Line data Source code
1 : //! Timeline repository implementation that keeps old data in layer files, and
2 : //! the recent changes in ephemeral files.
3 : //!
4 : //! See tenant/*_layer.rs files. The functions here are responsible for locating
5 : //! the correct layer for the get/put call, walking back the timeline branching
6 : //! history as needed.
7 : //!
8 : //! The files are stored in the .neon/tenants/<tenant_id>/timelines/<timeline_id>
9 : //! directory. See docs/pageserver-storage.md for how the files are managed.
10 : //! In addition to the layer files, there is a metadata file in the same
11 : //! directory that contains information about the timeline, in particular its
12 : //! parent timeline, and the last LSN that has been written to disk.
13 : //!
14 :
15 : use anyhow::{bail, Context};
16 : use arc_swap::ArcSwap;
17 : use camino::Utf8Path;
18 : use camino::Utf8PathBuf;
19 : use enumset::EnumSet;
20 : use futures::stream::FuturesUnordered;
21 : use futures::FutureExt;
22 : use futures::StreamExt;
23 : use pageserver_api::models;
24 : use pageserver_api::models::AuxFilePolicy;
25 : use pageserver_api::models::TimelineArchivalState;
26 : use pageserver_api::models::TimelineState;
27 : use pageserver_api::models::TopTenantShardItem;
28 : use pageserver_api::models::WalRedoManagerStatus;
29 : use pageserver_api::shard::ShardIdentity;
30 : use pageserver_api::shard::ShardStripeSize;
31 : use pageserver_api::shard::TenantShardId;
32 : use remote_storage::DownloadError;
33 : use remote_storage::GenericRemoteStorage;
34 : use remote_storage::TimeoutOrCancel;
35 : use std::collections::BTreeMap;
36 : use std::fmt;
37 : use std::sync::Weak;
38 : use std::time::SystemTime;
39 : use storage_broker::BrokerClientChannel;
40 : use tokio::io::BufReader;
41 : use tokio::sync::watch;
42 : use tokio::task::JoinSet;
43 : use tokio_util::sync::CancellationToken;
44 : use tracing::*;
45 : use upload_queue::NotInitialized;
46 : use utils::backoff;
47 : use utils::circuit_breaker::CircuitBreaker;
48 : use utils::completion;
49 : use utils::crashsafe::path_with_suffix_extension;
50 : use utils::failpoint_support;
51 : use utils::fs_ext;
52 : use utils::pausable_failpoint;
53 : use utils::sync::gate::Gate;
54 : use utils::sync::gate::GateGuard;
55 : use utils::timeout::timeout_cancellable;
56 : use utils::timeout::TimeoutCancellableError;
57 : use utils::zstd::create_zst_tarball;
58 : use utils::zstd::extract_zst_tarball;
59 :
60 : use self::config::AttachedLocationConfig;
61 : use self::config::AttachmentMode;
62 : use self::config::LocationConf;
63 : use self::config::TenantConf;
64 : use self::metadata::TimelineMetadata;
65 : use self::mgr::GetActiveTenantError;
66 : use self::mgr::GetTenantError;
67 : use self::remote_timeline_client::upload::upload_index_part;
68 : use self::remote_timeline_client::RemoteTimelineClient;
69 : use self::timeline::uninit::TimelineCreateGuard;
70 : use self::timeline::uninit::TimelineExclusionError;
71 : use self::timeline::uninit::UninitializedTimeline;
72 : use self::timeline::EvictionTaskTenantState;
73 : use self::timeline::GcCutoffs;
74 : use self::timeline::TimelineResources;
75 : use self::timeline::WaitLsnError;
76 : use crate::config::PageServerConf;
77 : use crate::context::{DownloadBehavior, RequestContext};
78 : use crate::deletion_queue::DeletionQueueClient;
79 : use crate::deletion_queue::DeletionQueueError;
80 : use crate::import_datadir;
81 : use crate::is_uninit_mark;
82 : use crate::l0_flush::L0FlushGlobalState;
83 : use crate::metrics::TENANT;
84 : use crate::metrics::{
85 : remove_tenant_metrics, BROKEN_TENANTS_SET, CIRCUIT_BREAKERS_BROKEN, CIRCUIT_BREAKERS_UNBROKEN,
86 : TENANT_STATE_METRIC, TENANT_SYNTHETIC_SIZE_METRIC,
87 : };
88 : use crate::repository::GcResult;
89 : use crate::task_mgr;
90 : use crate::task_mgr::TaskKind;
91 : use crate::tenant::config::LocationMode;
92 : use crate::tenant::config::TenantConfOpt;
93 : pub use crate::tenant::remote_timeline_client::index::IndexPart;
94 : use crate::tenant::remote_timeline_client::remote_initdb_archive_path;
95 : use crate::tenant::remote_timeline_client::MaybeDeletedIndexPart;
96 : use crate::tenant::remote_timeline_client::INITDB_PATH;
97 : use crate::tenant::storage_layer::DeltaLayer;
98 : use crate::tenant::storage_layer::ImageLayer;
99 : use crate::walredo;
100 : use crate::InitializationOrder;
101 : use std::collections::hash_map::Entry;
102 : use std::collections::HashMap;
103 : use std::collections::HashSet;
104 : use std::fmt::Debug;
105 : use std::fmt::Display;
106 : use std::fs;
107 : use std::fs::File;
108 : use std::sync::atomic::{AtomicU64, Ordering};
109 : use std::sync::Arc;
110 : use std::sync::Mutex;
111 : use std::time::{Duration, Instant};
112 :
113 : use crate::span;
114 : use crate::tenant::timeline::delete::DeleteTimelineFlow;
115 : use crate::tenant::timeline::uninit::cleanup_timeline_directory;
116 : use crate::virtual_file::VirtualFile;
117 : use crate::walredo::PostgresRedoManager;
118 : use crate::TEMP_FILE_SUFFIX;
119 : use once_cell::sync::Lazy;
120 : pub use pageserver_api::models::TenantState;
121 : use tokio::sync::Semaphore;
122 :
123 0 : static INIT_DB_SEMAPHORE: Lazy<Semaphore> = Lazy::new(|| Semaphore::new(8));
124 : use utils::{
125 : crashsafe,
126 : generation::Generation,
127 : id::TimelineId,
128 : lsn::{Lsn, RecordLsn},
129 : };
130 :
131 : pub mod blob_io;
132 : pub mod block_io;
133 : pub mod vectored_blob_io;
134 :
135 : pub mod disk_btree;
136 : pub(crate) mod ephemeral_file;
137 : pub mod layer_map;
138 :
139 : pub mod metadata;
140 : pub mod remote_timeline_client;
141 : pub mod storage_layer;
142 :
143 : pub mod checks;
144 : pub mod config;
145 : pub mod mgr;
146 : pub mod secondary;
147 : pub mod tasks;
148 : pub mod upload_queue;
149 :
150 : pub(crate) mod timeline;
151 :
152 : pub mod size;
153 :
154 : mod gc_block;
155 : pub(crate) mod throttle;
156 :
157 : pub(crate) use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
158 : pub(crate) use timeline::{LogicalSizeCalculationCause, PageReconstructError, Timeline};
159 :
160 : // re-export for use in walreceiver
161 : pub use crate::tenant::timeline::WalReceiverInfo;
162 :
163 : /// The "tenants" part of `tenants/<tenant>/timelines...`
164 : pub const TENANTS_SEGMENT_NAME: &str = "tenants";
165 :
166 : /// Parts of the `.neon/tenants/<tenant_id>/timelines/<timeline_id>` directory prefix.
167 : pub const TIMELINES_SEGMENT_NAME: &str = "timelines";
168 :
169 : /// References to shared objects that are passed into each tenant, such
170 : /// as the shared remote storage client and process initialization state.
171 : #[derive(Clone)]
172 : pub struct TenantSharedResources {
173 : pub broker_client: storage_broker::BrokerClientChannel,
174 : pub remote_storage: GenericRemoteStorage,
175 : pub deletion_queue_client: DeletionQueueClient,
176 : pub l0_flush_global_state: L0FlushGlobalState,
177 : }
178 :
179 : /// A [`Tenant`] is really an _attached_ tenant. The configuration
180 : /// for an attached tenant is a subset of the [`LocationConf`], represented
181 : /// in this struct.
182 : pub(super) struct AttachedTenantConf {
183 : tenant_conf: TenantConfOpt,
184 : location: AttachedLocationConfig,
185 : }
186 :
187 : impl AttachedTenantConf {
188 0 : fn new(tenant_conf: TenantConfOpt, location: AttachedLocationConfig) -> Self {
189 0 : Self {
190 0 : tenant_conf,
191 0 : location,
192 0 : }
193 0 : }
194 :
195 594 : fn try_from(location_conf: LocationConf) -> anyhow::Result<Self> {
196 594 : match &location_conf.mode {
197 594 : LocationMode::Attached(attach_conf) => Ok(Self {
198 594 : tenant_conf: location_conf.tenant_conf,
199 594 : location: *attach_conf,
200 594 : }),
201 : LocationMode::Secondary(_) => {
202 0 : anyhow::bail!("Attempted to construct AttachedTenantConf from a LocationConf in secondary mode")
203 : }
204 : }
205 594 : }
206 : }
207 : struct TimelinePreload {
208 : timeline_id: TimelineId,
209 : client: RemoteTimelineClient,
210 : index_part: Result<MaybeDeletedIndexPart, DownloadError>,
211 : }
212 :
213 : pub(crate) struct TenantPreload {
214 : timelines: HashMap<TimelineId, TimelinePreload>,
215 : }
216 :
217 : /// When we spawn a tenant, there is a special mode for tenant creation that
218 : /// avoids trying to read anything from remote storage.
219 : pub(crate) enum SpawnMode {
220 : /// Activate as soon as possible
221 : Eager,
222 : /// Lazy activation in the background, with the option to skip the queue if the need comes up
223 : Lazy,
224 : }
225 :
226 : ///
227 : /// Tenant consists of multiple timelines. Keep them in a hash table.
228 : ///
229 : pub struct Tenant {
230 : // Global pageserver config parameters
231 : pub conf: &'static PageServerConf,
232 :
233 : /// The value creation timestamp, used to measure activation delay, see:
234 : /// <https://github.com/neondatabase/neon/issues/4025>
235 : constructed_at: Instant,
236 :
237 : state: watch::Sender<TenantState>,
238 :
239 : // Overridden tenant-specific config parameters.
240 : // We keep TenantConfOpt sturct here to preserve the information
241 : // about parameters that are not set.
242 : // This is necessary to allow global config updates.
243 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
244 :
245 : tenant_shard_id: TenantShardId,
246 :
247 : // The detailed sharding information, beyond the number/count in tenant_shard_id
248 : shard_identity: ShardIdentity,
249 :
250 : /// The remote storage generation, used to protect S3 objects from split-brain.
251 : /// Does not change over the lifetime of the [`Tenant`] object.
252 : ///
253 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
254 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
255 : generation: Generation,
256 :
257 : timelines: Mutex<HashMap<TimelineId, Arc<Timeline>>>,
258 :
259 : /// During timeline creation, we first insert the TimelineId to the
260 : /// creating map, then `timelines`, then remove it from the creating map.
261 : /// **Lock order**: if acquring both, acquire`timelines` before `timelines_creating`
262 : timelines_creating: std::sync::Mutex<HashSet<TimelineId>>,
263 :
264 : // This mutex prevents creation of new timelines during GC.
265 : // Adding yet another mutex (in addition to `timelines`) is needed because holding
266 : // `timelines` mutex during all GC iteration
267 : // may block for a long time `get_timeline`, `get_timelines_state`,... and other operations
268 : // with timelines, which in turn may cause dropping replication connection, expiration of wait_for_lsn
269 : // timeout...
270 : gc_cs: tokio::sync::Mutex<()>,
271 : walredo_mgr: Option<Arc<WalRedoManager>>,
272 :
273 : // provides access to timeline data sitting in the remote storage
274 : pub(crate) remote_storage: GenericRemoteStorage,
275 :
276 : // Access to global deletion queue for when this tenant wants to schedule a deletion
277 : deletion_queue_client: DeletionQueueClient,
278 :
279 : /// Cached logical sizes updated updated on each [`Tenant::gather_size_inputs`].
280 : cached_logical_sizes: tokio::sync::Mutex<HashMap<(TimelineId, Lsn), u64>>,
281 : cached_synthetic_tenant_size: Arc<AtomicU64>,
282 :
283 : eviction_task_tenant_state: tokio::sync::Mutex<EvictionTaskTenantState>,
284 :
285 : /// Track repeated failures to compact, so that we can back off.
286 : /// Overhead of mutex is acceptable because compaction is done with a multi-second period.
287 : compaction_circuit_breaker: std::sync::Mutex<CircuitBreaker>,
288 :
289 : /// If the tenant is in Activating state, notify this to encourage it
290 : /// to proceed to Active as soon as possible, rather than waiting for lazy
291 : /// background warmup.
292 : pub(crate) activate_now_sem: tokio::sync::Semaphore,
293 :
294 : // Cancellation token fires when we have entered shutdown(). This is a parent of
295 : // Timelines' cancellation token.
296 : pub(crate) cancel: CancellationToken,
297 :
298 : // Users of the Tenant such as the page service must take this Gate to avoid
299 : // trying to use a Tenant which is shutting down.
300 : pub(crate) gate: Gate,
301 :
302 : /// Throttle applied at the top of [`Timeline::get`].
303 : /// All [`Tenant::timelines`] of a given [`Tenant`] instance share the same [`throttle::Throttle`] instance.
304 : pub(crate) timeline_get_throttle:
305 : Arc<throttle::Throttle<&'static crate::metrics::tenant_throttling::TimelineGet>>,
306 :
307 : /// An ongoing timeline detach concurrency limiter.
308 : ///
309 : /// As a tenant will likely be restarted as part of timeline detach ancestor it makes no sense
310 : /// to have two running at the same time. A different one can be started if an earlier one
311 : /// has failed for whatever reason.
312 : ongoing_timeline_detach: std::sync::Mutex<Option<(TimelineId, utils::completion::Barrier)>>,
313 :
314 : /// `index_part.json` based gc blocking reason tracking.
315 : ///
316 : /// New gc iterations must start a new iteration by acquiring `GcBlock::start` before
317 : /// proceeding.
318 : pub(crate) gc_block: gc_block::GcBlock,
319 :
320 : l0_flush_global_state: L0FlushGlobalState,
321 : }
322 :
323 : impl std::fmt::Debug for Tenant {
324 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
325 0 : write!(f, "{} ({})", self.tenant_shard_id, self.current_state())
326 0 : }
327 : }
328 :
329 : pub(crate) enum WalRedoManager {
330 : Prod(WalredoManagerId, PostgresRedoManager),
331 : #[cfg(test)]
332 : Test(harness::TestRedoManager),
333 : }
334 :
335 0 : #[derive(thiserror::Error, Debug)]
336 : #[error("pageserver is shutting down")]
337 : pub(crate) struct GlobalShutDown;
338 :
339 : impl WalRedoManager {
340 0 : pub(crate) fn new(mgr: PostgresRedoManager) -> Result<Arc<Self>, GlobalShutDown> {
341 0 : let id = WalredoManagerId::next();
342 0 : let arc = Arc::new(Self::Prod(id, mgr));
343 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
344 0 : match &mut *guard {
345 0 : Some(map) => {
346 0 : map.insert(id, Arc::downgrade(&arc));
347 0 : Ok(arc)
348 : }
349 0 : None => Err(GlobalShutDown),
350 : }
351 0 : }
352 : }
353 :
354 : impl Drop for WalRedoManager {
355 30 : fn drop(&mut self) {
356 30 : match self {
357 0 : Self::Prod(id, _) => {
358 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
359 0 : if let Some(map) = &mut *guard {
360 0 : map.remove(id).expect("new() registers, drop() unregisters");
361 0 : }
362 : }
363 : #[cfg(test)]
364 30 : Self::Test(_) => {
365 30 : // Not applicable to test redo manager
366 30 : }
367 : }
368 30 : }
369 : }
370 :
371 : /// Global registry of all walredo managers so that [`crate::shutdown_pageserver`] can shut down
372 : /// the walredo processes outside of the regular order.
373 : ///
374 : /// This is necessary to work around a systemd bug where it freezes if there are
375 : /// walredo processes left => <https://github.com/neondatabase/cloud/issues/11387>
376 : #[allow(clippy::type_complexity)]
377 : pub(crate) static WALREDO_MANAGERS: once_cell::sync::Lazy<
378 : Mutex<Option<HashMap<WalredoManagerId, Weak<WalRedoManager>>>>,
379 0 : > = once_cell::sync::Lazy::new(|| Mutex::new(Some(HashMap::new())));
380 : #[derive(PartialEq, Eq, Hash, Clone, Copy, Debug)]
381 : pub(crate) struct WalredoManagerId(u64);
382 : impl WalredoManagerId {
383 0 : pub fn next() -> Self {
384 : static NEXT: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(1);
385 0 : let id = NEXT.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
386 0 : if id == 0 {
387 0 : panic!("WalredoManagerId::new() returned 0, indicating wraparound, risking it's no longer unique");
388 0 : }
389 0 : Self(id)
390 0 : }
391 : }
392 :
393 : #[cfg(test)]
394 : impl From<harness::TestRedoManager> for WalRedoManager {
395 570 : fn from(mgr: harness::TestRedoManager) -> Self {
396 570 : Self::Test(mgr)
397 570 : }
398 : }
399 :
400 : impl WalRedoManager {
401 18 : pub(crate) async fn shutdown(&self) -> bool {
402 18 : match self {
403 0 : Self::Prod(_, mgr) => mgr.shutdown().await,
404 : #[cfg(test)]
405 : Self::Test(_) => {
406 : // Not applicable to test redo manager
407 18 : true
408 : }
409 : }
410 18 : }
411 :
412 0 : pub(crate) fn maybe_quiesce(&self, idle_timeout: Duration) {
413 0 : match self {
414 0 : Self::Prod(_, mgr) => mgr.maybe_quiesce(idle_timeout),
415 0 : #[cfg(test)]
416 0 : Self::Test(_) => {
417 0 : // Not applicable to test redo manager
418 0 : }
419 0 : }
420 0 : }
421 :
422 : /// # Cancel-Safety
423 : ///
424 : /// This method is cancellation-safe.
425 1038 : pub async fn request_redo(
426 1038 : &self,
427 1038 : key: crate::repository::Key,
428 1038 : lsn: Lsn,
429 1038 : base_img: Option<(Lsn, bytes::Bytes)>,
430 1038 : records: Vec<(Lsn, crate::walrecord::NeonWalRecord)>,
431 1038 : pg_version: u32,
432 1038 : ) -> Result<bytes::Bytes, walredo::Error> {
433 1038 : match self {
434 0 : Self::Prod(_, mgr) => {
435 0 : mgr.request_redo(key, lsn, base_img, records, pg_version)
436 0 : .await
437 : }
438 : #[cfg(test)]
439 1038 : Self::Test(mgr) => {
440 1038 : mgr.request_redo(key, lsn, base_img, records, pg_version)
441 0 : .await
442 : }
443 : }
444 1038 : }
445 :
446 0 : pub(crate) fn status(&self) -> Option<WalRedoManagerStatus> {
447 0 : match self {
448 0 : WalRedoManager::Prod(_, m) => Some(m.status()),
449 0 : #[cfg(test)]
450 0 : WalRedoManager::Test(_) => None,
451 0 : }
452 0 : }
453 : }
454 :
455 0 : #[derive(Debug, thiserror::Error, PartialEq, Eq)]
456 : pub enum GetTimelineError {
457 : #[error("Timeline is shutting down")]
458 : ShuttingDown,
459 : #[error("Timeline {tenant_id}/{timeline_id} is not active, state: {state:?}")]
460 : NotActive {
461 : tenant_id: TenantShardId,
462 : timeline_id: TimelineId,
463 : state: TimelineState,
464 : },
465 : #[error("Timeline {tenant_id}/{timeline_id} was not found")]
466 : NotFound {
467 : tenant_id: TenantShardId,
468 : timeline_id: TimelineId,
469 : },
470 : }
471 :
472 0 : #[derive(Debug, thiserror::Error)]
473 : pub enum LoadLocalTimelineError {
474 : #[error("FailedToLoad")]
475 : Load(#[source] anyhow::Error),
476 : #[error("FailedToResumeDeletion")]
477 : ResumeDeletion(#[source] anyhow::Error),
478 : }
479 :
480 0 : #[derive(thiserror::Error)]
481 : pub enum DeleteTimelineError {
482 : #[error("NotFound")]
483 : NotFound,
484 :
485 : #[error("HasChildren")]
486 : HasChildren(Vec<TimelineId>),
487 :
488 : #[error("Timeline deletion is already in progress")]
489 : AlreadyInProgress(Arc<tokio::sync::Mutex<DeleteTimelineFlow>>),
490 :
491 : #[error(transparent)]
492 : Other(#[from] anyhow::Error),
493 : }
494 :
495 : impl Debug for DeleteTimelineError {
496 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
497 0 : match self {
498 0 : Self::NotFound => write!(f, "NotFound"),
499 0 : Self::HasChildren(c) => f.debug_tuple("HasChildren").field(c).finish(),
500 0 : Self::AlreadyInProgress(_) => f.debug_tuple("AlreadyInProgress").finish(),
501 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
502 : }
503 0 : }
504 : }
505 :
506 0 : #[derive(thiserror::Error)]
507 : pub enum TimelineArchivalError {
508 : #[error("NotFound")]
509 : NotFound,
510 :
511 : #[error("Timeout")]
512 : Timeout,
513 :
514 : #[error("ancestor is archived: {}", .0)]
515 : HasArchivedParent(TimelineId),
516 :
517 : #[error("HasUnarchivedChildren")]
518 : HasUnarchivedChildren(Vec<TimelineId>),
519 :
520 : #[error("Timeline archival is already in progress")]
521 : AlreadyInProgress,
522 :
523 : #[error(transparent)]
524 : Other(#[from] anyhow::Error),
525 : }
526 :
527 : impl Debug for TimelineArchivalError {
528 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
529 0 : match self {
530 0 : Self::NotFound => write!(f, "NotFound"),
531 0 : Self::Timeout => write!(f, "Timeout"),
532 0 : Self::HasArchivedParent(p) => f.debug_tuple("HasArchivedParent").field(p).finish(),
533 0 : Self::HasUnarchivedChildren(c) => {
534 0 : f.debug_tuple("HasUnarchivedChildren").field(c).finish()
535 : }
536 0 : Self::AlreadyInProgress => f.debug_tuple("AlreadyInProgress").finish(),
537 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
538 : }
539 0 : }
540 : }
541 :
542 : pub enum SetStoppingError {
543 : AlreadyStopping(completion::Barrier),
544 : Broken,
545 : }
546 :
547 : impl Debug for SetStoppingError {
548 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
549 0 : match self {
550 0 : Self::AlreadyStopping(_) => f.debug_tuple("AlreadyStopping").finish(),
551 0 : Self::Broken => write!(f, "Broken"),
552 : }
553 0 : }
554 : }
555 :
556 0 : #[derive(thiserror::Error, Debug)]
557 : pub enum CreateTimelineError {
558 : #[error("creation of timeline with the given ID is in progress")]
559 : AlreadyCreating,
560 : #[error("timeline already exists with different parameters")]
561 : Conflict,
562 : #[error(transparent)]
563 : AncestorLsn(anyhow::Error),
564 : #[error("ancestor timeline is not active")]
565 : AncestorNotActive,
566 : #[error("tenant shutting down")]
567 : ShuttingDown,
568 : #[error(transparent)]
569 : Other(#[from] anyhow::Error),
570 : }
571 :
572 : #[derive(thiserror::Error, Debug)]
573 : enum InitdbError {
574 : Other(anyhow::Error),
575 : Cancelled,
576 : Spawn(std::io::Result<()>),
577 : Failed(std::process::ExitStatus, Vec<u8>),
578 : }
579 :
580 : impl fmt::Display for InitdbError {
581 0 : fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
582 0 : match self {
583 0 : InitdbError::Cancelled => write!(f, "Operation was cancelled"),
584 0 : InitdbError::Spawn(e) => write!(f, "Spawn error: {:?}", e),
585 0 : InitdbError::Failed(status, stderr) => write!(
586 0 : f,
587 0 : "Command failed with status {:?}: {}",
588 0 : status,
589 0 : String::from_utf8_lossy(stderr)
590 0 : ),
591 0 : InitdbError::Other(e) => write!(f, "Error: {:?}", e),
592 : }
593 0 : }
594 : }
595 :
596 : impl From<std::io::Error> for InitdbError {
597 0 : fn from(error: std::io::Error) -> Self {
598 0 : InitdbError::Spawn(Err(error))
599 0 : }
600 : }
601 :
602 : enum CreateTimelineCause {
603 : Load,
604 : Delete,
605 : }
606 :
607 0 : #[derive(thiserror::Error, Debug)]
608 : pub(crate) enum GcError {
609 : // The tenant is shutting down
610 : #[error("tenant shutting down")]
611 : TenantCancelled,
612 :
613 : // The tenant is shutting down
614 : #[error("timeline shutting down")]
615 : TimelineCancelled,
616 :
617 : // The tenant is in a state inelegible to run GC
618 : #[error("not active")]
619 : NotActive,
620 :
621 : // A requested GC cutoff LSN was invalid, for example it tried to move backwards
622 : #[error("not active")]
623 : BadLsn { why: String },
624 :
625 : // A remote storage error while scheduling updates after compaction
626 : #[error(transparent)]
627 : Remote(anyhow::Error),
628 :
629 : // An error reading while calculating GC cutoffs
630 : #[error(transparent)]
631 : GcCutoffs(PageReconstructError),
632 :
633 : // If GC was invoked for a particular timeline, this error means it didn't exist
634 : #[error("timeline not found")]
635 : TimelineNotFound,
636 : }
637 :
638 : impl From<PageReconstructError> for GcError {
639 0 : fn from(value: PageReconstructError) -> Self {
640 0 : match value {
641 0 : PageReconstructError::Cancelled => Self::TimelineCancelled,
642 0 : other => Self::GcCutoffs(other),
643 : }
644 0 : }
645 : }
646 :
647 : impl From<NotInitialized> for GcError {
648 0 : fn from(value: NotInitialized) -> Self {
649 0 : match value {
650 0 : NotInitialized::Uninitialized => GcError::Remote(value.into()),
651 0 : NotInitialized::Stopped | NotInitialized::ShuttingDown => GcError::TimelineCancelled,
652 : }
653 0 : }
654 : }
655 :
656 : impl From<timeline::layer_manager::Shutdown> for GcError {
657 0 : fn from(_: timeline::layer_manager::Shutdown) -> Self {
658 0 : GcError::TimelineCancelled
659 0 : }
660 : }
661 :
662 0 : #[derive(thiserror::Error, Debug)]
663 : pub(crate) enum LoadConfigError {
664 : #[error("TOML deserialization error: '{0}'")]
665 : DeserializeToml(#[from] toml_edit::de::Error),
666 :
667 : #[error("Config not found at {0}")]
668 : NotFound(Utf8PathBuf),
669 : }
670 :
671 : impl Tenant {
672 : /// Yet another helper for timeline initialization.
673 : ///
674 : /// - Initializes the Timeline struct and inserts it into the tenant's hash map
675 : /// - Scans the local timeline directory for layer files and builds the layer map
676 : /// - Downloads remote index file and adds remote files to the layer map
677 : /// - Schedules remote upload tasks for any files that are present locally but missing from remote storage.
678 : ///
679 : /// If the operation fails, the timeline is left in the tenant's hash map in Broken state. On success,
680 : /// it is marked as Active.
681 : #[allow(clippy::too_many_arguments)]
682 18 : async fn timeline_init_and_sync(
683 18 : &self,
684 18 : timeline_id: TimelineId,
685 18 : resources: TimelineResources,
686 18 : index_part: Option<IndexPart>,
687 18 : metadata: TimelineMetadata,
688 18 : ancestor: Option<Arc<Timeline>>,
689 18 : last_aux_file_policy: Option<AuxFilePolicy>,
690 18 : _ctx: &RequestContext,
691 18 : ) -> anyhow::Result<()> {
692 18 : let tenant_id = self.tenant_shard_id;
693 :
694 18 : let timeline = self.create_timeline_struct(
695 18 : timeline_id,
696 18 : &metadata,
697 18 : ancestor.clone(),
698 18 : resources,
699 18 : CreateTimelineCause::Load,
700 18 : // This could be derived from ancestor branch + index part. Though the only caller of `timeline_init_and_sync` is `load_remote_timeline`,
701 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.
702 18 : // Therefore, we pass this field explicitly for now, and remove it once we fully migrate to aux file v2.
703 18 : last_aux_file_policy,
704 18 : )?;
705 18 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
706 18 : anyhow::ensure!(
707 18 : disk_consistent_lsn.is_valid(),
708 0 : "Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn"
709 : );
710 18 : assert_eq!(
711 18 : disk_consistent_lsn,
712 18 : metadata.disk_consistent_lsn(),
713 0 : "these are used interchangeably"
714 : );
715 :
716 18 : if let Some(index_part) = index_part.as_ref() {
717 18 : timeline.remote_client.init_upload_queue(index_part)?;
718 :
719 18 : timeline
720 18 : .last_aux_file_policy
721 18 : .store(index_part.last_aux_file_policy());
722 : } else {
723 : // No data on the remote storage, but we have local metadata file. We can end up
724 : // here with timeline_create being interrupted before finishing index part upload.
725 : // By doing what we do here, the index part upload is retried.
726 : // If control plane retries timeline creation in the meantime, the mgmt API handler
727 : // for timeline creation will coalesce on the upload we queue here.
728 :
729 : // FIXME: this branch should be dead code as we no longer write local metadata.
730 :
731 0 : timeline
732 0 : .remote_client
733 0 : .init_upload_queue_for_empty_remote(&metadata)?;
734 0 : timeline
735 0 : .remote_client
736 0 : .schedule_index_upload_for_full_metadata_update(&metadata)?;
737 : }
738 :
739 18 : timeline
740 18 : .load_layer_map(disk_consistent_lsn, index_part)
741 17 : .await
742 18 : .with_context(|| {
743 0 : format!("Failed to load layermap for timeline {tenant_id}/{timeline_id}")
744 18 : })?;
745 :
746 : {
747 : // avoiding holding it across awaits
748 18 : let mut timelines_accessor = self.timelines.lock().unwrap();
749 18 : match timelines_accessor.entry(timeline_id) {
750 : // We should never try and load the same timeline twice during startup
751 : Entry::Occupied(_) => {
752 0 : unreachable!(
753 0 : "Timeline {tenant_id}/{timeline_id} already exists in the tenant map"
754 0 : );
755 : }
756 18 : Entry::Vacant(v) => {
757 18 : v.insert(Arc::clone(&timeline));
758 18 : timeline.maybe_spawn_flush_loop();
759 18 : }
760 18 : }
761 18 : };
762 18 :
763 18 : // Sanity check: a timeline should have some content.
764 18 : anyhow::ensure!(
765 18 : ancestor.is_some()
766 12 : || timeline
767 12 : .layers
768 12 : .read()
769 0 : .await
770 12 : .layer_map()
771 12 : .expect("currently loading, layer manager cannot be shutdown already")
772 12 : .iter_historic_layers()
773 12 : .next()
774 12 : .is_some(),
775 0 : "Timeline has no ancestor and no layer files"
776 : );
777 :
778 18 : Ok(())
779 18 : }
780 :
781 : /// Attach a tenant that's available in cloud storage.
782 : ///
783 : /// This returns quickly, after just creating the in-memory object
784 : /// Tenant struct and launching a background task to download
785 : /// the remote index files. On return, the tenant is most likely still in
786 : /// Attaching state, and it will become Active once the background task
787 : /// finishes. You can use wait_until_active() to wait for the task to
788 : /// complete.
789 : ///
790 : #[allow(clippy::too_many_arguments)]
791 0 : pub(crate) fn spawn(
792 0 : conf: &'static PageServerConf,
793 0 : tenant_shard_id: TenantShardId,
794 0 : resources: TenantSharedResources,
795 0 : attached_conf: AttachedTenantConf,
796 0 : shard_identity: ShardIdentity,
797 0 : init_order: Option<InitializationOrder>,
798 0 : mode: SpawnMode,
799 0 : ctx: &RequestContext,
800 0 : ) -> Result<Arc<Tenant>, GlobalShutDown> {
801 0 : let wal_redo_manager =
802 0 : WalRedoManager::new(PostgresRedoManager::new(conf, tenant_shard_id))?;
803 :
804 : let TenantSharedResources {
805 0 : broker_client,
806 0 : remote_storage,
807 0 : deletion_queue_client,
808 0 : l0_flush_global_state,
809 0 : } = resources;
810 0 :
811 0 : let attach_mode = attached_conf.location.attach_mode;
812 0 : let generation = attached_conf.location.generation;
813 0 :
814 0 : let tenant = Arc::new(Tenant::new(
815 0 : TenantState::Attaching,
816 0 : conf,
817 0 : attached_conf,
818 0 : shard_identity,
819 0 : Some(wal_redo_manager),
820 0 : tenant_shard_id,
821 0 : remote_storage.clone(),
822 0 : deletion_queue_client,
823 0 : l0_flush_global_state,
824 0 : ));
825 0 :
826 0 : // The attach task will carry a GateGuard, so that shutdown() reliably waits for it to drop out if
827 0 : // we shut down while attaching.
828 0 : let attach_gate_guard = tenant
829 0 : .gate
830 0 : .enter()
831 0 : .expect("We just created the Tenant: nothing else can have shut it down yet");
832 0 :
833 0 : // Do all the hard work in the background
834 0 : let tenant_clone = Arc::clone(&tenant);
835 0 : let ctx = ctx.detached_child(TaskKind::Attach, DownloadBehavior::Warn);
836 0 : task_mgr::spawn(
837 0 : &tokio::runtime::Handle::current(),
838 0 : TaskKind::Attach,
839 0 : tenant_shard_id,
840 0 : None,
841 0 : "attach tenant",
842 0 : async move {
843 0 :
844 0 : info!(
845 : ?attach_mode,
846 0 : "Attaching tenant"
847 : );
848 :
849 0 : let _gate_guard = attach_gate_guard;
850 0 :
851 0 : // Is this tenant being spawned as part of process startup?
852 0 : let starting_up = init_order.is_some();
853 0 : scopeguard::defer! {
854 0 : if starting_up {
855 0 : TENANT.startup_complete.inc();
856 0 : }
857 0 : }
858 :
859 : // Ideally we should use Tenant::set_broken_no_wait, but it is not supposed to be used when tenant is in loading state.
860 : enum BrokenVerbosity {
861 : Error,
862 : Info
863 : }
864 0 : let make_broken =
865 0 : |t: &Tenant, err: anyhow::Error, verbosity: BrokenVerbosity| {
866 0 : match verbosity {
867 : BrokenVerbosity::Info => {
868 0 : info!("attach cancelled, setting tenant state to Broken: {err}");
869 : },
870 : BrokenVerbosity::Error => {
871 0 : error!("attach failed, setting tenant state to Broken: {err:?}");
872 : }
873 : }
874 0 : t.state.send_modify(|state| {
875 0 : // The Stopping case is for when we have passed control on to DeleteTenantFlow:
876 0 : // if it errors, we will call make_broken when tenant is already in Stopping.
877 0 : assert!(
878 0 : matches!(*state, TenantState::Attaching | TenantState::Stopping { .. }),
879 0 : "the attach task owns the tenant state until activation is complete"
880 : );
881 :
882 0 : *state = TenantState::broken_from_reason(err.to_string());
883 0 : });
884 0 : };
885 :
886 : // TODO: should also be rejecting tenant conf changes that violate this check.
887 0 : if let Err(e) = crate::tenant::storage_layer::inmemory_layer::IndexEntry::validate_checkpoint_distance(tenant_clone.get_checkpoint_distance()) {
888 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
889 0 : return Ok(());
890 0 : }
891 0 :
892 0 : let mut init_order = init_order;
893 0 : // take the completion because initial tenant loading will complete when all of
894 0 : // these tasks complete.
895 0 : let _completion = init_order
896 0 : .as_mut()
897 0 : .and_then(|x| x.initial_tenant_load.take());
898 0 : let remote_load_completion = init_order
899 0 : .as_mut()
900 0 : .and_then(|x| x.initial_tenant_load_remote.take());
901 :
902 : enum AttachType<'a> {
903 : /// We are attaching this tenant lazily in the background.
904 : Warmup {
905 : _permit: tokio::sync::SemaphorePermit<'a>,
906 : during_startup: bool
907 : },
908 : /// We are attaching this tenant as soon as we can, because for example an
909 : /// endpoint tried to access it.
910 : OnDemand,
911 : /// During normal operations after startup, we are attaching a tenant, and
912 : /// eager attach was requested.
913 : Normal,
914 : }
915 :
916 0 : let attach_type = if matches!(mode, SpawnMode::Lazy) {
917 : // Before doing any I/O, wait for at least one of:
918 : // - A client attempting to access to this tenant (on-demand loading)
919 : // - A permit becoming available in the warmup semaphore (background warmup)
920 :
921 0 : tokio::select!(
922 0 : permit = tenant_clone.activate_now_sem.acquire() => {
923 0 : let _ = permit.expect("activate_now_sem is never closed");
924 0 : tracing::info!("Activating tenant (on-demand)");
925 0 : AttachType::OnDemand
926 : },
927 0 : permit = conf.concurrent_tenant_warmup.inner().acquire() => {
928 0 : let _permit = permit.expect("concurrent_tenant_warmup semaphore is never closed");
929 0 : tracing::info!("Activating tenant (warmup)");
930 0 : AttachType::Warmup {
931 0 : _permit,
932 0 : during_startup: init_order.is_some()
933 0 : }
934 : }
935 0 : _ = tenant_clone.cancel.cancelled() => {
936 : // This is safe, but should be pretty rare: it is interesting if a tenant
937 : // stayed in Activating for such a long time that shutdown found it in
938 : // that state.
939 0 : tracing::info!(state=%tenant_clone.current_state(), "Tenant shut down before activation");
940 : // Make the tenant broken so that set_stopping will not hang waiting for it to leave
941 : // the Attaching state. This is an over-reaction (nothing really broke, the tenant is
942 : // just shutting down), but ensures progress.
943 0 : make_broken(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"), BrokenVerbosity::Info);
944 0 : return Ok(());
945 : },
946 : )
947 : } else {
948 : // SpawnMode::{Create,Eager} always cause jumping ahead of the
949 : // concurrent_tenant_warmup queue
950 0 : AttachType::Normal
951 : };
952 :
953 0 : let preload = match &mode {
954 : SpawnMode::Eager | SpawnMode::Lazy => {
955 0 : let _preload_timer = TENANT.preload.start_timer();
956 0 : let res = tenant_clone
957 0 : .preload(&remote_storage, task_mgr::shutdown_token())
958 0 : .await;
959 0 : match res {
960 0 : Ok(p) => Some(p),
961 0 : Err(e) => {
962 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
963 0 : return Ok(());
964 : }
965 : }
966 : }
967 :
968 : };
969 :
970 : // Remote preload is complete.
971 0 : drop(remote_load_completion);
972 :
973 : // We will time the duration of the attach phase unless this is a creation (attach will do no work)
974 0 : let attached = {
975 0 : let _attach_timer = Some(TENANT.attach.start_timer());
976 0 : tenant_clone.attach(preload, &ctx).await
977 : };
978 :
979 0 : match attached {
980 : Ok(()) => {
981 0 : info!("attach finished, activating");
982 0 : tenant_clone.activate(broker_client, None, &ctx);
983 : }
984 0 : Err(e) => {
985 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
986 0 : }
987 : }
988 :
989 : // If we are doing an opportunistic warmup attachment at startup, initialize
990 : // logical size at the same time. This is better than starting a bunch of idle tenants
991 : // with cold caches and then coming back later to initialize their logical sizes.
992 : //
993 : // It also prevents the warmup proccess competing with the concurrency limit on
994 : // logical size calculations: if logical size calculation semaphore is saturated,
995 : // then warmup will wait for that before proceeding to the next tenant.
996 0 : if matches!(attach_type, AttachType::Warmup { during_startup: true, .. }) {
997 0 : let mut futs: FuturesUnordered<_> = tenant_clone.timelines.lock().unwrap().values().cloned().map(|t| t.await_initial_logical_size()).collect();
998 0 : tracing::info!("Waiting for initial logical sizes while warming up...");
999 0 : while futs.next().await.is_some() {}
1000 0 : tracing::info!("Warm-up complete");
1001 0 : }
1002 :
1003 0 : Ok(())
1004 0 : }
1005 0 : .instrument(tracing::info_span!(parent: None, "attach", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), gen=?generation)),
1006 : );
1007 0 : Ok(tenant)
1008 0 : }
1009 :
1010 570 : #[instrument(skip_all)]
1011 : pub(crate) async fn preload(
1012 : self: &Arc<Self>,
1013 : remote_storage: &GenericRemoteStorage,
1014 : cancel: CancellationToken,
1015 : ) -> anyhow::Result<TenantPreload> {
1016 : span::debug_assert_current_span_has_tenant_id();
1017 : // Get list of remote timelines
1018 : // download index files for every tenant timeline
1019 : info!("listing remote timelines");
1020 : let (remote_timeline_ids, other_keys) = remote_timeline_client::list_remote_timelines(
1021 : remote_storage,
1022 : self.tenant_shard_id,
1023 : cancel.clone(),
1024 : )
1025 : .await?;
1026 :
1027 : info!("found {} timelines", remote_timeline_ids.len(),);
1028 :
1029 : for k in other_keys {
1030 : warn!("Unexpected non timeline key {k}");
1031 : }
1032 :
1033 : Ok(TenantPreload {
1034 : timelines: Self::load_timeline_metadata(
1035 : self,
1036 : remote_timeline_ids,
1037 : remote_storage,
1038 : cancel,
1039 : )
1040 : .await?,
1041 : })
1042 : }
1043 :
1044 : ///
1045 : /// Background task that downloads all data for a tenant and brings it to Active state.
1046 : ///
1047 : /// No background tasks are started as part of this routine.
1048 : ///
1049 570 : async fn attach(
1050 570 : self: &Arc<Tenant>,
1051 570 : preload: Option<TenantPreload>,
1052 570 : ctx: &RequestContext,
1053 570 : ) -> anyhow::Result<()> {
1054 570 : span::debug_assert_current_span_has_tenant_id();
1055 570 :
1056 570 : failpoint_support::sleep_millis_async!("before-attaching-tenant");
1057 :
1058 570 : let Some(preload) = preload else {
1059 0 : anyhow::bail!("local-only deployment is no longer supported, https://github.com/neondatabase/neon/issues/5624");
1060 : };
1061 :
1062 570 : let mut timelines_to_resume_deletions = vec![];
1063 570 :
1064 570 : let mut remote_index_and_client = HashMap::new();
1065 570 : let mut timeline_ancestors = HashMap::new();
1066 570 : let mut existent_timelines = HashSet::new();
1067 588 : for (timeline_id, preload) in preload.timelines {
1068 18 : let index_part = match preload.index_part {
1069 18 : Ok(i) => {
1070 18 : debug!("remote index part exists for timeline {timeline_id}");
1071 : // We found index_part on the remote, this is the standard case.
