Line data Source code
1 : //!
2 : //! Timeline repository implementation that keeps old data in files on disk, and
3 : //! the recent changes in memory. See tenant/*_layer.rs files.
4 : //! The functions here are responsible for locating the correct layer for the
5 : //! get/put call, walking back the timeline branching history as needed.
6 : //!
7 : //! The files are stored in the .neon/tenants/<tenant_id>/timelines/<timeline_id>
8 : //! directory. See docs/pageserver-storage.md for how the files are managed.
9 : //! In addition to the layer files, there is a metadata file in the same
10 : //! directory that contains information about the timeline, in particular its
11 : //! parent timeline, and the last LSN that has been written to disk.
12 : //!
13 :
14 : use anyhow::{bail, Context};
15 : use arc_swap::ArcSwap;
16 : use camino::Utf8Path;
17 : use camino::Utf8PathBuf;
18 : use enumset::EnumSet;
19 : use futures::stream::FuturesUnordered;
20 : use futures::FutureExt;
21 : use futures::StreamExt;
22 : use pageserver_api::models;
23 : use pageserver_api::models::AuxFilePolicy;
24 : use pageserver_api::models::TimelineArchivalState;
25 : use pageserver_api::models::TimelineState;
26 : use pageserver_api::models::TopTenantShardItem;
27 : use pageserver_api::models::WalRedoManagerStatus;
28 : use pageserver_api::shard::ShardIdentity;
29 : use pageserver_api::shard::ShardStripeSize;
30 : use pageserver_api::shard::TenantShardId;
31 : use remote_storage::DownloadError;
32 : use remote_storage::GenericRemoteStorage;
33 : use remote_storage::TimeoutOrCancel;
34 : use std::collections::BTreeMap;
35 : use std::fmt;
36 : use std::sync::Weak;
37 : use std::time::SystemTime;
38 : use storage_broker::BrokerClientChannel;
39 : use tokio::io::BufReader;
40 : use tokio::sync::watch;
41 : use tokio::task::JoinSet;
42 : use tokio_util::sync::CancellationToken;
43 : use tracing::*;
44 : use utils::backoff;
45 : use utils::circuit_breaker::CircuitBreaker;
46 : use utils::completion;
47 : use utils::crashsafe::path_with_suffix_extension;
48 : use utils::failpoint_support;
49 : use utils::fs_ext;
50 : use utils::pausable_failpoint;
51 : use utils::sync::gate::Gate;
52 : use utils::sync::gate::GateGuard;
53 : use utils::timeout::timeout_cancellable;
54 : use utils::timeout::TimeoutCancellableError;
55 : use utils::zstd::create_zst_tarball;
56 : use utils::zstd::extract_zst_tarball;
57 :
58 : use self::config::AttachedLocationConfig;
59 : use self::config::AttachmentMode;
60 : use self::config::LocationConf;
61 : use self::config::TenantConf;
62 : use self::metadata::TimelineMetadata;
63 : use self::mgr::GetActiveTenantError;
64 : use self::mgr::GetTenantError;
65 : use self::remote_timeline_client::upload::upload_index_part;
66 : use self::remote_timeline_client::RemoteTimelineClient;
67 : use self::timeline::uninit::TimelineCreateGuard;
68 : use self::timeline::uninit::TimelineExclusionError;
69 : use self::timeline::uninit::UninitializedTimeline;
70 : use self::timeline::EvictionTaskTenantState;
71 : use self::timeline::GcCutoffs;
72 : use self::timeline::TimelineResources;
73 : use self::timeline::WaitLsnError;
74 : use crate::config::PageServerConf;
75 : use crate::context::{DownloadBehavior, RequestContext};
76 : use crate::deletion_queue::DeletionQueueClient;
77 : use crate::deletion_queue::DeletionQueueError;
78 : use crate::import_datadir;
79 : use crate::is_uninit_mark;
80 : use crate::l0_flush::L0FlushGlobalState;
81 : use crate::metrics::TENANT;
82 : use crate::metrics::{
83 : remove_tenant_metrics, BROKEN_TENANTS_SET, CIRCUIT_BREAKERS_BROKEN, CIRCUIT_BREAKERS_UNBROKEN,
84 : TENANT_STATE_METRIC, TENANT_SYNTHETIC_SIZE_METRIC,
85 : };
86 : use crate::repository::GcResult;
87 : use crate::task_mgr;
88 : use crate::task_mgr::TaskKind;
89 : use crate::tenant::config::LocationMode;
90 : use crate::tenant::config::TenantConfOpt;
91 : pub use crate::tenant::remote_timeline_client::index::IndexPart;
92 : use crate::tenant::remote_timeline_client::remote_initdb_archive_path;
93 : use crate::tenant::remote_timeline_client::MaybeDeletedIndexPart;
94 : use crate::tenant::remote_timeline_client::INITDB_PATH;
95 : use crate::tenant::storage_layer::DeltaLayer;
96 : use crate::tenant::storage_layer::ImageLayer;
97 : use crate::walredo;
98 : use crate::InitializationOrder;
99 : use std::collections::hash_map::Entry;
100 : use std::collections::HashMap;
101 : use std::collections::HashSet;
102 : use std::fmt::Debug;
103 : use std::fmt::Display;
104 : use std::fs;
105 : use std::fs::File;
106 : use std::sync::atomic::{AtomicU64, Ordering};
107 : use std::sync::Arc;
108 : use std::sync::Mutex;
109 : use std::time::{Duration, Instant};
110 :
111 : use crate::span;
112 : use crate::tenant::timeline::delete::DeleteTimelineFlow;
113 : use crate::tenant::timeline::uninit::cleanup_timeline_directory;
114 : use crate::virtual_file::VirtualFile;
115 : use crate::walredo::PostgresRedoManager;
116 : use crate::TEMP_FILE_SUFFIX;
117 : use once_cell::sync::Lazy;
118 : pub use pageserver_api::models::TenantState;
119 : use tokio::sync::Semaphore;
120 :
121 0 : static INIT_DB_SEMAPHORE: Lazy<Semaphore> = Lazy::new(|| Semaphore::new(8));
122 : use utils::{
123 : crashsafe,
124 : generation::Generation,
125 : id::TimelineId,
126 : lsn::{Lsn, RecordLsn},
127 : };
128 :
129 : pub mod blob_io;
130 : pub mod block_io;
131 : pub mod vectored_blob_io;
132 :
133 : pub mod disk_btree;
134 : pub(crate) mod ephemeral_file;
135 : pub mod layer_map;
136 :
137 : pub mod metadata;
138 : pub mod remote_timeline_client;
139 : pub mod storage_layer;
140 :
141 : pub mod config;
142 : pub mod mgr;
143 : pub mod secondary;
144 : pub mod tasks;
145 : pub mod upload_queue;
146 :
147 : pub(crate) mod timeline;
148 :
149 : pub mod size;
150 :
151 : pub(crate) mod throttle;
152 :
153 : pub(crate) use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
154 : pub(crate) use timeline::{LogicalSizeCalculationCause, PageReconstructError, Timeline};
155 :
156 : // re-export for use in walreceiver
157 : pub use crate::tenant::timeline::WalReceiverInfo;
158 :
159 : /// The "tenants" part of `tenants/<tenant>/timelines...`
160 : pub const TENANTS_SEGMENT_NAME: &str = "tenants";
161 :
162 : /// Parts of the `.neon/tenants/<tenant_id>/timelines/<timeline_id>` directory prefix.
163 : pub const TIMELINES_SEGMENT_NAME: &str = "timelines";
164 :
165 : /// References to shared objects that are passed into each tenant, such
166 : /// as the shared remote storage client and process initialization state.
167 : #[derive(Clone)]
168 : pub struct TenantSharedResources {
169 : pub broker_client: storage_broker::BrokerClientChannel,
170 : pub remote_storage: GenericRemoteStorage,
171 : pub deletion_queue_client: DeletionQueueClient,
172 : pub l0_flush_global_state: L0FlushGlobalState,
173 : }
174 :
175 : /// A [`Tenant`] is really an _attached_ tenant. The configuration
176 : /// for an attached tenant is a subset of the [`LocationConf`], represented
177 : /// in this struct.
178 : pub(super) struct AttachedTenantConf {
179 : tenant_conf: TenantConfOpt,
180 : location: AttachedLocationConfig,
181 : }
182 :
183 : impl AttachedTenantConf {
184 0 : fn new(tenant_conf: TenantConfOpt, location: AttachedLocationConfig) -> Self {
185 0 : Self {
186 0 : tenant_conf,
187 0 : location,
188 0 : }
189 0 : }
190 :
191 182 : fn try_from(location_conf: LocationConf) -> anyhow::Result<Self> {
192 182 : match &location_conf.mode {
193 182 : LocationMode::Attached(attach_conf) => Ok(Self {
194 182 : tenant_conf: location_conf.tenant_conf,
195 182 : location: *attach_conf,
196 182 : }),
197 : LocationMode::Secondary(_) => {
198 0 : anyhow::bail!("Attempted to construct AttachedTenantConf from a LocationConf in secondary mode")
199 : }
200 : }
201 182 : }
202 : }
203 : struct TimelinePreload {
204 : timeline_id: TimelineId,
205 : client: RemoteTimelineClient,
206 : index_part: Result<MaybeDeletedIndexPart, DownloadError>,
207 : }
208 :
209 : pub(crate) struct TenantPreload {
210 : timelines: HashMap<TimelineId, TimelinePreload>,
211 : }
212 :
213 : /// When we spawn a tenant, there is a special mode for tenant creation that
214 : /// avoids trying to read anything from remote storage.
215 : pub(crate) enum SpawnMode {
216 : /// Activate as soon as possible
217 : Eager,
218 : /// Lazy activation in the background, with the option to skip the queue if the need comes up
219 : Lazy,
220 : }
221 :
222 : ///
223 : /// Tenant consists of multiple timelines. Keep them in a hash table.
224 : ///
225 : pub struct Tenant {
226 : // Global pageserver config parameters
227 : pub conf: &'static PageServerConf,
228 :
229 : /// The value creation timestamp, used to measure activation delay, see:
230 : /// <https://github.com/neondatabase/neon/issues/4025>
231 : constructed_at: Instant,
232 :
233 : state: watch::Sender<TenantState>,
234 :
235 : // Overridden tenant-specific config parameters.
236 : // We keep TenantConfOpt sturct here to preserve the information
237 : // about parameters that are not set.
238 : // This is necessary to allow global config updates.
239 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
240 :
241 : tenant_shard_id: TenantShardId,
242 :
243 : // The detailed sharding information, beyond the number/count in tenant_shard_id
244 : shard_identity: ShardIdentity,
245 :
246 : /// The remote storage generation, used to protect S3 objects from split-brain.
247 : /// Does not change over the lifetime of the [`Tenant`] object.
248 : ///
249 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
250 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
251 : generation: Generation,
252 :
253 : timelines: Mutex<HashMap<TimelineId, Arc<Timeline>>>,
254 :
255 : /// During timeline creation, we first insert the TimelineId to the
256 : /// creating map, then `timelines`, then remove it from the creating map.
257 : /// **Lock order**: if acquring both, acquire`timelines` before `timelines_creating`
258 : timelines_creating: std::sync::Mutex<HashSet<TimelineId>>,
259 :
260 : // This mutex prevents creation of new timelines during GC.
261 : // Adding yet another mutex (in addition to `timelines`) is needed because holding
262 : // `timelines` mutex during all GC iteration
263 : // may block for a long time `get_timeline`, `get_timelines_state`,... and other operations
264 : // with timelines, which in turn may cause dropping replication connection, expiration of wait_for_lsn
265 : // timeout...
266 : gc_cs: tokio::sync::Mutex<()>,
267 : walredo_mgr: Option<Arc<WalRedoManager>>,
268 :
269 : // provides access to timeline data sitting in the remote storage
270 : pub(crate) remote_storage: GenericRemoteStorage,
271 :
272 : // Access to global deletion queue for when this tenant wants to schedule a deletion
273 : deletion_queue_client: DeletionQueueClient,
274 :
275 : /// Cached logical sizes updated updated on each [`Tenant::gather_size_inputs`].
276 : cached_logical_sizes: tokio::sync::Mutex<HashMap<(TimelineId, Lsn), u64>>,
277 : cached_synthetic_tenant_size: Arc<AtomicU64>,
278 :
279 : eviction_task_tenant_state: tokio::sync::Mutex<EvictionTaskTenantState>,
280 :
281 : /// Track repeated failures to compact, so that we can back off.
282 : /// Overhead of mutex is acceptable because compaction is done with a multi-second period.
283 : compaction_circuit_breaker: std::sync::Mutex<CircuitBreaker>,
284 :
285 : /// If the tenant is in Activating state, notify this to encourage it
286 : /// to proceed to Active as soon as possible, rather than waiting for lazy
287 : /// background warmup.
288 : pub(crate) activate_now_sem: tokio::sync::Semaphore,
289 :
290 : // Cancellation token fires when we have entered shutdown(). This is a parent of
291 : // Timelines' cancellation token.
292 : pub(crate) cancel: CancellationToken,
293 :
294 : // Users of the Tenant such as the page service must take this Gate to avoid
295 : // trying to use a Tenant which is shutting down.
296 : pub(crate) gate: Gate,
297 :
298 : /// Throttle applied at the top of [`Timeline::get`].
299 : /// All [`Tenant::timelines`] of a given [`Tenant`] instance share the same [`throttle::Throttle`] instance.
300 : pub(crate) timeline_get_throttle:
301 : Arc<throttle::Throttle<&'static crate::metrics::tenant_throttling::TimelineGet>>,
302 :
303 : /// An ongoing timeline detach must be checked during attempts to GC or compact a timeline.
304 : ongoing_timeline_detach: std::sync::Mutex<Option<(TimelineId, utils::completion::Barrier)>>,
305 :
306 : l0_flush_global_state: L0FlushGlobalState,
307 : }
308 :
309 : impl std::fmt::Debug for Tenant {
310 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
311 0 : write!(f, "{} ({})", self.tenant_shard_id, self.current_state())
312 0 : }
313 : }
314 :
315 : pub(crate) enum WalRedoManager {
316 : Prod(WalredoManagerId, PostgresRedoManager),
317 : #[cfg(test)]
318 : Test(harness::TestRedoManager),
319 : }
320 :
321 0 : #[derive(thiserror::Error, Debug)]
322 : #[error("pageserver is shutting down")]
323 : pub(crate) struct GlobalShutDown;
324 :
325 : impl WalRedoManager {
326 0 : pub(crate) fn new(mgr: PostgresRedoManager) -> Result<Arc<Self>, GlobalShutDown> {
327 0 : let id = WalredoManagerId::next();
328 0 : let arc = Arc::new(Self::Prod(id, mgr));
329 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
330 0 : match &mut *guard {
331 0 : Some(map) => {
332 0 : map.insert(id, Arc::downgrade(&arc));
333 0 : Ok(arc)
334 : }
335 0 : None => Err(GlobalShutDown),
336 : }
337 0 : }
338 : }
339 :
340 : impl Drop for WalRedoManager {
341 10 : fn drop(&mut self) {
342 10 : match self {
343 0 : Self::Prod(id, _) => {
344 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
345 0 : if let Some(map) = &mut *guard {
346 0 : map.remove(id).expect("new() registers, drop() unregisters");
347 0 : }
348 : }
349 : #[cfg(test)]
350 10 : Self::Test(_) => {
351 10 : // Not applicable to test redo manager
352 10 : }
353 : }
354 10 : }
355 : }
356 :
357 : /// Global registry of all walredo managers so that [`crate::shutdown_pageserver`] can shut down
358 : /// the walredo processes outside of the regular order.
359 : ///
360 : /// This is necessary to work around a systemd bug where it freezes if there are
361 : /// walredo processes left => <https://github.com/neondatabase/cloud/issues/11387>
362 : #[allow(clippy::type_complexity)]
363 : pub(crate) static WALREDO_MANAGERS: once_cell::sync::Lazy<
364 : Mutex<Option<HashMap<WalredoManagerId, Weak<WalRedoManager>>>>,
365 0 : > = once_cell::sync::Lazy::new(|| Mutex::new(Some(HashMap::new())));
366 : #[derive(PartialEq, Eq, Hash, Clone, Copy, Debug)]
367 : pub(crate) struct WalredoManagerId(u64);
368 : impl WalredoManagerId {
369 0 : pub fn next() -> Self {
370 0 : static NEXT: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(1);
371 0 : let id = NEXT.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
372 0 : if id == 0 {
373 0 : panic!("WalredoManagerId::new() returned 0, indicating wraparound, risking it's no longer unique");
374 0 : }
375 0 : Self(id)
376 0 : }
377 : }
378 :
379 : #[cfg(test)]
380 : impl From<harness::TestRedoManager> for WalRedoManager {
381 174 : fn from(mgr: harness::TestRedoManager) -> Self {
382 174 : Self::Test(mgr)
383 174 : }
384 : }
385 :
386 : impl WalRedoManager {
387 6 : pub(crate) async fn shutdown(&self) -> bool {
388 6 : match self {
389 0 : Self::Prod(_, mgr) => mgr.shutdown().await,
390 : #[cfg(test)]
391 : Self::Test(_) => {
392 : // Not applicable to test redo manager
393 6 : true
394 : }
395 : }
396 6 : }
397 :
398 0 : pub(crate) fn maybe_quiesce(&self, idle_timeout: Duration) {
399 0 : match self {
400 0 : Self::Prod(_, mgr) => mgr.maybe_quiesce(idle_timeout),
401 0 : #[cfg(test)]
402 0 : Self::Test(_) => {
403 0 : // Not applicable to test redo manager
404 0 : }
405 0 : }
406 0 : }
407 :
408 : /// # Cancel-Safety
409 : ///
410 : /// This method is cancellation-safe.
411 266 : pub async fn request_redo(
412 266 : &self,
413 266 : key: crate::repository::Key,
414 266 : lsn: Lsn,
415 266 : base_img: Option<(Lsn, bytes::Bytes)>,
416 266 : records: Vec<(Lsn, crate::walrecord::NeonWalRecord)>,
417 266 : pg_version: u32,
418 266 : ) -> Result<bytes::Bytes, walredo::Error> {
419 266 : match self {
420 0 : Self::Prod(_, mgr) => {
421 0 : mgr.request_redo(key, lsn, base_img, records, pg_version)
422 0 : .await
423 : }
424 : #[cfg(test)]
425 266 : Self::Test(mgr) => {
426 266 : mgr.request_redo(key, lsn, base_img, records, pg_version)
427 0 : .await
428 : }
429 : }
430 266 : }
431 :
432 0 : pub(crate) fn status(&self) -> Option<WalRedoManagerStatus> {
433 0 : match self {
434 0 : WalRedoManager::Prod(_, m) => Some(m.status()),
435 0 : #[cfg(test)]
436 0 : WalRedoManager::Test(_) => None,
437 0 : }
438 0 : }
439 : }
440 :
441 0 : #[derive(Debug, thiserror::Error, PartialEq, Eq)]
442 : pub enum GetTimelineError {
443 : #[error("Timeline is shutting down")]
444 : ShuttingDown,
445 : #[error("Timeline {tenant_id}/{timeline_id} is not active, state: {state:?}")]
446 : NotActive {
447 : tenant_id: TenantShardId,
448 : timeline_id: TimelineId,
449 : state: TimelineState,
450 : },
451 : #[error("Timeline {tenant_id}/{timeline_id} was not found")]
452 : NotFound {
453 : tenant_id: TenantShardId,
454 : timeline_id: TimelineId,
455 : },
456 : }
457 :
458 0 : #[derive(Debug, thiserror::Error)]
459 : pub enum LoadLocalTimelineError {
460 : #[error("FailedToLoad")]
461 : Load(#[source] anyhow::Error),
462 : #[error("FailedToResumeDeletion")]
463 : ResumeDeletion(#[source] anyhow::Error),
464 : }
465 :
466 0 : #[derive(thiserror::Error)]
467 : pub enum DeleteTimelineError {
468 : #[error("NotFound")]
469 : NotFound,
470 :
471 : #[error("HasChildren")]
472 : HasChildren(Vec<TimelineId>),
473 :
474 : #[error("Timeline deletion is already in progress")]
475 : AlreadyInProgress(Arc<tokio::sync::Mutex<DeleteTimelineFlow>>),
476 :
477 : #[error(transparent)]
478 : Other(#[from] anyhow::Error),
479 : }
480 :
481 : impl Debug for DeleteTimelineError {
482 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
483 0 : match self {
484 0 : Self::NotFound => write!(f, "NotFound"),
485 0 : Self::HasChildren(c) => f.debug_tuple("HasChildren").field(c).finish(),
486 0 : Self::AlreadyInProgress(_) => f.debug_tuple("AlreadyInProgress").finish(),
487 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
488 : }
489 0 : }
490 : }
491 :
492 : pub enum SetStoppingError {
493 : AlreadyStopping(completion::Barrier),
494 : Broken,
495 : }
496 :
497 : impl Debug for SetStoppingError {
498 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
499 0 : match self {
500 0 : Self::AlreadyStopping(_) => f.debug_tuple("AlreadyStopping").finish(),
501 0 : Self::Broken => write!(f, "Broken"),
502 : }
503 0 : }
504 : }
505 :
506 0 : #[derive(thiserror::Error, Debug)]
507 : pub enum CreateTimelineError {
508 : #[error("creation of timeline with the given ID is in progress")]
509 : AlreadyCreating,
510 : #[error("timeline already exists with different parameters")]
511 : Conflict,
512 : #[error(transparent)]
513 : AncestorLsn(anyhow::Error),
514 : #[error("ancestor timeline is not active")]
515 : AncestorNotActive,
516 : #[error("tenant shutting down")]
517 : ShuttingDown,
518 : #[error(transparent)]
519 : Other(#[from] anyhow::Error),
520 : }
521 :
522 : #[derive(thiserror::Error, Debug)]
523 : enum InitdbError {
524 : Other(anyhow::Error),
525 : Cancelled,
526 : Spawn(std::io::Result<()>),
527 : Failed(std::process::ExitStatus, Vec<u8>),
528 : }
529 :
530 : impl fmt::Display for InitdbError {
531 0 : fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
532 0 : match self {
533 0 : InitdbError::Cancelled => write!(f, "Operation was cancelled"),
534 0 : InitdbError::Spawn(e) => write!(f, "Spawn error: {:?}", e),
535 0 : InitdbError::Failed(status, stderr) => write!(
536 0 : f,
537 0 : "Command failed with status {:?}: {}",
538 0 : status,
539 0 : String::from_utf8_lossy(stderr)
540 0 : ),
541 0 : InitdbError::Other(e) => write!(f, "Error: {:?}", e),
542 : }
543 0 : }
544 : }
545 :
546 : impl From<std::io::Error> for InitdbError {
547 0 : fn from(error: std::io::Error) -> Self {
548 0 : InitdbError::Spawn(Err(error))
549 0 : }
550 : }
551 :
552 : enum CreateTimelineCause {
553 : Load,
554 : Delete,
555 : }
556 :
557 0 : #[derive(thiserror::Error, Debug)]
558 : pub(crate) enum GcError {
559 : // The tenant is shutting down
560 : #[error("tenant shutting down")]
561 : TenantCancelled,
562 :
563 : // The tenant is shutting down
564 : #[error("timeline shutting down")]
565 : TimelineCancelled,
566 :
567 : // The tenant is in a state inelegible to run GC
568 : #[error("not active")]
569 : NotActive,
570 :
571 : // A requested GC cutoff LSN was invalid, for example it tried to move backwards
572 : #[error("not active")]
573 : BadLsn { why: String },
574 :
575 : // A remote storage error while scheduling updates after compaction
576 : #[error(transparent)]
577 : Remote(anyhow::Error),
578 :
579 : // An error reading while calculating GC cutoffs
580 : #[error(transparent)]
581 : GcCutoffs(PageReconstructError),
582 :
583 : // If GC was invoked for a particular timeline, this error means it didn't exist
584 : #[error("timeline not found")]
585 : TimelineNotFound,
586 : }
587 :
588 : impl From<PageReconstructError> for GcError {
589 0 : fn from(value: PageReconstructError) -> Self {
590 0 : match value {
591 0 : PageReconstructError::Cancelled => Self::TimelineCancelled,
592 0 : other => Self::GcCutoffs(other),
593 : }
594 0 : }
595 : }
596 :
597 0 : #[derive(thiserror::Error, Debug)]
598 : pub(crate) enum LoadConfigError {
599 : #[error("TOML deserialization error: '{0}'")]
600 : DeserializeToml(#[from] toml_edit::de::Error),
601 :
602 : #[error("Config not found at {0}")]
603 : NotFound(Utf8PathBuf),
604 : }
605 :
606 : impl Tenant {
607 : /// Yet another helper for timeline initialization.
608 : ///
609 : /// - Initializes the Timeline struct and inserts it into the tenant's hash map
610 : /// - Scans the local timeline directory for layer files and builds the layer map
611 : /// - Downloads remote index file and adds remote files to the layer map
612 : /// - Schedules remote upload tasks for any files that are present locally but missing from remote storage.
613 : ///
614 : /// If the operation fails, the timeline is left in the tenant's hash map in Broken state. On success,
615 : /// it is marked as Active.
616 : #[allow(clippy::too_many_arguments)]
617 6 : async fn timeline_init_and_sync(
618 6 : &self,
619 6 : timeline_id: TimelineId,
620 6 : resources: TimelineResources,
621 6 : index_part: Option<IndexPart>,
622 6 : metadata: TimelineMetadata,
623 6 : ancestor: Option<Arc<Timeline>>,
624 6 : last_aux_file_policy: Option<AuxFilePolicy>,
625 6 : _ctx: &RequestContext,
626 6 : ) -> anyhow::Result<()> {
627 6 : let tenant_id = self.tenant_shard_id;
628 :
629 6 : let timeline = self.create_timeline_struct(
630 6 : timeline_id,
631 6 : &metadata,
632 6 : ancestor.clone(),
633 6 : resources,
634 6 : CreateTimelineCause::Load,
635 6 : // This could be derived from ancestor branch + index part. Though the only caller of `timeline_init_and_sync` is `load_remote_timeline`,
636 6 : // there will potentially be other caller of this function in the future, and we don't know whether `index_part` or `ancestor` takes precedence.
637 6 : // Therefore, we pass this field explicitly for now, and remove it once we fully migrate to aux file v2.
638 6 : last_aux_file_policy,
639 6 : )?;
640 6 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
641 6 : anyhow::ensure!(
642 6 : disk_consistent_lsn.is_valid(),
643 0 : "Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn"
644 : );
645 6 : assert_eq!(
646 6 : disk_consistent_lsn,
647 6 : metadata.disk_consistent_lsn(),
648 0 : "these are used interchangeably"
649 : );
650 :
651 6 : if let Some(index_part) = index_part.as_ref() {
652 6 : timeline.remote_client.init_upload_queue(index_part)?;
653 :
654 6 : timeline
655 6 : .last_aux_file_policy
656 6 : .store(index_part.last_aux_file_policy());
657 : } else {
658 : // No data on the remote storage, but we have local metadata file. We can end up
659 : // here with timeline_create being interrupted before finishing index part upload.
660 : // By doing what we do here, the index part upload is retried.
661 : // If control plane retries timeline creation in the meantime, the mgmt API handler
662 : // for timeline creation will coalesce on the upload we queue here.
663 :
664 : // FIXME: this branch should be dead code as we no longer write local metadata.
665 :
666 0 : timeline
667 0 : .remote_client
668 0 : .init_upload_queue_for_empty_remote(&metadata)?;
669 0 : timeline
670 0 : .remote_client
671 0 : .schedule_index_upload_for_full_metadata_update(&metadata)?;
672 : }
673 :
674 6 : timeline
675 6 : .load_layer_map(disk_consistent_lsn, index_part)
676 4 : .await
677 6 : .with_context(|| {
678 0 : format!("Failed to load layermap for timeline {tenant_id}/{timeline_id}")
679 6 : })?;
680 :
681 : {
682 : // avoiding holding it across awaits
683 6 : let mut timelines_accessor = self.timelines.lock().unwrap();
684 6 : match timelines_accessor.entry(timeline_id) {
685 : // We should never try and load the same timeline twice during startup
686 : Entry::Occupied(_) => {
687 0 : unreachable!(
688 0 : "Timeline {tenant_id}/{timeline_id} already exists in the tenant map"
689 0 : );
690 : }
691 6 : Entry::Vacant(v) => {
692 6 : v.insert(Arc::clone(&timeline));
693 6 : timeline.maybe_spawn_flush_loop();
694 6 : }
695 6 : }
696 6 : };
697 6 :
698 6 : // Sanity check: a timeline should have some content.
699 6 : anyhow::ensure!(
700 6 : ancestor.is_some()
701 4 : || timeline
702 4 : .layers
703 4 : .read()
704 0 : .await
705 4 : .layer_map()
706 4 : .iter_historic_layers()
707 4 : .next()
708 4 : .is_some(),
709 0 : "Timeline has no ancestor and no layer files"
710 : );
711 :
712 6 : Ok(())
713 6 : }
714 :
715 : /// Attach a tenant that's available in cloud storage.
716 : ///
717 : /// This returns quickly, after just creating the in-memory object
718 : /// Tenant struct and launching a background task to download
719 : /// the remote index files. On return, the tenant is most likely still in
720 : /// Attaching state, and it will become Active once the background task
721 : /// finishes. You can use wait_until_active() to wait for the task to
722 : /// complete.
723 : ///
724 : #[allow(clippy::too_many_arguments)]
725 0 : pub(crate) fn spawn(
726 0 : conf: &'static PageServerConf,
727 0 : tenant_shard_id: TenantShardId,
728 0 : resources: TenantSharedResources,
729 0 : attached_conf: AttachedTenantConf,
730 0 : shard_identity: ShardIdentity,
731 0 : init_order: Option<InitializationOrder>,
732 0 : mode: SpawnMode,
733 0 : ctx: &RequestContext,
734 0 : ) -> Result<Arc<Tenant>, GlobalShutDown> {
735 0 : let wal_redo_manager =
736 0 : WalRedoManager::new(PostgresRedoManager::new(conf, tenant_shard_id))?;
737 :
738 : let TenantSharedResources {
739 0 : broker_client,
740 0 : remote_storage,
741 0 : deletion_queue_client,
742 0 : l0_flush_global_state,
743 0 : } = resources;
744 0 :
745 0 : let attach_mode = attached_conf.location.attach_mode;
746 0 : let generation = attached_conf.location.generation;
747 0 :
748 0 : let tenant = Arc::new(Tenant::new(
749 0 : TenantState::Attaching,
750 0 : conf,
751 0 : attached_conf,
752 0 : shard_identity,
753 0 : Some(wal_redo_manager),
754 0 : tenant_shard_id,
755 0 : remote_storage.clone(),
756 0 : deletion_queue_client,
757 0 : l0_flush_global_state,
758 0 : ));
759 0 :
760 0 : // The attach task will carry a GateGuard, so that shutdown() reliably waits for it to drop out if
761 0 : // we shut down while attaching.
762 0 : let attach_gate_guard = tenant
763 0 : .gate
764 0 : .enter()
765 0 : .expect("We just created the Tenant: nothing else can have shut it down yet");
766 0 :
767 0 : // Do all the hard work in the background
768 0 : let tenant_clone = Arc::clone(&tenant);
769 0 : let ctx = ctx.detached_child(TaskKind::Attach, DownloadBehavior::Warn);
770 0 : task_mgr::spawn(
771 0 : &tokio::runtime::Handle::current(),
772 0 : TaskKind::Attach,
773 0 : Some(tenant_shard_id),
774 0 : None,
775 0 : "attach tenant",
776 0 : async move {
777 0 :
778 0 : info!(
779 : ?attach_mode,
780 0 : "Attaching tenant"
781 : );
782 :
783 0 : let _gate_guard = attach_gate_guard;
784 0 :
785 0 : // Is this tenant being spawned as part of process startup?
786 0 : let starting_up = init_order.is_some();
787 : scopeguard::defer! {
788 : if starting_up {
789 : TENANT.startup_complete.inc();
790 : }
791 : }
792 :
793 : // Ideally we should use Tenant::set_broken_no_wait, but it is not supposed to be used when tenant is in loading state.
794 : enum BrokenVerbosity {
795 : Error,
796 : Info
797 : }
798 0 : let make_broken =
799 0 : |t: &Tenant, err: anyhow::Error, verbosity: BrokenVerbosity| {
800 0 : match verbosity {
801 : BrokenVerbosity::Info => {
802 0 : info!("attach cancelled, setting tenant state to Broken: {err}");
803 : },
804 : BrokenVerbosity::Error => {
805 0 : error!("attach failed, setting tenant state to Broken: {err:?}");
806 : }
807 : }
808 0 : t.state.send_modify(|state| {
809 0 : // The Stopping case is for when we have passed control on to DeleteTenantFlow:
810 0 : // if it errors, we will call make_broken when tenant is already in Stopping.
811 0 : assert!(
812 0 : matches!(*state, TenantState::Attaching | TenantState::Stopping { .. }),
813 0 : "the attach task owns the tenant state until activation is complete"
814 : );
815 :
816 0 : *state = TenantState::broken_from_reason(err.to_string());
817 0 : });
818 0 : };
819 :
820 0 : let mut init_order = init_order;
821 0 : // take the completion because initial tenant loading will complete when all of
822 0 : // these tasks complete.
823 0 : let _completion = init_order
824 0 : .as_mut()
825 0 : .and_then(|x| x.initial_tenant_load.take());
826 0 : let remote_load_completion = init_order
827 0 : .as_mut()
828 0 : .and_then(|x| x.initial_tenant_load_remote.take());
829 :
830 : enum AttachType<'a> {
831 : /// We are attaching this tenant lazily in the background.
832 : Warmup {
833 : _permit: tokio::sync::SemaphorePermit<'a>,
834 : during_startup: bool
835 : },
836 : /// We are attaching this tenant as soon as we can, because for example an
837 : /// endpoint tried to access it.
838 : OnDemand,
839 : /// During normal operations after startup, we are attaching a tenant, and
840 : /// eager attach was requested.
841 : Normal,
842 : }
843 :
844 0 : let attach_type = if matches!(mode, SpawnMode::Lazy) {
845 : // Before doing any I/O, wait for at least one of:
846 : // - A client attempting to access to this tenant (on-demand loading)
847 : // - A permit becoming available in the warmup semaphore (background warmup)
848 :
849 : tokio::select!(
850 : permit = tenant_clone.activate_now_sem.acquire() => {
851 : let _ = permit.expect("activate_now_sem is never closed");
852 : tracing::info!("Activating tenant (on-demand)");
853 : AttachType::OnDemand
854 : },
855 : permit = conf.concurrent_tenant_warmup.inner().acquire() => {
856 : let _permit = permit.expect("concurrent_tenant_warmup semaphore is never closed");
857 : tracing::info!("Activating tenant (warmup)");
858 : AttachType::Warmup {
859 : _permit,
860 : during_startup: init_order.is_some()
861 : }
862 : }
863 : _ = tenant_clone.cancel.cancelled() => {
864 : // This is safe, but should be pretty rare: it is interesting if a tenant
865 : // stayed in Activating for such a long time that shutdown found it in
866 : // that state.
867 : tracing::info!(state=%tenant_clone.current_state(), "Tenant shut down before activation");
868 : // Make the tenant broken so that set_stopping will not hang waiting for it to leave
869 : // the Attaching state. This is an over-reaction (nothing really broke, the tenant is
870 : // just shutting down), but ensures progress.
871 : make_broken(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"), BrokenVerbosity::Info);
872 : return Ok(());
873 : },
874 : )
875 : } else {
876 : // SpawnMode::{Create,Eager} always cause jumping ahead of the
877 : // concurrent_tenant_warmup queue
878 0 : AttachType::Normal
879 : };
880 :
881 0 : let preload = match &mode {
882 : SpawnMode::Eager | SpawnMode::Lazy => {
883 0 : let _preload_timer = TENANT.preload.start_timer();
884 0 : let res = tenant_clone
885 0 : .preload(&remote_storage, task_mgr::shutdown_token())
886 0 : .await;
887 0 : match res {
888 0 : Ok(p) => Some(p),
889 0 : Err(e) => {
890 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
891 0 : return Ok(());
892 : }
893 : }
894 : }
895 :
896 : };
897 :
898 : // Remote preload is complete.
899 0 : drop(remote_load_completion);
900 :
901 : // We will time the duration of the attach phase unless this is a creation (attach will do no work)
902 0 : let attached = {
903 0 : let _attach_timer = Some(TENANT.attach.start_timer());
904 0 : tenant_clone.attach(preload, &ctx).await
905 : };
906 :
907 0 : match attached {
908 : Ok(()) => {
909 0 : info!("attach finished, activating");
910 0 : tenant_clone.activate(broker_client, None, &ctx);
911 : }
912 0 : Err(e) => {
913 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
914 0 : }
915 : }
916 :
917 : // If we are doing an opportunistic warmup attachment at startup, initialize
918 : // logical size at the same time. This is better than starting a bunch of idle tenants
919 : // with cold caches and then coming back later to initialize their logical sizes.
920 : //
921 : // It also prevents the warmup proccess competing with the concurrency limit on
922 : // logical size calculations: if logical size calculation semaphore is saturated,
923 : // then warmup will wait for that before proceeding to the next tenant.
924 0 : if matches!(attach_type, AttachType::Warmup { during_startup: true, .. }) {
925 0 : let mut futs: FuturesUnordered<_> = tenant_clone.timelines.lock().unwrap().values().cloned().map(|t| t.await_initial_logical_size()).collect();
926 0 : tracing::info!("Waiting for initial logical sizes while warming up...");
927 0 : while futs.next().await.is_some() {}
928 0 : tracing::info!("Warm-up complete");
929 0 : }
930 :
931 0 : Ok(())
932 0 : }
933 0 : .instrument(tracing::info_span!(parent: None, "attach", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), gen=?generation)),
934 : );
935 0 : Ok(tenant)
936 0 : }
937 :
938 348 : #[instrument(skip_all)]
939 : pub(crate) async fn preload(
940 : self: &Arc<Self>,
941 : remote_storage: &GenericRemoteStorage,
942 : cancel: CancellationToken,
943 : ) -> anyhow::Result<TenantPreload> {
944 : span::debug_assert_current_span_has_tenant_id();
945 : // Get list of remote timelines
946 : // download index files for every tenant timeline
947 : info!("listing remote timelines");
948 : let (remote_timeline_ids, other_keys) = remote_timeline_client::list_remote_timelines(
949 : remote_storage,
950 : self.tenant_shard_id,
951 : cancel.clone(),
952 : )
953 : .await?;
954 :
955 : info!("found {} timelines", remote_timeline_ids.len(),);
956 :
957 : for k in other_keys {
958 : warn!("Unexpected non timeline key {k}");
959 : }
960 :
961 : Ok(TenantPreload {
962 : timelines: Self::load_timeline_metadata(
963 : self,
964 : remote_timeline_ids,
965 : remote_storage,
966 : cancel,
967 : )
968 : .await?,
969 : })
970 : }
971 :
972 : ///
973 : /// Background task that downloads all data for a tenant and brings it to Active state.
974 : ///
975 : /// No background tasks are started as part of this routine.
976 : ///
977 174 : async fn attach(
978 174 : self: &Arc<Tenant>,
979 174 : preload: Option<TenantPreload>,
980 174 : ctx: &RequestContext,
981 174 : ) -> anyhow::Result<()> {
982 174 : span::debug_assert_current_span_has_tenant_id();
983 174 :
984 174 : failpoint_support::sleep_millis_async!("before-attaching-tenant");
985 :
986 174 : let Some(preload) = preload else {
987 0 : anyhow::bail!("local-only deployment is no longer supported, https://github.com/neondatabase/neon/issues/5624");
988 : };
989 :
990 174 : let mut timelines_to_resume_deletions = vec![];
991 174 :
992 174 : let mut remote_index_and_client = HashMap::new();
993 174 : let mut timeline_ancestors = HashMap::new();
994 174 : let mut existent_timelines = HashSet::new();
995 180 : for (timeline_id, preload) in preload.timelines {
996 6 : let index_part = match preload.index_part {
997 6 : Ok(i) => {
998 6 : debug!("remote index part exists for timeline {timeline_id}");
999 : // We found index_part on the remote, this is the standard case.
1000 6 : existent_timelines.insert(timeline_id);
1001 6 : i
1002 : }
1003 : Err(DownloadError::NotFound) => {
1004 : // There is no index_part on the remote. We only get here
1005 : // if there is some prefix for the timeline in the remote storage.
