Line data Source code
1 : //!
2 : //! Timeline repository implementation that keeps old data in files on disk, and
3 : //! the recent changes in memory. See tenant/*_layer.rs files.
4 : //! The functions here are responsible for locating the correct layer for the
5 : //! get/put call, walking back the timeline branching history as needed.
6 : //!
7 : //! The files are stored in the .neon/tenants/<tenant_id>/timelines/<timeline_id>
8 : //! directory. See docs/pageserver-storage.md for how the files are managed.
9 : //! In addition to the layer files, there is a metadata file in the same
10 : //! directory that contains information about the timeline, in particular its
11 : //! parent timeline, and the last LSN that has been written to disk.
12 : //!
13 :
14 : use anyhow::{bail, Context};
15 : use arc_swap::ArcSwap;
16 : use camino::Utf8Path;
17 : use camino::Utf8PathBuf;
18 : use enumset::EnumSet;
19 : use futures::stream::FuturesUnordered;
20 : use futures::FutureExt;
21 : use futures::StreamExt;
22 : use pageserver_api::models;
23 : use pageserver_api::models::AuxFilePolicy;
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::sync::Weak;
37 : use std::time::SystemTime;
38 : use storage_broker::BrokerClientChannel;
39 : use tokio::io::BufReader;
40 : use tokio::sync::watch;
41 : use tokio::task::JoinSet;
42 : use tokio_util::sync::CancellationToken;
43 : use tracing::*;
44 : use upload_queue::NotInitialized;
45 : use utils::backoff;
46 : use utils::circuit_breaker::CircuitBreaker;
47 : use utils::completion;
48 : use utils::crashsafe::path_with_suffix_extension;
49 : use utils::failpoint_support;
50 : use utils::fs_ext;
51 : use utils::pausable_failpoint;
52 : use utils::sync::gate::Gate;
53 : use utils::sync::gate::GateGuard;
54 : use utils::timeout::timeout_cancellable;
55 : use utils::timeout::TimeoutCancellableError;
56 : use utils::zstd::create_zst_tarball;
57 : use utils::zstd::extract_zst_tarball;
58 :
59 : use self::config::AttachedLocationConfig;
60 : use self::config::AttachmentMode;
61 : use self::config::LocationConf;
62 : use self::config::TenantConf;
63 : use self::metadata::TimelineMetadata;
64 : use self::mgr::GetActiveTenantError;
65 : use self::mgr::GetTenantError;
66 : use self::remote_timeline_client::upload::upload_index_part;
67 : use self::remote_timeline_client::RemoteTimelineClient;
68 : use self::timeline::uninit::TimelineCreateGuard;
69 : use self::timeline::uninit::TimelineExclusionError;
70 : use self::timeline::uninit::UninitializedTimeline;
71 : use self::timeline::EvictionTaskTenantState;
72 : use self::timeline::GcCutoffs;
73 : use self::timeline::TimelineResources;
74 : use self::timeline::WaitLsnError;
75 : use crate::config::PageServerConf;
76 : use crate::context::{DownloadBehavior, RequestContext};
77 : use crate::deletion_queue::DeletionQueueClient;
78 : use crate::deletion_queue::DeletionQueueError;
79 : use crate::import_datadir;
80 : use crate::is_uninit_mark;
81 : use crate::l0_flush::L0FlushGlobalState;
82 : use crate::metrics::TENANT;
83 : use crate::metrics::{
84 : remove_tenant_metrics, BROKEN_TENANTS_SET, CIRCUIT_BREAKERS_BROKEN, CIRCUIT_BREAKERS_UNBROKEN,
85 : TENANT_STATE_METRIC, TENANT_SYNTHETIC_SIZE_METRIC,
86 : };
87 : use crate::repository::GcResult;
88 : use crate::task_mgr;
89 : use crate::task_mgr::TaskKind;
90 : use crate::tenant::config::LocationMode;
91 : use crate::tenant::config::TenantConfOpt;
92 : pub use crate::tenant::remote_timeline_client::index::IndexPart;
93 : use crate::tenant::remote_timeline_client::remote_initdb_archive_path;
94 : use crate::tenant::remote_timeline_client::MaybeDeletedIndexPart;
95 : use crate::tenant::remote_timeline_client::INITDB_PATH;
96 : use crate::tenant::storage_layer::DeltaLayer;
97 : use crate::tenant::storage_layer::ImageLayer;
98 : use crate::walredo;
99 : use crate::InitializationOrder;
100 : use std::collections::hash_map::Entry;
101 : use std::collections::HashMap;
102 : use std::collections::HashSet;
103 : use std::fmt::Debug;
104 : use std::fmt::Display;
105 : use std::fs;
106 : use std::fs::File;
107 : use std::sync::atomic::{AtomicU64, Ordering};
108 : use std::sync::Arc;
109 : use std::sync::Mutex;
110 : use std::time::{Duration, Instant};
111 :
112 : use crate::span;
113 : use crate::tenant::timeline::delete::DeleteTimelineFlow;
114 : use crate::tenant::timeline::uninit::cleanup_timeline_directory;
115 : use crate::virtual_file::VirtualFile;
116 : use crate::walredo::PostgresRedoManager;
117 : use crate::TEMP_FILE_SUFFIX;
118 : use once_cell::sync::Lazy;
119 : pub use pageserver_api::models::TenantState;
120 : use tokio::sync::Semaphore;
121 :
122 0 : static INIT_DB_SEMAPHORE: Lazy<Semaphore> = Lazy::new(|| Semaphore::new(8));
123 : use utils::{
124 : crashsafe,
125 : generation::Generation,
126 : id::TimelineId,
127 : lsn::{Lsn, RecordLsn},
128 : };
129 :
130 : pub mod blob_io;
131 : pub mod block_io;
132 : pub mod vectored_blob_io;
133 :
134 : pub mod disk_btree;
135 : pub(crate) mod ephemeral_file;
136 : pub mod layer_map;
137 :
138 : pub mod metadata;
139 : pub mod remote_timeline_client;
140 : pub mod storage_layer;
141 :
142 : pub mod config;
143 : pub mod mgr;
144 : pub mod secondary;
145 : pub mod tasks;
146 : pub mod upload_queue;
147 :
148 : pub(crate) mod timeline;
149 :
150 : pub mod size;
151 :
152 : mod gc_block;
153 : pub(crate) mod throttle;
154 :
155 : pub(crate) use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
156 : pub(crate) use timeline::{LogicalSizeCalculationCause, PageReconstructError, Timeline};
157 :
158 : // re-export for use in walreceiver
159 : pub use crate::tenant::timeline::WalReceiverInfo;
160 :
161 : /// The "tenants" part of `tenants/<tenant>/timelines...`
162 : pub const TENANTS_SEGMENT_NAME: &str = "tenants";
163 :
164 : /// Parts of the `.neon/tenants/<tenant_id>/timelines/<timeline_id>` directory prefix.
165 : pub const TIMELINES_SEGMENT_NAME: &str = "timelines";
166 :
167 : /// References to shared objects that are passed into each tenant, such
168 : /// as the shared remote storage client and process initialization state.
169 : #[derive(Clone)]
170 : pub struct TenantSharedResources {
171 : pub broker_client: storage_broker::BrokerClientChannel,
172 : pub remote_storage: GenericRemoteStorage,
173 : pub deletion_queue_client: DeletionQueueClient,
174 : pub l0_flush_global_state: L0FlushGlobalState,
175 : }
176 :
177 : /// A [`Tenant`] is really an _attached_ tenant. The configuration
178 : /// for an attached tenant is a subset of the [`LocationConf`], represented
179 : /// in this struct.
180 : pub(super) struct AttachedTenantConf {
181 : tenant_conf: TenantConfOpt,
182 : location: AttachedLocationConfig,
183 : }
184 :
185 : impl AttachedTenantConf {
186 0 : fn new(tenant_conf: TenantConfOpt, location: AttachedLocationConfig) -> Self {
187 0 : Self {
188 0 : tenant_conf,
189 0 : location,
190 0 : }
191 0 : }
192 :
193 588 : fn try_from(location_conf: LocationConf) -> anyhow::Result<Self> {
194 588 : match &location_conf.mode {
195 588 : LocationMode::Attached(attach_conf) => Ok(Self {
196 588 : tenant_conf: location_conf.tenant_conf,
197 588 : location: *attach_conf,
198 588 : }),
199 : LocationMode::Secondary(_) => {
200 0 : anyhow::bail!("Attempted to construct AttachedTenantConf from a LocationConf in secondary mode")
201 : }
202 : }
203 588 : }
204 : }
205 : struct TimelinePreload {
206 : timeline_id: TimelineId,
207 : client: RemoteTimelineClient,
208 : index_part: Result<MaybeDeletedIndexPart, DownloadError>,
209 : }
210 :
211 : pub(crate) struct TenantPreload {
212 : timelines: HashMap<TimelineId, TimelinePreload>,
213 : }
214 :
215 : /// When we spawn a tenant, there is a special mode for tenant creation that
216 : /// avoids trying to read anything from remote storage.
217 : pub(crate) enum SpawnMode {
218 : /// Activate as soon as possible
219 : Eager,
220 : /// Lazy activation in the background, with the option to skip the queue if the need comes up
221 : Lazy,
222 : }
223 :
224 : ///
225 : /// Tenant consists of multiple timelines. Keep them in a hash table.
226 : ///
227 : pub struct Tenant {
228 : // Global pageserver config parameters
229 : pub conf: &'static PageServerConf,
230 :
231 : /// The value creation timestamp, used to measure activation delay, see:
232 : /// <https://github.com/neondatabase/neon/issues/4025>
233 : constructed_at: Instant,
234 :
235 : state: watch::Sender<TenantState>,
236 :
237 : // Overridden tenant-specific config parameters.
238 : // We keep TenantConfOpt sturct here to preserve the information
239 : // about parameters that are not set.
240 : // This is necessary to allow global config updates.
241 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
242 :
243 : tenant_shard_id: TenantShardId,
244 :
245 : // The detailed sharding information, beyond the number/count in tenant_shard_id
246 : shard_identity: ShardIdentity,
247 :
248 : /// The remote storage generation, used to protect S3 objects from split-brain.
249 : /// Does not change over the lifetime of the [`Tenant`] object.
250 : ///
251 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
252 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
253 : generation: Generation,
254 :
255 : timelines: Mutex<HashMap<TimelineId, Arc<Timeline>>>,
256 :
257 : /// During timeline creation, we first insert the TimelineId to the
258 : /// creating map, then `timelines`, then remove it from the creating map.
259 : /// **Lock order**: if acquring both, acquire`timelines` before `timelines_creating`
260 : timelines_creating: std::sync::Mutex<HashSet<TimelineId>>,
261 :
262 : // This mutex prevents creation of new timelines during GC.
263 : // Adding yet another mutex (in addition to `timelines`) is needed because holding
264 : // `timelines` mutex during all GC iteration
265 : // may block for a long time `get_timeline`, `get_timelines_state`,... and other operations
266 : // with timelines, which in turn may cause dropping replication connection, expiration of wait_for_lsn
267 : // timeout...
268 : gc_cs: tokio::sync::Mutex<()>,
269 : walredo_mgr: Option<Arc<WalRedoManager>>,
270 :
271 : // provides access to timeline data sitting in the remote storage
272 : pub(crate) remote_storage: GenericRemoteStorage,
273 :
274 : // Access to global deletion queue for when this tenant wants to schedule a deletion
275 : deletion_queue_client: DeletionQueueClient,
276 :
277 : /// Cached logical sizes updated updated on each [`Tenant::gather_size_inputs`].
278 : cached_logical_sizes: tokio::sync::Mutex<HashMap<(TimelineId, Lsn), u64>>,
279 : cached_synthetic_tenant_size: Arc<AtomicU64>,
280 :
281 : eviction_task_tenant_state: tokio::sync::Mutex<EvictionTaskTenantState>,
282 :
283 : /// Track repeated failures to compact, so that we can back off.
284 : /// Overhead of mutex is acceptable because compaction is done with a multi-second period.
285 : compaction_circuit_breaker: std::sync::Mutex<CircuitBreaker>,
286 :
287 : /// If the tenant is in Activating state, notify this to encourage it
288 : /// to proceed to Active as soon as possible, rather than waiting for lazy
289 : /// background warmup.
290 : pub(crate) activate_now_sem: tokio::sync::Semaphore,
291 :
292 : // Cancellation token fires when we have entered shutdown(). This is a parent of
293 : // Timelines' cancellation token.
294 : pub(crate) cancel: CancellationToken,
295 :
296 : // Users of the Tenant such as the page service must take this Gate to avoid
297 : // trying to use a Tenant which is shutting down.
298 : pub(crate) gate: Gate,
299 :
300 : /// Throttle applied at the top of [`Timeline::get`].
301 : /// All [`Tenant::timelines`] of a given [`Tenant`] instance share the same [`throttle::Throttle`] instance.
302 : pub(crate) timeline_get_throttle:
303 : Arc<throttle::Throttle<&'static crate::metrics::tenant_throttling::TimelineGet>>,
304 :
305 : /// An ongoing timeline detach concurrency limiter.
306 : ///
307 : /// As a tenant will likely be restarted as part of timeline detach ancestor it makes no sense
308 : /// to have two running at the same time. A different one can be started if an earlier one
309 : /// has failed for whatever reason.
310 : ongoing_timeline_detach: std::sync::Mutex<Option<(TimelineId, utils::completion::Barrier)>>,
311 :
312 : /// `index_part.json` based gc blocking reason tracking.
313 : ///
314 : /// New gc iterations must start a new iteration by acquiring `GcBlock::start` before
315 : /// proceeding.
316 : pub(crate) gc_block: gc_block::GcBlock,
317 :
318 : l0_flush_global_state: L0FlushGlobalState,
319 : }
320 :
321 : impl std::fmt::Debug for Tenant {
322 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
323 0 : write!(f, "{} ({})", self.tenant_shard_id, self.current_state())
324 0 : }
325 : }
326 :
327 : pub(crate) enum WalRedoManager {
328 : Prod(WalredoManagerId, PostgresRedoManager),
329 : #[cfg(test)]
330 : Test(harness::TestRedoManager),
331 : }
332 :
333 0 : #[derive(thiserror::Error, Debug)]
334 : #[error("pageserver is shutting down")]
335 : pub(crate) struct GlobalShutDown;
336 :
337 : impl WalRedoManager {
338 0 : pub(crate) fn new(mgr: PostgresRedoManager) -> Result<Arc<Self>, GlobalShutDown> {
339 0 : let id = WalredoManagerId::next();
340 0 : let arc = Arc::new(Self::Prod(id, mgr));
341 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
342 0 : match &mut *guard {
343 0 : Some(map) => {
344 0 : map.insert(id, Arc::downgrade(&arc));
345 0 : Ok(arc)
346 : }
347 0 : None => Err(GlobalShutDown),
348 : }
349 0 : }
350 : }
351 :
352 : impl Drop for WalRedoManager {
353 30 : fn drop(&mut self) {
354 30 : match self {
355 0 : Self::Prod(id, _) => {
356 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
357 0 : if let Some(map) = &mut *guard {
358 0 : map.remove(id).expect("new() registers, drop() unregisters");
359 0 : }
360 : }
361 : #[cfg(test)]
362 30 : Self::Test(_) => {
363 30 : // Not applicable to test redo manager
364 30 : }
365 : }
366 30 : }
367 : }
368 :
369 : /// Global registry of all walredo managers so that [`crate::shutdown_pageserver`] can shut down
370 : /// the walredo processes outside of the regular order.
371 : ///
372 : /// This is necessary to work around a systemd bug where it freezes if there are
373 : /// walredo processes left => <https://github.com/neondatabase/cloud/issues/11387>
374 : #[allow(clippy::type_complexity)]
375 : pub(crate) static WALREDO_MANAGERS: once_cell::sync::Lazy<
376 : Mutex<Option<HashMap<WalredoManagerId, Weak<WalRedoManager>>>>,
377 0 : > = once_cell::sync::Lazy::new(|| Mutex::new(Some(HashMap::new())));
378 : #[derive(PartialEq, Eq, Hash, Clone, Copy, Debug)]
379 : pub(crate) struct WalredoManagerId(u64);
380 : impl WalredoManagerId {
381 0 : pub fn next() -> Self {
382 0 : static NEXT: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(1);
383 0 : let id = NEXT.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
384 0 : if id == 0 {
385 0 : panic!("WalredoManagerId::new() returned 0, indicating wraparound, risking it's no longer unique");
386 0 : }
387 0 : Self(id)
388 0 : }
389 : }
390 :
391 : #[cfg(test)]
392 : impl From<harness::TestRedoManager> for WalRedoManager {
393 564 : fn from(mgr: harness::TestRedoManager) -> Self {
394 564 : Self::Test(mgr)
395 564 : }
396 : }
397 :
398 : impl WalRedoManager {
399 18 : pub(crate) async fn shutdown(&self) -> bool {
400 18 : match self {
401 0 : Self::Prod(_, mgr) => mgr.shutdown().await,
402 : #[cfg(test)]
403 : Self::Test(_) => {
404 : // Not applicable to test redo manager
405 18 : true
406 : }
407 : }
408 18 : }
409 :
410 0 : pub(crate) fn maybe_quiesce(&self, idle_timeout: Duration) {
411 0 : match self {
412 0 : Self::Prod(_, mgr) => mgr.maybe_quiesce(idle_timeout),
413 0 : #[cfg(test)]
414 0 : Self::Test(_) => {
415 0 : // Not applicable to test redo manager
416 0 : }
417 0 : }
418 0 : }
419 :
420 : /// # Cancel-Safety
421 : ///
422 : /// This method is cancellation-safe.
423 978 : pub async fn request_redo(
424 978 : &self,
425 978 : key: crate::repository::Key,
426 978 : lsn: Lsn,
427 978 : base_img: Option<(Lsn, bytes::Bytes)>,
428 978 : records: Vec<(Lsn, crate::walrecord::NeonWalRecord)>,
429 978 : pg_version: u32,
430 978 : ) -> Result<bytes::Bytes, walredo::Error> {
431 978 : match self {
432 0 : Self::Prod(_, mgr) => {
433 0 : mgr.request_redo(key, lsn, base_img, records, pg_version)
434 0 : .await
435 : }
436 : #[cfg(test)]
437 978 : Self::Test(mgr) => {
438 978 : mgr.request_redo(key, lsn, base_img, records, pg_version)
439 0 : .await
440 : }
441 : }
442 978 : }
443 :
444 0 : pub(crate) fn status(&self) -> Option<WalRedoManagerStatus> {
445 0 : match self {
446 0 : WalRedoManager::Prod(_, m) => Some(m.status()),
447 0 : #[cfg(test)]
448 0 : WalRedoManager::Test(_) => None,
449 0 : }
450 0 : }
451 : }
452 :
453 0 : #[derive(Debug, thiserror::Error, PartialEq, Eq)]
454 : pub enum GetTimelineError {
455 : #[error("Timeline is shutting down")]
456 : ShuttingDown,
457 : #[error("Timeline {tenant_id}/{timeline_id} is not active, state: {state:?}")]
458 : NotActive {
459 : tenant_id: TenantShardId,
460 : timeline_id: TimelineId,
461 : state: TimelineState,
462 : },
463 : #[error("Timeline {tenant_id}/{timeline_id} was not found")]
464 : NotFound {
465 : tenant_id: TenantShardId,
466 : timeline_id: TimelineId,
467 : },
468 : }
469 :
470 0 : #[derive(Debug, thiserror::Error)]
471 : pub enum LoadLocalTimelineError {
472 : #[error("FailedToLoad")]
473 : Load(#[source] anyhow::Error),
474 : #[error("FailedToResumeDeletion")]
475 : ResumeDeletion(#[source] anyhow::Error),
476 : }
477 :
478 0 : #[derive(thiserror::Error)]
479 : pub enum DeleteTimelineError {
480 : #[error("NotFound")]
481 : NotFound,
482 :
483 : #[error("HasChildren")]
484 : HasChildren(Vec<TimelineId>),
485 :
486 : #[error("Timeline deletion is already in progress")]
487 : AlreadyInProgress(Arc<tokio::sync::Mutex<DeleteTimelineFlow>>),
488 :
489 : #[error(transparent)]
490 : Other(#[from] anyhow::Error),
491 : }
492 :
493 : impl Debug for DeleteTimelineError {
494 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
495 0 : match self {
496 0 : Self::NotFound => write!(f, "NotFound"),
497 0 : Self::HasChildren(c) => f.debug_tuple("HasChildren").field(c).finish(),
498 0 : Self::AlreadyInProgress(_) => f.debug_tuple("AlreadyInProgress").finish(),
499 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
500 : }
501 0 : }
502 : }
503 :
504 0 : #[derive(thiserror::Error)]
505 : pub enum TimelineArchivalError {
506 : #[error("NotFound")]
507 : NotFound,
508 :
509 : #[error("Timeout")]
510 : Timeout,
511 :
512 : #[error("HasUnarchivedChildren")]
513 : HasUnarchivedChildren(Vec<TimelineId>),
514 :
515 : #[error("Timeline archival is already in progress")]
516 : AlreadyInProgress,
517 :
518 : #[error(transparent)]
519 : Other(#[from] anyhow::Error),
520 : }
521 :
522 : impl Debug for TimelineArchivalError {
523 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
524 0 : match self {
525 0 : Self::NotFound => write!(f, "NotFound"),
526 0 : Self::Timeout => write!(f, "Timeout"),
527 0 : Self::HasUnarchivedChildren(c) => {
528 0 : f.debug_tuple("HasUnarchivedChildren").field(c).finish()
529 : }
530 0 : Self::AlreadyInProgress => f.debug_tuple("AlreadyInProgress").finish(),
531 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
532 : }
533 0 : }
534 : }
535 :
536 : pub enum SetStoppingError {
537 : AlreadyStopping(completion::Barrier),
538 : Broken,
539 : }
540 :
541 : impl Debug for SetStoppingError {
542 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
543 0 : match self {
544 0 : Self::AlreadyStopping(_) => f.debug_tuple("AlreadyStopping").finish(),
545 0 : Self::Broken => write!(f, "Broken"),
546 : }
547 0 : }
548 : }
549 :
550 0 : #[derive(thiserror::Error, Debug)]
551 : pub enum CreateTimelineError {
552 : #[error("creation of timeline with the given ID is in progress")]
553 : AlreadyCreating,
554 : #[error("timeline already exists with different parameters")]
555 : Conflict,
556 : #[error(transparent)]
557 : AncestorLsn(anyhow::Error),
558 : #[error("ancestor timeline is not active")]
559 : AncestorNotActive,
560 : #[error("tenant shutting down")]
561 : ShuttingDown,
562 : #[error(transparent)]
563 : Other(#[from] anyhow::Error),
564 : }
565 :
566 : #[derive(thiserror::Error, Debug)]
567 : enum InitdbError {
568 : Other(anyhow::Error),
569 : Cancelled,
570 : Spawn(std::io::Result<()>),
571 : Failed(std::process::ExitStatus, Vec<u8>),
572 : }
573 :
574 : impl fmt::Display for InitdbError {
575 0 : fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
576 0 : match self {
577 0 : InitdbError::Cancelled => write!(f, "Operation was cancelled"),
578 0 : InitdbError::Spawn(e) => write!(f, "Spawn error: {:?}", e),
579 0 : InitdbError::Failed(status, stderr) => write!(
580 0 : f,
581 0 : "Command failed with status {:?}: {}",
582 0 : status,
583 0 : String::from_utf8_lossy(stderr)
584 0 : ),
585 0 : InitdbError::Other(e) => write!(f, "Error: {:?}", e),
586 : }
587 0 : }
588 : }
589 :
590 : impl From<std::io::Error> for InitdbError {
591 0 : fn from(error: std::io::Error) -> Self {
592 0 : InitdbError::Spawn(Err(error))
593 0 : }
594 : }
595 :
596 : enum CreateTimelineCause {
597 : Load,
598 : Delete,
599 : }
600 :
601 0 : #[derive(thiserror::Error, Debug)]
602 : pub(crate) enum GcError {
603 : // The tenant is shutting down
604 : #[error("tenant shutting down")]
605 : TenantCancelled,
606 :
607 : // The tenant is shutting down
608 : #[error("timeline shutting down")]
609 : TimelineCancelled,
610 :
611 : // The tenant is in a state inelegible to run GC
612 : #[error("not active")]
613 : NotActive,
614 :
615 : // A requested GC cutoff LSN was invalid, for example it tried to move backwards
616 : #[error("not active")]
617 : BadLsn { why: String },
618 :
619 : // A remote storage error while scheduling updates after compaction
620 : #[error(transparent)]
621 : Remote(anyhow::Error),
622 :
623 : // An error reading while calculating GC cutoffs
624 : #[error(transparent)]
625 : GcCutoffs(PageReconstructError),
626 :
627 : // If GC was invoked for a particular timeline, this error means it didn't exist
628 : #[error("timeline not found")]
629 : TimelineNotFound,
630 : }
631 :
632 : impl From<PageReconstructError> for GcError {
633 0 : fn from(value: PageReconstructError) -> Self {
634 0 : match value {
635 0 : PageReconstructError::Cancelled => Self::TimelineCancelled,
636 0 : other => Self::GcCutoffs(other),
637 : }
638 0 : }
639 : }
640 :
641 : impl From<NotInitialized> for GcError {
642 0 : fn from(value: NotInitialized) -> Self {
643 0 : match value {
644 0 : NotInitialized::Uninitialized => GcError::Remote(value.into()),
645 0 : NotInitialized::Stopped | NotInitialized::ShuttingDown => GcError::TimelineCancelled,
646 : }
647 0 : }
648 : }
649 :
650 : impl From<timeline::layer_manager::Shutdown> for GcError {
651 0 : fn from(_: timeline::layer_manager::Shutdown) -> Self {
652 0 : GcError::TimelineCancelled
653 0 : }
654 : }
655 :
656 0 : #[derive(thiserror::Error, Debug)]
657 : pub(crate) enum LoadConfigError {
658 : #[error("TOML deserialization error: '{0}'")]
659 : DeserializeToml(#[from] toml_edit::de::Error),
660 :
661 : #[error("Config not found at {0}")]
662 : NotFound(Utf8PathBuf),
663 : }
664 :
665 : impl Tenant {
666 : /// Yet another helper for timeline initialization.
667 : ///
668 : /// - Initializes the Timeline struct and inserts it into the tenant's hash map
669 : /// - Scans the local timeline directory for layer files and builds the layer map
670 : /// - Downloads remote index file and adds remote files to the layer map
671 : /// - Schedules remote upload tasks for any files that are present locally but missing from remote storage.
672 : ///
673 : /// If the operation fails, the timeline is left in the tenant's hash map in Broken state. On success,
674 : /// it is marked as Active.
675 : #[allow(clippy::too_many_arguments)]
676 18 : async fn timeline_init_and_sync(
677 18 : &self,
678 18 : timeline_id: TimelineId,
679 18 : resources: TimelineResources,
680 18 : index_part: Option<IndexPart>,
681 18 : metadata: TimelineMetadata,
682 18 : ancestor: Option<Arc<Timeline>>,
683 18 : last_aux_file_policy: Option<AuxFilePolicy>,
684 18 : _ctx: &RequestContext,
685 18 : ) -> anyhow::Result<()> {
686 18 : let tenant_id = self.tenant_shard_id;
687 :
688 18 : let timeline = self.create_timeline_struct(
689 18 : timeline_id,
690 18 : &metadata,
691 18 : ancestor.clone(),
692 18 : resources,
693 18 : CreateTimelineCause::Load,
694 18 : // This could be derived from ancestor branch + index part. Though the only caller of `timeline_init_and_sync` is `load_remote_timeline`,
695 18 : // there will potentially be other caller of this function in the future, and we don't know whether `index_part` or `ancestor` takes precedence.
696 18 : // Therefore, we pass this field explicitly for now, and remove it once we fully migrate to aux file v2.
697 18 : last_aux_file_policy,
698 18 : )?;
699 18 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
700 18 : anyhow::ensure!(
701 18 : disk_consistent_lsn.is_valid(),
702 0 : "Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn"
703 : );
704 18 : assert_eq!(
705 18 : disk_consistent_lsn,
706 18 : metadata.disk_consistent_lsn(),
707 0 : "these are used interchangeably"
708 : );
709 :
710 18 : if let Some(index_part) = index_part.as_ref() {
711 18 : timeline.remote_client.init_upload_queue(index_part)?;
712 :
713 18 : timeline
714 18 : .last_aux_file_policy
715 18 : .store(index_part.last_aux_file_policy());
716 : } else {
717 : // No data on the remote storage, but we have local metadata file. We can end up
718 : // here with timeline_create being interrupted before finishing index part upload.
719 : // By doing what we do here, the index part upload is retried.
720 : // If control plane retries timeline creation in the meantime, the mgmt API handler
721 : // for timeline creation will coalesce on the upload we queue here.
722 :
723 : // FIXME: this branch should be dead code as we no longer write local metadata.
724 :
725 0 : timeline
726 0 : .remote_client
727 0 : .init_upload_queue_for_empty_remote(&metadata)?;
728 0 : timeline
729 0 : .remote_client
730 0 : .schedule_index_upload_for_full_metadata_update(&metadata)?;
731 : }
732 :
733 18 : timeline
734 18 : .load_layer_map(disk_consistent_lsn, index_part)
735 15 : .await
736 18 : .with_context(|| {
737 0 : format!("Failed to load layermap for timeline {tenant_id}/{timeline_id}")
738 18 : })?;
739 :
740 : {
741 : // avoiding holding it across awaits
742 18 : let mut timelines_accessor = self.timelines.lock().unwrap();
743 18 : match timelines_accessor.entry(timeline_id) {
744 : // We should never try and load the same timeline twice during startup
745 : Entry::Occupied(_) => {
746 0 : unreachable!(
747 0 : "Timeline {tenant_id}/{timeline_id} already exists in the tenant map"
748 0 : );
749 : }
750 18 : Entry::Vacant(v) => {
751 18 : v.insert(Arc::clone(&timeline));
752 18 : timeline.maybe_spawn_flush_loop();
753 18 : }
754 18 : }
755 18 : };
756 18 :
757 18 : // Sanity check: a timeline should have some content.
758 18 : anyhow::ensure!(
759 18 : ancestor.is_some()
760 12 : || timeline
761 12 : .layers
762 12 : .read()
763 0 : .await
764 12 : .layer_map()
765 12 : .expect("currently loading, layer manager cannot be shutdown already")
766 12 : .iter_historic_layers()
767 12 : .next()
768 12 : .is_some(),
769 0 : "Timeline has no ancestor and no layer files"
770 : );
771 :
772 18 : Ok(())
773 18 : }
774 :
775 : /// Attach a tenant that's available in cloud storage.
776 : ///
777 : /// This returns quickly, after just creating the in-memory object
778 : /// Tenant struct and launching a background task to download
779 : /// the remote index files. On return, the tenant is most likely still in
780 : /// Attaching state, and it will become Active once the background task
781 : /// finishes. You can use wait_until_active() to wait for the task to
782 : /// complete.
783 : ///
784 : #[allow(clippy::too_many_arguments)]
785 0 : pub(crate) fn spawn(
786 0 : conf: &'static PageServerConf,
787 0 : tenant_shard_id: TenantShardId,
788 0 : resources: TenantSharedResources,
789 0 : attached_conf: AttachedTenantConf,
790 0 : shard_identity: ShardIdentity,
791 0 : init_order: Option<InitializationOrder>,
792 0 : mode: SpawnMode,
793 0 : ctx: &RequestContext,
794 0 : ) -> Result<Arc<Tenant>, GlobalShutDown> {
795 0 : let wal_redo_manager =
796 0 : WalRedoManager::new(PostgresRedoManager::new(conf, tenant_shard_id))?;
797 :
798 : let TenantSharedResources {
799 0 : broker_client,
800 0 : remote_storage,
801 0 : deletion_queue_client,
802 0 : l0_flush_global_state,
803 0 : } = resources;
804 0 :
805 0 : let attach_mode = attached_conf.location.attach_mode;
806 0 : let generation = attached_conf.location.generation;
807 0 :
808 0 : let tenant = Arc::new(Tenant::new(
809 0 : TenantState::Attaching,
810 0 : conf,
811 0 : attached_conf,
812 0 : shard_identity,
813 0 : Some(wal_redo_manager),
814 0 : tenant_shard_id,
815 0 : remote_storage.clone(),
816 0 : deletion_queue_client,
817 0 : l0_flush_global_state,
818 0 : ));
819 0 :
820 0 : // The attach task will carry a GateGuard, so that shutdown() reliably waits for it to drop out if
821 0 : // we shut down while attaching.
822 0 : let attach_gate_guard = tenant
823 0 : .gate
824 0 : .enter()
825 0 : .expect("We just created the Tenant: nothing else can have shut it down yet");
826 0 :
827 0 : // Do all the hard work in the background
828 0 : let tenant_clone = Arc::clone(&tenant);
829 0 : let ctx = ctx.detached_child(TaskKind::Attach, DownloadBehavior::Warn);
830 0 : task_mgr::spawn(
831 0 : &tokio::runtime::Handle::current(),
832 0 : TaskKind::Attach,
833 0 : tenant_shard_id,
834 0 : None,
835 0 : "attach tenant",
836 0 : async move {
837 0 :
838 0 : info!(
839 : ?attach_mode,
840 0 : "Attaching tenant"
841 : );
842 :
843 0 : let _gate_guard = attach_gate_guard;
844 0 :
845 0 : // Is this tenant being spawned as part of process startup?
846 0 : let starting_up = init_order.is_some();
847 : scopeguard::defer! {
848 : if starting_up {
849 : TENANT.startup_complete.inc();
850 : }
851 : }
852 :
853 : // Ideally we should use Tenant::set_broken_no_wait, but it is not supposed to be used when tenant is in loading state.
854 : enum BrokenVerbosity {
855 : Error,
856 : Info
857 : }
858 0 : let make_broken =
859 0 : |t: &Tenant, err: anyhow::Error, verbosity: BrokenVerbosity| {
860 0 : match verbosity {
861 : BrokenVerbosity::Info => {
862 0 : info!("attach cancelled, setting tenant state to Broken: {err}");
863 : },
864 : BrokenVerbosity::Error => {
865 0 : error!("attach failed, setting tenant state to Broken: {err:?}");
866 : }
867 : }
868 0 : t.state.send_modify(|state| {
869 0 : // The Stopping case is for when we have passed control on to DeleteTenantFlow:
870 0 : // if it errors, we will call make_broken when tenant is already in Stopping.
871 0 : assert!(
872 0 : matches!(*state, TenantState::Attaching | TenantState::Stopping { .. }),
873 0 : "the attach task owns the tenant state until activation is complete"
874 : );
875 :
876 0 : *state = TenantState::broken_from_reason(err.to_string());
877 0 : });
878 0 : };
879 :
880 : // TODO: should also be rejecting tenant conf changes that violate this check.
881 0 : if let Err(e) = crate::tenant::storage_layer::inmemory_layer::IndexEntry::validate_checkpoint_distance(tenant_clone.get_checkpoint_distance()) {
882 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
883 0 : return Ok(());
884 0 : }
885 0 :
886 0 : let mut init_order = init_order;
887 0 : // take the completion because initial tenant loading will complete when all of
888 0 : // these tasks complete.
889 0 : let _completion = init_order
890 0 : .as_mut()
891 0 : .and_then(|x| x.initial_tenant_load.take());
892 0 : let remote_load_completion = init_order
893 0 : .as_mut()
894 0 : .and_then(|x| x.initial_tenant_load_remote.take());
895 :
896 : enum AttachType<'a> {
897 : /// We are attaching this tenant lazily in the background.
898 : Warmup {
899 : _permit: tokio::sync::SemaphorePermit<'a>,
900 : during_startup: bool
901 : },
902 : /// We are attaching this tenant as soon as we can, because for example an
903 : /// endpoint tried to access it.
904 : OnDemand,
905 : /// During normal operations after startup, we are attaching a tenant, and
906 : /// eager attach was requested.
907 : Normal,
908 : }
909 :
910 0 : let attach_type = if matches!(mode, SpawnMode::Lazy) {
911 : // Before doing any I/O, wait for at least one of:
912 : // - A client attempting to access to this tenant (on-demand loading)
913 : // - A permit becoming available in the warmup semaphore (background warmup)
914 :
915 : tokio::select!(
916 : permit = tenant_clone.activate_now_sem.acquire() => {
917 : let _ = permit.expect("activate_now_sem is never closed");
918 : tracing::info!("Activating tenant (on-demand)");
919 : AttachType::OnDemand
920 : },
921 : permit = conf.concurrent_tenant_warmup.inner().acquire() => {
922 : let _permit = permit.expect("concurrent_tenant_warmup semaphore is never closed");
923 : tracing::info!("Activating tenant (warmup)");
924 : AttachType::Warmup {
925 : _permit,
926 : during_startup: init_order.is_some()
927 : }
928 : }
929 : _ = tenant_clone.cancel.cancelled() => {
930 : // This is safe, but should be pretty rare: it is interesting if a tenant
931 : // stayed in Activating for such a long time that shutdown found it in
932 : // that state.
933 : tracing::info!(state=%tenant_clone.current_state(), "Tenant shut down before activation");
934 : // Make the tenant broken so that set_stopping will not hang waiting for it to leave
935 : // the Attaching state. This is an over-reaction (nothing really broke, the tenant is
936 : // just shutting down), but ensures progress.
937 : make_broken(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"), BrokenVerbosity::Info);
938 : return Ok(());
939 : },
940 : )
941 : } else {
942 : // SpawnMode::{Create,Eager} always cause jumping ahead of the
943 : // concurrent_tenant_warmup queue
944 0 : AttachType::Normal
945 : };
946 :
947 0 : let preload = match &mode {
948 : SpawnMode::Eager | SpawnMode::Lazy => {
949 0 : let _preload_timer = TENANT.preload.start_timer();
950 0 : let res = tenant_clone
951 0 : .preload(&remote_storage, task_mgr::shutdown_token())
952 0 : .await;
953 0 : match res {
954 0 : Ok(p) => Some(p),
955 0 : Err(e) => {
956 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
957 0 : return Ok(());
958 : }
959 : }
960 : }
961 :
962 : };
963 :
964 : // Remote preload is complete.
965 0 : drop(remote_load_completion);
966 :
967 : // We will time the duration of the attach phase unless this is a creation (attach will do no work)
968 0 : let attached = {
969 0 : let _attach_timer = Some(TENANT.attach.start_timer());
970 0 : tenant_clone.attach(preload, &ctx).await
971 : };
972 :
973 0 : match attached {
974 : Ok(()) => {
975 0 : info!("attach finished, activating");
976 0 : tenant_clone.activate(broker_client, None, &ctx);
977 : }
978 0 : Err(e) => {
979 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
980 0 : }
981 : }
982 :
983 : // If we are doing an opportunistic warmup attachment at startup, initialize
984 : // logical size at the same time. This is better than starting a bunch of idle tenants
985 : // with cold caches and then coming back later to initialize their logical sizes.
986 : //
987 : // It also prevents the warmup proccess competing with the concurrency limit on
988 : // logical size calculations: if logical size calculation semaphore is saturated,
989 : // then warmup will wait for that before proceeding to the next tenant.
990 0 : if matches!(attach_type, AttachType::Warmup { during_startup: true, .. }) {
991 0 : let mut futs: FuturesUnordered<_> = tenant_clone.timelines.lock().unwrap().values().cloned().map(|t| t.await_initial_logical_size()).collect();
992 0 : tracing::info!("Waiting for initial logical sizes while warming up...");
993 0 : while futs.next().await.is_some() {}
994 0 : tracing::info!("Warm-up complete");
995 0 : }
996 :
997 0 : Ok(())
998 0 : }
999 0 : .instrument(tracing::info_span!(parent: None, "attach", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), gen=?generation)),
1000 : );
1001 0 : Ok(tenant)
1002 0 : }
1003 :
1004 1128 : #[instrument(skip_all)]
1005 : pub(crate) async fn preload(
1006 : self: &Arc<Self>,
1007 : remote_storage: &GenericRemoteStorage,
1008 : cancel: CancellationToken,
1009 : ) -> anyhow::Result<TenantPreload> {
1010 : span::debug_assert_current_span_has_tenant_id();
1011 : // Get list of remote timelines
1012 : // download index files for every tenant timeline
1013 : info!("listing remote timelines");
1014 : let (remote_timeline_ids, other_keys) = remote_timeline_client::list_remote_timelines(
1015 : remote_storage,
1016 : self.tenant_shard_id,
1017 : cancel.clone(),
1018 : )
1019 : .await?;
1020 :
1021 : info!("found {} timelines", remote_timeline_ids.len(),);
1022 :
1023 : for k in other_keys {
1024 : warn!("Unexpected non timeline key {k}");
1025 : }
1026 :
1027 : Ok(TenantPreload {
1028 : timelines: Self::load_timeline_metadata(
1029 : self,
1030 : remote_timeline_ids,
1031 : remote_storage,
1032 : cancel,
1033 : )
1034 : .await?,
1035 : })
1036 : }
1037 :
1038 : ///
1039 : /// Background task that downloads all data for a tenant and brings it to Active state.
1040 : ///
1041 : /// No background tasks are started as part of this routine.
1042 : ///
1043 564 : async fn attach(
1044 564 : self: &Arc<Tenant>,
1045 564 : preload: Option<TenantPreload>,
1046 564 : ctx: &RequestContext,
1047 564 : ) -> anyhow::Result<()> {
1048 564 : span::debug_assert_current_span_has_tenant_id();
1049 564 :
1050 564 : failpoint_support::sleep_millis_async!("before-attaching-tenant");
1051 :
1052 564 : let Some(preload) = preload else {
1053 0 : anyhow::bail!("local-only deployment is no longer supported, https://github.com/neondatabase/neon/issues/5624");
1054 : };
1055 :
1056 564 : let mut timelines_to_resume_deletions = vec![];
1057 564 :
1058 564 : let mut remote_index_and_client = HashMap::new();
1059 564 : let mut timeline_ancestors = HashMap::new();
1060 564 : let mut existent_timelines = HashSet::new();
1061 582 : for (timeline_id, preload) in preload.timelines {
1062 18 : let index_part = match preload.index_part {
1063 18 : Ok(i) => {
1064 18 : debug!("remote index part exists for timeline {timeline_id}");
1065 : // We found index_part on the remote, this is the standard case.