1072 18 : existent_timelines.insert(timeline_id);
1073 18 : i
1074 : }
1075 : Err(DownloadError::NotFound) => {
1076 : // There is no index_part on the remote. We only get here
1077 : // if there is some prefix for the timeline in the remote storage.
1078 : // This can e.g. be the initdb.tar.zst archive, maybe a
1079 : // remnant from a prior incomplete creation or deletion attempt.
1080 : // Delete the local directory as the deciding criterion for a
1081 : // timeline's existence is presence of index_part.
1082 0 : info!(%timeline_id, "index_part not found on remote");
1083 0 : continue;
1084 : }
1085 0 : Err(e) => {
1086 0 : // Some (possibly ephemeral) error happened during index_part download.
1087 0 : // Pretend the timeline exists to not delete the timeline directory,
1088 0 : // as it might be a temporary issue and we don't want to re-download
1089 0 : // everything after it resolves.
1090 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1091 :
1092 0 : existent_timelines.insert(timeline_id);
1093 0 : continue;
1094 : }
1095 : };
1096 18 : match index_part {
1097 18 : MaybeDeletedIndexPart::IndexPart(index_part) => {
1098 18 : timeline_ancestors.insert(timeline_id, index_part.metadata.clone());
1099 18 : remote_index_and_client.insert(timeline_id, (index_part, preload.client));
1100 18 : }
1101 0 : MaybeDeletedIndexPart::Deleted(index_part) => {
1102 0 : info!(
1103 0 : "timeline {} is deleted, picking to resume deletion",
1104 : timeline_id
1105 : );
1106 0 : timelines_to_resume_deletions.push((timeline_id, index_part, preload.client));
1107 : }
1108 : }
1109 : }
1110 :
1111 570 : let mut gc_blocks = HashMap::new();
1112 :
1113 : // For every timeline, download the metadata file, scan the local directory,
1114 : // and build a layer map that contains an entry for each remote and local
1115 : // layer file.
1116 570 : let sorted_timelines = tree_sort_timelines(timeline_ancestors, |m| m.ancestor_timeline())?;
1117 588 : for (timeline_id, remote_metadata) in sorted_timelines {
1118 18 : let (index_part, remote_client) = remote_index_and_client
1119 18 : .remove(&timeline_id)
1120 18 : .expect("just put it in above");
1121 :
1122 18 : if let Some(blocking) = index_part.gc_blocking.as_ref() {
1123 : // could just filter these away, but it helps while testing
1124 0 : anyhow::ensure!(
1125 0 : !blocking.reasons.is_empty(),
1126 0 : "index_part for {timeline_id} is malformed: it should not have gc blocking with zero reasons"
1127 : );
1128 0 : let prev = gc_blocks.insert(timeline_id, blocking.reasons);
1129 0 : assert!(prev.is_none());
1130 18 : }
1131 :
1132 : // TODO again handle early failure
1133 18 : self.load_remote_timeline(
1134 18 : timeline_id,
1135 18 : index_part,
1136 18 : remote_metadata,
1137 18 : TimelineResources {
1138 18 : remote_client,
1139 18 : timeline_get_throttle: self.timeline_get_throttle.clone(),
1140 18 : l0_flush_global_state: self.l0_flush_global_state.clone(),
1141 18 : },
1142 18 : ctx,
1143 18 : )
1144 31 : .await
1145 18 : .with_context(|| {
1146 0 : format!(
1147 0 : "failed to load remote timeline {} for tenant {}",
1148 0 : timeline_id, self.tenant_shard_id
1149 0 : )
1150 18 : })?;
1151 : }
1152 :
1153 : // Walk through deleted timelines, resume deletion
1154 570 : for (timeline_id, index_part, remote_timeline_client) in timelines_to_resume_deletions {
1155 0 : remote_timeline_client
1156 0 : .init_upload_queue_stopped_to_continue_deletion(&index_part)
1157 0 : .context("init queue stopped")
1158 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1159 :
1160 0 : DeleteTimelineFlow::resume_deletion(
1161 0 : Arc::clone(self),
1162 0 : timeline_id,
1163 0 : &index_part.metadata,
1164 0 : remote_timeline_client,
1165 0 : )
1166 0 : .instrument(tracing::info_span!("timeline_delete", %timeline_id))
1167 0 : .await
1168 0 : .context("resume_deletion")
1169 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1170 : }
1171 :
1172 : // The local filesystem contents are a cache of what's in the remote IndexPart;
1173 : // IndexPart is the source of truth.
1174 570 : self.clean_up_timelines(&existent_timelines)?;
1175 :
1176 570 : self.gc_block.set_scanned(gc_blocks);
1177 570 :
1178 570 : fail::fail_point!("attach-before-activate", |_| {
1179 0 : anyhow::bail!("attach-before-activate");
1180 570 : });
1181 570 : failpoint_support::sleep_millis_async!("attach-before-activate-sleep", &self.cancel);
1182 :
1183 570 : info!("Done");
1184 :
1185 570 : Ok(())
1186 570 : }
1187 :
1188 : /// Check for any local timeline directories that are temporary, or do not correspond to a
1189 : /// timeline that still exists: this can happen if we crashed during a deletion/creation, or
1190 : /// if a timeline was deleted while the tenant was attached to a different pageserver.
1191 570 : fn clean_up_timelines(&self, existent_timelines: &HashSet<TimelineId>) -> anyhow::Result<()> {
1192 570 : let timelines_dir = self.conf.timelines_path(&self.tenant_shard_id);
1193 :
1194 570 : let entries = match timelines_dir.read_dir_utf8() {
1195 570 : Ok(d) => d,
1196 0 : Err(e) => {
1197 0 : if e.kind() == std::io::ErrorKind::NotFound {
1198 0 : return Ok(());
1199 : } else {
1200 0 : return Err(e).context("list timelines directory for tenant");
1201 : }
1202 : }
1203 : };
1204 :
1205 594 : for entry in entries {
1206 24 : let entry = entry.context("read timeline dir entry")?;
1207 24 : let entry_path = entry.path();
1208 :
1209 24 : let purge = if crate::is_temporary(entry_path)
1210 : // TODO: remove uninit mark code (https://github.com/neondatabase/neon/issues/5718)
1211 24 : || is_uninit_mark(entry_path)
1212 24 : || crate::is_delete_mark(entry_path)
1213 : {
1214 0 : true
1215 : } else {
1216 24 : match TimelineId::try_from(entry_path.file_name()) {
1217 24 : Ok(i) => {
1218 24 : // Purge if the timeline ID does not exist in remote storage: remote storage is the authority.
1219 24 : !existent_timelines.contains(&i)
1220 : }
1221 0 : Err(e) => {
1222 0 : tracing::warn!(
1223 0 : "Unparseable directory in timelines directory: {entry_path}, ignoring ({e})"
1224 : );
1225 : // Do not purge junk: if we don't recognize it, be cautious and leave it for a human.
1226 0 : false
1227 : }
1228 : }
1229 : };
1230 :
1231 24 : if purge {
1232 6 : tracing::info!("Purging stale timeline dentry {entry_path}");
1233 6 : if let Err(e) = match entry.file_type() {
1234 6 : Ok(t) => if t.is_dir() {
1235 6 : std::fs::remove_dir_all(entry_path)
1236 : } else {
1237 0 : std::fs::remove_file(entry_path)
1238 : }
1239 6 : .or_else(fs_ext::ignore_not_found),
1240 0 : Err(e) => Err(e),
1241 : } {
1242 0 : tracing::warn!("Failed to purge stale timeline dentry {entry_path}: {e}");
1243 6 : }
1244 18 : }
1245 : }
1246 :
1247 570 : Ok(())
1248 570 : }
1249 :
1250 : /// Get sum of all remote timelines sizes
1251 : ///
1252 : /// This function relies on the index_part instead of listing the remote storage
1253 0 : pub fn remote_size(&self) -> u64 {
1254 0 : let mut size = 0;
1255 :
1256 0 : for timeline in self.list_timelines() {
1257 0 : size += timeline.remote_client.get_remote_physical_size();
1258 0 : }
1259 :
1260 0 : size
1261 0 : }
1262 :
1263 18 : #[instrument(skip_all, fields(timeline_id=%timeline_id))]
1264 : async fn load_remote_timeline(
1265 : &self,
1266 : timeline_id: TimelineId,
1267 : index_part: IndexPart,
1268 : remote_metadata: TimelineMetadata,
1269 : resources: TimelineResources,
1270 : ctx: &RequestContext,
1271 : ) -> anyhow::Result<()> {
1272 : span::debug_assert_current_span_has_tenant_id();
1273 :
1274 : info!("downloading index file for timeline {}", timeline_id);
1275 : tokio::fs::create_dir_all(self.conf.timeline_path(&self.tenant_shard_id, &timeline_id))
1276 : .await
1277 : .context("Failed to create new timeline directory")?;
1278 :
1279 : let ancestor = if let Some(ancestor_id) = remote_metadata.ancestor_timeline() {
1280 : let timelines = self.timelines.lock().unwrap();
1281 : Some(Arc::clone(timelines.get(&ancestor_id).ok_or_else(
1282 0 : || {
1283 0 : anyhow::anyhow!(
1284 0 : "cannot find ancestor timeline {ancestor_id} for timeline {timeline_id}"
1285 0 : )
1286 0 : },
1287 : )?))
1288 : } else {
1289 : None
1290 : };
1291 :
1292 : let last_aux_file_policy = index_part.last_aux_file_policy();
1293 :
1294 : self.timeline_init_and_sync(
1295 : timeline_id,
1296 : resources,
1297 : Some(index_part),
1298 : remote_metadata,
1299 : ancestor,
1300 : last_aux_file_policy,
1301 : ctx,
1302 : )
1303 : .await
1304 : }
1305 :
1306 570 : async fn load_timeline_metadata(
1307 570 : self: &Arc<Tenant>,
1308 570 : timeline_ids: HashSet<TimelineId>,
1309 570 : remote_storage: &GenericRemoteStorage,
1310 570 : cancel: CancellationToken,
1311 570 : ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
1312 570 : let mut part_downloads = JoinSet::new();
1313 588 : for timeline_id in timeline_ids {
1314 18 : let client = RemoteTimelineClient::new(
1315 18 : remote_storage.clone(),
1316 18 : self.deletion_queue_client.clone(),
1317 18 : self.conf,
1318 18 : self.tenant_shard_id,
1319 18 : timeline_id,
1320 18 : self.generation,
1321 18 : );
1322 18 : let cancel_clone = cancel.clone();
1323 18 : part_downloads.spawn(
1324 18 : async move {
1325 18 : debug!("starting index part download");
1326 :
1327 68 : let index_part = client.download_index_file(&cancel_clone).await;
1328 :
1329 18 : debug!("finished index part download");
1330 :
1331 18 : Result::<_, anyhow::Error>::Ok(TimelinePreload {
1332 18 : client,
1333 18 : timeline_id,
1334 18 : index_part,
1335 18 : })
1336 18 : }
1337 18 : .map(move |res| {
1338 18 : res.with_context(|| format!("download index part for timeline {timeline_id}"))
1339 18 : })
1340 18 : .instrument(info_span!("download_index_part", %timeline_id)),
1341 : );
1342 : }
1343 :
1344 570 : let mut timeline_preloads: HashMap<TimelineId, TimelinePreload> = HashMap::new();
1345 :
1346 : loop {
1347 588 : tokio::select!(
1348 588 : next = part_downloads.join_next() => {
1349 588 : match next {
1350 18 : Some(result) => {
1351 18 : let preload_result = result.context("join preload task")?;
1352 18 : let preload = preload_result?;
1353 18 : timeline_preloads.insert(preload.timeline_id, preload);
1354 : },
1355 : None => {
1356 570 : break;
1357 : }
1358 : }
1359 : },
1360 588 : _ = cancel.cancelled() => {
1361 0 : anyhow::bail!("Cancelled while waiting for remote index download")
1362 : }
1363 : )
1364 : }
1365 :
1366 570 : Ok(timeline_preloads)
1367 570 : }
1368 :
1369 0 : pub(crate) async fn apply_timeline_archival_config(
1370 0 : &self,
1371 0 : timeline_id: TimelineId,
1372 0 : state: TimelineArchivalState,
1373 0 : ) -> Result<(), TimelineArchivalError> {
1374 0 : info!("setting timeline archival config");
1375 0 : let timeline = {
1376 0 : let timelines = self.timelines.lock().unwrap();
1377 :
1378 0 : let Some(timeline) = timelines.get(&timeline_id) else {
1379 0 : return Err(TimelineArchivalError::NotFound);
1380 : };
1381 :
1382 0 : if state == TimelineArchivalState::Unarchived {
1383 0 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
1384 0 : if ancestor_timeline.is_archived() == Some(true) {
1385 0 : return Err(TimelineArchivalError::HasArchivedParent(
1386 0 : ancestor_timeline.timeline_id,
1387 0 : ));
1388 0 : }
1389 0 : }
1390 0 : }
1391 :
1392 : // Ensure that there are no non-archived child timelines
1393 0 : let children: Vec<TimelineId> = timelines
1394 0 : .iter()
1395 0 : .filter_map(|(id, entry)| {
1396 0 : if entry.get_ancestor_timeline_id() != Some(timeline_id) {
1397 0 : return None;
1398 0 : }
1399 0 : if entry.is_archived() == Some(true) {
1400 0 : return None;
1401 0 : }
1402 0 : Some(*id)
1403 0 : })
1404 0 : .collect();
1405 0 :
1406 0 : if !children.is_empty() && state == TimelineArchivalState::Archived {
1407 0 : return Err(TimelineArchivalError::HasUnarchivedChildren(children));
1408 0 : }
1409 0 : Arc::clone(timeline)
1410 : };
1411 :
1412 0 : let upload_needed = timeline
1413 0 : .remote_client
1414 0 : .schedule_index_upload_for_timeline_archival_state(state)?;
1415 :
1416 0 : if upload_needed {
1417 0 : info!("Uploading new state");
1418 : const MAX_WAIT: Duration = Duration::from_secs(10);
1419 0 : let Ok(v) =
1420 0 : tokio::time::timeout(MAX_WAIT, timeline.remote_client.wait_completion()).await
1421 : else {
1422 0 : tracing::warn!("reached timeout for waiting on upload queue");
1423 0 : return Err(TimelineArchivalError::Timeout);
1424 : };
1425 0 : v.map_err(|e| TimelineArchivalError::Other(anyhow::anyhow!(e)))?;
1426 0 : }
1427 0 : Ok(())
1428 0 : }
1429 :
1430 12 : pub(crate) fn tenant_shard_id(&self) -> TenantShardId {
1431 12 : self.tenant_shard_id
1432 12 : }
1433 :
1434 : /// Get Timeline handle for given Neon timeline ID.
1435 : /// This function is idempotent. It doesn't change internal state in any way.
1436 666 : pub fn get_timeline(
1437 666 : &self,
1438 666 : timeline_id: TimelineId,
1439 666 : active_only: bool,
1440 666 : ) -> Result<Arc<Timeline>, GetTimelineError> {
1441 666 : let timelines_accessor = self.timelines.lock().unwrap();
1442 666 : let timeline = timelines_accessor
1443 666 : .get(&timeline_id)
1444 666 : .ok_or(GetTimelineError::NotFound {
1445 666 : tenant_id: self.tenant_shard_id,
1446 666 : timeline_id,
1447 666 : })?;
1448 :
1449 660 : if active_only && !timeline.is_active() {
1450 0 : Err(GetTimelineError::NotActive {
1451 0 : tenant_id: self.tenant_shard_id,
1452 0 : timeline_id,
1453 0 : state: timeline.current_state(),
1454 0 : })
1455 : } else {
1456 660 : Ok(Arc::clone(timeline))
1457 : }
1458 666 : }
1459 :
1460 : /// Lists timelines the tenant contains.
1461 : /// Up to tenant's implementation to omit certain timelines that ar not considered ready for use.
1462 24 : pub fn list_timelines(&self) -> Vec<Arc<Timeline>> {
1463 24 : self.timelines
1464 24 : .lock()
1465 24 : .unwrap()
1466 24 : .values()
1467 24 : .map(Arc::clone)
1468 24 : .collect()
1469 24 : }
1470 :
1471 0 : pub fn list_timeline_ids(&self) -> Vec<TimelineId> {
1472 0 : self.timelines.lock().unwrap().keys().cloned().collect()
1473 0 : }
1474 :
1475 : /// This is used to create the initial 'main' timeline during bootstrapping,
1476 : /// or when importing a new base backup. The caller is expected to load an
1477 : /// initial image of the datadir to the new timeline after this.
1478 : ///
1479 : /// Until that happens, the on-disk state is invalid (disk_consistent_lsn=Lsn(0))
1480 : /// and the timeline will fail to load at a restart.
1481 : ///
1482 : /// For tests, use `DatadirModification::init_empty_test_timeline` + `commit` to setup the
1483 : /// minimum amount of keys required to get a writable timeline.
1484 : /// (Without it, `put` might fail due to `repartition` failing.)
1485 546 : pub(crate) async fn create_empty_timeline(
1486 546 : &self,
1487 546 : new_timeline_id: TimelineId,
1488 546 : initdb_lsn: Lsn,
1489 546 : pg_version: u32,
1490 546 : _ctx: &RequestContext,
1491 546 : ) -> anyhow::Result<UninitializedTimeline> {
1492 546 : anyhow::ensure!(
1493 546 : self.is_active(),
1494 0 : "Cannot create empty timelines on inactive tenant"
1495 : );
1496 :
1497 : // Protect against concurrent attempts to use this TimelineId
1498 546 : let create_guard = self.create_timeline_create_guard(new_timeline_id)?;
1499 :
1500 540 : let new_metadata = TimelineMetadata::new(
1501 540 : // Initialize disk_consistent LSN to 0, The caller must import some data to
1502 540 : // make it valid, before calling finish_creation()
1503 540 : Lsn(0),
1504 540 : None,
1505 540 : None,
1506 540 : Lsn(0),
1507 540 : initdb_lsn,
1508 540 : initdb_lsn,
1509 540 : pg_version,
1510 540 : );
1511 540 : self.prepare_new_timeline(
1512 540 : new_timeline_id,
1513 540 : &new_metadata,
1514 540 : create_guard,
1515 540 : initdb_lsn,
1516 540 : None,
1517 540 : None,
1518 540 : )
1519 0 : .await
1520 546 : }
1521 :
1522 : /// Helper for unit tests to create an empty timeline.
1523 : ///
1524 : /// The timeline is has state value `Active` but its background loops are not running.
1525 : // This makes the various functions which anyhow::ensure! for Active state work in tests.
1526 : // Our current tests don't need the background loops.
1527 : #[cfg(test)]
1528 516 : pub async fn create_test_timeline(
1529 516 : &self,
1530 516 : new_timeline_id: TimelineId,
1531 516 : initdb_lsn: Lsn,
1532 516 : pg_version: u32,
1533 516 : ctx: &RequestContext,
1534 516 : ) -> anyhow::Result<Arc<Timeline>> {
1535 516 : let uninit_tl = self
1536 516 : .create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
1537 0 : .await?;
1538 516 : let tline = uninit_tl.raw_timeline().expect("we just created it");
1539 516 : assert_eq!(tline.get_last_record_lsn(), Lsn(0));
1540 :
1541 : // Setup minimum keys required for the timeline to be usable.
1542 516 : let mut modification = tline.begin_modification(initdb_lsn);
1543 516 : modification
1544 516 : .init_empty_test_timeline()
1545 516 : .context("init_empty_test_timeline")?;
1546 516 : modification
1547 516 : .commit(ctx)
1548 498 : .await
1549 516 : .context("commit init_empty_test_timeline modification")?;
1550 :
1551 : // Flush to disk so that uninit_tl's check for valid disk_consistent_lsn passes.
1552 516 : tline.maybe_spawn_flush_loop();
1553 516 : tline.freeze_and_flush().await.context("freeze_and_flush")?;
1554 :
1555 : // Make sure the freeze_and_flush reaches remote storage.
1556 516 : tline.remote_client.wait_completion().await.unwrap();
1557 :
1558 516 : let tl = uninit_tl.finish_creation()?;
1559 : // The non-test code would call tl.activate() here.
1560 516 : tl.set_state(TimelineState::Active);
1561 516 : Ok(tl)
1562 516 : }
1563 :
1564 : /// Helper for unit tests to create a timeline with some pre-loaded states.
1565 : #[cfg(test)]
1566 : #[allow(clippy::too_many_arguments)]
1567 84 : pub async fn create_test_timeline_with_layers(
1568 84 : &self,
1569 84 : new_timeline_id: TimelineId,
1570 84 : initdb_lsn: Lsn,
1571 84 : pg_version: u32,
1572 84 : ctx: &RequestContext,
1573 84 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
1574 84 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
1575 84 : end_lsn: Lsn,
1576 84 : ) -> anyhow::Result<Arc<Timeline>> {
1577 : use checks::check_valid_layermap;
1578 : use itertools::Itertools;
1579 :
1580 84 : let tline = self
1581 84 : .create_test_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
1582 168 : .await?;
1583 84 : tline.force_advance_lsn(end_lsn);
1584 246 : for deltas in delta_layer_desc {
1585 162 : tline
1586 162 : .force_create_delta_layer(deltas, Some(initdb_lsn), ctx)
1587 486 : .await?;
1588 : }
1589 222 : for (lsn, images) in image_layer_desc {
1590 138 : tline
1591 138 : .force_create_image_layer(lsn, images, Some(initdb_lsn), ctx)
1592 807 : .await?;
1593 : }
1594 84 : let layer_names = tline
1595 84 : .layers
1596 84 : .read()
1597 0 : .await
1598 84 : .layer_map()
1599 84 : .unwrap()
1600 84 : .iter_historic_layers()
1601 384 : .map(|layer| layer.layer_name())
1602 84 : .collect_vec();
1603 84 : if let Some(err) = check_valid_layermap(&layer_names) {
1604 0 : bail!("invalid layermap: {err}");
1605 84 : }
1606 84 : Ok(tline)
1607 84 : }
1608 :
1609 : /// Create a new timeline.
1610 : ///
1611 : /// Returns the new timeline ID and reference to its Timeline object.
1612 : ///
1613 : /// If the caller specified the timeline ID to use (`new_timeline_id`), and timeline with
1614 : /// the same timeline ID already exists, returns CreateTimelineError::AlreadyExists.
1615 : #[allow(clippy::too_many_arguments)]
1616 0 : pub(crate) async fn create_timeline(
1617 0 : self: &Arc<Tenant>,
1618 0 : new_timeline_id: TimelineId,
1619 0 : ancestor_timeline_id: Option<TimelineId>,
1620 0 : mut ancestor_start_lsn: Option<Lsn>,
1621 0 : pg_version: u32,
1622 0 : load_existing_initdb: Option<TimelineId>,
1623 0 : broker_client: storage_broker::BrokerClientChannel,
1624 0 : ctx: &RequestContext,
1625 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
1626 0 : if !self.is_active() {
1627 0 : if matches!(self.current_state(), TenantState::Stopping { .. }) {
1628 0 : return Err(CreateTimelineError::ShuttingDown);
1629 : } else {
1630 0 : return Err(CreateTimelineError::Other(anyhow::anyhow!(
1631 0 : "Cannot create timelines on inactive tenant"
1632 0 : )));
1633 : }
1634 0 : }
1635 :
1636 0 : let _gate = self
1637 0 : .gate
1638 0 : .enter()
1639 0 : .map_err(|_| CreateTimelineError::ShuttingDown)?;
1640 :
1641 : // Get exclusive access to the timeline ID: this ensures that it does not already exist,
1642 : // and that no other creation attempts will be allowed in while we are working.
1643 0 : let create_guard = match self.create_timeline_create_guard(new_timeline_id) {
1644 0 : Ok(m) => m,
1645 : Err(TimelineExclusionError::AlreadyCreating) => {
1646 : // Creation is in progress, we cannot create it again, and we cannot
1647 : // check if this request matches the existing one, so caller must try
1648 : // again later.
1649 0 : return Err(CreateTimelineError::AlreadyCreating);
1650 : }
1651 0 : Err(TimelineExclusionError::Other(e)) => {
1652 0 : return Err(CreateTimelineError::Other(e));
1653 : }
1654 0 : Err(TimelineExclusionError::AlreadyExists(existing)) => {
1655 0 : debug!("timeline {new_timeline_id} already exists");
1656 :
1657 : // Idempotency: creating the same timeline twice is not an error, unless
1658 : // the second creation has different parameters.
1659 0 : if existing.get_ancestor_timeline_id() != ancestor_timeline_id
1660 0 : || existing.pg_version != pg_version
1661 0 : || (ancestor_start_lsn.is_some()
1662 0 : && ancestor_start_lsn != Some(existing.get_ancestor_lsn()))
1663 : {
1664 0 : return Err(CreateTimelineError::Conflict);
1665 0 : }
1666 0 :
1667 0 : // Wait for uploads to complete, so that when we return Ok, the timeline
1668 0 : // is known to be durable on remote storage. Just like we do at the end of
1669 0 : // this function, after we have created the timeline ourselves.
1670 0 : //
1671 0 : // We only really care that the initial version of `index_part.json` has
1672 0 : // been uploaded. That's enough to remember that the timeline
1673 0 : // exists. However, there is no function to wait specifically for that so
1674 0 : // we just wait for all in-progress uploads to finish.
1675 0 : existing
1676 0 : .remote_client
1677 0 : .wait_completion()
1678 0 : .await
1679 0 : .context("wait for timeline uploads to complete")?;
1680 :
1681 0 : return Ok(existing);
1682 : }
1683 : };
1684 :
1685 0 : pausable_failpoint!("timeline-creation-after-uninit");
1686 :
1687 0 : let loaded_timeline = match ancestor_timeline_id {
1688 0 : Some(ancestor_timeline_id) => {
1689 0 : let ancestor_timeline = self
1690 0 : .get_timeline(ancestor_timeline_id, false)
1691 0 : .context("Cannot branch off the timeline that's not present in pageserver")?;
1692 :
1693 : // instead of waiting around, just deny the request because ancestor is not yet
1694 : // ready for other purposes either.
1695 0 : if !ancestor_timeline.is_active() {
1696 0 : return Err(CreateTimelineError::AncestorNotActive);
1697 0 : }
1698 :
1699 0 : if let Some(lsn) = ancestor_start_lsn.as_mut() {
1700 0 : *lsn = lsn.align();
1701 0 :
1702 0 : let ancestor_ancestor_lsn = ancestor_timeline.get_ancestor_lsn();
1703 0 : if ancestor_ancestor_lsn > *lsn {
1704 : // can we safely just branch from the ancestor instead?
1705 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
1706 0 : "invalid start lsn {} for ancestor timeline {}: less than timeline ancestor lsn {}",
1707 0 : lsn,
1708 0 : ancestor_timeline_id,
1709 0 : ancestor_ancestor_lsn,
1710 0 : )));
1711 0 : }
1712 0 :
1713 0 : // Wait for the WAL to arrive and be processed on the parent branch up
1714 0 : // to the requested branch point. The repository code itself doesn't
1715 0 : // require it, but if we start to receive WAL on the new timeline,
1716 0 : // decoding the new WAL might need to look up previous pages, relation
1717 0 : // sizes etc. and that would get confused if the previous page versions
1718 0 : // are not in the repository yet.
1719 0 : ancestor_timeline
1720 0 : .wait_lsn(*lsn, timeline::WaitLsnWaiter::Tenant, ctx)
1721 0 : .await
1722 0 : .map_err(|e| match e {
1723 0 : e @ (WaitLsnError::Timeout(_) | WaitLsnError::BadState { .. }) => {
1724 0 : CreateTimelineError::AncestorLsn(anyhow::anyhow!(e))
1725 : }
1726 0 : WaitLsnError::Shutdown => CreateTimelineError::ShuttingDown,
1727 0 : })?;
1728 0 : }
1729 :
1730 0 : self.branch_timeline(
1731 0 : &ancestor_timeline,
1732 0 : new_timeline_id,
1733 0 : ancestor_start_lsn,
1734 0 : create_guard,
1735 0 : ctx,
1736 0 : )
1737 0 : .await?
1738 : }
1739 : None => {
1740 0 : self.bootstrap_timeline(
1741 0 : new_timeline_id,
1742 0 : pg_version,
1743 0 : load_existing_initdb,
1744 0 : create_guard,
1745 0 : ctx,
1746 0 : )
1747 0 : .await?
1748 : }
1749 : };
1750 :
1751 : // At this point we have dropped our guard on [`Self::timelines_creating`], and
1752 : // the timeline is visible in [`Self::timelines`], but it is _not_ durable yet. We must
1753 : // not send a success to the caller until it is. The same applies to handling retries,
1754 : // see the handling of [`TimelineExclusionError::AlreadyExists`] above.
1755 0 : let kind = ancestor_timeline_id
1756 0 : .map(|_| "branched")
1757 0 : .unwrap_or("bootstrapped");
1758 0 : loaded_timeline
1759 0 : .remote_client
1760 0 : .wait_completion()
1761 0 : .await
1762 0 : .with_context(|| format!("wait for {} timeline initial uploads to complete", kind))?;
1763 :
1764 0 : loaded_timeline.activate(self.clone(), broker_client, None, ctx);
1765 0 :
1766 0 : Ok(loaded_timeline)
1767 0 : }
1768 :
1769 0 : pub(crate) async fn delete_timeline(
1770 0 : self: Arc<Self>,
1771 0 : timeline_id: TimelineId,
1772 0 : ) -> Result<(), DeleteTimelineError> {
1773 0 : DeleteTimelineFlow::run(&self, timeline_id).await?;
1774 :
1775 0 : Ok(())
1776 0 : }
1777 :
1778 : /// perform one garbage collection iteration, removing old data files from disk.
1779 : /// this function is periodically called by gc task.
1780 : /// also it can be explicitly requested through page server api 'do_gc' command.
1781 : ///
1782 : /// `target_timeline_id` specifies the timeline to GC, or None for all.
1783 : ///
1784 : /// The `horizon` an `pitr` parameters determine how much WAL history needs to be retained.
1785 : /// Also known as the retention period, or the GC cutoff point. `horizon` specifies
1786 : /// the amount of history, as LSN difference from current latest LSN on each timeline.
1787 : /// `pitr` specifies the same as a time difference from the current time. The effective
1788 : /// GC cutoff point is determined conservatively by either `horizon` and `pitr`, whichever
1789 : /// requires more history to be retained.
1790 : //
1791 2262 : pub(crate) async fn gc_iteration(
1792 2262 : &self,
1793 2262 : target_timeline_id: Option<TimelineId>,
1794 2262 : horizon: u64,
1795 2262 : pitr: Duration,
1796 2262 : cancel: &CancellationToken,
1797 2262 : ctx: &RequestContext,
1798 2262 : ) -> Result<GcResult, GcError> {
1799 2262 : // Don't start doing work during shutdown
1800 2262 : if let TenantState::Stopping { .. } = self.current_state() {
1801 0 : return Ok(GcResult::default());
1802 2262 : }
1803 2262 :
1804 2262 : // there is a global allowed_error for this
1805 2262 : if !self.is_active() {
1806 0 : return Err(GcError::NotActive);
1807 2262 : }
1808 2262 :
1809 2262 : {
1810 2262 : let conf = self.tenant_conf.load();
1811 2262 :
1812 2262 : if !conf.location.may_delete_layers_hint() {
1813 0 : info!("Skipping GC in location state {:?}", conf.location);
1814 0 : return Ok(GcResult::default());
1815 2262 : }
1816 : }
1817 :
1818 2262 : let _guard = match self.gc_block.start().await {
1819 2262 : Ok(guard) => guard,
1820 0 : Err(reasons) => {
1821 0 : info!("Skipping GC: {reasons}");
1822 0 : return Ok(GcResult::default());
1823 : }
1824 : };
1825 :
1826 2262 : self.gc_iteration_internal(target_timeline_id, horizon, pitr, cancel, ctx)
1827 2254 : .await
1828 2262 : }
1829 :
1830 : /// Perform one compaction iteration.
1831 : /// This function is periodically called by compactor task.
1832 : /// Also it can be explicitly requested per timeline through page server
1833 : /// api's 'compact' command.
1834 : ///
1835 : /// Returns whether we have pending compaction task.
1836 0 : async fn compaction_iteration(
1837 0 : &self,
1838 0 : cancel: &CancellationToken,
1839 0 : ctx: &RequestContext,
1840 0 : ) -> Result<bool, timeline::CompactionError> {
1841 0 : // Don't start doing work during shutdown, or when broken, we do not need those in the logs
1842 0 : if !self.is_active() {
1843 0 : return Ok(false);
1844 0 : }
1845 0 :
1846 0 : {
1847 0 : let conf = self.tenant_conf.load();
1848 0 : if !conf.location.may_delete_layers_hint() || !conf.location.may_upload_layers_hint() {
1849 0 : info!("Skipping compaction in location state {:?}", conf.location);
1850 0 : return Ok(false);
1851 0 : }
1852 0 : }
1853 0 :
1854 0 : // Scan through the hashmap and collect a list of all the timelines,
1855 0 : // while holding the lock. Then drop the lock and actually perform the
1856 0 : // compactions. We don't want to block everything else while the
1857 0 : // compaction runs.
1858 0 : let timelines_to_compact = {
1859 0 : let timelines = self.timelines.lock().unwrap();
1860 0 : let timelines_to_compact = timelines
1861 0 : .iter()
1862 0 : .filter_map(|(timeline_id, timeline)| {
1863 0 : if timeline.is_active() {
1864 0 : Some((*timeline_id, timeline.clone()))
1865 : } else {
1866 0 : None
1867 : }
1868 0 : })
1869 0 : .collect::<Vec<_>>();
1870 0 : drop(timelines);
1871 0 : timelines_to_compact
1872 0 : };
1873 0 :
1874 0 : // Before doing any I/O work, check our circuit breaker
1875 0 : if self.compaction_circuit_breaker.lock().unwrap().is_broken() {
1876 0 : info!("Skipping compaction due to previous failures");
1877 0 : return Ok(false);
1878 0 : }
1879 0 :
1880 0 : let mut has_pending_task = false;
1881 :
1882 0 : for (timeline_id, timeline) in &timelines_to_compact {
1883 0 : has_pending_task |= timeline
1884 0 : .compact(cancel, EnumSet::empty(), ctx)
1885 0 : .instrument(info_span!("compact_timeline", %timeline_id))
1886 0 : .await
1887 0 : .inspect_err(|e| match e {
1888 0 : timeline::CompactionError::ShuttingDown => (),
1889 0 : timeline::CompactionError::Other(e) => {
1890 0 : self.compaction_circuit_breaker
1891 0 : .lock()
1892 0 : .unwrap()
1893 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, e);
1894 0 : }
1895 0 : })?;
1896 : }
1897 :
1898 0 : self.compaction_circuit_breaker
1899 0 : .lock()
1900 0 : .unwrap()
1901 0 : .success(&CIRCUIT_BREAKERS_UNBROKEN);
1902 0 :
1903 0 : Ok(has_pending_task)
1904 0 : }
1905 :
1906 : // Call through to all timelines to freeze ephemeral layers if needed. Usually
1907 : // this happens during ingest: this background housekeeping is for freezing layers
1908 : // that are open but haven't been written to for some time.
1909 0 : async fn ingest_housekeeping(&self) {
1910 0 : // Scan through the hashmap and collect a list of all the timelines,
1911 0 : // while holding the lock. Then drop the lock and actually perform the
1912 0 : // compactions. We don't want to block everything else while the
1913 0 : // compaction runs.
1914 0 : let timelines = {
1915 0 : self.timelines
1916 0 : .lock()
1917 0 : .unwrap()
1918 0 : .values()
1919 0 : .filter_map(|timeline| {
1920 0 : if timeline.is_active() {
1921 0 : Some(timeline.clone())
1922 : } else {
1923 0 : None
1924 : }
1925 0 : })
1926 0 : .collect::<Vec<_>>()
1927 : };
1928 :
1929 0 : for timeline in &timelines {
1930 0 : timeline.maybe_freeze_ephemeral_layer().await;
1931 : }
1932 0 : }
1933 :
1934 7350 : pub fn current_state(&self) -> TenantState {
1935 7350 : self.state.borrow().clone()
1936 7350 : }
1937 :
1938 5070 : pub fn is_active(&self) -> bool {
1939 5070 : self.current_state() == TenantState::Active
1940 5070 : }
1941 :
1942 0 : pub fn generation(&self) -> Generation {
1943 0 : self.generation
1944 0 : }
1945 :
1946 0 : pub(crate) fn wal_redo_manager_status(&self) -> Option<WalRedoManagerStatus> {
1947 0 : self.walredo_mgr.as_ref().and_then(|mgr| mgr.status())
1948 0 : }
1949 :
1950 : /// Changes tenant status to active, unless shutdown was already requested.
1951 : ///
1952 : /// `background_jobs_can_start` is an optional barrier set to a value during pageserver startup
1953 : /// to delay background jobs. Background jobs can be started right away when None is given.
1954 0 : fn activate(
1955 0 : self: &Arc<Self>,
1956 0 : broker_client: BrokerClientChannel,
1957 0 : background_jobs_can_start: Option<&completion::Barrier>,
1958 0 : ctx: &RequestContext,
1959 0 : ) {
1960 0 : span::debug_assert_current_span_has_tenant_id();
1961 0 :
1962 0 : let mut activating = false;
1963 0 : self.state.send_modify(|current_state| {
1964 : use pageserver_api::models::ActivatingFrom;
1965 0 : match &*current_state {
1966 : TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
1967 0 : panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {state:?}", state = current_state);
1968 : }
1969 0 : TenantState::Loading => {
1970 0 : *current_state = TenantState::Activating(ActivatingFrom::Loading);
1971 0 : }
1972 0 : TenantState::Attaching => {
1973 0 : *current_state = TenantState::Activating(ActivatingFrom::Attaching);
1974 0 : }
1975 : }
1976 0 : debug!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), "Activating tenant");
1977 0 : activating = true;
1978 0 : // Continue outside the closure. We need to grab timelines.lock()
1979 0 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
1980 0 : });
1981 0 :
1982 0 : if activating {
1983 0 : let timelines_accessor = self.timelines.lock().unwrap();
1984 0 : let timelines_to_activate = timelines_accessor
1985 0 : .values()
1986 0 : .filter(|timeline| !(timeline.is_broken() || timeline.is_stopping()));
1987 0 :
1988 0 : // Before activation, populate each Timeline's GcInfo with information about its children
1989 0 : self.initialize_gc_info(&timelines_accessor);
1990 0 :
1991 0 : // Spawn gc and compaction loops. The loops will shut themselves
1992 0 : // down when they notice that the tenant is inactive.
1993 0 : tasks::start_background_loops(self, background_jobs_can_start);
1994 0 :
1995 0 : let mut activated_timelines = 0;
1996 :
1997 0 : for timeline in timelines_to_activate {
1998 0 : timeline.activate(
1999 0 : self.clone(),
2000 0 : broker_client.clone(),
2001 0 : background_jobs_can_start,
2002 0 : ctx,
2003 0 : );
2004 0 : activated_timelines += 1;
2005 0 : }
2006 :
2007 0 : self.state.send_modify(move |current_state| {
2008 0 : assert!(
2009 0 : matches!(current_state, TenantState::Activating(_)),
2010 0 : "set_stopping and set_broken wait for us to leave Activating state",
2011 : );
2012 0 : *current_state = TenantState::Active;
2013 0 :
2014 0 : let elapsed = self.constructed_at.elapsed();
2015 0 : let total_timelines = timelines_accessor.len();
2016 0 :
2017 0 : // log a lot of stuff, because some tenants sometimes suffer from user-visible
2018 0 : // times to activate. see https://github.com/neondatabase/neon/issues/4025
2019 0 : info!(
2020 0 : since_creation_millis = elapsed.as_millis(),
2021 0 : tenant_id = %self.tenant_shard_id.tenant_id,
2022 0 : shard_id = %self.tenant_shard_id.shard_slug(),
2023 0 : activated_timelines,
2024 0 : total_timelines,
2025 0 : post_state = <&'static str>::from(&*current_state),
2026 0 : "activation attempt finished"
2027 : );
2028 :
2029 0 : TENANT.activation.observe(elapsed.as_secs_f64());
2030 0 : });
2031 0 : }
2032 0 : }
2033 :
2034 : /// Shutdown the tenant and join all of the spawned tasks.
2035 : ///
2036 : /// The method caters for all use-cases:
2037 : /// - pageserver shutdown (freeze_and_flush == true)
2038 : /// - detach + ignore (freeze_and_flush == false)
2039 : ///
2040 : /// This will attempt to shutdown even if tenant is broken.
2041 : ///
2042 : /// `shutdown_progress` is a [`completion::Barrier`] for the shutdown initiated by this call.
2043 : /// If the tenant is already shutting down, we return a clone of the first shutdown call's
2044 : /// `Barrier` as an `Err`. This not-first caller can use the returned barrier to join with
2045 : /// the ongoing shutdown.
2046 18 : async fn shutdown(
2047 18 : &self,
2048 18 : shutdown_progress: completion::Barrier,
2049 18 : shutdown_mode: timeline::ShutdownMode,
2050 18 : ) -> Result<(), completion::Barrier> {
2051 18 : span::debug_assert_current_span_has_tenant_id();
2052 :
2053 : // Set tenant (and its timlines) to Stoppping state.
2054 : //
2055 : // Since we can only transition into Stopping state after activation is complete,
2056 : // run it in a JoinSet so all tenants have a chance to stop before we get SIGKILLed.
2057 : //
2058 : // Transitioning tenants to Stopping state has a couple of non-obvious side effects:
2059 : // 1. Lock out any new requests to the tenants.
2060 : // 2. Signal cancellation to WAL receivers (we wait on it below).
2061 : // 3. Signal cancellation for other tenant background loops.
2062 : // 4. ???
2063 : //
2064 : // The waiting for the cancellation is not done uniformly.
2065 : // We certainly wait for WAL receivers to shut down.
2066 : // That is necessary so that no new data comes in before the freeze_and_flush.
2067 : // But the tenant background loops are joined-on in our caller.
2068 : // It's mesed up.
2069 : // we just ignore the failure to stop
2070 :
2071 : // If we're still attaching, fire the cancellation token early to drop out: this
2072 : // will prevent us flushing, but ensures timely shutdown if some I/O during attach
2073 : // is very slow.