1006 : // This can e.g. be the initdb.tar.zst archive, maybe a
1007 : // remnant from a prior incomplete creation or deletion attempt.
1008 : // Delete the local directory as the deciding criterion for a
1009 : // timeline's existence is presence of index_part.
1010 0 : info!(%timeline_id, "index_part not found on remote");
1011 0 : continue;
1012 : }
1013 0 : Err(e) => {
1014 0 : // Some (possibly ephemeral) error happened during index_part download.
1015 0 : // Pretend the timeline exists to not delete the timeline directory,
1016 0 : // as it might be a temporary issue and we don't want to re-download
1017 0 : // everything after it resolves.
1018 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1019 :
1020 0 : existent_timelines.insert(timeline_id);
1021 0 : continue;
1022 : }
1023 : };
1024 6 : match index_part {
1025 6 : MaybeDeletedIndexPart::IndexPart(index_part) => {
1026 6 : timeline_ancestors.insert(timeline_id, index_part.metadata.clone());
1027 6 : remote_index_and_client.insert(timeline_id, (index_part, preload.client));
1028 6 : }
1029 0 : MaybeDeletedIndexPart::Deleted(index_part) => {
1030 0 : info!(
1031 0 : "timeline {} is deleted, picking to resume deletion",
1032 : timeline_id
1033 : );
1034 0 : timelines_to_resume_deletions.push((timeline_id, index_part, preload.client));
1035 : }
1036 : }
1037 : }
1038 :
1039 : // For every timeline, download the metadata file, scan the local directory,
1040 : // and build a layer map that contains an entry for each remote and local
1041 : // layer file.
1042 174 : let sorted_timelines = tree_sort_timelines(timeline_ancestors, |m| m.ancestor_timeline())?;
1043 180 : for (timeline_id, remote_metadata) in sorted_timelines {
1044 6 : let (index_part, remote_client) = remote_index_and_client
1045 6 : .remove(&timeline_id)
1046 6 : .expect("just put it in above");
1047 6 :
1048 6 : // TODO again handle early failure
1049 6 : self.load_remote_timeline(
1050 6 : timeline_id,
1051 6 : index_part,
1052 6 : remote_metadata,
1053 6 : TimelineResources {
1054 6 : remote_client,
1055 6 : timeline_get_throttle: self.timeline_get_throttle.clone(),
1056 6 : l0_flush_global_state: self.l0_flush_global_state.clone(),
1057 6 : },
1058 6 : ctx,
1059 6 : )
1060 10 : .await
1061 6 : .with_context(|| {
1062 0 : format!(
1063 0 : "failed to load remote timeline {} for tenant {}",
1064 0 : timeline_id, self.tenant_shard_id
1065 0 : )
1066 6 : })?;
1067 : }
1068 :
1069 : // Walk through deleted timelines, resume deletion
1070 174 : for (timeline_id, index_part, remote_timeline_client) in timelines_to_resume_deletions {
1071 0 : remote_timeline_client
1072 0 : .init_upload_queue_stopped_to_continue_deletion(&index_part)
1073 0 : .context("init queue stopped")
1074 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1075 :
1076 0 : DeleteTimelineFlow::resume_deletion(
1077 0 : Arc::clone(self),
1078 0 : timeline_id,
1079 0 : &index_part.metadata,
1080 0 : remote_timeline_client,
1081 0 : )
1082 0 : .instrument(tracing::info_span!("timeline_delete", %timeline_id))
1083 0 : .await
1084 0 : .context("resume_deletion")
1085 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1086 : }
1087 :
1088 : // The local filesystem contents are a cache of what's in the remote IndexPart;
1089 : // IndexPart is the source of truth.
1090 174 : self.clean_up_timelines(&existent_timelines)?;
1091 :
1092 174 : fail::fail_point!("attach-before-activate", |_| {
1093 0 : anyhow::bail!("attach-before-activate");
1094 174 : });
1095 174 : failpoint_support::sleep_millis_async!("attach-before-activate-sleep", &self.cancel);
1096 :
1097 174 : info!("Done");
1098 :
1099 174 : Ok(())
1100 174 : }
1101 :
1102 : /// Check for any local timeline directories that are temporary, or do not correspond to a
1103 : /// timeline that still exists: this can happen if we crashed during a deletion/creation, or
1104 : /// if a timeline was deleted while the tenant was attached to a different pageserver.
1105 174 : fn clean_up_timelines(&self, existent_timelines: &HashSet<TimelineId>) -> anyhow::Result<()> {
1106 174 : let timelines_dir = self.conf.timelines_path(&self.tenant_shard_id);
1107 :
1108 174 : let entries = match timelines_dir.read_dir_utf8() {
1109 174 : Ok(d) => d,
1110 0 : Err(e) => {
1111 0 : if e.kind() == std::io::ErrorKind::NotFound {
1112 0 : return Ok(());
1113 : } else {
1114 0 : return Err(e).context("list timelines directory for tenant");
1115 : }
1116 : }
1117 : };
1118 :
1119 182 : for entry in entries {
1120 8 : let entry = entry.context("read timeline dir entry")?;
1121 8 : let entry_path = entry.path();
1122 :
1123 8 : let purge = if crate::is_temporary(entry_path)
1124 : // TODO: remove uninit mark code (https://github.com/neondatabase/neon/issues/5718)
1125 8 : || is_uninit_mark(entry_path)
1126 8 : || crate::is_delete_mark(entry_path)
1127 : {
1128 0 : true
1129 : } else {
1130 8 : match TimelineId::try_from(entry_path.file_name()) {
1131 8 : Ok(i) => {
1132 8 : // Purge if the timeline ID does not exist in remote storage: remote storage is the authority.
1133 8 : !existent_timelines.contains(&i)
1134 : }
1135 0 : Err(e) => {
1136 0 : tracing::warn!(
1137 0 : "Unparseable directory in timelines directory: {entry_path}, ignoring ({e})"
1138 : );
1139 : // Do not purge junk: if we don't recognize it, be cautious and leave it for a human.
1140 0 : false
1141 : }
1142 : }
1143 : };
1144 :
1145 8 : if purge {
1146 2 : tracing::info!("Purging stale timeline dentry {entry_path}");
1147 2 : if let Err(e) = match entry.file_type() {
1148 2 : Ok(t) => if t.is_dir() {
1149 2 : std::fs::remove_dir_all(entry_path)
1150 : } else {
1151 0 : std::fs::remove_file(entry_path)
1152 : }
1153 2 : .or_else(fs_ext::ignore_not_found),
1154 0 : Err(e) => Err(e),
1155 : } {
1156 0 : tracing::warn!("Failed to purge stale timeline dentry {entry_path}: {e}");
1157 2 : }
1158 6 : }
1159 : }
1160 :
1161 174 : Ok(())
1162 174 : }
1163 :
1164 : /// Get sum of all remote timelines sizes
1165 : ///
1166 : /// This function relies on the index_part instead of listing the remote storage
1167 0 : pub fn remote_size(&self) -> u64 {
1168 0 : let mut size = 0;
1169 :
1170 0 : for timeline in self.list_timelines() {
1171 0 : size += timeline.remote_client.get_remote_physical_size();
1172 0 : }
1173 :
1174 0 : size
1175 0 : }
1176 :
1177 12 : #[instrument(skip_all, fields(timeline_id=%timeline_id))]
1178 : async fn load_remote_timeline(
1179 : &self,
1180 : timeline_id: TimelineId,
1181 : index_part: IndexPart,
1182 : remote_metadata: TimelineMetadata,
1183 : resources: TimelineResources,
1184 : ctx: &RequestContext,
1185 : ) -> anyhow::Result<()> {
1186 : span::debug_assert_current_span_has_tenant_id();
1187 :
1188 : info!("downloading index file for timeline {}", timeline_id);
1189 : tokio::fs::create_dir_all(self.conf.timeline_path(&self.tenant_shard_id, &timeline_id))
1190 : .await
1191 : .context("Failed to create new timeline directory")?;
1192 :
1193 : let ancestor = if let Some(ancestor_id) = remote_metadata.ancestor_timeline() {
1194 : let timelines = self.timelines.lock().unwrap();
1195 : Some(Arc::clone(timelines.get(&ancestor_id).ok_or_else(
1196 0 : || {
1197 0 : anyhow::anyhow!(
1198 0 : "cannot find ancestor timeline {ancestor_id} for timeline {timeline_id}"
1199 0 : )
1200 0 : },
1201 : )?))
1202 : } else {
1203 : None
1204 : };
1205 :
1206 : let last_aux_file_policy = index_part.last_aux_file_policy();
1207 :
1208 : self.timeline_init_and_sync(
1209 : timeline_id,
1210 : resources,
1211 : Some(index_part),
1212 : remote_metadata,
1213 : ancestor,
1214 : last_aux_file_policy,
1215 : ctx,
1216 : )
1217 : .await
1218 : }
1219 :
1220 174 : async fn load_timeline_metadata(
1221 174 : self: &Arc<Tenant>,
1222 174 : timeline_ids: HashSet<TimelineId>,
1223 174 : remote_storage: &GenericRemoteStorage,
1224 174 : cancel: CancellationToken,
1225 174 : ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
1226 174 : let mut part_downloads = JoinSet::new();
1227 180 : for timeline_id in timeline_ids {
1228 6 : let client = RemoteTimelineClient::new(
1229 6 : remote_storage.clone(),
1230 6 : self.deletion_queue_client.clone(),
1231 6 : self.conf,
1232 6 : self.tenant_shard_id,
1233 6 : timeline_id,
1234 6 : self.generation,
1235 6 : );
1236 6 : let cancel_clone = cancel.clone();
1237 6 : part_downloads.spawn(
1238 6 : async move {
1239 6 : debug!("starting index part download");
1240 :
1241 23 : let index_part = client.download_index_file(&cancel_clone).await;
1242 :
1243 6 : debug!("finished index part download");
1244 :
1245 6 : Result::<_, anyhow::Error>::Ok(TimelinePreload {
1246 6 : client,
1247 6 : timeline_id,
1248 6 : index_part,
1249 6 : })
1250 6 : }
1251 6 : .map(move |res| {
1252 6 : res.with_context(|| format!("download index part for timeline {timeline_id}"))
1253 6 : })
1254 6 : .instrument(info_span!("download_index_part", %timeline_id)),
1255 : );
1256 : }
1257 :
1258 174 : let mut timeline_preloads: HashMap<TimelineId, TimelinePreload> = HashMap::new();
1259 :
1260 180 : loop {
1261 180 : tokio::select!(
1262 : next = part_downloads.join_next() => {
1263 : match next {
1264 : Some(result) => {
1265 : let preload_result = result.context("join preload task")?;
1266 : let preload = preload_result?;
1267 : timeline_preloads.insert(preload.timeline_id, preload);
1268 : },
1269 : None => {
1270 : break;
1271 : }
1272 : }
1273 : },
1274 : _ = cancel.cancelled() => {
1275 : anyhow::bail!("Cancelled while waiting for remote index download")
1276 : }
1277 180 : )
1278 180 : }
1279 :
1280 174 : Ok(timeline_preloads)
1281 174 : }
1282 :
1283 0 : pub(crate) async fn apply_timeline_archival_config(
1284 0 : &self,
1285 0 : timeline_id: TimelineId,
1286 0 : state: TimelineArchivalState,
1287 0 : ) -> anyhow::Result<()> {
1288 0 : let timeline = self
1289 0 : .get_timeline(timeline_id, false)
1290 0 : .context("Cannot apply timeline archival config to inexistent timeline")?;
1291 :
1292 0 : let upload_needed = timeline
1293 0 : .remote_client
1294 0 : .schedule_index_upload_for_timeline_archival_state(state)?;
1295 :
1296 0 : if upload_needed {
1297 : const MAX_WAIT: Duration = Duration::from_secs(10);
1298 0 : let Ok(v) =
1299 0 : tokio::time::timeout(MAX_WAIT, timeline.remote_client.wait_completion()).await
1300 : else {
1301 0 : tracing::warn!("reached timeout for waiting on upload queue");
1302 0 : bail!("reached timeout for upload queue flush");
1303 : };
1304 0 : v?;
1305 0 : }
1306 0 : Ok(())
1307 0 : }
1308 :
1309 4 : pub(crate) fn tenant_shard_id(&self) -> TenantShardId {
1310 4 : self.tenant_shard_id
1311 4 : }
1312 :
1313 : /// Get Timeline handle for given Neon timeline ID.
1314 : /// This function is idempotent. It doesn't change internal state in any way.
1315 222 : pub fn get_timeline(
1316 222 : &self,
1317 222 : timeline_id: TimelineId,
1318 222 : active_only: bool,
1319 222 : ) -> Result<Arc<Timeline>, GetTimelineError> {
1320 222 : let timelines_accessor = self.timelines.lock().unwrap();
1321 222 : let timeline = timelines_accessor
1322 222 : .get(&timeline_id)
1323 222 : .ok_or(GetTimelineError::NotFound {
1324 222 : tenant_id: self.tenant_shard_id,
1325 222 : timeline_id,
1326 222 : })?;
1327 :
1328 220 : if active_only && !timeline.is_active() {
1329 0 : Err(GetTimelineError::NotActive {
1330 0 : tenant_id: self.tenant_shard_id,
1331 0 : timeline_id,
1332 0 : state: timeline.current_state(),
1333 0 : })
1334 : } else {
1335 220 : Ok(Arc::clone(timeline))
1336 : }
1337 222 : }
1338 :
1339 : /// Lists timelines the tenant contains.
1340 : /// Up to tenant's implementation to omit certain timelines that ar not considered ready for use.
1341 8 : pub fn list_timelines(&self) -> Vec<Arc<Timeline>> {
1342 8 : self.timelines
1343 8 : .lock()
1344 8 : .unwrap()
1345 8 : .values()
1346 8 : .map(Arc::clone)
1347 8 : .collect()
1348 8 : }
1349 :
1350 0 : pub fn list_timeline_ids(&self) -> Vec<TimelineId> {
1351 0 : self.timelines.lock().unwrap().keys().cloned().collect()
1352 0 : }
1353 :
1354 : /// This is used to create the initial 'main' timeline during bootstrapping,
1355 : /// or when importing a new base backup. The caller is expected to load an
1356 : /// initial image of the datadir to the new timeline after this.
1357 : ///
1358 : /// Until that happens, the on-disk state is invalid (disk_consistent_lsn=Lsn(0))
1359 : /// and the timeline will fail to load at a restart.
1360 : ///
1361 : /// For tests, use `DatadirModification::init_empty_test_timeline` + `commit` to setup the
1362 : /// minimum amount of keys required to get a writable timeline.
1363 : /// (Without it, `put` might fail due to `repartition` failing.)
1364 166 : pub(crate) async fn create_empty_timeline(
1365 166 : &self,
1366 166 : new_timeline_id: TimelineId,
1367 166 : initdb_lsn: Lsn,
1368 166 : pg_version: u32,
1369 166 : _ctx: &RequestContext,
1370 166 : ) -> anyhow::Result<UninitializedTimeline> {
1371 166 : anyhow::ensure!(
1372 166 : self.is_active(),
1373 0 : "Cannot create empty timelines on inactive tenant"
1374 : );
1375 :
1376 : // Protect against concurrent attempts to use this TimelineId
1377 166 : let create_guard = self.create_timeline_create_guard(new_timeline_id)?;
1378 :
1379 164 : let new_metadata = TimelineMetadata::new(
1380 164 : // Initialize disk_consistent LSN to 0, The caller must import some data to
1381 164 : // make it valid, before calling finish_creation()
1382 164 : Lsn(0),
1383 164 : None,
1384 164 : None,
1385 164 : Lsn(0),
1386 164 : initdb_lsn,
1387 164 : initdb_lsn,
1388 164 : pg_version,
1389 164 : );
1390 164 : self.prepare_new_timeline(
1391 164 : new_timeline_id,
1392 164 : &new_metadata,
1393 164 : create_guard,
1394 164 : initdb_lsn,
1395 164 : None,
1396 164 : None,
1397 164 : )
1398 0 : .await
1399 166 : }
1400 :
1401 : /// Helper for unit tests to create an empty timeline.
1402 : ///
1403 : /// The timeline is has state value `Active` but its background loops are not running.
1404 : // This makes the various functions which anyhow::ensure! for Active state work in tests.
1405 : // Our current tests don't need the background loops.
1406 : #[cfg(test)]
1407 156 : pub async fn create_test_timeline(
1408 156 : &self,
1409 156 : new_timeline_id: TimelineId,
1410 156 : initdb_lsn: Lsn,
1411 156 : pg_version: u32,
1412 156 : ctx: &RequestContext,
1413 156 : ) -> anyhow::Result<Arc<Timeline>> {
1414 156 : let uninit_tl = self
1415 156 : .create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
1416 0 : .await?;
1417 156 : let tline = uninit_tl.raw_timeline().expect("we just created it");
1418 156 : assert_eq!(tline.get_last_record_lsn(), Lsn(0));
1419 :
1420 : // Setup minimum keys required for the timeline to be usable.
1421 156 : let mut modification = tline.begin_modification(initdb_lsn);
1422 156 : modification
1423 156 : .init_empty_test_timeline()
1424 156 : .context("init_empty_test_timeline")?;
1425 156 : modification
1426 156 : .commit(ctx)
1427 150 : .await
1428 156 : .context("commit init_empty_test_timeline modification")?;
1429 :
1430 : // Flush to disk so that uninit_tl's check for valid disk_consistent_lsn passes.
1431 156 : tline.maybe_spawn_flush_loop();
1432 156 : tline.freeze_and_flush().await.context("freeze_and_flush")?;
1433 :
1434 : // Make sure the freeze_and_flush reaches remote storage.
1435 156 : tline.remote_client.wait_completion().await.unwrap();
1436 :
1437 156 : let tl = uninit_tl.finish_creation()?;
1438 : // The non-test code would call tl.activate() here.
1439 156 : tl.set_state(TimelineState::Active);
1440 156 : Ok(tl)
1441 156 : }
1442 :
1443 : /// Helper for unit tests to create a timeline with some pre-loaded states.
1444 : #[cfg(test)]
1445 : #[allow(clippy::too_many_arguments)]
1446 22 : pub async fn create_test_timeline_with_layers(
1447 22 : &self,
1448 22 : new_timeline_id: TimelineId,
1449 22 : initdb_lsn: Lsn,
1450 22 : pg_version: u32,
1451 22 : ctx: &RequestContext,
1452 22 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
1453 22 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
1454 22 : end_lsn: Lsn,
1455 22 : ) -> anyhow::Result<Arc<Timeline>> {
1456 22 : let tline = self
1457 22 : .create_test_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
1458 44 : .await?;
1459 22 : tline.force_advance_lsn(end_lsn);
1460 62 : for deltas in delta_layer_desc {
1461 40 : tline
1462 40 : .force_create_delta_layer(deltas, Some(initdb_lsn), ctx)
1463 120 : .await?;
1464 : }
1465 60 : for (lsn, images) in image_layer_desc {
1466 38 : tline
1467 38 : .force_create_image_layer(lsn, images, Some(initdb_lsn), ctx)
1468 207 : .await?;
1469 : }
1470 22 : Ok(tline)
1471 22 : }
1472 :
1473 : /// Create a new timeline.
1474 : ///
1475 : /// Returns the new timeline ID and reference to its Timeline object.
1476 : ///
1477 : /// If the caller specified the timeline ID to use (`new_timeline_id`), and timeline with
1478 : /// the same timeline ID already exists, returns CreateTimelineError::AlreadyExists.
1479 : #[allow(clippy::too_many_arguments)]
1480 0 : pub(crate) async fn create_timeline(
1481 0 : self: &Arc<Tenant>,
1482 0 : new_timeline_id: TimelineId,
1483 0 : ancestor_timeline_id: Option<TimelineId>,
1484 0 : mut ancestor_start_lsn: Option<Lsn>,
1485 0 : pg_version: u32,
1486 0 : load_existing_initdb: Option<TimelineId>,
1487 0 : broker_client: storage_broker::BrokerClientChannel,
1488 0 : ctx: &RequestContext,
1489 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
1490 0 : if !self.is_active() {
1491 0 : if matches!(self.current_state(), TenantState::Stopping { .. }) {
1492 0 : return Err(CreateTimelineError::ShuttingDown);
1493 : } else {
1494 0 : return Err(CreateTimelineError::Other(anyhow::anyhow!(
1495 0 : "Cannot create timelines on inactive tenant"
1496 0 : )));
1497 : }
1498 0 : }
1499 :
1500 0 : let _gate = self
1501 0 : .gate
1502 0 : .enter()
1503 0 : .map_err(|_| CreateTimelineError::ShuttingDown)?;
1504 :
1505 : // Get exclusive access to the timeline ID: this ensures that it does not already exist,
1506 : // and that no other creation attempts will be allowed in while we are working.
1507 0 : let create_guard = match self.create_timeline_create_guard(new_timeline_id) {
1508 0 : Ok(m) => m,
1509 : Err(TimelineExclusionError::AlreadyCreating) => {
1510 : // Creation is in progress, we cannot create it again, and we cannot
1511 : // check if this request matches the existing one, so caller must try
1512 : // again later.
1513 0 : return Err(CreateTimelineError::AlreadyCreating);
1514 : }
1515 0 : Err(TimelineExclusionError::Other(e)) => {
1516 0 : return Err(CreateTimelineError::Other(e));
1517 : }
1518 0 : Err(TimelineExclusionError::AlreadyExists(existing)) => {
1519 0 : debug!("timeline {new_timeline_id} already exists");
1520 :
1521 : // Idempotency: creating the same timeline twice is not an error, unless
1522 : // the second creation has different parameters.
1523 0 : if existing.get_ancestor_timeline_id() != ancestor_timeline_id
1524 0 : || existing.pg_version != pg_version
1525 0 : || (ancestor_start_lsn.is_some()
1526 0 : && ancestor_start_lsn != Some(existing.get_ancestor_lsn()))
1527 : {
1528 0 : return Err(CreateTimelineError::Conflict);
1529 0 : }
1530 0 :
1531 0 : // Wait for uploads to complete, so that when we return Ok, the timeline
1532 0 : // is known to be durable on remote storage. Just like we do at the end of
1533 0 : // this function, after we have created the timeline ourselves.
1534 0 : //
1535 0 : // We only really care that the initial version of `index_part.json` has
1536 0 : // been uploaded. That's enough to remember that the timeline
1537 0 : // exists. However, there is no function to wait specifically for that so
1538 0 : // we just wait for all in-progress uploads to finish.
1539 0 : existing
1540 0 : .remote_client
1541 0 : .wait_completion()
1542 0 : .await
1543 0 : .context("wait for timeline uploads to complete")?;
1544 :
1545 0 : return Ok(existing);
1546 : }
1547 : };
1548 :
1549 : pausable_failpoint!("timeline-creation-after-uninit");
1550 :
1551 0 : let loaded_timeline = match ancestor_timeline_id {
1552 0 : Some(ancestor_timeline_id) => {
1553 0 : let ancestor_timeline = self
1554 0 : .get_timeline(ancestor_timeline_id, false)
1555 0 : .context("Cannot branch off the timeline that's not present in pageserver")?;
1556 :
1557 : // instead of waiting around, just deny the request because ancestor is not yet
1558 : // ready for other purposes either.
1559 0 : if !ancestor_timeline.is_active() {
1560 0 : return Err(CreateTimelineError::AncestorNotActive);
1561 0 : }
1562 :
1563 0 : if let Some(lsn) = ancestor_start_lsn.as_mut() {
1564 0 : *lsn = lsn.align();
1565 0 :
1566 0 : let ancestor_ancestor_lsn = ancestor_timeline.get_ancestor_lsn();
1567 0 : if ancestor_ancestor_lsn > *lsn {
1568 : // can we safely just branch from the ancestor instead?
1569 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
1570 0 : "invalid start lsn {} for ancestor timeline {}: less than timeline ancestor lsn {}",
1571 0 : lsn,
1572 0 : ancestor_timeline_id,
1573 0 : ancestor_ancestor_lsn,
1574 0 : )));
1575 0 : }
1576 0 :
1577 0 : // Wait for the WAL to arrive and be processed on the parent branch up
1578 0 : // to the requested branch point. The repository code itself doesn't
1579 0 : // require it, but if we start to receive WAL on the new timeline,
1580 0 : // decoding the new WAL might need to look up previous pages, relation
1581 0 : // sizes etc. and that would get confused if the previous page versions
1582 0 : // are not in the repository yet.
1583 0 : ancestor_timeline
1584 0 : .wait_lsn(*lsn, timeline::WaitLsnWaiter::Tenant, ctx)
1585 0 : .await
1586 0 : .map_err(|e| match e {
1587 0 : e @ (WaitLsnError::Timeout(_) | WaitLsnError::BadState { .. }) => {
1588 0 : CreateTimelineError::AncestorLsn(anyhow::anyhow!(e))
1589 : }
1590 0 : WaitLsnError::Shutdown => CreateTimelineError::ShuttingDown,
1591 0 : })?;
1592 0 : }
1593 :
1594 0 : self.branch_timeline(
1595 0 : &ancestor_timeline,
1596 0 : new_timeline_id,
1597 0 : ancestor_start_lsn,
1598 0 : create_guard,
1599 0 : ctx,
1600 0 : )
1601 0 : .await?
1602 : }
1603 : None => {
1604 0 : self.bootstrap_timeline(
1605 0 : new_timeline_id,
1606 0 : pg_version,
1607 0 : load_existing_initdb,
1608 0 : create_guard,
1609 0 : ctx,
1610 0 : )
1611 0 : .await?
1612 : }
1613 : };
1614 :
1615 : // At this point we have dropped our guard on [`Self::timelines_creating`], and
1616 : // the timeline is visible in [`Self::timelines`], but it is _not_ durable yet. We must
1617 : // not send a success to the caller until it is. The same applies to handling retries,
1618 : // see the handling of [`TimelineExclusionError::AlreadyExists`] above.
1619 0 : let kind = ancestor_timeline_id
1620 0 : .map(|_| "branched")
1621 0 : .unwrap_or("bootstrapped");
1622 0 : loaded_timeline
1623 0 : .remote_client
1624 0 : .wait_completion()
1625 0 : .await
1626 0 : .with_context(|| format!("wait for {} timeline initial uploads to complete", kind))?;
1627 :
1628 0 : loaded_timeline.activate(self.clone(), broker_client, None, ctx);
1629 0 :
1630 0 : Ok(loaded_timeline)
1631 0 : }
1632 :
1633 0 : pub(crate) async fn delete_timeline(
1634 0 : self: Arc<Self>,
1635 0 : timeline_id: TimelineId,
1636 0 : ) -> Result<(), DeleteTimelineError> {
1637 0 : DeleteTimelineFlow::run(&self, timeline_id).await?;
1638 :
1639 0 : Ok(())
1640 0 : }
1641 :
1642 : /// perform one garbage collection iteration, removing old data files from disk.
1643 : /// this function is periodically called by gc task.
1644 : /// also it can be explicitly requested through page server api 'do_gc' command.
1645 : ///
1646 : /// `target_timeline_id` specifies the timeline to GC, or None for all.
1647 : ///
1648 : /// The `horizon` an `pitr` parameters determine how much WAL history needs to be retained.
1649 : /// Also known as the retention period, or the GC cutoff point. `horizon` specifies
1650 : /// the amount of history, as LSN difference from current latest LSN on each timeline.
1651 : /// `pitr` specifies the same as a time difference from the current time. The effective
1652 : /// GC cutoff point is determined conservatively by either `horizon` and `pitr`, whichever
1653 : /// requires more history to be retained.
1654 : //
1655 754 : pub(crate) async fn gc_iteration(
1656 754 : &self,
1657 754 : target_timeline_id: Option<TimelineId>,
1658 754 : horizon: u64,
1659 754 : pitr: Duration,
1660 754 : cancel: &CancellationToken,
1661 754 : ctx: &RequestContext,
1662 754 : ) -> Result<GcResult, GcError> {
1663 754 : // Don't start doing work during shutdown
1664 754 : if let TenantState::Stopping { .. } = self.current_state() {
1665 0 : return Ok(GcResult::default());
1666 754 : }
1667 754 :
1668 754 : // there is a global allowed_error for this
1669 754 : if !self.is_active() {
1670 0 : return Err(GcError::NotActive);
1671 754 : }
1672 754 :
1673 754 : {
1674 754 : let conf = self.tenant_conf.load();
1675 754 :
1676 754 : if !conf.location.may_delete_layers_hint() {
1677 0 : info!("Skipping GC in location state {:?}", conf.location);
1678 0 : return Ok(GcResult::default());
1679 754 : }
1680 754 : }
1681 754 :
1682 754 : self.gc_iteration_internal(target_timeline_id, horizon, pitr, cancel, ctx)
1683 751 : .await
1684 754 : }
1685 :
1686 : /// Perform one compaction iteration.
1687 : /// This function is periodically called by compactor task.
1688 : /// Also it can be explicitly requested per timeline through page server
1689 : /// api's 'compact' command.
1690 : ///
1691 : /// Returns whether we have pending compaction task.
1692 0 : async fn compaction_iteration(
1693 0 : &self,
1694 0 : cancel: &CancellationToken,
1695 0 : ctx: &RequestContext,
1696 0 : ) -> Result<bool, timeline::CompactionError> {
1697 0 : // Don't start doing work during shutdown, or when broken, we do not need those in the logs
1698 0 : if !self.is_active() {
1699 0 : return Ok(false);
1700 0 : }
1701 0 :
1702 0 : {
1703 0 : let conf = self.tenant_conf.load();
1704 0 : if !conf.location.may_delete_layers_hint() || !conf.location.may_upload_layers_hint() {
1705 0 : info!("Skipping compaction in location state {:?}", conf.location);
1706 0 : return Ok(false);
1707 0 : }
1708 0 : }
1709 0 :
1710 0 : // Scan through the hashmap and collect a list of all the timelines,
1711 0 : // while holding the lock. Then drop the lock and actually perform the
1712 0 : // compactions. We don't want to block everything else while the
1713 0 : // compaction runs.
1714 0 : let timelines_to_compact = {
1715 0 : let timelines = self.timelines.lock().unwrap();
1716 0 : let timelines_to_compact = timelines
1717 0 : .iter()
1718 0 : .filter_map(|(timeline_id, timeline)| {
1719 0 : if timeline.is_active() {
1720 0 : Some((*timeline_id, timeline.clone()))
1721 : } else {
1722 0 : None
1723 : }
1724 0 : })
1725 0 : .collect::<Vec<_>>();
1726 0 : drop(timelines);
1727 0 : timelines_to_compact
1728 0 : };
1729 0 :
1730 0 : // Before doing any I/O work, check our circuit breaker
1731 0 : if self.compaction_circuit_breaker.lock().unwrap().is_broken() {
1732 0 : info!("Skipping compaction due to previous failures");
1733 0 : return Ok(false);
1734 0 : }
1735 0 :
1736 0 : let mut has_pending_task = false;
1737 :
1738 0 : for (timeline_id, timeline) in &timelines_to_compact {
1739 0 : has_pending_task |= timeline
1740 0 : .compact(cancel, EnumSet::empty(), ctx)
1741 0 : .instrument(info_span!("compact_timeline", %timeline_id))
1742 0 : .await
1743 0 : .inspect_err(|e| match e {
1744 0 : timeline::CompactionError::ShuttingDown => (),
1745 0 : timeline::CompactionError::Other(e) => {
1746 0 : self.compaction_circuit_breaker
1747 0 : .lock()
1748 0 : .unwrap()
1749 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, e);
1750 0 : }
1751 0 : })?;
1752 : }
1753 :
1754 0 : self.compaction_circuit_breaker
1755 0 : .lock()
1756 0 : .unwrap()
1757 0 : .success(&CIRCUIT_BREAKERS_UNBROKEN);
1758 0 :
1759 0 : Ok(has_pending_task)
1760 0 : }
1761 :
1762 : // Call through to all timelines to freeze ephemeral layers if needed. Usually
1763 : // this happens during ingest: this background housekeeping is for freezing layers
1764 : // that are open but haven't been written to for some time.
1765 0 : async fn ingest_housekeeping(&self) {
1766 0 : // Scan through the hashmap and collect a list of all the timelines,
1767 0 : // while holding the lock. Then drop the lock and actually perform the
1768 0 : // compactions. We don't want to block everything else while the
1769 0 : // compaction runs.
1770 0 : let timelines = {
1771 0 : self.timelines
1772 0 : .lock()
1773 0 : .unwrap()
1774 0 : .values()
1775 0 : .filter_map(|timeline| {
1776 0 : if timeline.is_active() {
1777 0 : Some(timeline.clone())
1778 : } else {
1779 0 : None
1780 : }
1781 0 : })
1782 0 : .collect::<Vec<_>>()
1783 : };
1784 :
1785 0 : for timeline in &timelines {
1786 0 : timeline.maybe_freeze_ephemeral_layer().await;
1787 : }
1788 0 : }
1789 :
1790 2434 : pub fn current_state(&self) -> TenantState {
1791 2434 : self.state.borrow().clone()
1792 2434 : }
1793 :
1794 1674 : pub fn is_active(&self) -> bool {
1795 1674 : self.current_state() == TenantState::Active
1796 1674 : }
1797 :
1798 0 : pub fn generation(&self) -> Generation {
1799 0 : self.generation
1800 0 : }
1801 :
1802 0 : pub(crate) fn wal_redo_manager_status(&self) -> Option<WalRedoManagerStatus> {
1803 0 : self.walredo_mgr.as_ref().and_then(|mgr| mgr.status())
1804 0 : }
1805 :
1806 : /// Changes tenant status to active, unless shutdown was already requested.
1807 : ///
1808 : /// `background_jobs_can_start` is an optional barrier set to a value during pageserver startup
1809 : /// to delay background jobs. Background jobs can be started right away when None is given.
1810 0 : fn activate(
1811 0 : self: &Arc<Self>,
1812 0 : broker_client: BrokerClientChannel,
1813 0 : background_jobs_can_start: Option<&completion::Barrier>,
1814 0 : ctx: &RequestContext,
1815 0 : ) {
1816 0 : span::debug_assert_current_span_has_tenant_id();
1817 0 :
1818 0 : let mut activating = false;
1819 0 : self.state.send_modify(|current_state| {
1820 0 : use pageserver_api::models::ActivatingFrom;
1821 0 : match &*current_state {
1822 : TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
1823 0 : panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {state:?}", state = current_state);
1824 : }
1825 0 : TenantState::Loading => {
1826 0 : *current_state = TenantState::Activating(ActivatingFrom::Loading);
1827 0 : }
1828 0 : TenantState::Attaching => {
1829 0 : *current_state = TenantState::Activating(ActivatingFrom::Attaching);
1830 0 : }
1831 : }
1832 0 : debug!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), "Activating tenant");
1833 0 : activating = true;
1834 0 : // Continue outside the closure. We need to grab timelines.lock()
1835 0 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
1836 0 : });
1837 0 :
1838 0 : if activating {
1839 0 : let timelines_accessor = self.timelines.lock().unwrap();
1840 0 : let timelines_to_activate = timelines_accessor
1841 0 : .values()
1842 0 : .filter(|timeline| !(timeline.is_broken() || timeline.is_stopping()));
1843 0 :
1844 0 : // Before activation, populate each Timeline's GcInfo with information about its children
1845 0 : self.initialize_gc_info(&timelines_accessor);
1846 0 :
1847 0 : // Spawn gc and compaction loops. The loops will shut themselves
1848 0 : // down when they notice that the tenant is inactive.
1849 0 : tasks::start_background_loops(self, background_jobs_can_start);
1850 0 :
1851 0 : let mut activated_timelines = 0;
1852 :
1853 0 : for timeline in timelines_to_activate {
1854 0 : timeline.activate(
1855 0 : self.clone(),
1856 0 : broker_client.clone(),
1857 0 : background_jobs_can_start,
1858 0 : ctx,
1859 0 : );
1860 0 : activated_timelines += 1;
1861 0 : }
1862 :
1863 0 : self.state.send_modify(move |current_state| {
1864 0 : assert!(
1865 0 : matches!(current_state, TenantState::Activating(_)),
1866 0 : "set_stopping and set_broken wait for us to leave Activating state",
1867 : );
1868 0 : *current_state = TenantState::Active;
1869 0 :
1870 0 : let elapsed = self.constructed_at.elapsed();
1871 0 : let total_timelines = timelines_accessor.len();
1872 0 :
1873 0 : // log a lot of stuff, because some tenants sometimes suffer from user-visible
1874 0 : // times to activate. see https://github.com/neondatabase/neon/issues/4025
1875 0 : info!(
1876 0 : since_creation_millis = elapsed.as_millis(),
1877 0 : tenant_id = %self.tenant_shard_id.tenant_id,
1878 0 : shard_id = %self.tenant_shard_id.shard_slug(),
1879 0 : activated_timelines,
1880 0 : total_timelines,
1881 0 : post_state = <&'static str>::from(&*current_state),
1882 0 : "activation attempt finished"
1883 : );
1884 :
1885 0 : TENANT.activation.observe(elapsed.as_secs_f64());
1886 0 : });
1887 0 : }
1888 0 : }
1889 :
1890 : /// Shutdown the tenant and join all of the spawned tasks.
1891 : ///
1892 : /// The method caters for all use-cases:
1893 : /// - pageserver shutdown (freeze_and_flush == true)
1894 : /// - detach + ignore (freeze_and_flush == false)
1895 : ///
1896 : /// This will attempt to shutdown even if tenant is broken.
1897 : ///
1898 : /// `shutdown_progress` is a [`completion::Barrier`] for the shutdown initiated by this call.
1899 : /// If the tenant is already shutting down, we return a clone of the first shutdown call's
1900 : /// `Barrier` as an `Err`. This not-first caller can use the returned barrier to join with
1901 : /// the ongoing shutdown.
1902 6 : async fn shutdown(
1903 6 : &self,
1904 6 : shutdown_progress: completion::Barrier,
1905 6 : shutdown_mode: timeline::ShutdownMode,
1906 6 : ) -> Result<(), completion::Barrier> {
1907 6 : span::debug_assert_current_span_has_tenant_id();
1908 :
1909 : // Set tenant (and its timlines) to Stoppping state.
1910 : //
1911 : // Since we can only transition into Stopping state after activation is complete,
1912 : // run it in a JoinSet so all tenants have a chance to stop before we get SIGKILLed.
1913 : //
1914 : // Transitioning tenants to Stopping state has a couple of non-obvious side effects:
1915 : // 1. Lock out any new requests to the tenants.
1916 : // 2. Signal cancellation to WAL receivers (we wait on it below).
1917 : // 3. Signal cancellation for other tenant background loops.
1918 : // 4. ???
1919 : //
1920 : // The waiting for the cancellation is not done uniformly.
1921 : // We certainly wait for WAL receivers to shut down.
1922 : // That is necessary so that no new data comes in before the freeze_and_flush.
1923 : // But the tenant background loops are joined-on in our caller.
1924 : // It's mesed up.
1925 : // we just ignore the failure to stop
1926 :
1927 : // If we're still attaching, fire the cancellation token early to drop out: this
1928 : // will prevent us flushing, but ensures timely shutdown if some I/O during attach
1929 : // is very slow.
1930 6 : let shutdown_mode = if matches!(self.current_state(), TenantState::Attaching) {
1931 0 : self.cancel.cancel();
1932 0 :
1933 0 : // Having fired our cancellation token, do not try and flush timelines: their cancellation tokens
1934 0 : // are children of ours, so their flush loops will have shut down already
1935 0 : timeline::ShutdownMode::Hard
1936 : } else {
1937 6 : shutdown_mode
1938 : };
1939 :
1940 6 : match self.set_stopping(shutdown_progress, false, false).await {
1941 6 : Ok(()) => {}
1942 0 : Err(SetStoppingError::Broken) => {
1943 0 : // assume that this is acceptable
1944 0 : }
1945 0 : Err(SetStoppingError::AlreadyStopping(other)) => {
1946 0 : // give caller the option to wait for this this shutdown
1947 0 : info!("Tenant::shutdown: AlreadyStopping");
1948 0 : return Err(other);
1949 : }
1950 : };
1951 :
1952 6 : let mut js = tokio::task::JoinSet::new();
1953 6 : {
1954 6 : let timelines = self.timelines.lock().unwrap();
1955 6 : timelines.values().for_each(|timeline| {
1956 6 : let timeline = Arc::clone(timeline);
1957 6 : let timeline_id = timeline.timeline_id;
1958 6 : let span = tracing::info_span!("timeline_shutdown", %timeline_id, ?shutdown_mode);
1959 11 : js.spawn(async move { timeline.shutdown(shutdown_mode).instrument(span).await });
1960 6 : })
1961 6 : };
1962 6 : // test_long_timeline_create_then_tenant_delete is leaning on this message
1963 6 : tracing::info!("Waiting for timelines...");
1964 12 : while let Some(res) = js.join_next().await {
1965 0 : match res {
1966 6 : Ok(()) => {}
1967 0 : Err(je) if je.is_cancelled() => unreachable!("no cancelling used"),
1968 0 : Err(je) if je.is_panic() => { /* logged already */ }
1969 0 : Err(je) => warn!("unexpected JoinError: {je:?}"),
1970 : }
1971 : }
1972 :
1973 : // We cancel the Tenant's cancellation token _after_ the timelines have all shut down. This permits
1974 : // them to continue to do work during their shutdown methods, e.g. flushing data.
