancel.is_cancelled() {
3982 0 : return Err(GcError::TenantCancelled);
3983 4 : }
3984 :
3985 : // grab mutex to prevent new timelines from being created here; avoid doing long operations
3986 : // because that will stall branch creation.
3987 4 : let gc_cs = self.gc_cs.lock().await;
3988 :
3989 : // Ok, we now know all the branch points.
3990 : // Update the GC information for each timeline.
3991 4 : let mut gc_timelines = Vec::with_capacity(timelines.len());
3992 8 : for timeline in timelines {
3993 : // We filtered the timeline list above
3994 4 : if let Some(target_timeline_id) = target_timeline_id {
3995 4 : assert_eq!(target_timeline_id, timeline.timeline_id);
3996 0 : }
3997 :
3998 : {
3999 4 : let mut target = timeline.gc_info.write().unwrap();
4000 4 :
4001 4 : // Cull any expired leases
4002 4 : let now = SystemTime::now();
4003 6 : target.leases.retain(|_, lease| !lease.is_expired(&now));
4004 4 :
4005 4 : timeline
4006 4 : .metrics
4007 4 : .valid_lsn_lease_count_gauge
4008 4 : .set(target.leases.len() as u64);
4009 :
4010 : // Look up parent's PITR cutoff to update the child's knowledge of whether it is within parent's PITR
4011 4 : if let Some(ancestor_id) = timeline.get_ancestor_timeline_id() {
4012 0 : if let Some(ancestor_gc_cutoffs) = gc_cutoffs.get(&ancestor_id) {
4013 0 : target.within_ancestor_pitr =
4014 0 : timeline.get_ancestor_lsn() >= ancestor_gc_cutoffs.time;
4015 0 : }
4016 4 : }
4017 :
4018 : // Update metrics that depend on GC state
4019 4 : timeline
4020 4 : .metrics
4021 4 : .archival_size
4022 4 : .set(if target.within_ancestor_pitr {
4023 0 : timeline.metrics.current_logical_size_gauge.get()
4024 : } else {
4025 4 : 0
4026 : });
4027 4 : timeline.metrics.pitr_history_size.set(
4028 4 : timeline
4029 4 : .get_last_record_lsn()
4030 4 : .checked_sub(target.cutoffs.time)
4031 4 : .unwrap_or(Lsn(0))
4032 4 : .0,
4033 4 : );
4034 :
4035 : // Apply the cutoffs we found to the Timeline's GcInfo. Why might we _not_ have cutoffs for a timeline?
4036 : // - this timeline was created while we were finding cutoffs
4037 : // - lsn for timestamp search fails for this timeline repeatedly
4038 4 : if let Some(cutoffs) = gc_cutoffs.get(&timeline.timeline_id) {
4039 4 : target.cutoffs = cutoffs.clone();
4040 4 : }
4041 : }
4042 :
4043 4 : gc_timelines.push(timeline);
4044 : }
4045 4 : drop(gc_cs);
4046 4 : Ok(gc_timelines)
4047 4 : }
4048 :
4049 : /// A substitute for `branch_timeline` for use in unit tests.
4050 : /// The returned timeline will have state value `Active` to make various `anyhow::ensure!()`
4051 : /// calls pass, but, we do not actually call `.activate()` under the hood. So, none of the
4052 : /// timeline background tasks are launched, except the flush loop.
4053 : #[cfg(test)]
4054 232 : async fn branch_timeline_test(
4055 232 : self: &Arc<Self>,
4056 232 : src_timeline: &Arc<Timeline>,
4057 232 : dst_id: TimelineId,
4058 232 : ancestor_lsn: Option<Lsn>,
4059 232 : ctx: &RequestContext,
4060 232 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
4061 232 : let tl = self
4062 232 : .branch_timeline_impl(src_timeline, dst_id, ancestor_lsn, ctx)
4063 185 : .await?
4064 228 : .into_timeline_for_test();
4065 228 : tl.set_state(TimelineState::Active);
4066 228 : Ok(tl)
4067 232 : }
4068 :
4069 : /// Helper for unit tests to branch a timeline with some pre-loaded states.
4070 : #[cfg(test)]
4071 : #[allow(clippy::too_many_arguments)]
4072 6 : pub async fn branch_timeline_test_with_layers(
4073 6 : self: &Arc<Self>,
4074 6 : src_timeline: &Arc<Timeline>,
4075 6 : dst_id: TimelineId,
4076 6 : ancestor_lsn: Option<Lsn>,
4077 6 : ctx: &RequestContext,
4078 6 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
4079 6 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
4080 6 : end_lsn: Lsn,
4081 6 : ) -> anyhow::Result<Arc<Timeline>> {
4082 : use checks::check_valid_layermap;
4083 : use itertools::Itertools;
4084 :
4085 6 : let tline = self
4086 6 : .branch_timeline_test(src_timeline, dst_id, ancestor_lsn, ctx)
4087 6 : .await?;
4088 6 : let ancestor_lsn = if let Some(ancestor_lsn) = ancestor_lsn {
4089 6 : ancestor_lsn
4090 : } else {
4091 0 : tline.get_last_record_lsn()
4092 : };
4093 6 : assert!(end_lsn >= ancestor_lsn);
4094 6 : tline.force_advance_lsn(end_lsn);
4095 12 : for deltas in delta_layer_desc {
4096 6 : tline
4097 6 : .force_create_delta_layer(deltas, Some(ancestor_lsn), ctx)
4098 18 : .await?;
4099 : }
4100 10 : for (lsn, images) in image_layer_desc {
4101 4 : tline
4102 4 : .force_create_image_layer(lsn, images, Some(ancestor_lsn), ctx)
4103 14 : .await?;
4104 : }
4105 6 : let layer_names = tline
4106 6 : .layers
4107 6 : .read()
4108 0 : .await
4109 6 : .layer_map()
4110 6 : .unwrap()
4111 6 : .iter_historic_layers()
4112 10 : .map(|layer| layer.layer_name())
4113 6 : .collect_vec();
4114 6 : if let Some(err) = check_valid_layermap(&layer_names) {
4115 0 : bail!("invalid layermap: {err}");
4116 6 : }
4117 6 : Ok(tline)
4118 6 : }
4119 :
4120 : /// Branch an existing timeline.
4121 0 : async fn branch_timeline(
4122 0 : self: &Arc<Self>,
4123 0 : src_timeline: &Arc<Timeline>,
4124 0 : dst_id: TimelineId,
4125 0 : start_lsn: Option<Lsn>,
4126 0 : ctx: &RequestContext,
4127 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4128 0 : self.branch_timeline_impl(src_timeline, dst_id, start_lsn, ctx)
4129 0 : .await
4130 0 : }
4131 :
4132 232 : async fn branch_timeline_impl(
4133 232 : self: &Arc<Self>,
4134 232 : src_timeline: &Arc<Timeline>,
4135 232 : dst_id: TimelineId,
4136 232 : start_lsn: Option<Lsn>,
4137 232 : _ctx: &RequestContext,
4138 232 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4139 232 : let src_id = src_timeline.timeline_id;
4140 :
4141 : // We will validate our ancestor LSN in this function. Acquire the GC lock so that
4142 : // this check cannot race with GC, and the ancestor LSN is guaranteed to remain
4143 : // valid while we are creating the branch.
4144 232 : let _gc_cs = self.gc_cs.lock().await;
4145 :
4146 : // If no start LSN is specified, we branch the new timeline from the source timeline's last record LSN
4147 232 : let start_lsn = start_lsn.unwrap_or_else(|| {
4148 2 : let lsn = src_timeline.get_last_record_lsn();
4149 2 : info!("branching timeline {dst_id} from timeline {src_id} at last record LSN: {lsn}");
4150 2 : lsn
4151 232 : });
4152 :
4153 : // we finally have determined the ancestor_start_lsn, so we can get claim exclusivity now
4154 232 : let timeline_create_guard = match self
4155 232 : .start_creating_timeline(
4156 232 : dst_id,
4157 232 : CreateTimelineIdempotency::Branch {
4158 232 : ancestor_timeline_id: src_timeline.timeline_id,
4159 232 : ancestor_start_lsn: start_lsn,
4160 232 : },
4161 232 : )
4162 185 : .await?
4163 : {
4164 232 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
4165 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
4166 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
4167 : }
4168 : };
4169 :
4170 : // Ensure that `start_lsn` is valid, i.e. the LSN is within the PITR
4171 : // horizon on the source timeline
4172 : //
4173 : // We check it against both the planned GC cutoff stored in 'gc_info',
4174 : // and the 'latest_gc_cutoff' of the last GC that was performed. The
4175 : // planned GC cutoff in 'gc_info' is normally larger than
4176 : // 'latest_gc_cutoff_lsn', but beware of corner cases like if you just
4177 : // changed the GC settings for the tenant to make the PITR window
4178 : // larger, but some of the data was already removed by an earlier GC
4179 : // iteration.
4180 :
4181 : // check against last actual 'latest_gc_cutoff' first
4182 232 : let latest_gc_cutoff_lsn = src_timeline.get_latest_gc_cutoff_lsn();
4183 232 : src_timeline
4184 232 : .check_lsn_is_in_scope(start_lsn, &latest_gc_cutoff_lsn)
4185 232 : .context(format!(
4186 232 : "invalid branch start lsn: less than latest GC cutoff {}",
4187 232 : *latest_gc_cutoff_lsn,
4188 232 : ))
4189 232 : .map_err(CreateTimelineError::AncestorLsn)?;
4190 :
4191 : // and then the planned GC cutoff
4192 : {
4193 228 : let gc_info = src_timeline.gc_info.read().unwrap();
4194 228 : let cutoff = gc_info.min_cutoff();
4195 228 : if start_lsn < cutoff {
4196 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
4197 0 : "invalid branch start lsn: less than planned GC cutoff {cutoff}"
4198 0 : )));
4199 228 : }
4200 228 : }
4201 228 :
4202 228 : //
4203 228 : // The branch point is valid, and we are still holding the 'gc_cs' lock
4204 228 : // so that GC cannot advance the GC cutoff until we are finished.
4205 228 : // Proceed with the branch creation.
4206 228 : //
4207 228 :
4208 228 : // Determine prev-LSN for the new timeline. We can only determine it if
4209 228 : // the timeline was branched at the current end of the source timeline.
4210 228 : let RecordLsn {
4211 228 : last: src_last,
4212 228 : prev: src_prev,
4213 228 : } = src_timeline.get_last_record_rlsn();
4214 228 : let dst_prev = if src_last == start_lsn {
4215 216 : Some(src_prev)
4216 : } else {
4217 12 : None
4218 : };
4219 :
4220 : // Create the metadata file, noting the ancestor of the new timeline.
4221 : // There is initially no data in it, but all the read-calls know to look
4222 : // into the ancestor.
4223 228 : let metadata = TimelineMetadata::new(
4224 228 : start_lsn,
4225 228 : dst_prev,
4226 228 : Some(src_id),
4227 228 : start_lsn,
4228 228 : *src_timeline.latest_gc_cutoff_lsn.read(), // FIXME: should we hold onto this guard longer?
4229 228 : src_timeline.initdb_lsn,
4230 228 : src_timeline.pg_version,
4231 228 : );
4232 :
4233 228 : let uninitialized_timeline = self
4234 228 : .prepare_new_timeline(
4235 228 : dst_id,
4236 228 : &metadata,
4237 228 : timeline_create_guard,
4238 228 : start_lsn + 1,
4239 228 : Some(Arc::clone(src_timeline)),
4240 228 : )
4241 0 : .await?;
4242 :
4243 228 : let new_timeline = uninitialized_timeline.finish_creation()?;
4244 :
4245 : // Root timeline gets its layers during creation and uploads them along with the metadata.
4246 : // A branch timeline though, when created, can get no writes for some time, hence won't get any layers created.
4247 : // We still need to upload its metadata eagerly: if other nodes `attach` the tenant and miss this timeline, their GC
4248 : // could get incorrect information and remove more layers, than needed.
4249 : // See also https://github.com/neondatabase/neon/issues/3865
4250 228 : new_timeline
4251 228 : .remote_client
4252 228 : .schedule_index_upload_for_full_metadata_update(&metadata)
4253 228 : .context("branch initial metadata upload")?;
4254 :
4255 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
4256 :
4257 228 : Ok(CreateTimelineResult::Created(new_timeline))
4258 232 : }
4259 :
4260 : /// For unit tests, make this visible so that other modules can directly create timelines
4261 : #[cfg(test)]
4262 2 : #[tracing::instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
4263 : pub(crate) async fn bootstrap_timeline_test(
4264 : self: &Arc<Self>,
4265 : timeline_id: TimelineId,
4266 : pg_version: u32,
4267 : load_existing_initdb: Option<TimelineId>,
4268 : ctx: &RequestContext,
4269 : ) -> anyhow::Result<Arc<Timeline>> {
4270 : self.bootstrap_timeline(timeline_id, pg_version, load_existing_initdb, ctx)
4271 : .await
4272 : .map_err(anyhow::Error::new)
4273 2 : .map(|r| r.into_timeline_for_test())
4274 : }
4275 :
4276 : /// Get exclusive access to the timeline ID for creation.
4277 : ///
4278 : /// Timeline-creating code paths must use this function before making changes
4279 : /// to in-memory or persistent state.
4280 : ///
4281 : /// The `state` parameter is a description of the timeline creation operation
4282 : /// we intend to perform.
4283 : /// If the timeline was already created in the meantime, we check whether this
4284 : /// request conflicts or is idempotent , based on `state`.
4285 418 : async fn start_creating_timeline(
4286 418 : &self,
4287 418 : new_timeline_id: TimelineId,
4288 418 : idempotency: CreateTimelineIdempotency,
4289 418 : ) -> Result<StartCreatingTimelineResult<'_>, CreateTimelineError> {
4290 418 : let allow_offloaded = false;
4291 418 : match self.create_timeline_create_guard(new_timeline_id, idempotency, allow_offloaded) {
4292 416 : Ok(create_guard) => {
4293 416 : pausable_failpoint!("timeline-creation-after-uninit");
4294 416 : Ok(StartCreatingTimelineResult::CreateGuard(create_guard))
4295 : }
4296 : Err(TimelineExclusionError::AlreadyCreating) => {
4297 : // Creation is in progress, we cannot create it again, and we cannot
4298 : // check if this request matches the existing one, so caller must try
4299 : // again later.
4300 0 : Err(CreateTimelineError::AlreadyCreating)
4301 : }
4302 0 : Err(TimelineExclusionError::Other(e)) => Err(CreateTimelineError::Other(e)),
4303 : Err(TimelineExclusionError::AlreadyExists {
4304 0 : existing: TimelineOrOffloaded::Offloaded(_existing),
4305 0 : ..
4306 0 : }) => {
4307 0 : info!("timeline already exists but is offloaded");
4308 0 : Err(CreateTimelineError::Conflict)
4309 : }
4310 : Err(TimelineExclusionError::AlreadyExists {
4311 2 : existing: TimelineOrOffloaded::Timeline(existing),
4312 2 : arg,
4313 2 : }) => {
4314 2 : {
4315 2 : let existing = &existing.create_idempotency;
4316 2 : let _span = info_span!("idempotency_check", ?existing, ?arg).entered();
4317 2 : debug!("timeline already exists");
4318 :
4319 2 : match (existing, &arg) {
4320 : // FailWithConflict => no idempotency check
4321 : (CreateTimelineIdempotency::FailWithConflict, _)
4322 : | (_, CreateTimelineIdempotency::FailWithConflict) => {
4323 2 : warn!("timeline already exists, failing request");
4324 2 : return Err(CreateTimelineError::Conflict);
4325 : }
4326 : // Idempotent <=> CreateTimelineIdempotency is identical
4327 0 : (x, y) if x == y => {
4328 0 : info!("timeline already exists and idempotency matches, succeeding request");
4329 : // fallthrough
4330 : }
4331 : (_, _) => {
4332 0 : warn!("idempotency conflict, failing request");
4333 0 : return Err(CreateTimelineError::Conflict);
4334 : }
4335 : }
4336 : }
4337 :
4338 0 : Ok(StartCreatingTimelineResult::Idempotent(existing))
4339 : }
4340 : }
4341 418 : }
4342 :
4343 0 : async fn upload_initdb(
4344 0 : &self,
4345 0 : timelines_path: &Utf8PathBuf,
4346 0 : pgdata_path: &Utf8PathBuf,
4347 0 : timeline_id: &TimelineId,
4348 0 : ) -> anyhow::Result<()> {
4349 0 : let temp_path = timelines_path.join(format!(
4350 0 : "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
4351 0 : ));
4352 0 :
4353 0 : scopeguard::defer! {
4354 0 : if let Err(e) = fs::remove_file(&temp_path) {
4355 0 : error!("Failed to remove temporary initdb archive '{temp_path}': {e}");
4356 0 : }
4357 0 : }
4358 :
4359 0 : let (pgdata_zstd, tar_zst_size) = create_zst_tarball(pgdata_path, &temp_path).await?;
4360 : const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
4361 0 : if tar_zst_size > INITDB_TAR_ZST_WARN_LIMIT {
4362 0 : warn!(
4363 0 : "compressed {temp_path} size of {tar_zst_size} is above limit {INITDB_TAR_ZST_WARN_LIMIT}."
4364 : );
4365 0 : }
4366 :
4367 0 : pausable_failpoint!("before-initdb-upload");
4368 :
4369 0 : backoff::retry(
4370 0 : || async {
4371 0 : self::remote_timeline_client::upload_initdb_dir(
4372 0 : &self.remote_storage,
4373 0 : &self.tenant_shard_id.tenant_id,
4374 0 : timeline_id,
4375 0 : pgdata_zstd.try_clone().await?,
4376 0 : tar_zst_size,
4377 0 : &self.cancel,
4378 : )
4379 0 : .await
4380 0 : },
4381 0 : |_| false,
4382 0 : 3,
4383 0 : u32::MAX,
4384 0 : "persist_initdb_tar_zst",
4385 0 : &self.cancel,
4386 0 : )
4387 0 : .await
4388 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
4389 0 : .and_then(|x| x)
4390 0 : }
4391 :
4392 : /// - run initdb to init temporary instance and get bootstrap data
4393 : /// - after initialization completes, tar up the temp dir and upload it to S3.
4394 2 : async fn bootstrap_timeline(
4395 2 : self: &Arc<Self>,
4396 2 : timeline_id: TimelineId,
4397 2 : pg_version: u32,
4398 2 : load_existing_initdb: Option<TimelineId>,
4399 2 : ctx: &RequestContext,
4400 2 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4401 2 : let timeline_create_guard = match self
4402 2 : .start_creating_timeline(
4403 2 : timeline_id,
4404 2 : CreateTimelineIdempotency::Bootstrap { pg_version },
4405 2 : )
4406 2 : .await?
4407 : {
4408 2 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
4409 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
4410 0 : return Ok(CreateTimelineResult::Idempotent(timeline))
4411 : }
4412 : };
4413 :
4414 : // create a `tenant/{tenant_id}/timelines/basebackup-{timeline_id}.{TEMP_FILE_SUFFIX}/`
4415 : // temporary directory for basebackup files for the given timeline.
4416 :
4417 2 : let timelines_path = self.conf.timelines_path(&self.tenant_shard_id);
4418 2 : let pgdata_path = path_with_suffix_extension(
4419 2 : timelines_path.join(format!("basebackup-{timeline_id}")),
4420 2 : TEMP_FILE_SUFFIX,
4421 2 : );
4422 2 :
4423 2 : // Remove whatever was left from the previous runs: safe because TimelineCreateGuard guarantees
4424 2 : // we won't race with other creations or existent timelines with the same path.
4425 2 : if pgdata_path.exists() {
4426 0 : fs::remove_dir_all(&pgdata_path).with_context(|| {
4427 0 : format!("Failed to remove already existing initdb directory: {pgdata_path}")
4428 0 : })?;
4429 2 : }
4430 :
4431 : // this new directory is very temporary, set to remove it immediately after bootstrap, we don't need it
4432 2 : scopeguard::defer! {
4433 2 : if let Err(e) = fs::remove_dir_all(&pgdata_path) {
4434 2 : // this is unlikely, but we will remove the directory on pageserver restart or another bootstrap call
4435 2 : error!("Failed to remove temporary initdb directory '{pgdata_path}': {e}");
4436 2 : }
4437 2 : }
4438 2 : if let Some(existing_initdb_timeline_id) = load_existing_initdb {
4439 2 : if existing_initdb_timeline_id != timeline_id {
4440 0 : let source_path = &remote_initdb_archive_path(
4441 0 : &self.tenant_shard_id.tenant_id,
4442 0 : &existing_initdb_timeline_id,
4443 0 : );
4444 0 : let dest_path =
4445 0 : &remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &timeline_id);
4446 0 :
4447 0 : // if this fails, it will get retried by retried control plane requests
4448 0 : self.remote_storage
4449 0 : .copy_object(source_path, dest_path, &self.cancel)
4450 0 : .await
4451 0 : .context("copy initdb tar")?;
4452 2 : }
4453 2 : let (initdb_tar_zst_path, initdb_tar_zst) =
4454 2 : self::remote_timeline_client::download_initdb_tar_zst(
4455 2 : self.conf,
4456 2 : &self.remote_storage,
4457 2 : &self.tenant_shard_id,
4458 2 : &existing_initdb_timeline_id,
4459 2 : &self.cancel,
4460 2 : )
4461 489 : .await
4462 2 : .context("download initdb tar")?;
4463 :
4464 2 : scopeguard::defer! {
4465 2 : if let Err(e) = fs::remove_file(&initdb_tar_zst_path) {
4466 2 : error!("Failed to remove temporary initdb archive '{initdb_tar_zst_path}': {e}");
4467 2 : }
4468 2 : }
4469 2 :
4470 2 : let buf_read =
4471 2 : BufReader::with_capacity(remote_timeline_client::BUFFER_SIZE, initdb_tar_zst);
4472 2 : extract_zst_tarball(&pgdata_path, buf_read)
4473 11023 : .await
4474 2 : .context("extract initdb tar")?;
4475 : } else {
4476 : // Init temporarily repo to get bootstrap data, this creates a directory in the `pgdata_path` path
4477 0 : run_initdb(self.conf, &pgdata_path, pg_version, &self.cancel)
4478 0 : .await
4479 0 : .context("run initdb")?;
4480 :
4481 : // Upload the created data dir to S3
4482 0 : if self.tenant_shard_id().is_shard_zero() {
4483 0 : self.upload_initdb(&timelines_path, &pgdata_path, &timeline_id)
4484 0 : .await?;
4485 0 : }
4486 : }
4487 2 : let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
4488 2 :
4489 2 : // Import the contents of the data directory at the initial checkpoint
4490 2 : // LSN, and any WAL after that.
4491 2 : // Initdb lsn will be equal to last_record_lsn which will be set after import.
4492 2 : // Because we know it upfront avoid having an option or dummy zero value by passing it to the metadata.
4493 2 : let new_metadata = TimelineMetadata::new(
4494 2 : Lsn(0),
4495 2 : None,
4496 2 : None,
4497 2 : Lsn(0),
4498 2 : pgdata_lsn,
4499 2 : pgdata_lsn,
4500 2 : pg_version,
4501 2 : );
4502 2 : let raw_timeline = self
4503 2 : .prepare_new_timeline(
4504 2 : timeline_id,
4505 2 : &new_metadata,
4506 2 : timeline_create_guard,
4507 2 : pgdata_lsn,
4508 2 : None,
4509 2 : )
4510 0 : .await?;
4511 :
4512 2 : let tenant_shard_id = raw_timeline.owning_tenant.tenant_shard_id;
4513 2 : let unfinished_timeline = raw_timeline.raw_timeline()?;
4514 :
4515 : // Flush the new layer files to disk, before we make the timeline as available to
4516 : // the outside world.
4517 : //
4518 : // Flush loop needs to be spawned in order to be able to flush.
4519 2 : unfinished_timeline.maybe_spawn_flush_loop();
4520 2 :
4521 2 : import_datadir::import_timeline_from_postgres_datadir(
4522 2 : unfinished_timeline,
4523 2 : &pgdata_path,
4524 2 : pgdata_lsn,
4525 2 : ctx,
4526 2 : )
4527 9481 : .await
4528 2 : .with_context(|| {
4529 0 : format!("Failed to import pgdatadir for timeline {tenant_shard_id}/{timeline_id}")
4530 2 : })?;
4531 :
4532 2 : fail::fail_point!("before-checkpoint-new-timeline", |_| {
4533 0 : Err(CreateTimelineError::Other(anyhow::anyhow!(
4534 0 : "failpoint before-checkpoint-new-timeline"
4535 0 : )))
4536 2 : });
4537 :
4538 2 : unfinished_timeline
4539 2 : .freeze_and_flush()
4540 2 : .await
4541 2 : .with_context(|| {
4542 0 : format!(
4543 0 : "Failed to flush after pgdatadir import for timeline {tenant_shard_id}/{timeline_id}"
4544 0 : )
4545 2 : })?;
4546 :
4547 : // All done!
4548 2 : let timeline = raw_timeline.finish_creation()?;
4549 :
4550 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
4551 :
4552 2 : Ok(CreateTimelineResult::Created(timeline))
4553 2 : }
4554 :
4555 412 : fn build_timeline_remote_client(&self, timeline_id: TimelineId) -> RemoteTimelineClient {
4556 412 : RemoteTimelineClient::new(
4557 412 : self.remote_storage.clone(),
4558 412 : self.deletion_queue_client.clone(),
4559 412 : self.conf,
4560 412 : self.tenant_shard_id,
4561 412 : timeline_id,
4562 412 : self.generation,
4563 412 : &self.tenant_conf.load().location,
4564 412 : )
4565 412 : }
4566 :
4567 : /// Call this before constructing a timeline, to build its required structures
4568 412 : fn build_timeline_resources(&self, timeline_id: TimelineId) -> TimelineResources {
4569 412 : TimelineResources {
4570 412 : remote_client: self.build_timeline_remote_client(timeline_id),
4571 412 : timeline_get_throttle: self.timeline_get_throttle.clone(),
4572 412 : l0_flush_global_state: self.l0_flush_global_state.clone(),
4573 412 : }
4574 412 : }
4575 :
4576 : /// Creates intermediate timeline structure and its files.
4577 : ///
4578 : /// An empty layer map is initialized, and new data and WAL can be imported starting
4579 : /// at 'disk_consistent_lsn'. After any initial data has been imported, call
4580 : /// `finish_creation` to insert the Timeline into the timelines map.
4581 412 : async fn prepare_new_timeline<'a>(
4582 412 : &'a self,
4583 412 : new_timeline_id: TimelineId,
4584 412 : new_metadata: &TimelineMetadata,
4585 412 : create_guard: TimelineCreateGuard<'a>,
4586 412 : start_lsn: Lsn,
4587 412 : ancestor: Option<Arc<Timeline>>,
4588 412 : ) -> anyhow::Result<UninitializedTimeline<'a>> {
4589 412 : let tenant_shard_id = self.tenant_shard_id;
4590 412 :
4591 412 : let resources = self.build_timeline_resources(new_timeline_id);
4592 412 : resources
4593 412 : .remote_client
4594 412 : .init_upload_queue_for_empty_remote(new_metadata)?;
4595 :
4596 412 : let timeline_struct = self
4597 412 : .create_timeline_struct(
4598 412 : new_timeline_id,
4599 412 : new_metadata,
4600 412 : ancestor,
4601 412 : resources,
4602 412 : CreateTimelineCause::Load,
4603 412 : create_guard.idempotency.clone(),
4604 412 : )
4605 412 : .context("Failed to create timeline data structure")?;
4606 :
4607 412 : timeline_struct.init_empty_layer_map(start_lsn);
4608 :
4609 412 : if let Err(e) = self
4610 412 : .create_timeline_files(&create_guard.timeline_path)
4611 0 : .await
4612 : {
4613 0 : error!("Failed to create initial files for timeline {tenant_shard_id}/{new_timeline_id}, cleaning up: {e:?}");
4614 0 : cleanup_timeline_directory(create_guard);
4615 0 : return Err(e);
4616 412 : }
4617 412 :
4618 412 : debug!(
4619 0 : "Successfully created initial files for timeline {tenant_shard_id}/{new_timeline_id}"
4620 : );
4621 :
4622 412 : Ok(UninitializedTimeline::new(
4623 412 : self,
4624 412 : new_timeline_id,
4625 412 : Some((timeline_struct, create_guard)),
4626 412 : ))
4627 412 : }
4628 :
4629 412 : async fn create_timeline_files(&self, timeline_path: &Utf8Path) -> anyhow::Result<()> {
4630 412 : crashsafe::create_dir(timeline_path).context("Failed to create timeline directory")?;
4631 :
4632 412 : fail::fail_point!("after-timeline-dir-creation", |_| {
4633 0 : anyhow::bail!("failpoint after-timeline-dir-creation");
4634 412 : });
4635 :
4636 412 : Ok(())
4637 412 : }
4638 :
4639 : /// Get a guard that provides exclusive access to the timeline directory, preventing
4640 : /// concurrent attempts to create the same timeline.
