Line data Source code
1 : //! Timeline repository implementation that keeps old data in layer files, and
2 : //! the recent changes in ephemeral files.
3 : //!
4 : //! See tenant/*_layer.rs files. The functions here are responsible for locating
5 : //! the correct layer for the get/put call, walking back the timeline branching
6 : //! history as needed.
7 : //!
8 : //! The files are stored in the .neon/tenants/<tenant_id>/timelines/<timeline_id>
9 : //! directory. See docs/pageserver-storage.md for how the files are managed.
10 : //! In addition to the layer files, there is a metadata file in the same
11 : //! directory that contains information about the timeline, in particular its
12 : //! parent timeline, and the last LSN that has been written to disk.
13 : //!
14 :
15 : use anyhow::{bail, Context};
16 : use arc_swap::ArcSwap;
17 : use camino::Utf8Path;
18 : use camino::Utf8PathBuf;
19 : use chrono::NaiveDateTime;
20 : use enumset::EnumSet;
21 : use futures::stream::FuturesUnordered;
22 : use futures::StreamExt;
23 : use pageserver_api::models;
24 : use pageserver_api::models::LsnLease;
25 : use pageserver_api::models::TimelineArchivalState;
26 : use pageserver_api::models::TimelineState;
27 : use pageserver_api::models::TopTenantShardItem;
28 : use pageserver_api::models::WalRedoManagerStatus;
29 : use pageserver_api::shard::ShardIdentity;
30 : use pageserver_api::shard::ShardStripeSize;
31 : use pageserver_api::shard::TenantShardId;
32 : use remote_storage::DownloadError;
33 : use remote_storage::GenericRemoteStorage;
34 : use remote_storage::TimeoutOrCancel;
35 : use remote_timeline_client::manifest::{
36 : OffloadedTimelineManifest, TenantManifest, LATEST_TENANT_MANIFEST_VERSION,
37 : };
38 : use remote_timeline_client::UploadQueueNotReadyError;
39 : use std::collections::BTreeMap;
40 : use std::fmt;
41 : use std::future::Future;
42 : use std::sync::atomic::AtomicBool;
43 : use std::sync::Weak;
44 : use std::time::SystemTime;
45 : use storage_broker::BrokerClientChannel;
46 : use timeline::offload::offload_timeline;
47 : use tokio::io::BufReader;
48 : use tokio::sync::watch;
49 : use tokio::task::JoinSet;
50 : use tokio_util::sync::CancellationToken;
51 : use tracing::*;
52 : use upload_queue::NotInitialized;
53 : use utils::backoff;
54 : use utils::circuit_breaker::CircuitBreaker;
55 : use utils::completion;
56 : use utils::crashsafe::path_with_suffix_extension;
57 : use utils::failpoint_support;
58 : use utils::fs_ext;
59 : use utils::pausable_failpoint;
60 : use utils::sync::gate::Gate;
61 : use utils::sync::gate::GateGuard;
62 : use utils::timeout::timeout_cancellable;
63 : use utils::timeout::TimeoutCancellableError;
64 : use utils::zstd::create_zst_tarball;
65 : use utils::zstd::extract_zst_tarball;
66 :
67 : use self::config::AttachedLocationConfig;
68 : use self::config::AttachmentMode;
69 : use self::config::LocationConf;
70 : use self::config::TenantConf;
71 : use self::metadata::TimelineMetadata;
72 : use self::mgr::GetActiveTenantError;
73 : use self::mgr::GetTenantError;
74 : use self::remote_timeline_client::upload::{upload_index_part, upload_tenant_manifest};
75 : use self::remote_timeline_client::{RemoteTimelineClient, WaitCompletionError};
76 : use self::timeline::uninit::TimelineCreateGuard;
77 : use self::timeline::uninit::TimelineExclusionError;
78 : use self::timeline::uninit::UninitializedTimeline;
79 : use self::timeline::EvictionTaskTenantState;
80 : use self::timeline::GcCutoffs;
81 : use self::timeline::TimelineDeleteProgress;
82 : use self::timeline::TimelineResources;
83 : use self::timeline::WaitLsnError;
84 : use crate::config::PageServerConf;
85 : use crate::context::{DownloadBehavior, RequestContext};
86 : use crate::deletion_queue::DeletionQueueClient;
87 : use crate::deletion_queue::DeletionQueueError;
88 : use crate::import_datadir;
89 : use crate::is_uninit_mark;
90 : use crate::l0_flush::L0FlushGlobalState;
91 : use crate::metrics::TENANT;
92 : use crate::metrics::{
93 : remove_tenant_metrics, BROKEN_TENANTS_SET, CIRCUIT_BREAKERS_BROKEN, CIRCUIT_BREAKERS_UNBROKEN,
94 : TENANT_STATE_METRIC, TENANT_SYNTHETIC_SIZE_METRIC,
95 : };
96 : use crate::task_mgr;
97 : use crate::task_mgr::TaskKind;
98 : use crate::tenant::config::LocationMode;
99 : use crate::tenant::config::TenantConfOpt;
100 : use crate::tenant::gc_result::GcResult;
101 : pub use crate::tenant::remote_timeline_client::index::IndexPart;
102 : use crate::tenant::remote_timeline_client::remote_initdb_archive_path;
103 : use crate::tenant::remote_timeline_client::MaybeDeletedIndexPart;
104 : use crate::tenant::remote_timeline_client::INITDB_PATH;
105 : use crate::tenant::storage_layer::DeltaLayer;
106 : use crate::tenant::storage_layer::ImageLayer;
107 : use crate::walingest::WalLagCooldown;
108 : use crate::walredo;
109 : use crate::InitializationOrder;
110 : use std::collections::hash_map::Entry;
111 : use std::collections::HashMap;
112 : use std::collections::HashSet;
113 : use std::fmt::Debug;
114 : use std::fmt::Display;
115 : use std::fs;
116 : use std::fs::File;
117 : use std::sync::atomic::{AtomicU64, Ordering};
118 : use std::sync::Arc;
119 : use std::sync::Mutex;
120 : use std::time::{Duration, Instant};
121 :
122 : use crate::span;
123 : use crate::tenant::timeline::delete::DeleteTimelineFlow;
124 : use crate::tenant::timeline::uninit::cleanup_timeline_directory;
125 : use crate::virtual_file::VirtualFile;
126 : use crate::walredo::PostgresRedoManager;
127 : use crate::TEMP_FILE_SUFFIX;
128 : use once_cell::sync::Lazy;
129 : pub use pageserver_api::models::TenantState;
130 : use tokio::sync::Semaphore;
131 :
132 0 : static INIT_DB_SEMAPHORE: Lazy<Semaphore> = Lazy::new(|| Semaphore::new(8));
133 : use utils::{
134 : crashsafe,
135 : generation::Generation,
136 : id::TimelineId,
137 : lsn::{Lsn, RecordLsn},
138 : };
139 :
140 : pub mod blob_io;
141 : pub mod block_io;
142 : pub mod vectored_blob_io;
143 :
144 : pub mod disk_btree;
145 : pub(crate) mod ephemeral_file;
146 : pub mod layer_map;
147 :
148 : pub mod metadata;
149 : pub mod remote_timeline_client;
150 : pub mod storage_layer;
151 :
152 : pub mod checks;
153 : pub mod config;
154 : pub mod mgr;
155 : pub mod secondary;
156 : pub mod tasks;
157 : pub mod upload_queue;
158 :
159 : pub(crate) mod timeline;
160 :
161 : pub mod size;
162 :
163 : mod gc_block;
164 : mod gc_result;
165 : pub(crate) mod throttle;
166 :
167 : pub(crate) use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
168 : pub(crate) use timeline::{LogicalSizeCalculationCause, PageReconstructError, Timeline};
169 :
170 : // re-export for use in walreceiver
171 : pub use crate::tenant::timeline::WalReceiverInfo;
172 :
173 : /// The "tenants" part of `tenants/<tenant>/timelines...`
174 : pub const TENANTS_SEGMENT_NAME: &str = "tenants";
175 :
176 : /// Parts of the `.neon/tenants/<tenant_id>/timelines/<timeline_id>` directory prefix.
177 : pub const TIMELINES_SEGMENT_NAME: &str = "timelines";
178 :
179 : /// References to shared objects that are passed into each tenant, such
180 : /// as the shared remote storage client and process initialization state.
181 : #[derive(Clone)]
182 : pub struct TenantSharedResources {
183 : pub broker_client: storage_broker::BrokerClientChannel,
184 : pub remote_storage: GenericRemoteStorage,
185 : pub deletion_queue_client: DeletionQueueClient,
186 : pub l0_flush_global_state: L0FlushGlobalState,
187 : }
188 :
189 : /// A [`Tenant`] is really an _attached_ tenant. The configuration
190 : /// for an attached tenant is a subset of the [`LocationConf`], represented
191 : /// in this struct.
192 : pub(super) struct AttachedTenantConf {
193 : tenant_conf: TenantConfOpt,
194 : location: AttachedLocationConfig,
195 : /// The deadline before which we are blocked from GC so that
196 : /// leases have a chance to be renewed.
197 : lsn_lease_deadline: Option<tokio::time::Instant>,
198 : }
199 :
200 : impl AttachedTenantConf {
201 192 : fn new(tenant_conf: TenantConfOpt, location: AttachedLocationConfig) -> Self {
202 : // Sets a deadline before which we cannot proceed to GC due to lsn lease.
203 : //
204 : // We do this as the leases mapping are not persisted to disk. By delaying GC by lease
205 : // length, we guarantee that all the leases we granted before will have a chance to renew
206 : // when we run GC for the first time after restart / transition from AttachedMulti to AttachedSingle.
207 192 : let lsn_lease_deadline = if location.attach_mode == AttachmentMode::Single {
208 192 : Some(
209 192 : tokio::time::Instant::now()
210 192 : + tenant_conf
211 192 : .lsn_lease_length
212 192 : .unwrap_or(LsnLease::DEFAULT_LENGTH),
213 192 : )
214 : } else {
215 : // We don't use `lsn_lease_deadline` to delay GC in AttachedMulti and AttachedStale
216 : // because we don't do GC in these modes.
217 0 : None
218 : };
219 :
220 192 : Self {
221 192 : tenant_conf,
222 192 : location,
223 192 : lsn_lease_deadline,
224 192 : }
225 192 : }
226 :
227 192 : fn try_from(location_conf: LocationConf) -> anyhow::Result<Self> {
228 192 : match &location_conf.mode {
229 192 : LocationMode::Attached(attach_conf) => {
230 192 : Ok(Self::new(location_conf.tenant_conf, *attach_conf))
231 : }
232 : LocationMode::Secondary(_) => {
233 0 : anyhow::bail!("Attempted to construct AttachedTenantConf from a LocationConf in secondary mode")
234 : }
235 : }
236 192 : }
237 :
238 762 : fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
239 762 : self.lsn_lease_deadline
240 762 : .map(|d| tokio::time::Instant::now() < d)
241 762 : .unwrap_or(false)
242 762 : }
243 : }
244 : struct TimelinePreload {
245 : timeline_id: TimelineId,
246 : client: RemoteTimelineClient,
247 : index_part: Result<MaybeDeletedIndexPart, DownloadError>,
248 : }
249 :
250 : pub(crate) struct TenantPreload {
251 : tenant_manifest: TenantManifest,
252 : timelines: HashMap<TimelineId, TimelinePreload>,
253 : }
254 :
255 : /// When we spawn a tenant, there is a special mode for tenant creation that
256 : /// avoids trying to read anything from remote storage.
257 : pub(crate) enum SpawnMode {
258 : /// Activate as soon as possible
259 : Eager,
260 : /// Lazy activation in the background, with the option to skip the queue if the need comes up
261 : Lazy,
262 : }
263 :
264 : ///
265 : /// Tenant consists of multiple timelines. Keep them in a hash table.
266 : ///
267 : pub struct Tenant {
268 : // Global pageserver config parameters
269 : pub conf: &'static PageServerConf,
270 :
271 : /// The value creation timestamp, used to measure activation delay, see:
272 : /// <https://github.com/neondatabase/neon/issues/4025>
273 : constructed_at: Instant,
274 :
275 : state: watch::Sender<TenantState>,
276 :
277 : // Overridden tenant-specific config parameters.
278 : // We keep TenantConfOpt sturct here to preserve the information
279 : // about parameters that are not set.
280 : // This is necessary to allow global config updates.
281 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
282 :
283 : tenant_shard_id: TenantShardId,
284 :
285 : // The detailed sharding information, beyond the number/count in tenant_shard_id
286 : shard_identity: ShardIdentity,
287 :
288 : /// The remote storage generation, used to protect S3 objects from split-brain.
289 : /// Does not change over the lifetime of the [`Tenant`] object.
290 : ///
291 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
292 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
293 : generation: Generation,
294 :
295 : timelines: Mutex<HashMap<TimelineId, Arc<Timeline>>>,
296 :
297 : /// During timeline creation, we first insert the TimelineId to the
298 : /// creating map, then `timelines`, then remove it from the creating map.
299 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
300 : timelines_creating: std::sync::Mutex<HashSet<TimelineId>>,
301 :
302 : /// Possibly offloaded and archived timelines
303 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
304 : timelines_offloaded: Mutex<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
305 :
306 : /// Serialize writes of the tenant manifest to remote storage. If there are concurrent operations
307 : /// affecting the manifest, such as timeline deletion and timeline offload, they must wait for
308 : /// each other (this could be optimized to coalesce writes if necessary).
309 : ///
310 : /// The contents of the Mutex are the last manifest we successfully uploaded
311 : tenant_manifest_upload: tokio::sync::Mutex<Option<TenantManifest>>,
312 :
313 : // This mutex prevents creation of new timelines during GC.
314 : // Adding yet another mutex (in addition to `timelines`) is needed because holding
315 : // `timelines` mutex during all GC iteration
316 : // may block for a long time `get_timeline`, `get_timelines_state`,... and other operations
317 : // with timelines, which in turn may cause dropping replication connection, expiration of wait_for_lsn
318 : // timeout...
319 : gc_cs: tokio::sync::Mutex<()>,
320 : walredo_mgr: Option<Arc<WalRedoManager>>,
321 :
322 : // provides access to timeline data sitting in the remote storage
323 : pub(crate) remote_storage: GenericRemoteStorage,
324 :
325 : // Access to global deletion queue for when this tenant wants to schedule a deletion
326 : deletion_queue_client: DeletionQueueClient,
327 :
328 : /// Cached logical sizes updated updated on each [`Tenant::gather_size_inputs`].
329 : cached_logical_sizes: tokio::sync::Mutex<HashMap<(TimelineId, Lsn), u64>>,
330 : cached_synthetic_tenant_size: Arc<AtomicU64>,
331 :
332 : eviction_task_tenant_state: tokio::sync::Mutex<EvictionTaskTenantState>,
333 :
334 : /// Track repeated failures to compact, so that we can back off.
335 : /// Overhead of mutex is acceptable because compaction is done with a multi-second period.
336 : compaction_circuit_breaker: std::sync::Mutex<CircuitBreaker>,
337 :
338 : /// If the tenant is in Activating state, notify this to encourage it
339 : /// to proceed to Active as soon as possible, rather than waiting for lazy
340 : /// background warmup.
341 : pub(crate) activate_now_sem: tokio::sync::Semaphore,
342 :
343 : /// Time it took for the tenant to activate. Zero if not active yet.
344 : attach_wal_lag_cooldown: Arc<std::sync::OnceLock<WalLagCooldown>>,
345 :
346 : // Cancellation token fires when we have entered shutdown(). This is a parent of
347 : // Timelines' cancellation token.
348 : pub(crate) cancel: CancellationToken,
349 :
350 : // Users of the Tenant such as the page service must take this Gate to avoid
351 : // trying to use a Tenant which is shutting down.
352 : pub(crate) gate: Gate,
353 :
354 : /// Throttle applied at the top of [`Timeline::get`].
355 : /// All [`Tenant::timelines`] of a given [`Tenant`] instance share the same [`throttle::Throttle`] instance.
356 : pub(crate) timeline_get_throttle:
357 : Arc<throttle::Throttle<crate::metrics::tenant_throttling::TimelineGet>>,
358 :
359 : /// An ongoing timeline detach concurrency limiter.
360 : ///
361 : /// As a tenant will likely be restarted as part of timeline detach ancestor it makes no sense
362 : /// to have two running at the same time. A different one can be started if an earlier one
363 : /// has failed for whatever reason.
364 : ongoing_timeline_detach: std::sync::Mutex<Option<(TimelineId, utils::completion::Barrier)>>,
365 :
366 : /// `index_part.json` based gc blocking reason tracking.
367 : ///
368 : /// New gc iterations must start a new iteration by acquiring `GcBlock::start` before
369 : /// proceeding.
370 : pub(crate) gc_block: gc_block::GcBlock,
371 :
372 : l0_flush_global_state: L0FlushGlobalState,
373 : }
374 :
375 : impl std::fmt::Debug for Tenant {
376 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
377 0 : write!(f, "{} ({})", self.tenant_shard_id, self.current_state())
378 0 : }
379 : }
380 :
381 : pub(crate) enum WalRedoManager {
382 : Prod(WalredoManagerId, PostgresRedoManager),
383 : #[cfg(test)]
384 : Test(harness::TestRedoManager),
385 : }
386 :
387 0 : #[derive(thiserror::Error, Debug)]
388 : #[error("pageserver is shutting down")]
389 : pub(crate) struct GlobalShutDown;
390 :
391 : impl WalRedoManager {
392 0 : pub(crate) fn new(mgr: PostgresRedoManager) -> Result<Arc<Self>, GlobalShutDown> {
393 0 : let id = WalredoManagerId::next();
394 0 : let arc = Arc::new(Self::Prod(id, mgr));
395 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
396 0 : match &mut *guard {
397 0 : Some(map) => {
398 0 : map.insert(id, Arc::downgrade(&arc));
399 0 : Ok(arc)
400 : }
401 0 : None => Err(GlobalShutDown),
402 : }
403 0 : }
404 : }
405 :
406 : impl Drop for WalRedoManager {
407 10 : fn drop(&mut self) {
408 10 : match self {
409 0 : Self::Prod(id, _) => {
410 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
411 0 : if let Some(map) = &mut *guard {
412 0 : map.remove(id).expect("new() registers, drop() unregisters");
413 0 : }
414 : }
415 : #[cfg(test)]
416 10 : Self::Test(_) => {
417 10 : // Not applicable to test redo manager
418 10 : }
419 : }
420 10 : }
421 : }
422 :
423 : /// Global registry of all walredo managers so that [`crate::shutdown_pageserver`] can shut down
424 : /// the walredo processes outside of the regular order.
425 : ///
426 : /// This is necessary to work around a systemd bug where it freezes if there are
427 : /// walredo processes left => <https://github.com/neondatabase/cloud/issues/11387>
428 : #[allow(clippy::type_complexity)]
429 : pub(crate) static WALREDO_MANAGERS: once_cell::sync::Lazy<
430 : Mutex<Option<HashMap<WalredoManagerId, Weak<WalRedoManager>>>>,
431 0 : > = once_cell::sync::Lazy::new(|| Mutex::new(Some(HashMap::new())));
432 : #[derive(PartialEq, Eq, Hash, Clone, Copy, Debug)]
433 : pub(crate) struct WalredoManagerId(u64);
434 : impl WalredoManagerId {
435 0 : pub fn next() -> Self {
436 : static NEXT: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(1);
437 0 : let id = NEXT.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
438 0 : if id == 0 {
439 0 : panic!("WalredoManagerId::new() returned 0, indicating wraparound, risking it's no longer unique");
440 0 : }
441 0 : Self(id)
442 0 : }
443 : }
444 :
445 : #[cfg(test)]
446 : impl From<harness::TestRedoManager> for WalRedoManager {
447 192 : fn from(mgr: harness::TestRedoManager) -> Self {
448 192 : Self::Test(mgr)
449 192 : }
450 : }
451 :
452 : impl WalRedoManager {
453 6 : pub(crate) async fn shutdown(&self) -> bool {
454 6 : match self {
455 0 : Self::Prod(_, mgr) => mgr.shutdown().await,
456 : #[cfg(test)]
457 : Self::Test(_) => {
458 : // Not applicable to test redo manager
459 6 : true
460 : }
461 : }
462 6 : }
463 :
464 0 : pub(crate) fn maybe_quiesce(&self, idle_timeout: Duration) {
465 0 : match self {
466 0 : Self::Prod(_, mgr) => mgr.maybe_quiesce(idle_timeout),
467 0 : #[cfg(test)]
468 0 : Self::Test(_) => {
469 0 : // Not applicable to test redo manager
470 0 : }
471 0 : }
472 0 : }
473 :
474 : /// # Cancel-Safety
475 : ///
476 : /// This method is cancellation-safe.
477 410 : pub async fn request_redo(
478 410 : &self,
479 410 : key: pageserver_api::key::Key,
480 410 : lsn: Lsn,
481 410 : base_img: Option<(Lsn, bytes::Bytes)>,
482 410 : records: Vec<(Lsn, pageserver_api::record::NeonWalRecord)>,
483 410 : pg_version: u32,
484 410 : ) -> Result<bytes::Bytes, walredo::Error> {
485 410 : match self {
486 0 : Self::Prod(_, mgr) => {
487 0 : mgr.request_redo(key, lsn, base_img, records, pg_version)
488 0 : .await
489 : }
490 : #[cfg(test)]
491 410 : Self::Test(mgr) => {
492 410 : mgr.request_redo(key, lsn, base_img, records, pg_version)
493 0 : .await
494 : }
495 : }
496 410 : }
497 :
498 0 : pub(crate) fn status(&self) -> Option<WalRedoManagerStatus> {
499 0 : match self {
500 0 : WalRedoManager::Prod(_, m) => Some(m.status()),
501 0 : #[cfg(test)]
502 0 : WalRedoManager::Test(_) => None,
503 0 : }
504 0 : }
505 : }
506 :
507 : /// A very lightweight memory representation of an offloaded timeline.
508 : ///
509 : /// We need to store the list of offloaded timelines so that we can perform operations on them,
510 : /// like unoffloading them, or (at a later date), decide to perform flattening.
511 : /// This type has a much smaller memory impact than [`Timeline`], and thus we can store many
512 : /// more offloaded timelines than we can manage ones that aren't.
513 : pub struct OffloadedTimeline {
514 : pub tenant_shard_id: TenantShardId,
515 : pub timeline_id: TimelineId,
516 : pub ancestor_timeline_id: Option<TimelineId>,
517 : /// Whether to retain the branch lsn at the ancestor or not
518 : pub ancestor_retain_lsn: Option<Lsn>,
519 :
520 : /// When the timeline was archived.
521 : ///
522 : /// Present for future flattening deliberations.
523 : pub archived_at: NaiveDateTime,
524 :
525 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
526 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
527 : pub delete_progress: TimelineDeleteProgress,
528 :
529 : /// Part of the `OffloadedTimeline` object's lifecycle: this needs to be set before we drop it
530 : pub deleted_from_ancestor: AtomicBool,
531 : }
532 :
533 : impl OffloadedTimeline {
534 : /// Obtains an offloaded timeline from a given timeline object.
535 : ///
536 : /// Returns `None` if the `archived_at` flag couldn't be obtained, i.e.
537 : /// the timeline is not in a stopped state.
538 : /// Panics if the timeline is not archived.
539 2 : fn from_timeline(timeline: &Timeline) -> Result<Self, UploadQueueNotReadyError> {
540 2 : let (ancestor_retain_lsn, ancestor_timeline_id) =
541 2 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
542 2 : let ancestor_lsn = timeline.get_ancestor_lsn();
543 2 : let ancestor_timeline_id = ancestor_timeline.timeline_id;
544 2 : let mut gc_info = ancestor_timeline.gc_info.write().unwrap();
545 2 : gc_info.insert_child(timeline.timeline_id, ancestor_lsn, MaybeOffloaded::Yes);
546 2 : (Some(ancestor_lsn), Some(ancestor_timeline_id))
547 : } else {
548 0 : (None, None)
549 : };
550 2 : let archived_at = timeline
551 2 : .remote_client
552 2 : .archived_at_stopped_queue()?
553 2 : .expect("must be called on an archived timeline");
554 2 : Ok(Self {
555 2 : tenant_shard_id: timeline.tenant_shard_id,
556 2 : timeline_id: timeline.timeline_id,
557 2 : ancestor_timeline_id,
558 2 : ancestor_retain_lsn,
559 2 : archived_at,
560 2 :
561 2 : delete_progress: timeline.delete_progress.clone(),
562 2 : deleted_from_ancestor: AtomicBool::new(false),
563 2 : })
564 2 : }
565 0 : fn from_manifest(tenant_shard_id: TenantShardId, manifest: &OffloadedTimelineManifest) -> Self {
566 0 : // We expect to reach this case in tenant loading, where the `retain_lsn` is populated in the parent's `gc_info`
567 0 : // by the `initialize_gc_info` function.
568 0 : let OffloadedTimelineManifest {
569 0 : timeline_id,
570 0 : ancestor_timeline_id,
571 0 : ancestor_retain_lsn,
572 0 : archived_at,
573 0 : } = *manifest;
574 0 : Self {
575 0 : tenant_shard_id,
576 0 : timeline_id,
577 0 : ancestor_timeline_id,
578 0 : ancestor_retain_lsn,
579 0 : archived_at,
580 0 : delete_progress: TimelineDeleteProgress::default(),
581 0 : deleted_from_ancestor: AtomicBool::new(false),
582 0 : }
583 0 : }
584 2 : fn manifest(&self) -> OffloadedTimelineManifest {
585 2 : let Self {
586 2 : timeline_id,
587 2 : ancestor_timeline_id,
588 2 : ancestor_retain_lsn,
589 2 : archived_at,
590 2 : ..
591 2 : } = self;
592 2 : OffloadedTimelineManifest {
593 2 : timeline_id: *timeline_id,
594 2 : ancestor_timeline_id: *ancestor_timeline_id,
595 2 : ancestor_retain_lsn: *ancestor_retain_lsn,
596 2 : archived_at: *archived_at,
597 2 : }
598 2 : }
599 : /// Delete this timeline's retain_lsn from its ancestor, if present in the given tenant
600 0 : fn delete_from_ancestor_with_timelines(
601 0 : &self,
602 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
603 0 : ) {
604 0 : if let (Some(_retain_lsn), Some(ancestor_timeline_id)) =
605 0 : (self.ancestor_retain_lsn, self.ancestor_timeline_id)
606 : {
607 0 : if let Some((_, ancestor_timeline)) = timelines
608 0 : .iter()
609 0 : .find(|(tid, _tl)| **tid == ancestor_timeline_id)
610 : {
611 0 : let removal_happened = ancestor_timeline
612 0 : .gc_info
613 0 : .write()
614 0 : .unwrap()
615 0 : .remove_child_offloaded(self.timeline_id);
616 0 : if !removal_happened {
617 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
618 0 : "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
619 0 : }
620 0 : }
621 0 : }
622 0 : self.deleted_from_ancestor.store(true, Ordering::Release);
623 0 : }
624 : /// Call [`Self::delete_from_ancestor_with_timelines`] instead if possible.
625 : ///
626 : /// As the entire tenant is being dropped, don't bother deregistering the `retain_lsn` from the ancestor.
627 2 : fn defuse_for_tenant_drop(&self) {
628 2 : self.deleted_from_ancestor.store(true, Ordering::Release);
629 2 : }
630 : }
631 :
632 : impl fmt::Debug for OffloadedTimeline {
633 0 : fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
634 0 : write!(f, "OffloadedTimeline<{}>", self.timeline_id)
635 0 : }
636 : }
637 :
638 : impl Drop for OffloadedTimeline {
639 2 : fn drop(&mut self) {
640 2 : if !self.deleted_from_ancestor.load(Ordering::Acquire) {
641 0 : tracing::warn!(
642 0 : "offloaded timeline {} was dropped without having cleaned it up at the ancestor",
643 : self.timeline_id
644 : );
645 2 : }
646 2 : }
647 : }
648 :
649 : #[derive(Copy, Clone, PartialEq, Eq, Hash, Debug)]
650 : pub enum MaybeOffloaded {
651 : Yes,
652 : No,
653 : }
654 :
655 : #[derive(Clone, Debug)]
656 : pub enum TimelineOrOffloaded {
657 : Timeline(Arc<Timeline>),
658 : Offloaded(Arc<OffloadedTimeline>),
659 : }
660 :
661 : impl TimelineOrOffloaded {
662 0 : pub fn arc_ref(&self) -> TimelineOrOffloadedArcRef<'_> {
663 0 : match self {
664 0 : TimelineOrOffloaded::Timeline(timeline) => {
665 0 : TimelineOrOffloadedArcRef::Timeline(timeline)
666 : }
667 0 : TimelineOrOffloaded::Offloaded(offloaded) => {
668 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded)
669 : }
670 : }
671 0 : }
672 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
673 0 : self.arc_ref().tenant_shard_id()
674 0 : }
675 0 : pub fn timeline_id(&self) -> TimelineId {
676 0 : self.arc_ref().timeline_id()
677 0 : }
678 2 : pub fn delete_progress(&self) -> &Arc<tokio::sync::Mutex<DeleteTimelineFlow>> {
679 2 : match self {
680 2 : TimelineOrOffloaded::Timeline(timeline) => &timeline.delete_progress,
681 0 : TimelineOrOffloaded::Offloaded(offloaded) => &offloaded.delete_progress,
682 : }
683 2 : }
684 0 : fn maybe_remote_client(&self) -> Option<Arc<RemoteTimelineClient>> {
685 0 : match self {
686 0 : TimelineOrOffloaded::Timeline(timeline) => Some(timeline.remote_client.clone()),
687 0 : TimelineOrOffloaded::Offloaded(_offloaded) => None,
688 : }
689 0 : }
690 : }
691 :
692 : pub enum TimelineOrOffloadedArcRef<'a> {
693 : Timeline(&'a Arc<Timeline>),
694 : Offloaded(&'a Arc<OffloadedTimeline>),
695 : }
696 :
697 : impl TimelineOrOffloadedArcRef<'_> {
698 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
699 0 : match self {
700 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.tenant_shard_id,
701 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.tenant_shard_id,
702 : }
703 0 : }
704 0 : pub fn timeline_id(&self) -> TimelineId {
705 0 : match self {
706 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.timeline_id,
707 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.timeline_id,
708 : }
709 0 : }
710 : }
711 :
712 : impl<'a> From<&'a Arc<Timeline>> for TimelineOrOffloadedArcRef<'a> {
713 0 : fn from(timeline: &'a Arc<Timeline>) -> Self {
714 0 : Self::Timeline(timeline)
715 0 : }
716 : }
717 :
718 : impl<'a> From<&'a Arc<OffloadedTimeline>> for TimelineOrOffloadedArcRef<'a> {
719 0 : fn from(timeline: &'a Arc<OffloadedTimeline>) -> Self {
720 0 : Self::Offloaded(timeline)
721 0 : }
722 : }
723 :
724 0 : #[derive(Debug, thiserror::Error, PartialEq, Eq)]
725 : pub enum GetTimelineError {
726 : #[error("Timeline is shutting down")]
727 : ShuttingDown,
728 : #[error("Timeline {tenant_id}/{timeline_id} is not active, state: {state:?}")]
729 : NotActive {
730 : tenant_id: TenantShardId,
731 : timeline_id: TimelineId,
732 : state: TimelineState,
733 : },
734 : #[error("Timeline {tenant_id}/{timeline_id} was not found")]
735 : NotFound {
736 : tenant_id: TenantShardId,
737 : timeline_id: TimelineId,
738 : },
739 : }
740 :
741 0 : #[derive(Debug, thiserror::Error)]
742 : pub enum LoadLocalTimelineError {
743 : #[error("FailedToLoad")]
744 : Load(#[source] anyhow::Error),
745 : #[error("FailedToResumeDeletion")]
746 : ResumeDeletion(#[source] anyhow::Error),
747 : }
748 :
749 0 : #[derive(thiserror::Error)]
750 : pub enum DeleteTimelineError {
751 : #[error("NotFound")]
752 : NotFound,
753 :
754 : #[error("HasChildren")]
755 : HasChildren(Vec<TimelineId>),
756 :
757 : #[error("Timeline deletion is already in progress")]
758 : AlreadyInProgress(Arc<tokio::sync::Mutex<DeleteTimelineFlow>>),
759 :
760 : #[error("Cancelled")]
761 : Cancelled,
762 :
763 : #[error(transparent)]
764 : Other(#[from] anyhow::Error),
765 : }
766 :
767 : impl Debug for DeleteTimelineError {
768 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
769 0 : match self {
770 0 : Self::NotFound => write!(f, "NotFound"),
771 0 : Self::HasChildren(c) => f.debug_tuple("HasChildren").field(c).finish(),
772 0 : Self::AlreadyInProgress(_) => f.debug_tuple("AlreadyInProgress").finish(),
773 0 : Self::Cancelled => f.debug_tuple("Cancelled").finish(),
774 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
775 : }
776 0 : }
777 : }
778 :
779 0 : #[derive(thiserror::Error)]
780 : pub enum TimelineArchivalError {
781 : #[error("NotFound")]
782 : NotFound,
783 :
784 : #[error("Timeout")]
785 : Timeout,
786 :
787 : #[error("Cancelled")]
788 : Cancelled,
789 :
790 : #[error("ancestor is archived: {}", .0)]
791 : HasArchivedParent(TimelineId),
792 :
793 : #[error("HasUnarchivedChildren")]
794 : HasUnarchivedChildren(Vec<TimelineId>),
795 :
796 : #[error("Timeline archival is already in progress")]
797 : AlreadyInProgress,
798 :
799 : #[error(transparent)]
800 : Other(anyhow::Error),
801 : }
802 :
803 0 : #[derive(thiserror::Error, Debug)]
804 : pub(crate) enum TenantManifestError {
805 : #[error("Remote storage error: {0}")]
806 : RemoteStorage(anyhow::Error),
807 :
808 : #[error("Cancelled")]
809 : Cancelled,
810 : }
811 :
812 : impl From<TenantManifestError> for TimelineArchivalError {
813 0 : fn from(e: TenantManifestError) -> Self {
814 0 : match e {
815 0 : TenantManifestError::RemoteStorage(e) => Self::Other(e),
816 0 : TenantManifestError::Cancelled => Self::Cancelled,
817 : }
818 0 : }
819 : }
820 :
821 : impl Debug for TimelineArchivalError {
822 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
823 0 : match self {
824 0 : Self::NotFound => write!(f, "NotFound"),
825 0 : Self::Timeout => write!(f, "Timeout"),
826 0 : Self::Cancelled => write!(f, "Cancelled"),
827 0 : Self::HasArchivedParent(p) => f.debug_tuple("HasArchivedParent").field(p).finish(),
828 0 : Self::HasUnarchivedChildren(c) => {
829 0 : f.debug_tuple("HasUnarchivedChildren").field(c).finish()
830 : }
831 0 : Self::AlreadyInProgress => f.debug_tuple("AlreadyInProgress").finish(),
832 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
833 : }
834 0 : }
835 : }
836 :
837 : pub enum SetStoppingError {
838 : AlreadyStopping(completion::Barrier),
839 : Broken,
840 : }
841 :
842 : impl Debug for SetStoppingError {
843 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
844 0 : match self {
845 0 : Self::AlreadyStopping(_) => f.debug_tuple("AlreadyStopping").finish(),
846 0 : Self::Broken => write!(f, "Broken"),
847 : }
848 0 : }
849 : }
850 :
851 : /// Arguments to [`Tenant::create_timeline`].
852 : ///
853 : /// Not usable as an idempotency key for timeline creation because if [`CreateTimelineParamsBranch::ancestor_start_lsn`]
854 : /// is `None`, the result of the timeline create call is not deterministic.
855 : ///
856 : /// See [`CreateTimelineIdempotency`] for an idempotency key.
857 : #[derive(Debug)]
858 : pub(crate) enum CreateTimelineParams {
859 : Bootstrap(CreateTimelineParamsBootstrap),
860 : Branch(CreateTimelineParamsBranch),
861 : }
862 :
863 : #[derive(Debug)]
864 : pub(crate) struct CreateTimelineParamsBootstrap {
865 : pub(crate) new_timeline_id: TimelineId,
866 : pub(crate) existing_initdb_timeline_id: Option<TimelineId>,
867 : pub(crate) pg_version: u32,
868 : }
869 :
870 : /// NB: See comment on [`CreateTimelineIdempotency::Branch`] for why there's no `pg_version` here.
871 : #[derive(Debug)]
872 : pub(crate) struct CreateTimelineParamsBranch {
873 : pub(crate) new_timeline_id: TimelineId,
874 : pub(crate) ancestor_timeline_id: TimelineId,
875 : pub(crate) ancestor_start_lsn: Option<Lsn>,
876 : }
877 :
878 : /// What is used to determine idempotency of a [`Tenant::create_timeline`] call in [`Tenant::start_creating_timeline`].
879 : ///
880 : /// Each [`Timeline`] object holds [`Self`] as an immutable property in [`Timeline::create_idempotency`].
881 : ///
882 : /// We lower timeline creation requests to [`Self`], and then use [`PartialEq::eq`] to compare [`Timeline::create_idempotency`] with the request.
883 : /// If they are equal, we return a reference to the existing timeline, otherwise it's an idempotency conflict.
884 : ///
885 : /// There is special treatment for [`Self::FailWithConflict`] to always return an idempotency conflict.
886 : /// It would be nice to have more advanced derive macros to make that special treatment declarative.
887 : ///
888 : /// Notes:
889 : /// - Unlike [`CreateTimelineParams`], ancestor LSN is fixed, so, branching will be at a deterministic LSN.
890 : /// - We make some trade-offs though, e.g., [`CreateTimelineParamsBootstrap::existing_initdb_timeline_id`]
891 : /// is not considered for idempotency. We can improve on this over time if we deem it necessary.
892 : ///
893 : #[derive(Debug, Clone, PartialEq, Eq)]
894 : pub(crate) enum CreateTimelineIdempotency {
895 : /// NB: special treatment, see comment in [`Self`].
896 : FailWithConflict,
897 : Bootstrap {
898 : pg_version: u32,
899 : },
900 : /// NB: branches always have the same `pg_version` as their ancestor.
901 : /// While [`pageserver_api::models::TimelineCreateRequestMode::Branch::pg_version`]
902 : /// exists as a field, and is set by cplane, it has always been ignored by pageserver when
903 : /// determining the child branch pg_version.
904 : Branch {
905 : ancestor_timeline_id: TimelineId,
906 : ancestor_start_lsn: Lsn,
907 : },
908 : }
909 :
910 : /// What is returned by [`Tenant::start_creating_timeline`].
911 : #[must_use]
912 : enum StartCreatingTimelineResult<'t> {
913 : CreateGuard(TimelineCreateGuard<'t>),
914 : Idempotent(Arc<Timeline>),
915 : }
916 :
917 : /// What is returned by [`Tenant::create_timeline`].
918 : enum CreateTimelineResult {
919 : Created(Arc<Timeline>),
920 : Idempotent(Arc<Timeline>),
921 : }
922 :
923 : impl CreateTimelineResult {
924 0 : fn discriminant(&self) -> &'static str {
925 0 : match self {
926 0 : Self::Created(_) => "Created",
927 0 : Self::Idempotent(_) => "Idempotent",
928 : }
929 0 : }
930 0 : fn timeline(&self) -> &Arc<Timeline> {
931 0 : match self {
932 0 : Self::Created(t) | Self::Idempotent(t) => t,
933 0 : }
934 0 : }
935 : /// Unit test timelines aren't activated, test has to do it if it needs to.
936 : #[cfg(test)]
937 230 : fn into_timeline_for_test(self) -> Arc<Timeline> {
938 230 : match self {
939 230 : Self::Created(t) | Self::Idempotent(t) => t,
940 230 : }
941 230 : }
942 : }
943 :
944 2 : #[derive(thiserror::Error, Debug)]
945 : pub enum CreateTimelineError {
946 : #[error("creation of timeline with the given ID is in progress")]
947 : AlreadyCreating,
948 : #[error("timeline already exists with different parameters")]
949 : Conflict,
950 : #[error(transparent)]
951 : AncestorLsn(anyhow::Error),
952 : #[error("ancestor timeline is not active")]
953 : AncestorNotActive,
954 : #[error("ancestor timeline is archived")]
955 : AncestorArchived,
956 : #[error("tenant shutting down")]
957 : ShuttingDown,
958 : #[error(transparent)]
959 : Other(#[from] anyhow::Error),
960 : }
961 :
962 : #[derive(thiserror::Error, Debug)]
963 : enum InitdbError {
964 : Other(anyhow::Error),
965 : Cancelled,
966 : Spawn(std::io::Result<()>),
967 : Failed(std::process::ExitStatus, Vec<u8>),
968 : }
969 :
970 : impl fmt::Display for InitdbError {
971 0 : fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
972 0 : match self {
973 0 : InitdbError::Cancelled => write!(f, "Operation was cancelled"),
974 0 : InitdbError::Spawn(e) => write!(f, "Spawn error: {:?}", e),
975 0 : InitdbError::Failed(status, stderr) => write!(
976 0 : f,
977 0 : "Command failed with status {:?}: {}",
978 0 : status,
979 0 : String::from_utf8_lossy(stderr)
980 0 : ),
981 0 : InitdbError::Other(e) => write!(f, "Error: {:?}", e),
982 : }
983 0 : }
984 : }
985 :
986 : impl From<std::io::Error> for InitdbError {
987 0 : fn from(error: std::io::Error) -> Self {
988 0 : InitdbError::Spawn(Err(error))
989 0 : }
990 : }
991 :
992 : enum CreateTimelineCause {
993 : Load,
994 : Delete,
995 : }
996 :
997 0 : #[derive(thiserror::Error, Debug)]
998 : pub(crate) enum GcError {
999 : // The tenant is shutting down
1000 : #[error("tenant shutting down")]
1001 : TenantCancelled,
1002 :
1003 : // The tenant is shutting down
1004 : #[error("timeline shutting down")]
1005 : TimelineCancelled,
1006 :
1007 : // The tenant is in a state inelegible to run GC
1008 : #[error("not active")]
1009 : NotActive,
1010 :
1011 : // A requested GC cutoff LSN was invalid, for example it tried to move backwards
1012 : #[error("not active")]
1013 : BadLsn { why: String },
1014 :
1015 : // A remote storage error while scheduling updates after compaction
1016 : #[error(transparent)]
1017 : Remote(anyhow::Error),
1018 :
1019 : // An error reading while calculating GC cutoffs
1020 : #[error(transparent)]
1021 : GcCutoffs(PageReconstructError),
1022 :
1023 : // If GC was invoked for a particular timeline, this error means it didn't exist
1024 : #[error("timeline not found")]
1025 : TimelineNotFound,
1026 : }
1027 :
1028 : impl From<PageReconstructError> for GcError {
1029 0 : fn from(value: PageReconstructError) -> Self {
1030 0 : match value {
1031 0 : PageReconstructError::Cancelled => Self::TimelineCancelled,
1032 0 : other => Self::GcCutoffs(other),
1033 : }
1034 0 : }
1035 : }
1036 :
1037 : impl From<NotInitialized> for GcError {
1038 0 : fn from(value: NotInitialized) -> Self {
1039 0 : match value {
1040 0 : NotInitialized::Uninitialized => GcError::Remote(value.into()),
1041 0 : NotInitialized::Stopped | NotInitialized::ShuttingDown => GcError::TimelineCancelled,
1042 : }
1043 0 : }
1044 : }
1045 :
1046 : impl From<timeline::layer_manager::Shutdown> for GcError {
1047 0 : fn from(_: timeline::layer_manager::Shutdown) -> Self {
1048 0 : GcError::TimelineCancelled
1049 0 : }
1050 : }
1051 :
1052 0 : #[derive(thiserror::Error, Debug)]
1053 : pub(crate) enum LoadConfigError {
1054 : #[error("TOML deserialization error: '{0}'")]
1055 : DeserializeToml(#[from] toml_edit::de::Error),
1056 :
1057 : #[error("Config not found at {0}")]
1058 : NotFound(Utf8PathBuf),
1059 : }
1060 :
1061 : impl Tenant {
1062 : /// Yet another helper for timeline initialization.
1063 : ///
1064 : /// - Initializes the Timeline struct and inserts it into the tenant's hash map
1065 : /// - Scans the local timeline directory for layer files and builds the layer map
1066 : /// - Downloads remote index file and adds remote files to the layer map
1067 : /// - Schedules remote upload tasks for any files that are present locally but missing from remote storage.
1068 : ///
1069 : /// If the operation fails, the timeline is left in the tenant's hash map in Broken state. On success,
1070 : /// it is marked as Active.
1071 : #[allow(clippy::too_many_arguments)]
1072 6 : async fn timeline_init_and_sync(
1073 6 : &self,
1074 6 : timeline_id: TimelineId,
1075 6 : resources: TimelineResources,
1076 6 : index_part: IndexPart,
1077 6 : metadata: TimelineMetadata,
1078 6 : ancestor: Option<Arc<Timeline>>,
1079 6 : _ctx: &RequestContext,
1080 6 : ) -> anyhow::Result<()> {
1081 6 : let tenant_id = self.tenant_shard_id;
1082 :
1083 6 : let idempotency = if metadata.ancestor_timeline().is_none() {
1084 4 : CreateTimelineIdempotency::Bootstrap {
1085 4 : pg_version: metadata.pg_version(),
1086 4 : }
1087 : } else {
1088 2 : CreateTimelineIdempotency::Branch {
1089 2 : ancestor_timeline_id: metadata.ancestor_timeline().unwrap(),
1090 2 : ancestor_start_lsn: metadata.ancestor_lsn(),
1091 2 : }
1092 : };
1093 :
1094 6 : let timeline = self.create_timeline_struct(
1095 6 : timeline_id,
1096 6 : &metadata,
1097 6 : ancestor.clone(),
1098 6 : resources,
1099 6 : CreateTimelineCause::Load,
1100 6 : idempotency.clone(),
1101 6 : )?;
1102 6 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
1103 6 : anyhow::ensure!(
1104 6 : disk_consistent_lsn.is_valid(),
1105 0 : "Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn"
1106 : );
1107 6 : assert_eq!(
1108 6 : disk_consistent_lsn,
1109 6 : metadata.disk_consistent_lsn(),
1110 0 : "these are used interchangeably"
1111 : );
1112 :
1113 6 : timeline.remote_client.init_upload_queue(&index_part)?;
1114 :
1115 6 : timeline
1116 6 : .load_layer_map(disk_consistent_lsn, index_part)
1117 3 : .await
1118 6 : .with_context(|| {
1119 0 : format!("Failed to load layermap for timeline {tenant_id}/{timeline_id}")
1120 6 : })?;
1121 :
1122 : {
1123 : // avoiding holding it across awaits
1124 6 : let mut timelines_accessor = self.timelines.lock().unwrap();
1125 6 : match timelines_accessor.entry(timeline_id) {
1126 : // We should never try and load the same timeline twice during startup
1127 : Entry::Occupied(_) => {
1128 0 : unreachable!(
1129 0 : "Timeline {tenant_id}/{timeline_id} already exists in the tenant map"
1130 0 : );
1131 : }
1132 6 : Entry::Vacant(v) => {
1133 6 : v.insert(Arc::clone(&timeline));
1134 6 : timeline.maybe_spawn_flush_loop();
1135 6 : }
1136 6 : }
1137 6 : };
1138 6 :
1139 6 : // Sanity check: a timeline should have some content.
1140 6 : anyhow::ensure!(
1141 6 : ancestor.is_some()
1142 4 : || timeline
1143 4 : .layers
1144 4 : .read()
1145 0 : .await
1146 4 : .layer_map()
1147 4 : .expect("currently loading, layer manager cannot be shutdown already")
1148 4 : .iter_historic_layers()
1149 4 : .next()
1150 4 : .is_some(),
1151 0 : "Timeline has no ancestor and no layer files"
1152 : );
1153 :
1154 6 : Ok(())
1155 6 : }
1156 :
1157 : /// Attach a tenant that's available in cloud storage.
1158 : ///
1159 : /// This returns quickly, after just creating the in-memory object
1160 : /// Tenant struct and launching a background task to download
1161 : /// the remote index files. On return, the tenant is most likely still in
1162 : /// Attaching state, and it will become Active once the background task
1163 : /// finishes. You can use wait_until_active() to wait for the task to
1164 : /// complete.