1975 6 : tracing::debug!("Cancelling CancellationToken");
1976 6 : self.cancel.cancel();
1977 6 :
1978 6 : // shutdown all tenant and timeline tasks: gc, compaction, page service
1979 6 : // No new tasks will be started for this tenant because it's in `Stopping` state.
1980 6 : //
1981 6 : // this will additionally shutdown and await all timeline tasks.
1982 6 : tracing::debug!("Waiting for tasks...");
1983 6 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), None).await;
1984 :
1985 6 : if let Some(walredo_mgr) = self.walredo_mgr.as_ref() {
1986 6 : walredo_mgr.shutdown().await;
1987 0 : }
1988 :
1989 : // Wait for any in-flight operations to complete
1990 6 : self.gate.close().await;
1991 :
1992 6 : remove_tenant_metrics(&self.tenant_shard_id);
1993 6 :
1994 6 : Ok(())
1995 6 : }
1996 :
1997 : /// Change tenant status to Stopping, to mark that it is being shut down.
1998 : ///
1999 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
2000 : ///
2001 : /// This function is not cancel-safe!
2002 : ///
2003 : /// `allow_transition_from_loading` is needed for the special case of loading task deleting the tenant.
2004 : /// `allow_transition_from_attaching` is needed for the special case of attaching deleted tenant.
2005 6 : async fn set_stopping(
2006 6 : &self,
2007 6 : progress: completion::Barrier,
2008 6 : allow_transition_from_loading: bool,
2009 6 : allow_transition_from_attaching: bool,
2010 6 : ) -> Result<(), SetStoppingError> {
2011 6 : let mut rx = self.state.subscribe();
2012 6 :
2013 6 : // cannot stop before we're done activating, so wait out until we're done activating
2014 6 : rx.wait_for(|state| match state {
2015 0 : TenantState::Attaching if allow_transition_from_attaching => true,
2016 : TenantState::Activating(_) | TenantState::Attaching => {
2017 0 : info!(
2018 0 : "waiting for {} to turn Active|Broken|Stopping",
2019 0 : <&'static str>::from(state)
2020 : );
2021 0 : false
2022 : }
2023 0 : TenantState::Loading => allow_transition_from_loading,
2024 6 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
2025 6 : })
2026 0 : .await
2027 6 : .expect("cannot drop self.state while on a &self method");
2028 6 :
2029 6 : // we now know we're done activating, let's see whether this task is the winner to transition into Stopping
2030 6 : let mut err = None;
2031 6 : let stopping = self.state.send_if_modified(|current_state| match current_state {
2032 : TenantState::Activating(_) => {
2033 0 : unreachable!("1we ensured above that we're done with activation, and, there is no re-activation")
2034 : }
2035 : TenantState::Attaching => {
2036 0 : if !allow_transition_from_attaching {
2037 0 : unreachable!("2we ensured above that we're done with activation, and, there is no re-activation")
2038 0 : };
2039 0 : *current_state = TenantState::Stopping { progress };
2040 0 : true
2041 : }
2042 : TenantState::Loading => {
2043 0 : if !allow_transition_from_loading {
2044 0 : unreachable!("3we ensured above that we're done with activation, and, there is no re-activation")
2045 0 : };
2046 0 : *current_state = TenantState::Stopping { progress };
2047 0 : true
2048 : }
2049 : TenantState::Active => {
2050 : // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
2051 : // are created after the transition to Stopping. That's harmless, as the Timelines
2052 : // won't be accessible to anyone afterwards, because the Tenant is in Stopping state.
2053 6 : *current_state = TenantState::Stopping { progress };
2054 6 : // Continue stopping outside the closure. We need to grab timelines.lock()
2055 6 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
2056 6 : true
2057 : }
2058 0 : TenantState::Broken { reason, .. } => {
2059 0 : info!(
2060 0 : "Cannot set tenant to Stopping state, it is in Broken state due to: {reason}"
2061 : );
2062 0 : err = Some(SetStoppingError::Broken);
2063 0 : false
2064 : }
2065 0 : TenantState::Stopping { progress } => {
2066 0 : info!("Tenant is already in Stopping state");
2067 0 : err = Some(SetStoppingError::AlreadyStopping(progress.clone()));
2068 0 : false
2069 : }
2070 6 : });
2071 6 : match (stopping, err) {
2072 6 : (true, None) => {} // continue
2073 0 : (false, Some(err)) => return Err(err),
2074 0 : (true, Some(_)) => unreachable!(
2075 0 : "send_if_modified closure must error out if not transitioning to Stopping"
2076 0 : ),
2077 0 : (false, None) => unreachable!(
2078 0 : "send_if_modified closure must return true if transitioning to Stopping"
2079 0 : ),
2080 : }
2081 :
2082 6 : let timelines_accessor = self.timelines.lock().unwrap();
2083 6 : let not_broken_timelines = timelines_accessor
2084 6 : .values()
2085 6 : .filter(|timeline| !timeline.is_broken());
2086 12 : for timeline in not_broken_timelines {
2087 6 : timeline.set_state(TimelineState::Stopping);
2088 6 : }
2089 6 : Ok(())
2090 6 : }
2091 :
2092 : /// Method for tenant::mgr to transition us into Broken state in case of a late failure in
2093 : /// `remove_tenant_from_memory`
2094 : ///
2095 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
2096 : ///
2097 : /// In tests, we also use this to set tenants to Broken state on purpose.
2098 0 : pub(crate) async fn set_broken(&self, reason: String) {
2099 0 : let mut rx = self.state.subscribe();
2100 0 :
2101 0 : // The load & attach routines own the tenant state until it has reached `Active`.
2102 0 : // So, wait until it's done.
2103 0 : rx.wait_for(|state| match state {
2104 : TenantState::Activating(_) | TenantState::Loading | TenantState::Attaching => {
2105 0 : info!(
2106 0 : "waiting for {} to turn Active|Broken|Stopping",
2107 0 : <&'static str>::from(state)
2108 : );
2109 0 : false
2110 : }
2111 0 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
2112 0 : })
2113 0 : .await
2114 0 : .expect("cannot drop self.state while on a &self method");
2115 0 :
2116 0 : // we now know we're done activating, let's see whether this task is the winner to transition into Broken
2117 0 : self.set_broken_no_wait(reason)
2118 0 : }
2119 :
2120 0 : pub(crate) fn set_broken_no_wait(&self, reason: impl Display) {
2121 0 : let reason = reason.to_string();
2122 0 : self.state.send_modify(|current_state| {
2123 0 : match *current_state {
2124 : TenantState::Activating(_) | TenantState::Loading | TenantState::Attaching => {
2125 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
2126 : }
2127 : TenantState::Active => {
2128 0 : if cfg!(feature = "testing") {
2129 0 : warn!("Changing Active tenant to Broken state, reason: {}", reason);
2130 0 : *current_state = TenantState::broken_from_reason(reason);
2131 : } else {
2132 0 : unreachable!("not allowed to call set_broken on Active tenants in non-testing builds")
2133 : }
2134 : }
2135 : TenantState::Broken { .. } => {
2136 0 : warn!("Tenant is already in Broken state");
2137 : }
2138 : // This is the only "expected" path, any other path is a bug.
2139 : TenantState::Stopping { .. } => {
2140 0 : warn!(
2141 0 : "Marking Stopping tenant as Broken state, reason: {}",
2142 : reason
2143 : );
2144 0 : *current_state = TenantState::broken_from_reason(reason);
2145 : }
2146 : }
2147 0 : });
2148 0 : }
2149 :
2150 0 : pub fn subscribe_for_state_updates(&self) -> watch::Receiver<TenantState> {
2151 0 : self.state.subscribe()
2152 0 : }
2153 :
2154 : /// The activate_now semaphore is initialized with zero units. As soon as
2155 : /// we add a unit, waiters will be able to acquire a unit and proceed.
2156 0 : pub(crate) fn activate_now(&self) {
2157 0 : self.activate_now_sem.add_permits(1);
2158 0 : }
2159 :
2160 0 : pub(crate) async fn wait_to_become_active(
2161 0 : &self,
2162 0 : timeout: Duration,
2163 0 : ) -> Result<(), GetActiveTenantError> {
2164 0 : let mut receiver = self.state.subscribe();
2165 0 : loop {
2166 0 : let current_state = receiver.borrow_and_update().clone();
2167 0 : match current_state {
2168 : TenantState::Loading | TenantState::Attaching | TenantState::Activating(_) => {
2169 : // in these states, there's a chance that we can reach ::Active
2170 0 : self.activate_now();
2171 0 : match timeout_cancellable(timeout, &self.cancel, receiver.changed()).await {
2172 0 : Ok(r) => {
2173 0 : r.map_err(
2174 0 : |_e: tokio::sync::watch::error::RecvError|
2175 : // Tenant existed but was dropped: report it as non-existent
2176 0 : GetActiveTenantError::NotFound(GetTenantError::NotFound(self.tenant_shard_id.tenant_id))
2177 0 : )?
2178 : }
2179 : Err(TimeoutCancellableError::Cancelled) => {
2180 0 : return Err(GetActiveTenantError::Cancelled);
2181 : }
2182 : Err(TimeoutCancellableError::Timeout) => {
2183 0 : return Err(GetActiveTenantError::WaitForActiveTimeout {
2184 0 : latest_state: Some(self.current_state()),
2185 0 : wait_time: timeout,
2186 0 : });
2187 : }
2188 : }
2189 : }
2190 : TenantState::Active { .. } => {
2191 0 : return Ok(());
2192 : }
2193 0 : TenantState::Broken { reason, .. } => {
2194 0 : // This is fatal, and reported distinctly from the general case of "will never be active" because
2195 0 : // it's logically a 500 to external API users (broken is always a bug).
2196 0 : return Err(GetActiveTenantError::Broken(reason));
2197 : }
2198 : TenantState::Stopping { .. } => {
2199 : // There's no chance the tenant can transition back into ::Active
2200 0 : return Err(GetActiveTenantError::WillNotBecomeActive(current_state));
2201 : }
2202 : }
2203 : }
2204 0 : }
2205 :
2206 0 : pub(crate) fn get_attach_mode(&self) -> AttachmentMode {
2207 0 : self.tenant_conf.load().location.attach_mode
2208 0 : }
2209 :
2210 : /// For API access: generate a LocationConfig equivalent to the one that would be used to
2211 : /// create a Tenant in the same state. Do not use this in hot paths: it's for relatively
2212 : /// rare external API calls, like a reconciliation at startup.
2213 0 : pub(crate) fn get_location_conf(&self) -> models::LocationConfig {
2214 0 : let conf = self.tenant_conf.load();
2215 :
2216 0 : let location_config_mode = match conf.location.attach_mode {
2217 0 : AttachmentMode::Single => models::LocationConfigMode::AttachedSingle,
2218 0 : AttachmentMode::Multi => models::LocationConfigMode::AttachedMulti,
2219 0 : AttachmentMode::Stale => models::LocationConfigMode::AttachedStale,
2220 : };
2221 :
2222 : // We have a pageserver TenantConf, we need the API-facing TenantConfig.
2223 0 : let tenant_config: models::TenantConfig = conf.tenant_conf.clone().into();
2224 0 :
2225 0 : models::LocationConfig {
2226 0 : mode: location_config_mode,
2227 0 : generation: self.generation.into(),
2228 0 : secondary_conf: None,
2229 0 : shard_number: self.shard_identity.number.0,
2230 0 : shard_count: self.shard_identity.count.literal(),
2231 0 : shard_stripe_size: self.shard_identity.stripe_size.0,
2232 0 : tenant_conf: tenant_config,
2233 0 : }
2234 0 : }
2235 :
2236 0 : pub(crate) fn get_tenant_shard_id(&self) -> &TenantShardId {
2237 0 : &self.tenant_shard_id
2238 0 : }
2239 :
2240 0 : pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
2241 0 : self.shard_identity.stripe_size
2242 0 : }
2243 :
2244 0 : pub(crate) fn get_generation(&self) -> Generation {
2245 0 : self.generation
2246 0 : }
2247 :
2248 : /// This function partially shuts down the tenant (it shuts down the Timelines) and is fallible,
2249 : /// and can leave the tenant in a bad state if it fails. The caller is responsible for
2250 : /// resetting this tenant to a valid state if we fail.
2251 0 : pub(crate) async fn split_prepare(
2252 0 : &self,
2253 0 : child_shards: &Vec<TenantShardId>,
2254 0 : ) -> anyhow::Result<()> {
2255 0 : let timelines = self.timelines.lock().unwrap().clone();
2256 0 : for timeline in timelines.values() {
2257 : // We do not block timeline creation/deletion during splits inside the pageserver: it is up to higher levels
2258 : // to ensure that they do not start a split if currently in the process of doing these.
2259 :
2260 : // Upload an index from the parent: this is partly to provide freshness for the
2261 : // child tenants that will copy it, and partly for general ease-of-debugging: there will
2262 : // always be a parent shard index in the same generation as we wrote the child shard index.
2263 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Uploading index");
2264 0 : timeline
2265 0 : .remote_client
2266 0 : .schedule_index_upload_for_file_changes()?;
2267 0 : timeline.remote_client.wait_completion().await?;
2268 :
2269 : // Shut down the timeline's remote client: this means that the indices we write
2270 : // for child shards will not be invalidated by the parent shard deleting layers.
2271 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Shutting down remote storage client");
2272 0 : timeline.remote_client.shutdown().await;
2273 :
2274 : // Download methods can still be used after shutdown, as they don't flow through the remote client's
2275 : // queue. In principal the RemoteTimelineClient could provide this without downloading it, but this
2276 : // operation is rare, so it's simpler to just download it (and robustly guarantees that the index
2277 : // we use here really is the remotely persistent one).
2278 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Downloading index_part from parent");
2279 0 : let result = timeline.remote_client
2280 0 : .download_index_file(&self.cancel)
2281 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))
2282 0 : .await?;
2283 0 : let index_part = match result {
2284 : MaybeDeletedIndexPart::Deleted(_) => {
2285 0 : anyhow::bail!("Timeline deletion happened concurrently with split")
2286 : }
2287 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
2288 : };
2289 :
2290 0 : for child_shard in child_shards {
2291 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Uploading index_part for child {}", child_shard.to_index());
2292 0 : upload_index_part(
2293 0 : &self.remote_storage,
2294 0 : child_shard,
2295 0 : &timeline.timeline_id,
2296 0 : self.generation,
2297 0 : &index_part,
2298 0 : &self.cancel,
2299 0 : )
2300 0 : .await?;
2301 : }
2302 : }
2303 :
2304 0 : Ok(())
2305 0 : }
2306 :
2307 0 : pub(crate) fn get_sizes(&self) -> TopTenantShardItem {
2308 0 : let mut result = TopTenantShardItem {
2309 0 : id: self.tenant_shard_id,
2310 0 : resident_size: 0,
2311 0 : physical_size: 0,
2312 0 : max_logical_size: 0,
2313 0 : };
2314 :
2315 0 : for timeline in self.timelines.lock().unwrap().values() {
2316 0 : result.resident_size += timeline.metrics.resident_physical_size_gauge.get();
2317 0 :
2318 0 : result.physical_size += timeline
2319 0 : .remote_client
2320 0 : .metrics
2321 0 : .remote_physical_size_gauge
2322 0 : .get();
2323 0 : result.max_logical_size = std::cmp::max(
2324 0 : result.max_logical_size,
2325 0 : timeline.metrics.current_logical_size_gauge.get(),
2326 0 : );
2327 0 : }
2328 :
2329 0 : result
2330 0 : }
2331 : }
2332 :
2333 : /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
2334 : /// perform a topological sort, so that the parent of each timeline comes
2335 : /// before the children.
2336 : /// E extracts the ancestor from T
2337 : /// This allows for T to be different. It can be TimelineMetadata, can be Timeline itself, etc.
2338 174 : fn tree_sort_timelines<T, E>(
2339 174 : timelines: HashMap<TimelineId, T>,
2340 174 : extractor: E,
2341 174 : ) -> anyhow::Result<Vec<(TimelineId, T)>>
2342 174 : where
2343 174 : E: Fn(&T) -> Option<TimelineId>,
2344 174 : {
2345 174 : let mut result = Vec::with_capacity(timelines.len());
2346 174 :
2347 174 : let mut now = Vec::with_capacity(timelines.len());
2348 174 : // (ancestor, children)
2349 174 : let mut later: HashMap<TimelineId, Vec<(TimelineId, T)>> =
2350 174 : HashMap::with_capacity(timelines.len());
2351 :
2352 180 : for (timeline_id, value) in timelines {
2353 6 : if let Some(ancestor_id) = extractor(&value) {
2354 2 : let children = later.entry(ancestor_id).or_default();
2355 2 : children.push((timeline_id, value));
2356 4 : } else {
2357 4 : now.push((timeline_id, value));
2358 4 : }
2359 : }
2360 :
2361 180 : while let Some((timeline_id, metadata)) = now.pop() {
2362 6 : result.push((timeline_id, metadata));
2363 : // All children of this can be loaded now
2364 6 : if let Some(mut children) = later.remove(&timeline_id) {
2365 2 : now.append(&mut children);
2366 4 : }
2367 : }
2368 :
2369 : // All timelines should be visited now. Unless there were timelines with missing ancestors.
2370 174 : if !later.is_empty() {
2371 0 : for (missing_id, orphan_ids) in later {
2372 0 : for (orphan_id, _) in orphan_ids {
2373 0 : error!("could not load timeline {orphan_id} because its ancestor timeline {missing_id} could not be loaded");
2374 : }
2375 : }
2376 0 : bail!("could not load tenant because some timelines are missing ancestors");
2377 174 : }
2378 174 :
2379 174 : Ok(result)
2380 174 : }
2381 :
2382 : impl Tenant {
2383 0 : pub fn tenant_specific_overrides(&self) -> TenantConfOpt {
2384 0 : self.tenant_conf.load().tenant_conf.clone()
2385 0 : }
2386 :
2387 0 : pub fn effective_config(&self) -> TenantConf {
2388 0 : self.tenant_specific_overrides()
2389 0 : .merge(self.conf.default_tenant_conf.clone())
2390 0 : }
2391 :
2392 0 : pub fn get_checkpoint_distance(&self) -> u64 {
2393 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2394 0 : tenant_conf
2395 0 : .checkpoint_distance
2396 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
2397 0 : }
2398 :
2399 0 : pub fn get_checkpoint_timeout(&self) -> Duration {
2400 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2401 0 : tenant_conf
2402 0 : .checkpoint_timeout
2403 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
2404 0 : }
2405 :
2406 0 : pub fn get_compaction_target_size(&self) -> u64 {
2407 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2408 0 : tenant_conf
2409 0 : .compaction_target_size
2410 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
2411 0 : }
2412 :
2413 0 : pub fn get_compaction_period(&self) -> Duration {
2414 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2415 0 : tenant_conf
2416 0 : .compaction_period
2417 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
2418 0 : }
2419 :
2420 0 : pub fn get_compaction_threshold(&self) -> usize {
2421 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2422 0 : tenant_conf
2423 0 : .compaction_threshold
2424 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
2425 0 : }
2426 :
2427 0 : pub fn get_gc_horizon(&self) -> u64 {
2428 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2429 0 : tenant_conf
2430 0 : .gc_horizon
2431 0 : .unwrap_or(self.conf.default_tenant_conf.gc_horizon)
2432 0 : }
2433 :
2434 0 : pub fn get_gc_period(&self) -> Duration {
2435 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2436 0 : tenant_conf
2437 0 : .gc_period
2438 0 : .unwrap_or(self.conf.default_tenant_conf.gc_period)
2439 0 : }
2440 :
2441 0 : pub fn get_image_creation_threshold(&self) -> usize {
2442 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2443 0 : tenant_conf
2444 0 : .image_creation_threshold
2445 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
2446 0 : }
2447 :
2448 0 : pub fn get_pitr_interval(&self) -> Duration {
2449 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2450 0 : tenant_conf
2451 0 : .pitr_interval
2452 0 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
2453 0 : }
2454 :
2455 0 : pub fn get_min_resident_size_override(&self) -> Option<u64> {
2456 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2457 0 : tenant_conf
2458 0 : .min_resident_size_override
2459 0 : .or(self.conf.default_tenant_conf.min_resident_size_override)
2460 0 : }
2461 :
2462 0 : pub fn get_heatmap_period(&self) -> Option<Duration> {
2463 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2464 0 : let heatmap_period = tenant_conf
2465 0 : .heatmap_period
2466 0 : .unwrap_or(self.conf.default_tenant_conf.heatmap_period);
2467 0 : if heatmap_period.is_zero() {
2468 0 : None
2469 : } else {
2470 0 : Some(heatmap_period)
2471 : }
2472 0 : }
2473 :
2474 0 : pub fn get_lsn_lease_length(&self) -> Duration {
2475 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2476 0 : tenant_conf
2477 0 : .lsn_lease_length
2478 0 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
2479 0 : }
2480 :
2481 0 : pub fn set_new_tenant_config(&self, new_tenant_conf: TenantConfOpt) {
2482 0 : // Use read-copy-update in order to avoid overwriting the location config
2483 0 : // state if this races with [`Tenant::set_new_location_config`]. Note that
2484 0 : // this race is not possible if both request types come from the storage
2485 0 : // controller (as they should!) because an exclusive op lock is required
2486 0 : // on the storage controller side.
2487 0 : self.tenant_conf.rcu(|inner| {
2488 0 : Arc::new(AttachedTenantConf {
2489 0 : tenant_conf: new_tenant_conf.clone(),
2490 0 : location: inner.location,
2491 0 : })
2492 0 : });
2493 0 :
2494 0 : self.tenant_conf_updated(&new_tenant_conf);
2495 0 : // Don't hold self.timelines.lock() during the notifies.
2496 0 : // There's no risk of deadlock right now, but there could be if we consolidate
2497 0 : // mutexes in struct Timeline in the future.
2498 0 : let timelines = self.list_timelines();
2499 0 : for timeline in timelines {
2500 0 : timeline.tenant_conf_updated(&new_tenant_conf);
2501 0 : }
2502 0 : }
2503 :
2504 8 : pub(crate) fn set_new_location_config(&self, new_conf: AttachedTenantConf) {
2505 8 : let new_tenant_conf = new_conf.tenant_conf.clone();
2506 8 :
2507 8 : self.tenant_conf.store(Arc::new(new_conf));
2508 8 :
2509 8 : self.tenant_conf_updated(&new_tenant_conf);
2510 8 : // Don't hold self.timelines.lock() during the notifies.
2511 8 : // There's no risk of deadlock right now, but there could be if we consolidate
2512 8 : // mutexes in struct Timeline in the future.
2513 8 : let timelines = self.list_timelines();
2514 16 : for timeline in timelines {
2515 8 : timeline.tenant_conf_updated(&new_tenant_conf);
2516 8 : }
2517 8 : }
2518 :
2519 182 : fn get_timeline_get_throttle_config(
2520 182 : psconf: &'static PageServerConf,
2521 182 : overrides: &TenantConfOpt,
2522 182 : ) -> throttle::Config {
2523 182 : overrides
2524 182 : .timeline_get_throttle
2525 182 : .clone()
2526 182 : .unwrap_or(psconf.default_tenant_conf.timeline_get_throttle.clone())
2527 182 : }
2528 :
2529 8 : pub(crate) fn tenant_conf_updated(&self, new_conf: &TenantConfOpt) {
2530 8 : let conf = Self::get_timeline_get_throttle_config(self.conf, new_conf);
2531 8 : self.timeline_get_throttle.reconfigure(conf)
2532 8 : }
2533 :
2534 : /// Helper function to create a new Timeline struct.
2535 : ///
2536 : /// The returned Timeline is in Loading state. The caller is responsible for
2537 : /// initializing any on-disk state, and for inserting the Timeline to the 'timelines'
2538 : /// map.
2539 : ///
2540 : /// `validate_ancestor == false` is used when a timeline is created for deletion
2541 : /// and we might not have the ancestor present anymore which is fine for to be
2542 : /// deleted timelines.
2543 398 : fn create_timeline_struct(
2544 398 : &self,
2545 398 : new_timeline_id: TimelineId,
2546 398 : new_metadata: &TimelineMetadata,
2547 398 : ancestor: Option<Arc<Timeline>>,
2548 398 : resources: TimelineResources,
2549 398 : cause: CreateTimelineCause,
2550 398 : last_aux_file_policy: Option<AuxFilePolicy>,
2551 398 : ) -> anyhow::Result<Arc<Timeline>> {
2552 398 : let state = match cause {
2553 : CreateTimelineCause::Load => {
2554 398 : let ancestor_id = new_metadata.ancestor_timeline();
2555 398 : anyhow::ensure!(
2556 398 : ancestor_id == ancestor.as_ref().map(|t| t.timeline_id),
2557 0 : "Timeline's {new_timeline_id} ancestor {ancestor_id:?} was not found"
2558 : );
2559 398 : TimelineState::Loading
2560 : }
2561 0 : CreateTimelineCause::Delete => TimelineState::Stopping,
2562 : };
2563 :
2564 398 : let pg_version = new_metadata.pg_version();
2565 398 :
2566 398 : let timeline = Timeline::new(
2567 398 : self.conf,
2568 398 : Arc::clone(&self.tenant_conf),
2569 398 : new_metadata,
2570 398 : ancestor,
2571 398 : new_timeline_id,
2572 398 : self.tenant_shard_id,
2573 398 : self.generation,
2574 398 : self.shard_identity,
2575 398 : self.walredo_mgr.clone(),
2576 398 : resources,
2577 398 : pg_version,
2578 398 : state,
2579 398 : last_aux_file_policy,
2580 398 : self.cancel.child_token(),
2581 398 : );
2582 398 :
2583 398 : Ok(timeline)
2584 398 : }
2585 :
2586 : // Allow too_many_arguments because a constructor's argument list naturally grows with the
2587 : // number of attributes in the struct: breaking these out into a builder wouldn't be helpful.
2588 : #[allow(clippy::too_many_arguments)]
2589 174 : fn new(
2590 174 : state: TenantState,
2591 174 : conf: &'static PageServerConf,
2592 174 : attached_conf: AttachedTenantConf,
2593 174 : shard_identity: ShardIdentity,
2594 174 : walredo_mgr: Option<Arc<WalRedoManager>>,
2595 174 : tenant_shard_id: TenantShardId,
2596 174 : remote_storage: GenericRemoteStorage,
2597 174 : deletion_queue_client: DeletionQueueClient,
2598 174 : l0_flush_global_state: L0FlushGlobalState,
2599 174 : ) -> Tenant {
2600 174 : debug_assert!(
2601 174 : !attached_conf.location.generation.is_none() || conf.control_plane_api.is_none()
2602 : );
2603 :
2604 174 : let (state, mut rx) = watch::channel(state);
2605 174 :
2606 174 : tokio::spawn(async move {
2607 174 : // reflect tenant state in metrics:
2608 174 : // - global per tenant state: TENANT_STATE_METRIC
2609 174 : // - "set" of broken tenants: BROKEN_TENANTS_SET
2610 174 : //
2611 174 : // set of broken tenants should not have zero counts so that it remains accessible for
2612 174 : // alerting.
2613 174 :
2614 174 : let tid = tenant_shard_id.to_string();
2615 174 : let shard_id = tenant_shard_id.shard_slug().to_string();
2616 174 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
2617 174 :
2618 346 : fn inspect_state(state: &TenantState) -> ([&'static str; 1], bool) {
2619 346 : ([state.into()], matches!(state, TenantState::Broken { .. }))
2620 346 : }
2621 174 :
2622 174 : let mut tuple = inspect_state(&rx.borrow_and_update());
2623 174 :
2624 174 : let is_broken = tuple.1;
2625 174 : let mut counted_broken = if is_broken {
2626 : // add the id to the set right away, there should not be any updates on the channel
2627 : // after before tenant is removed, if ever
2628 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
2629 0 : true
2630 : } else {
2631 174 : false
2632 : };
2633 :
2634 346 : loop {
2635 346 : let labels = &tuple.0;
2636 346 : let current = TENANT_STATE_METRIC.with_label_values(labels);
2637 346 : current.inc();
2638 346 :
2639 346 : if rx.changed().await.is_err() {
2640 : // tenant has been dropped
2641 16 : current.dec();
2642 16 : drop(BROKEN_TENANTS_SET.remove_label_values(set_key));
2643 16 : break;
2644 172 : }
2645 172 :
2646 172 : current.dec();
2647 172 : tuple = inspect_state(&rx.borrow_and_update());
2648 172 :
2649 172 : let is_broken = tuple.1;
2650 172 : if is_broken && !counted_broken {
2651 0 : counted_broken = true;
2652 0 : // insert the tenant_id (back) into the set while avoiding needless counter
2653 0 : // access
2654 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
2655 172 : }
2656 : }
2657 174 : });
2658 174 :
2659 174 : Tenant {
2660 174 : tenant_shard_id,
2661 174 : shard_identity,
2662 174 : generation: attached_conf.location.generation,
2663 174 : conf,
2664 174 : // using now here is good enough approximation to catch tenants with really long
2665 174 : // activation times.
2666 174 : constructed_at: Instant::now(),
2667 174 : timelines: Mutex::new(HashMap::new()),
2668 174 : timelines_creating: Mutex::new(HashSet::new()),
2669 174 : gc_cs: tokio::sync::Mutex::new(()),
2670 174 : walredo_mgr,
2671 174 : remote_storage,
2672 174 : deletion_queue_client,
2673 174 : state,
2674 174 : cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
2675 174 : cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
2676 174 : eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
2677 174 : compaction_circuit_breaker: std::sync::Mutex::new(CircuitBreaker::new(
2678 174 : format!("compaction-{tenant_shard_id}"),
2679 174 : 5,
2680 174 : // Compaction can be a very expensive operation, and might leak disk space. It also ought
2681 174 : // to be infallible, as long as remote storage is available. So if it repeatedly fails,
2682 174 : // use an extremely long backoff.
2683 174 : Some(Duration::from_secs(3600 * 24)),
2684 174 : )),
2685 174 : activate_now_sem: tokio::sync::Semaphore::new(0),
2686 174 : cancel: CancellationToken::default(),
2687 174 : gate: Gate::default(),
2688 174 : timeline_get_throttle: Arc::new(throttle::Throttle::new(
2689 174 : Tenant::get_timeline_get_throttle_config(conf, &attached_conf.tenant_conf),
2690 174 : &crate::metrics::tenant_throttling::TIMELINE_GET,
2691 174 : )),
2692 174 : tenant_conf: Arc::new(ArcSwap::from_pointee(attached_conf)),
2693 174 : ongoing_timeline_detach: std::sync::Mutex::default(),
2694 174 : l0_flush_global_state,
2695 174 : }
2696 174 : }
2697 :
2698 : /// Locate and load config
2699 0 : pub(super) fn load_tenant_config(
2700 0 : conf: &'static PageServerConf,
2701 0 : tenant_shard_id: &TenantShardId,
2702 0 : ) -> Result<LocationConf, LoadConfigError> {
2703 0 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
2704 0 :
2705 0 : info!("loading tenant configuration from {config_path}");
2706 :
2707 : // load and parse file
2708 0 : let config = fs::read_to_string(&config_path).map_err(|e| {
2709 0 : match e.kind() {
2710 : std::io::ErrorKind::NotFound => {
2711 : // The config should almost always exist for a tenant directory:
2712 : // - When attaching a tenant, the config is the first thing we write
2713 : // - When detaching a tenant, we atomically move the directory to a tmp location
2714 : // before deleting contents.
2715 : //
2716 : // The very rare edge case that can result in a missing config is if we crash during attach
2717 : // between creating directory and writing config. Callers should handle that as if the
2718 : // directory didn't exist.
2719 :
2720 0 : LoadConfigError::NotFound(config_path)
2721 : }
2722 : _ => {
2723 : // No IO errors except NotFound are acceptable here: other kinds of error indicate local storage or permissions issues
2724 : // that we cannot cleanly recover
2725 0 : crate::virtual_file::on_fatal_io_error(&e, "Reading tenant config file")
2726 : }
2727 : }
2728 0 : })?;
2729 :
2730 0 : Ok(toml_edit::de::from_str::<LocationConf>(&config)?)
2731 0 : }
2732 :
2733 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
2734 : pub(super) async fn persist_tenant_config(
2735 : conf: &'static PageServerConf,
2736 : tenant_shard_id: &TenantShardId,
2737 : location_conf: &LocationConf,
2738 : ) -> std::io::Result<()> {
2739 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
2740 :
2741 : Self::persist_tenant_config_at(tenant_shard_id, &config_path, location_conf).await
2742 : }
2743 :
2744 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
2745 : pub(super) async fn persist_tenant_config_at(
2746 : tenant_shard_id: &TenantShardId,
2747 : config_path: &Utf8Path,
2748 : location_conf: &LocationConf,
2749 : ) -> std::io::Result<()> {
2750 : debug!("persisting tenantconf to {config_path}");
2751 :
2752 : let mut conf_content = r#"# This file contains a specific per-tenant's config.
2753 : # It is read in case of pageserver restart.
2754 : "#
2755 : .to_string();
2756 :
2757 0 : fail::fail_point!("tenant-config-before-write", |_| {
2758 0 : Err(std::io::Error::new(
2759 0 : std::io::ErrorKind::Other,
2760 0 : "tenant-config-before-write",
2761 0 : ))
2762 0 : });
2763 :
2764 : // Convert the config to a toml file.
2765 : conf_content +=
2766 : &toml_edit::ser::to_string_pretty(&location_conf).expect("Config serialization failed");
2767 :
2768 : let temp_path = path_with_suffix_extension(config_path, TEMP_FILE_SUFFIX);
2769 :
2770 : let conf_content = conf_content.into_bytes();
2771 : VirtualFile::crashsafe_overwrite(config_path.to_owned(), temp_path, conf_content).await
2772 : }
2773 :
2774 : //
2775 : // How garbage collection works:
2776 : //
2777 : // +--bar------------->
2778 : // /
2779 : // +----+-----foo---------------->
2780 : // /
2781 : // ----main--+-------------------------->
2782 : // \
2783 : // +-----baz-------->
2784 : //
2785 : //
2786 : // 1. Grab 'gc_cs' mutex to prevent new timelines from being created while Timeline's
2787 : // `gc_infos` are being refreshed
2788 : // 2. Scan collected timelines, and on each timeline, make note of the
2789 : // all the points where other timelines have been branched off.
2790 : // We will refrain from removing page versions at those LSNs.
2791 : // 3. For each timeline, scan all layer files on the timeline.
2792 : // Remove all files for which a newer file exists and which
2793 : // don't cover any branch point LSNs.
2794 : //
2795 : // TODO:
2796 : // - if a relation has a non-incremental persistent layer on a child branch, then we
2797 : // don't need to keep that in the parent anymore. But currently
2798 : // we do.
2799 754 : async fn gc_iteration_internal(
2800 754 : &self,
2801 754 : target_timeline_id: Option<TimelineId>,
2802 754 : horizon: u64,
2803 754 : pitr: Duration,
2804 754 : cancel: &CancellationToken,
2805 754 : ctx: &RequestContext,
2806 754 : ) -> Result<GcResult, GcError> {
2807 754 : let mut totals: GcResult = Default::default();
2808 754 : let now = Instant::now();
2809 :
2810 754 : let gc_timelines = self
2811 754 : .refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
2812 751 : .await?;
2813 :
2814 754 : failpoint_support::sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
2815 :
2816 : // If there is nothing to GC, we don't want any messages in the INFO log.
2817 754 : if !gc_timelines.is_empty() {
2818 754 : info!("{} timelines need GC", gc_timelines.len());
2819 : } else {
2820 0 : debug!("{} timelines need GC", gc_timelines.len());
2821 : }
2822 :
2823 : // Perform GC for each timeline.
2824 : //
2825 : // Note that we don't hold the `Tenant::gc_cs` lock here because we don't want to delay the
2826 : // branch creation task, which requires the GC lock. A GC iteration can run concurrently
2827 : // with branch creation.
2828 : //
2829 : // See comments in [`Tenant::branch_timeline`] for more information about why branch
2830 : // creation task can run concurrently with timeline's GC iteration.
2831 1508 : for timeline in gc_timelines {
2832 754 : if cancel.is_cancelled() {
2833 : // We were requested to shut down. Stop and return with the progress we
2834 : // made.
2835 0 : break;
2836 754 : }
2837 754 : let result = match timeline.gc().await {
2838 : Err(GcError::TimelineCancelled) => {
2839 0 : if target_timeline_id.is_some() {
2840 : // If we were targetting this specific timeline, surface cancellation to caller
2841 0 : return Err(GcError::TimelineCancelled);
2842 : } else {
2843 : // A timeline may be shutting down independently of the tenant's lifecycle: we should
2844 : // skip past this and proceed to try GC on other timelines.
2845 0 : continue;
2846 : }
2847 : }
2848 754 : r => r?,
2849 : };
2850 754 : totals += result;
2851 : }
2852 :
2853 754 : totals.elapsed = now.elapsed();
2854 754 : Ok(totals)
2855 754 : }
2856 :
2857 : /// Refreshes the Timeline::gc_info for all timelines, returning the
2858 : /// vector of timelines which have [`Timeline::get_last_record_lsn`] past
2859 : /// [`Tenant::get_gc_horizon`].
2860 : ///
2861 : /// This is usually executed as part of periodic gc, but can now be triggered more often.
2862 0 : pub(crate) async fn refresh_gc_info(
2863 0 : &self,
2864 0 : cancel: &CancellationToken,
2865 0 : ctx: &RequestContext,
2866 0 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
2867 0 : // since this method can now be called at different rates than the configured gc loop, it
2868 0 : // might be that these configuration values get applied faster than what it was previously,
2869 0 : // since these were only read from the gc task.
2870 0 : let horizon = self.get_gc_horizon();
2871 0 : let pitr = self.get_pitr_interval();
2872 0 :
2873 0 : // refresh all timelines
2874 0 : let target_timeline_id = None;
2875 0 :
2876 0 : self.refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
2877 0 : .await
2878 0 : }
2879 :
2880 : /// Populate all Timelines' `GcInfo` with information about their children. We do not set the
2881 : /// PITR cutoffs here, because that requires I/O: this is done later, before GC, by [`Self::refresh_gc_info_internal`]
2882 : ///
2883 : /// Subsequently, parent-child relationships are updated incrementally during timeline creation/deletion.
2884 0 : fn initialize_gc_info(
2885 0 : &self,
2886 0 : timelines: &std::sync::MutexGuard<HashMap<TimelineId, Arc<Timeline>>>,
2887 0 : ) {
2888 0 : // This function must be called before activation: after activation timeline create/delete operations
2889 0 : // might happen, and this function is not safe to run concurrently with those.
2890 0 : assert!(!self.is_active());
2891 :
2892 : // Scan all timelines. For each timeline, remember the timeline ID and
2893 : // the branch point where it was created.
