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 enumset::EnumSet;
20 : use futures::stream::FuturesUnordered;
21 : use futures::StreamExt;
22 : use pageserver_api::models;
23 : use pageserver_api::models::LsnLease;
24 : use pageserver_api::models::TimelineArchivalState;
25 : use pageserver_api::models::TimelineState;
26 : use pageserver_api::models::TopTenantShardItem;
27 : use pageserver_api::models::WalRedoManagerStatus;
28 : use pageserver_api::shard::ShardIdentity;
29 : use pageserver_api::shard::ShardStripeSize;
30 : use pageserver_api::shard::TenantShardId;
31 : use remote_storage::DownloadError;
32 : use remote_storage::GenericRemoteStorage;
33 : use remote_storage::TimeoutOrCancel;
34 : use std::collections::BTreeMap;
35 : use std::fmt;
36 : use std::future::Future;
37 : use std::sync::Weak;
38 : use std::time::SystemTime;
39 : use storage_broker::BrokerClientChannel;
40 : use timeline::offload::offload_timeline;
41 : use tokio::io::BufReader;
42 : use tokio::sync::watch;
43 : use tokio::task::JoinSet;
44 : use tokio_util::sync::CancellationToken;
45 : use tracing::*;
46 : use upload_queue::NotInitialized;
47 : use utils::backoff;
48 : use utils::circuit_breaker::CircuitBreaker;
49 : use utils::completion;
50 : use utils::crashsafe::path_with_suffix_extension;
51 : use utils::failpoint_support;
52 : use utils::fs_ext;
53 : use utils::pausable_failpoint;
54 : use utils::sync::gate::Gate;
55 : use utils::sync::gate::GateGuard;
56 : use utils::timeout::timeout_cancellable;
57 : use utils::timeout::TimeoutCancellableError;
58 : use utils::zstd::create_zst_tarball;
59 : use utils::zstd::extract_zst_tarball;
60 :
61 : use self::config::AttachedLocationConfig;
62 : use self::config::AttachmentMode;
63 : use self::config::LocationConf;
64 : use self::config::TenantConf;
65 : use self::metadata::TimelineMetadata;
66 : use self::mgr::GetActiveTenantError;
67 : use self::mgr::GetTenantError;
68 : use self::remote_timeline_client::upload::upload_index_part;
69 : use self::remote_timeline_client::{RemoteTimelineClient, WaitCompletionError};
70 : use self::timeline::uninit::TimelineCreateGuard;
71 : use self::timeline::uninit::TimelineExclusionError;
72 : use self::timeline::uninit::UninitializedTimeline;
73 : use self::timeline::EvictionTaskTenantState;
74 : use self::timeline::GcCutoffs;
75 : use self::timeline::TimelineResources;
76 : use self::timeline::WaitLsnError;
77 : use crate::config::PageServerConf;
78 : use crate::context::{DownloadBehavior, RequestContext};
79 : use crate::deletion_queue::DeletionQueueClient;
80 : use crate::deletion_queue::DeletionQueueError;
81 : use crate::import_datadir;
82 : use crate::is_uninit_mark;
83 : use crate::l0_flush::L0FlushGlobalState;
84 : use crate::metrics::TENANT;
85 : use crate::metrics::{
86 : remove_tenant_metrics, BROKEN_TENANTS_SET, CIRCUIT_BREAKERS_BROKEN, CIRCUIT_BREAKERS_UNBROKEN,
87 : TENANT_STATE_METRIC, TENANT_SYNTHETIC_SIZE_METRIC,
88 : };
89 : use crate::repository::GcResult;
90 : use crate::task_mgr;
91 : use crate::task_mgr::TaskKind;
92 : use crate::tenant::config::LocationMode;
93 : use crate::tenant::config::TenantConfOpt;
94 : pub use crate::tenant::remote_timeline_client::index::IndexPart;
95 : use crate::tenant::remote_timeline_client::remote_initdb_archive_path;
96 : use crate::tenant::remote_timeline_client::MaybeDeletedIndexPart;
97 : use crate::tenant::remote_timeline_client::INITDB_PATH;
98 : use crate::tenant::storage_layer::DeltaLayer;
99 : use crate::tenant::storage_layer::ImageLayer;
100 : use crate::walingest::WalLagCooldown;
101 : use crate::walredo;
102 : use crate::InitializationOrder;
103 : use std::collections::hash_map::Entry;
104 : use std::collections::HashMap;
105 : use std::collections::HashSet;
106 : use std::fmt::Debug;
107 : use std::fmt::Display;
108 : use std::fs;
109 : use std::fs::File;
110 : use std::sync::atomic::{AtomicU64, Ordering};
111 : use std::sync::Arc;
112 : use std::sync::Mutex;
113 : use std::time::{Duration, Instant};
114 :
115 : use crate::span;
116 : use crate::tenant::timeline::delete::DeleteTimelineFlow;
117 : use crate::tenant::timeline::uninit::cleanup_timeline_directory;
118 : use crate::virtual_file::VirtualFile;
119 : use crate::walredo::PostgresRedoManager;
120 : use crate::TEMP_FILE_SUFFIX;
121 : use once_cell::sync::Lazy;
122 : pub use pageserver_api::models::TenantState;
123 : use tokio::sync::Semaphore;
124 :
125 0 : static INIT_DB_SEMAPHORE: Lazy<Semaphore> = Lazy::new(|| Semaphore::new(8));
126 : use utils::{
127 : crashsafe,
128 : generation::Generation,
129 : id::TimelineId,
130 : lsn::{Lsn, RecordLsn},
131 : };
132 :
133 : pub mod blob_io;
134 : pub mod block_io;
135 : pub mod vectored_blob_io;
136 :
137 : pub mod disk_btree;
138 : pub(crate) mod ephemeral_file;
139 : pub mod layer_map;
140 :
141 : pub mod metadata;
142 : pub mod remote_timeline_client;
143 : pub mod storage_layer;
144 :
145 : pub mod checks;
146 : pub mod config;
147 : pub mod mgr;
148 : pub mod secondary;
149 : pub mod tasks;
150 : pub mod upload_queue;
151 :
152 : pub(crate) mod timeline;
153 :
154 : pub mod size;
155 :
156 : mod gc_block;
157 : pub(crate) mod throttle;
158 :
159 : pub(crate) use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
160 : pub(crate) use timeline::{LogicalSizeCalculationCause, PageReconstructError, Timeline};
161 :
162 : // re-export for use in walreceiver
163 : pub use crate::tenant::timeline::WalReceiverInfo;
164 :
165 : /// The "tenants" part of `tenants/<tenant>/timelines...`
166 : pub const TENANTS_SEGMENT_NAME: &str = "tenants";
167 :
168 : /// Parts of the `.neon/tenants/<tenant_id>/timelines/<timeline_id>` directory prefix.
169 : pub const TIMELINES_SEGMENT_NAME: &str = "timelines";
170 :
171 : /// References to shared objects that are passed into each tenant, such
172 : /// as the shared remote storage client and process initialization state.
173 : #[derive(Clone)]
174 : pub struct TenantSharedResources {
175 : pub broker_client: storage_broker::BrokerClientChannel,
176 : pub remote_storage: GenericRemoteStorage,
177 : pub deletion_queue_client: DeletionQueueClient,
178 : pub l0_flush_global_state: L0FlushGlobalState,
179 : }
180 :
181 : /// A [`Tenant`] is really an _attached_ tenant. The configuration
182 : /// for an attached tenant is a subset of the [`LocationConf`], represented
183 : /// in this struct.
184 : pub(super) struct AttachedTenantConf {
185 : tenant_conf: TenantConfOpt,
186 : location: AttachedLocationConfig,
187 : /// The deadline before which we are blocked from GC so that
188 : /// leases have a chance to be renewed.
189 : lsn_lease_deadline: Option<tokio::time::Instant>,
190 : }
191 :
192 : impl AttachedTenantConf {
193 186 : fn new(tenant_conf: TenantConfOpt, location: AttachedLocationConfig) -> Self {
194 : // Sets a deadline before which we cannot proceed to GC due to lsn lease.
195 : //
196 : // We do this as the leases mapping are not persisted to disk. By delaying GC by lease
197 : // length, we guarantee that all the leases we granted before will have a chance to renew
198 : // when we run GC for the first time after restart / transition from AttachedMulti to AttachedSingle.
199 186 : let lsn_lease_deadline = if location.attach_mode == AttachmentMode::Single {
200 186 : Some(
201 186 : tokio::time::Instant::now()
202 186 : + tenant_conf
203 186 : .lsn_lease_length
204 186 : .unwrap_or(LsnLease::DEFAULT_LENGTH),
205 186 : )
206 : } else {
207 : // We don't use `lsn_lease_deadline` to delay GC in AttachedMulti and AttachedStale
208 : // because we don't do GC in these modes.
209 0 : None
210 : };
211 :
212 186 : Self {
213 186 : tenant_conf,
214 186 : location,
215 186 : lsn_lease_deadline,
216 186 : }
217 186 : }
218 :
219 186 : fn try_from(location_conf: LocationConf) -> anyhow::Result<Self> {
220 186 : match &location_conf.mode {
221 186 : LocationMode::Attached(attach_conf) => {
222 186 : Ok(Self::new(location_conf.tenant_conf, *attach_conf))
223 : }
224 : LocationMode::Secondary(_) => {
225 0 : anyhow::bail!("Attempted to construct AttachedTenantConf from a LocationConf in secondary mode")
226 : }
227 : }
228 186 : }
229 :
230 762 : fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
231 762 : self.lsn_lease_deadline
232 762 : .map(|d| tokio::time::Instant::now() < d)
233 762 : .unwrap_or(false)
234 762 : }
235 : }
236 : struct TimelinePreload {
237 : timeline_id: TimelineId,
238 : client: RemoteTimelineClient,
239 : index_part: Result<MaybeDeletedIndexPart, DownloadError>,
240 : }
241 :
242 : pub(crate) struct TenantPreload {
243 : timelines: HashMap<TimelineId, TimelinePreload>,
244 : }
245 :
246 : /// When we spawn a tenant, there is a special mode for tenant creation that
247 : /// avoids trying to read anything from remote storage.
248 : pub(crate) enum SpawnMode {
249 : /// Activate as soon as possible
250 : Eager,
251 : /// Lazy activation in the background, with the option to skip the queue if the need comes up
252 : Lazy,
253 : }
254 :
255 : ///
256 : /// Tenant consists of multiple timelines. Keep them in a hash table.
257 : ///
258 : pub struct Tenant {
259 : // Global pageserver config parameters
260 : pub conf: &'static PageServerConf,
261 :
262 : /// The value creation timestamp, used to measure activation delay, see:
263 : /// <https://github.com/neondatabase/neon/issues/4025>
264 : constructed_at: Instant,
265 :
266 : state: watch::Sender<TenantState>,
267 :
268 : // Overridden tenant-specific config parameters.
269 : // We keep TenantConfOpt sturct here to preserve the information
270 : // about parameters that are not set.
271 : // This is necessary to allow global config updates.
272 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
273 :
274 : tenant_shard_id: TenantShardId,
275 :
276 : // The detailed sharding information, beyond the number/count in tenant_shard_id
277 : shard_identity: ShardIdentity,
278 :
279 : /// The remote storage generation, used to protect S3 objects from split-brain.
280 : /// Does not change over the lifetime of the [`Tenant`] object.
281 : ///
282 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
283 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
284 : generation: Generation,
285 :
286 : timelines: Mutex<HashMap<TimelineId, Arc<Timeline>>>,
287 :
288 : /// During timeline creation, we first insert the TimelineId to the
289 : /// creating map, then `timelines`, then remove it from the creating map.
290 : /// **Lock order**: if acquiring both, acquire`timelines` before `timelines_creating`
291 : timelines_creating: std::sync::Mutex<HashSet<TimelineId>>,
292 :
293 : /// Possibly offloaded and archived timelines
294 : /// **Lock order**: if acquiring both, acquire`timelines` before `timelines_offloaded`
295 : timelines_offloaded: Mutex<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
296 :
297 : // This mutex prevents creation of new timelines during GC.
298 : // Adding yet another mutex (in addition to `timelines`) is needed because holding
299 : // `timelines` mutex during all GC iteration
300 : // may block for a long time `get_timeline`, `get_timelines_state`,... and other operations
301 : // with timelines, which in turn may cause dropping replication connection, expiration of wait_for_lsn
302 : // timeout...
303 : gc_cs: tokio::sync::Mutex<()>,
304 : walredo_mgr: Option<Arc<WalRedoManager>>,
305 :
306 : // provides access to timeline data sitting in the remote storage
307 : pub(crate) remote_storage: GenericRemoteStorage,
308 :
309 : // Access to global deletion queue for when this tenant wants to schedule a deletion
310 : deletion_queue_client: DeletionQueueClient,
311 :
312 : /// Cached logical sizes updated updated on each [`Tenant::gather_size_inputs`].
313 : cached_logical_sizes: tokio::sync::Mutex<HashMap<(TimelineId, Lsn), u64>>,
314 : cached_synthetic_tenant_size: Arc<AtomicU64>,
315 :
316 : eviction_task_tenant_state: tokio::sync::Mutex<EvictionTaskTenantState>,
317 :
318 : /// Track repeated failures to compact, so that we can back off.
319 : /// Overhead of mutex is acceptable because compaction is done with a multi-second period.
320 : compaction_circuit_breaker: std::sync::Mutex<CircuitBreaker>,
321 :
322 : /// If the tenant is in Activating state, notify this to encourage it
323 : /// to proceed to Active as soon as possible, rather than waiting for lazy
324 : /// background warmup.
325 : pub(crate) activate_now_sem: tokio::sync::Semaphore,
326 :
327 : /// Time it took for the tenant to activate. Zero if not active yet.
328 : attach_wal_lag_cooldown: Arc<std::sync::OnceLock<WalLagCooldown>>,
329 :
330 : // Cancellation token fires when we have entered shutdown(). This is a parent of
331 : // Timelines' cancellation token.
332 : pub(crate) cancel: CancellationToken,
333 :
334 : // Users of the Tenant such as the page service must take this Gate to avoid
335 : // trying to use a Tenant which is shutting down.
336 : pub(crate) gate: Gate,
337 :
338 : /// Throttle applied at the top of [`Timeline::get`].
339 : /// All [`Tenant::timelines`] of a given [`Tenant`] instance share the same [`throttle::Throttle`] instance.
340 : pub(crate) timeline_get_throttle:
341 : Arc<throttle::Throttle<crate::metrics::tenant_throttling::TimelineGet>>,
342 :
343 : /// An ongoing timeline detach concurrency limiter.
344 : ///
345 : /// As a tenant will likely be restarted as part of timeline detach ancestor it makes no sense
346 : /// to have two running at the same time. A different one can be started if an earlier one
347 : /// has failed for whatever reason.
348 : ongoing_timeline_detach: std::sync::Mutex<Option<(TimelineId, utils::completion::Barrier)>>,
349 :
350 : /// `index_part.json` based gc blocking reason tracking.
351 : ///
352 : /// New gc iterations must start a new iteration by acquiring `GcBlock::start` before
353 : /// proceeding.
354 : pub(crate) gc_block: gc_block::GcBlock,
355 :
356 : l0_flush_global_state: L0FlushGlobalState,
357 : }
358 :
359 : impl std::fmt::Debug for Tenant {
360 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
361 0 : write!(f, "{} ({})", self.tenant_shard_id, self.current_state())
362 0 : }
363 : }
364 :
365 : pub(crate) enum WalRedoManager {
366 : Prod(WalredoManagerId, PostgresRedoManager),
367 : #[cfg(test)]
368 : Test(harness::TestRedoManager),
369 : }
370 :
371 0 : #[derive(thiserror::Error, Debug)]
372 : #[error("pageserver is shutting down")]
373 : pub(crate) struct GlobalShutDown;
374 :
375 : impl WalRedoManager {
376 0 : pub(crate) fn new(mgr: PostgresRedoManager) -> Result<Arc<Self>, GlobalShutDown> {
377 0 : let id = WalredoManagerId::next();
378 0 : let arc = Arc::new(Self::Prod(id, mgr));
379 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
380 0 : match &mut *guard {
381 0 : Some(map) => {
382 0 : map.insert(id, Arc::downgrade(&arc));
383 0 : Ok(arc)
384 : }
385 0 : None => Err(GlobalShutDown),
386 : }
387 0 : }
388 : }
389 :
390 : impl Drop for WalRedoManager {
391 10 : fn drop(&mut self) {
392 10 : match self {
393 0 : Self::Prod(id, _) => {
394 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
395 0 : if let Some(map) = &mut *guard {
396 0 : map.remove(id).expect("new() registers, drop() unregisters");
397 0 : }
398 : }
399 : #[cfg(test)]
400 10 : Self::Test(_) => {
401 10 : // Not applicable to test redo manager
402 10 : }
403 : }
404 10 : }
405 : }
406 :
407 : /// Global registry of all walredo managers so that [`crate::shutdown_pageserver`] can shut down
408 : /// the walredo processes outside of the regular order.
409 : ///
410 : /// This is necessary to work around a systemd bug where it freezes if there are
411 : /// walredo processes left => <https://github.com/neondatabase/cloud/issues/11387>
412 : #[allow(clippy::type_complexity)]
413 : pub(crate) static WALREDO_MANAGERS: once_cell::sync::Lazy<
414 : Mutex<Option<HashMap<WalredoManagerId, Weak<WalRedoManager>>>>,
415 0 : > = once_cell::sync::Lazy::new(|| Mutex::new(Some(HashMap::new())));
416 : #[derive(PartialEq, Eq, Hash, Clone, Copy, Debug)]
417 : pub(crate) struct WalredoManagerId(u64);
418 : impl WalredoManagerId {
419 0 : pub fn next() -> Self {
420 : static NEXT: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(1);
421 0 : let id = NEXT.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
422 0 : if id == 0 {
423 0 : panic!("WalredoManagerId::new() returned 0, indicating wraparound, risking it's no longer unique");
424 0 : }
425 0 : Self(id)
426 0 : }
427 : }
428 :
429 : #[cfg(test)]
430 : impl From<harness::TestRedoManager> for WalRedoManager {
431 186 : fn from(mgr: harness::TestRedoManager) -> Self {
432 186 : Self::Test(mgr)
433 186 : }
434 : }
435 :
436 : impl WalRedoManager {
437 6 : pub(crate) async fn shutdown(&self) -> bool {
438 6 : match self {
439 0 : Self::Prod(_, mgr) => mgr.shutdown().await,
440 : #[cfg(test)]
441 : Self::Test(_) => {
442 : // Not applicable to test redo manager
443 6 : true
444 : }
445 : }
446 6 : }
447 :
448 0 : pub(crate) fn maybe_quiesce(&self, idle_timeout: Duration) {
449 0 : match self {
450 0 : Self::Prod(_, mgr) => mgr.maybe_quiesce(idle_timeout),
451 0 : #[cfg(test)]
452 0 : Self::Test(_) => {
453 0 : // Not applicable to test redo manager
454 0 : }
455 0 : }
456 0 : }
457 :
458 : /// # Cancel-Safety
459 : ///
460 : /// This method is cancellation-safe.
461 342 : pub async fn request_redo(
462 342 : &self,
463 342 : key: crate::repository::Key,
464 342 : lsn: Lsn,
465 342 : base_img: Option<(Lsn, bytes::Bytes)>,
466 342 : records: Vec<(Lsn, crate::walrecord::NeonWalRecord)>,
467 342 : pg_version: u32,
468 342 : ) -> Result<bytes::Bytes, walredo::Error> {
469 342 : match self {
470 0 : Self::Prod(_, mgr) => {
471 0 : mgr.request_redo(key, lsn, base_img, records, pg_version)
472 0 : .await
473 : }
474 : #[cfg(test)]
475 342 : Self::Test(mgr) => {
476 342 : mgr.request_redo(key, lsn, base_img, records, pg_version)
477 0 : .await
478 : }
479 : }
480 342 : }
481 :
482 0 : pub(crate) fn status(&self) -> Option<WalRedoManagerStatus> {
483 0 : match self {
484 0 : WalRedoManager::Prod(_, m) => Some(m.status()),
485 0 : #[cfg(test)]
486 0 : WalRedoManager::Test(_) => None,
487 0 : }
488 0 : }
489 : }
490 :
491 : pub struct OffloadedTimeline {
492 : pub tenant_shard_id: TenantShardId,
493 : pub timeline_id: TimelineId,
494 : pub ancestor_timeline_id: Option<TimelineId>,
495 : /// Whether to retain the branch lsn at the ancestor or not
496 : pub ancestor_retain_lsn: Option<Lsn>,
497 :
498 : // TODO: once we persist offloaded state, make this lazily constructed
499 : pub remote_client: Arc<RemoteTimelineClient>,
500 :
501 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
502 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
503 : pub delete_progress: Arc<tokio::sync::Mutex<DeleteTimelineFlow>>,
504 : }
505 :
506 : impl OffloadedTimeline {
507 0 : fn from_timeline(timeline: &Timeline) -> Self {
508 0 : let ancestor_retain_lsn = timeline
509 0 : .get_ancestor_timeline_id()
510 0 : .map(|_timeline_id| timeline.get_ancestor_lsn());
511 0 : Self {
512 0 : tenant_shard_id: timeline.tenant_shard_id,
513 0 : timeline_id: timeline.timeline_id,
514 0 : ancestor_timeline_id: timeline.get_ancestor_timeline_id(),
515 0 : ancestor_retain_lsn,
516 0 :
517 0 : remote_client: timeline.remote_client.clone(),
518 0 : delete_progress: timeline.delete_progress.clone(),
519 0 : }
520 0 : }
521 : }
522 :
523 : #[derive(Copy, Clone, PartialEq, Eq, Hash, Debug)]
524 : pub enum MaybeOffloaded {
525 : Yes,
526 : No,
527 : }
528 :
529 : #[derive(Clone)]
530 : pub enum TimelineOrOffloaded {
531 : Timeline(Arc<Timeline>),
532 : Offloaded(Arc<OffloadedTimeline>),
533 : }
534 :
535 : impl TimelineOrOffloaded {
536 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
537 0 : match self {
538 0 : TimelineOrOffloaded::Timeline(timeline) => timeline.tenant_shard_id,
539 0 : TimelineOrOffloaded::Offloaded(offloaded) => offloaded.tenant_shard_id,
540 : }
541 0 : }
542 0 : pub fn timeline_id(&self) -> TimelineId {
543 0 : match self {
544 0 : TimelineOrOffloaded::Timeline(timeline) => timeline.timeline_id,
545 0 : TimelineOrOffloaded::Offloaded(offloaded) => offloaded.timeline_id,
546 : }
547 0 : }
548 0 : pub fn delete_progress(&self) -> &Arc<tokio::sync::Mutex<DeleteTimelineFlow>> {
549 0 : match self {
550 0 : TimelineOrOffloaded::Timeline(timeline) => &timeline.delete_progress,
551 0 : TimelineOrOffloaded::Offloaded(offloaded) => &offloaded.delete_progress,
552 : }
553 0 : }
554 0 : pub fn remote_client(&self) -> &Arc<RemoteTimelineClient> {
555 0 : match self {
556 0 : TimelineOrOffloaded::Timeline(timeline) => &timeline.remote_client,
557 0 : TimelineOrOffloaded::Offloaded(offloaded) => &offloaded.remote_client,
558 : }
559 0 : }
560 : }
561 :
562 0 : #[derive(Debug, thiserror::Error, PartialEq, Eq)]
563 : pub enum GetTimelineError {
564 : #[error("Timeline is shutting down")]
565 : ShuttingDown,
566 : #[error("Timeline {tenant_id}/{timeline_id} is not active, state: {state:?}")]
567 : NotActive {
568 : tenant_id: TenantShardId,
569 : timeline_id: TimelineId,
570 : state: TimelineState,
571 : },
572 : #[error("Timeline {tenant_id}/{timeline_id} was not found")]
573 : NotFound {
574 : tenant_id: TenantShardId,
575 : timeline_id: TimelineId,
576 : },
577 : }
578 :
579 0 : #[derive(Debug, thiserror::Error)]
580 : pub enum LoadLocalTimelineError {
581 : #[error("FailedToLoad")]
582 : Load(#[source] anyhow::Error),
583 : #[error("FailedToResumeDeletion")]
584 : ResumeDeletion(#[source] anyhow::Error),
585 : }
586 :
587 0 : #[derive(thiserror::Error)]
588 : pub enum DeleteTimelineError {
589 : #[error("NotFound")]
590 : NotFound,
591 :
592 : #[error("HasChildren")]
593 : HasChildren(Vec<TimelineId>),
594 :
595 : #[error("Timeline deletion is already in progress")]
596 : AlreadyInProgress(Arc<tokio::sync::Mutex<DeleteTimelineFlow>>),
597 :
598 : #[error(transparent)]
599 : Other(#[from] anyhow::Error),
600 : }
601 :
602 : impl Debug for DeleteTimelineError {
603 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
604 0 : match self {
605 0 : Self::NotFound => write!(f, "NotFound"),
606 0 : Self::HasChildren(c) => f.debug_tuple("HasChildren").field(c).finish(),
607 0 : Self::AlreadyInProgress(_) => f.debug_tuple("AlreadyInProgress").finish(),
608 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
609 : }
610 0 : }
611 : }
612 :
613 0 : #[derive(thiserror::Error)]
614 : pub enum TimelineArchivalError {
615 : #[error("NotFound")]
616 : NotFound,
617 :
618 : #[error("Timeout")]
619 : Timeout,
620 :
621 : #[error("Cancelled")]
622 : Cancelled,
623 :
624 : #[error("ancestor is archived: {}", .0)]
625 : HasArchivedParent(TimelineId),
626 :
627 : #[error("HasUnarchivedChildren")]
628 : HasUnarchivedChildren(Vec<TimelineId>),
629 :
630 : #[error("Timeline archival is already in progress")]
631 : AlreadyInProgress,
632 :
633 : #[error(transparent)]
634 : Other(anyhow::Error),
635 : }
636 :
637 : impl Debug for TimelineArchivalError {
638 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
639 0 : match self {
640 0 : Self::NotFound => write!(f, "NotFound"),
641 0 : Self::Timeout => write!(f, "Timeout"),
642 0 : Self::Cancelled => write!(f, "Cancelled"),
643 0 : Self::HasArchivedParent(p) => f.debug_tuple("HasArchivedParent").field(p).finish(),
644 0 : Self::HasUnarchivedChildren(c) => {
645 0 : f.debug_tuple("HasUnarchivedChildren").field(c).finish()
646 : }
647 0 : Self::AlreadyInProgress => f.debug_tuple("AlreadyInProgress").finish(),
648 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
649 : }
650 0 : }
651 : }
652 :
653 : pub enum SetStoppingError {
654 : AlreadyStopping(completion::Barrier),
655 : Broken,
656 : }
657 :
658 : impl Debug for SetStoppingError {
659 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
660 0 : match self {
661 0 : Self::AlreadyStopping(_) => f.debug_tuple("AlreadyStopping").finish(),
662 0 : Self::Broken => write!(f, "Broken"),
663 : }
664 0 : }
665 : }
666 :
667 0 : #[derive(thiserror::Error, Debug)]
668 : pub enum CreateTimelineError {
669 : #[error("creation of timeline with the given ID is in progress")]
670 : AlreadyCreating,
671 : #[error("timeline already exists with different parameters")]
672 : Conflict,
673 : #[error(transparent)]
674 : AncestorLsn(anyhow::Error),
675 : #[error("ancestor timeline is not active")]
676 : AncestorNotActive,
677 : #[error("ancestor timeline is archived")]
678 : AncestorArchived,
679 : #[error("tenant shutting down")]
680 : ShuttingDown,
681 : #[error(transparent)]
682 : Other(#[from] anyhow::Error),
683 : }
684 :
685 : #[derive(thiserror::Error, Debug)]
686 : enum InitdbError {
687 : Other(anyhow::Error),
688 : Cancelled,
689 : Spawn(std::io::Result<()>),
690 : Failed(std::process::ExitStatus, Vec<u8>),
691 : }
692 :
693 : impl fmt::Display for InitdbError {
694 0 : fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
695 0 : match self {
696 0 : InitdbError::Cancelled => write!(f, "Operation was cancelled"),
697 0 : InitdbError::Spawn(e) => write!(f, "Spawn error: {:?}", e),
698 0 : InitdbError::Failed(status, stderr) => write!(
699 0 : f,
700 0 : "Command failed with status {:?}: {}",
701 0 : status,
702 0 : String::from_utf8_lossy(stderr)
703 0 : ),
704 0 : InitdbError::Other(e) => write!(f, "Error: {:?}", e),
705 : }
706 0 : }
707 : }
708 :
709 : impl From<std::io::Error> for InitdbError {
710 0 : fn from(error: std::io::Error) -> Self {
711 0 : InitdbError::Spawn(Err(error))
712 0 : }
713 : }
714 :
715 : enum CreateTimelineCause {
716 : Load,
717 : Delete,
718 : }
719 :
720 0 : #[derive(thiserror::Error, Debug)]
721 : pub(crate) enum GcError {
722 : // The tenant is shutting down
723 : #[error("tenant shutting down")]
724 : TenantCancelled,
725 :
726 : // The tenant is shutting down
727 : #[error("timeline shutting down")]
728 : TimelineCancelled,
729 :
730 : // The tenant is in a state inelegible to run GC
731 : #[error("not active")]
732 : NotActive,
733 :
734 : // A requested GC cutoff LSN was invalid, for example it tried to move backwards
735 : #[error("not active")]
736 : BadLsn { why: String },
737 :
738 : // A remote storage error while scheduling updates after compaction
739 : #[error(transparent)]
740 : Remote(anyhow::Error),
741 :
742 : // An error reading while calculating GC cutoffs
743 : #[error(transparent)]
744 : GcCutoffs(PageReconstructError),
745 :
746 : // If GC was invoked for a particular timeline, this error means it didn't exist
747 : #[error("timeline not found")]
748 : TimelineNotFound,
749 : }
750 :
751 : impl From<PageReconstructError> for GcError {
752 0 : fn from(value: PageReconstructError) -> Self {
753 0 : match value {
754 0 : PageReconstructError::Cancelled => Self::TimelineCancelled,
755 0 : other => Self::GcCutoffs(other),
756 : }
757 0 : }
758 : }
759 :
760 : impl From<NotInitialized> for GcError {
761 0 : fn from(value: NotInitialized) -> Self {
762 0 : match value {
763 0 : NotInitialized::Uninitialized => GcError::Remote(value.into()),
764 0 : NotInitialized::Stopped | NotInitialized::ShuttingDown => GcError::TimelineCancelled,
765 : }
766 0 : }
767 : }
768 :
769 : impl From<timeline::layer_manager::Shutdown> for GcError {
770 0 : fn from(_: timeline::layer_manager::Shutdown) -> Self {
771 0 : GcError::TimelineCancelled
772 0 : }
773 : }
774 :
775 0 : #[derive(thiserror::Error, Debug)]
776 : pub(crate) enum LoadConfigError {
777 : #[error("TOML deserialization error: '{0}'")]
778 : DeserializeToml(#[from] toml_edit::de::Error),
779 :
780 : #[error("Config not found at {0}")]
781 : NotFound(Utf8PathBuf),
782 : }
783 :
784 : impl Tenant {
785 : /// Yet another helper for timeline initialization.
786 : ///
787 : /// - Initializes the Timeline struct and inserts it into the tenant's hash map
788 : /// - Scans the local timeline directory for layer files and builds the layer map
789 : /// - Downloads remote index file and adds remote files to the layer map
790 : /// - Schedules remote upload tasks for any files that are present locally but missing from remote storage.
791 : ///
792 : /// If the operation fails, the timeline is left in the tenant's hash map in Broken state. On success,
793 : /// it is marked as Active.
794 : #[allow(clippy::too_many_arguments)]
795 6 : async fn timeline_init_and_sync(
796 6 : &self,
797 6 : timeline_id: TimelineId,
798 6 : resources: TimelineResources,
799 6 : index_part: Option<IndexPart>,
800 6 : metadata: TimelineMetadata,
801 6 : ancestor: Option<Arc<Timeline>>,
802 6 : _ctx: &RequestContext,
803 6 : ) -> anyhow::Result<()> {
804 6 : let tenant_id = self.tenant_shard_id;
805 :
806 6 : let timeline = self.create_timeline_struct(
807 6 : timeline_id,
808 6 : &metadata,
809 6 : ancestor.clone(),
810 6 : resources,
811 6 : CreateTimelineCause::Load,
812 6 : )?;
813 6 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
814 6 : anyhow::ensure!(
815 6 : disk_consistent_lsn.is_valid(),
816 0 : "Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn"
817 : );
818 6 : assert_eq!(
819 6 : disk_consistent_lsn,
820 6 : metadata.disk_consistent_lsn(),
821 0 : "these are used interchangeably"
822 : );
823 :
824 6 : if let Some(index_part) = index_part.as_ref() {
825 6 : timeline.remote_client.init_upload_queue(index_part)?;
826 : } else {
827 : // No data on the remote storage, but we have local metadata file. We can end up
828 : // here with timeline_create being interrupted before finishing index part upload.
829 : // By doing what we do here, the index part upload is retried.
830 : // If control plane retries timeline creation in the meantime, the mgmt API handler
831 : // for timeline creation will coalesce on the upload we queue here.
832 :
833 : // FIXME: this branch should be dead code as we no longer write local metadata.
834 :
835 0 : timeline
836 0 : .remote_client
837 0 : .init_upload_queue_for_empty_remote(&metadata)?;
838 0 : timeline
839 0 : .remote_client
840 0 : .schedule_index_upload_for_full_metadata_update(&metadata)?;
841 : }
842 :
843 6 : timeline
844 6 : .load_layer_map(disk_consistent_lsn, index_part)
845 6 : .await
846 6 : .with_context(|| {
847 0 : format!("Failed to load layermap for timeline {tenant_id}/{timeline_id}")
848 6 : })?;
849 :
850 : {
851 : // avoiding holding it across awaits
852 6 : let mut timelines_accessor = self.timelines.lock().unwrap();
853 6 : match timelines_accessor.entry(timeline_id) {
854 : // We should never try and load the same timeline twice during startup
855 : Entry::Occupied(_) => {
856 0 : unreachable!(
857 0 : "Timeline {tenant_id}/{timeline_id} already exists in the tenant map"
858 0 : );
859 : }
860 6 : Entry::Vacant(v) => {
861 6 : v.insert(Arc::clone(&timeline));
862 6 : timeline.maybe_spawn_flush_loop();
863 6 : }
864 6 : }
865 6 : };
866 6 :
867 6 : // Sanity check: a timeline should have some content.
868 6 : anyhow::ensure!(
869 6 : ancestor.is_some()
870 4 : || timeline
871 4 : .layers
872 4 : .read()
873 0 : .await
874 4 : .layer_map()
875 4 : .expect("currently loading, layer manager cannot be shutdown already")
876 4 : .iter_historic_layers()
877 4 : .next()
878 4 : .is_some(),
879 0 : "Timeline has no ancestor and no layer files"
880 : );
881 :
882 6 : Ok(())
883 6 : }
884 :
885 : /// Attach a tenant that's available in cloud storage.
886 : ///
887 : /// This returns quickly, after just creating the in-memory object
888 : /// Tenant struct and launching a background task to download
889 : /// the remote index files. On return, the tenant is most likely still in
890 : /// Attaching state, and it will become Active once the background task
891 : /// finishes. You can use wait_until_active() to wait for the task to
892 : /// complete.
893 : ///
894 : #[allow(clippy::too_many_arguments)]
895 0 : pub(crate) fn spawn(
896 0 : conf: &'static PageServerConf,
897 0 : tenant_shard_id: TenantShardId,
898 0 : resources: TenantSharedResources,
899 0 : attached_conf: AttachedTenantConf,
900 0 : shard_identity: ShardIdentity,
901 0 : init_order: Option<InitializationOrder>,
902 0 : mode: SpawnMode,
903 0 : ctx: &RequestContext,
904 0 : ) -> Result<Arc<Tenant>, GlobalShutDown> {
905 0 : let wal_redo_manager =
906 0 : WalRedoManager::new(PostgresRedoManager::new(conf, tenant_shard_id))?;
907 :
908 : let TenantSharedResources {
909 0 : broker_client,
910 0 : remote_storage,
911 0 : deletion_queue_client,
912 0 : l0_flush_global_state,
913 0 : } = resources;
914 0 :
915 0 : let attach_mode = attached_conf.location.attach_mode;
916 0 : let generation = attached_conf.location.generation;
917 0 :
918 0 : let tenant = Arc::new(Tenant::new(
919 0 : TenantState::Attaching,
920 0 : conf,
921 0 : attached_conf,
922 0 : shard_identity,
923 0 : Some(wal_redo_manager),
924 0 : tenant_shard_id,
925 0 : remote_storage.clone(),
926 0 : deletion_queue_client,
927 0 : l0_flush_global_state,
928 0 : ));
929 0 :
930 0 : // The attach task will carry a GateGuard, so that shutdown() reliably waits for it to drop out if
931 0 : // we shut down while attaching.
932 0 : let attach_gate_guard = tenant
933 0 : .gate
934 0 : .enter()
935 0 : .expect("We just created the Tenant: nothing else can have shut it down yet");
936 0 :
937 0 : // Do all the hard work in the background
938 0 : let tenant_clone = Arc::clone(&tenant);
939 0 : let ctx = ctx.detached_child(TaskKind::Attach, DownloadBehavior::Warn);
940 0 : task_mgr::spawn(
941 0 : &tokio::runtime::Handle::current(),
942 0 : TaskKind::Attach,
943 0 : tenant_shard_id,
944 0 : None,
945 0 : "attach tenant",
946 0 : async move {
947 0 :
948 0 : info!(
949 : ?attach_mode,
950 0 : "Attaching tenant"
951 : );
952 :
953 0 : let _gate_guard = attach_gate_guard;
954 0 :
955 0 : // Is this tenant being spawned as part of process startup?
956 0 : let starting_up = init_order.is_some();
957 0 : scopeguard::defer! {
958 0 : if starting_up {
959 0 : TENANT.startup_complete.inc();
960 0 : }
961 0 : }
962 :
963 : // Ideally we should use Tenant::set_broken_no_wait, but it is not supposed to be used when tenant is in loading state.
964 : enum BrokenVerbosity {
965 : Error,
966 : Info
967 : }
968 0 : let make_broken =
969 0 : |t: &Tenant, err: anyhow::Error, verbosity: BrokenVerbosity| {
970 0 : match verbosity {
971 : BrokenVerbosity::Info => {
972 0 : info!("attach cancelled, setting tenant state to Broken: {err}");
973 : },
974 : BrokenVerbosity::Error => {
975 0 : error!("attach failed, setting tenant state to Broken: {err:?}");
976 : }
977 : }
978 0 : t.state.send_modify(|state| {
979 0 : // The Stopping case is for when we have passed control on to DeleteTenantFlow:
980 0 : // if it errors, we will call make_broken when tenant is already in Stopping.
981 0 : assert!(
982 0 : matches!(*state, TenantState::Attaching | TenantState::Stopping { .. }),
983 0 : "the attach task owns the tenant state until activation is complete"
984 : );
985 :
986 0 : *state = TenantState::broken_from_reason(err.to_string());
987 0 : });
988 0 : };
989 :
990 : // TODO: should also be rejecting tenant conf changes that violate this check.
991 0 : if let Err(e) = crate::tenant::storage_layer::inmemory_layer::IndexEntry::validate_checkpoint_distance(tenant_clone.get_checkpoint_distance()) {
992 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
993 0 : return Ok(());
994 0 : }
995 0 :
996 0 : let mut init_order = init_order;
997 0 : // take the completion because initial tenant loading will complete when all of
998 0 : // these tasks complete.
999 0 : let _completion = init_order
1000 0 : .as_mut()
1001 0 : .and_then(|x| x.initial_tenant_load.take());
1002 0 : let remote_load_completion = init_order
1003 0 : .as_mut()
1004 0 : .and_then(|x| x.initial_tenant_load_remote.take());
1005 :
1006 : enum AttachType<'a> {
1007 : /// We are attaching this tenant lazily in the background.
1008 : Warmup {
1009 : _permit: tokio::sync::SemaphorePermit<'a>,
1010 : during_startup: bool
1011 : },
1012 : /// We are attaching this tenant as soon as we can, because for example an
1013 : /// endpoint tried to access it.
1014 : OnDemand,
1015 : /// During normal operations after startup, we are attaching a tenant, and
1016 : /// eager attach was requested.
1017 : Normal,
1018 : }
1019 :
1020 0 : let attach_type = if matches!(mode, SpawnMode::Lazy) {
1021 : // Before doing any I/O, wait for at least one of:
1022 : // - A client attempting to access to this tenant (on-demand loading)
1023 : // - A permit becoming available in the warmup semaphore (background warmup)
1024 :
1025 0 : tokio::select!(
1026 0 : permit = tenant_clone.activate_now_sem.acquire() => {
1027 0 : let _ = permit.expect("activate_now_sem is never closed");
1028 0 : tracing::info!("Activating tenant (on-demand)");
1029 0 : AttachType::OnDemand
1030 : },
1031 0 : permit = conf.concurrent_tenant_warmup.inner().acquire() => {
1032 0 : let _permit = permit.expect("concurrent_tenant_warmup semaphore is never closed");
1033 0 : tracing::info!("Activating tenant (warmup)");
1034 0 : AttachType::Warmup {
1035 0 : _permit,
1036 0 : during_startup: init_order.is_some()
1037 0 : }
1038 : }
1039 0 : _ = tenant_clone.cancel.cancelled() => {
1040 : // This is safe, but should be pretty rare: it is interesting if a tenant
1041 : // stayed in Activating for such a long time that shutdown found it in
1042 : // that state.
1043 0 : tracing::info!(state=%tenant_clone.current_state(), "Tenant shut down before activation");
1044 : // Make the tenant broken so that set_stopping will not hang waiting for it to leave
1045 : // the Attaching state. This is an over-reaction (nothing really broke, the tenant is
1046 : // just shutting down), but ensures progress.
1047 0 : make_broken(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"), BrokenVerbosity::Info);
1048 0 : return Ok(());
1049 : },
1050 : )
1051 : } else {
1052 : // SpawnMode::{Create,Eager} always cause jumping ahead of the
1053 : // concurrent_tenant_warmup queue
1054 0 : AttachType::Normal
1055 : };
1056 :
1057 0 : let preload = match &mode {
1058 : SpawnMode::Eager | SpawnMode::Lazy => {
1059 0 : let _preload_timer = TENANT.preload.start_timer();
1060 0 : let res = tenant_clone
1061 0 : .preload(&remote_storage, task_mgr::shutdown_token())
1062 0 : .await;
1063 0 : match res {
1064 0 : Ok(p) => Some(p),
1065 0 : Err(e) => {
1066 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1067 0 : return Ok(());
1068 : }
1069 : }
1070 : }
1071 :
1072 : };
1073 :
1074 : // Remote preload is complete.