4641 : ///
4642 : /// The `allow_offloaded` parameter controls whether to tolerate the existence of
4643 : /// offloaded timelines or not.
4644 418 : fn create_timeline_create_guard(
4645 418 : &self,
4646 418 : timeline_id: TimelineId,
4647 418 : idempotency: CreateTimelineIdempotency,
4648 418 : allow_offloaded: bool,
4649 418 : ) -> Result<TimelineCreateGuard, TimelineExclusionError> {
4650 418 : let tenant_shard_id = self.tenant_shard_id;
4651 418 :
4652 418 : let timeline_path = self.conf.timeline_path(&tenant_shard_id, &timeline_id);
4653 :
4654 418 : let create_guard = TimelineCreateGuard::new(
4655 418 : self,
4656 418 : timeline_id,
4657 418 : timeline_path.clone(),
4658 418 : idempotency,
4659 418 : allow_offloaded,
4660 418 : )?;
4661 :
4662 : // At this stage, we have got exclusive access to in-memory state for this timeline ID
4663 : // for creation.
4664 : // A timeline directory should never exist on disk already:
4665 : // - a previous failed creation would have cleaned up after itself
4666 : // - a pageserver restart would clean up timeline directories that don't have valid remote state
4667 : //
4668 : // Therefore it is an unexpected internal error to encounter a timeline directory already existing here,
4669 : // this error may indicate a bug in cleanup on failed creations.
4670 416 : if timeline_path.exists() {
4671 0 : return Err(TimelineExclusionError::Other(anyhow::anyhow!(
4672 0 : "Timeline directory already exists! This is a bug."
4673 0 : )));
4674 416 : }
4675 416 :
4676 416 : Ok(create_guard)
4677 418 : }
4678 :
4679 : /// Gathers inputs from all of the timelines to produce a sizing model input.
4680 : ///
4681 : /// Future is cancellation safe. Only one calculation can be running at once per tenant.
4682 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
4683 : pub async fn gather_size_inputs(
4684 : &self,
4685 : // `max_retention_period` overrides the cutoff that is used to calculate the size
4686 : // (only if it is shorter than the real cutoff).
4687 : max_retention_period: Option<u64>,
4688 : cause: LogicalSizeCalculationCause,
4689 : cancel: &CancellationToken,
4690 : ctx: &RequestContext,
4691 : ) -> Result<size::ModelInputs, size::CalculateSyntheticSizeError> {
4692 : let logical_sizes_at_once = self
4693 : .conf
4694 : .concurrent_tenant_size_logical_size_queries
4695 : .inner();
4696 :
4697 : // TODO: Having a single mutex block concurrent reads is not great for performance.
4698 : //
4699 : // But the only case where we need to run multiple of these at once is when we
4700 : // request a size for a tenant manually via API, while another background calculation
4701 : // is in progress (which is not a common case).
4702 : //
4703 : // See more for on the issue #2748 condenced out of the initial PR review.
4704 : let mut shared_cache = tokio::select! {
4705 : locked = self.cached_logical_sizes.lock() => locked,
4706 : _ = cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
4707 : _ = self.cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
4708 : };
4709 :
4710 : size::gather_inputs(
4711 : self,
4712 : logical_sizes_at_once,
4713 : max_retention_period,
4714 : &mut shared_cache,
4715 : cause,
4716 : cancel,
4717 : ctx,
4718 : )
4719 : .await
4720 : }
4721 :
4722 : /// Calculate synthetic tenant size and cache the result.
4723 : /// This is periodically called by background worker.
4724 : /// result is cached in tenant struct
4725 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
4726 : pub async fn calculate_synthetic_size(
4727 : &self,
4728 : cause: LogicalSizeCalculationCause,
4729 : cancel: &CancellationToken,
4730 : ctx: &RequestContext,
4731 : ) -> Result<u64, size::CalculateSyntheticSizeError> {
4732 : let inputs = self.gather_size_inputs(None, cause, cancel, ctx).await?;
4733 :
4734 : let size = inputs.calculate();
4735 :
4736 : self.set_cached_synthetic_size(size);
4737 :
4738 : Ok(size)
4739 : }
4740 :
4741 : /// Cache given synthetic size and update the metric value
4742 0 : pub fn set_cached_synthetic_size(&self, size: u64) {
4743 0 : self.cached_synthetic_tenant_size
4744 0 : .store(size, Ordering::Relaxed);
4745 0 :
4746 0 : // Only shard zero should be calculating synthetic sizes
4747 0 : debug_assert!(self.shard_identity.is_shard_zero());
4748 :
4749 0 : TENANT_SYNTHETIC_SIZE_METRIC
4750 0 : .get_metric_with_label_values(&[&self.tenant_shard_id.tenant_id.to_string()])
4751 0 : .unwrap()
4752 0 : .set(size);
4753 0 : }
4754 :
4755 0 : pub fn cached_synthetic_size(&self) -> u64 {
4756 0 : self.cached_synthetic_tenant_size.load(Ordering::Relaxed)
4757 0 : }
4758 :
4759 : /// Flush any in-progress layers, schedule uploads, and wait for uploads to complete.
4760 : ///
4761 : /// This function can take a long time: callers should wrap it in a timeout if calling
4762 : /// from an external API handler.
4763 : ///
4764 : /// Cancel-safety: cancelling this function may leave I/O running, but such I/O is
4765 : /// still bounded by tenant/timeline shutdown.
4766 0 : #[tracing::instrument(skip_all)]
4767 : pub(crate) async fn flush_remote(&self) -> anyhow::Result<()> {
4768 : let timelines = self.timelines.lock().unwrap().clone();
4769 :
4770 0 : async fn flush_timeline(_gate: GateGuard, timeline: Arc<Timeline>) -> anyhow::Result<()> {
4771 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Flushing...");
4772 0 : timeline.freeze_and_flush().await?;
4773 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Waiting for uploads...");
4774 0 : timeline.remote_client.wait_completion().await?;
4775 :
4776 0 : Ok(())
4777 0 : }
4778 :
4779 : // We do not use a JoinSet for these tasks, because we don't want them to be
4780 : // aborted when this function's future is cancelled: they should stay alive
4781 : // holding their GateGuard until they complete, to ensure their I/Os complete
4782 : // before Timeline shutdown completes.
4783 : let mut results = FuturesUnordered::new();
4784 :
4785 : for (_timeline_id, timeline) in timelines {
4786 : // Run each timeline's flush in a task holding the timeline's gate: this
4787 : // means that if this function's future is cancelled, the Timeline shutdown
4788 : // will still wait for any I/O in here to complete.
4789 : let Ok(gate) = timeline.gate.enter() else {
4790 : continue;
4791 : };
4792 0 : let jh = tokio::task::spawn(async move { flush_timeline(gate, timeline).await });
4793 : results.push(jh);
4794 : }
4795 :
4796 : while let Some(r) = results.next().await {
4797 : if let Err(e) = r {
4798 : if !e.is_cancelled() && !e.is_panic() {
4799 : tracing::error!("unexpected join error: {e:?}");
4800 : }
4801 : }
4802 : }
4803 :
4804 : // The flushes we did above were just writes, but the Tenant might have had
4805 : // pending deletions as well from recent compaction/gc: we want to flush those
4806 : // as well. This requires flushing the global delete queue. This is cheap
4807 : // because it's typically a no-op.
4808 : match self.deletion_queue_client.flush_execute().await {
4809 : Ok(_) => {}
4810 : Err(DeletionQueueError::ShuttingDown) => {}
4811 : }
4812 :
4813 : Ok(())
4814 : }
4815 :
4816 0 : pub(crate) fn get_tenant_conf(&self) -> TenantConfOpt {
4817 0 : self.tenant_conf.load().tenant_conf.clone()
4818 0 : }
4819 :
4820 : /// How much local storage would this tenant like to have? It can cope with
4821 : /// less than this (via eviction and on-demand downloads), but this function enables
4822 : /// the Tenant to advertise how much storage it would prefer to have to provide fast I/O
4823 : /// by keeping important things on local disk.
4824 : ///
4825 : /// This is a heuristic, not a guarantee: tenants that are long-idle will actually use less
4826 : /// than they report here, due to layer eviction. Tenants with many active branches may
4827 : /// actually use more than they report here.
4828 0 : pub(crate) fn local_storage_wanted(&self) -> u64 {
4829 0 : let timelines = self.timelines.lock().unwrap();
4830 0 :
4831 0 : // Heuristic: we use the max() of the timelines' visible sizes, rather than the sum. This
4832 0 : // reflects the observation that on tenants with multiple large branches, typically only one
4833 0 : // of them is used actively enough to occupy space on disk.
4834 0 : timelines
4835 0 : .values()
4836 0 : .map(|t| t.metrics.visible_physical_size_gauge.get())
4837 0 : .max()
4838 0 : .unwrap_or(0)
4839 0 : }
4840 :
4841 : /// Serialize and write the latest TenantManifest to remote storage.
4842 2 : pub(crate) async fn store_tenant_manifest(&self) -> Result<(), TenantManifestError> {
4843 : // Only one manifest write may be done at at time, and the contents of the manifest
4844 : // must be loaded while holding this lock. This makes it safe to call this function
4845 : // from anywhere without worrying about colliding updates.
4846 2 : let mut guard = tokio::select! {
4847 2 : g = self.tenant_manifest_upload.lock() => {
4848 2 : g
4849 : },
4850 2 : _ = self.cancel.cancelled() => {
4851 0 : return Err(TenantManifestError::Cancelled);
4852 : }
4853 : };
4854 :
4855 2 : let manifest = self.build_tenant_manifest();
4856 2 : if Some(&manifest) == (*guard).as_ref() {
4857 : // Optimisation: skip uploads that don't change anything.
4858 0 : return Ok(());
4859 2 : }
4860 2 :
4861 2 : upload_tenant_manifest(
4862 2 : &self.remote_storage,
4863 2 : &self.tenant_shard_id,
4864 2 : self.generation,
4865 2 : &manifest,
4866 2 : &self.cancel,
4867 2 : )
4868 8 : .await
4869 2 : .map_err(|e| {
4870 0 : if self.cancel.is_cancelled() {
4871 0 : TenantManifestError::Cancelled
4872 : } else {
4873 0 : TenantManifestError::RemoteStorage(e)
4874 : }
4875 2 : })?;
4876 :
4877 : // Store the successfully uploaded manifest, so that future callers can avoid
4878 : // re-uploading the same thing.
4879 2 : *guard = Some(manifest);
4880 2 :
4881 2 : Ok(())
4882 2 : }
4883 : }
4884 :
4885 : /// Create the cluster temporarily in 'initdbpath' directory inside the repository
4886 : /// to get bootstrap data for timeline initialization.
4887 0 : async fn run_initdb(
4888 0 : conf: &'static PageServerConf,
4889 0 : initdb_target_dir: &Utf8Path,
4890 0 : pg_version: u32,
4891 0 : cancel: &CancellationToken,
4892 0 : ) -> Result<(), InitdbError> {
4893 0 : let initdb_bin_path = conf
4894 0 : .pg_bin_dir(pg_version)
4895 0 : .map_err(InitdbError::Other)?
4896 0 : .join("initdb");
4897 0 : let initdb_lib_dir = conf.pg_lib_dir(pg_version).map_err(InitdbError::Other)?;
4898 0 : info!(
4899 0 : "running {} in {}, libdir: {}",
4900 : initdb_bin_path, initdb_target_dir, initdb_lib_dir,
4901 : );
4902 :
4903 0 : let _permit = INIT_DB_SEMAPHORE.acquire().await;
4904 :
4905 0 : let mut initdb_command = tokio::process::Command::new(&initdb_bin_path);
4906 0 : initdb_command
4907 0 : .args(["--pgdata", initdb_target_dir.as_ref()])
4908 0 : .args(["--username", &conf.superuser])
4909 0 : .args(["--encoding", "utf8"])
4910 0 : .args(["--locale", &conf.locale])
4911 0 : .arg("--no-instructions")
4912 0 : .arg("--no-sync")
4913 0 : .env_clear()
4914 0 : .env("LD_LIBRARY_PATH", &initdb_lib_dir)
4915 0 : .env("DYLD_LIBRARY_PATH", &initdb_lib_dir)
4916 0 : .stdin(std::process::Stdio::null())
4917 0 : // stdout invocation produces the same output every time, we don't need it
4918 0 : .stdout(std::process::Stdio::null())
4919 0 : // we would be interested in the stderr output, if there was any
4920 0 : .stderr(std::process::Stdio::piped());
4921 0 :
4922 0 : // Before version 14, only the libc provide was available.
4923 0 : if pg_version > 14 {
4924 : // Version 17 brought with it a builtin locale provider which only provides
4925 : // C and C.UTF-8. While being safer for collation purposes since it is
4926 : // guaranteed to be consistent throughout a major release, it is also more
4927 : // performant.
4928 0 : let locale_provider = if pg_version >= 17 { "builtin" } else { "libc" };
4929 :
4930 0 : initdb_command.args(["--locale-provider", locale_provider]);
4931 0 : }
4932 :
4933 0 : let initdb_proc = initdb_command.spawn()?;
4934 :
4935 : // Ideally we'd select here with the cancellation token, but the problem is that
4936 : // we can't safely terminate initdb: it launches processes of its own, and killing
4937 : // initdb doesn't kill them. After we return from this function, we want the target
4938 : // directory to be able to be cleaned up.
4939 : // See https://github.com/neondatabase/neon/issues/6385
4940 0 : let initdb_output = initdb_proc.wait_with_output().await?;
4941 0 : if !initdb_output.status.success() {
4942 0 : return Err(InitdbError::Failed(
4943 0 : initdb_output.status,
4944 0 : initdb_output.stderr,
4945 0 : ));
4946 0 : }
4947 0 :
4948 0 : // This isn't true cancellation support, see above. Still return an error to
4949 0 : // excercise the cancellation code path.
4950 0 : if cancel.is_cancelled() {
4951 0 : return Err(InitdbError::Cancelled);
4952 0 : }
4953 0 :
4954 0 : Ok(())
4955 0 : }
4956 :
4957 : /// Dump contents of a layer file to stdout.
4958 0 : pub async fn dump_layerfile_from_path(
4959 0 : path: &Utf8Path,
4960 0 : verbose: bool,
4961 0 : ctx: &RequestContext,
4962 0 : ) -> anyhow::Result<()> {
4963 : use std::os::unix::fs::FileExt;
4964 :
4965 : // All layer files start with a two-byte "magic" value, to identify the kind of
4966 : // file.
4967 0 : let file = File::open(path)?;
4968 0 : let mut header_buf = [0u8; 2];
4969 0 : file.read_exact_at(&mut header_buf, 0)?;
4970 :
4971 0 : match u16::from_be_bytes(header_buf) {
4972 : crate::IMAGE_FILE_MAGIC => {
4973 0 : ImageLayer::new_for_path(path, file)?
4974 0 : .dump(verbose, ctx)
4975 0 : .await?
4976 : }
4977 : crate::DELTA_FILE_MAGIC => {
4978 0 : DeltaLayer::new_for_path(path, file)?
4979 0 : .dump(verbose, ctx)
4980 0 : .await?
4981 : }
4982 0 : magic => bail!("unrecognized magic identifier: {:?}", magic),
4983 : }
4984 :
4985 0 : Ok(())
4986 0 : }
4987 :
4988 : #[cfg(test)]
4989 : pub(crate) mod harness {
4990 : use bytes::{Bytes, BytesMut};
4991 : use once_cell::sync::OnceCell;
4992 : use pageserver_api::models::ShardParameters;
4993 : use pageserver_api::shard::ShardIndex;
4994 : use utils::logging;
4995 :
4996 : use crate::deletion_queue::mock::MockDeletionQueue;
4997 : use crate::l0_flush::L0FlushConfig;
4998 : use crate::walredo::apply_neon;
4999 : use pageserver_api::key::Key;
5000 : use pageserver_api::record::NeonWalRecord;
5001 :
5002 : use super::*;
5003 : use hex_literal::hex;
5004 : use utils::id::TenantId;
5005 :
5006 : pub const TIMELINE_ID: TimelineId =
5007 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
5008 : pub const NEW_TIMELINE_ID: TimelineId =
5009 : TimelineId::from_array(hex!("AA223344556677881122334455667788"));
5010 :
5011 : /// Convenience function to create a page image with given string as the only content
5012 5028792 : pub fn test_img(s: &str) -> Bytes {
5013 5028792 : let mut buf = BytesMut::new();
5014 5028792 : buf.extend_from_slice(s.as_bytes());
5015 5028792 : buf.resize(64, 0);
5016 5028792 :
5017 5028792 : buf.freeze()
5018 5028792 : }
5019 :
5020 : impl From<TenantConf> for TenantConfOpt {
5021 192 : fn from(tenant_conf: TenantConf) -> Self {
5022 192 : Self {
5023 192 : checkpoint_distance: Some(tenant_conf.checkpoint_distance),
5024 192 : checkpoint_timeout: Some(tenant_conf.checkpoint_timeout),
5025 192 : compaction_target_size: Some(tenant_conf.compaction_target_size),
5026 192 : compaction_period: Some(tenant_conf.compaction_period),
5027 192 : compaction_threshold: Some(tenant_conf.compaction_threshold),
5028 192 : compaction_algorithm: Some(tenant_conf.compaction_algorithm),
5029 192 : gc_horizon: Some(tenant_conf.gc_horizon),
5030 192 : gc_period: Some(tenant_conf.gc_period),
5031 192 : image_creation_threshold: Some(tenant_conf.image_creation_threshold),
5032 192 : pitr_interval: Some(tenant_conf.pitr_interval),
5033 192 : walreceiver_connect_timeout: Some(tenant_conf.walreceiver_connect_timeout),
5034 192 : lagging_wal_timeout: Some(tenant_conf.lagging_wal_timeout),
5035 192 : max_lsn_wal_lag: Some(tenant_conf.max_lsn_wal_lag),
5036 192 : eviction_policy: Some(tenant_conf.eviction_policy),
5037 192 : min_resident_size_override: tenant_conf.min_resident_size_override,
5038 192 : evictions_low_residence_duration_metric_threshold: Some(
5039 192 : tenant_conf.evictions_low_residence_duration_metric_threshold,
5040 192 : ),
5041 192 : heatmap_period: Some(tenant_conf.heatmap_period),
5042 192 : lazy_slru_download: Some(tenant_conf.lazy_slru_download),
5043 192 : timeline_get_throttle: Some(tenant_conf.timeline_get_throttle),
5044 192 : image_layer_creation_check_threshold: Some(
5045 192 : tenant_conf.image_layer_creation_check_threshold,
5046 192 : ),
5047 192 : lsn_lease_length: Some(tenant_conf.lsn_lease_length),
5048 192 : lsn_lease_length_for_ts: Some(tenant_conf.lsn_lease_length_for_ts),
5049 192 : timeline_offloading: Some(tenant_conf.timeline_offloading),
5050 192 : }
5051 192 : }
5052 : }
5053 :
5054 : pub struct TenantHarness {
5055 : pub conf: &'static PageServerConf,
5056 : pub tenant_conf: TenantConf,
5057 : pub tenant_shard_id: TenantShardId,
5058 : pub generation: Generation,
5059 : pub shard: ShardIndex,
5060 : pub remote_storage: GenericRemoteStorage,
5061 : pub remote_fs_dir: Utf8PathBuf,
5062 : pub deletion_queue: MockDeletionQueue,
5063 : }
5064 :
5065 : static LOG_HANDLE: OnceCell<()> = OnceCell::new();
5066 :
5067 208 : pub(crate) fn setup_logging() {
5068 208 : LOG_HANDLE.get_or_init(|| {
5069 196 : logging::init(
5070 196 : logging::LogFormat::Test,
5071 196 : // enable it in case the tests exercise code paths that use
5072 196 : // debug_assert_current_span_has_tenant_and_timeline_id
5073 196 : logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
5074 196 : logging::Output::Stdout,
5075 196 : )
5076 196 : .expect("Failed to init test logging")
5077 208 : });
5078 208 : }
5079 :
5080 : impl TenantHarness {
5081 192 : pub async fn create_custom(
5082 192 : test_name: &'static str,
5083 192 : tenant_conf: TenantConf,
5084 192 : tenant_id: TenantId,
5085 192 : shard_identity: ShardIdentity,
5086 192 : generation: Generation,
5087 192 : ) -> anyhow::Result<Self> {
5088 192 : setup_logging();
5089 192 :
5090 192 : let repo_dir = PageServerConf::test_repo_dir(test_name);
5091 192 : let _ = fs::remove_dir_all(&repo_dir);
5092 192 : fs::create_dir_all(&repo_dir)?;
5093 :
5094 192 : let conf = PageServerConf::dummy_conf(repo_dir);
5095 192 : // Make a static copy of the config. This can never be free'd, but that's
5096 192 : // OK in a test.
5097 192 : let conf: &'static PageServerConf = Box::leak(Box::new(conf));
5098 192 :
5099 192 : let shard = shard_identity.shard_index();
5100 192 : let tenant_shard_id = TenantShardId {
5101 192 : tenant_id,
5102 192 : shard_number: shard.shard_number,
5103 192 : shard_count: shard.shard_count,
5104 192 : };
5105 192 : fs::create_dir_all(conf.tenant_path(&tenant_shard_id))?;
5106 192 : fs::create_dir_all(conf.timelines_path(&tenant_shard_id))?;
5107 :
5108 : use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
5109 192 : let remote_fs_dir = conf.workdir.join("localfs");
5110 192 : std::fs::create_dir_all(&remote_fs_dir).unwrap();
5111 192 : let config = RemoteStorageConfig {
5112 192 : storage: RemoteStorageKind::LocalFs {
5113 192 : local_path: remote_fs_dir.clone(),
5114 192 : },
5115 192 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
5116 192 : };
5117 192 : let remote_storage = GenericRemoteStorage::from_config(&config).await.unwrap();
5118 192 : let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()));
5119 192 :
5120 192 : Ok(Self {
5121 192 : conf,
5122 192 : tenant_conf,
5123 192 : tenant_shard_id,
5124 192 : generation,
5125 192 : shard,
5126 192 : remote_storage,
5127 192 : remote_fs_dir,
5128 192 : deletion_queue,
5129 192 : })
5130 192 : }
5131 :
5132 180 : pub async fn create(test_name: &'static str) -> anyhow::Result<Self> {
5133 180 : // Disable automatic GC and compaction to make the unit tests more deterministic.
5134 180 : // The tests perform them manually if needed.
5135 180 : let tenant_conf = TenantConf {
5136 180 : gc_period: Duration::ZERO,
5137 180 : compaction_period: Duration::ZERO,
5138 180 : ..TenantConf::default()
5139 180 : };
5140 180 : let tenant_id = TenantId::generate();
5141 180 : let shard = ShardIdentity::unsharded();
5142 180 : Self::create_custom(
5143 180 : test_name,
5144 180 : tenant_conf,
5145 180 : tenant_id,
5146 180 : shard,
5147 180 : Generation::new(0xdeadbeef),
5148 180 : )
5149 0 : .await
5150 180 : }
5151 :
5152 20 : pub fn span(&self) -> tracing::Span {
5153 20 : info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
5154 20 : }
5155 :
5156 192 : pub(crate) async fn load(&self) -> (Arc<Tenant>, RequestContext) {
5157 192 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
5158 192 : (
5159 192 : self.do_try_load(&ctx)
5160 1917 : .await
5161 192 : .expect("failed to load test tenant"),
5162 192 : ctx,
5163 192 : )
5164 192 : }
5165 :
5166 192 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5167 : pub(crate) async fn do_try_load(
5168 : &self,
5169 : ctx: &RequestContext,
5170 : ) -> anyhow::Result<Arc<Tenant>> {
5171 : let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
5172 :
5173 : let tenant = Arc::new(Tenant::new(
5174 : TenantState::Attaching,
5175 : self.conf,
5176 : AttachedTenantConf::try_from(LocationConf::attached_single(
5177 : TenantConfOpt::from(self.tenant_conf.clone()),
5178 : self.generation,
5179 : &ShardParameters::default(),
5180 : ))
5181 : .unwrap(),
5182 : // This is a legacy/test code path: sharding isn't supported here.
5183 : ShardIdentity::unsharded(),
5184 : Some(walredo_mgr),
5185 : self.tenant_shard_id,
5186 : self.remote_storage.clone(),
5187 : self.deletion_queue.new_client(),
5188 : // TODO: ideally we should run all unit tests with both configs
5189 : L0FlushGlobalState::new(L0FlushConfig::default()),
5190 : ));
5191 :
5192 : let preload = tenant
5193 : .preload(&self.remote_storage, CancellationToken::new())
5194 : .await?;
5195 : tenant.attach(Some(preload), ctx).await?;
5196 :
5197 : tenant.state.send_replace(TenantState::Active);
5198 : for timeline in tenant.timelines.lock().unwrap().values() {
5199 : timeline.set_state(TimelineState::Active);
5200 : }
5201 : Ok(tenant)
5202 : }
5203 :
5204 2 : pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
5205 2 : self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
5206 2 : }
5207 : }
5208 :
5209 : // Mock WAL redo manager that doesn't do much
5210 : pub(crate) struct TestRedoManager;
5211 :
5212 : impl TestRedoManager {
5213 : /// # Cancel-Safety
5214 : ///
5215 : /// This method is cancellation-safe.
5216 410 : pub async fn request_redo(
5217 410 : &self,
5218 410 : key: Key,
5219 410 : lsn: Lsn,
5220 410 : base_img: Option<(Lsn, Bytes)>,
5221 410 : records: Vec<(Lsn, NeonWalRecord)>,
5222 410 : _pg_version: u32,
5223 410 : ) -> Result<Bytes, walredo::Error> {
5224 570 : let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
5225 410 : if records_neon {
5226 : // For Neon wal records, we can decode without spawning postgres, so do so.
5227 410 : let mut page = match (base_img, records.first()) {
5228 344 : (Some((_lsn, img)), _) => {
5229 344 : let mut page = BytesMut::new();
5230 344 : page.extend_from_slice(&img);
5231 344 : page
5232 : }
5233 66 : (_, Some((_lsn, rec))) if rec.will_init() => BytesMut::new(),
5234 : _ => {
5235 0 : panic!("Neon WAL redo requires base image or will init record");
5236 : }
5237 : };
5238 :
5239 980 : for (record_lsn, record) in records {
5240 570 : apply_neon::apply_in_neon(&record, record_lsn, key, &mut page)?;
5241 : }
5242 410 : Ok(page.freeze())
5243 : } else {
5244 : // We never spawn a postgres walredo process in unit tests: just log what we might have done.
