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("ancestor is archived: {}", .0)]
513 : HasArchivedParent(TimelineId),
514 :
515 : #[error("HasUnarchivedChildren")]
516 : HasUnarchivedChildren(Vec<TimelineId>),
517 :
518 : #[error("Timeline archival is already in progress")]
519 : AlreadyInProgress,
520 :
521 : #[error(transparent)]
522 : Other(#[from] anyhow::Error),
523 : }
524 :
525 : impl Debug for TimelineArchivalError {
526 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
527 0 : match self {
528 0 : Self::NotFound => write!(f, "NotFound"),
529 0 : Self::Timeout => write!(f, "Timeout"),
530 0 : Self::HasArchivedParent(p) => f.debug_tuple("HasArchivedParent").field(p).finish(),
531 0 : Self::HasUnarchivedChildren(c) => {
532 0 : f.debug_tuple("HasUnarchivedChildren").field(c).finish()
533 : }
534 0 : Self::AlreadyInProgress => f.debug_tuple("AlreadyInProgress").finish(),
535 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
536 : }
537 0 : }
538 : }
539 :
540 : pub enum SetStoppingError {
541 : AlreadyStopping(completion::Barrier),
542 : Broken,
543 : }
544 :
545 : impl Debug for SetStoppingError {
546 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
547 0 : match self {
548 0 : Self::AlreadyStopping(_) => f.debug_tuple("AlreadyStopping").finish(),
549 0 : Self::Broken => write!(f, "Broken"),
550 : }
551 0 : }
552 : }
553 :
554 0 : #[derive(thiserror::Error, Debug)]
555 : pub enum CreateTimelineError {
556 : #[error("creation of timeline with the given ID is in progress")]
557 : AlreadyCreating,
558 : #[error("timeline already exists with different parameters")]
559 : Conflict,
560 : #[error(transparent)]
561 : AncestorLsn(anyhow::Error),
562 : #[error("ancestor timeline is not active")]
563 : AncestorNotActive,
564 : #[error("tenant shutting down")]
565 : ShuttingDown,
566 : #[error(transparent)]
567 : Other(#[from] anyhow::Error),
568 : }
569 :
570 : #[derive(thiserror::Error, Debug)]
571 : enum InitdbError {
572 : Other(anyhow::Error),
573 : Cancelled,
574 : Spawn(std::io::Result<()>),
575 : Failed(std::process::ExitStatus, Vec<u8>),
576 : }
577 :
578 : impl fmt::Display for InitdbError {
579 0 : fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
580 0 : match self {
581 0 : InitdbError::Cancelled => write!(f, "Operation was cancelled"),
582 0 : InitdbError::Spawn(e) => write!(f, "Spawn error: {:?}", e),
583 0 : InitdbError::Failed(status, stderr) => write!(
584 0 : f,
585 0 : "Command failed with status {:?}: {}",
586 0 : status,
587 0 : String::from_utf8_lossy(stderr)
588 0 : ),
589 0 : InitdbError::Other(e) => write!(f, "Error: {:?}", e),
590 : }
591 0 : }
592 : }
593 :
594 : impl From<std::io::Error> for InitdbError {
595 0 : fn from(error: std::io::Error) -> Self {
596 0 : InitdbError::Spawn(Err(error))
597 0 : }
598 : }
599 :
600 : enum CreateTimelineCause {
601 : Load,
602 : Delete,
603 : }
604 :
605 0 : #[derive(thiserror::Error, Debug)]
606 : pub(crate) enum GcError {
607 : // The tenant is shutting down
608 : #[error("tenant shutting down")]
609 : TenantCancelled,
610 :
611 : // The tenant is shutting down
612 : #[error("timeline shutting down")]
613 : TimelineCancelled,
614 :
615 : // The tenant is in a state inelegible to run GC
616 : #[error("not active")]
617 : NotActive,
618 :
619 : // A requested GC cutoff LSN was invalid, for example it tried to move backwards
620 : #[error("not active")]
621 : BadLsn { why: String },
622 :
623 : // A remote storage error while scheduling updates after compaction
624 : #[error(transparent)]
625 : Remote(anyhow::Error),
626 :
627 : // An error reading while calculating GC cutoffs
628 : #[error(transparent)]
629 : GcCutoffs(PageReconstructError),
630 :
631 : // If GC was invoked for a particular timeline, this error means it didn't exist
632 : #[error("timeline not found")]
633 : TimelineNotFound,
634 : }
635 :
636 : impl From<PageReconstructError> for GcError {
637 0 : fn from(value: PageReconstructError) -> Self {
638 0 : match value {
639 0 : PageReconstructError::Cancelled => Self::TimelineCancelled,
640 0 : other => Self::GcCutoffs(other),
641 : }
642 0 : }
643 : }
644 :
645 : impl From<NotInitialized> for GcError {
646 0 : fn from(value: NotInitialized) -> Self {
647 0 : match value {
648 0 : NotInitialized::Uninitialized => GcError::Remote(value.into()),
649 0 : NotInitialized::Stopped | NotInitialized::ShuttingDown => GcError::TimelineCancelled,
650 : }
651 0 : }
652 : }
653 :
654 : impl From<timeline::layer_manager::Shutdown> for GcError {
655 0 : fn from(_: timeline::layer_manager::Shutdown) -> Self {
656 0 : GcError::TimelineCancelled
657 0 : }
658 : }
659 :
660 0 : #[derive(thiserror::Error, Debug)]
661 : pub(crate) enum LoadConfigError {
662 : #[error("TOML deserialization error: '{0}'")]
663 : DeserializeToml(#[from] toml_edit::de::Error),
664 :
665 : #[error("Config not found at {0}")]
666 : NotFound(Utf8PathBuf),
667 : }
668 :
669 : impl Tenant {
670 : /// Yet another helper for timeline initialization.
671 : ///
672 : /// - Initializes the Timeline struct and inserts it into the tenant's hash map
673 : /// - Scans the local timeline directory for layer files and builds the layer map
674 : /// - Downloads remote index file and adds remote files to the layer map
675 : /// - Schedules remote upload tasks for any files that are present locally but missing from remote storage.
676 : ///
677 : /// If the operation fails, the timeline is left in the tenant's hash map in Broken state. On success,
678 : /// it is marked as Active.
679 : #[allow(clippy::too_many_arguments)]
680 18 : async fn timeline_init_and_sync(
681 18 : &self,
682 18 : timeline_id: TimelineId,
683 18 : resources: TimelineResources,
684 18 : index_part: Option<IndexPart>,
685 18 : metadata: TimelineMetadata,
686 18 : ancestor: Option<Arc<Timeline>>,
687 18 : last_aux_file_policy: Option<AuxFilePolicy>,
688 18 : _ctx: &RequestContext,
689 18 : ) -> anyhow::Result<()> {
690 18 : let tenant_id = self.tenant_shard_id;
691 :
692 18 : let timeline = self.create_timeline_struct(
693 18 : timeline_id,
694 18 : &metadata,
695 18 : ancestor.clone(),
696 18 : resources,
697 18 : CreateTimelineCause::Load,
698 18 : // This could be derived from ancestor branch + index part. Though the only caller of `timeline_init_and_sync` is `load_remote_timeline`,
699 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.
700 18 : // Therefore, we pass this field explicitly for now, and remove it once we fully migrate to aux file v2.
701 18 : last_aux_file_policy,
702 18 : )?;
703 18 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
704 18 : anyhow::ensure!(
705 18 : disk_consistent_lsn.is_valid(),
706 0 : "Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn"
707 : );
708 18 : assert_eq!(
709 18 : disk_consistent_lsn,
710 18 : metadata.disk_consistent_lsn(),
711 0 : "these are used interchangeably"
712 : );
713 :
714 18 : if let Some(index_part) = index_part.as_ref() {
715 18 : timeline.remote_client.init_upload_queue(index_part)?;
716 :
717 18 : timeline
718 18 : .last_aux_file_policy
719 18 : .store(index_part.last_aux_file_policy());
720 : } else {
721 : // No data on the remote storage, but we have local metadata file. We can end up
722 : // here with timeline_create being interrupted before finishing index part upload.
723 : // By doing what we do here, the index part upload is retried.
724 : // If control plane retries timeline creation in the meantime, the mgmt API handler
725 : // for timeline creation will coalesce on the upload we queue here.
726 :
727 : // FIXME: this branch should be dead code as we no longer write local metadata.
728 :
729 0 : timeline
730 0 : .remote_client
731 0 : .init_upload_queue_for_empty_remote(&metadata)?;
732 0 : timeline
733 0 : .remote_client
734 0 : .schedule_index_upload_for_full_metadata_update(&metadata)?;
735 : }
736 :
737 18 : timeline
738 18 : .load_layer_map(disk_consistent_lsn, index_part)
739 16 : .await
740 18 : .with_context(|| {
741 0 : format!("Failed to load layermap for timeline {tenant_id}/{timeline_id}")
742 18 : })?;
743 :
744 : {
745 : // avoiding holding it across awaits
746 18 : let mut timelines_accessor = self.timelines.lock().unwrap();
747 18 : match timelines_accessor.entry(timeline_id) {
748 : // We should never try and load the same timeline twice during startup
749 : Entry::Occupied(_) => {
750 0 : unreachable!(
751 0 : "Timeline {tenant_id}/{timeline_id} already exists in the tenant map"
752 0 : );
753 : }
754 18 : Entry::Vacant(v) => {
755 18 : v.insert(Arc::clone(&timeline));
756 18 : timeline.maybe_spawn_flush_loop();
757 18 : }
758 18 : }
759 18 : };
760 18 :
761 18 : // Sanity check: a timeline should have some content.
762 18 : anyhow::ensure!(
763 18 : ancestor.is_some()
764 12 : || timeline
765 12 : .layers
766 12 : .read()
767 0 : .await
768 12 : .layer_map()
769 12 : .expect("currently loading, layer manager cannot be shutdown already")
770 12 : .iter_historic_layers()
771 12 : .next()
772 12 : .is_some(),
773 0 : "Timeline has no ancestor and no layer files"
774 : );
775 :
776 18 : Ok(())
777 18 : }
778 :
779 : /// Attach a tenant that's available in cloud storage.
780 : ///
781 : /// This returns quickly, after just creating the in-memory object
782 : /// Tenant struct and launching a background task to download
783 : /// the remote index files. On return, the tenant is most likely still in
784 : /// Attaching state, and it will become Active once the background task
785 : /// finishes. You can use wait_until_active() to wait for the task to
786 : /// complete.
787 : ///
788 : #[allow(clippy::too_many_arguments)]
789 0 : pub(crate) fn spawn(
790 0 : conf: &'static PageServerConf,
791 0 : tenant_shard_id: TenantShardId,
792 0 : resources: TenantSharedResources,
793 0 : attached_conf: AttachedTenantConf,
794 0 : shard_identity: ShardIdentity,
795 0 : init_order: Option<InitializationOrder>,
796 0 : mode: SpawnMode,
797 0 : ctx: &RequestContext,
798 0 : ) -> Result<Arc<Tenant>, GlobalShutDown> {
799 0 : let wal_redo_manager =
800 0 : WalRedoManager::new(PostgresRedoManager::new(conf, tenant_shard_id))?;
801 :
802 : let TenantSharedResources {
803 0 : broker_client,
804 0 : remote_storage,
805 0 : deletion_queue_client,
806 0 : l0_flush_global_state,
807 0 : } = resources;
808 0 :
809 0 : let attach_mode = attached_conf.location.attach_mode;
810 0 : let generation = attached_conf.location.generation;
811 0 :
812 0 : let tenant = Arc::new(Tenant::new(
813 0 : TenantState::Attaching,
814 0 : conf,
815 0 : attached_conf,
816 0 : shard_identity,
817 0 : Some(wal_redo_manager),
818 0 : tenant_shard_id,
819 0 : remote_storage.clone(),
820 0 : deletion_queue_client,
821 0 : l0_flush_global_state,
822 0 : ));
823 0 :
824 0 : // The attach task will carry a GateGuard, so that shutdown() reliably waits for it to drop out if
825 0 : // we shut down while attaching.
826 0 : let attach_gate_guard = tenant
827 0 : .gate
828 0 : .enter()
829 0 : .expect("We just created the Tenant: nothing else can have shut it down yet");
830 0 :
831 0 : // Do all the hard work in the background
832 0 : let tenant_clone = Arc::clone(&tenant);
833 0 : let ctx = ctx.detached_child(TaskKind::Attach, DownloadBehavior::Warn);
834 0 : task_mgr::spawn(
835 0 : &tokio::runtime::Handle::current(),
836 0 : TaskKind::Attach,
837 0 : tenant_shard_id,
838 0 : None,
839 0 : "attach tenant",
840 0 : async move {
841 0 :
842 0 : info!(
843 : ?attach_mode,
844 0 : "Attaching tenant"
845 : );
846 :
847 0 : let _gate_guard = attach_gate_guard;
848 0 :
849 0 : // Is this tenant being spawned as part of process startup?
850 0 : let starting_up = init_order.is_some();
851 : scopeguard::defer! {
852 : if starting_up {
853 : TENANT.startup_complete.inc();
854 : }
855 : }
856 :
857 : // Ideally we should use Tenant::set_broken_no_wait, but it is not supposed to be used when tenant is in loading state.
858 : enum BrokenVerbosity {
859 : Error,
860 : Info
861 : }
862 0 : let make_broken =
863 0 : |t: &Tenant, err: anyhow::Error, verbosity: BrokenVerbosity| {
864 0 : match verbosity {
865 : BrokenVerbosity::Info => {
866 0 : info!("attach cancelled, setting tenant state to Broken: {err}");
867 : },
868 : BrokenVerbosity::Error => {
869 0 : error!("attach failed, setting tenant state to Broken: {err:?}");
870 : }
871 : }
872 0 : t.state.send_modify(|state| {
873 0 : // The Stopping case is for when we have passed control on to DeleteTenantFlow:
874 0 : // if it errors, we will call make_broken when tenant is already in Stopping.
875 0 : assert!(
876 0 : matches!(*state, TenantState::Attaching | TenantState::Stopping { .. }),
877 0 : "the attach task owns the tenant state until activation is complete"
878 : );
879 :
880 0 : *state = TenantState::broken_from_reason(err.to_string());
881 0 : });
882 0 : };
883 :
884 : // TODO: should also be rejecting tenant conf changes that violate this check.
885 0 : if let Err(e) = crate::tenant::storage_layer::inmemory_layer::IndexEntry::validate_checkpoint_distance(tenant_clone.get_checkpoint_distance()) {
886 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
887 0 : return Ok(());
888 0 : }
889 0 :
890 0 : let mut init_order = init_order;
891 0 : // take the completion because initial tenant loading will complete when all of
892 0 : // these tasks complete.
893 0 : let _completion = init_order
894 0 : .as_mut()
895 0 : .and_then(|x| x.initial_tenant_load.take());
896 0 : let remote_load_completion = init_order
897 0 : .as_mut()
898 0 : .and_then(|x| x.initial_tenant_load_remote.take());
899 :
900 : enum AttachType<'a> {
901 : /// We are attaching this tenant lazily in the background.
902 : Warmup {
903 : _permit: tokio::sync::SemaphorePermit<'a>,
904 : during_startup: bool
905 : },
906 : /// We are attaching this tenant as soon as we can, because for example an
907 : /// endpoint tried to access it.
908 : OnDemand,
909 : /// During normal operations after startup, we are attaching a tenant, and
910 : /// eager attach was requested.
911 : Normal,
912 : }
913 :
914 0 : let attach_type = if matches!(mode, SpawnMode::Lazy) {
915 : // Before doing any I/O, wait for at least one of:
916 : // - A client attempting to access to this tenant (on-demand loading)
917 : // - A permit becoming available in the warmup semaphore (background warmup)
918 :
919 : tokio::select!(
920 : permit = tenant_clone.activate_now_sem.acquire() => {
921 : let _ = permit.expect("activate_now_sem is never closed");
922 : tracing::info!("Activating tenant (on-demand)");
923 : AttachType::OnDemand
924 : },
925 : permit = conf.concurrent_tenant_warmup.inner().acquire() => {
926 : let _permit = permit.expect("concurrent_tenant_warmup semaphore is never closed");
927 : tracing::info!("Activating tenant (warmup)");
928 : AttachType::Warmup {
929 : _permit,
930 : during_startup: init_order.is_some()
931 : }
932 : }
933 : _ = tenant_clone.cancel.cancelled() => {
934 : // This is safe, but should be pretty rare: it is interesting if a tenant
935 : // stayed in Activating for such a long time that shutdown found it in
936 : // that state.
937 : tracing::info!(state=%tenant_clone.current_state(), "Tenant shut down before activation");
938 : // Make the tenant broken so that set_stopping will not hang waiting for it to leave
939 : // the Attaching state. This is an over-reaction (nothing really broke, the tenant is
940 : // just shutting down), but ensures progress.
941 : make_broken(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"), BrokenVerbosity::Info);
942 : return Ok(());
943 : },
944 : )
945 : } else {
946 : // SpawnMode::{Create,Eager} always cause jumping ahead of the
947 : // concurrent_tenant_warmup queue
948 0 : AttachType::Normal
949 : };
950 :
951 0 : let preload = match &mode {
952 : SpawnMode::Eager | SpawnMode::Lazy => {
953 0 : let _preload_timer = TENANT.preload.start_timer();
954 0 : let res = tenant_clone
955 0 : .preload(&remote_storage, task_mgr::shutdown_token())
956 0 : .await;
957 0 : match res {
958 0 : Ok(p) => Some(p),
959 0 : Err(e) => {
960 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
961 0 : return Ok(());
962 : }
963 : }
964 : }
965 :
966 : };
967 :
968 : // Remote preload is complete.
969 0 : drop(remote_load_completion);
970 :
971 : // We will time the duration of the attach phase unless this is a creation (attach will do no work)
972 0 : let attached = {
973 0 : let _attach_timer = Some(TENANT.attach.start_timer());
974 0 : tenant_clone.attach(preload, &ctx).await
975 : };
976 :
977 0 : match attached {
978 : Ok(()) => {
979 0 : info!("attach finished, activating");
980 0 : tenant_clone.activate(broker_client, None, &ctx);
981 : }
982 0 : Err(e) => {
983 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
984 0 : }
985 : }
986 :
987 : // If we are doing an opportunistic warmup attachment at startup, initialize
988 : // logical size at the same time. This is better than starting a bunch of idle tenants
989 : // with cold caches and then coming back later to initialize their logical sizes.
990 : //
991 : // It also prevents the warmup proccess competing with the concurrency limit on
992 : // logical size calculations: if logical size calculation semaphore is saturated,
993 : // then warmup will wait for that before proceeding to the next tenant.
994 0 : if matches!(attach_type, AttachType::Warmup { during_startup: true, .. }) {
995 0 : let mut futs: FuturesUnordered<_> = tenant_clone.timelines.lock().unwrap().values().cloned().map(|t| t.await_initial_logical_size()).collect();
996 0 : tracing::info!("Waiting for initial logical sizes while warming up...");
997 0 : while futs.next().await.is_some() {}
998 0 : tracing::info!("Warm-up complete");
999 0 : }
1000 :
1001 0 : Ok(())
1002 0 : }
1003 0 : .instrument(tracing::info_span!(parent: None, "attach", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), gen=?generation)),
1004 : );
1005 0 : Ok(tenant)
1006 0 : }
1007 :
1008 1128 : #[instrument(skip_all)]
1009 : pub(crate) async fn preload(
1010 : self: &Arc<Self>,
1011 : remote_storage: &GenericRemoteStorage,
1012 : cancel: CancellationToken,
1013 : ) -> anyhow::Result<TenantPreload> {
1014 : span::debug_assert_current_span_has_tenant_id();
1015 : // Get list of remote timelines
1016 : // download index files for every tenant timeline
1017 : info!("listing remote timelines");
1018 : let (remote_timeline_ids, other_keys) = remote_timeline_client::list_remote_timelines(
1019 : remote_storage,
1020 : self.tenant_shard_id,
1021 : cancel.clone(),
1022 : )
1023 : .await?;
1024 :
1025 : info!("found {} timelines", remote_timeline_ids.len(),);
1026 :
1027 : for k in other_keys {
1028 : warn!("Unexpected non timeline key {k}");
1029 : }
1030 :
1031 : Ok(TenantPreload {
1032 : timelines: Self::load_timeline_metadata(
1033 : self,
1034 : remote_timeline_ids,
1035 : remote_storage,
1036 : cancel,
1037 : )
1038 : .await?,
1039 : })
1040 : }
1041 :
1042 : ///
1043 : /// Background task that downloads all data for a tenant and brings it to Active state.
1044 : ///
1045 : /// No background tasks are started as part of this routine.
1046 : ///
1047 564 : async fn attach(
1048 564 : self: &Arc<Tenant>,
1049 564 : preload: Option<TenantPreload>,
1050 564 : ctx: &RequestContext,
1051 564 : ) -> anyhow::Result<()> {
1052 564 : span::debug_assert_current_span_has_tenant_id();
1053 564 :
1054 564 : failpoint_support::sleep_millis_async!("before-attaching-tenant");
1055 :
1056 564 : let Some(preload) = preload else {
1057 0 : anyhow::bail!("local-only deployment is no longer supported, https://github.com/neondatabase/neon/issues/5624");
1058 : };
1059 :
1060 564 : let mut timelines_to_resume_deletions = vec![];
1061 564 :
1062 564 : let mut remote_index_and_client = HashMap::new();
1063 564 : let mut timeline_ancestors = HashMap::new();
1064 564 : let mut existent_timelines = HashSet::new();
1065 582 : for (timeline_id, preload) in preload.timelines {
1066 18 : let index_part = match preload.index_part {
1067 18 : Ok(i) => {
1068 18 : debug!("remote index part exists for timeline {timeline_id}");
1069 : // We found index_part on the remote, this is the standard case.
1070 18 : existent_timelines.insert(timeline_id);
1071 18 : i
1072 : }
1073 : Err(DownloadError::NotFound) => {
1074 : // There is no index_part on the remote. We only get here
1075 : // if there is some prefix for the timeline in the remote storage.
1076 : // This can e.g. be the initdb.tar.zst archive, maybe a
1077 : // remnant from a prior incomplete creation or deletion attempt.
1078 : // Delete the local directory as the deciding criterion for a
1079 : // timeline's existence is presence of index_part.
1080 0 : info!(%timeline_id, "index_part not found on remote");
1081 0 : continue;
1082 : }
1083 0 : Err(e) => {
1084 0 : // Some (possibly ephemeral) error happened during index_part download.
1085 0 : // Pretend the timeline exists to not delete the timeline directory,
1086 0 : // as it might be a temporary issue and we don't want to re-download
1087 0 : // everything after it resolves.
1088 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1089 :
1090 0 : existent_timelines.insert(timeline_id);
1091 0 : continue;
1092 : }
1093 : };
1094 18 : match index_part {
1095 18 : MaybeDeletedIndexPart::IndexPart(index_part) => {
1096 18 : timeline_ancestors.insert(timeline_id, index_part.metadata.clone());
1097 18 : remote_index_and_client.insert(timeline_id, (index_part, preload.client));
1098 18 : }
1099 0 : MaybeDeletedIndexPart::Deleted(index_part) => {
1100 0 : info!(
1101 0 : "timeline {} is deleted, picking to resume deletion",
1102 : timeline_id
1103 : );
1104 0 : timelines_to_resume_deletions.push((timeline_id, index_part, preload.client));
1105 : }
1106 : }
1107 : }
1108 :
1109 564 : let mut gc_blocks = HashMap::new();
1110 :
1111 : // For every timeline, download the metadata file, scan the local directory,
1112 : // and build a layer map that contains an entry for each remote and local
1113 : // layer file.
1114 564 : let sorted_timelines = tree_sort_timelines(timeline_ancestors, |m| m.ancestor_timeline())?;
1115 582 : for (timeline_id, remote_metadata) in sorted_timelines {
1116 18 : let (index_part, remote_client) = remote_index_and_client
1117 18 : .remove(&timeline_id)
1118 18 : .expect("just put it in above");
1119 :
1120 18 : if let Some(blocking) = index_part.gc_blocking.as_ref() {
1121 : // could just filter these away, but it helps while testing
1122 0 : anyhow::ensure!(
1123 0 : !blocking.reasons.is_empty(),
1124 0 : "index_part for {timeline_id} is malformed: it should not have gc blocking with zero reasons"
1125 : );
1126 0 : let prev = gc_blocks.insert(timeline_id, blocking.reasons);
1127 0 : assert!(prev.is_none());
1128 18 : }
1129 :
1130 : // TODO again handle early failure
1131 18 : self.load_remote_timeline(
1132 18 : timeline_id,
1133 18 : index_part,
1134 18 : remote_metadata,
1135 18 : TimelineResources {
1136 18 : remote_client,
1137 18 : timeline_get_throttle: self.timeline_get_throttle.clone(),
1138 18 : l0_flush_global_state: self.l0_flush_global_state.clone(),
1139 18 : },
1140 18 : ctx,
1141 18 : )
1142 32 : .await
1143 18 : .with_context(|| {
1144 0 : format!(
1145 0 : "failed to load remote timeline {} for tenant {}",
1146 0 : timeline_id, self.tenant_shard_id
1147 0 : )
1148 18 : })?;
1149 : }
1150 :
1151 : // Walk through deleted timelines, resume deletion
1152 564 : for (timeline_id, index_part, remote_timeline_client) in timelines_to_resume_deletions {
1153 0 : remote_timeline_client
1154 0 : .init_upload_queue_stopped_to_continue_deletion(&index_part)
1155 0 : .context("init queue stopped")
1156 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1157 :
1158 0 : DeleteTimelineFlow::resume_deletion(
1159 0 : Arc::clone(self),
1160 0 : timeline_id,
1161 0 : &index_part.metadata,
1162 0 : remote_timeline_client,
1163 0 : )
1164 0 : .instrument(tracing::info_span!("timeline_delete", %timeline_id))
1165 0 : .await
1166 0 : .context("resume_deletion")
1167 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1168 : }
1169 :
1170 : // The local filesystem contents are a cache of what's in the remote IndexPart;
1171 : // IndexPart is the source of truth.
1172 564 : self.clean_up_timelines(&existent_timelines)?;
1173 :
1174 564 : self.gc_block.set_scanned(gc_blocks);
1175 564 :
1176 564 : fail::fail_point!("attach-before-activate", |_| {
1177 0 : anyhow::bail!("attach-before-activate");
1178 564 : });
1179 564 : failpoint_support::sleep_millis_async!("attach-before-activate-sleep", &self.cancel);
1180 :
1181 564 : info!("Done");
1182 :
1183 564 : Ok(())
1184 564 : }
1185 :
1186 : /// Check for any local timeline directories that are temporary, or do not correspond to a
1187 : /// timeline that still exists: this can happen if we crashed during a deletion/creation, or
1188 : /// if a timeline was deleted while the tenant was attached to a different pageserver.
1189 564 : fn clean_up_timelines(&self, existent_timelines: &HashSet<TimelineId>) -> anyhow::Result<()> {
1190 564 : let timelines_dir = self.conf.timelines_path(&self.tenant_shard_id);
1191 :
1192 564 : let entries = match timelines_dir.read_dir_utf8() {
1193 564 : Ok(d) => d,
1194 0 : Err(e) => {
1195 0 : if e.kind() == std::io::ErrorKind::NotFound {
1196 0 : return Ok(());
1197 : } else {
1198 0 : return Err(e).context("list timelines directory for tenant");
1199 : }
1200 : }
1201 : };
1202 :
1203 588 : for entry in entries {
1204 24 : let entry = entry.context("read timeline dir entry")?;
1205 24 : let entry_path = entry.path();
1206 :
1207 24 : let purge = if crate::is_temporary(entry_path)
1208 : // TODO: remove uninit mark code (https://github.com/neondatabase/neon/issues/5718)
1209 24 : || is_uninit_mark(entry_path)
1210 24 : || crate::is_delete_mark(entry_path)
1211 : {
1212 0 : true
1213 : } else {
1214 24 : match TimelineId::try_from(entry_path.file_name()) {
1215 24 : Ok(i) => {
1216 24 : // Purge if the timeline ID does not exist in remote storage: remote storage is the authority.
1217 24 : !existent_timelines.contains(&i)
1218 : }
1219 0 : Err(e) => {
1220 0 : tracing::warn!(
1221 0 : "Unparseable directory in timelines directory: {entry_path}, ignoring ({e})"
1222 : );
1223 : // Do not purge junk: if we don't recognize it, be cautious and leave it for a human.
1224 0 : false
1225 : }
1226 : }
1227 : };
1228 :
1229 24 : if purge {
1230 6 : tracing::info!("Purging stale timeline dentry {entry_path}");
1231 6 : if let Err(e) = match entry.file_type() {
1232 6 : Ok(t) => if t.is_dir() {
1233 6 : std::fs::remove_dir_all(entry_path)
1234 : } else {
1235 0 : std::fs::remove_file(entry_path)
1236 : }
1237 6 : .or_else(fs_ext::ignore_not_found),
1238 0 : Err(e) => Err(e),
1239 : } {
1240 0 : tracing::warn!("Failed to purge stale timeline dentry {entry_path}: {e}");
1241 6 : }
1242 18 : }
1243 : }
1244 :
1245 564 : Ok(())
1246 564 : }
1247 :
1248 : /// Get sum of all remote timelines sizes
1249 : ///
1250 : /// This function relies on the index_part instead of listing the remote storage
1251 0 : pub fn remote_size(&self) -> u64 {
1252 0 : let mut size = 0;
1253 :
1254 0 : for timeline in self.list_timelines() {
1255 0 : size += timeline.remote_client.get_remote_physical_size();
1256 0 : }
1257 :
1258 0 : size
1259 0 : }
1260 :
1261 36 : #[instrument(skip_all, fields(timeline_id=%timeline_id))]
1262 : async fn load_remote_timeline(
1263 : &self,
1264 : timeline_id: TimelineId,
1265 : index_part: IndexPart,
1266 : remote_metadata: TimelineMetadata,
1267 : resources: TimelineResources,
1268 : ctx: &RequestContext,
1269 : ) -> anyhow::Result<()> {
1270 : span::debug_assert_current_span_has_tenant_id();
1271 :
1272 : info!("downloading index file for timeline {}", timeline_id);
1273 : tokio::fs::create_dir_all(self.conf.timeline_path(&self.tenant_shard_id, &timeline_id))
1274 : .await
1275 : .context("Failed to create new timeline directory")?;
1276 :
1277 : let ancestor = if let Some(ancestor_id) = remote_metadata.ancestor_timeline() {
1278 : let timelines = self.timelines.lock().unwrap();
1279 : Some(Arc::clone(timelines.get(&ancestor_id).ok_or_else(
1280 0 : || {
1281 0 : anyhow::anyhow!(
1282 0 : "cannot find ancestor timeline {ancestor_id} for timeline {timeline_id}"
1283 0 : )
1284 0 : },
1285 : )?))
1286 : } else {
1287 : None
1288 : };
1289 :
1290 : let last_aux_file_policy = index_part.last_aux_file_policy();
1291 :
1292 : self.timeline_init_and_sync(
1293 : timeline_id,
1294 : resources,
1295 : Some(index_part),
1296 : remote_metadata,
1297 : ancestor,
1298 : last_aux_file_policy,
1299 : ctx,
1300 : )
1301 : .await
1302 : }
1303 :
1304 564 : async fn load_timeline_metadata(
1305 564 : self: &Arc<Tenant>,
1306 564 : timeline_ids: HashSet<TimelineId>,
1307 564 : remote_storage: &GenericRemoteStorage,
1308 564 : cancel: CancellationToken,
1309 564 : ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
1310 564 : let mut part_downloads = JoinSet::new();
1311 582 : for timeline_id in timeline_ids {
1312 18 : let client = RemoteTimelineClient::new(
1313 18 : remote_storage.clone(),
1314 18 : self.deletion_queue_client.clone(),
1315 18 : self.conf,
1316 18 : self.tenant_shard_id,
1317 18 : timeline_id,
1318 18 : self.generation,
1319 18 : );
1320 18 : let cancel_clone = cancel.clone();
1321 18 : part_downloads.spawn(
1322 18 : async move {
1323 18 : debug!("starting index part download");
1324 :
1325 66 : let index_part = client.download_index_file(&cancel_clone).await;
1326 :
1327 18 : debug!("finished index part download");
1328 :
1329 18 : Result::<_, anyhow::Error>::Ok(TimelinePreload {
1330 18 : client,
1331 18 : timeline_id,
1332 18 : index_part,
1333 18 : })
1334 18 : }
1335 18 : .map(move |res| {
1336 18 : res.with_context(|| format!("download index part for timeline {timeline_id}"))
1337 18 : })
1338 18 : .instrument(info_span!("download_index_part", %timeline_id)),
1339 : );
1340 : }
1341 :
1342 564 : let mut timeline_preloads: HashMap<TimelineId, TimelinePreload> = HashMap::new();
1343 :
1344 582 : loop {
1345 582 : tokio::select!(
1346 : next = part_downloads.join_next() => {
1347 : match next {
1348 : Some(result) => {
1349 : let preload_result = result.context("join preload task")?;
1350 : let preload = preload_result?;
1351 : timeline_preloads.insert(preload.timeline_id, preload);
1352 : },
1353 : None => {
1354 : break;
1355 : }
1356 : }
1357 : },
1358 : _ = cancel.cancelled() => {
1359 : anyhow::bail!("Cancelled while waiting for remote index download")
1360 : }
1361 582 : )
1362 582 : }
1363 :
1364 564 : Ok(timeline_preloads)
1365 564 : }
1366 :
1367 0 : pub(crate) async fn apply_timeline_archival_config(
1368 0 : &self,
1369 0 : timeline_id: TimelineId,
1370 0 : state: TimelineArchivalState,
1371 0 : ) -> Result<(), TimelineArchivalError> {
1372 0 : info!("setting timeline archival config");
1373 0 : let timeline = {
1374 0 : let timelines = self.timelines.lock().unwrap();
1375 :
1376 0 : let Some(timeline) = timelines.get(&timeline_id) else {
1377 0 : return Err(TimelineArchivalError::NotFound);
1378 : };
1379 :
1380 0 : if state == TimelineArchivalState::Unarchived {
1381 0 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
1382 0 : if ancestor_timeline.is_archived() == Some(true) {
1383 0 : return Err(TimelineArchivalError::HasArchivedParent(
1384 0 : ancestor_timeline.timeline_id,
1385 0 : ));
1386 0 : }
1387 0 : }
1388 0 : }
1389 :
1390 : // Ensure that there are no non-archived child timelines
1391 0 : let children: Vec<TimelineId> = timelines
1392 0 : .iter()
1393 0 : .filter_map(|(id, entry)| {
1394 0 : if entry.get_ancestor_timeline_id() != Some(timeline_id) {
1395 0 : return None;
1396 0 : }
1397 0 : if entry.is_archived() == Some(true) {
1398 0 : return None;
1399 0 : }
1400 0 : Some(*id)
1401 0 : })
1402 0 : .collect();
1403 0 :
1404 0 : if !children.is_empty() && state == TimelineArchivalState::Archived {
1405 0 : return Err(TimelineArchivalError::HasUnarchivedChildren(children));
1406 0 : }
1407 0 : Arc::clone(timeline)
1408 : };
1409 :
1410 0 : let upload_needed = timeline
1411 0 : .remote_client
1412 0 : .schedule_index_upload_for_timeline_archival_state(state)?;
1413 :
1414 0 : if upload_needed {
1415 0 : info!("Uploading new state");
1416 : const MAX_WAIT: Duration = Duration::from_secs(10);
1417 0 : let Ok(v) =
1418 0 : tokio::time::timeout(MAX_WAIT, timeline.remote_client.wait_completion()).await
1419 : else {
1420 0 : tracing::warn!("reached timeout for waiting on upload queue");
1421 0 : return Err(TimelineArchivalError::Timeout);
1422 : };
1423 0 : v.map_err(|e| TimelineArchivalError::Other(anyhow::anyhow!(e)))?;
1424 0 : }
1425 0 : Ok(())
1426 0 : }
1427 :
1428 12 : pub(crate) fn tenant_shard_id(&self) -> TenantShardId {
1429 12 : self.tenant_shard_id
1430 12 : }
1431 :
1432 : /// Get Timeline handle for given Neon timeline ID.
1433 : /// This function is idempotent. It doesn't change internal state in any way.
1434 666 : pub fn get_timeline(
1435 666 : &self,
1436 666 : timeline_id: TimelineId,
1437 666 : active_only: bool,
1438 666 : ) -> Result<Arc<Timeline>, GetTimelineError> {
1439 666 : let timelines_accessor = self.timelines.lock().unwrap();
1440 666 : let timeline = timelines_accessor
1441 666 : .get(&timeline_id)
1442 666 : .ok_or(GetTimelineError::NotFound {
1443 666 : tenant_id: self.tenant_shard_id,
1444 666 : timeline_id,
1445 666 : })?;
1446 :
1447 660 : if active_only && !timeline.is_active() {
1448 0 : Err(GetTimelineError::NotActive {
1449 0 : tenant_id: self.tenant_shard_id,
1450 0 : timeline_id,
1451 0 : state: timeline.current_state(),
1452 0 : })
1453 : } else {
1454 660 : Ok(Arc::clone(timeline))
1455 : }
1456 666 : }
1457 :
1458 : /// Lists timelines the tenant contains.
1459 : /// Up to tenant's implementation to omit certain timelines that ar not considered ready for use.
1460 24 : pub fn list_timelines(&self) -> Vec<Arc<Timeline>> {
1461 24 : self.timelines
1462 24 : .lock()
1463 24 : .unwrap()
1464 24 : .values()
1465 24 : .map(Arc::clone)
1466 24 : .collect()
1467 24 : }
1468 :
1469 0 : pub fn list_timeline_ids(&self) -> Vec<TimelineId> {
1470 0 : self.timelines.lock().unwrap().keys().cloned().collect()
1471 0 : }
1472 :
1473 : /// This is used to create the initial 'main' timeline during bootstrapping,
1474 : /// or when importing a new base backup. The caller is expected to load an
1475 : /// initial image of the datadir to the new timeline after this.
1476 : ///
1477 : /// Until that happens, the on-disk state is invalid (disk_consistent_lsn=Lsn(0))
1478 : /// and the timeline will fail to load at a restart.
1479 : ///
1480 : /// For tests, use `DatadirModification::init_empty_test_timeline` + `commit` to setup the
1481 : /// minimum amount of keys required to get a writable timeline.
1482 : /// (Without it, `put` might fail due to `repartition` failing.)