1075 0 : drop(remote_load_completion);
1076 0 :
1077 0 :
1078 0 : // We will time the duration of the attach phase unless this is a creation (attach will do no work)
1079 0 : let attach_start = std::time::Instant::now();
1080 0 : let attached = {
1081 0 : let _attach_timer = Some(TENANT.attach.start_timer());
1082 0 : tenant_clone.attach(preload, &ctx).await
1083 : };
1084 0 : let attach_duration = attach_start.elapsed();
1085 0 : _ = tenant_clone.attach_wal_lag_cooldown.set(WalLagCooldown::new(attach_start, attach_duration));
1086 0 :
1087 0 : match attached {
1088 : Ok(()) => {
1089 0 : info!("attach finished, activating");
1090 0 : tenant_clone.activate(broker_client, None, &ctx);
1091 : }
1092 0 : Err(e) => {
1093 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1094 0 : }
1095 : }
1096 :
1097 : // If we are doing an opportunistic warmup attachment at startup, initialize
1098 : // logical size at the same time. This is better than starting a bunch of idle tenants
1099 : // with cold caches and then coming back later to initialize their logical sizes.
1100 : //
1101 : // It also prevents the warmup proccess competing with the concurrency limit on
1102 : // logical size calculations: if logical size calculation semaphore is saturated,
1103 : // then warmup will wait for that before proceeding to the next tenant.
1104 0 : if matches!(attach_type, AttachType::Warmup { during_startup: true, .. }) {
1105 0 : let mut futs: FuturesUnordered<_> = tenant_clone.timelines.lock().unwrap().values().cloned().map(|t| t.await_initial_logical_size()).collect();
1106 0 : tracing::info!("Waiting for initial logical sizes while warming up...");
1107 0 : while futs.next().await.is_some() {}
1108 0 : tracing::info!("Warm-up complete");
1109 0 : }
1110 :
1111 0 : Ok(())
1112 0 : }
1113 0 : .instrument(tracing::info_span!(parent: None, "attach", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), gen=?generation)),
1114 : );
1115 0 : Ok(tenant)
1116 0 : }
1117 :
1118 186 : #[instrument(skip_all)]
1119 : pub(crate) async fn preload(
1120 : self: &Arc<Self>,
1121 : remote_storage: &GenericRemoteStorage,
1122 : cancel: CancellationToken,
1123 : ) -> anyhow::Result<TenantPreload> {
1124 : span::debug_assert_current_span_has_tenant_id();
1125 : // Get list of remote timelines
1126 : // download index files for every tenant timeline
1127 : info!("listing remote timelines");
1128 : let (remote_timeline_ids, other_keys) = remote_timeline_client::list_remote_timelines(
1129 : remote_storage,
1130 : self.tenant_shard_id,
1131 : cancel.clone(),
1132 : )
1133 : .await?;
1134 :
1135 : info!("found {} timelines", remote_timeline_ids.len(),);
1136 :
1137 : for k in other_keys {
1138 : warn!("Unexpected non timeline key {k}");
1139 : }
1140 :
1141 : Ok(TenantPreload {
1142 : timelines: self
1143 : .load_timelines_metadata(remote_timeline_ids, remote_storage, cancel)
1144 : .await?,
1145 : })
1146 : }
1147 :
1148 : ///
1149 : /// Background task that downloads all data for a tenant and brings it to Active state.
1150 : ///
1151 : /// No background tasks are started as part of this routine.
1152 : ///
1153 186 : async fn attach(
1154 186 : self: &Arc<Tenant>,
1155 186 : preload: Option<TenantPreload>,
1156 186 : ctx: &RequestContext,
1157 186 : ) -> anyhow::Result<()> {
1158 186 : span::debug_assert_current_span_has_tenant_id();
1159 186 :
1160 186 : failpoint_support::sleep_millis_async!("before-attaching-tenant");
1161 :
1162 186 : let Some(preload) = preload else {
1163 0 : anyhow::bail!("local-only deployment is no longer supported, https://github.com/neondatabase/neon/issues/5624");
1164 : };
1165 :
1166 186 : let mut timelines_to_resume_deletions = vec![];
1167 186 :
1168 186 : let mut remote_index_and_client = HashMap::new();
1169 186 : let mut timeline_ancestors = HashMap::new();
1170 186 : let mut existent_timelines = HashSet::new();
1171 192 : for (timeline_id, preload) in preload.timelines {
1172 6 : let index_part = match preload.index_part {
1173 6 : Ok(i) => {
1174 6 : debug!("remote index part exists for timeline {timeline_id}");
1175 : // We found index_part on the remote, this is the standard case.
1176 6 : existent_timelines.insert(timeline_id);
1177 6 : i
1178 : }
1179 : Err(DownloadError::NotFound) => {
1180 : // There is no index_part on the remote. We only get here
1181 : // if there is some prefix for the timeline in the remote storage.
1182 : // This can e.g. be the initdb.tar.zst archive, maybe a
1183 : // remnant from a prior incomplete creation or deletion attempt.
1184 : // Delete the local directory as the deciding criterion for a
1185 : // timeline's existence is presence of index_part.
1186 0 : info!(%timeline_id, "index_part not found on remote");
1187 0 : continue;
1188 : }
1189 0 : Err(e) => {
1190 0 : // Some (possibly ephemeral) error happened during index_part download.
1191 0 : // Pretend the timeline exists to not delete the timeline directory,
1192 0 : // as it might be a temporary issue and we don't want to re-download
1193 0 : // everything after it resolves.
1194 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1195 :
1196 0 : existent_timelines.insert(timeline_id);
1197 0 : continue;
1198 : }
1199 : };
1200 6 : match index_part {
1201 6 : MaybeDeletedIndexPart::IndexPart(index_part) => {
1202 6 : timeline_ancestors.insert(timeline_id, index_part.metadata.clone());
1203 6 : remote_index_and_client.insert(timeline_id, (index_part, preload.client));
1204 6 : }
1205 0 : MaybeDeletedIndexPart::Deleted(index_part) => {
1206 0 : info!(
1207 0 : "timeline {} is deleted, picking to resume deletion",
1208 : timeline_id
1209 : );
1210 0 : timelines_to_resume_deletions.push((timeline_id, index_part, preload.client));
1211 : }
1212 : }
1213 : }
1214 :
1215 186 : let mut gc_blocks = HashMap::new();
1216 :
1217 : // For every timeline, download the metadata file, scan the local directory,
1218 : // and build a layer map that contains an entry for each remote and local
1219 : // layer file.
1220 186 : let sorted_timelines = tree_sort_timelines(timeline_ancestors, |m| m.ancestor_timeline())?;
1221 192 : for (timeline_id, remote_metadata) in sorted_timelines {
1222 6 : let (index_part, remote_client) = remote_index_and_client
1223 6 : .remove(&timeline_id)
1224 6 : .expect("just put it in above");
1225 :
1226 6 : if let Some(blocking) = index_part.gc_blocking.as_ref() {
1227 : // could just filter these away, but it helps while testing
1228 0 : anyhow::ensure!(
1229 0 : !blocking.reasons.is_empty(),
1230 0 : "index_part for {timeline_id} is malformed: it should not have gc blocking with zero reasons"
1231 : );
1232 0 : let prev = gc_blocks.insert(timeline_id, blocking.reasons);
1233 0 : assert!(prev.is_none());
1234 6 : }
1235 :
1236 : // TODO again handle early failure
1237 6 : self.load_remote_timeline(
1238 6 : timeline_id,
1239 6 : index_part,
1240 6 : remote_metadata,
1241 6 : TimelineResources {
1242 6 : remote_client,
1243 6 : timeline_get_throttle: self.timeline_get_throttle.clone(),
1244 6 : l0_flush_global_state: self.l0_flush_global_state.clone(),
1245 6 : },
1246 6 : ctx,
1247 6 : )
1248 10 : .await
1249 6 : .with_context(|| {
1250 0 : format!(
1251 0 : "failed to load remote timeline {} for tenant {}",
1252 0 : timeline_id, self.tenant_shard_id
1253 0 : )
1254 6 : })?;
1255 : }
1256 :
1257 : // Walk through deleted timelines, resume deletion
1258 186 : for (timeline_id, index_part, remote_timeline_client) in timelines_to_resume_deletions {
1259 0 : remote_timeline_client
1260 0 : .init_upload_queue_stopped_to_continue_deletion(&index_part)
1261 0 : .context("init queue stopped")
1262 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1263 :
1264 0 : DeleteTimelineFlow::resume_deletion(
1265 0 : Arc::clone(self),
1266 0 : timeline_id,
1267 0 : &index_part.metadata,
1268 0 : remote_timeline_client,
1269 0 : )
1270 0 : .instrument(tracing::info_span!("timeline_delete", %timeline_id))
1271 0 : .await
1272 0 : .context("resume_deletion")
1273 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1274 : }
1275 :
1276 : // The local filesystem contents are a cache of what's in the remote IndexPart;
1277 : // IndexPart is the source of truth.
1278 186 : self.clean_up_timelines(&existent_timelines)?;
1279 :
1280 186 : self.gc_block.set_scanned(gc_blocks);
1281 186 :
1282 186 : fail::fail_point!("attach-before-activate", |_| {
1283 0 : anyhow::bail!("attach-before-activate");
1284 186 : });
1285 186 : failpoint_support::sleep_millis_async!("attach-before-activate-sleep", &self.cancel);
1286 :
1287 186 : info!("Done");
1288 :
1289 186 : Ok(())
1290 186 : }
1291 :
1292 : /// Check for any local timeline directories that are temporary, or do not correspond to a
1293 : /// timeline that still exists: this can happen if we crashed during a deletion/creation, or
1294 : /// if a timeline was deleted while the tenant was attached to a different pageserver.
1295 186 : fn clean_up_timelines(&self, existent_timelines: &HashSet<TimelineId>) -> anyhow::Result<()> {
1296 186 : let timelines_dir = self.conf.timelines_path(&self.tenant_shard_id);
1297 :
1298 186 : let entries = match timelines_dir.read_dir_utf8() {
1299 186 : Ok(d) => d,
1300 0 : Err(e) => {
1301 0 : if e.kind() == std::io::ErrorKind::NotFound {
1302 0 : return Ok(());
1303 : } else {
1304 0 : return Err(e).context("list timelines directory for tenant");
1305 : }
1306 : }
1307 : };
1308 :
1309 194 : for entry in entries {
1310 8 : let entry = entry.context("read timeline dir entry")?;
1311 8 : let entry_path = entry.path();
1312 :
1313 8 : let purge = if crate::is_temporary(entry_path)
1314 : // TODO: remove uninit mark code (https://github.com/neondatabase/neon/issues/5718)
1315 8 : || is_uninit_mark(entry_path)
1316 8 : || crate::is_delete_mark(entry_path)
1317 : {
1318 0 : true
1319 : } else {
1320 8 : match TimelineId::try_from(entry_path.file_name()) {
1321 8 : Ok(i) => {
1322 8 : // Purge if the timeline ID does not exist in remote storage: remote storage is the authority.
1323 8 : !existent_timelines.contains(&i)
1324 : }
1325 0 : Err(e) => {
1326 0 : tracing::warn!(
1327 0 : "Unparseable directory in timelines directory: {entry_path}, ignoring ({e})"
1328 : );
1329 : // Do not purge junk: if we don't recognize it, be cautious and leave it for a human.
1330 0 : false
1331 : }
1332 : }
1333 : };
1334 :
1335 8 : if purge {
1336 2 : tracing::info!("Purging stale timeline dentry {entry_path}");
1337 2 : if let Err(e) = match entry.file_type() {
1338 2 : Ok(t) => if t.is_dir() {
1339 2 : std::fs::remove_dir_all(entry_path)
1340 : } else {
1341 0 : std::fs::remove_file(entry_path)
1342 : }
1343 2 : .or_else(fs_ext::ignore_not_found),
1344 0 : Err(e) => Err(e),
1345 : } {
1346 0 : tracing::warn!("Failed to purge stale timeline dentry {entry_path}: {e}");
1347 2 : }
1348 6 : }
1349 : }
1350 :
1351 186 : Ok(())
1352 186 : }
1353 :
1354 : /// Get sum of all remote timelines sizes
1355 : ///
1356 : /// This function relies on the index_part instead of listing the remote storage
1357 0 : pub fn remote_size(&self) -> u64 {
1358 0 : let mut size = 0;
1359 :
1360 0 : for timeline in self.list_timelines() {
1361 0 : size += timeline.remote_client.get_remote_physical_size();
1362 0 : }
1363 :
1364 0 : size
1365 0 : }
1366 :
1367 6 : #[instrument(skip_all, fields(timeline_id=%timeline_id))]
1368 : async fn load_remote_timeline(
1369 : &self,
1370 : timeline_id: TimelineId,
1371 : index_part: IndexPart,
1372 : remote_metadata: TimelineMetadata,
1373 : resources: TimelineResources,
1374 : ctx: &RequestContext,
1375 : ) -> anyhow::Result<()> {
1376 : span::debug_assert_current_span_has_tenant_id();
1377 :
1378 : info!("downloading index file for timeline {}", timeline_id);
1379 : tokio::fs::create_dir_all(self.conf.timeline_path(&self.tenant_shard_id, &timeline_id))
1380 : .await
1381 : .context("Failed to create new timeline directory")?;
1382 :
1383 : let ancestor = if let Some(ancestor_id) = remote_metadata.ancestor_timeline() {
1384 : let timelines = self.timelines.lock().unwrap();
1385 : Some(Arc::clone(timelines.get(&ancestor_id).ok_or_else(
1386 0 : || {
1387 0 : anyhow::anyhow!(
1388 0 : "cannot find ancestor timeline {ancestor_id} for timeline {timeline_id}"
1389 0 : )
1390 0 : },
1391 : )?))
1392 : } else {
1393 : None
1394 : };
1395 :
1396 : self.timeline_init_and_sync(
1397 : timeline_id,
1398 : resources,
1399 : Some(index_part),
1400 : remote_metadata,
1401 : ancestor,
1402 : ctx,
1403 : )
1404 : .await
1405 : }
1406 :
1407 186 : async fn load_timelines_metadata(
1408 186 : self: &Arc<Tenant>,
1409 186 : timeline_ids: HashSet<TimelineId>,
1410 186 : remote_storage: &GenericRemoteStorage,
1411 186 : cancel: CancellationToken,
1412 186 : ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
1413 186 : let mut part_downloads = JoinSet::new();
1414 192 : for timeline_id in timeline_ids {
1415 6 : let cancel_clone = cancel.clone();
1416 6 : part_downloads.spawn(
1417 6 : self.load_timeline_metadata(timeline_id, remote_storage.clone(), cancel_clone)
1418 6 : .instrument(info_span!("download_index_part", %timeline_id)),
1419 : );
1420 : }
1421 :
1422 186 : let mut timeline_preloads: HashMap<TimelineId, TimelinePreload> = HashMap::new();
1423 :
1424 : loop {
1425 192 : tokio::select!(
1426 192 : next = part_downloads.join_next() => {
1427 192 : match next {
1428 6 : Some(result) => {
1429 6 : let preload = result.context("join preload task")?;
1430 6 : timeline_preloads.insert(preload.timeline_id, preload);
1431 : },
1432 : None => {
1433 186 : break;
1434 : }
1435 : }
1436 : },
1437 192 : _ = cancel.cancelled() => {
1438 0 : anyhow::bail!("Cancelled while waiting for remote index download")
1439 : }
1440 : )
1441 : }
1442 :
1443 186 : Ok(timeline_preloads)
1444 186 : }
1445 :
1446 6 : fn load_timeline_metadata(
1447 6 : self: &Arc<Tenant>,
1448 6 : timeline_id: TimelineId,
1449 6 : remote_storage: GenericRemoteStorage,
1450 6 : cancel: CancellationToken,
1451 6 : ) -> impl Future<Output = TimelinePreload> {
1452 6 : let client = RemoteTimelineClient::new(
1453 6 : remote_storage.clone(),
1454 6 : self.deletion_queue_client.clone(),
1455 6 : self.conf,
1456 6 : self.tenant_shard_id,
1457 6 : timeline_id,
1458 6 : self.generation,
1459 6 : );
1460 6 : async move {
1461 6 : debug_assert_current_span_has_tenant_and_timeline_id();
1462 6 : debug!("starting index part download");
1463 :
1464 11 : let index_part = client.download_index_file(&cancel).await;
1465 :
1466 6 : debug!("finished index part download");
1467 :
1468 6 : TimelinePreload {
1469 6 : client,
1470 6 : timeline_id,
1471 6 : index_part,
1472 6 : }
1473 6 : }
1474 6 : }
1475 :
1476 0 : fn check_to_be_archived_has_no_unarchived_children(
1477 0 : timeline_id: TimelineId,
1478 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
1479 0 : ) -> Result<(), TimelineArchivalError> {
1480 0 : let children: Vec<TimelineId> = timelines
1481 0 : .iter()
1482 0 : .filter_map(|(id, entry)| {
1483 0 : if entry.get_ancestor_timeline_id() != Some(timeline_id) {
1484 0 : return None;
1485 0 : }
1486 0 : if entry.is_archived() == Some(true) {
1487 0 : return None;
1488 0 : }
1489 0 : Some(*id)
1490 0 : })
1491 0 : .collect();
1492 0 :
1493 0 : if !children.is_empty() {
1494 0 : return Err(TimelineArchivalError::HasUnarchivedChildren(children));
1495 0 : }
1496 0 : Ok(())
1497 0 : }
1498 :
1499 0 : fn check_ancestor_of_to_be_unarchived_is_not_archived(
1500 0 : ancestor_timeline_id: TimelineId,
1501 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
1502 0 : offloaded_timelines: &std::sync::MutexGuard<
1503 0 : '_,
1504 0 : HashMap<TimelineId, Arc<OffloadedTimeline>>,
1505 0 : >,
1506 0 : ) -> Result<(), TimelineArchivalError> {
1507 0 : let has_archived_parent =
1508 0 : if let Some(ancestor_timeline) = timelines.get(&ancestor_timeline_id) {
1509 0 : ancestor_timeline.is_archived() == Some(true)
1510 0 : } else if offloaded_timelines.contains_key(&ancestor_timeline_id) {
1511 0 : true
1512 : } else {
1513 0 : error!("ancestor timeline {ancestor_timeline_id} not found");
1514 0 : if cfg!(debug_assertions) {
1515 0 : panic!("ancestor timeline {ancestor_timeline_id} not found");
1516 0 : }
1517 0 : return Err(TimelineArchivalError::NotFound);
1518 : };
1519 0 : if has_archived_parent {
1520 0 : return Err(TimelineArchivalError::HasArchivedParent(
1521 0 : ancestor_timeline_id,
1522 0 : ));
1523 0 : }
1524 0 : Ok(())
1525 0 : }
1526 :
1527 0 : fn check_to_be_unarchived_timeline_has_no_archived_parent(
1528 0 : timeline: &Arc<Timeline>,
1529 0 : ) -> Result<(), TimelineArchivalError> {
1530 0 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
1531 0 : if ancestor_timeline.is_archived() == Some(true) {
1532 0 : return Err(TimelineArchivalError::HasArchivedParent(
1533 0 : ancestor_timeline.timeline_id,
1534 0 : ));
1535 0 : }
1536 0 : }
1537 0 : Ok(())
1538 0 : }
1539 :
1540 : /// Loads the specified (offloaded) timeline from S3 and attaches it as a loaded timeline
1541 0 : async fn unoffload_timeline(
1542 0 : self: &Arc<Self>,
1543 0 : timeline_id: TimelineId,
1544 0 : broker_client: storage_broker::BrokerClientChannel,
1545 0 : ctx: RequestContext,
1546 0 : ) -> Result<Arc<Timeline>, TimelineArchivalError> {
1547 0 : info!("unoffloading timeline");
1548 0 : let cancel = self.cancel.clone();
1549 0 : let timeline_preload = self
1550 0 : .load_timeline_metadata(timeline_id, self.remote_storage.clone(), cancel)
1551 0 : .await;
1552 :
1553 0 : let index_part = match timeline_preload.index_part {
1554 0 : Ok(index_part) => {
1555 0 : debug!("remote index part exists for timeline {timeline_id}");
1556 0 : index_part
1557 : }
1558 : Err(DownloadError::NotFound) => {
1559 0 : error!(%timeline_id, "index_part not found on remote");
1560 0 : return Err(TimelineArchivalError::NotFound);
1561 : }
1562 0 : Err(DownloadError::Cancelled) => return Err(TimelineArchivalError::Cancelled),
1563 0 : Err(e) => {
1564 0 : // Some (possibly ephemeral) error happened during index_part download.
1565 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1566 0 : return Err(TimelineArchivalError::Other(
1567 0 : anyhow::Error::new(e).context("downloading index_part from remote storage"),
1568 0 : ));
1569 : }
1570 : };
1571 0 : let index_part = match index_part {
1572 0 : MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
1573 0 : MaybeDeletedIndexPart::Deleted(_index_part) => {
1574 0 : info!("timeline is deleted according to index_part.json");
1575 0 : return Err(TimelineArchivalError::NotFound);
1576 : }
1577 : };
1578 0 : let remote_metadata = index_part.metadata.clone();
1579 0 : let timeline_resources = self.build_timeline_resources(timeline_id);
1580 0 : self.load_remote_timeline(
1581 0 : timeline_id,
1582 0 : index_part,
1583 0 : remote_metadata,
1584 0 : timeline_resources,
1585 0 : &ctx,
1586 0 : )
1587 0 : .await
1588 0 : .with_context(|| {
1589 0 : format!(
1590 0 : "failed to load remote timeline {} for tenant {}",
1591 0 : timeline_id, self.tenant_shard_id
1592 0 : )
1593 0 : })
1594 0 : .map_err(TimelineArchivalError::Other)?;
1595 0 : let timelines = self.timelines.lock().unwrap();
1596 0 : let Some(timeline) = timelines.get(&timeline_id) else {
1597 0 : warn!("timeline not available directly after attach");
1598 0 : return Err(TimelineArchivalError::Other(anyhow::anyhow!(
1599 0 : "timeline not available directly after attach"
1600 0 : )));
1601 : };
1602 0 : let mut offloaded_timelines = self.timelines_offloaded.lock().unwrap();
1603 0 : if offloaded_timelines.remove(&timeline_id).is_none() {
1604 0 : warn!("timeline already removed from offloaded timelines");
1605 0 : }
1606 :
1607 : // Activate the timeline (if it makes sense)
1608 0 : if !(timeline.is_broken() || timeline.is_stopping()) {
1609 0 : let background_jobs_can_start = None;
1610 0 : timeline.activate(
1611 0 : self.clone(),
1612 0 : broker_client.clone(),
1613 0 : background_jobs_can_start,
1614 0 : &ctx,
1615 0 : );
1616 0 : }
1617 :
1618 0 : info!("timeline unoffloading complete");
1619 0 : Ok(Arc::clone(timeline))
1620 0 : }
1621 :
1622 0 : pub(crate) async fn apply_timeline_archival_config(
1623 0 : self: &Arc<Self>,
1624 0 : timeline_id: TimelineId,
1625 0 : new_state: TimelineArchivalState,
1626 0 : broker_client: storage_broker::BrokerClientChannel,
1627 0 : ctx: RequestContext,
1628 0 : ) -> Result<(), TimelineArchivalError> {
1629 0 : info!("setting timeline archival config");
1630 : // First part: figure out what is needed to do, and do validation
1631 0 : let timeline_or_unarchive_offloaded = 'outer: {
1632 0 : let timelines = self.timelines.lock().unwrap();
1633 :
1634 0 : let Some(timeline) = timelines.get(&timeline_id) else {
1635 0 : let offloaded_timelines = self.timelines_offloaded.lock().unwrap();
1636 0 : let Some(offloaded) = offloaded_timelines.get(&timeline_id) else {
1637 0 : return Err(TimelineArchivalError::NotFound);
1638 : };
1639 0 : if new_state == TimelineArchivalState::Archived {
1640 : // It's offloaded already, so nothing to do
1641 0 : return Ok(());
1642 0 : }
1643 0 : if let Some(ancestor_timeline_id) = offloaded.ancestor_timeline_id {
1644 0 : Self::check_ancestor_of_to_be_unarchived_is_not_archived(
1645 0 : ancestor_timeline_id,
1646 0 : &timelines,
1647 0 : &offloaded_timelines,
1648 0 : )?;
1649 0 : }
1650 0 : break 'outer None;
1651 : };
1652 :
1653 : // Do some validation. We release the timelines lock below, so there is potential
1654 : // for race conditions: these checks are more present to prevent misunderstandings of
1655 : // the API's capabilities, instead of serving as the sole way to defend their invariants.
1656 0 : match new_state {
1657 : TimelineArchivalState::Unarchived => {
1658 0 : Self::check_to_be_unarchived_timeline_has_no_archived_parent(timeline)?
1659 : }
1660 : TimelineArchivalState::Archived => {
1661 0 : Self::check_to_be_archived_has_no_unarchived_children(timeline_id, &timelines)?
1662 : }
1663 : }
1664 0 : Some(Arc::clone(timeline))
1665 : };
1666 :
1667 : // Second part: unoffload timeline (if needed)
1668 0 : let timeline = if let Some(timeline) = timeline_or_unarchive_offloaded {
1669 0 : timeline
1670 : } else {
1671 : // Turn offloaded timeline into a non-offloaded one
1672 0 : self.unoffload_timeline(timeline_id, broker_client, ctx)
1673 0 : .await?
1674 : };
1675 :
1676 : // Third part: upload new timeline archival state and block until it is present in S3
1677 0 : let upload_needed = match timeline
1678 0 : .remote_client
1679 0 : .schedule_index_upload_for_timeline_archival_state(new_state)
1680 : {
1681 0 : Ok(upload_needed) => upload_needed,
1682 0 : Err(e) => {
1683 0 : if timeline.cancel.is_cancelled() {
1684 0 : return Err(TimelineArchivalError::Cancelled);
1685 : } else {
1686 0 : return Err(TimelineArchivalError::Other(e));
1687 : }
1688 : }
1689 : };
1690 :
1691 0 : if upload_needed {
1692 0 : info!("Uploading new state");
1693 : const MAX_WAIT: Duration = Duration::from_secs(10);
1694 0 : let Ok(v) =
1695 0 : tokio::time::timeout(MAX_WAIT, timeline.remote_client.wait_completion()).await
1696 : else {
1697 0 : tracing::warn!("reached timeout for waiting on upload queue");
1698 0 : return Err(TimelineArchivalError::Timeout);
1699 : };
1700 0 : v.map_err(|e| match e {
1701 0 : WaitCompletionError::NotInitialized(e) => {
1702 0 : TimelineArchivalError::Other(anyhow::anyhow!(e))
1703 : }
1704 : WaitCompletionError::UploadQueueShutDownOrStopped => {
1705 0 : TimelineArchivalError::Cancelled
1706 : }
1707 0 : })?;
1708 0 : }
1709 0 : Ok(())
1710 0 : }
1711 :
1712 0 : pub fn get_offloaded_timeline(
1713 0 : &self,
1714 0 : timeline_id: TimelineId,
1715 0 : ) -> Result<Arc<OffloadedTimeline>, GetTimelineError> {
1716 0 : self.timelines_offloaded
1717 0 : .lock()
1718 0 : .unwrap()
1719 0 : .get(&timeline_id)
1720 0 : .map(Arc::clone)
1721 0 : .ok_or(GetTimelineError::NotFound {
1722 0 : tenant_id: self.tenant_shard_id,
1723 0 : timeline_id,
1724 0 : })
1725 0 : }
1726 :
1727 4 : pub(crate) fn tenant_shard_id(&self) -> TenantShardId {
1728 4 : self.tenant_shard_id
1729 4 : }
1730 :
1731 : /// Get Timeline handle for given Neon timeline ID.
1732 : /// This function is idempotent. It doesn't change internal state in any way.
1733 222 : pub fn get_timeline(
1734 222 : &self,
1735 222 : timeline_id: TimelineId,
1736 222 : active_only: bool,
1737 222 : ) -> Result<Arc<Timeline>, GetTimelineError> {
1738 222 : let timelines_accessor = self.timelines.lock().unwrap();
1739 222 : let timeline = timelines_accessor
1740 222 : .get(&timeline_id)
1741 222 : .ok_or(GetTimelineError::NotFound {
1742 222 : tenant_id: self.tenant_shard_id,
1743 222 : timeline_id,
1744 222 : })?;
1745 :
1746 220 : if active_only && !timeline.is_active() {
1747 0 : Err(GetTimelineError::NotActive {
1748 0 : tenant_id: self.tenant_shard_id,
1749 0 : timeline_id,
1750 0 : state: timeline.current_state(),
1751 0 : })
1752 : } else {
1753 220 : Ok(Arc::clone(timeline))
1754 : }
1755 222 : }
1756 :
1757 : /// Lists timelines the tenant contains.
1758 : /// It's up to callers to omit certain timelines that are not considered ready for use.
1759 0 : pub fn list_timelines(&self) -> Vec<Arc<Timeline>> {
1760 0 : self.timelines
1761 0 : .lock()
1762 0 : .unwrap()
1763 0 : .values()
1764 0 : .map(Arc::clone)
1765 0 : .collect()
1766 0 : }
1767 :
1768 : /// Lists timelines the tenant manages, including offloaded ones.
1769 : ///
1770 : /// It's up to callers to omit certain timelines that are not considered ready for use.
1771 0 : pub fn list_timelines_and_offloaded(
1772 0 : &self,
1773 0 : ) -> (Vec<Arc<Timeline>>, Vec<Arc<OffloadedTimeline>>) {
1774 0 : let timelines = self
1775 0 : .timelines
1776 0 : .lock()
1777 0 : .unwrap()
1778 0 : .values()
1779 0 : .map(Arc::clone)
1780 0 : .collect();
1781 0 : let offloaded = self
1782 0 : .timelines_offloaded
1783 0 : .lock()
1784 0 : .unwrap()
1785 0 : .values()
1786 0 : .map(Arc::clone)
1787 0 : .collect();
1788 0 : (timelines, offloaded)
1789 0 : }
1790 :
1791 0 : pub fn list_timeline_ids(&self) -> Vec<TimelineId> {
1792 0 : self.timelines.lock().unwrap().keys().cloned().collect()
1793 0 : }
1794 :
1795 : /// This is used to create the initial 'main' timeline during bootstrapping,
1796 : /// or when importing a new base backup. The caller is expected to load an
1797 : /// initial image of the datadir to the new timeline after this.
1798 : ///
1799 : /// Until that happens, the on-disk state is invalid (disk_consistent_lsn=Lsn(0))
1800 : /// and the timeline will fail to load at a restart.
1801 : ///
1802 : /// For tests, use `DatadirModification::init_empty_test_timeline` + `commit` to setup the
1803 : /// minimum amount of keys required to get a writable timeline.
1804 : /// (Without it, `put` might fail due to `repartition` failing.)
1805 178 : pub(crate) async fn create_empty_timeline(
1806 178 : &self,
1807 178 : new_timeline_id: TimelineId,
1808 178 : initdb_lsn: Lsn,
1809 178 : pg_version: u32,
1810 178 : _ctx: &RequestContext,
1811 178 : ) -> anyhow::Result<UninitializedTimeline> {
1812 178 : anyhow::ensure!(
1813 178 : self.is_active(),
1814 0 : "Cannot create empty timelines on inactive tenant"
1815 : );
1816 :
1817 : // Protect against concurrent attempts to use this TimelineId
1818 178 : let create_guard = self.create_timeline_create_guard(new_timeline_id)?;
1819 :
1820 176 : let new_metadata = TimelineMetadata::new(
1821 176 : // Initialize disk_consistent LSN to 0, The caller must import some data to
1822 176 : // make it valid, before calling finish_creation()
1823 176 : Lsn(0),
1824 176 : None,
1825 176 : None,
1826 176 : Lsn(0),
1827 176 : initdb_lsn,
1828 176 : initdb_lsn,
1829 176 : pg_version,
1830 176 : );
1831 176 : self.prepare_new_timeline(
1832 176 : new_timeline_id,
1833 176 : &new_metadata,
1834 176 : create_guard,
1835 176 : initdb_lsn,
1836 176 : None,
1837 176 : )
1838 0 : .await
1839 178 : }
1840 :
1841 : /// Helper for unit tests to create an empty timeline.
1842 : ///
1843 : /// The timeline is has state value `Active` but its background loops are not running.
1844 : // This makes the various functions which anyhow::ensure! for Active state work in tests.
1845 : // Our current tests don't need the background loops.
1846 : #[cfg(test)]
1847 168 : pub async fn create_test_timeline(
1848 168 : &self,
1849 168 : new_timeline_id: TimelineId,
1850 168 : initdb_lsn: Lsn,
1851 168 : pg_version: u32,
1852 168 : ctx: &RequestContext,
1853 168 : ) -> anyhow::Result<Arc<Timeline>> {
1854 168 : let uninit_tl = self
1855 168 : .create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
1856 0 : .await?;
1857 168 : let tline = uninit_tl.raw_timeline().expect("we just created it");
1858 168 : assert_eq!(tline.get_last_record_lsn(), Lsn(0));
1859 :
1860 : // Setup minimum keys required for the timeline to be usable.
1861 168 : let mut modification = tline.begin_modification(initdb_lsn);
1862 168 : modification
1863 168 : .init_empty_test_timeline()
1864 168 : .context("init_empty_test_timeline")?;
1865 168 : modification
1866 168 : .commit(ctx)
1867 162 : .await
1868 168 : .context("commit init_empty_test_timeline modification")?;
1869 :
1870 : // Flush to disk so that uninit_tl's check for valid disk_consistent_lsn passes.
1871 168 : tline.maybe_spawn_flush_loop();
1872 168 : tline.freeze_and_flush().await.context("freeze_and_flush")?;
1873 :
1874 : // Make sure the freeze_and_flush reaches remote storage.
1875 168 : tline.remote_client.wait_completion().await.unwrap();
1876 :
1877 168 : let tl = uninit_tl.finish_creation()?;
1878 : // The non-test code would call tl.activate() here.
1879 168 : tl.set_state(TimelineState::Active);
1880 168 : Ok(tl)
1881 168 : }
1882 :
1883 : /// Helper for unit tests to create a timeline with some pre-loaded states.
1884 : #[cfg(test)]
1885 : #[allow(clippy::too_many_arguments)]
1886 28 : pub async fn create_test_timeline_with_layers(
1887 28 : &self,
1888 28 : new_timeline_id: TimelineId,
1889 28 : initdb_lsn: Lsn,
1890 28 : pg_version: u32,
1891 28 : ctx: &RequestContext,
1892 28 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
1893 28 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
1894 28 : end_lsn: Lsn,
1895 28 : ) -> anyhow::Result<Arc<Timeline>> {
1896 : use checks::check_valid_layermap;
1897 : use itertools::Itertools;
1898 :
1899 28 : let tline = self
1900 28 : .create_test_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
1901 56 : .await?;
1902 28 : tline.force_advance_lsn(end_lsn);
1903 82 : for deltas in delta_layer_desc {
1904 54 : tline
1905 54 : .force_create_delta_layer(deltas, Some(initdb_lsn), ctx)
1906 162 : .await?;
1907 : }
1908 74 : for (lsn, images) in image_layer_desc {
1909 46 : tline
1910 46 : .force_create_image_layer(lsn, images, Some(initdb_lsn), ctx)
1911 269 : .await?;
1912 : }
1913 28 : let layer_names = tline
1914 28 : .layers
1915 28 : .read()
1916 0 : .await
1917 28 : .layer_map()
1918 28 : .unwrap()
1919 28 : .iter_historic_layers()
1920 128 : .map(|layer| layer.layer_name())
1921 28 : .collect_vec();
1922 28 : if let Some(err) = check_valid_layermap(&layer_names) {
1923 0 : bail!("invalid layermap: {err}");
1924 28 : }
1925 28 : Ok(tline)
1926 28 : }
1927 :
1928 : /// Create a new timeline.
1929 : ///
1930 : /// Returns the new timeline ID and reference to its Timeline object.
1931 : ///
1932 : /// If the caller specified the timeline ID to use (`new_timeline_id`), and timeline with
1933 : /// the same timeline ID already exists, returns CreateTimelineError::AlreadyExists.
1934 : #[allow(clippy::too_many_arguments)]
1935 0 : pub(crate) async fn create_timeline(
1936 0 : self: &Arc<Tenant>,
1937 0 : new_timeline_id: TimelineId,
1938 0 : ancestor_timeline_id: Option<TimelineId>,
1939 0 : mut ancestor_start_lsn: Option<Lsn>,
1940 0 : pg_version: u32,
1941 0 : load_existing_initdb: Option<TimelineId>,
1942 0 : broker_client: storage_broker::BrokerClientChannel,
1943 0 : ctx: &RequestContext,
1944 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
1945 0 : if !self.is_active() {
1946 0 : if matches!(self.current_state(), TenantState::Stopping { .. }) {
1947 0 : return Err(CreateTimelineError::ShuttingDown);
1948 : } else {
1949 0 : return Err(CreateTimelineError::Other(anyhow::anyhow!(
1950 0 : "Cannot create timelines on inactive tenant"
1951 0 : )));
1952 : }
1953 0 : }
1954 :
1955 0 : let _gate = self
1956 0 : .gate
1957 0 : .enter()
1958 0 : .map_err(|_| CreateTimelineError::ShuttingDown)?;
1959 :
1960 : // Get exclusive access to the timeline ID: this ensures that it does not already exist,
1961 : // and that no other creation attempts will be allowed in while we are working.
1962 0 : let create_guard = match self.create_timeline_create_guard(new_timeline_id) {
1963 0 : Ok(m) => m,
1964 : Err(TimelineExclusionError::AlreadyCreating) => {
1965 : // Creation is in progress, we cannot create it again, and we cannot
1966 : // check if this request matches the existing one, so caller must try
1967 : // again later.
1968 0 : return Err(CreateTimelineError::AlreadyCreating);
1969 : }
1970 0 : Err(TimelineExclusionError::Other(e)) => {
1971 0 : return Err(CreateTimelineError::Other(e));
1972 : }
1973 0 : Err(TimelineExclusionError::AlreadyExists(existing)) => {
1974 0 : debug!("timeline {new_timeline_id} already exists");
1975 :
1976 : // Idempotency: creating the same timeline twice is not an error, unless
1977 : // the second creation has different parameters.
1978 0 : if existing.get_ancestor_timeline_id() != ancestor_timeline_id
1979 0 : || existing.pg_version != pg_version
1980 0 : || (ancestor_start_lsn.is_some()
1981 0 : && ancestor_start_lsn != Some(existing.get_ancestor_lsn()))
1982 : {
1983 0 : return Err(CreateTimelineError::Conflict);
1984 0 : }
1985 0 :
1986 0 : // Wait for uploads to complete, so that when we return Ok, the timeline
1987 0 : // is known to be durable on remote storage. Just like we do at the end of
1988 0 : // this function, after we have created the timeline ourselves.
1989 0 : //
1990 0 : // We only really care that the initial version of `index_part.json` has
1991 0 : // been uploaded. That's enough to remember that the timeline
1992 0 : // exists. However, there is no function to wait specifically for that so
1993 0 : // we just wait for all in-progress uploads to finish.
1994 0 : existing
1995 0 : .remote_client
1996 0 : .wait_completion()
1997 0 : .await
1998 0 : .context("wait for timeline uploads to complete")?;
1999 :
2000 0 : return Ok(existing);
2001 : }
2002 : };
2003 :
2004 0 : pausable_failpoint!("timeline-creation-after-uninit");
2005 :
2006 0 : let loaded_timeline = match ancestor_timeline_id {
2007 0 : Some(ancestor_timeline_id) => {
2008 0 : let ancestor_timeline = self
2009 0 : .get_timeline(ancestor_timeline_id, false)
2010 0 : .context("Cannot branch off the timeline that's not present in pageserver")?;
2011 :
2012 : // instead of waiting around, just deny the request because ancestor is not yet
2013 : // ready for other purposes either.
2014 0 : if !ancestor_timeline.is_active() {
2015 0 : return Err(CreateTimelineError::AncestorNotActive);
2016 0 : }
2017 0 :
2018 0 : if ancestor_timeline.is_archived() == Some(true) {
2019 0 : info!("tried to branch archived timeline");
2020 0 : return Err(CreateTimelineError::AncestorArchived);
2021 0 : }
2022 :
2023 0 : if let Some(lsn) = ancestor_start_lsn.as_mut() {
2024 0 : *lsn = lsn.align();
2025 0 :
2026 0 : let ancestor_ancestor_lsn = ancestor_timeline.get_ancestor_lsn();
2027 0 : if ancestor_ancestor_lsn > *lsn {
2028 : // can we safely just branch from the ancestor instead?
2029 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
2030 0 : "invalid start lsn {} for ancestor timeline {}: less than timeline ancestor lsn {}",
2031 0 : lsn,
2032 0 : ancestor_timeline_id,
2033 0 : ancestor_ancestor_lsn,
2034 0 : )));
2035 0 : }
2036 0 :
2037 0 : // Wait for the WAL to arrive and be processed on the parent branch up
2038 0 : // to the requested branch point. The repository code itself doesn't
2039 0 : // require it, but if we start to receive WAL on the new timeline,
2040 0 : // decoding the new WAL might need to look up previous pages, relation
2041 0 : // sizes etc. and that would get confused if the previous page versions
2042 0 : // are not in the repository yet.
2043 0 : ancestor_timeline
2044 0 : .wait_lsn(*lsn, timeline::WaitLsnWaiter::Tenant, ctx)
2045 0 : .await
2046 0 : .map_err(|e| match e {
2047 0 : e @ (WaitLsnError::Timeout(_) | WaitLsnError::BadState { .. }) => {
2048 0 : CreateTimelineError::AncestorLsn(anyhow::anyhow!(e))
2049 : }
2050 0 : WaitLsnError::Shutdown => CreateTimelineError::ShuttingDown,
2051 0 : })?;
2052 0 : }
2053 :
2054 0 : self.branch_timeline(
2055 0 : &ancestor_timeline,
2056 0 : new_timeline_id,
2057 0 : ancestor_start_lsn,
2058 0 : create_guard,
2059 0 : ctx,
2060 0 : )
2061 0 : .await?
2062 : }
2063 : None => {
2064 0 : self.bootstrap_timeline(
2065 0 : new_timeline_id,
2066 0 : pg_version,
2067 0 : load_existing_initdb,
2068 0 : create_guard,
2069 0 : ctx,
2070 0 : )
2071 0 : .await?
2072 : }
2073 : };
2074 :
2075 : // At this point we have dropped our guard on [`Self::timelines_creating`], and
2076 : // the timeline is visible in [`Self::timelines`], but it is _not_ durable yet. We must
2077 : // not send a success to the caller until it is. The same applies to handling retries,
2078 : // see the handling of [`TimelineExclusionError::AlreadyExists`] above.
2079 0 : let kind = ancestor_timeline_id
2080 0 : .map(|_| "branched")
2081 0 : .unwrap_or("bootstrapped");
2082 0 : loaded_timeline
2083 0 : .remote_client
2084 0 : .wait_completion()
2085 0 : .await
2086 0 : .with_context(|| format!("wait for {} timeline initial uploads to complete", kind))?;
2087 :
2088 0 : loaded_timeline.activate(self.clone(), broker_client, None, ctx);
2089 0 :
2090 0 : Ok(loaded_timeline)
2091 0 : }
2092 :
2093 0 : pub(crate) async fn delete_timeline(
2094 0 : self: Arc<Self>,
2095 0 : timeline_id: TimelineId,
2096 0 : ) -> Result<(), DeleteTimelineError> {
2097 0 : DeleteTimelineFlow::run(&self, timeline_id).await?;
2098 :
2099 0 : Ok(())
2100 0 : }
2101 :
2102 : /// perform one garbage collection iteration, removing old data files from disk.