5245 0 : let s = format!(
5246 0 : "redo for {} to get to {}, with {} and {} records",
5247 0 : key,
5248 0 : lsn,
5249 0 : if base_img.is_some() {
5250 0 : "base image"
5251 : } else {
5252 0 : "no base image"
5253 : },
5254 0 : records.len()
5255 0 : );
5256 0 : println!("{s}");
5257 0 :
5258 0 : Ok(test_img(&s))
5259 : }
5260 410 : }
5261 : }
5262 : }
5263 :
5264 : #[cfg(test)]
5265 : mod tests {
5266 : use std::collections::{BTreeMap, BTreeSet};
5267 :
5268 : use super::*;
5269 : use crate::keyspace::KeySpaceAccum;
5270 : use crate::tenant::harness::*;
5271 : use crate::tenant::timeline::CompactFlags;
5272 : use crate::DEFAULT_PG_VERSION;
5273 : use bytes::{Bytes, BytesMut};
5274 : use hex_literal::hex;
5275 : use itertools::Itertools;
5276 : use pageserver_api::key::{Key, AUX_KEY_PREFIX, NON_INHERITED_RANGE};
5277 : use pageserver_api::keyspace::KeySpace;
5278 : use pageserver_api::models::{CompactionAlgorithm, CompactionAlgorithmSettings};
5279 : use pageserver_api::value::Value;
5280 : use pageserver_compaction::helpers::overlaps_with;
5281 : use rand::{thread_rng, Rng};
5282 : use storage_layer::PersistentLayerKey;
5283 : use tests::storage_layer::ValuesReconstructState;
5284 : use tests::timeline::{GetVectoredError, ShutdownMode};
5285 : use timeline::{CompactOptions, DeltaLayerTestDesc};
5286 : use utils::id::TenantId;
5287 :
5288 : #[cfg(feature = "testing")]
5289 : use pageserver_api::record::NeonWalRecord;
5290 : #[cfg(feature = "testing")]
5291 : use timeline::compaction::{KeyHistoryRetention, KeyLogAtLsn};
5292 : #[cfg(feature = "testing")]
5293 : use timeline::GcInfo;
5294 :
5295 : static TEST_KEY: Lazy<Key> =
5296 18 : Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));
5297 :
5298 : #[tokio::test]
5299 2 : async fn test_basic() -> anyhow::Result<()> {
5300 20 : let (tenant, ctx) = TenantHarness::create("test_basic").await?.load().await;
5301 2 : let tline = tenant
5302 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5303 6 : .await?;
5304 2 :
5305 2 : let mut writer = tline.writer().await;
5306 2 : writer
5307 2 : .put(
5308 2 : *TEST_KEY,
5309 2 : Lsn(0x10),
5310 2 : &Value::Image(test_img("foo at 0x10")),
5311 2 : &ctx,
5312 2 : )
5313 2 : .await?;
5314 2 : writer.finish_write(Lsn(0x10));
5315 2 : drop(writer);
5316 2 :
5317 2 : let mut writer = tline.writer().await;
5318 2 : writer
5319 2 : .put(
5320 2 : *TEST_KEY,
5321 2 : Lsn(0x20),
5322 2 : &Value::Image(test_img("foo at 0x20")),
5323 2 : &ctx,
5324 2 : )
5325 2 : .await?;
5326 2 : writer.finish_write(Lsn(0x20));
5327 2 : drop(writer);
5328 2 :
5329 2 : assert_eq!(
5330 2 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
5331 2 : test_img("foo at 0x10")
5332 2 : );
5333 2 : assert_eq!(
5334 2 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
5335 2 : test_img("foo at 0x10")
5336 2 : );
5337 2 : assert_eq!(
5338 2 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
5339 2 : test_img("foo at 0x20")
5340 2 : );
5341 2 :
5342 2 : Ok(())
5343 2 : }
5344 :
5345 : #[tokio::test]
5346 2 : async fn no_duplicate_timelines() -> anyhow::Result<()> {
5347 2 : let (tenant, ctx) = TenantHarness::create("no_duplicate_timelines")
5348 2 : .await?
5349 2 : .load()
5350 15 : .await;
5351 2 : let _ = tenant
5352 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5353 5 : .await?;
5354 2 :
5355 2 : match tenant
5356 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5357 2 : .await
5358 2 : {
5359 2 : Ok(_) => panic!("duplicate timeline creation should fail"),
5360 2 : Err(e) => assert_eq!(
5361 2 : e.to_string(),
5362 2 : "timeline already exists with different parameters".to_string()
5363 2 : ),
5364 2 : }
5365 2 :
5366 2 : Ok(())
5367 2 : }
5368 :
5369 : /// Convenience function to create a page image with given string as the only content
5370 10 : pub fn test_value(s: &str) -> Value {
5371 10 : let mut buf = BytesMut::new();
5372 10 : buf.extend_from_slice(s.as_bytes());
5373 10 : Value::Image(buf.freeze())
5374 10 : }
5375 :
5376 : ///
5377 : /// Test branch creation
5378 : ///
5379 : #[tokio::test]
5380 2 : async fn test_branch() -> anyhow::Result<()> {
5381 2 : use std::str::from_utf8;
5382 2 :
5383 20 : let (tenant, ctx) = TenantHarness::create("test_branch").await?.load().await;
5384 2 : let tline = tenant
5385 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5386 6 : .await?;
5387 2 : let mut writer = tline.writer().await;
5388 2 :
5389 2 : #[allow(non_snake_case)]
5390 2 : let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
5391 2 : #[allow(non_snake_case)]
5392 2 : let TEST_KEY_B: Key = Key::from_hex("110000000033333333444444445500000002").unwrap();
5393 2 :
5394 2 : // Insert a value on the timeline
5395 2 : writer
5396 2 : .put(TEST_KEY_A, Lsn(0x20), &test_value("foo at 0x20"), &ctx)
5397 2 : .await?;
5398 2 : writer
5399 2 : .put(TEST_KEY_B, Lsn(0x20), &test_value("foobar at 0x20"), &ctx)
5400 2 : .await?;
5401 2 : writer.finish_write(Lsn(0x20));
5402 2 :
5403 2 : writer
5404 2 : .put(TEST_KEY_A, Lsn(0x30), &test_value("foo at 0x30"), &ctx)
5405 2 : .await?;
5406 2 : writer.finish_write(Lsn(0x30));
5407 2 : writer
5408 2 : .put(TEST_KEY_A, Lsn(0x40), &test_value("foo at 0x40"), &ctx)
5409 2 : .await?;
5410 2 : writer.finish_write(Lsn(0x40));
5411 2 :
5412 2 : //assert_current_logical_size(&tline, Lsn(0x40));
5413 2 :
5414 2 : // Branch the history, modify relation differently on the new timeline
5415 2 : tenant
5416 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x30)), &ctx)
5417 2 : .await?;
5418 2 : let newtline = tenant
5419 2 : .get_timeline(NEW_TIMELINE_ID, true)
5420 2 : .expect("Should have a local timeline");
5421 2 : let mut new_writer = newtline.writer().await;
5422 2 : new_writer
5423 2 : .put(TEST_KEY_A, Lsn(0x40), &test_value("bar at 0x40"), &ctx)
5424 2 : .await?;
5425 2 : new_writer.finish_write(Lsn(0x40));
5426 2 :
5427 2 : // Check page contents on both branches
5428 2 : assert_eq!(
5429 2 : from_utf8(&tline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
5430 2 : "foo at 0x40"
5431 2 : );
5432 2 : assert_eq!(
5433 2 : from_utf8(&newtline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
5434 2 : "bar at 0x40"
5435 2 : );
5436 2 : assert_eq!(
5437 2 : from_utf8(&newtline.get(TEST_KEY_B, Lsn(0x40), &ctx).await?)?,
5438 2 : "foobar at 0x20"
5439 2 : );
5440 2 :
5441 2 : //assert_current_logical_size(&tline, Lsn(0x40));
5442 2 :
5443 2 : Ok(())
5444 2 : }
5445 :
5446 20 : async fn make_some_layers(
5447 20 : tline: &Timeline,
5448 20 : start_lsn: Lsn,
5449 20 : ctx: &RequestContext,
5450 20 : ) -> anyhow::Result<()> {
5451 20 : let mut lsn = start_lsn;
5452 : {
5453 20 : let mut writer = tline.writer().await;
5454 : // Create a relation on the timeline
5455 20 : writer
5456 20 : .put(
5457 20 : *TEST_KEY,
5458 20 : lsn,
5459 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5460 20 : ctx,
5461 20 : )
5462 10 : .await?;
5463 20 : writer.finish_write(lsn);
5464 20 : lsn += 0x10;
5465 20 : writer
5466 20 : .put(
5467 20 : *TEST_KEY,
5468 20 : lsn,
5469 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5470 20 : ctx,
5471 20 : )
5472 0 : .await?;
5473 20 : writer.finish_write(lsn);
5474 20 : lsn += 0x10;
5475 20 : }
5476 20 : tline.freeze_and_flush().await?;
5477 : {
5478 20 : let mut writer = tline.writer().await;
5479 20 : writer
5480 20 : .put(
5481 20 : *TEST_KEY,
5482 20 : lsn,
5483 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5484 20 : ctx,
5485 20 : )
5486 10 : .await?;
5487 20 : writer.finish_write(lsn);
5488 20 : lsn += 0x10;
5489 20 : writer
5490 20 : .put(
5491 20 : *TEST_KEY,
5492 20 : lsn,
5493 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5494 20 : ctx,
5495 20 : )
5496 0 : .await?;
5497 20 : writer.finish_write(lsn);
5498 20 : }
5499 21 : tline.freeze_and_flush().await.map_err(|e| e.into())
5500 20 : }
5501 :
5502 : #[tokio::test(start_paused = true)]
5503 2 : async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
5504 2 : let (tenant, ctx) =
5505 2 : TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
5506 2 : .await?
5507 2 : .load()
5508 20 : .await;
5509 2 : // Advance to the lsn lease deadline so that GC is not blocked by
5510 2 : // initial transition into AttachedSingle.
5511 2 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
5512 2 : tokio::time::resume();
5513 2 : let tline = tenant
5514 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5515 6 : .await?;
5516 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5517 2 :
5518 2 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
5519 2 : // FIXME: this doesn't actually remove any layer currently, given how the flushing
5520 2 : // and compaction works. But it does set the 'cutoff' point so that the cross check
5521 2 : // below should fail.
5522 2 : tenant
5523 2 : .gc_iteration(
5524 2 : Some(TIMELINE_ID),
5525 2 : 0x10,
5526 2 : Duration::ZERO,
5527 2 : &CancellationToken::new(),
5528 2 : &ctx,
5529 2 : )
5530 2 : .await?;
5531 2 :
5532 2 : // try to branch at lsn 25, should fail because we already garbage collected the data
5533 2 : match tenant
5534 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
5535 2 : .await
5536 2 : {
5537 2 : Ok(_) => panic!("branching should have failed"),
5538 2 : Err(err) => {
5539 2 : let CreateTimelineError::AncestorLsn(err) = err else {
5540 2 : panic!("wrong error type")
5541 2 : };
5542 2 : assert!(err.to_string().contains("invalid branch start lsn"));
5543 2 : assert!(err
5544 2 : .source()
5545 2 : .unwrap()
5546 2 : .to_string()
5547 2 : .contains("we might've already garbage collected needed data"))
5548 2 : }
5549 2 : }
5550 2 :
5551 2 : Ok(())
5552 2 : }
5553 :
5554 : #[tokio::test]
5555 2 : async fn test_prohibit_branch_creation_on_pre_initdb_lsn() -> anyhow::Result<()> {
5556 2 : let (tenant, ctx) =
5557 2 : TenantHarness::create("test_prohibit_branch_creation_on_pre_initdb_lsn")
5558 2 : .await?
5559 2 : .load()
5560 19 : .await;
5561 2 :
5562 2 : let tline = tenant
5563 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x50), DEFAULT_PG_VERSION, &ctx)
5564 5 : .await?;
5565 2 : // try to branch at lsn 0x25, should fail because initdb lsn is 0x50
5566 2 : match tenant
5567 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
5568 2 : .await
5569 2 : {
5570 2 : Ok(_) => panic!("branching should have failed"),
5571 2 : Err(err) => {
5572 2 : let CreateTimelineError::AncestorLsn(err) = err else {
5573 2 : panic!("wrong error type");
5574 2 : };
5575 2 : assert!(&err.to_string().contains("invalid branch start lsn"));
5576 2 : assert!(&err
5577 2 : .source()
5578 2 : .unwrap()
5579 2 : .to_string()
5580 2 : .contains("is earlier than latest GC cutoff"));
5581 2 : }
5582 2 : }
5583 2 :
5584 2 : Ok(())
5585 2 : }
5586 :
5587 : /*
5588 : // FIXME: This currently fails to error out. Calling GC doesn't currently
5589 : // remove the old value, we'd need to work a little harder
5590 : #[tokio::test]
5591 : async fn test_prohibit_get_for_garbage_collected_data() -> anyhow::Result<()> {
5592 : let repo =
5593 : RepoHarness::create("test_prohibit_get_for_garbage_collected_data")?
5594 : .load();
5595 :
5596 : let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION)?;
5597 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5598 :
5599 : repo.gc_iteration(Some(TIMELINE_ID), 0x10, Duration::ZERO)?;
5600 : let latest_gc_cutoff_lsn = tline.get_latest_gc_cutoff_lsn();
5601 : assert!(*latest_gc_cutoff_lsn > Lsn(0x25));
5602 : match tline.get(*TEST_KEY, Lsn(0x25)) {
5603 : Ok(_) => panic!("request for page should have failed"),
5604 : Err(err) => assert!(err.to_string().contains("not found at")),
5605 : }
5606 : Ok(())
5607 : }
5608 : */
5609 :
5610 : #[tokio::test]
5611 2 : async fn test_get_branchpoints_from_an_inactive_timeline() -> anyhow::Result<()> {
5612 2 : let (tenant, ctx) =
5613 2 : TenantHarness::create("test_get_branchpoints_from_an_inactive_timeline")
5614 2 : .await?
5615 2 : .load()
5616 20 : .await;
5617 2 : let tline = tenant
5618 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5619 6 : .await?;
5620 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5621 2 :
5622 2 : tenant
5623 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
5624 2 : .await?;
5625 2 : let newtline = tenant
5626 2 : .get_timeline(NEW_TIMELINE_ID, true)
5627 2 : .expect("Should have a local timeline");
5628 2 :
5629 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
5630 2 :
5631 2 : tline.set_broken("test".to_owned());
5632 2 :
5633 2 : tenant
5634 2 : .gc_iteration(
5635 2 : Some(TIMELINE_ID),
5636 2 : 0x10,
5637 2 : Duration::ZERO,
5638 2 : &CancellationToken::new(),
5639 2 : &ctx,
5640 2 : )
5641 2 : .await?;
5642 2 :
5643 2 : // The branchpoints should contain all timelines, even ones marked
5644 2 : // as Broken.
5645 2 : {
5646 2 : let branchpoints = &tline.gc_info.read().unwrap().retain_lsns;
5647 2 : assert_eq!(branchpoints.len(), 1);
5648 2 : assert_eq!(
5649 2 : branchpoints[0],
5650 2 : (Lsn(0x40), NEW_TIMELINE_ID, MaybeOffloaded::No)
5651 2 : );
5652 2 : }
5653 2 :
5654 2 : // You can read the key from the child branch even though the parent is
5655 2 : // Broken, as long as you don't need to access data from the parent.
5656 2 : assert_eq!(
5657 4 : newtline.get(*TEST_KEY, Lsn(0x70), &ctx).await?,
5658 2 : test_img(&format!("foo at {}", Lsn(0x70)))
5659 2 : );
5660 2 :
5661 2 : // This needs to traverse to the parent, and fails.
5662 2 : let err = newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await.unwrap_err();
5663 2 : assert!(
5664 2 : err.to_string().starts_with(&format!(
5665 2 : "bad state on timeline {}: Broken",
5666 2 : tline.timeline_id
5667 2 : )),
5668 2 : "{err}"
5669 2 : );
5670 2 :
5671 2 : Ok(())
5672 2 : }
5673 :
5674 : #[tokio::test]
5675 2 : async fn test_retain_data_in_parent_which_is_needed_for_child() -> anyhow::Result<()> {
5676 2 : let (tenant, ctx) =
5677 2 : TenantHarness::create("test_retain_data_in_parent_which_is_needed_for_child")
5678 2 : .await?
5679 2 : .load()
5680 17 : .await;
5681 2 : let tline = tenant
5682 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5683 6 : .await?;
5684 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5685 2 :
5686 2 : tenant
5687 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
5688 2 : .await?;
5689 2 : let newtline = tenant
5690 2 : .get_timeline(NEW_TIMELINE_ID, true)
5691 2 : .expect("Should have a local timeline");
5692 2 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
5693 2 : tenant
5694 2 : .gc_iteration(
5695 2 : Some(TIMELINE_ID),
5696 2 : 0x10,
5697 2 : Duration::ZERO,
5698 2 : &CancellationToken::new(),
5699 2 : &ctx,
5700 2 : )
5701 2 : .await?;
5702 4 : assert!(newtline.get(*TEST_KEY, Lsn(0x25), &ctx).await.is_ok());
5703 2 :
5704 2 : Ok(())
5705 2 : }
5706 : #[tokio::test]
5707 2 : async fn test_parent_keeps_data_forever_after_branching() -> anyhow::Result<()> {
5708 2 : let (tenant, ctx) = TenantHarness::create("test_parent_keeps_data_forever_after_branching")
5709 2 : .await?
5710 2 : .load()
5711 20 : .await;
5712 2 : let tline = tenant
5713 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5714 6 : .await?;
5715 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5716 2 :
5717 2 : tenant
5718 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
5719 2 : .await?;
5720 2 : let newtline = tenant
5721 2 : .get_timeline(NEW_TIMELINE_ID, true)
5722 2 : .expect("Should have a local timeline");
5723 2 :
5724 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
5725 2 :
5726 2 : // run gc on parent
5727 2 : tenant
5728 2 : .gc_iteration(
5729 2 : Some(TIMELINE_ID),
5730 2 : 0x10,
5731 2 : Duration::ZERO,
5732 2 : &CancellationToken::new(),
5733 2 : &ctx,
5734 2 : )
5735 2 : .await?;
5736 2 :
5737 2 : // Check that the data is still accessible on the branch.
5738 2 : assert_eq!(
5739 7 : newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
5740 2 : test_img(&format!("foo at {}", Lsn(0x40)))
5741 2 : );
5742 2 :
5743 2 : Ok(())
5744 2 : }
5745 :
5746 : #[tokio::test]
5747 2 : async fn timeline_load() -> anyhow::Result<()> {
5748 2 : const TEST_NAME: &str = "timeline_load";
5749 2 : let harness = TenantHarness::create(TEST_NAME).await?;
5750 2 : {
5751 20 : let (tenant, ctx) = harness.load().await;
5752 2 : let tline = tenant
5753 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x7000), DEFAULT_PG_VERSION, &ctx)
5754 6 : .await?;
5755 6 : make_some_layers(tline.as_ref(), Lsn(0x8000), &ctx).await?;
5756 2 : // so that all uploads finish & we can call harness.load() below again
5757 2 : tenant
5758 2 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
5759 2 : .instrument(harness.span())
5760 2 : .await
5761 2 : .ok()
5762 2 : .unwrap();
5763 2 : }
5764 2 :
5765 22 : let (tenant, _ctx) = harness.load().await;
5766 2 : tenant
5767 2 : .get_timeline(TIMELINE_ID, true)
5768 2 : .expect("cannot load timeline");
5769 2 :
5770 2 : Ok(())
5771 2 : }
5772 :
5773 : #[tokio::test]
5774 2 : async fn timeline_load_with_ancestor() -> anyhow::Result<()> {
5775 2 : const TEST_NAME: &str = "timeline_load_with_ancestor";
5776 2 : let harness = TenantHarness::create(TEST_NAME).await?;
5777 2 : // create two timelines
5778 2 : {
5779 19 : let (tenant, ctx) = harness.load().await;
5780 2 : let tline = tenant
5781 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5782 6 : .await?;
5783 2 :
5784 7 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5785 2 :
5786 2 : let child_tline = tenant
5787 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
5788 2 : .await?;
5789 2 : child_tline.set_state(TimelineState::Active);
5790 2 :
5791 2 : let newtline = tenant
5792 2 : .get_timeline(NEW_TIMELINE_ID, true)
5793 2 : .expect("Should have a local timeline");
5794 2 :
5795 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
5796 2 :
5797 2 : // so that all uploads finish & we can call harness.load() below again
5798 2 : tenant
5799 2 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
5800 2 : .instrument(harness.span())
5801 2 : .await
5802 2 : .ok()
5803 2 : .unwrap();
5804 2 : }
5805 2 :
5806 2 : // check that both of them are initially unloaded
5807 33 : let (tenant, _ctx) = harness.load().await;
5808 2 :
5809 2 : // check that both, child and ancestor are loaded
5810 2 : let _child_tline = tenant
5811 2 : .get_timeline(NEW_TIMELINE_ID, true)
5812 2 : .expect("cannot get child timeline loaded");
5813 2 :
5814 2 : let _ancestor_tline = tenant
5815 2 : .get_timeline(TIMELINE_ID, true)
5816 2 : .expect("cannot get ancestor timeline loaded");
5817 2 :
5818 2 : Ok(())
5819 2 : }
5820 :
5821 : #[tokio::test]
5822 2 : async fn delta_layer_dumping() -> anyhow::Result<()> {
5823 2 : use storage_layer::AsLayerDesc;
5824 2 : let (tenant, ctx) = TenantHarness::create("test_layer_dumping")
5825 2 : .await?
5826 2 : .load()
5827 20 : .await;
5828 2 : let tline = tenant
5829 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5830 6 : .await?;
5831 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5832 2 :
5833 2 : let layer_map = tline.layers.read().await;
5834 2 : let level0_deltas = layer_map
5835 2 : .layer_map()?
5836 2 : .level0_deltas()
5837 2 : .iter()
5838 4 : .map(|desc| layer_map.get_from_desc(desc))
5839 2 : .collect::<Vec<_>>();
5840 2 :
5841 2 : assert!(!level0_deltas.is_empty());
5842 2 :
5843 6 : for delta in level0_deltas {
5844 2 : // Ensure we are dumping a delta layer here
5845 4 : assert!(delta.layer_desc().is_delta);
5846 8 : delta.dump(true, &ctx).await.unwrap();
5847 2 : }
5848 2 :
5849 2 : Ok(())
5850 2 : }
5851 :
5852 : #[tokio::test]
5853 2 : async fn test_images() -> anyhow::Result<()> {
5854 20 : let (tenant, ctx) = TenantHarness::create("test_images").await?.load().await;
5855 2 : let tline = tenant
5856 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5857 6 : .await?;
5858 2 :
5859 2 : let mut writer = tline.writer().await;
5860 2 : writer
5861 2 : .put(
5862 2 : *TEST_KEY,
5863 2 : Lsn(0x10),
5864 2 : &Value::Image(test_img("foo at 0x10")),
5865 2 : &ctx,
5866 2 : )
5867 2 : .await?;
5868 2 : writer.finish_write(Lsn(0x10));
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(0x20),
5881 2 : &Value::Image(test_img("foo at 0x20")),
5882 2 : &ctx,
5883 2 : )
5884 2 : .await?;
5885 2 : writer.finish_write(Lsn(0x20));
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(0x30),
5898 2 : &Value::Image(test_img("foo at 0x30")),
5899 2 : &ctx,
5900 2 : )
5901 2 : .await?;
5902 2 : writer.finish_write(Lsn(0x30));
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 : let mut writer = tline.writer().await;
5911 2 : writer
5912 2 : .put(
5913 2 : *TEST_KEY,
5914 2 : Lsn(0x40),
5915 2 : &Value::Image(test_img("foo at 0x40")),
5916 2 : &ctx,
5917 2 : )
5918 2 : .await?;
5919 2 : writer.finish_write(Lsn(0x40));
5920 2 : drop(writer);
5921 2 :
5922 2 : tline.freeze_and_flush().await?;
5923 2 : tline
5924 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
5925 2 : .await?;
5926 2 :
5927 2 : assert_eq!(
5928 4 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
5929 2 : test_img("foo at 0x10")
5930 2 : );
5931 2 : assert_eq!(
5932 4 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
5933 2 : test_img("foo at 0x10")
5934 2 : );
5935 2 : assert_eq!(
5936 2 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
5937 2 : test_img("foo at 0x20")
5938 2 : );
5939 2 : assert_eq!(
5940 4 : tline.get(*TEST_KEY, Lsn(0x30), &ctx).await?,
5941 2 : test_img("foo at 0x30")
5942 2 : );
5943 2 : assert_eq!(
5944 4 : tline.get(*TEST_KEY, Lsn(0x40), &ctx).await?,
5945 2 : test_img("foo at 0x40")
5946 2 : );
5947 2 :
5948 2 : Ok(())
5949 2 : }
5950 :
5951 4 : async fn bulk_insert_compact_gc(
5952 4 : tenant: &Tenant,
5953 4 : timeline: &Arc<Timeline>,
5954 4 : ctx: &RequestContext,
5955 4 : lsn: Lsn,
5956 4 : repeat: usize,
5957 4 : key_count: usize,
5958 4 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
5959 4 : let compact = true;
5960 40718 : bulk_insert_maybe_compact_gc(tenant, timeline, ctx, lsn, repeat, key_count, compact).await
5961 4 : }
5962 :
5963 8 : async fn bulk_insert_maybe_compact_gc(
5964 8 : tenant: &Tenant,
5965 8 : timeline: &Arc<Timeline>,
5966 8 : ctx: &RequestContext,
5967 8 : mut lsn: Lsn,
5968 8 : repeat: usize,
5969 8 : key_count: usize,
5970 8 : compact: bool,
5971 8 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
5972 8 : let mut inserted: HashMap<Key, BTreeSet<Lsn>> = Default::default();
5973 8 :
5974 8 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5975 8 : let mut blknum = 0;
5976 8 :
5977 8 : // Enforce that key range is monotonously increasing
5978 8 : let mut keyspace = KeySpaceAccum::new();
5979 8 :
5980 8 : let cancel = CancellationToken::new();
5981 8 :
5982 8 : for _ in 0..repeat {
5983 400 : for _ in 0..key_count {
5984 4000000 : test_key.field6 = blknum;
5985 4000000 : let mut writer = timeline.writer().await;
5986 4000000 : writer
5987 4000000 : .put(
5988 4000000 : test_key,
5989 4000000 : lsn,
5990 4000000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5991 4000000 : ctx,
5992 4000000 : )
5993 3426 : .await?;
5994 4000000 : inserted.entry(test_key).or_default().insert(lsn);
5995 4000000 : writer.finish_write(lsn);
5996 4000000 : drop(writer);
5997 4000000 :
5998 4000000 : keyspace.add_key(test_key);
5999 4000000 :
6000 4000000 : lsn = Lsn(lsn.0 + 0x10);
6001 4000000 : blknum += 1;
6002 : }
6003 :
6004 400 : timeline.freeze_and_flush().await?;
6005 400 : if compact {
6006 : // this requires timeline to be &Arc<Timeline>
6007 8618 : timeline.compact(&cancel, EnumSet::empty(), ctx).await?;
6008 200 : }
6009 :
6010 : // this doesn't really need to use the timeline_id target, but it is closer to what it
6011 : // originally was.
6012 400 : let res = tenant
6013 400 : .gc_iteration(Some(timeline.timeline_id), 0, Duration::ZERO, &cancel, ctx)
6014 0 : .await?;
6015 :
6016 400 : assert_eq!(res.layers_removed, 0, "this never removes anything");
6017 : }
6018 :
6019 8 : Ok(inserted)
6020 8 : }
6021 :
6022 : //
6023 : // Insert 1000 key-value pairs with increasing keys, flush, compact, GC.
6024 : // Repeat 50 times.