1165 : ///
1166 : #[allow(clippy::too_many_arguments)]
1167 0 : pub(crate) fn spawn(
1168 0 : conf: &'static PageServerConf,
1169 0 : tenant_shard_id: TenantShardId,
1170 0 : resources: TenantSharedResources,
1171 0 : attached_conf: AttachedTenantConf,
1172 0 : shard_identity: ShardIdentity,
1173 0 : init_order: Option<InitializationOrder>,
1174 0 : mode: SpawnMode,
1175 0 : ctx: &RequestContext,
1176 0 : ) -> Result<Arc<Tenant>, GlobalShutDown> {
1177 0 : let wal_redo_manager =
1178 0 : WalRedoManager::new(PostgresRedoManager::new(conf, tenant_shard_id))?;
1179 :
1180 : let TenantSharedResources {
1181 0 : broker_client,
1182 0 : remote_storage,
1183 0 : deletion_queue_client,
1184 0 : l0_flush_global_state,
1185 0 : } = resources;
1186 0 :
1187 0 : let attach_mode = attached_conf.location.attach_mode;
1188 0 : let generation = attached_conf.location.generation;
1189 0 :
1190 0 : let tenant = Arc::new(Tenant::new(
1191 0 : TenantState::Attaching,
1192 0 : conf,
1193 0 : attached_conf,
1194 0 : shard_identity,
1195 0 : Some(wal_redo_manager),
1196 0 : tenant_shard_id,
1197 0 : remote_storage.clone(),
1198 0 : deletion_queue_client,
1199 0 : l0_flush_global_state,
1200 0 : ));
1201 0 :
1202 0 : // The attach task will carry a GateGuard, so that shutdown() reliably waits for it to drop out if
1203 0 : // we shut down while attaching.
1204 0 : let attach_gate_guard = tenant
1205 0 : .gate
1206 0 : .enter()
1207 0 : .expect("We just created the Tenant: nothing else can have shut it down yet");
1208 0 :
1209 0 : // Do all the hard work in the background
1210 0 : let tenant_clone = Arc::clone(&tenant);
1211 0 : let ctx = ctx.detached_child(TaskKind::Attach, DownloadBehavior::Warn);
1212 0 : task_mgr::spawn(
1213 0 : &tokio::runtime::Handle::current(),
1214 0 : TaskKind::Attach,
1215 0 : tenant_shard_id,
1216 0 : None,
1217 0 : "attach tenant",
1218 0 : async move {
1219 0 :
1220 0 : info!(
1221 : ?attach_mode,
1222 0 : "Attaching tenant"
1223 : );
1224 :
1225 0 : let _gate_guard = attach_gate_guard;
1226 0 :
1227 0 : // Is this tenant being spawned as part of process startup?
1228 0 : let starting_up = init_order.is_some();
1229 0 : scopeguard::defer! {
1230 0 : if starting_up {
1231 0 : TENANT.startup_complete.inc();
1232 0 : }
1233 0 : }
1234 :
1235 : // Ideally we should use Tenant::set_broken_no_wait, but it is not supposed to be used when tenant is in loading state.
1236 : enum BrokenVerbosity {
1237 : Error,
1238 : Info
1239 : }
1240 0 : let make_broken =
1241 0 : |t: &Tenant, err: anyhow::Error, verbosity: BrokenVerbosity| {
1242 0 : match verbosity {
1243 : BrokenVerbosity::Info => {
1244 0 : info!("attach cancelled, setting tenant state to Broken: {err}");
1245 : },
1246 : BrokenVerbosity::Error => {
1247 0 : error!("attach failed, setting tenant state to Broken: {err:?}");
1248 : }
1249 : }
1250 0 : t.state.send_modify(|state| {
1251 0 : // The Stopping case is for when we have passed control on to DeleteTenantFlow:
1252 0 : // if it errors, we will call make_broken when tenant is already in Stopping.
1253 0 : assert!(
1254 0 : matches!(*state, TenantState::Attaching | TenantState::Stopping { .. }),
1255 0 : "the attach task owns the tenant state until activation is complete"
1256 : );
1257 :
1258 0 : *state = TenantState::broken_from_reason(err.to_string());
1259 0 : });
1260 0 : };
1261 :
1262 : // TODO: should also be rejecting tenant conf changes that violate this check.
1263 0 : if let Err(e) = crate::tenant::storage_layer::inmemory_layer::IndexEntry::validate_checkpoint_distance(tenant_clone.get_checkpoint_distance()) {
1264 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1265 0 : return Ok(());
1266 0 : }
1267 0 :
1268 0 : let mut init_order = init_order;
1269 0 : // take the completion because initial tenant loading will complete when all of
1270 0 : // these tasks complete.
1271 0 : let _completion = init_order
1272 0 : .as_mut()
1273 0 : .and_then(|x| x.initial_tenant_load.take());
1274 0 : let remote_load_completion = init_order
1275 0 : .as_mut()
1276 0 : .and_then(|x| x.initial_tenant_load_remote.take());
1277 :
1278 : enum AttachType<'a> {
1279 : /// We are attaching this tenant lazily in the background.
1280 : Warmup {
1281 : _permit: tokio::sync::SemaphorePermit<'a>,
1282 : during_startup: bool
1283 : },
1284 : /// We are attaching this tenant as soon as we can, because for example an
1285 : /// endpoint tried to access it.
1286 : OnDemand,
1287 : /// During normal operations after startup, we are attaching a tenant, and
1288 : /// eager attach was requested.
1289 : Normal,
1290 : }
1291 :
1292 0 : let attach_type = if matches!(mode, SpawnMode::Lazy) {
1293 : // Before doing any I/O, wait for at least one of:
1294 : // - A client attempting to access to this tenant (on-demand loading)
1295 : // - A permit becoming available in the warmup semaphore (background warmup)
1296 :
1297 0 : tokio::select!(
1298 0 : permit = tenant_clone.activate_now_sem.acquire() => {
1299 0 : let _ = permit.expect("activate_now_sem is never closed");
1300 0 : tracing::info!("Activating tenant (on-demand)");
1301 0 : AttachType::OnDemand
1302 : },
1303 0 : permit = conf.concurrent_tenant_warmup.inner().acquire() => {
1304 0 : let _permit = permit.expect("concurrent_tenant_warmup semaphore is never closed");
1305 0 : tracing::info!("Activating tenant (warmup)");
1306 0 : AttachType::Warmup {
1307 0 : _permit,
1308 0 : during_startup: init_order.is_some()
1309 0 : }
1310 : }
1311 0 : _ = tenant_clone.cancel.cancelled() => {
1312 : // This is safe, but should be pretty rare: it is interesting if a tenant
1313 : // stayed in Activating for such a long time that shutdown found it in
1314 : // that state.
1315 0 : tracing::info!(state=%tenant_clone.current_state(), "Tenant shut down before activation");
1316 : // Make the tenant broken so that set_stopping will not hang waiting for it to leave
1317 : // the Attaching state. This is an over-reaction (nothing really broke, the tenant is
1318 : // just shutting down), but ensures progress.
1319 0 : make_broken(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"), BrokenVerbosity::Info);
1320 0 : return Ok(());
1321 : },
1322 : )
1323 : } else {
1324 : // SpawnMode::{Create,Eager} always cause jumping ahead of the
1325 : // concurrent_tenant_warmup queue
1326 0 : AttachType::Normal
1327 : };
1328 :
1329 0 : let preload = match &mode {
1330 : SpawnMode::Eager | SpawnMode::Lazy => {
1331 0 : let _preload_timer = TENANT.preload.start_timer();
1332 0 : let res = tenant_clone
1333 0 : .preload(&remote_storage, task_mgr::shutdown_token())
1334 0 : .await;
1335 0 : match res {
1336 0 : Ok(p) => Some(p),
1337 0 : Err(e) => {
1338 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1339 0 : return Ok(());
1340 : }
1341 : }
1342 : }
1343 :
1344 : };
1345 :
1346 : // Remote preload is complete.
1347 0 : drop(remote_load_completion);
1348 0 :
1349 0 :
1350 0 : // We will time the duration of the attach phase unless this is a creation (attach will do no work)
1351 0 : let attach_start = std::time::Instant::now();
1352 0 : let attached = {
1353 0 : let _attach_timer = Some(TENANT.attach.start_timer());
1354 0 : tenant_clone.attach(preload, &ctx).await
1355 : };
1356 0 : let attach_duration = attach_start.elapsed();
1357 0 : _ = tenant_clone.attach_wal_lag_cooldown.set(WalLagCooldown::new(attach_start, attach_duration));
1358 0 :
1359 0 : match attached {
1360 : Ok(()) => {
1361 0 : info!("attach finished, activating");
1362 0 : tenant_clone.activate(broker_client, None, &ctx);
1363 : }
1364 0 : Err(e) => {
1365 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1366 0 : }
1367 : }
1368 :
1369 : // If we are doing an opportunistic warmup attachment at startup, initialize
1370 : // logical size at the same time. This is better than starting a bunch of idle tenants
1371 : // with cold caches and then coming back later to initialize their logical sizes.
1372 : //
1373 : // It also prevents the warmup proccess competing with the concurrency limit on
1374 : // logical size calculations: if logical size calculation semaphore is saturated,
1375 : // then warmup will wait for that before proceeding to the next tenant.
1376 0 : if matches!(attach_type, AttachType::Warmup { during_startup: true, .. }) {
1377 0 : let mut futs: FuturesUnordered<_> = tenant_clone.timelines.lock().unwrap().values().cloned().map(|t| t.await_initial_logical_size()).collect();
1378 0 : tracing::info!("Waiting for initial logical sizes while warming up...");
1379 0 : while futs.next().await.is_some() {}
1380 0 : tracing::info!("Warm-up complete");
1381 0 : }
1382 :
1383 0 : Ok(())
1384 0 : }
1385 0 : .instrument(tracing::info_span!(parent: None, "attach", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), gen=?generation)),
1386 : );
1387 0 : Ok(tenant)
1388 0 : }
1389 :
1390 192 : #[instrument(skip_all)]
1391 : pub(crate) async fn preload(
1392 : self: &Arc<Self>,
1393 : remote_storage: &GenericRemoteStorage,
1394 : cancel: CancellationToken,
1395 : ) -> anyhow::Result<TenantPreload> {
1396 : span::debug_assert_current_span_has_tenant_id();
1397 : // Get list of remote timelines
1398 : // download index files for every tenant timeline
1399 : info!("listing remote timelines");
1400 : let (remote_timeline_ids, other_keys) = remote_timeline_client::list_remote_timelines(
1401 : remote_storage,
1402 : self.tenant_shard_id,
1403 : cancel.clone(),
1404 : )
1405 : .await?;
1406 : let (offloaded_add, tenant_manifest) =
1407 : match remote_timeline_client::download_tenant_manifest(
1408 : remote_storage,
1409 : &self.tenant_shard_id,
1410 : self.generation,
1411 : &cancel,
1412 : )
1413 : .await
1414 : {
1415 : Ok((tenant_manifest, _generation, _manifest_mtime)) => (
1416 : format!("{} offloaded", tenant_manifest.offloaded_timelines.len()),
1417 : tenant_manifest,
1418 : ),
1419 : Err(DownloadError::NotFound) => {
1420 : ("no manifest".to_string(), TenantManifest::empty())
1421 : }
1422 : Err(e) => Err(e)?,
1423 : };
1424 :
1425 : info!(
1426 : "found {} timelines, and {offloaded_add}",
1427 : remote_timeline_ids.len()
1428 : );
1429 :
1430 : for k in other_keys {
1431 : warn!("Unexpected non timeline key {k}");
1432 : }
1433 :
1434 : Ok(TenantPreload {
1435 : tenant_manifest,
1436 : timelines: self
1437 : .load_timelines_metadata(remote_timeline_ids, remote_storage, cancel)
1438 : .await?,
1439 : })
1440 : }
1441 :
1442 : ///
1443 : /// Background task that downloads all data for a tenant and brings it to Active state.
1444 : ///
1445 : /// No background tasks are started as part of this routine.
1446 : ///
1447 192 : async fn attach(
1448 192 : self: &Arc<Tenant>,
1449 192 : preload: Option<TenantPreload>,
1450 192 : ctx: &RequestContext,
1451 192 : ) -> anyhow::Result<()> {
1452 192 : span::debug_assert_current_span_has_tenant_id();
1453 192 :
1454 192 : failpoint_support::sleep_millis_async!("before-attaching-tenant");
1455 :
1456 192 : let Some(preload) = preload else {
1457 0 : anyhow::bail!("local-only deployment is no longer supported, https://github.com/neondatabase/neon/issues/5624");
1458 : };
1459 :
1460 192 : let mut offloaded_timeline_ids = HashSet::new();
1461 192 : let mut offloaded_timelines_list = Vec::new();
1462 192 : for timeline_manifest in preload.tenant_manifest.offloaded_timelines.iter() {
1463 0 : let timeline_id = timeline_manifest.timeline_id;
1464 0 : let offloaded_timeline =
1465 0 : OffloadedTimeline::from_manifest(self.tenant_shard_id, timeline_manifest);
1466 0 : offloaded_timelines_list.push((timeline_id, Arc::new(offloaded_timeline)));
1467 0 : offloaded_timeline_ids.insert(timeline_id);
1468 0 : }
1469 :
1470 192 : let mut timelines_to_resume_deletions = vec![];
1471 192 :
1472 192 : let mut remote_index_and_client = HashMap::new();
1473 192 : let mut timeline_ancestors = HashMap::new();
1474 192 : let mut existent_timelines = HashSet::new();
1475 198 : for (timeline_id, preload) in preload.timelines {
1476 6 : if offloaded_timeline_ids.remove(&timeline_id) {
1477 : // The timeline is offloaded, skip loading it.
1478 0 : continue;
1479 6 : }
1480 6 : let index_part = match preload.index_part {
1481 6 : Ok(i) => {
1482 6 : debug!("remote index part exists for timeline {timeline_id}");
1483 : // We found index_part on the remote, this is the standard case.
1484 6 : existent_timelines.insert(timeline_id);
1485 6 : i
1486 : }
1487 : Err(DownloadError::NotFound) => {
1488 : // There is no index_part on the remote. We only get here
1489 : // if there is some prefix for the timeline in the remote storage.
1490 : // This can e.g. be the initdb.tar.zst archive, maybe a
1491 : // remnant from a prior incomplete creation or deletion attempt.
1492 : // Delete the local directory as the deciding criterion for a
1493 : // timeline's existence is presence of index_part.
1494 0 : info!(%timeline_id, "index_part not found on remote");
1495 0 : continue;
1496 : }
1497 0 : Err(DownloadError::Fatal(why)) => {
1498 0 : // If, while loading one remote timeline, we saw an indication that our generation
1499 0 : // number is likely invalid, then we should not load the whole tenant.
1500 0 : error!(%timeline_id, "Fatal error loading timeline: {why}");
1501 0 : anyhow::bail!(why.to_string());
1502 : }
1503 0 : Err(e) => {
1504 0 : // Some (possibly ephemeral) error happened during index_part download.
1505 0 : // Pretend the timeline exists to not delete the timeline directory,
1506 0 : // as it might be a temporary issue and we don't want to re-download
1507 0 : // everything after it resolves.
1508 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1509 :
1510 0 : existent_timelines.insert(timeline_id);
1511 0 : continue;
1512 : }
1513 : };
1514 6 : match index_part {
1515 6 : MaybeDeletedIndexPart::IndexPart(index_part) => {
1516 6 : timeline_ancestors.insert(timeline_id, index_part.metadata.clone());
1517 6 : remote_index_and_client.insert(timeline_id, (index_part, preload.client));
1518 6 : }
1519 0 : MaybeDeletedIndexPart::Deleted(index_part) => {
1520 0 : info!(
1521 0 : "timeline {} is deleted, picking to resume deletion",
1522 : timeline_id
1523 : );
1524 0 : timelines_to_resume_deletions.push((timeline_id, index_part, preload.client));
1525 : }
1526 : }
1527 : }
1528 :
1529 192 : let mut gc_blocks = HashMap::new();
1530 :
1531 : // For every timeline, download the metadata file, scan the local directory,
1532 : // and build a layer map that contains an entry for each remote and local
1533 : // layer file.
1534 192 : let sorted_timelines = tree_sort_timelines(timeline_ancestors, |m| m.ancestor_timeline())?;
1535 198 : for (timeline_id, remote_metadata) in sorted_timelines {
1536 6 : let (index_part, remote_client) = remote_index_and_client
1537 6 : .remove(&timeline_id)
1538 6 : .expect("just put it in above");
1539 :
1540 6 : if let Some(blocking) = index_part.gc_blocking.as_ref() {
1541 : // could just filter these away, but it helps while testing
1542 0 : anyhow::ensure!(
1543 0 : !blocking.reasons.is_empty(),
1544 0 : "index_part for {timeline_id} is malformed: it should not have gc blocking with zero reasons"
1545 : );
1546 0 : let prev = gc_blocks.insert(timeline_id, blocking.reasons);
1547 0 : assert!(prev.is_none());
1548 6 : }
1549 :
1550 : // TODO again handle early failure
1551 6 : self.load_remote_timeline(
1552 6 : timeline_id,
1553 6 : index_part,
1554 6 : remote_metadata,
1555 6 : TimelineResources {
1556 6 : remote_client,
1557 6 : timeline_get_throttle: self.timeline_get_throttle.clone(),
1558 6 : l0_flush_global_state: self.l0_flush_global_state.clone(),
1559 6 : },
1560 6 : ctx,
1561 6 : )
1562 5 : .await
1563 6 : .with_context(|| {
1564 0 : format!(
1565 0 : "failed to load remote timeline {} for tenant {}",
1566 0 : timeline_id, self.tenant_shard_id
1567 0 : )
1568 6 : })?;
1569 : }
1570 :
1571 : // Walk through deleted timelines, resume deletion
1572 192 : for (timeline_id, index_part, remote_timeline_client) in timelines_to_resume_deletions {
1573 0 : remote_timeline_client
1574 0 : .init_upload_queue_stopped_to_continue_deletion(&index_part)
1575 0 : .context("init queue stopped")
1576 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1577 :
1578 0 : DeleteTimelineFlow::resume_deletion(
1579 0 : Arc::clone(self),
1580 0 : timeline_id,
1581 0 : &index_part.metadata,
1582 0 : remote_timeline_client,
1583 0 : )
1584 0 : .instrument(tracing::info_span!("timeline_delete", %timeline_id))
1585 0 : .await
1586 0 : .context("resume_deletion")
1587 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1588 : }
1589 : // Complete deletions for offloaded timeline id's.
1590 192 : offloaded_timelines_list
1591 192 : .retain(|(offloaded_id, offloaded)| {
1592 0 : // At this point, offloaded_timeline_ids has the list of all offloaded timelines
1593 0 : // without a prefix in S3, so they are inexistent.
1594 0 : // In the end, existence of a timeline is finally determined by the existence of an index-part.json in remote storage.
1595 0 : // If there is a dangling reference in another location, they need to be cleaned up.
1596 0 : let delete = offloaded_timeline_ids.contains(offloaded_id);
1597 0 : if delete {
1598 0 : tracing::info!("Removing offloaded timeline {offloaded_id} from manifest as no remote prefix was found");
1599 0 : offloaded.defuse_for_tenant_drop();
1600 0 : }
1601 0 : !delete
1602 192 : });
1603 192 : if !offloaded_timelines_list.is_empty() {
1604 0 : tracing::info!(
1605 0 : "Tenant has {} offloaded timelines",
1606 0 : offloaded_timelines_list.len()
1607 : );
1608 192 : }
1609 192 : {
1610 192 : let mut offloaded_timelines_accessor = self.timelines_offloaded.lock().unwrap();
1611 192 : offloaded_timelines_accessor.extend(offloaded_timelines_list.into_iter());
1612 192 : }
1613 192 : if !offloaded_timeline_ids.is_empty() {
1614 0 : self.store_tenant_manifest().await?;
1615 192 : }
1616 :
1617 : // The local filesystem contents are a cache of what's in the remote IndexPart;
1618 : // IndexPart is the source of truth.
1619 192 : self.clean_up_timelines(&existent_timelines)?;
1620 :
1621 192 : self.gc_block.set_scanned(gc_blocks);
1622 192 :
1623 192 : fail::fail_point!("attach-before-activate", |_| {
1624 0 : anyhow::bail!("attach-before-activate");
1625 192 : });
1626 192 : failpoint_support::sleep_millis_async!("attach-before-activate-sleep", &self.cancel);
1627 :
1628 192 : info!("Done");
1629 :
1630 192 : Ok(())
1631 192 : }
1632 :
1633 : /// Check for any local timeline directories that are temporary, or do not correspond to a
1634 : /// timeline that still exists: this can happen if we crashed during a deletion/creation, or
1635 : /// if a timeline was deleted while the tenant was attached to a different pageserver.
1636 192 : fn clean_up_timelines(&self, existent_timelines: &HashSet<TimelineId>) -> anyhow::Result<()> {
1637 192 : let timelines_dir = self.conf.timelines_path(&self.tenant_shard_id);
1638 :
1639 192 : let entries = match timelines_dir.read_dir_utf8() {
1640 192 : Ok(d) => d,
1641 0 : Err(e) => {
1642 0 : if e.kind() == std::io::ErrorKind::NotFound {
1643 0 : return Ok(());
1644 : } else {
1645 0 : return Err(e).context("list timelines directory for tenant");
1646 : }
1647 : }
1648 : };
1649 :
1650 200 : for entry in entries {
1651 8 : let entry = entry.context("read timeline dir entry")?;
1652 8 : let entry_path = entry.path();
1653 :
1654 8 : let purge = if crate::is_temporary(entry_path)
1655 : // TODO: remove uninit mark code (https://github.com/neondatabase/neon/issues/5718)
1656 8 : || is_uninit_mark(entry_path)
1657 8 : || crate::is_delete_mark(entry_path)
1658 : {
1659 0 : true
1660 : } else {
1661 8 : match TimelineId::try_from(entry_path.file_name()) {
1662 8 : Ok(i) => {
1663 8 : // Purge if the timeline ID does not exist in remote storage: remote storage is the authority.
1664 8 : !existent_timelines.contains(&i)
1665 : }
1666 0 : Err(e) => {
1667 0 : tracing::warn!(
1668 0 : "Unparseable directory in timelines directory: {entry_path}, ignoring ({e})"
1669 : );
1670 : // Do not purge junk: if we don't recognize it, be cautious and leave it for a human.
1671 0 : false
1672 : }
1673 : }
1674 : };
1675 :
1676 8 : if purge {
1677 2 : tracing::info!("Purging stale timeline dentry {entry_path}");
1678 2 : if let Err(e) = match entry.file_type() {
1679 2 : Ok(t) => if t.is_dir() {
1680 2 : std::fs::remove_dir_all(entry_path)
1681 : } else {
1682 0 : std::fs::remove_file(entry_path)
1683 : }
1684 2 : .or_else(fs_ext::ignore_not_found),
1685 0 : Err(e) => Err(e),
1686 : } {
1687 0 : tracing::warn!("Failed to purge stale timeline dentry {entry_path}: {e}");
1688 2 : }
1689 6 : }
1690 : }
1691 :
1692 192 : Ok(())
1693 192 : }
1694 :
1695 : /// Get sum of all remote timelines sizes
1696 : ///
1697 : /// This function relies on the index_part instead of listing the remote storage
1698 0 : pub fn remote_size(&self) -> u64 {
1699 0 : let mut size = 0;
1700 :
1701 0 : for timeline in self.list_timelines() {
1702 0 : size += timeline.remote_client.get_remote_physical_size();
1703 0 : }
1704 :
1705 0 : size
1706 0 : }
1707 :
1708 6 : #[instrument(skip_all, fields(timeline_id=%timeline_id))]
1709 : async fn load_remote_timeline(
1710 : &self,
1711 : timeline_id: TimelineId,
1712 : index_part: IndexPart,
1713 : remote_metadata: TimelineMetadata,
1714 : resources: TimelineResources,
1715 : ctx: &RequestContext,
1716 : ) -> anyhow::Result<()> {
1717 : span::debug_assert_current_span_has_tenant_id();
1718 :
1719 : info!("downloading index file for timeline {}", timeline_id);
1720 : tokio::fs::create_dir_all(self.conf.timeline_path(&self.tenant_shard_id, &timeline_id))
1721 : .await
1722 : .context("Failed to create new timeline directory")?;
1723 :
1724 : let ancestor = if let Some(ancestor_id) = remote_metadata.ancestor_timeline() {
1725 : let timelines = self.timelines.lock().unwrap();
1726 : Some(Arc::clone(timelines.get(&ancestor_id).ok_or_else(
1727 0 : || {
1728 0 : anyhow::anyhow!(
1729 0 : "cannot find ancestor timeline {ancestor_id} for timeline {timeline_id}"
1730 0 : )
1731 0 : },
1732 : )?))
1733 : } else {
1734 : None
1735 : };
1736 :
1737 : self.timeline_init_and_sync(
1738 : timeline_id,
1739 : resources,
1740 : index_part,
1741 : remote_metadata,
1742 : ancestor,
1743 : ctx,
1744 : )
1745 : .await
1746 : }
1747 :
1748 192 : async fn load_timelines_metadata(
1749 192 : self: &Arc<Tenant>,
1750 192 : timeline_ids: HashSet<TimelineId>,
1751 192 : remote_storage: &GenericRemoteStorage,
1752 192 : cancel: CancellationToken,
1753 192 : ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
1754 192 : let mut part_downloads = JoinSet::new();
1755 198 : for timeline_id in timeline_ids {
1756 6 : let cancel_clone = cancel.clone();
1757 6 : part_downloads.spawn(
1758 6 : self.load_timeline_metadata(timeline_id, remote_storage.clone(), cancel_clone)
1759 6 : .instrument(info_span!("download_index_part", %timeline_id)),
1760 : );
1761 : }
1762 :
1763 192 : let mut timeline_preloads: HashMap<TimelineId, TimelinePreload> = HashMap::new();
1764 :
1765 : loop {
1766 198 : tokio::select!(
1767 198 : next = part_downloads.join_next() => {
1768 198 : match next {
1769 6 : Some(result) => {
1770 6 : let preload = result.context("join preload task")?;
1771 6 : timeline_preloads.insert(preload.timeline_id, preload);
1772 : },
1773 : None => {
1774 192 : break;
1775 : }
1776 : }
1777 : },
1778 198 : _ = cancel.cancelled() => {
1779 0 : anyhow::bail!("Cancelled while waiting for remote index download")
1780 : }
1781 : )
1782 : }
1783 :
1784 192 : Ok(timeline_preloads)
1785 192 : }
1786 :
1787 6 : fn build_timeline_client(
1788 6 : &self,
1789 6 : timeline_id: TimelineId,
1790 6 : remote_storage: GenericRemoteStorage,
1791 6 : ) -> RemoteTimelineClient {
1792 6 : RemoteTimelineClient::new(
1793 6 : remote_storage.clone(),
1794 6 : self.deletion_queue_client.clone(),
1795 6 : self.conf,
1796 6 : self.tenant_shard_id,
1797 6 : timeline_id,
1798 6 : self.generation,
1799 6 : )
1800 6 : }
1801 :
1802 6 : fn load_timeline_metadata(
1803 6 : self: &Arc<Tenant>,
1804 6 : timeline_id: TimelineId,
1805 6 : remote_storage: GenericRemoteStorage,
1806 6 : cancel: CancellationToken,
1807 6 : ) -> impl Future<Output = TimelinePreload> {
1808 6 : let client = self.build_timeline_client(timeline_id, remote_storage);
1809 6 : async move {
1810 6 : debug_assert_current_span_has_tenant_and_timeline_id();
1811 6 : debug!("starting index part download");
1812 :
1813 6 : let index_part = client.download_index_file(&cancel).await;
1814 :
1815 6 : debug!("finished index part download");
1816 :
1817 6 : TimelinePreload {
1818 6 : client,
1819 6 : timeline_id,
1820 6 : index_part,
1821 6 : }
1822 6 : }
1823 6 : }
1824 :
1825 0 : fn check_to_be_archived_has_no_unarchived_children(
1826 0 : timeline_id: TimelineId,
1827 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
1828 0 : ) -> Result<(), TimelineArchivalError> {
1829 0 : let children: Vec<TimelineId> = timelines
1830 0 : .iter()
1831 0 : .filter_map(|(id, entry)| {
1832 0 : if entry.get_ancestor_timeline_id() != Some(timeline_id) {
1833 0 : return None;
1834 0 : }
1835 0 : if entry.is_archived() == Some(true) {
1836 0 : return None;
1837 0 : }
1838 0 : Some(*id)
1839 0 : })
1840 0 : .collect();
1841 0 :
1842 0 : if !children.is_empty() {
1843 0 : return Err(TimelineArchivalError::HasUnarchivedChildren(children));
1844 0 : }
1845 0 : Ok(())
1846 0 : }
1847 :
1848 0 : fn check_ancestor_of_to_be_unarchived_is_not_archived(
1849 0 : ancestor_timeline_id: TimelineId,
1850 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
1851 0 : offloaded_timelines: &std::sync::MutexGuard<
1852 0 : '_,
1853 0 : HashMap<TimelineId, Arc<OffloadedTimeline>>,
1854 0 : >,
1855 0 : ) -> Result<(), TimelineArchivalError> {
1856 0 : let has_archived_parent =
1857 0 : if let Some(ancestor_timeline) = timelines.get(&ancestor_timeline_id) {
1858 0 : ancestor_timeline.is_archived() == Some(true)
1859 0 : } else if offloaded_timelines.contains_key(&ancestor_timeline_id) {
1860 0 : true
1861 : } else {
1862 0 : error!("ancestor timeline {ancestor_timeline_id} not found");
1863 0 : if cfg!(debug_assertions) {
1864 0 : panic!("ancestor timeline {ancestor_timeline_id} not found");
1865 0 : }
1866 0 : return Err(TimelineArchivalError::NotFound);
1867 : };
1868 0 : if has_archived_parent {
1869 0 : return Err(TimelineArchivalError::HasArchivedParent(
1870 0 : ancestor_timeline_id,
1871 0 : ));
1872 0 : }
1873 0 : Ok(())
1874 0 : }
1875 :
1876 0 : fn check_to_be_unarchived_timeline_has_no_archived_parent(
1877 0 : timeline: &Arc<Timeline>,
1878 0 : ) -> Result<(), TimelineArchivalError> {
1879 0 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
1880 0 : if ancestor_timeline.is_archived() == Some(true) {
1881 0 : return Err(TimelineArchivalError::HasArchivedParent(
1882 0 : ancestor_timeline.timeline_id,
1883 0 : ));
1884 0 : }
1885 0 : }
1886 0 : Ok(())
1887 0 : }
1888 :
1889 : /// Loads the specified (offloaded) timeline from S3 and attaches it as a loaded timeline
1890 : ///
1891 : /// Counterpart to [`offload_timeline`].
1892 0 : async fn unoffload_timeline(
1893 0 : self: &Arc<Self>,
1894 0 : timeline_id: TimelineId,
1895 0 : broker_client: storage_broker::BrokerClientChannel,
1896 0 : ctx: RequestContext,
1897 0 : ) -> Result<Arc<Timeline>, TimelineArchivalError> {
1898 0 : info!("unoffloading timeline");
1899 :
1900 : // We activate the timeline below manually, so this must be called on an active timeline.
1901 : // We expect callers of this function to ensure this.
1902 0 : match self.current_state() {
1903 : TenantState::Activating { .. }
1904 : | TenantState::Attaching
1905 : | TenantState::Broken { .. } => {
1906 0 : panic!("Timeline expected to be active")
1907 : }
1908 0 : TenantState::Stopping { .. } => return Err(TimelineArchivalError::Cancelled),
1909 0 : TenantState::Active => {}
1910 0 : }
1911 0 : let cancel = self.cancel.clone();
1912 0 :
1913 0 : // Protect against concurrent attempts to use this TimelineId
1914 0 : // We don't care much about idempotency, as it's ensured a layer above.
1915 0 : let allow_offloaded = true;
1916 0 : let _create_guard = self
1917 0 : .create_timeline_create_guard(
1918 0 : timeline_id,
1919 0 : CreateTimelineIdempotency::FailWithConflict,
1920 0 : allow_offloaded,
1921 0 : )
1922 0 : .map_err(|err| match err {
1923 0 : TimelineExclusionError::AlreadyCreating => TimelineArchivalError::AlreadyInProgress,
1924 : TimelineExclusionError::AlreadyExists { .. } => {
1925 0 : TimelineArchivalError::Other(anyhow::anyhow!("Timeline already exists"))
1926 : }
1927 0 : TimelineExclusionError::Other(e) => TimelineArchivalError::Other(e),
1928 0 : })?;
1929 :
1930 0 : let timeline_preload = self
1931 0 : .load_timeline_metadata(timeline_id, self.remote_storage.clone(), cancel.clone())
1932 0 : .await;
1933 :
1934 0 : let index_part = match timeline_preload.index_part {
1935 0 : Ok(index_part) => {
1936 0 : debug!("remote index part exists for timeline {timeline_id}");
1937 0 : index_part
1938 : }
1939 : Err(DownloadError::NotFound) => {
1940 0 : error!(%timeline_id, "index_part not found on remote");
1941 0 : return Err(TimelineArchivalError::NotFound);
1942 : }
1943 0 : Err(DownloadError::Cancelled) => return Err(TimelineArchivalError::Cancelled),
1944 0 : Err(e) => {
1945 0 : // Some (possibly ephemeral) error happened during index_part download.
1946 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1947 0 : return Err(TimelineArchivalError::Other(
1948 0 : anyhow::Error::new(e).context("downloading index_part from remote storage"),
1949 0 : ));
1950 : }
1951 : };
1952 0 : let index_part = match index_part {
1953 0 : MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
1954 0 : MaybeDeletedIndexPart::Deleted(_index_part) => {
1955 0 : info!("timeline is deleted according to index_part.json");
1956 0 : return Err(TimelineArchivalError::NotFound);
1957 : }
1958 : };
1959 0 : let remote_metadata = index_part.metadata.clone();
1960 0 : let timeline_resources = self.build_timeline_resources(timeline_id);
1961 0 : self.load_remote_timeline(
1962 0 : timeline_id,
1963 0 : index_part,
1964 0 : remote_metadata,
1965 0 : timeline_resources,
1966 0 : &ctx,
1967 0 : )
1968 0 : .await
1969 0 : .with_context(|| {
1970 0 : format!(
1971 0 : "failed to load remote timeline {} for tenant {}",
1972 0 : timeline_id, self.tenant_shard_id
1973 0 : )
1974 0 : })
1975 0 : .map_err(TimelineArchivalError::Other)?;
1976 :
1977 0 : let timeline = {
1978 0 : let timelines = self.timelines.lock().unwrap();
1979 0 : let Some(timeline) = timelines.get(&timeline_id) else {
1980 0 : warn!("timeline not available directly after attach");
1981 : // This is not a panic because no locks are held between `load_remote_timeline`
1982 : // which puts the timeline into timelines, and our look into the timeline map.
1983 0 : return Err(TimelineArchivalError::Other(anyhow::anyhow!(
1984 0 : "timeline not available directly after attach"
1985 0 : )));
1986 : };
1987 0 : let mut offloaded_timelines = self.timelines_offloaded.lock().unwrap();
1988 0 : match offloaded_timelines.remove(&timeline_id) {
1989 0 : Some(offloaded) => {
1990 0 : offloaded.delete_from_ancestor_with_timelines(&timelines);
1991 0 : }
1992 0 : None => warn!("timeline already removed from offloaded timelines"),
1993 : }
1994 :
1995 0 : self.initialize_gc_info(&timelines, &offloaded_timelines, Some(timeline_id));
1996 0 :
1997 0 : Arc::clone(timeline)
1998 0 : };
1999 0 :
2000 0 : // Upload new list of offloaded timelines to S3
2001 0 : self.store_tenant_manifest().await?;
2002 :
2003 : // Activate the timeline (if it makes sense)
2004 0 : if !(timeline.is_broken() || timeline.is_stopping()) {
2005 0 : let background_jobs_can_start = None;
2006 0 : timeline.activate(
2007 0 : self.clone(),
2008 0 : broker_client.clone(),
2009 0 : background_jobs_can_start,
2010 0 : &ctx,
2011 0 : );
2012 0 : }
2013 :
2014 0 : info!("timeline unoffloading complete");
2015 0 : Ok(timeline)
2016 0 : }
2017 :
2018 0 : pub(crate) async fn apply_timeline_archival_config(
2019 0 : self: &Arc<Self>,
2020 0 : timeline_id: TimelineId,
2021 0 : new_state: TimelineArchivalState,
2022 0 : broker_client: storage_broker::BrokerClientChannel,
2023 0 : ctx: RequestContext,
2024 0 : ) -> Result<(), TimelineArchivalError> {
2025 0 : info!("setting timeline archival config");
2026 : // First part: figure out what is needed to do, and do validation
2027 0 : let timeline_or_unarchive_offloaded = 'outer: {
2028 0 : let timelines = self.timelines.lock().unwrap();
2029 :
2030 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2031 0 : let offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2032 0 : let Some(offloaded) = offloaded_timelines.get(&timeline_id) else {
2033 0 : return Err(TimelineArchivalError::NotFound);
2034 : };
2035 0 : if new_state == TimelineArchivalState::Archived {
2036 : // It's offloaded already, so nothing to do
2037 0 : return Ok(());
2038 0 : }
2039 0 : if let Some(ancestor_timeline_id) = offloaded.ancestor_timeline_id {
2040 0 : Self::check_ancestor_of_to_be_unarchived_is_not_archived(
2041 0 : ancestor_timeline_id,
2042 0 : &timelines,
2043 0 : &offloaded_timelines,
2044 0 : )?;
2045 0 : }
2046 0 : break 'outer None;
2047 : };
2048 :
2049 : // Do some validation. We release the timelines lock below, so there is potential
2050 : // for race conditions: these checks are more present to prevent misunderstandings of
2051 : // the API's capabilities, instead of serving as the sole way to defend their invariants.
2052 0 : match new_state {
2053 : TimelineArchivalState::Unarchived => {
2054 0 : Self::check_to_be_unarchived_timeline_has_no_archived_parent(timeline)?
2055 : }
2056 : TimelineArchivalState::Archived => {
2057 0 : Self::check_to_be_archived_has_no_unarchived_children(timeline_id, &timelines)?
2058 : }
2059 : }
2060 0 : Some(Arc::clone(timeline))
2061 : };
2062 :
2063 : // Second part: unoffload timeline (if needed)
2064 0 : let timeline = if let Some(timeline) = timeline_or_unarchive_offloaded {
2065 0 : timeline
2066 : } else {
2067 : // Turn offloaded timeline into a non-offloaded one
2068 0 : self.unoffload_timeline(timeline_id, broker_client, ctx)
2069 0 : .await?
2070 : };
2071 :
2072 : // Third part: upload new timeline archival state and block until it is present in S3
2073 0 : let upload_needed = match timeline
2074 0 : .remote_client
2075 0 : .schedule_index_upload_for_timeline_archival_state(new_state)
2076 : {
2077 0 : Ok(upload_needed) => upload_needed,
2078 0 : Err(e) => {
2079 0 : if timeline.cancel.is_cancelled() {
2080 0 : return Err(TimelineArchivalError::Cancelled);
2081 : } else {
2082 0 : return Err(TimelineArchivalError::Other(e));
2083 : }
2084 : }
2085 : };
2086 :
2087 0 : if upload_needed {
2088 0 : info!("Uploading new state");
2089 : const MAX_WAIT: Duration = Duration::from_secs(10);
2090 0 : let Ok(v) =
2091 0 : tokio::time::timeout(MAX_WAIT, timeline.remote_client.wait_completion()).await
2092 : else {
2093 0 : tracing::warn!("reached timeout for waiting on upload queue");
2094 0 : return Err(TimelineArchivalError::Timeout);
2095 : };
2096 0 : v.map_err(|e| match e {
2097 0 : WaitCompletionError::NotInitialized(e) => {
2098 0 : TimelineArchivalError::Other(anyhow::anyhow!(e))
2099 : }
2100 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2101 0 : TimelineArchivalError::Cancelled
2102 : }
2103 0 : })?;
2104 0 : }
2105 0 : Ok(())
2106 0 : }
2107 :
2108 2 : pub fn get_offloaded_timeline(
2109 2 : &self,
2110 2 : timeline_id: TimelineId,
2111 2 : ) -> Result<Arc<OffloadedTimeline>, GetTimelineError> {
2112 2 : self.timelines_offloaded
2113 2 : .lock()
2114 2 : .unwrap()
2115 2 : .get(&timeline_id)
2116 2 : .map(Arc::clone)
2117 2 : .ok_or(GetTimelineError::NotFound {
2118 2 : tenant_id: self.tenant_shard_id,
2119 2 : timeline_id,
2120 2 : })
2121 2 : }
2122 :
2123 4 : pub(crate) fn tenant_shard_id(&self) -> TenantShardId {
2124 4 : self.tenant_shard_id
2125 4 : }
2126 :
2127 : /// Get Timeline handle for given Neon timeline ID.
2128 : /// This function is idempotent. It doesn't change internal state in any way.
2129 222 : pub fn get_timeline(
2130 222 : &self,
2131 222 : timeline_id: TimelineId,
2132 222 : active_only: bool,
2133 222 : ) -> Result<Arc<Timeline>, GetTimelineError> {
2134 222 : let timelines_accessor = self.timelines.lock().unwrap();
2135 222 : let timeline = timelines_accessor
2136 222 : .get(&timeline_id)
2137 222 : .ok_or(GetTimelineError::NotFound {
2138 222 : tenant_id: self.tenant_shard_id,
2139 222 : timeline_id,
2140 222 : })?;
2141 :
2142 220 : if active_only && !timeline.is_active() {
2143 0 : Err(GetTimelineError::NotActive {
2144 0 : tenant_id: self.tenant_shard_id,
2145 0 : timeline_id,
2146 0 : state: timeline.current_state(),
2147 0 : })
2148 : } else {
2149 220 : Ok(Arc::clone(timeline))
2150 : }
2151 222 : }
2152 :
2153 : /// Lists timelines the tenant contains.
2154 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2155 0 : pub fn list_timelines(&self) -> Vec<Arc<Timeline>> {
2156 0 : self.timelines
2157 0 : .lock()
2158 0 : .unwrap()
2159 0 : .values()
2160 0 : .map(Arc::clone)
2161 0 : .collect()
2162 0 : }
2163 :
2164 : /// Lists timelines the tenant manages, including offloaded ones.
2165 : ///
2166 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2167 0 : pub fn list_timelines_and_offloaded(
2168 0 : &self,
2169 0 : ) -> (Vec<Arc<Timeline>>, Vec<Arc<OffloadedTimeline>>) {
2170 0 : let timelines = self
2171 0 : .timelines
2172 0 : .lock()
2173 0 : .unwrap()
2174 0 : .values()
2175 0 : .map(Arc::clone)
2176 0 : .collect();
2177 0 : let offloaded = self
2178 0 : .timelines_offloaded
2179 0 : .lock()
2180 0 : .unwrap()
2181 0 : .values()
2182 0 : .map(Arc::clone)
2183 0 : .collect();
2184 0 : (timelines, offloaded)
2185 0 : }
2186 :
2187 0 : pub fn list_timeline_ids(&self) -> Vec<TimelineId> {
2188 0 : self.timelines.lock().unwrap().keys().cloned().collect()
2189 0 : }
2190 :
2191 : /// This is used by tests & import-from-basebackup.
2192 : ///
2193 : /// The returned [`UninitializedTimeline`] contains no data nor metadata and it is in
2194 : /// a state that will fail [`Tenant::load_remote_timeline`] because `disk_consistent_lsn=Lsn(0)`.
2195 : ///
2196 : /// The caller is responsible for getting the timeline into a state that will be accepted
2197 : /// by [`Tenant::load_remote_timeline`] / [`Tenant::attach`].
2198 : /// Then they may call [`UninitializedTimeline::finish_creation`] to add the timeline
2199 : /// to the [`Tenant::timelines`].
2200 : ///
2201 : /// Tests should use `Tenant::create_test_timeline` to set up the minimum required metadata keys.
2202 184 : pub(crate) async fn create_empty_timeline(
2203 184 : &self,
2204 184 : new_timeline_id: TimelineId,
2205 184 : initdb_lsn: Lsn,
2206 184 : pg_version: u32,
2207 184 : _ctx: &RequestContext,
2208 184 : ) -> anyhow::Result<UninitializedTimeline> {
2209 184 : anyhow::ensure!(
2210 184 : self.is_active(),
2211 0 : "Cannot create empty timelines on inactive tenant"
2212 : );
2213 :
2214 : // Protect against concurrent attempts to use this TimelineId
2215 184 : let create_guard = match self
2216 184 : .start_creating_timeline(new_timeline_id, CreateTimelineIdempotency::FailWithConflict)
2217 173 : .await?
2218 : {
2219 182 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2220 : StartCreatingTimelineResult::Idempotent(_) => {
2221 0 : unreachable!("FailWithConflict implies we get an error instead")
2222 : }
2223 : };
2224 :
2225 182 : let new_metadata = TimelineMetadata::new(
2226 182 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2227 182 : // make it valid, before calling finish_creation()
2228 182 : Lsn(0),
2229 182 : None,
2230 182 : None,
2231 182 : Lsn(0),
2232 182 : initdb_lsn,
2233 182 : initdb_lsn,
2234 182 : pg_version,
2235 182 : );
2236 182 : self.prepare_new_timeline(
2237 182 : new_timeline_id,
2238 182 : &new_metadata,
2239 182 : create_guard,
2240 182 : initdb_lsn,
2241 182 : None,
2242 182 : )
2243 0 : .await
2244 184 : }
2245 :
2246 : /// Helper for unit tests to create an empty timeline.
2247 : ///
2248 : /// The timeline is has state value `Active` but its background loops are not running.
2249 : // This makes the various functions which anyhow::ensure! for Active state work in tests.
2250 : // Our current tests don't need the background loops.
2251 : #[cfg(test)]
2252 174 : pub async fn create_test_timeline(
2253 174 : &self,
2254 174 : new_timeline_id: TimelineId,
2255 174 : initdb_lsn: Lsn,
2256 174 : pg_version: u32,
2257 174 : ctx: &RequestContext,
2258 174 : ) -> anyhow::Result<Arc<Timeline>> {
2259 174 : let uninit_tl = self
2260 174 : .create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2261 165 : .await?;
2262 174 : let tline = uninit_tl.raw_timeline().expect("we just created it");
2263 174 : assert_eq!(tline.get_last_record_lsn(), Lsn(0));
2264 :
2265 : // Setup minimum keys required for the timeline to be usable.
2266 174 : let mut modification = tline.begin_modification(initdb_lsn);
2267 174 : modification
2268 174 : .init_empty_test_timeline()
2269 174 : .context("init_empty_test_timeline")?;
2270 174 : modification
2271 174 : .commit(ctx)
2272 168 : .await
2273 174 : .context("commit init_empty_test_timeline modification")?;
2274 :
2275 : // Flush to disk so that uninit_tl's check for valid disk_consistent_lsn passes.
2276 174 : tline.maybe_spawn_flush_loop();
2277 174 : tline.freeze_and_flush().await.context("freeze_and_flush")?;
2278 :
2279 : // Make sure the freeze_and_flush reaches remote storage.
2280 174 : tline.remote_client.wait_completion().await.unwrap();
2281 :
2282 174 : let tl = uninit_tl.finish_creation()?;
2283 : // The non-test code would call tl.activate() here.
2284 174 : tl.set_state(TimelineState::Active);
2285 174 : Ok(tl)
2286 174 : }
2287 :
2288 : /// Helper for unit tests to create a timeline with some pre-loaded states.
2289 : #[cfg(test)]
2290 : #[allow(clippy::too_many_arguments)]
2291 32 : pub async fn create_test_timeline_with_layers(
2292 32 : &self,
2293 32 : new_timeline_id: TimelineId,
2294 32 : initdb_lsn: Lsn,
2295 32 : pg_version: u32,
2296 32 : ctx: &RequestContext,
2297 32 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
2298 32 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
2299 32 : end_lsn: Lsn,
2300 32 : ) -> anyhow::Result<Arc<Timeline>> {
2301 : use checks::check_valid_layermap;
2302 : use itertools::Itertools;
2303 :
2304 32 : let tline = self
2305 32 : .create_test_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2306 95 : .await?;
2307 32 : tline.force_advance_lsn(end_lsn);
2308 100 : for deltas in delta_layer_desc {
2309 68 : tline
2310 68 : .force_create_delta_layer(deltas, Some(initdb_lsn), ctx)
2311 204 : .await?;
2312 : }
2313 80 : for (lsn, images) in image_layer_desc {
2314 48 : tline
2315 48 : .force_create_image_layer(lsn, images, Some(initdb_lsn), ctx)
2316 294 : .await?;
2317 : }
2318 32 : let layer_names = tline
2319 32 : .layers
2320 32 : .read()
2321 0 : .await
2322 32 : .layer_map()
2323 32 : .unwrap()
2324 32 : .iter_historic_layers()
2325 148 : .map(|layer| layer.layer_name())
2326 32 : .collect_vec();
2327 32 : if let Some(err) = check_valid_layermap(&layer_names) {
2328 0 : bail!("invalid layermap: {err}");
2329 32 : }
2330 32 : Ok(tline)
2331 32 : }
2332 :
2333 : /// Create a new timeline.