2103 : /// this function is periodically called by gc task.
2104 : /// also it can be explicitly requested through page server api 'do_gc' command.
2105 : ///
2106 : /// `target_timeline_id` specifies the timeline to GC, or None for all.
2107 : ///
2108 : /// The `horizon` an `pitr` parameters determine how much WAL history needs to be retained.
2109 : /// Also known as the retention period, or the GC cutoff point. `horizon` specifies
2110 : /// the amount of history, as LSN difference from current latest LSN on each timeline.
2111 : /// `pitr` specifies the same as a time difference from the current time. The effective
2112 : /// GC cutoff point is determined conservatively by either `horizon` and `pitr`, whichever
2113 : /// requires more history to be retained.
2114 : //
2115 754 : pub(crate) async fn gc_iteration(
2116 754 : &self,
2117 754 : target_timeline_id: Option<TimelineId>,
2118 754 : horizon: u64,
2119 754 : pitr: Duration,
2120 754 : cancel: &CancellationToken,
2121 754 : ctx: &RequestContext,
2122 754 : ) -> Result<GcResult, GcError> {
2123 754 : // Don't start doing work during shutdown
2124 754 : if let TenantState::Stopping { .. } = self.current_state() {
2125 0 : return Ok(GcResult::default());
2126 754 : }
2127 754 :
2128 754 : // there is a global allowed_error for this
2129 754 : if !self.is_active() {
2130 0 : return Err(GcError::NotActive);
2131 754 : }
2132 754 :
2133 754 : {
2134 754 : let conf = self.tenant_conf.load();
2135 754 :
2136 754 : if !conf.location.may_delete_layers_hint() {
2137 0 : info!("Skipping GC in location state {:?}", conf.location);
2138 0 : return Ok(GcResult::default());
2139 754 : }
2140 754 :
2141 754 : if conf.is_gc_blocked_by_lsn_lease_deadline() {
2142 750 : info!("Skipping GC because lsn lease deadline is not reached");
2143 750 : return Ok(GcResult::default());
2144 4 : }
2145 : }
2146 :
2147 4 : let _guard = match self.gc_block.start().await {
2148 4 : Ok(guard) => guard,
2149 0 : Err(reasons) => {
2150 0 : info!("Skipping GC: {reasons}");
2151 0 : return Ok(GcResult::default());
2152 : }
2153 : };
2154 :
2155 4 : self.gc_iteration_internal(target_timeline_id, horizon, pitr, cancel, ctx)
2156 4 : .await
2157 754 : }
2158 :
2159 : /// Perform one compaction iteration.
2160 : /// This function is periodically called by compactor task.
2161 : /// Also it can be explicitly requested per timeline through page server
2162 : /// api's 'compact' command.
2163 : ///
2164 : /// Returns whether we have pending compaction task.
2165 0 : async fn compaction_iteration(
2166 0 : self: &Arc<Self>,
2167 0 : cancel: &CancellationToken,
2168 0 : ctx: &RequestContext,
2169 0 : ) -> Result<bool, timeline::CompactionError> {
2170 0 : // Don't start doing work during shutdown, or when broken, we do not need those in the logs
2171 0 : if !self.is_active() {
2172 0 : return Ok(false);
2173 0 : }
2174 0 :
2175 0 : {
2176 0 : let conf = self.tenant_conf.load();
2177 0 : if !conf.location.may_delete_layers_hint() || !conf.location.may_upload_layers_hint() {
2178 0 : info!("Skipping compaction in location state {:?}", conf.location);
2179 0 : return Ok(false);
2180 0 : }
2181 0 : }
2182 0 :
2183 0 : // Scan through the hashmap and collect a list of all the timelines,
2184 0 : // while holding the lock. Then drop the lock and actually perform the
2185 0 : // compactions. We don't want to block everything else while the
2186 0 : // compaction runs.
2187 0 : let timelines_to_compact_or_offload;
2188 0 : {
2189 0 : let timelines = self.timelines.lock().unwrap();
2190 0 : timelines_to_compact_or_offload = timelines
2191 0 : .iter()
2192 0 : .filter_map(|(timeline_id, timeline)| {
2193 0 : let (is_active, can_offload) = (timeline.is_active(), timeline.can_offload());
2194 0 : let has_no_unoffloaded_children = {
2195 0 : !timelines
2196 0 : .iter()
2197 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(*timeline_id))
2198 : };
2199 0 : let can_offload =
2200 0 : can_offload && has_no_unoffloaded_children && self.conf.timeline_offloading;
2201 0 : if (is_active, can_offload) == (false, false) {
2202 0 : None
2203 : } else {
2204 0 : Some((*timeline_id, timeline.clone(), (is_active, can_offload)))
2205 : }
2206 0 : })
2207 0 : .collect::<Vec<_>>();
2208 0 : drop(timelines);
2209 0 : }
2210 0 :
2211 0 : // Before doing any I/O work, check our circuit breaker
2212 0 : if self.compaction_circuit_breaker.lock().unwrap().is_broken() {
2213 0 : info!("Skipping compaction due to previous failures");
2214 0 : return Ok(false);
2215 0 : }
2216 0 :
2217 0 : let mut has_pending_task = false;
2218 :
2219 0 : for (timeline_id, timeline, (can_compact, can_offload)) in &timelines_to_compact_or_offload
2220 : {
2221 0 : let pending_task_left = if *can_compact {
2222 : Some(
2223 0 : timeline
2224 0 : .compact(cancel, EnumSet::empty(), ctx)
2225 0 : .instrument(info_span!("compact_timeline", %timeline_id))
2226 0 : .await
2227 0 : .inspect_err(|e| match e {
2228 0 : timeline::CompactionError::ShuttingDown => (),
2229 0 : timeline::CompactionError::Other(e) => {
2230 0 : self.compaction_circuit_breaker
2231 0 : .lock()
2232 0 : .unwrap()
2233 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, e);
2234 0 : }
2235 0 : })?,
2236 : )
2237 : } else {
2238 0 : None
2239 : };
2240 0 : has_pending_task |= pending_task_left.unwrap_or(false);
2241 0 : if pending_task_left == Some(false) && *can_offload {
2242 0 : offload_timeline(self, timeline)
2243 0 : .instrument(info_span!("offload_timeline", %timeline_id))
2244 0 : .await
2245 0 : .map_err(timeline::CompactionError::Other)?;
2246 0 : }
2247 : }
2248 :
2249 0 : self.compaction_circuit_breaker
2250 0 : .lock()
2251 0 : .unwrap()
2252 0 : .success(&CIRCUIT_BREAKERS_UNBROKEN);
2253 0 :
2254 0 : Ok(has_pending_task)
2255 0 : }
2256 :
2257 : // Call through to all timelines to freeze ephemeral layers if needed. Usually
2258 : // this happens during ingest: this background housekeeping is for freezing layers
2259 : // that are open but haven't been written to for some time.
2260 0 : async fn ingest_housekeeping(&self) {
2261 0 : // Scan through the hashmap and collect a list of all the timelines,
2262 0 : // while holding the lock. Then drop the lock and actually perform the
2263 0 : // compactions. We don't want to block everything else while the
2264 0 : // compaction runs.
2265 0 : let timelines = {
2266 0 : self.timelines
2267 0 : .lock()
2268 0 : .unwrap()
2269 0 : .values()
2270 0 : .filter_map(|timeline| {
2271 0 : if timeline.is_active() {
2272 0 : Some(timeline.clone())
2273 : } else {
2274 0 : None
2275 : }
2276 0 : })
2277 0 : .collect::<Vec<_>>()
2278 : };
2279 :
2280 0 : for timeline in &timelines {
2281 0 : timeline.maybe_freeze_ephemeral_layer().await;
2282 : }
2283 0 : }
2284 :
2285 0 : pub fn timeline_has_no_attached_children(&self, timeline_id: TimelineId) -> bool {
2286 0 : let timelines = self.timelines.lock().unwrap();
2287 0 : !timelines
2288 0 : .iter()
2289 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(timeline_id))
2290 0 : }
2291 :
2292 1696 : pub fn current_state(&self) -> TenantState {
2293 1696 : self.state.borrow().clone()
2294 1696 : }
2295 :
2296 936 : pub fn is_active(&self) -> bool {
2297 936 : self.current_state() == TenantState::Active
2298 936 : }
2299 :
2300 0 : pub fn generation(&self) -> Generation {
2301 0 : self.generation
2302 0 : }
2303 :
2304 0 : pub(crate) fn wal_redo_manager_status(&self) -> Option<WalRedoManagerStatus> {
2305 0 : self.walredo_mgr.as_ref().and_then(|mgr| mgr.status())
2306 0 : }
2307 :
2308 : /// Changes tenant status to active, unless shutdown was already requested.
2309 : ///
2310 : /// `background_jobs_can_start` is an optional barrier set to a value during pageserver startup
2311 : /// to delay background jobs. Background jobs can be started right away when None is given.
2312 0 : fn activate(
2313 0 : self: &Arc<Self>,
2314 0 : broker_client: BrokerClientChannel,
2315 0 : background_jobs_can_start: Option<&completion::Barrier>,
2316 0 : ctx: &RequestContext,
2317 0 : ) {
2318 0 : span::debug_assert_current_span_has_tenant_id();
2319 0 :
2320 0 : let mut activating = false;
2321 0 : self.state.send_modify(|current_state| {
2322 : use pageserver_api::models::ActivatingFrom;
2323 0 : match &*current_state {
2324 : TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
2325 0 : panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {state:?}", state = current_state);
2326 : }
2327 0 : TenantState::Attaching => {
2328 0 : *current_state = TenantState::Activating(ActivatingFrom::Attaching);
2329 0 : }
2330 0 : }
2331 0 : debug!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), "Activating tenant");
2332 0 : activating = true;
2333 0 : // Continue outside the closure. We need to grab timelines.lock()
2334 0 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
2335 0 : });
2336 0 :
2337 0 : if activating {
2338 0 : let timelines_accessor = self.timelines.lock().unwrap();
2339 0 : let timelines_offloaded_accessor = self.timelines_offloaded.lock().unwrap();
2340 0 : let timelines_to_activate = timelines_accessor
2341 0 : .values()
2342 0 : .filter(|timeline| !(timeline.is_broken() || timeline.is_stopping()));
2343 0 :
2344 0 : // Before activation, populate each Timeline's GcInfo with information about its children
2345 0 : self.initialize_gc_info(&timelines_accessor, &timelines_offloaded_accessor);
2346 0 :
2347 0 : // Spawn gc and compaction loops. The loops will shut themselves
2348 0 : // down when they notice that the tenant is inactive.
2349 0 : tasks::start_background_loops(self, background_jobs_can_start);
2350 0 :
2351 0 : let mut activated_timelines = 0;
2352 :
2353 0 : for timeline in timelines_to_activate {
2354 0 : timeline.activate(
2355 0 : self.clone(),
2356 0 : broker_client.clone(),
2357 0 : background_jobs_can_start,
2358 0 : ctx,
2359 0 : );
2360 0 : activated_timelines += 1;
2361 0 : }
2362 :
2363 0 : self.state.send_modify(move |current_state| {
2364 0 : assert!(
2365 0 : matches!(current_state, TenantState::Activating(_)),
2366 0 : "set_stopping and set_broken wait for us to leave Activating state",
2367 : );
2368 0 : *current_state = TenantState::Active;
2369 0 :
2370 0 : let elapsed = self.constructed_at.elapsed();
2371 0 : let total_timelines = timelines_accessor.len();
2372 0 :
2373 0 : // log a lot of stuff, because some tenants sometimes suffer from user-visible
2374 0 : // times to activate. see https://github.com/neondatabase/neon/issues/4025
2375 0 : info!(
2376 0 : since_creation_millis = elapsed.as_millis(),
2377 0 : tenant_id = %self.tenant_shard_id.tenant_id,
2378 0 : shard_id = %self.tenant_shard_id.shard_slug(),
2379 0 : activated_timelines,
2380 0 : total_timelines,
2381 0 : post_state = <&'static str>::from(&*current_state),
2382 0 : "activation attempt finished"
2383 : );
2384 :
2385 0 : TENANT.activation.observe(elapsed.as_secs_f64());
2386 0 : });
2387 0 : }
2388 0 : }
2389 :
2390 : /// Shutdown the tenant and join all of the spawned tasks.
2391 : ///
2392 : /// The method caters for all use-cases:
2393 : /// - pageserver shutdown (freeze_and_flush == true)
2394 : /// - detach + ignore (freeze_and_flush == false)
2395 : ///
2396 : /// This will attempt to shutdown even if tenant is broken.
2397 : ///
2398 : /// `shutdown_progress` is a [`completion::Barrier`] for the shutdown initiated by this call.
2399 : /// If the tenant is already shutting down, we return a clone of the first shutdown call's
2400 : /// `Barrier` as an `Err`. This not-first caller can use the returned barrier to join with
2401 : /// the ongoing shutdown.
2402 6 : async fn shutdown(
2403 6 : &self,
2404 6 : shutdown_progress: completion::Barrier,
2405 6 : shutdown_mode: timeline::ShutdownMode,
2406 6 : ) -> Result<(), completion::Barrier> {
2407 6 : span::debug_assert_current_span_has_tenant_id();
2408 :
2409 : // Set tenant (and its timlines) to Stoppping state.
2410 : //
2411 : // Since we can only transition into Stopping state after activation is complete,
2412 : // run it in a JoinSet so all tenants have a chance to stop before we get SIGKILLed.
2413 : //
2414 : // Transitioning tenants to Stopping state has a couple of non-obvious side effects:
2415 : // 1. Lock out any new requests to the tenants.
2416 : // 2. Signal cancellation to WAL receivers (we wait on it below).
2417 : // 3. Signal cancellation for other tenant background loops.
2418 : // 4. ???
2419 : //
2420 : // The waiting for the cancellation is not done uniformly.
2421 : // We certainly wait for WAL receivers to shut down.
2422 : // That is necessary so that no new data comes in before the freeze_and_flush.
2423 : // But the tenant background loops are joined-on in our caller.
2424 : // It's mesed up.
2425 : // we just ignore the failure to stop
2426 :
2427 : // If we're still attaching, fire the cancellation token early to drop out: this
2428 : // will prevent us flushing, but ensures timely shutdown if some I/O during attach
2429 : // is very slow.
2430 6 : let shutdown_mode = if matches!(self.current_state(), TenantState::Attaching) {
2431 0 : self.cancel.cancel();
2432 0 :
2433 0 : // Having fired our cancellation token, do not try and flush timelines: their cancellation tokens
2434 0 : // are children of ours, so their flush loops will have shut down already
2435 0 : timeline::ShutdownMode::Hard
2436 : } else {
2437 6 : shutdown_mode
2438 : };
2439 :
2440 6 : match self.set_stopping(shutdown_progress, false, false).await {
2441 6 : Ok(()) => {}
2442 0 : Err(SetStoppingError::Broken) => {
2443 0 : // assume that this is acceptable
2444 0 : }
2445 0 : Err(SetStoppingError::AlreadyStopping(other)) => {
2446 0 : // give caller the option to wait for this this shutdown
2447 0 : info!("Tenant::shutdown: AlreadyStopping");
2448 0 : return Err(other);
2449 : }
2450 : };
2451 :
2452 6 : let mut js = tokio::task::JoinSet::new();
2453 6 : {
2454 6 : let timelines = self.timelines.lock().unwrap();
2455 6 : timelines.values().for_each(|timeline| {
2456 6 : let timeline = Arc::clone(timeline);
2457 6 : let timeline_id = timeline.timeline_id;
2458 6 : let span = tracing::info_span!("timeline_shutdown", %timeline_id, ?shutdown_mode);
2459 11 : js.spawn(async move { timeline.shutdown(shutdown_mode).instrument(span).await });
2460 6 : })
2461 6 : };
2462 6 : // test_long_timeline_create_then_tenant_delete is leaning on this message
2463 6 : tracing::info!("Waiting for timelines...");
2464 12 : while let Some(res) = js.join_next().await {
2465 0 : match res {
2466 6 : Ok(()) => {}
2467 0 : Err(je) if je.is_cancelled() => unreachable!("no cancelling used"),
2468 0 : Err(je) if je.is_panic() => { /* logged already */ }
2469 0 : Err(je) => warn!("unexpected JoinError: {je:?}"),
2470 : }
2471 : }
2472 :
2473 : // We cancel the Tenant's cancellation token _after_ the timelines have all shut down. This permits
2474 : // them to continue to do work during their shutdown methods, e.g. flushing data.
2475 6 : tracing::debug!("Cancelling CancellationToken");
2476 6 : self.cancel.cancel();
2477 6 :
2478 6 : // shutdown all tenant and timeline tasks: gc, compaction, page service
2479 6 : // No new tasks will be started for this tenant because it's in `Stopping` state.
2480 6 : //
2481 6 : // this will additionally shutdown and await all timeline tasks.
2482 6 : tracing::debug!("Waiting for tasks...");
2483 6 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), None).await;
2484 :
2485 6 : if let Some(walredo_mgr) = self.walredo_mgr.as_ref() {
2486 6 : walredo_mgr.shutdown().await;
2487 0 : }
2488 :
2489 : // Wait for any in-flight operations to complete
2490 6 : self.gate.close().await;
2491 :
2492 6 : remove_tenant_metrics(&self.tenant_shard_id);
2493 6 :
2494 6 : Ok(())
2495 6 : }
2496 :
2497 : /// Change tenant status to Stopping, to mark that it is being shut down.
2498 : ///
2499 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
2500 : ///
2501 : /// This function is not cancel-safe!
2502 : ///
2503 : /// `allow_transition_from_loading` is needed for the special case of loading task deleting the tenant.
2504 : /// `allow_transition_from_attaching` is needed for the special case of attaching deleted tenant.
2505 6 : async fn set_stopping(
2506 6 : &self,
2507 6 : progress: completion::Barrier,
2508 6 : _allow_transition_from_loading: bool,
2509 6 : allow_transition_from_attaching: bool,
2510 6 : ) -> Result<(), SetStoppingError> {
2511 6 : let mut rx = self.state.subscribe();
2512 6 :
2513 6 : // cannot stop before we're done activating, so wait out until we're done activating
2514 6 : rx.wait_for(|state| match state {
2515 0 : TenantState::Attaching if allow_transition_from_attaching => true,
2516 : TenantState::Activating(_) | TenantState::Attaching => {
2517 0 : info!(
2518 0 : "waiting for {} to turn Active|Broken|Stopping",
2519 0 : <&'static str>::from(state)
2520 : );
2521 0 : false
2522 : }
2523 6 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
2524 6 : })
2525 0 : .await
2526 6 : .expect("cannot drop self.state while on a &self method");
2527 6 :
2528 6 : // we now know we're done activating, let's see whether this task is the winner to transition into Stopping
2529 6 : let mut err = None;
2530 6 : let stopping = self.state.send_if_modified(|current_state| match current_state {
2531 : TenantState::Activating(_) => {
2532 0 : unreachable!("1we ensured above that we're done with activation, and, there is no re-activation")
2533 : }
2534 : TenantState::Attaching => {
2535 0 : if !allow_transition_from_attaching {
2536 0 : unreachable!("2we ensured above that we're done with activation, and, there is no re-activation")
2537 0 : };
2538 0 : *current_state = TenantState::Stopping { progress };
2539 0 : true
2540 : }
2541 : TenantState::Active => {
2542 : // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
2543 : // are created after the transition to Stopping. That's harmless, as the Timelines
2544 : // won't be accessible to anyone afterwards, because the Tenant is in Stopping state.
2545 6 : *current_state = TenantState::Stopping { progress };
2546 6 : // Continue stopping outside the closure. We need to grab timelines.lock()
2547 6 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
2548 6 : true
2549 : }
2550 0 : TenantState::Broken { reason, .. } => {
2551 0 : info!(
2552 0 : "Cannot set tenant to Stopping state, it is in Broken state due to: {reason}"
2553 : );
2554 0 : err = Some(SetStoppingError::Broken);
2555 0 : false
2556 : }
2557 0 : TenantState::Stopping { progress } => {
2558 0 : info!("Tenant is already in Stopping state");
2559 0 : err = Some(SetStoppingError::AlreadyStopping(progress.clone()));
2560 0 : false
2561 : }
2562 6 : });
2563 6 : match (stopping, err) {
2564 6 : (true, None) => {} // continue
2565 0 : (false, Some(err)) => return Err(err),
2566 0 : (true, Some(_)) => unreachable!(
2567 0 : "send_if_modified closure must error out if not transitioning to Stopping"
2568 0 : ),
2569 0 : (false, None) => unreachable!(
2570 0 : "send_if_modified closure must return true if transitioning to Stopping"
2571 0 : ),
2572 : }
2573 :
2574 6 : let timelines_accessor = self.timelines.lock().unwrap();
2575 6 : let not_broken_timelines = timelines_accessor
2576 6 : .values()
2577 6 : .filter(|timeline| !timeline.is_broken());
2578 12 : for timeline in not_broken_timelines {
2579 6 : timeline.set_state(TimelineState::Stopping);
2580 6 : }
2581 6 : Ok(())
2582 6 : }
2583 :
2584 : /// Method for tenant::mgr to transition us into Broken state in case of a late failure in
2585 : /// `remove_tenant_from_memory`
2586 : ///
2587 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
2588 : ///
2589 : /// In tests, we also use this to set tenants to Broken state on purpose.
2590 0 : pub(crate) async fn set_broken(&self, reason: String) {
2591 0 : let mut rx = self.state.subscribe();
2592 0 :
2593 0 : // The load & attach routines own the tenant state until it has reached `Active`.
2594 0 : // So, wait until it's done.
2595 0 : rx.wait_for(|state| match state {
2596 : TenantState::Activating(_) | TenantState::Attaching => {
2597 0 : info!(
2598 0 : "waiting for {} to turn Active|Broken|Stopping",
2599 0 : <&'static str>::from(state)
2600 : );
2601 0 : false
2602 : }
2603 0 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
2604 0 : })
2605 0 : .await
2606 0 : .expect("cannot drop self.state while on a &self method");
2607 0 :
2608 0 : // we now know we're done activating, let's see whether this task is the winner to transition into Broken
2609 0 : self.set_broken_no_wait(reason)
2610 0 : }
2611 :
2612 0 : pub(crate) fn set_broken_no_wait(&self, reason: impl Display) {
2613 0 : let reason = reason.to_string();
2614 0 : self.state.send_modify(|current_state| {
2615 0 : match *current_state {
2616 : TenantState::Activating(_) | TenantState::Attaching => {
2617 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
2618 : }
2619 : TenantState::Active => {
2620 0 : if cfg!(feature = "testing") {
2621 0 : warn!("Changing Active tenant to Broken state, reason: {}", reason);
2622 0 : *current_state = TenantState::broken_from_reason(reason);
2623 : } else {
2624 0 : unreachable!("not allowed to call set_broken on Active tenants in non-testing builds")
2625 : }
2626 : }
2627 : TenantState::Broken { .. } => {
2628 0 : warn!("Tenant is already in Broken state");
2629 : }
2630 : // This is the only "expected" path, any other path is a bug.
2631 : TenantState::Stopping { .. } => {
2632 0 : warn!(
2633 0 : "Marking Stopping tenant as Broken state, reason: {}",
2634 : reason
2635 : );
2636 0 : *current_state = TenantState::broken_from_reason(reason);
2637 : }
2638 : }
2639 0 : });
2640 0 : }
2641 :
2642 0 : pub fn subscribe_for_state_updates(&self) -> watch::Receiver<TenantState> {
2643 0 : self.state.subscribe()
2644 0 : }
2645 :
2646 : /// The activate_now semaphore is initialized with zero units. As soon as
2647 : /// we add a unit, waiters will be able to acquire a unit and proceed.
2648 0 : pub(crate) fn activate_now(&self) {
2649 0 : self.activate_now_sem.add_permits(1);
2650 0 : }
2651 :
2652 0 : pub(crate) async fn wait_to_become_active(
2653 0 : &self,
2654 0 : timeout: Duration,
2655 0 : ) -> Result<(), GetActiveTenantError> {
2656 0 : let mut receiver = self.state.subscribe();
2657 : loop {
2658 0 : let current_state = receiver.borrow_and_update().clone();
2659 0 : match current_state {
2660 : TenantState::Attaching | TenantState::Activating(_) => {
2661 : // in these states, there's a chance that we can reach ::Active
2662 0 : self.activate_now();
2663 0 : match timeout_cancellable(timeout, &self.cancel, receiver.changed()).await {
2664 0 : Ok(r) => {
2665 0 : r.map_err(
2666 0 : |_e: tokio::sync::watch::error::RecvError|
2667 : // Tenant existed but was dropped: report it as non-existent
2668 0 : GetActiveTenantError::NotFound(GetTenantError::NotFound(self.tenant_shard_id.tenant_id))
2669 0 : )?
2670 : }
2671 : Err(TimeoutCancellableError::Cancelled) => {
2672 0 : return Err(GetActiveTenantError::Cancelled);
2673 : }
2674 : Err(TimeoutCancellableError::Timeout) => {
2675 0 : return Err(GetActiveTenantError::WaitForActiveTimeout {
2676 0 : latest_state: Some(self.current_state()),
2677 0 : wait_time: timeout,
2678 0 : });
2679 : }
2680 : }
2681 : }
2682 : TenantState::Active { .. } => {
2683 0 : return Ok(());
2684 : }
2685 0 : TenantState::Broken { reason, .. } => {
2686 0 : // This is fatal, and reported distinctly from the general case of "will never be active" because
2687 0 : // it's logically a 500 to external API users (broken is always a bug).
2688 0 : return Err(GetActiveTenantError::Broken(reason));
2689 : }
2690 : TenantState::Stopping { .. } => {
2691 : // There's no chance the tenant can transition back into ::Active
2692 0 : return Err(GetActiveTenantError::WillNotBecomeActive(current_state));
2693 : }
2694 : }
2695 : }
2696 0 : }
2697 :
2698 0 : pub(crate) fn get_attach_mode(&self) -> AttachmentMode {
2699 0 : self.tenant_conf.load().location.attach_mode
2700 0 : }
2701 :
2702 : /// For API access: generate a LocationConfig equivalent to the one that would be used to
2703 : /// create a Tenant in the same state. Do not use this in hot paths: it's for relatively
2704 : /// rare external API calls, like a reconciliation at startup.
2705 0 : pub(crate) fn get_location_conf(&self) -> models::LocationConfig {
2706 0 : let conf = self.tenant_conf.load();
2707 :
2708 0 : let location_config_mode = match conf.location.attach_mode {
2709 0 : AttachmentMode::Single => models::LocationConfigMode::AttachedSingle,
2710 0 : AttachmentMode::Multi => models::LocationConfigMode::AttachedMulti,
2711 0 : AttachmentMode::Stale => models::LocationConfigMode::AttachedStale,
2712 : };
2713 :
2714 : // We have a pageserver TenantConf, we need the API-facing TenantConfig.
2715 0 : let tenant_config: models::TenantConfig = conf.tenant_conf.clone().into();
2716 0 :
2717 0 : models::LocationConfig {
2718 0 : mode: location_config_mode,
2719 0 : generation: self.generation.into(),
2720 0 : secondary_conf: None,
2721 0 : shard_number: self.shard_identity.number.0,
2722 0 : shard_count: self.shard_identity.count.literal(),
2723 0 : shard_stripe_size: self.shard_identity.stripe_size.0,
2724 0 : tenant_conf: tenant_config,
2725 0 : }
2726 0 : }
2727 :
2728 0 : pub(crate) fn get_tenant_shard_id(&self) -> &TenantShardId {
2729 0 : &self.tenant_shard_id
2730 0 : }
2731 :
2732 0 : pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
2733 0 : self.shard_identity.stripe_size
2734 0 : }
2735 :
2736 0 : pub(crate) fn get_generation(&self) -> Generation {
2737 0 : self.generation
2738 0 : }
2739 :
2740 : /// This function partially shuts down the tenant (it shuts down the Timelines) and is fallible,
2741 : /// and can leave the tenant in a bad state if it fails. The caller is responsible for
2742 : /// resetting this tenant to a valid state if we fail.
2743 0 : pub(crate) async fn split_prepare(
2744 0 : &self,
2745 0 : child_shards: &Vec<TenantShardId>,
2746 0 : ) -> anyhow::Result<()> {
2747 0 : let timelines = self.timelines.lock().unwrap().clone();
2748 0 : for timeline in timelines.values() {
2749 : // We do not block timeline creation/deletion during splits inside the pageserver: it is up to higher levels
2750 : // to ensure that they do not start a split if currently in the process of doing these.
2751 :
2752 : // Upload an index from the parent: this is partly to provide freshness for the
2753 : // child tenants that will copy it, and partly for general ease-of-debugging: there will
2754 : // always be a parent shard index in the same generation as we wrote the child shard index.
2755 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Uploading index");
2756 0 : timeline
2757 0 : .remote_client
2758 0 : .schedule_index_upload_for_file_changes()?;
2759 0 : timeline.remote_client.wait_completion().await?;
2760 :
2761 : // Shut down the timeline's remote client: this means that the indices we write
2762 : // for child shards will not be invalidated by the parent shard deleting layers.
2763 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Shutting down remote storage client");
2764 0 : timeline.remote_client.shutdown().await;
2765 :
2766 : // Download methods can still be used after shutdown, as they don't flow through the remote client's
2767 : // queue. In principal the RemoteTimelineClient could provide this without downloading it, but this
2768 : // operation is rare, so it's simpler to just download it (and robustly guarantees that the index
2769 : // we use here really is the remotely persistent one).
2770 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Downloading index_part from parent");
2771 0 : let result = timeline.remote_client
2772 0 : .download_index_file(&self.cancel)
2773 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=%timeline.timeline_id))
2774 0 : .await?;
2775 0 : let index_part = match result {
2776 : MaybeDeletedIndexPart::Deleted(_) => {
2777 0 : anyhow::bail!("Timeline deletion happened concurrently with split")
2778 : }
2779 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
2780 : };
2781 :
2782 0 : for child_shard in child_shards {
2783 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Uploading index_part for child {}", child_shard.to_index());
2784 0 : upload_index_part(
2785 0 : &self.remote_storage,
2786 0 : child_shard,
2787 0 : &timeline.timeline_id,
2788 0 : self.generation,
2789 0 : &index_part,
2790 0 : &self.cancel,
2791 0 : )
2792 0 : .await?;
2793 : }
2794 : }
2795 :
2796 0 : Ok(())
2797 0 : }
2798 :
2799 0 : pub(crate) fn get_sizes(&self) -> TopTenantShardItem {
2800 0 : let mut result = TopTenantShardItem {
2801 0 : id: self.tenant_shard_id,
2802 0 : resident_size: 0,
2803 0 : physical_size: 0,
2804 0 : max_logical_size: 0,
2805 0 : };
2806 :
2807 0 : for timeline in self.timelines.lock().unwrap().values() {
2808 0 : result.resident_size += timeline.metrics.resident_physical_size_gauge.get();
2809 0 :
2810 0 : result.physical_size += timeline
2811 0 : .remote_client
2812 0 : .metrics
2813 0 : .remote_physical_size_gauge
2814 0 : .get();
2815 0 : result.max_logical_size = std::cmp::max(
2816 0 : result.max_logical_size,
2817 0 : timeline.metrics.current_logical_size_gauge.get(),
2818 0 : );
2819 0 : }
2820 :
2821 0 : result
2822 0 : }
2823 : }
2824 :
2825 : /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
2826 : /// perform a topological sort, so that the parent of each timeline comes
2827 : /// before the children.
2828 : /// E extracts the ancestor from T
2829 : /// This allows for T to be different. It can be TimelineMetadata, can be Timeline itself, etc.
2830 186 : fn tree_sort_timelines<T, E>(
2831 186 : timelines: HashMap<TimelineId, T>,
2832 186 : extractor: E,
2833 186 : ) -> anyhow::Result<Vec<(TimelineId, T)>>
2834 186 : where
2835 186 : E: Fn(&T) -> Option<TimelineId>,
2836 186 : {
2837 186 : let mut result = Vec::with_capacity(timelines.len());
2838 186 :
2839 186 : let mut now = Vec::with_capacity(timelines.len());
2840 186 : // (ancestor, children)
2841 186 : let mut later: HashMap<TimelineId, Vec<(TimelineId, T)>> =
2842 186 : HashMap::with_capacity(timelines.len());
2843 :
2844 192 : for (timeline_id, value) in timelines {
2845 6 : if let Some(ancestor_id) = extractor(&value) {
2846 2 : let children = later.entry(ancestor_id).or_default();
2847 2 : children.push((timeline_id, value));
2848 4 : } else {
2849 4 : now.push((timeline_id, value));
2850 4 : }
2851 : }
2852 :
2853 192 : while let Some((timeline_id, metadata)) = now.pop() {
2854 6 : result.push((timeline_id, metadata));
2855 : // All children of this can be loaded now
2856 6 : if let Some(mut children) = later.remove(&timeline_id) {
2857 2 : now.append(&mut children);
2858 4 : }
2859 : }
2860 :
2861 : // All timelines should be visited now. Unless there were timelines with missing ancestors.
2862 186 : if !later.is_empty() {
2863 0 : for (missing_id, orphan_ids) in later {
2864 0 : for (orphan_id, _) in orphan_ids {
2865 0 : error!("could not load timeline {orphan_id} because its ancestor timeline {missing_id} could not be loaded");
2866 : }
2867 : }
2868 0 : bail!("could not load tenant because some timelines are missing ancestors");
2869 186 : }
2870 186 :
2871 186 : Ok(result)
2872 186 : }
2873 :
2874 : impl Tenant {
2875 0 : pub fn tenant_specific_overrides(&self) -> TenantConfOpt {
2876 0 : self.tenant_conf.load().tenant_conf.clone()
2877 0 : }
2878 :
2879 0 : pub fn effective_config(&self) -> TenantConf {
2880 0 : self.tenant_specific_overrides()
2881 0 : .merge(self.conf.default_tenant_conf.clone())
2882 0 : }
2883 :
2884 0 : pub fn get_checkpoint_distance(&self) -> u64 {
2885 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2886 0 : tenant_conf
2887 0 : .checkpoint_distance
2888 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
2889 0 : }
2890 :
2891 0 : pub fn get_checkpoint_timeout(&self) -> Duration {
2892 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2893 0 : tenant_conf
2894 0 : .checkpoint_timeout
2895 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
2896 0 : }
2897 :
2898 0 : pub fn get_compaction_target_size(&self) -> u64 {
2899 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2900 0 : tenant_conf
2901 0 : .compaction_target_size
2902 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
2903 0 : }
2904 :
2905 0 : pub fn get_compaction_period(&self) -> Duration {
2906 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2907 0 : tenant_conf
2908 0 : .compaction_period
2909 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
2910 0 : }
2911 :
2912 0 : pub fn get_compaction_threshold(&self) -> usize {
2913 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2914 0 : tenant_conf
2915 0 : .compaction_threshold
2916 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
2917 0 : }
2918 :
2919 0 : pub fn get_gc_horizon(&self) -> u64 {
2920 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2921 0 : tenant_conf
2922 0 : .gc_horizon
2923 0 : .unwrap_or(self.conf.default_tenant_conf.gc_horizon)
2924 0 : }
2925 :
2926 0 : pub fn get_gc_period(&self) -> Duration {
2927 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2928 0 : tenant_conf
2929 0 : .gc_period
2930 0 : .unwrap_or(self.conf.default_tenant_conf.gc_period)
2931 0 : }
2932 :
2933 0 : pub fn get_image_creation_threshold(&self) -> usize {
2934 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2935 0 : tenant_conf
2936 0 : .image_creation_threshold
2937 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
2938 0 : }
2939 :
2940 0 : pub fn get_pitr_interval(&self) -> Duration {
2941 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2942 0 : tenant_conf
2943 0 : .pitr_interval
2944 0 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
2945 0 : }
2946 :
2947 0 : pub fn get_min_resident_size_override(&self) -> Option<u64> {
2948 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2949 0 : tenant_conf
2950 0 : .min_resident_size_override
2951 0 : .or(self.conf.default_tenant_conf.min_resident_size_override)
2952 0 : }
2953 :
2954 0 : pub fn get_heatmap_period(&self) -> Option<Duration> {
2955 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2956 0 : let heatmap_period = tenant_conf
2957 0 : .heatmap_period
2958 0 : .unwrap_or(self.conf.default_tenant_conf.heatmap_period);
2959 0 : if heatmap_period.is_zero() {
2960 0 : None
2961 : } else {
2962 0 : Some(heatmap_period)
2963 : }
2964 0 : }
2965 :
2966 4 : pub fn get_lsn_lease_length(&self) -> Duration {
2967 4 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2968 4 : tenant_conf
2969 4 : .lsn_lease_length
2970 4 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
2971 4 : }
2972 :
2973 0 : pub fn set_new_tenant_config(&self, new_tenant_conf: TenantConfOpt) {
2974 0 : // Use read-copy-update in order to avoid overwriting the location config
2975 0 : // state if this races with [`Tenant::set_new_location_config`]. Note that
2976 0 : // this race is not possible if both request types come from the storage
2977 0 : // controller (as they should!) because an exclusive op lock is required
2978 0 : // on the storage controller side.
2979 0 : self.tenant_conf.rcu(|inner| {
2980 0 : Arc::new(AttachedTenantConf {
2981 0 : tenant_conf: new_tenant_conf.clone(),
2982 0 : location: inner.location,
2983 0 : // Attached location is not changed, no need to update lsn lease deadline.
2984 0 : lsn_lease_deadline: inner.lsn_lease_deadline,
2985 0 : })
2986 0 : });
2987 0 :
2988 0 : self.tenant_conf_updated(&new_tenant_conf);
2989 0 : // Don't hold self.timelines.lock() during the notifies.
2990 0 : // There's no risk of deadlock right now, but there could be if we consolidate
2991 0 : // mutexes in struct Timeline in the future.
2992 0 : let timelines = self.list_timelines();
2993 0 : for timeline in timelines {
2994 0 : timeline.tenant_conf_updated(&new_tenant_conf);
2995 0 : }
2996 0 : }
2997 :
2998 0 : pub(crate) fn set_new_location_config(&self, new_conf: AttachedTenantConf) {
2999 0 : let new_tenant_conf = new_conf.tenant_conf.clone();
3000 0 :
3001 0 : self.tenant_conf.store(Arc::new(new_conf));
3002 0 :
3003 0 : self.tenant_conf_updated(&new_tenant_conf);
3004 0 : // Don't hold self.timelines.lock() during the notifies.
3005 0 : // There's no risk of deadlock right now, but there could be if we consolidate
3006 0 : // mutexes in struct Timeline in the future.
3007 0 : let timelines = self.list_timelines();
3008 0 : for timeline in timelines {
3009 0 : timeline.tenant_conf_updated(&new_tenant_conf);
3010 0 : }
3011 0 : }
3012 :
3013 186 : fn get_timeline_get_throttle_config(
3014 186 : psconf: &'static PageServerConf,
3015 186 : overrides: &TenantConfOpt,
3016 186 : ) -> throttle::Config {
3017 186 : overrides
3018 186 : .timeline_get_throttle
3019 186 : .clone()
3020 186 : .unwrap_or(psconf.default_tenant_conf.timeline_get_throttle.clone())
3021 186 : }
3022 :
3023 0 : pub(crate) fn tenant_conf_updated(&self, new_conf: &TenantConfOpt) {
3024 0 : let conf = Self::get_timeline_get_throttle_config(self.conf, new_conf);
3025 0 : self.timeline_get_throttle.reconfigure(conf)
3026 0 : }
3027 :
3028 : /// Helper function to create a new Timeline struct.
3029 : ///
3030 : /// The returned Timeline is in Loading state. The caller is responsible for
3031 : /// initializing any on-disk state, and for inserting the Timeline to the 'timelines'
3032 : /// map.
3033 : ///
3034 : /// `validate_ancestor == false` is used when a timeline is created for deletion
3035 : /// and we might not have the ancestor present anymore which is fine for to be
3036 : /// deleted timelines.
3037 410 : fn create_timeline_struct(
3038 410 : &self,
3039 410 : new_timeline_id: TimelineId,
3040 410 : new_metadata: &TimelineMetadata,
3041 410 : ancestor: Option<Arc<Timeline>>,
3042 410 : resources: TimelineResources,
3043 410 : cause: CreateTimelineCause,
3044 410 : ) -> anyhow::Result<Arc<Timeline>> {
3045 410 : let state = match cause {
3046 : CreateTimelineCause::Load => {
3047 410 : let ancestor_id = new_metadata.ancestor_timeline();
3048 410 : anyhow::ensure!(
3049 410 : ancestor_id == ancestor.as_ref().map(|t| t.timeline_id),
3050 0 : "Timeline's {new_timeline_id} ancestor {ancestor_id:?} was not found"
3051 : );
3052 410 : TimelineState::Loading
3053 : }
3054 0 : CreateTimelineCause::Delete => TimelineState::Stopping,
3055 : };
3056 :
3057 410 : let pg_version = new_metadata.pg_version();
3058 410 :
3059 410 : let timeline = Timeline::new(
3060 410 : self.conf,
3061 410 : Arc::clone(&self.tenant_conf),
3062 410 : new_metadata,
3063 410 : ancestor,
3064 410 : new_timeline_id,
3065 410 : self.tenant_shard_id,
3066 410 : self.generation,
3067 410 : self.shard_identity,
3068 410 : self.walredo_mgr.clone(),
3069 410 : resources,
3070 410 : pg_version,
3071 410 : state,
3072 410 : self.attach_wal_lag_cooldown.clone(),
3073 410 : self.cancel.child_token(),
3074 410 : );
3075 410 :
3076 410 : Ok(timeline)
3077 410 : }
3078 :
3079 : // Allow too_many_arguments because a constructor's argument list naturally grows with the
3080 : // number of attributes in the struct: breaking these out into a builder wouldn't be helpful.
3081 : #[allow(clippy::too_many_arguments)]
3082 186 : fn new(
3083 186 : state: TenantState,
3084 186 : conf: &'static PageServerConf,
3085 186 : attached_conf: AttachedTenantConf,
3086 186 : shard_identity: ShardIdentity,
3087 186 : walredo_mgr: Option<Arc<WalRedoManager>>,
3088 186 : tenant_shard_id: TenantShardId,
3089 186 : remote_storage: GenericRemoteStorage,
3090 186 : deletion_queue_client: DeletionQueueClient,
3091 186 : l0_flush_global_state: L0FlushGlobalState,
3092 186 : ) -> Tenant {
3093 186 : debug_assert!(
3094 186 : !attached_conf.location.generation.is_none() || conf.control_plane_api.is_none()
3095 : );
3096 :
3097 186 : let (state, mut rx) = watch::channel(state);
3098 186 :
3099 186 : tokio::spawn(async move {
3100 186 : // reflect tenant state in metrics:
3101 186 : // - global per tenant state: TENANT_STATE_METRIC
3102 186 : // - "set" of broken tenants: BROKEN_TENANTS_SET
3103 186 : //
3104 186 : // set of broken tenants should not have zero counts so that it remains accessible for
3105 186 : // alerting.