2894 0 : let mut all_branchpoints: BTreeMap<TimelineId, Vec<(Lsn, TimelineId)>> = BTreeMap::new();
2895 0 : timelines.iter().for_each(|(timeline_id, timeline_entry)| {
2896 0 : if let Some(ancestor_timeline_id) = &timeline_entry.get_ancestor_timeline_id() {
2897 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
2898 0 : ancestor_children.push((timeline_entry.get_ancestor_lsn(), *timeline_id));
2899 0 : }
2900 0 : });
2901 0 :
2902 0 : // The number of bytes we always keep, irrespective of PITR: this is a constant across timelines
2903 0 : let horizon = self.get_gc_horizon();
2904 :
2905 : // Populate each timeline's GcInfo with information about its child branches
2906 0 : for timeline in timelines.values() {
2907 0 : let mut branchpoints: Vec<(Lsn, TimelineId)> = all_branchpoints
2908 0 : .remove(&timeline.timeline_id)
2909 0 : .unwrap_or_default();
2910 0 :
2911 0 : branchpoints.sort_by_key(|b| b.0);
2912 0 :
2913 0 : let mut target = timeline.gc_info.write().unwrap();
2914 0 :
2915 0 : target.retain_lsns = branchpoints;
2916 0 :
2917 0 : let space_cutoff = timeline
2918 0 : .get_last_record_lsn()
2919 0 : .checked_sub(horizon)
2920 0 : .unwrap_or(Lsn(0));
2921 0 :
2922 0 : target.cutoffs = GcCutoffs {
2923 0 : space: space_cutoff,
2924 0 : time: Lsn::INVALID,
2925 0 : };
2926 0 : }
2927 0 : }
2928 :
2929 754 : async fn refresh_gc_info_internal(
2930 754 : &self,
2931 754 : target_timeline_id: Option<TimelineId>,
2932 754 : horizon: u64,
2933 754 : pitr: Duration,
2934 754 : cancel: &CancellationToken,
2935 754 : ctx: &RequestContext,
2936 754 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
2937 754 : // before taking the gc_cs lock, do the heavier weight finding of gc_cutoff points for
2938 754 : // currently visible timelines.
2939 754 : let timelines = self
2940 754 : .timelines
2941 754 : .lock()
2942 754 : .unwrap()
2943 754 : .values()
2944 3310 : .filter(|tl| match target_timeline_id.as_ref() {
2945 3310 : Some(target) => &tl.timeline_id == target,
2946 0 : None => true,
2947 3310 : })
2948 754 : .cloned()
2949 754 : .collect::<Vec<_>>();
2950 754 :
2951 754 : if target_timeline_id.is_some() && timelines.is_empty() {
2952 : // We were to act on a particular timeline and it wasn't found
2953 0 : return Err(GcError::TimelineNotFound);
2954 754 : }
2955 754 :
2956 754 : let mut gc_cutoffs: HashMap<TimelineId, GcCutoffs> =
2957 754 : HashMap::with_capacity(timelines.len());
2958 :
2959 754 : for timeline in timelines.iter() {
2960 754 : let cutoff = timeline
2961 754 : .get_last_record_lsn()
2962 754 : .checked_sub(horizon)
2963 754 : .unwrap_or(Lsn(0));
2964 :
2965 754 : let cutoffs = timeline.find_gc_cutoffs(cutoff, pitr, cancel, ctx).await?;
2966 754 : let old = gc_cutoffs.insert(timeline.timeline_id, cutoffs);
2967 754 : assert!(old.is_none());
2968 : }
2969 :
2970 754 : if !self.is_active() || self.cancel.is_cancelled() {
2971 0 : return Err(GcError::TenantCancelled);
2972 754 : }
2973 :
2974 : // grab mutex to prevent new timelines from being created here; avoid doing long operations
2975 : // because that will stall branch creation.
2976 754 : let gc_cs = self.gc_cs.lock().await;
2977 :
2978 : // Paranoia check: it is critical that GcInfo's list of child timelines is correct, to avoid incorrectly GC'ing data they
2979 : // depend on. So although GcInfo is updated continuously by Timeline::new and Timeline::drop, we also calculate it here
2980 : // and fail out if it's inaccurate.
2981 : // (this can be removed later, it's a risk mitigation for https://github.com/neondatabase/neon/pull/8427)
2982 : {
2983 754 : let mut all_branchpoints: BTreeMap<TimelineId, Vec<(Lsn, TimelineId)>> =
2984 754 : BTreeMap::new();
2985 754 : timelines.iter().for_each(|timeline| {
2986 754 : if let Some(ancestor_timeline_id) = &timeline.get_ancestor_timeline_id() {
2987 100 : let ancestor_children =
2988 100 : all_branchpoints.entry(*ancestor_timeline_id).or_default();
2989 100 : ancestor_children.push((timeline.get_ancestor_lsn(), timeline.timeline_id));
2990 654 : }
2991 754 : });
2992 :
2993 1508 : for timeline in &timelines {
2994 754 : let mut branchpoints: Vec<(Lsn, TimelineId)> = all_branchpoints
2995 754 : .remove(&timeline.timeline_id)
2996 754 : .unwrap_or_default();
2997 754 :
2998 754 : branchpoints.sort_by_key(|b| b.0);
2999 754 :
3000 754 : let target = timeline.gc_info.read().unwrap();
3001 754 :
3002 754 : // We require that retain_lsns contains everything in `branchpoints`, but not that
3003 754 : // they are exactly equal: timeline deletions can race with us, so retain_lsns
3004 754 : // may contain some extra stuff. It is safe to have extra timelines in there, because it
3005 754 : // just means that we retain slightly more data than we otherwise might.
3006 754 : let have_branchpoints = target.retain_lsns.iter().copied().collect::<HashSet<_>>();
3007 754 : for b in &branchpoints {
3008 0 : if !have_branchpoints.contains(b) {
3009 0 : tracing::error!(
3010 0 : "Bug: `retain_lsns` is set incorrectly. Expected be {:?}, but found {:?}",
3011 0 : branchpoints,
3012 0 : target.retain_lsns
3013 : );
3014 0 : debug_assert!(false);
3015 : // Do not GC based on bad information!
3016 : // (ab-use an existing GcError type rather than adding a new one, since this is a
3017 : // "should never happen" check that will be removed soon).
3018 0 : return Err(GcError::Remote(anyhow::anyhow!(
3019 0 : "retain_lsns failed validation!"
3020 0 : )));
3021 0 : }
3022 : }
3023 : }
3024 : }
3025 :
3026 : // Ok, we now know all the branch points.
3027 : // Update the GC information for each timeline.
3028 754 : let mut gc_timelines = Vec::with_capacity(timelines.len());
3029 1508 : for timeline in timelines {
3030 : // We filtered the timeline list above
3031 754 : if let Some(target_timeline_id) = target_timeline_id {
3032 754 : assert_eq!(target_timeline_id, timeline.timeline_id);
3033 0 : }
3034 :
3035 : {
3036 754 : let mut target = timeline.gc_info.write().unwrap();
3037 754 :
3038 754 : // Cull any expired leases
3039 754 : let now = SystemTime::now();
3040 754 : target.leases.retain(|_, lease| !lease.is_expired(&now));
3041 754 :
3042 754 : timeline
3043 754 : .metrics
3044 754 : .valid_lsn_lease_count_gauge
3045 754 : .set(target.leases.len() as u64);
3046 :
3047 : // Look up parent's PITR cutoff to update the child's knowledge of whether it is within parent's PITR
3048 754 : if let Some(ancestor_id) = timeline.get_ancestor_timeline_id() {
3049 100 : if let Some(ancestor_gc_cutoffs) = gc_cutoffs.get(&ancestor_id) {
3050 0 : target.within_ancestor_pitr =
3051 0 : timeline.get_ancestor_lsn() >= ancestor_gc_cutoffs.time;
3052 100 : }
3053 654 : }
3054 :
3055 : // Update metrics that depend on GC state
3056 754 : timeline
3057 754 : .metrics
3058 754 : .archival_size
3059 754 : .set(if target.within_ancestor_pitr {
3060 0 : timeline.metrics.current_logical_size_gauge.get()
3061 : } else {
3062 754 : 0
3063 : });
3064 754 : timeline.metrics.pitr_history_size.set(
3065 754 : timeline
3066 754 : .get_last_record_lsn()
3067 754 : .checked_sub(target.cutoffs.time)
3068 754 : .unwrap_or(Lsn(0))
3069 754 : .0,
3070 754 : );
3071 :
3072 : // Apply the cutoffs we found to the Timeline's GcInfo. Why might we _not_ have cutoffs for a timeline?
3073 : // - this timeline was created while we were finding cutoffs
3074 : // - lsn for timestamp search fails for this timeline repeatedly
3075 754 : if let Some(cutoffs) = gc_cutoffs.get(&timeline.timeline_id) {
3076 754 : target.cutoffs = cutoffs.clone();
3077 754 : }
3078 : }
3079 :
3080 754 : gc_timelines.push(timeline);
3081 : }
3082 754 : drop(gc_cs);
3083 754 : Ok(gc_timelines)
3084 754 : }
3085 :
3086 : /// A substitute for `branch_timeline` for use in unit tests.
3087 : /// The returned timeline will have state value `Active` to make various `anyhow::ensure!()`
3088 : /// calls pass, but, we do not actually call `.activate()` under the hood. So, none of the
3089 : /// timeline background tasks are launched, except the flush loop.
3090 : #[cfg(test)]
3091 230 : async fn branch_timeline_test(
3092 230 : &self,
3093 230 : src_timeline: &Arc<Timeline>,
3094 230 : dst_id: TimelineId,
3095 230 : ancestor_lsn: Option<Lsn>,
3096 230 : ctx: &RequestContext,
3097 230 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
3098 230 : let create_guard = self.create_timeline_create_guard(dst_id).unwrap();
3099 230 : let tl = self
3100 230 : .branch_timeline_impl(src_timeline, dst_id, ancestor_lsn, create_guard, ctx)
3101 4 : .await?;
3102 226 : tl.set_state(TimelineState::Active);
3103 226 : Ok(tl)
3104 230 : }
3105 :
3106 : /// Helper for unit tests to branch a timeline with some pre-loaded states.
3107 : #[cfg(test)]
3108 : #[allow(clippy::too_many_arguments)]
3109 6 : pub async fn branch_timeline_test_with_layers(
3110 6 : &self,
3111 6 : src_timeline: &Arc<Timeline>,
3112 6 : dst_id: TimelineId,
3113 6 : ancestor_lsn: Option<Lsn>,
3114 6 : ctx: &RequestContext,
3115 6 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
3116 6 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
3117 6 : end_lsn: Lsn,
3118 6 : ) -> anyhow::Result<Arc<Timeline>> {
3119 6 : let tline = self
3120 6 : .branch_timeline_test(src_timeline, dst_id, ancestor_lsn, ctx)
3121 0 : .await?;
3122 6 : let ancestor_lsn = if let Some(ancestor_lsn) = ancestor_lsn {
3123 6 : ancestor_lsn
3124 : } else {
3125 0 : tline.get_last_record_lsn()
3126 : };
3127 6 : assert!(end_lsn >= ancestor_lsn);
3128 6 : tline.force_advance_lsn(end_lsn);
3129 12 : for deltas in delta_layer_desc {
3130 6 : tline
3131 6 : .force_create_delta_layer(deltas, Some(ancestor_lsn), ctx)
3132 18 : .await?;
3133 : }
3134 10 : for (lsn, images) in image_layer_desc {
3135 4 : tline
3136 4 : .force_create_image_layer(lsn, images, Some(ancestor_lsn), ctx)
3137 14 : .await?;
3138 : }
3139 6 : Ok(tline)
3140 6 : }
3141 :
3142 : /// Branch an existing timeline.
3143 : ///
3144 : /// The caller is responsible for activating the returned timeline.
3145 0 : async fn branch_timeline(
3146 0 : &self,
3147 0 : src_timeline: &Arc<Timeline>,
3148 0 : dst_id: TimelineId,
3149 0 : start_lsn: Option<Lsn>,
3150 0 : timeline_create_guard: TimelineCreateGuard<'_>,
3151 0 : ctx: &RequestContext,
3152 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
3153 0 : self.branch_timeline_impl(src_timeline, dst_id, start_lsn, timeline_create_guard, ctx)
3154 0 : .await
3155 0 : }
3156 :
3157 230 : async fn branch_timeline_impl(
3158 230 : &self,
3159 230 : src_timeline: &Arc<Timeline>,
3160 230 : dst_id: TimelineId,
3161 230 : start_lsn: Option<Lsn>,
3162 230 : timeline_create_guard: TimelineCreateGuard<'_>,
3163 230 : _ctx: &RequestContext,
3164 230 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
3165 230 : let src_id = src_timeline.timeline_id;
3166 :
3167 : // We will validate our ancestor LSN in this function. Acquire the GC lock so that
3168 : // this check cannot race with GC, and the ancestor LSN is guaranteed to remain
3169 : // valid while we are creating the branch.
3170 230 : let _gc_cs = self.gc_cs.lock().await;
3171 :
3172 : // If no start LSN is specified, we branch the new timeline from the source timeline's last record LSN
3173 230 : let start_lsn = start_lsn.unwrap_or_else(|| {
3174 2 : let lsn = src_timeline.get_last_record_lsn();
3175 2 : info!("branching timeline {dst_id} from timeline {src_id} at last record LSN: {lsn}");
3176 2 : lsn
3177 230 : });
3178 230 :
3179 230 : // Ensure that `start_lsn` is valid, i.e. the LSN is within the PITR
3180 230 : // horizon on the source timeline
3181 230 : //
3182 230 : // We check it against both the planned GC cutoff stored in 'gc_info',
3183 230 : // and the 'latest_gc_cutoff' of the last GC that was performed. The
3184 230 : // planned GC cutoff in 'gc_info' is normally larger than
3185 230 : // 'latest_gc_cutoff_lsn', but beware of corner cases like if you just
3186 230 : // changed the GC settings for the tenant to make the PITR window
3187 230 : // larger, but some of the data was already removed by an earlier GC
3188 230 : // iteration.
3189 230 :
3190 230 : // check against last actual 'latest_gc_cutoff' first
3191 230 : let latest_gc_cutoff_lsn = src_timeline.get_latest_gc_cutoff_lsn();
3192 230 : src_timeline
3193 230 : .check_lsn_is_in_scope(start_lsn, &latest_gc_cutoff_lsn)
3194 230 : .context(format!(
3195 230 : "invalid branch start lsn: less than latest GC cutoff {}",
3196 230 : *latest_gc_cutoff_lsn,
3197 230 : ))
3198 230 : .map_err(CreateTimelineError::AncestorLsn)?;
3199 :
3200 : // and then the planned GC cutoff
3201 : {
3202 226 : let gc_info = src_timeline.gc_info.read().unwrap();
3203 226 : let cutoff = gc_info.min_cutoff();
3204 226 : if start_lsn < cutoff {
3205 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
3206 0 : "invalid branch start lsn: less than planned GC cutoff {cutoff}"
3207 0 : )));
3208 226 : }
3209 226 : }
3210 226 :
3211 226 : //
3212 226 : // The branch point is valid, and we are still holding the 'gc_cs' lock
3213 226 : // so that GC cannot advance the GC cutoff until we are finished.
3214 226 : // Proceed with the branch creation.
3215 226 : //
3216 226 :
3217 226 : // Determine prev-LSN for the new timeline. We can only determine it if
3218 226 : // the timeline was branched at the current end of the source timeline.
3219 226 : let RecordLsn {
3220 226 : last: src_last,
3221 226 : prev: src_prev,
3222 226 : } = src_timeline.get_last_record_rlsn();
3223 226 : let dst_prev = if src_last == start_lsn {
3224 216 : Some(src_prev)
3225 : } else {
3226 10 : None
3227 : };
3228 :
3229 : // Create the metadata file, noting the ancestor of the new timeline.
3230 : // There is initially no data in it, but all the read-calls know to look
3231 : // into the ancestor.
3232 226 : let metadata = TimelineMetadata::new(
3233 226 : start_lsn,
3234 226 : dst_prev,
3235 226 : Some(src_id),
3236 226 : start_lsn,
3237 226 : *src_timeline.latest_gc_cutoff_lsn.read(), // FIXME: should we hold onto this guard longer?
3238 226 : src_timeline.initdb_lsn,
3239 226 : src_timeline.pg_version,
3240 226 : );
3241 :
3242 226 : let uninitialized_timeline = self
3243 226 : .prepare_new_timeline(
3244 226 : dst_id,
3245 226 : &metadata,
3246 226 : timeline_create_guard,
3247 226 : start_lsn + 1,
3248 226 : Some(Arc::clone(src_timeline)),
3249 226 : src_timeline.last_aux_file_policy.load(),
3250 226 : )
3251 0 : .await?;
3252 :
3253 226 : let new_timeline = uninitialized_timeline.finish_creation()?;
3254 :
3255 : // Root timeline gets its layers during creation and uploads them along with the metadata.
3256 : // A branch timeline though, when created, can get no writes for some time, hence won't get any layers created.
3257 : // We still need to upload its metadata eagerly: if other nodes `attach` the tenant and miss this timeline, their GC
3258 : // could get incorrect information and remove more layers, than needed.
3259 : // See also https://github.com/neondatabase/neon/issues/3865
3260 226 : new_timeline
3261 226 : .remote_client
3262 226 : .schedule_index_upload_for_full_metadata_update(&metadata)
3263 226 : .context("branch initial metadata upload")?;
3264 :
3265 226 : Ok(new_timeline)
3266 230 : }
3267 :
3268 : /// For unit tests, make this visible so that other modules can directly create timelines
3269 : #[cfg(test)]
3270 4 : #[tracing::instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
3271 : pub(crate) async fn bootstrap_timeline_test(
3272 : &self,
3273 : timeline_id: TimelineId,
3274 : pg_version: u32,
3275 : load_existing_initdb: Option<TimelineId>,
3276 : ctx: &RequestContext,
3277 : ) -> anyhow::Result<Arc<Timeline>> {
3278 : let create_guard = self.create_timeline_create_guard(timeline_id).unwrap();
3279 : self.bootstrap_timeline(
3280 : timeline_id,
3281 : pg_version,
3282 : load_existing_initdb,
3283 : create_guard,
3284 : ctx,
3285 : )
3286 : .await
3287 : }
3288 :
3289 0 : async fn upload_initdb(
3290 0 : &self,
3291 0 : timelines_path: &Utf8PathBuf,
3292 0 : pgdata_path: &Utf8PathBuf,
3293 0 : timeline_id: &TimelineId,
3294 0 : ) -> anyhow::Result<()> {
3295 0 : let temp_path = timelines_path.join(format!(
3296 0 : "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
3297 0 : ));
3298 :
3299 : scopeguard::defer! {
3300 : if let Err(e) = fs::remove_file(&temp_path) {
3301 : error!("Failed to remove temporary initdb archive '{temp_path}': {e}");
3302 : }
3303 : }
3304 :
3305 0 : let (pgdata_zstd, tar_zst_size) = create_zst_tarball(pgdata_path, &temp_path).await?;
3306 : const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
3307 0 : if tar_zst_size > INITDB_TAR_ZST_WARN_LIMIT {
3308 0 : warn!(
3309 0 : "compressed {temp_path} size of {tar_zst_size} is above limit {INITDB_TAR_ZST_WARN_LIMIT}."
3310 : );
3311 0 : }
3312 :
3313 : pausable_failpoint!("before-initdb-upload");
3314 :
3315 0 : backoff::retry(
3316 0 : || async {
3317 0 : self::remote_timeline_client::upload_initdb_dir(
3318 0 : &self.remote_storage,
3319 0 : &self.tenant_shard_id.tenant_id,
3320 0 : timeline_id,
3321 0 : pgdata_zstd.try_clone().await?,
3322 0 : tar_zst_size,
3323 0 : &self.cancel,
3324 0 : )
3325 0 : .await
3326 0 : },
3327 0 : |_| false,
3328 0 : 3,
3329 0 : u32::MAX,
3330 0 : "persist_initdb_tar_zst",
3331 0 : &self.cancel,
3332 0 : )
3333 0 : .await
3334 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
3335 0 : .and_then(|x| x)
3336 0 : }
3337 :
3338 : /// - run initdb to init temporary instance and get bootstrap data
3339 : /// - after initialization completes, tar up the temp dir and upload it to S3.
3340 : ///
3341 : /// The caller is responsible for activating the returned timeline.
3342 2 : async fn bootstrap_timeline(
3343 2 : &self,
3344 2 : timeline_id: TimelineId,
3345 2 : pg_version: u32,
3346 2 : load_existing_initdb: Option<TimelineId>,
3347 2 : timeline_create_guard: TimelineCreateGuard<'_>,
3348 2 : ctx: &RequestContext,
3349 2 : ) -> anyhow::Result<Arc<Timeline>> {
3350 2 : // create a `tenant/{tenant_id}/timelines/basebackup-{timeline_id}.{TEMP_FILE_SUFFIX}/`
3351 2 : // temporary directory for basebackup files for the given timeline.
3352 2 :
3353 2 : let timelines_path = self.conf.timelines_path(&self.tenant_shard_id);
3354 2 : let pgdata_path = path_with_suffix_extension(
3355 2 : timelines_path.join(format!("basebackup-{timeline_id}")),
3356 2 : TEMP_FILE_SUFFIX,
3357 2 : );
3358 2 :
3359 2 : // Remove whatever was left from the previous runs: safe because TimelineCreateGuard guarantees
3360 2 : // we won't race with other creations or existent timelines with the same path.
3361 2 : if pgdata_path.exists() {
3362 0 : fs::remove_dir_all(&pgdata_path).with_context(|| {
3363 0 : format!("Failed to remove already existing initdb directory: {pgdata_path}")
3364 0 : })?;
3365 2 : }
3366 :
3367 : // this new directory is very temporary, set to remove it immediately after bootstrap, we don't need it
3368 : scopeguard::defer! {
3369 : if let Err(e) = fs::remove_dir_all(&pgdata_path) {
3370 : // this is unlikely, but we will remove the directory on pageserver restart or another bootstrap call
3371 : error!("Failed to remove temporary initdb directory '{pgdata_path}': {e}");
3372 : }
3373 : }
3374 2 : if let Some(existing_initdb_timeline_id) = load_existing_initdb {
3375 2 : if existing_initdb_timeline_id != timeline_id {
3376 0 : let source_path = &remote_initdb_archive_path(
3377 0 : &self.tenant_shard_id.tenant_id,
3378 0 : &existing_initdb_timeline_id,
3379 0 : );
3380 0 : let dest_path =
3381 0 : &remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &timeline_id);
3382 0 :
3383 0 : // if this fails, it will get retried by retried control plane requests
3384 0 : self.remote_storage
3385 0 : .copy_object(source_path, dest_path, &self.cancel)
3386 0 : .await
3387 0 : .context("copy initdb tar")?;
3388 2 : }
3389 2 : let (initdb_tar_zst_path, initdb_tar_zst) =
3390 2 : self::remote_timeline_client::download_initdb_tar_zst(
3391 2 : self.conf,
3392 2 : &self.remote_storage,
3393 2 : &self.tenant_shard_id,
3394 2 : &existing_initdb_timeline_id,
3395 2 : &self.cancel,
3396 2 : )
3397 738 : .await
3398 2 : .context("download initdb tar")?;
3399 :
3400 : scopeguard::defer! {
3401 : if let Err(e) = fs::remove_file(&initdb_tar_zst_path) {
3402 : error!("Failed to remove temporary initdb archive '{initdb_tar_zst_path}': {e}");
3403 : }
3404 : }
3405 :
3406 2 : let buf_read =
3407 2 : BufReader::with_capacity(remote_timeline_client::BUFFER_SIZE, initdb_tar_zst);
3408 2 : extract_zst_tarball(&pgdata_path, buf_read)
3409 11105 : .await
3410 2 : .context("extract initdb tar")?;
3411 : } else {
3412 : // Init temporarily repo to get bootstrap data, this creates a directory in the `pgdata_path` path
3413 0 : run_initdb(self.conf, &pgdata_path, pg_version, &self.cancel).await?;
3414 :
3415 : // Upload the created data dir to S3
3416 0 : if self.tenant_shard_id().is_shard_zero() {
3417 0 : self.upload_initdb(&timelines_path, &pgdata_path, &timeline_id)
3418 0 : .await?;
3419 0 : }
3420 : }
3421 2 : let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
3422 2 :
3423 2 : // Import the contents of the data directory at the initial checkpoint
3424 2 : // LSN, and any WAL after that.
3425 2 : // Initdb lsn will be equal to last_record_lsn which will be set after import.
3426 2 : // Because we know it upfront avoid having an option or dummy zero value by passing it to the metadata.
3427 2 : let new_metadata = TimelineMetadata::new(
3428 2 : Lsn(0),
3429 2 : None,
3430 2 : None,
3431 2 : Lsn(0),
3432 2 : pgdata_lsn,
3433 2 : pgdata_lsn,
3434 2 : pg_version,
3435 2 : );
3436 2 : let raw_timeline = self
3437 2 : .prepare_new_timeline(
3438 2 : timeline_id,
3439 2 : &new_metadata,
3440 2 : timeline_create_guard,
3441 2 : pgdata_lsn,
3442 2 : None,
3443 2 : None,
3444 2 : )
3445 0 : .await?;
3446 :
3447 2 : let tenant_shard_id = raw_timeline.owning_tenant.tenant_shard_id;
3448 2 : let unfinished_timeline = raw_timeline.raw_timeline()?;
3449 :
3450 : // Flush the new layer files to disk, before we make the timeline as available to
3451 : // the outside world.
3452 : //
3453 : // Flush loop needs to be spawned in order to be able to flush.
3454 2 : unfinished_timeline.maybe_spawn_flush_loop();
3455 2 :
3456 2 : import_datadir::import_timeline_from_postgres_datadir(
3457 2 : unfinished_timeline,
3458 2 : &pgdata_path,
3459 2 : pgdata_lsn,
3460 2 : ctx,
3461 2 : )
3462 9194 : .await
3463 2 : .with_context(|| {
3464 0 : format!("Failed to import pgdatadir for timeline {tenant_shard_id}/{timeline_id}")
3465 2 : })?;
3466 :
3467 2 : fail::fail_point!("before-checkpoint-new-timeline", |_| {
3468 0 : anyhow::bail!("failpoint before-checkpoint-new-timeline");
3469 2 : });
3470 :
3471 2 : unfinished_timeline
3472 2 : .freeze_and_flush()
3473 2 : .await
3474 2 : .with_context(|| {
3475 0 : format!(
3476 0 : "Failed to flush after pgdatadir import for timeline {tenant_shard_id}/{timeline_id}"
3477 0 : )
3478 2 : })?;
3479 :
3480 : // All done!
3481 2 : let timeline = raw_timeline.finish_creation()?;
3482 :
3483 2 : Ok(timeline)
3484 2 : }
3485 :
3486 : /// Call this before constructing a timeline, to build its required structures
3487 392 : fn build_timeline_resources(&self, timeline_id: TimelineId) -> TimelineResources {
3488 392 : let remote_client = RemoteTimelineClient::new(
3489 392 : self.remote_storage.clone(),
3490 392 : self.deletion_queue_client.clone(),
3491 392 : self.conf,
3492 392 : self.tenant_shard_id,
3493 392 : timeline_id,
3494 392 : self.generation,
3495 392 : );
3496 392 : TimelineResources {
3497 392 : remote_client,
3498 392 : timeline_get_throttle: self.timeline_get_throttle.clone(),
3499 392 : l0_flush_global_state: self.l0_flush_global_state.clone(),
3500 392 : }
3501 392 : }
3502 :
3503 : /// Creates intermediate timeline structure and its files.
3504 : ///
3505 : /// An empty layer map is initialized, and new data and WAL can be imported starting
3506 : /// at 'disk_consistent_lsn'. After any initial data has been imported, call
3507 : /// `finish_creation` to insert the Timeline into the timelines map.
3508 392 : async fn prepare_new_timeline<'a>(
3509 392 : &'a self,
3510 392 : new_timeline_id: TimelineId,
3511 392 : new_metadata: &TimelineMetadata,
3512 392 : create_guard: TimelineCreateGuard<'a>,
3513 392 : start_lsn: Lsn,
3514 392 : ancestor: Option<Arc<Timeline>>,
3515 392 : last_aux_file_policy: Option<AuxFilePolicy>,
3516 392 : ) -> anyhow::Result<UninitializedTimeline> {
3517 392 : let tenant_shard_id = self.tenant_shard_id;
3518 392 :
3519 392 : let resources = self.build_timeline_resources(new_timeline_id);
3520 392 : resources
3521 392 : .remote_client
3522 392 : .init_upload_queue_for_empty_remote(new_metadata)?;
3523 :
3524 392 : let timeline_struct = self
3525 392 : .create_timeline_struct(
3526 392 : new_timeline_id,
3527 392 : new_metadata,
3528 392 : ancestor,
3529 392 : resources,
3530 392 : CreateTimelineCause::Load,
3531 392 : last_aux_file_policy,
3532 392 : )
3533 392 : .context("Failed to create timeline data structure")?;
3534 :
3535 392 : timeline_struct.init_empty_layer_map(start_lsn);
3536 :
3537 392 : if let Err(e) = self
3538 392 : .create_timeline_files(&create_guard.timeline_path)
3539 0 : .await
3540 : {
3541 0 : error!("Failed to create initial files for timeline {tenant_shard_id}/{new_timeline_id}, cleaning up: {e:?}");
3542 0 : cleanup_timeline_directory(create_guard);
3543 0 : return Err(e);
3544 392 : }
3545 392 :
3546 392 : debug!(
3547 0 : "Successfully created initial files for timeline {tenant_shard_id}/{new_timeline_id}"
3548 : );
3549 :
3550 392 : Ok(UninitializedTimeline::new(
3551 392 : self,
3552 392 : new_timeline_id,
3553 392 : Some((timeline_struct, create_guard)),
3554 392 : ))
3555 392 : }
3556 :
3557 392 : async fn create_timeline_files(&self, timeline_path: &Utf8Path) -> anyhow::Result<()> {
3558 392 : crashsafe::create_dir(timeline_path).context("Failed to create timeline directory")?;
3559 :
3560 392 : fail::fail_point!("after-timeline-dir-creation", |_| {
3561 0 : anyhow::bail!("failpoint after-timeline-dir-creation");
3562 392 : });
3563 :
3564 392 : Ok(())
3565 392 : }
3566 :
3567 : /// Get a guard that provides exclusive access to the timeline directory, preventing
3568 : /// concurrent attempts to create the same timeline.
3569 398 : fn create_timeline_create_guard(
3570 398 : &self,
3571 398 : timeline_id: TimelineId,
3572 398 : ) -> Result<TimelineCreateGuard, TimelineExclusionError> {
3573 398 : let tenant_shard_id = self.tenant_shard_id;
3574 398 :
3575 398 : let timeline_path = self.conf.timeline_path(&tenant_shard_id, &timeline_id);
3576 :
3577 398 : let create_guard = TimelineCreateGuard::new(self, timeline_id, timeline_path.clone())?;
3578 :
3579 : // At this stage, we have got exclusive access to in-memory state for this timeline ID
3580 : // for creation.
3581 : // A timeline directory should never exist on disk already:
3582 : // - a previous failed creation would have cleaned up after itself
3583 : // - a pageserver restart would clean up timeline directories that don't have valid remote state
3584 : //
3585 : // Therefore it is an unexpected internal error to encounter a timeline directory already existing here,
3586 : // this error may indicate a bug in cleanup on failed creations.
3587 396 : if timeline_path.exists() {
3588 0 : return Err(TimelineExclusionError::Other(anyhow::anyhow!(
3589 0 : "Timeline directory already exists! This is a bug."
3590 0 : )));
3591 396 : }
3592 396 :
3593 396 : Ok(create_guard)
3594 398 : }
3595 :
3596 : /// Gathers inputs from all of the timelines to produce a sizing model input.
3597 : ///
3598 : /// Future is cancellation safe. Only one calculation can be running at once per tenant.
3599 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
3600 : pub async fn gather_size_inputs(
3601 : &self,
3602 : // `max_retention_period` overrides the cutoff that is used to calculate the size
3603 : // (only if it is shorter than the real cutoff).
3604 : max_retention_period: Option<u64>,
3605 : cause: LogicalSizeCalculationCause,
3606 : cancel: &CancellationToken,
3607 : ctx: &RequestContext,
3608 : ) -> Result<size::ModelInputs, size::CalculateSyntheticSizeError> {
3609 : let logical_sizes_at_once = self
3610 : .conf
3611 : .concurrent_tenant_size_logical_size_queries
3612 : .inner();
3613 :
3614 : // TODO: Having a single mutex block concurrent reads is not great for performance.
3615 : //
3616 : // But the only case where we need to run multiple of these at once is when we
3617 : // request a size for a tenant manually via API, while another background calculation
3618 : // is in progress (which is not a common case).
3619 : //
3620 : // See more for on the issue #2748 condenced out of the initial PR review.
3621 : let mut shared_cache = tokio::select! {
3622 : locked = self.cached_logical_sizes.lock() => locked,
3623 : _ = cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
3624 : _ = self.cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
3625 : };
3626 :
3627 : size::gather_inputs(
3628 : self,
3629 : logical_sizes_at_once,
3630 : max_retention_period,
3631 : &mut shared_cache,
3632 : cause,
3633 : cancel,
3634 : ctx,
3635 : )
3636 : .await
3637 : }
3638 :
3639 : /// Calculate synthetic tenant size and cache the result.
3640 : /// This is periodically called by background worker.
3641 : /// result is cached in tenant struct
3642 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
3643 : pub async fn calculate_synthetic_size(
3644 : &self,
3645 : cause: LogicalSizeCalculationCause,
3646 : cancel: &CancellationToken,
3647 : ctx: &RequestContext,
3648 : ) -> Result<u64, size::CalculateSyntheticSizeError> {
3649 : let inputs = self.gather_size_inputs(None, cause, cancel, ctx).await?;
3650 :
3651 : let size = inputs.calculate();
3652 :
3653 : self.set_cached_synthetic_size(size);
3654 :
3655 : Ok(size)
3656 : }
3657 :
3658 : /// Cache given synthetic size and update the metric value
3659 0 : pub fn set_cached_synthetic_size(&self, size: u64) {
3660 0 : self.cached_synthetic_tenant_size
3661 0 : .store(size, Ordering::Relaxed);
3662 0 :
3663 0 : // Only shard zero should be calculating synthetic sizes
3664 0 : debug_assert!(self.shard_identity.is_shard_zero());
3665 :
3666 0 : TENANT_SYNTHETIC_SIZE_METRIC
3667 0 : .get_metric_with_label_values(&[&self.tenant_shard_id.tenant_id.to_string()])
3668 0 : .unwrap()
3669 0 : .set(size);
3670 0 : }
3671 :
3672 0 : pub fn cached_synthetic_size(&self) -> u64 {
3673 0 : self.cached_synthetic_tenant_size.load(Ordering::Relaxed)
3674 0 : }
3675 :
3676 : /// Flush any in-progress layers, schedule uploads, and wait for uploads to complete.
3677 : ///
3678 : /// This function can take a long time: callers should wrap it in a timeout if calling
3679 : /// from an external API handler.
3680 : ///
3681 : /// Cancel-safety: cancelling this function may leave I/O running, but such I/O is
3682 : /// still bounded by tenant/timeline shutdown.
3683 0 : #[tracing::instrument(skip_all)]
3684 : pub(crate) async fn flush_remote(&self) -> anyhow::Result<()> {
3685 : let timelines = self.timelines.lock().unwrap().clone();
3686 :
3687 0 : async fn flush_timeline(_gate: GateGuard, timeline: Arc<Timeline>) -> anyhow::Result<()> {
3688 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Flushing...");
3689 0 : timeline.freeze_and_flush().await?;
3690 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Waiting for uploads...");
3691 0 : timeline.remote_client.wait_completion().await?;
3692 :
3693 0 : Ok(())
3694 0 : }
3695 :
3696 : // We do not use a JoinSet for these tasks, because we don't want them to be
3697 : // aborted when this function's future is cancelled: they should stay alive
3698 : // holding their GateGuard until they complete, to ensure their I/Os complete
3699 : // before Timeline shutdown completes.
3700 : let mut results = FuturesUnordered::new();
3701 :
3702 : for (_timeline_id, timeline) in timelines {
3703 : // Run each timeline's flush in a task holding the timeline's gate: this
3704 : // means that if this function's future is cancelled, the Timeline shutdown
3705 : // will still wait for any I/O in here to complete.
3706 : let Ok(gate) = timeline.gate.enter() else {
3707 : continue;
3708 : };
3709 0 : let jh = tokio::task::spawn(async move { flush_timeline(gate, timeline).await });
3710 : results.push(jh);
3711 : }
3712 :
3713 : while let Some(r) = results.next().await {
3714 : if let Err(e) = r {
3715 : if !e.is_cancelled() && !e.is_panic() {
3716 : tracing::error!("unexpected join error: {e:?}");
3717 : }
3718 : }
3719 : }
3720 :
3721 : // The flushes we did above were just writes, but the Tenant might have had
3722 : // pending deletions as well from recent compaction/gc: we want to flush those
3723 : // as well. This requires flushing the global delete queue. This is cheap
3724 : // because it's typically a no-op.
3725 : match self.deletion_queue_client.flush_execute().await {
3726 : Ok(_) => {}
3727 : Err(DeletionQueueError::ShuttingDown) => {}
3728 : }
3729 :
3730 : Ok(())
3731 : }
3732 :
3733 0 : pub(crate) fn get_tenant_conf(&self) -> TenantConfOpt {
3734 0 : self.tenant_conf.load().tenant_conf.clone()
3735 0 : }
3736 : }
3737 :
3738 : /// Create the cluster temporarily in 'initdbpath' directory inside the repository
3739 : /// to get bootstrap data for timeline initialization.
3740 0 : async fn run_initdb(
3741 0 : conf: &'static PageServerConf,
3742 0 : initdb_target_dir: &Utf8Path,
3743 0 : pg_version: u32,
3744 0 : cancel: &CancellationToken,
3745 0 : ) -> Result<(), InitdbError> {
3746 0 : let initdb_bin_path = conf
3747 0 : .pg_bin_dir(pg_version)
3748 0 : .map_err(InitdbError::Other)?
3749 0 : .join("initdb");
3750 0 : let initdb_lib_dir = conf.pg_lib_dir(pg_version).map_err(InitdbError::Other)?;
3751 0 : info!(
3752 0 : "running {} in {}, libdir: {}",
3753 : initdb_bin_path, initdb_target_dir, initdb_lib_dir,
3754 : );
3755 :
3756 0 : let _permit = INIT_DB_SEMAPHORE.acquire().await;
3757 :
3758 0 : let initdb_command = tokio::process::Command::new(&initdb_bin_path)
3759 0 : .args(["-D", initdb_target_dir.as_ref()])
3760 0 : .args(["-U", &conf.superuser])
3761 0 : .args(["-E", "utf8"])
3762 0 : .arg("--no-instructions")
3763 0 : .arg("--no-sync")
3764 0 : .env_clear()
3765 0 : .env("LD_LIBRARY_PATH", &initdb_lib_dir)
3766 0 : .env("DYLD_LIBRARY_PATH", &initdb_lib_dir)
3767 0 : .stdin(std::process::Stdio::null())
3768 0 : // stdout invocation produces the same output every time, we don't need it
3769 0 : .stdout(std::process::Stdio::null())
3770 0 : // we would be interested in the stderr output, if there was any
3771 0 : .stderr(std::process::Stdio::piped())
3772 0 : .spawn()?;
3773 :
3774 : // Ideally we'd select here with the cancellation token, but the problem is that
3775 : // we can't safely terminate initdb: it launches processes of its own, and killing
3776 : // initdb doesn't kill them. After we return from this function, we want the target
3777 : // directory to be able to be cleaned up.
3778 : // See https://github.com/neondatabase/neon/issues/6385
3779 0 : let initdb_output = initdb_command.wait_with_output().await?;
3780 0 : if !initdb_output.status.success() {
3781 0 : return Err(InitdbError::Failed(
3782 0 : initdb_output.status,
3783 0 : initdb_output.stderr,
3784 0 : ));
3785 0 : }
3786 0 :
3787 0 : // This isn't true cancellation support, see above. Still return an error to
3788 0 : // excercise the cancellation code path.
3789 0 : if cancel.is_cancelled() {
3790 0 : return Err(InitdbError::Cancelled);
3791 0 : }
3792 0 :
3793 0 : Ok(())
3794 0 : }
3795 :
3796 : /// Dump contents of a layer file to stdout.
3797 0 : pub async fn dump_layerfile_from_path(
3798 0 : path: &Utf8Path,
3799 0 : verbose: bool,
3800 0 : ctx: &RequestContext,
3801 0 : ) -> anyhow::Result<()> {
3802 : use std::os::unix::fs::FileExt;
3803 :
3804 : // All layer files start with a two-byte "magic" value, to identify the kind of
3805 : // file.
3806 0 : let file = File::open(path)?;
3807 0 : let mut header_buf = [0u8; 2];
3808 0 : file.read_exact_at(&mut header_buf, 0)?;
3809 :
3810 0 : match u16::from_be_bytes(header_buf) {
3811 : crate::IMAGE_FILE_MAGIC => {
3812 0 : ImageLayer::new_for_path(path, file)?