2074 18 : let shutdown_mode = if matches!(self.current_state(), TenantState::Attaching) {
2075 0 : self.cancel.cancel();
2076 0 :
2077 0 : // Having fired our cancellation token, do not try and flush timelines: their cancellation tokens
2078 0 : // are children of ours, so their flush loops will have shut down already
2079 0 : timeline::ShutdownMode::Hard
2080 : } else {
2081 18 : shutdown_mode
2082 : };
2083 :
2084 18 : match self.set_stopping(shutdown_progress, false, false).await {
2085 18 : Ok(()) => {}
2086 0 : Err(SetStoppingError::Broken) => {
2087 0 : // assume that this is acceptable
2088 0 : }
2089 0 : Err(SetStoppingError::AlreadyStopping(other)) => {
2090 0 : // give caller the option to wait for this this shutdown
2091 0 : info!("Tenant::shutdown: AlreadyStopping");
2092 0 : return Err(other);
2093 : }
2094 : };
2095 :
2096 18 : let mut js = tokio::task::JoinSet::new();
2097 18 : {
2098 18 : let timelines = self.timelines.lock().unwrap();
2099 18 : timelines.values().for_each(|timeline| {
2100 18 : let timeline = Arc::clone(timeline);
2101 18 : let timeline_id = timeline.timeline_id;
2102 18 : let span = tracing::info_span!("timeline_shutdown", %timeline_id, ?shutdown_mode);
2103 36 : js.spawn(async move { timeline.shutdown(shutdown_mode).instrument(span).await });
2104 18 : })
2105 18 : };
2106 18 : // test_long_timeline_create_then_tenant_delete is leaning on this message
2107 18 : tracing::info!("Waiting for timelines...");
2108 36 : while let Some(res) = js.join_next().await {
2109 0 : match res {
2110 18 : Ok(()) => {}
2111 0 : Err(je) if je.is_cancelled() => unreachable!("no cancelling used"),
2112 0 : Err(je) if je.is_panic() => { /* logged already */ }
2113 0 : Err(je) => warn!("unexpected JoinError: {je:?}"),
2114 : }
2115 : }
2116 :
2117 : // We cancel the Tenant's cancellation token _after_ the timelines have all shut down. This permits
2118 : // them to continue to do work during their shutdown methods, e.g. flushing data.
2119 18 : tracing::debug!("Cancelling CancellationToken");
2120 18 : self.cancel.cancel();
2121 18 :
2122 18 : // shutdown all tenant and timeline tasks: gc, compaction, page service
2123 18 : // No new tasks will be started for this tenant because it's in `Stopping` state.
2124 18 : //
2125 18 : // this will additionally shutdown and await all timeline tasks.
2126 18 : tracing::debug!("Waiting for tasks...");
2127 18 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), None).await;
2128 :
2129 18 : if let Some(walredo_mgr) = self.walredo_mgr.as_ref() {
2130 18 : walredo_mgr.shutdown().await;
2131 0 : }
2132 :
2133 : // Wait for any in-flight operations to complete
2134 18 : self.gate.close().await;
2135 :
2136 18 : remove_tenant_metrics(&self.tenant_shard_id);
2137 18 :
2138 18 : Ok(())
2139 18 : }
2140 :
2141 : /// Change tenant status to Stopping, to mark that it is being shut down.
2142 : ///
2143 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
2144 : ///
2145 : /// This function is not cancel-safe!
2146 : ///
2147 : /// `allow_transition_from_loading` is needed for the special case of loading task deleting the tenant.
2148 : /// `allow_transition_from_attaching` is needed for the special case of attaching deleted tenant.
2149 18 : async fn set_stopping(
2150 18 : &self,
2151 18 : progress: completion::Barrier,
2152 18 : allow_transition_from_loading: bool,
2153 18 : allow_transition_from_attaching: bool,
2154 18 : ) -> Result<(), SetStoppingError> {
2155 18 : let mut rx = self.state.subscribe();
2156 18 :
2157 18 : // cannot stop before we're done activating, so wait out until we're done activating
2158 18 : rx.wait_for(|state| match state {
2159 0 : TenantState::Attaching if allow_transition_from_attaching => true,
2160 : TenantState::Activating(_) | TenantState::Attaching => {
2161 0 : info!(
2162 0 : "waiting for {} to turn Active|Broken|Stopping",
2163 0 : <&'static str>::from(state)
2164 : );
2165 0 : false
2166 : }
2167 0 : TenantState::Loading => allow_transition_from_loading,
2168 18 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
2169 18 : })
2170 0 : .await
2171 18 : .expect("cannot drop self.state while on a &self method");
2172 18 :
2173 18 : // we now know we're done activating, let's see whether this task is the winner to transition into Stopping
2174 18 : let mut err = None;
2175 18 : let stopping = self.state.send_if_modified(|current_state| match current_state {
2176 : TenantState::Activating(_) => {
2177 0 : unreachable!("1we ensured above that we're done with activation, and, there is no re-activation")
2178 : }
2179 : TenantState::Attaching => {
2180 0 : if !allow_transition_from_attaching {
2181 0 : unreachable!("2we ensured above that we're done with activation, and, there is no re-activation")
2182 0 : };
2183 0 : *current_state = TenantState::Stopping { progress };
2184 0 : true
2185 : }
2186 : TenantState::Loading => {
2187 0 : if !allow_transition_from_loading {
2188 0 : unreachable!("3we ensured above that we're done with activation, and, there is no re-activation")
2189 0 : };
2190 0 : *current_state = TenantState::Stopping { progress };
2191 0 : true
2192 : }
2193 : TenantState::Active => {
2194 : // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
2195 : // are created after the transition to Stopping. That's harmless, as the Timelines
2196 : // won't be accessible to anyone afterwards, because the Tenant is in Stopping state.
2197 18 : *current_state = TenantState::Stopping { progress };
2198 18 : // Continue stopping outside the closure. We need to grab timelines.lock()
2199 18 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
2200 18 : true
2201 : }
2202 0 : TenantState::Broken { reason, .. } => {
2203 0 : info!(
2204 0 : "Cannot set tenant to Stopping state, it is in Broken state due to: {reason}"
2205 : );
2206 0 : err = Some(SetStoppingError::Broken);
2207 0 : false
2208 : }
2209 0 : TenantState::Stopping { progress } => {
2210 0 : info!("Tenant is already in Stopping state");
2211 0 : err = Some(SetStoppingError::AlreadyStopping(progress.clone()));
2212 0 : false
2213 : }
2214 18 : });
2215 18 : match (stopping, err) {
2216 18 : (true, None) => {} // continue
2217 0 : (false, Some(err)) => return Err(err),
2218 0 : (true, Some(_)) => unreachable!(
2219 0 : "send_if_modified closure must error out if not transitioning to Stopping"
2220 0 : ),
2221 0 : (false, None) => unreachable!(
2222 0 : "send_if_modified closure must return true if transitioning to Stopping"
2223 0 : ),
2224 : }
2225 :
2226 18 : let timelines_accessor = self.timelines.lock().unwrap();
2227 18 : let not_broken_timelines = timelines_accessor
2228 18 : .values()
2229 18 : .filter(|timeline| !timeline.is_broken());
2230 36 : for timeline in not_broken_timelines {
2231 18 : timeline.set_state(TimelineState::Stopping);
2232 18 : }
2233 18 : Ok(())
2234 18 : }
2235 :
2236 : /// Method for tenant::mgr to transition us into Broken state in case of a late failure in
2237 : /// `remove_tenant_from_memory`
2238 : ///
2239 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
2240 : ///
2241 : /// In tests, we also use this to set tenants to Broken state on purpose.
2242 0 : pub(crate) async fn set_broken(&self, reason: String) {
2243 0 : let mut rx = self.state.subscribe();
2244 0 :
2245 0 : // The load & attach routines own the tenant state until it has reached `Active`.
2246 0 : // So, wait until it's done.
2247 0 : rx.wait_for(|state| match state {
2248 : TenantState::Activating(_) | TenantState::Loading | TenantState::Attaching => {
2249 0 : info!(
2250 0 : "waiting for {} to turn Active|Broken|Stopping",
2251 0 : <&'static str>::from(state)
2252 : );
2253 0 : false
2254 : }
2255 0 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
2256 0 : })
2257 0 : .await
2258 0 : .expect("cannot drop self.state while on a &self method");
2259 0 :
2260 0 : // we now know we're done activating, let's see whether this task is the winner to transition into Broken
2261 0 : self.set_broken_no_wait(reason)
2262 0 : }
2263 :
2264 0 : pub(crate) fn set_broken_no_wait(&self, reason: impl Display) {
2265 0 : let reason = reason.to_string();
2266 0 : self.state.send_modify(|current_state| {
2267 0 : match *current_state {
2268 : TenantState::Activating(_) | TenantState::Loading | TenantState::Attaching => {
2269 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
2270 : }
2271 : TenantState::Active => {
2272 0 : if cfg!(feature = "testing") {
2273 0 : warn!("Changing Active tenant to Broken state, reason: {}", reason);
2274 0 : *current_state = TenantState::broken_from_reason(reason);
2275 : } else {
2276 0 : unreachable!("not allowed to call set_broken on Active tenants in non-testing builds")
2277 : }
2278 : }
2279 : TenantState::Broken { .. } => {
2280 0 : warn!("Tenant is already in Broken state");
2281 : }
2282 : // This is the only "expected" path, any other path is a bug.
2283 : TenantState::Stopping { .. } => {
2284 0 : warn!(
2285 0 : "Marking Stopping tenant as Broken state, reason: {}",
2286 : reason
2287 : );
2288 0 : *current_state = TenantState::broken_from_reason(reason);
2289 : }
2290 : }
2291 0 : });
2292 0 : }
2293 :
2294 0 : pub fn subscribe_for_state_updates(&self) -> watch::Receiver<TenantState> {
2295 0 : self.state.subscribe()
2296 0 : }
2297 :
2298 : /// The activate_now semaphore is initialized with zero units. As soon as
2299 : /// we add a unit, waiters will be able to acquire a unit and proceed.
2300 0 : pub(crate) fn activate_now(&self) {
2301 0 : self.activate_now_sem.add_permits(1);
2302 0 : }
2303 :
2304 0 : pub(crate) async fn wait_to_become_active(
2305 0 : &self,
2306 0 : timeout: Duration,
2307 0 : ) -> Result<(), GetActiveTenantError> {
2308 0 : let mut receiver = self.state.subscribe();
2309 : loop {
2310 0 : let current_state = receiver.borrow_and_update().clone();
2311 0 : match current_state {
2312 : TenantState::Loading | TenantState::Attaching | TenantState::Activating(_) => {
2313 : // in these states, there's a chance that we can reach ::Active
2314 0 : self.activate_now();
2315 0 : match timeout_cancellable(timeout, &self.cancel, receiver.changed()).await {
2316 0 : Ok(r) => {
2317 0 : r.map_err(
2318 0 : |_e: tokio::sync::watch::error::RecvError|
2319 : // Tenant existed but was dropped: report it as non-existent
2320 0 : GetActiveTenantError::NotFound(GetTenantError::NotFound(self.tenant_shard_id.tenant_id))
2321 0 : )?
2322 : }
2323 : Err(TimeoutCancellableError::Cancelled) => {
2324 0 : return Err(GetActiveTenantError::Cancelled);
2325 : }
2326 : Err(TimeoutCancellableError::Timeout) => {
2327 0 : return Err(GetActiveTenantError::WaitForActiveTimeout {
2328 0 : latest_state: Some(self.current_state()),
2329 0 : wait_time: timeout,
2330 0 : });
2331 : }
2332 : }
2333 : }
2334 : TenantState::Active { .. } => {
2335 0 : return Ok(());
2336 : }
2337 0 : TenantState::Broken { reason, .. } => {
2338 0 : // This is fatal, and reported distinctly from the general case of "will never be active" because
2339 0 : // it's logically a 500 to external API users (broken is always a bug).
2340 0 : return Err(GetActiveTenantError::Broken(reason));
2341 : }
2342 : TenantState::Stopping { .. } => {
2343 : // There's no chance the tenant can transition back into ::Active
2344 0 : return Err(GetActiveTenantError::WillNotBecomeActive(current_state));
2345 : }
2346 : }
2347 : }
2348 0 : }
2349 :
2350 0 : pub(crate) fn get_attach_mode(&self) -> AttachmentMode {
2351 0 : self.tenant_conf.load().location.attach_mode
2352 0 : }
2353 :
2354 : /// For API access: generate a LocationConfig equivalent to the one that would be used to
2355 : /// create a Tenant in the same state. Do not use this in hot paths: it's for relatively
2356 : /// rare external API calls, like a reconciliation at startup.
2357 0 : pub(crate) fn get_location_conf(&self) -> models::LocationConfig {
2358 0 : let conf = self.tenant_conf.load();
2359 :
2360 0 : let location_config_mode = match conf.location.attach_mode {
2361 0 : AttachmentMode::Single => models::LocationConfigMode::AttachedSingle,
2362 0 : AttachmentMode::Multi => models::LocationConfigMode::AttachedMulti,
2363 0 : AttachmentMode::Stale => models::LocationConfigMode::AttachedStale,
2364 : };
2365 :
2366 : // We have a pageserver TenantConf, we need the API-facing TenantConfig.
2367 0 : let tenant_config: models::TenantConfig = conf.tenant_conf.clone().into();
2368 0 :
2369 0 : models::LocationConfig {
2370 0 : mode: location_config_mode,
2371 0 : generation: self.generation.into(),
2372 0 : secondary_conf: None,
2373 0 : shard_number: self.shard_identity.number.0,
2374 0 : shard_count: self.shard_identity.count.literal(),
2375 0 : shard_stripe_size: self.shard_identity.stripe_size.0,
2376 0 : tenant_conf: tenant_config,
2377 0 : }
2378 0 : }
2379 :
2380 0 : pub(crate) fn get_tenant_shard_id(&self) -> &TenantShardId {
2381 0 : &self.tenant_shard_id
2382 0 : }
2383 :
2384 0 : pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
2385 0 : self.shard_identity.stripe_size
2386 0 : }
2387 :
2388 0 : pub(crate) fn get_generation(&self) -> Generation {
2389 0 : self.generation
2390 0 : }
2391 :
2392 : /// This function partially shuts down the tenant (it shuts down the Timelines) and is fallible,
2393 : /// and can leave the tenant in a bad state if it fails. The caller is responsible for
2394 : /// resetting this tenant to a valid state if we fail.
2395 0 : pub(crate) async fn split_prepare(
2396 0 : &self,
2397 0 : child_shards: &Vec<TenantShardId>,
2398 0 : ) -> anyhow::Result<()> {
2399 0 : let timelines = self.timelines.lock().unwrap().clone();
2400 0 : for timeline in timelines.values() {
2401 : // We do not block timeline creation/deletion during splits inside the pageserver: it is up to higher levels
2402 : // to ensure that they do not start a split if currently in the process of doing these.
2403 :
2404 : // Upload an index from the parent: this is partly to provide freshness for the
2405 : // child tenants that will copy it, and partly for general ease-of-debugging: there will
2406 : // always be a parent shard index in the same generation as we wrote the child shard index.
2407 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Uploading index");
2408 0 : timeline
2409 0 : .remote_client
2410 0 : .schedule_index_upload_for_file_changes()?;
2411 0 : timeline.remote_client.wait_completion().await?;
2412 :
2413 : // Shut down the timeline's remote client: this means that the indices we write
2414 : // for child shards will not be invalidated by the parent shard deleting layers.
2415 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Shutting down remote storage client");
2416 0 : timeline.remote_client.shutdown().await;
2417 :
2418 : // Download methods can still be used after shutdown, as they don't flow through the remote client's
2419 : // queue. In principal the RemoteTimelineClient could provide this without downloading it, but this
2420 : // operation is rare, so it's simpler to just download it (and robustly guarantees that the index
2421 : // we use here really is the remotely persistent one).
2422 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Downloading index_part from parent");
2423 0 : let result = timeline.remote_client
2424 0 : .download_index_file(&self.cancel)
2425 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))
2426 0 : .await?;
2427 0 : let index_part = match result {
2428 : MaybeDeletedIndexPart::Deleted(_) => {
2429 0 : anyhow::bail!("Timeline deletion happened concurrently with split")
2430 : }
2431 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
2432 : };
2433 :
2434 0 : for child_shard in child_shards {
2435 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Uploading index_part for child {}", child_shard.to_index());
2436 0 : upload_index_part(
2437 0 : &self.remote_storage,
2438 0 : child_shard,
2439 0 : &timeline.timeline_id,
2440 0 : self.generation,
2441 0 : &index_part,
2442 0 : &self.cancel,
2443 0 : )
2444 0 : .await?;
2445 : }
2446 : }
2447 :
2448 0 : Ok(())
2449 0 : }
2450 :
2451 0 : pub(crate) fn get_sizes(&self) -> TopTenantShardItem {
2452 0 : let mut result = TopTenantShardItem {
2453 0 : id: self.tenant_shard_id,
2454 0 : resident_size: 0,
2455 0 : physical_size: 0,
2456 0 : max_logical_size: 0,
2457 0 : };
2458 :
2459 0 : for timeline in self.timelines.lock().unwrap().values() {
2460 0 : result.resident_size += timeline.metrics.resident_physical_size_gauge.get();
2461 0 :
2462 0 : result.physical_size += timeline
2463 0 : .remote_client
2464 0 : .metrics
2465 0 : .remote_physical_size_gauge
2466 0 : .get();
2467 0 : result.max_logical_size = std::cmp::max(
2468 0 : result.max_logical_size,
2469 0 : timeline.metrics.current_logical_size_gauge.get(),
2470 0 : );
2471 0 : }
2472 :
2473 0 : result
2474 0 : }
2475 : }
2476 :
2477 : /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
2478 : /// perform a topological sort, so that the parent of each timeline comes
2479 : /// before the children.
2480 : /// E extracts the ancestor from T
2481 : /// This allows for T to be different. It can be TimelineMetadata, can be Timeline itself, etc.
2482 570 : fn tree_sort_timelines<T, E>(
2483 570 : timelines: HashMap<TimelineId, T>,
2484 570 : extractor: E,
2485 570 : ) -> anyhow::Result<Vec<(TimelineId, T)>>
2486 570 : where
2487 570 : E: Fn(&T) -> Option<TimelineId>,
2488 570 : {
2489 570 : let mut result = Vec::with_capacity(timelines.len());
2490 570 :
2491 570 : let mut now = Vec::with_capacity(timelines.len());
2492 570 : // (ancestor, children)
2493 570 : let mut later: HashMap<TimelineId, Vec<(TimelineId, T)>> =
2494 570 : HashMap::with_capacity(timelines.len());
2495 :
2496 588 : for (timeline_id, value) in timelines {
2497 18 : if let Some(ancestor_id) = extractor(&value) {
2498 6 : let children = later.entry(ancestor_id).or_default();
2499 6 : children.push((timeline_id, value));
2500 12 : } else {
2501 12 : now.push((timeline_id, value));
2502 12 : }
2503 : }
2504 :
2505 588 : while let Some((timeline_id, metadata)) = now.pop() {
2506 18 : result.push((timeline_id, metadata));
2507 : // All children of this can be loaded now
2508 18 : if let Some(mut children) = later.remove(&timeline_id) {
2509 6 : now.append(&mut children);
2510 12 : }
2511 : }
2512 :
2513 : // All timelines should be visited now. Unless there were timelines with missing ancestors.
2514 570 : if !later.is_empty() {
2515 0 : for (missing_id, orphan_ids) in later {
2516 0 : for (orphan_id, _) in orphan_ids {
2517 0 : error!("could not load timeline {orphan_id} because its ancestor timeline {missing_id} could not be loaded");
2518 : }
2519 : }
2520 0 : bail!("could not load tenant because some timelines are missing ancestors");
2521 570 : }
2522 570 :
2523 570 : Ok(result)
2524 570 : }
2525 :
2526 : impl Tenant {
2527 0 : pub fn tenant_specific_overrides(&self) -> TenantConfOpt {
2528 0 : self.tenant_conf.load().tenant_conf.clone()
2529 0 : }
2530 :
2531 0 : pub fn effective_config(&self) -> TenantConf {
2532 0 : self.tenant_specific_overrides()
2533 0 : .merge(self.conf.default_tenant_conf.clone())
2534 0 : }
2535 :
2536 0 : pub fn get_checkpoint_distance(&self) -> u64 {
2537 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2538 0 : tenant_conf
2539 0 : .checkpoint_distance
2540 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
2541 0 : }
2542 :
2543 0 : pub fn get_checkpoint_timeout(&self) -> Duration {
2544 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2545 0 : tenant_conf
2546 0 : .checkpoint_timeout
2547 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
2548 0 : }
2549 :
2550 0 : pub fn get_compaction_target_size(&self) -> u64 {
2551 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2552 0 : tenant_conf
2553 0 : .compaction_target_size
2554 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
2555 0 : }
2556 :
2557 0 : pub fn get_compaction_period(&self) -> Duration {
2558 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2559 0 : tenant_conf
2560 0 : .compaction_period
2561 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
2562 0 : }
2563 :
2564 0 : pub fn get_compaction_threshold(&self) -> usize {
2565 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2566 0 : tenant_conf
2567 0 : .compaction_threshold
2568 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
2569 0 : }
2570 :
2571 0 : pub fn get_gc_horizon(&self) -> u64 {
2572 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2573 0 : tenant_conf
2574 0 : .gc_horizon
2575 0 : .unwrap_or(self.conf.default_tenant_conf.gc_horizon)
2576 0 : }
2577 :
2578 0 : pub fn get_gc_period(&self) -> Duration {
2579 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2580 0 : tenant_conf
2581 0 : .gc_period
2582 0 : .unwrap_or(self.conf.default_tenant_conf.gc_period)
2583 0 : }
2584 :
2585 0 : pub fn get_image_creation_threshold(&self) -> usize {
2586 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2587 0 : tenant_conf
2588 0 : .image_creation_threshold
2589 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
2590 0 : }
2591 :
2592 0 : pub fn get_pitr_interval(&self) -> Duration {
2593 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2594 0 : tenant_conf
2595 0 : .pitr_interval
2596 0 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
2597 0 : }
2598 :
2599 0 : pub fn get_min_resident_size_override(&self) -> Option<u64> {
2600 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2601 0 : tenant_conf
2602 0 : .min_resident_size_override
2603 0 : .or(self.conf.default_tenant_conf.min_resident_size_override)
2604 0 : }
2605 :
2606 0 : pub fn get_heatmap_period(&self) -> Option<Duration> {
2607 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2608 0 : let heatmap_period = tenant_conf
2609 0 : .heatmap_period
2610 0 : .unwrap_or(self.conf.default_tenant_conf.heatmap_period);
2611 0 : if heatmap_period.is_zero() {
2612 0 : None
2613 : } else {
2614 0 : Some(heatmap_period)
2615 : }
2616 0 : }
2617 :
2618 0 : pub fn get_lsn_lease_length(&self) -> Duration {
2619 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2620 0 : tenant_conf
2621 0 : .lsn_lease_length
2622 0 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
2623 0 : }
2624 :
2625 0 : pub fn set_new_tenant_config(&self, new_tenant_conf: TenantConfOpt) {
2626 0 : // Use read-copy-update in order to avoid overwriting the location config
2627 0 : // state if this races with [`Tenant::set_new_location_config`]. Note that
2628 0 : // this race is not possible if both request types come from the storage
2629 0 : // controller (as they should!) because an exclusive op lock is required
2630 0 : // on the storage controller side.
2631 0 : self.tenant_conf.rcu(|inner| {
2632 0 : Arc::new(AttachedTenantConf {
2633 0 : tenant_conf: new_tenant_conf.clone(),
2634 0 : location: inner.location,
2635 0 : })
2636 0 : });
2637 0 :
2638 0 : self.tenant_conf_updated(&new_tenant_conf);
2639 0 : // Don't hold self.timelines.lock() during the notifies.
2640 0 : // There's no risk of deadlock right now, but there could be if we consolidate
2641 0 : // mutexes in struct Timeline in the future.
2642 0 : let timelines = self.list_timelines();
2643 0 : for timeline in timelines {
2644 0 : timeline.tenant_conf_updated(&new_tenant_conf);
2645 0 : }
2646 0 : }
2647 :
2648 24 : pub(crate) fn set_new_location_config(&self, new_conf: AttachedTenantConf) {
2649 24 : let new_tenant_conf = new_conf.tenant_conf.clone();
2650 24 :
2651 24 : self.tenant_conf.store(Arc::new(new_conf));
2652 24 :
2653 24 : self.tenant_conf_updated(&new_tenant_conf);
2654 24 : // Don't hold self.timelines.lock() during the notifies.
2655 24 : // There's no risk of deadlock right now, but there could be if we consolidate
2656 24 : // mutexes in struct Timeline in the future.
2657 24 : let timelines = self.list_timelines();
2658 48 : for timeline in timelines {
2659 24 : timeline.tenant_conf_updated(&new_tenant_conf);
2660 24 : }
2661 24 : }
2662 :
2663 594 : fn get_timeline_get_throttle_config(
2664 594 : psconf: &'static PageServerConf,
2665 594 : overrides: &TenantConfOpt,
2666 594 : ) -> throttle::Config {
2667 594 : overrides
2668 594 : .timeline_get_throttle
2669 594 : .clone()
2670 594 : .unwrap_or(psconf.default_tenant_conf.timeline_get_throttle.clone())
2671 594 : }
2672 :
2673 24 : pub(crate) fn tenant_conf_updated(&self, new_conf: &TenantConfOpt) {
2674 24 : let conf = Self::get_timeline_get_throttle_config(self.conf, new_conf);
2675 24 : self.timeline_get_throttle.reconfigure(conf)
2676 24 : }
2677 :
2678 : /// Helper function to create a new Timeline struct.
2679 : ///
2680 : /// The returned Timeline is in Loading state. The caller is responsible for
2681 : /// initializing any on-disk state, and for inserting the Timeline to the 'timelines'
2682 : /// map.
2683 : ///
2684 : /// `validate_ancestor == false` is used when a timeline is created for deletion
2685 : /// and we might not have the ancestor present anymore which is fine for to be
2686 : /// deleted timelines.
2687 1242 : fn create_timeline_struct(
2688 1242 : &self,
2689 1242 : new_timeline_id: TimelineId,
2690 1242 : new_metadata: &TimelineMetadata,
2691 1242 : ancestor: Option<Arc<Timeline>>,
2692 1242 : resources: TimelineResources,
2693 1242 : cause: CreateTimelineCause,
2694 1242 : last_aux_file_policy: Option<AuxFilePolicy>,
2695 1242 : ) -> anyhow::Result<Arc<Timeline>> {
2696 1242 : let state = match cause {
2697 : CreateTimelineCause::Load => {
2698 1242 : let ancestor_id = new_metadata.ancestor_timeline();
2699 1242 : anyhow::ensure!(
2700 1242 : ancestor_id == ancestor.as_ref().map(|t| t.timeline_id),
2701 0 : "Timeline's {new_timeline_id} ancestor {ancestor_id:?} was not found"
2702 : );
2703 1242 : TimelineState::Loading
2704 : }
2705 0 : CreateTimelineCause::Delete => TimelineState::Stopping,
2706 : };
2707 :
2708 1242 : let pg_version = new_metadata.pg_version();
2709 1242 :
2710 1242 : let timeline = Timeline::new(
2711 1242 : self.conf,
2712 1242 : Arc::clone(&self.tenant_conf),
2713 1242 : new_metadata,
2714 1242 : ancestor,
2715 1242 : new_timeline_id,
2716 1242 : self.tenant_shard_id,
2717 1242 : self.generation,
2718 1242 : self.shard_identity,
2719 1242 : self.walredo_mgr.clone(),
2720 1242 : resources,
2721 1242 : pg_version,
2722 1242 : state,
2723 1242 : last_aux_file_policy,
2724 1242 : self.cancel.child_token(),
2725 1242 : );
2726 1242 :
2727 1242 : Ok(timeline)
2728 1242 : }
2729 :
2730 : // Allow too_many_arguments because a constructor's argument list naturally grows with the
2731 : // number of attributes in the struct: breaking these out into a builder wouldn't be helpful.
2732 : #[allow(clippy::too_many_arguments)]
2733 570 : fn new(
2734 570 : state: TenantState,
2735 570 : conf: &'static PageServerConf,
2736 570 : attached_conf: AttachedTenantConf,
2737 570 : shard_identity: ShardIdentity,
2738 570 : walredo_mgr: Option<Arc<WalRedoManager>>,
2739 570 : tenant_shard_id: TenantShardId,
2740 570 : remote_storage: GenericRemoteStorage,
2741 570 : deletion_queue_client: DeletionQueueClient,
2742 570 : l0_flush_global_state: L0FlushGlobalState,
2743 570 : ) -> Tenant {
2744 570 : debug_assert!(
2745 570 : !attached_conf.location.generation.is_none() || conf.control_plane_api.is_none()
2746 : );
2747 :
2748 570 : let (state, mut rx) = watch::channel(state);
2749 570 :
2750 570 : tokio::spawn(async move {
2751 570 : // reflect tenant state in metrics:
2752 570 : // - global per tenant state: TENANT_STATE_METRIC
2753 570 : // - "set" of broken tenants: BROKEN_TENANTS_SET
2754 570 : //
2755 570 : // set of broken tenants should not have zero counts so that it remains accessible for
2756 570 : // alerting.
2757 570 :
2758 570 : let tid = tenant_shard_id.to_string();
2759 570 : let shard_id = tenant_shard_id.shard_slug().to_string();
2760 570 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
2761 :
2762 1134 : fn inspect_state(state: &TenantState) -> ([&'static str; 1], bool) {
2763 1134 : ([state.into()], matches!(state, TenantState::Broken { .. }))
2764 1134 : }
2765 :
2766 570 : let mut tuple = inspect_state(&rx.borrow_and_update());
2767 570 :
2768 570 : let is_broken = tuple.1;
2769 570 : let mut counted_broken = if is_broken {
2770 : // add the id to the set right away, there should not be any updates on the channel
2771 : // after before tenant is removed, if ever
2772 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
2773 0 : true
2774 : } else {
2775 570 : false
2776 : };
2777 :
2778 : loop {
2779 1134 : let labels = &tuple.0;
2780 1134 : let current = TENANT_STATE_METRIC.with_label_values(labels);
2781 1134 : current.inc();
2782 1134 :
2783 1134 : if rx.changed().await.is_err() {
2784 : // tenant has been dropped
2785 48 : current.dec();
2786 48 : drop(BROKEN_TENANTS_SET.remove_label_values(set_key));
2787 48 : break;
2788 564 : }
2789 564 :
2790 564 : current.dec();
2791 564 : tuple = inspect_state(&rx.borrow_and_update());
2792 564 :
2793 564 : let is_broken = tuple.1;
2794 564 : if is_broken && !counted_broken {
2795 0 : counted_broken = true;
2796 0 : // insert the tenant_id (back) into the set while avoiding needless counter
2797 0 : // access
2798 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
2799 564 : }
2800 : }
2801 570 : });
2802 570 :
2803 570 : Tenant {
2804 570 : tenant_shard_id,
2805 570 : shard_identity,
2806 570 : generation: attached_conf.location.generation,
2807 570 : conf,
2808 570 : // using now here is good enough approximation to catch tenants with really long
2809 570 : // activation times.
2810 570 : constructed_at: Instant::now(),
2811 570 : timelines: Mutex::new(HashMap::new()),
2812 570 : timelines_creating: Mutex::new(HashSet::new()),
2813 570 : gc_cs: tokio::sync::Mutex::new(()),
2814 570 : walredo_mgr,
2815 570 : remote_storage,
2816 570 : deletion_queue_client,
2817 570 : state,
2818 570 : cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
2819 570 : cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
2820 570 : eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
2821 570 : compaction_circuit_breaker: std::sync::Mutex::new(CircuitBreaker::new(
2822 570 : format!("compaction-{tenant_shard_id}"),
2823 570 : 5,
2824 570 : // Compaction can be a very expensive operation, and might leak disk space. It also ought
2825 570 : // to be infallible, as long as remote storage is available. So if it repeatedly fails,
2826 570 : // use an extremely long backoff.
2827 570 : Some(Duration::from_secs(3600 * 24)),
2828 570 : )),
2829 570 : activate_now_sem: tokio::sync::Semaphore::new(0),
2830 570 : cancel: CancellationToken::default(),
2831 570 : gate: Gate::default(),
2832 570 : timeline_get_throttle: Arc::new(throttle::Throttle::new(
2833 570 : Tenant::get_timeline_get_throttle_config(conf, &attached_conf.tenant_conf),
2834 570 : &crate::metrics::tenant_throttling::TIMELINE_GET,
2835 570 : )),
2836 570 : tenant_conf: Arc::new(ArcSwap::from_pointee(attached_conf)),
2837 570 : ongoing_timeline_detach: std::sync::Mutex::default(),
2838 570 : gc_block: Default::default(),
2839 570 : l0_flush_global_state,
2840 570 : }
2841 570 : }
2842 :
2843 : /// Locate and load config
2844 0 : pub(super) fn load_tenant_config(
2845 0 : conf: &'static PageServerConf,
2846 0 : tenant_shard_id: &TenantShardId,
2847 0 : ) -> Result<LocationConf, LoadConfigError> {
2848 0 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
2849 0 :
2850 0 : info!("loading tenant configuration from {config_path}");
2851 :
2852 : // load and parse file
2853 0 : let config = fs::read_to_string(&config_path).map_err(|e| {
2854 0 : match e.kind() {
2855 : std::io::ErrorKind::NotFound => {
2856 : // The config should almost always exist for a tenant directory:
2857 : // - When attaching a tenant, the config is the first thing we write
2858 : // - When detaching a tenant, we atomically move the directory to a tmp location
2859 : // before deleting contents.
2860 : //
2861 : // The very rare edge case that can result in a missing config is if we crash during attach
2862 : // between creating directory and writing config. Callers should handle that as if the
2863 : // directory didn't exist.
2864 :
2865 0 : LoadConfigError::NotFound(config_path)
2866 : }
2867 : _ => {
2868 : // No IO errors except NotFound are acceptable here: other kinds of error indicate local storage or permissions issues
2869 : // that we cannot cleanly recover
2870 0 : crate::virtual_file::on_fatal_io_error(&e, "Reading tenant config file")
2871 : }
2872 : }
2873 0 : })?;
2874 :
2875 0 : Ok(toml_edit::de::from_str::<LocationConf>(&config)?)
2876 0 : }
2877 :
2878 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
2879 : pub(super) async fn persist_tenant_config(
2880 : conf: &'static PageServerConf,
2881 : tenant_shard_id: &TenantShardId,
2882 : location_conf: &LocationConf,
2883 : ) -> std::io::Result<()> {
2884 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
2885 :
2886 : Self::persist_tenant_config_at(tenant_shard_id, &config_path, location_conf).await
2887 : }
2888 :
2889 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
2890 : pub(super) async fn persist_tenant_config_at(
2891 : tenant_shard_id: &TenantShardId,
2892 : config_path: &Utf8Path,
2893 : location_conf: &LocationConf,
2894 : ) -> std::io::Result<()> {
2895 : debug!("persisting tenantconf to {config_path}");
2896 :
2897 : let mut conf_content = r#"# This file contains a specific per-tenant's config.
2898 : # It is read in case of pageserver restart.
2899 : "#
2900 : .to_string();
2901 :
2902 0 : fail::fail_point!("tenant-config-before-write", |_| {
2903 0 : Err(std::io::Error::new(
2904 0 : std::io::ErrorKind::Other,
2905 0 : "tenant-config-before-write",
2906 0 : ))
2907 0 : });
2908 :
2909 : // Convert the config to a toml file.
2910 : conf_content +=
2911 : &toml_edit::ser::to_string_pretty(&location_conf).expect("Config serialization failed");
2912 :
2913 : let temp_path = path_with_suffix_extension(config_path, TEMP_FILE_SUFFIX);
2914 :
2915 : let conf_content = conf_content.into_bytes();
2916 : VirtualFile::crashsafe_overwrite(config_path.to_owned(), temp_path, conf_content).await
2917 : }
2918 :
2919 : //
2920 : // How garbage collection works:
2921 : //
2922 : // +--bar------------->
2923 : // /
2924 : // +----+-----foo---------------->
2925 : // /
2926 : // ----main--+-------------------------->
2927 : // \
2928 : // +-----baz-------->
2929 : //
2930 : //
2931 : // 1. Grab 'gc_cs' mutex to prevent new timelines from being created while Timeline's
2932 : // `gc_infos` are being refreshed
2933 : // 2. Scan collected timelines, and on each timeline, make note of the
2934 : // all the points where other timelines have been branched off.
2935 : // We will refrain from removing page versions at those LSNs.
2936 : // 3. For each timeline, scan all layer files on the timeline.
2937 : // Remove all files for which a newer file exists and which
2938 : // don't cover any branch point LSNs.
2939 : //
2940 : // TODO:
2941 : // - if a relation has a non-incremental persistent layer on a child branch, then we
2942 : // don't need to keep that in the parent anymore. But currently
2943 : // we do.
2944 2262 : async fn gc_iteration_internal(
2945 2262 : &self,
2946 2262 : target_timeline_id: Option<TimelineId>,
2947 2262 : horizon: u64,
2948 2262 : pitr: Duration,
2949 2262 : cancel: &CancellationToken,
2950 2262 : ctx: &RequestContext,
2951 2262 : ) -> Result<GcResult, GcError> {
2952 2262 : let mut totals: GcResult = Default::default();
2953 2262 : let now = Instant::now();
2954 :
2955 2262 : let gc_timelines = self
2956 2262 : .refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
2957 2254 : .await?;
2958 :
2959 2262 : failpoint_support::sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
2960 :
2961 : // If there is nothing to GC, we don't want any messages in the INFO log.
2962 2262 : if !gc_timelines.is_empty() {
2963 2262 : info!("{} timelines need GC", gc_timelines.len());
2964 : } else {
2965 0 : debug!("{} timelines need GC", gc_timelines.len());
2966 : }
2967 :
2968 : // Perform GC for each timeline.
2969 : //
2970 : // Note that we don't hold the `Tenant::gc_cs` lock here because we don't want to delay the
2971 : // branch creation task, which requires the GC lock. A GC iteration can run concurrently
2972 : // with branch creation.
2973 : //
2974 : // See comments in [`Tenant::branch_timeline`] for more information about why branch
2975 : // creation task can run concurrently with timeline's GC iteration.
2976 4524 : for timeline in gc_timelines {
2977 2262 : if cancel.is_cancelled() {
2978 : // We were requested to shut down. Stop and return with the progress we
2979 : // made.
2980 0 : break;
2981 2262 : }
2982 2262 : let result = match timeline.gc().await {
2983 : Err(GcError::TimelineCancelled) => {
2984 0 : if target_timeline_id.is_some() {
2985 : // If we were targetting this specific timeline, surface cancellation to caller
2986 0 : return Err(GcError::TimelineCancelled);
2987 : } else {
2988 : // A timeline may be shutting down independently of the tenant's lifecycle: we should
2989 : // skip past this and proceed to try GC on other timelines.
2990 0 : continue;
2991 : }
2992 : }
2993 2262 : r => r?,
2994 : };
2995 2262 : totals += result;
2996 : }
2997 :
2998 2262 : totals.elapsed = now.elapsed();
2999 2262 : Ok(totals)
3000 2262 : }
3001 :
3002 : /// Refreshes the Timeline::gc_info for all timelines, returning the
3003 : /// vector of timelines which have [`Timeline::get_last_record_lsn`] past
3004 : /// [`Tenant::get_gc_horizon`].
3005 : ///
3006 : /// This is usually executed as part of periodic gc, but can now be triggered more often.
3007 0 : pub(crate) async fn refresh_gc_info(
3008 0 : &self,
3009 0 : cancel: &CancellationToken,
3010 0 : ctx: &RequestContext,
3011 0 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
3012 0 : // since this method can now be called at different rates than the configured gc loop, it
3013 0 : // might be that these configuration values get applied faster than what it was previously,
3014 0 : // since these were only read from the gc task.
3015 0 : let horizon = self.get_gc_horizon();
3016 0 : let pitr = self.get_pitr_interval();
3017 0 :
3018 0 : // refresh all timelines
3019 0 : let target_timeline_id = None;
3020 0 :
3021 0 : self.refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
3022 0 : .await
3023 0 : }
3024 :
3025 : /// Populate all Timelines' `GcInfo` with information about their children. We do not set the
3026 : /// PITR cutoffs here, because that requires I/O: this is done later, before GC, by [`Self::refresh_gc_info_internal`]
3027 : ///
3028 : /// Subsequently, parent-child relationships are updated incrementally during timeline creation/deletion.
3029 0 : fn initialize_gc_info(
3030 0 : &self,
3031 0 : timelines: &std::sync::MutexGuard<HashMap<TimelineId, Arc<Timeline>>>,
3032 0 : ) {
3033 0 : // This function must be called before activation: after activation timeline create/delete operations
3034 0 : // might happen, and this function is not safe to run concurrently with those.
3035 0 : assert!(!self.is_active());
3036 :
3037 : // Scan all timelines. For each timeline, remember the timeline ID and
3038 : // the branch point where it was created.
3039 0 : let mut all_branchpoints: BTreeMap<TimelineId, Vec<(Lsn, TimelineId)>> = BTreeMap::new();
3040 0 : timelines.iter().for_each(|(timeline_id, timeline_entry)| {
3041 0 : if let Some(ancestor_timeline_id) = &timeline_entry.get_ancestor_timeline_id() {
3042 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
3043 0 : ancestor_children.push((timeline_entry.get_ancestor_lsn(), *timeline_id));
3044 0 : }
3045 0 : });
3046 0 :
3047 0 : // The number of bytes we always keep, irrespective of PITR: this is a constant across timelines
3048 0 : let horizon = self.get_gc_horizon();
3049 :
3050 : // Populate each timeline's GcInfo with information about its child branches
3051 0 : for timeline in timelines.values() {
3052 0 : let mut branchpoints: Vec<(Lsn, TimelineId)> = all_branchpoints
3053 0 : .remove(&timeline.timeline_id)
3054 0 : .unwrap_or_default();
3055 0 :
3056 0 : branchpoints.sort_by_key(|b| b.0);
3057 0 :
3058 0 : let mut target = timeline.gc_info.write().unwrap();
3059 0 :
3060 0 : target.retain_lsns = branchpoints;
3061 0 :
3062 0 : let space_cutoff = timeline
3063 0 : .get_last_record_lsn()
3064 0 : .checked_sub(horizon)
3065 0 : .unwrap_or(Lsn(0));
3066 0 :
3067 0 : target.cutoffs = GcCutoffs {
3068 0 : space: space_cutoff,
3069 0 : time: Lsn::INVALID,
3070 0 : };
3071 0 : }
3072 0 : }
3073 :
3074 2262 : async fn refresh_gc_info_internal(
3075 2262 : &self,
3076 2262 : target_timeline_id: Option<TimelineId>,
3077 2262 : horizon: u64,
3078 2262 : pitr: Duration,
3079 2262 : cancel: &CancellationToken,
3080 2262 : ctx: &RequestContext,
3081 2262 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
3082 2262 : // before taking the gc_cs lock, do the heavier weight finding of gc_cutoff points for
3083 2262 : // currently visible timelines.