3106 186 :
3107 186 : let tid = tenant_shard_id.to_string();
3108 186 : let shard_id = tenant_shard_id.shard_slug().to_string();
3109 186 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
3110 :
3111 370 : fn inspect_state(state: &TenantState) -> ([&'static str; 1], bool) {
3112 370 : ([state.into()], matches!(state, TenantState::Broken { .. }))
3113 370 : }
3114 :
3115 186 : let mut tuple = inspect_state(&rx.borrow_and_update());
3116 186 :
3117 186 : let is_broken = tuple.1;
3118 186 : let mut counted_broken = if is_broken {
3119 : // add the id to the set right away, there should not be any updates on the channel
3120 : // after before tenant is removed, if ever
3121 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
3122 0 : true
3123 : } else {
3124 186 : false
3125 : };
3126 :
3127 : loop {
3128 370 : let labels = &tuple.0;
3129 370 : let current = TENANT_STATE_METRIC.with_label_values(labels);
3130 370 : current.inc();
3131 370 :
3132 370 : if rx.changed().await.is_err() {
3133 : // tenant has been dropped
3134 16 : current.dec();
3135 16 : drop(BROKEN_TENANTS_SET.remove_label_values(set_key));
3136 16 : break;
3137 184 : }
3138 184 :
3139 184 : current.dec();
3140 184 : tuple = inspect_state(&rx.borrow_and_update());
3141 184 :
3142 184 : let is_broken = tuple.1;
3143 184 : if is_broken && !counted_broken {
3144 0 : counted_broken = true;
3145 0 : // insert the tenant_id (back) into the set while avoiding needless counter
3146 0 : // access
3147 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
3148 184 : }
3149 : }
3150 186 : });
3151 186 :
3152 186 : Tenant {
3153 186 : tenant_shard_id,
3154 186 : shard_identity,
3155 186 : generation: attached_conf.location.generation,
3156 186 : conf,
3157 186 : // using now here is good enough approximation to catch tenants with really long
3158 186 : // activation times.
3159 186 : constructed_at: Instant::now(),
3160 186 : timelines: Mutex::new(HashMap::new()),
3161 186 : timelines_creating: Mutex::new(HashSet::new()),
3162 186 : timelines_offloaded: Mutex::new(HashMap::new()),
3163 186 : gc_cs: tokio::sync::Mutex::new(()),
3164 186 : walredo_mgr,
3165 186 : remote_storage,
3166 186 : deletion_queue_client,
3167 186 : state,
3168 186 : cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
3169 186 : cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
3170 186 : eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
3171 186 : compaction_circuit_breaker: std::sync::Mutex::new(CircuitBreaker::new(
3172 186 : format!("compaction-{tenant_shard_id}"),
3173 186 : 5,
3174 186 : // Compaction can be a very expensive operation, and might leak disk space. It also ought
3175 186 : // to be infallible, as long as remote storage is available. So if it repeatedly fails,
3176 186 : // use an extremely long backoff.
3177 186 : Some(Duration::from_secs(3600 * 24)),
3178 186 : )),
3179 186 : activate_now_sem: tokio::sync::Semaphore::new(0),
3180 186 : attach_wal_lag_cooldown: Arc::new(std::sync::OnceLock::new()),
3181 186 : cancel: CancellationToken::default(),
3182 186 : gate: Gate::default(),
3183 186 : timeline_get_throttle: Arc::new(throttle::Throttle::new(
3184 186 : Tenant::get_timeline_get_throttle_config(conf, &attached_conf.tenant_conf),
3185 186 : crate::metrics::tenant_throttling::TimelineGet::new(&tenant_shard_id),
3186 186 : )),
3187 186 : tenant_conf: Arc::new(ArcSwap::from_pointee(attached_conf)),
3188 186 : ongoing_timeline_detach: std::sync::Mutex::default(),
3189 186 : gc_block: Default::default(),
3190 186 : l0_flush_global_state,
3191 186 : }
3192 186 : }
3193 :
3194 : /// Locate and load config
3195 0 : pub(super) fn load_tenant_config(
3196 0 : conf: &'static PageServerConf,
3197 0 : tenant_shard_id: &TenantShardId,
3198 0 : ) -> Result<LocationConf, LoadConfigError> {
3199 0 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
3200 0 :
3201 0 : info!("loading tenant configuration from {config_path}");
3202 :
3203 : // load and parse file
3204 0 : let config = fs::read_to_string(&config_path).map_err(|e| {
3205 0 : match e.kind() {
3206 : std::io::ErrorKind::NotFound => {
3207 : // The config should almost always exist for a tenant directory:
3208 : // - When attaching a tenant, the config is the first thing we write
3209 : // - When detaching a tenant, we atomically move the directory to a tmp location
3210 : // before deleting contents.
3211 : //
3212 : // The very rare edge case that can result in a missing config is if we crash during attach
3213 : // between creating directory and writing config. Callers should handle that as if the
3214 : // directory didn't exist.
3215 :
3216 0 : LoadConfigError::NotFound(config_path)
3217 : }
3218 : _ => {
3219 : // No IO errors except NotFound are acceptable here: other kinds of error indicate local storage or permissions issues
3220 : // that we cannot cleanly recover
3221 0 : crate::virtual_file::on_fatal_io_error(&e, "Reading tenant config file")
3222 : }
3223 : }
3224 0 : })?;
3225 :
3226 0 : Ok(toml_edit::de::from_str::<LocationConf>(&config)?)
3227 0 : }
3228 :
3229 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
3230 : pub(super) async fn persist_tenant_config(
3231 : conf: &'static PageServerConf,
3232 : tenant_shard_id: &TenantShardId,
3233 : location_conf: &LocationConf,
3234 : ) -> std::io::Result<()> {
3235 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
3236 :
3237 : Self::persist_tenant_config_at(tenant_shard_id, &config_path, location_conf).await
3238 : }
3239 :
3240 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
3241 : pub(super) async fn persist_tenant_config_at(
3242 : tenant_shard_id: &TenantShardId,
3243 : config_path: &Utf8Path,
3244 : location_conf: &LocationConf,
3245 : ) -> std::io::Result<()> {
3246 : debug!("persisting tenantconf to {config_path}");
3247 :
3248 : let mut conf_content = r#"# This file contains a specific per-tenant's config.
3249 : # It is read in case of pageserver restart.
3250 : "#
3251 : .to_string();
3252 :
3253 0 : fail::fail_point!("tenant-config-before-write", |_| {
3254 0 : Err(std::io::Error::new(
3255 0 : std::io::ErrorKind::Other,
3256 0 : "tenant-config-before-write",
3257 0 : ))
3258 0 : });
3259 :
3260 : // Convert the config to a toml file.
3261 : conf_content +=
3262 : &toml_edit::ser::to_string_pretty(&location_conf).expect("Config serialization failed");
3263 :
3264 : let temp_path = path_with_suffix_extension(config_path, TEMP_FILE_SUFFIX);
3265 :
3266 : let conf_content = conf_content.into_bytes();
3267 : VirtualFile::crashsafe_overwrite(config_path.to_owned(), temp_path, conf_content).await
3268 : }
3269 :
3270 : //
3271 : // How garbage collection works:
3272 : //
3273 : // +--bar------------->
3274 : // /
3275 : // +----+-----foo---------------->
3276 : // /
3277 : // ----main--+-------------------------->
3278 : // \
3279 : // +-----baz-------->
3280 : //
3281 : //
3282 : // 1. Grab 'gc_cs' mutex to prevent new timelines from being created while Timeline's
3283 : // `gc_infos` are being refreshed
3284 : // 2. Scan collected timelines, and on each timeline, make note of the
3285 : // all the points where other timelines have been branched off.
3286 : // We will refrain from removing page versions at those LSNs.
3287 : // 3. For each timeline, scan all layer files on the timeline.
3288 : // Remove all files for which a newer file exists and which
3289 : // don't cover any branch point LSNs.
3290 : //
3291 : // TODO:
3292 : // - if a relation has a non-incremental persistent layer on a child branch, then we
3293 : // don't need to keep that in the parent anymore. But currently
3294 : // we do.
3295 4 : async fn gc_iteration_internal(
3296 4 : &self,
3297 4 : target_timeline_id: Option<TimelineId>,
3298 4 : horizon: u64,
3299 4 : pitr: Duration,
3300 4 : cancel: &CancellationToken,
3301 4 : ctx: &RequestContext,
3302 4 : ) -> Result<GcResult, GcError> {
3303 4 : let mut totals: GcResult = Default::default();
3304 4 : let now = Instant::now();
3305 :
3306 4 : let gc_timelines = self
3307 4 : .refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
3308 4 : .await?;
3309 :
3310 4 : failpoint_support::sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
3311 :
3312 : // If there is nothing to GC, we don't want any messages in the INFO log.
3313 4 : if !gc_timelines.is_empty() {
3314 4 : info!("{} timelines need GC", gc_timelines.len());
3315 : } else {
3316 0 : debug!("{} timelines need GC", gc_timelines.len());
3317 : }
3318 :
3319 : // Perform GC for each timeline.
3320 : //
3321 : // Note that we don't hold the `Tenant::gc_cs` lock here because we don't want to delay the
3322 : // branch creation task, which requires the GC lock. A GC iteration can run concurrently
3323 : // with branch creation.
3324 : //
3325 : // See comments in [`Tenant::branch_timeline`] for more information about why branch
3326 : // creation task can run concurrently with timeline's GC iteration.
3327 8 : for timeline in gc_timelines {
3328 4 : if cancel.is_cancelled() {
3329 : // We were requested to shut down. Stop and return with the progress we
3330 : // made.
3331 0 : break;
3332 4 : }
3333 4 : let result = match timeline.gc().await {
3334 : Err(GcError::TimelineCancelled) => {
3335 0 : if target_timeline_id.is_some() {
3336 : // If we were targetting this specific timeline, surface cancellation to caller
3337 0 : return Err(GcError::TimelineCancelled);
3338 : } else {
3339 : // A timeline may be shutting down independently of the tenant's lifecycle: we should
3340 : // skip past this and proceed to try GC on other timelines.
3341 0 : continue;
3342 : }
3343 : }
3344 4 : r => r?,
3345 : };
3346 4 : totals += result;
3347 : }
3348 :
3349 4 : totals.elapsed = now.elapsed();
3350 4 : Ok(totals)
3351 4 : }
3352 :
3353 : /// Refreshes the Timeline::gc_info for all timelines, returning the
3354 : /// vector of timelines which have [`Timeline::get_last_record_lsn`] past
3355 : /// [`Tenant::get_gc_horizon`].
3356 : ///
3357 : /// This is usually executed as part of periodic gc, but can now be triggered more often.
3358 0 : pub(crate) async fn refresh_gc_info(
3359 0 : &self,
3360 0 : cancel: &CancellationToken,
3361 0 : ctx: &RequestContext,
3362 0 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
3363 0 : // since this method can now be called at different rates than the configured gc loop, it
3364 0 : // might be that these configuration values get applied faster than what it was previously,
3365 0 : // since these were only read from the gc task.
3366 0 : let horizon = self.get_gc_horizon();
3367 0 : let pitr = self.get_pitr_interval();
3368 0 :
3369 0 : // refresh all timelines
3370 0 : let target_timeline_id = None;
3371 0 :
3372 0 : self.refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
3373 0 : .await
3374 0 : }
3375 :
3376 : /// Populate all Timelines' `GcInfo` with information about their children. We do not set the
3377 : /// PITR cutoffs here, because that requires I/O: this is done later, before GC, by [`Self::refresh_gc_info_internal`]
3378 : ///
3379 : /// Subsequently, parent-child relationships are updated incrementally inside [`Timeline::new`] and [`Timeline::drop`].
3380 0 : fn initialize_gc_info(
3381 0 : &self,
3382 0 : timelines: &std::sync::MutexGuard<HashMap<TimelineId, Arc<Timeline>>>,
3383 0 : timelines_offloaded: &std::sync::MutexGuard<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
3384 0 : ) {
3385 0 : // This function must be called before activation: after activation timeline create/delete operations
3386 0 : // might happen, and this function is not safe to run concurrently with those.
3387 0 : assert!(!self.is_active());
3388 :
3389 : // Scan all timelines. For each timeline, remember the timeline ID and
3390 : // the branch point where it was created.
3391 0 : let mut all_branchpoints: BTreeMap<TimelineId, Vec<(Lsn, TimelineId, MaybeOffloaded)>> =
3392 0 : BTreeMap::new();
3393 0 : timelines.iter().for_each(|(timeline_id, timeline_entry)| {
3394 0 : if let Some(ancestor_timeline_id) = &timeline_entry.get_ancestor_timeline_id() {
3395 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
3396 0 : ancestor_children.push((
3397 0 : timeline_entry.get_ancestor_lsn(),
3398 0 : *timeline_id,
3399 0 : MaybeOffloaded::No,
3400 0 : ));
3401 0 : }
3402 0 : });
3403 0 : timelines_offloaded
3404 0 : .iter()
3405 0 : .for_each(|(timeline_id, timeline_entry)| {
3406 0 : let Some(ancestor_timeline_id) = &timeline_entry.ancestor_timeline_id else {
3407 0 : return;
3408 : };
3409 0 : let Some(retain_lsn) = timeline_entry.ancestor_retain_lsn else {
3410 0 : return;
3411 : };
3412 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
3413 0 : ancestor_children.push((retain_lsn, *timeline_id, MaybeOffloaded::Yes));
3414 0 : });
3415 0 :
3416 0 : // The number of bytes we always keep, irrespective of PITR: this is a constant across timelines
3417 0 : let horizon = self.get_gc_horizon();
3418 :
3419 : // Populate each timeline's GcInfo with information about its child branches
3420 0 : for timeline in timelines.values() {
3421 0 : let mut branchpoints: Vec<(Lsn, TimelineId, MaybeOffloaded)> = all_branchpoints
3422 0 : .remove(&timeline.timeline_id)
3423 0 : .unwrap_or_default();
3424 0 :
3425 0 : branchpoints.sort_by_key(|b| b.0);
3426 0 :
3427 0 : let mut target = timeline.gc_info.write().unwrap();
3428 0 :
3429 0 : target.retain_lsns = branchpoints;
3430 0 :
3431 0 : let space_cutoff = timeline
3432 0 : .get_last_record_lsn()
3433 0 : .checked_sub(horizon)
3434 0 : .unwrap_or(Lsn(0));
3435 0 :
3436 0 : target.cutoffs = GcCutoffs {
3437 0 : space: space_cutoff,
3438 0 : time: Lsn::INVALID,
3439 0 : };
3440 0 : }
3441 0 : }
3442 :
3443 4 : async fn refresh_gc_info_internal(
3444 4 : &self,
3445 4 : target_timeline_id: Option<TimelineId>,
3446 4 : horizon: u64,
3447 4 : pitr: Duration,
3448 4 : cancel: &CancellationToken,
3449 4 : ctx: &RequestContext,
3450 4 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
3451 4 : // before taking the gc_cs lock, do the heavier weight finding of gc_cutoff points for
3452 4 : // currently visible timelines.
3453 4 : let timelines = self
3454 4 : .timelines
3455 4 : .lock()
3456 4 : .unwrap()
3457 4 : .values()
3458 4 : .filter(|tl| match target_timeline_id.as_ref() {
3459 4 : Some(target) => &tl.timeline_id == target,
3460 0 : None => true,
3461 4 : })
3462 4 : .cloned()
3463 4 : .collect::<Vec<_>>();
3464 4 :
3465 4 : if target_timeline_id.is_some() && timelines.is_empty() {
3466 : // We were to act on a particular timeline and it wasn't found
3467 0 : return Err(GcError::TimelineNotFound);
3468 4 : }
3469 4 :
3470 4 : let mut gc_cutoffs: HashMap<TimelineId, GcCutoffs> =
3471 4 : HashMap::with_capacity(timelines.len());
3472 :
3473 4 : for timeline in timelines.iter() {
3474 4 : let cutoff = timeline
3475 4 : .get_last_record_lsn()
3476 4 : .checked_sub(horizon)
3477 4 : .unwrap_or(Lsn(0));
3478 :
3479 4 : let cutoffs = timeline.find_gc_cutoffs(cutoff, pitr, cancel, ctx).await?;
3480 4 : let old = gc_cutoffs.insert(timeline.timeline_id, cutoffs);
3481 4 : assert!(old.is_none());
3482 : }
3483 :
3484 4 : if !self.is_active() || self.cancel.is_cancelled() {
3485 0 : return Err(GcError::TenantCancelled);
3486 4 : }
3487 :
3488 : // grab mutex to prevent new timelines from being created here; avoid doing long operations
3489 : // because that will stall branch creation.
3490 4 : let gc_cs = self.gc_cs.lock().await;
3491 :
3492 : // Ok, we now know all the branch points.
3493 : // Update the GC information for each timeline.
3494 4 : let mut gc_timelines = Vec::with_capacity(timelines.len());
3495 8 : for timeline in timelines {
3496 : // We filtered the timeline list above
3497 4 : if let Some(target_timeline_id) = target_timeline_id {
3498 4 : assert_eq!(target_timeline_id, timeline.timeline_id);
3499 0 : }
3500 :
3501 : {
3502 4 : let mut target = timeline.gc_info.write().unwrap();
3503 4 :
3504 4 : // Cull any expired leases
3505 4 : let now = SystemTime::now();
3506 6 : target.leases.retain(|_, lease| !lease.is_expired(&now));
3507 4 :
3508 4 : timeline
3509 4 : .metrics
3510 4 : .valid_lsn_lease_count_gauge
3511 4 : .set(target.leases.len() as u64);
3512 :
3513 : // Look up parent's PITR cutoff to update the child's knowledge of whether it is within parent's PITR
3514 4 : if let Some(ancestor_id) = timeline.get_ancestor_timeline_id() {
3515 0 : if let Some(ancestor_gc_cutoffs) = gc_cutoffs.get(&ancestor_id) {
3516 0 : target.within_ancestor_pitr =
3517 0 : timeline.get_ancestor_lsn() >= ancestor_gc_cutoffs.time;
3518 0 : }
3519 4 : }
3520 :
3521 : // Update metrics that depend on GC state
3522 4 : timeline
3523 4 : .metrics
3524 4 : .archival_size
3525 4 : .set(if target.within_ancestor_pitr {
3526 0 : timeline.metrics.current_logical_size_gauge.get()
3527 : } else {
3528 4 : 0
3529 : });
3530 4 : timeline.metrics.pitr_history_size.set(
3531 4 : timeline
3532 4 : .get_last_record_lsn()
3533 4 : .checked_sub(target.cutoffs.time)
3534 4 : .unwrap_or(Lsn(0))
3535 4 : .0,
3536 4 : );
3537 :
3538 : // Apply the cutoffs we found to the Timeline's GcInfo. Why might we _not_ have cutoffs for a timeline?
3539 : // - this timeline was created while we were finding cutoffs
3540 : // - lsn for timestamp search fails for this timeline repeatedly
3541 4 : if let Some(cutoffs) = gc_cutoffs.get(&timeline.timeline_id) {
3542 4 : target.cutoffs = cutoffs.clone();
3543 4 : }
3544 : }
3545 :
3546 4 : gc_timelines.push(timeline);
3547 : }
3548 4 : drop(gc_cs);
3549 4 : Ok(gc_timelines)
3550 4 : }
3551 :
3552 : /// A substitute for `branch_timeline` for use in unit tests.
3553 : /// The returned timeline will have state value `Active` to make various `anyhow::ensure!()`
3554 : /// calls pass, but, we do not actually call `.activate()` under the hood. So, none of the
3555 : /// timeline background tasks are launched, except the flush loop.
3556 : #[cfg(test)]
3557 230 : async fn branch_timeline_test(
3558 230 : &self,
3559 230 : src_timeline: &Arc<Timeline>,
3560 230 : dst_id: TimelineId,
3561 230 : ancestor_lsn: Option<Lsn>,
3562 230 : ctx: &RequestContext,
3563 230 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
3564 230 : let create_guard = self.create_timeline_create_guard(dst_id).unwrap();
3565 230 : let tl = self
3566 230 : .branch_timeline_impl(src_timeline, dst_id, ancestor_lsn, create_guard, ctx)
3567 4 : .await?;
3568 226 : tl.set_state(TimelineState::Active);
3569 226 : Ok(tl)
3570 230 : }
3571 :
3572 : /// Helper for unit tests to branch a timeline with some pre-loaded states.
3573 : #[cfg(test)]
3574 : #[allow(clippy::too_many_arguments)]
3575 6 : pub async fn branch_timeline_test_with_layers(
3576 6 : &self,
3577 6 : src_timeline: &Arc<Timeline>,
3578 6 : dst_id: TimelineId,
3579 6 : ancestor_lsn: Option<Lsn>,
3580 6 : ctx: &RequestContext,
3581 6 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
3582 6 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
3583 6 : end_lsn: Lsn,
3584 6 : ) -> anyhow::Result<Arc<Timeline>> {
3585 : use checks::check_valid_layermap;
3586 : use itertools::Itertools;
3587 :
3588 6 : let tline = self
3589 6 : .branch_timeline_test(src_timeline, dst_id, ancestor_lsn, ctx)
3590 0 : .await?;
3591 6 : let ancestor_lsn = if let Some(ancestor_lsn) = ancestor_lsn {
3592 6 : ancestor_lsn
3593 : } else {
3594 0 : tline.get_last_record_lsn()
3595 : };
3596 6 : assert!(end_lsn >= ancestor_lsn);
3597 6 : tline.force_advance_lsn(end_lsn);
3598 12 : for deltas in delta_layer_desc {
3599 6 : tline
3600 6 : .force_create_delta_layer(deltas, Some(ancestor_lsn), ctx)
3601 18 : .await?;
3602 : }
3603 10 : for (lsn, images) in image_layer_desc {
3604 4 : tline
3605 4 : .force_create_image_layer(lsn, images, Some(ancestor_lsn), ctx)
3606 14 : .await?;
3607 : }
3608 6 : let layer_names = tline
3609 6 : .layers
3610 6 : .read()
3611 0 : .await
3612 6 : .layer_map()
3613 6 : .unwrap()
3614 6 : .iter_historic_layers()
3615 10 : .map(|layer| layer.layer_name())
3616 6 : .collect_vec();
3617 6 : if let Some(err) = check_valid_layermap(&layer_names) {
3618 0 : bail!("invalid layermap: {err}");
3619 6 : }
3620 6 : Ok(tline)
3621 6 : }
3622 :
3623 : /// Branch an existing timeline.
3624 : ///
3625 : /// The caller is responsible for activating the returned timeline.
3626 0 : async fn branch_timeline(
3627 0 : &self,
3628 0 : src_timeline: &Arc<Timeline>,
3629 0 : dst_id: TimelineId,
3630 0 : start_lsn: Option<Lsn>,
3631 0 : timeline_create_guard: TimelineCreateGuard<'_>,
3632 0 : ctx: &RequestContext,
3633 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
3634 0 : self.branch_timeline_impl(src_timeline, dst_id, start_lsn, timeline_create_guard, ctx)
3635 0 : .await
3636 0 : }
3637 :
3638 230 : async fn branch_timeline_impl(
3639 230 : &self,
3640 230 : src_timeline: &Arc<Timeline>,
3641 230 : dst_id: TimelineId,
3642 230 : start_lsn: Option<Lsn>,
3643 230 : timeline_create_guard: TimelineCreateGuard<'_>,
3644 230 : _ctx: &RequestContext,
3645 230 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
3646 230 : let src_id = src_timeline.timeline_id;
3647 :
3648 : // We will validate our ancestor LSN in this function. Acquire the GC lock so that
3649 : // this check cannot race with GC, and the ancestor LSN is guaranteed to remain
3650 : // valid while we are creating the branch.
3651 230 : let _gc_cs = self.gc_cs.lock().await;
3652 :
3653 : // If no start LSN is specified, we branch the new timeline from the source timeline's last record LSN
3654 230 : let start_lsn = start_lsn.unwrap_or_else(|| {
3655 2 : let lsn = src_timeline.get_last_record_lsn();
3656 2 : info!("branching timeline {dst_id} from timeline {src_id} at last record LSN: {lsn}");
3657 2 : lsn
3658 230 : });
3659 230 :
3660 230 : // Ensure that `start_lsn` is valid, i.e. the LSN is within the PITR
3661 230 : // horizon on the source timeline
3662 230 : //
3663 230 : // We check it against both the planned GC cutoff stored in 'gc_info',
3664 230 : // and the 'latest_gc_cutoff' of the last GC that was performed. The
3665 230 : // planned GC cutoff in 'gc_info' is normally larger than
3666 230 : // 'latest_gc_cutoff_lsn', but beware of corner cases like if you just
3667 230 : // changed the GC settings for the tenant to make the PITR window
3668 230 : // larger, but some of the data was already removed by an earlier GC
3669 230 : // iteration.
3670 230 :
3671 230 : // check against last actual 'latest_gc_cutoff' first
3672 230 : let latest_gc_cutoff_lsn = src_timeline.get_latest_gc_cutoff_lsn();
3673 230 : src_timeline
3674 230 : .check_lsn_is_in_scope(start_lsn, &latest_gc_cutoff_lsn)
3675 230 : .context(format!(
3676 230 : "invalid branch start lsn: less than latest GC cutoff {}",
3677 230 : *latest_gc_cutoff_lsn,
3678 230 : ))
3679 230 : .map_err(CreateTimelineError::AncestorLsn)?;
3680 :
3681 : // and then the planned GC cutoff
3682 : {
3683 226 : let gc_info = src_timeline.gc_info.read().unwrap();
3684 226 : let cutoff = gc_info.min_cutoff();
3685 226 : if start_lsn < cutoff {
3686 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
3687 0 : "invalid branch start lsn: less than planned GC cutoff {cutoff}"
3688 0 : )));
3689 226 : }
3690 226 : }
3691 226 :
3692 226 : //
3693 226 : // The branch point is valid, and we are still holding the 'gc_cs' lock
3694 226 : // so that GC cannot advance the GC cutoff until we are finished.
3695 226 : // Proceed with the branch creation.
3696 226 : //
3697 226 :
3698 226 : // Determine prev-LSN for the new timeline. We can only determine it if
3699 226 : // the timeline was branched at the current end of the source timeline.
3700 226 : let RecordLsn {
3701 226 : last: src_last,
3702 226 : prev: src_prev,
3703 226 : } = src_timeline.get_last_record_rlsn();
3704 226 : let dst_prev = if src_last == start_lsn {
3705 216 : Some(src_prev)
3706 : } else {
3707 10 : None
3708 : };
3709 :
3710 : // Create the metadata file, noting the ancestor of the new timeline.
3711 : // There is initially no data in it, but all the read-calls know to look
3712 : // into the ancestor.
3713 226 : let metadata = TimelineMetadata::new(
3714 226 : start_lsn,
3715 226 : dst_prev,
3716 226 : Some(src_id),
3717 226 : start_lsn,
3718 226 : *src_timeline.latest_gc_cutoff_lsn.read(), // FIXME: should we hold onto this guard longer?
3719 226 : src_timeline.initdb_lsn,
3720 226 : src_timeline.pg_version,
3721 226 : );
3722 :
3723 226 : let uninitialized_timeline = self
3724 226 : .prepare_new_timeline(
3725 226 : dst_id,
3726 226 : &metadata,
3727 226 : timeline_create_guard,
3728 226 : start_lsn + 1,
3729 226 : Some(Arc::clone(src_timeline)),
3730 226 : )
3731 0 : .await?;
3732 :
3733 226 : let new_timeline = uninitialized_timeline.finish_creation()?;
3734 :
3735 : // Root timeline gets its layers during creation and uploads them along with the metadata.
3736 : // A branch timeline though, when created, can get no writes for some time, hence won't get any layers created.
3737 : // We still need to upload its metadata eagerly: if other nodes `attach` the tenant and miss this timeline, their GC
3738 : // could get incorrect information and remove more layers, than needed.
3739 : // See also https://github.com/neondatabase/neon/issues/3865
3740 226 : new_timeline
3741 226 : .remote_client
3742 226 : .schedule_index_upload_for_full_metadata_update(&metadata)
3743 226 : .context("branch initial metadata upload")?;
3744 :
3745 226 : Ok(new_timeline)
3746 230 : }
3747 :
3748 : /// For unit tests, make this visible so that other modules can directly create timelines
3749 : #[cfg(test)]
3750 2 : #[tracing::instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
3751 : pub(crate) async fn bootstrap_timeline_test(
3752 : &self,
3753 : timeline_id: TimelineId,
3754 : pg_version: u32,
3755 : load_existing_initdb: Option<TimelineId>,
3756 : ctx: &RequestContext,
3757 : ) -> anyhow::Result<Arc<Timeline>> {
3758 : let create_guard = self.create_timeline_create_guard(timeline_id).unwrap();
3759 : self.bootstrap_timeline(
3760 : timeline_id,
3761 : pg_version,
3762 : load_existing_initdb,
3763 : create_guard,
3764 : ctx,
3765 : )
3766 : .await
3767 : }
3768 :
3769 0 : async fn upload_initdb(
3770 0 : &self,
3771 0 : timelines_path: &Utf8PathBuf,
3772 0 : pgdata_path: &Utf8PathBuf,
3773 0 : timeline_id: &TimelineId,
3774 0 : ) -> anyhow::Result<()> {
3775 0 : let temp_path = timelines_path.join(format!(
3776 0 : "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
3777 0 : ));
3778 0 :
3779 0 : scopeguard::defer! {
3780 0 : if let Err(e) = fs::remove_file(&temp_path) {
3781 0 : error!("Failed to remove temporary initdb archive '{temp_path}': {e}");
3782 0 : }
3783 0 : }
3784 :
3785 0 : let (pgdata_zstd, tar_zst_size) = create_zst_tarball(pgdata_path, &temp_path).await?;
3786 : const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
3787 0 : if tar_zst_size > INITDB_TAR_ZST_WARN_LIMIT {
3788 0 : warn!(
3789 0 : "compressed {temp_path} size of {tar_zst_size} is above limit {INITDB_TAR_ZST_WARN_LIMIT}."
3790 : );
3791 0 : }
3792 :
3793 0 : pausable_failpoint!("before-initdb-upload");
3794 :
3795 0 : backoff::retry(
3796 0 : || async {
3797 0 : self::remote_timeline_client::upload_initdb_dir(
3798 0 : &self.remote_storage,
3799 0 : &self.tenant_shard_id.tenant_id,
3800 0 : timeline_id,
3801 0 : pgdata_zstd.try_clone().await?,
3802 0 : tar_zst_size,
3803 0 : &self.cancel,
3804 : )
3805 0 : .await
3806 0 : },
3807 0 : |_| false,
3808 0 : 3,
3809 0 : u32::MAX,
3810 0 : "persist_initdb_tar_zst",
3811 0 : &self.cancel,
3812 0 : )
3813 0 : .await
3814 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
3815 0 : .and_then(|x| x)
3816 0 : }
3817 :
3818 : /// - run initdb to init temporary instance and get bootstrap data
3819 : /// - after initialization completes, tar up the temp dir and upload it to S3.
3820 : ///
3821 : /// The caller is responsible for activating the returned timeline.
3822 2 : async fn bootstrap_timeline(
3823 2 : &self,
3824 2 : timeline_id: TimelineId,
3825 2 : pg_version: u32,
3826 2 : load_existing_initdb: Option<TimelineId>,
3827 2 : timeline_create_guard: TimelineCreateGuard<'_>,
3828 2 : ctx: &RequestContext,
3829 2 : ) -> anyhow::Result<Arc<Timeline>> {
3830 2 : // create a `tenant/{tenant_id}/timelines/basebackup-{timeline_id}.{TEMP_FILE_SUFFIX}/`
3831 2 : // temporary directory for basebackup files for the given timeline.
3832 2 :
3833 2 : let timelines_path = self.conf.timelines_path(&self.tenant_shard_id);
3834 2 : let pgdata_path = path_with_suffix_extension(
3835 2 : timelines_path.join(format!("basebackup-{timeline_id}")),
3836 2 : TEMP_FILE_SUFFIX,
3837 2 : );
3838 2 :
3839 2 : // Remove whatever was left from the previous runs: safe because TimelineCreateGuard guarantees
3840 2 : // we won't race with other creations or existent timelines with the same path.
3841 2 : if pgdata_path.exists() {
3842 0 : fs::remove_dir_all(&pgdata_path).with_context(|| {
3843 0 : format!("Failed to remove already existing initdb directory: {pgdata_path}")
3844 0 : })?;
3845 2 : }
3846 :
3847 : // this new directory is very temporary, set to remove it immediately after bootstrap, we don't need it
3848 2 : scopeguard::defer! {
3849 2 : if let Err(e) = fs::remove_dir_all(&pgdata_path) {
3850 2 : // this is unlikely, but we will remove the directory on pageserver restart or another bootstrap call
3851 2 : error!("Failed to remove temporary initdb directory '{pgdata_path}': {e}");
3852 2 : }
3853 2 : }
3854 2 : if let Some(existing_initdb_timeline_id) = load_existing_initdb {
3855 2 : if existing_initdb_timeline_id != timeline_id {
3856 0 : let source_path = &remote_initdb_archive_path(
3857 0 : &self.tenant_shard_id.tenant_id,
3858 0 : &existing_initdb_timeline_id,
3859 0 : );
3860 0 : let dest_path =
3861 0 : &remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &timeline_id);
3862 0 :
3863 0 : // if this fails, it will get retried by retried control plane requests
3864 0 : self.remote_storage
3865 0 : .copy_object(source_path, dest_path, &self.cancel)
3866 0 : .await
3867 0 : .context("copy initdb tar")?;
3868 2 : }
3869 2 : let (initdb_tar_zst_path, initdb_tar_zst) =
3870 2 : self::remote_timeline_client::download_initdb_tar_zst(
3871 2 : self.conf,
3872 2 : &self.remote_storage,
3873 2 : &self.tenant_shard_id,
3874 2 : &existing_initdb_timeline_id,
3875 2 : &self.cancel,
3876 2 : )
3877 363 : .await
3878 2 : .context("download initdb tar")?;
3879 :
3880 2 : scopeguard::defer! {
3881 2 : if let Err(e) = fs::remove_file(&initdb_tar_zst_path) {
3882 2 : error!("Failed to remove temporary initdb archive '{initdb_tar_zst_path}': {e}");
3883 2 : }
3884 2 : }
3885 2 :
3886 2 : let buf_read =
3887 2 : BufReader::with_capacity(remote_timeline_client::BUFFER_SIZE, initdb_tar_zst);
3888 2 : extract_zst_tarball(&pgdata_path, buf_read)
3889 10470 : .await
3890 2 : .context("extract initdb tar")?;
3891 : } else {
3892 : // Init temporarily repo to get bootstrap data, this creates a directory in the `pgdata_path` path
3893 0 : run_initdb(self.conf, &pgdata_path, pg_version, &self.cancel).await?;
3894 :
3895 : // Upload the created data dir to S3
3896 0 : if self.tenant_shard_id().is_shard_zero() {
3897 0 : self.upload_initdb(&timelines_path, &pgdata_path, &timeline_id)
3898 0 : .await?;
3899 0 : }
3900 : }
3901 2 : let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
3902 2 :
3903 2 : // Import the contents of the data directory at the initial checkpoint
3904 2 : // LSN, and any WAL after that.
3905 2 : // Initdb lsn will be equal to last_record_lsn which will be set after import.
3906 2 : // Because we know it upfront avoid having an option or dummy zero value by passing it to the metadata.
3907 2 : let new_metadata = TimelineMetadata::new(
3908 2 : Lsn(0),
3909 2 : None,
3910 2 : None,
3911 2 : Lsn(0),
3912 2 : pgdata_lsn,
3913 2 : pgdata_lsn,
3914 2 : pg_version,
3915 2 : );
3916 2 : let raw_timeline = self
3917 2 : .prepare_new_timeline(
3918 2 : timeline_id,
3919 2 : &new_metadata,
3920 2 : timeline_create_guard,
3921 2 : pgdata_lsn,
3922 2 : None,
3923 2 : )
3924 0 : .await?;
3925 :
3926 2 : let tenant_shard_id = raw_timeline.owning_tenant.tenant_shard_id;
3927 2 : let unfinished_timeline = raw_timeline.raw_timeline()?;
3928 :
3929 : // Flush the new layer files to disk, before we make the timeline as available to
3930 : // the outside world.
3931 : //
3932 : // Flush loop needs to be spawned in order to be able to flush.
3933 2 : unfinished_timeline.maybe_spawn_flush_loop();
3934 2 :
3935 2 : import_datadir::import_timeline_from_postgres_datadir(
3936 2 : unfinished_timeline,
3937 2 : &pgdata_path,
3938 2 : pgdata_lsn,
3939 2 : ctx,
3940 2 : )
3941 9122 : .await
3942 2 : .with_context(|| {
3943 0 : format!("Failed to import pgdatadir for timeline {tenant_shard_id}/{timeline_id}")
3944 2 : })?;
3945 :
3946 2 : fail::fail_point!("before-checkpoint-new-timeline", |_| {
3947 0 : anyhow::bail!("failpoint before-checkpoint-new-timeline");
3948 2 : });
3949 :
3950 2 : unfinished_timeline
3951 2 : .freeze_and_flush()
3952 2 : .await
3953 2 : .with_context(|| {
3954 0 : format!(
3955 0 : "Failed to flush after pgdatadir import for timeline {tenant_shard_id}/{timeline_id}"
3956 0 : )
3957 2 : })?;
3958 :
3959 : // All done!
3960 2 : let timeline = raw_timeline.finish_creation()?;
3961 :
3962 2 : Ok(timeline)
3963 2 : }
3964 :
3965 : /// Call this before constructing a timeline, to build its required structures
3966 404 : fn build_timeline_resources(&self, timeline_id: TimelineId) -> TimelineResources {
3967 404 : let remote_client = RemoteTimelineClient::new(
3968 404 : self.remote_storage.clone(),
3969 404 : self.deletion_queue_client.clone(),
3970 404 : self.conf,
3971 404 : self.tenant_shard_id,
3972 404 : timeline_id,
3973 404 : self.generation,
3974 404 : );
3975 404 : TimelineResources {
3976 404 : remote_client,
3977 404 : timeline_get_throttle: self.timeline_get_throttle.clone(),
3978 404 : l0_flush_global_state: self.l0_flush_global_state.clone(),
3979 404 : }
3980 404 : }
3981 :
3982 : /// Creates intermediate timeline structure and its files.
3983 : ///
3984 : /// An empty layer map is initialized, and new data and WAL can be imported starting
3985 : /// at 'disk_consistent_lsn'. After any initial data has been imported, call
3986 : /// `finish_creation` to insert the Timeline into the timelines map.
3987 404 : async fn prepare_new_timeline<'a>(
3988 404 : &'a self,
3989 404 : new_timeline_id: TimelineId,
3990 404 : new_metadata: &TimelineMetadata,
3991 404 : create_guard: TimelineCreateGuard<'a>,
3992 404 : start_lsn: Lsn,
3993 404 : ancestor: Option<Arc<Timeline>>,
3994 404 : ) -> anyhow::Result<UninitializedTimeline<'a>> {
3995 404 : let tenant_shard_id = self.tenant_shard_id;
3996 404 :
3997 404 : let resources = self.build_timeline_resources(new_timeline_id);
3998 404 : resources
3999 404 : .remote_client
4000 404 : .init_upload_queue_for_empty_remote(new_metadata)?;
4001 :
4002 404 : let timeline_struct = self
4003 404 : .create_timeline_struct(
4004 404 : new_timeline_id,
4005 404 : new_metadata,
4006 404 : ancestor,
4007 404 : resources,
4008 404 : CreateTimelineCause::Load,
4009 404 : )
4010 404 : .context("Failed to create timeline data structure")?;
4011 :
4012 404 : timeline_struct.init_empty_layer_map(start_lsn);
4013 :
4014 404 : if let Err(e) = self
4015 404 : .create_timeline_files(&create_guard.timeline_path)
4016 0 : .await
4017 : {
4018 0 : error!("Failed to create initial files for timeline {tenant_shard_id}/{new_timeline_id}, cleaning up: {e:?}");
4019 0 : cleanup_timeline_directory(create_guard);
4020 0 : return Err(e);
4021 404 : }
4022 404 :
4023 404 : debug!(
4024 0 : "Successfully created initial files for timeline {tenant_shard_id}/{new_timeline_id}"
4025 : );
4026 :
4027 404 : Ok(UninitializedTimeline::new(
4028 404 : self,
4029 404 : new_timeline_id,
4030 404 : Some((timeline_struct, create_guard)),
4031 404 : ))
4032 404 : }
4033 :
4034 404 : async fn create_timeline_files(&self, timeline_path: &Utf8Path) -> anyhow::Result<()> {
4035 404 : crashsafe::create_dir(timeline_path).context("Failed to create timeline directory")?;
4036 :
4037 404 : fail::fail_point!("after-timeline-dir-creation", |_| {
4038 0 : anyhow::bail!("failpoint after-timeline-dir-creation");
4039 404 : });
4040 :
4041 404 : Ok(())
4042 404 : }
4043 :
4044 : /// Get a guard that provides exclusive access to the timeline directory, preventing
4045 : /// concurrent attempts to create the same timeline.
4046 410 : fn create_timeline_create_guard(
4047 410 : &self,
4048 410 : timeline_id: TimelineId,
4049 410 : ) -> Result<TimelineCreateGuard, TimelineExclusionError> {
4050 410 : let tenant_shard_id = self.tenant_shard_id;
4051 410 :
4052 410 : let timeline_path = self.conf.timeline_path(&tenant_shard_id, &timeline_id);
4053 :
4054 410 : let create_guard = TimelineCreateGuard::new(self, timeline_id, timeline_path.clone())?;
4055 :
4056 : // At this stage, we have got exclusive access to in-memory state for this timeline ID
4057 : // for creation.
4058 : // A timeline directory should never exist on disk already:
4059 : // - a previous failed creation would have cleaned up after itself
4060 : // - a pageserver restart would clean up timeline directories that don't have valid remote state
4061 : //
4062 : // Therefore it is an unexpected internal error to encounter a timeline directory already existing here,
4063 : // this error may indicate a bug in cleanup on failed creations.
4064 408 : if timeline_path.exists() {
4065 0 : return Err(TimelineExclusionError::Other(anyhow::anyhow!(
4066 0 : "Timeline directory already exists! This is a bug."
4067 0 : )));
4068 408 : }
4069 408 :
4070 408 : Ok(create_guard)
4071 410 : }
4072 :
4073 : /// Gathers inputs from all of the timelines to produce a sizing model input.
4074 : ///
4075 : /// Future is cancellation safe. Only one calculation can be running at once per tenant.
4076 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
4077 : pub async fn gather_size_inputs(
4078 : &self,
4079 : // `max_retention_period` overrides the cutoff that is used to calculate the size
4080 : // (only if it is shorter than the real cutoff).