3813 0 : .dump(verbose, ctx)
3814 0 : .await?
3815 : }
3816 : crate::DELTA_FILE_MAGIC => {
3817 0 : DeltaLayer::new_for_path(path, file)?
3818 0 : .dump(verbose, ctx)
3819 0 : .await?
3820 : }
3821 0 : magic => bail!("unrecognized magic identifier: {:?}", magic),
3822 : }
3823 :
3824 0 : Ok(())
3825 0 : }
3826 :
3827 : #[cfg(test)]
3828 : pub(crate) mod harness {
3829 : use bytes::{Bytes, BytesMut};
3830 : use once_cell::sync::OnceCell;
3831 : use pageserver_api::models::ShardParameters;
3832 : use pageserver_api::shard::ShardIndex;
3833 : use utils::logging;
3834 :
3835 : use crate::deletion_queue::mock::MockDeletionQueue;
3836 : use crate::l0_flush::L0FlushConfig;
3837 : use crate::walredo::apply_neon;
3838 : use crate::{repository::Key, walrecord::NeonWalRecord};
3839 :
3840 : use super::*;
3841 : use hex_literal::hex;
3842 : use utils::id::TenantId;
3843 :
3844 : pub const TIMELINE_ID: TimelineId =
3845 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
3846 : pub const NEW_TIMELINE_ID: TimelineId =
3847 : TimelineId::from_array(hex!("AA223344556677881122334455667788"));
3848 :
3849 : /// Convenience function to create a page image with given string as the only content
3850 5028396 : pub fn test_img(s: &str) -> Bytes {
3851 5028396 : let mut buf = BytesMut::new();
3852 5028396 : buf.extend_from_slice(s.as_bytes());
3853 5028396 : buf.resize(64, 0);
3854 5028396 :
3855 5028396 : buf.freeze()
3856 5028396 : }
3857 :
3858 : impl From<TenantConf> for TenantConfOpt {
3859 174 : fn from(tenant_conf: TenantConf) -> Self {
3860 174 : Self {
3861 174 : checkpoint_distance: Some(tenant_conf.checkpoint_distance),
3862 174 : checkpoint_timeout: Some(tenant_conf.checkpoint_timeout),
3863 174 : compaction_target_size: Some(tenant_conf.compaction_target_size),
3864 174 : compaction_period: Some(tenant_conf.compaction_period),
3865 174 : compaction_threshold: Some(tenant_conf.compaction_threshold),
3866 174 : compaction_algorithm: Some(tenant_conf.compaction_algorithm),
3867 174 : gc_horizon: Some(tenant_conf.gc_horizon),
3868 174 : gc_period: Some(tenant_conf.gc_period),
3869 174 : image_creation_threshold: Some(tenant_conf.image_creation_threshold),
3870 174 : pitr_interval: Some(tenant_conf.pitr_interval),
3871 174 : walreceiver_connect_timeout: Some(tenant_conf.walreceiver_connect_timeout),
3872 174 : lagging_wal_timeout: Some(tenant_conf.lagging_wal_timeout),
3873 174 : max_lsn_wal_lag: Some(tenant_conf.max_lsn_wal_lag),
3874 174 : eviction_policy: Some(tenant_conf.eviction_policy),
3875 174 : min_resident_size_override: tenant_conf.min_resident_size_override,
3876 174 : evictions_low_residence_duration_metric_threshold: Some(
3877 174 : tenant_conf.evictions_low_residence_duration_metric_threshold,
3878 174 : ),
3879 174 : heatmap_period: Some(tenant_conf.heatmap_period),
3880 174 : lazy_slru_download: Some(tenant_conf.lazy_slru_download),
3881 174 : timeline_get_throttle: Some(tenant_conf.timeline_get_throttle),
3882 174 : image_layer_creation_check_threshold: Some(
3883 174 : tenant_conf.image_layer_creation_check_threshold,
3884 174 : ),
3885 174 : switch_aux_file_policy: Some(tenant_conf.switch_aux_file_policy),
3886 174 : lsn_lease_length: Some(tenant_conf.lsn_lease_length),
3887 174 : lsn_lease_length_for_ts: Some(tenant_conf.lsn_lease_length_for_ts),
3888 174 : }
3889 174 : }
3890 : }
3891 :
3892 : pub struct TenantHarness {
3893 : pub conf: &'static PageServerConf,
3894 : pub tenant_conf: TenantConf,
3895 : pub tenant_shard_id: TenantShardId,
3896 : pub generation: Generation,
3897 : pub shard: ShardIndex,
3898 : pub remote_storage: GenericRemoteStorage,
3899 : pub remote_fs_dir: Utf8PathBuf,
3900 : pub deletion_queue: MockDeletionQueue,
3901 : }
3902 :
3903 : static LOG_HANDLE: OnceCell<()> = OnceCell::new();
3904 :
3905 188 : pub(crate) fn setup_logging() {
3906 188 : LOG_HANDLE.get_or_init(|| {
3907 176 : logging::init(
3908 176 : logging::LogFormat::Test,
3909 176 : // enable it in case the tests exercise code paths that use
3910 176 : // debug_assert_current_span_has_tenant_and_timeline_id
3911 176 : logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
3912 176 : logging::Output::Stdout,
3913 176 : )
3914 176 : .expect("Failed to init test logging")
3915 188 : });
3916 188 : }
3917 :
3918 : impl TenantHarness {
3919 174 : pub async fn create_custom(
3920 174 : test_name: &'static str,
3921 174 : tenant_conf: TenantConf,
3922 174 : tenant_id: TenantId,
3923 174 : shard_identity: ShardIdentity,
3924 174 : generation: Generation,
3925 174 : ) -> anyhow::Result<Self> {
3926 174 : setup_logging();
3927 174 :
3928 174 : let repo_dir = PageServerConf::test_repo_dir(test_name);
3929 174 : let _ = fs::remove_dir_all(&repo_dir);
3930 174 : fs::create_dir_all(&repo_dir)?;
3931 :
3932 174 : let conf = PageServerConf::dummy_conf(repo_dir);
3933 174 : // Make a static copy of the config. This can never be free'd, but that's
3934 174 : // OK in a test.
3935 174 : let conf: &'static PageServerConf = Box::leak(Box::new(conf));
3936 174 :
3937 174 : let shard = shard_identity.shard_index();
3938 174 : let tenant_shard_id = TenantShardId {
3939 174 : tenant_id,
3940 174 : shard_number: shard.shard_number,
3941 174 : shard_count: shard.shard_count,
3942 174 : };
3943 174 : fs::create_dir_all(conf.tenant_path(&tenant_shard_id))?;
3944 174 : fs::create_dir_all(conf.timelines_path(&tenant_shard_id))?;
3945 :
3946 : use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
3947 174 : let remote_fs_dir = conf.workdir.join("localfs");
3948 174 : std::fs::create_dir_all(&remote_fs_dir).unwrap();
3949 174 : let config = RemoteStorageConfig {
3950 174 : storage: RemoteStorageKind::LocalFs {
3951 174 : local_path: remote_fs_dir.clone(),
3952 174 : },
3953 174 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
3954 174 : };
3955 174 : let remote_storage = GenericRemoteStorage::from_config(&config).await.unwrap();
3956 174 : let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()));
3957 174 :
3958 174 : Ok(Self {
3959 174 : conf,
3960 174 : tenant_conf,
3961 174 : tenant_shard_id,
3962 174 : generation,
3963 174 : shard,
3964 174 : remote_storage,
3965 174 : remote_fs_dir,
3966 174 : deletion_queue,
3967 174 : })
3968 174 : }
3969 :
3970 162 : pub async fn create(test_name: &'static str) -> anyhow::Result<Self> {
3971 162 : // Disable automatic GC and compaction to make the unit tests more deterministic.
3972 162 : // The tests perform them manually if needed.
3973 162 : let tenant_conf = TenantConf {
3974 162 : gc_period: Duration::ZERO,
3975 162 : compaction_period: Duration::ZERO,
3976 162 : ..TenantConf::default()
3977 162 : };
3978 162 : let tenant_id = TenantId::generate();
3979 162 : let shard = ShardIdentity::unsharded();
3980 162 : Self::create_custom(
3981 162 : test_name,
3982 162 : tenant_conf,
3983 162 : tenant_id,
3984 162 : shard,
3985 162 : Generation::new(0xdeadbeef),
3986 162 : )
3987 0 : .await
3988 162 : }
3989 :
3990 18 : pub fn span(&self) -> tracing::Span {
3991 18 : info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
3992 18 : }
3993 :
3994 174 : pub(crate) async fn load(&self) -> (Arc<Tenant>, RequestContext) {
3995 174 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
3996 174 : (
3997 174 : self.do_try_load(&ctx)
3998 699 : .await
3999 174 : .expect("failed to load test tenant"),
4000 174 : ctx,
4001 174 : )
4002 174 : }
4003 :
4004 348 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
4005 : pub(crate) async fn do_try_load(
4006 : &self,
4007 : ctx: &RequestContext,
4008 : ) -> anyhow::Result<Arc<Tenant>> {
4009 : let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
4010 :
4011 : let tenant = Arc::new(Tenant::new(
4012 : TenantState::Loading,
4013 : self.conf,
4014 : AttachedTenantConf::try_from(LocationConf::attached_single(
4015 : TenantConfOpt::from(self.tenant_conf.clone()),
4016 : self.generation,
4017 : &ShardParameters::default(),
4018 : ))
4019 : .unwrap(),
4020 : // This is a legacy/test code path: sharding isn't supported here.
4021 : ShardIdentity::unsharded(),
4022 : Some(walredo_mgr),
4023 : self.tenant_shard_id,
4024 : self.remote_storage.clone(),
4025 : self.deletion_queue.new_client(),
4026 : // TODO: ideally we should run all unit tests with both configs
4027 : L0FlushGlobalState::new(L0FlushConfig::default()),
4028 : ));
4029 :
4030 : let preload = tenant
4031 : .preload(&self.remote_storage, CancellationToken::new())
4032 : .await?;
4033 : tenant.attach(Some(preload), ctx).await?;
4034 :
4035 : tenant.state.send_replace(TenantState::Active);
4036 : for timeline in tenant.timelines.lock().unwrap().values() {
4037 : timeline.set_state(TimelineState::Active);
4038 : }
4039 : Ok(tenant)
4040 : }
4041 :
4042 2 : pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
4043 2 : self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
4044 2 : }
4045 : }
4046 :
4047 : // Mock WAL redo manager that doesn't do much
4048 : pub(crate) struct TestRedoManager;
4049 :
4050 : impl TestRedoManager {
4051 : /// # Cancel-Safety
4052 : ///
4053 : /// This method is cancellation-safe.
4054 266 : pub async fn request_redo(
4055 266 : &self,
4056 266 : key: Key,
4057 266 : lsn: Lsn,
4058 266 : base_img: Option<(Lsn, Bytes)>,
4059 266 : records: Vec<(Lsn, NeonWalRecord)>,
4060 266 : _pg_version: u32,
4061 266 : ) -> Result<Bytes, walredo::Error> {
4062 342 : let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
4063 266 : if records_neon {
4064 : // For Neon wal records, we can decode without spawning postgres, so do so.
4065 266 : let base_img = base_img.expect("Neon WAL redo requires base image").1;
4066 266 : let mut page = BytesMut::new();
4067 266 : page.extend_from_slice(&base_img);
4068 608 : for (record_lsn, record) in records {
4069 342 : apply_neon::apply_in_neon(&record, record_lsn, key, &mut page)?;
4070 : }
4071 266 : Ok(page.freeze())
4072 : } else {
4073 : // We never spawn a postgres walredo process in unit tests: just log what we might have done.
4074 0 : let s = format!(
4075 0 : "redo for {} to get to {}, with {} and {} records",
4076 0 : key,
4077 0 : lsn,
4078 0 : if base_img.is_some() {
4079 0 : "base image"
4080 : } else {
4081 0 : "no base image"
4082 : },
4083 0 : records.len()
4084 0 : );
4085 0 : println!("{s}");
4086 0 :
4087 0 : Ok(test_img(&s))
4088 : }
4089 266 : }
4090 : }
4091 : }
4092 :
4093 : #[cfg(test)]
4094 : mod tests {
4095 : use std::collections::BTreeMap;
4096 :
4097 : use super::*;
4098 : use crate::keyspace::KeySpaceAccum;
4099 : use crate::pgdatadir_mapping::AuxFilesDirectory;
4100 : use crate::repository::{Key, Value};
4101 : use crate::tenant::harness::*;
4102 : use crate::tenant::timeline::CompactFlags;
4103 : use crate::walrecord::NeonWalRecord;
4104 : use crate::DEFAULT_PG_VERSION;
4105 : use bytes::{Bytes, BytesMut};
4106 : use hex_literal::hex;
4107 : use itertools::Itertools;
4108 : use pageserver_api::key::{AUX_FILES_KEY, AUX_KEY_PREFIX, NON_INHERITED_RANGE};
4109 : use pageserver_api::keyspace::KeySpace;
4110 : use pageserver_api::models::{CompactionAlgorithm, CompactionAlgorithmSettings};
4111 : use rand::{thread_rng, Rng};
4112 : use storage_layer::PersistentLayerKey;
4113 : use tests::storage_layer::ValuesReconstructState;
4114 : use tests::timeline::{GetVectoredError, ShutdownMode};
4115 : use timeline::compaction::{KeyHistoryRetention, KeyLogAtLsn};
4116 : use timeline::{DeltaLayerTestDesc, GcInfo};
4117 : use utils::bin_ser::BeSer;
4118 : use utils::id::TenantId;
4119 :
4120 : static TEST_KEY: Lazy<Key> =
4121 18 : Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));
4122 :
4123 : #[tokio::test]
4124 2 : async fn test_basic() -> anyhow::Result<()> {
4125 8 : let (tenant, ctx) = TenantHarness::create("test_basic").await?.load().await;
4126 2 : let tline = tenant
4127 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
4128 4 : .await?;
4129 2 :
4130 2 : let mut writer = tline.writer().await;
4131 2 : writer
4132 2 : .put(
4133 2 : *TEST_KEY,
4134 2 : Lsn(0x10),
4135 2 : &Value::Image(test_img("foo at 0x10")),
4136 2 : &ctx,
4137 2 : )
4138 2 : .await?;
4139 2 : writer.finish_write(Lsn(0x10));
4140 2 : drop(writer);
4141 2 :
4142 2 : let mut writer = tline.writer().await;
4143 2 : writer
4144 2 : .put(
4145 2 : *TEST_KEY,
4146 2 : Lsn(0x20),
4147 2 : &Value::Image(test_img("foo at 0x20")),
4148 2 : &ctx,
4149 2 : )
4150 2 : .await?;
4151 2 : writer.finish_write(Lsn(0x20));
4152 2 : drop(writer);
4153 2 :
4154 2 : assert_eq!(
4155 2 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
4156 2 : test_img("foo at 0x10")
4157 2 : );
4158 2 : assert_eq!(
4159 2 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
4160 2 : test_img("foo at 0x10")
4161 2 : );
4162 2 : assert_eq!(
4163 2 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
4164 2 : test_img("foo at 0x20")
4165 2 : );
4166 2 :
4167 2 : Ok(())
4168 2 : }
4169 :
4170 : #[tokio::test]
4171 2 : async fn no_duplicate_timelines() -> anyhow::Result<()> {
4172 2 : let (tenant, ctx) = TenantHarness::create("no_duplicate_timelines")
4173 2 : .await?
4174 2 : .load()
4175 8 : .await;
4176 2 : let _ = tenant
4177 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4178 4 : .await?;
4179 2 :
4180 2 : match tenant
4181 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4182 2 : .await
4183 2 : {
4184 2 : Ok(_) => panic!("duplicate timeline creation should fail"),
4185 2 : Err(e) => assert_eq!(e.to_string(), "Already exists".to_string()),
4186 2 : }
4187 2 :
4188 2 : Ok(())
4189 2 : }
4190 :
4191 : /// Convenience function to create a page image with given string as the only content
4192 10 : pub fn test_value(s: &str) -> Value {
4193 10 : let mut buf = BytesMut::new();
4194 10 : buf.extend_from_slice(s.as_bytes());
4195 10 : Value::Image(buf.freeze())
4196 10 : }
4197 :
4198 : ///
4199 : /// Test branch creation
4200 : ///
4201 : #[tokio::test]
4202 2 : async fn test_branch() -> anyhow::Result<()> {
4203 2 : use std::str::from_utf8;
4204 2 :
4205 8 : let (tenant, ctx) = TenantHarness::create("test_branch").await?.load().await;
4206 2 : let tline = tenant
4207 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4208 4 : .await?;
4209 2 : let mut writer = tline.writer().await;
4210 2 :
4211 2 : #[allow(non_snake_case)]
4212 2 : let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
4213 2 : #[allow(non_snake_case)]
4214 2 : let TEST_KEY_B: Key = Key::from_hex("110000000033333333444444445500000002").unwrap();
4215 2 :
4216 2 : // Insert a value on the timeline
4217 2 : writer
4218 2 : .put(TEST_KEY_A, Lsn(0x20), &test_value("foo at 0x20"), &ctx)
4219 2 : .await?;
4220 2 : writer
4221 2 : .put(TEST_KEY_B, Lsn(0x20), &test_value("foobar at 0x20"), &ctx)
4222 2 : .await?;
4223 2 : writer.finish_write(Lsn(0x20));
4224 2 :
4225 2 : writer
4226 2 : .put(TEST_KEY_A, Lsn(0x30), &test_value("foo at 0x30"), &ctx)
4227 2 : .await?;
4228 2 : writer.finish_write(Lsn(0x30));
4229 2 : writer
4230 2 : .put(TEST_KEY_A, Lsn(0x40), &test_value("foo at 0x40"), &ctx)
4231 2 : .await?;
4232 2 : writer.finish_write(Lsn(0x40));
4233 2 :
4234 2 : //assert_current_logical_size(&tline, Lsn(0x40));
4235 2 :
4236 2 : // Branch the history, modify relation differently on the new timeline
4237 2 : tenant
4238 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x30)), &ctx)
4239 2 : .await?;
4240 2 : let newtline = tenant
4241 2 : .get_timeline(NEW_TIMELINE_ID, true)
4242 2 : .expect("Should have a local timeline");
4243 2 : let mut new_writer = newtline.writer().await;
4244 2 : new_writer
4245 2 : .put(TEST_KEY_A, Lsn(0x40), &test_value("bar at 0x40"), &ctx)
4246 2 : .await?;
4247 2 : new_writer.finish_write(Lsn(0x40));
4248 2 :
4249 2 : // Check page contents on both branches
4250 2 : assert_eq!(
4251 2 : from_utf8(&tline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
4252 2 : "foo at 0x40"
4253 2 : );
4254 2 : assert_eq!(
4255 2 : from_utf8(&newtline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
4256 2 : "bar at 0x40"
4257 2 : );
4258 2 : assert_eq!(
4259 2 : from_utf8(&newtline.get(TEST_KEY_B, Lsn(0x40), &ctx).await?)?,
4260 2 : "foobar at 0x20"
4261 2 : );
4262 2 :
4263 2 : //assert_current_logical_size(&tline, Lsn(0x40));
4264 2 :
4265 2 : Ok(())
4266 2 : }
4267 :
4268 20 : async fn make_some_layers(
4269 20 : tline: &Timeline,
4270 20 : start_lsn: Lsn,
4271 20 : ctx: &RequestContext,
4272 20 : ) -> anyhow::Result<()> {
4273 20 : let mut lsn = start_lsn;
4274 : {
4275 20 : let mut writer = tline.writer().await;
4276 : // Create a relation on the timeline
4277 20 : writer
4278 20 : .put(
4279 20 : *TEST_KEY,
4280 20 : lsn,
4281 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4282 20 : ctx,
4283 20 : )
4284 10 : .await?;
4285 20 : writer.finish_write(lsn);
4286 20 : lsn += 0x10;
4287 20 : writer
4288 20 : .put(
4289 20 : *TEST_KEY,
4290 20 : lsn,
4291 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4292 20 : ctx,
4293 20 : )
4294 0 : .await?;
4295 20 : writer.finish_write(lsn);
4296 20 : lsn += 0x10;
4297 20 : }
4298 21 : tline.freeze_and_flush().await?;
4299 : {
4300 20 : let mut writer = tline.writer().await;
4301 20 : writer
4302 20 : .put(
4303 20 : *TEST_KEY,
4304 20 : lsn,
4305 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4306 20 : ctx,
4307 20 : )
4308 10 : .await?;
4309 20 : writer.finish_write(lsn);
4310 20 : lsn += 0x10;
4311 20 : writer
4312 20 : .put(
4313 20 : *TEST_KEY,
4314 20 : lsn,
4315 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4316 20 : ctx,
4317 20 : )
4318 0 : .await?;
4319 20 : writer.finish_write(lsn);
4320 20 : }
4321 20 : tline.freeze_and_flush().await.map_err(|e| e.into())
4322 20 : }
4323 :
4324 : #[tokio::test]
4325 2 : async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
4326 2 : let (tenant, ctx) =
4327 2 : TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
4328 2 : .await?
4329 2 : .load()
4330 8 : .await;
4331 2 : let tline = tenant
4332 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4333 4 : .await?;
4334 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4335 2 :
4336 2 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
4337 2 : // FIXME: this doesn't actually remove any layer currently, given how the flushing
4338 2 : // and compaction works. But it does set the 'cutoff' point so that the cross check
4339 2 : // below should fail.
4340 2 : tenant
4341 2 : .gc_iteration(
4342 2 : Some(TIMELINE_ID),
4343 2 : 0x10,
4344 2 : Duration::ZERO,
4345 2 : &CancellationToken::new(),
4346 2 : &ctx,
4347 2 : )
4348 2 : .await?;
4349 2 :
4350 2 : // try to branch at lsn 25, should fail because we already garbage collected the data
4351 2 : match tenant
4352 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
4353 2 : .await
4354 2 : {
4355 2 : Ok(_) => panic!("branching should have failed"),
4356 2 : Err(err) => {
4357 2 : let CreateTimelineError::AncestorLsn(err) = err else {
4358 2 : panic!("wrong error type")
4359 2 : };
4360 2 : assert!(err.to_string().contains("invalid branch start lsn"));
4361 2 : assert!(err
4362 2 : .source()
4363 2 : .unwrap()
4364 2 : .to_string()
4365 2 : .contains("we might've already garbage collected needed data"))
4366 2 : }
4367 2 : }
4368 2 :
4369 2 : Ok(())
4370 2 : }
4371 :
4372 : #[tokio::test]
4373 2 : async fn test_prohibit_branch_creation_on_pre_initdb_lsn() -> anyhow::Result<()> {
4374 2 : let (tenant, ctx) =
4375 2 : TenantHarness::create("test_prohibit_branch_creation_on_pre_initdb_lsn")
4376 2 : .await?
4377 2 : .load()
4378 8 : .await;
4379 2 :
4380 2 : let tline = tenant
4381 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x50), DEFAULT_PG_VERSION, &ctx)
4382 4 : .await?;
4383 2 : // try to branch at lsn 0x25, should fail because initdb lsn is 0x50
4384 2 : match tenant
4385 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
4386 2 : .await
4387 2 : {
4388 2 : Ok(_) => panic!("branching should have failed"),
4389 2 : Err(err) => {
4390 2 : let CreateTimelineError::AncestorLsn(err) = err else {
4391 2 : panic!("wrong error type");
4392 2 : };
4393 2 : assert!(&err.to_string().contains("invalid branch start lsn"));
4394 2 : assert!(&err
4395 2 : .source()
4396 2 : .unwrap()
4397 2 : .to_string()
4398 2 : .contains("is earlier than latest GC cutoff"));
4399 2 : }
4400 2 : }
4401 2 :
4402 2 : Ok(())
4403 2 : }
4404 :
4405 : /*
4406 : // FIXME: This currently fails to error out. Calling GC doesn't currently
4407 : // remove the old value, we'd need to work a little harder
4408 : #[tokio::test]
4409 : async fn test_prohibit_get_for_garbage_collected_data() -> anyhow::Result<()> {
4410 : let repo =
4411 : RepoHarness::create("test_prohibit_get_for_garbage_collected_data")?
4412 : .load();
4413 :
4414 : let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION)?;
4415 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4416 :
4417 : repo.gc_iteration(Some(TIMELINE_ID), 0x10, Duration::ZERO)?;
4418 : let latest_gc_cutoff_lsn = tline.get_latest_gc_cutoff_lsn();
4419 : assert!(*latest_gc_cutoff_lsn > Lsn(0x25));
4420 : match tline.get(*TEST_KEY, Lsn(0x25)) {
4421 : Ok(_) => panic!("request for page should have failed"),
4422 : Err(err) => assert!(err.to_string().contains("not found at")),
4423 : }
4424 : Ok(())
4425 : }
4426 : */
4427 :
4428 : #[tokio::test]
4429 2 : async fn test_get_branchpoints_from_an_inactive_timeline() -> anyhow::Result<()> {
4430 2 : let (tenant, ctx) =
4431 2 : TenantHarness::create("test_get_branchpoints_from_an_inactive_timeline")
4432 2 : .await?
4433 2 : .load()
4434 8 : .await;
4435 2 : let tline = tenant
4436 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4437 4 : .await?;
4438 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4439 2 :
4440 2 : tenant
4441 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
4442 2 : .await?;
4443 2 : let newtline = tenant
4444 2 : .get_timeline(NEW_TIMELINE_ID, true)
4445 2 : .expect("Should have a local timeline");
4446 2 :
4447 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
4448 2 :
4449 2 : tline.set_broken("test".to_owned());
4450 2 :
4451 2 : tenant
4452 2 : .gc_iteration(
4453 2 : Some(TIMELINE_ID),
4454 2 : 0x10,
4455 2 : Duration::ZERO,
4456 2 : &CancellationToken::new(),
4457 2 : &ctx,
4458 2 : )
4459 2 : .await?;
4460 2 :
4461 2 : // The branchpoints should contain all timelines, even ones marked
4462 2 : // as Broken.
4463 2 : {
4464 2 : let branchpoints = &tline.gc_info.read().unwrap().retain_lsns;
4465 2 : assert_eq!(branchpoints.len(), 1);
4466 2 : assert_eq!(branchpoints[0], (Lsn(0x40), NEW_TIMELINE_ID));
4467 2 : }
4468 2 :
4469 2 : // You can read the key from the child branch even though the parent is
4470 2 : // Broken, as long as you don't need to access data from the parent.
4471 2 : assert_eq!(
4472 4 : newtline.get(*TEST_KEY, Lsn(0x70), &ctx).await?,
4473 2 : test_img(&format!("foo at {}", Lsn(0x70)))
4474 2 : );
4475 2 :
4476 2 : // This needs to traverse to the parent, and fails.
4477 2 : let err = newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await.unwrap_err();
4478 2 : assert!(err.to_string().starts_with(&format!(
4479 2 : "Bad state on timeline {}: Broken",
4480 2 : tline.timeline_id
4481 2 : )));
4482 2 :
4483 2 : Ok(())
4484 2 : }
4485 :
4486 : #[tokio::test]
4487 2 : async fn test_retain_data_in_parent_which_is_needed_for_child() -> anyhow::Result<()> {
4488 2 : let (tenant, ctx) =
4489 2 : TenantHarness::create("test_retain_data_in_parent_which_is_needed_for_child")
4490 2 : .await?
4491 2 : .load()
4492 8 : .await;
4493 2 : let tline = tenant
4494 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4495 4 : .await?;
4496 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4497 2 :
4498 2 : tenant
4499 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
4500 2 : .await?;
4501 2 : let newtline = tenant
4502 2 : .get_timeline(NEW_TIMELINE_ID, true)
4503 2 : .expect("Should have a local timeline");
4504 2 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
4505 2 : tenant
4506 2 : .gc_iteration(
4507 2 : Some(TIMELINE_ID),
4508 2 : 0x10,
4509 2 : Duration::ZERO,
4510 2 : &CancellationToken::new(),
4511 2 : &ctx,
4512 2 : )
4513 2 : .await?;
4514 4 : assert!(newtline.get(*TEST_KEY, Lsn(0x25), &ctx).await.is_ok());
4515 2 :
4516 2 : Ok(())
4517 2 : }
4518 : #[tokio::test]
4519 2 : async fn test_parent_keeps_data_forever_after_branching() -> anyhow::Result<()> {
4520 2 : let (tenant, ctx) = TenantHarness::create("test_parent_keeps_data_forever_after_branching")
4521 2 : .await?
4522 2 : .load()
4523 8 : .await;
4524 2 : let tline = tenant
4525 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4526 4 : .await?;
4527 7 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4528 2 :
4529 2 : tenant
4530 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
4531 2 : .await?;
4532 2 : let newtline = tenant
4533 2 : .get_timeline(NEW_TIMELINE_ID, true)
4534 2 : .expect("Should have a local timeline");
4535 2 :
4536 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
4537 2 :
4538 2 : // run gc on parent
4539 2 : tenant
4540 2 : .gc_iteration(
4541 2 : Some(TIMELINE_ID),
4542 2 : 0x10,
4543 2 : Duration::ZERO,
4544 2 : &CancellationToken::new(),
4545 2 : &ctx,
4546 2 : )
4547 2 : .await?;
4548 2 :
4549 2 : // Check that the data is still accessible on the branch.
4550 2 : assert_eq!(
4551 7 : newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
4552 2 : test_img(&format!("foo at {}", Lsn(0x40)))
4553 2 : );
4554 2 :
4555 2 : Ok(())
4556 2 : }
4557 :
4558 : #[tokio::test]
4559 2 : async fn timeline_load() -> anyhow::Result<()> {
4560 2 : const TEST_NAME: &str = "timeline_load";
4561 2 : let harness = TenantHarness::create(TEST_NAME).await?;
4562 2 : {
4563 8 : let (tenant, ctx) = harness.load().await;
4564 2 : let tline = tenant
4565 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x7000), DEFAULT_PG_VERSION, &ctx)
4566 4 : .await?;
4567 6 : make_some_layers(tline.as_ref(), Lsn(0x8000), &ctx).await?;
4568 2 : // so that all uploads finish & we can call harness.load() below again
4569 2 : tenant
4570 2 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
4571 2 : .instrument(harness.span())
4572 2 : .await
4573 2 : .ok()
4574 2 : .unwrap();
4575 2 : }
4576 2 :
4577 8 : let (tenant, _ctx) = harness.load().await;
4578 2 : tenant
4579 2 : .get_timeline(TIMELINE_ID, true)
4580 2 : .expect("cannot load timeline");
4581 2 :
4582 2 : Ok(())
4583 2 : }
4584 :
4585 : #[tokio::test]
4586 2 : async fn timeline_load_with_ancestor() -> anyhow::Result<()> {
4587 2 : const TEST_NAME: &str = "timeline_load_with_ancestor";
4588 2 : let harness = TenantHarness::create(TEST_NAME).await?;
4589 2 : // create two timelines
4590 2 : {
4591 8 : let (tenant, ctx) = harness.load().await;
4592 2 : let tline = tenant
4593 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4594 4 : .await?;
4595 2 :
4596 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4597 2 :
4598 2 : let child_tline = tenant
4599 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
4600 2 : .await?;
4601 2 : child_tline.set_state(TimelineState::Active);
4602 2 :
4603 2 : let newtline = tenant
4604 2 : .get_timeline(NEW_TIMELINE_ID, true)
4605 2 : .expect("Should have a local timeline");
4606 2 :
4607 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
4608 2 :
4609 2 : // so that all uploads finish & we can call harness.load() below again
4610 2 : tenant
4611 2 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
4612 2 : .instrument(harness.span())
4613 2 : .await
4614 2 : .ok()
4615 2 : .unwrap();
4616 2 : }
4617 2 :
4618 2 : // check that both of them are initially unloaded
4619 12 : let (tenant, _ctx) = harness.load().await;
4620 2 :
4621 2 : // check that both, child and ancestor are loaded
4622 2 : let _child_tline = tenant
4623 2 : .get_timeline(NEW_TIMELINE_ID, true)
4624 2 : .expect("cannot get child timeline loaded");
4625 2 :
4626 2 : let _ancestor_tline = tenant
4627 2 : .get_timeline(TIMELINE_ID, true)
4628 2 : .expect("cannot get ancestor timeline loaded");
4629 2 :
4630 2 : Ok(())
4631 2 : }
4632 :
4633 : #[tokio::test]
4634 2 : async fn delta_layer_dumping() -> anyhow::Result<()> {
4635 2 : use storage_layer::AsLayerDesc;
4636 2 : let (tenant, ctx) = TenantHarness::create("test_layer_dumping")
4637 2 : .await?
4638 2 : .load()
4639 8 : .await;
4640 2 : let tline = tenant
4641 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4642 4 : .await?;
4643 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4644 2 :
4645 2 : let layer_map = tline.layers.read().await;
4646 2 : let level0_deltas = layer_map
4647 2 : .layer_map()
4648 2 : .get_level0_deltas()
4649 2 : .into_iter()
4650 4 : .map(|desc| layer_map.get_from_desc(&desc))
4651 2 : .collect::<Vec<_>>();
4652 2 :
4653 2 : assert!(!level0_deltas.is_empty());
4654 2 :
4655 6 : for delta in level0_deltas {
4656 2 : // Ensure we are dumping a delta layer here
4657 4 : assert!(delta.layer_desc().is_delta);
4658 8 : delta.dump(true, &ctx).await.unwrap();
4659 2 : }
4660 2 :
4661 2 : Ok(())
4662 2 : }
4663 :
4664 : #[tokio::test]
4665 2 : async fn test_images() -> anyhow::Result<()> {
4666 8 : let (tenant, ctx) = TenantHarness::create("test_images").await?.load().await;
4667 2 : let tline = tenant
4668 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
4669 4 : .await?;
4670 2 :
4671 2 : let mut writer = tline.writer().await;
4672 2 : writer
4673 2 : .put(
4674 2 : *TEST_KEY,
4675 2 : Lsn(0x10),
4676 2 : &Value::Image(test_img("foo at 0x10")),
4677 2 : &ctx,
4678 2 : )
4679 2 : .await?;
4680 2 : writer.finish_write(Lsn(0x10));
4681 2 : drop(writer);
4682 2 :
4683 2 : tline.freeze_and_flush().await?;
4684 2 : tline
4685 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
4686 2 : .await?;
4687 2 :
4688 2 : let mut writer = tline.writer().await;
4689 2 : writer
4690 2 : .put(
4691 2 : *TEST_KEY,
4692 2 : Lsn(0x20),
4693 2 : &Value::Image(test_img("foo at 0x20")),
4694 2 : &ctx,
4695 2 : )
4696 2 : .await?;
4697 2 : writer.finish_write(Lsn(0x20));
4698 2 : drop(writer);
4699 2 :
4700 2 : tline.freeze_and_flush().await?;
4701 2 : tline
4702 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
4703 2 : .await?;
4704 2 :
4705 2 : let mut writer = tline.writer().await;
4706 2 : writer
4707 2 : .put(
4708 2 : *TEST_KEY,
4709 2 : Lsn(0x30),
4710 2 : &Value::Image(test_img("foo at 0x30")),
4711 2 : &ctx,
4712 2 : )
4713 2 : .await?;
4714 2 : writer.finish_write(Lsn(0x30));
4715 2 : drop(writer);
4716 2 :
4717 4 : tline.freeze_and_flush().await?;
4718 2 : tline
4719 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
4720 2 : .await?;
4721 2 :
4722 2 : let mut writer = tline.writer().await;
4723 2 : writer
4724 2 : .put(
4725 2 : *TEST_KEY,
4726 2 : Lsn(0x40),
4727 2 : &Value::Image(test_img("foo at 0x40")),
4728 2 : &ctx,
4729 2 : )
4730 2 : .await?;
4731 2 : writer.finish_write(Lsn(0x40));
4732 2 : drop(writer);
4733 2 :
4734 2 : tline.freeze_and_flush().await?;
4735 2 : tline
4736 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
4737 2 : .await?;
4738 2 :
4739 2 : assert_eq!(
4740 4 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
4741 2 : test_img("foo at 0x10")
4742 2 : );
4743 2 : assert_eq!(
4744 4 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
4745 2 : test_img("foo at 0x10")
4746 2 : );
4747 2 : assert_eq!(
4748 2 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
4749 2 : test_img("foo at 0x20")
4750 2 : );
4751 2 : assert_eq!(
4752 4 : tline.get(*TEST_KEY, Lsn(0x30), &ctx).await?,
4753 2 : test_img("foo at 0x30")
4754 2 : );
4755 2 : assert_eq!(
4756 4 : tline.get(*TEST_KEY, Lsn(0x40), &ctx).await?,
4757 2 : test_img("foo at 0x40")
4758 2 : );
4759 2 :
4760 2 : Ok(())
4761 2 : }
4762 :
4763 4 : async fn bulk_insert_compact_gc(
4764 4 : tenant: &Tenant,
4765 4 : timeline: &Arc<Timeline>,
4766 4 : ctx: &RequestContext,
4767 4 : lsn: Lsn,
4768 4 : repeat: usize,
4769 4 : key_count: usize,
4770 4 : ) -> anyhow::Result<()> {
4771 4 : let compact = true;
4772 75580 : bulk_insert_maybe_compact_gc(tenant, timeline, ctx, lsn, repeat, key_count, compact).await
4773 4 : }
4774 :
4775 8 : async fn bulk_insert_maybe_compact_gc(
4776 8 : tenant: &Tenant,
4777 8 : timeline: &Arc<Timeline>,
4778 8 : ctx: &RequestContext,
4779 8 : mut lsn: Lsn,
4780 8 : repeat: usize,
4781 8 : key_count: usize,
4782 8 : compact: bool,
4783 8 : ) -> anyhow::Result<()> {
4784 8 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
4785 8 : let mut blknum = 0;
4786 8 :
4787 8 : // Enforce that key range is monotonously increasing
4788 8 : let mut keyspace = KeySpaceAccum::new();
4789 8 :
4790 8 : let cancel = CancellationToken::new();
4791 8 :
4792 8 : for _ in 0..repeat {
4793 400 : for _ in 0..key_count {
4794 4000000 : test_key.field6 = blknum;
4795 4000000 : let mut writer = timeline.writer().await;
4796 4000000 : writer
4797 4000000 : .put(
4798 4000000 : test_key,
4799 4000000 : lsn,
4800 4000000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
4801 4000000 : ctx,
4802 4000000 : )
4803 2400 : .await?;
4804 4000000 : writer.finish_write(lsn);
4805 4000000 : drop(writer);
4806 4000000 :
4807 4000000 : keyspace.add_key(test_key);
4808 4000000 :
4809 4000000 : lsn = Lsn(lsn.0 + 0x10);
4810 4000000 : blknum += 1;
4811 : }
4812 :
4813 400 : timeline.freeze_and_flush().await?;
4814 400 : if compact {
4815 : // this requires timeline to be &Arc<Timeline>
4816 43180 : timeline.compact(&cancel, EnumSet::empty(), ctx).await?;
4817 200 : }
4818 :
4819 : // this doesn't really need to use the timeline_id target, but it is closer to what it
4820 : // originally was.
4821 400 : let res = tenant
4822 400 : .gc_iteration(Some(timeline.timeline_id), 0, Duration::ZERO, &cancel, ctx)
4823 400 : .await?;
4824 :
4825 400 : assert_eq!(res.layers_removed, 0, "this never removes anything");
4826 : }
4827 :
4828 8 : Ok(())
4829 8 : }
4830 :
4831 : //
4832 : // Insert 1000 key-value pairs with increasing keys, flush, compact, GC.
4833 : // Repeat 50 times.
4834 : //
4835 : #[tokio::test]
4836 2 : async fn test_bulk_insert() -> anyhow::Result<()> {
4837 2 : let harness = TenantHarness::create("test_bulk_insert").await?;
4838 8 : let (tenant, ctx) = harness.load().await;
4839 2 : let tline = tenant
4840 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
4841 4 : .await?;
4842 2 :
4843 2 : let lsn = Lsn(0x10);
4844 37790 : bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
4845 2 :
4846 2 : Ok(())
4847 2 : }
4848 :
4849 : // Test the vectored get real implementation against a simple sequential implementation.
4850 : //
4851 : // The test generates a keyspace by repeatedly flushing the in-memory layer and compacting.
4852 : // Projected to 2D the key space looks like below. Lsn grows upwards on the Y axis and keys
4853 : // grow to the right on the X axis.