3084 2262 : let timelines = self
3085 2262 : .timelines
3086 2262 : .lock()
3087 2262 : .unwrap()
3088 2262 : .values()
3089 9930 : .filter(|tl| match target_timeline_id.as_ref() {
3090 9930 : Some(target) => &tl.timeline_id == target,
3091 0 : None => true,
3092 9930 : })
3093 2262 : .cloned()
3094 2262 : .collect::<Vec<_>>();
3095 2262 :
3096 2262 : if target_timeline_id.is_some() && timelines.is_empty() {
3097 : // We were to act on a particular timeline and it wasn't found
3098 0 : return Err(GcError::TimelineNotFound);
3099 2262 : }
3100 2262 :
3101 2262 : let mut gc_cutoffs: HashMap<TimelineId, GcCutoffs> =
3102 2262 : HashMap::with_capacity(timelines.len());
3103 :
3104 2262 : for timeline in timelines.iter() {
3105 2262 : let cutoff = timeline
3106 2262 : .get_last_record_lsn()
3107 2262 : .checked_sub(horizon)
3108 2262 : .unwrap_or(Lsn(0));
3109 :
3110 2262 : let cutoffs = timeline.find_gc_cutoffs(cutoff, pitr, cancel, ctx).await?;
3111 2262 : let old = gc_cutoffs.insert(timeline.timeline_id, cutoffs);
3112 2262 : assert!(old.is_none());
3113 : }
3114 :
3115 2262 : if !self.is_active() || self.cancel.is_cancelled() {
3116 0 : return Err(GcError::TenantCancelled);
3117 2262 : }
3118 :
3119 : // grab mutex to prevent new timelines from being created here; avoid doing long operations
3120 : // because that will stall branch creation.
3121 2262 : let gc_cs = self.gc_cs.lock().await;
3122 :
3123 : // Ok, we now know all the branch points.
3124 : // Update the GC information for each timeline.
3125 2262 : let mut gc_timelines = Vec::with_capacity(timelines.len());
3126 4524 : for timeline in timelines {
3127 : // We filtered the timeline list above
3128 2262 : if let Some(target_timeline_id) = target_timeline_id {
3129 2262 : assert_eq!(target_timeline_id, timeline.timeline_id);
3130 0 : }
3131 :
3132 : {
3133 2262 : let mut target = timeline.gc_info.write().unwrap();
3134 2262 :
3135 2262 : // Cull any expired leases
3136 2262 : let now = SystemTime::now();
3137 2262 : target.leases.retain(|_, lease| !lease.is_expired(&now));
3138 2262 :
3139 2262 : timeline
3140 2262 : .metrics
3141 2262 : .valid_lsn_lease_count_gauge
3142 2262 : .set(target.leases.len() as u64);
3143 :
3144 : // Look up parent's PITR cutoff to update the child's knowledge of whether it is within parent's PITR
3145 2262 : if let Some(ancestor_id) = timeline.get_ancestor_timeline_id() {
3146 300 : if let Some(ancestor_gc_cutoffs) = gc_cutoffs.get(&ancestor_id) {
3147 0 : target.within_ancestor_pitr =
3148 0 : timeline.get_ancestor_lsn() >= ancestor_gc_cutoffs.time;
3149 300 : }
3150 1962 : }
3151 :
3152 : // Update metrics that depend on GC state
3153 2262 : timeline
3154 2262 : .metrics
3155 2262 : .archival_size
3156 2262 : .set(if target.within_ancestor_pitr {
3157 0 : timeline.metrics.current_logical_size_gauge.get()
3158 : } else {
3159 2262 : 0
3160 : });
3161 2262 : timeline.metrics.pitr_history_size.set(
3162 2262 : timeline
3163 2262 : .get_last_record_lsn()
3164 2262 : .checked_sub(target.cutoffs.time)
3165 2262 : .unwrap_or(Lsn(0))
3166 2262 : .0,
3167 2262 : );
3168 :
3169 : // Apply the cutoffs we found to the Timeline's GcInfo. Why might we _not_ have cutoffs for a timeline?
3170 : // - this timeline was created while we were finding cutoffs
3171 : // - lsn for timestamp search fails for this timeline repeatedly
3172 2262 : if let Some(cutoffs) = gc_cutoffs.get(&timeline.timeline_id) {
3173 2262 : target.cutoffs = cutoffs.clone();
3174 2262 : }
3175 : }
3176 :
3177 2262 : gc_timelines.push(timeline);
3178 : }
3179 2262 : drop(gc_cs);
3180 2262 : Ok(gc_timelines)
3181 2262 : }
3182 :
3183 : /// A substitute for `branch_timeline` for use in unit tests.
3184 : /// The returned timeline will have state value `Active` to make various `anyhow::ensure!()`
3185 : /// calls pass, but, we do not actually call `.activate()` under the hood. So, none of the
3186 : /// timeline background tasks are launched, except the flush loop.
3187 : #[cfg(test)]
3188 690 : async fn branch_timeline_test(
3189 690 : &self,
3190 690 : src_timeline: &Arc<Timeline>,
3191 690 : dst_id: TimelineId,
3192 690 : ancestor_lsn: Option<Lsn>,
3193 690 : ctx: &RequestContext,
3194 690 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
3195 690 : let create_guard = self.create_timeline_create_guard(dst_id).unwrap();
3196 690 : let tl = self
3197 690 : .branch_timeline_impl(src_timeline, dst_id, ancestor_lsn, create_guard, ctx)
3198 12 : .await?;
3199 678 : tl.set_state(TimelineState::Active);
3200 678 : Ok(tl)
3201 690 : }
3202 :
3203 : /// Helper for unit tests to branch a timeline with some pre-loaded states.
3204 : #[cfg(test)]
3205 : #[allow(clippy::too_many_arguments)]
3206 18 : pub async fn branch_timeline_test_with_layers(
3207 18 : &self,
3208 18 : src_timeline: &Arc<Timeline>,
3209 18 : dst_id: TimelineId,
3210 18 : ancestor_lsn: Option<Lsn>,
3211 18 : ctx: &RequestContext,
3212 18 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
3213 18 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
3214 18 : end_lsn: Lsn,
3215 18 : ) -> anyhow::Result<Arc<Timeline>> {
3216 : use checks::check_valid_layermap;
3217 : use itertools::Itertools;
3218 :
3219 18 : let tline = self
3220 18 : .branch_timeline_test(src_timeline, dst_id, ancestor_lsn, ctx)
3221 0 : .await?;
3222 18 : let ancestor_lsn = if let Some(ancestor_lsn) = ancestor_lsn {
3223 18 : ancestor_lsn
3224 : } else {
3225 0 : tline.get_last_record_lsn()
3226 : };
3227 18 : assert!(end_lsn >= ancestor_lsn);
3228 18 : tline.force_advance_lsn(end_lsn);
3229 36 : for deltas in delta_layer_desc {
3230 18 : tline
3231 18 : .force_create_delta_layer(deltas, Some(ancestor_lsn), ctx)
3232 54 : .await?;
3233 : }
3234 30 : for (lsn, images) in image_layer_desc {
3235 12 : tline
3236 12 : .force_create_image_layer(lsn, images, Some(ancestor_lsn), ctx)
3237 42 : .await?;
3238 : }
3239 18 : let layer_names = tline
3240 18 : .layers
3241 18 : .read()
3242 0 : .await
3243 18 : .layer_map()
3244 18 : .unwrap()
3245 18 : .iter_historic_layers()
3246 30 : .map(|layer| layer.layer_name())
3247 18 : .collect_vec();
3248 18 : if let Some(err) = check_valid_layermap(&layer_names) {
3249 0 : bail!("invalid layermap: {err}");
3250 18 : }
3251 18 : Ok(tline)
3252 18 : }
3253 :
3254 : /// Branch an existing timeline.
3255 : ///
3256 : /// The caller is responsible for activating the returned timeline.
3257 0 : async fn branch_timeline(
3258 0 : &self,
3259 0 : src_timeline: &Arc<Timeline>,
3260 0 : dst_id: TimelineId,
3261 0 : start_lsn: Option<Lsn>,
3262 0 : timeline_create_guard: TimelineCreateGuard<'_>,
3263 0 : ctx: &RequestContext,
3264 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
3265 0 : self.branch_timeline_impl(src_timeline, dst_id, start_lsn, timeline_create_guard, ctx)
3266 0 : .await
3267 0 : }
3268 :
3269 690 : async fn branch_timeline_impl(
3270 690 : &self,
3271 690 : src_timeline: &Arc<Timeline>,
3272 690 : dst_id: TimelineId,
3273 690 : start_lsn: Option<Lsn>,
3274 690 : timeline_create_guard: TimelineCreateGuard<'_>,
3275 690 : _ctx: &RequestContext,
3276 690 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
3277 690 : let src_id = src_timeline.timeline_id;
3278 :
3279 : // We will validate our ancestor LSN in this function. Acquire the GC lock so that
3280 : // this check cannot race with GC, and the ancestor LSN is guaranteed to remain
3281 : // valid while we are creating the branch.
3282 690 : let _gc_cs = self.gc_cs.lock().await;
3283 :
3284 : // If no start LSN is specified, we branch the new timeline from the source timeline's last record LSN
3285 690 : let start_lsn = start_lsn.unwrap_or_else(|| {
3286 6 : let lsn = src_timeline.get_last_record_lsn();
3287 6 : info!("branching timeline {dst_id} from timeline {src_id} at last record LSN: {lsn}");
3288 6 : lsn
3289 690 : });
3290 690 :
3291 690 : // Ensure that `start_lsn` is valid, i.e. the LSN is within the PITR
3292 690 : // horizon on the source timeline
3293 690 : //
3294 690 : // We check it against both the planned GC cutoff stored in 'gc_info',
3295 690 : // and the 'latest_gc_cutoff' of the last GC that was performed. The
3296 690 : // planned GC cutoff in 'gc_info' is normally larger than
3297 690 : // 'latest_gc_cutoff_lsn', but beware of corner cases like if you just
3298 690 : // changed the GC settings for the tenant to make the PITR window
3299 690 : // larger, but some of the data was already removed by an earlier GC
3300 690 : // iteration.
3301 690 :
3302 690 : // check against last actual 'latest_gc_cutoff' first
3303 690 : let latest_gc_cutoff_lsn = src_timeline.get_latest_gc_cutoff_lsn();
3304 690 : src_timeline
3305 690 : .check_lsn_is_in_scope(start_lsn, &latest_gc_cutoff_lsn)
3306 690 : .context(format!(
3307 690 : "invalid branch start lsn: less than latest GC cutoff {}",
3308 690 : *latest_gc_cutoff_lsn,
3309 690 : ))
3310 690 : .map_err(CreateTimelineError::AncestorLsn)?;
3311 :
3312 : // and then the planned GC cutoff
3313 : {
3314 678 : let gc_info = src_timeline.gc_info.read().unwrap();
3315 678 : let cutoff = gc_info.min_cutoff();
3316 678 : if start_lsn < cutoff {
3317 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
3318 0 : "invalid branch start lsn: less than planned GC cutoff {cutoff}"
3319 0 : )));
3320 678 : }
3321 678 : }
3322 678 :
3323 678 : //
3324 678 : // The branch point is valid, and we are still holding the 'gc_cs' lock
3325 678 : // so that GC cannot advance the GC cutoff until we are finished.
3326 678 : // Proceed with the branch creation.
3327 678 : //
3328 678 :
3329 678 : // Determine prev-LSN for the new timeline. We can only determine it if
3330 678 : // the timeline was branched at the current end of the source timeline.
3331 678 : let RecordLsn {
3332 678 : last: src_last,
3333 678 : prev: src_prev,
3334 678 : } = src_timeline.get_last_record_rlsn();
3335 678 : let dst_prev = if src_last == start_lsn {
3336 648 : Some(src_prev)
3337 : } else {
3338 30 : None
3339 : };
3340 :
3341 : // Create the metadata file, noting the ancestor of the new timeline.
3342 : // There is initially no data in it, but all the read-calls know to look
3343 : // into the ancestor.
3344 678 : let metadata = TimelineMetadata::new(
3345 678 : start_lsn,
3346 678 : dst_prev,
3347 678 : Some(src_id),
3348 678 : start_lsn,
3349 678 : *src_timeline.latest_gc_cutoff_lsn.read(), // FIXME: should we hold onto this guard longer?
3350 678 : src_timeline.initdb_lsn,
3351 678 : src_timeline.pg_version,
3352 678 : );
3353 :
3354 678 : let uninitialized_timeline = self
3355 678 : .prepare_new_timeline(
3356 678 : dst_id,
3357 678 : &metadata,
3358 678 : timeline_create_guard,
3359 678 : start_lsn + 1,
3360 678 : Some(Arc::clone(src_timeline)),
3361 678 : src_timeline.last_aux_file_policy.load(),
3362 678 : )
3363 0 : .await?;
3364 :
3365 678 : let new_timeline = uninitialized_timeline.finish_creation()?;
3366 :
3367 : // Root timeline gets its layers during creation and uploads them along with the metadata.
3368 : // A branch timeline though, when created, can get no writes for some time, hence won't get any layers created.
3369 : // We still need to upload its metadata eagerly: if other nodes `attach` the tenant and miss this timeline, their GC
3370 : // could get incorrect information and remove more layers, than needed.
3371 : // See also https://github.com/neondatabase/neon/issues/3865
3372 678 : new_timeline
3373 678 : .remote_client
3374 678 : .schedule_index_upload_for_full_metadata_update(&metadata)
3375 678 : .context("branch initial metadata upload")?;
3376 :
3377 678 : Ok(new_timeline)
3378 690 : }
3379 :
3380 : /// For unit tests, make this visible so that other modules can directly create timelines
3381 : #[cfg(test)]
3382 6 : #[tracing::instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
3383 : pub(crate) async fn bootstrap_timeline_test(
3384 : &self,
3385 : timeline_id: TimelineId,
3386 : pg_version: u32,
3387 : load_existing_initdb: Option<TimelineId>,
3388 : ctx: &RequestContext,
3389 : ) -> anyhow::Result<Arc<Timeline>> {
3390 : let create_guard = self.create_timeline_create_guard(timeline_id).unwrap();
3391 : self.bootstrap_timeline(
3392 : timeline_id,
3393 : pg_version,
3394 : load_existing_initdb,
3395 : create_guard,
3396 : ctx,
3397 : )
3398 : .await
3399 : }
3400 :
3401 0 : async fn upload_initdb(
3402 0 : &self,
3403 0 : timelines_path: &Utf8PathBuf,
3404 0 : pgdata_path: &Utf8PathBuf,
3405 0 : timeline_id: &TimelineId,
3406 0 : ) -> anyhow::Result<()> {
3407 0 : let temp_path = timelines_path.join(format!(
3408 0 : "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
3409 0 : ));
3410 0 :
3411 0 : scopeguard::defer! {
3412 0 : if let Err(e) = fs::remove_file(&temp_path) {
3413 0 : error!("Failed to remove temporary initdb archive '{temp_path}': {e}");
3414 0 : }
3415 0 : }
3416 :
3417 0 : let (pgdata_zstd, tar_zst_size) = create_zst_tarball(pgdata_path, &temp_path).await?;
3418 : const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
3419 0 : if tar_zst_size > INITDB_TAR_ZST_WARN_LIMIT {
3420 0 : warn!(
3421 0 : "compressed {temp_path} size of {tar_zst_size} is above limit {INITDB_TAR_ZST_WARN_LIMIT}."
3422 : );
3423 0 : }
3424 :
3425 0 : pausable_failpoint!("before-initdb-upload");
3426 :
3427 0 : backoff::retry(
3428 0 : || async {
3429 0 : self::remote_timeline_client::upload_initdb_dir(
3430 0 : &self.remote_storage,
3431 0 : &self.tenant_shard_id.tenant_id,
3432 0 : timeline_id,
3433 0 : pgdata_zstd.try_clone().await?,
3434 0 : tar_zst_size,
3435 0 : &self.cancel,
3436 : )
3437 0 : .await
3438 0 : },
3439 0 : |_| false,
3440 0 : 3,
3441 0 : u32::MAX,
3442 0 : "persist_initdb_tar_zst",
3443 0 : &self.cancel,
3444 0 : )
3445 0 : .await
3446 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
3447 0 : .and_then(|x| x)
3448 0 : }
3449 :
3450 : /// - run initdb to init temporary instance and get bootstrap data
3451 : /// - after initialization completes, tar up the temp dir and upload it to S3.
3452 : ///
3453 : /// The caller is responsible for activating the returned timeline.
3454 6 : async fn bootstrap_timeline(
3455 6 : &self,
3456 6 : timeline_id: TimelineId,
3457 6 : pg_version: u32,
3458 6 : load_existing_initdb: Option<TimelineId>,
3459 6 : timeline_create_guard: TimelineCreateGuard<'_>,
3460 6 : ctx: &RequestContext,
3461 6 : ) -> anyhow::Result<Arc<Timeline>> {
3462 6 : // create a `tenant/{tenant_id}/timelines/basebackup-{timeline_id}.{TEMP_FILE_SUFFIX}/`
3463 6 : // temporary directory for basebackup files for the given timeline.
3464 6 :
3465 6 : let timelines_path = self.conf.timelines_path(&self.tenant_shard_id);
3466 6 : let pgdata_path = path_with_suffix_extension(
3467 6 : timelines_path.join(format!("basebackup-{timeline_id}")),
3468 6 : TEMP_FILE_SUFFIX,
3469 6 : );
3470 6 :
3471 6 : // Remove whatever was left from the previous runs: safe because TimelineCreateGuard guarantees
3472 6 : // we won't race with other creations or existent timelines with the same path.
3473 6 : if pgdata_path.exists() {
3474 0 : fs::remove_dir_all(&pgdata_path).with_context(|| {
3475 0 : format!("Failed to remove already existing initdb directory: {pgdata_path}")
3476 0 : })?;
3477 6 : }
3478 :
3479 : // this new directory is very temporary, set to remove it immediately after bootstrap, we don't need it
3480 6 : scopeguard::defer! {
3481 6 : if let Err(e) = fs::remove_dir_all(&pgdata_path) {
3482 6 : // this is unlikely, but we will remove the directory on pageserver restart or another bootstrap call
3483 6 : error!("Failed to remove temporary initdb directory '{pgdata_path}': {e}");
3484 6 : }
3485 6 : }
3486 6 : if let Some(existing_initdb_timeline_id) = load_existing_initdb {
3487 6 : if existing_initdb_timeline_id != timeline_id {
3488 0 : let source_path = &remote_initdb_archive_path(
3489 0 : &self.tenant_shard_id.tenant_id,
3490 0 : &existing_initdb_timeline_id,
3491 0 : );
3492 0 : let dest_path =
3493 0 : &remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &timeline_id);
3494 0 :
3495 0 : // if this fails, it will get retried by retried control plane requests
3496 0 : self.remote_storage
3497 0 : .copy_object(source_path, dest_path, &self.cancel)
3498 0 : .await
3499 0 : .context("copy initdb tar")?;
3500 6 : }
3501 6 : let (initdb_tar_zst_path, initdb_tar_zst) =
3502 6 : self::remote_timeline_client::download_initdb_tar_zst(
3503 6 : self.conf,
3504 6 : &self.remote_storage,
3505 6 : &self.tenant_shard_id,
3506 6 : &existing_initdb_timeline_id,
3507 6 : &self.cancel,
3508 6 : )
3509 2206 : .await
3510 6 : .context("download initdb tar")?;
3511 :
3512 6 : scopeguard::defer! {
3513 6 : if let Err(e) = fs::remove_file(&initdb_tar_zst_path) {
3514 6 : error!("Failed to remove temporary initdb archive '{initdb_tar_zst_path}': {e}");
3515 6 : }
3516 6 : }
3517 6 :
3518 6 : let buf_read =
3519 6 : BufReader::with_capacity(remote_timeline_client::BUFFER_SIZE, initdb_tar_zst);
3520 6 : extract_zst_tarball(&pgdata_path, buf_read)
3521 31304 : .await
3522 6 : .context("extract initdb tar")?;
3523 : } else {
3524 : // Init temporarily repo to get bootstrap data, this creates a directory in the `pgdata_path` path
3525 0 : run_initdb(self.conf, &pgdata_path, pg_version, &self.cancel).await?;
3526 :
3527 : // Upload the created data dir to S3
3528 0 : if self.tenant_shard_id().is_shard_zero() {
3529 0 : self.upload_initdb(&timelines_path, &pgdata_path, &timeline_id)
3530 0 : .await?;
3531 0 : }
3532 : }
3533 6 : let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
3534 6 :
3535 6 : // Import the contents of the data directory at the initial checkpoint
3536 6 : // LSN, and any WAL after that.
3537 6 : // Initdb lsn will be equal to last_record_lsn which will be set after import.
3538 6 : // Because we know it upfront avoid having an option or dummy zero value by passing it to the metadata.
3539 6 : let new_metadata = TimelineMetadata::new(
3540 6 : Lsn(0),
3541 6 : None,
3542 6 : None,
3543 6 : Lsn(0),
3544 6 : pgdata_lsn,
3545 6 : pgdata_lsn,
3546 6 : pg_version,
3547 6 : );
3548 6 : let raw_timeline = self
3549 6 : .prepare_new_timeline(
3550 6 : timeline_id,
3551 6 : &new_metadata,
3552 6 : timeline_create_guard,
3553 6 : pgdata_lsn,
3554 6 : None,
3555 6 : None,
3556 6 : )
3557 0 : .await?;
3558 :
3559 6 : let tenant_shard_id = raw_timeline.owning_tenant.tenant_shard_id;
3560 6 : let unfinished_timeline = raw_timeline.raw_timeline()?;
3561 :
3562 : // Flush the new layer files to disk, before we make the timeline as available to
3563 : // the outside world.
3564 : //
3565 : // Flush loop needs to be spawned in order to be able to flush.
3566 6 : unfinished_timeline.maybe_spawn_flush_loop();
3567 6 :
3568 6 : import_datadir::import_timeline_from_postgres_datadir(
3569 6 : unfinished_timeline,
3570 6 : &pgdata_path,
3571 6 : pgdata_lsn,
3572 6 : ctx,
3573 6 : )
3574 28422 : .await
3575 6 : .with_context(|| {
3576 0 : format!("Failed to import pgdatadir for timeline {tenant_shard_id}/{timeline_id}")
3577 6 : })?;
3578 :
3579 6 : fail::fail_point!("before-checkpoint-new-timeline", |_| {
3580 0 : anyhow::bail!("failpoint before-checkpoint-new-timeline");
3581 6 : });
3582 :
3583 6 : unfinished_timeline
3584 6 : .freeze_and_flush()
3585 6 : .await
3586 6 : .with_context(|| {
3587 0 : format!(
3588 0 : "Failed to flush after pgdatadir import for timeline {tenant_shard_id}/{timeline_id}"
3589 0 : )
3590 6 : })?;
3591 :
3592 : // All done!
3593 6 : let timeline = raw_timeline.finish_creation()?;
3594 :
3595 6 : Ok(timeline)
3596 6 : }
3597 :
3598 : /// Call this before constructing a timeline, to build its required structures
3599 1224 : fn build_timeline_resources(&self, timeline_id: TimelineId) -> TimelineResources {
3600 1224 : let remote_client = RemoteTimelineClient::new(
3601 1224 : self.remote_storage.clone(),
3602 1224 : self.deletion_queue_client.clone(),
3603 1224 : self.conf,
3604 1224 : self.tenant_shard_id,
3605 1224 : timeline_id,
3606 1224 : self.generation,
3607 1224 : );
3608 1224 : TimelineResources {
3609 1224 : remote_client,
3610 1224 : timeline_get_throttle: self.timeline_get_throttle.clone(),
3611 1224 : l0_flush_global_state: self.l0_flush_global_state.clone(),
3612 1224 : }
3613 1224 : }
3614 :
3615 : /// Creates intermediate timeline structure and its files.
3616 : ///
3617 : /// An empty layer map is initialized, and new data and WAL can be imported starting
3618 : /// at 'disk_consistent_lsn'. After any initial data has been imported, call
3619 : /// `finish_creation` to insert the Timeline into the timelines map.
3620 1224 : async fn prepare_new_timeline<'a>(
3621 1224 : &'a self,
3622 1224 : new_timeline_id: TimelineId,
3623 1224 : new_metadata: &TimelineMetadata,
3624 1224 : create_guard: TimelineCreateGuard<'a>,
3625 1224 : start_lsn: Lsn,
3626 1224 : ancestor: Option<Arc<Timeline>>,
3627 1224 : last_aux_file_policy: Option<AuxFilePolicy>,
3628 1224 : ) -> anyhow::Result<UninitializedTimeline> {
3629 1224 : let tenant_shard_id = self.tenant_shard_id;
3630 1224 :
3631 1224 : let resources = self.build_timeline_resources(new_timeline_id);
3632 1224 : resources
3633 1224 : .remote_client
3634 1224 : .init_upload_queue_for_empty_remote(new_metadata)?;
3635 :
3636 1224 : let timeline_struct = self
3637 1224 : .create_timeline_struct(
3638 1224 : new_timeline_id,
3639 1224 : new_metadata,
3640 1224 : ancestor,
3641 1224 : resources,
3642 1224 : CreateTimelineCause::Load,
3643 1224 : last_aux_file_policy,
3644 1224 : )
3645 1224 : .context("Failed to create timeline data structure")?;
3646 :
3647 1224 : timeline_struct.init_empty_layer_map(start_lsn);
3648 :
3649 1224 : if let Err(e) = self
3650 1224 : .create_timeline_files(&create_guard.timeline_path)
3651 0 : .await
3652 : {
3653 0 : error!("Failed to create initial files for timeline {tenant_shard_id}/{new_timeline_id}, cleaning up: {e:?}");
3654 0 : cleanup_timeline_directory(create_guard);
3655 0 : return Err(e);
3656 1224 : }
3657 1224 :
3658 1224 : debug!(
3659 0 : "Successfully created initial files for timeline {tenant_shard_id}/{new_timeline_id}"
3660 : );
3661 :
3662 1224 : Ok(UninitializedTimeline::new(
3663 1224 : self,
3664 1224 : new_timeline_id,
3665 1224 : Some((timeline_struct, create_guard)),
3666 1224 : ))
3667 1224 : }
3668 :
3669 1224 : async fn create_timeline_files(&self, timeline_path: &Utf8Path) -> anyhow::Result<()> {
3670 1224 : crashsafe::create_dir(timeline_path).context("Failed to create timeline directory")?;
3671 :
3672 1224 : fail::fail_point!("after-timeline-dir-creation", |_| {
3673 0 : anyhow::bail!("failpoint after-timeline-dir-creation");
3674 1224 : });
3675 :
3676 1224 : Ok(())
3677 1224 : }
3678 :
3679 : /// Get a guard that provides exclusive access to the timeline directory, preventing
3680 : /// concurrent attempts to create the same timeline.
3681 1242 : fn create_timeline_create_guard(
3682 1242 : &self,
3683 1242 : timeline_id: TimelineId,
3684 1242 : ) -> Result<TimelineCreateGuard, TimelineExclusionError> {
3685 1242 : let tenant_shard_id = self.tenant_shard_id;
3686 1242 :
3687 1242 : let timeline_path = self.conf.timeline_path(&tenant_shard_id, &timeline_id);
3688 :
3689 1242 : let create_guard = TimelineCreateGuard::new(self, timeline_id, timeline_path.clone())?;
3690 :
3691 : // At this stage, we have got exclusive access to in-memory state for this timeline ID
3692 : // for creation.
3693 : // A timeline directory should never exist on disk already:
3694 : // - a previous failed creation would have cleaned up after itself
3695 : // - a pageserver restart would clean up timeline directories that don't have valid remote state
3696 : //
3697 : // Therefore it is an unexpected internal error to encounter a timeline directory already existing here,
3698 : // this error may indicate a bug in cleanup on failed creations.
3699 1236 : if timeline_path.exists() {
3700 0 : return Err(TimelineExclusionError::Other(anyhow::anyhow!(
3701 0 : "Timeline directory already exists! This is a bug."
3702 0 : )));
3703 1236 : }
3704 1236 :
3705 1236 : Ok(create_guard)
3706 1242 : }
3707 :
3708 : /// Gathers inputs from all of the timelines to produce a sizing model input.
3709 : ///
3710 : /// Future is cancellation safe. Only one calculation can be running at once per tenant.
3711 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
3712 : pub async fn gather_size_inputs(
3713 : &self,
3714 : // `max_retention_period` overrides the cutoff that is used to calculate the size
3715 : // (only if it is shorter than the real cutoff).
3716 : max_retention_period: Option<u64>,
3717 : cause: LogicalSizeCalculationCause,
3718 : cancel: &CancellationToken,
3719 : ctx: &RequestContext,
3720 : ) -> Result<size::ModelInputs, size::CalculateSyntheticSizeError> {
3721 : let logical_sizes_at_once = self
3722 : .conf
3723 : .concurrent_tenant_size_logical_size_queries
3724 : .inner();
3725 :
3726 : // TODO: Having a single mutex block concurrent reads is not great for performance.
3727 : //
3728 : // But the only case where we need to run multiple of these at once is when we
3729 : // request a size for a tenant manually via API, while another background calculation
3730 : // is in progress (which is not a common case).
3731 : //
3732 : // See more for on the issue #2748 condenced out of the initial PR review.
3733 : let mut shared_cache = tokio::select! {
3734 : locked = self.cached_logical_sizes.lock() => locked,
3735 : _ = cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
3736 : _ = self.cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
3737 : };
3738 :
3739 : size::gather_inputs(
3740 : self,
3741 : logical_sizes_at_once,
3742 : max_retention_period,
3743 : &mut shared_cache,
3744 : cause,
3745 : cancel,
3746 : ctx,
3747 : )
3748 : .await
3749 : }
3750 :
3751 : /// Calculate synthetic tenant size and cache the result.
3752 : /// This is periodically called by background worker.
3753 : /// result is cached in tenant struct
3754 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
3755 : pub async fn calculate_synthetic_size(
3756 : &self,
3757 : cause: LogicalSizeCalculationCause,
3758 : cancel: &CancellationToken,
3759 : ctx: &RequestContext,
3760 : ) -> Result<u64, size::CalculateSyntheticSizeError> {
3761 : let inputs = self.gather_size_inputs(None, cause, cancel, ctx).await?;
3762 :
3763 : let size = inputs.calculate();
3764 :
3765 : self.set_cached_synthetic_size(size);
3766 :
3767 : Ok(size)
3768 : }
3769 :
3770 : /// Cache given synthetic size and update the metric value
3771 0 : pub fn set_cached_synthetic_size(&self, size: u64) {
3772 0 : self.cached_synthetic_tenant_size
3773 0 : .store(size, Ordering::Relaxed);
3774 0 :
3775 0 : // Only shard zero should be calculating synthetic sizes
3776 0 : debug_assert!(self.shard_identity.is_shard_zero());
3777 :
3778 0 : TENANT_SYNTHETIC_SIZE_METRIC
3779 0 : .get_metric_with_label_values(&[&self.tenant_shard_id.tenant_id.to_string()])
3780 0 : .unwrap()
3781 0 : .set(size);
3782 0 : }
3783 :
3784 0 : pub fn cached_synthetic_size(&self) -> u64 {
3785 0 : self.cached_synthetic_tenant_size.load(Ordering::Relaxed)
3786 0 : }
3787 :
3788 : /// Flush any in-progress layers, schedule uploads, and wait for uploads to complete.
3789 : ///
3790 : /// This function can take a long time: callers should wrap it in a timeout if calling
3791 : /// from an external API handler.
3792 : ///
3793 : /// Cancel-safety: cancelling this function may leave I/O running, but such I/O is
3794 : /// still bounded by tenant/timeline shutdown.
3795 0 : #[tracing::instrument(skip_all)]
3796 : pub(crate) async fn flush_remote(&self) -> anyhow::Result<()> {
3797 : let timelines = self.timelines.lock().unwrap().clone();
3798 :
3799 0 : async fn flush_timeline(_gate: GateGuard, timeline: Arc<Timeline>) -> anyhow::Result<()> {
3800 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Flushing...");
3801 0 : timeline.freeze_and_flush().await?;
3802 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Waiting for uploads...");
3803 0 : timeline.remote_client.wait_completion().await?;
3804 :
3805 0 : Ok(())
3806 0 : }
3807 :
3808 : // We do not use a JoinSet for these tasks, because we don't want them to be
3809 : // aborted when this function's future is cancelled: they should stay alive
3810 : // holding their GateGuard until they complete, to ensure their I/Os complete
3811 : // before Timeline shutdown completes.
3812 : let mut results = FuturesUnordered::new();
3813 :
3814 : for (_timeline_id, timeline) in timelines {
3815 : // Run each timeline's flush in a task holding the timeline's gate: this
3816 : // means that if this function's future is cancelled, the Timeline shutdown
3817 : // will still wait for any I/O in here to complete.
3818 : let Ok(gate) = timeline.gate.enter() else {
3819 : continue;
3820 : };
3821 0 : let jh = tokio::task::spawn(async move { flush_timeline(gate, timeline).await });
3822 : results.push(jh);
3823 : }
3824 :
3825 : while let Some(r) = results.next().await {
3826 : if let Err(e) = r {
3827 : if !e.is_cancelled() && !e.is_panic() {
3828 : tracing::error!("unexpected join error: {e:?}");
3829 : }
3830 : }
3831 : }
3832 :
3833 : // The flushes we did above were just writes, but the Tenant might have had
3834 : // pending deletions as well from recent compaction/gc: we want to flush those
3835 : // as well. This requires flushing the global delete queue. This is cheap
3836 : // because it's typically a no-op.
3837 : match self.deletion_queue_client.flush_execute().await {
3838 : Ok(_) => {}
3839 : Err(DeletionQueueError::ShuttingDown) => {}
3840 : }
3841 :
3842 : Ok(())
3843 : }
3844 :
3845 0 : pub(crate) fn get_tenant_conf(&self) -> TenantConfOpt {
3846 0 : self.tenant_conf.load().tenant_conf.clone()
3847 0 : }
3848 :
3849 : /// How much local storage would this tenant like to have? It can cope with
3850 : /// less than this (via eviction and on-demand downloads), but this function enables
3851 : /// the Tenant to advertise how much storage it would prefer to have to provide fast I/O
3852 : /// by keeping important things on local disk.
3853 : ///
3854 : /// This is a heuristic, not a guarantee: tenants that are long-idle will actually use less
3855 : /// than they report here, due to layer eviction. Tenants with many active branches may
3856 : /// actually use more than they report here.
3857 0 : pub(crate) fn local_storage_wanted(&self) -> u64 {
3858 0 : let timelines = self.timelines.lock().unwrap();
3859 0 :
3860 0 : // Heuristic: we use the max() of the timelines' visible sizes, rather than the sum. This
3861 0 : // reflects the observation that on tenants with multiple large branches, typically only one
3862 0 : // of them is used actively enough to occupy space on disk.
3863 0 : timelines
3864 0 : .values()
3865 0 : .map(|t| t.metrics.visible_physical_size_gauge.get())
3866 0 : .max()
3867 0 : .unwrap_or(0)
3868 0 : }
3869 : }
3870 :
3871 : /// Create the cluster temporarily in 'initdbpath' directory inside the repository
3872 : /// to get bootstrap data for timeline initialization.
3873 0 : async fn run_initdb(
3874 0 : conf: &'static PageServerConf,
3875 0 : initdb_target_dir: &Utf8Path,
3876 0 : pg_version: u32,
3877 0 : cancel: &CancellationToken,
3878 0 : ) -> Result<(), InitdbError> {
3879 0 : let initdb_bin_path = conf
3880 0 : .pg_bin_dir(pg_version)
3881 0 : .map_err(InitdbError::Other)?
3882 0 : .join("initdb");
3883 0 : let initdb_lib_dir = conf.pg_lib_dir(pg_version).map_err(InitdbError::Other)?;
3884 0 : info!(
3885 0 : "running {} in {}, libdir: {}",
3886 : initdb_bin_path, initdb_target_dir, initdb_lib_dir,
3887 : );
3888 :
3889 0 : let _permit = INIT_DB_SEMAPHORE.acquire().await;
3890 :
3891 0 : let initdb_command = tokio::process::Command::new(&initdb_bin_path)
3892 0 : .args(["-D", initdb_target_dir.as_ref()])
3893 0 : .args(["-U", &conf.superuser])
3894 0 : .args(["-E", "utf8"])
3895 0 : .arg("--no-instructions")
3896 0 : .arg("--no-sync")
3897 0 : .env_clear()
3898 0 : .env("LD_LIBRARY_PATH", &initdb_lib_dir)
3899 0 : .env("DYLD_LIBRARY_PATH", &initdb_lib_dir)
3900 0 : .stdin(std::process::Stdio::null())
3901 0 : // stdout invocation produces the same output every time, we don't need it
3902 0 : .stdout(std::process::Stdio::null())
3903 0 : // we would be interested in the stderr output, if there was any
3904 0 : .stderr(std::process::Stdio::piped())
3905 0 : .spawn()?;
3906 :
3907 : // Ideally we'd select here with the cancellation token, but the problem is that
3908 : // we can't safely terminate initdb: it launches processes of its own, and killing
3909 : // initdb doesn't kill them. After we return from this function, we want the target
3910 : // directory to be able to be cleaned up.
3911 : // See https://github.com/neondatabase/neon/issues/6385
3912 0 : let initdb_output = initdb_command.wait_with_output().await?;
3913 0 : if !initdb_output.status.success() {
3914 0 : return Err(InitdbError::Failed(
3915 0 : initdb_output.status,
3916 0 : initdb_output.stderr,
3917 0 : ));
3918 0 : }
3919 0 :
3920 0 : // This isn't true cancellation support, see above. Still return an error to
3921 0 : // excercise the cancellation code path.
3922 0 : if cancel.is_cancelled() {
3923 0 : return Err(InitdbError::Cancelled);
3924 0 : }
3925 0 :
3926 0 : Ok(())
3927 0 : }
3928 :
3929 : /// Dump contents of a layer file to stdout.
3930 0 : pub async fn dump_layerfile_from_path(
3931 0 : path: &Utf8Path,
3932 0 : verbose: bool,
3933 0 : ctx: &RequestContext,
3934 0 : ) -> anyhow::Result<()> {
3935 : use std::os::unix::fs::FileExt;
3936 :
3937 : // All layer files start with a two-byte "magic" value, to identify the kind of
3938 : // file.
3939 0 : let file = File::open(path)?;
3940 0 : let mut header_buf = [0u8; 2];
3941 0 : file.read_exact_at(&mut header_buf, 0)?;
3942 :
3943 0 : match u16::from_be_bytes(header_buf) {
3944 : crate::IMAGE_FILE_MAGIC => {
3945 0 : ImageLayer::new_for_path(path, file)?
3946 0 : .dump(verbose, ctx)
3947 0 : .await?
3948 : }
3949 : crate::DELTA_FILE_MAGIC => {
3950 0 : DeltaLayer::new_for_path(path, file)?
3951 0 : .dump(verbose, ctx)
3952 0 : .await?
3953 : }
3954 0 : magic => bail!("unrecognized magic identifier: {:?}", magic),
3955 : }
3956 :
3957 0 : Ok(())
3958 0 : }
3959 :
3960 : #[cfg(test)]
3961 : pub(crate) mod harness {
3962 : use bytes::{Bytes, BytesMut};
3963 : use once_cell::sync::OnceCell;
3964 : use pageserver_api::models::ShardParameters;
3965 : use pageserver_api::shard::ShardIndex;
3966 : use utils::logging;
3967 :
3968 : use crate::deletion_queue::mock::MockDeletionQueue;
3969 : use crate::l0_flush::L0FlushConfig;
3970 : use crate::walredo::apply_neon;
3971 : use crate::{repository::Key, walrecord::NeonWalRecord};
3972 :
3973 : use super::*;
3974 : use hex_literal::hex;
3975 : use utils::id::TenantId;
3976 :
3977 : pub const TIMELINE_ID: TimelineId =
3978 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
3979 : pub const NEW_TIMELINE_ID: TimelineId =
3980 : TimelineId::from_array(hex!("AA223344556677881122334455667788"));
3981 :
3982 : /// Convenience function to create a page image with given string as the only content
3983 15086143 : pub fn test_img(s: &str) -> Bytes {
3984 15086143 : let mut buf = BytesMut::new();
3985 15086143 : buf.extend_from_slice(s.as_bytes());
3986 15086143 : buf.resize(64, 0);
3987 15086143 :
3988 15086143 : buf.freeze()
3989 15086143 : }
3990 :
3991 : impl From<TenantConf> for TenantConfOpt {
3992 570 : fn from(tenant_conf: TenantConf) -> Self {
3993 570 : Self {
3994 570 : checkpoint_distance: Some(tenant_conf.checkpoint_distance),
3995 570 : checkpoint_timeout: Some(tenant_conf.checkpoint_timeout),
3996 570 : compaction_target_size: Some(tenant_conf.compaction_target_size),
3997 570 : compaction_period: Some(tenant_conf.compaction_period),
3998 570 : compaction_threshold: Some(tenant_conf.compaction_threshold),
3999 570 : compaction_algorithm: Some(tenant_conf.compaction_algorithm),
4000 570 : gc_horizon: Some(tenant_conf.gc_horizon),
4001 570 : gc_period: Some(tenant_conf.gc_period),
4002 570 : image_creation_threshold: Some(tenant_conf.image_creation_threshold),
4003 570 : pitr_interval: Some(tenant_conf.pitr_interval),
4004 570 : walreceiver_connect_timeout: Some(tenant_conf.walreceiver_connect_timeout),
4005 570 : lagging_wal_timeout: Some(tenant_conf.lagging_wal_timeout),
4006 570 : max_lsn_wal_lag: Some(tenant_conf.max_lsn_wal_lag),
4007 570 : eviction_policy: Some(tenant_conf.eviction_policy),
4008 570 : min_resident_size_override: tenant_conf.min_resident_size_override,
4009 570 : evictions_low_residence_duration_metric_threshold: Some(
4010 570 : tenant_conf.evictions_low_residence_duration_metric_threshold,
4011 570 : ),
4012 570 : heatmap_period: Some(tenant_conf.heatmap_period),
4013 570 : lazy_slru_download: Some(tenant_conf.lazy_slru_download),
4014 570 : timeline_get_throttle: Some(tenant_conf.timeline_get_throttle),
4015 570 : image_layer_creation_check_threshold: Some(
4016 570 : tenant_conf.image_layer_creation_check_threshold,
4017 570 : ),
4018 570 : switch_aux_file_policy: Some(tenant_conf.switch_aux_file_policy),
4019 570 : lsn_lease_length: Some(tenant_conf.lsn_lease_length),
4020 570 : lsn_lease_length_for_ts: Some(tenant_conf.lsn_lease_length_for_ts),
4021 570 : }
4022 570 : }
4023 : }
4024 :
4025 : pub struct TenantHarness {
4026 : pub conf: &'static PageServerConf,
4027 : pub tenant_conf: TenantConf,
4028 : pub tenant_shard_id: TenantShardId,
4029 : pub generation: Generation,
4030 : pub shard: ShardIndex,
4031 : pub remote_storage: GenericRemoteStorage,
4032 : pub remote_fs_dir: Utf8PathBuf,
4033 : pub deletion_queue: MockDeletionQueue,
4034 : }
4035 :
4036 : static LOG_HANDLE: OnceCell<()> = OnceCell::new();
4037 :
4038 612 : pub(crate) fn setup_logging() {
4039 612 : LOG_HANDLE.get_or_init(|| {
4040 576 : logging::init(
4041 576 : logging::LogFormat::Test,
4042 576 : // enable it in case the tests exercise code paths that use
4043 576 : // debug_assert_current_span_has_tenant_and_timeline_id
4044 576 : logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
4045 576 : logging::Output::Stdout,
4046 576 : )
4047 576 : .expect("Failed to init test logging")
4048 612 : });
4049 612 : }
4050 :
4051 : impl TenantHarness {
4052 570 : pub async fn create_custom(
4053 570 : test_name: &'static str,
4054 570 : tenant_conf: TenantConf,
4055 570 : tenant_id: TenantId,
4056 570 : shard_identity: ShardIdentity,
4057 570 : generation: Generation,
4058 570 : ) -> anyhow::Result<Self> {
4059 570 : setup_logging();
4060 570 :
4061 570 : let repo_dir = PageServerConf::test_repo_dir(test_name);
4062 570 : let _ = fs::remove_dir_all(&repo_dir);
4063 570 : fs::create_dir_all(&repo_dir)?;
4064 :
4065 570 : let conf = PageServerConf::dummy_conf(repo_dir);
4066 570 : // Make a static copy of the config. This can never be free'd, but that's
4067 570 : // OK in a test.