1066 18 : existent_timelines.insert(timeline_id);
1067 18 : i
1068 : }
1069 : Err(DownloadError::NotFound) => {
1070 : // There is no index_part on the remote. We only get here
1071 : // if there is some prefix for the timeline in the remote storage.
1072 : // This can e.g. be the initdb.tar.zst archive, maybe a
1073 : // remnant from a prior incomplete creation or deletion attempt.
1074 : // Delete the local directory as the deciding criterion for a
1075 : // timeline's existence is presence of index_part.
1076 0 : info!(%timeline_id, "index_part not found on remote");
1077 0 : continue;
1078 : }
1079 0 : Err(e) => {
1080 0 : // Some (possibly ephemeral) error happened during index_part download.
1081 0 : // Pretend the timeline exists to not delete the timeline directory,
1082 0 : // as it might be a temporary issue and we don't want to re-download
1083 0 : // everything after it resolves.
1084 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1085 :
1086 0 : existent_timelines.insert(timeline_id);
1087 0 : continue;
1088 : }
1089 : };
1090 18 : match index_part {
1091 18 : MaybeDeletedIndexPart::IndexPart(index_part) => {
1092 18 : timeline_ancestors.insert(timeline_id, index_part.metadata.clone());
1093 18 : remote_index_and_client.insert(timeline_id, (index_part, preload.client));
1094 18 : }
1095 0 : MaybeDeletedIndexPart::Deleted(index_part) => {
1096 0 : info!(
1097 0 : "timeline {} is deleted, picking to resume deletion",
1098 : timeline_id
1099 : );
1100 0 : timelines_to_resume_deletions.push((timeline_id, index_part, preload.client));
1101 : }
1102 : }
1103 : }
1104 :
1105 564 : let mut gc_blocks = HashMap::new();
1106 :
1107 : // For every timeline, download the metadata file, scan the local directory,
1108 : // and build a layer map that contains an entry for each remote and local
1109 : // layer file.
1110 564 : let sorted_timelines = tree_sort_timelines(timeline_ancestors, |m| m.ancestor_timeline())?;
1111 582 : for (timeline_id, remote_metadata) in sorted_timelines {
1112 18 : let (index_part, remote_client) = remote_index_and_client
1113 18 : .remove(&timeline_id)
1114 18 : .expect("just put it in above");
1115 :
1116 18 : if let Some(blocking) = index_part.gc_blocking.as_ref() {
1117 : // could just filter these away, but it helps while testing
1118 0 : anyhow::ensure!(
1119 0 : !blocking.reasons.is_empty(),
1120 0 : "index_part for {timeline_id} is malformed: it should not have gc blocking with zero reasons"
1121 : );
1122 0 : let prev = gc_blocks.insert(timeline_id, blocking.reasons);
1123 0 : assert!(prev.is_none());
1124 18 : }
1125 :
1126 : // TODO again handle early failure
1127 18 : self.load_remote_timeline(
1128 18 : timeline_id,
1129 18 : index_part,
1130 18 : remote_metadata,
1131 18 : TimelineResources {
1132 18 : remote_client,
1133 18 : timeline_get_throttle: self.timeline_get_throttle.clone(),
1134 18 : l0_flush_global_state: self.l0_flush_global_state.clone(),
1135 18 : },
1136 18 : ctx,
1137 18 : )
1138 28 : .await
1139 18 : .with_context(|| {
1140 0 : format!(
1141 0 : "failed to load remote timeline {} for tenant {}",
1142 0 : timeline_id, self.tenant_shard_id
1143 0 : )
1144 18 : })?;
1145 : }
1146 :
1147 : // Walk through deleted timelines, resume deletion
1148 564 : for (timeline_id, index_part, remote_timeline_client) in timelines_to_resume_deletions {
1149 0 : remote_timeline_client
1150 0 : .init_upload_queue_stopped_to_continue_deletion(&index_part)
1151 0 : .context("init queue stopped")
1152 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1153 :
1154 0 : DeleteTimelineFlow::resume_deletion(
1155 0 : Arc::clone(self),
1156 0 : timeline_id,
1157 0 : &index_part.metadata,
1158 0 : remote_timeline_client,
1159 0 : )
1160 0 : .instrument(tracing::info_span!("timeline_delete", %timeline_id))
1161 0 : .await
1162 0 : .context("resume_deletion")
1163 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1164 : }
1165 :
1166 : // The local filesystem contents are a cache of what's in the remote IndexPart;
1167 : // IndexPart is the source of truth.
1168 564 : self.clean_up_timelines(&existent_timelines)?;
1169 :
1170 564 : self.gc_block.set_scanned(gc_blocks);
1171 564 :
1172 564 : fail::fail_point!("attach-before-activate", |_| {
1173 0 : anyhow::bail!("attach-before-activate");
1174 564 : });
1175 564 : failpoint_support::sleep_millis_async!("attach-before-activate-sleep", &self.cancel);
1176 :
1177 564 : info!("Done");
1178 :
1179 564 : Ok(())
1180 564 : }
1181 :
1182 : /// Check for any local timeline directories that are temporary, or do not correspond to a
1183 : /// timeline that still exists: this can happen if we crashed during a deletion/creation, or
1184 : /// if a timeline was deleted while the tenant was attached to a different pageserver.
1185 564 : fn clean_up_timelines(&self, existent_timelines: &HashSet<TimelineId>) -> anyhow::Result<()> {
1186 564 : let timelines_dir = self.conf.timelines_path(&self.tenant_shard_id);
1187 :
1188 564 : let entries = match timelines_dir.read_dir_utf8() {
1189 564 : Ok(d) => d,
1190 0 : Err(e) => {
1191 0 : if e.kind() == std::io::ErrorKind::NotFound {
1192 0 : return Ok(());
1193 : } else {
1194 0 : return Err(e).context("list timelines directory for tenant");
1195 : }
1196 : }
1197 : };
1198 :
1199 588 : for entry in entries {
1200 24 : let entry = entry.context("read timeline dir entry")?;
1201 24 : let entry_path = entry.path();
1202 :
1203 24 : let purge = if crate::is_temporary(entry_path)
1204 : // TODO: remove uninit mark code (https://github.com/neondatabase/neon/issues/5718)
1205 24 : || is_uninit_mark(entry_path)
1206 24 : || crate::is_delete_mark(entry_path)
1207 : {
1208 0 : true
1209 : } else {
1210 24 : match TimelineId::try_from(entry_path.file_name()) {
1211 24 : Ok(i) => {
1212 24 : // Purge if the timeline ID does not exist in remote storage: remote storage is the authority.
1213 24 : !existent_timelines.contains(&i)
1214 : }
1215 0 : Err(e) => {
1216 0 : tracing::warn!(
1217 0 : "Unparseable directory in timelines directory: {entry_path}, ignoring ({e})"
1218 : );
1219 : // Do not purge junk: if we don't recognize it, be cautious and leave it for a human.
1220 0 : false
1221 : }
1222 : }
1223 : };
1224 :
1225 24 : if purge {
1226 6 : tracing::info!("Purging stale timeline dentry {entry_path}");
1227 6 : if let Err(e) = match entry.file_type() {
1228 6 : Ok(t) => if t.is_dir() {
1229 6 : std::fs::remove_dir_all(entry_path)
1230 : } else {
1231 0 : std::fs::remove_file(entry_path)
1232 : }
1233 6 : .or_else(fs_ext::ignore_not_found),
1234 0 : Err(e) => Err(e),
1235 : } {
1236 0 : tracing::warn!("Failed to purge stale timeline dentry {entry_path}: {e}");
1237 6 : }
1238 18 : }
1239 : }
1240 :
1241 564 : Ok(())
1242 564 : }
1243 :
1244 : /// Get sum of all remote timelines sizes
1245 : ///
1246 : /// This function relies on the index_part instead of listing the remote storage
1247 0 : pub fn remote_size(&self) -> u64 {
1248 0 : let mut size = 0;
1249 :
1250 0 : for timeline in self.list_timelines() {
1251 0 : size += timeline.remote_client.get_remote_physical_size();
1252 0 : }
1253 :
1254 0 : size
1255 0 : }
1256 :
1257 36 : #[instrument(skip_all, fields(timeline_id=%timeline_id))]
1258 : async fn load_remote_timeline(
1259 : &self,
1260 : timeline_id: TimelineId,
1261 : index_part: IndexPart,
1262 : remote_metadata: TimelineMetadata,
1263 : resources: TimelineResources,
1264 : ctx: &RequestContext,
1265 : ) -> anyhow::Result<()> {
1266 : span::debug_assert_current_span_has_tenant_id();
1267 :
1268 : info!("downloading index file for timeline {}", timeline_id);
1269 : tokio::fs::create_dir_all(self.conf.timeline_path(&self.tenant_shard_id, &timeline_id))
1270 : .await
1271 : .context("Failed to create new timeline directory")?;
1272 :
1273 : let ancestor = if let Some(ancestor_id) = remote_metadata.ancestor_timeline() {
1274 : let timelines = self.timelines.lock().unwrap();
1275 : Some(Arc::clone(timelines.get(&ancestor_id).ok_or_else(
1276 0 : || {
1277 0 : anyhow::anyhow!(
1278 0 : "cannot find ancestor timeline {ancestor_id} for timeline {timeline_id}"
1279 0 : )
1280 0 : },
1281 : )?))
1282 : } else {
1283 : None
1284 : };
1285 :
1286 : let last_aux_file_policy = index_part.last_aux_file_policy();
1287 :
1288 : self.timeline_init_and_sync(
1289 : timeline_id,
1290 : resources,
1291 : Some(index_part),
1292 : remote_metadata,
1293 : ancestor,
1294 : last_aux_file_policy,
1295 : ctx,
1296 : )
1297 : .await
1298 : }
1299 :
1300 564 : async fn load_timeline_metadata(
1301 564 : self: &Arc<Tenant>,
1302 564 : timeline_ids: HashSet<TimelineId>,
1303 564 : remote_storage: &GenericRemoteStorage,
1304 564 : cancel: CancellationToken,
1305 564 : ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
1306 564 : let mut part_downloads = JoinSet::new();
1307 582 : for timeline_id in timeline_ids {
1308 18 : let client = RemoteTimelineClient::new(
1309 18 : remote_storage.clone(),
1310 18 : self.deletion_queue_client.clone(),
1311 18 : self.conf,
1312 18 : self.tenant_shard_id,
1313 18 : timeline_id,
1314 18 : self.generation,
1315 18 : );
1316 18 : let cancel_clone = cancel.clone();
1317 18 : part_downloads.spawn(
1318 18 : async move {
1319 18 : debug!("starting index part download");
1320 :
1321 71 : let index_part = client.download_index_file(&cancel_clone).await;
1322 :
1323 18 : debug!("finished index part download");
1324 :
1325 18 : Result::<_, anyhow::Error>::Ok(TimelinePreload {
1326 18 : client,
1327 18 : timeline_id,
1328 18 : index_part,
1329 18 : })
1330 18 : }
1331 18 : .map(move |res| {
1332 18 : res.with_context(|| format!("download index part for timeline {timeline_id}"))
1333 18 : })
1334 18 : .instrument(info_span!("download_index_part", %timeline_id)),
1335 : );
1336 : }
1337 :
1338 564 : let mut timeline_preloads: HashMap<TimelineId, TimelinePreload> = HashMap::new();
1339 :
1340 582 : loop {
1341 582 : tokio::select!(
1342 : next = part_downloads.join_next() => {
1343 : match next {
1344 : Some(result) => {
1345 : let preload_result = result.context("join preload task")?;
1346 : let preload = preload_result?;
1347 : timeline_preloads.insert(preload.timeline_id, preload);
1348 : },
1349 : None => {
1350 : break;
1351 : }
1352 : }
1353 : },
1354 : _ = cancel.cancelled() => {
1355 : anyhow::bail!("Cancelled while waiting for remote index download")
1356 : }
1357 582 : )
1358 582 : }
1359 :
1360 564 : Ok(timeline_preloads)
1361 564 : }
1362 :
1363 0 : pub(crate) async fn apply_timeline_archival_config(
1364 0 : &self,
1365 0 : timeline_id: TimelineId,
1366 0 : state: TimelineArchivalState,
1367 0 : ) -> Result<(), TimelineArchivalError> {
1368 0 : info!("setting timeline archival config");
1369 0 : let timeline = {
1370 0 : let timelines = self.timelines.lock().unwrap();
1371 :
1372 0 : let timeline = match timelines.get(&timeline_id) {
1373 0 : Some(t) => t,
1374 0 : None => return Err(TimelineArchivalError::NotFound),
1375 : };
1376 :
1377 : // Ensure that there are no non-archived child timelines
1378 0 : let children: Vec<TimelineId> = timelines
1379 0 : .iter()
1380 0 : .filter_map(|(id, entry)| {
1381 0 : if entry.get_ancestor_timeline_id() != Some(timeline_id) {
1382 0 : return None;
1383 0 : }
1384 0 : if entry.is_archived() == Some(true) {
1385 0 : return None;
1386 0 : }
1387 0 : Some(*id)
1388 0 : })
1389 0 : .collect();
1390 0 :
1391 0 : if !children.is_empty() && state == TimelineArchivalState::Archived {
1392 0 : return Err(TimelineArchivalError::HasUnarchivedChildren(children));
1393 0 : }
1394 0 : Arc::clone(timeline)
1395 : };
1396 :
1397 0 : let upload_needed = timeline
1398 0 : .remote_client
1399 0 : .schedule_index_upload_for_timeline_archival_state(state)?;
1400 :
1401 0 : if upload_needed {
1402 0 : info!("Uploading new state");
1403 : const MAX_WAIT: Duration = Duration::from_secs(10);
1404 0 : let Ok(v) =
1405 0 : tokio::time::timeout(MAX_WAIT, timeline.remote_client.wait_completion()).await
1406 : else {
1407 0 : tracing::warn!("reached timeout for waiting on upload queue");
1408 0 : return Err(TimelineArchivalError::Timeout);
1409 : };
1410 0 : v.map_err(|e| TimelineArchivalError::Other(anyhow::anyhow!(e)))?;
1411 0 : }
1412 0 : Ok(())
1413 0 : }
1414 :
1415 12 : pub(crate) fn tenant_shard_id(&self) -> TenantShardId {
1416 12 : self.tenant_shard_id
1417 12 : }
1418 :
1419 : /// Get Timeline handle for given Neon timeline ID.
1420 : /// This function is idempotent. It doesn't change internal state in any way.
1421 666 : pub fn get_timeline(
1422 666 : &self,
1423 666 : timeline_id: TimelineId,
1424 666 : active_only: bool,
1425 666 : ) -> Result<Arc<Timeline>, GetTimelineError> {
1426 666 : let timelines_accessor = self.timelines.lock().unwrap();
1427 666 : let timeline = timelines_accessor
1428 666 : .get(&timeline_id)
1429 666 : .ok_or(GetTimelineError::NotFound {
1430 666 : tenant_id: self.tenant_shard_id,
1431 666 : timeline_id,
1432 666 : })?;
1433 :
1434 660 : if active_only && !timeline.is_active() {
1435 0 : Err(GetTimelineError::NotActive {
1436 0 : tenant_id: self.tenant_shard_id,
1437 0 : timeline_id,
1438 0 : state: timeline.current_state(),
1439 0 : })
1440 : } else {
1441 660 : Ok(Arc::clone(timeline))
1442 : }
1443 666 : }
1444 :
1445 : /// Lists timelines the tenant contains.
1446 : /// Up to tenant's implementation to omit certain timelines that ar not considered ready for use.
1447 24 : pub fn list_timelines(&self) -> Vec<Arc<Timeline>> {
1448 24 : self.timelines
1449 24 : .lock()
1450 24 : .unwrap()
1451 24 : .values()
1452 24 : .map(Arc::clone)
1453 24 : .collect()
1454 24 : }
1455 :
1456 0 : pub fn list_timeline_ids(&self) -> Vec<TimelineId> {
1457 0 : self.timelines.lock().unwrap().keys().cloned().collect()
1458 0 : }
1459 :
1460 : /// This is used to create the initial 'main' timeline during bootstrapping,
1461 : /// or when importing a new base backup. The caller is expected to load an
1462 : /// initial image of the datadir to the new timeline after this.
1463 : ///
1464 : /// Until that happens, the on-disk state is invalid (disk_consistent_lsn=Lsn(0))
1465 : /// and the timeline will fail to load at a restart.
1466 : ///
1467 : /// For tests, use `DatadirModification::init_empty_test_timeline` + `commit` to setup the
1468 : /// minimum amount of keys required to get a writable timeline.
1469 : /// (Without it, `put` might fail due to `repartition` failing.)
1470 540 : pub(crate) async fn create_empty_timeline(
1471 540 : &self,
1472 540 : new_timeline_id: TimelineId,
1473 540 : initdb_lsn: Lsn,
1474 540 : pg_version: u32,
1475 540 : _ctx: &RequestContext,
1476 540 : ) -> anyhow::Result<UninitializedTimeline> {
1477 540 : anyhow::ensure!(
1478 540 : self.is_active(),
1479 0 : "Cannot create empty timelines on inactive tenant"
1480 : );
1481 :
1482 : // Protect against concurrent attempts to use this TimelineId
1483 540 : let create_guard = self.create_timeline_create_guard(new_timeline_id)?;
1484 :
1485 534 : let new_metadata = TimelineMetadata::new(
1486 534 : // Initialize disk_consistent LSN to 0, The caller must import some data to
1487 534 : // make it valid, before calling finish_creation()
1488 534 : Lsn(0),
1489 534 : None,
1490 534 : None,
1491 534 : Lsn(0),
1492 534 : initdb_lsn,
1493 534 : initdb_lsn,
1494 534 : pg_version,
1495 534 : );
1496 534 : self.prepare_new_timeline(
1497 534 : new_timeline_id,
1498 534 : &new_metadata,
1499 534 : create_guard,
1500 534 : initdb_lsn,
1501 534 : None,
1502 534 : None,
1503 534 : )
1504 0 : .await
1505 540 : }
1506 :
1507 : /// Helper for unit tests to create an empty timeline.
1508 : ///
1509 : /// The timeline is has state value `Active` but its background loops are not running.
1510 : // This makes the various functions which anyhow::ensure! for Active state work in tests.
1511 : // Our current tests don't need the background loops.
1512 : #[cfg(test)]
1513 510 : pub async fn create_test_timeline(
1514 510 : &self,
1515 510 : new_timeline_id: TimelineId,
1516 510 : initdb_lsn: Lsn,
1517 510 : pg_version: u32,
1518 510 : ctx: &RequestContext,
1519 510 : ) -> anyhow::Result<Arc<Timeline>> {
1520 510 : let uninit_tl = self
1521 510 : .create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
1522 0 : .await?;
1523 510 : let tline = uninit_tl.raw_timeline().expect("we just created it");
1524 510 : assert_eq!(tline.get_last_record_lsn(), Lsn(0));
1525 :
1526 : // Setup minimum keys required for the timeline to be usable.
1527 510 : let mut modification = tline.begin_modification(initdb_lsn);
1528 510 : modification
1529 510 : .init_empty_test_timeline()
1530 510 : .context("init_empty_test_timeline")?;
1531 510 : modification
1532 510 : .commit(ctx)
1533 492 : .await
1534 510 : .context("commit init_empty_test_timeline modification")?;
1535 :
1536 : // Flush to disk so that uninit_tl's check for valid disk_consistent_lsn passes.
1537 510 : tline.maybe_spawn_flush_loop();
1538 510 : tline.freeze_and_flush().await.context("freeze_and_flush")?;
1539 :
1540 : // Make sure the freeze_and_flush reaches remote storage.
1541 510 : tline.remote_client.wait_completion().await.unwrap();
1542 :
1543 510 : let tl = uninit_tl.finish_creation()?;
1544 : // The non-test code would call tl.activate() here.
1545 510 : tl.set_state(TimelineState::Active);
1546 510 : Ok(tl)
1547 510 : }
1548 :
1549 : /// Helper for unit tests to create a timeline with some pre-loaded states.
1550 : #[cfg(test)]
1551 : #[allow(clippy::too_many_arguments)]
1552 78 : pub async fn create_test_timeline_with_layers(
1553 78 : &self,
1554 78 : new_timeline_id: TimelineId,
1555 78 : initdb_lsn: Lsn,
1556 78 : pg_version: u32,
1557 78 : ctx: &RequestContext,
1558 78 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
1559 78 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
1560 78 : end_lsn: Lsn,
1561 78 : ) -> anyhow::Result<Arc<Timeline>> {
1562 78 : let tline = self
1563 78 : .create_test_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
1564 156 : .await?;
1565 78 : tline.force_advance_lsn(end_lsn);
1566 234 : for deltas in delta_layer_desc {
1567 156 : tline
1568 156 : .force_create_delta_layer(deltas, Some(initdb_lsn), ctx)
1569 468 : .await?;
1570 : }
1571 204 : for (lsn, images) in image_layer_desc {
1572 126 : tline
1573 126 : .force_create_image_layer(lsn, images, Some(initdb_lsn), ctx)
1574 717 : .await?;
1575 : }
1576 78 : Ok(tline)
1577 78 : }
1578 :
1579 : /// Create a new timeline.
1580 : ///
1581 : /// Returns the new timeline ID and reference to its Timeline object.
1582 : ///
1583 : /// If the caller specified the timeline ID to use (`new_timeline_id`), and timeline with
1584 : /// the same timeline ID already exists, returns CreateTimelineError::AlreadyExists.
1585 : #[allow(clippy::too_many_arguments)]
1586 0 : pub(crate) async fn create_timeline(
1587 0 : self: &Arc<Tenant>,
1588 0 : new_timeline_id: TimelineId,
1589 0 : ancestor_timeline_id: Option<TimelineId>,
1590 0 : mut ancestor_start_lsn: Option<Lsn>,
1591 0 : pg_version: u32,
1592 0 : load_existing_initdb: Option<TimelineId>,
1593 0 : broker_client: storage_broker::BrokerClientChannel,
1594 0 : ctx: &RequestContext,
1595 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
1596 0 : if !self.is_active() {
1597 0 : if matches!(self.current_state(), TenantState::Stopping { .. }) {
1598 0 : return Err(CreateTimelineError::ShuttingDown);
1599 : } else {
1600 0 : return Err(CreateTimelineError::Other(anyhow::anyhow!(
1601 0 : "Cannot create timelines on inactive tenant"
1602 0 : )));
1603 : }
1604 0 : }
1605 :
1606 0 : let _gate = self
1607 0 : .gate
1608 0 : .enter()
1609 0 : .map_err(|_| CreateTimelineError::ShuttingDown)?;
1610 :
1611 : // Get exclusive access to the timeline ID: this ensures that it does not already exist,
1612 : // and that no other creation attempts will be allowed in while we are working.
1613 0 : let create_guard = match self.create_timeline_create_guard(new_timeline_id) {
1614 0 : Ok(m) => m,
1615 : Err(TimelineExclusionError::AlreadyCreating) => {
1616 : // Creation is in progress, we cannot create it again, and we cannot
1617 : // check if this request matches the existing one, so caller must try
1618 : // again later.
1619 0 : return Err(CreateTimelineError::AlreadyCreating);
1620 : }
1621 0 : Err(TimelineExclusionError::Other(e)) => {
1622 0 : return Err(CreateTimelineError::Other(e));
1623 : }
1624 0 : Err(TimelineExclusionError::AlreadyExists(existing)) => {
1625 0 : debug!("timeline {new_timeline_id} already exists");
1626 :
1627 : // Idempotency: creating the same timeline twice is not an error, unless
1628 : // the second creation has different parameters.
1629 0 : if existing.get_ancestor_timeline_id() != ancestor_timeline_id
1630 0 : || existing.pg_version != pg_version
1631 0 : || (ancestor_start_lsn.is_some()
1632 0 : && ancestor_start_lsn != Some(existing.get_ancestor_lsn()))
1633 : {
1634 0 : return Err(CreateTimelineError::Conflict);
1635 0 : }
1636 0 :
1637 0 : // Wait for uploads to complete, so that when we return Ok, the timeline
1638 0 : // is known to be durable on remote storage. Just like we do at the end of
1639 0 : // this function, after we have created the timeline ourselves.
1640 0 : //
1641 0 : // We only really care that the initial version of `index_part.json` has
1642 0 : // been uploaded. That's enough to remember that the timeline
1643 0 : // exists. However, there is no function to wait specifically for that so
1644 0 : // we just wait for all in-progress uploads to finish.
1645 0 : existing
1646 0 : .remote_client
1647 0 : .wait_completion()
1648 0 : .await
1649 0 : .context("wait for timeline uploads to complete")?;
1650 :
1651 0 : return Ok(existing);
1652 : }
1653 : };
1654 :
1655 : pausable_failpoint!("timeline-creation-after-uninit");
1656 :
1657 0 : let loaded_timeline = match ancestor_timeline_id {
1658 0 : Some(ancestor_timeline_id) => {
1659 0 : let ancestor_timeline = self
1660 0 : .get_timeline(ancestor_timeline_id, false)
1661 0 : .context("Cannot branch off the timeline that's not present in pageserver")?;
1662 :
1663 : // instead of waiting around, just deny the request because ancestor is not yet
1664 : // ready for other purposes either.
1665 0 : if !ancestor_timeline.is_active() {
1666 0 : return Err(CreateTimelineError::AncestorNotActive);
1667 0 : }
1668 :
1669 0 : if let Some(lsn) = ancestor_start_lsn.as_mut() {
1670 0 : *lsn = lsn.align();
1671 0 :
1672 0 : let ancestor_ancestor_lsn = ancestor_timeline.get_ancestor_lsn();
1673 0 : if ancestor_ancestor_lsn > *lsn {
1674 : // can we safely just branch from the ancestor instead?
1675 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
1676 0 : "invalid start lsn {} for ancestor timeline {}: less than timeline ancestor lsn {}",
1677 0 : lsn,
1678 0 : ancestor_timeline_id,
1679 0 : ancestor_ancestor_lsn,
1680 0 : )));
1681 0 : }
1682 0 :
1683 0 : // Wait for the WAL to arrive and be processed on the parent branch up
1684 0 : // to the requested branch point. The repository code itself doesn't
1685 0 : // require it, but if we start to receive WAL on the new timeline,
1686 0 : // decoding the new WAL might need to look up previous pages, relation
1687 0 : // sizes etc. and that would get confused if the previous page versions
1688 0 : // are not in the repository yet.
1689 0 : ancestor_timeline
1690 0 : .wait_lsn(*lsn, timeline::WaitLsnWaiter::Tenant, ctx)
1691 0 : .await
1692 0 : .map_err(|e| match e {
1693 0 : e @ (WaitLsnError::Timeout(_) | WaitLsnError::BadState { .. }) => {
1694 0 : CreateTimelineError::AncestorLsn(anyhow::anyhow!(e))
1695 : }
1696 0 : WaitLsnError::Shutdown => CreateTimelineError::ShuttingDown,
1697 0 : })?;
1698 0 : }
1699 :
1700 0 : self.branch_timeline(
1701 0 : &ancestor_timeline,
1702 0 : new_timeline_id,
1703 0 : ancestor_start_lsn,
1704 0 : create_guard,
1705 0 : ctx,
1706 0 : )
1707 0 : .await?
1708 : }
1709 : None => {
1710 0 : self.bootstrap_timeline(
1711 0 : new_timeline_id,
1712 0 : pg_version,
1713 0 : load_existing_initdb,
1714 0 : create_guard,
1715 0 : ctx,
1716 0 : )
1717 0 : .await?
1718 : }
1719 : };
1720 :
1721 : // At this point we have dropped our guard on [`Self::timelines_creating`], and
1722 : // the timeline is visible in [`Self::timelines`], but it is _not_ durable yet. We must
1723 : // not send a success to the caller until it is. The same applies to handling retries,
1724 : // see the handling of [`TimelineExclusionError::AlreadyExists`] above.
1725 0 : let kind = ancestor_timeline_id
1726 0 : .map(|_| "branched")
1727 0 : .unwrap_or("bootstrapped");
1728 0 : loaded_timeline
1729 0 : .remote_client
1730 0 : .wait_completion()
1731 0 : .await
1732 0 : .with_context(|| format!("wait for {} timeline initial uploads to complete", kind))?;
1733 :
1734 0 : loaded_timeline.activate(self.clone(), broker_client, None, ctx);
1735 0 :
1736 0 : Ok(loaded_timeline)
1737 0 : }
1738 :
1739 0 : pub(crate) async fn delete_timeline(
1740 0 : self: Arc<Self>,
1741 0 : timeline_id: TimelineId,
1742 0 : ) -> Result<(), DeleteTimelineError> {
1743 0 : DeleteTimelineFlow::run(&self, timeline_id).await?;
1744 :
1745 0 : Ok(())
1746 0 : }
1747 :
1748 : /// perform one garbage collection iteration, removing old data files from disk.
1749 : /// this function is periodically called by gc task.
1750 : /// also it can be explicitly requested through page server api 'do_gc' command.
1751 : ///
1752 : /// `target_timeline_id` specifies the timeline to GC, or None for all.
1753 : ///
1754 : /// The `horizon` an `pitr` parameters determine how much WAL history needs to be retained.
1755 : /// Also known as the retention period, or the GC cutoff point. `horizon` specifies
1756 : /// the amount of history, as LSN difference from current latest LSN on each timeline.
1757 : /// `pitr` specifies the same as a time difference from the current time. The effective
1758 : /// GC cutoff point is determined conservatively by either `horizon` and `pitr`, whichever
1759 : /// requires more history to be retained.
1760 : //
1761 2262 : pub(crate) async fn gc_iteration(
1762 2262 : &self,
1763 2262 : target_timeline_id: Option<TimelineId>,
1764 2262 : horizon: u64,
1765 2262 : pitr: Duration,
1766 2262 : cancel: &CancellationToken,
1767 2262 : ctx: &RequestContext,
1768 2262 : ) -> Result<GcResult, GcError> {
1769 2262 : // Don't start doing work during shutdown
1770 2262 : if let TenantState::Stopping { .. } = self.current_state() {
1771 0 : return Ok(GcResult::default());
1772 2262 : }
1773 2262 :
1774 2262 : // there is a global allowed_error for this
1775 2262 : if !self.is_active() {
1776 0 : return Err(GcError::NotActive);
1777 2262 : }
1778 2262 :
1779 2262 : {
1780 2262 : let conf = self.tenant_conf.load();
1781 2262 :
1782 2262 : if !conf.location.may_delete_layers_hint() {
1783 0 : info!("Skipping GC in location state {:?}", conf.location);
1784 0 : return Ok(GcResult::default());
1785 2262 : }
1786 : }
1787 :
1788 2262 : let _guard = match self.gc_block.start().await {
1789 2262 : Ok(guard) => guard,
1790 0 : Err(reasons) => {
1791 0 : info!("Skipping GC: {reasons}");
1792 0 : return Ok(GcResult::default());
1793 : }
1794 : };
1795 :
1796 2262 : self.gc_iteration_internal(target_timeline_id, horizon, pitr, cancel, ctx)
1797 2171 : .await
1798 2262 : }
1799 :
1800 : /// Perform one compaction iteration.
1801 : /// This function is periodically called by compactor task.
1802 : /// Also it can be explicitly requested per timeline through page server
1803 : /// api's 'compact' command.
1804 : ///
1805 : /// Returns whether we have pending compaction task.
1806 0 : async fn compaction_iteration(
1807 0 : &self,
1808 0 : cancel: &CancellationToken,
1809 0 : ctx: &RequestContext,
1810 0 : ) -> Result<bool, timeline::CompactionError> {
1811 0 : // Don't start doing work during shutdown, or when broken, we do not need those in the logs
1812 0 : if !self.is_active() {
1813 0 : return Ok(false);
1814 0 : }
1815 0 :
1816 0 : {
1817 0 : let conf = self.tenant_conf.load();
1818 0 : if !conf.location.may_delete_layers_hint() || !conf.location.may_upload_layers_hint() {
1819 0 : info!("Skipping compaction in location state {:?}", conf.location);
1820 0 : return Ok(false);
1821 0 : }
1822 0 : }
1823 0 :
1824 0 : // Scan through the hashmap and collect a list of all the timelines,
1825 0 : // while holding the lock. Then drop the lock and actually perform the
1826 0 : // compactions. We don't want to block everything else while the
1827 0 : // compaction runs.
1828 0 : let timelines_to_compact = {
1829 0 : let timelines = self.timelines.lock().unwrap();
1830 0 : let timelines_to_compact = timelines
1831 0 : .iter()
1832 0 : .filter_map(|(timeline_id, timeline)| {
1833 0 : if timeline.is_active() {
1834 0 : Some((*timeline_id, timeline.clone()))
1835 : } else {
1836 0 : None
1837 : }
1838 0 : })
1839 0 : .collect::<Vec<_>>();
1840 0 : drop(timelines);
1841 0 : timelines_to_compact
1842 0 : };
1843 0 :
1844 0 : // Before doing any I/O work, check our circuit breaker
1845 0 : if self.compaction_circuit_breaker.lock().unwrap().is_broken() {
1846 0 : info!("Skipping compaction due to previous failures");
1847 0 : return Ok(false);
1848 0 : }
1849 0 :
1850 0 : let mut has_pending_task = false;
1851 :
1852 0 : for (timeline_id, timeline) in &timelines_to_compact {
1853 0 : has_pending_task |= timeline
1854 0 : .compact(cancel, EnumSet::empty(), ctx)
1855 0 : .instrument(info_span!("compact_timeline", %timeline_id))
1856 0 : .await
1857 0 : .inspect_err(|e| match e {
1858 0 : timeline::CompactionError::ShuttingDown => (),
1859 0 : timeline::CompactionError::Other(e) => {
1860 0 : self.compaction_circuit_breaker
1861 0 : .lock()
1862 0 : .unwrap()
1863 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, e);
1864 0 : }
1865 0 : })?;
1866 : }
1867 :
1868 0 : self.compaction_circuit_breaker
1869 0 : .lock()
1870 0 : .unwrap()
1871 0 : .success(&CIRCUIT_BREAKERS_UNBROKEN);
1872 0 :
1873 0 : Ok(has_pending_task)
1874 0 : }
1875 :
1876 : // Call through to all timelines to freeze ephemeral layers if needed. Usually
1877 : // this happens during ingest: this background housekeeping is for freezing layers
1878 : // that are open but haven't been written to for some time.
1879 0 : async fn ingest_housekeeping(&self) {
1880 0 : // Scan through the hashmap and collect a list of all the timelines,
1881 0 : // while holding the lock. Then drop the lock and actually perform the
1882 0 : // compactions. We don't want to block everything else while the
1883 0 : // compaction runs.
1884 0 : let timelines = {
1885 0 : self.timelines
1886 0 : .lock()
1887 0 : .unwrap()
1888 0 : .values()
1889 0 : .filter_map(|timeline| {
1890 0 : if timeline.is_active() {
1891 0 : Some(timeline.clone())
1892 : } else {
1893 0 : None
1894 : }
1895 0 : })
1896 0 : .collect::<Vec<_>>()
1897 : };
1898 :
1899 0 : for timeline in &timelines {
1900 0 : timeline.maybe_freeze_ephemeral_layer().await;
1901 : }
1902 0 : }
1903 :
1904 7344 : pub fn current_state(&self) -> TenantState {
1905 7344 : self.state.borrow().clone()
1906 7344 : }
1907 :
1908 5064 : pub fn is_active(&self) -> bool {
1909 5064 : self.current_state() == TenantState::Active
1910 5064 : }
1911 :
1912 0 : pub fn generation(&self) -> Generation {
1913 0 : self.generation
1914 0 : }
1915 :
1916 0 : pub(crate) fn wal_redo_manager_status(&self) -> Option<WalRedoManagerStatus> {
1917 0 : self.walredo_mgr.as_ref().and_then(|mgr| mgr.status())
1918 0 : }
1919 :
1920 : /// Changes tenant status to active, unless shutdown was already requested.
1921 : ///
1922 : /// `background_jobs_can_start` is an optional barrier set to a value during pageserver startup
1923 : /// to delay background jobs. Background jobs can be started right away when None is given.
1924 0 : fn activate(
1925 0 : self: &Arc<Self>,
1926 0 : broker_client: BrokerClientChannel,
1927 0 : background_jobs_can_start: Option<&completion::Barrier>,
1928 0 : ctx: &RequestContext,
1929 0 : ) {
1930 0 : span::debug_assert_current_span_has_tenant_id();
1931 0 :
1932 0 : let mut activating = false;
1933 0 : self.state.send_modify(|current_state| {
1934 0 : use pageserver_api::models::ActivatingFrom;
1935 0 : match &*current_state {
1936 : TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
1937 0 : panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {state:?}", state = current_state);
1938 : }
1939 0 : TenantState::Loading => {
1940 0 : *current_state = TenantState::Activating(ActivatingFrom::Loading);
1941 0 : }
1942 0 : TenantState::Attaching => {
1943 0 : *current_state = TenantState::Activating(ActivatingFrom::Attaching);
1944 0 : }
1945 : }
1946 0 : debug!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), "Activating tenant");
1947 0 : activating = true;
1948 0 : // Continue outside the closure. We need to grab timelines.lock()
1949 0 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
1950 0 : });
1951 0 :
1952 0 : if activating {
1953 0 : let timelines_accessor = self.timelines.lock().unwrap();
1954 0 : let timelines_to_activate = timelines_accessor
1955 0 : .values()
1956 0 : .filter(|timeline| !(timeline.is_broken() || timeline.is_stopping()));
1957 0 :
1958 0 : // Before activation, populate each Timeline's GcInfo with information about its children
1959 0 : self.initialize_gc_info(&timelines_accessor);
1960 0 :
1961 0 : // Spawn gc and compaction loops. The loops will shut themselves
1962 0 : // down when they notice that the tenant is inactive.
1963 0 : tasks::start_background_loops(self, background_jobs_can_start);
1964 0 :
1965 0 : let mut activated_timelines = 0;
1966 :
1967 0 : for timeline in timelines_to_activate {
1968 0 : timeline.activate(
1969 0 : self.clone(),
1970 0 : broker_client.clone(),
1971 0 : background_jobs_can_start,
1972 0 : ctx,
1973 0 : );
1974 0 : activated_timelines += 1;
1975 0 : }
1976 :
1977 0 : self.state.send_modify(move |current_state| {
1978 0 : assert!(
1979 0 : matches!(current_state, TenantState::Activating(_)),
1980 0 : "set_stopping and set_broken wait for us to leave Activating state",
1981 : );
1982 0 : *current_state = TenantState::Active;
1983 0 :
1984 0 : let elapsed = self.constructed_at.elapsed();
1985 0 : let total_timelines = timelines_accessor.len();
1986 0 :
1987 0 : // log a lot of stuff, because some tenants sometimes suffer from user-visible
1988 0 : // times to activate. see https://github.com/neondatabase/neon/issues/4025
1989 0 : info!(
1990 0 : since_creation_millis = elapsed.as_millis(),
1991 0 : tenant_id = %self.tenant_shard_id.tenant_id,
1992 0 : shard_id = %self.tenant_shard_id.shard_slug(),
1993 0 : activated_timelines,
1994 0 : total_timelines,
1995 0 : post_state = <&'static str>::from(&*current_state),
1996 0 : "activation attempt finished"
1997 : );
1998 :
1999 0 : TENANT.activation.observe(elapsed.as_secs_f64());
2000 0 : });
2001 0 : }
2002 0 : }
2003 :
2004 : /// Shutdown the tenant and join all of the spawned tasks.
2005 : ///
2006 : /// The method caters for all use-cases:
2007 : /// - pageserver shutdown (freeze_and_flush == true)
2008 : /// - detach + ignore (freeze_and_flush == false)
2009 : ///
2010 : /// This will attempt to shutdown even if tenant is broken.
2011 : ///
2012 : /// `shutdown_progress` is a [`completion::Barrier`] for the shutdown initiated by this call.
2013 : /// If the tenant is already shutting down, we return a clone of the first shutdown call's
2014 : /// `Barrier` as an `Err`. This not-first caller can use the returned barrier to join with
2015 : /// the ongoing shutdown.
2016 18 : async fn shutdown(
2017 18 : &self,
2018 18 : shutdown_progress: completion::Barrier,
2019 18 : shutdown_mode: timeline::ShutdownMode,
2020 18 : ) -> Result<(), completion::Barrier> {
2021 18 : span::debug_assert_current_span_has_tenant_id();
2022 :
2023 : // Set tenant (and its timlines) to Stoppping state.
2024 : //
2025 : // Since we can only transition into Stopping state after activation is complete,
2026 : // run it in a JoinSet so all tenants have a chance to stop before we get SIGKILLed.
2027 : //
2028 : // Transitioning tenants to Stopping state has a couple of non-obvious side effects:
2029 : // 1. Lock out any new requests to the tenants.
2030 : // 2. Signal cancellation to WAL receivers (we wait on it below).
2031 : // 3. Signal cancellation for other tenant background loops.
2032 : // 4. ???
2033 : //
2034 : // The waiting for the cancellation is not done uniformly.
2035 : // We certainly wait for WAL receivers to shut down.
2036 : // That is necessary so that no new data comes in before the freeze_and_flush.
2037 : // But the tenant background loops are joined-on in our caller.
2038 : // It's mesed up.
2039 : // we just ignore the failure to stop
2040 :
2041 : // If we're still attaching, fire the cancellation token early to drop out: this
2042 : // will prevent us flushing, but ensures timely shutdown if some I/O during attach
2043 : // is very slow.