6025 : //
6026 : #[tokio::test]
6027 2 : async fn test_bulk_insert() -> anyhow::Result<()> {
6028 2 : let harness = TenantHarness::create("test_bulk_insert").await?;
6029 20 : let (tenant, ctx) = harness.load().await;
6030 2 : let tline = tenant
6031 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6032 6 : .await?;
6033 2 :
6034 2 : let lsn = Lsn(0x10);
6035 20359 : bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
6036 2 :
6037 2 : Ok(())
6038 2 : }
6039 :
6040 : // Test the vectored get real implementation against a simple sequential implementation.
6041 : //
6042 : // The test generates a keyspace by repeatedly flushing the in-memory layer and compacting.
6043 : // Projected to 2D the key space looks like below. Lsn grows upwards on the Y axis and keys
6044 : // grow to the right on the X axis.
6045 : // [Delta]
6046 : // [Delta]
6047 : // [Delta]
6048 : // [Delta]
6049 : // ------------ Image ---------------
6050 : //
6051 : // After layer generation we pick the ranges to query as follows:
6052 : // 1. The beginning of each delta layer
6053 : // 2. At the seam between two adjacent delta layers
6054 : //
6055 : // There's one major downside to this test: delta layers only contains images,
6056 : // so the search can stop at the first delta layer and doesn't traverse any deeper.
6057 : #[tokio::test]
6058 2 : async fn test_get_vectored() -> anyhow::Result<()> {
6059 2 : let harness = TenantHarness::create("test_get_vectored").await?;
6060 20 : let (tenant, ctx) = harness.load().await;
6061 2 : let tline = tenant
6062 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6063 6 : .await?;
6064 2 :
6065 2 : let lsn = Lsn(0x10);
6066 20359 : let inserted = bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
6067 2 :
6068 2 : let guard = tline.layers.read().await;
6069 2 : let lm = guard.layer_map()?;
6070 2 :
6071 2 : lm.dump(true, &ctx).await?;
6072 2 :
6073 2 : let mut reads = Vec::new();
6074 2 : let mut prev = None;
6075 12 : lm.iter_historic_layers().for_each(|desc| {
6076 12 : if !desc.is_delta() {
6077 2 : prev = Some(desc.clone());
6078 2 : return;
6079 10 : }
6080 10 :
6081 10 : let start = desc.key_range.start;
6082 10 : let end = desc
6083 10 : .key_range
6084 10 : .start
6085 10 : .add(Timeline::MAX_GET_VECTORED_KEYS.try_into().unwrap());
6086 10 : reads.push(KeySpace {
6087 10 : ranges: vec![start..end],
6088 10 : });
6089 2 :
6090 10 : if let Some(prev) = &prev {
6091 10 : if !prev.is_delta() {
6092 10 : return;
6093 2 : }
6094 0 :
6095 0 : let first_range = Key {
6096 0 : field6: prev.key_range.end.field6 - 4,
6097 0 : ..prev.key_range.end
6098 0 : }..prev.key_range.end;
6099 0 :
6100 0 : let second_range = desc.key_range.start..Key {
6101 0 : field6: desc.key_range.start.field6 + 4,
6102 0 : ..desc.key_range.start
6103 0 : };
6104 0 :
6105 0 : reads.push(KeySpace {
6106 0 : ranges: vec![first_range, second_range],
6107 0 : });
6108 2 : };
6109 2 :
6110 2 : prev = Some(desc.clone());
6111 12 : });
6112 2 :
6113 2 : drop(guard);
6114 2 :
6115 2 : // Pick a big LSN such that we query over all the changes.
6116 2 : let reads_lsn = Lsn(u64::MAX - 1);
6117 2 :
6118 12 : for read in reads {
6119 10 : info!("Doing vectored read on {:?}", read);
6120 2 :
6121 10 : let vectored_res = tline
6122 10 : .get_vectored_impl(
6123 10 : read.clone(),
6124 10 : reads_lsn,
6125 10 : &mut ValuesReconstructState::new(),
6126 10 : &ctx,
6127 10 : )
6128 25 : .await;
6129 2 :
6130 10 : let mut expected_lsns: HashMap<Key, Lsn> = Default::default();
6131 10 : let mut expect_missing = false;
6132 10 : let mut key = read.start().unwrap();
6133 330 : while key != read.end().unwrap() {
6134 320 : if let Some(lsns) = inserted.get(&key) {
6135 320 : let expected_lsn = lsns.iter().rfind(|lsn| **lsn <= reads_lsn);
6136 320 : match expected_lsn {
6137 320 : Some(lsn) => {
6138 320 : expected_lsns.insert(key, *lsn);
6139 320 : }
6140 2 : None => {
6141 2 : expect_missing = true;
6142 0 : break;
6143 2 : }
6144 2 : }
6145 2 : } else {
6146 2 : expect_missing = true;
6147 0 : break;
6148 2 : }
6149 2 :
6150 320 : key = key.next();
6151 2 : }
6152 2 :
6153 10 : if expect_missing {
6154 2 : assert!(matches!(vectored_res, Err(GetVectoredError::MissingKey(_))));
6155 2 : } else {
6156 320 : for (key, image) in vectored_res? {
6157 320 : let expected_lsn = expected_lsns.get(&key).expect("determined above");
6158 320 : let expected_image = test_img(&format!("{} at {}", key.field6, expected_lsn));
6159 320 : assert_eq!(image?, expected_image);
6160 2 : }
6161 2 : }
6162 2 : }
6163 2 :
6164 2 : Ok(())
6165 2 : }
6166 :
6167 : #[tokio::test]
6168 2 : async fn test_get_vectored_aux_files() -> anyhow::Result<()> {
6169 2 : let harness = TenantHarness::create("test_get_vectored_aux_files").await?;
6170 2 :
6171 20 : let (tenant, ctx) = harness.load().await;
6172 2 : let tline = tenant
6173 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
6174 2 : .await?;
6175 2 : let tline = tline.raw_timeline().unwrap();
6176 2 :
6177 2 : let mut modification = tline.begin_modification(Lsn(0x1000));
6178 2 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
6179 2 : modification.set_lsn(Lsn(0x1008))?;
6180 2 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
6181 2 : modification.commit(&ctx).await?;
6182 2 :
6183 2 : let child_timeline_id = TimelineId::generate();
6184 2 : tenant
6185 2 : .branch_timeline_test(
6186 2 : tline,
6187 2 : child_timeline_id,
6188 2 : Some(tline.get_last_record_lsn()),
6189 2 : &ctx,
6190 2 : )
6191 2 : .await?;
6192 2 :
6193 2 : let child_timeline = tenant
6194 2 : .get_timeline(child_timeline_id, true)
6195 2 : .expect("Should have the branched timeline");
6196 2 :
6197 2 : let aux_keyspace = KeySpace {
6198 2 : ranges: vec![NON_INHERITED_RANGE],
6199 2 : };
6200 2 : let read_lsn = child_timeline.get_last_record_lsn();
6201 2 :
6202 2 : let vectored_res = child_timeline
6203 2 : .get_vectored_impl(
6204 2 : aux_keyspace.clone(),
6205 2 : read_lsn,
6206 2 : &mut ValuesReconstructState::new(),
6207 2 : &ctx,
6208 2 : )
6209 2 : .await;
6210 2 :
6211 2 : let images = vectored_res?;
6212 2 : assert!(images.is_empty());
6213 2 : Ok(())
6214 2 : }
6215 :
6216 : // Test that vectored get handles layer gaps correctly
6217 : // by advancing into the next ancestor timeline if required.
6218 : //
6219 : // The test generates timelines that look like the diagram below.
6220 : // We leave a gap in one of the L1 layers at `gap_at_key` (`/` in the diagram).
6221 : // The reconstruct data for that key lies in the ancestor timeline (`X` in the diagram).
6222 : //
6223 : // ```
6224 : //-------------------------------+
6225 : // ... |
6226 : // [ L1 ] |
6227 : // [ / L1 ] | Child Timeline
6228 : // ... |
6229 : // ------------------------------+
6230 : // [ X L1 ] | Parent Timeline
6231 : // ------------------------------+
6232 : // ```
6233 : #[tokio::test]
6234 2 : async fn test_get_vectored_key_gap() -> anyhow::Result<()> {
6235 2 : let tenant_conf = TenantConf {
6236 2 : // Make compaction deterministic
6237 2 : gc_period: Duration::ZERO,
6238 2 : compaction_period: Duration::ZERO,
6239 2 : // Encourage creation of L1 layers
6240 2 : checkpoint_distance: 16 * 1024,
6241 2 : compaction_target_size: 8 * 1024,
6242 2 : ..TenantConf::default()
6243 2 : };
6244 2 :
6245 2 : let harness = TenantHarness::create_custom(
6246 2 : "test_get_vectored_key_gap",
6247 2 : tenant_conf,
6248 2 : TenantId::generate(),
6249 2 : ShardIdentity::unsharded(),
6250 2 : Generation::new(0xdeadbeef),
6251 2 : )
6252 2 : .await?;
6253 20 : let (tenant, ctx) = harness.load().await;
6254 2 :
6255 2 : let mut current_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6256 2 : let gap_at_key = current_key.add(100);
6257 2 : let mut current_lsn = Lsn(0x10);
6258 2 :
6259 2 : const KEY_COUNT: usize = 10_000;
6260 2 :
6261 2 : let timeline_id = TimelineId::generate();
6262 2 : let current_timeline = tenant
6263 2 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
6264 6 : .await?;
6265 2 :
6266 2 : current_lsn += 0x100;
6267 2 :
6268 2 : let mut writer = current_timeline.writer().await;
6269 2 : writer
6270 2 : .put(
6271 2 : gap_at_key,
6272 2 : current_lsn,
6273 2 : &Value::Image(test_img(&format!("{} at {}", gap_at_key, current_lsn))),
6274 2 : &ctx,
6275 2 : )
6276 2 : .await?;
6277 2 : writer.finish_write(current_lsn);
6278 2 : drop(writer);
6279 2 :
6280 2 : let mut latest_lsns = HashMap::new();
6281 2 : latest_lsns.insert(gap_at_key, current_lsn);
6282 2 :
6283 2 : current_timeline.freeze_and_flush().await?;
6284 2 :
6285 2 : let child_timeline_id = TimelineId::generate();
6286 2 :
6287 2 : tenant
6288 2 : .branch_timeline_test(
6289 2 : ¤t_timeline,
6290 2 : child_timeline_id,
6291 2 : Some(current_lsn),
6292 2 : &ctx,
6293 2 : )
6294 2 : .await?;
6295 2 : let child_timeline = tenant
6296 2 : .get_timeline(child_timeline_id, true)
6297 2 : .expect("Should have the branched timeline");
6298 2 :
6299 20002 : for i in 0..KEY_COUNT {
6300 20000 : if current_key == gap_at_key {
6301 2 : current_key = current_key.next();
6302 2 : continue;
6303 19998 : }
6304 19998 :
6305 19998 : current_lsn += 0x10;
6306 2 :
6307 19998 : let mut writer = child_timeline.writer().await;
6308 19998 : writer
6309 19998 : .put(
6310 19998 : current_key,
6311 19998 : current_lsn,
6312 19998 : &Value::Image(test_img(&format!("{} at {}", current_key, current_lsn))),
6313 19998 : &ctx,
6314 19998 : )
6315 67 : .await?;
6316 19998 : writer.finish_write(current_lsn);
6317 19998 : drop(writer);
6318 19998 :
6319 19998 : latest_lsns.insert(current_key, current_lsn);
6320 19998 : current_key = current_key.next();
6321 19998 :
6322 19998 : // Flush every now and then to encourage layer file creation.
6323 19998 : if i % 500 == 0 {
6324 43 : child_timeline.freeze_and_flush().await?;
6325 19958 : }
6326 2 : }
6327 2 :
6328 3 : child_timeline.freeze_and_flush().await?;
6329 2 : let mut flags = EnumSet::new();
6330 2 : flags.insert(CompactFlags::ForceRepartition);
6331 2 : child_timeline
6332 2 : .compact(&CancellationToken::new(), flags, &ctx)
6333 1757 : .await?;
6334 2 :
6335 2 : let key_near_end = {
6336 2 : let mut tmp = current_key;
6337 2 : tmp.field6 -= 10;
6338 2 : tmp
6339 2 : };
6340 2 :
6341 2 : let key_near_gap = {
6342 2 : let mut tmp = gap_at_key;
6343 2 : tmp.field6 -= 10;
6344 2 : tmp
6345 2 : };
6346 2 :
6347 2 : let read = KeySpace {
6348 2 : ranges: vec![key_near_gap..gap_at_key.next(), key_near_end..current_key],
6349 2 : };
6350 2 : let results = child_timeline
6351 2 : .get_vectored_impl(
6352 2 : read.clone(),
6353 2 : current_lsn,
6354 2 : &mut ValuesReconstructState::new(),
6355 2 : &ctx,
6356 2 : )
6357 16 : .await?;
6358 2 :
6359 44 : for (key, img_res) in results {
6360 42 : let expected = test_img(&format!("{} at {}", key, latest_lsns[&key]));
6361 42 : assert_eq!(img_res?, expected);
6362 2 : }
6363 2 :
6364 2 : Ok(())
6365 2 : }
6366 :
6367 : // Test that vectored get descends into ancestor timelines correctly and
6368 : // does not return an image that's newer than requested.
6369 : //
6370 : // The diagram below ilustrates an interesting case. We have a parent timeline
6371 : // (top of the Lsn range) and a child timeline. The request key cannot be reconstructed
6372 : // from the child timeline, so the parent timeline must be visited. When advacing into
6373 : // the child timeline, the read path needs to remember what the requested Lsn was in
6374 : // order to avoid returning an image that's too new. The test below constructs such
6375 : // a timeline setup and does a few queries around the Lsn of each page image.
6376 : // ```
6377 : // LSN
6378 : // ^
6379 : // |
6380 : // |
6381 : // 500 | --------------------------------------> branch point
6382 : // 400 | X
6383 : // 300 | X
6384 : // 200 | --------------------------------------> requested lsn
6385 : // 100 | X
6386 : // |---------------------------------------> Key
6387 : // |
6388 : // ------> requested key
6389 : //
6390 : // Legend:
6391 : // * X - page images
6392 : // ```
6393 : #[tokio::test]
6394 2 : async fn test_get_vectored_ancestor_descent() -> anyhow::Result<()> {
6395 2 : let harness = TenantHarness::create("test_get_vectored_on_lsn_axis").await?;
6396 20 : let (tenant, ctx) = harness.load().await;
6397 2 :
6398 2 : let start_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6399 2 : let end_key = start_key.add(1000);
6400 2 : let child_gap_at_key = start_key.add(500);
6401 2 : let mut parent_gap_lsns: BTreeMap<Lsn, String> = BTreeMap::new();
6402 2 :
6403 2 : let mut current_lsn = Lsn(0x10);
6404 2 :
6405 2 : let timeline_id = TimelineId::generate();
6406 2 : let parent_timeline = tenant
6407 2 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
6408 6 : .await?;
6409 2 :
6410 2 : current_lsn += 0x100;
6411 2 :
6412 8 : for _ in 0..3 {
6413 6 : let mut key = start_key;
6414 6006 : while key < end_key {
6415 6000 : current_lsn += 0x10;
6416 6000 :
6417 6000 : let image_value = format!("{} at {}", child_gap_at_key, current_lsn);
6418 2 :
6419 6000 : let mut writer = parent_timeline.writer().await;
6420 6000 : writer
6421 6000 : .put(
6422 6000 : key,
6423 6000 : current_lsn,
6424 6000 : &Value::Image(test_img(&image_value)),
6425 6000 : &ctx,
6426 6000 : )
6427 6 : .await?;
6428 6000 : writer.finish_write(current_lsn);
6429 6000 :
6430 6000 : if key == child_gap_at_key {
6431 6 : parent_gap_lsns.insert(current_lsn, image_value);
6432 5994 : }
6433 2 :
6434 6000 : key = key.next();
6435 2 : }
6436 2 :
6437 6 : parent_timeline.freeze_and_flush().await?;
6438 2 : }
6439 2 :
6440 2 : let child_timeline_id = TimelineId::generate();
6441 2 :
6442 2 : let child_timeline = tenant
6443 2 : .branch_timeline_test(&parent_timeline, child_timeline_id, Some(current_lsn), &ctx)
6444 2 : .await?;
6445 2 :
6446 2 : let mut key = start_key;
6447 2002 : while key < end_key {
6448 2000 : if key == child_gap_at_key {
6449 2 : key = key.next();
6450 2 : continue;
6451 1998 : }
6452 1998 :
6453 1998 : current_lsn += 0x10;
6454 2 :
6455 1998 : let mut writer = child_timeline.writer().await;
6456 1998 : writer
6457 1998 : .put(
6458 1998 : key,
6459 1998 : current_lsn,
6460 1998 : &Value::Image(test_img(&format!("{} at {}", key, current_lsn))),
6461 1998 : &ctx,
6462 1998 : )
6463 17 : .await?;
6464 1998 : writer.finish_write(current_lsn);
6465 1998 :
6466 1998 : key = key.next();
6467 2 : }
6468 2 :
6469 2 : child_timeline.freeze_and_flush().await?;
6470 2 :
6471 2 : let lsn_offsets: [i64; 5] = [-10, -1, 0, 1, 10];
6472 2 : let mut query_lsns = Vec::new();
6473 6 : for image_lsn in parent_gap_lsns.keys().rev() {
6474 36 : for offset in lsn_offsets {
6475 30 : query_lsns.push(Lsn(image_lsn
6476 30 : .0
6477 30 : .checked_add_signed(offset)
6478 30 : .expect("Shouldn't overflow")));
6479 30 : }
6480 2 : }
6481 2 :
6482 32 : for query_lsn in query_lsns {
6483 30 : let results = child_timeline
6484 30 : .get_vectored_impl(
6485 30 : KeySpace {
6486 30 : ranges: vec![child_gap_at_key..child_gap_at_key.next()],
6487 30 : },
6488 30 : query_lsn,
6489 30 : &mut ValuesReconstructState::new(),
6490 30 : &ctx,
6491 30 : )
6492 29 : .await;
6493 2 :
6494 30 : let expected_item = parent_gap_lsns
6495 30 : .iter()
6496 30 : .rev()
6497 68 : .find(|(lsn, _)| **lsn <= query_lsn);
6498 30 :
6499 30 : info!(
6500 2 : "Doing vectored read at LSN {}. Expecting image to be: {:?}",
6501 2 : query_lsn, expected_item
6502 2 : );
6503 2 :
6504 30 : match expected_item {
6505 26 : Some((_, img_value)) => {
6506 26 : let key_results = results.expect("No vectored get error expected");
6507 26 : let key_result = &key_results[&child_gap_at_key];
6508 26 : let returned_img = key_result
6509 26 : .as_ref()
6510 26 : .expect("No page reconstruct error expected");
6511 26 :
6512 26 : info!(
6513 2 : "Vectored read at LSN {} returned image {}",
6514 0 : query_lsn,
6515 0 : std::str::from_utf8(returned_img)?
6516 2 : );
6517 26 : assert_eq!(*returned_img, test_img(img_value));
6518 2 : }
6519 2 : None => {
6520 4 : assert!(matches!(results, Err(GetVectoredError::MissingKey(_))));
6521 2 : }
6522 2 : }
6523 2 : }
6524 2 :
6525 2 : Ok(())
6526 2 : }
6527 :
6528 : #[tokio::test]
6529 2 : async fn test_random_updates() -> anyhow::Result<()> {
6530 2 : let names_algorithms = [
6531 2 : ("test_random_updates_legacy", CompactionAlgorithm::Legacy),
6532 2 : ("test_random_updates_tiered", CompactionAlgorithm::Tiered),
6533 2 : ];
6534 6 : for (name, algorithm) in names_algorithms {
6535 96245 : test_random_updates_algorithm(name, algorithm).await?;
6536 2 : }
6537 2 : Ok(())
6538 2 : }
6539 :
6540 4 : async fn test_random_updates_algorithm(
6541 4 : name: &'static str,
6542 4 : compaction_algorithm: CompactionAlgorithm,
6543 4 : ) -> anyhow::Result<()> {
6544 4 : let mut harness = TenantHarness::create(name).await?;
6545 4 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
6546 4 : kind: compaction_algorithm,
6547 4 : };
6548 40 : let (tenant, ctx) = harness.load().await;
6549 4 : let tline = tenant
6550 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6551 11 : .await?;
6552 :
6553 : const NUM_KEYS: usize = 1000;
6554 4 : let cancel = CancellationToken::new();
6555 4 :
6556 4 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6557 4 : let mut test_key_end = test_key;
6558 4 : test_key_end.field6 = NUM_KEYS as u32;
6559 4 : tline.add_extra_test_dense_keyspace(KeySpace::single(test_key..test_key_end));
6560 4 :
6561 4 : let mut keyspace = KeySpaceAccum::new();
6562 4 :
6563 4 : // Track when each page was last modified. Used to assert that
6564 4 : // a read sees the latest page version.
6565 4 : let mut updated = [Lsn(0); NUM_KEYS];
6566 4 :
6567 4 : let mut lsn = Lsn(0x10);
6568 : #[allow(clippy::needless_range_loop)]
6569 4004 : for blknum in 0..NUM_KEYS {
6570 4000 : lsn = Lsn(lsn.0 + 0x10);
6571 4000 : test_key.field6 = blknum as u32;
6572 4000 : let mut writer = tline.writer().await;
6573 4000 : writer
6574 4000 : .put(
6575 4000 : test_key,
6576 4000 : lsn,
6577 4000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6578 4000 : &ctx,
6579 4000 : )
6580 4 : .await?;
6581 4000 : writer.finish_write(lsn);
6582 4000 : updated[blknum] = lsn;
6583 4000 : drop(writer);
6584 4000 :
6585 4000 : keyspace.add_key(test_key);
6586 : }
6587 :
6588 204 : for _ in 0..50 {
6589 200200 : for _ in 0..NUM_KEYS {
6590 200000 : lsn = Lsn(lsn.0 + 0x10);
6591 200000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6592 200000 : test_key.field6 = blknum as u32;
6593 200000 : let mut writer = tline.writer().await;
6594 200000 : writer
6595 200000 : .put(
6596 200000 : test_key,
6597 200000 : lsn,
6598 200000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6599 200000 : &ctx,
6600 200000 : )
6601 198 : .await?;
6602 200000 : writer.finish_write(lsn);
6603 200000 : drop(writer);
6604 200000 : updated[blknum] = lsn;
6605 : }
6606 :
6607 : // Read all the blocks
6608 200000 : for (blknum, last_lsn) in updated.iter().enumerate() {
6609 200000 : test_key.field6 = blknum as u32;
6610 200000 : assert_eq!(
6611 200000 : tline.get(test_key, lsn, &ctx).await?,
6612 200000 : test_img(&format!("{} at {}", blknum, last_lsn))
6613 : );
6614 : }
6615 :
6616 : // Perform a cycle of flush, and GC
6617 203 : tline.freeze_and_flush().await?;
6618 200 : tenant
6619 200 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
6620 0 : .await?;
6621 : }
6622 :
6623 4 : Ok(())
6624 4 : }
6625 :
6626 : #[tokio::test]
6627 2 : async fn test_traverse_branches() -> anyhow::Result<()> {
6628 2 : let (tenant, ctx) = TenantHarness::create("test_traverse_branches")
6629 2 : .await?
6630 2 : .load()
6631 20 : .await;
6632 2 : let mut tline = tenant
6633 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6634 6 : .await?;
6635 2 :
6636 2 : const NUM_KEYS: usize = 1000;
6637 2 :
6638 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6639 2 :
6640 2 : let mut keyspace = KeySpaceAccum::new();
6641 2 :
6642 2 : let cancel = CancellationToken::new();
6643 2 :
6644 2 : // Track when each page was last modified. Used to assert that
6645 2 : // a read sees the latest page version.
6646 2 : let mut updated = [Lsn(0); NUM_KEYS];
6647 2 :
6648 2 : let mut lsn = Lsn(0x10);
6649 2 : #[allow(clippy::needless_range_loop)]
6650 2002 : for blknum in 0..NUM_KEYS {
6651 2000 : lsn = Lsn(lsn.0 + 0x10);
6652 2000 : test_key.field6 = blknum as u32;
6653 2000 : let mut writer = tline.writer().await;
6654 2000 : writer
6655 2000 : .put(
6656 2000 : test_key,
6657 2000 : lsn,
6658 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6659 2000 : &ctx,
6660 2000 : )
6661 2 : .await?;
6662 2000 : writer.finish_write(lsn);
6663 2000 : updated[blknum] = lsn;
6664 2000 : drop(writer);
6665 2000 :
6666 2000 : keyspace.add_key(test_key);
6667 2 : }
6668 2 :
6669 102 : for _ in 0..50 {
6670 100 : let new_tline_id = TimelineId::generate();
6671 100 : tenant
6672 100 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
6673 91 : .await?;
6674 100 : tline = tenant
6675 100 : .get_timeline(new_tline_id, true)
6676 100 : .expect("Should have the branched timeline");
6677 2 :
6678 100100 : for _ in 0..NUM_KEYS {
6679 100000 : lsn = Lsn(lsn.0 + 0x10);
6680 100000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6681 100000 : test_key.field6 = blknum as u32;
6682 100000 : let mut writer = tline.writer().await;
6683 100000 : writer
6684 100000 : .put(
6685 100000 : test_key,
6686 100000 : lsn,
6687 100000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6688 100000 : &ctx,
6689 100000 : )
6690 835 : .await?;
6691 100000 : println!("updating {} at {}", blknum, lsn);
6692 100000 : writer.finish_write(lsn);
6693 100000 : drop(writer);
6694 100000 : updated[blknum] = lsn;
6695 2 : }
6696 2 :
6697 2 : // Read all the blocks
6698 100000 : for (blknum, last_lsn) in updated.iter().enumerate() {
6699 100000 : test_key.field6 = blknum as u32;
6700 100000 : assert_eq!(
6701 100000 : tline.get(test_key, lsn, &ctx).await?,
6702 100000 : test_img(&format!("{} at {}", blknum, last_lsn))
6703 2 : );
6704 2 : }
6705 2 :
6706 2 : // Perform a cycle of flush, compact, and GC
6707 102 : tline.freeze_and_flush().await?;
6708 15066 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
6709 100 : tenant
6710 100 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
6711 2 : .await?;
6712 2 : }
6713 2 :
6714 2 : Ok(())
6715 2 : }
6716 :
6717 : #[tokio::test]
6718 2 : async fn test_traverse_ancestors() -> anyhow::Result<()> {
6719 2 : let (tenant, ctx) = TenantHarness::create("test_traverse_ancestors")
6720 2 : .await?
6721 2 : .load()
6722 19 : .await;
6723 2 : let mut tline = tenant
6724 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6725 6 : .await?;
6726 2 :
6727 2 : const NUM_KEYS: usize = 100;
6728 2 : const NUM_TLINES: usize = 50;
6729 2 :
6730 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6731 2 : // Track page mutation lsns across different timelines.
6732 2 : let mut updated = [[Lsn(0); NUM_KEYS]; NUM_TLINES];
6733 2 :
6734 2 : let mut lsn = Lsn(0x10);
6735 2 :
6736 2 : #[allow(clippy::needless_range_loop)]
6737 102 : for idx in 0..NUM_TLINES {
6738 100 : let new_tline_id = TimelineId::generate();
6739 100 : tenant
6740 100 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
6741 63 : .await?;
6742 100 : tline = tenant
6743 100 : .get_timeline(new_tline_id, true)
6744 100 : .expect("Should have the branched timeline");
6745 2 :
6746 10100 : for _ in 0..NUM_KEYS {
6747 10000 : lsn = Lsn(lsn.0 + 0x10);
6748 10000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6749 10000 : test_key.field6 = blknum as u32;
6750 10000 : let mut writer = tline.writer().await;
6751 10000 : writer
6752 10000 : .put(
6753 10000 : test_key,
6754 10000 : lsn,
6755 10000 : &Value::Image(test_img(&format!("{} {} at {}", idx, blknum, lsn))),
6756 10000 : &ctx,
6757 10000 : )
6758 114 : .await?;
6759 10000 : println!("updating [{}][{}] at {}", idx, blknum, lsn);
6760 10000 : writer.finish_write(lsn);
6761 10000 : drop(writer);
6762 10000 : updated[idx][blknum] = lsn;
6763 2 : }
6764 2 : }
6765 2 :
6766 2 : // Read pages from leaf timeline across all ancestors.