4854 : // [Delta]
4855 : // [Delta]
4856 : // [Delta]
4857 : // [Delta]
4858 : // ------------ Image ---------------
4859 : //
4860 : // After layer generation we pick the ranges to query as follows:
4861 : // 1. The beginning of each delta layer
4862 : // 2. At the seam between two adjacent delta layers
4863 : //
4864 : // There's one major downside to this test: delta layers only contains images,
4865 : // so the search can stop at the first delta layer and doesn't traverse any deeper.
4866 : #[tokio::test]
4867 2 : async fn test_get_vectored() -> anyhow::Result<()> {
4868 2 : let harness = TenantHarness::create("test_get_vectored").await?;
4869 8 : let (tenant, ctx) = harness.load().await;
4870 2 : let tline = tenant
4871 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
4872 4 : .await?;
4873 2 :
4874 2 : let lsn = Lsn(0x10);
4875 37790 : bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
4876 2 :
4877 2 : let guard = tline.layers.read().await;
4878 2 : guard.layer_map().dump(true, &ctx).await?;
4879 2 :
4880 2 : let mut reads = Vec::new();
4881 2 : let mut prev = None;
4882 12 : guard.layer_map().iter_historic_layers().for_each(|desc| {
4883 12 : if !desc.is_delta() {
4884 2 : prev = Some(desc.clone());
4885 2 : return;
4886 10 : }
4887 10 :
4888 10 : let start = desc.key_range.start;
4889 10 : let end = desc
4890 10 : .key_range
4891 10 : .start
4892 10 : .add(Timeline::MAX_GET_VECTORED_KEYS.try_into().unwrap());
4893 10 : reads.push(KeySpace {
4894 10 : ranges: vec![start..end],
4895 10 : });
4896 2 :
4897 10 : if let Some(prev) = &prev {
4898 10 : if !prev.is_delta() {
4899 10 : return;
4900 2 : }
4901 0 :
4902 0 : let first_range = Key {
4903 0 : field6: prev.key_range.end.field6 - 4,
4904 0 : ..prev.key_range.end
4905 0 : }..prev.key_range.end;
4906 0 :
4907 0 : let second_range = desc.key_range.start..Key {
4908 0 : field6: desc.key_range.start.field6 + 4,
4909 0 : ..desc.key_range.start
4910 0 : };
4911 0 :
4912 0 : reads.push(KeySpace {
4913 0 : ranges: vec![first_range, second_range],
4914 0 : });
4915 2 : };
4916 2 :
4917 2 : prev = Some(desc.clone());
4918 12 : });
4919 2 :
4920 2 : drop(guard);
4921 2 :
4922 2 : // Pick a big LSN such that we query over all the changes.
4923 2 : let reads_lsn = Lsn(u64::MAX - 1);
4924 2 :
4925 12 : for read in reads {
4926 10 : info!("Doing vectored read on {:?}", read);
4927 2 :
4928 10 : let vectored_res = tline
4929 10 : .get_vectored_impl(
4930 10 : read.clone(),
4931 10 : reads_lsn,
4932 10 : &mut ValuesReconstructState::new(),
4933 10 : &ctx,
4934 10 : )
4935 25 : .await;
4936 10 : tline
4937 10 : .validate_get_vectored_impl(&vectored_res, read, reads_lsn, &ctx)
4938 19 : .await;
4939 2 : }
4940 2 :
4941 2 : Ok(())
4942 2 : }
4943 :
4944 : #[tokio::test]
4945 2 : async fn test_get_vectored_aux_files() -> anyhow::Result<()> {
4946 2 : let harness = TenantHarness::create("test_get_vectored_aux_files").await?;
4947 2 :
4948 8 : let (tenant, ctx) = harness.load().await;
4949 2 : let tline = tenant
4950 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
4951 2 : .await?;
4952 2 : let tline = tline.raw_timeline().unwrap();
4953 2 :
4954 2 : let mut modification = tline.begin_modification(Lsn(0x1000));
4955 2 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
4956 2 : modification.set_lsn(Lsn(0x1008))?;
4957 2 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
4958 2 : modification.commit(&ctx).await?;
4959 2 :
4960 2 : let child_timeline_id = TimelineId::generate();
4961 2 : tenant
4962 2 : .branch_timeline_test(
4963 2 : tline,
4964 2 : child_timeline_id,
4965 2 : Some(tline.get_last_record_lsn()),
4966 2 : &ctx,
4967 2 : )
4968 2 : .await?;
4969 2 :
4970 2 : let child_timeline = tenant
4971 2 : .get_timeline(child_timeline_id, true)
4972 2 : .expect("Should have the branched timeline");
4973 2 :
4974 2 : let aux_keyspace = KeySpace {
4975 2 : ranges: vec![NON_INHERITED_RANGE],
4976 2 : };
4977 2 : let read_lsn = child_timeline.get_last_record_lsn();
4978 2 :
4979 2 : let vectored_res = child_timeline
4980 2 : .get_vectored_impl(
4981 2 : aux_keyspace.clone(),
4982 2 : read_lsn,
4983 2 : &mut ValuesReconstructState::new(),
4984 2 : &ctx,
4985 2 : )
4986 2 : .await;
4987 2 :
4988 2 : child_timeline
4989 2 : .validate_get_vectored_impl(&vectored_res, aux_keyspace, read_lsn, &ctx)
4990 2 : .await;
4991 2 :
4992 2 : let images = vectored_res?;
4993 2 : assert!(images.is_empty());
4994 2 : Ok(())
4995 2 : }
4996 :
4997 : // Test that vectored get handles layer gaps correctly
4998 : // by advancing into the next ancestor timeline if required.
4999 : //
5000 : // The test generates timelines that look like the diagram below.
5001 : // We leave a gap in one of the L1 layers at `gap_at_key` (`/` in the diagram).
5002 : // The reconstruct data for that key lies in the ancestor timeline (`X` in the diagram).
5003 : //
5004 : // ```
5005 : //-------------------------------+
5006 : // ... |
5007 : // [ L1 ] |
5008 : // [ / L1 ] | Child Timeline
5009 : // ... |
5010 : // ------------------------------+
5011 : // [ X L1 ] | Parent Timeline
5012 : // ------------------------------+
5013 : // ```
5014 : #[tokio::test]
5015 2 : async fn test_get_vectored_key_gap() -> anyhow::Result<()> {
5016 2 : let tenant_conf = TenantConf {
5017 2 : // Make compaction deterministic
5018 2 : gc_period: Duration::ZERO,
5019 2 : compaction_period: Duration::ZERO,
5020 2 : // Encourage creation of L1 layers
5021 2 : checkpoint_distance: 16 * 1024,
5022 2 : compaction_target_size: 8 * 1024,
5023 2 : ..TenantConf::default()
5024 2 : };
5025 2 :
5026 2 : let harness = TenantHarness::create_custom(
5027 2 : "test_get_vectored_key_gap",
5028 2 : tenant_conf,
5029 2 : TenantId::generate(),
5030 2 : ShardIdentity::unsharded(),
5031 2 : Generation::new(0xdeadbeef),
5032 2 : )
5033 2 : .await?;
5034 8 : let (tenant, ctx) = harness.load().await;
5035 2 :
5036 2 : let mut current_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5037 2 : let gap_at_key = current_key.add(100);
5038 2 : let mut current_lsn = Lsn(0x10);
5039 2 :
5040 2 : const KEY_COUNT: usize = 10_000;
5041 2 :
5042 2 : let timeline_id = TimelineId::generate();
5043 2 : let current_timeline = tenant
5044 2 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
5045 4 : .await?;
5046 2 :
5047 2 : current_lsn += 0x100;
5048 2 :
5049 2 : let mut writer = current_timeline.writer().await;
5050 2 : writer
5051 2 : .put(
5052 2 : gap_at_key,
5053 2 : current_lsn,
5054 2 : &Value::Image(test_img(&format!("{} at {}", gap_at_key, current_lsn))),
5055 2 : &ctx,
5056 2 : )
5057 2 : .await?;
5058 2 : writer.finish_write(current_lsn);
5059 2 : drop(writer);
5060 2 :
5061 2 : let mut latest_lsns = HashMap::new();
5062 2 : latest_lsns.insert(gap_at_key, current_lsn);
5063 2 :
5064 2 : current_timeline.freeze_and_flush().await?;
5065 2 :
5066 2 : let child_timeline_id = TimelineId::generate();
5067 2 :
5068 2 : tenant
5069 2 : .branch_timeline_test(
5070 2 : ¤t_timeline,
5071 2 : child_timeline_id,
5072 2 : Some(current_lsn),
5073 2 : &ctx,
5074 2 : )
5075 2 : .await?;
5076 2 : let child_timeline = tenant
5077 2 : .get_timeline(child_timeline_id, true)
5078 2 : .expect("Should have the branched timeline");
5079 2 :
5080 20002 : for i in 0..KEY_COUNT {
5081 20000 : if current_key == gap_at_key {
5082 2 : current_key = current_key.next();
5083 2 : continue;
5084 19998 : }
5085 19998 :
5086 19998 : current_lsn += 0x10;
5087 2 :
5088 19998 : let mut writer = child_timeline.writer().await;
5089 19998 : writer
5090 19998 : .put(
5091 19998 : current_key,
5092 19998 : current_lsn,
5093 19998 : &Value::Image(test_img(&format!("{} at {}", current_key, current_lsn))),
5094 19998 : &ctx,
5095 19998 : )
5096 77 : .await?;
5097 19998 : writer.finish_write(current_lsn);
5098 19998 : drop(writer);
5099 19998 :
5100 19998 : latest_lsns.insert(current_key, current_lsn);
5101 19998 : current_key = current_key.next();
5102 19998 :
5103 19998 : // Flush every now and then to encourage layer file creation.
5104 19998 : if i % 500 == 0 {
5105 43 : child_timeline.freeze_and_flush().await?;
5106 19958 : }
5107 2 : }
5108 2 :
5109 4 : child_timeline.freeze_and_flush().await?;
5110 2 : let mut flags = EnumSet::new();
5111 2 : flags.insert(CompactFlags::ForceRepartition);
5112 2 : child_timeline
5113 2 : .compact(&CancellationToken::new(), flags, &ctx)
5114 2327 : .await?;
5115 2 :
5116 2 : let key_near_end = {
5117 2 : let mut tmp = current_key;
5118 2 : tmp.field6 -= 10;
5119 2 : tmp
5120 2 : };
5121 2 :
5122 2 : let key_near_gap = {
5123 2 : let mut tmp = gap_at_key;
5124 2 : tmp.field6 -= 10;
5125 2 : tmp
5126 2 : };
5127 2 :
5128 2 : let read = KeySpace {
5129 2 : ranges: vec![key_near_gap..gap_at_key.next(), key_near_end..current_key],
5130 2 : };
5131 2 : let results = child_timeline
5132 2 : .get_vectored_impl(
5133 2 : read.clone(),
5134 2 : current_lsn,
5135 2 : &mut ValuesReconstructState::new(),
5136 2 : &ctx,
5137 2 : )
5138 15 : .await?;
5139 2 :
5140 44 : for (key, img_res) in results {
5141 42 : let expected = test_img(&format!("{} at {}", key, latest_lsns[&key]));
5142 42 : assert_eq!(img_res?, expected);
5143 2 : }
5144 2 :
5145 2 : Ok(())
5146 2 : }
5147 :
5148 : // Test that vectored get descends into ancestor timelines correctly and
5149 : // does not return an image that's newer than requested.
5150 : //
5151 : // The diagram below ilustrates an interesting case. We have a parent timeline
5152 : // (top of the Lsn range) and a child timeline. The request key cannot be reconstructed
5153 : // from the child timeline, so the parent timeline must be visited. When advacing into
5154 : // the child timeline, the read path needs to remember what the requested Lsn was in
5155 : // order to avoid returning an image that's too new. The test below constructs such
5156 : // a timeline setup and does a few queries around the Lsn of each page image.
5157 : // ```
5158 : // LSN
5159 : // ^
5160 : // |
5161 : // |
5162 : // 500 | --------------------------------------> branch point
5163 : // 400 | X
5164 : // 300 | X
5165 : // 200 | --------------------------------------> requested lsn
5166 : // 100 | X
5167 : // |---------------------------------------> Key
5168 : // |
5169 : // ------> requested key
5170 : //
5171 : // Legend:
5172 : // * X - page images
5173 : // ```
5174 : #[tokio::test]
5175 2 : async fn test_get_vectored_ancestor_descent() -> anyhow::Result<()> {
5176 2 : let harness = TenantHarness::create("test_get_vectored_on_lsn_axis").await?;
5177 8 : let (tenant, ctx) = harness.load().await;
5178 2 :
5179 2 : let start_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5180 2 : let end_key = start_key.add(1000);
5181 2 : let child_gap_at_key = start_key.add(500);
5182 2 : let mut parent_gap_lsns: BTreeMap<Lsn, String> = BTreeMap::new();
5183 2 :
5184 2 : let mut current_lsn = Lsn(0x10);
5185 2 :
5186 2 : let timeline_id = TimelineId::generate();
5187 2 : let parent_timeline = tenant
5188 2 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
5189 4 : .await?;
5190 2 :
5191 2 : current_lsn += 0x100;
5192 2 :
5193 8 : for _ in 0..3 {
5194 6 : let mut key = start_key;
5195 6006 : while key < end_key {
5196 6000 : current_lsn += 0x10;
5197 6000 :
5198 6000 : let image_value = format!("{} at {}", child_gap_at_key, current_lsn);
5199 2 :
5200 6000 : let mut writer = parent_timeline.writer().await;
5201 6000 : writer
5202 6000 : .put(
5203 6000 : key,
5204 6000 : current_lsn,
5205 6000 : &Value::Image(test_img(&image_value)),
5206 6000 : &ctx,
5207 6000 : )
5208 6 : .await?;
5209 6000 : writer.finish_write(current_lsn);
5210 6000 :
5211 6000 : if key == child_gap_at_key {
5212 6 : parent_gap_lsns.insert(current_lsn, image_value);
5213 5994 : }
5214 2 :
5215 6000 : key = key.next();
5216 2 : }
5217 2 :
5218 6 : parent_timeline.freeze_and_flush().await?;
5219 2 : }
5220 2 :
5221 2 : let child_timeline_id = TimelineId::generate();
5222 2 :
5223 2 : let child_timeline = tenant
5224 2 : .branch_timeline_test(&parent_timeline, child_timeline_id, Some(current_lsn), &ctx)
5225 2 : .await?;
5226 2 :
5227 2 : let mut key = start_key;
5228 2002 : while key < end_key {
5229 2000 : if key == child_gap_at_key {
5230 2 : key = key.next();
5231 2 : continue;
5232 1998 : }
5233 1998 :
5234 1998 : current_lsn += 0x10;
5235 2 :
5236 1998 : let mut writer = child_timeline.writer().await;
5237 1998 : writer
5238 1998 : .put(
5239 1998 : key,
5240 1998 : current_lsn,
5241 1998 : &Value::Image(test_img(&format!("{} at {}", key, current_lsn))),
5242 1998 : &ctx,
5243 1998 : )
5244 17 : .await?;
5245 1998 : writer.finish_write(current_lsn);
5246 1998 :
5247 1998 : key = key.next();
5248 2 : }
5249 2 :
5250 2 : child_timeline.freeze_and_flush().await?;
5251 2 :
5252 2 : let lsn_offsets: [i64; 5] = [-10, -1, 0, 1, 10];
5253 2 : let mut query_lsns = Vec::new();
5254 6 : for image_lsn in parent_gap_lsns.keys().rev() {
5255 36 : for offset in lsn_offsets {
5256 30 : query_lsns.push(Lsn(image_lsn
5257 30 : .0
5258 30 : .checked_add_signed(offset)
5259 30 : .expect("Shouldn't overflow")));
5260 30 : }
5261 2 : }
5262 2 :
5263 32 : for query_lsn in query_lsns {
5264 30 : let results = child_timeline
5265 30 : .get_vectored_impl(
5266 30 : KeySpace {
5267 30 : ranges: vec![child_gap_at_key..child_gap_at_key.next()],
5268 30 : },
5269 30 : query_lsn,
5270 30 : &mut ValuesReconstructState::new(),
5271 30 : &ctx,
5272 30 : )
5273 29 : .await;
5274 2 :
5275 30 : let expected_item = parent_gap_lsns
5276 30 : .iter()
5277 30 : .rev()
5278 68 : .find(|(lsn, _)| **lsn <= query_lsn);
5279 30 :
5280 30 : info!(
5281 2 : "Doing vectored read at LSN {}. Expecting image to be: {:?}",
5282 2 : query_lsn, expected_item
5283 2 : );
5284 2 :
5285 30 : match expected_item {
5286 26 : Some((_, img_value)) => {
5287 26 : let key_results = results.expect("No vectored get error expected");
5288 26 : let key_result = &key_results[&child_gap_at_key];
5289 26 : let returned_img = key_result
5290 26 : .as_ref()
5291 26 : .expect("No page reconstruct error expected");
5292 26 :
5293 26 : info!(
5294 2 : "Vectored read at LSN {} returned image {}",
5295 0 : query_lsn,
5296 0 : std::str::from_utf8(returned_img)?
5297 2 : );
5298 26 : assert_eq!(*returned_img, test_img(img_value));
5299 2 : }
5300 2 : None => {
5301 4 : assert!(matches!(results, Err(GetVectoredError::MissingKey(_))));
5302 2 : }
5303 2 : }
5304 2 : }
5305 2 :
5306 2 : Ok(())
5307 2 : }
5308 :
5309 : #[tokio::test]
5310 2 : async fn test_random_updates() -> anyhow::Result<()> {
5311 2 : let names_algorithms = [
5312 2 : ("test_random_updates_legacy", CompactionAlgorithm::Legacy),
5313 2 : ("test_random_updates_tiered", CompactionAlgorithm::Tiered),
5314 2 : ];
5315 6 : for (name, algorithm) in names_algorithms {
5316 46356 : test_random_updates_algorithm(name, algorithm).await?;
5317 2 : }
5318 2 : Ok(())
5319 2 : }
5320 :
5321 4 : async fn test_random_updates_algorithm(
5322 4 : name: &'static str,
5323 4 : compaction_algorithm: CompactionAlgorithm,
5324 4 : ) -> anyhow::Result<()> {
5325 4 : let mut harness = TenantHarness::create(name).await?;
5326 4 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
5327 4 : kind: compaction_algorithm,
5328 4 : };
5329 16 : let (tenant, ctx) = harness.load().await;
5330 4 : let tline = tenant
5331 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5332 7 : .await?;
5333 :
5334 : const NUM_KEYS: usize = 1000;
5335 4 : let cancel = CancellationToken::new();
5336 4 :
5337 4 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5338 4 : let mut test_key_end = test_key;
5339 4 : test_key_end.field6 = NUM_KEYS as u32;
5340 4 : tline.add_extra_test_dense_keyspace(KeySpace::single(test_key..test_key_end));
5341 4 :
5342 4 : let mut keyspace = KeySpaceAccum::new();
5343 4 :
5344 4 : // Track when each page was last modified. Used to assert that
5345 4 : // a read sees the latest page version.
5346 4 : let mut updated = [Lsn(0); NUM_KEYS];
5347 4 :
5348 4 : let mut lsn = Lsn(0x10);
5349 : #[allow(clippy::needless_range_loop)]
5350 4004 : for blknum in 0..NUM_KEYS {
5351 4000 : lsn = Lsn(lsn.0 + 0x10);
5352 4000 : test_key.field6 = blknum as u32;
5353 4000 : let mut writer = tline.writer().await;
5354 4000 : writer
5355 4000 : .put(
5356 4000 : test_key,
5357 4000 : lsn,
5358 4000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5359 4000 : &ctx,
5360 4000 : )
5361 4 : .await?;
5362 4000 : writer.finish_write(lsn);
5363 4000 : updated[blknum] = lsn;
5364 4000 : drop(writer);
5365 4000 :
5366 4000 : keyspace.add_key(test_key);
5367 : }
5368 :
5369 204 : for _ in 0..50 {
5370 200200 : for _ in 0..NUM_KEYS {
5371 200000 : lsn = Lsn(lsn.0 + 0x10);
5372 200000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
5373 200000 : test_key.field6 = blknum as u32;
5374 200000 : let mut writer = tline.writer().await;
5375 200000 : writer
5376 200000 : .put(
5377 200000 : test_key,
5378 200000 : lsn,
5379 200000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5380 200000 : &ctx,
5381 200000 : )
5382 198 : .await?;
5383 200000 : writer.finish_write(lsn);
5384 200000 : drop(writer);
5385 200000 : updated[blknum] = lsn;
5386 : }
5387 :
5388 : // Read all the blocks
5389 200000 : for (blknum, last_lsn) in updated.iter().enumerate() {
5390 200000 : test_key.field6 = blknum as u32;
5391 200000 : assert_eq!(
5392 200000 : tline.get(test_key, lsn, &ctx).await?,
5393 200000 : test_img(&format!("{} at {}", blknum, last_lsn))
5394 : );
5395 : }
5396 :
5397 : // Perform a cycle of flush, and GC
5398 203 : tline.freeze_and_flush().await?;
5399 200 : tenant
5400 200 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
5401 199 : .await?;
5402 : }
5403 :
5404 4 : Ok(())
5405 4 : }
5406 :
5407 : #[tokio::test]
5408 2 : async fn test_traverse_branches() -> anyhow::Result<()> {
5409 2 : let (tenant, ctx) = TenantHarness::create("test_traverse_branches")
5410 2 : .await?
5411 2 : .load()
5412 8 : .await;
5413 2 : let mut tline = tenant
5414 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5415 4 : .await?;
5416 2 :
5417 2 : const NUM_KEYS: usize = 1000;
5418 2 :
5419 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5420 2 :
5421 2 : let mut keyspace = KeySpaceAccum::new();
5422 2 :
5423 2 : let cancel = CancellationToken::new();
5424 2 :
5425 2 : // Track when each page was last modified. Used to assert that
5426 2 : // a read sees the latest page version.
5427 2 : let mut updated = [Lsn(0); NUM_KEYS];
5428 2 :
5429 2 : let mut lsn = Lsn(0x10);
5430 2 : #[allow(clippy::needless_range_loop)]
5431 2002 : for blknum in 0..NUM_KEYS {
5432 2000 : lsn = Lsn(lsn.0 + 0x10);
5433 2000 : test_key.field6 = blknum as u32;
5434 2000 : let mut writer = tline.writer().await;
5435 2000 : writer
5436 2000 : .put(
5437 2000 : test_key,
5438 2000 : lsn,
5439 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5440 2000 : &ctx,
5441 2000 : )
5442 2 : .await?;
5443 2000 : writer.finish_write(lsn);
5444 2000 : updated[blknum] = lsn;
5445 2000 : drop(writer);
5446 2000 :
5447 2000 : keyspace.add_key(test_key);
5448 2 : }
5449 2 :
5450 102 : for _ in 0..50 {
5451 100 : let new_tline_id = TimelineId::generate();
5452 100 : tenant
5453 100 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
5454 2 : .await?;
5455 100 : tline = tenant
5456 100 : .get_timeline(new_tline_id, true)
5457 100 : .expect("Should have the branched timeline");
5458 2 :
5459 100100 : for _ in 0..NUM_KEYS {
5460 100000 : lsn = Lsn(lsn.0 + 0x10);
5461 100000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
5462 100000 : test_key.field6 = blknum as u32;
5463 100000 : let mut writer = tline.writer().await;
5464 100000 : writer
5465 100000 : .put(
5466 100000 : test_key,
5467 100000 : lsn,
5468 100000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5469 100000 : &ctx,
5470 100000 : )
5471 851 : .await?;
5472 100000 : println!("updating {} at {}", blknum, lsn);
5473 100000 : writer.finish_write(lsn);
5474 100000 : drop(writer);
5475 100000 : updated[blknum] = lsn;
5476 2 : }
5477 2 :
5478 2 : // Read all the blocks
5479 100000 : for (blknum, last_lsn) in updated.iter().enumerate() {
5480 100000 : test_key.field6 = blknum as u32;
5481 100000 : assert_eq!(
5482 100000 : tline.get(test_key, lsn, &ctx).await?,
5483 100000 : test_img(&format!("{} at {}", blknum, last_lsn))
5484 2 : );
5485 2 : }
5486 2 :
5487 2 : // Perform a cycle of flush, compact, and GC
5488 102 : tline.freeze_and_flush().await?;
5489 15788 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
5490 100 : tenant
5491 100 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
5492 98 : .await?;
5493 2 : }
5494 2 :
5495 2 : Ok(())
5496 2 : }
5497 :
5498 : #[tokio::test]
5499 2 : async fn test_traverse_ancestors() -> anyhow::Result<()> {
5500 2 : let (tenant, ctx) = TenantHarness::create("test_traverse_ancestors")
5501 2 : .await?
5502 2 : .load()
5503 8 : .await;
5504 2 : let mut tline = tenant
5505 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5506 4 : .await?;
5507 2 :
5508 2 : const NUM_KEYS: usize = 100;
5509 2 : const NUM_TLINES: usize = 50;
5510 2 :
5511 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5512 2 : // Track page mutation lsns across different timelines.
5513 2 : let mut updated = [[Lsn(0); NUM_KEYS]; NUM_TLINES];
5514 2 :
5515 2 : let mut lsn = Lsn(0x10);
5516 2 :
5517 2 : #[allow(clippy::needless_range_loop)]
5518 102 : for idx in 0..NUM_TLINES {
5519 100 : let new_tline_id = TimelineId::generate();
5520 100 : tenant
5521 100 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
5522 2 : .await?;
5523 100 : tline = tenant
5524 100 : .get_timeline(new_tline_id, true)
5525 100 : .expect("Should have the branched timeline");
5526 2 :
5527 10100 : for _ in 0..NUM_KEYS {
5528 10000 : lsn = Lsn(lsn.0 + 0x10);
5529 10000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
5530 10000 : test_key.field6 = blknum as u32;
5531 10000 : let mut writer = tline.writer().await;
5532 10000 : writer
5533 10000 : .put(
5534 10000 : test_key,
5535 10000 : lsn,
5536 10000 : &Value::Image(test_img(&format!("{} {} at {}", idx, blknum, lsn))),
5537 10000 : &ctx,
5538 10000 : )
5539 88 : .await?;
5540 10000 : println!("updating [{}][{}] at {}", idx, blknum, lsn);
5541 10000 : writer.finish_write(lsn);
5542 10000 : drop(writer);
5543 10000 : updated[idx][blknum] = lsn;
5544 2 : }
5545 2 : }
5546 2 :
5547 2 : // Read pages from leaf timeline across all ancestors.
5548 100 : for (idx, lsns) in updated.iter().enumerate() {
5549 10000 : for (blknum, lsn) in lsns.iter().enumerate() {
5550 2 : // Skip empty mutations.
5551 10000 : if lsn.0 == 0 {
5552 3670 : continue;
5553 6330 : }
5554 6330 : println!("checking [{idx}][{blknum}] at {lsn}");
5555 6330 : test_key.field6 = blknum as u32;
5556 6330 : assert_eq!(
5557 6330 : tline.get(test_key, *lsn, &ctx).await?,
5558 6330 : test_img(&format!("{idx} {blknum} at {lsn}"))
5559 2 : );
5560 2 : }
5561 2 : }
5562 2 : Ok(())
5563 2 : }
5564 :
5565 : #[tokio::test]
5566 2 : async fn test_write_at_initdb_lsn_takes_optimization_code_path() -> anyhow::Result<()> {
5567 2 : let (tenant, ctx) = TenantHarness::create("test_empty_test_timeline_is_usable")
5568 2 : .await?
5569 2 : .load()
5570 8 : .await;
5571 2 :
5572 2 : let initdb_lsn = Lsn(0x20);
5573 2 : let utline = tenant
5574 2 : .create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)
5575 2 : .await?;
5576 2 : let tline = utline.raw_timeline().unwrap();
5577 2 :
5578 2 : // Spawn flush loop now so that we can set the `expect_initdb_optimization`
5579 2 : tline.maybe_spawn_flush_loop();
5580 2 :
5581 2 : // Make sure the timeline has the minimum set of required keys for operation.
5582 2 : // The only operation you can always do on an empty timeline is to `put` new data.
5583 2 : // Except if you `put` at `initdb_lsn`.
5584 2 : // In that case, there's an optimization to directly create image layers instead of delta layers.
5585 2 : // It uses `repartition()`, which assumes some keys to be present.
5586 2 : // Let's make sure the test timeline can handle that case.
5587 2 : {
5588 2 : let mut state = tline.flush_loop_state.lock().unwrap();
5589 2 : assert_eq!(
5590 2 : timeline::FlushLoopState::Running {
5591 2 : expect_initdb_optimization: false,
5592 2 : initdb_optimization_count: 0,
5593 2 : },
5594 2 : *state
5595 2 : );
5596 2 : *state = timeline::FlushLoopState::Running {
5597 2 : expect_initdb_optimization: true,
5598 2 : initdb_optimization_count: 0,
5599 2 : };
5600 2 : }
5601 2 :
5602 2 : // Make writes at the initdb_lsn. When we flush it below, it should be handled by the optimization.
5603 2 : // As explained above, the optimization requires some keys to be present.
5604 2 : // As per `create_empty_timeline` documentation, use init_empty to set them.
5605 2 : // This is what `create_test_timeline` does, by the way.
5606 2 : let mut modification = tline.begin_modification(initdb_lsn);
5607 2 : modification
5608 2 : .init_empty_test_timeline()
5609 2 : .context("init_empty_test_timeline")?;
5610 2 : modification
5611 2 : .commit(&ctx)
5612 2 : .await
5613 2 : .context("commit init_empty_test_timeline modification")?;
5614 2 :
5615 2 : // Do the flush. The flush code will check the expectations that we set above.
5616 2 : tline.freeze_and_flush().await?;
5617 2 :
5618 2 : // assert freeze_and_flush exercised the initdb optimization
5619 2 : {
5620 2 : let state = tline.flush_loop_state.lock().unwrap();
5621 2 : let timeline::FlushLoopState::Running {
5622 2 : expect_initdb_optimization,
5623 2 : initdb_optimization_count,
5624 2 : } = *state
5625 2 : else {
5626 2 : panic!("unexpected state: {:?}", *state);
5627 2 : };
5628 2 : assert!(expect_initdb_optimization);
5629 2 : assert!(initdb_optimization_count > 0);
5630 2 : }
5631 2 : Ok(())
5632 2 : }
5633 :
5634 : #[tokio::test]
5635 2 : async fn test_create_guard_crash() -> anyhow::Result<()> {
5636 2 : let name = "test_create_guard_crash";
5637 2 : let harness = TenantHarness::create(name).await?;
5638 2 : {
5639 8 : let (tenant, ctx) = harness.load().await;
5640 2 : let tline = tenant
5641 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
5642 2 : .await?;
5643 2 : // Leave the timeline ID in [`Tenant::timelines_creating`] to exclude attempting to create it again
5644 2 : let raw_tline = tline.raw_timeline().unwrap();
5645 2 : raw_tline
5646 2 : .shutdown(super::timeline::ShutdownMode::Hard)
5647 2 : .instrument(info_span!("test_shutdown", tenant_id=%raw_tline.tenant_shard_id, shard_id=%raw_tline.tenant_shard_id.shard_slug(), timeline_id=%TIMELINE_ID))
5648 2 : .await;
5649 2 : std::mem::forget(tline);
5650 2 : }
5651 2 :
5652 8 : let (tenant, _) = harness.load().await;
5653 2 : match tenant.get_timeline(TIMELINE_ID, false) {
5654 2 : Ok(_) => panic!("timeline should've been removed during load"),
5655 2 : Err(e) => {
5656 2 : assert_eq!(
5657 2 : e,
5658 2 : GetTimelineError::NotFound {
5659 2 : tenant_id: tenant.tenant_shard_id,
5660 2 : timeline_id: TIMELINE_ID,
5661 2 : }
5662 2 : )
5663 2 : }
5664 2 : }
5665 2 :
5666 2 : assert!(!harness
5667 2 : .conf
5668 2 : .timeline_path(&tenant.tenant_shard_id, &TIMELINE_ID)
5669 2 : .exists());
5670 2 :
5671 2 : Ok(())
5672 2 : }
5673 :
5674 : #[tokio::test]
5675 2 : async fn test_read_at_max_lsn() -> anyhow::Result<()> {
5676 2 : let names_algorithms = [
5677 2 : ("test_read_at_max_lsn_legacy", CompactionAlgorithm::Legacy),
5678 2 : ("test_read_at_max_lsn_tiered", CompactionAlgorithm::Tiered),
5679 2 : ];
5680 6 : for (name, algorithm) in names_algorithms {
5681 32833 : test_read_at_max_lsn_algorithm(name, algorithm).await?;
5682 2 : }
5683 2 : Ok(())
5684 2 : }
5685 :
5686 4 : async fn test_read_at_max_lsn_algorithm(
5687 4 : name: &'static str,
5688 4 : compaction_algorithm: CompactionAlgorithm,
5689 4 : ) -> anyhow::Result<()> {
5690 4 : let mut harness = TenantHarness::create(name).await?;
5691 4 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
5692 4 : kind: compaction_algorithm,
5693 4 : };
5694 16 : let (tenant, ctx) = harness.load().await;
5695 4 : let tline = tenant
5696 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5697 7 : .await?;
5698 :
5699 4 : let lsn = Lsn(0x10);
5700 4 : let compact = false;
5701 32400 : bulk_insert_maybe_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000, compact).await?;
5702 :
5703 4 : let test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5704 4 : let read_lsn = Lsn(u64::MAX - 1);
5705 :
5706 410 : let result = tline.get(test_key, read_lsn, &ctx).await;
5707 4 : assert!(result.is_ok(), "result is not Ok: {}", result.unwrap_err());
5708 :
5709 4 : Ok(())
5710 4 : }
5711 :
5712 : #[tokio::test]
5713 2 : async fn test_metadata_scan() -> anyhow::Result<()> {
5714 2 : let harness = TenantHarness::create("test_metadata_scan").await?;
5715 8 : let (tenant, ctx) = harness.load().await;
5716 2 : let tline = tenant
5717 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5718 4 : .await?;
5719 2 :
5720 2 : const NUM_KEYS: usize = 1000;
5721 2 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
5722 2 :
5723 2 : let cancel = CancellationToken::new();
5724 2 :
5725 2 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
5726 2 : base_key.field1 = AUX_KEY_PREFIX;
5727 2 : let mut test_key = base_key;
5728 2 :
5729 2 : // Track when each page was last modified. Used to assert that
5730 2 : // a read sees the latest page version.
5731 2 : let mut updated = [Lsn(0); NUM_KEYS];
5732 2 :
5733 2 : let mut lsn = Lsn(0x10);
5734 2 : #[allow(clippy::needless_range_loop)]
5735 2002 : for blknum in 0..NUM_KEYS {
5736 2000 : lsn = Lsn(lsn.0 + 0x10);
5737 2000 : test_key.field6 = (blknum * STEP) as u32;
5738 2000 : let mut writer = tline.writer().await;
5739 2000 : writer
5740 2000 : .put(
5741 2000 : test_key,
5742 2000 : lsn,
5743 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5744 2000 : &ctx,
5745 2000 : )
5746 2 : .await?;
5747 2000 : writer.finish_write(lsn);
5748 2000 : updated[blknum] = lsn;
5749 2000 : drop(writer);
5750 2 : }
5751 2 :
5752 2 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
5753 2 :
5754 24 : for iter in 0..=10 {
5755 2 : // Read all the blocks
5756 22000 : for (blknum, last_lsn) in updated.iter().enumerate() {
5757 22000 : test_key.field6 = (blknum * STEP) as u32;
5758 22000 : assert_eq!(
5759 22000 : tline.get(test_key, lsn, &ctx).await?,
5760 22000 : test_img(&format!("{} at {}", blknum, last_lsn))
5761 2 : );
5762 2 : }
5763 2 :
5764 22 : let mut cnt = 0;
5765 22000 : for (key, value) in tline
5766 22 : .get_vectored_impl(
5767 22 : keyspace.clone(),
5768 22 : lsn,
5769 22 : &mut ValuesReconstructState::default(),
5770 22 : &ctx,
5771 22 : )
5772 5552 : .await?
5773 2 : {
5774 22000 : let blknum = key.field6 as usize;
5775 22000 : let value = value?;
5776 22000 : assert!(blknum % STEP == 0);
5777 22000 : let blknum = blknum / STEP;
5778 22000 : assert_eq!(
5779 22000 : value,
5780 22000 : test_img(&format!("{} at {}", blknum, updated[blknum]))
5781 22000 : );
5782 22000 : cnt += 1;
5783 2 : }
5784 2 :
5785 22 : assert_eq!(cnt, NUM_KEYS);
5786 2 :
5787 22022 : for _ in 0..NUM_KEYS {
5788 22000 : lsn = Lsn(lsn.0 + 0x10);
5789 22000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
5790 22000 : test_key.field6 = (blknum * STEP) as u32;
5791 22000 : let mut writer = tline.writer().await;
5792 22000 : writer
5793 22000 : .put(
5794 22000 : test_key,
5795 22000 : lsn,
5796 22000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5797 22000 : &ctx,
5798 22000 : )
5799 170 : .await?;
5800 22000 : writer.finish_write(lsn);
5801 22000 : drop(writer);
5802 22000 : updated[blknum] = lsn;
5803 2 : }
5804 2 :
5805 2 : // Perform two cycles of flush, compact, and GC
5806 66 : for round in 0..2 {
5807 44 : tline.freeze_and_flush().await?;
5808 44 : tline
5809 44 : .compact(
5810 44 : &cancel,
5811 44 : if iter % 5 == 0 && round == 0 {
5812 6 : let mut flags = EnumSet::new();
5813 6 : flags.insert(CompactFlags::ForceImageLayerCreation);
5814 6 : flags.insert(CompactFlags::ForceRepartition);
5815 6 : flags
5816 2 : } else {
5817 38 : EnumSet::empty()
5818 2 : },
5819 44 : &ctx,
5820 2 : )
5821 8922 : .await?;
5822 44 : tenant
5823 44 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
5824 44 : .await?;
5825 2 : }
5826 2 : }
5827 2 :
5828 2 : Ok(())
5829 2 : }
5830 :
5831 : #[tokio::test]
5832 2 : async fn test_metadata_compaction_trigger() -> anyhow::Result<()> {
5833 2 : let harness = TenantHarness::create("test_metadata_compaction_trigger").await?;
5834 8 : let (tenant, ctx) = harness.load().await;
5835 2 : let tline = tenant
5836 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5837 4 : .await?;
5838 2 :
5839 2 : let cancel = CancellationToken::new();
5840 2 :
5841 2 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
5842 2 : base_key.field1 = AUX_KEY_PREFIX;
5843 2 : let test_key = base_key;
5844 2 : let mut lsn = Lsn(0x10);
5845 2 :
5846 42 : for _ in 0..20 {
5847 40 : lsn = Lsn(lsn.0 + 0x10);
5848 40 : let mut writer = tline.writer().await;
5849 40 : writer
5850 40 : .put(
5851 40 : test_key,
5852 40 : lsn,
5853 40 : &Value::Image(test_img(&format!("{} at {}", 0, lsn))),
5854 40 : &ctx,
5855 40 : )
5856 20 : .await?;
5857 40 : writer.finish_write(lsn);
5858 40 : drop(writer);
5859 40 : tline.freeze_and_flush().await?; // force create a delta layer
5860 2 : }
5861 2 :
5862 2 : let before_num_l0_delta_files = tline
5863 2 : .layers
5864 2 : .read()
5865 2 : .await
5866 2 : .layer_map()
5867 2 : .get_level0_deltas()
5868 2 : .len();
5869 2 :
5870 150 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
5871 2 :
5872 2 : let after_num_l0_delta_files = tline
5873 2 : .layers
5874 2 : .read()
5875 2 : .await
5876 2 : .layer_map()
5877 2 : .get_level0_deltas()
5878 2 : .len();
5879 2 :
5880 2 : assert!(after_num_l0_delta_files < before_num_l0_delta_files, "after_num_l0_delta_files={after_num_l0_delta_files}, before_num_l0_delta_files={before_num_l0_delta_files}");
5881 2 :
5882 2 : assert_eq!(
5883 4 : tline.get(test_key, lsn, &ctx).await?,
5884 2 : test_img(&format!("{} at {}", 0, lsn))
5885 2 : );
5886 2 :
5887 2 : Ok(())
5888 2 : }
5889 :
5890 : #[tokio::test]
5891 2 : async fn test_branch_copies_dirty_aux_file_flag() {
5892 2 : let harness = TenantHarness::create("test_branch_copies_dirty_aux_file_flag")
5893 2 : .await
5894 2 : .unwrap();
5895 2 :
5896 2 : // the default aux file policy to switch is v1 if not set by the admins
5897 2 : assert_eq!(
5898 2 : harness.tenant_conf.switch_aux_file_policy,
5899 2 : AuxFilePolicy::V1
5900 2 : );
5901 8 : let (tenant, ctx) = harness.load().await;
5902 2 :
5903 2 : let mut lsn = Lsn(0x08);
5904 2 :
5905 2 : let tline: Arc<Timeline> = tenant
5906 2 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
5907 4 : .await
5908 2 : .unwrap();
5909 2 :
5910 2 : // no aux file is written at this point, so the persistent flag should be unset
5911 2 : assert_eq!(tline.last_aux_file_policy.load(), None);
5912 2 :
5913 2 : {
5914 2 : lsn += 8;
5915 2 : let mut modification = tline.begin_modification(lsn);
5916 2 : modification
5917 2 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
5918 5 : .await
5919 2 : .unwrap();
5920 2 : modification.commit(&ctx).await.unwrap();
5921 2 : }
5922 2 :
5923 2 : // there is no tenant manager to pass the configuration through, so lets mimic it
5924 2 : tenant.set_new_location_config(
5925 2 : AttachedTenantConf::try_from(LocationConf::attached_single(
5926 2 : TenantConfOpt {
5927 2 : switch_aux_file_policy: Some(AuxFilePolicy::V2),
5928 2 : ..Default::default()
5929 2 : },
5930 2 : tenant.generation,
5931 2 : &pageserver_api::models::ShardParameters::default(),
5932 2 : ))
5933 2 : .unwrap(),
5934 2 : );
5935 2 :
5936 2 : assert_eq!(
5937 2 : tline.get_switch_aux_file_policy(),
5938 2 : AuxFilePolicy::V2,
5939 2 : "wanted state has been updated"
5940 2 : );
5941 2 : assert_eq!(
5942 2 : tline.last_aux_file_policy.load(),
5943 2 : Some(AuxFilePolicy::V1),
5944 2 : "aux file is written with switch_aux_file_policy unset (which is v1), so we should keep v1"
5945 2 : );
5946 2 :
5947 2 : // we can read everything from the storage
5948 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
5949 2 : assert_eq!(
5950 2 : files.get("pg_logical/mappings/test1"),
5951 2 : Some(&bytes::Bytes::from_static(b"first"))
5952 2 : );
5953 2 :
5954 2 : {
5955 2 : lsn += 8;
5956 2 : let mut modification = tline.begin_modification(lsn);
5957 2 : modification
5958 2 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
5959 2 : .await
5960 2 : .unwrap();
5961 2 : modification.commit(&ctx).await.unwrap();
5962 2 : }
5963 2 :
5964 2 : assert_eq!(
5965 2 : tline.last_aux_file_policy.load(),
5966 2 : Some(AuxFilePolicy::V1),
5967 2 : "keep v1 storage format when new files are written"
5968 2 : );
5969 2 :
5970 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
5971 2 : assert_eq!(
5972 2 : files.get("pg_logical/mappings/test2"),
5973 2 : Some(&bytes::Bytes::from_static(b"second"))
5974 2 : );
5975 2 :
5976 2 : let child = tenant
5977 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(lsn), &ctx)
5978 2 : .await
5979 2 : .unwrap();
5980 2 :
5981 2 : // child copies the last flag even if that is not on remote storage yet
5982 2 : assert_eq!(child.get_switch_aux_file_policy(), AuxFilePolicy::V2);
5983 2 : assert_eq!(child.last_aux_file_policy.load(), Some(AuxFilePolicy::V1));
5984 2 :
5985 2 : let files = child.list_aux_files(lsn, &ctx).await.unwrap();
5986 2 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
5987 2 : assert_eq!(files.get("pg_logical/mappings/test2"), None);
5988 2 :
5989 2 : // even if we crash here without flushing parent timeline with it's new
5990 2 : // last_aux_file_policy we are safe, because child was never meant to access ancestor's
5991 2 : // files. the ancestor can even switch back to V1 because of a migration safely.