2044 18 : let shutdown_mode = if matches!(self.current_state(), TenantState::Attaching) {
2045 0 : self.cancel.cancel();
2046 0 :
2047 0 : // Having fired our cancellation token, do not try and flush timelines: their cancellation tokens
2048 0 : // are children of ours, so their flush loops will have shut down already
2049 0 : timeline::ShutdownMode::Hard
2050 : } else {
2051 18 : shutdown_mode
2052 : };
2053 :
2054 18 : match self.set_stopping(shutdown_progress, false, false).await {
2055 18 : Ok(()) => {}
2056 0 : Err(SetStoppingError::Broken) => {
2057 0 : // assume that this is acceptable
2058 0 : }
2059 0 : Err(SetStoppingError::AlreadyStopping(other)) => {
2060 0 : // give caller the option to wait for this this shutdown
2061 0 : info!("Tenant::shutdown: AlreadyStopping");
2062 0 : return Err(other);
2063 : }
2064 : };
2065 :
2066 18 : let mut js = tokio::task::JoinSet::new();
2067 18 : {
2068 18 : let timelines = self.timelines.lock().unwrap();
2069 18 : timelines.values().for_each(|timeline| {
2070 18 : let timeline = Arc::clone(timeline);
2071 18 : let timeline_id = timeline.timeline_id;
2072 18 : let span = tracing::info_span!("timeline_shutdown", %timeline_id, ?shutdown_mode);
2073 37 : js.spawn(async move { timeline.shutdown(shutdown_mode).instrument(span).await });
2074 18 : })
2075 18 : };
2076 18 : // test_long_timeline_create_then_tenant_delete is leaning on this message
2077 18 : tracing::info!("Waiting for timelines...");
2078 36 : while let Some(res) = js.join_next().await {
2079 0 : match res {
2080 18 : Ok(()) => {}
2081 0 : Err(je) if je.is_cancelled() => unreachable!("no cancelling used"),
2082 0 : Err(je) if je.is_panic() => { /* logged already */ }
2083 0 : Err(je) => warn!("unexpected JoinError: {je:?}"),
2084 : }
2085 : }
2086 :
2087 : // We cancel the Tenant's cancellation token _after_ the timelines have all shut down. This permits
2088 : // them to continue to do work during their shutdown methods, e.g. flushing data.
2089 18 : tracing::debug!("Cancelling CancellationToken");
2090 18 : self.cancel.cancel();
2091 18 :
2092 18 : // shutdown all tenant and timeline tasks: gc, compaction, page service
2093 18 : // No new tasks will be started for this tenant because it's in `Stopping` state.
2094 18 : //
2095 18 : // this will additionally shutdown and await all timeline tasks.
2096 18 : tracing::debug!("Waiting for tasks...");
2097 18 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), None).await;
2098 :
2099 18 : if let Some(walredo_mgr) = self.walredo_mgr.as_ref() {
2100 18 : walredo_mgr.shutdown().await;
2101 0 : }
2102 :
2103 : // Wait for any in-flight operations to complete
2104 18 : self.gate.close().await;
2105 :
2106 18 : remove_tenant_metrics(&self.tenant_shard_id);
2107 18 :
2108 18 : Ok(())
2109 18 : }
2110 :
2111 : /// Change tenant status to Stopping, to mark that it is being shut down.
2112 : ///
2113 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
2114 : ///
2115 : /// This function is not cancel-safe!
2116 : ///
2117 : /// `allow_transition_from_loading` is needed for the special case of loading task deleting the tenant.
2118 : /// `allow_transition_from_attaching` is needed for the special case of attaching deleted tenant.
2119 18 : async fn set_stopping(
2120 18 : &self,
2121 18 : progress: completion::Barrier,
2122 18 : allow_transition_from_loading: bool,
2123 18 : allow_transition_from_attaching: bool,
2124 18 : ) -> Result<(), SetStoppingError> {
2125 18 : let mut rx = self.state.subscribe();
2126 18 :
2127 18 : // cannot stop before we're done activating, so wait out until we're done activating
2128 18 : rx.wait_for(|state| match state {
2129 0 : TenantState::Attaching if allow_transition_from_attaching => true,
2130 : TenantState::Activating(_) | TenantState::Attaching => {
2131 0 : info!(
2132 0 : "waiting for {} to turn Active|Broken|Stopping",
2133 0 : <&'static str>::from(state)
2134 : );
2135 0 : false
2136 : }
2137 0 : TenantState::Loading => allow_transition_from_loading,
2138 18 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
2139 18 : })
2140 0 : .await
2141 18 : .expect("cannot drop self.state while on a &self method");
2142 18 :
2143 18 : // we now know we're done activating, let's see whether this task is the winner to transition into Stopping
2144 18 : let mut err = None;
2145 18 : let stopping = self.state.send_if_modified(|current_state| match current_state {
2146 : TenantState::Activating(_) => {
2147 0 : unreachable!("1we ensured above that we're done with activation, and, there is no re-activation")
2148 : }
2149 : TenantState::Attaching => {
2150 0 : if !allow_transition_from_attaching {
2151 0 : unreachable!("2we ensured above that we're done with activation, and, there is no re-activation")
2152 0 : };
2153 0 : *current_state = TenantState::Stopping { progress };
2154 0 : true
2155 : }
2156 : TenantState::Loading => {
2157 0 : if !allow_transition_from_loading {
2158 0 : unreachable!("3we ensured above that we're done with activation, and, there is no re-activation")
2159 0 : };
2160 0 : *current_state = TenantState::Stopping { progress };
2161 0 : true
2162 : }
2163 : TenantState::Active => {
2164 : // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
2165 : // are created after the transition to Stopping. That's harmless, as the Timelines
2166 : // won't be accessible to anyone afterwards, because the Tenant is in Stopping state.
2167 18 : *current_state = TenantState::Stopping { progress };
2168 18 : // Continue stopping outside the closure. We need to grab timelines.lock()
2169 18 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
2170 18 : true
2171 : }
2172 0 : TenantState::Broken { reason, .. } => {
2173 0 : info!(
2174 0 : "Cannot set tenant to Stopping state, it is in Broken state due to: {reason}"
2175 : );
2176 0 : err = Some(SetStoppingError::Broken);
2177 0 : false
2178 : }
2179 0 : TenantState::Stopping { progress } => {
2180 0 : info!("Tenant is already in Stopping state");
2181 0 : err = Some(SetStoppingError::AlreadyStopping(progress.clone()));
2182 0 : false
2183 : }
2184 18 : });
2185 18 : match (stopping, err) {
2186 18 : (true, None) => {} // continue
2187 0 : (false, Some(err)) => return Err(err),
2188 0 : (true, Some(_)) => unreachable!(
2189 0 : "send_if_modified closure must error out if not transitioning to Stopping"
2190 0 : ),
2191 0 : (false, None) => unreachable!(
2192 0 : "send_if_modified closure must return true if transitioning to Stopping"
2193 0 : ),
2194 : }
2195 :
2196 18 : let timelines_accessor = self.timelines.lock().unwrap();
2197 18 : let not_broken_timelines = timelines_accessor
2198 18 : .values()
2199 18 : .filter(|timeline| !timeline.is_broken());
2200 36 : for timeline in not_broken_timelines {
2201 18 : timeline.set_state(TimelineState::Stopping);
2202 18 : }
2203 18 : Ok(())
2204 18 : }
2205 :
2206 : /// Method for tenant::mgr to transition us into Broken state in case of a late failure in
2207 : /// `remove_tenant_from_memory`
2208 : ///
2209 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
2210 : ///
2211 : /// In tests, we also use this to set tenants to Broken state on purpose.
2212 0 : pub(crate) async fn set_broken(&self, reason: String) {
2213 0 : let mut rx = self.state.subscribe();
2214 0 :
2215 0 : // The load & attach routines own the tenant state until it has reached `Active`.
2216 0 : // So, wait until it's done.
2217 0 : rx.wait_for(|state| match state {
2218 : TenantState::Activating(_) | TenantState::Loading | TenantState::Attaching => {
2219 0 : info!(
2220 0 : "waiting for {} to turn Active|Broken|Stopping",
2221 0 : <&'static str>::from(state)
2222 : );
2223 0 : false
2224 : }
2225 0 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
2226 0 : })
2227 0 : .await
2228 0 : .expect("cannot drop self.state while on a &self method");
2229 0 :
2230 0 : // we now know we're done activating, let's see whether this task is the winner to transition into Broken
2231 0 : self.set_broken_no_wait(reason)
2232 0 : }
2233 :
2234 0 : pub(crate) fn set_broken_no_wait(&self, reason: impl Display) {
2235 0 : let reason = reason.to_string();
2236 0 : self.state.send_modify(|current_state| {
2237 0 : match *current_state {
2238 : TenantState::Activating(_) | TenantState::Loading | TenantState::Attaching => {
2239 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
2240 : }
2241 : TenantState::Active => {
2242 0 : if cfg!(feature = "testing") {
2243 0 : warn!("Changing Active tenant to Broken state, reason: {}", reason);
2244 0 : *current_state = TenantState::broken_from_reason(reason);
2245 : } else {
2246 0 : unreachable!("not allowed to call set_broken on Active tenants in non-testing builds")
2247 : }
2248 : }
2249 : TenantState::Broken { .. } => {
2250 0 : warn!("Tenant is already in Broken state");
2251 : }
2252 : // This is the only "expected" path, any other path is a bug.
2253 : TenantState::Stopping { .. } => {
2254 0 : warn!(
2255 0 : "Marking Stopping tenant as Broken state, reason: {}",
2256 : reason
2257 : );
2258 0 : *current_state = TenantState::broken_from_reason(reason);
2259 : }
2260 : }
2261 0 : });
2262 0 : }
2263 :
2264 0 : pub fn subscribe_for_state_updates(&self) -> watch::Receiver<TenantState> {
2265 0 : self.state.subscribe()
2266 0 : }
2267 :
2268 : /// The activate_now semaphore is initialized with zero units. As soon as
2269 : /// we add a unit, waiters will be able to acquire a unit and proceed.
2270 0 : pub(crate) fn activate_now(&self) {
2271 0 : self.activate_now_sem.add_permits(1);
2272 0 : }
2273 :
2274 0 : pub(crate) async fn wait_to_become_active(
2275 0 : &self,
2276 0 : timeout: Duration,
2277 0 : ) -> Result<(), GetActiveTenantError> {
2278 0 : let mut receiver = self.state.subscribe();
2279 0 : loop {
2280 0 : let current_state = receiver.borrow_and_update().clone();
2281 0 : match current_state {
2282 : TenantState::Loading | TenantState::Attaching | TenantState::Activating(_) => {
2283 : // in these states, there's a chance that we can reach ::Active
2284 0 : self.activate_now();
2285 0 : match timeout_cancellable(timeout, &self.cancel, receiver.changed()).await {
2286 0 : Ok(r) => {
2287 0 : r.map_err(
2288 0 : |_e: tokio::sync::watch::error::RecvError|
2289 : // Tenant existed but was dropped: report it as non-existent
2290 0 : GetActiveTenantError::NotFound(GetTenantError::NotFound(self.tenant_shard_id.tenant_id))
2291 0 : )?
2292 : }
2293 : Err(TimeoutCancellableError::Cancelled) => {
2294 0 : return Err(GetActiveTenantError::Cancelled);
2295 : }
2296 : Err(TimeoutCancellableError::Timeout) => {
2297 0 : return Err(GetActiveTenantError::WaitForActiveTimeout {
2298 0 : latest_state: Some(self.current_state()),
2299 0 : wait_time: timeout,
2300 0 : });
2301 : }
2302 : }
2303 : }
2304 : TenantState::Active { .. } => {
2305 0 : return Ok(());
2306 : }
2307 0 : TenantState::Broken { reason, .. } => {
2308 0 : // This is fatal, and reported distinctly from the general case of "will never be active" because
2309 0 : // it's logically a 500 to external API users (broken is always a bug).
2310 0 : return Err(GetActiveTenantError::Broken(reason));
2311 : }
2312 : TenantState::Stopping { .. } => {
2313 : // There's no chance the tenant can transition back into ::Active
2314 0 : return Err(GetActiveTenantError::WillNotBecomeActive(current_state));
2315 : }
2316 : }
2317 : }
2318 0 : }
2319 :
2320 0 : pub(crate) fn get_attach_mode(&self) -> AttachmentMode {
2321 0 : self.tenant_conf.load().location.attach_mode
2322 0 : }
2323 :
2324 : /// For API access: generate a LocationConfig equivalent to the one that would be used to
2325 : /// create a Tenant in the same state. Do not use this in hot paths: it's for relatively
2326 : /// rare external API calls, like a reconciliation at startup.
2327 0 : pub(crate) fn get_location_conf(&self) -> models::LocationConfig {
2328 0 : let conf = self.tenant_conf.load();
2329 :
2330 0 : let location_config_mode = match conf.location.attach_mode {
2331 0 : AttachmentMode::Single => models::LocationConfigMode::AttachedSingle,
2332 0 : AttachmentMode::Multi => models::LocationConfigMode::AttachedMulti,
2333 0 : AttachmentMode::Stale => models::LocationConfigMode::AttachedStale,
2334 : };
2335 :
2336 : // We have a pageserver TenantConf, we need the API-facing TenantConfig.
2337 0 : let tenant_config: models::TenantConfig = conf.tenant_conf.clone().into();
2338 0 :
2339 0 : models::LocationConfig {
2340 0 : mode: location_config_mode,
2341 0 : generation: self.generation.into(),
2342 0 : secondary_conf: None,
2343 0 : shard_number: self.shard_identity.number.0,
2344 0 : shard_count: self.shard_identity.count.literal(),
2345 0 : shard_stripe_size: self.shard_identity.stripe_size.0,
2346 0 : tenant_conf: tenant_config,
2347 0 : }
2348 0 : }
2349 :
2350 0 : pub(crate) fn get_tenant_shard_id(&self) -> &TenantShardId {
2351 0 : &self.tenant_shard_id
2352 0 : }
2353 :
2354 0 : pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
2355 0 : self.shard_identity.stripe_size
2356 0 : }
2357 :
2358 0 : pub(crate) fn get_generation(&self) -> Generation {
2359 0 : self.generation
2360 0 : }
2361 :
2362 : /// This function partially shuts down the tenant (it shuts down the Timelines) and is fallible,
2363 : /// and can leave the tenant in a bad state if it fails. The caller is responsible for
2364 : /// resetting this tenant to a valid state if we fail.
2365 0 : pub(crate) async fn split_prepare(
2366 0 : &self,
2367 0 : child_shards: &Vec<TenantShardId>,
2368 0 : ) -> anyhow::Result<()> {
2369 0 : let timelines = self.timelines.lock().unwrap().clone();
2370 0 : for timeline in timelines.values() {
2371 : // We do not block timeline creation/deletion during splits inside the pageserver: it is up to higher levels
2372 : // to ensure that they do not start a split if currently in the process of doing these.
2373 :
2374 : // Upload an index from the parent: this is partly to provide freshness for the
2375 : // child tenants that will copy it, and partly for general ease-of-debugging: there will
2376 : // always be a parent shard index in the same generation as we wrote the child shard index.
2377 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Uploading index");
2378 0 : timeline
2379 0 : .remote_client
2380 0 : .schedule_index_upload_for_file_changes()?;
2381 0 : timeline.remote_client.wait_completion().await?;
2382 :
2383 : // Shut down the timeline's remote client: this means that the indices we write
2384 : // for child shards will not be invalidated by the parent shard deleting layers.
2385 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Shutting down remote storage client");
2386 0 : timeline.remote_client.shutdown().await;
2387 :
2388 : // Download methods can still be used after shutdown, as they don't flow through the remote client's
2389 : // queue. In principal the RemoteTimelineClient could provide this without downloading it, but this
2390 : // operation is rare, so it's simpler to just download it (and robustly guarantees that the index
2391 : // we use here really is the remotely persistent one).
2392 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Downloading index_part from parent");
2393 0 : let result = timeline.remote_client
2394 0 : .download_index_file(&self.cancel)
2395 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))
2396 0 : .await?;
2397 0 : let index_part = match result {
2398 : MaybeDeletedIndexPart::Deleted(_) => {
2399 0 : anyhow::bail!("Timeline deletion happened concurrently with split")
2400 : }
2401 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
2402 : };
2403 :
2404 0 : for child_shard in child_shards {
2405 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Uploading index_part for child {}", child_shard.to_index());
2406 0 : upload_index_part(
2407 0 : &self.remote_storage,
2408 0 : child_shard,
2409 0 : &timeline.timeline_id,
2410 0 : self.generation,
2411 0 : &index_part,
2412 0 : &self.cancel,
2413 0 : )
2414 0 : .await?;
2415 : }
2416 : }
2417 :
2418 0 : Ok(())
2419 0 : }
2420 :
2421 0 : pub(crate) fn get_sizes(&self) -> TopTenantShardItem {
2422 0 : let mut result = TopTenantShardItem {
2423 0 : id: self.tenant_shard_id,
2424 0 : resident_size: 0,
2425 0 : physical_size: 0,
2426 0 : max_logical_size: 0,
2427 0 : };
2428 :
2429 0 : for timeline in self.timelines.lock().unwrap().values() {
2430 0 : result.resident_size += timeline.metrics.resident_physical_size_gauge.get();
2431 0 :
2432 0 : result.physical_size += timeline
2433 0 : .remote_client
2434 0 : .metrics
2435 0 : .remote_physical_size_gauge
2436 0 : .get();
2437 0 : result.max_logical_size = std::cmp::max(
2438 0 : result.max_logical_size,
2439 0 : timeline.metrics.current_logical_size_gauge.get(),
2440 0 : );
2441 0 : }
2442 :
2443 0 : result
2444 0 : }
2445 : }
2446 :
2447 : /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
2448 : /// perform a topological sort, so that the parent of each timeline comes
2449 : /// before the children.
2450 : /// E extracts the ancestor from T
2451 : /// This allows for T to be different. It can be TimelineMetadata, can be Timeline itself, etc.
2452 564 : fn tree_sort_timelines<T, E>(
2453 564 : timelines: HashMap<TimelineId, T>,
2454 564 : extractor: E,
2455 564 : ) -> anyhow::Result<Vec<(TimelineId, T)>>
2456 564 : where
2457 564 : E: Fn(&T) -> Option<TimelineId>,
2458 564 : {
2459 564 : let mut result = Vec::with_capacity(timelines.len());
2460 564 :
2461 564 : let mut now = Vec::with_capacity(timelines.len());
2462 564 : // (ancestor, children)
2463 564 : let mut later: HashMap<TimelineId, Vec<(TimelineId, T)>> =
2464 564 : HashMap::with_capacity(timelines.len());
2465 :
2466 582 : for (timeline_id, value) in timelines {
2467 18 : if let Some(ancestor_id) = extractor(&value) {
2468 6 : let children = later.entry(ancestor_id).or_default();
2469 6 : children.push((timeline_id, value));
2470 12 : } else {
2471 12 : now.push((timeline_id, value));
2472 12 : }
2473 : }
2474 :
2475 582 : while let Some((timeline_id, metadata)) = now.pop() {
2476 18 : result.push((timeline_id, metadata));
2477 : // All children of this can be loaded now
2478 18 : if let Some(mut children) = later.remove(&timeline_id) {
2479 6 : now.append(&mut children);
2480 12 : }
2481 : }
2482 :
2483 : // All timelines should be visited now. Unless there were timelines with missing ancestors.
2484 564 : if !later.is_empty() {
2485 0 : for (missing_id, orphan_ids) in later {
2486 0 : for (orphan_id, _) in orphan_ids {
2487 0 : error!("could not load timeline {orphan_id} because its ancestor timeline {missing_id} could not be loaded");
2488 : }
2489 : }
2490 0 : bail!("could not load tenant because some timelines are missing ancestors");
2491 564 : }
2492 564 :
2493 564 : Ok(result)
2494 564 : }
2495 :
2496 : impl Tenant {
2497 0 : pub fn tenant_specific_overrides(&self) -> TenantConfOpt {
2498 0 : self.tenant_conf.load().tenant_conf.clone()
2499 0 : }
2500 :
2501 0 : pub fn effective_config(&self) -> TenantConf {
2502 0 : self.tenant_specific_overrides()
2503 0 : .merge(self.conf.default_tenant_conf.clone())
2504 0 : }
2505 :
2506 0 : pub fn get_checkpoint_distance(&self) -> u64 {
2507 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2508 0 : tenant_conf
2509 0 : .checkpoint_distance
2510 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
2511 0 : }
2512 :
2513 0 : pub fn get_checkpoint_timeout(&self) -> Duration {
2514 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2515 0 : tenant_conf
2516 0 : .checkpoint_timeout
2517 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
2518 0 : }
2519 :
2520 0 : pub fn get_compaction_target_size(&self) -> u64 {
2521 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2522 0 : tenant_conf
2523 0 : .compaction_target_size
2524 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
2525 0 : }
2526 :
2527 0 : pub fn get_compaction_period(&self) -> Duration {
2528 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2529 0 : tenant_conf
2530 0 : .compaction_period
2531 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
2532 0 : }
2533 :
2534 0 : pub fn get_compaction_threshold(&self) -> usize {
2535 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2536 0 : tenant_conf
2537 0 : .compaction_threshold
2538 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
2539 0 : }
2540 :
2541 0 : pub fn get_gc_horizon(&self) -> u64 {
2542 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2543 0 : tenant_conf
2544 0 : .gc_horizon
2545 0 : .unwrap_or(self.conf.default_tenant_conf.gc_horizon)
2546 0 : }
2547 :
2548 0 : pub fn get_gc_period(&self) -> Duration {
2549 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2550 0 : tenant_conf
2551 0 : .gc_period
2552 0 : .unwrap_or(self.conf.default_tenant_conf.gc_period)
2553 0 : }
2554 :
2555 0 : pub fn get_image_creation_threshold(&self) -> usize {
2556 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2557 0 : tenant_conf
2558 0 : .image_creation_threshold
2559 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
2560 0 : }
2561 :
2562 0 : pub fn get_pitr_interval(&self) -> Duration {
2563 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2564 0 : tenant_conf
2565 0 : .pitr_interval
2566 0 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
2567 0 : }
2568 :
2569 0 : pub fn get_min_resident_size_override(&self) -> Option<u64> {
2570 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2571 0 : tenant_conf
2572 0 : .min_resident_size_override
2573 0 : .or(self.conf.default_tenant_conf.min_resident_size_override)
2574 0 : }
2575 :
2576 0 : pub fn get_heatmap_period(&self) -> Option<Duration> {
2577 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2578 0 : let heatmap_period = tenant_conf
2579 0 : .heatmap_period
2580 0 : .unwrap_or(self.conf.default_tenant_conf.heatmap_period);
2581 0 : if heatmap_period.is_zero() {
2582 0 : None
2583 : } else {
2584 0 : Some(heatmap_period)
2585 : }
2586 0 : }
2587 :
2588 0 : pub fn get_lsn_lease_length(&self) -> Duration {
2589 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2590 0 : tenant_conf
2591 0 : .lsn_lease_length
2592 0 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
2593 0 : }
2594 :
2595 0 : pub fn set_new_tenant_config(&self, new_tenant_conf: TenantConfOpt) {
2596 0 : // Use read-copy-update in order to avoid overwriting the location config
2597 0 : // state if this races with [`Tenant::set_new_location_config`]. Note that
2598 0 : // this race is not possible if both request types come from the storage
2599 0 : // controller (as they should!) because an exclusive op lock is required
2600 0 : // on the storage controller side.
2601 0 : self.tenant_conf.rcu(|inner| {
2602 0 : Arc::new(AttachedTenantConf {
2603 0 : tenant_conf: new_tenant_conf.clone(),
2604 0 : location: inner.location,
2605 0 : })
2606 0 : });
2607 0 :
2608 0 : self.tenant_conf_updated(&new_tenant_conf);
2609 0 : // Don't hold self.timelines.lock() during the notifies.
2610 0 : // There's no risk of deadlock right now, but there could be if we consolidate
2611 0 : // mutexes in struct Timeline in the future.
2612 0 : let timelines = self.list_timelines();
2613 0 : for timeline in timelines {
2614 0 : timeline.tenant_conf_updated(&new_tenant_conf);
2615 0 : }
2616 0 : }
2617 :
2618 24 : pub(crate) fn set_new_location_config(&self, new_conf: AttachedTenantConf) {
2619 24 : let new_tenant_conf = new_conf.tenant_conf.clone();
2620 24 :
2621 24 : self.tenant_conf.store(Arc::new(new_conf));
2622 24 :
2623 24 : self.tenant_conf_updated(&new_tenant_conf);
2624 24 : // Don't hold self.timelines.lock() during the notifies.
2625 24 : // There's no risk of deadlock right now, but there could be if we consolidate
2626 24 : // mutexes in struct Timeline in the future.
2627 24 : let timelines = self.list_timelines();
2628 48 : for timeline in timelines {
2629 24 : timeline.tenant_conf_updated(&new_tenant_conf);
2630 24 : }
2631 24 : }
2632 :
2633 588 : fn get_timeline_get_throttle_config(
2634 588 : psconf: &'static PageServerConf,
2635 588 : overrides: &TenantConfOpt,
2636 588 : ) -> throttle::Config {
2637 588 : overrides
2638 588 : .timeline_get_throttle
2639 588 : .clone()
2640 588 : .unwrap_or(psconf.default_tenant_conf.timeline_get_throttle.clone())
2641 588 : }
2642 :
2643 24 : pub(crate) fn tenant_conf_updated(&self, new_conf: &TenantConfOpt) {
2644 24 : let conf = Self::get_timeline_get_throttle_config(self.conf, new_conf);
2645 24 : self.timeline_get_throttle.reconfigure(conf)
2646 24 : }
2647 :
2648 : /// Helper function to create a new Timeline struct.
2649 : ///
2650 : /// The returned Timeline is in Loading state. The caller is responsible for
2651 : /// initializing any on-disk state, and for inserting the Timeline to the 'timelines'
2652 : /// map.
2653 : ///
2654 : /// `validate_ancestor == false` is used when a timeline is created for deletion
2655 : /// and we might not have the ancestor present anymore which is fine for to be
2656 : /// deleted timelines.
2657 1236 : fn create_timeline_struct(
2658 1236 : &self,
2659 1236 : new_timeline_id: TimelineId,
2660 1236 : new_metadata: &TimelineMetadata,
2661 1236 : ancestor: Option<Arc<Timeline>>,
2662 1236 : resources: TimelineResources,
2663 1236 : cause: CreateTimelineCause,
2664 1236 : last_aux_file_policy: Option<AuxFilePolicy>,
2665 1236 : ) -> anyhow::Result<Arc<Timeline>> {
2666 1236 : let state = match cause {
2667 : CreateTimelineCause::Load => {
2668 1236 : let ancestor_id = new_metadata.ancestor_timeline();
2669 1236 : anyhow::ensure!(
2670 1236 : ancestor_id == ancestor.as_ref().map(|t| t.timeline_id),
2671 0 : "Timeline's {new_timeline_id} ancestor {ancestor_id:?} was not found"
2672 : );
2673 1236 : TimelineState::Loading
2674 : }
2675 0 : CreateTimelineCause::Delete => TimelineState::Stopping,
2676 : };
2677 :
2678 1236 : let pg_version = new_metadata.pg_version();
2679 1236 :
2680 1236 : let timeline = Timeline::new(
2681 1236 : self.conf,
2682 1236 : Arc::clone(&self.tenant_conf),
2683 1236 : new_metadata,
2684 1236 : ancestor,
2685 1236 : new_timeline_id,
2686 1236 : self.tenant_shard_id,
2687 1236 : self.generation,
2688 1236 : self.shard_identity,
2689 1236 : self.walredo_mgr.clone(),
2690 1236 : resources,
2691 1236 : pg_version,
2692 1236 : state,
2693 1236 : last_aux_file_policy,
2694 1236 : self.cancel.child_token(),
2695 1236 : );
2696 1236 :
2697 1236 : Ok(timeline)
2698 1236 : }
2699 :
2700 : // Allow too_many_arguments because a constructor's argument list naturally grows with the
2701 : // number of attributes in the struct: breaking these out into a builder wouldn't be helpful.
2702 : #[allow(clippy::too_many_arguments)]
2703 564 : fn new(
2704 564 : state: TenantState,
2705 564 : conf: &'static PageServerConf,
2706 564 : attached_conf: AttachedTenantConf,
2707 564 : shard_identity: ShardIdentity,
2708 564 : walredo_mgr: Option<Arc<WalRedoManager>>,
2709 564 : tenant_shard_id: TenantShardId,
2710 564 : remote_storage: GenericRemoteStorage,
2711 564 : deletion_queue_client: DeletionQueueClient,
2712 564 : l0_flush_global_state: L0FlushGlobalState,
2713 564 : ) -> Tenant {
2714 564 : debug_assert!(
2715 564 : !attached_conf.location.generation.is_none() || conf.control_plane_api.is_none()
2716 : );
2717 :
2718 564 : let (state, mut rx) = watch::channel(state);
2719 564 :
2720 564 : tokio::spawn(async move {
2721 564 : // reflect tenant state in metrics:
2722 564 : // - global per tenant state: TENANT_STATE_METRIC
2723 564 : // - "set" of broken tenants: BROKEN_TENANTS_SET
2724 564 : //
2725 564 : // set of broken tenants should not have zero counts so that it remains accessible for
2726 564 : // alerting.
2727 564 :
2728 564 : let tid = tenant_shard_id.to_string();
2729 564 : let shard_id = tenant_shard_id.shard_slug().to_string();
2730 564 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
2731 564 :
2732 1121 : fn inspect_state(state: &TenantState) -> ([&'static str; 1], bool) {
2733 1121 : ([state.into()], matches!(state, TenantState::Broken { .. }))
2734 1121 : }
2735 564 :
2736 564 : let mut tuple = inspect_state(&rx.borrow_and_update());
2737 564 :
2738 564 : let is_broken = tuple.1;
2739 564 : let mut counted_broken = if is_broken {
2740 : // add the id to the set right away, there should not be any updates on the channel
2741 : // after before tenant is removed, if ever
2742 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
2743 0 : true
2744 : } else {
2745 564 : false
2746 : };
2747 :
2748 1121 : loop {
2749 1121 : let labels = &tuple.0;
2750 1121 : let current = TENANT_STATE_METRIC.with_label_values(labels);
2751 1121 : current.inc();
2752 1121 :
2753 1121 : if rx.changed().await.is_err() {
2754 : // tenant has been dropped
2755 48 : current.dec();
2756 48 : drop(BROKEN_TENANTS_SET.remove_label_values(set_key));
2757 48 : break;
2758 557 : }
2759 557 :
2760 557 : current.dec();
2761 557 : tuple = inspect_state(&rx.borrow_and_update());
2762 557 :
2763 557 : let is_broken = tuple.1;
2764 557 : if is_broken && !counted_broken {
2765 0 : counted_broken = true;
2766 0 : // insert the tenant_id (back) into the set while avoiding needless counter
2767 0 : // access
2768 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
2769 557 : }
2770 : }
2771 564 : });
2772 564 :
2773 564 : Tenant {
2774 564 : tenant_shard_id,
2775 564 : shard_identity,
2776 564 : generation: attached_conf.location.generation,
2777 564 : conf,
2778 564 : // using now here is good enough approximation to catch tenants with really long
2779 564 : // activation times.
2780 564 : constructed_at: Instant::now(),
2781 564 : timelines: Mutex::new(HashMap::new()),
2782 564 : timelines_creating: Mutex::new(HashSet::new()),
2783 564 : gc_cs: tokio::sync::Mutex::new(()),
2784 564 : walredo_mgr,
2785 564 : remote_storage,
2786 564 : deletion_queue_client,
2787 564 : state,
2788 564 : cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
2789 564 : cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
2790 564 : eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
2791 564 : compaction_circuit_breaker: std::sync::Mutex::new(CircuitBreaker::new(
2792 564 : format!("compaction-{tenant_shard_id}"),
2793 564 : 5,
2794 564 : // Compaction can be a very expensive operation, and might leak disk space. It also ought
2795 564 : // to be infallible, as long as remote storage is available. So if it repeatedly fails,
2796 564 : // use an extremely long backoff.
2797 564 : Some(Duration::from_secs(3600 * 24)),
2798 564 : )),
2799 564 : activate_now_sem: tokio::sync::Semaphore::new(0),
2800 564 : cancel: CancellationToken::default(),
2801 564 : gate: Gate::default(),
2802 564 : timeline_get_throttle: Arc::new(throttle::Throttle::new(
2803 564 : Tenant::get_timeline_get_throttle_config(conf, &attached_conf.tenant_conf),
2804 564 : &crate::metrics::tenant_throttling::TIMELINE_GET,
2805 564 : )),
2806 564 : tenant_conf: Arc::new(ArcSwap::from_pointee(attached_conf)),
2807 564 : ongoing_timeline_detach: std::sync::Mutex::default(),
2808 564 : gc_block: Default::default(),
2809 564 : l0_flush_global_state,
2810 564 : }
2811 564 : }
2812 :
2813 : /// Locate and load config
2814 0 : pub(super) fn load_tenant_config(
2815 0 : conf: &'static PageServerConf,
2816 0 : tenant_shard_id: &TenantShardId,
2817 0 : ) -> Result<LocationConf, LoadConfigError> {
2818 0 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
2819 0 :
2820 0 : info!("loading tenant configuration from {config_path}");
2821 :
2822 : // load and parse file
2823 0 : let config = fs::read_to_string(&config_path).map_err(|e| {
2824 0 : match e.kind() {
2825 : std::io::ErrorKind::NotFound => {
2826 : // The config should almost always exist for a tenant directory:
2827 : // - When attaching a tenant, the config is the first thing we write
2828 : // - When detaching a tenant, we atomically move the directory to a tmp location
2829 : // before deleting contents.
2830 : //
2831 : // The very rare edge case that can result in a missing config is if we crash during attach
2832 : // between creating directory and writing config. Callers should handle that as if the
2833 : // directory didn't exist.
2834 :
2835 0 : LoadConfigError::NotFound(config_path)
2836 : }
2837 : _ => {
2838 : // No IO errors except NotFound are acceptable here: other kinds of error indicate local storage or permissions issues
2839 : // that we cannot cleanly recover
2840 0 : crate::virtual_file::on_fatal_io_error(&e, "Reading tenant config file")
2841 : }
2842 : }
2843 0 : })?;
2844 :
2845 0 : Ok(toml_edit::de::from_str::<LocationConf>(&config)?)
2846 0 : }
2847 :
2848 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
2849 : pub(super) async fn persist_tenant_config(
2850 : conf: &'static PageServerConf,
2851 : tenant_shard_id: &TenantShardId,
2852 : location_conf: &LocationConf,
2853 : ) -> std::io::Result<()> {
2854 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
2855 :
2856 : Self::persist_tenant_config_at(tenant_shard_id, &config_path, location_conf).await
2857 : }
2858 :
2859 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
2860 : pub(super) async fn persist_tenant_config_at(
2861 : tenant_shard_id: &TenantShardId,
2862 : config_path: &Utf8Path,
2863 : location_conf: &LocationConf,
2864 : ) -> std::io::Result<()> {
2865 : debug!("persisting tenantconf to {config_path}");
2866 :
2867 : let mut conf_content = r#"# This file contains a specific per-tenant's config.
2868 : # It is read in case of pageserver restart.
2869 : "#
2870 : .to_string();
2871 :
2872 0 : fail::fail_point!("tenant-config-before-write", |_| {
2873 0 : Err(std::io::Error::new(
2874 0 : std::io::ErrorKind::Other,
2875 0 : "tenant-config-before-write",
2876 0 : ))
2877 0 : });
2878 :
2879 : // Convert the config to a toml file.
2880 : conf_content +=
2881 : &toml_edit::ser::to_string_pretty(&location_conf).expect("Config serialization failed");
2882 :
2883 : let temp_path = path_with_suffix_extension(config_path, TEMP_FILE_SUFFIX);
2884 :
2885 : let conf_content = conf_content.into_bytes();
2886 : VirtualFile::crashsafe_overwrite(config_path.to_owned(), temp_path, conf_content).await
2887 : }
2888 :
2889 : //
2890 : // How garbage collection works:
2891 : //
2892 : // +--bar------------->
2893 : // /
2894 : // +----+-----foo---------------->
2895 : // /
2896 : // ----main--+-------------------------->
2897 : // \
2898 : // +-----baz-------->
2899 : //
2900 : //
2901 : // 1. Grab 'gc_cs' mutex to prevent new timelines from being created while Timeline's
2902 : // `gc_infos` are being refreshed
2903 : // 2. Scan collected timelines, and on each timeline, make note of the
2904 : // all the points where other timelines have been branched off.
2905 : // We will refrain from removing page versions at those LSNs.
2906 : // 3. For each timeline, scan all layer files on the timeline.
2907 : // Remove all files for which a newer file exists and which
2908 : // don't cover any branch point LSNs.
2909 : //
2910 : // TODO:
2911 : // - if a relation has a non-incremental persistent layer on a child branch, then we
2912 : // don't need to keep that in the parent anymore. But currently
2913 : // we do.
2914 2262 : async fn gc_iteration_internal(
2915 2262 : &self,
2916 2262 : target_timeline_id: Option<TimelineId>,
2917 2262 : horizon: u64,
2918 2262 : pitr: Duration,
2919 2262 : cancel: &CancellationToken,
2920 2262 : ctx: &RequestContext,
2921 2262 : ) -> Result<GcResult, GcError> {
2922 2262 : let mut totals: GcResult = Default::default();
2923 2262 : let now = Instant::now();
2924 :
2925 2262 : let gc_timelines = self
2926 2262 : .refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
2927 2169 : .await?;
2928 :
2929 2262 : failpoint_support::sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
2930 :
2931 : // If there is nothing to GC, we don't want any messages in the INFO log.
2932 2262 : if !gc_timelines.is_empty() {
2933 2262 : info!("{} timelines need GC", gc_timelines.len());
2934 : } else {
2935 0 : debug!("{} timelines need GC", gc_timelines.len());
2936 : }
2937 :
2938 : // Perform GC for each timeline.
2939 : //
2940 : // Note that we don't hold the `Tenant::gc_cs` lock here because we don't want to delay the
2941 : // branch creation task, which requires the GC lock. A GC iteration can run concurrently
2942 : // with branch creation.
2943 : //
2944 : // See comments in [`Tenant::branch_timeline`] for more information about why branch
2945 : // creation task can run concurrently with timeline's GC iteration.
2946 4524 : for timeline in gc_timelines {
2947 2262 : if cancel.is_cancelled() {
2948 : // We were requested to shut down. Stop and return with the progress we
2949 : // made.
2950 0 : break;
2951 2262 : }
2952 2262 : let result = match timeline.gc().await {
2953 : Err(GcError::TimelineCancelled) => {
2954 0 : if target_timeline_id.is_some() {
2955 : // If we were targetting this specific timeline, surface cancellation to caller
2956 0 : return Err(GcError::TimelineCancelled);
2957 : } else {
2958 : // A timeline may be shutting down independently of the tenant's lifecycle: we should
2959 : // skip past this and proceed to try GC on other timelines.
2960 0 : continue;
2961 : }
2962 : }
2963 2262 : r => r?,
2964 : };
2965 2262 : totals += result;
2966 : }
2967 :
2968 2262 : totals.elapsed = now.elapsed();
2969 2262 : Ok(totals)
2970 2262 : }
2971 :
2972 : /// Refreshes the Timeline::gc_info for all timelines, returning the
2973 : /// vector of timelines which have [`Timeline::get_last_record_lsn`] past
2974 : /// [`Tenant::get_gc_horizon`].
2975 : ///
2976 : /// This is usually executed as part of periodic gc, but can now be triggered more often.
2977 0 : pub(crate) async fn refresh_gc_info(
2978 0 : &self,
2979 0 : cancel: &CancellationToken,
2980 0 : ctx: &RequestContext,
2981 0 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
2982 0 : // since this method can now be called at different rates than the configured gc loop, it
2983 0 : // might be that these configuration values get applied faster than what it was previously,
2984 0 : // since these were only read from the gc task.
2985 0 : let horizon = self.get_gc_horizon();
2986 0 : let pitr = self.get_pitr_interval();
2987 0 :
2988 0 : // refresh all timelines
2989 0 : let target_timeline_id = None;
2990 0 :
2991 0 : self.refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
2992 0 : .await
2993 0 : }
2994 :
2995 : /// Populate all Timelines' `GcInfo` with information about their children. We do not set the
2996 : /// PITR cutoffs here, because that requires I/O: this is done later, before GC, by [`Self::refresh_gc_info_internal`]
2997 : ///
2998 : /// Subsequently, parent-child relationships are updated incrementally during timeline creation/deletion.
2999 0 : fn initialize_gc_info(
3000 0 : &self,
3001 0 : timelines: &std::sync::MutexGuard<HashMap<TimelineId, Arc<Timeline>>>,
3002 0 : ) {
3003 0 : // This function must be called before activation: after activation timeline create/delete operations
3004 0 : // might happen, and this function is not safe to run concurrently with those.
3005 0 : assert!(!self.is_active());
3006 :
3007 : // Scan all timelines. For each timeline, remember the timeline ID and
3008 : // the branch point where it was created.
3009 0 : let mut all_branchpoints: BTreeMap<TimelineId, Vec<(Lsn, TimelineId)>> = BTreeMap::new();
3010 0 : timelines.iter().for_each(|(timeline_id, timeline_entry)| {
3011 0 : if let Some(ancestor_timeline_id) = &timeline_entry.get_ancestor_timeline_id() {
3012 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
3013 0 : ancestor_children.push((timeline_entry.get_ancestor_lsn(), *timeline_id));
3014 0 : }
3015 0 : });
3016 0 :
3017 0 : // The number of bytes we always keep, irrespective of PITR: this is a constant across timelines
3018 0 : let horizon = self.get_gc_horizon();
3019 :
3020 : // Populate each timeline's GcInfo with information about its child branches
3021 0 : for timeline in timelines.values() {
3022 0 : let mut branchpoints: Vec<(Lsn, TimelineId)> = all_branchpoints
3023 0 : .remove(&timeline.timeline_id)
3024 0 : .unwrap_or_default();
3025 0 :
3026 0 : branchpoints.sort_by_key(|b| b.0);
3027 0 :
3028 0 : let mut target = timeline.gc_info.write().unwrap();
3029 0 :
3030 0 : target.retain_lsns = branchpoints;
3031 0 :
3032 0 : let space_cutoff = timeline
3033 0 : .get_last_record_lsn()
3034 0 : .checked_sub(horizon)
3035 0 : .unwrap_or(Lsn(0));
3036 0 :
3037 0 : target.cutoffs = GcCutoffs {
3038 0 : space: space_cutoff,
3039 0 : time: Lsn::INVALID,
3040 0 : };
3041 0 : }
3042 0 : }
3043 :
3044 2262 : async fn refresh_gc_info_internal(
3045 2262 : &self,
3046 2262 : target_timeline_id: Option<TimelineId>,
3047 2262 : horizon: u64,
3048 2262 : pitr: Duration,
3049 2262 : cancel: &CancellationToken,
3050 2262 : ctx: &RequestContext,
3051 2262 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
3052 2262 : // before taking the gc_cs lock, do the heavier weight finding of gc_cutoff points for
3053 2262 : // currently visible timelines.