6767 100 : for (idx, lsns) in updated.iter().enumerate() {
6768 10000 : for (blknum, lsn) in lsns.iter().enumerate() {
6769 2 : // Skip empty mutations.
6770 10000 : if lsn.0 == 0 {
6771 3602 : continue;
6772 6398 : }
6773 6398 : println!("checking [{idx}][{blknum}] at {lsn}");
6774 6398 : test_key.field6 = blknum as u32;
6775 6398 : assert_eq!(
6776 6398 : tline.get(test_key, *lsn, &ctx).await?,
6777 6398 : test_img(&format!("{idx} {blknum} at {lsn}"))
6778 2 : );
6779 2 : }
6780 2 : }
6781 2 : Ok(())
6782 2 : }
6783 :
6784 : #[tokio::test]
6785 2 : async fn test_write_at_initdb_lsn_takes_optimization_code_path() -> anyhow::Result<()> {
6786 2 : let (tenant, ctx) = TenantHarness::create("test_empty_test_timeline_is_usable")
6787 2 : .await?
6788 2 : .load()
6789 20 : .await;
6790 2 :
6791 2 : let initdb_lsn = Lsn(0x20);
6792 2 : let utline = tenant
6793 2 : .create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)
6794 2 : .await?;
6795 2 : let tline = utline.raw_timeline().unwrap();
6796 2 :
6797 2 : // Spawn flush loop now so that we can set the `expect_initdb_optimization`
6798 2 : tline.maybe_spawn_flush_loop();
6799 2 :
6800 2 : // Make sure the timeline has the minimum set of required keys for operation.
6801 2 : // The only operation you can always do on an empty timeline is to `put` new data.
6802 2 : // Except if you `put` at `initdb_lsn`.
6803 2 : // In that case, there's an optimization to directly create image layers instead of delta layers.
6804 2 : // It uses `repartition()`, which assumes some keys to be present.
6805 2 : // Let's make sure the test timeline can handle that case.
6806 2 : {
6807 2 : let mut state = tline.flush_loop_state.lock().unwrap();
6808 2 : assert_eq!(
6809 2 : timeline::FlushLoopState::Running {
6810 2 : expect_initdb_optimization: false,
6811 2 : initdb_optimization_count: 0,
6812 2 : },
6813 2 : *state
6814 2 : );
6815 2 : *state = timeline::FlushLoopState::Running {
6816 2 : expect_initdb_optimization: true,
6817 2 : initdb_optimization_count: 0,
6818 2 : };
6819 2 : }
6820 2 :
6821 2 : // Make writes at the initdb_lsn. When we flush it below, it should be handled by the optimization.
6822 2 : // As explained above, the optimization requires some keys to be present.
6823 2 : // As per `create_empty_timeline` documentation, use init_empty to set them.
6824 2 : // This is what `create_test_timeline` does, by the way.
6825 2 : let mut modification = tline.begin_modification(initdb_lsn);
6826 2 : modification
6827 2 : .init_empty_test_timeline()
6828 2 : .context("init_empty_test_timeline")?;
6829 2 : modification
6830 2 : .commit(&ctx)
6831 2 : .await
6832 2 : .context("commit init_empty_test_timeline modification")?;
6833 2 :
6834 2 : // Do the flush. The flush code will check the expectations that we set above.
6835 2 : tline.freeze_and_flush().await?;
6836 2 :
6837 2 : // assert freeze_and_flush exercised the initdb optimization
6838 2 : {
6839 2 : let state = tline.flush_loop_state.lock().unwrap();
6840 2 : let timeline::FlushLoopState::Running {
6841 2 : expect_initdb_optimization,
6842 2 : initdb_optimization_count,
6843 2 : } = *state
6844 2 : else {
6845 2 : panic!("unexpected state: {:?}", *state);
6846 2 : };
6847 2 : assert!(expect_initdb_optimization);
6848 2 : assert!(initdb_optimization_count > 0);
6849 2 : }
6850 2 : Ok(())
6851 2 : }
6852 :
6853 : #[tokio::test]
6854 2 : async fn test_create_guard_crash() -> anyhow::Result<()> {
6855 2 : let name = "test_create_guard_crash";
6856 2 : let harness = TenantHarness::create(name).await?;
6857 2 : {
6858 19 : let (tenant, ctx) = harness.load().await;
6859 2 : let tline = tenant
6860 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
6861 2 : .await?;
6862 2 : // Leave the timeline ID in [`Tenant::timelines_creating`] to exclude attempting to create it again
6863 2 : let raw_tline = tline.raw_timeline().unwrap();
6864 2 : raw_tline
6865 2 : .shutdown(super::timeline::ShutdownMode::Hard)
6866 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))
6867 2 : .await;
6868 2 : std::mem::forget(tline);
6869 2 : }
6870 2 :
6871 20 : let (tenant, _) = harness.load().await;
6872 2 : match tenant.get_timeline(TIMELINE_ID, false) {
6873 2 : Ok(_) => panic!("timeline should've been removed during load"),
6874 2 : Err(e) => {
6875 2 : assert_eq!(
6876 2 : e,
6877 2 : GetTimelineError::NotFound {
6878 2 : tenant_id: tenant.tenant_shard_id,
6879 2 : timeline_id: TIMELINE_ID,
6880 2 : }
6881 2 : )
6882 2 : }
6883 2 : }
6884 2 :
6885 2 : assert!(!harness
6886 2 : .conf
6887 2 : .timeline_path(&tenant.tenant_shard_id, &TIMELINE_ID)
6888 2 : .exists());
6889 2 :
6890 2 : Ok(())
6891 2 : }
6892 :
6893 : #[tokio::test]
6894 2 : async fn test_read_at_max_lsn() -> anyhow::Result<()> {
6895 2 : let names_algorithms = [
6896 2 : ("test_read_at_max_lsn_legacy", CompactionAlgorithm::Legacy),
6897 2 : ("test_read_at_max_lsn_tiered", CompactionAlgorithm::Tiered),
6898 2 : ];
6899 6 : for (name, algorithm) in names_algorithms {
6900 32561 : test_read_at_max_lsn_algorithm(name, algorithm).await?;
6901 2 : }
6902 2 : Ok(())
6903 2 : }
6904 :
6905 4 : async fn test_read_at_max_lsn_algorithm(
6906 4 : name: &'static str,
6907 4 : compaction_algorithm: CompactionAlgorithm,
6908 4 : ) -> anyhow::Result<()> {
6909 4 : let mut harness = TenantHarness::create(name).await?;
6910 4 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
6911 4 : kind: compaction_algorithm,
6912 4 : };
6913 40 : let (tenant, ctx) = harness.load().await;
6914 4 : let tline = tenant
6915 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6916 11 : .await?;
6917 :
6918 4 : let lsn = Lsn(0x10);
6919 4 : let compact = false;
6920 32100 : bulk_insert_maybe_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000, compact).await?;
6921 :
6922 4 : let test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6923 4 : let read_lsn = Lsn(u64::MAX - 1);
6924 :
6925 410 : let result = tline.get(test_key, read_lsn, &ctx).await;
6926 4 : assert!(result.is_ok(), "result is not Ok: {}", result.unwrap_err());
6927 :
6928 4 : Ok(())
6929 4 : }
6930 :
6931 : #[tokio::test]
6932 2 : async fn test_metadata_scan() -> anyhow::Result<()> {
6933 2 : let harness = TenantHarness::create("test_metadata_scan").await?;
6934 20 : let (tenant, ctx) = harness.load().await;
6935 2 : let tline = tenant
6936 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6937 6 : .await?;
6938 2 :
6939 2 : const NUM_KEYS: usize = 1000;
6940 2 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
6941 2 :
6942 2 : let cancel = CancellationToken::new();
6943 2 :
6944 2 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
6945 2 : base_key.field1 = AUX_KEY_PREFIX;
6946 2 : let mut test_key = base_key;
6947 2 :
6948 2 : // Track when each page was last modified. Used to assert that
6949 2 : // a read sees the latest page version.
6950 2 : let mut updated = [Lsn(0); NUM_KEYS];
6951 2 :
6952 2 : let mut lsn = Lsn(0x10);
6953 2 : #[allow(clippy::needless_range_loop)]
6954 2002 : for blknum in 0..NUM_KEYS {
6955 2000 : lsn = Lsn(lsn.0 + 0x10);
6956 2000 : test_key.field6 = (blknum * STEP) as u32;
6957 2000 : let mut writer = tline.writer().await;
6958 2000 : writer
6959 2000 : .put(
6960 2000 : test_key,
6961 2000 : lsn,
6962 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6963 2000 : &ctx,
6964 2000 : )
6965 2 : .await?;
6966 2000 : writer.finish_write(lsn);
6967 2000 : updated[blknum] = lsn;
6968 2000 : drop(writer);
6969 2 : }
6970 2 :
6971 2 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
6972 2 :
6973 24 : for iter in 0..=10 {
6974 2 : // Read all the blocks
6975 22000 : for (blknum, last_lsn) in updated.iter().enumerate() {
6976 22000 : test_key.field6 = (blknum * STEP) as u32;
6977 22000 : assert_eq!(
6978 22000 : tline.get(test_key, lsn, &ctx).await?,
6979 22000 : test_img(&format!("{} at {}", blknum, last_lsn))
6980 2 : );
6981 2 : }
6982 2 :
6983 22 : let mut cnt = 0;
6984 22000 : for (key, value) in tline
6985 22 : .get_vectored_impl(
6986 22 : keyspace.clone(),
6987 22 : lsn,
6988 22 : &mut ValuesReconstructState::default(),
6989 22 : &ctx,
6990 22 : )
6991 721 : .await?
6992 2 : {
6993 22000 : let blknum = key.field6 as usize;
6994 22000 : let value = value?;
6995 22000 : assert!(blknum % STEP == 0);
6996 22000 : let blknum = blknum / STEP;
6997 22000 : assert_eq!(
6998 22000 : value,
6999 22000 : test_img(&format!("{} at {}", blknum, updated[blknum]))
7000 22000 : );
7001 22000 : cnt += 1;
7002 2 : }
7003 2 :
7004 22 : assert_eq!(cnt, NUM_KEYS);
7005 2 :
7006 22022 : for _ in 0..NUM_KEYS {
7007 22000 : lsn = Lsn(lsn.0 + 0x10);
7008 22000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7009 22000 : test_key.field6 = (blknum * STEP) as u32;
7010 22000 : let mut writer = tline.writer().await;
7011 22000 : writer
7012 22000 : .put(
7013 22000 : test_key,
7014 22000 : lsn,
7015 22000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7016 22000 : &ctx,
7017 22000 : )
7018 111 : .await?;
7019 22000 : writer.finish_write(lsn);
7020 22000 : drop(writer);
7021 22000 : updated[blknum] = lsn;
7022 2 : }
7023 2 :
7024 2 : // Perform two cycles of flush, compact, and GC
7025 66 : for round in 0..2 {
7026 44 : tline.freeze_and_flush().await?;
7027 44 : tline
7028 44 : .compact(
7029 44 : &cancel,
7030 44 : if iter % 5 == 0 && round == 0 {
7031 6 : let mut flags = EnumSet::new();
7032 6 : flags.insert(CompactFlags::ForceImageLayerCreation);
7033 6 : flags.insert(CompactFlags::ForceRepartition);
7034 6 : flags
7035 2 : } else {
7036 38 : EnumSet::empty()
7037 2 : },
7038 44 : &ctx,
7039 2 : )
7040 6681 : .await?;
7041 44 : tenant
7042 44 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7043 2 : .await?;
7044 2 : }
7045 2 : }
7046 2 :
7047 2 : Ok(())
7048 2 : }
7049 :
7050 : #[tokio::test]
7051 2 : async fn test_metadata_compaction_trigger() -> anyhow::Result<()> {
7052 2 : let harness = TenantHarness::create("test_metadata_compaction_trigger").await?;
7053 20 : let (tenant, ctx) = harness.load().await;
7054 2 : let tline = tenant
7055 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7056 6 : .await?;
7057 2 :
7058 2 : let cancel = CancellationToken::new();
7059 2 :
7060 2 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7061 2 : base_key.field1 = AUX_KEY_PREFIX;
7062 2 : let test_key = base_key;
7063 2 : let mut lsn = Lsn(0x10);
7064 2 :
7065 42 : for _ in 0..20 {
7066 40 : lsn = Lsn(lsn.0 + 0x10);
7067 40 : let mut writer = tline.writer().await;
7068 40 : writer
7069 40 : .put(
7070 40 : test_key,
7071 40 : lsn,
7072 40 : &Value::Image(test_img(&format!("{} at {}", 0, lsn))),
7073 40 : &ctx,
7074 40 : )
7075 20 : .await?;
7076 40 : writer.finish_write(lsn);
7077 40 : drop(writer);
7078 40 : tline.freeze_and_flush().await?; // force create a delta layer
7079 2 : }
7080 2 :
7081 2 : let before_num_l0_delta_files =
7082 2 : tline.layers.read().await.layer_map()?.level0_deltas().len();
7083 2 :
7084 110 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
7085 2 :
7086 2 : let after_num_l0_delta_files = tline.layers.read().await.layer_map()?.level0_deltas().len();
7087 2 :
7088 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}");
7089 2 :
7090 2 : assert_eq!(
7091 4 : tline.get(test_key, lsn, &ctx).await?,
7092 2 : test_img(&format!("{} at {}", 0, lsn))
7093 2 : );
7094 2 :
7095 2 : Ok(())
7096 2 : }
7097 :
7098 : #[tokio::test]
7099 2 : async fn test_aux_file_e2e() {
7100 2 : let harness = TenantHarness::create("test_aux_file_e2e").await.unwrap();
7101 2 :
7102 20 : let (tenant, ctx) = harness.load().await;
7103 2 :
7104 2 : let mut lsn = Lsn(0x08);
7105 2 :
7106 2 : let tline: Arc<Timeline> = tenant
7107 2 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
7108 6 : .await
7109 2 : .unwrap();
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/test1", b"first", &ctx)
7116 2 : .await
7117 2 : .unwrap();
7118 2 : modification.commit(&ctx).await.unwrap();
7119 2 : }
7120 2 :
7121 2 : // we can read everything from the storage
7122 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
7123 2 : assert_eq!(
7124 2 : files.get("pg_logical/mappings/test1"),
7125 2 : Some(&bytes::Bytes::from_static(b"first"))
7126 2 : );
7127 2 :
7128 2 : {
7129 2 : lsn += 8;
7130 2 : let mut modification = tline.begin_modification(lsn);
7131 2 : modification
7132 2 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
7133 2 : .await
7134 2 : .unwrap();
7135 2 : modification.commit(&ctx).await.unwrap();
7136 2 : }
7137 2 :
7138 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
7139 2 : assert_eq!(
7140 2 : files.get("pg_logical/mappings/test2"),
7141 2 : Some(&bytes::Bytes::from_static(b"second"))
7142 2 : );
7143 2 :
7144 2 : let child = tenant
7145 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(lsn), &ctx)
7146 2 : .await
7147 2 : .unwrap();
7148 2 :
7149 2 : let files = child.list_aux_files(lsn, &ctx).await.unwrap();
7150 2 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
7151 2 : assert_eq!(files.get("pg_logical/mappings/test2"), None);
7152 2 : }
7153 :
7154 : #[tokio::test]
7155 2 : async fn test_metadata_image_creation() -> anyhow::Result<()> {
7156 2 : let harness = TenantHarness::create("test_metadata_image_creation").await?;
7157 20 : let (tenant, ctx) = harness.load().await;
7158 2 : let tline = tenant
7159 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7160 6 : .await?;
7161 2 :
7162 2 : const NUM_KEYS: usize = 1000;
7163 2 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
7164 2 :
7165 2 : let cancel = CancellationToken::new();
7166 2 :
7167 2 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7168 2 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
7169 2 : let mut test_key = base_key;
7170 2 : let mut lsn = Lsn(0x10);
7171 2 :
7172 8 : async fn scan_with_statistics(
7173 8 : tline: &Timeline,
7174 8 : keyspace: &KeySpace,
7175 8 : lsn: Lsn,
7176 8 : ctx: &RequestContext,
7177 8 : ) -> anyhow::Result<(BTreeMap<Key, Result<Bytes, PageReconstructError>>, usize)> {
7178 8 : let mut reconstruct_state = ValuesReconstructState::default();
7179 8 : let res = tline
7180 8 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
7181 244 : .await?;
7182 8 : Ok((res, reconstruct_state.get_delta_layers_visited() as usize))
7183 8 : }
7184 2 :
7185 2 : #[allow(clippy::needless_range_loop)]
7186 2002 : for blknum in 0..NUM_KEYS {
7187 2000 : lsn = Lsn(lsn.0 + 0x10);
7188 2000 : test_key.field6 = (blknum * STEP) as u32;
7189 2000 : let mut writer = tline.writer().await;
7190 2000 : writer
7191 2000 : .put(
7192 2000 : test_key,
7193 2000 : lsn,
7194 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7195 2000 : &ctx,
7196 2000 : )
7197 2 : .await?;
7198 2000 : writer.finish_write(lsn);
7199 2000 : drop(writer);
7200 2 : }
7201 2 :
7202 2 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
7203 2 :
7204 22 : for iter in 1..=10 {
7205 20020 : for _ in 0..NUM_KEYS {
7206 20000 : lsn = Lsn(lsn.0 + 0x10);
7207 20000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7208 20000 : test_key.field6 = (blknum * STEP) as u32;
7209 20000 : let mut writer = tline.writer().await;
7210 20000 : writer
7211 20000 : .put(
7212 20000 : test_key,
7213 20000 : lsn,
7214 20000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7215 20000 : &ctx,
7216 20000 : )
7217 35 : .await?;
7218 20000 : writer.finish_write(lsn);
7219 20000 : drop(writer);
7220 2 : }
7221 2 :
7222 20 : tline.freeze_and_flush().await?;
7223 2 :
7224 20 : if iter % 5 == 0 {
7225 4 : let (_, before_delta_file_accessed) =
7226 236 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
7227 4 : tline
7228 4 : .compact(
7229 4 : &cancel,
7230 4 : {
7231 4 : let mut flags = EnumSet::new();
7232 4 : flags.insert(CompactFlags::ForceImageLayerCreation);
7233 4 : flags.insert(CompactFlags::ForceRepartition);
7234 4 : flags
7235 4 : },
7236 4 : &ctx,
7237 4 : )
7238 4837 : .await?;
7239 4 : let (_, after_delta_file_accessed) =
7240 8 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
7241 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}");
7242 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.
7243 4 : assert!(
7244 4 : after_delta_file_accessed <= 2,
7245 2 : "after_delta_file_accessed={after_delta_file_accessed}"
7246 2 : );
7247 16 : }
7248 2 : }
7249 2 :
7250 2 : Ok(())
7251 2 : }
7252 :
7253 : #[tokio::test]
7254 2 : async fn test_vectored_missing_data_key_reads() -> anyhow::Result<()> {
7255 2 : let harness = TenantHarness::create("test_vectored_missing_data_key_reads").await?;
7256 20 : let (tenant, ctx) = harness.load().await;
7257 2 :
7258 2 : let base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7259 2 : let base_key_child = Key::from_hex("000000000033333333444444445500000001").unwrap();
7260 2 : let base_key_nonexist = Key::from_hex("000000000033333333444444445500000002").unwrap();
7261 2 :
7262 2 : let tline = tenant
7263 2 : .create_test_timeline_with_layers(
7264 2 : TIMELINE_ID,
7265 2 : Lsn(0x10),
7266 2 : DEFAULT_PG_VERSION,
7267 2 : &ctx,
7268 2 : Vec::new(), // delta layers
7269 2 : vec![(Lsn(0x20), vec![(base_key, test_img("data key 1"))])], // image layers
7270 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
7271 2 : )
7272 13 : .await?;
7273 2 : tline.add_extra_test_dense_keyspace(KeySpace::single(base_key..(base_key_nonexist.next())));
7274 2 :
7275 2 : let child = tenant
7276 2 : .branch_timeline_test_with_layers(
7277 2 : &tline,
7278 2 : NEW_TIMELINE_ID,
7279 2 : Some(Lsn(0x20)),
7280 2 : &ctx,
7281 2 : Vec::new(), // delta layers
7282 2 : vec![(Lsn(0x30), vec![(base_key_child, test_img("data key 2"))])], // image layers
7283 2 : Lsn(0x30),
7284 2 : )
7285 9 : .await
7286 2 : .unwrap();
7287 2 :
7288 2 : let lsn = Lsn(0x30);
7289 2 :
7290 2 : // test vectored get on parent timeline
7291 2 : assert_eq!(
7292 4 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
7293 2 : Some(test_img("data key 1"))
7294 2 : );
7295 2 : assert!(get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx)
7296 3 : .await
7297 2 : .unwrap_err()
7298 2 : .is_missing_key_error());
7299 2 : assert!(
7300 2 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx)
7301 2 : .await
7302 2 : .unwrap_err()
7303 2 : .is_missing_key_error()
7304 2 : );
7305 2 :
7306 2 : // test vectored get on child timeline
7307 2 : assert_eq!(
7308 2 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
7309 2 : Some(test_img("data key 1"))
7310 2 : );
7311 2 : assert_eq!(
7312 4 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
7313 2 : Some(test_img("data key 2"))
7314 2 : );
7315 2 : assert!(
7316 2 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx)
7317 2 : .await
7318 2 : .unwrap_err()
7319 2 : .is_missing_key_error()
7320 2 : );
7321 2 :
7322 2 : Ok(())
7323 2 : }
7324 :
7325 : #[tokio::test]
7326 2 : async fn test_vectored_missing_metadata_key_reads() -> anyhow::Result<()> {
7327 2 : let harness = TenantHarness::create("test_vectored_missing_metadata_key_reads").await?;
7328 20 : let (tenant, ctx) = harness.load().await;
7329 2 :
7330 2 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7331 2 : let base_key_child = Key::from_hex("620000000033333333444444445500000001").unwrap();
7332 2 : let base_key_nonexist = Key::from_hex("620000000033333333444444445500000002").unwrap();
7333 2 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
7334 2 :
7335 2 : let tline = tenant
7336 2 : .create_test_timeline_with_layers(
7337 2 : TIMELINE_ID,
7338 2 : Lsn(0x10),
7339 2 : DEFAULT_PG_VERSION,
7340 2 : &ctx,
7341 2 : Vec::new(), // delta layers
7342 2 : vec![(Lsn(0x20), vec![(base_key, test_img("metadata key 1"))])], // image layers
7343 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
7344 2 : )
7345 13 : .await?;
7346 2 :
7347 2 : let child = tenant
7348 2 : .branch_timeline_test_with_layers(
7349 2 : &tline,
7350 2 : NEW_TIMELINE_ID,
7351 2 : Some(Lsn(0x20)),
7352 2 : &ctx,
7353 2 : Vec::new(), // delta layers
7354 2 : vec![(
7355 2 : Lsn(0x30),
7356 2 : vec![(base_key_child, test_img("metadata key 2"))],
7357 2 : )], // image layers
7358 2 : Lsn(0x30),
7359 2 : )
7360 9 : .await
7361 2 : .unwrap();
7362 2 :
7363 2 : let lsn = Lsn(0x30);
7364 2 :
7365 2 : // test vectored get on parent timeline
7366 2 : assert_eq!(
7367 4 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
7368 2 : Some(test_img("metadata key 1"))
7369 2 : );
7370 2 : assert_eq!(
7371 2 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx).await?,
7372 2 : None
7373 2 : );
7374 2 : assert_eq!(
7375 2 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx).await?,
7376 2 : None
7377 2 : );
7378 2 :
7379 2 : // test vectored get on child timeline
7380 2 : assert_eq!(
7381 2 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
7382 2 : None
7383 2 : );
7384 2 : assert_eq!(
7385 4 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
7386 2 : Some(test_img("metadata key 2"))
7387 2 : );
7388 2 : assert_eq!(
7389 2 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx).await?,
7390 2 : None
7391 2 : );
7392 2 :
7393 2 : Ok(())
7394 2 : }
7395 :
7396 36 : async fn get_vectored_impl_wrapper(
7397 36 : tline: &Arc<Timeline>,
7398 36 : key: Key,
7399 36 : lsn: Lsn,
7400 36 : ctx: &RequestContext,
7401 36 : ) -> Result<Option<Bytes>, GetVectoredError> {
7402 36 : let mut reconstruct_state = ValuesReconstructState::new();
7403 36 : let mut res = tline
7404 36 : .get_vectored_impl(
7405 36 : KeySpace::single(key..key.next()),
7406 36 : lsn,
7407 36 : &mut reconstruct_state,
7408 36 : ctx,
7409 36 : )
7410 40 : .await?;
7411 30 : Ok(res.pop_last().map(|(k, v)| {
7412 18 : assert_eq!(k, key);
7413 18 : v.unwrap()
7414 30 : }))
7415 36 : }
7416 :
7417 : #[tokio::test]
7418 2 : async fn test_metadata_tombstone_reads() -> anyhow::Result<()> {
7419 2 : let harness = TenantHarness::create("test_metadata_tombstone_reads").await?;
7420 20 : let (tenant, ctx) = harness.load().await;
7421 2 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
7422 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
7423 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
7424 2 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
7425 2 :
7426 2 : // We emulate the situation that the compaction algorithm creates an image layer that removes the tombstones
7427 2 : // Lsn 0x30 key0, key3, no key1+key2
7428 2 : // Lsn 0x20 key1+key2 tomestones
7429 2 : // Lsn 0x10 key1 in image, key2 in delta
7430 2 : let tline = tenant
7431 2 : .create_test_timeline_with_layers(
7432 2 : TIMELINE_ID,
7433 2 : Lsn(0x10),
7434 2 : DEFAULT_PG_VERSION,
7435 2 : &ctx,
7436 2 : // delta layers
7437 2 : vec![
7438 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7439 2 : Lsn(0x10)..Lsn(0x20),
7440 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
7441 2 : ),
7442 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7443 2 : Lsn(0x20)..Lsn(0x30),
7444 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
7445 2 : ),
7446 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7447 2 : Lsn(0x20)..Lsn(0x30),
7448 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
7449 2 : ),
7450 2 : ],
7451 2 : // image layers
7452 2 : vec![
7453 2 : (Lsn(0x10), vec![(key1, test_img("metadata key 1"))]),
7454 2 : (
7455 2 : Lsn(0x30),
7456 2 : vec![
7457 2 : (key0, test_img("metadata key 0")),
7458 2 : (key3, test_img("metadata key 3")),
7459 2 : ],
7460 2 : ),
7461 2 : ],
7462 2 : Lsn(0x30),
7463 2 : )
7464 40 : .await?;
7465 2 :
7466 2 : let lsn = Lsn(0x30);
7467 2 : let old_lsn = Lsn(0x20);
7468 2 :
7469 2 : assert_eq!(
7470 4 : get_vectored_impl_wrapper(&tline, key0, lsn, &ctx).await?,
7471 2 : Some(test_img("metadata key 0"))
7472 2 : );
7473 2 : assert_eq!(
7474 2 : get_vectored_impl_wrapper(&tline, key1, lsn, &ctx).await?,
7475 2 : None,
7476 2 : );
7477 2 : assert_eq!(
7478 2 : get_vectored_impl_wrapper(&tline, key2, lsn, &ctx).await?,
7479 2 : None,
7480 2 : );
7481 2 : assert_eq!(
7482 8 : get_vectored_impl_wrapper(&tline, key1, old_lsn, &ctx).await?,
7483 2 : Some(Bytes::new()),
7484 2 : );
7485 2 : assert_eq!(
7486 7 : get_vectored_impl_wrapper(&tline, key2, old_lsn, &ctx).await?,
7487 2 : Some(Bytes::new()),
7488 2 : );
7489 2 : assert_eq!(
7490 2 : get_vectored_impl_wrapper(&tline, key3, lsn, &ctx).await?,
7491 2 : Some(test_img("metadata key 3"))
7492 2 : );
7493 2 :
7494 2 : Ok(())
7495 2 : }
7496 :
7497 : #[tokio::test]
7498 2 : async fn test_metadata_tombstone_image_creation() {
7499 2 : let harness = TenantHarness::create("test_metadata_tombstone_image_creation")
7500 2 : .await
7501 2 : .unwrap();
7502 20 : let (tenant, ctx) = harness.load().await;
7503 2 :
7504 2 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
7505 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
7506 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
7507 2 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
7508 2 :
7509 2 : let tline = tenant
7510 2 : .create_test_timeline_with_layers(
7511 2 : TIMELINE_ID,
7512 2 : Lsn(0x10),
7513 2 : DEFAULT_PG_VERSION,
7514 2 : &ctx,
7515 2 : // delta layers
7516 2 : vec![
7517 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7518 2 : Lsn(0x10)..Lsn(0x20),
7519 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
7520 2 : ),
7521 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7522 2 : Lsn(0x20)..Lsn(0x30),
7523 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
7524 2 : ),
7525 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7526 2 : Lsn(0x20)..Lsn(0x30),
7527 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
7528 2 : ),
7529 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7530 2 : Lsn(0x30)..Lsn(0x40),
7531 2 : vec![
7532 2 : (key0, Lsn(0x30), Value::Image(test_img("metadata key 0"))),
7533 2 : (key3, Lsn(0x30), Value::Image(test_img("metadata key 3"))),
7534 2 : ],
7535 2 : ),
7536 2 : ],
7537 2 : // image layers
7538 2 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
7539 2 : Lsn(0x40),
7540 2 : )
7541 37 : .await
7542 2 : .unwrap();
7543 2 :
7544 2 : let cancel = CancellationToken::new();
7545 2 :
7546 2 : tline
7547 2 : .compact(
7548 2 : &cancel,
7549 2 : {
7550 2 : let mut flags = EnumSet::new();
7551 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
7552 2 : flags.insert(CompactFlags::ForceRepartition);
7553 2 : flags
7554 2 : },
7555 2 : &ctx,
7556 2 : )
7557 62 : .await
7558 2 : .unwrap();
7559 2 :
7560 2 : // Image layers are created at last_record_lsn
7561 2 : let images = tline
7562 2 : .inspect_image_layers(Lsn(0x40), &ctx)
7563 8 : .await
7564 2 : .unwrap()
7565 2 : .into_iter()
7566 18 : .filter(|(k, _)| k.is_metadata_key())
7567 2 : .collect::<Vec<_>>();
7568 2 : assert_eq!(images.len(), 2); // the image layer should only contain two existing keys, tombstones should be removed.