1483 540 : pub(crate) async fn create_empty_timeline(
1484 540 : &self,
1485 540 : new_timeline_id: TimelineId,
1486 540 : initdb_lsn: Lsn,
1487 540 : pg_version: u32,
1488 540 : _ctx: &RequestContext,
1489 540 : ) -> anyhow::Result<UninitializedTimeline> {
1490 540 : anyhow::ensure!(
1491 540 : self.is_active(),
1492 0 : "Cannot create empty timelines on inactive tenant"
1493 : );
1494 :
1495 : // Protect against concurrent attempts to use this TimelineId
1496 540 : let create_guard = self.create_timeline_create_guard(new_timeline_id)?;
1497 :
1498 534 : let new_metadata = TimelineMetadata::new(
1499 534 : // Initialize disk_consistent LSN to 0, The caller must import some data to
1500 534 : // make it valid, before calling finish_creation()
1501 534 : Lsn(0),
1502 534 : None,
1503 534 : None,
1504 534 : Lsn(0),
1505 534 : initdb_lsn,
1506 534 : initdb_lsn,
1507 534 : pg_version,
1508 534 : );
1509 534 : self.prepare_new_timeline(
1510 534 : new_timeline_id,
1511 534 : &new_metadata,
1512 534 : create_guard,
1513 534 : initdb_lsn,
1514 534 : None,
1515 534 : None,
1516 534 : )
1517 0 : .await
1518 540 : }
1519 :
1520 : /// Helper for unit tests to create an empty timeline.
1521 : ///
1522 : /// The timeline is has state value `Active` but its background loops are not running.
1523 : // This makes the various functions which anyhow::ensure! for Active state work in tests.
1524 : // Our current tests don't need the background loops.
1525 : #[cfg(test)]
1526 510 : pub async fn create_test_timeline(
1527 510 : &self,
1528 510 : new_timeline_id: TimelineId,
1529 510 : initdb_lsn: Lsn,
1530 510 : pg_version: u32,
1531 510 : ctx: &RequestContext,
1532 510 : ) -> anyhow::Result<Arc<Timeline>> {
1533 510 : let uninit_tl = self
1534 510 : .create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
1535 0 : .await?;
1536 510 : let tline = uninit_tl.raw_timeline().expect("we just created it");
1537 510 : assert_eq!(tline.get_last_record_lsn(), Lsn(0));
1538 :
1539 : // Setup minimum keys required for the timeline to be usable.
1540 510 : let mut modification = tline.begin_modification(initdb_lsn);
1541 510 : modification
1542 510 : .init_empty_test_timeline()
1543 510 : .context("init_empty_test_timeline")?;
1544 510 : modification
1545 510 : .commit(ctx)
1546 492 : .await
1547 510 : .context("commit init_empty_test_timeline modification")?;
1548 :
1549 : // Flush to disk so that uninit_tl's check for valid disk_consistent_lsn passes.
1550 510 : tline.maybe_spawn_flush_loop();
1551 510 : tline.freeze_and_flush().await.context("freeze_and_flush")?;
1552 :
1553 : // Make sure the freeze_and_flush reaches remote storage.
1554 510 : tline.remote_client.wait_completion().await.unwrap();
1555 :
1556 510 : let tl = uninit_tl.finish_creation()?;
1557 : // The non-test code would call tl.activate() here.
1558 510 : tl.set_state(TimelineState::Active);
1559 510 : Ok(tl)
1560 510 : }
1561 :
1562 : /// Helper for unit tests to create a timeline with some pre-loaded states.
1563 : #[cfg(test)]
1564 : #[allow(clippy::too_many_arguments)]
1565 78 : pub async fn create_test_timeline_with_layers(
1566 78 : &self,
1567 78 : new_timeline_id: TimelineId,
1568 78 : initdb_lsn: Lsn,
1569 78 : pg_version: u32,
1570 78 : ctx: &RequestContext,
1571 78 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
1572 78 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
1573 78 : end_lsn: Lsn,
1574 78 : ) -> anyhow::Result<Arc<Timeline>> {
1575 78 : let tline = self
1576 78 : .create_test_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
1577 156 : .await?;
1578 78 : tline.force_advance_lsn(end_lsn);
1579 234 : for deltas in delta_layer_desc {
1580 156 : tline
1581 156 : .force_create_delta_layer(deltas, Some(initdb_lsn), ctx)
1582 468 : .await?;
1583 : }
1584 204 : for (lsn, images) in image_layer_desc {
1585 126 : tline
1586 126 : .force_create_image_layer(lsn, images, Some(initdb_lsn), ctx)
1587 717 : .await?;
1588 : }
1589 78 : Ok(tline)
1590 78 : }
1591 :
1592 : /// Create a new timeline.
1593 : ///
1594 : /// Returns the new timeline ID and reference to its Timeline object.
1595 : ///
1596 : /// If the caller specified the timeline ID to use (`new_timeline_id`), and timeline with
1597 : /// the same timeline ID already exists, returns CreateTimelineError::AlreadyExists.
1598 : #[allow(clippy::too_many_arguments)]
1599 0 : pub(crate) async fn create_timeline(
1600 0 : self: &Arc<Tenant>,
1601 0 : new_timeline_id: TimelineId,
1602 0 : ancestor_timeline_id: Option<TimelineId>,
1603 0 : mut ancestor_start_lsn: Option<Lsn>,
1604 0 : pg_version: u32,
1605 0 : load_existing_initdb: Option<TimelineId>,
1606 0 : broker_client: storage_broker::BrokerClientChannel,
1607 0 : ctx: &RequestContext,
1608 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
1609 0 : if !self.is_active() {
1610 0 : if matches!(self.current_state(), TenantState::Stopping { .. }) {
1611 0 : return Err(CreateTimelineError::ShuttingDown);
1612 : } else {
1613 0 : return Err(CreateTimelineError::Other(anyhow::anyhow!(
1614 0 : "Cannot create timelines on inactive tenant"
1615 0 : )));
1616 : }
1617 0 : }
1618 :
1619 0 : let _gate = self
1620 0 : .gate
1621 0 : .enter()
1622 0 : .map_err(|_| CreateTimelineError::ShuttingDown)?;
1623 :
1624 : // Get exclusive access to the timeline ID: this ensures that it does not already exist,
1625 : // and that no other creation attempts will be allowed in while we are working.
1626 0 : let create_guard = match self.create_timeline_create_guard(new_timeline_id) {
1627 0 : Ok(m) => m,
1628 : Err(TimelineExclusionError::AlreadyCreating) => {
1629 : // Creation is in progress, we cannot create it again, and we cannot
1630 : // check if this request matches the existing one, so caller must try
1631 : // again later.
1632 0 : return Err(CreateTimelineError::AlreadyCreating);
1633 : }
1634 0 : Err(TimelineExclusionError::Other(e)) => {
1635 0 : return Err(CreateTimelineError::Other(e));
1636 : }
1637 0 : Err(TimelineExclusionError::AlreadyExists(existing)) => {
1638 0 : debug!("timeline {new_timeline_id} already exists");
1639 :
1640 : // Idempotency: creating the same timeline twice is not an error, unless
1641 : // the second creation has different parameters.
1642 0 : if existing.get_ancestor_timeline_id() != ancestor_timeline_id
1643 0 : || existing.pg_version != pg_version
1644 0 : || (ancestor_start_lsn.is_some()
1645 0 : && ancestor_start_lsn != Some(existing.get_ancestor_lsn()))
1646 : {
1647 0 : return Err(CreateTimelineError::Conflict);
1648 0 : }
1649 0 :
1650 0 : // Wait for uploads to complete, so that when we return Ok, the timeline
1651 0 : // is known to be durable on remote storage. Just like we do at the end of
1652 0 : // this function, after we have created the timeline ourselves.
1653 0 : //
1654 0 : // We only really care that the initial version of `index_part.json` has
1655 0 : // been uploaded. That's enough to remember that the timeline
1656 0 : // exists. However, there is no function to wait specifically for that so
1657 0 : // we just wait for all in-progress uploads to finish.
1658 0 : existing
1659 0 : .remote_client
1660 0 : .wait_completion()
1661 0 : .await
1662 0 : .context("wait for timeline uploads to complete")?;
1663 :
1664 0 : return Ok(existing);
1665 : }
1666 : };
1667 :
1668 : pausable_failpoint!("timeline-creation-after-uninit");
1669 :
1670 0 : let loaded_timeline = match ancestor_timeline_id {
1671 0 : Some(ancestor_timeline_id) => {
1672 0 : let ancestor_timeline = self
1673 0 : .get_timeline(ancestor_timeline_id, false)
1674 0 : .context("Cannot branch off the timeline that's not present in pageserver")?;
1675 :
1676 : // instead of waiting around, just deny the request because ancestor is not yet
1677 : // ready for other purposes either.
1678 0 : if !ancestor_timeline.is_active() {
1679 0 : return Err(CreateTimelineError::AncestorNotActive);
1680 0 : }
1681 :
1682 0 : if let Some(lsn) = ancestor_start_lsn.as_mut() {
1683 0 : *lsn = lsn.align();
1684 0 :
1685 0 : let ancestor_ancestor_lsn = ancestor_timeline.get_ancestor_lsn();
1686 0 : if ancestor_ancestor_lsn > *lsn {
1687 : // can we safely just branch from the ancestor instead?
1688 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
1689 0 : "invalid start lsn {} for ancestor timeline {}: less than timeline ancestor lsn {}",
1690 0 : lsn,
1691 0 : ancestor_timeline_id,
1692 0 : ancestor_ancestor_lsn,
1693 0 : )));
1694 0 : }
1695 0 :
1696 0 : // Wait for the WAL to arrive and be processed on the parent branch up
1697 0 : // to the requested branch point. The repository code itself doesn't
1698 0 : // require it, but if we start to receive WAL on the new timeline,
1699 0 : // decoding the new WAL might need to look up previous pages, relation
1700 0 : // sizes etc. and that would get confused if the previous page versions
1701 0 : // are not in the repository yet.
1702 0 : ancestor_timeline
1703 0 : .wait_lsn(*lsn, timeline::WaitLsnWaiter::Tenant, ctx)
1704 0 : .await
1705 0 : .map_err(|e| match e {
1706 0 : e @ (WaitLsnError::Timeout(_) | WaitLsnError::BadState { .. }) => {
1707 0 : CreateTimelineError::AncestorLsn(anyhow::anyhow!(e))
1708 : }
1709 0 : WaitLsnError::Shutdown => CreateTimelineError::ShuttingDown,
1710 0 : })?;
1711 0 : }
1712 :
1713 0 : self.branch_timeline(
1714 0 : &ancestor_timeline,
1715 0 : new_timeline_id,
1716 0 : ancestor_start_lsn,
1717 0 : create_guard,
1718 0 : ctx,
1719 0 : )
1720 0 : .await?
1721 : }
1722 : None => {
1723 0 : self.bootstrap_timeline(
1724 0 : new_timeline_id,
1725 0 : pg_version,
1726 0 : load_existing_initdb,
1727 0 : create_guard,
1728 0 : ctx,
1729 0 : )
1730 0 : .await?
1731 : }
1732 : };
1733 :
1734 : // At this point we have dropped our guard on [`Self::timelines_creating`], and
1735 : // the timeline is visible in [`Self::timelines`], but it is _not_ durable yet. We must
1736 : // not send a success to the caller until it is. The same applies to handling retries,
1737 : // see the handling of [`TimelineExclusionError::AlreadyExists`] above.
1738 0 : let kind = ancestor_timeline_id
1739 0 : .map(|_| "branched")
1740 0 : .unwrap_or("bootstrapped");
1741 0 : loaded_timeline
1742 0 : .remote_client
1743 0 : .wait_completion()
1744 0 : .await
1745 0 : .with_context(|| format!("wait for {} timeline initial uploads to complete", kind))?;
1746 :
1747 0 : loaded_timeline.activate(self.clone(), broker_client, None, ctx);
1748 0 :
1749 0 : Ok(loaded_timeline)
1750 0 : }
1751 :
1752 0 : pub(crate) async fn delete_timeline(
1753 0 : self: Arc<Self>,
1754 0 : timeline_id: TimelineId,
1755 0 : ) -> Result<(), DeleteTimelineError> {
1756 0 : DeleteTimelineFlow::run(&self, timeline_id).await?;
1757 :
1758 0 : Ok(())
1759 0 : }
1760 :
1761 : /// perform one garbage collection iteration, removing old data files from disk.
1762 : /// this function is periodically called by gc task.
1763 : /// also it can be explicitly requested through page server api 'do_gc' command.
1764 : ///
1765 : /// `target_timeline_id` specifies the timeline to GC, or None for all.
1766 : ///
1767 : /// The `horizon` an `pitr` parameters determine how much WAL history needs to be retained.
1768 : /// Also known as the retention period, or the GC cutoff point. `horizon` specifies
1769 : /// the amount of history, as LSN difference from current latest LSN on each timeline.
1770 : /// `pitr` specifies the same as a time difference from the current time. The effective
1771 : /// GC cutoff point is determined conservatively by either `horizon` and `pitr`, whichever
1772 : /// requires more history to be retained.
1773 : //
1774 2262 : pub(crate) async fn gc_iteration(
1775 2262 : &self,
1776 2262 : target_timeline_id: Option<TimelineId>,
1777 2262 : horizon: u64,
1778 2262 : pitr: Duration,
1779 2262 : cancel: &CancellationToken,
1780 2262 : ctx: &RequestContext,
1781 2262 : ) -> Result<GcResult, GcError> {
1782 2262 : // Don't start doing work during shutdown
1783 2262 : if let TenantState::Stopping { .. } = self.current_state() {
1784 0 : return Ok(GcResult::default());
1785 2262 : }
1786 2262 :
1787 2262 : // there is a global allowed_error for this
1788 2262 : if !self.is_active() {
1789 0 : return Err(GcError::NotActive);
1790 2262 : }
1791 2262 :
1792 2262 : {
1793 2262 : let conf = self.tenant_conf.load();
1794 2262 :
1795 2262 : if !conf.location.may_delete_layers_hint() {
1796 0 : info!("Skipping GC in location state {:?}", conf.location);
1797 0 : return Ok(GcResult::default());
1798 2262 : }
1799 : }
1800 :
1801 2262 : let _guard = match self.gc_block.start().await {
1802 2262 : Ok(guard) => guard,
1803 0 : Err(reasons) => {
1804 0 : info!("Skipping GC: {reasons}");
1805 0 : return Ok(GcResult::default());
1806 : }
1807 : };
1808 :
1809 2262 : self.gc_iteration_internal(target_timeline_id, horizon, pitr, cancel, ctx)
1810 2235 : .await
1811 2262 : }
1812 :
1813 : /// Perform one compaction iteration.
1814 : /// This function is periodically called by compactor task.
1815 : /// Also it can be explicitly requested per timeline through page server
1816 : /// api's 'compact' command.
1817 : ///
1818 : /// Returns whether we have pending compaction task.
1819 0 : async fn compaction_iteration(
1820 0 : &self,
1821 0 : cancel: &CancellationToken,
1822 0 : ctx: &RequestContext,
1823 0 : ) -> Result<bool, timeline::CompactionError> {
1824 0 : // Don't start doing work during shutdown, or when broken, we do not need those in the logs
1825 0 : if !self.is_active() {
1826 0 : return Ok(false);
1827 0 : }
1828 0 :
1829 0 : {
1830 0 : let conf = self.tenant_conf.load();
1831 0 : if !conf.location.may_delete_layers_hint() || !conf.location.may_upload_layers_hint() {
1832 0 : info!("Skipping compaction in location state {:?}", conf.location);
1833 0 : return Ok(false);
1834 0 : }
1835 0 : }
1836 0 :
1837 0 : // Scan through the hashmap and collect a list of all the timelines,
1838 0 : // while holding the lock. Then drop the lock and actually perform the
1839 0 : // compactions. We don't want to block everything else while the
1840 0 : // compaction runs.
1841 0 : let timelines_to_compact = {
1842 0 : let timelines = self.timelines.lock().unwrap();
1843 0 : let timelines_to_compact = timelines
1844 0 : .iter()
1845 0 : .filter_map(|(timeline_id, timeline)| {
1846 0 : if timeline.is_active() {
1847 0 : Some((*timeline_id, timeline.clone()))
1848 : } else {
1849 0 : None
1850 : }
1851 0 : })
1852 0 : .collect::<Vec<_>>();
1853 0 : drop(timelines);
1854 0 : timelines_to_compact
1855 0 : };
1856 0 :
1857 0 : // Before doing any I/O work, check our circuit breaker
1858 0 : if self.compaction_circuit_breaker.lock().unwrap().is_broken() {
1859 0 : info!("Skipping compaction due to previous failures");
1860 0 : return Ok(false);
1861 0 : }
1862 0 :
1863 0 : let mut has_pending_task = false;
1864 :
1865 0 : for (timeline_id, timeline) in &timelines_to_compact {
1866 0 : has_pending_task |= timeline
1867 0 : .compact(cancel, EnumSet::empty(), ctx)
1868 0 : .instrument(info_span!("compact_timeline", %timeline_id))
1869 0 : .await
1870 0 : .inspect_err(|e| match e {
1871 0 : timeline::CompactionError::ShuttingDown => (),
1872 0 : timeline::CompactionError::Other(e) => {
1873 0 : self.compaction_circuit_breaker
1874 0 : .lock()
1875 0 : .unwrap()
1876 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, e);
1877 0 : }
1878 0 : })?;
1879 : }
1880 :
1881 0 : self.compaction_circuit_breaker
1882 0 : .lock()
1883 0 : .unwrap()
1884 0 : .success(&CIRCUIT_BREAKERS_UNBROKEN);
1885 0 :
1886 0 : Ok(has_pending_task)
1887 0 : }
1888 :
1889 : // Call through to all timelines to freeze ephemeral layers if needed. Usually
1890 : // this happens during ingest: this background housekeeping is for freezing layers
1891 : // that are open but haven't been written to for some time.
1892 0 : async fn ingest_housekeeping(&self) {
1893 0 : // Scan through the hashmap and collect a list of all the timelines,
1894 0 : // while holding the lock. Then drop the lock and actually perform the
1895 0 : // compactions. We don't want to block everything else while the
1896 0 : // compaction runs.
1897 0 : let timelines = {
1898 0 : self.timelines
1899 0 : .lock()
1900 0 : .unwrap()
1901 0 : .values()
1902 0 : .filter_map(|timeline| {
1903 0 : if timeline.is_active() {
1904 0 : Some(timeline.clone())
1905 : } else {
1906 0 : None
1907 : }
1908 0 : })
1909 0 : .collect::<Vec<_>>()
1910 : };
1911 :
1912 0 : for timeline in &timelines {
1913 0 : timeline.maybe_freeze_ephemeral_layer().await;
1914 : }
1915 0 : }
1916 :
1917 7344 : pub fn current_state(&self) -> TenantState {
1918 7344 : self.state.borrow().clone()
1919 7344 : }
1920 :
1921 5064 : pub fn is_active(&self) -> bool {
1922 5064 : self.current_state() == TenantState::Active
1923 5064 : }
1924 :
1925 0 : pub fn generation(&self) -> Generation {
1926 0 : self.generation
1927 0 : }
1928 :
1929 0 : pub(crate) fn wal_redo_manager_status(&self) -> Option<WalRedoManagerStatus> {
1930 0 : self.walredo_mgr.as_ref().and_then(|mgr| mgr.status())
1931 0 : }
1932 :
1933 : /// Changes tenant status to active, unless shutdown was already requested.
1934 : ///
1935 : /// `background_jobs_can_start` is an optional barrier set to a value during pageserver startup
1936 : /// to delay background jobs. Background jobs can be started right away when None is given.
1937 0 : fn activate(
1938 0 : self: &Arc<Self>,
1939 0 : broker_client: BrokerClientChannel,
1940 0 : background_jobs_can_start: Option<&completion::Barrier>,
1941 0 : ctx: &RequestContext,
1942 0 : ) {
1943 0 : span::debug_assert_current_span_has_tenant_id();
1944 0 :
1945 0 : let mut activating = false;
1946 0 : self.state.send_modify(|current_state| {
1947 0 : use pageserver_api::models::ActivatingFrom;
1948 0 : match &*current_state {
1949 : TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
1950 0 : panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {state:?}", state = current_state);
1951 : }
1952 0 : TenantState::Loading => {
1953 0 : *current_state = TenantState::Activating(ActivatingFrom::Loading);
1954 0 : }
1955 0 : TenantState::Attaching => {
1956 0 : *current_state = TenantState::Activating(ActivatingFrom::Attaching);
1957 0 : }
1958 : }
1959 0 : debug!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), "Activating tenant");
1960 0 : activating = true;
1961 0 : // Continue outside the closure. We need to grab timelines.lock()
1962 0 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
1963 0 : });
1964 0 :
1965 0 : if activating {
1966 0 : let timelines_accessor = self.timelines.lock().unwrap();
1967 0 : let timelines_to_activate = timelines_accessor
1968 0 : .values()
1969 0 : .filter(|timeline| !(timeline.is_broken() || timeline.is_stopping()));
1970 0 :
1971 0 : // Before activation, populate each Timeline's GcInfo with information about its children
1972 0 : self.initialize_gc_info(&timelines_accessor);
1973 0 :
1974 0 : // Spawn gc and compaction loops. The loops will shut themselves
1975 0 : // down when they notice that the tenant is inactive.
1976 0 : tasks::start_background_loops(self, background_jobs_can_start);
1977 0 :
1978 0 : let mut activated_timelines = 0;
1979 :
1980 0 : for timeline in timelines_to_activate {
1981 0 : timeline.activate(
1982 0 : self.clone(),
1983 0 : broker_client.clone(),
1984 0 : background_jobs_can_start,
1985 0 : ctx,
1986 0 : );
1987 0 : activated_timelines += 1;
1988 0 : }
1989 :
1990 0 : self.state.send_modify(move |current_state| {
1991 0 : assert!(
1992 0 : matches!(current_state, TenantState::Activating(_)),
1993 0 : "set_stopping and set_broken wait for us to leave Activating state",
1994 : );
1995 0 : *current_state = TenantState::Active;
1996 0 :
1997 0 : let elapsed = self.constructed_at.elapsed();
1998 0 : let total_timelines = timelines_accessor.len();
1999 0 :
2000 0 : // log a lot of stuff, because some tenants sometimes suffer from user-visible
2001 0 : // times to activate. see https://github.com/neondatabase/neon/issues/4025
2002 0 : info!(
2003 0 : since_creation_millis = elapsed.as_millis(),
2004 0 : tenant_id = %self.tenant_shard_id.tenant_id,
2005 0 : shard_id = %self.tenant_shard_id.shard_slug(),
2006 0 : activated_timelines,
2007 0 : total_timelines,
2008 0 : post_state = <&'static str>::from(&*current_state),
2009 0 : "activation attempt finished"
2010 : );
2011 :
2012 0 : TENANT.activation.observe(elapsed.as_secs_f64());
2013 0 : });
2014 0 : }
2015 0 : }
2016 :
2017 : /// Shutdown the tenant and join all of the spawned tasks.
2018 : ///
2019 : /// The method caters for all use-cases:
2020 : /// - pageserver shutdown (freeze_and_flush == true)
2021 : /// - detach + ignore (freeze_and_flush == false)
2022 : ///
2023 : /// This will attempt to shutdown even if tenant is broken.
2024 : ///
2025 : /// `shutdown_progress` is a [`completion::Barrier`] for the shutdown initiated by this call.
2026 : /// If the tenant is already shutting down, we return a clone of the first shutdown call's
2027 : /// `Barrier` as an `Err`. This not-first caller can use the returned barrier to join with
2028 : /// the ongoing shutdown.
2029 18 : async fn shutdown(
2030 18 : &self,
2031 18 : shutdown_progress: completion::Barrier,
2032 18 : shutdown_mode: timeline::ShutdownMode,
2033 18 : ) -> Result<(), completion::Barrier> {
2034 18 : span::debug_assert_current_span_has_tenant_id();
2035 :
2036 : // Set tenant (and its timlines) to Stoppping state.
2037 : //
2038 : // Since we can only transition into Stopping state after activation is complete,
2039 : // run it in a JoinSet so all tenants have a chance to stop before we get SIGKILLed.
2040 : //
2041 : // Transitioning tenants to Stopping state has a couple of non-obvious side effects:
2042 : // 1. Lock out any new requests to the tenants.
2043 : // 2. Signal cancellation to WAL receivers (we wait on it below).
2044 : // 3. Signal cancellation for other tenant background loops.
2045 : // 4. ???
2046 : //
2047 : // The waiting for the cancellation is not done uniformly.
2048 : // We certainly wait for WAL receivers to shut down.
2049 : // That is necessary so that no new data comes in before the freeze_and_flush.
2050 : // But the tenant background loops are joined-on in our caller.
2051 : // It's mesed up.
2052 : // we just ignore the failure to stop
2053 :
2054 : // If we're still attaching, fire the cancellation token early to drop out: this
2055 : // will prevent us flushing, but ensures timely shutdown if some I/O during attach
2056 : // is very slow.
2057 18 : let shutdown_mode = if matches!(self.current_state(), TenantState::Attaching) {
2058 0 : self.cancel.cancel();
2059 0 :
2060 0 : // Having fired our cancellation token, do not try and flush timelines: their cancellation tokens
2061 0 : // are children of ours, so their flush loops will have shut down already
2062 0 : timeline::ShutdownMode::Hard
2063 : } else {
2064 18 : shutdown_mode
2065 : };
2066 :
2067 18 : match self.set_stopping(shutdown_progress, false, false).await {
2068 18 : Ok(()) => {}
2069 0 : Err(SetStoppingError::Broken) => {
2070 0 : // assume that this is acceptable
2071 0 : }
2072 0 : Err(SetStoppingError::AlreadyStopping(other)) => {
2073 0 : // give caller the option to wait for this this shutdown
2074 0 : info!("Tenant::shutdown: AlreadyStopping");
2075 0 : return Err(other);
2076 : }
2077 : };
2078 :
2079 18 : let mut js = tokio::task::JoinSet::new();
2080 18 : {
2081 18 : let timelines = self.timelines.lock().unwrap();
2082 18 : timelines.values().for_each(|timeline| {
2083 18 : let timeline = Arc::clone(timeline);
2084 18 : let timeline_id = timeline.timeline_id;
2085 18 : let span = tracing::info_span!("timeline_shutdown", %timeline_id, ?shutdown_mode);
2086 36 : js.spawn(async move { timeline.shutdown(shutdown_mode).instrument(span).await });
2087 18 : })
2088 18 : };
2089 18 : // test_long_timeline_create_then_tenant_delete is leaning on this message
2090 18 : tracing::info!("Waiting for timelines...");
2091 36 : while let Some(res) = js.join_next().await {
2092 0 : match res {
2093 18 : Ok(()) => {}
2094 0 : Err(je) if je.is_cancelled() => unreachable!("no cancelling used"),
2095 0 : Err(je) if je.is_panic() => { /* logged already */ }
2096 0 : Err(je) => warn!("unexpected JoinError: {je:?}"),
2097 : }
2098 : }
2099 :
2100 : // We cancel the Tenant's cancellation token _after_ the timelines have all shut down. This permits
2101 : // them to continue to do work during their shutdown methods, e.g. flushing data.
2102 18 : tracing::debug!("Cancelling CancellationToken");
2103 18 : self.cancel.cancel();
2104 18 :
2105 18 : // shutdown all tenant and timeline tasks: gc, compaction, page service
2106 18 : // No new tasks will be started for this tenant because it's in `Stopping` state.
2107 18 : //
2108 18 : // this will additionally shutdown and await all timeline tasks.
2109 18 : tracing::debug!("Waiting for tasks...");
2110 18 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), None).await;
2111 :
2112 18 : if let Some(walredo_mgr) = self.walredo_mgr.as_ref() {
2113 18 : walredo_mgr.shutdown().await;
2114 0 : }
2115 :
2116 : // Wait for any in-flight operations to complete
2117 18 : self.gate.close().await;
2118 :
2119 18 : remove_tenant_metrics(&self.tenant_shard_id);
2120 18 :
2121 18 : Ok(())
2122 18 : }
2123 :
2124 : /// Change tenant status to Stopping, to mark that it is being shut down.
2125 : ///
2126 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
2127 : ///
2128 : /// This function is not cancel-safe!
2129 : ///
2130 : /// `allow_transition_from_loading` is needed for the special case of loading task deleting the tenant.
2131 : /// `allow_transition_from_attaching` is needed for the special case of attaching deleted tenant.
2132 18 : async fn set_stopping(
2133 18 : &self,
2134 18 : progress: completion::Barrier,
2135 18 : allow_transition_from_loading: bool,
2136 18 : allow_transition_from_attaching: bool,
2137 18 : ) -> Result<(), SetStoppingError> {
2138 18 : let mut rx = self.state.subscribe();
2139 18 :
2140 18 : // cannot stop before we're done activating, so wait out until we're done activating
2141 18 : rx.wait_for(|state| match state {
2142 0 : TenantState::Attaching if allow_transition_from_attaching => true,
2143 : TenantState::Activating(_) | TenantState::Attaching => {
2144 0 : info!(
2145 0 : "waiting for {} to turn Active|Broken|Stopping",
2146 0 : <&'static str>::from(state)
2147 : );
2148 0 : false
2149 : }
2150 0 : TenantState::Loading => allow_transition_from_loading,
2151 18 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
2152 18 : })
2153 0 : .await
2154 18 : .expect("cannot drop self.state while on a &self method");
2155 18 :
2156 18 : // we now know we're done activating, let's see whether this task is the winner to transition into Stopping
2157 18 : let mut err = None;
2158 18 : let stopping = self.state.send_if_modified(|current_state| match current_state {
2159 : TenantState::Activating(_) => {
2160 0 : unreachable!("1we ensured above that we're done with activation, and, there is no re-activation")
2161 : }
2162 : TenantState::Attaching => {
2163 0 : if !allow_transition_from_attaching {
2164 0 : unreachable!("2we ensured above that we're done with activation, and, there is no re-activation")
2165 0 : };
2166 0 : *current_state = TenantState::Stopping { progress };
2167 0 : true
2168 : }
2169 : TenantState::Loading => {
2170 0 : if !allow_transition_from_loading {
2171 0 : unreachable!("3we ensured above that we're done with activation, and, there is no re-activation")
2172 0 : };
2173 0 : *current_state = TenantState::Stopping { progress };
2174 0 : true
2175 : }
2176 : TenantState::Active => {
2177 : // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
2178 : // are created after the transition to Stopping. That's harmless, as the Timelines
2179 : // won't be accessible to anyone afterwards, because the Tenant is in Stopping state.
2180 18 : *current_state = TenantState::Stopping { progress };
2181 18 : // Continue stopping outside the closure. We need to grab timelines.lock()
2182 18 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
2183 18 : true
2184 : }
2185 0 : TenantState::Broken { reason, .. } => {
2186 0 : info!(
2187 0 : "Cannot set tenant to Stopping state, it is in Broken state due to: {reason}"
2188 : );
2189 0 : err = Some(SetStoppingError::Broken);
2190 0 : false
2191 : }
2192 0 : TenantState::Stopping { progress } => {
2193 0 : info!("Tenant is already in Stopping state");
2194 0 : err = Some(SetStoppingError::AlreadyStopping(progress.clone()));
2195 0 : false
2196 : }
2197 18 : });
2198 18 : match (stopping, err) {
2199 18 : (true, None) => {} // continue
2200 0 : (false, Some(err)) => return Err(err),
2201 0 : (true, Some(_)) => unreachable!(
2202 0 : "send_if_modified closure must error out if not transitioning to Stopping"
2203 0 : ),
2204 0 : (false, None) => unreachable!(
2205 0 : "send_if_modified closure must return true if transitioning to Stopping"
2206 0 : ),
2207 : }
2208 :
2209 18 : let timelines_accessor = self.timelines.lock().unwrap();
2210 18 : let not_broken_timelines = timelines_accessor
2211 18 : .values()
2212 18 : .filter(|timeline| !timeline.is_broken());
2213 36 : for timeline in not_broken_timelines {
2214 18 : timeline.set_state(TimelineState::Stopping);
2215 18 : }
2216 18 : Ok(())
2217 18 : }
2218 :
2219 : /// Method for tenant::mgr to transition us into Broken state in case of a late failure in
2220 : /// `remove_tenant_from_memory`
2221 : ///
2222 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
2223 : ///
2224 : /// In tests, we also use this to set tenants to Broken state on purpose.
2225 0 : pub(crate) async fn set_broken(&self, reason: String) {
2226 0 : let mut rx = self.state.subscribe();
2227 0 :
2228 0 : // The load & attach routines own the tenant state until it has reached `Active`.
2229 0 : // So, wait until it's done.
2230 0 : rx.wait_for(|state| match state {
2231 : TenantState::Activating(_) | TenantState::Loading | TenantState::Attaching => {
2232 0 : info!(
2233 0 : "waiting for {} to turn Active|Broken|Stopping",
2234 0 : <&'static str>::from(state)
2235 : );
2236 0 : false
2237 : }
2238 0 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
2239 0 : })
2240 0 : .await
2241 0 : .expect("cannot drop self.state while on a &self method");
2242 0 :
2243 0 : // we now know we're done activating, let's see whether this task is the winner to transition into Broken
2244 0 : self.set_broken_no_wait(reason)
2245 0 : }
2246 :
2247 0 : pub(crate) fn set_broken_no_wait(&self, reason: impl Display) {
2248 0 : let reason = reason.to_string();
2249 0 : self.state.send_modify(|current_state| {
2250 0 : match *current_state {
2251 : TenantState::Activating(_) | TenantState::Loading | TenantState::Attaching => {
2252 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
2253 : }
2254 : TenantState::Active => {
2255 0 : if cfg!(feature = "testing") {
2256 0 : warn!("Changing Active tenant to Broken state, reason: {}", reason);
2257 0 : *current_state = TenantState::broken_from_reason(reason);
2258 : } else {
2259 0 : unreachable!("not allowed to call set_broken on Active tenants in non-testing builds")
2260 : }
2261 : }
2262 : TenantState::Broken { .. } => {
2263 0 : warn!("Tenant is already in Broken state");
2264 : }
2265 : // This is the only "expected" path, any other path is a bug.
2266 : TenantState::Stopping { .. } => {
2267 0 : warn!(
2268 0 : "Marking Stopping tenant as Broken state, reason: {}",
2269 : reason
2270 : );
2271 0 : *current_state = TenantState::broken_from_reason(reason);
2272 : }
2273 : }
2274 0 : });
2275 0 : }
2276 :
2277 0 : pub fn subscribe_for_state_updates(&self) -> watch::Receiver<TenantState> {
2278 0 : self.state.subscribe()
2279 0 : }
2280 :
2281 : /// The activate_now semaphore is initialized with zero units. As soon as
2282 : /// we add a unit, waiters will be able to acquire a unit and proceed.
2283 0 : pub(crate) fn activate_now(&self) {
2284 0 : self.activate_now_sem.add_permits(1);
2285 0 : }
2286 :
2287 0 : pub(crate) async fn wait_to_become_active(
2288 0 : &self,
2289 0 : timeout: Duration,
2290 0 : ) -> Result<(), GetActiveTenantError> {
2291 0 : let mut receiver = self.state.subscribe();
2292 0 : loop {
2293 0 : let current_state = receiver.borrow_and_update().clone();
2294 0 : match current_state {
2295 : TenantState::Loading | TenantState::Attaching | TenantState::Activating(_) => {
2296 : // in these states, there's a chance that we can reach ::Active
2297 0 : self.activate_now();
2298 0 : match timeout_cancellable(timeout, &self.cancel, receiver.changed()).await {
2299 0 : Ok(r) => {
2300 0 : r.map_err(
2301 0 : |_e: tokio::sync::watch::error::RecvError|
2302 : // Tenant existed but was dropped: report it as non-existent
2303 0 : GetActiveTenantError::NotFound(GetTenantError::NotFound(self.tenant_shard_id.tenant_id))
2304 0 : )?
2305 : }
2306 : Err(TimeoutCancellableError::Cancelled) => {
2307 0 : return Err(GetActiveTenantError::Cancelled);
2308 : }
2309 : Err(TimeoutCancellableError::Timeout) => {
2310 0 : return Err(GetActiveTenantError::WaitForActiveTimeout {
2311 0 : latest_state: Some(self.current_state()),
2312 0 : wait_time: timeout,
2313 0 : });
2314 : }
2315 : }
2316 : }
2317 : TenantState::Active { .. } => {
2318 0 : return Ok(());
2319 : }
2320 0 : TenantState::Broken { reason, .. } => {
2321 0 : // This is fatal, and reported distinctly from the general case of "will never be active" because
2322 0 : // it's logically a 500 to external API users (broken is always a bug).
2323 0 : return Err(GetActiveTenantError::Broken(reason));
2324 : }
2325 : TenantState::Stopping { .. } => {
2326 : // There's no chance the tenant can transition back into ::Active
2327 0 : return Err(GetActiveTenantError::WillNotBecomeActive(current_state));
2328 : }
2329 : }
2330 : }
2331 0 : }
2332 :
2333 0 : pub(crate) fn get_attach_mode(&self) -> AttachmentMode {
2334 0 : self.tenant_conf.load().location.attach_mode
2335 0 : }
2336 :
2337 : /// For API access: generate a LocationConfig equivalent to the one that would be used to
2338 : /// create a Tenant in the same state. Do not use this in hot paths: it's for relatively
2339 : /// rare external API calls, like a reconciliation at startup.
2340 0 : pub(crate) fn get_location_conf(&self) -> models::LocationConfig {
2341 0 : let conf = self.tenant_conf.load();
2342 :
2343 0 : let location_config_mode = match conf.location.attach_mode {
2344 0 : AttachmentMode::Single => models::LocationConfigMode::AttachedSingle,
2345 0 : AttachmentMode::Multi => models::LocationConfigMode::AttachedMulti,
2346 0 : AttachmentMode::Stale => models::LocationConfigMode::AttachedStale,
2347 : };
2348 :
2349 : // We have a pageserver TenantConf, we need the API-facing TenantConfig.
2350 0 : let tenant_config: models::TenantConfig = conf.tenant_conf.clone().into();
2351 0 :
2352 0 : models::LocationConfig {
2353 0 : mode: location_config_mode,
2354 0 : generation: self.generation.into(),
2355 0 : secondary_conf: None,
2356 0 : shard_number: self.shard_identity.number.0,
2357 0 : shard_count: self.shard_identity.count.literal(),
2358 0 : shard_stripe_size: self.shard_identity.stripe_size.0,
2359 0 : tenant_conf: tenant_config,
2360 0 : }
2361 0 : }
2362 :
2363 0 : pub(crate) fn get_tenant_shard_id(&self) -> &TenantShardId {
2364 0 : &self.tenant_shard_id
2365 0 : }
2366 :
2367 0 : pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
2368 0 : self.shard_identity.stripe_size
2369 0 : }
2370 :
2371 0 : pub(crate) fn get_generation(&self) -> Generation {
2372 0 : self.generation
2373 0 : }
2374 :
2375 : /// This function partially shuts down the tenant (it shuts down the Timelines) and is fallible,
2376 : /// and can leave the tenant in a bad state if it fails. The caller is responsible for
2377 : /// resetting this tenant to a valid state if we fail.
2378 0 : pub(crate) async fn split_prepare(
2379 0 : &self,
2380 0 : child_shards: &Vec<TenantShardId>,
2381 0 : ) -> anyhow::Result<()> {
2382 0 : let timelines = self.timelines.lock().unwrap().clone();
2383 0 : for timeline in timelines.values() {
2384 : // We do not block timeline creation/deletion during splits inside the pageserver: it is up to higher levels
2385 : // to ensure that they do not start a split if currently in the process of doing these.