3054 2262 : let timelines = self
3055 2262 : .timelines
3056 2262 : .lock()
3057 2262 : .unwrap()
3058 2262 : .values()
3059 9930 : .filter(|tl| match target_timeline_id.as_ref() {
3060 9930 : Some(target) => &tl.timeline_id == target,
3061 0 : None => true,
3062 9930 : })
3063 2262 : .cloned()
3064 2262 : .collect::<Vec<_>>();
3065 2262 :
3066 2262 : if target_timeline_id.is_some() && timelines.is_empty() {
3067 : // We were to act on a particular timeline and it wasn't found
3068 0 : return Err(GcError::TimelineNotFound);
3069 2262 : }
3070 2262 :
3071 2262 : let mut gc_cutoffs: HashMap<TimelineId, GcCutoffs> =
3072 2262 : HashMap::with_capacity(timelines.len());
3073 :
3074 2262 : for timeline in timelines.iter() {
3075 2262 : let cutoff = timeline
3076 2262 : .get_last_record_lsn()
3077 2262 : .checked_sub(horizon)
3078 2262 : .unwrap_or(Lsn(0));
3079 :
3080 2262 : let cutoffs = timeline.find_gc_cutoffs(cutoff, pitr, cancel, ctx).await?;
3081 2262 : let old = gc_cutoffs.insert(timeline.timeline_id, cutoffs);
3082 2262 : assert!(old.is_none());
3083 : }
3084 :
3085 2262 : if !self.is_active() || self.cancel.is_cancelled() {
3086 0 : return Err(GcError::TenantCancelled);
3087 2262 : }
3088 :
3089 : // grab mutex to prevent new timelines from being created here; avoid doing long operations
3090 : // because that will stall branch creation.
3091 2262 : let gc_cs = self.gc_cs.lock().await;
3092 :
3093 : // Ok, we now know all the branch points.
3094 : // Update the GC information for each timeline.
3095 2262 : let mut gc_timelines = Vec::with_capacity(timelines.len());
3096 4524 : for timeline in timelines {
3097 : // We filtered the timeline list above
3098 2262 : if let Some(target_timeline_id) = target_timeline_id {
3099 2262 : assert_eq!(target_timeline_id, timeline.timeline_id);
3100 0 : }
3101 :
3102 : {
3103 2262 : let mut target = timeline.gc_info.write().unwrap();
3104 2262 :
3105 2262 : // Cull any expired leases
3106 2262 : let now = SystemTime::now();
3107 2262 : target.leases.retain(|_, lease| !lease.is_expired(&now));
3108 2262 :
3109 2262 : timeline
3110 2262 : .metrics
3111 2262 : .valid_lsn_lease_count_gauge
3112 2262 : .set(target.leases.len() as u64);
3113 :
3114 : // Look up parent's PITR cutoff to update the child's knowledge of whether it is within parent's PITR
3115 2262 : if let Some(ancestor_id) = timeline.get_ancestor_timeline_id() {
3116 300 : if let Some(ancestor_gc_cutoffs) = gc_cutoffs.get(&ancestor_id) {
3117 0 : target.within_ancestor_pitr =
3118 0 : timeline.get_ancestor_lsn() >= ancestor_gc_cutoffs.time;
3119 300 : }
3120 1962 : }
3121 :
3122 : // Update metrics that depend on GC state
3123 2262 : timeline
3124 2262 : .metrics
3125 2262 : .archival_size
3126 2262 : .set(if target.within_ancestor_pitr {
3127 0 : timeline.metrics.current_logical_size_gauge.get()
3128 : } else {
3129 2262 : 0
3130 : });
3131 2262 : timeline.metrics.pitr_history_size.set(
3132 2262 : timeline
3133 2262 : .get_last_record_lsn()
3134 2262 : .checked_sub(target.cutoffs.time)
3135 2262 : .unwrap_or(Lsn(0))
3136 2262 : .0,
3137 2262 : );
3138 :
3139 : // Apply the cutoffs we found to the Timeline's GcInfo. Why might we _not_ have cutoffs for a timeline?
3140 : // - this timeline was created while we were finding cutoffs
3141 : // - lsn for timestamp search fails for this timeline repeatedly
3142 2262 : if let Some(cutoffs) = gc_cutoffs.get(&timeline.timeline_id) {
3143 2262 : target.cutoffs = cutoffs.clone();
3144 2262 : }
3145 : }
3146 :
3147 2262 : gc_timelines.push(timeline);
3148 : }
3149 2262 : drop(gc_cs);
3150 2262 : Ok(gc_timelines)
3151 2262 : }
3152 :
3153 : /// A substitute for `branch_timeline` for use in unit tests.
3154 : /// The returned timeline will have state value `Active` to make various `anyhow::ensure!()`
3155 : /// calls pass, but, we do not actually call `.activate()` under the hood. So, none of the
3156 : /// timeline background tasks are launched, except the flush loop.
3157 : #[cfg(test)]
3158 690 : async fn branch_timeline_test(
3159 690 : &self,
3160 690 : src_timeline: &Arc<Timeline>,
3161 690 : dst_id: TimelineId,
3162 690 : ancestor_lsn: Option<Lsn>,
3163 690 : ctx: &RequestContext,
3164 690 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
3165 690 : let create_guard = self.create_timeline_create_guard(dst_id).unwrap();
3166 690 : let tl = self
3167 690 : .branch_timeline_impl(src_timeline, dst_id, ancestor_lsn, create_guard, ctx)
3168 12 : .await?;
3169 678 : tl.set_state(TimelineState::Active);
3170 678 : Ok(tl)
3171 690 : }
3172 :
3173 : /// Helper for unit tests to branch a timeline with some pre-loaded states.
3174 : #[cfg(test)]
3175 : #[allow(clippy::too_many_arguments)]
3176 18 : pub async fn branch_timeline_test_with_layers(
3177 18 : &self,
3178 18 : src_timeline: &Arc<Timeline>,
3179 18 : dst_id: TimelineId,
3180 18 : ancestor_lsn: Option<Lsn>,
3181 18 : ctx: &RequestContext,
3182 18 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
3183 18 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
3184 18 : end_lsn: Lsn,
3185 18 : ) -> anyhow::Result<Arc<Timeline>> {
3186 18 : let tline = self
3187 18 : .branch_timeline_test(src_timeline, dst_id, ancestor_lsn, ctx)
3188 0 : .await?;
3189 18 : let ancestor_lsn = if let Some(ancestor_lsn) = ancestor_lsn {
3190 18 : ancestor_lsn
3191 : } else {
3192 0 : tline.get_last_record_lsn()
3193 : };
3194 18 : assert!(end_lsn >= ancestor_lsn);
3195 18 : tline.force_advance_lsn(end_lsn);
3196 36 : for deltas in delta_layer_desc {
3197 18 : tline
3198 18 : .force_create_delta_layer(deltas, Some(ancestor_lsn), ctx)
3199 54 : .await?;
3200 : }
3201 30 : for (lsn, images) in image_layer_desc {
3202 12 : tline
3203 12 : .force_create_image_layer(lsn, images, Some(ancestor_lsn), ctx)
3204 42 : .await?;
3205 : }
3206 18 : Ok(tline)
3207 18 : }
3208 :
3209 : /// Branch an existing timeline.
3210 : ///
3211 : /// The caller is responsible for activating the returned timeline.
3212 0 : async fn branch_timeline(
3213 0 : &self,
3214 0 : src_timeline: &Arc<Timeline>,
3215 0 : dst_id: TimelineId,
3216 0 : start_lsn: Option<Lsn>,
3217 0 : timeline_create_guard: TimelineCreateGuard<'_>,
3218 0 : ctx: &RequestContext,
3219 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
3220 0 : self.branch_timeline_impl(src_timeline, dst_id, start_lsn, timeline_create_guard, ctx)
3221 0 : .await
3222 0 : }
3223 :
3224 690 : async fn branch_timeline_impl(
3225 690 : &self,
3226 690 : src_timeline: &Arc<Timeline>,
3227 690 : dst_id: TimelineId,
3228 690 : start_lsn: Option<Lsn>,
3229 690 : timeline_create_guard: TimelineCreateGuard<'_>,
3230 690 : _ctx: &RequestContext,
3231 690 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
3232 690 : let src_id = src_timeline.timeline_id;
3233 :
3234 : // We will validate our ancestor LSN in this function. Acquire the GC lock so that
3235 : // this check cannot race with GC, and the ancestor LSN is guaranteed to remain
3236 : // valid while we are creating the branch.
3237 690 : let _gc_cs = self.gc_cs.lock().await;
3238 :
3239 : // If no start LSN is specified, we branch the new timeline from the source timeline's last record LSN
3240 690 : let start_lsn = start_lsn.unwrap_or_else(|| {
3241 6 : let lsn = src_timeline.get_last_record_lsn();
3242 6 : info!("branching timeline {dst_id} from timeline {src_id} at last record LSN: {lsn}");
3243 6 : lsn
3244 690 : });
3245 690 :
3246 690 : // Ensure that `start_lsn` is valid, i.e. the LSN is within the PITR
3247 690 : // horizon on the source timeline
3248 690 : //
3249 690 : // We check it against both the planned GC cutoff stored in 'gc_info',
3250 690 : // and the 'latest_gc_cutoff' of the last GC that was performed. The
3251 690 : // planned GC cutoff in 'gc_info' is normally larger than
3252 690 : // 'latest_gc_cutoff_lsn', but beware of corner cases like if you just
3253 690 : // changed the GC settings for the tenant to make the PITR window
3254 690 : // larger, but some of the data was already removed by an earlier GC
3255 690 : // iteration.
3256 690 :
3257 690 : // check against last actual 'latest_gc_cutoff' first
3258 690 : let latest_gc_cutoff_lsn = src_timeline.get_latest_gc_cutoff_lsn();
3259 690 : src_timeline
3260 690 : .check_lsn_is_in_scope(start_lsn, &latest_gc_cutoff_lsn)
3261 690 : .context(format!(
3262 690 : "invalid branch start lsn: less than latest GC cutoff {}",
3263 690 : *latest_gc_cutoff_lsn,
3264 690 : ))
3265 690 : .map_err(CreateTimelineError::AncestorLsn)?;
3266 :
3267 : // and then the planned GC cutoff
3268 : {
3269 678 : let gc_info = src_timeline.gc_info.read().unwrap();
3270 678 : let cutoff = gc_info.min_cutoff();
3271 678 : if start_lsn < cutoff {
3272 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
3273 0 : "invalid branch start lsn: less than planned GC cutoff {cutoff}"
3274 0 : )));
3275 678 : }
3276 678 : }
3277 678 :
3278 678 : //
3279 678 : // The branch point is valid, and we are still holding the 'gc_cs' lock
3280 678 : // so that GC cannot advance the GC cutoff until we are finished.
3281 678 : // Proceed with the branch creation.
3282 678 : //
3283 678 :
3284 678 : // Determine prev-LSN for the new timeline. We can only determine it if
3285 678 : // the timeline was branched at the current end of the source timeline.
3286 678 : let RecordLsn {
3287 678 : last: src_last,
3288 678 : prev: src_prev,
3289 678 : } = src_timeline.get_last_record_rlsn();
3290 678 : let dst_prev = if src_last == start_lsn {
3291 648 : Some(src_prev)
3292 : } else {
3293 30 : None
3294 : };
3295 :
3296 : // Create the metadata file, noting the ancestor of the new timeline.
3297 : // There is initially no data in it, but all the read-calls know to look
3298 : // into the ancestor.
3299 678 : let metadata = TimelineMetadata::new(
3300 678 : start_lsn,
3301 678 : dst_prev,
3302 678 : Some(src_id),
3303 678 : start_lsn,
3304 678 : *src_timeline.latest_gc_cutoff_lsn.read(), // FIXME: should we hold onto this guard longer?
3305 678 : src_timeline.initdb_lsn,
3306 678 : src_timeline.pg_version,
3307 678 : );
3308 :
3309 678 : let uninitialized_timeline = self
3310 678 : .prepare_new_timeline(
3311 678 : dst_id,
3312 678 : &metadata,
3313 678 : timeline_create_guard,
3314 678 : start_lsn + 1,
3315 678 : Some(Arc::clone(src_timeline)),
3316 678 : src_timeline.last_aux_file_policy.load(),
3317 678 : )
3318 0 : .await?;
3319 :
3320 678 : let new_timeline = uninitialized_timeline.finish_creation()?;
3321 :
3322 : // Root timeline gets its layers during creation and uploads them along with the metadata.
3323 : // A branch timeline though, when created, can get no writes for some time, hence won't get any layers created.
3324 : // We still need to upload its metadata eagerly: if other nodes `attach` the tenant and miss this timeline, their GC
3325 : // could get incorrect information and remove more layers, than needed.
3326 : // See also https://github.com/neondatabase/neon/issues/3865
3327 678 : new_timeline
3328 678 : .remote_client
3329 678 : .schedule_index_upload_for_full_metadata_update(&metadata)
3330 678 : .context("branch initial metadata upload")?;
3331 :
3332 678 : Ok(new_timeline)
3333 690 : }
3334 :
3335 : /// For unit tests, make this visible so that other modules can directly create timelines
3336 : #[cfg(test)]
3337 12 : #[tracing::instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
3338 : pub(crate) async fn bootstrap_timeline_test(
3339 : &self,
3340 : timeline_id: TimelineId,
3341 : pg_version: u32,
3342 : load_existing_initdb: Option<TimelineId>,
3343 : ctx: &RequestContext,
3344 : ) -> anyhow::Result<Arc<Timeline>> {
3345 : let create_guard = self.create_timeline_create_guard(timeline_id).unwrap();
3346 : self.bootstrap_timeline(
3347 : timeline_id,
3348 : pg_version,
3349 : load_existing_initdb,
3350 : create_guard,
3351 : ctx,
3352 : )
3353 : .await
3354 : }
3355 :
3356 0 : async fn upload_initdb(
3357 0 : &self,
3358 0 : timelines_path: &Utf8PathBuf,
3359 0 : pgdata_path: &Utf8PathBuf,
3360 0 : timeline_id: &TimelineId,
3361 0 : ) -> anyhow::Result<()> {
3362 0 : let temp_path = timelines_path.join(format!(
3363 0 : "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
3364 0 : ));
3365 :
3366 : scopeguard::defer! {
3367 : if let Err(e) = fs::remove_file(&temp_path) {
3368 : error!("Failed to remove temporary initdb archive '{temp_path}': {e}");
3369 : }
3370 : }
3371 :
3372 0 : let (pgdata_zstd, tar_zst_size) = create_zst_tarball(pgdata_path, &temp_path).await?;
3373 : const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
3374 0 : if tar_zst_size > INITDB_TAR_ZST_WARN_LIMIT {
3375 0 : warn!(
3376 0 : "compressed {temp_path} size of {tar_zst_size} is above limit {INITDB_TAR_ZST_WARN_LIMIT}."
3377 : );
3378 0 : }
3379 :
3380 : pausable_failpoint!("before-initdb-upload");
3381 :
3382 0 : backoff::retry(
3383 0 : || async {
3384 0 : self::remote_timeline_client::upload_initdb_dir(
3385 0 : &self.remote_storage,
3386 0 : &self.tenant_shard_id.tenant_id,
3387 0 : timeline_id,
3388 0 : pgdata_zstd.try_clone().await?,
3389 0 : tar_zst_size,
3390 0 : &self.cancel,
3391 0 : )
3392 0 : .await
3393 0 : },
3394 0 : |_| false,
3395 0 : 3,
3396 0 : u32::MAX,
3397 0 : "persist_initdb_tar_zst",
3398 0 : &self.cancel,
3399 0 : )
3400 0 : .await
3401 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
3402 0 : .and_then(|x| x)
3403 0 : }
3404 :
3405 : /// - run initdb to init temporary instance and get bootstrap data
3406 : /// - after initialization completes, tar up the temp dir and upload it to S3.
3407 : ///
3408 : /// The caller is responsible for activating the returned timeline.
3409 6 : async fn bootstrap_timeline(
3410 6 : &self,
3411 6 : timeline_id: TimelineId,
3412 6 : pg_version: u32,
3413 6 : load_existing_initdb: Option<TimelineId>,
3414 6 : timeline_create_guard: TimelineCreateGuard<'_>,
3415 6 : ctx: &RequestContext,
3416 6 : ) -> anyhow::Result<Arc<Timeline>> {
3417 6 : // create a `tenant/{tenant_id}/timelines/basebackup-{timeline_id}.{TEMP_FILE_SUFFIX}/`
3418 6 : // temporary directory for basebackup files for the given timeline.
3419 6 :
3420 6 : let timelines_path = self.conf.timelines_path(&self.tenant_shard_id);
3421 6 : let pgdata_path = path_with_suffix_extension(
3422 6 : timelines_path.join(format!("basebackup-{timeline_id}")),
3423 6 : TEMP_FILE_SUFFIX,
3424 6 : );
3425 6 :
3426 6 : // Remove whatever was left from the previous runs: safe because TimelineCreateGuard guarantees
3427 6 : // we won't race with other creations or existent timelines with the same path.
3428 6 : if pgdata_path.exists() {
3429 0 : fs::remove_dir_all(&pgdata_path).with_context(|| {
3430 0 : format!("Failed to remove already existing initdb directory: {pgdata_path}")
3431 0 : })?;
3432 6 : }
3433 :
3434 : // this new directory is very temporary, set to remove it immediately after bootstrap, we don't need it
3435 : scopeguard::defer! {
3436 : if let Err(e) = fs::remove_dir_all(&pgdata_path) {
3437 : // this is unlikely, but we will remove the directory on pageserver restart or another bootstrap call
3438 : error!("Failed to remove temporary initdb directory '{pgdata_path}': {e}");
3439 : }
3440 : }
3441 6 : if let Some(existing_initdb_timeline_id) = load_existing_initdb {
3442 6 : if existing_initdb_timeline_id != timeline_id {
3443 0 : let source_path = &remote_initdb_archive_path(
3444 0 : &self.tenant_shard_id.tenant_id,
3445 0 : &existing_initdb_timeline_id,
3446 0 : );
3447 0 : let dest_path =
3448 0 : &remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &timeline_id);
3449 0 :
3450 0 : // if this fails, it will get retried by retried control plane requests
3451 0 : self.remote_storage
3452 0 : .copy_object(source_path, dest_path, &self.cancel)
3453 0 : .await
3454 0 : .context("copy initdb tar")?;
3455 6 : }
3456 6 : let (initdb_tar_zst_path, initdb_tar_zst) =
3457 6 : self::remote_timeline_client::download_initdb_tar_zst(
3458 6 : self.conf,
3459 6 : &self.remote_storage,
3460 6 : &self.tenant_shard_id,
3461 6 : &existing_initdb_timeline_id,
3462 6 : &self.cancel,
3463 6 : )
3464 1748 : .await
3465 6 : .context("download initdb tar")?;
3466 :
3467 : scopeguard::defer! {
3468 : if let Err(e) = fs::remove_file(&initdb_tar_zst_path) {
3469 : error!("Failed to remove temporary initdb archive '{initdb_tar_zst_path}': {e}");
3470 : }
3471 : }
3472 :
3473 6 : let buf_read =
3474 6 : BufReader::with_capacity(remote_timeline_client::BUFFER_SIZE, initdb_tar_zst);
3475 6 : extract_zst_tarball(&pgdata_path, buf_read)
3476 28777 : .await
3477 6 : .context("extract initdb tar")?;
3478 : } else {
3479 : // Init temporarily repo to get bootstrap data, this creates a directory in the `pgdata_path` path
3480 0 : run_initdb(self.conf, &pgdata_path, pg_version, &self.cancel).await?;
3481 :
3482 : // Upload the created data dir to S3
3483 0 : if self.tenant_shard_id().is_shard_zero() {
3484 0 : self.upload_initdb(&timelines_path, &pgdata_path, &timeline_id)
3485 0 : .await?;
3486 0 : }
3487 : }
3488 6 : let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
3489 6 :
3490 6 : // Import the contents of the data directory at the initial checkpoint
3491 6 : // LSN, and any WAL after that.
3492 6 : // Initdb lsn will be equal to last_record_lsn which will be set after import.
3493 6 : // Because we know it upfront avoid having an option or dummy zero value by passing it to the metadata.
3494 6 : let new_metadata = TimelineMetadata::new(
3495 6 : Lsn(0),
3496 6 : None,
3497 6 : None,
3498 6 : Lsn(0),
3499 6 : pgdata_lsn,
3500 6 : pgdata_lsn,
3501 6 : pg_version,
3502 6 : );
3503 6 : let raw_timeline = self
3504 6 : .prepare_new_timeline(
3505 6 : timeline_id,
3506 6 : &new_metadata,
3507 6 : timeline_create_guard,
3508 6 : pgdata_lsn,
3509 6 : None,
3510 6 : None,
3511 6 : )
3512 0 : .await?;
3513 :
3514 6 : let tenant_shard_id = raw_timeline.owning_tenant.tenant_shard_id;
3515 6 : let unfinished_timeline = raw_timeline.raw_timeline()?;
3516 :
3517 : // Flush the new layer files to disk, before we make the timeline as available to
3518 : // the outside world.
3519 : //
3520 : // Flush loop needs to be spawned in order to be able to flush.
3521 6 : unfinished_timeline.maybe_spawn_flush_loop();
3522 6 :
3523 6 : import_datadir::import_timeline_from_postgres_datadir(
3524 6 : unfinished_timeline,
3525 6 : &pgdata_path,
3526 6 : pgdata_lsn,
3527 6 : ctx,
3528 6 : )
3529 26243 : .await
3530 6 : .with_context(|| {
3531 0 : format!("Failed to import pgdatadir for timeline {tenant_shard_id}/{timeline_id}")
3532 6 : })?;
3533 :
3534 6 : fail::fail_point!("before-checkpoint-new-timeline", |_| {
3535 0 : anyhow::bail!("failpoint before-checkpoint-new-timeline");
3536 6 : });
3537 :
3538 6 : unfinished_timeline
3539 6 : .freeze_and_flush()
3540 6 : .await
3541 6 : .with_context(|| {
3542 0 : format!(
3543 0 : "Failed to flush after pgdatadir import for timeline {tenant_shard_id}/{timeline_id}"
3544 0 : )
3545 6 : })?;
3546 :
3547 : // All done!
3548 6 : let timeline = raw_timeline.finish_creation()?;
3549 :
3550 6 : Ok(timeline)
3551 6 : }
3552 :
3553 : /// Call this before constructing a timeline, to build its required structures
3554 1218 : fn build_timeline_resources(&self, timeline_id: TimelineId) -> TimelineResources {
3555 1218 : let remote_client = RemoteTimelineClient::new(
3556 1218 : self.remote_storage.clone(),
3557 1218 : self.deletion_queue_client.clone(),
3558 1218 : self.conf,
3559 1218 : self.tenant_shard_id,
3560 1218 : timeline_id,
3561 1218 : self.generation,
3562 1218 : );
3563 1218 : TimelineResources {
3564 1218 : remote_client,
3565 1218 : timeline_get_throttle: self.timeline_get_throttle.clone(),
3566 1218 : l0_flush_global_state: self.l0_flush_global_state.clone(),
3567 1218 : }
3568 1218 : }
3569 :
3570 : /// Creates intermediate timeline structure and its files.
3571 : ///
3572 : /// An empty layer map is initialized, and new data and WAL can be imported starting
3573 : /// at 'disk_consistent_lsn'. After any initial data has been imported, call
3574 : /// `finish_creation` to insert the Timeline into the timelines map.
3575 1218 : async fn prepare_new_timeline<'a>(
3576 1218 : &'a self,
3577 1218 : new_timeline_id: TimelineId,
3578 1218 : new_metadata: &TimelineMetadata,
3579 1218 : create_guard: TimelineCreateGuard<'a>,
3580 1218 : start_lsn: Lsn,
3581 1218 : ancestor: Option<Arc<Timeline>>,
3582 1218 : last_aux_file_policy: Option<AuxFilePolicy>,
3583 1218 : ) -> anyhow::Result<UninitializedTimeline> {
3584 1218 : let tenant_shard_id = self.tenant_shard_id;
3585 1218 :
3586 1218 : let resources = self.build_timeline_resources(new_timeline_id);
3587 1218 : resources
3588 1218 : .remote_client
3589 1218 : .init_upload_queue_for_empty_remote(new_metadata)?;
3590 :
3591 1218 : let timeline_struct = self
3592 1218 : .create_timeline_struct(
3593 1218 : new_timeline_id,
3594 1218 : new_metadata,
3595 1218 : ancestor,
3596 1218 : resources,
3597 1218 : CreateTimelineCause::Load,
3598 1218 : last_aux_file_policy,
3599 1218 : )
3600 1218 : .context("Failed to create timeline data structure")?;
3601 :
3602 1218 : timeline_struct.init_empty_layer_map(start_lsn);
3603 :
3604 1218 : if let Err(e) = self
3605 1218 : .create_timeline_files(&create_guard.timeline_path)
3606 0 : .await
3607 : {
3608 0 : error!("Failed to create initial files for timeline {tenant_shard_id}/{new_timeline_id}, cleaning up: {e:?}");
3609 0 : cleanup_timeline_directory(create_guard);
3610 0 : return Err(e);
3611 1218 : }
3612 1218 :
3613 1218 : debug!(
3614 0 : "Successfully created initial files for timeline {tenant_shard_id}/{new_timeline_id}"
3615 : );
3616 :
3617 1218 : Ok(UninitializedTimeline::new(
3618 1218 : self,
3619 1218 : new_timeline_id,
3620 1218 : Some((timeline_struct, create_guard)),
3621 1218 : ))
3622 1218 : }
3623 :
3624 1218 : async fn create_timeline_files(&self, timeline_path: &Utf8Path) -> anyhow::Result<()> {
3625 1218 : crashsafe::create_dir(timeline_path).context("Failed to create timeline directory")?;
3626 :
3627 1218 : fail::fail_point!("after-timeline-dir-creation", |_| {
3628 0 : anyhow::bail!("failpoint after-timeline-dir-creation");
3629 1218 : });
3630 :
3631 1218 : Ok(())
3632 1218 : }
3633 :
3634 : /// Get a guard that provides exclusive access to the timeline directory, preventing
3635 : /// concurrent attempts to create the same timeline.
3636 1236 : fn create_timeline_create_guard(
3637 1236 : &self,
3638 1236 : timeline_id: TimelineId,
3639 1236 : ) -> Result<TimelineCreateGuard, TimelineExclusionError> {
3640 1236 : let tenant_shard_id = self.tenant_shard_id;
3641 1236 :
3642 1236 : let timeline_path = self.conf.timeline_path(&tenant_shard_id, &timeline_id);
3643 :
3644 1236 : let create_guard = TimelineCreateGuard::new(self, timeline_id, timeline_path.clone())?;
3645 :
3646 : // At this stage, we have got exclusive access to in-memory state for this timeline ID
3647 : // for creation.
3648 : // A timeline directory should never exist on disk already:
3649 : // - a previous failed creation would have cleaned up after itself
3650 : // - a pageserver restart would clean up timeline directories that don't have valid remote state
3651 : //
3652 : // Therefore it is an unexpected internal error to encounter a timeline directory already existing here,
3653 : // this error may indicate a bug in cleanup on failed creations.
3654 1230 : if timeline_path.exists() {
3655 0 : return Err(TimelineExclusionError::Other(anyhow::anyhow!(
3656 0 : "Timeline directory already exists! This is a bug."
3657 0 : )));
3658 1230 : }
3659 1230 :
3660 1230 : Ok(create_guard)
3661 1236 : }
3662 :
3663 : /// Gathers inputs from all of the timelines to produce a sizing model input.
3664 : ///
3665 : /// Future is cancellation safe. Only one calculation can be running at once per tenant.
3666 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
3667 : pub async fn gather_size_inputs(
3668 : &self,
3669 : // `max_retention_period` overrides the cutoff that is used to calculate the size
3670 : // (only if it is shorter than the real cutoff).
3671 : max_retention_period: Option<u64>,
3672 : cause: LogicalSizeCalculationCause,
3673 : cancel: &CancellationToken,
3674 : ctx: &RequestContext,
3675 : ) -> Result<size::ModelInputs, size::CalculateSyntheticSizeError> {
3676 : let logical_sizes_at_once = self
3677 : .conf
3678 : .concurrent_tenant_size_logical_size_queries
3679 : .inner();
3680 :
3681 : // TODO: Having a single mutex block concurrent reads is not great for performance.
3682 : //
3683 : // But the only case where we need to run multiple of these at once is when we
3684 : // request a size for a tenant manually via API, while another background calculation
3685 : // is in progress (which is not a common case).
3686 : //
3687 : // See more for on the issue #2748 condenced out of the initial PR review.
3688 : let mut shared_cache = tokio::select! {
3689 : locked = self.cached_logical_sizes.lock() => locked,
3690 : _ = cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
3691 : _ = self.cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
3692 : };
3693 :
3694 : size::gather_inputs(
3695 : self,
3696 : logical_sizes_at_once,
3697 : max_retention_period,
3698 : &mut shared_cache,
3699 : cause,
3700 : cancel,
3701 : ctx,
3702 : )
3703 : .await
3704 : }
3705 :
3706 : /// Calculate synthetic tenant size and cache the result.
3707 : /// This is periodically called by background worker.
3708 : /// result is cached in tenant struct
3709 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
3710 : pub async fn calculate_synthetic_size(
3711 : &self,
3712 : cause: LogicalSizeCalculationCause,
3713 : cancel: &CancellationToken,
3714 : ctx: &RequestContext,
3715 : ) -> Result<u64, size::CalculateSyntheticSizeError> {
3716 : let inputs = self.gather_size_inputs(None, cause, cancel, ctx).await?;
3717 :
3718 : let size = inputs.calculate();
3719 :
3720 : self.set_cached_synthetic_size(size);
3721 :
3722 : Ok(size)
3723 : }
3724 :
3725 : /// Cache given synthetic size and update the metric value
3726 0 : pub fn set_cached_synthetic_size(&self, size: u64) {
3727 0 : self.cached_synthetic_tenant_size
3728 0 : .store(size, Ordering::Relaxed);
3729 0 :
3730 0 : // Only shard zero should be calculating synthetic sizes
3731 0 : debug_assert!(self.shard_identity.is_shard_zero());
3732 :
3733 0 : TENANT_SYNTHETIC_SIZE_METRIC
3734 0 : .get_metric_with_label_values(&[&self.tenant_shard_id.tenant_id.to_string()])
3735 0 : .unwrap()
3736 0 : .set(size);
3737 0 : }
3738 :
3739 0 : pub fn cached_synthetic_size(&self) -> u64 {
3740 0 : self.cached_synthetic_tenant_size.load(Ordering::Relaxed)
3741 0 : }
3742 :
3743 : /// Flush any in-progress layers, schedule uploads, and wait for uploads to complete.
3744 : ///
3745 : /// This function can take a long time: callers should wrap it in a timeout if calling
3746 : /// from an external API handler.
3747 : ///
3748 : /// Cancel-safety: cancelling this function may leave I/O running, but such I/O is
3749 : /// still bounded by tenant/timeline shutdown.
3750 0 : #[tracing::instrument(skip_all)]
3751 : pub(crate) async fn flush_remote(&self) -> anyhow::Result<()> {
3752 : let timelines = self.timelines.lock().unwrap().clone();
3753 :
3754 0 : async fn flush_timeline(_gate: GateGuard, timeline: Arc<Timeline>) -> anyhow::Result<()> {
3755 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Flushing...");
3756 0 : timeline.freeze_and_flush().await?;
3757 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Waiting for uploads...");
3758 0 : timeline.remote_client.wait_completion().await?;
3759 :
3760 0 : Ok(())
3761 0 : }
3762 :
3763 : // We do not use a JoinSet for these tasks, because we don't want them to be
3764 : // aborted when this function's future is cancelled: they should stay alive
3765 : // holding their GateGuard until they complete, to ensure their I/Os complete
3766 : // before Timeline shutdown completes.
3767 : let mut results = FuturesUnordered::new();
3768 :
3769 : for (_timeline_id, timeline) in timelines {
3770 : // Run each timeline's flush in a task holding the timeline's gate: this
3771 : // means that if this function's future is cancelled, the Timeline shutdown
3772 : // will still wait for any I/O in here to complete.
3773 : let Ok(gate) = timeline.gate.enter() else {
3774 : continue;
3775 : };
3776 0 : let jh = tokio::task::spawn(async move { flush_timeline(gate, timeline).await });
3777 : results.push(jh);
3778 : }
3779 :
3780 : while let Some(r) = results.next().await {
3781 : if let Err(e) = r {
3782 : if !e.is_cancelled() && !e.is_panic() {
3783 : tracing::error!("unexpected join error: {e:?}");
3784 : }
3785 : }
3786 : }
3787 :
3788 : // The flushes we did above were just writes, but the Tenant might have had
3789 : // pending deletions as well from recent compaction/gc: we want to flush those
3790 : // as well. This requires flushing the global delete queue. This is cheap
3791 : // because it's typically a no-op.
3792 : match self.deletion_queue_client.flush_execute().await {
3793 : Ok(_) => {}
3794 : Err(DeletionQueueError::ShuttingDown) => {}
3795 : }
3796 :
3797 : Ok(())
3798 : }
3799 :
3800 0 : pub(crate) fn get_tenant_conf(&self) -> TenantConfOpt {
3801 0 : self.tenant_conf.load().tenant_conf.clone()
3802 0 : }
3803 :
3804 : /// How much local storage would this tenant like to have? It can cope with
3805 : /// less than this (via eviction and on-demand downloads), but this function enables
3806 : /// the Tenant to advertise how much storage it would prefer to have to provide fast I/O
3807 : /// by keeping important things on local disk.
3808 : ///
3809 : /// This is a heuristic, not a guarantee: tenants that are long-idle will actually use less
3810 : /// than they report here, due to layer eviction. Tenants with many active branches may
3811 : /// actually use more than they report here.
3812 0 : pub(crate) fn local_storage_wanted(&self) -> u64 {
3813 0 : let timelines = self.timelines.lock().unwrap();
3814 0 :
3815 0 : // Heuristic: we use the max() of the timelines' visible sizes, rather than the sum. This
3816 0 : // reflects the observation that on tenants with multiple large branches, typically only one
3817 0 : // of them is used actively enough to occupy space on disk.
3818 0 : timelines
3819 0 : .values()
3820 0 : .map(|t| t.metrics.visible_physical_size_gauge.get())
3821 0 : .max()
3822 0 : .unwrap_or(0)
3823 0 : }
3824 : }
3825 :
3826 : /// Create the cluster temporarily in 'initdbpath' directory inside the repository
3827 : /// to get bootstrap data for timeline initialization.
3828 0 : async fn run_initdb(
3829 0 : conf: &'static PageServerConf,
3830 0 : initdb_target_dir: &Utf8Path,
3831 0 : pg_version: u32,
3832 0 : cancel: &CancellationToken,
3833 0 : ) -> Result<(), InitdbError> {
3834 0 : let initdb_bin_path = conf
3835 0 : .pg_bin_dir(pg_version)
3836 0 : .map_err(InitdbError::Other)?
3837 0 : .join("initdb");
3838 0 : let initdb_lib_dir = conf.pg_lib_dir(pg_version).map_err(InitdbError::Other)?;
3839 0 : info!(
3840 0 : "running {} in {}, libdir: {}",
3841 : initdb_bin_path, initdb_target_dir, initdb_lib_dir,
3842 : );
3843 :
3844 0 : let _permit = INIT_DB_SEMAPHORE.acquire().await;
3845 :
3846 0 : let initdb_command = tokio::process::Command::new(&initdb_bin_path)
3847 0 : .args(["-D", initdb_target_dir.as_ref()])
3848 0 : .args(["-U", &conf.superuser])
3849 0 : .args(["-E", "utf8"])
3850 0 : .arg("--no-instructions")
3851 0 : .arg("--no-sync")
3852 0 : .env_clear()
3853 0 : .env("LD_LIBRARY_PATH", &initdb_lib_dir)
3854 0 : .env("DYLD_LIBRARY_PATH", &initdb_lib_dir)
3855 0 : .stdin(std::process::Stdio::null())
3856 0 : // stdout invocation produces the same output every time, we don't need it
3857 0 : .stdout(std::process::Stdio::null())
3858 0 : // we would be interested in the stderr output, if there was any
3859 0 : .stderr(std::process::Stdio::piped())
3860 0 : .spawn()?;
3861 :
3862 : // Ideally we'd select here with the cancellation token, but the problem is that
3863 : // we can't safely terminate initdb: it launches processes of its own, and killing
3864 : // initdb doesn't kill them. After we return from this function, we want the target
3865 : // directory to be able to be cleaned up.
3866 : // See https://github.com/neondatabase/neon/issues/6385
3867 0 : let initdb_output = initdb_command.wait_with_output().await?;
3868 0 : if !initdb_output.status.success() {
3869 0 : return Err(InitdbError::Failed(
3870 0 : initdb_output.status,
3871 0 : initdb_output.stderr,
3872 0 : ));
3873 0 : }
3874 0 :
3875 0 : // This isn't true cancellation support, see above. Still return an error to
3876 0 : // excercise the cancellation code path.
3877 0 : if cancel.is_cancelled() {
3878 0 : return Err(InitdbError::Cancelled);
3879 0 : }
3880 0 :
3881 0 : Ok(())
3882 0 : }
3883 :
3884 : /// Dump contents of a layer file to stdout.
3885 0 : pub async fn dump_layerfile_from_path(
3886 0 : path: &Utf8Path,
3887 0 : verbose: bool,
3888 0 : ctx: &RequestContext,
3889 0 : ) -> anyhow::Result<()> {
3890 : use std::os::unix::fs::FileExt;
3891 :
3892 : // All layer files start with a two-byte "magic" value, to identify the kind of
3893 : // file.
3894 0 : let file = File::open(path)?;
3895 0 : let mut header_buf = [0u8; 2];
3896 0 : file.read_exact_at(&mut header_buf, 0)?;
3897 :
3898 0 : match u16::from_be_bytes(header_buf) {
3899 : crate::IMAGE_FILE_MAGIC => {
3900 0 : ImageLayer::new_for_path(path, file)?
3901 0 : .dump(verbose, ctx)
3902 0 : .await?
3903 : }
3904 : crate::DELTA_FILE_MAGIC => {
3905 0 : DeltaLayer::new_for_path(path, file)?
3906 0 : .dump(verbose, ctx)
3907 0 : .await?
3908 : }
3909 0 : magic => bail!("unrecognized magic identifier: {:?}", magic),
3910 : }
3911 :
3912 0 : Ok(())
3913 0 : }
3914 :
3915 : #[cfg(test)]
3916 : pub(crate) mod harness {
3917 : use bytes::{Bytes, BytesMut};
3918 : use once_cell::sync::OnceCell;
3919 : use pageserver_api::models::ShardParameters;
3920 : use pageserver_api::shard::ShardIndex;
3921 : use utils::logging;
3922 :
3923 : use crate::deletion_queue::mock::MockDeletionQueue;
3924 : use crate::l0_flush::L0FlushConfig;
3925 : use crate::walredo::apply_neon;
3926 : use crate::{repository::Key, walrecord::NeonWalRecord};
3927 :
3928 : use super::*;
3929 : use hex_literal::hex;
3930 : use utils::id::TenantId;
3931 :
3932 : pub const TIMELINE_ID: TimelineId =
3933 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
3934 : pub const NEW_TIMELINE_ID: TimelineId =
3935 : TimelineId::from_array(hex!("AA223344556677881122334455667788"));
3936 :
3937 : /// Convenience function to create a page image with given string as the only content
3938 15086214 : pub fn test_img(s: &str) -> Bytes {
3939 15086214 : let mut buf = BytesMut::new();
3940 15086214 : buf.extend_from_slice(s.as_bytes());
3941 15086214 : buf.resize(64, 0);
3942 15086214 :
3943 15086214 : buf.freeze()
3944 15086214 : }
3945 :
3946 : impl From<TenantConf> for TenantConfOpt {
3947 564 : fn from(tenant_conf: TenantConf) -> Self {
3948 564 : Self {
3949 564 : checkpoint_distance: Some(tenant_conf.checkpoint_distance),
3950 564 : checkpoint_timeout: Some(tenant_conf.checkpoint_timeout),
3951 564 : compaction_target_size: Some(tenant_conf.compaction_target_size),
3952 564 : compaction_period: Some(tenant_conf.compaction_period),
3953 564 : compaction_threshold: Some(tenant_conf.compaction_threshold),
3954 564 : compaction_algorithm: Some(tenant_conf.compaction_algorithm),
3955 564 : gc_horizon: Some(tenant_conf.gc_horizon),
3956 564 : gc_period: Some(tenant_conf.gc_period),
3957 564 : image_creation_threshold: Some(tenant_conf.image_creation_threshold),
3958 564 : pitr_interval: Some(tenant_conf.pitr_interval),
3959 564 : walreceiver_connect_timeout: Some(tenant_conf.walreceiver_connect_timeout),
3960 564 : lagging_wal_timeout: Some(tenant_conf.lagging_wal_timeout),
3961 564 : max_lsn_wal_lag: Some(tenant_conf.max_lsn_wal_lag),
3962 564 : eviction_policy: Some(tenant_conf.eviction_policy),
3963 564 : min_resident_size_override: tenant_conf.min_resident_size_override,
3964 564 : evictions_low_residence_duration_metric_threshold: Some(
3965 564 : tenant_conf.evictions_low_residence_duration_metric_threshold,
3966 564 : ),
3967 564 : heatmap_period: Some(tenant_conf.heatmap_period),
3968 564 : lazy_slru_download: Some(tenant_conf.lazy_slru_download),
3969 564 : timeline_get_throttle: Some(tenant_conf.timeline_get_throttle),
3970 564 : image_layer_creation_check_threshold: Some(
3971 564 : tenant_conf.image_layer_creation_check_threshold,
3972 564 : ),
3973 564 : switch_aux_file_policy: Some(tenant_conf.switch_aux_file_policy),
3974 564 : lsn_lease_length: Some(tenant_conf.lsn_lease_length),
3975 564 : lsn_lease_length_for_ts: Some(tenant_conf.lsn_lease_length_for_ts),
3976 564 : }
3977 564 : }
3978 : }
3979 :
3980 : pub struct TenantHarness {
3981 : pub conf: &'static PageServerConf,
3982 : pub tenant_conf: TenantConf,
3983 : pub tenant_shard_id: TenantShardId,
3984 : pub generation: Generation,
3985 : pub shard: ShardIndex,
3986 : pub remote_storage: GenericRemoteStorage,
3987 : pub remote_fs_dir: Utf8PathBuf,
3988 : pub deletion_queue: MockDeletionQueue,
3989 : }
3990 :
3991 : static LOG_HANDLE: OnceCell<()> = OnceCell::new();
3992 :
3993 606 : pub(crate) fn setup_logging() {
3994 606 : LOG_HANDLE.get_or_init(|| {
3995 570 : logging::init(
3996 570 : logging::LogFormat::Test,
3997 570 : // enable it in case the tests exercise code paths that use
3998 570 : // debug_assert_current_span_has_tenant_and_timeline_id
3999 570 : logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
4000 570 : logging::Output::Stdout,
4001 570 : )
4002 570 : .expect("Failed to init test logging")
4003 606 : });
4004 606 : }
4005 :
4006 : impl TenantHarness {
4007 564 : pub async fn create_custom(
4008 564 : test_name: &'static str,
4009 564 : tenant_conf: TenantConf,
4010 564 : tenant_id: TenantId,
4011 564 : shard_identity: ShardIdentity,
4012 564 : generation: Generation,
4013 564 : ) -> anyhow::Result<Self> {
4014 564 : setup_logging();
4015 564 :
4016 564 : let repo_dir = PageServerConf::test_repo_dir(test_name);
4017 564 : let _ = fs::remove_dir_all(&repo_dir);
4018 564 : fs::create_dir_all(&repo_dir)?;
4019 :
4020 564 : let conf = PageServerConf::dummy_conf(repo_dir);
4021 564 : // Make a static copy of the config. This can never be free'd, but that's
4022 564 : // OK in a test.