4068 570 : let conf: &'static PageServerConf = Box::leak(Box::new(conf));
4069 570 :
4070 570 : let shard = shard_identity.shard_index();
4071 570 : let tenant_shard_id = TenantShardId {
4072 570 : tenant_id,
4073 570 : shard_number: shard.shard_number,
4074 570 : shard_count: shard.shard_count,
4075 570 : };
4076 570 : fs::create_dir_all(conf.tenant_path(&tenant_shard_id))?;
4077 570 : fs::create_dir_all(conf.timelines_path(&tenant_shard_id))?;
4078 :
4079 : use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
4080 570 : let remote_fs_dir = conf.workdir.join("localfs");
4081 570 : std::fs::create_dir_all(&remote_fs_dir).unwrap();
4082 570 : let config = RemoteStorageConfig {
4083 570 : storage: RemoteStorageKind::LocalFs {
4084 570 : local_path: remote_fs_dir.clone(),
4085 570 : },
4086 570 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
4087 570 : };
4088 570 : let remote_storage = GenericRemoteStorage::from_config(&config).await.unwrap();
4089 570 : let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()));
4090 570 :
4091 570 : Ok(Self {
4092 570 : conf,
4093 570 : tenant_conf,
4094 570 : tenant_shard_id,
4095 570 : generation,
4096 570 : shard,
4097 570 : remote_storage,
4098 570 : remote_fs_dir,
4099 570 : deletion_queue,
4100 570 : })
4101 570 : }
4102 :
4103 534 : pub async fn create(test_name: &'static str) -> anyhow::Result<Self> {
4104 534 : // Disable automatic GC and compaction to make the unit tests more deterministic.
4105 534 : // The tests perform them manually if needed.
4106 534 : let tenant_conf = TenantConf {
4107 534 : gc_period: Duration::ZERO,
4108 534 : compaction_period: Duration::ZERO,
4109 534 : ..TenantConf::default()
4110 534 : };
4111 534 : let tenant_id = TenantId::generate();
4112 534 : let shard = ShardIdentity::unsharded();
4113 534 : Self::create_custom(
4114 534 : test_name,
4115 534 : tenant_conf,
4116 534 : tenant_id,
4117 534 : shard,
4118 534 : Generation::new(0xdeadbeef),
4119 534 : )
4120 0 : .await
4121 534 : }
4122 :
4123 60 : pub fn span(&self) -> tracing::Span {
4124 60 : info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
4125 60 : }
4126 :
4127 570 : pub(crate) async fn load(&self) -> (Arc<Tenant>, RequestContext) {
4128 570 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
4129 570 : (
4130 570 : self.do_try_load(&ctx)
4131 2289 : .await
4132 570 : .expect("failed to load test tenant"),
4133 570 : ctx,
4134 570 : )
4135 570 : }
4136 :
4137 570 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
4138 : pub(crate) async fn do_try_load(
4139 : &self,
4140 : ctx: &RequestContext,
4141 : ) -> anyhow::Result<Arc<Tenant>> {
4142 : let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
4143 :
4144 : let tenant = Arc::new(Tenant::new(
4145 : TenantState::Loading,
4146 : self.conf,
4147 : AttachedTenantConf::try_from(LocationConf::attached_single(
4148 : TenantConfOpt::from(self.tenant_conf.clone()),
4149 : self.generation,
4150 : &ShardParameters::default(),
4151 : ))
4152 : .unwrap(),
4153 : // This is a legacy/test code path: sharding isn't supported here.
4154 : ShardIdentity::unsharded(),
4155 : Some(walredo_mgr),
4156 : self.tenant_shard_id,
4157 : self.remote_storage.clone(),
4158 : self.deletion_queue.new_client(),
4159 : // TODO: ideally we should run all unit tests with both configs
4160 : L0FlushGlobalState::new(L0FlushConfig::default()),
4161 : ));
4162 :
4163 : let preload = tenant
4164 : .preload(&self.remote_storage, CancellationToken::new())
4165 : .await?;
4166 : tenant.attach(Some(preload), ctx).await?;
4167 :
4168 : tenant.state.send_replace(TenantState::Active);
4169 : for timeline in tenant.timelines.lock().unwrap().values() {
4170 : timeline.set_state(TimelineState::Active);
4171 : }
4172 : Ok(tenant)
4173 : }
4174 :
4175 6 : pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
4176 6 : self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
4177 6 : }
4178 : }
4179 :
4180 : // Mock WAL redo manager that doesn't do much
4181 : pub(crate) struct TestRedoManager;
4182 :
4183 : impl TestRedoManager {
4184 : /// # Cancel-Safety
4185 : ///
4186 : /// This method is cancellation-safe.
4187 1038 : pub async fn request_redo(
4188 1038 : &self,
4189 1038 : key: Key,
4190 1038 : lsn: Lsn,
4191 1038 : base_img: Option<(Lsn, Bytes)>,
4192 1038 : records: Vec<(Lsn, NeonWalRecord)>,
4193 1038 : _pg_version: u32,
4194 1038 : ) -> Result<Bytes, walredo::Error> {
4195 1386 : let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
4196 1038 : if records_neon {
4197 : // For Neon wal records, we can decode without spawning postgres, so do so.
4198 1038 : let mut page = match (base_img, records.first()) {
4199 1026 : (Some((_lsn, img)), _) => {
4200 1026 : let mut page = BytesMut::new();
4201 1026 : page.extend_from_slice(&img);
4202 1026 : page
4203 : }
4204 12 : (_, Some((_lsn, rec))) if rec.will_init() => BytesMut::new(),
4205 : _ => {
4206 0 : panic!("Neon WAL redo requires base image or will init record");
4207 : }
4208 : };
4209 :
4210 2424 : for (record_lsn, record) in records {
4211 1386 : apply_neon::apply_in_neon(&record, record_lsn, key, &mut page)?;
4212 : }
4213 1038 : Ok(page.freeze())
4214 : } else {
4215 : // We never spawn a postgres walredo process in unit tests: just log what we might have done.
4216 0 : let s = format!(
4217 0 : "redo for {} to get to {}, with {} and {} records",
4218 0 : key,
4219 0 : lsn,
4220 0 : if base_img.is_some() {
4221 0 : "base image"
4222 : } else {
4223 0 : "no base image"
4224 : },
4225 0 : records.len()
4226 0 : );
4227 0 : println!("{s}");
4228 0 :
4229 0 : Ok(test_img(&s))
4230 : }
4231 1038 : }
4232 : }
4233 : }
4234 :
4235 : #[cfg(test)]
4236 : mod tests {
4237 : use std::collections::{BTreeMap, BTreeSet};
4238 :
4239 : use super::*;
4240 : use crate::keyspace::KeySpaceAccum;
4241 : use crate::pgdatadir_mapping::AuxFilesDirectory;
4242 : use crate::repository::{Key, Value};
4243 : use crate::tenant::harness::*;
4244 : use crate::tenant::timeline::CompactFlags;
4245 : use crate::walrecord::NeonWalRecord;
4246 : use crate::DEFAULT_PG_VERSION;
4247 : use bytes::{Bytes, BytesMut};
4248 : use hex_literal::hex;
4249 : use itertools::Itertools;
4250 : use pageserver_api::key::{AUX_FILES_KEY, AUX_KEY_PREFIX, NON_INHERITED_RANGE};
4251 : use pageserver_api::keyspace::KeySpace;
4252 : use pageserver_api::models::{CompactionAlgorithm, CompactionAlgorithmSettings};
4253 : use rand::{thread_rng, Rng};
4254 : use storage_layer::PersistentLayerKey;
4255 : use tests::storage_layer::ValuesReconstructState;
4256 : use tests::timeline::{GetVectoredError, ShutdownMode};
4257 : use timeline::compaction::{KeyHistoryRetention, KeyLogAtLsn};
4258 : use timeline::{DeltaLayerTestDesc, GcInfo};
4259 : use utils::bin_ser::BeSer;
4260 : use utils::id::TenantId;
4261 :
4262 : static TEST_KEY: Lazy<Key> =
4263 54 : Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));
4264 :
4265 : #[tokio::test]
4266 6 : async fn test_basic() -> anyhow::Result<()> {
4267 24 : let (tenant, ctx) = TenantHarness::create("test_basic").await?.load().await;
4268 6 : let tline = tenant
4269 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
4270 12 : .await?;
4271 6 :
4272 6 : let mut writer = tline.writer().await;
4273 6 : writer
4274 6 : .put(
4275 6 : *TEST_KEY,
4276 6 : Lsn(0x10),
4277 6 : &Value::Image(test_img("foo at 0x10")),
4278 6 : &ctx,
4279 6 : )
4280 6 : .await?;
4281 6 : writer.finish_write(Lsn(0x10));
4282 6 : drop(writer);
4283 6 :
4284 6 : let mut writer = tline.writer().await;
4285 6 : writer
4286 6 : .put(
4287 6 : *TEST_KEY,
4288 6 : Lsn(0x20),
4289 6 : &Value::Image(test_img("foo at 0x20")),
4290 6 : &ctx,
4291 6 : )
4292 6 : .await?;
4293 6 : writer.finish_write(Lsn(0x20));
4294 6 : drop(writer);
4295 6 :
4296 6 : assert_eq!(
4297 6 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
4298 6 : test_img("foo at 0x10")
4299 6 : );
4300 6 : assert_eq!(
4301 6 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
4302 6 : test_img("foo at 0x10")
4303 6 : );
4304 6 : assert_eq!(
4305 6 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
4306 6 : test_img("foo at 0x20")
4307 6 : );
4308 6 :
4309 6 : Ok(())
4310 6 : }
4311 :
4312 : #[tokio::test]
4313 6 : async fn no_duplicate_timelines() -> anyhow::Result<()> {
4314 6 : let (tenant, ctx) = TenantHarness::create("no_duplicate_timelines")
4315 6 : .await?
4316 6 : .load()
4317 24 : .await;
4318 6 : let _ = tenant
4319 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4320 12 : .await?;
4321 6 :
4322 6 : match tenant
4323 6 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4324 6 : .await
4325 6 : {
4326 6 : Ok(_) => panic!("duplicate timeline creation should fail"),
4327 6 : Err(e) => assert_eq!(e.to_string(), "Already exists".to_string()),
4328 6 : }
4329 6 :
4330 6 : Ok(())
4331 6 : }
4332 :
4333 : /// Convenience function to create a page image with given string as the only content
4334 30 : pub fn test_value(s: &str) -> Value {
4335 30 : let mut buf = BytesMut::new();
4336 30 : buf.extend_from_slice(s.as_bytes());
4337 30 : Value::Image(buf.freeze())
4338 30 : }
4339 :
4340 : ///
4341 : /// Test branch creation
4342 : ///
4343 : #[tokio::test]
4344 6 : async fn test_branch() -> anyhow::Result<()> {
4345 6 : use std::str::from_utf8;
4346 6 :
4347 24 : let (tenant, ctx) = TenantHarness::create("test_branch").await?.load().await;
4348 6 : let tline = tenant
4349 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4350 12 : .await?;
4351 6 : let mut writer = tline.writer().await;
4352 6 :
4353 6 : #[allow(non_snake_case)]
4354 6 : let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
4355 6 : #[allow(non_snake_case)]
4356 6 : let TEST_KEY_B: Key = Key::from_hex("110000000033333333444444445500000002").unwrap();
4357 6 :
4358 6 : // Insert a value on the timeline
4359 6 : writer
4360 6 : .put(TEST_KEY_A, Lsn(0x20), &test_value("foo at 0x20"), &ctx)
4361 6 : .await?;
4362 6 : writer
4363 6 : .put(TEST_KEY_B, Lsn(0x20), &test_value("foobar at 0x20"), &ctx)
4364 6 : .await?;
4365 6 : writer.finish_write(Lsn(0x20));
4366 6 :
4367 6 : writer
4368 6 : .put(TEST_KEY_A, Lsn(0x30), &test_value("foo at 0x30"), &ctx)
4369 6 : .await?;
4370 6 : writer.finish_write(Lsn(0x30));
4371 6 : writer
4372 6 : .put(TEST_KEY_A, Lsn(0x40), &test_value("foo at 0x40"), &ctx)
4373 6 : .await?;
4374 6 : writer.finish_write(Lsn(0x40));
4375 6 :
4376 6 : //assert_current_logical_size(&tline, Lsn(0x40));
4377 6 :
4378 6 : // Branch the history, modify relation differently on the new timeline
4379 6 : tenant
4380 6 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x30)), &ctx)
4381 6 : .await?;
4382 6 : let newtline = tenant
4383 6 : .get_timeline(NEW_TIMELINE_ID, true)
4384 6 : .expect("Should have a local timeline");
4385 6 : let mut new_writer = newtline.writer().await;
4386 6 : new_writer
4387 6 : .put(TEST_KEY_A, Lsn(0x40), &test_value("bar at 0x40"), &ctx)
4388 6 : .await?;
4389 6 : new_writer.finish_write(Lsn(0x40));
4390 6 :
4391 6 : // Check page contents on both branches
4392 6 : assert_eq!(
4393 6 : from_utf8(&tline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
4394 6 : "foo at 0x40"
4395 6 : );
4396 6 : assert_eq!(
4397 6 : from_utf8(&newtline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
4398 6 : "bar at 0x40"
4399 6 : );
4400 6 : assert_eq!(
4401 6 : from_utf8(&newtline.get(TEST_KEY_B, Lsn(0x40), &ctx).await?)?,
4402 6 : "foobar at 0x20"
4403 6 : );
4404 6 :
4405 6 : //assert_current_logical_size(&tline, Lsn(0x40));
4406 6 :
4407 6 : Ok(())
4408 6 : }
4409 :
4410 60 : async fn make_some_layers(
4411 60 : tline: &Timeline,
4412 60 : start_lsn: Lsn,
4413 60 : ctx: &RequestContext,
4414 60 : ) -> anyhow::Result<()> {
4415 60 : let mut lsn = start_lsn;
4416 : {
4417 60 : let mut writer = tline.writer().await;
4418 : // Create a relation on the timeline
4419 60 : writer
4420 60 : .put(
4421 60 : *TEST_KEY,
4422 60 : lsn,
4423 60 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4424 60 : ctx,
4425 60 : )
4426 30 : .await?;
4427 60 : writer.finish_write(lsn);
4428 60 : lsn += 0x10;
4429 60 : writer
4430 60 : .put(
4431 60 : *TEST_KEY,
4432 60 : lsn,
4433 60 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4434 60 : ctx,
4435 60 : )
4436 0 : .await?;
4437 60 : writer.finish_write(lsn);
4438 60 : lsn += 0x10;
4439 60 : }
4440 61 : tline.freeze_and_flush().await?;
4441 : {
4442 60 : let mut writer = tline.writer().await;
4443 60 : writer
4444 60 : .put(
4445 60 : *TEST_KEY,
4446 60 : lsn,
4447 60 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4448 60 : ctx,
4449 60 : )
4450 30 : .await?;
4451 60 : writer.finish_write(lsn);
4452 60 : lsn += 0x10;
4453 60 : writer
4454 60 : .put(
4455 60 : *TEST_KEY,
4456 60 : lsn,
4457 60 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4458 60 : ctx,
4459 60 : )
4460 0 : .await?;
4461 60 : writer.finish_write(lsn);
4462 60 : }
4463 60 : tline.freeze_and_flush().await.map_err(|e| e.into())
4464 60 : }
4465 :
4466 : #[tokio::test]
4467 6 : async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
4468 6 : let (tenant, ctx) =
4469 6 : TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
4470 6 : .await?
4471 6 : .load()
4472 24 : .await;
4473 6 : let tline = tenant
4474 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4475 12 : .await?;
4476 18 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4477 6 :
4478 6 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
4479 6 : // FIXME: this doesn't actually remove any layer currently, given how the flushing
4480 6 : // and compaction works. But it does set the 'cutoff' point so that the cross check
4481 6 : // below should fail.
4482 6 : tenant
4483 6 : .gc_iteration(
4484 6 : Some(TIMELINE_ID),
4485 6 : 0x10,
4486 6 : Duration::ZERO,
4487 6 : &CancellationToken::new(),
4488 6 : &ctx,
4489 6 : )
4490 6 : .await?;
4491 6 :
4492 6 : // try to branch at lsn 25, should fail because we already garbage collected the data
4493 6 : match tenant
4494 6 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
4495 6 : .await
4496 6 : {
4497 6 : Ok(_) => panic!("branching should have failed"),
4498 6 : Err(err) => {
4499 6 : let CreateTimelineError::AncestorLsn(err) = err else {
4500 6 : panic!("wrong error type")
4501 6 : };
4502 6 : assert!(err.to_string().contains("invalid branch start lsn"));
4503 6 : assert!(err
4504 6 : .source()
4505 6 : .unwrap()
4506 6 : .to_string()
4507 6 : .contains("we might've already garbage collected needed data"))
4508 6 : }
4509 6 : }
4510 6 :
4511 6 : Ok(())
4512 6 : }
4513 :
4514 : #[tokio::test]
4515 6 : async fn test_prohibit_branch_creation_on_pre_initdb_lsn() -> anyhow::Result<()> {
4516 6 : let (tenant, ctx) =
4517 6 : TenantHarness::create("test_prohibit_branch_creation_on_pre_initdb_lsn")
4518 6 : .await?
4519 6 : .load()
4520 24 : .await;
4521 6 :
4522 6 : let tline = tenant
4523 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x50), DEFAULT_PG_VERSION, &ctx)
4524 12 : .await?;
4525 6 : // try to branch at lsn 0x25, should fail because initdb lsn is 0x50
4526 6 : match tenant
4527 6 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
4528 6 : .await
4529 6 : {
4530 6 : Ok(_) => panic!("branching should have failed"),
4531 6 : Err(err) => {
4532 6 : let CreateTimelineError::AncestorLsn(err) = err else {
4533 6 : panic!("wrong error type");
4534 6 : };
4535 6 : assert!(&err.to_string().contains("invalid branch start lsn"));
4536 6 : assert!(&err
4537 6 : .source()
4538 6 : .unwrap()
4539 6 : .to_string()
4540 6 : .contains("is earlier than latest GC cutoff"));
4541 6 : }
4542 6 : }
4543 6 :
4544 6 : Ok(())
4545 6 : }
4546 :
4547 : /*
4548 : // FIXME: This currently fails to error out. Calling GC doesn't currently
4549 : // remove the old value, we'd need to work a little harder
4550 : #[tokio::test]
4551 : async fn test_prohibit_get_for_garbage_collected_data() -> anyhow::Result<()> {
4552 : let repo =
4553 : RepoHarness::create("test_prohibit_get_for_garbage_collected_data")?
4554 : .load();
4555 :
4556 : let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION)?;
4557 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4558 :
4559 : repo.gc_iteration(Some(TIMELINE_ID), 0x10, Duration::ZERO)?;
4560 : let latest_gc_cutoff_lsn = tline.get_latest_gc_cutoff_lsn();
4561 : assert!(*latest_gc_cutoff_lsn > Lsn(0x25));
4562 : match tline.get(*TEST_KEY, Lsn(0x25)) {
4563 : Ok(_) => panic!("request for page should have failed"),
4564 : Err(err) => assert!(err.to_string().contains("not found at")),
4565 : }
4566 : Ok(())
4567 : }
4568 : */
4569 :
4570 : #[tokio::test]
4571 6 : async fn test_get_branchpoints_from_an_inactive_timeline() -> anyhow::Result<()> {
4572 6 : let (tenant, ctx) =
4573 6 : TenantHarness::create("test_get_branchpoints_from_an_inactive_timeline")
4574 6 : .await?
4575 6 : .load()
4576 24 : .await;
4577 6 : let tline = tenant
4578 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4579 12 : .await?;
4580 19 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4581 6 :
4582 6 : tenant
4583 6 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
4584 6 : .await?;
4585 6 : let newtline = tenant
4586 6 : .get_timeline(NEW_TIMELINE_ID, true)
4587 6 : .expect("Should have a local timeline");
4588 6 :
4589 18 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
4590 6 :
4591 6 : tline.set_broken("test".to_owned());
4592 6 :
4593 6 : tenant
4594 6 : .gc_iteration(
4595 6 : Some(TIMELINE_ID),
4596 6 : 0x10,
4597 6 : Duration::ZERO,
4598 6 : &CancellationToken::new(),
4599 6 : &ctx,
4600 6 : )
4601 6 : .await?;
4602 6 :
4603 6 : // The branchpoints should contain all timelines, even ones marked
4604 6 : // as Broken.
4605 6 : {
4606 6 : let branchpoints = &tline.gc_info.read().unwrap().retain_lsns;
4607 6 : assert_eq!(branchpoints.len(), 1);
4608 6 : assert_eq!(branchpoints[0], (Lsn(0x40), NEW_TIMELINE_ID));
4609 6 : }
4610 6 :
4611 6 : // You can read the key from the child branch even though the parent is
4612 6 : // Broken, as long as you don't need to access data from the parent.
4613 6 : assert_eq!(
4614 12 : newtline.get(*TEST_KEY, Lsn(0x70), &ctx).await?,
4615 6 : test_img(&format!("foo at {}", Lsn(0x70)))
4616 6 : );
4617 6 :
4618 6 : // This needs to traverse to the parent, and fails.
4619 6 : let err = newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await.unwrap_err();
4620 6 : assert!(
4621 6 : err.to_string().starts_with(&format!(
4622 6 : "bad state on timeline {}: Broken",
4623 6 : tline.timeline_id
4624 6 : )),
4625 6 : "{err}"
4626 6 : );
4627 6 :
4628 6 : Ok(())
4629 6 : }
4630 :
4631 : #[tokio::test]
4632 6 : async fn test_retain_data_in_parent_which_is_needed_for_child() -> anyhow::Result<()> {
4633 6 : let (tenant, ctx) =
4634 6 : TenantHarness::create("test_retain_data_in_parent_which_is_needed_for_child")
4635 6 : .await?
4636 6 : .load()
4637 24 : .await;
4638 6 : let tline = tenant
4639 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4640 12 : .await?;
4641 18 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4642 6 :
4643 6 : tenant
4644 6 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
4645 6 : .await?;
4646 6 : let newtline = tenant
4647 6 : .get_timeline(NEW_TIMELINE_ID, true)
4648 6 : .expect("Should have a local timeline");
4649 6 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
4650 6 : tenant
4651 6 : .gc_iteration(
4652 6 : Some(TIMELINE_ID),
4653 6 : 0x10,
4654 6 : Duration::ZERO,
4655 6 : &CancellationToken::new(),
4656 6 : &ctx,
4657 6 : )
4658 6 : .await?;
4659 12 : assert!(newtline.get(*TEST_KEY, Lsn(0x25), &ctx).await.is_ok());
4660 6 :
4661 6 : Ok(())
4662 6 : }
4663 : #[tokio::test]
4664 6 : async fn test_parent_keeps_data_forever_after_branching() -> anyhow::Result<()> {
4665 6 : let (tenant, ctx) = TenantHarness::create("test_parent_keeps_data_forever_after_branching")
4666 6 : .await?
4667 6 : .load()
4668 24 : .await;
4669 6 : let tline = tenant
4670 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4671 12 : .await?;
4672 18 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4673 6 :
4674 6 : tenant
4675 6 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
4676 6 : .await?;
4677 6 : let newtline = tenant
4678 6 : .get_timeline(NEW_TIMELINE_ID, true)
4679 6 : .expect("Should have a local timeline");
4680 6 :
4681 18 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
4682 6 :
4683 6 : // run gc on parent
4684 6 : tenant
4685 6 : .gc_iteration(
4686 6 : Some(TIMELINE_ID),
4687 6 : 0x10,
4688 6 : Duration::ZERO,
4689 6 : &CancellationToken::new(),
4690 6 : &ctx,
4691 6 : )
4692 6 : .await?;
4693 6 :
4694 6 : // Check that the data is still accessible on the branch.
4695 6 : assert_eq!(
4696 21 : newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
4697 6 : test_img(&format!("foo at {}", Lsn(0x40)))
4698 6 : );
4699 6 :
4700 6 : Ok(())
4701 6 : }
4702 :
4703 : #[tokio::test]
4704 6 : async fn timeline_load() -> anyhow::Result<()> {
4705 6 : const TEST_NAME: &str = "timeline_load";
4706 6 : let harness = TenantHarness::create(TEST_NAME).await?;
4707 6 : {
4708 24 : let (tenant, ctx) = harness.load().await;
4709 6 : let tline = tenant
4710 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x7000), DEFAULT_PG_VERSION, &ctx)
4711 12 : .await?;
4712 18 : make_some_layers(tline.as_ref(), Lsn(0x8000), &ctx).await?;
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 19 : let (tenant, _ctx) = harness.load().await;
4723 6 : tenant
4724 6 : .get_timeline(TIMELINE_ID, true)
4725 6 : .expect("cannot load timeline");
4726 6 :
4727 6 : Ok(())
4728 6 : }
4729 :
4730 : #[tokio::test]
4731 6 : async fn timeline_load_with_ancestor() -> anyhow::Result<()> {
4732 6 : const TEST_NAME: &str = "timeline_load_with_ancestor";
4733 6 : let harness = TenantHarness::create(TEST_NAME).await?;
4734 6 : // create two timelines
4735 6 : {
4736 24 : let (tenant, ctx) = harness.load().await;
4737 6 : let tline = tenant
4738 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4739 12 : .await?;
4740 6 :
4741 18 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4742 6 :
4743 6 : let child_tline = tenant
4744 6 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
4745 6 : .await?;
4746 6 : child_tline.set_state(TimelineState::Active);
4747 6 :
4748 6 : let newtline = tenant
4749 6 : .get_timeline(NEW_TIMELINE_ID, true)
4750 6 : .expect("Should have a local timeline");
4751 6 :
4752 18 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
4753 6 :
4754 6 : // so that all uploads finish & we can call harness.load() below again
4755 6 : tenant
4756 6 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
4757 6 : .instrument(harness.span())
4758 6 : .await
4759 6 : .ok()
4760 6 : .unwrap();
4761 6 : }
4762 6 :
4763 6 : // check that both of them are initially unloaded
4764 40 : let (tenant, _ctx) = harness.load().await;
4765 6 :
4766 6 : // check that both, child and ancestor are loaded
4767 6 : let _child_tline = tenant
4768 6 : .get_timeline(NEW_TIMELINE_ID, true)
4769 6 : .expect("cannot get child timeline loaded");
4770 6 :
4771 6 : let _ancestor_tline = tenant
4772 6 : .get_timeline(TIMELINE_ID, true)
4773 6 : .expect("cannot get ancestor timeline loaded");
4774 6 :
4775 6 : Ok(())
4776 6 : }
4777 :
4778 : #[tokio::test]
4779 6 : async fn delta_layer_dumping() -> anyhow::Result<()> {
4780 6 : use storage_layer::AsLayerDesc;
4781 6 : let (tenant, ctx) = TenantHarness::create("test_layer_dumping")
4782 6 : .await?
4783 6 : .load()
4784 24 : .await;
4785 6 : let tline = tenant
4786 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4787 12 : .await?;
4788 18 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4789 6 :
4790 6 : let layer_map = tline.layers.read().await;
4791 6 : let level0_deltas = layer_map
4792 6 : .layer_map()?
4793 6 : .level0_deltas()
4794 6 : .iter()
4795 12 : .map(|desc| layer_map.get_from_desc(desc))
4796 6 : .collect::<Vec<_>>();
4797 6 :
4798 6 : assert!(!level0_deltas.is_empty());
4799 6 :
4800 18 : for delta in level0_deltas {
4801 6 : // Ensure we are dumping a delta layer here
4802 12 : assert!(delta.layer_desc().is_delta);
4803 24 : delta.dump(true, &ctx).await.unwrap();
4804 6 : }
4805 6 :
4806 6 : Ok(())
4807 6 : }
4808 :
4809 : #[tokio::test]
4810 6 : async fn test_images() -> anyhow::Result<()> {
4811 24 : let (tenant, ctx) = TenantHarness::create("test_images").await?.load().await;
4812 6 : let tline = tenant
4813 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
4814 12 : .await?;
4815 6 :
4816 6 : let mut writer = tline.writer().await;
4817 6 : writer
4818 6 : .put(
4819 6 : *TEST_KEY,
4820 6 : Lsn(0x10),
4821 6 : &Value::Image(test_img("foo at 0x10")),
4822 6 : &ctx,
4823 6 : )
4824 6 : .await?;
4825 6 : writer.finish_write(Lsn(0x10));
4826 6 : drop(writer);
4827 6 :
4828 6 : tline.freeze_and_flush().await?;
4829 6 : tline
4830 6 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
4831 6 : .await?;
4832 6 :
4833 6 : let mut writer = tline.writer().await;
4834 6 : writer
4835 6 : .put(
4836 6 : *TEST_KEY,
4837 6 : Lsn(0x20),
4838 6 : &Value::Image(test_img("foo at 0x20")),
4839 6 : &ctx,
4840 6 : )
4841 6 : .await?;
4842 6 : writer.finish_write(Lsn(0x20));
4843 6 : drop(writer);
4844 6 :
4845 6 : tline.freeze_and_flush().await?;
4846 6 : tline
4847 6 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
4848 6 : .await?;
4849 6 :
4850 6 : let mut writer = tline.writer().await;
4851 6 : writer
4852 6 : .put(
4853 6 : *TEST_KEY,
4854 6 : Lsn(0x30),
4855 6 : &Value::Image(test_img("foo at 0x30")),
4856 6 : &ctx,
4857 6 : )
4858 6 : .await?;
4859 6 : writer.finish_write(Lsn(0x30));
4860 6 : drop(writer);
4861 6 :
4862 6 : tline.freeze_and_flush().await?;
4863 6 : tline
4864 6 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
4865 6 : .await?;
4866 6 :
4867 6 : let mut writer = tline.writer().await;
4868 6 : writer
4869 6 : .put(
4870 6 : *TEST_KEY,
4871 6 : Lsn(0x40),
4872 6 : &Value::Image(test_img("foo at 0x40")),
4873 6 : &ctx,
4874 6 : )
4875 6 : .await?;
4876 6 : writer.finish_write(Lsn(0x40));
4877 6 : drop(writer);
4878 6 :
4879 6 : tline.freeze_and_flush().await?;
4880 6 : tline
4881 6 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
4882 6 : .await?;
4883 6 :
4884 6 : assert_eq!(
4885 12 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
4886 6 : test_img("foo at 0x10")
4887 6 : );
4888 6 : assert_eq!(
4889 12 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
4890 6 : test_img("foo at 0x10")
4891 6 : );
4892 6 : assert_eq!(
4893 6 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
4894 6 : test_img("foo at 0x20")
4895 6 : );
4896 6 : assert_eq!(
4897 12 : tline.get(*TEST_KEY, Lsn(0x30), &ctx).await?,
4898 6 : test_img("foo at 0x30")
4899 6 : );
4900 6 : assert_eq!(
4901 12 : tline.get(*TEST_KEY, Lsn(0x40), &ctx).await?,
4902 6 : test_img("foo at 0x40")
4903 6 : );
4904 6 :
4905 6 : Ok(())
4906 6 : }
4907 :
4908 12 : async fn bulk_insert_compact_gc(
4909 12 : tenant: &Tenant,
4910 12 : timeline: &Arc<Timeline>,
4911 12 : ctx: &RequestContext,
4912 12 : lsn: Lsn,
4913 12 : repeat: usize,
4914 12 : key_count: usize,
4915 12 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
4916 12 : let compact = true;
4917 122753 : bulk_insert_maybe_compact_gc(tenant, timeline, ctx, lsn, repeat, key_count, compact).await
4918 12 : }
4919 :
4920 24 : async fn bulk_insert_maybe_compact_gc(
4921 24 : tenant: &Tenant,
4922 24 : timeline: &Arc<Timeline>,
4923 24 : ctx: &RequestContext,
4924 24 : mut lsn: Lsn,
4925 24 : repeat: usize,
4926 24 : key_count: usize,
4927 24 : compact: bool,
4928 24 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
4929 24 : let mut inserted: HashMap<Key, BTreeSet<Lsn>> = Default::default();
4930 24 :
4931 24 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
4932 24 : let mut blknum = 0;
4933 24 :
4934 24 : // Enforce that key range is monotonously increasing
4935 24 : let mut keyspace = KeySpaceAccum::new();
4936 24 :
4937 24 : let cancel = CancellationToken::new();
4938 24 :
4939 24 : for _ in 0..repeat {
4940 1200 : for _ in 0..key_count {
4941 12000000 : test_key.field6 = blknum;
4942 12000000 : let mut writer = timeline.writer().await;
4943 12000000 : writer
4944 12000000 : .put(
4945 12000000 : test_key,
4946 12000000 : lsn,
4947 12000000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
4948 12000000 : ctx,
4949 12000000 : )
4950 7788 : .await?;
4951 12000000 : inserted.entry(test_key).or_default().insert(lsn);
4952 12000000 : writer.finish_write(lsn);
4953 12000000 : drop(writer);
4954 12000000 :
4955 12000000 : keyspace.add_key(test_key);
4956 12000000 :
4957 12000000 : lsn = Lsn(lsn.0 + 0x10);
4958 12000000 : blknum += 1;
4959 : }
4960 :
4961 1200 : timeline.freeze_and_flush().await?;
4962 1200 : if compact {
4963 : // this requires timeline to be &Arc<Timeline>
4964 25854 : timeline.compact(&cancel, EnumSet::empty(), ctx).await?;
4965 600 : }
4966 :
4967 : // this doesn't really need to use the timeline_id target, but it is closer to what it
4968 : // originally was.
4969 1200 : let res = tenant
4970 1200 : .gc_iteration(Some(timeline.timeline_id), 0, Duration::ZERO, &cancel, ctx)
4971 1199 : .await?;
4972 :
4973 1200 : assert_eq!(res.layers_removed, 0, "this never removes anything");
4974 : }
4975 :
4976 24 : Ok(inserted)
4977 24 : }
4978 :
4979 : //
4980 : // Insert 1000 key-value pairs with increasing keys, flush, compact, GC.
4981 : // Repeat 50 times.
4982 : //
4983 : #[tokio::test]
4984 6 : async fn test_bulk_insert() -> anyhow::Result<()> {
4985 6 : let harness = TenantHarness::create("test_bulk_insert").await?;
4986 24 : let (tenant, ctx) = harness.load().await;
4987 6 : let tline = tenant
4988 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
4989 12 : .await?;
4990 6 :
4991 6 : let lsn = Lsn(0x10);
4992 61376 : bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
4993 6 :
4994 6 : Ok(())
4995 6 : }
4996 :
4997 : // Test the vectored get real implementation against a simple sequential implementation.
4998 : //
4999 : // The test generates a keyspace by repeatedly flushing the in-memory layer and compacting.
5000 : // Projected to 2D the key space looks like below. Lsn grows upwards on the Y axis and keys
5001 : // grow to the right on the X axis.
5002 : // [Delta]
5003 : // [Delta]
5004 : // [Delta]
5005 : // [Delta]
5006 : // ------------ Image ---------------
5007 : //
5008 : // After layer generation we pick the ranges to query as follows:
5009 : // 1. The beginning of each delta layer
5010 : // 2. At the seam between two adjacent delta layers
5011 : //
5012 : // There's one major downside to this test: delta layers only contains images,
5013 : // so the search can stop at the first delta layer and doesn't traverse any deeper.
5014 : #[tokio::test]
5015 6 : async fn test_get_vectored() -> anyhow::Result<()> {
5016 6 : let harness = TenantHarness::create("test_get_vectored").await?;
5017 24 : let (tenant, ctx) = harness.load().await;
5018 6 : let tline = tenant
5019 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5020 12 : .await?;
5021 6 :
5022 6 : let lsn = Lsn(0x10);
5023 61377 : let inserted = bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
5024 6 :
5025 6 : let guard = tline.layers.read().await;
5026 6 : let lm = guard.layer_map()?;
5027 6 :
5028 6 : lm.dump(true, &ctx).await?;
5029 6 :
5030 6 : let mut reads = Vec::new();
5031 6 : let mut prev = None;
5032 36 : lm.iter_historic_layers().for_each(|desc| {
5033 36 : if !desc.is_delta() {
5034 6 : prev = Some(desc.clone());
5035 6 : return;
5036 30 : }
5037 30 :
5038 30 : let start = desc.key_range.start;
5039 30 : let end = desc
5040 30 : .key_range
5041 30 : .start
5042 30 : .add(Timeline::MAX_GET_VECTORED_KEYS.try_into().unwrap());
5043 30 : reads.push(KeySpace {
5044 30 : ranges: vec![start..end],
5045 30 : });
5046 6 :
5047 30 : if let Some(prev) = &prev {
5048 30 : if !prev.is_delta() {
5049 30 : return;
5050 6 : }
5051 0 :
5052 0 : let first_range = Key {
5053 0 : field6: prev.key_range.end.field6 - 4,
5054 0 : ..prev.key_range.end
5055 0 : }..prev.key_range.end;
5056 0 :
5057 0 : let second_range = desc.key_range.start..Key {
5058 0 : field6: desc.key_range.start.field6 + 4,
5059 0 : ..desc.key_range.start
5060 0 : };
5061 0 :
5062 0 : reads.push(KeySpace {
5063 0 : ranges: vec![first_range, second_range],
5064 0 : });
5065 6 : };
5066 6 :
5067 6 : prev = Some(desc.clone());
5068 36 : });
5069 6 :
5070 6 : drop(guard);
5071 6 :
5072 6 : // Pick a big LSN such that we query over all the changes.
5073 6 : let reads_lsn = Lsn(u64::MAX - 1);
5074 6 :
5075 36 : for read in reads {
5076 30 : info!("Doing vectored read on {:?}", read);
5077 6 :
5078 30 : let vectored_res = tline
5079 30 : .get_vectored_impl(
5080 30 : read.clone(),
5081 30 : reads_lsn,
5082 30 : &mut ValuesReconstructState::new(),
5083 30 : &ctx,
5084 30 : )
5085 75 : .await;
5086 6 :
5087 30 : let mut expected_lsns: HashMap<Key, Lsn> = Default::default();
5088 30 : let mut expect_missing = false;
5089 30 : let mut key = read.start().unwrap();
5090 990 : while key != read.end().unwrap() {
5091 960 : if let Some(lsns) = inserted.get(&key) {
5092 960 : let expected_lsn = lsns.iter().rfind(|lsn| **lsn <= reads_lsn);
5093 960 : match expected_lsn {
5094 960 : Some(lsn) => {
5095 960 : expected_lsns.insert(key, *lsn);
5096 960 : }
5097 6 : None => {
5098 6 : expect_missing = true;
5099 0 : break;
5100 6 : }
5101 6 : }
5102 6 : } else {
5103 6 : expect_missing = true;
5104 0 : break;
5105 6 : }
5106 6 :
5107 960 : key = key.next();
5108 6 : }
5109 6 :
5110 30 : if expect_missing {
5111 6 : assert!(matches!(vectored_res, Err(GetVectoredError::MissingKey(_))));
5112 6 : } else {
5113 960 : for (key, image) in vectored_res? {
5114 960 : let expected_lsn = expected_lsns.get(&key).expect("determined above");
5115 960 : let expected_image = test_img(&format!("{} at {}", key.field6, expected_lsn));
5116 960 : assert_eq!(image?, expected_image);
5117 6 : }
5118 6 : }
5119 6 : }
5120 6 :
5121 6 : Ok(())
5122 6 : }
5123 :
5124 : #[tokio::test]
5125 6 : async fn test_get_vectored_aux_files() -> anyhow::Result<()> {
5126 6 : let harness = TenantHarness::create("test_get_vectored_aux_files").await?;
5127 6 :
5128 24 : let (tenant, ctx) = harness.load().await;
5129 6 : let tline = tenant
5130 6 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
5131 6 : .await?;
5132 6 : let tline = tline.raw_timeline().unwrap();
5133 6 :
5134 6 : let mut modification = tline.begin_modification(Lsn(0x1000));
5135 6 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
5136 6 : modification.set_lsn(Lsn(0x1008))?;
5137 6 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
5138 6 : modification.commit(&ctx).await?;
5139 6 :
5140 6 : let child_timeline_id = TimelineId::generate();
5141 6 : tenant
5142 6 : .branch_timeline_test(
5143 6 : tline,
5144 6 : child_timeline_id,
5145 6 : Some(tline.get_last_record_lsn()),
5146 6 : &ctx,
5147 6 : )
5148 6 : .await?;
5149 6 :
5150 6 : let child_timeline = tenant
5151 6 : .get_timeline(child_timeline_id, true)
5152 6 : .expect("Should have the branched timeline");
5153 6 :
5154 6 : let aux_keyspace = KeySpace {
5155 6 : ranges: vec![NON_INHERITED_RANGE],
5156 6 : };
5157 6 : let read_lsn = child_timeline.get_last_record_lsn();
5158 6 :
5159 6 : let vectored_res = child_timeline
5160 6 : .get_vectored_impl(
5161 6 : aux_keyspace.clone(),
5162 6 : read_lsn,
5163 6 : &mut ValuesReconstructState::new(),
5164 6 : &ctx,
5165 6 : )
5166 6 : .await;
5167 6 :
5168 6 : let images = vectored_res?;
5169 6 : assert!(images.is_empty());
5170 6 : Ok(())
5171 6 : }
5172 :
5173 : // Test that vectored get handles layer gaps correctly
5174 : // by advancing into the next ancestor timeline if required.
5175 : //
5176 : // The test generates timelines that look like the diagram below.
5177 : // We leave a gap in one of the L1 layers at `gap_at_key` (`/` in the diagram).
5178 : // The reconstruct data for that key lies in the ancestor timeline (`X` in the diagram).