2386 :
2387 : // Upload an index from the parent: this is partly to provide freshness for the
2388 : // child tenants that will copy it, and partly for general ease-of-debugging: there will
2389 : // always be a parent shard index in the same generation as we wrote the child shard index.
2390 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Uploading index");
2391 0 : timeline
2392 0 : .remote_client
2393 0 : .schedule_index_upload_for_file_changes()?;
2394 0 : timeline.remote_client.wait_completion().await?;
2395 :
2396 : // Shut down the timeline's remote client: this means that the indices we write
2397 : // for child shards will not be invalidated by the parent shard deleting layers.
2398 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Shutting down remote storage client");
2399 0 : timeline.remote_client.shutdown().await;
2400 :
2401 : // Download methods can still be used after shutdown, as they don't flow through the remote client's
2402 : // queue. In principal the RemoteTimelineClient could provide this without downloading it, but this
2403 : // operation is rare, so it's simpler to just download it (and robustly guarantees that the index
2404 : // we use here really is the remotely persistent one).
2405 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Downloading index_part from parent");
2406 0 : let result = timeline.remote_client
2407 0 : .download_index_file(&self.cancel)
2408 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))
2409 0 : .await?;
2410 0 : let index_part = match result {
2411 : MaybeDeletedIndexPart::Deleted(_) => {
2412 0 : anyhow::bail!("Timeline deletion happened concurrently with split")
2413 : }
2414 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
2415 : };
2416 :
2417 0 : for child_shard in child_shards {
2418 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Uploading index_part for child {}", child_shard.to_index());
2419 0 : upload_index_part(
2420 0 : &self.remote_storage,
2421 0 : child_shard,
2422 0 : &timeline.timeline_id,
2423 0 : self.generation,
2424 0 : &index_part,
2425 0 : &self.cancel,
2426 0 : )
2427 0 : .await?;
2428 : }
2429 : }
2430 :
2431 0 : Ok(())
2432 0 : }
2433 :
2434 0 : pub(crate) fn get_sizes(&self) -> TopTenantShardItem {
2435 0 : let mut result = TopTenantShardItem {
2436 0 : id: self.tenant_shard_id,
2437 0 : resident_size: 0,
2438 0 : physical_size: 0,
2439 0 : max_logical_size: 0,
2440 0 : };
2441 :
2442 0 : for timeline in self.timelines.lock().unwrap().values() {
2443 0 : result.resident_size += timeline.metrics.resident_physical_size_gauge.get();
2444 0 :
2445 0 : result.physical_size += timeline
2446 0 : .remote_client
2447 0 : .metrics
2448 0 : .remote_physical_size_gauge
2449 0 : .get();
2450 0 : result.max_logical_size = std::cmp::max(
2451 0 : result.max_logical_size,
2452 0 : timeline.metrics.current_logical_size_gauge.get(),
2453 0 : );
2454 0 : }
2455 :
2456 0 : result
2457 0 : }
2458 : }
2459 :
2460 : /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
2461 : /// perform a topological sort, so that the parent of each timeline comes
2462 : /// before the children.
2463 : /// E extracts the ancestor from T
2464 : /// This allows for T to be different. It can be TimelineMetadata, can be Timeline itself, etc.
2465 564 : fn tree_sort_timelines<T, E>(
2466 564 : timelines: HashMap<TimelineId, T>,
2467 564 : extractor: E,
2468 564 : ) -> anyhow::Result<Vec<(TimelineId, T)>>
2469 564 : where
2470 564 : E: Fn(&T) -> Option<TimelineId>,
2471 564 : {
2472 564 : let mut result = Vec::with_capacity(timelines.len());
2473 564 :
2474 564 : let mut now = Vec::with_capacity(timelines.len());
2475 564 : // (ancestor, children)
2476 564 : let mut later: HashMap<TimelineId, Vec<(TimelineId, T)>> =
2477 564 : HashMap::with_capacity(timelines.len());
2478 :
2479 582 : for (timeline_id, value) in timelines {
2480 18 : if let Some(ancestor_id) = extractor(&value) {
2481 6 : let children = later.entry(ancestor_id).or_default();
2482 6 : children.push((timeline_id, value));
2483 12 : } else {
2484 12 : now.push((timeline_id, value));
2485 12 : }
2486 : }
2487 :
2488 582 : while let Some((timeline_id, metadata)) = now.pop() {
2489 18 : result.push((timeline_id, metadata));
2490 : // All children of this can be loaded now
2491 18 : if let Some(mut children) = later.remove(&timeline_id) {
2492 6 : now.append(&mut children);
2493 12 : }
2494 : }
2495 :
2496 : // All timelines should be visited now. Unless there were timelines with missing ancestors.
2497 564 : if !later.is_empty() {
2498 0 : for (missing_id, orphan_ids) in later {
2499 0 : for (orphan_id, _) in orphan_ids {
2500 0 : error!("could not load timeline {orphan_id} because its ancestor timeline {missing_id} could not be loaded");
2501 : }
2502 : }
2503 0 : bail!("could not load tenant because some timelines are missing ancestors");
2504 564 : }
2505 564 :
2506 564 : Ok(result)
2507 564 : }
2508 :
2509 : impl Tenant {
2510 0 : pub fn tenant_specific_overrides(&self) -> TenantConfOpt {
2511 0 : self.tenant_conf.load().tenant_conf.clone()
2512 0 : }
2513 :
2514 0 : pub fn effective_config(&self) -> TenantConf {
2515 0 : self.tenant_specific_overrides()
2516 0 : .merge(self.conf.default_tenant_conf.clone())
2517 0 : }
2518 :
2519 0 : pub fn get_checkpoint_distance(&self) -> u64 {
2520 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2521 0 : tenant_conf
2522 0 : .checkpoint_distance
2523 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
2524 0 : }
2525 :
2526 0 : pub fn get_checkpoint_timeout(&self) -> Duration {
2527 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2528 0 : tenant_conf
2529 0 : .checkpoint_timeout
2530 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
2531 0 : }
2532 :
2533 0 : pub fn get_compaction_target_size(&self) -> u64 {
2534 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2535 0 : tenant_conf
2536 0 : .compaction_target_size
2537 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
2538 0 : }
2539 :
2540 0 : pub fn get_compaction_period(&self) -> Duration {
2541 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2542 0 : tenant_conf
2543 0 : .compaction_period
2544 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
2545 0 : }
2546 :
2547 0 : pub fn get_compaction_threshold(&self) -> usize {
2548 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2549 0 : tenant_conf
2550 0 : .compaction_threshold
2551 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
2552 0 : }
2553 :
2554 0 : pub fn get_gc_horizon(&self) -> u64 {
2555 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2556 0 : tenant_conf
2557 0 : .gc_horizon
2558 0 : .unwrap_or(self.conf.default_tenant_conf.gc_horizon)
2559 0 : }
2560 :
2561 0 : pub fn get_gc_period(&self) -> Duration {
2562 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2563 0 : tenant_conf
2564 0 : .gc_period
2565 0 : .unwrap_or(self.conf.default_tenant_conf.gc_period)
2566 0 : }
2567 :
2568 0 : pub fn get_image_creation_threshold(&self) -> usize {
2569 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2570 0 : tenant_conf
2571 0 : .image_creation_threshold
2572 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
2573 0 : }
2574 :
2575 0 : pub fn get_pitr_interval(&self) -> Duration {
2576 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2577 0 : tenant_conf
2578 0 : .pitr_interval
2579 0 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
2580 0 : }
2581 :
2582 0 : pub fn get_min_resident_size_override(&self) -> Option<u64> {
2583 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2584 0 : tenant_conf
2585 0 : .min_resident_size_override
2586 0 : .or(self.conf.default_tenant_conf.min_resident_size_override)
2587 0 : }
2588 :
2589 0 : pub fn get_heatmap_period(&self) -> Option<Duration> {
2590 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2591 0 : let heatmap_period = tenant_conf
2592 0 : .heatmap_period
2593 0 : .unwrap_or(self.conf.default_tenant_conf.heatmap_period);
2594 0 : if heatmap_period.is_zero() {
2595 0 : None
2596 : } else {
2597 0 : Some(heatmap_period)
2598 : }
2599 0 : }
2600 :
2601 0 : pub fn get_lsn_lease_length(&self) -> Duration {
2602 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2603 0 : tenant_conf
2604 0 : .lsn_lease_length
2605 0 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
2606 0 : }
2607 :
2608 0 : pub fn set_new_tenant_config(&self, new_tenant_conf: TenantConfOpt) {
2609 0 : // Use read-copy-update in order to avoid overwriting the location config
2610 0 : // state if this races with [`Tenant::set_new_location_config`]. Note that
2611 0 : // this race is not possible if both request types come from the storage
2612 0 : // controller (as they should!) because an exclusive op lock is required
2613 0 : // on the storage controller side.
2614 0 : self.tenant_conf.rcu(|inner| {
2615 0 : Arc::new(AttachedTenantConf {
2616 0 : tenant_conf: new_tenant_conf.clone(),
2617 0 : location: inner.location,
2618 0 : })
2619 0 : });
2620 0 :
2621 0 : self.tenant_conf_updated(&new_tenant_conf);
2622 0 : // Don't hold self.timelines.lock() during the notifies.
2623 0 : // There's no risk of deadlock right now, but there could be if we consolidate
2624 0 : // mutexes in struct Timeline in the future.
2625 0 : let timelines = self.list_timelines();
2626 0 : for timeline in timelines {
2627 0 : timeline.tenant_conf_updated(&new_tenant_conf);
2628 0 : }
2629 0 : }
2630 :
2631 24 : pub(crate) fn set_new_location_config(&self, new_conf: AttachedTenantConf) {
2632 24 : let new_tenant_conf = new_conf.tenant_conf.clone();
2633 24 :
2634 24 : self.tenant_conf.store(Arc::new(new_conf));
2635 24 :
2636 24 : self.tenant_conf_updated(&new_tenant_conf);
2637 24 : // Don't hold self.timelines.lock() during the notifies.
2638 24 : // There's no risk of deadlock right now, but there could be if we consolidate
2639 24 : // mutexes in struct Timeline in the future.
2640 24 : let timelines = self.list_timelines();
2641 48 : for timeline in timelines {
2642 24 : timeline.tenant_conf_updated(&new_tenant_conf);
2643 24 : }
2644 24 : }
2645 :
2646 588 : fn get_timeline_get_throttle_config(
2647 588 : psconf: &'static PageServerConf,
2648 588 : overrides: &TenantConfOpt,
2649 588 : ) -> throttle::Config {
2650 588 : overrides
2651 588 : .timeline_get_throttle
2652 588 : .clone()
2653 588 : .unwrap_or(psconf.default_tenant_conf.timeline_get_throttle.clone())
2654 588 : }
2655 :
2656 24 : pub(crate) fn tenant_conf_updated(&self, new_conf: &TenantConfOpt) {
2657 24 : let conf = Self::get_timeline_get_throttle_config(self.conf, new_conf);
2658 24 : self.timeline_get_throttle.reconfigure(conf)
2659 24 : }
2660 :
2661 : /// Helper function to create a new Timeline struct.
2662 : ///
2663 : /// The returned Timeline is in Loading state. The caller is responsible for
2664 : /// initializing any on-disk state, and for inserting the Timeline to the 'timelines'
2665 : /// map.
2666 : ///
2667 : /// `validate_ancestor == false` is used when a timeline is created for deletion
2668 : /// and we might not have the ancestor present anymore which is fine for to be
2669 : /// deleted timelines.
2670 1236 : fn create_timeline_struct(
2671 1236 : &self,
2672 1236 : new_timeline_id: TimelineId,
2673 1236 : new_metadata: &TimelineMetadata,
2674 1236 : ancestor: Option<Arc<Timeline>>,
2675 1236 : resources: TimelineResources,
2676 1236 : cause: CreateTimelineCause,
2677 1236 : last_aux_file_policy: Option<AuxFilePolicy>,
2678 1236 : ) -> anyhow::Result<Arc<Timeline>> {
2679 1236 : let state = match cause {
2680 : CreateTimelineCause::Load => {
2681 1236 : let ancestor_id = new_metadata.ancestor_timeline();
2682 1236 : anyhow::ensure!(
2683 1236 : ancestor_id == ancestor.as_ref().map(|t| t.timeline_id),
2684 0 : "Timeline's {new_timeline_id} ancestor {ancestor_id:?} was not found"
2685 : );
2686 1236 : TimelineState::Loading
2687 : }
2688 0 : CreateTimelineCause::Delete => TimelineState::Stopping,
2689 : };
2690 :
2691 1236 : let pg_version = new_metadata.pg_version();
2692 1236 :
2693 1236 : let timeline = Timeline::new(
2694 1236 : self.conf,
2695 1236 : Arc::clone(&self.tenant_conf),
2696 1236 : new_metadata,
2697 1236 : ancestor,
2698 1236 : new_timeline_id,
2699 1236 : self.tenant_shard_id,
2700 1236 : self.generation,
2701 1236 : self.shard_identity,
2702 1236 : self.walredo_mgr.clone(),
2703 1236 : resources,
2704 1236 : pg_version,
2705 1236 : state,
2706 1236 : last_aux_file_policy,
2707 1236 : self.cancel.child_token(),
2708 1236 : );
2709 1236 :
2710 1236 : Ok(timeline)
2711 1236 : }
2712 :
2713 : // Allow too_many_arguments because a constructor's argument list naturally grows with the
2714 : // number of attributes in the struct: breaking these out into a builder wouldn't be helpful.
2715 : #[allow(clippy::too_many_arguments)]
2716 564 : fn new(
2717 564 : state: TenantState,
2718 564 : conf: &'static PageServerConf,
2719 564 : attached_conf: AttachedTenantConf,
2720 564 : shard_identity: ShardIdentity,
2721 564 : walredo_mgr: Option<Arc<WalRedoManager>>,
2722 564 : tenant_shard_id: TenantShardId,
2723 564 : remote_storage: GenericRemoteStorage,
2724 564 : deletion_queue_client: DeletionQueueClient,
2725 564 : l0_flush_global_state: L0FlushGlobalState,
2726 564 : ) -> Tenant {
2727 564 : debug_assert!(
2728 564 : !attached_conf.location.generation.is_none() || conf.control_plane_api.is_none()
2729 : );
2730 :
2731 564 : let (state, mut rx) = watch::channel(state);
2732 564 :
2733 564 : tokio::spawn(async move {
2734 564 : // reflect tenant state in metrics:
2735 564 : // - global per tenant state: TENANT_STATE_METRIC
2736 564 : // - "set" of broken tenants: BROKEN_TENANTS_SET
2737 564 : //
2738 564 : // set of broken tenants should not have zero counts so that it remains accessible for
2739 564 : // alerting.
2740 564 :
2741 564 : let tid = tenant_shard_id.to_string();
2742 564 : let shard_id = tenant_shard_id.shard_slug().to_string();
2743 564 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
2744 564 :
2745 1122 : fn inspect_state(state: &TenantState) -> ([&'static str; 1], bool) {
2746 1122 : ([state.into()], matches!(state, TenantState::Broken { .. }))
2747 1122 : }
2748 564 :
2749 564 : let mut tuple = inspect_state(&rx.borrow_and_update());
2750 564 :
2751 564 : let is_broken = tuple.1;
2752 564 : let mut counted_broken = if is_broken {
2753 : // add the id to the set right away, there should not be any updates on the channel
2754 : // after before tenant is removed, if ever
2755 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
2756 0 : true
2757 : } else {
2758 564 : false
2759 : };
2760 :
2761 1122 : loop {
2762 1122 : let labels = &tuple.0;
2763 1122 : let current = TENANT_STATE_METRIC.with_label_values(labels);
2764 1122 : current.inc();
2765 1122 :
2766 1122 : if rx.changed().await.is_err() {
2767 : // tenant has been dropped
2768 48 : current.dec();
2769 48 : drop(BROKEN_TENANTS_SET.remove_label_values(set_key));
2770 48 : break;
2771 558 : }
2772 558 :
2773 558 : current.dec();
2774 558 : tuple = inspect_state(&rx.borrow_and_update());
2775 558 :
2776 558 : let is_broken = tuple.1;
2777 558 : if is_broken && !counted_broken {
2778 0 : counted_broken = true;
2779 0 : // insert the tenant_id (back) into the set while avoiding needless counter
2780 0 : // access
2781 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
2782 558 : }
2783 : }
2784 564 : });
2785 564 :
2786 564 : Tenant {
2787 564 : tenant_shard_id,
2788 564 : shard_identity,
2789 564 : generation: attached_conf.location.generation,
2790 564 : conf,
2791 564 : // using now here is good enough approximation to catch tenants with really long
2792 564 : // activation times.
2793 564 : constructed_at: Instant::now(),
2794 564 : timelines: Mutex::new(HashMap::new()),
2795 564 : timelines_creating: Mutex::new(HashSet::new()),
2796 564 : gc_cs: tokio::sync::Mutex::new(()),
2797 564 : walredo_mgr,
2798 564 : remote_storage,
2799 564 : deletion_queue_client,
2800 564 : state,
2801 564 : cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
2802 564 : cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
2803 564 : eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
2804 564 : compaction_circuit_breaker: std::sync::Mutex::new(CircuitBreaker::new(
2805 564 : format!("compaction-{tenant_shard_id}"),
2806 564 : 5,
2807 564 : // Compaction can be a very expensive operation, and might leak disk space. It also ought
2808 564 : // to be infallible, as long as remote storage is available. So if it repeatedly fails,
2809 564 : // use an extremely long backoff.
2810 564 : Some(Duration::from_secs(3600 * 24)),
2811 564 : )),
2812 564 : activate_now_sem: tokio::sync::Semaphore::new(0),
2813 564 : cancel: CancellationToken::default(),
2814 564 : gate: Gate::default(),
2815 564 : timeline_get_throttle: Arc::new(throttle::Throttle::new(
2816 564 : Tenant::get_timeline_get_throttle_config(conf, &attached_conf.tenant_conf),
2817 564 : &crate::metrics::tenant_throttling::TIMELINE_GET,
2818 564 : )),
2819 564 : tenant_conf: Arc::new(ArcSwap::from_pointee(attached_conf)),
2820 564 : ongoing_timeline_detach: std::sync::Mutex::default(),
2821 564 : gc_block: Default::default(),
2822 564 : l0_flush_global_state,
2823 564 : }
2824 564 : }
2825 :
2826 : /// Locate and load config
2827 0 : pub(super) fn load_tenant_config(
2828 0 : conf: &'static PageServerConf,
2829 0 : tenant_shard_id: &TenantShardId,
2830 0 : ) -> Result<LocationConf, LoadConfigError> {
2831 0 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
2832 0 :
2833 0 : info!("loading tenant configuration from {config_path}");
2834 :
2835 : // load and parse file
2836 0 : let config = fs::read_to_string(&config_path).map_err(|e| {
2837 0 : match e.kind() {
2838 : std::io::ErrorKind::NotFound => {
2839 : // The config should almost always exist for a tenant directory:
2840 : // - When attaching a tenant, the config is the first thing we write
2841 : // - When detaching a tenant, we atomically move the directory to a tmp location
2842 : // before deleting contents.
2843 : //
2844 : // The very rare edge case that can result in a missing config is if we crash during attach
2845 : // between creating directory and writing config. Callers should handle that as if the
2846 : // directory didn't exist.
2847 :
2848 0 : LoadConfigError::NotFound(config_path)
2849 : }
2850 : _ => {
2851 : // No IO errors except NotFound are acceptable here: other kinds of error indicate local storage or permissions issues
2852 : // that we cannot cleanly recover
2853 0 : crate::virtual_file::on_fatal_io_error(&e, "Reading tenant config file")
2854 : }
2855 : }
2856 0 : })?;
2857 :
2858 0 : Ok(toml_edit::de::from_str::<LocationConf>(&config)?)
2859 0 : }
2860 :
2861 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
2862 : pub(super) async fn persist_tenant_config(
2863 : conf: &'static PageServerConf,
2864 : tenant_shard_id: &TenantShardId,
2865 : location_conf: &LocationConf,
2866 : ) -> std::io::Result<()> {
2867 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
2868 :
2869 : Self::persist_tenant_config_at(tenant_shard_id, &config_path, location_conf).await
2870 : }
2871 :
2872 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
2873 : pub(super) async fn persist_tenant_config_at(
2874 : tenant_shard_id: &TenantShardId,
2875 : config_path: &Utf8Path,
2876 : location_conf: &LocationConf,
2877 : ) -> std::io::Result<()> {
2878 : debug!("persisting tenantconf to {config_path}");
2879 :
2880 : let mut conf_content = r#"# This file contains a specific per-tenant's config.
2881 : # It is read in case of pageserver restart.
2882 : "#
2883 : .to_string();
2884 :
2885 0 : fail::fail_point!("tenant-config-before-write", |_| {
2886 0 : Err(std::io::Error::new(
2887 0 : std::io::ErrorKind::Other,
2888 0 : "tenant-config-before-write",
2889 0 : ))
2890 0 : });
2891 :
2892 : // Convert the config to a toml file.
2893 : conf_content +=
2894 : &toml_edit::ser::to_string_pretty(&location_conf).expect("Config serialization failed");
2895 :
2896 : let temp_path = path_with_suffix_extension(config_path, TEMP_FILE_SUFFIX);
2897 :
2898 : let conf_content = conf_content.into_bytes();
2899 : VirtualFile::crashsafe_overwrite(config_path.to_owned(), temp_path, conf_content).await
2900 : }
2901 :
2902 : //
2903 : // How garbage collection works:
2904 : //
2905 : // +--bar------------->
2906 : // /
2907 : // +----+-----foo---------------->
2908 : // /
2909 : // ----main--+-------------------------->
2910 : // \
2911 : // +-----baz-------->
2912 : //
2913 : //
2914 : // 1. Grab 'gc_cs' mutex to prevent new timelines from being created while Timeline's
2915 : // `gc_infos` are being refreshed
2916 : // 2. Scan collected timelines, and on each timeline, make note of the
2917 : // all the points where other timelines have been branched off.
2918 : // We will refrain from removing page versions at those LSNs.
2919 : // 3. For each timeline, scan all layer files on the timeline.
2920 : // Remove all files for which a newer file exists and which
2921 : // don't cover any branch point LSNs.
2922 : //
2923 : // TODO:
2924 : // - if a relation has a non-incremental persistent layer on a child branch, then we
2925 : // don't need to keep that in the parent anymore. But currently
2926 : // we do.
2927 2262 : async fn gc_iteration_internal(
2928 2262 : &self,
2929 2262 : target_timeline_id: Option<TimelineId>,
2930 2262 : horizon: u64,
2931 2262 : pitr: Duration,
2932 2262 : cancel: &CancellationToken,
2933 2262 : ctx: &RequestContext,
2934 2262 : ) -> Result<GcResult, GcError> {
2935 2262 : let mut totals: GcResult = Default::default();
2936 2262 : let now = Instant::now();
2937 :
2938 2262 : let gc_timelines = self
2939 2262 : .refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
2940 2235 : .await?;
2941 :
2942 2262 : failpoint_support::sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
2943 :
2944 : // If there is nothing to GC, we don't want any messages in the INFO log.
2945 2262 : if !gc_timelines.is_empty() {
2946 2262 : info!("{} timelines need GC", gc_timelines.len());
2947 : } else {
2948 0 : debug!("{} timelines need GC", gc_timelines.len());
2949 : }
2950 :
2951 : // Perform GC for each timeline.
2952 : //
2953 : // Note that we don't hold the `Tenant::gc_cs` lock here because we don't want to delay the
2954 : // branch creation task, which requires the GC lock. A GC iteration can run concurrently
2955 : // with branch creation.
2956 : //
2957 : // See comments in [`Tenant::branch_timeline`] for more information about why branch
2958 : // creation task can run concurrently with timeline's GC iteration.
2959 4524 : for timeline in gc_timelines {
2960 2262 : if cancel.is_cancelled() {
2961 : // We were requested to shut down. Stop and return with the progress we
2962 : // made.
2963 0 : break;
2964 2262 : }
2965 2262 : let result = match timeline.gc().await {
2966 : Err(GcError::TimelineCancelled) => {
2967 0 : if target_timeline_id.is_some() {
2968 : // If we were targetting this specific timeline, surface cancellation to caller
2969 0 : return Err(GcError::TimelineCancelled);
2970 : } else {
2971 : // A timeline may be shutting down independently of the tenant's lifecycle: we should
2972 : // skip past this and proceed to try GC on other timelines.
2973 0 : continue;
2974 : }
2975 : }
2976 2262 : r => r?,
2977 : };
2978 2262 : totals += result;
2979 : }
2980 :
2981 2262 : totals.elapsed = now.elapsed();
2982 2262 : Ok(totals)
2983 2262 : }
2984 :
2985 : /// Refreshes the Timeline::gc_info for all timelines, returning the
2986 : /// vector of timelines which have [`Timeline::get_last_record_lsn`] past
2987 : /// [`Tenant::get_gc_horizon`].
2988 : ///
2989 : /// This is usually executed as part of periodic gc, but can now be triggered more often.
2990 0 : pub(crate) async fn refresh_gc_info(
2991 0 : &self,
2992 0 : cancel: &CancellationToken,
2993 0 : ctx: &RequestContext,
2994 0 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
2995 0 : // since this method can now be called at different rates than the configured gc loop, it
2996 0 : // might be that these configuration values get applied faster than what it was previously,
2997 0 : // since these were only read from the gc task.
2998 0 : let horizon = self.get_gc_horizon();
2999 0 : let pitr = self.get_pitr_interval();
3000 0 :
3001 0 : // refresh all timelines
3002 0 : let target_timeline_id = None;
3003 0 :
3004 0 : self.refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
3005 0 : .await
3006 0 : }
3007 :
3008 : /// Populate all Timelines' `GcInfo` with information about their children. We do not set the
3009 : /// PITR cutoffs here, because that requires I/O: this is done later, before GC, by [`Self::refresh_gc_info_internal`]
3010 : ///
3011 : /// Subsequently, parent-child relationships are updated incrementally during timeline creation/deletion.
3012 0 : fn initialize_gc_info(
3013 0 : &self,
3014 0 : timelines: &std::sync::MutexGuard<HashMap<TimelineId, Arc<Timeline>>>,
3015 0 : ) {
3016 0 : // This function must be called before activation: after activation timeline create/delete operations
3017 0 : // might happen, and this function is not safe to run concurrently with those.
3018 0 : assert!(!self.is_active());
3019 :
3020 : // Scan all timelines. For each timeline, remember the timeline ID and
3021 : // the branch point where it was created.
3022 0 : let mut all_branchpoints: BTreeMap<TimelineId, Vec<(Lsn, TimelineId)>> = BTreeMap::new();
3023 0 : timelines.iter().for_each(|(timeline_id, timeline_entry)| {
3024 0 : if let Some(ancestor_timeline_id) = &timeline_entry.get_ancestor_timeline_id() {
3025 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
3026 0 : ancestor_children.push((timeline_entry.get_ancestor_lsn(), *timeline_id));
3027 0 : }
3028 0 : });
3029 0 :
3030 0 : // The number of bytes we always keep, irrespective of PITR: this is a constant across timelines
3031 0 : let horizon = self.get_gc_horizon();
3032 :
3033 : // Populate each timeline's GcInfo with information about its child branches
3034 0 : for timeline in timelines.values() {
3035 0 : let mut branchpoints: Vec<(Lsn, TimelineId)> = all_branchpoints
3036 0 : .remove(&timeline.timeline_id)
3037 0 : .unwrap_or_default();
3038 0 :
3039 0 : branchpoints.sort_by_key(|b| b.0);
3040 0 :
3041 0 : let mut target = timeline.gc_info.write().unwrap();
3042 0 :
3043 0 : target.retain_lsns = branchpoints;
3044 0 :
3045 0 : let space_cutoff = timeline
3046 0 : .get_last_record_lsn()
3047 0 : .checked_sub(horizon)
3048 0 : .unwrap_or(Lsn(0));
3049 0 :
3050 0 : target.cutoffs = GcCutoffs {
3051 0 : space: space_cutoff,
3052 0 : time: Lsn::INVALID,
3053 0 : };
3054 0 : }
3055 0 : }
3056 :
3057 2262 : async fn refresh_gc_info_internal(
3058 2262 : &self,
3059 2262 : target_timeline_id: Option<TimelineId>,
3060 2262 : horizon: u64,
3061 2262 : pitr: Duration,
3062 2262 : cancel: &CancellationToken,
3063 2262 : ctx: &RequestContext,
3064 2262 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
3065 2262 : // before taking the gc_cs lock, do the heavier weight finding of gc_cutoff points for
3066 2262 : // currently visible timelines.
3067 2262 : let timelines = self
3068 2262 : .timelines
3069 2262 : .lock()
3070 2262 : .unwrap()
3071 2262 : .values()
3072 9930 : .filter(|tl| match target_timeline_id.as_ref() {
3073 9930 : Some(target) => &tl.timeline_id == target,
3074 0 : None => true,
3075 9930 : })
3076 2262 : .cloned()
3077 2262 : .collect::<Vec<_>>();
3078 2262 :
3079 2262 : if target_timeline_id.is_some() && timelines.is_empty() {
3080 : // We were to act on a particular timeline and it wasn't found
3081 0 : return Err(GcError::TimelineNotFound);
3082 2262 : }
3083 2262 :
3084 2262 : let mut gc_cutoffs: HashMap<TimelineId, GcCutoffs> =
3085 2262 : HashMap::with_capacity(timelines.len());
3086 :
3087 2262 : for timeline in timelines.iter() {
3088 2262 : let cutoff = timeline
3089 2262 : .get_last_record_lsn()
3090 2262 : .checked_sub(horizon)
3091 2262 : .unwrap_or(Lsn(0));
3092 :
3093 2262 : let cutoffs = timeline.find_gc_cutoffs(cutoff, pitr, cancel, ctx).await?;
3094 2262 : let old = gc_cutoffs.insert(timeline.timeline_id, cutoffs);
3095 2262 : assert!(old.is_none());
3096 : }
3097 :
3098 2262 : if !self.is_active() || self.cancel.is_cancelled() {
3099 0 : return Err(GcError::TenantCancelled);
3100 2262 : }
3101 :
3102 : // grab mutex to prevent new timelines from being created here; avoid doing long operations
3103 : // because that will stall branch creation.
3104 2262 : let gc_cs = self.gc_cs.lock().await;
3105 :
3106 : // Ok, we now know all the branch points.
3107 : // Update the GC information for each timeline.
3108 2262 : let mut gc_timelines = Vec::with_capacity(timelines.len());
3109 4524 : for timeline in timelines {
3110 : // We filtered the timeline list above
3111 2262 : if let Some(target_timeline_id) = target_timeline_id {
3112 2262 : assert_eq!(target_timeline_id, timeline.timeline_id);
3113 0 : }
3114 :
3115 : {
3116 2262 : let mut target = timeline.gc_info.write().unwrap();
3117 2262 :
3118 2262 : // Cull any expired leases
3119 2262 : let now = SystemTime::now();
3120 2262 : target.leases.retain(|_, lease| !lease.is_expired(&now));
3121 2262 :
3122 2262 : timeline
3123 2262 : .metrics
3124 2262 : .valid_lsn_lease_count_gauge
3125 2262 : .set(target.leases.len() as u64);
3126 :
3127 : // Look up parent's PITR cutoff to update the child's knowledge of whether it is within parent's PITR
3128 2262 : if let Some(ancestor_id) = timeline.get_ancestor_timeline_id() {
3129 300 : if let Some(ancestor_gc_cutoffs) = gc_cutoffs.get(&ancestor_id) {
3130 0 : target.within_ancestor_pitr =
3131 0 : timeline.get_ancestor_lsn() >= ancestor_gc_cutoffs.time;
3132 300 : }
3133 1962 : }
3134 :
3135 : // Update metrics that depend on GC state
3136 2262 : timeline
3137 2262 : .metrics
3138 2262 : .archival_size
3139 2262 : .set(if target.within_ancestor_pitr {
3140 0 : timeline.metrics.current_logical_size_gauge.get()
3141 : } else {
3142 2262 : 0
3143 : });
3144 2262 : timeline.metrics.pitr_history_size.set(
3145 2262 : timeline
3146 2262 : .get_last_record_lsn()
3147 2262 : .checked_sub(target.cutoffs.time)
3148 2262 : .unwrap_or(Lsn(0))
3149 2262 : .0,
3150 2262 : );
3151 :
3152 : // Apply the cutoffs we found to the Timeline's GcInfo. Why might we _not_ have cutoffs for a timeline?
3153 : // - this timeline was created while we were finding cutoffs
3154 : // - lsn for timestamp search fails for this timeline repeatedly
3155 2262 : if let Some(cutoffs) = gc_cutoffs.get(&timeline.timeline_id) {
3156 2262 : target.cutoffs = cutoffs.clone();
3157 2262 : }
3158 : }
3159 :
3160 2262 : gc_timelines.push(timeline);
3161 : }
3162 2262 : drop(gc_cs);
3163 2262 : Ok(gc_timelines)
3164 2262 : }
3165 :
3166 : /// A substitute for `branch_timeline` for use in unit tests.
3167 : /// The returned timeline will have state value `Active` to make various `anyhow::ensure!()`
3168 : /// calls pass, but, we do not actually call `.activate()` under the hood. So, none of the
3169 : /// timeline background tasks are launched, except the flush loop.
3170 : #[cfg(test)]
3171 690 : async fn branch_timeline_test(
3172 690 : &self,
3173 690 : src_timeline: &Arc<Timeline>,
3174 690 : dst_id: TimelineId,
3175 690 : ancestor_lsn: Option<Lsn>,
3176 690 : ctx: &RequestContext,
3177 690 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
3178 690 : let create_guard = self.create_timeline_create_guard(dst_id).unwrap();
3179 690 : let tl = self
3180 690 : .branch_timeline_impl(src_timeline, dst_id, ancestor_lsn, create_guard, ctx)
3181 12 : .await?;
3182 678 : tl.set_state(TimelineState::Active);
3183 678 : Ok(tl)
3184 690 : }
3185 :
3186 : /// Helper for unit tests to branch a timeline with some pre-loaded states.
3187 : #[cfg(test)]
3188 : #[allow(clippy::too_many_arguments)]
3189 18 : pub async fn branch_timeline_test_with_layers(
3190 18 : &self,
3191 18 : src_timeline: &Arc<Timeline>,
3192 18 : dst_id: TimelineId,
3193 18 : ancestor_lsn: Option<Lsn>,
3194 18 : ctx: &RequestContext,
3195 18 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
3196 18 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
3197 18 : end_lsn: Lsn,
3198 18 : ) -> anyhow::Result<Arc<Timeline>> {
3199 18 : let tline = self
3200 18 : .branch_timeline_test(src_timeline, dst_id, ancestor_lsn, ctx)
3201 0 : .await?;
3202 18 : let ancestor_lsn = if let Some(ancestor_lsn) = ancestor_lsn {
3203 18 : ancestor_lsn
3204 : } else {
3205 0 : tline.get_last_record_lsn()
3206 : };
3207 18 : assert!(end_lsn >= ancestor_lsn);
3208 18 : tline.force_advance_lsn(end_lsn);
3209 36 : for deltas in delta_layer_desc {
3210 18 : tline
3211 18 : .force_create_delta_layer(deltas, Some(ancestor_lsn), ctx)
3212 54 : .await?;
3213 : }
3214 30 : for (lsn, images) in image_layer_desc {
3215 12 : tline
3216 12 : .force_create_image_layer(lsn, images, Some(ancestor_lsn), ctx)
3217 42 : .await?;
3218 : }
3219 18 : Ok(tline)
3220 18 : }
3221 :
3222 : /// Branch an existing timeline.
3223 : ///
3224 : /// The caller is responsible for activating the returned timeline.
3225 0 : async fn branch_timeline(
3226 0 : &self,
3227 0 : src_timeline: &Arc<Timeline>,
3228 0 : dst_id: TimelineId,
3229 0 : start_lsn: Option<Lsn>,
3230 0 : timeline_create_guard: TimelineCreateGuard<'_>,
3231 0 : ctx: &RequestContext,
3232 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
3233 0 : self.branch_timeline_impl(src_timeline, dst_id, start_lsn, timeline_create_guard, ctx)
3234 0 : .await
3235 0 : }
3236 :
3237 690 : async fn branch_timeline_impl(
3238 690 : &self,
3239 690 : src_timeline: &Arc<Timeline>,
3240 690 : dst_id: TimelineId,
3241 690 : start_lsn: Option<Lsn>,
3242 690 : timeline_create_guard: TimelineCreateGuard<'_>,
3243 690 : _ctx: &RequestContext,
3244 690 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
3245 690 : let src_id = src_timeline.timeline_id;
3246 :
3247 : // We will validate our ancestor LSN in this function. Acquire the GC lock so that
3248 : // this check cannot race with GC, and the ancestor LSN is guaranteed to remain
3249 : // valid while we are creating the branch.
3250 690 : let _gc_cs = self.gc_cs.lock().await;
3251 :
3252 : // If no start LSN is specified, we branch the new timeline from the source timeline's last record LSN
3253 690 : let start_lsn = start_lsn.unwrap_or_else(|| {
3254 6 : let lsn = src_timeline.get_last_record_lsn();
3255 6 : info!("branching timeline {dst_id} from timeline {src_id} at last record LSN: {lsn}");
3256 6 : lsn
3257 690 : });
3258 690 :
3259 690 : // Ensure that `start_lsn` is valid, i.e. the LSN is within the PITR
3260 690 : // horizon on the source timeline
3261 690 : //
3262 690 : // We check it against both the planned GC cutoff stored in 'gc_info',
3263 690 : // and the 'latest_gc_cutoff' of the last GC that was performed. The
3264 690 : // planned GC cutoff in 'gc_info' is normally larger than
3265 690 : // 'latest_gc_cutoff_lsn', but beware of corner cases like if you just
3266 690 : // changed the GC settings for the tenant to make the PITR window
3267 690 : // larger, but some of the data was already removed by an earlier GC
3268 690 : // iteration.
3269 690 :
3270 690 : // check against last actual 'latest_gc_cutoff' first
3271 690 : let latest_gc_cutoff_lsn = src_timeline.get_latest_gc_cutoff_lsn();
3272 690 : src_timeline
3273 690 : .check_lsn_is_in_scope(start_lsn, &latest_gc_cutoff_lsn)
3274 690 : .context(format!(
3275 690 : "invalid branch start lsn: less than latest GC cutoff {}",
3276 690 : *latest_gc_cutoff_lsn,
3277 690 : ))
3278 690 : .map_err(CreateTimelineError::AncestorLsn)?;
3279 :
3280 : // and then the planned GC cutoff
3281 : {
3282 678 : let gc_info = src_timeline.gc_info.read().unwrap();
3283 678 : let cutoff = gc_info.min_cutoff();
3284 678 : if start_lsn < cutoff {
3285 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
3286 0 : "invalid branch start lsn: less than planned GC cutoff {cutoff}"
3287 0 : )));
3288 678 : }
3289 678 : }
3290 678 :
3291 678 : //
3292 678 : // The branch point is valid, and we are still holding the 'gc_cs' lock
3293 678 : // so that GC cannot advance the GC cutoff until we are finished.