4023 564 : let conf: &'static PageServerConf = Box::leak(Box::new(conf));
4024 564 :
4025 564 : let shard = shard_identity.shard_index();
4026 564 : let tenant_shard_id = TenantShardId {
4027 564 : tenant_id,
4028 564 : shard_number: shard.shard_number,
4029 564 : shard_count: shard.shard_count,
4030 564 : };
4031 564 : fs::create_dir_all(conf.tenant_path(&tenant_shard_id))?;
4032 564 : fs::create_dir_all(conf.timelines_path(&tenant_shard_id))?;
4033 :
4034 : use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
4035 564 : let remote_fs_dir = conf.workdir.join("localfs");
4036 564 : std::fs::create_dir_all(&remote_fs_dir).unwrap();
4037 564 : let config = RemoteStorageConfig {
4038 564 : storage: RemoteStorageKind::LocalFs {
4039 564 : local_path: remote_fs_dir.clone(),
4040 564 : },
4041 564 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
4042 564 : };
4043 564 : let remote_storage = GenericRemoteStorage::from_config(&config).await.unwrap();
4044 564 : let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()));
4045 564 :
4046 564 : Ok(Self {
4047 564 : conf,
4048 564 : tenant_conf,
4049 564 : tenant_shard_id,
4050 564 : generation,
4051 564 : shard,
4052 564 : remote_storage,
4053 564 : remote_fs_dir,
4054 564 : deletion_queue,
4055 564 : })
4056 564 : }
4057 :
4058 528 : pub async fn create(test_name: &'static str) -> anyhow::Result<Self> {
4059 528 : // Disable automatic GC and compaction to make the unit tests more deterministic.
4060 528 : // The tests perform them manually if needed.
4061 528 : let tenant_conf = TenantConf {
4062 528 : gc_period: Duration::ZERO,
4063 528 : compaction_period: Duration::ZERO,
4064 528 : ..TenantConf::default()
4065 528 : };
4066 528 : let tenant_id = TenantId::generate();
4067 528 : let shard = ShardIdentity::unsharded();
4068 528 : Self::create_custom(
4069 528 : test_name,
4070 528 : tenant_conf,
4071 528 : tenant_id,
4072 528 : shard,
4073 528 : Generation::new(0xdeadbeef),
4074 528 : )
4075 0 : .await
4076 528 : }
4077 :
4078 54 : pub fn span(&self) -> tracing::Span {
4079 54 : info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
4080 54 : }
4081 :
4082 564 : pub(crate) async fn load(&self) -> (Arc<Tenant>, RequestContext) {
4083 564 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
4084 564 : (
4085 564 : self.do_try_load(&ctx)
4086 2222 : .await
4087 564 : .expect("failed to load test tenant"),
4088 564 : ctx,
4089 564 : )
4090 564 : }
4091 :
4092 1128 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
4093 : pub(crate) async fn do_try_load(
4094 : &self,
4095 : ctx: &RequestContext,
4096 : ) -> anyhow::Result<Arc<Tenant>> {
4097 : let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
4098 :
4099 : let tenant = Arc::new(Tenant::new(
4100 : TenantState::Loading,
4101 : self.conf,
4102 : AttachedTenantConf::try_from(LocationConf::attached_single(
4103 : TenantConfOpt::from(self.tenant_conf.clone()),
4104 : self.generation,
4105 : &ShardParameters::default(),
4106 : ))
4107 : .unwrap(),
4108 : // This is a legacy/test code path: sharding isn't supported here.
4109 : ShardIdentity::unsharded(),
4110 : Some(walredo_mgr),
4111 : self.tenant_shard_id,
4112 : self.remote_storage.clone(),
4113 : self.deletion_queue.new_client(),
4114 : // TODO: ideally we should run all unit tests with both configs
4115 : L0FlushGlobalState::new(L0FlushConfig::default()),
4116 : ));
4117 :
4118 : let preload = tenant
4119 : .preload(&self.remote_storage, CancellationToken::new())
4120 : .await?;
4121 : tenant.attach(Some(preload), ctx).await?;
4122 :
4123 : tenant.state.send_replace(TenantState::Active);
4124 : for timeline in tenant.timelines.lock().unwrap().values() {
4125 : timeline.set_state(TimelineState::Active);
4126 : }
4127 : Ok(tenant)
4128 : }
4129 :
4130 6 : pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
4131 6 : self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
4132 6 : }
4133 : }
4134 :
4135 : // Mock WAL redo manager that doesn't do much
4136 : pub(crate) struct TestRedoManager;
4137 :
4138 : impl TestRedoManager {
4139 : /// # Cancel-Safety
4140 : ///
4141 : /// This method is cancellation-safe.
4142 978 : pub async fn request_redo(
4143 978 : &self,
4144 978 : key: Key,
4145 978 : lsn: Lsn,
4146 978 : base_img: Option<(Lsn, Bytes)>,
4147 978 : records: Vec<(Lsn, NeonWalRecord)>,
4148 978 : _pg_version: u32,
4149 978 : ) -> Result<Bytes, walredo::Error> {
4150 1326 : let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
4151 978 : if records_neon {
4152 : // For Neon wal records, we can decode without spawning postgres, so do so.
4153 978 : let base_img = base_img.expect("Neon WAL redo requires base image").1;
4154 978 : let mut page = BytesMut::new();
4155 978 : page.extend_from_slice(&base_img);
4156 2304 : for (record_lsn, record) in records {
4157 1326 : apply_neon::apply_in_neon(&record, record_lsn, key, &mut page)?;
4158 : }
4159 978 : Ok(page.freeze())
4160 : } else {
4161 : // We never spawn a postgres walredo process in unit tests: just log what we might have done.
4162 0 : let s = format!(
4163 0 : "redo for {} to get to {}, with {} and {} records",
4164 0 : key,
4165 0 : lsn,
4166 0 : if base_img.is_some() {
4167 0 : "base image"
4168 : } else {
4169 0 : "no base image"
4170 : },
4171 0 : records.len()
4172 0 : );
4173 0 : println!("{s}");
4174 0 :
4175 0 : Ok(test_img(&s))
4176 : }
4177 978 : }
4178 : }
4179 : }
4180 :
4181 : #[cfg(test)]
4182 : mod tests {
4183 : use std::collections::{BTreeMap, BTreeSet};
4184 :
4185 : use super::*;
4186 : use crate::keyspace::KeySpaceAccum;
4187 : use crate::pgdatadir_mapping::AuxFilesDirectory;
4188 : use crate::repository::{Key, Value};
4189 : use crate::tenant::harness::*;
4190 : use crate::tenant::timeline::CompactFlags;
4191 : use crate::walrecord::NeonWalRecord;
4192 : use crate::DEFAULT_PG_VERSION;
4193 : use bytes::{Bytes, BytesMut};
4194 : use hex_literal::hex;
4195 : use itertools::Itertools;
4196 : use pageserver_api::key::{AUX_FILES_KEY, AUX_KEY_PREFIX, NON_INHERITED_RANGE};
4197 : use pageserver_api::keyspace::KeySpace;
4198 : use pageserver_api::models::{CompactionAlgorithm, CompactionAlgorithmSettings};
4199 : use rand::{thread_rng, Rng};
4200 : use storage_layer::PersistentLayerKey;
4201 : use tests::storage_layer::ValuesReconstructState;
4202 : use tests::timeline::{GetVectoredError, ShutdownMode};
4203 : use timeline::compaction::{KeyHistoryRetention, KeyLogAtLsn};
4204 : use timeline::{DeltaLayerTestDesc, GcInfo};
4205 : use utils::bin_ser::BeSer;
4206 : use utils::id::TenantId;
4207 :
4208 : static TEST_KEY: Lazy<Key> =
4209 54 : Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));
4210 :
4211 : #[tokio::test]
4212 6 : async fn test_basic() -> anyhow::Result<()> {
4213 24 : let (tenant, ctx) = TenantHarness::create("test_basic").await?.load().await;
4214 6 : let tline = tenant
4215 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
4216 12 : .await?;
4217 6 :
4218 6 : let mut writer = tline.writer().await;
4219 6 : writer
4220 6 : .put(
4221 6 : *TEST_KEY,
4222 6 : Lsn(0x10),
4223 6 : &Value::Image(test_img("foo at 0x10")),
4224 6 : &ctx,
4225 6 : )
4226 6 : .await?;
4227 6 : writer.finish_write(Lsn(0x10));
4228 6 : drop(writer);
4229 6 :
4230 6 : let mut writer = tline.writer().await;
4231 6 : writer
4232 6 : .put(
4233 6 : *TEST_KEY,
4234 6 : Lsn(0x20),
4235 6 : &Value::Image(test_img("foo at 0x20")),
4236 6 : &ctx,
4237 6 : )
4238 6 : .await?;
4239 6 : writer.finish_write(Lsn(0x20));
4240 6 : drop(writer);
4241 6 :
4242 6 : assert_eq!(
4243 6 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
4244 6 : test_img("foo at 0x10")
4245 6 : );
4246 6 : assert_eq!(
4247 6 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
4248 6 : test_img("foo at 0x10")
4249 6 : );
4250 6 : assert_eq!(
4251 6 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
4252 6 : test_img("foo at 0x20")
4253 6 : );
4254 6 :
4255 6 : Ok(())
4256 6 : }
4257 :
4258 : #[tokio::test]
4259 6 : async fn no_duplicate_timelines() -> anyhow::Result<()> {
4260 6 : let (tenant, ctx) = TenantHarness::create("no_duplicate_timelines")
4261 6 : .await?
4262 6 : .load()
4263 24 : .await;
4264 6 : let _ = tenant
4265 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4266 12 : .await?;
4267 6 :
4268 6 : match tenant
4269 6 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4270 6 : .await
4271 6 : {
4272 6 : Ok(_) => panic!("duplicate timeline creation should fail"),
4273 6 : Err(e) => assert_eq!(e.to_string(), "Already exists".to_string()),
4274 6 : }
4275 6 :
4276 6 : Ok(())
4277 6 : }
4278 :
4279 : /// Convenience function to create a page image with given string as the only content
4280 30 : pub fn test_value(s: &str) -> Value {
4281 30 : let mut buf = BytesMut::new();
4282 30 : buf.extend_from_slice(s.as_bytes());
4283 30 : Value::Image(buf.freeze())
4284 30 : }
4285 :
4286 : ///
4287 : /// Test branch creation
4288 : ///
4289 : #[tokio::test]
4290 6 : async fn test_branch() -> anyhow::Result<()> {
4291 6 : use std::str::from_utf8;
4292 6 :
4293 24 : let (tenant, ctx) = TenantHarness::create("test_branch").await?.load().await;
4294 6 : let tline = tenant
4295 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4296 12 : .await?;
4297 6 : let mut writer = tline.writer().await;
4298 6 :
4299 6 : #[allow(non_snake_case)]
4300 6 : let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
4301 6 : #[allow(non_snake_case)]
4302 6 : let TEST_KEY_B: Key = Key::from_hex("110000000033333333444444445500000002").unwrap();
4303 6 :
4304 6 : // Insert a value on the timeline
4305 6 : writer
4306 6 : .put(TEST_KEY_A, Lsn(0x20), &test_value("foo at 0x20"), &ctx)
4307 6 : .await?;
4308 6 : writer
4309 6 : .put(TEST_KEY_B, Lsn(0x20), &test_value("foobar at 0x20"), &ctx)
4310 6 : .await?;
4311 6 : writer.finish_write(Lsn(0x20));
4312 6 :
4313 6 : writer
4314 6 : .put(TEST_KEY_A, Lsn(0x30), &test_value("foo at 0x30"), &ctx)
4315 6 : .await?;
4316 6 : writer.finish_write(Lsn(0x30));
4317 6 : writer
4318 6 : .put(TEST_KEY_A, Lsn(0x40), &test_value("foo at 0x40"), &ctx)
4319 6 : .await?;
4320 6 : writer.finish_write(Lsn(0x40));
4321 6 :
4322 6 : //assert_current_logical_size(&tline, Lsn(0x40));
4323 6 :
4324 6 : // Branch the history, modify relation differently on the new timeline
4325 6 : tenant
4326 6 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x30)), &ctx)
4327 6 : .await?;
4328 6 : let newtline = tenant
4329 6 : .get_timeline(NEW_TIMELINE_ID, true)
4330 6 : .expect("Should have a local timeline");
4331 6 : let mut new_writer = newtline.writer().await;
4332 6 : new_writer
4333 6 : .put(TEST_KEY_A, Lsn(0x40), &test_value("bar at 0x40"), &ctx)
4334 6 : .await?;
4335 6 : new_writer.finish_write(Lsn(0x40));
4336 6 :
4337 6 : // Check page contents on both branches
4338 6 : assert_eq!(
4339 6 : from_utf8(&tline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
4340 6 : "foo at 0x40"
4341 6 : );
4342 6 : assert_eq!(
4343 6 : from_utf8(&newtline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
4344 6 : "bar at 0x40"
4345 6 : );
4346 6 : assert_eq!(
4347 6 : from_utf8(&newtline.get(TEST_KEY_B, Lsn(0x40), &ctx).await?)?,
4348 6 : "foobar at 0x20"
4349 6 : );
4350 6 :
4351 6 : //assert_current_logical_size(&tline, Lsn(0x40));
4352 6 :
4353 6 : Ok(())
4354 6 : }
4355 :
4356 60 : async fn make_some_layers(
4357 60 : tline: &Timeline,
4358 60 : start_lsn: Lsn,
4359 60 : ctx: &RequestContext,
4360 60 : ) -> anyhow::Result<()> {
4361 60 : let mut lsn = start_lsn;
4362 : {
4363 60 : let mut writer = tline.writer().await;
4364 : // Create a relation on the timeline
4365 60 : writer
4366 60 : .put(
4367 60 : *TEST_KEY,
4368 60 : lsn,
4369 60 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4370 60 : ctx,
4371 60 : )
4372 30 : .await?;
4373 60 : writer.finish_write(lsn);
4374 60 : lsn += 0x10;
4375 60 : writer
4376 60 : .put(
4377 60 : *TEST_KEY,
4378 60 : lsn,
4379 60 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4380 60 : ctx,
4381 60 : )
4382 0 : .await?;
4383 60 : writer.finish_write(lsn);
4384 60 : lsn += 0x10;
4385 60 : }
4386 64 : tline.freeze_and_flush().await?;
4387 : {
4388 60 : let mut writer = tline.writer().await;
4389 60 : writer
4390 60 : .put(
4391 60 : *TEST_KEY,
4392 60 : lsn,
4393 60 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4394 60 : ctx,
4395 60 : )
4396 30 : .await?;
4397 60 : writer.finish_write(lsn);
4398 60 : lsn += 0x10;
4399 60 : writer
4400 60 : .put(
4401 60 : *TEST_KEY,
4402 60 : lsn,
4403 60 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4404 60 : ctx,
4405 60 : )
4406 0 : .await?;
4407 60 : writer.finish_write(lsn);
4408 60 : }
4409 60 : tline.freeze_and_flush().await.map_err(|e| e.into())
4410 60 : }
4411 :
4412 : #[tokio::test]
4413 6 : async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
4414 6 : let (tenant, ctx) =
4415 6 : TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
4416 6 : .await?
4417 6 : .load()
4418 24 : .await;
4419 6 : let tline = tenant
4420 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4421 12 : .await?;
4422 19 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4423 6 :
4424 6 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
4425 6 : // FIXME: this doesn't actually remove any layer currently, given how the flushing
4426 6 : // and compaction works. But it does set the 'cutoff' point so that the cross check
4427 6 : // below should fail.
4428 6 : tenant
4429 6 : .gc_iteration(
4430 6 : Some(TIMELINE_ID),
4431 6 : 0x10,
4432 6 : Duration::ZERO,
4433 6 : &CancellationToken::new(),
4434 6 : &ctx,
4435 6 : )
4436 6 : .await?;
4437 6 :
4438 6 : // try to branch at lsn 25, should fail because we already garbage collected the data
4439 6 : match tenant
4440 6 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
4441 6 : .await
4442 6 : {
4443 6 : Ok(_) => panic!("branching should have failed"),
4444 6 : Err(err) => {
4445 6 : let CreateTimelineError::AncestorLsn(err) = err else {
4446 6 : panic!("wrong error type")
4447 6 : };
4448 6 : assert!(err.to_string().contains("invalid branch start lsn"));
4449 6 : assert!(err
4450 6 : .source()
4451 6 : .unwrap()
4452 6 : .to_string()
4453 6 : .contains("we might've already garbage collected needed data"))
4454 6 : }
4455 6 : }
4456 6 :
4457 6 : Ok(())
4458 6 : }
4459 :
4460 : #[tokio::test]
4461 6 : async fn test_prohibit_branch_creation_on_pre_initdb_lsn() -> anyhow::Result<()> {
4462 6 : let (tenant, ctx) =
4463 6 : TenantHarness::create("test_prohibit_branch_creation_on_pre_initdb_lsn")
4464 6 : .await?
4465 6 : .load()
4466 24 : .await;
4467 6 :
4468 6 : let tline = tenant
4469 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x50), DEFAULT_PG_VERSION, &ctx)
4470 12 : .await?;
4471 6 : // try to branch at lsn 0x25, should fail because initdb lsn is 0x50
4472 6 : match tenant
4473 6 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
4474 6 : .await
4475 6 : {
4476 6 : Ok(_) => panic!("branching should have failed"),
4477 6 : Err(err) => {
4478 6 : let CreateTimelineError::AncestorLsn(err) = err else {
4479 6 : panic!("wrong error type");
4480 6 : };
4481 6 : assert!(&err.to_string().contains("invalid branch start lsn"));
4482 6 : assert!(&err
4483 6 : .source()
4484 6 : .unwrap()
4485 6 : .to_string()
4486 6 : .contains("is earlier than latest GC cutoff"));
4487 6 : }
4488 6 : }
4489 6 :
4490 6 : Ok(())
4491 6 : }
4492 :
4493 : /*
4494 : // FIXME: This currently fails to error out. Calling GC doesn't currently
4495 : // remove the old value, we'd need to work a little harder
4496 : #[tokio::test]
4497 : async fn test_prohibit_get_for_garbage_collected_data() -> anyhow::Result<()> {
4498 : let repo =
4499 : RepoHarness::create("test_prohibit_get_for_garbage_collected_data")?
4500 : .load();
4501 :
4502 : let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION)?;
4503 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4504 :
4505 : repo.gc_iteration(Some(TIMELINE_ID), 0x10, Duration::ZERO)?;
4506 : let latest_gc_cutoff_lsn = tline.get_latest_gc_cutoff_lsn();
4507 : assert!(*latest_gc_cutoff_lsn > Lsn(0x25));
4508 : match tline.get(*TEST_KEY, Lsn(0x25)) {
4509 : Ok(_) => panic!("request for page should have failed"),
4510 : Err(err) => assert!(err.to_string().contains("not found at")),
4511 : }
4512 : Ok(())
4513 : }
4514 : */
4515 :
4516 : #[tokio::test]
4517 6 : async fn test_get_branchpoints_from_an_inactive_timeline() -> anyhow::Result<()> {
4518 6 : let (tenant, ctx) =
4519 6 : TenantHarness::create("test_get_branchpoints_from_an_inactive_timeline")
4520 6 : .await?
4521 6 : .load()
4522 24 : .await;
4523 6 : let tline = tenant
4524 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4525 12 : .await?;
4526 18 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4527 6 :
4528 6 : tenant
4529 6 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
4530 6 : .await?;
4531 6 : let newtline = tenant
4532 6 : .get_timeline(NEW_TIMELINE_ID, true)
4533 6 : .expect("Should have a local timeline");
4534 6 :
4535 18 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
4536 6 :
4537 6 : tline.set_broken("test".to_owned());
4538 6 :
4539 6 : tenant
4540 6 : .gc_iteration(
4541 6 : Some(TIMELINE_ID),
4542 6 : 0x10,
4543 6 : Duration::ZERO,
4544 6 : &CancellationToken::new(),
4545 6 : &ctx,
4546 6 : )
4547 6 : .await?;
4548 6 :
4549 6 : // The branchpoints should contain all timelines, even ones marked
4550 6 : // as Broken.
4551 6 : {
4552 6 : let branchpoints = &tline.gc_info.read().unwrap().retain_lsns;
4553 6 : assert_eq!(branchpoints.len(), 1);
4554 6 : assert_eq!(branchpoints[0], (Lsn(0x40), NEW_TIMELINE_ID));
4555 6 : }
4556 6 :
4557 6 : // You can read the key from the child branch even though the parent is
4558 6 : // Broken, as long as you don't need to access data from the parent.
4559 6 : assert_eq!(
4560 12 : newtline.get(*TEST_KEY, Lsn(0x70), &ctx).await?,
4561 6 : test_img(&format!("foo at {}", Lsn(0x70)))
4562 6 : );
4563 6 :
4564 6 : // This needs to traverse to the parent, and fails.
4565 6 : let err = newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await.unwrap_err();
4566 6 : assert!(
4567 6 : err.to_string().starts_with(&format!(
4568 6 : "bad state on timeline {}: Broken",
4569 6 : tline.timeline_id
4570 6 : )),
4571 6 : "{err}"
4572 6 : );
4573 6 :
4574 6 : Ok(())
4575 6 : }
4576 :
4577 : #[tokio::test]
4578 6 : async fn test_retain_data_in_parent_which_is_needed_for_child() -> anyhow::Result<()> {
4579 6 : let (tenant, ctx) =
4580 6 : TenantHarness::create("test_retain_data_in_parent_which_is_needed_for_child")
4581 6 : .await?
4582 6 : .load()
4583 24 : .await;
4584 6 : let tline = tenant
4585 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4586 12 : .await?;
4587 19 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4588 6 :
4589 6 : tenant
4590 6 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
4591 6 : .await?;
4592 6 : let newtline = tenant
4593 6 : .get_timeline(NEW_TIMELINE_ID, true)
4594 6 : .expect("Should have a local timeline");
4595 6 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
4596 6 : tenant
4597 6 : .gc_iteration(
4598 6 : Some(TIMELINE_ID),
4599 6 : 0x10,
4600 6 : Duration::ZERO,
4601 6 : &CancellationToken::new(),
4602 6 : &ctx,
4603 6 : )
4604 6 : .await?;
4605 12 : assert!(newtline.get(*TEST_KEY, Lsn(0x25), &ctx).await.is_ok());
4606 6 :
4607 6 : Ok(())
4608 6 : }
4609 : #[tokio::test]
4610 6 : async fn test_parent_keeps_data_forever_after_branching() -> anyhow::Result<()> {
4611 6 : let (tenant, ctx) = TenantHarness::create("test_parent_keeps_data_forever_after_branching")
4612 6 : .await?
4613 6 : .load()
4614 24 : .await;
4615 6 : let tline = tenant
4616 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4617 12 : .await?;
4618 18 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4619 6 :
4620 6 : tenant
4621 6 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
4622 6 : .await?;
4623 6 : let newtline = tenant
4624 6 : .get_timeline(NEW_TIMELINE_ID, true)
4625 6 : .expect("Should have a local timeline");
4626 6 :
4627 18 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
4628 6 :
4629 6 : // run gc on parent
4630 6 : tenant
4631 6 : .gc_iteration(
4632 6 : Some(TIMELINE_ID),
4633 6 : 0x10,
4634 6 : Duration::ZERO,
4635 6 : &CancellationToken::new(),
4636 6 : &ctx,
4637 6 : )
4638 6 : .await?;
4639 6 :
4640 6 : // Check that the data is still accessible on the branch.
4641 6 : assert_eq!(
4642 21 : newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
4643 6 : test_img(&format!("foo at {}", Lsn(0x40)))
4644 6 : );
4645 6 :
4646 6 : Ok(())
4647 6 : }
4648 :
4649 : #[tokio::test]
4650 6 : async fn timeline_load() -> anyhow::Result<()> {
4651 6 : const TEST_NAME: &str = "timeline_load";
4652 6 : let harness = TenantHarness::create(TEST_NAME).await?;
4653 6 : {
4654 21 : let (tenant, ctx) = harness.load().await;
4655 6 : let tline = tenant
4656 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x7000), DEFAULT_PG_VERSION, &ctx)
4657 12 : .await?;
4658 18 : make_some_layers(tline.as_ref(), Lsn(0x8000), &ctx).await?;
4659 6 : // so that all uploads finish & we can call harness.load() below again
4660 6 : tenant
4661 6 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
4662 6 : .instrument(harness.span())
4663 6 : .await
4664 6 : .ok()
4665 6 : .unwrap();
4666 6 : }
4667 6 :
4668 19 : let (tenant, _ctx) = harness.load().await;
4669 6 : tenant
4670 6 : .get_timeline(TIMELINE_ID, true)
4671 6 : .expect("cannot load timeline");
4672 6 :
4673 6 : Ok(())
4674 6 : }
4675 :
4676 : #[tokio::test]
4677 6 : async fn timeline_load_with_ancestor() -> anyhow::Result<()> {
4678 6 : const TEST_NAME: &str = "timeline_load_with_ancestor";
4679 6 : let harness = TenantHarness::create(TEST_NAME).await?;
4680 6 : // create two timelines
4681 6 : {
4682 24 : let (tenant, ctx) = harness.load().await;
4683 6 : let tline = tenant
4684 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4685 12 : .await?;
4686 6 :
4687 18 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4688 6 :
4689 6 : let child_tline = tenant
4690 6 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
4691 6 : .await?;
4692 6 : child_tline.set_state(TimelineState::Active);
4693 6 :
4694 6 : let newtline = tenant
4695 6 : .get_timeline(NEW_TIMELINE_ID, true)
4696 6 : .expect("Should have a local timeline");
4697 6 :
4698 18 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
4699 6 :
4700 6 : // so that all uploads finish & we can call harness.load() below again
4701 6 : tenant
4702 6 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
4703 6 : .instrument(harness.span())
4704 7 : .await
4705 6 : .ok()
4706 6 : .unwrap();
4707 6 : }
4708 6 :
4709 6 : // check that both of them are initially unloaded
4710 35 : let (tenant, _ctx) = harness.load().await;
4711 6 :
4712 6 : // check that both, child and ancestor are loaded
4713 6 : let _child_tline = tenant
4714 6 : .get_timeline(NEW_TIMELINE_ID, true)
4715 6 : .expect("cannot get child timeline loaded");
4716 6 :
4717 6 : let _ancestor_tline = tenant
4718 6 : .get_timeline(TIMELINE_ID, true)
4719 6 : .expect("cannot get ancestor timeline loaded");
4720 6 :
4721 6 : Ok(())
4722 6 : }
4723 :
4724 : #[tokio::test]
4725 6 : async fn delta_layer_dumping() -> anyhow::Result<()> {
4726 6 : use storage_layer::AsLayerDesc;
4727 6 : let (tenant, ctx) = TenantHarness::create("test_layer_dumping")
4728 6 : .await?
4729 6 : .load()
4730 24 : .await;
4731 6 : let tline = tenant
4732 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4733 12 : .await?;
4734 20 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4735 6 :
4736 6 : let layer_map = tline.layers.read().await;
4737 6 : let level0_deltas = layer_map
4738 6 : .layer_map()?
4739 6 : .level0_deltas()
4740 6 : .iter()
4741 12 : .map(|desc| layer_map.get_from_desc(desc))
4742 6 : .collect::<Vec<_>>();
4743 6 :
4744 6 : assert!(!level0_deltas.is_empty());
4745 6 :
4746 18 : for delta in level0_deltas {
4747 6 : // Ensure we are dumping a delta layer here
4748 12 : assert!(delta.layer_desc().is_delta);
4749 24 : delta.dump(true, &ctx).await.unwrap();
4750 6 : }
4751 6 :
4752 6 : Ok(())
4753 6 : }
4754 :
4755 : #[tokio::test]
4756 6 : async fn test_images() -> anyhow::Result<()> {
4757 24 : let (tenant, ctx) = TenantHarness::create("test_images").await?.load().await;
4758 6 : let tline = tenant
4759 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
4760 12 : .await?;
4761 6 :
4762 6 : let mut writer = tline.writer().await;
4763 6 : writer
4764 6 : .put(
4765 6 : *TEST_KEY,
4766 6 : Lsn(0x10),
4767 6 : &Value::Image(test_img("foo at 0x10")),
4768 6 : &ctx,
4769 6 : )
4770 6 : .await?;
4771 6 : writer.finish_write(Lsn(0x10));
4772 6 : drop(writer);
4773 6 :
4774 6 : tline.freeze_and_flush().await?;
4775 6 : tline
4776 6 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
4777 6 : .await?;
4778 6 :
4779 6 : let mut writer = tline.writer().await;
4780 6 : writer
4781 6 : .put(
4782 6 : *TEST_KEY,
4783 6 : Lsn(0x20),
4784 6 : &Value::Image(test_img("foo at 0x20")),
4785 6 : &ctx,
4786 6 : )
4787 6 : .await?;
4788 6 : writer.finish_write(Lsn(0x20));
4789 6 : drop(writer);
4790 6 :
4791 7 : tline.freeze_and_flush().await?;
4792 6 : tline
4793 6 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
4794 6 : .await?;
4795 6 :
4796 6 : let mut writer = tline.writer().await;
4797 6 : writer
4798 6 : .put(
4799 6 : *TEST_KEY,
4800 6 : Lsn(0x30),
4801 6 : &Value::Image(test_img("foo at 0x30")),
4802 6 : &ctx,
4803 6 : )
4804 6 : .await?;
4805 6 : writer.finish_write(Lsn(0x30));
4806 6 : drop(writer);
4807 6 :
4808 6 : tline.freeze_and_flush().await?;
4809 6 : tline
4810 6 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
4811 6 : .await?;
4812 6 :
4813 6 : let mut writer = tline.writer().await;
4814 6 : writer
4815 6 : .put(
4816 6 : *TEST_KEY,
4817 6 : Lsn(0x40),
4818 6 : &Value::Image(test_img("foo at 0x40")),
4819 6 : &ctx,
4820 6 : )
4821 6 : .await?;
4822 6 : writer.finish_write(Lsn(0x40));
4823 6 : drop(writer);
4824 6 :
4825 6 : tline.freeze_and_flush().await?;
4826 6 : tline
4827 6 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
4828 6 : .await?;
4829 6 :
4830 6 : assert_eq!(
4831 12 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
4832 6 : test_img("foo at 0x10")
4833 6 : );
4834 6 : assert_eq!(
4835 12 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
4836 6 : test_img("foo at 0x10")
4837 6 : );
4838 6 : assert_eq!(
4839 6 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
4840 6 : test_img("foo at 0x20")
4841 6 : );
4842 6 : assert_eq!(
4843 12 : tline.get(*TEST_KEY, Lsn(0x30), &ctx).await?,
4844 6 : test_img("foo at 0x30")
4845 6 : );
4846 6 : assert_eq!(
4847 12 : tline.get(*TEST_KEY, Lsn(0x40), &ctx).await?,
4848 6 : test_img("foo at 0x40")
4849 6 : );
4850 6 :
4851 6 : Ok(())
4852 6 : }
4853 :
4854 12 : async fn bulk_insert_compact_gc(
4855 12 : tenant: &Tenant,
4856 12 : timeline: &Arc<Timeline>,
4857 12 : ctx: &RequestContext,
4858 12 : lsn: Lsn,
4859 12 : repeat: usize,
4860 12 : key_count: usize,
4861 12 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
4862 12 : let compact = true;
4863 226439 : bulk_insert_maybe_compact_gc(tenant, timeline, ctx, lsn, repeat, key_count, compact).await
4864 12 : }
4865 :
4866 24 : async fn bulk_insert_maybe_compact_gc(
4867 24 : tenant: &Tenant,
4868 24 : timeline: &Arc<Timeline>,
4869 24 : ctx: &RequestContext,
4870 24 : mut lsn: Lsn,
4871 24 : repeat: usize,
4872 24 : key_count: usize,
4873 24 : compact: bool,
4874 24 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
4875 24 : let mut inserted: HashMap<Key, BTreeSet<Lsn>> = Default::default();
4876 24 :
4877 24 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
4878 24 : let mut blknum = 0;
4879 24 :
4880 24 : // Enforce that key range is monotonously increasing
4881 24 : let mut keyspace = KeySpaceAccum::new();
4882 24 :
4883 24 : let cancel = CancellationToken::new();
4884 24 :
4885 24 : for _ in 0..repeat {
4886 1200 : for _ in 0..key_count {
4887 12000000 : test_key.field6 = blknum;
4888 12000000 : let mut writer = timeline.writer().await;
4889 12000000 : writer
4890 12000000 : .put(
4891 12000000 : test_key,
4892 12000000 : lsn,
4893 12000000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
4894 12000000 : ctx,
4895 12000000 : )
4896 7943 : .await?;
4897 12000000 : inserted.entry(test_key).or_default().insert(lsn);
4898 12000000 : writer.finish_write(lsn);
4899 12000000 : drop(writer);
4900 12000000 :
4901 12000000 : keyspace.add_key(test_key);
4902 12000000 :
4903 12000000 : lsn = Lsn(lsn.0 + 0x10);
4904 12000000 : blknum += 1;
4905 : }
4906 :
4907 1200 : timeline.freeze_and_flush().await?;
4908 1200 : if compact {
4909 : // this requires timeline to be &Arc<Timeline>
4910 129540 : timeline.compact(&cancel, EnumSet::empty(), ctx).await?;
4911 600 : }
4912 :
4913 : // this doesn't really need to use the timeline_id target, but it is closer to what it
4914 : // originally was.
4915 1200 : let res = tenant
4916 1200 : .gc_iteration(Some(timeline.timeline_id), 0, Duration::ZERO, &cancel, ctx)
4917 1198 : .await?;
4918 :
4919 1200 : assert_eq!(res.layers_removed, 0, "this never removes anything");
4920 : }
4921 :
4922 24 : Ok(inserted)
4923 24 : }
4924 :
4925 : //
4926 : // Insert 1000 key-value pairs with increasing keys, flush, compact, GC.
4927 : // Repeat 50 times.
4928 : //
4929 : #[tokio::test]
4930 6 : async fn test_bulk_insert() -> anyhow::Result<()> {
4931 6 : let harness = TenantHarness::create("test_bulk_insert").await?;
4932 24 : let (tenant, ctx) = harness.load().await;
4933 6 : let tline = tenant
4934 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
4935 12 : .await?;
4936 6 :
4937 6 : let lsn = Lsn(0x10);
4938 113220 : bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
4939 6 :
4940 6 : Ok(())
4941 6 : }
4942 :
4943 : // Test the vectored get real implementation against a simple sequential implementation.
4944 : //
4945 : // The test generates a keyspace by repeatedly flushing the in-memory layer and compacting.
4946 : // Projected to 2D the key space looks like below. Lsn grows upwards on the Y axis and keys
4947 : // grow to the right on the X axis.
4948 : // [Delta]
4949 : // [Delta]
4950 : // [Delta]
4951 : // [Delta]
4952 : // ------------ Image ---------------
4953 : //
4954 : // After layer generation we pick the ranges to query as follows:
4955 : // 1. The beginning of each delta layer
4956 : // 2. At the seam between two adjacent delta layers
4957 : //
4958 : // There's one major downside to this test: delta layers only contains images,
4959 : // so the search can stop at the first delta layer and doesn't traverse any deeper.
4960 : #[tokio::test]
4961 6 : async fn test_get_vectored() -> anyhow::Result<()> {
4962 6 : let harness = TenantHarness::create("test_get_vectored").await?;
4963 24 : let (tenant, ctx) = harness.load().await;
4964 6 : let tline = tenant
4965 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
4966 12 : .await?;
4967 6 :
4968 6 : let lsn = Lsn(0x10);
4969 113219 : let inserted = bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
4970 6 :
4971 6 : let guard = tline.layers.read().await;
4972 6 : let lm = guard.layer_map()?;
4973 6 :
4974 6 : lm.dump(true, &ctx).await?;
4975 6 :
4976 6 : let mut reads = Vec::new();
4977 6 : let mut prev = None;
4978 36 : lm.iter_historic_layers().for_each(|desc| {
4979 36 : if !desc.is_delta() {
4980 6 : prev = Some(desc.clone());
4981 6 : return;
4982 30 : }
4983 30 :
4984 30 : let start = desc.key_range.start;
4985 30 : let end = desc
4986 30 : .key_range
4987 30 : .start
4988 30 : .add(Timeline::MAX_GET_VECTORED_KEYS.try_into().unwrap());
4989 30 : reads.push(KeySpace {
4990 30 : ranges: vec![start..end],
4991 30 : });
4992 6 :
4993 30 : if let Some(prev) = &prev {
4994 30 : if !prev.is_delta() {
4995 30 : return;
4996 6 : }
4997 0 :
4998 0 : let first_range = Key {
4999 0 : field6: prev.key_range.end.field6 - 4,
5000 0 : ..prev.key_range.end
5001 0 : }..prev.key_range.end;
5002 0 :
5003 0 : let second_range = desc.key_range.start..Key {
5004 0 : field6: desc.key_range.start.field6 + 4,
5005 0 : ..desc.key_range.start
5006 0 : };
5007 0 :
5008 0 : reads.push(KeySpace {
5009 0 : ranges: vec![first_range, second_range],
5010 0 : });
5011 6 : };
5012 6 :
5013 6 : prev = Some(desc.clone());
5014 36 : });
5015 6 :
5016 6 : drop(guard);
5017 6 :
5018 6 : // Pick a big LSN such that we query over all the changes.
5019 6 : let reads_lsn = Lsn(u64::MAX - 1);
5020 6 :
5021 36 : for read in reads {
5022 30 : info!("Doing vectored read on {:?}", read);
5023 6 :
5024 30 : let vectored_res = tline
5025 30 : .get_vectored_impl(
5026 30 : read.clone(),
5027 30 : reads_lsn,
5028 30 : &mut ValuesReconstructState::new(),
5029 30 : &ctx,
5030 30 : )
5031 75 : .await;
5032 6 :
5033 30 : let mut expected_lsns: HashMap<Key, Lsn> = Default::default();
5034 30 : let mut expect_missing = false;
5035 30 : let mut key = read.start().unwrap();
5036 990 : while key != read.end().unwrap() {
5037 960 : if let Some(lsns) = inserted.get(&key) {
5038 960 : let expected_lsn = lsns.iter().rfind(|lsn| **lsn <= reads_lsn);
5039 960 : match expected_lsn {
5040 960 : Some(lsn) => {
5041 960 : expected_lsns.insert(key, *lsn);
5042 960 : }
5043 6 : None => {
5044 6 : expect_missing = true;
5045 0 : break;
5046 6 : }
5047 6 : }
5048 6 : } else {
5049 6 : expect_missing = true;
5050 0 : break;
5051 6 : }
5052 6 :
5053 960 : key = key.next();
5054 6 : }
5055 6 :
5056 30 : if expect_missing {
5057 6 : assert!(matches!(vectored_res, Err(GetVectoredError::MissingKey(_))));
5058 6 : } else {
5059 960 : for (key, image) in vectored_res? {
5060 960 : let expected_lsn = expected_lsns.get(&key).expect("determined above");
5061 960 : let expected_image = test_img(&format!("{} at {}", key.field6, expected_lsn));
5062 960 : assert_eq!(image?, expected_image);
5063 6 : }
5064 6 : }
5065 6 : }
5066 6 :
5067 6 : Ok(())
5068 6 : }
5069 :
5070 : #[tokio::test]
5071 6 : async fn test_get_vectored_aux_files() -> anyhow::Result<()> {
5072 6 : let harness = TenantHarness::create("test_get_vectored_aux_files").await?;
5073 6 :
5074 21 : let (tenant, ctx) = harness.load().await;
5075 6 : let tline = tenant
5076 6 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
5077 6 : .await?;
5078 6 : let tline = tline.raw_timeline().unwrap();
5079 6 :
5080 6 : let mut modification = tline.begin_modification(Lsn(0x1000));
5081 6 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
5082 6 : modification.set_lsn(Lsn(0x1008))?;
5083 6 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
5084 6 : modification.commit(&ctx).await?;
5085 6 :
5086 6 : let child_timeline_id = TimelineId::generate();
5087 6 : tenant
5088 6 : .branch_timeline_test(
5089 6 : tline,
5090 6 : child_timeline_id,
5091 6 : Some(tline.get_last_record_lsn()),
5092 6 : &ctx,
5093 6 : )
5094 6 : .await?;
5095 6 :
5096 6 : let child_timeline = tenant
5097 6 : .get_timeline(child_timeline_id, true)
5098 6 : .expect("Should have the branched timeline");
5099 6 :
5100 6 : let aux_keyspace = KeySpace {
5101 6 : ranges: vec![NON_INHERITED_RANGE],
5102 6 : };
5103 6 : let read_lsn = child_timeline.get_last_record_lsn();
5104 6 :
5105 6 : let vectored_res = child_timeline
5106 6 : .get_vectored_impl(
5107 6 : aux_keyspace.clone(),
5108 6 : read_lsn,
5109 6 : &mut ValuesReconstructState::new(),
5110 6 : &ctx,
5111 6 : )
5112 6 : .await;
5113 6 :
5114 6 : let images = vectored_res?;
5115 6 : assert!(images.is_empty());
5116 6 : Ok(())
5117 6 : }
5118 :
5119 : // Test that vectored get handles layer gaps correctly
5120 : // by advancing into the next ancestor timeline if required.