5179 : //
5180 : // ```
5181 : //-------------------------------+
5182 : // ... |
5183 : // [ L1 ] |
5184 : // [ / L1 ] | Child Timeline
5185 : // ... |
5186 : // ------------------------------+
5187 : // [ X L1 ] | Parent Timeline
5188 : // ------------------------------+
5189 : // ```
5190 : #[tokio::test]
5191 6 : async fn test_get_vectored_key_gap() -> anyhow::Result<()> {
5192 6 : let tenant_conf = TenantConf {
5193 6 : // Make compaction deterministic
5194 6 : gc_period: Duration::ZERO,
5195 6 : compaction_period: Duration::ZERO,
5196 6 : // Encourage creation of L1 layers
5197 6 : checkpoint_distance: 16 * 1024,
5198 6 : compaction_target_size: 8 * 1024,
5199 6 : ..TenantConf::default()
5200 6 : };
5201 6 :
5202 6 : let harness = TenantHarness::create_custom(
5203 6 : "test_get_vectored_key_gap",
5204 6 : tenant_conf,
5205 6 : TenantId::generate(),
5206 6 : ShardIdentity::unsharded(),
5207 6 : Generation::new(0xdeadbeef),
5208 6 : )
5209 6 : .await?;
5210 24 : let (tenant, ctx) = harness.load().await;
5211 6 :
5212 6 : let mut current_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5213 6 : let gap_at_key = current_key.add(100);
5214 6 : let mut current_lsn = Lsn(0x10);
5215 6 :
5216 6 : const KEY_COUNT: usize = 10_000;
5217 6 :
5218 6 : let timeline_id = TimelineId::generate();
5219 6 : let current_timeline = tenant
5220 6 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
5221 12 : .await?;
5222 6 :
5223 6 : current_lsn += 0x100;
5224 6 :
5225 6 : let mut writer = current_timeline.writer().await;
5226 6 : writer
5227 6 : .put(
5228 6 : gap_at_key,
5229 6 : current_lsn,
5230 6 : &Value::Image(test_img(&format!("{} at {}", gap_at_key, current_lsn))),
5231 6 : &ctx,
5232 6 : )
5233 6 : .await?;
5234 6 : writer.finish_write(current_lsn);
5235 6 : drop(writer);
5236 6 :
5237 6 : let mut latest_lsns = HashMap::new();
5238 6 : latest_lsns.insert(gap_at_key, current_lsn);
5239 6 :
5240 6 : current_timeline.freeze_and_flush().await?;
5241 6 :
5242 6 : let child_timeline_id = TimelineId::generate();
5243 6 :
5244 6 : tenant
5245 6 : .branch_timeline_test(
5246 6 : ¤t_timeline,
5247 6 : child_timeline_id,
5248 6 : Some(current_lsn),
5249 6 : &ctx,
5250 6 : )
5251 6 : .await?;
5252 6 : let child_timeline = tenant
5253 6 : .get_timeline(child_timeline_id, true)
5254 6 : .expect("Should have the branched timeline");
5255 6 :
5256 60006 : for i in 0..KEY_COUNT {
5257 60000 : if current_key == gap_at_key {
5258 6 : current_key = current_key.next();
5259 6 : continue;
5260 59994 : }
5261 59994 :
5262 59994 : current_lsn += 0x10;
5263 6 :
5264 59994 : let mut writer = child_timeline.writer().await;
5265 59994 : writer
5266 59994 : .put(
5267 59994 : current_key,
5268 59994 : current_lsn,
5269 59994 : &Value::Image(test_img(&format!("{} at {}", current_key, current_lsn))),
5270 59994 : &ctx,
5271 59994 : )
5272 229 : .await?;
5273 59994 : writer.finish_write(current_lsn);
5274 59994 : drop(writer);
5275 59994 :
5276 59994 : latest_lsns.insert(current_key, current_lsn);
5277 59994 : current_key = current_key.next();
5278 59994 :
5279 59994 : // Flush every now and then to encourage layer file creation.
5280 59994 : if i % 500 == 0 {
5281 140 : child_timeline.freeze_and_flush().await?;
5282 59874 : }
5283 6 : }
5284 6 :
5285 7 : child_timeline.freeze_and_flush().await?;
5286 6 : let mut flags = EnumSet::new();
5287 6 : flags.insert(CompactFlags::ForceRepartition);
5288 6 : child_timeline
5289 6 : .compact(&CancellationToken::new(), flags, &ctx)
5290 5598 : .await?;
5291 6 :
5292 6 : let key_near_end = {
5293 6 : let mut tmp = current_key;
5294 6 : tmp.field6 -= 10;
5295 6 : tmp
5296 6 : };
5297 6 :
5298 6 : let key_near_gap = {
5299 6 : let mut tmp = gap_at_key;
5300 6 : tmp.field6 -= 10;
5301 6 : tmp
5302 6 : };
5303 6 :
5304 6 : let read = KeySpace {
5305 6 : ranges: vec![key_near_gap..gap_at_key.next(), key_near_end..current_key],
5306 6 : };
5307 6 : let results = child_timeline
5308 6 : .get_vectored_impl(
5309 6 : read.clone(),
5310 6 : current_lsn,
5311 6 : &mut ValuesReconstructState::new(),
5312 6 : &ctx,
5313 6 : )
5314 45 : .await?;
5315 6 :
5316 132 : for (key, img_res) in results {
5317 126 : let expected = test_img(&format!("{} at {}", key, latest_lsns[&key]));
5318 126 : assert_eq!(img_res?, expected);
5319 6 : }
5320 6 :
5321 6 : Ok(())
5322 6 : }
5323 :
5324 : // Test that vectored get descends into ancestor timelines correctly and
5325 : // does not return an image that's newer than requested.
5326 : //
5327 : // The diagram below ilustrates an interesting case. We have a parent timeline
5328 : // (top of the Lsn range) and a child timeline. The request key cannot be reconstructed
5329 : // from the child timeline, so the parent timeline must be visited. When advacing into
5330 : // the child timeline, the read path needs to remember what the requested Lsn was in
5331 : // order to avoid returning an image that's too new. The test below constructs such
5332 : // a timeline setup and does a few queries around the Lsn of each page image.
5333 : // ```
5334 : // LSN
5335 : // ^
5336 : // |
5337 : // |
5338 : // 500 | --------------------------------------> branch point
5339 : // 400 | X
5340 : // 300 | X
5341 : // 200 | --------------------------------------> requested lsn
5342 : // 100 | X
5343 : // |---------------------------------------> Key
5344 : // |
5345 : // ------> requested key
5346 : //
5347 : // Legend:
5348 : // * X - page images
5349 : // ```
5350 : #[tokio::test]
5351 6 : async fn test_get_vectored_ancestor_descent() -> anyhow::Result<()> {
5352 6 : let harness = TenantHarness::create("test_get_vectored_on_lsn_axis").await?;
5353 24 : let (tenant, ctx) = harness.load().await;
5354 6 :
5355 6 : let start_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5356 6 : let end_key = start_key.add(1000);
5357 6 : let child_gap_at_key = start_key.add(500);
5358 6 : let mut parent_gap_lsns: BTreeMap<Lsn, String> = BTreeMap::new();
5359 6 :
5360 6 : let mut current_lsn = Lsn(0x10);
5361 6 :
5362 6 : let timeline_id = TimelineId::generate();
5363 6 : let parent_timeline = tenant
5364 6 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
5365 12 : .await?;
5366 6 :
5367 6 : current_lsn += 0x100;
5368 6 :
5369 24 : for _ in 0..3 {
5370 18 : let mut key = start_key;
5371 18018 : while key < end_key {
5372 18000 : current_lsn += 0x10;
5373 18000 :
5374 18000 : let image_value = format!("{} at {}", child_gap_at_key, current_lsn);
5375 6 :
5376 18000 : let mut writer = parent_timeline.writer().await;
5377 18000 : writer
5378 18000 : .put(
5379 18000 : key,
5380 18000 : current_lsn,
5381 18000 : &Value::Image(test_img(&image_value)),
5382 18000 : &ctx,
5383 18000 : )
5384 18 : .await?;
5385 18000 : writer.finish_write(current_lsn);
5386 18000 :
5387 18000 : if key == child_gap_at_key {
5388 18 : parent_gap_lsns.insert(current_lsn, image_value);
5389 17982 : }
5390 6 :
5391 18000 : key = key.next();
5392 6 : }
5393 6 :
5394 18 : parent_timeline.freeze_and_flush().await?;
5395 6 : }
5396 6 :
5397 6 : let child_timeline_id = TimelineId::generate();
5398 6 :
5399 6 : let child_timeline = tenant
5400 6 : .branch_timeline_test(&parent_timeline, child_timeline_id, Some(current_lsn), &ctx)
5401 6 : .await?;
5402 6 :
5403 6 : let mut key = start_key;
5404 6006 : while key < end_key {
5405 6000 : if key == child_gap_at_key {
5406 6 : key = key.next();
5407 6 : continue;
5408 5994 : }
5409 5994 :
5410 5994 : current_lsn += 0x10;
5411 6 :
5412 5994 : let mut writer = child_timeline.writer().await;
5413 5994 : writer
5414 5994 : .put(
5415 5994 : key,
5416 5994 : current_lsn,
5417 5994 : &Value::Image(test_img(&format!("{} at {}", key, current_lsn))),
5418 5994 : &ctx,
5419 5994 : )
5420 51 : .await?;
5421 5994 : writer.finish_write(current_lsn);
5422 5994 :
5423 5994 : key = key.next();
5424 6 : }
5425 6 :
5426 6 : child_timeline.freeze_and_flush().await?;
5427 6 :
5428 6 : let lsn_offsets: [i64; 5] = [-10, -1, 0, 1, 10];
5429 6 : let mut query_lsns = Vec::new();
5430 18 : for image_lsn in parent_gap_lsns.keys().rev() {
5431 108 : for offset in lsn_offsets {
5432 90 : query_lsns.push(Lsn(image_lsn
5433 90 : .0
5434 90 : .checked_add_signed(offset)
5435 90 : .expect("Shouldn't overflow")));
5436 90 : }
5437 6 : }
5438 6 :
5439 96 : for query_lsn in query_lsns {
5440 90 : let results = child_timeline
5441 90 : .get_vectored_impl(
5442 90 : KeySpace {
5443 90 : ranges: vec![child_gap_at_key..child_gap_at_key.next()],
5444 90 : },
5445 90 : query_lsn,
5446 90 : &mut ValuesReconstructState::new(),
5447 90 : &ctx,
5448 90 : )
5449 87 : .await;
5450 6 :
5451 90 : let expected_item = parent_gap_lsns
5452 90 : .iter()
5453 90 : .rev()
5454 204 : .find(|(lsn, _)| **lsn <= query_lsn);
5455 90 :
5456 90 : info!(
5457 6 : "Doing vectored read at LSN {}. Expecting image to be: {:?}",
5458 6 : query_lsn, expected_item
5459 6 : );
5460 6 :
5461 90 : match expected_item {
5462 78 : Some((_, img_value)) => {
5463 78 : let key_results = results.expect("No vectored get error expected");
5464 78 : let key_result = &key_results[&child_gap_at_key];
5465 78 : let returned_img = key_result
5466 78 : .as_ref()
5467 78 : .expect("No page reconstruct error expected");
5468 78 :
5469 78 : info!(
5470 6 : "Vectored read at LSN {} returned image {}",
5471 0 : query_lsn,
5472 0 : std::str::from_utf8(returned_img)?
5473 6 : );
5474 78 : assert_eq!(*returned_img, test_img(img_value));
5475 6 : }
5476 6 : None => {
5477 12 : assert!(matches!(results, Err(GetVectoredError::MissingKey(_))));
5478 6 : }
5479 6 : }
5480 6 : }
5481 6 :
5482 6 : Ok(())
5483 6 : }
5484 :
5485 : #[tokio::test]
5486 6 : async fn test_random_updates() -> anyhow::Result<()> {
5487 6 : let names_algorithms = [
5488 6 : ("test_random_updates_legacy", CompactionAlgorithm::Legacy),
5489 6 : ("test_random_updates_tiered", CompactionAlgorithm::Tiered),
5490 6 : ];
5491 18 : for (name, algorithm) in names_algorithms {
5492 289171 : test_random_updates_algorithm(name, algorithm).await?;
5493 6 : }
5494 6 : Ok(())
5495 6 : }
5496 :
5497 12 : async fn test_random_updates_algorithm(
5498 12 : name: &'static str,
5499 12 : compaction_algorithm: CompactionAlgorithm,
5500 12 : ) -> anyhow::Result<()> {
5501 12 : let mut harness = TenantHarness::create(name).await?;
5502 12 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
5503 12 : kind: compaction_algorithm,
5504 12 : };
5505 48 : let (tenant, ctx) = harness.load().await;
5506 12 : let tline = tenant
5507 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5508 21 : .await?;
5509 :
5510 : const NUM_KEYS: usize = 1000;
5511 12 : let cancel = CancellationToken::new();
5512 12 :
5513 12 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5514 12 : let mut test_key_end = test_key;
5515 12 : test_key_end.field6 = NUM_KEYS as u32;
5516 12 : tline.add_extra_test_dense_keyspace(KeySpace::single(test_key..test_key_end));
5517 12 :
5518 12 : let mut keyspace = KeySpaceAccum::new();
5519 12 :
5520 12 : // Track when each page was last modified. Used to assert that
5521 12 : // a read sees the latest page version.
5522 12 : let mut updated = [Lsn(0); NUM_KEYS];
5523 12 :
5524 12 : let mut lsn = Lsn(0x10);
5525 : #[allow(clippy::needless_range_loop)]
5526 12012 : for blknum in 0..NUM_KEYS {
5527 12000 : lsn = Lsn(lsn.0 + 0x10);
5528 12000 : test_key.field6 = blknum as u32;
5529 12000 : let mut writer = tline.writer().await;
5530 12000 : writer
5531 12000 : .put(
5532 12000 : test_key,
5533 12000 : lsn,
5534 12000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5535 12000 : &ctx,
5536 12000 : )
5537 12 : .await?;
5538 12000 : writer.finish_write(lsn);
5539 12000 : updated[blknum] = lsn;
5540 12000 : drop(writer);
5541 12000 :
5542 12000 : keyspace.add_key(test_key);
5543 : }
5544 :
5545 612 : for _ in 0..50 {
5546 600600 : for _ in 0..NUM_KEYS {
5547 600000 : lsn = Lsn(lsn.0 + 0x10);
5548 600000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
5549 600000 : test_key.field6 = blknum as u32;
5550 600000 : let mut writer = tline.writer().await;
5551 600000 : writer
5552 600000 : .put(
5553 600000 : test_key,
5554 600000 : lsn,
5555 600000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5556 600000 : &ctx,
5557 600000 : )
5558 609 : .await?;
5559 600000 : writer.finish_write(lsn);
5560 600000 : drop(writer);
5561 600000 : updated[blknum] = lsn;
5562 : }
5563 :
5564 : // Read all the blocks
5565 600000 : for (blknum, last_lsn) in updated.iter().enumerate() {
5566 600000 : test_key.field6 = blknum as u32;
5567 600000 : assert_eq!(
5568 600000 : tline.get(test_key, lsn, &ctx).await?,
5569 600000 : test_img(&format!("{} at {}", blknum, last_lsn))
5570 : );
5571 : }
5572 :
5573 : // Perform a cycle of flush, and GC
5574 608 : tline.freeze_and_flush().await?;
5575 600 : tenant
5576 600 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
5577 597 : .await?;
5578 : }
5579 :
5580 12 : Ok(())
5581 12 : }
5582 :
5583 : #[tokio::test]
5584 6 : async fn test_traverse_branches() -> anyhow::Result<()> {
5585 6 : let (tenant, ctx) = TenantHarness::create("test_traverse_branches")
5586 6 : .await?
5587 6 : .load()
5588 24 : .await;
5589 6 : let mut tline = tenant
5590 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5591 12 : .await?;
5592 6 :
5593 6 : const NUM_KEYS: usize = 1000;
5594 6 :
5595 6 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5596 6 :
5597 6 : let mut keyspace = KeySpaceAccum::new();
5598 6 :
5599 6 : let cancel = CancellationToken::new();
5600 6 :
5601 6 : // Track when each page was last modified. Used to assert that
5602 6 : // a read sees the latest page version.
5603 6 : let mut updated = [Lsn(0); NUM_KEYS];
5604 6 :
5605 6 : let mut lsn = Lsn(0x10);
5606 6 : #[allow(clippy::needless_range_loop)]
5607 6006 : for blknum in 0..NUM_KEYS {
5608 6000 : lsn = Lsn(lsn.0 + 0x10);
5609 6000 : test_key.field6 = blknum as u32;
5610 6000 : let mut writer = tline.writer().await;
5611 6000 : writer
5612 6000 : .put(
5613 6000 : test_key,
5614 6000 : lsn,
5615 6000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5616 6000 : &ctx,
5617 6000 : )
5618 6 : .await?;
5619 6000 : writer.finish_write(lsn);
5620 6000 : updated[blknum] = lsn;
5621 6000 : drop(writer);
5622 6000 :
5623 6000 : keyspace.add_key(test_key);
5624 6 : }
5625 6 :
5626 306 : for _ in 0..50 {
5627 300 : let new_tline_id = TimelineId::generate();
5628 300 : tenant
5629 300 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
5630 6 : .await?;
5631 300 : tline = tenant
5632 300 : .get_timeline(new_tline_id, true)
5633 300 : .expect("Should have the branched timeline");
5634 6 :
5635 300300 : for _ in 0..NUM_KEYS {
5636 300000 : lsn = Lsn(lsn.0 + 0x10);
5637 300000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
5638 300000 : test_key.field6 = blknum as u32;
5639 300000 : let mut writer = tline.writer().await;
5640 300000 : writer
5641 300000 : .put(
5642 300000 : test_key,
5643 300000 : lsn,
5644 300000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5645 300000 : &ctx,
5646 300000 : )
5647 2553 : .await?;
5648 300000 : println!("updating {} at {}", blknum, lsn);
5649 300000 : writer.finish_write(lsn);
5650 300000 : drop(writer);
5651 300000 : updated[blknum] = lsn;
5652 6 : }
5653 6 :
5654 6 : // Read all the blocks
5655 300000 : for (blknum, last_lsn) in updated.iter().enumerate() {
5656 300000 : test_key.field6 = blknum as u32;
5657 300000 : assert_eq!(
5658 300000 : tline.get(test_key, lsn, &ctx).await?,
5659 300000 : test_img(&format!("{} at {}", blknum, last_lsn))
5660 6 : );
5661 6 : }
5662 6 :
5663 6 : // Perform a cycle of flush, compact, and GC
5664 306 : tline.freeze_and_flush().await?;
5665 45241 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
5666 300 : tenant
5667 300 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
5668 303 : .await?;
5669 6 : }
5670 6 :
5671 6 : Ok(())
5672 6 : }
5673 :
5674 : #[tokio::test]
5675 6 : async fn test_traverse_ancestors() -> anyhow::Result<()> {
5676 6 : let (tenant, ctx) = TenantHarness::create("test_traverse_ancestors")
5677 6 : .await?
5678 6 : .load()
5679 24 : .await;
5680 6 : let mut tline = tenant
5681 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5682 12 : .await?;
5683 6 :
5684 6 : const NUM_KEYS: usize = 100;
5685 6 : const NUM_TLINES: usize = 50;
5686 6 :
5687 6 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5688 6 : // Track page mutation lsns across different timelines.
5689 6 : let mut updated = [[Lsn(0); NUM_KEYS]; NUM_TLINES];
5690 6 :
5691 6 : let mut lsn = Lsn(0x10);
5692 6 :
5693 6 : #[allow(clippy::needless_range_loop)]
5694 306 : for idx in 0..NUM_TLINES {
5695 300 : let new_tline_id = TimelineId::generate();
5696 300 : tenant
5697 300 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
5698 6 : .await?;
5699 300 : tline = tenant
5700 300 : .get_timeline(new_tline_id, true)
5701 300 : .expect("Should have the branched timeline");
5702 6 :
5703 30300 : for _ in 0..NUM_KEYS {
5704 30000 : lsn = Lsn(lsn.0 + 0x10);
5705 30000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
5706 30000 : test_key.field6 = blknum as u32;
5707 30000 : let mut writer = tline.writer().await;
5708 30000 : writer
5709 30000 : .put(
5710 30000 : test_key,
5711 30000 : lsn,
5712 30000 : &Value::Image(test_img(&format!("{} {} at {}", idx, blknum, lsn))),
5713 30000 : &ctx,
5714 30000 : )
5715 264 : .await?;
5716 30000 : println!("updating [{}][{}] at {}", idx, blknum, lsn);
5717 30000 : writer.finish_write(lsn);
5718 30000 : drop(writer);
5719 30000 : updated[idx][blknum] = lsn;
5720 6 : }
5721 6 : }
5722 6 :
5723 6 : // Read pages from leaf timeline across all ancestors.
5724 300 : for (idx, lsns) in updated.iter().enumerate() {
5725 30000 : for (blknum, lsn) in lsns.iter().enumerate() {
5726 6 : // Skip empty mutations.
5727 30000 : if lsn.0 == 0 {
5728 11039 : continue;
5729 18961 : }
5730 18961 : println!("checking [{idx}][{blknum}] at {lsn}");
5731 18961 : test_key.field6 = blknum as u32;
5732 18961 : assert_eq!(
5733 18961 : tline.get(test_key, *lsn, &ctx).await?,
5734 18961 : test_img(&format!("{idx} {blknum} at {lsn}"))
5735 6 : );
5736 6 : }
5737 6 : }
5738 6 : Ok(())
5739 6 : }
5740 :
5741 : #[tokio::test]
5742 6 : async fn test_write_at_initdb_lsn_takes_optimization_code_path() -> anyhow::Result<()> {
5743 6 : let (tenant, ctx) = TenantHarness::create("test_empty_test_timeline_is_usable")
5744 6 : .await?
5745 6 : .load()
5746 24 : .await;
5747 6 :
5748 6 : let initdb_lsn = Lsn(0x20);
5749 6 : let utline = tenant
5750 6 : .create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)
5751 6 : .await?;
5752 6 : let tline = utline.raw_timeline().unwrap();
5753 6 :
5754 6 : // Spawn flush loop now so that we can set the `expect_initdb_optimization`
5755 6 : tline.maybe_spawn_flush_loop();
5756 6 :
5757 6 : // Make sure the timeline has the minimum set of required keys for operation.
5758 6 : // The only operation you can always do on an empty timeline is to `put` new data.
5759 6 : // Except if you `put` at `initdb_lsn`.
5760 6 : // In that case, there's an optimization to directly create image layers instead of delta layers.
5761 6 : // It uses `repartition()`, which assumes some keys to be present.
5762 6 : // Let's make sure the test timeline can handle that case.
5763 6 : {
5764 6 : let mut state = tline.flush_loop_state.lock().unwrap();
5765 6 : assert_eq!(
5766 6 : timeline::FlushLoopState::Running {
5767 6 : expect_initdb_optimization: false,
5768 6 : initdb_optimization_count: 0,
5769 6 : },
5770 6 : *state
5771 6 : );
5772 6 : *state = timeline::FlushLoopState::Running {
5773 6 : expect_initdb_optimization: true,
5774 6 : initdb_optimization_count: 0,
5775 6 : };
5776 6 : }
5777 6 :
5778 6 : // Make writes at the initdb_lsn. When we flush it below, it should be handled by the optimization.
5779 6 : // As explained above, the optimization requires some keys to be present.
5780 6 : // As per `create_empty_timeline` documentation, use init_empty to set them.
5781 6 : // This is what `create_test_timeline` does, by the way.
5782 6 : let mut modification = tline.begin_modification(initdb_lsn);
5783 6 : modification
5784 6 : .init_empty_test_timeline()
5785 6 : .context("init_empty_test_timeline")?;
5786 6 : modification
5787 6 : .commit(&ctx)
5788 6 : .await
5789 6 : .context("commit init_empty_test_timeline modification")?;
5790 6 :
5791 6 : // Do the flush. The flush code will check the expectations that we set above.
5792 6 : tline.freeze_and_flush().await?;
5793 6 :
5794 6 : // assert freeze_and_flush exercised the initdb optimization
5795 6 : {
5796 6 : let state = tline.flush_loop_state.lock().unwrap();
5797 6 : let timeline::FlushLoopState::Running {
5798 6 : expect_initdb_optimization,
5799 6 : initdb_optimization_count,
5800 6 : } = *state
5801 6 : else {
5802 6 : panic!("unexpected state: {:?}", *state);
5803 6 : };
5804 6 : assert!(expect_initdb_optimization);
5805 6 : assert!(initdb_optimization_count > 0);
5806 6 : }
5807 6 : Ok(())
5808 6 : }
5809 :
5810 : #[tokio::test]
5811 6 : async fn test_create_guard_crash() -> anyhow::Result<()> {
5812 6 : let name = "test_create_guard_crash";
5813 6 : let harness = TenantHarness::create(name).await?;
5814 6 : {
5815 24 : let (tenant, ctx) = harness.load().await;
5816 6 : let tline = tenant
5817 6 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
5818 6 : .await?;
5819 6 : // Leave the timeline ID in [`Tenant::timelines_creating`] to exclude attempting to create it again
5820 6 : let raw_tline = tline.raw_timeline().unwrap();
5821 6 : raw_tline
5822 6 : .shutdown(super::timeline::ShutdownMode::Hard)
5823 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))
5824 6 : .await;
5825 6 : std::mem::forget(tline);
5826 6 : }
5827 6 :
5828 24 : let (tenant, _) = harness.load().await;
5829 6 : match tenant.get_timeline(TIMELINE_ID, false) {
5830 6 : Ok(_) => panic!("timeline should've been removed during load"),
5831 6 : Err(e) => {
5832 6 : assert_eq!(
5833 6 : e,
5834 6 : GetTimelineError::NotFound {
5835 6 : tenant_id: tenant.tenant_shard_id,
5836 6 : timeline_id: TIMELINE_ID,
5837 6 : }
5838 6 : )
5839 6 : }
5840 6 : }
5841 6 :
5842 6 : assert!(!harness
5843 6 : .conf
5844 6 : .timeline_path(&tenant.tenant_shard_id, &TIMELINE_ID)
5845 6 : .exists());
5846 6 :
5847 6 : Ok(())
5848 6 : }
5849 :
5850 : #[tokio::test]
5851 6 : async fn test_read_at_max_lsn() -> anyhow::Result<()> {
5852 6 : let names_algorithms = [
5853 6 : ("test_read_at_max_lsn_legacy", CompactionAlgorithm::Legacy),
5854 6 : ("test_read_at_max_lsn_tiered", CompactionAlgorithm::Tiered),
5855 6 : ];
5856 18 : for (name, algorithm) in names_algorithms {
5857 98199 : test_read_at_max_lsn_algorithm(name, algorithm).await?;
5858 6 : }
5859 6 : Ok(())
5860 6 : }
5861 :
5862 12 : async fn test_read_at_max_lsn_algorithm(
5863 12 : name: &'static str,
5864 12 : compaction_algorithm: CompactionAlgorithm,
5865 12 : ) -> anyhow::Result<()> {
5866 12 : let mut harness = TenantHarness::create(name).await?;
5867 12 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
5868 12 : kind: compaction_algorithm,
5869 12 : };
5870 48 : let (tenant, ctx) = harness.load().await;
5871 12 : let tline = tenant
5872 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5873 21 : .await?;
5874 :
5875 12 : let lsn = Lsn(0x10);
5876 12 : let compact = false;
5877 96900 : bulk_insert_maybe_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000, compact).await?;
5878 :
5879 12 : let test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5880 12 : let read_lsn = Lsn(u64::MAX - 1);
5881 :
5882 1230 : let result = tline.get(test_key, read_lsn, &ctx).await;
5883 12 : assert!(result.is_ok(), "result is not Ok: {}", result.unwrap_err());
5884 :
5885 12 : Ok(())
5886 12 : }
5887 :
5888 : #[tokio::test]
5889 6 : async fn test_metadata_scan() -> anyhow::Result<()> {
5890 6 : let harness = TenantHarness::create("test_metadata_scan").await?;
5891 24 : let (tenant, ctx) = harness.load().await;
5892 6 : let tline = tenant
5893 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5894 12 : .await?;
5895 6 :
5896 6 : const NUM_KEYS: usize = 1000;
5897 6 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
5898 6 :
5899 6 : let cancel = CancellationToken::new();
5900 6 :
5901 6 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
5902 6 : base_key.field1 = AUX_KEY_PREFIX;
5903 6 : let mut test_key = base_key;
5904 6 :
5905 6 : // Track when each page was last modified. Used to assert that
5906 6 : // a read sees the latest page version.
5907 6 : let mut updated = [Lsn(0); NUM_KEYS];
5908 6 :
5909 6 : let mut lsn = Lsn(0x10);
5910 6 : #[allow(clippy::needless_range_loop)]
5911 6006 : for blknum in 0..NUM_KEYS {
5912 6000 : lsn = Lsn(lsn.0 + 0x10);
5913 6000 : test_key.field6 = (blknum * STEP) as u32;
5914 6000 : let mut writer = tline.writer().await;
5915 6000 : writer
5916 6000 : .put(
5917 6000 : test_key,
5918 6000 : lsn,
5919 6000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5920 6000 : &ctx,
5921 6000 : )
5922 6 : .await?;
5923 6000 : writer.finish_write(lsn);
5924 6000 : updated[blknum] = lsn;
5925 6000 : drop(writer);
5926 6 : }
5927 6 :
5928 6 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
5929 6 :
5930 72 : for iter in 0..=10 {
5931 6 : // Read all the blocks
5932 66000 : for (blknum, last_lsn) in updated.iter().enumerate() {
5933 66000 : test_key.field6 = (blknum * STEP) as u32;
5934 66000 : assert_eq!(
5935 66000 : tline.get(test_key, lsn, &ctx).await?,
5936 66000 : test_img(&format!("{} at {}", blknum, last_lsn))
5937 6 : );
5938 6 : }
5939 6 :
5940 66 : let mut cnt = 0;
5941 66000 : for (key, value) in tline
5942 66 : .get_vectored_impl(
5943 66 : keyspace.clone(),
5944 66 : lsn,
5945 66 : &mut ValuesReconstructState::default(),
5946 66 : &ctx,
5947 66 : )
5948 12600 : .await?
5949 6 : {
5950 66000 : let blknum = key.field6 as usize;
5951 66000 : let value = value?;
5952 66000 : assert!(blknum % STEP == 0);
5953 66000 : let blknum = blknum / STEP;
5954 66000 : assert_eq!(
5955 66000 : value,
5956 66000 : test_img(&format!("{} at {}", blknum, updated[blknum]))
5957 66000 : );
5958 66000 : cnt += 1;
5959 6 : }
5960 6 :
5961 66 : assert_eq!(cnt, NUM_KEYS);
5962 6 :
5963 66066 : for _ in 0..NUM_KEYS {
5964 66000 : lsn = Lsn(lsn.0 + 0x10);
5965 66000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
5966 66000 : test_key.field6 = (blknum * STEP) as u32;
5967 66000 : let mut writer = tline.writer().await;
5968 66000 : writer
5969 66000 : .put(
5970 66000 : test_key,
5971 66000 : lsn,
5972 66000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5973 66000 : &ctx,
5974 66000 : )
5975 438 : .await?;
5976 66000 : writer.finish_write(lsn);
5977 66000 : drop(writer);
5978 66000 : updated[blknum] = lsn;
5979 6 : }
5980 6 :
5981 6 : // Perform two cycles of flush, compact, and GC
5982 198 : for round in 0..2 {
5983 132 : tline.freeze_and_flush().await?;
5984 132 : tline
5985 132 : .compact(
5986 132 : &cancel,
5987 132 : if iter % 5 == 0 && round == 0 {
5988 18 : let mut flags = EnumSet::new();
5989 18 : flags.insert(CompactFlags::ForceImageLayerCreation);
5990 18 : flags.insert(CompactFlags::ForceRepartition);
5991 18 : flags
5992 6 : } else {
5993 114 : EnumSet::empty()
5994 6 : },
5995 132 : &ctx,
5996 6 : )
5997 23963 : .await?;
5998 132 : tenant
5999 132 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
6000 126 : .await?;
6001 6 : }
6002 6 : }
6003 6 :
6004 6 : Ok(())
6005 6 : }
6006 :
6007 : #[tokio::test]
6008 6 : async fn test_metadata_compaction_trigger() -> anyhow::Result<()> {
6009 6 : let harness = TenantHarness::create("test_metadata_compaction_trigger").await?;
6010 24 : let (tenant, ctx) = harness.load().await;
6011 6 : let tline = tenant
6012 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6013 12 : .await?;
6014 6 :
6015 6 : let cancel = CancellationToken::new();
6016 6 :
6017 6 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
6018 6 : base_key.field1 = AUX_KEY_PREFIX;
6019 6 : let test_key = base_key;
6020 6 : let mut lsn = Lsn(0x10);
6021 6 :
6022 126 : for _ in 0..20 {
6023 120 : lsn = Lsn(lsn.0 + 0x10);
6024 120 : let mut writer = tline.writer().await;
6025 120 : writer
6026 120 : .put(
6027 120 : test_key,
6028 120 : lsn,
6029 120 : &Value::Image(test_img(&format!("{} at {}", 0, lsn))),
6030 120 : &ctx,
6031 120 : )
6032 60 : .await?;
6033 120 : writer.finish_write(lsn);
6034 120 : drop(writer);
6035 120 : tline.freeze_and_flush().await?; // force create a delta layer
6036 6 : }
6037 6 :
6038 6 : let before_num_l0_delta_files =
6039 6 : tline.layers.read().await.layer_map()?.level0_deltas().len();
6040 6 :
6041 330 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
6042 6 :
6043 6 : let after_num_l0_delta_files = tline.layers.read().await.layer_map()?.level0_deltas().len();
6044 6 :
6045 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}");
6046 6 :
6047 6 : assert_eq!(
6048 12 : tline.get(test_key, lsn, &ctx).await?,
6049 6 : test_img(&format!("{} at {}", 0, lsn))
6050 6 : );
6051 6 :
6052 6 : Ok(())
6053 6 : }
6054 :
6055 : #[tokio::test]
6056 6 : async fn test_branch_copies_dirty_aux_file_flag() {
6057 6 : let harness = TenantHarness::create("test_branch_copies_dirty_aux_file_flag")
6058 6 : .await
6059 6 : .unwrap();
6060 6 :
6061 6 : // the default aux file policy to switch is v2 if not set by the admins
6062 6 : assert_eq!(
6063 6 : harness.tenant_conf.switch_aux_file_policy,
6064 6 : AuxFilePolicy::default_tenant_config()
6065 6 : );
6066 24 : let (tenant, ctx) = harness.load().await;
6067 6 :
6068 6 : let mut lsn = Lsn(0x08);
6069 6 :
6070 6 : let tline: Arc<Timeline> = tenant
6071 6 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
6072 12 : .await
6073 6 : .unwrap();
6074 6 :
6075 6 : // no aux file is written at this point, so the persistent flag should be unset
6076 6 : assert_eq!(tline.last_aux_file_policy.load(), None);
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/test1", b"first", &ctx)
6083 6 : .await
6084 6 : .unwrap();
6085 6 : modification.commit(&ctx).await.unwrap();
6086 6 : }
6087 6 :
6088 6 : // there is no tenant manager to pass the configuration through, so lets mimic it
6089 6 : tenant.set_new_location_config(
6090 6 : AttachedTenantConf::try_from(LocationConf::attached_single(
6091 6 : TenantConfOpt {
6092 6 : switch_aux_file_policy: Some(AuxFilePolicy::V2),
6093 6 : ..Default::default()
6094 6 : },
6095 6 : tenant.generation,
6096 6 : &pageserver_api::models::ShardParameters::default(),
6097 6 : ))
6098 6 : .unwrap(),
6099 6 : );
6100 6 :
6101 6 : assert_eq!(
6102 6 : tline.get_switch_aux_file_policy(),
6103 6 : AuxFilePolicy::V2,
6104 6 : "wanted state has been updated"
6105 6 : );
6106 6 : assert_eq!(
6107 6 : tline.last_aux_file_policy.load(),
6108 6 : Some(AuxFilePolicy::V2),
6109 6 : "aux file is written with switch_aux_file_policy unset (which is v2), so we should use v2 there"
6110 6 : );
6111 6 :
6112 6 : // we can read everything from the storage
6113 6 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6114 6 : assert_eq!(
6115 6 : files.get("pg_logical/mappings/test1"),
6116 6 : Some(&bytes::Bytes::from_static(b"first"))
6117 6 : );
6118 6 :
6119 6 : {
6120 6 : lsn += 8;
6121 6 : let mut modification = tline.begin_modification(lsn);
6122 6 : modification
6123 6 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
6124 6 : .await
6125 6 : .unwrap();
6126 6 : modification.commit(&ctx).await.unwrap();
6127 6 : }
6128 6 :
6129 6 : assert_eq!(
6130 6 : tline.last_aux_file_policy.load(),
6131 6 : Some(AuxFilePolicy::V2),
6132 6 : "keep v2 storage format when new files are written"
6133 6 : );
6134 6 :
6135 6 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6136 6 : assert_eq!(
6137 6 : files.get("pg_logical/mappings/test2"),
6138 6 : Some(&bytes::Bytes::from_static(b"second"))
6139 6 : );
6140 6 :
6141 6 : let child = tenant
6142 6 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(lsn), &ctx)
6143 6 : .await
6144 6 : .unwrap();
6145 6 :
6146 6 : // child copies the last flag even if that is not on remote storage yet
6147 6 : assert_eq!(child.get_switch_aux_file_policy(), AuxFilePolicy::V2);
6148 6 : assert_eq!(child.last_aux_file_policy.load(), Some(AuxFilePolicy::V2));
6149 6 :
6150 6 : let files = child.list_aux_files(lsn, &ctx).await.unwrap();
6151 6 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
6152 6 : assert_eq!(files.get("pg_logical/mappings/test2"), None);
6153 6 :
6154 6 : // even if we crash here without flushing parent timeline with it's new
6155 6 : // last_aux_file_policy we are safe, because child was never meant to access ancestor's
6156 6 : // files. the ancestor can even switch back to V1 because of a migration safely.