4081 : max_retention_period: Option<u64>,
4082 : cause: LogicalSizeCalculationCause,
4083 : cancel: &CancellationToken,
4084 : ctx: &RequestContext,
4085 : ) -> Result<size::ModelInputs, size::CalculateSyntheticSizeError> {
4086 : let logical_sizes_at_once = self
4087 : .conf
4088 : .concurrent_tenant_size_logical_size_queries
4089 : .inner();
4090 :
4091 : // TODO: Having a single mutex block concurrent reads is not great for performance.
4092 : //
4093 : // But the only case where we need to run multiple of these at once is when we
4094 : // request a size for a tenant manually via API, while another background calculation
4095 : // is in progress (which is not a common case).
4096 : //
4097 : // See more for on the issue #2748 condenced out of the initial PR review.
4098 : let mut shared_cache = tokio::select! {
4099 : locked = self.cached_logical_sizes.lock() => locked,
4100 : _ = cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
4101 : _ = self.cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
4102 : };
4103 :
4104 : size::gather_inputs(
4105 : self,
4106 : logical_sizes_at_once,
4107 : max_retention_period,
4108 : &mut shared_cache,
4109 : cause,
4110 : cancel,
4111 : ctx,
4112 : )
4113 : .await
4114 : }
4115 :
4116 : /// Calculate synthetic tenant size and cache the result.
4117 : /// This is periodically called by background worker.
4118 : /// result is cached in tenant struct
4119 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
4120 : pub async fn calculate_synthetic_size(
4121 : &self,
4122 : cause: LogicalSizeCalculationCause,
4123 : cancel: &CancellationToken,
4124 : ctx: &RequestContext,
4125 : ) -> Result<u64, size::CalculateSyntheticSizeError> {
4126 : let inputs = self.gather_size_inputs(None, cause, cancel, ctx).await?;
4127 :
4128 : let size = inputs.calculate();
4129 :
4130 : self.set_cached_synthetic_size(size);
4131 :
4132 : Ok(size)
4133 : }
4134 :
4135 : /// Cache given synthetic size and update the metric value
4136 0 : pub fn set_cached_synthetic_size(&self, size: u64) {
4137 0 : self.cached_synthetic_tenant_size
4138 0 : .store(size, Ordering::Relaxed);
4139 0 :
4140 0 : // Only shard zero should be calculating synthetic sizes
4141 0 : debug_assert!(self.shard_identity.is_shard_zero());
4142 :
4143 0 : TENANT_SYNTHETIC_SIZE_METRIC
4144 0 : .get_metric_with_label_values(&[&self.tenant_shard_id.tenant_id.to_string()])
4145 0 : .unwrap()
4146 0 : .set(size);
4147 0 : }
4148 :
4149 0 : pub fn cached_synthetic_size(&self) -> u64 {
4150 0 : self.cached_synthetic_tenant_size.load(Ordering::Relaxed)
4151 0 : }
4152 :
4153 : /// Flush any in-progress layers, schedule uploads, and wait for uploads to complete.
4154 : ///
4155 : /// This function can take a long time: callers should wrap it in a timeout if calling
4156 : /// from an external API handler.
4157 : ///
4158 : /// Cancel-safety: cancelling this function may leave I/O running, but such I/O is
4159 : /// still bounded by tenant/timeline shutdown.
4160 0 : #[tracing::instrument(skip_all)]
4161 : pub(crate) async fn flush_remote(&self) -> anyhow::Result<()> {
4162 : let timelines = self.timelines.lock().unwrap().clone();
4163 :
4164 0 : async fn flush_timeline(_gate: GateGuard, timeline: Arc<Timeline>) -> anyhow::Result<()> {
4165 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Flushing...");
4166 0 : timeline.freeze_and_flush().await?;
4167 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Waiting for uploads...");
4168 0 : timeline.remote_client.wait_completion().await?;
4169 :
4170 0 : Ok(())
4171 0 : }
4172 :
4173 : // We do not use a JoinSet for these tasks, because we don't want them to be
4174 : // aborted when this function's future is cancelled: they should stay alive
4175 : // holding their GateGuard until they complete, to ensure their I/Os complete
4176 : // before Timeline shutdown completes.
4177 : let mut results = FuturesUnordered::new();
4178 :
4179 : for (_timeline_id, timeline) in timelines {
4180 : // Run each timeline's flush in a task holding the timeline's gate: this
4181 : // means that if this function's future is cancelled, the Timeline shutdown
4182 : // will still wait for any I/O in here to complete.
4183 : let Ok(gate) = timeline.gate.enter() else {
4184 : continue;
4185 : };
4186 0 : let jh = tokio::task::spawn(async move { flush_timeline(gate, timeline).await });
4187 : results.push(jh);
4188 : }
4189 :
4190 : while let Some(r) = results.next().await {
4191 : if let Err(e) = r {
4192 : if !e.is_cancelled() && !e.is_panic() {
4193 : tracing::error!("unexpected join error: {e:?}");
4194 : }
4195 : }
4196 : }
4197 :
4198 : // The flushes we did above were just writes, but the Tenant might have had
4199 : // pending deletions as well from recent compaction/gc: we want to flush those
4200 : // as well. This requires flushing the global delete queue. This is cheap
4201 : // because it's typically a no-op.
4202 : match self.deletion_queue_client.flush_execute().await {
4203 : Ok(_) => {}
4204 : Err(DeletionQueueError::ShuttingDown) => {}
4205 : }
4206 :
4207 : Ok(())
4208 : }
4209 :
4210 0 : pub(crate) fn get_tenant_conf(&self) -> TenantConfOpt {
4211 0 : self.tenant_conf.load().tenant_conf.clone()
4212 0 : }
4213 :
4214 : /// How much local storage would this tenant like to have? It can cope with
4215 : /// less than this (via eviction and on-demand downloads), but this function enables
4216 : /// the Tenant to advertise how much storage it would prefer to have to provide fast I/O
4217 : /// by keeping important things on local disk.
4218 : ///
4219 : /// This is a heuristic, not a guarantee: tenants that are long-idle will actually use less
4220 : /// than they report here, due to layer eviction. Tenants with many active branches may
4221 : /// actually use more than they report here.
4222 0 : pub(crate) fn local_storage_wanted(&self) -> u64 {
4223 0 : let timelines = self.timelines.lock().unwrap();
4224 0 :
4225 0 : // Heuristic: we use the max() of the timelines' visible sizes, rather than the sum. This
4226 0 : // reflects the observation that on tenants with multiple large branches, typically only one
4227 0 : // of them is used actively enough to occupy space on disk.
4228 0 : timelines
4229 0 : .values()
4230 0 : .map(|t| t.metrics.visible_physical_size_gauge.get())
4231 0 : .max()
4232 0 : .unwrap_or(0)
4233 0 : }
4234 : }
4235 :
4236 : /// Create the cluster temporarily in 'initdbpath' directory inside the repository
4237 : /// to get bootstrap data for timeline initialization.
4238 0 : async fn run_initdb(
4239 0 : conf: &'static PageServerConf,
4240 0 : initdb_target_dir: &Utf8Path,
4241 0 : pg_version: u32,
4242 0 : cancel: &CancellationToken,
4243 0 : ) -> Result<(), InitdbError> {
4244 0 : let initdb_bin_path = conf
4245 0 : .pg_bin_dir(pg_version)
4246 0 : .map_err(InitdbError::Other)?
4247 0 : .join("initdb");
4248 0 : let initdb_lib_dir = conf.pg_lib_dir(pg_version).map_err(InitdbError::Other)?;
4249 0 : info!(
4250 0 : "running {} in {}, libdir: {}",
4251 : initdb_bin_path, initdb_target_dir, initdb_lib_dir,
4252 : );
4253 :
4254 0 : let _permit = INIT_DB_SEMAPHORE.acquire().await;
4255 :
4256 0 : let initdb_command = tokio::process::Command::new(&initdb_bin_path)
4257 0 : .args(["--pgdata", initdb_target_dir.as_ref()])
4258 0 : .args(["--username", &conf.superuser])
4259 0 : .args(["--encoding", "utf8"])
4260 0 : .arg("--no-instructions")
4261 0 : .arg("--no-sync")
4262 0 : .env_clear()
4263 0 : .env("LD_LIBRARY_PATH", &initdb_lib_dir)
4264 0 : .env("DYLD_LIBRARY_PATH", &initdb_lib_dir)
4265 0 : .stdin(std::process::Stdio::null())
4266 0 : // stdout invocation produces the same output every time, we don't need it
4267 0 : .stdout(std::process::Stdio::null())
4268 0 : // we would be interested in the stderr output, if there was any
4269 0 : .stderr(std::process::Stdio::piped())
4270 0 : .spawn()?;
4271 :
4272 : // Ideally we'd select here with the cancellation token, but the problem is that
4273 : // we can't safely terminate initdb: it launches processes of its own, and killing
4274 : // initdb doesn't kill them. After we return from this function, we want the target
4275 : // directory to be able to be cleaned up.
4276 : // See https://github.com/neondatabase/neon/issues/6385
4277 0 : let initdb_output = initdb_command.wait_with_output().await?;
4278 0 : if !initdb_output.status.success() {
4279 0 : return Err(InitdbError::Failed(
4280 0 : initdb_output.status,
4281 0 : initdb_output.stderr,
4282 0 : ));
4283 0 : }
4284 0 :
4285 0 : // This isn't true cancellation support, see above. Still return an error to
4286 0 : // excercise the cancellation code path.
4287 0 : if cancel.is_cancelled() {
4288 0 : return Err(InitdbError::Cancelled);
4289 0 : }
4290 0 :
4291 0 : Ok(())
4292 0 : }
4293 :
4294 : /// Dump contents of a layer file to stdout.
4295 0 : pub async fn dump_layerfile_from_path(
4296 0 : path: &Utf8Path,
4297 0 : verbose: bool,
4298 0 : ctx: &RequestContext,
4299 0 : ) -> anyhow::Result<()> {
4300 : use std::os::unix::fs::FileExt;
4301 :
4302 : // All layer files start with a two-byte "magic" value, to identify the kind of
4303 : // file.
4304 0 : let file = File::open(path)?;
4305 0 : let mut header_buf = [0u8; 2];
4306 0 : file.read_exact_at(&mut header_buf, 0)?;
4307 :
4308 0 : match u16::from_be_bytes(header_buf) {
4309 : crate::IMAGE_FILE_MAGIC => {
4310 0 : ImageLayer::new_for_path(path, file)?
4311 0 : .dump(verbose, ctx)
4312 0 : .await?
4313 : }
4314 : crate::DELTA_FILE_MAGIC => {
4315 0 : DeltaLayer::new_for_path(path, file)?
4316 0 : .dump(verbose, ctx)
4317 0 : .await?
4318 : }
4319 0 : magic => bail!("unrecognized magic identifier: {:?}", magic),
4320 : }
4321 :
4322 0 : Ok(())
4323 0 : }
4324 :
4325 : #[cfg(test)]
4326 : pub(crate) mod harness {
4327 : use bytes::{Bytes, BytesMut};
4328 : use once_cell::sync::OnceCell;
4329 : use pageserver_api::models::ShardParameters;
4330 : use pageserver_api::shard::ShardIndex;
4331 : use utils::logging;
4332 :
4333 : use crate::deletion_queue::mock::MockDeletionQueue;
4334 : use crate::l0_flush::L0FlushConfig;
4335 : use crate::walredo::apply_neon;
4336 : use crate::{repository::Key, walrecord::NeonWalRecord};
4337 :
4338 : use super::*;
4339 : use hex_literal::hex;
4340 : use utils::id::TenantId;
4341 :
4342 : pub const TIMELINE_ID: TimelineId =
4343 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
4344 : pub const NEW_TIMELINE_ID: TimelineId =
4345 : TimelineId::from_array(hex!("AA223344556677881122334455667788"));
4346 :
4347 : /// Convenience function to create a page image with given string as the only content
4348 5028788 : pub fn test_img(s: &str) -> Bytes {
4349 5028788 : let mut buf = BytesMut::new();
4350 5028788 : buf.extend_from_slice(s.as_bytes());
4351 5028788 : buf.resize(64, 0);
4352 5028788 :
4353 5028788 : buf.freeze()
4354 5028788 : }
4355 :
4356 : impl From<TenantConf> for TenantConfOpt {
4357 186 : fn from(tenant_conf: TenantConf) -> Self {
4358 186 : Self {
4359 186 : checkpoint_distance: Some(tenant_conf.checkpoint_distance),
4360 186 : checkpoint_timeout: Some(tenant_conf.checkpoint_timeout),
4361 186 : compaction_target_size: Some(tenant_conf.compaction_target_size),
4362 186 : compaction_period: Some(tenant_conf.compaction_period),
4363 186 : compaction_threshold: Some(tenant_conf.compaction_threshold),
4364 186 : compaction_algorithm: Some(tenant_conf.compaction_algorithm),
4365 186 : gc_horizon: Some(tenant_conf.gc_horizon),
4366 186 : gc_period: Some(tenant_conf.gc_period),
4367 186 : image_creation_threshold: Some(tenant_conf.image_creation_threshold),
4368 186 : pitr_interval: Some(tenant_conf.pitr_interval),
4369 186 : walreceiver_connect_timeout: Some(tenant_conf.walreceiver_connect_timeout),
4370 186 : lagging_wal_timeout: Some(tenant_conf.lagging_wal_timeout),
4371 186 : max_lsn_wal_lag: Some(tenant_conf.max_lsn_wal_lag),
4372 186 : eviction_policy: Some(tenant_conf.eviction_policy),
4373 186 : min_resident_size_override: tenant_conf.min_resident_size_override,
4374 186 : evictions_low_residence_duration_metric_threshold: Some(
4375 186 : tenant_conf.evictions_low_residence_duration_metric_threshold,
4376 186 : ),
4377 186 : heatmap_period: Some(tenant_conf.heatmap_period),
4378 186 : lazy_slru_download: Some(tenant_conf.lazy_slru_download),
4379 186 : timeline_get_throttle: Some(tenant_conf.timeline_get_throttle),
4380 186 : image_layer_creation_check_threshold: Some(
4381 186 : tenant_conf.image_layer_creation_check_threshold,
4382 186 : ),
4383 186 : switch_aux_file_policy: Some(tenant_conf.switch_aux_file_policy),
4384 186 : lsn_lease_length: Some(tenant_conf.lsn_lease_length),
4385 186 : lsn_lease_length_for_ts: Some(tenant_conf.lsn_lease_length_for_ts),
4386 186 : }
4387 186 : }
4388 : }
4389 :
4390 : pub struct TenantHarness {
4391 : pub conf: &'static PageServerConf,
4392 : pub tenant_conf: TenantConf,
4393 : pub tenant_shard_id: TenantShardId,
4394 : pub generation: Generation,
4395 : pub shard: ShardIndex,
4396 : pub remote_storage: GenericRemoteStorage,
4397 : pub remote_fs_dir: Utf8PathBuf,
4398 : pub deletion_queue: MockDeletionQueue,
4399 : }
4400 :
4401 : static LOG_HANDLE: OnceCell<()> = OnceCell::new();
4402 :
4403 202 : pub(crate) fn setup_logging() {
4404 202 : LOG_HANDLE.get_or_init(|| {
4405 190 : logging::init(
4406 190 : logging::LogFormat::Test,
4407 190 : // enable it in case the tests exercise code paths that use
4408 190 : // debug_assert_current_span_has_tenant_and_timeline_id
4409 190 : logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
4410 190 : logging::Output::Stdout,
4411 190 : )
4412 190 : .expect("Failed to init test logging")
4413 202 : });
4414 202 : }
4415 :
4416 : impl TenantHarness {
4417 186 : pub async fn create_custom(
4418 186 : test_name: &'static str,
4419 186 : tenant_conf: TenantConf,
4420 186 : tenant_id: TenantId,
4421 186 : shard_identity: ShardIdentity,
4422 186 : generation: Generation,
4423 186 : ) -> anyhow::Result<Self> {
4424 186 : setup_logging();
4425 186 :
4426 186 : let repo_dir = PageServerConf::test_repo_dir(test_name);
4427 186 : let _ = fs::remove_dir_all(&repo_dir);
4428 186 : fs::create_dir_all(&repo_dir)?;
4429 :
4430 186 : let conf = PageServerConf::dummy_conf(repo_dir);
4431 186 : // Make a static copy of the config. This can never be free'd, but that's
4432 186 : // OK in a test.
4433 186 : let conf: &'static PageServerConf = Box::leak(Box::new(conf));
4434 186 :
4435 186 : let shard = shard_identity.shard_index();
4436 186 : let tenant_shard_id = TenantShardId {
4437 186 : tenant_id,
4438 186 : shard_number: shard.shard_number,
4439 186 : shard_count: shard.shard_count,
4440 186 : };
4441 186 : fs::create_dir_all(conf.tenant_path(&tenant_shard_id))?;
4442 186 : fs::create_dir_all(conf.timelines_path(&tenant_shard_id))?;
4443 :
4444 : use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
4445 186 : let remote_fs_dir = conf.workdir.join("localfs");
4446 186 : std::fs::create_dir_all(&remote_fs_dir).unwrap();
4447 186 : let config = RemoteStorageConfig {
4448 186 : storage: RemoteStorageKind::LocalFs {
4449 186 : local_path: remote_fs_dir.clone(),
4450 186 : },
4451 186 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
4452 186 : };
4453 186 : let remote_storage = GenericRemoteStorage::from_config(&config).await.unwrap();
4454 186 : let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()));
4455 186 :
4456 186 : Ok(Self {
4457 186 : conf,
4458 186 : tenant_conf,
4459 186 : tenant_shard_id,
4460 186 : generation,
4461 186 : shard,
4462 186 : remote_storage,
4463 186 : remote_fs_dir,
4464 186 : deletion_queue,
4465 186 : })
4466 186 : }
4467 :
4468 174 : pub async fn create(test_name: &'static str) -> anyhow::Result<Self> {
4469 174 : // Disable automatic GC and compaction to make the unit tests more deterministic.
4470 174 : // The tests perform them manually if needed.
4471 174 : let tenant_conf = TenantConf {
4472 174 : gc_period: Duration::ZERO,
4473 174 : compaction_period: Duration::ZERO,
4474 174 : ..TenantConf::default()
4475 174 : };
4476 174 : let tenant_id = TenantId::generate();
4477 174 : let shard = ShardIdentity::unsharded();
4478 174 : Self::create_custom(
4479 174 : test_name,
4480 174 : tenant_conf,
4481 174 : tenant_id,
4482 174 : shard,
4483 174 : Generation::new(0xdeadbeef),
4484 174 : )
4485 0 : .await
4486 174 : }
4487 :
4488 20 : pub fn span(&self) -> tracing::Span {
4489 20 : info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
4490 20 : }
4491 :
4492 186 : pub(crate) async fn load(&self) -> (Arc<Tenant>, RequestContext) {
4493 186 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
4494 186 : (
4495 186 : self.do_try_load(&ctx)
4496 730 : .await
4497 186 : .expect("failed to load test tenant"),
4498 186 : ctx,
4499 186 : )
4500 186 : }
4501 :
4502 186 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
4503 : pub(crate) async fn do_try_load(
4504 : &self,
4505 : ctx: &RequestContext,
4506 : ) -> anyhow::Result<Arc<Tenant>> {
4507 : let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
4508 :
4509 : let tenant = Arc::new(Tenant::new(
4510 : TenantState::Attaching,
4511 : self.conf,
4512 : AttachedTenantConf::try_from(LocationConf::attached_single(
4513 : TenantConfOpt::from(self.tenant_conf.clone()),
4514 : self.generation,
4515 : &ShardParameters::default(),
4516 : ))
4517 : .unwrap(),
4518 : // This is a legacy/test code path: sharding isn't supported here.
4519 : ShardIdentity::unsharded(),
4520 : Some(walredo_mgr),
4521 : self.tenant_shard_id,
4522 : self.remote_storage.clone(),
4523 : self.deletion_queue.new_client(),
4524 : // TODO: ideally we should run all unit tests with both configs
4525 : L0FlushGlobalState::new(L0FlushConfig::default()),
4526 : ));
4527 :
4528 : let preload = tenant
4529 : .preload(&self.remote_storage, CancellationToken::new())
4530 : .await?;
4531 : tenant.attach(Some(preload), ctx).await?;
4532 :
4533 : tenant.state.send_replace(TenantState::Active);
4534 : for timeline in tenant.timelines.lock().unwrap().values() {
4535 : timeline.set_state(TimelineState::Active);
4536 : }
4537 : Ok(tenant)
4538 : }
4539 :
4540 2 : pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
4541 2 : self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
4542 2 : }
4543 : }
4544 :
4545 : // Mock WAL redo manager that doesn't do much
4546 : pub(crate) struct TestRedoManager;
4547 :
4548 : impl TestRedoManager {
4549 : /// # Cancel-Safety
4550 : ///
4551 : /// This method is cancellation-safe.
4552 342 : pub async fn request_redo(
4553 342 : &self,
4554 342 : key: Key,
4555 342 : lsn: Lsn,
4556 342 : base_img: Option<(Lsn, Bytes)>,
4557 342 : records: Vec<(Lsn, NeonWalRecord)>,
4558 342 : _pg_version: u32,
4559 342 : ) -> Result<Bytes, walredo::Error> {
4560 458 : let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
4561 342 : if records_neon {
4562 : // For Neon wal records, we can decode without spawning postgres, so do so.
4563 342 : let mut page = match (base_img, records.first()) {
4564 338 : (Some((_lsn, img)), _) => {
4565 338 : let mut page = BytesMut::new();
4566 338 : page.extend_from_slice(&img);
4567 338 : page
4568 : }
4569 4 : (_, Some((_lsn, rec))) if rec.will_init() => BytesMut::new(),
4570 : _ => {
4571 0 : panic!("Neon WAL redo requires base image or will init record");
4572 : }
4573 : };
4574 :
4575 800 : for (record_lsn, record) in records {
4576 458 : apply_neon::apply_in_neon(&record, record_lsn, key, &mut page)?;
4577 : }
4578 342 : Ok(page.freeze())
4579 : } else {
4580 : // We never spawn a postgres walredo process in unit tests: just log what we might have done.
4581 0 : let s = format!(
4582 0 : "redo for {} to get to {}, with {} and {} records",
4583 0 : key,
4584 0 : lsn,
4585 0 : if base_img.is_some() {
4586 0 : "base image"
4587 : } else {
4588 0 : "no base image"
4589 : },
4590 0 : records.len()
4591 0 : );
4592 0 : println!("{s}");
4593 0 :
4594 0 : Ok(test_img(&s))
4595 : }
4596 342 : }
4597 : }
4598 : }
4599 :
4600 : #[cfg(test)]
4601 : mod tests {
4602 : use std::collections::{BTreeMap, BTreeSet};
4603 :
4604 : use super::*;
4605 : use crate::keyspace::KeySpaceAccum;
4606 : use crate::repository::{Key, Value};
4607 : use crate::tenant::harness::*;
4608 : use crate::tenant::timeline::CompactFlags;
4609 : use crate::walrecord::NeonWalRecord;
4610 : use crate::DEFAULT_PG_VERSION;
4611 : use bytes::{Bytes, BytesMut};
4612 : use hex_literal::hex;
4613 : use itertools::Itertools;
4614 : use pageserver_api::key::{AUX_KEY_PREFIX, NON_INHERITED_RANGE};
4615 : use pageserver_api::keyspace::KeySpace;
4616 : use pageserver_api::models::{CompactionAlgorithm, CompactionAlgorithmSettings};
4617 : use rand::{thread_rng, Rng};
4618 : use storage_layer::PersistentLayerKey;
4619 : use tests::storage_layer::ValuesReconstructState;
4620 : use tests::timeline::{GetVectoredError, ShutdownMode};
4621 : use timeline::compaction::{KeyHistoryRetention, KeyLogAtLsn};
4622 : use timeline::{DeltaLayerTestDesc, GcInfo};
4623 : use utils::id::TenantId;
4624 :
4625 : static TEST_KEY: Lazy<Key> =
4626 18 : Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));
4627 :
4628 : #[tokio::test]
4629 2 : async fn test_basic() -> anyhow::Result<()> {
4630 8 : let (tenant, ctx) = TenantHarness::create("test_basic").await?.load().await;
4631 2 : let tline = tenant
4632 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
4633 4 : .await?;
4634 2 :
4635 2 : let mut writer = tline.writer().await;
4636 2 : writer
4637 2 : .put(
4638 2 : *TEST_KEY,
4639 2 : Lsn(0x10),
4640 2 : &Value::Image(test_img("foo at 0x10")),
4641 2 : &ctx,
4642 2 : )
4643 2 : .await?;
4644 2 : writer.finish_write(Lsn(0x10));
4645 2 : drop(writer);
4646 2 :
4647 2 : let mut writer = tline.writer().await;
4648 2 : writer
4649 2 : .put(
4650 2 : *TEST_KEY,
4651 2 : Lsn(0x20),
4652 2 : &Value::Image(test_img("foo at 0x20")),
4653 2 : &ctx,
4654 2 : )
4655 2 : .await?;
4656 2 : writer.finish_write(Lsn(0x20));
4657 2 : drop(writer);
4658 2 :
4659 2 : assert_eq!(
4660 2 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
4661 2 : test_img("foo at 0x10")
4662 2 : );
4663 2 : assert_eq!(
4664 2 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
4665 2 : test_img("foo at 0x10")
4666 2 : );
4667 2 : assert_eq!(
4668 2 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
4669 2 : test_img("foo at 0x20")
4670 2 : );
4671 2 :
4672 2 : Ok(())
4673 2 : }
4674 :
4675 : #[tokio::test]
4676 2 : async fn no_duplicate_timelines() -> anyhow::Result<()> {
4677 2 : let (tenant, ctx) = TenantHarness::create("no_duplicate_timelines")
4678 2 : .await?
4679 2 : .load()
4680 8 : .await;
4681 2 : let _ = tenant
4682 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4683 4 : .await?;
4684 2 :
4685 2 : match tenant
4686 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4687 2 : .await
4688 2 : {
4689 2 : Ok(_) => panic!("duplicate timeline creation should fail"),
4690 2 : Err(e) => assert_eq!(e.to_string(), "Already exists".to_string()),
4691 2 : }
4692 2 :
4693 2 : Ok(())
4694 2 : }
4695 :
4696 : /// Convenience function to create a page image with given string as the only content
4697 10 : pub fn test_value(s: &str) -> Value {
4698 10 : let mut buf = BytesMut::new();
4699 10 : buf.extend_from_slice(s.as_bytes());
4700 10 : Value::Image(buf.freeze())
4701 10 : }
4702 :
4703 : ///
4704 : /// Test branch creation
4705 : ///
4706 : #[tokio::test]
4707 2 : async fn test_branch() -> anyhow::Result<()> {
4708 2 : use std::str::from_utf8;
4709 2 :
4710 7 : let (tenant, ctx) = TenantHarness::create("test_branch").await?.load().await;
4711 2 : let tline = tenant
4712 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4713 4 : .await?;
4714 2 : let mut writer = tline.writer().await;
4715 2 :
4716 2 : #[allow(non_snake_case)]
4717 2 : let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
4718 2 : #[allow(non_snake_case)]
4719 2 : let TEST_KEY_B: Key = Key::from_hex("110000000033333333444444445500000002").unwrap();
4720 2 :
4721 2 : // Insert a value on the timeline
4722 2 : writer
4723 2 : .put(TEST_KEY_A, Lsn(0x20), &test_value("foo at 0x20"), &ctx)
4724 2 : .await?;
4725 2 : writer
4726 2 : .put(TEST_KEY_B, Lsn(0x20), &test_value("foobar at 0x20"), &ctx)
4727 2 : .await?;
4728 2 : writer.finish_write(Lsn(0x20));
4729 2 :
4730 2 : writer
4731 2 : .put(TEST_KEY_A, Lsn(0x30), &test_value("foo at 0x30"), &ctx)
4732 2 : .await?;
4733 2 : writer.finish_write(Lsn(0x30));
4734 2 : writer
4735 2 : .put(TEST_KEY_A, Lsn(0x40), &test_value("foo at 0x40"), &ctx)
4736 2 : .await?;
4737 2 : writer.finish_write(Lsn(0x40));
4738 2 :
4739 2 : //assert_current_logical_size(&tline, Lsn(0x40));
4740 2 :
4741 2 : // Branch the history, modify relation differently on the new timeline
4742 2 : tenant
4743 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x30)), &ctx)
4744 2 : .await?;
4745 2 : let newtline = tenant
4746 2 : .get_timeline(NEW_TIMELINE_ID, true)
4747 2 : .expect("Should have a local timeline");
4748 2 : let mut new_writer = newtline.writer().await;
4749 2 : new_writer
4750 2 : .put(TEST_KEY_A, Lsn(0x40), &test_value("bar at 0x40"), &ctx)
4751 2 : .await?;
4752 2 : new_writer.finish_write(Lsn(0x40));
4753 2 :
4754 2 : // Check page contents on both branches
4755 2 : assert_eq!(
4756 2 : from_utf8(&tline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
4757 2 : "foo at 0x40"
4758 2 : );
4759 2 : assert_eq!(
4760 2 : from_utf8(&newtline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
4761 2 : "bar at 0x40"
4762 2 : );
4763 2 : assert_eq!(
4764 2 : from_utf8(&newtline.get(TEST_KEY_B, Lsn(0x40), &ctx).await?)?,
4765 2 : "foobar at 0x20"
4766 2 : );
4767 2 :
4768 2 : //assert_current_logical_size(&tline, Lsn(0x40));
4769 2 :
4770 2 : Ok(())
4771 2 : }
4772 :
4773 20 : async fn make_some_layers(
4774 20 : tline: &Timeline,
4775 20 : start_lsn: Lsn,
4776 20 : ctx: &RequestContext,
4777 20 : ) -> anyhow::Result<()> {
4778 20 : let mut lsn = start_lsn;
4779 : {
4780 20 : let mut writer = tline.writer().await;
4781 : // Create a relation on the timeline
4782 20 : writer
4783 20 : .put(
4784 20 : *TEST_KEY,
4785 20 : lsn,
4786 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4787 20 : ctx,
4788 20 : )
4789 10 : .await?;
4790 20 : writer.finish_write(lsn);
4791 20 : lsn += 0x10;
4792 20 : writer
4793 20 : .put(
4794 20 : *TEST_KEY,
4795 20 : lsn,
4796 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4797 20 : ctx,
4798 20 : )
4799 0 : .await?;
4800 20 : writer.finish_write(lsn);
4801 20 : lsn += 0x10;
4802 20 : }
4803 21 : tline.freeze_and_flush().await?;
4804 : {
4805 20 : let mut writer = tline.writer().await;
4806 20 : writer
4807 20 : .put(
4808 20 : *TEST_KEY,
4809 20 : lsn,
4810 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4811 20 : ctx,
4812 20 : )
4813 10 : .await?;
4814 20 : writer.finish_write(lsn);
4815 20 : lsn += 0x10;
4816 20 : writer
4817 20 : .put(
4818 20 : *TEST_KEY,
4819 20 : lsn,
4820 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4821 20 : ctx,
4822 20 : )
4823 0 : .await?;
4824 20 : writer.finish_write(lsn);
4825 20 : }
4826 21 : tline.freeze_and_flush().await.map_err(|e| e.into())
4827 20 : }
4828 :
4829 : #[tokio::test(start_paused = true)]
4830 2 : async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
4831 2 : let (tenant, ctx) =
4832 2 : TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
4833 2 : .await?
4834 2 : .load()
4835 7 : .await;
4836 2 : // Advance to the lsn lease deadline so that GC is not blocked by
4837 2 : // initial transition into AttachedSingle.
4838 2 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
4839 2 : tokio::time::resume();
4840 2 : let tline = tenant
4841 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4842 4 : .await?;
4843 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4844 2 :
4845 2 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
4846 2 : // FIXME: this doesn't actually remove any layer currently, given how the flushing
4847 2 : // and compaction works. But it does set the 'cutoff' point so that the cross check
4848 2 : // below should fail.
4849 2 : tenant
4850 2 : .gc_iteration(
4851 2 : Some(TIMELINE_ID),
4852 2 : 0x10,
4853 2 : Duration::ZERO,
4854 2 : &CancellationToken::new(),
4855 2 : &ctx,
4856 2 : )
4857 2 : .await?;
4858 2 :
4859 2 : // try to branch at lsn 25, should fail because we already garbage collected the data
4860 2 : match tenant
4861 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
4862 2 : .await
4863 2 : {
4864 2 : Ok(_) => panic!("branching should have failed"),
4865 2 : Err(err) => {
4866 2 : let CreateTimelineError::AncestorLsn(err) = err else {
4867 2 : panic!("wrong error type")
4868 2 : };
4869 2 : assert!(err.to_string().contains("invalid branch start lsn"));
4870 2 : assert!(err
4871 2 : .source()
4872 2 : .unwrap()
4873 2 : .to_string()
4874 2 : .contains("we might've already garbage collected needed data"))
4875 2 : }
4876 2 : }
4877 2 :
4878 2 : Ok(())
4879 2 : }
4880 :
4881 : #[tokio::test]
4882 2 : async fn test_prohibit_branch_creation_on_pre_initdb_lsn() -> anyhow::Result<()> {
4883 2 : let (tenant, ctx) =
4884 2 : TenantHarness::create("test_prohibit_branch_creation_on_pre_initdb_lsn")
4885 2 : .await?
4886 2 : .load()
4887 6 : .await;
4888 2 :
4889 2 : let tline = tenant
4890 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x50), DEFAULT_PG_VERSION, &ctx)
4891 4 : .await?;
4892 2 : // try to branch at lsn 0x25, should fail because initdb lsn is 0x50
4893 2 : match tenant
4894 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
4895 2 : .await
4896 2 : {
4897 2 : Ok(_) => panic!("branching should have failed"),
4898 2 : Err(err) => {
4899 2 : let CreateTimelineError::AncestorLsn(err) = err else {
4900 2 : panic!("wrong error type");
4901 2 : };
4902 2 : assert!(&err.to_string().contains("invalid branch start lsn"));
4903 2 : assert!(&err
4904 2 : .source()
4905 2 : .unwrap()
4906 2 : .to_string()
4907 2 : .contains("is earlier than latest GC cutoff"));
4908 2 : }
4909 2 : }
4910 2 :
4911 2 : Ok(())
4912 2 : }
4913 :
4914 : /*
4915 : // FIXME: This currently fails to error out. Calling GC doesn't currently
4916 : // remove the old value, we'd need to work a little harder
4917 : #[tokio::test]
4918 : async fn test_prohibit_get_for_garbage_collected_data() -> anyhow::Result<()> {
4919 : let repo =
4920 : RepoHarness::create("test_prohibit_get_for_garbage_collected_data")?
4921 : .load();
4922 :
4923 : let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION)?;
4924 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4925 :
4926 : repo.gc_iteration(Some(TIMELINE_ID), 0x10, Duration::ZERO)?;
4927 : let latest_gc_cutoff_lsn = tline.get_latest_gc_cutoff_lsn();
4928 : assert!(*latest_gc_cutoff_lsn > Lsn(0x25));
4929 : match tline.get(*TEST_KEY, Lsn(0x25)) {
4930 : Ok(_) => panic!("request for page should have failed"),
4931 : Err(err) => assert!(err.to_string().contains("not found at")),
4932 : }
4933 : Ok(())
4934 : }
4935 : */
4936 :
4937 : #[tokio::test]
4938 2 : async fn test_get_branchpoints_from_an_inactive_timeline() -> anyhow::Result<()> {
4939 2 : let (tenant, ctx) =
4940 2 : TenantHarness::create("test_get_branchpoints_from_an_inactive_timeline")
4941 2 : .await?
4942 2 : .load()
4943 8 : .await;
4944 2 : let tline = tenant
4945 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4946 4 : .await?;
4947 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4948 2 :
4949 2 : tenant
4950 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
4951 2 : .await?;
4952 2 : let newtline = tenant
4953 2 : .get_timeline(NEW_TIMELINE_ID, true)
4954 2 : .expect("Should have a local timeline");
4955 2 :
4956 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
4957 2 :
4958 2 : tline.set_broken("test".to_owned());
4959 2 :
4960 2 : tenant
4961 2 : .gc_iteration(
4962 2 : Some(TIMELINE_ID),
4963 2 : 0x10,
4964 2 : Duration::ZERO,
4965 2 : &CancellationToken::new(),
4966 2 : &ctx,
4967 2 : )
4968 2 : .await?;
4969 2 :
4970 2 : // The branchpoints should contain all timelines, even ones marked
4971 2 : // as Broken.
4972 2 : {
4973 2 : let branchpoints = &tline.gc_info.read().unwrap().retain_lsns;
4974 2 : assert_eq!(branchpoints.len(), 1);
4975 2 : assert_eq!(
4976 2 : branchpoints[0],
4977 2 : (Lsn(0x40), NEW_TIMELINE_ID, MaybeOffloaded::No)
4978 2 : );
4979 2 : }
4980 2 :
4981 2 : // You can read the key from the child branch even though the parent is
4982 2 : // Broken, as long as you don't need to access data from the parent.
4983 2 : assert_eq!(
4984 4 : newtline.get(*TEST_KEY, Lsn(0x70), &ctx).await?,
4985 2 : test_img(&format!("foo at {}", Lsn(0x70)))
4986 2 : );
4987 2 :
4988 2 : // This needs to traverse to the parent, and fails.
4989 2 : let err = newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await.unwrap_err();
4990 2 : assert!(
4991 2 : err.to_string().starts_with(&format!(
4992 2 : "bad state on timeline {}: Broken",
4993 2 : tline.timeline_id
4994 2 : )),
4995 2 : "{err}"
4996 2 : );
4997 2 :
4998 2 : Ok(())
4999 2 : }
5000 :
5001 : #[tokio::test]
5002 2 : async fn test_retain_data_in_parent_which_is_needed_for_child() -> anyhow::Result<()> {
5003 2 : let (tenant, ctx) =
5004 2 : TenantHarness::create("test_retain_data_in_parent_which_is_needed_for_child")
5005 2 : .await?
5006 2 : .load()
5007 7 : .await;
5008 2 : let tline = tenant
5009 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5010 4 : .await?;
5011 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5012 2 :
5013 2 : tenant
5014 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
5015 2 : .await?;
5016 2 : let newtline = tenant
5017 2 : .get_timeline(NEW_TIMELINE_ID, true)
5018 2 : .expect("Should have a local timeline");
5019 2 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
5020 2 : tenant
5021 2 : .gc_iteration(
5022 2 : Some(TIMELINE_ID),
5023 2 : 0x10,
5024 2 : Duration::ZERO,
5025 2 : &CancellationToken::new(),
5026 2 : &ctx,
5027 2 : )
5028 2 : .await?;
5029 4 : assert!(newtline.get(*TEST_KEY, Lsn(0x25), &ctx).await.is_ok());
5030 2 :
5031 2 : Ok(())
5032 2 : }
5033 : #[tokio::test]
5034 2 : async fn test_parent_keeps_data_forever_after_branching() -> anyhow::Result<()> {
5035 2 : let (tenant, ctx) = TenantHarness::create("test_parent_keeps_data_forever_after_branching")
5036 2 : .await?
5037 2 : .load()
5038 8 : .await;
5039 2 : let tline = tenant
5040 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5041 4 : .await?;
5042 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5043 2 :
5044 2 : tenant
5045 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
5046 2 : .await?;
5047 2 : let newtline = tenant
5048 2 : .get_timeline(NEW_TIMELINE_ID, true)
5049 2 : .expect("Should have a local timeline");
5050 2 :
5051 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
5052 2 :
5053 2 : // run gc on parent
5054 2 : tenant
5055 2 : .gc_iteration(
5056 2 : Some(TIMELINE_ID),
5057 2 : 0x10,
5058 2 : Duration::ZERO,
5059 2 : &CancellationToken::new(),
5060 2 : &ctx,
5061 2 : )
5062 2 : .await?;
5063 2 :
5064 2 : // Check that the data is still accessible on the branch.
5065 2 : assert_eq!(
5066 7 : newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
5067 2 : test_img(&format!("foo at {}", Lsn(0x40)))
5068 2 : );
5069 2 :
5070 2 : Ok(())
5071 2 : }
5072 :
5073 : #[tokio::test]
5074 2 : async fn timeline_load() -> anyhow::Result<()> {
5075 2 : const TEST_NAME: &str = "timeline_load";
5076 2 : let harness = TenantHarness::create(TEST_NAME).await?;
5077 2 : {
5078 8 : let (tenant, ctx) = harness.load().await;
5079 2 : let tline = tenant
5080 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x7000), DEFAULT_PG_VERSION, &ctx)
5081 4 : .await?;
5082 6 : make_some_layers(tline.as_ref(), Lsn(0x8000), &ctx).await?;
5083 2 : // so that all uploads finish & we can call harness.load() below again
5084 2 : tenant
5085 2 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
5086 2 : .instrument(harness.span())
5087 2 : .await
5088 2 : .ok()
5089 2 : .unwrap();
5090 2 : }
5091 2 :
5092 8 : let (tenant, _ctx) = harness.load().await;
5093 2 : tenant
5094 2 : .get_timeline(TIMELINE_ID, true)
5095 2 : .expect("cannot load timeline");
5096 2 :
5097 2 : Ok(())
5098 2 : }
5099 :
5100 : #[tokio::test]
5101 2 : async fn timeline_load_with_ancestor() -> anyhow::Result<()> {
5102 2 : const TEST_NAME: &str = "timeline_load_with_ancestor";
5103 2 : let harness = TenantHarness::create(TEST_NAME).await?;
5104 2 : // create two timelines
5105 2 : {
5106 8 : let (tenant, ctx) = harness.load().await;
5107 2 : let tline = tenant
5108 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5109 4 : .await?;
5110 2 :
5111 8 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5112 2 :
5113 2 : let child_tline = tenant
5114 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
5115 2 : .await?;
5116 2 : child_tline.set_state(TimelineState::Active);
5117 2 :
5118 2 : let newtline = tenant
5119 2 : .get_timeline(NEW_TIMELINE_ID, true)
5120 2 : .expect("Should have a local timeline");
5121 2 :
5122 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
5123 2 :
5124 2 : // so that all uploads finish & we can call harness.load() below again
5125 2 : tenant
5126 2 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
5127 2 : .instrument(harness.span())
5128 3 : .await
5129 2 : .ok()
5130 2 : .unwrap();
5131 2 : }
5132 2 :
5133 2 : // check that both of them are initially unloaded
5134 9 : let (tenant, _ctx) = harness.load().await;
5135 2 :
5136 2 : // check that both, child and ancestor are loaded
5137 2 : let _child_tline = tenant
5138 2 : .get_timeline(NEW_TIMELINE_ID, true)
5139 2 : .expect("cannot get child timeline loaded");
5140 2 :
5141 2 : let _ancestor_tline = tenant
5142 2 : .get_timeline(TIMELINE_ID, true)
5143 2 : .expect("cannot get ancestor timeline loaded");
5144 2 :
5145 2 : Ok(())
5146 2 : }
5147 :
5148 : #[tokio::test]
5149 2 : async fn delta_layer_dumping() -> anyhow::Result<()> {
5150 2 : use storage_layer::AsLayerDesc;
5151 2 : let (tenant, ctx) = TenantHarness::create("test_layer_dumping")
5152 2 : .await?