7569 2 : }
7570 :
7571 : #[tokio::test]
7572 2 : async fn test_metadata_tombstone_empty_image_creation() {
7573 2 : let harness = TenantHarness::create("test_metadata_tombstone_empty_image_creation")
7574 2 : .await
7575 2 : .unwrap();
7576 20 : let (tenant, ctx) = harness.load().await;
7577 2 :
7578 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
7579 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
7580 2 :
7581 2 : let tline = tenant
7582 2 : .create_test_timeline_with_layers(
7583 2 : TIMELINE_ID,
7584 2 : Lsn(0x10),
7585 2 : DEFAULT_PG_VERSION,
7586 2 : &ctx,
7587 2 : // delta layers
7588 2 : vec![
7589 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7590 2 : Lsn(0x10)..Lsn(0x20),
7591 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
7592 2 : ),
7593 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7594 2 : Lsn(0x20)..Lsn(0x30),
7595 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
7596 2 : ),
7597 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7598 2 : Lsn(0x20)..Lsn(0x30),
7599 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
7600 2 : ),
7601 2 : ],
7602 2 : // image layers
7603 2 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
7604 2 : Lsn(0x30),
7605 2 : )
7606 31 : .await
7607 2 : .unwrap();
7608 2 :
7609 2 : let cancel = CancellationToken::new();
7610 2 :
7611 2 : tline
7612 2 : .compact(
7613 2 : &cancel,
7614 2 : {
7615 2 : let mut flags = EnumSet::new();
7616 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
7617 2 : flags.insert(CompactFlags::ForceRepartition);
7618 2 : flags
7619 2 : },
7620 2 : &ctx,
7621 2 : )
7622 50 : .await
7623 2 : .unwrap();
7624 2 :
7625 2 : // Image layers are created at last_record_lsn
7626 2 : let images = tline
7627 2 : .inspect_image_layers(Lsn(0x30), &ctx)
7628 4 : .await
7629 2 : .unwrap()
7630 2 : .into_iter()
7631 14 : .filter(|(k, _)| k.is_metadata_key())
7632 2 : .collect::<Vec<_>>();
7633 2 : assert_eq!(images.len(), 0); // the image layer should not contain tombstones, or it is not created
7634 2 : }
7635 :
7636 : #[tokio::test]
7637 2 : async fn test_simple_bottom_most_compaction_images() -> anyhow::Result<()> {
7638 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_images").await?;
7639 20 : let (tenant, ctx) = harness.load().await;
7640 2 :
7641 102 : fn get_key(id: u32) -> Key {
7642 102 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7643 102 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7644 102 : key.field6 = id;
7645 102 : key
7646 102 : }
7647 2 :
7648 2 : // We create
7649 2 : // - one bottom-most image layer,
7650 2 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
7651 2 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
7652 2 : // - a delta layer D3 above the horizon.
7653 2 : //
7654 2 : // | D3 |
7655 2 : // | D1 |
7656 2 : // -| |-- gc horizon -----------------
7657 2 : // | | | D2 |
7658 2 : // --------- img layer ------------------
7659 2 : //
7660 2 : // What we should expact from this compaction is:
7661 2 : // | D3 |
7662 2 : // | Part of D1 |
7663 2 : // --------- img layer with D1+D2 at GC horizon------------------
7664 2 :
7665 2 : // img layer at 0x10
7666 2 : let img_layer = (0..10)
7667 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
7668 2 : .collect_vec();
7669 2 :
7670 2 : let delta1 = vec![
7671 2 : (
7672 2 : get_key(1),
7673 2 : Lsn(0x20),
7674 2 : Value::Image(Bytes::from("value 1@0x20")),
7675 2 : ),
7676 2 : (
7677 2 : get_key(2),
7678 2 : Lsn(0x30),
7679 2 : Value::Image(Bytes::from("value 2@0x30")),
7680 2 : ),
7681 2 : (
7682 2 : get_key(3),
7683 2 : Lsn(0x40),
7684 2 : Value::Image(Bytes::from("value 3@0x40")),
7685 2 : ),
7686 2 : ];
7687 2 : let delta2 = vec![
7688 2 : (
7689 2 : get_key(5),
7690 2 : Lsn(0x20),
7691 2 : Value::Image(Bytes::from("value 5@0x20")),
7692 2 : ),
7693 2 : (
7694 2 : get_key(6),
7695 2 : Lsn(0x20),
7696 2 : Value::Image(Bytes::from("value 6@0x20")),
7697 2 : ),
7698 2 : ];
7699 2 : let delta3 = vec![
7700 2 : (
7701 2 : get_key(8),
7702 2 : Lsn(0x48),
7703 2 : Value::Image(Bytes::from("value 8@0x48")),
7704 2 : ),
7705 2 : (
7706 2 : get_key(9),
7707 2 : Lsn(0x48),
7708 2 : Value::Image(Bytes::from("value 9@0x48")),
7709 2 : ),
7710 2 : ];
7711 2 :
7712 2 : let tline = tenant
7713 2 : .create_test_timeline_with_layers(
7714 2 : TIMELINE_ID,
7715 2 : Lsn(0x10),
7716 2 : DEFAULT_PG_VERSION,
7717 2 : &ctx,
7718 2 : vec![
7719 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
7720 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
7721 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
7722 2 : ], // delta layers
7723 2 : vec![(Lsn(0x10), img_layer)], // image layers
7724 2 : Lsn(0x50),
7725 2 : )
7726 49 : .await?;
7727 2 : {
7728 2 : // Update GC info
7729 2 : let mut guard = tline.gc_info.write().unwrap();
7730 2 : guard.cutoffs.time = Lsn(0x30);
7731 2 : guard.cutoffs.space = Lsn(0x30);
7732 2 : }
7733 2 :
7734 2 : let expected_result = [
7735 2 : Bytes::from_static(b"value 0@0x10"),
7736 2 : Bytes::from_static(b"value 1@0x20"),
7737 2 : Bytes::from_static(b"value 2@0x30"),
7738 2 : Bytes::from_static(b"value 3@0x40"),
7739 2 : Bytes::from_static(b"value 4@0x10"),
7740 2 : Bytes::from_static(b"value 5@0x20"),
7741 2 : Bytes::from_static(b"value 6@0x20"),
7742 2 : Bytes::from_static(b"value 7@0x10"),
7743 2 : Bytes::from_static(b"value 8@0x48"),
7744 2 : Bytes::from_static(b"value 9@0x48"),
7745 2 : ];
7746 2 :
7747 20 : for (idx, expected) in expected_result.iter().enumerate() {
7748 20 : assert_eq!(
7749 20 : tline
7750 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7751 30 : .await
7752 20 : .unwrap(),
7753 2 : expected
7754 2 : );
7755 2 : }
7756 2 :
7757 2 : let cancel = CancellationToken::new();
7758 2 : tline
7759 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
7760 56 : .await
7761 2 : .unwrap();
7762 2 :
7763 20 : for (idx, expected) in expected_result.iter().enumerate() {
7764 20 : assert_eq!(
7765 20 : tline
7766 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7767 20 : .await
7768 20 : .unwrap(),
7769 2 : expected
7770 2 : );
7771 2 : }
7772 2 :
7773 2 : // Check if the image layer at the GC horizon contains exactly what we want
7774 2 : let image_at_gc_horizon = tline
7775 2 : .inspect_image_layers(Lsn(0x30), &ctx)
7776 2 : .await
7777 2 : .unwrap()
7778 2 : .into_iter()
7779 34 : .filter(|(k, _)| k.is_metadata_key())
7780 2 : .collect::<Vec<_>>();
7781 2 :
7782 2 : assert_eq!(image_at_gc_horizon.len(), 10);
7783 2 : let expected_result = [
7784 2 : Bytes::from_static(b"value 0@0x10"),
7785 2 : Bytes::from_static(b"value 1@0x20"),
7786 2 : Bytes::from_static(b"value 2@0x30"),
7787 2 : Bytes::from_static(b"value 3@0x10"),
7788 2 : Bytes::from_static(b"value 4@0x10"),
7789 2 : Bytes::from_static(b"value 5@0x20"),
7790 2 : Bytes::from_static(b"value 6@0x20"),
7791 2 : Bytes::from_static(b"value 7@0x10"),
7792 2 : Bytes::from_static(b"value 8@0x10"),
7793 2 : Bytes::from_static(b"value 9@0x10"),
7794 2 : ];
7795 22 : for idx in 0..10 {
7796 20 : assert_eq!(
7797 20 : image_at_gc_horizon[idx],
7798 20 : (get_key(idx as u32), expected_result[idx].clone())
7799 20 : );
7800 2 : }
7801 2 :
7802 2 : // Check if old layers are removed / new layers have the expected LSN
7803 2 : let all_layers = inspect_and_sort(&tline, None).await;
7804 2 : assert_eq!(
7805 2 : all_layers,
7806 2 : vec![
7807 2 : // Image layer at GC horizon
7808 2 : PersistentLayerKey {
7809 2 : key_range: Key::MIN..Key::MAX,
7810 2 : lsn_range: Lsn(0x30)..Lsn(0x31),
7811 2 : is_delta: false
7812 2 : },
7813 2 : // The delta layer below the horizon
7814 2 : PersistentLayerKey {
7815 2 : key_range: get_key(3)..get_key(4),
7816 2 : lsn_range: Lsn(0x30)..Lsn(0x48),
7817 2 : is_delta: true
7818 2 : },
7819 2 : // The delta3 layer that should not be picked for the compaction
7820 2 : PersistentLayerKey {
7821 2 : key_range: get_key(8)..get_key(10),
7822 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
7823 2 : is_delta: true
7824 2 : }
7825 2 : ]
7826 2 : );
7827 2 :
7828 2 : // increase GC horizon and compact again
7829 2 : {
7830 2 : // Update GC info
7831 2 : let mut guard = tline.gc_info.write().unwrap();
7832 2 : guard.cutoffs.time = Lsn(0x40);
7833 2 : guard.cutoffs.space = Lsn(0x40);
7834 2 : }
7835 2 : tline
7836 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
7837 45 : .await
7838 2 : .unwrap();
7839 2 :
7840 2 : Ok(())
7841 2 : }
7842 :
7843 : #[cfg(feature = "testing")]
7844 : #[tokio::test]
7845 2 : async fn test_neon_test_record() -> anyhow::Result<()> {
7846 2 : let harness = TenantHarness::create("test_neon_test_record").await?;
7847 19 : let (tenant, ctx) = harness.load().await;
7848 2 :
7849 24 : fn get_key(id: u32) -> Key {
7850 24 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7851 24 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7852 24 : key.field6 = id;
7853 24 : key
7854 24 : }
7855 2 :
7856 2 : let delta1 = vec![
7857 2 : (
7858 2 : get_key(1),
7859 2 : Lsn(0x20),
7860 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
7861 2 : ),
7862 2 : (
7863 2 : get_key(1),
7864 2 : Lsn(0x30),
7865 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
7866 2 : ),
7867 2 : (get_key(2), Lsn(0x10), Value::Image("0x10".into())),
7868 2 : (
7869 2 : get_key(2),
7870 2 : Lsn(0x20),
7871 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
7872 2 : ),
7873 2 : (
7874 2 : get_key(2),
7875 2 : Lsn(0x30),
7876 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
7877 2 : ),
7878 2 : (get_key(3), Lsn(0x10), Value::Image("0x10".into())),
7879 2 : (
7880 2 : get_key(3),
7881 2 : Lsn(0x20),
7882 2 : Value::WalRecord(NeonWalRecord::wal_clear("c")),
7883 2 : ),
7884 2 : (get_key(4), Lsn(0x10), Value::Image("0x10".into())),
7885 2 : (
7886 2 : get_key(4),
7887 2 : Lsn(0x20),
7888 2 : Value::WalRecord(NeonWalRecord::wal_init("i")),
7889 2 : ),
7890 2 : ];
7891 2 : let image1 = vec![(get_key(1), "0x10".into())];
7892 2 :
7893 2 : let tline = tenant
7894 2 : .create_test_timeline_with_layers(
7895 2 : TIMELINE_ID,
7896 2 : Lsn(0x10),
7897 2 : DEFAULT_PG_VERSION,
7898 2 : &ctx,
7899 2 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
7900 2 : Lsn(0x10)..Lsn(0x40),
7901 2 : delta1,
7902 2 : )], // delta layers
7903 2 : vec![(Lsn(0x10), image1)], // image layers
7904 2 : Lsn(0x50),
7905 2 : )
7906 19 : .await?;
7907 2 :
7908 2 : assert_eq!(
7909 8 : tline.get(get_key(1), Lsn(0x50), &ctx).await?,
7910 2 : Bytes::from_static(b"0x10,0x20,0x30")
7911 2 : );
7912 2 : assert_eq!(
7913 2 : tline.get(get_key(2), Lsn(0x50), &ctx).await?,
7914 2 : Bytes::from_static(b"0x10,0x20,0x30")
7915 2 : );
7916 2 :
7917 2 : // Need to remove the limit of "Neon WAL redo requires base image".
7918 2 :
7919 2 : // assert_eq!(tline.get(get_key(3), Lsn(0x50), &ctx).await?, Bytes::new());
7920 2 : // assert_eq!(tline.get(get_key(4), Lsn(0x50), &ctx).await?, Bytes::new());
7921 2 :
7922 2 : Ok(())
7923 2 : }
7924 :
7925 : #[tokio::test(start_paused = true)]
7926 2 : async fn test_lsn_lease() -> anyhow::Result<()> {
7927 2 : let (tenant, ctx) = TenantHarness::create("test_lsn_lease")
7928 2 : .await
7929 2 : .unwrap()
7930 2 : .load()
7931 20 : .await;
7932 2 : // Advance to the lsn lease deadline so that GC is not blocked by
7933 2 : // initial transition into AttachedSingle.
7934 2 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
7935 2 : tokio::time::resume();
7936 2 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7937 2 :
7938 2 : let end_lsn = Lsn(0x100);
7939 2 : let image_layers = (0x20..=0x90)
7940 2 : .step_by(0x10)
7941 16 : .map(|n| {
7942 16 : (
7943 16 : Lsn(n),
7944 16 : vec![(key, test_img(&format!("data key at {:x}", n)))],
7945 16 : )
7946 16 : })
7947 2 : .collect();
7948 2 :
7949 2 : let timeline = tenant
7950 2 : .create_test_timeline_with_layers(
7951 2 : TIMELINE_ID,
7952 2 : Lsn(0x10),
7953 2 : DEFAULT_PG_VERSION,
7954 2 : &ctx,
7955 2 : Vec::new(),
7956 2 : image_layers,
7957 2 : end_lsn,
7958 2 : )
7959 62 : .await?;
7960 2 :
7961 2 : let leased_lsns = [0x30, 0x50, 0x70];
7962 2 : let mut leases = Vec::new();
7963 6 : leased_lsns.iter().for_each(|n| {
7964 6 : leases.push(
7965 6 : timeline
7966 6 : .init_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)
7967 6 : .expect("lease request should succeed"),
7968 6 : );
7969 6 : });
7970 2 :
7971 2 : let updated_lease_0 = timeline
7972 2 : .renew_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)
7973 2 : .expect("lease renewal should succeed");
7974 2 : assert_eq!(
7975 2 : updated_lease_0.valid_until, leases[0].valid_until,
7976 2 : " Renewing with shorter lease should not change the lease."
7977 2 : );
7978 2 :
7979 2 : let updated_lease_1 = timeline
7980 2 : .renew_lsn_lease(
7981 2 : Lsn(leased_lsns[1]),
7982 2 : timeline.get_lsn_lease_length() * 2,
7983 2 : &ctx,
7984 2 : )
7985 2 : .expect("lease renewal should succeed");
7986 2 : assert!(
7987 2 : updated_lease_1.valid_until > leases[1].valid_until,
7988 2 : "Renewing with a long lease should renew lease with later expiration time."
7989 2 : );
7990 2 :
7991 2 : // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
7992 2 : info!(
7993 2 : "latest_gc_cutoff_lsn: {}",
7994 0 : *timeline.get_latest_gc_cutoff_lsn()
7995 2 : );
7996 2 : timeline.force_set_disk_consistent_lsn(end_lsn);
7997 2 :
7998 2 : let res = tenant
7999 2 : .gc_iteration(
8000 2 : Some(TIMELINE_ID),
8001 2 : 0,
8002 2 : Duration::ZERO,
8003 2 : &CancellationToken::new(),
8004 2 : &ctx,
8005 2 : )
8006 2 : .await
8007 2 : .unwrap();
8008 2 :
8009 2 : // Keeping everything <= Lsn(0x80) b/c leases:
8010 2 : // 0/10: initdb layer
8011 2 : // (0/20..=0/70).step_by(0x10): image layers added when creating the timeline.
8012 2 : assert_eq!(res.layers_needed_by_leases, 7);
8013 2 : // Keeping 0/90 b/c it is the latest layer.
8014 2 : assert_eq!(res.layers_not_updated, 1);
8015 2 : // Removed 0/80.
8016 2 : assert_eq!(res.layers_removed, 1);
8017 2 :
8018 2 : // Make lease on a already GC-ed LSN.
8019 2 : // 0/80 does not have a valid lease + is below latest_gc_cutoff
8020 2 : assert!(Lsn(0x80) < *timeline.get_latest_gc_cutoff_lsn());
8021 2 : timeline
8022 2 : .init_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx)
8023 2 : .expect_err("lease request on GC-ed LSN should fail");
8024 2 :
8025 2 : // Should still be able to renew a currently valid lease
8026 2 : // Assumption: original lease to is still valid for 0/50.
8027 2 : // (use `Timeline::init_lsn_lease` for testing so it always does validation)
8028 2 : timeline
8029 2 : .init_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)
8030 2 : .expect("lease renewal with validation should succeed");
8031 2 :
8032 2 : Ok(())
8033 2 : }
8034 :
8035 : #[cfg(feature = "testing")]
8036 : #[tokio::test]
8037 2 : async fn test_simple_bottom_most_compaction_deltas_1() -> anyhow::Result<()> {
8038 2 : test_simple_bottom_most_compaction_deltas_helper(
8039 2 : "test_simple_bottom_most_compaction_deltas_1",
8040 2 : false,
8041 2 : )
8042 248 : .await
8043 2 : }
8044 :
8045 : #[cfg(feature = "testing")]
8046 : #[tokio::test]
8047 2 : async fn test_simple_bottom_most_compaction_deltas_2() -> anyhow::Result<()> {
8048 2 : test_simple_bottom_most_compaction_deltas_helper(
8049 2 : "test_simple_bottom_most_compaction_deltas_2",
8050 2 : true,
8051 2 : )
8052 226 : .await
8053 2 : }
8054 :
8055 : #[cfg(feature = "testing")]
8056 4 : async fn test_simple_bottom_most_compaction_deltas_helper(
8057 4 : test_name: &'static str,
8058 4 : use_delta_bottom_layer: bool,
8059 4 : ) -> anyhow::Result<()> {
8060 4 : let harness = TenantHarness::create(test_name).await?;
8061 40 : let (tenant, ctx) = harness.load().await;
8062 :
8063 276 : fn get_key(id: u32) -> Key {
8064 276 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8065 276 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8066 276 : key.field6 = id;
8067 276 : key
8068 276 : }
8069 :
8070 : // We create
8071 : // - one bottom-most image layer,
8072 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
8073 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
8074 : // - a delta layer D3 above the horizon.
8075 : //
8076 : // | D3 |
8077 : // | D1 |
8078 : // -| |-- gc horizon -----------------
8079 : // | | | D2 |
8080 : // --------- img layer ------------------
8081 : //
8082 : // What we should expact from this compaction is:
8083 : // | D3 |
8084 : // | Part of D1 |
8085 : // --------- img layer with D1+D2 at GC horizon------------------
8086 :
8087 : // img layer at 0x10
8088 4 : let img_layer = (0..10)
8089 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8090 4 : .collect_vec();
8091 4 : // or, delta layer at 0x10 if `use_delta_bottom_layer` is true
8092 4 : let delta4 = (0..10)
8093 40 : .map(|id| {
8094 40 : (
8095 40 : get_key(id),
8096 40 : Lsn(0x08),
8097 40 : Value::WalRecord(NeonWalRecord::wal_init(format!("value {id}@0x10"))),
8098 40 : )
8099 40 : })
8100 4 : .collect_vec();
8101 4 :
8102 4 : let delta1 = vec![
8103 4 : (
8104 4 : get_key(1),
8105 4 : Lsn(0x20),
8106 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8107 4 : ),
8108 4 : (
8109 4 : get_key(2),
8110 4 : Lsn(0x30),
8111 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8112 4 : ),
8113 4 : (
8114 4 : get_key(3),
8115 4 : Lsn(0x28),
8116 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8117 4 : ),
8118 4 : (
8119 4 : get_key(3),
8120 4 : Lsn(0x30),
8121 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8122 4 : ),
8123 4 : (
8124 4 : get_key(3),
8125 4 : Lsn(0x40),
8126 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
8127 4 : ),
8128 4 : ];
8129 4 : let delta2 = vec![
8130 4 : (
8131 4 : get_key(5),
8132 4 : Lsn(0x20),
8133 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8134 4 : ),
8135 4 : (
8136 4 : get_key(6),
8137 4 : Lsn(0x20),
8138 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8139 4 : ),
8140 4 : ];
8141 4 : let delta3 = vec![
8142 4 : (
8143 4 : get_key(8),
8144 4 : Lsn(0x48),
8145 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8146 4 : ),
8147 4 : (
8148 4 : get_key(9),
8149 4 : Lsn(0x48),
8150 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8151 4 : ),
8152 4 : ];
8153 :
8154 4 : let tline = if use_delta_bottom_layer {
8155 2 : tenant
8156 2 : .create_test_timeline_with_layers(
8157 2 : TIMELINE_ID,
8158 2 : Lsn(0x08),
8159 2 : DEFAULT_PG_VERSION,
8160 2 : &ctx,
8161 2 : vec![
8162 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8163 2 : Lsn(0x08)..Lsn(0x10),
8164 2 : delta4,
8165 2 : ),
8166 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8167 2 : Lsn(0x20)..Lsn(0x48),
8168 2 : delta1,
8169 2 : ),
8170 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8171 2 : Lsn(0x20)..Lsn(0x48),
8172 2 : delta2,
8173 2 : ),
8174 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8175 2 : Lsn(0x48)..Lsn(0x50),
8176 2 : delta3,
8177 2 : ),
8178 2 : ], // delta layers
8179 2 : vec![], // image layers
8180 2 : Lsn(0x50),
8181 2 : )
8182 30 : .await?
8183 : } else {
8184 2 : tenant
8185 2 : .create_test_timeline_with_layers(
8186 2 : TIMELINE_ID,
8187 2 : Lsn(0x10),
8188 2 : DEFAULT_PG_VERSION,
8189 2 : &ctx,
8190 2 : vec![
8191 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8192 2 : Lsn(0x10)..Lsn(0x48),
8193 2 : delta1,
8194 2 : ),
8195 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8196 2 : Lsn(0x10)..Lsn(0x48),
8197 2 : delta2,
8198 2 : ),
8199 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8200 2 : Lsn(0x48)..Lsn(0x50),
8201 2 : delta3,
8202 2 : ),
8203 2 : ], // delta layers
8204 2 : vec![(Lsn(0x10), img_layer)], // image layers
8205 2 : Lsn(0x50),
8206 2 : )
8207 49 : .await?