2334 : ///
2335 : /// Returns the new timeline ID and reference to its Timeline object.
2336 : ///
2337 : /// If the caller specified the timeline ID to use (`new_timeline_id`), and timeline with
2338 : /// the same timeline ID already exists, returns CreateTimelineError::AlreadyExists.
2339 : #[allow(clippy::too_many_arguments)]
2340 0 : pub(crate) async fn create_timeline(
2341 0 : self: &Arc<Tenant>,
2342 0 : params: CreateTimelineParams,
2343 0 : broker_client: storage_broker::BrokerClientChannel,
2344 0 : ctx: &RequestContext,
2345 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
2346 0 : if !self.is_active() {
2347 0 : if matches!(self.current_state(), TenantState::Stopping { .. }) {
2348 0 : return Err(CreateTimelineError::ShuttingDown);
2349 : } else {
2350 0 : return Err(CreateTimelineError::Other(anyhow::anyhow!(
2351 0 : "Cannot create timelines on inactive tenant"
2352 0 : )));
2353 : }
2354 0 : }
2355 :
2356 0 : let _gate = self
2357 0 : .gate
2358 0 : .enter()
2359 0 : .map_err(|_| CreateTimelineError::ShuttingDown)?;
2360 :
2361 0 : let result: CreateTimelineResult = match params {
2362 : CreateTimelineParams::Bootstrap(CreateTimelineParamsBootstrap {
2363 0 : new_timeline_id,
2364 0 : existing_initdb_timeline_id,
2365 0 : pg_version,
2366 0 : }) => {
2367 0 : self.bootstrap_timeline(
2368 0 : new_timeline_id,
2369 0 : pg_version,
2370 0 : existing_initdb_timeline_id,
2371 0 : ctx,
2372 0 : )
2373 0 : .await?
2374 : }
2375 : CreateTimelineParams::Branch(CreateTimelineParamsBranch {
2376 0 : new_timeline_id,
2377 0 : ancestor_timeline_id,
2378 0 : mut ancestor_start_lsn,
2379 : }) => {
2380 0 : let ancestor_timeline = self
2381 0 : .get_timeline(ancestor_timeline_id, false)
2382 0 : .context("Cannot branch off the timeline that's not present in pageserver")?;
2383 :
2384 : // instead of waiting around, just deny the request because ancestor is not yet
2385 : // ready for other purposes either.
2386 0 : if !ancestor_timeline.is_active() {
2387 0 : return Err(CreateTimelineError::AncestorNotActive);
2388 0 : }
2389 0 :
2390 0 : if ancestor_timeline.is_archived() == Some(true) {
2391 0 : info!("tried to branch archived timeline");
2392 0 : return Err(CreateTimelineError::AncestorArchived);
2393 0 : }
2394 :
2395 0 : if let Some(lsn) = ancestor_start_lsn.as_mut() {
2396 0 : *lsn = lsn.align();
2397 0 :
2398 0 : let ancestor_ancestor_lsn = ancestor_timeline.get_ancestor_lsn();
2399 0 : if ancestor_ancestor_lsn > *lsn {
2400 : // can we safely just branch from the ancestor instead?
2401 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
2402 0 : "invalid start lsn {} for ancestor timeline {}: less than timeline ancestor lsn {}",
2403 0 : lsn,
2404 0 : ancestor_timeline_id,
2405 0 : ancestor_ancestor_lsn,
2406 0 : )));
2407 0 : }
2408 0 :
2409 0 : // Wait for the WAL to arrive and be processed on the parent branch up
2410 0 : // to the requested branch point. The repository code itself doesn't
2411 0 : // require it, but if we start to receive WAL on the new timeline,
2412 0 : // decoding the new WAL might need to look up previous pages, relation
2413 0 : // sizes etc. and that would get confused if the previous page versions
2414 0 : // are not in the repository yet.
2415 0 : ancestor_timeline
2416 0 : .wait_lsn(*lsn, timeline::WaitLsnWaiter::Tenant, ctx)
2417 0 : .await
2418 0 : .map_err(|e| match e {
2419 0 : e @ (WaitLsnError::Timeout(_) | WaitLsnError::BadState { .. }) => {
2420 0 : CreateTimelineError::AncestorLsn(anyhow::anyhow!(e))
2421 : }
2422 0 : WaitLsnError::Shutdown => CreateTimelineError::ShuttingDown,
2423 0 : })?;
2424 0 : }
2425 :
2426 0 : self.branch_timeline(&ancestor_timeline, new_timeline_id, ancestor_start_lsn, ctx)
2427 0 : .await?
2428 : }
2429 : };
2430 :
2431 : // At this point we have dropped our guard on [`Self::timelines_creating`], and
2432 : // the timeline is visible in [`Self::timelines`], but it is _not_ durable yet. We must
2433 : // not send a success to the caller until it is. The same applies to idempotent retries.
2434 : //
2435 : // TODO: the timeline is already visible in [`Self::timelines`]; a caller could incorrectly
2436 : // assume that, because they can see the timeline via API, that the creation is done and
2437 : // that it is durable. Ideally, we would keep the timeline hidden (in [`Self::timelines_creating`])
2438 : // until it is durable, e.g., by extending the time we hold the creation guard. This also
2439 : // interacts with UninitializedTimeline and is generally a bit tricky.
2440 : //
2441 : // To re-emphasize: the only correct way to create a timeline is to repeat calling the
2442 : // creation API until it returns success. Only then is durability guaranteed.
2443 0 : info!(creation_result=%result.discriminant(), "waiting for timeline to be durable");
2444 0 : result
2445 0 : .timeline()
2446 0 : .remote_client
2447 0 : .wait_completion()
2448 0 : .await
2449 0 : .context("wait for timeline initial uploads to complete")?;
2450 :
2451 : // The creating task is responsible for activating the timeline.
2452 : // We do this after `wait_completion()` so that we don't spin up tasks that start
2453 : // doing stuff before the IndexPart is durable in S3, which is done by the previous section.
2454 0 : let activated_timeline = match result {
2455 0 : CreateTimelineResult::Created(timeline) => {
2456 0 : timeline.activate(self.clone(), broker_client, None, ctx);
2457 0 : timeline
2458 : }
2459 0 : CreateTimelineResult::Idempotent(timeline) => {
2460 0 : info!(
2461 0 : "request was deemed idempotent, activation will be done by the creating task"
2462 : );
2463 0 : timeline
2464 : }
2465 : };
2466 :
2467 0 : Ok(activated_timeline)
2468 0 : }
2469 :
2470 0 : pub(crate) async fn delete_timeline(
2471 0 : self: Arc<Self>,
2472 0 : timeline_id: TimelineId,
2473 0 : ) -> Result<(), DeleteTimelineError> {
2474 0 : DeleteTimelineFlow::run(&self, timeline_id).await?;
2475 :
2476 0 : Ok(())
2477 0 : }
2478 :
2479 : /// perform one garbage collection iteration, removing old data files from disk.
2480 : /// this function is periodically called by gc task.
2481 : /// also it can be explicitly requested through page server api 'do_gc' command.
2482 : ///
2483 : /// `target_timeline_id` specifies the timeline to GC, or None for all.
2484 : ///
2485 : /// The `horizon` an `pitr` parameters determine how much WAL history needs to be retained.
2486 : /// Also known as the retention period, or the GC cutoff point. `horizon` specifies
2487 : /// the amount of history, as LSN difference from current latest LSN on each timeline.
2488 : /// `pitr` specifies the same as a time difference from the current time. The effective
2489 : /// GC cutoff point is determined conservatively by either `horizon` and `pitr`, whichever
2490 : /// requires more history to be retained.
2491 : //
2492 754 : pub(crate) async fn gc_iteration(
2493 754 : &self,
2494 754 : target_timeline_id: Option<TimelineId>,
2495 754 : horizon: u64,
2496 754 : pitr: Duration,
2497 754 : cancel: &CancellationToken,
2498 754 : ctx: &RequestContext,
2499 754 : ) -> Result<GcResult, GcError> {
2500 754 : // Don't start doing work during shutdown
2501 754 : if let TenantState::Stopping { .. } = self.current_state() {
2502 0 : return Ok(GcResult::default());
2503 754 : }
2504 754 :
2505 754 : // there is a global allowed_error for this
2506 754 : if !self.is_active() {
2507 0 : return Err(GcError::NotActive);
2508 754 : }
2509 754 :
2510 754 : {
2511 754 : let conf = self.tenant_conf.load();
2512 754 :
2513 754 : if !conf.location.may_delete_layers_hint() {
2514 0 : info!("Skipping GC in location state {:?}", conf.location);
2515 0 : return Ok(GcResult::default());
2516 754 : }
2517 754 :
2518 754 : if conf.is_gc_blocked_by_lsn_lease_deadline() {
2519 750 : info!("Skipping GC because lsn lease deadline is not reached");
2520 750 : return Ok(GcResult::default());
2521 4 : }
2522 : }
2523 :
2524 4 : let _guard = match self.gc_block.start().await {
2525 4 : Ok(guard) => guard,
2526 0 : Err(reasons) => {
2527 0 : info!("Skipping GC: {reasons}");
2528 0 : return Ok(GcResult::default());
2529 : }
2530 : };
2531 :
2532 4 : self.gc_iteration_internal(target_timeline_id, horizon, pitr, cancel, ctx)
2533 4 : .await
2534 754 : }
2535 :
2536 : /// Perform one compaction iteration.
2537 : /// This function is periodically called by compactor task.
2538 : /// Also it can be explicitly requested per timeline through page server
2539 : /// api's 'compact' command.
2540 : ///
2541 : /// Returns whether we have pending compaction task.
2542 0 : async fn compaction_iteration(
2543 0 : self: &Arc<Self>,
2544 0 : cancel: &CancellationToken,
2545 0 : ctx: &RequestContext,
2546 0 : ) -> Result<bool, timeline::CompactionError> {
2547 0 : // Don't start doing work during shutdown, or when broken, we do not need those in the logs
2548 0 : if !self.is_active() {
2549 0 : return Ok(false);
2550 0 : }
2551 0 :
2552 0 : {
2553 0 : let conf = self.tenant_conf.load();
2554 0 : if !conf.location.may_delete_layers_hint() || !conf.location.may_upload_layers_hint() {
2555 0 : info!("Skipping compaction in location state {:?}", conf.location);
2556 0 : return Ok(false);
2557 0 : }
2558 0 : }
2559 0 :
2560 0 : // Scan through the hashmap and collect a list of all the timelines,
2561 0 : // while holding the lock. Then drop the lock and actually perform the
2562 0 : // compactions. We don't want to block everything else while the
2563 0 : // compaction runs.
2564 0 : let timelines_to_compact_or_offload;
2565 0 : {
2566 0 : let timelines = self.timelines.lock().unwrap();
2567 0 : timelines_to_compact_or_offload = timelines
2568 0 : .iter()
2569 0 : .filter_map(|(timeline_id, timeline)| {
2570 0 : let (is_active, (can_offload, _)) =
2571 0 : (timeline.is_active(), timeline.can_offload());
2572 0 : let has_no_unoffloaded_children = {
2573 0 : !timelines
2574 0 : .iter()
2575 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(*timeline_id))
2576 : };
2577 0 : let config_allows_offload = self.conf.timeline_offloading
2578 0 : || self
2579 0 : .tenant_conf
2580 0 : .load()
2581 0 : .tenant_conf
2582 0 : .timeline_offloading
2583 0 : .unwrap_or_default();
2584 0 : let can_offload =
2585 0 : can_offload && has_no_unoffloaded_children && config_allows_offload;
2586 0 : if (is_active, can_offload) == (false, false) {
2587 0 : None
2588 : } else {
2589 0 : Some((*timeline_id, timeline.clone(), (is_active, can_offload)))
2590 : }
2591 0 : })
2592 0 : .collect::<Vec<_>>();
2593 0 : drop(timelines);
2594 0 : }
2595 0 :
2596 0 : // Before doing any I/O work, check our circuit breaker
2597 0 : if self.compaction_circuit_breaker.lock().unwrap().is_broken() {
2598 0 : info!("Skipping compaction due to previous failures");
2599 0 : return Ok(false);
2600 0 : }
2601 0 :
2602 0 : let mut has_pending_task = false;
2603 :
2604 0 : for (timeline_id, timeline, (can_compact, can_offload)) in &timelines_to_compact_or_offload
2605 : {
2606 0 : let pending_task_left = if *can_compact {
2607 : Some(
2608 0 : timeline
2609 0 : .compact(cancel, EnumSet::empty(), ctx)
2610 0 : .instrument(info_span!("compact_timeline", %timeline_id))
2611 0 : .await
2612 0 : .inspect_err(|e| match e {
2613 0 : timeline::CompactionError::ShuttingDown => (),
2614 0 : timeline::CompactionError::Offload(_) => {
2615 0 : // Failures to offload timelines do not trip the circuit breaker, because
2616 0 : // they do not do lots of writes the way compaction itself does: it is cheap
2617 0 : // to retry, and it would be bad to stop all compaction because of an issue with offloading.
2618 0 : }
2619 0 : timeline::CompactionError::Other(e) => {
2620 0 : self.compaction_circuit_breaker
2621 0 : .lock()
2622 0 : .unwrap()
2623 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, e);
2624 0 : }
2625 0 : })?,
2626 : )
2627 : } else {
2628 0 : None
2629 : };
2630 0 : has_pending_task |= pending_task_left.unwrap_or(false);
2631 0 : if pending_task_left == Some(false) && *can_offload {
2632 0 : offload_timeline(self, timeline)
2633 0 : .instrument(info_span!("offload_timeline", %timeline_id))
2634 0 : .await?;
2635 0 : }
2636 : }
2637 :
2638 0 : self.compaction_circuit_breaker
2639 0 : .lock()
2640 0 : .unwrap()
2641 0 : .success(&CIRCUIT_BREAKERS_UNBROKEN);
2642 0 :
2643 0 : Ok(has_pending_task)
2644 0 : }
2645 :
2646 : // Call through to all timelines to freeze ephemeral layers if needed. Usually
2647 : // this happens during ingest: this background housekeeping is for freezing layers
2648 : // that are open but haven't been written to for some time.
2649 0 : async fn ingest_housekeeping(&self) {
2650 0 : // Scan through the hashmap and collect a list of all the timelines,
2651 0 : // while holding the lock. Then drop the lock and actually perform the
2652 0 : // compactions. We don't want to block everything else while the
2653 0 : // compaction runs.
2654 0 : let timelines = {
2655 0 : self.timelines
2656 0 : .lock()
2657 0 : .unwrap()
2658 0 : .values()
2659 0 : .filter_map(|timeline| {
2660 0 : if timeline.is_active() {
2661 0 : Some(timeline.clone())
2662 : } else {
2663 0 : None
2664 : }
2665 0 : })
2666 0 : .collect::<Vec<_>>()
2667 : };
2668 :
2669 0 : for timeline in &timelines {
2670 0 : timeline.maybe_freeze_ephemeral_layer().await;
2671 : }
2672 0 : }
2673 :
2674 0 : pub fn timeline_has_no_attached_children(&self, timeline_id: TimelineId) -> bool {
2675 0 : let timelines = self.timelines.lock().unwrap();
2676 0 : !timelines
2677 0 : .iter()
2678 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(timeline_id))
2679 0 : }
2680 :
2681 1702 : pub fn current_state(&self) -> TenantState {
2682 1702 : self.state.borrow().clone()
2683 1702 : }
2684 :
2685 942 : pub fn is_active(&self) -> bool {
2686 942 : self.current_state() == TenantState::Active
2687 942 : }
2688 :
2689 0 : pub fn generation(&self) -> Generation {
2690 0 : self.generation
2691 0 : }
2692 :
2693 0 : pub(crate) fn wal_redo_manager_status(&self) -> Option<WalRedoManagerStatus> {
2694 0 : self.walredo_mgr.as_ref().and_then(|mgr| mgr.status())
2695 0 : }
2696 :
2697 : /// Changes tenant status to active, unless shutdown was already requested.
2698 : ///
2699 : /// `background_jobs_can_start` is an optional barrier set to a value during pageserver startup
2700 : /// to delay background jobs. Background jobs can be started right away when None is given.
2701 0 : fn activate(
2702 0 : self: &Arc<Self>,
2703 0 : broker_client: BrokerClientChannel,
2704 0 : background_jobs_can_start: Option<&completion::Barrier>,
2705 0 : ctx: &RequestContext,
2706 0 : ) {
2707 0 : span::debug_assert_current_span_has_tenant_id();
2708 0 :
2709 0 : let mut activating = false;
2710 0 : self.state.send_modify(|current_state| {
2711 : use pageserver_api::models::ActivatingFrom;
2712 0 : match &*current_state {
2713 : TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
2714 0 : panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {state:?}", state = current_state);
2715 : }
2716 0 : TenantState::Attaching => {
2717 0 : *current_state = TenantState::Activating(ActivatingFrom::Attaching);
2718 0 : }
2719 0 : }
2720 0 : debug!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), "Activating tenant");
2721 0 : activating = true;
2722 0 : // Continue outside the closure. We need to grab timelines.lock()
2723 0 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
2724 0 : });
2725 0 :
2726 0 : if activating {
2727 0 : let timelines_accessor = self.timelines.lock().unwrap();
2728 0 : let timelines_offloaded_accessor = self.timelines_offloaded.lock().unwrap();
2729 0 : let timelines_to_activate = timelines_accessor
2730 0 : .values()
2731 0 : .filter(|timeline| !(timeline.is_broken() || timeline.is_stopping()));
2732 0 :
2733 0 : // Before activation, populate each Timeline's GcInfo with information about its children
2734 0 : self.initialize_gc_info(&timelines_accessor, &timelines_offloaded_accessor, None);
2735 0 :
2736 0 : // Spawn gc and compaction loops. The loops will shut themselves
2737 0 : // down when they notice that the tenant is inactive.
2738 0 : tasks::start_background_loops(self, background_jobs_can_start);
2739 0 :
2740 0 : let mut activated_timelines = 0;
2741 :
2742 0 : for timeline in timelines_to_activate {
2743 0 : timeline.activate(
2744 0 : self.clone(),
2745 0 : broker_client.clone(),
2746 0 : background_jobs_can_start,
2747 0 : ctx,
2748 0 : );
2749 0 : activated_timelines += 1;
2750 0 : }
2751 :
2752 0 : self.state.send_modify(move |current_state| {
2753 0 : assert!(
2754 0 : matches!(current_state, TenantState::Activating(_)),
2755 0 : "set_stopping and set_broken wait for us to leave Activating state",
2756 : );
2757 0 : *current_state = TenantState::Active;
2758 0 :
2759 0 : let elapsed = self.constructed_at.elapsed();
2760 0 : let total_timelines = timelines_accessor.len();
2761 0 :
2762 0 : // log a lot of stuff, because some tenants sometimes suffer from user-visible
2763 0 : // times to activate. see https://github.com/neondatabase/neon/issues/4025
2764 0 : info!(
2765 0 : since_creation_millis = elapsed.as_millis(),
2766 0 : tenant_id = %self.tenant_shard_id.tenant_id,
2767 0 : shard_id = %self.tenant_shard_id.shard_slug(),
2768 0 : activated_timelines,
2769 0 : total_timelines,
2770 0 : post_state = <&'static str>::from(&*current_state),
2771 0 : "activation attempt finished"
2772 : );
2773 :
2774 0 : TENANT.activation.observe(elapsed.as_secs_f64());
2775 0 : });
2776 0 : }
2777 0 : }
2778 :
2779 : /// Shutdown the tenant and join all of the spawned tasks.
2780 : ///
2781 : /// The method caters for all use-cases:
2782 : /// - pageserver shutdown (freeze_and_flush == true)
2783 : /// - detach + ignore (freeze_and_flush == false)
2784 : ///
2785 : /// This will attempt to shutdown even if tenant is broken.
2786 : ///
2787 : /// `shutdown_progress` is a [`completion::Barrier`] for the shutdown initiated by this call.
2788 : /// If the tenant is already shutting down, we return a clone of the first shutdown call's
2789 : /// `Barrier` as an `Err`. This not-first caller can use the returned barrier to join with
2790 : /// the ongoing shutdown.
2791 6 : async fn shutdown(
2792 6 : &self,
2793 6 : shutdown_progress: completion::Barrier,
2794 6 : shutdown_mode: timeline::ShutdownMode,
2795 6 : ) -> Result<(), completion::Barrier> {
2796 6 : span::debug_assert_current_span_has_tenant_id();
2797 :
2798 : // Set tenant (and its timlines) to Stoppping state.
2799 : //
2800 : // Since we can only transition into Stopping state after activation is complete,
2801 : // run it in a JoinSet so all tenants have a chance to stop before we get SIGKILLed.
2802 : //
2803 : // Transitioning tenants to Stopping state has a couple of non-obvious side effects:
2804 : // 1. Lock out any new requests to the tenants.
2805 : // 2. Signal cancellation to WAL receivers (we wait on it below).
2806 : // 3. Signal cancellation for other tenant background loops.
2807 : // 4. ???
2808 : //
2809 : // The waiting for the cancellation is not done uniformly.
2810 : // We certainly wait for WAL receivers to shut down.
2811 : // That is necessary so that no new data comes in before the freeze_and_flush.
2812 : // But the tenant background loops are joined-on in our caller.
2813 : // It's mesed up.
2814 : // we just ignore the failure to stop
2815 :
2816 : // If we're still attaching, fire the cancellation token early to drop out: this
2817 : // will prevent us flushing, but ensures timely shutdown if some I/O during attach
2818 : // is very slow.
2819 6 : let shutdown_mode = if matches!(self.current_state(), TenantState::Attaching) {
2820 0 : self.cancel.cancel();
2821 0 :
2822 0 : // Having fired our cancellation token, do not try and flush timelines: their cancellation tokens
2823 0 : // are children of ours, so their flush loops will have shut down already
2824 0 : timeline::ShutdownMode::Hard
2825 : } else {
2826 6 : shutdown_mode
2827 : };
2828 :
2829 6 : match self.set_stopping(shutdown_progress, false, false).await {
2830 6 : Ok(()) => {}
2831 0 : Err(SetStoppingError::Broken) => {
2832 0 : // assume that this is acceptable
2833 0 : }
2834 0 : Err(SetStoppingError::AlreadyStopping(other)) => {
2835 0 : // give caller the option to wait for this this shutdown
2836 0 : info!("Tenant::shutdown: AlreadyStopping");
2837 0 : return Err(other);
2838 : }
2839 : };
2840 :
2841 6 : let mut js = tokio::task::JoinSet::new();
2842 6 : {
2843 6 : let timelines = self.timelines.lock().unwrap();
2844 6 : timelines.values().for_each(|timeline| {
2845 6 : let timeline = Arc::clone(timeline);
2846 6 : let timeline_id = timeline.timeline_id;
2847 6 : let span = tracing::info_span!("timeline_shutdown", %timeline_id, ?shutdown_mode);
2848 10 : js.spawn(async move { timeline.shutdown(shutdown_mode).instrument(span).await });
2849 6 : });
2850 6 : }
2851 6 : {
2852 6 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
2853 6 : timelines_offloaded.values().for_each(|timeline| {
2854 0 : timeline.defuse_for_tenant_drop();
2855 6 : });
2856 6 : }
2857 6 : // test_long_timeline_create_then_tenant_delete is leaning on this message
2858 6 : tracing::info!("Waiting for timelines...");
2859 12 : while let Some(res) = js.join_next().await {
2860 0 : match res {
2861 6 : Ok(()) => {}
2862 0 : Err(je) if je.is_cancelled() => unreachable!("no cancelling used"),
2863 0 : Err(je) if je.is_panic() => { /* logged already */ }
2864 0 : Err(je) => warn!("unexpected JoinError: {je:?}"),
2865 : }
2866 : }
2867 :
2868 : // We cancel the Tenant's cancellation token _after_ the timelines have all shut down. This permits
2869 : // them to continue to do work during their shutdown methods, e.g. flushing data.
2870 6 : tracing::debug!("Cancelling CancellationToken");
2871 6 : self.cancel.cancel();
2872 6 :
2873 6 : // shutdown all tenant and timeline tasks: gc, compaction, page service
2874 6 : // No new tasks will be started for this tenant because it's in `Stopping` state.
2875 6 : //
2876 6 : // this will additionally shutdown and await all timeline tasks.
2877 6 : tracing::debug!("Waiting for tasks...");
2878 6 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), None).await;
2879 :
2880 6 : if let Some(walredo_mgr) = self.walredo_mgr.as_ref() {
2881 6 : walredo_mgr.shutdown().await;
2882 0 : }
2883 :
2884 : // Wait for any in-flight operations to complete
2885 6 : self.gate.close().await;
2886 :
2887 6 : remove_tenant_metrics(&self.tenant_shard_id);
2888 6 :
2889 6 : Ok(())
2890 6 : }
2891 :
2892 : /// Change tenant status to Stopping, to mark that it is being shut down.
2893 : ///
2894 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
2895 : ///
2896 : /// This function is not cancel-safe!
2897 : ///
2898 : /// `allow_transition_from_loading` is needed for the special case of loading task deleting the tenant.
2899 : /// `allow_transition_from_attaching` is needed for the special case of attaching deleted tenant.
2900 6 : async fn set_stopping(
2901 6 : &self,
2902 6 : progress: completion::Barrier,
2903 6 : _allow_transition_from_loading: bool,
2904 6 : allow_transition_from_attaching: bool,
2905 6 : ) -> Result<(), SetStoppingError> {
2906 6 : let mut rx = self.state.subscribe();
2907 6 :
2908 6 : // cannot stop before we're done activating, so wait out until we're done activating
2909 6 : rx.wait_for(|state| match state {
2910 0 : TenantState::Attaching if allow_transition_from_attaching => true,
2911 : TenantState::Activating(_) | TenantState::Attaching => {
2912 0 : info!(
2913 0 : "waiting for {} to turn Active|Broken|Stopping",
2914 0 : <&'static str>::from(state)
2915 : );
2916 0 : false
2917 : }
2918 6 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
2919 6 : })
2920 0 : .await
2921 6 : .expect("cannot drop self.state while on a &self method");
2922 6 :
2923 6 : // we now know we're done activating, let's see whether this task is the winner to transition into Stopping
2924 6 : let mut err = None;
2925 6 : let stopping = self.state.send_if_modified(|current_state| match current_state {
2926 : TenantState::Activating(_) => {
2927 0 : unreachable!("1we ensured above that we're done with activation, and, there is no re-activation")
2928 : }
2929 : TenantState::Attaching => {
2930 0 : if !allow_transition_from_attaching {
2931 0 : unreachable!("2we ensured above that we're done with activation, and, there is no re-activation")
2932 0 : };
2933 0 : *current_state = TenantState::Stopping { progress };
2934 0 : true
2935 : }
2936 : TenantState::Active => {
2937 : // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
2938 : // are created after the transition to Stopping. That's harmless, as the Timelines
2939 : // won't be accessible to anyone afterwards, because the Tenant is in Stopping state.
2940 6 : *current_state = TenantState::Stopping { progress };
2941 6 : // Continue stopping outside the closure. We need to grab timelines.lock()
2942 6 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
2943 6 : true
2944 : }
2945 0 : TenantState::Broken { reason, .. } => {
2946 0 : info!(
2947 0 : "Cannot set tenant to Stopping state, it is in Broken state due to: {reason}"
2948 : );
2949 0 : err = Some(SetStoppingError::Broken);
2950 0 : false
2951 : }
2952 0 : TenantState::Stopping { progress } => {
2953 0 : info!("Tenant is already in Stopping state");
2954 0 : err = Some(SetStoppingError::AlreadyStopping(progress.clone()));
2955 0 : false
2956 : }
2957 6 : });
2958 6 : match (stopping, err) {
2959 6 : (true, None) => {} // continue
2960 0 : (false, Some(err)) => return Err(err),
2961 0 : (true, Some(_)) => unreachable!(
2962 0 : "send_if_modified closure must error out if not transitioning to Stopping"
2963 0 : ),
2964 0 : (false, None) => unreachable!(
2965 0 : "send_if_modified closure must return true if transitioning to Stopping"
2966 0 : ),
2967 : }
2968 :
2969 6 : let timelines_accessor = self.timelines.lock().unwrap();
2970 6 : let not_broken_timelines = timelines_accessor
2971 6 : .values()
2972 6 : .filter(|timeline| !timeline.is_broken());
2973 12 : for timeline in not_broken_timelines {
2974 6 : timeline.set_state(TimelineState::Stopping);
2975 6 : }
2976 6 : Ok(())
2977 6 : }
2978 :
2979 : /// Method for tenant::mgr to transition us into Broken state in case of a late failure in
2980 : /// `remove_tenant_from_memory`
2981 : ///
2982 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
2983 : ///
2984 : /// In tests, we also use this to set tenants to Broken state on purpose.
2985 0 : pub(crate) async fn set_broken(&self, reason: String) {
2986 0 : let mut rx = self.state.subscribe();
2987 0 :
2988 0 : // The load & attach routines own the tenant state until it has reached `Active`.
2989 0 : // So, wait until it's done.
2990 0 : rx.wait_for(|state| match state {
2991 : TenantState::Activating(_) | TenantState::Attaching => {
2992 0 : info!(
2993 0 : "waiting for {} to turn Active|Broken|Stopping",
2994 0 : <&'static str>::from(state)
2995 : );
2996 0 : false
2997 : }
2998 0 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
2999 0 : })
3000 0 : .await
3001 0 : .expect("cannot drop self.state while on a &self method");
3002 0 :
3003 0 : // we now know we're done activating, let's see whether this task is the winner to transition into Broken
3004 0 : self.set_broken_no_wait(reason)
3005 0 : }
3006 :
3007 0 : pub(crate) fn set_broken_no_wait(&self, reason: impl Display) {
3008 0 : let reason = reason.to_string();
3009 0 : self.state.send_modify(|current_state| {
3010 0 : match *current_state {
3011 : TenantState::Activating(_) | TenantState::Attaching => {
3012 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
3013 : }
3014 : TenantState::Active => {
3015 0 : if cfg!(feature = "testing") {
3016 0 : warn!("Changing Active tenant to Broken state, reason: {}", reason);
3017 0 : *current_state = TenantState::broken_from_reason(reason);
3018 : } else {
3019 0 : unreachable!("not allowed to call set_broken on Active tenants in non-testing builds")
3020 : }
3021 : }
3022 : TenantState::Broken { .. } => {
3023 0 : warn!("Tenant is already in Broken state");
3024 : }
3025 : // This is the only "expected" path, any other path is a bug.
3026 : TenantState::Stopping { .. } => {
3027 0 : warn!(
3028 0 : "Marking Stopping tenant as Broken state, reason: {}",
3029 : reason
3030 : );
3031 0 : *current_state = TenantState::broken_from_reason(reason);
3032 : }
3033 : }
3034 0 : });
3035 0 : }
3036 :
3037 0 : pub fn subscribe_for_state_updates(&self) -> watch::Receiver<TenantState> {
3038 0 : self.state.subscribe()
3039 0 : }
3040 :
3041 : /// The activate_now semaphore is initialized with zero units. As soon as
3042 : /// we add a unit, waiters will be able to acquire a unit and proceed.
3043 0 : pub(crate) fn activate_now(&self) {
3044 0 : self.activate_now_sem.add_permits(1);
3045 0 : }
3046 :
3047 0 : pub(crate) async fn wait_to_become_active(
3048 0 : &self,
3049 0 : timeout: Duration,
3050 0 : ) -> Result<(), GetActiveTenantError> {
3051 0 : let mut receiver = self.state.subscribe();
3052 : loop {
3053 0 : let current_state = receiver.borrow_and_update().clone();
3054 0 : match current_state {
3055 : TenantState::Attaching | TenantState::Activating(_) => {
3056 : // in these states, there's a chance that we can reach ::Active
3057 0 : self.activate_now();
3058 0 : match timeout_cancellable(timeout, &self.cancel, receiver.changed()).await {
3059 0 : Ok(r) => {
3060 0 : r.map_err(
3061 0 : |_e: tokio::sync::watch::error::RecvError|
3062 : // Tenant existed but was dropped: report it as non-existent
3063 0 : GetActiveTenantError::NotFound(GetTenantError::NotFound(self.tenant_shard_id.tenant_id))
3064 0 : )?
3065 : }
3066 : Err(TimeoutCancellableError::Cancelled) => {
3067 0 : return Err(GetActiveTenantError::Cancelled);
3068 : }
3069 : Err(TimeoutCancellableError::Timeout) => {
3070 0 : return Err(GetActiveTenantError::WaitForActiveTimeout {
3071 0 : latest_state: Some(self.current_state()),
3072 0 : wait_time: timeout,
3073 0 : });
3074 : }
3075 : }
3076 : }
3077 : TenantState::Active { .. } => {
3078 0 : return Ok(());
3079 : }
3080 0 : TenantState::Broken { reason, .. } => {
3081 0 : // This is fatal, and reported distinctly from the general case of "will never be active" because
3082 0 : // it's logically a 500 to external API users (broken is always a bug).
3083 0 : return Err(GetActiveTenantError::Broken(reason));
3084 : }
3085 : TenantState::Stopping { .. } => {
3086 : // There's no chance the tenant can transition back into ::Active
3087 0 : return Err(GetActiveTenantError::WillNotBecomeActive(current_state));
3088 : }
3089 : }
3090 : }
3091 0 : }
3092 :
3093 0 : pub(crate) fn get_attach_mode(&self) -> AttachmentMode {
3094 0 : self.tenant_conf.load().location.attach_mode
3095 0 : }
3096 :
3097 : /// For API access: generate a LocationConfig equivalent to the one that would be used to
3098 : /// create a Tenant in the same state. Do not use this in hot paths: it's for relatively
3099 : /// rare external API calls, like a reconciliation at startup.
3100 0 : pub(crate) fn get_location_conf(&self) -> models::LocationConfig {
3101 0 : let conf = self.tenant_conf.load();
3102 :
3103 0 : let location_config_mode = match conf.location.attach_mode {
3104 0 : AttachmentMode::Single => models::LocationConfigMode::AttachedSingle,
3105 0 : AttachmentMode::Multi => models::LocationConfigMode::AttachedMulti,
3106 0 : AttachmentMode::Stale => models::LocationConfigMode::AttachedStale,
3107 : };
3108 :
3109 : // We have a pageserver TenantConf, we need the API-facing TenantConfig.
3110 0 : let tenant_config: models::TenantConfig = conf.tenant_conf.clone().into();
3111 0 :
3112 0 : models::LocationConfig {
3113 0 : mode: location_config_mode,
3114 0 : generation: self.generation.into(),
3115 0 : secondary_conf: None,
3116 0 : shard_number: self.shard_identity.number.0,
3117 0 : shard_count: self.shard_identity.count.literal(),
3118 0 : shard_stripe_size: self.shard_identity.stripe_size.0,
3119 0 : tenant_conf: tenant_config,
3120 0 : }
3121 0 : }
3122 :
3123 0 : pub(crate) fn get_tenant_shard_id(&self) -> &TenantShardId {
3124 0 : &self.tenant_shard_id
3125 0 : }
3126 :
3127 0 : pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
3128 0 : self.shard_identity.stripe_size
3129 0 : }
3130 :
3131 0 : pub(crate) fn get_generation(&self) -> Generation {
3132 0 : self.generation
3133 0 : }
3134 :
3135 : /// This function partially shuts down the tenant (it shuts down the Timelines) and is fallible,
3136 : /// and can leave the tenant in a bad state if it fails. The caller is responsible for
3137 : /// resetting this tenant to a valid state if we fail.
3138 0 : pub(crate) async fn split_prepare(
3139 0 : &self,
3140 0 : child_shards: &Vec<TenantShardId>,
3141 0 : ) -> anyhow::Result<()> {
3142 0 : let (timelines, offloaded) = {
3143 0 : let timelines = self.timelines.lock().unwrap();
3144 0 : let offloaded = self.timelines_offloaded.lock().unwrap();
3145 0 : (timelines.clone(), offloaded.clone())
3146 0 : };
3147 0 : let timelines_iter = timelines
3148 0 : .values()
3149 0 : .map(TimelineOrOffloadedArcRef::<'_>::from)
3150 0 : .chain(
3151 0 : offloaded
3152 0 : .values()
3153 0 : .map(TimelineOrOffloadedArcRef::<'_>::from),
3154 0 : );
3155 0 : for timeline in timelines_iter {
3156 : // We do not block timeline creation/deletion during splits inside the pageserver: it is up to higher levels
3157 : // to ensure that they do not start a split if currently in the process of doing these.
3158 :
3159 0 : let timeline_id = timeline.timeline_id();
3160 :
3161 0 : if let TimelineOrOffloadedArcRef::Timeline(timeline) = timeline {
3162 : // Upload an index from the parent: this is partly to provide freshness for the
3163 : // child tenants that will copy it, and partly for general ease-of-debugging: there will
3164 : // always be a parent shard index in the same generation as we wrote the child shard index.
3165 0 : tracing::info!(%timeline_id, "Uploading index");
3166 0 : timeline
3167 0 : .remote_client
3168 0 : .schedule_index_upload_for_file_changes()?;
3169 0 : timeline.remote_client.wait_completion().await?;
3170 0 : }
3171 :
3172 0 : let remote_client = match timeline {
3173 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.remote_client.clone(),
3174 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => {
3175 0 : let remote_client = self
3176 0 : .build_timeline_client(offloaded.timeline_id, self.remote_storage.clone());
3177 0 : Arc::new(remote_client)
3178 : }
3179 : };
3180 :
3181 : // Shut down the timeline's remote client: this means that the indices we write
3182 : // for child shards will not be invalidated by the parent shard deleting layers.
3183 0 : tracing::info!(%timeline_id, "Shutting down remote storage client");
3184 0 : remote_client.shutdown().await;
3185 :
3186 : // Download methods can still be used after shutdown, as they don't flow through the remote client's
3187 : // queue. In principal the RemoteTimelineClient could provide this without downloading it, but this
3188 : // operation is rare, so it's simpler to just download it (and robustly guarantees that the index
3189 : // we use here really is the remotely persistent one).
3190 0 : tracing::info!(%timeline_id, "Downloading index_part from parent");
3191 0 : let result = remote_client
3192 0 : .download_index_file(&self.cancel)
3193 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))
3194 0 : .await?;
3195 0 : let index_part = match result {
3196 : MaybeDeletedIndexPart::Deleted(_) => {
3197 0 : anyhow::bail!("Timeline deletion happened concurrently with split")
3198 : }
3199 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
3200 : };
3201 :
3202 0 : for child_shard in child_shards {
3203 0 : tracing::info!(%timeline_id, "Uploading index_part for child {}", child_shard.to_index());
3204 0 : upload_index_part(
3205 0 : &self.remote_storage,
3206 0 : child_shard,
3207 0 : &timeline_id,
3208 0 : self.generation,
3209 0 : &index_part,
3210 0 : &self.cancel,
3211 0 : )
3212 0 : .await?;
3213 : }
3214 : }
3215 :
3216 0 : let tenant_manifest = self.build_tenant_manifest();
3217 0 : for child_shard in child_shards {
3218 0 : tracing::info!(
3219 0 : "Uploading tenant manifest for child {}",
3220 0 : child_shard.to_index()
3221 : );
3222 0 : upload_tenant_manifest(
3223 0 : &self.remote_storage,
3224 0 : child_shard,
3225 0 : self.generation,
3226 0 : &tenant_manifest,
3227 0 : &self.cancel,
3228 0 : )
3229 0 : .await?;
3230 : }
3231 :
3232 0 : Ok(())
3233 0 : }
3234 :
3235 0 : pub(crate) fn get_sizes(&self) -> TopTenantShardItem {
3236 0 : let mut result = TopTenantShardItem {
3237 0 : id: self.tenant_shard_id,
3238 0 : resident_size: 0,
3239 0 : physical_size: 0,
3240 0 : max_logical_size: 0,
3241 0 : };
3242 :
3243 0 : for timeline in self.timelines.lock().unwrap().values() {
3244 0 : result.resident_size += timeline.metrics.resident_physical_size_gauge.get();
3245 0 :
3246 0 : result.physical_size += timeline
3247 0 : .remote_client
3248 0 : .metrics
3249 0 : .remote_physical_size_gauge
3250 0 : .get();
3251 0 : result.max_logical_size = std::cmp::max(
3252 0 : result.max_logical_size,
3253 0 : timeline.metrics.current_logical_size_gauge.get(),
3254 0 : );
3255 0 : }
3256 :
3257 0 : result
3258 0 : }
3259 : }
3260 :
3261 : /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
3262 : /// perform a topological sort, so that the parent of each timeline comes
3263 : /// before the children.
3264 : /// E extracts the ancestor from T
3265 : /// This allows for T to be different. It can be TimelineMetadata, can be Timeline itself, etc.
3266 192 : fn tree_sort_timelines<T, E>(
3267 192 : timelines: HashMap<TimelineId, T>,
3268 192 : extractor: E,
3269 192 : ) -> anyhow::Result<Vec<(TimelineId, T)>>
3270 192 : where
3271 192 : E: Fn(&T) -> Option<TimelineId>,
3272 192 : {
3273 192 : let mut result = Vec::with_capacity(timelines.len());
3274 192 :
3275 192 : let mut now = Vec::with_capacity(timelines.len());
3276 192 : // (ancestor, children)
3277 192 : let mut later: HashMap<TimelineId, Vec<(TimelineId, T)>> =
3278 192 : HashMap::with_capacity(timelines.len());
3279 :
3280 198 : for (timeline_id, value) in timelines {
3281 6 : if let Some(ancestor_id) = extractor(&value) {
3282 2 : let children = later.entry(ancestor_id).or_default();
3283 2 : children.push((timeline_id, value));
3284 4 : } else {
3285 4 : now.push((timeline_id, value));
3286 4 : }
3287 : }
3288 :
3289 198 : while let Some((timeline_id, metadata)) = now.pop() {
3290 6 : result.push((timeline_id, metadata));
3291 : // All children of this can be loaded now
3292 6 : if let Some(mut children) = later.remove(&timeline_id) {
3293 2 : now.append(&mut children);
3294 4 : }
3295 : }
3296 :
3297 : // All timelines should be visited now. Unless there were timelines with missing ancestors.
3298 192 : if !later.is_empty() {
3299 0 : for (missing_id, orphan_ids) in later {
3300 0 : for (orphan_id, _) in orphan_ids {
3301 0 : error!("could not load timeline {orphan_id} because its ancestor timeline {missing_id} could not be loaded");
3302 : }
3303 : }
3304 0 : bail!("could not load tenant because some timelines are missing ancestors");
3305 192 : }
3306 192 :
3307 192 : Ok(result)
3308 192 : }
3309 :
3310 : impl Tenant {
3311 0 : pub fn tenant_specific_overrides(&self) -> TenantConfOpt {
3312 0 : self.tenant_conf.load().tenant_conf.clone()
3313 0 : }
3314 :
3315 0 : pub fn effective_config(&self) -> TenantConf {
3316 0 : self.tenant_specific_overrides()
3317 0 : .merge(self.conf.default_tenant_conf.clone())
3318 0 : }
3319 :
3320 0 : pub fn get_checkpoint_distance(&self) -> u64 {
3321 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3322 0 : tenant_conf
3323 0 : .checkpoint_distance
3324 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
3325 0 : }
3326 :
3327 0 : pub fn get_checkpoint_timeout(&self) -> Duration {
3328 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3329 0 : tenant_conf
3330 0 : .checkpoint_timeout
3331 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
3332 0 : }
3333 :
3334 0 : pub fn get_compaction_target_size(&self) -> u64 {
3335 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3336 0 : tenant_conf
3337 0 : .compaction_target_size
3338 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
3339 0 : }
3340 :
3341 0 : pub fn get_compaction_period(&self) -> Duration {
3342 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3343 0 : tenant_conf
3344 0 : .compaction_period
3345 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
3346 0 : }
3347 :
3348 0 : pub fn get_compaction_threshold(&self) -> usize {
3349 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3350 0 : tenant_conf
3351 0 : .compaction_threshold
3352 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
3353 0 : }
3354 :
3355 0 : pub fn get_gc_horizon(&self) -> u64 {
3356 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3357 0 : tenant_conf
3358 0 : .gc_horizon
3359 0 : .unwrap_or(self.conf.default_tenant_conf.gc_horizon)
3360 0 : }
3361 :
3362 0 : pub fn get_gc_period(&self) -> Duration {
3363 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3364 0 : tenant_conf
3365 0 : .gc_period
3366 0 : .unwrap_or(self.conf.default_tenant_conf.gc_period)
3367 0 : }
3368 :
3369 0 : pub fn get_image_creation_threshold(&self) -> usize {
3370 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3371 0 : tenant_conf
3372 0 : .image_creation_threshold
3373 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
3374 0 : }
3375 :
3376 0 : pub fn get_pitr_interval(&self) -> Duration {
3377 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3378 0 : tenant_conf
3379 0 : .pitr_interval
3380 0 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
3381 0 : }
3382 :
3383 0 : pub fn get_min_resident_size_override(&self) -> Option<u64> {
3384 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3385 0 : tenant_conf
3386 0 : .min_resident_size_override
3387 0 : .or(self.conf.default_tenant_conf.min_resident_size_override)
3388 0 : }
3389 :
3390 0 : pub fn get_heatmap_period(&self) -> Option<Duration> {
3391 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3392 0 : let heatmap_period = tenant_conf
3393 0 : .heatmap_period
3394 0 : .unwrap_or(self.conf.default_tenant_conf.heatmap_period);
3395 0 : if heatmap_period.is_zero() {
3396 0 : None
3397 : } else {
3398 0 : Some(heatmap_period)
3399 : }
3400 0 : }
3401 :
3402 4 : pub fn get_lsn_lease_length(&self) -> Duration {
3403 4 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3404 4 : tenant_conf
3405 4 : .lsn_lease_length
3406 4 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
3407 4 : }
3408 :
3409 : /// Generate an up-to-date TenantManifest based on the state of this Tenant.
3410 2 : fn build_tenant_manifest(&self) -> TenantManifest {
3411 2 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
3412 2 :
3413 2 : let mut timeline_manifests = timelines_offloaded
3414 2 : .iter()
3415 2 : .map(|(_timeline_id, offloaded)| offloaded.manifest())
3416 2 : .collect::<Vec<_>>();
3417 2 : // Sort the manifests so that our output is deterministic
3418 2 : timeline_manifests.sort_by_key(|timeline_manifest| timeline_manifest.timeline_id);
3419 2 :
3420 2 : TenantManifest {
3421 2 : version: LATEST_TENANT_MANIFEST_VERSION,
3422 2 : offloaded_timelines: timeline_manifests,
3423 2 : }
3424 2 : }
3425 :
3426 0 : pub fn set_new_tenant_config(&self, new_tenant_conf: TenantConfOpt) {
3427 0 : // Use read-copy-update in order to avoid overwriting the location config
3428 0 : // state if this races with [`Tenant::set_new_location_config`]. Note that
3429 0 : // this race is not possible if both request types come from the storage
3430 0 : // controller (as they should!) because an exclusive op lock is required
3431 0 : // on the storage controller side.
3432 0 : self.tenant_conf.rcu(|inner| {
3433 0 : Arc::new(AttachedTenantConf {
3434 0 : tenant_conf: new_tenant_conf.clone(),
3435 0 : location: inner.location,
3436 0 : // Attached location is not changed, no need to update lsn lease deadline.
3437 0 : lsn_lease_deadline: inner.lsn_lease_deadline,
3438 0 : })
3439 0 : });
3440 0 :
3441 0 : self.tenant_conf_updated(&new_tenant_conf);
3442 0 : // Don't hold self.timelines.lock() during the notifies.
3443 0 : // There's no risk of deadlock right now, but there could be if we consolidate
3444 0 : // mutexes in struct Timeline in the future.