5153 2 : .load()
5154 8 : .await;
5155 2 : let tline = tenant
5156 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5157 4 : .await?;
5158 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5159 2 :
5160 2 : let layer_map = tline.layers.read().await;
5161 2 : let level0_deltas = layer_map
5162 2 : .layer_map()?
5163 2 : .level0_deltas()
5164 2 : .iter()
5165 4 : .map(|desc| layer_map.get_from_desc(desc))
5166 2 : .collect::<Vec<_>>();
5167 2 :
5168 2 : assert!(!level0_deltas.is_empty());
5169 2 :
5170 6 : for delta in level0_deltas {
5171 2 : // Ensure we are dumping a delta layer here
5172 4 : assert!(delta.layer_desc().is_delta);
5173 8 : delta.dump(true, &ctx).await.unwrap();
5174 2 : }
5175 2 :
5176 2 : Ok(())
5177 2 : }
5178 :
5179 : #[tokio::test]
5180 2 : async fn test_images() -> anyhow::Result<()> {
5181 8 : let (tenant, ctx) = TenantHarness::create("test_images").await?.load().await;
5182 2 : let tline = tenant
5183 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5184 4 : .await?;
5185 2 :
5186 2 : let mut writer = tline.writer().await;
5187 2 : writer
5188 2 : .put(
5189 2 : *TEST_KEY,
5190 2 : Lsn(0x10),
5191 2 : &Value::Image(test_img("foo at 0x10")),
5192 2 : &ctx,
5193 2 : )
5194 2 : .await?;
5195 2 : writer.finish_write(Lsn(0x10));
5196 2 : drop(writer);
5197 2 :
5198 2 : tline.freeze_and_flush().await?;
5199 2 : tline
5200 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
5201 2 : .await?;
5202 2 :
5203 2 : let mut writer = tline.writer().await;
5204 2 : writer
5205 2 : .put(
5206 2 : *TEST_KEY,
5207 2 : Lsn(0x20),
5208 2 : &Value::Image(test_img("foo at 0x20")),
5209 2 : &ctx,
5210 2 : )
5211 2 : .await?;
5212 2 : writer.finish_write(Lsn(0x20));
5213 2 : drop(writer);
5214 2 :
5215 2 : tline.freeze_and_flush().await?;
5216 2 : tline
5217 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
5218 2 : .await?;
5219 2 :
5220 2 : let mut writer = tline.writer().await;
5221 2 : writer
5222 2 : .put(
5223 2 : *TEST_KEY,
5224 2 : Lsn(0x30),
5225 2 : &Value::Image(test_img("foo at 0x30")),
5226 2 : &ctx,
5227 2 : )
5228 2 : .await?;
5229 2 : writer.finish_write(Lsn(0x30));
5230 2 : drop(writer);
5231 2 :
5232 2 : tline.freeze_and_flush().await?;
5233 2 : tline
5234 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
5235 2 : .await?;
5236 2 :
5237 2 : let mut writer = tline.writer().await;
5238 2 : writer
5239 2 : .put(
5240 2 : *TEST_KEY,
5241 2 : Lsn(0x40),
5242 2 : &Value::Image(test_img("foo at 0x40")),
5243 2 : &ctx,
5244 2 : )
5245 2 : .await?;
5246 2 : writer.finish_write(Lsn(0x40));
5247 2 : drop(writer);
5248 2 :
5249 2 : tline.freeze_and_flush().await?;
5250 2 : tline
5251 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
5252 2 : .await?;
5253 2 :
5254 2 : assert_eq!(
5255 4 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
5256 2 : test_img("foo at 0x10")
5257 2 : );
5258 2 : assert_eq!(
5259 4 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
5260 2 : test_img("foo at 0x10")
5261 2 : );
5262 2 : assert_eq!(
5263 2 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
5264 2 : test_img("foo at 0x20")
5265 2 : );
5266 2 : assert_eq!(
5267 4 : tline.get(*TEST_KEY, Lsn(0x30), &ctx).await?,
5268 2 : test_img("foo at 0x30")
5269 2 : );
5270 2 : assert_eq!(
5271 4 : tline.get(*TEST_KEY, Lsn(0x40), &ctx).await?,
5272 2 : test_img("foo at 0x40")
5273 2 : );
5274 2 :
5275 2 : Ok(())
5276 2 : }
5277 :
5278 4 : async fn bulk_insert_compact_gc(
5279 4 : tenant: &Tenant,
5280 4 : timeline: &Arc<Timeline>,
5281 4 : ctx: &RequestContext,
5282 4 : lsn: Lsn,
5283 4 : repeat: usize,
5284 4 : key_count: usize,
5285 4 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
5286 4 : let compact = true;
5287 40718 : bulk_insert_maybe_compact_gc(tenant, timeline, ctx, lsn, repeat, key_count, compact).await
5288 4 : }
5289 :
5290 8 : async fn bulk_insert_maybe_compact_gc(
5291 8 : tenant: &Tenant,
5292 8 : timeline: &Arc<Timeline>,
5293 8 : ctx: &RequestContext,
5294 8 : mut lsn: Lsn,
5295 8 : repeat: usize,
5296 8 : key_count: usize,
5297 8 : compact: bool,
5298 8 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
5299 8 : let mut inserted: HashMap<Key, BTreeSet<Lsn>> = Default::default();
5300 8 :
5301 8 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5302 8 : let mut blknum = 0;
5303 8 :
5304 8 : // Enforce that key range is monotonously increasing
5305 8 : let mut keyspace = KeySpaceAccum::new();
5306 8 :
5307 8 : let cancel = CancellationToken::new();
5308 8 :
5309 8 : for _ in 0..repeat {
5310 400 : for _ in 0..key_count {
5311 4000000 : test_key.field6 = blknum;
5312 4000000 : let mut writer = timeline.writer().await;
5313 4000000 : writer
5314 4000000 : .put(
5315 4000000 : test_key,
5316 4000000 : lsn,
5317 4000000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5318 4000000 : ctx,
5319 4000000 : )
5320 3426 : .await?;
5321 4000000 : inserted.entry(test_key).or_default().insert(lsn);
5322 4000000 : writer.finish_write(lsn);
5323 4000000 : drop(writer);
5324 4000000 :
5325 4000000 : keyspace.add_key(test_key);
5326 4000000 :
5327 4000000 : lsn = Lsn(lsn.0 + 0x10);
5328 4000000 : blknum += 1;
5329 : }
5330 :
5331 400 : timeline.freeze_and_flush().await?;
5332 400 : if compact {
5333 : // this requires timeline to be &Arc<Timeline>
5334 8618 : timeline.compact(&cancel, EnumSet::empty(), ctx).await?;
5335 200 : }
5336 :
5337 : // this doesn't really need to use the timeline_id target, but it is closer to what it
5338 : // originally was.
5339 400 : let res = tenant
5340 400 : .gc_iteration(Some(timeline.timeline_id), 0, Duration::ZERO, &cancel, ctx)
5341 0 : .await?;
5342 :
5343 400 : assert_eq!(res.layers_removed, 0, "this never removes anything");
5344 : }
5345 :
5346 8 : Ok(inserted)
5347 8 : }
5348 :
5349 : //
5350 : // Insert 1000 key-value pairs with increasing keys, flush, compact, GC.
5351 : // Repeat 50 times.
5352 : //
5353 : #[tokio::test]
5354 2 : async fn test_bulk_insert() -> anyhow::Result<()> {
5355 2 : let harness = TenantHarness::create("test_bulk_insert").await?;
5356 8 : let (tenant, ctx) = harness.load().await;
5357 2 : let tline = tenant
5358 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5359 4 : .await?;
5360 2 :
5361 2 : let lsn = Lsn(0x10);
5362 20359 : bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
5363 2 :
5364 2 : Ok(())
5365 2 : }
5366 :
5367 : // Test the vectored get real implementation against a simple sequential implementation.
5368 : //
5369 : // The test generates a keyspace by repeatedly flushing the in-memory layer and compacting.
5370 : // Projected to 2D the key space looks like below. Lsn grows upwards on the Y axis and keys
5371 : // grow to the right on the X axis.
5372 : // [Delta]
5373 : // [Delta]
5374 : // [Delta]
5375 : // [Delta]
5376 : // ------------ Image ---------------
5377 : //
5378 : // After layer generation we pick the ranges to query as follows:
5379 : // 1. The beginning of each delta layer
5380 : // 2. At the seam between two adjacent delta layers
5381 : //
5382 : // There's one major downside to this test: delta layers only contains images,
5383 : // so the search can stop at the first delta layer and doesn't traverse any deeper.
5384 : #[tokio::test]
5385 2 : async fn test_get_vectored() -> anyhow::Result<()> {
5386 2 : let harness = TenantHarness::create("test_get_vectored").await?;
5387 8 : let (tenant, ctx) = harness.load().await;
5388 2 : let tline = tenant
5389 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5390 4 : .await?;
5391 2 :
5392 2 : let lsn = Lsn(0x10);
5393 20359 : let inserted = bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
5394 2 :
5395 2 : let guard = tline.layers.read().await;
5396 2 : let lm = guard.layer_map()?;
5397 2 :
5398 2 : lm.dump(true, &ctx).await?;
5399 2 :
5400 2 : let mut reads = Vec::new();
5401 2 : let mut prev = None;
5402 12 : lm.iter_historic_layers().for_each(|desc| {
5403 12 : if !desc.is_delta() {
5404 2 : prev = Some(desc.clone());
5405 2 : return;
5406 10 : }
5407 10 :
5408 10 : let start = desc.key_range.start;
5409 10 : let end = desc
5410 10 : .key_range
5411 10 : .start
5412 10 : .add(Timeline::MAX_GET_VECTORED_KEYS.try_into().unwrap());
5413 10 : reads.push(KeySpace {
5414 10 : ranges: vec![start..end],
5415 10 : });
5416 2 :
5417 10 : if let Some(prev) = &prev {
5418 10 : if !prev.is_delta() {
5419 10 : return;
5420 2 : }
5421 0 :
5422 0 : let first_range = Key {
5423 0 : field6: prev.key_range.end.field6 - 4,
5424 0 : ..prev.key_range.end
5425 0 : }..prev.key_range.end;
5426 0 :
5427 0 : let second_range = desc.key_range.start..Key {
5428 0 : field6: desc.key_range.start.field6 + 4,
5429 0 : ..desc.key_range.start
5430 0 : };
5431 0 :
5432 0 : reads.push(KeySpace {
5433 0 : ranges: vec![first_range, second_range],
5434 0 : });
5435 2 : };
5436 2 :
5437 2 : prev = Some(desc.clone());
5438 12 : });
5439 2 :
5440 2 : drop(guard);
5441 2 :
5442 2 : // Pick a big LSN such that we query over all the changes.
5443 2 : let reads_lsn = Lsn(u64::MAX - 1);
5444 2 :
5445 12 : for read in reads {
5446 10 : info!("Doing vectored read on {:?}", read);
5447 2 :
5448 10 : let vectored_res = tline
5449 10 : .get_vectored_impl(
5450 10 : read.clone(),
5451 10 : reads_lsn,
5452 10 : &mut ValuesReconstructState::new(),
5453 10 : &ctx,
5454 10 : )
5455 25 : .await;
5456 2 :
5457 10 : let mut expected_lsns: HashMap<Key, Lsn> = Default::default();
5458 10 : let mut expect_missing = false;
5459 10 : let mut key = read.start().unwrap();
5460 330 : while key != read.end().unwrap() {
5461 320 : if let Some(lsns) = inserted.get(&key) {
5462 320 : let expected_lsn = lsns.iter().rfind(|lsn| **lsn <= reads_lsn);
5463 320 : match expected_lsn {
5464 320 : Some(lsn) => {
5465 320 : expected_lsns.insert(key, *lsn);
5466 320 : }
5467 2 : None => {
5468 2 : expect_missing = true;
5469 0 : break;
5470 2 : }
5471 2 : }
5472 2 : } else {
5473 2 : expect_missing = true;
5474 0 : break;
5475 2 : }
5476 2 :
5477 320 : key = key.next();
5478 2 : }
5479 2 :
5480 10 : if expect_missing {
5481 2 : assert!(matches!(vectored_res, Err(GetVectoredError::MissingKey(_))));
5482 2 : } else {
5483 320 : for (key, image) in vectored_res? {
5484 320 : let expected_lsn = expected_lsns.get(&key).expect("determined above");
5485 320 : let expected_image = test_img(&format!("{} at {}", key.field6, expected_lsn));
5486 320 : assert_eq!(image?, expected_image);
5487 2 : }
5488 2 : }
5489 2 : }
5490 2 :
5491 2 : Ok(())
5492 2 : }
5493 :
5494 : #[tokio::test]
5495 2 : async fn test_get_vectored_aux_files() -> anyhow::Result<()> {
5496 2 : let harness = TenantHarness::create("test_get_vectored_aux_files").await?;
5497 2 :
5498 8 : let (tenant, ctx) = harness.load().await;
5499 2 : let tline = tenant
5500 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
5501 2 : .await?;
5502 2 : let tline = tline.raw_timeline().unwrap();
5503 2 :
5504 2 : let mut modification = tline.begin_modification(Lsn(0x1000));
5505 2 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
5506 2 : modification.set_lsn(Lsn(0x1008))?;
5507 2 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
5508 2 : modification.commit(&ctx).await?;
5509 2 :
5510 2 : let child_timeline_id = TimelineId::generate();
5511 2 : tenant
5512 2 : .branch_timeline_test(
5513 2 : tline,
5514 2 : child_timeline_id,
5515 2 : Some(tline.get_last_record_lsn()),
5516 2 : &ctx,
5517 2 : )
5518 2 : .await?;
5519 2 :
5520 2 : let child_timeline = tenant
5521 2 : .get_timeline(child_timeline_id, true)
5522 2 : .expect("Should have the branched timeline");
5523 2 :
5524 2 : let aux_keyspace = KeySpace {
5525 2 : ranges: vec![NON_INHERITED_RANGE],
5526 2 : };
5527 2 : let read_lsn = child_timeline.get_last_record_lsn();
5528 2 :
5529 2 : let vectored_res = child_timeline
5530 2 : .get_vectored_impl(
5531 2 : aux_keyspace.clone(),
5532 2 : read_lsn,
5533 2 : &mut ValuesReconstructState::new(),
5534 2 : &ctx,
5535 2 : )
5536 2 : .await;
5537 2 :
5538 2 : let images = vectored_res?;
5539 2 : assert!(images.is_empty());
5540 2 : Ok(())
5541 2 : }
5542 :
5543 : // Test that vectored get handles layer gaps correctly
5544 : // by advancing into the next ancestor timeline if required.
5545 : //
5546 : // The test generates timelines that look like the diagram below.
5547 : // We leave a gap in one of the L1 layers at `gap_at_key` (`/` in the diagram).
5548 : // The reconstruct data for that key lies in the ancestor timeline (`X` in the diagram).
5549 : //
5550 : // ```
5551 : //-------------------------------+
5552 : // ... |
5553 : // [ L1 ] |
5554 : // [ / L1 ] | Child Timeline
5555 : // ... |
5556 : // ------------------------------+
5557 : // [ X L1 ] | Parent Timeline
5558 : // ------------------------------+
5559 : // ```
5560 : #[tokio::test]
5561 2 : async fn test_get_vectored_key_gap() -> anyhow::Result<()> {
5562 2 : let tenant_conf = TenantConf {
5563 2 : // Make compaction deterministic
5564 2 : gc_period: Duration::ZERO,
5565 2 : compaction_period: Duration::ZERO,
5566 2 : // Encourage creation of L1 layers
5567 2 : checkpoint_distance: 16 * 1024,
5568 2 : compaction_target_size: 8 * 1024,
5569 2 : ..TenantConf::default()
5570 2 : };
5571 2 :
5572 2 : let harness = TenantHarness::create_custom(
5573 2 : "test_get_vectored_key_gap",
5574 2 : tenant_conf,
5575 2 : TenantId::generate(),
5576 2 : ShardIdentity::unsharded(),
5577 2 : Generation::new(0xdeadbeef),
5578 2 : )
5579 2 : .await?;
5580 8 : let (tenant, ctx) = harness.load().await;
5581 2 :
5582 2 : let mut current_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5583 2 : let gap_at_key = current_key.add(100);
5584 2 : let mut current_lsn = Lsn(0x10);
5585 2 :
5586 2 : const KEY_COUNT: usize = 10_000;
5587 2 :
5588 2 : let timeline_id = TimelineId::generate();
5589 2 : let current_timeline = tenant
5590 2 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
5591 4 : .await?;
5592 2 :
5593 2 : current_lsn += 0x100;
5594 2 :
5595 2 : let mut writer = current_timeline.writer().await;
5596 2 : writer
5597 2 : .put(
5598 2 : gap_at_key,
5599 2 : current_lsn,
5600 2 : &Value::Image(test_img(&format!("{} at {}", gap_at_key, current_lsn))),
5601 2 : &ctx,
5602 2 : )
5603 2 : .await?;
5604 2 : writer.finish_write(current_lsn);
5605 2 : drop(writer);
5606 2 :
5607 2 : let mut latest_lsns = HashMap::new();
5608 2 : latest_lsns.insert(gap_at_key, current_lsn);
5609 2 :
5610 2 : current_timeline.freeze_and_flush().await?;
5611 2 :
5612 2 : let child_timeline_id = TimelineId::generate();
5613 2 :
5614 2 : tenant
5615 2 : .branch_timeline_test(
5616 2 : ¤t_timeline,
5617 2 : child_timeline_id,
5618 2 : Some(current_lsn),
5619 2 : &ctx,
5620 2 : )
5621 2 : .await?;
5622 2 : let child_timeline = tenant
5623 2 : .get_timeline(child_timeline_id, true)
5624 2 : .expect("Should have the branched timeline");
5625 2 :
5626 20002 : for i in 0..KEY_COUNT {
5627 20000 : if current_key == gap_at_key {
5628 2 : current_key = current_key.next();
5629 2 : continue;
5630 19998 : }
5631 19998 :
5632 19998 : current_lsn += 0x10;
5633 2 :
5634 19998 : let mut writer = child_timeline.writer().await;
5635 19998 : writer
5636 19998 : .put(
5637 19998 : current_key,
5638 19998 : current_lsn,
5639 19998 : &Value::Image(test_img(&format!("{} at {}", current_key, current_lsn))),
5640 19998 : &ctx,
5641 19998 : )
5642 67 : .await?;
5643 19998 : writer.finish_write(current_lsn);
5644 19998 : drop(writer);
5645 19998 :
5646 19998 : latest_lsns.insert(current_key, current_lsn);
5647 19998 : current_key = current_key.next();
5648 19998 :
5649 19998 : // Flush every now and then to encourage layer file creation.
5650 19998 : if i % 500 == 0 {
5651 40 : child_timeline.freeze_and_flush().await?;
5652 19958 : }
5653 2 : }
5654 2 :
5655 2 : child_timeline.freeze_and_flush().await?;
5656 2 : let mut flags = EnumSet::new();
5657 2 : flags.insert(CompactFlags::ForceRepartition);
5658 2 : child_timeline
5659 2 : .compact(&CancellationToken::new(), flags, &ctx)
5660 1757 : .await?;
5661 2 :
5662 2 : let key_near_end = {
5663 2 : let mut tmp = current_key;
5664 2 : tmp.field6 -= 10;
5665 2 : tmp
5666 2 : };
5667 2 :
5668 2 : let key_near_gap = {
5669 2 : let mut tmp = gap_at_key;
5670 2 : tmp.field6 -= 10;
5671 2 : tmp
5672 2 : };
5673 2 :
5674 2 : let read = KeySpace {
5675 2 : ranges: vec![key_near_gap..gap_at_key.next(), key_near_end..current_key],
5676 2 : };
5677 2 : let results = child_timeline
5678 2 : .get_vectored_impl(
5679 2 : read.clone(),
5680 2 : current_lsn,
5681 2 : &mut ValuesReconstructState::new(),
5682 2 : &ctx,
5683 2 : )
5684 16 : .await?;
5685 2 :
5686 44 : for (key, img_res) in results {
5687 42 : let expected = test_img(&format!("{} at {}", key, latest_lsns[&key]));
5688 42 : assert_eq!(img_res?, expected);
5689 2 : }
5690 2 :
5691 2 : Ok(())
5692 2 : }
5693 :
5694 : // Test that vectored get descends into ancestor timelines correctly and
5695 : // does not return an image that's newer than requested.
5696 : //
5697 : // The diagram below ilustrates an interesting case. We have a parent timeline
5698 : // (top of the Lsn range) and a child timeline. The request key cannot be reconstructed
5699 : // from the child timeline, so the parent timeline must be visited. When advacing into
5700 : // the child timeline, the read path needs to remember what the requested Lsn was in
5701 : // order to avoid returning an image that's too new. The test below constructs such
5702 : // a timeline setup and does a few queries around the Lsn of each page image.
5703 : // ```
5704 : // LSN
5705 : // ^
5706 : // |
5707 : // |
5708 : // 500 | --------------------------------------> branch point
5709 : // 400 | X
5710 : // 300 | X
5711 : // 200 | --------------------------------------> requested lsn
5712 : // 100 | X
5713 : // |---------------------------------------> Key
5714 : // |
5715 : // ------> requested key
5716 : //
5717 : // Legend:
5718 : // * X - page images
5719 : // ```
5720 : #[tokio::test]
5721 2 : async fn test_get_vectored_ancestor_descent() -> anyhow::Result<()> {
5722 2 : let harness = TenantHarness::create("test_get_vectored_on_lsn_axis").await?;
5723 8 : let (tenant, ctx) = harness.load().await;
5724 2 :
5725 2 : let start_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5726 2 : let end_key = start_key.add(1000);
5727 2 : let child_gap_at_key = start_key.add(500);
5728 2 : let mut parent_gap_lsns: BTreeMap<Lsn, String> = BTreeMap::new();
5729 2 :
5730 2 : let mut current_lsn = Lsn(0x10);
5731 2 :
5732 2 : let timeline_id = TimelineId::generate();
5733 2 : let parent_timeline = tenant
5734 2 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
5735 4 : .await?;
5736 2 :
5737 2 : current_lsn += 0x100;
5738 2 :
5739 8 : for _ in 0..3 {
5740 6 : let mut key = start_key;
5741 6006 : while key < end_key {
5742 6000 : current_lsn += 0x10;
5743 6000 :
5744 6000 : let image_value = format!("{} at {}", child_gap_at_key, current_lsn);
5745 2 :
5746 6000 : let mut writer = parent_timeline.writer().await;
5747 6000 : writer
5748 6000 : .put(
5749 6000 : key,
5750 6000 : current_lsn,
5751 6000 : &Value::Image(test_img(&image_value)),
5752 6000 : &ctx,
5753 6000 : )
5754 6 : .await?;
5755 6000 : writer.finish_write(current_lsn);
5756 6000 :
5757 6000 : if key == child_gap_at_key {
5758 6 : parent_gap_lsns.insert(current_lsn, image_value);
5759 5994 : }
5760 2 :
5761 6000 : key = key.next();
5762 2 : }
5763 2 :
5764 6 : parent_timeline.freeze_and_flush().await?;
5765 2 : }
5766 2 :
5767 2 : let child_timeline_id = TimelineId::generate();
5768 2 :
5769 2 : let child_timeline = tenant
5770 2 : .branch_timeline_test(&parent_timeline, child_timeline_id, Some(current_lsn), &ctx)
5771 2 : .await?;
5772 2 :
5773 2 : let mut key = start_key;
5774 2002 : while key < end_key {
5775 2000 : if key == child_gap_at_key {
5776 2 : key = key.next();
5777 2 : continue;
5778 1998 : }
5779 1998 :
5780 1998 : current_lsn += 0x10;
5781 2 :
5782 1998 : let mut writer = child_timeline.writer().await;
5783 1998 : writer
5784 1998 : .put(
5785 1998 : key,
5786 1998 : current_lsn,
5787 1998 : &Value::Image(test_img(&format!("{} at {}", key, current_lsn))),
5788 1998 : &ctx,
5789 1998 : )
5790 17 : .await?;
5791 1998 : writer.finish_write(current_lsn);
5792 1998 :
5793 1998 : key = key.next();
5794 2 : }
5795 2 :
5796 2 : child_timeline.freeze_and_flush().await?;
5797 2 :
5798 2 : let lsn_offsets: [i64; 5] = [-10, -1, 0, 1, 10];
5799 2 : let mut query_lsns = Vec::new();
5800 6 : for image_lsn in parent_gap_lsns.keys().rev() {
5801 36 : for offset in lsn_offsets {
5802 30 : query_lsns.push(Lsn(image_lsn
5803 30 : .0
5804 30 : .checked_add_signed(offset)
5805 30 : .expect("Shouldn't overflow")));
5806 30 : }
5807 2 : }
5808 2 :
5809 32 : for query_lsn in query_lsns {
5810 30 : let results = child_timeline
5811 30 : .get_vectored_impl(
5812 30 : KeySpace {
5813 30 : ranges: vec![child_gap_at_key..child_gap_at_key.next()],
5814 30 : },
5815 30 : query_lsn,
5816 30 : &mut ValuesReconstructState::new(),
5817 30 : &ctx,
5818 30 : )
5819 29 : .await;
5820 2 :
5821 30 : let expected_item = parent_gap_lsns
5822 30 : .iter()
5823 30 : .rev()
5824 68 : .find(|(lsn, _)| **lsn <= query_lsn);
5825 30 :
5826 30 : info!(
5827 2 : "Doing vectored read at LSN {}. Expecting image to be: {:?}",
5828 2 : query_lsn, expected_item
5829 2 : );
5830 2 :
5831 30 : match expected_item {
5832 26 : Some((_, img_value)) => {
5833 26 : let key_results = results.expect("No vectored get error expected");
5834 26 : let key_result = &key_results[&child_gap_at_key];
5835 26 : let returned_img = key_result
5836 26 : .as_ref()
5837 26 : .expect("No page reconstruct error expected");
5838 26 :
5839 26 : info!(
5840 2 : "Vectored read at LSN {} returned image {}",
5841 0 : query_lsn,
5842 0 : std::str::from_utf8(returned_img)?
5843 2 : );
5844 26 : assert_eq!(*returned_img, test_img(img_value));
5845 2 : }
5846 2 : None => {
5847 4 : assert!(matches!(results, Err(GetVectoredError::MissingKey(_))));
5848 2 : }
5849 2 : }
5850 2 : }
5851 2 :
5852 2 : Ok(())
5853 2 : }
5854 :
5855 : #[tokio::test]
5856 2 : async fn test_random_updates() -> anyhow::Result<()> {
5857 2 : let names_algorithms = [
5858 2 : ("test_random_updates_legacy", CompactionAlgorithm::Legacy),
5859 2 : ("test_random_updates_tiered", CompactionAlgorithm::Tiered),
5860 2 : ];
5861 6 : for (name, algorithm) in names_algorithms {
5862 96267 : test_random_updates_algorithm(name, algorithm).await?;
5863 2 : }
5864 2 : Ok(())
5865 2 : }
5866 :
5867 4 : async fn test_random_updates_algorithm(
5868 4 : name: &'static str,
5869 4 : compaction_algorithm: CompactionAlgorithm,
5870 4 : ) -> anyhow::Result<()> {
5871 4 : let mut harness = TenantHarness::create(name).await?;
5872 4 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
5873 4 : kind: compaction_algorithm,
5874 4 : };
5875 16 : let (tenant, ctx) = harness.load().await;
5876 4 : let tline = tenant
5877 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5878 7 : .await?;
5879 :
5880 : const NUM_KEYS: usize = 1000;
5881 4 : let cancel = CancellationToken::new();
5882 4 :
5883 4 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5884 4 : let mut test_key_end = test_key;
5885 4 : test_key_end.field6 = NUM_KEYS as u32;
5886 4 : tline.add_extra_test_dense_keyspace(KeySpace::single(test_key..test_key_end));
5887 4 :
5888 4 : let mut keyspace = KeySpaceAccum::new();
5889 4 :
5890 4 : // Track when each page was last modified. Used to assert that
5891 4 : // a read sees the latest page version.
5892 4 : let mut updated = [Lsn(0); NUM_KEYS];
5893 4 :
5894 4 : let mut lsn = Lsn(0x10);
5895 : #[allow(clippy::needless_range_loop)]
5896 4004 : for blknum in 0..NUM_KEYS {
5897 4000 : lsn = Lsn(lsn.0 + 0x10);
5898 4000 : test_key.field6 = blknum as u32;
5899 4000 : let mut writer = tline.writer().await;
5900 4000 : writer
5901 4000 : .put(
5902 4000 : test_key,
5903 4000 : lsn,
5904 4000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5905 4000 : &ctx,
5906 4000 : )
5907 4 : .await?;
5908 4000 : writer.finish_write(lsn);
5909 4000 : updated[blknum] = lsn;
5910 4000 : drop(writer);
5911 4000 :
5912 4000 : keyspace.add_key(test_key);
5913 : }
5914 :
5915 204 : for _ in 0..50 {
5916 200200 : for _ in 0..NUM_KEYS {
5917 200000 : lsn = Lsn(lsn.0 + 0x10);
5918 200000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
5919 200000 : test_key.field6 = blknum as u32;
5920 200000 : let mut writer = tline.writer().await;
5921 200000 : writer
5922 200000 : .put(
5923 200000 : test_key,
5924 200000 : lsn,
5925 200000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5926 200000 : &ctx,
5927 200000 : )
5928 198 : .await?;
5929 200000 : writer.finish_write(lsn);
5930 200000 : drop(writer);
5931 200000 : updated[blknum] = lsn;
5932 : }
5933 :
5934 : // Read all the blocks
5935 200000 : for (blknum, last_lsn) in updated.iter().enumerate() {
5936 200000 : test_key.field6 = blknum as u32;
5937 200000 : assert_eq!(
5938 200000 : tline.get(test_key, lsn, &ctx).await?,
5939 200000 : test_img(&format!("{} at {}", blknum, last_lsn))
5940 : );
5941 : }
5942 :
5943 : // Perform a cycle of flush, and GC
5944 201 : tline.freeze_and_flush().await?;
5945 200 : tenant
5946 200 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
5947 0 : .await?;
5948 : }
5949 :
5950 4 : Ok(())
5951 4 : }
5952 :
5953 : #[tokio::test]
5954 2 : async fn test_traverse_branches() -> anyhow::Result<()> {
5955 2 : let (tenant, ctx) = TenantHarness::create("test_traverse_branches")
5956 2 : .await?
5957 2 : .load()
5958 8 : .await;
5959 2 : let mut tline = tenant
5960 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5961 4 : .await?;
5962 2 :
5963 2 : const NUM_KEYS: usize = 1000;
5964 2 :
5965 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5966 2 :
5967 2 : let mut keyspace = KeySpaceAccum::new();
5968 2 :
5969 2 : let cancel = CancellationToken::new();
5970 2 :
5971 2 : // Track when each page was last modified. Used to assert that
5972 2 : // a read sees the latest page version.
5973 2 : let mut updated = [Lsn(0); NUM_KEYS];
5974 2 :
5975 2 : let mut lsn = Lsn(0x10);
5976 2 : #[allow(clippy::needless_range_loop)]
5977 2002 : for blknum in 0..NUM_KEYS {
5978 2000 : lsn = Lsn(lsn.0 + 0x10);
5979 2000 : test_key.field6 = blknum as u32;
5980 2000 : let mut writer = tline.writer().await;
5981 2000 : writer
5982 2000 : .put(
5983 2000 : test_key,
5984 2000 : lsn,
5985 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5986 2000 : &ctx,
5987 2000 : )
5988 2 : .await?;
5989 2000 : writer.finish_write(lsn);
5990 2000 : updated[blknum] = lsn;
5991 2000 : drop(writer);
5992 2000 :
5993 2000 : keyspace.add_key(test_key);
5994 2 : }
5995 2 :
5996 102 : for _ in 0..50 {
5997 100 : let new_tline_id = TimelineId::generate();
5998 100 : tenant
5999 100 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
6000 2 : .await?;
6001 100 : tline = tenant
6002 100 : .get_timeline(new_tline_id, true)
6003 100 : .expect("Should have the branched timeline");
6004 2 :
6005 100100 : for _ in 0..NUM_KEYS {
6006 100000 : lsn = Lsn(lsn.0 + 0x10);
6007 100000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6008 100000 : test_key.field6 = blknum as u32;
6009 100000 : let mut writer = tline.writer().await;
6010 100000 : writer
6011 100000 : .put(
6012 100000 : test_key,
6013 100000 : lsn,
6014 100000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6015 100000 : &ctx,
6016 100000 : )
6017 631 : .await?;
6018 100000 : println!("updating {} at {}", blknum, lsn);
6019 100000 : writer.finish_write(lsn);
6020 100000 : drop(writer);
6021 100000 : updated[blknum] = lsn;
6022 2 : }
6023 2 :
6024 2 : // Read all the blocks
6025 100000 : for (blknum, last_lsn) in updated.iter().enumerate() {
6026 100000 : test_key.field6 = blknum as u32;
6027 100000 : assert_eq!(
6028 100000 : tline.get(test_key, lsn, &ctx).await?,
6029 100000 : test_img(&format!("{} at {}", blknum, last_lsn))
6030 2 : );
6031 2 : }
6032 2 :
6033 2 : // Perform a cycle of flush, compact, and GC
6034 103 : tline.freeze_and_flush().await?;
6035 14963 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
6036 100 : tenant
6037 100 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
6038 2 : .await?;
6039 2 : }
6040 2 :
6041 2 : Ok(())
6042 2 : }
6043 :
6044 : #[tokio::test]
6045 2 : async fn test_traverse_ancestors() -> anyhow::Result<()> {
6046 2 : let (tenant, ctx) = TenantHarness::create("test_traverse_ancestors")
6047 2 : .await?
6048 2 : .load()
6049 7 : .await;
6050 2 : let mut tline = tenant
6051 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6052 4 : .await?;
6053 2 :
6054 2 : const NUM_KEYS: usize = 100;
6055 2 : const NUM_TLINES: usize = 50;
6056 2 :
6057 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6058 2 : // Track page mutation lsns across different timelines.
6059 2 : let mut updated = [[Lsn(0); NUM_KEYS]; NUM_TLINES];
6060 2 :
6061 2 : let mut lsn = Lsn(0x10);
6062 2 :
6063 2 : #[allow(clippy::needless_range_loop)]
6064 102 : for idx in 0..NUM_TLINES {
6065 100 : let new_tline_id = TimelineId::generate();
6066 100 : tenant
6067 100 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
6068 2 : .await?;
6069 100 : tline = tenant
6070 100 : .get_timeline(new_tline_id, true)
6071 100 : .expect("Should have the branched timeline");
6072 2 :
6073 10100 : for _ in 0..NUM_KEYS {
6074 10000 : lsn = Lsn(lsn.0 + 0x10);
6075 10000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6076 10000 : test_key.field6 = blknum as u32;
6077 10000 : let mut writer = tline.writer().await;
6078 10000 : writer
6079 10000 : .put(
6080 10000 : test_key,
6081 10000 : lsn,
6082 10000 : &Value::Image(test_img(&format!("{} {} at {}", idx, blknum, lsn))),
6083 10000 : &ctx,
6084 10000 : )
6085 88 : .await?;
6086 10000 : println!("updating [{}][{}] at {}", idx, blknum, lsn);
6087 10000 : writer.finish_write(lsn);
6088 10000 : drop(writer);
6089 10000 : updated[idx][blknum] = lsn;
6090 2 : }
6091 2 : }
6092 2 :
6093 2 : // Read pages from leaf timeline across all ancestors.
6094 100 : for (idx, lsns) in updated.iter().enumerate() {
6095 10000 : for (blknum, lsn) in lsns.iter().enumerate() {
6096 2 : // Skip empty mutations.
6097 10000 : if lsn.0 == 0 {
6098 3606 : continue;
6099 6394 : }
6100 6394 : println!("checking [{idx}][{blknum}] at {lsn}");
6101 6394 : test_key.field6 = blknum as u32;
6102 6394 : assert_eq!(
6103 6394 : tline.get(test_key, *lsn, &ctx).await?,
6104 6394 : test_img(&format!("{idx} {blknum} at {lsn}"))
6105 2 : );
6106 2 : }
6107 2 : }
6108 2 : Ok(())
6109 2 : }
6110 :
6111 : #[tokio::test]
6112 2 : async fn test_write_at_initdb_lsn_takes_optimization_code_path() -> anyhow::Result<()> {
6113 2 : let (tenant, ctx) = TenantHarness::create("test_empty_test_timeline_is_usable")
6114 2 : .await?
6115 2 : .load()
6116 8 : .await;
6117 2 :
6118 2 : let initdb_lsn = Lsn(0x20);
6119 2 : let utline = tenant
6120 2 : .create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)
6121 2 : .await?;
6122 2 : let tline = utline.raw_timeline().unwrap();
6123 2 :
6124 2 : // Spawn flush loop now so that we can set the `expect_initdb_optimization`
6125 2 : tline.maybe_spawn_flush_loop();
6126 2 :
6127 2 : // Make sure the timeline has the minimum set of required keys for operation.
6128 2 : // The only operation you can always do on an empty timeline is to `put` new data.
6129 2 : // Except if you `put` at `initdb_lsn`.
6130 2 : // In that case, there's an optimization to directly create image layers instead of delta layers.
6131 2 : // It uses `repartition()`, which assumes some keys to be present.
6132 2 : // Let's make sure the test timeline can handle that case.
6133 2 : {
6134 2 : let mut state = tline.flush_loop_state.lock().unwrap();
6135 2 : assert_eq!(
6136 2 : timeline::FlushLoopState::Running {
6137 2 : expect_initdb_optimization: false,
6138 2 : initdb_optimization_count: 0,
6139 2 : },
6140 2 : *state
6141 2 : );
6142 2 : *state = timeline::FlushLoopState::Running {
6143 2 : expect_initdb_optimization: true,
6144 2 : initdb_optimization_count: 0,
6145 2 : };
6146 2 : }
6147 2 :
6148 2 : // Make writes at the initdb_lsn. When we flush it below, it should be handled by the optimization.
6149 2 : // As explained above, the optimization requires some keys to be present.
6150 2 : // As per `create_empty_timeline` documentation, use init_empty to set them.
6151 2 : // This is what `create_test_timeline` does, by the way.
6152 2 : let mut modification = tline.begin_modification(initdb_lsn);
6153 2 : modification
6154 2 : .init_empty_test_timeline()
6155 2 : .context("init_empty_test_timeline")?;
6156 2 : modification
6157 2 : .commit(&ctx)
6158 2 : .await
6159 2 : .context("commit init_empty_test_timeline modification")?;
6160 2 :
6161 2 : // Do the flush. The flush code will check the expectations that we set above.
6162 2 : tline.freeze_and_flush().await?;
6163 2 :
6164 2 : // assert freeze_and_flush exercised the initdb optimization
6165 2 : {
6166 2 : let state = tline.flush_loop_state.lock().unwrap();
6167 2 : let timeline::FlushLoopState::Running {
6168 2 : expect_initdb_optimization,
6169 2 : initdb_optimization_count,
6170 2 : } = *state
6171 2 : else {
6172 2 : panic!("unexpected state: {:?}", *state);
6173 2 : };
6174 2 : assert!(expect_initdb_optimization);
6175 2 : assert!(initdb_optimization_count > 0);
6176 2 : }
6177 2 : Ok(())
6178 2 : }
6179 :
6180 : #[tokio::test]
6181 2 : async fn test_create_guard_crash() -> anyhow::Result<()> {
6182 2 : let name = "test_create_guard_crash";
6183 2 : let harness = TenantHarness::create(name).await?;
6184 2 : {
6185 8 : let (tenant, ctx) = harness.load().await;
6186 2 : let tline = tenant
6187 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
6188 2 : .await?;
6189 2 : // Leave the timeline ID in [`Tenant::timelines_creating`] to exclude attempting to create it again
6190 2 : let raw_tline = tline.raw_timeline().unwrap();
6191 2 : raw_tline
6192 2 : .shutdown(super::timeline::ShutdownMode::Hard)
6193 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))
6194 2 : .await;
6195 2 : std::mem::forget(tline);
6196 2 : }
6197 2 :
6198 8 : let (tenant, _) = harness.load().await;
6199 2 : match tenant.get_timeline(TIMELINE_ID, false) {
6200 2 : Ok(_) => panic!("timeline should've been removed during load"),
6201 2 : Err(e) => {
6202 2 : assert_eq!(
6203 2 : e,
6204 2 : GetTimelineError::NotFound {
6205 2 : tenant_id: tenant.tenant_shard_id,
6206 2 : timeline_id: TIMELINE_ID,
6207 2 : }
6208 2 : )
6209 2 : }
6210 2 : }
6211 2 :
6212 2 : assert!(!harness
6213 2 : .conf
6214 2 : .timeline_path(&tenant.tenant_shard_id, &TIMELINE_ID)
6215 2 : .exists());
6216 2 :
6217 2 : Ok(())
6218 2 : }
6219 :
6220 : #[tokio::test]
6221 2 : async fn test_read_at_max_lsn() -> anyhow::Result<()> {
6222 2 : let names_algorithms = [
6223 2 : ("test_read_at_max_lsn_legacy", CompactionAlgorithm::Legacy),
6224 2 : ("test_read_at_max_lsn_tiered", CompactionAlgorithm::Tiered),
6225 2 : ];
6226 6 : for (name, algorithm) in names_algorithms {
6227 32533 : test_read_at_max_lsn_algorithm(name, algorithm).await?;
6228 2 : }
6229 2 : Ok(())
6230 2 : }
6231 :
6232 4 : async fn test_read_at_max_lsn_algorithm(
6233 4 : name: &'static str,
6234 4 : compaction_algorithm: CompactionAlgorithm,
6235 4 : ) -> anyhow::Result<()> {
6236 4 : let mut harness = TenantHarness::create(name).await?;
6237 4 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
6238 4 : kind: compaction_algorithm,
6239 4 : };
6240 16 : let (tenant, ctx) = harness.load().await;
6241 4 : let tline = tenant
6242 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6243 7 : .await?;
6244 :
6245 4 : let lsn = Lsn(0x10);
6246 4 : let compact = false;
6247 32100 : bulk_insert_maybe_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000, compact).await?;
6248 :
6249 4 : let test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6250 4 : let read_lsn = Lsn(u64::MAX - 1);
6251 :
6252 410 : let result = tline.get(test_key, read_lsn, &ctx).await;
6253 4 : assert!(result.is_ok(), "result is not Ok: {}", result.unwrap_err());
6254 :
6255 4 : Ok(())
6256 4 : }
6257 :
6258 : #[tokio::test]
6259 2 : async fn test_metadata_scan() -> anyhow::Result<()> {
6260 2 : let harness = TenantHarness::create("test_metadata_scan").await?;
6261 8 : let (tenant, ctx) = harness.load().await;
6262 2 : let tline = tenant
6263 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6264 4 : .await?;
6265 2 :
6266 2 : const NUM_KEYS: usize = 1000;
6267 2 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
6268 2 :
6269 2 : let cancel = CancellationToken::new();
6270 2 :
6271 2 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
6272 2 : base_key.field1 = AUX_KEY_PREFIX;
6273 2 : let mut test_key = base_key;
6274 2 :
6275 2 : // Track when each page was last modified. Used to assert that
6276 2 : // a read sees the latest page version.