3294 678 : // Proceed with the branch creation.
3295 678 : //
3296 678 :
3297 678 : // Determine prev-LSN for the new timeline. We can only determine it if
3298 678 : // the timeline was branched at the current end of the source timeline.
3299 678 : let RecordLsn {
3300 678 : last: src_last,
3301 678 : prev: src_prev,
3302 678 : } = src_timeline.get_last_record_rlsn();
3303 678 : let dst_prev = if src_last == start_lsn {
3304 648 : Some(src_prev)
3305 : } else {
3306 30 : None
3307 : };
3308 :
3309 : // Create the metadata file, noting the ancestor of the new timeline.
3310 : // There is initially no data in it, but all the read-calls know to look
3311 : // into the ancestor.
3312 678 : let metadata = TimelineMetadata::new(
3313 678 : start_lsn,
3314 678 : dst_prev,
3315 678 : Some(src_id),
3316 678 : start_lsn,
3317 678 : *src_timeline.latest_gc_cutoff_lsn.read(), // FIXME: should we hold onto this guard longer?
3318 678 : src_timeline.initdb_lsn,
3319 678 : src_timeline.pg_version,
3320 678 : );
3321 :
3322 678 : let uninitialized_timeline = self
3323 678 : .prepare_new_timeline(
3324 678 : dst_id,
3325 678 : &metadata,
3326 678 : timeline_create_guard,
3327 678 : start_lsn + 1,
3328 678 : Some(Arc::clone(src_timeline)),
3329 678 : src_timeline.last_aux_file_policy.load(),
3330 678 : )
3331 0 : .await?;
3332 :
3333 678 : let new_timeline = uninitialized_timeline.finish_creation()?;
3334 :
3335 : // Root timeline gets its layers during creation and uploads them along with the metadata.
3336 : // A branch timeline though, when created, can get no writes for some time, hence won't get any layers created.
3337 : // We still need to upload its metadata eagerly: if other nodes `attach` the tenant and miss this timeline, their GC
3338 : // could get incorrect information and remove more layers, than needed.
3339 : // See also https://github.com/neondatabase/neon/issues/3865
3340 678 : new_timeline
3341 678 : .remote_client
3342 678 : .schedule_index_upload_for_full_metadata_update(&metadata)
3343 678 : .context("branch initial metadata upload")?;
3344 :
3345 678 : Ok(new_timeline)
3346 690 : }
3347 :
3348 : /// For unit tests, make this visible so that other modules can directly create timelines
3349 : #[cfg(test)]
3350 12 : #[tracing::instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
3351 : pub(crate) async fn bootstrap_timeline_test(
3352 : &self,
3353 : timeline_id: TimelineId,
3354 : pg_version: u32,
3355 : load_existing_initdb: Option<TimelineId>,
3356 : ctx: &RequestContext,
3357 : ) -> anyhow::Result<Arc<Timeline>> {
3358 : let create_guard = self.create_timeline_create_guard(timeline_id).unwrap();
3359 : self.bootstrap_timeline(
3360 : timeline_id,
3361 : pg_version,
3362 : load_existing_initdb,
3363 : create_guard,
3364 : ctx,
3365 : )
3366 : .await
3367 : }
3368 :
3369 0 : async fn upload_initdb(
3370 0 : &self,
3371 0 : timelines_path: &Utf8PathBuf,
3372 0 : pgdata_path: &Utf8PathBuf,
3373 0 : timeline_id: &TimelineId,
3374 0 : ) -> anyhow::Result<()> {
3375 0 : let temp_path = timelines_path.join(format!(
3376 0 : "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
3377 0 : ));
3378 :
3379 : scopeguard::defer! {
3380 : if let Err(e) = fs::remove_file(&temp_path) {
3381 : error!("Failed to remove temporary initdb archive '{temp_path}': {e}");
3382 : }
3383 : }
3384 :
3385 0 : let (pgdata_zstd, tar_zst_size) = create_zst_tarball(pgdata_path, &temp_path).await?;
3386 : const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
3387 0 : if tar_zst_size > INITDB_TAR_ZST_WARN_LIMIT {
3388 0 : warn!(
3389 0 : "compressed {temp_path} size of {tar_zst_size} is above limit {INITDB_TAR_ZST_WARN_LIMIT}."
3390 : );
3391 0 : }
3392 :
3393 : pausable_failpoint!("before-initdb-upload");
3394 :
3395 0 : backoff::retry(
3396 0 : || async {
3397 0 : self::remote_timeline_client::upload_initdb_dir(
3398 0 : &self.remote_storage,
3399 0 : &self.tenant_shard_id.tenant_id,
3400 0 : timeline_id,
3401 0 : pgdata_zstd.try_clone().await?,
3402 0 : tar_zst_size,
3403 0 : &self.cancel,
3404 0 : )
3405 0 : .await
3406 0 : },
3407 0 : |_| false,
3408 0 : 3,
3409 0 : u32::MAX,
3410 0 : "persist_initdb_tar_zst",
3411 0 : &self.cancel,
3412 0 : )
3413 0 : .await
3414 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
3415 0 : .and_then(|x| x)
3416 0 : }
3417 :
3418 : /// - run initdb to init temporary instance and get bootstrap data
3419 : /// - after initialization completes, tar up the temp dir and upload it to S3.
3420 : ///
3421 : /// The caller is responsible for activating the returned timeline.
3422 6 : async fn bootstrap_timeline(
3423 6 : &self,
3424 6 : timeline_id: TimelineId,
3425 6 : pg_version: u32,
3426 6 : load_existing_initdb: Option<TimelineId>,
3427 6 : timeline_create_guard: TimelineCreateGuard<'_>,
3428 6 : ctx: &RequestContext,
3429 6 : ) -> anyhow::Result<Arc<Timeline>> {
3430 6 : // create a `tenant/{tenant_id}/timelines/basebackup-{timeline_id}.{TEMP_FILE_SUFFIX}/`
3431 6 : // temporary directory for basebackup files for the given timeline.
3432 6 :
3433 6 : let timelines_path = self.conf.timelines_path(&self.tenant_shard_id);
3434 6 : let pgdata_path = path_with_suffix_extension(
3435 6 : timelines_path.join(format!("basebackup-{timeline_id}")),
3436 6 : TEMP_FILE_SUFFIX,
3437 6 : );
3438 6 :
3439 6 : // Remove whatever was left from the previous runs: safe because TimelineCreateGuard guarantees
3440 6 : // we won't race with other creations or existent timelines with the same path.
3441 6 : if pgdata_path.exists() {
3442 0 : fs::remove_dir_all(&pgdata_path).with_context(|| {
3443 0 : format!("Failed to remove already existing initdb directory: {pgdata_path}")
3444 0 : })?;
3445 6 : }
3446 :
3447 : // this new directory is very temporary, set to remove it immediately after bootstrap, we don't need it
3448 : scopeguard::defer! {
3449 : if let Err(e) = fs::remove_dir_all(&pgdata_path) {
3450 : // this is unlikely, but we will remove the directory on pageserver restart or another bootstrap call
3451 : error!("Failed to remove temporary initdb directory '{pgdata_path}': {e}");
3452 : }
3453 : }
3454 6 : if let Some(existing_initdb_timeline_id) = load_existing_initdb {
3455 6 : if existing_initdb_timeline_id != timeline_id {
3456 0 : let source_path = &remote_initdb_archive_path(
3457 0 : &self.tenant_shard_id.tenant_id,
3458 0 : &existing_initdb_timeline_id,
3459 0 : );
3460 0 : let dest_path =
3461 0 : &remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &timeline_id);
3462 0 :
3463 0 : // if this fails, it will get retried by retried control plane requests
3464 0 : self.remote_storage
3465 0 : .copy_object(source_path, dest_path, &self.cancel)
3466 0 : .await
3467 0 : .context("copy initdb tar")?;
3468 6 : }
3469 6 : let (initdb_tar_zst_path, initdb_tar_zst) =
3470 6 : self::remote_timeline_client::download_initdb_tar_zst(
3471 6 : self.conf,
3472 6 : &self.remote_storage,
3473 6 : &self.tenant_shard_id,
3474 6 : &existing_initdb_timeline_id,
3475 6 : &self.cancel,
3476 6 : )
3477 1923 : .await
3478 6 : .context("download initdb tar")?;
3479 :
3480 : scopeguard::defer! {
3481 : if let Err(e) = fs::remove_file(&initdb_tar_zst_path) {
3482 : error!("Failed to remove temporary initdb archive '{initdb_tar_zst_path}': {e}");
3483 : }
3484 : }
3485 :
3486 6 : let buf_read =
3487 6 : BufReader::with_capacity(remote_timeline_client::BUFFER_SIZE, initdb_tar_zst);
3488 6 : extract_zst_tarball(&pgdata_path, buf_read)
3489 30859 : .await
3490 6 : .context("extract initdb tar")?;
3491 : } else {
3492 : // Init temporarily repo to get bootstrap data, this creates a directory in the `pgdata_path` path
3493 0 : run_initdb(self.conf, &pgdata_path, pg_version, &self.cancel).await?;
3494 :
3495 : // Upload the created data dir to S3
3496 0 : if self.tenant_shard_id().is_shard_zero() {
3497 0 : self.upload_initdb(&timelines_path, &pgdata_path, &timeline_id)
3498 0 : .await?;
3499 0 : }
3500 : }
3501 6 : let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
3502 6 :
3503 6 : // Import the contents of the data directory at the initial checkpoint
3504 6 : // LSN, and any WAL after that.
3505 6 : // Initdb lsn will be equal to last_record_lsn which will be set after import.
3506 6 : // Because we know it upfront avoid having an option or dummy zero value by passing it to the metadata.
3507 6 : let new_metadata = TimelineMetadata::new(
3508 6 : Lsn(0),
3509 6 : None,
3510 6 : None,
3511 6 : Lsn(0),
3512 6 : pgdata_lsn,
3513 6 : pgdata_lsn,
3514 6 : pg_version,
3515 6 : );
3516 6 : let raw_timeline = self
3517 6 : .prepare_new_timeline(
3518 6 : timeline_id,
3519 6 : &new_metadata,
3520 6 : timeline_create_guard,
3521 6 : pgdata_lsn,
3522 6 : None,
3523 6 : None,
3524 6 : )
3525 0 : .await?;
3526 :
3527 6 : let tenant_shard_id = raw_timeline.owning_tenant.tenant_shard_id;
3528 6 : let unfinished_timeline = raw_timeline.raw_timeline()?;
3529 :
3530 : // Flush the new layer files to disk, before we make the timeline as available to
3531 : // the outside world.
3532 : //
3533 : // Flush loop needs to be spawned in order to be able to flush.
3534 6 : unfinished_timeline.maybe_spawn_flush_loop();
3535 6 :
3536 6 : import_datadir::import_timeline_from_postgres_datadir(
3537 6 : unfinished_timeline,
3538 6 : &pgdata_path,
3539 6 : pgdata_lsn,
3540 6 : ctx,
3541 6 : )
3542 28917 : .await
3543 6 : .with_context(|| {
3544 0 : format!("Failed to import pgdatadir for timeline {tenant_shard_id}/{timeline_id}")
3545 6 : })?;
3546 :
3547 6 : fail::fail_point!("before-checkpoint-new-timeline", |_| {
3548 0 : anyhow::bail!("failpoint before-checkpoint-new-timeline");
3549 6 : });
3550 :
3551 6 : unfinished_timeline
3552 6 : .freeze_and_flush()
3553 6 : .await
3554 6 : .with_context(|| {
3555 0 : format!(
3556 0 : "Failed to flush after pgdatadir import for timeline {tenant_shard_id}/{timeline_id}"
3557 0 : )
3558 6 : })?;
3559 :
3560 : // All done!
3561 6 : let timeline = raw_timeline.finish_creation()?;
3562 :
3563 6 : Ok(timeline)
3564 6 : }
3565 :
3566 : /// Call this before constructing a timeline, to build its required structures
3567 1218 : fn build_timeline_resources(&self, timeline_id: TimelineId) -> TimelineResources {
3568 1218 : let remote_client = RemoteTimelineClient::new(
3569 1218 : self.remote_storage.clone(),
3570 1218 : self.deletion_queue_client.clone(),
3571 1218 : self.conf,
3572 1218 : self.tenant_shard_id,
3573 1218 : timeline_id,
3574 1218 : self.generation,
3575 1218 : );
3576 1218 : TimelineResources {
3577 1218 : remote_client,
3578 1218 : timeline_get_throttle: self.timeline_get_throttle.clone(),
3579 1218 : l0_flush_global_state: self.l0_flush_global_state.clone(),
3580 1218 : }
3581 1218 : }
3582 :
3583 : /// Creates intermediate timeline structure and its files.
3584 : ///
3585 : /// An empty layer map is initialized, and new data and WAL can be imported starting
3586 : /// at 'disk_consistent_lsn'. After any initial data has been imported, call
3587 : /// `finish_creation` to insert the Timeline into the timelines map.
3588 1218 : async fn prepare_new_timeline<'a>(
3589 1218 : &'a self,
3590 1218 : new_timeline_id: TimelineId,
3591 1218 : new_metadata: &TimelineMetadata,
3592 1218 : create_guard: TimelineCreateGuard<'a>,
3593 1218 : start_lsn: Lsn,
3594 1218 : ancestor: Option<Arc<Timeline>>,
3595 1218 : last_aux_file_policy: Option<AuxFilePolicy>,
3596 1218 : ) -> anyhow::Result<UninitializedTimeline> {
3597 1218 : let tenant_shard_id = self.tenant_shard_id;
3598 1218 :
3599 1218 : let resources = self.build_timeline_resources(new_timeline_id);
3600 1218 : resources
3601 1218 : .remote_client
3602 1218 : .init_upload_queue_for_empty_remote(new_metadata)?;
3603 :
3604 1218 : let timeline_struct = self
3605 1218 : .create_timeline_struct(
3606 1218 : new_timeline_id,
3607 1218 : new_metadata,
3608 1218 : ancestor,
3609 1218 : resources,
3610 1218 : CreateTimelineCause::Load,
3611 1218 : last_aux_file_policy,
3612 1218 : )
3613 1218 : .context("Failed to create timeline data structure")?;
3614 :
3615 1218 : timeline_struct.init_empty_layer_map(start_lsn);
3616 :
3617 1218 : if let Err(e) = self
3618 1218 : .create_timeline_files(&create_guard.timeline_path)
3619 0 : .await
3620 : {
3621 0 : error!("Failed to create initial files for timeline {tenant_shard_id}/{new_timeline_id}, cleaning up: {e:?}");
3622 0 : cleanup_timeline_directory(create_guard);
3623 0 : return Err(e);
3624 1218 : }
3625 1218 :
3626 1218 : debug!(
3627 0 : "Successfully created initial files for timeline {tenant_shard_id}/{new_timeline_id}"
3628 : );
3629 :
3630 1218 : Ok(UninitializedTimeline::new(
3631 1218 : self,
3632 1218 : new_timeline_id,
3633 1218 : Some((timeline_struct, create_guard)),
3634 1218 : ))
3635 1218 : }
3636 :
3637 1218 : async fn create_timeline_files(&self, timeline_path: &Utf8Path) -> anyhow::Result<()> {
3638 1218 : crashsafe::create_dir(timeline_path).context("Failed to create timeline directory")?;
3639 :
3640 1218 : fail::fail_point!("after-timeline-dir-creation", |_| {
3641 0 : anyhow::bail!("failpoint after-timeline-dir-creation");
3642 1218 : });
3643 :
3644 1218 : Ok(())
3645 1218 : }
3646 :
3647 : /// Get a guard that provides exclusive access to the timeline directory, preventing
3648 : /// concurrent attempts to create the same timeline.
3649 1236 : fn create_timeline_create_guard(
3650 1236 : &self,
3651 1236 : timeline_id: TimelineId,
3652 1236 : ) -> Result<TimelineCreateGuard, TimelineExclusionError> {
3653 1236 : let tenant_shard_id = self.tenant_shard_id;
3654 1236 :
3655 1236 : let timeline_path = self.conf.timeline_path(&tenant_shard_id, &timeline_id);
3656 :
3657 1236 : let create_guard = TimelineCreateGuard::new(self, timeline_id, timeline_path.clone())?;
3658 :
3659 : // At this stage, we have got exclusive access to in-memory state for this timeline ID
3660 : // for creation.
3661 : // A timeline directory should never exist on disk already:
3662 : // - a previous failed creation would have cleaned up after itself
3663 : // - a pageserver restart would clean up timeline directories that don't have valid remote state
3664 : //
3665 : // Therefore it is an unexpected internal error to encounter a timeline directory already existing here,
3666 : // this error may indicate a bug in cleanup on failed creations.
3667 1230 : if timeline_path.exists() {
3668 0 : return Err(TimelineExclusionError::Other(anyhow::anyhow!(
3669 0 : "Timeline directory already exists! This is a bug."
3670 0 : )));
3671 1230 : }
3672 1230 :
3673 1230 : Ok(create_guard)
3674 1236 : }
3675 :
3676 : /// Gathers inputs from all of the timelines to produce a sizing model input.
3677 : ///
3678 : /// Future is cancellation safe. Only one calculation can be running at once per tenant.
3679 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
3680 : pub async fn gather_size_inputs(
3681 : &self,
3682 : // `max_retention_period` overrides the cutoff that is used to calculate the size
3683 : // (only if it is shorter than the real cutoff).
3684 : max_retention_period: Option<u64>,
3685 : cause: LogicalSizeCalculationCause,
3686 : cancel: &CancellationToken,
3687 : ctx: &RequestContext,
3688 : ) -> Result<size::ModelInputs, size::CalculateSyntheticSizeError> {
3689 : let logical_sizes_at_once = self
3690 : .conf
3691 : .concurrent_tenant_size_logical_size_queries
3692 : .inner();
3693 :
3694 : // TODO: Having a single mutex block concurrent reads is not great for performance.
3695 : //
3696 : // But the only case where we need to run multiple of these at once is when we
3697 : // request a size for a tenant manually via API, while another background calculation
3698 : // is in progress (which is not a common case).
3699 : //
3700 : // See more for on the issue #2748 condenced out of the initial PR review.
3701 : let mut shared_cache = tokio::select! {
3702 : locked = self.cached_logical_sizes.lock() => locked,
3703 : _ = cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
3704 : _ = self.cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
3705 : };
3706 :
3707 : size::gather_inputs(
3708 : self,
3709 : logical_sizes_at_once,
3710 : max_retention_period,
3711 : &mut shared_cache,
3712 : cause,
3713 : cancel,
3714 : ctx,
3715 : )
3716 : .await
3717 : }
3718 :
3719 : /// Calculate synthetic tenant size and cache the result.
3720 : /// This is periodically called by background worker.
3721 : /// result is cached in tenant struct
3722 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
3723 : pub async fn calculate_synthetic_size(
3724 : &self,
3725 : cause: LogicalSizeCalculationCause,
3726 : cancel: &CancellationToken,
3727 : ctx: &RequestContext,
3728 : ) -> Result<u64, size::CalculateSyntheticSizeError> {
3729 : let inputs = self.gather_size_inputs(None, cause, cancel, ctx).await?;
3730 :
3731 : let size = inputs.calculate();
3732 :
3733 : self.set_cached_synthetic_size(size);
3734 :
3735 : Ok(size)
3736 : }
3737 :
3738 : /// Cache given synthetic size and update the metric value
3739 0 : pub fn set_cached_synthetic_size(&self, size: u64) {
3740 0 : self.cached_synthetic_tenant_size
3741 0 : .store(size, Ordering::Relaxed);
3742 0 :
3743 0 : // Only shard zero should be calculating synthetic sizes
3744 0 : debug_assert!(self.shard_identity.is_shard_zero());
3745 :
3746 0 : TENANT_SYNTHETIC_SIZE_METRIC
3747 0 : .get_metric_with_label_values(&[&self.tenant_shard_id.tenant_id.to_string()])
3748 0 : .unwrap()
3749 0 : .set(size);
3750 0 : }
3751 :
3752 0 : pub fn cached_synthetic_size(&self) -> u64 {
3753 0 : self.cached_synthetic_tenant_size.load(Ordering::Relaxed)
3754 0 : }
3755 :
3756 : /// Flush any in-progress layers, schedule uploads, and wait for uploads to complete.
3757 : ///
3758 : /// This function can take a long time: callers should wrap it in a timeout if calling
3759 : /// from an external API handler.
3760 : ///
3761 : /// Cancel-safety: cancelling this function may leave I/O running, but such I/O is
3762 : /// still bounded by tenant/timeline shutdown.
3763 0 : #[tracing::instrument(skip_all)]
3764 : pub(crate) async fn flush_remote(&self) -> anyhow::Result<()> {
3765 : let timelines = self.timelines.lock().unwrap().clone();
3766 :
3767 0 : async fn flush_timeline(_gate: GateGuard, timeline: Arc<Timeline>) -> anyhow::Result<()> {
3768 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Flushing...");
3769 0 : timeline.freeze_and_flush().await?;
3770 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Waiting for uploads...");
3771 0 : timeline.remote_client.wait_completion().await?;
3772 :
3773 0 : Ok(())
3774 0 : }
3775 :
3776 : // We do not use a JoinSet for these tasks, because we don't want them to be
3777 : // aborted when this function's future is cancelled: they should stay alive
3778 : // holding their GateGuard until they complete, to ensure their I/Os complete
3779 : // before Timeline shutdown completes.
3780 : let mut results = FuturesUnordered::new();
3781 :
3782 : for (_timeline_id, timeline) in timelines {
3783 : // Run each timeline's flush in a task holding the timeline's gate: this
3784 : // means that if this function's future is cancelled, the Timeline shutdown
3785 : // will still wait for any I/O in here to complete.
3786 : let Ok(gate) = timeline.gate.enter() else {
3787 : continue;
3788 : };
3789 0 : let jh = tokio::task::spawn(async move { flush_timeline(gate, timeline).await });
3790 : results.push(jh);
3791 : }
3792 :
3793 : while let Some(r) = results.next().await {
3794 : if let Err(e) = r {
3795 : if !e.is_cancelled() && !e.is_panic() {
3796 : tracing::error!("unexpected join error: {e:?}");
3797 : }
3798 : }
3799 : }
3800 :
3801 : // The flushes we did above were just writes, but the Tenant might have had
3802 : // pending deletions as well from recent compaction/gc: we want to flush those
3803 : // as well. This requires flushing the global delete queue. This is cheap
3804 : // because it's typically a no-op.
3805 : match self.deletion_queue_client.flush_execute().await {
3806 : Ok(_) => {}
3807 : Err(DeletionQueueError::ShuttingDown) => {}
3808 : }
3809 :
3810 : Ok(())
3811 : }
3812 :
3813 0 : pub(crate) fn get_tenant_conf(&self) -> TenantConfOpt {
3814 0 : self.tenant_conf.load().tenant_conf.clone()
3815 0 : }
3816 :
3817 : /// How much local storage would this tenant like to have? It can cope with
3818 : /// less than this (via eviction and on-demand downloads), but this function enables
3819 : /// the Tenant to advertise how much storage it would prefer to have to provide fast I/O
3820 : /// by keeping important things on local disk.
3821 : ///
3822 : /// This is a heuristic, not a guarantee: tenants that are long-idle will actually use less
3823 : /// than they report here, due to layer eviction. Tenants with many active branches may
3824 : /// actually use more than they report here.
3825 0 : pub(crate) fn local_storage_wanted(&self) -> u64 {
3826 0 : let timelines = self.timelines.lock().unwrap();
3827 0 :
3828 0 : // Heuristic: we use the max() of the timelines' visible sizes, rather than the sum. This
3829 0 : // reflects the observation that on tenants with multiple large branches, typically only one
3830 0 : // of them is used actively enough to occupy space on disk.
3831 0 : timelines
3832 0 : .values()
3833 0 : .map(|t| t.metrics.visible_physical_size_gauge.get())
3834 0 : .max()
3835 0 : .unwrap_or(0)
3836 0 : }
3837 : }
3838 :
3839 : /// Create the cluster temporarily in 'initdbpath' directory inside the repository
3840 : /// to get bootstrap data for timeline initialization.
3841 0 : async fn run_initdb(
3842 0 : conf: &'static PageServerConf,
3843 0 : initdb_target_dir: &Utf8Path,
3844 0 : pg_version: u32,
3845 0 : cancel: &CancellationToken,
3846 0 : ) -> Result<(), InitdbError> {
3847 0 : let initdb_bin_path = conf
3848 0 : .pg_bin_dir(pg_version)
3849 0 : .map_err(InitdbError::Other)?
3850 0 : .join("initdb");
3851 0 : let initdb_lib_dir = conf.pg_lib_dir(pg_version).map_err(InitdbError::Other)?;
3852 0 : info!(
3853 0 : "running {} in {}, libdir: {}",
3854 : initdb_bin_path, initdb_target_dir, initdb_lib_dir,
3855 : );
3856 :
3857 0 : let _permit = INIT_DB_SEMAPHORE.acquire().await;
3858 :
3859 0 : let initdb_command = tokio::process::Command::new(&initdb_bin_path)
3860 0 : .args(["-D", initdb_target_dir.as_ref()])
3861 0 : .args(["-U", &conf.superuser])
3862 0 : .args(["-E", "utf8"])
3863 0 : .arg("--no-instructions")
3864 0 : .arg("--no-sync")
3865 0 : .env_clear()
3866 0 : .env("LD_LIBRARY_PATH", &initdb_lib_dir)
3867 0 : .env("DYLD_LIBRARY_PATH", &initdb_lib_dir)
3868 0 : .stdin(std::process::Stdio::null())
3869 0 : // stdout invocation produces the same output every time, we don't need it
3870 0 : .stdout(std::process::Stdio::null())
3871 0 : // we would be interested in the stderr output, if there was any
3872 0 : .stderr(std::process::Stdio::piped())
3873 0 : .spawn()?;
3874 :
3875 : // Ideally we'd select here with the cancellation token, but the problem is that
3876 : // we can't safely terminate initdb: it launches processes of its own, and killing
3877 : // initdb doesn't kill them. After we return from this function, we want the target
3878 : // directory to be able to be cleaned up.
3879 : // See https://github.com/neondatabase/neon/issues/6385
3880 0 : let initdb_output = initdb_command.wait_with_output().await?;
3881 0 : if !initdb_output.status.success() {
3882 0 : return Err(InitdbError::Failed(
3883 0 : initdb_output.status,
3884 0 : initdb_output.stderr,
3885 0 : ));
3886 0 : }
3887 0 :
3888 0 : // This isn't true cancellation support, see above. Still return an error to
3889 0 : // excercise the cancellation code path.
3890 0 : if cancel.is_cancelled() {
3891 0 : return Err(InitdbError::Cancelled);
3892 0 : }
3893 0 :
3894 0 : Ok(())
3895 0 : }
3896 :
3897 : /// Dump contents of a layer file to stdout.
3898 0 : pub async fn dump_layerfile_from_path(
3899 0 : path: &Utf8Path,
3900 0 : verbose: bool,
3901 0 : ctx: &RequestContext,
3902 0 : ) -> anyhow::Result<()> {
3903 : use std::os::unix::fs::FileExt;
3904 :
3905 : // All layer files start with a two-byte "magic" value, to identify the kind of
3906 : // file.
3907 0 : let file = File::open(path)?;
3908 0 : let mut header_buf = [0u8; 2];
3909 0 : file.read_exact_at(&mut header_buf, 0)?;
3910 :
3911 0 : match u16::from_be_bytes(header_buf) {
3912 : crate::IMAGE_FILE_MAGIC => {
3913 0 : ImageLayer::new_for_path(path, file)?
3914 0 : .dump(verbose, ctx)
3915 0 : .await?
3916 : }
3917 : crate::DELTA_FILE_MAGIC => {
3918 0 : DeltaLayer::new_for_path(path, file)?
3919 0 : .dump(verbose, ctx)
3920 0 : .await?
3921 : }
3922 0 : magic => bail!("unrecognized magic identifier: {:?}", magic),
3923 : }
3924 :
3925 0 : Ok(())
3926 0 : }
3927 :
3928 : #[cfg(test)]
3929 : pub(crate) mod harness {
3930 : use bytes::{Bytes, BytesMut};
3931 : use once_cell::sync::OnceCell;
3932 : use pageserver_api::models::ShardParameters;
3933 : use pageserver_api::shard::ShardIndex;
3934 : use utils::logging;
3935 :
3936 : use crate::deletion_queue::mock::MockDeletionQueue;
3937 : use crate::l0_flush::L0FlushConfig;
3938 : use crate::walredo::apply_neon;
3939 : use crate::{repository::Key, walrecord::NeonWalRecord};
3940 :
3941 : use super::*;
3942 : use hex_literal::hex;
3943 : use utils::id::TenantId;
3944 :
3945 : pub const TIMELINE_ID: TimelineId =
3946 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
3947 : pub const NEW_TIMELINE_ID: TimelineId =
3948 : TimelineId::from_array(hex!("AA223344556677881122334455667788"));
3949 :
3950 : /// Convenience function to create a page image with given string as the only content
3951 15086096 : pub fn test_img(s: &str) -> Bytes {
3952 15086096 : let mut buf = BytesMut::new();
3953 15086096 : buf.extend_from_slice(s.as_bytes());
3954 15086096 : buf.resize(64, 0);
3955 15086096 :
3956 15086096 : buf.freeze()
3957 15086096 : }
3958 :
3959 : impl From<TenantConf> for TenantConfOpt {
3960 564 : fn from(tenant_conf: TenantConf) -> Self {
3961 564 : Self {
3962 564 : checkpoint_distance: Some(tenant_conf.checkpoint_distance),
3963 564 : checkpoint_timeout: Some(tenant_conf.checkpoint_timeout),
3964 564 : compaction_target_size: Some(tenant_conf.compaction_target_size),
3965 564 : compaction_period: Some(tenant_conf.compaction_period),
3966 564 : compaction_threshold: Some(tenant_conf.compaction_threshold),
3967 564 : compaction_algorithm: Some(tenant_conf.compaction_algorithm),
3968 564 : gc_horizon: Some(tenant_conf.gc_horizon),
3969 564 : gc_period: Some(tenant_conf.gc_period),
3970 564 : image_creation_threshold: Some(tenant_conf.image_creation_threshold),
3971 564 : pitr_interval: Some(tenant_conf.pitr_interval),
3972 564 : walreceiver_connect_timeout: Some(tenant_conf.walreceiver_connect_timeout),
3973 564 : lagging_wal_timeout: Some(tenant_conf.lagging_wal_timeout),
3974 564 : max_lsn_wal_lag: Some(tenant_conf.max_lsn_wal_lag),
3975 564 : eviction_policy: Some(tenant_conf.eviction_policy),
3976 564 : min_resident_size_override: tenant_conf.min_resident_size_override,
3977 564 : evictions_low_residence_duration_metric_threshold: Some(
3978 564 : tenant_conf.evictions_low_residence_duration_metric_threshold,
3979 564 : ),
3980 564 : heatmap_period: Some(tenant_conf.heatmap_period),
3981 564 : lazy_slru_download: Some(tenant_conf.lazy_slru_download),
3982 564 : timeline_get_throttle: Some(tenant_conf.timeline_get_throttle),
3983 564 : image_layer_creation_check_threshold: Some(
3984 564 : tenant_conf.image_layer_creation_check_threshold,
3985 564 : ),
3986 564 : switch_aux_file_policy: Some(tenant_conf.switch_aux_file_policy),
3987 564 : lsn_lease_length: Some(tenant_conf.lsn_lease_length),
3988 564 : lsn_lease_length_for_ts: Some(tenant_conf.lsn_lease_length_for_ts),
3989 564 : }
3990 564 : }
3991 : }
3992 :
3993 : pub struct TenantHarness {
3994 : pub conf: &'static PageServerConf,
3995 : pub tenant_conf: TenantConf,
3996 : pub tenant_shard_id: TenantShardId,
3997 : pub generation: Generation,
3998 : pub shard: ShardIndex,
3999 : pub remote_storage: GenericRemoteStorage,
4000 : pub remote_fs_dir: Utf8PathBuf,
4001 : pub deletion_queue: MockDeletionQueue,
4002 : }
4003 :
4004 : static LOG_HANDLE: OnceCell<()> = OnceCell::new();
4005 :
4006 606 : pub(crate) fn setup_logging() {
4007 606 : LOG_HANDLE.get_or_init(|| {
4008 570 : logging::init(
4009 570 : logging::LogFormat::Test,
4010 570 : // enable it in case the tests exercise code paths that use
4011 570 : // debug_assert_current_span_has_tenant_and_timeline_id
4012 570 : logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
4013 570 : logging::Output::Stdout,
4014 570 : )
4015 570 : .expect("Failed to init test logging")
4016 606 : });
4017 606 : }
4018 :
4019 : impl TenantHarness {
4020 564 : pub async fn create_custom(
4021 564 : test_name: &'static str,
4022 564 : tenant_conf: TenantConf,
4023 564 : tenant_id: TenantId,
4024 564 : shard_identity: ShardIdentity,
4025 564 : generation: Generation,
4026 564 : ) -> anyhow::Result<Self> {
4027 564 : setup_logging();
4028 564 :
4029 564 : let repo_dir = PageServerConf::test_repo_dir(test_name);
4030 564 : let _ = fs::remove_dir_all(&repo_dir);
4031 564 : fs::create_dir_all(&repo_dir)?;
4032 :
4033 564 : let conf = PageServerConf::dummy_conf(repo_dir);
4034 564 : // Make a static copy of the config. This can never be free'd, but that's
4035 564 : // OK in a test.
4036 564 : let conf: &'static PageServerConf = Box::leak(Box::new(conf));
4037 564 :
4038 564 : let shard = shard_identity.shard_index();
4039 564 : let tenant_shard_id = TenantShardId {
4040 564 : tenant_id,
4041 564 : shard_number: shard.shard_number,
4042 564 : shard_count: shard.shard_count,
4043 564 : };
4044 564 : fs::create_dir_all(conf.tenant_path(&tenant_shard_id))?;
4045 564 : fs::create_dir_all(conf.timelines_path(&tenant_shard_id))?;
4046 :
4047 : use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
4048 564 : let remote_fs_dir = conf.workdir.join("localfs");
4049 564 : std::fs::create_dir_all(&remote_fs_dir).unwrap();
4050 564 : let config = RemoteStorageConfig {
4051 564 : storage: RemoteStorageKind::LocalFs {
4052 564 : local_path: remote_fs_dir.clone(),
4053 564 : },
4054 564 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
4055 564 : };
4056 564 : let remote_storage = GenericRemoteStorage::from_config(&config).await.unwrap();
4057 564 : let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()));
4058 564 :
4059 564 : Ok(Self {
4060 564 : conf,
4061 564 : tenant_conf,
4062 564 : tenant_shard_id,
4063 564 : generation,
4064 564 : shard,
4065 564 : remote_storage,
4066 564 : remote_fs_dir,
4067 564 : deletion_queue,
4068 564 : })
4069 564 : }
4070 :
4071 528 : pub async fn create(test_name: &'static str) -> anyhow::Result<Self> {
4072 528 : // Disable automatic GC and compaction to make the unit tests more deterministic.
4073 528 : // The tests perform them manually if needed.
4074 528 : let tenant_conf = TenantConf {
4075 528 : gc_period: Duration::ZERO,
4076 528 : compaction_period: Duration::ZERO,
4077 528 : ..TenantConf::default()
4078 528 : };
4079 528 : let tenant_id = TenantId::generate();
4080 528 : let shard = ShardIdentity::unsharded();
4081 528 : Self::create_custom(
4082 528 : test_name,
4083 528 : tenant_conf,
4084 528 : tenant_id,
4085 528 : shard,
4086 528 : Generation::new(0xdeadbeef),
4087 528 : )
4088 0 : .await
4089 528 : }
4090 :
4091 60 : pub fn span(&self) -> tracing::Span {
4092 60 : info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
4093 60 : }
4094 :
4095 564 : pub(crate) async fn load(&self) -> (Arc<Tenant>, RequestContext) {
4096 564 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
4097 564 : (
4098 564 : self.do_try_load(&ctx)
4099 2259 : .await
4100 564 : .expect("failed to load test tenant"),
4101 564 : ctx,
4102 564 : )
4103 564 : }
4104 :
4105 1128 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
4106 : pub(crate) async fn do_try_load(
4107 : &self,
4108 : ctx: &RequestContext,
4109 : ) -> anyhow::Result<Arc<Tenant>> {
4110 : let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
4111 :
4112 : let tenant = Arc::new(Tenant::new(
4113 : TenantState::Loading,
4114 : self.conf,
4115 : AttachedTenantConf::try_from(LocationConf::attached_single(
4116 : TenantConfOpt::from(self.tenant_conf.clone()),
4117 : self.generation,
4118 : &ShardParameters::default(),
4119 : ))
4120 : .unwrap(),
4121 : // This is a legacy/test code path: sharding isn't supported here.
4122 : ShardIdentity::unsharded(),
4123 : Some(walredo_mgr),
4124 : self.tenant_shard_id,
4125 : self.remote_storage.clone(),
4126 : self.deletion_queue.new_client(),
4127 : // TODO: ideally we should run all unit tests with both configs
4128 : L0FlushGlobalState::new(L0FlushConfig::default()),
4129 : ));
4130 :
4131 : let preload = tenant
4132 : .preload(&self.remote_storage, CancellationToken::new())
4133 : .await?;
4134 : tenant.attach(Some(preload), ctx).await?;
4135 :
4136 : tenant.state.send_replace(TenantState::Active);
4137 : for timeline in tenant.timelines.lock().unwrap().values() {
4138 : timeline.set_state(TimelineState::Active);
4139 : }
4140 : Ok(tenant)
4141 : }
4142 :
4143 6 : pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
4144 6 : self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
4145 6 : }
4146 : }
4147 :
4148 : // Mock WAL redo manager that doesn't do much
4149 : pub(crate) struct TestRedoManager;
4150 :
4151 : impl TestRedoManager {
4152 : /// # Cancel-Safety
4153 : ///
4154 : /// This method is cancellation-safe.
4155 978 : pub async fn request_redo(
4156 978 : &self,
4157 978 : key: Key,
4158 978 : lsn: Lsn,
4159 978 : base_img: Option<(Lsn, Bytes)>,
4160 978 : records: Vec<(Lsn, NeonWalRecord)>,
4161 978 : _pg_version: u32,
4162 978 : ) -> Result<Bytes, walredo::Error> {
4163 1326 : let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
4164 978 : if records_neon {
4165 : // For Neon wal records, we can decode without spawning postgres, so do so.