5121 : //
5122 : // The test generates timelines that look like the diagram below.
5123 : // We leave a gap in one of the L1 layers at `gap_at_key` (`/` in the diagram).
5124 : // The reconstruct data for that key lies in the ancestor timeline (`X` in the diagram).
5125 : //
5126 : // ```
5127 : //-------------------------------+
5128 : // ... |
5129 : // [ L1 ] |
5130 : // [ / L1 ] | Child Timeline
5131 : // ... |
5132 : // ------------------------------+
5133 : // [ X L1 ] | Parent Timeline
5134 : // ------------------------------+
5135 : // ```
5136 : #[tokio::test]
5137 6 : async fn test_get_vectored_key_gap() -> anyhow::Result<()> {
5138 6 : let tenant_conf = TenantConf {
5139 6 : // Make compaction deterministic
5140 6 : gc_period: Duration::ZERO,
5141 6 : compaction_period: Duration::ZERO,
5142 6 : // Encourage creation of L1 layers
5143 6 : checkpoint_distance: 16 * 1024,
5144 6 : compaction_target_size: 8 * 1024,
5145 6 : ..TenantConf::default()
5146 6 : };
5147 6 :
5148 6 : let harness = TenantHarness::create_custom(
5149 6 : "test_get_vectored_key_gap",
5150 6 : tenant_conf,
5151 6 : TenantId::generate(),
5152 6 : ShardIdentity::unsharded(),
5153 6 : Generation::new(0xdeadbeef),
5154 6 : )
5155 6 : .await?;
5156 24 : let (tenant, ctx) = harness.load().await;
5157 6 :
5158 6 : let mut current_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5159 6 : let gap_at_key = current_key.add(100);
5160 6 : let mut current_lsn = Lsn(0x10);
5161 6 :
5162 6 : const KEY_COUNT: usize = 10_000;
5163 6 :
5164 6 : let timeline_id = TimelineId::generate();
5165 6 : let current_timeline = tenant
5166 6 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
5167 12 : .await?;
5168 6 :
5169 6 : current_lsn += 0x100;
5170 6 :
5171 6 : let mut writer = current_timeline.writer().await;
5172 6 : writer
5173 6 : .put(
5174 6 : gap_at_key,
5175 6 : current_lsn,
5176 6 : &Value::Image(test_img(&format!("{} at {}", gap_at_key, current_lsn))),
5177 6 : &ctx,
5178 6 : )
5179 6 : .await?;
5180 6 : writer.finish_write(current_lsn);
5181 6 : drop(writer);
5182 6 :
5183 6 : let mut latest_lsns = HashMap::new();
5184 6 : latest_lsns.insert(gap_at_key, current_lsn);
5185 6 :
5186 6 : current_timeline.freeze_and_flush().await?;
5187 6 :
5188 6 : let child_timeline_id = TimelineId::generate();
5189 6 :
5190 6 : tenant
5191 6 : .branch_timeline_test(
5192 6 : ¤t_timeline,
5193 6 : child_timeline_id,
5194 6 : Some(current_lsn),
5195 6 : &ctx,
5196 6 : )
5197 6 : .await?;
5198 6 : let child_timeline = tenant
5199 6 : .get_timeline(child_timeline_id, true)
5200 6 : .expect("Should have the branched timeline");
5201 6 :
5202 60006 : for i in 0..KEY_COUNT {
5203 60000 : if current_key == gap_at_key {
5204 6 : current_key = current_key.next();
5205 6 : continue;
5206 59994 : }
5207 59994 :
5208 59994 : current_lsn += 0x10;
5209 6 :
5210 59994 : let mut writer = child_timeline.writer().await;
5211 59994 : writer
5212 59994 : .put(
5213 59994 : current_key,
5214 59994 : current_lsn,
5215 59994 : &Value::Image(test_img(&format!("{} at {}", current_key, current_lsn))),
5216 59994 : &ctx,
5217 59994 : )
5218 202 : .await?;
5219 59994 : writer.finish_write(current_lsn);
5220 59994 : drop(writer);
5221 59994 :
5222 59994 : latest_lsns.insert(current_key, current_lsn);
5223 59994 : current_key = current_key.next();
5224 59994 :
5225 59994 : // Flush every now and then to encourage layer file creation.
5226 59994 : if i % 500 == 0 {
5227 127 : child_timeline.freeze_and_flush().await?;
5228 59874 : }
5229 6 : }
5230 6 :
5231 6 : child_timeline.freeze_and_flush().await?;
5232 6 : let mut flags = EnumSet::new();
5233 6 : flags.insert(CompactFlags::ForceRepartition);
5234 6 : child_timeline
5235 6 : .compact(&CancellationToken::new(), flags, &ctx)
5236 6981 : .await?;
5237 6 :
5238 6 : let key_near_end = {
5239 6 : let mut tmp = current_key;
5240 6 : tmp.field6 -= 10;
5241 6 : tmp
5242 6 : };
5243 6 :
5244 6 : let key_near_gap = {
5245 6 : let mut tmp = gap_at_key;
5246 6 : tmp.field6 -= 10;
5247 6 : tmp
5248 6 : };
5249 6 :
5250 6 : let read = KeySpace {
5251 6 : ranges: vec![key_near_gap..gap_at_key.next(), key_near_end..current_key],
5252 6 : };
5253 6 : let results = child_timeline
5254 6 : .get_vectored_impl(
5255 6 : read.clone(),
5256 6 : current_lsn,
5257 6 : &mut ValuesReconstructState::new(),
5258 6 : &ctx,
5259 6 : )
5260 45 : .await?;
5261 6 :
5262 132 : for (key, img_res) in results {
5263 126 : let expected = test_img(&format!("{} at {}", key, latest_lsns[&key]));
5264 126 : assert_eq!(img_res?, expected);
5265 6 : }
5266 6 :
5267 6 : Ok(())
5268 6 : }
5269 :
5270 : // Test that vectored get descends into ancestor timelines correctly and
5271 : // does not return an image that's newer than requested.
5272 : //
5273 : // The diagram below ilustrates an interesting case. We have a parent timeline
5274 : // (top of the Lsn range) and a child timeline. The request key cannot be reconstructed
5275 : // from the child timeline, so the parent timeline must be visited. When advacing into
5276 : // the child timeline, the read path needs to remember what the requested Lsn was in
5277 : // order to avoid returning an image that's too new. The test below constructs such
5278 : // a timeline setup and does a few queries around the Lsn of each page image.
5279 : // ```
5280 : // LSN
5281 : // ^
5282 : // |
5283 : // |
5284 : // 500 | --------------------------------------> branch point
5285 : // 400 | X
5286 : // 300 | X
5287 : // 200 | --------------------------------------> requested lsn
5288 : // 100 | X
5289 : // |---------------------------------------> Key
5290 : // |
5291 : // ------> requested key
5292 : //
5293 : // Legend:
5294 : // * X - page images
5295 : // ```
5296 : #[tokio::test]
5297 6 : async fn test_get_vectored_ancestor_descent() -> anyhow::Result<()> {
5298 6 : let harness = TenantHarness::create("test_get_vectored_on_lsn_axis").await?;
5299 24 : let (tenant, ctx) = harness.load().await;
5300 6 :
5301 6 : let start_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5302 6 : let end_key = start_key.add(1000);
5303 6 : let child_gap_at_key = start_key.add(500);
5304 6 : let mut parent_gap_lsns: BTreeMap<Lsn, String> = BTreeMap::new();
5305 6 :
5306 6 : let mut current_lsn = Lsn(0x10);
5307 6 :
5308 6 : let timeline_id = TimelineId::generate();
5309 6 : let parent_timeline = tenant
5310 6 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
5311 12 : .await?;
5312 6 :
5313 6 : current_lsn += 0x100;
5314 6 :
5315 24 : for _ in 0..3 {
5316 18 : let mut key = start_key;
5317 18018 : while key < end_key {
5318 18000 : current_lsn += 0x10;
5319 18000 :
5320 18000 : let image_value = format!("{} at {}", child_gap_at_key, current_lsn);
5321 6 :
5322 18000 : let mut writer = parent_timeline.writer().await;
5323 18000 : writer
5324 18000 : .put(
5325 18000 : key,
5326 18000 : current_lsn,
5327 18000 : &Value::Image(test_img(&image_value)),
5328 18000 : &ctx,
5329 18000 : )
5330 18 : .await?;
5331 18000 : writer.finish_write(current_lsn);
5332 18000 :
5333 18000 : if key == child_gap_at_key {
5334 18 : parent_gap_lsns.insert(current_lsn, image_value);
5335 17982 : }
5336 6 :
5337 18000 : key = key.next();
5338 6 : }
5339 6 :
5340 18 : parent_timeline.freeze_and_flush().await?;
5341 6 : }
5342 6 :
5343 6 : let child_timeline_id = TimelineId::generate();
5344 6 :
5345 6 : let child_timeline = tenant
5346 6 : .branch_timeline_test(&parent_timeline, child_timeline_id, Some(current_lsn), &ctx)
5347 6 : .await?;
5348 6 :
5349 6 : let mut key = start_key;
5350 6006 : while key < end_key {
5351 6000 : if key == child_gap_at_key {
5352 6 : key = key.next();
5353 6 : continue;
5354 5994 : }
5355 5994 :
5356 5994 : current_lsn += 0x10;
5357 6 :
5358 5994 : let mut writer = child_timeline.writer().await;
5359 5994 : writer
5360 5994 : .put(
5361 5994 : key,
5362 5994 : current_lsn,
5363 5994 : &Value::Image(test_img(&format!("{} at {}", key, current_lsn))),
5364 5994 : &ctx,
5365 5994 : )
5366 51 : .await?;
5367 5994 : writer.finish_write(current_lsn);
5368 5994 :
5369 5994 : key = key.next();
5370 6 : }
5371 6 :
5372 6 : child_timeline.freeze_and_flush().await?;
5373 6 :
5374 6 : let lsn_offsets: [i64; 5] = [-10, -1, 0, 1, 10];
5375 6 : let mut query_lsns = Vec::new();
5376 18 : for image_lsn in parent_gap_lsns.keys().rev() {
5377 108 : for offset in lsn_offsets {
5378 90 : query_lsns.push(Lsn(image_lsn
5379 90 : .0
5380 90 : .checked_add_signed(offset)
5381 90 : .expect("Shouldn't overflow")));
5382 90 : }
5383 6 : }
5384 6 :
5385 96 : for query_lsn in query_lsns {
5386 90 : let results = child_timeline
5387 90 : .get_vectored_impl(
5388 90 : KeySpace {
5389 90 : ranges: vec![child_gap_at_key..child_gap_at_key.next()],
5390 90 : },
5391 90 : query_lsn,
5392 90 : &mut ValuesReconstructState::new(),
5393 90 : &ctx,
5394 90 : )
5395 87 : .await;
5396 6 :
5397 90 : let expected_item = parent_gap_lsns
5398 90 : .iter()
5399 90 : .rev()
5400 204 : .find(|(lsn, _)| **lsn <= query_lsn);
5401 90 :
5402 90 : info!(
5403 6 : "Doing vectored read at LSN {}. Expecting image to be: {:?}",
5404 6 : query_lsn, expected_item
5405 6 : );
5406 6 :
5407 90 : match expected_item {
5408 78 : Some((_, img_value)) => {
5409 78 : let key_results = results.expect("No vectored get error expected");
5410 78 : let key_result = &key_results[&child_gap_at_key];
5411 78 : let returned_img = key_result
5412 78 : .as_ref()
5413 78 : .expect("No page reconstruct error expected");
5414 78 :
5415 78 : info!(
5416 6 : "Vectored read at LSN {} returned image {}",
5417 0 : query_lsn,
5418 0 : std::str::from_utf8(returned_img)?
5419 6 : );
5420 78 : assert_eq!(*returned_img, test_img(img_value));
5421 6 : }
5422 6 : None => {
5423 12 : assert!(matches!(results, Err(GetVectoredError::MissingKey(_))));
5424 6 : }
5425 6 : }
5426 6 : }
5427 6 :
5428 6 : Ok(())
5429 6 : }
5430 :
5431 : #[tokio::test]
5432 6 : async fn test_random_updates() -> anyhow::Result<()> {
5433 6 : let names_algorithms = [
5434 6 : ("test_random_updates_legacy", CompactionAlgorithm::Legacy),
5435 6 : ("test_random_updates_tiered", CompactionAlgorithm::Tiered),
5436 6 : ];
5437 18 : for (name, algorithm) in names_algorithms {
5438 289215 : test_random_updates_algorithm(name, algorithm).await?;
5439 6 : }
5440 6 : Ok(())
5441 6 : }
5442 :
5443 12 : async fn test_random_updates_algorithm(
5444 12 : name: &'static str,
5445 12 : compaction_algorithm: CompactionAlgorithm,
5446 12 : ) -> anyhow::Result<()> {
5447 12 : let mut harness = TenantHarness::create(name).await?;
5448 12 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
5449 12 : kind: compaction_algorithm,
5450 12 : };
5451 48 : let (tenant, ctx) = harness.load().await;
5452 12 : let tline = tenant
5453 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5454 21 : .await?;
5455 :
5456 : const NUM_KEYS: usize = 1000;
5457 12 : let cancel = CancellationToken::new();
5458 12 :
5459 12 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5460 12 : let mut test_key_end = test_key;
5461 12 : test_key_end.field6 = NUM_KEYS as u32;
5462 12 : tline.add_extra_test_dense_keyspace(KeySpace::single(test_key..test_key_end));
5463 12 :
5464 12 : let mut keyspace = KeySpaceAccum::new();
5465 12 :
5466 12 : // Track when each page was last modified. Used to assert that
5467 12 : // a read sees the latest page version.
5468 12 : let mut updated = [Lsn(0); NUM_KEYS];
5469 12 :
5470 12 : let mut lsn = Lsn(0x10);
5471 : #[allow(clippy::needless_range_loop)]
5472 12012 : for blknum in 0..NUM_KEYS {
5473 12000 : lsn = Lsn(lsn.0 + 0x10);
5474 12000 : test_key.field6 = blknum as u32;
5475 12000 : let mut writer = tline.writer().await;
5476 12000 : writer
5477 12000 : .put(
5478 12000 : test_key,
5479 12000 : lsn,
5480 12000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5481 12000 : &ctx,
5482 12000 : )
5483 12 : .await?;
5484 12000 : writer.finish_write(lsn);
5485 12000 : updated[blknum] = lsn;
5486 12000 : drop(writer);
5487 12000 :
5488 12000 : keyspace.add_key(test_key);
5489 : }
5490 :
5491 612 : for _ in 0..50 {
5492 600600 : for _ in 0..NUM_KEYS {
5493 600000 : lsn = Lsn(lsn.0 + 0x10);
5494 600000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
5495 600000 : test_key.field6 = blknum as u32;
5496 600000 : let mut writer = tline.writer().await;
5497 600000 : writer
5498 600000 : .put(
5499 600000 : test_key,
5500 600000 : lsn,
5501 600000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5502 600000 : &ctx,
5503 600000 : )
5504 714 : .await?;
5505 600000 : writer.finish_write(lsn);
5506 600000 : drop(writer);
5507 600000 : updated[blknum] = lsn;
5508 : }
5509 :
5510 : // Read all the blocks
5511 600000 : for (blknum, last_lsn) in updated.iter().enumerate() {
5512 600000 : test_key.field6 = blknum as u32;
5513 600000 : assert_eq!(
5514 600000 : tline.get(test_key, lsn, &ctx).await?,
5515 600000 : test_img(&format!("{} at {}", blknum, last_lsn))
5516 : );
5517 : }
5518 :
5519 : // Perform a cycle of flush, and GC
5520 612 : tline.freeze_and_flush().await?;
5521 600 : tenant
5522 600 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
5523 578 : .await?;
5524 : }
5525 :
5526 12 : Ok(())
5527 12 : }
5528 :
5529 : #[tokio::test]
5530 6 : async fn test_traverse_branches() -> anyhow::Result<()> {
5531 6 : let (tenant, ctx) = TenantHarness::create("test_traverse_branches")
5532 6 : .await?
5533 6 : .load()
5534 21 : .await;
5535 6 : let mut tline = tenant
5536 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5537 12 : .await?;
5538 6 :
5539 6 : const NUM_KEYS: usize = 1000;
5540 6 :
5541 6 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5542 6 :
5543 6 : let mut keyspace = KeySpaceAccum::new();
5544 6 :
5545 6 : let cancel = CancellationToken::new();
5546 6 :
5547 6 : // Track when each page was last modified. Used to assert that
5548 6 : // a read sees the latest page version.
5549 6 : let mut updated = [Lsn(0); NUM_KEYS];
5550 6 :
5551 6 : let mut lsn = Lsn(0x10);
5552 6 : #[allow(clippy::needless_range_loop)]
5553 6006 : for blknum in 0..NUM_KEYS {
5554 6000 : lsn = Lsn(lsn.0 + 0x10);
5555 6000 : test_key.field6 = blknum as u32;
5556 6000 : let mut writer = tline.writer().await;
5557 6000 : writer
5558 6000 : .put(
5559 6000 : test_key,
5560 6000 : lsn,
5561 6000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5562 6000 : &ctx,
5563 6000 : )
5564 6 : .await?;
5565 6000 : writer.finish_write(lsn);
5566 6000 : updated[blknum] = lsn;
5567 6000 : drop(writer);
5568 6000 :
5569 6000 : keyspace.add_key(test_key);
5570 6 : }
5571 6 :
5572 306 : for _ in 0..50 {
5573 300 : let new_tline_id = TimelineId::generate();
5574 300 : tenant
5575 300 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
5576 6 : .await?;
5577 300 : tline = tenant
5578 300 : .get_timeline(new_tline_id, true)
5579 300 : .expect("Should have the branched timeline");
5580 6 :
5581 300300 : for _ in 0..NUM_KEYS {
5582 300000 : lsn = Lsn(lsn.0 + 0x10);
5583 300000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
5584 300000 : test_key.field6 = blknum as u32;
5585 300000 : let mut writer = tline.writer().await;
5586 300000 : writer
5587 300000 : .put(
5588 300000 : test_key,
5589 300000 : lsn,
5590 300000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5591 300000 : &ctx,
5592 300000 : )
5593 2469 : .await?;
5594 300000 : println!("updating {} at {}", blknum, lsn);
5595 300000 : writer.finish_write(lsn);
5596 300000 : drop(writer);
5597 300000 : updated[blknum] = lsn;
5598 6 : }
5599 6 :
5600 6 : // Read all the blocks
5601 300000 : for (blknum, last_lsn) in updated.iter().enumerate() {
5602 300000 : test_key.field6 = blknum as u32;
5603 300000 : assert_eq!(
5604 300000 : tline.get(test_key, lsn, &ctx).await?,
5605 300000 : test_img(&format!("{} at {}", blknum, last_lsn))
5606 6 : );
5607 6 : }
5608 6 :
5609 6 : // Perform a cycle of flush, compact, and GC
5610 307 : tline.freeze_and_flush().await?;
5611 45346 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
5612 300 : tenant
5613 300 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
5614 260 : .await?;
5615 6 : }
5616 6 :
5617 6 : Ok(())
5618 6 : }
5619 :
5620 : #[tokio::test]
5621 6 : async fn test_traverse_ancestors() -> anyhow::Result<()> {
5622 6 : let (tenant, ctx) = TenantHarness::create("test_traverse_ancestors")
5623 6 : .await?
5624 6 : .load()
5625 24 : .await;
5626 6 : let mut tline = tenant
5627 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5628 12 : .await?;
5629 6 :
5630 6 : const NUM_KEYS: usize = 100;
5631 6 : const NUM_TLINES: usize = 50;
5632 6 :
5633 6 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5634 6 : // Track page mutation lsns across different timelines.
5635 6 : let mut updated = [[Lsn(0); NUM_KEYS]; NUM_TLINES];
5636 6 :
5637 6 : let mut lsn = Lsn(0x10);
5638 6 :
5639 6 : #[allow(clippy::needless_range_loop)]
5640 306 : for idx in 0..NUM_TLINES {
5641 300 : let new_tline_id = TimelineId::generate();
5642 300 : tenant
5643 300 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
5644 6 : .await?;
5645 300 : tline = tenant
5646 300 : .get_timeline(new_tline_id, true)
5647 300 : .expect("Should have the branched timeline");
5648 6 :
5649 30300 : for _ in 0..NUM_KEYS {
5650 30000 : lsn = Lsn(lsn.0 + 0x10);
5651 30000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
5652 30000 : test_key.field6 = blknum as u32;
5653 30000 : let mut writer = tline.writer().await;
5654 30000 : writer
5655 30000 : .put(
5656 30000 : test_key,
5657 30000 : lsn,
5658 30000 : &Value::Image(test_img(&format!("{} {} at {}", idx, blknum, lsn))),
5659 30000 : &ctx,
5660 30000 : )
5661 264 : .await?;
5662 30000 : println!("updating [{}][{}] at {}", idx, blknum, lsn);
5663 30000 : writer.finish_write(lsn);
5664 30000 : drop(writer);
5665 30000 : updated[idx][blknum] = lsn;
5666 6 : }
5667 6 : }
5668 6 :
5669 6 : // Read pages from leaf timeline across all ancestors.
5670 300 : for (idx, lsns) in updated.iter().enumerate() {
5671 30000 : for (blknum, lsn) in lsns.iter().enumerate() {
5672 6 : // Skip empty mutations.
5673 30000 : if lsn.0 == 0 {
5674 10968 : continue;
5675 19032 : }
5676 19032 : println!("checking [{idx}][{blknum}] at {lsn}");
5677 19032 : test_key.field6 = blknum as u32;
5678 19032 : assert_eq!(
5679 19032 : tline.get(test_key, *lsn, &ctx).await?,
5680 19032 : test_img(&format!("{idx} {blknum} at {lsn}"))
5681 6 : );
5682 6 : }
5683 6 : }
5684 6 : Ok(())
5685 6 : }
5686 :
5687 : #[tokio::test]
5688 6 : async fn test_write_at_initdb_lsn_takes_optimization_code_path() -> anyhow::Result<()> {
5689 6 : let (tenant, ctx) = TenantHarness::create("test_empty_test_timeline_is_usable")
5690 6 : .await?
5691 6 : .load()
5692 24 : .await;
5693 6 :
5694 6 : let initdb_lsn = Lsn(0x20);
5695 6 : let utline = tenant
5696 6 : .create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)
5697 6 : .await?;
5698 6 : let tline = utline.raw_timeline().unwrap();
5699 6 :
5700 6 : // Spawn flush loop now so that we can set the `expect_initdb_optimization`
5701 6 : tline.maybe_spawn_flush_loop();
5702 6 :
5703 6 : // Make sure the timeline has the minimum set of required keys for operation.
5704 6 : // The only operation you can always do on an empty timeline is to `put` new data.
5705 6 : // Except if you `put` at `initdb_lsn`.
5706 6 : // In that case, there's an optimization to directly create image layers instead of delta layers.
5707 6 : // It uses `repartition()`, which assumes some keys to be present.
5708 6 : // Let's make sure the test timeline can handle that case.
5709 6 : {
5710 6 : let mut state = tline.flush_loop_state.lock().unwrap();
5711 6 : assert_eq!(
5712 6 : timeline::FlushLoopState::Running {
5713 6 : expect_initdb_optimization: false,
5714 6 : initdb_optimization_count: 0,
5715 6 : },
5716 6 : *state
5717 6 : );
5718 6 : *state = timeline::FlushLoopState::Running {
5719 6 : expect_initdb_optimization: true,
5720 6 : initdb_optimization_count: 0,
5721 6 : };
5722 6 : }
5723 6 :
5724 6 : // Make writes at the initdb_lsn. When we flush it below, it should be handled by the optimization.
5725 6 : // As explained above, the optimization requires some keys to be present.
5726 6 : // As per `create_empty_timeline` documentation, use init_empty to set them.
5727 6 : // This is what `create_test_timeline` does, by the way.
5728 6 : let mut modification = tline.begin_modification(initdb_lsn);
5729 6 : modification
5730 6 : .init_empty_test_timeline()
5731 6 : .context("init_empty_test_timeline")?;
5732 6 : modification
5733 6 : .commit(&ctx)
5734 6 : .await
5735 6 : .context("commit init_empty_test_timeline modification")?;
5736 6 :
5737 6 : // Do the flush. The flush code will check the expectations that we set above.
5738 6 : tline.freeze_and_flush().await?;
5739 6 :
5740 6 : // assert freeze_and_flush exercised the initdb optimization
5741 6 : {
5742 6 : let state = tline.flush_loop_state.lock().unwrap();
5743 6 : let timeline::FlushLoopState::Running {
5744 6 : expect_initdb_optimization,
5745 6 : initdb_optimization_count,
5746 6 : } = *state
5747 6 : else {
5748 6 : panic!("unexpected state: {:?}", *state);
5749 6 : };
5750 6 : assert!(expect_initdb_optimization);
5751 6 : assert!(initdb_optimization_count > 0);
5752 6 : }
5753 6 : Ok(())
5754 6 : }
5755 :
5756 : #[tokio::test]
5757 6 : async fn test_create_guard_crash() -> anyhow::Result<()> {
5758 6 : let name = "test_create_guard_crash";
5759 6 : let harness = TenantHarness::create(name).await?;
5760 6 : {
5761 24 : let (tenant, ctx) = harness.load().await;
5762 6 : let tline = tenant
5763 6 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
5764 6 : .await?;
5765 6 : // Leave the timeline ID in [`Tenant::timelines_creating`] to exclude attempting to create it again
5766 6 : let raw_tline = tline.raw_timeline().unwrap();
5767 6 : raw_tline
5768 6 : .shutdown(super::timeline::ShutdownMode::Hard)
5769 6 : .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))
5770 6 : .await;
5771 6 : std::mem::forget(tline);
5772 6 : }
5773 6 :
5774 23 : let (tenant, _) = harness.load().await;
5775 6 : match tenant.get_timeline(TIMELINE_ID, false) {
5776 6 : Ok(_) => panic!("timeline should've been removed during load"),
5777 6 : Err(e) => {
5778 6 : assert_eq!(
5779 6 : e,
5780 6 : GetTimelineError::NotFound {
5781 6 : tenant_id: tenant.tenant_shard_id,
5782 6 : timeline_id: TIMELINE_ID,
5783 6 : }
5784 6 : )
5785 6 : }
5786 6 : }
5787 6 :
5788 6 : assert!(!harness
5789 6 : .conf
5790 6 : .timeline_path(&tenant.tenant_shard_id, &TIMELINE_ID)
5791 6 : .exists());
5792 6 :
5793 6 : Ok(())
5794 6 : }
5795 :
5796 : #[tokio::test]
5797 6 : async fn test_read_at_max_lsn() -> anyhow::Result<()> {
5798 6 : let names_algorithms = [
5799 6 : ("test_read_at_max_lsn_legacy", CompactionAlgorithm::Legacy),
5800 6 : ("test_read_at_max_lsn_tiered", CompactionAlgorithm::Tiered),
5801 6 : ];
5802 18 : for (name, algorithm) in names_algorithms {
5803 98198 : test_read_at_max_lsn_algorithm(name, algorithm).await?;
5804 6 : }
5805 6 : Ok(())
5806 6 : }
5807 :
5808 12 : async fn test_read_at_max_lsn_algorithm(
5809 12 : name: &'static str,
5810 12 : compaction_algorithm: CompactionAlgorithm,
5811 12 : ) -> anyhow::Result<()> {
5812 12 : let mut harness = TenantHarness::create(name).await?;
5813 12 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
5814 12 : kind: compaction_algorithm,
5815 12 : };
5816 48 : let (tenant, ctx) = harness.load().await;
5817 12 : let tline = tenant
5818 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5819 21 : .await?;
5820 :
5821 12 : let lsn = Lsn(0x10);
5822 12 : let compact = false;
5823 96899 : bulk_insert_maybe_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000, compact).await?;
5824 :
5825 12 : let test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5826 12 : let read_lsn = Lsn(u64::MAX - 1);
5827 :
5828 1230 : let result = tline.get(test_key, read_lsn, &ctx).await;
5829 12 : assert!(result.is_ok(), "result is not Ok: {}", result.unwrap_err());
5830 :
5831 12 : Ok(())
5832 12 : }
5833 :
5834 : #[tokio::test]
5835 6 : async fn test_metadata_scan() -> anyhow::Result<()> {
5836 6 : let harness = TenantHarness::create("test_metadata_scan").await?;
5837 24 : let (tenant, ctx) = harness.load().await;
5838 6 : let tline = tenant
5839 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5840 12 : .await?;
5841 6 :
5842 6 : const NUM_KEYS: usize = 1000;
5843 6 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
5844 6 :
5845 6 : let cancel = CancellationToken::new();
5846 6 :
5847 6 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
5848 6 : base_key.field1 = AUX_KEY_PREFIX;
5849 6 : let mut test_key = base_key;
5850 6 :
5851 6 : // Track when each page was last modified. Used to assert that
5852 6 : // a read sees the latest page version.
5853 6 : let mut updated = [Lsn(0); NUM_KEYS];
5854 6 :
5855 6 : let mut lsn = Lsn(0x10);
5856 6 : #[allow(clippy::needless_range_loop)]
5857 6006 : for blknum in 0..NUM_KEYS {
5858 6000 : lsn = Lsn(lsn.0 + 0x10);
5859 6000 : test_key.field6 = (blknum * STEP) as u32;
5860 6000 : let mut writer = tline.writer().await;
5861 6000 : writer
5862 6000 : .put(
5863 6000 : test_key,
5864 6000 : lsn,
5865 6000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5866 6000 : &ctx,
5867 6000 : )
5868 6 : .await?;
5869 6000 : writer.finish_write(lsn);
5870 6000 : updated[blknum] = lsn;
5871 6000 : drop(writer);
5872 6 : }
5873 6 :
5874 6 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
5875 6 :
5876 72 : for iter in 0..=10 {
5877 6 : // Read all the blocks
5878 66000 : for (blknum, last_lsn) in updated.iter().enumerate() {
5879 66000 : test_key.field6 = (blknum * STEP) as u32;
5880 66000 : assert_eq!(
5881 66000 : tline.get(test_key, lsn, &ctx).await?,
5882 66000 : test_img(&format!("{} at {}", blknum, last_lsn))
5883 6 : );
5884 6 : }
5885 6 :
5886 66 : let mut cnt = 0;
5887 66000 : for (key, value) in tline
5888 66 : .get_vectored_impl(
5889 66 : keyspace.clone(),
5890 66 : lsn,
5891 66 : &mut ValuesReconstructState::default(),
5892 66 : &ctx,
5893 66 : )
5894 12568 : .await?
5895 6 : {
5896 66000 : let blknum = key.field6 as usize;
5897 66000 : let value = value?;
5898 66000 : assert!(blknum % STEP == 0);
5899 66000 : let blknum = blknum / STEP;
5900 66000 : assert_eq!(
5901 66000 : value,
5902 66000 : test_img(&format!("{} at {}", blknum, updated[blknum]))
5903 66000 : );
5904 66000 : cnt += 1;
5905 6 : }
5906 6 :
5907 66 : assert_eq!(cnt, NUM_KEYS);
5908 6 :
5909 66066 : for _ in 0..NUM_KEYS {
5910 66000 : lsn = Lsn(lsn.0 + 0x10);
5911 66000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
5912 66000 : test_key.field6 = (blknum * STEP) as u32;
5913 66000 : let mut writer = tline.writer().await;
5914 66000 : writer
5915 66000 : .put(
5916 66000 : test_key,
5917 66000 : lsn,
5918 66000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5919 66000 : &ctx,
5920 66000 : )
5921 468 : .await?;
5922 66000 : writer.finish_write(lsn);
5923 66000 : drop(writer);
5924 66000 : updated[blknum] = lsn;
5925 6 : }
5926 6 :
5927 6 : // Perform two cycles of flush, compact, and GC
5928 198 : for round in 0..2 {
5929 132 : tline.freeze_and_flush().await?;
5930 132 : tline
5931 132 : .compact(
5932 132 : &cancel,
5933 132 : if iter % 5 == 0 && round == 0 {
5934 18 : let mut flags = EnumSet::new();
5935 18 : flags.insert(CompactFlags::ForceImageLayerCreation);
5936 18 : flags.insert(CompactFlags::ForceRepartition);
5937 18 : flags
5938 6 : } else {
5939 114 : EnumSet::empty()
5940 6 : },
5941 132 : &ctx,
5942 6 : )
5943 25246 : .await?;
5944 132 : tenant
5945 132 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
5946 108 : .await?;
5947 6 : }
5948 6 : }
5949 6 :
5950 6 : Ok(())
5951 6 : }
5952 :
5953 : #[tokio::test]
5954 6 : async fn test_metadata_compaction_trigger() -> anyhow::Result<()> {
5955 6 : let harness = TenantHarness::create("test_metadata_compaction_trigger").await?;
5956 21 : let (tenant, ctx) = harness.load().await;
5957 6 : let tline = tenant
5958 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5959 12 : .await?;
5960 6 :
5961 6 : let cancel = CancellationToken::new();
5962 6 :
5963 6 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
5964 6 : base_key.field1 = AUX_KEY_PREFIX;
5965 6 : let test_key = base_key;
5966 6 : let mut lsn = Lsn(0x10);
5967 6 :
5968 126 : for _ in 0..20 {
5969 120 : lsn = Lsn(lsn.0 + 0x10);
5970 120 : let mut writer = tline.writer().await;
5971 120 : writer
5972 120 : .put(
5973 120 : test_key,
5974 120 : lsn,
5975 120 : &Value::Image(test_img(&format!("{} at {}", 0, lsn))),
5976 120 : &ctx,
5977 120 : )
5978 60 : .await?;
5979 120 : writer.finish_write(lsn);
5980 120 : drop(writer);
5981 126 : tline.freeze_and_flush().await?; // force create a delta layer
5982 6 : }
5983 6 :
5984 6 : let before_num_l0_delta_files =
5985 6 : tline.layers.read().await.layer_map()?.level0_deltas().len();
5986 6 :
5987 450 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
5988 6 :
5989 6 : let after_num_l0_delta_files = tline.layers.read().await.layer_map()?.level0_deltas().len();
5990 6 :
5991 6 : 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}");
5992 6 :
5993 6 : assert_eq!(
5994 12 : tline.get(test_key, lsn, &ctx).await?,
5995 6 : test_img(&format!("{} at {}", 0, lsn))
5996 6 : );
5997 6 :
5998 6 : Ok(())
5999 6 : }
6000 :
6001 : #[tokio::test]
6002 6 : async fn test_branch_copies_dirty_aux_file_flag() {
6003 6 : let harness = TenantHarness::create("test_branch_copies_dirty_aux_file_flag")
6004 6 : .await
6005 6 : .unwrap();
6006 6 :
6007 6 : // the default aux file policy to switch is v2 if not set by the admins
6008 6 : assert_eq!(
6009 6 : harness.tenant_conf.switch_aux_file_policy,
6010 6 : AuxFilePolicy::default_tenant_config()
6011 6 : );
6012 24 : let (tenant, ctx) = harness.load().await;
6013 6 :
6014 6 : let mut lsn = Lsn(0x08);
6015 6 :
6016 6 : let tline: Arc<Timeline> = tenant
6017 6 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
6018 12 : .await
6019 6 : .unwrap();
6020 6 :
6021 6 : // no aux file is written at this point, so the persistent flag should be unset
6022 6 : assert_eq!(tline.last_aux_file_policy.load(), None);
6023 6 :
6024 6 : {
6025 6 : lsn += 8;
6026 6 : let mut modification = tline.begin_modification(lsn);
6027 6 : modification
6028 6 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
6029 6 : .await
6030 6 : .unwrap();
6031 6 : modification.commit(&ctx).await.unwrap();
6032 6 : }
6033 6 :
6034 6 : // there is no tenant manager to pass the configuration through, so lets mimic it
6035 6 : tenant.set_new_location_config(
6036 6 : AttachedTenantConf::try_from(LocationConf::attached_single(
6037 6 : TenantConfOpt {
6038 6 : switch_aux_file_policy: Some(AuxFilePolicy::V2),
6039 6 : ..Default::default()
6040 6 : },
6041 6 : tenant.generation,
6042 6 : &pageserver_api::models::ShardParameters::default(),
6043 6 : ))
6044 6 : .unwrap(),
6045 6 : );
6046 6 :
6047 6 : assert_eq!(
6048 6 : tline.get_switch_aux_file_policy(),
6049 6 : AuxFilePolicy::V2,
6050 6 : "wanted state has been updated"
6051 6 : );
6052 6 : assert_eq!(
6053 6 : tline.last_aux_file_policy.load(),
6054 6 : Some(AuxFilePolicy::V2),
6055 6 : "aux file is written with switch_aux_file_policy unset (which is v2), so we should use v2 there"
6056 6 : );
6057 6 :
6058 6 : // we can read everything from the storage
6059 6 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6060 6 : assert_eq!(
6061 6 : files.get("pg_logical/mappings/test1"),
6062 6 : Some(&bytes::Bytes::from_static(b"first"))
6063 6 : );
6064 6 :
6065 6 : {
6066 6 : lsn += 8;
6067 6 : let mut modification = tline.begin_modification(lsn);
6068 6 : modification
6069 6 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
6070 6 : .await
6071 6 : .unwrap();
6072 6 : modification.commit(&ctx).await.unwrap();
6073 6 : }
6074 6 :
6075 6 : assert_eq!(
6076 6 : tline.last_aux_file_policy.load(),
6077 6 : Some(AuxFilePolicy::V2),
6078 6 : "keep v2 storage format when new files are written"
6079 6 : );
6080 6 :
6081 6 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6082 6 : assert_eq!(
6083 6 : files.get("pg_logical/mappings/test2"),
6084 6 : Some(&bytes::Bytes::from_static(b"second"))
6085 6 : );
6086 6 :
6087 6 : let child = tenant
6088 6 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(lsn), &ctx)
6089 6 : .await
6090 6 : .unwrap();
6091 6 :
6092 6 : // child copies the last flag even if that is not on remote storage yet
6093 6 : assert_eq!(child.get_switch_aux_file_policy(), AuxFilePolicy::V2);
6094 6 : assert_eq!(child.last_aux_file_policy.load(), Some(AuxFilePolicy::V2));
6095 6 :
6096 6 : let files = child.list_aux_files(lsn, &ctx).await.unwrap();
6097 6 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
6098 6 : assert_eq!(files.get("pg_logical/mappings/test2"), None);
6099 6 :
6100 6 : // even if we crash here without flushing parent timeline with it's new
6101 6 : // last_aux_file_policy we are safe, because child was never meant to access ancestor's
6102 6 : // files. the ancestor can even switch back to V1 because of a migration safely.