6157 6 : }
6158 :
6159 : #[tokio::test]
6160 6 : async fn aux_file_policy_switch() {
6161 6 : let mut harness = TenantHarness::create("aux_file_policy_switch")
6162 6 : .await
6163 6 : .unwrap();
6164 6 : harness.tenant_conf.switch_aux_file_policy = AuxFilePolicy::CrossValidation; // set to cross-validation mode
6165 24 : let (tenant, ctx) = harness.load().await;
6166 6 :
6167 6 : let mut lsn = Lsn(0x08);
6168 6 :
6169 6 : let tline: Arc<Timeline> = tenant
6170 6 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
6171 12 : .await
6172 6 : .unwrap();
6173 6 :
6174 6 : assert_eq!(
6175 6 : tline.last_aux_file_policy.load(),
6176 6 : None,
6177 6 : "no aux file is written so it should be unset"
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/test1", b"first", &ctx)
6185 15 : .await
6186 6 : .unwrap();
6187 6 : modification.commit(&ctx).await.unwrap();
6188 6 : }
6189 6 :
6190 6 : // there is no tenant manager to pass the configuration through, so lets mimic it
6191 6 : tenant.set_new_location_config(
6192 6 : AttachedTenantConf::try_from(LocationConf::attached_single(
6193 6 : TenantConfOpt {
6194 6 : switch_aux_file_policy: Some(AuxFilePolicy::V2),
6195 6 : ..Default::default()
6196 6 : },
6197 6 : tenant.generation,
6198 6 : &pageserver_api::models::ShardParameters::default(),
6199 6 : ))
6200 6 : .unwrap(),
6201 6 : );
6202 6 :
6203 6 : assert_eq!(
6204 6 : tline.get_switch_aux_file_policy(),
6205 6 : AuxFilePolicy::V2,
6206 6 : "wanted state has been updated"
6207 6 : );
6208 6 : assert_eq!(
6209 6 : tline.last_aux_file_policy.load(),
6210 6 : Some(AuxFilePolicy::CrossValidation),
6211 6 : "dirty index_part.json reflected state is yet to be updated"
6212 6 : );
6213 6 :
6214 6 : // we can still read the auxfile v1 before we ingest anything new
6215 6 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6216 6 : assert_eq!(
6217 6 : files.get("pg_logical/mappings/test1"),
6218 6 : Some(&bytes::Bytes::from_static(b"first"))
6219 6 : );
6220 6 :
6221 6 : {
6222 6 : lsn += 8;
6223 6 : let mut modification = tline.begin_modification(lsn);
6224 6 : modification
6225 6 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
6226 6 : .await
6227 6 : .unwrap();
6228 6 : modification.commit(&ctx).await.unwrap();
6229 6 : }
6230 6 :
6231 6 : assert_eq!(
6232 6 : tline.last_aux_file_policy.load(),
6233 6 : Some(AuxFilePolicy::V2),
6234 6 : "ingesting a file should apply the wanted switch state when applicable"
6235 6 : );
6236 6 :
6237 6 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6238 6 : assert_eq!(
6239 6 : files.get("pg_logical/mappings/test1"),
6240 6 : Some(&bytes::Bytes::from_static(b"first")),
6241 6 : "cross validation writes to both v1 and v2 so this should be available in v2"
6242 6 : );
6243 6 : assert_eq!(
6244 6 : files.get("pg_logical/mappings/test2"),
6245 6 : Some(&bytes::Bytes::from_static(b"second"))
6246 6 : );
6247 6 :
6248 6 : // mimic again by trying to flip it from V2 to V1 (not switched to while ingesting a file)
6249 6 : tenant.set_new_location_config(
6250 6 : AttachedTenantConf::try_from(LocationConf::attached_single(
6251 6 : TenantConfOpt {
6252 6 : switch_aux_file_policy: Some(AuxFilePolicy::V1),
6253 6 : ..Default::default()
6254 6 : },
6255 6 : tenant.generation,
6256 6 : &pageserver_api::models::ShardParameters::default(),
6257 6 : ))
6258 6 : .unwrap(),
6259 6 : );
6260 6 :
6261 6 : {
6262 6 : lsn += 8;
6263 6 : let mut modification = tline.begin_modification(lsn);
6264 6 : modification
6265 6 : .put_file("pg_logical/mappings/test2", b"third", &ctx)
6266 6 : .await
6267 6 : .unwrap();
6268 6 : modification.commit(&ctx).await.unwrap();
6269 6 : }
6270 6 :
6271 6 : assert_eq!(
6272 6 : tline.get_switch_aux_file_policy(),
6273 6 : AuxFilePolicy::V1,
6274 6 : "wanted state has been updated again, even if invalid request"
6275 6 : );
6276 6 :
6277 6 : assert_eq!(
6278 6 : tline.last_aux_file_policy.load(),
6279 6 : Some(AuxFilePolicy::V2),
6280 6 : "ingesting a file should apply the wanted switch state when applicable"
6281 6 : );
6282 6 :
6283 6 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6284 6 : assert_eq!(
6285 6 : files.get("pg_logical/mappings/test1"),
6286 6 : Some(&bytes::Bytes::from_static(b"first"))
6287 6 : );
6288 6 : assert_eq!(
6289 6 : files.get("pg_logical/mappings/test2"),
6290 6 : Some(&bytes::Bytes::from_static(b"third"))
6291 6 : );
6292 6 :
6293 6 : // mimic again by trying to flip it from from V1 to V2 (not switched to while ingesting a file)
6294 6 : tenant.set_new_location_config(
6295 6 : AttachedTenantConf::try_from(LocationConf::attached_single(
6296 6 : TenantConfOpt {
6297 6 : switch_aux_file_policy: Some(AuxFilePolicy::V2),
6298 6 : ..Default::default()
6299 6 : },
6300 6 : tenant.generation,
6301 6 : &pageserver_api::models::ShardParameters::default(),
6302 6 : ))
6303 6 : .unwrap(),
6304 6 : );
6305 6 :
6306 6 : {
6307 6 : lsn += 8;
6308 6 : let mut modification = tline.begin_modification(lsn);
6309 6 : modification
6310 6 : .put_file("pg_logical/mappings/test3", b"last", &ctx)
6311 6 : .await
6312 6 : .unwrap();
6313 6 : modification.commit(&ctx).await.unwrap();
6314 6 : }
6315 6 :
6316 6 : assert_eq!(tline.get_switch_aux_file_policy(), AuxFilePolicy::V2);
6317 6 :
6318 6 : assert_eq!(tline.last_aux_file_policy.load(), Some(AuxFilePolicy::V2));
6319 6 :
6320 6 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6321 6 : assert_eq!(
6322 6 : files.get("pg_logical/mappings/test1"),
6323 6 : Some(&bytes::Bytes::from_static(b"first"))
6324 6 : );
6325 6 : assert_eq!(
6326 6 : files.get("pg_logical/mappings/test2"),
6327 6 : Some(&bytes::Bytes::from_static(b"third"))
6328 6 : );
6329 6 : assert_eq!(
6330 6 : files.get("pg_logical/mappings/test3"),
6331 6 : Some(&bytes::Bytes::from_static(b"last"))
6332 6 : );
6333 6 : }
6334 :
6335 : #[tokio::test]
6336 6 : async fn aux_file_policy_force_switch() {
6337 6 : let mut harness = TenantHarness::create("aux_file_policy_force_switch")
6338 6 : .await
6339 6 : .unwrap();
6340 6 : harness.tenant_conf.switch_aux_file_policy = AuxFilePolicy::V1;
6341 24 : let (tenant, ctx) = harness.load().await;
6342 6 :
6343 6 : let mut lsn = Lsn(0x08);
6344 6 :
6345 6 : let tline: Arc<Timeline> = tenant
6346 6 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
6347 12 : .await
6348 6 : .unwrap();
6349 6 :
6350 6 : assert_eq!(
6351 6 : tline.last_aux_file_policy.load(),
6352 6 : None,
6353 6 : "no aux file is written so it should be unset"
6354 6 : );
6355 6 :
6356 6 : {
6357 6 : lsn += 8;
6358 6 : let mut modification = tline.begin_modification(lsn);
6359 6 : modification
6360 6 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
6361 15 : .await
6362 6 : .unwrap();
6363 6 : modification.commit(&ctx).await.unwrap();
6364 6 : }
6365 6 :
6366 6 : tline.do_switch_aux_policy(AuxFilePolicy::V2).unwrap();
6367 6 :
6368 6 : assert_eq!(
6369 6 : tline.last_aux_file_policy.load(),
6370 6 : Some(AuxFilePolicy::V2),
6371 6 : "dirty index_part.json reflected state is yet to be updated"
6372 6 : );
6373 6 :
6374 6 : // lose all data from v1
6375 6 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6376 6 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
6377 6 :
6378 6 : {
6379 6 : lsn += 8;
6380 6 : let mut modification = tline.begin_modification(lsn);
6381 6 : modification
6382 6 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
6383 6 : .await
6384 6 : .unwrap();
6385 6 : modification.commit(&ctx).await.unwrap();
6386 6 : }
6387 6 :
6388 6 : // read data ingested in v2
6389 6 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6390 6 : assert_eq!(
6391 6 : files.get("pg_logical/mappings/test2"),
6392 6 : Some(&bytes::Bytes::from_static(b"second"))
6393 6 : );
6394 6 : // lose all data from v1
6395 6 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
6396 6 : }
6397 :
6398 : #[tokio::test]
6399 6 : async fn aux_file_policy_auto_detect() {
6400 6 : let mut harness = TenantHarness::create("aux_file_policy_auto_detect")
6401 6 : .await
6402 6 : .unwrap();
6403 6 : harness.tenant_conf.switch_aux_file_policy = AuxFilePolicy::V2; // set to cross-validation mode
6404 24 : let (tenant, ctx) = harness.load().await;
6405 6 :
6406 6 : let mut lsn = Lsn(0x08);
6407 6 :
6408 6 : let tline: Arc<Timeline> = tenant
6409 6 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
6410 12 : .await
6411 6 : .unwrap();
6412 6 :
6413 6 : assert_eq!(
6414 6 : tline.last_aux_file_policy.load(),
6415 6 : None,
6416 6 : "no aux file is written so it should be unset"
6417 6 : );
6418 6 :
6419 6 : {
6420 6 : lsn += 8;
6421 6 : let mut modification = tline.begin_modification(lsn);
6422 6 : let buf = AuxFilesDirectory::ser(&AuxFilesDirectory {
6423 6 : files: vec![(
6424 6 : "test_file".to_string(),
6425 6 : Bytes::copy_from_slice(b"test_file"),
6426 6 : )]
6427 6 : .into_iter()
6428 6 : .collect(),
6429 6 : })
6430 6 : .unwrap();
6431 6 : modification.put_for_test(AUX_FILES_KEY, Value::Image(Bytes::from(buf)));
6432 6 : modification.commit(&ctx).await.unwrap();
6433 6 : }
6434 6 :
6435 6 : {
6436 6 : lsn += 8;
6437 6 : let mut modification = tline.begin_modification(lsn);
6438 6 : modification
6439 6 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
6440 6 : .await
6441 6 : .unwrap();
6442 6 : modification.commit(&ctx).await.unwrap();
6443 6 : }
6444 6 :
6445 6 : assert_eq!(
6446 6 : tline.last_aux_file_policy.load(),
6447 6 : Some(AuxFilePolicy::V1),
6448 6 : "keep using v1 because there are aux files writting with v1"
6449 6 : );
6450 6 :
6451 6 : // we can still read the auxfile v1
6452 6 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6453 6 : assert_eq!(
6454 6 : files.get("pg_logical/mappings/test1"),
6455 6 : Some(&bytes::Bytes::from_static(b"first"))
6456 6 : );
6457 6 : assert_eq!(
6458 6 : files.get("test_file"),
6459 6 : Some(&bytes::Bytes::from_static(b"test_file"))
6460 6 : );
6461 6 : }
6462 :
6463 : #[tokio::test]
6464 6 : async fn test_metadata_image_creation() -> anyhow::Result<()> {
6465 6 : let harness = TenantHarness::create("test_metadata_image_creation").await?;
6466 24 : let (tenant, ctx) = harness.load().await;
6467 6 : let tline = tenant
6468 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6469 12 : .await?;
6470 6 :
6471 6 : const NUM_KEYS: usize = 1000;
6472 6 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
6473 6 :
6474 6 : let cancel = CancellationToken::new();
6475 6 :
6476 6 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
6477 6 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
6478 6 : let mut test_key = base_key;
6479 6 : let mut lsn = Lsn(0x10);
6480 6 :
6481 24 : async fn scan_with_statistics(
6482 24 : tline: &Timeline,
6483 24 : keyspace: &KeySpace,
6484 24 : lsn: Lsn,
6485 24 : ctx: &RequestContext,
6486 24 : ) -> anyhow::Result<(BTreeMap<Key, Result<Bytes, PageReconstructError>>, usize)> {
6487 24 : let mut reconstruct_state = ValuesReconstructState::default();
6488 24 : let res = tline
6489 24 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
6490 4040 : .await?;
6491 24 : Ok((res, reconstruct_state.get_delta_layers_visited() as usize))
6492 24 : }
6493 6 :
6494 6 : #[allow(clippy::needless_range_loop)]
6495 6006 : for blknum in 0..NUM_KEYS {
6496 6000 : lsn = Lsn(lsn.0 + 0x10);
6497 6000 : test_key.field6 = (blknum * STEP) as u32;
6498 6000 : let mut writer = tline.writer().await;
6499 6000 : writer
6500 6000 : .put(
6501 6000 : test_key,
6502 6000 : lsn,
6503 6000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6504 6000 : &ctx,
6505 6000 : )
6506 6 : .await?;
6507 6000 : writer.finish_write(lsn);
6508 6000 : drop(writer);
6509 6 : }
6510 6 :
6511 6 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
6512 6 :
6513 66 : for iter in 1..=10 {
6514 60060 : for _ in 0..NUM_KEYS {
6515 60000 : lsn = Lsn(lsn.0 + 0x10);
6516 60000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6517 60000 : test_key.field6 = (blknum * STEP) as u32;
6518 60000 : let mut writer = tline.writer().await;
6519 60000 : writer
6520 60000 : .put(
6521 60000 : test_key,
6522 60000 : lsn,
6523 60000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6524 60000 : &ctx,
6525 60000 : )
6526 102 : .await?;
6527 60000 : writer.finish_write(lsn);
6528 60000 : drop(writer);
6529 6 : }
6530 6 :
6531 60 : tline.freeze_and_flush().await?;
6532 6 :
6533 60 : if iter % 5 == 0 {
6534 12 : let (_, before_delta_file_accessed) =
6535 4014 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
6536 12 : tline
6537 12 : .compact(
6538 12 : &cancel,
6539 12 : {
6540 12 : let mut flags = EnumSet::new();
6541 12 : flags.insert(CompactFlags::ForceImageLayerCreation);
6542 12 : flags.insert(CompactFlags::ForceRepartition);
6543 12 : flags
6544 12 : },
6545 12 : &ctx,
6546 12 : )
6547 17256 : .await?;
6548 12 : let (_, after_delta_file_accessed) =
6549 26 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
6550 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}");
6551 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.
6552 12 : assert!(
6553 12 : after_delta_file_accessed <= 2,
6554 6 : "after_delta_file_accessed={after_delta_file_accessed}"
6555 6 : );
6556 48 : }
6557 6 : }
6558 6 :
6559 6 : Ok(())
6560 6 : }
6561 :
6562 : #[tokio::test]
6563 6 : async fn test_vectored_missing_data_key_reads() -> anyhow::Result<()> {
6564 6 : let harness = TenantHarness::create("test_vectored_missing_data_key_reads").await?;
6565 24 : let (tenant, ctx) = harness.load().await;
6566 6 :
6567 6 : let base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
6568 6 : let base_key_child = Key::from_hex("000000000033333333444444445500000001").unwrap();
6569 6 : let base_key_nonexist = Key::from_hex("000000000033333333444444445500000002").unwrap();
6570 6 :
6571 6 : let tline = tenant
6572 6 : .create_test_timeline_with_layers(
6573 6 : TIMELINE_ID,
6574 6 : Lsn(0x10),
6575 6 : DEFAULT_PG_VERSION,
6576 6 : &ctx,
6577 6 : Vec::new(), // delta layers
6578 6 : vec![(Lsn(0x20), vec![(base_key, test_img("data key 1"))])], // image layers
6579 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
6580 6 : )
6581 33 : .await?;
6582 6 : tline.add_extra_test_dense_keyspace(KeySpace::single(base_key..(base_key_nonexist.next())));
6583 6 :
6584 6 : let child = tenant
6585 6 : .branch_timeline_test_with_layers(
6586 6 : &tline,
6587 6 : NEW_TIMELINE_ID,
6588 6 : Some(Lsn(0x20)),
6589 6 : &ctx,
6590 6 : Vec::new(), // delta layers
6591 6 : vec![(Lsn(0x30), vec![(base_key_child, test_img("data key 2"))])], // image layers
6592 6 : Lsn(0x30),
6593 6 : )
6594 21 : .await
6595 6 : .unwrap();
6596 6 :
6597 6 : let lsn = Lsn(0x30);
6598 6 :
6599 6 : // test vectored get on parent timeline
6600 6 : assert_eq!(
6601 12 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
6602 6 : Some(test_img("data key 1"))
6603 6 : );
6604 6 : assert!(get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx)
6605 9 : .await
6606 6 : .unwrap_err()
6607 6 : .is_missing_key_error());
6608 6 : assert!(
6609 6 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx)
6610 6 : .await
6611 6 : .unwrap_err()
6612 6 : .is_missing_key_error()
6613 6 : );
6614 6 :
6615 6 : // test vectored get on child timeline
6616 6 : assert_eq!(
6617 6 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
6618 6 : Some(test_img("data key 1"))
6619 6 : );
6620 6 : assert_eq!(
6621 12 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
6622 6 : Some(test_img("data key 2"))
6623 6 : );
6624 6 : assert!(
6625 6 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx)
6626 6 : .await
6627 6 : .unwrap_err()
6628 6 : .is_missing_key_error()
6629 6 : );
6630 6 :
6631 6 : Ok(())
6632 6 : }
6633 :
6634 : #[tokio::test]
6635 6 : async fn test_vectored_missing_metadata_key_reads() -> anyhow::Result<()> {
6636 6 : let harness = TenantHarness::create("test_vectored_missing_metadata_key_reads").await?;
6637 24 : let (tenant, ctx) = harness.load().await;
6638 6 :
6639 6 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
6640 6 : let base_key_child = Key::from_hex("620000000033333333444444445500000001").unwrap();
6641 6 : let base_key_nonexist = Key::from_hex("620000000033333333444444445500000002").unwrap();
6642 6 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
6643 6 :
6644 6 : let tline = tenant
6645 6 : .create_test_timeline_with_layers(
6646 6 : TIMELINE_ID,
6647 6 : Lsn(0x10),
6648 6 : DEFAULT_PG_VERSION,
6649 6 : &ctx,
6650 6 : Vec::new(), // delta layers
6651 6 : vec![(Lsn(0x20), vec![(base_key, test_img("metadata key 1"))])], // image layers
6652 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
6653 6 : )
6654 33 : .await?;
6655 6 :
6656 6 : let child = tenant
6657 6 : .branch_timeline_test_with_layers(
6658 6 : &tline,
6659 6 : NEW_TIMELINE_ID,
6660 6 : Some(Lsn(0x20)),
6661 6 : &ctx,
6662 6 : Vec::new(), // delta layers
6663 6 : vec![(
6664 6 : Lsn(0x30),
6665 6 : vec![(base_key_child, test_img("metadata key 2"))],
6666 6 : )], // image layers
6667 6 : Lsn(0x30),
6668 6 : )
6669 21 : .await
6670 6 : .unwrap();
6671 6 :
6672 6 : let lsn = Lsn(0x30);
6673 6 :
6674 6 : // test vectored get on parent timeline
6675 6 : assert_eq!(
6676 12 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
6677 6 : Some(test_img("metadata key 1"))
6678 6 : );
6679 6 : assert_eq!(
6680 6 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx).await?,
6681 6 : None
6682 6 : );
6683 6 : assert_eq!(
6684 6 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx).await?,
6685 6 : None
6686 6 : );
6687 6 :
6688 6 : // test vectored get on child timeline
6689 6 : assert_eq!(
6690 6 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
6691 6 : None
6692 6 : );
6693 6 : assert_eq!(
6694 12 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
6695 6 : Some(test_img("metadata key 2"))
6696 6 : );
6697 6 : assert_eq!(
6698 6 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx).await?,
6699 6 : None
6700 6 : );
6701 6 :
6702 6 : Ok(())
6703 6 : }
6704 :
6705 108 : async fn get_vectored_impl_wrapper(
6706 108 : tline: &Arc<Timeline>,
6707 108 : key: Key,
6708 108 : lsn: Lsn,
6709 108 : ctx: &RequestContext,
6710 108 : ) -> Result<Option<Bytes>, GetVectoredError> {
6711 108 : let mut reconstruct_state = ValuesReconstructState::new();
6712 108 : let mut res = tline
6713 108 : .get_vectored_impl(
6714 108 : KeySpace::single(key..key.next()),
6715 108 : lsn,
6716 108 : &mut reconstruct_state,
6717 108 : ctx,
6718 108 : )
6719 99 : .await?;
6720 90 : Ok(res.pop_last().map(|(k, v)| {
6721 54 : assert_eq!(k, key);
6722 54 : v.unwrap()
6723 90 : }))
6724 108 : }
6725 :
6726 : #[tokio::test]
6727 6 : async fn test_metadata_tombstone_reads() -> anyhow::Result<()> {
6728 6 : let harness = TenantHarness::create("test_metadata_tombstone_reads").await?;
6729 24 : let (tenant, ctx) = harness.load().await;
6730 6 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
6731 6 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
6732 6 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
6733 6 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
6734 6 :
6735 6 : // We emulate the situation that the compaction algorithm creates an image layer that removes the tombstones
6736 6 : // Lsn 0x30 key0, key3, no key1+key2
6737 6 : // Lsn 0x20 key1+key2 tomestones
6738 6 : // Lsn 0x10 key1 in image, key2 in delta
6739 6 : let tline = tenant
6740 6 : .create_test_timeline_with_layers(
6741 6 : TIMELINE_ID,
6742 6 : Lsn(0x10),
6743 6 : DEFAULT_PG_VERSION,
6744 6 : &ctx,
6745 6 : // delta layers
6746 6 : vec![
6747 6 : DeltaLayerTestDesc::new_with_inferred_key_range(
6748 6 : Lsn(0x10)..Lsn(0x20),
6749 6 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
6750 6 : ),
6751 6 : DeltaLayerTestDesc::new_with_inferred_key_range(
6752 6 : Lsn(0x20)..Lsn(0x30),
6753 6 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
6754 6 : ),
6755 6 : DeltaLayerTestDesc::new_with_inferred_key_range(
6756 6 : Lsn(0x20)..Lsn(0x30),
6757 6 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
6758 6 : ),
6759 6 : ],
6760 6 : // image layers
6761 6 : vec![
6762 6 : (Lsn(0x10), vec![(key1, test_img("metadata key 1"))]),
6763 6 : (
6764 6 : Lsn(0x30),
6765 6 : vec![
6766 6 : (key0, test_img("metadata key 0")),
6767 6 : (key3, test_img("metadata key 3")),
6768 6 : ],
6769 6 : ),
6770 6 : ],
6771 6 : Lsn(0x30),
6772 6 : )
6773 114 : .await?;
6774 6 :
6775 6 : let lsn = Lsn(0x30);
6776 6 : let old_lsn = Lsn(0x20);
6777 6 :
6778 6 : assert_eq!(
6779 12 : get_vectored_impl_wrapper(&tline, key0, lsn, &ctx).await?,
6780 6 : Some(test_img("metadata key 0"))
6781 6 : );
6782 6 : assert_eq!(
6783 6 : get_vectored_impl_wrapper(&tline, key1, lsn, &ctx).await?,
6784 6 : None,
6785 6 : );
6786 6 : assert_eq!(
6787 6 : get_vectored_impl_wrapper(&tline, key2, lsn, &ctx).await?,
6788 6 : None,
6789 6 : );
6790 6 : assert_eq!(
6791 12 : get_vectored_impl_wrapper(&tline, key1, old_lsn, &ctx).await?,
6792 6 : Some(Bytes::new()),
6793 6 : );
6794 6 : assert_eq!(
6795 12 : get_vectored_impl_wrapper(&tline, key2, old_lsn, &ctx).await?,
6796 6 : Some(Bytes::new()),
6797 6 : );
6798 6 : assert_eq!(
6799 6 : get_vectored_impl_wrapper(&tline, key3, lsn, &ctx).await?,
6800 6 : Some(test_img("metadata key 3"))
6801 6 : );
6802 6 :
6803 6 : Ok(())
6804 6 : }
6805 :
6806 : #[tokio::test]
6807 6 : async fn test_metadata_tombstone_image_creation() {
6808 6 : let harness = TenantHarness::create("test_metadata_tombstone_image_creation")
6809 6 : .await
6810 6 : .unwrap();
6811 24 : let (tenant, ctx) = harness.load().await;
6812 6 :
6813 6 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
6814 6 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
6815 6 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
6816 6 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
6817 6 :
6818 6 : let tline = tenant
6819 6 : .create_test_timeline_with_layers(
6820 6 : TIMELINE_ID,
6821 6 : Lsn(0x10),
6822 6 : DEFAULT_PG_VERSION,
6823 6 : &ctx,
6824 6 : // delta layers
6825 6 : vec![
6826 6 : DeltaLayerTestDesc::new_with_inferred_key_range(
6827 6 : Lsn(0x10)..Lsn(0x20),
6828 6 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
6829 6 : ),
6830 6 : DeltaLayerTestDesc::new_with_inferred_key_range(
6831 6 : Lsn(0x20)..Lsn(0x30),
6832 6 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
6833 6 : ),
6834 6 : DeltaLayerTestDesc::new_with_inferred_key_range(
6835 6 : Lsn(0x20)..Lsn(0x30),
6836 6 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
6837 6 : ),
6838 6 : DeltaLayerTestDesc::new_with_inferred_key_range(
6839 6 : Lsn(0x30)..Lsn(0x40),
6840 6 : vec![
6841 6 : (key0, Lsn(0x30), Value::Image(test_img("metadata key 0"))),
6842 6 : (key3, Lsn(0x30), Value::Image(test_img("metadata key 3"))),
6843 6 : ],
6844 6 : ),
6845 6 : ],
6846 6 : // image layers
6847 6 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
6848 6 : Lsn(0x40),
6849 6 : )
6850 105 : .await
6851 6 : .unwrap();
6852 6 :
6853 6 : let cancel = CancellationToken::new();
6854 6 :
6855 6 : tline
6856 6 : .compact(
6857 6 : &cancel,
6858 6 : {
6859 6 : let mut flags = EnumSet::new();
6860 6 : flags.insert(CompactFlags::ForceImageLayerCreation);
6861 6 : flags.insert(CompactFlags::ForceRepartition);
6862 6 : flags
6863 6 : },
6864 6 : &ctx,
6865 6 : )
6866 165 : .await
6867 6 : .unwrap();
6868 6 :
6869 6 : // Image layers are created at last_record_lsn
6870 6 : let images = tline
6871 6 : .inspect_image_layers(Lsn(0x40), &ctx)
6872 24 : .await
6873 6 : .unwrap()
6874 6 : .into_iter()
6875 54 : .filter(|(k, _)| k.is_metadata_key())
6876 6 : .collect::<Vec<_>>();
6877 6 : assert_eq!(images.len(), 2); // the image layer should only contain two existing keys, tombstones should be removed.
6878 6 : }
6879 :
6880 : #[tokio::test]
6881 6 : async fn test_metadata_tombstone_empty_image_creation() {
6882 6 : let harness = TenantHarness::create("test_metadata_tombstone_empty_image_creation")
6883 6 : .await
6884 6 : .unwrap();
6885 24 : let (tenant, ctx) = harness.load().await;
6886 6 :
6887 6 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
6888 6 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
6889 6 :
6890 6 : let tline = tenant
6891 6 : .create_test_timeline_with_layers(
6892 6 : TIMELINE_ID,
6893 6 : Lsn(0x10),
6894 6 : DEFAULT_PG_VERSION,
6895 6 : &ctx,
6896 6 : // delta layers
6897 6 : vec![
6898 6 : DeltaLayerTestDesc::new_with_inferred_key_range(
6899 6 : Lsn(0x10)..Lsn(0x20),
6900 6 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
6901 6 : ),
6902 6 : DeltaLayerTestDesc::new_with_inferred_key_range(
6903 6 : Lsn(0x20)..Lsn(0x30),
6904 6 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
6905 6 : ),
6906 6 : DeltaLayerTestDesc::new_with_inferred_key_range(
6907 6 : Lsn(0x20)..Lsn(0x30),
6908 6 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
6909 6 : ),
6910 6 : ],
6911 6 : // image layers
6912 6 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
6913 6 : Lsn(0x30),
6914 6 : )
6915 87 : .await
6916 6 : .unwrap();
6917 6 :
6918 6 : let cancel = CancellationToken::new();
6919 6 :
6920 6 : tline
6921 6 : .compact(
6922 6 : &cancel,
6923 6 : {
6924 6 : let mut flags = EnumSet::new();
6925 6 : flags.insert(CompactFlags::ForceImageLayerCreation);
6926 6 : flags.insert(CompactFlags::ForceRepartition);
6927 6 : flags
6928 6 : },
6929 6 : &ctx,
6930 6 : )
6931 129 : .await
6932 6 : .unwrap();
6933 6 :
6934 6 : // Image layers are created at last_record_lsn
6935 6 : let images = tline
6936 6 : .inspect_image_layers(Lsn(0x30), &ctx)
6937 12 : .await
6938 6 : .unwrap()
6939 6 : .into_iter()
6940 42 : .filter(|(k, _)| k.is_metadata_key())
6941 6 : .collect::<Vec<_>>();
6942 6 : assert_eq!(images.len(), 0); // the image layer should not contain tombstones, or it is not created
6943 6 : }
6944 :
6945 : #[tokio::test]
6946 6 : async fn test_simple_bottom_most_compaction_images() -> anyhow::Result<()> {
6947 6 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_images").await?;
6948 24 : let (tenant, ctx) = harness.load().await;
6949 6 :
6950 306 : fn get_key(id: u32) -> Key {
6951 306 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
6952 306 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
6953 306 : key.field6 = id;
6954 306 : key
6955 306 : }
6956 6 :
6957 6 : // We create
6958 6 : // - one bottom-most image layer,
6959 6 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
6960 6 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
6961 6 : // - a delta layer D3 above the horizon.
6962 6 : //
6963 6 : // | D3 |
6964 6 : // | D1 |
6965 6 : // -| |-- gc horizon -----------------
6966 6 : // | | | D2 |
6967 6 : // --------- img layer ------------------
6968 6 : //
6969 6 : // What we should expact from this compaction is:
6970 6 : // | D3 |
6971 6 : // | Part of D1 |
6972 6 : // --------- img layer with D1+D2 at GC horizon------------------
6973 6 :
6974 6 : // img layer at 0x10
6975 6 : let img_layer = (0..10)
6976 60 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
6977 6 : .collect_vec();
6978 6 :
6979 6 : let delta1 = vec![
6980 6 : (
6981 6 : get_key(1),
6982 6 : Lsn(0x20),
6983 6 : Value::Image(Bytes::from("value 1@0x20")),
6984 6 : ),
6985 6 : (
6986 6 : get_key(2),
6987 6 : Lsn(0x30),
6988 6 : Value::Image(Bytes::from("value 2@0x30")),
6989 6 : ),
6990 6 : (
6991 6 : get_key(3),
6992 6 : Lsn(0x40),
6993 6 : Value::Image(Bytes::from("value 3@0x40")),
6994 6 : ),
6995 6 : ];
6996 6 : let delta2 = vec![
6997 6 : (
6998 6 : get_key(5),
6999 6 : Lsn(0x20),
7000 6 : Value::Image(Bytes::from("value 5@0x20")),
7001 6 : ),
7002 6 : (
7003 6 : get_key(6),
7004 6 : Lsn(0x20),
7005 6 : Value::Image(Bytes::from("value 6@0x20")),
7006 6 : ),
7007 6 : ];
7008 6 : let delta3 = vec![
7009 6 : (
7010 6 : get_key(8),
7011 6 : Lsn(0x48),
7012 6 : Value::Image(Bytes::from("value 8@0x48")),
7013 6 : ),
7014 6 : (
7015 6 : get_key(9),
7016 6 : Lsn(0x48),
7017 6 : Value::Image(Bytes::from("value 9@0x48")),
7018 6 : ),
7019 6 : ];
7020 6 :
7021 6 : let tline = tenant
7022 6 : .create_test_timeline_with_layers(
7023 6 : TIMELINE_ID,
7024 6 : Lsn(0x10),
7025 6 : DEFAULT_PG_VERSION,
7026 6 : &ctx,
7027 6 : vec![
7028 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
7029 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
7030 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
7031 6 : ], // delta layers
7032 6 : vec![(Lsn(0x10), img_layer)], // image layers
7033 6 : Lsn(0x50),
7034 6 : )
7035 141 : .await?;
7036 6 : {
7037 6 : // Update GC info
7038 6 : let mut guard = tline.gc_info.write().unwrap();
7039 6 : guard.cutoffs.time = Lsn(0x30);
7040 6 : guard.cutoffs.space = Lsn(0x30);
7041 6 : }
7042 6 :
7043 6 : let expected_result = [
7044 6 : Bytes::from_static(b"value 0@0x10"),
7045 6 : Bytes::from_static(b"value 1@0x20"),
7046 6 : Bytes::from_static(b"value 2@0x30"),
7047 6 : Bytes::from_static(b"value 3@0x40"),
7048 6 : Bytes::from_static(b"value 4@0x10"),
7049 6 : Bytes::from_static(b"value 5@0x20"),
7050 6 : Bytes::from_static(b"value 6@0x20"),
7051 6 : Bytes::from_static(b"value 7@0x10"),
7052 6 : Bytes::from_static(b"value 8@0x48"),
7053 6 : Bytes::from_static(b"value 9@0x48"),
7054 6 : ];
7055 6 :
7056 60 : for (idx, expected) in expected_result.iter().enumerate() {
7057 60 : assert_eq!(
7058 60 : tline
7059 60 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7060 69 : .await
7061 60 : .unwrap(),
7062 6 : expected
7063 6 : );
7064 6 : }
7065 6 :
7066 6 : let cancel = CancellationToken::new();
7067 6 : tline
7068 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7069 159 : .await
7070 6 : .unwrap();
7071 6 :
7072 60 : for (idx, expected) in expected_result.iter().enumerate() {
7073 60 : assert_eq!(
7074 60 : tline
7075 60 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7076 48 : .await
7077 60 : .unwrap(),
7078 6 : expected
7079 6 : );
7080 6 : }
7081 6 :
7082 6 : // Check if the image layer at the GC horizon contains exactly what we want
7083 6 : let image_at_gc_horizon = tline
7084 6 : .inspect_image_layers(Lsn(0x30), &ctx)
7085 6 : .await
7086 6 : .unwrap()
7087 6 : .into_iter()
7088 102 : .filter(|(k, _)| k.is_metadata_key())
7089 6 : .collect::<Vec<_>>();
7090 6 :
7091 6 : assert_eq!(image_at_gc_horizon.len(), 10);
7092 6 : let expected_result = [
7093 6 : Bytes::from_static(b"value 0@0x10"),
7094 6 : Bytes::from_static(b"value 1@0x20"),
7095 6 : Bytes::from_static(b"value 2@0x30"),
7096 6 : Bytes::from_static(b"value 3@0x10"),
7097 6 : Bytes::from_static(b"value 4@0x10"),
7098 6 : Bytes::from_static(b"value 5@0x20"),
7099 6 : Bytes::from_static(b"value 6@0x20"),
7100 6 : Bytes::from_static(b"value 7@0x10"),
7101 6 : Bytes::from_static(b"value 8@0x10"),
7102 6 : Bytes::from_static(b"value 9@0x10"),
7103 6 : ];
7104 66 : for idx in 0..10 {
7105 60 : assert_eq!(
7106 60 : image_at_gc_horizon[idx],
7107 60 : (get_key(idx as u32), expected_result[idx].clone())
7108 60 : );
7109 6 : }
7110 6 :
7111 6 : // Check if old layers are removed / new layers have the expected LSN
7112 6 : let mut all_layers = tline.inspect_historic_layers().await.unwrap();
7113 12 : all_layers.sort_by(|k1, k2| {
7114 12 : (
7115 12 : k1.is_delta,
7116 12 : k1.key_range.start,
7117 12 : k1.key_range.end,
7118 12 : k1.lsn_range.start,
7119 12 : k1.lsn_range.end,
7120 12 : )
7121 12 : .cmp(&(
7122 12 : k2.is_delta,
7123 12 : k2.key_range.start,
7124 12 : k2.key_range.end,
7125 12 : k2.lsn_range.start,
7126 12 : k2.lsn_range.end,
7127 12 : ))
7128 12 : });
7129 6 : assert_eq!(
7130 6 : all_layers,
7131 6 : vec![
7132 6 : // Image layer at GC horizon
7133 6 : PersistentLayerKey {
7134 6 : key_range: Key::MIN..Key::MAX,
7135 6 : lsn_range: Lsn(0x30)..Lsn(0x31),
7136 6 : is_delta: false
7137 6 : },
7138 6 : // The delta layer below the horizon
7139 6 : PersistentLayerKey {
7140 6 : key_range: get_key(3)..get_key(4),
7141 6 : lsn_range: Lsn(0x30)..Lsn(0x48),
7142 6 : is_delta: true
7143 6 : },
7144 6 : // The delta3 layer that should not be picked for the compaction
7145 6 : PersistentLayerKey {
7146 6 : key_range: get_key(8)..get_key(10),
7147 6 : lsn_range: Lsn(0x48)..Lsn(0x50),
7148 6 : is_delta: true
7149 6 : }
7150 6 : ]
7151 6 : );
7152 6 :
7153 6 : // increase GC horizon and compact again
7154 6 : {
7155 6 : // Update GC info
7156 6 : let mut guard = tline.gc_info.write().unwrap();
7157 6 : guard.cutoffs.time = Lsn(0x40);
7158 6 : guard.cutoffs.space = Lsn(0x40);
7159 6 : }
7160 6 : tline
7161 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7162 126 : .await
7163 6 : .unwrap();
7164 6 :
7165 6 : Ok(())
7166 6 : }
7167 :
7168 : #[tokio::test]
7169 6 : async fn test_neon_test_record() -> anyhow::Result<()> {
7170 6 : let harness = TenantHarness::create("test_neon_test_record").await?;
7171 24 : let (tenant, ctx) = harness.load().await;
7172 6 :
7173 72 : fn get_key(id: u32) -> Key {
7174 72 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7175 72 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7176 72 : key.field6 = id;
7177 72 : key
7178 72 : }
7179 6 :
7180 6 : let delta1 = vec![
7181 6 : (
7182 6 : get_key(1),
7183 6 : Lsn(0x20),
7184 6 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
7185 6 : ),
7186 6 : (
7187 6 : get_key(1),
7188 6 : Lsn(0x30),
7189 6 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
7190 6 : ),
7191 6 : (get_key(2), Lsn(0x10), Value::Image("0x10".into())),
7192 6 : (
7193 6 : get_key(2),
7194 6 : Lsn(0x20),
7195 6 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
7196 6 : ),
7197 6 : (
7198 6 : get_key(2),
7199 6 : Lsn(0x30),
7200 6 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
7201 6 : ),
7202 6 : (get_key(3), Lsn(0x10), Value::Image("0x10".into())),
7203 6 : (
7204 6 : get_key(3),
7205 6 : Lsn(0x20),
7206 6 : Value::WalRecord(NeonWalRecord::wal_clear()),
7207 6 : ),
7208 6 : (get_key(4), Lsn(0x10), Value::Image("0x10".into())),
7209 6 : (
7210 6 : get_key(4),
7211 6 : Lsn(0x20),
7212 6 : Value::WalRecord(NeonWalRecord::wal_init()),
7213 6 : ),
7214 6 : ];
7215 6 : let image1 = vec![(get_key(1), "0x10".into())];
7216 6 :
7217 6 : let tline = tenant
7218 6 : .create_test_timeline_with_layers(
7219 6 : TIMELINE_ID,
7220 6 : Lsn(0x10),
7221 6 : DEFAULT_PG_VERSION,
7222 6 : &ctx,
7223 6 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
7224 6 : Lsn(0x10)..Lsn(0x40),
7225 6 : delta1,
7226 6 : )], // delta layers
7227 6 : vec![(Lsn(0x10), image1)], // image layers
7228 6 : Lsn(0x50),
7229 6 : )
7230 51 : .await?;
7231 6 :
7232 6 : assert_eq!(
7233 24 : tline.get(get_key(1), Lsn(0x50), &ctx).await?,
7234 6 : Bytes::from_static(b"0x10,0x20,0x30")
7235 6 : );
7236 6 : assert_eq!(
7237 6 : tline.get(get_key(2), Lsn(0x50), &ctx).await?,
7238 6 : Bytes::from_static(b"0x10,0x20,0x30")
7239 6 : );
7240 6 :
7241 6 : // Need to remove the limit of "Neon WAL redo requires base image".
7242 6 :
7243 6 : // assert_eq!(tline.get(get_key(3), Lsn(0x50), &ctx).await?, Bytes::new());
7244 6 : // assert_eq!(tline.get(get_key(4), Lsn(0x50), &ctx).await?, Bytes::new());
7245 6 :
7246 6 : Ok(())
7247 6 : }
7248 :
7249 : #[tokio::test]
7250 6 : async fn test_lsn_lease() -> anyhow::Result<()> {
7251 24 : let (tenant, ctx) = TenantHarness::create("test_lsn_lease").await?.load().await;
7252 6 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7253 6 :
7254 6 : let end_lsn = Lsn(0x100);
7255 6 : let image_layers = (0x20..=0x90)
7256 6 : .step_by(0x10)
7257 48 : .map(|n| {
7258 48 : (
7259 48 : Lsn(n),
7260 48 : vec![(key, test_img(&format!("data key at {:x}", n)))],
7261 48 : )
7262 48 : })
7263 6 : .collect();
7264 6 :
7265 6 : let timeline = tenant
7266 6 : .create_test_timeline_with_layers(
7267 6 : TIMELINE_ID,
7268 6 : Lsn(0x10),
7269 6 : DEFAULT_PG_VERSION,
7270 6 : &ctx,
7271 6 : Vec::new(),
7272 6 : image_layers,
7273 6 : end_lsn,
7274 6 : )
7275 180 : .await?;
7276 6 :
7277 6 : let leased_lsns = [0x30, 0x50, 0x70];
7278 6 : let mut leases = Vec::new();
7279 18 : let _: anyhow::Result<_> = leased_lsns.iter().try_for_each(|n| {
7280 18 : leases.push(timeline.make_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)?);
7281 18 : Ok(())
7282 18 : });
7283 6 :
7284 6 : // Renewing with shorter lease should not change the lease.
7285 6 : let updated_lease_0 =
7286 6 : timeline.make_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)?;
7287 6 : assert_eq!(updated_lease_0.valid_until, leases[0].valid_until);
7288 6 :
7289 6 : // Renewing with a long lease should renew lease with later expiration time.
7290 6 : let updated_lease_1 = timeline.make_lsn_lease(
7291 6 : Lsn(leased_lsns[1]),
7292 6 : timeline.get_lsn_lease_length() * 2,
7293 6 : &ctx,
7294 6 : )?;
7295 6 :
7296 6 : assert!(updated_lease_1.valid_until > leases[1].valid_until);
7297 6 :
7298 6 : // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
7299 6 : info!(
7300 6 : "latest_gc_cutoff_lsn: {}",
7301 0 : *timeline.get_latest_gc_cutoff_lsn()
7302 6 : );
7303 6 : timeline.force_set_disk_consistent_lsn(end_lsn);
7304 6 :
7305 6 : let res = tenant
7306 6 : .gc_iteration(
7307 6 : Some(TIMELINE_ID),
7308 6 : 0,
7309 6 : Duration::ZERO,
7310 6 : &CancellationToken::new(),
7311 6 : &ctx,
7312 6 : )
7313 6 : .await?;
7314 6 :
7315 6 : // Keeping everything <= Lsn(0x80) b/c leases:
7316 6 : // 0/10: initdb layer
7317 6 : // (0/20..=0/70).step_by(0x10): image layers added when creating the timeline.
7318 6 : assert_eq!(res.layers_needed_by_leases, 7);
7319 6 : // Keeping 0/90 b/c it is the latest layer.
7320 6 : assert_eq!(res.layers_not_updated, 1);
7321 6 : // Removed 0/80.
7322 6 : assert_eq!(res.layers_removed, 1);
7323 6 :
7324 6 : // Make lease on a already GC-ed LSN.
7325 6 : // 0/80 does not have a valid lease + is below latest_gc_cutoff
7326 6 : assert!(Lsn(0x80) < *timeline.get_latest_gc_cutoff_lsn());
7327 6 : let res = timeline.make_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx);
7328 6 : assert!(res.is_err());
7329 6 :
7330 6 : // Should still be able to renew a currently valid lease
7331 6 : // Assumption: original lease to is still valid for 0/50.
7332 6 : let _ =
7333 6 : timeline.make_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)?;
7334 6 :
7335 6 : Ok(())
7336 6 : }
7337 :
7338 : #[tokio::test]
7339 6 : async fn test_simple_bottom_most_compaction_deltas() -> anyhow::Result<()> {
7340 6 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_deltas").await?;
7341 24 : let (tenant, ctx) = harness.load().await;
7342 6 :
7343 354 : fn get_key(id: u32) -> Key {
7344 354 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7345 354 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7346 354 : key.field6 = id;
7347 354 : key
7348 354 : }
7349 6 :
7350 6 : // We create
7351 6 : // - one bottom-most image layer,
7352 6 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
7353 6 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
7354 6 : // - a delta layer D3 above the horizon.