4166 978 : let base_img = base_img.expect("Neon WAL redo requires base image").1;
4167 978 : let mut page = BytesMut::new();
4168 978 : page.extend_from_slice(&base_img);
4169 2304 : for (record_lsn, record) in records {
4170 1326 : apply_neon::apply_in_neon(&record, record_lsn, key, &mut page)?;
4171 : }
4172 978 : Ok(page.freeze())
4173 : } else {
4174 : // We never spawn a postgres walredo process in unit tests: just log what we might have done.
4175 0 : let s = format!(
4176 0 : "redo for {} to get to {}, with {} and {} records",
4177 0 : key,
4178 0 : lsn,
4179 0 : if base_img.is_some() {
4180 0 : "base image"
4181 : } else {
4182 0 : "no base image"
4183 : },
4184 0 : records.len()
4185 0 : );
4186 0 : println!("{s}");
4187 0 :
4188 0 : Ok(test_img(&s))
4189 : }
4190 978 : }
4191 : }
4192 : }
4193 :
4194 : #[cfg(test)]
4195 : mod tests {
4196 : use std::collections::{BTreeMap, BTreeSet};
4197 :
4198 : use super::*;
4199 : use crate::keyspace::KeySpaceAccum;
4200 : use crate::pgdatadir_mapping::AuxFilesDirectory;
4201 : use crate::repository::{Key, Value};
4202 : use crate::tenant::harness::*;
4203 : use crate::tenant::timeline::CompactFlags;
4204 : use crate::walrecord::NeonWalRecord;
4205 : use crate::DEFAULT_PG_VERSION;
4206 : use bytes::{Bytes, BytesMut};
4207 : use hex_literal::hex;
4208 : use itertools::Itertools;
4209 : use pageserver_api::key::{AUX_FILES_KEY, AUX_KEY_PREFIX, NON_INHERITED_RANGE};
4210 : use pageserver_api::keyspace::KeySpace;
4211 : use pageserver_api::models::{CompactionAlgorithm, CompactionAlgorithmSettings};
4212 : use rand::{thread_rng, Rng};
4213 : use storage_layer::PersistentLayerKey;
4214 : use tests::storage_layer::ValuesReconstructState;
4215 : use tests::timeline::{GetVectoredError, ShutdownMode};
4216 : use timeline::compaction::{KeyHistoryRetention, KeyLogAtLsn};
4217 : use timeline::{DeltaLayerTestDesc, GcInfo};
4218 : use utils::bin_ser::BeSer;
4219 : use utils::id::TenantId;
4220 :
4221 : static TEST_KEY: Lazy<Key> =
4222 54 : Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));
4223 :
4224 : #[tokio::test]
4225 6 : async fn test_basic() -> anyhow::Result<()> {
4226 24 : let (tenant, ctx) = TenantHarness::create("test_basic").await?.load().await;
4227 6 : let tline = tenant
4228 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
4229 12 : .await?;
4230 6 :
4231 6 : let mut writer = tline.writer().await;
4232 6 : writer
4233 6 : .put(
4234 6 : *TEST_KEY,
4235 6 : Lsn(0x10),
4236 6 : &Value::Image(test_img("foo at 0x10")),
4237 6 : &ctx,
4238 6 : )
4239 6 : .await?;
4240 6 : writer.finish_write(Lsn(0x10));
4241 6 : drop(writer);
4242 6 :
4243 6 : let mut writer = tline.writer().await;
4244 6 : writer
4245 6 : .put(
4246 6 : *TEST_KEY,
4247 6 : Lsn(0x20),
4248 6 : &Value::Image(test_img("foo at 0x20")),
4249 6 : &ctx,
4250 6 : )
4251 6 : .await?;
4252 6 : writer.finish_write(Lsn(0x20));
4253 6 : drop(writer);
4254 6 :
4255 6 : assert_eq!(
4256 6 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
4257 6 : test_img("foo at 0x10")
4258 6 : );
4259 6 : assert_eq!(
4260 6 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
4261 6 : test_img("foo at 0x10")
4262 6 : );
4263 6 : assert_eq!(
4264 6 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
4265 6 : test_img("foo at 0x20")
4266 6 : );
4267 6 :
4268 6 : Ok(())
4269 6 : }
4270 :
4271 : #[tokio::test]
4272 6 : async fn no_duplicate_timelines() -> anyhow::Result<()> {
4273 6 : let (tenant, ctx) = TenantHarness::create("no_duplicate_timelines")
4274 6 : .await?
4275 6 : .load()
4276 24 : .await;
4277 6 : let _ = tenant
4278 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4279 12 : .await?;
4280 6 :
4281 6 : match tenant
4282 6 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4283 6 : .await
4284 6 : {
4285 6 : Ok(_) => panic!("duplicate timeline creation should fail"),
4286 6 : Err(e) => assert_eq!(e.to_string(), "Already exists".to_string()),
4287 6 : }
4288 6 :
4289 6 : Ok(())
4290 6 : }
4291 :
4292 : /// Convenience function to create a page image with given string as the only content
4293 30 : pub fn test_value(s: &str) -> Value {
4294 30 : let mut buf = BytesMut::new();
4295 30 : buf.extend_from_slice(s.as_bytes());
4296 30 : Value::Image(buf.freeze())
4297 30 : }
4298 :
4299 : ///
4300 : /// Test branch creation
4301 : ///
4302 : #[tokio::test]
4303 6 : async fn test_branch() -> anyhow::Result<()> {
4304 6 : use std::str::from_utf8;
4305 6 :
4306 24 : let (tenant, ctx) = TenantHarness::create("test_branch").await?.load().await;
4307 6 : let tline = tenant
4308 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4309 12 : .await?;
4310 6 : let mut writer = tline.writer().await;
4311 6 :
4312 6 : #[allow(non_snake_case)]
4313 6 : let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
4314 6 : #[allow(non_snake_case)]
4315 6 : let TEST_KEY_B: Key = Key::from_hex("110000000033333333444444445500000002").unwrap();
4316 6 :
4317 6 : // Insert a value on the timeline
4318 6 : writer
4319 6 : .put(TEST_KEY_A, Lsn(0x20), &test_value("foo at 0x20"), &ctx)
4320 6 : .await?;
4321 6 : writer
4322 6 : .put(TEST_KEY_B, Lsn(0x20), &test_value("foobar at 0x20"), &ctx)
4323 6 : .await?;
4324 6 : writer.finish_write(Lsn(0x20));
4325 6 :
4326 6 : writer
4327 6 : .put(TEST_KEY_A, Lsn(0x30), &test_value("foo at 0x30"), &ctx)
4328 6 : .await?;
4329 6 : writer.finish_write(Lsn(0x30));
4330 6 : writer
4331 6 : .put(TEST_KEY_A, Lsn(0x40), &test_value("foo at 0x40"), &ctx)
4332 6 : .await?;
4333 6 : writer.finish_write(Lsn(0x40));
4334 6 :
4335 6 : //assert_current_logical_size(&tline, Lsn(0x40));
4336 6 :
4337 6 : // Branch the history, modify relation differently on the new timeline
4338 6 : tenant
4339 6 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x30)), &ctx)
4340 6 : .await?;
4341 6 : let newtline = tenant
4342 6 : .get_timeline(NEW_TIMELINE_ID, true)
4343 6 : .expect("Should have a local timeline");
4344 6 : let mut new_writer = newtline.writer().await;
4345 6 : new_writer
4346 6 : .put(TEST_KEY_A, Lsn(0x40), &test_value("bar at 0x40"), &ctx)
4347 6 : .await?;
4348 6 : new_writer.finish_write(Lsn(0x40));
4349 6 :
4350 6 : // Check page contents on both branches
4351 6 : assert_eq!(
4352 6 : from_utf8(&tline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
4353 6 : "foo at 0x40"
4354 6 : );
4355 6 : assert_eq!(
4356 6 : from_utf8(&newtline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
4357 6 : "bar at 0x40"
4358 6 : );
4359 6 : assert_eq!(
4360 6 : from_utf8(&newtline.get(TEST_KEY_B, Lsn(0x40), &ctx).await?)?,
4361 6 : "foobar at 0x20"
4362 6 : );
4363 6 :
4364 6 : //assert_current_logical_size(&tline, Lsn(0x40));
4365 6 :
4366 6 : Ok(())
4367 6 : }
4368 :
4369 60 : async fn make_some_layers(
4370 60 : tline: &Timeline,
4371 60 : start_lsn: Lsn,
4372 60 : ctx: &RequestContext,
4373 60 : ) -> anyhow::Result<()> {
4374 60 : let mut lsn = start_lsn;
4375 : {
4376 60 : let mut writer = tline.writer().await;
4377 : // Create a relation on the timeline
4378 60 : writer
4379 60 : .put(
4380 60 : *TEST_KEY,
4381 60 : lsn,
4382 60 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4383 60 : ctx,
4384 60 : )
4385 30 : .await?;
4386 60 : writer.finish_write(lsn);
4387 60 : lsn += 0x10;
4388 60 : writer
4389 60 : .put(
4390 60 : *TEST_KEY,
4391 60 : lsn,
4392 60 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4393 60 : ctx,
4394 60 : )
4395 0 : .await?;
4396 60 : writer.finish_write(lsn);
4397 60 : lsn += 0x10;
4398 60 : }
4399 61 : tline.freeze_and_flush().await?;
4400 : {
4401 60 : let mut writer = tline.writer().await;
4402 60 : writer
4403 60 : .put(
4404 60 : *TEST_KEY,
4405 60 : lsn,
4406 60 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4407 60 : ctx,
4408 60 : )
4409 30 : .await?;
4410 60 : writer.finish_write(lsn);
4411 60 : lsn += 0x10;
4412 60 : writer
4413 60 : .put(
4414 60 : *TEST_KEY,
4415 60 : lsn,
4416 60 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4417 60 : ctx,
4418 60 : )
4419 0 : .await?;
4420 60 : writer.finish_write(lsn);
4421 60 : }
4422 61 : tline.freeze_and_flush().await.map_err(|e| e.into())
4423 60 : }
4424 :
4425 : #[tokio::test]
4426 6 : async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
4427 6 : let (tenant, ctx) =
4428 6 : TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
4429 6 : .await?
4430 6 : .load()
4431 24 : .await;
4432 6 : let tline = tenant
4433 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4434 12 : .await?;
4435 18 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4436 6 :
4437 6 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
4438 6 : // FIXME: this doesn't actually remove any layer currently, given how the flushing
4439 6 : // and compaction works. But it does set the 'cutoff' point so that the cross check
4440 6 : // below should fail.
4441 6 : tenant
4442 6 : .gc_iteration(
4443 6 : Some(TIMELINE_ID),
4444 6 : 0x10,
4445 6 : Duration::ZERO,
4446 6 : &CancellationToken::new(),
4447 6 : &ctx,
4448 6 : )
4449 6 : .await?;
4450 6 :
4451 6 : // try to branch at lsn 25, should fail because we already garbage collected the data
4452 6 : match tenant
4453 6 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
4454 6 : .await
4455 6 : {
4456 6 : Ok(_) => panic!("branching should have failed"),
4457 6 : Err(err) => {
4458 6 : let CreateTimelineError::AncestorLsn(err) = err else {
4459 6 : panic!("wrong error type")
4460 6 : };
4461 6 : assert!(err.to_string().contains("invalid branch start lsn"));
4462 6 : assert!(err
4463 6 : .source()
4464 6 : .unwrap()
4465 6 : .to_string()
4466 6 : .contains("we might've already garbage collected needed data"))
4467 6 : }
4468 6 : }
4469 6 :
4470 6 : Ok(())
4471 6 : }
4472 :
4473 : #[tokio::test]
4474 6 : async fn test_prohibit_branch_creation_on_pre_initdb_lsn() -> anyhow::Result<()> {
4475 6 : let (tenant, ctx) =
4476 6 : TenantHarness::create("test_prohibit_branch_creation_on_pre_initdb_lsn")
4477 6 : .await?
4478 6 : .load()
4479 24 : .await;
4480 6 :
4481 6 : let tline = tenant
4482 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x50), DEFAULT_PG_VERSION, &ctx)
4483 12 : .await?;
4484 6 : // try to branch at lsn 0x25, should fail because initdb lsn is 0x50
4485 6 : match tenant
4486 6 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
4487 6 : .await
4488 6 : {
4489 6 : Ok(_) => panic!("branching should have failed"),
4490 6 : Err(err) => {
4491 6 : let CreateTimelineError::AncestorLsn(err) = err else {
4492 6 : panic!("wrong error type");
4493 6 : };
4494 6 : assert!(&err.to_string().contains("invalid branch start lsn"));
4495 6 : assert!(&err
4496 6 : .source()
4497 6 : .unwrap()
4498 6 : .to_string()
4499 6 : .contains("is earlier than latest GC cutoff"));
4500 6 : }
4501 6 : }
4502 6 :
4503 6 : Ok(())
4504 6 : }
4505 :
4506 : /*
4507 : // FIXME: This currently fails to error out. Calling GC doesn't currently
4508 : // remove the old value, we'd need to work a little harder
4509 : #[tokio::test]
4510 : async fn test_prohibit_get_for_garbage_collected_data() -> anyhow::Result<()> {
4511 : let repo =
4512 : RepoHarness::create("test_prohibit_get_for_garbage_collected_data")?
4513 : .load();
4514 :
4515 : let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION)?;
4516 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4517 :
4518 : repo.gc_iteration(Some(TIMELINE_ID), 0x10, Duration::ZERO)?;
4519 : let latest_gc_cutoff_lsn = tline.get_latest_gc_cutoff_lsn();
4520 : assert!(*latest_gc_cutoff_lsn > Lsn(0x25));
4521 : match tline.get(*TEST_KEY, Lsn(0x25)) {
4522 : Ok(_) => panic!("request for page should have failed"),
4523 : Err(err) => assert!(err.to_string().contains("not found at")),
4524 : }
4525 : Ok(())
4526 : }
4527 : */
4528 :
4529 : #[tokio::test]
4530 6 : async fn test_get_branchpoints_from_an_inactive_timeline() -> anyhow::Result<()> {
4531 6 : let (tenant, ctx) =
4532 6 : TenantHarness::create("test_get_branchpoints_from_an_inactive_timeline")
4533 6 : .await?
4534 6 : .load()
4535 24 : .await;
4536 6 : let tline = tenant
4537 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4538 12 : .await?;
4539 18 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4540 6 :
4541 6 : tenant
4542 6 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
4543 6 : .await?;
4544 6 : let newtline = tenant
4545 6 : .get_timeline(NEW_TIMELINE_ID, true)
4546 6 : .expect("Should have a local timeline");
4547 6 :
4548 18 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
4549 6 :
4550 6 : tline.set_broken("test".to_owned());
4551 6 :
4552 6 : tenant
4553 6 : .gc_iteration(
4554 6 : Some(TIMELINE_ID),
4555 6 : 0x10,
4556 6 : Duration::ZERO,
4557 6 : &CancellationToken::new(),
4558 6 : &ctx,
4559 6 : )
4560 6 : .await?;
4561 6 :
4562 6 : // The branchpoints should contain all timelines, even ones marked
4563 6 : // as Broken.
4564 6 : {
4565 6 : let branchpoints = &tline.gc_info.read().unwrap().retain_lsns;
4566 6 : assert_eq!(branchpoints.len(), 1);
4567 6 : assert_eq!(branchpoints[0], (Lsn(0x40), NEW_TIMELINE_ID));
4568 6 : }
4569 6 :
4570 6 : // You can read the key from the child branch even though the parent is
4571 6 : // Broken, as long as you don't need to access data from the parent.
4572 6 : assert_eq!(
4573 12 : newtline.get(*TEST_KEY, Lsn(0x70), &ctx).await?,
4574 6 : test_img(&format!("foo at {}", Lsn(0x70)))
4575 6 : );
4576 6 :
4577 6 : // This needs to traverse to the parent, and fails.
4578 6 : let err = newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await.unwrap_err();
4579 6 : assert!(
4580 6 : err.to_string().starts_with(&format!(
4581 6 : "bad state on timeline {}: Broken",
4582 6 : tline.timeline_id
4583 6 : )),
4584 6 : "{err}"
4585 6 : );
4586 6 :
4587 6 : Ok(())
4588 6 : }
4589 :
4590 : #[tokio::test]
4591 6 : async fn test_retain_data_in_parent_which_is_needed_for_child() -> anyhow::Result<()> {
4592 6 : let (tenant, ctx) =
4593 6 : TenantHarness::create("test_retain_data_in_parent_which_is_needed_for_child")
4594 6 : .await?
4595 6 : .load()
4596 24 : .await;
4597 6 : let tline = tenant
4598 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4599 12 : .await?;
4600 19 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4601 6 :
4602 6 : tenant
4603 6 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
4604 6 : .await?;
4605 6 : let newtline = tenant
4606 6 : .get_timeline(NEW_TIMELINE_ID, true)
4607 6 : .expect("Should have a local timeline");
4608 6 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
4609 6 : tenant
4610 6 : .gc_iteration(
4611 6 : Some(TIMELINE_ID),
4612 6 : 0x10,
4613 6 : Duration::ZERO,
4614 6 : &CancellationToken::new(),
4615 6 : &ctx,
4616 6 : )
4617 6 : .await?;
4618 12 : assert!(newtline.get(*TEST_KEY, Lsn(0x25), &ctx).await.is_ok());
4619 6 :
4620 6 : Ok(())
4621 6 : }
4622 : #[tokio::test]
4623 6 : async fn test_parent_keeps_data_forever_after_branching() -> anyhow::Result<()> {
4624 6 : let (tenant, ctx) = TenantHarness::create("test_parent_keeps_data_forever_after_branching")
4625 6 : .await?
4626 6 : .load()
4627 24 : .await;
4628 6 : let tline = tenant
4629 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4630 12 : .await?;
4631 18 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4632 6 :
4633 6 : tenant
4634 6 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
4635 6 : .await?;
4636 6 : let newtline = tenant
4637 6 : .get_timeline(NEW_TIMELINE_ID, true)
4638 6 : .expect("Should have a local timeline");
4639 6 :
4640 18 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
4641 6 :
4642 6 : // run gc on parent
4643 6 : tenant
4644 6 : .gc_iteration(
4645 6 : Some(TIMELINE_ID),
4646 6 : 0x10,
4647 6 : Duration::ZERO,
4648 6 : &CancellationToken::new(),
4649 6 : &ctx,
4650 6 : )
4651 6 : .await?;
4652 6 :
4653 6 : // Check that the data is still accessible on the branch.
4654 6 : assert_eq!(
4655 21 : newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
4656 6 : test_img(&format!("foo at {}", Lsn(0x40)))
4657 6 : );
4658 6 :
4659 6 : Ok(())
4660 6 : }
4661 :
4662 : #[tokio::test]
4663 6 : async fn timeline_load() -> anyhow::Result<()> {
4664 6 : const TEST_NAME: &str = "timeline_load";
4665 6 : let harness = TenantHarness::create(TEST_NAME).await?;
4666 6 : {
4667 24 : let (tenant, ctx) = harness.load().await;
4668 6 : let tline = tenant
4669 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x7000), DEFAULT_PG_VERSION, &ctx)
4670 12 : .await?;
4671 18 : make_some_layers(tline.as_ref(), Lsn(0x8000), &ctx).await?;
4672 6 : // so that all uploads finish & we can call harness.load() below again
4673 6 : tenant
4674 6 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
4675 6 : .instrument(harness.span())
4676 6 : .await
4677 6 : .ok()
4678 6 : .unwrap();
4679 6 : }
4680 6 :
4681 23 : let (tenant, _ctx) = harness.load().await;
4682 6 : tenant
4683 6 : .get_timeline(TIMELINE_ID, true)
4684 6 : .expect("cannot load timeline");
4685 6 :
4686 6 : Ok(())
4687 6 : }
4688 :
4689 : #[tokio::test]
4690 6 : async fn timeline_load_with_ancestor() -> anyhow::Result<()> {
4691 6 : const TEST_NAME: &str = "timeline_load_with_ancestor";
4692 6 : let harness = TenantHarness::create(TEST_NAME).await?;
4693 6 : // create two timelines
4694 6 : {
4695 24 : let (tenant, ctx) = harness.load().await;
4696 6 : let tline = tenant
4697 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4698 12 : .await?;
4699 6 :
4700 18 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4701 6 :
4702 6 : let child_tline = tenant
4703 6 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
4704 6 : .await?;
4705 6 : child_tline.set_state(TimelineState::Active);
4706 6 :
4707 6 : let newtline = tenant
4708 6 : .get_timeline(NEW_TIMELINE_ID, true)
4709 6 : .expect("Should have a local timeline");
4710 6 :
4711 18 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
4712 6 :
4713 6 : // so that all uploads finish & we can call harness.load() below again
4714 6 : tenant
4715 6 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
4716 6 : .instrument(harness.span())
4717 6 : .await
4718 6 : .ok()
4719 6 : .unwrap();
4720 6 : }
4721 6 :
4722 6 : // check that both of them are initially unloaded
4723 35 : let (tenant, _ctx) = harness.load().await;
4724 6 :
4725 6 : // check that both, child and ancestor are loaded
4726 6 : let _child_tline = tenant
4727 6 : .get_timeline(NEW_TIMELINE_ID, true)
4728 6 : .expect("cannot get child timeline loaded");
4729 6 :
4730 6 : let _ancestor_tline = tenant
4731 6 : .get_timeline(TIMELINE_ID, true)
4732 6 : .expect("cannot get ancestor timeline loaded");
4733 6 :
4734 6 : Ok(())
4735 6 : }
4736 :
4737 : #[tokio::test]
4738 6 : async fn delta_layer_dumping() -> anyhow::Result<()> {
4739 6 : use storage_layer::AsLayerDesc;
4740 6 : let (tenant, ctx) = TenantHarness::create("test_layer_dumping")
4741 6 : .await?
4742 6 : .load()
4743 24 : .await;
4744 6 : let tline = tenant
4745 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4746 12 : .await?;
4747 19 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4748 6 :
4749 6 : let layer_map = tline.layers.read().await;
4750 6 : let level0_deltas = layer_map
4751 6 : .layer_map()?
4752 6 : .level0_deltas()
4753 6 : .iter()
4754 12 : .map(|desc| layer_map.get_from_desc(desc))
4755 6 : .collect::<Vec<_>>();
4756 6 :
4757 6 : assert!(!level0_deltas.is_empty());
4758 6 :
4759 18 : for delta in level0_deltas {
4760 6 : // Ensure we are dumping a delta layer here
4761 12 : assert!(delta.layer_desc().is_delta);
4762 24 : delta.dump(true, &ctx).await.unwrap();
4763 6 : }
4764 6 :
4765 6 : Ok(())
4766 6 : }
4767 :
4768 : #[tokio::test]
4769 6 : async fn test_images() -> anyhow::Result<()> {
4770 24 : let (tenant, ctx) = TenantHarness::create("test_images").await?.load().await;
4771 6 : let tline = tenant
4772 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
4773 12 : .await?;
4774 6 :
4775 6 : let mut writer = tline.writer().await;
4776 6 : writer
4777 6 : .put(
4778 6 : *TEST_KEY,
4779 6 : Lsn(0x10),
4780 6 : &Value::Image(test_img("foo at 0x10")),
4781 6 : &ctx,
4782 6 : )
4783 6 : .await?;
4784 6 : writer.finish_write(Lsn(0x10));
4785 6 : drop(writer);
4786 6 :
4787 7 : tline.freeze_and_flush().await?;
4788 6 : tline
4789 6 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
4790 6 : .await?;
4791 6 :
4792 6 : let mut writer = tline.writer().await;
4793 6 : writer
4794 6 : .put(
4795 6 : *TEST_KEY,
4796 6 : Lsn(0x20),
4797 6 : &Value::Image(test_img("foo at 0x20")),
4798 6 : &ctx,
4799 6 : )
4800 6 : .await?;
4801 6 : writer.finish_write(Lsn(0x20));
4802 6 : drop(writer);
4803 6 :
4804 7 : tline.freeze_and_flush().await?;
4805 6 : tline
4806 6 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
4807 6 : .await?;
4808 6 :
4809 6 : let mut writer = tline.writer().await;
4810 6 : writer
4811 6 : .put(
4812 6 : *TEST_KEY,
4813 6 : Lsn(0x30),
4814 6 : &Value::Image(test_img("foo at 0x30")),
4815 6 : &ctx,
4816 6 : )
4817 6 : .await?;
4818 6 : writer.finish_write(Lsn(0x30));
4819 6 : drop(writer);
4820 6 :
4821 6 : tline.freeze_and_flush().await?;
4822 6 : tline
4823 6 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
4824 6 : .await?;
4825 6 :
4826 6 : let mut writer = tline.writer().await;
4827 6 : writer
4828 6 : .put(
4829 6 : *TEST_KEY,
4830 6 : Lsn(0x40),
4831 6 : &Value::Image(test_img("foo at 0x40")),
4832 6 : &ctx,
4833 6 : )
4834 6 : .await?;
4835 6 : writer.finish_write(Lsn(0x40));
4836 6 : drop(writer);
4837 6 :
4838 6 : tline.freeze_and_flush().await?;
4839 6 : tline
4840 6 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
4841 6 : .await?;
4842 6 :
4843 6 : assert_eq!(
4844 12 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
4845 6 : test_img("foo at 0x10")
4846 6 : );
4847 6 : assert_eq!(
4848 12 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
4849 6 : test_img("foo at 0x10")
4850 6 : );
4851 6 : assert_eq!(
4852 6 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
4853 6 : test_img("foo at 0x20")
4854 6 : );
4855 6 : assert_eq!(
4856 12 : tline.get(*TEST_KEY, Lsn(0x30), &ctx).await?,
4857 6 : test_img("foo at 0x30")
4858 6 : );
4859 6 : assert_eq!(
4860 12 : tline.get(*TEST_KEY, Lsn(0x40), &ctx).await?,
4861 6 : test_img("foo at 0x40")
4862 6 : );
4863 6 :
4864 6 : Ok(())
4865 6 : }
4866 :
4867 12 : async fn bulk_insert_compact_gc(
4868 12 : tenant: &Tenant,
4869 12 : timeline: &Arc<Timeline>,
4870 12 : ctx: &RequestContext,
4871 12 : lsn: Lsn,
4872 12 : repeat: usize,
4873 12 : key_count: usize,
4874 12 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
4875 12 : let compact = true;
4876 226438 : bulk_insert_maybe_compact_gc(tenant, timeline, ctx, lsn, repeat, key_count, compact).await
4877 12 : }
4878 :
4879 24 : async fn bulk_insert_maybe_compact_gc(
4880 24 : tenant: &Tenant,
4881 24 : timeline: &Arc<Timeline>,
4882 24 : ctx: &RequestContext,
4883 24 : mut lsn: Lsn,
4884 24 : repeat: usize,
4885 24 : key_count: usize,
4886 24 : compact: bool,
4887 24 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
4888 24 : let mut inserted: HashMap<Key, BTreeSet<Lsn>> = Default::default();
4889 24 :
4890 24 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
4891 24 : let mut blknum = 0;
4892 24 :
4893 24 : // Enforce that key range is monotonously increasing
4894 24 : let mut keyspace = KeySpaceAccum::new();
4895 24 :
4896 24 : let cancel = CancellationToken::new();
4897 24 :
4898 24 : for _ in 0..repeat {
4899 1200 : for _ in 0..key_count {
4900 12000000 : test_key.field6 = blknum;
4901 12000000 : let mut writer = timeline.writer().await;
4902 12000000 : writer
4903 12000000 : .put(
4904 12000000 : test_key,
4905 12000000 : lsn,
4906 12000000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
4907 12000000 : ctx,
4908 12000000 : )
4909 7788 : .await?;
4910 12000000 : inserted.entry(test_key).or_default().insert(lsn);
4911 12000000 : writer.finish_write(lsn);
4912 12000000 : drop(writer);
4913 12000000 :
4914 12000000 : keyspace.add_key(test_key);
4915 12000000 :
4916 12000000 : lsn = Lsn(lsn.0 + 0x10);
4917 12000000 : blknum += 1;
4918 : }
4919 :
4920 1200 : timeline.freeze_and_flush().await?;
4921 1200 : if compact {
4922 : // this requires timeline to be &Arc<Timeline>
4923 129540 : timeline.compact(&cancel, EnumSet::empty(), ctx).await?;
4924 600 : }
4925 :
4926 : // this doesn't really need to use the timeline_id target, but it is closer to what it
4927 : // originally was.
4928 1200 : let res = tenant
4929 1200 : .gc_iteration(Some(timeline.timeline_id), 0, Duration::ZERO, &cancel, ctx)
4930 1196 : .await?;
4931 :
4932 1200 : assert_eq!(res.layers_removed, 0, "this never removes anything");
4933 : }
4934 :
4935 24 : Ok(inserted)
4936 24 : }
4937 :
4938 : //
4939 : // Insert 1000 key-value pairs with increasing keys, flush, compact, GC.
4940 : // Repeat 50 times.
4941 : //
4942 : #[tokio::test]
4943 6 : async fn test_bulk_insert() -> anyhow::Result<()> {
4944 6 : let harness = TenantHarness::create("test_bulk_insert").await?;
4945 21 : let (tenant, ctx) = harness.load().await;
4946 6 : let tline = tenant
4947 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
4948 12 : .await?;
4949 6 :
4950 6 : let lsn = Lsn(0x10);
4951 113219 : bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
4952 6 :
4953 6 : Ok(())
4954 6 : }
4955 :
4956 : // Test the vectored get real implementation against a simple sequential implementation.
4957 : //
4958 : // The test generates a keyspace by repeatedly flushing the in-memory layer and compacting.
4959 : // Projected to 2D the key space looks like below. Lsn grows upwards on the Y axis and keys
4960 : // grow to the right on the X axis.
4961 : // [Delta]
4962 : // [Delta]
4963 : // [Delta]
4964 : // [Delta]
4965 : // ------------ Image ---------------
4966 : //
4967 : // After layer generation we pick the ranges to query as follows:
4968 : // 1. The beginning of each delta layer
4969 : // 2. At the seam between two adjacent delta layers
4970 : //
4971 : // There's one major downside to this test: delta layers only contains images,
4972 : // so the search can stop at the first delta layer and doesn't traverse any deeper.
4973 : #[tokio::test]
4974 6 : async fn test_get_vectored() -> anyhow::Result<()> {
4975 6 : let harness = TenantHarness::create("test_get_vectored").await?;
4976 24 : let (tenant, ctx) = harness.load().await;
4977 6 : let tline = tenant
4978 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
4979 12 : .await?;
4980 6 :
4981 6 : let lsn = Lsn(0x10);
4982 113219 : let inserted = bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
4983 6 :
4984 6 : let guard = tline.layers.read().await;
4985 6 : let lm = guard.layer_map()?;
4986 6 :
4987 6 : lm.dump(true, &ctx).await?;
4988 6 :
4989 6 : let mut reads = Vec::new();
4990 6 : let mut prev = None;
4991 36 : lm.iter_historic_layers().for_each(|desc| {
4992 36 : if !desc.is_delta() {
4993 6 : prev = Some(desc.clone());
4994 6 : return;
4995 30 : }
4996 30 :
4997 30 : let start = desc.key_range.start;
4998 30 : let end = desc
4999 30 : .key_range
5000 30 : .start
5001 30 : .add(Timeline::MAX_GET_VECTORED_KEYS.try_into().unwrap());
5002 30 : reads.push(KeySpace {
5003 30 : ranges: vec![start..end],
5004 30 : });
5005 6 :
5006 30 : if let Some(prev) = &prev {
5007 30 : if !prev.is_delta() {
5008 30 : return;
5009 6 : }
5010 0 :
5011 0 : let first_range = Key {
5012 0 : field6: prev.key_range.end.field6 - 4,
5013 0 : ..prev.key_range.end
5014 0 : }..prev.key_range.end;
5015 0 :
5016 0 : let second_range = desc.key_range.start..Key {
5017 0 : field6: desc.key_range.start.field6 + 4,
5018 0 : ..desc.key_range.start
5019 0 : };
5020 0 :
5021 0 : reads.push(KeySpace {
5022 0 : ranges: vec![first_range, second_range],
5023 0 : });
5024 6 : };
5025 6 :
5026 6 : prev = Some(desc.clone());
5027 36 : });
5028 6 :
5029 6 : drop(guard);
5030 6 :
5031 6 : // Pick a big LSN such that we query over all the changes.
5032 6 : let reads_lsn = Lsn(u64::MAX - 1);
5033 6 :
5034 36 : for read in reads {
5035 30 : info!("Doing vectored read on {:?}", read);
5036 6 :
5037 30 : let vectored_res = tline
5038 30 : .get_vectored_impl(
5039 30 : read.clone(),
5040 30 : reads_lsn,
5041 30 : &mut ValuesReconstructState::new(),
5042 30 : &ctx,
5043 30 : )
5044 75 : .await;
5045 6 :
5046 30 : let mut expected_lsns: HashMap<Key, Lsn> = Default::default();
5047 30 : let mut expect_missing = false;
5048 30 : let mut key = read.start().unwrap();
5049 990 : while key != read.end().unwrap() {
5050 960 : if let Some(lsns) = inserted.get(&key) {
5051 960 : let expected_lsn = lsns.iter().rfind(|lsn| **lsn <= reads_lsn);
5052 960 : match expected_lsn {
5053 960 : Some(lsn) => {
5054 960 : expected_lsns.insert(key, *lsn);
5055 960 : }
5056 6 : None => {
5057 6 : expect_missing = true;
5058 0 : break;
5059 6 : }
5060 6 : }
5061 6 : } else {
5062 6 : expect_missing = true;
5063 0 : break;
5064 6 : }
5065 6 :
5066 960 : key = key.next();
5067 6 : }
5068 6 :
5069 30 : if expect_missing {
5070 6 : assert!(matches!(vectored_res, Err(GetVectoredError::MissingKey(_))));
5071 6 : } else {
5072 960 : for (key, image) in vectored_res? {
5073 960 : let expected_lsn = expected_lsns.get(&key).expect("determined above");
5074 960 : let expected_image = test_img(&format!("{} at {}", key.field6, expected_lsn));
5075 960 : assert_eq!(image?, expected_image);
5076 6 : }
5077 6 : }
5078 6 : }
5079 6 :
5080 6 : Ok(())
5081 6 : }
5082 :
5083 : #[tokio::test]
5084 6 : async fn test_get_vectored_aux_files() -> anyhow::Result<()> {
5085 6 : let harness = TenantHarness::create("test_get_vectored_aux_files").await?;
5086 6 :
5087 24 : let (tenant, ctx) = harness.load().await;
5088 6 : let tline = tenant
5089 6 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
5090 6 : .await?;
5091 6 : let tline = tline.raw_timeline().unwrap();
5092 6 :
5093 6 : let mut modification = tline.begin_modification(Lsn(0x1000));
5094 6 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
5095 6 : modification.set_lsn(Lsn(0x1008))?;
5096 6 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
5097 6 : modification.commit(&ctx).await?;
5098 6 :
5099 6 : let child_timeline_id = TimelineId::generate();
5100 6 : tenant
5101 6 : .branch_timeline_test(
5102 6 : tline,
5103 6 : child_timeline_id,
5104 6 : Some(tline.get_last_record_lsn()),
5105 6 : &ctx,
5106 6 : )
5107 6 : .await?;
5108 6 :
5109 6 : let child_timeline = tenant
5110 6 : .get_timeline(child_timeline_id, true)
5111 6 : .expect("Should have the branched timeline");
5112 6 :
5113 6 : let aux_keyspace = KeySpace {
5114 6 : ranges: vec![NON_INHERITED_RANGE],
5115 6 : };
5116 6 : let read_lsn = child_timeline.get_last_record_lsn();
5117 6 :
5118 6 : let vectored_res = child_timeline
5119 6 : .get_vectored_impl(
5120 6 : aux_keyspace.clone(),
5121 6 : read_lsn,
5122 6 : &mut ValuesReconstructState::new(),
5123 6 : &ctx,
5124 6 : )
5125 6 : .await;
5126 6 :
5127 6 : let images = vectored_res?;
5128 6 : assert!(images.is_empty());
5129 6 : Ok(())
5130 6 : }
5131 :
5132 : // Test that vectored get handles layer gaps correctly
5133 : // by advancing into the next ancestor timeline if required.
5134 : //
5135 : // The test generates timelines that look like the diagram below.
5136 : // We leave a gap in one of the L1 layers at `gap_at_key` (`/` in the diagram).
5137 : // The reconstruct data for that key lies in the ancestor timeline (`X` in the diagram).
5138 : //
5139 : // ```
5140 : //-------------------------------+
5141 : // ... |
5142 : // [ L1 ] |
5143 : // [ / L1 ] | Child Timeline
5144 : // ... |
5145 : // ------------------------------+
5146 : // [ X L1 ] | Parent Timeline
5147 : // ------------------------------+
5148 : // ```
5149 : #[tokio::test]
5150 6 : async fn test_get_vectored_key_gap() -> anyhow::Result<()> {
5151 6 : let tenant_conf = TenantConf {
5152 6 : // Make compaction deterministic
5153 6 : gc_period: Duration::ZERO,
5154 6 : compaction_period: Duration::ZERO,
5155 6 : // Encourage creation of L1 layers
5156 6 : checkpoint_distance: 16 * 1024,
5157 6 : compaction_target_size: 8 * 1024,
5158 6 : ..TenantConf::default()
5159 6 : };
5160 6 :
5161 6 : let harness = TenantHarness::create_custom(
5162 6 : "test_get_vectored_key_gap",
5163 6 : tenant_conf,
5164 6 : TenantId::generate(),
5165 6 : ShardIdentity::unsharded(),
5166 6 : Generation::new(0xdeadbeef),
5167 6 : )
5168 6 : .await?;
5169 24 : let (tenant, ctx) = harness.load().await;
5170 6 :
5171 6 : let mut current_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5172 6 : let gap_at_key = current_key.add(100);
5173 6 : let mut current_lsn = Lsn(0x10);
5174 6 :
5175 6 : const KEY_COUNT: usize = 10_000;
5176 6 :
5177 6 : let timeline_id = TimelineId::generate();
5178 6 : let current_timeline = tenant
5179 6 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
5180 12 : .await?;
5181 6 :
5182 6 : current_lsn += 0x100;
5183 6 :
5184 6 : let mut writer = current_timeline.writer().await;
5185 6 : writer
5186 6 : .put(
5187 6 : gap_at_key,
5188 6 : current_lsn,
5189 6 : &Value::Image(test_img(&format!("{} at {}", gap_at_key, current_lsn))),
5190 6 : &ctx,
5191 6 : )
5192 6 : .await?;
5193 6 : writer.finish_write(current_lsn);
5194 6 : drop(writer);
5195 6 :
5196 6 : let mut latest_lsns = HashMap::new();
5197 6 : latest_lsns.insert(gap_at_key, current_lsn);
5198 6 :
5199 6 : current_timeline.freeze_and_flush().await?;
5200 6 :
5201 6 : let child_timeline_id = TimelineId::generate();
5202 6 :
5203 6 : tenant
5204 6 : .branch_timeline_test(
5205 6 : ¤t_timeline,
5206 6 : child_timeline_id,
5207 6 : Some(current_lsn),
5208 6 : &ctx,
5209 6 : )
5210 6 : .await?;
5211 6 : let child_timeline = tenant
5212 6 : .get_timeline(child_timeline_id, true)
5213 6 : .expect("Should have the branched timeline");
5214 6 :
5215 60006 : for i in 0..KEY_COUNT {
5216 60000 : if current_key == gap_at_key {
5217 6 : current_key = current_key.next();
5218 6 : continue;
5219 59994 : }
5220 59994 :
5221 59994 : current_lsn += 0x10;
5222 6 :
5223 59994 : let mut writer = child_timeline.writer().await;
5224 59994 : writer
5225 59994 : .put(
5226 59994 : current_key,
5227 59994 : current_lsn,
5228 59994 : &Value::Image(test_img(&format!("{} at {}", current_key, current_lsn))),
5229 59994 : &ctx,
5230 59994 : )
5231 220 : .await?;
5232 59994 : writer.finish_write(current_lsn);
5233 59994 : drop(writer);
5234 59994 :
5235 59994 : latest_lsns.insert(current_key, current_lsn);
5236 59994 : current_key = current_key.next();
5237 59994 :
5238 59994 : // Flush every now and then to encourage layer file creation.