3445 0 : let timelines = self.list_timelines();
3446 0 : for timeline in timelines {
3447 0 : timeline.tenant_conf_updated(&new_tenant_conf);
3448 0 : }
3449 0 : }
3450 :
3451 0 : pub(crate) fn set_new_location_config(&self, new_conf: AttachedTenantConf) {
3452 0 : let new_tenant_conf = new_conf.tenant_conf.clone();
3453 0 :
3454 0 : self.tenant_conf.store(Arc::new(new_conf));
3455 0 :
3456 0 : self.tenant_conf_updated(&new_tenant_conf);
3457 0 : // Don't hold self.timelines.lock() during the notifies.
3458 0 : // There's no risk of deadlock right now, but there could be if we consolidate
3459 0 : // mutexes in struct Timeline in the future.
3460 0 : let timelines = self.list_timelines();
3461 0 : for timeline in timelines {
3462 0 : timeline.tenant_conf_updated(&new_tenant_conf);
3463 0 : }
3464 0 : }
3465 :
3466 192 : fn get_timeline_get_throttle_config(
3467 192 : psconf: &'static PageServerConf,
3468 192 : overrides: &TenantConfOpt,
3469 192 : ) -> throttle::Config {
3470 192 : overrides
3471 192 : .timeline_get_throttle
3472 192 : .clone()
3473 192 : .unwrap_or(psconf.default_tenant_conf.timeline_get_throttle.clone())
3474 192 : }
3475 :
3476 0 : pub(crate) fn tenant_conf_updated(&self, new_conf: &TenantConfOpt) {
3477 0 : let conf = Self::get_timeline_get_throttle_config(self.conf, new_conf);
3478 0 : self.timeline_get_throttle.reconfigure(conf)
3479 0 : }
3480 :
3481 : /// Helper function to create a new Timeline struct.
3482 : ///
3483 : /// The returned Timeline is in Loading state. The caller is responsible for
3484 : /// initializing any on-disk state, and for inserting the Timeline to the 'timelines'
3485 : /// map.
3486 : ///
3487 : /// `validate_ancestor == false` is used when a timeline is created for deletion
3488 : /// and we might not have the ancestor present anymore which is fine for to be
3489 : /// deleted timelines.
3490 418 : fn create_timeline_struct(
3491 418 : &self,
3492 418 : new_timeline_id: TimelineId,
3493 418 : new_metadata: &TimelineMetadata,
3494 418 : ancestor: Option<Arc<Timeline>>,
3495 418 : resources: TimelineResources,
3496 418 : cause: CreateTimelineCause,
3497 418 : create_idempotency: CreateTimelineIdempotency,
3498 418 : ) -> anyhow::Result<Arc<Timeline>> {
3499 418 : let state = match cause {
3500 : CreateTimelineCause::Load => {
3501 418 : let ancestor_id = new_metadata.ancestor_timeline();
3502 418 : anyhow::ensure!(
3503 418 : ancestor_id == ancestor.as_ref().map(|t| t.timeline_id),
3504 0 : "Timeline's {new_timeline_id} ancestor {ancestor_id:?} was not found"
3505 : );
3506 418 : TimelineState::Loading
3507 : }
3508 0 : CreateTimelineCause::Delete => TimelineState::Stopping,
3509 : };
3510 :
3511 418 : let pg_version = new_metadata.pg_version();
3512 418 :
3513 418 : let timeline = Timeline::new(
3514 418 : self.conf,
3515 418 : Arc::clone(&self.tenant_conf),
3516 418 : new_metadata,
3517 418 : ancestor,
3518 418 : new_timeline_id,
3519 418 : self.tenant_shard_id,
3520 418 : self.generation,
3521 418 : self.shard_identity,
3522 418 : self.walredo_mgr.clone(),
3523 418 : resources,
3524 418 : pg_version,
3525 418 : state,
3526 418 : self.attach_wal_lag_cooldown.clone(),
3527 418 : create_idempotency,
3528 418 : self.cancel.child_token(),
3529 418 : );
3530 418 :
3531 418 : Ok(timeline)
3532 418 : }
3533 :
3534 : // Allow too_many_arguments because a constructor's argument list naturally grows with the
3535 : // number of attributes in the struct: breaking these out into a builder wouldn't be helpful.
3536 : #[allow(clippy::too_many_arguments)]
3537 192 : fn new(
3538 192 : state: TenantState,
3539 192 : conf: &'static PageServerConf,
3540 192 : attached_conf: AttachedTenantConf,
3541 192 : shard_identity: ShardIdentity,
3542 192 : walredo_mgr: Option<Arc<WalRedoManager>>,
3543 192 : tenant_shard_id: TenantShardId,
3544 192 : remote_storage: GenericRemoteStorage,
3545 192 : deletion_queue_client: DeletionQueueClient,
3546 192 : l0_flush_global_state: L0FlushGlobalState,
3547 192 : ) -> Tenant {
3548 192 : debug_assert!(
3549 192 : !attached_conf.location.generation.is_none() || conf.control_plane_api.is_none()
3550 : );
3551 :
3552 192 : let (state, mut rx) = watch::channel(state);
3553 192 :
3554 192 : tokio::spawn(async move {
3555 192 : // reflect tenant state in metrics:
3556 192 : // - global per tenant state: TENANT_STATE_METRIC
3557 192 : // - "set" of broken tenants: BROKEN_TENANTS_SET
3558 192 : //
3559 192 : // set of broken tenants should not have zero counts so that it remains accessible for
3560 192 : // alerting.
3561 192 :
3562 192 : let tid = tenant_shard_id.to_string();
3563 192 : let shard_id = tenant_shard_id.shard_slug().to_string();
3564 192 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
3565 :
3566 384 : fn inspect_state(state: &TenantState) -> ([&'static str; 1], bool) {
3567 384 : ([state.into()], matches!(state, TenantState::Broken { .. }))
3568 384 : }
3569 :
3570 192 : let mut tuple = inspect_state(&rx.borrow_and_update());
3571 192 :
3572 192 : let is_broken = tuple.1;
3573 192 : let mut counted_broken = if is_broken {
3574 : // add the id to the set right away, there should not be any updates on the channel
3575 : // after before tenant is removed, if ever
3576 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
3577 0 : true
3578 : } else {
3579 192 : false
3580 : };
3581 :
3582 : loop {
3583 384 : let labels = &tuple.0;
3584 384 : let current = TENANT_STATE_METRIC.with_label_values(labels);
3585 384 : current.inc();
3586 384 :
3587 384 : if rx.changed().await.is_err() {
3588 : // tenant has been dropped
3589 16 : current.dec();
3590 16 : drop(BROKEN_TENANTS_SET.remove_label_values(set_key));
3591 16 : break;
3592 192 : }
3593 192 :
3594 192 : current.dec();
3595 192 : tuple = inspect_state(&rx.borrow_and_update());
3596 192 :
3597 192 : let is_broken = tuple.1;
3598 192 : if is_broken && !counted_broken {
3599 0 : counted_broken = true;
3600 0 : // insert the tenant_id (back) into the set while avoiding needless counter
3601 0 : // access
3602 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
3603 192 : }
3604 : }
3605 192 : });
3606 192 :
3607 192 : Tenant {
3608 192 : tenant_shard_id,
3609 192 : shard_identity,
3610 192 : generation: attached_conf.location.generation,
3611 192 : conf,
3612 192 : // using now here is good enough approximation to catch tenants with really long
3613 192 : // activation times.
3614 192 : constructed_at: Instant::now(),
3615 192 : timelines: Mutex::new(HashMap::new()),
3616 192 : timelines_creating: Mutex::new(HashSet::new()),
3617 192 : timelines_offloaded: Mutex::new(HashMap::new()),
3618 192 : tenant_manifest_upload: Default::default(),
3619 192 : gc_cs: tokio::sync::Mutex::new(()),
3620 192 : walredo_mgr,
3621 192 : remote_storage,
3622 192 : deletion_queue_client,
3623 192 : state,
3624 192 : cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
3625 192 : cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
3626 192 : eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
3627 192 : compaction_circuit_breaker: std::sync::Mutex::new(CircuitBreaker::new(
3628 192 : format!("compaction-{tenant_shard_id}"),
3629 192 : 5,
3630 192 : // Compaction can be a very expensive operation, and might leak disk space. It also ought
3631 192 : // to be infallible, as long as remote storage is available. So if it repeatedly fails,
3632 192 : // use an extremely long backoff.
3633 192 : Some(Duration::from_secs(3600 * 24)),
3634 192 : )),
3635 192 : activate_now_sem: tokio::sync::Semaphore::new(0),
3636 192 : attach_wal_lag_cooldown: Arc::new(std::sync::OnceLock::new()),
3637 192 : cancel: CancellationToken::default(),
3638 192 : gate: Gate::default(),
3639 192 : timeline_get_throttle: Arc::new(throttle::Throttle::new(
3640 192 : Tenant::get_timeline_get_throttle_config(conf, &attached_conf.tenant_conf),
3641 192 : crate::metrics::tenant_throttling::TimelineGet::new(&tenant_shard_id),
3642 192 : )),
3643 192 : tenant_conf: Arc::new(ArcSwap::from_pointee(attached_conf)),
3644 192 : ongoing_timeline_detach: std::sync::Mutex::default(),
3645 192 : gc_block: Default::default(),
3646 192 : l0_flush_global_state,
3647 192 : }
3648 192 : }
3649 :
3650 : /// Locate and load config
3651 0 : pub(super) fn load_tenant_config(
3652 0 : conf: &'static PageServerConf,
3653 0 : tenant_shard_id: &TenantShardId,
3654 0 : ) -> Result<LocationConf, LoadConfigError> {
3655 0 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
3656 0 :
3657 0 : info!("loading tenant configuration from {config_path}");
3658 :
3659 : // load and parse file
3660 0 : let config = fs::read_to_string(&config_path).map_err(|e| {
3661 0 : match e.kind() {
3662 : std::io::ErrorKind::NotFound => {
3663 : // The config should almost always exist for a tenant directory:
3664 : // - When attaching a tenant, the config is the first thing we write
3665 : // - When detaching a tenant, we atomically move the directory to a tmp location
3666 : // before deleting contents.
3667 : //
3668 : // The very rare edge case that can result in a missing config is if we crash during attach
3669 : // between creating directory and writing config. Callers should handle that as if the
3670 : // directory didn't exist.
3671 :
3672 0 : LoadConfigError::NotFound(config_path)
3673 : }
3674 : _ => {
3675 : // No IO errors except NotFound are acceptable here: other kinds of error indicate local storage or permissions issues
3676 : // that we cannot cleanly recover
3677 0 : crate::virtual_file::on_fatal_io_error(&e, "Reading tenant config file")
3678 : }
3679 : }
3680 0 : })?;
3681 :
3682 0 : Ok(toml_edit::de::from_str::<LocationConf>(&config)?)
3683 0 : }
3684 :
3685 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
3686 : pub(super) async fn persist_tenant_config(
3687 : conf: &'static PageServerConf,
3688 : tenant_shard_id: &TenantShardId,
3689 : location_conf: &LocationConf,
3690 : ) -> std::io::Result<()> {
3691 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
3692 :
3693 : Self::persist_tenant_config_at(tenant_shard_id, &config_path, location_conf).await
3694 : }
3695 :
3696 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
3697 : pub(super) async fn persist_tenant_config_at(
3698 : tenant_shard_id: &TenantShardId,
3699 : config_path: &Utf8Path,
3700 : location_conf: &LocationConf,
3701 : ) -> std::io::Result<()> {
3702 : debug!("persisting tenantconf to {config_path}");
3703 :
3704 : let mut conf_content = r#"# This file contains a specific per-tenant's config.
3705 : # It is read in case of pageserver restart.
3706 : "#
3707 : .to_string();
3708 :
3709 0 : fail::fail_point!("tenant-config-before-write", |_| {
3710 0 : Err(std::io::Error::new(
3711 0 : std::io::ErrorKind::Other,
3712 0 : "tenant-config-before-write",
3713 0 : ))
3714 0 : });
3715 :
3716 : // Convert the config to a toml file.
3717 : conf_content +=
3718 : &toml_edit::ser::to_string_pretty(&location_conf).expect("Config serialization failed");
3719 :
3720 : let temp_path = path_with_suffix_extension(config_path, TEMP_FILE_SUFFIX);
3721 :
3722 : let conf_content = conf_content.into_bytes();
3723 : VirtualFile::crashsafe_overwrite(config_path.to_owned(), temp_path, conf_content).await
3724 : }
3725 :
3726 : //
3727 : // How garbage collection works:
3728 : //
3729 : // +--bar------------->
3730 : // /
3731 : // +----+-----foo---------------->
3732 : // /
3733 : // ----main--+-------------------------->
3734 : // \
3735 : // +-----baz-------->
3736 : //
3737 : //
3738 : // 1. Grab 'gc_cs' mutex to prevent new timelines from being created while Timeline's
3739 : // `gc_infos` are being refreshed
3740 : // 2. Scan collected timelines, and on each timeline, make note of the
3741 : // all the points where other timelines have been branched off.
3742 : // We will refrain from removing page versions at those LSNs.
3743 : // 3. For each timeline, scan all layer files on the timeline.
3744 : // Remove all files for which a newer file exists and which
3745 : // don't cover any branch point LSNs.
3746 : //
3747 : // TODO:
3748 : // - if a relation has a non-incremental persistent layer on a child branch, then we
3749 : // don't need to keep that in the parent anymore. But currently
3750 : // we do.
3751 4 : async fn gc_iteration_internal(
3752 4 : &self,
3753 4 : target_timeline_id: Option<TimelineId>,
3754 4 : horizon: u64,
3755 4 : pitr: Duration,
3756 4 : cancel: &CancellationToken,
3757 4 : ctx: &RequestContext,
3758 4 : ) -> Result<GcResult, GcError> {
3759 4 : let mut totals: GcResult = Default::default();
3760 4 : let now = Instant::now();
3761 :
3762 4 : let gc_timelines = self
3763 4 : .refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
3764 4 : .await?;
3765 :
3766 4 : failpoint_support::sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
3767 :
3768 : // If there is nothing to GC, we don't want any messages in the INFO log.
3769 4 : if !gc_timelines.is_empty() {
3770 4 : info!("{} timelines need GC", gc_timelines.len());
3771 : } else {
3772 0 : debug!("{} timelines need GC", gc_timelines.len());
3773 : }
3774 :
3775 : // Perform GC for each timeline.
3776 : //
3777 : // Note that we don't hold the `Tenant::gc_cs` lock here because we don't want to delay the
3778 : // branch creation task, which requires the GC lock. A GC iteration can run concurrently
3779 : // with branch creation.
3780 : //
3781 : // See comments in [`Tenant::branch_timeline`] for more information about why branch
3782 : // creation task can run concurrently with timeline's GC iteration.
3783 8 : for timeline in gc_timelines {
3784 4 : if cancel.is_cancelled() {
3785 : // We were requested to shut down. Stop and return with the progress we
3786 : // made.
3787 0 : break;
3788 4 : }
3789 4 : let result = match timeline.gc().await {
3790 : Err(GcError::TimelineCancelled) => {
3791 0 : if target_timeline_id.is_some() {
3792 : // If we were targetting this specific timeline, surface cancellation to caller
3793 0 : return Err(GcError::TimelineCancelled);
3794 : } else {
3795 : // A timeline may be shutting down independently of the tenant's lifecycle: we should
3796 : // skip past this and proceed to try GC on other timelines.
3797 0 : continue;
3798 : }
3799 : }
3800 4 : r => r?,
3801 : };
3802 4 : totals += result;
3803 : }
3804 :
3805 4 : totals.elapsed = now.elapsed();
3806 4 : Ok(totals)
3807 4 : }
3808 :
3809 : /// Refreshes the Timeline::gc_info for all timelines, returning the
3810 : /// vector of timelines which have [`Timeline::get_last_record_lsn`] past
3811 : /// [`Tenant::get_gc_horizon`].
3812 : ///
3813 : /// This is usually executed as part of periodic gc, but can now be triggered more often.
3814 0 : pub(crate) async fn refresh_gc_info(
3815 0 : &self,
3816 0 : cancel: &CancellationToken,
3817 0 : ctx: &RequestContext,
3818 0 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
3819 0 : // since this method can now be called at different rates than the configured gc loop, it
3820 0 : // might be that these configuration values get applied faster than what it was previously,
3821 0 : // since these were only read from the gc task.
3822 0 : let horizon = self.get_gc_horizon();
3823 0 : let pitr = self.get_pitr_interval();
3824 0 :
3825 0 : // refresh all timelines
3826 0 : let target_timeline_id = None;
3827 0 :
3828 0 : self.refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
3829 0 : .await
3830 0 : }
3831 :
3832 : /// Populate all Timelines' `GcInfo` with information about their children. We do not set the
3833 : /// PITR cutoffs here, because that requires I/O: this is done later, before GC, by [`Self::refresh_gc_info_internal`]
3834 : ///
3835 : /// Subsequently, parent-child relationships are updated incrementally inside [`Timeline::new`] and [`Timeline::drop`].
3836 0 : fn initialize_gc_info(
3837 0 : &self,
3838 0 : timelines: &std::sync::MutexGuard<HashMap<TimelineId, Arc<Timeline>>>,
3839 0 : timelines_offloaded: &std::sync::MutexGuard<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
3840 0 : restrict_to_timeline: Option<TimelineId>,
3841 0 : ) {
3842 0 : if restrict_to_timeline.is_none() {
3843 : // This function must be called before activation: after activation timeline create/delete operations
3844 : // might happen, and this function is not safe to run concurrently with those.
3845 0 : assert!(!self.is_active());
3846 0 : }
3847 :
3848 : // Scan all timelines. For each timeline, remember the timeline ID and
3849 : // the branch point where it was created.
3850 0 : let mut all_branchpoints: BTreeMap<TimelineId, Vec<(Lsn, TimelineId, MaybeOffloaded)>> =
3851 0 : BTreeMap::new();
3852 0 : timelines.iter().for_each(|(timeline_id, timeline_entry)| {
3853 0 : if let Some(ancestor_timeline_id) = &timeline_entry.get_ancestor_timeline_id() {
3854 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
3855 0 : ancestor_children.push((
3856 0 : timeline_entry.get_ancestor_lsn(),
3857 0 : *timeline_id,
3858 0 : MaybeOffloaded::No,
3859 0 : ));
3860 0 : }
3861 0 : });
3862 0 : timelines_offloaded
3863 0 : .iter()
3864 0 : .for_each(|(timeline_id, timeline_entry)| {
3865 0 : let Some(ancestor_timeline_id) = &timeline_entry.ancestor_timeline_id else {
3866 0 : return;
3867 : };
3868 0 : let Some(retain_lsn) = timeline_entry.ancestor_retain_lsn else {
3869 0 : return;
3870 : };
3871 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
3872 0 : ancestor_children.push((retain_lsn, *timeline_id, MaybeOffloaded::Yes));
3873 0 : });
3874 0 :
3875 0 : // The number of bytes we always keep, irrespective of PITR: this is a constant across timelines
3876 0 : let horizon = self.get_gc_horizon();
3877 :
3878 : // Populate each timeline's GcInfo with information about its child branches
3879 0 : let timelines_to_write = if let Some(timeline_id) = restrict_to_timeline {
3880 0 : itertools::Either::Left(timelines.get(&timeline_id).into_iter())
3881 : } else {
3882 0 : itertools::Either::Right(timelines.values())
3883 : };
3884 0 : for timeline in timelines_to_write {
3885 0 : let mut branchpoints: Vec<(Lsn, TimelineId, MaybeOffloaded)> = all_branchpoints
3886 0 : .remove(&timeline.timeline_id)
3887 0 : .unwrap_or_default();
3888 0 :
3889 0 : branchpoints.sort_by_key(|b| b.0);
3890 0 :
3891 0 : let mut target = timeline.gc_info.write().unwrap();
3892 0 :
3893 0 : target.retain_lsns = branchpoints;
3894 0 :
3895 0 : let space_cutoff = timeline
3896 0 : .get_last_record_lsn()
3897 0 : .checked_sub(horizon)
3898 0 : .unwrap_or(Lsn(0));
3899 0 :
3900 0 : target.cutoffs = GcCutoffs {
3901 0 : space: space_cutoff,
3902 0 : time: Lsn::INVALID,
3903 0 : };
3904 0 : }
3905 0 : }
3906 :
3907 4 : async fn refresh_gc_info_internal(
3908 4 : &self,
3909 4 : target_timeline_id: Option<TimelineId>,
3910 4 : horizon: u64,
3911 4 : pitr: Duration,
3912 4 : cancel: &CancellationToken,
3913 4 : ctx: &RequestContext,
3914 4 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
3915 4 : // before taking the gc_cs lock, do the heavier weight finding of gc_cutoff points for
3916 4 : // currently visible timelines.
3917 4 : let timelines = self
3918 4 : .timelines
3919 4 : .lock()
3920 4 : .unwrap()
3921 4 : .values()
3922 4 : .filter(|tl| match target_timeline_id.as_ref() {
3923 4 : Some(target) => &tl.timeline_id == target,
3924 0 : None => true,
3925 4 : })
3926 4 : .cloned()
3927 4 : .collect::<Vec<_>>();
3928 4 :
3929 4 : if target_timeline_id.is_some() && timelines.is_empty() {
3930 : // We were to act on a particular timeline and it wasn't found
3931 0 : return Err(GcError::TimelineNotFound);
3932 4 : }
3933 4 :
3934 4 : let mut gc_cutoffs: HashMap<TimelineId, GcCutoffs> =
3935 4 : HashMap::with_capacity(timelines.len());
3936 :
3937 4 : for timeline in timelines.iter() {
3938 4 : let cutoff = timeline
3939 4 : .get_last_record_lsn()
3940 4 : .checked_sub(horizon)
3941 4 : .unwrap_or(Lsn(0));
3942 :
3943 4 : let cutoffs = timeline.find_gc_cutoffs(cutoff, pitr, cancel, ctx).await?;
3944 4 : let old = gc_cutoffs.insert(timeline.timeline_id, cutoffs);
3945 4 : assert!(old.is_none());
3946 : }
3947 :
3948 4 : if !self.is_active() || self.cancel.is_cancelled() {
3949 0 : return Err(GcError::TenantCancelled);
3950 4 : }
3951 :
3952 : // grab mutex to prevent new timelines from being created here; avoid doing long operations
3953 : // because that will stall branch creation.
3954 4 : let gc_cs = self.gc_cs.lock().await;
3955 :
3956 : // Ok, we now know all the branch points.
3957 : // Update the GC information for each timeline.
3958 4 : let mut gc_timelines = Vec::with_capacity(timelines.len());
3959 8 : for timeline in timelines {
3960 : // We filtered the timeline list above
3961 4 : if let Some(target_timeline_id) = target_timeline_id {
3962 4 : assert_eq!(target_timeline_id, timeline.timeline_id);
3963 0 : }
3964 :
3965 : {
3966 4 : let mut target = timeline.gc_info.write().unwrap();
3967 4 :
3968 4 : // Cull any expired leases
3969 4 : let now = SystemTime::now();
3970 6 : target.leases.retain(|_, lease| !lease.is_expired(&now));
3971 4 :
3972 4 : timeline
3973 4 : .metrics
3974 4 : .valid_lsn_lease_count_gauge
3975 4 : .set(target.leases.len() as u64);
3976 :
3977 : // Look up parent's PITR cutoff to update the child's knowledge of whether it is within parent's PITR
3978 4 : if let Some(ancestor_id) = timeline.get_ancestor_timeline_id() {
3979 0 : if let Some(ancestor_gc_cutoffs) = gc_cutoffs.get(&ancestor_id) {
3980 0 : target.within_ancestor_pitr =
3981 0 : timeline.get_ancestor_lsn() >= ancestor_gc_cutoffs.time;
3982 0 : }
3983 4 : }
3984 :
3985 : // Update metrics that depend on GC state
3986 4 : timeline
3987 4 : .metrics
3988 4 : .archival_size
3989 4 : .set(if target.within_ancestor_pitr {
3990 0 : timeline.metrics.current_logical_size_gauge.get()
3991 : } else {
3992 4 : 0
3993 : });
3994 4 : timeline.metrics.pitr_history_size.set(
3995 4 : timeline
3996 4 : .get_last_record_lsn()
3997 4 : .checked_sub(target.cutoffs.time)
3998 4 : .unwrap_or(Lsn(0))
3999 4 : .0,
4000 4 : );
4001 :
4002 : // Apply the cutoffs we found to the Timeline's GcInfo. Why might we _not_ have cutoffs for a timeline?
4003 : // - this timeline was created while we were finding cutoffs
4004 : // - lsn for timestamp search fails for this timeline repeatedly
4005 4 : if let Some(cutoffs) = gc_cutoffs.get(&timeline.timeline_id) {
4006 4 : target.cutoffs = cutoffs.clone();
4007 4 : }
4008 : }
4009 :
4010 4 : gc_timelines.push(timeline);
4011 : }
4012 4 : drop(gc_cs);
4013 4 : Ok(gc_timelines)
4014 4 : }
4015 :
4016 : /// A substitute for `branch_timeline` for use in unit tests.
4017 : /// The returned timeline will have state value `Active` to make various `anyhow::ensure!()`
4018 : /// calls pass, but, we do not actually call `.activate()` under the hood. So, none of the
4019 : /// timeline background tasks are launched, except the flush loop.
4020 : #[cfg(test)]
4021 232 : async fn branch_timeline_test(
4022 232 : self: &Arc<Self>,
4023 232 : src_timeline: &Arc<Timeline>,
4024 232 : dst_id: TimelineId,
4025 232 : ancestor_lsn: Option<Lsn>,
4026 232 : ctx: &RequestContext,
4027 232 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
4028 232 : let tl = self
4029 232 : .branch_timeline_impl(src_timeline, dst_id, ancestor_lsn, ctx)
4030 185 : .await?
4031 228 : .into_timeline_for_test();
4032 228 : tl.set_state(TimelineState::Active);
4033 228 : Ok(tl)
4034 232 : }
4035 :
4036 : /// Helper for unit tests to branch a timeline with some pre-loaded states.
4037 : #[cfg(test)]
4038 : #[allow(clippy::too_many_arguments)]
4039 6 : pub async fn branch_timeline_test_with_layers(
4040 6 : self: &Arc<Self>,
4041 6 : src_timeline: &Arc<Timeline>,
4042 6 : dst_id: TimelineId,
4043 6 : ancestor_lsn: Option<Lsn>,
4044 6 : ctx: &RequestContext,
4045 6 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
4046 6 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
4047 6 : end_lsn: Lsn,
4048 6 : ) -> anyhow::Result<Arc<Timeline>> {
4049 : use checks::check_valid_layermap;
4050 : use itertools::Itertools;
4051 :
4052 6 : let tline = self
4053 6 : .branch_timeline_test(src_timeline, dst_id, ancestor_lsn, ctx)
4054 5 : .await?;
4055 6 : let ancestor_lsn = if let Some(ancestor_lsn) = ancestor_lsn {
4056 6 : ancestor_lsn
4057 : } else {
4058 0 : tline.get_last_record_lsn()
4059 : };
4060 6 : assert!(end_lsn >= ancestor_lsn);
4061 6 : tline.force_advance_lsn(end_lsn);
4062 12 : for deltas in delta_layer_desc {
4063 6 : tline
4064 6 : .force_create_delta_layer(deltas, Some(ancestor_lsn), ctx)
4065 18 : .await?;
4066 : }
4067 10 : for (lsn, images) in image_layer_desc {
4068 4 : tline
4069 4 : .force_create_image_layer(lsn, images, Some(ancestor_lsn), ctx)
4070 14 : .await?;
4071 : }
4072 6 : let layer_names = tline
4073 6 : .layers
4074 6 : .read()
4075 0 : .await
4076 6 : .layer_map()
4077 6 : .unwrap()
4078 6 : .iter_historic_layers()
4079 10 : .map(|layer| layer.layer_name())
4080 6 : .collect_vec();
4081 6 : if let Some(err) = check_valid_layermap(&layer_names) {
4082 0 : bail!("invalid layermap: {err}");
4083 6 : }
4084 6 : Ok(tline)
4085 6 : }
4086 :
4087 : /// Branch an existing timeline.
4088 0 : async fn branch_timeline(
4089 0 : self: &Arc<Self>,
4090 0 : src_timeline: &Arc<Timeline>,
4091 0 : dst_id: TimelineId,
4092 0 : start_lsn: Option<Lsn>,
4093 0 : ctx: &RequestContext,
4094 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4095 0 : self.branch_timeline_impl(src_timeline, dst_id, start_lsn, ctx)
4096 0 : .await
4097 0 : }
4098 :
4099 232 : async fn branch_timeline_impl(
4100 232 : self: &Arc<Self>,
4101 232 : src_timeline: &Arc<Timeline>,
4102 232 : dst_id: TimelineId,
4103 232 : start_lsn: Option<Lsn>,
4104 232 : _ctx: &RequestContext,
4105 232 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4106 232 : let src_id = src_timeline.timeline_id;
4107 :
4108 : // We will validate our ancestor LSN in this function. Acquire the GC lock so that
4109 : // this check cannot race with GC, and the ancestor LSN is guaranteed to remain
4110 : // valid while we are creating the branch.
4111 232 : let _gc_cs = self.gc_cs.lock().await;
4112 :
4113 : // If no start LSN is specified, we branch the new timeline from the source timeline's last record LSN
4114 232 : let start_lsn = start_lsn.unwrap_or_else(|| {
4115 2 : let lsn = src_timeline.get_last_record_lsn();
4116 2 : info!("branching timeline {dst_id} from timeline {src_id} at last record LSN: {lsn}");
4117 2 : lsn
4118 232 : });
4119 :
4120 : // we finally have determined the ancestor_start_lsn, so we can get claim exclusivity now
4121 232 : let timeline_create_guard = match self
4122 232 : .start_creating_timeline(
4123 232 : dst_id,
4124 232 : CreateTimelineIdempotency::Branch {
4125 232 : ancestor_timeline_id: src_timeline.timeline_id,
4126 232 : ancestor_start_lsn: start_lsn,
4127 232 : },
4128 232 : )
4129 185 : .await?
4130 : {
4131 232 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
4132 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
4133 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
4134 : }
4135 : };
4136 :
4137 : // Ensure that `start_lsn` is valid, i.e. the LSN is within the PITR
4138 : // horizon on the source timeline
4139 : //
4140 : // We check it against both the planned GC cutoff stored in 'gc_info',
4141 : // and the 'latest_gc_cutoff' of the last GC that was performed. The
4142 : // planned GC cutoff in 'gc_info' is normally larger than
4143 : // 'latest_gc_cutoff_lsn', but beware of corner cases like if you just
4144 : // changed the GC settings for the tenant to make the PITR window
4145 : // larger, but some of the data was already removed by an earlier GC
4146 : // iteration.
4147 :
4148 : // check against last actual 'latest_gc_cutoff' first
4149 232 : let latest_gc_cutoff_lsn = src_timeline.get_latest_gc_cutoff_lsn();
4150 232 : src_timeline
4151 232 : .check_lsn_is_in_scope(start_lsn, &latest_gc_cutoff_lsn)
4152 232 : .context(format!(
4153 232 : "invalid branch start lsn: less than latest GC cutoff {}",
4154 232 : *latest_gc_cutoff_lsn,
4155 232 : ))
4156 232 : .map_err(CreateTimelineError::AncestorLsn)?;
4157 :
4158 : // and then the planned GC cutoff
4159 : {
4160 228 : let gc_info = src_timeline.gc_info.read().unwrap();
4161 228 : let cutoff = gc_info.min_cutoff();
4162 228 : if start_lsn < cutoff {
4163 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
4164 0 : "invalid branch start lsn: less than planned GC cutoff {cutoff}"
4165 0 : )));
4166 228 : }
4167 228 : }
4168 228 :
4169 228 : //
4170 228 : // The branch point is valid, and we are still holding the 'gc_cs' lock
4171 228 : // so that GC cannot advance the GC cutoff until we are finished.
4172 228 : // Proceed with the branch creation.
4173 228 : //
4174 228 :
4175 228 : // Determine prev-LSN for the new timeline. We can only determine it if
4176 228 : // the timeline was branched at the current end of the source timeline.
4177 228 : let RecordLsn {
4178 228 : last: src_last,
4179 228 : prev: src_prev,
4180 228 : } = src_timeline.get_last_record_rlsn();
4181 228 : let dst_prev = if src_last == start_lsn {
4182 216 : Some(src_prev)
4183 : } else {
4184 12 : None
4185 : };
4186 :
4187 : // Create the metadata file, noting the ancestor of the new timeline.
4188 : // There is initially no data in it, but all the read-calls know to look
4189 : // into the ancestor.
4190 228 : let metadata = TimelineMetadata::new(
4191 228 : start_lsn,
4192 228 : dst_prev,
4193 228 : Some(src_id),
4194 228 : start_lsn,
4195 228 : *src_timeline.latest_gc_cutoff_lsn.read(), // FIXME: should we hold onto this guard longer?
4196 228 : src_timeline.initdb_lsn,
4197 228 : src_timeline.pg_version,
4198 228 : );
4199 :
4200 228 : let uninitialized_timeline = self
4201 228 : .prepare_new_timeline(
4202 228 : dst_id,
4203 228 : &metadata,
4204 228 : timeline_create_guard,
4205 228 : start_lsn + 1,
4206 228 : Some(Arc::clone(src_timeline)),
4207 228 : )
4208 0 : .await?;
4209 :
4210 228 : let new_timeline = uninitialized_timeline.finish_creation()?;
4211 :
4212 : // Root timeline gets its layers during creation and uploads them along with the metadata.
4213 : // A branch timeline though, when created, can get no writes for some time, hence won't get any layers created.
4214 : // We still need to upload its metadata eagerly: if other nodes `attach` the tenant and miss this timeline, their GC
4215 : // could get incorrect information and remove more layers, than needed.
4216 : // See also https://github.com/neondatabase/neon/issues/3865
4217 228 : new_timeline
4218 228 : .remote_client
4219 228 : .schedule_index_upload_for_full_metadata_update(&metadata)
4220 228 : .context("branch initial metadata upload")?;
4221 :
4222 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
4223 :
4224 228 : Ok(CreateTimelineResult::Created(new_timeline))
4225 232 : }
4226 :
4227 : /// For unit tests, make this visible so that other modules can directly create timelines
4228 : #[cfg(test)]
4229 2 : #[tracing::instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
4230 : pub(crate) async fn bootstrap_timeline_test(
4231 : self: &Arc<Self>,
4232 : timeline_id: TimelineId,
4233 : pg_version: u32,
4234 : load_existing_initdb: Option<TimelineId>,
4235 : ctx: &RequestContext,
4236 : ) -> anyhow::Result<Arc<Timeline>> {
4237 : self.bootstrap_timeline(timeline_id, pg_version, load_existing_initdb, ctx)
4238 : .await
4239 : .map_err(anyhow::Error::new)
4240 2 : .map(|r| r.into_timeline_for_test())
4241 : }
4242 :
4243 : /// Get exclusive access to the timeline ID for creation.
4244 : ///
4245 : /// Timeline-creating code paths must use this function before making changes
4246 : /// to in-memory or persistent state.
4247 : ///
4248 : /// The `state` parameter is a description of the timeline creation operation
4249 : /// we intend to perform.
4250 : /// If the timeline was already created in the meantime, we check whether this
4251 : /// request conflicts or is idempotent , based on `state`.
4252 418 : async fn start_creating_timeline(
4253 418 : &self,
4254 418 : new_timeline_id: TimelineId,
4255 418 : idempotency: CreateTimelineIdempotency,
4256 418 : ) -> Result<StartCreatingTimelineResult<'_>, CreateTimelineError> {
4257 418 : let allow_offloaded = false;
4258 418 : match self.create_timeline_create_guard(new_timeline_id, idempotency, allow_offloaded) {
4259 416 : Ok(create_guard) => {
4260 416 : pausable_failpoint!("timeline-creation-after-uninit");
4261 416 : Ok(StartCreatingTimelineResult::CreateGuard(create_guard))
4262 : }
4263 : Err(TimelineExclusionError::AlreadyCreating) => {
4264 : // Creation is in progress, we cannot create it again, and we cannot
4265 : // check if this request matches the existing one, so caller must try
4266 : // again later.
4267 0 : Err(CreateTimelineError::AlreadyCreating)
4268 : }
4269 0 : Err(TimelineExclusionError::Other(e)) => Err(CreateTimelineError::Other(e)),
4270 : Err(TimelineExclusionError::AlreadyExists {
4271 0 : existing: TimelineOrOffloaded::Offloaded(_existing),
4272 0 : ..
4273 0 : }) => {
4274 0 : info!("timeline already exists but is offloaded");
4275 0 : Err(CreateTimelineError::Conflict)
4276 : }
4277 : Err(TimelineExclusionError::AlreadyExists {
4278 2 : existing: TimelineOrOffloaded::Timeline(existing),
4279 2 : arg,
4280 2 : }) => {
4281 2 : {
4282 2 : let existing = &existing.create_idempotency;
4283 2 : let _span = info_span!("idempotency_check", ?existing, ?arg).entered();
4284 2 : debug!("timeline already exists");
4285 :
4286 2 : match (existing, &arg) {
4287 : // FailWithConflict => no idempotency check
4288 : (CreateTimelineIdempotency::FailWithConflict, _)
4289 : | (_, CreateTimelineIdempotency::FailWithConflict) => {
4290 2 : warn!("timeline already exists, failing request");
4291 2 : return Err(CreateTimelineError::Conflict);
4292 : }
4293 : // Idempotent <=> CreateTimelineIdempotency is identical
4294 0 : (x, y) if x == y => {
4295 0 : info!("timeline already exists and idempotency matches, succeeding request");
4296 : // fallthrough
4297 : }
4298 : (_, _) => {
4299 0 : warn!("idempotency conflict, failing request");
4300 0 : return Err(CreateTimelineError::Conflict);
4301 : }
4302 : }
4303 : }
4304 :
4305 0 : Ok(StartCreatingTimelineResult::Idempotent(existing))
4306 : }
4307 : }
4308 418 : }
4309 :
4310 0 : async fn upload_initdb(
4311 0 : &self,
4312 0 : timelines_path: &Utf8PathBuf,
4313 0 : pgdata_path: &Utf8PathBuf,
4314 0 : timeline_id: &TimelineId,
4315 0 : ) -> anyhow::Result<()> {
4316 0 : let temp_path = timelines_path.join(format!(
4317 0 : "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
4318 0 : ));
4319 0 :
4320 0 : scopeguard::defer! {
4321 0 : if let Err(e) = fs::remove_file(&temp_path) {
4322 0 : error!("Failed to remove temporary initdb archive '{temp_path}': {e}");
4323 0 : }
4324 0 : }
4325 :
4326 0 : let (pgdata_zstd, tar_zst_size) = create_zst_tarball(pgdata_path, &temp_path).await?;
4327 : const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
4328 0 : if tar_zst_size > INITDB_TAR_ZST_WARN_LIMIT {
4329 0 : warn!(
4330 0 : "compressed {temp_path} size of {tar_zst_size} is above limit {INITDB_TAR_ZST_WARN_LIMIT}."
4331 : );
4332 0 : }
4333 :
4334 0 : pausable_failpoint!("before-initdb-upload");
4335 :
4336 0 : backoff::retry(
4337 0 : || async {
4338 0 : self::remote_timeline_client::upload_initdb_dir(
4339 0 : &self.remote_storage,
4340 0 : &self.tenant_shard_id.tenant_id,
4341 0 : timeline_id,
4342 0 : pgdata_zstd.try_clone().await?,
4343 0 : tar_zst_size,
4344 0 : &self.cancel,
4345 : )
4346 0 : .await
4347 0 : },
4348 0 : |_| false,
4349 0 : 3,
4350 0 : u32::MAX,
4351 0 : "persist_initdb_tar_zst",
4352 0 : &self.cancel,
4353 0 : )
4354 0 : .await
4355 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
4356 0 : .and_then(|x| x)
4357 0 : }
4358 :
4359 : /// - run initdb to init temporary instance and get bootstrap data
4360 : /// - after initialization completes, tar up the temp dir and upload it to S3.
4361 2 : async fn bootstrap_timeline(
4362 2 : self: &Arc<Self>,
4363 2 : timeline_id: TimelineId,
4364 2 : pg_version: u32,
4365 2 : load_existing_initdb: Option<TimelineId>,
4366 2 : ctx: &RequestContext,
4367 2 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4368 2 : let timeline_create_guard = match self
4369 2 : .start_creating_timeline(
4370 2 : timeline_id,
4371 2 : CreateTimelineIdempotency::Bootstrap { pg_version },
4372 2 : )
4373 1 : .await?
4374 : {
4375 2 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
4376 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
4377 0 : return Ok(CreateTimelineResult::Idempotent(timeline))
4378 : }
4379 : };
4380 :
4381 : // create a `tenant/{tenant_id}/timelines/basebackup-{timeline_id}.{TEMP_FILE_SUFFIX}/`
4382 : // temporary directory for basebackup files for the given timeline.
4383 :
4384 2 : let timelines_path = self.conf.timelines_path(&self.tenant_shard_id);
4385 2 : let pgdata_path = path_with_suffix_extension(
4386 2 : timelines_path.join(format!("basebackup-{timeline_id}")),
4387 2 : TEMP_FILE_SUFFIX,
4388 2 : );
4389 2 :
4390 2 : // Remove whatever was left from the previous runs: safe because TimelineCreateGuard guarantees
4391 2 : // we won't race with other creations or existent timelines with the same path.
4392 2 : if pgdata_path.exists() {
4393 0 : fs::remove_dir_all(&pgdata_path).with_context(|| {
4394 0 : format!("Failed to remove already existing initdb directory: {pgdata_path}")
4395 0 : })?;
4396 2 : }
4397 :
4398 : // this new directory is very temporary, set to remove it immediately after bootstrap, we don't need it
4399 2 : scopeguard::defer! {
4400 2 : if let Err(e) = fs::remove_dir_all(&pgdata_path) {
4401 2 : // this is unlikely, but we will remove the directory on pageserver restart or another bootstrap call
4402 2 : error!("Failed to remove temporary initdb directory '{pgdata_path}': {e}");
4403 2 : }
4404 2 : }
4405 2 : if let Some(existing_initdb_timeline_id) = load_existing_initdb {
4406 2 : if existing_initdb_timeline_id != timeline_id {
4407 0 : let source_path = &remote_initdb_archive_path(
4408 0 : &self.tenant_shard_id.tenant_id,
4409 0 : &existing_initdb_timeline_id,
4410 0 : );
4411 0 : let dest_path =
4412 0 : &remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &timeline_id);
4413 0 :
4414 0 : // if this fails, it will get retried by retried control plane requests
4415 0 : self.remote_storage
4416 0 : .copy_object(source_path, dest_path, &self.cancel)
4417 0 : .await
4418 0 : .context("copy initdb tar")?;
4419 2 : }
4420 2 : let (initdb_tar_zst_path, initdb_tar_zst) =
4421 2 : self::remote_timeline_client::download_initdb_tar_zst(
4422 2 : self.conf,
4423 2 : &self.remote_storage,
4424 2 : &self.tenant_shard_id,
4425 2 : &existing_initdb_timeline_id,
4426 2 : &self.cancel,
4427 2 : )
4428 268 : .await
4429 2 : .context("download initdb tar")?;
4430 :
4431 2 : scopeguard::defer! {
4432 2 : if let Err(e) = fs::remove_file(&initdb_tar_zst_path) {
4433 2 : error!("Failed to remove temporary initdb archive '{initdb_tar_zst_path}': {e}");
4434 2 : }
4435 2 : }
4436 2 :
4437 2 : let buf_read =
4438 2 : BufReader::with_capacity(remote_timeline_client::BUFFER_SIZE, initdb_tar_zst);
4439 2 : extract_zst_tarball(&pgdata_path, buf_read)
4440 9512 : .await
4441 2 : .context("extract initdb tar")?;
4442 : } else {
4443 : // Init temporarily repo to get bootstrap data, this creates a directory in the `pgdata_path` path
4444 0 : run_initdb(self.conf, &pgdata_path, pg_version, &self.cancel)
4445 0 : .await
4446 0 : .context("run initdb")?;
4447 :
4448 : // Upload the created data dir to S3
4449 0 : if self.tenant_shard_id().is_shard_zero() {
4450 0 : self.upload_initdb(&timelines_path, &pgdata_path, &timeline_id)
4451 0 : .await?;
4452 0 : }
4453 : }
4454 2 : let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
4455 2 :
4456 2 : // Import the contents of the data directory at the initial checkpoint
4457 2 : // LSN, and any WAL after that.
4458 2 : // Initdb lsn will be equal to last_record_lsn which will be set after import.
4459 2 : // Because we know it upfront avoid having an option or dummy zero value by passing it to the metadata.
4460 2 : let new_metadata = TimelineMetadata::new(
4461 2 : Lsn(0),
4462 2 : None,
4463 2 : None,
4464 2 : Lsn(0),
4465 2 : pgdata_lsn,
4466 2 : pgdata_lsn,
4467 2 : pg_version,
4468 2 : );
4469 2 : let raw_timeline = self
4470 2 : .prepare_new_timeline(
4471 2 : timeline_id,
4472 2 : &new_metadata,
4473 2 : timeline_create_guard,
4474 2 : pgdata_lsn,
4475 2 : None,
4476 2 : )
4477 0 : .await?;
4478 :
4479 2 : let tenant_shard_id = raw_timeline.owning_tenant.tenant_shard_id;
4480 2 : let unfinished_timeline = raw_timeline.raw_timeline()?;
4481 :
4482 : // Flush the new layer files to disk, before we make the timeline as available to
4483 : // the outside world.
4484 : //
4485 : // Flush loop needs to be spawned in order to be able to flush.
4486 2 : unfinished_timeline.maybe_spawn_flush_loop();
4487 2 :
4488 2 : import_datadir::import_timeline_from_postgres_datadir(
4489 2 : unfinished_timeline,
4490 2 : &pgdata_path,
4491 2 : pgdata_lsn,
4492 2 : ctx,
4493 2 : )
4494 7880 : .await
4495 2 : .with_context(|| {
4496 0 : format!("Failed to import pgdatadir for timeline {tenant_shard_id}/{timeline_id}")
4497 2 : })?;
4498 :
4499 2 : fail::fail_point!("before-checkpoint-new-timeline", |_| {
4500 0 : Err(CreateTimelineError::Other(anyhow::anyhow!(
4501 0 : "failpoint before-checkpoint-new-timeline"
4502 0 : )))
4503 2 : });
4504 :
4505 2 : unfinished_timeline
4506 2 : .freeze_and_flush()
4507 2 : .await
4508 2 : .with_context(|| {
4509 0 : format!(
4510 0 : "Failed to flush after pgdatadir import for timeline {tenant_shard_id}/{timeline_id}"
4511 0 : )
4512 2 : })?;
4513 :
4514 : // All done!
4515 2 : let timeline = raw_timeline.finish_creation()?;
4516 :
4517 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
4518 :
4519 2 : Ok(CreateTimelineResult::Created(timeline))
4520 2 : }
4521 :
4522 412 : fn build_timeline_remote_client(&self, timeline_id: TimelineId) -> RemoteTimelineClient {
4523 412 : RemoteTimelineClient::new(
4524 412 : self.remote_storage.clone(),
4525 412 : self.deletion_queue_client.clone(),
4526 412 : self.conf,
4527 412 : self.tenant_shard_id,
4528 412 : timeline_id,
4529 412 : self.generation,
4530 412 : )
4531 412 : }
4532 :
4533 : /// Call this before constructing a timeline, to build its required structures
4534 412 : fn build_timeline_resources(&self, timeline_id: TimelineId) -> TimelineResources {
4535 412 : TimelineResources {
4536 412 : remote_client: self.build_timeline_remote_client(timeline_id),
4537 412 : timeline_get_throttle: self.timeline_get_throttle.clone(),
4538 412 : l0_flush_global_state: self.l0_flush_global_state.clone(),
4539 412 : }
4540 412 : }
4541 :
4542 : /// Creates intermediate timeline structure and its files.
4543 : ///
4544 : /// An empty layer map is initialized, and new data and WAL can be imported starting
4545 : /// at 'disk_consistent_lsn'. After any initial data has been imported, call
4546 : /// `finish_creation` to insert the Timeline into the timelines map.