8208 : };
8209 4 : {
8210 4 : // Update GC info
8211 4 : let mut guard = tline.gc_info.write().unwrap();
8212 4 : *guard = GcInfo {
8213 4 : retain_lsns: vec![],
8214 4 : cutoffs: GcCutoffs {
8215 4 : time: Lsn(0x30),
8216 4 : space: Lsn(0x30),
8217 4 : },
8218 4 : leases: Default::default(),
8219 4 : within_ancestor_pitr: false,
8220 4 : };
8221 4 : }
8222 4 :
8223 4 : let expected_result = [
8224 4 : Bytes::from_static(b"value 0@0x10"),
8225 4 : Bytes::from_static(b"value 1@0x10@0x20"),
8226 4 : Bytes::from_static(b"value 2@0x10@0x30"),
8227 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
8228 4 : Bytes::from_static(b"value 4@0x10"),
8229 4 : Bytes::from_static(b"value 5@0x10@0x20"),
8230 4 : Bytes::from_static(b"value 6@0x10@0x20"),
8231 4 : Bytes::from_static(b"value 7@0x10"),
8232 4 : Bytes::from_static(b"value 8@0x10@0x48"),
8233 4 : Bytes::from_static(b"value 9@0x10@0x48"),
8234 4 : ];
8235 4 :
8236 4 : let expected_result_at_gc_horizon = [
8237 4 : Bytes::from_static(b"value 0@0x10"),
8238 4 : Bytes::from_static(b"value 1@0x10@0x20"),
8239 4 : Bytes::from_static(b"value 2@0x10@0x30"),
8240 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
8241 4 : Bytes::from_static(b"value 4@0x10"),
8242 4 : Bytes::from_static(b"value 5@0x10@0x20"),
8243 4 : Bytes::from_static(b"value 6@0x10@0x20"),
8244 4 : Bytes::from_static(b"value 7@0x10"),
8245 4 : Bytes::from_static(b"value 8@0x10"),
8246 4 : Bytes::from_static(b"value 9@0x10"),
8247 4 : ];
8248 :
8249 44 : for idx in 0..10 {
8250 40 : assert_eq!(
8251 40 : tline
8252 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8253 60 : .await
8254 40 : .unwrap(),
8255 40 : &expected_result[idx]
8256 : );
8257 40 : assert_eq!(
8258 40 : tline
8259 40 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
8260 31 : .await
8261 40 : .unwrap(),
8262 40 : &expected_result_at_gc_horizon[idx]
8263 : );
8264 : }
8265 :
8266 4 : let cancel = CancellationToken::new();
8267 4 : tline
8268 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8269 114 : .await
8270 4 : .unwrap();
8271 :
8272 44 : for idx in 0..10 {
8273 40 : assert_eq!(
8274 40 : tline
8275 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8276 40 : .await
8277 40 : .unwrap(),
8278 40 : &expected_result[idx]
8279 : );
8280 40 : assert_eq!(
8281 40 : tline
8282 40 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
8283 20 : .await
8284 40 : .unwrap(),
8285 40 : &expected_result_at_gc_horizon[idx]
8286 : );
8287 : }
8288 :
8289 : // increase GC horizon and compact again
8290 4 : {
8291 4 : // Update GC info
8292 4 : let mut guard = tline.gc_info.write().unwrap();
8293 4 : guard.cutoffs.time = Lsn(0x40);
8294 4 : guard.cutoffs.space = Lsn(0x40);
8295 4 : }
8296 4 : tline
8297 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8298 90 : .await
8299 4 : .unwrap();
8300 4 :
8301 4 : Ok(())
8302 4 : }
8303 :
8304 : #[cfg(feature = "testing")]
8305 : #[tokio::test]
8306 2 : async fn test_generate_key_retention() -> anyhow::Result<()> {
8307 2 : let harness = TenantHarness::create("test_generate_key_retention").await?;
8308 19 : let (tenant, ctx) = harness.load().await;
8309 2 : let tline = tenant
8310 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
8311 6 : .await?;
8312 2 : tline.force_advance_lsn(Lsn(0x70));
8313 2 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
8314 2 : let history = vec![
8315 2 : (
8316 2 : key,
8317 2 : Lsn(0x10),
8318 2 : Value::WalRecord(NeonWalRecord::wal_init("0x10")),
8319 2 : ),
8320 2 : (
8321 2 : key,
8322 2 : Lsn(0x20),
8323 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8324 2 : ),
8325 2 : (
8326 2 : key,
8327 2 : Lsn(0x30),
8328 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8329 2 : ),
8330 2 : (
8331 2 : key,
8332 2 : Lsn(0x40),
8333 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8334 2 : ),
8335 2 : (
8336 2 : key,
8337 2 : Lsn(0x50),
8338 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
8339 2 : ),
8340 2 : (
8341 2 : key,
8342 2 : Lsn(0x60),
8343 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8344 2 : ),
8345 2 : (
8346 2 : key,
8347 2 : Lsn(0x70),
8348 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8349 2 : ),
8350 2 : (
8351 2 : key,
8352 2 : Lsn(0x80),
8353 2 : Value::Image(Bytes::copy_from_slice(
8354 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8355 2 : )),
8356 2 : ),
8357 2 : (
8358 2 : key,
8359 2 : Lsn(0x90),
8360 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8361 2 : ),
8362 2 : ];
8363 2 : let res = tline
8364 2 : .generate_key_retention(
8365 2 : key,
8366 2 : &history,
8367 2 : Lsn(0x60),
8368 2 : &[Lsn(0x20), Lsn(0x40), Lsn(0x50)],
8369 2 : 3,
8370 2 : None,
8371 2 : )
8372 2 : .await
8373 2 : .unwrap();
8374 2 : let expected_res = KeyHistoryRetention {
8375 2 : below_horizon: vec![
8376 2 : (
8377 2 : Lsn(0x20),
8378 2 : KeyLogAtLsn(vec![(
8379 2 : Lsn(0x20),
8380 2 : Value::Image(Bytes::from_static(b"0x10;0x20")),
8381 2 : )]),
8382 2 : ),
8383 2 : (
8384 2 : Lsn(0x40),
8385 2 : KeyLogAtLsn(vec![
8386 2 : (
8387 2 : Lsn(0x30),
8388 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8389 2 : ),
8390 2 : (
8391 2 : Lsn(0x40),
8392 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8393 2 : ),
8394 2 : ]),
8395 2 : ),
8396 2 : (
8397 2 : Lsn(0x50),
8398 2 : KeyLogAtLsn(vec![(
8399 2 : Lsn(0x50),
8400 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40;0x50")),
8401 2 : )]),
8402 2 : ),
8403 2 : (
8404 2 : Lsn(0x60),
8405 2 : KeyLogAtLsn(vec![(
8406 2 : Lsn(0x60),
8407 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8408 2 : )]),
8409 2 : ),
8410 2 : ],
8411 2 : above_horizon: KeyLogAtLsn(vec![
8412 2 : (
8413 2 : Lsn(0x70),
8414 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8415 2 : ),
8416 2 : (
8417 2 : Lsn(0x80),
8418 2 : Value::Image(Bytes::copy_from_slice(
8419 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8420 2 : )),
8421 2 : ),
8422 2 : (
8423 2 : Lsn(0x90),
8424 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8425 2 : ),
8426 2 : ]),
8427 2 : };
8428 2 : assert_eq!(res, expected_res);
8429 2 :
8430 2 : // We expect GC-compaction to run with the original GC. This would create a situation that
8431 2 : // the original GC algorithm removes some delta layers b/c there are full image coverage,
8432 2 : // therefore causing some keys to have an incomplete history below the lowest retain LSN.
8433 2 : // For example, we have
8434 2 : // ```plain
8435 2 : // init delta @ 0x10, image @ 0x20, delta @ 0x30 (gc_horizon), image @ 0x40.
8436 2 : // ```
8437 2 : // Now the GC horizon moves up, and we have
8438 2 : // ```plain
8439 2 : // init delta @ 0x10, image @ 0x20, delta @ 0x30, image @ 0x40 (gc_horizon)
8440 2 : // ```
8441 2 : // The original GC algorithm kicks in, and removes delta @ 0x10, image @ 0x20.
8442 2 : // We will end up with
8443 2 : // ```plain
8444 2 : // delta @ 0x30, image @ 0x40 (gc_horizon)
8445 2 : // ```
8446 2 : // Now we run the GC-compaction, and this key does not have a full history.
8447 2 : // We should be able to handle this partial history and drop everything before the
8448 2 : // gc_horizon image.
8449 2 :
8450 2 : let history = vec![
8451 2 : (
8452 2 : key,
8453 2 : Lsn(0x20),
8454 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8455 2 : ),
8456 2 : (
8457 2 : key,
8458 2 : Lsn(0x30),
8459 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8460 2 : ),
8461 2 : (
8462 2 : key,
8463 2 : Lsn(0x40),
8464 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
8465 2 : ),
8466 2 : (
8467 2 : key,
8468 2 : Lsn(0x50),
8469 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
8470 2 : ),
8471 2 : (
8472 2 : key,
8473 2 : Lsn(0x60),
8474 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8475 2 : ),
8476 2 : (
8477 2 : key,
8478 2 : Lsn(0x70),
8479 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8480 2 : ),
8481 2 : (
8482 2 : key,
8483 2 : Lsn(0x80),
8484 2 : Value::Image(Bytes::copy_from_slice(
8485 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8486 2 : )),
8487 2 : ),
8488 2 : (
8489 2 : key,
8490 2 : Lsn(0x90),
8491 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8492 2 : ),
8493 2 : ];
8494 2 : let res = tline
8495 2 : .generate_key_retention(key, &history, Lsn(0x60), &[Lsn(0x40), Lsn(0x50)], 3, None)
8496 2 : .await
8497 2 : .unwrap();
8498 2 : let expected_res = KeyHistoryRetention {
8499 2 : below_horizon: vec![
8500 2 : (
8501 2 : Lsn(0x40),
8502 2 : KeyLogAtLsn(vec![(
8503 2 : Lsn(0x40),
8504 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
8505 2 : )]),
8506 2 : ),
8507 2 : (
8508 2 : Lsn(0x50),
8509 2 : KeyLogAtLsn(vec![(
8510 2 : Lsn(0x50),
8511 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
8512 2 : )]),
8513 2 : ),
8514 2 : (
8515 2 : Lsn(0x60),
8516 2 : KeyLogAtLsn(vec![(
8517 2 : Lsn(0x60),
8518 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8519 2 : )]),
8520 2 : ),
8521 2 : ],
8522 2 : above_horizon: KeyLogAtLsn(vec![
8523 2 : (
8524 2 : Lsn(0x70),
8525 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8526 2 : ),
8527 2 : (
8528 2 : Lsn(0x80),
8529 2 : Value::Image(Bytes::copy_from_slice(
8530 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8531 2 : )),
8532 2 : ),
8533 2 : (
8534 2 : Lsn(0x90),
8535 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8536 2 : ),
8537 2 : ]),
8538 2 : };
8539 2 : assert_eq!(res, expected_res);
8540 2 :
8541 2 : // In case of branch compaction, the branch itself does not have the full history, and we need to provide
8542 2 : // the ancestor image in the test case.
8543 2 :
8544 2 : let history = vec![
8545 2 : (
8546 2 : key,
8547 2 : Lsn(0x20),
8548 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8549 2 : ),
8550 2 : (
8551 2 : key,
8552 2 : Lsn(0x30),
8553 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8554 2 : ),
8555 2 : (
8556 2 : key,
8557 2 : Lsn(0x40),
8558 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8559 2 : ),
8560 2 : (
8561 2 : key,
8562 2 : Lsn(0x70),
8563 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8564 2 : ),
8565 2 : ];
8566 2 : let res = tline
8567 2 : .generate_key_retention(
8568 2 : key,
8569 2 : &history,
8570 2 : Lsn(0x60),
8571 2 : &[],
8572 2 : 3,
8573 2 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
8574 2 : )
8575 2 : .await
8576 2 : .unwrap();
8577 2 : let expected_res = KeyHistoryRetention {
8578 2 : below_horizon: vec![(
8579 2 : Lsn(0x60),
8580 2 : KeyLogAtLsn(vec![(
8581 2 : Lsn(0x60),
8582 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")), // use the ancestor image to reconstruct the page
8583 2 : )]),
8584 2 : )],
8585 2 : above_horizon: KeyLogAtLsn(vec![(
8586 2 : Lsn(0x70),
8587 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8588 2 : )]),
8589 2 : };
8590 2 : assert_eq!(res, expected_res);
8591 2 :
8592 2 : let history = vec![
8593 2 : (
8594 2 : key,
8595 2 : Lsn(0x20),
8596 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8597 2 : ),
8598 2 : (
8599 2 : key,
8600 2 : Lsn(0x40),
8601 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8602 2 : ),
8603 2 : (
8604 2 : key,
8605 2 : Lsn(0x60),
8606 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8607 2 : ),
8608 2 : (
8609 2 : key,
8610 2 : Lsn(0x70),
8611 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8612 2 : ),
8613 2 : ];
8614 2 : let res = tline
8615 2 : .generate_key_retention(
8616 2 : key,
8617 2 : &history,
8618 2 : Lsn(0x60),
8619 2 : &[Lsn(0x30)],
8620 2 : 3,
8621 2 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
8622 2 : )
8623 2 : .await
8624 2 : .unwrap();
8625 2 : let expected_res = KeyHistoryRetention {
8626 2 : below_horizon: vec![
8627 2 : (
8628 2 : Lsn(0x30),
8629 2 : KeyLogAtLsn(vec![(
8630 2 : Lsn(0x20),
8631 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8632 2 : )]),
8633 2 : ),
8634 2 : (
8635 2 : Lsn(0x60),
8636 2 : KeyLogAtLsn(vec![(
8637 2 : Lsn(0x60),
8638 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x40;0x60")),
8639 2 : )]),
8640 2 : ),
8641 2 : ],
8642 2 : above_horizon: KeyLogAtLsn(vec![(
8643 2 : Lsn(0x70),
8644 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8645 2 : )]),
8646 2 : };
8647 2 : assert_eq!(res, expected_res);
8648 2 :
8649 2 : Ok(())
8650 2 : }
8651 :
8652 : #[cfg(feature = "testing")]
8653 : #[tokio::test]
8654 2 : async fn test_simple_bottom_most_compaction_with_retain_lsns() -> anyhow::Result<()> {
8655 2 : let harness =
8656 2 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns").await?;
8657 20 : let (tenant, ctx) = harness.load().await;
8658 2 :
8659 518 : fn get_key(id: u32) -> Key {
8660 518 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8661 518 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8662 518 : key.field6 = id;
8663 518 : key
8664 518 : }
8665 2 :
8666 2 : let img_layer = (0..10)
8667 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8668 2 : .collect_vec();
8669 2 :
8670 2 : let delta1 = vec![
8671 2 : (
8672 2 : get_key(1),
8673 2 : Lsn(0x20),
8674 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8675 2 : ),
8676 2 : (
8677 2 : get_key(2),
8678 2 : Lsn(0x30),
8679 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8680 2 : ),
8681 2 : (
8682 2 : get_key(3),
8683 2 : Lsn(0x28),
8684 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8685 2 : ),
8686 2 : (
8687 2 : get_key(3),
8688 2 : Lsn(0x30),
8689 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8690 2 : ),
8691 2 : (
8692 2 : get_key(3),
8693 2 : Lsn(0x40),
8694 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
8695 2 : ),
8696 2 : ];
8697 2 : let delta2 = vec![
8698 2 : (
8699 2 : get_key(5),
8700 2 : Lsn(0x20),
8701 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8702 2 : ),
8703 2 : (
8704 2 : get_key(6),
8705 2 : Lsn(0x20),
8706 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8707 2 : ),
8708 2 : ];
8709 2 : let delta3 = vec![
8710 2 : (
8711 2 : get_key(8),
8712 2 : Lsn(0x48),
8713 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8714 2 : ),
8715 2 : (
8716 2 : get_key(9),
8717 2 : Lsn(0x48),
8718 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8719 2 : ),
8720 2 : ];
8721 2 :
8722 2 : let tline = tenant
8723 2 : .create_test_timeline_with_layers(
8724 2 : TIMELINE_ID,
8725 2 : Lsn(0x10),
8726 2 : DEFAULT_PG_VERSION,
8727 2 : &ctx,
8728 2 : vec![
8729 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
8730 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
8731 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
8732 2 : ], // delta layers
8733 2 : vec![(Lsn(0x10), img_layer)], // image layers
8734 2 : Lsn(0x50),
8735 2 : )
8736 49 : .await?;
8737 2 : {
8738 2 : // Update GC info
8739 2 : let mut guard = tline.gc_info.write().unwrap();
8740 2 : *guard = GcInfo {
8741 2 : retain_lsns: vec![
8742 2 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
8743 2 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
8744 2 : ],
8745 2 : cutoffs: GcCutoffs {
8746 2 : time: Lsn(0x30),
8747 2 : space: Lsn(0x30),
8748 2 : },
8749 2 : leases: Default::default(),
8750 2 : within_ancestor_pitr: false,
8751 2 : };
8752 2 : }
8753 2 :
8754 2 : let expected_result = [
8755 2 : Bytes::from_static(b"value 0@0x10"),
8756 2 : Bytes::from_static(b"value 1@0x10@0x20"),
8757 2 : Bytes::from_static(b"value 2@0x10@0x30"),
8758 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
8759 2 : Bytes::from_static(b"value 4@0x10"),
8760 2 : Bytes::from_static(b"value 5@0x10@0x20"),
8761 2 : Bytes::from_static(b"value 6@0x10@0x20"),
8762 2 : Bytes::from_static(b"value 7@0x10"),
8763 2 : Bytes::from_static(b"value 8@0x10@0x48"),
8764 2 : Bytes::from_static(b"value 9@0x10@0x48"),
8765 2 : ];
8766 2 :
8767 2 : let expected_result_at_gc_horizon = [
8768 2 : Bytes::from_static(b"value 0@0x10"),
8769 2 : Bytes::from_static(b"value 1@0x10@0x20"),
8770 2 : Bytes::from_static(b"value 2@0x10@0x30"),
8771 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
8772 2 : Bytes::from_static(b"value 4@0x10"),
8773 2 : Bytes::from_static(b"value 5@0x10@0x20"),
8774 2 : Bytes::from_static(b"value 6@0x10@0x20"),
8775 2 : Bytes::from_static(b"value 7@0x10"),
8776 2 : Bytes::from_static(b"value 8@0x10"),
8777 2 : Bytes::from_static(b"value 9@0x10"),
8778 2 : ];
8779 2 :
8780 2 : let expected_result_at_lsn_20 = [
8781 2 : Bytes::from_static(b"value 0@0x10"),
8782 2 : Bytes::from_static(b"value 1@0x10@0x20"),
8783 2 : Bytes::from_static(b"value 2@0x10"),
8784 2 : Bytes::from_static(b"value 3@0x10"),
8785 2 : Bytes::from_static(b"value 4@0x10"),
8786 2 : Bytes::from_static(b"value 5@0x10@0x20"),
8787 2 : Bytes::from_static(b"value 6@0x10@0x20"),
8788 2 : Bytes::from_static(b"value 7@0x10"),
8789 2 : Bytes::from_static(b"value 8@0x10"),
8790 2 : Bytes::from_static(b"value 9@0x10"),
8791 2 : ];
8792 2 :
8793 2 : let expected_result_at_lsn_10 = [
8794 2 : Bytes::from_static(b"value 0@0x10"),
8795 2 : Bytes::from_static(b"value 1@0x10"),
8796 2 : Bytes::from_static(b"value 2@0x10"),
8797 2 : Bytes::from_static(b"value 3@0x10"),
8798 2 : Bytes::from_static(b"value 4@0x10"),
8799 2 : Bytes::from_static(b"value 5@0x10"),
8800 2 : Bytes::from_static(b"value 6@0x10"),
8801 2 : Bytes::from_static(b"value 7@0x10"),
8802 2 : Bytes::from_static(b"value 8@0x10"),
8803 2 : Bytes::from_static(b"value 9@0x10"),
8804 2 : ];
8805 2 :
8806 12 : let verify_result = || async {
8807 12 : let gc_horizon = {
8808 12 : let gc_info = tline.gc_info.read().unwrap();
8809 12 : gc_info.cutoffs.time
8810 2 : };
8811 132 : for idx in 0..10 {
8812 120 : assert_eq!(
8813 120 : tline
8814 120 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8815 121 : .await
8816 120 : .unwrap(),
8817 120 : &expected_result[idx]
8818 2 : );
8819 120 : assert_eq!(
8820 120 : tline
8821 120 : .get(get_key(idx as u32), gc_horizon, &ctx)
8822 93 : .await
8823 120 : .unwrap(),
8824 120 : &expected_result_at_gc_horizon[idx]
8825 2 : );
8826 120 : assert_eq!(
8827 120 : tline
8828 120 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
8829 86 : .await
8830 120 : .unwrap(),
8831 120 : &expected_result_at_lsn_20[idx]
8832 2 : );
8833 120 : assert_eq!(
8834 120 : tline
8835 120 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
8836 61 : .await
8837 120 : .unwrap(),
8838 120 : &expected_result_at_lsn_10[idx]
8839 2 : );
8840 2 : }
8841 24 : };
8842 2 :
8843 69 : verify_result().await;
8844 2 :
8845 2 : let cancel = CancellationToken::new();
8846 2 : let mut dryrun_flags = EnumSet::new();
8847 2 : dryrun_flags.insert(CompactFlags::DryRun);
8848 2 :
8849 2 : tline
8850 2 : .compact_with_gc(
8851 2 : &cancel,
8852 2 : CompactOptions {
8853 2 : flags: dryrun_flags,
8854 2 : compact_range: None,
8855 2 : },
8856 2 : &ctx,
8857 2 : )
8858 48 : .await
8859 2 : .unwrap();
8860 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
8861 2 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
8862 57 : verify_result().await;
8863 2 :
8864 2 : tline
8865 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8866 53 : .await
8867 2 : .unwrap();
8868 64 : verify_result().await;
8869 2 :
8870 2 : // compact again
8871 2 : tline
8872 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8873 40 : .await
8874 2 : .unwrap();
8875 57 : verify_result().await;
8876 2 :
8877 2 : // increase GC horizon and compact again
8878 2 : {
8879 2 : // Update GC info
8880 2 : let mut guard = tline.gc_info.write().unwrap();
8881 2 : guard.cutoffs.time = Lsn(0x38);
8882 2 : guard.cutoffs.space = Lsn(0x38);
8883 2 : }
8884 2 : tline
8885 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8886 40 : .await
8887 2 : .unwrap();
8888 57 : verify_result().await; // no wals between 0x30 and 0x38, so we should obtain the same result
8889 2 :
8890 2 : // not increasing the GC horizon and compact again
8891 2 : tline
8892 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8893 40 : .await
8894 2 : .unwrap();
8895 57 : verify_result().await;
8896 2 :
8897 2 : Ok(())
8898 2 : }
8899 :
8900 : #[cfg(feature = "testing")]
8901 : #[tokio::test]
8902 2 : async fn test_simple_bottom_most_compaction_with_retain_lsns_single_key() -> anyhow::Result<()>
8903 2 : {
8904 2 : let harness =
8905 2 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns_single_key")
8906 2 : .await?;
8907 20 : let (tenant, ctx) = harness.load().await;
8908 2 :
8909 352 : fn get_key(id: u32) -> Key {
8910 352 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8911 352 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8912 352 : key.field6 = id;
8913 352 : key
8914 352 : }
8915 2 :
8916 2 : let img_layer = (0..10)
8917 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8918 2 : .collect_vec();
8919 2 :
8920 2 : let delta1 = vec![
8921 2 : (
8922 2 : get_key(1),
8923 2 : Lsn(0x20),
8924 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8925 2 : ),
8926 2 : (
8927 2 : get_key(1),
8928 2 : Lsn(0x28),
8929 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8930 2 : ),
8931 2 : ];
8932 2 : let delta2 = vec![
8933 2 : (
8934 2 : get_key(1),
8935 2 : Lsn(0x30),
8936 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8937 2 : ),
8938 2 : (
8939 2 : get_key(1),
8940 2 : Lsn(0x38),
8941 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
8942 2 : ),
8943 2 : ];
8944 2 : let delta3 = vec![
8945 2 : (
8946 2 : get_key(8),
8947 2 : Lsn(0x48),
8948 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8949 2 : ),
8950 2 : (
8951 2 : get_key(9),
8952 2 : Lsn(0x48),
8953 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8954 2 : ),
8955 2 : ];
8956 2 :
8957 2 : let tline = tenant
8958 2 : .create_test_timeline_with_layers(
8959 2 : TIMELINE_ID,
8960 2 : Lsn(0x10),
8961 2 : DEFAULT_PG_VERSION,
8962 2 : &ctx,
8963 2 : vec![
8964 2 : // delta1 and delta 2 only contain a single key but multiple updates
8965 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x30), delta1),
8966 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
8967 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x50), delta3),
8968 2 : ], // delta layers
8969 2 : vec![(Lsn(0x10), img_layer)], // image layers
8970 2 : Lsn(0x50),
8971 2 : )
8972 49 : .await?;
8973 2 : {
8974 2 : // Update GC info
8975 2 : let mut guard = tline.gc_info.write().unwrap();
8976 2 : *guard = GcInfo {
8977 2 : retain_lsns: vec![
8978 2 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
8979 2 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
8980 2 : ],
8981 2 : cutoffs: GcCutoffs {
8982 2 : time: Lsn(0x30),
8983 2 : space: Lsn(0x30),
8984 2 : },
8985 2 : leases: Default::default(),
8986 2 : within_ancestor_pitr: false,
8987 2 : };
8988 2 : }
8989 2 :
8990 2 : let expected_result = [
8991 2 : Bytes::from_static(b"value 0@0x10"),
8992 2 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
8993 2 : Bytes::from_static(b"value 2@0x10"),
8994 2 : Bytes::from_static(b"value 3@0x10"),
8995 2 : Bytes::from_static(b"value 4@0x10"),
8996 2 : Bytes::from_static(b"value 5@0x10"),
8997 2 : Bytes::from_static(b"value 6@0x10"),
8998 2 : Bytes::from_static(b"value 7@0x10"),
8999 2 : Bytes::from_static(b"value 8@0x10@0x48"),
9000 2 : Bytes::from_static(b"value 9@0x10@0x48"),
9001 2 : ];
9002 2 :
9003 2 : let expected_result_at_gc_horizon = [
9004 2 : Bytes::from_static(b"value 0@0x10"),
9005 2 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
9006 2 : Bytes::from_static(b"value 2@0x10"),
9007 2 : Bytes::from_static(b"value 3@0x10"),
9008 2 : Bytes::from_static(b"value 4@0x10"),
9009 2 : Bytes::from_static(b"value 5@0x10"),
9010 2 : Bytes::from_static(b"value 6@0x10"),
9011 2 : Bytes::from_static(b"value 7@0x10"),
9012 2 : Bytes::from_static(b"value 8@0x10"),
9013 2 : Bytes::from_static(b"value 9@0x10"),
9014 2 : ];
9015 2 :
9016 2 : let expected_result_at_lsn_20 = [
9017 2 : Bytes::from_static(b"value 0@0x10"),
9018 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9019 2 : Bytes::from_static(b"value 2@0x10"),
9020 2 : Bytes::from_static(b"value 3@0x10"),
9021 2 : Bytes::from_static(b"value 4@0x10"),
9022 2 : Bytes::from_static(b"value 5@0x10"),
9023 2 : Bytes::from_static(b"value 6@0x10"),
9024 2 : Bytes::from_static(b"value 7@0x10"),
9025 2 : Bytes::from_static(b"value 8@0x10"),
9026 2 : Bytes::from_static(b"value 9@0x10"),
9027 2 : ];
9028 2 :
9029 2 : let expected_result_at_lsn_10 = [
9030 2 : Bytes::from_static(b"value 0@0x10"),
9031 2 : Bytes::from_static(b"value 1@0x10"),
9032 2 : Bytes::from_static(b"value 2@0x10"),
9033 2 : Bytes::from_static(b"value 3@0x10"),
9034 2 : Bytes::from_static(b"value 4@0x10"),
9035 2 : Bytes::from_static(b"value 5@0x10"),
9036 2 : Bytes::from_static(b"value 6@0x10"),
9037 2 : Bytes::from_static(b"value 7@0x10"),
9038 2 : Bytes::from_static(b"value 8@0x10"),
9039 2 : Bytes::from_static(b"value 9@0x10"),
9040 2 : ];
9041 2 :
9042 8 : let verify_result = || async {
9043 8 : let gc_horizon = {
9044 8 : let gc_info = tline.gc_info.read().unwrap();
9045 8 : gc_info.cutoffs.time
9046 2 : };
9047 88 : for idx in 0..10 {
9048 80 : assert_eq!(
9049 80 : tline
9050 80 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9051 76 : .await
9052 80 : .unwrap(),
9053 80 : &expected_result[idx]
9054 2 : );
9055 80 : assert_eq!(
9056 80 : tline
9057 80 : .get(get_key(idx as u32), gc_horizon, &ctx)
9058 46 : .await
9059 80 : .unwrap(),
9060 80 : &expected_result_at_gc_horizon[idx]
9061 2 : );
9062 80 : assert_eq!(
9063 80 : tline
9064 80 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
9065 48 : .await
9066 80 : .unwrap(),
9067 80 : &expected_result_at_lsn_20[idx]
9068 2 : );
9069 80 : assert_eq!(
9070 80 : tline
9071 80 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
9072 41 : .await
9073 80 : .unwrap(),
9074 80 : &expected_result_at_lsn_10[idx]
9075 2 : );
9076 2 : }
9077 16 : };
9078 2 :
9079 61 : verify_result().await;
9080 2 :
9081 2 : let cancel = CancellationToken::new();
9082 2 : let mut dryrun_flags = EnumSet::new();
9083 2 : dryrun_flags.insert(CompactFlags::DryRun);
9084 2 :
9085 2 : tline
9086 2 : .compact_with_gc(
9087 2 : &cancel,
9088 2 : CompactOptions {
9089 2 : flags: dryrun_flags,
9090 2 : compact_range: None,
9091 2 : },
9092 2 : &ctx,
9093 2 : )
9094 54 : .await
9095 2 : .unwrap();
9096 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
9097 2 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
9098 49 : verify_result().await;
9099 2 :
9100 2 : tline
9101 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9102 60 : .await
9103 2 : .unwrap();
9104 54 : verify_result().await;
9105 2 :
9106 2 : // compact again
9107 2 : tline
9108 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9109 40 : .await
9110 2 : .unwrap();
9111 47 : verify_result().await;
9112 2 :
9113 2 : Ok(())
9114 2 : }
9115 :
9116 : #[cfg(feature = "testing")]
9117 : #[tokio::test]
9118 2 : async fn test_simple_bottom_most_compaction_on_branch() -> anyhow::Result<()> {
9119 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_on_branch").await?;
9120 17 : let (tenant, ctx) = harness.load().await;
9121 2 :
9122 126 : fn get_key(id: u32) -> Key {
9123 126 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
9124 126 : key.field6 = id;
9125 126 : key
9126 126 : }
9127 2 :
9128 2 : let img_layer = (0..10)
9129 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9130 2 : .collect_vec();
9131 2 :
9132 2 : let delta1 = vec![
9133 2 : (
9134 2 : get_key(1),
9135 2 : Lsn(0x20),
9136 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9137 2 : ),
9138 2 : (
9139 2 : get_key(2),
9140 2 : Lsn(0x30),
9141 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9142 2 : ),
9143 2 : (
9144 2 : get_key(3),
9145 2 : Lsn(0x28),
9146 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9147 2 : ),
9148 2 : (
9149 2 : get_key(3),
9150 2 : Lsn(0x30),
9151 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9152 2 : ),
9153 2 : (
9154 2 : get_key(3),
9155 2 : Lsn(0x40),
9156 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9157 2 : ),
9158 2 : ];
9159 2 : let delta2 = vec![
9160 2 : (
9161 2 : get_key(5),
9162 2 : Lsn(0x20),
9163 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9164 2 : ),
9165 2 : (
9166 2 : get_key(6),
9167 2 : Lsn(0x20),
9168 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9169 2 : ),
9170 2 : ];
9171 2 : let delta3 = vec![
9172 2 : (
9173 2 : get_key(8),
9174 2 : Lsn(0x48),
9175 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9176 2 : ),
9177 2 : (
9178 2 : get_key(9),
9179 2 : Lsn(0x48),
9180 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9181 2 : ),
9182 2 : ];
9183 2 :
9184 2 : let parent_tline = tenant
9185 2 : .create_test_timeline_with_layers(
9186 2 : TIMELINE_ID,
9187 2 : Lsn(0x10),
9188 2 : DEFAULT_PG_VERSION,
9189 2 : &ctx,
9190 2 : vec![], // delta layers
9191 2 : vec![(Lsn(0x18), img_layer)], // image layers
9192 2 : Lsn(0x18),
9193 2 : )
9194 30 : .await?;
9195 2 :
9196 2 : parent_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
9197 2 :
9198 2 : let branch_tline = tenant
9199 2 : .branch_timeline_test_with_layers(
9200 2 : &parent_tline,
9201 2 : NEW_TIMELINE_ID,
9202 2 : Some(Lsn(0x18)),
9203 2 : &ctx,
9204 2 : vec![
9205 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
9206 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
9207 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
9208 2 : ], // delta layers
9209 2 : vec![], // image layers
9210 2 : Lsn(0x50),
9211 2 : )
9212 20 : .await?;
9213 2 :
9214 2 : branch_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
9215 2 :
9216 2 : {
9217 2 : // Update GC info
9218 2 : let mut guard = parent_tline.gc_info.write().unwrap();
9219 2 : *guard = GcInfo {
9220 2 : retain_lsns: vec![(Lsn(0x18), branch_tline.timeline_id, MaybeOffloaded::No)],
9221 2 : cutoffs: GcCutoffs {
9222 2 : time: Lsn(0x10),
9223 2 : space: Lsn(0x10),
9224 2 : },
9225 2 : leases: Default::default(),
9226 2 : within_ancestor_pitr: false,
9227 2 : };
9228 2 : }
9229 2 :
9230 2 : {
9231 2 : // Update GC info
9232 2 : let mut guard = branch_tline.gc_info.write().unwrap();
9233 2 : *guard = GcInfo {
9234 2 : retain_lsns: vec![(Lsn(0x40), branch_tline.timeline_id, MaybeOffloaded::No)],
9235 2 : cutoffs: GcCutoffs {
9236 2 : time: Lsn(0x50),
9237 2 : space: Lsn(0x50),
9238 2 : },
9239 2 : leases: Default::default(),
9240 2 : within_ancestor_pitr: false,
9241 2 : };
9242 2 : }
9243 2 :
9244 2 : let expected_result_at_gc_horizon = [
9245 2 : Bytes::from_static(b"value 0@0x10"),
9246 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9247 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9248 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9249 2 : Bytes::from_static(b"value 4@0x10"),
9250 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9251 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9252 2 : Bytes::from_static(b"value 7@0x10"),
9253 2 : Bytes::from_static(b"value 8@0x10@0x48"),
9254 2 : Bytes::from_static(b"value 9@0x10@0x48"),
9255 2 : ];
9256 2 :
9257 2 : let expected_result_at_lsn_40 = [
9258 2 : Bytes::from_static(b"value 0@0x10"),
9259 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9260 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9261 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9262 2 : Bytes::from_static(b"value 4@0x10"),
9263 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9264 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9265 2 : Bytes::from_static(b"value 7@0x10"),
9266 2 : Bytes::from_static(b"value 8@0x10"),
9267 2 : Bytes::from_static(b"value 9@0x10"),
9268 2 : ];
9269 2 :
9270 4 : let verify_result = || async {
9271 44 : for idx in 0..10 {
9272 40 : assert_eq!(
9273 40 : branch_tline
9274 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9275 52 : .await
9276 40 : .unwrap(),
9277 40 : &expected_result_at_gc_horizon[idx]
9278 2 : );
9279 40 : assert_eq!(
9280 40 : branch_tline
9281 40 : .get(get_key(idx as u32), Lsn(0x40), &ctx)
9282 30 : .await
9283 40 : .unwrap(),
9284 40 : &expected_result_at_lsn_40[idx]
9285 2 : );
9286 2 : }
9287 8 : };
9288 2 :
9289 45 : verify_result().await;
9290 2 :
9291 2 : let cancel = CancellationToken::new();
9292 2 : branch_tline
9293 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9294 22 : .await
9295 2 : .unwrap();
9296 2 :
9297 37 : verify_result().await;
9298 2 :
9299 2 : Ok(())
9300 2 : }
9301 :
9302 : // Regression test for https://github.com/neondatabase/neon/issues/9012
9303 : // Create an image arrangement where we have to read at different LSN ranges
9304 : // from a delta layer. This is achieved by overlapping an image layer on top of
9305 : // a delta layer. Like so:
9306 : //
9307 : // A B
9308 : // +----------------+ -> delta_layer
9309 : // | | ^ lsn
9310 : // | =========|-> nested_image_layer |
9311 : // | C | |
9312 : // +----------------+ |
9313 : // ======== -> baseline_image_layer +-------> key
9314 : //
9315 : //
9316 : // When querying the key range [A, B) we need to read at different LSN ranges
9317 : // for [A, C) and [C, B). This test checks that the described edge case is handled correctly.