6103 6 : }
6104 :
6105 : #[tokio::test]
6106 6 : async fn aux_file_policy_switch() {
6107 6 : let mut harness = TenantHarness::create("aux_file_policy_switch")
6108 6 : .await
6109 6 : .unwrap();
6110 6 : harness.tenant_conf.switch_aux_file_policy = AuxFilePolicy::CrossValidation; // set to cross-validation mode
6111 24 : let (tenant, ctx) = harness.load().await;
6112 6 :
6113 6 : let mut lsn = Lsn(0x08);
6114 6 :
6115 6 : let tline: Arc<Timeline> = tenant
6116 6 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
6117 12 : .await
6118 6 : .unwrap();
6119 6 :
6120 6 : assert_eq!(
6121 6 : tline.last_aux_file_policy.load(),
6122 6 : None,
6123 6 : "no aux file is written so it should be unset"
6124 6 : );
6125 6 :
6126 6 : {
6127 6 : lsn += 8;
6128 6 : let mut modification = tline.begin_modification(lsn);
6129 6 : modification
6130 6 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
6131 15 : .await
6132 6 : .unwrap();
6133 6 : modification.commit(&ctx).await.unwrap();
6134 6 : }
6135 6 :
6136 6 : // there is no tenant manager to pass the configuration through, so lets mimic it
6137 6 : tenant.set_new_location_config(
6138 6 : AttachedTenantConf::try_from(LocationConf::attached_single(
6139 6 : TenantConfOpt {
6140 6 : switch_aux_file_policy: Some(AuxFilePolicy::V2),
6141 6 : ..Default::default()
6142 6 : },
6143 6 : tenant.generation,
6144 6 : &pageserver_api::models::ShardParameters::default(),
6145 6 : ))
6146 6 : .unwrap(),
6147 6 : );
6148 6 :
6149 6 : assert_eq!(
6150 6 : tline.get_switch_aux_file_policy(),
6151 6 : AuxFilePolicy::V2,
6152 6 : "wanted state has been updated"
6153 6 : );
6154 6 : assert_eq!(
6155 6 : tline.last_aux_file_policy.load(),
6156 6 : Some(AuxFilePolicy::CrossValidation),
6157 6 : "dirty index_part.json reflected state is yet to be updated"
6158 6 : );
6159 6 :
6160 6 : // we can still read the auxfile v1 before we ingest anything new
6161 6 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6162 6 : assert_eq!(
6163 6 : files.get("pg_logical/mappings/test1"),
6164 6 : Some(&bytes::Bytes::from_static(b"first"))
6165 6 : );
6166 6 :
6167 6 : {
6168 6 : lsn += 8;
6169 6 : let mut modification = tline.begin_modification(lsn);
6170 6 : modification
6171 6 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
6172 6 : .await
6173 6 : .unwrap();
6174 6 : modification.commit(&ctx).await.unwrap();
6175 6 : }
6176 6 :
6177 6 : assert_eq!(
6178 6 : tline.last_aux_file_policy.load(),
6179 6 : Some(AuxFilePolicy::V2),
6180 6 : "ingesting a file should apply the wanted switch state when applicable"
6181 6 : );
6182 6 :
6183 6 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6184 6 : assert_eq!(
6185 6 : files.get("pg_logical/mappings/test1"),
6186 6 : Some(&bytes::Bytes::from_static(b"first")),
6187 6 : "cross validation writes to both v1 and v2 so this should be available in v2"
6188 6 : );
6189 6 : assert_eq!(
6190 6 : files.get("pg_logical/mappings/test2"),
6191 6 : Some(&bytes::Bytes::from_static(b"second"))
6192 6 : );
6193 6 :
6194 6 : // mimic again by trying to flip it from V2 to V1 (not switched to while ingesting a file)
6195 6 : tenant.set_new_location_config(
6196 6 : AttachedTenantConf::try_from(LocationConf::attached_single(
6197 6 : TenantConfOpt {
6198 6 : switch_aux_file_policy: Some(AuxFilePolicy::V1),
6199 6 : ..Default::default()
6200 6 : },
6201 6 : tenant.generation,
6202 6 : &pageserver_api::models::ShardParameters::default(),
6203 6 : ))
6204 6 : .unwrap(),
6205 6 : );
6206 6 :
6207 6 : {
6208 6 : lsn += 8;
6209 6 : let mut modification = tline.begin_modification(lsn);
6210 6 : modification
6211 6 : .put_file("pg_logical/mappings/test2", b"third", &ctx)
6212 6 : .await
6213 6 : .unwrap();
6214 6 : modification.commit(&ctx).await.unwrap();
6215 6 : }
6216 6 :
6217 6 : assert_eq!(
6218 6 : tline.get_switch_aux_file_policy(),
6219 6 : AuxFilePolicy::V1,
6220 6 : "wanted state has been updated again, even if invalid request"
6221 6 : );
6222 6 :
6223 6 : assert_eq!(
6224 6 : tline.last_aux_file_policy.load(),
6225 6 : Some(AuxFilePolicy::V2),
6226 6 : "ingesting a file should apply the wanted switch state when applicable"
6227 6 : );
6228 6 :
6229 6 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6230 6 : assert_eq!(
6231 6 : files.get("pg_logical/mappings/test1"),
6232 6 : Some(&bytes::Bytes::from_static(b"first"))
6233 6 : );
6234 6 : assert_eq!(
6235 6 : files.get("pg_logical/mappings/test2"),
6236 6 : Some(&bytes::Bytes::from_static(b"third"))
6237 6 : );
6238 6 :
6239 6 : // mimic again by trying to flip it from from V1 to V2 (not switched to while ingesting a file)
6240 6 : tenant.set_new_location_config(
6241 6 : AttachedTenantConf::try_from(LocationConf::attached_single(
6242 6 : TenantConfOpt {
6243 6 : switch_aux_file_policy: Some(AuxFilePolicy::V2),
6244 6 : ..Default::default()
6245 6 : },
6246 6 : tenant.generation,
6247 6 : &pageserver_api::models::ShardParameters::default(),
6248 6 : ))
6249 6 : .unwrap(),
6250 6 : );
6251 6 :
6252 6 : {
6253 6 : lsn += 8;
6254 6 : let mut modification = tline.begin_modification(lsn);
6255 6 : modification
6256 6 : .put_file("pg_logical/mappings/test3", b"last", &ctx)
6257 6 : .await
6258 6 : .unwrap();
6259 6 : modification.commit(&ctx).await.unwrap();
6260 6 : }
6261 6 :
6262 6 : assert_eq!(tline.get_switch_aux_file_policy(), AuxFilePolicy::V2);
6263 6 :
6264 6 : assert_eq!(tline.last_aux_file_policy.load(), Some(AuxFilePolicy::V2));
6265 6 :
6266 6 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6267 6 : assert_eq!(
6268 6 : files.get("pg_logical/mappings/test1"),
6269 6 : Some(&bytes::Bytes::from_static(b"first"))
6270 6 : );
6271 6 : assert_eq!(
6272 6 : files.get("pg_logical/mappings/test2"),
6273 6 : Some(&bytes::Bytes::from_static(b"third"))
6274 6 : );
6275 6 : assert_eq!(
6276 6 : files.get("pg_logical/mappings/test3"),
6277 6 : Some(&bytes::Bytes::from_static(b"last"))
6278 6 : );
6279 6 : }
6280 :
6281 : #[tokio::test]
6282 6 : async fn aux_file_policy_force_switch() {
6283 6 : let mut harness = TenantHarness::create("aux_file_policy_force_switch")
6284 6 : .await
6285 6 : .unwrap();
6286 6 : harness.tenant_conf.switch_aux_file_policy = AuxFilePolicy::V1;
6287 24 : let (tenant, ctx) = harness.load().await;
6288 6 :
6289 6 : let mut lsn = Lsn(0x08);
6290 6 :
6291 6 : let tline: Arc<Timeline> = tenant
6292 6 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
6293 12 : .await
6294 6 : .unwrap();
6295 6 :
6296 6 : assert_eq!(
6297 6 : tline.last_aux_file_policy.load(),
6298 6 : None,
6299 6 : "no aux file is written so it should be unset"
6300 6 : );
6301 6 :
6302 6 : {
6303 6 : lsn += 8;
6304 6 : let mut modification = tline.begin_modification(lsn);
6305 6 : modification
6306 6 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
6307 15 : .await
6308 6 : .unwrap();
6309 6 : modification.commit(&ctx).await.unwrap();
6310 6 : }
6311 6 :
6312 6 : tline.do_switch_aux_policy(AuxFilePolicy::V2).unwrap();
6313 6 :
6314 6 : assert_eq!(
6315 6 : tline.last_aux_file_policy.load(),
6316 6 : Some(AuxFilePolicy::V2),
6317 6 : "dirty index_part.json reflected state is yet to be updated"
6318 6 : );
6319 6 :
6320 6 : // lose all data from v1
6321 6 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6322 6 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
6323 6 :
6324 6 : {
6325 6 : lsn += 8;
6326 6 : let mut modification = tline.begin_modification(lsn);
6327 6 : modification
6328 6 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
6329 6 : .await
6330 6 : .unwrap();
6331 6 : modification.commit(&ctx).await.unwrap();
6332 6 : }
6333 6 :
6334 6 : // read data ingested in v2
6335 6 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6336 6 : assert_eq!(
6337 6 : files.get("pg_logical/mappings/test2"),
6338 6 : Some(&bytes::Bytes::from_static(b"second"))
6339 6 : );
6340 6 : // lose all data from v1
6341 6 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
6342 6 : }
6343 :
6344 : #[tokio::test]
6345 6 : async fn aux_file_policy_auto_detect() {
6346 6 : let mut harness = TenantHarness::create("aux_file_policy_auto_detect")
6347 6 : .await
6348 6 : .unwrap();
6349 6 : harness.tenant_conf.switch_aux_file_policy = AuxFilePolicy::V2; // set to cross-validation mode
6350 24 : let (tenant, ctx) = harness.load().await;
6351 6 :
6352 6 : let mut lsn = Lsn(0x08);
6353 6 :
6354 6 : let tline: Arc<Timeline> = tenant
6355 6 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
6356 12 : .await
6357 6 : .unwrap();
6358 6 :
6359 6 : assert_eq!(
6360 6 : tline.last_aux_file_policy.load(),
6361 6 : None,
6362 6 : "no aux file is written so it should be unset"
6363 6 : );
6364 6 :
6365 6 : {
6366 6 : lsn += 8;
6367 6 : let mut modification = tline.begin_modification(lsn);
6368 6 : let buf = AuxFilesDirectory::ser(&AuxFilesDirectory {
6369 6 : files: vec![(
6370 6 : "test_file".to_string(),
6371 6 : Bytes::copy_from_slice(b"test_file"),
6372 6 : )]
6373 6 : .into_iter()
6374 6 : .collect(),
6375 6 : })
6376 6 : .unwrap();
6377 6 : modification.put_for_test(AUX_FILES_KEY, Value::Image(Bytes::from(buf)));
6378 6 : modification.commit(&ctx).await.unwrap();
6379 6 : }
6380 6 :
6381 6 : {
6382 6 : lsn += 8;
6383 6 : let mut modification = tline.begin_modification(lsn);
6384 6 : modification
6385 6 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
6386 6 : .await
6387 6 : .unwrap();
6388 6 : modification.commit(&ctx).await.unwrap();
6389 6 : }
6390 6 :
6391 6 : assert_eq!(
6392 6 : tline.last_aux_file_policy.load(),
6393 6 : Some(AuxFilePolicy::V1),
6394 6 : "keep using v1 because there are aux files writting with v1"
6395 6 : );
6396 6 :
6397 6 : // we can still read the auxfile v1
6398 6 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6399 6 : assert_eq!(
6400 6 : files.get("pg_logical/mappings/test1"),
6401 6 : Some(&bytes::Bytes::from_static(b"first"))
6402 6 : );
6403 6 : assert_eq!(
6404 6 : files.get("test_file"),
6405 6 : Some(&bytes::Bytes::from_static(b"test_file"))
6406 6 : );
6407 6 : }
6408 :
6409 : #[tokio::test]
6410 6 : async fn test_metadata_image_creation() -> anyhow::Result<()> {
6411 6 : let harness = TenantHarness::create("test_metadata_image_creation").await?;
6412 24 : let (tenant, ctx) = harness.load().await;
6413 6 : let tline = tenant
6414 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6415 12 : .await?;
6416 6 :
6417 6 : const NUM_KEYS: usize = 1000;
6418 6 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
6419 6 :
6420 6 : let cancel = CancellationToken::new();
6421 6 :
6422 6 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
6423 6 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
6424 6 : let mut test_key = base_key;
6425 6 : let mut lsn = Lsn(0x10);
6426 6 :
6427 24 : async fn scan_with_statistics(
6428 24 : tline: &Timeline,
6429 24 : keyspace: &KeySpace,
6430 24 : lsn: Lsn,
6431 24 : ctx: &RequestContext,
6432 24 : ) -> anyhow::Result<(BTreeMap<Key, Result<Bytes, PageReconstructError>>, usize)> {
6433 24 : let mut reconstruct_state = ValuesReconstructState::default();
6434 24 : let res = tline
6435 24 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
6436 4023 : .await?;
6437 24 : Ok((res, reconstruct_state.get_delta_layers_visited() as usize))
6438 24 : }
6439 6 :
6440 6 : #[allow(clippy::needless_range_loop)]
6441 6006 : for blknum in 0..NUM_KEYS {
6442 6000 : lsn = Lsn(lsn.0 + 0x10);
6443 6000 : test_key.field6 = (blknum * STEP) as u32;
6444 6000 : let mut writer = tline.writer().await;
6445 6000 : writer
6446 6000 : .put(
6447 6000 : test_key,
6448 6000 : lsn,
6449 6000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6450 6000 : &ctx,
6451 6000 : )
6452 6 : .await?;
6453 6000 : writer.finish_write(lsn);
6454 6000 : drop(writer);
6455 6 : }
6456 6 :
6457 6 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
6458 6 :
6459 66 : for iter in 1..=10 {
6460 60060 : for _ in 0..NUM_KEYS {
6461 60000 : lsn = Lsn(lsn.0 + 0x10);
6462 60000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6463 60000 : test_key.field6 = (blknum * STEP) as u32;
6464 60000 : let mut writer = tline.writer().await;
6465 60000 : writer
6466 60000 : .put(
6467 60000 : test_key,
6468 60000 : lsn,
6469 60000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6470 60000 : &ctx,
6471 60000 : )
6472 103 : .await?;
6473 60000 : writer.finish_write(lsn);
6474 60000 : drop(writer);
6475 6 : }
6476 6 :
6477 60 : tline.freeze_and_flush().await?;
6478 6 :
6479 60 : if iter % 5 == 0 {
6480 12 : let (_, before_delta_file_accessed) =
6481 3997 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
6482 12 : tline
6483 12 : .compact(
6484 12 : &cancel,
6485 12 : {
6486 12 : let mut flags = EnumSet::new();
6487 12 : flags.insert(CompactFlags::ForceImageLayerCreation);
6488 12 : flags.insert(CompactFlags::ForceRepartition);
6489 12 : flags
6490 12 : },
6491 12 : &ctx,
6492 12 : )
6493 18565 : .await?;
6494 12 : let (_, after_delta_file_accessed) =
6495 26 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
6496 12 : 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}");
6497 6 : // 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.
6498 12 : assert!(
6499 12 : after_delta_file_accessed <= 2,
6500 6 : "after_delta_file_accessed={after_delta_file_accessed}"
6501 6 : );
6502 48 : }
6503 6 : }
6504 6 :
6505 6 : Ok(())
6506 6 : }
6507 :
6508 : #[tokio::test]
6509 6 : async fn test_vectored_missing_data_key_reads() -> anyhow::Result<()> {
6510 6 : let harness = TenantHarness::create("test_vectored_missing_data_key_reads").await?;
6511 24 : let (tenant, ctx) = harness.load().await;
6512 6 :
6513 6 : let base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
6514 6 : let base_key_child = Key::from_hex("000000000033333333444444445500000001").unwrap();
6515 6 : let base_key_nonexist = Key::from_hex("000000000033333333444444445500000002").unwrap();
6516 6 :
6517 6 : let tline = tenant
6518 6 : .create_test_timeline_with_layers(
6519 6 : TIMELINE_ID,
6520 6 : Lsn(0x10),
6521 6 : DEFAULT_PG_VERSION,
6522 6 : &ctx,
6523 6 : Vec::new(), // delta layers
6524 6 : vec![(Lsn(0x20), vec![(base_key, test_img("data key 1"))])], // image layers
6525 6 : 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
6526 6 : )
6527 33 : .await?;
6528 6 : tline.add_extra_test_dense_keyspace(KeySpace::single(base_key..(base_key_nonexist.next())));
6529 6 :
6530 6 : let child = tenant
6531 6 : .branch_timeline_test_with_layers(
6532 6 : &tline,
6533 6 : NEW_TIMELINE_ID,
6534 6 : Some(Lsn(0x20)),
6535 6 : &ctx,
6536 6 : Vec::new(), // delta layers
6537 6 : vec![(Lsn(0x30), vec![(base_key_child, test_img("data key 2"))])], // image layers
6538 6 : Lsn(0x30),
6539 6 : )
6540 21 : .await
6541 6 : .unwrap();
6542 6 :
6543 6 : let lsn = Lsn(0x30);
6544 6 :
6545 6 : // test vectored get on parent timeline
6546 6 : assert_eq!(
6547 12 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
6548 6 : Some(test_img("data key 1"))
6549 6 : );
6550 6 : assert!(get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx)
6551 9 : .await
6552 6 : .unwrap_err()
6553 6 : .is_missing_key_error());
6554 6 : assert!(
6555 6 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx)
6556 6 : .await
6557 6 : .unwrap_err()
6558 6 : .is_missing_key_error()
6559 6 : );
6560 6 :
6561 6 : // test vectored get on child timeline
6562 6 : assert_eq!(
6563 6 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
6564 6 : Some(test_img("data key 1"))
6565 6 : );
6566 6 : assert_eq!(
6567 12 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
6568 6 : Some(test_img("data key 2"))
6569 6 : );
6570 6 : assert!(
6571 6 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx)
6572 6 : .await
6573 6 : .unwrap_err()
6574 6 : .is_missing_key_error()
6575 6 : );
6576 6 :
6577 6 : Ok(())
6578 6 : }
6579 :
6580 : #[tokio::test]
6581 6 : async fn test_vectored_missing_metadata_key_reads() -> anyhow::Result<()> {
6582 6 : let harness = TenantHarness::create("test_vectored_missing_metadata_key_reads").await?;
6583 24 : let (tenant, ctx) = harness.load().await;
6584 6 :
6585 6 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
6586 6 : let base_key_child = Key::from_hex("620000000033333333444444445500000001").unwrap();
6587 6 : let base_key_nonexist = Key::from_hex("620000000033333333444444445500000002").unwrap();
6588 6 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
6589 6 :
6590 6 : let tline = tenant
6591 6 : .create_test_timeline_with_layers(
6592 6 : TIMELINE_ID,
6593 6 : Lsn(0x10),
6594 6 : DEFAULT_PG_VERSION,
6595 6 : &ctx,
6596 6 : Vec::new(), // delta layers
6597 6 : vec![(Lsn(0x20), vec![(base_key, test_img("metadata key 1"))])], // image layers
6598 6 : 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
6599 6 : )
6600 33 : .await?;
6601 6 :
6602 6 : let child = tenant
6603 6 : .branch_timeline_test_with_layers(
6604 6 : &tline,
6605 6 : NEW_TIMELINE_ID,
6606 6 : Some(Lsn(0x20)),
6607 6 : &ctx,
6608 6 : Vec::new(), // delta layers
6609 6 : vec![(
6610 6 : Lsn(0x30),
6611 6 : vec![(base_key_child, test_img("metadata key 2"))],
6612 6 : )], // image layers
6613 6 : Lsn(0x30),
6614 6 : )
6615 21 : .await
6616 6 : .unwrap();
6617 6 :
6618 6 : let lsn = Lsn(0x30);
6619 6 :
6620 6 : // test vectored get on parent timeline
6621 6 : assert_eq!(
6622 12 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
6623 6 : Some(test_img("metadata key 1"))
6624 6 : );
6625 6 : assert_eq!(
6626 6 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx).await?,
6627 6 : None
6628 6 : );
6629 6 : assert_eq!(
6630 6 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx).await?,
6631 6 : None
6632 6 : );
6633 6 :
6634 6 : // test vectored get on child timeline
6635 6 : assert_eq!(
6636 6 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
6637 6 : None
6638 6 : );
6639 6 : assert_eq!(
6640 12 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
6641 6 : Some(test_img("metadata key 2"))
6642 6 : );
6643 6 : assert_eq!(
6644 6 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx).await?,
6645 6 : None
6646 6 : );
6647 6 :
6648 6 : Ok(())
6649 6 : }
6650 :
6651 108 : async fn get_vectored_impl_wrapper(
6652 108 : tline: &Arc<Timeline>,
6653 108 : key: Key,
6654 108 : lsn: Lsn,
6655 108 : ctx: &RequestContext,
6656 108 : ) -> Result<Option<Bytes>, GetVectoredError> {
6657 108 : let mut reconstruct_state = ValuesReconstructState::new();
6658 108 : let mut res = tline
6659 108 : .get_vectored_impl(
6660 108 : KeySpace::single(key..key.next()),
6661 108 : lsn,
6662 108 : &mut reconstruct_state,
6663 108 : ctx,
6664 108 : )
6665 99 : .await?;
6666 90 : Ok(res.pop_last().map(|(k, v)| {
6667 54 : assert_eq!(k, key);
6668 54 : v.unwrap()
6669 90 : }))
6670 108 : }
6671 :
6672 : #[tokio::test]
6673 6 : async fn test_metadata_tombstone_reads() -> anyhow::Result<()> {
6674 6 : let harness = TenantHarness::create("test_metadata_tombstone_reads").await?;
6675 24 : let (tenant, ctx) = harness.load().await;
6676 6 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
6677 6 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
6678 6 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
6679 6 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
6680 6 :
6681 6 : // We emulate the situation that the compaction algorithm creates an image layer that removes the tombstones
6682 6 : // Lsn 0x30 key0, key3, no key1+key2
6683 6 : // Lsn 0x20 key1+key2 tomestones
6684 6 : // Lsn 0x10 key1 in image, key2 in delta
6685 6 : let tline = tenant
6686 6 : .create_test_timeline_with_layers(
6687 6 : TIMELINE_ID,
6688 6 : Lsn(0x10),
6689 6 : DEFAULT_PG_VERSION,
6690 6 : &ctx,
6691 6 : // delta layers
6692 6 : vec![
6693 6 : DeltaLayerTestDesc::new_with_inferred_key_range(
6694 6 : Lsn(0x10)..Lsn(0x20),
6695 6 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
6696 6 : ),
6697 6 : DeltaLayerTestDesc::new_with_inferred_key_range(
6698 6 : Lsn(0x20)..Lsn(0x30),
6699 6 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
6700 6 : ),
6701 6 : DeltaLayerTestDesc::new_with_inferred_key_range(
6702 6 : Lsn(0x20)..Lsn(0x30),
6703 6 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
6704 6 : ),
6705 6 : ],
6706 6 : // image layers
6707 6 : vec![
6708 6 : (Lsn(0x10), vec![(key1, test_img("metadata key 1"))]),
6709 6 : (
6710 6 : Lsn(0x30),
6711 6 : vec![
6712 6 : (key0, test_img("metadata key 0")),
6713 6 : (key3, test_img("metadata key 3")),
6714 6 : ],
6715 6 : ),
6716 6 : ],
6717 6 : Lsn(0x30),
6718 6 : )
6719 114 : .await?;
6720 6 :
6721 6 : let lsn = Lsn(0x30);
6722 6 : let old_lsn = Lsn(0x20);
6723 6 :
6724 6 : assert_eq!(
6725 12 : get_vectored_impl_wrapper(&tline, key0, lsn, &ctx).await?,
6726 6 : Some(test_img("metadata key 0"))
6727 6 : );
6728 6 : assert_eq!(
6729 6 : get_vectored_impl_wrapper(&tline, key1, lsn, &ctx).await?,
6730 6 : None,
6731 6 : );
6732 6 : assert_eq!(
6733 6 : get_vectored_impl_wrapper(&tline, key2, lsn, &ctx).await?,
6734 6 : None,
6735 6 : );
6736 6 : assert_eq!(
6737 12 : get_vectored_impl_wrapper(&tline, key1, old_lsn, &ctx).await?,
6738 6 : Some(Bytes::new()),
6739 6 : );
6740 6 : assert_eq!(
6741 12 : get_vectored_impl_wrapper(&tline, key2, old_lsn, &ctx).await?,
6742 6 : Some(Bytes::new()),
6743 6 : );
6744 6 : assert_eq!(
6745 6 : get_vectored_impl_wrapper(&tline, key3, lsn, &ctx).await?,
6746 6 : Some(test_img("metadata key 3"))
6747 6 : );
6748 6 :
6749 6 : Ok(())
6750 6 : }
6751 :
6752 : #[tokio::test]
6753 6 : async fn test_metadata_tombstone_image_creation() {
6754 6 : let harness = TenantHarness::create("test_metadata_tombstone_image_creation")
6755 6 : .await
6756 6 : .unwrap();
6757 24 : let (tenant, ctx) = harness.load().await;
6758 6 :
6759 6 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
6760 6 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
6761 6 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
6762 6 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
6763 6 :
6764 6 : let tline = tenant
6765 6 : .create_test_timeline_with_layers(
6766 6 : TIMELINE_ID,
6767 6 : Lsn(0x10),
6768 6 : DEFAULT_PG_VERSION,
6769 6 : &ctx,
6770 6 : // delta layers
6771 6 : vec![
6772 6 : DeltaLayerTestDesc::new_with_inferred_key_range(
6773 6 : Lsn(0x10)..Lsn(0x20),
6774 6 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
6775 6 : ),
6776 6 : DeltaLayerTestDesc::new_with_inferred_key_range(
6777 6 : Lsn(0x20)..Lsn(0x30),
6778 6 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
6779 6 : ),
6780 6 : DeltaLayerTestDesc::new_with_inferred_key_range(
6781 6 : Lsn(0x20)..Lsn(0x30),
6782 6 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
6783 6 : ),
6784 6 : DeltaLayerTestDesc::new_with_inferred_key_range(
6785 6 : Lsn(0x30)..Lsn(0x40),
6786 6 : vec![
6787 6 : (key0, Lsn(0x30), Value::Image(test_img("metadata key 0"))),
6788 6 : (key3, Lsn(0x30), Value::Image(test_img("metadata key 3"))),
6789 6 : ],
6790 6 : ),
6791 6 : ],
6792 6 : // image layers
6793 6 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
6794 6 : Lsn(0x40),
6795 6 : )
6796 105 : .await
6797 6 : .unwrap();
6798 6 :
6799 6 : let cancel = CancellationToken::new();
6800 6 :
6801 6 : tline
6802 6 : .compact(
6803 6 : &cancel,
6804 6 : {
6805 6 : let mut flags = EnumSet::new();
6806 6 : flags.insert(CompactFlags::ForceImageLayerCreation);
6807 6 : flags.insert(CompactFlags::ForceRepartition);
6808 6 : flags
6809 6 : },
6810 6 : &ctx,
6811 6 : )
6812 165 : .await
6813 6 : .unwrap();
6814 6 :
6815 6 : // Image layers are created at last_record_lsn
6816 6 : let images = tline
6817 6 : .inspect_image_layers(Lsn(0x40), &ctx)
6818 24 : .await
6819 6 : .unwrap()
6820 6 : .into_iter()
6821 54 : .filter(|(k, _)| k.is_metadata_key())
6822 6 : .collect::<Vec<_>>();
6823 6 : assert_eq!(images.len(), 2); // the image layer should only contain two existing keys, tombstones should be removed.
6824 6 : }
6825 :
6826 : #[tokio::test]
6827 6 : async fn test_metadata_tombstone_empty_image_creation() {
6828 6 : let harness = TenantHarness::create("test_metadata_tombstone_empty_image_creation")
6829 6 : .await
6830 6 : .unwrap();
6831 24 : let (tenant, ctx) = harness.load().await;
6832 6 :
6833 6 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
6834 6 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
6835 6 :
6836 6 : let tline = tenant
6837 6 : .create_test_timeline_with_layers(
6838 6 : TIMELINE_ID,
6839 6 : Lsn(0x10),
6840 6 : DEFAULT_PG_VERSION,
6841 6 : &ctx,
6842 6 : // delta layers
6843 6 : vec![
6844 6 : DeltaLayerTestDesc::new_with_inferred_key_range(
6845 6 : Lsn(0x10)..Lsn(0x20),
6846 6 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
6847 6 : ),
6848 6 : DeltaLayerTestDesc::new_with_inferred_key_range(
6849 6 : Lsn(0x20)..Lsn(0x30),
6850 6 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
6851 6 : ),
6852 6 : DeltaLayerTestDesc::new_with_inferred_key_range(
6853 6 : Lsn(0x20)..Lsn(0x30),
6854 6 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
6855 6 : ),
6856 6 : ],
6857 6 : // image layers
6858 6 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
6859 6 : Lsn(0x30),
6860 6 : )
6861 87 : .await
6862 6 : .unwrap();
6863 6 :
6864 6 : let cancel = CancellationToken::new();
6865 6 :
6866 6 : tline
6867 6 : .compact(
6868 6 : &cancel,
6869 6 : {
6870 6 : let mut flags = EnumSet::new();
6871 6 : flags.insert(CompactFlags::ForceImageLayerCreation);
6872 6 : flags.insert(CompactFlags::ForceRepartition);
6873 6 : flags
6874 6 : },
6875 6 : &ctx,
6876 6 : )
6877 129 : .await
6878 6 : .unwrap();
6879 6 :
6880 6 : // Image layers are created at last_record_lsn
6881 6 : let images = tline
6882 6 : .inspect_image_layers(Lsn(0x30), &ctx)
6883 12 : .await
6884 6 : .unwrap()
6885 6 : .into_iter()
6886 42 : .filter(|(k, _)| k.is_metadata_key())
6887 6 : .collect::<Vec<_>>();
6888 6 : assert_eq!(images.len(), 0); // the image layer should not contain tombstones, or it is not created
6889 6 : }
6890 :
6891 : #[tokio::test]
6892 6 : async fn test_simple_bottom_most_compaction_images() -> anyhow::Result<()> {
6893 6 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_images").await?;
6894 24 : let (tenant, ctx) = harness.load().await;
6895 6 :
6896 294 : fn get_key(id: u32) -> Key {
6897 294 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
6898 294 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
6899 294 : key.field6 = id;
6900 294 : key
6901 294 : }
6902 6 :
6903 6 : // We create
6904 6 : // - one bottom-most image layer,
6905 6 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
6906 6 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
6907 6 : // - a delta layer D3 above the horizon.
6908 6 : //
6909 6 : // | D3 |
6910 6 : // | D1 |
6911 6 : // -| |-- gc horizon -----------------
6912 6 : // | | | D2 |
6913 6 : // --------- img layer ------------------
6914 6 : //
6915 6 : // What we should expact from this compaction is:
6916 6 : // | D3 |
6917 6 : // | Part of D1 |
6918 6 : // --------- img layer with D1+D2 at GC horizon------------------
6919 6 :
6920 6 : // img layer at 0x10
6921 6 : let img_layer = (0..10)
6922 60 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
6923 6 : .collect_vec();
6924 6 :
6925 6 : let delta1 = vec![
6926 6 : (
6927 6 : get_key(1),
6928 6 : Lsn(0x20),
6929 6 : Value::Image(Bytes::from("value 1@0x20")),
6930 6 : ),
6931 6 : (
6932 6 : get_key(2),
6933 6 : Lsn(0x30),
6934 6 : Value::Image(Bytes::from("value 2@0x30")),
6935 6 : ),
6936 6 : (
6937 6 : get_key(3),
6938 6 : Lsn(0x40),
6939 6 : Value::Image(Bytes::from("value 3@0x40")),
6940 6 : ),
6941 6 : ];
6942 6 : let delta2 = vec![
6943 6 : (
6944 6 : get_key(5),
6945 6 : Lsn(0x20),
6946 6 : Value::Image(Bytes::from("value 5@0x20")),
6947 6 : ),
6948 6 : (
6949 6 : get_key(6),
6950 6 : Lsn(0x20),
6951 6 : Value::Image(Bytes::from("value 6@0x20")),
6952 6 : ),
6953 6 : ];
6954 6 : let delta3 = vec![
6955 6 : (
6956 6 : get_key(8),
6957 6 : Lsn(0x48),
6958 6 : Value::Image(Bytes::from("value 8@0x48")),
6959 6 : ),
6960 6 : (
6961 6 : get_key(9),
6962 6 : Lsn(0x48),
6963 6 : Value::Image(Bytes::from("value 9@0x48")),
6964 6 : ),
6965 6 : ];
6966 6 :
6967 6 : let tline = tenant
6968 6 : .create_test_timeline_with_layers(
6969 6 : TIMELINE_ID,
6970 6 : Lsn(0x10),
6971 6 : DEFAULT_PG_VERSION,
6972 6 : &ctx,
6973 6 : vec![
6974 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
6975 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
6976 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
6977 6 : ], // delta layers
6978 6 : vec![(Lsn(0x10), img_layer)], // image layers
6979 6 : Lsn(0x50),
6980 6 : )
6981 141 : .await?;
6982 6 : {
6983 6 : // Update GC info
6984 6 : let mut guard = tline.gc_info.write().unwrap();
6985 6 : guard.cutoffs.time = Lsn(0x30);
6986 6 : guard.cutoffs.space = Lsn(0x30);
6987 6 : }
6988 6 :
6989 6 : let expected_result = [
6990 6 : Bytes::from_static(b"value 0@0x10"),
6991 6 : Bytes::from_static(b"value 1@0x20"),
6992 6 : Bytes::from_static(b"value 2@0x30"),
6993 6 : Bytes::from_static(b"value 3@0x40"),
6994 6 : Bytes::from_static(b"value 4@0x10"),
6995 6 : Bytes::from_static(b"value 5@0x20"),
6996 6 : Bytes::from_static(b"value 6@0x20"),
6997 6 : Bytes::from_static(b"value 7@0x10"),
6998 6 : Bytes::from_static(b"value 8@0x48"),
6999 6 : Bytes::from_static(b"value 9@0x48"),
7000 6 : ];
7001 6 :
7002 60 : for (idx, expected) in expected_result.iter().enumerate() {
7003 60 : assert_eq!(
7004 60 : tline
7005 60 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7006 69 : .await
7007 60 : .unwrap(),
7008 6 : expected
7009 6 : );
7010 6 : }
7011 6 :
7012 6 : let cancel = CancellationToken::new();
7013 6 : tline
7014 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7015 159 : .await
7016 6 : .unwrap();
7017 6 :
7018 60 : for (idx, expected) in expected_result.iter().enumerate() {
7019 60 : assert_eq!(
7020 60 : tline
7021 60 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7022 48 : .await
7023 60 : .unwrap(),
7024 6 : expected
7025 6 : );
7026 6 : }
7027 6 :
7028 6 : // Check if the image layer at the GC horizon contains exactly what we want
7029 6 : let image_at_gc_horizon = tline
7030 6 : .inspect_image_layers(Lsn(0x30), &ctx)
7031 6 : .await
7032 6 : .unwrap()
7033 6 : .into_iter()
7034 102 : .filter(|(k, _)| k.is_metadata_key())
7035 6 : .collect::<Vec<_>>();
7036 6 :
7037 6 : assert_eq!(image_at_gc_horizon.len(), 10);
7038 6 : let expected_result = [
7039 6 : Bytes::from_static(b"value 0@0x10"),
7040 6 : Bytes::from_static(b"value 1@0x20"),
7041 6 : Bytes::from_static(b"value 2@0x30"),
7042 6 : Bytes::from_static(b"value 3@0x10"),
7043 6 : Bytes::from_static(b"value 4@0x10"),
7044 6 : Bytes::from_static(b"value 5@0x20"),
7045 6 : Bytes::from_static(b"value 6@0x20"),
7046 6 : Bytes::from_static(b"value 7@0x10"),
7047 6 : Bytes::from_static(b"value 8@0x10"),
7048 6 : Bytes::from_static(b"value 9@0x10"),
7049 6 : ];
7050 66 : for idx in 0..10 {
7051 60 : assert_eq!(
7052 60 : image_at_gc_horizon[idx],
7053 60 : (get_key(idx as u32), expected_result[idx].clone())
7054 60 : );
7055 6 : }
7056 6 :
7057 6 : // Check if old layers are removed / new layers have the expected LSN
7058 6 : let mut all_layers = tline.inspect_historic_layers().await.unwrap();
7059 12 : all_layers.sort_by(|k1, k2| {
7060 12 : (
7061 12 : k1.is_delta,
7062 12 : k1.key_range.start,
7063 12 : k1.key_range.end,
7064 12 : k1.lsn_range.start,
7065 12 : k1.lsn_range.end,
7066 12 : )
7067 12 : .cmp(&(
7068 12 : k2.is_delta,
7069 12 : k2.key_range.start,
7070 12 : k2.key_range.end,
7071 12 : k2.lsn_range.start,
7072 12 : k2.lsn_range.end,
7073 12 : ))
7074 12 : });
7075 6 : assert_eq!(
7076 6 : all_layers,
7077 6 : vec![
7078 6 : // Image layer at GC horizon
7079 6 : PersistentLayerKey {
7080 6 : key_range: Key::MIN..Key::NON_L0_MAX,
7081 6 : lsn_range: Lsn(0x30)..Lsn(0x31),
7082 6 : is_delta: false
7083 6 : },
7084 6 : // The delta layer covers the full range (with the layer key hack to avoid being recognized as L0)
7085 6 : PersistentLayerKey {
7086 6 : key_range: Key::MIN..Key::NON_L0_MAX,
7087 6 : lsn_range: Lsn(0x30)..Lsn(0x48),
7088 6 : is_delta: true
7089 6 : },
7090 6 : // The delta3 layer that should not be picked for the compaction
7091 6 : PersistentLayerKey {
7092 6 : key_range: get_key(8)..get_key(10),
7093 6 : lsn_range: Lsn(0x48)..Lsn(0x50),
7094 6 : is_delta: true
7095 6 : }
7096 6 : ]
7097 6 : );
7098 6 :
7099 6 : // increase GC horizon and compact again
7100 6 : {
7101 6 : // Update GC info
7102 6 : let mut guard = tline.gc_info.write().unwrap();
7103 6 : guard.cutoffs.time = Lsn(0x40);
7104 6 : guard.cutoffs.space = Lsn(0x40);
7105 6 : }
7106 6 : tline
7107 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7108 129 : .await
7109 6 : .unwrap();
7110 6 :
7111 6 : Ok(())
7112 6 : }
7113 :
7114 : #[tokio::test]
7115 6 : async fn test_neon_test_record() -> anyhow::Result<()> {
7116 6 : let harness = TenantHarness::create("test_neon_test_record").await?;
7117 24 : let (tenant, ctx) = harness.load().await;
7118 6 :
7119 72 : fn get_key(id: u32) -> Key {
7120 72 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7121 72 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7122 72 : key.field6 = id;
7123 72 : key
7124 72 : }
7125 6 :
7126 6 : let delta1 = vec![
7127 6 : (
7128 6 : get_key(1),
7129 6 : Lsn(0x20),
7130 6 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
7131 6 : ),
7132 6 : (
7133 6 : get_key(1),
7134 6 : Lsn(0x30),
7135 6 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
7136 6 : ),
7137 6 : (get_key(2), Lsn(0x10), Value::Image("0x10".into())),
7138 6 : (
7139 6 : get_key(2),
7140 6 : Lsn(0x20),
7141 6 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
7142 6 : ),
7143 6 : (
7144 6 : get_key(2),
7145 6 : Lsn(0x30),
7146 6 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
7147 6 : ),
7148 6 : (get_key(3), Lsn(0x10), Value::Image("0x10".into())),
7149 6 : (
7150 6 : get_key(3),
7151 6 : Lsn(0x20),
7152 6 : Value::WalRecord(NeonWalRecord::wal_clear()),
7153 6 : ),
7154 6 : (get_key(4), Lsn(0x10), Value::Image("0x10".into())),
7155 6 : (
7156 6 : get_key(4),
7157 6 : Lsn(0x20),
7158 6 : Value::WalRecord(NeonWalRecord::wal_init()),
7159 6 : ),
7160 6 : ];
7161 6 : let image1 = vec![(get_key(1), "0x10".into())];
7162 6 :
7163 6 : let tline = tenant
7164 6 : .create_test_timeline_with_layers(
7165 6 : TIMELINE_ID,
7166 6 : Lsn(0x10),
7167 6 : DEFAULT_PG_VERSION,
7168 6 : &ctx,
7169 6 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
7170 6 : Lsn(0x10)..Lsn(0x40),
7171 6 : delta1,
7172 6 : )], // delta layers
7173 6 : vec![(Lsn(0x10), image1)], // image layers
7174 6 : Lsn(0x50),
7175 6 : )
7176 51 : .await?;
7177 6 :
7178 6 : assert_eq!(
7179 24 : tline.get(get_key(1), Lsn(0x50), &ctx).await?,
7180 6 : Bytes::from_static(b"0x10,0x20,0x30")
7181 6 : );
7182 6 : assert_eq!(
7183 6 : tline.get(get_key(2), Lsn(0x50), &ctx).await?,
7184 6 : Bytes::from_static(b"0x10,0x20,0x30")
7185 6 : );
7186 6 :
7187 6 : // Need to remove the limit of "Neon WAL redo requires base image".
7188 6 :
7189 6 : // assert_eq!(tline.get(get_key(3), Lsn(0x50), &ctx).await?, Bytes::new());
7190 6 : // assert_eq!(tline.get(get_key(4), Lsn(0x50), &ctx).await?, Bytes::new());
7191 6 :
7192 6 : Ok(())
7193 6 : }
7194 :
7195 : #[tokio::test]
7196 6 : async fn test_lsn_lease() -> anyhow::Result<()> {
7197 24 : let (tenant, ctx) = TenantHarness::create("test_lsn_lease").await?.load().await;
7198 6 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7199 6 :
7200 6 : let end_lsn = Lsn(0x100);
7201 6 : let image_layers = (0x20..=0x90)
7202 6 : .step_by(0x10)
7203 48 : .map(|n| {
7204 48 : (
7205 48 : Lsn(n),
7206 48 : vec![(key, test_img(&format!("data key at {:x}", n)))],
7207 48 : )
7208 48 : })
7209 6 : .collect();
7210 6 :
7211 6 : let timeline = tenant
7212 6 : .create_test_timeline_with_layers(
7213 6 : TIMELINE_ID,
7214 6 : Lsn(0x10),
7215 6 : DEFAULT_PG_VERSION,
7216 6 : &ctx,
7217 6 : Vec::new(),
7218 6 : image_layers,
7219 6 : end_lsn,
7220 6 : )
7221 180 : .await?;
7222 6 :
7223 6 : let leased_lsns = [0x30, 0x50, 0x70];
7224 6 : let mut leases = Vec::new();
7225 18 : let _: anyhow::Result<_> = leased_lsns.iter().try_for_each(|n| {
7226 18 : leases.push(timeline.make_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)?);
7227 18 : Ok(())
7228 18 : });
7229 6 :
7230 6 : // Renewing with shorter lease should not change the lease.
7231 6 : let updated_lease_0 =
7232 6 : timeline.make_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)?;
7233 6 : assert_eq!(updated_lease_0.valid_until, leases[0].valid_until);
7234 6 :
7235 6 : // Renewing with a long lease should renew lease with later expiration time.