4547 412 : async fn prepare_new_timeline<'a>(
4548 412 : &'a self,
4549 412 : new_timeline_id: TimelineId,
4550 412 : new_metadata: &TimelineMetadata,
4551 412 : create_guard: TimelineCreateGuard<'a>,
4552 412 : start_lsn: Lsn,
4553 412 : ancestor: Option<Arc<Timeline>>,
4554 412 : ) -> anyhow::Result<UninitializedTimeline<'a>> {
4555 412 : let tenant_shard_id = self.tenant_shard_id;
4556 412 :
4557 412 : let resources = self.build_timeline_resources(new_timeline_id);
4558 412 : resources
4559 412 : .remote_client
4560 412 : .init_upload_queue_for_empty_remote(new_metadata)?;
4561 :
4562 412 : let timeline_struct = self
4563 412 : .create_timeline_struct(
4564 412 : new_timeline_id,
4565 412 : new_metadata,
4566 412 : ancestor,
4567 412 : resources,
4568 412 : CreateTimelineCause::Load,
4569 412 : create_guard.idempotency.clone(),
4570 412 : )
4571 412 : .context("Failed to create timeline data structure")?;
4572 :
4573 412 : timeline_struct.init_empty_layer_map(start_lsn);
4574 :
4575 412 : if let Err(e) = self
4576 412 : .create_timeline_files(&create_guard.timeline_path)
4577 0 : .await
4578 : {
4579 0 : error!("Failed to create initial files for timeline {tenant_shard_id}/{new_timeline_id}, cleaning up: {e:?}");
4580 0 : cleanup_timeline_directory(create_guard);
4581 0 : return Err(e);
4582 412 : }
4583 412 :
4584 412 : debug!(
4585 0 : "Successfully created initial files for timeline {tenant_shard_id}/{new_timeline_id}"
4586 : );
4587 :
4588 412 : Ok(UninitializedTimeline::new(
4589 412 : self,
4590 412 : new_timeline_id,
4591 412 : Some((timeline_struct, create_guard)),
4592 412 : ))
4593 412 : }
4594 :
4595 412 : async fn create_timeline_files(&self, timeline_path: &Utf8Path) -> anyhow::Result<()> {
4596 412 : crashsafe::create_dir(timeline_path).context("Failed to create timeline directory")?;
4597 :
4598 412 : fail::fail_point!("after-timeline-dir-creation", |_| {
4599 0 : anyhow::bail!("failpoint after-timeline-dir-creation");
4600 412 : });
4601 :
4602 412 : Ok(())
4603 412 : }
4604 :
4605 : /// Get a guard that provides exclusive access to the timeline directory, preventing
4606 : /// concurrent attempts to create the same timeline.
4607 : ///
4608 : /// The `allow_offloaded` parameter controls whether to tolerate the existence of
4609 : /// offloaded timelines or not.
4610 418 : fn create_timeline_create_guard(
4611 418 : &self,
4612 418 : timeline_id: TimelineId,
4613 418 : idempotency: CreateTimelineIdempotency,
4614 418 : allow_offloaded: bool,
4615 418 : ) -> Result<TimelineCreateGuard, TimelineExclusionError> {
4616 418 : let tenant_shard_id = self.tenant_shard_id;
4617 418 :
4618 418 : let timeline_path = self.conf.timeline_path(&tenant_shard_id, &timeline_id);
4619 :
4620 418 : let create_guard = TimelineCreateGuard::new(
4621 418 : self,
4622 418 : timeline_id,
4623 418 : timeline_path.clone(),
4624 418 : idempotency,
4625 418 : allow_offloaded,
4626 418 : )?;
4627 :
4628 : // At this stage, we have got exclusive access to in-memory state for this timeline ID
4629 : // for creation.
4630 : // A timeline directory should never exist on disk already:
4631 : // - a previous failed creation would have cleaned up after itself
4632 : // - a pageserver restart would clean up timeline directories that don't have valid remote state
4633 : //
4634 : // Therefore it is an unexpected internal error to encounter a timeline directory already existing here,
4635 : // this error may indicate a bug in cleanup on failed creations.
4636 416 : if timeline_path.exists() {
4637 0 : return Err(TimelineExclusionError::Other(anyhow::anyhow!(
4638 0 : "Timeline directory already exists! This is a bug."
4639 0 : )));
4640 416 : }
4641 416 :
4642 416 : Ok(create_guard)
4643 418 : }
4644 :
4645 : /// Gathers inputs from all of the timelines to produce a sizing model input.
4646 : ///
4647 : /// Future is cancellation safe. Only one calculation can be running at once per tenant.
4648 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
4649 : pub async fn gather_size_inputs(
4650 : &self,
4651 : // `max_retention_period` overrides the cutoff that is used to calculate the size
4652 : // (only if it is shorter than the real cutoff).
4653 : max_retention_period: Option<u64>,
4654 : cause: LogicalSizeCalculationCause,
4655 : cancel: &CancellationToken,
4656 : ctx: &RequestContext,
4657 : ) -> Result<size::ModelInputs, size::CalculateSyntheticSizeError> {
4658 : let logical_sizes_at_once = self
4659 : .conf
4660 : .concurrent_tenant_size_logical_size_queries
4661 : .inner();
4662 :
4663 : // TODO: Having a single mutex block concurrent reads is not great for performance.
4664 : //
4665 : // But the only case where we need to run multiple of these at once is when we
4666 : // request a size for a tenant manually via API, while another background calculation
4667 : // is in progress (which is not a common case).
4668 : //
4669 : // See more for on the issue #2748 condenced out of the initial PR review.
4670 : let mut shared_cache = tokio::select! {
4671 : locked = self.cached_logical_sizes.lock() => locked,
4672 : _ = cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
4673 : _ = self.cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
4674 : };
4675 :
4676 : size::gather_inputs(
4677 : self,
4678 : logical_sizes_at_once,
4679 : max_retention_period,
4680 : &mut shared_cache,
4681 : cause,
4682 : cancel,
4683 : ctx,
4684 : )
4685 : .await
4686 : }
4687 :
4688 : /// Calculate synthetic tenant size and cache the result.
4689 : /// This is periodically called by background worker.
4690 : /// result is cached in tenant struct
4691 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
4692 : pub async fn calculate_synthetic_size(
4693 : &self,
4694 : cause: LogicalSizeCalculationCause,
4695 : cancel: &CancellationToken,
4696 : ctx: &RequestContext,
4697 : ) -> Result<u64, size::CalculateSyntheticSizeError> {
4698 : let inputs = self.gather_size_inputs(None, cause, cancel, ctx).await?;
4699 :
4700 : let size = inputs.calculate();
4701 :
4702 : self.set_cached_synthetic_size(size);
4703 :
4704 : Ok(size)
4705 : }
4706 :
4707 : /// Cache given synthetic size and update the metric value
4708 0 : pub fn set_cached_synthetic_size(&self, size: u64) {
4709 0 : self.cached_synthetic_tenant_size
4710 0 : .store(size, Ordering::Relaxed);
4711 0 :
4712 0 : // Only shard zero should be calculating synthetic sizes
4713 0 : debug_assert!(self.shard_identity.is_shard_zero());
4714 :
4715 0 : TENANT_SYNTHETIC_SIZE_METRIC
4716 0 : .get_metric_with_label_values(&[&self.tenant_shard_id.tenant_id.to_string()])
4717 0 : .unwrap()
4718 0 : .set(size);
4719 0 : }
4720 :
4721 0 : pub fn cached_synthetic_size(&self) -> u64 {
4722 0 : self.cached_synthetic_tenant_size.load(Ordering::Relaxed)
4723 0 : }
4724 :
4725 : /// Flush any in-progress layers, schedule uploads, and wait for uploads to complete.
4726 : ///
4727 : /// This function can take a long time: callers should wrap it in a timeout if calling
4728 : /// from an external API handler.
4729 : ///
4730 : /// Cancel-safety: cancelling this function may leave I/O running, but such I/O is
4731 : /// still bounded by tenant/timeline shutdown.
4732 0 : #[tracing::instrument(skip_all)]
4733 : pub(crate) async fn flush_remote(&self) -> anyhow::Result<()> {
4734 : let timelines = self.timelines.lock().unwrap().clone();
4735 :
4736 0 : async fn flush_timeline(_gate: GateGuard, timeline: Arc<Timeline>) -> anyhow::Result<()> {
4737 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Flushing...");
4738 0 : timeline.freeze_and_flush().await?;
4739 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Waiting for uploads...");
4740 0 : timeline.remote_client.wait_completion().await?;
4741 :
4742 0 : Ok(())
4743 0 : }
4744 :
4745 : // We do not use a JoinSet for these tasks, because we don't want them to be
4746 : // aborted when this function's future is cancelled: they should stay alive
4747 : // holding their GateGuard until they complete, to ensure their I/Os complete
4748 : // before Timeline shutdown completes.
4749 : let mut results = FuturesUnordered::new();
4750 :
4751 : for (_timeline_id, timeline) in timelines {
4752 : // Run each timeline's flush in a task holding the timeline's gate: this
4753 : // means that if this function's future is cancelled, the Timeline shutdown
4754 : // will still wait for any I/O in here to complete.
4755 : let Ok(gate) = timeline.gate.enter() else {
4756 : continue;
4757 : };
4758 0 : let jh = tokio::task::spawn(async move { flush_timeline(gate, timeline).await });
4759 : results.push(jh);
4760 : }
4761 :
4762 : while let Some(r) = results.next().await {
4763 : if let Err(e) = r {
4764 : if !e.is_cancelled() && !e.is_panic() {
4765 : tracing::error!("unexpected join error: {e:?}");
4766 : }
4767 : }
4768 : }
4769 :
4770 : // The flushes we did above were just writes, but the Tenant might have had
4771 : // pending deletions as well from recent compaction/gc: we want to flush those
4772 : // as well. This requires flushing the global delete queue. This is cheap
4773 : // because it's typically a no-op.
4774 : match self.deletion_queue_client.flush_execute().await {
4775 : Ok(_) => {}
4776 : Err(DeletionQueueError::ShuttingDown) => {}
4777 : }
4778 :
4779 : Ok(())
4780 : }
4781 :
4782 0 : pub(crate) fn get_tenant_conf(&self) -> TenantConfOpt {
4783 0 : self.tenant_conf.load().tenant_conf.clone()
4784 0 : }
4785 :
4786 : /// How much local storage would this tenant like to have? It can cope with
4787 : /// less than this (via eviction and on-demand downloads), but this function enables
4788 : /// the Tenant to advertise how much storage it would prefer to have to provide fast I/O
4789 : /// by keeping important things on local disk.
4790 : ///
4791 : /// This is a heuristic, not a guarantee: tenants that are long-idle will actually use less
4792 : /// than they report here, due to layer eviction. Tenants with many active branches may
4793 : /// actually use more than they report here.
4794 0 : pub(crate) fn local_storage_wanted(&self) -> u64 {
4795 0 : let timelines = self.timelines.lock().unwrap();
4796 0 :
4797 0 : // Heuristic: we use the max() of the timelines' visible sizes, rather than the sum. This
4798 0 : // reflects the observation that on tenants with multiple large branches, typically only one
4799 0 : // of them is used actively enough to occupy space on disk.
4800 0 : timelines
4801 0 : .values()
4802 0 : .map(|t| t.metrics.visible_physical_size_gauge.get())
4803 0 : .max()
4804 0 : .unwrap_or(0)
4805 0 : }
4806 :
4807 : /// Serialize and write the latest TenantManifest to remote storage.
4808 2 : pub(crate) async fn store_tenant_manifest(&self) -> Result<(), TenantManifestError> {
4809 : // Only one manifest write may be done at at time, and the contents of the manifest
4810 : // must be loaded while holding this lock. This makes it safe to call this function
4811 : // from anywhere without worrying about colliding updates.
4812 2 : let mut guard = tokio::select! {
4813 2 : g = self.tenant_manifest_upload.lock() => {
4814 2 : g
4815 : },
4816 2 : _ = self.cancel.cancelled() => {
4817 0 : return Err(TenantManifestError::Cancelled);
4818 : }
4819 : };
4820 :
4821 2 : let manifest = self.build_tenant_manifest();
4822 2 : if Some(&manifest) == (*guard).as_ref() {
4823 : // Optimisation: skip uploads that don't change anything.
4824 0 : return Ok(());
4825 2 : }
4826 2 :
4827 2 : upload_tenant_manifest(
4828 2 : &self.remote_storage,
4829 2 : &self.tenant_shard_id,
4830 2 : self.generation,
4831 2 : &manifest,
4832 2 : &self.cancel,
4833 2 : )
4834 6 : .await
4835 2 : .map_err(|e| {
4836 0 : if self.cancel.is_cancelled() {
4837 0 : TenantManifestError::Cancelled
4838 : } else {
4839 0 : TenantManifestError::RemoteStorage(e)
4840 : }
4841 2 : })?;
4842 :
4843 : // Store the successfully uploaded manifest, so that future callers can avoid
4844 : // re-uploading the same thing.
4845 2 : *guard = Some(manifest);
4846 2 :
4847 2 : Ok(())
4848 2 : }
4849 : }
4850 :
4851 : /// Create the cluster temporarily in 'initdbpath' directory inside the repository
4852 : /// to get bootstrap data for timeline initialization.
4853 0 : async fn run_initdb(
4854 0 : conf: &'static PageServerConf,
4855 0 : initdb_target_dir: &Utf8Path,
4856 0 : pg_version: u32,
4857 0 : cancel: &CancellationToken,
4858 0 : ) -> Result<(), InitdbError> {
4859 0 : let initdb_bin_path = conf
4860 0 : .pg_bin_dir(pg_version)
4861 0 : .map_err(InitdbError::Other)?
4862 0 : .join("initdb");
4863 0 : let initdb_lib_dir = conf.pg_lib_dir(pg_version).map_err(InitdbError::Other)?;
4864 0 : info!(
4865 0 : "running {} in {}, libdir: {}",
4866 : initdb_bin_path, initdb_target_dir, initdb_lib_dir,
4867 : );
4868 :
4869 0 : let _permit = INIT_DB_SEMAPHORE.acquire().await;
4870 :
4871 0 : let mut initdb_command = tokio::process::Command::new(&initdb_bin_path);
4872 0 : initdb_command
4873 0 : .args(["--pgdata", initdb_target_dir.as_ref()])
4874 0 : .args(["--username", &conf.superuser])
4875 0 : .args(["--encoding", "utf8"])
4876 0 : .args(["--locale", &conf.locale])
4877 0 : .arg("--no-instructions")
4878 0 : .arg("--no-sync")
4879 0 : .env_clear()
4880 0 : .env("LD_LIBRARY_PATH", &initdb_lib_dir)
4881 0 : .env("DYLD_LIBRARY_PATH", &initdb_lib_dir)
4882 0 : .stdin(std::process::Stdio::null())
4883 0 : // stdout invocation produces the same output every time, we don't need it
4884 0 : .stdout(std::process::Stdio::null())
4885 0 : // we would be interested in the stderr output, if there was any
4886 0 : .stderr(std::process::Stdio::piped());
4887 0 :
4888 0 : // Before version 14, only the libc provide was available.
4889 0 : if pg_version > 14 {
4890 : // Version 17 brought with it a builtin locale provider which only provides
4891 : // C and C.UTF-8. While being safer for collation purposes since it is
4892 : // guaranteed to be consistent throughout a major release, it is also more
4893 : // performant.
4894 0 : let locale_provider = if pg_version >= 17 { "builtin" } else { "libc" };
4895 :
4896 0 : initdb_command.args(["--locale-provider", locale_provider]);
4897 0 : }
4898 :
4899 0 : let initdb_proc = initdb_command.spawn()?;
4900 :
4901 : // Ideally we'd select here with the cancellation token, but the problem is that
4902 : // we can't safely terminate initdb: it launches processes of its own, and killing
4903 : // initdb doesn't kill them. After we return from this function, we want the target
4904 : // directory to be able to be cleaned up.
4905 : // See https://github.com/neondatabase/neon/issues/6385
4906 0 : let initdb_output = initdb_proc.wait_with_output().await?;
4907 0 : if !initdb_output.status.success() {
4908 0 : return Err(InitdbError::Failed(
4909 0 : initdb_output.status,
4910 0 : initdb_output.stderr,
4911 0 : ));
4912 0 : }
4913 0 :
4914 0 : // This isn't true cancellation support, see above. Still return an error to
4915 0 : // excercise the cancellation code path.
4916 0 : if cancel.is_cancelled() {
4917 0 : return Err(InitdbError::Cancelled);
4918 0 : }
4919 0 :
4920 0 : Ok(())
4921 0 : }
4922 :
4923 : /// Dump contents of a layer file to stdout.
4924 0 : pub async fn dump_layerfile_from_path(
4925 0 : path: &Utf8Path,
4926 0 : verbose: bool,
4927 0 : ctx: &RequestContext,
4928 0 : ) -> anyhow::Result<()> {
4929 : use std::os::unix::fs::FileExt;
4930 :
4931 : // All layer files start with a two-byte "magic" value, to identify the kind of
4932 : // file.
4933 0 : let file = File::open(path)?;
4934 0 : let mut header_buf = [0u8; 2];
4935 0 : file.read_exact_at(&mut header_buf, 0)?;
4936 :
4937 0 : match u16::from_be_bytes(header_buf) {
4938 : crate::IMAGE_FILE_MAGIC => {
4939 0 : ImageLayer::new_for_path(path, file)?
4940 0 : .dump(verbose, ctx)
4941 0 : .await?
4942 : }
4943 : crate::DELTA_FILE_MAGIC => {
4944 0 : DeltaLayer::new_for_path(path, file)?
4945 0 : .dump(verbose, ctx)
4946 0 : .await?
4947 : }
4948 0 : magic => bail!("unrecognized magic identifier: {:?}", magic),
4949 : }
4950 :
4951 0 : Ok(())
4952 0 : }
4953 :
4954 : #[cfg(test)]
4955 : pub(crate) mod harness {
4956 : use bytes::{Bytes, BytesMut};
4957 : use once_cell::sync::OnceCell;
4958 : use pageserver_api::models::ShardParameters;
4959 : use pageserver_api::shard::ShardIndex;
4960 : use utils::logging;
4961 :
4962 : use crate::deletion_queue::mock::MockDeletionQueue;
4963 : use crate::l0_flush::L0FlushConfig;
4964 : use crate::walredo::apply_neon;
4965 : use pageserver_api::key::Key;
4966 : use pageserver_api::record::NeonWalRecord;
4967 :
4968 : use super::*;
4969 : use hex_literal::hex;
4970 : use utils::id::TenantId;
4971 :
4972 : pub const TIMELINE_ID: TimelineId =
4973 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
4974 : pub const NEW_TIMELINE_ID: TimelineId =
4975 : TimelineId::from_array(hex!("AA223344556677881122334455667788"));
4976 :
4977 : /// Convenience function to create a page image with given string as the only content
4978 5028700 : pub fn test_img(s: &str) -> Bytes {
4979 5028700 : let mut buf = BytesMut::new();
4980 5028700 : buf.extend_from_slice(s.as_bytes());
4981 5028700 : buf.resize(64, 0);
4982 5028700 :
4983 5028700 : buf.freeze()
4984 5028700 : }
4985 :
4986 : impl From<TenantConf> for TenantConfOpt {
4987 192 : fn from(tenant_conf: TenantConf) -> Self {
4988 192 : Self {
4989 192 : checkpoint_distance: Some(tenant_conf.checkpoint_distance),
4990 192 : checkpoint_timeout: Some(tenant_conf.checkpoint_timeout),
4991 192 : compaction_target_size: Some(tenant_conf.compaction_target_size),
4992 192 : compaction_period: Some(tenant_conf.compaction_period),
4993 192 : compaction_threshold: Some(tenant_conf.compaction_threshold),
4994 192 : compaction_algorithm: Some(tenant_conf.compaction_algorithm),
4995 192 : gc_horizon: Some(tenant_conf.gc_horizon),
4996 192 : gc_period: Some(tenant_conf.gc_period),
4997 192 : image_creation_threshold: Some(tenant_conf.image_creation_threshold),
4998 192 : pitr_interval: Some(tenant_conf.pitr_interval),
4999 192 : walreceiver_connect_timeout: Some(tenant_conf.walreceiver_connect_timeout),
5000 192 : lagging_wal_timeout: Some(tenant_conf.lagging_wal_timeout),
5001 192 : max_lsn_wal_lag: Some(tenant_conf.max_lsn_wal_lag),
5002 192 : eviction_policy: Some(tenant_conf.eviction_policy),
5003 192 : min_resident_size_override: tenant_conf.min_resident_size_override,
5004 192 : evictions_low_residence_duration_metric_threshold: Some(
5005 192 : tenant_conf.evictions_low_residence_duration_metric_threshold,
5006 192 : ),
5007 192 : heatmap_period: Some(tenant_conf.heatmap_period),
5008 192 : lazy_slru_download: Some(tenant_conf.lazy_slru_download),
5009 192 : timeline_get_throttle: Some(tenant_conf.timeline_get_throttle),
5010 192 : image_layer_creation_check_threshold: Some(
5011 192 : tenant_conf.image_layer_creation_check_threshold,
5012 192 : ),
5013 192 : lsn_lease_length: Some(tenant_conf.lsn_lease_length),
5014 192 : lsn_lease_length_for_ts: Some(tenant_conf.lsn_lease_length_for_ts),
5015 192 : timeline_offloading: Some(tenant_conf.timeline_offloading),
5016 192 : }
5017 192 : }
5018 : }
5019 :
5020 : pub struct TenantHarness {
5021 : pub conf: &'static PageServerConf,
5022 : pub tenant_conf: TenantConf,
5023 : pub tenant_shard_id: TenantShardId,
5024 : pub generation: Generation,
5025 : pub shard: ShardIndex,
5026 : pub remote_storage: GenericRemoteStorage,
5027 : pub remote_fs_dir: Utf8PathBuf,
5028 : pub deletion_queue: MockDeletionQueue,
5029 : }
5030 :
5031 : static LOG_HANDLE: OnceCell<()> = OnceCell::new();
5032 :
5033 208 : pub(crate) fn setup_logging() {
5034 208 : LOG_HANDLE.get_or_init(|| {
5035 196 : logging::init(
5036 196 : logging::LogFormat::Test,
5037 196 : // enable it in case the tests exercise code paths that use
5038 196 : // debug_assert_current_span_has_tenant_and_timeline_id
5039 196 : logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
5040 196 : logging::Output::Stdout,
5041 196 : )
5042 196 : .expect("Failed to init test logging")
5043 208 : });
5044 208 : }
5045 :
5046 : impl TenantHarness {
5047 192 : pub async fn create_custom(
5048 192 : test_name: &'static str,
5049 192 : tenant_conf: TenantConf,
5050 192 : tenant_id: TenantId,
5051 192 : shard_identity: ShardIdentity,
5052 192 : generation: Generation,
5053 192 : ) -> anyhow::Result<Self> {
5054 192 : setup_logging();
5055 192 :
5056 192 : let repo_dir = PageServerConf::test_repo_dir(test_name);
5057 192 : let _ = fs::remove_dir_all(&repo_dir);
5058 192 : fs::create_dir_all(&repo_dir)?;
5059 :
5060 192 : let conf = PageServerConf::dummy_conf(repo_dir);
5061 192 : // Make a static copy of the config. This can never be free'd, but that's
5062 192 : // OK in a test.
5063 192 : let conf: &'static PageServerConf = Box::leak(Box::new(conf));
5064 192 :
5065 192 : let shard = shard_identity.shard_index();
5066 192 : let tenant_shard_id = TenantShardId {
5067 192 : tenant_id,
5068 192 : shard_number: shard.shard_number,
5069 192 : shard_count: shard.shard_count,
5070 192 : };
5071 192 : fs::create_dir_all(conf.tenant_path(&tenant_shard_id))?;
5072 192 : fs::create_dir_all(conf.timelines_path(&tenant_shard_id))?;
5073 :
5074 : use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
5075 192 : let remote_fs_dir = conf.workdir.join("localfs");
5076 192 : std::fs::create_dir_all(&remote_fs_dir).unwrap();
5077 192 : let config = RemoteStorageConfig {
5078 192 : storage: RemoteStorageKind::LocalFs {
5079 192 : local_path: remote_fs_dir.clone(),
5080 192 : },
5081 192 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
5082 192 : };
5083 192 : let remote_storage = GenericRemoteStorage::from_config(&config).await.unwrap();
5084 192 : let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()));
5085 192 :
5086 192 : Ok(Self {
5087 192 : conf,
5088 192 : tenant_conf,
5089 192 : tenant_shard_id,
5090 192 : generation,
5091 192 : shard,
5092 192 : remote_storage,
5093 192 : remote_fs_dir,
5094 192 : deletion_queue,
5095 192 : })
5096 192 : }
5097 :
5098 180 : pub async fn create(test_name: &'static str) -> anyhow::Result<Self> {
5099 180 : // Disable automatic GC and compaction to make the unit tests more deterministic.
5100 180 : // The tests perform them manually if needed.
5101 180 : let tenant_conf = TenantConf {
5102 180 : gc_period: Duration::ZERO,
5103 180 : compaction_period: Duration::ZERO,
5104 180 : ..TenantConf::default()
5105 180 : };
5106 180 : let tenant_id = TenantId::generate();
5107 180 : let shard = ShardIdentity::unsharded();
5108 180 : Self::create_custom(
5109 180 : test_name,
5110 180 : tenant_conf,
5111 180 : tenant_id,
5112 180 : shard,
5113 180 : Generation::new(0xdeadbeef),
5114 180 : )
5115 0 : .await
5116 180 : }
5117 :
5118 20 : pub fn span(&self) -> tracing::Span {
5119 20 : info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
5120 20 : }
5121 :
5122 192 : pub(crate) async fn load(&self) -> (Arc<Tenant>, RequestContext) {
5123 192 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
5124 192 : (
5125 192 : self.do_try_load(&ctx)
5126 1865 : .await
5127 192 : .expect("failed to load test tenant"),
5128 192 : ctx,
5129 192 : )
5130 192 : }
5131 :
5132 192 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5133 : pub(crate) async fn do_try_load(
5134 : &self,
5135 : ctx: &RequestContext,
5136 : ) -> anyhow::Result<Arc<Tenant>> {
5137 : let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
5138 :
5139 : let tenant = Arc::new(Tenant::new(
5140 : TenantState::Attaching,
5141 : self.conf,
5142 : AttachedTenantConf::try_from(LocationConf::attached_single(
5143 : TenantConfOpt::from(self.tenant_conf.clone()),
5144 : self.generation,
5145 : &ShardParameters::default(),
5146 : ))
5147 : .unwrap(),
5148 : // This is a legacy/test code path: sharding isn't supported here.
5149 : ShardIdentity::unsharded(),
5150 : Some(walredo_mgr),
5151 : self.tenant_shard_id,
5152 : self.remote_storage.clone(),
5153 : self.deletion_queue.new_client(),
5154 : // TODO: ideally we should run all unit tests with both configs
5155 : L0FlushGlobalState::new(L0FlushConfig::default()),
5156 : ));
5157 :
5158 : let preload = tenant
5159 : .preload(&self.remote_storage, CancellationToken::new())
5160 : .await?;
5161 : tenant.attach(Some(preload), ctx).await?;
5162 :
5163 : tenant.state.send_replace(TenantState::Active);
5164 : for timeline in tenant.timelines.lock().unwrap().values() {
5165 : timeline.set_state(TimelineState::Active);
5166 : }
5167 : Ok(tenant)
5168 : }
5169 :
5170 2 : pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
5171 2 : self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
5172 2 : }
5173 : }
5174 :
5175 : // Mock WAL redo manager that doesn't do much
5176 : pub(crate) struct TestRedoManager;
5177 :
5178 : impl TestRedoManager {
5179 : /// # Cancel-Safety
5180 : ///
5181 : /// This method is cancellation-safe.
5182 410 : pub async fn request_redo(
5183 410 : &self,
5184 410 : key: Key,
5185 410 : lsn: Lsn,
5186 410 : base_img: Option<(Lsn, Bytes)>,
5187 410 : records: Vec<(Lsn, NeonWalRecord)>,
5188 410 : _pg_version: u32,
5189 410 : ) -> Result<Bytes, walredo::Error> {
5190 570 : let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
5191 410 : if records_neon {
5192 : // For Neon wal records, we can decode without spawning postgres, so do so.
5193 410 : let mut page = match (base_img, records.first()) {
5194 344 : (Some((_lsn, img)), _) => {
5195 344 : let mut page = BytesMut::new();
5196 344 : page.extend_from_slice(&img);
5197 344 : page
5198 : }
5199 66 : (_, Some((_lsn, rec))) if rec.will_init() => BytesMut::new(),
5200 : _ => {
5201 0 : panic!("Neon WAL redo requires base image or will init record");
5202 : }
5203 : };
5204 :
5205 980 : for (record_lsn, record) in records {
5206 570 : apply_neon::apply_in_neon(&record, record_lsn, key, &mut page)?;
5207 : }
5208 410 : Ok(page.freeze())
5209 : } else {
5210 : // We never spawn a postgres walredo process in unit tests: just log what we might have done.
5211 0 : let s = format!(
5212 0 : "redo for {} to get to {}, with {} and {} records",
5213 0 : key,
5214 0 : lsn,
5215 0 : if base_img.is_some() {
5216 0 : "base image"
5217 : } else {
5218 0 : "no base image"
5219 : },
5220 0 : records.len()
5221 0 : );
5222 0 : println!("{s}");
5223 0 :
5224 0 : Ok(test_img(&s))
5225 : }
5226 410 : }
5227 : }
5228 : }
5229 :
5230 : #[cfg(test)]
5231 : mod tests {
5232 : use std::collections::{BTreeMap, BTreeSet};
5233 :
5234 : use super::*;
5235 : use crate::keyspace::KeySpaceAccum;
5236 : use crate::tenant::harness::*;
5237 : use crate::tenant::timeline::CompactFlags;
5238 : use crate::DEFAULT_PG_VERSION;
5239 : use bytes::{Bytes, BytesMut};
5240 : use hex_literal::hex;
5241 : use itertools::Itertools;
5242 : use pageserver_api::key::{Key, AUX_KEY_PREFIX, NON_INHERITED_RANGE};
5243 : use pageserver_api::keyspace::KeySpace;
5244 : use pageserver_api::models::{CompactionAlgorithm, CompactionAlgorithmSettings};
5245 : use pageserver_api::value::Value;
5246 : use pageserver_compaction::helpers::overlaps_with;
5247 : use rand::{thread_rng, Rng};
5248 : use storage_layer::PersistentLayerKey;
5249 : use tests::storage_layer::ValuesReconstructState;
5250 : use tests::timeline::{GetVectoredError, ShutdownMode};
5251 : use timeline::DeltaLayerTestDesc;
5252 : use utils::id::TenantId;
5253 :
5254 : #[cfg(feature = "testing")]
5255 : use pageserver_api::record::NeonWalRecord;
5256 : #[cfg(feature = "testing")]
5257 : use timeline::compaction::{KeyHistoryRetention, KeyLogAtLsn};
5258 : #[cfg(feature = "testing")]
5259 : use timeline::GcInfo;
5260 :
5261 : static TEST_KEY: Lazy<Key> =
5262 18 : Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));
5263 :
5264 : #[tokio::test]
5265 2 : async fn test_basic() -> anyhow::Result<()> {
5266 14 : let (tenant, ctx) = TenantHarness::create("test_basic").await?.load().await;
5267 2 : let tline = tenant
5268 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5269 6 : .await?;
5270 2 :
5271 2 : let mut writer = tline.writer().await;
5272 2 : writer
5273 2 : .put(
5274 2 : *TEST_KEY,
5275 2 : Lsn(0x10),
5276 2 : &Value::Image(test_img("foo at 0x10")),
5277 2 : &ctx,
5278 2 : )
5279 2 : .await?;
5280 2 : writer.finish_write(Lsn(0x10));
5281 2 : drop(writer);
5282 2 :
5283 2 : let mut writer = tline.writer().await;
5284 2 : writer
5285 2 : .put(
5286 2 : *TEST_KEY,
5287 2 : Lsn(0x20),
5288 2 : &Value::Image(test_img("foo at 0x20")),
5289 2 : &ctx,
5290 2 : )
5291 2 : .await?;
5292 2 : writer.finish_write(Lsn(0x20));
5293 2 : drop(writer);
5294 2 :
5295 2 : assert_eq!(
5296 2 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
5297 2 : test_img("foo at 0x10")
5298 2 : );
5299 2 : assert_eq!(
5300 2 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
5301 2 : test_img("foo at 0x10")
5302 2 : );
5303 2 : assert_eq!(
5304 2 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
5305 2 : test_img("foo at 0x20")
5306 2 : );
5307 2 :
5308 2 : Ok(())
5309 2 : }
5310 :
5311 : #[tokio::test]
5312 2 : async fn no_duplicate_timelines() -> anyhow::Result<()> {
5313 2 : let (tenant, ctx) = TenantHarness::create("no_duplicate_timelines")
5314 2 : .await?
5315 2 : .load()
5316 20 : .await;
5317 2 : let _ = tenant
5318 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5319 6 : .await?;
5320 2 :
5321 2 : match tenant
5322 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5323 2 : .await
5324 2 : {
5325 2 : Ok(_) => panic!("duplicate timeline creation should fail"),
5326 2 : Err(e) => assert_eq!(
5327 2 : e.to_string(),
5328 2 : "timeline already exists with different parameters".to_string()
5329 2 : ),
5330 2 : }
5331 2 :
5332 2 : Ok(())
5333 2 : }
5334 :
5335 : /// Convenience function to create a page image with given string as the only content
5336 10 : pub fn test_value(s: &str) -> Value {
5337 10 : let mut buf = BytesMut::new();
5338 10 : buf.extend_from_slice(s.as_bytes());
5339 10 : Value::Image(buf.freeze())
5340 10 : }
5341 :
5342 : ///
5343 : /// Test branch creation
5344 : ///
5345 : #[tokio::test]
5346 2 : async fn test_branch() -> anyhow::Result<()> {
5347 2 : use std::str::from_utf8;
5348 2 :
5349 20 : let (tenant, ctx) = TenantHarness::create("test_branch").await?.load().await;
5350 2 : let tline = tenant
5351 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5352 6 : .await?;
5353 2 : let mut writer = tline.writer().await;
5354 2 :
5355 2 : #[allow(non_snake_case)]
5356 2 : let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
5357 2 : #[allow(non_snake_case)]
5358 2 : let TEST_KEY_B: Key = Key::from_hex("110000000033333333444444445500000002").unwrap();
5359 2 :
5360 2 : // Insert a value on the timeline
5361 2 : writer
5362 2 : .put(TEST_KEY_A, Lsn(0x20), &test_value("foo at 0x20"), &ctx)
5363 2 : .await?;
5364 2 : writer
5365 2 : .put(TEST_KEY_B, Lsn(0x20), &test_value("foobar at 0x20"), &ctx)
5366 2 : .await?;
5367 2 : writer.finish_write(Lsn(0x20));
5368 2 :
5369 2 : writer
5370 2 : .put(TEST_KEY_A, Lsn(0x30), &test_value("foo at 0x30"), &ctx)
5371 2 : .await?;
5372 2 : writer.finish_write(Lsn(0x30));
5373 2 : writer
5374 2 : .put(TEST_KEY_A, Lsn(0x40), &test_value("foo at 0x40"), &ctx)
5375 2 : .await?;
5376 2 : writer.finish_write(Lsn(0x40));
5377 2 :
5378 2 : //assert_current_logical_size(&tline, Lsn(0x40));
5379 2 :
5380 2 : // Branch the history, modify relation differently on the new timeline
5381 2 : tenant
5382 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x30)), &ctx)
5383 2 : .await?;
5384 2 : let newtline = tenant
5385 2 : .get_timeline(NEW_TIMELINE_ID, true)
5386 2 : .expect("Should have a local timeline");
5387 2 : let mut new_writer = newtline.writer().await;
5388 2 : new_writer
5389 2 : .put(TEST_KEY_A, Lsn(0x40), &test_value("bar at 0x40"), &ctx)
5390 2 : .await?;
5391 2 : new_writer.finish_write(Lsn(0x40));
5392 2 :
5393 2 : // Check page contents on both branches
5394 2 : assert_eq!(
5395 2 : from_utf8(&tline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
5396 2 : "foo at 0x40"
5397 2 : );
5398 2 : assert_eq!(
5399 2 : from_utf8(&newtline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
5400 2 : "bar at 0x40"
5401 2 : );
5402 2 : assert_eq!(
5403 2 : from_utf8(&newtline.get(TEST_KEY_B, Lsn(0x40), &ctx).await?)?,
5404 2 : "foobar at 0x20"
5405 2 : );
5406 2 :
5407 2 : //assert_current_logical_size(&tline, Lsn(0x40));
5408 2 :
5409 2 : Ok(())
5410 2 : }
5411 :
5412 20 : async fn make_some_layers(
5413 20 : tline: &Timeline,
5414 20 : start_lsn: Lsn,
5415 20 : ctx: &RequestContext,
5416 20 : ) -> anyhow::Result<()> {
5417 20 : let mut lsn = start_lsn;
5418 : {
5419 20 : let mut writer = tline.writer().await;
5420 : // Create a relation on the timeline
5421 20 : writer
5422 20 : .put(
5423 20 : *TEST_KEY,
5424 20 : lsn,
5425 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5426 20 : ctx,
5427 20 : )
5428 10 : .await?;
5429 20 : writer.finish_write(lsn);
5430 20 : lsn += 0x10;
5431 20 : writer
5432 20 : .put(
5433 20 : *TEST_KEY,
5434 20 : lsn,
5435 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5436 20 : ctx,
5437 20 : )
5438 0 : .await?;
5439 20 : writer.finish_write(lsn);
5440 20 : lsn += 0x10;
5441 20 : }
5442 20 : tline.freeze_and_flush().await?;
5443 : {
5444 20 : let mut writer = tline.writer().await;
5445 20 : writer
5446 20 : .put(
5447 20 : *TEST_KEY,
5448 20 : lsn,
5449 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5450 20 : ctx,
5451 20 : )
5452 10 : .await?;
5453 20 : writer.finish_write(lsn);
5454 20 : lsn += 0x10;
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 0 : .await?;
5463 20 : writer.finish_write(lsn);
5464 20 : }
5465 21 : tline.freeze_and_flush().await.map_err(|e| e.into())
5466 20 : }
5467 :
5468 : #[tokio::test(start_paused = true)]
5469 2 : async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
5470 2 : let (tenant, ctx) =
5471 2 : TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
5472 2 : .await?
5473 2 : .load()
5474 20 : .await;
5475 2 : // Advance to the lsn lease deadline so that GC is not blocked by
5476 2 : // initial transition into AttachedSingle.
5477 2 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
5478 2 : tokio::time::resume();
5479 2 : let tline = tenant
5480 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5481 6 : .await?;
5482 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5483 2 :
5484 2 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
5485 2 : // FIXME: this doesn't actually remove any layer currently, given how the flushing
5486 2 : // and compaction works. But it does set the 'cutoff' point so that the cross check
5487 2 : // below should fail.
5488 2 : tenant
5489 2 : .gc_iteration(
5490 2 : Some(TIMELINE_ID),
5491 2 : 0x10,
5492 2 : Duration::ZERO,
5493 2 : &CancellationToken::new(),
5494 2 : &ctx,
5495 2 : )
5496 2 : .await?;
5497 2 :
5498 2 : // try to branch at lsn 25, should fail because we already garbage collected the data
5499 2 : match tenant
5500 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
5501 2 : .await
5502 2 : {
5503 2 : Ok(_) => panic!("branching should have failed"),
5504 2 : Err(err) => {
5505 2 : let CreateTimelineError::AncestorLsn(err) = err else {
5506 2 : panic!("wrong error type")
5507 2 : };
5508 2 : assert!(err.to_string().contains("invalid branch start lsn"));
5509 2 : assert!(err
5510 2 : .source()
5511 2 : .unwrap()
5512 2 : .to_string()
5513 2 : .contains("we might've already garbage collected needed data"))
5514 2 : }
5515 2 : }
5516 2 :
5517 2 : Ok(())
5518 2 : }
5519 :
5520 : #[tokio::test]
5521 2 : async fn test_prohibit_branch_creation_on_pre_initdb_lsn() -> anyhow::Result<()> {
5522 2 : let (tenant, ctx) =
5523 2 : TenantHarness::create("test_prohibit_branch_creation_on_pre_initdb_lsn")
5524 2 : .await?
5525 2 : .load()
5526 20 : .await;
5527 2 :
5528 2 : let tline = tenant
5529 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x50), DEFAULT_PG_VERSION, &ctx)
5530 5 : .await?;
5531 2 : // try to branch at lsn 0x25, should fail because initdb lsn is 0x50
5532 2 : match tenant
5533 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
5534 2 : .await
5535 2 : {
5536 2 : Ok(_) => panic!("branching should have failed"),
5537 2 : Err(err) => {
5538 2 : let CreateTimelineError::AncestorLsn(err) = err else {
5539 2 : panic!("wrong error type");
5540 2 : };
5541 2 : assert!(&err.to_string().contains("invalid branch start lsn"));
5542 2 : assert!(&err
5543 2 : .source()
5544 2 : .unwrap()
5545 2 : .to_string()
5546 2 : .contains("is earlier than latest GC cutoff"));
5547 2 : }
5548 2 : }
5549 2 :
5550 2 : Ok(())
5551 2 : }
5552 :
5553 : /*
5554 : // FIXME: This currently fails to error out. Calling GC doesn't currently
5555 : // remove the old value, we'd need to work a little harder
5556 : #[tokio::test]
5557 : async fn test_prohibit_get_for_garbage_collected_data() -> anyhow::Result<()> {
5558 : let repo =
5559 : RepoHarness::create("test_prohibit_get_for_garbage_collected_data")?
5560 : .load();
5561 :
5562 : let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION)?;
5563 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5564 :
5565 : repo.gc_iteration(Some(TIMELINE_ID), 0x10, Duration::ZERO)?;
5566 : let latest_gc_cutoff_lsn = tline.get_latest_gc_cutoff_lsn();
5567 : assert!(*latest_gc_cutoff_lsn > Lsn(0x25));
5568 : match tline.get(*TEST_KEY, Lsn(0x25)) {
5569 : Ok(_) => panic!("request for page should have failed"),
5570 : Err(err) => assert!(err.to_string().contains("not found at")),
5571 : }
5572 : Ok(())
5573 : }
5574 : */
5575 :
5576 : #[tokio::test]
5577 2 : async fn test_get_branchpoints_from_an_inactive_timeline() -> anyhow::Result<()> {
5578 2 : let (tenant, ctx) =
5579 2 : TenantHarness::create("test_get_branchpoints_from_an_inactive_timeline")
5580 2 : .await?
5581 2 : .load()
5582 20 : .await;
5583 2 : let tline = tenant
5584 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5585 6 : .await?;
5586 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5587 2 :
5588 2 : tenant
5589 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
5590 2 : .await?;
5591 2 : let newtline = tenant
5592 2 : .get_timeline(NEW_TIMELINE_ID, true)
5593 2 : .expect("Should have a local timeline");
5594 2 :
5595 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
5596 2 :
5597 2 : tline.set_broken("test".to_owned());
5598 2 :
5599 2 : tenant
5600 2 : .gc_iteration(
5601 2 : Some(TIMELINE_ID),
5602 2 : 0x10,
5603 2 : Duration::ZERO,
5604 2 : &CancellationToken::new(),
5605 2 : &ctx,
5606 2 : )
5607 2 : .await?;
5608 2 :
5609 2 : // The branchpoints should contain all timelines, even ones marked
5610 2 : // as Broken.
5611 2 : {
5612 2 : let branchpoints = &tline.gc_info.read().unwrap().retain_lsns;
5613 2 : assert_eq!(branchpoints.len(), 1);
5614 2 : assert_eq!(
5615 2 : branchpoints[0],
5616 2 : (Lsn(0x40), NEW_TIMELINE_ID, MaybeOffloaded::No)
5617 2 : );
5618 2 : }
5619 2 :
5620 2 : // You can read the key from the child branch even though the parent is
5621 2 : // Broken, as long as you don't need to access data from the parent.
5622 2 : assert_eq!(
5623 4 : newtline.get(*TEST_KEY, Lsn(0x70), &ctx).await?,
5624 2 : test_img(&format!("foo at {}", Lsn(0x70)))
5625 2 : );
5626 2 :
5627 2 : // This needs to traverse to the parent, and fails.
5628 2 : let err = newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await.unwrap_err();
5629 2 : assert!(
5630 2 : err.to_string().starts_with(&format!(
5631 2 : "bad state on timeline {}: Broken",
5632 2 : tline.timeline_id
5633 2 : )),
5634 2 : "{err}"
5635 2 : );
5636 2 :
5637 2 : Ok(())
5638 2 : }
5639 :
5640 : #[tokio::test]
5641 2 : async fn test_retain_data_in_parent_which_is_needed_for_child() -> anyhow::Result<()> {
5642 2 : let (tenant, ctx) =
5643 2 : TenantHarness::create("test_retain_data_in_parent_which_is_needed_for_child")
5644 2 : .await?
5645 2 : .load()
5646 20 : .await;
5647 2 : let tline = tenant
5648 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5649 6 : .await?;
5650 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5651 2 :
5652 2 : tenant
5653 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
5654 2 : .await?;
5655 2 : let newtline = tenant
5656 2 : .get_timeline(NEW_TIMELINE_ID, true)
5657 2 : .expect("Should have a local timeline");
5658 2 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
5659 2 : tenant
5660 2 : .gc_iteration(
5661 2 : Some(TIMELINE_ID),
5662 2 : 0x10,
5663 2 : Duration::ZERO,
5664 2 : &CancellationToken::new(),
5665 2 : &ctx,
5666 2 : )
5667 2 : .await?;
5668 4 : assert!(newtline.get(*TEST_KEY, Lsn(0x25), &ctx).await.is_ok());
5669 2 :
5670 2 : Ok(())
5671 2 : }
5672 : #[tokio::test]
5673 2 : async fn test_parent_keeps_data_forever_after_branching() -> anyhow::Result<()> {
5674 2 : let (tenant, ctx) = TenantHarness::create("test_parent_keeps_data_forever_after_branching")
5675 2 : .await?
5676 2 : .load()
5677 20 : .await;
5678 2 : let tline = tenant
5679 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5680 6 : .await?;
5681 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5682 2 :
5683 2 : tenant
5684 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
5685 2 : .await?;
5686 2 : let newtline = tenant
5687 2 : .get_timeline(NEW_TIMELINE_ID, true)
5688 2 : .expect("Should have a local timeline");
5689 2 :
5690 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
5691 2 :
5692 2 : // run gc on parent
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 2 :
5703 2 : // Check that the data is still accessible on the branch.
5704 2 : assert_eq!(
5705 7 : newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
5706 2 : test_img(&format!("foo at {}", Lsn(0x40)))
5707 2 : );
5708 2 :
5709 2 : Ok(())
5710 2 : }
5711 :
5712 : #[tokio::test]
5713 2 : async fn timeline_load() -> anyhow::Result<()> {
5714 2 : const TEST_NAME: &str = "timeline_load";
5715 2 : let harness = TenantHarness::create(TEST_NAME).await?;
5716 2 : {
5717 20 : let (tenant, ctx) = harness.load().await;
5718 2 : let tline = tenant
5719 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x7000), DEFAULT_PG_VERSION, &ctx)
5720 6 : .await?;
5721 6 : make_some_layers(tline.as_ref(), Lsn(0x8000), &ctx).await?;
5722 2 : // so that all uploads finish & we can call harness.load() below again
5723 2 : tenant
5724 2 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
5725 2 : .instrument(harness.span())
5726 2 : .await
5727 2 : .ok()
5728 2 : .unwrap();
5729 2 : }
5730 2 :
5731 17 : let (tenant, _ctx) = harness.load().await;
5732 2 : tenant
5733 2 : .get_timeline(TIMELINE_ID, true)
5734 2 : .expect("cannot load timeline");
5735 2 :
5736 2 : Ok(())
5737 2 : }
5738 :
5739 : #[tokio::test]
5740 2 : async fn timeline_load_with_ancestor() -> anyhow::Result<()> {
5741 2 : const TEST_NAME: &str = "timeline_load_with_ancestor";
5742 2 : let harness = TenantHarness::create(TEST_NAME).await?;
5743 2 : // create two timelines
5744 2 : {
5745 11 : let (tenant, ctx) = harness.load().await;
5746 2 : let tline = tenant
5747 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5748 5 : .await?;
5749 2 :
5750 7 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5751 2 :
5752 2 : let child_tline = tenant
5753 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
5754 2 : .await?;
5755 2 : child_tline.set_state(TimelineState::Active);
5756 2 :
5757 2 : let newtline = tenant
5758 2 : .get_timeline(NEW_TIMELINE_ID, true)
5759 2 : .expect("Should have a local timeline");
5760 2 :
5761 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
5762 2 :
5763 2 : // so that all uploads finish & we can call harness.load() below again
5764 2 : tenant
5765 2 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
5766 2 : .instrument(harness.span())
5767 2 : .await
5768 2 : .ok()
5769 2 : .unwrap();
5770 2 : }
5771 2 :
5772 2 : // check that both of them are initially unloaded
5773 6 : let (tenant, _ctx) = harness.load().await;
5774 2 :
5775 2 : // check that both, child and ancestor are loaded
5776 2 : let _child_tline = tenant
5777 2 : .get_timeline(NEW_TIMELINE_ID, true)
5778 2 : .expect("cannot get child timeline loaded");
5779 2 :
5780 2 : let _ancestor_tline = tenant
5781 2 : .get_timeline(TIMELINE_ID, true)
5782 2 : .expect("cannot get ancestor timeline loaded");
5783 2 :
5784 2 : Ok(())
5785 2 : }
5786 :
5787 : #[tokio::test]
5788 2 : async fn delta_layer_dumping() -> anyhow::Result<()> {
5789 2 : use storage_layer::AsLayerDesc;
5790 2 : let (tenant, ctx) = TenantHarness::create("test_layer_dumping")
5791 2 : .await?