5239 59994 : if i % 500 == 0 {
5240 128 : child_timeline.freeze_and_flush().await?;
5241 59874 : }
5242 6 : }
5243 6 :
5244 6 : child_timeline.freeze_and_flush().await?;
5245 6 : let mut flags = EnumSet::new();
5246 6 : flags.insert(CompactFlags::ForceRepartition);
5247 6 : child_timeline
5248 6 : .compact(&CancellationToken::new(), flags, &ctx)
5249 6981 : .await?;
5250 6 :
5251 6 : let key_near_end = {
5252 6 : let mut tmp = current_key;
5253 6 : tmp.field6 -= 10;
5254 6 : tmp
5255 6 : };
5256 6 :
5257 6 : let key_near_gap = {
5258 6 : let mut tmp = gap_at_key;
5259 6 : tmp.field6 -= 10;
5260 6 : tmp
5261 6 : };
5262 6 :
5263 6 : let read = KeySpace {
5264 6 : ranges: vec![key_near_gap..gap_at_key.next(), key_near_end..current_key],
5265 6 : };
5266 6 : let results = child_timeline
5267 6 : .get_vectored_impl(
5268 6 : read.clone(),
5269 6 : current_lsn,
5270 6 : &mut ValuesReconstructState::new(),
5271 6 : &ctx,
5272 6 : )
5273 45 : .await?;
5274 6 :
5275 132 : for (key, img_res) in results {
5276 126 : let expected = test_img(&format!("{} at {}", key, latest_lsns[&key]));
5277 126 : assert_eq!(img_res?, expected);
5278 6 : }
5279 6 :
5280 6 : Ok(())
5281 6 : }
5282 :
5283 : // Test that vectored get descends into ancestor timelines correctly and
5284 : // does not return an image that's newer than requested.
5285 : //
5286 : // The diagram below ilustrates an interesting case. We have a parent timeline
5287 : // (top of the Lsn range) and a child timeline. The request key cannot be reconstructed
5288 : // from the child timeline, so the parent timeline must be visited. When advacing into
5289 : // the child timeline, the read path needs to remember what the requested Lsn was in
5290 : // order to avoid returning an image that's too new. The test below constructs such
5291 : // a timeline setup and does a few queries around the Lsn of each page image.
5292 : // ```
5293 : // LSN
5294 : // ^
5295 : // |
5296 : // |
5297 : // 500 | --------------------------------------> branch point
5298 : // 400 | X
5299 : // 300 | X
5300 : // 200 | --------------------------------------> requested lsn
5301 : // 100 | X
5302 : // |---------------------------------------> Key
5303 : // |
5304 : // ------> requested key
5305 : //
5306 : // Legend:
5307 : // * X - page images
5308 : // ```
5309 : #[tokio::test]
5310 6 : async fn test_get_vectored_ancestor_descent() -> anyhow::Result<()> {
5311 6 : let harness = TenantHarness::create("test_get_vectored_on_lsn_axis").await?;
5312 24 : let (tenant, ctx) = harness.load().await;
5313 6 :
5314 6 : let start_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5315 6 : let end_key = start_key.add(1000);
5316 6 : let child_gap_at_key = start_key.add(500);
5317 6 : let mut parent_gap_lsns: BTreeMap<Lsn, String> = BTreeMap::new();
5318 6 :
5319 6 : let mut current_lsn = Lsn(0x10);
5320 6 :
5321 6 : let timeline_id = TimelineId::generate();
5322 6 : let parent_timeline = tenant
5323 6 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
5324 12 : .await?;
5325 6 :
5326 6 : current_lsn += 0x100;
5327 6 :
5328 24 : for _ in 0..3 {
5329 18 : let mut key = start_key;
5330 18018 : while key < end_key {
5331 18000 : current_lsn += 0x10;
5332 18000 :
5333 18000 : let image_value = format!("{} at {}", child_gap_at_key, current_lsn);
5334 6 :
5335 18000 : let mut writer = parent_timeline.writer().await;
5336 18000 : writer
5337 18000 : .put(
5338 18000 : key,
5339 18000 : current_lsn,
5340 18000 : &Value::Image(test_img(&image_value)),
5341 18000 : &ctx,
5342 18000 : )
5343 18 : .await?;
5344 18000 : writer.finish_write(current_lsn);
5345 18000 :
5346 18000 : if key == child_gap_at_key {
5347 18 : parent_gap_lsns.insert(current_lsn, image_value);
5348 17982 : }
5349 6 :
5350 18000 : key = key.next();
5351 6 : }
5352 6 :
5353 18 : parent_timeline.freeze_and_flush().await?;
5354 6 : }
5355 6 :
5356 6 : let child_timeline_id = TimelineId::generate();
5357 6 :
5358 6 : let child_timeline = tenant
5359 6 : .branch_timeline_test(&parent_timeline, child_timeline_id, Some(current_lsn), &ctx)
5360 6 : .await?;
5361 6 :
5362 6 : let mut key = start_key;
5363 6006 : while key < end_key {
5364 6000 : if key == child_gap_at_key {
5365 6 : key = key.next();
5366 6 : continue;
5367 5994 : }
5368 5994 :
5369 5994 : current_lsn += 0x10;
5370 6 :
5371 5994 : let mut writer = child_timeline.writer().await;
5372 5994 : writer
5373 5994 : .put(
5374 5994 : key,
5375 5994 : current_lsn,
5376 5994 : &Value::Image(test_img(&format!("{} at {}", key, current_lsn))),
5377 5994 : &ctx,
5378 5994 : )
5379 51 : .await?;
5380 5994 : writer.finish_write(current_lsn);
5381 5994 :
5382 5994 : key = key.next();
5383 6 : }
5384 6 :
5385 6 : child_timeline.freeze_and_flush().await?;
5386 6 :
5387 6 : let lsn_offsets: [i64; 5] = [-10, -1, 0, 1, 10];
5388 6 : let mut query_lsns = Vec::new();
5389 18 : for image_lsn in parent_gap_lsns.keys().rev() {
5390 108 : for offset in lsn_offsets {
5391 90 : query_lsns.push(Lsn(image_lsn
5392 90 : .0
5393 90 : .checked_add_signed(offset)
5394 90 : .expect("Shouldn't overflow")));
5395 90 : }
5396 6 : }
5397 6 :
5398 96 : for query_lsn in query_lsns {
5399 90 : let results = child_timeline
5400 90 : .get_vectored_impl(
5401 90 : KeySpace {
5402 90 : ranges: vec![child_gap_at_key..child_gap_at_key.next()],
5403 90 : },
5404 90 : query_lsn,
5405 90 : &mut ValuesReconstructState::new(),
5406 90 : &ctx,
5407 90 : )
5408 87 : .await;
5409 6 :
5410 90 : let expected_item = parent_gap_lsns
5411 90 : .iter()
5412 90 : .rev()
5413 204 : .find(|(lsn, _)| **lsn <= query_lsn);
5414 90 :
5415 90 : info!(
5416 6 : "Doing vectored read at LSN {}. Expecting image to be: {:?}",
5417 6 : query_lsn, expected_item
5418 6 : );
5419 6 :
5420 90 : match expected_item {
5421 78 : Some((_, img_value)) => {
5422 78 : let key_results = results.expect("No vectored get error expected");
5423 78 : let key_result = &key_results[&child_gap_at_key];
5424 78 : let returned_img = key_result
5425 78 : .as_ref()
5426 78 : .expect("No page reconstruct error expected");
5427 78 :
5428 78 : info!(
5429 6 : "Vectored read at LSN {} returned image {}",
5430 0 : query_lsn,
5431 0 : std::str::from_utf8(returned_img)?
5432 6 : );
5433 78 : assert_eq!(*returned_img, test_img(img_value));
5434 6 : }
5435 6 : None => {
5436 12 : assert!(matches!(results, Err(GetVectoredError::MissingKey(_))));
5437 6 : }
5438 6 : }
5439 6 : }
5440 6 :
5441 6 : Ok(())
5442 6 : }
5443 :
5444 : #[tokio::test]
5445 6 : async fn test_random_updates() -> anyhow::Result<()> {
5446 6 : let names_algorithms = [
5447 6 : ("test_random_updates_legacy", CompactionAlgorithm::Legacy),
5448 6 : ("test_random_updates_tiered", CompactionAlgorithm::Tiered),
5449 6 : ];
5450 18 : for (name, algorithm) in names_algorithms {
5451 289192 : test_random_updates_algorithm(name, algorithm).await?;
5452 6 : }
5453 6 : Ok(())
5454 6 : }
5455 :
5456 12 : async fn test_random_updates_algorithm(
5457 12 : name: &'static str,
5458 12 : compaction_algorithm: CompactionAlgorithm,
5459 12 : ) -> anyhow::Result<()> {
5460 12 : let mut harness = TenantHarness::create(name).await?;
5461 12 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
5462 12 : kind: compaction_algorithm,
5463 12 : };
5464 48 : let (tenant, ctx) = harness.load().await;
5465 12 : let tline = tenant
5466 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5467 21 : .await?;
5468 :
5469 : const NUM_KEYS: usize = 1000;
5470 12 : let cancel = CancellationToken::new();
5471 12 :
5472 12 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5473 12 : let mut test_key_end = test_key;
5474 12 : test_key_end.field6 = NUM_KEYS as u32;
5475 12 : tline.add_extra_test_dense_keyspace(KeySpace::single(test_key..test_key_end));
5476 12 :
5477 12 : let mut keyspace = KeySpaceAccum::new();
5478 12 :
5479 12 : // Track when each page was last modified. Used to assert that
5480 12 : // a read sees the latest page version.
5481 12 : let mut updated = [Lsn(0); NUM_KEYS];
5482 12 :
5483 12 : let mut lsn = Lsn(0x10);
5484 : #[allow(clippy::needless_range_loop)]
5485 12012 : for blknum in 0..NUM_KEYS {
5486 12000 : lsn = Lsn(lsn.0 + 0x10);
5487 12000 : test_key.field6 = blknum as u32;
5488 12000 : let mut writer = tline.writer().await;
5489 12000 : writer
5490 12000 : .put(
5491 12000 : test_key,
5492 12000 : lsn,
5493 12000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5494 12000 : &ctx,
5495 12000 : )
5496 12 : .await?;
5497 12000 : writer.finish_write(lsn);
5498 12000 : updated[blknum] = lsn;
5499 12000 : drop(writer);
5500 12000 :
5501 12000 : keyspace.add_key(test_key);
5502 : }
5503 :
5504 612 : for _ in 0..50 {
5505 600600 : for _ in 0..NUM_KEYS {
5506 600000 : lsn = Lsn(lsn.0 + 0x10);
5507 600000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
5508 600000 : test_key.field6 = blknum as u32;
5509 600000 : let mut writer = tline.writer().await;
5510 600000 : writer
5511 600000 : .put(
5512 600000 : test_key,
5513 600000 : lsn,
5514 600000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5515 600000 : &ctx,
5516 600000 : )
5517 594 : .await?;
5518 600000 : writer.finish_write(lsn);
5519 600000 : drop(writer);
5520 600000 : updated[blknum] = lsn;
5521 : }
5522 :
5523 : // Read all the blocks
5524 600000 : for (blknum, last_lsn) in updated.iter().enumerate() {
5525 600000 : test_key.field6 = blknum as u32;
5526 600000 : assert_eq!(
5527 600000 : tline.get(test_key, lsn, &ctx).await?,
5528 600000 : test_img(&format!("{} at {}", blknum, last_lsn))
5529 : );
5530 : }
5531 :
5532 : // Perform a cycle of flush, and GC
5533 607 : tline.freeze_and_flush().await?;
5534 600 : tenant
5535 600 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
5536 595 : .await?;
5537 : }
5538 :
5539 12 : Ok(())
5540 12 : }
5541 :
5542 : #[tokio::test]
5543 6 : async fn test_traverse_branches() -> anyhow::Result<()> {
5544 6 : let (tenant, ctx) = TenantHarness::create("test_traverse_branches")
5545 6 : .await?
5546 6 : .load()
5547 24 : .await;
5548 6 : let mut tline = tenant
5549 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5550 12 : .await?;
5551 6 :
5552 6 : const NUM_KEYS: usize = 1000;
5553 6 :
5554 6 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5555 6 :
5556 6 : let mut keyspace = KeySpaceAccum::new();
5557 6 :
5558 6 : let cancel = CancellationToken::new();
5559 6 :
5560 6 : // Track when each page was last modified. Used to assert that
5561 6 : // a read sees the latest page version.
5562 6 : let mut updated = [Lsn(0); NUM_KEYS];
5563 6 :
5564 6 : let mut lsn = Lsn(0x10);
5565 6 : #[allow(clippy::needless_range_loop)]
5566 6006 : for blknum in 0..NUM_KEYS {
5567 6000 : lsn = Lsn(lsn.0 + 0x10);
5568 6000 : test_key.field6 = blknum as u32;
5569 6000 : let mut writer = tline.writer().await;
5570 6000 : writer
5571 6000 : .put(
5572 6000 : test_key,
5573 6000 : lsn,
5574 6000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5575 6000 : &ctx,
5576 6000 : )
5577 6 : .await?;
5578 6000 : writer.finish_write(lsn);
5579 6000 : updated[blknum] = lsn;
5580 6000 : drop(writer);
5581 6000 :
5582 6000 : keyspace.add_key(test_key);
5583 6 : }
5584 6 :
5585 306 : for _ in 0..50 {
5586 300 : let new_tline_id = TimelineId::generate();
5587 300 : tenant
5588 300 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
5589 6 : .await?;
5590 300 : tline = tenant
5591 300 : .get_timeline(new_tline_id, true)
5592 300 : .expect("Should have the branched timeline");
5593 6 :
5594 300300 : for _ in 0..NUM_KEYS {
5595 300000 : lsn = Lsn(lsn.0 + 0x10);
5596 300000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
5597 300000 : test_key.field6 = blknum as u32;
5598 300000 : let mut writer = tline.writer().await;
5599 300000 : writer
5600 300000 : .put(
5601 300000 : test_key,
5602 300000 : lsn,
5603 300000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5604 300000 : &ctx,
5605 300000 : )
5606 2540 : .await?;
5607 300000 : println!("updating {} at {}", blknum, lsn);
5608 300000 : writer.finish_write(lsn);
5609 300000 : drop(writer);
5610 300000 : updated[blknum] = lsn;
5611 6 : }
5612 6 :
5613 6 : // Read all the blocks
5614 300000 : for (blknum, last_lsn) in updated.iter().enumerate() {
5615 300000 : test_key.field6 = blknum as u32;
5616 300000 : assert_eq!(
5617 300000 : tline.get(test_key, lsn, &ctx).await?,
5618 300000 : test_img(&format!("{} at {}", blknum, last_lsn))
5619 6 : );
5620 6 : }
5621 6 :
5622 6 : // Perform a cycle of flush, compact, and GC
5623 304 : tline.freeze_and_flush().await?;
5624 45352 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
5625 300 : tenant
5626 300 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
5627 296 : .await?;
5628 6 : }
5629 6 :
5630 6 : Ok(())
5631 6 : }
5632 :
5633 : #[tokio::test]
5634 6 : async fn test_traverse_ancestors() -> anyhow::Result<()> {
5635 6 : let (tenant, ctx) = TenantHarness::create("test_traverse_ancestors")
5636 6 : .await?
5637 6 : .load()
5638 24 : .await;
5639 6 : let mut tline = tenant
5640 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5641 12 : .await?;
5642 6 :
5643 6 : const NUM_KEYS: usize = 100;
5644 6 : const NUM_TLINES: usize = 50;
5645 6 :
5646 6 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5647 6 : // Track page mutation lsns across different timelines.
5648 6 : let mut updated = [[Lsn(0); NUM_KEYS]; NUM_TLINES];
5649 6 :
5650 6 : let mut lsn = Lsn(0x10);
5651 6 :
5652 6 : #[allow(clippy::needless_range_loop)]
5653 306 : for idx in 0..NUM_TLINES {
5654 300 : let new_tline_id = TimelineId::generate();
5655 300 : tenant
5656 300 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
5657 6 : .await?;
5658 300 : tline = tenant
5659 300 : .get_timeline(new_tline_id, true)
5660 300 : .expect("Should have the branched timeline");
5661 6 :
5662 30300 : for _ in 0..NUM_KEYS {
5663 30000 : lsn = Lsn(lsn.0 + 0x10);
5664 30000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
5665 30000 : test_key.field6 = blknum as u32;
5666 30000 : let mut writer = tline.writer().await;
5667 30000 : writer
5668 30000 : .put(
5669 30000 : test_key,
5670 30000 : lsn,
5671 30000 : &Value::Image(test_img(&format!("{} {} at {}", idx, blknum, lsn))),
5672 30000 : &ctx,
5673 30000 : )
5674 264 : .await?;
5675 30000 : println!("updating [{}][{}] at {}", idx, blknum, lsn);
5676 30000 : writer.finish_write(lsn);
5677 30000 : drop(writer);
5678 30000 : updated[idx][blknum] = lsn;
5679 6 : }
5680 6 : }
5681 6 :
5682 6 : // Read pages from leaf timeline across all ancestors.
5683 300 : for (idx, lsns) in updated.iter().enumerate() {
5684 30000 : for (blknum, lsn) in lsns.iter().enumerate() {
5685 6 : // Skip empty mutations.
5686 30000 : if lsn.0 == 0 {
5687 11086 : continue;
5688 18914 : }
5689 18914 : println!("checking [{idx}][{blknum}] at {lsn}");
5690 18914 : test_key.field6 = blknum as u32;
5691 18914 : assert_eq!(
5692 18914 : tline.get(test_key, *lsn, &ctx).await?,
5693 18914 : test_img(&format!("{idx} {blknum} at {lsn}"))
5694 6 : );
5695 6 : }
5696 6 : }
5697 6 : Ok(())
5698 6 : }
5699 :
5700 : #[tokio::test]
5701 6 : async fn test_write_at_initdb_lsn_takes_optimization_code_path() -> anyhow::Result<()> {
5702 6 : let (tenant, ctx) = TenantHarness::create("test_empty_test_timeline_is_usable")
5703 6 : .await?
5704 6 : .load()
5705 24 : .await;
5706 6 :
5707 6 : let initdb_lsn = Lsn(0x20);
5708 6 : let utline = tenant
5709 6 : .create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)
5710 6 : .await?;
5711 6 : let tline = utline.raw_timeline().unwrap();
5712 6 :
5713 6 : // Spawn flush loop now so that we can set the `expect_initdb_optimization`
5714 6 : tline.maybe_spawn_flush_loop();
5715 6 :
5716 6 : // Make sure the timeline has the minimum set of required keys for operation.
5717 6 : // The only operation you can always do on an empty timeline is to `put` new data.
5718 6 : // Except if you `put` at `initdb_lsn`.
5719 6 : // In that case, there's an optimization to directly create image layers instead of delta layers.
5720 6 : // It uses `repartition()`, which assumes some keys to be present.
5721 6 : // Let's make sure the test timeline can handle that case.
5722 6 : {
5723 6 : let mut state = tline.flush_loop_state.lock().unwrap();
5724 6 : assert_eq!(
5725 6 : timeline::FlushLoopState::Running {
5726 6 : expect_initdb_optimization: false,
5727 6 : initdb_optimization_count: 0,
5728 6 : },
5729 6 : *state
5730 6 : );
5731 6 : *state = timeline::FlushLoopState::Running {
5732 6 : expect_initdb_optimization: true,
5733 6 : initdb_optimization_count: 0,
5734 6 : };
5735 6 : }
5736 6 :
5737 6 : // Make writes at the initdb_lsn. When we flush it below, it should be handled by the optimization.
5738 6 : // As explained above, the optimization requires some keys to be present.
5739 6 : // As per `create_empty_timeline` documentation, use init_empty to set them.
5740 6 : // This is what `create_test_timeline` does, by the way.
5741 6 : let mut modification = tline.begin_modification(initdb_lsn);
5742 6 : modification
5743 6 : .init_empty_test_timeline()
5744 6 : .context("init_empty_test_timeline")?;
5745 6 : modification
5746 6 : .commit(&ctx)
5747 6 : .await
5748 6 : .context("commit init_empty_test_timeline modification")?;
5749 6 :
5750 6 : // Do the flush. The flush code will check the expectations that we set above.
5751 6 : tline.freeze_and_flush().await?;
5752 6 :
5753 6 : // assert freeze_and_flush exercised the initdb optimization
5754 6 : {
5755 6 : let state = tline.flush_loop_state.lock().unwrap();
5756 6 : let timeline::FlushLoopState::Running {
5757 6 : expect_initdb_optimization,
5758 6 : initdb_optimization_count,
5759 6 : } = *state
5760 6 : else {
5761 6 : panic!("unexpected state: {:?}", *state);
5762 6 : };
5763 6 : assert!(expect_initdb_optimization);
5764 6 : assert!(initdb_optimization_count > 0);
5765 6 : }
5766 6 : Ok(())
5767 6 : }
5768 :
5769 : #[tokio::test]
5770 6 : async fn test_create_guard_crash() -> anyhow::Result<()> {
5771 6 : let name = "test_create_guard_crash";
5772 6 : let harness = TenantHarness::create(name).await?;
5773 6 : {
5774 24 : let (tenant, ctx) = harness.load().await;
5775 6 : let tline = tenant
5776 6 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
5777 6 : .await?;
5778 6 : // Leave the timeline ID in [`Tenant::timelines_creating`] to exclude attempting to create it again
5779 6 : let raw_tline = tline.raw_timeline().unwrap();
5780 6 : raw_tline
5781 6 : .shutdown(super::timeline::ShutdownMode::Hard)
5782 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))
5783 6 : .await;
5784 6 : std::mem::forget(tline);
5785 6 : }
5786 6 :
5787 22 : let (tenant, _) = harness.load().await;
5788 6 : match tenant.get_timeline(TIMELINE_ID, false) {
5789 6 : Ok(_) => panic!("timeline should've been removed during load"),
5790 6 : Err(e) => {
5791 6 : assert_eq!(
5792 6 : e,
5793 6 : GetTimelineError::NotFound {
5794 6 : tenant_id: tenant.tenant_shard_id,
5795 6 : timeline_id: TIMELINE_ID,
5796 6 : }
5797 6 : )
5798 6 : }
5799 6 : }
5800 6 :
5801 6 : assert!(!harness
5802 6 : .conf
5803 6 : .timeline_path(&tenant.tenant_shard_id, &TIMELINE_ID)
5804 6 : .exists());
5805 6 :
5806 6 : Ok(())
5807 6 : }
5808 :
5809 : #[tokio::test]
5810 6 : async fn test_read_at_max_lsn() -> anyhow::Result<()> {
5811 6 : let names_algorithms = [
5812 6 : ("test_read_at_max_lsn_legacy", CompactionAlgorithm::Legacy),
5813 6 : ("test_read_at_max_lsn_tiered", CompactionAlgorithm::Tiered),
5814 6 : ];
5815 18 : for (name, algorithm) in names_algorithms {
5816 98197 : test_read_at_max_lsn_algorithm(name, algorithm).await?;
5817 6 : }
5818 6 : Ok(())
5819 6 : }
5820 :
5821 12 : async fn test_read_at_max_lsn_algorithm(
5822 12 : name: &'static str,
5823 12 : compaction_algorithm: CompactionAlgorithm,
5824 12 : ) -> anyhow::Result<()> {
5825 12 : let mut harness = TenantHarness::create(name).await?;
5826 12 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
5827 12 : kind: compaction_algorithm,
5828 12 : };
5829 48 : let (tenant, ctx) = harness.load().await;
5830 12 : let tline = tenant
5831 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5832 21 : .await?;
5833 :
5834 12 : let lsn = Lsn(0x10);
5835 12 : let compact = false;
5836 96898 : bulk_insert_maybe_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000, compact).await?;
5837 :
5838 12 : let test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5839 12 : let read_lsn = Lsn(u64::MAX - 1);
5840 :
5841 1230 : let result = tline.get(test_key, read_lsn, &ctx).await;
5842 12 : assert!(result.is_ok(), "result is not Ok: {}", result.unwrap_err());
5843 :
5844 12 : Ok(())
5845 12 : }
5846 :
5847 : #[tokio::test]
5848 6 : async fn test_metadata_scan() -> anyhow::Result<()> {
5849 6 : let harness = TenantHarness::create("test_metadata_scan").await?;
5850 24 : let (tenant, ctx) = harness.load().await;
5851 6 : let tline = tenant
5852 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5853 12 : .await?;
5854 6 :
5855 6 : const NUM_KEYS: usize = 1000;
5856 6 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
5857 6 :
5858 6 : let cancel = CancellationToken::new();
5859 6 :
5860 6 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
5861 6 : base_key.field1 = AUX_KEY_PREFIX;
5862 6 : let mut test_key = base_key;
5863 6 :
5864 6 : // Track when each page was last modified. Used to assert that
5865 6 : // a read sees the latest page version.
5866 6 : let mut updated = [Lsn(0); NUM_KEYS];
5867 6 :
5868 6 : let mut lsn = Lsn(0x10);
5869 6 : #[allow(clippy::needless_range_loop)]
5870 6006 : for blknum in 0..NUM_KEYS {
5871 6000 : lsn = Lsn(lsn.0 + 0x10);
5872 6000 : test_key.field6 = (blknum * STEP) as u32;
5873 6000 : let mut writer = tline.writer().await;
5874 6000 : writer
5875 6000 : .put(
5876 6000 : test_key,
5877 6000 : lsn,
5878 6000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5879 6000 : &ctx,
5880 6000 : )
5881 6 : .await?;
5882 6000 : writer.finish_write(lsn);
5883 6000 : updated[blknum] = lsn;
5884 6000 : drop(writer);
5885 6 : }
5886 6 :
5887 6 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
5888 6 :
5889 72 : for iter in 0..=10 {
5890 6 : // Read all the blocks
5891 66000 : for (blknum, last_lsn) in updated.iter().enumerate() {
5892 66000 : test_key.field6 = (blknum * STEP) as u32;
5893 66000 : assert_eq!(
5894 66000 : tline.get(test_key, lsn, &ctx).await?,
5895 66000 : test_img(&format!("{} at {}", blknum, last_lsn))
5896 6 : );
5897 6 : }
5898 6 :
5899 66 : let mut cnt = 0;
5900 66000 : for (key, value) in tline
5901 66 : .get_vectored_impl(
5902 66 : keyspace.clone(),
5903 66 : lsn,
5904 66 : &mut ValuesReconstructState::default(),
5905 66 : &ctx,
5906 66 : )
5907 12613 : .await?
5908 6 : {
5909 66000 : let blknum = key.field6 as usize;
5910 66000 : let value = value?;
5911 66000 : assert!(blknum % STEP == 0);
5912 66000 : let blknum = blknum / STEP;
5913 66000 : assert_eq!(
5914 66000 : value,
5915 66000 : test_img(&format!("{} at {}", blknum, updated[blknum]))
5916 66000 : );
5917 66000 : cnt += 1;
5918 6 : }
5919 6 :
5920 66 : assert_eq!(cnt, NUM_KEYS);
5921 6 :
5922 66066 : for _ in 0..NUM_KEYS {
5923 66000 : lsn = Lsn(lsn.0 + 0x10);
5924 66000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
5925 66000 : test_key.field6 = (blknum * STEP) as u32;
5926 66000 : let mut writer = tline.writer().await;
5927 66000 : writer
5928 66000 : .put(
5929 66000 : test_key,
5930 66000 : lsn,
5931 66000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5932 66000 : &ctx,
5933 66000 : )
5934 440 : .await?;
5935 66000 : writer.finish_write(lsn);
5936 66000 : drop(writer);
5937 66000 : updated[blknum] = lsn;
5938 6 : }
5939 6 :
5940 6 : // Perform two cycles of flush, compact, and GC
5941 198 : for round in 0..2 {
5942 132 : tline.freeze_and_flush().await?;
5943 132 : tline
5944 132 : .compact(
5945 132 : &cancel,
5946 132 : if iter % 5 == 0 && round == 0 {
5947 18 : let mut flags = EnumSet::new();
5948 18 : flags.insert(CompactFlags::ForceImageLayerCreation);
5949 18 : flags.insert(CompactFlags::ForceRepartition);
5950 18 : flags
5951 6 : } else {
5952 114 : EnumSet::empty()
5953 6 : },
5954 132 : &ctx,
5955 6 : )
5956 25343 : .await?;
5957 132 : tenant
5958 132 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
5959 121 : .await?;
5960 6 : }
5961 6 : }
5962 6 :
5963 6 : Ok(())
5964 6 : }
5965 :
5966 : #[tokio::test]
5967 6 : async fn test_metadata_compaction_trigger() -> anyhow::Result<()> {
5968 6 : let harness = TenantHarness::create("test_metadata_compaction_trigger").await?;
5969 24 : let (tenant, ctx) = harness.load().await;
5970 6 : let tline = tenant
5971 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5972 12 : .await?;
5973 6 :
5974 6 : let cancel = CancellationToken::new();
5975 6 :
5976 6 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
5977 6 : base_key.field1 = AUX_KEY_PREFIX;
5978 6 : let test_key = base_key;
5979 6 : let mut lsn = Lsn(0x10);
5980 6 :
5981 126 : for _ in 0..20 {
5982 120 : lsn = Lsn(lsn.0 + 0x10);
5983 120 : let mut writer = tline.writer().await;
5984 120 : writer
5985 120 : .put(
5986 120 : test_key,
5987 120 : lsn,
5988 120 : &Value::Image(test_img(&format!("{} at {}", 0, lsn))),
5989 120 : &ctx,
5990 120 : )
5991 60 : .await?;
5992 120 : writer.finish_write(lsn);
5993 120 : drop(writer);
5994 120 : tline.freeze_and_flush().await?; // force create a delta layer
5995 6 : }
5996 6 :
5997 6 : let before_num_l0_delta_files =
5998 6 : tline.layers.read().await.layer_map()?.level0_deltas().len();
5999 6 :
6000 450 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
6001 6 :
6002 6 : let after_num_l0_delta_files = tline.layers.read().await.layer_map()?.level0_deltas().len();
6003 6 :
6004 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}");
6005 6 :
6006 6 : assert_eq!(
6007 12 : tline.get(test_key, lsn, &ctx).await?,
6008 6 : test_img(&format!("{} at {}", 0, lsn))
6009 6 : );
6010 6 :
6011 6 : Ok(())
6012 6 : }
6013 :
6014 : #[tokio::test]
6015 6 : async fn test_branch_copies_dirty_aux_file_flag() {
6016 6 : let harness = TenantHarness::create("test_branch_copies_dirty_aux_file_flag")
6017 6 : .await
6018 6 : .unwrap();
6019 6 :
6020 6 : // the default aux file policy to switch is v2 if not set by the admins
6021 6 : assert_eq!(
6022 6 : harness.tenant_conf.switch_aux_file_policy,
6023 6 : AuxFilePolicy::default_tenant_config()
6024 6 : );
6025 24 : let (tenant, ctx) = harness.load().await;
6026 6 :
6027 6 : let mut lsn = Lsn(0x08);
6028 6 :
6029 6 : let tline: Arc<Timeline> = tenant
6030 6 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
6031 12 : .await
6032 6 : .unwrap();
6033 6 :
6034 6 : // no aux file is written at this point, so the persistent flag should be unset
6035 6 : assert_eq!(tline.last_aux_file_policy.load(), None);
6036 6 :
6037 6 : {
6038 6 : lsn += 8;
6039 6 : let mut modification = tline.begin_modification(lsn);
6040 6 : modification
6041 6 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
6042 6 : .await
6043 6 : .unwrap();
6044 6 : modification.commit(&ctx).await.unwrap();
6045 6 : }
6046 6 :
6047 6 : // there is no tenant manager to pass the configuration through, so lets mimic it
6048 6 : tenant.set_new_location_config(
6049 6 : AttachedTenantConf::try_from(LocationConf::attached_single(
6050 6 : TenantConfOpt {
6051 6 : switch_aux_file_policy: Some(AuxFilePolicy::V2),
6052 6 : ..Default::default()
6053 6 : },
6054 6 : tenant.generation,
6055 6 : &pageserver_api::models::ShardParameters::default(),
6056 6 : ))
6057 6 : .unwrap(),
6058 6 : );
6059 6 :
6060 6 : assert_eq!(
6061 6 : tline.get_switch_aux_file_policy(),
6062 6 : AuxFilePolicy::V2,
6063 6 : "wanted state has been updated"
6064 6 : );
6065 6 : assert_eq!(
6066 6 : tline.last_aux_file_policy.load(),
6067 6 : Some(AuxFilePolicy::V2),
6068 6 : "aux file is written with switch_aux_file_policy unset (which is v2), so we should use v2 there"
6069 6 : );
6070 6 :
6071 6 : // we can read everything from the storage
6072 6 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6073 6 : assert_eq!(
6074 6 : files.get("pg_logical/mappings/test1"),
6075 6 : Some(&bytes::Bytes::from_static(b"first"))
6076 6 : );
6077 6 :
6078 6 : {
6079 6 : lsn += 8;
6080 6 : let mut modification = tline.begin_modification(lsn);
6081 6 : modification
6082 6 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
6083 6 : .await
6084 6 : .unwrap();
6085 6 : modification.commit(&ctx).await.unwrap();
6086 6 : }
6087 6 :
6088 6 : assert_eq!(
6089 6 : tline.last_aux_file_policy.load(),
6090 6 : Some(AuxFilePolicy::V2),
6091 6 : "keep v2 storage format when new files are written"
6092 6 : );
6093 6 :
6094 6 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6095 6 : assert_eq!(
6096 6 : files.get("pg_logical/mappings/test2"),
6097 6 : Some(&bytes::Bytes::from_static(b"second"))
6098 6 : );
6099 6 :
6100 6 : let child = tenant
6101 6 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(lsn), &ctx)
6102 6 : .await
6103 6 : .unwrap();
6104 6 :
6105 6 : // child copies the last flag even if that is not on remote storage yet
6106 6 : assert_eq!(child.get_switch_aux_file_policy(), AuxFilePolicy::V2);
6107 6 : assert_eq!(child.last_aux_file_policy.load(), Some(AuxFilePolicy::V2));
6108 6 :
6109 6 : let files = child.list_aux_files(lsn, &ctx).await.unwrap();
6110 6 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
6111 6 : assert_eq!(files.get("pg_logical/mappings/test2"), None);
6112 6 :
6113 6 : // even if we crash here without flushing parent timeline with it's new
6114 6 : // last_aux_file_policy we are safe, because child was never meant to access ancestor's
6115 6 : // files. the ancestor can even switch back to V1 because of a migration safely.