7355 6 : //
7356 6 : // | D3 |
7357 6 : // | D1 |
7358 6 : // -| |-- gc horizon -----------------
7359 6 : // | | | D2 |
7360 6 : // --------- img layer ------------------
7361 6 : //
7362 6 : // What we should expact from this compaction is:
7363 6 : // | D3 |
7364 6 : // | Part of D1 |
7365 6 : // --------- img layer with D1+D2 at GC horizon------------------
7366 6 :
7367 6 : // img layer at 0x10
7368 6 : let img_layer = (0..10)
7369 60 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
7370 6 : .collect_vec();
7371 6 :
7372 6 : let delta1 = vec![
7373 6 : (
7374 6 : get_key(1),
7375 6 : Lsn(0x20),
7376 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7377 6 : ),
7378 6 : (
7379 6 : get_key(2),
7380 6 : Lsn(0x30),
7381 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7382 6 : ),
7383 6 : (
7384 6 : get_key(3),
7385 6 : Lsn(0x28),
7386 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
7387 6 : ),
7388 6 : (
7389 6 : get_key(3),
7390 6 : Lsn(0x30),
7391 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7392 6 : ),
7393 6 : (
7394 6 : get_key(3),
7395 6 : Lsn(0x40),
7396 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
7397 6 : ),
7398 6 : ];
7399 6 : let delta2 = vec![
7400 6 : (
7401 6 : get_key(5),
7402 6 : Lsn(0x20),
7403 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7404 6 : ),
7405 6 : (
7406 6 : get_key(6),
7407 6 : Lsn(0x20),
7408 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7409 6 : ),
7410 6 : ];
7411 6 : let delta3 = vec![
7412 6 : (
7413 6 : get_key(8),
7414 6 : Lsn(0x48),
7415 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
7416 6 : ),
7417 6 : (
7418 6 : get_key(9),
7419 6 : Lsn(0x48),
7420 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
7421 6 : ),
7422 6 : ];
7423 6 :
7424 6 : let tline = tenant
7425 6 : .create_test_timeline_with_layers(
7426 6 : TIMELINE_ID,
7427 6 : Lsn(0x10),
7428 6 : DEFAULT_PG_VERSION,
7429 6 : &ctx,
7430 6 : vec![
7431 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
7432 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
7433 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
7434 6 : ], // delta layers
7435 6 : vec![(Lsn(0x10), img_layer)], // image layers
7436 6 : Lsn(0x50),
7437 6 : )
7438 141 : .await?;
7439 6 : {
7440 6 : // Update GC info
7441 6 : let mut guard = tline.gc_info.write().unwrap();
7442 6 : *guard = GcInfo {
7443 6 : retain_lsns: vec![],
7444 6 : cutoffs: GcCutoffs {
7445 6 : time: Lsn(0x30),
7446 6 : space: Lsn(0x30),
7447 6 : },
7448 6 : leases: Default::default(),
7449 6 : within_ancestor_pitr: false,
7450 6 : };
7451 6 : }
7452 6 :
7453 6 : let expected_result = [
7454 6 : Bytes::from_static(b"value 0@0x10"),
7455 6 : Bytes::from_static(b"value 1@0x10@0x20"),
7456 6 : Bytes::from_static(b"value 2@0x10@0x30"),
7457 6 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
7458 6 : Bytes::from_static(b"value 4@0x10"),
7459 6 : Bytes::from_static(b"value 5@0x10@0x20"),
7460 6 : Bytes::from_static(b"value 6@0x10@0x20"),
7461 6 : Bytes::from_static(b"value 7@0x10"),
7462 6 : Bytes::from_static(b"value 8@0x10@0x48"),
7463 6 : Bytes::from_static(b"value 9@0x10@0x48"),
7464 6 : ];
7465 6 :
7466 6 : let expected_result_at_gc_horizon = [
7467 6 : Bytes::from_static(b"value 0@0x10"),
7468 6 : Bytes::from_static(b"value 1@0x10@0x20"),
7469 6 : Bytes::from_static(b"value 2@0x10@0x30"),
7470 6 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
7471 6 : Bytes::from_static(b"value 4@0x10"),
7472 6 : Bytes::from_static(b"value 5@0x10@0x20"),
7473 6 : Bytes::from_static(b"value 6@0x10@0x20"),
7474 6 : Bytes::from_static(b"value 7@0x10"),
7475 6 : Bytes::from_static(b"value 8@0x10"),
7476 6 : Bytes::from_static(b"value 9@0x10"),
7477 6 : ];
7478 6 :
7479 66 : for idx in 0..10 {
7480 60 : assert_eq!(
7481 60 : tline
7482 60 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7483 93 : .await
7484 60 : .unwrap(),
7485 60 : &expected_result[idx]
7486 6 : );
7487 60 : assert_eq!(
7488 60 : tline
7489 60 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
7490 45 : .await
7491 60 : .unwrap(),
7492 60 : &expected_result_at_gc_horizon[idx]
7493 6 : );
7494 6 : }
7495 6 :
7496 6 : let cancel = CancellationToken::new();
7497 6 : tline
7498 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7499 156 : .await
7500 6 : .unwrap();
7501 6 :
7502 66 : for idx in 0..10 {
7503 60 : assert_eq!(
7504 60 : tline
7505 60 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7506 60 : .await
7507 60 : .unwrap(),
7508 60 : &expected_result[idx]
7509 6 : );
7510 60 : assert_eq!(
7511 60 : tline
7512 60 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
7513 30 : .await
7514 60 : .unwrap(),
7515 60 : &expected_result_at_gc_horizon[idx]
7516 6 : );
7517 6 : }
7518 6 :
7519 6 : // increase GC horizon and compact again
7520 6 : {
7521 6 : // Update GC info
7522 6 : let mut guard = tline.gc_info.write().unwrap();
7523 6 : guard.cutoffs.time = Lsn(0x40);
7524 6 : guard.cutoffs.space = Lsn(0x40);
7525 6 : }
7526 6 : tline
7527 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7528 126 : .await
7529 6 : .unwrap();
7530 6 :
7531 6 : Ok(())
7532 6 : }
7533 :
7534 : #[tokio::test]
7535 6 : async fn test_generate_key_retention() -> anyhow::Result<()> {
7536 6 : let harness = TenantHarness::create("test_generate_key_retention").await?;
7537 24 : let (tenant, ctx) = harness.load().await;
7538 6 : let tline = tenant
7539 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7540 12 : .await?;
7541 6 : tline.force_advance_lsn(Lsn(0x70));
7542 6 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7543 6 : let history = vec![
7544 6 : (
7545 6 : key,
7546 6 : Lsn(0x10),
7547 6 : Value::Image(Bytes::copy_from_slice(b"0x10")),
7548 6 : ),
7549 6 : (
7550 6 : key,
7551 6 : Lsn(0x20),
7552 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7553 6 : ),
7554 6 : (
7555 6 : key,
7556 6 : Lsn(0x30),
7557 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
7558 6 : ),
7559 6 : (
7560 6 : key,
7561 6 : Lsn(0x40),
7562 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
7563 6 : ),
7564 6 : (
7565 6 : key,
7566 6 : Lsn(0x50),
7567 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
7568 6 : ),
7569 6 : (
7570 6 : key,
7571 6 : Lsn(0x60),
7572 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7573 6 : ),
7574 6 : (
7575 6 : key,
7576 6 : Lsn(0x70),
7577 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7578 6 : ),
7579 6 : (
7580 6 : key,
7581 6 : Lsn(0x80),
7582 6 : Value::Image(Bytes::copy_from_slice(
7583 6 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
7584 6 : )),
7585 6 : ),
7586 6 : (
7587 6 : key,
7588 6 : Lsn(0x90),
7589 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
7590 6 : ),
7591 6 : ];
7592 6 : let res = tline
7593 6 : .generate_key_retention(
7594 6 : key,
7595 6 : &history,
7596 6 : Lsn(0x60),
7597 6 : &[Lsn(0x20), Lsn(0x40), Lsn(0x50)],
7598 6 : 3,
7599 6 : None,
7600 6 : )
7601 6 : .await
7602 6 : .unwrap();
7603 6 : let expected_res = KeyHistoryRetention {
7604 6 : below_horizon: vec![
7605 6 : (
7606 6 : Lsn(0x20),
7607 6 : KeyLogAtLsn(vec![(
7608 6 : Lsn(0x20),
7609 6 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20")),
7610 6 : )]),
7611 6 : ),
7612 6 : (
7613 6 : Lsn(0x40),
7614 6 : KeyLogAtLsn(vec![
7615 6 : (
7616 6 : Lsn(0x30),
7617 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
7618 6 : ),
7619 6 : (
7620 6 : Lsn(0x40),
7621 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
7622 6 : ),
7623 6 : ]),
7624 6 : ),
7625 6 : (
7626 6 : Lsn(0x50),
7627 6 : KeyLogAtLsn(vec![(
7628 6 : Lsn(0x50),
7629 6 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40;0x50")),
7630 6 : )]),
7631 6 : ),
7632 6 : (
7633 6 : Lsn(0x60),
7634 6 : KeyLogAtLsn(vec![(
7635 6 : Lsn(0x60),
7636 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7637 6 : )]),
7638 6 : ),
7639 6 : ],
7640 6 : above_horizon: KeyLogAtLsn(vec![
7641 6 : (
7642 6 : Lsn(0x70),
7643 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7644 6 : ),
7645 6 : (
7646 6 : Lsn(0x80),
7647 6 : Value::Image(Bytes::copy_from_slice(
7648 6 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
7649 6 : )),
7650 6 : ),
7651 6 : (
7652 6 : Lsn(0x90),
7653 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
7654 6 : ),
7655 6 : ]),
7656 6 : };
7657 6 : assert_eq!(res, expected_res);
7658 6 :
7659 6 : // We expect GC-compaction to run with the original GC. This would create a situation that
7660 6 : // the original GC algorithm removes some delta layers b/c there are full image coverage,
7661 6 : // therefore causing some keys to have an incomplete history below the lowest retain LSN.
7662 6 : // For example, we have
7663 6 : // ```plain
7664 6 : // init delta @ 0x10, image @ 0x20, delta @ 0x30 (gc_horizon), image @ 0x40.
7665 6 : // ```
7666 6 : // Now the GC horizon moves up, and we have
7667 6 : // ```plain
7668 6 : // init delta @ 0x10, image @ 0x20, delta @ 0x30, image @ 0x40 (gc_horizon)
7669 6 : // ```
7670 6 : // The original GC algorithm kicks in, and removes delta @ 0x10, image @ 0x20.
7671 6 : // We will end up with
7672 6 : // ```plain
7673 6 : // delta @ 0x30, image @ 0x40 (gc_horizon)
7674 6 : // ```
7675 6 : // Now we run the GC-compaction, and this key does not have a full history.
7676 6 : // We should be able to handle this partial history and drop everything before the
7677 6 : // gc_horizon image.
7678 6 :
7679 6 : let history = vec![
7680 6 : (
7681 6 : key,
7682 6 : Lsn(0x20),
7683 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7684 6 : ),
7685 6 : (
7686 6 : key,
7687 6 : Lsn(0x30),
7688 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
7689 6 : ),
7690 6 : (
7691 6 : key,
7692 6 : Lsn(0x40),
7693 6 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
7694 6 : ),
7695 6 : (
7696 6 : key,
7697 6 : Lsn(0x50),
7698 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
7699 6 : ),
7700 6 : (
7701 6 : key,
7702 6 : Lsn(0x60),
7703 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7704 6 : ),
7705 6 : (
7706 6 : key,
7707 6 : Lsn(0x70),
7708 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7709 6 : ),
7710 6 : (
7711 6 : key,
7712 6 : Lsn(0x80),
7713 6 : Value::Image(Bytes::copy_from_slice(
7714 6 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
7715 6 : )),
7716 6 : ),
7717 6 : (
7718 6 : key,
7719 6 : Lsn(0x90),
7720 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
7721 6 : ),
7722 6 : ];
7723 6 : let res = tline
7724 6 : .generate_key_retention(key, &history, Lsn(0x60), &[Lsn(0x40), Lsn(0x50)], 3, None)
7725 6 : .await
7726 6 : .unwrap();
7727 6 : let expected_res = KeyHistoryRetention {
7728 6 : below_horizon: vec![
7729 6 : (
7730 6 : Lsn(0x40),
7731 6 : KeyLogAtLsn(vec![(
7732 6 : Lsn(0x40),
7733 6 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
7734 6 : )]),
7735 6 : ),
7736 6 : (
7737 6 : Lsn(0x50),
7738 6 : KeyLogAtLsn(vec![(
7739 6 : Lsn(0x50),
7740 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
7741 6 : )]),
7742 6 : ),
7743 6 : (
7744 6 : Lsn(0x60),
7745 6 : KeyLogAtLsn(vec![(
7746 6 : Lsn(0x60),
7747 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7748 6 : )]),
7749 6 : ),
7750 6 : ],
7751 6 : above_horizon: KeyLogAtLsn(vec![
7752 6 : (
7753 6 : Lsn(0x70),
7754 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7755 6 : ),
7756 6 : (
7757 6 : Lsn(0x80),
7758 6 : Value::Image(Bytes::copy_from_slice(
7759 6 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
7760 6 : )),
7761 6 : ),
7762 6 : (
7763 6 : Lsn(0x90),
7764 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
7765 6 : ),
7766 6 : ]),
7767 6 : };
7768 6 : assert_eq!(res, expected_res);
7769 6 :
7770 6 : // In case of branch compaction, the branch itself does not have the full history, and we need to provide
7771 6 : // the ancestor image in the test case.
7772 6 :
7773 6 : let history = vec![
7774 6 : (
7775 6 : key,
7776 6 : Lsn(0x20),
7777 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7778 6 : ),
7779 6 : (
7780 6 : key,
7781 6 : Lsn(0x30),
7782 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
7783 6 : ),
7784 6 : (
7785 6 : key,
7786 6 : Lsn(0x40),
7787 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
7788 6 : ),
7789 6 : (
7790 6 : key,
7791 6 : Lsn(0x70),
7792 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7793 6 : ),
7794 6 : ];
7795 6 : let res = tline
7796 6 : .generate_key_retention(
7797 6 : key,
7798 6 : &history,
7799 6 : Lsn(0x60),
7800 6 : &[],
7801 6 : 3,
7802 6 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
7803 6 : )
7804 6 : .await
7805 6 : .unwrap();
7806 6 : let expected_res = KeyHistoryRetention {
7807 6 : below_horizon: vec![(
7808 6 : Lsn(0x60),
7809 6 : KeyLogAtLsn(vec![(
7810 6 : Lsn(0x60),
7811 6 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")), // use the ancestor image to reconstruct the page
7812 6 : )]),
7813 6 : )],
7814 6 : above_horizon: KeyLogAtLsn(vec![(
7815 6 : Lsn(0x70),
7816 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7817 6 : )]),
7818 6 : };
7819 6 : assert_eq!(res, expected_res);
7820 6 :
7821 6 : let history = vec![
7822 6 : (
7823 6 : key,
7824 6 : Lsn(0x20),
7825 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7826 6 : ),
7827 6 : (
7828 6 : key,
7829 6 : Lsn(0x40),
7830 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
7831 6 : ),
7832 6 : (
7833 6 : key,
7834 6 : Lsn(0x60),
7835 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7836 6 : ),
7837 6 : (
7838 6 : key,
7839 6 : Lsn(0x70),
7840 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7841 6 : ),
7842 6 : ];
7843 6 : let res = tline
7844 6 : .generate_key_retention(
7845 6 : key,
7846 6 : &history,
7847 6 : Lsn(0x60),
7848 6 : &[Lsn(0x30)],
7849 6 : 3,
7850 6 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
7851 6 : )
7852 6 : .await
7853 6 : .unwrap();
7854 6 : let expected_res = KeyHistoryRetention {
7855 6 : below_horizon: vec![
7856 6 : (
7857 6 : Lsn(0x30),
7858 6 : KeyLogAtLsn(vec![(
7859 6 : Lsn(0x20),
7860 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7861 6 : )]),
7862 6 : ),
7863 6 : (
7864 6 : Lsn(0x60),
7865 6 : KeyLogAtLsn(vec![(
7866 6 : Lsn(0x60),
7867 6 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x40;0x60")),
7868 6 : )]),
7869 6 : ),
7870 6 : ],
7871 6 : above_horizon: KeyLogAtLsn(vec![(
7872 6 : Lsn(0x70),
7873 6 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7874 6 : )]),
7875 6 : };
7876 6 : assert_eq!(res, expected_res);
7877 6 :
7878 6 : Ok(())
7879 6 : }
7880 :
7881 : #[tokio::test]
7882 6 : async fn test_simple_bottom_most_compaction_with_retain_lsns() -> anyhow::Result<()> {
7883 6 : let harness =
7884 6 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns").await?;
7885 24 : let (tenant, ctx) = harness.load().await;
7886 6 :
7887 1554 : fn get_key(id: u32) -> Key {
7888 1554 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7889 1554 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7890 1554 : key.field6 = id;
7891 1554 : key
7892 1554 : }
7893 6 :
7894 6 : let img_layer = (0..10)
7895 60 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
7896 6 : .collect_vec();
7897 6 :
7898 6 : let delta1 = vec![
7899 6 : (
7900 6 : get_key(1),
7901 6 : Lsn(0x20),
7902 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7903 6 : ),
7904 6 : (
7905 6 : get_key(2),
7906 6 : Lsn(0x30),
7907 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7908 6 : ),
7909 6 : (
7910 6 : get_key(3),
7911 6 : Lsn(0x28),
7912 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
7913 6 : ),
7914 6 : (
7915 6 : get_key(3),
7916 6 : Lsn(0x30),
7917 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7918 6 : ),
7919 6 : (
7920 6 : get_key(3),
7921 6 : Lsn(0x40),
7922 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
7923 6 : ),
7924 6 : ];
7925 6 : let delta2 = vec![
7926 6 : (
7927 6 : get_key(5),
7928 6 : Lsn(0x20),
7929 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7930 6 : ),
7931 6 : (
7932 6 : get_key(6),
7933 6 : Lsn(0x20),
7934 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7935 6 : ),
7936 6 : ];
7937 6 : let delta3 = vec![
7938 6 : (
7939 6 : get_key(8),
7940 6 : Lsn(0x48),
7941 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
7942 6 : ),
7943 6 : (
7944 6 : get_key(9),
7945 6 : Lsn(0x48),
7946 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
7947 6 : ),
7948 6 : ];
7949 6 :
7950 6 : let tline = tenant
7951 6 : .create_test_timeline_with_layers(
7952 6 : TIMELINE_ID,
7953 6 : Lsn(0x10),
7954 6 : DEFAULT_PG_VERSION,
7955 6 : &ctx,
7956 6 : vec![
7957 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
7958 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
7959 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
7960 6 : ], // delta layers
7961 6 : vec![(Lsn(0x10), img_layer)], // image layers
7962 6 : Lsn(0x50),
7963 6 : )
7964 141 : .await?;
7965 6 : {
7966 6 : // Update GC info
7967 6 : let mut guard = tline.gc_info.write().unwrap();
7968 6 : *guard = GcInfo {
7969 6 : retain_lsns: vec![
7970 6 : (Lsn(0x10), tline.timeline_id),
7971 6 : (Lsn(0x20), tline.timeline_id),
7972 6 : ],
7973 6 : cutoffs: GcCutoffs {
7974 6 : time: Lsn(0x30),
7975 6 : space: Lsn(0x30),
7976 6 : },
7977 6 : leases: Default::default(),
7978 6 : within_ancestor_pitr: false,
7979 6 : };
7980 6 : }
7981 6 :
7982 6 : let expected_result = [
7983 6 : Bytes::from_static(b"value 0@0x10"),
7984 6 : Bytes::from_static(b"value 1@0x10@0x20"),
7985 6 : Bytes::from_static(b"value 2@0x10@0x30"),
7986 6 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
7987 6 : Bytes::from_static(b"value 4@0x10"),
7988 6 : Bytes::from_static(b"value 5@0x10@0x20"),
7989 6 : Bytes::from_static(b"value 6@0x10@0x20"),
7990 6 : Bytes::from_static(b"value 7@0x10"),
7991 6 : Bytes::from_static(b"value 8@0x10@0x48"),
7992 6 : Bytes::from_static(b"value 9@0x10@0x48"),
7993 6 : ];
7994 6 :
7995 6 : let expected_result_at_gc_horizon = [
7996 6 : Bytes::from_static(b"value 0@0x10"),
7997 6 : Bytes::from_static(b"value 1@0x10@0x20"),
7998 6 : Bytes::from_static(b"value 2@0x10@0x30"),
7999 6 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
8000 6 : Bytes::from_static(b"value 4@0x10"),
8001 6 : Bytes::from_static(b"value 5@0x10@0x20"),
8002 6 : Bytes::from_static(b"value 6@0x10@0x20"),
8003 6 : Bytes::from_static(b"value 7@0x10"),
8004 6 : Bytes::from_static(b"value 8@0x10"),
8005 6 : Bytes::from_static(b"value 9@0x10"),
8006 6 : ];
8007 6 :
8008 6 : let expected_result_at_lsn_20 = [
8009 6 : Bytes::from_static(b"value 0@0x10"),
8010 6 : Bytes::from_static(b"value 1@0x10@0x20"),
8011 6 : Bytes::from_static(b"value 2@0x10"),
8012 6 : Bytes::from_static(b"value 3@0x10"),
8013 6 : Bytes::from_static(b"value 4@0x10"),
8014 6 : Bytes::from_static(b"value 5@0x10@0x20"),
8015 6 : Bytes::from_static(b"value 6@0x10@0x20"),
8016 6 : Bytes::from_static(b"value 7@0x10"),
8017 6 : Bytes::from_static(b"value 8@0x10"),
8018 6 : Bytes::from_static(b"value 9@0x10"),
8019 6 : ];
8020 6 :
8021 6 : let expected_result_at_lsn_10 = [
8022 6 : Bytes::from_static(b"value 0@0x10"),
8023 6 : Bytes::from_static(b"value 1@0x10"),
8024 6 : Bytes::from_static(b"value 2@0x10"),
8025 6 : Bytes::from_static(b"value 3@0x10"),
8026 6 : Bytes::from_static(b"value 4@0x10"),
8027 6 : Bytes::from_static(b"value 5@0x10"),
8028 6 : Bytes::from_static(b"value 6@0x10"),
8029 6 : Bytes::from_static(b"value 7@0x10"),
8030 6 : Bytes::from_static(b"value 8@0x10"),
8031 6 : Bytes::from_static(b"value 9@0x10"),
8032 6 : ];
8033 6 :
8034 36 : let verify_result = || async {
8035 36 : let gc_horizon = {
8036 36 : let gc_info = tline.gc_info.read().unwrap();
8037 36 : gc_info.cutoffs.time
8038 6 : };
8039 396 : for idx in 0..10 {
8040 360 : assert_eq!(
8041 360 : tline
8042 360 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8043 372 : .await
8044 360 : .unwrap(),
8045 360 : &expected_result[idx]
8046 6 : );
8047 360 : assert_eq!(
8048 360 : tline
8049 360 : .get(get_key(idx as u32), gc_horizon, &ctx)
8050 288 : .await
8051 360 : .unwrap(),
8052 360 : &expected_result_at_gc_horizon[idx]
8053 6 : );
8054 360 : assert_eq!(
8055 360 : tline
8056 360 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
8057 243 : .await
8058 360 : .unwrap(),
8059 360 : &expected_result_at_lsn_20[idx]
8060 6 : );
8061 360 : assert_eq!(
8062 360 : tline
8063 360 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
8064 180 : .await
8065 360 : .unwrap(),
8066 360 : &expected_result_at_lsn_10[idx]
8067 6 : );
8068 6 : }
8069 72 : };
8070 6 :
8071 207 : verify_result().await;
8072 6 :
8073 6 : let cancel = CancellationToken::new();
8074 6 : let mut dryrun_flags = EnumSet::new();
8075 6 : dryrun_flags.insert(CompactFlags::DryRun);
8076 6 :
8077 6 : tline
8078 6 : .compact_with_gc(&cancel, dryrun_flags, &ctx)
8079 132 : .await
8080 6 : .unwrap();
8081 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
8082 6 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
8083 171 : verify_result().await;
8084 6 :
8085 6 : tline
8086 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8087 150 : .await
8088 6 : .unwrap();
8089 192 : verify_result().await;
8090 6 :
8091 6 : // compact again
8092 6 : tline
8093 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8094 114 : .await
8095 6 : .unwrap();
8096 171 : verify_result().await;
8097 6 :
8098 6 : // increase GC horizon and compact again
8099 6 : {
8100 6 : // Update GC info
8101 6 : let mut guard = tline.gc_info.write().unwrap();
8102 6 : guard.cutoffs.time = Lsn(0x38);
8103 6 : guard.cutoffs.space = Lsn(0x38);
8104 6 : }
8105 6 : tline
8106 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8107 117 : .await
8108 6 : .unwrap();
8109 171 : verify_result().await; // no wals between 0x30 and 0x38, so we should obtain the same result
8110 6 :
8111 6 : // not increasing the GC horizon and compact again
8112 6 : tline
8113 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8114 117 : .await
8115 6 : .unwrap();
8116 171 : verify_result().await;
8117 6 :
8118 6 : Ok(())
8119 6 : }
8120 :
8121 : #[tokio::test]
8122 6 : async fn test_simple_bottom_most_compaction_with_retain_lsns_single_key() -> anyhow::Result<()>
8123 6 : {
8124 6 : let harness =
8125 6 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns_single_key")
8126 6 : .await?;
8127 24 : let (tenant, ctx) = harness.load().await;
8128 6 :
8129 1056 : fn get_key(id: u32) -> Key {
8130 1056 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8131 1056 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8132 1056 : key.field6 = id;
8133 1056 : key
8134 1056 : }
8135 6 :
8136 6 : let img_layer = (0..10)
8137 60 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8138 6 : .collect_vec();
8139 6 :
8140 6 : let delta1 = vec![
8141 6 : (
8142 6 : get_key(1),
8143 6 : Lsn(0x20),
8144 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8145 6 : ),
8146 6 : (
8147 6 : get_key(1),
8148 6 : Lsn(0x28),
8149 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8150 6 : ),
8151 6 : ];
8152 6 : let delta2 = vec![
8153 6 : (
8154 6 : get_key(1),
8155 6 : Lsn(0x30),
8156 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8157 6 : ),
8158 6 : (
8159 6 : get_key(1),
8160 6 : Lsn(0x38),
8161 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
8162 6 : ),
8163 6 : ];
8164 6 : let delta3 = vec![
8165 6 : (
8166 6 : get_key(8),
8167 6 : Lsn(0x48),
8168 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8169 6 : ),
8170 6 : (
8171 6 : get_key(9),
8172 6 : Lsn(0x48),
8173 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8174 6 : ),
8175 6 : ];
8176 6 :
8177 6 : let tline = tenant
8178 6 : .create_test_timeline_with_layers(
8179 6 : TIMELINE_ID,
8180 6 : Lsn(0x10),
8181 6 : DEFAULT_PG_VERSION,
8182 6 : &ctx,
8183 6 : vec![
8184 6 : // delta1 and delta 2 only contain a single key but multiple updates
8185 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x30), delta1),
8186 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
8187 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x50), delta3),
8188 6 : ], // delta layers
8189 6 : vec![(Lsn(0x10), img_layer)], // image layers
8190 6 : Lsn(0x50),
8191 6 : )
8192 141 : .await?;
8193 6 : {
8194 6 : // Update GC info
8195 6 : let mut guard = tline.gc_info.write().unwrap();
8196 6 : *guard = GcInfo {
8197 6 : retain_lsns: vec![
8198 6 : (Lsn(0x10), tline.timeline_id),
8199 6 : (Lsn(0x20), tline.timeline_id),
8200 6 : ],
8201 6 : cutoffs: GcCutoffs {
8202 6 : time: Lsn(0x30),
8203 6 : space: Lsn(0x30),
8204 6 : },
8205 6 : leases: Default::default(),
8206 6 : within_ancestor_pitr: false,
8207 6 : };
8208 6 : }
8209 6 :
8210 6 : let expected_result = [
8211 6 : Bytes::from_static(b"value 0@0x10"),
8212 6 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
8213 6 : Bytes::from_static(b"value 2@0x10"),
8214 6 : Bytes::from_static(b"value 3@0x10"),
8215 6 : Bytes::from_static(b"value 4@0x10"),
8216 6 : Bytes::from_static(b"value 5@0x10"),
8217 6 : Bytes::from_static(b"value 6@0x10"),
8218 6 : Bytes::from_static(b"value 7@0x10"),
8219 6 : Bytes::from_static(b"value 8@0x10@0x48"),
8220 6 : Bytes::from_static(b"value 9@0x10@0x48"),
8221 6 : ];
8222 6 :
8223 6 : let expected_result_at_gc_horizon = [
8224 6 : Bytes::from_static(b"value 0@0x10"),
8225 6 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
8226 6 : Bytes::from_static(b"value 2@0x10"),
8227 6 : Bytes::from_static(b"value 3@0x10"),
8228 6 : Bytes::from_static(b"value 4@0x10"),
8229 6 : Bytes::from_static(b"value 5@0x10"),
8230 6 : Bytes::from_static(b"value 6@0x10"),
8231 6 : Bytes::from_static(b"value 7@0x10"),
8232 6 : Bytes::from_static(b"value 8@0x10"),
8233 6 : Bytes::from_static(b"value 9@0x10"),
8234 6 : ];
8235 6 :
8236 6 : let expected_result_at_lsn_20 = [
8237 6 : Bytes::from_static(b"value 0@0x10"),
8238 6 : Bytes::from_static(b"value 1@0x10@0x20"),
8239 6 : Bytes::from_static(b"value 2@0x10"),
8240 6 : Bytes::from_static(b"value 3@0x10"),
8241 6 : Bytes::from_static(b"value 4@0x10"),
8242 6 : Bytes::from_static(b"value 5@0x10"),
8243 6 : Bytes::from_static(b"value 6@0x10"),
8244 6 : Bytes::from_static(b"value 7@0x10"),
8245 6 : Bytes::from_static(b"value 8@0x10"),
8246 6 : Bytes::from_static(b"value 9@0x10"),
8247 6 : ];
8248 6 :
8249 6 : let expected_result_at_lsn_10 = [
8250 6 : Bytes::from_static(b"value 0@0x10"),
8251 6 : Bytes::from_static(b"value 1@0x10"),
8252 6 : Bytes::from_static(b"value 2@0x10"),
8253 6 : Bytes::from_static(b"value 3@0x10"),
8254 6 : Bytes::from_static(b"value 4@0x10"),
8255 6 : Bytes::from_static(b"value 5@0x10"),
8256 6 : Bytes::from_static(b"value 6@0x10"),
8257 6 : Bytes::from_static(b"value 7@0x10"),
8258 6 : Bytes::from_static(b"value 8@0x10"),
8259 6 : Bytes::from_static(b"value 9@0x10"),
8260 6 : ];
8261 6 :
8262 24 : let verify_result = || async {
8263 24 : let gc_horizon = {
8264 24 : let gc_info = tline.gc_info.read().unwrap();
8265 24 : gc_info.cutoffs.time
8266 6 : };
8267 264 : for idx in 0..10 {
8268 240 : assert_eq!(
8269 240 : tline
8270 240 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8271 222 : .await
8272 240 : .unwrap(),
8273 240 : &expected_result[idx]
8274 6 : );
8275 240 : assert_eq!(
8276 240 : tline
8277 240 : .get(get_key(idx as u32), gc_horizon, &ctx)
8278 147 : .await
8279 240 : .unwrap(),
8280 240 : &expected_result_at_gc_horizon[idx]
8281 6 : );
8282 240 : assert_eq!(
8283 240 : tline
8284 240 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
8285 138 : .await
8286 240 : .unwrap(),
8287 240 : &expected_result_at_lsn_20[idx]
8288 6 : );
8289 240 : assert_eq!(
8290 240 : tline
8291 240 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
8292 123 : .await
8293 240 : .unwrap(),
8294 240 : &expected_result_at_lsn_10[idx]
8295 6 : );
8296 6 : }
8297 48 : };
8298 6 :
8299 183 : verify_result().await;
8300 6 :
8301 6 : let cancel = CancellationToken::new();
8302 6 : let mut dryrun_flags = EnumSet::new();
8303 6 : dryrun_flags.insert(CompactFlags::DryRun);
8304 6 :
8305 6 : tline
8306 6 : .compact_with_gc(&cancel, dryrun_flags, &ctx)
8307 135 : .await
8308 6 : .unwrap();
8309 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
8310 6 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
8311 147 : verify_result().await;
8312 6 :
8313 6 : tline
8314 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8315 153 : .await
8316 6 : .unwrap();
8317 159 : verify_result().await;
8318 6 :
8319 6 : // compact again
8320 6 : tline
8321 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8322 117 : .await
8323 6 : .unwrap();
8324 141 : verify_result().await;
8325 6 :
8326 6 : Ok(())
8327 6 : }
8328 :
8329 : #[tokio::test]
8330 6 : async fn test_simple_bottom_most_compaction_on_branch() -> anyhow::Result<()> {
8331 6 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_on_branch").await?;
8332 24 : let (tenant, ctx) = harness.load().await;
8333 6 :
8334 378 : fn get_key(id: u32) -> Key {
8335 378 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
8336 378 : key.field6 = id;
8337 378 : key
8338 378 : }
8339 6 :
8340 6 : let img_layer = (0..10)
8341 60 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8342 6 : .collect_vec();
8343 6 :
8344 6 : let delta1 = vec![
8345 6 : (
8346 6 : get_key(1),
8347 6 : Lsn(0x20),
8348 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8349 6 : ),
8350 6 : (
8351 6 : get_key(2),
8352 6 : Lsn(0x30),
8353 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8354 6 : ),
8355 6 : (
8356 6 : get_key(3),
8357 6 : Lsn(0x28),
8358 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8359 6 : ),
8360 6 : (
8361 6 : get_key(3),
8362 6 : Lsn(0x30),
8363 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8364 6 : ),
8365 6 : (
8366 6 : get_key(3),
8367 6 : Lsn(0x40),
8368 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
8369 6 : ),
8370 6 : ];
8371 6 : let delta2 = vec![
8372 6 : (
8373 6 : get_key(5),
8374 6 : Lsn(0x20),
8375 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8376 6 : ),
8377 6 : (
8378 6 : get_key(6),
8379 6 : Lsn(0x20),
8380 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8381 6 : ),
8382 6 : ];
8383 6 : let delta3 = vec![
8384 6 : (
8385 6 : get_key(8),
8386 6 : Lsn(0x48),
8387 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8388 6 : ),
8389 6 : (
8390 6 : get_key(9),
8391 6 : Lsn(0x48),
8392 6 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8393 6 : ),
8394 6 : ];
8395 6 :
8396 6 : let parent_tline = tenant
8397 6 : .create_test_timeline_with_layers(
8398 6 : TIMELINE_ID,
8399 6 : Lsn(0x10),
8400 6 : DEFAULT_PG_VERSION,
8401 6 : &ctx,
8402 6 : vec![], // delta layers
8403 6 : vec![(Lsn(0x18), img_layer)], // image layers
8404 6 : Lsn(0x18),
8405 6 : )
8406 87 : .await?;
8407 6 :
8408 6 : parent_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
8409 6 :
8410 6 : let branch_tline = tenant
8411 6 : .branch_timeline_test_with_layers(
8412 6 : &parent_tline,
8413 6 : NEW_TIMELINE_ID,
8414 6 : Some(Lsn(0x18)),
8415 6 : &ctx,
8416 6 : vec![
8417 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
8418 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
8419 6 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
8420 6 : ], // delta layers
8421 6 : vec![], // image layers
8422 6 : Lsn(0x50),
8423 6 : )
8424 54 : .await?;
8425 6 :
8426 6 : branch_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
8427 6 :
8428 6 : {
8429 6 : // Update GC info
8430 6 : let mut guard = parent_tline.gc_info.write().unwrap();
8431 6 : *guard = GcInfo {
8432 6 : retain_lsns: vec![(Lsn(0x18), branch_tline.timeline_id)],
8433 6 : cutoffs: GcCutoffs {
8434 6 : time: Lsn(0x10),
8435 6 : space: Lsn(0x10),
8436 6 : },
8437 6 : leases: Default::default(),
8438 6 : within_ancestor_pitr: false,
8439 6 : };
8440 6 : }
8441 6 :
8442 6 : {
8443 6 : // Update GC info
8444 6 : let mut guard = branch_tline.gc_info.write().unwrap();
8445 6 : *guard = GcInfo {
8446 6 : retain_lsns: vec![(Lsn(0x40), branch_tline.timeline_id)],
8447 6 : cutoffs: GcCutoffs {
8448 6 : time: Lsn(0x50),
8449 6 : space: Lsn(0x50),
8450 6 : },
8451 6 : leases: Default::default(),
8452 6 : within_ancestor_pitr: false,
8453 6 : };
8454 6 : }
8455 6 :
8456 6 : let expected_result_at_gc_horizon = [
8457 6 : Bytes::from_static(b"value 0@0x10"),
8458 6 : Bytes::from_static(b"value 1@0x10@0x20"),
8459 6 : Bytes::from_static(b"value 2@0x10@0x30"),
8460 6 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
8461 6 : Bytes::from_static(b"value 4@0x10"),
8462 6 : Bytes::from_static(b"value 5@0x10@0x20"),
8463 6 : Bytes::from_static(b"value 6@0x10@0x20"),
8464 6 : Bytes::from_static(b"value 7@0x10"),
8465 6 : Bytes::from_static(b"value 8@0x10@0x48"),
8466 6 : Bytes::from_static(b"value 9@0x10@0x48"),
8467 6 : ];
8468 6 :
8469 6 : let expected_result_at_lsn_40 = [
8470 6 : Bytes::from_static(b"value 0@0x10"),
8471 6 : Bytes::from_static(b"value 1@0x10@0x20"),
8472 6 : Bytes::from_static(b"value 2@0x10@0x30"),
8473 6 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
8474 6 : Bytes::from_static(b"value 4@0x10"),
8475 6 : Bytes::from_static(b"value 5@0x10@0x20"),
8476 6 : Bytes::from_static(b"value 6@0x10@0x20"),
8477 6 : Bytes::from_static(b"value 7@0x10"),
8478 6 : Bytes::from_static(b"value 8@0x10"),
8479 6 : Bytes::from_static(b"value 9@0x10"),
8480 6 : ];
8481 6 :
8482 12 : let verify_result = || async {
8483 132 : for idx in 0..10 {
8484 120 : assert_eq!(
8485 120 : branch_tline
8486 120 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8487 150 : .await
8488 120 : .unwrap(),
8489 120 : &expected_result_at_gc_horizon[idx]
8490 6 : );
8491 120 : assert_eq!(
8492 120 : branch_tline
8493 120 : .get(get_key(idx as u32), Lsn(0x40), &ctx)
8494 99 : .await
8495 120 : .unwrap(),
8496 120 : &expected_result_at_lsn_40[idx]
8497 6 : );
8498 6 : }
8499 24 : };
8500 6 :
8501 138 : verify_result().await;
8502 6 :
8503 6 : let cancel = CancellationToken::new();
8504 6 : branch_tline
8505 6 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8506 48 : .await
8507 6 : .unwrap();
8508 6 :
8509 111 : verify_result().await;
8510 6 :
8511 6 : Ok(())
8512 6 : }
8513 :
8514 : // Regression test for https://github.com/neondatabase/neon/issues/9012
8515 : // Create an image arrangement where we have to read at different LSN ranges
8516 : // from a delta layer. This is achieved by overlapping an image layer on top of
8517 : // a delta layer. Like so:
8518 : //
8519 : // A B
8520 : // +----------------+ -> delta_layer
8521 : // | | ^ lsn
8522 : // | =========|-> nested_image_layer |
8523 : // | C | |
8524 : // +----------------+ |
8525 : // ======== -> baseline_image_layer +-------> key
8526 : //
8527 : //
8528 : // When querying the key range [A, B) we need to read at different LSN ranges
8529 : // for [A, C) and [C, B). This test checks that the described edge case is handled correctly.
8530 : #[tokio::test]
8531 6 : async fn test_vectored_read_with_nested_image_layer() -> anyhow::Result<()> {
8532 6 : let harness = TenantHarness::create("test_vectored_read_with_nested_image_layer").await?;
8533 24 : let (tenant, ctx) = harness.load().await;
8534 6 :
8535 6 : let will_init_keys = [2, 6];
8536 132 : fn get_key(id: u32) -> Key {
8537 132 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
8538 132 : key.field6 = id;
8539 132 : key
8540 132 : }
8541 6 :
8542 6 : let mut expected_key_values = HashMap::new();
8543 6 :
8544 6 : let baseline_image_layer_lsn = Lsn(0x10);
8545 6 : let mut baseline_img_layer = Vec::new();
8546 36 : for i in 0..5 {
8547 30 : let key = get_key(i);
8548 30 : let value = format!("value {i}@{baseline_image_layer_lsn}");
8549 30 :
8550 30 : let removed = expected_key_values.insert(key, value.clone());
8551 30 : assert!(removed.is_none());
8552 6 :
8553 30 : baseline_img_layer.push((key, Bytes::from(value)));
8554 6 : }
8555 6 :
8556 6 : let nested_image_layer_lsn = Lsn(0x50);
8557 6 : let mut nested_img_layer = Vec::new();
8558 36 : for i in 5..10 {
8559 30 : let key = get_key(i);
8560 30 : let value = format!("value {i}@{nested_image_layer_lsn}");
8561 30 :
8562 30 : let removed = expected_key_values.insert(key, value.clone());
8563 30 : assert!(removed.is_none());
8564 6 :
8565 30 : nested_img_layer.push((key, Bytes::from(value)));
8566 6 : }
8567 6 :
8568 6 : let mut delta_layer_spec = Vec::default();
8569 6 : let delta_layer_start_lsn = Lsn(0x20);
8570 6 : let mut delta_layer_end_lsn = delta_layer_start_lsn;
8571 6 :
8572 66 : for i in 0..10 {
8573 60 : let key = get_key(i);
8574 60 : let key_in_nested = nested_img_layer
8575 60 : .iter()
8576 240 : .any(|(key_with_img, _)| *key_with_img == key);
8577 60 : let lsn = {
8578 60 : if key_in_nested {
8579 30 : Lsn(nested_image_layer_lsn.0 + 0x10)
8580 6 : } else {
8581 30 : delta_layer_start_lsn
8582 6 : }
8583 6 : };
8584 6 :
8585 60 : let will_init = will_init_keys.contains(&i);
8586 60 : if will_init {
8587 12 : delta_layer_spec.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init())));
8588 12 :
8589 12 : expected_key_values.insert(key, "".to_string());
8590 48 : } else {
8591 48 : let delta = format!("@{lsn}");
8592 48 : delta_layer_spec.push((
8593 48 : key,
8594 48 : lsn,
8595 48 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
8596 48 : ));
8597 48 :
8598 48 : expected_key_values
8599 48 : .get_mut(&key)
8600 48 : .expect("An image exists for each key")
8601 48 : .push_str(delta.as_str());
8602 48 : }
8603 60 : delta_layer_end_lsn = std::cmp::max(delta_layer_start_lsn, lsn);
8604 6 : }
8605 6 :
8606 6 : delta_layer_end_lsn = Lsn(delta_layer_end_lsn.0 + 1);
8607 6 :
8608 6 : assert!(
8609 6 : nested_image_layer_lsn > delta_layer_start_lsn
8610 6 : && nested_image_layer_lsn < delta_layer_end_lsn
8611 6 : );
8612 6 :
8613 6 : let tline = tenant
8614 6 : .create_test_timeline_with_layers(
8615 6 : TIMELINE_ID,
8616 6 : baseline_image_layer_lsn,
8617 6 : DEFAULT_PG_VERSION,
8618 6 : &ctx,
8619 6 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
8620 6 : delta_layer_start_lsn..delta_layer_end_lsn,
8621 6 : delta_layer_spec,
8622 6 : )], // delta layers
8623 6 : vec![
8624 6 : (baseline_image_layer_lsn, baseline_img_layer),
8625 6 : (nested_image_layer_lsn, nested_img_layer),
8626 6 : ], // image layers
8627 6 : delta_layer_end_lsn,
8628 6 : )
8629 120 : .await?;
8630 6 :
8631 6 : let keyspace = KeySpace::single(get_key(0)..get_key(10));
8632 6 : let results = tline
8633 6 : .get_vectored(keyspace, delta_layer_end_lsn, &ctx)
8634 43 : .await
8635 6 : .expect("No vectored errors");
8636 66 : for (key, res) in results {
8637 60 : let value = res.expect("No key errors");
8638 60 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
8639 60 : assert_eq!(value, Bytes::from(expected_value));
8640 6 : }
8641 6 :
8642 6 : Ok(())
8643 6 : }
8644 : }
|