5792 2 : .load()
5793 20 : .await;
5794 2 : let tline = tenant
5795 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5796 6 : .await?;
5797 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5798 2 :
5799 2 : let layer_map = tline.layers.read().await;
5800 2 : let level0_deltas = layer_map
5801 2 : .layer_map()?
5802 2 : .level0_deltas()
5803 2 : .iter()
5804 4 : .map(|desc| layer_map.get_from_desc(desc))
5805 2 : .collect::<Vec<_>>();
5806 2 :
5807 2 : assert!(!level0_deltas.is_empty());
5808 2 :
5809 6 : for delta in level0_deltas {
5810 2 : // Ensure we are dumping a delta layer here
5811 4 : assert!(delta.layer_desc().is_delta);
5812 8 : delta.dump(true, &ctx).await.unwrap();
5813 2 : }
5814 2 :
5815 2 : Ok(())
5816 2 : }
5817 :
5818 : #[tokio::test]
5819 2 : async fn test_images() -> anyhow::Result<()> {
5820 20 : let (tenant, ctx) = TenantHarness::create("test_images").await?.load().await;
5821 2 : let tline = tenant
5822 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5823 6 : .await?;
5824 2 :
5825 2 : let mut writer = tline.writer().await;
5826 2 : writer
5827 2 : .put(
5828 2 : *TEST_KEY,
5829 2 : Lsn(0x10),
5830 2 : &Value::Image(test_img("foo at 0x10")),
5831 2 : &ctx,
5832 2 : )
5833 2 : .await?;
5834 2 : writer.finish_write(Lsn(0x10));
5835 2 : drop(writer);
5836 2 :
5837 3 : tline.freeze_and_flush().await?;
5838 2 : tline
5839 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
5840 2 : .await?;
5841 2 :
5842 2 : let mut writer = tline.writer().await;
5843 2 : writer
5844 2 : .put(
5845 2 : *TEST_KEY,
5846 2 : Lsn(0x20),
5847 2 : &Value::Image(test_img("foo at 0x20")),
5848 2 : &ctx,
5849 2 : )
5850 2 : .await?;
5851 2 : writer.finish_write(Lsn(0x20));
5852 2 : drop(writer);
5853 2 :
5854 2 : tline.freeze_and_flush().await?;
5855 2 : tline
5856 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
5857 2 : .await?;
5858 2 :
5859 2 : let mut writer = tline.writer().await;
5860 2 : writer
5861 2 : .put(
5862 2 : *TEST_KEY,
5863 2 : Lsn(0x30),
5864 2 : &Value::Image(test_img("foo at 0x30")),
5865 2 : &ctx,
5866 2 : )
5867 2 : .await?;
5868 2 : writer.finish_write(Lsn(0x30));
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(0x40),
5881 2 : &Value::Image(test_img("foo at 0x40")),
5882 2 : &ctx,
5883 2 : )
5884 2 : .await?;
5885 2 : writer.finish_write(Lsn(0x40));
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 : assert_eq!(
5894 4 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
5895 2 : test_img("foo at 0x10")
5896 2 : );
5897 2 : assert_eq!(
5898 4 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
5899 2 : test_img("foo at 0x10")
5900 2 : );
5901 2 : assert_eq!(
5902 2 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
5903 2 : test_img("foo at 0x20")
5904 2 : );
5905 2 : assert_eq!(
5906 4 : tline.get(*TEST_KEY, Lsn(0x30), &ctx).await?,
5907 2 : test_img("foo at 0x30")
5908 2 : );
5909 2 : assert_eq!(
5910 4 : tline.get(*TEST_KEY, Lsn(0x40), &ctx).await?,
5911 2 : test_img("foo at 0x40")
5912 2 : );
5913 2 :
5914 2 : Ok(())
5915 2 : }
5916 :
5917 4 : async fn bulk_insert_compact_gc(
5918 4 : tenant: &Tenant,
5919 4 : timeline: &Arc<Timeline>,
5920 4 : ctx: &RequestContext,
5921 4 : lsn: Lsn,
5922 4 : repeat: usize,
5923 4 : key_count: usize,
5924 4 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
5925 4 : let compact = true;
5926 40718 : bulk_insert_maybe_compact_gc(tenant, timeline, ctx, lsn, repeat, key_count, compact).await
5927 4 : }
5928 :
5929 8 : async fn bulk_insert_maybe_compact_gc(
5930 8 : tenant: &Tenant,
5931 8 : timeline: &Arc<Timeline>,
5932 8 : ctx: &RequestContext,
5933 8 : mut lsn: Lsn,
5934 8 : repeat: usize,
5935 8 : key_count: usize,
5936 8 : compact: bool,
5937 8 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
5938 8 : let mut inserted: HashMap<Key, BTreeSet<Lsn>> = Default::default();
5939 8 :
5940 8 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5941 8 : let mut blknum = 0;
5942 8 :
5943 8 : // Enforce that key range is monotonously increasing
5944 8 : let mut keyspace = KeySpaceAccum::new();
5945 8 :
5946 8 : let cancel = CancellationToken::new();
5947 8 :
5948 8 : for _ in 0..repeat {
5949 400 : for _ in 0..key_count {
5950 4000000 : test_key.field6 = blknum;
5951 4000000 : let mut writer = timeline.writer().await;
5952 4000000 : writer
5953 4000000 : .put(
5954 4000000 : test_key,
5955 4000000 : lsn,
5956 4000000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5957 4000000 : ctx,
5958 4000000 : )
5959 3426 : .await?;
5960 4000000 : inserted.entry(test_key).or_default().insert(lsn);
5961 4000000 : writer.finish_write(lsn);
5962 4000000 : drop(writer);
5963 4000000 :
5964 4000000 : keyspace.add_key(test_key);
5965 4000000 :
5966 4000000 : lsn = Lsn(lsn.0 + 0x10);
5967 4000000 : blknum += 1;
5968 : }
5969 :
5970 400 : timeline.freeze_and_flush().await?;
5971 400 : if compact {
5972 : // this requires timeline to be &Arc<Timeline>
5973 8618 : timeline.compact(&cancel, EnumSet::empty(), ctx).await?;
5974 200 : }
5975 :
5976 : // this doesn't really need to use the timeline_id target, but it is closer to what it
5977 : // originally was.
5978 400 : let res = tenant
5979 400 : .gc_iteration(Some(timeline.timeline_id), 0, Duration::ZERO, &cancel, ctx)
5980 0 : .await?;
5981 :
5982 400 : assert_eq!(res.layers_removed, 0, "this never removes anything");
5983 : }
5984 :
5985 8 : Ok(inserted)
5986 8 : }
5987 :
5988 : //
5989 : // Insert 1000 key-value pairs with increasing keys, flush, compact, GC.
5990 : // Repeat 50 times.
5991 : //
5992 : #[tokio::test]
5993 2 : async fn test_bulk_insert() -> anyhow::Result<()> {
5994 2 : let harness = TenantHarness::create("test_bulk_insert").await?;
5995 20 : let (tenant, ctx) = harness.load().await;
5996 2 : let tline = tenant
5997 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5998 6 : .await?;
5999 2 :
6000 2 : let lsn = Lsn(0x10);
6001 20359 : bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
6002 2 :
6003 2 : Ok(())
6004 2 : }
6005 :
6006 : // Test the vectored get real implementation against a simple sequential implementation.
6007 : //
6008 : // The test generates a keyspace by repeatedly flushing the in-memory layer and compacting.
6009 : // Projected to 2D the key space looks like below. Lsn grows upwards on the Y axis and keys
6010 : // grow to the right on the X axis.
6011 : // [Delta]
6012 : // [Delta]
6013 : // [Delta]
6014 : // [Delta]
6015 : // ------------ Image ---------------
6016 : //
6017 : // After layer generation we pick the ranges to query as follows:
6018 : // 1. The beginning of each delta layer
6019 : // 2. At the seam between two adjacent delta layers
6020 : //
6021 : // There's one major downside to this test: delta layers only contains images,
6022 : // so the search can stop at the first delta layer and doesn't traverse any deeper.
6023 : #[tokio::test]
6024 2 : async fn test_get_vectored() -> anyhow::Result<()> {
6025 2 : let harness = TenantHarness::create("test_get_vectored").await?;
6026 14 : let (tenant, ctx) = harness.load().await;
6027 2 : let tline = tenant
6028 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6029 5 : .await?;
6030 2 :
6031 2 : let lsn = Lsn(0x10);
6032 20359 : let inserted = bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
6033 2 :
6034 2 : let guard = tline.layers.read().await;
6035 2 : let lm = guard.layer_map()?;
6036 2 :
6037 2 : lm.dump(true, &ctx).await?;
6038 2 :
6039 2 : let mut reads = Vec::new();
6040 2 : let mut prev = None;
6041 12 : lm.iter_historic_layers().for_each(|desc| {
6042 12 : if !desc.is_delta() {
6043 2 : prev = Some(desc.clone());
6044 2 : return;
6045 10 : }
6046 10 :
6047 10 : let start = desc.key_range.start;
6048 10 : let end = desc
6049 10 : .key_range
6050 10 : .start
6051 10 : .add(Timeline::MAX_GET_VECTORED_KEYS.try_into().unwrap());
6052 10 : reads.push(KeySpace {
6053 10 : ranges: vec![start..end],
6054 10 : });
6055 2 :
6056 10 : if let Some(prev) = &prev {
6057 10 : if !prev.is_delta() {
6058 10 : return;
6059 2 : }
6060 0 :
6061 0 : let first_range = Key {
6062 0 : field6: prev.key_range.end.field6 - 4,
6063 0 : ..prev.key_range.end
6064 0 : }..prev.key_range.end;
6065 0 :
6066 0 : let second_range = desc.key_range.start..Key {
6067 0 : field6: desc.key_range.start.field6 + 4,
6068 0 : ..desc.key_range.start
6069 0 : };
6070 0 :
6071 0 : reads.push(KeySpace {
6072 0 : ranges: vec![first_range, second_range],
6073 0 : });
6074 2 : };
6075 2 :
6076 2 : prev = Some(desc.clone());
6077 12 : });
6078 2 :
6079 2 : drop(guard);
6080 2 :
6081 2 : // Pick a big LSN such that we query over all the changes.
6082 2 : let reads_lsn = Lsn(u64::MAX - 1);
6083 2 :
6084 12 : for read in reads {
6085 10 : info!("Doing vectored read on {:?}", read);
6086 2 :
6087 10 : let vectored_res = tline
6088 10 : .get_vectored_impl(
6089 10 : read.clone(),
6090 10 : reads_lsn,
6091 10 : &mut ValuesReconstructState::new(),
6092 10 : &ctx,
6093 10 : )
6094 25 : .await;
6095 2 :
6096 10 : let mut expected_lsns: HashMap<Key, Lsn> = Default::default();
6097 10 : let mut expect_missing = false;
6098 10 : let mut key = read.start().unwrap();
6099 330 : while key != read.end().unwrap() {
6100 320 : if let Some(lsns) = inserted.get(&key) {
6101 320 : let expected_lsn = lsns.iter().rfind(|lsn| **lsn <= reads_lsn);
6102 320 : match expected_lsn {
6103 320 : Some(lsn) => {
6104 320 : expected_lsns.insert(key, *lsn);
6105 320 : }
6106 2 : None => {
6107 2 : expect_missing = true;
6108 0 : break;
6109 2 : }
6110 2 : }
6111 2 : } else {
6112 2 : expect_missing = true;
6113 0 : break;
6114 2 : }
6115 2 :
6116 320 : key = key.next();
6117 2 : }
6118 2 :
6119 10 : if expect_missing {
6120 2 : assert!(matches!(vectored_res, Err(GetVectoredError::MissingKey(_))));
6121 2 : } else {
6122 320 : for (key, image) in vectored_res? {
6123 320 : let expected_lsn = expected_lsns.get(&key).expect("determined above");
6124 320 : let expected_image = test_img(&format!("{} at {}", key.field6, expected_lsn));
6125 320 : assert_eq!(image?, expected_image);
6126 2 : }
6127 2 : }
6128 2 : }
6129 2 :
6130 2 : Ok(())
6131 2 : }
6132 :
6133 : #[tokio::test]
6134 2 : async fn test_get_vectored_aux_files() -> anyhow::Result<()> {
6135 2 : let harness = TenantHarness::create("test_get_vectored_aux_files").await?;
6136 2 :
6137 20 : let (tenant, ctx) = harness.load().await;
6138 2 : let tline = tenant
6139 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
6140 2 : .await?;
6141 2 : let tline = tline.raw_timeline().unwrap();
6142 2 :
6143 2 : let mut modification = tline.begin_modification(Lsn(0x1000));
6144 2 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
6145 2 : modification.set_lsn(Lsn(0x1008))?;
6146 2 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
6147 2 : modification.commit(&ctx).await?;
6148 2 :
6149 2 : let child_timeline_id = TimelineId::generate();
6150 2 : tenant
6151 2 : .branch_timeline_test(
6152 2 : tline,
6153 2 : child_timeline_id,
6154 2 : Some(tline.get_last_record_lsn()),
6155 2 : &ctx,
6156 2 : )
6157 2 : .await?;
6158 2 :
6159 2 : let child_timeline = tenant
6160 2 : .get_timeline(child_timeline_id, true)
6161 2 : .expect("Should have the branched timeline");
6162 2 :
6163 2 : let aux_keyspace = KeySpace {
6164 2 : ranges: vec![NON_INHERITED_RANGE],
6165 2 : };
6166 2 : let read_lsn = child_timeline.get_last_record_lsn();
6167 2 :
6168 2 : let vectored_res = child_timeline
6169 2 : .get_vectored_impl(
6170 2 : aux_keyspace.clone(),
6171 2 : read_lsn,
6172 2 : &mut ValuesReconstructState::new(),
6173 2 : &ctx,
6174 2 : )
6175 2 : .await;
6176 2 :
6177 2 : let images = vectored_res?;
6178 2 : assert!(images.is_empty());
6179 2 : Ok(())
6180 2 : }
6181 :
6182 : // Test that vectored get handles layer gaps correctly
6183 : // by advancing into the next ancestor timeline if required.
6184 : //
6185 : // The test generates timelines that look like the diagram below.
6186 : // We leave a gap in one of the L1 layers at `gap_at_key` (`/` in the diagram).
6187 : // The reconstruct data for that key lies in the ancestor timeline (`X` in the diagram).
6188 : //
6189 : // ```
6190 : //-------------------------------+
6191 : // ... |
6192 : // [ L1 ] |
6193 : // [ / L1 ] | Child Timeline
6194 : // ... |
6195 : // ------------------------------+
6196 : // [ X L1 ] | Parent Timeline
6197 : // ------------------------------+
6198 : // ```
6199 : #[tokio::test]
6200 2 : async fn test_get_vectored_key_gap() -> anyhow::Result<()> {
6201 2 : let tenant_conf = TenantConf {
6202 2 : // Make compaction deterministic
6203 2 : gc_period: Duration::ZERO,
6204 2 : compaction_period: Duration::ZERO,
6205 2 : // Encourage creation of L1 layers
6206 2 : checkpoint_distance: 16 * 1024,
6207 2 : compaction_target_size: 8 * 1024,
6208 2 : ..TenantConf::default()
6209 2 : };
6210 2 :
6211 2 : let harness = TenantHarness::create_custom(
6212 2 : "test_get_vectored_key_gap",
6213 2 : tenant_conf,
6214 2 : TenantId::generate(),
6215 2 : ShardIdentity::unsharded(),
6216 2 : Generation::new(0xdeadbeef),
6217 2 : )
6218 2 : .await?;
6219 20 : let (tenant, ctx) = harness.load().await;
6220 2 :
6221 2 : let mut current_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6222 2 : let gap_at_key = current_key.add(100);
6223 2 : let mut current_lsn = Lsn(0x10);
6224 2 :
6225 2 : const KEY_COUNT: usize = 10_000;
6226 2 :
6227 2 : let timeline_id = TimelineId::generate();
6228 2 : let current_timeline = tenant
6229 2 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
6230 5 : .await?;
6231 2 :
6232 2 : current_lsn += 0x100;
6233 2 :
6234 2 : let mut writer = current_timeline.writer().await;
6235 2 : writer
6236 2 : .put(
6237 2 : gap_at_key,
6238 2 : current_lsn,
6239 2 : &Value::Image(test_img(&format!("{} at {}", gap_at_key, current_lsn))),
6240 2 : &ctx,
6241 2 : )
6242 2 : .await?;
6243 2 : writer.finish_write(current_lsn);
6244 2 : drop(writer);
6245 2 :
6246 2 : let mut latest_lsns = HashMap::new();
6247 2 : latest_lsns.insert(gap_at_key, current_lsn);
6248 2 :
6249 2 : current_timeline.freeze_and_flush().await?;
6250 2 :
6251 2 : let child_timeline_id = TimelineId::generate();
6252 2 :
6253 2 : tenant
6254 2 : .branch_timeline_test(
6255 2 : ¤t_timeline,
6256 2 : child_timeline_id,
6257 2 : Some(current_lsn),
6258 2 : &ctx,
6259 2 : )
6260 2 : .await?;
6261 2 : let child_timeline = tenant
6262 2 : .get_timeline(child_timeline_id, true)
6263 2 : .expect("Should have the branched timeline");
6264 2 :
6265 20002 : for i in 0..KEY_COUNT {
6266 20000 : if current_key == gap_at_key {
6267 2 : current_key = current_key.next();
6268 2 : continue;
6269 19998 : }
6270 19998 :
6271 19998 : current_lsn += 0x10;
6272 2 :
6273 19998 : let mut writer = child_timeline.writer().await;
6274 19998 : writer
6275 19998 : .put(
6276 19998 : current_key,
6277 19998 : current_lsn,
6278 19998 : &Value::Image(test_img(&format!("{} at {}", current_key, current_lsn))),
6279 19998 : &ctx,
6280 19998 : )
6281 67 : .await?;
6282 19998 : writer.finish_write(current_lsn);
6283 19998 : drop(writer);
6284 19998 :
6285 19998 : latest_lsns.insert(current_key, current_lsn);
6286 19998 : current_key = current_key.next();
6287 19998 :
6288 19998 : // Flush every now and then to encourage layer file creation.
6289 19998 : if i % 500 == 0 {
6290 42 : child_timeline.freeze_and_flush().await?;
6291 19958 : }
6292 2 : }
6293 2 :
6294 2 : child_timeline.freeze_and_flush().await?;
6295 2 : let mut flags = EnumSet::new();
6296 2 : flags.insert(CompactFlags::ForceRepartition);
6297 2 : child_timeline
6298 2 : .compact(&CancellationToken::new(), flags, &ctx)
6299 1757 : .await?;
6300 2 :
6301 2 : let key_near_end = {
6302 2 : let mut tmp = current_key;
6303 2 : tmp.field6 -= 10;
6304 2 : tmp
6305 2 : };
6306 2 :
6307 2 : let key_near_gap = {
6308 2 : let mut tmp = gap_at_key;
6309 2 : tmp.field6 -= 10;
6310 2 : tmp
6311 2 : };
6312 2 :
6313 2 : let read = KeySpace {
6314 2 : ranges: vec![key_near_gap..gap_at_key.next(), key_near_end..current_key],
6315 2 : };
6316 2 : let results = child_timeline
6317 2 : .get_vectored_impl(
6318 2 : read.clone(),
6319 2 : current_lsn,
6320 2 : &mut ValuesReconstructState::new(),
6321 2 : &ctx,
6322 2 : )
6323 16 : .await?;
6324 2 :
6325 44 : for (key, img_res) in results {
6326 42 : let expected = test_img(&format!("{} at {}", key, latest_lsns[&key]));
6327 42 : assert_eq!(img_res?, expected);
6328 2 : }
6329 2 :
6330 2 : Ok(())
6331 2 : }
6332 :
6333 : // Test that vectored get descends into ancestor timelines correctly and
6334 : // does not return an image that's newer than requested.
6335 : //
6336 : // The diagram below ilustrates an interesting case. We have a parent timeline
6337 : // (top of the Lsn range) and a child timeline. The request key cannot be reconstructed
6338 : // from the child timeline, so the parent timeline must be visited. When advacing into
6339 : // the child timeline, the read path needs to remember what the requested Lsn was in
6340 : // order to avoid returning an image that's too new. The test below constructs such
6341 : // a timeline setup and does a few queries around the Lsn of each page image.
6342 : // ```
6343 : // LSN
6344 : // ^
6345 : // |
6346 : // |
6347 : // 500 | --------------------------------------> branch point
6348 : // 400 | X
6349 : // 300 | X
6350 : // 200 | --------------------------------------> requested lsn
6351 : // 100 | X
6352 : // |---------------------------------------> Key
6353 : // |
6354 : // ------> requested key
6355 : //
6356 : // Legend:
6357 : // * X - page images
6358 : // ```
6359 : #[tokio::test]
6360 2 : async fn test_get_vectored_ancestor_descent() -> anyhow::Result<()> {
6361 2 : let harness = TenantHarness::create("test_get_vectored_on_lsn_axis").await?;
6362 20 : let (tenant, ctx) = harness.load().await;
6363 2 :
6364 2 : let start_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6365 2 : let end_key = start_key.add(1000);
6366 2 : let child_gap_at_key = start_key.add(500);
6367 2 : let mut parent_gap_lsns: BTreeMap<Lsn, String> = BTreeMap::new();
6368 2 :
6369 2 : let mut current_lsn = Lsn(0x10);
6370 2 :
6371 2 : let timeline_id = TimelineId::generate();
6372 2 : let parent_timeline = tenant
6373 2 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
6374 6 : .await?;
6375 2 :
6376 2 : current_lsn += 0x100;
6377 2 :
6378 8 : for _ in 0..3 {
6379 6 : let mut key = start_key;
6380 6006 : while key < end_key {
6381 6000 : current_lsn += 0x10;
6382 6000 :
6383 6000 : let image_value = format!("{} at {}", child_gap_at_key, current_lsn);
6384 2 :
6385 6000 : let mut writer = parent_timeline.writer().await;
6386 6000 : writer
6387 6000 : .put(
6388 6000 : key,
6389 6000 : current_lsn,
6390 6000 : &Value::Image(test_img(&image_value)),
6391 6000 : &ctx,
6392 6000 : )
6393 6 : .await?;
6394 6000 : writer.finish_write(current_lsn);
6395 6000 :
6396 6000 : if key == child_gap_at_key {
6397 6 : parent_gap_lsns.insert(current_lsn, image_value);
6398 5994 : }
6399 2 :
6400 6000 : key = key.next();
6401 2 : }
6402 2 :
6403 6 : parent_timeline.freeze_and_flush().await?;
6404 2 : }
6405 2 :
6406 2 : let child_timeline_id = TimelineId::generate();
6407 2 :
6408 2 : let child_timeline = tenant
6409 2 : .branch_timeline_test(&parent_timeline, child_timeline_id, Some(current_lsn), &ctx)
6410 2 : .await?;
6411 2 :
6412 2 : let mut key = start_key;
6413 2002 : while key < end_key {
6414 2000 : if key == child_gap_at_key {
6415 2 : key = key.next();
6416 2 : continue;
6417 1998 : }
6418 1998 :
6419 1998 : current_lsn += 0x10;
6420 2 :
6421 1998 : let mut writer = child_timeline.writer().await;
6422 1998 : writer
6423 1998 : .put(
6424 1998 : key,
6425 1998 : current_lsn,
6426 1998 : &Value::Image(test_img(&format!("{} at {}", key, current_lsn))),
6427 1998 : &ctx,
6428 1998 : )
6429 17 : .await?;
6430 1998 : writer.finish_write(current_lsn);
6431 1998 :
6432 1998 : key = key.next();
6433 2 : }
6434 2 :
6435 2 : child_timeline.freeze_and_flush().await?;
6436 2 :
6437 2 : let lsn_offsets: [i64; 5] = [-10, -1, 0, 1, 10];
6438 2 : let mut query_lsns = Vec::new();
6439 6 : for image_lsn in parent_gap_lsns.keys().rev() {
6440 36 : for offset in lsn_offsets {
6441 30 : query_lsns.push(Lsn(image_lsn
6442 30 : .0
6443 30 : .checked_add_signed(offset)
6444 30 : .expect("Shouldn't overflow")));
6445 30 : }
6446 2 : }
6447 2 :
6448 32 : for query_lsn in query_lsns {
6449 30 : let results = child_timeline
6450 30 : .get_vectored_impl(
6451 30 : KeySpace {
6452 30 : ranges: vec![child_gap_at_key..child_gap_at_key.next()],
6453 30 : },
6454 30 : query_lsn,
6455 30 : &mut ValuesReconstructState::new(),
6456 30 : &ctx,
6457 30 : )
6458 29 : .await;
6459 2 :
6460 30 : let expected_item = parent_gap_lsns
6461 30 : .iter()
6462 30 : .rev()
6463 68 : .find(|(lsn, _)| **lsn <= query_lsn);
6464 30 :
6465 30 : info!(
6466 2 : "Doing vectored read at LSN {}. Expecting image to be: {:?}",
6467 2 : query_lsn, expected_item
6468 2 : );
6469 2 :
6470 30 : match expected_item {
6471 26 : Some((_, img_value)) => {
6472 26 : let key_results = results.expect("No vectored get error expected");
6473 26 : let key_result = &key_results[&child_gap_at_key];
6474 26 : let returned_img = key_result
6475 26 : .as_ref()
6476 26 : .expect("No page reconstruct error expected");
6477 26 :
6478 26 : info!(
6479 2 : "Vectored read at LSN {} returned image {}",
6480 0 : query_lsn,
6481 0 : std::str::from_utf8(returned_img)?
6482 2 : );
6483 26 : assert_eq!(*returned_img, test_img(img_value));
6484 2 : }
6485 2 : None => {
6486 4 : assert!(matches!(results, Err(GetVectoredError::MissingKey(_))));
6487 2 : }
6488 2 : }
6489 2 : }
6490 2 :
6491 2 : Ok(())
6492 2 : }
6493 :
6494 : #[tokio::test]
6495 2 : async fn test_random_updates() -> anyhow::Result<()> {
6496 2 : let names_algorithms = [
6497 2 : ("test_random_updates_legacy", CompactionAlgorithm::Legacy),
6498 2 : ("test_random_updates_tiered", CompactionAlgorithm::Tiered),
6499 2 : ];
6500 6 : for (name, algorithm) in names_algorithms {
6501 96217 : test_random_updates_algorithm(name, algorithm).await?;
6502 2 : }
6503 2 : Ok(())
6504 2 : }
6505 :
6506 4 : async fn test_random_updates_algorithm(
6507 4 : name: &'static str,
6508 4 : compaction_algorithm: CompactionAlgorithm,
6509 4 : ) -> anyhow::Result<()> {
6510 4 : let mut harness = TenantHarness::create(name).await?;
6511 4 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
6512 4 : kind: compaction_algorithm,
6513 4 : };
6514 40 : let (tenant, ctx) = harness.load().await;
6515 4 : let tline = tenant
6516 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6517 9 : .await?;
6518 :
6519 : const NUM_KEYS: usize = 1000;
6520 4 : let cancel = CancellationToken::new();
6521 4 :
6522 4 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6523 4 : let mut test_key_end = test_key;
6524 4 : test_key_end.field6 = NUM_KEYS as u32;
6525 4 : tline.add_extra_test_dense_keyspace(KeySpace::single(test_key..test_key_end));
6526 4 :
6527 4 : let mut keyspace = KeySpaceAccum::new();
6528 4 :
6529 4 : // Track when each page was last modified. Used to assert that
6530 4 : // a read sees the latest page version.
6531 4 : let mut updated = [Lsn(0); NUM_KEYS];
6532 4 :
6533 4 : let mut lsn = Lsn(0x10);
6534 : #[allow(clippy::needless_range_loop)]
6535 4004 : for blknum in 0..NUM_KEYS {
6536 4000 : lsn = Lsn(lsn.0 + 0x10);
6537 4000 : test_key.field6 = blknum as u32;
6538 4000 : let mut writer = tline.writer().await;
6539 4000 : writer
6540 4000 : .put(
6541 4000 : test_key,
6542 4000 : lsn,
6543 4000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6544 4000 : &ctx,
6545 4000 : )
6546 4 : .await?;
6547 4000 : writer.finish_write(lsn);
6548 4000 : updated[blknum] = lsn;
6549 4000 : drop(writer);
6550 4000 :
6551 4000 : keyspace.add_key(test_key);
6552 : }
6553 :
6554 204 : for _ in 0..50 {
6555 200200 : for _ in 0..NUM_KEYS {
6556 200000 : lsn = Lsn(lsn.0 + 0x10);
6557 200000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6558 200000 : test_key.field6 = blknum as u32;
6559 200000 : let mut writer = tline.writer().await;
6560 200000 : writer
6561 200000 : .put(
6562 200000 : test_key,
6563 200000 : lsn,
6564 200000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6565 200000 : &ctx,
6566 200000 : )
6567 198 : .await?;
6568 200000 : writer.finish_write(lsn);
6569 200000 : drop(writer);
6570 200000 : updated[blknum] = lsn;
6571 : }
6572 :
6573 : // Read all the blocks
6574 200000 : for (blknum, last_lsn) in updated.iter().enumerate() {
6575 200000 : test_key.field6 = blknum as u32;
6576 200000 : assert_eq!(
6577 200000 : tline.get(test_key, lsn, &ctx).await?,
6578 200000 : test_img(&format!("{} at {}", blknum, last_lsn))
6579 : );
6580 : }
6581 :
6582 : // Perform a cycle of flush, and GC
6583 201 : tline.freeze_and_flush().await?;
6584 200 : tenant
6585 200 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
6586 0 : .await?;
6587 : }
6588 :
6589 4 : Ok(())
6590 4 : }
6591 :
6592 : #[tokio::test]
6593 2 : async fn test_traverse_branches() -> anyhow::Result<()> {
6594 2 : let (tenant, ctx) = TenantHarness::create("test_traverse_branches")
6595 2 : .await?
6596 2 : .load()
6597 20 : .await;
6598 2 : let mut tline = tenant
6599 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6600 6 : .await?;
6601 2 :
6602 2 : const NUM_KEYS: usize = 1000;
6603 2 :
6604 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6605 2 :
6606 2 : let mut keyspace = KeySpaceAccum::new();
6607 2 :
6608 2 : let cancel = CancellationToken::new();
6609 2 :
6610 2 : // Track when each page was last modified. Used to assert that
6611 2 : // a read sees the latest page version.
6612 2 : let mut updated = [Lsn(0); NUM_KEYS];
6613 2 :
6614 2 : let mut lsn = Lsn(0x10);
6615 2 : #[allow(clippy::needless_range_loop)]
6616 2002 : for blknum in 0..NUM_KEYS {
6617 2000 : lsn = Lsn(lsn.0 + 0x10);
6618 2000 : test_key.field6 = blknum as u32;
6619 2000 : let mut writer = tline.writer().await;
6620 2000 : writer
6621 2000 : .put(
6622 2000 : test_key,
6623 2000 : lsn,
6624 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6625 2000 : &ctx,
6626 2000 : )
6627 2 : .await?;
6628 2000 : writer.finish_write(lsn);
6629 2000 : updated[blknum] = lsn;
6630 2000 : drop(writer);
6631 2000 :
6632 2000 : keyspace.add_key(test_key);
6633 2 : }
6634 2 :
6635 102 : for _ in 0..50 {
6636 100 : let new_tline_id = TimelineId::generate();
6637 100 : tenant
6638 100 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
6639 91 : .await?;
6640 100 : tline = tenant
6641 100 : .get_timeline(new_tline_id, true)
6642 100 : .expect("Should have the branched timeline");
6643 2 :
6644 100100 : for _ in 0..NUM_KEYS {
6645 100000 : lsn = Lsn(lsn.0 + 0x10);
6646 100000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6647 100000 : test_key.field6 = blknum as u32;
6648 100000 : let mut writer = tline.writer().await;
6649 100000 : writer
6650 100000 : .put(
6651 100000 : test_key,
6652 100000 : lsn,
6653 100000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6654 100000 : &ctx,
6655 100000 : )
6656 792 : .await?;
6657 100000 : println!("updating {} at {}", blknum, lsn);
6658 100000 : writer.finish_write(lsn);
6659 100000 : drop(writer);
6660 100000 : updated[blknum] = lsn;
6661 2 : }
6662 2 :
6663 2 : // Read all the blocks
6664 100000 : for (blknum, last_lsn) in updated.iter().enumerate() {
6665 100000 : test_key.field6 = blknum as u32;
6666 100000 : assert_eq!(
6667 100000 : tline.get(test_key, lsn, &ctx).await?,
6668 100000 : test_img(&format!("{} at {}", blknum, last_lsn))
6669 2 : );
6670 2 : }
6671 2 :
6672 2 : // Perform a cycle of flush, compact, and GC
6673 102 : tline.freeze_and_flush().await?;
6674 15044 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
6675 100 : tenant
6676 100 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
6677 2 : .await?;
6678 2 : }
6679 2 :
6680 2 : Ok(())
6681 2 : }
6682 :
6683 : #[tokio::test]
6684 2 : async fn test_traverse_ancestors() -> anyhow::Result<()> {
6685 2 : let (tenant, ctx) = TenantHarness::create("test_traverse_ancestors")
6686 2 : .await?
6687 2 : .load()
6688 20 : .await;
6689 2 : let mut tline = tenant
6690 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6691 6 : .await?;
6692 2 :
6693 2 : const NUM_KEYS: usize = 100;
6694 2 : const NUM_TLINES: usize = 50;
6695 2 :
6696 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6697 2 : // Track page mutation lsns across different timelines.
6698 2 : let mut updated = [[Lsn(0); NUM_KEYS]; NUM_TLINES];
6699 2 :
6700 2 : let mut lsn = Lsn(0x10);
6701 2 :
6702 2 : #[allow(clippy::needless_range_loop)]
6703 102 : for idx in 0..NUM_TLINES {
6704 100 : let new_tline_id = TimelineId::generate();
6705 100 : tenant
6706 100 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
6707 64 : .await?;
6708 100 : tline = tenant
6709 100 : .get_timeline(new_tline_id, true)
6710 100 : .expect("Should have the branched timeline");
6711 2 :
6712 10100 : for _ in 0..NUM_KEYS {
6713 10000 : lsn = Lsn(lsn.0 + 0x10);
6714 10000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6715 10000 : test_key.field6 = blknum as u32;
6716 10000 : let mut writer = tline.writer().await;
6717 10000 : writer
6718 10000 : .put(
6719 10000 : test_key,
6720 10000 : lsn,
6721 10000 : &Value::Image(test_img(&format!("{} {} at {}", idx, blknum, lsn))),
6722 10000 : &ctx,
6723 10000 : )
6724 108 : .await?;
6725 10000 : println!("updating [{}][{}] at {}", idx, blknum, lsn);
6726 10000 : writer.finish_write(lsn);
6727 10000 : drop(writer);
6728 10000 : updated[idx][blknum] = lsn;
6729 2 : }
6730 2 : }
6731 2 :
6732 2 : // Read pages from leaf timeline across all ancestors.
6733 100 : for (idx, lsns) in updated.iter().enumerate() {
6734 10000 : for (blknum, lsn) in lsns.iter().enumerate() {
6735 2 : // Skip empty mutations.
6736 10000 : if lsn.0 == 0 {
6737 3694 : continue;
6738 6306 : }
6739 6306 : println!("checking [{idx}][{blknum}] at {lsn}");
6740 6306 : test_key.field6 = blknum as u32;
6741 6306 : assert_eq!(
6742 6306 : tline.get(test_key, *lsn, &ctx).await?,
6743 6306 : test_img(&format!("{idx} {blknum} at {lsn}"))
6744 2 : );
6745 2 : }
6746 2 : }
6747 2 : Ok(())
6748 2 : }
6749 :
6750 : #[tokio::test]
6751 2 : async fn test_write_at_initdb_lsn_takes_optimization_code_path() -> anyhow::Result<()> {
6752 2 : let (tenant, ctx) = TenantHarness::create("test_empty_test_timeline_is_usable")
6753 2 : .await?
6754 2 : .load()
6755 20 : .await;
6756 2 :
6757 2 : let initdb_lsn = Lsn(0x20);
6758 2 : let utline = tenant
6759 2 : .create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)
6760 2 : .await?;
6761 2 : let tline = utline.raw_timeline().unwrap();
6762 2 :
6763 2 : // Spawn flush loop now so that we can set the `expect_initdb_optimization`
6764 2 : tline.maybe_spawn_flush_loop();
6765 2 :
6766 2 : // Make sure the timeline has the minimum set of required keys for operation.
6767 2 : // The only operation you can always do on an empty timeline is to `put` new data.
6768 2 : // Except if you `put` at `initdb_lsn`.
6769 2 : // In that case, there's an optimization to directly create image layers instead of delta layers.
6770 2 : // It uses `repartition()`, which assumes some keys to be present.
6771 2 : // Let's make sure the test timeline can handle that case.
6772 2 : {
6773 2 : let mut state = tline.flush_loop_state.lock().unwrap();
6774 2 : assert_eq!(
6775 2 : timeline::FlushLoopState::Running {
6776 2 : expect_initdb_optimization: false,
6777 2 : initdb_optimization_count: 0,
6778 2 : },
6779 2 : *state
6780 2 : );
6781 2 : *state = timeline::FlushLoopState::Running {
6782 2 : expect_initdb_optimization: true,
6783 2 : initdb_optimization_count: 0,
6784 2 : };
6785 2 : }
6786 2 :
6787 2 : // Make writes at the initdb_lsn. When we flush it below, it should be handled by the optimization.
6788 2 : // As explained above, the optimization requires some keys to be present.
6789 2 : // As per `create_empty_timeline` documentation, use init_empty to set them.
6790 2 : // This is what `create_test_timeline` does, by the way.
6791 2 : let mut modification = tline.begin_modification(initdb_lsn);
6792 2 : modification
6793 2 : .init_empty_test_timeline()
6794 2 : .context("init_empty_test_timeline")?;
6795 2 : modification
6796 2 : .commit(&ctx)
6797 2 : .await
6798 2 : .context("commit init_empty_test_timeline modification")?;
6799 2 :
6800 2 : // Do the flush. The flush code will check the expectations that we set above.
6801 2 : tline.freeze_and_flush().await?;
6802 2 :
6803 2 : // assert freeze_and_flush exercised the initdb optimization
6804 2 : {
6805 2 : let state = tline.flush_loop_state.lock().unwrap();
6806 2 : let timeline::FlushLoopState::Running {
6807 2 : expect_initdb_optimization,
6808 2 : initdb_optimization_count,
6809 2 : } = *state
6810 2 : else {
6811 2 : panic!("unexpected state: {:?}", *state);
6812 2 : };
6813 2 : assert!(expect_initdb_optimization);
6814 2 : assert!(initdb_optimization_count > 0);
6815 2 : }
6816 2 : Ok(())
6817 2 : }
6818 :
6819 : #[tokio::test]
6820 2 : async fn test_create_guard_crash() -> anyhow::Result<()> {
6821 2 : let name = "test_create_guard_crash";
6822 2 : let harness = TenantHarness::create(name).await?;
6823 2 : {
6824 20 : let (tenant, ctx) = harness.load().await;
6825 2 : let tline = tenant
6826 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
6827 2 : .await?;
6828 2 : // Leave the timeline ID in [`Tenant::timelines_creating`] to exclude attempting to create it again
6829 2 : let raw_tline = tline.raw_timeline().unwrap();
6830 2 : raw_tline
6831 2 : .shutdown(super::timeline::ShutdownMode::Hard)
6832 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))
6833 2 : .await;
6834 2 : std::mem::forget(tline);
6835 2 : }
6836 2 :
6837 20 : let (tenant, _) = harness.load().await;
6838 2 : match tenant.get_timeline(TIMELINE_ID, false) {
6839 2 : Ok(_) => panic!("timeline should've been removed during load"),
6840 2 : Err(e) => {
6841 2 : assert_eq!(
6842 2 : e,
6843 2 : GetTimelineError::NotFound {
6844 2 : tenant_id: tenant.tenant_shard_id,
6845 2 : timeline_id: TIMELINE_ID,
6846 2 : }
6847 2 : )
6848 2 : }
6849 2 : }
6850 2 :
6851 2 : assert!(!harness
6852 2 : .conf
6853 2 : .timeline_path(&tenant.tenant_shard_id, &TIMELINE_ID)
6854 2 : .exists());
6855 2 :
6856 2 : Ok(())
6857 2 : }
6858 :
6859 : #[tokio::test]
6860 2 : async fn test_read_at_max_lsn() -> anyhow::Result<()> {
6861 2 : let names_algorithms = [
6862 2 : ("test_read_at_max_lsn_legacy", CompactionAlgorithm::Legacy),
6863 2 : ("test_read_at_max_lsn_tiered", CompactionAlgorithm::Tiered),
6864 2 : ];
6865 6 : for (name, algorithm) in names_algorithms {
6866 32561 : test_read_at_max_lsn_algorithm(name, algorithm).await?;
6867 2 : }
6868 2 : Ok(())
6869 2 : }
6870 :
6871 4 : async fn test_read_at_max_lsn_algorithm(
6872 4 : name: &'static str,
6873 4 : compaction_algorithm: CompactionAlgorithm,
6874 4 : ) -> anyhow::Result<()> {
6875 4 : let mut harness = TenantHarness::create(name).await?;
6876 4 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
6877 4 : kind: compaction_algorithm,
6878 4 : };
6879 40 : let (tenant, ctx) = harness.load().await;
6880 4 : let tline = tenant
6881 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6882 11 : .await?;
6883 :
6884 4 : let lsn = Lsn(0x10);
6885 4 : let compact = false;
6886 32100 : bulk_insert_maybe_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000, compact).await?;
6887 :
6888 4 : let test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6889 4 : let read_lsn = Lsn(u64::MAX - 1);
6890 :
6891 410 : let result = tline.get(test_key, read_lsn, &ctx).await;
6892 4 : assert!(result.is_ok(), "result is not Ok: {}", result.unwrap_err());
6893 :
6894 4 : Ok(())
6895 4 : }
6896 :
6897 : #[tokio::test]
6898 2 : async fn test_metadata_scan() -> anyhow::Result<()> {
6899 2 : let harness = TenantHarness::create("test_metadata_scan").await?;
6900 20 : let (tenant, ctx) = harness.load().await;
6901 2 : let tline = tenant
6902 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6903 6 : .await?;
6904 2 :
6905 2 : const NUM_KEYS: usize = 1000;
6906 2 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
6907 2 :
6908 2 : let cancel = CancellationToken::new();
6909 2 :
6910 2 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
6911 2 : base_key.field1 = AUX_KEY_PREFIX;
6912 2 : let mut test_key = base_key;
6913 2 :
6914 2 : // Track when each page was last modified. Used to assert that
6915 2 : // a read sees the latest page version.
6916 2 : let mut updated = [Lsn(0); NUM_KEYS];
6917 2 :
6918 2 : let mut lsn = Lsn(0x10);
6919 2 : #[allow(clippy::needless_range_loop)]
6920 2002 : for blknum in 0..NUM_KEYS {
6921 2000 : lsn = Lsn(lsn.0 + 0x10);
6922 2000 : test_key.field6 = (blknum * STEP) as u32;
6923 2000 : let mut writer = tline.writer().await;
6924 2000 : writer
6925 2000 : .put(
6926 2000 : test_key,
6927 2000 : lsn,
6928 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6929 2000 : &ctx,
6930 2000 : )
6931 2 : .await?;
6932 2000 : writer.finish_write(lsn);
6933 2000 : updated[blknum] = lsn;
6934 2000 : drop(writer);
6935 2 : }
6936 2 :
6937 2 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
6938 2 :
6939 24 : for iter in 0..=10 {
6940 2 : // Read all the blocks
6941 22000 : for (blknum, last_lsn) in updated.iter().enumerate() {
6942 22000 : test_key.field6 = (blknum * STEP) as u32;
6943 22000 : assert_eq!(
6944 22000 : tline.get(test_key, lsn, &ctx).await?,
6945 22000 : test_img(&format!("{} at {}", blknum, last_lsn))
6946 2 : );
6947 2 : }
6948 2 :
6949 22 : let mut cnt = 0;
6950 22000 : for (key, value) in tline
6951 22 : .get_vectored_impl(
6952 22 : keyspace.clone(),
6953 22 : lsn,
6954 22 : &mut ValuesReconstructState::default(),
6955 22 : &ctx,
6956 22 : )
6957 736 : .await?