6277 2 : let mut updated = [Lsn(0); NUM_KEYS];
6278 2 :
6279 2 : let mut lsn = Lsn(0x10);
6280 2 : #[allow(clippy::needless_range_loop)]
6281 2002 : for blknum in 0..NUM_KEYS {
6282 2000 : lsn = Lsn(lsn.0 + 0x10);
6283 2000 : test_key.field6 = (blknum * STEP) as u32;
6284 2000 : let mut writer = tline.writer().await;
6285 2000 : writer
6286 2000 : .put(
6287 2000 : test_key,
6288 2000 : lsn,
6289 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6290 2000 : &ctx,
6291 2000 : )
6292 2 : .await?;
6293 2000 : writer.finish_write(lsn);
6294 2000 : updated[blknum] = lsn;
6295 2000 : drop(writer);
6296 2 : }
6297 2 :
6298 2 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
6299 2 :
6300 24 : for iter in 0..=10 {
6301 2 : // Read all the blocks
6302 22000 : for (blknum, last_lsn) in updated.iter().enumerate() {
6303 22000 : test_key.field6 = (blknum * STEP) as u32;
6304 22000 : assert_eq!(
6305 22000 : tline.get(test_key, lsn, &ctx).await?,
6306 22000 : test_img(&format!("{} at {}", blknum, last_lsn))
6307 2 : );
6308 2 : }
6309 2 :
6310 22 : let mut cnt = 0;
6311 22000 : for (key, value) in tline
6312 22 : .get_vectored_impl(
6313 22 : keyspace.clone(),
6314 22 : lsn,
6315 22 : &mut ValuesReconstructState::default(),
6316 22 : &ctx,
6317 22 : )
6318 1513 : .await?
6319 2 : {
6320 22000 : let blknum = key.field6 as usize;
6321 22000 : let value = value?;
6322 22000 : assert!(blknum % STEP == 0);
6323 22000 : let blknum = blknum / STEP;
6324 22000 : assert_eq!(
6325 22000 : value,
6326 22000 : test_img(&format!("{} at {}", blknum, updated[blknum]))
6327 22000 : );
6328 22000 : cnt += 1;
6329 2 : }
6330 2 :
6331 22 : assert_eq!(cnt, NUM_KEYS);
6332 2 :
6333 22022 : for _ in 0..NUM_KEYS {
6334 22000 : lsn = Lsn(lsn.0 + 0x10);
6335 22000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6336 22000 : test_key.field6 = (blknum * STEP) as u32;
6337 22000 : let mut writer = tline.writer().await;
6338 22000 : writer
6339 22000 : .put(
6340 22000 : test_key,
6341 22000 : lsn,
6342 22000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6343 22000 : &ctx,
6344 22000 : )
6345 128 : .await?;
6346 22000 : writer.finish_write(lsn);
6347 22000 : drop(writer);
6348 22000 : updated[blknum] = lsn;
6349 2 : }
6350 2 :
6351 2 : // Perform two cycles of flush, compact, and GC
6352 66 : for round in 0..2 {
6353 44 : tline.freeze_and_flush().await?;
6354 44 : tline
6355 44 : .compact(
6356 44 : &cancel,
6357 44 : if iter % 5 == 0 && round == 0 {
6358 6 : let mut flags = EnumSet::new();
6359 6 : flags.insert(CompactFlags::ForceImageLayerCreation);
6360 6 : flags.insert(CompactFlags::ForceRepartition);
6361 6 : flags
6362 2 : } else {
6363 38 : EnumSet::empty()
6364 2 : },
6365 44 : &ctx,
6366 2 : )
6367 7003 : .await?;
6368 44 : tenant
6369 44 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
6370 2 : .await?;
6371 2 : }
6372 2 : }
6373 2 :
6374 2 : Ok(())
6375 2 : }
6376 :
6377 : #[tokio::test]
6378 2 : async fn test_metadata_compaction_trigger() -> anyhow::Result<()> {
6379 2 : let harness = TenantHarness::create("test_metadata_compaction_trigger").await?;
6380 8 : let (tenant, ctx) = harness.load().await;
6381 2 : let tline = tenant
6382 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6383 4 : .await?;
6384 2 :
6385 2 : let cancel = CancellationToken::new();
6386 2 :
6387 2 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
6388 2 : base_key.field1 = AUX_KEY_PREFIX;
6389 2 : let test_key = base_key;
6390 2 : let mut lsn = Lsn(0x10);
6391 2 :
6392 42 : for _ in 0..20 {
6393 40 : lsn = Lsn(lsn.0 + 0x10);
6394 40 : let mut writer = tline.writer().await;
6395 40 : writer
6396 40 : .put(
6397 40 : test_key,
6398 40 : lsn,
6399 40 : &Value::Image(test_img(&format!("{} at {}", 0, lsn))),
6400 40 : &ctx,
6401 40 : )
6402 20 : .await?;
6403 40 : writer.finish_write(lsn);
6404 40 : drop(writer);
6405 40 : tline.freeze_and_flush().await?; // force create a delta layer
6406 2 : }
6407 2 :
6408 2 : let before_num_l0_delta_files =
6409 2 : tline.layers.read().await.layer_map()?.level0_deltas().len();
6410 2 :
6411 110 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
6412 2 :
6413 2 : let after_num_l0_delta_files = tline.layers.read().await.layer_map()?.level0_deltas().len();
6414 2 :
6415 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}");
6416 2 :
6417 2 : assert_eq!(
6418 4 : tline.get(test_key, lsn, &ctx).await?,
6419 2 : test_img(&format!("{} at {}", 0, lsn))
6420 2 : );
6421 2 :
6422 2 : Ok(())
6423 2 : }
6424 :
6425 : #[tokio::test]
6426 2 : async fn test_aux_file_e2e() {
6427 2 : let harness = TenantHarness::create("test_aux_file_e2e").await.unwrap();
6428 2 :
6429 8 : let (tenant, ctx) = harness.load().await;
6430 2 :
6431 2 : let mut lsn = Lsn(0x08);
6432 2 :
6433 2 : let tline: Arc<Timeline> = tenant
6434 2 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
6435 4 : .await
6436 2 : .unwrap();
6437 2 :
6438 2 : {
6439 2 : lsn += 8;
6440 2 : let mut modification = tline.begin_modification(lsn);
6441 2 : modification
6442 2 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
6443 2 : .await
6444 2 : .unwrap();
6445 2 : modification.commit(&ctx).await.unwrap();
6446 2 : }
6447 2 :
6448 2 : // we can read everything from the storage
6449 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6450 2 : assert_eq!(
6451 2 : files.get("pg_logical/mappings/test1"),
6452 2 : Some(&bytes::Bytes::from_static(b"first"))
6453 2 : );
6454 2 :
6455 2 : {
6456 2 : lsn += 8;
6457 2 : let mut modification = tline.begin_modification(lsn);
6458 2 : modification
6459 2 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
6460 2 : .await
6461 2 : .unwrap();
6462 2 : modification.commit(&ctx).await.unwrap();
6463 2 : }
6464 2 :
6465 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6466 2 : assert_eq!(
6467 2 : files.get("pg_logical/mappings/test2"),
6468 2 : Some(&bytes::Bytes::from_static(b"second"))
6469 2 : );
6470 2 :
6471 2 : let child = tenant
6472 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(lsn), &ctx)
6473 2 : .await
6474 2 : .unwrap();
6475 2 :
6476 2 : let files = child.list_aux_files(lsn, &ctx).await.unwrap();
6477 2 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
6478 2 : assert_eq!(files.get("pg_logical/mappings/test2"), None);
6479 2 : }
6480 :
6481 : #[tokio::test]
6482 2 : async fn test_metadata_image_creation() -> anyhow::Result<()> {
6483 2 : let harness = TenantHarness::create("test_metadata_image_creation").await?;
6484 8 : let (tenant, ctx) = harness.load().await;
6485 2 : let tline = tenant
6486 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6487 4 : .await?;
6488 2 :
6489 2 : const NUM_KEYS: usize = 1000;
6490 2 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
6491 2 :
6492 2 : let cancel = CancellationToken::new();
6493 2 :
6494 2 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
6495 2 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
6496 2 : let mut test_key = base_key;
6497 2 : let mut lsn = Lsn(0x10);
6498 2 :
6499 8 : async fn scan_with_statistics(
6500 8 : tline: &Timeline,
6501 8 : keyspace: &KeySpace,
6502 8 : lsn: Lsn,
6503 8 : ctx: &RequestContext,
6504 8 : ) -> anyhow::Result<(BTreeMap<Key, Result<Bytes, PageReconstructError>>, usize)> {
6505 8 : let mut reconstruct_state = ValuesReconstructState::default();
6506 8 : let res = tline
6507 8 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
6508 689 : .await?;
6509 8 : Ok((res, reconstruct_state.get_delta_layers_visited() as usize))
6510 8 : }
6511 2 :
6512 2 : #[allow(clippy::needless_range_loop)]
6513 2002 : for blknum in 0..NUM_KEYS {
6514 2000 : lsn = Lsn(lsn.0 + 0x10);
6515 2000 : test_key.field6 = (blknum * STEP) as u32;
6516 2000 : let mut writer = tline.writer().await;
6517 2000 : writer
6518 2000 : .put(
6519 2000 : test_key,
6520 2000 : lsn,
6521 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6522 2000 : &ctx,
6523 2000 : )
6524 2 : .await?;
6525 2000 : writer.finish_write(lsn);
6526 2000 : drop(writer);
6527 2 : }
6528 2 :
6529 2 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
6530 2 :
6531 22 : for iter in 1..=10 {
6532 20020 : for _ in 0..NUM_KEYS {
6533 20000 : lsn = Lsn(lsn.0 + 0x10);
6534 20000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6535 20000 : test_key.field6 = (blknum * STEP) as u32;
6536 20000 : let mut writer = tline.writer().await;
6537 20000 : writer
6538 20000 : .put(
6539 20000 : test_key,
6540 20000 : lsn,
6541 20000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6542 20000 : &ctx,
6543 20000 : )
6544 35 : .await?;
6545 20000 : writer.finish_write(lsn);
6546 20000 : drop(writer);
6547 2 : }
6548 2 :
6549 20 : tline.freeze_and_flush().await?;
6550 2 :
6551 20 : if iter % 5 == 0 {
6552 4 : let (_, before_delta_file_accessed) =
6553 681 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
6554 4 : tline
6555 4 : .compact(
6556 4 : &cancel,
6557 4 : {
6558 4 : let mut flags = EnumSet::new();
6559 4 : flags.insert(CompactFlags::ForceImageLayerCreation);
6560 4 : flags.insert(CompactFlags::ForceRepartition);
6561 4 : flags
6562 4 : },
6563 4 : &ctx,
6564 4 : )
6565 5076 : .await?;
6566 4 : let (_, after_delta_file_accessed) =
6567 8 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
6568 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}");
6569 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.
6570 4 : assert!(
6571 4 : after_delta_file_accessed <= 2,
6572 2 : "after_delta_file_accessed={after_delta_file_accessed}"
6573 2 : );
6574 16 : }
6575 2 : }
6576 2 :
6577 2 : Ok(())
6578 2 : }
6579 :
6580 : #[tokio::test]
6581 2 : async fn test_vectored_missing_data_key_reads() -> anyhow::Result<()> {
6582 2 : let harness = TenantHarness::create("test_vectored_missing_data_key_reads").await?;
6583 8 : let (tenant, ctx) = harness.load().await;
6584 2 :
6585 2 : let base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
6586 2 : let base_key_child = Key::from_hex("000000000033333333444444445500000001").unwrap();
6587 2 : let base_key_nonexist = Key::from_hex("000000000033333333444444445500000002").unwrap();
6588 2 :
6589 2 : let tline = tenant
6590 2 : .create_test_timeline_with_layers(
6591 2 : TIMELINE_ID,
6592 2 : Lsn(0x10),
6593 2 : DEFAULT_PG_VERSION,
6594 2 : &ctx,
6595 2 : Vec::new(), // delta layers
6596 2 : vec![(Lsn(0x20), vec![(base_key, test_img("data key 1"))])], // image layers
6597 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
6598 2 : )
6599 11 : .await?;
6600 2 : tline.add_extra_test_dense_keyspace(KeySpace::single(base_key..(base_key_nonexist.next())));
6601 2 :
6602 2 : let child = tenant
6603 2 : .branch_timeline_test_with_layers(
6604 2 : &tline,
6605 2 : NEW_TIMELINE_ID,
6606 2 : Some(Lsn(0x20)),
6607 2 : &ctx,
6608 2 : Vec::new(), // delta layers
6609 2 : vec![(Lsn(0x30), vec![(base_key_child, test_img("data key 2"))])], // image layers
6610 2 : Lsn(0x30),
6611 2 : )
6612 7 : .await
6613 2 : .unwrap();
6614 2 :
6615 2 : let lsn = Lsn(0x30);
6616 2 :
6617 2 : // test vectored get on parent timeline
6618 2 : assert_eq!(
6619 4 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
6620 2 : Some(test_img("data key 1"))
6621 2 : );
6622 2 : assert!(get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx)
6623 3 : .await
6624 2 : .unwrap_err()
6625 2 : .is_missing_key_error());
6626 2 : assert!(
6627 2 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx)
6628 2 : .await
6629 2 : .unwrap_err()
6630 2 : .is_missing_key_error()
6631 2 : );
6632 2 :
6633 2 : // test vectored get on child timeline
6634 2 : assert_eq!(
6635 2 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
6636 2 : Some(test_img("data key 1"))
6637 2 : );
6638 2 : assert_eq!(
6639 4 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
6640 2 : Some(test_img("data key 2"))
6641 2 : );
6642 2 : assert!(
6643 2 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx)
6644 2 : .await
6645 2 : .unwrap_err()
6646 2 : .is_missing_key_error()
6647 2 : );
6648 2 :
6649 2 : Ok(())
6650 2 : }
6651 :
6652 : #[tokio::test]
6653 2 : async fn test_vectored_missing_metadata_key_reads() -> anyhow::Result<()> {
6654 2 : let harness = TenantHarness::create("test_vectored_missing_metadata_key_reads").await?;
6655 7 : let (tenant, ctx) = harness.load().await;
6656 2 :
6657 2 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
6658 2 : let base_key_child = Key::from_hex("620000000033333333444444445500000001").unwrap();
6659 2 : let base_key_nonexist = Key::from_hex("620000000033333333444444445500000002").unwrap();
6660 2 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
6661 2 :
6662 2 : let tline = tenant
6663 2 : .create_test_timeline_with_layers(
6664 2 : TIMELINE_ID,
6665 2 : Lsn(0x10),
6666 2 : DEFAULT_PG_VERSION,
6667 2 : &ctx,
6668 2 : Vec::new(), // delta layers
6669 2 : vec![(Lsn(0x20), vec![(base_key, test_img("metadata key 1"))])], // image layers
6670 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
6671 2 : )
6672 11 : .await?;
6673 2 :
6674 2 : let child = tenant
6675 2 : .branch_timeline_test_with_layers(
6676 2 : &tline,
6677 2 : NEW_TIMELINE_ID,
6678 2 : Some(Lsn(0x20)),
6679 2 : &ctx,
6680 2 : Vec::new(), // delta layers
6681 2 : vec![(
6682 2 : Lsn(0x30),
6683 2 : vec![(base_key_child, test_img("metadata key 2"))],
6684 2 : )], // image layers
6685 2 : Lsn(0x30),
6686 2 : )
6687 7 : .await
6688 2 : .unwrap();
6689 2 :
6690 2 : let lsn = Lsn(0x30);
6691 2 :
6692 2 : // test vectored get on parent timeline
6693 2 : assert_eq!(
6694 4 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
6695 2 : Some(test_img("metadata key 1"))
6696 2 : );
6697 2 : assert_eq!(
6698 2 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx).await?,
6699 2 : None
6700 2 : );
6701 2 : assert_eq!(
6702 2 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx).await?,
6703 2 : None
6704 2 : );
6705 2 :
6706 2 : // test vectored get on child timeline
6707 2 : assert_eq!(
6708 2 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
6709 2 : None
6710 2 : );
6711 2 : assert_eq!(
6712 4 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
6713 2 : Some(test_img("metadata key 2"))
6714 2 : );
6715 2 : assert_eq!(
6716 2 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx).await?,
6717 2 : None
6718 2 : );
6719 2 :
6720 2 : Ok(())
6721 2 : }
6722 :
6723 36 : async fn get_vectored_impl_wrapper(
6724 36 : tline: &Arc<Timeline>,
6725 36 : key: Key,
6726 36 : lsn: Lsn,
6727 36 : ctx: &RequestContext,
6728 36 : ) -> Result<Option<Bytes>, GetVectoredError> {
6729 36 : let mut reconstruct_state = ValuesReconstructState::new();
6730 36 : let mut res = tline
6731 36 : .get_vectored_impl(
6732 36 : KeySpace::single(key..key.next()),
6733 36 : lsn,
6734 36 : &mut reconstruct_state,
6735 36 : ctx,
6736 36 : )
6737 33 : .await?;
6738 30 : Ok(res.pop_last().map(|(k, v)| {
6739 18 : assert_eq!(k, key);
6740 18 : v.unwrap()
6741 30 : }))
6742 36 : }
6743 :
6744 : #[tokio::test]
6745 2 : async fn test_metadata_tombstone_reads() -> anyhow::Result<()> {
6746 2 : let harness = TenantHarness::create("test_metadata_tombstone_reads").await?;
6747 8 : let (tenant, ctx) = harness.load().await;
6748 2 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
6749 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
6750 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
6751 2 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
6752 2 :
6753 2 : // We emulate the situation that the compaction algorithm creates an image layer that removes the tombstones
6754 2 : // Lsn 0x30 key0, key3, no key1+key2
6755 2 : // Lsn 0x20 key1+key2 tomestones
6756 2 : // Lsn 0x10 key1 in image, key2 in delta
6757 2 : let tline = tenant
6758 2 : .create_test_timeline_with_layers(
6759 2 : TIMELINE_ID,
6760 2 : Lsn(0x10),
6761 2 : DEFAULT_PG_VERSION,
6762 2 : &ctx,
6763 2 : // delta layers
6764 2 : vec![
6765 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6766 2 : Lsn(0x10)..Lsn(0x20),
6767 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
6768 2 : ),
6769 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6770 2 : Lsn(0x20)..Lsn(0x30),
6771 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
6772 2 : ),
6773 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6774 2 : Lsn(0x20)..Lsn(0x30),
6775 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
6776 2 : ),
6777 2 : ],
6778 2 : // image layers
6779 2 : vec![
6780 2 : (Lsn(0x10), vec![(key1, test_img("metadata key 1"))]),
6781 2 : (
6782 2 : Lsn(0x30),
6783 2 : vec![
6784 2 : (key0, test_img("metadata key 0")),
6785 2 : (key3, test_img("metadata key 3")),
6786 2 : ],
6787 2 : ),
6788 2 : ],
6789 2 : Lsn(0x30),
6790 2 : )
6791 38 : .await?;
6792 2 :
6793 2 : let lsn = Lsn(0x30);
6794 2 : let old_lsn = Lsn(0x20);
6795 2 :
6796 2 : assert_eq!(
6797 4 : get_vectored_impl_wrapper(&tline, key0, lsn, &ctx).await?,
6798 2 : Some(test_img("metadata key 0"))
6799 2 : );
6800 2 : assert_eq!(
6801 2 : get_vectored_impl_wrapper(&tline, key1, lsn, &ctx).await?,
6802 2 : None,
6803 2 : );
6804 2 : assert_eq!(
6805 2 : get_vectored_impl_wrapper(&tline, key2, lsn, &ctx).await?,
6806 2 : None,
6807 2 : );
6808 2 : assert_eq!(
6809 4 : get_vectored_impl_wrapper(&tline, key1, old_lsn, &ctx).await?,
6810 2 : Some(Bytes::new()),
6811 2 : );
6812 2 : assert_eq!(
6813 4 : get_vectored_impl_wrapper(&tline, key2, old_lsn, &ctx).await?,
6814 2 : Some(Bytes::new()),
6815 2 : );
6816 2 : assert_eq!(
6817 2 : get_vectored_impl_wrapper(&tline, key3, lsn, &ctx).await?,
6818 2 : Some(test_img("metadata key 3"))
6819 2 : );
6820 2 :
6821 2 : Ok(())
6822 2 : }
6823 :
6824 : #[tokio::test]
6825 2 : async fn test_metadata_tombstone_image_creation() {
6826 2 : let harness = TenantHarness::create("test_metadata_tombstone_image_creation")
6827 2 : .await
6828 2 : .unwrap();
6829 8 : let (tenant, ctx) = harness.load().await;
6830 2 :
6831 2 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
6832 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
6833 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
6834 2 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
6835 2 :
6836 2 : let tline = tenant
6837 2 : .create_test_timeline_with_layers(
6838 2 : TIMELINE_ID,
6839 2 : Lsn(0x10),
6840 2 : DEFAULT_PG_VERSION,
6841 2 : &ctx,
6842 2 : // delta layers
6843 2 : vec![
6844 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6845 2 : Lsn(0x10)..Lsn(0x20),
6846 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
6847 2 : ),
6848 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6849 2 : Lsn(0x20)..Lsn(0x30),
6850 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
6851 2 : ),
6852 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6853 2 : Lsn(0x20)..Lsn(0x30),
6854 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
6855 2 : ),
6856 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6857 2 : Lsn(0x30)..Lsn(0x40),
6858 2 : vec![
6859 2 : (key0, Lsn(0x30), Value::Image(test_img("metadata key 0"))),
6860 2 : (key3, Lsn(0x30), Value::Image(test_img("metadata key 3"))),
6861 2 : ],
6862 2 : ),
6863 2 : ],
6864 2 : // image layers
6865 2 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
6866 2 : Lsn(0x40),
6867 2 : )
6868 35 : .await
6869 2 : .unwrap();
6870 2 :
6871 2 : let cancel = CancellationToken::new();
6872 2 :
6873 2 : tline
6874 2 : .compact(
6875 2 : &cancel,
6876 2 : {
6877 2 : let mut flags = EnumSet::new();
6878 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
6879 2 : flags.insert(CompactFlags::ForceRepartition);
6880 2 : flags
6881 2 : },
6882 2 : &ctx,
6883 2 : )
6884 55 : .await
6885 2 : .unwrap();
6886 2 :
6887 2 : // Image layers are created at last_record_lsn
6888 2 : let images = tline
6889 2 : .inspect_image_layers(Lsn(0x40), &ctx)
6890 8 : .await
6891 2 : .unwrap()
6892 2 : .into_iter()
6893 18 : .filter(|(k, _)| k.is_metadata_key())
6894 2 : .collect::<Vec<_>>();
6895 2 : assert_eq!(images.len(), 2); // the image layer should only contain two existing keys, tombstones should be removed.
6896 2 : }
6897 :
6898 : #[tokio::test]
6899 2 : async fn test_metadata_tombstone_empty_image_creation() {
6900 2 : let harness = TenantHarness::create("test_metadata_tombstone_empty_image_creation")
6901 2 : .await
6902 2 : .unwrap();
6903 8 : let (tenant, ctx) = harness.load().await;
6904 2 :
6905 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
6906 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
6907 2 :
6908 2 : let tline = tenant
6909 2 : .create_test_timeline_with_layers(
6910 2 : TIMELINE_ID,
6911 2 : Lsn(0x10),
6912 2 : DEFAULT_PG_VERSION,
6913 2 : &ctx,
6914 2 : // delta layers
6915 2 : vec![
6916 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6917 2 : Lsn(0x10)..Lsn(0x20),
6918 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
6919 2 : ),
6920 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6921 2 : Lsn(0x20)..Lsn(0x30),
6922 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
6923 2 : ),
6924 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6925 2 : Lsn(0x20)..Lsn(0x30),
6926 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
6927 2 : ),
6928 2 : ],
6929 2 : // image layers
6930 2 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
6931 2 : Lsn(0x30),
6932 2 : )
6933 29 : .await
6934 2 : .unwrap();
6935 2 :
6936 2 : let cancel = CancellationToken::new();
6937 2 :
6938 2 : tline
6939 2 : .compact(
6940 2 : &cancel,
6941 2 : {
6942 2 : let mut flags = EnumSet::new();
6943 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
6944 2 : flags.insert(CompactFlags::ForceRepartition);
6945 2 : flags
6946 2 : },
6947 2 : &ctx,
6948 2 : )
6949 43 : .await
6950 2 : .unwrap();
6951 2 :
6952 2 : // Image layers are created at last_record_lsn
6953 2 : let images = tline
6954 2 : .inspect_image_layers(Lsn(0x30), &ctx)
6955 4 : .await
6956 2 : .unwrap()
6957 2 : .into_iter()
6958 14 : .filter(|(k, _)| k.is_metadata_key())
6959 2 : .collect::<Vec<_>>();
6960 2 : assert_eq!(images.len(), 0); // the image layer should not contain tombstones, or it is not created
6961 2 : }
6962 :
6963 : #[tokio::test]
6964 2 : async fn test_simple_bottom_most_compaction_images() -> anyhow::Result<()> {
6965 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_images").await?;
6966 8 : let (tenant, ctx) = harness.load().await;
6967 2 :
6968 102 : fn get_key(id: u32) -> Key {
6969 102 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
6970 102 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
6971 102 : key.field6 = id;
6972 102 : key
6973 102 : }
6974 2 :
6975 2 : // We create
6976 2 : // - one bottom-most image layer,
6977 2 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
6978 2 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
6979 2 : // - a delta layer D3 above the horizon.
6980 2 : //
6981 2 : // | D3 |
6982 2 : // | D1 |
6983 2 : // -| |-- gc horizon -----------------
6984 2 : // | | | D2 |
6985 2 : // --------- img layer ------------------
6986 2 : //
6987 2 : // What we should expact from this compaction is:
6988 2 : // | D3 |
6989 2 : // | Part of D1 |
6990 2 : // --------- img layer with D1+D2 at GC horizon------------------
6991 2 :
6992 2 : // img layer at 0x10
6993 2 : let img_layer = (0..10)
6994 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
6995 2 : .collect_vec();
6996 2 :
6997 2 : let delta1 = vec![
6998 2 : (
6999 2 : get_key(1),
7000 2 : Lsn(0x20),
7001 2 : Value::Image(Bytes::from("value 1@0x20")),
7002 2 : ),
7003 2 : (
7004 2 : get_key(2),
7005 2 : Lsn(0x30),
7006 2 : Value::Image(Bytes::from("value 2@0x30")),
7007 2 : ),
7008 2 : (
7009 2 : get_key(3),
7010 2 : Lsn(0x40),
7011 2 : Value::Image(Bytes::from("value 3@0x40")),
7012 2 : ),
7013 2 : ];
7014 2 : let delta2 = vec![
7015 2 : (
7016 2 : get_key(5),
7017 2 : Lsn(0x20),
7018 2 : Value::Image(Bytes::from("value 5@0x20")),
7019 2 : ),
7020 2 : (
7021 2 : get_key(6),
7022 2 : Lsn(0x20),
7023 2 : Value::Image(Bytes::from("value 6@0x20")),
7024 2 : ),
7025 2 : ];
7026 2 : let delta3 = vec![
7027 2 : (
7028 2 : get_key(8),
7029 2 : Lsn(0x48),
7030 2 : Value::Image(Bytes::from("value 8@0x48")),
7031 2 : ),
7032 2 : (
7033 2 : get_key(9),
7034 2 : Lsn(0x48),
7035 2 : Value::Image(Bytes::from("value 9@0x48")),
7036 2 : ),
7037 2 : ];
7038 2 :
7039 2 : let tline = tenant
7040 2 : .create_test_timeline_with_layers(
7041 2 : TIMELINE_ID,
7042 2 : Lsn(0x10),
7043 2 : DEFAULT_PG_VERSION,
7044 2 : &ctx,
7045 2 : vec![
7046 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
7047 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
7048 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
7049 2 : ], // delta layers
7050 2 : vec![(Lsn(0x10), img_layer)], // image layers
7051 2 : Lsn(0x50),
7052 2 : )
7053 47 : .await?;
7054 2 : {
7055 2 : // Update GC info
7056 2 : let mut guard = tline.gc_info.write().unwrap();
7057 2 : guard.cutoffs.time = Lsn(0x30);
7058 2 : guard.cutoffs.space = Lsn(0x30);
7059 2 : }
7060 2 :
7061 2 : let expected_result = [
7062 2 : Bytes::from_static(b"value 0@0x10"),
7063 2 : Bytes::from_static(b"value 1@0x20"),
7064 2 : Bytes::from_static(b"value 2@0x30"),
7065 2 : Bytes::from_static(b"value 3@0x40"),
7066 2 : Bytes::from_static(b"value 4@0x10"),
7067 2 : Bytes::from_static(b"value 5@0x20"),
7068 2 : Bytes::from_static(b"value 6@0x20"),
7069 2 : Bytes::from_static(b"value 7@0x10"),
7070 2 : Bytes::from_static(b"value 8@0x48"),
7071 2 : Bytes::from_static(b"value 9@0x48"),
7072 2 : ];
7073 2 :
7074 20 : for (idx, expected) in expected_result.iter().enumerate() {
7075 20 : assert_eq!(
7076 20 : tline
7077 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7078 23 : .await
7079 20 : .unwrap(),
7080 2 : expected
7081 2 : );
7082 2 : }
7083 2 :
7084 2 : let cancel = CancellationToken::new();
7085 2 : tline
7086 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7087 57 : .await
7088 2 : .unwrap();
7089 2 :
7090 20 : for (idx, expected) in expected_result.iter().enumerate() {
7091 20 : assert_eq!(
7092 20 : tline
7093 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7094 16 : .await
7095 20 : .unwrap(),
7096 2 : expected
7097 2 : );
7098 2 : }
7099 2 :
7100 2 : // Check if the image layer at the GC horizon contains exactly what we want
7101 2 : let image_at_gc_horizon = tline
7102 2 : .inspect_image_layers(Lsn(0x30), &ctx)
7103 2 : .await
7104 2 : .unwrap()
7105 2 : .into_iter()
7106 34 : .filter(|(k, _)| k.is_metadata_key())
7107 2 : .collect::<Vec<_>>();
7108 2 :
7109 2 : assert_eq!(image_at_gc_horizon.len(), 10);
7110 2 : let expected_result = [
7111 2 : Bytes::from_static(b"value 0@0x10"),
7112 2 : Bytes::from_static(b"value 1@0x20"),
7113 2 : Bytes::from_static(b"value 2@0x30"),
7114 2 : Bytes::from_static(b"value 3@0x10"),
7115 2 : Bytes::from_static(b"value 4@0x10"),
7116 2 : Bytes::from_static(b"value 5@0x20"),
7117 2 : Bytes::from_static(b"value 6@0x20"),
7118 2 : Bytes::from_static(b"value 7@0x10"),
7119 2 : Bytes::from_static(b"value 8@0x10"),
7120 2 : Bytes::from_static(b"value 9@0x10"),
7121 2 : ];
7122 22 : for idx in 0..10 {
7123 20 : assert_eq!(
7124 20 : image_at_gc_horizon[idx],
7125 20 : (get_key(idx as u32), expected_result[idx].clone())
7126 20 : );
7127 2 : }
7128 2 :
7129 2 : // Check if old layers are removed / new layers have the expected LSN
7130 2 : let mut all_layers = tline.inspect_historic_layers().await.unwrap();
7131 4 : all_layers.sort_by(|k1, k2| {
7132 4 : (
7133 4 : k1.is_delta,
7134 4 : k1.key_range.start,
7135 4 : k1.key_range.end,
7136 4 : k1.lsn_range.start,
7137 4 : k1.lsn_range.end,
7138 4 : )
7139 4 : .cmp(&(
7140 4 : k2.is_delta,
7141 4 : k2.key_range.start,
7142 4 : k2.key_range.end,
7143 4 : k2.lsn_range.start,
7144 4 : k2.lsn_range.end,
7145 4 : ))
7146 4 : });
7147 2 : assert_eq!(
7148 2 : all_layers,
7149 2 : vec![
7150 2 : // Image layer at GC horizon
7151 2 : PersistentLayerKey {
7152 2 : key_range: Key::MIN..Key::MAX,
7153 2 : lsn_range: Lsn(0x30)..Lsn(0x31),
7154 2 : is_delta: false
7155 2 : },
7156 2 : // The delta layer below the horizon
7157 2 : PersistentLayerKey {
7158 2 : key_range: get_key(3)..get_key(4),
7159 2 : lsn_range: Lsn(0x30)..Lsn(0x48),
7160 2 : is_delta: true
7161 2 : },
7162 2 : // The delta3 layer that should not be picked for the compaction
7163 2 : PersistentLayerKey {
7164 2 : key_range: get_key(8)..get_key(10),
7165 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
7166 2 : is_delta: true
7167 2 : }
7168 2 : ]
7169 2 : );
7170 2 :
7171 2 : // increase GC horizon and compact again
7172 2 : {
7173 2 : // Update GC info
7174 2 : let mut guard = tline.gc_info.write().unwrap();
7175 2 : guard.cutoffs.time = Lsn(0x40);
7176 2 : guard.cutoffs.space = Lsn(0x40);
7177 2 : }
7178 2 : tline
7179 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7180 44 : .await
7181 2 : .unwrap();
7182 2 :
7183 2 : Ok(())
7184 2 : }
7185 :
7186 : #[tokio::test]
7187 2 : async fn test_neon_test_record() -> anyhow::Result<()> {
7188 2 : let harness = TenantHarness::create("test_neon_test_record").await?;
7189 8 : let (tenant, ctx) = harness.load().await;
7190 2 :
7191 24 : fn get_key(id: u32) -> Key {
7192 24 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7193 24 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7194 24 : key.field6 = id;
7195 24 : key
7196 24 : }
7197 2 :
7198 2 : let delta1 = vec![
7199 2 : (
7200 2 : get_key(1),
7201 2 : Lsn(0x20),
7202 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
7203 2 : ),
7204 2 : (
7205 2 : get_key(1),
7206 2 : Lsn(0x30),
7207 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
7208 2 : ),
7209 2 : (get_key(2), Lsn(0x10), Value::Image("0x10".into())),
7210 2 : (
7211 2 : get_key(2),
7212 2 : Lsn(0x20),
7213 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
7214 2 : ),
7215 2 : (
7216 2 : get_key(2),
7217 2 : Lsn(0x30),
7218 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
7219 2 : ),
7220 2 : (get_key(3), Lsn(0x10), Value::Image("0x10".into())),
7221 2 : (
7222 2 : get_key(3),
7223 2 : Lsn(0x20),
7224 2 : Value::WalRecord(NeonWalRecord::wal_clear()),
7225 2 : ),
7226 2 : (get_key(4), Lsn(0x10), Value::Image("0x10".into())),
7227 2 : (
7228 2 : get_key(4),
7229 2 : Lsn(0x20),
7230 2 : Value::WalRecord(NeonWalRecord::wal_init()),
7231 2 : ),
7232 2 : ];
7233 2 : let image1 = vec![(get_key(1), "0x10".into())];
7234 2 :
7235 2 : let tline = tenant
7236 2 : .create_test_timeline_with_layers(
7237 2 : TIMELINE_ID,
7238 2 : Lsn(0x10),
7239 2 : DEFAULT_PG_VERSION,
7240 2 : &ctx,
7241 2 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
7242 2 : Lsn(0x10)..Lsn(0x40),
7243 2 : delta1,
7244 2 : )], // delta layers
7245 2 : vec![(Lsn(0x10), image1)], // image layers
7246 2 : Lsn(0x50),
7247 2 : )
7248 17 : .await?;
7249 2 :
7250 2 : assert_eq!(
7251 8 : tline.get(get_key(1), Lsn(0x50), &ctx).await?,
7252 2 : Bytes::from_static(b"0x10,0x20,0x30")
7253 2 : );
7254 2 : assert_eq!(
7255 2 : tline.get(get_key(2), Lsn(0x50), &ctx).await?,
7256 2 : Bytes::from_static(b"0x10,0x20,0x30")
7257 2 : );
7258 2 :
7259 2 : // Need to remove the limit of "Neon WAL redo requires base image".
7260 2 :
7261 2 : // assert_eq!(tline.get(get_key(3), Lsn(0x50), &ctx).await?, Bytes::new());
7262 2 : // assert_eq!(tline.get(get_key(4), Lsn(0x50), &ctx).await?, Bytes::new());
7263 2 :
7264 2 : Ok(())
7265 2 : }
7266 :
7267 : #[tokio::test(start_paused = true)]
7268 2 : async fn test_lsn_lease() -> anyhow::Result<()> {
7269 2 : let (tenant, ctx) = TenantHarness::create("test_lsn_lease")
7270 2 : .await
7271 2 : .unwrap()
7272 2 : .load()
7273 8 : .await;
7274 2 : // Advance to the lsn lease deadline so that GC is not blocked by
7275 2 : // initial transition into AttachedSingle.
7276 2 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
7277 2 : tokio::time::resume();
7278 2 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7279 2 :
7280 2 : let end_lsn = Lsn(0x100);
7281 2 : let image_layers = (0x20..=0x90)
7282 2 : .step_by(0x10)
7283 16 : .map(|n| {
7284 16 : (
7285 16 : Lsn(n),
7286 16 : vec![(key, test_img(&format!("data key at {:x}", n)))],
7287 16 : )
7288 16 : })
7289 2 : .collect();
7290 2 :
7291 2 : let timeline = tenant
7292 2 : .create_test_timeline_with_layers(
7293 2 : TIMELINE_ID,
7294 2 : Lsn(0x10),
7295 2 : DEFAULT_PG_VERSION,
7296 2 : &ctx,
7297 2 : Vec::new(),
7298 2 : image_layers,
7299 2 : end_lsn,
7300 2 : )
7301 60 : .await?;
7302 2 :
7303 2 : let leased_lsns = [0x30, 0x50, 0x70];
7304 2 : let mut leases = Vec::new();
7305 6 : leased_lsns.iter().for_each(|n| {
7306 6 : leases.push(
7307 6 : timeline
7308 6 : .init_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)
7309 6 : .expect("lease request should succeed"),
7310 6 : );
7311 6 : });
7312 2 :
7313 2 : let updated_lease_0 = timeline
7314 2 : .renew_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)
7315 2 : .expect("lease renewal should succeed");
7316 2 : assert_eq!(
7317 2 : updated_lease_0.valid_until, leases[0].valid_until,
7318 2 : " Renewing with shorter lease should not change the lease."
7319 2 : );
7320 2 :
7321 2 : let updated_lease_1 = timeline
7322 2 : .renew_lsn_lease(
7323 2 : Lsn(leased_lsns[1]),
7324 2 : timeline.get_lsn_lease_length() * 2,
7325 2 : &ctx,
7326 2 : )
7327 2 : .expect("lease renewal should succeed");
7328 2 : assert!(
7329 2 : updated_lease_1.valid_until > leases[1].valid_until,
7330 2 : "Renewing with a long lease should renew lease with later expiration time."
7331 2 : );
7332 2 :
7333 2 : // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
7334 2 : info!(
7335 2 : "latest_gc_cutoff_lsn: {}",
7336 0 : *timeline.get_latest_gc_cutoff_lsn()
7337 2 : );
7338 2 : timeline.force_set_disk_consistent_lsn(end_lsn);
7339 2 :
7340 2 : let res = tenant
7341 2 : .gc_iteration(
7342 2 : Some(TIMELINE_ID),
7343 2 : 0,
7344 2 : Duration::ZERO,
7345 2 : &CancellationToken::new(),
7346 2 : &ctx,
7347 2 : )
7348 2 : .await
7349 2 : .unwrap();
7350 2 :
7351 2 : // Keeping everything <= Lsn(0x80) b/c leases:
7352 2 : // 0/10: initdb layer
7353 2 : // (0/20..=0/70).step_by(0x10): image layers added when creating the timeline.
7354 2 : assert_eq!(res.layers_needed_by_leases, 7);
7355 2 : // Keeping 0/90 b/c it is the latest layer.
7356 2 : assert_eq!(res.layers_not_updated, 1);
7357 2 : // Removed 0/80.
7358 2 : assert_eq!(res.layers_removed, 1);
7359 2 :
7360 2 : // Make lease on a already GC-ed LSN.
7361 2 : // 0/80 does not have a valid lease + is below latest_gc_cutoff
7362 2 : assert!(Lsn(0x80) < *timeline.get_latest_gc_cutoff_lsn());
7363 2 : timeline
7364 2 : .init_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx)
7365 2 : .expect_err("lease request on GC-ed LSN should fail");
7366 2 :
7367 2 : // Should still be able to renew a currently valid lease
7368 2 : // Assumption: original lease to is still valid for 0/50.
7369 2 : // (use `Timeline::init_lsn_lease` for testing so it always does validation)
7370 2 : timeline
7371 2 : .init_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)
7372 2 : .expect("lease renewal with validation should succeed");
7373 2 :
7374 2 : Ok(())
7375 2 : }
7376 :
7377 : #[tokio::test]
7378 2 : async fn test_simple_bottom_most_compaction_deltas() -> anyhow::Result<()> {
7379 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_deltas").await?;
7380 8 : let (tenant, ctx) = harness.load().await;
7381 2 :
7382 118 : fn get_key(id: u32) -> Key {
7383 118 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7384 118 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7385 118 : key.field6 = id;
7386 118 : key
7387 118 : }
7388 2 :
7389 2 : // We create
7390 2 : // - one bottom-most image layer,
7391 2 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
7392 2 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
7393 2 : // - a delta layer D3 above the horizon.