5992 2 : }
5993 :
5994 : #[tokio::test]
5995 2 : async fn aux_file_policy_switch() {
5996 2 : let mut harness = TenantHarness::create("aux_file_policy_switch")
5997 2 : .await
5998 2 : .unwrap();
5999 2 : harness.tenant_conf.switch_aux_file_policy = AuxFilePolicy::CrossValidation; // set to cross-validation mode
6000 8 : let (tenant, ctx) = harness.load().await;
6001 2 :
6002 2 : let mut lsn = Lsn(0x08);
6003 2 :
6004 2 : let tline: Arc<Timeline> = tenant
6005 2 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
6006 4 : .await
6007 2 : .unwrap();
6008 2 :
6009 2 : assert_eq!(
6010 2 : tline.last_aux_file_policy.load(),
6011 2 : None,
6012 2 : "no aux file is written so it should be unset"
6013 2 : );
6014 2 :
6015 2 : {
6016 2 : lsn += 8;
6017 2 : let mut modification = tline.begin_modification(lsn);
6018 2 : modification
6019 2 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
6020 5 : .await
6021 2 : .unwrap();
6022 2 : modification.commit(&ctx).await.unwrap();
6023 2 : }
6024 2 :
6025 2 : // there is no tenant manager to pass the configuration through, so lets mimic it
6026 2 : tenant.set_new_location_config(
6027 2 : AttachedTenantConf::try_from(LocationConf::attached_single(
6028 2 : TenantConfOpt {
6029 2 : switch_aux_file_policy: Some(AuxFilePolicy::V2),
6030 2 : ..Default::default()
6031 2 : },
6032 2 : tenant.generation,
6033 2 : &pageserver_api::models::ShardParameters::default(),
6034 2 : ))
6035 2 : .unwrap(),
6036 2 : );
6037 2 :
6038 2 : assert_eq!(
6039 2 : tline.get_switch_aux_file_policy(),
6040 2 : AuxFilePolicy::V2,
6041 2 : "wanted state has been updated"
6042 2 : );
6043 2 : assert_eq!(
6044 2 : tline.last_aux_file_policy.load(),
6045 2 : Some(AuxFilePolicy::CrossValidation),
6046 2 : "dirty index_part.json reflected state is yet to be updated"
6047 2 : );
6048 2 :
6049 2 : // we can still read the auxfile v1 before we ingest anything new
6050 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6051 2 : assert_eq!(
6052 2 : files.get("pg_logical/mappings/test1"),
6053 2 : Some(&bytes::Bytes::from_static(b"first"))
6054 2 : );
6055 2 :
6056 2 : {
6057 2 : lsn += 8;
6058 2 : let mut modification = tline.begin_modification(lsn);
6059 2 : modification
6060 2 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
6061 2 : .await
6062 2 : .unwrap();
6063 2 : modification.commit(&ctx).await.unwrap();
6064 2 : }
6065 2 :
6066 2 : assert_eq!(
6067 2 : tline.last_aux_file_policy.load(),
6068 2 : Some(AuxFilePolicy::V2),
6069 2 : "ingesting a file should apply the wanted switch state when applicable"
6070 2 : );
6071 2 :
6072 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6073 2 : assert_eq!(
6074 2 : files.get("pg_logical/mappings/test1"),
6075 2 : Some(&bytes::Bytes::from_static(b"first")),
6076 2 : "cross validation writes to both v1 and v2 so this should be available in v2"
6077 2 : );
6078 2 : assert_eq!(
6079 2 : files.get("pg_logical/mappings/test2"),
6080 2 : Some(&bytes::Bytes::from_static(b"second"))
6081 2 : );
6082 2 :
6083 2 : // mimic again by trying to flip it from V2 to V1 (not switched to while ingesting a file)
6084 2 : tenant.set_new_location_config(
6085 2 : AttachedTenantConf::try_from(LocationConf::attached_single(
6086 2 : TenantConfOpt {
6087 2 : switch_aux_file_policy: Some(AuxFilePolicy::V1),
6088 2 : ..Default::default()
6089 2 : },
6090 2 : tenant.generation,
6091 2 : &pageserver_api::models::ShardParameters::default(),
6092 2 : ))
6093 2 : .unwrap(),
6094 2 : );
6095 2 :
6096 2 : {
6097 2 : lsn += 8;
6098 2 : let mut modification = tline.begin_modification(lsn);
6099 2 : modification
6100 2 : .put_file("pg_logical/mappings/test2", b"third", &ctx)
6101 2 : .await
6102 2 : .unwrap();
6103 2 : modification.commit(&ctx).await.unwrap();
6104 2 : }
6105 2 :
6106 2 : assert_eq!(
6107 2 : tline.get_switch_aux_file_policy(),
6108 2 : AuxFilePolicy::V1,
6109 2 : "wanted state has been updated again, even if invalid request"
6110 2 : );
6111 2 :
6112 2 : assert_eq!(
6113 2 : tline.last_aux_file_policy.load(),
6114 2 : Some(AuxFilePolicy::V2),
6115 2 : "ingesting a file should apply the wanted switch state when applicable"
6116 2 : );
6117 2 :
6118 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6119 2 : assert_eq!(
6120 2 : files.get("pg_logical/mappings/test1"),
6121 2 : Some(&bytes::Bytes::from_static(b"first"))
6122 2 : );
6123 2 : assert_eq!(
6124 2 : files.get("pg_logical/mappings/test2"),
6125 2 : Some(&bytes::Bytes::from_static(b"third"))
6126 2 : );
6127 2 :
6128 2 : // mimic again by trying to flip it from from V1 to V2 (not switched to while ingesting a file)
6129 2 : tenant.set_new_location_config(
6130 2 : AttachedTenantConf::try_from(LocationConf::attached_single(
6131 2 : TenantConfOpt {
6132 2 : switch_aux_file_policy: Some(AuxFilePolicy::V2),
6133 2 : ..Default::default()
6134 2 : },
6135 2 : tenant.generation,
6136 2 : &pageserver_api::models::ShardParameters::default(),
6137 2 : ))
6138 2 : .unwrap(),
6139 2 : );
6140 2 :
6141 2 : {
6142 2 : lsn += 8;
6143 2 : let mut modification = tline.begin_modification(lsn);
6144 2 : modification
6145 2 : .put_file("pg_logical/mappings/test3", b"last", &ctx)
6146 2 : .await
6147 2 : .unwrap();
6148 2 : modification.commit(&ctx).await.unwrap();
6149 2 : }
6150 2 :
6151 2 : assert_eq!(tline.get_switch_aux_file_policy(), AuxFilePolicy::V2);
6152 2 :
6153 2 : assert_eq!(tline.last_aux_file_policy.load(), Some(AuxFilePolicy::V2));
6154 2 :
6155 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6156 2 : assert_eq!(
6157 2 : files.get("pg_logical/mappings/test1"),
6158 2 : Some(&bytes::Bytes::from_static(b"first"))
6159 2 : );
6160 2 : assert_eq!(
6161 2 : files.get("pg_logical/mappings/test2"),
6162 2 : Some(&bytes::Bytes::from_static(b"third"))
6163 2 : );
6164 2 : assert_eq!(
6165 2 : files.get("pg_logical/mappings/test3"),
6166 2 : Some(&bytes::Bytes::from_static(b"last"))
6167 2 : );
6168 2 : }
6169 :
6170 : #[tokio::test]
6171 2 : async fn aux_file_policy_force_switch() {
6172 2 : let mut harness = TenantHarness::create("aux_file_policy_force_switch")
6173 2 : .await
6174 2 : .unwrap();
6175 2 : harness.tenant_conf.switch_aux_file_policy = AuxFilePolicy::V1;
6176 8 : let (tenant, ctx) = harness.load().await;
6177 2 :
6178 2 : let mut lsn = Lsn(0x08);
6179 2 :
6180 2 : let tline: Arc<Timeline> = tenant
6181 2 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
6182 4 : .await
6183 2 : .unwrap();
6184 2 :
6185 2 : assert_eq!(
6186 2 : tline.last_aux_file_policy.load(),
6187 2 : None,
6188 2 : "no aux file is written so it should be unset"
6189 2 : );
6190 2 :
6191 2 : {
6192 2 : lsn += 8;
6193 2 : let mut modification = tline.begin_modification(lsn);
6194 2 : modification
6195 2 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
6196 5 : .await
6197 2 : .unwrap();
6198 2 : modification.commit(&ctx).await.unwrap();
6199 2 : }
6200 2 :
6201 2 : tline.do_switch_aux_policy(AuxFilePolicy::V2).unwrap();
6202 2 :
6203 2 : assert_eq!(
6204 2 : tline.last_aux_file_policy.load(),
6205 2 : Some(AuxFilePolicy::V2),
6206 2 : "dirty index_part.json reflected state is yet to be updated"
6207 2 : );
6208 2 :
6209 2 : // lose all data from v1
6210 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6211 2 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
6212 2 :
6213 2 : {
6214 2 : lsn += 8;
6215 2 : let mut modification = tline.begin_modification(lsn);
6216 2 : modification
6217 2 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
6218 2 : .await
6219 2 : .unwrap();
6220 2 : modification.commit(&ctx).await.unwrap();
6221 2 : }
6222 2 :
6223 2 : // read data ingested in v2
6224 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6225 2 : assert_eq!(
6226 2 : files.get("pg_logical/mappings/test2"),
6227 2 : Some(&bytes::Bytes::from_static(b"second"))
6228 2 : );
6229 2 : // lose all data from v1
6230 2 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
6231 2 : }
6232 :
6233 : #[tokio::test]
6234 2 : async fn aux_file_policy_auto_detect() {
6235 2 : let mut harness = TenantHarness::create("aux_file_policy_auto_detect")
6236 2 : .await
6237 2 : .unwrap();
6238 2 : harness.tenant_conf.switch_aux_file_policy = AuxFilePolicy::V2; // set to cross-validation mode
6239 8 : let (tenant, ctx) = harness.load().await;
6240 2 :
6241 2 : let mut lsn = Lsn(0x08);
6242 2 :
6243 2 : let tline: Arc<Timeline> = tenant
6244 2 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
6245 4 : .await
6246 2 : .unwrap();
6247 2 :
6248 2 : assert_eq!(
6249 2 : tline.last_aux_file_policy.load(),
6250 2 : None,
6251 2 : "no aux file is written so it should be unset"
6252 2 : );
6253 2 :
6254 2 : {
6255 2 : lsn += 8;
6256 2 : let mut modification = tline.begin_modification(lsn);
6257 2 : let buf = AuxFilesDirectory::ser(&AuxFilesDirectory {
6258 2 : files: vec![(
6259 2 : "test_file".to_string(),
6260 2 : Bytes::copy_from_slice(b"test_file"),
6261 2 : )]
6262 2 : .into_iter()
6263 2 : .collect(),
6264 2 : })
6265 2 : .unwrap();
6266 2 : modification.put_for_test(AUX_FILES_KEY, Value::Image(Bytes::from(buf)));
6267 2 : modification.commit(&ctx).await.unwrap();
6268 2 : }
6269 2 :
6270 2 : {
6271 2 : lsn += 8;
6272 2 : let mut modification = tline.begin_modification(lsn);
6273 2 : modification
6274 2 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
6275 2 : .await
6276 2 : .unwrap();
6277 2 : modification.commit(&ctx).await.unwrap();
6278 2 : }
6279 2 :
6280 2 : assert_eq!(
6281 2 : tline.last_aux_file_policy.load(),
6282 2 : Some(AuxFilePolicy::V1),
6283 2 : "keep using v1 because there are aux files writting with v1"
6284 2 : );
6285 2 :
6286 2 : // we can still read the auxfile v1
6287 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6288 2 : assert_eq!(
6289 2 : files.get("pg_logical/mappings/test1"),
6290 2 : Some(&bytes::Bytes::from_static(b"first"))
6291 2 : );
6292 2 : assert_eq!(
6293 2 : files.get("test_file"),
6294 2 : Some(&bytes::Bytes::from_static(b"test_file"))
6295 2 : );
6296 2 : }
6297 :
6298 : #[tokio::test]
6299 2 : async fn test_metadata_image_creation() -> anyhow::Result<()> {
6300 2 : let harness = TenantHarness::create("test_metadata_image_creation").await?;
6301 8 : let (tenant, ctx) = harness.load().await;
6302 2 : let tline = tenant
6303 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6304 4 : .await?;
6305 2 :
6306 2 : const NUM_KEYS: usize = 1000;
6307 2 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
6308 2 :
6309 2 : let cancel = CancellationToken::new();
6310 2 :
6311 2 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
6312 2 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
6313 2 : let mut test_key = base_key;
6314 2 : let mut lsn = Lsn(0x10);
6315 2 :
6316 8 : async fn scan_with_statistics(
6317 8 : tline: &Timeline,
6318 8 : keyspace: &KeySpace,
6319 8 : lsn: Lsn,
6320 8 : ctx: &RequestContext,
6321 8 : ) -> anyhow::Result<(BTreeMap<Key, Result<Bytes, PageReconstructError>>, usize)> {
6322 8 : let mut reconstruct_state = ValuesReconstructState::default();
6323 8 : let res = tline
6324 8 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
6325 1674 : .await?;
6326 8 : Ok((res, reconstruct_state.get_delta_layers_visited() as usize))
6327 8 : }
6328 2 :
6329 2 : #[allow(clippy::needless_range_loop)]
6330 2002 : for blknum in 0..NUM_KEYS {
6331 2000 : lsn = Lsn(lsn.0 + 0x10);
6332 2000 : test_key.field6 = (blknum * STEP) as u32;
6333 2000 : let mut writer = tline.writer().await;
6334 2000 : writer
6335 2000 : .put(
6336 2000 : test_key,
6337 2000 : lsn,
6338 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6339 2000 : &ctx,
6340 2000 : )
6341 2 : .await?;
6342 2000 : writer.finish_write(lsn);
6343 2000 : drop(writer);
6344 2 : }
6345 2 :
6346 2 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
6347 2 :
6348 22 : for iter in 1..=10 {
6349 20020 : for _ in 0..NUM_KEYS {
6350 20000 : lsn = Lsn(lsn.0 + 0x10);
6351 20000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6352 20000 : test_key.field6 = (blknum * STEP) as u32;
6353 20000 : let mut writer = tline.writer().await;
6354 20000 : writer
6355 20000 : .put(
6356 20000 : test_key,
6357 20000 : lsn,
6358 20000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6359 20000 : &ctx,
6360 20000 : )
6361 19 : .await?;
6362 20000 : writer.finish_write(lsn);
6363 20000 : drop(writer);
6364 2 : }
6365 2 :
6366 20 : tline.freeze_and_flush().await?;
6367 2 :
6368 20 : if iter % 5 == 0 {
6369 4 : let (_, before_delta_file_accessed) =
6370 1666 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
6371 4 : tline
6372 4 : .compact(
6373 4 : &cancel,
6374 4 : {
6375 4 : let mut flags = EnumSet::new();
6376 4 : flags.insert(CompactFlags::ForceImageLayerCreation);
6377 4 : flags.insert(CompactFlags::ForceRepartition);
6378 4 : flags
6379 4 : },
6380 4 : &ctx,
6381 4 : )
6382 6517 : .await?;
6383 4 : let (_, after_delta_file_accessed) =
6384 8 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
6385 4 : assert!(after_delta_file_accessed < before_delta_file_accessed, "after_delta_file_accessed={after_delta_file_accessed}, before_delta_file_accessed={before_delta_file_accessed}");
6386 2 : // Given that we already produced an image layer, there should be no delta layer needed for the scan, but still setting a low threshold there for unforeseen circumstances.
6387 4 : assert!(
6388 4 : after_delta_file_accessed <= 2,
6389 2 : "after_delta_file_accessed={after_delta_file_accessed}"
6390 2 : );
6391 16 : }
6392 2 : }
6393 2 :
6394 2 : Ok(())
6395 2 : }
6396 :
6397 : #[tokio::test]
6398 2 : async fn test_vectored_missing_data_key_reads() -> anyhow::Result<()> {
6399 2 : let harness = TenantHarness::create("test_vectored_missing_data_key_reads").await?;
6400 8 : let (tenant, ctx) = harness.load().await;
6401 2 :
6402 2 : let base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
6403 2 : let base_key_child = Key::from_hex("000000000033333333444444445500000001").unwrap();
6404 2 : let base_key_nonexist = Key::from_hex("000000000033333333444444445500000002").unwrap();
6405 2 :
6406 2 : let tline = tenant
6407 2 : .create_test_timeline_with_layers(
6408 2 : TIMELINE_ID,
6409 2 : Lsn(0x10),
6410 2 : DEFAULT_PG_VERSION,
6411 2 : &ctx,
6412 2 : Vec::new(), // delta layers
6413 2 : vec![(Lsn(0x20), vec![(base_key, test_img("data key 1"))])], // image layers
6414 2 : Lsn(0x20), // it's fine to not advance LSN to 0x30 while using 0x30 to get below because `get_vectored_impl` does not wait for LSN
6415 2 : )
6416 11 : .await?;
6417 2 : tline.add_extra_test_dense_keyspace(KeySpace::single(base_key..(base_key_nonexist.next())));
6418 2 :
6419 2 : let child = tenant
6420 2 : .branch_timeline_test_with_layers(
6421 2 : &tline,
6422 2 : NEW_TIMELINE_ID,
6423 2 : Some(Lsn(0x20)),
6424 2 : &ctx,
6425 2 : Vec::new(), // delta layers
6426 2 : vec![(Lsn(0x30), vec![(base_key_child, test_img("data key 2"))])], // image layers
6427 2 : Lsn(0x30),
6428 2 : )
6429 7 : .await
6430 2 : .unwrap();
6431 2 :
6432 2 : let lsn = Lsn(0x30);
6433 2 :
6434 2 : // test vectored get on parent timeline
6435 2 : assert_eq!(
6436 4 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
6437 2 : Some(test_img("data key 1"))
6438 2 : );
6439 2 : assert!(get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx)
6440 3 : .await
6441 2 : .unwrap_err()
6442 2 : .is_missing_key_error());
6443 2 : assert!(
6444 2 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx)
6445 2 : .await
6446 2 : .unwrap_err()
6447 2 : .is_missing_key_error()
6448 2 : );
6449 2 :
6450 2 : // test vectored get on child timeline
6451 2 : assert_eq!(
6452 2 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
6453 2 : Some(test_img("data key 1"))
6454 2 : );
6455 2 : assert_eq!(
6456 4 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
6457 2 : Some(test_img("data key 2"))
6458 2 : );
6459 2 : assert!(
6460 2 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx)
6461 2 : .await
6462 2 : .unwrap_err()
6463 2 : .is_missing_key_error()
6464 2 : );
6465 2 :
6466 2 : Ok(())
6467 2 : }
6468 :
6469 : #[tokio::test]
6470 2 : async fn test_vectored_missing_metadata_key_reads() -> anyhow::Result<()> {
6471 2 : let harness = TenantHarness::create("test_vectored_missing_metadata_key_reads").await?;
6472 8 : let (tenant, ctx) = harness.load().await;
6473 2 :
6474 2 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
6475 2 : let base_key_child = Key::from_hex("620000000033333333444444445500000001").unwrap();
6476 2 : let base_key_nonexist = Key::from_hex("620000000033333333444444445500000002").unwrap();
6477 2 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
6478 2 :
6479 2 : let tline = tenant
6480 2 : .create_test_timeline_with_layers(
6481 2 : TIMELINE_ID,
6482 2 : Lsn(0x10),
6483 2 : DEFAULT_PG_VERSION,
6484 2 : &ctx,
6485 2 : Vec::new(), // delta layers
6486 2 : vec![(Lsn(0x20), vec![(base_key, test_img("metadata key 1"))])], // image layers
6487 2 : Lsn(0x20), // it's fine to not advance LSN to 0x30 while using 0x30 to get below because `get_vectored_impl` does not wait for LSN
6488 2 : )
6489 11 : .await?;
6490 2 :
6491 2 : let child = tenant
6492 2 : .branch_timeline_test_with_layers(
6493 2 : &tline,
6494 2 : NEW_TIMELINE_ID,
6495 2 : Some(Lsn(0x20)),
6496 2 : &ctx,
6497 2 : Vec::new(), // delta layers
6498 2 : vec![(
6499 2 : Lsn(0x30),
6500 2 : vec![(base_key_child, test_img("metadata key 2"))],
6501 2 : )], // image layers
6502 2 : Lsn(0x30),
6503 2 : )
6504 7 : .await
6505 2 : .unwrap();
6506 2 :
6507 2 : let lsn = Lsn(0x30);
6508 2 :
6509 2 : // test vectored get on parent timeline
6510 2 : assert_eq!(
6511 4 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
6512 2 : Some(test_img("metadata key 1"))
6513 2 : );
6514 2 : assert_eq!(
6515 2 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx).await?,
6516 2 : None
6517 2 : );
6518 2 : assert_eq!(
6519 2 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx).await?,
6520 2 : None
6521 2 : );
6522 2 :
6523 2 : // test vectored get on child timeline
6524 2 : assert_eq!(
6525 2 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
6526 2 : None
6527 2 : );
6528 2 : assert_eq!(
6529 4 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
6530 2 : Some(test_img("metadata key 2"))
6531 2 : );
6532 2 : assert_eq!(
6533 2 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx).await?,
6534 2 : None
6535 2 : );
6536 2 :
6537 2 : Ok(())
6538 2 : }
6539 :
6540 36 : async fn get_vectored_impl_wrapper(
6541 36 : tline: &Arc<Timeline>,
6542 36 : key: Key,
6543 36 : lsn: Lsn,
6544 36 : ctx: &RequestContext,
6545 36 : ) -> Result<Option<Bytes>, GetVectoredError> {
6546 36 : let mut reconstruct_state = ValuesReconstructState::new();
6547 36 : let mut res = tline
6548 36 : .get_vectored_impl(
6549 36 : KeySpace::single(key..key.next()),
6550 36 : lsn,
6551 36 : &mut reconstruct_state,
6552 36 : ctx,
6553 36 : )
6554 33 : .await?;
6555 30 : Ok(res.pop_last().map(|(k, v)| {
6556 18 : assert_eq!(k, key);
6557 18 : v.unwrap()
6558 30 : }))
6559 36 : }
6560 :
6561 : #[tokio::test]
6562 2 : async fn test_metadata_tombstone_reads() -> anyhow::Result<()> {
6563 2 : let harness = TenantHarness::create("test_metadata_tombstone_reads").await?;
6564 8 : let (tenant, ctx) = harness.load().await;
6565 2 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
6566 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
6567 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
6568 2 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
6569 2 :
6570 2 : // We emulate the situation that the compaction algorithm creates an image layer that removes the tombstones
6571 2 : // Lsn 0x30 key0, key3, no key1+key2
6572 2 : // Lsn 0x20 key1+key2 tomestones
6573 2 : // Lsn 0x10 key1 in image, key2 in delta
6574 2 : let tline = tenant
6575 2 : .create_test_timeline_with_layers(
6576 2 : TIMELINE_ID,
6577 2 : Lsn(0x10),
6578 2 : DEFAULT_PG_VERSION,
6579 2 : &ctx,
6580 2 : // delta layers
6581 2 : vec![
6582 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6583 2 : Lsn(0x10)..Lsn(0x20),
6584 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
6585 2 : ),
6586 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6587 2 : Lsn(0x20)..Lsn(0x30),
6588 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
6589 2 : ),
6590 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6591 2 : Lsn(0x20)..Lsn(0x30),
6592 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
6593 2 : ),
6594 2 : ],
6595 2 : // image layers
6596 2 : vec![
6597 2 : (Lsn(0x10), vec![(key1, test_img("metadata key 1"))]),
6598 2 : (
6599 2 : Lsn(0x30),
6600 2 : vec![
6601 2 : (key0, test_img("metadata key 0")),
6602 2 : (key3, test_img("metadata key 3")),
6603 2 : ],
6604 2 : ),
6605 2 : ],
6606 2 : Lsn(0x30),
6607 2 : )
6608 38 : .await?;
6609 2 :
6610 2 : let lsn = Lsn(0x30);
6611 2 : let old_lsn = Lsn(0x20);
6612 2 :
6613 2 : assert_eq!(
6614 4 : get_vectored_impl_wrapper(&tline, key0, lsn, &ctx).await?,
6615 2 : Some(test_img("metadata key 0"))
6616 2 : );
6617 2 : assert_eq!(
6618 2 : get_vectored_impl_wrapper(&tline, key1, lsn, &ctx).await?,
6619 2 : None,
6620 2 : );
6621 2 : assert_eq!(
6622 2 : get_vectored_impl_wrapper(&tline, key2, lsn, &ctx).await?,
6623 2 : None,
6624 2 : );
6625 2 : assert_eq!(
6626 4 : get_vectored_impl_wrapper(&tline, key1, old_lsn, &ctx).await?,
6627 2 : Some(Bytes::new()),
6628 2 : );
6629 2 : assert_eq!(
6630 4 : get_vectored_impl_wrapper(&tline, key2, old_lsn, &ctx).await?,
6631 2 : Some(Bytes::new()),
6632 2 : );
6633 2 : assert_eq!(
6634 2 : get_vectored_impl_wrapper(&tline, key3, lsn, &ctx).await?,
6635 2 : Some(test_img("metadata key 3"))
6636 2 : );
6637 2 :
6638 2 : Ok(())
6639 2 : }
6640 :
6641 : #[tokio::test]
6642 2 : async fn test_metadata_tombstone_image_creation() {
6643 2 : let harness = TenantHarness::create("test_metadata_tombstone_image_creation")
6644 2 : .await
6645 2 : .unwrap();
6646 8 : let (tenant, ctx) = harness.load().await;
6647 2 :
6648 2 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
6649 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
6650 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
6651 2 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
6652 2 :
6653 2 : let tline = tenant
6654 2 : .create_test_timeline_with_layers(
6655 2 : TIMELINE_ID,
6656 2 : Lsn(0x10),
6657 2 : DEFAULT_PG_VERSION,
6658 2 : &ctx,
6659 2 : // delta layers
6660 2 : vec![
6661 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6662 2 : Lsn(0x10)..Lsn(0x20),
6663 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
6664 2 : ),
6665 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6666 2 : Lsn(0x20)..Lsn(0x30),
6667 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
6668 2 : ),
6669 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6670 2 : Lsn(0x20)..Lsn(0x30),
6671 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
6672 2 : ),
6673 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6674 2 : Lsn(0x30)..Lsn(0x40),
6675 2 : vec![
6676 2 : (key0, Lsn(0x30), Value::Image(test_img("metadata key 0"))),
6677 2 : (key3, Lsn(0x30), Value::Image(test_img("metadata key 3"))),
6678 2 : ],
6679 2 : ),
6680 2 : ],
6681 2 : // image layers
6682 2 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
6683 2 : Lsn(0x40),
6684 2 : )
6685 35 : .await
6686 2 : .unwrap();
6687 2 :
6688 2 : let cancel = CancellationToken::new();
6689 2 :
6690 2 : tline
6691 2 : .compact(
6692 2 : &cancel,
6693 2 : {
6694 2 : let mut flags = EnumSet::new();
6695 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
6696 2 : flags.insert(CompactFlags::ForceRepartition);
6697 2 : flags
6698 2 : },
6699 2 : &ctx,
6700 2 : )
6701 58 : .await
6702 2 : .unwrap();
6703 2 :
6704 2 : // Image layers are created at last_record_lsn
6705 2 : let images = tline
6706 2 : .inspect_image_layers(Lsn(0x40), &ctx)
6707 8 : .await
6708 2 : .unwrap()
6709 2 : .into_iter()
6710 20 : .filter(|(k, _)| k.is_metadata_key())
6711 2 : .collect::<Vec<_>>();
6712 2 : assert_eq!(images.len(), 2); // the image layer should only contain two existing keys, tombstones should be removed.
6713 2 : }
6714 :
6715 : #[tokio::test]
6716 2 : async fn test_metadata_tombstone_empty_image_creation() {
6717 2 : let harness = TenantHarness::create("test_metadata_tombstone_empty_image_creation")
6718 2 : .await
6719 2 : .unwrap();
6720 8 : let (tenant, ctx) = harness.load().await;
6721 2 :
6722 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
6723 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
6724 2 :
6725 2 : let tline = tenant
6726 2 : .create_test_timeline_with_layers(
6727 2 : TIMELINE_ID,
6728 2 : Lsn(0x10),
6729 2 : DEFAULT_PG_VERSION,
6730 2 : &ctx,
6731 2 : // delta layers
6732 2 : vec![
6733 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6734 2 : Lsn(0x10)..Lsn(0x20),
6735 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
6736 2 : ),
6737 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6738 2 : Lsn(0x20)..Lsn(0x30),
6739 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
6740 2 : ),
6741 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6742 2 : Lsn(0x20)..Lsn(0x30),
6743 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
6744 2 : ),
6745 2 : ],
6746 2 : // image layers
6747 2 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
6748 2 : Lsn(0x30),
6749 2 : )
6750 29 : .await
6751 2 : .unwrap();
6752 2 :
6753 2 : let cancel = CancellationToken::new();
6754 2 :
6755 2 : tline
6756 2 : .compact(
6757 2 : &cancel,
6758 2 : {
6759 2 : let mut flags = EnumSet::new();
6760 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
6761 2 : flags.insert(CompactFlags::ForceRepartition);
6762 2 : flags
6763 2 : },
6764 2 : &ctx,
6765 2 : )
6766 46 : .await
6767 2 : .unwrap();
6768 2 :
6769 2 : // Image layers are created at last_record_lsn
6770 2 : let images = tline
6771 2 : .inspect_image_layers(Lsn(0x30), &ctx)
6772 4 : .await
6773 2 : .unwrap()
6774 2 : .into_iter()
6775 16 : .filter(|(k, _)| k.is_metadata_key())
6776 2 : .collect::<Vec<_>>();
6777 2 : assert_eq!(images.len(), 0); // the image layer should not contain tombstones, or it is not created
6778 2 : }
6779 :
6780 : #[tokio::test]
6781 2 : async fn test_simple_bottom_most_compaction_images() -> anyhow::Result<()> {
6782 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_images").await?;
6783 8 : let (tenant, ctx) = harness.load().await;
6784 2 :
6785 102 : fn get_key(id: u32) -> Key {
6786 102 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
6787 102 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
6788 102 : key.field6 = id;
6789 102 : key
6790 102 : }
6791 2 :
6792 2 : // We create
6793 2 : // - one bottom-most image layer,
6794 2 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
6795 2 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
6796 2 : // - a delta layer D3 above the horizon.
6797 2 : //
6798 2 : // | D3 |
6799 2 : // | D1 |
6800 2 : // -| |-- gc horizon -----------------
6801 2 : // | | | D2 |
6802 2 : // --------- img layer ------------------
6803 2 : //
6804 2 : // What we should expact from this compaction is:
6805 2 : // | D3 |
6806 2 : // | Part of D1 |
6807 2 : // --------- img layer with D1+D2 at GC horizon------------------
6808 2 :
6809 2 : // img layer at 0x10
6810 2 : let img_layer = (0..10)
6811 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
6812 2 : .collect_vec();
6813 2 :
6814 2 : let delta1 = vec![
6815 2 : (
6816 2 : get_key(1),
6817 2 : Lsn(0x20),
6818 2 : Value::Image(Bytes::from("value 1@0x20")),
6819 2 : ),
6820 2 : (
6821 2 : get_key(2),
6822 2 : Lsn(0x30),
6823 2 : Value::Image(Bytes::from("value 2@0x30")),
6824 2 : ),
6825 2 : (
6826 2 : get_key(3),
6827 2 : Lsn(0x40),
6828 2 : Value::Image(Bytes::from("value 3@0x40")),
6829 2 : ),
6830 2 : ];
6831 2 : let delta2 = vec![
6832 2 : (
6833 2 : get_key(5),
6834 2 : Lsn(0x20),
6835 2 : Value::Image(Bytes::from("value 5@0x20")),
6836 2 : ),
6837 2 : (
6838 2 : get_key(6),
6839 2 : Lsn(0x20),
6840 2 : Value::Image(Bytes::from("value 6@0x20")),
6841 2 : ),
6842 2 : ];
6843 2 : let delta3 = vec![
6844 2 : (
6845 2 : get_key(8),
6846 2 : Lsn(0x48),
6847 2 : Value::Image(Bytes::from("value 8@0x48")),
6848 2 : ),
6849 2 : (
6850 2 : get_key(9),
6851 2 : Lsn(0x48),
6852 2 : Value::Image(Bytes::from("value 9@0x48")),
6853 2 : ),
6854 2 : ];
6855 2 :
6856 2 : let tline = tenant
6857 2 : .create_test_timeline_with_layers(
6858 2 : TIMELINE_ID,
6859 2 : Lsn(0x10),
6860 2 : DEFAULT_PG_VERSION,
6861 2 : &ctx,
6862 2 : vec![
6863 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
6864 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
6865 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
6866 2 : ], // delta layers
6867 2 : vec![(Lsn(0x10), img_layer)], // image layers
6868 2 : Lsn(0x50),
6869 2 : )
6870 47 : .await?;
6871 2 : {
6872 2 : // Update GC info
6873 2 : let mut guard = tline.gc_info.write().unwrap();
6874 2 : guard.cutoffs.time = Lsn(0x30);
6875 2 : guard.cutoffs.space = Lsn(0x30);
6876 2 : }
6877 2 :
6878 2 : let expected_result = [
6879 2 : Bytes::from_static(b"value 0@0x10"),
6880 2 : Bytes::from_static(b"value 1@0x20"),
6881 2 : Bytes::from_static(b"value 2@0x30"),
6882 2 : Bytes::from_static(b"value 3@0x40"),
6883 2 : Bytes::from_static(b"value 4@0x10"),
6884 2 : Bytes::from_static(b"value 5@0x20"),
6885 2 : Bytes::from_static(b"value 6@0x20"),
6886 2 : Bytes::from_static(b"value 7@0x10"),
6887 2 : Bytes::from_static(b"value 8@0x48"),
6888 2 : Bytes::from_static(b"value 9@0x48"),
6889 2 : ];
6890 2 :
6891 20 : for (idx, expected) in expected_result.iter().enumerate() {
6892 20 : assert_eq!(
6893 20 : tline
6894 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
6895 23 : .await
6896 20 : .unwrap(),
6897 2 : expected
6898 2 : );
6899 2 : }
6900 2 :
6901 2 : let cancel = CancellationToken::new();
6902 55 : tline.compact_with_gc(&cancel, &ctx).await.unwrap();
6903 2 :
6904 20 : for (idx, expected) in expected_result.iter().enumerate() {
6905 20 : assert_eq!(
6906 20 : tline
6907 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
6908 16 : .await
6909 20 : .unwrap(),
6910 2 : expected
6911 2 : );
6912 2 : }
6913 2 :
6914 2 : // Check if the image layer at the GC horizon contains exactly what we want
6915 2 : let image_at_gc_horizon = tline
6916 2 : .inspect_image_layers(Lsn(0x30), &ctx)
6917 2 : .await
6918 2 : .unwrap()
6919 2 : .into_iter()
6920 36 : .filter(|(k, _)| k.is_metadata_key())
6921 2 : .collect::<Vec<_>>();
6922 2 :
6923 2 : assert_eq!(image_at_gc_horizon.len(), 10);
6924 2 : let expected_result = [
6925 2 : Bytes::from_static(b"value 0@0x10"),
6926 2 : Bytes::from_static(b"value 1@0x20"),
6927 2 : Bytes::from_static(b"value 2@0x30"),
6928 2 : Bytes::from_static(b"value 3@0x10"),
6929 2 : Bytes::from_static(b"value 4@0x10"),
6930 2 : Bytes::from_static(b"value 5@0x20"),
6931 2 : Bytes::from_static(b"value 6@0x20"),
6932 2 : Bytes::from_static(b"value 7@0x10"),
6933 2 : Bytes::from_static(b"value 8@0x10"),
6934 2 : Bytes::from_static(b"value 9@0x10"),
6935 2 : ];
6936 22 : for idx in 0..10 {
6937 20 : assert_eq!(
6938 20 : image_at_gc_horizon[idx],
6939 20 : (get_key(idx as u32), expected_result[idx].clone())
6940 20 : );
6941 2 : }
6942 2 :
6943 2 : // Check if old layers are removed / new layers have the expected LSN
6944 2 : let mut all_layers = tline.inspect_historic_layers().await.unwrap();
6945 4 : all_layers.sort_by(|k1, k2| {
6946 4 : (
6947 4 : k1.is_delta,
6948 4 : k1.key_range.start,
6949 4 : k1.key_range.end,
6950 4 : k1.lsn_range.start,
6951 4 : k1.lsn_range.end,
6952 4 : )
6953 4 : .cmp(&(
6954 4 : k2.is_delta,
6955 4 : k2.key_range.start,
6956 4 : k2.key_range.end,
6957 4 : k2.lsn_range.start,
6958 4 : k2.lsn_range.end,
6959 4 : ))
6960 4 : });
6961 2 : assert_eq!(
6962 2 : all_layers,
6963 2 : vec![
6964 2 : // Image layer at GC horizon
6965 2 : PersistentLayerKey {
6966 2 : key_range: {
6967 2 : let mut key = Key::MAX;
6968 2 : key.field6 -= 1;
6969 2 : Key::MIN..key
6970 2 : },
6971 2 : lsn_range: Lsn(0x30)..Lsn(0x31),
6972 2 : is_delta: false
6973 2 : },
6974 2 : // The delta layer that is cut in the middle
6975 2 : PersistentLayerKey {
6976 2 : key_range: get_key(3)..get_key(4),
6977 2 : lsn_range: Lsn(0x30)..Lsn(0x41),
6978 2 : is_delta: true
6979 2 : },
6980 2 : // The delta3 layer that should not be picked for the compaction
6981 2 : PersistentLayerKey {
6982 2 : key_range: get_key(8)..get_key(10),
6983 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
6984 2 : is_delta: true
6985 2 : }
6986 2 : ]
6987 2 : );
6988 2 :
6989 2 : // increase GC horizon and compact again
6990 2 : {
6991 2 : // Update GC info
6992 2 : let mut guard = tline.gc_info.write().unwrap();
6993 2 : guard.cutoffs.time = Lsn(0x40);
6994 2 : guard.cutoffs.space = Lsn(0x40);
6995 2 : }
6996 44 : tline.compact_with_gc(&cancel, &ctx).await.unwrap();
6997 2 :
6998 2 : Ok(())
6999 2 : }
7000 :
7001 : #[tokio::test]
7002 2 : async fn test_neon_test_record() -> anyhow::Result<()> {
7003 2 : let harness = TenantHarness::create("test_neon_test_record").await?;
7004 8 : let (tenant, ctx) = harness.load().await;
7005 2 :
7006 24 : fn get_key(id: u32) -> Key {
7007 24 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7008 24 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7009 24 : key.field6 = id;
7010 24 : key
7011 24 : }
7012 2 :
7013 2 : let delta1 = vec![
7014 2 : (
7015 2 : get_key(1),
7016 2 : Lsn(0x20),
7017 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
7018 2 : ),
7019 2 : (
7020 2 : get_key(1),
7021 2 : Lsn(0x30),
7022 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
7023 2 : ),
7024 2 : (get_key(2), Lsn(0x10), Value::Image("0x10".into())),
7025 2 : (
7026 2 : get_key(2),
7027 2 : Lsn(0x20),
7028 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
7029 2 : ),
7030 2 : (
7031 2 : get_key(2),
7032 2 : Lsn(0x30),
7033 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
7034 2 : ),
7035 2 : (get_key(3), Lsn(0x10), Value::Image("0x10".into())),
7036 2 : (
7037 2 : get_key(3),
7038 2 : Lsn(0x20),
7039 2 : Value::WalRecord(NeonWalRecord::wal_clear()),
7040 2 : ),
7041 2 : (get_key(4), Lsn(0x10), Value::Image("0x10".into())),
7042 2 : (
7043 2 : get_key(4),
7044 2 : Lsn(0x20),
7045 2 : Value::WalRecord(NeonWalRecord::wal_init()),
7046 2 : ),
7047 2 : ];
7048 2 : let image1 = vec![(get_key(1), "0x10".into())];
7049 2 :
7050 2 : let tline = tenant
7051 2 : .create_test_timeline_with_layers(
7052 2 : TIMELINE_ID,
7053 2 : Lsn(0x10),
7054 2 : DEFAULT_PG_VERSION,
7055 2 : &ctx,
7056 2 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
7057 2 : Lsn(0x10)..Lsn(0x40),
7058 2 : delta1,
7059 2 : )], // delta layers
7060 2 : vec![(Lsn(0x10), image1)], // image layers
7061 2 : Lsn(0x50),
7062 2 : )
7063 17 : .await?;
7064 2 :
7065 2 : assert_eq!(
7066 8 : tline.get(get_key(1), Lsn(0x50), &ctx).await?,
7067 2 : Bytes::from_static(b"0x10,0x20,0x30")
7068 2 : );
7069 2 : assert_eq!(
7070 2 : tline.get(get_key(2), Lsn(0x50), &ctx).await?,
7071 2 : Bytes::from_static(b"0x10,0x20,0x30")
7072 2 : );
7073 2 :
7074 2 : // Need to remove the limit of "Neon WAL redo requires base image".