6958 2 : {
6959 22000 : let blknum = key.field6 as usize;
6960 22000 : let value = value?;
6961 22000 : assert!(blknum % STEP == 0);
6962 22000 : let blknum = blknum / STEP;
6963 22000 : assert_eq!(
6964 22000 : value,
6965 22000 : test_img(&format!("{} at {}", blknum, updated[blknum]))
6966 22000 : );
6967 22000 : cnt += 1;
6968 2 : }
6969 2 :
6970 22 : assert_eq!(cnt, NUM_KEYS);
6971 2 :
6972 22022 : for _ in 0..NUM_KEYS {
6973 22000 : lsn = Lsn(lsn.0 + 0x10);
6974 22000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6975 22000 : test_key.field6 = (blknum * STEP) as u32;
6976 22000 : let mut writer = tline.writer().await;
6977 22000 : writer
6978 22000 : .put(
6979 22000 : test_key,
6980 22000 : lsn,
6981 22000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6982 22000 : &ctx,
6983 22000 : )
6984 111 : .await?;
6985 22000 : writer.finish_write(lsn);
6986 22000 : drop(writer);
6987 22000 : updated[blknum] = lsn;
6988 2 : }
6989 2 :
6990 2 : // Perform two cycles of flush, compact, and GC
6991 66 : for round in 0..2 {
6992 44 : tline.freeze_and_flush().await?;
6993 44 : tline
6994 44 : .compact(
6995 44 : &cancel,
6996 44 : if iter % 5 == 0 && round == 0 {
6997 6 : let mut flags = EnumSet::new();
6998 6 : flags.insert(CompactFlags::ForceImageLayerCreation);
6999 6 : flags.insert(CompactFlags::ForceRepartition);
7000 6 : flags
7001 2 : } else {
7002 38 : EnumSet::empty()
7003 2 : },
7004 44 : &ctx,
7005 2 : )
7006 6672 : .await?;
7007 44 : tenant
7008 44 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7009 2 : .await?;
7010 2 : }
7011 2 : }
7012 2 :
7013 2 : Ok(())
7014 2 : }
7015 :
7016 : #[tokio::test]
7017 2 : async fn test_metadata_compaction_trigger() -> anyhow::Result<()> {
7018 2 : let harness = TenantHarness::create("test_metadata_compaction_trigger").await?;
7019 19 : let (tenant, ctx) = harness.load().await;
7020 2 : let tline = tenant
7021 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7022 6 : .await?;
7023 2 :
7024 2 : let cancel = CancellationToken::new();
7025 2 :
7026 2 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7027 2 : base_key.field1 = AUX_KEY_PREFIX;
7028 2 : let test_key = base_key;
7029 2 : let mut lsn = Lsn(0x10);
7030 2 :
7031 42 : for _ in 0..20 {
7032 40 : lsn = Lsn(lsn.0 + 0x10);
7033 40 : let mut writer = tline.writer().await;
7034 40 : writer
7035 40 : .put(
7036 40 : test_key,
7037 40 : lsn,
7038 40 : &Value::Image(test_img(&format!("{} at {}", 0, lsn))),
7039 40 : &ctx,
7040 40 : )
7041 20 : .await?;
7042 40 : writer.finish_write(lsn);
7043 40 : drop(writer);
7044 40 : tline.freeze_and_flush().await?; // force create a delta layer
7045 2 : }
7046 2 :
7047 2 : let before_num_l0_delta_files =
7048 2 : tline.layers.read().await.layer_map()?.level0_deltas().len();
7049 2 :
7050 110 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
7051 2 :
7052 2 : let after_num_l0_delta_files = tline.layers.read().await.layer_map()?.level0_deltas().len();
7053 2 :
7054 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}");
7055 2 :
7056 2 : assert_eq!(
7057 4 : tline.get(test_key, lsn, &ctx).await?,
7058 2 : test_img(&format!("{} at {}", 0, lsn))
7059 2 : );
7060 2 :
7061 2 : Ok(())
7062 2 : }
7063 :
7064 : #[tokio::test]
7065 2 : async fn test_aux_file_e2e() {
7066 2 : let harness = TenantHarness::create("test_aux_file_e2e").await.unwrap();
7067 2 :
7068 20 : let (tenant, ctx) = harness.load().await;
7069 2 :
7070 2 : let mut lsn = Lsn(0x08);
7071 2 :
7072 2 : let tline: Arc<Timeline> = tenant
7073 2 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
7074 6 : .await
7075 2 : .unwrap();
7076 2 :
7077 2 : {
7078 2 : lsn += 8;
7079 2 : let mut modification = tline.begin_modification(lsn);
7080 2 : modification
7081 2 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
7082 2 : .await
7083 2 : .unwrap();
7084 2 : modification.commit(&ctx).await.unwrap();
7085 2 : }
7086 2 :
7087 2 : // we can read everything from the storage
7088 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
7089 2 : assert_eq!(
7090 2 : files.get("pg_logical/mappings/test1"),
7091 2 : Some(&bytes::Bytes::from_static(b"first"))
7092 2 : );
7093 2 :
7094 2 : {
7095 2 : lsn += 8;
7096 2 : let mut modification = tline.begin_modification(lsn);
7097 2 : modification
7098 2 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
7099 2 : .await
7100 2 : .unwrap();
7101 2 : modification.commit(&ctx).await.unwrap();
7102 2 : }
7103 2 :
7104 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
7105 2 : assert_eq!(
7106 2 : files.get("pg_logical/mappings/test2"),
7107 2 : Some(&bytes::Bytes::from_static(b"second"))
7108 2 : );
7109 2 :
7110 2 : let child = tenant
7111 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(lsn), &ctx)
7112 2 : .await
7113 2 : .unwrap();
7114 2 :
7115 2 : let files = child.list_aux_files(lsn, &ctx).await.unwrap();
7116 2 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
7117 2 : assert_eq!(files.get("pg_logical/mappings/test2"), None);
7118 2 : }
7119 :
7120 : #[tokio::test]
7121 2 : async fn test_metadata_image_creation() -> anyhow::Result<()> {
7122 2 : let harness = TenantHarness::create("test_metadata_image_creation").await?;
7123 20 : let (tenant, ctx) = harness.load().await;
7124 2 : let tline = tenant
7125 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7126 6 : .await?;
7127 2 :
7128 2 : const NUM_KEYS: usize = 1000;
7129 2 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
7130 2 :
7131 2 : let cancel = CancellationToken::new();
7132 2 :
7133 2 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7134 2 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
7135 2 : let mut test_key = base_key;
7136 2 : let mut lsn = Lsn(0x10);
7137 2 :
7138 8 : async fn scan_with_statistics(
7139 8 : tline: &Timeline,
7140 8 : keyspace: &KeySpace,
7141 8 : lsn: Lsn,
7142 8 : ctx: &RequestContext,
7143 8 : ) -> anyhow::Result<(BTreeMap<Key, Result<Bytes, PageReconstructError>>, usize)> {
7144 8 : let mut reconstruct_state = ValuesReconstructState::default();
7145 8 : let res = tline
7146 8 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
7147 250 : .await?;
7148 8 : Ok((res, reconstruct_state.get_delta_layers_visited() as usize))
7149 8 : }
7150 2 :
7151 2 : #[allow(clippy::needless_range_loop)]
7152 2002 : for blknum in 0..NUM_KEYS {
7153 2000 : lsn = Lsn(lsn.0 + 0x10);
7154 2000 : test_key.field6 = (blknum * STEP) as u32;
7155 2000 : let mut writer = tline.writer().await;
7156 2000 : writer
7157 2000 : .put(
7158 2000 : test_key,
7159 2000 : lsn,
7160 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7161 2000 : &ctx,
7162 2000 : )
7163 2 : .await?;
7164 2000 : writer.finish_write(lsn);
7165 2000 : drop(writer);
7166 2 : }
7167 2 :
7168 2 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
7169 2 :
7170 22 : for iter in 1..=10 {
7171 20020 : for _ in 0..NUM_KEYS {
7172 20000 : lsn = Lsn(lsn.0 + 0x10);
7173 20000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7174 20000 : test_key.field6 = (blknum * STEP) as u32;
7175 20000 : let mut writer = tline.writer().await;
7176 20000 : writer
7177 20000 : .put(
7178 20000 : test_key,
7179 20000 : lsn,
7180 20000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7181 20000 : &ctx,
7182 20000 : )
7183 35 : .await?;
7184 20000 : writer.finish_write(lsn);
7185 20000 : drop(writer);
7186 2 : }
7187 2 :
7188 20 : tline.freeze_and_flush().await?;
7189 2 :
7190 20 : if iter % 5 == 0 {
7191 4 : let (_, before_delta_file_accessed) =
7192 242 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
7193 4 : tline
7194 4 : .compact(
7195 4 : &cancel,
7196 4 : {
7197 4 : let mut flags = EnumSet::new();
7198 4 : flags.insert(CompactFlags::ForceImageLayerCreation);
7199 4 : flags.insert(CompactFlags::ForceRepartition);
7200 4 : flags
7201 4 : },
7202 4 : &ctx,
7203 4 : )
7204 4818 : .await?;
7205 4 : let (_, after_delta_file_accessed) =
7206 8 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
7207 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}");
7208 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.
7209 4 : assert!(
7210 4 : after_delta_file_accessed <= 2,
7211 2 : "after_delta_file_accessed={after_delta_file_accessed}"
7212 2 : );
7213 16 : }
7214 2 : }
7215 2 :
7216 2 : Ok(())
7217 2 : }
7218 :
7219 : #[tokio::test]
7220 2 : async fn test_vectored_missing_data_key_reads() -> anyhow::Result<()> {
7221 2 : let harness = TenantHarness::create("test_vectored_missing_data_key_reads").await?;
7222 17 : let (tenant, ctx) = harness.load().await;
7223 2 :
7224 2 : let base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7225 2 : let base_key_child = Key::from_hex("000000000033333333444444445500000001").unwrap();
7226 2 : let base_key_nonexist = Key::from_hex("000000000033333333444444445500000002").unwrap();
7227 2 :
7228 2 : let tline = tenant
7229 2 : .create_test_timeline_with_layers(
7230 2 : TIMELINE_ID,
7231 2 : Lsn(0x10),
7232 2 : DEFAULT_PG_VERSION,
7233 2 : &ctx,
7234 2 : Vec::new(), // delta layers
7235 2 : vec![(Lsn(0x20), vec![(base_key, test_img("data key 1"))])], // image layers
7236 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
7237 2 : )
7238 12 : .await?;
7239 2 : tline.add_extra_test_dense_keyspace(KeySpace::single(base_key..(base_key_nonexist.next())));
7240 2 :
7241 2 : let child = tenant
7242 2 : .branch_timeline_test_with_layers(
7243 2 : &tline,
7244 2 : NEW_TIMELINE_ID,
7245 2 : Some(Lsn(0x20)),
7246 2 : &ctx,
7247 2 : Vec::new(), // delta layers
7248 2 : vec![(Lsn(0x30), vec![(base_key_child, test_img("data key 2"))])], // image layers
7249 2 : Lsn(0x30),
7250 2 : )
7251 9 : .await
7252 2 : .unwrap();
7253 2 :
7254 2 : let lsn = Lsn(0x30);
7255 2 :
7256 2 : // test vectored get on parent timeline
7257 2 : assert_eq!(
7258 4 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
7259 2 : Some(test_img("data key 1"))
7260 2 : );
7261 2 : assert!(get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx)
7262 3 : .await
7263 2 : .unwrap_err()
7264 2 : .is_missing_key_error());
7265 2 : assert!(
7266 2 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx)
7267 2 : .await
7268 2 : .unwrap_err()
7269 2 : .is_missing_key_error()
7270 2 : );
7271 2 :
7272 2 : // test vectored get on child timeline
7273 2 : assert_eq!(
7274 2 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
7275 2 : Some(test_img("data key 1"))
7276 2 : );
7277 2 : assert_eq!(
7278 4 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
7279 2 : Some(test_img("data key 2"))
7280 2 : );
7281 2 : assert!(
7282 2 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx)
7283 2 : .await
7284 2 : .unwrap_err()
7285 2 : .is_missing_key_error()
7286 2 : );
7287 2 :
7288 2 : Ok(())
7289 2 : }
7290 :
7291 : #[tokio::test]
7292 2 : async fn test_vectored_missing_metadata_key_reads() -> anyhow::Result<()> {
7293 2 : let harness = TenantHarness::create("test_vectored_missing_metadata_key_reads").await?;
7294 20 : let (tenant, ctx) = harness.load().await;
7295 2 :
7296 2 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7297 2 : let base_key_child = Key::from_hex("620000000033333333444444445500000001").unwrap();
7298 2 : let base_key_nonexist = Key::from_hex("620000000033333333444444445500000002").unwrap();
7299 2 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
7300 2 :
7301 2 : let tline = tenant
7302 2 : .create_test_timeline_with_layers(
7303 2 : TIMELINE_ID,
7304 2 : Lsn(0x10),
7305 2 : DEFAULT_PG_VERSION,
7306 2 : &ctx,
7307 2 : Vec::new(), // delta layers
7308 2 : vec![(Lsn(0x20), vec![(base_key, test_img("metadata key 1"))])], // image layers
7309 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
7310 2 : )
7311 13 : .await?;
7312 2 :
7313 2 : let child = tenant
7314 2 : .branch_timeline_test_with_layers(
7315 2 : &tline,
7316 2 : NEW_TIMELINE_ID,
7317 2 : Some(Lsn(0x20)),
7318 2 : &ctx,
7319 2 : Vec::new(), // delta layers
7320 2 : vec![(
7321 2 : Lsn(0x30),
7322 2 : vec![(base_key_child, test_img("metadata key 2"))],
7323 2 : )], // image layers
7324 2 : Lsn(0x30),
7325 2 : )
7326 9 : .await
7327 2 : .unwrap();
7328 2 :
7329 2 : let lsn = Lsn(0x30);
7330 2 :
7331 2 : // test vectored get on parent timeline
7332 2 : assert_eq!(
7333 4 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
7334 2 : Some(test_img("metadata key 1"))
7335 2 : );
7336 2 : assert_eq!(
7337 2 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx).await?,
7338 2 : None
7339 2 : );
7340 2 : assert_eq!(
7341 2 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx).await?,
7342 2 : None
7343 2 : );
7344 2 :
7345 2 : // test vectored get on child timeline
7346 2 : assert_eq!(
7347 2 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
7348 2 : None
7349 2 : );
7350 2 : assert_eq!(
7351 4 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
7352 2 : Some(test_img("metadata key 2"))
7353 2 : );
7354 2 : assert_eq!(
7355 2 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx).await?,
7356 2 : None
7357 2 : );
7358 2 :
7359 2 : Ok(())
7360 2 : }
7361 :
7362 36 : async fn get_vectored_impl_wrapper(
7363 36 : tline: &Arc<Timeline>,
7364 36 : key: Key,
7365 36 : lsn: Lsn,
7366 36 : ctx: &RequestContext,
7367 36 : ) -> Result<Option<Bytes>, GetVectoredError> {
7368 36 : let mut reconstruct_state = ValuesReconstructState::new();
7369 36 : let mut res = tline
7370 36 : .get_vectored_impl(
7371 36 : KeySpace::single(key..key.next()),
7372 36 : lsn,
7373 36 : &mut reconstruct_state,
7374 36 : ctx,
7375 36 : )
7376 40 : .await?;
7377 30 : Ok(res.pop_last().map(|(k, v)| {
7378 18 : assert_eq!(k, key);
7379 18 : v.unwrap()
7380 30 : }))
7381 36 : }
7382 :
7383 : #[tokio::test]
7384 2 : async fn test_metadata_tombstone_reads() -> anyhow::Result<()> {
7385 2 : let harness = TenantHarness::create("test_metadata_tombstone_reads").await?;
7386 20 : let (tenant, ctx) = harness.load().await;
7387 2 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
7388 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
7389 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
7390 2 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
7391 2 :
7392 2 : // We emulate the situation that the compaction algorithm creates an image layer that removes the tombstones
7393 2 : // Lsn 0x30 key0, key3, no key1+key2
7394 2 : // Lsn 0x20 key1+key2 tomestones
7395 2 : // Lsn 0x10 key1 in image, key2 in delta
7396 2 : let tline = tenant
7397 2 : .create_test_timeline_with_layers(
7398 2 : TIMELINE_ID,
7399 2 : Lsn(0x10),
7400 2 : DEFAULT_PG_VERSION,
7401 2 : &ctx,
7402 2 : // delta layers
7403 2 : vec![
7404 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7405 2 : Lsn(0x10)..Lsn(0x20),
7406 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
7407 2 : ),
7408 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7409 2 : Lsn(0x20)..Lsn(0x30),
7410 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
7411 2 : ),
7412 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7413 2 : Lsn(0x20)..Lsn(0x30),
7414 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
7415 2 : ),
7416 2 : ],
7417 2 : // image layers
7418 2 : vec![
7419 2 : (Lsn(0x10), vec![(key1, test_img("metadata key 1"))]),
7420 2 : (
7421 2 : Lsn(0x30),
7422 2 : vec![
7423 2 : (key0, test_img("metadata key 0")),
7424 2 : (key3, test_img("metadata key 3")),
7425 2 : ],
7426 2 : ),
7427 2 : ],
7428 2 : Lsn(0x30),
7429 2 : )
7430 40 : .await?;
7431 2 :
7432 2 : let lsn = Lsn(0x30);
7433 2 : let old_lsn = Lsn(0x20);
7434 2 :
7435 2 : assert_eq!(
7436 4 : get_vectored_impl_wrapper(&tline, key0, lsn, &ctx).await?,
7437 2 : Some(test_img("metadata key 0"))
7438 2 : );
7439 2 : assert_eq!(
7440 2 : get_vectored_impl_wrapper(&tline, key1, lsn, &ctx).await?,
7441 2 : None,
7442 2 : );
7443 2 : assert_eq!(
7444 2 : get_vectored_impl_wrapper(&tline, key2, lsn, &ctx).await?,
7445 2 : None,
7446 2 : );
7447 2 : assert_eq!(
7448 8 : get_vectored_impl_wrapper(&tline, key1, old_lsn, &ctx).await?,
7449 2 : Some(Bytes::new()),
7450 2 : );
7451 2 : assert_eq!(
7452 7 : get_vectored_impl_wrapper(&tline, key2, old_lsn, &ctx).await?,
7453 2 : Some(Bytes::new()),
7454 2 : );
7455 2 : assert_eq!(
7456 2 : get_vectored_impl_wrapper(&tline, key3, lsn, &ctx).await?,
7457 2 : Some(test_img("metadata key 3"))
7458 2 : );
7459 2 :
7460 2 : Ok(())
7461 2 : }
7462 :
7463 : #[tokio::test]
7464 2 : async fn test_metadata_tombstone_image_creation() {
7465 2 : let harness = TenantHarness::create("test_metadata_tombstone_image_creation")
7466 2 : .await
7467 2 : .unwrap();
7468 20 : let (tenant, ctx) = harness.load().await;
7469 2 :
7470 2 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
7471 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
7472 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
7473 2 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
7474 2 :
7475 2 : let tline = tenant
7476 2 : .create_test_timeline_with_layers(
7477 2 : TIMELINE_ID,
7478 2 : Lsn(0x10),
7479 2 : DEFAULT_PG_VERSION,
7480 2 : &ctx,
7481 2 : // delta layers
7482 2 : vec![
7483 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7484 2 : Lsn(0x10)..Lsn(0x20),
7485 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
7486 2 : ),
7487 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7488 2 : Lsn(0x20)..Lsn(0x30),
7489 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
7490 2 : ),
7491 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7492 2 : Lsn(0x20)..Lsn(0x30),
7493 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
7494 2 : ),
7495 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7496 2 : Lsn(0x30)..Lsn(0x40),
7497 2 : vec![
7498 2 : (key0, Lsn(0x30), Value::Image(test_img("metadata key 0"))),
7499 2 : (key3, Lsn(0x30), Value::Image(test_img("metadata key 3"))),
7500 2 : ],
7501 2 : ),
7502 2 : ],
7503 2 : // image layers
7504 2 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
7505 2 : Lsn(0x40),
7506 2 : )
7507 37 : .await
7508 2 : .unwrap();
7509 2 :
7510 2 : let cancel = CancellationToken::new();
7511 2 :
7512 2 : tline
7513 2 : .compact(
7514 2 : &cancel,
7515 2 : {
7516 2 : let mut flags = EnumSet::new();
7517 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
7518 2 : flags.insert(CompactFlags::ForceRepartition);
7519 2 : flags
7520 2 : },
7521 2 : &ctx,
7522 2 : )
7523 62 : .await
7524 2 : .unwrap();
7525 2 :
7526 2 : // Image layers are created at last_record_lsn
7527 2 : let images = tline
7528 2 : .inspect_image_layers(Lsn(0x40), &ctx)
7529 8 : .await
7530 2 : .unwrap()
7531 2 : .into_iter()
7532 18 : .filter(|(k, _)| k.is_metadata_key())
7533 2 : .collect::<Vec<_>>();
7534 2 : assert_eq!(images.len(), 2); // the image layer should only contain two existing keys, tombstones should be removed.
7535 2 : }
7536 :
7537 : #[tokio::test]
7538 2 : async fn test_metadata_tombstone_empty_image_creation() {
7539 2 : let harness = TenantHarness::create("test_metadata_tombstone_empty_image_creation")
7540 2 : .await
7541 2 : .unwrap();
7542 20 : let (tenant, ctx) = harness.load().await;
7543 2 :
7544 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
7545 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
7546 2 :
7547 2 : let tline = tenant
7548 2 : .create_test_timeline_with_layers(
7549 2 : TIMELINE_ID,
7550 2 : Lsn(0x10),
7551 2 : DEFAULT_PG_VERSION,
7552 2 : &ctx,
7553 2 : // delta layers
7554 2 : vec![
7555 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7556 2 : Lsn(0x10)..Lsn(0x20),
7557 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
7558 2 : ),
7559 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7560 2 : Lsn(0x20)..Lsn(0x30),
7561 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
7562 2 : ),
7563 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7564 2 : Lsn(0x20)..Lsn(0x30),
7565 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
7566 2 : ),
7567 2 : ],
7568 2 : // image layers
7569 2 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
7570 2 : Lsn(0x30),
7571 2 : )
7572 31 : .await
7573 2 : .unwrap();
7574 2 :
7575 2 : let cancel = CancellationToken::new();
7576 2 :
7577 2 : tline
7578 2 : .compact(
7579 2 : &cancel,
7580 2 : {
7581 2 : let mut flags = EnumSet::new();
7582 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
7583 2 : flags.insert(CompactFlags::ForceRepartition);
7584 2 : flags
7585 2 : },
7586 2 : &ctx,
7587 2 : )
7588 50 : .await
7589 2 : .unwrap();
7590 2 :
7591 2 : // Image layers are created at last_record_lsn
7592 2 : let images = tline
7593 2 : .inspect_image_layers(Lsn(0x30), &ctx)
7594 4 : .await
7595 2 : .unwrap()
7596 2 : .into_iter()
7597 14 : .filter(|(k, _)| k.is_metadata_key())
7598 2 : .collect::<Vec<_>>();
7599 2 : assert_eq!(images.len(), 0); // the image layer should not contain tombstones, or it is not created
7600 2 : }
7601 :
7602 : #[tokio::test]
7603 2 : async fn test_simple_bottom_most_compaction_images() -> anyhow::Result<()> {
7604 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_images").await?;
7605 20 : let (tenant, ctx) = harness.load().await;
7606 2 :
7607 102 : fn get_key(id: u32) -> Key {
7608 102 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7609 102 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7610 102 : key.field6 = id;
7611 102 : key
7612 102 : }
7613 2 :
7614 2 : // We create
7615 2 : // - one bottom-most image layer,
7616 2 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
7617 2 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
7618 2 : // - a delta layer D3 above the horizon.
7619 2 : //
7620 2 : // | D3 |
7621 2 : // | D1 |
7622 2 : // -| |-- gc horizon -----------------
7623 2 : // | | | D2 |
7624 2 : // --------- img layer ------------------
7625 2 : //
7626 2 : // What we should expact from this compaction is:
7627 2 : // | D3 |
7628 2 : // | Part of D1 |
7629 2 : // --------- img layer with D1+D2 at GC horizon------------------
7630 2 :
7631 2 : // img layer at 0x10
7632 2 : let img_layer = (0..10)
7633 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
7634 2 : .collect_vec();
7635 2 :
7636 2 : let delta1 = vec![
7637 2 : (
7638 2 : get_key(1),
7639 2 : Lsn(0x20),
7640 2 : Value::Image(Bytes::from("value 1@0x20")),
7641 2 : ),
7642 2 : (
7643 2 : get_key(2),
7644 2 : Lsn(0x30),
7645 2 : Value::Image(Bytes::from("value 2@0x30")),
7646 2 : ),
7647 2 : (
7648 2 : get_key(3),
7649 2 : Lsn(0x40),
7650 2 : Value::Image(Bytes::from("value 3@0x40")),
7651 2 : ),
7652 2 : ];
7653 2 : let delta2 = vec![
7654 2 : (
7655 2 : get_key(5),
7656 2 : Lsn(0x20),
7657 2 : Value::Image(Bytes::from("value 5@0x20")),
7658 2 : ),
7659 2 : (
7660 2 : get_key(6),
7661 2 : Lsn(0x20),
7662 2 : Value::Image(Bytes::from("value 6@0x20")),
7663 2 : ),
7664 2 : ];
7665 2 : let delta3 = vec![
7666 2 : (
7667 2 : get_key(8),
7668 2 : Lsn(0x48),
7669 2 : Value::Image(Bytes::from("value 8@0x48")),
7670 2 : ),
7671 2 : (
7672 2 : get_key(9),
7673 2 : Lsn(0x48),
7674 2 : Value::Image(Bytes::from("value 9@0x48")),
7675 2 : ),
7676 2 : ];
7677 2 :
7678 2 : let tline = tenant
7679 2 : .create_test_timeline_with_layers(
7680 2 : TIMELINE_ID,
7681 2 : Lsn(0x10),
7682 2 : DEFAULT_PG_VERSION,
7683 2 : &ctx,
7684 2 : vec![
7685 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
7686 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
7687 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
7688 2 : ], // delta layers
7689 2 : vec![(Lsn(0x10), img_layer)], // image layers
7690 2 : Lsn(0x50),
7691 2 : )
7692 49 : .await?;
7693 2 : {
7694 2 : // Update GC info
7695 2 : let mut guard = tline.gc_info.write().unwrap();
7696 2 : guard.cutoffs.time = Lsn(0x30);
7697 2 : guard.cutoffs.space = Lsn(0x30);
7698 2 : }
7699 2 :
7700 2 : let expected_result = [
7701 2 : Bytes::from_static(b"value 0@0x10"),
7702 2 : Bytes::from_static(b"value 1@0x20"),
7703 2 : Bytes::from_static(b"value 2@0x30"),
7704 2 : Bytes::from_static(b"value 3@0x40"),
7705 2 : Bytes::from_static(b"value 4@0x10"),
7706 2 : Bytes::from_static(b"value 5@0x20"),
7707 2 : Bytes::from_static(b"value 6@0x20"),
7708 2 : Bytes::from_static(b"value 7@0x10"),
7709 2 : Bytes::from_static(b"value 8@0x48"),
7710 2 : Bytes::from_static(b"value 9@0x48"),
7711 2 : ];
7712 2 :
7713 20 : for (idx, expected) in expected_result.iter().enumerate() {
7714 20 : assert_eq!(
7715 20 : tline
7716 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7717 30 : .await
7718 20 : .unwrap(),
7719 2 : expected
7720 2 : );
7721 2 : }
7722 2 :
7723 2 : let cancel = CancellationToken::new();
7724 2 : tline
7725 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7726 56 : .await
7727 2 : .unwrap();
7728 2 :
7729 20 : for (idx, expected) in expected_result.iter().enumerate() {
7730 20 : assert_eq!(
7731 20 : tline
7732 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7733 20 : .await
7734 20 : .unwrap(),
7735 2 : expected
7736 2 : );
7737 2 : }
7738 2 :
7739 2 : // Check if the image layer at the GC horizon contains exactly what we want
7740 2 : let image_at_gc_horizon = tline
7741 2 : .inspect_image_layers(Lsn(0x30), &ctx)
7742 2 : .await
7743 2 : .unwrap()
7744 2 : .into_iter()
7745 34 : .filter(|(k, _)| k.is_metadata_key())
7746 2 : .collect::<Vec<_>>();
7747 2 :
7748 2 : assert_eq!(image_at_gc_horizon.len(), 10);
7749 2 : let expected_result = [
7750 2 : Bytes::from_static(b"value 0@0x10"),
7751 2 : Bytes::from_static(b"value 1@0x20"),
7752 2 : Bytes::from_static(b"value 2@0x30"),
7753 2 : Bytes::from_static(b"value 3@0x10"),
7754 2 : Bytes::from_static(b"value 4@0x10"),
7755 2 : Bytes::from_static(b"value 5@0x20"),
7756 2 : Bytes::from_static(b"value 6@0x20"),
7757 2 : Bytes::from_static(b"value 7@0x10"),
7758 2 : Bytes::from_static(b"value 8@0x10"),
7759 2 : Bytes::from_static(b"value 9@0x10"),
7760 2 : ];
7761 22 : for idx in 0..10 {
7762 20 : assert_eq!(
7763 20 : image_at_gc_horizon[idx],
7764 20 : (get_key(idx as u32), expected_result[idx].clone())
7765 20 : );
7766 2 : }
7767 2 :
7768 2 : // Check if old layers are removed / new layers have the expected LSN
7769 2 : let all_layers = inspect_and_sort(&tline, None).await;
7770 2 : assert_eq!(
7771 2 : all_layers,
7772 2 : vec![
7773 2 : // Image layer at GC horizon
7774 2 : PersistentLayerKey {
7775 2 : key_range: Key::MIN..Key::MAX,
7776 2 : lsn_range: Lsn(0x30)..Lsn(0x31),
7777 2 : is_delta: false
7778 2 : },
7779 2 : // The delta layer below the horizon
7780 2 : PersistentLayerKey {
7781 2 : key_range: get_key(3)..get_key(4),
7782 2 : lsn_range: Lsn(0x30)..Lsn(0x48),
7783 2 : is_delta: true
7784 2 : },
7785 2 : // The delta3 layer that should not be picked for the compaction
7786 2 : PersistentLayerKey {
7787 2 : key_range: get_key(8)..get_key(10),
7788 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
7789 2 : is_delta: true
7790 2 : }
7791 2 : ]
7792 2 : );
7793 2 :
7794 2 : // increase GC horizon and compact again
7795 2 : {
7796 2 : // Update GC info
7797 2 : let mut guard = tline.gc_info.write().unwrap();
7798 2 : guard.cutoffs.time = Lsn(0x40);
7799 2 : guard.cutoffs.space = Lsn(0x40);
7800 2 : }
7801 2 : tline
7802 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7803 43 : .await
7804 2 : .unwrap();
7805 2 :
7806 2 : Ok(())
7807 2 : }
7808 :
7809 : #[cfg(feature = "testing")]
7810 : #[tokio::test]
7811 2 : async fn test_neon_test_record() -> anyhow::Result<()> {
7812 2 : let harness = TenantHarness::create("test_neon_test_record").await?;
7813 20 : let (tenant, ctx) = harness.load().await;
7814 2 :
7815 24 : fn get_key(id: u32) -> Key {
7816 24 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7817 24 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7818 24 : key.field6 = id;
7819 24 : key
7820 24 : }
7821 2 :
7822 2 : let delta1 = vec![
7823 2 : (
7824 2 : get_key(1),
7825 2 : Lsn(0x20),
7826 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
7827 2 : ),
7828 2 : (
7829 2 : get_key(1),
7830 2 : Lsn(0x30),
7831 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
7832 2 : ),
7833 2 : (get_key(2), Lsn(0x10), Value::Image("0x10".into())),
7834 2 : (
7835 2 : get_key(2),
7836 2 : Lsn(0x20),
7837 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
7838 2 : ),
7839 2 : (
7840 2 : get_key(2),
7841 2 : Lsn(0x30),
7842 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
7843 2 : ),
7844 2 : (get_key(3), Lsn(0x10), Value::Image("0x10".into())),
7845 2 : (
7846 2 : get_key(3),
7847 2 : Lsn(0x20),
7848 2 : Value::WalRecord(NeonWalRecord::wal_clear("c")),
7849 2 : ),
7850 2 : (get_key(4), Lsn(0x10), Value::Image("0x10".into())),
7851 2 : (
7852 2 : get_key(4),
7853 2 : Lsn(0x20),
7854 2 : Value::WalRecord(NeonWalRecord::wal_init("i")),
7855 2 : ),
7856 2 : ];
7857 2 : let image1 = vec![(get_key(1), "0x10".into())];
7858 2 :
7859 2 : let tline = tenant
7860 2 : .create_test_timeline_with_layers(
7861 2 : TIMELINE_ID,
7862 2 : Lsn(0x10),
7863 2 : DEFAULT_PG_VERSION,
7864 2 : &ctx,
7865 2 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
7866 2 : Lsn(0x10)..Lsn(0x40),
7867 2 : delta1,
7868 2 : )], // delta layers
7869 2 : vec![(Lsn(0x10), image1)], // image layers
7870 2 : Lsn(0x50),
7871 2 : )
7872 19 : .await?;
7873 2 :
7874 2 : assert_eq!(
7875 8 : tline.get(get_key(1), Lsn(0x50), &ctx).await?,
7876 2 : Bytes::from_static(b"0x10,0x20,0x30")
7877 2 : );
7878 2 : assert_eq!(
7879 2 : tline.get(get_key(2), Lsn(0x50), &ctx).await?,
7880 2 : Bytes::from_static(b"0x10,0x20,0x30")
7881 2 : );
7882 2 :
7883 2 : // Need to remove the limit of "Neon WAL redo requires base image".
7884 2 :
7885 2 : // assert_eq!(tline.get(get_key(3), Lsn(0x50), &ctx).await?, Bytes::new());
7886 2 : // assert_eq!(tline.get(get_key(4), Lsn(0x50), &ctx).await?, Bytes::new());
7887 2 :
7888 2 : Ok(())
7889 2 : }
7890 :
7891 : #[tokio::test(start_paused = true)]
7892 2 : async fn test_lsn_lease() -> anyhow::Result<()> {
7893 2 : let (tenant, ctx) = TenantHarness::create("test_lsn_lease")
7894 2 : .await
7895 2 : .unwrap()
7896 2 : .load()
7897 20 : .await;
7898 2 : // Advance to the lsn lease deadline so that GC is not blocked by
7899 2 : // initial transition into AttachedSingle.
7900 2 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
7901 2 : tokio::time::resume();
7902 2 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7903 2 :
7904 2 : let end_lsn = Lsn(0x100);
7905 2 : let image_layers = (0x20..=0x90)
7906 2 : .step_by(0x10)
7907 16 : .map(|n| {
7908 16 : (
7909 16 : Lsn(n),
7910 16 : vec![(key, test_img(&format!("data key at {:x}", n)))],
7911 16 : )
7912 16 : })
7913 2 : .collect();
7914 2 :
7915 2 : let timeline = tenant
7916 2 : .create_test_timeline_with_layers(
7917 2 : TIMELINE_ID,
7918 2 : Lsn(0x10),
7919 2 : DEFAULT_PG_VERSION,
7920 2 : &ctx,
7921 2 : Vec::new(),
7922 2 : image_layers,
7923 2 : end_lsn,
7924 2 : )
7925 62 : .await?;
7926 2 :
7927 2 : let leased_lsns = [0x30, 0x50, 0x70];
7928 2 : let mut leases = Vec::new();
7929 6 : leased_lsns.iter().for_each(|n| {
7930 6 : leases.push(
7931 6 : timeline
7932 6 : .init_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)
7933 6 : .expect("lease request should succeed"),
7934 6 : );
7935 6 : });
7936 2 :
7937 2 : let updated_lease_0 = timeline
7938 2 : .renew_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)
7939 2 : .expect("lease renewal should succeed");
7940 2 : assert_eq!(
7941 2 : updated_lease_0.valid_until, leases[0].valid_until,
7942 2 : " Renewing with shorter lease should not change the lease."
7943 2 : );
7944 2 :
7945 2 : let updated_lease_1 = timeline
7946 2 : .renew_lsn_lease(
7947 2 : Lsn(leased_lsns[1]),
7948 2 : timeline.get_lsn_lease_length() * 2,
7949 2 : &ctx,
7950 2 : )
7951 2 : .expect("lease renewal should succeed");
7952 2 : assert!(
7953 2 : updated_lease_1.valid_until > leases[1].valid_until,
7954 2 : "Renewing with a long lease should renew lease with later expiration time."
7955 2 : );
7956 2 :
7957 2 : // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
7958 2 : info!(
7959 2 : "latest_gc_cutoff_lsn: {}",
7960 0 : *timeline.get_latest_gc_cutoff_lsn()
7961 2 : );
7962 2 : timeline.force_set_disk_consistent_lsn(end_lsn);
7963 2 :
7964 2 : let res = tenant
7965 2 : .gc_iteration(
7966 2 : Some(TIMELINE_ID),
7967 2 : 0,
7968 2 : Duration::ZERO,
7969 2 : &CancellationToken::new(),
7970 2 : &ctx,
7971 2 : )
7972 2 : .await
7973 2 : .unwrap();
7974 2 :
7975 2 : // Keeping everything <= Lsn(0x80) b/c leases:
7976 2 : // 0/10: initdb layer
7977 2 : // (0/20..=0/70).step_by(0x10): image layers added when creating the timeline.
7978 2 : assert_eq!(res.layers_needed_by_leases, 7);
7979 2 : // Keeping 0/90 b/c it is the latest layer.
7980 2 : assert_eq!(res.layers_not_updated, 1);
7981 2 : // Removed 0/80.
7982 2 : assert_eq!(res.layers_removed, 1);
7983 2 :
7984 2 : // Make lease on a already GC-ed LSN.
7985 2 : // 0/80 does not have a valid lease + is below latest_gc_cutoff
7986 2 : assert!(Lsn(0x80) < *timeline.get_latest_gc_cutoff_lsn());
7987 2 : timeline
7988 2 : .init_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx)
7989 2 : .expect_err("lease request on GC-ed LSN should fail");
7990 2 :
7991 2 : // Should still be able to renew a currently valid lease
7992 2 : // Assumption: original lease to is still valid for 0/50.
7993 2 : // (use `Timeline::init_lsn_lease` for testing so it always does validation)
7994 2 : timeline
7995 2 : .init_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)
7996 2 : .expect("lease renewal with validation should succeed");
7997 2 :
7998 2 : Ok(())
7999 2 : }
8000 :
8001 : #[cfg(feature = "testing")]
8002 : #[tokio::test]
8003 2 : async fn test_simple_bottom_most_compaction_deltas_1() -> anyhow::Result<()> {
8004 2 : test_simple_bottom_most_compaction_deltas_helper(
8005 2 : "test_simple_bottom_most_compaction_deltas_1",
8006 2 : false,
8007 2 : )
8008 240 : .await
8009 2 : }
8010 :
8011 : #[cfg(feature = "testing")]
8012 : #[tokio::test]
8013 2 : async fn test_simple_bottom_most_compaction_deltas_2() -> anyhow::Result<()> {
8014 2 : test_simple_bottom_most_compaction_deltas_helper(
8015 2 : "test_simple_bottom_most_compaction_deltas_2",
8016 2 : true,
8017 2 : )
8018 225 : .await
8019 2 : }
8020 :
8021 : #[cfg(feature = "testing")]
8022 4 : async fn test_simple_bottom_most_compaction_deltas_helper(
8023 4 : test_name: &'static str,
8024 4 : use_delta_bottom_layer: bool,
8025 4 : ) -> anyhow::Result<()> {
8026 4 : let harness = TenantHarness::create(test_name).await?;
8027 40 : let (tenant, ctx) = harness.load().await;
8028 :
8029 276 : fn get_key(id: u32) -> Key {
8030 276 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8031 276 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8032 276 : key.field6 = id;
8033 276 : key
8034 276 : }
8035 :
8036 : // We create
8037 : // - one bottom-most image layer,
8038 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
8039 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
8040 : // - a delta layer D3 above the horizon.
8041 : //
8042 : // | D3 |
8043 : // | D1 |
8044 : // -| |-- gc horizon -----------------
8045 : // | | | D2 |
8046 : // --------- img layer ------------------
8047 : //
8048 : // What we should expact from this compaction is:
8049 : // | D3 |
8050 : // | Part of D1 |
8051 : // --------- img layer with D1+D2 at GC horizon------------------
8052 :
8053 : // img layer at 0x10
8054 4 : let img_layer = (0..10)
8055 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8056 4 : .collect_vec();
8057 4 : // or, delta layer at 0x10 if `use_delta_bottom_layer` is true
8058 4 : let delta4 = (0..10)
8059 40 : .map(|id| {
8060 40 : (
8061 40 : get_key(id),
8062 40 : Lsn(0x08),
8063 40 : Value::WalRecord(NeonWalRecord::wal_init(format!("value {id}@0x10"))),
8064 40 : )
8065 40 : })
8066 4 : .collect_vec();
8067 4 :
8068 4 : let delta1 = vec![
8069 4 : (
8070 4 : get_key(1),
8071 4 : Lsn(0x20),
8072 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8073 4 : ),
8074 4 : (
8075 4 : get_key(2),
8076 4 : Lsn(0x30),
8077 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8078 4 : ),
8079 4 : (
8080 4 : get_key(3),
8081 4 : Lsn(0x28),
8082 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8083 4 : ),
8084 4 : (
8085 4 : get_key(3),
8086 4 : Lsn(0x30),
8087 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8088 4 : ),
8089 4 : (
8090 4 : get_key(3),
8091 4 : Lsn(0x40),
8092 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
8093 4 : ),
8094 4 : ];
8095 4 : let delta2 = vec![
8096 4 : (
8097 4 : get_key(5),
8098 4 : Lsn(0x20),
8099 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8100 4 : ),
8101 4 : (
8102 4 : get_key(6),
8103 4 : Lsn(0x20),
8104 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8105 4 : ),
8106 4 : ];
8107 4 : let delta3 = vec![
8108 4 : (
8109 4 : get_key(8),
8110 4 : Lsn(0x48),
8111 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8112 4 : ),
8113 4 : (
8114 4 : get_key(9),
8115 4 : Lsn(0x48),
8116 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8117 4 : ),
8118 4 : ];
8119 :
8120 4 : let tline = if use_delta_bottom_layer {
8121 2 : tenant
8122 2 : .create_test_timeline_with_layers(
8123 2 : TIMELINE_ID,
8124 2 : Lsn(0x08),
8125 2 : DEFAULT_PG_VERSION,
8126 2 : &ctx,
8127 2 : vec![
8128 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8129 2 : Lsn(0x08)..Lsn(0x10),
8130 2 : delta4,
8131 2 : ),
8132 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8133 2 : Lsn(0x20)..Lsn(0x48),
8134 2 : delta1,
8135 2 : ),
8136 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8137 2 : Lsn(0x20)..Lsn(0x48),
8138 2 : delta2,
8139 2 : ),
8140 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8141 2 : Lsn(0x48)..Lsn(0x50),
8142 2 : delta3,
8143 2 : ),
8144 2 : ], // delta layers
8145 2 : vec![], // image layers
8146 2 : Lsn(0x50),
8147 2 : )
8148 30 : .await?
8149 : } else {
8150 2 : tenant
8151 2 : .create_test_timeline_with_layers(
8152 2 : TIMELINE_ID,
8153 2 : Lsn(0x10),
8154 2 : DEFAULT_PG_VERSION,
8155 2 : &ctx,
8156 2 : vec![
8157 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8158 2 : Lsn(0x10)..Lsn(0x48),
8159 2 : delta1,
8160 2 : ),
8161 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8162 2 : Lsn(0x10)..Lsn(0x48),
8163 2 : delta2,
8164 2 : ),
8165 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8166 2 : Lsn(0x48)..Lsn(0x50),
8167 2 : delta3,
8168 2 : ),
8169 2 : ], // delta layers
8170 2 : vec![(Lsn(0x10), img_layer)], // image layers
8171 2 : Lsn(0x50),
8172 2 : )
8173 49 : .await?
8174 : };
8175 4 : {
8176 4 : // Update GC info
8177 4 : let mut guard = tline.gc_info.write().unwrap();
8178 4 : *guard = GcInfo {
8179 4 : retain_lsns: vec![],
8180 4 : cutoffs: GcCutoffs {
8181 4 : time: Lsn(0x30),
8182 4 : space: Lsn(0x30),
8183 4 : },
8184 4 : leases: Default::default(),
8185 4 : within_ancestor_pitr: false,
8186 4 : };
8187 4 : }
8188 4 :
8189 4 : let expected_result = [
8190 4 : Bytes::from_static(b"value 0@0x10"),
8191 4 : Bytes::from_static(b"value 1@0x10@0x20"),
8192 4 : Bytes::from_static(b"value 2@0x10@0x30"),
8193 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
8194 4 : Bytes::from_static(b"value 4@0x10"),
8195 4 : Bytes::from_static(b"value 5@0x10@0x20"),
8196 4 : Bytes::from_static(b"value 6@0x10@0x20"),
8197 4 : Bytes::from_static(b"value 7@0x10"),
8198 4 : Bytes::from_static(b"value 8@0x10@0x48"),
8199 4 : Bytes::from_static(b"value 9@0x10@0x48"),
8200 4 : ];
8201 4 :
8202 4 : let expected_result_at_gc_horizon = [
8203 4 : Bytes::from_static(b"value 0@0x10"),
8204 4 : Bytes::from_static(b"value 1@0x10@0x20"),
8205 4 : Bytes::from_static(b"value 2@0x10@0x30"),
8206 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
8207 4 : Bytes::from_static(b"value 4@0x10"),
8208 4 : Bytes::from_static(b"value 5@0x10@0x20"),
8209 4 : Bytes::from_static(b"value 6@0x10@0x20"),
8210 4 : Bytes::from_static(b"value 7@0x10"),
8211 4 : Bytes::from_static(b"value 8@0x10"),
8212 4 : Bytes::from_static(b"value 9@0x10"),
8213 4 : ];
8214 :
8215 44 : for idx in 0..10 {
8216 40 : assert_eq!(
8217 40 : tline
8218 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8219 60 : .await
8220 40 : .unwrap(),
8221 40 : &expected_result[idx]
8222 : );
8223 40 : assert_eq!(
8224 40 : tline
8225 40 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
8226 31 : .await
8227 40 : .unwrap(),
8228 40 : &expected_result_at_gc_horizon[idx]
8229 : );
8230 : }
8231 :
8232 4 : let cancel = CancellationToken::new();
8233 4 : tline
8234 4 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8235 108 : .await
8236 4 : .unwrap();
8237 :
8238 44 : for idx in 0..10 {
8239 40 : assert_eq!(
8240 40 : tline
8241 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8242 40 : .await
8243 40 : .unwrap(),
8244 40 : &expected_result[idx]
8245 : );
8246 40 : assert_eq!(
8247 40 : tline
8248 40 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
8249 20 : .await
8250 40 : .unwrap(),
8251 40 : &expected_result_at_gc_horizon[idx]
8252 : );
8253 : }
8254 :
8255 : // increase GC horizon and compact again
8256 4 : {
8257 4 : // Update GC info
8258 4 : let mut guard = tline.gc_info.write().unwrap();
8259 4 : guard.cutoffs.time = Lsn(0x40);
8260 4 : guard.cutoffs.space = Lsn(0x40);
8261 4 : }
8262 4 : tline
8263 4 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8264 87 : .await
8265 4 : .unwrap();
8266 4 :
8267 4 : Ok(())
8268 4 : }
8269 :
8270 : #[cfg(feature = "testing")]
8271 : #[tokio::test]
8272 2 : async fn test_generate_key_retention() -> anyhow::Result<()> {
8273 2 : let harness = TenantHarness::create("test_generate_key_retention").await?;
8274 20 : let (tenant, ctx) = harness.load().await;
8275 2 : let tline = tenant
8276 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
8277 6 : .await?;
8278 2 : tline.force_advance_lsn(Lsn(0x70));
8279 2 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
8280 2 : let history = vec![
8281 2 : (
8282 2 : key,
8283 2 : Lsn(0x10),
8284 2 : Value::WalRecord(NeonWalRecord::wal_init("0x10")),
8285 2 : ),
8286 2 : (
8287 2 : key,
8288 2 : Lsn(0x20),
8289 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8290 2 : ),
8291 2 : (
8292 2 : key,
8293 2 : Lsn(0x30),
8294 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8295 2 : ),
8296 2 : (
8297 2 : key,
8298 2 : Lsn(0x40),
8299 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8300 2 : ),
8301 2 : (
8302 2 : key,
8303 2 : Lsn(0x50),
8304 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
8305 2 : ),
8306 2 : (
8307 2 : key,
8308 2 : Lsn(0x60),
8309 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8310 2 : ),
8311 2 : (
8312 2 : key,
8313 2 : Lsn(0x70),
8314 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8315 2 : ),
8316 2 : (
8317 2 : key,
8318 2 : Lsn(0x80),
8319 2 : Value::Image(Bytes::copy_from_slice(
8320 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8321 2 : )),
8322 2 : ),
8323 2 : (
8324 2 : key,
8325 2 : Lsn(0x90),
8326 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8327 2 : ),
8328 2 : ];
8329 2 : let res = tline
8330 2 : .generate_key_retention(
8331 2 : key,
8332 2 : &history,
8333 2 : Lsn(0x60),
8334 2 : &[Lsn(0x20), Lsn(0x40), Lsn(0x50)],
8335 2 : 3,
8336 2 : None,
8337 2 : )
8338 2 : .await
8339 2 : .unwrap();
8340 2 : let expected_res = KeyHistoryRetention {
8341 2 : below_horizon: vec![
8342 2 : (
8343 2 : Lsn(0x20),
8344 2 : KeyLogAtLsn(vec![(
8345 2 : Lsn(0x20),
8346 2 : Value::Image(Bytes::from_static(b"0x10;0x20")),
8347 2 : )]),
8348 2 : ),
8349 2 : (
8350 2 : Lsn(0x40),
8351 2 : KeyLogAtLsn(vec![
8352 2 : (
8353 2 : Lsn(0x30),
8354 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8355 2 : ),
8356 2 : (
8357 2 : Lsn(0x40),
8358 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8359 2 : ),
8360 2 : ]),
8361 2 : ),
8362 2 : (
8363 2 : Lsn(0x50),
8364 2 : KeyLogAtLsn(vec![(
8365 2 : Lsn(0x50),
8366 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40;0x50")),
8367 2 : )]),
8368 2 : ),
8369 2 : (
8370 2 : Lsn(0x60),
8371 2 : KeyLogAtLsn(vec![(
8372 2 : Lsn(0x60),
8373 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8374 2 : )]),
8375 2 : ),
8376 2 : ],
8377 2 : above_horizon: KeyLogAtLsn(vec![
8378 2 : (
8379 2 : Lsn(0x70),
8380 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8381 2 : ),
8382 2 : (
8383 2 : Lsn(0x80),
8384 2 : Value::Image(Bytes::copy_from_slice(
8385 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8386 2 : )),
8387 2 : ),
8388 2 : (
8389 2 : Lsn(0x90),
8390 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8391 2 : ),
8392 2 : ]),
8393 2 : };
8394 2 : assert_eq!(res, expected_res);
8395 2 :
8396 2 : // We expect GC-compaction to run with the original GC. This would create a situation that
8397 2 : // the original GC algorithm removes some delta layers b/c there are full image coverage,
8398 2 : // therefore causing some keys to have an incomplete history below the lowest retain LSN.