9318 : #[cfg(feature = "testing")]
9319 : #[tokio::test]
9320 2 : async fn test_vectored_read_with_nested_image_layer() -> anyhow::Result<()> {
9321 2 : let harness = TenantHarness::create("test_vectored_read_with_nested_image_layer").await?;
9322 20 : let (tenant, ctx) = harness.load().await;
9323 2 :
9324 2 : let will_init_keys = [2, 6];
9325 44 : fn get_key(id: u32) -> Key {
9326 44 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
9327 44 : key.field6 = id;
9328 44 : key
9329 44 : }
9330 2 :
9331 2 : let mut expected_key_values = HashMap::new();
9332 2 :
9333 2 : let baseline_image_layer_lsn = Lsn(0x10);
9334 2 : let mut baseline_img_layer = Vec::new();
9335 12 : for i in 0..5 {
9336 10 : let key = get_key(i);
9337 10 : let value = format!("value {i}@{baseline_image_layer_lsn}");
9338 10 :
9339 10 : let removed = expected_key_values.insert(key, value.clone());
9340 10 : assert!(removed.is_none());
9341 2 :
9342 10 : baseline_img_layer.push((key, Bytes::from(value)));
9343 2 : }
9344 2 :
9345 2 : let nested_image_layer_lsn = Lsn(0x50);
9346 2 : let mut nested_img_layer = Vec::new();
9347 12 : for i in 5..10 {
9348 10 : let key = get_key(i);
9349 10 : let value = format!("value {i}@{nested_image_layer_lsn}");
9350 10 :
9351 10 : let removed = expected_key_values.insert(key, value.clone());
9352 10 : assert!(removed.is_none());
9353 2 :
9354 10 : nested_img_layer.push((key, Bytes::from(value)));
9355 2 : }
9356 2 :
9357 2 : let mut delta_layer_spec = Vec::default();
9358 2 : let delta_layer_start_lsn = Lsn(0x20);
9359 2 : let mut delta_layer_end_lsn = delta_layer_start_lsn;
9360 2 :
9361 22 : for i in 0..10 {
9362 20 : let key = get_key(i);
9363 20 : let key_in_nested = nested_img_layer
9364 20 : .iter()
9365 80 : .any(|(key_with_img, _)| *key_with_img == key);
9366 20 : let lsn = {
9367 20 : if key_in_nested {
9368 10 : Lsn(nested_image_layer_lsn.0 + 0x10)
9369 2 : } else {
9370 10 : delta_layer_start_lsn
9371 2 : }
9372 2 : };
9373 2 :
9374 20 : let will_init = will_init_keys.contains(&i);
9375 20 : if will_init {
9376 4 : delta_layer_spec.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init(""))));
9377 4 :
9378 4 : expected_key_values.insert(key, "".to_string());
9379 16 : } else {
9380 16 : let delta = format!("@{lsn}");
9381 16 : delta_layer_spec.push((
9382 16 : key,
9383 16 : lsn,
9384 16 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
9385 16 : ));
9386 16 :
9387 16 : expected_key_values
9388 16 : .get_mut(&key)
9389 16 : .expect("An image exists for each key")
9390 16 : .push_str(delta.as_str());
9391 16 : }
9392 20 : delta_layer_end_lsn = std::cmp::max(delta_layer_start_lsn, lsn);
9393 2 : }
9394 2 :
9395 2 : delta_layer_end_lsn = Lsn(delta_layer_end_lsn.0 + 1);
9396 2 :
9397 2 : assert!(
9398 2 : nested_image_layer_lsn > delta_layer_start_lsn
9399 2 : && nested_image_layer_lsn < delta_layer_end_lsn
9400 2 : );
9401 2 :
9402 2 : let tline = tenant
9403 2 : .create_test_timeline_with_layers(
9404 2 : TIMELINE_ID,
9405 2 : baseline_image_layer_lsn,
9406 2 : DEFAULT_PG_VERSION,
9407 2 : &ctx,
9408 2 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
9409 2 : delta_layer_start_lsn..delta_layer_end_lsn,
9410 2 : delta_layer_spec,
9411 2 : )], // delta layers
9412 2 : vec![
9413 2 : (baseline_image_layer_lsn, baseline_img_layer),
9414 2 : (nested_image_layer_lsn, nested_img_layer),
9415 2 : ], // image layers
9416 2 : delta_layer_end_lsn,
9417 2 : )
9418 42 : .await?;
9419 2 :
9420 2 : let keyspace = KeySpace::single(get_key(0)..get_key(10));
9421 2 : let results = tline
9422 2 : .get_vectored(keyspace, delta_layer_end_lsn, &ctx)
9423 13 : .await
9424 2 : .expect("No vectored errors");
9425 22 : for (key, res) in results {
9426 20 : let value = res.expect("No key errors");
9427 20 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
9428 20 : assert_eq!(value, Bytes::from(expected_value));
9429 2 : }
9430 2 :
9431 2 : Ok(())
9432 2 : }
9433 :
9434 142 : fn sort_layer_key(k1: &PersistentLayerKey, k2: &PersistentLayerKey) -> std::cmp::Ordering {
9435 142 : (
9436 142 : k1.is_delta,
9437 142 : k1.key_range.start,
9438 142 : k1.key_range.end,
9439 142 : k1.lsn_range.start,
9440 142 : k1.lsn_range.end,
9441 142 : )
9442 142 : .cmp(&(
9443 142 : k2.is_delta,
9444 142 : k2.key_range.start,
9445 142 : k2.key_range.end,
9446 142 : k2.lsn_range.start,
9447 142 : k2.lsn_range.end,
9448 142 : ))
9449 142 : }
9450 :
9451 12 : async fn inspect_and_sort(
9452 12 : tline: &Arc<Timeline>,
9453 12 : filter: Option<std::ops::Range<Key>>,
9454 12 : ) -> Vec<PersistentLayerKey> {
9455 12 : let mut all_layers = tline.inspect_historic_layers().await.unwrap();
9456 12 : if let Some(filter) = filter {
9457 64 : all_layers.retain(|layer| overlaps_with(&layer.key_range, &filter));
9458 10 : }
9459 12 : all_layers.sort_by(sort_layer_key);
9460 12 : all_layers
9461 12 : }
9462 :
9463 : #[cfg(feature = "testing")]
9464 10 : fn check_layer_map_key_eq(
9465 10 : mut left: Vec<PersistentLayerKey>,
9466 10 : mut right: Vec<PersistentLayerKey>,
9467 10 : ) {
9468 10 : left.sort_by(sort_layer_key);
9469 10 : right.sort_by(sort_layer_key);
9470 10 : if left != right {
9471 0 : eprintln!("---LEFT---");
9472 0 : for left in left.iter() {
9473 0 : eprintln!("{}", left);
9474 0 : }
9475 0 : eprintln!("---RIGHT---");
9476 0 : for right in right.iter() {
9477 0 : eprintln!("{}", right);
9478 0 : }
9479 0 : assert_eq!(left, right);
9480 10 : }
9481 10 : }
9482 :
9483 : #[cfg(feature = "testing")]
9484 : #[tokio::test]
9485 2 : async fn test_simple_partial_bottom_most_compaction() -> anyhow::Result<()> {
9486 2 : let harness = TenantHarness::create("test_simple_partial_bottom_most_compaction").await?;
9487 20 : let (tenant, ctx) = harness.load().await;
9488 2 :
9489 182 : fn get_key(id: u32) -> Key {
9490 182 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9491 182 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9492 182 : key.field6 = id;
9493 182 : key
9494 182 : }
9495 2 :
9496 2 : // img layer at 0x10
9497 2 : let img_layer = (0..10)
9498 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9499 2 : .collect_vec();
9500 2 :
9501 2 : let delta1 = vec![
9502 2 : (
9503 2 : get_key(1),
9504 2 : Lsn(0x20),
9505 2 : Value::Image(Bytes::from("value 1@0x20")),
9506 2 : ),
9507 2 : (
9508 2 : get_key(2),
9509 2 : Lsn(0x30),
9510 2 : Value::Image(Bytes::from("value 2@0x30")),
9511 2 : ),
9512 2 : (
9513 2 : get_key(3),
9514 2 : Lsn(0x40),
9515 2 : Value::Image(Bytes::from("value 3@0x40")),
9516 2 : ),
9517 2 : ];
9518 2 : let delta2 = vec![
9519 2 : (
9520 2 : get_key(5),
9521 2 : Lsn(0x20),
9522 2 : Value::Image(Bytes::from("value 5@0x20")),
9523 2 : ),
9524 2 : (
9525 2 : get_key(6),
9526 2 : Lsn(0x20),
9527 2 : Value::Image(Bytes::from("value 6@0x20")),
9528 2 : ),
9529 2 : ];
9530 2 : let delta3 = vec![
9531 2 : (
9532 2 : get_key(8),
9533 2 : Lsn(0x48),
9534 2 : Value::Image(Bytes::from("value 8@0x48")),
9535 2 : ),
9536 2 : (
9537 2 : get_key(9),
9538 2 : Lsn(0x48),
9539 2 : Value::Image(Bytes::from("value 9@0x48")),
9540 2 : ),
9541 2 : ];
9542 2 :
9543 2 : let tline = tenant
9544 2 : .create_test_timeline_with_layers(
9545 2 : TIMELINE_ID,
9546 2 : Lsn(0x10),
9547 2 : DEFAULT_PG_VERSION,
9548 2 : &ctx,
9549 2 : vec![
9550 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
9551 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
9552 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
9553 2 : ], // delta layers
9554 2 : vec![(Lsn(0x10), img_layer)], // image layers
9555 2 : Lsn(0x50),
9556 2 : )
9557 49 : .await?;
9558 2 :
9559 2 : {
9560 2 : // Update GC info
9561 2 : let mut guard = tline.gc_info.write().unwrap();
9562 2 : *guard = GcInfo {
9563 2 : retain_lsns: vec![(Lsn(0x20), tline.timeline_id, MaybeOffloaded::No)],
9564 2 : cutoffs: GcCutoffs {
9565 2 : time: Lsn(0x30),
9566 2 : space: Lsn(0x30),
9567 2 : },
9568 2 : leases: Default::default(),
9569 2 : within_ancestor_pitr: false,
9570 2 : };
9571 2 : }
9572 2 :
9573 2 : let cancel = CancellationToken::new();
9574 2 :
9575 2 : // Do a partial compaction on key range 0..2
9576 2 : tline
9577 2 : .partial_compact_with_gc(get_key(0)..get_key(2), &cancel, EnumSet::new(), &ctx)
9578 27 : .await
9579 2 : .unwrap();
9580 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
9581 2 : check_layer_map_key_eq(
9582 2 : all_layers,
9583 2 : vec![
9584 2 : // newly-generated image layer for the partial compaction range 0-2
9585 2 : PersistentLayerKey {
9586 2 : key_range: get_key(0)..get_key(2),
9587 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9588 2 : is_delta: false,
9589 2 : },
9590 2 : PersistentLayerKey {
9591 2 : key_range: get_key(0)..get_key(10),
9592 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
9593 2 : is_delta: false,
9594 2 : },
9595 2 : // delta1 is split and the second part is rewritten
9596 2 : PersistentLayerKey {
9597 2 : key_range: get_key(2)..get_key(4),
9598 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9599 2 : is_delta: true,
9600 2 : },
9601 2 : PersistentLayerKey {
9602 2 : key_range: get_key(5)..get_key(7),
9603 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9604 2 : is_delta: true,
9605 2 : },
9606 2 : PersistentLayerKey {
9607 2 : key_range: get_key(8)..get_key(10),
9608 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
9609 2 : is_delta: true,
9610 2 : },
9611 2 : ],
9612 2 : );
9613 2 :
9614 2 : // Do a partial compaction on key range 2..4
9615 2 : tline
9616 2 : .partial_compact_with_gc(get_key(2)..get_key(4), &cancel, EnumSet::new(), &ctx)
9617 18 : .await
9618 2 : .unwrap();
9619 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
9620 2 : check_layer_map_key_eq(
9621 2 : all_layers,
9622 2 : vec![
9623 2 : PersistentLayerKey {
9624 2 : key_range: get_key(0)..get_key(2),
9625 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9626 2 : is_delta: false,
9627 2 : },
9628 2 : PersistentLayerKey {
9629 2 : key_range: get_key(0)..get_key(10),
9630 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
9631 2 : is_delta: false,
9632 2 : },
9633 2 : // image layer generated for the compaction range 2-4
9634 2 : PersistentLayerKey {
9635 2 : key_range: get_key(2)..get_key(4),
9636 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9637 2 : is_delta: false,
9638 2 : },
9639 2 : // we have key2/key3 above the retain_lsn, so we still need this delta layer
9640 2 : PersistentLayerKey {
9641 2 : key_range: get_key(2)..get_key(4),
9642 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9643 2 : is_delta: true,
9644 2 : },
9645 2 : PersistentLayerKey {
9646 2 : key_range: get_key(5)..get_key(7),
9647 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9648 2 : is_delta: true,
9649 2 : },
9650 2 : PersistentLayerKey {
9651 2 : key_range: get_key(8)..get_key(10),
9652 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
9653 2 : is_delta: true,
9654 2 : },
9655 2 : ],
9656 2 : );
9657 2 :
9658 2 : // Do a partial compaction on key range 4..9
9659 2 : tline
9660 2 : .partial_compact_with_gc(get_key(4)..get_key(9), &cancel, EnumSet::new(), &ctx)
9661 24 : .await
9662 2 : .unwrap();
9663 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
9664 2 : check_layer_map_key_eq(
9665 2 : all_layers,
9666 2 : vec![
9667 2 : PersistentLayerKey {
9668 2 : key_range: get_key(0)..get_key(2),
9669 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9670 2 : is_delta: false,
9671 2 : },
9672 2 : PersistentLayerKey {
9673 2 : key_range: get_key(0)..get_key(10),
9674 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
9675 2 : is_delta: false,
9676 2 : },
9677 2 : PersistentLayerKey {
9678 2 : key_range: get_key(2)..get_key(4),
9679 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9680 2 : is_delta: false,
9681 2 : },
9682 2 : PersistentLayerKey {
9683 2 : key_range: get_key(2)..get_key(4),
9684 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9685 2 : is_delta: true,
9686 2 : },
9687 2 : // image layer generated for this compaction range
9688 2 : PersistentLayerKey {
9689 2 : key_range: get_key(4)..get_key(9),
9690 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9691 2 : is_delta: false,
9692 2 : },
9693 2 : PersistentLayerKey {
9694 2 : key_range: get_key(8)..get_key(10),
9695 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
9696 2 : is_delta: true,
9697 2 : },
9698 2 : ],
9699 2 : );
9700 2 :
9701 2 : // Do a partial compaction on key range 9..10
9702 2 : tline
9703 2 : .partial_compact_with_gc(get_key(9)..get_key(10), &cancel, EnumSet::new(), &ctx)
9704 10 : .await
9705 2 : .unwrap();
9706 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
9707 2 : check_layer_map_key_eq(
9708 2 : all_layers,
9709 2 : vec![
9710 2 : PersistentLayerKey {
9711 2 : key_range: get_key(0)..get_key(2),
9712 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9713 2 : is_delta: false,
9714 2 : },
9715 2 : PersistentLayerKey {
9716 2 : key_range: get_key(0)..get_key(10),
9717 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
9718 2 : is_delta: false,
9719 2 : },
9720 2 : PersistentLayerKey {
9721 2 : key_range: get_key(2)..get_key(4),
9722 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9723 2 : is_delta: false,
9724 2 : },
9725 2 : PersistentLayerKey {
9726 2 : key_range: get_key(2)..get_key(4),
9727 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9728 2 : is_delta: true,
9729 2 : },
9730 2 : PersistentLayerKey {
9731 2 : key_range: get_key(4)..get_key(9),
9732 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9733 2 : is_delta: false,
9734 2 : },
9735 2 : // image layer generated for the compaction range
9736 2 : PersistentLayerKey {
9737 2 : key_range: get_key(9)..get_key(10),
9738 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9739 2 : is_delta: false,
9740 2 : },
9741 2 : PersistentLayerKey {
9742 2 : key_range: get_key(8)..get_key(10),
9743 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
9744 2 : is_delta: true,
9745 2 : },
9746 2 : ],
9747 2 : );
9748 2 :
9749 2 : // Do a partial compaction on key range 0..10, all image layers below LSN 20 can be replaced with new ones.
9750 2 : tline
9751 2 : .partial_compact_with_gc(get_key(0)..get_key(10), &cancel, EnumSet::new(), &ctx)
9752 52 : .await
9753 2 : .unwrap();
9754 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
9755 2 : check_layer_map_key_eq(
9756 2 : all_layers,
9757 2 : vec![
9758 2 : // aha, we removed all unnecessary image/delta layers and got a very clean layer map!
9759 2 : PersistentLayerKey {
9760 2 : key_range: get_key(0)..get_key(10),
9761 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9762 2 : is_delta: false,
9763 2 : },
9764 2 : PersistentLayerKey {
9765 2 : key_range: get_key(2)..get_key(4),
9766 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9767 2 : is_delta: true,
9768 2 : },
9769 2 : PersistentLayerKey {
9770 2 : key_range: get_key(8)..get_key(10),
9771 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
9772 2 : is_delta: true,
9773 2 : },
9774 2 : ],
9775 2 : );
9776 2 :
9777 2 : Ok(())
9778 2 : }
9779 :
9780 : #[cfg(feature = "testing")]
9781 : #[tokio::test]
9782 2 : async fn test_timeline_offload_retain_lsn() -> anyhow::Result<()> {
9783 2 : let harness = TenantHarness::create("test_timeline_offload_retain_lsn")
9784 2 : .await
9785 2 : .unwrap();
9786 20 : let (tenant, ctx) = harness.load().await;
9787 2 : let tline_parent = tenant
9788 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
9789 6 : .await
9790 2 : .unwrap();
9791 2 : let tline_child = tenant
9792 2 : .branch_timeline_test(&tline_parent, NEW_TIMELINE_ID, Some(Lsn(0x20)), &ctx)
9793 2 : .await
9794 2 : .unwrap();
9795 2 : {
9796 2 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
9797 2 : assert_eq!(
9798 2 : gc_info_parent.retain_lsns,
9799 2 : vec![(Lsn(0x20), tline_child.timeline_id, MaybeOffloaded::No)]
9800 2 : );
9801 2 : }
9802 2 : // We have to directly call the remote_client instead of using the archive function to avoid constructing broker client...
9803 2 : tline_child
9804 2 : .remote_client
9805 2 : .schedule_index_upload_for_timeline_archival_state(TimelineArchivalState::Archived)
9806 2 : .unwrap();
9807 2 : tline_child.remote_client.wait_completion().await.unwrap();
9808 2 : offload_timeline(&tenant, &tline_child)
9809 2 : .instrument(tracing::info_span!(parent: None, "offload_test", tenant_id=%"test", shard_id=%"test", timeline_id=%"test"))
9810 16 : .await.unwrap();
9811 2 : let child_timeline_id = tline_child.timeline_id;
9812 2 : Arc::try_unwrap(tline_child).unwrap();
9813 2 :
9814 2 : {
9815 2 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
9816 2 : assert_eq!(
9817 2 : gc_info_parent.retain_lsns,
9818 2 : vec![(Lsn(0x20), child_timeline_id, MaybeOffloaded::Yes)]
9819 2 : );
9820 2 : }
9821 2 :
9822 2 : tenant
9823 2 : .get_offloaded_timeline(child_timeline_id)
9824 2 : .unwrap()
9825 2 : .defuse_for_tenant_drop();
9826 2 :
9827 2 : Ok(())
9828 2 : }
9829 : }
|