7236 6 : let updated_lease_1 = timeline.make_lsn_lease(
7237 6 : Lsn(leased_lsns[1]),
7238 6 : timeline.get_lsn_lease_length() * 2,
7239 6 : &ctx,
7240 6 : )?;
7241 6 :
7242 6 : assert!(updated_lease_1.valid_until > leases[1].valid_until);
7243 6 :
7244 6 : // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
7245 6 : info!(
7246 6 : "latest_gc_cutoff_lsn: {}",
7247 0 : *timeline.get_latest_gc_cutoff_lsn()
7248 6 : );
7249 6 : timeline.force_set_disk_consistent_lsn(end_lsn);
7250 6 :
7251 6 : let res = tenant
7252 6 : .gc_iteration(
7253 6 : Some(TIMELINE_ID),
7254 6 : 0,
7255 6 : Duration::ZERO,
7256 6 : &CancellationToken::new(),
7257 6 : &ctx,
7258 6 : )
7259 6 : .await?;
7260 6 :
7261 6 : // Keeping everything <= Lsn(0x80) b/c leases:
7262 6 : // 0/10: initdb layer
7263 6 : // (0/20..=0/70).step_by(0x10): image layers added when creating the timeline.
7264 6 : assert_eq!(res.layers_needed_by_leases, 7);
7265 6 : // Keeping 0/90 b/c it is the latest layer.
7266 6 : assert_eq!(res.layers_not_updated, 1);
7267 6 : // Removed 0/80.
7268 6 : assert_eq!(res.layers_removed, 1);
7269 6 :
7270 6 : // Make lease on a already GC-ed LSN.
7271 6 : // 0/80 does not have a valid lease + is below latest_gc_cutoff
7272 6 : assert!(Lsn(0x80) < *timeline.get_latest_gc_cutoff_lsn());
7273 6 : let res = timeline.make_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx);
7274 6 : assert!(res.is_err());
7275 6 :
7276 6 : // Should still be able to renew a currently valid lease
7277 6 : // Assumption: original lease to is still valid for 0/50.
7278 6 : let _ =
7279 6 : timeline.make_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)?;
7280 6 :
7281 6 : Ok(())
7282 6 : }
7283 :
7284 : #[tokio::test]
7285 6 : async fn test_simple_bottom_most_compaction_deltas() -> anyhow::Result<()> {
7286 6 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_deltas").await?;
7287 23 : let (tenant, ctx) = harness.load().await;
7288 6 :
7289 354 : fn get_key(id: u32) -> Key {
7290 354 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7291 354 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7292 354 : key.field6 = id;
7293 354 : key
7294 354 : }
7295 6 :
7296 6 : // We create
7297 6 : // - one bottom-most image layer,
7298 6 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
7299 6 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
7300 6 : // - a delta layer D3 above the horizon.
7301 6 : //
7302 6 : // | D3 |
7303 6 : // | D1 |
7304 6 : // -| |-- gc horizon -----------------
7305 6 : // | | | D2 |
7306 6 : // --------- img layer ------------------
7307 6 : //
7308 6 : // What we should expact from this compaction is:
7309 6 : // | D3 |
7310 6 : // | Part of D1 |
7311 6 : // --------- img layer with D1+D2 at GC horizon------------------
7312 6 :
7313 6 : // img layer at 0x10
7314 6 : let img_layer = (0..10)
7315 60 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
7316 6 : .collect_vec();
7317 6 :
7318 6 : let delta1 = vec![
7319 6 : (
7320 6 : get_key(1),
7321 6 : Lsn(0x20),
7322 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7323 6 : ),
7324 6 : (
7325 6 : get_key(2),
7326 6 : Lsn(0x30),
7327 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7328 6 : ),
7329 6 : (
7330 6 : get_key(3),
7331 6 : Lsn(0x28),
7332 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
7333 6 : ),
7334 6 : (
7335 6 : get_key(3),
7336 6 : Lsn(0x30),
7337 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7338 6 : ),
7339 6 : (
7340 6 : get_key(3),
7341 6 : Lsn(0x40),
7342 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
7343 6 : ),
7344 6 : ];
7345 6 : let delta2 = vec![
7346 6 : (
7347 6 : get_key(5),
7348 6 : Lsn(0x20),
7349 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7350 6 : ),
7351 6 : (
7352 6 : get_key(6),
7353 6 : Lsn(0x20),
7354 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7355 6 : ),
7356 6 : ];
7357 6 : let delta3 = vec![
7358 6 : (
7359 6 : get_key(8),
7360 6 : Lsn(0x48),
7361 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
7362 6 : ),
7363 6 : (
7364 6 : get_key(9),
7365 6 : Lsn(0x48),
7366 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
7367 6 : ),
7368 6 : ];
7369 6 :
7370 6 : let tline = tenant
7371 6 : .create_test_timeline_with_layers(
7372 6 : TIMELINE_ID,
7373 6 : Lsn(0x10),
7374 6 : DEFAULT_PG_VERSION,
7375 6 : &ctx,
7376 6 : vec![
7377 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
7378 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
7379 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
7380 6 : ], // delta layers
7381 6 : vec![(Lsn(0x10), img_layer)], // image layers
7382 6 : Lsn(0x50),
7383 6 : )
7384 141 : .await?;
7385 6 : {
7386 6 : // Update GC info
7387 6 : let mut guard = tline.gc_info.write().unwrap();
7388 6 : *guard = GcInfo {
7389 6 : retain_lsns: vec![],
7390 6 : cutoffs: GcCutoffs {
7391 6 : time: Lsn(0x30),
7392 6 : space: Lsn(0x30),
7393 6 : },
7394 6 : leases: Default::default(),
7395 6 : within_ancestor_pitr: false,
7396 6 : };
7397 6 : }
7398 6 :
7399 6 : let expected_result = [
7400 6 : Bytes::from_static(b"value 0@0x10"),
7401 6 : Bytes::from_static(b"value 1@0x10@0x20"),
7402 6 : Bytes::from_static(b"value 2@0x10@0x30"),
7403 6 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
7404 6 : Bytes::from_static(b"value 4@0x10"),
7405 6 : Bytes::from_static(b"value 5@0x10@0x20"),
7406 6 : Bytes::from_static(b"value 6@0x10@0x20"),
7407 6 : Bytes::from_static(b"value 7@0x10"),
7408 6 : Bytes::from_static(b"value 8@0x10@0x48"),
7409 6 : Bytes::from_static(b"value 9@0x10@0x48"),
7410 6 : ];
7411 6 :
7412 6 : let expected_result_at_gc_horizon = [
7413 6 : Bytes::from_static(b"value 0@0x10"),
7414 6 : Bytes::from_static(b"value 1@0x10@0x20"),
7415 6 : Bytes::from_static(b"value 2@0x10@0x30"),
7416 6 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
7417 6 : Bytes::from_static(b"value 4@0x10"),
7418 6 : Bytes::from_static(b"value 5@0x10@0x20"),
7419 6 : Bytes::from_static(b"value 6@0x10@0x20"),
7420 6 : Bytes::from_static(b"value 7@0x10"),
7421 6 : Bytes::from_static(b"value 8@0x10"),
7422 6 : Bytes::from_static(b"value 9@0x10"),
7423 6 : ];
7424 6 :
7425 66 : for idx in 0..10 {
7426 60 : assert_eq!(
7427 60 : tline
7428 60 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7429 93 : .await
7430 60 : .unwrap(),
7431 60 : &expected_result[idx]
7432 6 : );
7433 60 : assert_eq!(
7434 60 : tline
7435 60 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
7436 45 : .await
7437 60 : .unwrap(),
7438 60 : &expected_result_at_gc_horizon[idx]
7439 6 : );
7440 6 : }
7441 6 :
7442 6 : let cancel = CancellationToken::new();
7443 6 : tline
7444 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7445 156 : .await
7446 6 : .unwrap();
7447 6 :
7448 66 : for idx in 0..10 {
7449 60 : assert_eq!(
7450 60 : tline
7451 60 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7452 63 : .await
7453 60 : .unwrap(),
7454 60 : &expected_result[idx]
7455 6 : );
7456 60 : assert_eq!(
7457 60 : tline
7458 60 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
7459 30 : .await
7460 60 : .unwrap(),
7461 60 : &expected_result_at_gc_horizon[idx]
7462 6 : );
7463 6 : }
7464 6 :
7465 6 : // increase GC horizon and compact again
7466 6 : {
7467 6 : // Update GC info
7468 6 : let mut guard = tline.gc_info.write().unwrap();
7469 6 : guard.cutoffs.time = Lsn(0x40);
7470 6 : guard.cutoffs.space = Lsn(0x40);
7471 6 : }
7472 6 : tline
7473 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7474 126 : .await
7475 6 : .unwrap();
7476 6 :
7477 6 : Ok(())
7478 6 : }
7479 :
7480 : #[tokio::test]
7481 6 : async fn test_generate_key_retention() -> anyhow::Result<()> {
7482 6 : let harness = TenantHarness::create("test_generate_key_retention").await?;
7483 24 : let (tenant, ctx) = harness.load().await;
7484 6 : let tline = tenant
7485 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7486 12 : .await?;
7487 6 : tline.force_advance_lsn(Lsn(0x70));
7488 6 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7489 6 : let history = vec![
7490 6 : (
7491 6 : key,
7492 6 : Lsn(0x10),
7493 6 : Value::Image(Bytes::copy_from_slice(b"0x10")),
7494 6 : ),
7495 6 : (
7496 6 : key,
7497 6 : Lsn(0x20),
7498 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7499 6 : ),
7500 6 : (
7501 6 : key,
7502 6 : Lsn(0x30),
7503 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
7504 6 : ),
7505 6 : (
7506 6 : key,
7507 6 : Lsn(0x40),
7508 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
7509 6 : ),
7510 6 : (
7511 6 : key,
7512 6 : Lsn(0x50),
7513 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
7514 6 : ),
7515 6 : (
7516 6 : key,
7517 6 : Lsn(0x60),
7518 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7519 6 : ),
7520 6 : (
7521 6 : key,
7522 6 : Lsn(0x70),
7523 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7524 6 : ),
7525 6 : (
7526 6 : key,
7527 6 : Lsn(0x80),
7528 6 : Value::Image(Bytes::copy_from_slice(
7529 6 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
7530 6 : )),
7531 6 : ),
7532 6 : (
7533 6 : key,
7534 6 : Lsn(0x90),
7535 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
7536 6 : ),
7537 6 : ];
7538 6 : let res = tline
7539 6 : .generate_key_retention(
7540 6 : key,
7541 6 : &history,
7542 6 : Lsn(0x60),
7543 6 : &[Lsn(0x20), Lsn(0x40), Lsn(0x50)],
7544 6 : 3,
7545 6 : None,
7546 6 : )
7547 6 : .await
7548 6 : .unwrap();
7549 6 : let expected_res = KeyHistoryRetention {
7550 6 : below_horizon: vec![
7551 6 : (
7552 6 : Lsn(0x20),
7553 6 : KeyLogAtLsn(vec![(
7554 6 : Lsn(0x20),
7555 6 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20")),
7556 6 : )]),
7557 6 : ),
7558 6 : (
7559 6 : Lsn(0x40),
7560 6 : KeyLogAtLsn(vec![
7561 6 : (
7562 6 : Lsn(0x30),
7563 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
7564 6 : ),
7565 6 : (
7566 6 : Lsn(0x40),
7567 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
7568 6 : ),
7569 6 : ]),
7570 6 : ),
7571 6 : (
7572 6 : Lsn(0x50),
7573 6 : KeyLogAtLsn(vec![(
7574 6 : Lsn(0x50),
7575 6 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40;0x50")),
7576 6 : )]),
7577 6 : ),
7578 6 : (
7579 6 : Lsn(0x60),
7580 6 : KeyLogAtLsn(vec![(
7581 6 : Lsn(0x60),
7582 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7583 6 : )]),
7584 6 : ),
7585 6 : ],
7586 6 : above_horizon: KeyLogAtLsn(vec![
7587 6 : (
7588 6 : Lsn(0x70),
7589 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7590 6 : ),
7591 6 : (
7592 6 : Lsn(0x80),
7593 6 : Value::Image(Bytes::copy_from_slice(
7594 6 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
7595 6 : )),
7596 6 : ),
7597 6 : (
7598 6 : Lsn(0x90),
7599 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
7600 6 : ),
7601 6 : ]),
7602 6 : };
7603 6 : assert_eq!(res, expected_res);
7604 6 :
7605 6 : // We expect GC-compaction to run with the original GC. This would create a situation that
7606 6 : // the original GC algorithm removes some delta layers b/c there are full image coverage,
7607 6 : // therefore causing some keys to have an incomplete history below the lowest retain LSN.
7608 6 : // For example, we have
7609 6 : // ```plain
7610 6 : // init delta @ 0x10, image @ 0x20, delta @ 0x30 (gc_horizon), image @ 0x40.
7611 6 : // ```
7612 6 : // Now the GC horizon moves up, and we have
7613 6 : // ```plain
7614 6 : // init delta @ 0x10, image @ 0x20, delta @ 0x30, image @ 0x40 (gc_horizon)
7615 6 : // ```
7616 6 : // The original GC algorithm kicks in, and removes delta @ 0x10, image @ 0x20.
7617 6 : // We will end up with
7618 6 : // ```plain
7619 6 : // delta @ 0x30, image @ 0x40 (gc_horizon)
7620 6 : // ```
7621 6 : // Now we run the GC-compaction, and this key does not have a full history.
7622 6 : // We should be able to handle this partial history and drop everything before the
7623 6 : // gc_horizon image.
7624 6 :
7625 6 : let history = vec![
7626 6 : (
7627 6 : key,
7628 6 : Lsn(0x20),
7629 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7630 6 : ),
7631 6 : (
7632 6 : key,
7633 6 : Lsn(0x30),
7634 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
7635 6 : ),
7636 6 : (
7637 6 : key,
7638 6 : Lsn(0x40),
7639 6 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
7640 6 : ),
7641 6 : (
7642 6 : key,
7643 6 : Lsn(0x50),
7644 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
7645 6 : ),
7646 6 : (
7647 6 : key,
7648 6 : Lsn(0x60),
7649 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7650 6 : ),
7651 6 : (
7652 6 : key,
7653 6 : Lsn(0x70),
7654 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7655 6 : ),
7656 6 : (
7657 6 : key,
7658 6 : Lsn(0x80),
7659 6 : Value::Image(Bytes::copy_from_slice(
7660 6 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
7661 6 : )),
7662 6 : ),
7663 6 : (
7664 6 : key,
7665 6 : Lsn(0x90),
7666 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
7667 6 : ),
7668 6 : ];
7669 6 : let res = tline
7670 6 : .generate_key_retention(key, &history, Lsn(0x60), &[Lsn(0x40), Lsn(0x50)], 3, None)
7671 6 : .await
7672 6 : .unwrap();
7673 6 : let expected_res = KeyHistoryRetention {
7674 6 : below_horizon: vec![
7675 6 : (
7676 6 : Lsn(0x40),
7677 6 : KeyLogAtLsn(vec![(
7678 6 : Lsn(0x40),
7679 6 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
7680 6 : )]),
7681 6 : ),
7682 6 : (
7683 6 : Lsn(0x50),
7684 6 : KeyLogAtLsn(vec![(
7685 6 : Lsn(0x50),
7686 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
7687 6 : )]),
7688 6 : ),
7689 6 : (
7690 6 : Lsn(0x60),
7691 6 : KeyLogAtLsn(vec![(
7692 6 : Lsn(0x60),
7693 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7694 6 : )]),
7695 6 : ),
7696 6 : ],
7697 6 : above_horizon: KeyLogAtLsn(vec![
7698 6 : (
7699 6 : Lsn(0x70),
7700 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7701 6 : ),
7702 6 : (
7703 6 : Lsn(0x80),
7704 6 : Value::Image(Bytes::copy_from_slice(
7705 6 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
7706 6 : )),
7707 6 : ),
7708 6 : (
7709 6 : Lsn(0x90),
7710 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
7711 6 : ),
7712 6 : ]),
7713 6 : };
7714 6 : assert_eq!(res, expected_res);
7715 6 :
7716 6 : // In case of branch compaction, the branch itself does not have the full history, and we need to provide
7717 6 : // the ancestor image in the test case.
7718 6 :
7719 6 : let history = vec![
7720 6 : (
7721 6 : key,
7722 6 : Lsn(0x20),
7723 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7724 6 : ),
7725 6 : (
7726 6 : key,
7727 6 : Lsn(0x30),
7728 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
7729 6 : ),
7730 6 : (
7731 6 : key,
7732 6 : Lsn(0x40),
7733 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
7734 6 : ),
7735 6 : (
7736 6 : key,
7737 6 : Lsn(0x70),
7738 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7739 6 : ),
7740 6 : ];
7741 6 : let res = tline
7742 6 : .generate_key_retention(
7743 6 : key,
7744 6 : &history,
7745 6 : Lsn(0x60),
7746 6 : &[],
7747 6 : 3,
7748 6 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
7749 6 : )
7750 6 : .await
7751 6 : .unwrap();
7752 6 : let expected_res = KeyHistoryRetention {
7753 6 : below_horizon: vec![(
7754 6 : Lsn(0x60),
7755 6 : KeyLogAtLsn(vec![(
7756 6 : Lsn(0x60),
7757 6 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")), // use the ancestor image to reconstruct the page
7758 6 : )]),
7759 6 : )],
7760 6 : above_horizon: KeyLogAtLsn(vec![(
7761 6 : Lsn(0x70),
7762 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7763 6 : )]),
7764 6 : };
7765 6 : assert_eq!(res, expected_res);
7766 6 :
7767 6 : let history = vec![
7768 6 : (
7769 6 : key,
7770 6 : Lsn(0x20),
7771 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7772 6 : ),
7773 6 : (
7774 6 : key,
7775 6 : Lsn(0x40),
7776 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
7777 6 : ),
7778 6 : (
7779 6 : key,
7780 6 : Lsn(0x60),
7781 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7782 6 : ),
7783 6 : (
7784 6 : key,
7785 6 : Lsn(0x70),
7786 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7787 6 : ),
7788 6 : ];
7789 6 : let res = tline
7790 6 : .generate_key_retention(
7791 6 : key,
7792 6 : &history,
7793 6 : Lsn(0x60),
7794 6 : &[Lsn(0x30)],
7795 6 : 3,
7796 6 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
7797 6 : )
7798 6 : .await
7799 6 : .unwrap();
7800 6 : let expected_res = KeyHistoryRetention {
7801 6 : below_horizon: vec![
7802 6 : (
7803 6 : Lsn(0x30),
7804 6 : KeyLogAtLsn(vec![(
7805 6 : Lsn(0x20),
7806 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7807 6 : )]),
7808 6 : ),
7809 6 : (
7810 6 : Lsn(0x60),
7811 6 : KeyLogAtLsn(vec![(
7812 6 : Lsn(0x60),
7813 6 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x40;0x60")),
7814 6 : )]),
7815 6 : ),
7816 6 : ],
7817 6 : above_horizon: KeyLogAtLsn(vec![(
7818 6 : Lsn(0x70),
7819 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7820 6 : )]),
7821 6 : };
7822 6 : assert_eq!(res, expected_res);
7823 6 :
7824 6 : Ok(())
7825 6 : }
7826 :
7827 : #[tokio::test]
7828 6 : async fn test_simple_bottom_most_compaction_with_retain_lsns() -> anyhow::Result<()> {
7829 6 : let harness =
7830 6 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns").await?;
7831 24 : let (tenant, ctx) = harness.load().await;
7832 6 :
7833 1554 : fn get_key(id: u32) -> Key {
7834 1554 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7835 1554 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7836 1554 : key.field6 = id;
7837 1554 : key
7838 1554 : }
7839 6 :
7840 6 : let img_layer = (0..10)
7841 60 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
7842 6 : .collect_vec();
7843 6 :
7844 6 : let delta1 = vec![
7845 6 : (
7846 6 : get_key(1),
7847 6 : Lsn(0x20),
7848 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7849 6 : ),
7850 6 : (
7851 6 : get_key(2),
7852 6 : Lsn(0x30),
7853 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7854 6 : ),
7855 6 : (
7856 6 : get_key(3),
7857 6 : Lsn(0x28),
7858 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
7859 6 : ),
7860 6 : (
7861 6 : get_key(3),
7862 6 : Lsn(0x30),
7863 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7864 6 : ),
7865 6 : (
7866 6 : get_key(3),
7867 6 : Lsn(0x40),
7868 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
7869 6 : ),
7870 6 : ];
7871 6 : let delta2 = vec![
7872 6 : (
7873 6 : get_key(5),
7874 6 : Lsn(0x20),
7875 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7876 6 : ),
7877 6 : (
7878 6 : get_key(6),
7879 6 : Lsn(0x20),
7880 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7881 6 : ),
7882 6 : ];
7883 6 : let delta3 = vec![
7884 6 : (
7885 6 : get_key(8),
7886 6 : Lsn(0x48),
7887 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
7888 6 : ),
7889 6 : (
7890 6 : get_key(9),
7891 6 : Lsn(0x48),
7892 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
7893 6 : ),
7894 6 : ];
7895 6 :
7896 6 : let tline = tenant
7897 6 : .create_test_timeline_with_layers(
7898 6 : TIMELINE_ID,
7899 6 : Lsn(0x10),
7900 6 : DEFAULT_PG_VERSION,
7901 6 : &ctx,
7902 6 : vec![
7903 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
7904 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
7905 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
7906 6 : ], // delta layers
7907 6 : vec![(Lsn(0x10), img_layer)], // image layers
7908 6 : Lsn(0x50),
7909 6 : )
7910 141 : .await?;
7911 6 : {
7912 6 : // Update GC info
7913 6 : let mut guard = tline.gc_info.write().unwrap();
7914 6 : *guard = GcInfo {
7915 6 : retain_lsns: vec![
7916 6 : (Lsn(0x10), tline.timeline_id),
7917 6 : (Lsn(0x20), tline.timeline_id),
7918 6 : ],
7919 6 : cutoffs: GcCutoffs {
7920 6 : time: Lsn(0x30),
7921 6 : space: Lsn(0x30),
7922 6 : },
7923 6 : leases: Default::default(),
7924 6 : within_ancestor_pitr: false,
7925 6 : };
7926 6 : }
7927 6 :
7928 6 : let expected_result = [
7929 6 : Bytes::from_static(b"value 0@0x10"),
7930 6 : Bytes::from_static(b"value 1@0x10@0x20"),
7931 6 : Bytes::from_static(b"value 2@0x10@0x30"),
7932 6 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
7933 6 : Bytes::from_static(b"value 4@0x10"),
7934 6 : Bytes::from_static(b"value 5@0x10@0x20"),
7935 6 : Bytes::from_static(b"value 6@0x10@0x20"),
7936 6 : Bytes::from_static(b"value 7@0x10"),
7937 6 : Bytes::from_static(b"value 8@0x10@0x48"),
7938 6 : Bytes::from_static(b"value 9@0x10@0x48"),
7939 6 : ];
7940 6 :
7941 6 : let expected_result_at_gc_horizon = [
7942 6 : Bytes::from_static(b"value 0@0x10"),
7943 6 : Bytes::from_static(b"value 1@0x10@0x20"),
7944 6 : Bytes::from_static(b"value 2@0x10@0x30"),
7945 6 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
7946 6 : Bytes::from_static(b"value 4@0x10"),
7947 6 : Bytes::from_static(b"value 5@0x10@0x20"),
7948 6 : Bytes::from_static(b"value 6@0x10@0x20"),
7949 6 : Bytes::from_static(b"value 7@0x10"),
7950 6 : Bytes::from_static(b"value 8@0x10"),
7951 6 : Bytes::from_static(b"value 9@0x10"),
7952 6 : ];
7953 6 :
7954 6 : let expected_result_at_lsn_20 = [
7955 6 : Bytes::from_static(b"value 0@0x10"),
7956 6 : Bytes::from_static(b"value 1@0x10@0x20"),
7957 6 : Bytes::from_static(b"value 2@0x10"),
7958 6 : Bytes::from_static(b"value 3@0x10"),
7959 6 : Bytes::from_static(b"value 4@0x10"),
7960 6 : Bytes::from_static(b"value 5@0x10@0x20"),
7961 6 : Bytes::from_static(b"value 6@0x10@0x20"),
7962 6 : Bytes::from_static(b"value 7@0x10"),
7963 6 : Bytes::from_static(b"value 8@0x10"),
7964 6 : Bytes::from_static(b"value 9@0x10"),
7965 6 : ];
7966 6 :
7967 6 : let expected_result_at_lsn_10 = [
7968 6 : Bytes::from_static(b"value 0@0x10"),
7969 6 : Bytes::from_static(b"value 1@0x10"),
7970 6 : Bytes::from_static(b"value 2@0x10"),
7971 6 : Bytes::from_static(b"value 3@0x10"),
7972 6 : Bytes::from_static(b"value 4@0x10"),
7973 6 : Bytes::from_static(b"value 5@0x10"),
7974 6 : Bytes::from_static(b"value 6@0x10"),
7975 6 : Bytes::from_static(b"value 7@0x10"),
7976 6 : Bytes::from_static(b"value 8@0x10"),
7977 6 : Bytes::from_static(b"value 9@0x10"),
7978 6 : ];
7979 6 :
7980 36 : let verify_result = || async {
7981 36 : let gc_horizon = {
7982 36 : let gc_info = tline.gc_info.read().unwrap();
7983 36 : gc_info.cutoffs.time
7984 36 : };
7985 396 : for idx in 0..10 {
7986 360 : assert_eq!(
7987 360 : tline
7988 360 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7989 366 : .await
7990 360 : .unwrap(),
7991 360 : &expected_result[idx]
7992 36 : );
7993 360 : assert_eq!(
7994 360 : tline
7995 360 : .get(get_key(idx as u32), gc_horizon, &ctx)
7996 288 : .await
7997 360 : .unwrap(),
7998 360 : &expected_result_at_gc_horizon[idx]
7999 36 : );
8000 360 : assert_eq!(
8001 360 : tline
8002 360 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
8003 243 : .await
8004 360 : .unwrap(),
8005 360 : &expected_result_at_lsn_20[idx]
8006 36 : );
8007 360 : assert_eq!(
8008 360 : tline
8009 360 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
8010 189 : .await
8011 360 : .unwrap(),
8012 360 : &expected_result_at_lsn_10[idx]
8013 36 : );
8014 36 : }
8015 36 : };
8016 6 :
8017 210 : verify_result().await;
8018 6 :
8019 6 : let cancel = CancellationToken::new();
8020 6 : let mut dryrun_flags = EnumSet::new();
8021 6 : dryrun_flags.insert(CompactFlags::DryRun);
8022 6 :
8023 6 : tline
8024 6 : .compact_with_gc(&cancel, dryrun_flags, &ctx)
8025 129 : .await
8026 6 : .unwrap();
8027 6 : // 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
8028 6 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
8029 171 : verify_result().await;
8030 6 :
8031 6 : tline
8032 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8033 150 : .await
8034 6 : .unwrap();
8035 192 : verify_result().await;
8036 6 :
8037 6 : // compact again
8038 6 : tline
8039 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8040 117 : .await
8041 6 : .unwrap();
8042 171 : verify_result().await;
8043 6 :
8044 6 : // increase GC horizon and compact again
8045 6 : {
8046 6 : // Update GC info
8047 6 : let mut guard = tline.gc_info.write().unwrap();
8048 6 : guard.cutoffs.time = Lsn(0x38);
8049 6 : guard.cutoffs.space = Lsn(0x38);
8050 6 : }
8051 6 : tline
8052 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8053 117 : .await
8054 6 : .unwrap();
8055 171 : verify_result().await; // no wals between 0x30 and 0x38, so we should obtain the same result
8056 6 :
8057 6 : // not increasing the GC horizon and compact again
8058 6 : tline
8059 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8060 117 : .await
8061 6 : .unwrap();
8062 171 : verify_result().await;
8063 6 :
8064 6 : Ok(())
8065 6 : }
8066 :
8067 : #[tokio::test]
8068 6 : async fn test_simple_bottom_most_compaction_with_retain_lsns_single_key() -> anyhow::Result<()>
8069 6 : {
8070 6 : let harness =
8071 6 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns_single_key")
8072 6 : .await?;
8073 23 : let (tenant, ctx) = harness.load().await;
8074 6 :
8075 1056 : fn get_key(id: u32) -> Key {
8076 1056 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8077 1056 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8078 1056 : key.field6 = id;
8079 1056 : key
8080 1056 : }
8081 6 :
8082 6 : let img_layer = (0..10)
8083 60 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8084 6 : .collect_vec();
8085 6 :
8086 6 : let delta1 = vec![
8087 6 : (
8088 6 : get_key(1),
8089 6 : Lsn(0x20),
8090 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8091 6 : ),
8092 6 : (
8093 6 : get_key(1),
8094 6 : Lsn(0x28),
8095 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8096 6 : ),
8097 6 : ];
8098 6 : let delta2 = vec![
8099 6 : (
8100 6 : get_key(1),
8101 6 : Lsn(0x30),
8102 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8103 6 : ),
8104 6 : (
8105 6 : get_key(1),
8106 6 : Lsn(0x38),
8107 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
8108 6 : ),
8109 6 : ];
8110 6 : let delta3 = vec![
8111 6 : (
8112 6 : get_key(8),
8113 6 : Lsn(0x48),
8114 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8115 6 : ),
8116 6 : (
8117 6 : get_key(9),
8118 6 : Lsn(0x48),
8119 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8120 6 : ),
8121 6 : ];
8122 6 :
8123 6 : let tline = tenant
8124 6 : .create_test_timeline_with_layers(
8125 6 : TIMELINE_ID,
8126 6 : Lsn(0x10),
8127 6 : DEFAULT_PG_VERSION,
8128 6 : &ctx,
8129 6 : vec![
8130 6 : // delta1 and delta 2 only contain a single key but multiple updates
8131 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x30), delta1),
8132 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
8133 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x50), delta3),
8134 6 : ], // delta layers
8135 6 : vec![(Lsn(0x10), img_layer)], // image layers
8136 6 : Lsn(0x50),
8137 6 : )
8138 141 : .await?;
8139 6 : {
8140 6 : // Update GC info
8141 6 : let mut guard = tline.gc_info.write().unwrap();
8142 6 : *guard = GcInfo {
8143 6 : retain_lsns: vec![
8144 6 : (Lsn(0x10), tline.timeline_id),
8145 6 : (Lsn(0x20), tline.timeline_id),
8146 6 : ],
8147 6 : cutoffs: GcCutoffs {
8148 6 : time: Lsn(0x30),
8149 6 : space: Lsn(0x30),
8150 6 : },
8151 6 : leases: Default::default(),
8152 6 : within_ancestor_pitr: false,
8153 6 : };
8154 6 : }
8155 6 :
8156 6 : let expected_result = [
8157 6 : Bytes::from_static(b"value 0@0x10"),
8158 6 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
8159 6 : Bytes::from_static(b"value 2@0x10"),
8160 6 : Bytes::from_static(b"value 3@0x10"),
8161 6 : Bytes::from_static(b"value 4@0x10"),
8162 6 : Bytes::from_static(b"value 5@0x10"),
8163 6 : Bytes::from_static(b"value 6@0x10"),
8164 6 : Bytes::from_static(b"value 7@0x10"),
8165 6 : Bytes::from_static(b"value 8@0x10@0x48"),
8166 6 : Bytes::from_static(b"value 9@0x10@0x48"),
8167 6 : ];
8168 6 :
8169 6 : let expected_result_at_gc_horizon = [
8170 6 : Bytes::from_static(b"value 0@0x10"),
8171 6 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
8172 6 : Bytes::from_static(b"value 2@0x10"),
8173 6 : Bytes::from_static(b"value 3@0x10"),
8174 6 : Bytes::from_static(b"value 4@0x10"),
8175 6 : Bytes::from_static(b"value 5@0x10"),
8176 6 : Bytes::from_static(b"value 6@0x10"),
8177 6 : Bytes::from_static(b"value 7@0x10"),
8178 6 : Bytes::from_static(b"value 8@0x10"),
8179 6 : Bytes::from_static(b"value 9@0x10"),
8180 6 : ];
8181 6 :
8182 6 : let expected_result_at_lsn_20 = [
8183 6 : Bytes::from_static(b"value 0@0x10"),
8184 6 : Bytes::from_static(b"value 1@0x10@0x20"),
8185 6 : Bytes::from_static(b"value 2@0x10"),
8186 6 : Bytes::from_static(b"value 3@0x10"),
8187 6 : Bytes::from_static(b"value 4@0x10"),
8188 6 : Bytes::from_static(b"value 5@0x10"),
8189 6 : Bytes::from_static(b"value 6@0x10"),
8190 6 : Bytes::from_static(b"value 7@0x10"),
8191 6 : Bytes::from_static(b"value 8@0x10"),
8192 6 : Bytes::from_static(b"value 9@0x10"),
8193 6 : ];
8194 6 :
8195 6 : let expected_result_at_lsn_10 = [
8196 6 : Bytes::from_static(b"value 0@0x10"),
8197 6 : Bytes::from_static(b"value 1@0x10"),
8198 6 : Bytes::from_static(b"value 2@0x10"),
8199 6 : Bytes::from_static(b"value 3@0x10"),
8200 6 : Bytes::from_static(b"value 4@0x10"),
8201 6 : Bytes::from_static(b"value 5@0x10"),
8202 6 : Bytes::from_static(b"value 6@0x10"),
8203 6 : Bytes::from_static(b"value 7@0x10"),
8204 6 : Bytes::from_static(b"value 8@0x10"),
8205 6 : Bytes::from_static(b"value 9@0x10"),
8206 6 : ];
8207 6 :
8208 24 : let verify_result = || async {
8209 24 : let gc_horizon = {
8210 24 : let gc_info = tline.gc_info.read().unwrap();
8211 24 : gc_info.cutoffs.time
8212 24 : };
8213 264 : for idx in 0..10 {
8214 240 : assert_eq!(
8215 240 : tline
8216 240 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8217 228 : .await
8218 240 : .unwrap(),
8219 240 : &expected_result[idx]
8220 24 : );
8221 240 : assert_eq!(
8222 240 : tline
8223 240 : .get(get_key(idx as u32), gc_horizon, &ctx)
8224 141 : .await
8225 240 : .unwrap(),
8226 240 : &expected_result_at_gc_horizon[idx]
8227 24 : );
8228 240 : assert_eq!(
8229 240 : tline
8230 240 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
8231 141 : .await
8232 240 : .unwrap(),
8233 240 : &expected_result_at_lsn_20[idx]
8234 24 : );
8235 240 : assert_eq!(
8236 240 : tline
8237 240 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
8238 120 : .await
8239 240 : .unwrap(),
8240 240 : &expected_result_at_lsn_10[idx]
8241 24 : );
8242 24 : }
8243 24 : };
8244 6 :
8245 183 : verify_result().await;
8246 6 :
8247 6 : let cancel = CancellationToken::new();
8248 6 : let mut dryrun_flags = EnumSet::new();
8249 6 : dryrun_flags.insert(CompactFlags::DryRun);
8250 6 :
8251 6 : tline
8252 6 : .compact_with_gc(&cancel, dryrun_flags, &ctx)
8253 135 : .await
8254 6 : .unwrap();
8255 6 : // 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
8256 6 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
8257 147 : verify_result().await;
8258 6 :
8259 6 : tline
8260 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8261 153 : .await
8262 6 : .unwrap();
8263 159 : verify_result().await;
8264 6 :
8265 6 : // compact again
8266 6 : tline
8267 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8268 117 : .await
8269 6 : .unwrap();
8270 141 : verify_result().await;
8271 6 :
8272 6 : Ok(())
8273 6 : }
8274 :
8275 : #[tokio::test]
8276 6 : async fn test_simple_bottom_most_compaction_on_branch() -> anyhow::Result<()> {
8277 6 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_on_branch").await?;
8278 24 : let (tenant, ctx) = harness.load().await;
8279 6 :
8280 378 : fn get_key(id: u32) -> Key {
8281 378 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
8282 378 : key.field6 = id;
8283 378 : key
8284 378 : }
8285 6 :
8286 6 : let img_layer = (0..10)
8287 60 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8288 6 : .collect_vec();
8289 6 :
8290 6 : let delta1 = vec![
8291 6 : (
8292 6 : get_key(1),
8293 6 : Lsn(0x20),
8294 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8295 6 : ),
8296 6 : (
8297 6 : get_key(2),
8298 6 : Lsn(0x30),
8299 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8300 6 : ),
8301 6 : (
8302 6 : get_key(3),
8303 6 : Lsn(0x28),
8304 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8305 6 : ),
8306 6 : (
8307 6 : get_key(3),
8308 6 : Lsn(0x30),
8309 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8310 6 : ),
8311 6 : (
8312 6 : get_key(3),
8313 6 : Lsn(0x40),
8314 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
8315 6 : ),
8316 6 : ];
8317 6 : let delta2 = vec![
8318 6 : (
8319 6 : get_key(5),
8320 6 : Lsn(0x20),
8321 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8322 6 : ),
8323 6 : (
8324 6 : get_key(6),
8325 6 : Lsn(0x20),
8326 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8327 6 : ),
8328 6 : ];
8329 6 : let delta3 = vec![
8330 6 : (
8331 6 : get_key(8),
8332 6 : Lsn(0x48),
8333 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8334 6 : ),
8335 6 : (
8336 6 : get_key(9),
8337 6 : Lsn(0x48),
8338 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8339 6 : ),
8340 6 : ];
8341 6 :
8342 6 : let parent_tline = tenant
8343 6 : .create_test_timeline_with_layers(
8344 6 : TIMELINE_ID,
8345 6 : Lsn(0x10),
8346 6 : DEFAULT_PG_VERSION,
8347 6 : &ctx,
8348 6 : vec![], // delta layers
8349 6 : vec![(Lsn(0x18), img_layer)], // image layers
8350 6 : Lsn(0x18),
8351 6 : )
8352 87 : .await?;
8353 6 :
8354 6 : parent_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
8355 6 :
8356 6 : let branch_tline = tenant
8357 6 : .branch_timeline_test_with_layers(
8358 6 : &parent_tline,
8359 6 : NEW_TIMELINE_ID,
8360 6 : Some(Lsn(0x18)),
8361 6 : &ctx,
8362 6 : vec![
8363 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
8364 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
8365 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
8366 6 : ], // delta layers
8367 6 : vec![], // image layers
8368 6 : Lsn(0x50),
8369 6 : )
8370 54 : .await?;
8371 6 :
8372 6 : branch_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
8373 6 :
8374 6 : {
8375 6 : // Update GC info
8376 6 : let mut guard = parent_tline.gc_info.write().unwrap();
8377 6 : *guard = GcInfo {
8378 6 : retain_lsns: vec![(Lsn(0x18), branch_tline.timeline_id)],
8379 6 : cutoffs: GcCutoffs {
8380 6 : time: Lsn(0x10),
8381 6 : space: Lsn(0x10),
8382 6 : },
8383 6 : leases: Default::default(),
8384 6 : within_ancestor_pitr: false,
8385 6 : };
8386 6 : }
8387 6 :
8388 6 : {
8389 6 : // Update GC info
8390 6 : let mut guard = branch_tline.gc_info.write().unwrap();
8391 6 : *guard = GcInfo {
8392 6 : retain_lsns: vec![(Lsn(0x40), branch_tline.timeline_id)],
8393 6 : cutoffs: GcCutoffs {
8394 6 : time: Lsn(0x50),
8395 6 : space: Lsn(0x50),
8396 6 : },
8397 6 : leases: Default::default(),
8398 6 : within_ancestor_pitr: false,
8399 6 : };
8400 6 : }
8401 6 :
8402 6 : let expected_result_at_gc_horizon = [
8403 6 : Bytes::from_static(b"value 0@0x10"),
8404 6 : Bytes::from_static(b"value 1@0x10@0x20"),
8405 6 : Bytes::from_static(b"value 2@0x10@0x30"),
8406 6 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
8407 6 : Bytes::from_static(b"value 4@0x10"),
8408 6 : Bytes::from_static(b"value 5@0x10@0x20"),
8409 6 : Bytes::from_static(b"value 6@0x10@0x20"),
8410 6 : Bytes::from_static(b"value 7@0x10"),
8411 6 : Bytes::from_static(b"value 8@0x10@0x48"),
8412 6 : Bytes::from_static(b"value 9@0x10@0x48"),
8413 6 : ];
8414 6 :
8415 6 : let expected_result_at_lsn_40 = [
8416 6 : Bytes::from_static(b"value 0@0x10"),
8417 6 : Bytes::from_static(b"value 1@0x10@0x20"),
8418 6 : Bytes::from_static(b"value 2@0x10@0x30"),
8419 6 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
8420 6 : Bytes::from_static(b"value 4@0x10"),
8421 6 : Bytes::from_static(b"value 5@0x10@0x20"),
8422 6 : Bytes::from_static(b"value 6@0x10@0x20"),
8423 6 : Bytes::from_static(b"value 7@0x10"),
8424 6 : Bytes::from_static(b"value 8@0x10"),
8425 6 : Bytes::from_static(b"value 9@0x10"),
8426 6 : ];
8427 6 :
8428 12 : let verify_result = || async {
8429 132 : for idx in 0..10 {
8430 120 : assert_eq!(
8431 120 : branch_tline
8432 120 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8433 156 : .await
8434 120 : .unwrap(),
8435 120 : &expected_result_at_gc_horizon[idx]
8436 12 : );
8437 120 : assert_eq!(
8438 120 : branch_tline
8439 120 : .get(get_key(idx as u32), Lsn(0x40), &ctx)
8440 93 : .await
8441 120 : .unwrap(),
8442 120 : &expected_result_at_lsn_40[idx]
8443 12 : );
8444 12 : }
8445 12 : };
8446 6 :
8447 138 : verify_result().await;
8448 6 :
8449 6 : let cancel = CancellationToken::new();
8450 6 : branch_tline
8451 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8452 48 : .await
8453 6 : .unwrap();
8454 6 :
8455 111 : verify_result().await;
8456 6 :
8457 6 : Ok(())
8458 6 : }
8459 : }
|