8399 2 : // For example, we have
8400 2 : // ```plain
8401 2 : // init delta @ 0x10, image @ 0x20, delta @ 0x30 (gc_horizon), image @ 0x40.
8402 2 : // ```
8403 2 : // Now the GC horizon moves up, and we have
8404 2 : // ```plain
8405 2 : // init delta @ 0x10, image @ 0x20, delta @ 0x30, image @ 0x40 (gc_horizon)
8406 2 : // ```
8407 2 : // The original GC algorithm kicks in, and removes delta @ 0x10, image @ 0x20.
8408 2 : // We will end up with
8409 2 : // ```plain
8410 2 : // delta @ 0x30, image @ 0x40 (gc_horizon)
8411 2 : // ```
8412 2 : // Now we run the GC-compaction, and this key does not have a full history.
8413 2 : // We should be able to handle this partial history and drop everything before the
8414 2 : // gc_horizon image.
8415 2 :
8416 2 : let history = vec![
8417 2 : (
8418 2 : key,
8419 2 : Lsn(0x20),
8420 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8421 2 : ),
8422 2 : (
8423 2 : key,
8424 2 : Lsn(0x30),
8425 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8426 2 : ),
8427 2 : (
8428 2 : key,
8429 2 : Lsn(0x40),
8430 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
8431 2 : ),
8432 2 : (
8433 2 : key,
8434 2 : Lsn(0x50),
8435 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
8436 2 : ),
8437 2 : (
8438 2 : key,
8439 2 : Lsn(0x60),
8440 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8441 2 : ),
8442 2 : (
8443 2 : key,
8444 2 : Lsn(0x70),
8445 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8446 2 : ),
8447 2 : (
8448 2 : key,
8449 2 : Lsn(0x80),
8450 2 : Value::Image(Bytes::copy_from_slice(
8451 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8452 2 : )),
8453 2 : ),
8454 2 : (
8455 2 : key,
8456 2 : Lsn(0x90),
8457 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8458 2 : ),
8459 2 : ];
8460 2 : let res = tline
8461 2 : .generate_key_retention(key, &history, Lsn(0x60), &[Lsn(0x40), Lsn(0x50)], 3, None)
8462 2 : .await
8463 2 : .unwrap();
8464 2 : let expected_res = KeyHistoryRetention {
8465 2 : below_horizon: vec![
8466 2 : (
8467 2 : Lsn(0x40),
8468 2 : KeyLogAtLsn(vec![(
8469 2 : Lsn(0x40),
8470 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
8471 2 : )]),
8472 2 : ),
8473 2 : (
8474 2 : Lsn(0x50),
8475 2 : KeyLogAtLsn(vec![(
8476 2 : Lsn(0x50),
8477 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
8478 2 : )]),
8479 2 : ),
8480 2 : (
8481 2 : Lsn(0x60),
8482 2 : KeyLogAtLsn(vec![(
8483 2 : Lsn(0x60),
8484 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8485 2 : )]),
8486 2 : ),
8487 2 : ],
8488 2 : above_horizon: KeyLogAtLsn(vec![
8489 2 : (
8490 2 : Lsn(0x70),
8491 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8492 2 : ),
8493 2 : (
8494 2 : Lsn(0x80),
8495 2 : Value::Image(Bytes::copy_from_slice(
8496 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8497 2 : )),
8498 2 : ),
8499 2 : (
8500 2 : Lsn(0x90),
8501 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8502 2 : ),
8503 2 : ]),
8504 2 : };
8505 2 : assert_eq!(res, expected_res);
8506 2 :
8507 2 : // In case of branch compaction, the branch itself does not have the full history, and we need to provide
8508 2 : // the ancestor image in the test case.
8509 2 :
8510 2 : let history = vec![
8511 2 : (
8512 2 : key,
8513 2 : Lsn(0x20),
8514 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8515 2 : ),
8516 2 : (
8517 2 : key,
8518 2 : Lsn(0x30),
8519 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8520 2 : ),
8521 2 : (
8522 2 : key,
8523 2 : Lsn(0x40),
8524 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8525 2 : ),
8526 2 : (
8527 2 : key,
8528 2 : Lsn(0x70),
8529 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8530 2 : ),
8531 2 : ];
8532 2 : let res = tline
8533 2 : .generate_key_retention(
8534 2 : key,
8535 2 : &history,
8536 2 : Lsn(0x60),
8537 2 : &[],
8538 2 : 3,
8539 2 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
8540 2 : )
8541 2 : .await
8542 2 : .unwrap();
8543 2 : let expected_res = KeyHistoryRetention {
8544 2 : below_horizon: vec![(
8545 2 : Lsn(0x60),
8546 2 : KeyLogAtLsn(vec![(
8547 2 : Lsn(0x60),
8548 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")), // use the ancestor image to reconstruct the page
8549 2 : )]),
8550 2 : )],
8551 2 : above_horizon: KeyLogAtLsn(vec![(
8552 2 : Lsn(0x70),
8553 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8554 2 : )]),
8555 2 : };
8556 2 : assert_eq!(res, expected_res);
8557 2 :
8558 2 : let history = vec![
8559 2 : (
8560 2 : key,
8561 2 : Lsn(0x20),
8562 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8563 2 : ),
8564 2 : (
8565 2 : key,
8566 2 : Lsn(0x40),
8567 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8568 2 : ),
8569 2 : (
8570 2 : key,
8571 2 : Lsn(0x60),
8572 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8573 2 : ),
8574 2 : (
8575 2 : key,
8576 2 : Lsn(0x70),
8577 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8578 2 : ),
8579 2 : ];
8580 2 : let res = tline
8581 2 : .generate_key_retention(
8582 2 : key,
8583 2 : &history,
8584 2 : Lsn(0x60),
8585 2 : &[Lsn(0x30)],
8586 2 : 3,
8587 2 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
8588 2 : )
8589 2 : .await
8590 2 : .unwrap();
8591 2 : let expected_res = KeyHistoryRetention {
8592 2 : below_horizon: vec![
8593 2 : (
8594 2 : Lsn(0x30),
8595 2 : KeyLogAtLsn(vec![(
8596 2 : Lsn(0x20),
8597 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8598 2 : )]),
8599 2 : ),
8600 2 : (
8601 2 : Lsn(0x60),
8602 2 : KeyLogAtLsn(vec![(
8603 2 : Lsn(0x60),
8604 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x40;0x60")),
8605 2 : )]),
8606 2 : ),
8607 2 : ],
8608 2 : above_horizon: KeyLogAtLsn(vec![(
8609 2 : Lsn(0x70),
8610 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8611 2 : )]),
8612 2 : };
8613 2 : assert_eq!(res, expected_res);
8614 2 :
8615 2 : Ok(())
8616 2 : }
8617 :
8618 : #[cfg(feature = "testing")]
8619 : #[tokio::test]
8620 2 : async fn test_simple_bottom_most_compaction_with_retain_lsns() -> anyhow::Result<()> {
8621 2 : let harness =
8622 2 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns").await?;
8623 19 : let (tenant, ctx) = harness.load().await;
8624 2 :
8625 518 : fn get_key(id: u32) -> Key {
8626 518 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8627 518 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8628 518 : key.field6 = id;
8629 518 : key
8630 518 : }
8631 2 :
8632 2 : let img_layer = (0..10)
8633 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8634 2 : .collect_vec();
8635 2 :
8636 2 : let delta1 = vec![
8637 2 : (
8638 2 : get_key(1),
8639 2 : Lsn(0x20),
8640 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8641 2 : ),
8642 2 : (
8643 2 : get_key(2),
8644 2 : Lsn(0x30),
8645 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8646 2 : ),
8647 2 : (
8648 2 : get_key(3),
8649 2 : Lsn(0x28),
8650 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8651 2 : ),
8652 2 : (
8653 2 : get_key(3),
8654 2 : Lsn(0x30),
8655 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8656 2 : ),
8657 2 : (
8658 2 : get_key(3),
8659 2 : Lsn(0x40),
8660 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
8661 2 : ),
8662 2 : ];
8663 2 : let delta2 = vec![
8664 2 : (
8665 2 : get_key(5),
8666 2 : Lsn(0x20),
8667 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8668 2 : ),
8669 2 : (
8670 2 : get_key(6),
8671 2 : Lsn(0x20),
8672 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8673 2 : ),
8674 2 : ];
8675 2 : let delta3 = vec![
8676 2 : (
8677 2 : get_key(8),
8678 2 : Lsn(0x48),
8679 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8680 2 : ),
8681 2 : (
8682 2 : get_key(9),
8683 2 : Lsn(0x48),
8684 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8685 2 : ),
8686 2 : ];
8687 2 :
8688 2 : let tline = tenant
8689 2 : .create_test_timeline_with_layers(
8690 2 : TIMELINE_ID,
8691 2 : Lsn(0x10),
8692 2 : DEFAULT_PG_VERSION,
8693 2 : &ctx,
8694 2 : vec![
8695 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
8696 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
8697 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
8698 2 : ], // delta layers
8699 2 : vec![(Lsn(0x10), img_layer)], // image layers
8700 2 : Lsn(0x50),
8701 2 : )
8702 49 : .await?;
8703 2 : {
8704 2 : // Update GC info
8705 2 : let mut guard = tline.gc_info.write().unwrap();
8706 2 : *guard = GcInfo {
8707 2 : retain_lsns: vec![
8708 2 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
8709 2 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
8710 2 : ],
8711 2 : cutoffs: GcCutoffs {
8712 2 : time: Lsn(0x30),
8713 2 : space: Lsn(0x30),
8714 2 : },
8715 2 : leases: Default::default(),
8716 2 : within_ancestor_pitr: false,
8717 2 : };
8718 2 : }
8719 2 :
8720 2 : let expected_result = [
8721 2 : Bytes::from_static(b"value 0@0x10"),
8722 2 : Bytes::from_static(b"value 1@0x10@0x20"),
8723 2 : Bytes::from_static(b"value 2@0x10@0x30"),
8724 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
8725 2 : Bytes::from_static(b"value 4@0x10"),
8726 2 : Bytes::from_static(b"value 5@0x10@0x20"),
8727 2 : Bytes::from_static(b"value 6@0x10@0x20"),
8728 2 : Bytes::from_static(b"value 7@0x10"),
8729 2 : Bytes::from_static(b"value 8@0x10@0x48"),
8730 2 : Bytes::from_static(b"value 9@0x10@0x48"),
8731 2 : ];
8732 2 :
8733 2 : let expected_result_at_gc_horizon = [
8734 2 : Bytes::from_static(b"value 0@0x10"),
8735 2 : Bytes::from_static(b"value 1@0x10@0x20"),
8736 2 : Bytes::from_static(b"value 2@0x10@0x30"),
8737 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
8738 2 : Bytes::from_static(b"value 4@0x10"),
8739 2 : Bytes::from_static(b"value 5@0x10@0x20"),
8740 2 : Bytes::from_static(b"value 6@0x10@0x20"),
8741 2 : Bytes::from_static(b"value 7@0x10"),
8742 2 : Bytes::from_static(b"value 8@0x10"),
8743 2 : Bytes::from_static(b"value 9@0x10"),
8744 2 : ];
8745 2 :
8746 2 : let expected_result_at_lsn_20 = [
8747 2 : Bytes::from_static(b"value 0@0x10"),
8748 2 : Bytes::from_static(b"value 1@0x10@0x20"),
8749 2 : Bytes::from_static(b"value 2@0x10"),
8750 2 : Bytes::from_static(b"value 3@0x10"),
8751 2 : Bytes::from_static(b"value 4@0x10"),
8752 2 : Bytes::from_static(b"value 5@0x10@0x20"),
8753 2 : Bytes::from_static(b"value 6@0x10@0x20"),
8754 2 : Bytes::from_static(b"value 7@0x10"),
8755 2 : Bytes::from_static(b"value 8@0x10"),
8756 2 : Bytes::from_static(b"value 9@0x10"),
8757 2 : ];
8758 2 :
8759 2 : let expected_result_at_lsn_10 = [
8760 2 : Bytes::from_static(b"value 0@0x10"),
8761 2 : Bytes::from_static(b"value 1@0x10"),
8762 2 : Bytes::from_static(b"value 2@0x10"),
8763 2 : Bytes::from_static(b"value 3@0x10"),
8764 2 : Bytes::from_static(b"value 4@0x10"),
8765 2 : Bytes::from_static(b"value 5@0x10"),
8766 2 : Bytes::from_static(b"value 6@0x10"),
8767 2 : Bytes::from_static(b"value 7@0x10"),
8768 2 : Bytes::from_static(b"value 8@0x10"),
8769 2 : Bytes::from_static(b"value 9@0x10"),
8770 2 : ];
8771 2 :
8772 12 : let verify_result = || async {
8773 12 : let gc_horizon = {
8774 12 : let gc_info = tline.gc_info.read().unwrap();
8775 12 : gc_info.cutoffs.time
8776 2 : };
8777 132 : for idx in 0..10 {
8778 120 : assert_eq!(
8779 120 : tline
8780 120 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8781 121 : .await
8782 120 : .unwrap(),
8783 120 : &expected_result[idx]
8784 2 : );
8785 120 : assert_eq!(
8786 120 : tline
8787 120 : .get(get_key(idx as u32), gc_horizon, &ctx)
8788 93 : .await
8789 120 : .unwrap(),
8790 120 : &expected_result_at_gc_horizon[idx]
8791 2 : );
8792 120 : assert_eq!(
8793 120 : tline
8794 120 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
8795 86 : .await
8796 120 : .unwrap(),
8797 120 : &expected_result_at_lsn_20[idx]
8798 2 : );
8799 120 : assert_eq!(
8800 120 : tline
8801 120 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
8802 61 : .await
8803 120 : .unwrap(),
8804 120 : &expected_result_at_lsn_10[idx]
8805 2 : );
8806 2 : }
8807 24 : };
8808 2 :
8809 69 : verify_result().await;
8810 2 :
8811 2 : let cancel = CancellationToken::new();
8812 2 : let mut dryrun_flags = EnumSet::new();
8813 2 : dryrun_flags.insert(CompactFlags::DryRun);
8814 2 :
8815 2 : tline
8816 2 : .compact_with_gc(&cancel, dryrun_flags, &ctx)
8817 45 : .await
8818 2 : .unwrap();
8819 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
8820 2 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
8821 57 : verify_result().await;
8822 2 :
8823 2 : tline
8824 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8825 50 : .await
8826 2 : .unwrap();
8827 64 : verify_result().await;
8828 2 :
8829 2 : // compact again
8830 2 : tline
8831 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8832 40 : .await
8833 2 : .unwrap();
8834 57 : verify_result().await;
8835 2 :
8836 2 : // increase GC horizon and compact again
8837 2 : {
8838 2 : // Update GC info
8839 2 : let mut guard = tline.gc_info.write().unwrap();
8840 2 : guard.cutoffs.time = Lsn(0x38);
8841 2 : guard.cutoffs.space = Lsn(0x38);
8842 2 : }
8843 2 : tline
8844 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8845 39 : .await
8846 2 : .unwrap();
8847 57 : verify_result().await; // no wals between 0x30 and 0x38, so we should obtain the same result
8848 2 :
8849 2 : // not increasing the GC horizon and compact again
8850 2 : tline
8851 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8852 40 : .await
8853 2 : .unwrap();
8854 57 : verify_result().await;
8855 2 :
8856 2 : Ok(())
8857 2 : }
8858 :
8859 : #[cfg(feature = "testing")]
8860 : #[tokio::test]
8861 2 : async fn test_simple_bottom_most_compaction_with_retain_lsns_single_key() -> anyhow::Result<()>
8862 2 : {
8863 2 : let harness =
8864 2 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns_single_key")
8865 2 : .await?;
8866 20 : let (tenant, ctx) = harness.load().await;
8867 2 :
8868 352 : fn get_key(id: u32) -> Key {
8869 352 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8870 352 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8871 352 : key.field6 = id;
8872 352 : key
8873 352 : }
8874 2 :
8875 2 : let img_layer = (0..10)
8876 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8877 2 : .collect_vec();
8878 2 :
8879 2 : let delta1 = vec![
8880 2 : (
8881 2 : get_key(1),
8882 2 : Lsn(0x20),
8883 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8884 2 : ),
8885 2 : (
8886 2 : get_key(1),
8887 2 : Lsn(0x28),
8888 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8889 2 : ),
8890 2 : ];
8891 2 : let delta2 = vec![
8892 2 : (
8893 2 : get_key(1),
8894 2 : Lsn(0x30),
8895 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8896 2 : ),
8897 2 : (
8898 2 : get_key(1),
8899 2 : Lsn(0x38),
8900 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
8901 2 : ),
8902 2 : ];
8903 2 : let delta3 = vec![
8904 2 : (
8905 2 : get_key(8),
8906 2 : Lsn(0x48),
8907 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8908 2 : ),
8909 2 : (
8910 2 : get_key(9),
8911 2 : Lsn(0x48),
8912 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8913 2 : ),
8914 2 : ];
8915 2 :
8916 2 : let tline = tenant
8917 2 : .create_test_timeline_with_layers(
8918 2 : TIMELINE_ID,
8919 2 : Lsn(0x10),
8920 2 : DEFAULT_PG_VERSION,
8921 2 : &ctx,
8922 2 : vec![
8923 2 : // delta1 and delta 2 only contain a single key but multiple updates
8924 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x30), delta1),
8925 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
8926 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x50), delta3),
8927 2 : ], // delta layers
8928 2 : vec![(Lsn(0x10), img_layer)], // image layers
8929 2 : Lsn(0x50),
8930 2 : )
8931 49 : .await?;
8932 2 : {
8933 2 : // Update GC info
8934 2 : let mut guard = tline.gc_info.write().unwrap();
8935 2 : *guard = GcInfo {
8936 2 : retain_lsns: vec![
8937 2 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
8938 2 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
8939 2 : ],
8940 2 : cutoffs: GcCutoffs {
8941 2 : time: Lsn(0x30),
8942 2 : space: Lsn(0x30),
8943 2 : },
8944 2 : leases: Default::default(),
8945 2 : within_ancestor_pitr: false,
8946 2 : };
8947 2 : }
8948 2 :
8949 2 : let expected_result = [
8950 2 : Bytes::from_static(b"value 0@0x10"),
8951 2 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
8952 2 : Bytes::from_static(b"value 2@0x10"),
8953 2 : Bytes::from_static(b"value 3@0x10"),
8954 2 : Bytes::from_static(b"value 4@0x10"),
8955 2 : Bytes::from_static(b"value 5@0x10"),
8956 2 : Bytes::from_static(b"value 6@0x10"),
8957 2 : Bytes::from_static(b"value 7@0x10"),
8958 2 : Bytes::from_static(b"value 8@0x10@0x48"),
8959 2 : Bytes::from_static(b"value 9@0x10@0x48"),
8960 2 : ];
8961 2 :
8962 2 : let expected_result_at_gc_horizon = [
8963 2 : Bytes::from_static(b"value 0@0x10"),
8964 2 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
8965 2 : Bytes::from_static(b"value 2@0x10"),
8966 2 : Bytes::from_static(b"value 3@0x10"),
8967 2 : Bytes::from_static(b"value 4@0x10"),
8968 2 : Bytes::from_static(b"value 5@0x10"),
8969 2 : Bytes::from_static(b"value 6@0x10"),
8970 2 : Bytes::from_static(b"value 7@0x10"),
8971 2 : Bytes::from_static(b"value 8@0x10"),
8972 2 : Bytes::from_static(b"value 9@0x10"),
8973 2 : ];
8974 2 :
8975 2 : let expected_result_at_lsn_20 = [
8976 2 : Bytes::from_static(b"value 0@0x10"),
8977 2 : Bytes::from_static(b"value 1@0x10@0x20"),
8978 2 : Bytes::from_static(b"value 2@0x10"),
8979 2 : Bytes::from_static(b"value 3@0x10"),
8980 2 : Bytes::from_static(b"value 4@0x10"),
8981 2 : Bytes::from_static(b"value 5@0x10"),
8982 2 : Bytes::from_static(b"value 6@0x10"),
8983 2 : Bytes::from_static(b"value 7@0x10"),
8984 2 : Bytes::from_static(b"value 8@0x10"),
8985 2 : Bytes::from_static(b"value 9@0x10"),
8986 2 : ];
8987 2 :
8988 2 : let expected_result_at_lsn_10 = [
8989 2 : Bytes::from_static(b"value 0@0x10"),
8990 2 : Bytes::from_static(b"value 1@0x10"),
8991 2 : Bytes::from_static(b"value 2@0x10"),
8992 2 : Bytes::from_static(b"value 3@0x10"),
8993 2 : Bytes::from_static(b"value 4@0x10"),
8994 2 : Bytes::from_static(b"value 5@0x10"),
8995 2 : Bytes::from_static(b"value 6@0x10"),
8996 2 : Bytes::from_static(b"value 7@0x10"),
8997 2 : Bytes::from_static(b"value 8@0x10"),
8998 2 : Bytes::from_static(b"value 9@0x10"),
8999 2 : ];
9000 2 :
9001 8 : let verify_result = || async {
9002 8 : let gc_horizon = {
9003 8 : let gc_info = tline.gc_info.read().unwrap();
9004 8 : gc_info.cutoffs.time
9005 2 : };
9006 88 : for idx in 0..10 {
9007 80 : assert_eq!(
9008 80 : tline
9009 80 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9010 76 : .await
9011 80 : .unwrap(),
9012 80 : &expected_result[idx]
9013 2 : );
9014 80 : assert_eq!(
9015 80 : tline
9016 80 : .get(get_key(idx as u32), gc_horizon, &ctx)
9017 46 : .await
9018 80 : .unwrap(),
9019 80 : &expected_result_at_gc_horizon[idx]
9020 2 : );
9021 80 : assert_eq!(
9022 80 : tline
9023 80 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
9024 47 : .await
9025 80 : .unwrap(),
9026 80 : &expected_result_at_lsn_20[idx]
9027 2 : );
9028 80 : assert_eq!(
9029 80 : tline
9030 80 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
9031 41 : .await
9032 80 : .unwrap(),
9033 80 : &expected_result_at_lsn_10[idx]
9034 2 : );
9035 2 : }
9036 16 : };
9037 2 :
9038 61 : verify_result().await;
9039 2 :
9040 2 : let cancel = CancellationToken::new();
9041 2 : let mut dryrun_flags = EnumSet::new();
9042 2 : dryrun_flags.insert(CompactFlags::DryRun);
9043 2 :
9044 2 : tline
9045 2 : .compact_with_gc(&cancel, dryrun_flags, &ctx)
9046 45 : .await
9047 2 : .unwrap();
9048 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
9049 2 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
9050 49 : verify_result().await;
9051 2 :
9052 2 : tline
9053 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
9054 51 : .await
9055 2 : .unwrap();
9056 53 : verify_result().await;
9057 2 :
9058 2 : // compact again
9059 2 : tline
9060 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
9061 40 : .await
9062 2 : .unwrap();
9063 47 : verify_result().await;
9064 2 :
9065 2 : Ok(())
9066 2 : }
9067 :
9068 : #[cfg(feature = "testing")]
9069 : #[tokio::test]
9070 2 : async fn test_simple_bottom_most_compaction_on_branch() -> anyhow::Result<()> {
9071 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_on_branch").await?;
9072 20 : let (tenant, ctx) = harness.load().await;
9073 2 :
9074 126 : fn get_key(id: u32) -> Key {
9075 126 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
9076 126 : key.field6 = id;
9077 126 : key
9078 126 : }
9079 2 :
9080 2 : let img_layer = (0..10)
9081 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9082 2 : .collect_vec();
9083 2 :
9084 2 : let delta1 = vec![
9085 2 : (
9086 2 : get_key(1),
9087 2 : Lsn(0x20),
9088 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9089 2 : ),
9090 2 : (
9091 2 : get_key(2),
9092 2 : Lsn(0x30),
9093 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9094 2 : ),
9095 2 : (
9096 2 : get_key(3),
9097 2 : Lsn(0x28),
9098 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9099 2 : ),
9100 2 : (
9101 2 : get_key(3),
9102 2 : Lsn(0x30),
9103 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9104 2 : ),
9105 2 : (
9106 2 : get_key(3),
9107 2 : Lsn(0x40),
9108 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9109 2 : ),
9110 2 : ];
9111 2 : let delta2 = vec![
9112 2 : (
9113 2 : get_key(5),
9114 2 : Lsn(0x20),
9115 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9116 2 : ),
9117 2 : (
9118 2 : get_key(6),
9119 2 : Lsn(0x20),
9120 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9121 2 : ),
9122 2 : ];
9123 2 : let delta3 = vec![
9124 2 : (
9125 2 : get_key(8),
9126 2 : Lsn(0x48),
9127 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9128 2 : ),
9129 2 : (
9130 2 : get_key(9),
9131 2 : Lsn(0x48),
9132 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9133 2 : ),
9134 2 : ];
9135 2 :
9136 2 : let parent_tline = tenant
9137 2 : .create_test_timeline_with_layers(
9138 2 : TIMELINE_ID,
9139 2 : Lsn(0x10),
9140 2 : DEFAULT_PG_VERSION,
9141 2 : &ctx,
9142 2 : vec![], // delta layers
9143 2 : vec![(Lsn(0x18), img_layer)], // image layers
9144 2 : Lsn(0x18),
9145 2 : )
9146 31 : .await?;
9147 2 :
9148 2 : parent_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
9149 2 :
9150 2 : let branch_tline = tenant
9151 2 : .branch_timeline_test_with_layers(
9152 2 : &parent_tline,
9153 2 : NEW_TIMELINE_ID,
9154 2 : Some(Lsn(0x18)),
9155 2 : &ctx,
9156 2 : vec![
9157 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
9158 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
9159 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
9160 2 : ], // delta layers
9161 2 : vec![], // image layers
9162 2 : Lsn(0x50),
9163 2 : )
9164 19 : .await?;
9165 2 :
9166 2 : branch_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
9167 2 :
9168 2 : {
9169 2 : // Update GC info
9170 2 : let mut guard = parent_tline.gc_info.write().unwrap();
9171 2 : *guard = GcInfo {
9172 2 : retain_lsns: vec![(Lsn(0x18), branch_tline.timeline_id, MaybeOffloaded::No)],
9173 2 : cutoffs: GcCutoffs {
9174 2 : time: Lsn(0x10),
9175 2 : space: Lsn(0x10),
9176 2 : },
9177 2 : leases: Default::default(),
9178 2 : within_ancestor_pitr: false,
9179 2 : };
9180 2 : }
9181 2 :
9182 2 : {
9183 2 : // Update GC info
9184 2 : let mut guard = branch_tline.gc_info.write().unwrap();
9185 2 : *guard = GcInfo {
9186 2 : retain_lsns: vec![(Lsn(0x40), branch_tline.timeline_id, MaybeOffloaded::No)],
9187 2 : cutoffs: GcCutoffs {
9188 2 : time: Lsn(0x50),
9189 2 : space: Lsn(0x50),
9190 2 : },
9191 2 : leases: Default::default(),
9192 2 : within_ancestor_pitr: false,
9193 2 : };
9194 2 : }
9195 2 :
9196 2 : let expected_result_at_gc_horizon = [
9197 2 : Bytes::from_static(b"value 0@0x10"),
9198 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9199 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9200 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9201 2 : Bytes::from_static(b"value 4@0x10"),
9202 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9203 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9204 2 : Bytes::from_static(b"value 7@0x10"),
9205 2 : Bytes::from_static(b"value 8@0x10@0x48"),
9206 2 : Bytes::from_static(b"value 9@0x10@0x48"),
9207 2 : ];
9208 2 :
9209 2 : let expected_result_at_lsn_40 = [
9210 2 : Bytes::from_static(b"value 0@0x10"),
9211 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9212 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9213 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9214 2 : Bytes::from_static(b"value 4@0x10"),
9215 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9216 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9217 2 : Bytes::from_static(b"value 7@0x10"),
9218 2 : Bytes::from_static(b"value 8@0x10"),
9219 2 : Bytes::from_static(b"value 9@0x10"),
9220 2 : ];
9221 2 :
9222 4 : let verify_result = || async {
9223 44 : for idx in 0..10 {
9224 40 : assert_eq!(
9225 40 : branch_tline
9226 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9227 52 : .await
9228 40 : .unwrap(),
9229 40 : &expected_result_at_gc_horizon[idx]
9230 2 : );
9231 40 : assert_eq!(
9232 40 : branch_tline
9233 40 : .get(get_key(idx as u32), Lsn(0x40), &ctx)
9234 31 : .await
9235 40 : .unwrap(),
9236 40 : &expected_result_at_lsn_40[idx]
9237 2 : );
9238 2 : }
9239 8 : };
9240 2 :
9241 46 : verify_result().await;
9242 2 :
9243 2 : let cancel = CancellationToken::new();
9244 2 : branch_tline
9245 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
9246 19 : .await
9247 2 : .unwrap();
9248 2 :
9249 37 : verify_result().await;
9250 2 :
9251 2 : Ok(())
9252 2 : }
9253 :
9254 : // Regression test for https://github.com/neondatabase/neon/issues/9012
9255 : // Create an image arrangement where we have to read at different LSN ranges
9256 : // from a delta layer. This is achieved by overlapping an image layer on top of
9257 : // a delta layer. Like so:
9258 : //
9259 : // A B
9260 : // +----------------+ -> delta_layer
9261 : // | | ^ lsn
9262 : // | =========|-> nested_image_layer |
9263 : // | C | |
9264 : // +----------------+ |
9265 : // ======== -> baseline_image_layer +-------> key
9266 : //
9267 : //
9268 : // When querying the key range [A, B) we need to read at different LSN ranges
9269 : // for [A, C) and [C, B). This test checks that the described edge case is handled correctly.
9270 : #[cfg(feature = "testing")]
9271 : #[tokio::test]
9272 2 : async fn test_vectored_read_with_nested_image_layer() -> anyhow::Result<()> {
9273 2 : let harness = TenantHarness::create("test_vectored_read_with_nested_image_layer").await?;
9274 20 : let (tenant, ctx) = harness.load().await;
9275 2 :
9276 2 : let will_init_keys = [2, 6];
9277 44 : fn get_key(id: u32) -> Key {
9278 44 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
9279 44 : key.field6 = id;
9280 44 : key
9281 44 : }
9282 2 :
9283 2 : let mut expected_key_values = HashMap::new();
9284 2 :
9285 2 : let baseline_image_layer_lsn = Lsn(0x10);
9286 2 : let mut baseline_img_layer = Vec::new();
9287 12 : for i in 0..5 {
9288 10 : let key = get_key(i);
9289 10 : let value = format!("value {i}@{baseline_image_layer_lsn}");
9290 10 :
9291 10 : let removed = expected_key_values.insert(key, value.clone());
9292 10 : assert!(removed.is_none());
9293 2 :
9294 10 : baseline_img_layer.push((key, Bytes::from(value)));
9295 2 : }
9296 2 :
9297 2 : let nested_image_layer_lsn = Lsn(0x50);
9298 2 : let mut nested_img_layer = Vec::new();
9299 12 : for i in 5..10 {
9300 10 : let key = get_key(i);
9301 10 : let value = format!("value {i}@{nested_image_layer_lsn}");
9302 10 :
9303 10 : let removed = expected_key_values.insert(key, value.clone());
9304 10 : assert!(removed.is_none());
9305 2 :
9306 10 : nested_img_layer.push((key, Bytes::from(value)));
9307 2 : }
9308 2 :
9309 2 : let mut delta_layer_spec = Vec::default();
9310 2 : let delta_layer_start_lsn = Lsn(0x20);
9311 2 : let mut delta_layer_end_lsn = delta_layer_start_lsn;
9312 2 :
9313 22 : for i in 0..10 {
9314 20 : let key = get_key(i);
9315 20 : let key_in_nested = nested_img_layer
9316 20 : .iter()
9317 80 : .any(|(key_with_img, _)| *key_with_img == key);
9318 20 : let lsn = {
9319 20 : if key_in_nested {
9320 10 : Lsn(nested_image_layer_lsn.0 + 0x10)
9321 2 : } else {
9322 10 : delta_layer_start_lsn
9323 2 : }
9324 2 : };
9325 2 :
9326 20 : let will_init = will_init_keys.contains(&i);
9327 20 : if will_init {
9328 4 : delta_layer_spec.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init(""))));
9329 4 :
9330 4 : expected_key_values.insert(key, "".to_string());
9331 16 : } else {
9332 16 : let delta = format!("@{lsn}");
9333 16 : delta_layer_spec.push((
9334 16 : key,
9335 16 : lsn,
9336 16 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
9337 16 : ));
9338 16 :
9339 16 : expected_key_values
9340 16 : .get_mut(&key)
9341 16 : .expect("An image exists for each key")
9342 16 : .push_str(delta.as_str());
9343 16 : }
9344 20 : delta_layer_end_lsn = std::cmp::max(delta_layer_start_lsn, lsn);
9345 2 : }
9346 2 :
9347 2 : delta_layer_end_lsn = Lsn(delta_layer_end_lsn.0 + 1);
9348 2 :
9349 2 : assert!(
9350 2 : nested_image_layer_lsn > delta_layer_start_lsn
9351 2 : && nested_image_layer_lsn < delta_layer_end_lsn
9352 2 : );
9353 2 :
9354 2 : let tline = tenant
9355 2 : .create_test_timeline_with_layers(
9356 2 : TIMELINE_ID,
9357 2 : baseline_image_layer_lsn,
9358 2 : DEFAULT_PG_VERSION,
9359 2 : &ctx,
9360 2 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
9361 2 : delta_layer_start_lsn..delta_layer_end_lsn,
9362 2 : delta_layer_spec,
9363 2 : )], // delta layers
9364 2 : vec![
9365 2 : (baseline_image_layer_lsn, baseline_img_layer),
9366 2 : (nested_image_layer_lsn, nested_img_layer),
9367 2 : ], // image layers
9368 2 : delta_layer_end_lsn,
9369 2 : )
9370 42 : .await?;
9371 2 :
9372 2 : let keyspace = KeySpace::single(get_key(0)..get_key(10));
9373 2 : let results = tline
9374 2 : .get_vectored(keyspace, delta_layer_end_lsn, &ctx)
9375 13 : .await
9376 2 : .expect("No vectored errors");
9377 22 : for (key, res) in results {
9378 20 : let value = res.expect("No key errors");
9379 20 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
9380 20 : assert_eq!(value, Bytes::from(expected_value));
9381 2 : }
9382 2 :
9383 2 : Ok(())
9384 2 : }
9385 :
9386 142 : fn sort_layer_key(k1: &PersistentLayerKey, k2: &PersistentLayerKey) -> std::cmp::Ordering {
9387 142 : (
9388 142 : k1.is_delta,
9389 142 : k1.key_range.start,
9390 142 : k1.key_range.end,
9391 142 : k1.lsn_range.start,
9392 142 : k1.lsn_range.end,
9393 142 : )
9394 142 : .cmp(&(
9395 142 : k2.is_delta,
9396 142 : k2.key_range.start,
9397 142 : k2.key_range.end,
9398 142 : k2.lsn_range.start,
9399 142 : k2.lsn_range.end,
9400 142 : ))
9401 142 : }
9402 :
9403 12 : async fn inspect_and_sort(
9404 12 : tline: &Arc<Timeline>,
9405 12 : filter: Option<std::ops::Range<Key>>,
9406 12 : ) -> Vec<PersistentLayerKey> {
9407 12 : let mut all_layers = tline.inspect_historic_layers().await.unwrap();
9408 12 : if let Some(filter) = filter {
9409 64 : all_layers.retain(|layer| overlaps_with(&layer.key_range, &filter));
9410 10 : }
9411 12 : all_layers.sort_by(sort_layer_key);
9412 12 : all_layers
9413 12 : }
9414 :
9415 : #[cfg(feature = "testing")]
9416 10 : fn check_layer_map_key_eq(
9417 10 : mut left: Vec<PersistentLayerKey>,
9418 10 : mut right: Vec<PersistentLayerKey>,
9419 10 : ) {
9420 10 : left.sort_by(sort_layer_key);
9421 10 : right.sort_by(sort_layer_key);
9422 10 : if left != right {
9423 0 : eprintln!("---LEFT---");
9424 0 : for left in left.iter() {
9425 0 : eprintln!("{}", left);
9426 0 : }
9427 0 : eprintln!("---RIGHT---");
9428 0 : for right in right.iter() {
9429 0 : eprintln!("{}", right);
9430 0 : }
9431 0 : assert_eq!(left, right);
9432 10 : }
9433 10 : }
9434 :
9435 : #[cfg(feature = "testing")]
9436 : #[tokio::test]
9437 2 : async fn test_simple_partial_bottom_most_compaction() -> anyhow::Result<()> {
9438 2 : let harness = TenantHarness::create("test_simple_partial_bottom_most_compaction").await?;
9439 20 : let (tenant, ctx) = harness.load().await;
9440 2 :
9441 182 : fn get_key(id: u32) -> Key {
9442 182 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9443 182 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9444 182 : key.field6 = id;
9445 182 : key
9446 182 : }
9447 2 :
9448 2 : // img layer at 0x10
9449 2 : let img_layer = (0..10)
9450 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9451 2 : .collect_vec();
9452 2 :
9453 2 : let delta1 = vec![
9454 2 : (
9455 2 : get_key(1),
9456 2 : Lsn(0x20),
9457 2 : Value::Image(Bytes::from("value 1@0x20")),
9458 2 : ),
9459 2 : (
9460 2 : get_key(2),
9461 2 : Lsn(0x30),
9462 2 : Value::Image(Bytes::from("value 2@0x30")),
9463 2 : ),
9464 2 : (
9465 2 : get_key(3),
9466 2 : Lsn(0x40),
9467 2 : Value::Image(Bytes::from("value 3@0x40")),
9468 2 : ),
9469 2 : ];
9470 2 : let delta2 = vec![
9471 2 : (
9472 2 : get_key(5),
9473 2 : Lsn(0x20),
9474 2 : Value::Image(Bytes::from("value 5@0x20")),
9475 2 : ),
9476 2 : (
9477 2 : get_key(6),
9478 2 : Lsn(0x20),
9479 2 : Value::Image(Bytes::from("value 6@0x20")),
9480 2 : ),
9481 2 : ];
9482 2 : let delta3 = vec![
9483 2 : (
9484 2 : get_key(8),
9485 2 : Lsn(0x48),
9486 2 : Value::Image(Bytes::from("value 8@0x48")),
9487 2 : ),
9488 2 : (
9489 2 : get_key(9),
9490 2 : Lsn(0x48),
9491 2 : Value::Image(Bytes::from("value 9@0x48")),
9492 2 : ),
9493 2 : ];
9494 2 :
9495 2 : let tline = tenant
9496 2 : .create_test_timeline_with_layers(
9497 2 : TIMELINE_ID,
9498 2 : Lsn(0x10),
9499 2 : DEFAULT_PG_VERSION,
9500 2 : &ctx,
9501 2 : vec![
9502 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
9503 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
9504 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
9505 2 : ], // delta layers
9506 2 : vec![(Lsn(0x10), img_layer)], // image layers
9507 2 : Lsn(0x50),
9508 2 : )
9509 49 : .await?;
9510 2 :
9511 2 : {
9512 2 : // Update GC info
9513 2 : let mut guard = tline.gc_info.write().unwrap();
9514 2 : *guard = GcInfo {
9515 2 : retain_lsns: vec![(Lsn(0x20), tline.timeline_id, MaybeOffloaded::No)],
9516 2 : cutoffs: GcCutoffs {
9517 2 : time: Lsn(0x30),
9518 2 : space: Lsn(0x30),
9519 2 : },
9520 2 : leases: Default::default(),
9521 2 : within_ancestor_pitr: false,
9522 2 : };
9523 2 : }
9524 2 :
9525 2 : let cancel = CancellationToken::new();
9526 2 :
9527 2 : // Do a partial compaction on key range 0..2
9528 2 : tline
9529 2 : .partial_compact_with_gc(get_key(0)..get_key(2), &cancel, EnumSet::new(), &ctx)
9530 27 : .await
9531 2 : .unwrap();
9532 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
9533 2 : check_layer_map_key_eq(
9534 2 : all_layers,
9535 2 : vec![
9536 2 : // newly-generated image layer for the partial compaction range 0-2
9537 2 : PersistentLayerKey {
9538 2 : key_range: get_key(0)..get_key(2),
9539 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9540 2 : is_delta: false,
9541 2 : },
9542 2 : PersistentLayerKey {
9543 2 : key_range: get_key(0)..get_key(10),
9544 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
9545 2 : is_delta: false,
9546 2 : },
9547 2 : // delta1 is split and the second part is rewritten
9548 2 : PersistentLayerKey {
9549 2 : key_range: get_key(2)..get_key(4),
9550 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9551 2 : is_delta: true,
9552 2 : },
9553 2 : PersistentLayerKey {
9554 2 : key_range: get_key(5)..get_key(7),
9555 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9556 2 : is_delta: true,
9557 2 : },
9558 2 : PersistentLayerKey {
9559 2 : key_range: get_key(8)..get_key(10),
9560 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
9561 2 : is_delta: true,
9562 2 : },
9563 2 : ],
9564 2 : );
9565 2 :
9566 2 : // Do a partial compaction on key range 2..4
9567 2 : tline
9568 2 : .partial_compact_with_gc(get_key(2)..get_key(4), &cancel, EnumSet::new(), &ctx)
9569 19 : .await
9570 2 : .unwrap();
9571 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
9572 2 : check_layer_map_key_eq(
9573 2 : all_layers,
9574 2 : vec![
9575 2 : PersistentLayerKey {
9576 2 : key_range: get_key(0)..get_key(2),
9577 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9578 2 : is_delta: false,
9579 2 : },
9580 2 : PersistentLayerKey {
9581 2 : key_range: get_key(0)..get_key(10),
9582 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
9583 2 : is_delta: false,
9584 2 : },
9585 2 : // image layer generated for the compaction range 2-4
9586 2 : PersistentLayerKey {
9587 2 : key_range: get_key(2)..get_key(4),
9588 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9589 2 : is_delta: false,
9590 2 : },
9591 2 : // we have key2/key3 above the retain_lsn, so we still need this delta layer
9592 2 : PersistentLayerKey {
9593 2 : key_range: get_key(2)..get_key(4),
9594 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9595 2 : is_delta: true,
9596 2 : },
9597 2 : PersistentLayerKey {
9598 2 : key_range: get_key(5)..get_key(7),
9599 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9600 2 : is_delta: true,
9601 2 : },
9602 2 : PersistentLayerKey {
9603 2 : key_range: get_key(8)..get_key(10),
9604 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
9605 2 : is_delta: true,
9606 2 : },
9607 2 : ],
9608 2 : );
9609 2 :
9610 2 : // Do a partial compaction on key range 4..9
9611 2 : tline
9612 2 : .partial_compact_with_gc(get_key(4)..get_key(9), &cancel, EnumSet::new(), &ctx)
9613 24 : .await
9614 2 : .unwrap();
9615 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
9616 2 : check_layer_map_key_eq(
9617 2 : all_layers,
9618 2 : vec![
9619 2 : PersistentLayerKey {
9620 2 : key_range: get_key(0)..get_key(2),
9621 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9622 2 : is_delta: false,
9623 2 : },
9624 2 : PersistentLayerKey {
9625 2 : key_range: get_key(0)..get_key(10),
9626 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
9627 2 : is_delta: false,
9628 2 : },
9629 2 : PersistentLayerKey {
9630 2 : key_range: get_key(2)..get_key(4),
9631 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9632 2 : is_delta: false,
9633 2 : },
9634 2 : PersistentLayerKey {
9635 2 : key_range: get_key(2)..get_key(4),
9636 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9637 2 : is_delta: true,
9638 2 : },
9639 2 : // image layer generated for this compaction range
9640 2 : PersistentLayerKey {
9641 2 : key_range: get_key(4)..get_key(9),
9642 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9643 2 : is_delta: false,
9644 2 : },
9645 2 : PersistentLayerKey {
9646 2 : key_range: get_key(8)..get_key(10),
9647 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
9648 2 : is_delta: true,
9649 2 : },
9650 2 : ],
9651 2 : );
9652 2 :
9653 2 : // Do a partial compaction on key range 9..10
9654 2 : tline
9655 2 : .partial_compact_with_gc(get_key(9)..get_key(10), &cancel, EnumSet::new(), &ctx)
9656 10 : .await
9657 2 : .unwrap();
9658 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
9659 2 : check_layer_map_key_eq(
9660 2 : all_layers,
9661 2 : vec![
9662 2 : PersistentLayerKey {
9663 2 : key_range: get_key(0)..get_key(2),
9664 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9665 2 : is_delta: false,
9666 2 : },
9667 2 : PersistentLayerKey {
9668 2 : key_range: get_key(0)..get_key(10),
9669 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
9670 2 : is_delta: false,
9671 2 : },
9672 2 : PersistentLayerKey {
9673 2 : key_range: get_key(2)..get_key(4),
9674 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
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(0x48),
9680 2 : is_delta: true,
9681 2 : },
9682 2 : PersistentLayerKey {
9683 2 : key_range: get_key(4)..get_key(9),
9684 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9685 2 : is_delta: false,
9686 2 : },
9687 2 : // image layer generated for the compaction range
9688 2 : PersistentLayerKey {
9689 2 : key_range: get_key(9)..get_key(10),
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 0..10, all image layers below LSN 20 can be replaced with new ones.
9702 2 : tline
9703 2 : .partial_compact_with_gc(get_key(0)..get_key(10), &cancel, EnumSet::new(), &ctx)
9704 54 : .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 : // aha, we removed all unnecessary image/delta layers and got a very clean layer map!
9711 2 : PersistentLayerKey {
9712 2 : key_range: get_key(0)..get_key(10),
9713 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9714 2 : is_delta: false,
9715 2 : },
9716 2 : PersistentLayerKey {
9717 2 : key_range: get_key(2)..get_key(4),
9718 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9719 2 : is_delta: true,
9720 2 : },
9721 2 : PersistentLayerKey {
9722 2 : key_range: get_key(8)..get_key(10),
9723 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
9724 2 : is_delta: true,
9725 2 : },
9726 2 : ],
9727 2 : );
9728 2 :
9729 2 : Ok(())
9730 2 : }
9731 :
9732 : #[cfg(feature = "testing")]
9733 : #[tokio::test]
9734 2 : async fn test_timeline_offload_retain_lsn() -> anyhow::Result<()> {
9735 2 : let harness = TenantHarness::create("test_timeline_offload_retain_lsn")
9736 2 : .await
9737 2 : .unwrap();
9738 19 : let (tenant, ctx) = harness.load().await;
9739 2 : let tline_parent = tenant
9740 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
9741 6 : .await
9742 2 : .unwrap();
9743 2 : let tline_child = tenant
9744 2 : .branch_timeline_test(&tline_parent, NEW_TIMELINE_ID, Some(Lsn(0x20)), &ctx)
9745 2 : .await
9746 2 : .unwrap();
9747 2 : {
9748 2 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
9749 2 : assert_eq!(
9750 2 : gc_info_parent.retain_lsns,
9751 2 : vec![(Lsn(0x20), tline_child.timeline_id, MaybeOffloaded::No)]
9752 2 : );
9753 2 : }
9754 2 : // We have to directly call the remote_client instead of using the archive function to avoid constructing broker client...
9755 2 : tline_child
9756 2 : .remote_client
9757 2 : .schedule_index_upload_for_timeline_archival_state(TimelineArchivalState::Archived)
9758 2 : .unwrap();
9759 2 : tline_child.remote_client.wait_completion().await.unwrap();
9760 2 : offload_timeline(&tenant, &tline_child)
9761 2 : .instrument(tracing::info_span!(parent: None, "offload_test", tenant_id=%"test", shard_id=%"test", timeline_id=%"test"))
9762 14 : .await.unwrap();
9763 2 : let child_timeline_id = tline_child.timeline_id;
9764 2 : Arc::try_unwrap(tline_child).unwrap();
9765 2 :
9766 2 : {
9767 2 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
9768 2 : assert_eq!(
9769 2 : gc_info_parent.retain_lsns,
9770 2 : vec![(Lsn(0x20), child_timeline_id, MaybeOffloaded::Yes)]
9771 2 : );
9772 2 : }
9773 2 :
9774 2 : tenant
9775 2 : .get_offloaded_timeline(child_timeline_id)
9776 2 : .unwrap()
9777 2 : .defuse_for_tenant_drop();
9778 2 :
9779 2 : Ok(())
9780 2 : }
9781 : }
|