7075 2 :
7076 2 : // assert_eq!(tline.get(get_key(3), Lsn(0x50), &ctx).await?, Bytes::new());
7077 2 : // assert_eq!(tline.get(get_key(4), Lsn(0x50), &ctx).await?, Bytes::new());
7078 2 :
7079 2 : Ok(())
7080 2 : }
7081 :
7082 : #[tokio::test]
7083 2 : async fn test_lsn_lease() -> anyhow::Result<()> {
7084 8 : let (tenant, ctx) = TenantHarness::create("test_lsn_lease").await?.load().await;
7085 2 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7086 2 :
7087 2 : let end_lsn = Lsn(0x100);
7088 2 : let image_layers = (0x20..=0x90)
7089 2 : .step_by(0x10)
7090 16 : .map(|n| {
7091 16 : (
7092 16 : Lsn(n),
7093 16 : vec![(key, test_img(&format!("data key at {:x}", n)))],
7094 16 : )
7095 16 : })
7096 2 : .collect();
7097 2 :
7098 2 : let timeline = tenant
7099 2 : .create_test_timeline_with_layers(
7100 2 : TIMELINE_ID,
7101 2 : Lsn(0x10),
7102 2 : DEFAULT_PG_VERSION,
7103 2 : &ctx,
7104 2 : Vec::new(),
7105 2 : image_layers,
7106 2 : end_lsn,
7107 2 : )
7108 60 : .await?;
7109 2 :
7110 2 : let leased_lsns = [0x30, 0x50, 0x70];
7111 2 : let mut leases = Vec::new();
7112 6 : let _: anyhow::Result<_> = leased_lsns.iter().try_for_each(|n| {
7113 6 : leases.push(timeline.make_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)?);
7114 6 : Ok(())
7115 6 : });
7116 2 :
7117 2 : // Renewing with shorter lease should not change the lease.
7118 2 : let updated_lease_0 =
7119 2 : timeline.make_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)?;
7120 2 : assert_eq!(updated_lease_0.valid_until, leases[0].valid_until);
7121 2 :
7122 2 : // Renewing with a long lease should renew lease with later expiration time.
7123 2 : let updated_lease_1 = timeline.make_lsn_lease(
7124 2 : Lsn(leased_lsns[1]),
7125 2 : timeline.get_lsn_lease_length() * 2,
7126 2 : &ctx,
7127 2 : )?;
7128 2 :
7129 2 : assert!(updated_lease_1.valid_until > leases[1].valid_until);
7130 2 :
7131 2 : // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
7132 2 : info!(
7133 2 : "latest_gc_cutoff_lsn: {}",
7134 0 : *timeline.get_latest_gc_cutoff_lsn()
7135 2 : );
7136 2 : timeline.force_set_disk_consistent_lsn(end_lsn);
7137 2 :
7138 2 : let res = tenant
7139 2 : .gc_iteration(
7140 2 : Some(TIMELINE_ID),
7141 2 : 0,
7142 2 : Duration::ZERO,
7143 2 : &CancellationToken::new(),
7144 2 : &ctx,
7145 2 : )
7146 2 : .await?;
7147 2 :
7148 2 : // Keeping everything <= Lsn(0x80) b/c leases:
7149 2 : // 0/10: initdb layer
7150 2 : // (0/20..=0/70).step_by(0x10): image layers added when creating the timeline.
7151 2 : assert_eq!(res.layers_needed_by_leases, 7);
7152 2 : // Keeping 0/90 b/c it is the latest layer.
7153 2 : assert_eq!(res.layers_not_updated, 1);
7154 2 : // Removed 0/80.
7155 2 : assert_eq!(res.layers_removed, 1);
7156 2 :
7157 2 : // Make lease on a already GC-ed LSN.
7158 2 : // 0/80 does not have a valid lease + is below latest_gc_cutoff
7159 2 : assert!(Lsn(0x80) < *timeline.get_latest_gc_cutoff_lsn());
7160 2 : let res = timeline.make_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx);
7161 2 : assert!(res.is_err());
7162 2 :
7163 2 : // Should still be able to renew a currently valid lease
7164 2 : // Assumption: original lease to is still valid for 0/50.
7165 2 : let _ =
7166 2 : timeline.make_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)?;
7167 2 :
7168 2 : Ok(())
7169 2 : }
7170 :
7171 : #[tokio::test]
7172 2 : async fn test_simple_bottom_most_compaction_deltas() -> anyhow::Result<()> {
7173 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_deltas").await?;
7174 8 : let (tenant, ctx) = harness.load().await;
7175 2 :
7176 118 : fn get_key(id: u32) -> Key {
7177 118 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7178 118 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7179 118 : key.field6 = id;
7180 118 : key
7181 118 : }
7182 2 :
7183 2 : // We create
7184 2 : // - one bottom-most image layer,
7185 2 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
7186 2 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
7187 2 : // - a delta layer D3 above the horizon.
7188 2 : //
7189 2 : // | D3 |
7190 2 : // | D1 |
7191 2 : // -| |-- gc horizon -----------------
7192 2 : // | | | D2 |
7193 2 : // --------- img layer ------------------
7194 2 : //
7195 2 : // What we should expact from this compaction is:
7196 2 : // | D3 |
7197 2 : // | Part of D1 |
7198 2 : // --------- img layer with D1+D2 at GC horizon------------------
7199 2 :
7200 2 : // img layer at 0x10
7201 2 : let img_layer = (0..10)
7202 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
7203 2 : .collect_vec();
7204 2 :
7205 2 : let delta1 = vec![
7206 2 : (
7207 2 : get_key(1),
7208 2 : Lsn(0x20),
7209 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7210 2 : ),
7211 2 : (
7212 2 : get_key(2),
7213 2 : Lsn(0x30),
7214 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7215 2 : ),
7216 2 : (
7217 2 : get_key(3),
7218 2 : Lsn(0x28),
7219 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
7220 2 : ),
7221 2 : (
7222 2 : get_key(3),
7223 2 : Lsn(0x30),
7224 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7225 2 : ),
7226 2 : (
7227 2 : get_key(3),
7228 2 : Lsn(0x40),
7229 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
7230 2 : ),
7231 2 : ];
7232 2 : let delta2 = vec![
7233 2 : (
7234 2 : get_key(5),
7235 2 : Lsn(0x20),
7236 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7237 2 : ),
7238 2 : (
7239 2 : get_key(6),
7240 2 : Lsn(0x20),
7241 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7242 2 : ),
7243 2 : ];
7244 2 : let delta3 = vec![
7245 2 : (
7246 2 : get_key(8),
7247 2 : Lsn(0x48),
7248 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
7249 2 : ),
7250 2 : (
7251 2 : get_key(9),
7252 2 : Lsn(0x48),
7253 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
7254 2 : ),
7255 2 : ];
7256 2 :
7257 2 : let tline = tenant
7258 2 : .create_test_timeline_with_layers(
7259 2 : TIMELINE_ID,
7260 2 : Lsn(0x10),
7261 2 : DEFAULT_PG_VERSION,
7262 2 : &ctx,
7263 2 : vec![
7264 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
7265 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
7266 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
7267 2 : ], // delta layers
7268 2 : vec![(Lsn(0x10), img_layer)], // image layers
7269 2 : Lsn(0x50),
7270 2 : )
7271 47 : .await?;
7272 2 : {
7273 2 : // Update GC info
7274 2 : let mut guard = tline.gc_info.write().unwrap();
7275 2 : *guard = GcInfo {
7276 2 : retain_lsns: vec![],
7277 2 : cutoffs: GcCutoffs {
7278 2 : time: Lsn(0x30),
7279 2 : space: Lsn(0x30),
7280 2 : },
7281 2 : leases: Default::default(),
7282 2 : within_ancestor_pitr: false,
7283 2 : };
7284 2 : }
7285 2 :
7286 2 : let expected_result = [
7287 2 : Bytes::from_static(b"value 0@0x10"),
7288 2 : Bytes::from_static(b"value 1@0x10@0x20"),
7289 2 : Bytes::from_static(b"value 2@0x10@0x30"),
7290 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
7291 2 : Bytes::from_static(b"value 4@0x10"),
7292 2 : Bytes::from_static(b"value 5@0x10@0x20"),
7293 2 : Bytes::from_static(b"value 6@0x10@0x20"),
7294 2 : Bytes::from_static(b"value 7@0x10"),
7295 2 : Bytes::from_static(b"value 8@0x10@0x48"),
7296 2 : Bytes::from_static(b"value 9@0x10@0x48"),
7297 2 : ];
7298 2 :
7299 2 : let expected_result_at_gc_horizon = [
7300 2 : Bytes::from_static(b"value 0@0x10"),
7301 2 : Bytes::from_static(b"value 1@0x10@0x20"),
7302 2 : Bytes::from_static(b"value 2@0x10@0x30"),
7303 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
7304 2 : Bytes::from_static(b"value 4@0x10"),
7305 2 : Bytes::from_static(b"value 5@0x10@0x20"),
7306 2 : Bytes::from_static(b"value 6@0x10@0x20"),
7307 2 : Bytes::from_static(b"value 7@0x10"),
7308 2 : Bytes::from_static(b"value 8@0x10"),
7309 2 : Bytes::from_static(b"value 9@0x10"),
7310 2 : ];
7311 2 :
7312 22 : for idx in 0..10 {
7313 20 : assert_eq!(
7314 20 : tline
7315 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7316 31 : .await
7317 20 : .unwrap(),
7318 20 : &expected_result[idx]
7319 2 : );
7320 20 : assert_eq!(
7321 20 : tline
7322 20 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
7323 15 : .await
7324 20 : .unwrap(),
7325 20 : &expected_result_at_gc_horizon[idx]
7326 2 : );
7327 2 : }
7328 2 :
7329 2 : let cancel = CancellationToken::new();
7330 54 : tline.compact_with_gc(&cancel, &ctx).await.unwrap();
7331 2 :
7332 22 : for idx in 0..10 {
7333 20 : assert_eq!(
7334 20 : tline
7335 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7336 20 : .await
7337 20 : .unwrap(),
7338 20 : &expected_result[idx]
7339 2 : );
7340 20 : assert_eq!(
7341 20 : tline
7342 20 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
7343 10 : .await
7344 20 : .unwrap(),
7345 20 : &expected_result_at_gc_horizon[idx]
7346 2 : );
7347 2 : }
7348 2 :
7349 2 : // increase GC horizon and compact again
7350 2 : {
7351 2 : // Update GC info
7352 2 : let mut guard = tline.gc_info.write().unwrap();
7353 2 : guard.cutoffs.time = Lsn(0x40);
7354 2 : guard.cutoffs.space = Lsn(0x40);
7355 2 : }
7356 44 : tline.compact_with_gc(&cancel, &ctx).await.unwrap();
7357 2 :
7358 2 : Ok(())
7359 2 : }
7360 :
7361 : #[tokio::test]
7362 2 : async fn test_generate_key_retention() -> anyhow::Result<()> {
7363 2 : let harness = TenantHarness::create("test_generate_key_retention").await?;
7364 8 : let (tenant, ctx) = harness.load().await;
7365 2 : let tline = tenant
7366 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7367 4 : .await?;
7368 2 : tline.force_advance_lsn(Lsn(0x70));
7369 2 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7370 2 : let history = vec![
7371 2 : (
7372 2 : key,
7373 2 : Lsn(0x10),
7374 2 : Value::Image(Bytes::copy_from_slice(b"0x10")),
7375 2 : ),
7376 2 : (
7377 2 : key,
7378 2 : Lsn(0x20),
7379 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7380 2 : ),
7381 2 : (
7382 2 : key,
7383 2 : Lsn(0x30),
7384 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
7385 2 : ),
7386 2 : (
7387 2 : key,
7388 2 : Lsn(0x40),
7389 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
7390 2 : ),
7391 2 : (
7392 2 : key,
7393 2 : Lsn(0x50),
7394 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
7395 2 : ),
7396 2 : (
7397 2 : key,
7398 2 : Lsn(0x60),
7399 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7400 2 : ),
7401 2 : (
7402 2 : key,
7403 2 : Lsn(0x70),
7404 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7405 2 : ),
7406 2 : (
7407 2 : key,
7408 2 : Lsn(0x80),
7409 2 : Value::Image(Bytes::copy_from_slice(
7410 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
7411 2 : )),
7412 2 : ),
7413 2 : (
7414 2 : key,
7415 2 : Lsn(0x90),
7416 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
7417 2 : ),
7418 2 : ];
7419 2 : let res = tline
7420 2 : .generate_key_retention(
7421 2 : key,
7422 2 : &history,
7423 2 : Lsn(0x60),
7424 2 : &[Lsn(0x20), Lsn(0x40), Lsn(0x50)],
7425 2 : 3,
7426 2 : None,
7427 2 : )
7428 2 : .await
7429 2 : .unwrap();
7430 2 : let expected_res = KeyHistoryRetention {
7431 2 : below_horizon: vec![
7432 2 : (
7433 2 : Lsn(0x20),
7434 2 : KeyLogAtLsn(vec![(
7435 2 : Lsn(0x20),
7436 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20")),
7437 2 : )]),
7438 2 : ),
7439 2 : (
7440 2 : Lsn(0x40),
7441 2 : KeyLogAtLsn(vec![
7442 2 : (
7443 2 : Lsn(0x30),
7444 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
7445 2 : ),
7446 2 : (
7447 2 : Lsn(0x40),
7448 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
7449 2 : ),
7450 2 : ]),
7451 2 : ),
7452 2 : (
7453 2 : Lsn(0x50),
7454 2 : KeyLogAtLsn(vec![(
7455 2 : Lsn(0x50),
7456 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40;0x50")),
7457 2 : )]),
7458 2 : ),
7459 2 : (
7460 2 : Lsn(0x60),
7461 2 : KeyLogAtLsn(vec![(
7462 2 : Lsn(0x60),
7463 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7464 2 : )]),
7465 2 : ),
7466 2 : ],
7467 2 : above_horizon: KeyLogAtLsn(vec![
7468 2 : (
7469 2 : Lsn(0x70),
7470 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7471 2 : ),
7472 2 : (
7473 2 : Lsn(0x80),
7474 2 : Value::Image(Bytes::copy_from_slice(
7475 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
7476 2 : )),
7477 2 : ),
7478 2 : (
7479 2 : Lsn(0x90),
7480 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
7481 2 : ),
7482 2 : ]),
7483 2 : };
7484 2 : assert_eq!(res, expected_res);
7485 2 :
7486 2 : // We expect GC-compaction to run with the original GC. This would create a situation that
7487 2 : // the original GC algorithm removes some delta layers b/c there are full image coverage,
7488 2 : // therefore causing some keys to have an incomplete history below the lowest retain LSN.
7489 2 : // For example, we have
7490 2 : // ```plain
7491 2 : // init delta @ 0x10, image @ 0x20, delta @ 0x30 (gc_horizon), image @ 0x40.
7492 2 : // ```
7493 2 : // Now the GC horizon moves up, and we have
7494 2 : // ```plain
7495 2 : // init delta @ 0x10, image @ 0x20, delta @ 0x30, image @ 0x40 (gc_horizon)
7496 2 : // ```
7497 2 : // The original GC algorithm kicks in, and removes delta @ 0x10, image @ 0x20.
7498 2 : // We will end up with
7499 2 : // ```plain
7500 2 : // delta @ 0x30, image @ 0x40 (gc_horizon)
7501 2 : // ```
7502 2 : // Now we run the GC-compaction, and this key does not have a full history.
7503 2 : // We should be able to handle this partial history and drop everything before the
7504 2 : // gc_horizon image.
7505 2 :
7506 2 : let history = vec![
7507 2 : (
7508 2 : key,
7509 2 : Lsn(0x20),
7510 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7511 2 : ),
7512 2 : (
7513 2 : key,
7514 2 : Lsn(0x30),
7515 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
7516 2 : ),
7517 2 : (
7518 2 : key,
7519 2 : Lsn(0x40),
7520 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
7521 2 : ),
7522 2 : (
7523 2 : key,
7524 2 : Lsn(0x50),
7525 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
7526 2 : ),
7527 2 : (
7528 2 : key,
7529 2 : Lsn(0x60),
7530 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7531 2 : ),
7532 2 : (
7533 2 : key,
7534 2 : Lsn(0x70),
7535 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7536 2 : ),
7537 2 : (
7538 2 : key,
7539 2 : Lsn(0x80),
7540 2 : Value::Image(Bytes::copy_from_slice(
7541 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
7542 2 : )),
7543 2 : ),
7544 2 : (
7545 2 : key,
7546 2 : Lsn(0x90),
7547 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
7548 2 : ),
7549 2 : ];
7550 2 : let res = tline
7551 2 : .generate_key_retention(key, &history, Lsn(0x60), &[Lsn(0x40), Lsn(0x50)], 3, None)
7552 2 : .await
7553 2 : .unwrap();
7554 2 : let expected_res = KeyHistoryRetention {
7555 2 : below_horizon: vec![
7556 2 : (
7557 2 : Lsn(0x40),
7558 2 : KeyLogAtLsn(vec![(
7559 2 : Lsn(0x40),
7560 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
7561 2 : )]),
7562 2 : ),
7563 2 : (
7564 2 : Lsn(0x50),
7565 2 : KeyLogAtLsn(vec![(
7566 2 : Lsn(0x50),
7567 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
7568 2 : )]),
7569 2 : ),
7570 2 : (
7571 2 : Lsn(0x60),
7572 2 : KeyLogAtLsn(vec![(
7573 2 : Lsn(0x60),
7574 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7575 2 : )]),
7576 2 : ),
7577 2 : ],
7578 2 : above_horizon: KeyLogAtLsn(vec![
7579 2 : (
7580 2 : Lsn(0x70),
7581 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7582 2 : ),
7583 2 : (
7584 2 : Lsn(0x80),
7585 2 : Value::Image(Bytes::copy_from_slice(
7586 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
7587 2 : )),
7588 2 : ),
7589 2 : (
7590 2 : Lsn(0x90),
7591 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
7592 2 : ),
7593 2 : ]),
7594 2 : };
7595 2 : assert_eq!(res, expected_res);
7596 2 :
7597 2 : // In case of branch compaction, the branch itself does not have the full history, and we need to provide
7598 2 : // the ancestor image in the test case.
7599 2 :
7600 2 : let history = vec![
7601 2 : (
7602 2 : key,
7603 2 : Lsn(0x20),
7604 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7605 2 : ),
7606 2 : (
7607 2 : key,
7608 2 : Lsn(0x30),
7609 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
7610 2 : ),
7611 2 : (
7612 2 : key,
7613 2 : Lsn(0x40),
7614 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
7615 2 : ),
7616 2 : (
7617 2 : key,
7618 2 : Lsn(0x70),
7619 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7620 2 : ),
7621 2 : ];
7622 2 : let res = tline
7623 2 : .generate_key_retention(
7624 2 : key,
7625 2 : &history,
7626 2 : Lsn(0x60),
7627 2 : &[],
7628 2 : 3,
7629 2 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
7630 2 : )
7631 2 : .await
7632 2 : .unwrap();
7633 2 : let expected_res = KeyHistoryRetention {
7634 2 : below_horizon: vec![(
7635 2 : Lsn(0x60),
7636 2 : KeyLogAtLsn(vec![(
7637 2 : Lsn(0x60),
7638 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")), // use the ancestor image to reconstruct the page
7639 2 : )]),
7640 2 : )],
7641 2 : above_horizon: KeyLogAtLsn(vec![(
7642 2 : Lsn(0x70),
7643 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7644 2 : )]),
7645 2 : };
7646 2 : assert_eq!(res, expected_res);
7647 2 :
7648 2 : let history = vec![
7649 2 : (
7650 2 : key,
7651 2 : Lsn(0x20),
7652 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7653 2 : ),
7654 2 : (
7655 2 : key,
7656 2 : Lsn(0x40),
7657 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
7658 2 : ),
7659 2 : (
7660 2 : key,
7661 2 : Lsn(0x60),
7662 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7663 2 : ),
7664 2 : (
7665 2 : key,
7666 2 : Lsn(0x70),
7667 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7668 2 : ),
7669 2 : ];
7670 2 : let res = tline
7671 2 : .generate_key_retention(
7672 2 : key,
7673 2 : &history,
7674 2 : Lsn(0x60),
7675 2 : &[Lsn(0x30)],
7676 2 : 3,
7677 2 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
7678 2 : )
7679 2 : .await
7680 2 : .unwrap();
7681 2 : let expected_res = KeyHistoryRetention {
7682 2 : below_horizon: vec![
7683 2 : (
7684 2 : Lsn(0x30),
7685 2 : KeyLogAtLsn(vec![(
7686 2 : Lsn(0x20),
7687 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7688 2 : )]),
7689 2 : ),
7690 2 : (
7691 2 : Lsn(0x60),
7692 2 : KeyLogAtLsn(vec![(
7693 2 : Lsn(0x60),
7694 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x40;0x60")),
7695 2 : )]),
7696 2 : ),
7697 2 : ],
7698 2 : above_horizon: KeyLogAtLsn(vec![(
7699 2 : Lsn(0x70),
7700 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7701 2 : )]),
7702 2 : };
7703 2 : assert_eq!(res, expected_res);
7704 2 :
7705 2 : Ok(())
7706 2 : }
7707 :
7708 : #[tokio::test]
7709 2 : async fn test_simple_bottom_most_compaction_with_retain_lsns() -> anyhow::Result<()> {
7710 2 : let harness =
7711 2 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns").await?;
7712 8 : let (tenant, ctx) = harness.load().await;
7713 2 :
7714 438 : fn get_key(id: u32) -> Key {
7715 438 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7716 438 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7717 438 : key.field6 = id;
7718 438 : key
7719 438 : }
7720 2 :
7721 2 : let img_layer = (0..10)
7722 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
7723 2 : .collect_vec();
7724 2 :
7725 2 : let delta1 = vec![
7726 2 : (
7727 2 : get_key(1),
7728 2 : Lsn(0x20),
7729 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7730 2 : ),
7731 2 : (
7732 2 : get_key(2),
7733 2 : Lsn(0x30),
7734 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7735 2 : ),
7736 2 : (
7737 2 : get_key(3),
7738 2 : Lsn(0x28),
7739 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
7740 2 : ),
7741 2 : (
7742 2 : get_key(3),
7743 2 : Lsn(0x30),
7744 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7745 2 : ),
7746 2 : (
7747 2 : get_key(3),
7748 2 : Lsn(0x40),
7749 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
7750 2 : ),
7751 2 : ];
7752 2 : let delta2 = vec![
7753 2 : (
7754 2 : get_key(5),
7755 2 : Lsn(0x20),
7756 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7757 2 : ),
7758 2 : (
7759 2 : get_key(6),
7760 2 : Lsn(0x20),
7761 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7762 2 : ),
7763 2 : ];
7764 2 : let delta3 = vec![
7765 2 : (
7766 2 : get_key(8),
7767 2 : Lsn(0x48),
7768 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
7769 2 : ),
7770 2 : (
7771 2 : get_key(9),
7772 2 : Lsn(0x48),
7773 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
7774 2 : ),
7775 2 : ];
7776 2 :
7777 2 : let tline = tenant
7778 2 : .create_test_timeline_with_layers(
7779 2 : TIMELINE_ID,
7780 2 : Lsn(0x10),
7781 2 : DEFAULT_PG_VERSION,
7782 2 : &ctx,
7783 2 : vec![
7784 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
7785 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
7786 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
7787 2 : ], // delta layers
7788 2 : vec![(Lsn(0x10), img_layer)], // image layers
7789 2 : Lsn(0x50),
7790 2 : )
7791 47 : .await?;
7792 2 : {
7793 2 : // Update GC info
7794 2 : let mut guard = tline.gc_info.write().unwrap();
7795 2 : *guard = GcInfo {
7796 2 : retain_lsns: vec![
7797 2 : (Lsn(0x10), tline.timeline_id),
7798 2 : (Lsn(0x20), tline.timeline_id),
7799 2 : ],
7800 2 : cutoffs: GcCutoffs {
7801 2 : time: Lsn(0x30),
7802 2 : space: Lsn(0x30),
7803 2 : },
7804 2 : leases: Default::default(),
7805 2 : within_ancestor_pitr: false,
7806 2 : };
7807 2 : }
7808 2 :
7809 2 : let expected_result = [
7810 2 : Bytes::from_static(b"value 0@0x10"),
7811 2 : Bytes::from_static(b"value 1@0x10@0x20"),
7812 2 : Bytes::from_static(b"value 2@0x10@0x30"),
7813 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
7814 2 : Bytes::from_static(b"value 4@0x10"),
7815 2 : Bytes::from_static(b"value 5@0x10@0x20"),
7816 2 : Bytes::from_static(b"value 6@0x10@0x20"),
7817 2 : Bytes::from_static(b"value 7@0x10"),
7818 2 : Bytes::from_static(b"value 8@0x10@0x48"),
7819 2 : Bytes::from_static(b"value 9@0x10@0x48"),
7820 2 : ];
7821 2 :
7822 2 : let expected_result_at_gc_horizon = [
7823 2 : Bytes::from_static(b"value 0@0x10"),
7824 2 : Bytes::from_static(b"value 1@0x10@0x20"),
7825 2 : Bytes::from_static(b"value 2@0x10@0x30"),
7826 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
7827 2 : Bytes::from_static(b"value 4@0x10"),
7828 2 : Bytes::from_static(b"value 5@0x10@0x20"),
7829 2 : Bytes::from_static(b"value 6@0x10@0x20"),
7830 2 : Bytes::from_static(b"value 7@0x10"),
7831 2 : Bytes::from_static(b"value 8@0x10"),
7832 2 : Bytes::from_static(b"value 9@0x10"),
7833 2 : ];
7834 2 :
7835 2 : let expected_result_at_lsn_20 = [
7836 2 : Bytes::from_static(b"value 0@0x10"),
7837 2 : Bytes::from_static(b"value 1@0x10@0x20"),
7838 2 : Bytes::from_static(b"value 2@0x10"),
7839 2 : Bytes::from_static(b"value 3@0x10"),
7840 2 : Bytes::from_static(b"value 4@0x10"),
7841 2 : Bytes::from_static(b"value 5@0x10@0x20"),
7842 2 : Bytes::from_static(b"value 6@0x10@0x20"),
7843 2 : Bytes::from_static(b"value 7@0x10"),
7844 2 : Bytes::from_static(b"value 8@0x10"),
7845 2 : Bytes::from_static(b"value 9@0x10"),
7846 2 : ];
7847 2 :
7848 2 : let expected_result_at_lsn_10 = [
7849 2 : Bytes::from_static(b"value 0@0x10"),
7850 2 : Bytes::from_static(b"value 1@0x10"),
7851 2 : Bytes::from_static(b"value 2@0x10"),
7852 2 : Bytes::from_static(b"value 3@0x10"),
7853 2 : Bytes::from_static(b"value 4@0x10"),
7854 2 : Bytes::from_static(b"value 5@0x10"),
7855 2 : Bytes::from_static(b"value 6@0x10"),
7856 2 : Bytes::from_static(b"value 7@0x10"),
7857 2 : Bytes::from_static(b"value 8@0x10"),
7858 2 : Bytes::from_static(b"value 9@0x10"),
7859 2 : ];
7860 2 :
7861 10 : let verify_result = || async {
7862 10 : let gc_horizon = {
7863 10 : let gc_info = tline.gc_info.read().unwrap();
7864 10 : gc_info.cutoffs.time
7865 10 : };
7866 110 : for idx in 0..10 {
7867 100 : assert_eq!(
7868 100 : tline
7869 100 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7870 123 : .await
7871 100 : .unwrap(),
7872 100 : &expected_result[idx]
7873 10 : );
7874 100 : assert_eq!(
7875 100 : tline
7876 100 : .get(get_key(idx as u32), gc_horizon, &ctx)
7877 77 : .await
7878 100 : .unwrap(),
7879 100 : &expected_result_at_gc_horizon[idx]
7880 10 : );
7881 100 : assert_eq!(
7882 100 : tline
7883 100 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
7884 66 : .await
7885 100 : .unwrap(),
7886 100 : &expected_result_at_lsn_20[idx]
7887 10 : );
7888 100 : assert_eq!(
7889 100 : tline
7890 100 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
7891 53 : .await
7892 100 : .unwrap(),
7893 100 : &expected_result_at_lsn_10[idx]
7894 10 : );
7895 10 : }
7896 10 : };
7897 2 :
7898 70 : verify_result().await;
7899 2 :
7900 2 : let cancel = CancellationToken::new();
7901 73 : tline.compact_with_gc(&cancel, &ctx).await.unwrap();
7902 72 : verify_result().await;
7903 2 :
7904 2 : // compact again
7905 55 : tline.compact_with_gc(&cancel, &ctx).await.unwrap();
7906 63 : verify_result().await;
7907 2 :
7908 2 : // increase GC horizon and compact again
7909 2 : {
7910 2 : // Update GC info
7911 2 : let mut guard = tline.gc_info.write().unwrap();
7912 2 : guard.cutoffs.time = Lsn(0x38);
7913 2 : guard.cutoffs.space = Lsn(0x38);
7914 2 : }
7915 44 : tline.compact_with_gc(&cancel, &ctx).await.unwrap();
7916 57 : verify_result().await; // no wals between 0x30 and 0x38, so we should obtain the same result
7917 2 :
7918 2 : // not increasing the GC horizon and compact again
7919 44 : tline.compact_with_gc(&cancel, &ctx).await.unwrap();
7920 57 : verify_result().await;
7921 2 :
7922 2 : Ok(())
7923 2 : }
7924 :
7925 : #[tokio::test]
7926 2 : async fn test_simple_bottom_most_compaction_on_branch() -> anyhow::Result<()> {
7927 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_on_branch").await?;
7928 8 : let (tenant, ctx) = harness.load().await;
7929 2 :
7930 126 : fn get_key(id: u32) -> Key {
7931 126 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7932 126 : key.field6 = id;
7933 126 : key
7934 126 : }
7935 2 :
7936 2 : let img_layer = (0..10)
7937 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
7938 2 : .collect_vec();
7939 2 :
7940 2 : let delta1 = vec![
7941 2 : (
7942 2 : get_key(1),
7943 2 : Lsn(0x20),
7944 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7945 2 : ),
7946 2 : (
7947 2 : get_key(2),
7948 2 : Lsn(0x30),
7949 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7950 2 : ),
7951 2 : (
7952 2 : get_key(3),
7953 2 : Lsn(0x28),
7954 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
7955 2 : ),
7956 2 : (
7957 2 : get_key(3),
7958 2 : Lsn(0x30),
7959 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7960 2 : ),
7961 2 : (
7962 2 : get_key(3),
7963 2 : Lsn(0x40),
7964 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
7965 2 : ),
7966 2 : ];
7967 2 : let delta2 = vec![
7968 2 : (
7969 2 : get_key(5),
7970 2 : Lsn(0x20),
7971 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7972 2 : ),
7973 2 : (
7974 2 : get_key(6),
7975 2 : Lsn(0x20),
7976 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7977 2 : ),
7978 2 : ];
7979 2 : let delta3 = vec![
7980 2 : (
7981 2 : get_key(8),
7982 2 : Lsn(0x48),
7983 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
7984 2 : ),
7985 2 : (
7986 2 : get_key(9),
7987 2 : Lsn(0x48),
7988 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
7989 2 : ),
7990 2 : ];
7991 2 :
7992 2 : let parent_tline = tenant
7993 2 : .create_test_timeline_with_layers(
7994 2 : TIMELINE_ID,
7995 2 : Lsn(0x10),
7996 2 : DEFAULT_PG_VERSION,
7997 2 : &ctx,
7998 2 : vec![], // delta layers
7999 2 : vec![(Lsn(0x18), img_layer)], // image layers
8000 2 : Lsn(0x18),
8001 2 : )
8002 29 : .await?;
8003 2 :
8004 2 : parent_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
8005 2 :
8006 2 : let branch_tline = tenant
8007 2 : .branch_timeline_test_with_layers(
8008 2 : &parent_tline,
8009 2 : NEW_TIMELINE_ID,
8010 2 : Some(Lsn(0x18)),
8011 2 : &ctx,
8012 2 : vec![
8013 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
8014 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
8015 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
8016 2 : ], // delta layers
8017 2 : vec![], // image layers
8018 2 : Lsn(0x50),
8019 2 : )
8020 18 : .await?;
8021 2 :
8022 2 : branch_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
8023 2 :
8024 2 : {
8025 2 : // Update GC info
8026 2 : let mut guard = parent_tline.gc_info.write().unwrap();
8027 2 : *guard = GcInfo {
8028 2 : retain_lsns: vec![(Lsn(0x18), branch_tline.timeline_id)],
8029 2 : cutoffs: GcCutoffs {
8030 2 : time: Lsn(0x10),
8031 2 : space: Lsn(0x10),
8032 2 : },
8033 2 : leases: Default::default(),
8034 2 : within_ancestor_pitr: false,
8035 2 : };
8036 2 : }
8037 2 :
8038 2 : {
8039 2 : // Update GC info
8040 2 : let mut guard = branch_tline.gc_info.write().unwrap();
8041 2 : *guard = GcInfo {
8042 2 : retain_lsns: vec![(Lsn(0x40), branch_tline.timeline_id)],
8043 2 : cutoffs: GcCutoffs {
8044 2 : time: Lsn(0x50),
8045 2 : space: Lsn(0x50),
8046 2 : },
8047 2 : leases: Default::default(),
8048 2 : within_ancestor_pitr: false,
8049 2 : };
8050 2 : }
8051 2 :
8052 2 : let expected_result_at_gc_horizon = [
8053 2 : Bytes::from_static(b"value 0@0x10"),
8054 2 : Bytes::from_static(b"value 1@0x10@0x20"),
8055 2 : Bytes::from_static(b"value 2@0x10@0x30"),
8056 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
8057 2 : Bytes::from_static(b"value 4@0x10"),
8058 2 : Bytes::from_static(b"value 5@0x10@0x20"),
8059 2 : Bytes::from_static(b"value 6@0x10@0x20"),
8060 2 : Bytes::from_static(b"value 7@0x10"),
8061 2 : Bytes::from_static(b"value 8@0x10@0x48"),
8062 2 : Bytes::from_static(b"value 9@0x10@0x48"),
8063 2 : ];
8064 2 :
8065 2 : let expected_result_at_lsn_40 = [
8066 2 : Bytes::from_static(b"value 0@0x10"),
8067 2 : Bytes::from_static(b"value 1@0x10@0x20"),
8068 2 : Bytes::from_static(b"value 2@0x10@0x30"),
8069 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
8070 2 : Bytes::from_static(b"value 4@0x10"),
8071 2 : Bytes::from_static(b"value 5@0x10@0x20"),
8072 2 : Bytes::from_static(b"value 6@0x10@0x20"),
8073 2 : Bytes::from_static(b"value 7@0x10"),
8074 2 : Bytes::from_static(b"value 8@0x10"),
8075 2 : Bytes::from_static(b"value 9@0x10"),
8076 2 : ];
8077 2 :
8078 4 : let verify_result = || async {
8079 44 : for idx in 0..10 {
8080 40 : assert_eq!(
8081 40 : branch_tline
8082 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8083 59 : .await
8084 40 : .unwrap(),
8085 40 : &expected_result_at_gc_horizon[idx]
8086 4 : );
8087 40 : assert_eq!(
8088 40 : branch_tline
8089 40 : .get(get_key(idx as u32), Lsn(0x40), &ctx)
8090 32 : .await
8091 40 : .unwrap(),
8092 40 : &expected_result_at_lsn_40[idx]
8093 4 : );
8094 4 : }
8095 4 : };
8096 2 :
8097 46 : verify_result().await;
8098 2 :
8099 2 : let cancel = CancellationToken::new();
8100 35 : branch_tline.compact_with_gc(&cancel, &ctx).await.unwrap();
8101 2 :
8102 45 : verify_result().await;
8103 2 :
8104 2 : Ok(())
8105 2 : }
8106 : }
|