6116 6 : }
6117 :
6118 : #[tokio::test]
6119 6 : async fn aux_file_policy_switch() {
6120 6 : let mut harness = TenantHarness::create("aux_file_policy_switch")
6121 6 : .await
6122 6 : .unwrap();
6123 6 : harness.tenant_conf.switch_aux_file_policy = AuxFilePolicy::CrossValidation; // set to cross-validation mode
6124 24 : let (tenant, ctx) = harness.load().await;
6125 6 :
6126 6 : let mut lsn = Lsn(0x08);
6127 6 :
6128 6 : let tline: Arc<Timeline> = tenant
6129 6 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
6130 12 : .await
6131 6 : .unwrap();
6132 6 :
6133 6 : assert_eq!(
6134 6 : tline.last_aux_file_policy.load(),
6135 6 : None,
6136 6 : "no aux file is written so it should be unset"
6137 6 : );
6138 6 :
6139 6 : {
6140 6 : lsn += 8;
6141 6 : let mut modification = tline.begin_modification(lsn);
6142 6 : modification
6143 6 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
6144 15 : .await
6145 6 : .unwrap();
6146 6 : modification.commit(&ctx).await.unwrap();
6147 6 : }
6148 6 :
6149 6 : // there is no tenant manager to pass the configuration through, so lets mimic it
6150 6 : tenant.set_new_location_config(
6151 6 : AttachedTenantConf::try_from(LocationConf::attached_single(
6152 6 : TenantConfOpt {
6153 6 : switch_aux_file_policy: Some(AuxFilePolicy::V2),
6154 6 : ..Default::default()
6155 6 : },
6156 6 : tenant.generation,
6157 6 : &pageserver_api::models::ShardParameters::default(),
6158 6 : ))
6159 6 : .unwrap(),
6160 6 : );
6161 6 :
6162 6 : assert_eq!(
6163 6 : tline.get_switch_aux_file_policy(),
6164 6 : AuxFilePolicy::V2,
6165 6 : "wanted state has been updated"
6166 6 : );
6167 6 : assert_eq!(
6168 6 : tline.last_aux_file_policy.load(),
6169 6 : Some(AuxFilePolicy::CrossValidation),
6170 6 : "dirty index_part.json reflected state is yet to be updated"
6171 6 : );
6172 6 :
6173 6 : // we can still read the auxfile v1 before we ingest anything new
6174 6 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6175 6 : assert_eq!(
6176 6 : files.get("pg_logical/mappings/test1"),
6177 6 : Some(&bytes::Bytes::from_static(b"first"))
6178 6 : );
6179 6 :
6180 6 : {
6181 6 : lsn += 8;
6182 6 : let mut modification = tline.begin_modification(lsn);
6183 6 : modification
6184 6 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
6185 6 : .await
6186 6 : .unwrap();
6187 6 : modification.commit(&ctx).await.unwrap();
6188 6 : }
6189 6 :
6190 6 : assert_eq!(
6191 6 : tline.last_aux_file_policy.load(),
6192 6 : Some(AuxFilePolicy::V2),
6193 6 : "ingesting a file should apply the wanted switch state when applicable"
6194 6 : );
6195 6 :
6196 6 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6197 6 : assert_eq!(
6198 6 : files.get("pg_logical/mappings/test1"),
6199 6 : Some(&bytes::Bytes::from_static(b"first")),
6200 6 : "cross validation writes to both v1 and v2 so this should be available in v2"
6201 6 : );
6202 6 : assert_eq!(
6203 6 : files.get("pg_logical/mappings/test2"),
6204 6 : Some(&bytes::Bytes::from_static(b"second"))
6205 6 : );
6206 6 :
6207 6 : // mimic again by trying to flip it from V2 to V1 (not switched to while ingesting a file)
6208 6 : tenant.set_new_location_config(
6209 6 : AttachedTenantConf::try_from(LocationConf::attached_single(
6210 6 : TenantConfOpt {
6211 6 : switch_aux_file_policy: Some(AuxFilePolicy::V1),
6212 6 : ..Default::default()
6213 6 : },
6214 6 : tenant.generation,
6215 6 : &pageserver_api::models::ShardParameters::default(),
6216 6 : ))
6217 6 : .unwrap(),
6218 6 : );
6219 6 :
6220 6 : {
6221 6 : lsn += 8;
6222 6 : let mut modification = tline.begin_modification(lsn);
6223 6 : modification
6224 6 : .put_file("pg_logical/mappings/test2", b"third", &ctx)
6225 6 : .await
6226 6 : .unwrap();
6227 6 : modification.commit(&ctx).await.unwrap();
6228 6 : }
6229 6 :
6230 6 : assert_eq!(
6231 6 : tline.get_switch_aux_file_policy(),
6232 6 : AuxFilePolicy::V1,
6233 6 : "wanted state has been updated again, even if invalid request"
6234 6 : );
6235 6 :
6236 6 : assert_eq!(
6237 6 : tline.last_aux_file_policy.load(),
6238 6 : Some(AuxFilePolicy::V2),
6239 6 : "ingesting a file should apply the wanted switch state when applicable"
6240 6 : );
6241 6 :
6242 6 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6243 6 : assert_eq!(
6244 6 : files.get("pg_logical/mappings/test1"),
6245 6 : Some(&bytes::Bytes::from_static(b"first"))
6246 6 : );
6247 6 : assert_eq!(
6248 6 : files.get("pg_logical/mappings/test2"),
6249 6 : Some(&bytes::Bytes::from_static(b"third"))
6250 6 : );
6251 6 :
6252 6 : // mimic again by trying to flip it from from V1 to V2 (not switched to while ingesting a file)
6253 6 : tenant.set_new_location_config(
6254 6 : AttachedTenantConf::try_from(LocationConf::attached_single(
6255 6 : TenantConfOpt {
6256 6 : switch_aux_file_policy: Some(AuxFilePolicy::V2),
6257 6 : ..Default::default()
6258 6 : },
6259 6 : tenant.generation,
6260 6 : &pageserver_api::models::ShardParameters::default(),
6261 6 : ))
6262 6 : .unwrap(),
6263 6 : );
6264 6 :
6265 6 : {
6266 6 : lsn += 8;
6267 6 : let mut modification = tline.begin_modification(lsn);
6268 6 : modification
6269 6 : .put_file("pg_logical/mappings/test3", b"last", &ctx)
6270 6 : .await
6271 6 : .unwrap();
6272 6 : modification.commit(&ctx).await.unwrap();
6273 6 : }
6274 6 :
6275 6 : assert_eq!(tline.get_switch_aux_file_policy(), AuxFilePolicy::V2);
6276 6 :
6277 6 : assert_eq!(tline.last_aux_file_policy.load(), Some(AuxFilePolicy::V2));
6278 6 :
6279 6 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6280 6 : assert_eq!(
6281 6 : files.get("pg_logical/mappings/test1"),
6282 6 : Some(&bytes::Bytes::from_static(b"first"))
6283 6 : );
6284 6 : assert_eq!(
6285 6 : files.get("pg_logical/mappings/test2"),
6286 6 : Some(&bytes::Bytes::from_static(b"third"))
6287 6 : );
6288 6 : assert_eq!(
6289 6 : files.get("pg_logical/mappings/test3"),
6290 6 : Some(&bytes::Bytes::from_static(b"last"))
6291 6 : );
6292 6 : }
6293 :
6294 : #[tokio::test]
6295 6 : async fn aux_file_policy_force_switch() {
6296 6 : let mut harness = TenantHarness::create("aux_file_policy_force_switch")
6297 6 : .await
6298 6 : .unwrap();
6299 6 : harness.tenant_conf.switch_aux_file_policy = AuxFilePolicy::V1;
6300 24 : let (tenant, ctx) = harness.load().await;
6301 6 :
6302 6 : let mut lsn = Lsn(0x08);
6303 6 :
6304 6 : let tline: Arc<Timeline> = tenant
6305 6 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
6306 12 : .await
6307 6 : .unwrap();
6308 6 :
6309 6 : assert_eq!(
6310 6 : tline.last_aux_file_policy.load(),
6311 6 : None,
6312 6 : "no aux file is written so it should be unset"
6313 6 : );
6314 6 :
6315 6 : {
6316 6 : lsn += 8;
6317 6 : let mut modification = tline.begin_modification(lsn);
6318 6 : modification
6319 6 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
6320 15 : .await
6321 6 : .unwrap();
6322 6 : modification.commit(&ctx).await.unwrap();
6323 6 : }
6324 6 :
6325 6 : tline.do_switch_aux_policy(AuxFilePolicy::V2).unwrap();
6326 6 :
6327 6 : assert_eq!(
6328 6 : tline.last_aux_file_policy.load(),
6329 6 : Some(AuxFilePolicy::V2),
6330 6 : "dirty index_part.json reflected state is yet to be updated"
6331 6 : );
6332 6 :
6333 6 : // lose all data from v1
6334 6 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6335 6 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
6336 6 :
6337 6 : {
6338 6 : lsn += 8;
6339 6 : let mut modification = tline.begin_modification(lsn);
6340 6 : modification
6341 6 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
6342 6 : .await
6343 6 : .unwrap();
6344 6 : modification.commit(&ctx).await.unwrap();
6345 6 : }
6346 6 :
6347 6 : // read data ingested in v2
6348 6 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6349 6 : assert_eq!(
6350 6 : files.get("pg_logical/mappings/test2"),
6351 6 : Some(&bytes::Bytes::from_static(b"second"))
6352 6 : );
6353 6 : // lose all data from v1
6354 6 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
6355 6 : }
6356 :
6357 : #[tokio::test]
6358 6 : async fn aux_file_policy_auto_detect() {
6359 6 : let mut harness = TenantHarness::create("aux_file_policy_auto_detect")
6360 6 : .await
6361 6 : .unwrap();
6362 6 : harness.tenant_conf.switch_aux_file_policy = AuxFilePolicy::V2; // set to cross-validation mode
6363 24 : let (tenant, ctx) = harness.load().await;
6364 6 :
6365 6 : let mut lsn = Lsn(0x08);
6366 6 :
6367 6 : let tline: Arc<Timeline> = tenant
6368 6 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
6369 12 : .await
6370 6 : .unwrap();
6371 6 :
6372 6 : assert_eq!(
6373 6 : tline.last_aux_file_policy.load(),
6374 6 : None,
6375 6 : "no aux file is written so it should be unset"
6376 6 : );
6377 6 :
6378 6 : {
6379 6 : lsn += 8;
6380 6 : let mut modification = tline.begin_modification(lsn);
6381 6 : let buf = AuxFilesDirectory::ser(&AuxFilesDirectory {
6382 6 : files: vec![(
6383 6 : "test_file".to_string(),
6384 6 : Bytes::copy_from_slice(b"test_file"),
6385 6 : )]
6386 6 : .into_iter()
6387 6 : .collect(),
6388 6 : })
6389 6 : .unwrap();
6390 6 : modification.put_for_test(AUX_FILES_KEY, Value::Image(Bytes::from(buf)));
6391 6 : modification.commit(&ctx).await.unwrap();
6392 6 : }
6393 6 :
6394 6 : {
6395 6 : lsn += 8;
6396 6 : let mut modification = tline.begin_modification(lsn);
6397 6 : modification
6398 6 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
6399 6 : .await
6400 6 : .unwrap();
6401 6 : modification.commit(&ctx).await.unwrap();
6402 6 : }
6403 6 :
6404 6 : assert_eq!(
6405 6 : tline.last_aux_file_policy.load(),
6406 6 : Some(AuxFilePolicy::V1),
6407 6 : "keep using v1 because there are aux files writting with v1"
6408 6 : );
6409 6 :
6410 6 : // we can still read the auxfile v1
6411 6 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6412 6 : assert_eq!(
6413 6 : files.get("pg_logical/mappings/test1"),
6414 6 : Some(&bytes::Bytes::from_static(b"first"))
6415 6 : );
6416 6 : assert_eq!(
6417 6 : files.get("test_file"),
6418 6 : Some(&bytes::Bytes::from_static(b"test_file"))
6419 6 : );
6420 6 : }
6421 :
6422 : #[tokio::test]
6423 6 : async fn test_metadata_image_creation() -> anyhow::Result<()> {
6424 6 : let harness = TenantHarness::create("test_metadata_image_creation").await?;
6425 24 : let (tenant, ctx) = harness.load().await;
6426 6 : let tline = tenant
6427 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6428 12 : .await?;
6429 6 :
6430 6 : const NUM_KEYS: usize = 1000;
6431 6 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
6432 6 :
6433 6 : let cancel = CancellationToken::new();
6434 6 :
6435 6 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
6436 6 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
6437 6 : let mut test_key = base_key;
6438 6 : let mut lsn = Lsn(0x10);
6439 6 :
6440 24 : async fn scan_with_statistics(
6441 24 : tline: &Timeline,
6442 24 : keyspace: &KeySpace,
6443 24 : lsn: Lsn,
6444 24 : ctx: &RequestContext,
6445 24 : ) -> anyhow::Result<(BTreeMap<Key, Result<Bytes, PageReconstructError>>, usize)> {
6446 24 : let mut reconstruct_state = ValuesReconstructState::default();
6447 24 : let res = tline
6448 24 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
6449 4098 : .await?;
6450 24 : Ok((res, reconstruct_state.get_delta_layers_visited() as usize))
6451 24 : }
6452 6 :
6453 6 : #[allow(clippy::needless_range_loop)]
6454 6006 : for blknum in 0..NUM_KEYS {
6455 6000 : lsn = Lsn(lsn.0 + 0x10);
6456 6000 : test_key.field6 = (blknum * STEP) as u32;
6457 6000 : let mut writer = tline.writer().await;
6458 6000 : writer
6459 6000 : .put(
6460 6000 : test_key,
6461 6000 : lsn,
6462 6000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6463 6000 : &ctx,
6464 6000 : )
6465 6 : .await?;
6466 6000 : writer.finish_write(lsn);
6467 6000 : drop(writer);
6468 6 : }
6469 6 :
6470 6 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
6471 6 :
6472 66 : for iter in 1..=10 {
6473 60060 : for _ in 0..NUM_KEYS {
6474 60000 : lsn = Lsn(lsn.0 + 0x10);
6475 60000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6476 60000 : test_key.field6 = (blknum * STEP) as u32;
6477 60000 : let mut writer = tline.writer().await;
6478 60000 : writer
6479 60000 : .put(
6480 60000 : test_key,
6481 60000 : lsn,
6482 60000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6483 60000 : &ctx,
6484 60000 : )
6485 104 : .await?;
6486 60000 : writer.finish_write(lsn);
6487 60000 : drop(writer);
6488 6 : }
6489 6 :
6490 60 : tline.freeze_and_flush().await?;
6491 6 :
6492 60 : if iter % 5 == 0 {
6493 12 : let (_, before_delta_file_accessed) =
6494 4073 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
6495 12 : tline
6496 12 : .compact(
6497 12 : &cancel,
6498 12 : {
6499 12 : let mut flags = EnumSet::new();
6500 12 : flags.insert(CompactFlags::ForceImageLayerCreation);
6501 12 : flags.insert(CompactFlags::ForceRepartition);
6502 12 : flags
6503 12 : },
6504 12 : &ctx,
6505 12 : )
6506 18598 : .await?;
6507 12 : let (_, after_delta_file_accessed) =
6508 25 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
6509 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}");
6510 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.
6511 12 : assert!(
6512 12 : after_delta_file_accessed <= 2,
6513 6 : "after_delta_file_accessed={after_delta_file_accessed}"
6514 6 : );
6515 48 : }
6516 6 : }
6517 6 :
6518 6 : Ok(())
6519 6 : }
6520 :
6521 : #[tokio::test]
6522 6 : async fn test_vectored_missing_data_key_reads() -> anyhow::Result<()> {
6523 6 : let harness = TenantHarness::create("test_vectored_missing_data_key_reads").await?;
6524 24 : let (tenant, ctx) = harness.load().await;
6525 6 :
6526 6 : let base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
6527 6 : let base_key_child = Key::from_hex("000000000033333333444444445500000001").unwrap();
6528 6 : let base_key_nonexist = Key::from_hex("000000000033333333444444445500000002").unwrap();
6529 6 :
6530 6 : let tline = tenant
6531 6 : .create_test_timeline_with_layers(
6532 6 : TIMELINE_ID,
6533 6 : Lsn(0x10),
6534 6 : DEFAULT_PG_VERSION,
6535 6 : &ctx,
6536 6 : Vec::new(), // delta layers
6537 6 : vec![(Lsn(0x20), vec![(base_key, test_img("data key 1"))])], // image layers
6538 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
6539 6 : )
6540 33 : .await?;
6541 6 : tline.add_extra_test_dense_keyspace(KeySpace::single(base_key..(base_key_nonexist.next())));
6542 6 :
6543 6 : let child = tenant
6544 6 : .branch_timeline_test_with_layers(
6545 6 : &tline,
6546 6 : NEW_TIMELINE_ID,
6547 6 : Some(Lsn(0x20)),
6548 6 : &ctx,
6549 6 : Vec::new(), // delta layers
6550 6 : vec![(Lsn(0x30), vec![(base_key_child, test_img("data key 2"))])], // image layers
6551 6 : Lsn(0x30),
6552 6 : )
6553 21 : .await
6554 6 : .unwrap();
6555 6 :
6556 6 : let lsn = Lsn(0x30);
6557 6 :
6558 6 : // test vectored get on parent timeline
6559 6 : assert_eq!(
6560 12 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
6561 6 : Some(test_img("data key 1"))
6562 6 : );
6563 6 : assert!(get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx)
6564 9 : .await
6565 6 : .unwrap_err()
6566 6 : .is_missing_key_error());
6567 6 : assert!(
6568 6 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx)
6569 6 : .await
6570 6 : .unwrap_err()
6571 6 : .is_missing_key_error()
6572 6 : );
6573 6 :
6574 6 : // test vectored get on child timeline
6575 6 : assert_eq!(
6576 6 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
6577 6 : Some(test_img("data key 1"))
6578 6 : );
6579 6 : assert_eq!(
6580 12 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
6581 6 : Some(test_img("data key 2"))
6582 6 : );
6583 6 : assert!(
6584 6 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx)
6585 6 : .await
6586 6 : .unwrap_err()
6587 6 : .is_missing_key_error()
6588 6 : );
6589 6 :
6590 6 : Ok(())
6591 6 : }
6592 :
6593 : #[tokio::test]
6594 6 : async fn test_vectored_missing_metadata_key_reads() -> anyhow::Result<()> {
6595 6 : let harness = TenantHarness::create("test_vectored_missing_metadata_key_reads").await?;
6596 24 : let (tenant, ctx) = harness.load().await;
6597 6 :
6598 6 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
6599 6 : let base_key_child = Key::from_hex("620000000033333333444444445500000001").unwrap();
6600 6 : let base_key_nonexist = Key::from_hex("620000000033333333444444445500000002").unwrap();
6601 6 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
6602 6 :
6603 6 : let tline = tenant
6604 6 : .create_test_timeline_with_layers(
6605 6 : TIMELINE_ID,
6606 6 : Lsn(0x10),
6607 6 : DEFAULT_PG_VERSION,
6608 6 : &ctx,
6609 6 : Vec::new(), // delta layers
6610 6 : vec![(Lsn(0x20), vec![(base_key, test_img("metadata key 1"))])], // image layers
6611 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
6612 6 : )
6613 33 : .await?;
6614 6 :
6615 6 : let child = tenant
6616 6 : .branch_timeline_test_with_layers(
6617 6 : &tline,
6618 6 : NEW_TIMELINE_ID,
6619 6 : Some(Lsn(0x20)),
6620 6 : &ctx,
6621 6 : Vec::new(), // delta layers
6622 6 : vec![(
6623 6 : Lsn(0x30),
6624 6 : vec![(base_key_child, test_img("metadata key 2"))],
6625 6 : )], // image layers
6626 6 : Lsn(0x30),
6627 6 : )
6628 21 : .await
6629 6 : .unwrap();
6630 6 :
6631 6 : let lsn = Lsn(0x30);
6632 6 :
6633 6 : // test vectored get on parent timeline
6634 6 : assert_eq!(
6635 12 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
6636 6 : Some(test_img("metadata key 1"))
6637 6 : );
6638 6 : assert_eq!(
6639 6 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx).await?,
6640 6 : None
6641 6 : );
6642 6 : assert_eq!(
6643 6 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx).await?,
6644 6 : None
6645 6 : );
6646 6 :
6647 6 : // test vectored get on child timeline
6648 6 : assert_eq!(
6649 6 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
6650 6 : None
6651 6 : );
6652 6 : assert_eq!(
6653 12 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
6654 6 : Some(test_img("metadata key 2"))
6655 6 : );
6656 6 : assert_eq!(
6657 6 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx).await?,
6658 6 : None
6659 6 : );
6660 6 :
6661 6 : Ok(())
6662 6 : }
6663 :
6664 108 : async fn get_vectored_impl_wrapper(
6665 108 : tline: &Arc<Timeline>,
6666 108 : key: Key,
6667 108 : lsn: Lsn,
6668 108 : ctx: &RequestContext,
6669 108 : ) -> Result<Option<Bytes>, GetVectoredError> {
6670 108 : let mut reconstruct_state = ValuesReconstructState::new();
6671 108 : let mut res = tline
6672 108 : .get_vectored_impl(
6673 108 : KeySpace::single(key..key.next()),
6674 108 : lsn,
6675 108 : &mut reconstruct_state,
6676 108 : ctx,
6677 108 : )
6678 99 : .await?;
6679 90 : Ok(res.pop_last().map(|(k, v)| {
6680 54 : assert_eq!(k, key);
6681 54 : v.unwrap()
6682 90 : }))
6683 108 : }
6684 :
6685 : #[tokio::test]
6686 6 : async fn test_metadata_tombstone_reads() -> anyhow::Result<()> {
6687 6 : let harness = TenantHarness::create("test_metadata_tombstone_reads").await?;
6688 24 : let (tenant, ctx) = harness.load().await;
6689 6 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
6690 6 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
6691 6 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
6692 6 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
6693 6 :
6694 6 : // We emulate the situation that the compaction algorithm creates an image layer that removes the tombstones
6695 6 : // Lsn 0x30 key0, key3, no key1+key2
6696 6 : // Lsn 0x20 key1+key2 tomestones
6697 6 : // Lsn 0x10 key1 in image, key2 in delta
6698 6 : let tline = tenant
6699 6 : .create_test_timeline_with_layers(
6700 6 : TIMELINE_ID,
6701 6 : Lsn(0x10),
6702 6 : DEFAULT_PG_VERSION,
6703 6 : &ctx,
6704 6 : // delta layers
6705 6 : vec![
6706 6 : DeltaLayerTestDesc::new_with_inferred_key_range(
6707 6 : Lsn(0x10)..Lsn(0x20),
6708 6 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
6709 6 : ),
6710 6 : DeltaLayerTestDesc::new_with_inferred_key_range(
6711 6 : Lsn(0x20)..Lsn(0x30),
6712 6 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
6713 6 : ),
6714 6 : DeltaLayerTestDesc::new_with_inferred_key_range(
6715 6 : Lsn(0x20)..Lsn(0x30),
6716 6 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
6717 6 : ),
6718 6 : ],
6719 6 : // image layers
6720 6 : vec![
6721 6 : (Lsn(0x10), vec![(key1, test_img("metadata key 1"))]),
6722 6 : (
6723 6 : Lsn(0x30),
6724 6 : vec![
6725 6 : (key0, test_img("metadata key 0")),
6726 6 : (key3, test_img("metadata key 3")),
6727 6 : ],
6728 6 : ),
6729 6 : ],
6730 6 : Lsn(0x30),
6731 6 : )
6732 114 : .await?;
6733 6 :
6734 6 : let lsn = Lsn(0x30);
6735 6 : let old_lsn = Lsn(0x20);
6736 6 :
6737 6 : assert_eq!(
6738 12 : get_vectored_impl_wrapper(&tline, key0, lsn, &ctx).await?,
6739 6 : Some(test_img("metadata key 0"))
6740 6 : );
6741 6 : assert_eq!(
6742 6 : get_vectored_impl_wrapper(&tline, key1, lsn, &ctx).await?,
6743 6 : None,
6744 6 : );
6745 6 : assert_eq!(
6746 6 : get_vectored_impl_wrapper(&tline, key2, lsn, &ctx).await?,
6747 6 : None,
6748 6 : );
6749 6 : assert_eq!(
6750 12 : get_vectored_impl_wrapper(&tline, key1, old_lsn, &ctx).await?,
6751 6 : Some(Bytes::new()),
6752 6 : );
6753 6 : assert_eq!(
6754 12 : get_vectored_impl_wrapper(&tline, key2, old_lsn, &ctx).await?,
6755 6 : Some(Bytes::new()),
6756 6 : );
6757 6 : assert_eq!(
6758 6 : get_vectored_impl_wrapper(&tline, key3, lsn, &ctx).await?,
6759 6 : Some(test_img("metadata key 3"))
6760 6 : );
6761 6 :
6762 6 : Ok(())
6763 6 : }
6764 :
6765 : #[tokio::test]
6766 6 : async fn test_metadata_tombstone_image_creation() {
6767 6 : let harness = TenantHarness::create("test_metadata_tombstone_image_creation")
6768 6 : .await
6769 6 : .unwrap();
6770 24 : let (tenant, ctx) = harness.load().await;
6771 6 :
6772 6 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
6773 6 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
6774 6 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
6775 6 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
6776 6 :
6777 6 : let tline = tenant
6778 6 : .create_test_timeline_with_layers(
6779 6 : TIMELINE_ID,
6780 6 : Lsn(0x10),
6781 6 : DEFAULT_PG_VERSION,
6782 6 : &ctx,
6783 6 : // delta layers
6784 6 : vec![
6785 6 : DeltaLayerTestDesc::new_with_inferred_key_range(
6786 6 : Lsn(0x10)..Lsn(0x20),
6787 6 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
6788 6 : ),
6789 6 : DeltaLayerTestDesc::new_with_inferred_key_range(
6790 6 : Lsn(0x20)..Lsn(0x30),
6791 6 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
6792 6 : ),
6793 6 : DeltaLayerTestDesc::new_with_inferred_key_range(
6794 6 : Lsn(0x20)..Lsn(0x30),
6795 6 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
6796 6 : ),
6797 6 : DeltaLayerTestDesc::new_with_inferred_key_range(
6798 6 : Lsn(0x30)..Lsn(0x40),
6799 6 : vec![
6800 6 : (key0, Lsn(0x30), Value::Image(test_img("metadata key 0"))),
6801 6 : (key3, Lsn(0x30), Value::Image(test_img("metadata key 3"))),
6802 6 : ],
6803 6 : ),
6804 6 : ],
6805 6 : // image layers
6806 6 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
6807 6 : Lsn(0x40),
6808 6 : )
6809 105 : .await
6810 6 : .unwrap();
6811 6 :
6812 6 : let cancel = CancellationToken::new();
6813 6 :
6814 6 : tline
6815 6 : .compact(
6816 6 : &cancel,
6817 6 : {
6818 6 : let mut flags = EnumSet::new();
6819 6 : flags.insert(CompactFlags::ForceImageLayerCreation);
6820 6 : flags.insert(CompactFlags::ForceRepartition);
6821 6 : flags
6822 6 : },
6823 6 : &ctx,
6824 6 : )
6825 165 : .await
6826 6 : .unwrap();
6827 6 :
6828 6 : // Image layers are created at last_record_lsn
6829 6 : let images = tline
6830 6 : .inspect_image_layers(Lsn(0x40), &ctx)
6831 24 : .await
6832 6 : .unwrap()
6833 6 : .into_iter()
6834 54 : .filter(|(k, _)| k.is_metadata_key())
6835 6 : .collect::<Vec<_>>();
6836 6 : assert_eq!(images.len(), 2); // the image layer should only contain two existing keys, tombstones should be removed.
6837 6 : }
6838 :
6839 : #[tokio::test]
6840 6 : async fn test_metadata_tombstone_empty_image_creation() {
6841 6 : let harness = TenantHarness::create("test_metadata_tombstone_empty_image_creation")
6842 6 : .await
6843 6 : .unwrap();
6844 24 : let (tenant, ctx) = harness.load().await;
6845 6 :
6846 6 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
6847 6 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
6848 6 :
6849 6 : let tline = tenant
6850 6 : .create_test_timeline_with_layers(
6851 6 : TIMELINE_ID,
6852 6 : Lsn(0x10),
6853 6 : DEFAULT_PG_VERSION,
6854 6 : &ctx,
6855 6 : // delta layers
6856 6 : vec![
6857 6 : DeltaLayerTestDesc::new_with_inferred_key_range(
6858 6 : Lsn(0x10)..Lsn(0x20),
6859 6 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
6860 6 : ),
6861 6 : DeltaLayerTestDesc::new_with_inferred_key_range(
6862 6 : Lsn(0x20)..Lsn(0x30),
6863 6 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
6864 6 : ),
6865 6 : DeltaLayerTestDesc::new_with_inferred_key_range(
6866 6 : Lsn(0x20)..Lsn(0x30),
6867 6 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
6868 6 : ),
6869 6 : ],
6870 6 : // image layers
6871 6 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
6872 6 : Lsn(0x30),
6873 6 : )
6874 87 : .await
6875 6 : .unwrap();
6876 6 :
6877 6 : let cancel = CancellationToken::new();
6878 6 :
6879 6 : tline
6880 6 : .compact(
6881 6 : &cancel,
6882 6 : {
6883 6 : let mut flags = EnumSet::new();
6884 6 : flags.insert(CompactFlags::ForceImageLayerCreation);
6885 6 : flags.insert(CompactFlags::ForceRepartition);
6886 6 : flags
6887 6 : },
6888 6 : &ctx,
6889 6 : )
6890 129 : .await
6891 6 : .unwrap();
6892 6 :
6893 6 : // Image layers are created at last_record_lsn
6894 6 : let images = tline
6895 6 : .inspect_image_layers(Lsn(0x30), &ctx)
6896 12 : .await
6897 6 : .unwrap()
6898 6 : .into_iter()
6899 42 : .filter(|(k, _)| k.is_metadata_key())
6900 6 : .collect::<Vec<_>>();
6901 6 : assert_eq!(images.len(), 0); // the image layer should not contain tombstones, or it is not created
6902 6 : }
6903 :
6904 : #[tokio::test]
6905 6 : async fn test_simple_bottom_most_compaction_images() -> anyhow::Result<()> {
6906 6 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_images").await?;
6907 23 : let (tenant, ctx) = harness.load().await;
6908 6 :
6909 294 : fn get_key(id: u32) -> Key {
6910 294 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
6911 294 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
6912 294 : key.field6 = id;
6913 294 : key
6914 294 : }
6915 6 :
6916 6 : // We create
6917 6 : // - one bottom-most image layer,
6918 6 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
6919 6 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
6920 6 : // - a delta layer D3 above the horizon.
6921 6 : //
6922 6 : // | D3 |
6923 6 : // | D1 |
6924 6 : // -| |-- gc horizon -----------------
6925 6 : // | | | D2 |
6926 6 : // --------- img layer ------------------
6927 6 : //
6928 6 : // What we should expact from this compaction is:
6929 6 : // | D3 |
6930 6 : // | Part of D1 |
6931 6 : // --------- img layer with D1+D2 at GC horizon------------------
6932 6 :
6933 6 : // img layer at 0x10
6934 6 : let img_layer = (0..10)
6935 60 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
6936 6 : .collect_vec();
6937 6 :
6938 6 : let delta1 = vec![
6939 6 : (
6940 6 : get_key(1),
6941 6 : Lsn(0x20),
6942 6 : Value::Image(Bytes::from("value 1@0x20")),
6943 6 : ),
6944 6 : (
6945 6 : get_key(2),
6946 6 : Lsn(0x30),
6947 6 : Value::Image(Bytes::from("value 2@0x30")),
6948 6 : ),
6949 6 : (
6950 6 : get_key(3),
6951 6 : Lsn(0x40),
6952 6 : Value::Image(Bytes::from("value 3@0x40")),
6953 6 : ),
6954 6 : ];
6955 6 : let delta2 = vec![
6956 6 : (
6957 6 : get_key(5),
6958 6 : Lsn(0x20),
6959 6 : Value::Image(Bytes::from("value 5@0x20")),
6960 6 : ),
6961 6 : (
6962 6 : get_key(6),
6963 6 : Lsn(0x20),
6964 6 : Value::Image(Bytes::from("value 6@0x20")),
6965 6 : ),
6966 6 : ];
6967 6 : let delta3 = vec![
6968 6 : (
6969 6 : get_key(8),
6970 6 : Lsn(0x48),
6971 6 : Value::Image(Bytes::from("value 8@0x48")),
6972 6 : ),
6973 6 : (
6974 6 : get_key(9),
6975 6 : Lsn(0x48),
6976 6 : Value::Image(Bytes::from("value 9@0x48")),
6977 6 : ),
6978 6 : ];
6979 6 :
6980 6 : let tline = tenant
6981 6 : .create_test_timeline_with_layers(
6982 6 : TIMELINE_ID,
6983 6 : Lsn(0x10),
6984 6 : DEFAULT_PG_VERSION,
6985 6 : &ctx,
6986 6 : vec![
6987 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
6988 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
6989 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
6990 6 : ], // delta layers
6991 6 : vec![(Lsn(0x10), img_layer)], // image layers
6992 6 : Lsn(0x50),
6993 6 : )
6994 141 : .await?;
6995 6 : {
6996 6 : // Update GC info
6997 6 : let mut guard = tline.gc_info.write().unwrap();
6998 6 : guard.cutoffs.time = Lsn(0x30);
6999 6 : guard.cutoffs.space = Lsn(0x30);
7000 6 : }
7001 6 :
7002 6 : let expected_result = [
7003 6 : Bytes::from_static(b"value 0@0x10"),
7004 6 : Bytes::from_static(b"value 1@0x20"),
7005 6 : Bytes::from_static(b"value 2@0x30"),
7006 6 : Bytes::from_static(b"value 3@0x40"),
7007 6 : Bytes::from_static(b"value 4@0x10"),
7008 6 : Bytes::from_static(b"value 5@0x20"),
7009 6 : Bytes::from_static(b"value 6@0x20"),
7010 6 : Bytes::from_static(b"value 7@0x10"),
7011 6 : Bytes::from_static(b"value 8@0x48"),
7012 6 : Bytes::from_static(b"value 9@0x48"),
7013 6 : ];
7014 6 :
7015 60 : for (idx, expected) in expected_result.iter().enumerate() {
7016 60 : assert_eq!(
7017 60 : tline
7018 60 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7019 69 : .await
7020 60 : .unwrap(),
7021 6 : expected
7022 6 : );
7023 6 : }
7024 6 :
7025 6 : let cancel = CancellationToken::new();
7026 6 : tline
7027 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7028 159 : .await
7029 6 : .unwrap();
7030 6 :
7031 60 : for (idx, expected) in expected_result.iter().enumerate() {
7032 60 : assert_eq!(
7033 60 : tline
7034 60 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7035 48 : .await
7036 60 : .unwrap(),
7037 6 : expected
7038 6 : );
7039 6 : }
7040 6 :
7041 6 : // Check if the image layer at the GC horizon contains exactly what we want
7042 6 : let image_at_gc_horizon = tline
7043 6 : .inspect_image_layers(Lsn(0x30), &ctx)
7044 6 : .await
7045 6 : .unwrap()
7046 6 : .into_iter()
7047 102 : .filter(|(k, _)| k.is_metadata_key())
7048 6 : .collect::<Vec<_>>();
7049 6 :
7050 6 : assert_eq!(image_at_gc_horizon.len(), 10);
7051 6 : let expected_result = [
7052 6 : Bytes::from_static(b"value 0@0x10"),
7053 6 : Bytes::from_static(b"value 1@0x20"),
7054 6 : Bytes::from_static(b"value 2@0x30"),
7055 6 : Bytes::from_static(b"value 3@0x10"),
7056 6 : Bytes::from_static(b"value 4@0x10"),
7057 6 : Bytes::from_static(b"value 5@0x20"),
7058 6 : Bytes::from_static(b"value 6@0x20"),
7059 6 : Bytes::from_static(b"value 7@0x10"),
7060 6 : Bytes::from_static(b"value 8@0x10"),
7061 6 : Bytes::from_static(b"value 9@0x10"),
7062 6 : ];
7063 66 : for idx in 0..10 {
7064 60 : assert_eq!(
7065 60 : image_at_gc_horizon[idx],
7066 60 : (get_key(idx as u32), expected_result[idx].clone())
7067 60 : );
7068 6 : }
7069 6 :
7070 6 : // Check if old layers are removed / new layers have the expected LSN
7071 6 : let mut all_layers = tline.inspect_historic_layers().await.unwrap();
7072 12 : all_layers.sort_by(|k1, k2| {
7073 12 : (
7074 12 : k1.is_delta,
7075 12 : k1.key_range.start,
7076 12 : k1.key_range.end,
7077 12 : k1.lsn_range.start,
7078 12 : k1.lsn_range.end,
7079 12 : )
7080 12 : .cmp(&(
7081 12 : k2.is_delta,
7082 12 : k2.key_range.start,
7083 12 : k2.key_range.end,
7084 12 : k2.lsn_range.start,
7085 12 : k2.lsn_range.end,
7086 12 : ))
7087 12 : });
7088 6 : assert_eq!(
7089 6 : all_layers,
7090 6 : vec![
7091 6 : // Image layer at GC horizon
7092 6 : PersistentLayerKey {
7093 6 : key_range: Key::MIN..Key::NON_L0_MAX,
7094 6 : lsn_range: Lsn(0x30)..Lsn(0x31),
7095 6 : is_delta: false
7096 6 : },
7097 6 : // The delta layer covers the full range (with the layer key hack to avoid being recognized as L0)
7098 6 : PersistentLayerKey {
7099 6 : key_range: Key::MIN..Key::NON_L0_MAX,
7100 6 : lsn_range: Lsn(0x30)..Lsn(0x48),
7101 6 : is_delta: true
7102 6 : },
7103 6 : // The delta3 layer that should not be picked for the compaction
7104 6 : PersistentLayerKey {
7105 6 : key_range: get_key(8)..get_key(10),
7106 6 : lsn_range: Lsn(0x48)..Lsn(0x50),
7107 6 : is_delta: true
7108 6 : }
7109 6 : ]
7110 6 : );
7111 6 :
7112 6 : // increase GC horizon and compact again
7113 6 : {
7114 6 : // Update GC info
7115 6 : let mut guard = tline.gc_info.write().unwrap();
7116 6 : guard.cutoffs.time = Lsn(0x40);
7117 6 : guard.cutoffs.space = Lsn(0x40);
7118 6 : }
7119 6 : tline
7120 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7121 129 : .await
7122 6 : .unwrap();
7123 6 :
7124 6 : Ok(())
7125 6 : }
7126 :
7127 : #[tokio::test]
7128 6 : async fn test_neon_test_record() -> anyhow::Result<()> {
7129 6 : let harness = TenantHarness::create("test_neon_test_record").await?;
7130 24 : let (tenant, ctx) = harness.load().await;
7131 6 :
7132 72 : fn get_key(id: u32) -> Key {
7133 72 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7134 72 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7135 72 : key.field6 = id;
7136 72 : key
7137 72 : }
7138 6 :
7139 6 : let delta1 = vec![
7140 6 : (
7141 6 : get_key(1),
7142 6 : Lsn(0x20),
7143 6 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
7144 6 : ),
7145 6 : (
7146 6 : get_key(1),
7147 6 : Lsn(0x30),
7148 6 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
7149 6 : ),
7150 6 : (get_key(2), Lsn(0x10), Value::Image("0x10".into())),
7151 6 : (
7152 6 : get_key(2),
7153 6 : Lsn(0x20),
7154 6 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
7155 6 : ),
7156 6 : (
7157 6 : get_key(2),
7158 6 : Lsn(0x30),
7159 6 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
7160 6 : ),
7161 6 : (get_key(3), Lsn(0x10), Value::Image("0x10".into())),
7162 6 : (
7163 6 : get_key(3),
7164 6 : Lsn(0x20),
7165 6 : Value::WalRecord(NeonWalRecord::wal_clear()),
7166 6 : ),
7167 6 : (get_key(4), Lsn(0x10), Value::Image("0x10".into())),
7168 6 : (
7169 6 : get_key(4),
7170 6 : Lsn(0x20),
7171 6 : Value::WalRecord(NeonWalRecord::wal_init()),
7172 6 : ),
7173 6 : ];
7174 6 : let image1 = vec![(get_key(1), "0x10".into())];
7175 6 :
7176 6 : let tline = tenant
7177 6 : .create_test_timeline_with_layers(
7178 6 : TIMELINE_ID,
7179 6 : Lsn(0x10),
7180 6 : DEFAULT_PG_VERSION,
7181 6 : &ctx,
7182 6 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
7183 6 : Lsn(0x10)..Lsn(0x40),
7184 6 : delta1,
7185 6 : )], // delta layers
7186 6 : vec![(Lsn(0x10), image1)], // image layers
7187 6 : Lsn(0x50),
7188 6 : )
7189 51 : .await?;
7190 6 :
7191 6 : assert_eq!(
7192 24 : tline.get(get_key(1), Lsn(0x50), &ctx).await?,
7193 6 : Bytes::from_static(b"0x10,0x20,0x30")
7194 6 : );
7195 6 : assert_eq!(
7196 6 : tline.get(get_key(2), Lsn(0x50), &ctx).await?,
7197 6 : Bytes::from_static(b"0x10,0x20,0x30")
7198 6 : );
7199 6 :
7200 6 : // Need to remove the limit of "Neon WAL redo requires base image".