7394 2 : //
7395 2 : // | D3 |
7396 2 : // | D1 |
7397 2 : // -| |-- gc horizon -----------------
7398 2 : // | | | D2 |
7399 2 : // --------- img layer ------------------
7400 2 : //
7401 2 : // What we should expact from this compaction is:
7402 2 : // | D3 |
7403 2 : // | Part of D1 |
7404 2 : // --------- img layer with D1+D2 at GC horizon------------------
7405 2 :
7406 2 : // img layer at 0x10
7407 2 : let img_layer = (0..10)
7408 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
7409 2 : .collect_vec();
7410 2 :
7411 2 : let delta1 = vec![
7412 2 : (
7413 2 : get_key(1),
7414 2 : Lsn(0x20),
7415 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7416 2 : ),
7417 2 : (
7418 2 : get_key(2),
7419 2 : Lsn(0x30),
7420 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7421 2 : ),
7422 2 : (
7423 2 : get_key(3),
7424 2 : Lsn(0x28),
7425 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
7426 2 : ),
7427 2 : (
7428 2 : get_key(3),
7429 2 : Lsn(0x30),
7430 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7431 2 : ),
7432 2 : (
7433 2 : get_key(3),
7434 2 : Lsn(0x40),
7435 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
7436 2 : ),
7437 2 : ];
7438 2 : let delta2 = vec![
7439 2 : (
7440 2 : get_key(5),
7441 2 : Lsn(0x20),
7442 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7443 2 : ),
7444 2 : (
7445 2 : get_key(6),
7446 2 : Lsn(0x20),
7447 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7448 2 : ),
7449 2 : ];
7450 2 : let delta3 = vec![
7451 2 : (
7452 2 : get_key(8),
7453 2 : Lsn(0x48),
7454 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
7455 2 : ),
7456 2 : (
7457 2 : get_key(9),
7458 2 : Lsn(0x48),
7459 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
7460 2 : ),
7461 2 : ];
7462 2 :
7463 2 : let tline = tenant
7464 2 : .create_test_timeline_with_layers(
7465 2 : TIMELINE_ID,
7466 2 : Lsn(0x10),
7467 2 : DEFAULT_PG_VERSION,
7468 2 : &ctx,
7469 2 : vec![
7470 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
7471 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
7472 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
7473 2 : ], // delta layers
7474 2 : vec![(Lsn(0x10), img_layer)], // image layers
7475 2 : Lsn(0x50),
7476 2 : )
7477 47 : .await?;
7478 2 : {
7479 2 : // Update GC info
7480 2 : let mut guard = tline.gc_info.write().unwrap();
7481 2 : *guard = GcInfo {
7482 2 : retain_lsns: vec![],
7483 2 : cutoffs: GcCutoffs {
7484 2 : time: Lsn(0x30),
7485 2 : space: Lsn(0x30),
7486 2 : },
7487 2 : leases: Default::default(),
7488 2 : within_ancestor_pitr: false,
7489 2 : };
7490 2 : }
7491 2 :
7492 2 : let expected_result = [
7493 2 : Bytes::from_static(b"value 0@0x10"),
7494 2 : Bytes::from_static(b"value 1@0x10@0x20"),
7495 2 : Bytes::from_static(b"value 2@0x10@0x30"),
7496 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
7497 2 : Bytes::from_static(b"value 4@0x10"),
7498 2 : Bytes::from_static(b"value 5@0x10@0x20"),
7499 2 : Bytes::from_static(b"value 6@0x10@0x20"),
7500 2 : Bytes::from_static(b"value 7@0x10"),
7501 2 : Bytes::from_static(b"value 8@0x10@0x48"),
7502 2 : Bytes::from_static(b"value 9@0x10@0x48"),
7503 2 : ];
7504 2 :
7505 2 : let expected_result_at_gc_horizon = [
7506 2 : Bytes::from_static(b"value 0@0x10"),
7507 2 : Bytes::from_static(b"value 1@0x10@0x20"),
7508 2 : Bytes::from_static(b"value 2@0x10@0x30"),
7509 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
7510 2 : Bytes::from_static(b"value 4@0x10"),
7511 2 : Bytes::from_static(b"value 5@0x10@0x20"),
7512 2 : Bytes::from_static(b"value 6@0x10@0x20"),
7513 2 : Bytes::from_static(b"value 7@0x10"),
7514 2 : Bytes::from_static(b"value 8@0x10"),
7515 2 : Bytes::from_static(b"value 9@0x10"),
7516 2 : ];
7517 2 :
7518 22 : for idx in 0..10 {
7519 20 : assert_eq!(
7520 20 : tline
7521 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7522 31 : .await
7523 20 : .unwrap(),
7524 20 : &expected_result[idx]
7525 2 : );
7526 20 : assert_eq!(
7527 20 : tline
7528 20 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
7529 15 : .await
7530 20 : .unwrap(),
7531 20 : &expected_result_at_gc_horizon[idx]
7532 2 : );
7533 2 : }
7534 2 :
7535 2 : let cancel = CancellationToken::new();
7536 2 : tline
7537 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7538 53 : .await
7539 2 : .unwrap();
7540 2 :
7541 22 : for idx in 0..10 {
7542 20 : assert_eq!(
7543 20 : tline
7544 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7545 20 : .await
7546 20 : .unwrap(),
7547 20 : &expected_result[idx]
7548 2 : );
7549 20 : assert_eq!(
7550 20 : tline
7551 20 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
7552 10 : .await
7553 20 : .unwrap(),
7554 20 : &expected_result_at_gc_horizon[idx]
7555 2 : );
7556 2 : }
7557 2 :
7558 2 : // increase GC horizon and compact again
7559 2 : {
7560 2 : // Update GC info
7561 2 : let mut guard = tline.gc_info.write().unwrap();
7562 2 : guard.cutoffs.time = Lsn(0x40);
7563 2 : guard.cutoffs.space = Lsn(0x40);
7564 2 : }
7565 2 : tline
7566 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7567 45 : .await
7568 2 : .unwrap();
7569 2 :
7570 2 : Ok(())
7571 2 : }
7572 :
7573 : #[tokio::test]
7574 2 : async fn test_generate_key_retention() -> anyhow::Result<()> {
7575 2 : let harness = TenantHarness::create("test_generate_key_retention").await?;
7576 8 : let (tenant, ctx) = harness.load().await;
7577 2 : let tline = tenant
7578 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7579 4 : .await?;
7580 2 : tline.force_advance_lsn(Lsn(0x70));
7581 2 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7582 2 : let history = vec![
7583 2 : (
7584 2 : key,
7585 2 : Lsn(0x10),
7586 2 : Value::Image(Bytes::copy_from_slice(b"0x10")),
7587 2 : ),
7588 2 : (
7589 2 : key,
7590 2 : Lsn(0x20),
7591 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7592 2 : ),
7593 2 : (
7594 2 : key,
7595 2 : Lsn(0x30),
7596 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
7597 2 : ),
7598 2 : (
7599 2 : key,
7600 2 : Lsn(0x40),
7601 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
7602 2 : ),
7603 2 : (
7604 2 : key,
7605 2 : Lsn(0x50),
7606 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
7607 2 : ),
7608 2 : (
7609 2 : key,
7610 2 : Lsn(0x60),
7611 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7612 2 : ),
7613 2 : (
7614 2 : key,
7615 2 : Lsn(0x70),
7616 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7617 2 : ),
7618 2 : (
7619 2 : key,
7620 2 : Lsn(0x80),
7621 2 : Value::Image(Bytes::copy_from_slice(
7622 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
7623 2 : )),
7624 2 : ),
7625 2 : (
7626 2 : key,
7627 2 : Lsn(0x90),
7628 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
7629 2 : ),
7630 2 : ];
7631 2 : let res = tline
7632 2 : .generate_key_retention(
7633 2 : key,
7634 2 : &history,
7635 2 : Lsn(0x60),
7636 2 : &[Lsn(0x20), Lsn(0x40), Lsn(0x50)],
7637 2 : 3,
7638 2 : None,
7639 2 : )
7640 2 : .await
7641 2 : .unwrap();
7642 2 : let expected_res = KeyHistoryRetention {
7643 2 : below_horizon: vec![
7644 2 : (
7645 2 : Lsn(0x20),
7646 2 : KeyLogAtLsn(vec![(
7647 2 : Lsn(0x20),
7648 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20")),
7649 2 : )]),
7650 2 : ),
7651 2 : (
7652 2 : Lsn(0x40),
7653 2 : KeyLogAtLsn(vec![
7654 2 : (
7655 2 : Lsn(0x30),
7656 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
7657 2 : ),
7658 2 : (
7659 2 : Lsn(0x40),
7660 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
7661 2 : ),
7662 2 : ]),
7663 2 : ),
7664 2 : (
7665 2 : Lsn(0x50),
7666 2 : KeyLogAtLsn(vec![(
7667 2 : Lsn(0x50),
7668 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40;0x50")),
7669 2 : )]),
7670 2 : ),
7671 2 : (
7672 2 : Lsn(0x60),
7673 2 : KeyLogAtLsn(vec![(
7674 2 : Lsn(0x60),
7675 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7676 2 : )]),
7677 2 : ),
7678 2 : ],
7679 2 : above_horizon: KeyLogAtLsn(vec![
7680 2 : (
7681 2 : Lsn(0x70),
7682 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7683 2 : ),
7684 2 : (
7685 2 : Lsn(0x80),
7686 2 : Value::Image(Bytes::copy_from_slice(
7687 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
7688 2 : )),
7689 2 : ),
7690 2 : (
7691 2 : Lsn(0x90),
7692 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
7693 2 : ),
7694 2 : ]),
7695 2 : };
7696 2 : assert_eq!(res, expected_res);
7697 2 :
7698 2 : // We expect GC-compaction to run with the original GC. This would create a situation that
7699 2 : // the original GC algorithm removes some delta layers b/c there are full image coverage,
7700 2 : // therefore causing some keys to have an incomplete history below the lowest retain LSN.
7701 2 : // For example, we have
7702 2 : // ```plain
7703 2 : // init delta @ 0x10, image @ 0x20, delta @ 0x30 (gc_horizon), image @ 0x40.
7704 2 : // ```
7705 2 : // Now the GC horizon moves up, and we have
7706 2 : // ```plain
7707 2 : // init delta @ 0x10, image @ 0x20, delta @ 0x30, image @ 0x40 (gc_horizon)
7708 2 : // ```
7709 2 : // The original GC algorithm kicks in, and removes delta @ 0x10, image @ 0x20.
7710 2 : // We will end up with
7711 2 : // ```plain
7712 2 : // delta @ 0x30, image @ 0x40 (gc_horizon)
7713 2 : // ```
7714 2 : // Now we run the GC-compaction, and this key does not have a full history.
7715 2 : // We should be able to handle this partial history and drop everything before the
7716 2 : // gc_horizon image.
7717 2 :
7718 2 : let history = vec![
7719 2 : (
7720 2 : key,
7721 2 : Lsn(0x20),
7722 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7723 2 : ),
7724 2 : (
7725 2 : key,
7726 2 : Lsn(0x30),
7727 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
7728 2 : ),
7729 2 : (
7730 2 : key,
7731 2 : Lsn(0x40),
7732 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
7733 2 : ),
7734 2 : (
7735 2 : key,
7736 2 : Lsn(0x50),
7737 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
7738 2 : ),
7739 2 : (
7740 2 : key,
7741 2 : Lsn(0x60),
7742 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7743 2 : ),
7744 2 : (
7745 2 : key,
7746 2 : Lsn(0x70),
7747 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7748 2 : ),
7749 2 : (
7750 2 : key,
7751 2 : Lsn(0x80),
7752 2 : Value::Image(Bytes::copy_from_slice(
7753 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
7754 2 : )),
7755 2 : ),
7756 2 : (
7757 2 : key,
7758 2 : Lsn(0x90),
7759 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
7760 2 : ),
7761 2 : ];
7762 2 : let res = tline
7763 2 : .generate_key_retention(key, &history, Lsn(0x60), &[Lsn(0x40), Lsn(0x50)], 3, None)
7764 2 : .await
7765 2 : .unwrap();
7766 2 : let expected_res = KeyHistoryRetention {
7767 2 : below_horizon: vec![
7768 2 : (
7769 2 : Lsn(0x40),
7770 2 : KeyLogAtLsn(vec![(
7771 2 : Lsn(0x40),
7772 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
7773 2 : )]),
7774 2 : ),
7775 2 : (
7776 2 : Lsn(0x50),
7777 2 : KeyLogAtLsn(vec![(
7778 2 : Lsn(0x50),
7779 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
7780 2 : )]),
7781 2 : ),
7782 2 : (
7783 2 : Lsn(0x60),
7784 2 : KeyLogAtLsn(vec![(
7785 2 : Lsn(0x60),
7786 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7787 2 : )]),
7788 2 : ),
7789 2 : ],
7790 2 : above_horizon: KeyLogAtLsn(vec![
7791 2 : (
7792 2 : Lsn(0x70),
7793 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7794 2 : ),
7795 2 : (
7796 2 : Lsn(0x80),
7797 2 : Value::Image(Bytes::copy_from_slice(
7798 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
7799 2 : )),
7800 2 : ),
7801 2 : (
7802 2 : Lsn(0x90),
7803 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
7804 2 : ),
7805 2 : ]),
7806 2 : };
7807 2 : assert_eq!(res, expected_res);
7808 2 :
7809 2 : // In case of branch compaction, the branch itself does not have the full history, and we need to provide
7810 2 : // the ancestor image in the test case.
7811 2 :
7812 2 : let history = vec![
7813 2 : (
7814 2 : key,
7815 2 : Lsn(0x20),
7816 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7817 2 : ),
7818 2 : (
7819 2 : key,
7820 2 : Lsn(0x30),
7821 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
7822 2 : ),
7823 2 : (
7824 2 : key,
7825 2 : Lsn(0x40),
7826 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
7827 2 : ),
7828 2 : (
7829 2 : key,
7830 2 : Lsn(0x70),
7831 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7832 2 : ),
7833 2 : ];
7834 2 : let res = tline
7835 2 : .generate_key_retention(
7836 2 : key,
7837 2 : &history,
7838 2 : Lsn(0x60),
7839 2 : &[],
7840 2 : 3,
7841 2 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
7842 2 : )
7843 2 : .await
7844 2 : .unwrap();
7845 2 : let expected_res = KeyHistoryRetention {
7846 2 : below_horizon: vec![(
7847 2 : Lsn(0x60),
7848 2 : KeyLogAtLsn(vec![(
7849 2 : Lsn(0x60),
7850 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")), // use the ancestor image to reconstruct the page
7851 2 : )]),
7852 2 : )],
7853 2 : above_horizon: KeyLogAtLsn(vec![(
7854 2 : Lsn(0x70),
7855 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7856 2 : )]),
7857 2 : };
7858 2 : assert_eq!(res, expected_res);
7859 2 :
7860 2 : let history = vec![
7861 2 : (
7862 2 : key,
7863 2 : Lsn(0x20),
7864 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7865 2 : ),
7866 2 : (
7867 2 : key,
7868 2 : Lsn(0x40),
7869 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
7870 2 : ),
7871 2 : (
7872 2 : key,
7873 2 : Lsn(0x60),
7874 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7875 2 : ),
7876 2 : (
7877 2 : key,
7878 2 : Lsn(0x70),
7879 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7880 2 : ),
7881 2 : ];
7882 2 : let res = tline
7883 2 : .generate_key_retention(
7884 2 : key,
7885 2 : &history,
7886 2 : Lsn(0x60),
7887 2 : &[Lsn(0x30)],
7888 2 : 3,
7889 2 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
7890 2 : )
7891 2 : .await
7892 2 : .unwrap();
7893 2 : let expected_res = KeyHistoryRetention {
7894 2 : below_horizon: vec![
7895 2 : (
7896 2 : Lsn(0x30),
7897 2 : KeyLogAtLsn(vec![(
7898 2 : Lsn(0x20),
7899 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7900 2 : )]),
7901 2 : ),
7902 2 : (
7903 2 : Lsn(0x60),
7904 2 : KeyLogAtLsn(vec![(
7905 2 : Lsn(0x60),
7906 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x40;0x60")),
7907 2 : )]),
7908 2 : ),
7909 2 : ],
7910 2 : above_horizon: KeyLogAtLsn(vec![(
7911 2 : Lsn(0x70),
7912 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7913 2 : )]),
7914 2 : };
7915 2 : assert_eq!(res, expected_res);
7916 2 :
7917 2 : Ok(())
7918 2 : }
7919 :
7920 : #[tokio::test]
7921 2 : async fn test_simple_bottom_most_compaction_with_retain_lsns() -> anyhow::Result<()> {
7922 2 : let harness =
7923 2 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns").await?;
7924 8 : let (tenant, ctx) = harness.load().await;
7925 2 :
7926 518 : fn get_key(id: u32) -> Key {
7927 518 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7928 518 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7929 518 : key.field6 = id;
7930 518 : key
7931 518 : }
7932 2 :
7933 2 : let img_layer = (0..10)
7934 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
7935 2 : .collect_vec();
7936 2 :
7937 2 : let delta1 = vec![
7938 2 : (
7939 2 : get_key(1),
7940 2 : Lsn(0x20),
7941 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7942 2 : ),
7943 2 : (
7944 2 : get_key(2),
7945 2 : Lsn(0x30),
7946 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7947 2 : ),
7948 2 : (
7949 2 : get_key(3),
7950 2 : Lsn(0x28),
7951 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
7952 2 : ),
7953 2 : (
7954 2 : get_key(3),
7955 2 : Lsn(0x30),
7956 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7957 2 : ),
7958 2 : (
7959 2 : get_key(3),
7960 2 : Lsn(0x40),
7961 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
7962 2 : ),
7963 2 : ];
7964 2 : let delta2 = vec![
7965 2 : (
7966 2 : get_key(5),
7967 2 : Lsn(0x20),
7968 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7969 2 : ),
7970 2 : (
7971 2 : get_key(6),
7972 2 : Lsn(0x20),
7973 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7974 2 : ),
7975 2 : ];
7976 2 : let delta3 = vec![
7977 2 : (
7978 2 : get_key(8),
7979 2 : Lsn(0x48),
7980 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
7981 2 : ),
7982 2 : (
7983 2 : get_key(9),
7984 2 : Lsn(0x48),
7985 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
7986 2 : ),
7987 2 : ];
7988 2 :
7989 2 : let tline = tenant
7990 2 : .create_test_timeline_with_layers(
7991 2 : TIMELINE_ID,
7992 2 : Lsn(0x10),
7993 2 : DEFAULT_PG_VERSION,
7994 2 : &ctx,
7995 2 : vec![
7996 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
7997 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
7998 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
7999 2 : ], // delta layers
8000 2 : vec![(Lsn(0x10), img_layer)], // image layers
8001 2 : Lsn(0x50),
8002 2 : )
8003 47 : .await?;
8004 2 : {
8005 2 : // Update GC info
8006 2 : let mut guard = tline.gc_info.write().unwrap();
8007 2 : *guard = GcInfo {
8008 2 : retain_lsns: vec![
8009 2 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
8010 2 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
8011 2 : ],
8012 2 : cutoffs: GcCutoffs {
8013 2 : time: Lsn(0x30),
8014 2 : space: Lsn(0x30),
8015 2 : },
8016 2 : leases: Default::default(),
8017 2 : within_ancestor_pitr: false,
8018 2 : };
8019 2 : }
8020 2 :
8021 2 : let expected_result = [
8022 2 : Bytes::from_static(b"value 0@0x10"),
8023 2 : Bytes::from_static(b"value 1@0x10@0x20"),
8024 2 : Bytes::from_static(b"value 2@0x10@0x30"),
8025 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
8026 2 : Bytes::from_static(b"value 4@0x10"),
8027 2 : Bytes::from_static(b"value 5@0x10@0x20"),
8028 2 : Bytes::from_static(b"value 6@0x10@0x20"),
8029 2 : Bytes::from_static(b"value 7@0x10"),
8030 2 : Bytes::from_static(b"value 8@0x10@0x48"),
8031 2 : Bytes::from_static(b"value 9@0x10@0x48"),
8032 2 : ];
8033 2 :
8034 2 : let expected_result_at_gc_horizon = [
8035 2 : Bytes::from_static(b"value 0@0x10"),
8036 2 : Bytes::from_static(b"value 1@0x10@0x20"),
8037 2 : Bytes::from_static(b"value 2@0x10@0x30"),
8038 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
8039 2 : Bytes::from_static(b"value 4@0x10"),
8040 2 : Bytes::from_static(b"value 5@0x10@0x20"),
8041 2 : Bytes::from_static(b"value 6@0x10@0x20"),
8042 2 : Bytes::from_static(b"value 7@0x10"),
8043 2 : Bytes::from_static(b"value 8@0x10"),
8044 2 : Bytes::from_static(b"value 9@0x10"),
8045 2 : ];
8046 2 :
8047 2 : let expected_result_at_lsn_20 = [
8048 2 : Bytes::from_static(b"value 0@0x10"),
8049 2 : Bytes::from_static(b"value 1@0x10@0x20"),
8050 2 : Bytes::from_static(b"value 2@0x10"),
8051 2 : Bytes::from_static(b"value 3@0x10"),
8052 2 : Bytes::from_static(b"value 4@0x10"),
8053 2 : Bytes::from_static(b"value 5@0x10@0x20"),
8054 2 : Bytes::from_static(b"value 6@0x10@0x20"),
8055 2 : Bytes::from_static(b"value 7@0x10"),
8056 2 : Bytes::from_static(b"value 8@0x10"),
8057 2 : Bytes::from_static(b"value 9@0x10"),
8058 2 : ];
8059 2 :
8060 2 : let expected_result_at_lsn_10 = [
8061 2 : Bytes::from_static(b"value 0@0x10"),
8062 2 : Bytes::from_static(b"value 1@0x10"),
8063 2 : Bytes::from_static(b"value 2@0x10"),
8064 2 : Bytes::from_static(b"value 3@0x10"),
8065 2 : Bytes::from_static(b"value 4@0x10"),
8066 2 : Bytes::from_static(b"value 5@0x10"),
8067 2 : Bytes::from_static(b"value 6@0x10"),
8068 2 : Bytes::from_static(b"value 7@0x10"),
8069 2 : Bytes::from_static(b"value 8@0x10"),
8070 2 : Bytes::from_static(b"value 9@0x10"),
8071 2 : ];
8072 2 :
8073 12 : let verify_result = || async {
8074 12 : let gc_horizon = {
8075 12 : let gc_info = tline.gc_info.read().unwrap();
8076 12 : gc_info.cutoffs.time
8077 2 : };
8078 132 : for idx in 0..10 {
8079 120 : assert_eq!(
8080 120 : tline
8081 120 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8082 121 : .await
8083 120 : .unwrap(),
8084 120 : &expected_result[idx]
8085 2 : );
8086 120 : assert_eq!(
8087 120 : tline
8088 120 : .get(get_key(idx as u32), gc_horizon, &ctx)
8089 93 : .await
8090 120 : .unwrap(),
8091 120 : &expected_result_at_gc_horizon[idx]
8092 2 : );
8093 120 : assert_eq!(
8094 120 : tline
8095 120 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
8096 86 : .await
8097 120 : .unwrap(),
8098 120 : &expected_result_at_lsn_20[idx]
8099 2 : );
8100 120 : assert_eq!(
8101 120 : tline
8102 120 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
8103 61 : .await
8104 120 : .unwrap(),
8105 120 : &expected_result_at_lsn_10[idx]
8106 2 : );
8107 2 : }
8108 24 : };
8109 2 :
8110 69 : verify_result().await;
8111 2 :
8112 2 : let cancel = CancellationToken::new();
8113 2 : let mut dryrun_flags = EnumSet::new();
8114 2 : dryrun_flags.insert(CompactFlags::DryRun);
8115 2 :
8116 2 : tline
8117 2 : .compact_with_gc(&cancel, dryrun_flags, &ctx)
8118 48 : .await
8119 2 : .unwrap();
8120 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
8121 2 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
8122 57 : verify_result().await;
8123 2 :
8124 2 : tline
8125 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8126 53 : .await
8127 2 : .unwrap();
8128 64 : verify_result().await;
8129 2 :
8130 2 : // compact again
8131 2 : tline
8132 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8133 40 : .await
8134 2 : .unwrap();
8135 57 : verify_result().await;
8136 2 :
8137 2 : // increase GC horizon and compact again
8138 2 : {
8139 2 : // Update GC info
8140 2 : let mut guard = tline.gc_info.write().unwrap();
8141 2 : guard.cutoffs.time = Lsn(0x38);
8142 2 : guard.cutoffs.space = Lsn(0x38);
8143 2 : }
8144 2 : tline
8145 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8146 39 : .await
8147 2 : .unwrap();
8148 57 : verify_result().await; // no wals between 0x30 and 0x38, so we should obtain the same result
8149 2 :
8150 2 : // not increasing the GC horizon and compact again
8151 2 : tline
8152 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8153 40 : .await
8154 2 : .unwrap();
8155 57 : verify_result().await;
8156 2 :
8157 2 : Ok(())
8158 2 : }
8159 :
8160 : #[tokio::test]
8161 2 : async fn test_simple_bottom_most_compaction_with_retain_lsns_single_key() -> anyhow::Result<()>
8162 2 : {
8163 2 : let harness =
8164 2 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns_single_key")
8165 2 : .await?;
8166 8 : let (tenant, ctx) = harness.load().await;
8167 2 :
8168 352 : fn get_key(id: u32) -> Key {
8169 352 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8170 352 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8171 352 : key.field6 = id;
8172 352 : key
8173 352 : }
8174 2 :
8175 2 : let img_layer = (0..10)
8176 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8177 2 : .collect_vec();
8178 2 :
8179 2 : let delta1 = vec![
8180 2 : (
8181 2 : get_key(1),
8182 2 : Lsn(0x20),
8183 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8184 2 : ),
8185 2 : (
8186 2 : get_key(1),
8187 2 : Lsn(0x28),
8188 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8189 2 : ),
8190 2 : ];
8191 2 : let delta2 = vec![
8192 2 : (
8193 2 : get_key(1),
8194 2 : Lsn(0x30),
8195 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8196 2 : ),
8197 2 : (
8198 2 : get_key(1),
8199 2 : Lsn(0x38),
8200 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
8201 2 : ),
8202 2 : ];
8203 2 : let delta3 = vec![
8204 2 : (
8205 2 : get_key(8),
8206 2 : Lsn(0x48),
8207 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8208 2 : ),
8209 2 : (
8210 2 : get_key(9),
8211 2 : Lsn(0x48),
8212 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8213 2 : ),
8214 2 : ];
8215 2 :
8216 2 : let tline = tenant
8217 2 : .create_test_timeline_with_layers(
8218 2 : TIMELINE_ID,
8219 2 : Lsn(0x10),
8220 2 : DEFAULT_PG_VERSION,
8221 2 : &ctx,
8222 2 : vec![
8223 2 : // delta1 and delta 2 only contain a single key but multiple updates
8224 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x30), delta1),
8225 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
8226 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x50), delta3),
8227 2 : ], // delta layers
8228 2 : vec![(Lsn(0x10), img_layer)], // image layers
8229 2 : Lsn(0x50),
8230 2 : )
8231 47 : .await?;
8232 2 : {
8233 2 : // Update GC info
8234 2 : let mut guard = tline.gc_info.write().unwrap();
8235 2 : *guard = GcInfo {
8236 2 : retain_lsns: vec![
8237 2 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
8238 2 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
8239 2 : ],
8240 2 : cutoffs: GcCutoffs {
8241 2 : time: Lsn(0x30),
8242 2 : space: Lsn(0x30),
8243 2 : },
8244 2 : leases: Default::default(),
8245 2 : within_ancestor_pitr: false,
8246 2 : };
8247 2 : }
8248 2 :
8249 2 : let expected_result = [
8250 2 : Bytes::from_static(b"value 0@0x10"),
8251 2 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
8252 2 : Bytes::from_static(b"value 2@0x10"),
8253 2 : Bytes::from_static(b"value 3@0x10"),
8254 2 : Bytes::from_static(b"value 4@0x10"),
8255 2 : Bytes::from_static(b"value 5@0x10"),
8256 2 : Bytes::from_static(b"value 6@0x10"),
8257 2 : Bytes::from_static(b"value 7@0x10"),
8258 2 : Bytes::from_static(b"value 8@0x10@0x48"),
8259 2 : Bytes::from_static(b"value 9@0x10@0x48"),
8260 2 : ];
8261 2 :
8262 2 : let expected_result_at_gc_horizon = [
8263 2 : Bytes::from_static(b"value 0@0x10"),
8264 2 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
8265 2 : Bytes::from_static(b"value 2@0x10"),
8266 2 : Bytes::from_static(b"value 3@0x10"),
8267 2 : Bytes::from_static(b"value 4@0x10"),
8268 2 : Bytes::from_static(b"value 5@0x10"),
8269 2 : Bytes::from_static(b"value 6@0x10"),
8270 2 : Bytes::from_static(b"value 7@0x10"),
8271 2 : Bytes::from_static(b"value 8@0x10"),
8272 2 : Bytes::from_static(b"value 9@0x10"),
8273 2 : ];
8274 2 :
8275 2 : let expected_result_at_lsn_20 = [
8276 2 : Bytes::from_static(b"value 0@0x10"),
8277 2 : Bytes::from_static(b"value 1@0x10@0x20"),
8278 2 : Bytes::from_static(b"value 2@0x10"),
8279 2 : Bytes::from_static(b"value 3@0x10"),
8280 2 : Bytes::from_static(b"value 4@0x10"),
8281 2 : Bytes::from_static(b"value 5@0x10"),
8282 2 : Bytes::from_static(b"value 6@0x10"),
8283 2 : Bytes::from_static(b"value 7@0x10"),
8284 2 : Bytes::from_static(b"value 8@0x10"),
8285 2 : Bytes::from_static(b"value 9@0x10"),
8286 2 : ];
8287 2 :
8288 2 : let expected_result_at_lsn_10 = [
8289 2 : Bytes::from_static(b"value 0@0x10"),
8290 2 : Bytes::from_static(b"value 1@0x10"),
8291 2 : Bytes::from_static(b"value 2@0x10"),
8292 2 : Bytes::from_static(b"value 3@0x10"),
8293 2 : Bytes::from_static(b"value 4@0x10"),
8294 2 : Bytes::from_static(b"value 5@0x10"),
8295 2 : Bytes::from_static(b"value 6@0x10"),
8296 2 : Bytes::from_static(b"value 7@0x10"),
8297 2 : Bytes::from_static(b"value 8@0x10"),
8298 2 : Bytes::from_static(b"value 9@0x10"),
8299 2 : ];
8300 2 :
8301 8 : let verify_result = || async {
8302 8 : let gc_horizon = {
8303 8 : let gc_info = tline.gc_info.read().unwrap();
8304 8 : gc_info.cutoffs.time
8305 2 : };
8306 88 : for idx in 0..10 {
8307 80 : assert_eq!(
8308 80 : tline
8309 80 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8310 75 : .await
8311 80 : .unwrap(),
8312 80 : &expected_result[idx]
8313 2 : );
8314 80 : assert_eq!(
8315 80 : tline
8316 80 : .get(get_key(idx as u32), gc_horizon, &ctx)
8317 47 : .await
8318 80 : .unwrap(),
8319 80 : &expected_result_at_gc_horizon[idx]
8320 2 : );
8321 80 : assert_eq!(
8322 80 : tline
8323 80 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
8324 46 : .await
8325 80 : .unwrap(),
8326 80 : &expected_result_at_lsn_20[idx]
8327 2 : );
8328 80 : assert_eq!(
8329 80 : tline
8330 80 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
8331 42 : .await
8332 80 : .unwrap(),
8333 80 : &expected_result_at_lsn_10[idx]
8334 2 : );
8335 2 : }
8336 16 : };
8337 2 :
8338 61 : verify_result().await;
8339 2 :
8340 2 : let cancel = CancellationToken::new();
8341 2 : let mut dryrun_flags = EnumSet::new();
8342 2 : dryrun_flags.insert(CompactFlags::DryRun);
8343 2 :
8344 2 : tline
8345 2 : .compact_with_gc(&cancel, dryrun_flags, &ctx)
8346 45 : .await
8347 2 : .unwrap();
8348 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
8349 2 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
8350 49 : verify_result().await;
8351 2 :
8352 2 : tline
8353 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8354 51 : .await
8355 2 : .unwrap();
8356 53 : verify_result().await;
8357 2 :
8358 2 : // compact again
8359 2 : tline
8360 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8361 39 : .await
8362 2 : .unwrap();
8363 47 : verify_result().await;
8364 2 :
8365 2 : Ok(())
8366 2 : }
8367 :
8368 : #[tokio::test]
8369 2 : async fn test_simple_bottom_most_compaction_on_branch() -> anyhow::Result<()> {
8370 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_on_branch").await?;
8371 8 : let (tenant, ctx) = harness.load().await;
8372 2 :
8373 126 : fn get_key(id: u32) -> Key {
8374 126 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
8375 126 : key.field6 = id;
8376 126 : key
8377 126 : }
8378 2 :
8379 2 : let img_layer = (0..10)
8380 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8381 2 : .collect_vec();
8382 2 :
8383 2 : let delta1 = vec![
8384 2 : (
8385 2 : get_key(1),
8386 2 : Lsn(0x20),
8387 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8388 2 : ),
8389 2 : (
8390 2 : get_key(2),
8391 2 : Lsn(0x30),
8392 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8393 2 : ),
8394 2 : (
8395 2 : get_key(3),
8396 2 : Lsn(0x28),
8397 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8398 2 : ),
8399 2 : (
8400 2 : get_key(3),
8401 2 : Lsn(0x30),
8402 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8403 2 : ),
8404 2 : (
8405 2 : get_key(3),
8406 2 : Lsn(0x40),
8407 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
8408 2 : ),
8409 2 : ];
8410 2 : let delta2 = vec![
8411 2 : (
8412 2 : get_key(5),
8413 2 : Lsn(0x20),
8414 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8415 2 : ),
8416 2 : (
8417 2 : get_key(6),
8418 2 : Lsn(0x20),
8419 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8420 2 : ),
8421 2 : ];
8422 2 : let delta3 = vec![
8423 2 : (
8424 2 : get_key(8),
8425 2 : Lsn(0x48),
8426 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8427 2 : ),
8428 2 : (
8429 2 : get_key(9),
8430 2 : Lsn(0x48),
8431 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8432 2 : ),
8433 2 : ];
8434 2 :
8435 2 : let parent_tline = tenant
8436 2 : .create_test_timeline_with_layers(
8437 2 : TIMELINE_ID,
8438 2 : Lsn(0x10),
8439 2 : DEFAULT_PG_VERSION,
8440 2 : &ctx,
8441 2 : vec![], // delta layers
8442 2 : vec![(Lsn(0x18), img_layer)], // image layers
8443 2 : Lsn(0x18),
8444 2 : )
8445 29 : .await?;
8446 2 :
8447 2 : parent_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
8448 2 :
8449 2 : let branch_tline = tenant
8450 2 : .branch_timeline_test_with_layers(
8451 2 : &parent_tline,
8452 2 : NEW_TIMELINE_ID,
8453 2 : Some(Lsn(0x18)),
8454 2 : &ctx,
8455 2 : vec![
8456 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
8457 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
8458 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
8459 2 : ], // delta layers
8460 2 : vec![], // image layers
8461 2 : Lsn(0x50),
8462 2 : )
8463 18 : .await?;
8464 2 :
8465 2 : branch_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
8466 2 :
8467 2 : {
8468 2 : // Update GC info
8469 2 : let mut guard = parent_tline.gc_info.write().unwrap();
8470 2 : *guard = GcInfo {
8471 2 : retain_lsns: vec![(Lsn(0x18), branch_tline.timeline_id, MaybeOffloaded::No)],
8472 2 : cutoffs: GcCutoffs {
8473 2 : time: Lsn(0x10),
8474 2 : space: Lsn(0x10),
8475 2 : },
8476 2 : leases: Default::default(),
8477 2 : within_ancestor_pitr: false,
8478 2 : };
8479 2 : }
8480 2 :
8481 2 : {
8482 2 : // Update GC info
8483 2 : let mut guard = branch_tline.gc_info.write().unwrap();
8484 2 : *guard = GcInfo {
8485 2 : retain_lsns: vec![(Lsn(0x40), branch_tline.timeline_id, MaybeOffloaded::No)],
8486 2 : cutoffs: GcCutoffs {
8487 2 : time: Lsn(0x50),
8488 2 : space: Lsn(0x50),
8489 2 : },
8490 2 : leases: Default::default(),
8491 2 : within_ancestor_pitr: false,
8492 2 : };
8493 2 : }
8494 2 :
8495 2 : let expected_result_at_gc_horizon = [
8496 2 : Bytes::from_static(b"value 0@0x10"),
8497 2 : Bytes::from_static(b"value 1@0x10@0x20"),
8498 2 : Bytes::from_static(b"value 2@0x10@0x30"),
8499 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
8500 2 : Bytes::from_static(b"value 4@0x10"),
8501 2 : Bytes::from_static(b"value 5@0x10@0x20"),
8502 2 : Bytes::from_static(b"value 6@0x10@0x20"),
8503 2 : Bytes::from_static(b"value 7@0x10"),
8504 2 : Bytes::from_static(b"value 8@0x10@0x48"),
8505 2 : Bytes::from_static(b"value 9@0x10@0x48"),
8506 2 : ];
8507 2 :
8508 2 : let expected_result_at_lsn_40 = [
8509 2 : Bytes::from_static(b"value 0@0x10"),
8510 2 : Bytes::from_static(b"value 1@0x10@0x20"),
8511 2 : Bytes::from_static(b"value 2@0x10@0x30"),
8512 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
8513 2 : Bytes::from_static(b"value 4@0x10"),
8514 2 : Bytes::from_static(b"value 5@0x10@0x20"),
8515 2 : Bytes::from_static(b"value 6@0x10@0x20"),
8516 2 : Bytes::from_static(b"value 7@0x10"),
8517 2 : Bytes::from_static(b"value 8@0x10"),
8518 2 : Bytes::from_static(b"value 9@0x10"),
8519 2 : ];
8520 2 :
8521 4 : let verify_result = || async {
8522 44 : for idx in 0..10 {
8523 40 : assert_eq!(
8524 40 : branch_tline
8525 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8526 52 : .await
8527 40 : .unwrap(),
8528 40 : &expected_result_at_gc_horizon[idx]
8529 2 : );
8530 40 : assert_eq!(
8531 40 : branch_tline
8532 40 : .get(get_key(idx as u32), Lsn(0x40), &ctx)
8533 31 : .await
8534 40 : .unwrap(),
8535 40 : &expected_result_at_lsn_40[idx]
8536 2 : );
8537 2 : }
8538 8 : };
8539 2 :
8540 46 : verify_result().await;
8541 2 :
8542 2 : let cancel = CancellationToken::new();
8543 2 : branch_tline
8544 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8545 19 : .await
8546 2 : .unwrap();
8547 2 :
8548 37 : verify_result().await;
8549 2 :
8550 2 : Ok(())
8551 2 : }
8552 :
8553 : // Regression test for https://github.com/neondatabase/neon/issues/9012
8554 : // Create an image arrangement where we have to read at different LSN ranges
8555 : // from a delta layer. This is achieved by overlapping an image layer on top of
8556 : // a delta layer. Like so:
8557 : //
8558 : // A B
8559 : // +----------------+ -> delta_layer
8560 : // | | ^ lsn
8561 : // | =========|-> nested_image_layer |
8562 : // | C | |
8563 : // +----------------+ |
8564 : // ======== -> baseline_image_layer +-------> key
8565 : //
8566 : //
8567 : // When querying the key range [A, B) we need to read at different LSN ranges
8568 : // for [A, C) and [C, B). This test checks that the described edge case is handled correctly.
8569 : #[tokio::test]
8570 2 : async fn test_vectored_read_with_nested_image_layer() -> anyhow::Result<()> {
8571 2 : let harness = TenantHarness::create("test_vectored_read_with_nested_image_layer").await?;
8572 8 : let (tenant, ctx) = harness.load().await;
8573 2 :
8574 2 : let will_init_keys = [2, 6];
8575 44 : fn get_key(id: u32) -> Key {
8576 44 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
8577 44 : key.field6 = id;
8578 44 : key
8579 44 : }
8580 2 :
8581 2 : let mut expected_key_values = HashMap::new();
8582 2 :
8583 2 : let baseline_image_layer_lsn = Lsn(0x10);
8584 2 : let mut baseline_img_layer = Vec::new();
8585 12 : for i in 0..5 {
8586 10 : let key = get_key(i);
8587 10 : let value = format!("value {i}@{baseline_image_layer_lsn}");
8588 10 :
8589 10 : let removed = expected_key_values.insert(key, value.clone());
8590 10 : assert!(removed.is_none());
8591 2 :
8592 10 : baseline_img_layer.push((key, Bytes::from(value)));
8593 2 : }
8594 2 :
8595 2 : let nested_image_layer_lsn = Lsn(0x50);
8596 2 : let mut nested_img_layer = Vec::new();
8597 12 : for i in 5..10 {
8598 10 : let key = get_key(i);
8599 10 : let value = format!("value {i}@{nested_image_layer_lsn}");
8600 10 :
8601 10 : let removed = expected_key_values.insert(key, value.clone());
8602 10 : assert!(removed.is_none());
8603 2 :
8604 10 : nested_img_layer.push((key, Bytes::from(value)));
8605 2 : }
8606 2 :
8607 2 : let mut delta_layer_spec = Vec::default();
8608 2 : let delta_layer_start_lsn = Lsn(0x20);
8609 2 : let mut delta_layer_end_lsn = delta_layer_start_lsn;
8610 2 :
8611 22 : for i in 0..10 {
8612 20 : let key = get_key(i);
8613 20 : let key_in_nested = nested_img_layer
8614 20 : .iter()
8615 80 : .any(|(key_with_img, _)| *key_with_img == key);
8616 20 : let lsn = {
8617 20 : if key_in_nested {
8618 10 : Lsn(nested_image_layer_lsn.0 + 0x10)
8619 2 : } else {
8620 10 : delta_layer_start_lsn
8621 2 : }
8622 2 : };
8623 2 :
8624 20 : let will_init = will_init_keys.contains(&i);
8625 20 : if will_init {
8626 4 : delta_layer_spec.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init())));
8627 4 :
8628 4 : expected_key_values.insert(key, "".to_string());
8629 16 : } else {
8630 16 : let delta = format!("@{lsn}");
8631 16 : delta_layer_spec.push((
8632 16 : key,
8633 16 : lsn,
8634 16 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
8635 16 : ));
8636 16 :
8637 16 : expected_key_values
8638 16 : .get_mut(&key)
8639 16 : .expect("An image exists for each key")
8640 16 : .push_str(delta.as_str());
8641 16 : }
8642 20 : delta_layer_end_lsn = std::cmp::max(delta_layer_start_lsn, lsn);
8643 2 : }
8644 2 :
8645 2 : delta_layer_end_lsn = Lsn(delta_layer_end_lsn.0 + 1);
8646 2 :
8647 2 : assert!(
8648 2 : nested_image_layer_lsn > delta_layer_start_lsn
8649 2 : && nested_image_layer_lsn < delta_layer_end_lsn
8650 2 : );
8651 2 :
8652 2 : let tline = tenant
8653 2 : .create_test_timeline_with_layers(
8654 2 : TIMELINE_ID,
8655 2 : baseline_image_layer_lsn,
8656 2 : DEFAULT_PG_VERSION,
8657 2 : &ctx,
8658 2 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
8659 2 : delta_layer_start_lsn..delta_layer_end_lsn,
8660 2 : delta_layer_spec,
8661 2 : )], // delta layers
8662 2 : vec![
8663 2 : (baseline_image_layer_lsn, baseline_img_layer),
8664 2 : (nested_image_layer_lsn, nested_img_layer),
8665 2 : ], // image layers
8666 2 : delta_layer_end_lsn,
8667 2 : )
8668 40 : .await?;
8669 2 :
8670 2 : let keyspace = KeySpace::single(get_key(0)..get_key(10));
8671 2 : let results = tline
8672 2 : .get_vectored(keyspace, delta_layer_end_lsn, &ctx)
8673 13 : .await
8674 2 : .expect("No vectored errors");
8675 22 : for (key, res) in results {
8676 20 : let value = res.expect("No key errors");
8677 20 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
8678 20 : assert_eq!(value, Bytes::from(expected_value));
8679 2 : }
8680 2 :
8681 2 : Ok(())
8682 2 : }
8683 : }
|