7201 6 :
7202 6 : // assert_eq!(tline.get(get_key(3), Lsn(0x50), &ctx).await?, Bytes::new());
7203 6 : // assert_eq!(tline.get(get_key(4), Lsn(0x50), &ctx).await?, Bytes::new());
7204 6 :
7205 6 : Ok(())
7206 6 : }
7207 :
7208 : #[tokio::test]
7209 6 : async fn test_lsn_lease() -> anyhow::Result<()> {
7210 24 : let (tenant, ctx) = TenantHarness::create("test_lsn_lease").await?.load().await;
7211 6 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7212 6 :
7213 6 : let end_lsn = Lsn(0x100);
7214 6 : let image_layers = (0x20..=0x90)
7215 6 : .step_by(0x10)
7216 48 : .map(|n| {
7217 48 : (
7218 48 : Lsn(n),
7219 48 : vec![(key, test_img(&format!("data key at {:x}", n)))],
7220 48 : )
7221 48 : })
7222 6 : .collect();
7223 6 :
7224 6 : let timeline = tenant
7225 6 : .create_test_timeline_with_layers(
7226 6 : TIMELINE_ID,
7227 6 : Lsn(0x10),
7228 6 : DEFAULT_PG_VERSION,
7229 6 : &ctx,
7230 6 : Vec::new(),
7231 6 : image_layers,
7232 6 : end_lsn,
7233 6 : )
7234 180 : .await?;
7235 6 :
7236 6 : let leased_lsns = [0x30, 0x50, 0x70];
7237 6 : let mut leases = Vec::new();
7238 18 : let _: anyhow::Result<_> = leased_lsns.iter().try_for_each(|n| {
7239 18 : leases.push(timeline.make_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)?);
7240 18 : Ok(())
7241 18 : });
7242 6 :
7243 6 : // Renewing with shorter lease should not change the lease.
7244 6 : let updated_lease_0 =
7245 6 : timeline.make_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)?;
7246 6 : assert_eq!(updated_lease_0.valid_until, leases[0].valid_until);
7247 6 :
7248 6 : // Renewing with a long lease should renew lease with later expiration time.
7249 6 : let updated_lease_1 = timeline.make_lsn_lease(
7250 6 : Lsn(leased_lsns[1]),
7251 6 : timeline.get_lsn_lease_length() * 2,
7252 6 : &ctx,
7253 6 : )?;
7254 6 :
7255 6 : assert!(updated_lease_1.valid_until > leases[1].valid_until);
7256 6 :
7257 6 : // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
7258 6 : info!(
7259 6 : "latest_gc_cutoff_lsn: {}",
7260 0 : *timeline.get_latest_gc_cutoff_lsn()
7261 6 : );
7262 6 : timeline.force_set_disk_consistent_lsn(end_lsn);
7263 6 :
7264 6 : let res = tenant
7265 6 : .gc_iteration(
7266 6 : Some(TIMELINE_ID),
7267 6 : 0,
7268 6 : Duration::ZERO,
7269 6 : &CancellationToken::new(),
7270 6 : &ctx,
7271 6 : )
7272 6 : .await?;
7273 6 :
7274 6 : // Keeping everything <= Lsn(0x80) b/c leases:
7275 6 : // 0/10: initdb layer
7276 6 : // (0/20..=0/70).step_by(0x10): image layers added when creating the timeline.
7277 6 : assert_eq!(res.layers_needed_by_leases, 7);
7278 6 : // Keeping 0/90 b/c it is the latest layer.
7279 6 : assert_eq!(res.layers_not_updated, 1);
7280 6 : // Removed 0/80.
7281 6 : assert_eq!(res.layers_removed, 1);
7282 6 :
7283 6 : // Make lease on a already GC-ed LSN.
7284 6 : // 0/80 does not have a valid lease + is below latest_gc_cutoff
7285 6 : assert!(Lsn(0x80) < *timeline.get_latest_gc_cutoff_lsn());
7286 6 : let res = timeline.make_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx);
7287 6 : assert!(res.is_err());
7288 6 :
7289 6 : // Should still be able to renew a currently valid lease
7290 6 : // Assumption: original lease to is still valid for 0/50.
7291 6 : let _ =
7292 6 : timeline.make_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)?;
7293 6 :
7294 6 : Ok(())
7295 6 : }
7296 :
7297 : #[tokio::test]
7298 6 : async fn test_simple_bottom_most_compaction_deltas() -> anyhow::Result<()> {
7299 6 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_deltas").await?;
7300 24 : let (tenant, ctx) = harness.load().await;
7301 6 :
7302 354 : fn get_key(id: u32) -> Key {
7303 354 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7304 354 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7305 354 : key.field6 = id;
7306 354 : key
7307 354 : }
7308 6 :
7309 6 : // We create
7310 6 : // - one bottom-most image layer,
7311 6 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
7312 6 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
7313 6 : // - a delta layer D3 above the horizon.
7314 6 : //
7315 6 : // | D3 |
7316 6 : // | D1 |
7317 6 : // -| |-- gc horizon -----------------
7318 6 : // | | | D2 |
7319 6 : // --------- img layer ------------------
7320 6 : //
7321 6 : // What we should expact from this compaction is:
7322 6 : // | D3 |
7323 6 : // | Part of D1 |
7324 6 : // --------- img layer with D1+D2 at GC horizon------------------
7325 6 :
7326 6 : // img layer at 0x10
7327 6 : let img_layer = (0..10)
7328 60 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
7329 6 : .collect_vec();
7330 6 :
7331 6 : let delta1 = vec![
7332 6 : (
7333 6 : get_key(1),
7334 6 : Lsn(0x20),
7335 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7336 6 : ),
7337 6 : (
7338 6 : get_key(2),
7339 6 : Lsn(0x30),
7340 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7341 6 : ),
7342 6 : (
7343 6 : get_key(3),
7344 6 : Lsn(0x28),
7345 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
7346 6 : ),
7347 6 : (
7348 6 : get_key(3),
7349 6 : Lsn(0x30),
7350 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7351 6 : ),
7352 6 : (
7353 6 : get_key(3),
7354 6 : Lsn(0x40),
7355 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
7356 6 : ),
7357 6 : ];
7358 6 : let delta2 = vec![
7359 6 : (
7360 6 : get_key(5),
7361 6 : Lsn(0x20),
7362 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7363 6 : ),
7364 6 : (
7365 6 : get_key(6),
7366 6 : Lsn(0x20),
7367 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7368 6 : ),
7369 6 : ];
7370 6 : let delta3 = vec![
7371 6 : (
7372 6 : get_key(8),
7373 6 : Lsn(0x48),
7374 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
7375 6 : ),
7376 6 : (
7377 6 : get_key(9),
7378 6 : Lsn(0x48),
7379 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
7380 6 : ),
7381 6 : ];
7382 6 :
7383 6 : let tline = tenant
7384 6 : .create_test_timeline_with_layers(
7385 6 : TIMELINE_ID,
7386 6 : Lsn(0x10),
7387 6 : DEFAULT_PG_VERSION,
7388 6 : &ctx,
7389 6 : vec![
7390 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
7391 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
7392 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
7393 6 : ], // delta layers
7394 6 : vec![(Lsn(0x10), img_layer)], // image layers
7395 6 : Lsn(0x50),
7396 6 : )
7397 141 : .await?;
7398 6 : {
7399 6 : // Update GC info
7400 6 : let mut guard = tline.gc_info.write().unwrap();
7401 6 : *guard = GcInfo {
7402 6 : retain_lsns: vec![],
7403 6 : cutoffs: GcCutoffs {
7404 6 : time: Lsn(0x30),
7405 6 : space: Lsn(0x30),
7406 6 : },
7407 6 : leases: Default::default(),
7408 6 : within_ancestor_pitr: false,
7409 6 : };
7410 6 : }
7411 6 :
7412 6 : let expected_result = [
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@0x40"),
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@0x48"),
7422 6 : Bytes::from_static(b"value 9@0x10@0x48"),
7423 6 : ];
7424 6 :
7425 6 : let expected_result_at_gc_horizon = [
7426 6 : Bytes::from_static(b"value 0@0x10"),
7427 6 : Bytes::from_static(b"value 1@0x10@0x20"),
7428 6 : Bytes::from_static(b"value 2@0x10@0x30"),
7429 6 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
7430 6 : Bytes::from_static(b"value 4@0x10"),
7431 6 : Bytes::from_static(b"value 5@0x10@0x20"),
7432 6 : Bytes::from_static(b"value 6@0x10@0x20"),
7433 6 : Bytes::from_static(b"value 7@0x10"),
7434 6 : Bytes::from_static(b"value 8@0x10"),
7435 6 : Bytes::from_static(b"value 9@0x10"),
7436 6 : ];
7437 6 :
7438 66 : for idx in 0..10 {
7439 60 : assert_eq!(
7440 60 : tline
7441 60 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7442 93 : .await
7443 60 : .unwrap(),
7444 60 : &expected_result[idx]
7445 6 : );
7446 60 : assert_eq!(
7447 60 : tline
7448 60 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
7449 45 : .await
7450 60 : .unwrap(),
7451 60 : &expected_result_at_gc_horizon[idx]
7452 6 : );
7453 6 : }
7454 6 :
7455 6 : let cancel = CancellationToken::new();
7456 6 : tline
7457 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7458 156 : .await
7459 6 : .unwrap();
7460 6 :
7461 66 : for idx in 0..10 {
7462 60 : assert_eq!(
7463 60 : tline
7464 60 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7465 63 : .await
7466 60 : .unwrap(),
7467 60 : &expected_result[idx]
7468 6 : );
7469 60 : assert_eq!(
7470 60 : tline
7471 60 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
7472 30 : .await
7473 60 : .unwrap(),
7474 60 : &expected_result_at_gc_horizon[idx]
7475 6 : );
7476 6 : }
7477 6 :
7478 6 : // increase GC horizon and compact again
7479 6 : {
7480 6 : // Update GC info
7481 6 : let mut guard = tline.gc_info.write().unwrap();
7482 6 : guard.cutoffs.time = Lsn(0x40);
7483 6 : guard.cutoffs.space = Lsn(0x40);
7484 6 : }
7485 6 : tline
7486 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7487 126 : .await
7488 6 : .unwrap();
7489 6 :
7490 6 : Ok(())
7491 6 : }
7492 :
7493 : #[tokio::test]
7494 6 : async fn test_generate_key_retention() -> anyhow::Result<()> {
7495 6 : let harness = TenantHarness::create("test_generate_key_retention").await?;
7496 24 : let (tenant, ctx) = harness.load().await;
7497 6 : let tline = tenant
7498 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7499 12 : .await?;
7500 6 : tline.force_advance_lsn(Lsn(0x70));
7501 6 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7502 6 : let history = vec![
7503 6 : (
7504 6 : key,
7505 6 : Lsn(0x10),
7506 6 : Value::Image(Bytes::copy_from_slice(b"0x10")),
7507 6 : ),
7508 6 : (
7509 6 : key,
7510 6 : Lsn(0x20),
7511 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7512 6 : ),
7513 6 : (
7514 6 : key,
7515 6 : Lsn(0x30),
7516 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
7517 6 : ),
7518 6 : (
7519 6 : key,
7520 6 : Lsn(0x40),
7521 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
7522 6 : ),
7523 6 : (
7524 6 : key,
7525 6 : Lsn(0x50),
7526 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
7527 6 : ),
7528 6 : (
7529 6 : key,
7530 6 : Lsn(0x60),
7531 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7532 6 : ),
7533 6 : (
7534 6 : key,
7535 6 : Lsn(0x70),
7536 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7537 6 : ),
7538 6 : (
7539 6 : key,
7540 6 : Lsn(0x80),
7541 6 : Value::Image(Bytes::copy_from_slice(
7542 6 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
7543 6 : )),
7544 6 : ),
7545 6 : (
7546 6 : key,
7547 6 : Lsn(0x90),
7548 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
7549 6 : ),
7550 6 : ];
7551 6 : let res = tline
7552 6 : .generate_key_retention(
7553 6 : key,
7554 6 : &history,
7555 6 : Lsn(0x60),
7556 6 : &[Lsn(0x20), Lsn(0x40), Lsn(0x50)],
7557 6 : 3,
7558 6 : None,
7559 6 : )
7560 6 : .await
7561 6 : .unwrap();
7562 6 : let expected_res = KeyHistoryRetention {
7563 6 : below_horizon: vec![
7564 6 : (
7565 6 : Lsn(0x20),
7566 6 : KeyLogAtLsn(vec![(
7567 6 : Lsn(0x20),
7568 6 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20")),
7569 6 : )]),
7570 6 : ),
7571 6 : (
7572 6 : Lsn(0x40),
7573 6 : KeyLogAtLsn(vec![
7574 6 : (
7575 6 : Lsn(0x30),
7576 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
7577 6 : ),
7578 6 : (
7579 6 : Lsn(0x40),
7580 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
7581 6 : ),
7582 6 : ]),
7583 6 : ),
7584 6 : (
7585 6 : Lsn(0x50),
7586 6 : KeyLogAtLsn(vec![(
7587 6 : Lsn(0x50),
7588 6 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40;0x50")),
7589 6 : )]),
7590 6 : ),
7591 6 : (
7592 6 : Lsn(0x60),
7593 6 : KeyLogAtLsn(vec![(
7594 6 : Lsn(0x60),
7595 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7596 6 : )]),
7597 6 : ),
7598 6 : ],
7599 6 : above_horizon: KeyLogAtLsn(vec![
7600 6 : (
7601 6 : Lsn(0x70),
7602 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7603 6 : ),
7604 6 : (
7605 6 : Lsn(0x80),
7606 6 : Value::Image(Bytes::copy_from_slice(
7607 6 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
7608 6 : )),
7609 6 : ),
7610 6 : (
7611 6 : Lsn(0x90),
7612 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
7613 6 : ),
7614 6 : ]),
7615 6 : };
7616 6 : assert_eq!(res, expected_res);
7617 6 :
7618 6 : // We expect GC-compaction to run with the original GC. This would create a situation that
7619 6 : // the original GC algorithm removes some delta layers b/c there are full image coverage,
7620 6 : // therefore causing some keys to have an incomplete history below the lowest retain LSN.
7621 6 : // For example, we have
7622 6 : // ```plain
7623 6 : // init delta @ 0x10, image @ 0x20, delta @ 0x30 (gc_horizon), image @ 0x40.
7624 6 : // ```
7625 6 : // Now the GC horizon moves up, and we have
7626 6 : // ```plain
7627 6 : // init delta @ 0x10, image @ 0x20, delta @ 0x30, image @ 0x40 (gc_horizon)
7628 6 : // ```
7629 6 : // The original GC algorithm kicks in, and removes delta @ 0x10, image @ 0x20.
7630 6 : // We will end up with
7631 6 : // ```plain
7632 6 : // delta @ 0x30, image @ 0x40 (gc_horizon)
7633 6 : // ```
7634 6 : // Now we run the GC-compaction, and this key does not have a full history.
7635 6 : // We should be able to handle this partial history and drop everything before the
7636 6 : // gc_horizon image.
7637 6 :
7638 6 : let history = vec![
7639 6 : (
7640 6 : key,
7641 6 : Lsn(0x20),
7642 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7643 6 : ),
7644 6 : (
7645 6 : key,
7646 6 : Lsn(0x30),
7647 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
7648 6 : ),
7649 6 : (
7650 6 : key,
7651 6 : Lsn(0x40),
7652 6 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
7653 6 : ),
7654 6 : (
7655 6 : key,
7656 6 : Lsn(0x50),
7657 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
7658 6 : ),
7659 6 : (
7660 6 : key,
7661 6 : Lsn(0x60),
7662 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7663 6 : ),
7664 6 : (
7665 6 : key,
7666 6 : Lsn(0x70),
7667 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7668 6 : ),
7669 6 : (
7670 6 : key,
7671 6 : Lsn(0x80),
7672 6 : Value::Image(Bytes::copy_from_slice(
7673 6 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
7674 6 : )),
7675 6 : ),
7676 6 : (
7677 6 : key,
7678 6 : Lsn(0x90),
7679 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
7680 6 : ),
7681 6 : ];
7682 6 : let res = tline
7683 6 : .generate_key_retention(key, &history, Lsn(0x60), &[Lsn(0x40), Lsn(0x50)], 3, None)
7684 6 : .await
7685 6 : .unwrap();
7686 6 : let expected_res = KeyHistoryRetention {
7687 6 : below_horizon: vec![
7688 6 : (
7689 6 : Lsn(0x40),
7690 6 : KeyLogAtLsn(vec![(
7691 6 : Lsn(0x40),
7692 6 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
7693 6 : )]),
7694 6 : ),
7695 6 : (
7696 6 : Lsn(0x50),
7697 6 : KeyLogAtLsn(vec![(
7698 6 : Lsn(0x50),
7699 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
7700 6 : )]),
7701 6 : ),
7702 6 : (
7703 6 : Lsn(0x60),
7704 6 : KeyLogAtLsn(vec![(
7705 6 : Lsn(0x60),
7706 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7707 6 : )]),
7708 6 : ),
7709 6 : ],
7710 6 : above_horizon: KeyLogAtLsn(vec![
7711 6 : (
7712 6 : Lsn(0x70),
7713 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7714 6 : ),
7715 6 : (
7716 6 : Lsn(0x80),
7717 6 : Value::Image(Bytes::copy_from_slice(
7718 6 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
7719 6 : )),
7720 6 : ),
7721 6 : (
7722 6 : Lsn(0x90),
7723 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
7724 6 : ),
7725 6 : ]),
7726 6 : };
7727 6 : assert_eq!(res, expected_res);
7728 6 :
7729 6 : // In case of branch compaction, the branch itself does not have the full history, and we need to provide
7730 6 : // the ancestor image in the test case.
7731 6 :
7732 6 : let history = vec![
7733 6 : (
7734 6 : key,
7735 6 : Lsn(0x20),
7736 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7737 6 : ),
7738 6 : (
7739 6 : key,
7740 6 : Lsn(0x30),
7741 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
7742 6 : ),
7743 6 : (
7744 6 : key,
7745 6 : Lsn(0x40),
7746 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
7747 6 : ),
7748 6 : (
7749 6 : key,
7750 6 : Lsn(0x70),
7751 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7752 6 : ),
7753 6 : ];
7754 6 : let res = tline
7755 6 : .generate_key_retention(
7756 6 : key,
7757 6 : &history,
7758 6 : Lsn(0x60),
7759 6 : &[],
7760 6 : 3,
7761 6 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
7762 6 : )
7763 6 : .await
7764 6 : .unwrap();
7765 6 : let expected_res = KeyHistoryRetention {
7766 6 : below_horizon: vec![(
7767 6 : Lsn(0x60),
7768 6 : KeyLogAtLsn(vec![(
7769 6 : Lsn(0x60),
7770 6 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")), // use the ancestor image to reconstruct the page
7771 6 : )]),
7772 6 : )],
7773 6 : above_horizon: KeyLogAtLsn(vec![(
7774 6 : Lsn(0x70),
7775 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7776 6 : )]),
7777 6 : };
7778 6 : assert_eq!(res, expected_res);
7779 6 :
7780 6 : let history = vec![
7781 6 : (
7782 6 : key,
7783 6 : Lsn(0x20),
7784 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7785 6 : ),
7786 6 : (
7787 6 : key,
7788 6 : Lsn(0x40),
7789 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
7790 6 : ),
7791 6 : (
7792 6 : key,
7793 6 : Lsn(0x60),
7794 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7795 6 : ),
7796 6 : (
7797 6 : key,
7798 6 : Lsn(0x70),
7799 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7800 6 : ),
7801 6 : ];
7802 6 : let res = tline
7803 6 : .generate_key_retention(
7804 6 : key,
7805 6 : &history,
7806 6 : Lsn(0x60),
7807 6 : &[Lsn(0x30)],
7808 6 : 3,
7809 6 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
7810 6 : )
7811 6 : .await
7812 6 : .unwrap();
7813 6 : let expected_res = KeyHistoryRetention {
7814 6 : below_horizon: vec![
7815 6 : (
7816 6 : Lsn(0x30),
7817 6 : KeyLogAtLsn(vec![(
7818 6 : Lsn(0x20),
7819 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7820 6 : )]),
7821 6 : ),
7822 6 : (
7823 6 : Lsn(0x60),
7824 6 : KeyLogAtLsn(vec![(
7825 6 : Lsn(0x60),
7826 6 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x40;0x60")),
7827 6 : )]),
7828 6 : ),
7829 6 : ],
7830 6 : above_horizon: KeyLogAtLsn(vec![(
7831 6 : Lsn(0x70),
7832 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7833 6 : )]),
7834 6 : };
7835 6 : assert_eq!(res, expected_res);
7836 6 :
7837 6 : Ok(())
7838 6 : }
7839 :
7840 : #[tokio::test]
7841 6 : async fn test_simple_bottom_most_compaction_with_retain_lsns() -> anyhow::Result<()> {
7842 6 : let harness =
7843 6 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns").await?;
7844 24 : let (tenant, ctx) = harness.load().await;
7845 6 :
7846 1554 : fn get_key(id: u32) -> Key {
7847 1554 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7848 1554 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7849 1554 : key.field6 = id;
7850 1554 : key
7851 1554 : }
7852 6 :
7853 6 : let img_layer = (0..10)
7854 60 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
7855 6 : .collect_vec();
7856 6 :
7857 6 : let delta1 = vec![
7858 6 : (
7859 6 : get_key(1),
7860 6 : Lsn(0x20),
7861 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7862 6 : ),
7863 6 : (
7864 6 : get_key(2),
7865 6 : Lsn(0x30),
7866 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7867 6 : ),
7868 6 : (
7869 6 : get_key(3),
7870 6 : Lsn(0x28),
7871 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
7872 6 : ),
7873 6 : (
7874 6 : get_key(3),
7875 6 : Lsn(0x30),
7876 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7877 6 : ),
7878 6 : (
7879 6 : get_key(3),
7880 6 : Lsn(0x40),
7881 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
7882 6 : ),
7883 6 : ];
7884 6 : let delta2 = vec![
7885 6 : (
7886 6 : get_key(5),
7887 6 : Lsn(0x20),
7888 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7889 6 : ),
7890 6 : (
7891 6 : get_key(6),
7892 6 : Lsn(0x20),
7893 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7894 6 : ),
7895 6 : ];
7896 6 : let delta3 = vec![
7897 6 : (
7898 6 : get_key(8),
7899 6 : Lsn(0x48),
7900 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
7901 6 : ),
7902 6 : (
7903 6 : get_key(9),
7904 6 : Lsn(0x48),
7905 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
7906 6 : ),
7907 6 : ];
7908 6 :
7909 6 : let tline = tenant
7910 6 : .create_test_timeline_with_layers(
7911 6 : TIMELINE_ID,
7912 6 : Lsn(0x10),
7913 6 : DEFAULT_PG_VERSION,
7914 6 : &ctx,
7915 6 : vec![
7916 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
7917 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
7918 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
7919 6 : ], // delta layers
7920 6 : vec![(Lsn(0x10), img_layer)], // image layers
7921 6 : Lsn(0x50),
7922 6 : )
7923 141 : .await?;
7924 6 : {
7925 6 : // Update GC info
7926 6 : let mut guard = tline.gc_info.write().unwrap();
7927 6 : *guard = GcInfo {
7928 6 : retain_lsns: vec![
7929 6 : (Lsn(0x10), tline.timeline_id),
7930 6 : (Lsn(0x20), tline.timeline_id),
7931 6 : ],
7932 6 : cutoffs: GcCutoffs {
7933 6 : time: Lsn(0x30),
7934 6 : space: Lsn(0x30),
7935 6 : },
7936 6 : leases: Default::default(),
7937 6 : within_ancestor_pitr: false,
7938 6 : };
7939 6 : }
7940 6 :
7941 6 : let expected_result = [
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@0x40"),
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@0x48"),
7951 6 : Bytes::from_static(b"value 9@0x10@0x48"),
7952 6 : ];
7953 6 :
7954 6 : let expected_result_at_gc_horizon = [
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@0x30"),
7958 6 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
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_20 = [
7968 6 : Bytes::from_static(b"value 0@0x10"),
7969 6 : Bytes::from_static(b"value 1@0x10@0x20"),
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@0x20"),
7974 6 : Bytes::from_static(b"value 6@0x10@0x20"),
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 6 : let expected_result_at_lsn_10 = [
7981 6 : Bytes::from_static(b"value 0@0x10"),
7982 6 : Bytes::from_static(b"value 1@0x10"),
7983 6 : Bytes::from_static(b"value 2@0x10"),
7984 6 : Bytes::from_static(b"value 3@0x10"),
7985 6 : Bytes::from_static(b"value 4@0x10"),
7986 6 : Bytes::from_static(b"value 5@0x10"),
7987 6 : Bytes::from_static(b"value 6@0x10"),
7988 6 : Bytes::from_static(b"value 7@0x10"),
7989 6 : Bytes::from_static(b"value 8@0x10"),
7990 6 : Bytes::from_static(b"value 9@0x10"),
7991 6 : ];
7992 6 :
7993 36 : let verify_result = || async {
7994 36 : let gc_horizon = {
7995 36 : let gc_info = tline.gc_info.read().unwrap();
7996 36 : gc_info.cutoffs.time
7997 36 : };
7998 396 : for idx in 0..10 {
7999 360 : assert_eq!(
8000 360 : tline
8001 360 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8002 366 : .await
8003 360 : .unwrap(),
8004 360 : &expected_result[idx]
8005 36 : );
8006 360 : assert_eq!(
8007 360 : tline
8008 360 : .get(get_key(idx as u32), gc_horizon, &ctx)
8009 288 : .await
8010 360 : .unwrap(),
8011 360 : &expected_result_at_gc_horizon[idx]
8012 36 : );
8013 360 : assert_eq!(
8014 360 : tline
8015 360 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
8016 243 : .await
8017 360 : .unwrap(),
8018 360 : &expected_result_at_lsn_20[idx]
8019 36 : );
8020 360 : assert_eq!(
8021 360 : tline
8022 360 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
8023 189 : .await
8024 360 : .unwrap(),
8025 360 : &expected_result_at_lsn_10[idx]
8026 36 : );
8027 36 : }
8028 36 : };
8029 6 :
8030 210 : verify_result().await;
8031 6 :
8032 6 : let cancel = CancellationToken::new();
8033 6 : let mut dryrun_flags = EnumSet::new();
8034 6 : dryrun_flags.insert(CompactFlags::DryRun);
8035 6 :
8036 6 : tline
8037 6 : .compact_with_gc(&cancel, dryrun_flags, &ctx)
8038 129 : .await
8039 6 : .unwrap();
8040 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
8041 6 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
8042 171 : verify_result().await;
8043 6 :
8044 6 : tline
8045 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8046 150 : .await
8047 6 : .unwrap();
8048 192 : verify_result().await;
8049 6 :
8050 6 : // compact again
8051 6 : tline
8052 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8053 117 : .await
8054 6 : .unwrap();
8055 171 : verify_result().await;
8056 6 :
8057 6 : // increase GC horizon and compact again
8058 6 : {
8059 6 : // Update GC info
8060 6 : let mut guard = tline.gc_info.write().unwrap();
8061 6 : guard.cutoffs.time = Lsn(0x38);
8062 6 : guard.cutoffs.space = Lsn(0x38);
8063 6 : }
8064 6 : tline
8065 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8066 117 : .await
8067 6 : .unwrap();
8068 171 : verify_result().await; // no wals between 0x30 and 0x38, so we should obtain the same result
8069 6 :
8070 6 : // not increasing the GC horizon and compact again
8071 6 : tline
8072 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8073 117 : .await
8074 6 : .unwrap();
8075 171 : verify_result().await;
8076 6 :
8077 6 : Ok(())
8078 6 : }
8079 :
8080 : #[tokio::test]
8081 6 : async fn test_simple_bottom_most_compaction_with_retain_lsns_single_key() -> anyhow::Result<()>
8082 6 : {
8083 6 : let harness =
8084 6 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns_single_key")
8085 6 : .await?;
8086 24 : let (tenant, ctx) = harness.load().await;
8087 6 :
8088 1056 : fn get_key(id: u32) -> Key {
8089 1056 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8090 1056 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8091 1056 : key.field6 = id;
8092 1056 : key
8093 1056 : }
8094 6 :
8095 6 : let img_layer = (0..10)
8096 60 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8097 6 : .collect_vec();
8098 6 :
8099 6 : let delta1 = vec![
8100 6 : (
8101 6 : get_key(1),
8102 6 : Lsn(0x20),
8103 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8104 6 : ),
8105 6 : (
8106 6 : get_key(1),
8107 6 : Lsn(0x28),
8108 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8109 6 : ),
8110 6 : ];
8111 6 : let delta2 = vec![
8112 6 : (
8113 6 : get_key(1),
8114 6 : Lsn(0x30),
8115 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8116 6 : ),
8117 6 : (
8118 6 : get_key(1),
8119 6 : Lsn(0x38),
8120 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
8121 6 : ),
8122 6 : ];
8123 6 : let delta3 = vec![
8124 6 : (
8125 6 : get_key(8),
8126 6 : Lsn(0x48),
8127 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8128 6 : ),
8129 6 : (
8130 6 : get_key(9),
8131 6 : Lsn(0x48),
8132 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8133 6 : ),
8134 6 : ];
8135 6 :
8136 6 : let tline = tenant
8137 6 : .create_test_timeline_with_layers(
8138 6 : TIMELINE_ID,
8139 6 : Lsn(0x10),
8140 6 : DEFAULT_PG_VERSION,
8141 6 : &ctx,
8142 6 : vec![
8143 6 : // delta1 and delta 2 only contain a single key but multiple updates
8144 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x30), delta1),
8145 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
8146 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x50), delta3),
8147 6 : ], // delta layers
8148 6 : vec![(Lsn(0x10), img_layer)], // image layers
8149 6 : Lsn(0x50),
8150 6 : )
8151 141 : .await?;
8152 6 : {
8153 6 : // Update GC info
8154 6 : let mut guard = tline.gc_info.write().unwrap();
8155 6 : *guard = GcInfo {
8156 6 : retain_lsns: vec![
8157 6 : (Lsn(0x10), tline.timeline_id),
8158 6 : (Lsn(0x20), tline.timeline_id),
8159 6 : ],
8160 6 : cutoffs: GcCutoffs {
8161 6 : time: Lsn(0x30),
8162 6 : space: Lsn(0x30),
8163 6 : },
8164 6 : leases: Default::default(),
8165 6 : within_ancestor_pitr: false,
8166 6 : };
8167 6 : }
8168 6 :
8169 6 : let expected_result = [
8170 6 : Bytes::from_static(b"value 0@0x10"),
8171 6 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
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@0x48"),
8179 6 : Bytes::from_static(b"value 9@0x10@0x48"),
8180 6 : ];
8181 6 :
8182 6 : let expected_result_at_gc_horizon = [
8183 6 : Bytes::from_static(b"value 0@0x10"),
8184 6 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
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_20 = [
8196 6 : Bytes::from_static(b"value 0@0x10"),
8197 6 : Bytes::from_static(b"value 1@0x10@0x20"),
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 6 : let expected_result_at_lsn_10 = [
8209 6 : Bytes::from_static(b"value 0@0x10"),
8210 6 : Bytes::from_static(b"value 1@0x10"),
8211 6 : Bytes::from_static(b"value 2@0x10"),
8212 6 : Bytes::from_static(b"value 3@0x10"),
8213 6 : Bytes::from_static(b"value 4@0x10"),
8214 6 : Bytes::from_static(b"value 5@0x10"),
8215 6 : Bytes::from_static(b"value 6@0x10"),
8216 6 : Bytes::from_static(b"value 7@0x10"),
8217 6 : Bytes::from_static(b"value 8@0x10"),
8218 6 : Bytes::from_static(b"value 9@0x10"),
8219 6 : ];
8220 6 :
8221 24 : let verify_result = || async {
8222 24 : let gc_horizon = {
8223 24 : let gc_info = tline.gc_info.read().unwrap();
8224 24 : gc_info.cutoffs.time
8225 24 : };
8226 264 : for idx in 0..10 {
8227 240 : assert_eq!(
8228 240 : tline
8229 240 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8230 228 : .await
8231 240 : .unwrap(),
8232 240 : &expected_result[idx]
8233 24 : );
8234 240 : assert_eq!(
8235 240 : tline
8236 240 : .get(get_key(idx as u32), gc_horizon, &ctx)
8237 141 : .await
8238 240 : .unwrap(),
8239 240 : &expected_result_at_gc_horizon[idx]
8240 24 : );
8241 240 : assert_eq!(
8242 240 : tline
8243 240 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
8244 141 : .await
8245 240 : .unwrap(),
8246 240 : &expected_result_at_lsn_20[idx]
8247 24 : );
8248 240 : assert_eq!(
8249 240 : tline
8250 240 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
8251 120 : .await
8252 240 : .unwrap(),
8253 240 : &expected_result_at_lsn_10[idx]
8254 24 : );
8255 24 : }
8256 24 : };
8257 6 :
8258 183 : verify_result().await;
8259 6 :
8260 6 : let cancel = CancellationToken::new();
8261 6 : let mut dryrun_flags = EnumSet::new();
8262 6 : dryrun_flags.insert(CompactFlags::DryRun);
8263 6 :
8264 6 : tline
8265 6 : .compact_with_gc(&cancel, dryrun_flags, &ctx)
8266 135 : .await
8267 6 : .unwrap();
8268 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
8269 6 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
8270 147 : verify_result().await;
8271 6 :
8272 6 : tline
8273 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8274 153 : .await
8275 6 : .unwrap();
8276 159 : verify_result().await;
8277 6 :
8278 6 : // compact again
8279 6 : tline
8280 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8281 117 : .await
8282 6 : .unwrap();
8283 141 : verify_result().await;
8284 6 :
8285 6 : Ok(())
8286 6 : }
8287 :
8288 : #[tokio::test]
8289 6 : async fn test_simple_bottom_most_compaction_on_branch() -> anyhow::Result<()> {
8290 6 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_on_branch").await?;
8291 24 : let (tenant, ctx) = harness.load().await;
8292 6 :
8293 378 : fn get_key(id: u32) -> Key {
8294 378 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
8295 378 : key.field6 = id;
8296 378 : key
8297 378 : }
8298 6 :
8299 6 : let img_layer = (0..10)
8300 60 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8301 6 : .collect_vec();
8302 6 :
8303 6 : let delta1 = vec![
8304 6 : (
8305 6 : get_key(1),
8306 6 : Lsn(0x20),
8307 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8308 6 : ),
8309 6 : (
8310 6 : get_key(2),
8311 6 : Lsn(0x30),
8312 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8313 6 : ),
8314 6 : (
8315 6 : get_key(3),
8316 6 : Lsn(0x28),
8317 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8318 6 : ),
8319 6 : (
8320 6 : get_key(3),
8321 6 : Lsn(0x30),
8322 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8323 6 : ),
8324 6 : (
8325 6 : get_key(3),
8326 6 : Lsn(0x40),
8327 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
8328 6 : ),
8329 6 : ];
8330 6 : let delta2 = vec![
8331 6 : (
8332 6 : get_key(5),
8333 6 : Lsn(0x20),
8334 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8335 6 : ),
8336 6 : (
8337 6 : get_key(6),
8338 6 : Lsn(0x20),
8339 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8340 6 : ),
8341 6 : ];
8342 6 : let delta3 = vec![
8343 6 : (
8344 6 : get_key(8),
8345 6 : Lsn(0x48),
8346 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8347 6 : ),
8348 6 : (
8349 6 : get_key(9),
8350 6 : Lsn(0x48),
8351 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8352 6 : ),
8353 6 : ];
8354 6 :
8355 6 : let parent_tline = tenant
8356 6 : .create_test_timeline_with_layers(
8357 6 : TIMELINE_ID,
8358 6 : Lsn(0x10),
8359 6 : DEFAULT_PG_VERSION,
8360 6 : &ctx,
8361 6 : vec![], // delta layers
8362 6 : vec![(Lsn(0x18), img_layer)], // image layers
8363 6 : Lsn(0x18),
8364 6 : )
8365 87 : .await?;
8366 6 :
8367 6 : parent_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
8368 6 :
8369 6 : let branch_tline = tenant
8370 6 : .branch_timeline_test_with_layers(
8371 6 : &parent_tline,
8372 6 : NEW_TIMELINE_ID,
8373 6 : Some(Lsn(0x18)),
8374 6 : &ctx,
8375 6 : vec![
8376 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
8377 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
8378 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
8379 6 : ], // delta layers
8380 6 : vec![], // image layers
8381 6 : Lsn(0x50),
8382 6 : )
8383 54 : .await?;
8384 6 :
8385 6 : branch_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
8386 6 :
8387 6 : {
8388 6 : // Update GC info
8389 6 : let mut guard = parent_tline.gc_info.write().unwrap();
8390 6 : *guard = GcInfo {
8391 6 : retain_lsns: vec![(Lsn(0x18), branch_tline.timeline_id)],
8392 6 : cutoffs: GcCutoffs {
8393 6 : time: Lsn(0x10),
8394 6 : space: Lsn(0x10),
8395 6 : },
8396 6 : leases: Default::default(),
8397 6 : within_ancestor_pitr: false,
8398 6 : };
8399 6 : }
8400 6 :
8401 6 : {
8402 6 : // Update GC info
8403 6 : let mut guard = branch_tline.gc_info.write().unwrap();
8404 6 : *guard = GcInfo {
8405 6 : retain_lsns: vec![(Lsn(0x40), branch_tline.timeline_id)],
8406 6 : cutoffs: GcCutoffs {
8407 6 : time: Lsn(0x50),
8408 6 : space: Lsn(0x50),
8409 6 : },
8410 6 : leases: Default::default(),
8411 6 : within_ancestor_pitr: false,
8412 6 : };
8413 6 : }
8414 6 :
8415 6 : let expected_result_at_gc_horizon = [
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@0x48"),
8425 6 : Bytes::from_static(b"value 9@0x10@0x48"),
8426 6 : ];
8427 6 :
8428 6 : let expected_result_at_lsn_40 = [
8429 6 : Bytes::from_static(b"value 0@0x10"),
8430 6 : Bytes::from_static(b"value 1@0x10@0x20"),
8431 6 : Bytes::from_static(b"value 2@0x10@0x30"),
8432 6 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
8433 6 : Bytes::from_static(b"value 4@0x10"),
8434 6 : Bytes::from_static(b"value 5@0x10@0x20"),
8435 6 : Bytes::from_static(b"value 6@0x10@0x20"),
8436 6 : Bytes::from_static(b"value 7@0x10"),
8437 6 : Bytes::from_static(b"value 8@0x10"),
8438 6 : Bytes::from_static(b"value 9@0x10"),
8439 6 : ];
8440 6 :
8441 12 : let verify_result = || async {
8442 132 : for idx in 0..10 {
8443 120 : assert_eq!(
8444 120 : branch_tline
8445 120 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8446 156 : .await
8447 120 : .unwrap(),
8448 120 : &expected_result_at_gc_horizon[idx]
8449 12 : );
8450 120 : assert_eq!(
8451 120 : branch_tline
8452 120 : .get(get_key(idx as u32), Lsn(0x40), &ctx)
8453 93 : .await
8454 120 : .unwrap(),
8455 120 : &expected_result_at_lsn_40[idx]
8456 12 : );
8457 12 : }
8458 12 : };
8459 6 :
8460 138 : verify_result().await;
8461 6 :
8462 6 : let cancel = CancellationToken::new();
8463 6 : branch_tline
8464 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8465 48 : .await
8466 6 : .unwrap();
8467 6 :
8468 111 : verify_result().await;
8469 6 :
8470 6 : Ok(())
8471 6 : }
8472 : }
|
